arxiv_id stringlengths 9 15 | arxiv_category stringclasses 5 values | category stringclasses 2 values | text stringlengths 0 1.19M |
|---|---|---|---|
0705.0815 | cs.NI | cyber | # Netsukuku
Close the world,
Open the next
http://netsukuku.freaknet.org
###### Abstract
Netsukuku is a P2P network system designed to handle a large number of nodes with minimal CPU and memory resources. It can be easily used to build a worldwide distributed, anonymous and not controlled network, separated from the Internet, without the support of any servers, ISPs or authority controls.
In this document, we give a generic and non technical description of the Netsukuku network, emphasizing its main ideas and features.
## 1 The old wired
The Internet is a hierarchic network managed by multinational companies and organisations supported by governments. Each bit of Internet traffic passes through proprietary backbones and routers.
The Internet Service Providers give the connectivity to all the users, who are in the lowest rank of this hierarchic pyramid. For this reason, the Internet is not a global network created by the users and shared between them. The people can join the Net only in accordance with the restrictive conditions and terms imposed by the Majors.
The Internet represents, today, the means to access information, knowledge and communication. About 1 billion of people can connect to this great proprietary network, but the remaining 5 billion of people, which don’t have enough economic resource, are still waiting the multinationals to supply a service within their reach.
The Internet was born with the intent of warranting a secure and unattackable communication between the various nodes of the network, but now, paradoxically, an ISP has the power to cut out of the Internet entire nations by simply stopping to give its services.
Beside that, Internet is not anonymous: the ISP can trace back and analyse the traffic of data going through their servers, without any limits.
The centralised and hierarchical structure of the Internet creates, as a consequence, other identical systems, based on it, i.e. the DNS. The servers of the Domain Name System are managed by different ISPs, as well and the domains are literally sold through a similar centralised system. This kind of structures allows, in a very simple and efficient way, to physically localise any computers, which are connected to the Internet, in a very short time and without any particular efforts.
In China, the whole net is constantly watched by several computers filtering the Internet traffic: a Chinese will never be able to see or came to know about a site containing some keywords, such as “democracy”, censored by the government. He’ll never be able to express his own ideas on the net, e.g. about his government’s policy, without risking till the death penalty.
Internet was born to satisfy the military needs of security for the administration of the American defence, not to ensure freedom of communication and information: in order to communicate between each other the Internet users are obliged to submit themselves to the control and to the support of big multinationals, whose only aim is to expand their own hegemony.
As long as all the efforts to bring more freedom, privacy and accessibility in the Internet face aversions, fears, contrary interests of governments and private companies, the very alternative solution to this problem is to let the users migrate toward a distributed, decentralised and fully efficient net, in which all the users interact at the same level, with no privilege and no conditioning means, as authentic citizens of a free world wide community.
## 2 The Netsukuku wired
Netsukuku is a mesh network or a p2p net system that generates and sustains itself autonomously. It is designed to handle an unlimited number of nodes with minimal CPU and memory resources. Thanks to this feature it can be easily used to build a worldwide distributed, anonymous and not controlled network, separated from the Internet, without the support of any servers, ISPs or control authorities.
This net is composed by computers linked physically each other, therefore it isn’t build upon any existing network. Netsukuku builds only the routes which connects all the computers of the net.
In other words, Netsukuku replaces the level 3 of the model iso/osi with another routing protocol.
Being Netsukuku a distributed and decentralised net, it is possible to implement real distributed systems on it, e.g. the Abnormal Netsukuku Domain Name Anarchy (ANDNA)\[[4](#bib.bib4)\] which will replace the actual hierarchic and centralised system of DNS.
### 2.1 Gandhi
The main characteristic of Netsukuku is its self-management: the network dynamically configures itself without any external interventions. All the nodes share the same privileges and limitations, giving their contribution to sustain and expand Netsukuku. The more the nodes increase in number the more the net grows and becomes efficient.
The total decentralisation and distribution allows Netsukuku to be neither controlled nor destroyed: the only way to manipulate or demolish it, is to knock physically down each single node composing the network.
### 2.2 No name, no identity
Inside Netsukuku everyone, in any place, at any moment, can hook immediately to the net without coming trough any bureaucratic or legal compliance.
Every elements of the net is extremely dynamic and it’s never the same. The IP address identifying a computer is chosen randomly, therefore it’s extremely difficult to associate it to a particular physical place; furthermore since the routes are composed by a huge number of nodes, it becomes a titanic task to trace a specific node.
The traffic of the nodes is protected by a complete cryptographic layer\[[5](#bib.bib5)\], which ensures a strong anonymity and security.
### 2.3 So, what is it?
Netsukuku is a mesh network built on top of its own dynamic routing protocol.
Currently there are a wide number of dynamic routing protocols, but they are solely utilised to create small and medium nets. The routers of Internet are also managed by different protocols such as the OSPF, the RIP, or the BGP. They are based on different classical graph algorithms, which are able to find out the best path to reach a node in a given graph. However, all these protocols require a very high waste of CPU and memory. Indeed, the Internet routers are computers specifically dedicated to the executions of these algorithms. It would be impossible to implement one of these protocols to create a maintain a mesh network large as the whole Internet.
The Netsukuku protocol structures the entire net as a fractal\[[3](#bib.bib3)\] and, in order to calculate all the needed routes it makes use of a particular algorithm called Quantum Shortest Path Netsukuku (QSPN)\[[2](#bib.bib2)\].
A fractal is a structure which can be compressed infinitely, because every part of itself is composed by the same fractal. Thus its high compression level gives the ability to store the whole Netsukuku map in just few Kilobytes.
On the other hand, the QSPN is an algorithm that has to be executed by the network itself. The nodes, in order to execute it, have just to send and receive the Tracer Packets, without using heavy computational resources.
### 2.4 Netsukuku the wireless
The best medium to establish physical connections between nodes is the wifi. When Netsukuku will be widely adopted, its users will have just to set up their own wifi antennas on a well exposed place (i.e. windows or roofs), linking themselves to the other Netsukuku users, placed within their radio ranges. Nowadays, there is a great variety of wifi technologies which allow to link two nodes distant kilometres each other. A whole city can be easily covered by placing just one node in each neighbourhood.
Moreover, it is possible to use virtual tunnels over the Internet to temporarily replace missing physical links. This subject is described in \[[6](#bib.bib6)\].
## References
- \[1\] Netsukuku website: http://netsukuku.freaknet.org/
- \[2\] QSPN document: http://netsukuku.freaknet.org/doc/main_doc/qspn.pdf
- \[3\] Netsukuku topology document: http://netsukuku.freaknet.org/doc/main_doc/topology.pdf
- \[4\] ANDNA document: http://netsukuku.freaknet.org/doc/main_doc/andna.pdf
- \[5\] Carciofo NTK_RFC: http://lab.dyne.org/Ntk_carciofo
- \[6\] Internet and Netsukuku: http://netsukuku.freaknet.org/doc/main_doc/inetntk.pdf
`^_^` |
0705.2786 | cs.OS | cyber | <sup>1</sup>
<sup>1</sup>institutetext: Faculty of Mathematics, Physics and Chemistry, Silesian University, Bankowa 14, 40-007 Katowice, Poland. Email:
<sup>1</sup>
<sup>1</sup>email: jjw@us.edu.pl
# Virtualization: A double-edged sword
Joachim J. Włodarz
###### Abstract
Virtualization became recently a hot topic once again, after being dormant for more than twenty years. In the meantime, it has been almost forgotten, that virtual machines are not so perfect isolating environments as it seems, when looking at the principles. These lessons were already learnt earlier when the first virtualized systems have been exposed to real life usage. Contemporary virtualization software enables instant creation and destruction of virtual machines on a host, live migration from one host to another, execution history manipulation, etc. These features are very useful in practice, but also causing headaches among security specialists, especially in current hostile network environments. In the present contribution we discuss the principles, potential benefits and risks of virtualization in a déjà vu perspective, related to previous experiences with virtualization in the mainframe era.
## 1 Introduction
Recently, a vigorous interest in various virtualization techniques has been observed. Rather than being only a buzzword, virtualization became since 2005 a main direction in the evolution of the whole IT market. According to recent predictions by the Gartner Group, operating system virtualization and I/O virtualization will become mainstream by 2010 \[[1](#bib.bib1)\].
Traditionally, the leadership in providing virtualization based solutions belonged to IBM, beginning with the VM/370 operating system for System/370 mainframes introduced in 1972 \[[2](#bib.bib2)\]. With the advent of relatively cheap microprocessor based computers within the next decade, the IT industry shifted their interests to personal computing, based on individual hardware computers rather than some virtual machines. But at the beginning of the 21st century, most personal workstations and low-end servers became powerful enough to carry multiple operating systems running concurrently on the same hardware, which rised the interests in virtualization once again, this time on the commodity x86 platform. Moreover, it turned out from experience that setting up an extensive IT infrastructure with a lot of dedicated hardware put together with the appropriate, also dedicated software, is highly inefficient and also very expensive in management. Virtualization, leading generally to decoupling of software from the underlying hardware, seems to have the potential for substantial improvements in such situations.
In the meantime, virtualization software and virtualized systems became already a target of new security threats, with possibly disastrous consequences. In contrary to the widespread belief that, e.g., virtual machines should be way more secure than their physical equivalents, virtualized computer systems may be sometimes less secure and even create new security challenges.
In the present contribution we discuss the benefits as well as the risks and problems of employing virtualization based IT solutions. The discussion is undoubtedly biased due to the authors’ experiences with virtualization, first on various IBM mainframes many years ago and more recently also on the x86 platform, where many problems appeared as déjà vu. As already emphasized in the title, virtualization is like a double-edged sword: a powerful weapon, but also quite dangerous for the warrior.
## 2 Background
When browsing through the literature, one could find several definitions for virtualization, depending on the particular subject and its context. Generally, virtualization creates new entities in a computer system which are substitutes for the real ones, with the same functionality, interfaces and behavior, except the timings which are usually different. The virtual resources, created through such a substitution process, have to be eventually mapped to appropriate real resources, possibly going through more than one level of indirection, if necessary.
The first and also the most important concept of a virtual entity in computer systems was virtual memory, introduced and developed around 1956 by Fritz-Rudolf Güntsch \[[4](#bib.bib4)\] to unify various kinds of memory devices within an experimental computer system built at TU Berlin. A couple of years later the first commercial computer with memory virtualization onboard was already built: the famous Atlas Computer \[[5](#bib.bib5)\] developed at the University of Manchester with collaboration of Ferranti/ICL. Although influential, the Atlas computer was known not to be working well in practical environments, mostly due to limited performance of the hardware available at this time.
A more systematic approach to virtualization has been taken by IBM engineers working on the System/360, the first computer family ”architected” according to common, precisely defined principles of operation. In order to develop an interactive, multi-user system based on the already existing S/360 operating systems, facilitate time-sharing and at the same time also protect the users from each other, they decided to provide each user with a dedicated ”pseudo-machine”, a fake System/360 computer capable to run any other S/360 operating system. The term ”pseudo-machine” has been renamed soon to ”virtual machine”, when it turned out that the latter term has been already used earlier at IBM Research for an M44/44x experimental system, aimed at providing partially virtualized “7044-like” virtual machines on a modified IBM 7044 36-bit mainframe \[[6](#bib.bib6)\].
The intended full hardware virtualization per time-sharing user was quite demanding and needed an appropriate support from the hardware itself to run with acceptable performance. Such hardware support appeared in only one System/360 computer: the famous System/360 Model 67. Regular hardware support for virtualization became reality a couple of years later, within the System/370 family and not without exercising some pressure from the “VM community” inside and outside IBM \[[7](#bib.bib7)\]. The first officially released version of VM/370 was able to provide an exact functional copy of the underlying System/370 hardware, according to the virtual machine configuration defined in the system directory. A particular user could even run another copy of VM/370 on the virtual hardware, e.g., to test a new version of system software without disrupting the running system.
The architecture of a “classical” VM/370 system, nowadays known rather as hypervisor Type I architecture according to a widely accepted classification, introduced by Robert P. Goldberg in his PhD thesis \[[8](#bib.bib8)\], is sketched in Fig.[1](#S2.F1 "Figure 1 ‣ 2 Background ‣ Virtualization: A double-edged sword") below.
[Refer to caption](/html/0705.2786/assets/x1.png)
Figure 1: The “classical” VM/370 hypervisor (Type I)
The VM/370 hypervisor, or Control Program (CP) in the original IBM terminology, represents here the software layer implementing the virtual machines, each having exactly the same instruction set architecture (ISA) as the underlying System/370 hardware. This approach enabled a peaceful coexistence of several operating systems with different needs and applications, running in parallel on the same hardware machine. In the example presented in Fig.[1](#S2.F1 "Figure 1 ‣ 2 Background ‣ Virtualization: A double-edged sword"), one virtual machine runs CMS, a simple operating system intended for individual usage, side-by-side with the mammoth-size MVS, a general purpose multitasking and multiuser operating system and even a second-level VM/370, running concurrently with its own virtual machines. A direct communication with the hypervisor was possible, and sometimes also necessary, through the built-in CP command interface set up for each virtual machine according to the assigned priviledge level. The seemingly perfectly isolated virtual machines, running under the VM/370 hypervisor with carefully assigned priviledges, proved in practice to be totally insecure for intended intrusions, enabling even a total system penetration from a running VM, with direct access to the real hardware and in the real supervisor state \[[9](#bib.bib9)\].
Virtual computer systems could be also created via emulation, where appropriate software imitates another computer system or its part, e.g. CPU. In principle, any operating environment could be emulated within another operating environment, including exact copy of the hardware used by the operating system executing the emulator software itself. Emulators were originally intended to permit the execution of programs written for another computer \[[10](#bib.bib10)\], i.e. for another ISA than employed on the host computer system, but there is certainly nothing wrong with emulation of the same ISA and the related I/O hardware, giving effectively another kind of hypervisor: the so called Type II hypervisor \[[8](#bib.bib8)\].
Emulators and Type II hypervisors may run as ordinary applications within the host computer operating system, being totally decoupled from the real hardware. Although it may be desirable e.g. for debugging purposes, it means usually a very slow execution within the virtual machine. A considerable speedup could be provided by the help of real hardware, made available in parallel to the operating system services, usually through a specialized “Hybrid Virtualization” host operating system driver, see the sketch in Fig.[2](#S2.F2 "Figure 2 ‣ 2 Background ‣ Virtualization: A double-edged sword").
[Refer to caption](/html/0705.2786/assets/x2.png)
Figure 2: A Type II hypervisor
Yet another approach to virtualization, a so called container-based virtualization, is based on the mechanisms employed by contemporary operating systems to establish a stardard execution environment for processes: the “process virtual machine”. Augmented with mechanisms for hiding and/or rewriting system-level data, the operating system kernel itself could be used as a kind of “hypervisor” to confine the processes within well-separated “execution containers”. As shown in Fig.[3](#S2.F3 "Figure 3 ‣ 2 Background ‣ Virtualization: A double-edged sword"), the virtualized system contains in this case a fully priviledged “Host VM” used mainly for system management and several “Guest VMs”, which are seemingly equivalent to separate hosts.
[Refer to caption](/html/0705.2786/assets/x3.png)
Figure 3: Container-based operating system level virtualization
The main advantage of container-based virtualization is high performance and scalability with only a minimal overhead introduced through the virtualization. Many contemporary operating systems offer this possibility as a built-in or addon feature, e.g., Zones/BrandZ in Solaris, Secure Resource Partitions in HP-UX or VServer and OpenVZ/Virtuozzo for Linux.
## 3 x86 Virtualization
The constantly increasing performance of x86-based computers within the last decade rised also the interest in virtualization on this platform. The Intel Pentium “classical” IA-32 architecture has beed designed rather to remain compatible with previous Intel processors than being virtualizable, with exception of the “virtual 86” sub-mode intended for running “Multiple Virtual DOS Machines” in protected mode. Therefore, it is quite challenging to implement virtualization software on this platform (cf. \[[11](#bib.bib11), [12](#bib.bib12)\] and references therein).
Due to the recent high demand for virtualization, the newest Intel and AMD processors include a series of architectural extensions providing hardware support for full virtualization, which simplifies also considerably the implementation of x86-based “true” hypervisors (cf.\[[13](#bib.bib13), [14](#bib.bib14)\]).
The most hardware independent way of virtualization is pure emulation, where no instructions are being passed directly to the host hardware. Several emulators exists for the x86 platform, the most advanced: Bochs \[[15](#bib.bib15)\] and QEMU \[[16](#bib.bib16)\] are available for many host platforms, including x86 itself. QEMU enhanced with kqemu module (a HV driver) works efectively as a Type II hypervisor with almost native speed. The Kernel-based Virtual Machine, available with recent Linux kernels, provides full virtualization support with QEMU user-level frontend \[[17](#bib.bib17)\].
VMware Inc. \[[18](#bib.bib18)\], a pioneer of virtualization on the x86 platform, offers since 1999 a whole line of proprietary x86 virtualization software. The lowest-level products for workstations (VMware Player) and servers (VMware Server, earlier known as GSX Server) are available as freeware for MS Windows and also Linux. Similar virtualization products are offered also by Microsoft: Virtual PC and Virtual Server, which became also freely available last year as direct competitors to VMware products on the MS Windows platform.
The Xen hypervisor, developed initially as research project at the University of Cambridge \[[19](#bib.bib19)\], seems to be the most promising and dynamically evolving open-source virtualization software. Working generally as a Type I hypervisor, Xen provides paravirtualized hardware for its virtual machines when full virtualization is not possible, e.g. when working on a Pentium host.
Paravirtualization enables a cooperation between the guest operating system and the hypervisor to circumvent virtualization problems and increase performance. The operating systems intended for the paravirtualized Xen virtual machines have to be ported to the Xen architecture, which may become problematic in case of proprietary operating systems. On the other hand, paravirtualized operating systems could perform even better than their native versions running directly on the host hardware, because the hypervisor may e.g. warrant apropriate resource allocations \[[20](#bib.bib20)\].
Virtual machines established under the Xen 3 hypervisor may be “live migrated” to another Xen 3 host computer, with only a very short application-level downtime around 100-300 ms in a local network with available network storage \[[21](#bib.bib21)\]. It has been demonstrated recently, that a long-haul live migration is fairly possible even at a planetary level, with downtimes rising only to 1-2 seconds for the distance between Amsterdam and San Diego \[[22](#bib.bib22)\].
## 4 Applications and Benefits
Virtualization technologies can be used for many purposes. In the mainframe era the virtualized resources were localized on a single host computer and accesible only to a relatively small number of users. Today, the possibility of migration of virtual machines from host to host, even in a “live” state, makes a tremendous difference. A virtual machine could be instantly created on some host, then moved elsewhere, e.g., when the host becomes overloaded, or completely destroyed if not needed. With a proper configuration, virtualized environments exhibit a sound level of isolation from each other, hence a crash occuring in one environment should remain unnoticed outside, which is a desired property for testbeds and security systems.
### 4.1 Hardware Replacement
Virtual hardware may easily replace real hardware consuming precious datacenter resouces like floor space, electricity or manpower. This is especially important for sites running older equipment, which may be not easy to maintain, and at some point become even impossible to run, e.g., due to the lack of appropriate spare parts on the market. Virtualization is often used for server consolidation, i.e. a reduction of the total number of servers or even separate datacenters to a required minimum. At the same time, software which became incompatible with the new hardware and/or its configuration could be deployed within specialized virtual machines when needed, e.g., for archival purposes. The possibility to migrate “live machines” across networks gives additional flexibility in load balancing and disaster recovery, enabling seamless movement of workloads between distant datacenters practically without downtimes.
### 4.2 Testing and Debugging
Many virtualization systems, particularily emulators, were invented for testing and debugging purposes, especially in the operating system kernel development, where the experimental system have to be rebooted again and again. Also, getting some data for debugging may be not easy on a real hardware, requiring sometimes arcane equipment. On emulated hardware, the same task could be a trivial recording to some log file. With appropriate builtin hooks, any virtualization environment may serve as a powerful debugger, with possibility of a “time travel”, i.e. arbitrary navigation in the execution history \[[23](#bib.bib23)\].
### 4.3 Education and E-Learning
Virtual hardware is a very attractive possibility to have handy in any educational computer laboratory, because the students may experiment with various operating systems and network setups, without fighting with problems related to running a multiplatform laboratory based on real hardware \[[24](#bib.bib24)\]. For e-Learning purposes, such a “virtual laboratory” could be made accesible remotely, e.g. from a standard web browser \[[25](#bib.bib25)\].
### 4.4 Security Systems
The isolation of applications running within a standard operating system is usually based only of the process abstraction, with many facilities available for data sharing, also in a system-wide manner. Virtual machine technology may be therefore used to additionally isolate sensitive environments and minimize the risk of compromising the entire system due to problems in one of the environments.
Network services are usually regarded as particularily sensitive to security problems, related to the direct contact with the possible hostile network surrounding. Therefore, a separate network frontend virtual machine could be added to a virtualized system to handle the network connections and redirect them appropriately. With intelligent network hardware dedicated to this virtual machine and a minimal hardened operating system with integrated firewall, e.g. OpenBSD with pf, the overhead should be negigible \[[26](#bib.bib26)\].
Much more demanding would be a virtual secure file server, established to enforce a directory and file access control policy, even if the main system become compromised \[[27](#bib.bib27)\]. It has been recently demonstrated that a more elaborated configuration, with several dedicated virtual machines and checkpointing, allows for an almost instant, automatical recovery after intrusion detection \[[28](#bib.bib28)\].
## 5 The Other Edge of Virtualization
Even in the mainframe era virtual machines were considered not secure enough for hosting sensitive data, mainly due to the I/O handling in VM/370 \[[9](#bib.bib9)\]. It turned out that a “hardening” by closing known security holes does not guarantee the absence of other security flaws in the hardened system, therefore a “security retrofit program” has been started to design a formally verified “security kernel” for VM/370 \[[29](#bib.bib29)\].
There is no doubt that many hypervisors proliferating nowadays on the x86 platform need much more than only closing the detected security holes on the fly. Without similar, specific “security retrofit” programs it would be pretty hard to avoid security problems. As shown in a very recent study by Tavis Ormandy \[[30](#bib.bib30)\], the present state of security exposure from implementation flaws in popular virtualization software is quite alarming. With exception of Xen, all tested packages contained exploitable security flaws, allowing a potential attacker to escape reliably from the confining virtual machine.
The mobility of virtual machines, which could be easily migrated along multiple domains, adds its own vulnerabilities. E.g., a virtual machine may be moved into a faked recipient host (“kidnapped”) or compromised while stored semewhere in transit. The ease of manipulations involving virtual machines, which may be created on the fly, possibly immitating another virtual/physical machine, then (mis)used for some purpose, and subsequently destroyed or moved away could be also a source of many security problems, unknown when working on dedicated hardware computers. Static security arrangements are rather useless in this context.
As demonstrated recently \[[31](#bib.bib31)\], it is perfectly possible to hide malware in a virtual machine running in parallel or force the host operating system into a virtual machine, keeping the malware running outside. With latest hardware enhancements for virtualization, it is also possible to do that on the fly, without disrupting the running operating system \[[32](#bib.bib32), [33](#bib.bib33)\].
## 6 Conclusions
When properly prepared and properly done, virtualization offers many benefits in comparison with equivalent, contemporary available hardware based solutions. In the next few years virtualization will become a standard way of computing in the enterprise, academia and personal use of computers due to substantial savings and enormous flexibility.
The installation of operating systems and applications will quickly become obsolete due to the possibility of preparation and distribution of virtual appliances, tailored to the needs and ready to run on any computer with the right hypervisor onboard.
When using virtualization for security, a security enhanced implementation with multilevel security \[[34](#bib.bib34)\] would be definitely recommended. A secure hypervisor architecture, sHype developed at IBM \[[35](#bib.bib35)\], targeted at the Common Criteria EAL4 assurance level, is now a part of the Xen open-source hypervisor \[[36](#bib.bib36)\]. There are plans \[[37](#bib.bib37)\] for further elevation of the Xen assurance level at EAL5 and then EAL6. A prototype system for enforcing mandatory access control policies across a distributed set of virtual machines has been also tested recently with the Xen hypervisor \[[38](#bib.bib38)\]. Therefore Xen seems to be now the most advanced hypervisor with a proven track record and clear evolution path in virtualization security.
Last but not least, virtual machines are in principle nothing more than big fat programs with a complex inner life, executed on a usually thin hypervisor layer, which in turn is in principle nothing more than a specialized operating system. Looking from this perspective, there is no doubt, that virtualization software have to be always kept up-to-date and possibly equipped with a proper security add-on, exactly like ordinary operating systems. And the virtual appliances/machines have to be treated like ordinary executables, which are potentially corrupted, infected or malicious.
## References
- \[1\]
Enck, J., Dawson, P., Phelps, J.R.: Predicts 2007: Brace Yourself for the Next Wave of Server Technology. Gartner RAS Core Research Note G00143973, 8 November 2006.
- \[2\]
Creasy, R.J.: The origin of the VM/370 time-sharing system. IBM J. Res. Dev. 25:483–490, 1981.
- \[3\]
MacDonald, N.: Securing Virtualization and Virtualizing Security. ITxpo 2007, San Francisco CA, 22–26 April 2007.
- \[4\]
Jessen, E.: Die Entwicklung des virtuellen Speichers. Informatik Spektrum 19:216–219, 1996.
- \[5\]
Sumner, F.H.: The Atlas Computer. In: The First Computers: History and Architectures. , edited by R. Rojas and U. Hashagen, MIT Press, 2000. (ISBN 0-262-18197-5)
- \[6\]
Parmelee, R.P., Peterson, T.I., Tillman, C.C. and Hatfield, D.J.: Virtual storage and virtual machine concepts. IBM Systems J. 11:99–130, 1972.
- \[7\]
Varian, M.: VM and the VM Community: Past, Present, and Future. SHARE 89, August 1997.
- \[8\]
Goldberg, R.P.: Architectural Principles for Virtual Computer Systems. Division of Engineering and Applied Physics, Harvard Univ., Cambridge Mass., 1972. (PhD Thesis)
- \[9\]
Attanasio, C.R., Markstein, P.W. and Philipps, R.J.: Penetrating an Operating System: A Study of VM/370 integrity. IBM Systems J. 15:102–116, 1976.
- \[10\]
Mallach, E.G.: Emulation: A Survey. Honeywell Comp. J. 6:287–297, 1972.
- \[11\]
Robin, J.S., Irvine, C.E: Analysis of the Intel Pentium’s Ability to Support a Secure Virtual Machine Monitor. In: Proc. of the Ninth USENIX Security Symposium, Denver CO, Aug. 14-17, 2000.
- \[12\]
Adams, K., Agesen, O.: A Comparison of Software and Hardware Techniques for x86 Virtualization. Operating Systems Rev. 40(5):2–13, 2006.
- \[13\]
Intel: Intel Virtualization Technology. Intel Technology Journal 10(3) 2006. (whole issue)
- \[14\]
AMD: SVM Architecture Reference Manual. AMD 2005.
- \[15\]
The Bochs Project: Bochs: The Open Source IA-32 Emulation Project. http://bochs.sourceforge.net
- \[16\]
Bellard, F.: QEMU. http://www.qemu.org
- \[17\]
Qumranet: KVM Whitepaper. Qumranet 2006.
- \[18\]
VMware: VMware: Virtualization, Virtual Machine and Virtual Server Consolidation. http://www.vmware.com
- \[19\]
Barham, P., Dragovic, B., Fraser, K., Hand, S., Harris, T., Ho, A., Pratt, I., Warfield, A., and Neugebauer, R.: Xen and The Art of Virtualization. In: Proc. of the ACM Symposium on Operating System Principles, 2003.
- \[20\]
Mergen, M.F., Uhlig, V., Krieger, O. and Xenidis, J.: Virtualization for High-Performance Computing. Operating Systems Review 40(2):8–11, 2006.
- \[21\]
Clark, C., Fraser, K., Hand, S., Hansen, J.G., Jul, E., Limpach, Ch., Pratt, I., and Warfield, A.: Live Migration of Virtual Machines. In: 2nd Symp. on Networked System Design and Implementation, 2005.
- \[22\]
Travostino, F., Daspit, P., Gommans, L., Jog., Ch., de Laat, C., Mambretti, J., Monga, I., van Oudernaarde, B., Raghunath, S., and Wang, P.Y.: Seamless live migration of virtual machines over the MAN/WAN. Future Generation Computer Systems 22:901–907, 2006.
- \[23\]
King, S.T., Dunlap, G.W., and Chen, P.M.: Debugging operating systems with time-traveling virtual machines. In: Proc. of the 2005 Annual USENIX Technical Conference, April 2005.
- \[24\]
Bullers, W.I., Burd, S. and Seazzu, A.F.: Virtual Machines - An Idea Whose Time Has Returned: Application to Network, Security and Database Courses. ACM SIGSE Bulletin 38(1):102–106, 2006.
- \[25\]
Damiani, E., Frati, F., and Rebeccani, D.: The Open Source Virtual Lab: a Case Study. In: Proc. of the Workshop on Free and Open Source Learning Environments and Tools FOSLET06, Lugano CH, 2006.
- \[26\]
Prevelakis, V.: The Virtual Firewall. ;login: The USENIX Magazine 30(6):27–34, 2005.
- \[27\]
Zhao, X., Borders, K., and Prakash, A.: Towards Protecting Sensitive Files in a Compromised System. In: Proc. of the 3rd IEEE Int. Security in Storage Workshop, pp.21–28, 2005.
- \[28\]
Matthews, J.N., Herne, J.J., Deshane, T.M., Jablonski, P.A., Cherian, L.R., and McCabe, M.T.: Data Protection and Rapid Recovery From Attack With A Virtual Private File Server and Virtual Machine Appliances. In: Proc. of the IASTED Int. Conf. on Communication, Network and Information Security, pp.170–181, Nov. 2005.
- \[29\]
Gold, B.D., Linde, R.R., Schaefer, M., and Scheid, J.F.: VM/370 Security Retrofit Program. In: ACM Proc. of the 1977 Annual Conference, pp.411–418, 1977.
- \[30\]
Ormandy, T.: An Empirical Study into the Security Exposure to Hosts of Hostile Virtualized Environments. In: CanSecWest 2007, Vancouver BC, Apr. 18-20, 2007.
- \[31\]
King, S.T., Chen, P.M., Wang, Y-M., Verbowski, C., Wang, H.J., and Lorch, J.R. SubVirt: Implementing malware with virtual machines. In: Proc. of the 2006 IEEE Symposium on Security and Privacy, pp.314-327, May 2006.
- \[32\]
Rutkowska, J. Subverting Vista Kernel for Fun and Profit Black Hat USA 2006
- \[33\]
Zovi, D.D. Hardware Virtualization Based Rootkits Black Hat USA 2006
- \[34\]
Karger, P.A.: Multi-Level Security Requirements for Hypervisors. In: Proc. of the 21st Annual Computer Security Applications Conference, Tucson AZ, pp.267–275, Dec. 5-9, 2005.
- \[35\]
Sailer, R., Valdez, E., Jaeger, T., Perez, R., van Doorn, L. Griffin, J.L., and Berger, S.: sHype: Secure Hypervisor Approach to Trusted Virtualized Systems. IBM Research Report RC23511, 2005.
- \[36\]
Sailer, R., Jaeger, T., Valdez, E., Caceres, R., Perez, R., Berger, S., Griffin, J.L., and van Doorn, L.: Building a MAC-based Security Architecture for the Xen Open-Source Hypervisor. In: Proc. of the 21st Annual Computer Security Applications Conference, Tucson AZ, pp.276–285, Dec. 5-9, 2005.
- \[37\]
McDermott, J., Kang, M.: An Open-Source High Robustness Virtual Machine Monitor. The 22st Annual Computer Security Applications Conference, Miami Beach FL, Dec. 11-15, Works in Progress Session, WiP06-06, 2006.
- \[38\]
McCune, J.M., Jaeger, T., Berger, S., Caceres, R., and Sailer, R.: Shamon: A System for Distributed Mandatory Access Control. In: Proc. of the 22st Annual Computer Security Applications Conference, Miami Beach FL, pp.23–32, Dec. 11-15, 2006. |
0705.0820 | cs.NI | cyber | # Abnormal Netsukuku Domain Name Anarchy
http://netsukuku.freaknet.org
AlpT (@freaknet.org)
###### Abstract
In this document, we present the Abnormal Netsukuku Domain Name Anarchy system. ANDNA is the distributed, non hierarchical and decentralised system of hostname management used in the Netsukuku network.
## 1 Preface
We’re assuming that you already know the basics of the Netsukuku topology. If not, read the topology document first: \[[2](#bib.bib2)\].
## 2 Introduction
ANDNA is the distributed, non hierarchical and decentralised system of hostname management in Netsukuku. It substitutes the DNS.
The ANDNA database is scattered inside all Netsukuku network. In the worst case, every node will have to use few hundred kilobytes of memory.
## 3 ANDNA
ANDNA is based on the principle that every hostname $`h`$ is a string of maximum 512 bytes. By taking the hash $`H\hspace{0pt}{(h)}`$ of the hostname, we obtain a number. We consider this number an IP and we call *hash node* the node which owns the same IP. Every hash node will keep a small portion of the distributed database of ANDNA.
```
Node n
ip: 123.123.123.123
hash( hostname: "netsukuku" ) == 11.22.33.44
||
||
Node i
ip: 11.22.33.44
{ [ Andna database of the node i ] }
{ hash("netsukuku") ---> 123.123.123.123 }
```
The registration of the hostname $`h`$, by the node $`n`$ proceeds as follow:
1. 1\.
$`n`$ computes $`i = {H\hspace{0pt}{(h)}}`$
2. 2\.
$`n`$ sends a registration request to $`i`$
3. 3\.
$`i`$ associates in its database $`H\hspace{0pt}{(h)}`$ to the IP of $`n`$, which is $`I\hspace{0pt}P\hspace{0pt}{(n)}`$.
The resolution of the hostname $`h`$ by the node $`m`$:
1. 1\.
$`m`$ computes $`i = {H\hspace{0pt}{(h)}}`$
2. 2\.
$`m`$ sends a resolution query to $`i`$
3. 3\.
$`i`$ sends back to $`m`$ the IP associated to $`H\hspace{0pt}{(h)}`$, that in this example is $`I\hspace{0pt}P\hspace{0pt}{(n)}`$.
### 3.1 Hash gnode
It is highly probable that the hash $`H\hspace{0pt}{(h)}`$ doesn’t corresponds to any active node: if there are a total of $`2^{3}\hspace{0pt}0`$ active nodes and we are using IP of 32 bits, then by taking the 32 bit hash of $`h`$ the probability is equal to $`25\%`$. Moreover, even the actives node can also exit from the network.
For this reasons, the registration of a hostname will be kept by all the nodes of a single gnodes. This ensures the distribution and the redundancy of the database. The gnode associated to $`H\hspace{0pt}{(h)}`$ is the *hash gnode*, and it is simply the gnode of level 1 of the hash node $`H\hspace{0pt}{(h)}`$. We will denote the former with $`g\hspace{0pt}H\hspace{0pt}{(h)}`$ and the latter with $`n\hspace{0pt}H\hspace{0pt}{(h)}`$. In symbols we can write: $`{n\hspace{0pt}H\hspace{0pt}{(h)}} \in {g\hspace{0pt}H\hspace{0pt}{(h)}}`$.
### 3.2 Approximated hash gnode
Since even the hash gnodes cannot exist, an approximation strategy is utilised: the nearest active gnode to the hash gnode becomes the *rounded hash gnode*, which will substitute in full the original (non existent) hash gnode. The rounded hash gnode is denoted with $`r\hspace{0pt}g\hspace{0pt}H\hspace{0pt}{(h)}`$.
Generally, when we are referring to the gnode, which has accepted a registration, there is no difference between the two kind of gnodes, they are always called hash gnodes.
### 3.3 Hook
When a node hooks to Netsukuku, it becomes automatically part of a hash gnode, thus it will also perform the *andna hook*. By hooking, it will get from its rnodes all the caches and databases, which are already distributed inside its hash gnode.
### 3.4 Cryptographic ANDNA
Before making a request to ANDNA, a node generates a couple of RSA keys, i.e. a public one (*pub key*) and a private (*priv key*). The size of the pub key will be limited due to reasons of space. The request of a hostname made to ANDNA will be signed with the private key and in the same request the public key will be attached. In this way, the node will be able to certify the true identity of its future requests.
### 3.5 Hostnames limit
The maximum number of hostnames, which can be registered is 256. This limit is necessary to prevent the massive registration of hostnames, which may lead to a overload of the ANDNA system. One its side effect is to slow down the action of spammers, which registers thousands of common words to resell them.
#### 3.5.1 Implementation of the limit
The *counter node* is a node with an ip equal to the hash of the public key of the node, which has registered the hostname $`h`$. We’ll denote the counter node associated to the node $`n`$ with $`c\hspace{0pt}r\hspace{0pt}{(n)}`$. Thus, for every node $`n`$, which has registered a hostname $`h`$, exists its corresponding counter node $`c\hspace{0pt}r\hspace{0pt}{(n)}`$.
The counter node $`c\hspace{0pt}r\hspace{0pt}{(n)}`$ will keep the number of hostnames registered by $`n`$.
When a hash gnode receives the registration request sent by $`n`$, it contacts $`c\hspace{0pt}r\hspace{0pt}{(n)}`$, which replies by telling how many hostnames have been already registered by the $`n`$. If the $`n`$ has not exceeded its limit, then $`c\hspace{0pt}r\hspace{0pt}{(n)}`$ will increment its counter and will allow the hash gnode to register the hostname.
$`c\hspace{0pt}r\hspace{0pt}{(n)}`$ is activated by the check request the hash gnode sends. $`n`$ has to keep the $`c\hspace{0pt}r\hspace{0pt}{(n)}`$ active following the same rules of the hibernation (see the chapter below). Practically, if $`c\hspace{0pt}r\hspace{0pt}{(n)}`$ doesn’t receives any more check requests, it will deactivate itself, and all the registered hostnames will become invalid.
The same distribution technique used for the hash gnode is used for the counter node: there is a whole gnode of counter nodes, which is called, indeed, counter gnode.
### 3.6 Registration step by step
1. 1\.
The node $`n`$ wants to register its hostname $`h`$,
2. 2\.
it finds the rounded hash gnode $`r\hspace{0pt}g\hspace{0pt}H\hspace{0pt}{(h)}`$ and contacts, randomly, one of its node. Let this node be $`y`$.
3. 3\.
$`n`$ sends the registration request to $`y`$. The request includes the public key of the node $`n`$. The pkt is also signed with the private key of $`n`$.
4. 4\.
The node $`y`$ verifies to be effectively the nearest gnode to the hash gnode, on the contrary it rejects the request. The signature validity is also checked.
5. 5\.
The node $`y`$ contacts the counter gnode $`c\hspace{0pt}r\hspace{0pt}{(n)}`$ and sends to it the IP, the hash of th hostname to be registered and a copy of the registration request itself.
6. 6\.
$`c\hspace{0pt}r\hspace{0pt}{(n)}`$ checks the data and gives its ok.
7. 7\.
The node $`y`$, after the affirmative reply, accepts the registration request and adds the entry in its database, storing the date of registration,
8. 8\.
$`y`$ forwards to its gnode the registration request.
9. 9\.
The nodes, which receive the forwarded request, will check its validity and store the entry in their db.
### 3.7 Hostname hibernation
The hash gnode keeps the hostname in an hibernated state for about 30 days since the moment of its registration or last update. The expiration time is very long to stabilise the domains. In this way, even someone attacks a node to steal its domain, it will have to wait 30 days to fully have it.
When the expiration time runs out, all the expired hostnames are deleted and substituted with the others in queue.
A node has to send an update request for each of its hostnames, each time it changes ip and before the hibernation time expires, in this way it’s hostname won’t be deleted.
The packet of the update request has an id, which is equal to the number of updates already sent. The pkt is also signed with the private key of the node to warrant the true identity of the request. The pkt is sent to any node of the hash gnode, which will send a copy of the request to the counter gnode, in order to verify if it is still active and that the entry related to the hostname being updated exists. On the contrary, the update request is rejected.
### 3.8 Hash gnodes mutation
If a generical rounded gnode is overpassed by a new gnode, which is nearer to the hash gnode, it will exchange its role with that of the second one, i.e. the new gnode will become the rounded hash gnode.
This transition takes place passively. When the register node will update its hostname, it will directly contact the new rounded gnode. Since the hostname stored inside the old rounded gnode won’t be updated, it will be dropped.
While the hostname has not been updated yet, all the nodes trying to resolve it, will find the new rounded gnode as the gnode nearest to the hash gnode and so they’ll send their resolution queries to the new gnode.
When the new rounded gnode receives the first query, it will notice that it is indeed a rounded gnode and to reply to the query, it it will retrieve from the old hash gnode the andna cache related to the hostname to resolve.
In this way, the registration of that hostname is automatically transfered from the old gnode into the new one.
If an attacker is able to send a registration request to the new hash gnode, before that the rightful owner of the hostname updates it or before that the passive transfer occurs, it will be able to steal it.
To prevent this, the new hash gnode will always double check a registration request by contacting the old hash gnode.
### 3.9 Yet another queue
Every node is free to choose any hostname, even if the hostname has been already chosen by another node.
The andna cache will keep a queue of MAX_ANDNA_QUEUE (5) entries. When the hostname on top of the queue expires, it will be automatically substituted by the next hostname.
### 3.10 Distributed cache for hostname resolution
In order to optimise the resolution of a hostname, a simple strategy is used: a node, each time it resolves a hostname, stores the result in a cache. For each next resolution of the same hostname, the node has already the result in its cache.
Since the resolution packet contains the latest time when the hostname has been registered or updated, an entry in the cache will expires exactly at the same time of the ANDNA hostname expiration.
The *resolved hnames* cache is readable by any node.
A node $`x`$, exploiting this feature, can ask to any node $`y`$, randomly chosen inside its same gnode, to resolve for itself the given hostname. The node $`y`$, will search in its resolved cache the hostname and on negative result the node will resolve it in the standard way, sending the result to the node $`x`$. These technique avoids the overload of the hash gnodes keeping very famous hostnames.
### 3.11 noituloser emantsoh esreveR
If a node wants to know all the hostnames associated to an ip, it will directly contact the node which owns that ip.
### 3.12 DNS wrapper
The DNS wrapper listen to the resolution queries sent by localhost and wraps them: it sends to the ANDNA daemon the hostnames to resolve and receives the their resolutions.
Thanks to the wrapper it is possible to use the ANDNA without modifying any preexistent programs.
See the ANDNS RFC \[[3](#bib.bib3)\] and the ANDNA manual \[[4](#bib.bib4)\].
## 4 Scattered Name Service Disgregation
The updated SNSD RFC can be found here: \[[5](#bib.bib5)\]
The Scattered Name Service Disgregation is the ANDNA equivalent of the SRV Record\[[6](#bib.bib6)\] of the Internet Domain Name System.
SNSD isn’t the same of the “SRV Record”, in fact, it has its own unique features.
With the SNSD it is possible to associate IPs and hostnames to another hostname.
Each assigned record has a service number, in this way the IPs and hostnames which have the same service number are grouped in an array. In the resolution request the client will specify the service number too, therefore it will get the record of the specified service number which is associated to the hostname. Example:
1. 1\.
The node X has registered the hostname ”angelica”. The default IP of “angelica” is “1.2.3.4”.
2. 2\.
X associates the “depausceve” hostname to the ‘http’ service number (80) of “angelica”.
3. 3\.
X associates the “11.22.33.44” IP to the ‘ftp’ service number (21) of “angelica”.
When the node Y resolves normally “angelica”, it gets 1.2.3.4, but when its web browser tries to resolve it, it asks for the record associated to the ‘http’ service, therefore the resolution will return ”depausceve”. The browser will resolve ”depausceve” and will finally contact the server. When the ftp client of Y will try to resolve ”angelica”, it will get the ”11.22.33.44” IP.
The node associated to a SNSD record is called ”SNSD node”. In this example ”depausceve” and 11.22.33.44 are SNSD nodes.
The node which registers the records and keeps the registration of the main hostname is always called ”register node”, but it can also be named ”Zero SNSD node”, in fact, it corresponds to the most general SNSD record: the service number 0.
Note that with the SNSD, the NTK_RFC 0004 will be completely deprecated.
### 4.1 Service, priority and weight number
#### 4.1.1 Service number
The service number specifies the scope of a SNSD record. The IP associated to the service number ‘x’ will be returned only to a resolution request which has the same service number.
A service number is the port number of a specific service. The port of the service can be retrieved from /etc/services.
The service number 0 corresponds to a normal ANDNA record. The relative IP will be returned to a general resolution request.
#### 4.1.2 Priority
The SNSD record has also a priority number. This number specifies the priority of the record inside its service array.
The client will contact first the SNSD nodes which have the higher priority, and only if they are unreachable, it will try to contact the other nodes which have a lower priority.
#### 4.1.3 Weight
The weight number, associated to each SNSD record, is used when there are more than one records which have the same priority number. In this case, this is how the client chooses which record using to contact the servers:
The client asks ANDNA the resolution request and it gets, for example, 8 different records.
The first record which will be used by the client is chosen in a pseudo-random manner: each record has a probability to be picked, which is proportional to its weight number, therefore the records with the heavier weight are more likely to be picked.
Note that if the records have the same priority, then the choice is completely random.
It is also possible to use a weight equal to zero to disable a record.
The weight number has to be less than 128.
### 4.2 SNSD Registration
The registration method of a SNSD record is similar to that described in the NTK_RFC 0004.
It is possible to associate up to 16 records to a single hostname. The maximum number of total records which can be registered is 256.
The registration of the SNSD records is performed by the same register node. The hash node which receives the registration won’t contact the counter node, because the hostname is already registered and it doesn’t need to verify anything about it. It has only to check the validity of the signature.
The register node can also choose to use an optional SNSD feature to be sure that a SNSD hostname is always associated to its trusted machine. In this case, the register node needs the ANDNA pubkey of the SNSD node to send a periodical challenge to the node. If the node fails to reply, the register node will send to ANDNA a delete request for the relative SNSD record.
The registration of SNSD records of hostnames which are only queued in the andna queue is discarded.
Practically, the steps necessary to register a SNSD record are:
1. 1\.
Modify the `/etc/netsukuku/snsd_nodes` file.
```
register_node# cd /etc/netsukuku/
register_node# cat snsd_nodes
#
# SNSD nodes file
#
# The format is:
# hostname:snsd_hostname:service:priority:weight[:pub_key_file]
# or
# hostname:snsd_ip:service:priority:weight[:pub_key_file]
#
# The ‘pub_key_file’ parameter is optional. If you specify it, NetsukukuD will
# check periodically ‘snsd_hostname’ and it will verify if it is always the
# same machine. If it isn’t, the relative snsd will be deleted.
#
depausceve:pippo:http:1
depausceve:1.2.3.4:21:0
angelica:frenzu:ssh:1:/etc/netsukuku/snsd/frenzu.pubk
register_node#
register_node# scp frenzu:/usr/share/andna_lcl_keyring snsd/frenzu.pubk
```
2. 2\.
Send a SIGHUP to the NetsukukuD of the register node:
```
register_node# killall -HUP ntkd
# or, alternatively
register_node# rc.ntk reload
```
#### 4.2.1 Zero SNSD IP
The main IP associated to a normal hostname has these default values:
```
IPΨ = register_node IPΨ# This value can’t be changed
service = 0
priority = 16
weight = 1
```
It is possible to associate other SNSD records in the service 0, but it isn’t allowed to change the main IP. The main IP can only be the IP of the register_node.
Although it isn’t possible to set a different association for the main IP, it can be disabled by setting its weight number to 0.
The string used to change the priority and weight value of the main IP is:
```
hostname:hostname:0:priority:weight
# For example:
register_node# echo depausceve:depausceve:0:23:12 >> /etc/netsukuku/snsd_nodes
```
#### 4.2.2 SNSD chain
Since it is possible to assign different aliases and backup IPs to the zero record, there is the possibility to create a SNSD chain. For example:
```
depausceve registers: depausceve:80 --> pippo
pippo registers: pippo:0 --> frenzu
frenzu registers: frenzu:0 --> angelica
```
However the SNSD chains are ignored, only the first resolution is considered valid. Since in the zero service there’s always the main IP, the resolution is always performed.
In this case (”depausceve:80 –¿ pippo:0”) the resolution will return the main IP of ”pippo:0”.
The reply to a resolution request of service zero, returns always IPs and not hostnames.
## References
- \[1\] Netsukuku website: http://netsukuku.freaknet.org/
- \[2\] Netsukuku topology document: http://netsukuku.freaknet.org/doc/main_doc/topology.pdf
- \[3\] ANDNS RFC: http://lab.dyne.org/Ntk_andna_and_dns
- \[4\] ANDNA manual page: http://netsukuku.freaknet.org/doc/manuals/html/andna.html
- \[5\] SNSD RFC: http://lab.dyne.org/Ntk_SNSD
- \[6\] SRV record: http://www.ietf.org/rfc/rfc2782.txt
http://en.wikipedia.org/wiki/SRV_record
`^_^` |
0704.3228 | cs.NI | cyber | # Technical Report v2 Characterization of P2P IPTV Traffic: Scaling Analysis
Thomas Silverston, Olivier Fourmaux and Kavé Salamatian
Université Pierre et Marie Curie - Paris VI
Laboratoire LIP6/CNRS, UMR 7606
104 avenue du président Kennedy 75016 Paris, France
Email: {thomas.silverston, olivier.fourmaux, kave.salamatian}@lip6.fr
###### Abstract
P2P IPTV applications arise on the Internet and will be massively used in the future. It is expected that P2P IPTV will contribute to increase the overall Internet traffic. In this context, it is important to measure the impact of P2P IPTV on the networks and to characterize this traffic. During the 2006 FIFA World Cup, we performed an extensive measurement campaign. We measured network traffic generated by broadcasting soccer games by the most popular P2P IPTV applications, namely PPLive, PPStream, SOPCast and TVAnts. From the collected data, we characterized the P2P IPTV traffic structure at different time scales by using wavelet based transform method. To the best of our knowledge, this is the first work, which presents a complete multiscale analysis of the P2P IPTV traffic.
Our results show that the scaling properties of the TCP traffic present periodic behavior whereas the UDP traffic is stationary and lead to long-range depedency characteristics. For all the applications, the download traffic has different characteristics than the upload traffic. The signaling traffic has a significant impact on the download traffic but it has negligible impact on the upload. Both sides of the traffic and its granularity has to be taken into account to design accurate P2P IPTV traffic models.
## 1 Introduction
P2P live streaming applications like P2P IPTV are emerging on the Internet and will be massively used in the future. The P2P traffic counts already for a large part of the Internet traffic and this is mainly due to P2P file-sharing applications as BitTorrent \[[1](#bib.bib1)\] or eDonkey \[[2](#bib.bib2)\]. Video streaming services like Youtube \[[3](#bib.bib3)\] appeared only a few months ago but contribute already to an important part of the Internet traffic. It is expected that P2P IPTV will largely contribute to increase the overall Internet traffic. It is therefore important to study P2P IPTV traffic and to characterize its properties.
The characterization of P2P IPTV traffic will allow us to understand its impact on the network. P2P IPTV applications have stringent QoS constraints (e.g. bandwidth, delay, jitter) and their traffic characterization will enable to understand their exact needs in network resources. The knowledge of the traffic properties enables the development of synthetic traffic generation models that are key input parameters when modeling or simulating these systems. Indeed, the modeling or simulating steps are necessary to design judiciously applications. From a traffic engineering point of view, well understanding P2P IPTV traffic is essential for Internet service providers to forecast their internal traffic and ensure a good provisioning of their network. And last but not least, global knowledge of the traffic properties will highlight some drawbacks of the applications and will make it possible to improve the design of new P2P IPTV architectures. For instance, an important concern of these systems is the scalability. The traffic characterization may help estimate the impact of overhead traffic generated by the signaling.
In this paper, we present a multiscale analysis of the structure of the traffic generated by the most popular P2P IPTV applications, namely PPLive \[[4](#bib.bib4)\], PPStream \[[5](#bib.bib5)\], SOPCast \[[6](#bib.bib6)\] and TVAnts \[[7](#bib.bib7)\].
During the 2006 FIFA World Cup, we performed an extensive measurement campaign. We measured the network traffic generated by broadcasting soccer games by the previously mentioned applications. The multiscale behavior of the collected traffic is analyzed using a wavelet transform based tool. In this paper, we characterize the network traffic of P2P IPTV systems at different time scales and compare their properties. To the best of our knowledge, this is the first work that does a comparative multiscale characterization of P2P IPTV traffic.
Our multiscale P2P IPTV traffic analysis shows significant differences in the scaling behaviors of TCP and UDP traffic. the TCP traffic presents periodic behavior while the UDP traffic is stationary and presents long-range dependency characteristics, which will affect the quality of the video reception. The signaling traffic has an impact on the download traffic but it has negligible impact on the upload traffic. The upload traffic generated by P2P IPTV systems have different scaling characteristics compare to the download traffic and both sides of the traffic has to be taken into account to design judiciously P2P IPTV traffic models. Moreover, the traffic granularity has to be considered while using traffic models to simulate these systems.
The rest of the paper is organized as follows. Firstly, we present the related work in section [2](#S2 "2 Related Work ‣ Technical Report v2 Characterization of P2P IPTV Traffic: Scaling Analysis"). In section [3](#S3 "3 Experiments ‣ Technical Report v2 Characterization of P2P IPTV Traffic: Scaling Analysis"), we give an overview of the measured applications and describe our measurement experiment setup. In section [4](#S4 "4 Analysis Methodology ‣ Technical Report v2 Characterization of P2P IPTV Traffic: Scaling Analysis"), we present our methodology to analyze the traffic at different time scales. We present our P2P IPTV traffic analysis in section [5](#S5 "5 Results Analysis ‣ Technical Report v2 Characterization of P2P IPTV Traffic: Scaling Analysis") and discuss the results in section [6](#S6 "6 Results Discussion ‣ Technical Report v2 Characterization of P2P IPTV Traffic: Scaling Analysis"). Finally, we conclude the paper and give perspectives in section [7](#S7 "7 Conclusion ‣ Technical Report v2 Characterization of P2P IPTV Traffic: Scaling Analysis").
## 2 Related Work
Nowadays, an increasing number of P2P IPTV measurement studies is conducted to analyze the mechanisms of such systems.
Zhang et. al \[[8](#bib.bib8)\] present the first measurement results about their protocol Donet \[[9](#bib.bib9)\], which were deployed on the Internet and called Coolstreaming. They provide network statistics, like user’s behavior in the whole system and the quality of video reception.
Hei et al. \[[10](#bib.bib10)\] \[[11](#bib.bib11)\] made a complete measurement of the popular PPLive application. They made active measurements by instrumentalizing their own crawler and give many architecture and overlay details like buffer size or number of peers in the networks. Vu et al. \[[12](#bib.bib12)\] made also active measurements of the PPLive system and derive mathematical models for the distributions of channel population size or session length.
In our previous work \[[13](#bib.bib13)\], we passively measured the network traffic generated by several popular applications during a worldwide event. We compared the measured applications by inferring their underlying mechanisms and highlight their design differences and similarities. Ali et al. \[[14](#bib.bib14)\] made passive measurements of PPLive and SOPCast applications and analyze the performance and characteristics of such systems.
Still in their previously mentioned works, Ali et al. provide their own methodology to study the data exchanges of such P2P applications. Based on their measurement studies, Hei et al. \[[15](#bib.bib15)\] developed also a methodology to estimate the overall perceived video quality throughout the network.
All these works studied P2P IPTV systems by measuring the traffic and tried to infer their mechanisms, but they did not characterize the correlation structure of the generated traffic at different time scales to understand its properties and its impact on the network.
## 3 Experiments
### 3.1 P2P IPTV applications overview
For our P2P IPTV traffic measurement experiments, we chose four applications, namely PPLive, PPStream, SOPCast and TVAnts because they were very popular on the Internet. Whenever these applications are freely available, their source codes are not open and their exact implementation details and used protocols are still widely unknown. We can only rely on reverse engineering to understand their transmission mechanisms.
All these applications claim to use swarming protocol like Donet. Similarly to BitTorrent, video data flows are divided into data chunks and each peer downloads the chunck of data to other peers concurrently. The peers know how to download the video data chunks by exchanging randomly with other peers information about the data chunks they have or neighbor peers they know. With this signaling traffic, each peer discovers iteratively new peers, new available data chunks and is able to download video from several peers. In these P2P protocols, there are two kinds of traffic: video traffic where peers exchange data chunks with each other and signaling traffic where peers exchange information to get the data.
As we show in \[[13](#bib.bib13)\], all the applications transport video and signaling traffics differently: PPStream uses exclusively TCP for all traffics while PPLive adds UDP for some signaling traffic. SOPCast uses almost entirely UDP and TVAnts is more balanced between TCP ($`\approx`$ 75%) and UDP for all kinds of traffic.
In the next section, we will present the measurement experiments platform we used to collect the P2P IPTV traffic.
### 3.2 Measurement experiments platform
[Refer to caption](/html/0704.3228/assets/x1.png)
Figure 1: Measurement experiments platform. Each node is a common PC directly connected to the Internet via campus network.
Table 1: Packet traces summary
<table>
<tbody>
<tr>
<td></td>
<td>PPLive</td>
<td>PPStream</td>
<td>SOPCast</td>
<td>TVAnts</td>
</tr>
<tr>
<td>Duration (s)</td>
<td>13,321</td>
<td>12,375</td>
<td>12,198</td>
<td>13,358</td>
</tr>
<tr>
<td>Size (MB)</td>
<td>6,339</td>
<td>4,121</td>
<td>5,475</td>
<td>3,992</td>
</tr>
<tr>
<td>Download(%)</td>
<td>14.11</td>
<td>20.50</td>
<td>16.13</td>
<td>24.76</td>
</tr>
<tr>
<td>TCP</td>
<td>14.09</td>
<td>20.50</td>
<td>0.23</td>
<td>14.71</td>
</tr>
<tr>
<td>UDP</td>
<td>0.02</td>
<td>0</td>
<td>15.90</td>
<td>10.05</td>
</tr>
<tr>
<td>Upload(%)</td>
<td>85.89</td>
<td>79.50</td>
<td>83.87</td>
<td>75.24</td>
</tr>
<tr>
<td>TCP</td>
<td>85.81</td>
<td>79.50</td>
<td>3.89</td>
<td>61.67</td>
</tr>
<tr>
<td>UDP</td>
<td>0.08</td>
<td>0</td>
<td>79.98</td>
<td>13.57</td>
</tr>
</tbody>
</table>
Our measurement experiments take place during the 2006 FIFA World Cup from 09 June to 09 July. We collected a huge amount of data, measuring most of the World Cup soccer games with different applications at the same time and under different network environments: campus Ethernet access and residential ADSL access.
In this paper, we focus on four packet traces collected on June 30 in our campus network: one for each measured application. In all our data, we selected these packet traces because they are well representative of all of them. The traces are made available on our traces sharing service \[[16](#bib.bib16)\]. Two soccer games were scheduled, one in the afternoon (Germany vs. Argentine) measured by PPStream and SOPCast and one in the evening (Italy vs. Ukraine) measured by PPLive and TVAnts. We selected the four traces from different applications to be able to characterize the P2P IPTV traffic without being closely related to the design of the applications.
Our measurement experiment set up is described on Fig. [1](#S3.F1 "Figure 1 ‣ 3.2 Measurement experiments platform ‣ 3 Experiments ‣ Technical Report v2 Characterization of P2P IPTV Traffic: Scaling Analysis"). To collect the packets, we used two personal computers with 1.8GHz CPU and common graphic card capabilities. The operating system running on the PCs was Windows XP. The PCs (nodes) were situated in our campus network and were directly connected to the Internet with 100Mbps Ethernet access. During a game, each node was running a P2P IPTV application and we used tcpdump on each measuring node to collect all the packets. For all the measurement experiments, the consumed bandwidth was always relatively low and does not exceed 10Mbps. The Ethernet cards did not suffer any packet loss and captured all the packets. For all the experiments, nodes were watching CCTV5, a Chinese TV channel available for all the measured applications. It was important to watch the same TV channel with all the applications to assure that the behavior of users will be similar in each trace. For example, during the advertisement, whatever the applications, an user may stop watching the channel and switches the application off and then switch it on a few minutes later. All the applications used MPEG4 video encoding.
Our platform has high-speed access and our observations can not be directly generalized to residential peers with common access to the Internet (e.g. 20/1 Mbps or 512/128 Kbps). However, residential network capacities are quickly increasing and will have such high-speed access in only a few years when P2P IPTV would be commonly used.
Table [1](#S3.T1 "Table 1 ‣ 3.2 Measurement experiments platform ‣ 3 Experiments ‣ Technical Report v2 Characterization of P2P IPTV Traffic: Scaling Analysis") summarizes the four presented traces. At first, we can notice in these traces that PPLive, TVAnts and PPStream use massively TCP whereas only SOPCast uses mainly UDP. The duration of the traces is longer than the duration of a soccer game ($`\approx`$ 105 minutes). We chose to collect the traffic a few minutes before and after the games to capture all the effects that the live interest of a soccer game could produce on the behavior of users (e.g. flash crowds).
## 4 Analysis Methodology
[Refer to caption](/html/0704.3228/assets/x2.png)
\(a\)
PPLive
[Refer to caption](/html/0704.3228/assets/x3.png)
\(b\)
SOPCast
[Refer to caption](/html/0704.3228/assets/x4.png)
\(c\)
PPStream
[Refer to caption](/html/0704.3228/assets/x5.png)
\(d\)
TVAnts
Figure 2: Packet distributions for all the applications
### 4.1 Video and signaling traffics
The packet size distributions for all the applications are presented in Fig. [2](#S4.F2 "Figure 2 ‣ 4 Analysis Methodology ‣ Technical Report v2 Characterization of P2P IPTV Traffic: Scaling Analysis"). PPLive has 75% of its packets larger than 1300 Bytes and the other packets are very small ($`\approx`$ 17% are 100 Bytes length). Differently, SOPCast counts only about 30% of large packets ($`>`$ 1300 Bytes) and 60% of its packets are very small ($`<`$100 Bytes). PPStream counts more than 40% of packets bigger than 1400 Bytes, almost 40% of its packets are 100 Bytes and a small part is 1000 Bytes. TVAnts packets seems more balanced in three equal parts: one part is 100 Bytes, a second part is 1100 Bytes and the third part is 1400 Bytes.
All the packets distributions of these applications are different but we can distinguish two sets of packets within these packet distributions: small-size packets ($`<`$ 200 Bytes) and large-size packets ($`>`$ 1000 Bytes). As we explained in section [3](#S3 "3 Experiments ‣ Technical Report v2 Characterization of P2P IPTV Traffic: Scaling Analysis"), the studied P2P applications generate two kinds of traffic: video and signaling.
It is expected that the video traffic is essentially composed of large-size packets. Most of the video packets should belong to the large-size packets set ($`>`$ 1000 Bytes). The video data are also delay sensitive since video packets have to respect stringent playback delay for the users get a smooth quality video.
The signaling packets should belong to the small-size packets set ($`<`$ 200 Bytes). The signaling traffic of P2P IPTV systems is not expected to be delay sensitive because it is used to exchange information about peers or data availability but not for interactive commands as for Video on-Demand systems like Joost \[[17](#bib.bib17)\]. In video on-demand systems, the users may want to read the video forward or backward instantaneously. In the case of P2P IPTV, it is not possible to have this kind of interactive commands since the data flows are broadcasted in live.
The signaling and video traffics have not the same characteristics as packet size or delay constraints and they would have a different impact on the network. We have therefore to separate video and signaling traffic and to analyze them separately.
### 4.2 Signaling traffic filtering heuristic
Table 2: Signaling Traffic Ratio
<table>
<tbody>
<tr>
<td></td>
<td>PPLive</td>
<td>PPStream</td>
<td>SOPCast</td>
<td>TVAnts</td>
</tr>
<tr>
<td>Total</td>
<td>4.1%</td>
<td>13.6%</td>
<td>19.3%</td>
<td>10.2%</td>
</tr>
<tr>
<td>Upload</td>
<td>2.2%</td>
<td>10.8%</td>
<td>13.6%</td>
<td>7.8%</td>
</tr>
<tr>
<td>Download</td>
<td>19.2%</td>
<td>25.8%</td>
<td>48.5%</td>
<td>18.0%</td>
</tr>
</tbody>
</table>
By observing that the video packets size should be larger than 1000 Bytes and signaling packets should be much smaller ($`<`$ 200 Bytes), we used the simple heuristic proposed by Hei. \[[10](#bib.bib10)\] to separate the video traffic from the overall traffic. The heuristic works as follows: for a session (same IP addresses and ports), we counted the number of packets bigger or equal than 1000 Bytes. If a session had at least 10 large packets, then it was labeled as a video session and we removed small packets ($`<`$ 1000 Bytes) for this session. At the end, we removed all the non video-sessions from the trace to obtain video traffic.
At the end, we removed all the non video-sessions from the trace and all the labeled video sessions compose the video traffic.
Table [2](#S4.T2 "Table 2 ‣ 4.2 Signaling traffic filtering heuristic ‣ 4 Analysis Methodology ‣ Technical Report v2 Characterization of P2P IPTV Traffic: Scaling Analysis") summarizes the signaling ratio for all the applications in the entire traces and in both upload or download directions. As an example, for PPLive, the signaling traffic represents 4.1% of the total traffic. If we consider both traffic directions, the signaling traffic represents 2.2% of the upload traffic whereas it represents 19.2% of the download traffic. The signaling traffic ratios will be discussed more deeply in the results analysis section ($`§`$ [5](#S5 "5 Results Analysis ‣ Technical Report v2 Characterization of P2P IPTV Traffic: Scaling Analysis")).
The next section is dedicated to validate the heuristic used to filter the signaling traffic.
### 4.3 Validation of the signaling heuristic
The measured applications are proprietary and we do not have any implementation detail. We do not know exactly if a packet is for signaling traffic or video traffic. The heuristic used to filter the signaling traffic relies mainly on the packet size. The heuristic will perform well if it removes only signaling packets without removing any part of the video traffic.
From Table [2](#S4.T2 "Table 2 ‣ 4.2 Signaling traffic filtering heuristic ‣ 4 Analysis Methodology ‣ Technical Report v2 Characterization of P2P IPTV Traffic: Scaling Analysis"), we notice that the download signaling traffic ratio computed by the heuristic for SOPCast is 48.5%, which is a high ratio indicating that half of the traffic would have been signaling traffic. This high ratio of download signaling traffic may eventually come from the inaccuracy of the heuristic.
To validate the filtering heuristic, we compute the resulting video bitrate after removing the signaling traffic. Realistic computed video bitrate will confirm the efficiency of the heuristic. We compute the video bitrate for download traffic because it is the only video traffic we can deduce. The download traffic is provided by other peers on the Internet and all the video flows received by our controlled nodes compose the downloaded video. The video upload traffic is not provided to a single consumer peer. We can not estimate the video bitrate received by remote peers because they mix video flows from many other providers peers to receive the entire video. In the following, we compute for all the traces the video download bitrate received by our controlled nodes by removing signaling traffic with the presented heuristic.
Fig. [2(b)](#S4.F2.sf2 "In Figure 2 ‣ 4 Analysis Methodology ‣ Technical Report v2 Characterization of P2P IPTV Traffic: Scaling Analysis") indicates that SOPCast has more than 60% of its packets smaller than 100 Bytes. These packet are certainly signaling packets according to their size. The design of SOPCast introduces a large amount of signaling packets and the heuristic removes these packets from the video traffic.
Table [1](#S3.T1 "Table 1 ‣ 3.2 Measurement experiments platform ‣ 3 Experiments ‣ Technical Report v2 Characterization of P2P IPTV Traffic: Scaling Analysis") shows that the SOPCast download traffic represents 16.13% of the collected traffic (5,475 MBytes) and the download signaling ratio represents 48.5% of the overall download traffic. The average bandwidth used by SOPCast to download the video at its bitrate can be computed by dividing the amount of received video data by the measurement experiment duration (12,198s).
<table>
<tbody>
<tr>
<td></td>
<td>(5475∗220∗8∗0.1613)∗(1−0.485)12198≈305Kbps5475superscript22080.161310.48512198305𝐾𝑏𝑝𝑠\frac{(5475*2^{20}*8*0.1613)*(1-0.485)}{12198}\approx 305Kbps</td>
<td></td>
</tr>
</tbody>
</table>
For SOPCast, the video download speed (305Kbps) computed with the heuristic is realistic for downloading a video broadcasted in the Internet.
Table [3](#S4.T3 "Table 3 ‣ 4.3 Validation of the signaling heuristic ‣ 4 Analysis Methodology ‣ Technical Report v2 Characterization of P2P IPTV Traffic: Scaling Analysis") shows the computed download speeds for all the applications.
Table 3: Video Download Bitrate computed with the signaling filtering heuristic
<table>
<thead>
<tr>
<th></th>
<th>PPLive</th>
<th>PPStream</th>
<th>SOPCast</th>
<th>TVAnts</th>
</tr>
</thead>
<tbody>
<tr>
<td>Video Bitrate</td>
<td rowspan="2">445</td>
<td rowspan="2">415</td>
<td rowspan="2">305/360</td>
<td rowspan="2">500</td>
</tr>
<tr>
<td>(Kbps)</td>
</tr>
</tbody>
</table>
All the computed download speeds are realistic and it confirms that the heuristic used to filter signaling traffic performs well and is efficient to distinguish signaling traffic from the overall traffic. Our node running SOPCast receives really a lot of signaling packets from other peers in the Internet and this high ratio is not distorted by the heuristic.
To have a finer understanding of the high signaling traffic ratio of SOPCast, we show in our previous work \[[13](#bib.bib13)\] that SOPCast received no video traffic during 30 minutes. The video source suffered troubles during this period of time. In this period, our node was only receiving signaling traffic from other peers requesting video data chunks, but no video chunks were received. Since our node has high bandwidth capabilities, lots of peers were requesting video generating a large amount of signaling traffic. This phenomenon increased artificially the SOPCast signaling traffic and explains why the signaling traffic ratio SOPCast experiments in the download direction was so high. If we take into account this lack of video data during 30mn in our calculation, we obtain 360Kbps video download bitrate, which is a more realistic bitrate closer to all the other computed video bitrate.
In this section, we validate the choice of the heuristic used to filter signaling traffic from the entire traces. In the following, for each trace, we will consider the overall traffic generated by the applications and the video traffic without any signaling packet deduced by using the signaling filtering heuristic.
To characterize the correlation structure of the generated traffic at different time scales, we analyzed the traffic using a wavelet based transform method. To this end, we used a tool that is presented in the next section.
### 4.4 Multiscale traffic analysis
We analyze the measured P2P IPTV traffic at different time scales to characterize this traffic and its properties. To this end, we compute the energy spectrum of the traffic at different time scales using a wavelet based transform method. The smaller time scales analyzed is the 20 milliseconds intervals as we observed that inter arrivals are rarely below this value (our bin duration is 20ms). In each interval, we counted the number of packet arrivals in both directions (i.e. upload and download). We only counted arrivals of packets with data payload and do not take into account empty TCP Acknowledgment packets.
Logscale Diagram Estimate \[[18](#bib.bib18)\] (LDestimate) is based on discrete wavelet transform and allows to analyze the scaling behavior of the packet traffic. LDestimate produces a logarithmic plot of the data energy spectrum.
For all the produced logscale diagrams, the X-axes are the octaves of the traffic, which are the time scales of the packet arrivals. The right-most part of the graph is relative to large time scales and the left part is relative to small time scales. The Y-axes are the data energy spectra. A logscale diagram can be understood as follows: an octave $`j`$ (X-axes) is a time scale of the packet traffic energy spectrum. Since our bin is 20 ms, the octave $`j = 8`$ means the time scale $`t = {{2^{8} \ast 20}\hspace{0pt}m\hspace{0pt}s} = {5.12\hspace{0pt}s}`$.
LDEstimate is a tool that allow to visually observed the properties of the measured traffic. In a produced diagram, a bump in the energy spectrum indicates a possible periodic behavior of the traffic, a constant energy spectrum a possible memoryless process and a linear increase indicates a possible long-range dependency of the traffic (LRD).
## 5 Results Analysis
[Refer to caption](/html/0704.3228/assets/x6.png)
\(a\)
Upload: overall packet traffic (signaling and video traffic)
[Refer to caption](/html/0704.3228/assets/x7.png)
\(b\)
Upload: video packet traffic
[Refer to caption](/html/0704.3228/assets/x8.png)
\(c\)
Download: overall packet traffic (signaling and video traffic)
[Refer to caption](/html/0704.3228/assets/x9.png)
\(d\)
Download: video packet traffic
Figure 3: PPlive packet traffics energy spectra. Bin duration is 20ms. (Ex: Octave $`j = 8`$ is for scale process at $`t = {{2^{8} \ast 20}\hspace{0pt}m\hspace{0pt}s} = {5.12\hspace{0pt}s}`$).
[Refer to caption](/html/0704.3228/assets/x10.png)
\(a\)
Upload: overall packet traffic (signaling and video traffic)
[Refer to caption](/html/0704.3228/assets/x11.png)
\(b\)
Upload: video packet traffic
[Refer to caption](/html/0704.3228/assets/x12.png)
\(c\)
Download: overall packet traffic (signaling and video traffic)
[Refer to caption](/html/0704.3228/assets/x13.png)
\(d\)
Download: video packet traffic
Figure 4: SOPCast packet traffics energy spectra. Bin duration is 20ms. (Ex: Octave $`j = 8`$ is for scale process at $`t = {{2^{8} \ast 20}\hspace{0pt}m\hspace{0pt}s} = {5.12\hspace{0pt}s}`$).
[Refer to caption](/html/0704.3228/assets/x14.png)
\(a\)
Upload: overall packet traffic (signaling and video traffic)
[Refer to caption](/html/0704.3228/assets/x15.png)
\(b\)
Upload: video packet traffic
[Refer to caption](/html/0704.3228/assets/x16.png)
\(c\)
Download: overall packet traffic (signaling and video traffic)
[Refer to caption](/html/0704.3228/assets/x17.png)
\(d\)
Download: video packet traffic
Figure 5: PPStream packet traffics energy spectra. Bin duration is 20ms. (Ex: Octave $`j = 8`$ is for scale process at $`t = {{2^{8} \ast 20}\hspace{0pt}m\hspace{0pt}s} = {5.12\hspace{0pt}s}`$).
[Refer to caption](/html/0704.3228/assets/x18.png)
\(a\)
Upload: overall packet traffic (signaling and video traffic)
[Refer to caption](/html/0704.3228/assets/x19.png)
\(b\)
Upload: video packet traffic
[Refer to caption](/html/0704.3228/assets/x20.png)
\(c\)
Download: overall packet traffic (signaling and video traffic)
[Refer to caption](/html/0704.3228/assets/x21.png)
\(d\)
Download: video packet traffic
Figure 6: TVAnts packet traffics energy spectra. Bin duration is 20ms. (Ex: Octave $`j = 8`$ is for scale process at $`t = {{2^{8} \ast 20}\hspace{0pt}m\hspace{0pt}s} = {5.12\hspace{0pt}s}`$).
### 5.1 Presentation
In the rest of the paper, we will refer to the traffic trace of an application by the name of the application. For example, we will refer to SOPCast packet trace by SOPCast.
For each application, we study the traffic by separating the upload traffic and the download traffic. In both traffic directions, we separated the video traffic from the overall traffic by using the filtering heuristic presented in section [4.2](#S4.SS2 "4.2 Signaling traffic filtering heuristic ‣ 4 Analysis Methodology ‣ Technical Report v2 Characterization of P2P IPTV Traffic: Scaling Analysis"). Then, each application is characterized by four distinct logscale diagrams: overall upload traffic, video upload traffic, overall download traffic and video download traffic. Fig. [3](#S5.F3 "Figure 3 ‣ 5 Results Analysis ‣ Technical Report v2 Characterization of P2P IPTV Traffic: Scaling Analysis") presents the logscale diagrams of the energy spectra for PPLive, Fig. [4](#S5.F4 "Figure 4 ‣ 5 Results Analysis ‣ Technical Report v2 Characterization of P2P IPTV Traffic: Scaling Analysis") the energy spectra for SOPCast, Fig. [5](#S5.F5 "Figure 5 ‣ 5 Results Analysis ‣ Technical Report v2 Characterization of P2P IPTV Traffic: Scaling Analysis") for PPStream and Fig. [6](#S5.F6 "Figure 6 ‣ 5 Results Analysis ‣ Technical Report v2 Characterization of P2P IPTV Traffic: Scaling Analysis") for TVAnts.
Table [1](#S3.T1 "Table 1 ‣ 3.2 Measurement experiments platform ‣ 3 Experiments ‣ Technical Report v2 Characterization of P2P IPTV Traffic: Scaling Analysis") indicated previously that three of the measured applications use massively TCP (PPLive, PPStream and TVAnts) whereas only SOPCast uses mainly UDP. We will refer to an application using massively TCP as TCP application and UDP application for an application using UDP.
In the following, we will present traffic differences between TCP and UDP applications in section [5.2](#S5.SS2 "5.2 TCP applications vs. UDP applications ‣ 5 Results Analysis ‣ Technical Report v2 Characterization of P2P IPTV Traffic: Scaling Analysis"). In section [5.3](#S5.SS3 "5.3 Impact of the signaling traffic ‣ 5 Results Analysis ‣ Technical Report v2 Characterization of P2P IPTV Traffic: Scaling Analysis") we will highlight the impact of the signaling traffic for P2P IPTV applications. Then we discuss the stationnarity of the traffic in section [5.4](#S5.SS4 "5.4 Traffic stationnarity ‣ 5 Results Analysis ‣ Technical Report v2 Characterization of P2P IPTV Traffic: Scaling Analysis"). We will extend our findings in section [5.5](#S5.SS5 "5.5 Analysis of the top download flows ‣ 5 Results Analysis ‣ Technical Report v2 Characterization of P2P IPTV Traffic: Scaling Analysis"). The results are summarized and discussed in section [6](#S6 "6 Results Discussion ‣ Technical Report v2 Characterization of P2P IPTV Traffic: Scaling Analysis").
### 5.2 TCP applications vs. UDP applications
For the TCP applications (PPLive, PPStream and TVAnts), the two upload energy spectra look similar for all the time scales (Fig. [3(a)](#S5.F3.sf1 "In Figure 3 ‣ 5 Results Analysis ‣ Technical Report v2 Characterization of P2P IPTV Traffic: Scaling Analysis") and [3(b)](#S5.F3.sf2 "In Figure 3 ‣ 5 Results Analysis ‣ Technical Report v2 Characterization of P2P IPTV Traffic: Scaling Analysis"), Fig. [5(a)](#S5.F5.sf1 "In Figure 5 ‣ 5 Results Analysis ‣ Technical Report v2 Characterization of P2P IPTV Traffic: Scaling Analysis") and [5(b)](#S5.F5.sf2 "In Figure 5 ‣ 5 Results Analysis ‣ Technical Report v2 Characterization of P2P IPTV Traffic: Scaling Analysis"), Fig. [6(a)](#S5.F6.sf1 "In Figure 6 ‣ 5 Results Analysis ‣ Technical Report v2 Characterization of P2P IPTV Traffic: Scaling Analysis") and [6(b)](#S5.F6.sf2 "In Figure 6 ‣ 5 Results Analysis ‣ Technical Report v2 Characterization of P2P IPTV Traffic: Scaling Analysis")) The two download energy spectra look similar until $`j = 9`$ and after they are different (Fig. [3(c)](#S5.F3.sf3 "In Figure 3 ‣ 5 Results Analysis ‣ Technical Report v2 Characterization of P2P IPTV Traffic: Scaling Analysis") and [3(d)](#S5.F3.sf4 "In Figure 3 ‣ 5 Results Analysis ‣ Technical Report v2 Characterization of P2P IPTV Traffic: Scaling Analysis"), Fig. [5(c)](#S5.F5.sf3 "In Figure 5 ‣ 5 Results Analysis ‣ Technical Report v2 Characterization of P2P IPTV Traffic: Scaling Analysis") and [5(d)](#S5.F5.sf4 "In Figure 5 ‣ 5 Results Analysis ‣ Technical Report v2 Characterization of P2P IPTV Traffic: Scaling Analysis"), Fig. [6(c)](#S5.F6.sf3 "In Figure 6 ‣ 5 Results Analysis ‣ Technical Report v2 Characterization of P2P IPTV Traffic: Scaling Analysis") and [6(d)](#S5.F6.sf4 "In Figure 6 ‣ 5 Results Analysis ‣ Technical Report v2 Characterization of P2P IPTV Traffic: Scaling Analysis")). The upload energy spectra of TCP applications ([3(a)](#S5.F3.sf1 "In Figure 3 ‣ 5 Results Analysis ‣ Technical Report v2 Characterization of P2P IPTV Traffic: Scaling Analysis") and [3(b)](#S5.F3.sf2 "In Figure 3 ‣ 5 Results Analysis ‣ Technical Report v2 Characterization of P2P IPTV Traffic: Scaling Analysis"), [5(a)](#S5.F5.sf1 "In Figure 5 ‣ 5 Results Analysis ‣ Technical Report v2 Characterization of P2P IPTV Traffic: Scaling Analysis") and [5(b)](#S5.F5.sf2 "In Figure 5 ‣ 5 Results Analysis ‣ Technical Report v2 Characterization of P2P IPTV Traffic: Scaling Analysis"), [6(a)](#S5.F6.sf1 "In Figure 6 ‣ 5 Results Analysis ‣ Technical Report v2 Characterization of P2P IPTV Traffic: Scaling Analysis") and [6(b)](#S5.F6.sf2 "In Figure 6 ‣ 5 Results Analysis ‣ Technical Report v2 Characterization of P2P IPTV Traffic: Scaling Analysis")) do not look like their download energy spectra ([3(c)](#S5.F3.sf3 "In Figure 3 ‣ 5 Results Analysis ‣ Technical Report v2 Characterization of P2P IPTV Traffic: Scaling Analysis") and [3(d)](#S5.F3.sf4 "In Figure 3 ‣ 5 Results Analysis ‣ Technical Report v2 Characterization of P2P IPTV Traffic: Scaling Analysis"), [5(c)](#S5.F5.sf3 "In Figure 5 ‣ 5 Results Analysis ‣ Technical Report v2 Characterization of P2P IPTV Traffic: Scaling Analysis") and [5(d)](#S5.F5.sf4 "In Figure 5 ‣ 5 Results Analysis ‣ Technical Report v2 Characterization of P2P IPTV Traffic: Scaling Analysis"), ([6(c)](#S5.F6.sf3 "In Figure 6 ‣ 5 Results Analysis ‣ Technical Report v2 Characterization of P2P IPTV Traffic: Scaling Analysis") and [6(d)](#S5.F6.sf4 "In Figure 6 ‣ 5 Results Analysis ‣ Technical Report v2 Characterization of P2P IPTV Traffic: Scaling Analysis")).
All the TCP applications have similar energy spectra whatever the kind of traffic and direction (e.g. overall or video). The observations for TCP applications may be generalized for all of those we measured (PPLive, PPStream and TVAnts).
For UDP applications (SOPCast Fig. [4](#S5.F4 "Figure 4 ‣ 5 Results Analysis ‣ Technical Report v2 Characterization of P2P IPTV Traffic: Scaling Analysis")), the four energy spectra look similar whatever the traffic direction or the traffic nature (e.g. overall or video traffics).
The energy spectra of TCP applications (Fig. [3](#S5.F3 "Figure 3 ‣ 5 Results Analysis ‣ Technical Report v2 Characterization of P2P IPTV Traffic: Scaling Analysis"), [5](#S5.F5 "Figure 5 ‣ 5 Results Analysis ‣ Technical Report v2 Characterization of P2P IPTV Traffic: Scaling Analysis") and [6](#S5.F6 "Figure 6 ‣ 5 Results Analysis ‣ Technical Report v2 Characterization of P2P IPTV Traffic: Scaling Analysis")) are different from the energy spectra of UDP applications (Fig. [4](#S5.F4 "Figure 4 ‣ 5 Results Analysis ‣ Technical Report v2 Characterization of P2P IPTV Traffic: Scaling Analysis")).
Regarding TCP applications energy spectra more precisely, we can observe an energy bump in all the logscale diagrams about time scale $`j = 8`$ (5.12s). The energy bump is more clearly defined in upload traffic (Fig. [3(a)](#S5.F3.sf1 "In Figure 3 ‣ 5 Results Analysis ‣ Technical Report v2 Characterization of P2P IPTV Traffic: Scaling Analysis") and [3(b)](#S5.F3.sf2 "In Figure 3 ‣ 5 Results Analysis ‣ Technical Report v2 Characterization of P2P IPTV Traffic: Scaling Analysis"), Fig. [5(a)](#S5.F5.sf1 "In Figure 5 ‣ 5 Results Analysis ‣ Technical Report v2 Characterization of P2P IPTV Traffic: Scaling Analysis") and [5(b)](#S5.F5.sf2 "In Figure 5 ‣ 5 Results Analysis ‣ Technical Report v2 Characterization of P2P IPTV Traffic: Scaling Analysis"), Fig. [6(a)](#S5.F6.sf1 "In Figure 6 ‣ 5 Results Analysis ‣ Technical Report v2 Characterization of P2P IPTV Traffic: Scaling Analysis") and [6(b)](#S5.F6.sf2 "In Figure 6 ‣ 5 Results Analysis ‣ Technical Report v2 Characterization of P2P IPTV Traffic: Scaling Analysis")) than download traffic (Fig. [3(c)](#S5.F3.sf3 "In Figure 3 ‣ 5 Results Analysis ‣ Technical Report v2 Characterization of P2P IPTV Traffic: Scaling Analysis") and [3(d)](#S5.F3.sf4 "In Figure 3 ‣ 5 Results Analysis ‣ Technical Report v2 Characterization of P2P IPTV Traffic: Scaling Analysis"), Fig. [5(c)](#S5.F5.sf3 "In Figure 5 ‣ 5 Results Analysis ‣ Technical Report v2 Characterization of P2P IPTV Traffic: Scaling Analysis") and [5(d)](#S5.F5.sf4 "In Figure 5 ‣ 5 Results Analysis ‣ Technical Report v2 Characterization of P2P IPTV Traffic: Scaling Analysis"), Fig. [6(c)](#S5.F6.sf3 "In Figure 6 ‣ 5 Results Analysis ‣ Technical Report v2 Characterization of P2P IPTV Traffic: Scaling Analysis") and [6(d)](#S5.F6.sf4 "In Figure 6 ‣ 5 Results Analysis ‣ Technical Report v2 Characterization of P2P IPTV Traffic: Scaling Analysis")).
The energy bump exists for all the applications using mostly TCP. The energy bump indicates a possible periodic behavior at these time scales whatever the traffic direction or its nature. The energy bump phenomenon has to be confirmed by studying the stationnarity of the traffic. We study traffic stationnarity in section [5.4](#S5.SS4 "5.4 Traffic stationnarity ‣ 5 Results Analysis ‣ Technical Report v2 Characterization of P2P IPTV Traffic: Scaling Analysis"). However, the stationnarity analysis shows that the energy bumps observed in the spectra are essential phenomena and not simply artifact coming from non-stationnarity.
The energy bumps are observed for all the TCP applications but not for the UDP applications. The well known TCP mechanisms used to transport data and TCP retransmission mechanisms could lead to such periodic traffic behaviors. However, the periodic behaviors are observed for time scale $`j = 8`$ (5.12s). A 5 seconds duration is a very long duration for TCP mechanisms. It is not so obvious that this periodic behavior is provided by TCP mechanisms.
The periodic behaviors could also come from the video broadcasted in the network. However, SOPCast does not show any energy bump in its energy spectra and SOPCast broadcasts also video in the network.
Currently, we can not surely establish the source of these periodic behaviors. It does not seem to come from TCP mechanisms nor broadcasted video. We are still investigating the periodic behavior we observed in the TCP applications energy spectra.
The energy bumps are characteristics shared by all the measured applications using massively TCP. They illustrate how the application design may impact the properties of the generated network traffic.
### 5.3 Impact of the signaling traffic
For all the applications, whatever the transport protocol they use, their video upload energy spectra look like their overall upload energy spectra. This is illustrated on Fig. [3(a)](#S5.F3.sf1 "In Figure 3 ‣ 5 Results Analysis ‣ Technical Report v2 Characterization of P2P IPTV Traffic: Scaling Analysis") and [3(b)](#S5.F3.sf2 "In Figure 3 ‣ 5 Results Analysis ‣ Technical Report v2 Characterization of P2P IPTV Traffic: Scaling Analysis"), [4(a)](#S5.F4.sf1 "In Figure 4 ‣ 5 Results Analysis ‣ Technical Report v2 Characterization of P2P IPTV Traffic: Scaling Analysis") and [4(b)](#S5.F4.sf2 "In Figure 4 ‣ 5 Results Analysis ‣ Technical Report v2 Characterization of P2P IPTV Traffic: Scaling Analysis"), [5(a)](#S5.F5.sf1 "In Figure 5 ‣ 5 Results Analysis ‣ Technical Report v2 Characterization of P2P IPTV Traffic: Scaling Analysis") and [5(b)](#S5.F5.sf2 "In Figure 5 ‣ 5 Results Analysis ‣ Technical Report v2 Characterization of P2P IPTV Traffic: Scaling Analysis"), [6(a)](#S5.F6.sf1 "In Figure 6 ‣ 5 Results Analysis ‣ Technical Report v2 Characterization of P2P IPTV Traffic: Scaling Analysis") and [6(b)](#S5.F6.sf2 "In Figure 6 ‣ 5 Results Analysis ‣ Technical Report v2 Characterization of P2P IPTV Traffic: Scaling Analysis"). Removing the signaling traffic has no impact on the upload traffic.
Regarding the download traffic, the video download energy spectra (Fig. [3(d)](#S5.F3.sf4 "In Figure 3 ‣ 5 Results Analysis ‣ Technical Report v2 Characterization of P2P IPTV Traffic: Scaling Analysis") [4(d)](#S5.F4.sf4 "In Figure 4 ‣ 5 Results Analysis ‣ Technical Report v2 Characterization of P2P IPTV Traffic: Scaling Analysis") [5(d)](#S5.F5.sf4 "In Figure 5 ‣ 5 Results Analysis ‣ Technical Report v2 Characterization of P2P IPTV Traffic: Scaling Analysis") and [6(d)](#S5.F6.sf4 "In Figure 6 ‣ 5 Results Analysis ‣ Technical Report v2 Characterization of P2P IPTV Traffic: Scaling Analysis")) are different from their corresponding overall download energy spectra (Fig. [3(c)](#S5.F3.sf3 "In Figure 3 ‣ 5 Results Analysis ‣ Technical Report v2 Characterization of P2P IPTV Traffic: Scaling Analysis") [4(c)](#S5.F4.sf3 "In Figure 4 ‣ 5 Results Analysis ‣ Technical Report v2 Characterization of P2P IPTV Traffic: Scaling Analysis") [5(c)](#S5.F5.sf3 "In Figure 5 ‣ 5 Results Analysis ‣ Technical Report v2 Characterization of P2P IPTV Traffic: Scaling Analysis") and [6(c)](#S5.F6.sf3 "In Figure 6 ‣ 5 Results Analysis ‣ Technical Report v2 Characterization of P2P IPTV Traffic: Scaling Analysis")). Removing signaling traffic from the download traffic has clearly an impact on the download traffic because it modifies the download energy spectra.
Table [2](#S4.T2 "Table 2 ‣ 4.2 Signaling traffic filtering heuristic ‣ 4 Analysis Methodology ‣ Technical Report v2 Characterization of P2P IPTV Traffic: Scaling Analysis") shows the signaling traffic ratio for all the applications in upload and download. For all the applications, the signaling traffic represents a larger part in the download traffic than the upload traffic.
The upload signaling traffic is only provided by our controlled node to other peers in the Internet. Since our node has high bandwidth capabilities, it serves video to many other peers. Table [1](#S3.T1 "Table 1 ‣ 3.2 Measurement experiments platform ‣ 3 Experiments ‣ Technical Report v2 Characterization of P2P IPTV Traffic: Scaling Analysis") indicates that the amount of upload traffic is 3 to 6 times larger than the amount of download traffic. The signaling traffic sent by our node to other peers in the Internet counts only for a small part of the overall upload traffic.
The download signaling traffic is provided by other peers in the Internet to our controlled node. Our node just needs to download the video at the video bitrate. The download signaling traffic coming from many other peers counts for a large part of the overall download traffic. This explains the impact of signaling traffic on the download traffic.
To summarize our observations, the signaling traffic has no impact on the upload traffic but it has an impact on the download traffic. In the following, we will discuss the stationnarity of the traffic that will help to better characterize the properties of the P2P IPTV signaling traffic.
### 5.4 Traffic stationnarity
[Refer to caption](/html/0704.3228/assets/x22.png)
\(a\)
Upload: overall packet traffic (signaling and video traffic)
[Refer to caption](/html/0704.3228/assets/x23.png)
\(b\)
Upload: video packet traffic
[Refer to caption](/html/0704.3228/assets/x24.png)
\(c\)
Download: overall packet traffic (signaling and video traffic)
[Refer to caption](/html/0704.3228/assets/x25.png)
\(d\)
Download: video packet traffic
Figure 7: TCP applications: PPLive. Traffic stationarity for the three equal parts of the traffic. The blue solid line is the first part, the dashed red line is the second part and the dash-dotted green line is the third part of the traffic.
[Refer to caption](/html/0704.3228/assets/x26.png)
\(a\)
Upload: overall packet traffic (signaling and video traffic)
[Refer to caption](/html/0704.3228/assets/x27.png)
\(b\)
Upload: video packet traffic
[Refer to caption](/html/0704.3228/assets/x28.png)
\(c\)
Download: overall packet traffic (signaling and video traffic)
[Refer to caption](/html/0704.3228/assets/x29.png)
\(d\)
Download: video packet traffic
Figure 8: UDP applications: SOPCast. Traffic stationarity for the three equal parts of the traffic. The blue solid line is the first part, the dashed red line is the second part and the dash-dotted green line is the third part of the traffic.
From all the upload energy spectra (Fig [3(a)](#S5.F3.sf1 "In Figure 3 ‣ 5 Results Analysis ‣ Technical Report v2 Characterization of P2P IPTV Traffic: Scaling Analysis") [3(b)](#S5.F3.sf2 "In Figure 3 ‣ 5 Results Analysis ‣ Technical Report v2 Characterization of P2P IPTV Traffic: Scaling Analysis"), Fig [4(a)](#S5.F4.sf1 "In Figure 4 ‣ 5 Results Analysis ‣ Technical Report v2 Characterization of P2P IPTV Traffic: Scaling Analysis") [4(b)](#S5.F4.sf2 "In Figure 4 ‣ 5 Results Analysis ‣ Technical Report v2 Characterization of P2P IPTV Traffic: Scaling Analysis"), Fig [5(a)](#S5.F5.sf1 "In Figure 5 ‣ 5 Results Analysis ‣ Technical Report v2 Characterization of P2P IPTV Traffic: Scaling Analysis") [5(b)](#S5.F5.sf2 "In Figure 5 ‣ 5 Results Analysis ‣ Technical Report v2 Characterization of P2P IPTV Traffic: Scaling Analysis"), and Fig [6(a)](#S5.F6.sf1 "In Figure 6 ‣ 5 Results Analysis ‣ Technical Report v2 Characterization of P2P IPTV Traffic: Scaling Analysis") [6(b)](#S5.F6.sf2 "In Figure 6 ‣ 5 Results Analysis ‣ Technical Report v2 Characterization of P2P IPTV Traffic: Scaling Analysis")), we observe a linear increase, starting from $`j = 6`$ for SOPCast, from $`j = 9`$ for PPLive, from $`j = 12`$ for PPStream or $`j = 10`$ for TVAnts. We also show that the linear increase exists in the download traffic (Fig. [3(c)](#S5.F3.sf3 "In Figure 3 ‣ 5 Results Analysis ‣ Technical Report v2 Characterization of P2P IPTV Traffic: Scaling Analysis") [4(c)](#S5.F4.sf3 "In Figure 4 ‣ 5 Results Analysis ‣ Technical Report v2 Characterization of P2P IPTV Traffic: Scaling Analysis") [5(c)](#S5.F5.sf3 "In Figure 5 ‣ 5 Results Analysis ‣ Technical Report v2 Characterization of P2P IPTV Traffic: Scaling Analysis") and [6(c)](#S5.F6.sf3 "In Figure 6 ‣ 5 Results Analysis ‣ Technical Report v2 Characterization of P2P IPTV Traffic: Scaling Analysis")) but it is modified (Fig. [4(d)](#S5.F4.sf4 "In Figure 4 ‣ 5 Results Analysis ‣ Technical Report v2 Characterization of P2P IPTV Traffic: Scaling Analysis")) or wasted (Fig. [3(d)](#S5.F3.sf4 "In Figure 3 ‣ 5 Results Analysis ‣ Technical Report v2 Characterization of P2P IPTV Traffic: Scaling Analysis") [5(d)](#S5.F5.sf4 "In Figure 5 ‣ 5 Results Analysis ‣ Technical Report v2 Characterization of P2P IPTV Traffic: Scaling Analysis") [6(d)](#S5.F6.sf4 "In Figure 6 ‣ 5 Results Analysis ‣ Technical Report v2 Characterization of P2P IPTV Traffic: Scaling Analysis")) when removing signaling traffic.
We already show that signaling traffic has an impact on the download traffic because it count for a larger amount of data. The signaling traffic may also lead to the linear increase in the download energy spectra.
A linear increase indicates a possible long-range dependency of the traffic (LRD). It means that the traffic fluctuates largely and is not predictable. With such traffic fluctuations, it becomes impossible to forecast the traffic behavior or to make network provisioning.
The signaling traffic is used by peers to get the video chunks they need. If the signaling traffic is responsible for LRD, the signaling traffic generates itself the troubles to download the video. In other words, the design of the applications would be not efficient if the signaling traffic generates the long-range dependency of the traffic.
For a P2P network, a LRD traffic indicates, there is no stability in the network traffic. Then, it becomes a hard task to provide QoS parameters (delay, bandwidth, jiiter) to users because networks conditions are always changing.
Regarding P2P IPTV systems, which are delay sensitive, the traffic LRD has to be avoided because it will directly affect and decrease the quality of video reception. For example, under a high churn of peers, each peer has always to discover new peers and to establish new partnerships with other peers to receive the video data chunks. In this case, there would be no stability in the overlay network because of the disruptions in the peers connections and communications. This would lead to non-predictable traffic and long-range dependency of the traffic.
In this section, we want to know if we really observe a long-range dependency of the traffic. To this end, we study the stationnarity of the traffic. We split each trace in three equal parts, and we used the previous wavelet transform method to analyze all the parts of the traffic. We visually control the traffic stability with LDEstimate.
Due to space limitations, we only present one example for TCP applications (PPLive, Fig. [10](#Sx1.F10 "Figure 10 ‣ Appendix ‣ Technical Report v2 Characterization of P2P IPTV Traffic: Scaling Analysis")) and one example for UDP application (SOPCast, Fig. [11](#Sx1.F11 "Figure 11 ‣ Appendix ‣ Technical Report v2 Characterization of P2P IPTV Traffic: Scaling Analysis")). The other TCP applications energy spectra are similar to the TCP example (Fig. [10](#Sx1.F10 "Figure 10 ‣ Appendix ‣ Technical Report v2 Characterization of P2P IPTV Traffic: Scaling Analysis")) and can be found in the appendix.
In each logscale diagram, we plot the three parts of the trace. The blue solid line is the first part of the trace, the dashed red line is the second part and the green dash-dotted line is the third part of the trace. The traffic is stationary if each part of the traffic looks like the other parts and has the same energy level.
For TCP applications (Fig. [10](#Sx1.F10 "Figure 10 ‣ Appendix ‣ Technical Report v2 Characterization of P2P IPTV Traffic: Scaling Analysis")), the three parts of the energy spectra are similar at small time scales but become different at large time scales (about $`j = 10`$). There is stationnarity in the traffic until $`j = 10`$. The traffic is not stationary beyond $`j = 10`$.
As shown in section [5.2](#S5.SS2 "5.2 TCP applications vs. UDP applications ‣ 5 Results Analysis ‣ Technical Report v2 Characterization of P2P IPTV Traffic: Scaling Analysis"), the TCP applications experiment a bump in their energy spectra at time scale $`j = 8`$, whatever the traffic direction or its nature. At this time scale, the traffic is stationary and it demonstrates that the observed energy bumps are essential phenomenon and not simply artifact from non-stationnarity. On the contrary, the linear increase observed from $`j = 9`$ is not a long-range dependency of the traffic.
For UDP applications (Fig. [11](#Sx1.F11 "Figure 11 ‣ Appendix ‣ Technical Report v2 Characterization of P2P IPTV Traffic: Scaling Analysis")), the three parts of the traffic are similar and increase linearly. The traffic of UDP applications is stationary. The linear increase characterizes a long-range dependency of the UDP applications traffic represented by SOPCast. Removing signaling traffic modifies the linear increase in the SOPCast download traffic. For UDP applications, signaling traffic may lead the to the long-range dependency in the traffic.
In this section, we show an important design difference between TCP and UDP applications. The traffic of UDP applications is stationary and present long-range dependency whereas the traffic of TCP applications is not stationary and does not experiments long-range dependency.
### 5.5 Analysis of the top download flows
[Refer to caption](/html/0704.3228/assets/x30.png)
\(a\)
PPLive Download: overall packet traffic (signaling and video traffic)
[Refer to caption](/html/0704.3228/assets/x31.png)
\(b\)
PPLive Download: video packet traffic
[Refer to caption](/html/0704.3228/assets/x32.png)
\(c\)
SOPCast Download: overall packet traffic (signaling and video traffic)
[Refer to caption](/html/0704.3228/assets/x33.png)
\(d\)
SOPCast Download: video packet traffic
Figure 9: Top download flows
In the previous experiments, we make our observations based on the aggregate traffic received and sent by our controlled peers situated in our campus network. Our nodes has high bandwidth capacities and they play an important role in duplicating data to a large amount of peers. The high-speed access of our nodes help to get the video efficiently compare to a residential access to the Internet, in which received packets are limited by a smaller download rate.
To extend our previous findings without being strongly related to our network environment, we want to analyze the properties of a single flow instead of observing the aggregate traffic in our contoled peers. Whatever the network environment is, a peer will try to download the video at its bitrate. On the contrary, the number of duplicate flows a peer could upload is directly related to the capacities of the network environment. Thus, we limit the scope of this experiment to the download traffic flows because this is a general observation of the traffic that does not directly depend on the network environment.
For all the traces, we isolated the top peer that sent the biggest amount of data to our nodes. We refer to these resulting flows as top download flows since the data transported by these flows are downloaded by our controlled nodes.
Table [4](#S5.T4 "Table 4 ‣ 5.5 Analysis of the top download flows ‣ 5 Results Analysis ‣ Technical Report v2 Characterization of P2P IPTV Traffic: Scaling Analysis") summarizes the amount of data carried by each top flow for all the applications.
Table 4: Top download flows
<table>
<thead>
<tr>
<th></th>
<th colspan="2">Overall traffic</th>
<th colspan="2">Video traffic</th>
</tr>
</thead>
<tbody>
<tr>
<td></td>
<td>Volume (MB)</td>
<td>#Packets</td>
<td>Volume (MB)</td>
<td>#Packets</td>
</tr>
<tr>
<td>PPLive</td>
<td>106.30</td>
<td>82,428</td>
<td>105.97</td>
<td>76,797</td>
</tr>
<tr>
<td>PPStream</td>
<td>13.21</td>
<td>23,951</td>
<td>11.79</td>
<td>9,809</td>
</tr>
<tr>
<td>SOPCast</td>
<td>61.91</td>
<td>63,054</td>
<td>59.20</td>
<td>47,187</td>
</tr>
<tr>
<td>TVAnts</td>
<td>53.86</td>
<td>41,740</td>
<td>52.13</td>
<td>37,460</td>
</tr>
</tbody>
</table>
We already show in \[[13](#bib.bib13)\] that all the applications do not implement the same mechanisms to download the video. According to the applications, the video can be received from only a few provider peers at the same time or from many peers and the video peers session durations are various. This explains why the amount of data transported by the top flows are different for all the applications.
The top download flows sent by the top peer to our nodes will be analyzed by using wavelet based transform method with LDEstimate, similarly to the previous experiments. Due to space limitation, we only present on Fig. [13](#Sx1.F13 "Figure 13 ‣ Appendix ‣ Technical Report v2 Characterization of P2P IPTV Traffic: Scaling Analysis"), the energy spectra for a single TCP application (PPLive Fig [12(a)](#Sx1.F12.sf1 "In Figure 12 ‣ Appendix ‣ Technical Report v2 Characterization of P2P IPTV Traffic: Scaling Analysis") and [12(b)](#Sx1.F12.sf2 "In Figure 12 ‣ Appendix ‣ Technical Report v2 Characterization of P2P IPTV Traffic: Scaling Analysis")) and for UDP application (SOPCast Fig. [12(c)](#Sx1.F12.sf3 "In Figure 12 ‣ Appendix ‣ Technical Report v2 Characterization of P2P IPTV Traffic: Scaling Analysis") and [12(d)](#Sx1.F12.sf4 "In Figure 12 ‣ Appendix ‣ Technical Report v2 Characterization of P2P IPTV Traffic: Scaling Analysis")). The other TCP applications plots are similar to the presented TCP application and can be found in the appendix.
We notice for the two applications that their video energy spectra look similar to their overall energy spectra. This was expected because these flows are sent by the top contributor peers and transport almost entirely video packets and not signaling packets. Removing signaling traffic on these flows can only have a limited impact, depending on the signaling packets in the flows.
For example, the top download flow of SOPCast transport 15,867 packets of signaling traffic ($`63054 - 47187`$) counting for 2.71 MBytes whereas PPLive top download flow transport only 5,731 signaling packets ($`82428 - 76797`$) counting for 0.33 MBytes.
Regarding TCP applications Fig. [12(a)](#Sx1.F12.sf1 "In Figure 12 ‣ Appendix ‣ Technical Report v2 Characterization of P2P IPTV Traffic: Scaling Analysis") and [12(b)](#Sx1.F12.sf2 "In Figure 12 ‣ Appendix ‣ Technical Report v2 Characterization of P2P IPTV Traffic: Scaling Analysis"), until time scale $`j = 10`$, the energy spectra of the top flow look similar to the aggregate traffic. Beyond this time scale, the energy spectra of the top download flow are different from the aggregate traffic because the energy spectra of the top flow are increasing.
With UDP applications, until time scale $`j = 11`$, the energy spectra of the top flow are different from the aggregate traffic. Fig. [12(c)](#Sx1.F12.sf3 "In Figure 12 ‣ Appendix ‣ Technical Report v2 Characterization of P2P IPTV Traffic: Scaling Analysis") and [12(d)](#Sx1.F12.sf4 "In Figure 12 ‣ Appendix ‣ Technical Report v2 Characterization of P2P IPTV Traffic: Scaling Analysis") show an energy bump at time scale $`j = 3`$, then the energy spectra increase slightly from $`j = 4`$ to $`j = 11`$. Beyond $`j = 11`$, we observe the linear increase usually observed for the UDP energy spectra.
In this experiment, we observe that the top flows in the download traffic do not have the same scaling properties as the aggregate download traffic. We did the same experiments for the 10th top download flows (i.e. the 10th flow according to data volume transported). The 10th top flows present the same scaling properties as the top flows. The plots for the 10th top flows can be shown in the appendix.
The aggregate traffic is not only the mix of every single flow. The granularity of the P2P IPTV traffic has to be taken into account when designing P2P IPTV traffic models.
## 6 Results Discussion
In this work, we analyzed the P2P IPTV traffic by using a wavelet based transform method. This allows us to characterize this traffic and to understand its properties and impact on the network. Thanks to our original P2P IPTV traffic analysis, we have many new findings and observations that have to be summarized and discussed.
First of all, we observed that the energy spectra of TCP applications are different from the energy spectra of UDP applications (section [5.2](#S5.SS2 "5.2 TCP applications vs. UDP applications ‣ 5 Results Analysis ‣ Technical Report v2 Characterization of P2P IPTV Traffic: Scaling Analysis")). One of the most relevant difference is the energy bump observed in the spectra of TCP applications at time scale $`j = 8`$ (5.12s), which indicates a possible periodic behavior in the traffic. Intuitively, we could believe these differences come from the two different transport protocols used. However, a 5 seconds periodic behaviors is a very long duration for TCP mechanisms and TCP should not be the responsible of this periodic behavior. With a simple application design difference, the scaling properties of the generated traffic do not have the same impact on the network.
Secondly, for all the applications, the signaling traffic represents a larger part in the download traffic than the upload traffic (section [5.3](#S5.SS3 "5.3 Impact of the signaling traffic ‣ 5 Results Analysis ‣ Technical Report v2 Characterization of P2P IPTV Traffic: Scaling Analysis")). The signaling traffic has clearly an impact on the scaling properties of the download traffic and has no impact on the upload traffic. This observation is important since signaling traffic is necessary to coordinate the data exchanges in such P2P systems. For scalability reasons, the amount of signaling traffic has to be kept as low as possible. The download signaling traffic comes from other peers on the Internet that request the video data. Efforts have to be made to reduce the number of packets sent by the signaling protocol to get the video data and to preserve the scalability of these systems in the network.
Then, the previous observation highlights an important point when modeling P2P IPTV traffic: the download traffic has not the same properties as the upload traffic. The differences between both sides of the traffic (i.e. upload and download) have to be taken into account carefully when designing synthetic traffic generation models.
The generated traffic of TCP applications is not stationary beyond time scale $`j = 10`$. On the contrary, the traffic of UDP applications is stationary. As shown in section [5.4](#S5.SS4 "5.4 Traffic stationnarity ‣ 5 Results Analysis ‣ Technical Report v2 Characterization of P2P IPTV Traffic: Scaling Analysis"), the stationnarity experiment proves that the signaling traffic of UDP application involves long-range dependency in the download traffic. The UDP application experiments also long-range dependency in the upload traffic. In presence of traffic LRD, the network conditions are always changing and it becomes a hard task to provide QoS parameters as delay for users to get good quality video.
This finding highlights the not so trivial choice of transport protocols for P2P IPTV traffic. It is usually admitted that the non-elastic data transfer -as video- has to rely on UDP but we show that UDP traffic may lead to trouble in the network traffic.
Finally, the aggregate download traffic of P2P IPTV systems has not exactly the same scaling properties as the top download flow (section [5.5](#S5.SS5 "5.5 Analysis of the top download flows ‣ 5 Results Analysis ‣ Technical Report v2 Characterization of P2P IPTV Traffic: Scaling Analysis")). The granularity of the traffic has to be taken into account when designing P2P IPTV traffic models. A P2P IPTV traffic model based only on flows properties would fail to capture the global characteristics of the aggregate traffic. The use of an inappropriate traffic model would lead to wrong results when simulating new architectures with such significant input parameter.
## 7 Conclusion
In this paper, we analyzed network traffic generated by P2P IPTV applications. We performed an extensive measurement campaign during the 2006 FIFA World Cup and we measured the most popular P2P IPTV applications on the Internet. We used wavelet transform based method to study the P2P IPTV traffic at different time scales and to characterize its properties.
Our multiscale traffic analysis show how different are the scaling properties of the TCP and UDP traffics. For all the applications, the signaling traffic has a significant impact on the download traffic but not on the upload traffic. It involves scalability concerns regarding the P2P IPTV signaling protocols used to download the video data. The UDP traffic is stationary and leads to long-range dependency of the traffic. The choice of UDP as transport protocol for non-elastic transfers in P2P networks becomes not so trivial since the traffic LRD indicates that the traffic is not predictable in the network. The scaling properties of the download traffic are different from the upload traffic. The traffic granularity and both traffic directions have to be taken into account to model P2P IPTV traffic accurately.
Currently, we are analyzing the traffic collected during other games and under different network environments to extend our observations. It will allow us to have a finer analysis of our findings and could also help to answer to the open questions introduced by this work. In a long-term work, the characterization of the P2P IPTV traffic will help us to accurately model and simulate such systems.
## References
- \[1\] B. Cohen, “Incentives build robustness in bittorrent,” 2003.
- \[2\] http://www.edonkey2000.com.
- \[3\] http://www.youtube.com.
- \[4\] http://www.pplive.com.
- \[5\] http://www.ppstream.com.
- \[6\] http://www.sopcast.com.
- \[7\] http://www.tvants.com.
- \[8\] X. Zhang, J. Liu, and B. Li, “On large-scale peer-to-peer live video distribution: Coolstreaming and its preliminary experimental results,” in *Proc. MMSP*, 2005.
- \[9\] X. Zhang, J. Liu, B. Li, and T. P. Yum, “Coolstreaming/donet: A data-driven overlay network for peer-to-peer live media streaming,” in *Proc. Infocom*, 2005.
- \[10\] X. Hei, C. Liang, J. Liang, Y. Liu, and K. W. Ross, “Insights into pplive: A measurement study of a large-scale p2p iptv system,” in *Proc. of IPTV Workshop, International World Wide Web Conference*, 2006.
- \[11\]
X. Hei, C. Liang, J. Liang, Y. Liu, and K. W. Ross, “A measurement study of a large-scale p2p iptv system,” to appear in *IEEE Transactions on Multimedia*, 2007.
- \[12\] L. Vu, I. Gupta, J. Liang, and K. Nahrstedt, “Measurement and modeling of a large-scale overlay for multimedia streaming,” in *Proc. of QSHINE’07, International Conference on Heterogeneous Networking for Quality, Reliability, Security and Robustness*, 2007.
- \[13\] T. Silverston and O. Fourmaux, “Measuring p2p iptv systems,” in *Proc. of NOSSDAV’07, International Workshop on Network and Operating Systems Support for Digital Audio & Video*, 2007.
- \[14\] S. Ali, A. Mathur, and H. Zhang, “Measurement of commercial peer-to-peer live video streaming,” in *Proc. of Workshop in Recent Advances in Peer-to-Peer Streaming*, 2006.
- \[15\] X. Hei, Y. Liu, and K. W. Ross, “Inferring network-wide quality in p2p live streaming systems.” \[Online\]. Available: http://cis.poly.edu/ ross/papers/buffermap.pdf
- \[16\] http://content.lip6.fr/traces/.
- \[17\] http://www.joost.com.
- \[18\] D. Veitch and P. Abry, “Matlab code for the wavelet based analysis of scaling processes.” \[Online\]. Available: http://www.cubinlab.ee.mu.oz.au/$`\sim`$darryl
## Appendix
[Refer to caption](/html/0704.3228/assets/x34.png)
\(a\)
Upload: overall packet traffic (signaling and video traffic)
[Refer to caption](/html/0704.3228/assets/x35.png)
\(b\)
Upload: video packet traffic
[Refer to caption](/html/0704.3228/assets/x36.png)
\(c\)
Download: overall packet traffic (signaling and video traffic)
[Refer to caption](/html/0704.3228/assets/x37.png)
\(d\)
Download: video packet traffic
Figure 10: PPStream: Traffic stationarity for the three equal parts of the traffic. The blue solid line is the first part, the dashed red line is the second part and the dash-dotted green line is the third part of the traffic.
[Refer to caption](/html/0704.3228/assets/x38.png)
\(a\)
Upload: overall packet traffic (signaling and video traffic)
[Refer to caption](/html/0704.3228/assets/x39.png)
\(b\)
Upload: video packet traffic
[Refer to caption](/html/0704.3228/assets/x40.png)
\(c\)
Download: overall packet traffic (signaling and video traffic)
[Refer to caption](/html/0704.3228/assets/x41.png)
\(d\)
Download: video packet traffic
Figure 11: TVAnts: Traffic stationarity for the three equal parts of the traffic. The blue solid line is the first part, the dashed red line is the second part and the dash-dotted green line is the third part of the traffic.
[Refer to caption](/html/0704.3228/assets/x42.png)
\(a\)
PPStream Download: overall packet traffic (signaling and video traffic)
[Refer to caption](/html/0704.3228/assets/x43.png)
\(b\)
PPStream Download: video packet traffic
[Refer to caption](/html/0704.3228/assets/x44.png)
\(c\)
TVAnts Download: overall packet traffic (signaling and video traffic)
[Refer to caption](/html/0704.3228/assets/x45.png)
\(d\)
TVAnts Download: video packet traffic
Figure 12: Top download flows: PPStream: (a) and (b). TVAnts: (c) and (d)
[Refer to caption](/html/0704.3228/assets/x46.png)
\(a\)
PPLive Download: overall packet traffic (signaling and video traffic)
[Refer to caption](/html/0704.3228/assets/x47.png)
\(b\)
PPLive Download: video packet traffic
[Refer to caption](/html/0704.3228/assets/x48.png)
\(c\)
SOPCast Download: overall packet traffic (signaling and video traffic)
[Refer to caption](/html/0704.3228/assets/x49.png)
\(d\)
SOPCast Download: video packet traffic
[Refer to caption](/html/0704.3228/assets/x50.png)
\(e\)
PPStream Download: overall packet traffic (signaling and video traffic)
[Refer to caption](/html/0704.3228/assets/x51.png)
\(f\)
PPStream Download: video packet traffic
[Refer to caption](/html/0704.3228/assets/x52.png)
\(g\)
TVAnts Download: overall packet traffic (signaling and video traffic)
[Refer to caption](/html/0704.3228/assets/x53.png)
\(h\)
TVAnts Download: video packet traffic
Figure 13: 10th top download flows: PPLive: (a) and (b). SOPcast: (c) and (d). PPStream: (e) and (f). TVAnts: (g) and (h) |
0705.0819 | cs.NI | cyber | # Netsukuku topology
http://netsukuku.freaknet.org
AlpT (@freaknet.org)
###### Abstract
In this document, we describe the fractal structure of the Netsukuku topology. Moreover, we show how it is possible to use the QSPN v2 on the high levels of the fractal.
This document is part of Netsukuku.
Copyright ©2007 Andrea Lo Pumo aka AlpT $`<`$alpt@freaknet.org$`>`$. All rights reserved.
This document is free; you can redistribute it and/or modify it under the terms of the GNU General Public License as published by the Free Software Foundation; either version 2 of the License, or (at your option) any later version.
This document is distributed in the hope that it will be useful, but WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for more details.
You should have received a copy of the GNU General Public License along with this document; if not, write to the Free Software Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
###### Contents
1. [](#S1 "In Netsukuku topology")
1 Preface
2. [](#S2 "In Netsukuku topology")
2 The general idea
3. [](#S3 "In Netsukuku topology")
3 Basic definitions
4. [](#S4 "In Netsukuku topology")
4 Network topology
1. [](#S4.SS1 "In 4 Network topology ‣ Netsukuku topology")
4.1 Fractal topology
1. [](#S4.SS1.SSS1 "In 4.1 Fractal topology ‣ 4 Network topology ‣ Netsukuku topology")
4.1.1 Level 1
2. [](#S4.SS1.SSS2 "In 4.1 Fractal topology ‣ 4 Network topology ‣ Netsukuku topology")
4.1.2 Level n
3. [](#S4.SS1.SSS3 "In 4.1 Fractal topology ‣ 4 Network topology ‣ Netsukuku topology")
4.1.3 Membership
2. [](#S4.SS2 "In 4 Network topology ‣ Netsukuku topology")
4.2 Fractal map
1. [](#S4.SS2.SSS1 "In 4.2 Fractal map ‣ 4 Network topology ‣ Netsukuku topology")
4.2.1 IP v4 and v6
2. [](#S4.SS2.SSS2 "In 4.2 Fractal map ‣ 4 Network topology ‣ Netsukuku topology")
4.2.2 Internal and external map
3. [](#S4.SS2.SSS3 "In 4.2 Fractal map ‣ 4 Network topology ‣ Netsukuku topology")
4.2.3 Bnode map
3. [](#S4.SS3 "In 4 Network topology ‣ Netsukuku topology")
4.3 CIDR routing
5. [](#S5 "In Netsukuku topology")
5 Tracer Packets in high levels
1. [](#S5.SS1 "In 5 Tracer Packets in high levels ‣ Netsukuku topology")
5.1 A gnode is a node
1. [](#S5.SS1.SSS1 "In 5.1 A gnode is a node ‣ 5 Tracer Packets in high levels ‣ Netsukuku topology")
5.1.1 Example of a wandering TP
6. [](#S6 "In Netsukuku topology")
6 QSPN v2 in high levels
1. [](#S6.SS1 "In 6 QSPN v2 in high levels ‣ Netsukuku topology")
6.1 Endless loops
2. [](#S6.SS2 "In 6 QSPN v2 in high levels ‣ Netsukuku topology")
6.2 Route Efficiency Measure
3. [](#S6.SS3 "In 6 QSPN v2 in high levels ‣ Netsukuku topology")
6.3 Flat levels
4. [](#S6.SS4 "In 6 QSPN v2 in high levels ‣ Netsukuku topology")
6.4 Unique flood
7. [](#S7 "In Netsukuku topology")
7 Network dynamics
1. [](#S7.SS1 "In 7 Network dynamics ‣ Netsukuku topology")
7.1 Radar
2. [](#S7.SS2 "In 7 Network dynamics ‣ Netsukuku topology")
7.2 Level 0
3. [](#S7.SS3 "In 7 Network dynamics ‣ Netsukuku topology")
7.3 Level n
4. [](#S7.SS4 "In 7 Network dynamics ‣ Netsukuku topology")
7.4 Hooking phase
5. [](#S7.SS5 "In 7 Network dynamics ‣ Netsukuku topology")
7.5 Gnode hook
8. [](#S8 "In Netsukuku topology")
8 ChangeLog
## 1 Preface
We’re assuming that you already know the basics of the QSPN. If not, read the QSPN document first: \[[1](#bib.bib1)\].
## 2 The general idea
The aim of Netsukuku is to be a (physical) scalable mesh network, completely distributed and decentralised, anonymous and autonomous.
The software, which must be executed by every node of the net, has to be unobtrusive. It has to use very few CPU and memory resources, in this way it will be possible to run it inside low-performance computers, like Access Points, embedded devices and old computers.
If this requirements are met, Netsukuku can be easily used to build a worldwide distributed, anonymous and not controlled network, separated from the Internet, without the support of any servers, ISPs or control authorities.
## 3 Basic definitions
Node
We call *node* any computer that is hooked up to the Netsukuku network.
Rnode
stands for remote node: given a node X, it is any other node directly linked to X, i.e. it’s a neighbour of X.
Map
A map is a file, kept by each node, which contains all the necessary information about the network, f.e. routes and nodes status.
Example:
[Refer to caption](/html/0705.0819/assets/fig/segABC.png)
Figure 1: The nodes A,B and C
A is the rnode of B.
B is the rnode of A and C.
C is the rnode of B.
## 4 Network topology
A simple topology, which doesn’t impose any structure on the network, can be memorised with a simple map. In this map, all the information regarding the nodes of the network have to be memorised. Surely, this kind of map cannot be utilised by Netsukuku, because it would require too much memory. For example, even if we store just one route to reach one node and even if this route costs one byte, we would need 1Gb of memory for a network composed by $`10^{9}`$ nodes (the current Internet).
For this reason, it’s necessary to structure the network in a convenient topology.
### 4.1 Fractal topology
#### 4.1.1 Level 1
First of all we’ll subdivide the network in groups of 256 nodes and we’ll use the following definitions:
Gnode
means group node. It is a group of nodes, i.e. a set of nodes. Each node of the network belongs to just one gnode.
A gnode contains a maximum of 256 nodes.
By writing $`n \in G`$ we mean that the node $`n`$ belongs to the gnode $`G`$.
Bnode
stands for border node. It is a node which belongs to a gnode G, but that is also directly linked to at least one node of another gnode, i.e. some of its rnodes belongs to different gnodes than its.
By writing $`b \in G`$ we mean that the bnode $`b`$ belongs to the gnode $`G`$.
Example:
[Refer to caption](/html/0705.0819/assets/fig/bnode.png)
Figure 2: The bnode A and B, belonging respectively to the gnode G and $`G'`$
$`A \in G`$, A is a node belonging to the gnode G, its rnode is B.
$`B \in G'`$, B is a node belonging to the gnode $`G'`$, its rnode is A.
A is a bnode of G, while B is a bnode of $`G'`$.
#### 4.1.2 Level n
We further subdivide the network topology in *groups of 256 groups of nodes* and we continue to name them as gnode.
At this point, we repeat recursively this subdivision process until we can group all the nodes of the network into a single gnode.
Doing so, we’ve structured the network in $`n + 1`$ levels (from $`0`$ to $`n`$).
In the base level (level 0), there are 256 single nodes.
In the first level (level 1), there are 256 normal gnodes. Each of them contains 256 single nodes.
In the second (level 2), 256 gnodes of level 1 forms a single *group of groups of nodes*.
In the third (level 3), there are 256 groups of 256 groups of 256 groups of 256 nodes.
Continuing in this way, we arrive at the last level (level $`n`$), where there is a single group which contains the whole network.
The QSPN algorithm is able to operate independently on any level, considering each gnode as a single node of level 0. For this reason, we can view the Netsukuku topology as a fractal, where each level is composed by single nodes.
#### Example
Figure [3](#S4.F3 "Figure 3 ‣ Example ‣ 4.1 Fractal topology ‣ 4 Network topology ‣ Netsukuku topology")
<sup>1</sup>
<sup>1</sup>1this figure has been taken from: http://www.ian.org/FX/Plugins.html
is an example of the fractal topology of Netsukuku.
[Refer to caption](/html/0705.0819/assets/fig/fractal_circle.png)
Figure 3: An example of the netsukuku topology structure
In this topology, each gnode contains four nodes, i.e. each group contains four elements. The network is structured in 6 levels.
The red elements, are single nodes (level 0).
Four nodes forms a single group of nodes (level 1).
A single bright green circle is a group of groups of nodes (level 2).
The dark green circles are groups of groups of groups of nodes (level 3).
The dark blue circle are groups of groups of groups of groups of nodes (level 4).
Finally, the bright blue circle is the gnode which contains the whole network (level 5).
#### 4.1.3 Membership
Let’s assign a numeric ID to each (g)gnode, starting from the last level:
1. 1\.
in the last level ($`n`$) there’s only one giant gnode, thus we assign to it the ID “0”. Our global ID will be:
<table>
<tbody>
<tr>
<td></td>
<td>00</td>
<td></td>
</tr>
</tbody>
</table>
2. 2\.
in $`n - 1`$ there are 256 gnodes, which belongs to the gnode 0 of level $`n`$, thus we assign them the IDs from $`0`$ to $`255`$. The global ID becomes:
<table>
<tbody>
<tr>
<td></td>
<td>0⋅i0≤i≤255⋅0𝑖0𝑖2550\cdot i\quad 0\leq i\leq 255</td>
<td></td>
</tr>
</tbody>
</table>
3. 3\.
we repeat the step 2 recursively gaining an ID of this form:
<table>
<tbody>
<tr>
<td></td>
<td>0⋅in−1⋅in−2⋅⋯⋅i00≤ij≤255, 0≤j≤n−1formulae-sequence⋅0subscript𝑖𝑛1subscript𝑖𝑛2⋯subscript𝑖00subscript𝑖𝑗255 0𝑗𝑛10\cdot i_{n-1}\cdot i_{n-2}\cdot\dots\cdot i_{0}\quad 0\leq i_{j}\leq 255,\;0\leq j\leq n-1</td>
<td></td>
</tr>
</tbody>
</table>
4. 4\.
since the last level is always $`0`$, we’ll omit it and we’ll consider only the first $`n`$ levels.
In a network with a maximum of $`2^{32}`$ nodes (the maximum allowed by the ipv4), there would be five levels ($`n = 4`$), where each gnode will be composed by 256 nodes. Therefore, the ID will be in the usual IP form:
<table>
<tbody>
<tr>
<td></td>
<td>0…255⋅0…255⋅0…255⋅0…255⋅⋅⋅0…2550…2550…2550…2550\dots 255\cdot 0\dots 255\cdot 0\dots 255\cdot 0\dots 255</td>
<td></td>
</tr>
</tbody>
</table>
For example, a single node of level 0 of the network is:
<table>
<tbody>
<tr>
<td></td>
<td>3⋅41⋅5⋅12⋅3415123\cdot 41\cdot 5\cdot 12</td>
<td></td>
</tr>
</tbody>
</table>
That said, each gnode of the network belongs to only one combination of gnodes of the various levels. In our previous example we have:
<table>
<tbody>
<tr>
<td></td>
<td>g3=3subscript𝑔33\displaystyle g_{3}=3</td>
<td></td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td>g2=41subscript𝑔241\displaystyle g_{2}=41</td>
<td></td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td>g1=5subscript𝑔15\displaystyle g_{1}=5</td>
<td></td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td>g0=12subscript𝑔012\displaystyle g_{0}=12</td>
<td></td>
</tr>
</tbody>
</table>
where each $`g_{i}`$ corresponds to the gnode ID of the level $`i`$. Note that $`g_{0}`$ is the ID attributed to the single node, at level 0.
### 4.2 Fractal map
The advantages of using a fractal topology are clear.
The node $`N`$, instead of memorising information about each node of the whole network, will keep only that regarding the gnodes where it belongs to. Suppose the node $`N`$ had this ID:
<table>
<tbody>
<tr>
<td></td>
<td>g3⋅g2⋅g1⋅g0⋅subscript𝑔3subscript𝑔2subscript𝑔1subscript𝑔0g_{3}\cdot g_{2}\cdot g_{1}\cdot g_{0}</td>
<td></td>
</tr>
</tbody>
</table>
It will store in memory information regarding:
1. 1\.
the 256 single nodes which belongs to its same gnode of level 1, or in other words, the 256 nodes of the gnode $`g_{1}`$,
2. 2\.
the 256 gnodes gnodes which belongs to its same gnode of level 2, of in other words, the 256 gnodes of the gnode $`g_{2}`$,
3. 3\.
finally, the 256 gnodes which belongs to the gnode $`g_{3}`$.
Note that doing so, the node $`N`$ will be blind to all the other gnodes. For example, it won’t know any information regarding the single nodes which belong to all the other gnodes of level 1 different from $`g_{1}`$.
Even with this lack of knowledge, as we’ll see later, the node $`N`$ is still able to reach all the other nodes of the network. In conclusion, $`N`$ only needs $`256\hspace{0pt}n`$ entries in its map, instead of $`2^{32}`$. To clarify the ideas suppose that each entry costs one byte. In the plain topology we needed $`4\hspace{0pt}G\hspace{0pt}b`$, while in the fractal one we just need $`{{256 \cdot 4}\hspace{0pt}b} = {1\hspace{0pt}K\hspace{0pt}b}`$.
#### 4.2.1 IP v4 and v6
Netsukuku is both compatible with ipv4 and ipv6.
In ipv4 there are a maximum of $`2^{32}`$ IPs, thus we have five levels $`n = 4`$.
In ipv6 there are a maximum of $`2^{128}`$ IPs, thus $`n = 16`$.
#### 4.2.2 Internal and external map
For simplicity we divide the map of the node $`N`$, in the *internal map* and in the *external* one. The internal map contains information regarding the nodes belonging to $`g_{1}`$. The external map describes all the other levels of the topology.
#### 4.2.3 Bnode map
The bnode map of the node $`N`$ contains the information regarding the bnodes of each level where $`N`$ belongs. Some examples to clarify the ideas:
suppose that $`N = {g_{3} \cdot g_{2} \cdot g_{1} \cdot g_{0}}`$
- •
a bnode of level 0 is a single node linked with two nodes of two different gnodes of level 1.
- •
the bnodes of level 0, known by $`N`$, are only that which belong to the gnode $`g_{0}`$. They are all the nodes of $`g_{0}`$ which are linked to at least a gnode different from $`g_{1}`$.
### 4.3 CIDR routing
The QSPN, for each level, will build the routes necessary to connect each (g)node to all the other (g)nodes of the same level. The routes will be saved in the maps of each node.
If the node $`N = {g_{3} \cdot g_{2} \cdot g_{1} \cdot g_{0}}`$ wants to reach a node $`M`$ which belongs to different gnodes, f.e. $`M = {g_{3} \cdot g_{2} \cdot h_{1} \cdot h_{0}}`$, it will add a CIDR\[[3](#bib.bib3)\] route in the routing table of the kernel:
*all the packets whose destination is $`{g_{3} \cdot g_{2} \cdot h_{1} \cdot 0}\hspace{0pt}\ldots\hspace{0pt}255`$ will be forwarded to the gateway $`X`$*.
We’ll see later how the gateway $`X`$ is chosen.
## 5 Tracer Packets in high levels
In the QSPN document \[[1](#bib.bib1)\], we’ve seen how a Tracer Packet works in a network composed by single nodes, i.e. a gnode of level 0.
We’ll now study its way of working on higher levels.
### 5.1 A gnode is a node
In the abstract sense a single node is an entity which:
1. 1\.
receives input from its links
2. 2\.
stores it in its memory
3. 3\.
computes it
4. 4\.
and sends the output of the computation over some of its links
Thus any other entity which performs the same operations can be thought as a single node.
A gnode $`G`$ can act as a single node too.
1. 1\.
A bnode $`I`$, which belongs to $`G`$, receives an input from its links. We call $`I`$ the *ingress* (b)node.
2. 2\.
this input is flooded to all the nodes, of any level, of the gnode. The nodes will memorize the information contained in the input.
Note that the flood is not a TP. The flooded pkt will be received only once by each node.
3. 3\.
A bnode $`O`$, of the same gnode, which is different from $`I`$, receives the flooded input and computes it. We call $`O`$ the *egress* (b)node.
4. 4\.
The bnode $`O`$ sends the output of its computation to its external links, i.e. avoiding those links which connect it to nodes of $`G`$ or to the same gnode which sent the input to $`I`$.
#### 5.1.1 Example of a wandering TP
Consider the network in figure [4](#S5.F4 "Figure 4 ‣ 5.1.1 Example of a wandering TP ‣ 5.1 A gnode is a node ‣ 5 Tracer Packets in high levels ‣ Netsukuku topology").
[Refer to caption](/html/0705.0819/assets/fig/qspn_g3.png)
Figure 4: The gnodes $`G_{1}`$, $`G_{2}`$ and $`G_{3}`$
$`G_{1}`$, $`G_{2}`$, $`G_{3}`$ are gnodes of level 1. $`A_{1}`$ and $`A_{2}`$ belong to $`G_{1}`$. $`B_{1}`$ and $`B_{2}`$ belong to $`G_{2}`$. $`C_{1}`$ belongs to $`G_{3}`$.
Suppose the node $`A_{1}`$ sends a Tracer Packet in level 1
<sup>2</sup>
<sup>2</sup>2Note that any node can send a TP in any level. In this case, we are considering $`A_{1}`$, which is a bnode, to simplify the example
. The following will happen:
1. 1\.
The TP is flooded in $`G_{1}`$.
2. 2\.
$`A_{2}`$ receives the TP and appends in it the ID of the gnode $`G_{1}`$.
3. 3\.
$`A_{2}`$ sends the TP to $`B_{1}`$.
4. 4\.
$`B_{1}`$ receives it, updates its maps and floods it in $`G_{2}`$.
5. 5\.
All the node belonging to $`G_{2}`$ will receive the TP, updating their maps.
6. 6\.
This same procedure is reiterated from step 2, i.e. $`B_{2}`$ receives the TP, appends the ID of $`G_{2}`$ and so on until $`C_{1}`$ receives it.
7. 7\.
$`C_{1}`$, noticing that its gnode hasn’t any links to other gnodes than $`G_{2}`$, will bounce back the TP to $`B_{2}`$ and at the same time will flood $`G_{3}`$.
8. 8\.
The TP, with the same procedure, will return back to $`A_{1}`$, completing the TP cycle.
## 6 QSPN v2 in high levels
In order to use the $`Q^{2}`$ (QSPN v2) in high levels, we need to be sure that a TP, flooded inside a lower gnode, will reach once and only once all the nodes of the same gnode. Moreover, a good metric needs to be defined for the high levels: what is the rtt (Round-Trip Time) and bandwidth capacity of the link between two gnodes, how is it measured?
### 6.1 Endless loops
Consider the situation in figure [5](#S6.F5 "Figure 5 ‣ 6.1 Endless loops ‣ 6 QSPN v2 in high levels ‣ Netsukuku topology").
[Refer to caption](/html/0705.0819/assets/fig/3bnodes.png)
Figure 5: Three bnodes, forming a cycle
A and B are bnodes of the node $`G_{1}`$, while $`C`$ is a bnode of the gnode $`G_{2}`$. Suppose C sends a TP to B. B floods it inside $`G_{1}`$, thus $`A`$ receives it. At this point, $`A`$ sends the packet to all its external links, and thus to $`C`$. $`C`$ will send the packet again to $`B`$ and the cycle will continue.
This situation can be avoided if an ingress bnode, before forwarding a packet to a an external gnode $`F`$, checks if the packet has already traversed $`F`$ itself. If it has, the bnode won’t forward it to $`F`$.
### 6.2 Route Efficiency Measure
We will refer to REM (Route Efficiency Measure), as a value characterising the quality of a route. REM can be calculated in various ways, f.e. by taking in account the total rtt and the bw capacity of the route. We denote the REM of a route $`r`$ as $`R\hspace{0pt}E\hspace{0pt}M\hspace{0pt}{(r)}`$.
### 6.3 Flat levels
From the point of view of the QSPN v2, the levels are “flattened”, because the propagation of CTPs
<sup>3</sup>
<sup>3</sup>3Continuous Tracer Packet, see \[[1](#bib.bib1)\]
ignore the subdivisions of the network in gnodes and levels. The structure of the network remains unchanged, subdvided fractally in gnodes, but the exploration of the CTP doesn’t. A CTP doesn’t consider a gnode as a single, homogenous entity. It propagates itself between the single nodes, tracing a precise route formed by single hops, until it is considered interesting. In this way every node will have its own personal route to reach a specific gnode.
The memory limits are respected, because the CTP, once it jumped over the border of a gnode, deletes the specific details that were useful by the nodes of the gnode.
We’ll know describe the rules of the Flat levels.
Suppose we want to explore the level $`n \geq 1`$ (we already know how to explore $`n = 0`$). Let $`G_{1},\ldots,G_{r}`$ be the gnode of level $`n`$. The generic rule of the QSPN v2 in high levels is:
> When an ingress bnode $`b`$ of $`G`$ receives a CTP from a neighbouring bnode of $`H`$, it appends in the packet its IP. The packet, called *Locked Tracer Packet* is then flooded inside $`G`$. The LTP of level $`n`$, is a packet that is propagated with the same rules of the CTP, i.e if it isn’t interesting it’s dropped, but that isn’t updated while it explores the levels inferior to $`n`$. When the egress bnode $`b'`$, of $`G`$ receives the LTP, it changes it in a CTP, appends the ID of $`G`$ and forwards it to all its neighbouring gnodes.
Detailed rules:
1. 1\.
When exploring the level $`n`$, each inferior level must been already explored.
To accomplish this, it sufficies that a node, before sending/forwarding a CTP of level $`n`$ has already sent/forwarded at least a CTP of level $`n - 1`$. If it hasn’t, it will queue the CTP of level $`n`$.
2. 2\.
Suppose that the ingress bnode $`b`$ belonging to a gnode $`G`$ of level $`n`$, has received a CTP of the same level, sent by the gnode $`G'`$. The following will happen: $`b`$ makes copy of the CTP, marking it as a LTP, appends in it its IP and floods it in $`G`$. If $`b`$ is also linked to a gnode different from $`G'`$, it will respect the rule [3](#S6.I1.i3 "item 3 ‣ 6.3 Flat levels ‣ 6 QSPN v2 in high levels ‣ Netsukuku topology") for egress bnodes.
3. 3\.
Suppose that the egress bnode $`b'`$ of $`G`$ receives the LTP sent by $`b`$ (see above). $`b'`$ transforms the LTP in a CTP, deletes the IP added by $`b`$ and sets (in the TP) the REM of $`G\rightarrow G'`$ as:
<table>
<tbody>
<tr>
<td></td>
<td>REM(G→G′):=REM(b′→b)assign𝑅𝐸𝑀→𝐺superscript𝐺′𝑅𝐸𝑀→superscript𝑏′𝑏REM(G\rightarrow G^{\prime}):=REM(b^{\prime}\rightarrow b)</td>
<td></td>
</tr>
</tbody>
</table>
Finally, it forwards the CTP to all its external rnodes.
4. 4\.
A node $`n`$ of $`G`$, uses the received LTP to learn that $`b`$ is a bnode bordering on $`G'`$ and to find out the route of level $`n`$ contained in the packet. It considers the REM of the route to reach $`G'`$ equal to
<table>
<tbody>
<tr>
<td></td>
<td>REM(n→b)𝑅𝐸𝑀→𝑛𝑏REM(n\rightarrow b)</td>
<td></td>
</tr>
</tbody>
</table>
in other words, it saves in its map that $`{R\hspace{0pt}E\hspace{0pt}M\hspace{0pt}{({n\rightarrow G})}}:={R\hspace{0pt}E\hspace{0pt}M\hspace{0pt}{({n\rightarrow b})}}`$.
$`n`$ forwards the LTP to its rnodes without modifying it.
5. 5\.
The LTP respects the same rule of the QSPN v2: it is forwarded by a node only if it is carries interesting information, otherwise it is dropped. A node $`n`$ consider the LTP interesting if one of the following conditions
<sup>4</sup>
<sup>4</sup>4this isn’t the complete list of conditions
is met:
- •
The LTP contains the IP of a bnode $`b`$, which wasn’t already known by $`n`$.
- •
The LTP contains a bnode already known by $`n`$, but that borders on a new gnode $`g`$: with this LTP, the node $`n`$ learns that it can reach $`g`$ by passing through the bnode.
- •
The LTP contains the information of the death of a bnode which was used by $`n`$ to reach some gnodes.
- •
The LTP contains a new gnode, which was previously unknown.
- •
The LTP contains the information of the death of a gnode.
- •
The LTP contains an improved route to reach a gnode.
Note: the first three conditions are used by the nodes to build and update the bnode map.
6. 6\.
Recall the situation of rule [3](#S6.I1.i3 "item 3 ‣ 6.3 Flat levels ‣ 6 QSPN v2 in high levels ‣ Netsukuku topology").
When the value $`R\hspace{0pt}E\hspace{0pt}M\hspace{0pt}{({b'\rightarrow b})}`$ changes considerably, both $`b`$ and $`b'`$ sends a CTP of level $`n`$ to their external nodes
<sup>5</sup>
<sup>5</sup>5A CTP of level $`n - 1`$ tells $`b`$ and $`b'`$ of the REM change
. The CTP contains the new $`R\hspace{0pt}E\hspace{0pt}M`$ value. Every node which used $`G`$ as hop of some routes, will consider the CTP interesting.
This rule permits the update of high levels.
7. 7\.
When a bnode $`b`$ (of any level)
<sup>6</sup>
<sup>6</sup>6a bnode $`b^{\operatorname{\prime\prime}}`$, which is a gnode of level $`l \geq 1`$, loses its external link to the gnode $`H`$ of level $`l + 1`$, when all its bnodes of level $`l - 1`$ loses their link to $`H`$. This is a recursive definition.
loses one of its external links of level $`n`$, a CTP is sent in the level $`n - 1`$. The CTP contains the information regarding the lost link. The CTP is forwarded by all the (g)nodes that used the bnode $`b`$ in one of their saved routes to reach the lost link.
If $`b`$ was the unique bnode of its gnode $`G`$ bordering to gnode $`G'`$ of level $`n`$, then the link $`G\leftrightarrow G'`$ becomes broken. In this case a CTP of level $`n`$ is sent by $`G'`$ and by the other bnodes of $`G`$
<sup>7</sup>
<sup>7</sup>7they will know of the broken link when receiving the CTP of level $`n - 1`$
.
A dying bnode is equivalent to a bnode which loses all its external links. The only difference is that a CTP of level 0 is sent to inform the nodes of its gnode.
We’ll now describe some rules by giving an example (see figure [6](#S6.F6 "Figure 6 ‣ 6.3 Flat levels ‣ 6 QSPN v2 in high levels ‣ Netsukuku topology")). For simplicity, we’ll use the IP notation: $`11.22.33`$ indicates the single node $`33`$ belonging to the gnode $`22`$, belonging to the ggnode $`11`$. Using the symbolic notation: $`11.22.33 \in 11.22 \in 11`$. We are also assuming that the routes are symmetric.
[Refer to caption](/html/0705.0819/assets/fig/segG3.png)
Figure 6: Level 3
*CASE 1: A new gnode join.*
We are exploring the level 3, which is formed by the gnodes $`11,\;22,\;33,\;44`$. The level 2, of each gnode, has been already explored.
- •
The gnode $`44`$ has just created the link to the gnode $`33`$. The new physical link is maintained by the two single bnodes $`33.33.33,\;44.44.44`$.
- •
The bnode $`33.33.33`$ creates a CTP of level 3, appends in it the ID “$`33`$” and sends it to $`44.44.44`$.
- •
The bnode $`44.44.44`$ receives it.
- –
It reflects back a new CTP to $`33.33.33`$, appending in it the ID “$`44`$”.
- –
It appends in the received CTP, its IP, marks it as a LTP and sends it to its gnode. The LTP contains the path $`33\rightarrow 44.44.44`$.
- \*
The node $`44.44.x`$, receiving the LTP, learns the following: $`44.44`$ is a bnode of level 2 that can be used to reach $`33`$, and $`44.44.44`$ is a bnode of its own gnode which borders on $`33`$. It also learns its own REM for the route to $`33`$:
<table>
<tbody>
<tr>
<td></td>
<td>REM(44.44.x→33):=REM(44.44.x→44.44.44)REM(44.44.x\rightarrow 33):=REM(44.44.x\rightarrow 44.44.44)</td>
<td></td>
</tr>
</tbody>
</table>
$`44.44.x`$ continues to forward the LTP (without modifying it).
- \*
The LTP reaches the node $`44.y.z`$, node of $`44.y`$. $`44.y.z`$, from the LTP learns that: $`44.44`$ is a bnode of level 2 that can be used to reach $`33`$. Since the level $`2`$ has been already explored, $`44.y.z`$ already knows a route to reach $`44.44`$.
- \*
The LTP having reached all the nodes of $`44`$ ceases to being forwarded.
- •
The bnode $`33.33.33`$ receives the reflected CTP sent by $`44.44.44`$.
- –
It appends in it its IP, marks it as a LTP and sends it to its gnode.
- \*
The LTP is propagated in $`33`$ (in the same way of the previous LTP).
- •
The bnode $`33.r.s`$, which borders on $`22`$, receives the LTP.
- –
It modifies it in a CTP: it removes the IP $`33.33.33`$ and appends the ID $`33`$. The CTP is now $`44\rightarrow 33`$. Finally, it sets the first REM of the CTP:
<table>
<tbody>
<tr>
<td></td>
<td>REMCTP(33↔︎44):=REM(33.r.s↔︎33.33) see note REM_{CTP}(33\leftrightarrow 44):=REM(33.r.s\leftrightarrow 33.33)\;\;\;\textrm{ see note }</td>
<td></td>
</tr>
</tbody>
</table>
The CTP is then sent to $`22.22.22`$, bnode of $`22`$.
- •
The bnode $`22.22.22`$ receives the CTP. It sends a LTP in $`22`$.
- •
The bnode $`22.t.u`$, which borders on $`11`$, receives the LTP. It appends $`22`$ in it and sends it to $`11`$. The CTP is now: $`44\rightarrow 33\rightarrow 22`$.
- •
A bnode of $`11`$ receives it, and reflect it to $`22`$. It sends also the LTP to $`11`$.
- •
…
- •
The reflected CTP is finally received by the node $`44.44.44`$, which sends the LTP to $`44`$ and reflects the CTP to $`33.33.33`$. The bnode $`33.33.33`$, drops immediately the CTP, because it doesn’t contain any interesting information.
- •
the exploration is completed.
### 6.4 Unique flood
Suppose that a TP has been flooded inside the gnode $`G`$. Suppose also that the node $`n \in G`$ receives two duplicate packets.
The $`Q^{2}`$ instructs the node $`n`$ to keep forwarding only the interesting packets. A packet, which is a perfect copy of a packet already received, is always uninteresting. Therefore, the node $`n`$ will drop the second copy of the received TP.
However, in this case the node $`n`$ should not consider the REM values saved in the TP, because two TPs, which have crossed the same gnodes in the same order, might have different REM values.
In conclusion, whenever two or more TPs, which have crossed the same high level route, are received by a node of $`G`$, only one of them will be forwarded. In this way, the successive nodes will receive only one copy of the same packet.
## 7 Network dynamics
When a part of the network changes considerably, the maps of the involved levels must be updated.
### 7.1 Radar
Every node has its own radar, which periodically sends a broadcast request to all its (physical) near nodes. By collecting the replies, the radar is able to determine the active rnodes of the node and the quality of its links.
### 7.2 Level 0
In level 0, a CTP is sent every time a node joins the network, dies or every time the change of the quality of a link exceeds a predefined delta value. The CTP is also is restricted to the gnode where it has been originated: a bnode won’t send it to external nodes.
When a node joins, it won’t send a CTP, only its rnodes will. Their CTP will be directed to the node. More formally:
if the node $`n`$ joins, that is if the node $`n`$ creates a link with the nodes $`r_{1},\ldots,r_{n}`$, with $`n \geq 1`$, then a CTP will be sent by each rnode $`r_{i}`$ *only* to $`n`$. This means that if $`r_{j}`$ is connected to a node $`s`$, when $`n`$ joins, $`r_{j}`$ will send the CTP only to $`n`$ and not to $`s`$.
This saves the propagation of two CTP in the following situation:
<table>
<tbody>
<tr>
<td></td>
<td>…↔︎A↔︎B↔︎C↔︎…↔︎…𝐴↔︎𝐵↔︎𝐶↔︎…\dots\leftrightarrow A\leftrightarrow B\leftrightarrow C\leftrightarrow\dots</td>
<td></td>
</tr>
</tbody>
</table>
suppose that the node $`B`$ joins. As a consequence of the rule we’ve described, the node $`C`$ will send a CTP to $`B`$ and $`A`$ to $`B`$, thus only two distinct CTP are generated. Their path is $`A\rightarrow B\rightarrow C\rightarrow\ldots`$ and the reverse. If instead the node $`B`$ sends a CTP to $`C`$ and then $`A`$ to $`B`$, two distinct CTP will explore the same verse: $`B\rightarrow C\rightarrow\ldots`$ and $`A\rightarrow B\rightarrow C\rightarrow\ldots`$, the same is for the reverse, thus, in total, four distinct CTP are propagated.
Let $`A\overset{l}{\leftrightarrow}B`$ be a link. If its quality changes $`B`$ will send a CTP to $`A`$ and vice-versa.
When a node dies, all its rnodes will send a CTP to all their rnodes. The CTP will include, as the first hop, the ID of the dead node, with a flag which indicates its death. For example:
<table>
<tbody>
<tr>
<td></td>
<td>A↔︎B↔︎C↔︎…↔︎𝐴𝐵↔︎𝐶↔︎…A\leftrightarrow B\leftrightarrow C\leftrightarrow\dots</td>
<td></td>
</tr>
</tbody>
</table>
if $`A`$ dies, then $`B`$ will send the following CTP to $`C`$: $`\overset{+}{A}\rightarrow B`$
In order to prevent false positives, the nodes won’t immediately send the CTP, but will wait a small amount of time. Only if the change persists, they will send it.
### 7.3 Level n
The dynamics for the update of high levels are mainly governed by rule [6](#S6.I1.i6 "item 6 ‣ 6.3 Flat levels ‣ 6 QSPN v2 in high levels ‣ Netsukuku topology") and [7](#S6.I1.i7 "item 7 ‣ 6.3 Flat levels ‣ 6 QSPN v2 in high levels ‣ Netsukuku topology") of flat levels (see [6.3](#S6.SS3 "6.3 Flat levels ‣ 6 QSPN v2 in high levels ‣ Netsukuku topology")). In this paragraph, we’ll specify some details.
Rule [6](#S6.I1.i6 "item 6 ‣ 6.3 Flat levels ‣ 6 QSPN v2 in high levels ‣ Netsukuku topology") says: “when the value $`R\hspace{0pt}E\hspace{0pt}M\hspace{0pt}{({b'\rightarrow b})}`$ changes considerably, a new CTP is sent”. By considerably we mean that the difference of the current REM value with the previous one exceedes a predefined threshold. More precisely: let $`A\rightarrow B`$ be a route of level $`n`$. If
<table>
<tbody>
<tr>
<td></td>
<td>|REMnow(A→B)−REMprev(A→B)|>Δn𝑅𝐸subscript𝑀now→𝐴𝐵𝑅𝐸subscript𝑀𝑝𝑟𝑒𝑣→𝐴𝐵subscriptΔ𝑛|REM_{\textrm{now}}(A\rightarrow B)-REM_{prev}(A\rightarrow B)|>\Delta_{n}</td>
<td></td>
</tr>
</tbody>
</table>
then the route has changed considerably. $`\Delta_{n}`$ is proportional to $`n`$, in this way the nodes will be more sensible to minor changes of lower levels and they will consider only relevant changes of higher levels.
Also in high levels, a bnode won’t immediately send a new CTP, but it will delay it for a small amount of time.
### 7.4 Hooking phase
A new node joins the network when it has been able to create at least one physical link to an active Netsukuku node and when it has correctly executed the hooking procedure. In this paragraph, we describe loosely the hooking phase of a new node.
Suppose that the node $`n`$ has established a physical link to at least one Netsukuku node. In order to become an active Netsukuku node, $`n`$ has to *hook* to its rnodes.
During the hook, $`n`$ will exchange vital information with its rnodes, it will choose its new IP and it will finally become part of a gnode.
The hook procedure is formed by these general steps:
1. 1\.
The node $`n`$ chooses an IP in the range of $`10.0.0.1 \leq {I\hspace{0pt}P} \leq 10.0.0.255`$.
2. 2\.
It launches the first radar to see what its rnodes are. If not a single node is found, it creates a new gnode and ends hooking phase.
3. 3\.
At this point, $`n`$ asks to its nearer rnode the list of all the available free nodes presents inside the gnode of the rnode.
If the rnode rejects the request (the gnode might be full), the node $`n`$ contacts another rnode.
4. 4\.
$`n`$ chooses an IP from the received list of free nodes and sets it on its network interface.
5. 5\.
$`n`$ will then download the external map from the same rnode. Looking at the external map, it will be able to determine if it has to create a new gnode. If it has, it creates it and ends the hooking.
6. 6\.
$`n`$ gets the internal and the bnode map from the same rnode.
7. 7\.
$`n`$ launches a second radar and updates its routing table.
8. 8\.
All the rnodes of $`n`$ send a CTP to update the maps.
### 7.5 Gnode hook
When a node creates a new gnode, it will choose a random gnode ID, and thus a random ip.
Suppose that two isolated gnodes get the same gnode ID. When they will be linked, they’ll enter in conflict.
The solution to this problem is to let each new gnode hook as a normal node would. You can find more information about this in the NTK_RFC 001\[[4](#bib.bib4)\].
## 8 ChangeLog
- •
`March 2007`
- –
Description of the Flat levels (sec. [6.3](#S6.SS3 "6.3 Flat levels ‣ 6 QSPN v2 in high levels ‣ Netsukuku topology"))
- –
Section [7.2](#S7.SS2 "7.2 Level 0 ‣ 7 Network dynamics ‣ Netsukuku topology") “Network dynamics - Level 0” expanded.
- –
Section [7.3](#S7.SS3 "7.3 Level n ‣ 7 Network dynamics ‣ Netsukuku topology") “Network dynamics - Level n” updated: the references to the pre-Flatlevels REM metric have been removed.
- •
`October 2006`
Initial release.
## References
- \[1\] QSPN document: http://netsukuku.freaknet.org/doc/main_doc/qspn.pdf
- \[2\] Netsukuku website: http://netsukuku.freaknet.org/
- \[3\] CIDR routing: http://en.wikipedia.org/wiki/Classless_Inter-Domain_Routing
- \[4\] NTK_RFC 001: http://lab.dyne.org/Ntk_gnodes_contiguity
`^_^` |
0705.2819 | cs.NI | cyber | # An Autonomous Distributed Admission Control Scheme for IEEE 802.11 DCF
Preetam Patil
Varsha Apte
Department of CSE Indian Institute of Technology - Bombay, India <pvpatil@iitb.ac.in> Department of CSE Indian Institute of Technology - Bombay, India <varsha@cse.iitb.ac.in>
###### Abstract
Admission control as a mechanism for providing QoS requires an accurate description of the requested flow as well as already admitted flows. Since 802.11 WLAN capacity is shared between flows belonging to all stations, admission control requires knowledge of all flows in the WLAN. Further, estimation of the load-dependent WLAN capacity through analytical model requires inputs about channel data rate, payload size and the number of stations. These factors combined point to a centralized admission control whereas for 802.11 DCF it is ideally performed in a distributed manner. The use of measurements from the channel avoids explicit inputs about the state of the channel described above. BUFFET, a model based measurement-assisted distributed admission control scheme for DCF proposed in this paper relies on measurements to derive model inputs and predict WLAN saturation, thereby maintaining average delay within acceptable limits. Being measurement based, it adapts to a combination of data rates and payload sizes, making it completely autonomous and distributed. Performance analysis using OPNET simulations suggests that BUFFET is able to ensure average delay under 7$`m\hspace{0pt}s`$ at a near-optimal throughput.
###### category:
C.2.1 Computer-Communication Networks Network Architecture and Design
###### keywords:
Admission Control, Measurements, Wireless LANs, Analytical Models, Simulations
<sup>†</sup>
<sup>†</sup>conference:
<sup>†</sup>
<sup>†</sup>terms: Performance
## 1 Introduction
With the widespread use of WLANs based on IEEE 802.11 distributed coordination function (DCF), efforts are on to improve the Quality of Service (QoS) offered by WLANs. The most important component of QoS is the delay experienced by packets. While real-time flows have strict requirements on delays, all applications remain sensitive to high and variable delays.
The proposed QoS-oriented 802.11e standard provides prioritized access through the Enhanced Distributed Channel Access (EDCA), but 802.11e devices are not widely available. Moreover, the QoS provision of 802.11e EDCA depends on appropriate configuration of the tunable parameters and admission control, otherwise its performance degrades to that of DCF. On the other hand, it has been argued \[[13](#bib.bib13)\] that DCF is capable of providing acceptable delays as long as the load on WLAN is maintained within the capacity of the WLAN.
Provision of QoS in communication systems necessarily involves maintaining the load within the finite capacity of the system. This task is performed by the call
<sup>1</sup>
<sup>1</sup>1the term ‘call’ is used synonymously with ‘flow’ in this context.
admission control (CAC) mechanism based on a selected criterion. Admission control can be performed on a per-flow or per-host basis, either before admitting an entity, or, in some cases, even after admission, if it is clear that the desired (or guaranteed) QoS requirements can not be met.
Various models proposed for DCF and EDCA help predict the achievable throughput and delay \[[4](#bib.bib4), [7](#bib.bib7)\]. However, the application of these models for admission control requires an exact description of the traffic parameters such as packet arrival rate, average packet size, as well as WLAN parameters. Our previous experience \[[10](#bib.bib10)\] suggests that describing the packet stream at the link layer is difficult due to the diverse application characteristics as well as control overheads of the intermediate layers.
In addition, since the WLAN capacity is shared between all stations, the capacity computation requires the statistics of all flows in the WLAN that in turn lends itself to a centralized mode of admission control.
In order to preserve the advantages of the distributed operation of DCF, our endeavor is to design an autonomous, distributed admission control that requires minimal inputs and is able to deduce current state of the WLAN from relevant measurements. The use of measurements from the channel or the WLAN interface will serve two purposes:
- $`\bullet`$
help the station estimate the aggregate statistics for its admitted flows that are difficult to characterize.
- $`\bullet`$
help the station deduce traffic statistics for other stations’ flows without using any message passing.
In this paper, we present one such distributed admission control scheme named BUFFET. In the next section we present a summary of related work that motivates the need for current work. Section [3](#S3 "3 Admission Control in 802.11 WLANs ‣ An Autonomous Distributed Admission Control Scheme for IEEE 802.11 DCF") presents the admission control problem and the analytical framework for the solution. Section [4](#S4 "4 Model based Distributed Admission Control (BUFFET) ‣ An Autonomous Distributed Admission Control Scheme for IEEE 802.11 DCF") describes the algorithm in detail as well as the two competing approaches we use for performance comparison. Performance analysis of BUFFET and other approaches is presented in Section [5](#S5 "5 Performance Analysis ‣ An Autonomous Distributed Admission Control Scheme for IEEE 802.11 DCF"). We conclude with a discussion of performance results and future work in Section [6](#S6 "6 Conclusion ‣ An Autonomous Distributed Admission Control Scheme for IEEE 802.11 DCF").
## 2 Related Work
Bianchi and Tinnirello \[[5](#bib.bib5)\] use the collision probability $`p`$ derived from the measured transmission probability to estimate the number $`n`$ of competing stations. Pong and Moors \[[11](#bib.bib11)\] propose a call admission control for EDCA based on the saturation model in \[[4](#bib.bib4)\]. Per-flow measurement of the collision probability $`p`$ is used to estimate the throughput achievable by the flow. A limitation of saturation model based CAC is that the model exaggerates the effect of contention, especially at higher $`n`$.
The centralized CAC for EDCA proposed by Kong et al. \[[9](#bib.bib9)\] uses the measured channel utilization to estimate the achievable bandwidth (fair share of the throughput) for the requested flow based on a non-saturation model. The CAC for EDCA proposed by Bai et al. \[[2](#bib.bib2)\] attempts to keep the queue utilization ($`\rho`$) below a threshold. $`\rho`$ is computed using regression analysis and an analytical model from the measured value of $`\tau`$ (the transmission probability by a station in a slot) and load specification.
An important shortcoming of the CAC mechanisms listed above is that they require exact specification of packet arrival rates (except saturation model based CACs) and payload size *for all flows*. It implies a centralized CAC mechanism that uses and stores this information for admission decision.
It is possible that a flow obtains more than its fair share of bandwidth (WLAN throughput/$`n`$) without violating the QoS of other flows as long as the WLAN is not saturated. A uniform throughput sharing assumption \[[11](#bib.bib11), [9](#bib.bib9)\] results in rejecting such flows, even if there is spare capacity.
Channel utilization (fraction of channel time used by transmissions) threshold based CAC has been explored in \[[6](#bib.bib6), [12](#bib.bib12), [8](#bib.bib8), [3](#bib.bib3)\]. The CAC proposed by Chou et al. \[[6](#bib.bib6)\] maintains the allocated airtime below a threshold, but the airtime computation excludes the overheads of contention mechanism. Admission control for DCF proposed in \[[12](#bib.bib12)\] combines channel utilization (including the requested flow) threshold based CAC for real-time traffic and rate control for best-effort traffic. The CAC scheme for EDCA in \[[8](#bib.bib8)\] uses the measured utilization to decide on accepting a new flow or stopping low-priority flows. The work in \[[3](#bib.bib3)\] evaluates two threshold-based schemes for infrastructure WLANs, based on channel utilization and access point queue size respectively.
Performance of threshold based CAC schemes is contingent on the selection of the correct threshold especially in realistic WLAN scenarios because the optimum value of the threshold depends on payload, channel data rate, and number of stations.
## 3 Admission Control in 802.11 WLANs
The link capacity of a 802.11 WLAN varies depending on traffic parameters such as the number of stations, packet size, and channel data rate \[[13](#bib.bib13)\]. The admission control for WLANs is further complicated by the requirement of predicting the capacity or the delay at the 802.11 MAC. The admission control objective in this paper is to keep the average delay for all flows within acceptable limits. Thus the admission control provides statistical QoS guarantees.
### 3.1 Requirements and desired properties of distributed CAC
We start with the following design objectives for the distributed CAC mechanism:
- $`\bullet`$
The algorithm is to run at every wireless station without requiring any centralized control and coordination.
- $`\bullet`$
No knowledge of global parameters would be assumed; measurements are made locally at the WLAN interface.
- $`\bullet`$
The measurements or the algorithm will not necessitate any change to the 802.11 protocol.
The following are the desired properties of a measurement-based admission control algorithm:
- $`\bullet`$
The algorithm should be responsive to changing load and number of stations.
- $`\bullet`$
It should adapt to varying data rates selected by stations based on channel quality.
- $`\bullet`$
It should not depend on accurate statistical characteristics of all flows.
- $`\bullet`$
It should be scalable with respect to the number of flows as well as stations.
### 3.2 Using analytical model of 802.11 MAC to predict saturation
The delay experienced by a packet is the queueing delay at WLAN interface plus the time to transmit the packet (including contention and collisions, if any). This queue can be modeled as an $`M/G/1`$ queue assuming Poisson arrival process. The service rate of the queue is however dependent on the arrival rate $`\lambda`$. As illustrated by the delay vs. load curve in Fig. [1](#S3.F1 "Figure 1 ‣ 3.2 Using analytical model of 802.11 MAC to predict saturation ‣ 3 Admission Control in 802.11 WLANs ‣ An Autonomous Distributed Admission Control Scheme for IEEE 802.11 DCF"), the delay is close to nominal packet transmission time at moderate loads whereas it increases by an order of magnitude after the WLAN saturates (e.g., at 29% load in Fig. [1](#S3.F1 "Figure 1 ‣ 3.2 Using analytical model of 802.11 MAC to predict saturation ‣ 3 Admission Control in 802.11 WLANs ‣ An Autonomous Distributed Admission Control Scheme for IEEE 802.11 DCF")). Thus prevention of WLAN saturation has the desired effect of maintaining average delay within acceptable limits.
[Refer to caption](/html/0705.2819/assets/x1.png)
Figure 1: $`\gamma`$ and delay co-relation
Avoiding saturation requires predicting it in real time given the state of current load, requested load, and WLAN parameters. For this purpose, we use a Discrete Time Markov Chain (DTMC) based analytical model we have developed. It is an extended version of the model in \[[4](#bib.bib4)\] and is applicable to saturated as well as non-saturated WLANs. We add a state called *idle* to the single-station DTMC in \[[4](#bib.bib4)\]. A station, after completing a post-backoff (mandatory backoff after a successful transmission), examines the state of the interface queue. We define a probability $`\gamma`$ as the probability of queue being empty with which the station enters *idle* state after post-backoff. The estimated $`\gamma`$ after accounting for an incoming flow can be used as an indicator of saturation (as illustrated in Fig. [1](#S3.F1 "Figure 1 ‣ 3.2 Using analytical model of 802.11 MAC to predict saturation ‣ 3 Admission Control in 802.11 WLANs ‣ An Autonomous Distributed Admission Control Scheme for IEEE 802.11 DCF")). $`\gamma`$ can be obtained from the model as a function $`\Psi\hspace{0pt}{(\lambda,n,T_{s})}`$ (equation ([6](#A1.E6 "In A.2 Calculation of 𝛾 ‣ Appendix A DTMC model for non-saturated 802.11 DCF ‣ An Autonomous Distributed Admission Control Scheme for IEEE 802.11 DCF"))) of three variables $`\lambda`$ (packet arrival rate), $`n`$ and $`T_{s}`$ (duration of successful transmission). The details of the DTMC and computation of $`\gamma`$ are presented in the Appendix. Based on this argument, the CAC algorithm follows.
## 4 Model based Distributed Admission Control (BUFFET)
In a distributed CAC scheme, a station may not be aware of the flows belonging to other stations, neither will it know the instantaneous data rates adopted by individual stations. However, the station is capable of listening to all transmissions and derive inference about the current load on the WLAN as well as WLAN parameters. The load/population dependent inputs to the model are generated by combining the measurements with the flow specifications provided by the application for the requested flow. The expected value of $`\gamma`$ computed using the model is used to decide on accepting/rejecting the flow. The point of departure for BUFFET as compared to other techniques is that no external inputs other than the flow specification for the requested flow are required.
### 4.1 WLAN interface measurements
We follow the convention of denoting the measurement samples by $`\hat{}`$ (e.g., $`\hat{T}`$) and running averages by $`\overline{}`$ (e.g., $`\overline{T}`$).
#### 4.1.1 Frame transmission rate $`{\hat{R}}_{t\hspace{0pt}x}`$
As the station has no means of measuring the packet arrival rate at other stations, we derive the aggregate packet arrival rate to the WLAN, $`\lambda_{M\hspace{0pt}A\hspace{0pt}C}`$ from the measured rate of packet transmissions (successful and collisions) $`{\overline{R}}_{t\hspace{0pt}x}`$.
#### 4.1.2 Average transmission slot duration $`{\hat{T}}_{t\hspace{0pt}x}`$
The throughput of a non-saturated WLAN is greatly influenced by the average duration of a frame transmission which in turn depends on the average frame size for all frames (including higher layer control frames) and PHY data rates used by the transmitting station.
A radio interface is capable of measuring the average duration $`{\hat{T}}_{t\hspace{0pt}x}`$ of transmission. This single measurement abstracts out the effect of the two important variable parameters mentioned above and it suffices because the model requires just the duration of successful and collided transmissions ($`T_{s}`$ and $`T_{c}`$ respectively).
#### 4.1.3 The number of stations with active flows, $`n`$
The number of active stations ($`n`$) is determined from the number of unique transmitters on the channel.
The measurement samples are updated every $`T_{u\hspace{0pt}p\hspace{0pt}d\hspace{0pt}a\hspace{0pt}t\hspace{0pt}e}`$. In order to reduce the effect of short-term dynamics of traffic and channel conditions, we maintain their exponential weighted average with smoothing parameter $`\alpha`$.
<table>
<tbody>
<tr>
<td></td>
<td>T¯txsubscript¯𝑇𝑡𝑥\displaystyle\bar{T}_{tx}</td>
<td>←αT¯tx+(1−α)T^tx←absent𝛼subscript¯𝑇𝑡𝑥1𝛼subscript^𝑇𝑡𝑥\displaystyle\leftarrow\alpha\bar{T}_{tx}+(1-\alpha)\hat{T}_{tx}</td>
<td></td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td>R¯txsubscript¯𝑅𝑡𝑥\displaystyle\bar{R}_{tx}</td>
<td>←αR¯tx+(1−α)R^tx←absent𝛼subscript¯𝑅𝑡𝑥1𝛼subscript^𝑅𝑡𝑥\displaystyle\leftarrow\alpha\bar{R}_{tx}+(1-\alpha)\hat{R}_{tx}</td>
<td></td>
</tr>
</tbody>
</table>
Assuming that the new flow is from an inactive station,
<table>
<tbody>
<tr>
<td></td>
<td>n←n+1←𝑛𝑛1n\leftarrow n+1</td>
<td></td>
</tr>
</tbody>
</table>
### 4.2 Input flow specification
The description of the traffic offered at the link-layer (referred to as $`F\hspace{0pt}l\hspace{0pt}o\hspace{0pt}w\hspace{0pt}S\hspace{0pt}p\hspace{0pt}e\hspace{0pt}c`$) by a new flow will be provided by specifying the following parameters:
- $`\bullet`$
packet arrival rate $`\lambda_{f\hspace{0pt}l\hspace{0pt}o\hspace{0pt}w}`$
- $`\bullet`$
average payload size in bits, $`P\hspace{0pt}A\hspace{0pt}Y\hspace{0pt}L\hspace{0pt}O\hspace{0pt}A\hspace{0pt}D_{f\hspace{0pt}l\hspace{0pt}o\hspace{0pt}w}`$
It should be noted that BUFFET makes use of the declared $`F\hspace{0pt}l\hspace{0pt}o\hspace{0pt}w\hspace{0pt}S\hspace{0pt}p\hspace{0pt}e\hspace{0pt}c`$ only while admitting that particular flow. For the previously admitted flows, the aggregate flow statistics are derived from channel measurements and thus inaccuracy as well as change in $`F\hspace{0pt}l\hspace{0pt}o\hspace{0pt}w\hspace{0pt}S\hspace{0pt}p\hspace{0pt}e\hspace{0pt}c`$ will be automatically adjusted before admitting the next flow.
### 4.3 Deriving model inputs from measurements and $`F\hspace{0pt}l\hspace{0pt}o\hspace{0pt}w\hspace{0pt}S\hspace{0pt}p\hspace{0pt}e\hspace{0pt}c`$
For a moderately loaded WLAN in a steady state, all arrived packets at the interface queues are successfully transmitted on the channel. However, as we are considering random packet arrival processes, momentary queue buildup can happen when collisions occur. Therefore we approximate packet arrival rate to the WLAN to be:
<table>
<tbody>
<tr>
<td></td>
<td>λMAC=Rsucc+Rcoll=R¯txsubscript𝜆𝑀𝐴𝐶subscript𝑅𝑠𝑢𝑐𝑐subscript𝑅𝑐𝑜𝑙𝑙subscript¯𝑅𝑡𝑥\lambda_{MAC}=R_{succ}+R_{coll}=\bar{R}_{tx}</td>
<td></td>
</tr>
</tbody>
</table>
All admitted flows are able to obtain their required throughput that may be different from their fair share as long as the WLAN throughput is less than the capacity. As an approximation, the model assumes a uniform $`\lambda`$ at every station. This approximation does not affect the accuracy of the results as we are not using a station’s fair share of throughput for admission decision. Thus, accounting for the new flow being admitted, $`\lambda`$ per station is then averaged as
<table>
<tbody>
<tr>
<td></td>
<td>λnew=λMACn+λflownsubscript𝜆𝑛𝑒𝑤subscript𝜆𝑀𝐴𝐶𝑛subscript𝜆𝑓𝑙𝑜𝑤𝑛\lambda_{new}=\frac{\lambda_{MAC}}{n}+\frac{\lambda_{flow}}{n}</td>
<td></td>
<td>(1)</td>
</tr>
</tbody>
</table>
For a non-saturated WLAN, we ignore the effect of collision on the measured frame duration. We factor the payload and data rate for the new flow by calculating $`T_{s}^{f\hspace{0pt}l\hspace{0pt}o\hspace{0pt}w}`$, $`T_{s}`$ and $`T_{c}`$ as follows:
<table>
<tbody>
<tr>
<td></td>
<td>Tsflow=DIFS+PHY_HDR+(MAC_HDR+PAYLOADflow)/R+SIFS+PHY_HDR+ACKTs=(λMACnT¯tx+λflownTsflow)(λMACn+λflown)Tc=Ts−(PHY_HDR+ACK+SIFS)superscriptsubscript𝑇𝑠𝑓𝑙𝑜𝑤𝐷𝐼𝐹𝑆𝑃𝐻𝑌_𝐻𝐷𝑅𝑀𝐴𝐶_𝐻𝐷𝑅𝑃𝐴𝑌𝐿𝑂𝐴subscript𝐷𝑓𝑙𝑜𝑤𝑅𝑆𝐼𝐹𝑆𝑃𝐻𝑌_𝐻𝐷𝑅𝐴𝐶𝐾subscript𝑇𝑠subscript𝜆𝑀𝐴𝐶𝑛subscript¯𝑇𝑡𝑥subscript𝜆𝑓𝑙𝑜𝑤𝑛superscriptsubscript𝑇𝑠𝑓𝑙𝑜𝑤subscript𝜆𝑀𝐴𝐶𝑛subscript𝜆𝑓𝑙𝑜𝑤𝑛subscript𝑇𝑐subscript𝑇𝑠𝑃𝐻𝑌_𝐻𝐷𝑅𝐴𝐶𝐾𝑆𝐼𝐹𝑆\begin{split}T_{s}^{flow}&=DIFS+PHY\_HDR\\ &\quad+(MAC\_HDR+PAYLOAD_{flow})/R\\ &\quad+SIFS+PHY\_HDR+ACK\\ T_{s}&=\frac{(\frac{\lambda_{MAC}}{n}\bar{T}_{tx}+\frac{\lambda_{flow}}{n}T_{s}^{flow})}{(\frac{\lambda_{MAC}}{n}+\frac{\lambda_{flow}}{n})}\\ T_{c}&=T_{s}-(PHY\_HDR+ACK+SIFS)\end{split}</td>
<td></td>
<td>(2)</td>
</tr>
</tbody>
</table>
The $`P\hspace{0pt}H\hspace{0pt}Y\hspace{0pt}\_\hspace{0pt}H\hspace{0pt}D\hspace{0pt}R`$ and $`A\hspace{0pt}C\hspace{0pt}K`$ in the above equations are expressed as their respective durations and $`R`$ is the PHY data rate used by the station.
### 4.4 Admission Criterion
As described earlier, a non-zero value of $`\gamma`$ indicates that the WLAN is not saturated. We use $`\gamma_{n\hspace{0pt}e\hspace{0pt}w}`$ predicted by the model as an indicator of saturation:
<table>
<tbody>
<tr>
<td></td>
<td>γnew=Ψ(λnew,n,Ts)subscript𝛾𝑛𝑒𝑤Ψsubscript𝜆𝑛𝑒𝑤𝑛subscript𝑇𝑠\gamma_{new}=\Psi(\lambda_{new},n,T_{s})</td>
<td></td>
<td>(3)</td>
</tr>
</tbody>
</table>
We admit a new flow only if the predicted value of $`\gamma_{n\hspace{0pt}e\hspace{0pt}w}`$ is non-zero. The BUFFET algorithm is illustrated in Fig. [2](#S4.F2 "Figure 2 ‣ 4.4 Admission Criterion ‣ 4 Model based Distributed Admission Control (BUFFET) ‣ An Autonomous Distributed Admission Control Scheme for IEEE 802.11 DCF").
[img.svg](data:image...)
[img.svg](data:image...)
Figure 2: Admission control flowchart
### 4.5 Description of other CAC mechanisms for performance comparison
We compare BUFFET with centralized CAC mechanisms belonging to two other categories, namely saturation model based CACs and threshold based CACs. Although they are originally proposed for EDCA, we adapt them to DCF by considering only one access category as described next.
#### 4.5.1 Call Admission Control based on saturation model (TPUTSAT)
According to the CAC mechanism proposed by Pong and Moors \[[11](#bib.bib11)\] based on Bianchi’s saturation model \[[4](#bib.bib4)\] each station computes the probability $`\tau`$ of a transmission in a slot from the measured probability $`p`$ of collision as
<table>
<tbody>
<tr>
<td></td>
<td>τ=2(1−2p)(1−2p)(W+1)+pW(1−(2p)m)𝜏212𝑝12𝑝𝑊1𝑝𝑊1superscript2𝑝𝑚\tau=\frac{2(1-2p)}{(1-2p)(W+1)+pW(1-(2p)^{m})}</td>
<td></td>
</tr>
</tbody>
</table>
from which $`P_{t\hspace{0pt}r}`$, $`P_{s}`$ and $`T_{s\hspace{0pt}l\hspace{0pt}o\hspace{0pt}t}`$ are obtained. A flow is admitted if the achievable throughput by a station is sufficient to meet the throughput demand:
<table>
<tbody>
<tr>
<td></td>
<td>Sflowsubscript𝑆𝑓𝑙𝑜𝑤\displaystyle S_{flow}</td>
<td>=τ(1−τ)(n−1)PAYLOADflowTslotabsent𝜏superscript1𝜏𝑛1𝑃𝐴𝑌𝐿𝑂𝐴subscript𝐷𝑓𝑙𝑜𝑤subscript𝑇𝑠𝑙𝑜𝑡\displaystyle=\frac{\tau(1-\tau)^{(n-1)}PAYLOAD_{flow}}{T_{slot}}</td>
<td></td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td>Admit if</td>
<td>Sflow>=Requested throughputsubscript𝑆𝑓𝑙𝑜𝑤Requested throughput\displaystyle\quad S_{flow}>=\textrm{Requested throughput}</td>
<td></td>
</tr>
</tbody>
</table>
The packet arrival rate $`\lambda`$ does not need to be supplied for computation of $`S_{f\hspace{0pt}l\hspace{0pt}o\hspace{0pt}w}`$ as it is the throughput at saturation.
#### 4.5.2 Threshold based admission control (AIRTIME)
An airtime allocation and admission control is proposed in \[[6](#bib.bib6)\]. Without consideration for parameterized QoS, the airtime required per second by a flow $`j`$ from station $`i`$ is
<table>
<tbody>
<tr>
<td></td>
<td>ri,j=si,jRisubscript𝑟𝑖𝑗subscript𝑠𝑖𝑗subscript𝑅𝑖r_{i,j}=\frac{s_{i,j}}{R_{i}}</td>
<td></td>
</tr>
</tbody>
</table>
where $`s_{i,j}`$ is throughput requirement of flow $`\{ i,j\}`$ and $`R_{i}`$ is the PHY data rate used by station $`i`$. Assuming the knowledge about all admitted flows, a new flow $`q`$ from station $`p`$ is admitted if:
<table>
<tbody>
<tr>
<td></td>
<td>rp,q+∑i∑jri,j≤EAsubscript𝑟𝑝𝑞subscript𝑖subscript𝑗subscript𝑟𝑖𝑗𝐸𝐴r_{p,q}+\sum_{i}\sum_{j}r_{i,j}\leq~{}EA</td>
<td></td>
</tr>
</tbody>
</table>
where $`E\hspace{0pt}A`$ is the effective airtime ratio or airtime threshold that excludes the control overhead of the resource allocation mechanism.
## 5 Performance Analysis
We analyze the performance of BUFFET, TPUTSAT and AIRTIME through simulations using the OPNET 11.5 modeler \[[1](#bib.bib1)\] according to the parameters given in Table [1](#S5.T1 "Table 1 ‣ 5 Performance Analysis ‣ An Autonomous Distributed Admission Control Scheme for IEEE 802.11 DCF"). In each scenario, a new station requests a flow every 10 seconds. All the flows have fixed payload size and Poisson packet arrivals unless mentioned otherwise. We compare the number of admitted flows (throughput) and average delay after the time when either admission control kicks in or saturation sets in.
<table>
<thead>
<tr>
<th>Area</th>
<th>50x50 m</th>
</tr>
</thead>
<tbody>
<tr>
<th>Number of stations in WLAN</th>
<td>10, 20, 40, 60</td>
</tr>
<tr>
<th>Smoothing parameter α𝛼\alpha</th>
<td>0.8</td>
</tr>
<tr>
<th>Update interval Tupdatesubscript𝑇𝑢𝑝𝑑𝑎𝑡𝑒T_{update}</th>
<td>1 sec.</td>
</tr>
<tr>
<th>Payload size</th>
<td>100, 250, 500 and 1500 B</td>
</tr>
<tr>
<th>Packet arrival process</th>
<td>Poisson, CBR</td>
</tr>
<tr>
<th>PHY data rate</th>
<td>2, 11 Mbps DSSS</td>
</tr>
</tbody>
</table>
Table 1: Simulation Parameters
The delay vs. simulation time curves in Fig. [3](#S5.F3 "Figure 3 ‣ 5 Performance Analysis ‣ An Autonomous Distributed Admission Control Scheme for IEEE 802.11 DCF") with and without CAC (BUFFET) illustrate the working of the CAC mechanism. At 170 seconds, BUFFET determines that the requested flow would cause saturation and hence starts rejecting flows. Accepting flows beyond this point causes the delay to rise sharply.
[Refer to caption](/html/0705.2819/assets/x3.png)
Figure 3: Delays with and without CAC (BUFFET)
Table [2](#S5.T2 "Table 2 ‣ 5 Performance Analysis ‣ An Autonomous Distributed Admission Control Scheme for IEEE 802.11 DCF") summarizes the delay and throughput obtained with BUFFET, TPUTSAT and AIRTIME for representative scenarios. BUFFET is able to maintain the average delay under 7$`m\hspace{0pt}s`$ for all scenarios. More importantly, this consistent delay performance is achieved at a throughput close to the optimum. For example, in scenario-1, BUFFET admits 27 flows; AIRTIME with a threshold of 0.09 admits 30 flows but at the cost of WLAN saturation. This aspect is pictorially depicted by Fig. [4](#S5.F4 "Figure 4 ‣ 5 Performance Analysis ‣ An Autonomous Distributed Admission Control Scheme for IEEE 802.11 DCF") for scenario-3 which shows that BUFFET achieves high utilization at low delays, managing a good balance between delay and utilization.
Table [2](#S5.T2 "Table 2 ‣ 5 Performance Analysis ‣ An Autonomous Distributed Admission Control Scheme for IEEE 802.11 DCF") suggests that the delay and throughput for AIRTIME is a function of the airtime threshold. The optimum threshold itself is variable across scenarios due to the effect of payload size and channel data rate on resource allocation overheads. Therefore, setting a correct threshold is essential for good performance of AIRTIME.
On the other hand, both BUFFET and TPUTSAT avoid saturation and provide low delays without depending on a threshold. TPUTSAT being based on a saturation model provides marginally lower delays but conservatively admits fewer number of flows. This effect is more pronounced for higher $`n`$ when saturation models tend to overestimate the effect of collision and contention. For instance, for a WLAN size of 60 stations(scenario-5) TPUTSAT admits 40% fewer flows than BUFFET.
[Refer to caption](/html/0705.2819/assets/x4.png)
Figure 4: AIRTIME, BUFFET and TPUTSAT delay-throughput comparison
<table>
<tbody>
<tr>
<th></th>
<td colspan="5">scenario-1</td>
<td colspan="5">scenario-2</td>
<td colspan="5">scenario-3</td>
</tr>
<tr>
<th>Sim. parameters</th>
<td colspan="5">(11, 100, 32)</td>
<td colspan="5">(11, 500, 105)</td>
<td colspan="5">(11, 1500, 172)</td>
</tr>
<tr>
<th rowspan="2">CAC scheme</th>
<td rowspan="2">B</td>
<td rowspan="2">T</td>
<td colspan="3">Airtime</td>
<td rowspan="2">B</td>
<td rowspan="2">T</td>
<td colspan="3">Airtime</td>
<td rowspan="2">B</td>
<td rowspan="2">T</td>
<td colspan="3">Airtime</td>
</tr>
<tr>
<td>0.07</td>
<td>0.08</td>
<td>0.09</td>
<td>0.21</td>
<td>0.26</td>
<td>0.31</td>
<td>0.42</td>
<td>0.48</td>
<td>0.54</td>
</tr>
<tr>
<th>Admitted flows</th>
<td>27</td>
<td>23</td>
<td>24</td>
<td>27</td>
<td>30</td>
<td>28</td>
<td>22</td>
<td>21</td>
<td>27</td>
<td>32</td>
<td>29</td>
<td>22</td>
<td>26</td>
<td>30</td>
<td>34</td>
</tr>
<tr>
<th>Delay (ms𝑚𝑠ms)</th>
<td>4.36</td>
<td>1.71</td>
<td>1.90</td>
<td>4.36</td>
<td>4055</td>
<td>3.41</td>
<td>1.94</td>
<td>1.82</td>
<td>2.92</td>
<td>1856</td>
<td>4.82</td>
<td>2.99</td>
<td>3.76</td>
<td>5.55</td>
<td>652</td>
</tr>
</tbody>
</table>
\(a\)
Effect of payload size
<table>
<tbody>
<tr>
<th></th>
<td colspan="5">scenario-4</td>
<td colspan="5">scenario-5</td>
<td colspan="5">scenario-6</td>
</tr>
<tr>
<th>Sim. parameters</th>
<td colspan="5">(2, 500, 33)</td>
<td colspan="5">(11, 500, 57)</td>
<td colspan="5">(11, 500, 400)</td>
</tr>
<tr>
<th rowspan="2">CAC scheme</th>
<td rowspan="2">B</td>
<td rowspan="2">T</td>
<td colspan="3">Airtime</td>
<td rowspan="2">B</td>
<td rowspan="2">T</td>
<td colspan="3">Airtime</td>
<td rowspan="2">B</td>
<td rowspan="2">T</td>
<td colspan="3">Airtime</td>
</tr>
<tr>
<td>0.47</td>
<td>0.54</td>
<td>0.61</td>
<td>0.23</td>
<td>0.26</td>
<td>0.29</td>
<td>0.26</td>
<td>0.31</td>
<td>0.36</td>
</tr>
<tr>
<th>Admitted flows</th>
<td>30</td>
<td>22</td>
<td>28</td>
<td>32</td>
<td>36</td>
<td>50</td>
<td>31</td>
<td>44</td>
<td>50</td>
<td>55</td>
<td>8</td>
<td>8</td>
<td>7</td>
<td>8</td>
<td>9</td>
</tr>
<tr>
<th>Delay (ms𝑚𝑠ms)</th>
<td>6.52</td>
<td>4.24</td>
<td>5.62</td>
<td>8.23</td>
<td>2502</td>
<td>3.03</td>
<td>1.51</td>
<td>2.11</td>
<td>3.03</td>
<td>801</td>
<td>5.83</td>
<td>5.83</td>
<td>3.06</td>
<td>5.83</td>
<td>412</td>
</tr>
</tbody>
</table>
\(b\)
Effect of data rate and flow bandwidth
<table>
<tbody>
<tr>
<th></th>
<td colspan="5">scenario-7</td>
</tr>
<tr>
<th>Sim. parameters</th>
<td colspan="5">(11, 500, 105)</td>
</tr>
<tr>
<th rowspan="2">CAC scheme</th>
<td rowspan="2">B</td>
<td rowspan="2">T</td>
<td colspan="3">Airtime</td>
</tr>
<tr>
<td>0.22</td>
<td>0.26</td>
<td>0.30</td>
</tr>
<tr>
<th>Admitted flows</th>
<td>24</td>
<td>26</td>
<td>22</td>
<td>27</td>
<td>31</td>
</tr>
<tr>
<th>Delay (ms𝑚𝑠ms)</th>
<td>1.36</td>
<td>1.68</td>
<td>1.29</td>
<td>1.87</td>
<td>1651</td>
</tr>
</tbody>
</table>
\(c\)
CBR flows
<table>
<tbody>
<tr>
<td>Sim. parameters:</td>
<td>(PHY rate (Mbps), Payload size (B), Per-flow bandwidth)</td>
</tr>
<tr>
<td>CAC scheme:</td>
<td>B: BUFFET, T: TPUTSAT</td>
</tr>
<tr>
<td></td>
<td>Airtime: AIRTIME with threshold</td>
</tr>
</tbody>
</table>
Table 2: Admitted flows and delay comparison for representative scenarios
For CBR flows (Table. [2(c)](#S5.T2.st3 "In Table 2 ‣ 5 Performance Analysis ‣ An Autonomous Distributed Admission Control Scheme for IEEE 802.11 DCF")) BUFFET conservatively admits fewer flows than TPUTSAT owing to the assumption of Poisson packet arrival. The loss of throughput is however marginal. Lower measured probability of collision due to regular packet arrivals helps TPUTSAT admit more flows.
<table>
<tbody>
<tr>
<td colspan="3">Flow type-1</td>
<td colspan="3">Flow type-2</td>
<td></td>
<td></td>
</tr>
<tr>
<td>PHY Rate</td>
<td>Payload (B)</td>
<td>Flow B/W (Kbps)</td>
<td>PHY Rate</td>
<td>Payload (B)</td>
<td>Flow B/W (Kbps)</td>
<td>Admitted flows</td>
<td>Delay (ms𝑚𝑠ms)</td>
</tr>
<tr>
<td>11</td>
<td>500</td>
<td>100</td>
<td>2</td>
<td>500</td>
<td>33</td>
<td>30</td>
<td>4.83</td>
</tr>
<tr>
<td>2</td>
<td>500</td>
<td>33</td>
<td>11</td>
<td>1500</td>
<td>172</td>
<td>30</td>
<td>6.76</td>
</tr>
<tr>
<td>11</td>
<td>100</td>
<td>32</td>
<td>11</td>
<td>1500</td>
<td>172</td>
<td>28</td>
<td>7.03</td>
</tr>
<tr>
<td>11</td>
<td>500</td>
<td>100</td>
<td>11</td>
<td>100</td>
<td>32</td>
<td>29</td>
<td>4.06</td>
</tr>
</tbody>
</table>
Table 3: BUFFET with for non-uniform flows
As mentioned before, BUFFET adapts to non-uniform payload sizes as well as data rates through $`T_{t\hspace{0pt}x}`$ measurements. To verify this, we conducted another set of simulations with BUFFET for two different types of flows as listed in Table [3](#S5.T3 "Table 3 ‣ 5 Performance Analysis ‣ An Autonomous Distributed Admission Control Scheme for IEEE 802.11 DCF"). The first 20 flows requested are of type-1 and next 20 flows are of type-2. Delays in this case too are less than 7$`m\hspace{0pt}s`$, confirming that BUFFET works well without any configuration even for combinations of diverse data rates and application types.
BUFFET is therefore ideal for realistic WLAN deployments with diverse applications and channel conditions, providing a fully distributed, zero-configuration autonomous setup.
## 6 Conclusion
In this work, we propose an autonomous distributed admission control named BUFFET for 802.11 DCF that is based on an analytical model. In order to keep the average delay within acceptable limits, BUFFET admits a flow only if it does not lead to WLAN saturation, an indicator of which is a parameter $`\gamma`$ predicted by the model. BUFFET is able to derive all inputs to the model from the measurements (frame transmission rate, average transmission duration and number of stations) and requested FlowSpec.
Performance analysis through OPNET simulations suggests that BUFFET is able to provide consistent sub-7$`m\hspace{0pt}s`$ delay while achieving near-optimal throughput. We also compare the performance of BUFFET with two other admission control schemes, one based on saturation throughput (TPUTSAT) and the other based on airtime threshold (AIRTIME). TPUTSAT is found to be too conservative in admitting flows, especially for higher number of stations. Configuration of correct threshold (which itself is widely variable based on load and data rate) is essential for correct operation of AIRTIME.
The fully distributed nature of BUFFET, wherein it is able to deduce information about already admitted flows, coupled with its ability to work correctly for a combination of diverse data rates and payload sizes makes it ideal for zero-configuration self-regulating distributed WLAN setup.
We are currently implementing BUFFET for Atheros chipset based 802.11g WLAN cards on GNU/Linux systems. Applying the algorithm to 802.11e EDCA by extending the model and using similar measurements per access category is another future direction we are pursuing.
## References
- \[1\] OPNET Technologies, Inc. http://www.opnet.com.
- \[2\] A. Bai, T. Skeie, and P. Engelstad. A model-based admission control for 802.11e EDCA using delay predictions.
In Proceedings of IPCCC, 2007.
- \[3\] A. Bazzi, M. Diolaiti, and G. Pasolini. Measurement based call admission control strategies in infrastructured IEEE 802.11 WLANs.
In Proceedings of PIMRC, volume 3, pages 2093–2098, 2005.
- \[4\] G. Bianchi. Performance analysis of the IEEE 802.11 distributed coordination function.
IEEE JSAC, 18(3):535–547, March 2000.
- \[5\] G. Bianchi and I. Tinnirello. Kalman filter estimation of the number of competing terminals in an IEEE 802.11 network.
In Proceedings of INFOCOM, volume 2, pages 844–852, 2003.
- \[6\] C. T. Chou, S. Shankar, and K. Shin. Achieving per-stream QoS with distributed airtime allocation and admission control in IEEE 802.11e wireless LANs.
In Proceeding of INFOCOM, volume 3, pages 1584–1595, 2005.
- \[7\] P. Engelstad and O. Osterbo. Delay and throughput analysis of IEEE 802.11e EDCA with starvation prediction.
In Proceedings of LCN, pages 647–655, 2005.
- \[8\] D. Gu and J. Zhang. A new measurement-based admission control method for IEEE802.11 wireless local area networks.
In Proceedings of PIMRC, volume 3, pages 2009–2013, 2003.
- \[9\] Z. N. Kong, D. Tsang, and B. Bensaou. Measurement-assisted model-based call admission control for IEEE 802.11e WLAN contention-based channel access.
In Proceedings of LANMAN, pages 55–60, 2004.
- \[10\] P. Patil and V. Apte. Sizing of IEEE 802.11 wireless LANs.
In Proceedings of ACM WMASH, pages 96–99, 2005.
- \[11\] D. Pong and T. Moors. Call admission control for IEEE 802.11 contention access mechanism.
In Proceedings of GLOBECOM ’03, volume 1, pages 174–178, 2003.
- \[12\] H. Zhai, X. Chen, and Y. Fang. A call admission and rate control scheme for multimedia support over IEEE 802.11 wireless LANs.
In Proceedings of QSHINE, pages 76–83, 2004.
- \[13\] H. Zhai, X. Chen, and Y. Fang. How well can the IEEE 802.11 wireless LAN support quality of service?
IEEE Transactions on Wireless Communications, 4(6):3084–3094, November 2005.
## Appendix A DTMC model for non-saturated 802.11 DCF
We model the behavior of a single station using a Discrete Time Markov Chain (DTMC). The model is based on the model proposed by Bianchi \[[4](#bib.bib4)\] for saturated WLANs using a DTMC. To account for non-saturated conditions we add a state called *idle* to this DTMC (Fig. [5](#A1.F5 "Figure 5 ‣ Appendix A DTMC model for non-saturated 802.11 DCF ‣ An Autonomous Distributed Admission Control Scheme for IEEE 802.11 DCF")). The station after completing a packet transmission performs a mandatory backoff with random backoff counter picked from $`(0,{{C\hspace{0pt}W_{M\hspace{0pt}I\hspace{0pt}N}} - 1})`$
<sup>2</sup>
<sup>2</sup>2We follow the same terminology as that used in \[[4](#bib.bib4)\].
(*post-backoff*). If the MAC transmission queue is empty after post-backoff, it goes in *idle* state. It remains in this state till the end of the slot corresponding to the first packet arrival. We define the probability $`\gamma`$ as the probability that the station queue is non-empty after the post-backoff. At saturation, this probability becomes zero.
Accordingly, there are four possibilities after completion of post-backoff:
- •
Station already has packets in its queue; it transmits back-to-back packets after post-backoff without going in *idle* state.
- •
A packet arrived in *idle* state, and during a silent slot on the channel is transmitted in the next slot.
- •
A packet arrived in *idle* state and during a slot corresponding to a transmission, before the transmission could be sensed (initial CCATime). The station is required to backoff in this case, similar to post-backoff.
- •
A packet arrived in *idle* state and during a slot corresponding to a (successful or collided) transmission but after the transmission has been sensed. The station is allowed to transmit after the end of the ongoing transmission and the subsequent DIFS silence.
[Refer to caption](/html/0705.2819/assets/x5.png)
Figure 5: Markov chain for a single station
The state of the DTMC is defined by the tuple $`\{{s\hspace{0pt}{(t)}},{b\hspace{0pt}{(t)}}\}`$ as defined in \[[4](#bib.bib4)\] or $`i\hspace{0pt}d\hspace{0pt}l\hspace{0pt}e`$. Similar to \[[4](#bib.bib4)\], we assume a constant and independent conditional collision probability $`p`$ and ignore the relevant retry limits (ShortRetryLimit or LongRetryLimit). In addition, we assume that packet arrival to the MAC are Poisson with rate $`\lambda`$. We define $`P_{a\hspace{0pt}s},P_{a\hspace{0pt}c}`$, $`P_{a\hspace{0pt}i}`$ and $`P_{a}`$ as the probabilities of a packet arrival during successful transmission slot, collision slot, idle slot and a generic slot respectively.
<table>
<tbody>
<tr>
<td></td>
<td>Pas=1−e−λTsPac=1−e−λTcPai=1−e−λσPa=PtrPsPas+Ptr(1−Ps)Pac+(1−Ptr)Paisubscript𝑃𝑎𝑠1superscript𝑒𝜆subscript𝑇𝑠subscript𝑃𝑎𝑐1superscript𝑒𝜆subscript𝑇𝑐subscript𝑃𝑎𝑖1superscript𝑒𝜆𝜎subscript𝑃𝑎subscript𝑃𝑡𝑟subscript𝑃𝑠subscript𝑃𝑎𝑠subscript𝑃𝑡𝑟1subscript𝑃𝑠subscript𝑃𝑎𝑐1subscript𝑃𝑡𝑟subscript𝑃𝑎𝑖\begin{split}P_{as}&=1-e^{-\lambda T_{s}}\\ P_{ac}&=1-e^{-\lambda T_{c}}\\ P_{ai}&=1-e^{-\lambda\sigma}\\ P_{a}&=P_{tr}P_{s}P_{as}+P_{tr}(1-P_{s})P_{ac}+(1-P_{tr})P_{ai}\end{split}</td>
<td></td>
<td>(4)</td>
</tr>
</tbody>
</table>
The single-step transition probabilities are:
<table>
<tbody>
<tr>
<td></td>
<td colspan="2">P{i,k|i,k+1}=1,k∈(0,Wi−2),i∈(1,m)P{0,k|i,0}=(1−p)W−1,k∈(1,W−1),i∈(0,m)P{i,k|i−1,0}=pWi,k∈(0,Wi−1),i∈(1,m)P{m,k|m,0}=pWm,k∈(0,Wm−1)P{idle|idle}=1−PaP{0,k|idle}=Ptr(1−e−λCCA)W,k∈(1,W−1)\displaystyle\begin{split}P\{i,k~{}|~{}i,k+1\}&=1,\quad k\in(0,W_{i}-2),~{}i\in(1,m)\\ P\{0,k~{}|~{}i,0\}&=\frac{(1-p)}{W-1},\quad k\in(1,W-1),~{}i\in(0,m)\\ P\{i,k~{}|~{}i-1,0\}&=\frac{p}{W_{i}},\quad k\in(0,W_{i}-1),~{}i\in(1,m)\\ P\{m,k~{}|~{}m,0\}&=\frac{p}{W_{m}},\quad k\in(0,W_{m}-1)\\ P\{idle~{}|~{}idle\}&=1-P_{a}\\ P\{0,k~{}|~{}idle\}&=\frac{P_{tr}(1-e^{-\lambda~{}CCA})}{W},\quad k\in(1,W-1)\\ \end{split}</td>
<td></td>
</tr>
</tbody>
</table>
<table>
<tbody>
<tr>
<td></td>
<td colspan="2">P{0,0|idle}=PtrPs(e−λCCA−e−λTs)+Ptr(1−Ps)(e−λCCA−e−λTc)+(1−Ptr)Pai+Ptr(1−e−λCCA)WP{idle|0,1)=γP{0,0|0,1}=1−γP{0,k|0,k+1}=1,k∈(1,W−1)\displaystyle\begin{split}P\{0,0~{}|~{}idle\}&=P_{tr}P_{s}(e^{-\lambda~{}CCA}-e^{-\lambda~{}T_{s}})\\ &\quad+P_{tr}(1-P_{s})(e^{-\lambda~{}CCA}-e^{-\lambda~{}T_{c}})\\ &\quad+(1-P_{tr})P_{ai}+\frac{P_{tr}(1-e^{-\lambda~{}CCA})}{W}\\ P\{idle~{}|~{}0,1)&=\gamma\\ P\{0,0~{}|~{}0,1\}&=1-\gamma\\ P\{0,k~{}|~{}0,k+1\}&=1,\quad k\in(1,W-1)\end{split}</td>
<td></td>
</tr>
</tbody>
</table>
This is an aperiodic, irreducible Markov chain for which steady state probabilities are known to exist. Applying the normalization condition
<table>
<tbody>
<tr>
<td></td>
<td>1=∑i=0m∑k=0Wibi,k+bidle1superscriptsubscript𝑖0𝑚superscriptsubscript𝑘0subscript𝑊𝑖subscript𝑏𝑖𝑘subscript𝑏𝑖𝑑𝑙𝑒1=\sum_{i=0}^{m}\sum_{k=0}^{W_{i}}b_{i,k}+b_{idle}</td>
<td></td>
<td>(5)</td>
</tr>
</tbody>
</table>
and solving the chain, the probabilities $`\tau`$, $`P_{t\hspace{0pt}r}`$, $`P_{s}`$ and throughput $`S`$ are obtained similar to \[[4](#bib.bib4)\].
### A.1 Calculation of MAC service time ($`D_{M\hspace{0pt}A\hspace{0pt}C}`$)
The MAC service time is the duration from the time a packet becomes head of the queue to the time it is acknowledged. The number of transmissions before the packet is successfully transmitted follows a modified geometric distribution with parameter $`p`$. The delays and probabilities for the four possibilities after successful transmission mentioned above are calculated as follows:
<table>
<tbody>
<tr>
<td></td>
<td>Db2bsubscript𝐷𝑏2𝑏\displaystyle D_{b2b}</td>
<td>=Ts+Tcp1−p+BOslotsabsentsubscript𝑇𝑠subscript𝑇𝑐𝑝1𝑝𝐵subscript𝑂𝑠𝑙𝑜𝑡𝑠\displaystyle=T_{s}+\frac{T_{c}~{}p}{1-p}+BO_{slots}</td>
<td></td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td>Pb2bsubscript𝑃𝑏2𝑏\displaystyle P_{b2b}</td>
<td>=1−γabsent1𝛾\displaystyle=1-\gamma</td>
<td></td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td>Dnob2b_idlesubscript𝐷𝑛𝑜𝑏2𝑏_𝑖𝑑𝑙𝑒\displaystyle D_{nob2b\_idle}</td>
<td>=Db2b−WTslot2absentsubscript𝐷𝑏2𝑏𝑊subscript𝑇𝑠𝑙𝑜𝑡2\displaystyle=D_{b2b}-\frac{W~{}T_{slot}}{2}</td>
<td></td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td>Pnob2b_idlesubscript𝑃𝑛𝑜𝑏2𝑏_𝑖𝑑𝑙𝑒\displaystyle P_{nob2b\_idle}</td>
<td>=γ(1−Ptr)PaiPaabsent𝛾1subscript𝑃𝑡𝑟subscript𝑃𝑎𝑖subscript𝑃𝑎\displaystyle=\frac{\gamma(1-P_{tr})P_{ai}}{P_{a}}</td>
<td></td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td>Dnob2b_txsubscript𝐷𝑛𝑜𝑏2𝑏_𝑡𝑥\displaystyle D_{nob2b\_tx}</td>
<td>=Db2b−WTslot2+[PsTs+(1−Ps)Tc]2absentsubscript𝐷𝑏2𝑏𝑊subscript𝑇𝑠𝑙𝑜𝑡2delimited-[]subscript𝑃𝑠subscript𝑇𝑠1subscript𝑃𝑠subscript𝑇𝑐2\displaystyle=D_{b2b}-\frac{W~{}T_{slot}}{2}+\frac{[P_{s}~{}T_{s}+(1-P_{s})~{}T_{c}]}{2}</td>
<td></td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td>Pnob2b_txsubscript𝑃𝑛𝑜𝑏2𝑏_𝑡𝑥\displaystyle P_{nob2b\_tx}</td>
<td>=γ[PtrPs(e−λCCA−e−λTs)]Paabsent𝛾delimited-[]subscript𝑃𝑡𝑟subscript𝑃𝑠superscript𝑒𝜆𝐶𝐶𝐴superscript𝑒𝜆subscript𝑇𝑠subscript𝑃𝑎\displaystyle=\frac{\gamma[P_{tr}P_{s}(e^{-\lambda~{}CCA}-e^{-\lambda~{}T_{s}})]}{P_{a}}</td>
<td></td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td></td>
<td>+γ[Ptr(1−Ps)(e−λCCA−e−λTc)]Pa𝛾delimited-[]subscript𝑃𝑡𝑟1subscript𝑃𝑠superscript𝑒𝜆𝐶𝐶𝐴superscript𝑒𝜆subscript𝑇𝑐subscript𝑃𝑎\displaystyle\quad+\frac{\gamma[P_{tr}(1-P_{s})(e^{-\lambda~{}CCA}-e^{-\lambda~{}T_{c}})]}{P_{a}}</td>
<td></td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td>Dnob2b_ccasubscript𝐷𝑛𝑜𝑏2𝑏_𝑐𝑐𝑎\displaystyle D_{nob2b\_cca}</td>
<td>=Db2b+PsTs+(1−Ps)Tcabsentsubscript𝐷𝑏2𝑏subscript𝑃𝑠subscript𝑇𝑠1subscript𝑃𝑠subscript𝑇𝑐\displaystyle=D_{b2b}+P_{s}~{}T_{s}+(1-P_{s})~{}T_{c}</td>
<td></td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td>Pnob2b_ccasubscript𝑃𝑛𝑜𝑏2𝑏_𝑐𝑐𝑎\displaystyle P_{nob2b\_cca}</td>
<td>=γPtr(1−e−λCCA)Paabsent𝛾subscript𝑃𝑡𝑟1superscript𝑒𝜆𝐶𝐶𝐴subscript𝑃𝑎\displaystyle=\frac{\gamma P_{tr}(1-e^{-\lambda~{}CCA})}{P_{a}}</td>
<td></td>
</tr>
</tbody>
</table>
where $`{B\hspace{0pt}O_{s\hspace{0pt}l\hspace{0pt}o\hspace{0pt}t\hspace{0pt}s}} = {\frac{{({1 - p})}\hspace{0pt}W\hspace{0pt}T_{s\hspace{0pt}l\hspace{0pt}o\hspace{0pt}t}}{2}\hspace{0pt}\left\lbrack {\frac{2\hspace{0pt}{({1 - {2\hspace{0pt}p^{m}}})}}{1 - {2\hspace{0pt}p}} + \frac{\lbrack{{{({2\hspace{0pt}p})}^{m}\hspace{0pt}{({2 - p})}} - {({1 - p})}}\rbrack}{{({1 - p})}^{2}}} \right\rbrack}`$ and $`T_{s\hspace{0pt}l\hspace{0pt}o\hspace{0pt}t}`$ is the average duration of a logical slot. The average MAC service time, $`D_{M\hspace{0pt}A\hspace{0pt}C}`$, is the conditional average of the above.
### A.2 Calculation of $`\gamma`$
Treating the WLAN interface as an $`M/G/1`$ queue, from the definition of queue utilization ($`\rho`$) and $`\gamma`$, we obtain
<table>
<tbody>
<tr>
<td></td>
<td colspan="2">ρ=min(1,λ/μ)=min(1,λDMAC)γ=P{MAC queue is empty}=1−ρ𝜌min1𝜆𝜇min1𝜆subscript𝐷𝑀𝐴𝐶𝛾𝑃MAC queue is empty1𝜌\displaystyle\begin{split}\rho&=\textrm{min}(1,\lambda/\mu)=\textrm{min}(1,\lambda~{}D_{MAC})\\ \gamma&=P\{\text{MAC queue is empty}\}=1-\rho\end{split}</td>
<td></td>
<td>(6)</td>
</tr>
</tbody>
</table>
The values of $`\gamma`$, $`\tau`$ and $`D_{M\hspace{0pt}A\hspace{0pt}C}`$ are obtained numerically through successive iterations. For convenience, we express $`\gamma`$ as a function of three load-dependent variables:
<table>
<tbody>
<tr>
<td></td>
<td>γ=Ψ(λ,n,Ts)𝛾Ψ𝜆𝑛subscript𝑇𝑠\gamma=\Psi(\lambda,n,T_{s})</td>
<td></td>
<td>(7)</td>
</tr>
</tbody>
</table> |
0705.1939 | cs.NI | cyber | # Towards Informative Statistical Flow Inversion
Richard Clegg
Hamed Haddadi
Raul Landa
Miguel Rio
University College London United Kingdom [\[rclegg—hamed—rlanda—mrio\]@ee.ucl.ac.uk](mailto:%5Brclegg%E2%80%94hamed%E2%80%94rlanda%E2%80%94mrio%5D@ee.ucl.ac.uk)
###### Abstract
A problem which has recently attracted research attention is that of estimating the distribution of flow sizes in internet traffic. On high traffic links it is sometimes impossible to record every packet. Researchers have approached the problem of estimating flow lengths from sampled packet data in two separate ways. Firstly, different sampling methodologies can be tried to more accurately measure the desired system parameters. One such method is the sample-and-hold method where, if a packet is sampled, all subsequent packets in that flow are sampled. Secondly, statistical methods can be used to “invert” the sampled data and produce an estimate of flow lengths from a sample.
In this paper we propose, implement and test two variants on the sample-and-hold method. In addition we show how the sample-and-hold method can be inverted to get an estimation of the genuine distribution of flow sizes. Experiments are carried out on real network traces to compare standard packet sampling with three variants of sample-and-hold. The methods are compared for their ability to reconstruct the genuine distribution of flow sizes in the traffic.
###### category:
G.3 Probability and Statistics Distribution functions
###### category:
C.2.3 Computer-Communication Networks Network Operations
###### keywords:
Network monitoring
###### keywords:
Sampling, Inference, Inversion
<sup>†</sup>
<sup>†</sup>terms: Statistical Inversion, Measurement
## 1 Introduction
Routers at the core of the internet deal with millions of packets per second on multiple interfaces. From a network operations perspective, it is vital for the administrators to be aware of the volume and types of the packets that are traversing their networks. In order to achieve this objective, routers are required to collect management information but it is impossible to keep a record of all the packets. Thus, given the vast amount of information that needs to be collected, routers sample the traffic stream. This means that only a subset of the packets traversing any interface of the router are processed. Today, the most commonly implemented technique is packet sampling, where 1 packet out of every $`N`$ is chosen on a random or periodic basis, and integrated into a flow record in the router memory.
In many practical cases, $`1/N`$ packet sampling is followed by multiplication of the recovered statistics by $`N`$ (N-multiplication). This simple technique can be used to recover a number of packet level statistics of interest. For example, the number of SYN packets, TCP packets, ICMP packets or packets to or from given destinations in the original trace can be estimated by this process. However, the distribution of flow lengths can not be recovered by this procedure (see Section [1.3](#S1.SS3 "1.3 Packet sampling ‣ 1 Introduction ‣ Towards Informative Statistical Flow Inversion")).
The problem at the heart of this paper is that of recovering the distribution of flow lengths from sampled data. The flow inversion problem amounts to mathematically compensating for the effects of sampling in order to estimate the distribution of flow lengths which would have been observed in the original data. There has been great research activity around the flow distribution inversion \[[1](#bib.bib1), [2](#bib.bib2), [3](#bib.bib3)\] and this is discussed in Section [1.5](#S1.SS5 "1.5 Related work ‣ 1 Introduction ‣ Towards Informative Statistical Flow Inversion"). Based on the previous studies on analysis of the NetFlow performance \[[4](#bib.bib4)\], it is evident that N-multiplication upsets the flow level statistics of the original data stream \[[5](#bib.bib5)\].
### 1.1 Outline
This paper focuses on considering sampling methods and their use to estimate the flow length distribution in real traffic. Packet sampling has many useful statistical properties, but it is a hard problem to recover the flow distribution from packet-sampled data. This is sometimes called the flow inversion problem. In this paper we investigate three techniques based on the idea of the sample-and-hold method \[[6](#bib.bib6)\]. This method was originally conceived to track the largest flows in the traffic (with applications related to billing) \[[7](#bib.bib7)\].
Section [1.3](#S1.SS3 "1.3 Packet sampling ‣ 1 Introduction ‣ Towards Informative Statistical Flow Inversion") gives some basic information about packet sampling as applied in real situations. Section [1.5](#S1.SS5 "1.5 Related work ‣ 1 Introduction ‣ Towards Informative Statistical Flow Inversion") discusses other work on the problem of inferring flow distributions from sampled data. In Section [2](#S2 "2 Methodology ‣ Towards Informative Statistical Flow Inversion") we describe the sampling methods we use and the inversion procedure used to recover the original flow distribution. There is also a brief overview on the router memory and resources requirements. In Section [3](#S3 "3 Results ‣ Towards Informative Statistical Flow Inversion") we have applied our proposed algorithms on packet traces from a backbone network and have looked at the performance of our algorithm. In Section [4](#S4 "4 Conclusions and future work ‣ Towards Informative Statistical Flow Inversion") we have summarized our results and discuss potential avenues for future work.
### 1.2 Definitions
The traffic on the Internet is carried in form of Internet Protocol (IP) packets and transmitted to the destination on a hop-by-hop basis by Internet routers. In order to keep account of the packets belonging to the same application, the concept of a flow is defined by router manufacturers. A flow is usually defined as a group of packets that have the same 5-tuple (IP protocol, source address, source port, destination address, destination port).
<sup>1</sup>
<sup>1</sup>1The industry sometimes uses other definitions such as 7-tuple format, however we choose the 5-tuple format which is more commonly used in the research context.
Usually, core Internet routers carry a large number of flows at any given time. This pressure on the router is controlled by using strict rules to remove from router memory (*export*) the statistics, and thus keep the router memory buffer and CPU resources available to deal with changes in traffic patterns by avoiding the handling of excessively large tables of flow records. Cisco NetFlow \[[8](#bib.bib8)\], the dominant standard on today’s routers, uses the following criteria for expiring flows in the cache entries:
1. 1\.
Flows which have been idle for a specified time are expired and removed from the cache (15 seconds is default).
2. 2\.
Long lived flows are expired and removed from the cache (30 minutes is default).
3. 3\.
As the cache becomes full a number of heuristics are applied to aggressively age groups of flows simultaneously.
4. 4\.
TCP connections which have reached the end of byte stream (FIN) or which have been reset (RST) will be expired.
As will be seen in Section [2.2](#S2.SS2 "2.2 Flow termination dilemma ‣ 2 Methodology ‣ Towards Informative Statistical Flow Inversion"), the selection of these parameters can greatly affect the nature of the sampled traffic.
After the flow records are terminated, they are grouped together and exported to an external aggregation point through a UDP (User datagram Protocol) stream. The collection of these NetFlow records enables system administrators to have a view of general trends in spatial traffic distribution, network host behavior, traffic matrix estimation, anomaly detection \[[9](#bib.bib9)\] and other relevant measurements. However, the effectiveness of these applications is contingent upon the quality of the flow level statics recovered from the actual network measurements \[[10](#bib.bib10)\].
The flow distribution is the distribution of flow lengths in a given traffic trace. The lengths are usually expressed in packets but sometimes in bytes. This can be thought of as the probability that a given flow has a particular length. That is, the distribution is $`(\theta_{1},\ldots,\theta_{M})`$ where
<table>
<tbody>
<tr>
<td></td>
<td>θi=ℙ[Randomly chosen flow is of length i].subscript𝜃𝑖ℙdelimited-[]Randomly chosen flow is of length i\theta_{i}=\mathbb{P}\left[\text{Randomly chosen flow is of length $i$}\right].</td>
<td></td>
</tr>
</tbody>
</table>
### 1.3 Packet sampling
In an analysis by Cisco \[[11](#bib.bib11)\] one NetFlow-enabled access router used up to 68% of its CPU on processing flow records when an average of 65,000 flows was kept in memory. When sampling was used, this utilization was decreased by more than 82%. There are three constraints on a core router which lead to the use packet sampling: the size of the record buffer, the CPU speed and the flow record look-up time. In packet sampling, in order to relax the pressure on the router while collecting measurements, $`1`$ in $`N`$ packets are chosen, and the rest are discarded. Sometimes this is done in a periodic way with every $`N`$th packet sampled. However, in the literature, independent and identically distributed (iid) sampling with a fixed probability $`p`$ is often considered. The differences between periodic sampling and iid sampling can be important. Roughan \[[12](#bib.bib12)\] has shown iid sampling is useful in active probing and the concepts are also applicable to the case of passive measurement.
There are many advantages to iid packet sampling and it preserves many important characteristics of the traffic. However, this sampling does not preserve the flow length distribution. The reason for this should be clear but an example is illustrative. Imagine a situation where the flow distribution is such that half the flows in the original trace are of length two and half are of length one ($`\theta_{1} = 0.5`$ and $`\theta_{2} = 0.5`$). Imagine these packets are sampled in an iid manner with $`p = 0.5`$. Half of the flows of length one will be sampled but only one quarter of the flows of length two will have both packets sampled. Another half of the flows of length two will have just one packet sampled and a final quarter will have no packets sampled. In the final sample the flow distribution will be ($`\theta_{1}' = 0.8`$ and $`\theta_{2}' = 0.2`$). The problem of flow inversion is, therefore, defined as the problem of recovering the original distribution ($`\theta_{i}`$) from the sampled traffic.
The choice of sampling strategy will have a large impact on the quality of the data obtained from the network. This is why, to an extent, thereason why the usability of NetFlow sampled data has been questioned by researchers \[[13](#bib.bib13)\]. The problems with packet sampling are twofoldstem from the followinf effects it has on sampled flows:
1. 1\.
It is easy to miss short flows altogether. This is due to the fact that many flows be only a few packets long, and they may be temporally correlated. Thus, these constituent packets may cluster together and totally evade the sampling process.
2. 2\.
It can be difficult to estimate the length of long flows. The major problem is that, for each flow, only a small subset of packets are seen with a given probability $`p`$ equal to the sampling rate. Thus, it is not clear how many packets actually were present, out of which a given number $`X_{i}`$ were been seen.
3. 3\.
Flows may be mis-ranked. This means that, even though flow $`A`$ may seem to be larger than flow $`B`$ in the sampled statistics, this is not necessarily the case in reality \[[14](#bib.bib14)\].
4. 4\.
Large flows may be split into smaller ones (creating sparse flows) \[[2](#bib.bib2)\]. This is due to the fact that some long flows have a bursty nature, and thus may include long periods of inactivity. During these periods, they might be mistakenly expired, and any new packets belonging to the same flow are mistakenly classified as part of a new flow.
For applications such as billing and monitoring, the naïve inversion method of division of the final statistics by the sampling rate, or basically multiplying the final data by ($`1/p`$) will simply lead to inaccurate results as pointed in \[[5](#bib.bib5)\].
On the other hand, the most important problem of the current NetFlow implement ation of packet sampling is the fact that many flows are not sampled at all, as none of their packets are selected for sampling. In many cases, particularly in short-lived flows (like web and email applications, where a group of packets are sent together to reply a query), only one packet of the flow is captured. This results in NetFlow reports being dominated by single packet flows.
### 1.4 Practical implementation of sampling techniques
Even though sampling sampling techniques are used in order to simplify the processing of data collected at a router, in practice it is a complicated process in itself. In a simple implementation of packet sampling in a router, there are various points to be considered. A core or large access Internet router must constantly accomodate memory and processor resource constraints. Even though it is desirable to keep a large number of records in the flow cache, the fast growth of this number makes flow look-up and update a challenge. Thus, when sampling is implemented, the operator has to decide on a few parameters.
1. 1\.
Sampling rate: The sampling rate has a direct impact on the quantity and the quality of the information formed from the data.
2. 2\.
Flow time-out: The length of the time-out can have an impact on intermittent traffic flows, such as peer-to-peer file sharing or Instant Messaging, where the flows may not by transmitting packets at full rate the whole time.
3. 3\.
Flow expiry: If many large flows are active, the flow cache of the router becomes progressively full, leaving no space for new flows. In order to avoid this, a value must be chosen for the expiration (timeout) of the flows.
4. 4\.
Flow export frequency: If this is done too frequently, it increases the processing load on the router. However, if it is not done often enough, the loss of the UDP export packets in the path can effect the quality of the gathered statistics.
5. 5\.
Flow cache size: The number of flows which are kept at the router plays a critical role on its performance. If this number is too large, flow look-up time becomes a difficult. On the other hand, if it is too small, many flows are bound to be dropped and frequent expiry and time-out of flows will be needed.
It can be observed that optimally setting all these values can be a challenging task for an operator. Changes in any of the above parameters can effect the length and number of the flows which are reported by a router. In Section [2.2](#S2.SS2 "2.2 Flow termination dilemma ‣ 2 Methodology ‣ Towards Informative Statistical Flow Inversion") we look at some of the effects of the mentioned parameters.
### 1.5 Related work
Hohn and Veitch \[[3](#bib.bib3)\] considered in some depth the problem of producing an estimate of flow distribution from sampled packets. They first looked at methods for “inversion” to recreate the original flow distribution from the sampled packet data. They use two schemes to recreate the flow distribution from the packet sampled data, the first based upon a binomial sum and the second upon a Cauchy integral. These schemes can successfully recover the flow distribution for short length flows if the sampling rate $`p`$ is relatively high (for example, more than half the packets sampled is ideal). This is not a flaw in the methods described, but a fundamental limitation in the amount of information which can be retrieved from packets sampled in this manner.
Following this, their paper proposes a flow sampling model which can be used in an offline analysis of flow records formed from an unsampled packet stream. In this method, all the packets are recorded and formed into flows. Then, a subset of these recreated flows are sampled using iid sampling with a given probability $`p`$. This sampling method proves extremely successful at recreating the flow distribution even when the sampling ratio $`p`$ is relatively small (say $`p = 0.001`$). However, the intensive computing and memory requirements makes the implementation of such a scheme on high speed routers a challenge.
Duffield et al \[[2](#bib.bib2)\] have looked at recovering the flow length distributions from a sampled packet trace. A scaling based, Maximum Likelihood Estimation method is proposed and, due to its complexity, an iterative Expectation Maximization algorithm is tested on the available trace files. The biggest issue encountered by the authors is the complexity of the process and the adjustment of the lowest order weights to reflect the underlying distributions. It is re-established by the authors that the estimation of flow level statistics from packet sampled data remains an open question.
In a subsequent paper, \[[15](#bib.bib15)\] the authors introduce threshold sampling as a sampling scheme that optimally controls the expected volume of samples and the variance of estimators over any classification of flows. The proposed scheme has packet capturing performed at routers, followed by flow formation and export and staging at a mediation station, and aggregation of records at a measurement collector.
Ribeiro et al \[[1](#bib.bib1)\] use several methods to estimate the flow distribution from sampled packets. They make use of several features of the TCP protocol, including the SYN flag, and the fact that sequence numbers can give information about the number of bytes between sampled packets. Their work uses maximum likelihood estimators to fit a the distribution of flow lengths up to some maximum flow length (maximum flow lengths of twenty and one hundred are used in the paper). The sequence numbers in particular prove helpful in extracting information about these short and mid-length flows. In addition they use a technique based on the Cramér–Rao bound to investigate the best possible (lowest variance) performance of unbiased flow distribution estimators given assumptions about the information available.
Estan and Varghese \[[6](#bib.bib6)\] propose two algorithms for identifying the large flows: *sample and hold* and multistage filters, which take a constant number of memory references per packet and use a small amount of memory. If $`M`$ is the available memory, the errors of the algorithms are proportional to $`1/M`$; by contrast, the error of an algorithm based on classical sampling is proportional to $`1/\sqrt{(}M)`$, thus providing much less accuracy for the same amount of memory. This scheme is intended for billing schemes where large flows are of higher interest to the operator.
Estan et al \[[16](#bib.bib16)\] have proposed an improvement to NetFlow by adapting the sampling rate, enabling the router to keep a pre-determined number of flows in the cache. As a result of change in sampling rate, at each stage a normalization step is performed which ignores the packets that would not have been sampled if the lower sampling rates had been chosen. This scheme produces more concise but less accurate reports, due to reduction in the collected information. The constant change in sampling rate and renormalization stage can be an exploitable threat to the router, allowing the degradation of its performance performance under some attack scenarios.
Barakat et al \[[14](#bib.bib14)\] study the possibility of detection and ranking of the largest flows on a link. A comparison is made between the blind ranking method and study how to detect and rank the largest flows on a link. The results indicate that at sampling rates of higher than 1 in a 100 it is difficult to identify the top flow with both methods.
Although not strictly relevant to the inversion problem it is worth noting that there is considerable research interest in the distribution of flow lengths in internet traffic. This is because the flow lengths are generally held to be heavy-tailled \[[17](#bib.bib17)\], that is they follow a distribution such that
<table>
<tbody>
<tr>
<td></td>
<td>ℙ[Length of flow>x]∼x−α,similar-toℙdelimited-[]Length of flow𝑥superscript𝑥𝛼\mathbb{P}\left[\text{Length of flow}>x\right]\sim x^{-\alpha},</td>
<td></td>
</tr>
</tbody>
</table>
where $`\alpha \in {(0,2)}`$ and $`\sim`$ means asymptotically proportional to as $`x\rightarrow\infty`$. This means that it is not sufficient simply to look at the flows under a given length, extremely long flows will also play an important part in the make-up of the traffic.
## 2 Methodology
Our results are based on a 30 minute long trace from an OC-48 link on the CAIDA \[[18](#bib.bib18)\] network on $`24`$th April 2003. The trace contains 47,047,240 packets from which an average 83% are TCP, 7% are UDP. The rest are usually other network layer protocols, such as ICMP.
The sampling strategies used in this paper are referred to as
1. 1\.
packet sampling,
2. 2\.
sample-and-hold (by byte),
3. 3\.
sample-and-hold (by packet) and
4. 4\.
sample-and-hold (by SYN).
Sample-and-hold (by byte) is the original sample-and-hold technique developed in \[[6](#bib.bib6)\]. Packet sampling as has been previously described, is the commonly used technique of sampling each packet in an independent manner with a given probability $`p`$. This can be contrasted with techniques which are also commonly used whereby for a given $`n`$, every $`n`$th packet is sampled. The three sample-and-hold techniques are described in the next section.
### 2.1 Sample-and-hold techniques
Sample-and-hold (by byte) is a sampling technique developed in Estan-Varghese \[[6](#bib.bib6)\]. In this technique the router keeps track of certain flows and samples every packet on these flows until their expiry. The technique was developed with the aim of producing a sampling method in which flows that carry greater traffic volume (sometimes called “elephant” flows) are more likely to be ampled than smaller flows. Once a flow is expired, due to one of the reasons discussed in Section [1.2](#S1.SS2 "1.2 Definitions ‣ 1 Introduction ‣ Towards Informative Statistical Flow Inversion"), it is marked for export and kept in the router cache, until a relatively large set of flows is ready for export to an external aggregation point. The process proceeds, in a packet by packet basis, as follows. When a packet is seen which is part of a flow being tracked, that packet is sampled. If the packet is part of a flow which is not being tracked, then there is a probability that this packet will be sampled and the flow will be added to the list of flows being tracked. Let $`b`$ be the length of the packet being considered in bytes. Let $`p`$ be a constant in $`(0,1)`$. The probability of starting to sample this flow at the packet under consideration is $`p_{p} = {1 - {({1 - p})}^{b}}`$. This is equivalent to considering sampling every byte with probability $`p`$.
Sample-and-hold (by packet) is an obvious variant of this technique where the probability of beginning to sample a flow at a given packet is a constant $`p`$. This is equivalent to the technique from Estan–Varghese but with the probability fixed rather than depending on the length in bytes of the packet.
Sample-and-hold (by SYN) is another sample-and-hold variant based on the Transmission Control Protocol (TCP). A valid TCP flow is expected to begin with exactly one packet with the SYN flag set. If a packet is not part of the set of flows being tracked and it has the SYN flag set then, with probability $`p`$, that packet is sampled and that flow added to the list of flows being tracked.
The idea is that this SYN based sampling is as close as possible to a version of the flow-based sampling suggested by Hohn–Veitch \[[3](#bib.bib3)\] which can be implemented without recording every packet and producing flows from them before sampling. Of course, in any given traffic trace, some TCP flows will have their SYN flag before the trace collection started. Other flows may have more than one SYN flag. This was observed previously by Duffield et al \[[2](#bib.bib2)\]. In their packet traces, they determined the proportion of those TCP flows containing at least one SYN packet that contained exactly one SYN packet. For one data set it was 98.8%; for another one it was 94.6%.
In the CAIDA data investigated here, 7% of flows which contained one SYN packet contained at least one other.
It should be noted that not all traffic in the traces analyzed is TCP traffic, and the Sample-and-hold (by SYN) method can only produce an estimate of the distribution of TCP flows. However, we have examined our algorithm on TCP since more than 90% of the traffic in our trace is TCP.
### 2.2 Flow termination dilemma
Memory constraints prevent routers from keeping flows active for long spans of time. The flow lifetime in the router cache is configurable by the user. If sampling is not used, it is impossible to keep the flows in the buffer for more than a few minutes on a heavily utilized router. For example, authors in \[[10](#bib.bib10)\] use a flow expiry timeout of 2 seconds, which they find to be the maximum before flow loss rates reach unacceptable levels. Figures [1](#S2.F1 "Figure 1 ‣ 2.2 Flow termination dilemma ‣ 2 Methodology ‣ Towards Informative Statistical Flow Inversion") and [2](#S2.F2 "Figure 2 ‣ 2.2 Flow termination dilemma ‣ 2 Methodology ‣ Towards Informative Statistical Flow Inversion") show the effects of the buffer size on the accuracy of the flow reports. It can be seen that the longer expiry times consistently help pick out more long flows. Additionally, it can be seen that the distribution obtained using a longer expiry time is more consistent with the straight line graph expected of a heavy-tailed flow distribution, as discussed in Section [1.5](#S1.SS5 "1.5 Related work ‣ 1 Introduction ‣ Towards Informative Statistical Flow Inversion").
[Refer to caption](/html/0705.1939/assets/x1.png)
Figure 1: Changes in the flow buffer can slice up or join flows.
If the flow buffer memory in the router is chosen to be that of length $`t`$, then each section of flow $`F\hspace{0pt}2`$ will be reported as a separate flow, and flow $`F\hspace{0pt}1`$ will also be sliced up into smaller flows, creating so-called sparse flows. However if the router can afford to have a larger time out for the flows ($`T`$ in Figure [1](#S2.F1 "Figure 1 ‣ 2.2 Flow termination dilemma ‣ 2 Methodology ‣ Towards Informative Statistical Flow Inversion")), even if the smaller flows of $`F\hspace{0pt}2`$ are in reality individual, unrelated flows (though very unlikely due to the vast number of source ports available for TCP packets at least), they are reported as a single cumulative flow. Flow $`F\hspace{0pt}1`$ will be correctly reported as a complete flow.
[Refer to caption](/html/0705.1939/assets/x2.png)
Figure 2: Number of observed flows of a given length for two different timeout values on the CAIDA data.
Figure [3](#S2.F3 "Figure 3 ‣ 2.2 Flow termination dilemma ‣ 2 Methodology ‣ Towards Informative Statistical Flow Inversion") displays the complimentary cumulative distribution function (CCDF) of flow lengths on the CAIDA data for two different flow expiry lengths. This shows clearly that, as would be expected, a shorter expiry time reduces the probability that the longer flows can be seen.
[Refer to caption](/html/0705.1939/assets/x3.png)
Figure 3: The complementary cumulative distribution function for flow lengths on the CAIDA data using two different flow expiry times.
### 2.3 Flow construction from trace files
To build flow records out of trace files, we emulated the operation of NetFlow on a general purpose computer, relaxing the real-time memory requirements usually imposed on routers. Thus, we were able to greatly extend the amount of time that detailed flow records are kept in memory, and thus construct a baseline of unsampled measurements with which the results of our inversion procedures could be tested. However, the sheer amount of packets in some high-speed Internet core traces means that we cannot process all of it in one go. To address this, we divide time into *analysis windows*, over which flows are considered independently. For the 30 minute CAIDA datasets, however, we only used one analysis window per trace.
The algorithm we used for the sample-and-hold techniques of Section [1.4](#S1.SS4 "1.4 Practical implementation of sampling techniques ‣ 1 Introduction ‣ Towards Informative Statistical Flow Inversion") is detailed in Algorithm [2.3](#S2.SS3 "2.3 Flow construction from trace files ‣ 2 Methodology ‣ Towards Informative Statistical Flow Inversion") with the variables explained in Table [1](#S2.T1 "Table 1 ‣ 2.3 Flow construction from trace files ‣ 2 Methodology ‣ Towards Informative Statistical Flow Inversion"). The algorithm for packet sampling is similar but simpler, since it does not need to track flows.
Basically, a trace file is explored and each of its packets considered in turn. If the packet belongs to a flow that had been previously selected for sampling, its information is aggregated into the current flow tables. If it is not, its flow is sampled with a probability $`p`$. This probability is calculated on the basis of the sampling technique: in the case of sample-and-hold (by packet), the same $`p`$ is used for all packets, while in sample-and-hold (by byte) the probability of sampling a packet is a function of the packet length. If a given packet is selected, its 5-tuple $`\phi`$ is tracked, so that new packets with this same $`\phi`$ and within the flow expiry timeout $`t_{t}`$ are considered part of the same flow.
There are two conditions that trigger complete flow buffer exports in Algorithm [2.3](#S2.SS3 "2.3 Flow construction from trace files ‣ 2 Methodology ‣ Towards Informative Statistical Flow Inversion"). The first one, implemented by the boolean function FlowBufferFull(), represents the flow buffer reaching its maximum capacity
<sup>1</sup>
<sup>1</sup>1As Algorithm [2.3](#S2.SS3 "2.3 Flow construction from trace files ‣ 2 Methodology ‣ Towards Informative Statistical Flow Inversion") was implemented for emulation, the size of the flow buffer is not a hard parameter, but can be modified as appropriate.
. The other one, corresponding to FlowExportTimerExpired(), represents the expiry of a *flow analysis window*, that is, a periodic event over which the flow collection process is restarted. When a buffer export occurs (independently of its triggering condition) then the system stops tracking all flows and writes the flow statistics to disk. This means that, for every $`\psi \in \Psi`$, the 3-tuple ($`\psi`$, $`T_{p}\hspace{0pt}{(\psi)}`$, $`T_{b}\hspace{0pt}{(\psi)}`$) is written to a file and the relevant data structures in program memory are cleared.
It is informative to consider the difference between the 5-tuple $`\phi`$ and the flow identifier $`\psi`$ in Algorithm [2.3](#S2.SS3 "2.3 Flow construction from trace files ‣ 2 Methodology ‣ Towards Informative Statistical Flow Inversion"). If the buffer export timeout $`t_{w}`$ is significantly longer than $`t_{t}`$, it may be possible to encounter two (or more) different flows on the same 5-tuple $`\phi`$ during the same analysis window (because the flow has been timed out and exported then seen again). Thus, $`\psi`$ appends a discriminating string to $`\phi`$, so that the identity of both flows is maintained. As a result of this procedure, after a flow expires its statistics are no longer updated, and it will be analyzed as a separate entity from other flows on the same $`\phi`$. Of course, by choosing $`t_{t} > t_{w}`$, individual flow expiration can be completely bypassed, and by setting $`t_{w}`$ larger than the time spanned by the trace being analyzed, flow analysis window exports can be completely bypassed as well. This gives Algorithm [2.3](#S2.SS3 "2.3 Flow construction from trace files ‣ 2 Methodology ‣ Towards Informative Statistical Flow Inversion") full flexibility to explore the influence of $`t_{t}`$, $`t_{w}`$ and $`N_{f}`$ on the sampled flow statistics and our proposed inversion techniques.
<table>
<tbody>
<tr>
<th>ϕitalic-ϕ\phi</th>
<td>5-tuple corresponding to packet P𝑃P</td>
</tr>
<tr>
<th>t𝑡t</th>
<td>Capture time of packet P𝑃P</td>
</tr>
<tr>
<th>Nbsubscript𝑁𝑏N_{b}</th>
<td>Number of bytes in packet P𝑃P</td>
</tr>
<tr>
<th>ts(ϕ)subscript𝑡𝑠italic-ϕt_{s}(\phi)</th>
<td>Trace time since the last packet on 5-tuple ϕitalic-ϕ\phi was seen</td>
</tr>
<tr>
<th>ttsubscript𝑡𝑡t_{t}</th>
<td>Flow expiry timeout</td>
</tr>
<tr>
<th>ψ𝜓\psi</th>
<td>Flow identification number</td>
</tr>
<tr>
<th>ΨΨ\Psi</th>
<td>Set of all ψ𝜓\psi</td>
</tr>
<tr>
<th>Tp(ψ)subscript𝑇𝑝𝜓T_{p}(\psi)</th>
<td>Total number of packets in flow ψ𝜓\psi</td>
</tr>
<tr>
<th>Tb(ψ)subscript𝑇𝑏𝜓T_{b}(\psi)</th>
<td>Total number of bytes in flow ψ𝜓\psi</td>
</tr>
<tr>
<th>p𝑝p</th>
<td>Probability of starting to follow flow ϕitalic-ϕ\phi</td>
</tr>
<tr>
<th>twsubscript𝑡𝑤t_{w}</th>
<td>Flow buffer export timeout</td>
</tr>
<tr>
<th>Nfsubscript𝑁𝑓N_{f}</th>
<td>Flow buffer size in records</td>
</tr>
</tbody>
</table>
Table 1: Variables for the flow construction algorithm
Algorithm 2.1: BuildFlows
($`t\hspace{0pt}r\hspace{0pt}a\hspace{0pt}c\hspace{0pt}e`$)
$`\begin{matrix}
{{\text{while~}\text{packetsLeft}}\hspace{0pt}{({t\hspace{0pt}r\hspace{0pt}a\hspace{0pt}c\hspace{0pt}e})}} & \\
& \\
{{\ \text{do~}}\hspace{0pt}\left\{ \begin{matrix}
{P\leftarrow{\text{ReadPacket}\hspace{0pt}{({t\hspace{0pt}r\hspace{0pt}a\hspace{0pt}c\hspace{0pt}e})}}} & \\
{{(\phi,t,N_{b})}\leftarrow{\text{DecodePacket}\hspace{0pt}{(P)}}} & \\
& \\
{{\text{if~}\text{FlowIsBeingTracked}}\hspace{0pt}{(\phi)}} & \\
{\ \text{then~}} & \\
\left\{ \begin{matrix}
{\text{comment:}\hspace{0pt}\text{Has the flow expired?}} & \\
{\text{if~}\hspace{0pt}{({{t_{s}\hspace{0pt}{(\phi)}} > t_{t}})}} & \\
\left\{ \begin{array}{lc}
{\psi\leftarrow{\text{GetFlowID}\hspace{0pt}{(\phi)}}} & \\
{\text{TerminateFlow}\hspace{0pt}{(\psi)}} & \\
{\psi\leftarrow{\text{CreateFlow}\hspace{0pt}{(\phi)}}} &
\end{array} \right. & \\
& \\
{{t_{s}\hspace{0pt}{(\phi)}}\leftarrow t} & \\
{{T_{p}\hspace{0pt}{(\psi)}}\leftarrow{{T_{p}\hspace{0pt}{(\psi)}} + 1}} & \\
{{T_{b}\hspace{0pt}{(\psi)}}\leftarrow{{T_{b}\hspace{0pt}{(\psi)}} + N_{b}}} &
\end{matrix} \right. & \\
{\ \text{else~}} & \\
\left\{ \begin{matrix}
{\text{comment:}\hspace{0pt}\text{Is the flow going to be sampled?}} & \\
{{\text{if~}\text{flowSelectedForSampling}}\hspace{0pt}{(p,N_{b})}} & \\
& \\
{{\ \text{then~}}\hspace{0pt}\left\{ \begin{array}{lc}
{{t_{s}\hspace{0pt}{(\phi)}}\leftarrow t} & \\
{\psi\leftarrow{\text{CreateFlow}\hspace{0pt}{(\phi)}}} & \\
{\Psi\leftarrow\psi} & \\
{{T_{p}\hspace{0pt}{(\psi)}}\leftarrow 1} & \\
{{T_{b}\hspace{0pt}{(\psi)}}\leftarrow N_{b}} &
\end{array} \right.} &
\end{matrix} \right. & \\
& \\
{{\text{if~}\text{FlowBufferFull}}\hspace{0pt}{({|\Psi|},N_{f})}} & \\
{{{\ \text{or~}}\text{FlowExportTimerExpired}}\hspace{0pt}{(t,t_{w})}} & \\
{{\ \text{then~}}\hspace{0pt}\left\{ \begin{array}{lc}
{\text{ExportFlowBuffer}\hspace{0pt}{()}} & \\
{\text{ResetFlowBuffer}\hspace{0pt}{()}} &
\end{array} \right.} &
\end{matrix} \right.} &
\end{matrix}`$
For the rest of this paper we use a relatively large timeout $`t_{t}`$ of five minutes for the flows. Even though this may be longer than the value usually applied in routers (of around fifteen to thirty seconds) it helps avoid unnecessary flow splitting. In this paper we use $`t_{t} = t_{w} = 5`$ minutes, so that all flow information is reset every 5 minutes. After this is done, a secon post-processing step is done where the output of this process is integrated as a single 30 miutes long file. This is done to reduce memory consumption while avoiding dropping flows due to lack of buffer space.
### 2.4 Inverting the sampled data
Two methods for inverting packet sampled data are given by Hohn and Veitch \[[3](#bib.bib3)\]. These techniques, while mathematically sound, are problematic in realistic cases. In particular, they are numerically unstable when estimating longer flows or rates of sampling with small values of $`p`$ (where $`p`$ is significantly less than $`1/2`$). No results are presented here for inverting packet sampling but for an excellent discussion of the problem, the reader is referred to \[[3](#bib.bib3)\].
Inverting the sample-and-hold (by packet or by byte) is, on the other hand, a new problem.
For each flow which is not being sampled, there is a per-packet probability $`p`$ that the flow will start being sampled at that point. Define $`q = {1 - p}`$. Let $`(\theta_{1},\theta_{2},\ldots)`$ be the original flow length distribution before sampling, where $`\theta_{i}`$ is the probability that a randomly chosen flow is exactly $`i`$ packets long.
Let $`\theta_{i}'`$ be the probability that the algorithm would start sampling a randomly chosen stream $`i`$ packets from the end of the stream. This is the probability that $`i`$ packets are sampled. Note that $`\theta_{0}' \neq 0`$.
<table>
<tbody>
<tr>
<td></td>
<td>θi′={∑j=i∞pqj−iθji>0∑j=0∞qjθji=0subscriptsuperscript𝜃′𝑖casessuperscriptsubscript𝑗𝑖𝑝superscript𝑞𝑗𝑖subscript𝜃𝑗𝑖0superscriptsubscript𝑗0superscript𝑞𝑗subscript𝜃𝑗𝑖0\theta^{\prime}_{i}=\begin{cases}\sum_{j=i}^{\infty}pq^{j-i}\theta_{j}&i>0\\ \sum_{j=0}^{\infty}q^{j}\theta_{j}&i=0\end{cases}</td>
<td></td>
</tr>
</tbody>
</table>
Let $`X_{i}`$, $`i \in {\mathbb{N}}`$ be the distribution of flow lengths which can actually be observed. This can be thought of as the distribution $`\theta_{i}'`$ without the probability of zero length flows.
<table>
<tbody>
<tr>
<td></td>
<td>Xisubscript𝑋𝑖\displaystyle X_{i}</td>
<td>=ℙ[Sample length = i|Sample length >0]absentℙdelimited-[]conditionalSample length = iSample length >0\displaystyle=\mathbb{P}\left[\text{Sample length = $i$}|\text{Sample length $>0$}\right]</td>
<td></td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td></td>
<td>=θi′∑k=1∞θk′absentsubscriptsuperscript𝜃′𝑖superscriptsubscript𝑘1subscriptsuperscript𝜃′𝑘\displaystyle=\frac{\theta^{\prime}_{i}}{\sum_{k=1}^{\infty}\theta^{\prime}_{k}}</td>
<td></td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td></td>
<td>=∑j=i∞pqj−iθj∑k=1∞∑j=k∞pqj−kθjabsentsuperscriptsubscript𝑗𝑖𝑝superscript𝑞𝑗𝑖subscript𝜃𝑗superscriptsubscript𝑘1superscriptsubscript𝑗𝑘𝑝superscript𝑞𝑗𝑘subscript𝜃𝑗\displaystyle=\frac{\sum_{j=i}^{\infty}pq^{j-i}\theta_{j}}{\sum_{k=1}^{\infty}\sum_{j=k}^{\infty}pq^{j-k}\theta_{j}}</td>
<td></td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td></td>
<td>=∑j=i∞qjθjqi∑j=1∞qjθj∑k=1jq−kabsentsuperscriptsubscript𝑗𝑖superscript𝑞𝑗subscript𝜃𝑗superscript𝑞𝑖superscriptsubscript𝑗1superscript𝑞𝑗subscript𝜃𝑗superscriptsubscript𝑘1𝑗superscript𝑞𝑘\displaystyle=\frac{\sum_{j=i}^{\infty}q^{j}\theta_{j}}{q^{i}\sum_{j=1}^{\infty}q^{j}\theta_{j}\sum_{k=1}^{j}q^{-k}}</td>
<td></td>
</tr>
</tbody>
</table>
The sum $`\sum_{k = 1}^{j}q^{- k}`$ can be evaluated giving,
<table>
<tbody>
<tr>
<td></td>
<td>Xisubscript𝑋𝑖\displaystyle X_{i}</td>
<td>=(1−q)∑j=i∞qjθjqi∑j=1∞qj(q−j−1)absent1𝑞superscriptsubscript𝑗𝑖superscript𝑞𝑗subscript𝜃𝑗superscript𝑞𝑖superscriptsubscript𝑗1superscript𝑞𝑗superscript𝑞𝑗1\displaystyle=\frac{(1-q)\sum_{j=i}^{\infty}q^{j}\theta_{j}}{q^{i}\sum_{j=1}^{\infty}q^{j}(q^{-j}-1)}</td>
<td></td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td></td>
<td>=(1−q)∑j=i∞qjθjqi∑j=1∞(1−qj)θjabsent1𝑞superscriptsubscript𝑗𝑖superscript𝑞𝑗subscript𝜃𝑗superscript𝑞𝑖superscriptsubscript𝑗11superscript𝑞𝑗subscript𝜃𝑗\displaystyle=\frac{(1-q)\sum_{j=i}^{\infty}q^{j}\theta_{j}}{q^{i}\sum_{j=1}^{\infty}(1-q^{j})\theta_{j}}</td>
<td></td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td></td>
<td>=(1−q)∑j=i∞qjθjqi[1−∑j=1∞qjθj].absent1𝑞superscriptsubscript𝑗𝑖superscript𝑞𝑗subscript𝜃𝑗superscript𝑞𝑖delimited-[]1superscriptsubscript𝑗1superscript𝑞𝑗subscript𝜃𝑗\displaystyle=\frac{(1-q)\sum_{j=i}^{\infty}q^{j}\theta_{j}}{q^{i}[1-\sum_{j=1}^{\infty}q^{j}\theta_{j}]}.</td>
<td></td>
</tr>
</tbody>
</table>
Setting $`i = 1`$ and rearranging gives
<table>
<tbody>
<tr>
<td></td>
<td>∑j=1∞qjθj=qX11−q+qX1.superscriptsubscript𝑗1superscript𝑞𝑗subscript𝜃𝑗𝑞subscript𝑋11𝑞𝑞subscript𝑋1\sum_{j=1}^{\infty}q^{j}\theta_{j}=\frac{qX_{1}}{1-q+qX_{1}}.</td>
<td></td>
</tr>
</tbody>
</table>
Let $`C = {{({1 - q})}/{({1 - \frac{q\hspace{0pt}X_{1}}{{1 - q} + {q\hspace{0pt}X_{1}}}})}} = {({{1 - q} + {q\hspace{0pt}X_{1}}})}`$ and therefore
<table>
<tbody>
<tr>
<td></td>
<td>qiXi=C∑j=i∞qjθj.superscript𝑞𝑖subscript𝑋𝑖𝐶superscriptsubscript𝑗𝑖superscript𝑞𝑗subscript𝜃𝑗q^{i}X_{i}=C\sum_{j=i}^{\infty}q^{j}\theta_{j}.</td>
<td></td>
</tr>
</tbody>
</table>
Giving the final estimate
<table>
<tbody>
<tr>
<td></td>
<td>θisubscript𝜃𝑖\displaystyle\theta_{i}</td>
<td>=Xi−qXi+1Cabsentsubscript𝑋𝑖𝑞subscript𝑋𝑖1𝐶\displaystyle=\frac{X_{i}-qX_{i+1}}{C}</td>
<td></td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td></td>
<td>=Xi−qXi+11−q+qX1.absentsubscript𝑋𝑖𝑞subscript𝑋𝑖11𝑞𝑞subscript𝑋1\displaystyle=\frac{X_{i}-qX_{i+1}}{1-q+qX_{1}}.</td>
<td></td>
</tr>
</tbody>
</table>
This method has certain obvious weaknesses. The factor $`{1 - q} + {q\hspace{0pt}X_{1}}`$ is simply a normalization factor, the method wholly relies on the difference between $`X_{i}`$ and $`X_{i + 1}`$. It is relatively insensitive to the particular value of $`p`$ when $`p`$ is near zero (which it would be for typical sampling rates) since the difference between $`X_{i} - {0.99\hspace{0pt}X_{i + 1}}`$ and $`X_{i} - {0.999\hspace{0pt}X_{i + 1}}`$ is usually not great. However, particularly at large flows this creates problems. In particular, if $`X_{i + 1} > X_{i}`$ then the method will produce a negative estimate for the probability. This problem can be offset to some extent by pooling adjacent estimates so that, instead of estimating the probability that a flow has exactly length $`i`$, instead an estimate is given of the probability that the flow has a length in some range $`i,{i + 1},\ldots,{i + n}`$. This is discussed in the next section.
We did not find any obvious method of inverting the original Estan sample-and-hold (by byte). The method used in this paper is simply to assume that the data was obtained from sample-and-hold by packet with $`p`$ as the probability of sampling a packet of mean packet length using Estan’s method.
For SYN based sampling, the assumption that each TCP flow begins with exactly one SYN flag implies that no inversion should be needed. Unfortunately, the SYN based sampling will only sample TCP flows and can provide no information about the distribution of UDP flows. This is a weakness of the method.
### 2.5 Logarithmic binning
When examining the flow distribution, particularly for long flows, it is likely to be of more interest to know how many flows have a length in a given range, rather than the number of flows with a specific length. Therefore, we have used a pooling technique to average data using a logarithmic scale. The data relating to flow lengths is averaged over bins which contain data on a set of flow lengths (for example, one bin from all lengths from 1000 to 1100). The size of the bins are chosen so that they have a constant width (or as nearly as possible given they are integer valued) on a logarithmic scale. This technique is sometimes known as logarithmic binning.
Logarithmic binning is a simple way of smoothing sample data which is distributed on a logscale. Let $`x_{k}`$ be the number of observations of a flow with length $`k`$. Let $`m`$ be the largest flow observed. The valued $`x_{1},x_{2},{\ldots\hspace{0pt}x_{m}}`$ will be combined into $`n < m`$ observations $`X_{1},X_{2},\ldots,X_{n}`$. Now, let $`i_{0},i_{1},i_{2},\ldots,i_{n}`$ be some series of integers such that $`i_{0} < i_{1} < {i_{2}\hspace{0pt}\ldots}`$ with $`i_{0} = 1`$, $`i_{n}`$ is larger than the largest flow length observed and $`i_{k + 1}/i_{k}`$ is approximately constant for large $`k`$. Now we can derive a series $`X_{1},X_{2},\ldots,X_{n}`$ giving the average number of observations in the range $`\lbrack i_{k - 1},i_{k})`$ – note that the integer $`i_{k}`$ is not in this range (but will be in the range of $`X_{k + 1}`$.
<table>
<tbody>
<tr>
<td></td>
<td>Xk=∑j=ik−1ik−1xjik−ik−1,subscript𝑋𝑘superscriptsubscript𝑗subscript𝑖𝑘1subscript𝑖𝑘1subscript𝑥𝑗subscript𝑖𝑘subscript𝑖𝑘1X_{k}=\frac{\sum_{j=i_{k-1}}^{i_{k}-1}x_{j}}{i_{k}-i_{k-1}},</td>
<td></td>
</tr>
</tbody>
</table>
for $`k = {1,2,\ldots,n}`$. Note that, for display purposes, it makes sense to show the observation $`X_{k}`$ as occurring in the range $`i_{k - 1} - 0.5`$ to $`i_{k} - 0.5`$.
Figure [4](#S2.F4 "Figure 4 ‣ 2.5 Logarithmic binning ‣ 2 Methodology ‣ Towards Informative Statistical Flow Inversion") shows the results of logarithmic binning on one of the data sets from Figure [2](#S2.F2 "Figure 2 ‣ 2.2 Flow termination dilemma ‣ 2 Methodology ‣ Towards Informative Statistical Flow Inversion"). The technique has two advantages for this study. Firstly, it produces clearer information for large flows. In the large flows regime, we usually observe either one flow or no flows, and this can make the graph on that region harder to interpret. However, the logarithmic binning allows the graph to convey information about how many flows are in a given range, including those long-flow regimes where simple plots are usually uninformative. This also gives a clearer idea of the heavy-tailed nature of flow lengths. Secondly, it pools those estimators for which there is most uncertainty. Estimating the probability that there is a flow exactly (say) 10,005 packets long is a difficult task, requiring vast amounts of data and processing power. On the other hand, estimating the expected number of flows which are between 10,000 and 11,000 packets long allows the pooling of estimators to produce a more accurate estimate using a smaller data sets, and with lesser computational demands.
[Refer to caption](/html/0705.1939/assets/x4.png)
Figure 4: The effects of logarithmic binning.
[Refer to caption](/html/0705.1939/assets/x5.png)
Figure 5: Figure [2](#S2.F2 "Figure 2 ‣ 2.2 Flow termination dilemma ‣ 2 Methodology ‣ Towards Informative Statistical Flow Inversion") replotted with logarithmic binning.
## 3 Results
The results in this paper are all obtained on real network trace data. The traces are sampled using the techniques described in the previous section. The flow distribution is then produced on the sampled data after sampling inversion techniques are applied (where such techniques are available) in order to recreate the original flow distribution. This is compared with the correct flow distribution obtained from the unsampled data. In order to assist comparison, the sampling rates are chosen so that approximately one packet in every one hundred is sampled. That is, the methods used are all sampling approximately the same percentage of the data set and the storage requirements for each sampling method would not be dissimilar.
The logarithmic binning method is a simple but invaluable tool for the investigation of sampled flow length distribution. In addition to being a useful method for presenting the data it enables the pooling of otherwise unreliable estimates to get a reliable estimate over a range of values.
Packet sampling is an attractive sampling scheme for many purposes. It allows recovery of many important properties of the data, however, it is difficult to recover flow based information. Three sample-and-hold based schemes are used here, based upon the original sample-and-hold described by Estan and Varghese \[[6](#bib.bib6)\] (which is here referred to as sample-and-hold by byte). Like packet sampling, sample-and-hold can be implemented in a practical setting (for example in firmware) \[[19](#bib.bib19)\].
### 3.1 Packet sampling
As previously stated, inversion results are not given here for packet based sampling. This is due to the extreme difficulty of producing a flow distribution over the full range of possible flow lengths from packet sampled data (see the discussion in Section [1.5](#S1.SS5 "1.5 Related work ‣ 1 Introduction ‣ Towards Informative Statistical Flow Inversion")). The sampling was performed to get one packet in one hundred by setting $`p = 0.01`$. This gives $`425,014`$ packets sampled in $`207,126`$ flows, a mean of $`2.1`$ packets per flow.
[Refer to caption](/html/0705.1939/assets/x6.png)
Figure 6: The impact of packet sampling with $`p = 0.01`$ on the flow distribution.
### 3.2 Sample-and-hold (by packet)
The value of $`p`$ was adjusted so that approximately one packet in every one hundred was sampled. The value of $`p`$ used was $`0.000014`$ and this gave $`413,702`$ packets and $`614`$ flows (a mean flow length of 674 packets per flow).
[Refer to caption](/html/0705.1939/assets/x7.png)
Figure 7: Cumulative density function for packet based sample-and-hold with sampling of approximately one packet in every one hundred on the CAIDA data.
[Refer to caption](/html/0705.1939/assets/x8.png)
Figure 8: Density function for packet based sample-and-hold with sampling of approximately one packet in every one hundred on the CAIDA data.
At a higher sampling rate the inversion algorithm can be seen to be very good indeed. With $`p = 0.001`$ I $`10,333,134`$ of $`47,047,240`$ packets were sampled on the CAIDA trace. This is a very high rate of sampling but suitable for an initial test of the sampling algorithm.
Figure [9](#S3.F9 "Figure 9 ‣ 3.2 Sample-and-hold (by packet) ‣ 3 Results ‣ Towards Informative Statistical Flow Inversion") shows the density function for this experiment before and after inversion compared with the unsampled data. Figure [10](#S3.F10 "Figure 10 ‣ 3.2 Sample-and-hold (by packet) ‣ 3 Results ‣ Towards Informative Statistical Flow Inversion") shows the distribution function for the same experiment.
[Refer to caption](/html/0705.1939/assets/x9.png)
Figure 9: Density function for packet based sample-and-hold with $`p = 0.001`$ on the CAIDA data.
[Refer to caption](/html/0705.1939/assets/x10.png)
Figure 10: Cumulative density function for packet based sample-and-hold with $`p = 0.001`$ on the CAIDA data.
The sample-and-hold by packet method focuses on large flows. At the high sampling rate in Figures [10](#S3.F10 "Figure 10 ‣ 3.2 Sample-and-hold (by packet) ‣ 3 Results ‣ Towards Informative Statistical Flow Inversion") and [9](#S3.F9 "Figure 9 ‣ 3.2 Sample-and-hold (by packet) ‣ 3 Results ‣ Towards Informative Statistical Flow Inversion") the sample-and-hold by packet method inverts almost precisely to the original distribution. However, nearly one in five packets were sampled and this is an unrealistically high sampling rate for a highly loaded router. At the more realistic sampling rates shown in Figures [7](#S3.F7 "Figure 7 ‣ 3.2 Sample-and-hold (by packet) ‣ 3 Results ‣ Towards Informative Statistical Flow Inversion") and [8](#S3.F8 "Figure 8 ‣ 3.2 Sample-and-hold (by packet) ‣ 3 Results ‣ Towards Informative Statistical Flow Inversion") the algorithm still performs relatively well, particularly at higher flow lengths. The distribution here was recovered from only 614 sampled flows. Another potentially useful property of the sample-and-hold by packet is that the packets sampled can be resampled to create a sample which would have been obtained from packet sampling with probability $`p`$ (where this is the same $`p`$ used for sample-and-hold by packet in the first place). This is done by sampling the first packet in each flow and then performing packet sampling with probability $`p`$ on all subsequent packets.
### 3.3 Sample-and-hold (by byte)
Figure [11](#S3.F11 "Figure 11 ‣ 3.3 Sample-and-hold (by byte) ‣ 3 Results ‣ Towards Informative Statistical Flow Inversion") shows the results from sampling the CAIDA data using the sample-and-hold (by byte) method as proposed by Estan and Varghese \[[6](#bib.bib6)\]. The $`p`$ value has been tuned so that approximately one in one hundred packets are sampled. This gave $`521,337`$ packets sampled in total over $`527`$ flows, a mean flow length of $`989`$ packets per flow.
[Refer to caption](/html/0705.1939/assets/x11.png)
Figure 11: Cumulative density function for byte based sample-and-hold with sampling of approximately one packet in every one hundred on the CAIDA data.
Sample-and-hold by byte was engineered to focus on the largest flows (sometimes referred to in the literature as “elephant” flows). It is no surprise that the flow distribution in Figure [11](#S3.F11 "Figure 11 ‣ 3.3 Sample-and-hold (by byte) ‣ 3 Results ‣ Towards Informative Statistical Flow Inversion") shows this clearly. The largest flows are tracked. The inversion algorithm in this paper was designed to work with the packet based sample-and-hold and it is no surprise that it does not perform particularly well in this case. As discussed, negative values occur in the predicted “probabilities” and the “distribution” does not total to one as it should. This is a result of applying an algorithm which is not quite appropriate for the data set. Of course these issues could be fixed by forcing a minimum of zero and normalizing the distribution artificially. Nonetheless, the inverted distribution is an improvement over the original at longer flow lengths although performs poorly over short flows. The distribution in Figure [11](#S3.F11 "Figure 11 ‣ 3.3 Sample-and-hold (by byte) ‣ 3 Results ‣ Towards Informative Statistical Flow Inversion") is reconstructed from only 527 flows so it is, perhaps, impressive that it is as close to the original as it is.
The advantages of sample-and-hold by byte are that it focuses clearly on those “elephant” flows which can dominate traffic. It has been previously studied in the literature and implemented in software for real sampling applications. On the other hand, the disadvantages are that no good inversion algorithm exists as yet. In addition there is no obvious way to recover a packet sampled data set from the sample-and-hold by byte data.
### 3.4 Sample-and-hold (by SYN)
Sample and hold by SYN was run with $`p`$ tuned to get approximately one packet in every one hundred. If the assumption of one SYN packet per flow was correct this would simply mean setting $`p = 0.01`$ but, as previously discussed, this assumption is not met in the real data. In this sample, $`520,116`$ packets were sampled in $`68,618`$ flows, a mean of $`7.6`$ packets per flow.
[Refer to caption](/html/0705.1939/assets/x12.png)
Figure 12: Cumulative density function for SYN based sample-and-hold with sampling of approximately one packet in every one hundred on the CAIDA data.
[Refer to caption](/html/0705.1939/assets/x13.png)
Figure 13: Cumulative density function for SYN based sample-and-hold with sampling of approximately one packet in every one hundred on the CAIDA data.
The sample-and-hold by SYN is actually surprisingly good at recovering the sampled distribution as can be seen in Figures [12](#S3.F12 "Figure 12 ‣ 3.4 Sample-and-hold (by SYN) ‣ 3 Results ‣ Towards Informative Statistical Flow Inversion") and [13](#S3.F13 "Figure 13 ‣ 3.4 Sample-and-hold (by SYN) ‣ 3 Results ‣ Towards Informative Statistical Flow Inversion"). However, this is somewhat misleading. As can be seen if Figure [13](#S3.F13 "Figure 13 ‣ 3.4 Sample-and-hold (by SYN) ‣ 3 Results ‣ Towards Informative Statistical Flow Inversion"), the distribution from all SYN flows (effectively SYN sampling at a rate of 1) is, in fact, very different from the distribution of all the flows. This is not unexpected. It seems that in this trace the non SYN flows (mostly UDP and some ICMP) are shorter and, in particular there are more flows which are just one or two packets long.
It has already been noted that SYN sampling does not provide an unbiased estimate of SYN flows because, in reality, flows can have more than one SYN packet. Many of these flows with more than one SYN packet are short flows (perhaps because a flow with multiple SYN packets can result from trying to initiate a connection to a machine which is not responding). When SYN sample-and-hold is used then it will be more likely to sample those flows with multiple SYN packets (unless the sample rate is one, of course). The presence of such a protocol behaviour has fortuitously cancelled out the error in the other direction and the good recovery of the flow distribution is a product of two errors in opposite directions cancelling rather than a true measure of the success of the algorithm. This gives a large element of uncertainty to the use of SYN sampling as a method for recovering flow distributions since the basic assumption (that TCP flows begin with a single SYN packet) is not met in real data.
A major disadvantage of sample-and-hold by SYN packet is that it cannot provide information about non TCP packets. A major advantage is that it gives an approximate flow distribution with no need for inversion because it is an approximation to flow sampling.
## 4 Conclusions and future work
Producing flow distributions for sampled packet traces is a difficult problem. Several authors have approached the problem of producing flow distributions from traces sampled using standard packet sampling. However, different sampling methods can be used to provide a sample which makes the recovery of the flow distribution easier while at the same time not putting an undue requirement on memory and storage for the hardware performing the sampling.
Of the sample-and-hold based methods studied here all have advantages and disadvantages. The byte based sample-and-hold originally proposed in \[[6](#bib.bib6)\] was intended to focus on the longest flows and does this better than any other method. An inversion to recover the original flow distribution has been attempted in this paper and is partially successful. Further work in this area might improve this algorithm.
Packet based sample-and-hold has two advantages. Firstly, it can be inverted well to produce a reasonable estimate of the flow distribution even for relatively low sampling rates (approximately one in one hundred packets sampled). Secondly, it can be resampled to get a packet sample and recover those quantities which can be measured at the packet, not flow, level. A disadvantage is that the estimated probabilities are not guaranteed positive.
SYN based sample-and-hold is near to the original flow-based sampling proposed by Hohn and Veitch \[[3](#bib.bib3)\] but has problems due to packets with more than one SYN flag. It is possible that further work could correct for this problem. However, the problem that this technique will only ever be useful for TCP traffic remains a major issue.
Our future work on this topic will focus on two main issues. Firstly, there is more information to be gained from other parts of the TCP header, notably the authors know of no results which use the FIN or RST flag for inversion. While these are problematic since not every flow terminates correctly, still it would seem that valuable information is contained in these flags. Secondly, using multiple sample sources could be a rich topic for research. Some work has already been done in this area: \[[1](#bib.bib1), Section 6\] provides a start on this topic focusing on packet sampling and \[[20](#bib.bib20)\] provides another approach. Investigating different sampling techniques which might take advantage of network topology (for example, if samples are available from two directions on the same link) could provide more information which might be used to develop better sampling techniques and also to provide more information for the inversion problem.
## Acknowledgments
The authors would like to acknowledge CAIDA \[[18](#bib.bib18)\] for providing the trace files. This work is conducted under the MASTS project (EPSRC grant GR/T10503).
## References
- \[1\] Bruno Ribeiro, Don Towsley, Tao Ye, and Jean Bolot. Fisher information of sampled packets: an application to flow size estimation.
In IMC ’06: Proceedings of the 6th ACM SIGCOMM on Internet measurement, pages 15–26, New York, NY, USA, 2006. ACM Press.
- \[2\] Nick Duffield, Carsten Lund, and Mikkel Thorup. Estimating flow distributions from sampled flow statistics.
IEEE/ACM Trans. Netw., 13(5):933–946, 2005.
- \[3\] Nicolas Hohn and Darryl Veitch. Inverting sampled traffic.
In IMC ’03: Proceedings of the 3rd ACM SIGCOMM conference on Internet measurement, pages 222–233, New York, NY, USA, 2003. ACM Press.
- \[4\] Baek-Young Choi and Supratik Bhattacharyya. Observations on cisco sampled netFlow.
SIGMETRICS Perform. Eval. Rev., 33(3):18–23, 2005.
- \[5\] Hamed Haddadi, Raul Landa, Miguel Rio, and Saleem Bhatti. Revisiting the issues on NetFlow sample and export performance, 2007.
arxiv.org/abs/0704.0730.
- \[6\] Cristian Estan and George Varghese. New directions in traffic measurement and accounting.
In SIGCOMM ’02: Proceedings of the 2002 conference on Applications, technologies, architectures, and protocols for computer communications, pages 323–336, New York, NY, USA, 2002. ACM Press.
- \[7\] Cristian Estan and George Varghese. New directions in traffic measurement and accounting: Focusing on the elephants, ignoring the mice.
ACM Trans. Comput. Syst., 21(3):270–313, 2003.
- \[8\] Cisco IOS NetFlow.
www.cisco.com.
- \[9\] Haakon Ringberg, Augustin Soule, Jennifer Rexford, and Christophe Diot. Sensitivity of PCA for traffic anomaly detection. Technical Report CR-PRL-2006-11-0001, Thomson lab, 2006.
- \[10\] Daniela Brauckhoff, Bernhard Tellenbach, Arno Wagner, Martin May, and Anukool Lakhina. Impact of packet sampling on anomaly detection metrics.
In IMC ’06: Proceedings of the 6th ACM SIGCOMM on Internet measurement, pages 159–164, New York, NY, USA, 2006. ACM Press.
- \[11\] NetFlow performance analysis.
www.cisco.com/en/US/tech/tk812/
technologies_white_paper0900aecd802a0eb9
.shtml.
- \[12\] Matthew Roughan. A comparison of Poisson and uniform sampling for active measurements.
IEEE Journal on Selected Areas in Communications, 24(12):2299–2312, 2006.
- \[13\] Robin Sommer and Anja Feldmann. NetFlow: information loss or win?
In IMW ’02: Proceedings of the 2nd ACM SIGCOMM Workshop on Internet measurment, pages 173–174, New York, NY, USA, 2002. ACM Press.
- \[14\] Chadi Barakat, Gianluca Iannaccone, and Christophe Diot. Ranking flows from sampled traffic.
In CoNEXT’05: Proceedings of the 2005 ACM conference on Emerging network experiment and technology, pages 188–199, New York, NY, USA, 2005. ACM Press.
- \[15\] N. Duffield, C. Lund, and M. Thorup. Learn more, sample less: Control of volume and variance in network measurement. Information Theory, IEEE Transactions on, Vol.51, Iss.5, May 2005 Pages: 1756- 1775.
- \[16\] Cristian Estan, Ken Keys, David Moore, and George Varghese. Building a better NetFlow.
In SIGCOMM ’04: Proceedings of the 2004 conference on Applications, technologies, architectures, and protocols for computer communications, pages 245–256, New York, NY, USA, 2004. ACM Press.
- \[17\] W. Willinger, M. S. Taqqu, R. Sherman, and D. V. Wilson. Self-similarity through high-variability: statistical analysis of Ethernet LAN traffic at the source level.
IEEE/ACM Trans. on Networking, 5(1):71–86, 1997.
- \[18\] Cooperative Association for Internet Data Analysis (CAIDA).
www.caida.org.
- \[19\] Flowmon website.
www.flowmon.org/flowmon-probe/flowmon-fw.
- \[20\] Gion Reto Cantieni, Gianluca Iannaccone, Chadi Barakat, Christophe Diot, and Patrick Thiran. Reformulating the monitor placement problem: Optimal network-wide sampling. 2005. Intel Research Technical Report.
\balancecolumns |
0705.3740 | cs.CR | cyber | # Optimal Iris Fuzzy Sketches
J. Bringer, H. Chabanne, G. Cohen, B. Kindarji and G. Zémor J. Bringer and H. Chabanne are with Sagem Défense Sécurité, Eragny, France.G. Cohen is with ENST, Département Informatique et Réseaux, Paris, France.G. Zémor is with Institut de Mathématiques de Bordeaux, Université de Bordeaux I, Bordeaux, France.
###### Abstract
Fuzzy sketches, introduced as a link between biometry and cryptography, are a way of handling biometric data matching as an error correction issue. We focus here on iris biometrics and look for the best error-correcting code in that respect. We show that two-dimensional iterative min-sum decoding leads to results near the theoretical limits. In particular, we experiment our techniques on the Iris Challenge Evaluation (ICE) database and validate our findings.
Keywords. Iris, biometry, fuzzy sketches, min-sum decoding.
## 1 Introduction
Fuzzy Sketches have been introduced to handle differences occurring between two captures of biometric data, viewed as errors over a codeword. Many papers give applications of this technique for cryptographic purposes \[[1](#bib.bib1), [3](#bib.bib3), [2](#bib.bib2), [6](#bib.bib6), [7](#bib.bib7), [9](#bib.bib9), [17](#bib.bib17)\] but only a few investigate what are the best codes for this decoding problem, e.g. \[[8](#bib.bib8)\], and how to find them. This issue is addressed here.
### 1.1 Biometric matching and errors correction
Typically, a biometric-based recognition scheme consists of two phases: The enrollment phase where a biometric template $`b`$ is measured from a user $`U`$ and then registered in a token or a database. The second phase – the verification – captures a new biometric sample $`b'`$ from $`U`$ and compares it to the reference data via a matching function. According to some underlying measure $`\mu`$ and some recognition threshold $`\tau`$, $`b'`$ will be accepted as a biometric measure of $`U`$ if $`{\mu\hspace{0pt}{(b,b')}} \leq \tau`$, else rejected. Mainly two kinds of errors are associated to this scheme: False Reject (FR), when a matching user, i.e. a legitimate user, is rejected; False Acceptance (FA), when a non-matching one, e.g. an impostor, is accepted.
Note that, when the threshold increases, the FR’s rate (FRR) decreases while the FA’s rate (FAR) grows, and conversely.
Our methods will resort to information theory and coding. For more background, notation and classical results, the reader is refered to \[[4](#bib.bib4)\] and \[[11](#bib.bib11)\] in these two fields respectively.
Assuming that the templates live in the Hamming space $`\mathcal{H} = {\{ 0,1\}}^{n}`$ equipped with the Hamming distance $`d_{\mathcal{H}}`$, the main idea of fuzzy sketches, as introduced in \[[9](#bib.bib9)\], is to convert the matching step into an error-correcting one. Let $`C`$ be an error-correcting code included in $`\mathcal{H}`$:
- •
During the enrollment phase, one stores $`z = {c \oplus b}`$, where $`c`$ is a random codeword in $`C`$,
- •
During the verification phase, one tries to correct the corrupted codeword $`{z \oplus b'} = {c \oplus {({b \oplus b'})}}`$. Note that when the Hamming distance $`d_{\mathcal{H}}\hspace{0pt}{(b,b')}`$ is small, recovering $`c`$ from $`c \oplus {({b \oplus b'})}`$ is, in principle, possible.
The correction capacity of $`C`$ may thus be equal to $`\tau`$ if we do not want to alter the FRR and the FAR of the system. Unfortunately, the difference between two measures of one biometric source can be very important, whereas the correction capacity of a code is structurally constrained. Moreover, the size of the code should not be too small, to prevent $`z`$ from revealing too much information about the template $`b`$.
### 1.2 Organization of this work
In a first part, we look for theoretical limits. We first modelize our problem with a binary erasure-and-error channel. Given a database of biometric data, we then give a method for finding an upper bound on the underlying error correction capacity.
In a second part, restricting ourselves to iris biometric data and illustrating our method with iterative min-sum decoding of product codes, we provide parameters that put our performances close to the theoretical limit.
## 2 Model
We consider two separate channels with a noise model based on the differences between any two biometric templates.
- •
The first channel, called the matching channel, is generated by errors $`b \oplus b'`$ where $`b`$ and $`b'`$ come from the same user $`U`$.
- •
The second channel, the non-matching channel, is generated by errors where $`b`$ and $`b'`$ come from different biometric sources.
In a practical biometric system, the number of errors in the matching channel is on average lower than in the non-matching channel.
Moreover, the templates are not restricted to a constant length. Indeed, when a sensor captures biometric data, we want to keep the maximum quantity of information but it is rarely possible to capture the same amount of data twice – for instance an iris may be occulted by eyelids – hence the templates are of variable length. This variability can be smoothed by forming a list of erasures, i.e. the list of coordinates where they occur. More precisely, in coding theory, an erasure in the received message is an unknown symbol at a known location. We thus have an erasure-and-error decoding problem on the matching channel. Simultaneously, to keep the FAR low, we want a decoding success to be unlikely on the non-matching channel : to this end we impose bounds on the correction capacity.
In the sequel, we deal with binary templates with at most $`N`$ bits and assume, for the theoretical analysis that follows, that the probabilities of error and erasure on each bit are independent. Note that resorting to interleaving makes this hypothesis valid for all practical purposes.
### 2.1 Theoretical limit
Our goal is to estimate the capacity, in the Shannon sense \[[15](#bib.bib15)\], of the matching channel when we work with a code of a given dimension. Namely, we want to know the maximum number of errors and erasures between two biometric measures that we can manage with fuzzy sketches for this code.
Starting with a representative range of matching biometric data, the theorem below gives an easy way to estimate the lowest achievable FRR. The idea is to check whether the best possible code with the best generic decoding algorithm, i.e. a maximum-likelihood (ML) decoding algorithm (which systematically outputs the most likely codeword), would succeed in correcting the errors.
######
Theorem 1
Let $`k \in {\mathbb{N}}^{\ast}`$, $`C`$ be a binary code of length $`N`$ and size $`2^{k}`$, and $`m`$ a random received message, from a random codeword of $`C`$, of length $`N`$ with $`w_{n}`$ errors and $`w_{e}`$ erasures. Assume that $`C`$ is an optimal code with respect to $`N`$ and $`k`$, equipped with an ML decoder.
If $`\frac{w_{n}}{N - w_{e}} > \theta`$ then $`m`$ is only decodable with a negligible probability, where $`\theta`$ is such that the Hamming sphere of radius $`{({N - w_{e}})}\hspace{0pt}\theta`$ in $`{\mathbb{F}}_{2}^{N - w_{e}}`$ contains $`2^{N - w_{e} - k}`$ elements.
Proof. In the case of errors only (i.e. no erasures) with error-rate $`p:={w_{e}/N}`$ , the canonical second theorem of Shannon asserts that there are families of codes with (transmission) rate $`R:={k/n}`$ coming arbitrarily close to the channel capacity $`\kappa\hspace{0pt}{(p)}`$, decodable with ML-decoding and a vanishing (in $`N`$) word error probability $`P_{e}`$.
In this case, $`{\kappa\hspace{0pt}{(p)}} = {1 - {h\hspace{0pt}{(p)}}}`$, where $`h\hspace{0pt}{(p)}`$ is the (binary) entropy function ($`\log`$’s are to the base 2):
<table>
<tbody>
<tr>
<td></td>
<td>h(x)=−xlogx−(1−x)log(1−x).ℎ𝑥𝑥𝑥1𝑥1𝑥h(x)=-x\log x-(1-x)\log(1-x).</td>
<td></td>
</tr>
</tbody>
</table>
Furthermore, $`P_{e}`$ displays a threshold phenomenon: for any rate arbitrarily close to, but above capacity and any family of codes, $`P_{e}`$ tends to 1 when $`N`$ grows.
Equivalently, given $`R`$, there exists an error-rate threshold of
<table>
<tbody>
<tr>
<td></td>
<td>p=h−1(1−R),𝑝superscriptℎ11𝑅p=h^{-1}(1-R),</td>
<td></td>
</tr>
</tbody>
</table>
$`h^{- 1}`$ being the inverse of the entropy function.
Back to the errors-and-erasures setting now. Our problem is to decode to the codeword nearest to the received word on the nonerased positions.
Thus we are now faced with a punctured code with length $`N - w_{e}`$, size $`2^{k}`$, transmission rate $`R':={k/{({N - w_{e}})}}`$ and required to sustain an error-rate $`p':=\frac{w_{n}}{N - w_{e}}`$.
By the previous discussion, if
<table>
<tbody>
<tr>
<td></td>
<td>p′>θ:=h−1(1−R′),superscript𝑝′𝜃assignsuperscriptℎ11superscript𝑅′p^{\prime}>\theta:=h^{-1}(1-R^{\prime}),</td>
<td></td>
</tr>
</tbody>
</table>
NO code and NO decoding procedure exist with a non-vanishing probability of success.
To conclude the proof, use the classical Stirling approximation for the size of a Hamming sphere of radius $`\alpha\hspace{0pt}M`$ in $`{\mathbb{F}}_{2}^{M}`$ by $`2^{h\hspace{0pt}{({\alpha\hspace{0pt}M})}}`$.
Practical implications of this theorem are illustrated in Table [1](#S3.T1 "Table 1 ‣ 3.3 Results on ICE database ‣ 3 Application ‣ Optimal Iris Fuzzy Sketches"), Sec. [3.3](#S3.SS3 "3.3 Results on ICE database ‣ 3 Application ‣ Optimal Iris Fuzzy Sketches").
## 3 Application
### 3.1 Description of the two-dimensional iterative min-sum decoding algorithm
A binary linear error-correcting code $`C`$ is a vector subspace of $`{\mathbb{F}}_{2}^{N}`$. The minimum distance $`d_{m\hspace{0pt}i\hspace{0pt}n}`$ of $`C`$ is the smallest Hamming distance between two distinct codewords. When $`k`$ is the dimension of the subspace $`C`$, i.e. when it contains $`2^{k}`$ codewords, $`C`$ is denoted by $`{\lbrack N,k,d_{m\hspace{0pt}i\hspace{0pt}n}\rbrack}_{2}`$. The correction capacity $`t`$ of $`C`$ is the radius of the largest Hamming ball for which, for any $`x \in {\mathbb{F}}_{2}^{N}`$, there is at most one codeword in the ball of radius $`t`$ centered on $`x`$. Clearly, $`t = {\lfloor{{({d_{m\hspace{0pt}i\hspace{0pt}n} - 1})}/2}\rfloor}`$. An altered codeword with $`w_{n}`$ errors and $`w_{e}`$ erasures can always be corrected (by ML decoding) provided $`{{2\hspace{0pt}w_{n}} + w_{e}} < d_{m\hspace{0pt}i\hspace{0pt}n}`$. However, if the code admits an iterative decoding algorithm, practical results overtake this limitation.
We will work with product codes together with a specific iterative decoding algorithm described below. A product code $`C = {C_{1} \otimes C_{2}}`$ is constructed from two codes: $`C_{1}\hspace{0pt}{\lbrack N_{1},k_{1},d_{1}\rbrack}_{2}`$ and $`C_{2}\hspace{0pt}{\lbrack N_{2},k_{2},d_{2}\rbrack}_{2}`$. The codewords of $`C`$ can be viewed as matrices of size $`N_{2} \times N_{1}`$ whose rows are codewords of $`C_{1}`$ and columns are codewords of $`C_{2}`$. This yields a $`\lbrack{N_{1} \times N_{2}},{k_{1} \times k_{2}},{d_{1} \times d_{2}}\rbrack`$ code. When $`k_{1}`$ and $`k_{2}`$ are small enough for $`C_{1}`$ and $`C_{2}`$ to be decoded exhaustively a very efficient iterative decoding algorithm is available, namely the min-sum decoding algorithm. Min-sum decoding of LDPC codes was developed by Wiberg \[[18](#bib.bib18)\] as a particular instance of message passing algorithms. In a somewhat different setting it was also proposed by Tanner \[[16](#bib.bib16)\] for decoding generalized LDPC (Tanner) codes. The variant we will be using is close to Tanner’s algorithm and is adapted to product codes. Min-sum is usually considered to perform slightly worse than the more classical sum-product message passing algorithm on the Gaussian, or binary-symmetric channels, but it is specially adapted to our case where knowledge of the channel is poor, and the emphasis is simply to use the Hamming distance as the appropriate basic cost function.
Let $`(x_{i\hspace{0pt}j})`$ be a vector of $`{\{ 0,1\}}^{N_{1} \times N_{2}}`$. The min-sum algorithm associates to every coordinate $`x_{i\hspace{0pt}j}`$ a cost function $`\kappa_{i\hspace{0pt}j}`$ for every iteration of the algorithm. The cost functions are defined on the set $`\{ 0,1\}`$. The initial cost function $`\kappa_{i\hspace{0pt}j}^{0}`$ is defined by $`{\kappa_{i\hspace{0pt}j}^{0}\hspace{0pt}{(x)}} = 0`$ if the received symbol on coordinate $`({i\hspace{0pt}j})`$ is $`x`$ and $`{\kappa_{i\hspace{0pt}j}^{0}\hspace{0pt}{(x)}} = 1`$ if the received symbol is $`1 - x`$.
A row iteration of the algorithm takes an input cost function $`\kappa_{i\hspace{0pt}j}^{i\hspace{0pt}n}`$ and produces an output cost function $`\kappa_{i\hspace{0pt}j}^{o\hspace{0pt}u\hspace{0pt}t}`$. The algorithm first computes, for every row $`i`$ and for every codeword $`c = {({c_{1}\hspace{0pt}\ldots\hspace{0pt}c_{N_{1}}})}`$ of $`C_{1}`$, the sum
<table>
<tbody>
<tr>
<td></td>
<td>κi(c)=∑j=1N1κijin(cj)subscript𝜅𝑖𝑐superscriptsubscript𝑗1subscript𝑁1superscriptsubscript𝜅𝑖𝑗𝑖𝑛subscript𝑐𝑗\kappa_{i}(c)=\sum_{j=1}^{N_{1}}\kappa_{ij}^{in}(c_{j})</td>
<td></td>
</tr>
</tbody>
</table>
which should be understood as the cost of putting codeword $`c`$ on row $`i`$. The algorithm then computes, for every $`i,j`$, $`\kappa_{i\hspace{0pt}j}^{o\hspace{0pt}u\hspace{0pt}t}`$ defined as the following min, over the set of codewords of $`C_{1}`$,
<table>
<tbody>
<tr>
<td></td>
<td>κijout(x)=minc∈C1,cj=xκi(c).superscriptsubscript𝜅𝑖𝑗𝑜𝑢𝑡𝑥subscriptformulae-sequence𝑐subscript𝐶1subscript𝑐𝑗𝑥subscript𝜅𝑖𝑐\kappa_{ij}^{out}(x)=\min_{c\in C_{1},c_{j}=x}\kappa_{i}(c).</td>
<td></td>
</tr>
</tbody>
</table>
This last quantity should be thought of as the minimum cost of putting the symbol $`x`$ on coordinate $`({i\hspace{0pt}j})`$ while satisfying the row constraint.
A column iteration of the algorithm is analogous to a row iteration, with simply the roles of the row and column indexes reversed, and code $`C_{2}`$ replacing code $`C_{1}`$. Precisely we have
<table>
<tbody>
<tr>
<td></td>
<td>κj(c)=∑i=1N2κijin(ci)subscript𝜅𝑗𝑐superscriptsubscript𝑖1subscript𝑁2superscriptsubscript𝜅𝑖𝑗𝑖𝑛subscript𝑐𝑖\kappa_{j}(c)=\sum_{i=1}^{N_{2}}\kappa_{ij}^{in}(c_{i})</td>
<td></td>
</tr>
</tbody>
</table>
and
<table>
<tbody>
<tr>
<td></td>
<td>κijout(x)=minc∈C2,ci=xκj(c).superscriptsubscript𝜅𝑖𝑗𝑜𝑢𝑡𝑥subscriptformulae-sequence𝑐subscript𝐶2subscript𝑐𝑖𝑥subscript𝜅𝑗𝑐\kappa_{ij}^{out}(x)=\min_{c\in C_{2},c_{i}=x}\kappa_{j}(c).</td>
<td></td>
</tr>
</tbody>
</table>
<table>
<tbody>
<tr>
<td></td>
<td></td>
<td></td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td>i(⋮κi1in⋯κiN1in⋮)𝑖missing-subexpression⋮missing-subexpressionmissing-subexpressionmissing-subexpressionmissing-subexpressionsuperscriptsubscript𝜅𝑖1𝑖𝑛⋯superscriptsubscript𝜅𝑖subscript𝑁1𝑖𝑛missing-subexpressionmissing-subexpressionmissing-subexpressionmissing-subexpression⋮missing-subexpression\displaystyle i\left(\begin{array}[]{ccc}&\vdots&\\ \hline\cr\kappa_{i1}^{in}&\cdots&\kappa_{iN_{1}}^{in}\\ \hline\cr&\vdots&\\ \end{array}\right)</td>
<td></td>
</tr>
</tbody>
</table>
<table>
<tbody>
<tr>
<td></td>
<td>⇓⇓\Downarrow\hskip 28.45274pt</td>
<td></td>
</tr>
</tbody>
</table>
<table>
<tbody>
<tr>
<td></td>
<td>κijout(x)=minc∈C1,cj=x∑k=1N1κikin(ck)superscriptsubscript𝜅𝑖𝑗𝑜𝑢𝑡𝑥subscriptformulae-sequence𝑐subscript𝐶1subscript𝑐𝑗𝑥superscriptsubscript𝑘1subscript𝑁1superscriptsubscript𝜅𝑖𝑘𝑖𝑛subscript𝑐𝑘\kappa_{ij}^{out}(x)=\min_{c\in C_{1},c_{j}=x}\sum_{k=1}^{N_{1}}\kappa_{ik}^{in}(c_{k})</td>
<td></td>
</tr>
</tbody>
</table>
<table>
<tbody>
<tr>
<td></td>
<td>⇓⇓\Downarrow\hskip 28.45274pt</td>
<td></td>
</tr>
</tbody>
</table>
<table>
<tbody>
<tr>
<td></td>
<td>i(⋮⋯κijout⋯⋮)𝑖missing-subexpression⋮missing-subexpressionmissing-subexpressionmissing-subexpressionmissing-subexpression⋯superscriptsubscript𝜅𝑖𝑗𝑜𝑢𝑡⋯missing-subexpressionmissing-subexpressionmissing-subexpressionmissing-subexpression⋮missing-subexpression\displaystyle i\left(\begin{array}[]{ccc}&\vdots&\\ \hline\cr\cdots&\kappa_{ij}^{out}&\cdots\\ \hline\cr&\vdots&\\ \end{array}\right)\hskip 11.38092pt</td>
<td></td>
</tr>
</tbody>
</table>
<table>
<tbody>
<tr>
<td></td>
<td>⇓j⇓𝑗\begin{array}[]{c}\Downarrow\\[11.38109pt] j\end{array}\hskip 28.45274pt</td>
<td></td>
</tr>
</tbody>
</table>
<table>
<tbody>
<tr>
<td></td>
<td>(κ1jin⋮⋯⋮⋯⋮κN2jin)missing-subexpressionsuperscriptsubscript𝜅1𝑗𝑖𝑛missing-subexpressionmissing-subexpression⋮missing-subexpression⋯⋮⋯missing-subexpression⋮missing-subexpressionmissing-subexpressionsuperscriptsubscript𝜅subscript𝑁2𝑗𝑖𝑛missing-subexpression\displaystyle\left(\begin{array}[]{c|c|c}&\kappa_{1j}^{in}&\\ &\vdots&\\ \cdots&\vdots&\cdots\\ &\vdots&\\ &\kappa_{N_{2}j}^{in}&\\ \end{array}\right)\hskip 11.38092pt</td>
<td></td>
</tr>
</tbody>
</table>
<table>
<tbody>
<tr>
<td></td>
<td>⇓⇓\Downarrow\hskip 28.45274pt</td>
<td></td>
</tr>
</tbody>
</table>
<table>
<tbody>
<tr>
<td></td>
<td>κijout(x)=minc∈C2,ci=x∑l=1N2κljin(cl)superscriptsubscript𝜅𝑖𝑗𝑜𝑢𝑡𝑥subscriptformulae-sequence𝑐subscript𝐶2subscript𝑐𝑖𝑥superscriptsubscript𝑙1subscript𝑁2superscriptsubscript𝜅𝑙𝑗𝑖𝑛subscript𝑐𝑙\kappa_{ij}^{out}(x)=\min_{c\in C_{2},c_{i}=x}\sum_{l=1}^{N_{2}}\kappa_{lj}^{in}(c_{l})</td>
<td></td>
</tr>
</tbody>
</table>
<table>
<tbody>
<tr>
<td></td>
<td>⇓j⇓𝑗\begin{array}[]{c}\Downarrow\\[11.38109pt] j\end{array}\hskip 28.45274pt</td>
<td></td>
</tr>
</tbody>
</table>
<table>
<tbody>
<tr>
<td></td>
<td>(⋮⋮⋯κijout⋯⋮⋮)missing-subexpression⋮missing-subexpressionmissing-subexpression⋮missing-subexpression⋯superscriptsubscript𝜅𝑖𝑗𝑜𝑢𝑡⋯missing-subexpression⋮missing-subexpressionmissing-subexpression⋮missing-subexpression\displaystyle\left(\begin{array}[]{c|c|c}&\vdots&\\ &\vdots&\\ \cdots&\kappa_{ij}^{out}&\cdots\\ &\vdots&\\ &\vdots&\\ \end{array}\right)\hskip 11.38092pt</td>
<td></td>
</tr>
</tbody>
</table>
Figure 1: A row iteration followed by a column one
The algorithm alternates row and column iterations as illustrated by Fig. [1](#S3.F1 "Figure 1 ‣ 3.1 Description of the two-dimensional iterative min-sum decoding algorithm ‣ 3 Application ‣ Optimal Iris Fuzzy Sketches"). After a given number of iterations (or before, if we find a codeword) it stops, and the value of every symbol $`x_{i\hspace{0pt}j}`$ is put at $`x_{i\hspace{0pt}j} = x`$ if $`{\kappa_{i\hspace{0pt}j}^{o\hspace{0pt}u\hspace{0pt}t}\hspace{0pt}{(x)}} < {\kappa_{i\hspace{0pt}j}^{o\hspace{0pt}u\hspace{0pt}t}\hspace{0pt}{({1 - x})}}`$. If $`{\kappa_{i\hspace{0pt}j}^{o\hspace{0pt}u\hspace{0pt}t}\hspace{0pt}{(x)}} = {\kappa_{i\hspace{0pt}j}^{o\hspace{0pt}u\hspace{0pt}t}\hspace{0pt}{({1 - x})}}`$ then the value of $`x_{i\hspace{0pt}j}`$ stays undecided (or erased).
The following theorem is fairly straightforward to prove and illustrates the power of min-sum decoding.
######
Theorem 2
If the number of errors is less than $`{d_{1}\hspace{0pt}d_{2}}/2`$, then two iterations of min-sum decoding of the product code $`C_{1} \otimes C_{2}`$ recover the correct codeword.
### 3.2 Our setting
To validate our approach, we now present the results of experiments on a practical iris database where we obtain correction performances close to the theoretical limit.
The database used for these experiments is the ICE (Phase I) database \[[10](#bib.bib10), [13](#bib.bib13)\] which contained 2953 images from 244 different eyes. A 256-byte (2048 bits) iris template, together with a 256-byte mask, is computed from each iris image using the algorithm reported in \[[5](#bib.bib5)\]; the mask filters out the unreliable bits, i.e. stores the erasures indices of the iris template. The database is taken without any modification but two slight corrections: one eye is suppressed due to a very low quality and the side of another eye has been switched from left to right. Hence we keep 2952 images. Note that in the database, the number of images provided for each eye is variable: so the number of intra-eye matching verifications between two iris codes from the same eye is not constant. The same holds for the inter-eye matching between two iris codes from different eyes. Among all the combinations, its gives a set of 29827 intra-eye matching and about 4 million of inter-eye matching to check.
The classical way to compare two iris codes $`I_{1},I_{2}`$ with masks $`M_{1},M_{2}`$ is to compute the relative Hamming distance
<table>
<tbody>
<tr>
<td></td>
<td>‖(I1⊕I2)∩M1∩M2‖‖M1∩M2‖normdirect-sumsubscript𝐼1subscript𝐼2subscript𝑀1subscript𝑀2normsubscript𝑀1subscript𝑀2\frac{||(I_{1}\oplus I_{2})\cap M_{1}\cap M_{2}||}{||M_{1}\cap M_{2}||}</td>
<td></td>
</tr>
</tbody>
</table>
for some rotations of the second template – to deal with the iris orientation’s variation – and to keep the lowest score. It gives the following distributions of matching scores (cf. Fig [2](#S3.F2 "Figure 2 ‣ 3.2 Our setting ‣ 3 Application ‣ Optimal Iris Fuzzy Sketches"))
[Refer to caption](/html/0705.3740/assets/x1.png)
Figure 2: Inter-eyes and intra-eye distributions
where we see an overlap between the two curves. We also see that the number of errors to handle in the matching channel is large (for instance at least $`29\%`$ of errors for a FRR lower than $`5\%`$). On this channel, an additional difficulty originates from the number of erasures which varies from 512 to 1977.
### 3.3 Results on ICE database
We have experimented with the algorithm described in section [3.1](#S3.SS1 "3.1 Description of the two-dimensional iterative min-sum decoding algorithm ‣ 3 Application ‣ Optimal Iris Fuzzy Sketches") on this database with a particular choice for the code. In fact, the product code is constructed to fit with an array of 2048 bits, by using Reed-Muller codes \[[12](#bib.bib12), [14](#bib.bib14)\] of order 1 which are known to have good weight distributions. A binary Reed-Muller code of order 1 in $`m`$ variables, abbreviated as $`R\hspace{0pt}M\hspace{0pt}{(1,m)}`$, is a $`{\lbrack 2^{m},{m + 1},2^{m - 1}\rbrack}_{2}`$ code. We chose to combine the $`R\hspace{0pt}M\hspace{0pt}{(1,6)}`$ with the $`R\hspace{0pt}M\hspace{0pt}{(1,5)}`$, leading to a product code of dimension 42 and codewords of length $`64 \times 32`$.
As the density of errors and erasures in an iris code can be very high in some regions, we also added a randomly chosen interleaver to break this structure and increase the efficiency of the decoding algorithm. In so doing, we succeeded in obtaining a FRR of about $`5.62\%`$ for a very small FAR (lower than $`10^{- 5}`$). This is in fact very close to the FAR obtained in a classical matching configuration for a similar FRR.
The overall size of the code could appear small from a cryptographic point of view, but following the theoretical analysis of section [2.1](#S2.SS1 "2.1 Theoretical limit ‣ 2 Model ‣ Optimal Iris Fuzzy Sketches"), it is difficult to expect much more while achieving a low FRR on this database. Indeed, from the distribution of errors and erasures on the matching channel, we obtain by Theorem [1](#Thmthm1 "Theorem 1 ‣ 2.1 Theoretical limit ‣ 2 Model ‣ Optimal Iris Fuzzy Sketches") the practical limits which are reported in Table [1](#S3.T1 "Table 1 ‣ 3.3 Results on ICE database ‣ 3 Application ‣ Optimal Iris Fuzzy Sketches").
Table 1: Theoretical limits on ICE database
<table>
<tbody>
<tr>
<th>Code’s dimension</th>
<td>Best theoretical FRR</td>
</tr>
<tr>
<th>42</th>
<td>2.49%percent2.492.49\%</td>
</tr>
<tr>
<th>64</th>
<td>3.76%percent3.763.76\%</td>
</tr>
<tr>
<th>80</th>
<td>4.87%percent4.874.87\%</td>
</tr>
<tr>
<th>128</th>
<td>9.10%percent9.109.10\%</td>
</tr>
</tbody>
</table>
Remark. In \[[8](#bib.bib8)\], the fuzzy sketch scheme is applied with a concatenated error-correcting code combining a Hadamard code and a Reed-Solomon code. More precisely, the authors use a Reed-Solomon code of length 32 over $`{\mathbb{F}}_{2^{7}}`$ (with a correction capacity $`t_{R\hspace{0pt}S} < 16`$) and a Hadamard code of order 6 and length 64 (with a correction capacity $`t_{H} = 15`$): a codeword of 2048 bits is in fact constructed as a set of 32 blocks of 64 bits where each block is a codeword of the underlying Hadamard code. As explained in \[[8](#bib.bib8)\], the Hadamard code is introduced to deal with the background errors and the Reed-Solomon code to deal with the bursts (e.g. caused by eyelashes, reflections, $`\ldots`$).
Note that in this scheme, the model is not exactly the same as ours, as the masks are not taken into account. Moreover, the quality of the database used in \[[8](#bib.bib8)\] is better than for the ICE database. Actually, \[[8](#bib.bib8)\] reports very good results on their experiments with a database of 700 images, but the codes do not seem appropriate to our case as our experiment on the ICE database gave a too large rate of FR (e.g. $`10\%`$ of FR with $`0.80\%`$ of FA), even for the smallest possible dimension of the Reed-Solomon code when $`t_{R\hspace{0pt}S} = 15`$.
## 4 Conclusion
We derived explicit upper bounds on the correction capacity of Fuzzy Sketches on iris-based biometrics. We then showed how the two-dimensional iterative min-sum decoding algorithm achieves correction performance close to the optimal decoding rate. Our results were validated on a typical iris database.
## References
- \[1\] X. Boyen, Y. Dodis, J. Katz, R. Ostrovsky, and A. Smith. Secure remote authentication using biometric data.
In R. Cramer, editor, EUROCRYPT, volume 3494 of Lecture Notes in Computer Science, pages 147–163. Springer, 2005.
- \[2\] G. Cohen and G. Zémor. Generalized coset schemes for the wire-tap channel: application to biometrics.
In ISIT, 2004.
- \[3\] G. Cohen and G. Zémor. The wire-tap channel applied to biometrics.
In ISITA, 2004.
- \[4\] T. M. Cover and J. A. Thomas.
Elements of Information Theory.
Wiley, 2006.
- \[5\] J. Daugman. The importance of being random: statistical principles of iris recognition.
Pattern Recognition, 36(2):279–291, 2003.
- \[6\] G. I. Davida and Y. Frankel. Perfectly secure authorization and passive identification for an error tolerant biometric system.
In M. Walker, editor, IMA Int. Conf., volume 1746 of Lecture Notes in Computer Science, pages 104–113. Springer, 1999.
- \[7\] Y. Dodis, L. Reyzin, and A. Smith. Fuzzy extractors: How to generate strong keys from biometrics and other noisy data.
In C. Cachin and J. Camenisch, editors, EUROCRYPT, volume 3027 of Lecture Notes in Computer Science, pages 523–540. Springer, 2004.
- \[8\] F. Hao, R. Anderson, and J. Daugman. Combining crypto with biometrics effectively.
IEEE Transactions on Computers, 55(9):1081–1088, 2006.
- \[9\] A. Juels and M. Wattenberg. A fuzzy commitment scheme.
In ACM Conference on Computer and Communications Security, pages 28–36, 1999.
- \[10\] Xiaomei Liu, Kevin W. Bowyer, and Patrick J. Flynn. Iris Recognition and Verification Experiments with Improved Segmentation Method.
In Fourth IEEE Workshop on Automatic Identification Advanced Technologies (AutoID), 17-18 October 2005, Buffalo, New York, 2005.
- \[11\] F. J. MacWilliams and N. J. A. Sloane.
The Theory of Error-correcting Codes.
North-Holland, 1988.
- \[12\] D.E. Muller. Application of boolean algebra to switching circuit design and to error detection.
IEEE Trans. on Electronic Computers, 3:6–12, 1954.
- \[13\] National Institute of Science and Technology (NIST). Iris Challenge Evaluation.
http://iris.nist.gov/ICE, 2005.
- \[14\] I.S. Reed. A class of multiple-error-correcting codes and their decoding scheme.
IEEE Trans. on Information Theory, 4:38–42, 1954.
- \[15\] Claude E. Shannon. A mathematical theory of communication.
The Bell System Technical Journal, 27, 1948.
http://cm.bell-labs.com/cm/ms/what/shannonday/paper.html.
- \[16\] R. M. Tanner. A recursive approach to low-complexity codes.
IEEE Trans. on Information Theory, 27:533–547, 1981.
- \[17\] P. Tuyls and J. Goseling. Capacity and examples of template-protecting biometric authentication systems.
In D. Maltoni and A. K. Jain, editors, ECCV Workshop BioAW, volume 3087 of Lecture Notes in Computer Science, pages 158–170. Springer, 2004.
- \[18\] N. Wiberg.
Codes and Decoding on general Graphs.
PhD thesis, Linkoping University, Linkoping, Sweden, 1996. |
0704.1068 | cs.NI | cyber | Fast paths in large-scale dynamic road networks
Giacomo Nannicini<sup>1,2</sup>, Philippe Baptiste<sup>1</sup>, Gilles Barbier<sup>2</sup>, Daniel Krob<sup>1</sup>, Leo Liberti<sup>1</sup>
- <sup>1</sup>
LIX, École Polytechnique, F-91128 Palaiseau, France
Email:{giacomon,baptiste,dk,liberti}@lix.polytechnique.fr
- <sup>2</sup>
Mediamobile, 10 rue d’Oradour sur Glane, Paris, France
Email:{giacomo.nannicini,gilles.barbier,contact}@v-trafic.com
###### Abstract
Efficiently computing fast paths in large-scale dynamic road networks (where dynamic traffic information is known over a part of the network) is a practical problem faced by several traffic information service providers who wish to offer a realistic fast path computation to GPS terminal enabled vehicles. The heuristic solution method we propose is based on a highway hierarchy-based shortest path algorithm for static large-scale networks; we maintain a static highway hierarchy and perform each query on the dynamically evaluated network.
## 1 Introduction
Several cars are now fitted with a Global Positioning System (GPS) terminal which gives the exact geographic location of the vehicle on the surface of the earth. All of these GPS terminals are now endowed with detailed road network databases which allow them to compute the shortest path (in terms of distance) between the current vehicle location (source) and another location given by the driver (destination). Naturally, drivers are more interested in the source-destination fastest path (i.e. shortest in terms of travelling time). The greatest difficulty to overcome is that the travelling time depends heavily on the amount of traffic on the chosen road. Currently, some state agencies as well as commercial enterprises are charged with monitoring the traffic situation in certain pre-determined strategic places. Furthermore, traffic reports are collected from police cars as well as some taxi services. The dynamic traffic information, however, is as yet limited to a small proportion of the whole road network.
The problem faced by traffic information providers is currently that of offering GPS terminal enabled drivers a source-destination path subject to the following constraints: (a) the path should be fast in terms of travelling time subject to dynamic traffic information being available on part of the road network; (b) traffic information data are updated approximately each minute; (c) answers to path queries should be computed in real time. Given the data communication time and other overheads, constraint (c) practically asks for a shortest path computation time of no more than 1 second. Constraint (b) poses a serious problem, because it implies that the fastest source-destination path may change each minute, giving an on-line dimension to the problem. A source-destination query spanning several hundred kilometers, which would take several hours to travel, would need a system recomputing the fastest path each minute; this in turn would mean keeping track of each query for potentially several hours. As the estimated computational cost of this requirement is superior to the resources usually devoted to the task, a system based on dynamic traffic information will not, in practice, ever compute the on-line fastest path. As a typical national road network for a large European country usually counts several million junctions and road segments, constraint (c) implies that a straight Dijkstra’s algorithm is not a viable option. In view of constraint (a), in our solution method fast paths can be efficiently computed by means of a point-to-point hierarchy-based shortest path algorithm for static large-scale networks, where the hierarchy is built using static information and each query is answered on the dynamically evaluated network.
This paper makes two original scientific contributions (i) We extend a known hierarchy-based shortest path algorithm for static large-scale undirected graphs (the Highway Hierarchies algorithm \[[SS05](#bib.bibx17)\]) to the directed case. The method has been developed and tested on real road network data taken from the TeleAtlas France database \[[NV05](#bib.bibx15)\]. We note that the original authors of \[[SS05](#bib.bibx17)\] have extended the algorithm to work on directed graphs in a slightly different way than ours (see \[[SS06](#bib.bibx18)\]). (ii) We propose a method for efficiently finding fast paths on a large-scale dynamic road network where arc travelling times are updated in quasi real-time (meaning very often but not continuously).
In the rest of this section, we discuss the state of the art as regards shortest path algorithms in dynamic and large-scale networks, and we describe the proposed solution. The rest of the paper is organized as follows. In Section [2](#S2 "2 Highway Hierarchies algorithm on dynamic directed graphs") we briefly review the highway hierarchy-based shortest path algorithm for static large-scale networks, which is one of the important building blocks of our method, and discuss the extension of the existing shortest-path algorithm to the directed case. Section [3](#S3 "3 Computational results") discusses the computational results, and Section [4](#S4 "4 Conclusion") concludes the paper.
### 1.1 Shortest path algorithms in road networks
The problem of computing fastest paths in graphs whose arc weights change over time is termed the Dynamic Shortest Path Problem (DSPP) \[[BRTed](#bib.bibx2)\]. The work that laid the foundations for solving the DSPP is \[[CH66](#bib.bibx3)\] (a good review of this paper can be found in \[[Dre69](#bib.bibx9)\], p. 407): Dijkstra’s algorithm is extended to the dynamic case through a recursion formula based on the assumption that the network $`G = {(V,A)}`$ has the FIFO property: for each pair of time instants $`t,t'`$ with $`t < t'`$:
<table>
<tbody>
<tr>
<td></td>
<td>∀(u,v)∈Aτuv(t)+t≤τuv(t′)+t′,for-all𝑢𝑣𝐴subscript𝜏𝑢𝑣𝑡𝑡subscript𝜏𝑢𝑣superscript𝑡′superscript𝑡′\forall\;(u,v)\in A\ \tau_{uv}(t)+t\leq\tau_{uv}(t^{\prime})+t^{\prime},</td>
<td></td>
</tr>
</tbody>
</table>
where $`\tau_{u\hspace{0pt}v}\hspace{0pt}{(t)}`$ is the travelling time on the arc $`(u,v)`$ starting from $`u`$ at time $`t`$. The FIFO property is also called the non-overtaking property, because it basically says that if $`A`$ leaves $`u`$ at time $`t`$ and $`B`$ at time $`t' > t`$, $`B`$ cannot arrive at $`v`$ before $`A`$ using the arc $`(u,v)`$. The shortest path problem in dynamic FIFO networks is therefore polynomially solvable \[[Cha98](#bib.bibx4)\], even in the presence of traffic lights \[[AOPS03](#bib.bibx1)\]. Dijkstra’s algorithm applied to dynamic FIFO networks has been optimized in various ways \[[BRTed](#bib.bibx2), [Cha98](#bib.bibx4)\]; the $`A^{\ast}`$ one-to-one shortest path algorithm has also been extended to dynamic networks \[[CS02](#bib.bibx5)\]. The DSPP is NP-hard in non-FIFO networks \[[Dea04](#bib.bibx7)\].
Although in this paper we do not assume any knowledge about the statistical distribution of the arc weights in time, it is worth mentioning that a considerable amount of work has been carried out for computing shortest paths in stochastic networks. A good review is \[[FHK<sup>+</sup>05](#bib.bibx11)\].
The computation of exact shortest paths in large-scale static networks has received a good deal of attention \[[CZ01](#bib.bibx6)\]. The established practice is to delegate a considerable amount of computation to a preprocessing phase (which may be very slow) and then perform fast source-destination shortest path queries on the pre-processed data. Recently, the concept of highway hierarchy was proposed in \[[SS05](#bib.bibx17), [Sch05](#bib.bibx16), [SS06](#bib.bibx18)\]. A highway hierarchy of $`L`$ levels of a graph $`G = {(V,A)}`$ is a sequence of graphs $`G = {G^{0},\ldots,G^{L}}`$ with vertex sets $`{V^{0} = V},{V^{1} \supseteq \ldots \supseteq V^{L}}`$ and arc sets $`{A^{0} = A},{A^{1} \supseteq \ldots \supseteq A^{L}}`$; each arc has maximum hierarchy level (the maximum $`i`$ such that it belongs to $`A^{i}`$) such that for all pairs of vertices there exists between them a shortest path $`(a_{1},\ldots,a_{k})`$, where $`a_{i}`$ are the consecutive path arcs, whose search level first increases and then decreases, and each arc’s search level is not greater than its maximum hierarchy level. A more precise description is given in Section [2](#S2 "2 Highway Hierarchies algorithm on dynamic directed graphs"). The $`A^{\ast}`$ algorithm has also been extended to use a concept, reach, which has turned out to be closely related to highway hierarchies (see \[[GKW05](#bib.bibx12)\]).
### 1.2 Description of the solution method
The solution method we propose in this paper efficiently finds fast paths by deploying Dijkstra-like queries on a highway hierarchy built using the static arc weights found in the road network database, but used with the dynamic arc weights reflecting quasi real-time traffic observations. This implies using two main building blocks: highway hierarchy construction (the Highway Hierarchies
<sup>1</sup>
<sup>1</sup>1From now on, simply HH
algorithm extended to directed graphs), and the query algorithm. Consequently, the implementation is a complex piece of software whose architecture has been detailed in the appendix.
- •
Highway hierarchy. Apply the directed graph extension of the HH algorithm (see Section [2](#S2 "2 Highway Hierarchies algorithm on dynamic directed graphs")) to construct a highway hierarchy using the static road network information. In particular, arc travelling times are average estimations found in the database. This is a preprocessing step that has to be performed only when the topology of the road network changes. The CPU time taken for this step is not an issue.
- •
Efficient path queries. Efficiently address source-destination fast path requests by employing a multi-level bidirectional Dijkstra’s algorithm on the dynamically evaluated graph using the highway hierarchy structure constructed during preprocessing. This algorithm is carried out each time a path request is issued; its running time must be as fast as possible, in any case not over 1 second.
## 2 Highway Hierarchies algorithm on dynamic directed graphs
The Highway Hierarchies algorithm \[[SS05](#bib.bibx17), [Sch05](#bib.bibx16)\] is a fast, hierarchy-based shortest paths algorithm which works on static undirected graphs. HH algorithm is specially suited to efficiently finding shortest paths in large-scale networks. Since the HH algorithm is one of our main building blocks, we briefly review the necessary concepts.
The Highway Hierarchies algorithm is heavily based on Dijkstra’s algorithm \[[Dij59](#bib.bibx8)\], which finds the tree of all shortest paths from a root vertex $`r`$ to all other vertices $`v \in V`$ of a weighted digraph $`G = {(V,A)}`$ by maintaining a heap $`H`$ of reached vertices $`u`$ with their associated (current) shortest path length $`c\hspace{0pt}{(u)}`$ (elements of the heap are denoted by $`\lbrack u,{c\hspace{0pt}{(u)}}\rbrack`$. Vertices which have not yet entered the heap (i.e. which are still unvisited) are unreached, and vertices which have already exited the heap (i.e. for which a shortest path has already been found) are settled. Dijkstra’s algorithm is as follows.
1. 1\.
Let $`H = {\{{\lbrack r,0\rbrack}\}}`$.
2. 2\.
If $`H = \varnothing`$, terminate.
3. 3\.
Let $`u`$ be the vertex in $`H`$ with minimum associated path length $`c\hspace{0pt}{(u)}`$.
4. 4\.
Let $`H = {H \smallsetminus {\{ u\}}}`$.
5. 5\.
For all $`v \in {\delta^{+}\hspace{0pt}{(u)}}`$, if $`{{c\hspace{0pt}{(u)}} + \tau_{u\hspace{0pt}v}} < {c\hspace{0pt}{(v)}}`$ then let $`H = {{({H \smallsetminus {\{{\lbrack v,{c\hspace{0pt}{(v)}}\rbrack}\}}})} \cup {\{{\lbrack v,{{c\hspace{0pt}{(u)}} + \tau_{u\hspace{0pt}v}}\rbrack}\}}}`$.
6. 6\.
Go to 2.
A bidirectional Dijkstra algorithm works by keeping track of two Dijkstra search scopes: one from the source, and one from the destination working on the reverse graph. When the two search scopes meet it can be shown that the shortest path passes through a vertex that has been reached from both nodes (\[[Sch05](#bib.bibx16)\], p. 30). A set of shortest paths is
canonical
<sup>2</sup>
<sup>2</sup>
2
Dijkstra’s algorithm can easily be modified to output a canonical shortest paths tree (see \[[Sch05](#bib.bibx16)\], Appendix A.1 — can be downloaded from http://algo2.iti.uka.de/schultes/hwy/).
if, for any shortest path $`p = {\langle u_{1},\ldots,u_{i},\ldots,{u_{j}\hspace{0pt}\ldots},u_{k}\rangle}`$ in the set, the canonical shortest path between $`u_{i}`$ and $`u_{j}`$ is a subpath of $`p`$.
The HH algorithm works in two stages: a time-consuming pre-processing stage to be carried out only once, and a fast query stage to be executed at each shortest path request. Let $`G^{0} = G`$. During the first stage, a highway hierarchy is constructed, where each hierarchy level $`G^{l}`$, for $`l \leq L`$, is a modified subgraph of the previous level graph $`G^{l - 1}`$ such that no canonical shortest path in $`G^{l - 1}`$ lies entirely outside the current level for all sufficiently distant path endpoints: this ensures that all queries between far endpoints on level $`l - 1`$ are mostly carried out on level $`l`$, which is smaller, thus speeding up the search. Each shortest path query is executed by a multi-level bidirectional Dijkstra algorithm: two searches are started from the source and from the destination, and the query is completed shortly after the search scopes have met; at no time do the search scopes decrease hierarchical level. Intuitively, path optimality is due to the fact that by hierarchy construction there exist no canonical shortest path of the form $`\langle a_{1},\ldots,a_{i},\ldots,{a_{j}\hspace{0pt}\ldots},{a_{k}\hspace{0pt}\ldots}\rangle`$, where $`{a_{i},a_{j},a_{k}} \in A`$ and the search level of $`a_{j}`$ is lower than the level of both $`a_{i},a_{k}`$; besides, each arc’s search level is always lower or equal to that arc’s maximum level, which is computed during the hierarchy construction phase and is equal to the maximum level $`l`$ such that the arc belongs to $`G^{l}`$. The speed of the query is due to the fact that the search scopes occur mostly on a high hierarchy level, with fewer arcs and nodes than in the original graph.
### 2.1 Highway hierarchy
As the construction of the highway hierarchy is the most complicated part of HH algorithm, we endeavour to explain its main traits in more detail. Given a local extensionality parameter $`H`$ (which measures the degree at which shortest path queries are satisfied without stepping up hierarchical levels) and the maximum number of hierarchy levels $`L`$, the iterative method to build the next highway level $`l + 1`$ starting from a given level graph $`G^{l}`$ is as follows:
1. 1\.
For each $`v \in V`$, build the neighbourhood $`N_{H}^{l}\hspace{0pt}{(v)}`$ of all vertices reached from $`v`$ with a simple Dijkstra search in the $`l`$-th level graph up to and including the $`H`$-st settled vertex. This defines the local extensionality of each vertex, i.e. the extent to which the query “stays on level $`l`$”.
2. 2\.
For each $`v \in V`$:
1. \(a\)
Build a partial shortest path tree $`B\hspace{0pt}{(v)}`$ from $`v`$, assigning a status to each vertex. The initial status for $`v`$ is “active”. The vertex status is inherited from the parent vertex whenever a vertex is reached or settled. A vertex $`w`$ which is settled on the shortest path $`\langle v,u,\ldots,w\rangle`$ (where $`v \neq u \neq w`$) becomes “passive” if
<table>
<tbody>
<tr>
<td></td>
<td>|NHl(u)∩NHl(w)|≤1.superscriptsubscript𝑁𝐻𝑙𝑢superscriptsubscript𝑁𝐻𝑙𝑤1|N_{H}^{l}(u)\cap N_{H}^{l}(w)|\leq 1.</td>
<td></td>
<td>(1)</td>
</tr>
</tbody>
</table>
The partial shortest path tree is complete when there are no more active reached but unsettled vertices left.
2. \(b\)
From each leaf $`t`$ of $`B\hspace{0pt}{(v)}`$, iterate backwards along the branch from $`t`$ to $`v`$: all arcs $`(u,w)`$ such that $`u \notin {N_{H}^{l}\hspace{0pt}{(t)}}`$ and $`w \notin {N_{H}^{l}\hspace{0pt}{(v)}}`$, as well as their adjacent vertices $`u,w`$, are raised to the next hierarchy level $`l + 1`$.
3. 3\.
Select a set of bypassable nodes on level $`l + 1`$; intuitively, these nodes have low degree, so that the benefit of skipping them during a search outweights the drawbacks (i.e., the fact that we have to add shortcuts to preserve the algorithm’s correctness). Specifically, for a given set $`B_{l + 1} \subset V_{l + 1}`$ of bypassable nodes, we define the set $`S_{l + 1}`$ of shortcut edges that bypass the nodes in $`B_{l + 1}`$: for each path $`p = {(s,b_{1},b_{2},\ldots,b_{k},t)}`$ with $`{s,t} \in {V_{l + 1} \smallsetminus {B\hspace{0pt}l}}`$ and $`{b_{i} \in B_{l + 1}},{1 \leq i \leq k}`$, the set $`S_{l + 1}`$ contains an edge $`(s,t)`$ with $`{c\hspace{0pt}{(s,t)}} = {c\hspace{0pt}{(p)}}`$. The core $`G_{l + 1}' = {(V_{l + 1}',E_{l + 1}')}`$ of level $`l + 1`$ is defined as:$`V_{l + 1}' = {V_{l + 1} \smallsetminus B_{l + 1}}`$, $`E_{l + 1}' = {{({E_{l + 1} \cap {({V_{l + 1}' \times V_{l + 1}'})}})} \cup S_{l + 1}}`$.
The result of the contraction is the contracted highway network $`G_{l + 1}'`$, which can be used as input for the following iteration of the construction procedure. It is worth noting that higher level graphs may be disconnected even though the original graph is connected.
######
2.1 Example
Take the directed graph $`G = {(V,A)}`$ given in Fig. [1](#S2.F1 "Figure 1 ‣ 2.1 Example ‣ 2.1 Highway hierarchy ‣ 2 Highway Hierarchies algorithm on dynamic directed graphs") (above). We are going to construct a road hierarchy with $`H = 3`$ and $`L = 1`$ on $`G`$. First we compute $`N_{3}^{0}\hspace{0pt}{(v)}`$ for all $`v \in V = {\{ v_{0},\ldots,v_{6}\}}`$.
<table>
<thead>
<tr>
<th>v𝑣v</th>
<th>N30(v)superscriptsubscript𝑁30𝑣N_{3}^{0}(v)</th>
</tr>
</thead>
<tbody>
<tr>
<td>v0,v1,v2subscript𝑣0subscript𝑣1subscript𝑣2v_{0},v_{1},v_{2}</td>
<td>{0,1,2}012\{0,1,2\}</td>
</tr>
<tr>
<td>v3,v4,v5subscript𝑣3subscript𝑣4subscript𝑣5v_{3},v_{4},v_{5}</td>
<td>{3,4,5}345\{3,4,5\}</td>
</tr>
<tr>
<td>v6subscript𝑣6v_{6}</td>
<td>{3,5,6}356\{3,5,6\}</td>
</tr>
</tbody>
</table>
Next, we compute $`B\hspace{0pt}{(v)}`$ for all $`v \in V`$ and raise the hierarchy level of the relevant arcs from the leaves to $`B\hspace{0pt}{(v)}`$ to $`v`$. We only discuss the computation of $`B\hspace{0pt}{(v_{0})}`$ in detail as the others are similar.
1. 1\.
Vertex $`v_{0}`$ is initialized as an active vertex.
2. 2\.
Dijkstra’s algorithm is started.
1. \(a\)
$`v_{0}`$ is settled (cost $`0`$) on the empty path, so the passivity condition ([1](#S2.E1 "In item 2a ‣ item 2 ‣ 2.1 Highway hierarchy ‣ 2 Highway Hierarchies algorithm on dynamic directed graphs")) does not apply;
2. \(b\)
$`v_{1}`$ and $`v_{2}`$ are reached from $`v_{0}`$ with costs resp. $`1`$ and $`2`$, and inherit its active status;
3. \(c\)
$`v_{1}`$ is settled (cost $`1`$) on the path $`\langle v_{0},v_{1}\rangle`$ and condition ([1](#S2.E1 "In item 2a ‣ item 2 ‣ 2.1 Highway hierarchy ‣ 2 Highway Hierarchies algorithm on dynamic directed graphs")) does not apply;
4. \(d\)
$`v_{6}`$ is reached from $`v_{1}`$ with cost $`{1 + 4} = 5`$ and set to active;
5. \(e\)
$`v_{2}`$ (cost $`2`$) is settled on $`\langle v_{0},v_{2}\rangle`$;
6. \(f\)
$`v_{4}`$ is reached from $`2`$ with cost $`{2 + 6} = 8`$ and set to active;
7. \(g\)
$`v_{6}`$ (cost $`5`$) is settled on the path $`\langle v_{0},v_{1},v_{6}\rangle`$: since $`{{N_{3}^{0}\hspace{0pt}{(v_{1})}} \cap {N_{3}^{0}\hspace{0pt}{(v_{6})}}} = \varnothing`$, condition ([1](#S2.E1 "In item 2a ‣ item 2 ‣ 2.1 Highway hierarchy ‣ 2 Highway Hierarchies algorithm on dynamic directed graphs")) is verified, and $`v_{6}`$ is labeled passive;
8. \(h\)
$`v_{3}`$ is reached from $`v_{6}`$ with cost $`{1 + 4 + 4} = 9`$ and set to passive.
9. \(i\)
$`v_{4}`$ (cost $`8`$) is settled on the path $`\langle v_{0},v_{2},v_{4}\rangle`$: since $`{{N_{3}^{0}\hspace{0pt}{(v_{2})}} \cap {N_{3}^{0}\hspace{0pt}{(v_{4})}}} = \varnothing`$, condition ([1](#S2.E1 "In item 2a ‣ item 2 ‣ 2.1 Highway hierarchy ‣ 2 Highway Hierarchies algorithm on dynamic directed graphs")) is verified, and $`v_{4}`$ is labeled passive;
10. \(j\)
$`v_{5}`$ is reached from $`v_{4}`$ with cost $`{2 + 6 + 2} = 10`$ and set to passive;
11. \(k\)
the only unsettled vertices are $`v_{3}`$ and $`v_{5}`$. Since both are reached and passive, the search terminates.
3. 3\.
The leaf vertices of $`B\hspace{0pt}{(v_{0})}`$ are $`v_{4}`$ and $`v_{6}`$.
1. \(a\)
From $`t = v_{4}`$, we iterate backwards along the arcs on the path $`\langle v_{0},v_{2},v_{4}\rangle`$: the arc $`(v_{2},v_{4})`$ has the property that $`v_{2} \notin {N_{3}^{0}\hspace{0pt}{(v_{4})}}`$ and $`v_{4} \notin {N_{3}^{0}\hspace{0pt}{(v_{2})}}`$, so its hierarchy level is raised to $`{l + 1} = 1`$ (the other arc on the path, $`(v_{0},v_{2})`$, stays at level $`l = 0`$);
2. \(b\)
from $`t = v_{6}`$, we iterate backwards along the arcs on the path $`\langle v_{0},v_{1},v_{6}\rangle`$: the arc $`(v_{1},v_{6})`$ has the property that $`v_{1} \notin {N_{3}^{0}\hspace{0pt}{(v_{6})}}`$ and $`v_{6} \notin {N_{3}^{0}\hspace{0pt}{(v_{1})}}`$, so its hierarchy level is raised to $`1`$ (the other arc on the path stays at level $`0`$).
Fig. [1](#S2.F1 "Figure 1 ‣ 2.1 Example ‣ 2.1 Highway hierarchy ‣ 2 Highway Hierarchies algorithm on dynamic directed graphs") shows the hierarchy at level 1.
[Refer to caption](/html/0704.1068/assets/x1.png) [Refer to caption](/html/0704.1068/assets/x2.png)
Figure 1:
The graph of Example [2.1](#S2.Thmresult1 "2.1 Example ‣ 2.1 Highway hierarchy ‣ 2 Highway Hierarchies algorithm on dynamic directed graphs") (left) and its highway hierarchy for $`{H = 3},{L = 1}`$ (right): the dashed lines indicate arcs at level 0, the solid lines indicate arcs at level 1.
### 2.2 Extension to directed graphs
The original description of the HH algorithm \[[SS05](#bib.bibx17)\] applies to undirected graphs only; in this section we provide an extension to the directed case. It should be noted that the HH algorithm was extended to the directed case by the authors (see \[[SS06](#bib.bibx18)\]) in a way which is very similar to that described here. However, we believe our slightly different exposition helps to clarify these ideas considerably.
The algorithm for hierarchy construction, as explained in Section [2.1](#S2.SS1 "2.1 Highway hierarchy ‣ 2 Highway Hierarchies algorithm on dynamic directed graphs"), works with both undirected and directed graphs. However, storing all neighbourhoods $`N_{H}^{l}\hspace{0pt}{(v)}`$ for each $`v`$ and $`l`$ has prohibitive memory requirements. Thus, the original HH implementation for checking whether a vertex $`v`$ is in $`N_{H}^{l}\hspace{0pt}{(u)}`$ is based on comparing the distance $`d\hspace{0pt}{(u,v)}`$ with the “distance-to-border” (also called slack) from $`u`$ to the border of its neighbourhood $`N_{H}^{l}\hspace{0pt}{(u)}`$. The “distance-to-border” $`d_{H}^{l}\hspace{0pt}{(u)}`$ is a measure of a neighbourhood’s radius, and is defined as the distance $`d\hspace{0pt}{(u,v)}`$ where $`v`$ is the farthest node in $`N_{H}^{l}\hspace{0pt}{(u)}`$, i.e. the cost of the shortest path from $`u`$ to the $`H`$-th settled node when applying Dijkstra’s algorithm on node $`u`$ at level $`l`$. This is the basis of the slack-based method in \[[Sch05](#bib.bibx16)\], p. 19 (from which we draw our notation). In the partial shortest paths tree $`B\hspace{0pt}{(s_{0})}`$ computed in Step [2a](#S2.I2.i2.I1.i1 "item 2a ‣ item 2 ‣ 2.1 Highway hierarchy ‣ 2 Highway Hierarchies algorithm on dynamic directed graphs") of the algorithm in Section [2.1](#S2.SS1 "2.1 Highway hierarchy ‣ 2 Highway Hierarchies algorithm on dynamic directed graphs"), the slack $`\Delta\hspace{0pt}{(u)}`$ is recursively computed for all $`u \in {B\hspace{0pt}{(s_{0})}}`$ starting from the leaves $`t_{0}`$ of $`B\hspace{0pt}{(s_{0})}`$, as follows.
1. 1\.
Initialise a FIFO queue $`Q`$ to contain all nodes $`u`$ of $`B\hspace{0pt}{(s_{0})}`$, ordering them from the farthest one to the nearest one with respect to $`s_{0}`$.
2. 2\.
Set $`{\Delta\hspace{0pt}{(u)}} = {d_{H}^{l}\hspace{0pt}{(u)}}`$ for $`u`$ a leaf of $`B\hspace{0pt}{(s_{0})}`$ and $`+ \infty`$ otherwise.
3. 3\.
If $`Q`$ is empty, terminate.
4. 4\.
Remove $`u`$ from $`Q`$, and let $`p`$ be its predecessor in $`B\hspace{0pt}{(s_{0})}`$.
5. 5\.
If $`{\Delta\hspace{0pt}{(p)}} = {+ \infty}`$ and $`p \notin {N_{H}^{l}\hspace{0pt}{(s_{0})}}`$, $`p`$ is added to $`Q`$.
6. 6\.
Let $`{\Delta\hspace{0pt}{(p)}} = {\min{({\Delta\hspace{0pt}{(p)}},{{\Delta\hspace{0pt}{(u)}} - {d\hspace{0pt}{(p,u)}}})}}`$.
7. 7\.
If $`{\Delta\hspace{0pt}{(p)}} < 0`$, the edge $`(p,u)`$ is lifted to the higher hierarchical level.
8. 8\.
Return to Step [3](#S2.I4.i3 "item 3 ‣ 2.2 Extension to directed graphs ‣ 2 Highway Hierarchies algorithm on dynamic directed graphs").
The algorithm works because Thm. 2 in \[[Sch05](#bib.bibx16)\] proves that condition $`{\Delta\hspace{0pt}{(p)}} < 0`$ is equivalent to the condition of Step [2b](#S2.I2.i2.I1.i2 "item 2b ‣ item 2 ‣ 2.1 Highway hierarchy ‣ 2 Highway Hierarchies algorithm on dynamic directed graphs") of the algorithm in Section [2.1](#S2.SS1 "2.1 Highway hierarchy ‣ 2 Highway Hierarchies algorithm on dynamic directed graphs"). The cited theorem is based on the following assumption:
<table>
<tbody>
<tr>
<td></td>
<td>∀u∈V(u∉NHl(t0)→dHl(t0)−d(u,t0)<0).for-all𝑢𝑉𝑢superscriptsubscript𝑁𝐻𝑙subscript𝑡0→superscriptsubscript𝑑𝐻𝑙subscript𝑡0𝑑𝑢subscript𝑡00\forall u\in V\;(u\not\in N_{H}^{l}(t_{0})\rightarrow d_{H}^{l}(t_{0})-d(u,t_{0})<0).</td>
<td></td>
<td>(2)</td>
</tr>
</tbody>
</table>
This condition may fail to hold for directed graphs, since $`{d\hspace{0pt}{(u,t_{0})}} \neq {d\hspace{0pt}{(t_{0},u)}}`$.
To make Assumption [2](#S2.E2 "In 2.2 Extension to directed graphs ‣ 2 Highway Hierarchies algorithm on dynamic directed graphs") hold, we have to consider a neighbourhood radius computed on the reverse graph, that is the graph $`\overline{G} = {(V,\overline{A})}`$ such that $`{{(u,v)} \in \overline{A}}\Leftrightarrow{{(v,u)} \in A}`$. Thus, we modified the original implementation to compute, for each node, a reverse neighbourhood $`{\overline{N}}_{H}^{l}\hspace{0pt}{(v)}`$ (see Figure [2](#S2.F2 "Figure 2 ‣ 2.2 Extension to directed graphs ‣ 2 Highway Hierarchies algorithm on dynamic directed graphs")), so that we are able to store the corresponding reverse neighbourhood radius $`{{\overline{d}}_{H}^{l}\hspace{0pt}{(u)}\hspace{0pt}{\forall u}} \in V`$. We replace Step [2](#S2.I4.i2 "item 2 ‣ 2.2 Extension to directed graphs ‣ 2 Highway Hierarchies algorithm on dynamic directed graphs") in the algorithm above by:
> [2](#S2.I4.i2 "item 2 ‣ 2.2 Extension to directed graphs ‣ 2 Highway Hierarchies algorithm on dynamic directed graphs")a. Set $`{\Delta\hspace{0pt}{(u)}} = {{\overline{d}}_{H}^{l}\hspace{0pt}{(u)}}`$ for $`u`$ a leaf of $`B\hspace{0pt}{(s_{0})}`$ and $`+ \infty`$ otherwise.
We are now going to prove our key lemma.
######
2.2 Lemma
Let $`{u,s} \in V`$ and $`t`$ a leaf in $`B\hspace{0pt}{(s)}`$. If $`u \notin {{\overline{N}}_{H}^{l}\hspace{0pt}{(t)}}`$ then $`{{{\overline{d}}_{H}^{l}\hspace{0pt}{(t)}} - {d\hspace{0pt}{(u,t)}}} < 0`$.
###### Proof.
Suppose $`{d\hspace{0pt}{(u,t)}} \leq {{\overline{d}}_{H}^{l}\hspace{0pt}{(t)}}`$. By definition, this means that there is a shortest path in $`{\overline{N}}_{H}^{l}\hspace{0pt}{(t)}`$ which connects $`u`$ to $`t`$. Therefore, $`u \in {{\overline{N}}_{H}^{l}\hspace{0pt}{(t)}}`$ against the hypothesis. ∎
It is now straightforward to amend Thm. 2 in \[[Sch05](#bib.bibx16)\] to hold in the directed case; all other theorems in \[[Sch05](#bib.bibx16)\] need similar modifications, replacing $`N_{H}^{l}\hspace{0pt}{(t)}`$ with $`{\overline{N}}_{H}^{l}\hspace{0pt}{(t)}`$ and $`d_{H}^{l}\hspace{0pt}{(t)}`$ with $`{\overline{d}}_{H}^{l}\hspace{0pt}{(t)}`$ whenever $`t`$ is target node or is “on the right side” of a path - it will always be clear from the context. The query algorithm must me modified to cope with these differences, using $`{\overline{d}}_{H}^{l}\hspace{0pt}{(t)}`$ instead of $`d_{H}^{l}\hspace{0pt}{(t)}`$ whenever we are searching in the backwards direction.
[Refer to caption](/html/0704.1068/assets/x3.png)
Figure 2: An example which shows neighbourhoods and reverse neighbourhoods with $`H = 3`$; only solid arcs are lifted to a higher level in the hierarchy. Note that arcs $`(p,t)`$ and $`(p',t)`$ are not lifted even if $`{p,p'} \notin {N_{H}^{l}\hspace{0pt}{(t)}}`$; this is because $`{p,p'} \in {{\overline{N}}_{H}^{l}\hspace{0pt}{(t)}}`$, and for target node we consider the reverse neighbourhood.
Interestingly, the problem with the slack-based method was first detected when our original implementation of the HH algorithm failed to construct a correct hierarchy for the Paris urban area. This shows that the extension of the algorithm to the directed case actually arises from real needs.
### 2.3 Heuristic application to dynamic networks
The original Highway Hierarchies algorithm, as described above, finds shortest paths in networks whose arc weights do not change in time. By forsaking the optimality guarantee, we adapt the algorithm to the case of networks whose arc weights are updated in quasi real-time. Whereas the highway hierarchy is constructed using the static arc travelling times from the road network database, each point-to-point path query is deployed on a dynamically evaluated version of the highway hierarchy where the arcs are weighted using the quasi real-time traffic information. In particular, in all tests that involved a comparison with neighbourhood radius we use the static arc travelling times, while for all evaluations of path lengths or of node distances we use the real-time (dynamic) travelling times. This means that the static travelling times are used to determine neighbourhood’s crossings, and thus to determine when to switch to a higher level in the hierarchy, while the key for the priority queue for HH algorithm is computed using only dynamic travelling times.
## 3 Computational results
In this section we discuss the computational results obtained by our implementation. As there seems to be no other readily available software with equivalent functionality, the computational results are not comparative. However, we establish the quality of the heuristic solutions by comparing them against the fastest paths found by a plain Dijkstra’s algorithm. We mention here two different approaches: dynamic highway-node routing (\[[SS07](#bib.bibx19)\]), which uses a selection of nodes operated by the HH algorithm to build an overlay graph (see \[[HSW06](#bib.bibx13)\]), and dynamic ALT (\[[DW07](#bib.bibx10)\]), which is a dynamic landmark-based implementation of $`A^{\ast}`$. Both approaches, however, although very performing with respect to query times, require a computationally heavy update phase (which takes time in the order of minutes), and thus are not suitable for our scenario, where, supposedly, each arc can have its cost changed every 2 minutes (roughly).
We performed the tests on the entire road network of France, using a highway hierarchy with $`H = 65`$ and $`L = 9`$. The original network has $`7778913`$ junctions and $`17154042`$ road segments; the number of nodes and arcs in each level is as follows.
<table>
<thead>
<tr>
<th>level</th>
<th>0</th>
<th>1</th>
<th>2</th>
<th>3</th>
<th>4</th>
<th>5</th>
<th>6</th>
<th>7</th>
<th>8</th>
<th>9</th>
</tr>
</thead>
<tbody>
<tr>
<th>nodes</th>
<td>7778913</td>
<td>1517291</td>
<td>433286</td>
<td>182474</td>
<td>91888</td>
<td>53376</td>
<td>34116</td>
<td>23338</td>
<td>16445</td>
<td>11790</td>
</tr>
<tr>
<th>arcs</th>
<td>17154042</td>
<td>3461385</td>
<td>1283000</td>
<td>583380</td>
<td>308249</td>
<td>183659</td>
<td>119524</td>
<td>81170</td>
<td>57235</td>
<td>41092</td>
</tr>
</tbody>
</table>
We show the results for queries on the full graph without dynamic travelling times in Table [1](#S3.T1 "Table 1 ‣ 3 Computational results"); in this case, all paths computed with the HH algorithm are fastest paths. In Table [2](#S3.T2 "Table 2 ‣ 3 Computational results"), instead, we record our results on a graph with dynamic travelling times; we also report the relative distance of the solution found with our heuristic version of the HH algorithm and the fastest path computed with Dijkstra, and, for comparative reasons, the results of the naive approach which consists in computing the traffic-free optimal solution with the HH algorithm (i.e., on the static graph) and then applying dynamic times on the so-found solution. Dynamic travelling times were taken choosing, for each query, one out of five sets of values recorded in different times of the day for each of the $`29384`$ arcs with dynamic information.
Although this number is small with respect to the total number of arcs in the graph, it should be noted that most of these arcs correspond to very important road segments (highways and national roads). All arcs $`(i,j)`$ that did not have a dynamic travelling time were assigned a different weight at each query, chosen at random with a uniform distribution over $`\lbrack\tau_{i\hspace{0pt}j},{15\hspace{0pt}\tau_{i\hspace{0pt}j}}\rbrack`$, where $`\tau_{i\hspace{0pt}j}`$ is the reference time for arc $`(i,j)`$. This choice has been made in order to recreate a difficult scenario for the query algorithm: even if the number of arcs with real traffic information is still small, it is going to increase rapidly as the means for obtaining dynamic information increase (e.g. number of road cameras, etc.), and thus, to simulate an instance where most arcs have their travelling time changed several times per hour, we generated each arc’s cost at random. The interval $`\lbrack\tau_{i\hspace{0pt}j},{15\hspace{0pt}\tau_{i\hspace{0pt}j}}\rbrack`$ is simply a rough estimation of a likely cost interval, based on the analysis of historical data. All tables report average values over 5000 queries. All computational results in Table [1](#S3.T1 "Table 1 ‣ 3 Computational results") and [2](#S3.T2 "Table 2 ‣ 3 Computational results") have been obtained on a multiprocessor Intel Xeon 2.6 GHz with 8GB RAM running Microsoft Windows Server 2003, compiling with Miscrosoft Visual Studio 2005 and optimization level 2.
Computational results show that, although with no guarantee of optimality, our heuristic version of the algorithm works well in practice, with $`0.55\%`$ average deviation from the optimal solution and a recorded maximum deviation of $`17.59\%`$; query times do not seem to be influenced by our changes with respect to the original version of the algorithm. The naive approach of computing the shortest path on the static graph, and then applying dynamic times, records an average error of $`2\%`$, but it has a much higher variance, and a maximum error of $`27.95\%`$; although the average error is not high, it’s still almost $`4`$ times the average error of the more sofisticated approach, and the high variance suggests lack of stability in the solution’s quality. The low value recorded for the average error with the naive approach (in absolute terms) can be explained as a consequence of the following two facts: travelling times generated at random on arcs without real-time traffic information cannot simulate real traffic situation, because they lack spatial coherence (i.e. they do not simulate congested nearby zones) and traffic behaviour information (i.e. the fact that during peak hours important road segments are likely to be congested, while less important roads are not), thus making the task of finding a fast path easier; besides, the average query on such a large graph corresponds to a very long path (296 minutes on the traffic-free graph, 2356 minutes on the dynamic graph), and on long paths it is usually necessary to use highways or national roads regardless of their congestion status - which is exactly what the HH algorithm does. This last sentence is supported by the fact that, if we consider only the $`500`$ shortest queries in terms of path length, the average error of the naive approach increases to $`3.60\%`$, while the average error of the heuristic version increases to $`0.97\%`$; this is in accord with the fact that on short paths the influence of traffic is greater, because alternative routes that do not use highways are more appealing, while on long paths using highways is often a necessary step. However, in relative terms, the heuristic version of the HH algorithm performs significantly better than the naive approach proposed for comparison, and we expect the difference to increase (in favour of the heuristic algorithm) if applied to a graph fully covered with real traffic information.
Figure [3](#S3.F3 "Figure 3 ‣ 3 Computational results") shows how the optimal and the heuristic path may differ; since the hierarchy built on the static graph emphasizes important roads, the heuristic algorithm applied on the dynamically weighted graph still tends to use highways and national roads even when they are congested (up to a certain degree), thus sometimes losing optimality.
<table>
<tbody>
<tr>
<th></th>
<td>Dijkstra’s algorithm</td>
<td>HH algorithm</td>
</tr>
<tr>
<th># settled nodes</th>
<td>2275563</td>
<td>18966</td>
</tr>
<tr>
<th># explored nodes</th>
<td>2587112</td>
<td>36200</td>
</tr>
<tr>
<th>query time [sec]</th>
<td>11.830</td>
<td>0.099</td>
</tr>
</tbody>
</table>
Table 1: Computational results on the static graph: average values
<table>
<tbody>
<tr>
<th></th>
<td>Dijkstra’s algorithm</td>
<td>HH algorithm</td>
<td>HH algorithm</td>
</tr>
<tr>
<th></th>
<td></td>
<td>naive approach</td>
<td>heuristic version</td>
</tr>
<tr>
<th># settled nodes</th>
<td>2280872</td>
<td>19174</td>
<td>19099</td>
</tr>
<tr>
<th># explored nodes</th>
<td>2594361</td>
<td>36581</td>
<td>36492</td>
</tr>
<tr>
<th>query time [sec]</th>
<td>11.917</td>
<td>0.100</td>
<td>0.099</td>
</tr>
<tr>
<th>distance from optimum (variance)</th>
<td>0%</td>
<td>2.00% (5.00)</td>
<td>0.55% (0.45)</td>
</tr>
</tbody>
</table>
Table 2: Computational results on the graph with dynamic times: average values
[Refer to caption](/html/0704.1068/assets/x4.png)
Figure 3: Fast paths calculated with different algorithms; each number identifies a path, paths are partially overlapping. 1: Dijkstra’s algorithm (optimal solution) with real-time arc costs; dynamic travelling time: 24 minutes and 6 seconds. 2: HH algorithm (heuristic solution) with real-time arc costs; dynamic travelling time: 25 minutes and 5 seconds. 3: HH algorithm without real-time arc costs (traffic-free optimal solution); dynamic travelling time: 37 minutes and 5 seconds.
[Refer to caption](/html/0704.1068/assets/x5.png)
Figure 4: Highway hierarchy near the Champs Elysées, Paris; colour intensity and line width increase with hierarchy level, in that order (a wide line with a lighter colour has a higher hierarchy level than a thin line with dark colour).
## 4 Conclusion
We present a heuristic algorithm for efficiently finding fast paths in large-scale partially dynamically weighted road networks, and benchmark its application on real-world data. The proposed solution is based on fast multi-level bidirectional Dijkstra queries on a highway hierarchy built on the statically weighted version of the network using the Highway Hierarchies algorithm, and deployed using the dynamic arc weights. Computational results show that, although with no guarantee of optimality, the proposed solution works well in practice, computing near-optimal fast paths quickly enough for our purposes.
## Acknowledgements
We are grateful to Ms. Annabel Chevaux, Mr. Benjamin Simon and Mr. Benjamin Becquet for invaluable practical help with Oracle and the real data, and to the rest of the Mediamobile’s energetic and youthful staff for being always friendly and helpful.
## References
- \[AOPS03\] R.K. Ahuja, J.B. Orlin, S. Pallottino, and M.G. Scutellà. Dynamic shortest paths minimizing travel times and costs.
Networks, 41(4):197–205, 2003.
- \[BRTed\] L.S. Buriol, M.G.C. Resende, and M. Thorup. Speeding up dynamic shortest path algorithms.
INFORMS Journal on Computing, submitted.
- \[CH66\] K.L. Cooke and E. Halsey. The shortest route through a network with time-dependent internodal transit times.
Journal of Mathematical Analysis and Applications, 14:493–498, 1966.
- \[Cha98\] I. Chabini. Discrete dynamic shortest path problems in transportation applications: complexity and algorithms with optimal run time.
Transportation Research Records, 1645:170–175, 1998.
- \[CS02\] I. Chabini and L. Shan. Adaptations of the $`A^{\ast}`$ algorithm for the computation of fastest paths in deterministic discrete-time dynamic networks.
IEEE Transactions on Intelligent Transportation Systems, 3(1):60–74, 2002.
- \[CZ01\] E.P.F. Chan and N. Zhang. Finding shortest paths in large network systems.
In GIS ’01: Proceedings of the 9th ACM international symposium on Advances in geographic information systems, pages 160–166, New York, NY, USA, 2001. ACM Press.
- \[Dea04\] B.C. Dean. Shortest paths in fifo time-dependent networks: theory and algorithms. Technical report, MIT, Cambridge MA, 2004.
- \[Dij59\] E.W. Dijkstra. A note on two problems in connexion with graphs.
Numerische Mathematik, 1:269–271, 1959.
- \[Dre69\] S.E. Dreyfus. An appraisal of some shortest-path algorithms.
Operations Research, 17(3):395–412, 1969.
- \[DW07\] D. Delling and D. Wagner. Landmark-based routing in dynamic graphs.
In WEA 2007, volume 4525 of Lecture Notes in Computer Science. Springer, 2007.
- \[FHK<sup>+</sup>05\] T. Flatberg, G. Hasle, O. Kloster, E.J. Nilssen, and A. Riise. Dynamic and stochastic aspects in vehicle routing – a literature survey. Technical Report STF90A05413, SINTEF, Oslo, Norway, 2005.
- \[GKW05\] A.V. Goldberg, H. Kaplan, and R.F. Werneck. Reach for $`A^{\ast}`$: Efficient point-to-point shortest path algorithms. Technical Report MSR-TR-2005-132, Microsoft Research, 2005.
- \[HSW06\] M. Holzer, F. Schulz, and D. Wagner. Engineering multi-level overlay graphs for shortest-path queries.
In SIAM, volume 129 of Lecture Notes in Computer Science, pages 156–170. Springer, 2006.
- \[Ker04\] B. S. Kerner.
The Physics of Traffic.
Springer, Berlin, 2004.
- \[NV05\] TeleAtlas NV.
Tele Atlas Multinet ShapeFile 4.3.1 Format Specifications.
TeleAtlas NV, May 2005.
- \[Sch05\] D. Schultes. Fast and exact shortest path queries using highway hierarchies.
Master Thesis, Informatik, Universität des Saarlandes, June 2005.
- \[SS05\] P. Sanders and D. Schultes. Highway hierarchies hasten exact shortest path queries.
In G. Stølting Brodal and S. Leonardi, editors, ESA, volume 3669 of Lecture Notes in Computer Science, pages 568–579. Springer, 2005.
- \[SS06\] P. Sanders and D. Schultes. Engineering highway hierarchies.
In ESA 2006, volume 4168 of Lecture Notes in Computer Science, pages 804–816. Springer, 2006.
- \[SS07\] P. Sanders and D. Schultes. Dynamic highway-node routing.
In WEA 2007, volume 4525 of Lecture Notes in Computer Science, pages 66–79. Springer, 2007. |
0706.0430 | cs.CR | cyber | <sup>1</sup>
<sup>1</sup>institutetext: Computer Laboratory
JJ Thomson Avenue, Cambridge CB3 0FD, UK
shishir.nagaraja@cl.cam.ac.uk
# Anonymity in the wild: Mixes on unstructured networks
Shishir Nagaraja
###### Abstract
As decentralized computing scenarios get ever more popular, unstructured topologies are natural candidates to consider running mix networks upon. We consider mix network topologies where mixes are placed on the nodes of an unstructured network, such as social networks and scale-free random networks. We explore the efficiency and traffic analysis resistance properties of mix networks based on unstructured topologies as opposed to theoretically optimal structured topologies, under high latency conditions. We consider a mix of directed and undirected network models, as well as one real world case study – the LiveJournal friendship network topology. Our analysis indicates that mix-networks based on scale-free and small-world topologies have, firstly, mix-route lengths that are roughly comparable to those in expander graphs; second, that compromise of the most central nodes has little effect on anonymization properties, and third, batch sizes required for warding off intersection attacks need to be an order of magnitude higher in unstructured networks in comparison with expander graph topologies.
## 1 Introduction
As governments pursue large scale surveillance and censorship programs, anonymity in online communication mechanisms is an increasingly important requirement. Anonymous communications are also useful in building resistance against a global passive adversary who can subject the targets to traffic analysis. Often, an attacker will try to destabilize a network by building a dossier of the most central nodes, and attacking ones on the top of the list. Traffic analysis of inter-node communication offers basic tools to collect necessary intelligence in order to plan an attack.
Seminal work by Chaum \[[Cha81](#bib.bibx5)\] introduced mix networks as a technique to provide anonymous communications where messages are relayed through a sequence of intermediate nodes called mixes, to make the task of tracing them through the network as difficult as possible. The essential idea is to make the inputs of each mix bit-wise unlinkable to its outputs.
Anonymity research conducted since, can be classified into low-latency or real time systems primarily for Internet browsing such as onion routing \[[STRL00](#bib.bibx31)\] and high-latency or non-real time systems such as mixminion \[[DDM03](#bib.bibx8)\].
The topology of a mix network plays an important role in its efficiency and traffic analysis resistance properties. The mainstream design paradigm that has emerged so far is that of structured network topologies based on regular graphs. The theory is that such topologies are amenable to theoretical analysis that proves they have optimal expansion properties. This leads to a mix network design that is highly efficient and resistant to traffic analysis. Examples are onion-routing systems such as TOR \[[DMS04](#bib.bibx10)\]that use a complete graph topology, where a mix can contact every other mix in the network. While such models are theoretically elegant, the assumption that every node in the network is equally resourced (as regular graphs necessitate) to handle network traffic loads is their main drawback.
An alternate paradigm is topology based on unstructured networks, such as those inspired from social networks. The argument in their favor being that the incentive to carry traffic is clear and simple - friends carry each-others traffic. Moreover, no additional resources go into constructing an overlay network since the pre-existing topology is used by the mix network as well, which works well for power constrained environments such as adhoc networks and sensor networks. Legal considerations play an important role too. It is not enough to merely have a large number of mixes. When hassled by legal requests (such as a subpoena to hand-over mix server logs to the police), a mix-network where friends route each others traffic, is likely to have a higher proportion of servers in operation, as opposed to a synthetic network.
A comparison between the two paradigms needs to address mix-network efficiency, resilience to corrupt nodes and the loss of anonymity from statistical disclosure attacks.
In this paper we analyze various types of unstructured networks, especially social networks and evaluate their suitability as mix topologies. We discuss the reasons behind using social networks to route mix traffic and we analyze the suitability of various types of model networks to routing mix traffic and offer a comparison between them. We also analyze the theoretical bounds on anonymity such networks can provide in terms of mixing speed and resistance to traffic analysis. We apply concepts from spectral graph theory to derive the route length necessary to provide maximal anonymity.
This paper is organized as follows: Section 3 discusses the various topologies used in our analysis. Section 4, lays out the evaluation framework to measure the traffic analysis resistance of various topologies. Section 5 discusses the application of the framework to various topologies and the results obtained. Finally, we offer our conclusions in section 6.
## 2 Related work
Danezis \[[Dan03](#bib.bibx7)\] explored the anonymity provided by expander graph topologies, this is one of the main sources of inspiration for our work. He established the thoretical bounds of anonymity for expander graphs, and also showed that they were optimal.
Borisov \[[Bor05](#bib.bibx3)\] analyzes anonymous communications over a De Bruijn graph topology overlay network. He analyzes the deBruijn graph topology and comments on their successful mixing capabilities.
## 3 Network models
In this section we give a brief introduction to the network models we wish to analyze as candidates for mix network topologies.
### 3.1 Erdös-Rényi model of random networks
On the earliest models for heterogeneous networks is the Erdös-Rényi (ER )model \[[ER59](#bib.bibx12)\]. Although seldom found in real world networks, their use has been popularised by the work of Eschenauer and Gligor \[[EG02](#bib.bibx11)\] is designing a key management scheme for sensor networks.
Here, we start from $`N`$ vertices without any edges. Subsequently, edges connecting two randomly chosen vertices are added as the result of a Bernoulli trial, with a parameter $`p`$. It generates random networks with no particular structural bias. The average degree $`{\langle k\rangle} = {{2\hspace{0pt}L}/N}`$ where $`L`$ is the total number of edges, can also be used as a control parameter. ER model networks have a logarithmically increasing $`l`$, a normal degree distribution, and a clustering coefficient close to zero.
### 3.2 Scale-free networks with linear preferential attachment
A number of popular peer-to-peer systems are found to have heterogeneous topologies with heavy tailed degree distributions. The work of Ripeanu \[[RFI02](#bib.bibx26)\] shows that two popular systems, Gnutella \[[KM02](#bib.bibx16)\] and Freenet \[[CSWH00](#bib.bibx6)\], have power-law degree distributions.
A variable X is said to follow a heavy tail distribution if $`{P\hspace{0pt}r\hspace{0pt}{\lbrack{X > x}\rbrack}} \sim {x^{- k}\hspace{0pt}L\hspace{0pt}{(x)}}`$ where $`k \in \Re^{+}`$ and $`L\hspace{0pt}{(x)}`$ is a slowly varying function so that $`{\lim_{x\rightarrow\infty}\frac{L\hspace{0pt}{({t\hspace{0pt}x})}}{L\hspace{0pt}{(x)}}}\rightarrow 1`$. A power-law distribution is simply a variation of the above where one studies $`{P\hspace{0pt}r\hspace{0pt}{\lbrack{X = x}\rbrack}} \sim x^{- {({k + 1})}} = x^{- \alpha}`$. The degree of a node is the number of links it has to other nodes in the network. If the degree distribution of a network follows a power-law distribution it is known as a scale-free network. The power-law in the degree or link distribution reflects the presence of central individuals who interact with many others on a continual basis and play a key role in relaying information.
We denote a scale-free network generated by preferential attachment, by $`G_{m,N}\hspace{0pt}{(V,E)}`$ where $`m`$ is the number of initial nodes created at time=$`t_{0}`$ and $`N`$ is the total number of nodes in the network. At every time step $`{t_{i},i} \geq 0`$, one node is added to the network. For every node $`v`$ added, we create $`m`$ edges from the $`v`$ to existing nodes in the network according to the following linear preferential attachment function due to Barabasi and Albert \[[AB02](#bib.bibx1)\]:
<table>
<tbody>
<tr>
<td></td>
<td>Pr[(v,i)]=ki/∑jkj𝑃𝑟delimited-[]𝑣𝑖subscript𝑘𝑖subscript𝑗subscript𝑘𝑗Pr[(v,i)]=k_{i}/\sum_{j}k_{j}</td>
<td></td>
</tr>
</tbody>
</table>
where $`k_{i}`$ is the degree of node $`i`$. We continue until $`{|V|} = N`$.
### 3.3 Scale-free random graph topology
An alternate way of constructing a large scale-free network is to create a network with a given power-law degree sequence that is random in all other aspects. Aiello et.al. \[[ACL00](#bib.bibx2)\] propose such a random graph model inspired by massive AT&T call graphs, with two parameters $`\alpha`$ and $`\beta`$. Where, $`\alpha`$ gives the fraction of nodes with degree 1 and $`\beta`$ defines the exponent of the power-law function. Then, if $`y`$ be the number of vertices of degree $`x > 0`$, $`x`$ and $`y`$ satisfy $`{l\hspace{0pt}o\hspace{0pt}g\hspace{0pt}{(y)}} = {\alpha - {\beta\hspace{0pt}l\hspace{0pt}o\hspace{0pt}g\hspace{0pt}{(x)}}}`$.
### 3.4 Klienberg-Watts-Strogatz(KWS) small world topology
Our next network model is inspired by the network of social contacts. It is well known that any two people are linked by a chain of half a dozen others who are pairwise acquainted – known as the ‘small-world’ phenomenon. This idea was popularised by Milgram in the 60s \[[Mil67](#bib.bibx17)\].
The KWS graph topology models a small world network that encapsulates the following: a network rich in local connections, with a few long range connections. The network generation starts from a N by N lattice each point representing an individual in a social network. The lattice distance $`{d\hspace{0pt}{({(i,j)},{(k,l)})}} = {{|{k - i}|} + {|{l - j}|}}`$. For a parameter $`p`$, every node $`u`$ has a directed link to every other node $`v`$ within $`{d\hspace{0pt}{(u,v)}} \leq p`$. For parameters $`q`$ and $`r`$, we construct $`q`$ long range directed links from $`u`$ to a node $`v`$ with a probability distribution $`{\lbrack{P\hspace{0pt}r\hspace{0pt}{(u,v)}}\rbrack} = \frac{{(d{(u,v)})}^{(} - r)}{\sum_{v}{(d{(u,v)})}^{(} - r)}`$.
Low $`r`$ values means long-range connections, whereas higher values lead to preferential connections in the vicinity of $`u`$.
### 3.5 LiveJournal (LJ)
In order to test our ideas on a real world unstructured network, we turned to a large-scale social network called LiveJournal (LJ). LiveJournal is a social networking and blogging site with several million members and a large collection of user defined communities. LiveJournal allows members to maintain journals, individual and group blogs, and – most importantly for our study here – it allows people to declare which other members are their friends. Using a web crawler called touchgraph (http://www.touchgraph.com), we traced the LJ network to the online friendship network. The snapshot of the network we use in our analysis has 3,746,240 nodes and 27,430,000 edges.
A mix server bundled along with a future LiveJournal client acts as the basis of mix deployment. Mix circuits are built on top of the social network topology.
### 3.6 Expander graphs
Danezis \[[Dan03](#bib.bibx7)\] previously analyzed the use of expander graph topologies to construct mix networks. Expanders are well known to have excellent expansion properties. We include this as a baseline comparison against theoretical structured topologies. An expander graph $`G_{N,D}`$ has a homogeneous topology with $`N`$ nodes each with a degree $`D`$.
## 4 Evaluation framework for measuring traffic analysis resistance
Before we set out the evaluation framework, we first clarify what we mean by “anonymity” in this paper. The focus of this work is on message anonymity \[[SD02](#bib.bibx29)\]: given a message, the attacker should not be able to determine who sent it to whom. There are other definitions such as relationship anonymity defined by Pfitzmann et. al. \[[PH00](#bib.bibx24)\].
The objective of our analysis is to determine how the topology of a mix network affects the amount of effort on the attacker’s part to uniquely identify communication endpoints using traffic analysis attacks alone. The effectiveness of such attacks depends heavily on the topology of the underlying network. If the attacker is not able to reduce anonymity beyond his or her initial knowledge then the mix network is said to be resistant to traffic analysis attacks under the given threat model.
The attacker might also employ side channel analysis on the end-points before the data enters the mix network, we do not consider such attacks here. Side channel information might be timestamps or other information related to the protocol or mechanism in use. Attacks using such information can be used to link messages to the communication end-points, and are known as traffic confirmation attacks \[[RSG98](#bib.bibx28)\], their effectiveness depends on the mixes’ batching and flushing strategy.
### 4.1 Threat Model
Throughout this paper we consider the adversarial context of a global passive adversary.
### 4.2 Measuring anonymity
There are several ways one can express the anonymity a system provides. In our analysis we use a quantitative method due to Serjantov and Danezis \[[SD02](#bib.bibx29)\], based on the following definition: “Anonymity of a system may be defined as the amount of information the attacker is missing to uniquely identify an actor’s link to an action”. In information theoretic terms, the anonymity of the system $`\mathcal{A}`$, is the entropy $`\mathcal{E}`$, of the probability distribution over all the actors $`\alpha_{i}`$, that they committed a specific action.
<table>
<tbody>
<tr>
<td></td>
<td>𝒜=ℰ[αi]=−∑iPr[αi]log2Pr[αi]𝒜ℰdelimited-[]subscript𝛼𝑖subscript𝑖𝑃𝑟delimited-[]subscript𝛼𝑖𝑙𝑜subscript𝑔2𝑃𝑟delimited-[]subscript𝛼𝑖\mathcal{A}=\mathcal{E}[\alpha_{i}]=-\sum_{i}Pr[\alpha_{i}]log_{2}Pr[\alpha_{i}]</td>
<td></td>
<td>(1)</td>
</tr>
</tbody>
</table>
This gives the number of bits of information, with a negative sign, that the attacker is missing before they can uniquely identify a sender or a receiver.
### 4.3 Modeling mix route selection
In order to understand the maximal anonymity provided by a mix network we use Markov chains to model the route selection process, as they closely match the way mixes are selected to form a mix route.
The process of selecting a mix route of length $`k`$ by selecting $`k`$ random nodes in the mix network, is equivalent to first selecting a random mix node, and, then a random neighbour of the first mix, repeating this process $`k - 2`$ times. Hence we may model the route selection process as a random walk on the underlying graph, with the various states of the Markov chain process being the mix nodes of the network.
### 4.4 Measuring mix network efficiency
In analyzing the anonymity provided by a particular network topology we need to examine the probability that a specific message is at a particular node at a certain time. In order to link the sender and the receiver to a particular message, the attacker must retrace the steps taken by the message through the mix network starting from the receiver. Let the mix network be an undirected graph $`G\hspace{0pt}{(V,E)}`$. If messages $`m_{i\hspace{0pt}j}`$ are inserted at node $`i`$ destined for $`j`$, then for a message $`m_{x}^{t}`$ at node $`x`$ at time $`t`$, the attacker must link $`m_{x}^{t}`$ to $`m_{i\hspace{0pt}j}`$. Note that $`m_{x}^{t}`$ might either be in the edge or the core of the mix network.
Applying the above mentioned information theoretic metric we have:
<table>
<tbody>
<tr>
<td></td>
<td>𝒜=ℰ(pij)𝒜ℰsubscript𝑝𝑖𝑗\mathcal{A}=\mathcal{E}(p_{ij})</td>
<td></td>
</tr>
</tbody>
</table>
where $`p_{i\hspace{0pt}j} = {P\hspace{0pt}r\hspace{0pt}{\lbrack{m_{x}^{t}\hspace{0pt}i\hspace{0pt}s\hspace{0pt}m_{i\hspace{0pt}j}}\rbrack}}`$ is the probability distribution over all the nodes in $`V`$.
Suppose a message is inserted into the mix network through a randomly chosen node. Then after an infinite number of steps, the probability that the message is present on any randomly chosen node in the network is given by stationary distribution of the Markov chain $`\pi`$. Let $`q^{(0)}`$ be the initial probability distribution describing the node on which message $`m`$ is introduced into the mix network, this is equivalent to the distribution of input load across the nodes in network. $`q^{(t)}`$ then, is the probability distribution of the node on which the message is present after $`t`$ steps. (this is also known as the state probability vector of the Markov chain at time $`t \geq 0`$). With increasing $`t`$ one would like to see that $`q^{(t)}`$ merges with $`\pi`$. The rate at which this takes place is known as the convergence rate of the Markov chain, and the difference itself is called the relative point-wise distance defined as:
<table>
<tbody>
<tr>
<td></td>
<td>Δ(t)=maxi|qit−πi|πiΔ𝑡𝑚𝑎subscript𝑥𝑖superscriptsubscript𝑞𝑖𝑡subscript𝜋𝑖subscript𝜋𝑖\Delta(t)=max_{i}\frac{|q_{i}^{t}-\pi_{i}|}{\pi_{i}}</td>
<td></td>
<td>(2)</td>
</tr>
</tbody>
</table>
The smaller the relative point-wise distance, faster the convergence, and more efficient the mix network. It is now easy to see that the maximum anonymity the network can provide is the entropy of the stationary distribution of the chain.
<table>
<tbody>
<tr>
<td></td>
<td>𝒜network=ℰ(π)subscript𝒜𝑛𝑒𝑡𝑤𝑜𝑟𝑘ℰ𝜋\mathcal{A}_{network}=\mathcal{E}(\pi)</td>
<td></td>
<td>(3)</td>
</tr>
</tbody>
</table>
.
When $`P`$ is the transition matrix of the chain it is well known that $`P`$ has $`n`$ real eigen-vectors $`\pi_{i}`$ and $`n`$ eigenvalues $`\lambda_{i}`$ \[[Wes01](#bib.bibx32)\].
By using the relation $`q^{(t)} = q^{(0)}P^{(}t)`$, we calculate the probability distribution of a message being on a node after having transited a mix route of length $`t`$.
### 4.5 Compromised mixes
Suppose a subset of mixes are taken over by an adversary. Then a compromised mix route is defined as a mix circuit that is solely composed of compromised mix nodes. Then, what is the probability that a randomly chosen mix route is compromised?
A network topology with poor expansion properties (or lower *eigen-value gap* $`\epsilon = {1 - \lambda_{2}}`$) tends to have relatively ’localized’ mix routes, so that, given the first mix of a route, there exists a subset of mixes within the network that have a higher chance of being on the route than others.
The spectral theory of graphs lends us a few tools, namely chernoff bounds, in quantifying this risk. Suppose $`S`$ is the set of subverted nodes, and $`\pi_{S}`$ the corresponding probability mass of the stationary distribution $`\pi`$. The upper bound of the probability that a mix route (random walk) of length $`t`$ goes through $`t_{S}`$ nodes of $`S`$ is given by Gilbert \[[Gil98](#bib.bibx13)\]: $`{P\hspace{0pt}r\hspace{0pt}{\lbrack{t_{A} = t}\rbrack}} \leq {\left( {1 + \frac{{({1 - {\pi\hspace{0pt}{(A)}}})}\hspace{0pt}\epsilon}{10}} \right)\hspace{0pt}e^{- {t\hspace{0pt}\frac{{({1 - {\pi\hspace{0pt}{(A)}}})}^{2}\hspace{0pt}\epsilon}{20}}}}`$. However as Danezis \[[Dan03](#bib.bibx7)\] notes, given that this probability exponentially decreases with increase in $`t`$, a small increase in route length will successfully mitigate this risk.
What is more relevant in the context of unstructured networks, is the presence of ’hub’ nodes and ’weak-ties’. Hubs \[[New03](#bib.bibx23)\] are special nodes that owing to their position in the network topology handle large amounts of traffic reducing. Similarly, weak-ties \[[Gra73](#bib.bibx14)\] are edges responsible for significantly reducing average path-lengths in networks of tightly knit communities such as social networks. The risk of compromised mix routes is significantly higher in a topology where hubs only connect to other hubs, and handle most of the network traffic. If an attacker can locate and strategically target mix nodes that also play the role of a hub, then the percentage of mix routes under risk can be significant.
Hence, we simulated a large number of random walks for various topologies presented in section [3](#S3 "3 Network models ‣ Anonymity in the wild: Mixes on unstructured networks"), of different lengths, and make a recommendation on the route length to mitigate this risk in section [5.1](#S5.SS1 "5.1 Simulation parameters ‣ 5 Results and Discussion ‣ Anonymity in the wild: Mixes on unstructured networks").
### 4.6 Intersection attacks
The term intersection attack was introduced by Berthold et.al. \[[BPS00](#bib.bibx4)\]. These attacks involve the detection of the preferential use of a mix route. If for some reason, a sender under attack sends more traffic along a specific route much more often than other routes, then a simple intersection attack is carried out by intersecting the set of possible next-hop mixes of every mix with the set of possible next-hop destinations of previous messages. The the actual path of a message will then become apparent unless the network has countermeasures against observability.
If each link from a mix node is used to flush messages to its neighbours, then the potential for the simplest of intersection attacks can be greatly reduced \[[KAP02](#bib.bibx15)\]. So, for a given node $`i`$, we wish to calculate the probability that any out going link remains unused during a flushing cycle. If each mix node receives $`b`$ messages per batch, then each of these will appear on a particular outgoing link $`j`$ with a binomial probability distribution $`p_{i} = {{1/d}\hspace{0pt}e\hspace{0pt}g_{i}}`$. Danezis \[[Dan03](#bib.bibx7)\] then calculates the volume of incoming traffic required so that the probability of any out going link being unused is negligible.
<table>
<tbody>
<tr>
<td></td>
<td>b=9f2(1−pipi)𝑏9superscript𝑓21subscript𝑝𝑖subscript𝑝𝑖b=\frac{9}{f^{2}}\left(\frac{1-p_{i}}{p_{i}}\right)</td>
<td></td>
<td>(4)</td>
</tr>
</tbody>
</table>
where $`f`$ is the percentage deviation of traffic output on a particular link of $`i`$ in a given flushing cycle from the mean traffic output.
Combining this with $`p_{m\hspace{0pt}i\hspace{0pt}n}`$, the probability associated with the highest degree node in the mix network, we can derive the amount of genuine traffic to be mixed together.
The prevention of basic intersection attacks as a system design criteria can be traced back to the work of Reiter and Rubin \[[RR98](#bib.bibx27)\].
## 5 Results and Discussion
### 5.1 Simulation parameters
In all the synthetically generated networks we considered, we $`N \approxeq 5000`$ nodes. The parameters used for each of them are listed below.
We model scale-free networks with linear preferential attachment with $`m`$ links per node and average node degree $`\langle d\rangle`$ ; $`2 \leq m \leq 7`$ and $`4 \leq {\langle d\rangle} \leq 14`$.
Next we model scale-free random networks which have a scale-free degree sequence but which are random in all other respects. Generated with parameters $`\alpha = 0.25`$, $`\beta = 0.25`$ and Average node degree $`4 \leq {\langle d\rangle} \leq 14`$. See section [3.3](#S3.SS3 "3.3 Scale-free random graph topology ‣ 3 Network models ‣ Anonymity in the wild: Mixes on unstructured networks") for an explanation of $`\alpha`$ and $`\beta`$
Klienberg-Watts-Strogatz model of directed social network ties is analyzed next, generated with parameters $`r`$, the lattice radius within which each node creates direct links to all its neighbors. $`q`$ is the number of weak ties. We used $`1 \leq r \leq 4`$ and $`2 \leq q \leq 10`$.
Our next network is based on our primary source data, obtained by web-crawling the LiveJournal site. The snapshot of the network we use in our analysis has 3,746,240 nodes and 27,430,000 edges.
Finally we analyze two theoretical topologies, one degree heterogeneous and the other degree homogeneous, to offer a baseline comparison against ER graph and constant expander graph topologies.
The ER graph is created with each edge formation as the result of a Poisson distribution of $`p = 0.0028`$ with $`{\langle d\rangle} = 14`$.
The constant expander graph is created with each node having $`D = 14`$ edges. Motwani et.al. \[[MR95](#bib.bibx19)\] prove a relation between the second eigenvalue $`\lambda_{2}`$ of the transition matrix of a constant expander graph and the degree $`D`$ of a node $`\lambda_{2} \geq \frac{2\hspace{0pt}\sqrt{D - 1}}{D}`$. We can then use the result of Sinclair \[[Sin93](#bib.bibx30)\] connecting $`\lambda_{2}`$, random walk length $`t`$ and convergence rate $`\Delta\hspace{0pt}{(t)}`$, namely $`{\Delta\hspace{0pt}{(t)}} \leq \frac{\lambda_{2}^{t}}{\min_{i \in V}\pi_{i}}`$. For $`D = 14`$, we have a constant expander graph with theoretical minimum second eigen-value of $`\lambda_{2} \geq 0.5527708`$, converging to maximal anonymity state in approximately 4 steps. This forms the baseline against which we compare all the other topologies.
### 5.2 Efficiency
We can now comment on the efficiency and recommended mix route lengths for various network topologies by comparing them to our baselines.
[Refer to caption](/html/0706.0430/assets/x1.png)
<table>
<thead>
<tr>
<th>Network (N=5000𝑁5000N=5000)</th>
<th>⟨d⟩orDdelimited-⟨⟩𝑑𝑜𝑟𝐷\langle d\rangle orD</th>
<th>t𝑡t</th>
<th>𝒜networksubscript𝒜𝑛𝑒𝑡𝑤𝑜𝑟𝑘\mathcal{A}_{network}</th>
</tr>
</thead>
<tbody>
<tr>
<td>SFR</td>
<td>2</td>
<td> 8</td>
<td>11.4383</td>
</tr>
<tr>
<td></td>
<td>3</td>
<td> 7</td>
<td>11.5626</td>
</tr>
<tr>
<td></td>
<td>4</td>
<td> 6</td>
<td>11.5958</td>
</tr>
<tr>
<td></td>
<td>5</td>
<td> 6</td>
<td>11.6135</td>
</tr>
<tr>
<td></td>
<td>6</td>
<td> 5</td>
<td>11.6351</td>
</tr>
<tr>
<td>ER</td>
<td>14</td>
<td>7</td>
<td>12.2339</td>
</tr>
<tr>
<td>Expander</td>
<td>14</td>
<td>4</td>
<td>12.2877</td>
</tr>
</tbody>
</table>
Figure 1: Convergence rates: Efficiency and maximal anonymity for Scale-free random, ER and Constant expander graph topologies
The efficiency of mix topologies based on a scale-free random networks is shown in Figure [1](#S5.F1 "Figure 1 ‣ 5.2 Efficiency ‣ 5 Results and Discussion ‣ Anonymity in the wild: Mixes on unstructured networks"). It plots the anonymity achieved against increasing random walk lengths. Maximal anonymity is calculated using equation [3](#S4.E3 "In 4.4 Measuring mix network efficiency ‣ 4 Evaluation framework for measuring traffic analysis resistance ‣ Anonymity in the wild: Mixes on unstructured networks") is the entropy of the probability distribution of the chain at convergence.
Our calculations show that maximal anonymity is reached in just 6 steps in the medium density case $`{\langle d\rangle} \geq 4`$, as opposed to 4 steps in to 4 steps in an expander graph topology. It turns out that social collaboration networks \[[New01a](#bib.bibx20), [New01c](#bib.bibx22), [New01b](#bib.bibx21)\] with scale-free characteristics have average degrees in the range of $`4 \leq {\langle d\rangle} \leq 18`$. This suggests, firstly, that efficient mix networks can be designed using scale-free random networks, and second, that mildly denser scale-free networks are more suitable for building mix networks than sparser ones.
While this is an encouraging initial result, it is important to strike a note of caution. Scale-free random graphs only model the scale-free aspect of degree distribution, while being random in every other way. However most real world unstructured networks have several other non-random characteristics apart from their degree distributions.
A number of real world unstructured networks are not scale-free, hence we included the Klienberg-Watts-Strogatz(KWS) network topology, as it explicitly models the presence of weak ties in a network. We experiment with a number of parameter configurations; selecting $`r = 1`$ and $`r = 4`$ to model low and high richness in local links or ’strong ties’ between nodes; and $`2 \leq q \leq 10`$ the number of short cuts or ’weak ties’, between mix nodes. Figure [2](#S5.F2 "Figure 2 ‣ 5.2 Efficiency ‣ 5 Results and Discussion ‣ Anonymity in the wild: Mixes on unstructured networks")-a plots mix-route length vs mix network anonymity, for the KWS topology. When the topology is poor in local links, it seems to converge in 7 to 8 hops, given enough short cuts. However, if the network invests a large amount of resources into local connections forming relatively tightly knit communities, then regardless of the amount of shortcuts, convergence is not achieved until 62 hops!
[Refer to caption](/html/0706.0430/assets/x2.png)
\(a\) Klienberg-Watts-Strogatz model
[Refer to caption](/html/0706.0430/assets/x3.png)
\(b\) Scale-free network with preferential attachment
Figure 2: Mean entropy vs mix-route length
Our final model network topology is the scale-free network based on linear preferential attachment, which has attracted much attention in the complex networks literature. This topology models a scale-free network where hubs are connected to other hubs, a pattern that is repeatedly observed in many real world scale-free networks. The parameter $`m`$ controls graph sparsity, random walk and convergence results are shown in figure [2](#S5.F2 "Figure 2 ‣ 5.2 Efficiency ‣ 5 Results and Discussion ‣ Anonymity in the wild: Mixes on unstructured networks")-b. Our simulations show that while very sparse topologies converge in 10 to 15 hops, topologies that are relatively dense converge within 6 or so hops, this is comparable to the optimal 4 hops of a constant expander graph.
[Refer to caption](/html/0706.0430/assets/x4.png)
Figure 3: Mean entropy vs mix-route length for LiveJournal
<table>
<thead>
<tr>
<th>SF</th>
<th>m𝑚m</th>
<th>t𝑡t</th>
<th>𝒜networksubscript𝒜𝑛𝑒𝑡𝑤𝑜𝑟𝑘\mathcal{A}_{network}</th>
</tr>
</thead>
<tbody>
<tr>
<td></td>
<td>2</td>
<td>15</td>
<td>11.5852</td>
</tr>
<tr>
<td></td>
<td>3</td>
<td>10</td>
<td>11.6961</td>
</tr>
<tr>
<td></td>
<td>4</td>
<td>6</td>
<td>11.7293</td>
</tr>
<tr>
<td></td>
<td>5</td>
<td>6</td>
<td>11.7687</td>
</tr>
<tr>
<td></td>
<td>6</td>
<td>6</td>
<td>11.7953</td>
</tr>
<tr>
<td></td>
<td>7</td>
<td>6</td>
<td>11.8090</td>
</tr>
<tr>
<td>KWS</td>
<td>q=2,p=1formulae-sequence𝑞2𝑝1q=2,p=1</td>
<td>11</td>
<td>12.2945</td>
</tr>
<tr>
<td></td>
<td>q=10,p=1formulae-sequence𝑞10𝑝1q=10,p=1</td>
<td>5</td>
<td>12.2939</td>
</tr>
<tr>
<td></td>
<td>q=2,p=4formulae-sequence𝑞2𝑝4q=2,p=4</td>
<td>63</td>
<td>11.6440</td>
</tr>
<tr>
<td></td>
<td>q=10,p=4formulae-sequence𝑞10𝑝4q=10,p=4</td>
<td>63</td>
<td>11.6380</td>
</tr>
<tr>
<td>ER</td>
<td>⟨d⟩=14delimited-⟨⟩𝑑14\langle d\rangle=14</td>
<td>7</td>
<td>12.2339</td>
</tr>
<tr>
<td>Expander</td>
<td>D=14𝐷14D=14</td>
<td>4</td>
<td>12.2877</td>
</tr>
</tbody>
</table>
Figure 4: Convergence rates: Efficiency and maximal anonymity for linear Scale-free, KWS, ER and constant expander graph topologies
Next, we considered our primary data source the LiveJournal graph with a little less than 4 million nodes. Figure [4](#S5.F4 "Figure 4 ‣ 5.2 Efficiency ‣ 5 Results and Discussion ‣ Anonymity in the wild: Mixes on unstructured networks") shows the convergence rate of mix routes, which we note converges to the stationary distribution in around 11 hops. While this seems a high number in comparison to expander graphs (converging in 4 hops), we also note that the entropy achieved by the random walk in 4 hops in LJ is $`\mathcal{A}_{L\hspace{0pt}J}^{4} = 15.56`$. To obtain the equivalent on an expander graph topology we would only need $`2^{15.56}`$ or 48309 nodes. On the face of it the design decision seems really simple, to go for a structured expander graph topology. We argue a different view: A successful mix network design must also consider liability management issues arising from running a mix. Considering that aspect, topology links backed up by social capital are likely to be more robust than those of an optimal topology, but where nodes quickly buckle under legal pressure. We propose, running mixes on the nodes of the LJ topology bundled along with a future LiveJournal client. Nodes only allow incoming traffic from their neighbors, and will only direct outgoing traffic to their neighbors.
In this context, an interesting question is why nodes would process traffic that didn’t originate from their neighbors, and especially so in the face of legal hassles?
We offer the following reasoning: Humans making decisions on whether or not to run a mix server, will have to consider the following costs. They benefit in the long term, from processing traffic for unknown nodes in order to generate a diverse user base, the need for which is well illustrated in Dingledine and Mathewson \[[DM06](#bib.bibx9)\]. However this only holds if other mixes cooperative accordingly. Then there is the immediate social benefit of having processed traffic for your friends. The success of the system then depends on the extent to which individual nodes perceive the costs of litigation pressure to be less than the total of immediate social benefit and the long term benefit of a diverse user base. Psychology studies tell us that humans involved in taking security decisions weigh short and long term benefits differently. It should also be interesting to investigate whether the idea of running a mix to primarily process traffic for your friends is an effective tool for seeding indirect reciprocity in a mix network where cooperation flourishes.
### 5.3 Compromised mix nodes
As explained in our evaluation framework, compromised nodes can lead to compromised routes. This presents a special challenge in unstructured networks where $`\pi_{A}`$, the probability mass of the stationary distribution $`\pi`$, corresponding to set of compromised nodes $`A`$, can be significant for topological reasons.
To measure the robustness to nodes being strategically compromised by an attacker, we simulated 100000 random walks of different lengths, for each of our network topologies, in the range indicated by efficiency considerations of the previous section $`3 \leq t \leq 6`$, and measured the fraction that passed through compromised nodes. The set of compromised nodes is chosen to consist of the nodes with the highest degrees in the network. In each case, for mix routes greater than 4 hops the probability of existence of a compromised mix route is negligible. Fig [5](#Pt0.A1.F5 "Figure 5 ‣ Appendix 0.A Mix-route compromise on linear preferential attachment scale-free networks ‣ Anonymity in the wild: Mixes on unstructured networks") in the appendix confirms that the threat of mix route compromise can be successfully reduced by suitably increasing the mix-route length.
### 5.4 Intersection attacks
Using equation [4](#S4.E4 "In 4.6 Intersection attacks ‣ 4 Evaluation framework for measuring traffic analysis resistance ‣ Anonymity in the wild: Mixes on unstructured networks") we consider the required batch sizes for a threshold mix, so that the traffic output on any link in the mix network does not deviate by more than 5% from the mean traffic output on that link. For $`f = 5`$ we calculate the number of messages that must be received in each mixing cycle in table [1](#S5.T1 "Table 1 ‣ 5.4 Intersection attacks ‣ 5 Results and Discussion ‣ Anonymity in the wild: Mixes on unstructured networks").
<table>
<thead>
<tr>
<th>Network</th>
<th>⟨d⟩delimited-⟨⟩𝑑\langle d\rangle</th>
<th>pminsubscript𝑝𝑚𝑖𝑛p_{min}</th>
<th>Batch size</th>
</tr>
</thead>
<tbody>
<tr>
<th>SFR</th>
<th>2</th>
<td>0.0344</td>
<td>10.08</td>
</tr>
<tr>
<th></th>
<th>3</th>
<td>0.0222</td>
<td>15.84</td>
</tr>
<tr>
<th></th>
<th>4</th>
<td>0.0243</td>
<td>14.4</td>
</tr>
<tr>
<th></th>
<th>5</th>
<td>0.0192</td>
<td>18.36</td>
</tr>
<tr>
<th></th>
<th>6</th>
<td>0.0135</td>
<td>26.28</td>
</tr>
<tr>
<th></th>
<th>7</th>
<td>0.0125</td>
<td>28.44</td>
</tr>
<tr>
<th>KWS</th>
<th>27 (q=1,r=1formulae-sequence𝑞1𝑟1q=1,r=1)</th>
<td>0.0294</td>
<td>11.88</td>
</tr>
<tr>
<th></th>
<th>43 (q=10,r=1formulae-sequence𝑞10𝑟1q=10,r=1)</th>
<td>0.0169</td>
<td>20.88</td>
</tr>
<tr>
<th></th>
<th>26 (q=1,r=4formulae-sequence𝑞1𝑟4q=1,r=4)</th>
<td>0.0333</td>
<td>10.44</td>
</tr>
<tr>
<th></th>
<th>28 (q=10,r=4formulae-sequence𝑞10𝑟4q=10,r=4)</th>
<td>0.0294</td>
<td>11.88</td>
</tr>
<tr>
<th>SF-linear</th>
<th>4</th>
<td>0.0048</td>
<td>74.16</td>
</tr>
<tr>
<th></th>
<th>6</th>
<td>0.0048</td>
<td>74.16</td>
</tr>
<tr>
<th></th>
<th>8</th>
<td>0.0041</td>
<td>86.04</td>
</tr>
<tr>
<th></th>
<th>10</th>
<td>0.0038</td>
<td>93.6</td>
</tr>
<tr>
<th></th>
<th>12</th>
<td>0.0037</td>
<td>96.12</td>
</tr>
<tr>
<th></th>
<th>14</th>
<td>0.0031</td>
<td>112.32</td>
</tr>
<tr>
<th>LJ</th>
<th>7.3221</th>
<td>0.00857</td>
<td>41.64</td>
</tr>
<tr>
<th>ER</th>
<th>14</th>
<td>0.0333</td>
<td>10.44</td>
</tr>
<tr>
<th>Expander</th>
<th>14</th>
<td>0.0714</td>
<td>4.68</td>
</tr>
</tbody>
</table>
Table 1: Batch sizes required to prevent intersection attacks
From table [1](#S5.T1 "Table 1 ‣ 5.4 Intersection attacks ‣ 5 Results and Discussion ‣ Anonymity in the wild: Mixes on unstructured networks") it is clear that scale-free random networks and KWS both require a batch size that is 4-5 times that of expander graphs. Whether social networks can produce enough ’chatter’ to feed genuine traffic into the mix network is an open question.
Our theoretical base line of ER network topology does slightly better at a little over twice that. More significantly, the LJ network has a batch size of almost 9 times the required batch size for expander graphs. Scale-free networks with linear preferential attachment are the worst performing, requiring a batch size almost 20 times larger than expanders. We think that the exceptionally high value of batch size in LJ network is due to its size of four million or so nodes. While does not mean that LJ is inherently unsuitable as a mix network topology, but it certainly indicates a scalability limit with the deployment of mixes on LJ nodes, as proposed earlier.
## 6 Conclusions
We have analyzed a comprehensive set of network topologies from the perspective of efficiency, maximal anonymity, compromised nodes and simple intersection attacks in comparison with (provably optimal) expander graphs.
To the standard threat model of the global passive adversary, we have added real world issues such as liability management and the need for clear incentives for carrying traffic under the pressure of legal threats, and discussed our simulation results in this context.
We considered topologies with two important characteristics found in empirical studies of large-scale unstructured networks: scale-freeness (scale-free random graph) and the small-world property (Klienberg-Watts-Strogatz (KWS) graph). In both the topologies, we can recommend mix route lengths for achieving 95% of maximal anonymity, that is only a few hops larger than the optimal route length found in expander graph topologies. Currently deployed mix networks such as TOR have around 540 volunteers. To increase the scale of such mix deployments the Internet, we believe the way forward (for high latency systems only) is to use online social networks. The minimum mix route must have three mixes to allow sender and receiver anonymity. For this length, a mix network constructed by placing mixes on the nodes of a social network such as LiveJournal can achieve far higher maximal anonymity as per the entropy metric we have used. We argue that including network incentives within a framework does not allow the construction of structured overlay mix topologies that can robustly withstand the threat of legal action. By moving to social networks, we make a start on tapping the social capital underlying node-node interaction to encourage users to deploy and run mixes with policies that reflect this aspect.
We also found that subverted nodes, either compromised randomly, or by strategic choice, on the basis of their degrees has little effect on the efficiency of a mix network. This is because the route length required to mitigate that risk is less than the recommended length for achieving efficient convergence rates.
We also analyzed scale-free and the small-world topologies for their robustness to attacks based on traffic load patterns observable on their out-going and in-coming links. Both the scale-free random graph topology and the KWS topology turn out to require almost 5 times as much traffic as corresponding expander graph topology. This suggests the need for further tests to see if enough genuine traffic is generated in online social network interaction, to satisfy the minimum batch sizes required for preventing the most basic versions of these attacks.
We conclude that, unstructured networks based on large-scale topologies are indeed very promising, we have outlined the merits and challenges these topologies present to the design of mix networks for anonymous communication.
## 7 Acknowledgements
The authors are grateful to Ross Anderson for reviews on early versions of the paper, and to George Danezis and Roger Dingledine, for thought provoking discussions.
## Appendix 0.A Mix-route compromise on linear preferential attachment scale-free networks
In this section we sketch a few analytical results concerning mix-route compromise in BA scale-free networks.
Let $`B`$ be the set of compromised high vertex-order centrality nodes. For a route to be fully compromised, all intermediate nodes must be in $`B`$. We then wish to calculate,
$`P{(C_{l})} = {\lbrack Pr{(Random - Walk{(v_{1}\ldots.v_{l})})}\rbrack}\forall v_{1}\ldots v_{l} \subseteq B`$.
It is straightforward to see that if $`l > {|B|}`$ then P(C)=0. In BA scale-free networks, all hubs(high vertex-order) nodes are connected to each other. Hence,
<table>
<tbody>
<tr>
<td></td>
<td>P(C)=|B|−1∏j∈Bkj𝑃𝐶𝐵1subscriptproduct𝑗𝐵subscript𝑘𝑗P(C)=\frac{|B|-1}{\prod_{{\scriptscriptstyle j\in B}}k_{j}}</td>
<td></td>
</tr>
</tbody>
</table>
[Refer to caption](/html/0706.0430/assets/x5.png)
\(a\) length=$`3`$
[Refer to caption](/html/0706.0430/assets/x6.png)
\(b\) length=$`4`$
[Refer to caption](/html/0706.0430/assets/x7.png)
\(c\) length=$`5`$
[Refer to caption](/html/0706.0430/assets/x8.png)
\(d\) length=$`6`$
Figure 5: Probability of mix-route compromise vs no. of corrupt nodes
## Appendix 0.B Convergence rate and network size in scale-free random networks
Simulations conducted in this paper have not accounted for the effect of varying network size on the convergence rate of the respective topologies. We address this, by offering a simple conductance based proof that the second eigen-value of a scale-free network is a independent of the network size. See \[[Ran06](#bib.bibx25)\] for a review of the conductance based technique as well as others.
We denote a scale-free network generated by preferential attachment, by $`G_{m,n}\hspace{0pt}{(V,E)}`$ where $`m`$ is the number of initial nodes created at time=$`t_{0}`$ and $`n`$ is the total number of nodes in the network. At every time step $`{t_{i},i} \geq 0`$, $`m`$ nodes are added to the networks. For every node added, we create $`m`$ edges from the node to existing nodes in the network. We continue until $`{|V|} = n`$.
Next, there is an intimate relationship between the rate of convergence and a certain structural property called the conductance of the underlying graph. Consider a randomly chosen sub-graph $`S`$ of $`G\hspace{0pt}{(V,E)}`$. Suppose a random walk on the graph visits node $`i`$ $`i \in S`$. What is the probability that the walk exits $`S`$ in a single hop. If conductance is small, then a walk would tend to “get stuck” in $`S`$, whereas if conductance is large it easily “flows” out of $`S`$.
Formally, for $`S \subset G`$, the volume of S is $`{v\hspace{0pt}o\hspace{0pt}l_{G}\hspace{0pt}{(S)}} = {\sum_{u \in S}{d_{G}\hspace{0pt}{(u)}}}`$, where $`d_{G}\hspace{0pt}{(u)}`$ is the degree of node $`u`$. The cutset of $`S`$, $`C_{G}\hspace{0pt}{(S,\overline{S})}`$, is the multiset of edges with one endpoint in S and the other endpoint in $`\overline{S}`$. The textbook definition of conductance $`\Phi_{G}`$ of the graph $`G`$ is the following:
<table>
<tbody>
<tr>
<td></td>
<td>ΦG=minS⊂V,volG(S)≤volG(V)/2|CG(S,S¯)|volG(S)subscriptΦ𝐺subscriptformulae-sequence𝑆𝑉𝑣𝑜subscript𝑙𝐺𝑆𝑣𝑜subscript𝑙𝐺𝑉2subscript𝐶𝐺𝑆¯𝑆𝑣𝑜subscript𝑙𝐺𝑆\Phi_{G}=\min_{S\subset V,vol_{G}(S)\leq vol_{G}(V)/2}\frac{|C_{G}(S,\overline{S})|}{vol_{G}(S)}</td>
<td></td>
<td>(5)</td>
</tr>
</tbody>
</table>
\[[MPS03](#bib.bibx18)\] prove that the conductance of a scale-free network is a constant. Specifically, $`{\forall m} \geq {2\hspace{0pt}a\hspace{0pt}n\hspace{0pt}d\hspace{0pt}c} < {{2\hspace{0pt}{({d - 1})}} - 1}`$, $`{\exists\alpha} = {\alpha\hspace{0pt}{(d,c)}}`$ such that
<table>
<tbody>
<tr>
<td></td>
<td>Φ=αm+αΦ𝛼𝑚𝛼\Phi=\frac{\alpha}{m+\alpha}</td>
<td></td>
<td>(6)</td>
</tr>
</tbody>
</table>
From \[[Sin93](#bib.bibx30)\] we have the following bound for $`\lambda_{2}`$:
<table>
<tbody>
<tr>
<td></td>
<td>1−2Φ≤λ2≤1−Φ2/212Φsubscript𝜆21superscriptΦ221-2\Phi\leq\lambda_{2}\leq 1-\Phi^{2}/2</td>
<td></td>
<td>(7)</td>
</tr>
</tbody>
</table>
Substituting for $`\Phi`$ from equation [6](#Pt0.A2.E6 "In Appendix 0.B Convergence rate and network size in scale-free random networks ‣ Anonymity in the wild: Mixes on unstructured networks") in equation [7](#Pt0.A2.E7 "In Appendix 0.B Convergence rate and network size in scale-free random networks ‣ Anonymity in the wild: Mixes on unstructured networks"), it is easy to see that $`\lambda_{2}`$ is a constant.
## References
- \[AB02\] R. Albert and A. Barabási. Statistical mechanics of complex networks, 2002.
- \[ACL00\] William Aiello, Fan Chung, and Linyuan Lu. A random graph model for massive graphs.
In STOC ’00: Proceedings of the thirty-second annual ACM symposium on Theory of computing, pages 171–180, New York, NY, USA, 2000. ACM Press.
- \[Bor05\] Nikita Borisov. Phd thesis: Anonymous routing in structured peer-to-peer overlays, April 2005.
- \[BPS00\] Oliver Berthold, Andreas Pfitzmann, and Ronny Standtke. The disadvantages of free MIX routes and how to overcome them.
In H. Federrath, editor, Proceedings of Designing Privacy Enhancing Technologies: Workshop on Design Issues in Anonymity and Unobservability, pages 30–45. Springer-Verlag, LNCS 2009, July 2000.
- \[Cha81\] David Chaum. Untraceable electronic mail, return addresses, and digital pseudonyms.
Communications of the ACM, 4(2), February 1981.
- \[CSWH00\] Ian Clarke, Oskar Sandberg, Brandon Wiley, and Theodore Hong. Freenet: A distributed anonymous information storage and retrieval system.
In Proceedings of Designing Privacy Enhancing Technologies: Workshop on Design Issues in Anonymity and Unobservability, pages 46–66, July 2000.
- \[Dan03\] George Danezis. Mix-networks with restricted routes.
In Roger Dingledine, editor, Proceedings of Privacy Enhancing Technologies workshop (PET 2003). Springer-Verlag, LNCS 2760, March 2003.
- \[DDM03\] George Danezis, Roger Dingledine, and Nick Mathewson. Mixminion: Design of a type iii anonymous remailer protocol.
In IEEE Symposium on Security and Privacy, pages 2–15, 2003.
- \[DM06\] Roger Dingledine and Nick Mathewson. Anonymity loves company: Usability and the network effect.
In Proceedings of the Fifth Workshop on the Economics of Information Security (WEIS 2006), Cambridge, UK, June 2006.
- \[DMS04\] Roger Dingledine, Nick Mathewson, and Paul Syverson. Tor: The second-generation onion router.
In Proceedings of the 13th USENIX Security Symposium, August 2004.
- \[EG02\] Laurent Eschenauer and Virgil D. Gligor. A key-management scheme for distributed sensor networks.
In CCS ’02: Proceedings of the 9th ACM conference on Computer and communications security, pages 41–47, New York, NY, USA, 2002. ACM Press.
- \[ER59\] P. Erdos and A. R nyi. On random graphs.
Publicationes Mathemticae (Debrecen), 6:290–297, 1959.
- \[Gil98\] David Gillman. A chernoff bound for random walks on expander graphs.
SIAM J. Comput., 27(4):1203–1220, 1998.
- \[Gra73\] Mark S. Granovetter. The strength of weak ties.
The American Journal of Sociology, 78(6):1360–1380, 1973.
- \[KAP02\] Dogan Kesdogan, Dakshi Agrawal, and Stefan Penz. Limits of anonymity in open environments.
In Fabien Petitcolas, editor, Proceedings of Information Hiding Workshop (IH 2002). Springer-Verlag, LNCS 2578, October 2002.
- \[KM02\] T. Klingberg and R. Manfredi. ”gnutella 0.6”, June 2002.
- \[Mil67\] Stanley Milgram. The small world problem.
Psychology Today, 2:60–67, 1967.
- \[MPS03\] Milena Mihail, Christos Papadimitriou, and Amin Saberi. On certain connectivity properties of the internet topology.
In FOCS ’03: Proceedings of the 44th Annual IEEE Symposium on Foundations of Computer Science, page 28, Washington, DC, USA, 2003. IEEE Computer Society.
- \[MR95\] Rajeev Motwani and Prabhakar Raghavan.
Randomized Algorithms.
1\. Cambridge Univ. Press, 1995. Motwani.
- \[New01a\] M. E. Newman. The structure of scientific collaboration networks.
Proc Natl Acad Sci U S A, 98(2):404–409, January 2001.
- \[New01b\] M. E. J. Newman. Scientific collaboration networks. ii. shortest paths, weighted networks, and centrality.
Phys. Rev. E, 64(1):016132, Jun 2001.
- \[New01c\] M. E. J. Newman. Scientific collaboration networks. i. network construction and fundamental results.
Phys. Rev. E, 64(1):016131, Jun 2001.
- \[New03\] M. E. J. Newman. The structure and function of complex networks.
SIAM Review, 45(2):167–256, 2003.
- \[PH00\] Andreas Pfitzmann and Marit Hansen. Anonymity, unobservability, and pseudonymity: A consolidated proposal for terminology. Draft, July 2000.
- \[Ran06\] Dana Randall. Rapidly mixing markov chains with applications in computer science and physics.
Computing in Science and Engineering, 8(2):30–41, 2006.
- \[RFI02\] Matei Ripeanu, Ian Foster, and Adriana Iamnitchi. ”mapping the gnutella network: Properties of large-scale peer-to-peer systems and implications for system design”.
IEEE Internet Computing Journal, 6(1), August 2002.
- \[RR98\] Michael K. Reiter and Aviel D. Rubin. Crowds: anonymity for web transactions.
ACM Trans. Inf. Syst. Secur., 1(1):66–92, 1998.
- \[RSG98\] Michael G. Reed, Paul F. Syverson, and David M. Goldschlag. Anonymous connections and onion routing.
IEEE Journal on Selected Areas in Communications, 16(4), 1998.
- \[SD02\] Andrei Serjantov and George Danezis. Towards an information theoretic metric for anonymity.
In Paul Syverson and Roger Dingledine, editors, Privacy Enhancing Technologies, volume 2482 of LNCS, San Francisco, CA, April 2002.
- \[Sin93\] Alistair Sinclair.
Algorithms for random generation and counting: a Markov chain approach.
Birkhauser Verlag, Basel, Switzerland, Switzerland, 1993.
- \[STRL00\] Paul Syverson, Gene Tsudik, Michael Reed, and Carl Landwehr. Towards an Analysis of Onion Routing Security.
In H. Federrath, editor, Proceedings of Designing Privacy Enhancing Technologies: Workshop on Design Issues in Anonymity and Unobservability, pages 96–114. Springer-Verlag, LNCS 2009, July 2000.
- \[Wes01\] Douglas B West.
Introduction to Graph Theory.
Prentice Hall, second edition, 2001. |
0705.4185 | cs.CR | cyber | # Secure Two-party Protocols for Point Inclusion Problem
###### Abstract.
It is well known that, in theory, the general secure multi-party computation problem is solvable using circuit evaluation protocols. However, the communication complexity of the resulting protocols depend on the size of the circuit that expresses the functionality to be computed and hence can be impractical. Hence special solutions are needed for specific problems for efficiency reasons. The point inclusion problem in computational geometry is a special multiparty computation and has got many applications. Previous protocols for the secure point inclusion problem are not adequate. In this paper we modify some known solutions to the point inclusion problem in computational geometry to the frame work of secure two-party computation.
Tony Thomas
Department of Mathematics
Korea Advanced Institute of Science and Technology
Daejeon, 305-701, Republic of Korea
Email: tonythomasiit@yahoo.com
keywords: multiparty computation, point inclusion problem, computational geometry
## 1. Introduction
The rapid growth of networks has opened up tremendous opportunities for cooperative computation, where the output depends on the private inputs of several entities. These computations could even occur between mutually untrusted entities or competitors. The problem is trivial if the context allows to have a trusted entity that would know the inputs from all the participants; however if the context disallows this, then the techniques of secure multi-party computation are used to provide useful solutions.
Generally speaking, a secure multi-party computation problem deals with computing a function in a distributed network where each participant holds one of the inputs, ensuring that no more information is revealed to a participant in the computation than that can be computed from that participant’s input and output. The history of the multi-party computation problem is extensive since it was introduced by Yao \[[1](#bib.bib1)\] and extended by Goldreich, Micali, and Wigderson \[[2](#bib.bib2)\] and by many others. These works use a similar methodology: each functionality $`F`$ is represented as a Boolean circuit, and then the parties run a protocol for every gate in the circuit. The protocols it generates depend on the size of the circuit. This size depends on the size of the input and on the complexity of expressing F as a circuit. If the functionality F is complicated, using the circuit evaluation protocol will typically not be practical. Therefore, Goldreich \[[3](#bib.bib3)\] pointed out that using the solutions from these general results for special cases of multi-party computation could be impractical; special efficient solutions should be developed for specific problems. This is the motivation for seeking solutions to specific cooperative computational problems, in which the solutions are more efficient than the general theoretical solutions. To this end some problems such as comparing two private numbers \[[4](#bib.bib4), [5](#bib.bib5), [6](#bib.bib6), [7](#bib.bib7)\], privacy preserving data mining \[[8](#bib.bib8), [9](#bib.bib9)\], comparing information \[[10](#bib.bib10)\], privacy preserving geometric computation \[[11](#bib.bib11)\], privacy preserving cooperative scientific computation \[[12](#bib.bib12), [13](#bib.bib13)\], privacy preserving database query \[[14](#bib.bib14)\], privacy preserving auction \[[15](#bib.bib15)\], privacy preserving statistical analysis \[[16](#bib.bib16), [17](#bib.bib17)\], privacy preserving set operations \[[18](#bib.bib18)\] have been investigated.
In secure multi-party computational geometry we seek secure protocols for several geometric problem like point inclusion problem, intersection of two shapes, range searching problem etc where the data is shared by two or more entities. In this paper, we construct secure two-party protocols for the point inclusion problem in star-shaped domains and more complex polygonal domains. Here, one entity Alice has a point $`M`$, and Bob has a polygon $`P`$. Their aim is to determine whether $`M`$ is inside $`P`$, or not without revealing to each other their private inputs.
We outline the related work in Section 2. In Section 3, we introduce our adversary models as well as the cryptographic tools used in the subsequent sections. In sections 4, we study the point inclusion problem in star-shaped domains and in Section 5, we consider more general polygonal domains. The paper concludes with some remarks in Section 6.
## 2. Related Work
The secure multiparty computational geometry has got wide applications in the fields of military, computer graphics etc. The study of secure multiparty computational geometry was initiated by Atallah et al. \[[11](#bib.bib11)\] with their work on secure point inclusion problem and polygonal intersection problem. Their protocol for the point inclusion problem is applicable to simple polygonal domain and has complexity $`O\hspace{0pt}{(n)}`$ where $`n`$ is the number of edges of the polygon. Later Li et al. \[[19](#bib.bib19)\] studied the point inclusion problem for circular domain. However, their solution is not secure in the sense that each party gets additional information regarding the location of the other party’s object. Moreover, their solution is highly inefficient. A more efficient protocol for the point inclusion problem in a circular domain was recently proposed by Luo et al. \[[20](#bib.bib20)\].
In this paper we consider the point inclusion problem in a star-shaped domain and a more general polygonal domain (can have several disconnected nested components). Two protocols for the star shaped domain with round complexities $`O\hspace{0pt}{(n)}`$ and $`O\hspace{0pt}{({\log n})}`$ respectively, and a protocols for more general polygonal domain with round complexity $`O\hspace{0pt}{(n)}`$, where $`n`$ is the number of vertices are given.
## 3. Preliminaries
In this section we state our security assumptions and list the building block for our protocols.
### 3.1. Security Assumption
We assume that all parties are semihonest. A semi-honest party is the one who follows the protocol correctly with the exception that it keeps a record of all its intermediate computations and might derive the other parties inputs from the record.
The existing protocols listed below serve as important building blocks for our protocols.
### 3.2. Homomorphic Encryption Schemes
An encryption scheme is homomorphic if for some operations $`\oplus`$ and $`\otimes`$, $`{{{E_{k}\hspace{0pt}{(x)}} \otimes E_{k}}\hspace{0pt}{(y)}} = {E_{k}\hspace{0pt}{({x \oplus y})}}`$, where $`x`$ and $`y`$ are two elements from the message space and $`k`$ is the key. Many such systems exist, and examples include the systems by Benaloh \[[21](#bib.bib21)\], Naccache and Stern \[[22](#bib.bib22)\], Okamoto and Uchiyama \[[23](#bib.bib23)\], Paillier \[[24](#bib.bib24)\], to mention a few. A useful property of homomorphic encryption schemes is that an addition operation can be conduced based on the encrypted data without decrypting them.
### 3.3. Yao’s Millionaire Protocol
The purpose of this protocol is to compare two private numbers and to determine which one is larger without revealing the numbers. This was first proposed by Yao \[[1](#bib.bib1)\] and is referred as Yao’s Millionaire Problem (because two millionaires wish to know who is richer, without revealing any other information about their net wealth). The early cryptographic solution by Yao \[[1](#bib.bib1)\] uses an untrusted third party and has communication complexity that is exponential in the number of bits of the numbers involved. Cachin proposed a solution \[[15](#bib.bib15)\] based on an untrusted third party that can misbehave on its own (for the purpose of illegally obtaining information about Alice’s or Bob’s private vectors) but does not collude with either participant. The communication complexity of Cachin’s scheme is $`O\hspace{0pt}{(l)}`$, where $`l`$ is the number of bits of each input number. Recently many efficient protocols which do not need a third party have been suggested by various authors \[[4](#bib.bib4), [5](#bib.bib5)\].
### 3.4. Scalar Product Protocol
Let Alice has a vector $`X = {(x_{1},\ldots,x_{n})}`$ and Bob has a vector $`Y = {(y_{1},\ldots,y_{n})}`$. The scalar product protocol is to securely compute the scalar (dot) product of $`X`$ and $`Y`$, given by $`X.{Y = {\sum\limits_{k = 1}^{n}{x_{i}\hspace{0pt}y_{i}}}}`$.
In \[[11](#bib.bib11)\] Du and Atallah considered a slightly different and more general form of the scalar product protocol in which Alice has the vector $`X`$ and Bob has the vector $`Y`$, and the goal of the protocol is for Alice (but not Bob) to get $`X.{Y + V}`$ where $`V`$ is random and known to Bob only. Their protocols can be easily modified to work for the version of the problem where the random $`V`$ is given ahead of time as part of Bob’s data (the special case $`V = 0`$ puts us back to the usual scalar product definition). They had developed two protocols for it. Secure protocols for the scalar product problem can be found in \[[11](#bib.bib11), [25](#bib.bib25), [26](#bib.bib26)\].
## 4. Point Inclusion in Star-shaped Domain
In this section, we study the point inclusion problem in a star-shaped polygonal domain.
Problem: Let Alice has a point $`M`$ and Bob has a star-shaped polygon $`P`$ with vertices $`P_{i}`$, for $`1 \leq i \leq n`$, where the vertices are named in the anticlockwise direction. Alice and Bob want to securely check whether $`M`$ lies inside (including boundary) $`P`$ or not.
Since $`P`$ is a star-shaped polygon, it contains a point $`Q`$ such the line segments joining $`Q`$ to $`P_{i}`$ for $`1 \leq i \leq n`$ lies entirely in $`P`$. We have the following algorithm for point inclusion from \[[27](#bib.bib27)\].
The Point Inclusion Protocol Without Privacy
1. \(1\)
Determine by binary search the wedge in which $`M`$ lies. $`M`$ lies in the wedge bounded by the rays $`\overset{\rightarrow}{Q\hspace{0pt}P_{i}}`$ and $`\overset{\rightarrow}{Q\hspace{0pt}P_{i+1}}`$ if and only if the angle formed by $`M`$, $`Q`$ and $`P_{i}`$ is a left turn and the angle formed by $`M`$, $`Q`$ and $`P_{i + 1}`$ is a right turn.
2. \(2\)
Once $`P_{i}`$ and $`P_{i + 1}`$ are found, then $`M`$ is internal if and only if the angle formed by $`P_{i}`$, $`P_{i + 1}`$ and $`M`$ is a left turn.
######
Theorem 4.1
.
\[[27](#bib.bib27)\] The inclusion question can be answered in $`O\hspace{0pt}{({\log n})}`$ time, given $`O\hspace{0pt}{(n)}`$ space and $`O\hspace{0pt}{(n)}`$ processing time.
To decide whether the angle $`\angle\hspace{0pt}P_{1}\hspace{0pt}P_{2}\hspace{0pt}P_{3}`$ is a right or left turn corresponds to evaluating a $`3 \times 3`$ determinant in the points’ coordinates. Let $`P_{i} = {(a_{i},b_{i})}`$ for $`1 \leq i \leq 3`$. The determinant
<table>
<tbody>
<tr>
<td></td>
<td>D(P1,P2,P3)=|a1b11a2b21a3b31|𝐷subscript𝑃1subscript𝑃2subscript𝑃3subscript𝑎1subscript𝑏11subscript𝑎2subscript𝑏21subscript𝑎3subscript𝑏31D(P_{1},P_{2},P_{3})=\left|\begin{array}[]{ccc}a_{1}&b_{1}&1\\ a_{2}&b_{2}&1\\ a_{3}&b_{3}&1\\ \end{array}\right|</td>
<td></td>
</tr>
</tbody>
</table>
gives twice the signed area of the triangle $`\bigtriangleup\hspace{0pt}P_{1}\hspace{0pt}P_{2}\hspace{0pt}P_{3}`$, where the sign is $`+`$ if and only if $`(P_{1},P_{2},P_{3})`$ forms a counterclockwise cycle.
Let the coordinates of $`M`$ be $`(a,b)`$ and that of $`Q`$ be $`(s,t)`$ with respect to some coordinate system known to both Alice and Bob. Now Bob chooses a new coordinate system with origin at $`Q`$ and axes parallel to the original axes. Let the co-ordinates of $`P_{i}`$ with respect to the new coordinate axes be $`(a_{i},b_{i})`$ for $`1 \leq i \leq n`$. The new coordinates of $`M`$ becomes $`({a - s},{a - t})`$. Now the angle $`\angle\hspace{0pt}M\hspace{0pt}Q\hspace{0pt}P_{i}`$ is a right turn or left turn according as the determinant
<table>
<tbody>
<tr>
<td></td>
<td>D(M,Q,Pi)=|a−sb−t1001aibi1|𝐷𝑀𝑄subscript𝑃𝑖𝑎𝑠𝑏𝑡1001subscript𝑎𝑖subscript𝑏𝑖1D(M,Q,P_{i})=\left|\begin{array}[]{ccc}a-s&b-t&1\\ 0&0&1\\ a_{i}&b_{i}&1\\ \end{array}\right|</td>
<td></td>
</tr>
</tbody>
</table>
is positive or negative. For $`1 \leq i \leq n`$, let $`A = {(a,b,1)}`$, $`B_{i} = {({- b_{i}},a_{i},{{s\hspace{0pt}b_{i}} - {t\hspace{0pt}a_{i}}})}`$ and $`C_{i} = {({(b_{i} - b_{i + 1})}, - {(a_{i} - a_{i + 1})}, - s{(b_{i} - b_{i + 1})} + t{(a_{i} - a_{i + 1})} + {(a_{i}b_{i + 1} - b_{i}a_{i + 1})}}`$. Now we have,
<table>
<tbody>
<tr>
<td></td>
<td>D(M,Q,Pi)𝐷𝑀𝑄subscript𝑃𝑖\displaystyle D(M,Q,P_{i})</td>
<td>=\displaystyle=</td>
<td>−(a−s)bi+(b−t)ai=−abi+bai+(sbi−tai)𝑎𝑠subscript𝑏𝑖𝑏𝑡subscript𝑎𝑖𝑎subscript𝑏𝑖𝑏subscript𝑎𝑖𝑠subscript𝑏𝑖𝑡subscript𝑎𝑖\displaystyle-(a-s)b_{i}+(b-t)a_{i}=-ab_{i}+ba_{i}+(sb_{i}-ta_{i})</td>
<td></td>
</tr>
<tr>
<td></td>
<td></td>
<td>=\displaystyle=</td>
<td>(a,b,1).(−bi,ai,sbi−tai)=A.Bi.formulae-sequence𝑎𝑏1subscript𝑏𝑖subscript𝑎𝑖𝑠subscript𝑏𝑖𝑡subscript𝑎𝑖𝐴subscript𝐵𝑖\displaystyle(a,b,1).(-b_{i},a_{i},sb_{i}-ta_{i})=A.B_{i}.</td>
<td></td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td>D(Pi,Pi+1,M)𝐷subscript𝑃𝑖subscript𝑃𝑖1𝑀\displaystyle D(P_{i},P_{i+1},M)</td>
<td>=\displaystyle=</td>
<td>(a−s)(bi−bi+1)−(b−t)(ai−ai+1)+(aibi+1−biai+1)𝑎𝑠subscript𝑏𝑖subscript𝑏𝑖1𝑏𝑡subscript𝑎𝑖subscript𝑎𝑖1subscript𝑎𝑖subscript𝑏𝑖1subscript𝑏𝑖subscript𝑎𝑖1\displaystyle(a-s)(b_{i}-b_{i+1})-(b-t)(a_{i}-a_{i+1})+(a_{i}b_{i+1}-b_{i}a_{i+1})</td>
<td></td>
</tr>
<tr>
<td></td>
<td></td>
<td>=\displaystyle=</td>
<td>a(bi−bi+1)−b(ai−ai+1)−s(bi−bi+1)+t(ai−ai+1)𝑎subscript𝑏𝑖subscript𝑏𝑖1𝑏subscript𝑎𝑖subscript𝑎𝑖1𝑠subscript𝑏𝑖subscript𝑏𝑖1𝑡subscript𝑎𝑖subscript𝑎𝑖1\displaystyle a(b_{i}-b_{i+1})-b(a_{i}-a_{i+1})-s(b_{i}-b_{i+1})+t(a_{i}-a_{i+1})</td>
<td></td>
</tr>
<tr>
<td></td>
<td></td>
<td></td>
<td>+(aibi+1−biai+1)subscript𝑎𝑖subscript𝑏𝑖1subscript𝑏𝑖subscript𝑎𝑖1\displaystyle+(a_{i}b_{i+1}-b_{i}a_{i+1})</td>
<td></td>
</tr>
<tr>
<td></td>
<td></td>
<td>=\displaystyle=</td>
<td>(a,b,1).((bi−bi+1),−(ai−ai+1),−s(bi−bi+1)+t(ai−ai+1)\displaystyle(a,b,1).((b_{i}-b_{i+1}),-(a_{i}-a_{i+1}),-s(b_{i}-b_{i+1})+t(a_{i}-a_{i+1})</td>
<td></td>
</tr>
<tr>
<td></td>
<td></td>
<td></td>
<td>+(aibi+1−biai+1)=A.Ci.formulae-sequencesubscript𝑎𝑖subscript𝑏𝑖1subscript𝑏𝑖subscript𝑎𝑖1𝐴subscript𝐶𝑖\displaystyle+(a_{i}b_{i+1}-b_{i}a_{i+1})=A.C_{i}.</td>
<td></td>
</tr>
</tbody>
</table>
The point $`M`$ lies in the wedge bounded by the rays $`\overset{\rightarrow}{Q\hspace{0pt}P_{i}}`$ and $`\overset{\rightarrow}{Q\hspace{0pt}P_{i+1}}`$ if and only if $`A.{B_{i} \leq 0}`$ and $`A.{B_{i + 1} \geq 0}`$ and if it happens to lie in that wedge, it lies inside the polygon if and only if $`A.{C_{i} \leq 0}`$. Note that Alice has the vector $`A`$ and Bob has the vectors $`B_{i}`$ and $`C_{i}`$ for $`1 \leq i \leq n`$. We now give the corresponding secure protocol for the point inclusion problem.
The Secure Point Inclusion Protocol 4.1
1. \(1\)
For $`i = {1,\ldots,n}`$, Alice and Bob do the following:
1. \(a\)
Bob computes $`B_{i}`$, $`C_{i}`$ and chooses at random $`V_{i}`$ and $`W_{i}`$.
2. \(b\)
Alice engages in two secure scalar product protocols with Bob and gets $`{U_{i} = A}.{B_{i} + V_{i}}`$ and $`{Z_{i} = A}.{C_{i} + W_{i}}`$.
3. \(c\)
Alice compares $`U_{i}`$ with $`V_{i}`$ and $`Z_{i}`$ with $`W_{i}`$ using millionaire protocol with Bob.
2. \(2\)
Alice identifies the index, $`i = j`$ at which $`U_{j} < V_{j}`$ and $`U_{j + 1} > V_{j + 1}`$.
3. \(3\)
Alice looks at the millionaire protocol output for the pair $`Z_{j}`$ and $`W_{j}`$. If $`Z_{j}`$ was smaller than $`W_{j}`$ then the point is inside else it is outside.
4. \(4\)
Alice communicates the result to Bob.
Analysis of the Protocol 4.1
######
Theorem 4.2
.
The Protocol 4.1 is correct, secure and has round complexity $`O\hspace{0pt}{(n)}`$.
###### Proof.
Correctness: Using the millionaire protocol, in Step 2 Alice identifies the wedge in which the point $`M`$ lies and in Step 3 she checks whether the point $`M`$ lies inside or outside the polygon. The correctness of the protocol follows from the correctness of the corresponding insecure protocol.
Security: The security of the protocol immediately follows from the privacy of the secure scalar product protocol and that of the secure protocol for the millionaire problem. Also, Alice does not reveal to Bob the wedge in which $`M`$ lies, and so Bob will not get any idea about the location of the point $`M`$.
Round Complexity: It is easy to see that the round complexity of the protocol is $`O\hspace{0pt}{(n)}`$. ∎
Using Binary Search to Reduce Round Complexity
Now, we will incorporate binary search in the above protocol to reduce its round complexity to $`O\hspace{0pt}{({\log n})}`$.
Let $`E`$ be a homomorphic commutative encryption scheme. That is if $`(E_{A},D_{A})`$ and $`(E_{B},D_{B})`$ be the encryption and decryption pairs of Alice and Bob corresponding to their keys and let $`E = E_{A}`$ or $`E_{B}`$, then
1. \(1\)
$`{E_{A}\hspace{0pt}{({E_{B}\hspace{0pt}{(x)}})}} = {E_{B}\hspace{0pt}{({E_{A}\hspace{0pt}{(x)}})}}`$;
2. \(2\)
$`E{(x)} \ast E{(y)} = E{(x.y)}`$.
Given $`U = {(u_{1},\ldots,u_{n})}`$, let $`{E\hspace{0pt}{(S)}} = {({E\hspace{0pt}{(u_{1})}},\ldots,{E\hspace{0pt}{(u_{n})}})}`$. We now give the modified secure protocol for the point inclusion problem.
The Secure Point Inclusion Protocol 4.2
1. \(1\)
For $`1 \leq i \leq n`$, Bob computes $`b_{i} = {E_{B}\hspace{0pt}{(B_{i})}}`$, and $`c_{i} = {E_{B}\hspace{0pt}{(C_{i})}}`$.
2. \(2\)
Bob sends $`(b_{1},\ldots,b_{n})`$ and $`(c_{1},\ldots,c_{n})`$ to Alice.
3. \(3\)
Alice picks an $`r`$ randomly such that $`1 < r < n`$ and cyclically rotates the lists obtained from Bob by $`r`$ positions to get $`(b_{1 + r},\ldots,b_{r})`$ and $`(c_{1 + r},\ldots,c_{r})`$.
4. \(4\)
Alice sends $`({E_{A}\hspace{0pt}{(b_{1 + r})}},\ldots,{E_{A}\hspace{0pt}{(b_{r})}})`$ and $`({E_{A}\hspace{0pt}{(c_{1 + r})}},\ldots,{E_{A}\hspace{0pt}{(c_{r})}})`$ to Bob.
5. \(5\)
Bob decrypts the list obtained from Alice with his private key $`D_{B}`$ and obtains
<table>
<tbody>
<tr>
<td></td>
<td>(DB(EA(b1+r)),…,DB(EA(br)))subscript𝐷𝐵subscript𝐸𝐴subscript𝑏1𝑟…subscript𝐷𝐵subscript𝐸𝐴subscript𝑏𝑟\displaystyle(D_{B}(E_{A}(b_{1+r})),\dots,D_{B}(E_{A}(b_{r})))</td>
<td>=\displaystyle=</td>
<td>(EA(B1+r),…,EA(Br)),subscript𝐸𝐴subscript𝐵1𝑟…subscript𝐸𝐴subscript𝐵𝑟\displaystyle(E_{A}(B_{1+r}),\dots,E_{A}(B_{r})),</td>
<td></td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td>(DB(EA(c1+r)),…,DB(EA(cr)))subscript𝐷𝐵subscript𝐸𝐴subscript𝑐1𝑟…subscript𝐷𝐵subscript𝐸𝐴subscript𝑐𝑟\displaystyle(D_{B}(E_{A}(c_{1+r})),\dots,D_{B}(E_{A}(c_{r})))</td>
<td>=\displaystyle=</td>
<td>(EA(C1+r),…,EA(Cr)).subscript𝐸𝐴subscript𝐶1𝑟…subscript𝐸𝐴subscript𝐶𝑟\displaystyle(E_{A}(C_{1+r}),\dots,E_{A}(C_{r})).</td>
<td></td>
</tr>
</tbody>
</table>
6. \(6\)
Alice computes $`E_{A}\hspace{0pt}{(A)}`$.
7. \(7\)
Alice identifies the index, $`i = j`$ for which $`A.{B_{j} < 0}`$ and $`A.{B_{j + 1} > 0}`$ using the following sub protocol in the binary search.
1. \(a\)
For each index $`k`$ Alice picks up in the binary search, Bob picks a random $`r_{k} > 0`$ encrypts with his key and sends Alice $`E_{B}\hspace{0pt}{(r_{k})}`$.
2. \(b\)
Alice encrypts with her key and sends back to Bob $`E_{A}\hspace{0pt}{({E_{B}\hspace{0pt}{(r_{k})}})}`$.
3. \(c\)
Bob decrypts and obtains $`{D_{B}\hspace{0pt}{({E_{A}\hspace{0pt}{({E_{B}\hspace{0pt}{(r_{k})}})}})}} = {E_{A}\hspace{0pt}{(r_{k})}}`$.
4. \(d\)
Bob computes $`{{{E_{B}\hspace{0pt}{(r_{k})}} \ast E_{B}}\hspace{0pt}{(B_{k})}} = {E_{B}\hspace{0pt}{({r_{k}\hspace{0pt}B_{k}})}}`$.
5. \(e\)
Alice engages in a secure scalar product protocol with Bob and obtains
$`E_{A}{(A)} \ast E_{A}{(r_{k}B_{k})} = E_{A}{(r_{k}{(A.B_{k})})}`$.
6. \(f\)
Alice decrypts and obtains $`D_{A}{(E_{A}{(r_{k}{(A.B_{k})})} = r_{k}{(A.B_{k})}}`$ and checks whether it is positive or not.
8. \(8\)
Alice checks whether $`A.D_{j}`$ is negative or positive using a similar sub protocol as in Step 7. If it is negative, the point is inside else it is outside.
9. \(9\)
Alice communicates the result to Bob.
Analysis of the Protocol 4.2
######
Theorem 4.3
.
The Protocol 4.2 is correct, secure and has round complexity $`O\hspace{0pt}{({\log n})}`$.
###### Proof.
Correctness: It is clear that, in Step 5, Bob gets the encryption of the vectors $`B_{i}`$ and $`C_{i}`$ for $`1 \leq i \leq n`$ with the key of Alice. For each index $`k`$ occurring in the binary search, Alice has $`E_{A}\hspace{0pt}{(A)}`$ and Bob has $`E_{A}\hspace{0pt}{({r_{k}\hspace{0pt}B_{k}})}`$. Using the secure scalar product protocol she obtains $`{{E_{A}\hspace{0pt}{(A)}} \ast E_{A}}\hspace{0pt}{({r_{k}\hspace{0pt}B_{k}})}`$, which is equal to $`E_{A}{(r_{k}A.B_{k})}`$ from the homomorphic property of the encryption scheme. By decryption using her private key Alice gets $`r_{k}{(A.B_{k})}`$ and she can check whether $`A.{B_{k} \geq 0}`$, since $`r_{k} > 0`$. Thus Alice can identify the wedge in which the point $`M`$ lies. Similarly, once the wedge is identified, she can check whether the point lies inside the polygon or not. Thus the correctness of the protocol follows from the correctness of the corresponding insecure protocol.
Security: Since Bob is sending $`B_{i}`$ and $`C_{i}`$ for $`1 \leq i \leq n`$, after encryption with his key, Alice will not get any information about the private data of Bob. Since Alice rotates the list of $`B_{i}`$ and $`C_{i}`$ after masking with her key, Bob will not get any idea of the specific $`B_{i}`$ and $`C_{i}`$ Alice is using in the binary search in Step 7. Hence, Bob will not get any idea of the wedge in which the point $`M`$ lies. The privacy of the secure scalar product protocol guarantees the privacy of the individual inputs during the scalar product computation in Step 7 and Step 8. Also since $`r_{k}`$ is random known only to Bob, the only information Alice can get from the scalar product is its sign.
Round Complexity: As Alice is using binary search in the identification of the wedge in which the point $`M`$ lies, it is clear that the round complexity of the protocol is $`O\hspace{0pt}{({\log n})}`$, since the complexity of the binary search is $`O\hspace{0pt}{({\log n})}`$. ∎
## 5. Point Inclusion in More General Polygonal Domain
In this section, we consider an algorithm for the point inclusion problem for a more general polygonal domain given in \[[28](#bib.bib28)\]. This domain is more general than any of the domains so far considered in the context of secure point inclusion problem.
Problem: Alice has a point $`M`$ and Bob has a polygon $`P`$ that may have multiple disconnected nested components, with vertices $`P_{1},\ldots,P_{n}`$. Alice and Bob wants to securely check whether $`M`$ lies inside (including boundary) $`P`$ or not.
The characteristic function, $`\chi\hspace{0pt}{(M)}`$ of the polygon $`P`$ is defined as,
<table>
<tbody>
<tr>
<td></td>
<td>χ(M)={1ifMlies on or insideP;0otherwise,𝜒𝑀cases1if𝑀lies on or inside𝑃otherwise0otherwiseotherwise\chi(M)=\begin{cases}1~{}~{}\text{if}~{}~{}M~{}~{}\text{lies on or inside}~{}~{}P;\\ 0~{}~{}\text{otherwise},\end{cases}</td>
<td></td>
</tr>
</tbody>
</table>
where $`M \in {\mathbb{R}}^{2}`$. Let $`0 < \theta < {2\hspace{0pt}\pi}`$, be the included angle (edges swept inside the polygon) at a vertex $`V`$. The extension to $`\infty`$ in both directions of the edges incident on the vertex $`V`$ divide the plane into $`4`$ wedges. If $`\theta < \pi`$ (convex vertex), there are two wedges with angle $`\theta`$ and two wedges with angle $`\pi - \theta`$. We call the wedges with angle $`\theta`$ as inner and those with angle $`\pi - \theta`$ as outer. If $`\theta > \pi`$ (concave vertex), there are two wedges with angle $`{2\hspace{0pt}\pi} - \theta`$ and two wedges with angle $`\theta - \pi`$. In this case, we call the wedges with angle $`{2\hspace{0pt}\pi} - \theta`$ as inner and those with angle $`\theta - \pi`$ as outer.
We assume for convenience that the point $`M`$ does not lie on any of the four rays emanating from any of the vertices of the polygon. The case in which $`M`$ lies on a ray can be easily handled separately. Now, the cross function, $`\rho_{_{V}}\hspace{0pt}{(M)}`$ of a point $`M`$ with respect to a vertex $`V`$ of the polygon is defined as
<table>
<tbody>
<tr>
<td></td>
<td>ρV(M)={12−θ2πifθ<πandMis in an inner wedge;−θ2πifθ<πandMis in an outer wedge;θ2π−12ifθ>πandMis in an inner wedge;θ2πifθ>πandMis in an outer wedge.subscript𝜌𝑉𝑀cases12𝜃2𝜋if𝜃𝜋and𝑀is in an inner wedgeotherwise𝜃2𝜋if𝜃𝜋and𝑀is in an outer wedgeotherwise𝜃2𝜋12if𝜃𝜋and𝑀is in an inner wedgeotherwise𝜃2𝜋if𝜃𝜋and𝑀is in an outer wedgeotherwise\rho_{{}_{V}}(M)=\begin{cases}\frac{1}{2}-\frac{\theta}{2\pi}~{}~{}\text{if}~{}~{}\theta<\pi~{}~{}\text{and}~{}~{}M~{}~{}\text{is in an inner wedge};\\ -\frac{\theta}{2\pi}~{}~{}\hskip 11.38092pt\text{if}~{}~{}\theta<\pi~{}~{}\text{and}~{}~{}M~{}~{}\text{is in an outer wedge};\\ \frac{\theta}{2\pi}-\frac{1}{2}~{}~{}\text{if}~{}~{}\theta>\pi~{}~{}\text{and}~{}~{}M~{}~{}\text{is in an inner wedge};\\ \frac{\theta}{2\pi}~{}~{}\hskip 17.07182pt\text{if}~{}~{}\theta>\pi~{}~{}\text{and}~{}~{}M~{}~{}\text{is in an outer wedge}.\end{cases}\\ \vspace*{1cm}</td>
<td></td>
</tr>
</tbody>
</table>
######
Theorem 5.1
.
\[[28](#bib.bib28)\] The characteristic function of the whole polygon is the sum of the cross functions of its vertices. That is
<table>
<tbody>
<tr>
<td></td>
<td>χ(M)=∑V=P1PnρV(M),∀M∈ℝ2.formulae-sequence𝜒𝑀superscriptsubscript𝑉subscript𝑃1subscript𝑃𝑛subscript𝜌𝑉𝑀for-all𝑀superscriptℝ2\chi(M)=\sum_{V=P_{1}}^{P_{n}}\rho_{{}_{V}}(M),~{}~{}~{}~{}~{}\forall M\in\mathbb{R}^{2}.</td>
<td></td>
</tr>
</tbody>
</table>
Before we give the secure protocol for the point inclusion problem, we outline a way for Alice to securely identify whether her point lies in an inner or outer wedge corresponding to a vertex $`V`$. Bob chooses four points $`V_{1},V_{2},V_{3}`$ and $`V_{4}`$ on the four rays emanating from the vertex $`V`$. Without loss of generality let us suppose that $`\overset{\rightarrow}{V\hspace{0pt}V_{1}}`$ and $`\overset{\rightarrow}{V\hspace{0pt}V_{2}}`$ bound one inner wedge and $`\overset{\rightarrow}{V\hspace{0pt}V_{3}}`$ and $`\overset{\rightarrow}{V\hspace{0pt}V_{4}}`$ bound the other one. Now Alice and Bob engages in a secure protocol (as described in the previous section) and Alice checks whether $`M`$ is inside any of these two wedges. If that is the cases $`M`$ is inside an inner wedge else $`M`$ is inside an outer wedge.
For $`1 \leq i \leq n`$, let $`\theta_{i}`$ be the included angle at the vertex $`P_{i}`$. We now give a secure protocol for the point inclusion problem.
The Secure Point Inclusion Protocol 5.1
1. \(1\)
Bob computes $`\theta = {\sum\limits_{i = 1}^{n}{{({- 1})}^{m_{i}}\hspace{0pt}\frac{\theta_{i}}{2\hspace{0pt}\pi}}}`$, where $`m_{i} = 0`$ if $`V_{i}`$ is a convex vertex ($`\theta_{i} < \pi`$) and $`m_{i} = 1`$, otherwise.
2. \(2\)
For $`1 \leq i \leq n`$ Alice and Bob do the following.
1. \(a\)
For the vertex $`V_{i}`$, Alice checks whether $`M`$ lies inside an inner or outer wedge using the protocol described above.
2. \(b\)
If the wedge is inner, Alice assigns $`u_{i} = \frac{1}{2}`$, else she assigns $`u_{i} = 0`$.
3. \(c\)
If the edge is convex Bob assigns $`v_{i} = 1`$, else he assigns $`v_{i} = {- 1}`$.
3. \(3\)
Alice assigns $`U = {(u_{1},\ldots,u_{n})}`$.
4. \(4\)
Bob assigns $`V = {(v_{1},\ldots,v_{n})}`$.
5. \(5\)
Bob engages in a secure scalar product protocol with Alice and gets $`U.V`$.
6. \(6\)
Bob computes $`{{\chi\hspace{0pt}{(M)}} = U}.{V + \theta}`$.
7. \(7\)
Bob communicates the result to Alice.
### 5.1. Analysis of the Protocol 5.1
######
Theorem 5.2
.
The Protocol 5.1 is correct, secure and has round complexity $`O\hspace{0pt}{(n)}`$.
###### Proof.
Correctness: Let $`E_{1}`$ be the set of convex vertices where the point $`M`$ lies in an inner wedge, $`E_{2}`$ be the set of convex vertices where the point $`M`$ lies in an outer wedge, $`E_{3}`$ be the set of concave vertices where the point $`M`$ lies in an inner wedge and $`E_{4}`$ be the set of concave vertices where the point $`M`$ lies in an outer wedge. Then we have,
<table>
<tbody>
<tr>
<td></td>
<td>χ(M)𝜒𝑀\displaystyle\chi(M)</td>
<td>=\displaystyle=</td>
<td>∑V=P1PnρV(M)superscriptsubscript𝑉subscript𝑃1subscript𝑃𝑛subscript𝜌𝑉𝑀\displaystyle\sum_{V=P_{1}}^{P_{n}}\rho_{{}_{V}}(M)</td>
<td></td>
</tr>
<tr>
<td></td>
<td></td>
<td>=\displaystyle=</td>
<td>∑Vi∈E1ρVi(M)+∑Vi∈E2ρVi(M)+∑Vi∈E3ρVi(M)+∑Vi∈E4ρVi(M)subscriptsubscript𝑉𝑖subscript𝐸1subscript𝜌subscript𝑉𝑖𝑀subscriptsubscript𝑉𝑖subscript𝐸2subscript𝜌subscript𝑉𝑖𝑀subscriptsubscript𝑉𝑖subscript𝐸3subscript𝜌subscript𝑉𝑖𝑀subscriptsubscript𝑉𝑖subscript𝐸4subscript𝜌subscript𝑉𝑖𝑀\displaystyle\sum_{V_{i}\in E_{1}}\rho_{{}_{V_{i}}}(M)+\sum_{V_{i}\in E_{2}}\rho_{{}_{V_{i}}}(M)+\sum_{V_{i}\in E_{3}}\rho_{{}_{V_{i}}}(M)+\sum_{V_{i}\in E_{4}}\rho_{{}_{V_{i}}}(M)</td>
<td></td>
</tr>
<tr>
<td></td>
<td></td>
<td>=\displaystyle=</td>
<td>∑Vi∈E112−θi2π+∑Vi∈E2−θi2π+∑Vi∈E1θi2π−12+∑Vi∈E1θi2πsubscriptsubscript𝑉𝑖subscript𝐸112subscript𝜃𝑖2𝜋subscriptsubscript𝑉𝑖subscript𝐸2subscript𝜃𝑖2𝜋subscriptsubscript𝑉𝑖subscript𝐸1subscript𝜃𝑖2𝜋12subscriptsubscript𝑉𝑖subscript𝐸1subscript𝜃𝑖2𝜋\displaystyle\sum_{V_{i}\in E_{1}}\frac{1}{2}-\frac{\theta_{i}}{2\pi}+\sum_{V_{i}\in E_{2}}-\frac{\theta_{i}}{2\pi}+\sum_{V_{i}\in E_{1}}\frac{\theta_{i}}{2\pi}-\frac{1}{2}+\sum_{V_{i}\in E_{1}}\frac{\theta_{i}}{2\pi}</td>
<td></td>
</tr>
<tr>
<td></td>
<td></td>
<td>=\displaystyle=</td>
<td>∑i=1n(−1)miθi2π+U.Vformulae-sequencesuperscriptsubscript𝑖1𝑛superscript1subscript𝑚𝑖subscript𝜃𝑖2𝜋𝑈𝑉\displaystyle\sum_{i=1}^{n}(-1)^{m_{i}}\frac{\theta_{i}}{2\pi}+U.V</td>
<td></td>
</tr>
</tbody>
</table>
Thus Bob can compute $`\chi\hspace{0pt}{(M)}`$ and hence the protocol is correct.
Security: The security of the protocol immediately follows from the privacy of the secure scalar product protocol and that of the secure protocol for the millionaire problem.
Round Complexity: It is clear that the round complexity of the protocol is $`O\hspace{0pt}{(n)}`$. ∎
## 6. Conclusion
In this paper, we studied the point inclusion problem for polygons in $`2`$ dimension. The Protocol 4.2 for the star shaped domain has far better round complexity than the existing protocols. The Protocol 5.1 for the general polygonal domains is applicable for a large class of polygonal domains than the existing protocols. We hope to extend these ideas to more general domains and to higher dimensions.
## References
- \[1\] A.C. Yao. Protocols for secure computations. In Proceedings of the 23rd Annual IEEE Symposium on Foundations of Computer Science, Chicago, USA, pp.160-164, 1982.
- \[2\] O. Goldreich, S. Micali and A. Wigderson. How to play any mental game. In Proceedings of the 19th annual ACM Conference on Theory of computing, New York, pp.218-229, 1987.
- \[3\] O. Goldreich. Secure multi-party computation (working draft). Available from
http://citeseer.ist.psu.edu/goldreich98secure.html, 1998.
- \[4\] M. Fischlin. A cost-effecive pay-per multiplication comparison method for millonaires. Progress in Cryptology-CT-RSA 2001: Cryptographers Track at RSA Conference, pp.457-472, San Francisco: USA, April 8-12, 2001.
- \[5\] L. Yonglong, H. Liusheng. An efficient private comparison problem. Preprint.
- \[6\] I. Ioannidis, A. Grama. An efficient protocol for Yao’s millionaires’ problem, Proceedings of the 36th Hawaii International Conference on System Sciences, Jan 6-9, 2003.
- \[7\] H. Y. Lin, W. G. Tzeng. An efficient solution to the millionaires’ problem based on hormomorphic encryption. Lecture notes in Computer Science 3531, Springer-Verlag, pp.456-466, 2005.
- \[8\]
Y. Lindell, B. Pinkas. Privacy preserving data mining. Journal of Cryptology, 15(3), pp.177-206, 2002.
- \[9\]
R. Agrawal, R. Srikant. Privacy preserving data mining. In Proc. the 2000 ACM SIGMOND on Management of Data, ACM Press, pp.439-450, 2000.
- \[10\] R. Fagin, M. Naor and P. Winkler. Comparing information without leaking it. Communication of the ACM, 39, pp.77-85, 1996.
- \[11\] W. Du, M. J. Atallah. Secure Multiparty Computational Geometry, Lecture Notes in Computer Science, Proceedings, LNCS 2125, pp.165-179, Algorithms and Data Structures : 7th International Workshop, WADS 2001, Providence, RI, USA, August, 8-10, 2001,
- \[12\]
S. Goldwasser, Y. Lindell. Secure comptutation without agreement. Proc. 16th DISC Lecture Notes in Computer Science, 2508: 17-32, 2002.
- \[13\] W. Du and M. J. Atallah. Privacy-preserving cooperative scientific computations. In 14th IEEE Computer Security Foundations Workshop, Nova Scotia, Canada, June 11-13, 2001.
- \[14\] W. Du, M. J. Atallah, F. Kerschbaum. Protocols for secure remote database access with approximate matching. Technical report, 2001.
- \[15\] C. Cachin. Efficient private bidding and auctions with an oblivious third party. In Proceedings of the 6th ACM conference on Computer and communications security, pp. 120-127, Singapore, November 1-4, 1999.
- \[16\] W. Du, Mikhail J. Atallah. Privacy-Preserving Cooperative Statistical Analysis, Proceedings of the 17th Annual Computer Security Applications Conference, pp.102-110, New Orleans, USA, Dec 10-14, 2001.
- \[17\] W. Du, Y.S. Han, and S. Chen. Privacy-Preserving Multivariate Statistical Analysis: Linear Regression and Classification, Proc. 2004 SIAM Int l Conf. Data Mining (SDM04), Apr. 2004.
- \[18\] L. Kissner, D. Song. Privacy Preserving Set Operations. Advances in Cryptology: CRYPTO 2005, Lecture Notes in Computer Science, 3621, pp.241-257, 2005.
- \[19\]
S. D. Li, Y. Q. Dai. Secure two-party computational geometry. Journal of Computer Science and Technology, 20(2): pp.258-263, 2005.
- \[20\]
Y. L. Luo, L. S. Huang, H Zhong. Secure two-party point-circle inclusion problem. Journal of Computer Science and Technology, 22(1), pp.88-91, 2007.
- \[21\] J. Benaloh. Dense probabilistic encryption. In Proceedings of the Workshop on Selected Areas of Cryptography, pp.120-128, Kingston, ON, May 1994.
- \[22\] D. Naccache and J. Stern. A new cryptosystem based on higher residues. In Pro- ceedings of the 5th ACM Conference on Computer and Communications Security, pp.59-66, 1998.
- \[23\] T. Okamoto and S. Uchiyama. An efficient public-key cryptosystem. In Advances in Cryptology, EUROCRYPT 98, pp.308-318, 1998.
- \[24\] P. Paillier. Public-key cryptosystems based on composite degree residue classes. In Advances in Cryptology, EUROCRYPT 99, Lecture Notes in Computer Science, 1592 , pp 223-238, 1999.
- \[25\] B. Goethals, S. Laur, H. Lipmaa and T. Mielikainen. On Private Scalar Product Computation for Privacy-Preserving Data Mining, The 7th Annual International Conference in Information Security and Cryptology (ICISC 2004), Lecture Notes in Computer Science 3506 , pp.104-120, Seoul, Korea, December 2-3, 2004. Springer-Verlag.
- \[26\] Z.Yang, R. N. Wright, H. Subramaniam. Experimental Analysis of a Privacy-Preserving Scalar Product Protocol, International Journal of Computer Systems Science and Engineering, Vol. 21, No. 1, pp. 47 52, 2006.
- \[27\] F. P. Preparata, M. I. Shamos. Computational Geometry, An Introduction, Springer.
- \[28\] Wm. R. Franklin. Polygon Properties Calculated from the vertex neighborhoods, Annual Symposium on Computational Geometry, Proceedings of the third annual symposium on Computational geometry, Waterloo, Ontario, Canada, pp.110-118 Year of Publication, 1987. |
0706.0580 | cs.NI | cyber | # Efficient Batch Update of Unique Identifiers in a Distributed Hash Table for Resources in a Mobile Host
Yoo Chung
###### Abstract
Resources in a distributed system can be identified using identifiers based on random numbers. When using a distributed hash table to resolve such identifiers to network locations, the straightforward approach is to store the network location directly in the hash table entry associated with an identifier. When a mobile host contains a large number of resources, this requires that all of the associated hash table entries must be updated when its network address changes.
We propose an alternative approach where we store a host identifier in the entry associated with a resource identifier and the actual network address of the host in a separate host entry. This can drastically reduce the time required for updating the distributed hash table when a mobile host changes its network address. We also investigate under which circumstances our approach should or should not be used. We evaluate and confirm the usefulness of our approach with experiments run on top of OpenDHT.
## 1 Introduction
A distributed system needs a way to identify computing resources that it uses. A common way to identify them is to use a URL \[[4](#bib.bib4)\], which comprises a network address for the host and a path within the host.
<sup>1</sup>
<sup>1</sup>1In http://example.invalid/1/2/3/, for example, the network address for the host would be example.invalid while /1/2/3/ would be the path within the host.
However, some authors have argued that resource identifiers should contain little or no information, since otherwise an identifier would become invalid whenever there are changes in network locations, storage locations, naming policy, organization, etc. \[[3](#bib.bib3), [18](#bib.bib18), [19](#bib.bib19)\]
Using random numbers for resource identification avoids including such information in the identifier. It also makes it easy to allocate identifiers without having to manage the identifier space carefully, since identifier duplication is virtually impossible with a large enough identifier space.
<sup>2</sup>
<sup>2</sup>2Given a 160-bit identifier space, the probability that even a single duplicate identifier would be generated over a 100-year time period with a billion identifiers being generated per second is about $`10^{- 12}`$.
However, unique identifiers based on random numbers need to be resolved to actual network locations, which are often composed of a host address and a path within the host, in order for the represented resource to be used. One way to resolve such identifiers is to use distributed hash tables \[[1](#bib.bib1), [17](#bib.bib17)\]. Most distributed hash tables are scalable and are well suited for storing data indexed by random identifiers.
The most straightforward way to index a resource in a distributed hash table is to simply store an entry in the table with the resource identifier as the key and the network location of the resource as the value. However, this results in performance and latency issues when a mobile host contains many resources, since the network location in each entry must be updated independently whenever the host moves.
For data, this problem can be alleviated by using replication \[[5](#bib.bib5), [16](#bib.bib16)\]. However, there are many cases when replication is not a feasible option, some of which are listed below:
- •
Owners of nodes in a distributed hash table may not be willing to contribute large amounts of storage to store data for other people. For example, while they may be willing to store network locations of home videos, they may not be be willing to store the video files themselves.
- •
An identifier needs to identify a specific master copy of a file in a mobile host in order to ensure that updates to the file are immediately available to the host.
- •
The resource in question is inherently not replicable. For example, it could be a network service or sensor specific to a mobile host.
Instead of replication, an alternative approach is to use Mobile IP \[[12](#bib.bib12)\] or the Host Identity Protocol \[rfc4423\] to preserve the network address of mobile hosts. This requires support in the operating system and the network infrastructure. Such support is not widespread, however, so this approach may not be desirable for applications using distributed hash tables.
In this paper, we propose the use of indirect entries in a distributed hash table. An indirect entry contains a host identifier and a host-specific path. A host identifier is a random number which identifies a specific mobile host and is the key to a host entry in the distributed hash table. The network address of the mobile host is obtained from the host entry, which gives the actual network location when combined with the host-specific path. When a mobile host moves, only its host entry needs to be updated.
The remainder of the paper is organized as follows. We discuss related work in section [2](#S2 "2 Related work ‣ Efficient Batch Update of Unique Identifiers in a Distributed Hash Table for Resources in a Mobile Host"). Section [3](#S3 "3 Batch update for mobile hosts ‣ Efficient Batch Update of Unique Identifiers in a Distributed Hash Table for Resources in a Mobile Host") describes our proposal and discusses the circumstances under which it should or should not be used. We evaluate it against the straightforward approach in section [4](#S4 "4 Evaluation ‣ Efficient Batch Update of Unique Identifiers in a Distributed Hash Table for Resources in a Mobile Host") and conclude in section [5](#S5 "5 Conclusions ‣ Efficient Batch Update of Unique Identifiers in a Distributed Hash Table for Resources in a Mobile Host").
## 2 Related work
The use of random numbers for globally unique identifiers is not uncommon, which takes advantage of the fact that the probability of two different resources being assigned the same random number is extremely low for a large enough identifier space. For example, X.667 defines random-number-based UUIDs \[[21](#bib.bib21)\], while SPKI/SDSI uses hashes of public keys, which for identification purposes are similar to random numbers \[[6](#bib.bib6)\].
Ballintijn et al. argue that resource naming should be decoupled from resource identification \[[2](#bib.bib2)\]. Resources are named with human-friendly names, which are based on DNS \[[10](#bib.bib10)\], while identification is done with object handles, which are globally unique identifiers that need not contain network locations. They use DNS to resolve human-friendly names to object handles and a location service to resolve object handles to network locations. The location service uses a hierarchical architecture for resolving object handles. This two-level approach allows the naming of resources without worrying about replication or migration and the identification of resources without worrying about naming policies.
Walfish et al. argue for the use of semantic-free references for identifying web documents instead of URLs \[[19](#bib.bib19)\]. The reason is that changes in naming policies or ownership of DNS domain names often result in previous URLs pointing to unrelated or non-existent documents, even when the original documents still exist. Semantic-free references are hashes of public keys or other data, and are resolved to URLs using a distributed hash table based on Chord \[[17](#bib.bib17)\]. Using semantic-free references would allow web documents to link to each other without worrying about changes in the URLs of the documents.
Distributed hash tables, also called peer-to-peer structured overlay networks, are distributed systems which map a uniform distribution of identifiers to nodes in the system \[[1](#bib.bib1), [17](#bib.bib17), [22](#bib.bib22)\]. Nodes act as peers, with no node having to play a special role, and a distributed hash table can continue operation even as nodes join or leave the system. Lookups and updates to a distributed hash table are scalable, typically taking time logarithmic to the number of nodes in the system. We experimentally evaluated our work using OpenDHT \[[15](#bib.bib15)\], which is a public distributed hash table service based on Bamboo \[[14](#bib.bib14)\].
There has also been research on implementing distributed hash tables on top of mobile ad hoc networks \[[8](#bib.bib8), [9](#bib.bib9)\]. As with Mobile IP \[[12](#bib.bib12)\] and HIP \[rfc4423\], hosts in mobile ad hoc networks do not change their network address with movement, so there would be no need to update entries in a distributed hash table used for resolving resource identifiers. However, almost the entire Internet is not part of a mobile ad hoc network, so it is of little help to applications that need to run on current networks.
## 3 Batch update for mobile hosts
[Refer to caption](/html/0706.0580/assets/x1.png)
Figure 1: Representation of when identifiers are mapped directly to resources in a single mobile host. Circles denote identifiers, diamonds denote direct entries in the hash table, and squares denote network locations. All of the hash table entries must be updated whenever the mobile host moves.
The most straightforward way to map a unique identifier to an actual network location is to store a *direct entry* in the distributed hash table for each identifier, with the identifier as the key and the actual network location as the value. Resolution is done by simply looking up the identifier in the distributed hash table and using the resulting network location. Figure [1](#S3.F1 "Figure 1 ‣ 3 Batch update for mobile hosts ‣ Efficient Batch Update of Unique Identifiers in a Distributed Hash Table for Resources in a Mobile Host") illustrates this approach.
When the network address of a mobile host changes, this approach requires that entries for every resource contained by the host be updated independently. For a distributed hash table consisting of a constant number of nodes, this requires time linear to the number of resources in the mobile host. When the host contains a large number of resources, this can result in an unacceptably large delay before identifiers can be resolved to their updated location.
Instead of storing the network location directly in a hash table entry for a resource identifier, we propose the alternative approach of storing both an location-independent identifier for the mobile host and the host-specific path in an *indirect entry*. The *host identifier* identifies the mobile host which contains the resource and are random numbers as in ordinary resource identifiers. The distributed hash table contains a host entry which maps this identifier to the network address of the mobile host. The path identifies the specific resource within the host.
In this approach, we first find the corresponding indirect entry for the given resource identifier in the distributed hash table. Once the indirect entry is found, we find the corresponding host entry for the included host identifier. We then combine the network address of the host in the host entry and the path in the indirect entry to construct the network location of the desired resource.
<sup>3</sup>
<sup>3</sup>3When network locations are given as a HTTP URL, the network address of the host would be an IP address and the port number, while the path would simply be the URL path.
This requires two lookups to the distributed hash table, compared to a single lookup required for direct entries.
[Refer to caption](/html/0706.0580/assets/x2.png)
Figure 2: Representation of when identifiers are mapped indirectly to resources in a single mobile host. Circles denote identifiers, diamonds denote indirect entries in the hash table, the triangle denotes the host entry, and squares denote network locations. Only the host entry needs to be updated whenever the mobile host moves.
However, updating the distributed hash table when a mobile host changes its network address is much more efficient when using indirect entries compared to using direct entries. Unlike with direct entries, where every entry must be updated independently, only a single host entry needs to be updated when using indirect entries. This can greatly reduce the delay during which resource identifiers cannot be resolved to their correct network locations. Figure [2](#S3.F2 "Figure 2 ‣ 3 Batch update for mobile hosts ‣ Efficient Batch Update of Unique Identifiers in a Distributed Hash Table for Resources in a Mobile Host") illustrates the approach using indirect entries.
### 3.1 Using direct and indirect entries together
If a host contains only a very small number of resources or almost never changes its network address, then using direct entries would be more efficient because of the smaller lookup overhead. On the other hand, using indirect entries drastically reduces the update latency for a mobile host which contains a large number of resources and changes its network address frequently. Fortunately, both types of entries can be used simultaneously in a single distributed hash table.
Entries in the distributed hash table can be prepended by a magic number which identifies the type of entry they are. The magic numbers are used to distinguish among direct, indirect, and host entries. They also serve to prevent potential conflicts when the same distributed hash table is used for other applications besides resource identifier resolution. Table [1](#S3.T1 "Table 1 ‣ 3.1 Using direct and indirect entries together ‣ 3 Batch update for mobile hosts ‣ Efficient Batch Update of Unique Identifiers in a Distributed Hash Table for Resources in a Mobile Host") shows the entry types and their contents, while figure [3](#S3.F3 "Figure 3 ‣ 3.1 Using direct and indirect entries together ‣ 3 Batch update for mobile hosts ‣ Efficient Batch Update of Unique Identifiers in a Distributed Hash Table for Resources in a Mobile Host") describes the resolution procedure.
<table>
<thead>
<tr>
<th>Type</th>
<th>Content</th>
</tr>
</thead>
<tbody>
<tr>
<td>Direct entry</td>
<td>MD, network location</td>
</tr>
<tr>
<td>Indirect entry</td>
<td>MI, host identifier, path</td>
</tr>
<tr>
<td>Host entry</td>
<td>MH, host network address</td>
</tr>
</tbody>
</table>
Table 1: Entry types and their contents. MD, MI, MH are magic numbers for direct, indirect, and host entries, respectively.
1\.
Find entry indexed by the resource identifier in the distributed hash table.
2\.
If entry is direct entry, return with included network location.
3\.
If entry is indirect entry,
\(a\)
Find host entry indexed by the included host identifier.
\(b\)
Combine network address of host in the host entry and the path of the resource in the indirect entry to construct the network location of the resource.
\(c\)
Return with the network location.
4\.
Otherwise, return that the resource cannot be found.
Figure 3: Resource identifier resolution procedure. This procedure does not treat a host as a resource (extending it so that it does is trivial).
### 3.2 When to use direct or indirect entries
A host can choose whether to use direct or indirect entries for its resources depending on which approach performs better for its needs. But under which circumstances should the host choose which approach? This section discusses this in terms of lookup overhead and update latency.
[Refer to caption](/html/0706.0580/assets/x3.png)
Figure 4: Types of delays during concurrent get operations. $`c_{g}`$ is the minimum delay between the operations, while $`c_{r}`$ is the delay due to network latency.
Since get and put operations to the distributed hash table can be pipelined, where multiple operations may be handled concurrently as in figure [4](#S3.F4 "Figure 4 ‣ 3.2 When to use direct or indirect entries ‣ 3 Batch update for mobile hosts ‣ Efficient Batch Update of Unique Identifiers in a Distributed Hash Table for Resources in a Mobile Host"), we will consider the time costs $`c_{g}`$ or $`c_{p}`$ for an individual get or put operation separately from the fixed time costs $`c_{r}`$ or $`c_{q}`$ due to network latency in the get or put operation, which not only comes from accessing the distributed hash table externally but also from the communication among the distributed hash table nodes. We will assume that the number of nodes in the distributed hash table is constant so that these costs are also essentially constant.
When using direct entries, all entries referencing resources in a given host must be updated independently. With $`n`$ resources in a host, the migration time $`c_{m,d}`$ required to update all of the entries when it changes its network address is
<table>
<tbody>
<tr>
<td></td>
<td>cm,d=ncp+cqsubscript𝑐𝑚𝑑𝑛subscript𝑐𝑝subscript𝑐𝑞c_{m,d}=nc_{p}+c_{q}</td>
<td></td>
<td>(1)</td>
</tr>
</tbody>
</table>
On the other hand, only a single host entry needs to be updated when using indirect entries, so the migration time $`c_{m,i}`$ in this case is
<table>
<tbody>
<tr>
<td></td>
<td>cm,i=cp+cqsubscript𝑐𝑚𝑖subscript𝑐𝑝subscript𝑐𝑞c_{m,i}=c_{p}+c_{q}</td>
<td></td>
<td>(2)</td>
</tr>
</tbody>
</table>
A direct entry requires only a single get operation to resolve an identifier, whereas an indirect entry requires that it get the indirect entry first and then obtain the appropriate host entry (the two get operations cannot be done concurrently since the second operation is done based on the result from the first operation), so the respective lookup times $`c_{l,d}`$ and $`c_{l,i}`$ are
<table>
<tbody>
<tr>
<td></td>
<td>cl,dsubscript𝑐𝑙𝑑\displaystyle c_{l,d}</td>
<td>=\displaystyle=</td>
<td>cg+crsubscript𝑐𝑔subscript𝑐𝑟\displaystyle c_{g}+c_{r}</td>
<td></td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td>cl,isubscript𝑐𝑙𝑖\displaystyle c_{l,i}</td>
<td>=\displaystyle=</td>
<td>2(cg+cr)2subscript𝑐𝑔subscript𝑐𝑟\displaystyle 2(c_{g}+c_{r})</td>
<td></td>
</tr>
</tbody>
</table>
If there are $`r_{l}`$ lookups per unit time and $`r_{m}`$ migrations per unit time, then the overall time costs $`C_{d}`$ and $`C_{i}`$ per unit time when using direct and indirect entries, respectively, are
<table>
<tbody>
<tr>
<td></td>
<td>Cd=rlcl,d+rmcm,d=rl(cg+cr)+rm(ncp+cq)subscript𝐶𝑑subscript𝑟𝑙subscript𝑐𝑙𝑑subscript𝑟𝑚subscript𝑐𝑚𝑑subscript𝑟𝑙subscript𝑐𝑔subscript𝑐𝑟subscript𝑟𝑚𝑛subscript𝑐𝑝subscript𝑐𝑞C_{d}=r_{l}c_{l,d}+r_{m}c_{m,d}=r_{l}(c_{g}+c_{r})+r_{m}(nc_{p}+c_{q})</td>
<td></td>
</tr>
</tbody>
</table>
<table>
<tbody>
<tr>
<td></td>
<td>Ci=rlcl,i+rmcm,i=2rl(cg+cr)+rm(cp+cq)subscript𝐶𝑖subscript𝑟𝑙subscript𝑐𝑙𝑖subscript𝑟𝑚subscript𝑐𝑚𝑖2subscript𝑟𝑙subscript𝑐𝑔subscript𝑐𝑟subscript𝑟𝑚subscript𝑐𝑝subscript𝑐𝑞C_{i}=r_{l}c_{l,i}+r_{m}c_{m,i}=2r_{l}(c_{g}+c_{r})+r_{m}(c_{p}+c_{q})</td>
<td></td>
</tr>
</tbody>
</table>
When minimizing the overall time cost, it is better to use indirect entries when
<table>
<tbody>
<tr>
<td></td>
<td>Cisubscript𝐶𝑖\displaystyle C_{i}</td>
<td><\displaystyle<</td>
<td>Cdsubscript𝐶𝑑\displaystyle C_{d}</td>
<td></td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td>rl(cg+cr)subscript𝑟𝑙subscript𝑐𝑔subscript𝑐𝑟\displaystyle r_{l}(c_{g}+c_{r})</td>
<td><\displaystyle<</td>
<td>rm(n−1)cpsubscript𝑟𝑚𝑛1subscript𝑐𝑝\displaystyle r_{m}(n-1)c_{p}</td>
<td></td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td>rlrmsubscript𝑟𝑙subscript𝑟𝑚\displaystyle\frac{r_{l}}{r_{m}}</td>
<td><\displaystyle<</td>
<td>(n−1)cpcg+cr𝑛1subscript𝑐𝑝subscript𝑐𝑔subscript𝑐𝑟\displaystyle\frac{(n-1)c_{p}}{c_{g}+c_{r}}</td>
<td></td>
</tr>
</tbody>
</table>
One may also wish to give more weight to reducing migration times or lookup times. If we set the weights $`w_{m}`$ and $`w_{l}`$ by how much importance we attach to reducing migration times or lookup times, respectively, we can compute the weighted time costs as
<table>
<tbody>
<tr>
<td></td>
<td>Cd′=wlrlcl,d+wmrmcm,dsuperscriptsubscript𝐶𝑑′subscript𝑤𝑙subscript𝑟𝑙subscript𝑐𝑙𝑑subscript𝑤𝑚subscript𝑟𝑚subscript𝑐𝑚𝑑C_{d}^{\prime}=w_{l}r_{l}c_{l,d}+w_{m}r_{m}c_{m,d}</td>
<td></td>
</tr>
</tbody>
</table>
<table>
<tbody>
<tr>
<td></td>
<td>Ci′=wlrlcl,i+wmrmcm,isuperscriptsubscript𝐶𝑖′subscript𝑤𝑙subscript𝑟𝑙subscript𝑐𝑙𝑖subscript𝑤𝑚subscript𝑟𝑚subscript𝑐𝑚𝑖C_{i}^{\prime}=w_{l}r_{l}c_{l,i}+w_{m}r_{m}c_{m,i}</td>
<td></td>
</tr>
</tbody>
</table>
and then indirect entries should be used when
<table>
<tbody>
<tr>
<td></td>
<td>wlwm⋅rlrm<(n−1)cpcg+cr⋅subscript𝑤𝑙subscript𝑤𝑚subscript𝑟𝑙subscript𝑟𝑚𝑛1subscript𝑐𝑝subscript𝑐𝑔subscript𝑐𝑟\frac{w_{l}}{w_{m}}\cdot\frac{r_{l}}{r_{m}}<\frac{(n-1)c_{p}}{c_{g}+c_{r}}</td>
<td></td>
<td>(3)</td>
</tr>
</tbody>
</table>
Assuming a large $`n`$, with $`W`$ denoting the relative importance of reducing lookup times compared to migration times and $`R`$ denoting how often lookups occur compared to migrations, equation ([3](#S3.E3 "In 3.2 When to use direct or indirect entries ‣ 3 Batch update for mobile hosts ‣ Efficient Batch Update of Unique Identifiers in a Distributed Hash Table for Resources in a Mobile Host")) can be approximately rewritten as
<table>
<tbody>
<tr>
<td></td>
<td>WR<ncpcg+cr𝑊𝑅𝑛subscript𝑐𝑝subscript𝑐𝑔subscript𝑐𝑟WR<\frac{nc_{p}}{c_{g}+c_{r}}</td>
<td></td>
<td>(4)</td>
</tr>
</tbody>
</table>
Equation ([4](#S3.E4 "In 3.2 When to use direct or indirect entries ‣ 3 Batch update for mobile hosts ‣ Efficient Batch Update of Unique Identifiers in a Distributed Hash Table for Resources in a Mobile Host")) agrees with our intuition that direct entries should be used when migration times do not matter or when migrations are rare, and that indirect entries should be used when migration times do matter and happen often for mobile hosts with a large number of resources. It also gives a concrete forumula for deciding whether to use direct or indirect entries.
## 4 Evaluation
[Refer to caption](/html/0706.0580/assets/x4.png)
Figure 5: Experimental setup. The mobile host, which can move around and change its network address, accesses OpenDHT through a gateway which is one of the nodes in the distributed hash table.
In order to evaluate how using direct and indirect entries perform in a real network, we conducted experiments on OpenDHT \[[15](#bib.bib15)\]. OpenDHT is a public distributed hash table service which runs on about 200 nodes in PlanetLab \[fiuczynski:sigops2006\]. We used the service by selecting a single gateway to the distributed hash table and accessing it with the XML-RPC \[[20](#bib.bib20)\] interface throughout our experiments. Here the mobile host is not part of the distributed hash table, similar to how clients are separate from the distributed hash tables in SFR \[[19](#bib.bib19)\] and CoDoNS \[[13](#bib.bib13)\].
<sup>4</sup>
<sup>4</sup>4In cases where mobile hosts are part of the distributed hash table, the overhead for updating the routing tables should also be considered.
Figure [5](#S4.F5 "Figure 5 ‣ 4 Evaluation ‣ Efficient Batch Update of Unique Identifiers in a Distributed Hash Table for Resources in a Mobile Host") illustrates our experimental setup.
We compared lookup times and migration times when using direct entries and indirect entries for a single host. The host, a 2.16 GHz Intel Core 2 Duo with 1GB of memory connected to the Internet via Ethernet, was migrated between two network addresses. Resource identifiers were mapped to URLs that point to files. A URL was stored directly in a direct entry, while only the URL path and a host identifier was stored in an indirect entry, with the IP address of the host being stored in a host entry.
<table>
<thead>
<tr>
<th></th>
<th>Lookup time (s)</th>
</tr>
</thead>
<tbody>
<tr>
<th>Direct</th>
<td>0.53±0.63plus-or-minus0.530.630.53\pm 0.63</td>
</tr>
<tr>
<th>Indirect</th>
<td>1.13±0.60plus-or-minus1.130.601.13\pm 0.60</td>
</tr>
</tbody>
</table>
Table 2: Lookup times and their standard deviations.
We first measured the lookup times for resolving an identifier to a URL. Since lookup for a direct entry requires exactly a single get operation and lookup for an indirect entry requires exactly two get operations, lookup times do not depend on the number of resources in a host.
<sup>5</sup>
<sup>5</sup>5The time for a get operation should be constant for a distributed hash table with a fixed number of nodes, since network latency dominates the time.
Thus we measured the average lookup times required by direct and indirect entries by first inserting entries for 5000 resource identifiers into the distributed hash table and then resolving randomly selected identifiers 2000 times for each case. As expected, the average lookup time for indirect entries was roughly twice that of direct entries as can be seen in table [2](#S4.T2 "Table 2 ‣ 4 Evaluation ‣ Efficient Batch Update of Unique Identifiers in a Distributed Hash Table for Resources in a Mobile Host").
Next, we measured the migration times when using direct or indirect entries with varying numbers of resources in the host. For each number of resources, we first put in the entries for each resource into the distributed hash table. We then migrated the host 100 times and measured the average migration time.
When updating direct entries, 100 entries were updated concurrently, which is much faster than updating each entry one by one. Using significantly larger amounts of concurrency was problematic because the gateway to OpenDHT had problems handling the number of connections.
Also, we selected the entry with the largest remaining time-to-live value when retrieving entries from OpenDHT. This entry is the one that is most up-to-date since we used a fixed TTL value for all entries. We did not have to worry about individual entries becoming large enough to skew the results
<sup>6</sup>
<sup>6</sup>6OpenDHT returns all unexpired values that have been associated with a key.
since we alternated the host between only two network addresses.
[Refer to caption](/html/0706.0580/assets/x5.png)
Figure 6: Migration times for up to 100 resources in the mobile host. The error bars denote the standard deviation for each case.
[Refer to caption](/html/0706.0580/assets/x6.png)
Figure 7: Migration times for up to 1000 resources in the mobile host. The error bars denote the standard deviation for each case.
[Refer to caption](/html/0706.0580/assets/x7.png)
Figure 8: Migration times for up to 5000 resources in the mobile host. The error bars denote the standard deviation for each case.
Our results for the migration times are shown in figures [6](#S4.F6 "Figure 6 ‣ 4 Evaluation ‣ Efficient Batch Update of Unique Identifiers in a Distributed Hash Table for Resources in a Mobile Host"), [7](#S4.F7 "Figure 7 ‣ 4 Evaluation ‣ Efficient Batch Update of Unique Identifiers in a Distributed Hash Table for Resources in a Mobile Host"), and [8](#S4.F8 "Figure 8 ‣ 4 Evaluation ‣ Efficient Batch Update of Unique Identifiers in a Distributed Hash Table for Resources in a Mobile Host"), where the host contained up to 100, 1000, and 5000 resources, respectively. We can see that migrating direct entries takes time linear to the number of entries in the host, as is expected from equation ([1](#S3.E1 "In 3.2 When to use direct or indirect entries ‣ 3 Batch update for mobile hosts ‣ Efficient Batch Update of Unique Identifiers in a Distributed Hash Table for Resources in a Mobile Host")). On the other hand, the time required for updating indirect entries is essentially constant, as is expected from equation ([2](#S3.E2 "In 3.2 When to use direct or indirect entries ‣ 3 Batch update for mobile hosts ‣ Efficient Batch Update of Unique Identifiers in a Distributed Hash Table for Resources in a Mobile Host")).
While migration with indirect entries took only about a second, migration with direct entries took over 4 minutes and a half with 5000 resources contained in the host. Even with only 10 resources, it took about 9 seconds longer to update direct entries compared to the one second it takes to migrate with indirect entries, which can be a significant difference for interactive applications.
Mobile hosts could contain even more resources than what was tried in our experiments. For example, the home directory in a personal machine of one of the authors contains more than 60,000 files. A straightforward extrapolation from our results suggests that this case would require almost an hour for migration when using direct entries.
These results show that the drastic reduction in migration time by using indirect entries over direct entries can be worth the small increase in lookup time required when resolving indirect entries.
## 5 Conclusions
When identifying resources in a distributed system using identifiers based on random numbers, the most straightforward way to resolve identifiers with a distributed hash table is to store the network location directly in the entry keyed by the identifier. However, when a mobile host which contains multiple non-replicable resources changes its network address, all of the associated entries in the distributed hash table must be updated.
When the number of resources in the mobile host is large, updating all of the entries so that remote hosts can properly use resources in the mobile host can take a long time. Therefore we proposed an alternative approach, where the entry keyed by the resource identifier contained only a host identifier and host-specific path for the resource, and the host identifier itself is a key to a host entry containing the actual network address for the mobile host.
With our proposed approach, only the host entry needs to be updated when the mobile host changes its network address. This can drastically reduce the delay during which its resources cannot be resolved to their current network locations, as was shown theoretically and experimentally. However, there is a small increase in the time required for resolving identifiers with our approach, so we also discussed under which circumstances it should not be used.
In our work, we only consider whether to use direct or indirect entries given static lookup and migration rates. It would be interesting to see how an adaptive system could dynamically adjust the approach used in order to achieve optimal performance with changing lookup and migration rates, since such a system would also have to consider the overhead for switching between the two.
It may also be possible to apply system-specific optimizations to our approach. For example, while our approach can be applied to any type of distributed hash table, it could be possible to reduce the lookup overhead when it is applied on top of OpenDHT by taking advantage of the ReDiR framework.
We plan to apply our approach in a decentralized and unified naming system we are developing, where it would improve performance for identifying resources such as files and network services located inside mobile devices in a persistent and location-independent manner.
## References
- \[1\] Stephanos Androutsellis-Theotokis and Diomidis Spinellis. A survey of peer-to-peer content distribution technologies.
ACM Computing Surveys, 36(4):335–371, December 2004.
- \[2\] Gerco Ballintijn, Maarten van Steen, and Andrew S. Tanenbaum. Scalable human-friendly resource names.
IEEE Internet Computing, 5(5):20–27, September 2001.
- \[3\] Tim Berners-Lee. Cool URIs don’t change.
http://www.w3.org/Provider/Style/URI.html, 1998.
- \[4\] Tim Berners-Lee, Larry Masinter, and Mark McCahill. RFC 1738: Uniform resource locators (URL), December 1994.
- \[5\] Frank Dabek, M. Frans Kaashoek, David Karger, Robert Morris, and Ion Stoica. Wide-area cooperative storage with CFS.
In Proceedings of the Eighteenth ACM Symposium on Operating Systems Principles, pages 202–215, 2001.
- \[6\] Carl M. Ellison, Bill Frantz, Butler Lampson, Ron Rivest, Brian Thomas, and Tatu Ylonen. RFC 2693: SPKI certificate theory, September 1999.
- \[7\] Marc E. Fiuczynski. PlanetLab: Overview, history, and future directions.
ACM SIGOPS Operating Systems Review, 40(1):6–10, January 2006.
- \[8\] Tobias Heer, Stefan Götz, Simon Rieche, and Klaus Wehrle. Adapting distributed hash tables for mobile ad hoc networks.
In Proceedings of the 4th Annual IEEE International Conference on Pervasive Computing and Communications Workshops. IEEE Computer Society Press, March 2006.
- \[9\] Olaf Landsiedel, Stefan Götz, and Klaus Wehrle. A churn and mobility resistant approach for DHTs.
In Proceedings of the 1st International Workshop on Decentralized Resource Sharing in Mobile Computing and Networking, pages 42–47. ACM Press, 2006.
- \[10\] Paul V. Mockapetris and Kevin J. Dunlap. Development of the domain name system.
ACM SIGCOMM Computer Communication Review, 18(4):123–133, August 1988.
- \[11\] Robert Moskowitz and Pekka Nikander. RFC 4423: Host identity protocol (HIP) architecture, May 2006.
- \[12\] Charles E. Perkins. RFC 3344: IP mobility support for IPv4, August 2002.
- \[13\] Venugopalan Ramasubramanian and Emin Gün Sirer. The design and implementation of a next generation name service for the Internet.
ACM SIGCOMM Computer Communication Review, 34(4):331–342, October 2004.
- \[14\] Sean Rhea, Dennis Geels, Timothy Roscoe, and John Kubiatowicz. Handling churn in a DHT.
In Proceedings of the USENIX Annual Technical Conference, June 2004.
- \[15\] Sean Rhea, Brighten Godfrey, Brad Karp, John Kubiatowicz, Sylvia Ratnasamy, Scott Shenker, Ion Stoica, and Harlan Yu. OpenDHT: A public DHT service and its uses.
In Proceedings of the 2005 Conference on Applications, Technologies, Architectures, and Protocols for Computer Communications, pages 73–84. ACM Press, 2005.
- \[16\] Antony Rowstron and Peter Druschel. Storage management and caching in PAST, a large-scale, persistent peer-to-peer storage utility.
In Proceedings of the Eighteenth ACM Symposium on Operating Systems Principles, pages 188–201, 2001.
- \[17\] Ion Stoica, Robert Morris, David Liben-Nowell, David R. Karger, M. Frans Kaashoek, Frank Dabek, and Hari Balakrishnan. Chord: A scalable peer-to-peer lookup protocol for internet applications.
IEEE/ACM Transactions on Networking, 11(1):17–32, February 2003.
- \[18\] Maarten van Steen, Franz J. Hauck, Philip Homburg, and Andrew S. Tanenbaum. Locating objects in wide-area systems.
IEEE Communications Magazine, 36(1):104–109, January 1998.
- \[19\] Michael Walfish, Hari Balakrishnan, and Scott Shenker. Untangling the Web from DNS.
In Proceedings of the First Symposium on Networked Systems Design and Implementation, pages 225–238, March 2004.
- \[20\] Dave Winer. XML-RPC specification.
http://www.xmlrpc.com/spec, June 1999.
- \[21\] Procedures for the operation of OSI registration authorities: Generation and registration of universally unique identifiers (UUIDs) and their use as ASN.1 object identifier components. ITU-T Recommendation X.667, September 2004.
- \[22\] Ben Y. Zhao, Ling Huang, Jeremy Stribling, Sean C. Rhea, Anthony D. Joseph, and John D. Kubiatowicz. Tapestry: A resilient global-scale overlay for service deployment.
IEEE Journal on Selected Areas in Communications, 22(1):41–53, January 2004. |
0706.1755 | cs.CR | cyber | # FreeBSD Mandatory Access Control Usage for Implementing Enterprise Security Policies
Kirill Bolshakov Elena Reshetova Saint-Petersburg State University of Aerospace Instrumentation
kirill.bolshakov@ieee.org elena.reshetova@gmail.com
###### Abstract
FreeBSD was one of the first widely deployed free operating systems to provide mandatory access control. It supports a number of classic MAC models. This tutorial paper addresses exploiting this implementation to enforce typical enterprise security policies of varying complexities.
## I Introduction
Security needs of organizations are becoming more and more sophisticated nowadays. Most general-purpose operating systems (GPOS) provide access control policies to meet these needs. There are cases when the traditionally deployed Discretionary Access Control (DAC) rules are not sufficient: they tend to quickly become unmanageable in the case of large installations, and also are not enough for controlling information flows. This is when the Mandatory Access Control (MAC) comes in: it provides for better manageability and directly targets the information flows. In their turn, the information flows address the confidentiality and integrity needs of information security within an organization. Until very recently, the GPOSes tended to provide various flavors of DAC only. The FreeBSD OS \[[1](#bib.bib1)\] was one of the first widely deployed open source GPOSes to support MAC \[[2](#bib.bib2), [3](#bib.bib3)\]. In this paper, a number of organizational policy examples are implemented in the environment of the FreeBSD MAC.
The authors strongly believe that in order to implement a sound MAC policy it is important to understand MAC’s mathematical foundations. These foundations were set by Denning in \[[4](#bib.bib4)\]. There also exists a terminology confusion between MAC and LBAC (lattice-based access control). These models are the same, because MAC security labels \[[5](#bib.bib5)\] directly correspond to security classes of lattice-based models (this was also pointed to by Sandhu \[[6](#bib.bib6)\] and Osborn \[[7](#bib.bib7)\]).
Let us first address the definition of the information flow. According to Denning and Sandhu, the security policies regulate how the information “flows from one object to another”. A typical object is a shared memory segment, a file system object or a network packet. Obviously, controlling the information flows is important to prevent the leakage of the confidential information, the one usually sought by insiders. Another goal is the forgery prevention, so that no untrusted reports are ever submitted to the top level of the organization hierarchy, and no top-ranking company officers take any unchecked or untrusted information into account during decision making.
To implement the information flow control, every object is assigned a security label (also called a security classification), implemented by the FreeBSD file label. When the information flows from one object into another, *an information flow* from the security class of the first object to the security class of the second one also takes place. Whether the information flow is allowed is regulated by the relation between the object security classes. The subjects are the entities performing the information transfer between the objects. In our case, a subject appears when a user logs in to the system and is assigned a set of privileges. As we are considering MAC, the set of privileges is rendered as security clearance. It is implemented by the FreeBSD user label.
This paper is organized as follows. In the next section an example of an organization and its document flow is described. The following sections implement organization’s information security goals, which gradually increase in complexity. The information security goals specify the target effect: preserving data and process integrity, restricting access to the confidential information, or implementing a consulting services policy. For every security goal, a corresponding classic MAC model or a combination of them is chosen. The models are then implemented in the FreeBSD MAC framework.
## II The Example
The example organization we consider is a very small technology company. There are only six positions within the organization. The organization chart and the exchange of files corresponding to the information flows are shown in Fig. [1](#S2.F1 "Figure 1 ‣ II The Example ‣ FreeBSD Mandatory Access Control Usage for Implementing Enterprise Security Policies"). The document flow is bidirectional.
[Refer to caption](/html/0706.1755/assets/x1.png)
Figure 1: Example Organization Chart and Document Flows
The file system folder layout reflects this, and the folders are listed in Table [I](#S2.T1 "Table I ‣ II The Example ‣ FreeBSD Mandatory Access Control Usage for Implementing Enterprise Security Policies"). Operations will be applied to the files in these folders. The Temp folder is used for file exchange purposes. The administrator may consider configuring this folder with the “sticky” attribute (so that the files could be deleted by their owners only). The “U” (from “Untrusted”) folders are used by Mary, Alice and Robert for delivering documents to Jane and Paul, respectively. After the managers review these untrusted documents and deem them trustworthy, the files may be copied to the corresponding no-prefix folders while increasing their security class. This makes these documents available for further processing to the higher levels of hierarchy: for example, they may be included into Jane’s or Paul’s reports for John.
In FreeBSD, the MAC labels are assigned to the user accounts via the notion of *login classes.* The mapping of user accounts to their login classes is given in the master.passwd file, where the user accounts are traditionally stored. The login classes are listed alongside the group information. The login classes themselves are specified in the login.conf file. In the examples, we will list only the MAC-related configuration of the login classes. The MAC labels used in FreeBSD have the syntax of policy1/qualifier1,policy2/qualifier2,…, where the policy component describes the policy module (e.g., Biba or MLS), and the qualifier describes a corresponding grade, compartments, and a combined grade-compartment range \[[8](#bib.bib8)\].
Table I: Folder Layout
<table>
<tbody>
<tr>
<td>AccountingGoals</td>
<td>StrategicTechnologyGoals</td>
</tr>
<tr>
<td>AccountingReports</td>
<td>SummarySalesReports</td>
</tr>
<tr>
<td>SummaryTechnicalReports</td>
<td>TechnicalReports</td>
</tr>
<tr>
<td>FeatureRequests</td>
<td>Temp</td>
</tr>
<tr>
<td>SalesGoals</td>
<td>UAccountingReports</td>
</tr>
<tr>
<td>SalesReports</td>
<td>USalesReports</td>
</tr>
<tr>
<td>StrategicSalesGoals</td>
<td>UTechnicalReports</td>
</tr>
</tbody>
</table>
In the following sections the increasingly complex information flow control goals are considered. The goals are then translated into a MAC policy sufficient to fulfill them, and then into a FreeBSD implementation.
## III Protecting the Integrity
Suppose that the organization needs to protect itself from making decisions based on the untrusted data. That is, John should not make decisions based on the data directly received from Alice or Robert. This corresponds to the John’s subject (which is his user login) having a higher integrity label: he should be prevented from inadvertently reading a document of the lower integrity label than his own. Also, John’s requests should propagate through the organization unchanged: his documents cannot be tampered with by the subjects with a lower integrity. In the information flow sense, the information is allowed to flow from the higher integrity to the lower integrity objects, but the reverse flow is disallowed. This is the Biba model with the liberal $`\star`$-property \[[6](#bib.bib6)\]. It is directly supported in FreeBSD by the mac_biba module. The assignment of the labels to the files in the case of the Biba model support will look like shown in Table [VII](#S3.T7 "Table VII ‣ III Protecting the Integrity ‣ FreeBSD Mandatory Access Control Usage for Implementing Enterprise Security Policies"), and the user label assignment is shown in Table [VII](#S3.T7 "Table VII ‣ III Protecting the Integrity ‣ FreeBSD Mandatory Access Control Usage for Implementing Enterprise Security Policies"). The Temp folder is excluded from the policy by setting its grade to equal, because it serves as a document exchange location.
Table II: Files: Biba
Table III: Files: Biba and MLS
Table IV: Files: Biba, MLS and Compartments
<table>
<tbody>
<tr>
<td>AccountingGoals</td>
<td>biba/2</td>
</tr>
<tr>
<td>AccountingReports</td>
<td>biba/5</td>
</tr>
<tr>
<td>SummaryTechnicalReports</td>
<td>biba/10</td>
</tr>
<tr>
<td>FeatureRequests</td>
<td>biba/2</td>
</tr>
<tr>
<td>SalesGoals</td>
<td>biba/2</td>
</tr>
<tr>
<td>SalesReports</td>
<td>biba/5</td>
</tr>
<tr>
<td>StrategicSalesGoals</td>
<td>biba/5</td>
</tr>
<tr>
<td>StrategicTechnologyGoals</td>
<td>biba/5</td>
</tr>
<tr>
<td>SummarySalesReports</td>
<td>biba/10</td>
</tr>
<tr>
<td>TechnicalReports</td>
<td>biba/5</td>
</tr>
<tr>
<td>Temp</td>
<td>biba/equal</td>
</tr>
<tr>
<td>UAccountingReports</td>
<td>biba/2</td>
</tr>
<tr>
<td>USalesReports</td>
<td>biba/2</td>
</tr>
<tr>
<td>UTechnicalReports</td>
<td>biba/2</td>
</tr>
</tbody>
</table>
<table>
<tbody>
<tr>
<td>AccountingGoals</td>
<td>biba/2,mls/low</td>
</tr>
<tr>
<td>AccountingReports</td>
<td>biba/5,mls/low</td>
</tr>
<tr>
<td>SummaryTechnicalReports</td>
<td>biba/10,mls/50</td>
</tr>
<tr>
<td>FeatureRequests</td>
<td>biba/2,mls/50</td>
</tr>
<tr>
<td>SalesGoals</td>
<td>biba/2,mls/low</td>
</tr>
<tr>
<td>SalesReports</td>
<td>biba/5,mls/low</td>
</tr>
<tr>
<td>StrategicSalesGoals</td>
<td>biba/5,mls/low</td>
</tr>
<tr>
<td>StrategicTechnologyGoals</td>
<td>biba/5,mls/50</td>
</tr>
<tr>
<td>SummarySalesReports</td>
<td>biba/10,mls/low</td>
</tr>
<tr>
<td>TechnicalReports</td>
<td>biba/5,mls/50</td>
</tr>
<tr>
<td>Temp</td>
<td>biba/equal,mls/equal</td>
</tr>
<tr>
<td>UAccountingReports</td>
<td>biba/2,mls/low</td>
</tr>
<tr>
<td>USalesReports</td>
<td>biba/2,mls/low</td>
</tr>
<tr>
<td>UTechnicalReports</td>
<td>biba/2,mls/50</td>
</tr>
</tbody>
</table>
<table>
<tbody>
<tr>
<td>AccountingPlans</td>
<td>biba/2,mls/50:1</td>
</tr>
<tr>
<td>AccountingReports</td>
<td>biba/5,mls/50:1</td>
</tr>
<tr>
<td>CommonTechnicalReports</td>
<td>biba/10,mls/50:2</td>
</tr>
<tr>
<td>FeatureRequests</td>
<td>biba/2,mls/50:2</td>
</tr>
<tr>
<td>FinancialReports</td>
<td>biba/10,mls/50:1</td>
</tr>
<tr>
<td>SalesPlans</td>
<td>biba/2,mls/50:1</td>
</tr>
<tr>
<td>SalesReports</td>
<td>biba/5,mls/50:1</td>
</tr>
<tr>
<td>StrategicMarketingGoals</td>
<td>biba/5,mls/50:1</td>
</tr>
<tr>
<td>StrategicTechnologyGoals</td>
<td>biba/5,mls/50:2</td>
</tr>
<tr>
<td>TechnicalReports</td>
<td>biba/5,mls/50:2</td>
</tr>
<tr>
<td>Temp</td>
<td>biba/equal,mls/equal</td>
</tr>
<tr>
<td>UAccountingReports</td>
<td>biba/2,mls/50:1</td>
</tr>
<tr>
<td>USalesReports</td>
<td>biba/2,mls/50:1</td>
</tr>
<tr>
<td>UTechnicalReports</td>
<td>biba/2,mls/50:2</td>
</tr>
</tbody>
</table>
Table V: Users: Biba
Table VI: Users: Biba and MLS
Table VII: Users: Biba, MLS and Compartments
<table>
<tbody>
<tr>
<td>John</td>
<td>biba/10(10-10)</td>
</tr>
<tr>
<td>Jane</td>
<td>biba/5(2-10)</td>
</tr>
<tr>
<td>Paul</td>
<td>biba/5(2-10)</td>
</tr>
<tr>
<td>Alice</td>
<td>biba/2(2-2)</td>
</tr>
<tr>
<td>Mary</td>
<td>biba/2(2-2)</td>
</tr>
<tr>
<td>Robert</td>
<td>biba/2(2-2)</td>
</tr>
</tbody>
</table>
<table>
<tbody>
<tr>
<td>John</td>
<td>biba/10(10-10),mls/100(100-100)</td>
</tr>
<tr>
<td>John.Sales</td>
<td>biba/10(10-10),mls/low(low-low)</td>
</tr>
<tr>
<td>John.Engineering</td>
<td>biba/10(10-10),mls/50(50-50)</td>
</tr>
<tr>
<td>Jane</td>
<td>biba/5(2-10),mls/low(low-low)</td>
</tr>
<tr>
<td>Paul</td>
<td>biba/5(2-10),mls/50(50-50)</td>
</tr>
<tr>
<td>Alice</td>
<td>biba/2(2-2),mls/low(low-low)</td>
</tr>
<tr>
<td>Mary</td>
<td>biba/2(2-2),mls/low(low-low)</td>
</tr>
<tr>
<td>Robert</td>
<td>biba/2(2-2),mls/50(50-50)</td>
</tr>
</tbody>
</table>
<table>
<tbody>
<tr>
<td>John</td>
<td>biba/10(10-10),mls/100:1+2(100-100:1+2)</td>
</tr>
<tr>
<td>John.Lower</td>
<td>biba/10(10-10),mls/50:1+2(50-50:1+2)</td>
</tr>
<tr>
<td>Jane</td>
<td>biba/5(2-10),mls/50:1(50:1-50:1)</td>
</tr>
<tr>
<td>Paul</td>
<td>biba/5(2-10),mls/50:2(50:2-50:2)</td>
</tr>
<tr>
<td>Alice</td>
<td>biba/2(2-2),mls/50:1(50:1-50:1)</td>
</tr>
<tr>
<td>Mary</td>
<td>biba/2(2-2),mls/50:1(50:1-50:1)</td>
</tr>
<tr>
<td>Robert</td>
<td>biba/2(2-2),mls/50:2(50:2-50:2)</td>
</tr>
</tbody>
</table>
While the Biba policy enforces the integrity, it also disallows the reverse flow of information. It is obvious that the management should be aware of what is going on in the managed departments, so the reverse flow is required for the considered example. This is achievable in FreeBSD by adding *ranged labels*. The ranged labels allow the subjects to change both their grade and compartment within some range: the grade in the numeric one and the compartment in the set one. The syntax of these labels is biba/effectivegrade:effectivecompartments(lograde:locompartments-higrade:hicompartments). If John’s subject were assigned the biba/10(10-10) label, there would be no way for him to access the document labeled biba/2. But if the subject’s label were biba/10(2-10), he would be able to downgrade it to the necessary grade of 2 to read the document. For the case of the considered example, this situation is avoided. Instead, we use the notion of a *trusted entity*. This entity is a subject that can promote the document integrity grade, and is trusted by the higher integrity subject. In our example, this subject’s user should be able to verify the documents submitted by the subjects with the lower integrity, and if possible and needed, to promote the document integrity grade. Jane and Paul are assigned these privilege and responsibility, and are assigned the label of biba/5(2-10). Thus, Jane can downgrade her integrity grade to 2 to read Mary’s report, decide that it can be trusted and is worth for John to see, and then copy it to John while promoting its integrity grade to 10. This is achievable by the following sequence of actions:
- •
Mary creates a Report1 file in the UAccountingReports folder.
- •
She moves the file to the Temp folder.
- •
Jane’s current integrity label is biba/5(2-10). She downgrades it to biba/2 using a command like setpmac biba/2 sh, and copies the Report1 file to her home folder. The ownership of the document’s copy belongs to Jane.
- •
Now Jane can read the report and check it for mistakes or forged data.
- •
After she decides that the document can be forwarded to John, she changes its integrity label to biba/10 by using the setfmac biba/10 Report1 command. She then promotes her own clearance to biba/10 as well, and moves the document to the SummarySalesReports folder. That folder has the integrity label of biba/10 and is accessible to John.
- •
Now John can read the document.
The described policy protects the integrity of the data, and makes sure that the top management decisions are not affected by any unverified data. We proceed by adding the confidentiality to the organization’s information security goals.
## IV Adding Confidentiality Levels
As the company we consider is a technology one, it is rather possible that it generates the intellectual property (IP). The Engineering department (in our case, Paul and Robert) generates the company’s IP. John should have access to it, while Jane and her team should not. However, John may have access to the more confidential information than Paul. For example, he may be working with the company’s partners on the subject of a common patent or elaborating a shared strategic technology plan. Also, not only should the confidential documents be kept away from the persons with the lower clearance levels, but the persons with the higher clearance levels should be prevented from disclosing the confidential information they know, either knowingly or inadvertently. The policy to govern this kind of information flow may look like the following:
- •
A confidential object with the higher security classification cannot be read by a subject with the lower clearance.
- •
A subject with the higher clearance cannot write to an object with the lower security classification.
This policy matches the one provided by the Bell-LaPadula model with the liberal $`\star`$-property \[[9](#bib.bib9)\] (or the closely related BLP model \[[6](#bib.bib6)\]). The Bell-LaPadula model assigns a clearance label per every confidentiality grade in the organization. It is very similar in its structure to the Biba model. For the considered case, there should be three clearance labels: Secret, Confidential, and Public, with the latter one assigned to Jane and her team. The Bell-LaPadula model is supported by the FreeBSD mac_mls module (further referred to as MLS).
Also, it is still required to preserve the integrity, and do not loose the work done in the previous section. FreeBSD allows this by providing the capability of running multiple MAC policy modules simultaneously. Thus, it is possible for the Biba and the Bell-LaPadula policies to be in effect at the same time, and to preserve both integrity and confidentiality.
For the FreeBSD implementation, we map the Secret clearance to the grade of 100, Confidential to 50, and Public to low. The updated file labels are shown in Table [VII](#S3.T7 "Table VII ‣ III Protecting the Integrity ‣ FreeBSD Mandatory Access Control Usage for Implementing Enterprise Security Policies"), and the user labels are shown in Table [VII](#S3.T7 "Table VII ‣ III Protecting the Integrity ‣ FreeBSD Mandatory Access Control Usage for Implementing Enterprise Security Policies").
Compare the user labels assignment for the Biba and the Bell-LaPadula policies. There are two new user logins for the Bell-LaPadula model: John.Sales and John.Engineering. As per the confidentiality configuration, while being in the CEO position, John cannot directly publish the information for the Sales and the Engineering, for his clearance level is the highest one. By logging in as the lower clearance subject, John will be unable to access his documents, and this will prevent him from inadvertently (or any malicious software on his computer from knowingly) disclosing secret and confidential information to the lower clearance subjects.
## V Splitting the Competence Areas
Jane has the aggregate data of the sales activity of the company. While the data of every individual salesperson cannot give the whole picture, the aggregate reveals the current state of the company sales. Thus, it might be desirable to classify the aggregate information as Confidential. In the considered example, simply adding an MLS grade would not help, because the MLS grades are totally ordered, and the new labels will conflict with the Engineering department: either the engineers will be able to read the Sales’ documents, or vice versa (this depends on whose grade is higher). In other words, the Sales and the Engineering should have their own hierarchies with a single point of intersection: John should have access to the documents of both departments. This corresponds to the set-based domination relation: for two sets $`S_{1}`$ and $`S_{2}`$, $`S_{1}`$ dominates $`S_{2}`$ (further written as $`S_{1} \succ S_{2}`$) iff $`S_{2} \subset S_{1}`$. Obviously, incomparable labels are supported by this definition: $`{\{ 1,2,3\}} \subset {\{ 1,2,3,5\}}`$, but $`\{ 5,9\}`$ and $`\{ 5,6,7\}`$ are incomparable. In the considered example, three labels are sufficient: $`{{S_{\text{Sales}} = {\{ 1\}}},{S_{\text{Eng}} = {\{ 2\}}}},`$ and $`S = {\{ 1,2\}}`$, where label $`S`$ belongs to John, so that he can access documents belonging to any of the departments: $`S \succ S_{\text{Sales}}`$ and $`{S \succ S_{\text{Eng}}}.`$
FreeBSD supports the set-based domination relation by the notion of *compartments*. Compartments are the sets of integers, which can be added to the Biba or the MLS labels. In the example, the labels are directly mapped to the compartments. The MLS parts of the labels may be rearranged, as they are not used for distinguishing the Engineering and the Sales any longer. Instead, the MLS grades set the confidentiality grades within individual compartments. The resulting folder and user labels are shown in Tables [VII](#S3.T7 "Table VII ‣ III Protecting the Integrity ‣ FreeBSD Mandatory Access Control Usage for Implementing Enterprise Security Policies") and [VII](#S3.T7 "Table VII ‣ III Protecting the Integrity ‣ FreeBSD Mandatory Access Control Usage for Implementing Enterprise Security Policies"), respectively. The John.Sales and John.Engineering logins are removed, while a new one is added: John.Lower. This new login has the same MLS grade as Jane’s and Paul’s ones, and is used for creating the documents with their confidentiality classification.
With these settings, John is the only person who accesses the secret information. All other users have access to the confidential information, and none of them deals with creating publicly accessible documents. John has access to both compartments. While it is possible for him to disclose the information of one compartment to another one, this will not violate the confidentiality grades, and John is considered a trusted entity for managing the inter-department information flows.
## VI The Chinese Wall Policy
The Chinese Wall policy first formalized by Brewer and Nash \[[10](#bib.bib10)\] is oriented toward the commercial sector. This policy provides for prevention of information flows, which cause conflicts of interest for individual consultants working for the same consulting company. As soon as a consultant has obtained access to the confidential information of a client company, he or she must have their access rights to all other companies from this industry sector revoked. However, the consultant must still have access to the public information of all client companies. The consultant starts with the access rights to all information of all client companies. As soon as the consultant accesses the information of a company $`C`$ from an industry $`I`$, he is then denied access to the confidential information of all other companies from the industry $`I`$. However, the consultant’s access to any information of any company from other industries is allowed. Thus, the conflict of interest does not take place. The deficiencies and important enhancements of the original Chinese Wall policy model are described by Sandhu in \[[11](#bib.bib11)\]. In the same paper, the author demonstrates that the enhanced Chinese Wall policy can be represented in the Bell-LaPadula model, and is therefore a lattice-based policy. While discussing a possible implementation of this policy on FreeBSD, we will follow the definitions given in \[[6](#bib.bib6)\].
Let the classes of conflict of interest (i.e. industries) be represented as a set $`{I = \left\{ I_{1},I_{2},\ldots,I_{N} \right\}},`$ and the number of companies in a single industry assumed equal and be denoted by $`C`$. A clearance label will be then represented by an $`N`$-element vector $`{{l = {\{{l\hspace{0pt}{\lbrack 1\rbrack}},{l\hspace{0pt}{\lbrack 2\rbrack}},\ldots,{l\hspace{0pt}{\lbrack N\rbrack}}\}}},{{\forall k} \in \overline{1,N}}}:{{l\hspace{0pt}{\lbrack k\rbrack}} \in {I_{k} \cup {\{\bot\}}}}`$, where $`\bot`$ in position $`k`$ denotes the public information of any client company belonging to the industry $`k`$ ($`\bot`$ is to be read as null). For example, a label for an object containing confidential information from the company 3 in the industry 2 and the company 2 from the industry 4 is represented as $`\lbrack\bot,3,\bot,2,\ldots,\bot\rbrack`$. There is a dominance relation for these labels defined as follows: $`l^{1} \geq l^{2}`$ if $`\forall k \in \overline{1,N}:\left( l^{1}{\lbrack k\rbrack} = l^{2}{\lbrack k\rbrack} \right) \vee \left( l^{1}{\lbrack k\rbrack} \neq \bot \land l^{2}{\lbrack k\rbrack} = \bot \right).`$ Thus, $`\lbrack 1,1,2\rbrack > \lbrack 1,\bot,2\rbrack`$ and $`\lbrack 1,1,\bot\rbrack > \lbrack 1,\bot,\bot\rbrack`$, but $`\lbrack 1,1,2\rbrack`$ and $`\lbrack 1,2,2\rbrack`$ are incomparable. The label of all nulls $`\lbrack\bot,\ldots,\bot\rbrack`$ is dominated by all other labels. The highest label is denoted by SYSHIGH. The sample lattice for the case of 2 industries and 2 companies per industry is shown in Fig. [2](#S6.F2 "Figure 2 ‣ VI The Chinese Wall Policy ‣ FreeBSD Mandatory Access Control Usage for Implementing Enterprise Security Policies").
\psfrag{EaaR8WYHeq74GME}\[cc\]\[cc\]{\$\mbox{SYSHIGH}\$}\psfrag{sbGvYCyCBhoPN18Wb}\[cc\]\[cc\]{\$\left\[1,1\right\]\$}\psfrag{hWwKpYpSapezvtPBE}\[cc\]\[cc\]{\$\left\[1,2\right\]\$}\psfrag{t0CKNgt5UjzblEjut}\[cc\]\[cc\]{\$\left\[2,1\right\]\$}\psfrag{qOtXiJ7lVlTyJYOJL}\[cc\]\[cc\]{\$\left\[2,2\right\]\$}\psfrag{yCGsl3nqX1m5eDWpR5Nc}\[cc\]\[cc\]{\$\left\[1,\bot\right\]\$}\psfrag{LRUk54AWWycck_G4TqiT}\[cc\]\[cc\]{\$\left\[\bot,1\right\]\$}\psfrag{dkMy9m9CCgwgL6rIhvgC}\[cc\]\[cc\]{\$\left\[2,\bot\right\]\$}\psfrag{qDVkRyoragSrXim2wkvP}\[cc\]\[cc\]{\$\left\[\bot,2\right\]\$}\psfrag{Mrl07bWju4spovG3EkBxFnU}\[cc\]\[cc\]{\$\left\[\bot,\bot\right\]\$}\psfrag{Sxn7SnSbWKbC2on4p7LawOrV}\[cc\]\[cc\]{\$\mbox{mls/high:1+2+3+4}\$}\psfrag{kaefOglSUFf385_XR7}\[cc\]\[cc\]{\$\mbox{mls/20:1+3}\$}\psfrag{xRBNE8km2AGDuKPRkq}\[cc\]\[cc\]{\$\mbox{mls/20:1+2}\$}\psfrag{lfxWs6Xrbl9dA9qIaL}\[cc\]\[cc\]{\$\mbox{mls/20:3+4}\$}\psfrag{M5qxp30lvURZYljl7S}\[cc\]\[cc\]{\$\mbox{mls/20:2+4}\$}\psfrag{ssKJ_ov587y1ubbx}\[cc\]\[cc\]{\$\mbox{mls/10:1}\$}\psfrag{cHV1HP6OjMKr9EY0}\[cc\]\[cc\]{\$\mbox{mls/10:3}\$}\psfrag{BlbtzOOesbDJSpkO}\[cc\]\[cc\]{\$\mbox{mls/10:4}\$}\psfrag{ISH9FmBxfmyfeFRQ}\[cc\]\[cc\]{\$\mbox{mls/10:2}\$}\psfrag{Td9nvyi0HrF1a0g}\[cc\]\[cc\]{\$\mbox{mls/low}\$}\includegraphics\[height=216.81pt,width=433.62pt\]{fig_CW.ps}
Figure 2: Example Chinese Wall Lattice and Its Mapping to FreeBSD Compartments
It is important to understand how the lattice structure is completed by defining the class-combining join operator $`\oplus`$. Two labels $`l^{1}`$ and $`l^{2}`$ are called *compatible* if $`\forall k \in \overline{1,N}:l^{1}{\lbrack k\rbrack} = l^{2}{\lbrack k\rbrack} \vee l^{1}{\lbrack k\rbrack} = \bot \vee l^{2}{\lbrack k\rbrack} = \bot.`$ This also means that all comparable labels are compatible, but there are incomparable labels which are compatible, like $`\lbrack 1,\bot,2\rbrack`$ and $`\lbrack\bot,3,2\rbrack.`$ Incompatible labels cannot be combined, while for compatible labels the following rule exists: $`{{l^{1} \oplus l^{2}} = \left\{ {{\forall k} \in \overline{1,N}}:{{\text{if}\hspace{0pt}l^{1}\hspace{0pt}{\lbrack k\rbrack}} \neq {\bot{\text{then}\hspace{0pt}l^{1}\hspace{0pt}{\lbrack k\rbrack}\hspace{0pt}\text{else}\hspace{0pt}l^{2}\hspace{0pt}{\lbrack k\rbrack}}}} \right\}}.`$ For example, $`{{\lbrack\bot,3,1\rbrack \oplus \lbrack 2,\bot,1\rbrack} = \lbrack 2,3,1\rbrack}.`$ We will denote the labels $`l^{1}`$ and $`l^{2}`$ as being compatible by $`\hat{C}\hspace{0pt}{(l^{1},l^{2})}`$. Understanding this definition will be required during feasibility analysis for the FreeBSD implementation of the policy.
As a user progresses upwards the domination relation over the lattice, his rights become more and more restrictive. The user leaves a trail of subjects during this progress. For example, for user Mary, the access to public information is granted to the subject “Mary”. As soon as Mary starts working on a project with a bank A, a new subject “Mary.Banks.A” is created for her. When she starts working on a project with an oil company B, another subject is created for her: “Mary.Banks.A.Oil.B”. While in a session with “Mary.Banks.A.Oil.B” subject, Mary can read the public documents from the banking and oil industries, but cannot write to them. To write to the public documents for the oil industry, she has to close her current session and initiate a new one with the subject “Mary”.
The assignment of the user names may be assumed to be performed manually (for example, by a system administrator). The FreeBSD implementation of the Chinese Wall policy is based on this fact. All possible login classes should be created in advance, so that adding a new user will not lead to rebuilding the login classes. The set of login classes required is $`{({C + 1})}^{N},`$ and one of those classes is a public one. The number of compartments required is $`C \times N`$. The FreeBSD MAC implementation imposes a limit on the number of compartments: it should be lower than 256. This limitation should be taken into account when checking a specific organization Chinese Wall policy for feasibility. To prevent the user from applying the dominance relation to the classes dominated according to the compartment set inclusion relation, MLS labels are added. The number of MLS grades should be $`N + 1`$ (taking into account the public one, and with the exception of the SYSHIGH class).
The rule for labeling the login classes on FreeBSD closely follows the $`\oplus`$ operator considered earlier in this section. The labeling algorithm is shown in Fig. [3](#S6.F3 "Figure 3 ‣ VI The Chinese Wall Policy ‣ FreeBSD Mandatory Access Control Usage for Implementing Enterprise Security Policies"). As a result of applying this algorithm to the original lattice, a set of triplets is constructed. A triplet contains an MLS grade, a FreeBSD compartments set, and a corresponding original Chinese Wall label. The algorithm iterates over the lattice by examining the set of labels with a fixed non-$`\bot`$ set of positions on every step. The previously constructed triplets are examined for containing the original labels dominated by the current one. The corresponding FreeBSD compartment sets are combined and added to the currently constructed triplet.
1:
2:
$`L^{1}\leftarrow\left\{ l:\exists i:{(l{\lbrack i\rbrack} \neq \bot)} \land {(\forall j \neq i:l{\lbrack j\rbrack} = \bot)} \right\},M^{1}\leftarrow\varnothing,i\leftarrow 1`$
3:
4:
for all $`l \in L^{1}`$ do
5:
6:
$`M^{1}\leftarrow{M^{1} \cup {\langle 10,i,l\rangle}}`$
7:
8:
$`i\leftarrow{i + 1}`$
9:
end for
10:
11:
for $`i\leftarrow{2,N}`$ do
12:
13:
$`g\leftarrow{\{{{\text{if}\hspace{0pt}i} \neq {{{N\hspace{0pt}\text{then}\hspace{0pt}i} \times 10}\hspace{0pt}\text{else}\hspace{0pt}\text{high}}}\}}`$
14:
15:
$`X^{i}\leftarrow`$ set of all $`i`$-subsets of $`{{{\{ 1,\ldots,N\}},L^{i}}\leftarrow\varnothing},{M^{i}\leftarrow\varnothing}`$
16:
17:
for all $`x \in X^{i}`$ do $`L^{i}\leftarrow L^{i} \cup \left\{ l:{(\forall j \in x:l{\lbrack j\rbrack} \neq \bot)} \land {(\forall j \notin x:l{\lbrack j\rbrack} = \bot)},j = \overline{1,N} \right\}`$
18:
Now $`L^{i}`$ contains all lattice node labels with $`i`$ positions being not $`\bot`$.
19:
end for
20:
21:
for all $`l \in L^{i}`$ do
22:
23:
$`M'\leftarrow{\{{{\forall m'} \in M^{i - 1}}:{m' = {{\langle g',f',l'\rangle} \land {\hat{C}\hspace{0pt}{(l,l')}}}}\}}`$
24:
25:
$`M^{i}\leftarrow{M^{i} \cup {\{{\langle g,{\bigcup_{m' \in M'}f'},l\rangle}\}}}`$
26:
end for
27:
end for
28:
29:
$`M\leftarrow{\bigcup_{i}M^{i}}`$
Figure 3: FreeBSD Label Generation for Chinese Wall Policy
The system administrator has to create a set of login classes for the combined Bell-LaPadula and compartment labels. When a user starts working with a company from a new industry, the administrator provides the user name corresponding to the changed Chinese Wall lattice label.
## VII Conclusion
The set of the policies implemented by the FreeBSD/TrustedBSD project allows to implement not only the Bell-LaPadula and Biba models. By reusing the result of the Chinese Wall model being a lattice-based one, it can be implemented in the FreeBSD MAC framework as well. Also, the level of MAC support allows implementing trusted entities, so that the information flows are possible both ways, but in a controllable manner, so that the policies are not violated. The combination of a number of MAC models provides a very flexible platform for satisfying the needs of various organizations. The authors hope that the proper illustration of FreeBSD MAC support capabilities will promote the usage of mandatory access control in the commercial sector.
## Acknowledgment
The authors are grateful to Alexey Fomin whose comments and suggestions have significantly improved the paper.
## References
- \[1\] “FreeBSD project. FreeBSD home page,” http://www.FreeBSD.org/.
- \[2\] R. N. M. Watson, “TrustedBSD: Adding trusted operating system features to FreeBSD.” in *USENIX Annual Technical Conference, FREENIX Track*, C. Cole, Ed. USENIX, 2001, pp. 15–28.
- \[3\] R. Watson, W. Morrison, C. Vance, and B. Feldman, “The TrustedBSD MAC framework: Extensible kernel access control for FreeBSD 5.0.” in *USENIX Annual Technical Conference, FREENIX Track*. USENIX, 2003, pp. 285–296.
- \[4\] D. E. Denning, “A lattice model of secure information flow.” *Commun. ACM*, vol. 19, no. 5, pp. 236–243, 1976.
- \[5\] *Department of Defense Trusted Computer System Evaluation Criteria. DoD 5200.28-STD*, Department of Defense National Computer Security Center, 1985.
- \[6\] R. S. Sandhu, “Lattice-based access control models.” *IEEE Computer*, vol. 26, no. 11, pp. 9–19, 1993.
- \[7\] S. L. Osborn, “Information flow analysis of an RBAC system.” in *SACMAT*, 2002, pp. 163–168.
- \[8\] *The FreeBSD Documentation Project. FreeBSD Handbook*, The FreeBSD Project, http://www.FreeBSD.org/doc/en_US.ISO8859-1/books/handbook/.
- \[9\] D. E. Bell and L. J. LaPadula, “Secure computer systems: A mathematical model, volume II.” *Journal of Computer Security*, vol. 4, no. 2/3, pp. 229–263, 1996.
- \[10\] D. F. C. Brewer and M. J. Nash, “The Chinese Wall security policy.” in *IEEE Symposium on Security and Privacy*, 1989, pp. 206–214.
- \[11\] R. S. Sandhu, “Lattice-based enforcement of Chinese Walls.” *Computers & Security*, vol. 11, no. 8, pp. 753–763, 1992. |
0705.0817 | cs.NI | cyber | # Quantum Shortest Path Netsukuku
AlpT (@freaknet.org)
http://netsukuku.freaknet.org
AlpT (@freaknet.org)
###### Abstract
This document describes the QSPN, the routing discovery algorithm used by Netsukuku. Through a deductive analysis the main proprieties of the QSPN are shown. Moreover, a second version of the algorithm, is presented.
This document is part of Netsukuku.
Copyright ©2007 Andrea Lo Pumo aka AlpT $`<`$alpt@freaknet.org$`>`$. All rights reserved.
This document is free; you can redistribute it and/or modify it under the terms of the GNU General Public License as published by the Free Software Foundation; either version 2 of the License, or (at your option) any later version.
This document is distributed in the hope that it will be useful, but WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for more details.
You should have received a copy of the GNU General Public License along with this document; if not, write to the Free Software Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
###### Contents
1. [](#S1 "In Quantum Shortest Path Netsukuku")
1 Preface
2. [](#S2 "In Quantum Shortest Path Netsukuku")
2 The general idea
1. [](#S2.SS1 "In 2 The general idea ‣ Quantum Shortest Path Netsukuku")
2.1 The network model
2. [](#S2.SS2 "In 2 The general idea ‣ Quantum Shortest Path Netsukuku")
2.2 The routing algorithm
3. [](#S2.SS3 "In 2 The general idea ‣ Quantum Shortest Path Netsukuku")
2.3 The QSPN
3. [](#S3 "In Quantum Shortest Path Netsukuku")
3 Network topology
1. [](#S3.SS1 "In 3 Network topology ‣ Quantum Shortest Path Netsukuku")
3.1 Fractal topology
4. [](#S4 "In Quantum Shortest Path Netsukuku")
4 Tracer Packet
1. [](#S4.SS1 "In 4 Tracer Packet ‣ Quantum Shortest Path Netsukuku")
4.1 Tracer Packet flood
2. [](#S4.SS2 "In 4 Tracer Packet ‣ Quantum Shortest Path Netsukuku")
4.2 Proprieties of the tracer packet
5. [](#S5 "In Quantum Shortest Path Netsukuku")
5 Routes of a graph
6. [](#S6 "In Quantum Shortest Path Netsukuku")
6 Acyclic Tracer Packet flood
7. [](#S7 "In Quantum Shortest Path Netsukuku")
7 Routes simplification
1. [](#S7.SS1 "In 7 Routes simplification ‣ Quantum Shortest Path Netsukuku")
7.1 Simplification rules
2. [](#S7.SS2 "In 7 Routes simplification ‣ Quantum Shortest Path Netsukuku")
7.2 General results
3. [](#S7.SS3 "In 7 Routes simplification ‣ Quantum Shortest Path Netsukuku")
7.3 The question
8. [](#S8 "In Quantum Shortest Path Netsukuku")
8 Continuous Tracer Packet
1. [](#S8.SS1 "In 8 Continuous Tracer Packet ‣ Quantum Shortest Path Netsukuku")
8.1 Reflected CTP
9. [](#S9 "In Quantum Shortest Path Netsukuku")
9 QSPN v2
1. [](#S9.SS1 "In 9 QSPN v2 ‣ Quantum Shortest Path Netsukuku")
9.1 Interesting information
2. [](#S9.SS2 "In 9 QSPN v2 ‣ Quantum Shortest Path Netsukuku")
9.2 Live routes simplification
3. [](#S9.SS3 "In 9 QSPN v2 ‣ Quantum Shortest Path Netsukuku")
9.3 Cyclicity
1. [](#S9.SS3.SSS1 "In 9.3 Cyclicity ‣ 9 QSPN v2 ‣ Quantum Shortest Path Netsukuku")
9.3.1 Subcycles examples
4. [](#S9.SS4 "In 9 QSPN v2 ‣ Quantum Shortest Path Netsukuku")
9.4 Finiteness
5. [](#S9.SS5 "In 9 QSPN v2 ‣ Quantum Shortest Path Netsukuku")
9.5 Routes limit
6. [](#S9.SS6 "In 9 QSPN v2 ‣ Quantum Shortest Path Netsukuku")
9.6 Scalability
1. [](#S9.SS6.SSS1 "In 9.6 Scalability ‣ 9 QSPN v2 ‣ Quantum Shortest Path Netsukuku")
9.6.1 TP Classes
2. [](#S9.SS6.SSS2 "In 9.6 Scalability ‣ 9 QSPN v2 ‣ Quantum Shortest Path Netsukuku")
9.6.2 Subcycle filter
3. [](#S9.SS6.SSS3 "In 9.6 Scalability ‣ 9 QSPN v2 ‣ Quantum Shortest Path Netsukuku")
9.6.3 Efficiency order
7. [](#S9.SS7 "In 9 QSPN v2 ‣ Quantum Shortest Path Netsukuku")
9.7 Bandwidth issues
8. [](#S9.SS8 "In 9 QSPN v2 ‣ Quantum Shortest Path Netsukuku")
9.8 Worst case
10. [](#S10 "In Quantum Shortest Path Netsukuku")
10 QSPN v1
1. [](#S10.SS1 "In 10 QSPN v1 ‣ Quantum Shortest Path Netsukuku")
10.1 $`\mathbf{Q}`$ vs $`\mathbf{Q^{2}}`$
11. [](#S11 "In Quantum Shortest Path Netsukuku")
11 Network dynamics
1. [](#S11.SS1 "In 11 Network dynamics ‣ Quantum Shortest Path Netsukuku")
11.1 Extended Tracer Packet
12. [](#S12 "In Quantum Shortest Path Netsukuku")
12 QSPN optimisations
1. [](#S12.SS1 "In 12 QSPN optimisations ‣ Quantum Shortest Path Netsukuku")
12.1 Rtt and bandwidth
1. [](#S12.SS1.SSS1 "In 12.1 Rtt and bandwidth ‣ 12 QSPN optimisations ‣ Quantum Shortest Path Netsukuku")
12.1.1 Rtt delay
2. [](#S12.SS1.SSS2 "In 12.1 Rtt and bandwidth ‣ 12 QSPN optimisations ‣ Quantum Shortest Path Netsukuku")
12.1.2 Asymmetry in $`Q^{2}`$
2. [](#S12.SS2 "In 12 QSPN optimisations ‣ Quantum Shortest Path Netsukuku")
12.2 Disjoint routes
3. [](#S12.SS3 "In 12 QSPN optimisations ‣ Quantum Shortest Path Netsukuku")
12.3 Cryptographic QSPN
13. [](#S13 "In Quantum Shortest Path Netsukuku")
13 Simulating the QSPN v2
14. [](#S14 "In Quantum Shortest Path Netsukuku")
14 TODO
15. [](#S15 "In Quantum Shortest Path Netsukuku")
15 ChangeLog
## 1 Preface
The first part of the document describes the reasoning which led us to the construction of the current form of the QSPN v2. If you are just interested in the description of the QSPN v1 and v2 and you already know the concept of the Tracer Packet, you can directly skip to section [8](#S8 "8 Continuous Tracer Packet ‣ Quantum Shortest Path Netsukuku").
## 2 The general idea
The aim of Netsukuku is to be a (physical) scalable mesh network, completely distributed and decentralised, anonymous and autonomous.
The software, which must be executed by every node of the net, has to be unobtrusive. It has to use very few CPU and memory resources, in this way it will be possible to run it inside low-performance computers, like Access Points, embedded devices and old computers.
If this requirements are met, Netsukuku can be easily used to build a worldwide distributed, anonymous and not controlled network, separated from the Internet, without the support of any servers, ISPs or control authorities.
### 2.1 The network model
Netsukuku prioritises the stability and the scalability of net: the network has to be able to grow to even $`2^{2^{7}}`$ nodes.
A completely dynamic network would requires rapid and frequent updates of the routes and this is in contrast with the stability and the scalability requirements of Netsukuku. For this reason, we restrict Netsukuku to the case where a node won’t change its physical location quickly or often.
This assumption is licit, because the location of a wifi node mounted on top of a building won’t change and its only dynamic actions would be the joining and the disconnection to and from the network and the changes of the quality of its wifi links. However, there are some consequences of this assumption:
1. 1\.
Mobiles node aren’t supported by Netsukuku algorithms.
<sup>1</sup>
<sup>1</sup>1It is possible to use other mesh network protocols designed for mobility in conjunction with Netsukuku, in the same way they are used in conjunction with the Internet (f.e. see http://olsrd.org ).
2. 2\.
The network isn’t updated quickly: several minutes may be required before all the nodes become aware of a change of the network (new nodes have joined, more efficient routes have become available, …). However, when a node joins the network, it can reach all the other nodes from the first instant, using the routes of its neighbours.
### 2.2 The routing algorithm
One of the most important parts of Netsukuku, is the routing discovery algorithm, which is responsible to find all the most efficient routes of the network. These routes will permit to each node to reach any other node.
The routing algorithm must be capable to find the routes without overloading the network or the nodes’ CPU and memory resources.
### 2.3 The QSPN
Netsukuku implements its own algorithm, the *QSPN* (Quantum Shortest Path Netsukuku). The name derives from the way of working of its principal component: the *TP* (Tracer Packet), a packet which gains a “quantum” of information at each hop.
The QSPN is based on the assumptions described in section [2.1](#S2.SS1 "2.1 The network model ‣ 2 The general idea ‣ Quantum Shortest Path Netsukuku").
## 3 Network topology
The QSPN alone wouldn’t be capable of handling the whole network, because it would still require too much memory. For example, even if we store just one route to reach one node and even if this route costs one byte, we would need 1Gb of memory for a network composed by $`10^{9}`$ nodes (the current Internet).
For this reason, it’s necessary to structure the network in a convenient topology.
### 3.1 Fractal topology
Netsukuku, adopts a fractal like structure: 256 nodes are grouped inside a *group node* (gnode), 256 group nodes are grouped in a single *group of group nodes* (ggnode), 256 group of group nodes are grouped in a gggnode, and so on. (We won’t analyse the topology of Netsukuku. You can find more information about it in the proper document: \[[2](#bib.bib2)\]).
Since each gnode acts as a single real node, the QSPN is able to operate independently on each level of the fractal.
Since in each level there are a maximum of 256 (g)nodes, the QSPN will always operate on a maximum of 256 (g)nodes, therefore we would need just to be sure that it works as expected on every cases of a graph composed by $`\leq 256`$ nodes. By the way, we’ll directly analyse the general case.
For the sake of simplicity, in this paper, we will assume to operate on level 0 (the level formed by 256 single nodes).
## 4 Tracer Packet
A *TP* (Tracer Packet) is the fundamental concept on which the QSPN is based: it is a packet which stores in its body the IDs of the traversed hops.
### 4.1 Tracer Packet flood
A TP isn’t sent to a specific destination but instead, it is used to flood the network. By saying “the node A sends a TP” we mean that “the node A is starting a TP flood”.
A TP flood passes only once through each node of the net: a node which receives a TP will forward it to all its neighbours, except the one from which it received the TP. Once a node has forwarded a TP, it will not forward any other TPs of the same flood.
### 4.2 Proprieties of the tracer packet
1. 1\.
A node $`D`$ which received a TP, can know the exact route covered by the TP. Therefore, $`D`$ can know the route to reach the source node $`S`$, which sent the TP, and the routes to reach the nodes standing in the middle of the route.
For example, suppose that the TP received by $`D`$ is: $`\left\{ S,A,B,C,D \right\}`$. By looking at the packet $`D`$ will know that the route to reach $`B`$ is $`C\rightarrow B`$, to reach $`A`$ is $`C\rightarrow B\rightarrow A`$, and finally to reach $`S`$ is $`C\rightarrow B\rightarrow A\rightarrow S`$. The same also applies for all the other nodes which received the TP, f.e, $`B`$ knows that its route to reach $`S`$ is $`A\rightarrow S`$.
2. 2\.
The *bouquet of $`S`$* is the set of all the TPs which will be forwarded or sent by the node $`S`$ during the flood. The first TP of this bouquet received by a generic node $`D`$, will be the TP which covered the fastest route which connects $`S`$ to $`D`$. The fastest $`S\rightarrow D`$ route is the route with the minimum *rtt* (Round-Trip Time) between $`S`$ and $`D`$. This property is also valid if $`S`$ is the node which started the TP flood, i.e. the first node which sent the first bouquet of the TP flood.
#### Example
[Refer to caption](/html/0705.0817/assets/x1.png)
Figure 1: A simple graph
Suppose that $`D`$ sends a TP. The TP will cover this routes: $`D\rightarrow E\rightarrow F`$ and $`D\rightarrow C\rightarrow B\rightarrow A`$. When the TP reaches the node $`F`$ and the node $`A`$, the flood will stop, because either $`A`$ and $`F`$ won’t be able to forward the TP to any other node.
At the end, $`A`$ will know the route $`A\rightarrow B\rightarrow C\rightarrow D`$ and $`F`$ will know the route $`F\rightarrow E\rightarrow D`$.
## 5 Routes of a graph
Given a graph $`\mathbf{G}`$ we want to find all the existing routes between a node and all the other nodes.
Let $`N`$ be a generic node. Starting from $`N`$ we explore the entire graph until we re-enter in a cycle already visited or we cannot proceed any further. This approach is similar to the Depth-First Search\[[3](#bib.bib3)\] algorithm, but instead of searching for a specific goal, we just traverse the entire graph. Note that a cycle is traversed only once, because we need non redundant routes. In other words, if we already know the $`S\rightarrow A\rightarrow B\rightarrow C\rightarrow D`$ route, it’s useless to known that we can reach $`D`$ with the $`S\rightarrow A\rightarrow B\rightarrow C\rightarrow A\rightarrow B\rightarrow C\rightarrow D`$ route.
This is the pseudo code of the algorithm:
```
generate_routes(G) {
forall node in G
/* Starts the exploration of the graph from the ‘‘node’’ of the
graph ‘‘G’’ and print all its routes */
walk(node, node)
}
/* Print all the routes which start from the node ‘N’ */
walk(N, branch) {
deepened=0
forall L in N.links
/* L is a neighbour of N */
if(L in branch)
/* If ‘‘L’’ is already contained in the explored
branch, we’ve found a cycle. Since we just need to
traverse only once a cycle, we skip this ‘‘L’’ node
and continue to consider the other neighbours
of N */
continue;
newbranch=branch + L /* Append in the explored branch the
‘‘L’’ node. */
walk(L, newbranch) /* Recursively explore the new
branch */
/* Indicate that we’ve deepened in the graph at least once */
deepened=1
if(!deepened)
/* We haven’t deepened in the above for, this means that the
current branch can’t be explored anymore, therefore it is a
valid route. Print it */
print branch
}
```
A proof of concept of the above algorithm has been implemented in Awk \[[4](#bib.bib4)\].
### Example
Consider this graph:
[Refer to caption](/html/0705.0817/assets/x2.png)
Figure 2: A simple graph with one segment and one cycle
Given this graph as input the algorithm will output:
<table>
<tbody>
<tr>
<td></td>
<td>A→B→D→C→𝐴𝐵→𝐷→𝐶\displaystyle A\rightarrow B\rightarrow D\rightarrow C</td>
<td></td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td>A→B→D→E→𝐴𝐵→𝐷→𝐸\displaystyle A\rightarrow B\rightarrow D\rightarrow E</td>
<td></td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td>A→C→D→B→𝐴𝐶→𝐷→𝐵\displaystyle A\rightarrow C\rightarrow D\rightarrow B</td>
<td></td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td>A→C→D→E→𝐴𝐶→𝐷→𝐸\displaystyle A\rightarrow C\rightarrow D\rightarrow E</td>
<td></td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td>B→A→C→D→E→𝐵𝐴→𝐶→𝐷→𝐸\displaystyle B\rightarrow A\rightarrow C\rightarrow D\rightarrow E</td>
<td></td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td>B→D→C→A→𝐵𝐷→𝐶→𝐴\displaystyle B\rightarrow D\rightarrow C\rightarrow A</td>
<td></td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td>B→D→E→𝐵𝐷→𝐸\displaystyle B\rightarrow D\rightarrow E</td>
<td></td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td>C→A→B→D→E→𝐶𝐴→𝐵→𝐷→𝐸\displaystyle C\rightarrow A\rightarrow B\rightarrow D\rightarrow E</td>
<td></td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td>C→D→B→A→𝐶𝐷→𝐵→𝐴\displaystyle C\rightarrow D\rightarrow B\rightarrow A</td>
<td></td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td>C→D→E→𝐶𝐷→𝐸\displaystyle C\rightarrow D\rightarrow E</td>
<td></td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td>D→B→A→C→𝐷𝐵→𝐴→𝐶\displaystyle D\rightarrow B\rightarrow A\rightarrow C</td>
<td></td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td>D→C→A→B→𝐷𝐶→𝐴→𝐵\displaystyle D\rightarrow C\rightarrow A\rightarrow B</td>
<td></td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td>D→E→𝐷𝐸\displaystyle D\rightarrow E</td>
<td></td>
</tr>
</tbody>
</table>
## 6 Acyclic Tracer Packet flood
We can consider each route given by the output of the above algorithm as a single Tracer Packet. In fact, it is possible to implement the same algorithm using a slightly modified version of the TP flood, called the Acyclic TP flood:
The flood is not restricted like in a normal TP flood: one or more ATP can pass from the same node. The end of the flood is given by this rule: a node will not forward to any of its neighbours the ATP if its node ID is already present in the route contained in the body of the packet. With this rule an ATP can walk in a cycle only once, hence the name. Finally, like in the normal TP, a node doesn’t forward the ATP to the neighbour from which it has received the packet itself.
If every node of the network sends an ATP flood, then every node will get all the possible routes to reach any other node.
As you can see, the ATP flood performs a “live” version of the algorithm described in section [5](#S5 "5 Routes of a graph ‣ Quantum Shortest Path Netsukuku"). Obviously this is far from an efficient routing discovery algorithm, but it represents a good start.
## 7 Routes simplification
Looking carefully at the example output ([2](#S5.F2 "Figure 2 ‣ Example ‣ 5 Routes of a graph ‣ Quantum Shortest Path Netsukuku")) of the Generate Route algorithm, we can notice that many routes are higly redundant, in other words, some routes are almost the same. Consider for example the following four routes:
<table>
<tbody>
<tr>
<td></td>
<td>A→B→D→E→𝐴𝐵→𝐷→𝐸\displaystyle A\rightarrow B\rightarrow D\rightarrow E</td>
<td></td>
<td>(1)</td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td>D→E→𝐷𝐸\displaystyle D\rightarrow E</td>
<td></td>
<td>(2)</td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td>A→B→D→C→𝐴𝐵→𝐷→𝐶\displaystyle A\rightarrow B\rightarrow D\rightarrow C</td>
<td></td>
<td>(3)</td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td>D→C→A→B→𝐷𝐶→𝐴→𝐵\displaystyle D\rightarrow C\rightarrow A\rightarrow B</td>
<td></td>
<td>(4)</td>
</tr>
</tbody>
</table>
As we’ve seen in the previous section [6](#S6 "6 Acyclic Tracer Packet flood ‣ Quantum Shortest Path Netsukuku"), we can consider these routes as effective Tracer Packets. In this example, the TP (1) cover the same route of the TP (2). Therefore we can save one TP by just sending the TP (1), which will traverse the route (2) too.
The TP (3) covers part of the TP (4), thus we can simplify the two of them by just sending a TP which cover this route: $`A\rightarrow B\rightarrow D\rightarrow C\rightarrow A\rightarrow B`$.
Continuing in this process we can further simplify the two TP:
<table>
<tbody>
<tr>
<td></td>
<td>ABDCAB+ABDE⇒ABDCABDE⇒𝐴𝐵𝐷𝐶𝐴𝐵𝐴𝐵𝐷𝐸𝐴𝐵𝐷𝐶𝐴𝐵𝐷𝐸ABDCAB+ABDE\Rightarrow ABDCABDE</td>
<td></td>
</tr>
</tbody>
</table>
Thus, from the initial four TPs we’ve found a unique TP which gives the same routes of the original ones.
### 7.1 Simplification rules
We can derive some rules to simplify routes.
Since we can represent a route as a string where each symbol is a node, we can also describe the routes simplification as a series of operations on strings.
In the following rules, each letter found in an expression represents a generic string, which may be also the NULL string, f.e. the “$`X\hspace{0pt}X`$” string can be anything like $`f\hspace{0pt}o\hspace{0pt}o\hspace{0pt}f\hspace{0pt}o\hspace{0pt}o`$ or $`1234512345`$.
The $`c\hspace{0pt}\ldots\hspace{0pt}c`$ expression represents a cycle, where the $`c`$ character refers to just one node, and not to an entire string.
XY+YZ $`\Rightarrow`$ XYZ
If two routes share respectively the ending and the starting part, they can be merged into a unique route. Example:
<table>
<tbody>
<tr>
<td></td>
<td>ABCDE+CDEKRE⇒ABCDEKRE⇒𝐴𝐵𝐶𝐷𝐸𝐶𝐷𝐸𝐾𝑅𝐸𝐴𝐵𝐶𝐷𝐸𝐾𝑅𝐸ABCDE+CDEKRE\Rightarrow ABCDEKRE</td>
<td></td>
</tr>
</tbody>
</table>
YXZ + X $`\Rightarrow`$ YXZ
Example:
<table>
<tbody>
<tr>
<td></td>
<td>123ABCXYZ+ABC⇒123ABCXYZ⇒123𝐴𝐵𝐶𝑋𝑌𝑍𝐴𝐵𝐶123𝐴𝐵𝐶𝑋𝑌𝑍123ABCXYZ+ABC\Rightarrow 123ABCXYZ</td>
<td></td>
</tr>
</tbody>
</table>
Xc…c + XcY $`\Rightarrow`$ Xc…cY
Example:
<table>
<tbody>
<tr>
<td></td>
<td>123ABCDA+123A987⇒123ABCDA987⇒123𝐴𝐵𝐶𝐷𝐴123𝐴987123𝐴𝐵𝐶𝐷𝐴987123ABCDA+123A987\Rightarrow 123ABCDA987</td>
<td></td>
</tr>
</tbody>
</table>
c…cZ + YcZ $`\Rightarrow`$ Yc…cZ
Example:
<table>
<tbody>
<tr>
<td></td>
<td>ABCDA123+987A123⇒987ABCDA123⇒𝐴𝐵𝐶𝐷𝐴123987𝐴123987𝐴𝐵𝐶𝐷𝐴123ABCDA123+987A123\Rightarrow 987ABCDA123</td>
<td></td>
</tr>
</tbody>
</table>
c…c + YcZ $`\Rightarrow`$ Yc…cZ
Example:
<table>
<tbody>
<tr>
<td></td>
<td>ABCDA+987A123⇒987ABCDA123⇒𝐴𝐵𝐶𝐷𝐴987𝐴123987𝐴𝐵𝐶𝐷𝐴123ABCDA+987A123\Rightarrow 987ABCDA123</td>
<td></td>
</tr>
</tbody>
</table>
Invalid route
A route must not be in the form of:
<table>
<tbody>
<tr>
<td></td>
<td>XacaY𝑋𝑎𝑐𝑎𝑌XacaY</td>
<td></td>
</tr>
</tbody>
</table>
where $`a`$ and $`c`$ are two nodes. A simplification, which gives a route of this form, is not considered valid. This is because a TP must not change its verse while traversing a network.
All these rules can be applied recursively to the routes of a graph, until they cannot be simplified anymore.
A proof of concept of the above algorithm has been implemented in Awk \[[5](#bib.bib5)\].
### Example
Simplifying all the routes of the example [2](#S5.F2 "Figure 2 ‣ Example ‣ 5 Routes of a graph ‣ Quantum Shortest Path Netsukuku"), we obtain just these two TPs:
<table>
<tbody>
<tr>
<td></td>
<td>A→B→D→C→A→B→D→E→𝐴𝐵→𝐷→𝐶→𝐴→𝐵→𝐷→𝐸\displaystyle A\rightarrow B\rightarrow D\rightarrow C\rightarrow A\rightarrow B\rightarrow D\rightarrow E</td>
<td></td>
<td>(5)</td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td>A→C→D→B→A→C→D→E→𝐴𝐶→𝐷→𝐵→𝐴→𝐶→𝐷→𝐸\displaystyle A\rightarrow C\rightarrow D\rightarrow B\rightarrow A\rightarrow C\rightarrow D\rightarrow E</td>
<td></td>
<td>(6)</td>
</tr>
</tbody>
</table>
You can verify that all the routes listed in [2](#S5.F2 "Figure 2 ‣ Example ‣ 5 Routes of a graph ‣ Quantum Shortest Path Netsukuku") are contained in these two simplified TPs.
### 7.2 General results
By looking at many different simplifications, we can recognize some general rules:
1. 1\.
For each TP there has to be its inverse. For example, if there’s a TP which covered the route $`12345`$, then there has to be at least the TP which covers the inverse route $`54321`$.
2. 2\.
In a segment, to give all the routes to all the nodes, it is sufficient that the two extremes sends a TP. Example:
[Refer to caption](/html/0705.0817/assets/x3.png)
Figure 3: A segment
in this case, if $`A`$ and $`F`$ send a TP, all the routes will be generated, since the two TP would be: $`A\hspace{0pt}B\hspace{0pt}C\hspace{0pt}D\hspace{0pt}E\hspace{0pt}F`$ and $`F\hspace{0pt}E\hspace{0pt}D\hspace{0pt}C\hspace{0pt}B\hspace{0pt}A`$. You can verify that in these two TP, there are contained all the routes of the segment.
3. 3\.
In a cycle, just two TP are needed, and one is the reverse of the other. The first can be constructed in this way:
- •
Choose a node of the cycle, this will be the pivot node.
- •
Start from one neighbour of the pivot and write sequencially all the other nodes until you return to the pivot (but do not include it). Call this string $`C`$.
- •
The TP will be:
<table>
<tbody>
<tr>
<td></td>
<td>CpC𝐶𝑝𝐶CpC</td>
<td></td>
</tr>
</tbody>
</table>
where $`p`$ is the pivot node.
Example:
[Refer to caption](/html/0705.0817/assets/x4.png)
Figure 4: A cycle
if we choose the node $`D`$ as the pivot, we can write the TP as:
<table>
<tbody>
<tr>
<td></td>
<td>EFABCDEFABC𝐸𝐹𝐴𝐵𝐶𝐷𝐸𝐹𝐴𝐵𝐶EFABCDEFABC</td>
<td></td>
</tr>
</tbody>
</table>
and its reverse:
<table>
<tbody>
<tr>
<td></td>
<td>CBAFEDCBAFE𝐶𝐵𝐴𝐹𝐸𝐷𝐶𝐵𝐴𝐹𝐸CBAFEDCBAFE</td>
<td></td>
</tr>
</tbody>
</table>
These two TPs will give all the routes to all the nodes of the cycle.
### 7.3 The question
Can we implement a “live” version of the Simplify Route algorithm like we did with the Generate Route one?
The reply is ahead.
## 8 Continuous Tracer Packet
A Continuous Tracer Packet (CTP) is an extension of the TP flood: a node will always forward a TP to all its neighbours, excepting the one from which it has received the TP. If a node is an extreme of a segment, i.e. a node with just one link, it will erase the route stored in the body of the TP and will forward back the TP.
In short, a CTP is a TP flood which will never end, thus it will continue to explore all the infinite combination of routes.
### Example
Consider this graph.
[Refer to caption](/html/0705.0817/assets/x5.png)
Figure 5: The simplest cyclic graph
If $`A`$ sends a CTP flood, there will be two CTPs that will explore respectively these routes:
<table>
<tbody>
<tr>
<td></td>
<td>A→B→C→A→B→C→A→B→C→A→B→C→…→𝐴𝐵→𝐶→𝐴→𝐵→𝐶→𝐴→𝐵→𝐶→𝐴→𝐵→𝐶→…A\rightarrow B\rightarrow C\rightarrow A\rightarrow B\rightarrow C\rightarrow A\rightarrow B\rightarrow C\rightarrow A\rightarrow B\rightarrow C\rightarrow\dots</td>
<td></td>
</tr>
</tbody>
</table>
<table>
<tbody>
<tr>
<td></td>
<td>A→C→B→A→C→B→A→C→B→A→C→B→…→𝐴𝐶→𝐵→𝐴→𝐶→𝐵→𝐴→𝐶→𝐵→𝐴→𝐶→𝐵→…A\rightarrow C\rightarrow B\rightarrow A\rightarrow C\rightarrow B\rightarrow A\rightarrow C\rightarrow B\rightarrow A\rightarrow C\rightarrow B\rightarrow\dots</td>
<td></td>
</tr>
</tbody>
</table>
### 8.1 Reflected CTP
Suppose that the node $`N`$ has just one link. $`N`$, before back forwarding the received CTP, erases the route contained in the body, because the nodes preceding it, already know this same route.
For example, consider this segment:
<table>
<tbody>
<tr>
<td></td>
<td>⋯↔︎A↔︎B↔︎C↔︎N↔︎⋯𝐴↔︎𝐵↔︎𝐶↔︎𝑁\cdots\leftrightarrow A\leftrightarrow B\leftrightarrow C\leftrightarrow N</td>
<td></td>
</tr>
</tbody>
</table>
If $`N`$ hadn’t erased the route received in the CTP, $`A`$ would have received the following CTP:
<table>
<tbody>
<tr>
<td></td>
<td>⋯→A→B→C→N→C→B→A→⋯𝐴→𝐵→𝐶→𝑁→𝐶→𝐵→𝐴\cdots\rightarrow A\rightarrow B\rightarrow C\rightarrow N\rightarrow C\rightarrow B\rightarrow A</td>
<td></td>
</tr>
</tbody>
</table>
This packet contains the route $`C\rightarrow N\rightarrow C`$, which is invalid, as explained in section [](#S7.I1.ix6 "item Invalid route ‣ 7.1 Simplification rules ‣ 7 Routes simplification ‣ Quantum Shortest Path Netsukuku")
Invalid route
. The valid parts of the packet are: $`\cdots\rightarrow A\rightarrow B\rightarrow C\rightarrow N`$ and $`N\rightarrow C\rightarrow B\rightarrow A`$. For this reason, when $`N`$ receives the first part, it will send a new, empty CTP.
## 9 QSPN v2
The second version of the QSPN
<sup>2</sup>
<sup>2</sup>2The short name of the QSPN v2 is $`Q^{2}`$
can be described in a single phrase:
*A Continuous Tracer Packet will continue to roam inside*
*the network until it carries interesting information.*
### 9.1 Interesting information
A node considers a received CTP interesting when its body contains at least a new route, i.e. a route that the node didn’t previously know. In other words, if a CTP contains routes already known by the node, it is considered uninteresting.
When a node receives an interesting CTP, it forwards the packet to all its neighbours, excepting the one from which it has received the CTP. If, instead, the CTP is uninteresting, it will drop the packet.
Note that if a CTP is uninteresting for the node $`N`$, then it is also uninteresting for all the other nodes. This is because an uninteresting CTP contains routes which has been previously received, memorised and forwarded by the node $`N`$. Therefore all the other nodes already know the same routes too.
#### Example
Consider this graph.
[Refer to caption](/html/0705.0817/assets/x6.png)
Figure 6: The A-B-C-A cycle
Suppose that $`A`$ sends a CTP. The two CTPs, after having covered the following two paths will stop:
<table>
<tbody>
<tr>
<td></td>
<td>A→B→C→A→B→C→𝐴𝐵→𝐶→𝐴→𝐵→𝐶A\rightarrow B\rightarrow C\rightarrow A\rightarrow B\rightarrow C</td>
<td></td>
</tr>
</tbody>
</table>
<table>
<tbody>
<tr>
<td></td>
<td>A→C→B→A→C→B→𝐴𝐶→𝐵→𝐴→𝐶→𝐵A\rightarrow C\rightarrow B\rightarrow A\rightarrow C\rightarrow B</td>
<td></td>
</tr>
</tbody>
</table>
Let’s analyze the first CTP step by step, considering that before $`A`$ sent the CTP, none knew any route.
A $`\rightarrow`$ B
At this point $`B`$ doesn’t know any route to reach $`A`$, therefore it considers this CTP as interesting and forwards it to $`C`$.
A $`\rightarrow`$ B $`\rightarrow`$ C
By looking at this packet $`C`$ learns a route to reach $`B`$ and $`A`$.
A $`\rightarrow`$ B $`\rightarrow`$ C $`\rightarrow`$ A
The node $`A`$ learns a route to reach $`C`$.
A $`\rightarrow`$ B $`\rightarrow`$ C $`\rightarrow`$ A $`\rightarrow`$ B
The node $`B`$ learns a route to reach $`C`$.
A $`\rightarrow`$ B $`\rightarrow`$ C $`\rightarrow`$ A $`\rightarrow`$ B $`\rightarrow`$ C
Finally, $`C`$ drops the packet, because it already knows all the routes contained in it.
From this example we can derive a general result: a CTP will always terminate in a cycle.
### 9.2 Live routes simplification
The QSPN v2 is the “live” version of the Simplify Route algorithm (section [7](#S7 "7 Routes simplification ‣ Quantum Shortest Path Netsukuku")).
The CTP flood of the QSPN v2 explores the entire graph, but unlike the ATP (section [6](#S6 "6 Acyclic Tracer Packet flood ‣ Quantum Shortest Path Netsukuku")), it drops the TPs which contains redundant routes, thus only the simplified, non redundant routes survives and continue to explore the graph.
### 9.3 Cyclicity
When a CTP reaches the extremity of a segment, it is back forwarded, thus it’s as if the extreme nodes had a link with themselves.
[Refer to caption](/html/0705.0817/assets/x7.png)
Figure 7: A segment as viewed from a CTP
From the point of view of a CTP, even a segment is a cycle, therefore, for a CTP, any connected graph is formed just by cycles.
For this reason, a CTP will explore any combination of cycles of the graph.
#### 9.3.1 Subcycles examples
These examples highlights some subcycles of a simple graph.
[\[Uncaptioned image\]](/html/0705.0817/assets/x8.png)
A CTP would explore all these cycles.
### 9.4 Finiteness
$`Q^{2}`$ will finish the exploration of the graph in a finite amount of time, i.e. the flood will terminate.
As we’ve seen in the example [9.1](#S9.SS1.SSSx1 "Example ‣ 9.1 Interesting information ‣ 9 QSPN v2 ‣ Quantum Shortest Path Netsukuku"), a CTP flood of a cycle will always terminate. Moreover in section [9.3](#S9.SS3 "9.3 Cyclicity ‣ 9 QSPN v2 ‣ Quantum Shortest Path Netsukuku") we’ve noticed that, from the point of view of a CTP, any connected graph is formed by a combination of cycles. Therefore, a CTP flood of a graph will always terminate in a finite amount of time.
### 9.5 Routes limit
Even if $`Q^{2}`$ is finite, it still generates too many routes and packets. Therefore we need to limit the exploration of the graph.
An efficient and elegant solution is to further define what the “interesting information” is:
Let each node of the network keep a maximum of *MaxRoutes* routes in its memory. A node considers a received CTP interesting when its body contains at least a route which is more efficient than the previously memorised routes. The efficiency of a route can be quantified with a convenient parameter, f.e. the rtt or the bandwidth capacity. If the node has reached the *MaxRoutes* limit, it will substitute the old route with the more efficient one.
Note that this definition is more general than the previous. Indeed, if the node $`S`$ doesn’t know the route to reach $`D`$, the efficiency of the route $`S\rightarrow D`$ is equal to $`0`$.
A node can also keep in memory more than *MaxRoutes* routes, because this limit applies only to the number of routes which will be used to evaluate the received CTP.
### 9.6 Scalability
We will now exploit the bouquet property of the Tracer Packets, which has been described in section [4.2](#S4.SS2 "4.2 Proprieties of the tracer packet ‣ 4 Tracer Packet ‣ Quantum Shortest Path Netsukuku").
Suppose that in our network every link has the same bandwidth capacity and that the generic node $`D`$ doesn’t know any route to reach the node $`T`$. If a TP, received by the node $`D`$, contains a new route $`t`$ that connects $`D\rightarrow T`$, then we can deduce, by the bouquet property, that $`t`$ is the fastest route between $`D`$ and $`T`$.
The immediate consequence is that $`D`$ will receive all the other $`D\rightarrow T`$ routes in order of efficiency: the first is, as we’ve seen, the best route, the second one will be slower than the first but surely better than any other, and so on.
#### 9.6.1 TP Classes
Two different routes can be very similar, because they can differ only in a small part. Two routes which differs of just one hop, are almost identical. For this reason, other than the best $`D\rightarrow T`$ route, the CTP will also explore all the other routes which are almost identical to it.
We can thus order all the TPs which $`D`$ will receive into classes. The first class, denoted with $`\lbrack 1\rbrack`$, contains the TP which have covered the best route and all the others similar to it, i.e. all the other routes with the same number of hop and a similar trtt (total round trip time). The second class $`\lbrack 2\rbrack`$, contains the TP which have covered routes which are less efficient than those contained in $`\lbrack 1\rbrack`$ but are more efficient than those of class $`\lbrack 3\rbrack`$. More generally we can say that the $`\lbrack{n\hspace{0pt}\text{-th}}\rbrack`$ class contains the routes that, if included in a list of all the routes of the graph, ordered in decrescent order of efficiency, will be listed starting from the position $`{{({n - 1})}\hspace{0pt}c} + 1`$. Where $`c`$ is the numbers of routes contained in each class.
In the classes we are including *routes* and not *tracer packets*, because a TP may contain more than one route.
#### 9.6.2 Subcycle filter
Each node of the graph acts as a filter for all the subcycles containing it.
Suppose the node $`D`$ is contained in the subcycle $`\sigma`$ and that a CTP $`t`$ enters in it (through another node). If the node $`D`$ has the *MaxRoutes* limit set, it will memorise, for each node of the network, only the first *MaxRoutes* received routes, while the rest will be disregarded, and not forwarded.
The CTP $`t`$ won’t be forwarded by the node $`D`$, if it contains routes which exceed the *MaxRoutes* limit, but this is true for all the nodes of $`\sigma`$, therefore $`t`$ won’t even be able to escape from the subcycle $`\sigma`$. However, this also means that all the CTPs, which are in a higher class than that of $`t`$, won’t be allowed to pass from $`\sigma`$.
Since this happens for all the subcycles of the graph, we can conclude that at worst, the number of CTPs increases in polynomial time with the increase of subcycles.
#### Example
Consider this graph.
[Refer to caption](/html/0705.0817/assets/x9.png)
Figure 8: Sequential composition of cycles
Let’s analyse the difference between a CTP without limits and the same CTP with *MaxRoutes* set.
Unlimited CTP
$`A`$ sends a CTP. Suppose for simplicity that this CTP won’t loop inside a cycle.
$`D`$ receives $`2`$ packets from $`A`$, $`G`$ $`2 \ast 2`$ from $`D`$ and finally $`L`$ $`2 \ast 2 \ast 2`$ from $`D`$.
$`L`$ sends back all the received CTPs, thus $`G`$ gets $`2^{4}`$ CTPs from $`L`$, $`D`$ gets $`2^{5}`$ and finally $`A`$ receives $`2^{6} = 64`$ packets. Obviously this is too much for this simple graph.
Limited CTP
Let’s use $`= 1`$ and suppose that the CTP won’t loop inside a cycle. We’ll write $`D_{A}\hspace{0pt}{(n,z)}`$ to indicate that $`D`$ received $`n`$ CTPs from the node $`A`$ but has kept and forwarded only $`z`$ packets.
$`A`$ sends a CTP.
<table>
<tbody>
<tr>
<td></td>
<td>A∅(0,2)→DA(2,2)→GD(4,3)→LG(6,4)→→subscript𝐴02subscript𝐷𝐴22→subscript𝐺𝐷43→subscript𝐿𝐺64→absent\displaystyle A_{\emptyset}(0,2)\rightarrow D_{A}(2,2)\rightarrow G_{D}(4,3)\rightarrow L_{G}(6,4)\rightarrow</td>
<td></td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td>GL(8,5)→DG(10,6)→AD(12,0)→subscript𝐺𝐿85subscript𝐷𝐺106→subscript𝐴𝐷120\displaystyle\qquad G_{L}(8,5)\rightarrow D_{G}(10,6)\rightarrow A_{D}(12,0)</td>
<td></td>
</tr>
</tbody>
</table>
At the end, A gets $`12`$ packets.
Each node forwards $`{p/2} + 1`$ packets, where $`p`$ is the number of received CTPs. This is because the first two packets give to the node a new route, while the other two, and the successive ones, cover a superflous route. For example, consider
<table>
<tbody>
<tr>
<td></td>
<td>A→C→D→E→G¯¯→𝐴𝐶→𝐷→𝐸→𝐺\displaystyle\underline{A\rightarrow C\rightarrow D\rightarrow E\rightarrow G}</td>
<td></td>
<td>(7)</td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td>A→C→D→F→G¯→𝐴𝐶→𝐷→¯→𝐹𝐺\displaystyle A\rightarrow C\rightarrow D\rightarrow\underline{F\rightarrow G}</td>
<td></td>
<td>(8)</td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td>A→B→D¯→E→G→𝐴¯→𝐵𝐷→𝐸→𝐺\displaystyle A\rightarrow\underline{B\rightarrow D}\rightarrow E\rightarrow G</td>
<td></td>
<td>(9)</td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td>A→B→D→F→G!→𝐴𝐵→𝐷→𝐹→𝐺\displaystyle A\rightarrow B\rightarrow D\rightarrow F\rightarrow G\mathbb{!}</td>
<td></td>
<td>(10)</td>
</tr>
</tbody>
</table>
The underlined routes are the new route for $`G`$. As you can see, in the CTP (10) G doesn’t find any new route, so it drops the packet and doesn’t forward it.
#### 9.6.3 Efficiency order
We’ve noted in section [9.6](#S9.SS6 "9.6 Scalability ‣ 9 QSPN v2 ‣ Quantum Shortest Path Netsukuku") that any node will receive the CTPs in order of efficiency, thus we are sure that only the first *MaxRoutes* received routes (which point to a specific node) are meaningfull, while all the successive ones are uninteresting and should be dropped. This is important, because each node will first receive *MaxRoutes* interesting routes and then the uninteresting ones, that will be dropped. Doing so, an uninteresting route won’t be forwarded before its interesting correspondent, and as soon it is recognised it will be dropped.
To clarify this concept suppose that the routes aren’t received in order of efficiency and that $`= 2`$. Then suppose that node $`D`$ doesn’t have any route yet and receives a CTP of class $`\lbrack 7\rbrack`$. $`D`$ will consider this CTP as interesting and forward it, because it’s the first one it receives.
In conclusion, the CTP of class $`\lbrack 7\rbrack`$ would be allowed to be propagated among the network even if the *MaxRoutes* has been set to $`2`$ routes. Instead, if $`D`$ receives the packet in an ordered manner, it will first get the CTPs of class $`\lbrack 1\rbrack`$, then those of class $`\lbrack 2\rbrack`$ and so on. For this reason the CTP of class $`\lbrack 7\rbrack`$. and all the other which exceed the *MaxRoutes* limit, won’t be allowed to pass from $`D`$.
### 9.7 Bandwidth issues
Until now we’ve supposed that every link of the network has the same bandwidth capacity. However, we’ll see in section [12.1](#S12.SS1 "12.1 Rtt and bandwidth ‣ 12 QSPN optimisations ‣ Quantum Shortest Path Netsukuku") that the QSPN can be used in the general case.
### 9.8 Worst case
The graph, formed by $`n`$ nodes, which has the maximum number of cycles is the worst case for $`Q^{2}`$, because, if not limited, it will have to explore any combination of cycles.
Such graph is the *complete graph* \[[7](#bib.bib7)\], and the total number of its subcycles is:
<table>
<tbody>
<tr>
<td></td>
<td>∑k=3n12(nk)(k−1)!superscriptsubscript𝑘3𝑛12binomial𝑛𝑘𝑘1\sum_{k=3}^{n}\,\frac{1}{2}\binom{n}{k}(k-1)!</td>
<td></td>
</tr>
</tbody>
</table>
## 10 QSPN v1
It isn’t necessary to read this paragraph in order to understand the rest of the paper. If you aren’t interesent in the QSPN v1, just skip over.
The QSPN v1 is a restricted case of $`Q^{2}`$. It is divided in two phases. The first one is called qspn_close: a node sends a QTP (QSPN Tracer Packet) called qclose, this node becomes a qspn starter. A qclose is a modified form of tracer packet. A node $`N`$, which receives a qclose from the link $`l`$, marks as “closed” the same link $`l`$ and forwards the packet to all its other neighbours. All the following qclose packets received by the same node $`N`$, will be forwarded only to the links which have not been already closed.
During the graph exploration, some nodes will close all their links. These nodes are called *extreme nodes*. When a node becomes an extreme node, it will send another type of tracer packet, called qspn_open (which is also the name of the second phase) to all its neighbours, except the one from which it received the last qclose packet (let’s call this neighbour $`L`$). The qopen packet sent to $`L`$ is empty, while those sent to the other neighbours contains the body of the last received qclose packet.
The qopen behaves as the qclose: it ”opens” the links, however the nodes which have all their links opened won’t forward any other packets.
Example:
[Refer to caption](/html/0705.0817/assets/x10.png)
Figure 9: Example of a QSPN v1 graph exploration
Consider figure [9](#S10.F9 "Figure 9 ‣ 10 QSPN v1 ‣ Quantum Shortest Path Netsukuku").
- •
The node $`E`$ sends a qclose. It is now a qspn starter.
- •
Suppose that the node A receives the qclose before $`C`$. $`A`$ closes the link $`E\rightarrow A`$ and forwards the qclose to $`B`$, $`C`$ and $`D`$.
- •
$`C`$ receives the qclose from $`E`$, closes the link $`E\rightarrow C`$ and forwards it to A and $`D`$.
- •
$`C`$ receives the qclose from $`A`$ and closes the link.
- •
$`B`$ and $`D`$ have received from $`A`$ the qclose and close the respective links.
- •
Suppose that $`B`$ is the first to forwards the qclose to $`F`$.
- •
$`D`$ forwards the qclose to $`F`$, but at the same time $`F`$ forwards it to $`D`$.
- •
$`D`$ receives the qclose from $`B`$, too.
- •
$`D`$ and $`F`$ have all the links closed. They send a qopen.
- •
The qopen propagates itself in the opposite sense.
- •
The qopen ends. Each node has the routes to reach all the other nodes.
The qspn_close phase can be seen as a CTP with the added rule that when two CTPs collide, they will be converted to two normal TPs (the qopen phase).
### 10.1 $`\mathbf{Q}`$ vs $`\mathbf{Q}^{\mathbf{2}}`$
These are the substantial differences between Q (QSPN v1) and $`Q^{2}`$ (QSPN v2):
1. 1\.
Q generates less packets than $`Q^{2}`$, because in the qopen phase it uses normal TPs which expires quickly. The side effect of this behaviour is that Q may not discover all the best routes. However Q gives at least one route to reach each node of the graph.
2. 2\.
Q uses less memory than $`Q^{2}`$, because it just keeps a forwarding table, instead, $`Q^{2}`$ needs to memorize MaxRoutes complete routes to evaluate the successive CTPs. By the way, this difference is minimal.
3. 3\.
$`Q^{2}`$ doesn’t need synchronization. the CTPs doesn’t need to have an ID, thus many nodes can send simultaneously or asynchronously a CTP without creating any problem.
This isn’t the same in Q, which requires a strict synchronization between the nodes: two nodes can send a qclose only at the same time.
4. 4\.
This is a consequence of the propriety described above: every time a node joins the net or dies or its rtt/bw capacity changes, it is possible to immediately send a CTP. Indeed, if the changes in the local gnode regard that node only, the CTP will be like a normal Tracer Packet (see <a href="#S4.SS1" style="font-size:90%;" title="4.1 Tracer Packet flood ‣ 4 Tracer Packet ‣ Quantum Shortest Path Netsukuku">4.1</a>).
5. 5\.
$`Q^{2}`$ is easier and simpler than Q to be implemented. In general this means that the code of $`Q^{2}`$ will have less bugs.
From this comparison we can conclude that $`Q^{2}`$ is preferable over Q.
## 11 Network dynamics
The QSPN v2 defined until now is not suitable for dynamic networks. As example, consider this problem:
suppose that the whole graph has been already explored, and thus every node has at least one route to reach all the other nodes. Consider the case when the efficiency of a link, f.e. $`N\leftrightarrow P`$, worsens. $`N`$, in order to update the maps of the other nodes, sends a CTP to its neighbour $`P`$, and $`P`$ forwards it to its neighbours. However this CTP will be immediately dropped! Indeed, the nodes will consider this CTP not interesting, because the contained $`\ldots\rightarrow P\rightarrow N`$ route is less efficient than the old one, which has been saved during the last graph exploration.
### 11.1 Extended Tracer Packet
The ETP solves the problem of how the graph should be re-explored to update the maps of the nodes interested to a network change. Its way of working is based on a simple observation:
The first QSPN exploration distributes, among the nodes, information describing the network topology. When a change in the network occurs, only the information stored in the nodes affected by the change must be updated. The unaffected nodes will still have up to date information that they can simply redistribute with the use of the Extended Tracer Packets.
An ETP is an Acyclic Tracer Packet,
<sup>3</sup>
<sup>3</sup>3 An ATP (see paragraph [6](#S6 "6 Acyclic Tracer Packet flood ‣ Quantum Shortest Path Netsukuku")) is a normal TP with the following rule: a node drops the received ATP if its node ID is already present in the route contained in the body of the packet. Note also, that since it is a normal TP, it is not reflected back, when it reaches the end of a segment.
which contains a portion of a map. Since a map is a set of routes and a route can be described by a TP, the ETP can be considered as different TPs packed togheter. Each TP of the ETP is then subjected to the rules of the QSPN v2.
In order to give an exact definition of an ETP, we must examine each case of network change.
Worsened link
Suppose that the link $`A\overset{l}{\leftrightarrow}B`$ worsens.
Let’s analyse what $`B`$ will do (the situation is symmetric for $`A`$).
$`B`$, if interested in the change, will create an ETP containing all its old routes, affected by the change, and all the backup routes used as substitute for the old ones. The ETP will be sent to all its neighbours, except $`A`$. In detail, $`B`$ will use the following algorithm:
1. 1\.
If at least one of the primary routes
<sup>4</sup>
<sup>4</sup>4 a route src$`\rightarrow{d\hspace{0pt}s\hspace{0pt}t}`$ is called primary if it is among the first MaxRoutes routes of type src$`\rightarrow{d\hspace{0pt}s\hspace{0pt}t}`$
saved in the map of $`B`$ and different from the route $`B\rightarrow A`$, uses the link $`l`$, then $`B`$ creates an empty ETP, otherwise the algorithm halts, i.e. $`B`$ won’t do anything.
2. 2\.
If the empty ETP has been created, $`B`$ updates its maps: suppose that the route $`r`$, passing through the link $`l`$, had a total rtt $`t_{0}\hspace{0pt}(r)`$. If the rtt of the link $`l`$ before the change was $`t_{0}\hspace{0pt}(l)`$ and now is $`t_{1}\hspace{0pt}(l)`$, where $`{t_{0}\hspace{0pt}(l)} < {t_{1}\hspace{0pt}(l)}`$, then $`{t_{1}\hspace{0pt}(r)}:={{{t_{0}\hspace{0pt}(r)} - {t_{0}\hspace{0pt}(l)}} + {t_{1}\hspace{0pt}(l)}}`$.
For the bandwidth we have:
<table>
<tbody>
<tr>
<td></td>
<td>b0(r) the total bandwidth of the route r, before the changesubscript𝑏0𝑟 the total bandwidth of the route r, before the change\displaystyle b_{0}(r)\;\;\textrm{ the total bandwidth of the route $r$, before the change}</td>
<td></td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td>b0(l)>b1(l)the bw of l has worsend during the changesubscript𝑏0𝑙subscript𝑏1𝑙the bw of l has worsend during the change\displaystyle b_{0}(l)>b_{1}(l)\;\;\textrm{the bw of $l$ has worsend during the change}</td>
<td></td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td>b1(r):=min{b0(r),b1(l)}assignsubscript𝑏1𝑟subscript𝑏0𝑟subscript𝑏1𝑙\displaystyle b_{1}(r):=\min\{b_{0}(r),b_{1}(l)\}</td>
<td></td>
</tr>
</tbody>
</table>
The routes are then sorted.
3. 3\.
$`B`$ creates the temporary set $`Q`$, containing all the primary routes passing through the link $`l`$. From $`Q`$ it creates the set $`R`$, where
<table>
<tbody>
<tr>
<td></td>
<td>R={r∈M|∃q∈Q:dst(r)=dst(q)}𝑅conditional-set𝑟𝑀:𝑞𝑄dst𝑟dst𝑞R=\{r\in M\;|\;\exists q\in Q:\;\textrm{dst}(r)=\textrm{dst}(q)\}</td>
<td></td>
</tr>
</tbody>
</table>
where $`M`$ is the set of all the primary routes of the map, and $`\text{dst}\hspace{0pt}(r)`$ is the destination of the route $`r`$. In other words, $`R`$ is the set of primary routes having the same destination of at least one route of $`Q`$. Note that $`Q \subseteq R`$. Each route $`r \in R`$ is saved as $`\left( {\text{dst}\hspace{0pt}(r)},{\text{rem}\hspace{0pt}(r)},{\text{tpmask}\hspace{0pt}(r)} \right)`$, where $`\text{rem}\hspace{0pt}(r)`$ is the Route Efficiency Measure, and $`\text{tpmask}\hspace{0pt}(r)`$ is a bitmask of 256 bits, where the bit at the i-th position indicates if the node i is an hop of the route $`r`$.
4. 4\.
$`B`$ fills the ETP:
1. \(a\)
it adds in it the set $`R`$
2. \(b\)
it appends the ID of $`A`$, along with the efficiency value of the link $`l`$, and, as usual, its ID.
3. \(c\)
it sets to 1 the *flag of interest*.
5. 5\.
Finally, $`B`$ sends the ETP to all its neighbours, except $`A`$.
Suppose that the neighbor $`C`$ of $`B`$ has received the ETP. $`C`$ will examine the ETP and, if considered interesting, it will update its map and forward the ETP to the other neighbours, as follow:
1. 1\.
If the ID of the node $`C`$ is already present in the received ETP, then $`C`$ immediately drops the ETP, and skips all the following steps
<sup>5</sup>
<sup>5</sup>5This is the acyclic rule
.
2. 2\.
Let $`R`$ be the set of routes contained in the ETP received by $`C`$.
Let $`M`$ be the set of all primary routes contained in the map of $`C`$.
For each route $`r \in R`$, the node $`C`$ looks for a route $`m \in M`$ such that
<table>
<tbody>
<tr>
<td></td>
<td>dst(m)=dst(r),tpmask(m)=tpmask(r)formulae-sequencedst𝑚dst𝑟tpmask𝑚tpmask𝑟\textrm{dst}(m)=\textrm{dst}(r),\;\;\textrm{tpmask}(m)=\textrm{tpmask}(r)</td>
<td></td>
</tr>
</tbody>
</table>
If $`m`$ exists, then $`C`$ sets $`{\text{rem}\hspace{0pt}(m)}:={\text{rem}\hspace{0pt}(r)}`$
<sup>6</sup>
<sup>6</sup>6with this operation we are actually replacing $`m`$ with $`r`$, in the map $`M`$
. Otherwise, $`r`$ is copied in the temporary set $`R'`$.
$`M`$ is sorted, i.e. the routes of the map of $`C`$ are sorted in order of efficiency.
3. 3\.
For all $`r' \in R'`$,
1. \(a\)
if $`r'`$ is a better alternative to at least one primary route $`m' \in M`$ such that $`{\text{dst}\hspace{0pt}\left( r' \right)} = {\text{dst}\hspace{0pt}\left( m' \right)}`$, then $`r'`$ is saved in the map of $`C`$ (note
<sup>7</sup>
<sup>7</sup>7When saving a route $`r`$ of the ETP in the map, we must consider the hops covered by the ETP, so the real saved route is $`r\leftarrow\text{hop}_{1}\leftarrow\text{hop}_{2}\leftarrow\ldots\leftarrow\text{hop}_{n}`$. In this case we’ll have $`r\leftarrow B\leftarrow A`$.
),
2. \(b\)
otherwise, $`r'`$ is removed from $`R`$.
Note
<sup>8</sup>
<sup>8</sup>8this step implements the QSPN v2 rules: only good routes are kept, the other are discarded. Notice the extension: if the ETP had only one route, it would be almost equal to a CTP (the CTP doesn’t have the acyclic rule)
4. 4\.
If $`R`$ is now empty, i.e. all its routes have been removed, then $`C`$ considers the ETP as uninteresting. Let’s suppose for now that it is interesting: the *flag of interest* remains set to 1.
5. 5\.
$`C`$ packs the ETP with the previously modified set $`R`$, and adds its ID. The ETP is sent to all its neighbours, except $`B`$
The rnodes of $`C`$ will use this same procedure. In this way, the ETP will continue to be propagated until it is considered interesting.
Let’s suppose now that a node $`N`$ receives the ETP and considers it uninteresting. $`N`$ won’t just drop the ETP, but will also send back another ETP containing its own routes. The reason is simple: $`N`$ considers the received ETP uninteresting, this means that $`N`$ isn’t affected by the change of the link $`l`$, i.e. all its primary routes don’t pass through $`l`$ and thus are still optimal. Therefore, $`N`$ will send back its routes, hoping that they will be useful to the nodes affected by the change. In detail, this is what will happen when $`N`$ receives the ETP:
1. 1\.
$`N`$ receives the ETP from the node $`L`$, and considers it uninteresting.
2. 2\.
Let $`R`$ be the set of routes contained in the ETP.
Let $`M`$ be the set of all primary routes contained in the map of $`N`$.
$`N`$ creates the following set:
<table>
<tbody>
<tr>
<td></td>
<td>S={m∈M|∃r∈R:dst(m)=dst(r)}𝑆conditional-set𝑚𝑀:𝑟𝑅dst𝑚dst𝑟S=\{m\in M\;|\;\exists r\in R:\;\textrm{dst}(m)=\textrm{dst}(r)\}</td>
<td></td>
</tr>
</tbody>
</table>
3. 3\.
$`N`$ creates the new ETP, appending in it the set $`S`$ and its ID. The *flag of interest* of this ETP is set to 0.
4. 4\.
The ETP is sent to $`L`$.
At this point, the new ETP created by $`N`$, will propagate back in the same fashion of the previous ETP (see page <a href="#S11.I1.ix1" style="font-size:90%;" title="item Worsened link ‣ 11.1 Extended Tracer Packet ‣ 11 Network dynamics ‣ Quantum Shortest Path Netsukuku"></a>
Worsened link
), i.e. until considered interesting. The only difference is that when a node considers it uninteresting, it is just dropped
<sup>9</sup>
<sup>9</sup>9the node will drop the ETP if the *flag of interest* is set to 0 and if it is uninteresting
.
A node dies
Suppose the node $`A`$ dies. Each neighbour $`B`$ of $`A`$ will send an ETP. The ETPs are generated and propagated with the algorithms described in the worsened link case (page <a href="#S11.I1.ix1" style="font-size:90%;" title="item Worsened link ‣ 11.1 Extended Tracer Packet ‣ 11 Network dynamics ‣ Quantum Shortest Path Netsukuku"></a>
Worsened link
), the differences are:
1. 1\.
Instead of considering the routes passing through the worsened link we consider the routes passing from the dead node.
2. 2\.
Suppose that the node $`N`$ receives the ETP from its neighbour $`L`$ and considers it uninteresting. $`N`$ sends back to $`L`$ the new ETP to share its routes among the interested node. However, unlike the case for the worsened link, $`N`$ creates also a new simple TP, where it writes only the information of the death of $`A`$. $`N`$ sends this TP to all its neighbours, except $`L`$.
This simple TP will be propagated with the rules of the QSPN v2. For this reason and since it carries only one useful information (the death of $`A`$), each node will receive it just once (the second time it will be dropped). This simple TP serves to inform the nodes, unaffected by the network change, of the death of $`A`$.
Improved link
Suppose that the link $`A\overset{l}{\leftrightarrow}B`$ improves.
Let’s examine the events, starting from $`A`$, keeping in mind that the situation is symmetric for $`B`$.
Since the link $`l`$ improved, it may be possible for $`B`$ to use it to improve some of its routes. For this reason, $`A`$ will send to it an ETP with all the routes of its map, except those of the form $`A\rightarrow B\rightarrow\rightarrow\ldots`$. If $`B`$ finds something of interest, it will forwards the ETP. In detail:
1. 1\.
Let $`M`$ be the set of all primary routes contained in the map of $`A`$.
$`A`$ creates the following set:
<table>
<tbody>
<tr>
<td></td>
<td>R={m∈M|gw(m)≠B}𝑅conditional-set𝑚𝑀gw𝑚𝐵R=\{m\in M\;|\;\textrm{gw}(m)\neq B\}</td>
<td></td>
</tr>
</tbody>
</table>
where $`\text{gw}\hspace{0pt}(m)`$ is the first hop of the route $`m`$. Each route $`r \in R`$ is saved as $`\left( {\text{dst}\hspace{0pt}(r)},{\text{rem}\hspace{0pt}(r)},{\text{tpmask}\hspace{0pt}(r)} \right)`$.
2. 2\.
$`A`$ creates the ETP:
1. \(a\)
it writes in it the set $`R`$
2. \(b\)
it appends the its node ID, along with the efficiency value of the link $`l`$.
3. \(c\)
it sets to 1 the *flag of interest*.
3. 3\.
It sends the ETP to $`B`$
At this point, the ETP is propagated exactly in the same way of the worsened link case (see page <a href="#S11.I1.ix1" style="font-size:90%;" title="item Worsened link ‣ 11.1 Extended Tracer Packet ‣ 11 Network dynamics ‣ Quantum Shortest Path Netsukuku"></a>
Worsened link
).
A new node joins
Suppose the node $`A`$ is joining the network. Its neighbours are $`B_{1},B_{2},\ldots,B_{n}`$, which are all already hooked, i.e. they aren’t joining. Then,
1. 1\.
Each neighbour $`B_{i}`$ sends its whole map to $`A`$
2. 2\.
$`A`$ waits until the maps of all its neighbours are received.
3. 3\.
The maps are “merged” into a single map, which becomes the map of $`A`$. In simple words, the merge of two maps result in a map having only the best routes of the two.
4. 4\.
If the neighbours of $`A`$ are more than one, i.e. $`n > 1`$, then $`A`$ sends, to each of them, an ETP containing all the primary routes of its map.
5. 5\.
The ETPs are propagated exactly in the same way of the worsened link case (see page <a href="#S11.I1.ix1" style="font-size:90%;" title="item Worsened link ‣ 11.1 Extended Tracer Packet ‣ 11 Network dynamics ‣ Quantum Shortest Path Netsukuku"></a>
Worsened link
).
Note that this case extends the hooking procedure.
Broken link
The case where the link $`A\overset{l}{\leftrightarrow}B`$ becomes invalid, is handled in the same way of the worsened link case (see page <a href="#S11.I1.ix1" style="font-size:90%;" title="item Worsened link ‣ 11.1 Extended Tracer Packet ‣ 11 Network dynamics ‣ Quantum Shortest Path Netsukuku"></a>
Worsened link
), because we can consider $`l`$ as infinitely worsened.
New link
The case where a new link $`A\overset{l}{\leftrightarrow}B`$ is established between, is handled in the same way of the improved link case (see page <a href="#S11.I1.ix3" style="font-size:90%;" title="item Improved link ‣ 11.1 Extended Tracer Packet ‣ 11 Network dynamics ‣ Quantum Shortest Path Netsukuku"></a>
Improved link
), because we can consider $`l`$ as infinitely improved.
## 12 QSPN optimisations
### 12.1 Rtt and bandwidth
The bandwith capacity of a route can be used as a parameter of its efficiency. In this section we’ll analyse the implications for the QSPN. For more information about the bandwidth management in Netsukuku you can read the NTK_RFC 002 \[[9](#bib.bib9)\].
#### 12.1.1 Rtt delay
Each node of the network will delay the forwarding of a received CTP by a time inversely proportional to its upload bandwidth. In this way the CTPs will continue to be received in order of efficiency (see section <a href="#S9.SS6.SSS3" style="font-size:90%;" title="9.6.3 Efficiency order ‣ 9.6 Scalability ‣ 9 QSPN v2 ‣ Quantum Shortest Path Netsukuku">9.6.3</a>). The side effect of this rule is that the extreme cases will be ignored, i.e. a route with a very low rtt but with a very poor bw, or a route with an optimal bw but with a very high rtt. However, in the “real world” these extreme cases are rare, because the rtt and the bw are often related.
#### 12.1.2 Asymmetry in $`Q^{2}`$
The QSPN v2 is a very flexible algorithm that can be adapted to a large range of cases. Indeed, with a minimal added overhead, it is possible to achieve asymmetric routing discovery, i.e. a discovery that discerns the upload bandwidth of a route from its download one.
We call this extension the *asymmetric QSPN v2*, while we refer to the old one as *symmetric $`Q^{2}`$*.
1. 1\.
First of all, it is necessary to define further the “interesting information”. A CTP will be considered interesting, not only when it contains interesting (see <a href="#S9.SS1" style="font-size:90%;" title="9.1 Interesting information ‣ 9 QSPN v2 ‣ Quantum Shortest Path Netsukuku">9.1</a>) download routes, but also upload ones. In other words, we consider the upload sense of a route too.
For example, suppose that the node $`A`$ received the CTP $`A\hspace{0pt}B\hspace{0pt}C\hspace{0pt}D\hspace{0pt}A\hspace{0pt}E\hspace{0pt}R\hspace{0pt}T\hspace{0pt}A`$. In this case $`A`$ will know two dinstinct upload routes: $`A\rightarrow B\rightarrow C\rightarrow D`$ and $`A\rightarrow E\rightarrow R\rightarrow T`$. Instead, in the classic CTP, $`A`$ would have known only $`A\rightarrow T\rightarrow R\rightarrow E`$ and $`A\rightarrow D\rightarrow C\rightarrow B`$.
2. 2\.
Secondly, since we are considering the reverse (upload) routes too, we have to remove the restriction imposed on the CTP, which has been described in section <a href="#S8.SS1" style="font-size:90%;" title="8.1 Reflected CTP ‣ 8 Continuous Tracer Packet ‣ Quantum Shortest Path Netsukuku">8.1</a>. The body of the CTP reflected from the extreme of a segment won’t be erased, thus it will contain the old routes too. This is because the old routes can contain interesting information about upload routes. For example, consider this segment:
<table>
<tbody>
<tr>
<td></td>
<td>⋯↔︎A↔︎B↔︎C↔︎N↔︎⋯𝐴↔︎𝐵↔︎𝐶↔︎𝑁\cdots\leftrightarrow A\leftrightarrow B\leftrightarrow C\leftrightarrow N</td>
<td></td>
</tr>
</tbody>
</table>
If $`N`$ doesn’t erase the route received in the CTP, $`A`$ will receive the following CTP:
<table>
<tbody>
<tr>
<td></td>
<td>⋯→A→B→C→N→C→B→A→⋯𝐴→𝐵→𝐶→𝑁→𝐶→𝐵→𝐴\cdots\rightarrow A\rightarrow B\rightarrow C\rightarrow N\rightarrow C\rightarrow B\rightarrow A</td>
<td></td>
</tr>
</tbody>
</table>
In this case $`A`$ will know the following upload route:
<table>
<tbody>
<tr>
<td></td>
<td>A→B→C→N→𝐴𝐵→𝐶→𝑁A\rightarrow B\rightarrow C\rightarrow N</td>
<td></td>
</tr>
</tbody>
</table>
When parsing a CTP, a node will recognize the part of the routes which are in the form of $`X\hspace{0pt}a\hspace{0pt}c\hspace{0pt}a\hspace{0pt}Y`$, where $`a`$, $`c`$ are two nodes and $`X`$ and $`Y`$ are two generic routes. The packet will then be split in $`X\hspace{0pt}a\hspace{0pt}c`$ and $`c\hspace{0pt}a\hspace{0pt}Y`$.
At this point we’ve finished. In fact, we are sure to receive at least one upload route per node because a CTP traverses each path first in one sense and then in the opposite. The CTP information filter, will allow us to receive only the best routes. However, since the Rtt Delay (<a href="#S12.SS1.SSS1" style="font-size:90%;" title="12.1.1 Rtt delay ‣ 12.1 Rtt and bandwidth ‣ 12 QSPN optimisations ‣ Quantum Shortest Path Netsukuku">12.1.1</a>) is tuned for download routes only, it is possible that some upload paths will be ignored.
It is interesting to note that in the majority of cases, the number of CTPs will remain equal to that of the symmetric $`Q^{2}`$.
### 12.2 Disjoint routes
The routing table of each node should be differentiated, i.e. it should not contain redundant routes.
For example, consider these $`S\rightarrow D`$ routes:
<table>
<tbody>
<tr>
<td></td>
<td>SBCFG1G2G3G4G5G6G7…G19D𝑆𝐵𝐶𝐹subscript𝐺1subscript𝐺2subscript𝐺3subscript𝐺4subscript𝐺5subscript𝐺6subscript𝐺7…subscript𝐺19𝐷\displaystyle SBCFG_{1}G_{2}G_{3}G_{4}G_{5}G_{6}G_{7}\dots G_{19}D</td>
<td></td>
<td>(11)</td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td>SRTEG1G2G3G4G5G6G7…G19D𝑆𝑅𝑇𝐸subscript𝐺1subscript𝐺2subscript𝐺3subscript𝐺4subscript𝐺5subscript𝐺6subscript𝐺7…subscript𝐺19𝐷\displaystyle SRTEG_{1}G_{2}G_{3}G_{4}G_{5}G_{6}G_{7}\dots G_{19}D</td>
<td></td>
<td>(12)</td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td>SZXMNO1O2O3O4O5D𝑆𝑍𝑋𝑀𝑁subscript𝑂1subscript𝑂2subscript𝑂3subscript𝑂4subscript𝑂5𝐷\displaystyle SZXMNO_{1}O_{2}O_{3}O_{4}O_{5}D</td>
<td></td>
<td>(13)</td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td>SQPVY1Y2Y3Y4D𝑆𝑄𝑃𝑉subscript𝑌1subscript𝑌2subscript𝑌3subscript𝑌4𝐷\displaystyle SQPVY_{1}Y_{2}Y_{3}Y_{4}D</td>
<td></td>
<td>(14)</td>
</tr>
</tbody>
</table>
The first two are almost identical, indeed they differ only in the first three hops. The last two are, instead, totally different from all the others.
Since the first two routes are redundant, the node $`S`$ should keep in memory only one of them, saving up space for the others non-redundant routes.
Keeping redundant routes in the routing table isn’t optimal, because if one of the routes fails, then there’s a high probability that all the other redundant routes will fail too. Moreover when implementing the multipath routing to load balance the traffic there won’t be any significative improvements.
$`Q^{2}`$ itself should avoid to spread redundant routes. In order to achieve this result, we refine the efficiency value associated to a route. Suppose we want to affect the efficiency value $`R_{e}`$ assigned to the route $`R`$:
1. 1\.
let $`0 \leq {s\hspace{0pt}(R,S)} \leq 1`$ be the similarity level of the route $`R`$ with $`S`$.
2. 2\.
for each memorised route $`S`$ we compute $`s\hspace{0pt}(R,S)`$ and if we find a route $`S`$ such that $`{s\hspace{0pt}(R,S)} > 0.5`$ we go to step 3.
3. 3\.
we set
<table>
<tbody>
<tr>
<td></td>
<td>Re=Re1−s(R,S)ksubscript𝑅𝑒subscript𝑅𝑒1𝑠𝑅𝑆𝑘R_{e}=R_{e}\frac{1-s(R,S)}{k}</td>
<td></td>
</tr>
</tbody>
</table>
where $`k`$ is an appropriate coefficient.
As explained in section <a href="#S9.SS5" style="font-size:90%;" title="9.5 Routes limit ‣ 9 QSPN v2 ‣ Quantum Shortest Path Netsukuku">9.5</a> the efficiency of a route is used as a parameter to evaluate its interest, therefore the more a route is similar to a memorised route the more its efficiency will decrease. Hence it will be considered less interesting.
Note that this is a generalization of concept of interesting route defined in section <a href="#S9.SS5" style="font-size:90%;" title="9.5 Routes limit ‣ 9 QSPN v2 ‣ Quantum Shortest Path Netsukuku">9.5</a>, in fact, when $`R`$ and $`S`$ are equal, we’ll have
<table>
<tbody>
<tr>
<td></td>
<td>s(R,S)=1𝑠𝑅𝑆1s(R,S)=1</td>
<td></td>
</tr>
</tbody>
</table>
so the $`R_{e}`$ value will be equal to $`0`$.
### 12.3 Cryptographic QSPN
A node could easily forge a TP, injecting in the network false routes and links information. The attack would just create a temporary local damage, thanks to the distributed nature of the QSPN. However the optimal solution is to prevent these attacks.
A node, whenever joins Netsukuku, generates a new RSA key pair. The continuous generation of keys prevents the leakage of the node’s anonymity. The node shares its public key to all the other nodes of its group node.
Each entry appended in a TP is then signed with its private key, doing so the other nodes will be able to prove its authenticity and validate the path covered by the TP.
The size required to store the signatures in the TP can be kept costant using the aggregate signature system \[[10](#bib.bib10)\] \[[11](#bib.bib11)\].
## 13 Simulating the QSPN v2
As a proof of concept, we’ve a written the q2sim.py\[[6](#bib.bib6)\], a simulator implementing the core of the QSPN v2.
The q2sim is an event-oriented Discrete Event Simulator. Each (event,time) pair is pushed in a priority queue. The main loop of the program retrieves from the queue the event having the lowest ‘time’ value. This ”popped” event is executed. In this case, the events are the packets sent on the network.
In this paragraph we’ll analyze the results of various simulations.
TP flux
The *TP flux* of a node $`n`$, is the number of distinct
<sup>10</sup>
<sup>10</sup>10To clarify what we mean with “distinct TPs”: suppose $`n`$ has $`6`$ neighbours, it receives only two TPs and sends only a new one, then $`{\Phi\hspace{0pt}{(n)}} \neq {5 + 5 + 6}`$, but $`{\Phi\hspace{0pt}{(n)}} = 3`$.
TP packets which have been forwarded by $`n`$ during the entire QSPN exploration. it is indicated with $`\Phi\hspace{0pt}(n)`$.
Mean TP flux
Given $`k`$ nodes $`n_{1},\ldots,n_{k}`$, their mean TP flux is:
<table>
<tbody>
<tr>
<td></td>
<td>Φm(n1,…,nk)=∑i=1kΦ(ni)ksubscriptΦ𝑚subscript𝑛1…subscript𝑛𝑘superscriptsubscript𝑖1𝑘Φsubscript𝑛𝑖𝑘\Phi_{m}({n_{1},\dots,n_{k}})=\frac{\sum_{i=1}^{k}\Phi(n_{i})}{k}</td>
<td></td>
</tr>
</tbody>
</table>
Starter node
A starter node is a node which sends the first TP in a graph, not yet explored by the $`Q^{2}`$. There can be more than one simultaneous starter node.
In all the following tests the *MaxRoutes* limit has been set to 1.
As a general result, the mean TP flux is proportional to the number of subcycles present in the graph. Thus the maximum mean TP flux is reached in a complete graph, where each node is connected to all the others. In a complete graph of $`k`$ nodes the mean TP flux is approximately:
<table>
<tbody>
<tr>
<td></td>
<td>Φm≈k−1,Φm≤kformulae-sequencesubscriptΦ𝑚𝑘1subscriptΦ𝑚𝑘\Phi_{m}\approx k-1,\;\;\Phi_{m}\leq k</td>
<td></td>
</tr>
</tbody>
</table>
The figure below shows this result. In the $`x`$ axis, an integer point corresponds to the number of nodes of the complete graph. The adjacent points of the graph have been connected with straight line segments.
[\[Uncaptioned image\]](/html/0705.0817/assets/x11.png)
By increasing the number of starter nodes the mean TP flux increases slightly until it reaches $`\Phi_{m} = k`$. Indeed, if all the nodes of the graph are starters, then each of them will send a TP to all the other nodes. The increasing of the number of starter nodes decreases the time required to complete the exploration of the graph. The figure below shows this result. In the $`x`$ axis, an integer point corresponds to the number of starter nodes. Each plotted line corresponds to a complete graph with different number of nodes. The first bisector has been plotted to facilitate the reading of the graph.
[\[Uncaptioned image\]](/html/0705.0817/assets/x12.png)
The complete graph is the worst case for the $`Q^{2}`$, therefore in the general case the mean TP flux will be:
<table>
<tbody>
<tr>
<td></td>
<td>Φm≤nsubscriptΦ𝑚𝑛\Phi_{m}\leq n</td>
<td></td>
</tr>
</tbody>
</table>
where $`n`$ is the number of nodes of the graph. For example, in random graphs with increasing number of nodes, the mean TP flux will assume the distribution shown in the figure below:
[\[Uncaptioned image\]](/html/0705.0817/assets/x13.png)
The same happens for a mesh graph. The mesh graph used in the simulations is a square grid where every intersection represents a node.
If the graph is uniformly connected, as in the case of a mesh graph, then the distribution of single TP fluxes is uniform too. The majority of single TP fluxes are near to the mean TP flux, while the highest TP fluxes are shared between different nodes, i.e. there aren’t few unlucky nodes which have to bear a high TP flux. In the following figure, we can observe the TP flux distribution of two different QSPN exploration of a mesh graph with $`11 \times 11`$ nodes. In the $`x`$ axis, an integer point corresponds to a single node of the graph. Adjacent points are connected by a straight line.
[\[Uncaptioned image\]](/html/0705.0817/assets/x14.png)
In the first exploration, plotted in red, the starter node (node 40) and the eight nodes surrounding him (node 50,51,52,28,29,30,39,41) have the highest TP flux. The TP flux decreases with the increase of the distance from the starter node. Since the graph is a $`11 \times 11`$ grid, there are $`11`$ classes where the TP flux falls and for this reason in the graph there are exactly $`11`$ slopes. The nodes with lower TP fluxes are those with the least number of links, i.e. the nodes on the edge of the mesh, which have only 3 links (and not 4). The four nodes at the vertices of the grid (node 0,10,110,120), which have just 2 links, register the lowest TP flux.
The second exploration, plotted in blue, simulates the scenario where 32 links of the graph change and consequently the interested nodes send a new CTP to inform the other nodes and update their maps. As supposed, the mean TP flux (26.80) is heavily lower than that of the initial exploration (82.90), because the majority of routes are already known by the nodes. Even the execution time is lower (2.6s vs 3.5s). Note that this scenario is the most common in real world, because in general, the nodes won’t join or die at once, but progressively.
## 14 TODO
1. 1\.
The $`Q^{2}`$ must be implemented in the ntkd daemon.
2. 2\.
Improve, test and implement the Caustic Routing: http://lab.dyne.org/Ntk_caustic_routing
3. 3\.
Research a “mobile QSPN”
## 15 ChangeLog
- •
`April 2007`
- –
New section: “Network dynamics” (<a href="#S11" style="font-size:90%;" title="11 Network dynamics ‣ Quantum Shortest Path Netsukuku">11</a>)
- –
Description of the ETP (sec. <a href="#S11.SS1" style="font-size:90%;" title="11.1 Extended Tracer Packet ‣ 11 Network dynamics ‣ Quantum Shortest Path Netsukuku">11.1</a>)
- –
Link ID section remove. With the ETP they are no more necessary.
- –
More detailed description of the QSPN v1 (sec. <a href="#S10" style="font-size:90%;" title="10 QSPN v1 ‣ Quantum Shortest Path Netsukuku">10</a>)
- –
Subsection “QSPN v2 - High levels” removed. It was redundant with the topology document\[[2](#bib.bib2)\]
- •
`October 2006`
Initial release.
## References
- \[1\]
Netsukuku website: http://netsukuku.freaknet.org/
- \[2\]
Netsukuku topology document: http://netsukuku.freaknet.org/doc/main_doc/topology.pdf
- \[3\]
Depth-First Search: http://en.wikipedia.org/wiki/Depth-first_search
- \[4\]
Generate Routes in Awk:
http://cvs.hinezumi.org/viewcvs/netsukuku/proto/doc/qspn/generate_routes.awk
- \[5\]
Simplify Routes in Awk:
http://cvs.hinezumi.org/viewcvs/netsukuku/proto/doc/qspn/simplify_routes.awk
- \[6\]
QSPN v2 simulator:
http://cvs.hinezumi.org/viewcvs/netsukuku/proto/doc/qspn/q2sim.py
- \[7\]
Complete graph:
http://mathworld.wolfram.com/CompleteGraph.html
- \[8\]
Network simulator:
http://www-mash.cs.berkeley.edu/ns/
- \[9\]
NTK_RFC 002: http://lab.dyne.org/Ntk_bandwidth_measurement
- \[10\]
A Survey of Two Signature Aggregation Techniques:
http://crypto.stanford.edu/ dabo/abstracts/aggsurvey.html
- \[11\]
Aggregate and Verifiably Encrypted Signatures from Bilinear Maps:
http://crypto.stanford.edu/ dabo/abstracts/aggreg.html
`^_^` |
0705.3683 | cs.CR | cyber | # Power-Efficient Direct-Voting Assurance for Data Fusion in Wireless Sensor Networks
<sup>†</sup>
<sup>†</sup>thanks: The preliminary result of this work is presented at the IEEE International Conference on Sensor Networks, Ubiquitous, and Trustworthy Computing (SUTC2006), Taiwan, R.O.C.. This work was supported by the National Science Council of Taiwan, R.O.C., under grants NSC 94-2213-E-305-002 and NSC 94-2213-E-305-001.
Hung-Ta Pai and Yunghsiang S. Han The authors are with the Graduate Institute of Communication Engineering, National Taipei University, Taiwan (E-mail: htpai@mail.ntpu.edu.tw; yshan@mail.ntpu.edu.tw).
###### Abstract
Wireless sensor networks place sensors into an area to collect data and send them back to a base station. Data fusion, which fuses the collected data before they are sent to the base station, is usually implemented over the network. Since the sensor is typically placed in locations accessible to malicious attackers, information assurance of the data fusion process is very important. A witness-based approach has been proposed to validate the fusion data. In this approach, the base station receives the fusion data and ”votes” on the data from a randomly chosen sensor node. The vote comes from other sensor nodes, called ”witnesses,” to verify the correctness of the fusion data. Because the base station obtains the vote through the chosen node, the chosen node could forge the vote if it is compromised. Thus, the witness node must encrypt the vote to prevent this forgery. Compared with the vote, the encryption requires more bits, increasing transmission burden from the chosen node to the base station. The chosen node consumes more power. This work improves the witness-based approach using direct voting mechanism such that the proposed scheme has better performance in terms of assurance, overhead, and delay. The witness node transmits the vote directly to the base station. Forgery is not a problem in this scheme. Moreover, fewer bits are necessary to represent the vote, significantly reducing the power consumption. Performance analysis and simulation results indicate that the proposed approach can achieve a $`40`$ times better overhead than the witness-based approach.
###### Index Terms:
wireless sensor networks, data fusion assurance, power-efficient, voting mechanism, witness
## I Introduction
Wireless sensor networks (WSNs) comprise many tiny, low-cost, battery-powered sensors in a small area \[[1](#bib.bib1), [5](#bib.bib5), [22](#bib.bib22), [11](#bib.bib11), [21](#bib.bib21), [12](#bib.bib12), [13](#bib.bib13)\]. The sensors detect environmental variations and then transmit the detection results to other sensors or a base station \[[2](#bib.bib2), [4](#bib.bib4), [24](#bib.bib24), [6](#bib.bib6), [7](#bib.bib7)\]. One or several sensors then collect the detection results from other sensors. The collected data must be processed by the sensor to reduce the transmission burden before they are transmitted to the base station. This process is called data fusion and the sensor performing data fusion is the fusion node. The fusion data may be sent from the fusion node to the base station through multiple hops \[[10](#bib.bib10)\] or a direct link \[[19](#bib.bib19)\].
Although fusion significantly lowers the traffic between the fusion node and the base station, the fusion node is more critical and vulnerable to malicious attacks than non-fusion sensors \[[20](#bib.bib20), [16](#bib.bib16), [15](#bib.bib15)\]. If a fusion node is compromised, then the base station cannot ensure the correctness of the fusion data sent to it. This problem of fusion data assurance arises because the detection results are not sent directly to the base station, so the fusion result cannot usually be verified.
This problem can be resolved in two ways: One is hardware-based \[[3](#bib.bib3), [14](#bib.bib14)\] and the other is software-based \[[17](#bib.bib17), [8](#bib.bib8), [9](#bib.bib9)\]. Since the hardware-based approach requires extra circuits to detect or frustrate the compromised node, the cost and continual power consumption of sensors are increased but still cannot guarantee protection against all attacks. Conversely, the software methods generally require no or little extra hardware for data assurance. However, as mentioned in \[[17](#bib.bib17), [8](#bib.bib8)\], several copies of the fusion data must be sent to the base station, the power consumption for the data transmission is very high.
The witness-based approach presented by Du et al. \[[9](#bib.bib9)\] does not have this difficulty. Several fusion nodes are used to fuse the collected data and have the ability to communicate with the base station. Only one node is chosen to transmit the fusion result to the base station. The other fusion nodes, serving as witnesses, encrypt the fusion results to message authentication codes (MACs). The MACs are then sent to the base station through the chosen fusion node. Finally, the base station utilizes the received MACs to verify the received fusion data. The verification could be wrong since the chosen node could be compromised and forge MACs. The correctness of the verification depends not only on the number of malicious fusion nodes, but also on the length of the MAC. A long MAC increases the reliability of the verification. However, the transmission of the long MAC imposes a high communication burden. If the received fusion result at the base station cannot pass the verification, then a polling scheme is started to determine whether any valid fusion result is available at the other fusion nodes. In addition to the fusion result sent by the malicious fusion node, several copies of the correct fusion result may also have to be transmitted to the base station. The transmission of the correct fusion result consumes the power of the uncompromised fusion node.
This work develops a new data fusion assurance to improve the witness-based method of Du et al. \[[9](#bib.bib9)\]. The correctness of the verification in the proposed scheme depends only on the number of compromised fusion nodes. As in the witness-based approach, a fusion node is selected to transmit the fusion result, while other fusion nodes serve as witnesses. Nevertheless, the base station obtains votes contributing to the transmitted fusion result directly from the witness nodes. No valid fusion data are available if the transmitted fusion data are not approved by a pre-set number of witness nodes. Based on this voting mechanism, two schemes are addressed: one needs variant rounds of voting and the other needs only one round of voting. In the variant-round scheme, only one copy of the correct fusion data provided by one uncompromised fusion node is transmitted to the base station. Analytical and simulation results reveal that the proposed scheme is up to $`40`$ times better on the overhead than that of Du et al. \[[9](#bib.bib9)\]. In the one-round scheme, the base station polls each sensor once at most. The maximum delay of the one-round scheme is much less than the variant-round scheme.
The remainder of this work is organized as follows. Section [1](#S2.F1 "Figure 1 ‣ II Data Fusion Assurance Problem and The Previous Work ‣ Power-Efficient Direct-Voting Assurance for Data Fusion in Wireless Sensor Networks The preliminary result of this work is presented at the IEEE International Conference on Sensor Networks, Ubiquitous, and Trustworthy Computing (SUTC2006), Taiwan, R.O.C.. This work was supported by the National Science Council of Taiwan, R.O.C., under grants NSC 94-2213-E-305-002 and NSC 94-2213-E-305-001.") briefly addresses the problem of data fusion assurance in WSNs and previous works on the problem. Section [III](#S3 "III Improved Voting Mechanism ‣ Power-Efficient Direct-Voting Assurance for Data Fusion in Wireless Sensor Networks The preliminary result of this work is presented at the IEEE International Conference on Sensor Networks, Ubiquitous, and Trustworthy Computing (SUTC2006), Taiwan, R.O.C.. This work was supported by the National Science Council of Taiwan, R.O.C., under grants NSC 94-2213-E-305-002 and NSC 94-2213-E-305-001.") describes the variant-round scheme and analyzes its performance in terms of overhead and delay. The description of the one-round scheme is also given. Section [IV](#S4 "IV Performance Evaluation ‣ Power-Efficient Direct-Voting Assurance for Data Fusion in Wireless Sensor Networks The preliminary result of this work is presented at the IEEE International Conference on Sensor Networks, Ubiquitous, and Trustworthy Computing (SUTC2006), Taiwan, R.O.C.. This work was supported by the National Science Council of Taiwan, R.O.C., under grants NSC 94-2213-E-305-002 and NSC 94-2213-E-305-001.") presents a performance evaluation of the proposed approach. Concluding remarks and suggestions for future work are presented in Section [V](#S5 "V Conclusions and Future Work ‣ Power-Efficient Direct-Voting Assurance for Data Fusion in Wireless Sensor Networks The preliminary result of this work is presented at the IEEE International Conference on Sensor Networks, Ubiquitous, and Trustworthy Computing (SUTC2006), Taiwan, R.O.C.. This work was supported by the National Science Council of Taiwan, R.O.C., under grants NSC 94-2213-E-305-002 and NSC 94-2213-E-305-001.").
## II Data Fusion Assurance Problem and The Previous Work
[Refer to caption](/html/0705.3683/assets/x1.png)
Figure 1: Structure of a wireless sensor network for distributed detection using $`N`$ sensors and $`M`$ fusion nodes
Figure [1](#S2.F1 "Figure 1 ‣ II Data Fusion Assurance Problem and The Previous Work ‣ Power-Efficient Direct-Voting Assurance for Data Fusion in Wireless Sensor Networks The preliminary result of this work is presented at the IEEE International Conference on Sensor Networks, Ubiquitous, and Trustworthy Computing (SUTC2006), Taiwan, R.O.C.. This work was supported by the National Science Council of Taiwan, R.O.C., under grants NSC 94-2213-E-305-002 and NSC 94-2213-E-305-001.") depicts a wireless sensor network for distributed detection with $`N`$ sensors for collecting environment variation data, and a fusion center for making a final decision of detections. This network architecture is similar to the so-called SENsor with Mobile Access (SENMA) \[[28](#bib.bib28), [23](#bib.bib23)\], Message Ferry \[[29](#bib.bib29)\], and Data Mule \[[18](#bib.bib18)\]. At the $`j`$th sensor, one observation $`y_{j}`$ is undertaken for one of phenomena $`H_{i}`$, where $`i = {1,2,\ldots,L}`$. If the detection (raw) data are transmitted to the fusion nodes without any processing, then the transmission imposes a very high communication burden. Hence, each sensor must make a local decision based on the raw data before transmission. The decisions
<sup>1</sup>
<sup>1</sup>1These decisions could be compressed data whose sizes depend on the applications of the WSNs.
, $`v_{j}`$, $`j = {1,2,\ldots,N}`$, can be represented with fewer symbols than the raw data. The sensor then transmits the local decision to $`M`$ fusion nodes using broadcast. The fusion node can combine all of the local decisions to yield a final result, and directly communicate with the base station. Finally, one of the fusion nodes is specified to send the final result to the base station. Unless all of the fusion nodes or all of the sensors fail, this detection and fusion scheme can guarantee that the base station can obtain the detection result. However, the accuracy of the result is not certain.
Two problems must be solved to ensure that the base station obtain the correct result. First, every fusion node must correctly fuse all of the local decisions, which also implies that all of the fusion results must be the same. Several algorithms have been proposed to deal with this problem \[[25](#bib.bib25), [7](#bib.bib7), [27](#bib.bib27), [26](#bib.bib26)\]. This work assumes that this problem has been solved. The second problem concerns assurance of the fusion result. The transmission between the fusion node and the base station is assumed herein to be error-free. Since some fusion nodes may be compromised, the fusion node chosen by the base station to transmit the fusion result may be one of the compromised nodes. Malicious data may be sent by the compromised node, and the base station cannot discover the compromised nodes from the normal fusion nodes since the data detected by the sensor are not sent directly to the base station. Consequently, the result obtained at the base station may be incorrect.
Du et al. \[[9](#bib.bib9)\] presented a witness-based approach to ensure the correctness of the fusion result. All fusion nodes, other than the chosen node, acts as witnesses of the transmitted fusion result. The witness nodes encrypt their own fusion results to MACs with private keys shared with the base station, and send the MACs, as ”votes”, to the chosen node. In the $`T + 1`$ out of $`M`$ voting scheme, the chosen node collects all MACs from the witness nodes, and transmits them with its own fusion result to the base station. The base station can determine the received data whether the fusion result from the chosen node is accurate. If the fusion result is supported by at least $`T`$ MACs, then the base station accepts the fusion result. Normally, $`T > {\lfloor{M/2}\rfloor}`$. However, although the number of compromised nodes $`C < T`$, the accepted fusion result is not always correct. If the chosen node is compromised, then it may forge the fusion result and the MAC. Let the size of each MAC be $`k_{w}`$ (bits). Since the number of the transmitted MACs is $`M - 1`$, the number of the transmitted bits, in addition to the fusion result, is $`{({M - 1})}\hspace{0pt}k_{w}`$. The probability that the base station accepts the forged fusion result is given by
<table>
<tbody>
<tr>
<td></td>
<td>Pe=∑i=TM−1(M−1i)(12kw)i(1−12kw)M−i−1.subscript𝑃𝑒superscriptsubscript𝑖𝑇𝑀1𝑀1𝑖superscript1superscript2subscript𝑘𝑤𝑖superscript11superscript2subscript𝑘𝑤𝑀𝑖1P_{e}=\sum_{i=T}^{M-1}\left(\begin{array}[]{c}M-1\\ i\end{array}\right)\left(\frac{1}{2^{k_{w}}}\right)^{i}\left(1-\frac{1}{2^{k_{w}}}\right)^{M-i-1}.</td>
<td></td>
</tr>
</tbody>
</table>
For instance, consider the majority voting rule in which $`{T + 1} \geq {\lceil{M/2}\rceil}`$. To ensure that $`P_{e} \leq 2^{- 10}`$, set $`k_{w} = {\lceil{2\hspace{0pt}\left( {{10/{({M - 1})}} + 1} \right)}\rceil}`$. Additionally, although only one copy of the fusion result is sent to the base station in this witness-based approach, the witness nodes still requires significant communication bandwidth because the MACs of the fusion results are transmitted.
To compare the proposed scheme fairly with the witness-based approach, the overhead is defined as the total number of bits, except the bits for one copy of the correct fusion result, transmitted to the base station by uncompromised nodes during the data assurance process.
<sup>2</sup>
<sup>2</sup>2The method does not consider the power consumption of all compromised nodes since they are not useful to the WSN.
The overhead can therefore be regarded as the useful power consumed for the data assurance. Moreover, the round delay is defined as the number of rounds
<sup>3</sup>
<sup>3</sup>3Or the number of fusion results sent to the base station.
required to collect all MACs (votes) from the witness nodes and the polling delay the number of votes (including all agree and disagree voting).
<sup>4</sup>
<sup>4</sup>4The overall time delay is then can be derived by these two delays.
The overhead of the witness-based approach given in \[[9](#bib.bib9)\] is then derived as follows. If the received fusion result is not accepted, then the base station may start a polling mechanism to seek the correct fusion result. The base station randomly specifies another fusion node. The new chosen node then sends its fusion result and all MACs from the witness nodes to the base station.
<sup>5</sup>
<sup>5</sup>5All MACs must be sent to the base station again to avoid the denial of service since the previously chosen compromised fusion node might modify MACs it sent to the base station. This action is not clearly presented in \[[9](#bib.bib9)\].
When the number of compromised fusion nodes, $`C`$, is greater than $`M - T - 1`$ but less than $`T + 1`$,
<sup>6</sup>
<sup>6</sup>6The $`C`$ compromised nodes are assumed to collude to forge a wrong fusion result. Hence, if $`C \geq {T + 1}`$, then they can successfully forge a wrong fusion result, and the base station accepts the forged result.
the fusion result is invalid and $`M - T`$ fusion nodes are chosen to transmit their fusion results to the base station in the polling process. Since the overhead defined only considers the power consumption of uncompromised nodes, the number of uncompromised nodes among the $`M - T`$ fusion nodes must be discovered. The probability that $`i`$ of the $`M - T`$ fusion nodes are uncompromised is given by
<table>
<tbody>
<tr>
<td></td>
<td>Pw(i)subscript𝑃𝑤𝑖\displaystyle P_{w}(i)</td>
<td>=\displaystyle{}={}</td>
<td>(M−Ci)(CM−T−i)(MM−T)𝑀𝐶𝑖𝐶𝑀𝑇𝑖𝑀𝑀𝑇\displaystyle\frac{\left(\begin{array}[]{c}M-C\\ i\end{array}\right)\left(\begin{array}[]{c}C\\ M-T-i\end{array}\right)}{\left(\begin{array}[]{c}M\\ M-T\end{array}\right)}</td>
<td></td>
</tr>
<tr>
<td></td>
<td></td>
<td>=\displaystyle{}={}</td>
<td>(M−Ti)(TM−C−i)(MC),𝑀𝑇𝑖𝑇𝑀𝐶𝑖𝑀𝐶\displaystyle\frac{\left(\begin{array}[]{c}M-T\\ i\end{array}\right)\left(\begin{array}[]{c}T\\ M-C-i\end{array}\right)}{\left(\begin{array}[]{c}M\\ C\end{array}\right)},</td>
<td></td>
</tr>
</tbody>
</table>
where $`0 \leq i \leq {M - T}`$. Let $`K`$ be the number of bits representing the fusion result. The average overhead is thus
<table>
<tbody>
<tr>
<td></td>
<td>Ow(M,T,C)=∑i=1M−TPw(i)[(M−1)ikw+K(i−1)] (bits),subscript𝑂𝑤𝑀𝑇𝐶superscriptsubscript𝑖1𝑀𝑇subscript𝑃𝑤𝑖delimited-[]𝑀1𝑖subscript𝑘𝑤𝐾𝑖1 (bits)\displaystyle O_{w}(M,T,C)=\sum_{i=1}^{M-T}P_{w}(i)\left[(M-1)ik_{w}+K(i-1)\right]\mbox{ (bits)},</td>
<td></td>
<td>(3)</td>
</tr>
</tbody>
</table>
where $`i - 1`$ is used because the overhead is defined such that one copy of correct fusion result does not count on. Equation ([3](#S2.E3 "In II Data Fusion Assurance Problem and The Previous Work ‣ Power-Efficient Direct-Voting Assurance for Data Fusion in Wireless Sensor Networks The preliminary result of this work is presented at the IEEE International Conference on Sensor Networks, Ubiquitous, and Trustworthy Computing (SUTC2006), Taiwan, R.O.C.. This work was supported by the National Science Council of Taiwan, R.O.C., under grants NSC 94-2213-E-305-002 and NSC 94-2213-E-305-001.")) indicates that the number of the correct fusion results transmitted by the uncompromised fusion nodes may be up to $`M - T`$. In other words, the power of the uncompromised nodes is significantly wasted. Moreover, because each chosen node has to collects all MACs from the witness nodes, the average round delay, $`R_{w}`$, is $`M - T`$ and the average polling delay $`D_{w}`$ is $`{({M - T})}\hspace{0pt}{({M - 1})}`$.
Conversely, when the number of uncompromised nodes is greater than $`T`$, the base station can obtain the correct fusion result. If the base station gets the correction result at round $`i`$, meaning that the chosen fusion nodes from round $`1`$ to $`i - 1`$ are compromised, then the average round delay can be given by
<table>
<tbody>
<tr>
<td></td>
<td>Rw=∑i=1C+1i(M−iC−i+1)(MC)subscript𝑅𝑤superscriptsubscript𝑖1𝐶1𝑖𝑀𝑖𝐶𝑖1𝑀𝐶R_{w}=\sum_{i=1}^{C+1}i\frac{{\left(\begin{array}[]{c}M-i\\ C-i+1\end{array}\right)}}{{\left(\begin{array}[]{c}M\\ C\end{array}\right)}}</td>
<td></td>
</tr>
</tbody>
</table>
and the average polling delay is $`R_{w}\hspace{0pt}{({M - 1})}`$. The average overhead is given by $`{({M - 1})}\hspace{0pt}k_{w}`$. Notice that the maximum round delay is $`C + 1`$. When the fusion result is valid, $`40`$ and $`60`$ bits must be transmitted to the base station when $`M = 11`$ ($`k_{w} = 4`$, i.e., $`P_{e} \leq 2^{- 10}`$) and $`M = 21`$ ($`k_{w} = 3`$, i.e., $`P_{e} \leq 2^{- 10}`$), respectively, setting $`K = 0`$ (in practice, $`K > 0`$). A large amount of power must be consumed for this transmission, significantly reducing the lifetime of the fusion node. The problem of power consumption is even worse when the fusion result is invalid. For example, the maximum average overhead is about $`109`$ and $`314`$ bits for $`M = 11`$ ($`C = 5`$ and $`T = 6`$) and $`M = 21`$ ($`C = 10`$ and $`T = 11`$), respectively. Therefore, the witness-based approach must be enhanced.
## III Improved Voting Mechanism
The voting mechanism in the witness-based approach is designed according to the MAC of the fusion result at each witness node. This design is reasonable when the witness node does not know about the fusion result at the chosen node. However, in practice, the base station can transmit the fusion result of the chosen node to the witness or the witness node is in the communication range of the chosen node and the base station. Therefore, the witness node can obtain the transmitted fusion result from the chosen node through the base station or overhearing. The witness node then can compare the transmitted fusion result with its own fusion result. Finally, the witness node can send its vote (agreement or disagreement) on the transmitted result directly to the base station, rather than through the chosen node.
The base station has to set up a group key for all fusion nodes to ensure that the direct voting mechanism works.
<sup>7</sup>
<sup>7</sup>7We assume that all witnesses can overhear the fusion result sent by the chosen node. If it is not the case, the direct voting mechanism needs to be slightly modified.
When a fusion node wishes to send its fusion result to the base station, it adopts the group key to encrypt the result, and other fusion nodes serving as witness nodes can decode the encrypted result. The witness node then starts to vote on the transmitted result. Two data fusion assurance schemes are proposed based on the voting mechanism using a group key.
###
III-A
Variant-round Scheme
In this scheme, the base station needs to ask the witness node whether it agrees or disagrees with the transmitted fusion result. The witness node then sends its vote to the base station. If the transmitted fusion result is not supported by at least $`T`$ witness nodes, then the base station might have to select a witness node that does not agree with the transmitted result as the next chosen node. The detail steps of the scheme are given as follows:
Step 1:
The base station chooses a fusion node. Other fusion nodes serve as witness nodes. Define a set of witness nodes that includes all witness nodes and let the nodes in the set be randomly ordered. Denote $`M' = {M - 1}`$ as the size of the witness set in the current round.
Step 2:
The chosen node transmits its fusion result to the base station.
Step 3:
The base station polls the node in the witness set by following the order of the witness nodes. The polling-for-vote
<sup>8</sup>
<sup>8</sup>8The rest of this work utilizes ”voting”, ”polling” and ”polling-for-vote” interchangeably.
process does not stop until
- •
$`T`$ witness nodes agree with the transmitted fusion result (agreeing nodes), where $`1 \leq T \leq {M - 1}`$,
- •
$`{M' - T} + 1`$ witness nodes disagree with the transmitted fusion result (disagreeing nodes), or
- •
all witness nodes have been polled.
Step 4:
Represent $`A`$ as the number of witness nodes that agree with the transmitted fusion result.
- •
If $`A = T`$, then the transmitted fusion result passes the verification of the fusion result. Stop the polling.
- •
If $`{M' - T - 1} < A < T`$, then no reliable fusion result is valid. Stop the polling.
- •
If $`A \leq {M' - T - 1}`$, then exclude the $`A`$ agreeing witness nodes from the witness set. Let the first node that disagrees with the transmitted fusion result be the chosen node to transmit its fusion result. Thus, the updated size of the witness set, $`M'`$, becomes to $`M' - A - 1`$
<sup>9</sup>
<sup>9</sup>9The number of nodes performing the polling-for-vote at next round becomes to $`M' - A`$.
. Go to Step 2 for the next round of the polling.
###
III-B
Analysis of the Variant-round Scheme
This analysis assumes that the compromised node always transmits the forged fusion result while the compromised node is chosen to send its fusion result. When a compromised node serves as a witness node, it always disagrees with the correct fusion result, and agrees with the forged fusion result with a probability $`P_{f}`$. If the compromised node attempts to make the base station accept the forged fusion result, then it always agrees with the fusion result transmitted by other compromised nodes, i.e., $`P_{f} = 1`$, and at most two rounds of voting have to be run. Conversely, if the compromised node wants to make the polling-for-vote process run as long as possible, then it always disagrees with the transmitted fusion result, i.e., $`P_{f} = 0`$. The performance of the variant-round scheme when $`P_{f} = 0`$ is analyzed next. Appendix describes the performance analysis of the variant-round scheme when $`P_{f} = 1`$. When $`P_{f} \neq {0,1}`$, the overheads are presented by computer simulations in next section. Since all compromised nodes (uncompromised nodes) have the same behavior in the following analysis, the analysis can be treated as the problem of counting for $`C`$ black balls (compromised nodes) and $`M - C`$ white balls (uncompromised nodes) together.
If the compromised node always disagrees with the transmitted fusion result, then no forged fusion result is accepted. Two cases must be addressed:
Case 1
$`C \geq {M - T}`$,
Case 2
$`C < {M - T}`$.
Note that the valid fusion result is not available in Case $`1`$.
Assume that the chosen node at the first round is compromised. The probability that the chosen node is compromised at the first round is given by $`C/M`$. The first-round voting finishes when $`M - T`$ witness nodes do not agree with the transmitted fusion result, as described in Step 3. Thus, the polling order (i.e., the order of witness nodes as described in Step 1) determines the number of witness nodes that the base station has to poll at this round of voting. The number of possible polling orders in the sense of the black-white-ball model is given by
<table>
<tbody>
<tr>
<td></td>
<td>Πv1c1=(M−1)!(C−1)!(M−C)!=(M−1C−1),superscriptsubscriptΠ𝑣1𝑐1𝑀1𝐶1𝑀𝐶𝑀1𝐶1\Pi_{v1}^{c1}=\frac{(M-1)!}{(C-1)!(M-C)!}=\left(\begin{array}[]{c}M-1\\ C-1\end{array}\right),</td>
<td></td>
</tr>
</tbody>
</table>
where the subscript, $`v\hspace{0pt}1`$, denotes the first case of the variant-round scheme, and the superscript, $`c\hspace{0pt}1`$, represents the first round of voting when the chosen node is compromised. Since the chosen node at the first-round voting is compromised, the polling stops after $`M - T`$ witness nodes have been polled. Moreover, the number of unpolled nodes is $`{M - 1 - {({M - T})}} = {T - 1}`$, and the number of uncompromised nodes among the unpolled nodes is $`M - C - i`$ if there are $`i`$ uncompromised nodes polled. Thus, the probability that $`i`$ of the $`M - T`$ polled nodes are uncompromised, where $`0 \leq i \leq {M - T}`$,
<sup>10</sup>
<sup>10</sup>10Actually, $`{\max{\{{{M - T - C} + 1},0\}}} \leq i`$. Since $`\binom{a}{b}`$ is $`0`$ when $`b > a`$, $`0`$ is adopted as the lower bound of $`i`$.
is then written by
<table>
<tbody>
<tr>
<td></td>
<td>1Πv1c1(M−Ti)(T−1M−C−i).1superscriptsubscriptΠ𝑣1𝑐1𝑀𝑇𝑖𝑇1𝑀𝐶𝑖\frac{1}{\Pi_{v1}^{c1}}{\left(\begin{array}[]{c}M-T\\ i\end{array}\right)\left(\begin{array}[]{c}T-1\\ M-C-i\end{array}\right)}.</td>
<td></td>
</tr>
</tbody>
</table>
No node is excluded from the witness set. The number of compromised nodes and the size of the witness set at the second round become $`C - 1`$ and $`M - 2`$, respectively. The average overhead, when the chosen node at the first round is compromised, is expressed recursively by
<table>
<tbody>
<tr>
<td></td>
<td>Ov1c(M,T,C,0)=∑i=0M−T1Πv1c1(M−Ti)(T−1M−C−i)ik′+Ov1(M−1,T,C−1,0),superscriptsubscript𝑂𝑣1𝑐𝑀𝑇𝐶0superscriptsubscript𝑖0𝑀𝑇1superscriptsubscriptΠ𝑣1𝑐1𝑀𝑇𝑖𝑇1𝑀𝐶𝑖𝑖superscript𝑘′subscript𝑂𝑣1𝑀1𝑇𝐶10O_{v1}^{c}(M,T,C,0)=\sum_{i=0}^{M-T}\frac{1}{\Pi_{v1}^{c1}}{\left(\begin{array}[]{c}M-T\\ i\end{array}\right)\left(\begin{array}[]{c}T-1\\ M-C-i\end{array}\right)}ik^{\prime}+O_{v1}(M-1,T,C-1,0),</td>
<td></td>
<td>(4)</td>
</tr>
</tbody>
</table>
where $`O_{v\hspace{0pt}1}\hspace{0pt}{({M - 1},T,{C - 1},0)}`$ represents the average overhead when the number of fusion nodes is $`M - 1`$ and the number of compromised nodes is $`C - 1`$, and $`k`$ ($`k'`$) is the number of bits that a witness must send to the base station while it agrees (disagrees) with the transmitted fusion result.
<sup>11</sup>
<sup>11</sup>11The bits sent to the base station when a node agrees with the fusion result are separated from those sent when the node disagrees with the result, since the node can be silenced when it agrees with the result, while only a few bits are sent when it disagrees.
Moreover, the average round delay and the average polling delay under the same condition are represented by
<table>
<tbody>
<tr>
<td></td>
<td>Rv1c(M,T,C,0)=1+Rv1(M−1,T,C−1,0),superscriptsubscript𝑅𝑣1𝑐𝑀𝑇𝐶01subscript𝑅𝑣1𝑀1𝑇𝐶10R_{v1}^{c}(M,T,C,0)=1+R_{v1}(M-1,T,C-1,0),</td>
<td></td>
<td>(5)</td>
</tr>
</tbody>
</table>
and
<table>
<tbody>
<tr>
<td></td>
<td>Dv1c(M,T,C,0)=M−T+Dv1(M−1,T,C−1,0),superscriptsubscript𝐷𝑣1𝑐𝑀𝑇𝐶0𝑀𝑇subscript𝐷𝑣1𝑀1𝑇𝐶10D_{v1}^{c}(M,T,C,0)=M-T+D_{v1}(M-1,T,C-1,0),</td>
<td></td>
<td>(6)</td>
</tr>
</tbody>
</table>
where $`R_{v\hspace{0pt}1}\hspace{0pt}{({M - 1},T,{C - 1},0)}`$ and $`D_{v\hspace{0pt}1}\hspace{0pt}{({M - 1},T,{C - 1},0)}`$ are the average round delay and polling delay, respectively, when the number of fusion nodes performing the polling-for-vote is $`M - 1`$ and the number of compromised nodes among them is $`C - 1`$.
Next, suppose that the chosen node at the first round is not compromised, which has a probability of $`{({M - C})}/M`$. The number of the possible polling orders is given by
<table>
<tbody>
<tr>
<td></td>
<td>Πv1u1=(M−1C),superscriptsubscriptΠ𝑣1𝑢1𝑀1𝐶\Pi_{v1}^{u1}=\left(\begin{array}[]{c}M-1\\ C\end{array}\right),</td>
<td></td>
</tr>
</tbody>
</table>
where the superscript, $`u\hspace{0pt}1`$, denotes the first round of polling while the chosen node is uncompromised. When the polling stops at witness node $`j`$, the node does not agree with the transmitted result and the base station has polled $`M - T`$ disagreeing nodes (including witness node $`j`$). Moreover, $`M - j - 1`$ nodes are unpolled, and $`{C - {({M - T})}} = {{T + C} - M}`$ of these are compromised. Since the witness set has $`M - C - 1`$ uncompromised nodes, the maximum number of polled witness nodes is $`{({M - C - 1})} + {({M - T})}`$. Thus, the probability that the polling stops at the $`j`$th witness node, where $`{M - T} \leq j \leq {{({M - C - 1})} + {({M - T})}} = {{2\hspace{0pt}M} - T - C - 1}`$, is given by
<table>
<tbody>
<tr>
<td></td>
<td>Pv1u1(j)=1Πv1u1(j−1M−T−1)(M−j−1T+C−M).superscriptsubscript𝑃𝑣1𝑢1𝑗1superscriptsubscriptΠ𝑣1𝑢1𝑗1𝑀𝑇1𝑀𝑗1𝑇𝐶𝑀P_{v1}^{u1}(j)=\frac{1}{\Pi_{v1}^{u1}}\left(\begin{array}[]{c}j-1\\ M-T-1\end{array}\right)\left(\begin{array}[]{c}M-j-1\\ T+C-M\end{array}\right).</td>
<td></td>
</tr>
</tbody>
</table>
Since the number of unpolled nodes $`M - j - 1`$ is less than $`T`$, these nodes will never be polled in the following runs before the voting mechanism stops. Hence, only compromised nodes are polled after the first round and then no uncompromised nodes are further polled. Note that the number of uncompromised nodes among the polled nodes is $`{j - {({M - T})}} = {{j - M} + T}`$. Therefore, the average overhead, when the chosen node at the first round is not compromised, is then given by
<table>
<tbody>
<tr>
<td></td>
<td>Ov1u(M,T,C,0)=∑j=M−T2M−T−C−1Pv1u1(j)(j−M+T)k.superscriptsubscript𝑂𝑣1𝑢𝑀𝑇𝐶0superscriptsubscript𝑗𝑀𝑇2𝑀𝑇𝐶1superscriptsubscript𝑃𝑣1𝑢1𝑗𝑗𝑀𝑇𝑘O_{v1}^{u}(M,T,C,0)=\sum_{j=M-T}^{2M-T-C-1}P_{v1}^{u1}(j)(j-M+T)k.</td>
<td></td>
<td>(7)</td>
</tr>
</tbody>
</table>
Moreover, the size of the witness set after the first round becomes $`M' = {{{({M - j - 1})} + {({M - T})}} - 1} = {{2\hspace{0pt}M} - T - j - 2}`$ (the total number of unpolled and disagreeing nodes minus 1). The polling process stops if
<table>
<tbody>
<tr>
<td></td>
<td></td>
<td></td>
<td>M′≤T−1superscript𝑀′𝑇1\displaystyle M^{\prime}\leq T-1</td>
<td></td>
</tr>
<tr>
<td></td>
<td></td>
<td>⇔⇔\displaystyle\Leftrightarrow</td>
<td>2M−T−j−2≤T−12𝑀𝑇𝑗2𝑇1\displaystyle 2M-T-j-2\leq T-1</td>
<td></td>
</tr>
<tr>
<td></td>
<td></td>
<td>⇔⇔\displaystyle\Leftrightarrow</td>
<td>j≥2M−2T−1.𝑗2𝑀2𝑇1\displaystyle j\geq 2M-2T-1.</td>
<td></td>
</tr>
</tbody>
</table>
Otherwise, the size is decreased by $`1`$ in each following round until it becomes $`T`$. Consequently, the total number of the following rounds is $`{{{2\hspace{0pt}M} - T - j - 2 - T} + 1} = {{2\hspace{0pt}M} - {2\hspace{0pt}T} - j - 1}`$ and then the number of the total rounds is $`{{{2\hspace{0pt}M} - {2\hspace{0pt}T} - j - 1} + 1} = {{2\hspace{0pt}M} - {2\hspace{0pt}T} - j}`$. The average round delay is then represented by
<table>
<tbody>
<tr>
<td></td>
<td>Rv1u(M,T,C,0)=∑j=M−T2M−2T−2Pv1u1(j)(2M−2T−j)+∑j=2M−2T−12M−T−C−1Pv1u1(j).superscriptsubscript𝑅𝑣1𝑢𝑀𝑇𝐶0superscriptsubscript𝑗𝑀𝑇2𝑀2𝑇2superscriptsubscript𝑃𝑣1𝑢1𝑗2𝑀2𝑇𝑗superscriptsubscript𝑗2𝑀2𝑇12𝑀𝑇𝐶1superscriptsubscript𝑃𝑣1𝑢1𝑗R_{v1}^{u}(M,T,C,0)=\sum_{j=M-T}^{2M-2T-2}P_{v1}^{u1}(j)(2M-2T-j)+\sum_{j=2M-2T-1}^{2M-T-C-1}P_{v1}^{u1}(j).</td>
<td></td>
<td>(8)</td>
</tr>
</tbody>
</table>
Since the voting process will stop when $`{M' - T} + 1`$ nodes are polled at each round, the polling delay is
<table>
<tbody>
<tr>
<td></td>
<td>Dv1u(M,T,C,0)superscriptsubscript𝐷𝑣1𝑢𝑀𝑇𝐶0\displaystyle D_{v1}^{u}(M,T,C,0)</td>
<td>=\displaystyle=</td>
<td>∑j=M−T2M−T−C−1Pv1u1(j)jsuperscriptsubscript𝑗𝑀𝑇2𝑀𝑇𝐶1superscriptsubscript𝑃𝑣1𝑢1𝑗𝑗\displaystyle\sum_{j=M-T}^{2M-T-C-1}P_{v1}^{u1}(j)j</td>
<td></td>
<td rowspan="2">(9)</td>
</tr>
<tr>
<td></td>
<td></td>
<td>+\displaystyle+</td>
<td>∑j=M−T2M−2T−2Pv1u1(j)(2M−2T−j)(2M−2T−j−1)2.subscriptsuperscript2𝑀2𝑇2𝑗𝑀𝑇superscriptsubscript𝑃𝑣1𝑢1𝑗2𝑀2𝑇𝑗2𝑀2𝑇𝑗12\displaystyle\sum^{2M-2T-2}_{j=M-T}P_{v1}^{u1}(j)\frac{(2M-2T-j)(2M-2T-j-1)}{2}.</td>
<td></td>
</tr>
</tbody>
</table>
Equations ([4](#S3.E4 "In III-B Analysis of the Variant-round Scheme ‣ III Improved Voting Mechanism ‣ Power-Efficient Direct-Voting Assurance for Data Fusion in Wireless Sensor Networks The preliminary result of this work is presented at the IEEE International Conference on Sensor Networks, Ubiquitous, and Trustworthy Computing (SUTC2006), Taiwan, R.O.C.. This work was supported by the National Science Council of Taiwan, R.O.C., under grants NSC 94-2213-E-305-002 and NSC 94-2213-E-305-001.")) to ([9](#S3.E9 "In III-B Analysis of the Variant-round Scheme ‣ III Improved Voting Mechanism ‣ Power-Efficient Direct-Voting Assurance for Data Fusion in Wireless Sensor Networks The preliminary result of this work is presented at the IEEE International Conference on Sensor Networks, Ubiquitous, and Trustworthy Computing (SUTC2006), Taiwan, R.O.C.. This work was supported by the National Science Council of Taiwan, R.O.C., under grants NSC 94-2213-E-305-002 and NSC 94-2213-E-305-001.")), and the initial conditions then give the average overhead and the average delays of Case 1 as
<table>
<tbody>
<tr>
<td></td>
<td>Ov1(M,T,C,0)={0M≤TCMOv1c(M,T,C,0)+M−CMOv1u(M,T,C,0) else,subscript𝑂𝑣1𝑀𝑇𝐶0cases0𝑀𝑇𝐶𝑀superscriptsubscript𝑂𝑣1𝑐𝑀𝑇𝐶0𝑀𝐶𝑀superscriptsubscript𝑂𝑣1𝑢𝑀𝑇𝐶0 else\displaystyle O_{v1}(M,T,C,0)=\left\{\begin{array}[]{ll}0&M\leq T\\ \frac{C}{M}O_{v1}^{c}(M,T,C,0)+\frac{M-C}{M}O_{v1}^{u}(M,T,C,0)&\mbox{ else}\end{array},\right.</td>
<td></td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td>Rv1(M,T,C,0)={0M≤TCMRv1c(M,T,C,0)+(M−C)MRv1u(M,T,C,0) else,subscript𝑅𝑣1𝑀𝑇𝐶0cases0𝑀𝑇𝐶𝑀superscriptsubscript𝑅𝑣1𝑐𝑀𝑇𝐶0𝑀𝐶𝑀superscriptsubscript𝑅𝑣1𝑢𝑀𝑇𝐶0 else\displaystyle R_{v1}(M,T,C,0)=\left\{\begin{array}[]{ll}0&M\leq T\\ \frac{C}{M}R_{v1}^{c}(M,T,C,0)+\frac{(M-C)}{M}R_{v1}^{u}(M,T,C,0)&\mbox{ else}\end{array},\right.</td>
<td></td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td>Dv1(M,T,C,0)={0M≤TCMDv1c(M,T,C,0)+(M−C)MDv1u(M,T,C,0) else.subscript𝐷𝑣1𝑀𝑇𝐶0cases0𝑀𝑇𝐶𝑀superscriptsubscript𝐷𝑣1𝑐𝑀𝑇𝐶0𝑀𝐶𝑀superscriptsubscript𝐷𝑣1𝑢𝑀𝑇𝐶0 else\displaystyle D_{v1}(M,T,C,0)=\left\{\begin{array}[]{ll}0&M\leq T\\ \frac{C}{M}D_{v1}^{c}(M,T,C,0)+\frac{(M-C)}{M}D_{v1}^{u}(M,T,C,0)&\mbox{ else}\end{array}.\right.</td>
<td></td>
</tr>
</tbody>
</table>
The second case, i.e., $`C < {M - T}`$, produces a valid fusion result. Similarly, if the chosen node at the first-round polling is compromised, then the polling stops after $`M - T`$ witness nodes are polled. The average overhead and the average delays, when the chosen node at the first round is compromised, are expressed respectively as
<table>
<tbody>
<tr>
<td></td>
<td>Ov2c(M,T,C,0)=∑i=M−T−C+1M−T1Πv1c1(M−Ti)(T−1M−C−i)ik′+Ov2(M−1,T,C−1,0),superscriptsubscript𝑂𝑣2𝑐𝑀𝑇𝐶0superscriptsubscript𝑖𝑀𝑇𝐶1𝑀𝑇1superscriptsubscriptΠ𝑣1𝑐1𝑀𝑇𝑖𝑇1𝑀𝐶𝑖𝑖superscript𝑘′subscript𝑂𝑣2𝑀1𝑇𝐶10\displaystyle O_{v2}^{c}(M,T,C,0)=\sum_{i=M-T-C+1}^{M-T}\frac{1}{\Pi_{v1}^{c1}}{\left(\begin{array}[]{c}M-T\\ i\end{array}\right)\left(\begin{array}[]{c}T-1\\ M-C-i\end{array}\right)}ik^{\prime}+O_{v2}(M-1,T,C-1,0),</td>
<td></td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td>Rv2c(M,T,C,0)=1+Rv2(M−1,T,C−1,0),superscriptsubscript𝑅𝑣2𝑐𝑀𝑇𝐶01subscript𝑅𝑣2𝑀1𝑇𝐶10\displaystyle R_{v2}^{c}(M,T,C,0)=1+R_{v2}(M-1,T,C-1,0),</td>
<td></td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td>Dv2c(M,T,C,0)=M−T+Dv2(M−1,T,C−1,0).superscriptsubscript𝐷𝑣2𝑐𝑀𝑇𝐶0𝑀𝑇subscript𝐷𝑣2𝑀1𝑇𝐶10\displaystyle D_{v2}^{c}(M,T,C,0)=M-T+D_{v2}(M-1,T,C-1,0).</td>
<td></td>
</tr>
</tbody>
</table>
Only one round of polling is needed when the chosen node is uncompromised at the first round. When the polling stops at witness node $`j`$, the node agrees with the transmitted result and the base station has polled $`T`$ agreeing nodes (including witness node $`j`$). Moreover, $`M - j - 1`$ nodes are unpolled, and $`M - 1 - C - T`$ of these are uncompromised. The probability that the polling process ends at the $`j`$th witness node, where $`T \leq j \leq {T + C}`$, is given by
<table>
<tbody>
<tr>
<td></td>
<td>Pv2u1(j)=1Πv1u1(j−1T−1)(M−j−1M−C−T−1).superscriptsubscript𝑃𝑣2𝑢1𝑗1superscriptsubscriptΠ𝑣1𝑢1𝑗1𝑇1𝑀𝑗1𝑀𝐶𝑇1P_{v2}^{u1}(j)=\frac{1}{\Pi_{v1}^{u1}}\left(\begin{array}[]{c}j-1\\ T-1\end{array}\right)\left(\begin{array}[]{c}M-j-1\\ M-C-T-1\end{array}\right).</td>
<td></td>
</tr>
</tbody>
</table>
The number of polled uncompromised nodes is $`T`$. The average overhead when the chosen node is uncompromised at the first round is represented as
<table>
<tbody>
<tr>
<td></td>
<td>Ov2u(M,T,C,0)=∑j=TT+CPv2u1(j)Tk,superscriptsubscript𝑂𝑣2𝑢𝑀𝑇𝐶0superscriptsubscript𝑗𝑇𝑇𝐶superscriptsubscript𝑃𝑣2𝑢1𝑗𝑇𝑘O_{v2}^{u}(M,T,C,0)=\sum_{j=T}^{T+C}P_{v2}^{u1}(j)Tk,</td>
<td></td>
</tr>
</tbody>
</table>
and the average polling delay is
<table>
<tbody>
<tr>
<td></td>
<td>Dv2u(M,T,C,0)=∑j=TT+CPv2u1(j)j.superscriptsubscript𝐷𝑣2𝑢𝑀𝑇𝐶0superscriptsubscript𝑗𝑇𝑇𝐶superscriptsubscript𝑃𝑣2𝑢1𝑗𝑗D_{v2}^{u}(M,T,C,0)=\sum_{j=T}^{T+C}P_{v2}^{u1}(j)j.</td>
<td></td>
</tr>
</tbody>
</table>
Consequently, the average overhead $`O_{v\hspace{0pt}2}\hspace{0pt}{(M,T,C,0)}`$, the average round delay $`R_{v\hspace{0pt}2}\hspace{0pt}{(M,T,C,0)}`$, and the average polling delay $`D_{v\hspace{0pt}2}\hspace{0pt}{(M,T,C,0)}`$ can be represented as
<table>
<tbody>
<tr>
<td></td>
<td>Ov2(M,T,C,0)={0M≤TTkM>T and C=0CMOv2c(M,T,C,0)+M−CMOv2u(M,T,C,0) else,subscript𝑂𝑣2𝑀𝑇𝐶0cases0𝑀𝑇𝑇𝑘𝑀𝑇 and 𝐶0𝐶𝑀superscriptsubscript𝑂𝑣2𝑐𝑀𝑇𝐶0𝑀𝐶𝑀superscriptsubscript𝑂𝑣2𝑢𝑀𝑇𝐶0 else\displaystyle O_{v2}(M,T,C,0)=\left\{\begin{array}[]{ll}0&M\leq T\\ Tk&M>T\mbox{ and }C=0\\ \frac{C}{M}O_{v2}^{c}(M,T,C,0)+\frac{M-C}{M}O_{v2}^{u}(M,T,C,0)&\mbox{ else}\end{array}\right.,</td>
<td></td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td>Rv2(M,T,C,0)={0M≤T1M>T and C=0CMRv2c(M,T,C,0)+M−CM else,subscript𝑅𝑣2𝑀𝑇𝐶0cases0𝑀𝑇1𝑀𝑇 and 𝐶0𝐶𝑀superscriptsubscript𝑅𝑣2𝑐𝑀𝑇𝐶0𝑀𝐶𝑀 else\displaystyle R_{v2}(M,T,C,0)=\left\{\begin{array}[]{ll}0&M\leq T\\ 1&M>T\mbox{ and }C=0\\ \frac{C}{M}R_{v2}^{c}(M,T,C,0)+\frac{M-C}{M}&\mbox{ else}\end{array}\right.,</td>
<td></td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td>Dv2(M,T,C,0)={0M≤TTM>T and C=0CMDv2c(M,T,C,0)+M−CMDv2u(M,T,C,0) else.subscript𝐷𝑣2𝑀𝑇𝐶0cases0𝑀𝑇𝑇𝑀𝑇 and 𝐶0𝐶𝑀superscriptsubscript𝐷𝑣2𝑐𝑀𝑇𝐶0𝑀𝐶𝑀superscriptsubscript𝐷𝑣2𝑢𝑀𝑇𝐶0 else\displaystyle D_{v2}(M,T,C,0)=\left\{\begin{array}[]{ll}0&M\leq T\\ T&M>T\mbox{ and }C=0\\ \frac{C}{M}D_{v2}^{c}(M,T,C,0)+\frac{M-C}{M}D_{v2}^{u}(M,T,C,0)&\mbox{ else}\end{array}\right..</td>
<td></td>
</tr>
</tbody>
</table>
An interesting property of the variant-round scheme is that throughout the polling process, at most one fusion result is transmitted from all of the uncompromised nodes to the base station. Hence, the overhead of the scheme is independent of the size of the fusion result.
<sup>12</sup>
<sup>12</sup>12Note that the overhead is defined as the total number of bits transmitted minus the bits of the correct fusion result.
This claim can be proven by the following argument. In the case of a valid fusion result, when the fusion result from an uncompromised node is sent to the base station at a round of polling, the polling process stops at this round and the valid fusion result is obtained by the base station. Accordingly, only one valid fusion result is sent to the base station by all uncompromised nodes. In the case of no valid fusion result, the number of uncompromised nodes as witnesses is less than $`T`$. If a round of the polling process is the first time that the chosen node is uncompromised, this round terminates when either all witness nodes have been polled or $`{M' - T} + 1`$ witness nodes disagree with the transmitted fusion result. In the former case, the polling process stops. In the latter case, another polling round is required. Importantly, in the next round all uncompromised nodes will be the last $`T - 1`$ nodes in the witness set and will not then be chosen to send any fusion result before the polling process is completed.
<sup>13</sup>
<sup>13</sup>13In the next round, all uncompromised polled nodes are deleted from the witness set according to the scheme.
Therefore, only one fusion result is sent by all uncompromised nodes when no valid result can be obtained by the base station.
###
III-C
One-round Scheme
The number of rounds in the above scheme is not fixed. Hence, the delay varies. Variant delay is not desired in some applications such as real-time systems. This work proposes another scheme that is based on the improved voting mechanism. In this scheme the base station may receive different fusion results from the witness nodes. It requires that all received fusion results be stored. This scheme has a fixed delay and is summarized as follows:
Step 1:
The base station randomly chooses a fusion node. Other fusion nodes serve as witness nodes. Define a set of witness nodes that includes all of the witness nodes and let the nodes in the set be randomly ordered.
Step 2:
The chosen node transmits its fusion result to the base station. The base station sets the fusion result as a potential voting result and the number of agreeing votes for the fusion result is set to be zero.
Step 3:
The base station polls the nodes, with the voting result, one by one in the witness set, following the order of the witness nodes. The witness node compares its fusion result with the voting result.
- •
If the witness node agrees with the voting result, it sends an agreeing vote to the base station. The base station increases the number of agreeing votes for the voting result by one.
- •
If the witness node does not agree with the voting result, it transmits its fusion result to the base station.
- –
If the fusion result has been stored in the base station, then the base station increases the number of agreeing votes for the fusion result by one.
- –
If the fusion result has not been stored in the base station, then the base station stores the fusion result and the number of agreeing votes for the fusion result is set to be zero.
The base station sets the voting result to the received fusion result with the maximum number of agreeing votes to poll the next witness node. If two or more fusion results received the maximum umber of votes, then the voting result is set to the result with the most recent vote. The polling stops when any received fusion result receives $`T`$ votes or when the number of unpolled nodes plus the maximum number of votes for the results recorded at the base station is less than $`T`$.
## IV Performance Evaluation
In this section numerical and computer simulations are performed to evaluate the performance of the proposed schemes. The performances of the proposed variant-round scheme are numerically calculated by the results given in Section [III](#S3 "III Improved Voting Mechanism ‣ Power-Efficient Direct-Voting Assurance for Data Fusion in Wireless Sensor Networks The preliminary result of this work is presented at the IEEE International Conference on Sensor Networks, Ubiquitous, and Trustworthy Computing (SUTC2006), Taiwan, R.O.C.. This work was supported by the National Science Council of Taiwan, R.O.C., under grants NSC 94-2213-E-305-002 and NSC 94-2213-E-305-001.") when $`P_{f} = 0`$ and $`1`$. The performances of the variant-round scheme when $`P_{f} \neq 0`$ or $`1`$ and that of the one-round scheme are evaluated using Monte Carlo computer simulations. The proposed schemes are compared using the witness-based approach in terms of overhead, average round delay, and average polling delay performances. For the witness-based approach, the size of each MAC, $`k_{w}`$, is assumed to be four bits. In the evaluation of the overhead of the variant-round scheme, the size of fusion result is zero (i.e., $`K = 0`$), such that the witness-based approach has the best performance. As stated at the end of Subsection [](#S3.SS2 "III-B Analysis of the Variant-round Scheme ‣ III Improved Voting Mechanism ‣ Power-Efficient Direct-Voting Assurance for Data Fusion in Wireless Sensor Networks The preliminary result of this work is presented at the IEEE International Conference on Sensor Networks, Ubiquitous, and Trustworthy Computing (SUTC2006), Taiwan, R.O.C.. This work was supported by the National Science Council of Taiwan, R.O.C., under grants NSC 94-2213-E-305-002 and NSC 94-2213-E-305-001.")
III-B
, the overhead of the variant-round scheme is independent of the size of the fusion result. Under this assumption, the variant-round scheme is inferred to have better overhead performance than the witness-based approach for all sizes of the fusion result whenever it outperforms with a zero-sized the fusion result. All results are presented for number of nodes $`M = 11`$.
<sup>14</sup>
<sup>14</sup>14Similar results can be obtained for $`M = 21`$, but omitted because of the page limit.
Figure [2](#S4.F2 "Figure 2 ‣ IV Performance Evaluation ‣ Power-Efficient Direct-Voting Assurance for Data Fusion in Wireless Sensor Networks The preliminary result of this work is presented at the IEEE International Conference on Sensor Networks, Ubiquitous, and Trustworthy Computing (SUTC2006), Taiwan, R.O.C.. This work was supported by the National Science Council of Taiwan, R.O.C., under grants NSC 94-2213-E-305-002 and NSC 94-2213-E-305-001.") shows the overheads of the variant-round scheme when $`P_{f} = 0`$. As in the example in Section [1](#S2.F1 "Figure 1 ‣ II Data Fusion Assurance Problem and The Previous Work ‣ Power-Efficient Direct-Voting Assurance for Data Fusion in Wireless Sensor Networks The preliminary result of this work is presented at the IEEE International Conference on Sensor Networks, Ubiquitous, and Trustworthy Computing (SUTC2006), Taiwan, R.O.C.. This work was supported by the National Science Council of Taiwan, R.O.C., under grants NSC 94-2213-E-305-002 and NSC 94-2213-E-305-001."), $`{k,k'} = 4`$ is set first and then set $`{k = 1},{k' = 0}`$ and $`{k = 0},{k' = 1}`$ are considered. This figure indicates that setting $`k = 1`$ and $`k' = 0`$ significantly reduces the overhead of the proposed scheme, independently of the fusion result at the base station. Setting $`k = 0`$ and $`k' = 1`$ further reduces the overhead of the proposed scheme. Next, the variant-round (VR) scheme is compared with the witness-based approach. In the variant-round scheme, $`k = 0`$ and $`k' = 1`$ are set. Figure [3](#S4.F3 "Figure 3 ‣ IV Performance Evaluation ‣ Power-Efficient Direct-Voting Assurance for Data Fusion in Wireless Sensor Networks The preliminary result of this work is presented at the IEEE International Conference on Sensor Networks, Ubiquitous, and Trustworthy Computing (SUTC2006), Taiwan, R.O.C.. This work was supported by the National Science Council of Taiwan, R.O.C., under grants NSC 94-2213-E-305-002 and NSC 94-2213-E-305-001.") compares overheads. The variant-round scheme significantly outperforms the witness-based approach regardless of the fusion result at the base station. For example, according to Fig. [3](#S4.F3 "Figure 3 ‣ IV Performance Evaluation ‣ Power-Efficient Direct-Voting Assurance for Data Fusion in Wireless Sensor Networks The preliminary result of this work is presented at the IEEE International Conference on Sensor Networks, Ubiquitous, and Trustworthy Computing (SUTC2006), Taiwan, R.O.C.. This work was supported by the National Science Council of Taiwan, R.O.C., under grants NSC 94-2213-E-305-002 and NSC 94-2213-E-305-001."), when $`T = 5`$ and $`M = 11`$ the variant-round scheme is almost $`40`$ times better than the witness-based approach given in \[[9](#bib.bib9)\] in terms of overhead performance for $`C = 1`$ and $`2.`$ Figure [4](#S4.F4 "Figure 4 ‣ IV Performance Evaluation ‣ Power-Efficient Direct-Voting Assurance for Data Fusion in Wireless Sensor Networks The preliminary result of this work is presented at the IEEE International Conference on Sensor Networks, Ubiquitous, and Trustworthy Computing (SUTC2006), Taiwan, R.O.C.. This work was supported by the National Science Council of Taiwan, R.O.C., under grants NSC 94-2213-E-305-002 and NSC 94-2213-E-305-001.") compares the average round delays of the proposed variant-round scheme and the witness-based approach. This figure demonstrates that the average round delays of the proposed scheme are smaller than those of the witness-based approach when the base station can obtain the valid fusion results; however, they perform equally when the base station obtains invalid fusion results. Figure [5](#S4.F5 "Figure 5 ‣ IV Performance Evaluation ‣ Power-Efficient Direct-Voting Assurance for Data Fusion in Wireless Sensor Networks The preliminary result of this work is presented at the IEEE International Conference on Sensor Networks, Ubiquitous, and Trustworthy Computing (SUTC2006), Taiwan, R.O.C.. This work was supported by the National Science Council of Taiwan, R.O.C., under grants NSC 94-2213-E-305-002 and NSC 94-2213-E-305-001.") compares the average polling delays of the proposed variant-round scheme and the witness-based approach. In Fig. [5](#S4.F5 "Figure 5 ‣ IV Performance Evaluation ‣ Power-Efficient Direct-Voting Assurance for Data Fusion in Wireless Sensor Networks The preliminary result of this work is presented at the IEEE International Conference on Sensor Networks, Ubiquitous, and Trustworthy Computing (SUTC2006), Taiwan, R.O.C.. This work was supported by the National Science Council of Taiwan, R.O.C., under grants NSC 94-2213-E-305-002 and NSC 94-2213-E-305-001."), the proposed scheme has much smaller average polling delays than the witness-based approach - unlike average round delay performance. For example, when $`T = 10`$ and $`M = 11`$, the proposed variant-round scheme is almost five times better than the witness-based approach in terms of average polling delay performance for $`C = 4`$ and $`5.`$ Accordingly, the proposed scheme outperforms the witness-based approach given in \[[9](#bib.bib9)\] in terms of overhead and delay.
[Refer to caption](/html/0705.3683/assets/x2.png)
Figure 2: The overheads of the variant-round scheme, for $`M = 11`$ and $`P_{f} = 0`$, (a) when $`T = 5`$ (valid fusion result) and (b) when $`T = 10`$ (invalid fusion result).
[Refer to caption](/html/0705.3683/assets/x3.png)
Figure 3: Overhead comparison between the variant-round (VR) scheme and the witness-based approach given in \[[9](#bib.bib9)\], for $`M = 11`$ and $`P_{f} = {1,0}`$, (a) when $`T = 5`$ (valid fusion result) and (b) when $`T = 10`$ (invalid fusion result).
[Refer to caption](/html/0705.3683/assets/x4.png)
Figure 4: Average round delay comparison between the variant-round (VR) scheme and the witness-based approach given in \[[9](#bib.bib9)\], for $`M = 11`$ and $`P_{f} = {1,0}`$, (a) when $`T = 5`$ (valid fusion result) and (b) when $`T = 10`$ (invalid fusion result).
[Refer to caption](/html/0705.3683/assets/x5.png)
Figure 5: Average polling delay comparison between the variant-round (VR) scheme and the witness-based approach given in \[[9](#bib.bib9)\], for $`M = 11`$ and $`P_{f} = {1,0}`$, (a) when $`T = 5`$ (valid fusion result) and (b) when $`T = 10`$ (invalid fusion result).
The following computer simulations evaluate the variant-round scheme when $`{P_{f} = {0.25,0.5}},`$ and $`0.75`$ by performing $`10000`$ Monte Carlo tests for each simulation. In the first set of simulations, the witness-based approach given in \[[9](#bib.bib9)\] is compared with the proposed variant-round scheme when $`{P_{f} = {0.25,0.5}},`$ and $`0.75`$. Figure [6](#S4.F6 "Figure 6 ‣ IV Performance Evaluation ‣ Power-Efficient Direct-Voting Assurance for Data Fusion in Wireless Sensor Networks The preliminary result of this work is presented at the IEEE International Conference on Sensor Networks, Ubiquitous, and Trustworthy Computing (SUTC2006), Taiwan, R.O.C.. This work was supported by the National Science Council of Taiwan, R.O.C., under grants NSC 94-2213-E-305-002 and NSC 94-2213-E-305-001.") presents the results for $`M = 11`$, $`k = 0`$ and $`k' = 1`$. The variant-round scheme outperforms the witness-based approach in every $`P_{f}`$ simulated. For example, in Fig. [6](#S4.F6 "Figure 6 ‣ IV Performance Evaluation ‣ Power-Efficient Direct-Voting Assurance for Data Fusion in Wireless Sensor Networks The preliminary result of this work is presented at the IEEE International Conference on Sensor Networks, Ubiquitous, and Trustworthy Computing (SUTC2006), Taiwan, R.O.C.. This work was supported by the National Science Council of Taiwan, R.O.C., under grants NSC 94-2213-E-305-002 and NSC 94-2213-E-305-001."), when $`T = 5`$ the proposed variant-round scheme is almost $`40`$ times better than the witness-based approach given in \[[9](#bib.bib9)\] in terms of overhead performance for $`C = {1,2}`$ and $`P_{f} = {0.25,0.5,0.75}`$. Finally, Table [I](#S4.T1 "TABLE I ‣ IV Performance Evaluation ‣ Power-Efficient Direct-Voting Assurance for Data Fusion in Wireless Sensor Networks The preliminary result of this work is presented at the IEEE International Conference on Sensor Networks, Ubiquitous, and Trustworthy Computing (SUTC2006), Taiwan, R.O.C.. This work was supported by the National Science Council of Taiwan, R.O.C., under grants NSC 94-2213-E-305-002 and NSC 94-2213-E-305-001.") summarizes the maximum average overhead of the variant-round scheme and the witness-based approach for $`M = 11`$ and $`12`$.
TABLE I: Maximum average overheads (bits) for variant-round (VR) scheme and witness-based approach given in \[[9](#bib.bib9)\]
<table>
<thead>
<tr>
<th></th>
<th>Witness-based</th>
<th>VR, Pf=0subscript𝑃𝑓0P_{f}=0</th>
<th>VR, Pf=0.25subscript𝑃𝑓0.25P_{f}=0.25</th>
<th>VR, Pf=0.5subscript𝑃𝑓0.5P_{f}=0.5</th>
<th>VR, Pf=0.75subscript𝑃𝑓0.75P_{f}=0.75</th>
<th>VR, Pf=1subscript𝑃𝑓1P_{f}=1</th>
</tr>
</thead>
<tbody>
<tr>
<th>M=11𝑀11M=11</th>
<th>109</th>
<td>2.5</td>
<td>2.5</td>
<td>2.4</td>
<td>2.7</td>
<td>3.2</td>
</tr>
<tr>
<th>M=21𝑀21M=21</th>
<th>314</th>
<td>4.9</td>
<td>4.7</td>
<td>4.7</td>
<td>4.9</td>
<td>6.2</td>
</tr>
</tbody>
</table>
In the second set of simulations, the average numbers of bits sent by uncompromised nodes in the one-round scheme is evaluated. When the compromised node does not agree with the voting result, the fusion result transmitted by the compromised node is different from other fusion results and the size of the fusion result is $`K = 48`$. Fig. [7](#S4.F7 "Figure 7 ‣ IV Performance Evaluation ‣ Power-Efficient Direct-Voting Assurance for Data Fusion in Wireless Sensor Networks The preliminary result of this work is presented at the IEEE International Conference on Sensor Networks, Ubiquitous, and Trustworthy Computing (SUTC2006), Taiwan, R.O.C.. This work was supported by the National Science Council of Taiwan, R.O.C., under grants NSC 94-2213-E-305-002 and NSC 94-2213-E-305-001.") shows the results for $`M = 11`$ when $`P_{f} = {0,0.5}`$, and $`1`$. In Fig. [7](#S4.F7 "Figure 7 ‣ IV Performance Evaluation ‣ Power-Efficient Direct-Voting Assurance for Data Fusion in Wireless Sensor Networks The preliminary result of this work is presented at the IEEE International Conference on Sensor Networks, Ubiquitous, and Trustworthy Computing (SUTC2006), Taiwan, R.O.C.. This work was supported by the National Science Council of Taiwan, R.O.C., under grants NSC 94-2213-E-305-002 and NSC 94-2213-E-305-001.")(a), when the base station can obtain a valid fusion result, the bits transmitted by the uncompromised nodes to the base station in the one-round scheme increase with the number of compromised nodes $`C`$, as expected. However, they are smaller than those in the witness-based approach. Notably, in this case, the bits transmitted by uncompromised nodes in the witness-based approach is constant, since once an uncompromised node is polled, the polling process is completed. For small $`C`$, such as $`C = {0,1,2}`$ or $`3`$, the number of bits transmitted by uncompromised nodes in the one-round scheme is about half that of those in the witness-based approach. Additionally, the performance of the one-round scheme when $`P_{f} = 1`$ is the worst in all three simulations, since any compromised node that agrees with the forged result will sometimes make the forged result with the largest number of votes and force the base station to use it as the temporary voting result to poll the next node. Then, the next uncompromised node needs to transmit its fusion result to the base station instead of only sending a agreeing vote and the total number of bits transmitted by the uncompromised nodes increases.
According to Fig. [7](#S4.F7 "Figure 7 ‣ IV Performance Evaluation ‣ Power-Efficient Direct-Voting Assurance for Data Fusion in Wireless Sensor Networks The preliminary result of this work is presented at the IEEE International Conference on Sensor Networks, Ubiquitous, and Trustworthy Computing (SUTC2006), Taiwan, R.O.C.. This work was supported by the National Science Council of Taiwan, R.O.C., under grants NSC 94-2213-E-305-002 and NSC 94-2213-E-305-001.")(b), when the base station cannot obtain a valid fusion result, the number of bits transmitted by the uncompromised nodes to the base station in the one-round scheme decreases with the number of compromised nodes $`C`$, except for $`P_{f} = 1`$. This phenomenon is caused by the fact that the scheme stops when the number of unpolled nodes plus the maximum number of votes for any result recorded at the base station is less than $`T`$. When $`T = {M - 1}`$ as simulated, the recoding of two different results at the base station stops the polling process. Recall that when $`P_{f} = 1`$, the only way to stop the polling process is for one fusion result to be sent by an uncompromised node and the other to be sent by a compromised node, and for the transmitted bits of the uncompromised nodes to be the same for all $`C`$. If $`P_{f} \neq 1`$, the two compromised nodes may yield two different results, and no bit is transmitted by the uncompromised node. This concludes the simulation results. This subfigure reveals that the one-round scheme outperforms when $`C`$ is small but loses when $`C`$ exceeds $`T/2`$. Similarly, the maximum average number of bits in the one-round scheme is compared with that of the witness-based approach for $`M = 11`$ and $`12`$, and the results are given in Table [II](#S4.T2 "TABLE II ‣ IV Performance Evaluation ‣ Power-Efficient Direct-Voting Assurance for Data Fusion in Wireless Sensor Networks The preliminary result of this work is presented at the IEEE International Conference on Sensor Networks, Ubiquitous, and Trustworthy Computing (SUTC2006), Taiwan, R.O.C.. This work was supported by the National Science Council of Taiwan, R.O.C., under grants NSC 94-2213-E-305-002 and NSC 94-2213-E-305-001."). This table indicates that the maximum average number of bits transmitted in the one-round scheme is much lower than that in the witness-based approach.
TABLE II: Maximum average overheads (bits) for one-round (OR) scheme and witness-based approach given in \[[9](#bib.bib9)\]
<table>
<thead>
<tr>
<th></th>
<th>Witness-based</th>
<th>OR, Pf=0subscript𝑃𝑓0P_{f}=0</th>
<th>OR, Pf=0.5subscript𝑃𝑓0.5P_{f}=0.5</th>
<th>OR, Pf=1subscript𝑃𝑓1P_{f}=1</th>
</tr>
</thead>
<tbody>
<tr>
<th>M=11𝑀11M=11</th>
<th>240</th>
<td>71.5</td>
<td>86.7</td>
<td>128.6</td>
</tr>
<tr>
<th>M=21𝑀21M=21</th>
<th>566</th>
<td>76.1</td>
<td>106.4</td>
<td>205.7</td>
</tr>
</tbody>
</table>
[Refer to caption](/html/0705.3683/assets/x6.png)
Figure 6: Overhead comparison between the variant-round (VR) scheme and the witness-based approach given in \[[9](#bib.bib9)\], for $`M = 11`$ and $`P_{f} = {0.25,0.5,0.75}`$, (a) when $`T = 5`$ (valid fusion result) and (b) when $`T = 10`$ (invalid fusion result).
[Refer to caption](/html/0705.3683/assets/x7.png)
Figure 7: Overhead comparison between the one-round (OR) scheme and the witness-based approach given in \[[9](#bib.bib9)\], for $`M = 11`$ and $`P_{f} = {0,0.5,1}`$, (a) when $`T = 5`$ (valid fusion result) and (b) when $`T = 10`$ (invalid fusion result).
## V Conclusions and Future Work
This work proposes a power-efficient scheme for data fusion assurance, in which the base station in the wireless sensor network collects the fusion data and the votes on the data directly from the fusion nodes. The proposed scheme is more reliable with less assurance overhead and delay than the witness-based approach. That is, the power and delay for the transmission of the fusion result and the votes are significantly decreased.
In the future, we will discover the performance when a node is compromised with some probability, both statistically dependent and independent to other nodes. Moreover, the propose scheme cannot be applied to multi-hop WSNs. We will develop other schemes based on the direct voting mechanism for the multi-hop WSN.
Performance Analysis of the Variant-round Scheme when $`P_{f} = 1`$
In this analysis we assume that $`{{2\hspace{0pt}T} + 1} \geq M`$. Since when $`C > T`$ the base station will get a forged fusion result and this should be avoided, we are only considering two cases:
Case 1
$`{M - T} \leq C \leq T`$,
Case 2
$`C < {M - T}`$.
Case $`1`$ does not produce a valid fusion result. Assume that the chosen node at the first round of polling is compromised. The probability that the chosen node is compromised at the first round is given by $`C/M`$. The first-round polling-for-vote process finishes when $`M - T`$ witness nodes do not agree with the transmitted fusion result, as described in Step 3. When the polling stops at the $`i`$th witness node, the $`i`$th witness node (uncompromised) does not agree with the transmitted result and the base station has polled $`M - T`$ disagreeing nodes (including the $`i`$th witness node). Moreover, $`M - i - 1`$ nodes are unpolled, and $`{M - C - {({M - T})}} = {T - C}`$ of them are uncompromised. Since the witness set contains $`C - 1`$ compromised nodes, the maximum number of polled witness nodes is $`{{({C - 1})} + {({M - T})}} = {{{M - T} + C} - 1}`$. Hence, the probability that the polling stops at the $`i`$th witness node, where $`{M - T} \leq i \leq {{{M - T} + C} - 1}`$, is given by
<table>
<tbody>
<tr>
<td></td>
<td>Pv1c1(i)=1Πv1c1(i−1M−T−1)(M−i−1T−C).superscriptsubscript𝑃𝑣1𝑐1𝑖1superscriptsubscriptΠ𝑣1𝑐1𝑖1𝑀𝑇1𝑀𝑖1𝑇𝐶P_{v1}^{c1}(i)=\frac{1}{\Pi_{v1}^{c1}}\left(\begin{array}[]{c}i-1\\ M-T-1\end{array}\right)\left(\begin{array}[]{c}M-i-1\\ T-C\end{array}\right).</td>
<td></td>
</tr>
</tbody>
</table>
The number of agreeing nodes, $`A`$, equals to $`i - {({M - T})}`$. The first witness node that disagreed with the transmitted fusion result at the first round becomes the new chosen node at the second round as stated in Step 4. The size of the witness set becomes $`M' = {{({M - 1})} - {({{i - M} + T})} - 1} = {{2\hspace{0pt}M} - T - i - 2}`$. If $`M' < T`$, which is equivalent to $`{{({M - 1})} - T - 1} < A`$, then the polling stops as described in the second part of Step 4. Otherwise, since the previous chosen fusion node is compromised, the current chosen fusion node is not compromised. At the second round, the witness set now has $`C' = {C - 1 - {({i - {({M - T})}})}} = {{M + C} - T - i - 1}`$ compromised nodes and $`M - C - 1`$ uncompromised nodes. As implied in Step 3, the base station first polls the $`M - T - 1`$ witness nodes that disagreed with the previous transmitted fusion result in the first round. The witness set contains $`M - i - 1`$ unpolled nodes of which $`{M - C - {({M - T})}} = {T - C}`$ are uncompromised. Consequently, the number of the possible polling orders is given by
<table>
<tbody>
<tr>
<td></td>
<td>Πv1c2=(M−i−1T−C).superscriptsubscriptΠ𝑣1𝑐2𝑀𝑖1𝑇𝐶\Pi_{v1}^{c2}=\left(\begin{array}[]{c}M-i-1\\ T-C\end{array}\right).</td>
<td></td>
</tr>
</tbody>
</table>
When the second round of polling stops at the $`j`$th witness node, the $`j`$th witness node (which is a compromised node) does not agree with the transmitted result. The base station has polled $`{M' - T} + 1`$ disagreeing nodes (including the $`j`$th witness node) and $`{j - {({{M' - T} + 1})}} = {{{j - M'} + T} - 1}`$ agreeing (uncompromised) nodes. The first $`M - T - 1`$ of the $`j`$ polled nodes are uncompromised, and the rest $`j - {({M - T - 1})} - 1`$ nodes (including the $`j`$th witness node) may include the uncompromised or compromised nodes. Since the $`j`$ polled nodes include $`{M' - T} + 1`$ disagreeing (compromised) nodes, the $`M' - j`$ unpolled nodes include $`C' - {({{M' - T} + 1})}`$ compromised nodes. The probability that the second polling process stops at the $`j`$th witness node, where $`j \geq {{({{M' - T} + 1})} + {({M - T - 1})}} = {{M' + M} - {2\hspace{0pt}T}}`$ and $`j \leq {{({{M' - T} + 1})} + {({M - C - 1})}} = {{{M - C} + M'} - T}`$ (disagreeing nodes and uncompromised nodes), is given by
<table>
<tbody>
<tr>
<td></td>
<td>Pv1c2(j)superscriptsubscript𝑃𝑣1𝑐2𝑗\displaystyle P_{v1}^{c2}(j)</td>
<td>=\displaystyle=</td>
<td>1Πv1c2(j−(M−T−1)−1M′−T)(M′−jC′−(M′−T+1))1superscriptsubscriptΠ𝑣1𝑐2𝑗𝑀𝑇11superscript𝑀′𝑇superscript𝑀′𝑗superscript𝐶′superscript𝑀′𝑇1\displaystyle\frac{1}{\Pi_{v1}^{c2}}\left(\begin{array}[]{c}j-(M-T-1)-1\\ M^{\prime}-T\end{array}\right)\left(\begin{array}[]{c}M^{\prime}-j\\ C^{\prime}-(M^{\prime}-T+1)\end{array}\right)</td>
<td></td>
</tr>
<tr>
<td></td>
<td></td>
<td>=\displaystyle=</td>
<td>1Πv1c2(j−M+T2M−2T−i−2)(2M−T−i−j−2T+C−M).1superscriptsubscriptΠ𝑣1𝑐2𝑗𝑀𝑇2𝑀2𝑇𝑖22𝑀𝑇𝑖𝑗2𝑇𝐶𝑀\displaystyle\frac{1}{\Pi_{v1}^{c2}}\left(\begin{array}[]{c}j-M+T\\ 2M-2T-i-2\end{array}\right)\left(\begin{array}[]{c}2M-T-i-j-2\\ T+C-M\end{array}\right).</td>
<td></td>
</tr>
</tbody>
</table>
Since $`A`$ is at least $`M - T - 1`$ at the second round, it is easy to see that $`{M' - T - 1} < A`$ and the scheme will stop after the second round. Therefore, the average overhead is
<table>
<tbody>
<tr>
<td></td>
<td>Ov1c(M,T,C,1)=∑i=M−TM−T+C−1Pv1c1(i)[(M−T)k′+∑j=M′+M−2TM−C+M′−TPv1c2(j)(j−M′+T−1)k].superscriptsubscript𝑂𝑣1𝑐𝑀𝑇𝐶1superscriptsubscript𝑖𝑀𝑇𝑀𝑇𝐶1superscriptsubscript𝑃𝑣1𝑐1𝑖delimited-[]𝑀𝑇superscript𝑘′superscriptsubscript𝑗superscript𝑀′𝑀2𝑇𝑀𝐶superscript𝑀′𝑇superscriptsubscript𝑃𝑣1𝑐2𝑗𝑗superscript𝑀′𝑇1𝑘O_{v1}^{c}(M,T,C,1)=\sum_{i=M-T}^{M-T+C-1}P_{v1}^{c1}(i)\left[(M-T)k^{\prime}+\sum_{j=M^{\prime}+M-2T}^{M-C+M^{\prime}-T}P_{v1}^{c2}(j)(j-M^{\prime}+T-1)k\right].</td>
<td></td>
<td>(19)</td>
</tr>
</tbody>
</table>
As mentioned before, when $`M' < T`$, i.e., $`{{2\hspace{0pt}M} - {2\hspace{0pt}T} - 2} < i`$, the scheme will stop after the first round. Hence, the average round delay is
<table>
<tbody>
<tr>
<td></td>
<td>Rv1c(M,T,C,1)=∑i=M−T2M−2T−22Pv1c1(i)+∑i=2M−2T−1M−T+C−1Pv1c1(i),superscriptsubscript𝑅𝑣1𝑐𝑀𝑇𝐶1superscriptsubscript𝑖𝑀𝑇2𝑀2𝑇22superscriptsubscript𝑃𝑣1𝑐1𝑖superscriptsubscript𝑖2𝑀2𝑇1𝑀𝑇𝐶1superscriptsubscript𝑃𝑣1𝑐1𝑖R_{v1}^{c}(M,T,C,1)=\sum_{i=M-T}^{2M-2T-2}2P_{v1}^{c1}(i)+\sum_{i=2M-2T-1}^{M-T+C-1}P_{v1}^{c1}(i),</td>
<td></td>
<td>(20)</td>
</tr>
</tbody>
</table>
and the average polling delay is
<table>
<tbody>
<tr>
<td></td>
<td>Dv1c(M,T,C,1)=∑i=M−T2M−2T−2Pv1c1(i)[i+∑j=M′+M−2TM−C+M′−TjPv1c2(j)]+∑i=2M−2T−1M−T+C−1iPv1c1(i).superscriptsubscript𝐷𝑣1𝑐𝑀𝑇𝐶1superscriptsubscript𝑖𝑀𝑇2𝑀2𝑇2superscriptsubscript𝑃𝑣1𝑐1𝑖delimited-[]𝑖superscriptsubscript𝑗superscript𝑀′𝑀2𝑇𝑀𝐶superscript𝑀′𝑇𝑗superscriptsubscript𝑃𝑣1𝑐2𝑗superscriptsubscript𝑖2𝑀2𝑇1𝑀𝑇𝐶1𝑖superscriptsubscript𝑃𝑣1𝑐1𝑖D_{v1}^{c}(M,T,C,1)=\sum_{i=M-T}^{2M-2T-2}P_{v1}^{c1}(i)\left[i+\sum_{j=M^{\prime}+M-2T}^{M-C+M^{\prime}-T}jP_{v1}^{c2}(j)\right]+\sum_{i=2M-2T-1}^{M-T+C-1}iP_{v1}^{c1}(i).</td>
<td></td>
<td>(21)</td>
</tr>
</tbody>
</table>
In the first case, the probability that the chosen node is not compromised at the first round is given by $`{({M - C})}/M`$. The number of the possible polling orders is written as
<table>
<tbody>
<tr>
<td></td>
<td>Πv1u1=(M−1C).superscriptsubscriptΠ𝑣1𝑢1𝑀1𝐶\Pi_{v1}^{u1}=\left(\begin{array}[]{c}M-1\\ C\end{array}\right).</td>
<td></td>
</tr>
</tbody>
</table>
Thus, the probability that the polling stops at the $`i`$th witness node, where $`{M - T} \leq i \leq {{({M - C - 1})} + {({M - T})}} = {{2\hspace{0pt}M} - T - C - 1}`$, is expressed as
<table>
<tbody>
<tr>
<td></td>
<td>Pv1u1(i)=1Πv1u1(i−1M−T−1)(M−i−1T+C−M).superscriptsubscript𝑃𝑣1𝑢1𝑖1superscriptsubscriptΠ𝑣1𝑢1𝑖1𝑀𝑇1𝑀𝑖1𝑇𝐶𝑀P_{v1}^{u1}(i)=\frac{1}{\Pi_{v1}^{u1}}\left(\begin{array}[]{c}i-1\\ M-T-1\end{array}\right)\left(\begin{array}[]{c}M-i-1\\ T+C-M\end{array}\right).</td>
<td></td>
</tr>
</tbody>
</table>
At the second round the base station chooses the first witness nodes that disagreed with the transmitted fusion result sent in the previous round. The size of the witness set becomes $`M' = {{2\hspace{0pt}M} - T - i - 2}`$. The witness set has $`{M - C - {({i - {({M - T})}})} - 1} = {{2\hspace{0pt}M} - C - T - i - 1}`$ uncompromised nodes. Similarly, the polling stops if $`M' < T`$. Otherwise, since the previous chosen fusion node is not compromised, the current chosen fusion node is compromised. Thus, the number of the possible polling orders is given by
<table>
<tbody>
<tr>
<td></td>
<td>Πv1u2=(M−i−1T+C−M).superscriptsubscriptΠ𝑣1𝑢2𝑀𝑖1𝑇𝐶𝑀\Pi_{v1}^{u2}=\left(\begin{array}[]{c}M-i-1\\ T+C-M\end{array}\right).</td>
<td></td>
</tr>
</tbody>
</table>
When the second round of polling stops at the $`j`$th witness node, the $`j`$ polled nodes contain $`{M' - T} + 1`$ disagreeing (uncompromised) nodes and the unpolled nodes include $`{2\hspace{0pt}M} - C - T - i - 1 - {({{M' - T} + 1})}`$ uncompromised nodes. The probability that the second round of polling finishes at the $`j`$th witness node, where $`{{({{M' - T} + 1})} + {({C - 1})}} = {{M' - T} + C} \geq j \geq {{({{M' - T} + 1})} + {({M - T - 1})}} = {{M' + M} - {2\hspace{0pt}T}}`$, is given by
<table>
<tbody>
<tr>
<td></td>
<td>Pv1u2(j)superscriptsubscript𝑃𝑣1𝑢2𝑗\displaystyle P_{v1}^{u2}(j)</td>
<td>=\displaystyle=</td>
<td>1Πv1u2(j−(M−T−1)−1M′−T)(M′−j2M−C−T−i−1−(M′−T+1))1superscriptsubscriptΠ𝑣1𝑢2𝑗𝑀𝑇11superscript𝑀′𝑇superscript𝑀′𝑗2𝑀𝐶𝑇𝑖1superscript𝑀′𝑇1\displaystyle\frac{1}{\Pi_{v1}^{u2}}\left(\begin{array}[]{c}j-(M-T-1)-1\\ M^{\prime}-T\end{array}\right)\left(\begin{array}[]{c}M^{\prime}-j\\ 2M-C-T-i-1-(M^{\prime}-T+1)\end{array}\right)</td>
<td></td>
</tr>
<tr>
<td></td>
<td></td>
<td>=\displaystyle=</td>
<td>1Πv1u2(j−M+T2M−2T−i−2)(2M−T−i−j−2T−C).1superscriptsubscriptΠ𝑣1𝑢2𝑗𝑀𝑇2𝑀2𝑇𝑖22𝑀𝑇𝑖𝑗2𝑇𝐶\displaystyle\frac{1}{\Pi_{v1}^{u2}}\left(\begin{array}[]{c}j-M+T\\ 2M-2T-i-2\end{array}\right)\left(\begin{array}[]{c}2M-T-i-j-2\\ T-C\end{array}\right).</td>
<td></td>
</tr>
</tbody>
</table>
Similarly, it is easy to see that the scheme will stop after the second round. Therefore, the average overhead is given by
<table>
<tbody>
<tr>
<td></td>
<td>Ov1u(M,T,C,1)=∑i=M−T2M−T−C−1Pv1u1(i)[(i−M+T)k+∑j=M′+M−2TM′−T+CPv1u2(j)(M′−T+1)k′].superscriptsubscript𝑂𝑣1𝑢𝑀𝑇𝐶1superscriptsubscript𝑖𝑀𝑇2𝑀𝑇𝐶1superscriptsubscript𝑃𝑣1𝑢1𝑖delimited-[]𝑖𝑀𝑇𝑘superscriptsubscript𝑗superscript𝑀′𝑀2𝑇superscript𝑀′𝑇𝐶superscriptsubscript𝑃𝑣1𝑢2𝑗superscript𝑀′𝑇1superscript𝑘′O_{v1}^{u}(M,T,C,1)=\sum_{i=M-T}^{2M-T-C-1}P_{v1}^{u1}(i)\left[(i-M+T)k+\sum_{j=M^{\prime}+M-2T}^{M^{\prime}-T+C}P_{v1}^{u2}(j)(M^{\prime}-T+1)k^{\prime}\right].</td>
<td></td>
<td>(24)</td>
</tr>
</tbody>
</table>
Again, when $`M' < T`$, i.e., $`{{2\hspace{0pt}M} - {2\hspace{0pt}T} - 2} < i`$, the scheme will stop after the first round. Hence, the average round delay is
<table>
<tbody>
<tr>
<td></td>
<td>Rv1u(M,T,C,1)=∑i=M−T2M−2T−22Pv1u1(i)+∑i=2M−2T−12M−T−C−1Pv1u1(i),superscriptsubscript𝑅𝑣1𝑢𝑀𝑇𝐶1superscriptsubscript𝑖𝑀𝑇2𝑀2𝑇22superscriptsubscript𝑃𝑣1𝑢1𝑖superscriptsubscript𝑖2𝑀2𝑇12𝑀𝑇𝐶1superscriptsubscript𝑃𝑣1𝑢1𝑖R_{v1}^{u}(M,T,C,1)=\sum_{i=M-T}^{2M-2T-2}2P_{v1}^{u1}(i)+\sum_{i=2M-2T-1}^{2M-T-C-1}P_{v1}^{u1}(i),</td>
<td></td>
<td>(25)</td>
</tr>
</tbody>
</table>
and the average polling delay is
<table>
<tbody>
<tr>
<td></td>
<td>Dv1u(M,T,C,1)=∑i=M−T2M−2T−2Pv1u1(i)[i+∑j=M′+M−2TM′−T+CjPv1u2(j)]+∑i=2M−2T−1M−T+C−1iPv1u1(i).superscriptsubscript𝐷𝑣1𝑢𝑀𝑇𝐶1superscriptsubscript𝑖𝑀𝑇2𝑀2𝑇2superscriptsubscript𝑃𝑣1𝑢1𝑖delimited-[]𝑖superscriptsubscript𝑗superscript𝑀′𝑀2𝑇superscript𝑀′𝑇𝐶𝑗superscriptsubscript𝑃𝑣1𝑢2𝑗superscriptsubscript𝑖2𝑀2𝑇1𝑀𝑇𝐶1𝑖superscriptsubscript𝑃𝑣1𝑢1𝑖D_{v1}^{u}(M,T,C,1)=\sum_{i=M-T}^{2M-2T-2}P_{v1}^{u1}(i)\left[i+\sum_{j=M^{\prime}+M-2T}^{M^{\prime}-T+C}jP_{v1}^{u2}(j)\right]+\sum_{i=2M-2T-1}^{M-T+C-1}iP_{v1}^{u1}(i).</td>
<td></td>
<td>(26)</td>
</tr>
</tbody>
</table>
From ([19](#A0.E19 "In V Conclusions and Future Work ‣ Power-Efficient Direct-Voting Assurance for Data Fusion in Wireless Sensor Networks The preliminary result of this work is presented at the IEEE International Conference on Sensor Networks, Ubiquitous, and Trustworthy Computing (SUTC2006), Taiwan, R.O.C.. This work was supported by the National Science Council of Taiwan, R.O.C., under grants NSC 94-2213-E-305-002 and NSC 94-2213-E-305-001.")) and ([26](#A0.E26 "In V Conclusions and Future Work ‣ Power-Efficient Direct-Voting Assurance for Data Fusion in Wireless Sensor Networks The preliminary result of this work is presented at the IEEE International Conference on Sensor Networks, Ubiquitous, and Trustworthy Computing (SUTC2006), Taiwan, R.O.C.. This work was supported by the National Science Council of Taiwan, R.O.C., under grants NSC 94-2213-E-305-002 and NSC 94-2213-E-305-001.")), we have the average overhead and the average delays of Case 1 as
<table>
<tbody>
<tr>
<td></td>
<td>Ov1(M,T,C,1)=CMOv1c(M,T,C,1)+M−CMOv1u(M,T,C,1),subscript𝑂𝑣1𝑀𝑇𝐶1𝐶𝑀superscriptsubscript𝑂𝑣1𝑐𝑀𝑇𝐶1𝑀𝐶𝑀superscriptsubscript𝑂𝑣1𝑢𝑀𝑇𝐶1O_{v1}(M,T,C,1)=\frac{C}{M}O_{v1}^{c}(M,T,C,1)+\frac{M-C}{M}O_{v1}^{u}(M,T,C,1),</td>
<td></td>
<td>(27)</td>
</tr>
</tbody>
</table>
<table>
<tbody>
<tr>
<td></td>
<td>Rv1(M,T,C,1)=CMRv1c(M,T,C,1)+M−CMRv1u(M,T,C,1),subscript𝑅𝑣1𝑀𝑇𝐶1𝐶𝑀superscriptsubscript𝑅𝑣1𝑐𝑀𝑇𝐶1𝑀𝐶𝑀superscriptsubscript𝑅𝑣1𝑢𝑀𝑇𝐶1R_{v1}(M,T,C,1)=\frac{C}{M}R_{v1}^{c}(M,T,C,1)+\frac{M-C}{M}R_{v1}^{u}(M,T,C,1),</td>
<td></td>
<td>(28)</td>
</tr>
</tbody>
</table>
and
<table>
<tbody>
<tr>
<td></td>
<td>Dv1(M,T,C,1)=CMDv1c(M,T,C,1)+M−CMDv1u(M,T,C,1).subscript𝐷𝑣1𝑀𝑇𝐶1𝐶𝑀superscriptsubscript𝐷𝑣1𝑐𝑀𝑇𝐶1𝑀𝐶𝑀superscriptsubscript𝐷𝑣1𝑢𝑀𝑇𝐶1D_{v1}(M,T,C,1)=\frac{C}{M}D_{v1}^{c}(M,T,C,1)+\frac{M-C}{M}D_{v1}^{u}(M,T,C,1).</td>
<td></td>
<td>(29)</td>
</tr>
</tbody>
</table>
A valid fusion result is available in the second case. The first-round process is similar to the first-round process in the first case when the chosen node is compromised at the first round. The probability that the $`T`$th witness node agrees with the transmitted fusion result at witness node $`j`$ of the second-round polling process is
<table>
<tbody>
<tr>
<td></td>
<td>Pv2c2(j)superscriptsubscript𝑃𝑣2𝑐2𝑗\displaystyle P_{v2}^{c2}(j)</td>
<td>=\displaystyle=</td>
<td>1Πv1c2(j−(M−T−1)−1T−(M−T−1)−1)(2M−T−i−j−2M−C−(T+1))1superscriptsubscriptΠ𝑣1𝑐2𝑗𝑀𝑇11𝑇𝑀𝑇112𝑀𝑇𝑖𝑗2𝑀𝐶𝑇1\displaystyle\frac{1}{\Pi_{v1}^{c2}}\left(\begin{array}[]{c}j-(M-T-1)-1\\ T-(M-T-1)-1\end{array}\right)\left(\begin{array}[]{c}2M-T-i-j-2\\ M-C-(T+1)\end{array}\right)</td>
<td></td>
</tr>
<tr>
<td></td>
<td></td>
<td>=\displaystyle=</td>
<td>1Πv1c2(j−M+T2T−M)(2M−T−i−j−2M−C−T−1).1superscriptsubscriptΠ𝑣1𝑐2𝑗𝑀𝑇2𝑇𝑀2𝑀𝑇𝑖𝑗2𝑀𝐶𝑇1\displaystyle\frac{1}{\Pi_{v1}^{c2}}\left(\begin{array}[]{c}j-M+T\\ 2T-M\end{array}\right)\left(\begin{array}[]{c}2M-T-i-j-2\\ M-C-T-1\end{array}\right).</td>
<td></td>
</tr>
</tbody>
</table>
Significantly, if $`M = {{2\hspace{0pt}T} + 1}`$, then $`j = T`$. Otherwise, $`T \leq j \leq {{C - {({i - {({M - T})}})} - 1} + T} = {{M + C} - i - 1}`$. Therefore, the average overhead is expressed as
<table>
<tbody>
<tr>
<td></td>
<td>Ov2c(M,T,C,1)={∑i=M−TM−T+C−1Pv1c1(i)[(M−T)k′+Tk]=(M−T)k′+TkM=2T+1∑i=M−TM−T+C−1Pv1c1(i)[(M−T)k′+∑j=TM+C−i−1Pv2c2(j)Tk]else.superscriptsubscript𝑂𝑣2𝑐𝑀𝑇𝐶1casessuperscriptsubscript𝑖𝑀𝑇𝑀𝑇𝐶1superscriptsubscript𝑃𝑣1𝑐1𝑖delimited-[]𝑀𝑇superscript𝑘′𝑇𝑘𝑀𝑇superscript𝑘′𝑇𝑘𝑀2𝑇1superscriptsubscript𝑖𝑀𝑇𝑀𝑇𝐶1superscriptsubscript𝑃𝑣1𝑐1𝑖delimited-[]𝑀𝑇superscript𝑘′superscriptsubscript𝑗𝑇𝑀𝐶𝑖1superscriptsubscript𝑃𝑣2𝑐2𝑗𝑇𝑘else\displaystyle O_{v2}^{c}(M,T,C,1)=\left\{\begin{array}[]{ll}\sum_{i=M-T}^{M-T+C-1}P_{v1}^{c1}(i)\left[(M-T)k^{\prime}+Tk\right]=(M-T)k^{\prime}+Tk&M=2T+1\\ \sum_{i=M-T}^{M-T+C-1}P_{v1}^{c1}(i)\left[(M-T)k^{\prime}+\sum_{j=T}^{M+C-i-1}P_{v2}^{c2}(j)Tk\right]&\mbox{else}\end{array}\right..</td>
<td></td>
<td>(34)</td>
</tr>
</tbody>
</table>
Hence, the average round delay is
<table>
<tbody>
<tr>
<td></td>
<td>Rv2c(M,T,C,1)=∑i=M−TM−T+C−12Pv1c1(i)superscriptsubscript𝑅𝑣2𝑐𝑀𝑇𝐶1superscriptsubscript𝑖𝑀𝑇𝑀𝑇𝐶12superscriptsubscript𝑃𝑣1𝑐1𝑖R_{v2}^{c}(M,T,C,1)=\sum_{i=M-T}^{M-T+C-1}2P_{v1}^{c1}(i)</td>
<td></td>
<td>(35)</td>
</tr>
</tbody>
</table>
and the average polling delay is
<table>
<tbody>
<tr>
<td></td>
<td>Dv2c(M,T,C,1)=∑i=M−TM−T+C−1Pv1c1(i)[i+∑j=TM+C−i−1jPv2c2(j)].superscriptsubscript𝐷𝑣2𝑐𝑀𝑇𝐶1superscriptsubscript𝑖𝑀𝑇𝑀𝑇𝐶1superscriptsubscript𝑃𝑣1𝑐1𝑖delimited-[]𝑖superscriptsubscript𝑗𝑇𝑀𝐶𝑖1𝑗superscriptsubscript𝑃𝑣2𝑐2𝑗D_{v2}^{c}(M,T,C,1)=\sum_{i=M-T}^{M-T+C-1}P_{v1}^{c1}(i)\left[i+\sum_{j=T}^{M+C-i-1}jP_{v2}^{c2}(j)\right].</td>
<td></td>
<td>(36)</td>
</tr>
</tbody>
</table>
Only one round of polling is needed when the chosen node is uncompromised at the first round. The probability that the polling process ends at the witness node $`i`$, where $`T \leq i \leq {T + C}`$, is written as
<table>
<tbody>
<tr>
<td></td>
<td>Pv2u1(i)=1Πv1u1(i−1T−1)(M−i−1M−C−T−1).superscriptsubscript𝑃𝑣2𝑢1𝑖1superscriptsubscriptΠ𝑣1𝑢1𝑖1𝑇1𝑀𝑖1𝑀𝐶𝑇1P_{v2}^{u1}(i)=\frac{1}{\Pi_{v1}^{u1}}\left(\begin{array}[]{c}i-1\\ T-1\end{array}\right)\left(\begin{array}[]{c}M-i-1\\ M-C-T-1\end{array}\right).</td>
<td></td>
</tr>
</tbody>
</table>
The average overhead, and the delays are given by
<table>
<tbody>
<tr>
<td></td>
<td>Ov2u(M,T,C,1)=∑i=TT+CPv2u1(i)Tk,superscriptsubscript𝑂𝑣2𝑢𝑀𝑇𝐶1superscriptsubscript𝑖𝑇𝑇𝐶superscriptsubscript𝑃𝑣2𝑢1𝑖𝑇𝑘O_{v2}^{u}(M,T,C,1)=\sum_{i=T}^{T+C}P_{v2}^{u1}(i)Tk,</td>
<td></td>
<td>(37)</td>
</tr>
</tbody>
</table>
<table>
<tbody>
<tr>
<td></td>
<td>Rv2u(M,T,C,1)=1,superscriptsubscript𝑅𝑣2𝑢𝑀𝑇𝐶11R_{v2}^{u}(M,T,C,1)=1,</td>
<td></td>
<td>(38)</td>
</tr>
</tbody>
</table>
and
<table>
<tbody>
<tr>
<td></td>
<td>Dv2u(M,T,C,1)=∑i=TT+CiPv2u1(i).superscriptsubscript𝐷𝑣2𝑢𝑀𝑇𝐶1superscriptsubscript𝑖𝑇𝑇𝐶𝑖superscriptsubscript𝑃𝑣2𝑢1𝑖D_{v2}^{u}(M,T,C,1)=\sum_{i=T}^{T+C}iP_{v2}^{u1}(i).</td>
<td></td>
<td>(39)</td>
</tr>
</tbody>
</table>
From ([34](#A0.E34 "In V Conclusions and Future Work ‣ Power-Efficient Direct-Voting Assurance for Data Fusion in Wireless Sensor Networks The preliminary result of this work is presented at the IEEE International Conference on Sensor Networks, Ubiquitous, and Trustworthy Computing (SUTC2006), Taiwan, R.O.C.. This work was supported by the National Science Council of Taiwan, R.O.C., under grants NSC 94-2213-E-305-002 and NSC 94-2213-E-305-001.")) and ([39](#A0.E39 "In V Conclusions and Future Work ‣ Power-Efficient Direct-Voting Assurance for Data Fusion in Wireless Sensor Networks The preliminary result of this work is presented at the IEEE International Conference on Sensor Networks, Ubiquitous, and Trustworthy Computing (SUTC2006), Taiwan, R.O.C.. This work was supported by the National Science Council of Taiwan, R.O.C., under grants NSC 94-2213-E-305-002 and NSC 94-2213-E-305-001.")), the average overhead, the delays in Case 2 are given by
<table>
<tbody>
<tr>
<td></td>
<td>Ov2(M,T,C,1)=CMOv2c(M,T,C,1)+M−CMOv2u(M,T,C,1),subscript𝑂𝑣2𝑀𝑇𝐶1𝐶𝑀superscriptsubscript𝑂𝑣2𝑐𝑀𝑇𝐶1𝑀𝐶𝑀superscriptsubscript𝑂𝑣2𝑢𝑀𝑇𝐶1\displaystyle O_{v2}(M,T,C,1)=\frac{C}{M}O_{v2}^{c}(M,T,C,1)+\frac{M-C}{M}O_{v2}^{u}(M,T,C,1),</td>
<td></td>
</tr>
</tbody>
</table>
<table>
<tbody>
<tr>
<td></td>
<td>Rv2(M,T,C,1)=CMRv2c(M,T,C,1)+M−CMRv2u(M,T,C,1),subscript𝑅𝑣2𝑀𝑇𝐶1𝐶𝑀superscriptsubscript𝑅𝑣2𝑐𝑀𝑇𝐶1𝑀𝐶𝑀superscriptsubscript𝑅𝑣2𝑢𝑀𝑇𝐶1\displaystyle R_{v2}(M,T,C,1)=\frac{C}{M}R_{v2}^{c}(M,T,C,1)+\frac{M-C}{M}R_{v2}^{u}(M,T,C,1),</td>
<td></td>
</tr>
</tbody>
</table>
and
<table>
<tbody>
<tr>
<td></td>
<td>Dv2(M,T,C,1)=CMDv2c(M,T,C,1)+M−CMDv2u(M,T,C,1).subscript𝐷𝑣2𝑀𝑇𝐶1𝐶𝑀superscriptsubscript𝐷𝑣2𝑐𝑀𝑇𝐶1𝑀𝐶𝑀superscriptsubscript𝐷𝑣2𝑢𝑀𝑇𝐶1\displaystyle D_{v2}(M,T,C,1)=\frac{C}{M}D_{v2}^{c}(M,T,C,1)+\frac{M-C}{M}D_{v2}^{u}(M,T,C,1).</td>
<td></td>
</tr>
</tbody>
</table>
## References
- \[1\] I. F. Akyildiz, W. Su, Y. Sankarasubramaniam, and E. Cayirci, “Wireless sensor networks: a survey,” *IEEE Commun. Mag.*, vol. 40, pp. 102–114, Aug. 2002.
- \[2\] S. A. Aldosari and J. M. F. Moura, “Detection in decentralized sensor networks,” in *Proc. ICASSP 2004*, Montreal, Canada, May 2004.
- \[3\] R. Anderson and M. Kuhn, “Tamper resistance - a cautionary note,” in *Proc. 2nd Useix Workshop on Electronic Commerce*, Oakland, CA, Nov. 1996, pp. 1–11.
- \[4\] J.-F. Chamberland and V. V. Veeravalli, “Aymptotic results for decentralized detection in power constrained wireless sensor networks,” *IEEE J. Select. Areas Commun.*, vol. 22, no. 6, pp. 1007–1015, Aug. 2004.
- \[5\] D. Culler, D. Estrin, and M. Srivastava, “Overview of sensor networks,” *IEEE Computer*, vol. 37, no. 8, pp. 41–49, Aug. 2004.
- \[6\] L. Dan, K. D. Wong, H. H. Yu, and A. M. Sayeed, “Detection, classification, and tracking of targets,” *IEEE Trans. Signal Processing*, vol. 19, no. 3, pp. 17–29, Mar. 2002.
- \[7\] A. D’Costa, V. Ramachandran, and A. M. Sayeed, “Distributed classification of gaussian space-time sources in wireless sensor networks,” *IEEE J. Select. Areas Commun.*, vol. 22, no. 6, pp. 1026–1036, Aug. 2004.
- \[8\] J. Deng, R. Han, and S. Mishra, “A performance evaluation of intrusion-tolerant routing in wireless sensor networks,” in *Proc. IPSN 2003*, Palo Alto, CA, 2003, pp. 349–364.
- \[9\] W. Du, J. Deng, Y. S. Han, and P. K. Varshney, “A witness-based approach for data fusion assurance in wireless sensor networks,” in *Proc. GLOBECOM 2003*, vol. 3, San Francisco, CA, Dec. 2003, pp. 1435–1439.
- \[10\] Y. Lin, B. Chen, and P. K. Varshney, “Decision fusion rules in multi-hop wireless sensor networks,” *IEEE Trans. Aerosp. Electron. Syst.*, vol. 41, no. 2, pp. 475–488, Apr. 2005.
- \[11\]
J. N. Al-Karaki and A. E. Kamal, “Routing techniques in wireless sensor networks: a survey,” *IEEE Wireless Commun. Mag.*, vol. 11, no. 6, pp. 6–28, Dec. 2004.
- \[12\] D. Niculescu, “Positioning in ad hoc sensor networks,” *IEEE Network*, vol. 18, no. 4, pp. 24–29, Jul.-Aug. 2004.
- \[13\] ——, “Communication paradigms for sensor networks,” *IEEE Commun. Mag.*, vol. 43, no. 3, pp. 116–122, Mar. 2005.
- \[14\] S. Olariu, A. Wadaa, L. Wilson, and M. Eltoweissy, “Wireless sensor networks: leveraging the virtual infrastructure,” *IEEE Network*, vol. 18, no. 4, pp. 51–56, Apr. 2004.
- \[15\] S. Olariu and Q. Xu, “Information assurance in wireless sensor networks,” in *Proc. IEEE IPDPS’05*, Denver, CO, Apr. 2005, pp. 236a–236a.
- \[16\] A. Perrig, J. A. Stankovic, and D. Wagner, “Security in wireless sensor networks,” *Communications of the ACM*, vol. 47, no. 6, pp. 53–57, June 2004.
- \[17\]
B. Przydatek, D. Song, and A. Perrig, “SIA: secure information aggregation in sensor networks,” in *Proc. SenSys 2003*, Los Angeles, CA, Nov. 2003, pp. 255–265.
- \[18\] R. C. Shah, S. Roy, S. Jain, and W. Brunette, “Data mules: modeling a three-tier architecture for sparse sensor networks,” in *Proc. IEEE Workshop on Sensor Network Protocols and Applications*, Anchorage, AK, May 2003, pp. 30–41.
- \[19\] C.-C. Shen, C. Srisathapornphat, and C. Jaikaeo, “Sensor information networking architecture and applications,” *IEEE Personal Commun. Mag.*, vol. 8, no. 4, pp. 52–59, Aug. 2001.
- \[20\] P. Sholander, A. Harris, and J. Brown, “Intersensor information assurance for dod tactical networks,” in *Proc. MILCOM 2002*, vol. 2, Anaheim, CA, Oct. 2002, pp. 1456–1461.
- \[21\] F. Sivrikaya and B. Yener, “Time synchronization in sensor networks: a survey,” *IEEE Network*, vol. 18, no. 4, pp. 45–50, Jul.-Aug. 2004.
- \[22\] K. Sohrabi, W. Merrill, J. Elson, L. Girod, F. Newberg, and W. Kaiser, “Methods for scalable self-assembly of ad hoc wireless sensor networks,” *IEEE Trans. Mobile Comput.*, vol. 3, no. 4, pp. 317–331, Oct.-Dec. 2004.
- \[23\] L. Tong, Q. Zhao, and S. Adireddy, “Sensor networks with mobile agents,” in *Proc. IEEE MILCOM’03*, Boston, MA, Oct. 2003, pp. 688–693.
- \[24\] J. N. Tsitsiklis, “Decentralized detection by a large number of sensors,” *Mathematics of Conrol, Signals, and Systems*, vol. 1, no. 2, pp. 167–182, 1988.
- \[25\] P. K. Varshney, *Distributed Detection and Data Fusion*. New York: Springer, 1997.
- \[26\] T.-Y. Wang, Y. S. Han, and P. K. Varshney, “A combined decision fusion and channel coding scheme for fault-tolerant classification in wireless wensor networks,” in *Proc. ICASSP 2004*, Montreal, Canada, May 2004.
- \[27\] T.-Y. Wang, Y. S. Han, P. K. Varshney, and P.-N. Chen, “Distributed fault-tolerant classification in wireless sensor networks,” *IEEE J. Select. Areas Commun.*, vol. 23, no. 4, pp. 724–734, Apr. 2005.
- \[28\] Z. Yang and L. Tong, “Cooperative sensor networks with misinformed nodes,” *IEEE Trans. Inform. Theory*, vol. 51, no. 12, pp. 4118–4133, Dec. 2005.
- \[29\] W. Zhao and M. H. Ammar, “Message ferrying: proactive routing in highly-partitioned wireless ad hoc networks,” in *Proc. IEEE Workshop on Future Trends in Distributed Computing Systems*, Puerto Rico, May 2003, pp. 308–314. |
0705.0425 | cs.NI | cyber | INSTITUT NATIONAL DE RECHERCHE EN INFORMATIQUE ET EN AUTOMATIQUE
Batch Processor Sharing with
Hyper-Exponential Service Time
Natalia Osipova N° 6180
April 2007
Batch Processor Sharing with
Hyper-Exponential Service Time
Natalia Osipova
<sup>†</sup>
<sup>†</sup>thanks: PhD student at INRIA Sophia Antipolis, France, e-mail: Natalia.Osipova@sophia.inria.fr
<sup>†</sup>
<sup>†</sup>thanks: The work was supported by France Telecom R&D Grant “Modélisation et Gestion du Trafic Réseaux Internet” no. 42937433.
Thème COM — Systèmes communicants
Projet MAESTRO
Rapport de recherche n° 6180 — April 2007 — ?? pages
Abstract: We study Processor-Sharing queueing model with the hyper-exponential service time distribution and Poisson batch arrival process. One of the main goals to study Batch Processor-Sharing (BPS) is the possibility of its application to size-based scheduling, which is used in differentiation between short and long flows in the Internet. In the case of hyper-exponential service time distribution we find an analytical expression for the expected conditional response time for the BPS queue. We show that the expected conditional response time is a concave function of the service time. We apply the obtained results to the Two Level Processor-Sharing (TLPS) model with the hyper-exponential service time distribution and find an expression of the expected response time for the TLPS model. The TLPS scheduling discipline can be applied to size-based differentiation in TCP/IP networks and Web server request handling.
Key-words: Batch Processor Sharing, Two Level Processor sharing, Hyper-Exponential distribution, Laplace transform, Cauchy matrices.
La File d’Attente Avec Service à Temps Partagé Avec Arrivées en Rafale
Résumé : Nous étudions la file d’attente avec service à temps partagé avec arrivées en rafale, BPS (“Batch Processor Sharing”), lorsque les temps de service ont une distribution hyper-exponentielle et que le processus d’arrivée des rafales est Poisson. Une des motivations de cette étude est que la file BPS apparaît naturellement lorsque l’on étudie les procédures d’ordonnancement favorisant les connexions courtes sur un réseau internet. Dans le cas d’une distribution hyper-exponentielle des temps de service nous trouvons une expression analytique pour le temps de réponse moyen conditionné sur le temps de service pour une file d’attente BPS. Nous montrons que ce temps de réponse moyen conditionné est une fonction concave du temps de service. Nous appliquons les résultats obtenus à la file d’attente munie d’une politique à Temps Partagées avec Deux-Niveaux “Two Level Processor Sharing” (TLPS) avec distribution hyper-exponentielle des temps de service. Nous trouvons l’expression du temps de réponse moyen pour la file TLPS. La discipline de service TLPS peut être utilisée pour ordonner l’accès aux ressources en fonction de la taille dans un réseau TCP/IP ou sur un serveur Web.
Mots-clés : La file d’attente avec service à temps partagé avec arrivées en rafale, la file d’attente munie d’une politique à Temps Partagées avec Deux-Niveaux, réseau TCP/IP, distribution hyper-exponentielle, Laplace transforment, Matrices de Cauchy.
## 1 Introduction
The Processor-Sharing (PS) queueing systems are now often used to model communication and computer systems. The PS systems were first introduced by Kleinrock (see \[[6](#bib.bib6)\] and references therein). Under the PS policy each job receives an equal share of the processor.
PS with batch arrivals (BPS) is not yet characterized fully. Kleinrock *et al.* \[[5](#bib.bib5)\] first studied BPS. They found that the derivative of the expected response time satisfies the integral equation and found the analytical solution in the case when the job size (service time) distribution function has the form $`{F\hspace{0pt}{(x)}} = {1 - {p\hspace{0pt}{(x)}\hspace{0pt}e^{- {\mu\hspace{0pt}x}}}}`$ where $`p\hspace{0pt}{(x)}`$ is a polynomial.
Bansal \[[4](#bib.bib4)\], using Kleinrock’s integral equation, obtained the solution for the Laplace transform of the expected conditional service time as a solution of the system of linear equations, when the job size distribution is a hyper-exponential distribution. Also he considers distributions with a rational Laplace transform. Rege and Sengupta \[[12](#bib.bib12)\] obtained the expression for the response time in condition upon the number of customers in the system. Feng and Mishra \[[8](#bib.bib8)\] provided bounds for the expected conditional response time, the bounds depend on the second moment of the service time distribution. Avrachenkov *et al.* \[[2](#bib.bib2)\] proved existence and uniqueness of the solution of the Kleinrock’s integral equation and provided asymptotic analysis and bounds on the expected conditional response time.
We study the BPS model with the hyper-exponential service time distribution. For this distribution we provide the solution of the Kleinrock’s integral equation according to the derivative of the expected conditional response time. We prove the concavity of the expected conditional sojourn time function for the BPS model with the hyper-exponential job size distribution function. We note that the concavity of the expected conditional sojourn time for the BPS with the hyper-exponential job size distribution was proven by another method in \[[10](#bib.bib10)\].
One of the main goals to study BPS is the possibility of its application to age-based scheduling and the possibility to take into account the burstiness of the arrival process. Bursty arrivals often occur in such modern systems as web server. Age-based scheduling is used in differentiation of short and long flows in the Internet. A quite general set of age-based scheduling mechanisms was introduced by Kleinrock and termed as Multy-Level PS (MLPS).
In MLPS jobs are classified into different classes depending on their attained amount of service. Jobs within the same class are served according to FCFS (First Come First Serve), PS (Processor Sharing) or FB (Foreground Background) policy. The classes themselves are served according to the FB policy, so that the priority is given to the jobs with small sizes.
We study the particular case of MLPS, Two-Level PS (TLPS) scheduling mechanism. It is based on the differentiation of jobs according to some threshold and gives priority to jobs with small sizes. The TLPS scheduling mechanism could be used to model such applications as size based differentiation in TCP/IP networks and Web server request differentiation.
It was shown in \[[1](#bib.bib1)\] that when a job size distribution has a decreasing hazard rate, then with the selection of the threshold the expected sojourn time of the TLPS system could be reduced in comparison to standard PS system.
The distribution of file sizes in the Internet has a decreasing hazard rate and often modelled with heavy-tailed distributions. In \[[9](#bib.bib9)\] it is shown that heavy-tailed distributions such as Pareto distribution could be approximated with the hyper-exponential distributions with a significant number of phases.
Therefore, we study the TLPS model with the hyper-exponential service time distribution. We apply the results of the BPS queueing model to the TLPS model and find an expression of expected conditional sojourn time for the TLPS model.
The paper is organized as follows. In Section [2](#S2 "2 The Analysis of the Batch Arrival Processor Sharing Queue") the BPS scheduling mechanism is considered in the case when the job service time distribution is the hyper-exponential distribution. For this case we provide the solution of the Kleinrock’s integral equation. In Section [3](#S3 "3 The Analysis of the Two Level Processor Sharing Model") the results obtained for the BPS model are applied to the TLPS model, where the job size distribution is also hyper-exponential. An analytical expression of the expected sojourn time is found. We put some technical proofs in the Appendix.
## 2 The Analysis of the Batch Arrival Processor Sharing Queue
### 2.1 Main definitions
Let us consider an M/G/1 system with batch arrivals and Processor-Sharing (PS) queueing discipline. The batches arrive according to a Poisson process with arrival rate $`\lambda`$. Let $`\overline{n} > 0`$ be the average size of a batch. Let $`b > 0`$ be the average number of jobs that arrive with (and in addition to) the tagged job.
Let $`B\hspace{0pt}{(x)}`$ be the required job size (service time) distribution and $`{\overline{B}\hspace{0pt}{(x)}} = {1 - {B\hspace{0pt}{(x)}}}`$ be its complementary distribution function.
The load is given by $`\rho = {\lambda\hspace{0pt}\overline{n}\hspace{0pt}m}`$, with $`m = {\int_{0}^{\infty}{x\hspace{0pt}{dB}\hspace{0pt}{(x)}}}`$. We consider that the system is stable, $`\rho < 1`$.
Let $`\alpha\hspace{0pt}{(x)}`$ be the expected conditional response time for a job with service time $`x`$ and $`\alpha'\hspace{0pt}{(x)}`$ be its derivative. Kleinrock showed in \[[6](#bib.bib6)\] that $`\alpha'\hspace{0pt}{(x)}`$ satisfies the following integro-differential equation
<table>
<tbody>
<tr>
<td></td>
<td>α′(x)=λn¯∫0∞α′(y)B¯(x+y)𝑑y+λn¯∫0xα′(y)B¯(x−y)𝑑y+bB¯(x)+1.superscript𝛼′𝑥𝜆¯𝑛superscriptsubscript0superscript𝛼′𝑦¯𝐵𝑥𝑦differential-d𝑦𝜆¯𝑛superscriptsubscript0𝑥superscript𝛼′𝑦¯𝐵𝑥𝑦differential-d𝑦𝑏¯𝐵𝑥1\displaystyle\alpha^{\prime}(x)=\lambda\overline{n}\int_{0}^{\infty}\alpha^{\prime}(y)\overline{B}(x+y)dy+\lambda\overline{n}\int_{0}^{x}\alpha^{\prime}(y)\overline{B}(x-y)dy+b\overline{B}(x)+1.</td>
<td></td>
<td>(1)</td>
</tr>
</tbody>
</table>
We are interested in the case when the job size distribution is a hyper-exponential function
<table>
<tbody>
<tr>
<td></td>
<td>B(x)=1−∑i=1Npie−μix,∑ipi=1,pi>0,μi>0,i=1,…,N,1<N≤∞.formulae-sequence𝐵𝑥1superscriptsubscript𝑖1𝑁subscript𝑝𝑖superscript𝑒subscript𝜇𝑖𝑥formulae-sequencesubscript𝑖subscript𝑝𝑖1formulae-sequencesubscript𝑝𝑖0formulae-sequencesubscript𝜇𝑖0formulae-sequence𝑖1…𝑁1𝑁\displaystyle B(x)=1-\sum_{i=1}^{N}{p_{i}\,e^{-\mu_{i}\,x}},\,\,{\sum_{i}{p_{i}}=1},\,\,p_{i}>0,\,\mu_{i}>0,\quad i=1,...,N,\quad 1<N\leq\infty.</td>
<td></td>
<td>(2)</td>
</tr>
</tbody>
</table>
By $`\sum_{i}`$ and $`\prod_{i}`$ we mean $`\sum_{i = 1}^{N}`$ and $`\prod_{i = 1}^{N}`$. By $`\sum_{i \neq j}`$ or $`\prod_{i \neq j}`$ we mean $`\sum_{{i = {1,\ldots,N}},{i \neq j}}`$ and $`\prod_{{i = {1,\ldots,N}},{i \neq j}}`$. By $`\forall i`$ we mean $`i = {1,\ldots,N}`$.
Without loss of generality, we can assume that
<table>
<tbody>
<tr>
<td></td>
<td>0<μN<μN−1<…<μ2<μ1<∞.0subscript𝜇𝑁subscript𝜇𝑁1…subscript𝜇2subscript𝜇1\displaystyle 0<\mu_{N}<\mu_{N-1}<...<\mu_{2}<\mu_{1}<\infty.</td>
<td></td>
<td>(3)</td>
</tr>
</tbody>
</table>
### 2.2 The expected conditional sojourn time for the BPS model
Let us first prove additional Lemma.
######
Lemma 1
.
The zeros $`b_{i}`$ of the rational function
<table>
<tbody>
<tr>
<td></td>
<td>Ψ(s)=def1−λn¯∑ipis+μi=∏i(s+bi)∏i(s+μi)superscript𝑑𝑒𝑓Ψ𝑠1𝜆¯𝑛subscript𝑖subscript𝑝𝑖𝑠subscript𝜇𝑖subscriptproduct𝑖𝑠subscript𝑏𝑖subscriptproduct𝑖𝑠subscript𝜇𝑖\displaystyle\Psi(s)=^{def}1-\lambda\overline{n}\sum_{i}{\frac{p_{i}}{s+\mu_{i}}}=\frac{\prod_{i}(s+b_{i})}{\prod_{i}(s+\mu_{i})}</td>
<td></td>
<td>(4)</td>
</tr>
</tbody>
</table>
are all real, distinct, positive and satisfy the following inequalities:
<table>
<tbody>
<tr>
<td></td>
<td>0<bN<μN,μi+1<bi<μi,i=1,…,N−1.formulae-sequence0subscript𝑏𝑁subscript𝜇𝑁subscript𝜇𝑖1subscript𝑏𝑖subscript𝜇𝑖𝑖1…𝑁1\displaystyle{}0<b_{N}<\mu_{N},\,\,\,\mu_{i+1}<b_{i}<\mu_{i},\quad i=1,...,N-1.</td>
<td></td>
<td>(5)</td>
</tr>
</tbody>
</table>
###### *Proof.*.
Let us study the following equation
<table>
<tbody>
<tr>
<td></td>
<td>Ψ(s)=0.Ψ𝑠0\Psi(s)=0.</td>
<td></td>
</tr>
</tbody>
</table>
Following the approach of \[[7](#bib.bib7)\] we get that it has $`N_{1}`$ roots $`{{{- b_{i}},i} = 1},{\ldots,N_{1}}`$, where $`N_{1}`$ is the number of distinct elements within $`\mu_{i}`$. As $`{{\mu_{i},i} = 1},{\ldots,N}`$ are all different, $`{\mu_{i} \neq \mu_{j}},{i \neq j}`$, then $`N_{1} = N`$ and there are $`N`$ different roots $`- b_{i}`$. All $`- b_{i}`$ are real, distinct, negative, satisfy the following inequalities: $`0 > {- b_{N}} > {- \mu_{N}}`$, $`{{- \mu_{i + 1}} > {- b_{i}} > {- \mu_{i}}},{i = {1,\ldots,{N - 1}}}`$. With this we prove the statement of Lemma [1](#ThmTheorem1 "Lemma 1. ‣ 2.2 The expected conditional sojourn time for the BPS model ‣ 2 The Analysis of the Batch Arrival Processor Sharing Queue"). So, then it is possible to present $`\Psi\hspace{0pt}{(s)}`$ in the following form:
<table>
<tbody>
<tr>
<td></td>
<td>Ψ(s)=∏i(s+bi)∏i(s+μi),Ψ𝑠subscriptproduct𝑖𝑠subscript𝑏𝑖subscriptproduct𝑖𝑠subscript𝜇𝑖\Psi(s)=\frac{\prod_{i}(s+b_{i})}{\prod_{i}(s+\mu_{i})},</td>
<td></td>
</tr>
</tbody>
</table>
and from here we have (LABEL:eq:b_i_find_eq). ∎
Before presenting our main result let us prove an auxiliary Lemma.
######
Lemma 2
.
The solution of the following system of linear equations:
<table>
<tbody>
<tr>
<td></td>
<td>∑jxjμq2−bj2=1,q=1,…,N,formulae-sequencesubscript𝑗subscript𝑥𝑗superscriptsubscript𝜇𝑞2superscriptsubscript𝑏𝑗21𝑞1…𝑁{}\begin{array}[]{l}\displaystyle\sum_{j}\frac{x_{j}}{\mu_{q}^{2}-b_{j}^{2}}=1,\quad q=1,...,N,\end{array}</td>
<td></td>
<td>(6)</td>
</tr>
</tbody>
</table>
is given by
<table>
<tbody>
<tr>
<td></td>
<td>xk=∏q=1,…,N(μq2−bk2)∏q≠k(bq2−bk2),k=1,…,N.formulae-sequencesubscript𝑥𝑘subscriptproduct𝑞1…𝑁superscriptsubscript𝜇𝑞2superscriptsubscript𝑏𝑘2subscriptproduct𝑞𝑘superscriptsubscript𝑏𝑞2superscriptsubscript𝑏𝑘2𝑘1…𝑁\displaystyle{}x_{k}=\frac{\displaystyle\prod_{q=1,...,N}{(\mu_{q}^{2}-b_{k}^{2})}}{\displaystyle\prod_{q\neq k}{(b_{q}^{2}-b_{k}^{2})}},\quad k=1,...,N.</td>
<td></td>
<td>(7)</td>
</tr>
</tbody>
</table>
###### *Proof.*.
The proof is given in the appendix. ∎
######
Corollary 3
.
The solution of equation ([6](#S2.E6 "In Lemma 2. ‣ 2.2 The expected conditional sojourn time for the BPS model ‣ 2 The Analysis of the Batch Arrival Processor Sharing Queue")) is positive. Namely, $`x_{k} > 0`$ for $`k = {1,\ldots,N}`$.
###### *Proof.*.
As (LABEL:eq:mu_cond) and ([5](#S2.E5 "In Lemma 1. ‣ 2.2 The expected conditional sojourn time for the BPS model ‣ 2 The Analysis of the Batch Arrival Processor Sharing Queue")), then the statement of the Corollary holds. ∎
Now we can prove our main result.
######
Theorem 4
.
The expected conditional response time for BPS queue with hyper-exponential job size distribution function as in (LABEL:eq:B(x)) is given by:
<table>
<tbody>
<tr>
<td></td>
<td>α(x)=c0x−∑kckbke−bkx+∑kckbk,α(0)=0,formulae-sequence𝛼𝑥subscript𝑐0𝑥subscript𝑘subscript𝑐𝑘subscript𝑏𝑘superscript𝑒subscript𝑏𝑘𝑥subscript𝑘subscript𝑐𝑘subscript𝑏𝑘𝛼00\displaystyle\alpha(x)=c_{0}x-\sum_{k}{\frac{c_{k}}{b_{k}}e^{-b_{k}x}}+\sum_{k}\frac{c_{k}}{b_{k}},\quad\alpha(0)=0,</td>
<td></td>
<td>(8)</td>
</tr>
</tbody>
</table>
where
<table>
<tbody>
<tr>
<td></td>
<td>c0=11−ρ,subscript𝑐011𝜌\displaystyle c_{0}=\frac{1}{1-\rho},</td>
<td></td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td>ck=b2λn¯(∏q(μq2−bk2)bk∏q≠k(bq2−bk2)),subscript𝑐𝑘𝑏2𝜆¯𝑛subscriptproduct𝑞superscriptsubscript𝜇𝑞2superscriptsubscript𝑏𝑘2subscript𝑏𝑘subscriptproduct𝑞𝑘superscriptsubscript𝑏𝑞2superscriptsubscript𝑏𝑘2\displaystyle c_{k}=\frac{b}{2\lambda\overline{n}}\left(\frac{\displaystyle\prod_{q}{(\mu_{q}^{2}-b_{k}^{2})}}{b_{k}\displaystyle\prod_{q\neq k}{(b_{q}^{2}-b_{k}^{2})}}\right),</td>
<td></td>
</tr>
</tbody>
</table>
and where $`b_{k}`$ are the solutions of the following equation:
<table>
<tbody>
<tr>
<td></td>
<td>1−λn¯∑ipis+μi=0,1𝜆¯𝑛subscript𝑖subscript𝑝𝑖𝑠subscript𝜇𝑖01-\lambda\overline{n}\sum_{i}{\frac{p_{i}}{s+\mu_{i}}}=0,</td>
<td></td>
</tr>
</tbody>
</table>
and are all positive, distinct, real, satisfy the following inequalities
<table>
<tbody>
<tr>
<td></td>
<td>0<bN<μN,μi+1<bi<μi,i=1,…,N−1.formulae-sequence0subscript𝑏𝑁subscript𝜇𝑁subscript𝜇𝑖1subscript𝑏𝑖subscript𝜇𝑖𝑖1…𝑁10<b_{N}<\mu_{N},\,\,\,\mu_{i+1}<b_{i}<\mu_{i},\quad i=1,...,N-1.</td>
<td></td>
</tr>
</tbody>
</table>
###### *Proof.*.
We can rewrite integral equation ([1](#S2.E1 "In 2.1 Main definitions ‣ 2 The Analysis of the Batch Arrival Processor Sharing Queue")) in the following way:
<table>
<tbody>
<tr>
<td></td>
<td>α′(x)=λn¯∫0∞α′(y)∑ipie−μi(x+y)dy+λn¯∫0xα′(y)B¯(x−y)𝑑y+bB¯(x)+1,superscript𝛼′𝑥𝜆¯𝑛superscriptsubscript0superscript𝛼′𝑦subscript𝑖subscript𝑝𝑖superscript𝑒subscript𝜇𝑖𝑥𝑦𝑑𝑦𝜆¯𝑛superscriptsubscript0𝑥superscript𝛼′𝑦¯𝐵𝑥𝑦differential-d𝑦𝑏¯𝐵𝑥1\displaystyle{\alpha^{\prime}(x)=\lambda\overline{n}\int_{0}^{\infty}\alpha^{\prime}(y)\sum_{i}p_{i}e^{-\mu_{i}(x+y)}dy+\lambda\overline{n}\int_{0}^{x}\alpha^{\prime}(y)\overline{B}(x-y)dy+b\overline{B}(x)+1},</td>
<td></td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td>α′(x)=λn¯∑ipie−μix∫0∞α′(y)e−μiy𝑑y+λn¯∫0xα′(y)B¯(x−y)𝑑y+bB¯(x)+1.superscript𝛼′𝑥𝜆¯𝑛subscript𝑖subscript𝑝𝑖superscript𝑒subscript𝜇𝑖𝑥superscriptsubscript0superscript𝛼′𝑦superscript𝑒subscript𝜇𝑖𝑦differential-d𝑦𝜆¯𝑛superscriptsubscript0𝑥superscript𝛼′𝑦¯𝐵𝑥𝑦differential-d𝑦𝑏¯𝐵𝑥1\displaystyle{\alpha^{\prime}(x)=\lambda\overline{n}\sum_{i}p_{i}e^{-\mu_{i}x}\int_{0}^{\infty}\alpha^{\prime}(y)e^{-\mu_{i}y}dy+\lambda\overline{n}\int_{0}^{x}\alpha^{\prime}(y)\overline{B}(x-y)dy+b\overline{B}(x)+1}.</td>
<td></td>
</tr>
</tbody>
</table>
We note that in the latter equation $`\int_{0}^{\infty}{\alpha'\hspace{0pt}{(y)}\hspace{0pt}e^{- {\mu_{i}\hspace{0pt}y}}\hspace{0pt}{dy}}`$, $`i = {1,\ldots,N}`$ are the Laplace transforms of $`\alpha'\hspace{0pt}{(y)}`$ evaluated at $`{{\mu_{i},i} = 1},{\ldots,N}`$. Denote
<table>
<tbody>
<tr>
<td></td>
<td>Li=∫0∞α′(y)e−μiy𝑑y,i=1,…,N.formulae-sequencesubscript𝐿𝑖superscriptsubscript0superscript𝛼′𝑦superscript𝑒subscript𝜇𝑖𝑦differential-d𝑦𝑖1…𝑁\displaystyle L_{i}=\int_{0}^{\infty}\alpha^{\prime}(y){e^{-\mu_{i}y}dy},\quad i=1,...,N.</td>
<td></td>
</tr>
</tbody>
</table>
Then, we have
<table>
<tbody>
<tr>
<td></td>
<td>α′(x)=λn¯∑ipiLie−μix+λn¯∫0xα′(y)B¯(x−y)𝑑y+bB¯(x)+1.superscript𝛼′𝑥𝜆¯𝑛subscript𝑖subscript𝑝𝑖subscript𝐿𝑖superscript𝑒subscript𝜇𝑖𝑥𝜆¯𝑛superscriptsubscript0𝑥superscript𝛼′𝑦¯𝐵𝑥𝑦differential-d𝑦𝑏¯𝐵𝑥1\displaystyle\alpha^{\prime}(x)=\lambda\overline{n}\sum_{i}p_{i}L_{i}e^{-\mu_{i}x}+{\lambda\overline{n}\int_{0}^{x}\alpha^{\prime}(y){\overline{B}(x-y)}dy+b\overline{B}(x)+1}.</td>
<td></td>
</tr>
</tbody>
</table>
Now taking the Laplace transform of the above equation and using the convolution property, we get the following equation. Here we denote $`L_{\alpha'}\hspace{0pt}{(s)}`$ the Laplace transform of $`\alpha'\hspace{0pt}{(x)}`$.
<table>
<tbody>
<tr>
<td></td>
<td></td>
<td></td>
<td>Lα′(s)=λn¯∑ipiLis+μi+λn¯∑ipiLα′(s)s+μi+b∑ipis+μi+1s,subscript𝐿superscript𝛼′𝑠𝜆¯𝑛subscript𝑖subscript𝑝𝑖subscript𝐿𝑖𝑠subscript𝜇𝑖𝜆¯𝑛subscript𝑖subscript𝑝𝑖subscript𝐿superscript𝛼′𝑠𝑠subscript𝜇𝑖𝑏subscript𝑖subscript𝑝𝑖𝑠subscript𝜇𝑖1𝑠\displaystyle{L_{\alpha^{\prime}}(s)}=\lambda\overline{n}\sum_{i}{\frac{{p_{i}}L_{i}}{s+\mu_{i}}}+\lambda\overline{n}\sum_{i}{\frac{{p_{i}}L_{\alpha^{\prime}}(s)}{s+\mu_{i}}}+{b}\sum_{i}{\frac{p_{i}}{s+\mu_{i}}}+{\frac{1}{s}},</td>
<td></td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td>⟹⟹\displaystyle\Longrightarrow</td>
<td></td>
<td>Lα′(s)(1−λn¯∑ipis+μi)=λn¯∑ipiLis+μi+b∑ipis+μi+1s.subscript𝐿superscript𝛼′𝑠1𝜆¯𝑛subscript𝑖subscript𝑝𝑖𝑠subscript𝜇𝑖𝜆¯𝑛subscript𝑖subscript𝑝𝑖subscript𝐿𝑖𝑠subscript𝜇𝑖𝑏subscript𝑖subscript𝑝𝑖𝑠subscript𝜇𝑖1𝑠\displaystyle L_{\alpha^{\prime}}(s)\left(1-\lambda\overline{n}\sum_{i}{\frac{p_{i}}{s+\mu_{i}}}\right)=\lambda\overline{n}\sum_{i}{\frac{p_{i}L_{i}}{s+\mu_{i}}}+{b}\sum_{i}{\frac{p_{i}}{s+\mu_{i}}}+{\frac{1}{s}}.</td>
<td></td>
</tr>
</tbody>
</table>
Let us note that $`{{L_{\alpha'}\hspace{0pt}{(\mu_{i})}} = L_{i}},{i = {1,\ldots,N}}`$. According to Lemma [1](#ThmTheorem1 "Lemma 1. ‣ 2.2 The expected conditional sojourn time for the BPS model ‣ 2 The Analysis of the Batch Arrival Processor Sharing Queue") and (LABEL:eq:b_i_find_eq), the following equation holds:
<table>
<tbody>
<tr>
<td></td>
<td></td>
<td></td>
<td>Lα′(s)∏i(s+bi)∏i(s+μi)=λn¯∑ipiLis+μi+b∑ipis+μi+1s,subscript𝐿superscript𝛼′𝑠subscriptproduct𝑖𝑠subscript𝑏𝑖subscriptproduct𝑖𝑠subscript𝜇𝑖𝜆¯𝑛subscript𝑖subscript𝑝𝑖subscript𝐿𝑖𝑠subscript𝜇𝑖𝑏subscript𝑖subscript𝑝𝑖𝑠subscript𝜇𝑖1𝑠\displaystyle L_{\alpha^{\prime}}(s)\frac{\prod_{i}(s+b_{i})}{\prod_{i}(s+\mu_{i})}=\lambda\overline{n}\sum_{i}{\frac{p_{i}L_{i}}{s+\mu_{i}}}+{b}\sum_{i}{\frac{p_{i}}{s+\mu_{i}}}+{\frac{1}{s}},</td>
<td></td>
<td>(9)</td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td>⟹⟹\displaystyle\Longrightarrow</td>
<td></td>
<td>Lα′(s)=λn¯∑ipiLi∏k≠i(s+μk)∏i(s+bi)+b∑ipi∏k≠i(s+μk)∏i(s+bi)+1s∏i(s+μi)∏i(s+bi).subscript𝐿superscript𝛼′𝑠𝜆¯𝑛subscript𝑖subscript𝑝𝑖subscript𝐿𝑖subscriptproduct𝑘𝑖𝑠subscript𝜇𝑘subscriptproduct𝑖𝑠subscript𝑏𝑖𝑏subscript𝑖subscript𝑝𝑖subscriptproduct𝑘𝑖𝑠subscript𝜇𝑘subscriptproduct𝑖𝑠subscript𝑏𝑖1𝑠subscriptproduct𝑖𝑠subscript𝜇𝑖subscriptproduct𝑖𝑠subscript𝑏𝑖\displaystyle L_{\alpha^{\prime}}(s)=\lambda\overline{n}\sum_{i}{p_{i}L_{i}{\frac{\prod_{k\neq i}(s+\mu_{k})}{\prod_{i}(s+b_{i})}}}+{b}\sum_{i}{p_{i}}{\frac{\prod_{k\neq i}(s+\mu_{k})}{\prod_{i}(s+b_{i})}}+{\frac{1}{s}}{\frac{\prod_{i}(s+\mu_{i})}{\prod_{i}(s+b_{i})}}.{}</td>
<td></td>
<td>(10)</td>
</tr>
</tbody>
</table>
From here we see that there exist such $`c_{k}`$ that:
<table>
<tbody>
<tr>
<td></td>
<td>Lα′(s)=c0s+∑kcks+bk.subscript𝐿superscript𝛼′𝑠subscript𝑐0𝑠subscript𝑘subscript𝑐𝑘𝑠subscript𝑏𝑘\displaystyle L_{\alpha^{\prime}}(s)=\frac{c_{0}}{s}+\sum_{k}\frac{c_{k}}{s+b_{k}}.</td>
<td></td>
<td>(11)</td>
</tr>
</tbody>
</table>
Then, taking the inversion of the Laplace transform, we have the expression for $`\alpha'\hspace{0pt}{(x)}`$:
<table>
<tbody>
<tr>
<td></td>
<td>α′(x)=c0+∑kcke−bkx.superscript𝛼′𝑥subscript𝑐0subscript𝑘subscript𝑐𝑘superscript𝑒subscript𝑏𝑘𝑥\displaystyle{\alpha^{\prime}}(x)={c_{0}}+\sum_{k}{c_{k}}e^{-b_{k}x}.</td>
<td></td>
</tr>
</tbody>
</table>
and as $`{\alpha\hspace{0pt}{(0)}} = 0`$, then for $`\alpha\hspace{0pt}{(x)}`$ we have
<table>
<tbody>
<tr>
<td></td>
<td>α(x)=c0x−∑kckbke−bkx+∑kckbk,α(0)=0.formulae-sequence𝛼𝑥subscript𝑐0𝑥subscript𝑘subscript𝑐𝑘subscript𝑏𝑘superscript𝑒subscript𝑏𝑘𝑥subscript𝑘subscript𝑐𝑘subscript𝑏𝑘𝛼00\displaystyle\alpha(x)=c_{0}x-\sum_{k}{\frac{c_{k}}{b_{k}}e^{-b_{k}x}}+\sum_{k}\frac{c_{k}}{b_{k}},\quad\alpha(0)=0.</td>
<td></td>
</tr>
</tbody>
</table>
Now let us find $`c_{k}`$. To find $`c_{0}`$ let us multiply both parts of the equation (LABEL:eq:L_alpha_diff_2) by $`s`$ and find the value at the point $`s = 0`$. Then, we have that
<table>
<tbody>
<tr>
<td></td>
<td>c0=Lα′(s)s|s=0=∏iμi∏ibi.subscript𝑐0evaluated-atsubscript𝐿superscript𝛼′𝑠𝑠𝑠0subscriptproduct𝑖subscript𝜇𝑖subscriptproduct𝑖subscript𝑏𝑖\displaystyle c_{0}=L_{\alpha^{\prime}}(s)s|_{s=0}={\frac{\prod_{i}\mu_{i}}{\prod_{i}b_{i}}}.</td>
<td></td>
<td>(12)</td>
</tr>
</tbody>
</table>
From (LABEL:eq:b_i_find_eq) we have the following:
<table>
<tbody>
<tr>
<td></td>
<td>∏ibi∏iμi=Ψ(s)|s=0=1−λn¯∑ipiμi=1−λn¯m=1−ρ.subscriptproduct𝑖subscript𝑏𝑖subscriptproduct𝑖subscript𝜇𝑖evaluated-atΨ𝑠𝑠01𝜆¯𝑛subscript𝑖subscript𝑝𝑖subscript𝜇𝑖1𝜆¯𝑛𝑚1𝜌\displaystyle\frac{\prod_{i}b_{i}}{\prod_{i}\mu_{i}}=\Psi(s)|_{s=0}=1-\lambda\overline{n}\sum_{i}{\frac{p_{i}}{\mu_{i}}}=1-\lambda\overline{n}m={1-\rho}.</td>
<td></td>
</tr>
</tbody>
</table>
So, then
<table>
<tbody>
<tr>
<td></td>
<td>c0=11−ρ.subscript𝑐011𝜌\displaystyle c_{0}=\frac{1}{1-\rho}.</td>
<td></td>
</tr>
</tbody>
</table>
Let us find the other coefficients $`c_{i}`$. We denote:
<table>
<tbody>
<tr>
<td></td>
<td></td>
<td></td>
<td>Lα′∗(s)=∑icis+bi,superscriptsubscript𝐿superscript𝛼′𝑠subscript𝑖subscript𝑐𝑖𝑠subscript𝑏𝑖\displaystyle L_{\alpha^{\prime}}^{*}(s)=\sum_{i}\frac{c_{i}}{s+b_{i}},</td>
<td></td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td>⟹⟹\displaystyle\Longrightarrow</td>
<td></td>
<td>Lα′(s)=c0s+∑icis+bi=c0s+Lα′∗(s).subscript𝐿superscript𝛼′𝑠subscript𝑐0𝑠subscript𝑖subscript𝑐𝑖𝑠subscript𝑏𝑖subscript𝑐0𝑠superscriptsubscript𝐿superscript𝛼′𝑠\displaystyle L_{\alpha^{\prime}}(s)=\frac{c_{0}}{s}+\sum_{i}\frac{c_{i}}{s+b_{i}}=\frac{c_{0}}{s}+L_{\alpha^{\prime}}^{*}(s).</td>
<td></td>
</tr>
</tbody>
</table>
According to the system ([9](#S2.E9 "In Proof.. ‣ 2.2 The expected conditional sojourn time for the BPS model ‣ 2 The Analysis of the Batch Arrival Processor Sharing Queue")), ([10](#S2.E10 "In Proof.. ‣ 2.2 The expected conditional sojourn time for the BPS model ‣ 2 The Analysis of the Batch Arrival Processor Sharing Queue")) we have the following:
<table>
<tbody>
<tr>
<td></td>
<td></td>
<td></td>
<td>Lj∗=∑iciμj+bi,j=1,…,N,formulae-sequencesuperscriptsubscript𝐿𝑗subscript𝑖subscript𝑐𝑖subscript𝜇𝑗subscript𝑏𝑖𝑗1…𝑁\displaystyle L_{j}^{*}=\sum_{i}\frac{c_{i}}{\mu_{j}+b_{i}},\quad j=1,...,N,</td>
<td></td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td>(9)⟹⟹9absent\displaystyle(\ref{eq:L_alpha_diff_mult})\Longrightarrow</td>
<td></td>
<td>Lα′∗(s)∏i(s+bi)∏i(s+μi)=λn¯∑ipiLi∗s+μi+b∑ipis+μi,superscriptsubscript𝐿superscript𝛼′𝑠subscriptproduct𝑖𝑠subscript𝑏𝑖subscriptproduct𝑖𝑠subscript𝜇𝑖𝜆¯𝑛subscript𝑖subscript𝑝𝑖superscriptsubscript𝐿𝑖𝑠subscript𝜇𝑖𝑏subscript𝑖subscript𝑝𝑖𝑠subscript𝜇𝑖\displaystyle L_{\alpha^{\prime}}^{*}(s)\frac{\prod_{i}(s+b_{i})}{\prod_{i}(s+\mu_{i})}=\lambda\overline{n}\sum_{i}{\frac{p_{i}L_{i}^{*}}{s+\mu_{i}}}+{b}\sum_{i}{\frac{p_{i}}{s+\mu_{i}}},</td>
<td></td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td>⟹⟹\displaystyle\Longrightarrow</td>
<td></td>
<td>Lα′∗(s)∏i(s+bi)∏i(s+μi)(s+μq)|s=−μq=λn¯∑ipiLi∗s+μi(s+μq)|s=−μq+evaluated-atsuperscriptsubscript𝐿superscript𝛼′𝑠subscriptproduct𝑖𝑠subscript𝑏𝑖subscriptproduct𝑖𝑠subscript𝜇𝑖𝑠subscript𝜇𝑞𝑠subscript𝜇𝑞limit-fromevaluated-at𝜆¯𝑛subscript𝑖subscript𝑝𝑖superscriptsubscript𝐿𝑖𝑠subscript𝜇𝑖𝑠subscript𝜇𝑞𝑠subscript𝜇𝑞\displaystyle L_{\alpha^{\prime}}^{*}(s)\frac{\prod_{i}(s+b_{i})}{\prod_{i}(s+\mu_{i})}(s+\mu_{q})|_{s=-\mu_{q}}=\lambda\overline{n}\sum_{i}{\frac{p_{i}L_{i}^{*}}{s+\mu_{i}}}(s+\mu_{q})|_{s=-\mu_{q}}+</td>
<td></td>
</tr>
<tr>
<td></td>
<td></td>
<td></td>
<td>+b∑ipis+μi(s+μq)|s=−μq,q=1,…,N,formulae-sequenceevaluated-at𝑏subscript𝑖subscript𝑝𝑖𝑠subscript𝜇𝑖𝑠subscript𝜇𝑞𝑠subscript𝜇𝑞𝑞1…𝑁\displaystyle\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad+{b}\sum_{i}{\frac{p_{i}}{s+\mu_{i}}}(s+\mu_{q})|_{s=-\mu_{q}},\quad q=1,...,N,</td>
<td></td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td>⟹⟹\displaystyle\Longrightarrow</td>
<td></td>
<td>∑jcjbj−μq∏i(bi−μq)∏i≠q(μi−μq)=λn¯pqLq∗+bpq,q=1,…,N,formulae-sequencesubscript𝑗subscript𝑐𝑗subscript𝑏𝑗subscript𝜇𝑞subscriptproduct𝑖subscript𝑏𝑖subscript𝜇𝑞subscriptproduct𝑖𝑞subscript𝜇𝑖subscript𝜇𝑞𝜆¯𝑛subscript𝑝𝑞superscriptsubscript𝐿𝑞𝑏subscript𝑝𝑞𝑞1…𝑁\displaystyle\sum_{j}\frac{c_{j}}{b_{j}-\mu_{q}}\frac{\prod_{i}(b_{i}-\mu_{q})}{\prod_{i\neq q}(\mu_{i}-\mu_{q})}=\lambda\overline{n}{p_{q}L_{q}^{*}}+b{p_{q}},\quad q=1,...,N,</td>
<td></td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td>⟹⟹\displaystyle\Longrightarrow</td>
<td></td>
<td>∑jcjbj−μq∏i(bi−μq)∏i≠q(μi−μq)=λn¯pq∑jcjbj+μq+bpq,q=1,…,N.formulae-sequencesubscript𝑗subscript𝑐𝑗subscript𝑏𝑗subscript𝜇𝑞subscriptproduct𝑖subscript𝑏𝑖subscript𝜇𝑞subscriptproduct𝑖𝑞subscript𝜇𝑖subscript𝜇𝑞𝜆¯𝑛subscript𝑝𝑞subscript𝑗subscript𝑐𝑗subscript𝑏𝑗subscript𝜇𝑞𝑏subscript𝑝𝑞𝑞1…𝑁\displaystyle\sum_{j}\frac{c_{j}}{b_{j}-\mu_{q}}\frac{\prod_{i}(b_{i}-\mu_{q})}{\prod_{i\neq q}(\mu_{i}-\mu_{q})}=\lambda\overline{n}p_{q}\sum_{j}\frac{c_{j}}{b_{j}+\mu_{q}}+b{p_{q}},\quad q=1,...,N.</td>
<td></td>
</tr>
</tbody>
</table>
Let us notice that from (LABEL:eq:b_i_find_eq) we have the following:
<table>
<tbody>
<tr>
<td></td>
<td>∏i(bi−μq)∏i≠q(μi−μq)=∏i(s+bi)∏i(s+μi)(s+μq)|s=−μq=Ψ(s)(s+μq)|s=−μq=subscriptproduct𝑖subscript𝑏𝑖subscript𝜇𝑞subscriptproduct𝑖𝑞subscript𝜇𝑖subscript𝜇𝑞evaluated-atsubscriptproduct𝑖𝑠subscript𝑏𝑖subscriptproduct𝑖𝑠subscript𝜇𝑖𝑠subscript𝜇𝑞𝑠subscript𝜇𝑞evaluated-atΨ𝑠𝑠subscript𝜇𝑞𝑠subscript𝜇𝑞absent\displaystyle\frac{\prod_{i}(b_{i}-\mu_{q})}{\prod_{i\neq q}(\mu_{i}-\mu_{q})}=\frac{\prod_{i}(s+b_{i})}{\prod_{i}(s+\mu_{i})}(s+\mu_{q})|_{s=-\mu_{q}}=\Psi(s)(s+\mu_{q})|_{s=-\mu_{q}}=</td>
<td></td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td>=(1−λn¯∑ipis+μi)(s+μq)|s=−μq=−λn¯pq,q=1,…,N,formulae-sequenceabsentevaluated-at1𝜆¯𝑛subscript𝑖subscript𝑝𝑖𝑠subscript𝜇𝑖𝑠subscript𝜇𝑞𝑠subscript𝜇𝑞𝜆¯𝑛subscript𝑝𝑞𝑞1…𝑁\displaystyle=\left(1-\lambda\overline{n}\sum_{i}{\frac{p_{i}}{s+\mu_{i}}}\right)(s+\mu_{q})|_{s=-\mu_{q}}=-\lambda\overline{n}\,p_{q},\quad q=1,...,N,</td>
<td></td>
</tr>
</tbody>
</table>
then
<table>
<tbody>
<tr>
<td></td>
<td>⟹⟹\displaystyle\Longrightarrow</td>
<td></td>
<td>∑jcjbj−μq(−λn¯pq)=λn¯pq∑jcjbj+μq+bpq,q=1,…,N,formulae-sequencesubscript𝑗subscript𝑐𝑗subscript𝑏𝑗subscript𝜇𝑞𝜆¯𝑛subscript𝑝𝑞𝜆¯𝑛subscript𝑝𝑞subscript𝑗subscript𝑐𝑗subscript𝑏𝑗subscript𝜇𝑞𝑏subscript𝑝𝑞𝑞1…𝑁\displaystyle\sum_{j}\frac{c_{j}}{b_{j}-\mu_{q}}(-\lambda\overline{n}\,p_{q})=\lambda\overline{n}\,p_{q}\sum_{j}\frac{c_{j}}{b_{j}+\mu_{q}}+{b}{p_{q}},\quad q=1,...,N,</td>
<td></td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td>⟹⟹\displaystyle\Longrightarrow</td>
<td></td>
<td>∑jcjμq−bj−∑jcjμq+bj=bλn¯,q=1,…,N,formulae-sequencesubscript𝑗subscript𝑐𝑗subscript𝜇𝑞subscript𝑏𝑗subscript𝑗subscript𝑐𝑗subscript𝜇𝑞subscript𝑏𝑗𝑏𝜆¯𝑛𝑞1…𝑁\displaystyle\sum_{j}\frac{c_{j}}{\mu_{q}-b_{j}}-\sum_{j}\frac{c_{j}}{\mu_{q}+b_{j}}=\frac{b}{\lambda\overline{n}},\quad q=1,...,N,</td>
<td></td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td>⟹⟹\displaystyle\Longrightarrow</td>
<td></td>
<td>∑jcjbjμq2−bj2=b2λn¯,q=1,…,N.formulae-sequencesubscript𝑗subscript𝑐𝑗subscript𝑏𝑗superscriptsubscript𝜇𝑞2superscriptsubscript𝑏𝑗2𝑏2𝜆¯𝑛𝑞1…𝑁\displaystyle\sum_{j}\frac{c_{j}b_{j}}{\mu_{q}^{2}-b_{j}^{2}}={\frac{b}{2\lambda\overline{n}}},\quad q=1,...,N.</td>
<td></td>
</tr>
</tbody>
</table>
So, $`c_{j}`$ are solutions of the following linear system:
<table>
<tbody>
<tr>
<td></td>
<td>∑jcjbjμq2−bj2=b2λn¯,q=1,…,N.formulae-sequencesubscript𝑗subscript𝑐𝑗subscript𝑏𝑗superscriptsubscript𝜇𝑞2superscriptsubscript𝑏𝑗2𝑏2𝜆¯𝑛𝑞1…𝑁{}\begin{array}[]{l}\displaystyle\sum_{j}\frac{c_{j}b_{j}}{\mu_{q}^{2}-b_{j}^{2}}={\frac{b}{2\lambda\overline{n}}},\quad q=1,...,N.\end{array}</td>
<td></td>
<td>(13)</td>
</tr>
</tbody>
</table>
If we denote
<table>
<tbody>
<tr>
<td></td>
<td>xk=ckbkb2λn¯,k=1,…,N,formulae-sequencesubscript𝑥𝑘subscript𝑐𝑘subscript𝑏𝑘𝑏2𝜆¯𝑛𝑘1…𝑁\displaystyle x_{k}=\frac{c_{k}b_{k}}{\frac{b}{2\lambda\overline{n}}},\quad k=1,...,N,</td>
<td></td>
</tr>
</tbody>
</table>
the system ([13](#S2.E13 "In Proof.. ‣ 2.2 The expected conditional sojourn time for the BPS model ‣ 2 The Analysis of the Batch Arrival Processor Sharing Queue")) will take the form ([6](#S2.E6 "In Lemma 2. ‣ 2.2 The expected conditional sojourn time for the BPS model ‣ 2 The Analysis of the Batch Arrival Processor Sharing Queue")) and by Lemma [2](#ThmTheorem2 "Lemma 2. ‣ 2.2 The expected conditional sojourn time for the BPS model ‣ 2 The Analysis of the Batch Arrival Processor Sharing Queue") for $`c_{j}`$ the final result is as follows:
<table>
<tbody>
<tr>
<td></td>
<td>ck=b2λn¯(xkbk)=b2λn¯(∏q(μq2−bk2)bk∏q≠k(bq2−bk2)),k=1,…,N.formulae-sequencesubscript𝑐𝑘𝑏2𝜆¯𝑛subscript𝑥𝑘subscript𝑏𝑘𝑏2𝜆¯𝑛subscriptproduct𝑞superscriptsubscript𝜇𝑞2superscriptsubscript𝑏𝑘2subscript𝑏𝑘subscriptproduct𝑞𝑘superscriptsubscript𝑏𝑞2superscriptsubscript𝑏𝑘2𝑘1…𝑁\displaystyle c_{k}={\frac{b}{2\lambda\overline{n}}}\left({\frac{x_{k}}{b_{k}}}\right)={\frac{b}{2\lambda\overline{n}}}\left(\frac{\displaystyle\prod_{q}{(\mu_{q}^{2}-b_{k}^{2})}}{b_{k}\displaystyle\prod_{q\neq k}{(b_{q}^{2}-b_{k}^{2})}}\right),\quad k=1,...,N.</td>
<td></td>
</tr>
</tbody>
</table>
This completes the proof of Theorem [4](#ThmTheorem4 "Theorem 4. ‣ 2.2 The expected conditional sojourn time for the BPS model ‣ 2 The Analysis of the Batch Arrival Processor Sharing Queue"). ∎
######
Corollary 5
.
The expected conditional sojourn time function in the BPS system with hyper-exponential job size distribution as in (LABEL:eq:B(x)) is a strictly concave function with respect to job sizes.
###### *Proof.*.
The function
<table>
<tbody>
<tr>
<td></td>
<td>α(x)=c0x−∑kckbke−bkx+∑kckbk,α(0)=0formulae-sequence𝛼𝑥subscript𝑐0𝑥subscript𝑘subscript𝑐𝑘subscript𝑏𝑘superscript𝑒subscript𝑏𝑘𝑥subscript𝑘subscript𝑐𝑘subscript𝑏𝑘𝛼00\displaystyle\alpha(x)=c_{0}x-\sum_{k}{\frac{c_{k}}{b_{k}}e^{-b_{k}x}}+\sum_{k}\frac{c_{k}}{b_{k}},\quad\alpha(0)=0</td>
<td></td>
</tr>
</tbody>
</table>
is a strictly concave function if $`{\alpha^{\operatorname{\prime\prime}}\hspace{0pt}{(x)}} < 0`$.
<table>
<tbody>
<tr>
<td></td>
<td>α′′(x)=−∑kckbke−bkx<0superscript𝛼′′𝑥subscript𝑘subscript𝑐𝑘subscript𝑏𝑘superscript𝑒subscript𝑏𝑘𝑥0\displaystyle\alpha^{\prime\prime}(x)=-\sum_{k}{{c_{k}}{b_{k}}e^{-b_{k}x}}<0</td>
<td></td>
</tr>
</tbody>
</table>
as $`{c_{k} > 0},{{b_{k} > 0},{k = {1,\ldots,N}}}`$, which follows from $`{b > 0},{\overline{n} > 0}`$, Corollary [3](#ThmTheorem3 "Corollary 3. ‣ 2.2 The expected conditional sojourn time for the BPS model ‣ 2 The Analysis of the Batch Arrival Processor Sharing Queue") and Lemma [1](#ThmTheorem1 "Lemma 1. ‣ 2.2 The expected conditional sojourn time for the BPS model ‣ 2 The Analysis of the Batch Arrival Processor Sharing Queue"). ∎
######
Corollary 6
.
The expected sojourn time in the BPS system with hyper-exponential job size distribution as in (LABEL:eq:B(x)) is given by
<table>
<tbody>
<tr>
<td></td>
<td>T¯BPS=m1−ρ+∑i,jpicjμi+bj.superscript¯𝑇𝐵𝑃𝑆𝑚1𝜌subscript𝑖𝑗subscript𝑝𝑖subscript𝑐𝑗subscript𝜇𝑖subscript𝑏𝑗\displaystyle\overline{T}^{BPS}=\frac{m}{1-\rho}+\sum_{i,j}{\frac{p_{i}c_{j}}{\mu_{i}+b_{j}}}.</td>
<td></td>
<td>(14)</td>
</tr>
</tbody>
</table>
###### *Proof.*.
As the expected sojourn time $`{\overline{T}}^{B\hspace{0pt}P\hspace{0pt}S}`$ is given by
<table>
<tbody>
<tr>
<td></td>
<td>T¯BPS=∫0∞α′(x)B¯(x)𝑑x,superscript¯𝑇𝐵𝑃𝑆superscriptsubscript0superscript𝛼′𝑥¯𝐵𝑥differential-d𝑥\displaystyle\overline{T}^{BPS}=\int_{0}^{\infty}\alpha^{\prime}(x)\overline{B}(x)dx,</td>
<td></td>
</tr>
</tbody>
</table>
then using (LABEL:eq:alpha(x)) we receive the statement of the Corollary. ∎
## 3 The Analysis of the Two Level Processor Sharing Model
### 3.1 Main definitions
We study the Two Level Processor Sharing (TLPS) scheduling discipline with the hyper-exponential job size distribution. The model description is as follows.
Jobs arrive to the system according to a Poisson process with rate $`\lambda`$. Let $`\theta`$ be a given threshold. The jobs in the system that attained a service less than $`\theta`$ are assigned to the high priority queue. If in addition there are jobs with attained service greater than $`\theta`$, such a job is separated into two parts. The first part of size $`\theta`$ is assigned to the high priority queue and the second part of size $`x - \theta`$ waits in the lower priority queue. The low priority queue is served when the high priority queue is empty. Both queues are served according to the Processor Sharing (PS) discipline.
Let us denote the job size distribution by $`F\hspace{0pt}{(x)}`$. By $`{\overline{F}\hspace{0pt}{(x)}} = {1 - {F\hspace{0pt}{(x)}}}`$ we denote the complementary distribution function. We consider the case, when $`F\hspace{0pt}{(x)}`$ is a hyper-exponential distribution function, namely
<table>
<tbody>
<tr>
<td></td>
<td>F(x)=1−∑i=1Npi~e−μix,∑ipi~=1,pi~>0,μi>0,i=1,…,N,1<N≤∞.formulae-sequence𝐹𝑥1superscriptsubscript𝑖1𝑁~subscript𝑝𝑖superscript𝑒subscript𝜇𝑖𝑥formulae-sequencesubscript𝑖~subscript𝑝𝑖1formulae-sequence~subscript𝑝𝑖0formulae-sequencesubscript𝜇𝑖0formulae-sequence𝑖1…𝑁1𝑁{}F(x)=1-\sum_{i=1}^{N}{\widetilde{p_{i}}e^{-\mu_{i}x}},\quad\sum_{i}{\widetilde{p_{i}}}=1,\,\,\widetilde{p_{i}}>0,\,\,\mu_{i}>0,\quad i=1,...,N,\quad 1<N\leq\infty.</td>
<td></td>
<td>(15)</td>
</tr>
</tbody>
</table>
The mean job size is given by $`m = {\int_{0}^{\infty}{x\hspace{0pt}{dF}\hspace{0pt}{(x)}}}`$ and the system load is $`\rho = {\lambda\hspace{0pt}m}`$. We assume that the system is stable ($`\rho < 1`$) and is in steady state.
### 3.2 The expected conditional sojourn time for the TLPS model
Let us denote by $`{\overline{T}}^{T\hspace{0pt}L\hspace{0pt}P\hspace{0pt}S}\hspace{0pt}{(x)}`$ the expected conditional sojourn time in the TLPS system for a job of size $`x`$ and by $`\overline{T}\hspace{0pt}{(\theta)}`$ the expected sojourn time of the system.
According to \[[6](#bib.bib6)\] the expected conditional sojourn time of the system is given by:
<table>
<tbody>
<tr>
<td></td>
<td>T¯TLPS(x)={x1−ρθ,x∈[0,θ],W¯(θ)+θ+α(x−θ)1−ρθ,x∈(θ,∞),superscript¯𝑇𝑇𝐿𝑃𝑆𝑥cases𝑥1subscript𝜌𝜃𝑥0𝜃¯𝑊𝜃𝜃𝛼𝑥𝜃1subscript𝜌𝜃𝑥𝜃\displaystyle\overline{T}^{TLPS}(x)=\left\{\begin{array}[]{l l}{\displaystyle\frac{x}{1-\rho_{\theta}}},&x\in[0,\theta],\\ {\displaystyle\frac{\overline{W}(\theta)+\theta+\alpha(x-\theta)}{1-\rho_{\theta}}},&x\in(\theta,\infty),\end{array}\right.</td>
<td></td>
</tr>
</tbody>
</table>
where $`\rho_{\theta}`$ is the utilization factor for the truncated distribution $`\rho_{\theta} = {\lambda\hspace{0pt}\overline{X_{\theta}^{1}}}`$, the $`n`$-th moment of the distribution truncated at $`\theta`$ is
<table>
<tbody>
<tr>
<td></td>
<td>Xθn¯=∫0θnyn−1F¯(y)𝑑y,¯superscriptsubscript𝑋𝜃𝑛superscriptsubscript0𝜃𝑛superscript𝑦𝑛1¯𝐹𝑦differential-d𝑦{}{\overline{X_{\theta}^{n}}}=\int_{0}^{\theta}\!ny^{n-1}\overline{F}(y)dy,</td>
<td></td>
</tr>
</tbody>
</table>
$`{({{\overline{W}\hspace{0pt}{(\theta)}} + \theta})}/{({1 - \rho_{\theta}})}`$ expresses the time needed to reach the low priority queue. This time consists of the time $`\theta/{({1 - \rho_{\theta}})}`$ spent in the high priority queue, where the flow is served up to the threshold $`\theta`$, plus the time $`{\overline{W}\hspace{0pt}{(\theta)}}/{({1 - \rho_{\theta}})}`$ which is spent waiting for the high priority queue to empty. Here $`{\overline{W}\hspace{0pt}{(\theta)}} = \frac{\lambda\hspace{0pt}\overline{X_{\theta}^{2}}}{2\hspace{0pt}{({1 - \rho_{\theta}})}}`$. The remaining term $`{\alpha\hspace{0pt}{({x - \theta})}}/{({1 - \rho_{\theta}})}`$ is the time spent in the low priority queue.
To find $`\alpha\hspace{0pt}{(x)}`$ we use the interpretation of the lower priority queue as a PS system with batch arrivals.
Let us denote
<table>
<tbody>
<tr>
<td></td>
<td>Fθi¯=pi~e−μiθ,i=1,…,N,formulae-sequence¯superscriptsubscript𝐹𝜃𝑖~subscript𝑝𝑖superscript𝑒subscript𝜇𝑖𝜃𝑖1…𝑁\overline{F_{\theta}^{i}}=\widetilde{p_{i}}e^{-\mu_{i}\theta},\quad i=1,...,N,</td>
<td></td>
</tr>
</tbody>
</table>
and prove the following Theorem:
######
Theorem 7
.
In TLPS priority queue with the hyper-exponential job size distribution as in ([15](#S3.E15 "In 3.1 Main definitions ‣ 3 The Analysis of the Two Level Processor Sharing Model")):
<table>
<tbody>
<tr>
<td></td>
<td>α(x)=c0(θ)x−∑kck(θ)bk(θ)e−bk(θ)x+∑kck(θ)bk(θ),α(0)=0,formulae-sequence𝛼𝑥subscript𝑐0𝜃𝑥subscript𝑘subscript𝑐𝑘𝜃subscript𝑏𝑘𝜃superscript𝑒subscript𝑏𝑘𝜃𝑥subscript𝑘subscript𝑐𝑘𝜃subscript𝑏𝑘𝜃𝛼00\displaystyle\alpha(x)=c_{0}(\theta)x-\sum_{k}{\frac{c_{k}(\theta)}{b_{k}(\theta)}e^{-b_{k}(\theta)x}}+\sum_{k}\frac{c_{k}(\theta)}{b_{k}(\theta)},\quad\alpha(0)=0,{}</td>
<td></td>
</tr>
</tbody>
</table>
where
<table>
<tbody>
<tr>
<td></td>
<td>c0(θ)=1−ρθ1−ρ,subscript𝑐0𝜃1subscript𝜌𝜃1𝜌\displaystyle c_{0}(\theta)=\frac{1-\rho_{\theta}}{1-\rho},</td>
<td></td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td>ck(θ)=b2λn¯(∏q=1,…,N(μq2−bk2(θ))bk(θ)∏q≠k(bq2(θ)−bk2(θ))),k=1,…,N,formulae-sequencesubscript𝑐𝑘𝜃𝑏2𝜆¯𝑛subscriptproduct𝑞1…𝑁superscriptsubscript𝜇𝑞2superscriptsubscript𝑏𝑘2𝜃subscript𝑏𝑘𝜃subscriptproduct𝑞𝑘superscriptsubscript𝑏𝑞2𝜃superscriptsubscript𝑏𝑘2𝜃𝑘1…𝑁\displaystyle c_{k}(\theta)={\frac{b}{2\lambda\overline{n}}}\left(\frac{\displaystyle\prod_{q=1,...,N}{(\mu_{q}^{2}-b_{k}^{2}(\theta))}}{b_{k}(\theta)\displaystyle\prod_{q\neq k}{(b_{q}^{2}(\theta)-b_{k}^{2}(\theta))}}\right),\quad k=1,...,N,{}</td>
<td></td>
</tr>
</tbody>
</table>
and $`b_{i}\hspace{0pt}{(\theta)}`$ are roots of
<table>
<tbody>
<tr>
<td></td>
<td>1−λ1−ρθ∑iFθi¯s+μi=0,1𝜆1subscript𝜌𝜃subscript𝑖¯superscriptsubscript𝐹𝜃𝑖𝑠subscript𝜇𝑖0\displaystyle 1-\frac{\lambda}{1-\rho_{\theta}}\sum_{i}{\frac{\overline{F_{\theta}^{i}}}{s+\mu_{i}}}=0,</td>
<td></td>
</tr>
</tbody>
</table>
and satisfy the following inequalities:
<table>
<tbody>
<tr>
<td></td>
<td>0<bN(θ)<μN,μi+1<bi(θ)<μi,i=1,…,N−1.formulae-sequence0subscript𝑏𝑁𝜃subscript𝜇𝑁subscript𝜇𝑖1subscript𝑏𝑖𝜃subscript𝜇𝑖𝑖1…𝑁10<b_{N}(\theta)<\mu_{N},\,\,\,\mu_{i+1}<b_{i}(\theta)<\mu_{i},\quad i=1,...,N-1.</td>
<td></td>
</tr>
</tbody>
</table>
###### *Proof.*.
As was shown in \[[6](#bib.bib6)\], $`{\alpha'\hspace{0pt}{(x)}} = {{{d\hspace{0pt}\alpha}/d}\hspace{0pt}x}`$ is the solution of the integral equation we looked at in Section [2](#S2 "2 The Analysis of the Batch Arrival Processor Sharing Queue")
<table>
<tbody>
<tr>
<td></td>
<td>α′(x)=λn¯∫0∞α′(y)B¯(x+y)𝑑y+λn¯∫0xα′(y)B¯(x−y)𝑑y+bB¯(x)+1.superscript𝛼′𝑥𝜆¯𝑛superscriptsubscript0superscript𝛼′𝑦¯𝐵𝑥𝑦differential-d𝑦𝜆¯𝑛superscriptsubscript0𝑥superscript𝛼′𝑦¯𝐵𝑥𝑦differential-d𝑦𝑏¯𝐵𝑥1{\alpha^{\prime}(x)=\lambda\overline{n}\int_{0}^{\infty}\alpha^{\prime}(y)\overline{B}(x+y)dy+\lambda\overline{n}\int_{0}^{x}\alpha^{\prime}(y)\overline{B}(x-y)dy+b\overline{B}(x)+1}.</td>
<td></td>
</tr>
</tbody>
</table>
Here the average batch size is given by $`\overline{n} = \frac{1 - {F\hspace{0pt}{(\theta)}}}{1 - \rho_{\theta}}`$ and the average number of jobs that arrive to the low priority queue in addition to the tagged job is given by $`b = \frac{2\hspace{0pt}\lambda\hspace{0pt}\left( {1 - {F\hspace{0pt}(\theta)}} \right)\hspace{0pt}\left( {{\overline{W}\hspace{0pt}{(\theta)}} + \theta} \right)}{({1 - \rho_{\theta}})}`$. The complementary truncated distribution is $`{\overline{B}\hspace{0pt}{(x)}} = \frac{1 - {F\hspace{0pt}{({x + \theta})}}}{1 - {F\hspace{0pt}{(\theta)}}}`$, then:
<table>
<tbody>
<tr>
<td></td>
<td>B¯(x)=F¯(x+θ)F¯(θ)=1F¯(θ)∑ipi~e−μi(x+θ)=∑ipi~e−μiθF¯(θ)e−μix=∑ipie−μix,¯𝐵𝑥¯𝐹𝑥𝜃¯𝐹𝜃1¯𝐹𝜃subscript𝑖~subscript𝑝𝑖superscript𝑒subscript𝜇𝑖𝑥𝜃subscript𝑖~subscript𝑝𝑖superscript𝑒subscript𝜇𝑖𝜃¯𝐹𝜃superscript𝑒subscript𝜇𝑖𝑥subscript𝑖subscript𝑝𝑖superscript𝑒subscript𝜇𝑖𝑥\displaystyle\overline{B}(x)=\frac{\overline{F}(x+\theta)}{\overline{F}(\theta)}=\frac{1}{\overline{F}(\theta)}{\sum_{i}{\widetilde{p_{i}}e^{-\mu_{i}(x+\theta)}}}=\sum_{i}\frac{\widetilde{p_{i}}e^{-\mu_{i}\theta}}{\overline{F}(\theta)}e^{-\mu_{i}x}=\sum_{i}{p_{i}}e^{-\mu_{i}x},</td>
<td></td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td>pi=1F¯(θ)pi~e−μiθ=1F¯(θ)Fθi¯,i=1,…,N.formulae-sequencesubscript𝑝𝑖1¯𝐹𝜃~subscript𝑝𝑖superscript𝑒subscript𝜇𝑖𝜃1¯𝐹𝜃¯superscriptsubscript𝐹𝜃𝑖𝑖1…𝑁\displaystyle p_{i}=\frac{1}{\overline{F}(\theta)}\widetilde{p_{i}}e^{-\mu_{i}\theta}=\frac{1}{\overline{F}(\theta)}\overline{F_{\theta}^{i}},\quad i=1,...,N.</td>
<td></td>
</tr>
</tbody>
</table>
We apply the results of Theorem [4](#ThmTheorem4 "Theorem 4. ‣ 2.2 The expected conditional sojourn time for the BPS model ‣ 2 The Analysis of the Batch Arrival Processor Sharing Queue") for the TLPS model. To calculate $`c_{0}`$ we use (LABEL:eq:c0_L) and (LABEL:eq:b_i_find_eq) and get the following
<table>
<tbody>
<tr>
<td></td>
<td>1c0(θ)=∏ibi(θ)∏iμi=Ψ(s)|s=0=1−λn¯∑ipiμi=1−λ1−ρθ∑iFθi¯μi=1−λ(m−Xθ1¯)1−ρθ=1−ρ1−ρθ.1subscript𝑐0𝜃subscriptproduct𝑖subscript𝑏𝑖𝜃subscriptproduct𝑖subscript𝜇𝑖evaluated-atΨ𝑠𝑠01𝜆¯𝑛subscript𝑖subscript𝑝𝑖subscript𝜇𝑖1𝜆1subscript𝜌𝜃subscript𝑖¯superscriptsubscript𝐹𝜃𝑖subscript𝜇𝑖1𝜆𝑚¯subscriptsuperscript𝑋1𝜃1subscript𝜌𝜃1𝜌1subscript𝜌𝜃\displaystyle\frac{1}{c_{0}(\theta)}=\frac{\prod_{i}b_{i}(\theta)}{\prod_{i}\mu_{i}}=\Psi(s)|_{s=0}=1-\lambda\overline{n}\sum_{i}{\frac{p_{i}}{\mu_{i}}}=1-\frac{\lambda}{1-\rho_{\theta}}\sum_{i}{\frac{\overline{F_{\theta}^{i}}}{\mu_{i}}}=1-\frac{\lambda(m-\overline{X^{1}_{\theta}})}{1-\rho_{\theta}}=\frac{1-\rho}{1-\rho_{\theta}}.</td>
<td></td>
</tr>
</tbody>
</table>
So, then
<table>
<tbody>
<tr>
<td></td>
<td>c0(θ)=1−ρθ1−ρ.subscript𝑐0𝜃1subscript𝜌𝜃1𝜌\displaystyle c_{0}(\theta)=\frac{1-\rho_{\theta}}{1-\rho}.</td>
<td></td>
</tr>
</tbody>
</table>
For $`{c_{k}\hspace{0pt}{(\theta)}},{b_{k}\hspace{0pt}{(\theta)}}`$ we use the results of Theorem [4](#ThmTheorem4 "Theorem 4. ‣ 2.2 The expected conditional sojourn time for the BPS model ‣ 2 The Analysis of the Batch Arrival Processor Sharing Queue") and get the statement of Theorem [7](#ThmTheorem7 "Theorem 7. ‣ 3.2 The expected conditional sojourn time for the TLPS model ‣ 3 The Analysis of the Two Level Processor Sharing Model"). ∎
######
Corollary 8
.
The coefficients $`{{c_{k}\hspace{0pt}{(\theta)}} > 0},{k = {1,\ldots,N}}`$ if $`\theta > 0`$.
###### *Proof.*.
As (LABEL:eq:mu_cond), $`{0 < {b_{N}\hspace{0pt}{(\theta)}} < \mu_{N}},{\mu_{i + 1} < {b_{i}\hspace{0pt}{(\theta)}} < \mu_{i}},{i = {1,\ldots,{N - 1}}}`$ and $`b > 0`$, $`\overline{n} > 0`$ when $`\theta > 0`$, then the statement of Corollary holds. ∎
######
Corollary 9
.
For the TLPS queue with hyper-exponential job size distribution as in ([15](#S3.E15 "In 3.1 Main definitions ‣ 3 The Analysis of the Two Level Processor Sharing Model")) the function $`\alpha\hspace{0pt}{(x)}`$ is a strictly concave function with respect to job sizes with positive values of $`\theta`$.
###### *Proof.*.
The function
<table>
<tbody>
<tr>
<td></td>
<td>α(x)=c0(θ)x−∑kck(θ)bk(θ)e−bk(θ)x+∑kck(θ)bk(θ),α(0)=0formulae-sequence𝛼𝑥subscript𝑐0𝜃𝑥subscript𝑘subscript𝑐𝑘𝜃subscript𝑏𝑘𝜃superscript𝑒subscript𝑏𝑘𝜃𝑥subscript𝑘subscript𝑐𝑘𝜃subscript𝑏𝑘𝜃𝛼00\displaystyle\alpha(x)=c_{0}(\theta)x-\sum_{k}{\frac{c_{k}(\theta)}{b_{k}(\theta)}e^{-b_{k}(\theta)x}}+\sum_{k}\frac{c_{k}(\theta)}{b_{k}(\theta)},\quad\alpha(0)=0</td>
<td></td>
</tr>
</tbody>
</table>
is a strictly concave function if $`{\alpha^{\operatorname{\prime\prime}}\hspace{0pt}{(x)}} < 0`$.
<table>
<tbody>
<tr>
<td></td>
<td>α′′(x)=−∑kck(θ)bk(θ)e−bk(θ)x<0superscript𝛼′′𝑥subscript𝑘subscript𝑐𝑘𝜃subscript𝑏𝑘𝜃superscript𝑒subscript𝑏𝑘𝜃𝑥0\alpha^{\prime\prime}(x)=-\sum_{k}{{c_{k}(\theta)}{b_{k}(\theta)}e^{-b_{k}(\theta)x}}<0</td>
<td></td>
</tr>
</tbody>
</table>
as $`{{{c_{k}\hspace{0pt}{(\theta)}} > 0},{{{b_{k}\hspace{0pt}{(\theta)}} > 0},{k = {1,\ldots,N}}}},`$ if $`\theta > 0`$ as it follows from Corollary [8](#ThmTheorem8 "Corollary 8. ‣ 3.2 The expected conditional sojourn time for the TLPS model ‣ 3 The Analysis of the Two Level Processor Sharing Model") and Theorem [7](#ThmTheorem7 "Theorem 7. ‣ 3.2 The expected conditional sojourn time for the TLPS model ‣ 3 The Analysis of the Two Level Processor Sharing Model"). ∎
######
Corollary 10
.
The expected conditional sojourn time for the TLPS queue with hyper-exponential job size distribution as in ([15](#S3.E15 "In 3.1 Main definitions ‣ 3 The Analysis of the Two Level Processor Sharing Model")) is not a concave function with respect to job sizes.
###### *Proof.*.
The function $`{\overline{T}}^{T\hspace{0pt}L\hspace{0pt}P\hspace{0pt}S}\hspace{0pt}{(x)}`$ is a concave function, when
<table>
<tbody>
<tr>
<td></td>
<td>T¯TLPS(x)′|x=θ−0≥T¯TLPS(x)′|x=θ+0\displaystyle\overline{T}^{TLPS}\!\,{}^{\prime}(x)|_{x=\theta-0}\geq\overline{T}^{TLPS}\!\,{}^{\prime}(x)|_{x=\theta+0}</td>
<td></td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td>11−ρθ|x=θ−0≥α′(x−θ)1−ρθ|x=θ+0evaluated-at11subscript𝜌𝜃𝑥𝜃0evaluated-atsuperscript𝛼′𝑥𝜃1subscript𝜌𝜃𝑥𝜃0\displaystyle\frac{1}{1-\rho_{\theta}}|_{x=\theta-0}\geq{\displaystyle\frac{\alpha^{\prime}(x-\theta)}{1-\rho_{\theta}}}|_{x=\theta+0}</td>
<td></td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td>1≥α′(0+).1superscript𝛼′limit-from0\displaystyle 1\geq{\alpha^{\prime}(0+)}.</td>
<td></td>
</tr>
</tbody>
</table>
As shown in \[[2](#bib.bib2)\],
<table>
<tbody>
<tr>
<td></td>
<td>α(x)≥x1−ρ,𝛼𝑥𝑥1𝜌\displaystyle\alpha(x)\geq\frac{x}{1-\rho},</td>
<td></td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td>α′(x)≥11−ρ>1.superscript𝛼′𝑥11𝜌1\displaystyle\alpha^{\prime}(x)\geq\frac{1}{1-\rho}>1.</td>
<td></td>
</tr>
</tbody>
</table>
Then, it follows that $`{\overline{T}}^{T\hspace{0pt}L\hspace{0pt}P\hspace{0pt}S}\hspace{0pt}{(x)}`$ is not a concave function. ∎
### 3.3 The expected sojourn time for the TLPS model
######
Theorem 11
.
The expected sojourn time in the TLPS system with hyper-exponential distribution function ([15](#S3.E15 "In 3.1 Main definitions ‣ 3 The Analysis of the Two Level Processor Sharing Model")) is given by the following equation:
<table>
<tbody>
<tr>
<td></td>
<td>T¯(θ)=Xθ1¯+W¯(θ)F¯(θ)1−ρθ+(m−Xθ1¯)1−ρ+(W¯(θ)+θ)1−ρθ∑i,jFθi¯bj(θ)(μi+bj(θ))∏q(μq2−bj2(θ))∏q≠j(bq2(θ)−bj2(θ)),¯𝑇𝜃¯subscriptsuperscript𝑋1𝜃¯𝑊𝜃¯𝐹𝜃1subscript𝜌𝜃𝑚¯subscriptsuperscript𝑋1𝜃1𝜌¯𝑊𝜃𝜃1subscript𝜌𝜃subscript𝑖𝑗¯superscriptsubscript𝐹𝜃𝑖subscript𝑏𝑗𝜃subscript𝜇𝑖subscript𝑏𝑗𝜃subscriptproduct𝑞superscriptsubscript𝜇𝑞2superscriptsubscript𝑏𝑗2𝜃subscriptproduct𝑞𝑗superscriptsubscript𝑏𝑞2𝜃superscriptsubscript𝑏𝑗2𝜃\displaystyle{\overline{T}(\theta)}=\frac{{\overline{X^{1}_{\theta}}}+\overline{W}(\theta)\overline{F}(\theta)}{1-\rho_{\theta}}+{\frac{(m-{\overline{X^{1}_{\theta}}})}{1-\rho}}+{\frac{(\overline{W}(\theta)+\theta)}{1-\rho_{\theta}}}\sum_{i,j}\frac{\overline{F_{\theta}^{i}}}{b_{j}(\theta)(\mu_{i}+b_{j}(\theta))}\frac{\displaystyle\prod_{q}{(\mu_{q}^{2}-b_{j}^{2}(\theta))}}{\displaystyle\prod_{q\neq j}{(b_{q}^{2}(\theta)-b_{j}^{2}(\theta))}},</td>
<td></td>
</tr>
</tbody>
</table>
where $`b_{i}\hspace{0pt}{(\theta)}`$ are defined as in Theorem [7](#ThmTheorem7 "Theorem 7. ‣ 3.2 The expected conditional sojourn time for the TLPS model ‣ 3 The Analysis of the Two Level Processor Sharing Model").
###### *Proof.*.
According to \[[6](#bib.bib6)\] and Theorem [7](#ThmTheorem7 "Theorem 7. ‣ 3.2 The expected conditional sojourn time for the TLPS model ‣ 3 The Analysis of the Two Level Processor Sharing Model") we have the following:
<table>
<tbody>
<tr>
<td></td>
<td></td>
<td></td>
<td>T¯(θ)=Xθ1¯+W¯(θ)F¯(θ)1−ρθ+11−ρθT¯BPS(θ),¯𝑇𝜃¯superscriptsubscript𝑋𝜃1¯𝑊𝜃¯𝐹𝜃1subscript𝜌𝜃11subscript𝜌𝜃superscript¯𝑇𝐵𝑃𝑆𝜃\displaystyle\overline{T}(\theta)=\frac{{\overline{X_{\theta}^{1}}}+\overline{W}(\theta)\overline{F}(\theta)}{1-\rho_{\theta}}+{\frac{1}{1-\rho_{\theta}}}\overline{T}^{BPS}(\theta),</td>
<td></td>
</tr>
<tr>
<td></td>
<td></td>
<td></td>
<td>T¯BPS(θ)=∫θ∞α(x−θ)𝑑F(x)=∫0∞α′(x)F¯(x+θ)𝑑x,superscript¯𝑇𝐵𝑃𝑆𝜃superscriptsubscript𝜃𝛼𝑥𝜃differential-d𝐹𝑥superscriptsubscript0superscript𝛼′𝑥¯𝐹𝑥𝜃differential-d𝑥\displaystyle\overline{T}^{BPS}(\theta)={\int_{\theta}^{\infty}{{\alpha(x-\theta)}}dF(x)=\int_{0}^{\infty}\alpha^{\prime}(x)\overline{F}(x+\theta)dx},</td>
<td></td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td>(Theorem7)⟹⟹𝑇ℎ𝑒𝑜𝑟𝑒𝑚7absent\displaystyle(Theorem\leavevmode\nobreak\ {\ref{th:TLPS_alpha}})\Longrightarrow</td>
<td></td>
<td>T¯BPS(θ)=c0(θ)(m−Xθ1¯)+∑i,jFθi¯cj(θ)μi+bj(θ),superscript¯𝑇𝐵𝑃𝑆𝜃subscript𝑐0𝜃𝑚¯subscriptsuperscript𝑋1𝜃subscript𝑖𝑗¯superscriptsubscript𝐹𝜃𝑖subscript𝑐𝑗𝜃subscript𝜇𝑖subscript𝑏𝑗𝜃\displaystyle\overline{T}^{BPS}(\theta)=c_{0}(\theta)(m-{\overline{X^{1}_{\theta}}})+\sum_{i,j}\frac{\overline{F_{\theta}^{i}}c_{j}(\theta)}{\mu_{i}+b_{j}(\theta)},</td>
<td></td>
</tr>
</tbody>
</table>
then
<table>
<tbody>
<tr>
<td></td>
<td>T¯(θ)=Xθ1¯+W¯(θ)F¯(θ)1−ρθ+(m−Xθ1¯)1−ρ+(W¯(θ)+θ)1−ρθ∑i,jFθi¯bj(θ)(μi+bj(θ))∏q(μq2−bj2(θ))∏q≠j(bq2(θ)−bj2(θ)).¯𝑇𝜃¯subscriptsuperscript𝑋1𝜃¯𝑊𝜃¯𝐹𝜃1subscript𝜌𝜃𝑚¯subscriptsuperscript𝑋1𝜃1𝜌¯𝑊𝜃𝜃1subscript𝜌𝜃subscript𝑖𝑗¯superscriptsubscript𝐹𝜃𝑖subscript𝑏𝑗𝜃subscript𝜇𝑖subscript𝑏𝑗𝜃subscriptproduct𝑞superscriptsubscript𝜇𝑞2superscriptsubscript𝑏𝑗2𝜃subscriptproduct𝑞𝑗superscriptsubscript𝑏𝑞2𝜃superscriptsubscript𝑏𝑗2𝜃\displaystyle{\overline{T}(\theta)}=\frac{{\overline{X^{1}_{\theta}}}+\overline{W}(\theta)\overline{F}(\theta)}{1-\rho_{\theta}}+{\frac{(m-{\overline{X^{1}_{\theta}}})}{1-\rho}}+{\frac{(\overline{W}(\theta)+\theta)}{1-\rho_{\theta}}}\sum_{i,j}\frac{\overline{F_{\theta}^{i}}}{b_{j}(\theta)(\mu_{i}+b_{j}(\theta))}\frac{\displaystyle\prod_{q}{(\mu_{q}^{2}-b_{j}^{2}(\theta))}}{\displaystyle\prod_{q\neq j}{(b_{q}^{2}(\theta)-b_{j}^{2}(\theta))}}.</td>
<td></td>
</tr>
</tbody>
</table>
∎
## 4 Conclusion
We study the BPS queueing model, when the job size distribution is hyper-exponential, and we find an analytical expression of the expected conditional response time and for the expected sojourn time. We show that the function of the expected conditional sojourn time in the BPS system with hyper-exponential job size distribution is a concave function with respect to job sizes. We apply the results obtained for the BPS model to the TLPS scheduling mechanism with the hyper-exponential job size distribution and we find the expressions of the expected conditional response time and expected response time for the TLPS model.
## Acknowledgment
I would like to thank K. Avrachenkov and P. Brown for fruitful discussions and suggestions.
## 5 Appendix
Lemma [2](#ThmTheorem2 "Lemma 2. ‣ 2.2 The expected conditional sojourn time for the BPS model ‣ 2 The Analysis of the Batch Arrival Processor Sharing Queue"). The solution of the following system of linear equations
<table>
<tbody>
<tr>
<td></td>
<td>∑jxjμq2−bj2=1,q=1,…,N,formulae-sequencesubscript𝑗subscript𝑥𝑗superscriptsubscript𝜇𝑞2superscriptsubscript𝑏𝑗21𝑞1…𝑁{}\begin{array}[]{l}\displaystyle\sum_{j}\frac{x_{j}}{\mu_{q}^{2}-b_{j}^{2}}=1,\quad q=1,...,N,\end{array}</td>
<td></td>
<td>(17)</td>
</tr>
</tbody>
</table>
is given by:
<table>
<tbody>
<tr>
<td></td>
<td>xk=∏q=1,…,N(μq2−bk2)∏q≠k(bq2−bk2).subscript𝑥𝑘subscriptproduct𝑞1…𝑁superscriptsubscript𝜇𝑞2superscriptsubscript𝑏𝑘2subscriptproduct𝑞𝑘superscriptsubscript𝑏𝑞2superscriptsubscript𝑏𝑘2x_{k}=\frac{\displaystyle\prod_{q=1,...,N}{(\mu_{q}^{2}-b_{k}^{2})}}{\displaystyle\prod_{q\neq k}{(b_{q}^{2}-b_{k}^{2})}}.</td>
<td></td>
<td>(18)</td>
</tr>
</tbody>
</table>
###### *Proof.*.
Let
<table>
<tbody>
<tr>
<td></td>
<td>x=[x1,x2,…,xN]T,𝑥superscriptsubscript𝑥1subscript𝑥2…subscript𝑥𝑁𝑇\displaystyle x=[x_{1},x_{2},...,x_{N}]^{T},</td>
<td></td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td>1¯=[1,1,…,1]T|1×N¯1evaluated-atsuperscript11…1𝑇1𝑁\displaystyle\underline{1}=[1,1,...,1]^{T}|_{1\times N}</td>
<td></td>
</tr>
</tbody>
</table>
be two vectors of size $`N`$,
<table>
<tbody>
<tr>
<td></td>
<td>D=(1μ12−b121μ12−b22…1μ12−bN21μ22−b121μ22−b22…1μ22−bN2…..1μN2−b121μN2−b22…1μN2−bN2)n×N𝐷subscript1superscriptsubscript𝜇12superscriptsubscript𝑏121superscriptsubscript𝜇12superscriptsubscript𝑏22…1superscriptsubscript𝜇12superscriptsubscript𝑏𝑁21superscriptsubscript𝜇22superscriptsubscript𝑏121superscriptsubscript𝜇22superscriptsubscript𝑏22…1superscriptsubscript𝜇22superscriptsubscript𝑏𝑁2…missing-subexpressionmissing-subexpressionmissing-subexpression1superscriptsubscript𝜇𝑁2superscriptsubscript𝑏121superscriptsubscript𝜇𝑁2superscriptsubscript𝑏22…1superscriptsubscript𝜇𝑁2superscriptsubscript𝑏𝑁2𝑛𝑁D=\left(\begin{array}[]{l l l l }\frac{1}{\mu_{1}^{2}-b_{1}^{2}}&\frac{1}{\mu_{1}^{2}-b_{2}^{2}}&...&\frac{1}{\mu_{1}^{2}-b_{N}^{2}}\\ \frac{1}{\mu_{2}^{2}-b_{1}^{2}}&\frac{1}{\mu_{2}^{2}-b_{2}^{2}}&...&\frac{1}{\mu_{2}^{2}-b_{N}^{2}}\\ .....\\ \frac{1}{\mu_{N}^{2}-b_{1}^{2}}&\frac{1}{\mu_{N}^{2}-b_{2}^{2}}&...&\frac{1}{\mu_{N}^{2}-b_{N}^{2}}\\ \end{array}\right)_{n\times N}</td>
<td></td>
</tr>
</tbody>
</table>
be the matrix of size $`N \times N`$. Then equation ([17](#S5.E17 "In 5 Appendix")) could be rewritten as
<table>
<tbody>
<tr>
<td></td>
<td>Dx=1¯.𝐷𝑥¯1\displaystyle Dx=\underline{1}.</td>
<td></td>
</tr>
</tbody>
</table>
Applying the Cramer formulas \[[11](#bib.bib11)\] we obtain:
<table>
<tbody>
<tr>
<td></td>
<td>xk=detDkdetD=det[D[1],…D[k−1],1¯,D[k+1],…D[N]]detD,k=1,…,N,formulae-sequencesubscript𝑥𝑘subscript𝐷𝑘𝐷subscript𝐷delimited-[]1…subscript𝐷delimited-[]𝑘1¯1subscript𝐷delimited-[]𝑘1…subscript𝐷delimited-[]𝑁𝐷𝑘1…𝑁\displaystyle x_{k}=\frac{\det D_{k}}{\det D}=\frac{\det{[D_{[1]},...D_{[k-1]},\underline{1},D_{[k+1]},...D_{[N]}]}}{\det{D}},\quad k=1,...,N,{}</td>
<td></td>
<td>(19)</td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td>Dk=[D[1],…D[k−1],1¯,D[k+1],…D[N]].subscript𝐷𝑘subscript𝐷delimited-[]1…subscript𝐷delimited-[]𝑘1¯1subscript𝐷delimited-[]𝑘1…subscript𝐷delimited-[]𝑁\displaystyle D_{k}=[D_{[1]},...D_{[k-1]},\underline{1},D_{[k+1]},...D_{[N]}].</td>
<td></td>
<td></td>
</tr>
</tbody>
</table>
Since $`D`$ is a Cauchy matrix, its determinant is known \[[11](#bib.bib11)\]:
<table>
<tbody>
<tr>
<td></td>
<td>detD=∏1≤j<k≤N(μj2−μk2)(bk2−bj2)∏j,k=1,…,N(μj2−bk2).𝐷subscriptproduct1𝑗𝑘𝑁superscriptsubscript𝜇𝑗2superscriptsubscript𝜇𝑘2superscriptsubscript𝑏𝑘2superscriptsubscript𝑏𝑗2subscriptproductformulae-sequence𝑗𝑘1…𝑁superscriptsubscript𝜇𝑗2superscriptsubscript𝑏𝑘2{}\displaystyle\det{D}=\frac{\displaystyle\prod_{1\leq j<k\leq N}{(\mu_{j}^{2}-\mu_{k}^{2})(b_{k}^{2}-b_{j}^{2})}}{\displaystyle\prod_{j,k=1,...,N}{(\mu_{j}^{2}-b_{k}^{2})}}\,\,.</td>
<td></td>
<td>(20)</td>
</tr>
</tbody>
</table>
As $`0 < b_{N} < \mu_{N} < \ldots < \mu_{i + 1} < b_{i} < \mu_{i} < \ldots < b_{1} < \mu_{1}`$ , then $`{\det D} > 0`$ and we can use Cramer formulas to calculate $`x_{k}`$. Let us find $`{\det D_{k}} = {\det{\lbrack D_{\lbrack 1\rbrack},{\ldots\hspace{0pt}D_{\lbrack{k - 1}\rbrack}},\underset{¯}{1},D_{\lbrack{k + 1}\rbrack},{\ldots\hspace{0pt}D_{\lbrack N\rbrack}}\rbrack}}`$.
<table>
<tbody>
<tr>
<td></td>
<td>Dk=(1μ12−b12…1μ12−bk−1211μ12−bk+12…1μ12−bN21μ22−b12…1μ22−bk−1211μ22−bk+12…1μ22−bN2…..1μN2−b12…1μN2−bk−1211μN2−bk+12…1μN2−bN2)N×Nsubscript𝐷𝑘subscript1superscriptsubscript𝜇12superscriptsubscript𝑏12…1superscriptsubscript𝜇12superscriptsubscript𝑏𝑘1211superscriptsubscript𝜇12superscriptsubscript𝑏𝑘12…1superscriptsubscript𝜇12superscriptsubscript𝑏𝑁21superscriptsubscript𝜇22superscriptsubscript𝑏12…1superscriptsubscript𝜇22superscriptsubscript𝑏𝑘1211superscriptsubscript𝜇22superscriptsubscript𝑏𝑘12…1superscriptsubscript𝜇22superscriptsubscript𝑏𝑁2…missing-subexpressionmissing-subexpressionmissing-subexpressionmissing-subexpressionmissing-subexpressionmissing-subexpression1superscriptsubscript𝜇𝑁2superscriptsubscript𝑏12…1superscriptsubscript𝜇𝑁2superscriptsubscript𝑏𝑘1211superscriptsubscript𝜇𝑁2superscriptsubscript𝑏𝑘12…1superscriptsubscript𝜇𝑁2superscriptsubscript𝑏𝑁2𝑁𝑁D_{k}=\left(\begin{array}[]{l l l l l l l}\frac{1}{\mu_{1}^{2}-b_{1}^{2}}&...&\frac{1}{\mu_{1}^{2}-b_{k-1}^{2}}&1&\frac{1}{\mu_{1}^{2}-b_{k+1}^{2}}&...&\frac{1}{\mu_{1}^{2}-b_{N}^{2}}\\ \frac{1}{\mu_{2}^{2}-b_{1}^{2}}&...&\frac{1}{\mu_{2}^{2}-b_{k-1}^{2}}&1&\frac{1}{\mu_{2}^{2}-b_{k+1}^{2}}&...&\frac{1}{\mu_{2}^{2}-b_{N}^{2}}\\ .....\\ \frac{1}{\mu_{N}^{2}-b_{1}^{2}}&...&\frac{1}{\mu_{N}^{2}-b_{k-1}^{2}}&1&\frac{1}{\mu_{N}^{2}-b_{k+1}^{2}}&...&\frac{1}{\mu_{N}^{2}-b_{N}^{2}}\\ \end{array}\right)_{{N\times N}}</td>
<td></td>
</tr>
</tbody>
</table>
<table>
<tbody>
<tr>
<td></td>
<td>detDk=|1μ12−b12…1μ12−bk−1211μ12−bk+12…1μ12−bN21μ22−b12…1μ22−bk−1211μ22−bk+12…1μ22−bN2…..1μN2−b12…1μN2−bk−1211μN2−bk+12…1μN2−bN2|N×Nsubscript𝐷𝑘subscript1superscriptsubscript𝜇12superscriptsubscript𝑏12…1superscriptsubscript𝜇12superscriptsubscript𝑏𝑘1211superscriptsubscript𝜇12superscriptsubscript𝑏𝑘12…1superscriptsubscript𝜇12superscriptsubscript𝑏𝑁21superscriptsubscript𝜇22superscriptsubscript𝑏12…1superscriptsubscript𝜇22superscriptsubscript𝑏𝑘1211superscriptsubscript𝜇22superscriptsubscript𝑏𝑘12…1superscriptsubscript𝜇22superscriptsubscript𝑏𝑁2…missing-subexpressionmissing-subexpressionmissing-subexpressionmissing-subexpressionmissing-subexpressionmissing-subexpression1superscriptsubscript𝜇𝑁2superscriptsubscript𝑏12…1superscriptsubscript𝜇𝑁2superscriptsubscript𝑏𝑘1211superscriptsubscript𝜇𝑁2superscriptsubscript𝑏𝑘12…1superscriptsubscript𝜇𝑁2superscriptsubscript𝑏𝑁2𝑁𝑁\det{D_{k}}=\left|\begin{array}[]{l l l l l l l}\frac{1}{\mu_{1}^{2}-b_{1}^{2}}&...&\frac{1}{\mu_{1}^{2}-b_{k-1}^{2}}&1&\frac{1}{\mu_{1}^{2}-b_{k+1}^{2}}&...&\frac{1}{\mu_{1}^{2}-b_{N}^{2}}\\ \frac{1}{\mu_{2}^{2}-b_{1}^{2}}&...&\frac{1}{\mu_{2}^{2}-b_{k-1}^{2}}&1&\frac{1}{\mu_{2}^{2}-b_{k+1}^{2}}&...&\frac{1}{\mu_{2}^{2}-b_{N}^{2}}\\ .....\\ \frac{1}{\mu_{N}^{2}-b_{1}^{2}}&...&\frac{1}{\mu_{N}^{2}-b_{k-1}^{2}}&1&\frac{1}{\mu_{N}^{2}-b_{k+1}^{2}}&...&\frac{1}{\mu_{N}^{2}-b_{N}^{2}}\\ \end{array}\right|_{N\times N}</td>
<td></td>
</tr>
</tbody>
</table>
<table>
<tbody>
<tr>
<td></td>
<td>=(−1)k−1|11μ12−b12…1μ12−bk−121μ12−bk+12…1μ12−bN211μ22−b12…1μ22−bk−121μ22−bk+12…1μ22−bN2…..11μN2−b12…1μN2−bk−121μN2−bk+12…1μN2−bN2|N×Nabsentsuperscript1𝑘1subscript11superscriptsubscript𝜇12superscriptsubscript𝑏12…1superscriptsubscript𝜇12superscriptsubscript𝑏𝑘121superscriptsubscript𝜇12superscriptsubscript𝑏𝑘12…1superscriptsubscript𝜇12superscriptsubscript𝑏𝑁211superscriptsubscript𝜇22superscriptsubscript𝑏12…1superscriptsubscript𝜇22superscriptsubscript𝑏𝑘121superscriptsubscript𝜇22superscriptsubscript𝑏𝑘12…1superscriptsubscript𝜇22superscriptsubscript𝑏𝑁2…missing-subexpressionmissing-subexpressionmissing-subexpressionmissing-subexpressionmissing-subexpressionmissing-subexpression11superscriptsubscript𝜇𝑁2superscriptsubscript𝑏12…1superscriptsubscript𝜇𝑁2superscriptsubscript𝑏𝑘121superscriptsubscript𝜇𝑁2superscriptsubscript𝑏𝑘12…1superscriptsubscript𝜇𝑁2superscriptsubscript𝑏𝑁2𝑁𝑁=(-1)^{k-1}\left|\begin{array}[]{l l l l l l l}1&\frac{1}{\mu_{1}^{2}-b_{1}^{2}}&...&\frac{1}{\mu_{1}^{2}-b_{k-1}^{2}}&\frac{1}{\mu_{1}^{2}-b_{k+1}^{2}}&...&\frac{1}{\mu_{1}^{2}-b_{N}^{2}}\\ 1&\frac{1}{\mu_{2}^{2}-b_{1}^{2}}&...&\frac{1}{\mu_{2}^{2}-b_{k-1}^{2}}&\frac{1}{\mu_{2}^{2}-b_{k+1}^{2}}&...&\frac{1}{\mu_{2}^{2}-b_{N}^{2}}\\ .....\\ 1&\frac{1}{\mu_{N}^{2}-b_{1}^{2}}&...&\frac{1}{\mu_{N}^{2}-b_{k-1}^{2}}&\frac{1}{\mu_{N}^{2}-b_{k+1}^{2}}&...&\frac{1}{\mu_{N}^{2}-b_{N}^{2}}\\ \end{array}\right|_{N\times N}</td>
<td></td>
</tr>
</tbody>
</table>
To simplify the following computations let us introduce the following notations:
<table>
<tbody>
<tr>
<td></td>
<td>βi=−bi−12,i=2,…,k,formulae-sequencesubscript𝛽𝑖superscriptsubscript𝑏𝑖12𝑖2…𝑘\displaystyle\beta_{i}=-b_{i-1}^{2},\quad i=2,...,k,</td>
<td></td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td>βi=−bi2,i=k+1,…,N.formulae-sequencesubscript𝛽𝑖superscriptsubscript𝑏𝑖2𝑖𝑘1…𝑁\displaystyle\beta_{i}=-b_{i}^{2},\quad i=k+1,...,N.</td>
<td></td>
</tr>
</tbody>
</table>
Then, we have
<table>
<tbody>
<tr>
<td></td>
<td>detDk=(−1)k−1|11μ12+β2…1μ12+βk1μ12+βk+1…1μ12+βN11μ22+β2…1μ22+βk1μ22+βk+1…1μ22+βN…..11μN2+β22…1μN2+βk1μN2+βk+1…1μN2+βN|N×Nsubscript𝐷𝑘superscript1𝑘1subscript11superscriptsubscript𝜇12subscript𝛽2…1superscriptsubscript𝜇12subscript𝛽𝑘1superscriptsubscript𝜇12subscript𝛽𝑘1…1superscriptsubscript𝜇12subscript𝛽𝑁11superscriptsubscript𝜇22subscript𝛽2…1superscriptsubscript𝜇22subscript𝛽𝑘1superscriptsubscript𝜇22subscript𝛽𝑘1…1superscriptsubscript𝜇22subscript𝛽𝑁…missing-subexpressionmissing-subexpressionmissing-subexpressionmissing-subexpressionmissing-subexpressionmissing-subexpression11superscriptsubscript𝜇𝑁2superscriptsubscript𝛽22…1superscriptsubscript𝜇𝑁2subscript𝛽𝑘1superscriptsubscript𝜇𝑁2subscript𝛽𝑘1…1superscriptsubscript𝜇𝑁2subscript𝛽𝑁𝑁𝑁\det{D_{k}}=(-1)^{k-1}\left|\begin{array}[]{l l l l l l l}1&\frac{1}{\mu_{1}^{2}+\beta_{2}}&...&\frac{1}{\mu_{1}^{2}+\beta_{k}}&\frac{1}{\mu_{1}^{2}+\beta_{k+1}}&...&\frac{1}{\mu_{1}^{2}+\beta_{N}}\\ 1&\frac{1}{\mu_{2}^{2}+\beta_{2}}&...&\frac{1}{\mu_{2}^{2}+\beta_{k}}&\frac{1}{\mu_{2}^{2}+\beta_{k+1}}&...&\frac{1}{\mu_{2}^{2}+\beta_{N}}\\ .....\\ 1&\frac{1}{\mu_{N}^{2}+\beta_{2}^{2}}&...&\frac{1}{\mu_{N}^{2}+\beta_{k}}&\frac{1}{\mu_{N}^{2}+\beta_{k+1}}&...&\frac{1}{\mu_{N}^{2}+\beta_{N}}\\ \end{array}\right|_{N\times N}</td>
<td></td>
</tr>
</tbody>
</table>
<table>
<tbody>
<tr>
<td></td>
<td>detDk=(−1)k−1|11μ12+β2…1μ12+βk…1μ12+βN0μ12−μ22(μ22+β2)(μ12+β2)…μ12−μ22(μ22+βk)(μ12+βk)…μ12−μ22(μ22+βN)(μ12+βN)…..0μ12−μN2(μN2+β2)(μ12+β2)…μ12−μN2(μN2+βk)(μ12+βk)…μ12−μN2(μN2+βN)(μ12+βN)|N×Nsubscript𝐷𝑘superscript1𝑘1subscript11superscriptsubscript𝜇12subscript𝛽2…1superscriptsubscript𝜇12subscript𝛽𝑘…1superscriptsubscript𝜇12subscript𝛽𝑁0superscriptsubscript𝜇12superscriptsubscript𝜇22superscriptsubscript𝜇22subscript𝛽2superscriptsubscript𝜇12subscript𝛽2…superscriptsubscript𝜇12superscriptsubscript𝜇22superscriptsubscript𝜇22subscript𝛽𝑘superscriptsubscript𝜇12subscript𝛽𝑘…superscriptsubscript𝜇12superscriptsubscript𝜇22superscriptsubscript𝜇22subscript𝛽𝑁superscriptsubscript𝜇12subscript𝛽𝑁…missing-subexpressionmissing-subexpressionmissing-subexpressionmissing-subexpressionmissing-subexpression0superscriptsubscript𝜇12superscriptsubscript𝜇𝑁2superscriptsubscript𝜇𝑁2subscript𝛽2superscriptsubscript𝜇12subscript𝛽2…superscriptsubscript𝜇12superscriptsubscript𝜇𝑁2superscriptsubscript𝜇𝑁2subscript𝛽𝑘superscriptsubscript𝜇12subscript𝛽𝑘…superscriptsubscript𝜇12superscriptsubscript𝜇𝑁2superscriptsubscript𝜇𝑁2subscript𝛽𝑁superscriptsubscript𝜇12subscript𝛽𝑁𝑁𝑁\det{D_{k}}=(-1)^{k-1}\left|\begin{array}[]{l l l l l l}1&\frac{1}{\mu_{1}^{2}+\beta_{2}}&...&\frac{1}{\mu_{1}^{2}+\beta_{k}}&...&\frac{1}{\mu_{1}^{2}+\beta_{N}}\\ 0&\frac{\mu_{1}^{2}-\mu_{2}^{2}}{(\mu_{2}^{2}+\beta_{2})(\mu_{1}^{2}+\beta_{2})}&...&\frac{\mu_{1}^{2}-\mu_{2}^{2}}{(\mu_{2}^{2}+\beta_{k})(\mu_{1}^{2}+\beta_{k})}&...&\frac{\mu_{1}^{2}-\mu_{2}^{2}}{(\mu_{2}^{2}+\beta_{N})(\mu_{1}^{2}+\beta_{N})}\\ .....\\ 0&\frac{\mu_{1}^{2}-\mu_{N}^{2}}{(\mu_{N}^{2}+\beta_{2})(\mu_{1}^{2}+\beta_{2})}&...&\frac{\mu_{1}^{2}-\mu_{N}^{2}}{(\mu_{N}^{2}+\beta_{k})(\mu_{1}^{2}+\beta_{k})}&...&\frac{\mu_{1}^{2}-\mu_{N}^{2}}{(\mu_{N}^{2}+\beta_{N})(\mu_{1}^{2}+\beta_{N})}\\ \end{array}\right|_{N\times N}</td>
<td></td>
</tr>
</tbody>
</table>
<table>
<tbody>
<tr>
<td></td>
<td>=(−1)k−1(μ12−μ22)…(μ12−μN2)(μ12+β2)…(μ12+βN)|1μ22+β2…1μ22+βk1μ22+βk+1…1μ22+βN…..1μN2+β22…1μN2+βk1μN2+βk+1…1μN2+βN|(N−1)×(N−1)absentsuperscript1𝑘1superscriptsubscript𝜇12superscriptsubscript𝜇22…superscriptsubscript𝜇12superscriptsubscript𝜇𝑁2superscriptsubscript𝜇12subscript𝛽2…superscriptsubscript𝜇12subscript𝛽𝑁subscript1superscriptsubscript𝜇22subscript𝛽2…1superscriptsubscript𝜇22subscript𝛽𝑘1superscriptsubscript𝜇22subscript𝛽𝑘1…1superscriptsubscript𝜇22subscript𝛽𝑁…missing-subexpressionmissing-subexpressionmissing-subexpressionmissing-subexpressionmissing-subexpression1superscriptsubscript𝜇𝑁2superscriptsubscript𝛽22…1superscriptsubscript𝜇𝑁2subscript𝛽𝑘1superscriptsubscript𝜇𝑁2subscript𝛽𝑘1…1superscriptsubscript𝜇𝑁2subscript𝛽𝑁𝑁1𝑁1=(-1)^{k-1}\frac{(\mu_{1}^{2}-\mu_{2}^{2})...(\mu_{1}^{2}-\mu_{N}^{2})}{(\mu_{1}^{2}+\beta_{2})...(\mu_{1}^{2}+\beta_{N})}\left|\begin{array}[]{l l l l l l}\frac{1}{\mu_{2}^{2}+\beta_{2}}&...&\frac{1}{\mu_{2}^{2}+\beta_{k}}&\frac{1}{\mu_{2}^{2}+\beta_{k+1}}&...&\frac{1}{\mu_{2}^{2}+\beta_{N}}\\ .....\\ \frac{1}{\mu_{N}^{2}+\beta_{2}^{2}}&...&\frac{1}{\mu_{N}^{2}+\beta_{k}}&\frac{1}{\mu_{N}^{2}+\beta_{k+1}}&...&\frac{1}{\mu_{N}^{2}+\beta_{N}}\\ \end{array}\right|_{(N-1)\times(N-1)}</td>
<td></td>
</tr>
</tbody>
</table>
So, as the above matrix under the sign of determinant is a Cauchy matrix of size $`N - 1`$, the following equation holds:
<table>
<tbody>
<tr>
<td></td>
<td>detDk=(−1)k−1(μ12−μ22)…(μ12−μN2)(μ12+β2)…(μ12+βN)∏2≤j<q≤N(μj2−μq2)(βj−βq)∏j,q=2,…,N(μj2+βq).subscript𝐷𝑘superscript1𝑘1superscriptsubscript𝜇12superscriptsubscript𝜇22…superscriptsubscript𝜇12superscriptsubscript𝜇𝑁2superscriptsubscript𝜇12subscript𝛽2…superscriptsubscript𝜇12subscript𝛽𝑁subscriptproduct2𝑗𝑞𝑁superscriptsubscript𝜇𝑗2superscriptsubscript𝜇𝑞2subscript𝛽𝑗subscript𝛽𝑞subscriptproductformulae-sequence𝑗𝑞2…𝑁superscriptsubscript𝜇𝑗2subscript𝛽𝑞\det{D_{k}}=(-1)^{k-1}\frac{(\mu_{1}^{2}-\mu_{2}^{2})...(\mu_{1}^{2}-\mu_{N}^{2})}{(\mu_{1}^{2}+\beta_{2})...(\mu_{1}^{2}+\beta_{N})}\frac{\displaystyle\prod_{2\leq j<q\leq N}{(\mu_{j}^{2}-\mu_{q}^{2})(\beta_{j}-\beta_{q})}}{\displaystyle\prod_{j,q=2,...,N}{(\mu_{j}^{2}+\beta_{q})}}\,\,.</td>
<td></td>
</tr>
</tbody>
</table>
Let us recall that $`{\beta_{i} = {- b_{i - 1}^{2}}}\quad{{i = {2,k}},{{\beta_{i} = {- b_{i}^{2}}}\quad{i = {{k + 1},\ldots,N}}}}`$, then
<table>
<tbody>
<tr>
<td></td>
<td>detDksubscript𝐷𝑘\displaystyle\det{D_{k}}</td>
<td>=\displaystyle=</td>
<td>(−1)k−1∏q=2,…,N(μ12−μq2)∏q=2,…,N(μ12+βq)∏2≤j<q≤N(μj2−μq2)(βj−βq)∏j,q=2,…,N(μj2+βq)superscript1𝑘1subscriptproduct𝑞2…𝑁superscriptsubscript𝜇12superscriptsubscript𝜇𝑞2subscriptproduct𝑞2…𝑁superscriptsubscript𝜇12subscript𝛽𝑞subscriptproduct2𝑗𝑞𝑁superscriptsubscript𝜇𝑗2superscriptsubscript𝜇𝑞2subscript𝛽𝑗subscript𝛽𝑞subscriptproductformulae-sequence𝑗𝑞2…𝑁superscriptsubscript𝜇𝑗2subscript𝛽𝑞\displaystyle(-1)^{k-1}\frac{\displaystyle\prod_{q=2,...,N}(\mu_{1}^{2}-\mu_{q}^{2})}{\displaystyle\prod_{q=2,...,N}(\mu_{1}^{2}+\beta_{q})}\frac{\displaystyle\prod_{2\leq j<q\leq N}{(\mu_{j}^{2}-\mu_{q}^{2})(\beta_{j}-\beta_{q})}}{\displaystyle\prod_{j,q=2,...,N}{(\mu_{j}^{2}+\beta_{q})}}</td>
<td></td>
</tr>
<tr>
<td></td>
<td></td>
<td>=\displaystyle=</td>
<td>(−1)k−1∏1≤j<q≤N(μj2−μq2)∏2≤j<q≤N(βj−βq)∏j=1,…,N,q=2,…,N(μj2+βq)superscript1𝑘1subscriptproduct1𝑗𝑞𝑁superscriptsubscript𝜇𝑗2superscriptsubscript𝜇𝑞2subscriptproduct2𝑗𝑞𝑁subscript𝛽𝑗subscript𝛽𝑞subscriptproductformulae-sequence𝑗1…𝑁𝑞2…𝑁superscriptsubscript𝜇𝑗2subscript𝛽𝑞\displaystyle(-1)^{k-1}\frac{\displaystyle\prod_{1\leq j<q\leq N}{(\mu_{j}^{2}-\mu_{q}^{2})\prod_{2\leq j<q\leq N}{(\beta_{j}-\beta_{q})}}}{\displaystyle\prod_{j=1,...,N,q=2,...,N}{(\mu_{j}^{2}+\beta_{q})}}</td>
<td></td>
</tr>
<tr>
<td></td>
<td></td>
<td>=\displaystyle=</td>
<td>(−1)k−1∏1≤j<q≤N(μj2−μq2)∏1≤j<q≤N,j,q≠k(bq2−bj2)∏j,q=1,…,N,q≠k(μj2−bq2)superscript1𝑘1subscriptproduct1𝑗𝑞𝑁superscriptsubscript𝜇𝑗2superscriptsubscript𝜇𝑞2subscriptproductformulae-sequence1𝑗𝑞𝑁𝑗𝑞𝑘superscriptsubscript𝑏𝑞2superscriptsubscript𝑏𝑗2subscriptproductformulae-sequence𝑗𝑞1…𝑁𝑞𝑘superscriptsubscript𝜇𝑗2superscriptsubscript𝑏𝑞2\displaystyle(-1)^{k-1}\frac{\displaystyle\prod_{1\leq j<q\leq N}{(\mu_{j}^{2}-\mu_{q}^{2})\prod_{1\leq j<q\leq N,\,j,q\neq k}(b_{q}^{2}-b_{j}^{2})}}{\displaystyle\prod_{j,q=1,...,N,q\neq k}{(\mu_{j}^{2}-b_{q}^{2})}}</td>
<td></td>
</tr>
<tr>
<td></td>
<td></td>
<td>=\displaystyle=</td>
<td>(−1)k−1∏1≤j<q≤N(μj2−μq2)∏1≤j<q≤N(bq2−bj2)∏j,q=1,…,N,q≠k(μj2−bq2)(−1)k−1∏q=1,…,N,q≠k(bq2−bk2)superscript1𝑘1subscriptproduct1𝑗𝑞𝑁superscriptsubscript𝜇𝑗2superscriptsubscript𝜇𝑞2subscriptproduct1𝑗𝑞𝑁superscriptsubscript𝑏𝑞2superscriptsubscript𝑏𝑗2subscriptproductformulae-sequence𝑗𝑞1…𝑁𝑞𝑘superscriptsubscript𝜇𝑗2superscriptsubscript𝑏𝑞2superscript1𝑘1subscriptproductformulae-sequence𝑞1…𝑁𝑞𝑘superscriptsubscript𝑏𝑞2superscriptsubscript𝑏𝑘2\displaystyle(-1)^{k-1}\frac{\displaystyle\prod_{1\leq j<q\leq N}{(\mu_{j}^{2}-\mu_{q}^{2})\prod_{1\leq j<q\leq N}(b_{q}^{2}-b_{j}^{2})}}{\displaystyle\prod_{j,q=1,...,N,q\neq k}{(\mu_{j}^{2}-b_{q}^{2})}(-1)^{k-1}\displaystyle\prod_{q=1,...,N,\,q\neq k}{(b_{q}^{2}-b_{k}^{2})}}</td>
<td></td>
</tr>
<tr>
<td></td>
<td></td>
<td>=\displaystyle=</td>
<td>∏1≤j<q≤N(μj2−μq2)(bq2−bj2)∏j=1,…,N(μj2−bk2)∏j,q=1,…,N(μj2−bq2)∏q=1,…,N,q≠k(bq2−bk2).subscriptproduct1𝑗𝑞𝑁superscriptsubscript𝜇𝑗2superscriptsubscript𝜇𝑞2superscriptsubscript𝑏𝑞2superscriptsubscript𝑏𝑗2subscriptproduct𝑗1…𝑁superscriptsubscript𝜇𝑗2superscriptsubscript𝑏𝑘2subscriptproductformulae-sequence𝑗𝑞1…𝑁superscriptsubscript𝜇𝑗2superscriptsubscript𝑏𝑞2subscriptproductformulae-sequence𝑞1…𝑁𝑞𝑘superscriptsubscript𝑏𝑞2superscriptsubscript𝑏𝑘2\displaystyle\frac{\displaystyle\prod_{1\leq j<q\leq N}{(\mu_{j}^{2}-\mu_{q}^{2})(b_{q}^{2}-b_{j}^{2})}\displaystyle\prod_{j=1,...,N}{(\mu_{j}^{2}-b_{k}^{2})}}{\displaystyle\prod_{j,q=1,...,N}{(\mu_{j}^{2}-b_{q}^{2})}\displaystyle\prod_{q=1,...,N,\,q\neq k}{(b_{q}^{2}-b_{k}^{2})}}.</td>
<td></td>
</tr>
</tbody>
</table>
Then, from ([19](#S5.E19 "In Proof.. ‣ 5 Appendix")) and ([20](#S5.E20 "In Proof.. ‣ 5 Appendix")), we have the following expressions for $`x_{k}`$:
<table>
<tbody>
<tr>
<td></td>
<td>xk=∏q=1,…,N(μq2−bk2)∏q=1,…,N,q≠k(bq2−bk2),subscript𝑥𝑘subscriptproduct𝑞1…𝑁superscriptsubscript𝜇𝑞2superscriptsubscript𝑏𝑘2subscriptproductformulae-sequence𝑞1…𝑁𝑞𝑘superscriptsubscript𝑏𝑞2superscriptsubscript𝑏𝑘2x_{k}=\frac{\displaystyle\prod_{q=1,...,N}{(\mu_{q}^{2}-b_{k}^{2})}}{\displaystyle\prod_{q=1,...,N,\,q\neq k}{(b_{q}^{2}-b_{k}^{2})}},</td>
<td></td>
</tr>
</tbody>
</table>
what proves Lemma [2](#ThmTheorem2 "Lemma 2. ‣ 2.2 The expected conditional sojourn time for the BPS model ‣ 2 The Analysis of the Batch Arrival Processor Sharing Queue"). ∎
## References
- \[1\] S. Aalto and U. Ayesta, “Mean delay analysis of multilevel processor sharing disciplines”, In Proceedings of IEEE INFOCOM 2006.
- \[2\] K. Avrachenkov, U. Ayesta, P. Brown, “Batch Arrival Processor-Sharing with Application to Multi-Level Processor-Sharing Scheduling”, Queueing Systems 50, pp.459-480, 2005.
- \[3\] K. Avrachenkov, P. Brown, N. Osipova, “Optimal choice of Threshold in Two Level Processor Sharing”, preprint. Also available as INRIA Research Report RR-6215, http://hal.inria.fr/inria-00153966/
- \[4\] N. Bansal. “Analysis of the M/G/1 Processor-Sharing Queue with Bulk Arrivals”, Operation Research Letters, 31(5), September 2003, pp.401-405.
- \[5\] Kleinrock, L., R. R. Muntz, and E. Rodemich, “The Processor-Sharing Queueing Model for Time-Shared Systems with Bulk Arrivals,” Networks Journal,1,1-13 (1971).
- \[6\]
L. Kleinrock, Queueing systems, vol. 2, John Wiley and Sons, 1976.
- \[7\] G. Fayolle, I. Mitrani, and R. Iasnogorodski, “Sharing a processor among many job classes”, Journal of the ACM, 27(3):519-532, 1980.
- \[8\] H. Feng and V. Misra, “Asymptotic bounds for $`M^{X}/G/1`$ processor sharing queues”, Technical report CUCS-00-04, Columbia University, jul.2003
- \[9\]
A. Feldmann, W. Whitt, “Fitting mixtures of exponentials to long-tail distributions to analyze network performance models”, Performance Evaluation, v.31, 1998, pp.245-258.
- \[10\] J. Kim and B. Kim, “Concavity of the conditional mean sojourn time in the M/G/1 processor-sharing queue with batch arrivals”, preprint.
- \[11\] A.G. Kurosh, “Higher algebra”, MIR, 1972.
- \[12\] K.M. Rege and B. Sengupta, “The M/G/1 Processor-Sharing Queue with Bulk Arrivals”. In Proceedings of Modelling and Evaluation of ATM Networks 1993, pp.417-432.
###### Contents
1. [](#S1)
1 Introduction
2. [](#S2)
2 The Analysis of the Batch Arrival Processor Sharing Queue
1. [](#S2.SS1 "In 2 The Analysis of the Batch Arrival Processor Sharing Queue")
2.1 Main definitions
2. [](#S2.SS2 "In 2 The Analysis of the Batch Arrival Processor Sharing Queue")
2.2 The expected conditional sojourn time for the BPS model
3. [](#S3)
3 The Analysis of the Two Level Processor Sharing Model
1. [](#S3.SS1 "In 3 The Analysis of the Two Level Processor Sharing Model")
3.1 Main definitions
2. [](#S3.SS2 "In 3 The Analysis of the Two Level Processor Sharing Model")
3.2 The expected conditional sojourn time for the TLPS model
3. [](#S3.SS3 "In 3 The Analysis of the Two Level Processor Sharing Model")
3.3 The expected sojourn time for the TLPS model
4. [](#S4)
4 Conclusion
5. [](#S5)
5 Appendix |
0706.1395 | cs.NI | cyber | # Opportunistic Network Coding for Video Streaming over Wireless
Hulya Seferoglu, Athina Markopoulou
Electrical Engineering and Computer Science Dept.
University of California, Irvine
Irvine, CA, USA
{hseferog,athina}@uci.edu
###### Abstract
In this paper, we study video streaming over wireless networks with network coding capabilities. We build upon recent work, which demonstrated that network coding can increase throughput over a broadcast medium, by mixing packets from different flows into a single packet, thus increasing the information content per transmission. Our key insight is that, when the transmitted flows are video streams, network codes should be selected so as to maximize not only the network throughput but also the video quality. We propose video-aware opportunistic network coding schemes that take into account both (i) the decodability of network codes by several receivers and (ii) the importance and deadlines of video packets. Simulation results show that our schemes significantly improve both video quality and throughput.
###### Index Terms:
Video streaming, wireless networks, mesh networks, network coding.
## I Introduction
Providing high quality video over wireless networks is a challenging problem, due to both the erratic and time-varying nature of a wireless channel and the stringent delivery requirements of media traffic. Developments in video compression and streaming, wireless networking, and cross-layer design, are continuously advancing the state-of-the art in wireless video \[[1](#bib.bib1), [2](#bib.bib2)\]. In this paper, we propose a novel technique for video streaming in a wireless environment inspired by the emerging paradigm of network coding \[[3](#bib.bib3), [4](#bib.bib4)\].
Our work builds on recent work in \[[6](#bib.bib6), [5](#bib.bib5)\] that used network coding to improve throughput in a wireless mesh network. In particular, \[[6](#bib.bib6), [5](#bib.bib5)\] proposed that wireless routers mix packets from different flows, so as to increase the information content of each -broadcast- transmission and therefore the throughput for data applications. In this paper, we build on this idea, and propose a network coding and scheduling scheme for transmitting several video streams over a wireless mesh network.
Our key insight is that the transmission of video streams in a network coding-capable wireless network should be optimized not only for network throughput but also, and more importantly, for video quality. The fact that video packets have unequal importance is well understood and extensively studied in the video streaming community, e.g. for rate-distortion optimized streaming \[[7](#bib.bib7), [8](#bib.bib8), [9](#bib.bib9)\]. The fact that mixing different information flows can increase throughput in multicast networks is well understood in the network coding community \[[3](#bib.bib3), [4](#bib.bib4), [10](#bib.bib10), [11](#bib.bib11)\]. Our work bridges the gap between the two approaches, and proposes a new video-aware scheme for network coding and packet scheduling that improves both aspects, namely video quality and throughput.
We consider a wireless mesh network, in which routers can mix different incoming flows/streams, using simple network coding operations (XOR). The resulting network code is broadcasted to the neighborhood of the router. Nodes in the same neighborhood listen to each other’s transmission and store overheard packets; these are used later to decode received coded packets and also to construct new coded packets. The core question in this architecture is how to select the best -according to an appropriate metric- network code for transmission among all possible codes. In \[[6](#bib.bib6), [5](#bib.bib5)\], a transmitting node chooses a network code that can be decoded by several neighbors at the same time slot; this policy increases the utility of each transmission thus leading to throughput benefits. However, when the transmitted flows are video streams, this is not necessarily the best choice. Video quality can be improved by intelligently selecting network codes that combine those video packets that are decodable by several neighbors but also contribute the most to video quality. In other words, when video streams are transmitted, it is not only the quantity but also the quality/content of information transferred that should be taken into account in the selection of network codes. In this paper, we develop schemes for network code selection and packet scheduling that take into account both (i) the importance and deadlines of video packets and (ii) the network state and the received/overheard packets in the neighborhood.
The paper is organized as follows. Section [II](#S2 "II Related Work ‣ Opportunistic Network Coding for Video Streaming over Wireless") discusses related work. Section [III](#S3 "III System Overview ‣ Opportunistic Network Coding for Video Streaming over Wireless") gives an overview of the system model. Section [IV](#S4 "IV Coding Algorithms ‣ Opportunistic Network Coding for Video Streaming over Wireless") presents the algorithms for network coding. Section [V](#S5 "V Performance Evaluation ‣ Opportunistic Network Coding for Video Streaming over Wireless") presents simulation results that demonstrate the benefits of the proposed algorithms over baseline schemes, in terms of video quality and application-level throughput. Section [VI](#S6 "VI Discussion ‣ Opportunistic Network Coding for Video Streaming over Wireless") discusses open issues and ongoing work. Section [VII](#S7 "VII Conclusion ‣ Opportunistic Network Coding for Video Streaming over Wireless") concludes the paper.
## II Related Work
This work combines ideas and techniques from two bodies of work: video streaming and network coding.
Several network-adaptive techniques have been proposed to support streaming media over unreliable and/or time-varying networks \[[12](#bib.bib12)\]. Supporting video over wireless is particularly challenging due to the limited, time-varying resources of the wireless channel \[[1](#bib.bib1), [2](#bib.bib2)\]. There is a large body of work on cross-layer design for video over wireless, including \[[13](#bib.bib13), [14](#bib.bib14), [15](#bib.bib15), [16](#bib.bib16), [17](#bib.bib17)\]. Packet scheduling is an important control at the medium access control layer. The problem of rate-distortion optimized packet scheduling has been studied in the RaDiO family of techniques \[[7](#bib.bib7), [9](#bib.bib9), [8](#bib.bib8)\]: in every transmission opportunity, media units are selected for transmission so as to maximize the expected quality of received video subject to a constraint in the transmission rate, and taking into account transmission errors, delays and decoding dependencies. Cross-layer approaches exploit the fact that packets in a video stream have different importance and therefore should be treated differently by network mechanisms.
Independently, the network coding paradigm has emerged from the pioneering work in \[[10](#bib.bib10), [11](#bib.bib11)\], which showed that, in multicast networks where intermediate nodes do simple linear operations on incoming packets, one can achieve the min-cut throughput of the network to each receiver. The linearly combined packets can be utilized at the receivers to recover the original packets by solving a set of linear equations over a finite field. This breakthrough idea inspired significant effort in several directions \[[3](#bib.bib3), [4](#bib.bib4)\], including studying topologies beyond multicast, such as unicast \[[18](#bib.bib18), [19](#bib.bib19), [20](#bib.bib20)\] and broadcast scenarios. The broadcast nature of the wireless medium offers an opportunity for exploiting the throughput benefits of network coding \[[21](#bib.bib21), [22](#bib.bib22)\]. The recent work in \[[6](#bib.bib6), [5](#bib.bib5)\] applied these ideas from the network coding community in the context of wireless mesh networks. \[[5](#bib.bib5)\] implemented a pseudo-broadcast mechanism for 802.11 together with opportunistic listening and a coding layer between IP and MAC that is used to detect coding opportunities and pack packets from different flows into a single transmission, thus increasing network throughput.
Our paper introduces a novel technique for video streaming over wireless that combines the above two approaches. On one hand, we build on \[[6](#bib.bib6), [5](#bib.bib5)\] to exploit the broadcast nature of the wireless medium and use network coding to pack several packets from different streams into a single code for transmission, thus increasing throughput. On the other hand, we construct and select network codes taking into account the importance of video packets (in terms of video distortion and playout deadlines) within the same stream, as well as their contribution to the total throughput and video quality. This combined approach allows us to achieve significant video quality improvement while still maintaining the throughput benefits.
## III System Overview
[Refer to caption](/html/0706.1395/assets/x1.png)
Figure 1: A wireless mesh network
We consider video streaming over wireless mesh networks where intermediate nodes (wireless mesh routers) are able to forward packets to other intermediate nodes and/or clients, as shown in Fig. [1](#S3.F1 "Figure 1 ‣ III System Overview ‣ Opportunistic Network Coding for Video Streaming over Wireless"). In this paper, we propose algorithms that can be used at the intermediate node to maximize video quality and throughput. We assume that intermediate nodes can perform simple network coding operations (bit-wise XOR) and combine packets from several incoming streams into a single outgoing packet. This packet is broadcasted to the entire neighborhood, thus reaching several nodes at the same time. We assume that nodes can overhear all transmissions in their neighborhood, whether they are intended for them or not; they can decode a network-coded packet using overheard packets. The idea of combining network coding with broadcast to increase the information content per transmission, is well understood in the network coding community. This idea has been recently applied in 802.11-based multi-hop wireless networks and throughput benefits have been demonstrated for data applications \[[6](#bib.bib6), [5](#bib.bib5)\].
Our key observation is that, when the transmitted flows are video streams, this is not necessarily the best choice and video quality must also be considered. The importance and deadlines of video packets must be taken into account to intelligently select those network codes that contribute the most to the quality of video streams. In this paper, we develop schemes for network coding across different flows, and packet selection within each flow, to improve both video quality and throughput.
Code Selection at the Intermediate Node. Let us consider an intermediate node that receives $`N`$ packets from different video streams and forwards them to $`N`$ nodes in its neighborhood. The intermediate node maintains a transmission (Tx) queue with incoming video packets. At a given time slot a packet is selected from the Tx queue for transmission. The selected packet is called the primary packet and its destination node is called the target node. Depending on the network coding scheme, the primary packet may be the first packet from the head of the queue, or any packet in Tx queue that is marked as active (i.e. not transmitted within the last round-trip time). In addition to the primary packet, all packets in the queue are considered as candidate side packets, i.e. candidates for a transmission in the same time slot together with the primary packet. The primary and the side packets are all XOR-ed together into a single packet, called the network code.
<sup>1</sup>
<sup>1</sup>1The primary packet can be thought as the main packet we try to transmit during a time slot; this is the packet that would be normally transmitted by a FIFO policy. Side packets are opportunistically transmitted together with the primary packet; they are useful to nodes other than the target node.
The core question then is:
> which network code (i.e. XOR of the primary and side packets) to select and transmit so as to maximize the total video quality and throughput.
The algorithms addressing this question are the main part of this paper, and will be discussed separately in the next section ([IV](#S4 "IV Coding Algorithms ‣ Opportunistic Network Coding for Video Streaming over Wireless")). In the rest of this section, we describe the remaining components and functions of the system. The terminology is summarized in Table [I](#S3.T1 "TABLE I ‣ III System Overview ‣ Opportunistic Network Coding for Video Streaming over Wireless").
TABLE I: Terminology
<table>
<thead>
<tr>
<th>Term</th>
<th>Definition</th>
</tr>
</thead>
<tbody>
<tr>
<td>Primary Packet</td>
<td>The packet selected from the Tx queue before network coding. It must be included in all network codes. It can be thought as the main packet we try to transmit in a given time-slot.</td>
</tr>
<tr>
<td>Side Packet</td>
<td>Packet in the Tx queue, other than the primary, included in the network code.</td>
</tr>
<tr>
<td>Active Packet</td>
<td>Packet in the Tx queue that can be considered as primary. (Not transmitted within the last RTT.)</td>
</tr>
<tr>
<td>Inactive packet</td>
<td>Packet in the Tx queue that cannot be considered as primary. (It has already been transmitted within the last RTT, and the acknowledgement is still pending.)</td>
</tr>
<tr>
<td>Network Code</td>
<td>The primary and side packets XOR-ed together into a single packet.</td>
</tr>
<tr>
<td>Target Node</td>
<td>The intended recipient of the primary packet.</td>
</tr>
<tr>
<td>Tx Queue</td>
<td>The output queue of the transmitting node.</td>
</tr>
<tr>
<td>Rx Buffer</td>
<td>The receive queue of the receiving node. It stores received packets, destined to this node.</td>
</tr>
<tr>
<td>Virtual Buffer</td>
<td>Also maintained at a receiving node. It stores overheard packets, destined to other nodes.</td>
</tr>
</tbody>
</table>
ACKs and Other Functions at the Receiving Nodes. Once the network code is chosen, it is broadcasted to all nodes in the neighborhood. Depending on the channel conditions, some nodes successfully receive it. When the target node receives it, it decodes it (which is guaranteed by the construction of the network code in the next section), stores the primary packet in its receive (Rx) buffer, and sends an acknowledgement (ACK) back to the intermediate node. Nodes, other than the target node, overhear the transmitted packet and try to decode it; if they overhear a new packet destined to them, they store it in their Rx buffer and send an ACK back to the intermediate node; if they obtain a packet destined for another node, they store it in their virtual buffer. An overheard packet stays in the virtual buffer until an ACK from the target is overheard or until its deadline expires.
Retransmissions and Active/Inactive Packets at the Intermediate Node. The intermediate node waits for a mean round-trip time (RTT) from the time it transmits the network code until it receives an ACK. During that period, all packets that were part of the code stay in the Tx queue but are marked as inactive. Inactive packets are not considered for primary transmission (in order to avoid unnecessary duplicate transmissions) but are still considered as candidates for side packets (to increase coding opportunities). When the transmitter receives an ACK, it removes the corresponding packet from the Tx queue. If an RTT expires without receiving an ACK, the packet is marked as active again and the process is repeated. A packet stays in the Tx queue until either it is successfully transmitted or its deadline expires; when either of these occur, the packet is removed from the transmission buffer.
Notice that considering any active packet as primary, better utilizes the bandwidth but may lead to reordering. Although this may be a concern for TCP, it is clearly better for video that requires timely delivery and can reorder packets at the playout buffer.
Requirements. We assume the following capabilities available at our system. First, broadcast is needed to harvest the benefits of network coding. Although wireless is inherently a broadcast medium, this may be hidden by some communication protocols. We assume that some broadcast capability is available, e.g. 802.11 broadcast or pseudo-broadcast as implemented in \[[6](#bib.bib6), [5](#bib.bib5)\]. Second, nodes need to know the contents of the virtual buffers of all their neighbors, in order to code. In our simulations we assume perfect knowledge of the contents of all virtual buffers. This can be achieved by exchanging and guessing this information, as in \[[6](#bib.bib6), [5](#bib.bib5)\]; in practice, there will be some error and a slight degradation in the overall performance. Third, nodes must be capable of coding/decoding in real time, which is a realistic assumption for simple (bit-wise XOR) operations. Finally, we assume a Tx queue with video packets only.
## IV Coding Algorithms
The main questions in this system have to do with the construction and selection of network codes. The code construction problem is concerned with finding candidate codes that guarantee decodability by the target and several other nodes. The code selection problem is concerned with finding the best among the candidate codes to optimize video quality and throughput. The first proposed algorithm, NCV, achieves the same throughput gains as in \[[5](#bib.bib5)\] but also intelligently chooses the packets to improve video quality. The second algorithm, NCVD, uses NCV as a building block but considers more coding options, thus further improving video quality and throughput.
###
IV-A
NCV Algorithm: Network Coding for Video
Assume that there are several video streams coming to an intermediate node that can be mixed together. Depending on the content of virtual buffers at the clients, there may be several combinations of these streams, i.e. several network coding opportunities. The main idea behind the Network Coding for Video (NCV) algorithm is to select the best network code to improve video quality. Let us demonstrate this idea through an example.
######
Example 1
Consider the example shown in Fig. [1](#S3.F1 "Figure 1 ‣ III System Overview ‣ Opportunistic Network Coding for Video Streaming over Wireless") and let us focus on a single-hop shown in more detail in Fig. [2](#S4.F2 "Figure 2 ‣ IV-A NCV Algorithm: Network Coding for Video ‣ IV Coding Algorithms ‣ Opportunistic Network Coding for Video Streaming over Wireless"). Node $`I`$ receives three independent video streams, e.g. from the Internet through the gateway, destined to its neighbors $`A,B,C`$. $`I`$ maintains a FIFO Tx queue that stores packets $`\{ A_{1},A_{2},\ldots\}`$ destined to node $`A`$, $`\{ B_{1},B_{2},\ldots\}`$ destined to node $`B`$, and $`\{ C_{1},C_{2},\ldots\}`$ destined to node $`C`$. Fig. [2](#S4.F2 "Figure 2 ‣ IV-A NCV Algorithm: Network Coding for Video ‣ IV Coding Algorithms ‣ Opportunistic Network Coding for Video Streaming over Wireless") also shows the contents of the virtual buffers at each client: node $`A`$ has overheard packets $`\{ B_{1},C_{1}\}`$ and nodes $`B`$ and $`C`$ have both overheard packet $`A_{1}`$, from previous transmissions. $`A_{1}`$ is the first active packet from head of the queue and is selected as the primary packet. Any packet (active or inactive) in the output queue, other than $`A_{1}`$, can be chosen as a side packet, on the condition that the constructed network code should be decoded at node $`A`$, i.e. $`A_{1}`$ can be retrieved. To satisfy this condition, side packets that will be used in the network code should already be available at node $`A`$; in other words, the decodability of a network code depends on the overheard packets at node $`A`$. Network codes $`c_{1} = A_{1}`$, $`c_{2} = {A_{1}\hspace{0pt}{\bigoplus B_{1}}}`$, $`c_{3} = {A_{1}\hspace{0pt}{\bigoplus C_{1}}}`$, and $`c_{4} = {A_{1}\hspace{0pt}{\bigoplus{B_{1}\hspace{0pt}{\bigoplus C_{1}}}}}`$ can all be decoded by $`A`$ and thus are eligible network codes. $`\square`$
[Refer to caption](/html/0706.1395/assets/x2.png)
Figure 2: Example of Network Coding for Video (NCV)
The Code Construction Problem. More generally, consider an intermediate node $`I`$ with $`N`$ clients $`\{ n_{1},n_{2},\ldots,n_{N}\}`$. There are $`\Phi`$ packets in the Tx queue, which are denoted with $`\{ p_{1},p_{2},\ldots,p_{\Phi}\}`$. Choose the first active packet, $`p_{i}`$, from the head of the Tx FIFO queue as the primary packet with target node $`n_{m}`$. $`I`$ will construct and transmit a network code, which consists of $`p_{i}`$ XOR-ed together with some side packets. The network code should be constructed so as to guarantee decodability of the primary packet at its target node $`n_{m}`$. For $`p_{i}`$ to be decodable at $`n_{m}`$, all $`n - 1`$ side packets must be among the overheard packets at $`n_{m}`$. Assume that $`\Psi_{m}`$ packets are overheard at node $`n_{m}`$ and denoted by $`\{\nu_{m,1},\nu_{m,2},\ldots,\nu_{m,\Psi_{m}}\}`$. Therefore, the candidate network codes are:
<table>
<tbody>
<tr>
<td></td>
<td>cki={pi}⋃Skm,k=1,2,…,2Ψmformulae-sequencesuperscriptsubscript𝑐𝑘𝑖subscript𝑝𝑖superscriptsubscript𝑆𝑘𝑚𝑘12…superscript2subscriptΨ𝑚c_{k}^{i}=\{p_{i}\}\bigcup S_{k}^{m},k=1,2,...,2^{\Psi_{m}}</td>
<td></td>
</tr>
</tbody>
</table>
where $`S_{k}^{m}`$ is the $`k^{t\hspace{0pt}h}`$ subset of $`\{\nu_{m,1},\nu_{m,2},\ldots,\nu_{m,\Psi_{m}}\}`$. Note that, since linear operations are limited to bit-wise XOR, a network code $`p_{1}\hspace{0pt}{\bigoplus{p_{2}\hspace{0pt}{\bigoplus{\ldots\hspace{0pt}{\bigoplus p_{k}}}}}}`$ is completely specified by the set of packets $`\{ p_{1},p_{2},\ldots,p_{k}\}`$ that are XOR-ed together. The complexity of considering all possible network codes is discussed in section [](#S6.SS2 "VI-B Complexity Analysis ‣ VI Discussion ‣ Opportunistic Network Coding for Video Streaming over Wireless")
VI-B
. The next step, is to select the best among all candidate codes.
Example 1 Continued. Node $`A`$ can get packet $`A_{1}`$ from all possible network codes. Codes $`c_{2}`$ and $`c_{3}`$ improve the video quality at node sets $`\{ A,B\}`$ and $`\{ A,C\}`$, respectively. It is clear that $`c_{2}`$ and $`c_{3}`$ are better codes than $`c_{1}`$ and $`c_{4}`$ both for throughput (they are useful to two instead of one node) and video quality. Comparing $`c_{2}`$ to $`c_{3}`$, we observe that they are equivalent in terms of throughput but they may contribute differently to video quality depending on the content of video packets $`A_{1},B_{1},C_{1}`$. Deciding which candidate code to select between $`c_{2} = {A_{1}\hspace{0pt}{\bigoplus B_{1}}}`$ and $`c_{3} = {A_{1}\hspace{0pt}{\bigoplus C_{1}}}`$ should depend on the importance and urgency of the original video packets $`B_{1}`$ and $`C_{1}`$. NCV exploits this observation. $`\square`$
Algorithm 1
The NCV Algorithm
1:
Initialization: $`I_{m\hspace{0pt}a\hspace{0pt}x}^{i} = 0`$, $`c_{m\hspace{0pt}a\hspace{0pt}x}^{i} = \varnothing`$
2:
Choose the first head-of-queue active packet as primary $`p_{i}`$.
3:
Let $`n_{m}`$ be the target node of $`p_{i}`$. Let $`\left\{ \nu_{m,1},\ldots,\nu_{m,\Psi_{m}} \right\}`$ be the overheard packets at $`n_{m}`$.
4:
for $`k = {1\hspace{0pt}\ldots\hspace{0pt}2^{\Psi_{m}}}`$ do
5:
$`c_{k}^{i} = {\left\{ p_{i} \right\}\hspace{0pt}\left. \bigcup S_{k}^{m} \right.}`$
6:
Calculate $`I_{k}^{i}`$ with Eq. (<a href="#S4.E3" style="font-size:80%;" title="In IV-A NCV Algorithm: Network Coding for Video ‣ IV Coding Algorithms ‣ Opportunistic Network Coding for Video Streaming over Wireless">3</a>)
7:
if $`I_{k}^{i} > I_{m\hspace{0pt}a\hspace{0pt}x}^{i}`$ then
8:
$`I_{m\hspace{0pt}a\hspace{0pt}x}^{i} = I_{k}^{i}`$, $`c_{m\hspace{0pt}a\hspace{0pt}x}^{i} = c_{k}^{i}`$
9:
end if
10:
end for
11:
Choose $`c_{m\hspace{0pt}a\hspace{0pt}x}^{i}`$ as the network code. XOR all packets and transmit
The Code Selection Problem. In order to choose the best code, we first need to define a metric that captures the contribution of each candidate code to video quality improvement. Assume that $`p_{i}`$ is the primary packet targeted to node $`n_{m}`$, and $`{\{ c_{k}^{i}\}}_{k = 1}^{k = 2^{\Psi_{m}}}`$ are all the candidate codes. Let $`I_{k}^{i}\hspace{0pt}{(n_{m})}`$ be the improvement in video quality at client $`n_{m}`$, when $`c_{k}^{i}`$ is received and decoded:
<table>
<tbody>
<tr>
<td></td>
<td>Iki(nm)=∑l=1Lk(1−elk)Δlkglkdlk,superscriptsubscript𝐼𝑘𝑖subscript𝑛𝑚superscriptsubscript𝑙1subscript𝐿𝑘1superscriptsubscript𝑒𝑙𝑘superscriptsubscriptΔ𝑙𝑘superscriptsubscript𝑔𝑙𝑘superscriptsubscript𝑑𝑙𝑘I_{k}^{i}(n_{m})=\sum_{l=1}^{L_{k}}(1-e_{l}^{k})\Delta_{l}^{k}g_{l}^{k}d_{l}^{k},</td>
<td></td>
<td>(1)</td>
</tr>
</tbody>
</table>
where each factor in this formula is defined as follows:
- •
$`L_{k}`$ is the number of original packets included in network code $`c_{k}^{i}`$.
<sup>2</sup>
<sup>2</sup>2Notice that at most one out of these $`L_{k}`$ packets can be useful to a particular node $`n_{m}`$, but different packets are useful to different nodes.
- •
$`d_{l}^{k}`$ and $`g_{l}^{k}`$ are indicator functions that express whether code $`k`$ is useful for node $`n_{m}`$. We define $`d_{l}^{k} = 1`$ if $`c_{k}^{i}`$ is decodable at node $`n_{m}`$, or 0 otherwise. We define $`g_{l}^{k} = 1`$ if packet $`l`$ is targeted to node $`n_{m}`$, or $`0`$ otherwise.
- •
$`\Delta_{l}^{k}`$ is the improvement in video quality (SNR) if packet $`l`$ is received correctly and on time at client $`n_{m}`$. To compute $`\Delta_{l}^{k}`$, we decode the entire video sequence with this packet missing and we compute the resulting distortion.
<sup>3</sup>
<sup>3</sup>3This is an approximation as the actual distortion that may also depend on the delivery status of prior and subsequent NALs. The distortion model can be extended to capture these loss correlations \[[23](#bib.bib23)\]. Furthermore, we assume that distortions caused by loss of multiple packets are additive, which is reasonable for sparse losses. These approximations reduce the computational complexity by separating the total distortion function into a set of individual packet distortion functions and optimizing for each one of them.
- •
$`e_{l}^{k}`$ is the loss probability of packet $`l`$ due to channel errors or latency:
<table>
<tbody>
<tr>
<td></td>
<td>elk=∫τ∞pF(t)𝑑t+(1−∫τ∞pF(t)𝑑t)εF(s).superscriptsubscript𝑒𝑙𝑘superscriptsubscript𝜏subscript𝑝𝐹𝑡differential-d𝑡1superscriptsubscript𝜏subscript𝑝𝐹𝑡differential-d𝑡subscript𝜀𝐹𝑠e_{l}^{k}=\int_{\tau}^{\infty}p_{F}(t)dt+\left(1-\int_{\tau}^{\infty}p_{F}(t)dt\right)\varepsilon_{F}(s).</td>
<td></td>
<td>(2)</td>
</tr>
</tbody>
</table>
The first part in Eq.([2](#S4.E2 "In 4th item ‣ IV-A NCV Algorithm: Network Coding for Video ‣ IV Coding Algorithms ‣ Opportunistic Network Coding for Video Streaming over Wireless")) describes the probability of a packet arriving late; $`\tau`$ is the remaining time until the playout deadline and $`p_{F}\hspace{0pt}{(t)}`$ is the distribution of the forward-trip time. The second part describes the loss probability (of a packet that is still on time) due to effects of the wireless channel, such as noise, fading, interference, etc; $`\varepsilon_{F}\hspace{0pt}{(s)}`$ is the loss probability at state $`s`$ of the channel.
After defining the contribution of code $`c_{k}^{i}`$ to a single node $`n_{m}`$, $`I_{k}^{i}\hspace{0pt}{(n_{m})}`$, we define the total video quality improvement of code $`c_{k}^{i}`$ as the sum of the video quality improvements at all clients $`n_{1},{\ldots\hspace{0pt}n_{N}}`$,
<sup>4</sup>
<sup>4</sup>4If the quality of the encoded video sequences to different clients $`n_{m}`$ are significantly different from each other, then the terms $`I_{k}^{i}\hspace{0pt}{(n_{m})}`$ should be normalized by the average PSNR per sequence, before adding them up.
due to code $`c_{k}^{i}`$:
<table>
<tbody>
<tr>
<td></td>
<td>Iki=∑m=1NIki(nm),superscriptsubscript𝐼𝑘𝑖superscriptsubscript𝑚1𝑁superscriptsubscript𝐼𝑘𝑖subscript𝑛𝑚I_{k}^{i}=\sum_{m=1}^{N}I_{k}^{i}(n_{m}),</td>
<td></td>
<td>(3)</td>
</tr>
</tbody>
</table>
The NCV algorithm is summarized in Alg. ([1](#alg1 "Algorithm 1 ‣ IV-A NCV Algorithm: Network Coding for Video ‣ IV Coding Algorithms ‣ Opportunistic Network Coding for Video Streaming over Wireless")). At each time slot, the NCV algorithm chooses the primary packet $`p_{i}`$ and constructs all candidate network codes $`{}c_{k{k = 1}}^{i{k = 2^{\Psi_{m}}}}`$. Among all candidate network codes, NCV chooses the code that maximizes the total video quality improvement:
<table>
<tbody>
<tr>
<td></td>
<td>maxkIkisubscript𝑘superscriptsubscript𝐼𝑘𝑖\max_{k}I_{k}^{i}</td>
<td></td>
<td>(4)</td>
</tr>
</tbody>
</table>
Notice that, depending on the contents of the virtual buffers, it is possible that no side packets can be used together with a given primary packet $`p_{i}`$. In that case, the network code is simply $`{{\{ p_{i}\}} \cup \varnothing} = {\{ p_{i}\}}`$, and we transmit only the primary packet alone.
###
IV-B
NCVD Algorithm: looking into the queue in Depth
[Refer to caption](/html/0706.1395/assets/x3.png)
Figure 3: Example of NCVD
As described in the previous section, NCV selects the primary packet from the head of the queue but ignoring packets marked as “inactive”. This limits the candidate codes to those that are decodable for this single primary packet. The second algorithm, NCVD looks into the entire Tx queue (“in depth”) and considers all, not just the head-of-line, packet as candidates for the primary packet, thus increasing the options for candidate codes, which eventually leads to a better choice for the metric of interest. Note that a different set of candidate codes can be constructed for each primary packet. Let us explain NCVD through the following example.
######
Example 2
Let us look at Fig. [3](#S4.F3 "Figure 3 ‣ IV-B NCVD Algorithm: looking into the queue in Depth ‣ IV Coding Algorithms ‣ Opportunistic Network Coding for Video Streaming over Wireless"). The topology is the same as in Fig. [2](#S4.F2 "Figure 2 ‣ IV-A NCV Algorithm: Network Coding for Video ‣ IV Coding Algorithms ‣ Opportunistic Network Coding for Video Streaming over Wireless"), but the contents of the Tx queue and of the virtual buffers are different. Assume that all packets are active packets, i.e. they can all be considered as primary. One option is to select the head-of-line packet $`A_{1}`$ as the primary packet. As discussed in Example 1, the best codes for this primary packet are $`c_{3} = {A_{1}\hspace{0pt}{\bigoplus{C\hspace{0pt}1}}}`$ or $`c_{4} = {A_{1}\hspace{0pt}{\bigoplus{B_{1}\hspace{0pt}{\bigoplus C_{1}}}}}`$. A different choice is to select $`B_{1}`$ as the primary packet, which leads to completely different set of candidate network codes (listed on the Fig. [3](#S4.F3 "Figure 3 ‣ IV-B NCVD Algorithm: looking into the queue in Depth ‣ IV Coding Algorithms ‣ Opportunistic Network Coding for Video Streaming over Wireless")). Code $`c_{4}' = {B_{1}\hspace{0pt}{\bigoplus{C_{1}\hspace{0pt}{\bigoplus A_{2}}}}}`$ achieves the maximum throughput improvement, and potentially the maximum video quality achievement, depending on the importance and urgency of all packets. This example demonstrates that increasing our options of primary packet, increases the set of candidate codes, and thus can potentially improve both throughput and video quality. $`\square`$
More generally, NCVD constructs candidate codes $`{{c_{k}^{i},k} = 1},{2,\ldots,2^{\Psi_{m}}}`$ for each candidate primary packet $`p_{i}`$ in the Tx queue. Among all constructed codes, NCVD selects the code that maximizes the total improvement in video quality for all clients:
<table>
<tbody>
<tr>
<td></td>
<td>maxpimaxk(Iki),subscriptsubscript𝑝𝑖subscript𝑘superscriptsubscript𝐼𝑘𝑖\max_{p_{i}}\max_{k}(I_{k}^{i}),</td>
<td></td>
<td>(5)</td>
</tr>
</tbody>
</table>
Algorithm [2](#alg2 "Algorithm 2 ‣ IV-B NCVD Algorithm: looking into the queue in Depth ‣ IV Coding Algorithms ‣ Opportunistic Network Coding for Video Streaming over Wireless") summarizes NCVD.
NCVD can be parameterized by the depth $`d`$ of the Tx queue, that is considered in the selection of the primary packet. NCVD($`d = 1`$) is simply NCV, while NCVD($`d = \infty`$) considers all packets in the Tx queue. The larger the value of $`d`$, the more coding options, the better the performance of NCVD. Because queue sizes are small for real time applications, we can focus on NCVD($`d = \infty`$), which we simply refer to as NCVD.
Algorithm 2
The NCVD Algorithm
1:
Initialization: $`c_{m\hspace{0pt}a\hspace{0pt}x} = \varnothing`$, $`I_{m\hspace{0pt}a\hspace{0pt}x} = 0`$
2:
for every packet $`i = 1,\ldots,,\Phi`$ from the head of Tx queue do
3:
Consider this packet, $`p_{i}`$, as candidate for primary
4:
Construct all possible codes $`c_{k}^{i}`$ for $`p_{i}`$
5:
Determine the max improvement $`I_{m\hspace{0pt}a\hspace{0pt}x}^{i} = {\max_{k}I_{k}^{i}}`$
6:
and the corresponding code $`c_{k}^{i}`$: $`k = {a\hspace{0pt}r\hspace{0pt}g\hspace{0pt}m\hspace{0pt}a\hspace{0pt}x\hspace{0pt}I_{k}^{i}}`$ as in NCV
7:
if $`I_{m\hspace{0pt}a\hspace{0pt}x}^{i} > I_{m\hspace{0pt}a\hspace{0pt}x}`$ then
8:
$`I_{m\hspace{0pt}a\hspace{0pt}x} = I_{m\hspace{0pt}a\hspace{0pt}x}^{i}`$, $`c_{m\hspace{0pt}a\hspace{0pt}x} = c_{k}^{i}`$
9:
end if
10:
end for
11:
Choose $`c_{m\hspace{0pt}a\hspace{0pt}x}`$ as the network code. XOR all packets and transmit.
## V Performance Evaluation
In this section, we evaluate the performance of the proposed schemes (NCV and NCVD) in terms of video quality and network throughput under different conditions. We compare them to two baseline schemes, one without coding (noNC) and one with network coding optimized for throughput (NCT) as in \[[5](#bib.bib5)\]. Simulation results show that NCV and NCVD can significantly improve video quality and application-level throughout, without compromising MAC-level throughput.
###
V-A
Simulation Setup
Let us first describe the communication scenario, wireless channel models and video sequences used, and the baselines algorithms for comparison.
####
V-A1
Single-Hop Scenario
In this paper, we focus on the evaluation of the single-hop scenario shown in Fig. [2](#S4.F2 "Figure 2 ‣ IV-A NCV Algorithm: Network Coding for Video ‣ IV Coding Algorithms ‣ Opportunistic Network Coding for Video Streaming over Wireless"): the intermediate node $`I`$ receives different video streams which it forwards downstream towards their destinations $`A,B,C`$. $`I`$ can apply different schemes for network coding and packet scheduling in the downlink. The downlink rate is selected to be 300 kbps, and each video stream has a rate of 70 kbps. We assume that the three video streams are the only ones using the downlink, hence, there is no congestion. However, packets may still be lost due to error on the wireless channel, according to models described in the next paragraph, and can also experience a random delay, which we assume to be exponentially distributed with $`4\hspace{0pt}m\hspace{0pt}s`$ average. The delay budget (playout deadline) for this single hop is set to $`100\hspace{0pt}m\hspace{0pt}s`$, to allow for queueing, transmission, propagation and a few retransmissions. We also perform simulations for different delay budgets, from 50ms to 200ms. ACKs and retransmissions are used to combat errors, as explained in the system overview.
####
V-A2
Wireless Channel Model
We consider two channel models to evaluate our algorithms in different conditions. In both cases, packet loss is independent and identically distributed across the three sessions (from $`I`$ to $`A`$, from $`I`$ to $`B`$, and from $`I`$ to $`C`$). Below we specify the packet loss model within a session.
Model I: Packets transmitted in each link experience i.i.d. loss. with a fixed loss probability. We simulated a wide range of effective packet loss rates from 1% up to 20%. The effective loss rate depends on the use of retransmissions, FEC and other mechanisms that mask the error rate on the channel.
Model II: A flat Rayleigh fading channel is modeled by a finite-state Markov chain (MC), where the fading is approximated as a discrete time Markov process with time discretized typically to the channel coherence time \[[24](#bib.bib24)\]. The set of all possible fading gains (equivalently, SNR or BER levels) is modeled as the states of the MC, and the channel variations are modeled as the transitions between these states that occur at each interval with certain probabilities. We used a two-state (Gilbert-Elliot) model, characterized by the following parameters specified in \[[25](#bib.bib25)\]: (i) the bit-error-rate (BER) of each state, which depends on the channel signal-to-noise (SNR) level (ii) the state duration and transition probabilities, which depend on the channel coherence time, and in turn on the speed of mobiles and the channel frequency. We assume nomadic or pedestrian wireless clients with 3 km/h speed, and 2.4 GHz channel frequency, which results in an average coherence time of 21 ms. This duration ensures that the channel remains static during a packet transmission. In our experiments, we consider different channel quality levels assigning average SNR levels from the set $`\{ 3,5,7,9\}`$ dB, resulting in effective packet loss rates from $`1\%`$ to $`35\%`$.
####
V-A3
Video Sequences
As our test sequence, we used standard sequences: *Carphone*, *Foreman*, *Mother & Daughter*. These were QCIF sequences encoded using the JM 8.6 version of the H.264/AVC codec \[[26](#bib.bib26), [27](#bib.bib27)\]. The group of pictures consisted of I and P frames, one I every 10 frames. All encoded sequences had data rates of 70 kbps each and frame rate of 30 fps. Each frame consists of at least one slice. Each slice was packetized into an independent NAL (network abstraction layer) unit, of size 250B. There are two reasons for this choice of packet size. First, due to the time varying nature of wireless channels, it is preferred to have short packet transmissions to avoid the channel variation during a packet s transmission. Second, using a fixed size (on the average) simplifies the network coding (bit-wise XOR) operations requiring small or no padding of the packets.
The average peak-signal-to-noise (PSNR) ratio for the encoded sequences, Carphone, Foreman and Mother& Daughter, were $`29.95\hspace{0pt}d\hspace{0pt}B`$, $`28.70\hspace{0pt}d\hspace{0pt}B`$ and $`40.74\hspace{0pt}d\hspace{0pt}B`$ respectively; these PSNR values, of the encoded sequences before transmission, are denoted as “No Error” in the first row of Table [II](#S5.T2 "TABLE II ‣ V-B1 Video Quality Improvements ‣ V-B Simulation Results ‣ V Performance Evaluation ‣ Opportunistic Network Coding for Video Streaming over Wireless") and correspond to the top lines of Fig. [4](#S5.F4 "Figure 4 ‣ V-B1 Video Quality Improvements ‣ V-B Simulation Results ‣ V Performance Evaluation ‣ Opportunistic Network Coding for Video Streaming over Wireless") . We repeated and concatenated the standard sequences to create longer test sequences of duration 30sec each.
We simulated packet loss by erasing the corresponding NAL units from the RTP stream, according to the packet traces produced by the network simulation. At the receiver side, we decoded the remaining RTP stream with standard error concealment enabled. When an entire frame was lost, we used copy-concealment.
####
V-A4
Algorithms under Comparison
We compare our algorithms, NCV and NCVD, against two baseline algorithms for packet scheduling: no Network Coding (noNC) and Network Coding for Throughput (NCT), which are described next.
No Network Coding (noNC): This is a FIFO Tx queue without network coding. Consider again Example [1](#Thmexample1 "Example 1 ‣ IV-A NCV Algorithm: Network Coding for Video ‣ IV Coding Algorithms ‣ Opportunistic Network Coding for Video Streaming over Wireless") and Fig. [2](#S4.F2 "Figure 2 ‣ IV-A NCV Algorithm: Network Coding for Video ‣ IV Coding Algorithms ‣ Opportunistic Network Coding for Video Streaming over Wireless"): node $`I`$ stores packets for all three streams destined to nodes $`A,B,C`$. In every time slot, $`I`$ transmits the first packet from the head of the queue. It may require several consecutive retransmissions until the head-of-line packet is successfully transmitted; in 802.11, there is an upper limit in the maximum number of allowed retransmissions. In order to conduct a fair comparison with our schemes, we slightly modify (improve) this scenario by using the same scheme as described in section [III](#S3 "III System Overview ‣ Opportunistic Network Coding for Video Streaming over Wireless").
<sup>5</sup>
<sup>5</sup>5The summary of the scheme is as follows. After transmission, a packet is marked as *inactive* and is not transmitted as primary for a time duration of a mean RTT; during that period, other packets are transmitted from the FIFO as primary, thus better utilizing the channel. After an RTT, if an ACK is still not received, the packet is marked as *active* and considered again for transmission. Packets whose playout deadlines have expired are removed from the Tx queue.
Network Coding for Throughput (NCT): This is an improved version of the algorithm proposed in \[[5](#bib.bib5)\]. The packet transmission mechanism is the same as in the noNC scheme, but network coding is used to maximize throughput, as follows. The packet at the head of the Tx queue is selected as a primary packet; side packets are chosen to be XOR-ed together with the primary packet so as to construct a network code that is useful to the maximum number of receivers possible at this time slot.
There are two improvements in NCT compared to the coding algorithm in \[[5](#bib.bib5)\] that allow NCT to achieve even higher throughput than \[[5](#bib.bib5)\]. First, NCT follows the same ACK and retransmission mechanism described in section [III](#S3 "III System Overview ‣ Opportunistic Network Coding for Video Streaming over Wireless") and repeated in Footnote [5](#footnote5 "footnote 5 ‣ V-A4 Algorithms under Comparison ‣ V-A Simulation Setup ‣ V Performance Evaluation ‣ Opportunistic Network Coding for Video Streaming over Wireless"): packets with pending acknowledgments are marked as inactive for one RTT, while the channel is used to transmit other packets as primary. In \[[5](#bib.bib5)\] and in general MAC retransmissions, a packet stays at the head of the queue blocking other packets, until it goes through successfully or it exceeds the maximum number of retransmissions. Head-of-line blocking avoids reordering, which may be an issue for TCP traffic. However, in the context of video streaming, the playout buffer can handle reordering. Another difference is that NCT uses an improved version of the coding procedure in \[[5](#bib.bib5)\]: NCT considers all possible subsets of the candidate side packets thus maximizing the number of receivers that can decode; while \[[5](#bib.bib5)\] considers side packets in a sequential order, thus sacrificing some throughput for reduced complexity. Therefore, we use NCT as our baseline for the maximum achievable throughput per transmission using network coding.
The main difference between NCV/NCVD and NCT is that our schemes select side packets to maximize video quality while NCT maximizes throughput. A secondary difference, is that we consider all packets in a queue as candidates for side packets, while NCT, consistently with \[[5](#bib.bib5)\], considers only the earliest packet per flow.
###
V-B
Simulation Results
In this section, we present simulation results that compare our schemes to the baselines and demonstrate that NCV and NCVD can improve video quality and application-level throughout, without compromising MAC-level throughput. We report simulation results for the single-hop scenario of Fig. [2](#S4.F2 "Figure 2 ‣ IV-A NCV Algorithm: Network Coding for Video ‣ IV Coding Algorithms ‣ Opportunistic Network Coding for Video Streaming over Wireless"), when node $`I`$ streams sequences *Carphone*, *Foreman*, and *Mother and Daughter* streamed to clients $`A,B,C`$, respectively.
####
V-B1
Video Quality Improvements
Fig. [4](#S5.F4 "Figure 4 ‣ V-B1 Video Quality Improvements ‣ V-B Simulation Results ‣ V Performance Evaluation ‣ Opportunistic Network Coding for Video Streaming over Wireless") shows the video quality experienced by the clients (PSNR over frame number for parts of the sequences) for the four algorithms under comparison, namely noNC, NCT, NCV, NCVD, as well as for the encoded sequences before transmission (noError). The simulation is performed for wireless channel *Model I* at packet loss rate of $`9.4\%`$ with 100 ms delay budget; for comparison, the same wireless channel trace is used as input to all 4 algorithms. As expected, there are time periods, during which the channel is bad, the quality degrades for all algorithms. However, the degradation for NCV and NCVD is much less than for NCT and noNC, because NCV and NCVD select network codes to protect and deliver the most important packets on time, thus improving the video quality; in contrast, NCT and noNC treat all packets similarly.
The average PSNR for each sequence and algorithm is summarized in Table [II](#S5.T2 "TABLE II ‣ V-B1 Video Quality Improvements ‣ V-B Simulation Results ‣ V Performance Evaluation ‣ Opportunistic Network Coding for Video Streaming over Wireless"). As expected, the noNC scheme performs poorly. NCT improves over noNC because it delivers more packets per time slot. NCV improves over NCT because it chooses the most important video packets; although the number of packets does not increase over NCT, their quality does. NCVD further improves over NCV because it considers more candidate codes and opportunities. These numbers are compared to the original encoded sequence (NoError).
<table>
<tbody>
<tr>
<td></td>
</tr>
</tbody>
</table>
[Refer to caption](/html/0706.1395/assets/x4.png)
\(a\)
Carphone
[Refer to caption](/html/0706.1395/assets/x5.png)
\(b\)
Foreman
[Refer to caption](/html/0706.1395/assets/x6.png)
\(c\)
Mother & Daughter
Figure 4: PSNR per frame for (parts of) the test sequences, for channel Model I with packet loss $`9.4\%`$ and delay budget 100ms. Four schemes (noNC, NCT, NCV, NVCD) are compared. “NoError” refers to the encoded sequences without any loss. The average PSNR values (averaged over the entire sequence) are summarized in Table [II](#S5.T2 "TABLE II ‣ V-B1 Video Quality Improvements ‣ V-B Simulation Results ‣ V Performance Evaluation ‣ Opportunistic Network Coding for Video Streaming over Wireless")).
TABLE II: Average PSNR for the scenario of Fig.[4](#S5.F4 "Figure 4 ‣ V-B1 Video Quality Improvements ‣ V-B Simulation Results ‣ V Performance Evaluation ‣ Opportunistic Network Coding for Video Streaming over Wireless") ($`70\hspace{0pt}k\hspace{0pt}b\hspace{0pt}p\hspace{0pt}s`$ video rate, channel Model I with $`9.4\%`$ loss, $`100\hspace{0pt}m\hspace{0pt}s`$ playout deadline)
<table>
<thead>
<tr>
<th>avg PSNR (dB)</th>
<th>Carphone</th>
<th>Foreman</th>
<th>Mother&Daughter</th>
</tr>
</thead>
<tbody>
<tr>
<th>No Error</th>
<td>29.95</td>
<td>28.70</td>
<td>40.74</td>
</tr>
<tr>
<th>NCVD</th>
<td>26.32</td>
<td>26.08</td>
<td>32.87</td>
</tr>
<tr>
<th>NCV</th>
<td>23.99</td>
<td>25.03</td>
<td>32.62</td>
</tr>
<tr>
<th>NCT</th>
<td>22.40</td>
<td>22.76</td>
<td>30.81</td>
</tr>
<tr>
<th>noNC</th>
<td>22.08</td>
<td>21.59</td>
<td>26.92</td>
</tr>
</tbody>
</table>
Fig. [5](#S5.F5 "Figure 5 ‣ V-B1 Video Quality Improvements ‣ V-B Simulation Results ‣ V Performance Evaluation ‣ Opportunistic Network Coding for Video Streaming over Wireless") focuses on the client that receives the sequence *Foreman*. The same scenario as in Fig. [4](#S5.F4 "Figure 4 ‣ V-B1 Video Quality Improvements ‣ V-B Simulation Results ‣ V Performance Evaluation ‣ Opportunistic Network Coding for Video Streaming over Wireless") is considered, but with loss rates varying from 1% to 20%. Fig. [5](#S5.F5 "Figure 5 ‣ V-B1 Video Quality Improvements ‣ V-B Simulation Results ‣ V Performance Evaluation ‣ Opportunistic Network Coding for Video Streaming over Wireless") shows the average PSNR for each value of packet loss rate and for each algorithm. Clearly, NCV and NCVD outperform NCT (by $`2.5 - 3.5`$dB) and noNC (up to $`3.5 - 5`$ dB) for all packet loss rates. Another observation (from this and other figures omitted for lack of space) is that the PSNR gain of NCV and NCVD is larger for medium than for very low and very high packet loss rates. For low loss rates, most packets are transmitted successfully, while for high loss rates most packets are lost. In both cases, the number of network coding opportunities decrease. However, even then, the proposed algorithms still achieve a considerable PSNR improvement. The upper part of Fig. [5](#S5.F5 "Figure 5 ‣ V-B1 Video Quality Improvements ‣ V-B Simulation Results ‣ V Performance Evaluation ‣ Opportunistic Network Coding for Video Streaming over Wireless") shows the video quality for the Foreman sequence, and the lower part shows the PSNR averaged across all sequences.
Fig. [6](#S5.F6 "Figure 6 ‣ V-B1 Video Quality Improvements ‣ V-B Simulation Results ‣ V Performance Evaluation ‣ Opportunistic Network Coding for Video Streaming over Wireless") evaluates the same scenario as in Fig.[5](#S5.F5 "Figure 5 ‣ V-B1 Video Quality Improvements ‣ V-B Simulation Results ‣ V Performance Evaluation ‣ Opportunistic Network Coding for Video Streaming over Wireless"), but for the second wireless channel *Model II*. This is the two-state model, parameterized by the SNR levels (resulting in loss rates from 1% to 35%). The compared algorithms are ranked similar to the previous case and the PSNR improvements from NCV and NCVD are still high. However, there is less improvement compared to channel *Model I*. The difference can be explained by the network coding and code selection opportunities. In *Model I*, consecutive transmissions to the same client experience independent loss, and links to different clients are independent from each other. This independence results in all virtual buffers having roughly the same number of packets, over short time periods. However, in *Model II*, the channel alternates between a good and a bad state; while in a bad state, a client is more likely to experience consecutive losses, while clients in a good state are more likely to receive consecutive successful packets. This results in an unbalanced number of packets in the virtual buffers. Since network code construction directly depends on the number of packets in virtual buffers, there are less network coding opportunities, hence less network code selection possibilities. Even when there are less opportunities, we still observe significant quality improvement: NCV and NCVD improves up to 2 and 3dB over NCT and up to 3 and 4dB over noNC, respectively.
[Refer to caption](/html/0706.1395/assets/x7.png)
[Refer to caption](/html/0706.1395/assets/x8.png)
Figure 5: Video quality observed over wireless channel *Model I*. Average PSNR for *Foreman*, and Average PSNR, averaged across all three sequences.
[Refer to caption](/html/0706.1395/assets/x9.png)
[Refer to caption](/html/0706.1395/assets/x10.png)
Figure 6: Video quality observed over wireless channel *Model II*. Average PSNR for *Foreman*, Average PSNR, averaged across all three sequences.
In the previous scenarios, we considered a delay budget of 100ms. We now consider channel Model I at $`9.4\%`$ packet loss rate, with delay values ranging from $`50`$ to $`200`$ms. Fig. [7](#S5.F7 "Figure 7 ‣ V-B1 Video Quality Improvements ‣ V-B Simulation Results ‣ V Performance Evaluation ‣ Opportunistic Network Coding for Video Streaming over Wireless") depicts the PSNR values for each scenario (first averaged over each sequence and then across the three sequences, to summarize the overall improvement). The figure shows that both NCV and NCVD improves video quality for the entire range of delay values. The improvement is smaller for a tight delay budget, because a tight delay constraint limits the number of retransmissions and the lifetime of packets both at the Tx queue and the virtual buffers, thus decreasing network coding and selection opportunities. However, even with tight delay constraints, there is significant video quality improvement from NCV and NCVD compared to NCT and noNC.
[Refer to caption](/html/0706.1395/assets/x11.png)
Figure 7: PSNR values (averaged over each sequence and across sequences) for different delay budgets. Wireless Model I is considered.
####
V-B2
Throughput Improvements
<table>
<tbody>
<tr>
<td></td>
</tr>
</tbody>
</table>
[Refer to caption](/html/0706.1395/assets/x12.png)
\(a\)
Application-level throughput over channel Model I
[Refer to caption](/html/0706.1395/assets/x13.png)
\(b\)
Application-level throughput over channel Model II
[Refer to caption](/html/0706.1395/assets/x14.png)
\(c\)
MAC-level throughput over channel Model I
Figure 8: Total throughput (added over all three streams) achieved by the four algorithms.
The video-aware schemes improve video quality because they explicitly take it into account in the code selection. In this section, we show that, our schemes also significantly improve application-level throughput.
Application throughput. Fig. [8](#S5.F8 "Figure 8 ‣ V-B2 Throughput Improvements ‣ V-B Simulation Results ‣ V Performance Evaluation ‣ Opportunistic Network Coding for Video Streaming over Wireless")(a) and (b) show the total throughput as seen by the application-layer (i.e. NAL units per sec) added over all clients. The figure clearly shows that NCV and NCVD achieve higher throughput as compared to NCT and noNC. The main reason is that NCV and NCVD do not select codes consisting of packets whose deadlines are within one transmission time, while NCT and noNC transmit all packets. Late packets does not contribute to application-level throughput, because those packets are discarded at the client even if they are received successfully. NCV and NCVD transmit more useful packets to the client, even though the number of transmitted packets may be comparable to NCT. Application-level throughput is the most relevant notion of throughput in our context, as in \[[5](#bib.bib5)\].
MAC Throughput. For completeness, we also show that the MAC-layer throughput of NCV and NCVD is very close to that of NCT, which is specifically designed for maximum throughput. Fig. [8](#S5.F8 "Figure 8 ‣ V-B2 Throughput Improvements ‣ V-B Simulation Results ‣ V Performance Evaluation ‣ Opportunistic Network Coding for Video Streaming over Wireless")(c) shows the MAC-level throughput for the four schemes. As expected, NCT, NCV and NCVD all achieve higher MAC-level throughput than noNC; this is because all three schemes use network coding across streams, which increases the information content per transmission. The second and more interesting observation is that NCV and NCT achieve almost the same MAC level throughput, for loss rates up to $`9.4\%`$, and NCT achieves slightly higher throughput for larger loss rates. The reason is that NCV and NCT use the same network code construction schemes but different network code selection mechanisms; therefore, the amount of data delivered at the MAC layer is almost the same for low packet loss rates. For higher packet loss rates, there are more packets with urgent deadline in the Tx queue. NCT transmits all possible side packets even if the their deadline is within one transmission time. NCV selects both primary and side packets considering the deadline and importance; therefore, NCV may purposely not choose the code with the largest number of packets, if one or more packets in this code are going to be useless at the receiver. Instead, NCV sends the most useful and effective packets to the client. In summary, NCV achieves the same MAC-level throughput with NCT for loss rates up to 9.4%, and slightly less for higher loss rates. However, and more importantly, NCV always achieves higher application-level throughput than NCT.
NCVD. A third observation from Fig.[8](#S5.F8 "Figure 8 ‣ V-B2 Throughput Improvements ‣ V-B Simulation Results ‣ V Performance Evaluation ‣ Opportunistic Network Coding for Video Streaming over Wireless")(c) is that NCVD achieves not only the highest application-level throughput but also the highest MAC throughput for most loss rates. This is explained by the fact that NCVD looks into the entire queue and has more options to choose from, both in terms of video quality and in terms of absolute number of packets. For very high loss rates (above $`20\%`$) NCT achieves higher MAC throughput because NCVD prefers to optimize the code selection for the application level.
Small Queue Sizes. We finally looked at the buffer occupancy at the transmitter and at the clients and observed that they were really small, in the order of 5-10 packets, for the simulated scenarios and delay budgets considered. Furthermore, we observed that NCV and NCVD further reduces the queue size, compared to NCT and noNC. This is intuitively expected: NCV and NCVD deliver more packets successfully to the client, thus there are less packets waiting for transmission in Tx. Maintaining short buffers has several positive implications. First, short queues are good for bottlenecks shared with TCP. Second, having a small Tx queue allows to consider all packets in the queue for network codes and enjoy the performance gains of NCVD without increase in complexity, even for $`d = \infty`$. Finally, having small virtual buffers means that there are only a few overheard packets to consider in the construction of the network codes, which significantly decreases the complexity.
## VI Discussion
###
VI-A
Video Quality and Throughput
Our metric is the improvement in video quality for every single node, in Eq. ([1](#S4.E1 "In IV-A NCV Algorithm: Network Coding for Video ‣ IV Coding Algorithms ‣ Opportunistic Network Coding for Video Streaming over Wireless")), added over all nodes, in Eq. ([3](#S4.E3 "In IV-A NCV Algorithm: Network Coding for Video ‣ IV Coding Algorithms ‣ Opportunistic Network Coding for Video Streaming over Wireless")). Although, throughput is not explicitly mentioned, it is implicitly captured. E.g. if we omit the distortion term $`{({1 - e_{l}^{k}})}\hspace{0pt}\Delta_{l}^{k}`$ from Eq. ([1](#S4.E1 "In IV-A NCV Algorithm: Network Coding for Video ‣ IV Coding Algorithms ‣ Opportunistic Network Coding for Video Streaming over Wireless")), Eq. ([3](#S4.E3 "In IV-A NCV Algorithm: Network Coding for Video ‣ IV Coding Algorithms ‣ Opportunistic Network Coding for Video Streaming over Wireless")) simply becomes $`I_{k}^{i} = {\sum_{m = 1}^{N}{\sum_{l = 1}^{L_{k}}{g_{l}^{k}\hspace{0pt}d_{l}^{k}}}}`$, which counts the absolute number of packets delivered in a single transmission, i.e. throughput, similarly to NCT and to \[[5](#bib.bib5)\]. Therefore, our metric successfully captures both the number and the quality of packets in a code. Furthermore, our schemes choose “better-quality” packets and improve the application-level throughput, as already shown in the performance evaluation.
###
VI-B
Complexity Analysis
NCV (in Alg. [1](#alg1 "Algorithm 1 ‣ IV-A NCV Algorithm: Network Coding for Video ‣ IV Coding Algorithms ‣ Opportunistic Network Coding for Video Streaming over Wireless")) constructs candidate network codes and selects the one that maximizes video quality improvement. The main complexity comes from considering all possible candidate codes. Selecting side packets among all possible subsets of overheard packets at the target node, is clearly exponential in the size of the virtual buffer. However, the complexity of NCV is no worse than the complexity of NCT: they both consider all possible codes but they evaluate them using a different metric. Another important observation is that real-time delay requirements significantly reduce the number of packets in the virtual buffers and therefore the complexity; e.g. in our simulations, a delay budget of 100ms resulted in at most $`5`$ packets in the virtual buffers. Thus, the brute-force approach is feasible for real-time applications.
For a larger delay budget, approximation algorithms for NCV and NCT can be developed as follows. Recall that in each time slot, first we pick the primary packet, then we look at all overheard packets at the target client as candidate side packets. Conflicts between side packets can be represented using a graph, whose vertices represent the overheard packets and edges represent their conflict. Each vertex has a weight corresponding to the expected improvement from including this packet into the code. Two vertices are connected through an edge if the corresponding packets cannot be used together in the same code. The problem of selecting side packets to maximize quality improvement is then reduced to selecting vertices that are not connected to maximize the total weight. This problem is known as the maximum weight independent set problem and is the complement of the vertex cover problem. Although these problems are NP-complete, they are also well-studied, and approximation algorithms can be found in the literature, \[[28](#bib.bib28)\].
Finally, NCVD (Alg. [2](#alg2 "Algorithm 2 ‣ IV-B NCVD Algorithm: looking into the queue in Depth ‣ IV Coding Algorithms ‣ Opportunistic Network Coding for Video Streaming over Wireless")) runs NCV for each packet (considered as primary) in the Tx queue and selects the best overall code. The NCVD complexity is linear in the number of packets in the Tx queue, which is also small for real-time applications. The dominant part is still due to the NCV part.
###
VI-C
Ongoing and Future Work
We are currently extending our approach to several directions. We are experimenting with additional topologies/scenarios, beyond the last-hop one-directional scenario discussed in this paper. We are also exploring extensions of our framework to capture (i) the dependency among video packets (ii) the benefit of overheard packets and (iii) code selection across multiple time slots.
## VII Conclusion
In this paper, we proposed a novel approach to opportunistic video coding for video streaming over wireless networks that takes into account the importance of video packets in network code selection. Simulation results show that the proposed schemes improve video quality up to $`3 - 5`$dB compared to baseline schemes. Furthermore, they significantly improve the application-level throughput and achieve the same or similar levels of MAC throughput.
## References
- \[1\]
N. Färber and B. Girod, “Wireless Video,” in A. Reibman, M.-T. Sun (eds.), Compressed Video over Networks, M. Dekker, 2000.
- \[2\]
Special issue on “Advances in Wireless Video,” in IEEE Wireless Communications Magazine, Aug. 2005.
- \[3\]
“The network coding webpage,” http://www.netcod.org.
- \[4\]
C. Fragouli. J. Widmer and J.Y. LeBoudec, “Network coding: an instant primer,” in Proc. of ACM SIGCOMM CCR, vol. 36(1), pp. 63-68, Jan. 2006.
- \[5\] S. Katti, H. Rahul, W. Hu, D. Katabi, M. Medard, and J. Crowcroft, “XORs in the air: Practical network coding,” *in Proc. of ACM SIGCOMM*, vol. 36(4), pp. 243-254, Pisa, Italy, Sept. 2006.
- \[6\] S. Katti, D. Katabi, W. Hu, H. Rahul, and M. Medard, “The Importance of Being Opportunistic: Practical Network Coding For Wireless Environments,” *in Proc. of Allerton Conference*, 2005.
- \[7\]
P. Chou and Z. Miao, “Rate-distortion optimized streaming of packetized media,” in IEEE Trans. on Multimedia, vol. 8(2), pp. 390- 404, Apr. 2006.
- \[8\]
J. Chakareski, J. Apostolopoulos, S. Wee, W. Tan, and B. Girod, “Rate-Distortion Hint Tracks for Adaptive Video Streaming,” in IEEE Trans. on Circuits and Systems for Video Technology, vol. 15(10), pp. 1257-1269, Oct. 2005.
- \[9\]
M. Kalman and B. Girod, “Rate-Distortion Optimized Video Streaming with Multiple Deadlines for Low Latency Applications,” in Proc. IEEE Packet Video Workshop, Irvine, CA, Dec. 2004.
- \[10\] R. Ahlswede, N. Cai, S. R. Li, and R. W. Yeung, “Network information flow,” *in IEEE Trans. on Information Theory*, vol. 46(4), pp. 1204-1216, July 2000.
- \[11\] S.R. Li, R.W. Yeung, and N. Cai,“Linear network coding,” *in IEEE Trans. on Information Theory*, vol. 49(2), pp. 371-381, Feb. 2003.
- \[12\]
B. Girod, J. Chakareski, M. Kalman, Y.J. Liang, E. Setton and R. Zhang, “Advances in Network-adaptive Video Streaming,” in Proc. of IWDC 2002, Capri, Italy, Sept. 2002.
- \[13\]
S. Shankar, Z. Hu, and M. Van der Schaar, “Cross Layer Optimized Transmission of Wavelet Video over IEEE 802.11a/e WLANs,” in Proc. IEEE Packet Video 2004, Irvine, CA, Dec. 2004.
- \[14\]
Special issue on “Multimedia over Broadband Wireless Networks,” in IEEE Network Magazine, vol. 19(1), pp. 25-25, Jan.-Feb 2005.
- \[15\]
M. Van der Schaar, S. Shankar,“Cross-Layer Wireless Multimedia Transmission: Challenges, principles, and new paradigms,” in IEEE Wireless Comm. Magazine, vol. 12(4), pp. 50-58, Aug. 2005.
- \[16\]
E. Setton, T. Yoo, X. Zhu, A. Goldsmith and B. Girod, “Cross-layer design of ad hoc networks for real-time video streaming,” in IEEE Wireless Comm. Magazine, vol. 12(4), pp. 59-65, Aug. 2005.
- \[17\]
G.M. Su, Z. Han, M. Wu, and R. Liu, “Multi-user Cross-Layer Resource Allocation for Video Transmission over Wireless Networks,” in IEEE Network Magazine, vol.20(2), pp.21-27, Mar-Apr 2006.
- \[18\] Z. Li, B. Li, “Network coding: The case for multiple unicast sessions,” *in Proc. of Allerton Conference*, Sept. 2004.
- \[19\] Y. Wu, P.A. Chou, and S. Y. Kung, “Information exchange in wireless network coding and physical layer broadcast,” *in Proc. of IEEE CISS*, Baltimore, MD, March 2005.
- \[20\] T. Ho and R. Koetter, “Online incremental network coding for multiple unicasts,” *in DIMACS WG on Network Coding*, Jan. 2005.
- \[21\] S. Deb, M. Effros, T. Ho, D.R. Karger, R. Koetter, D.S. Lun, M. Medard, and N. Ratnakar, “Network coding for wireless applications: A brief tutorial,” *in Proc. of IWWAN ’05*, London, UK.
- \[22\] A. Ramamoorthy, J. Shi, and R. Wesel, “On the capacity of network coding for wireless networks,” *in Allerton*, 2003.
- \[23\]
Y. Liang, J. Apostolopoulos and B. Girod, “Analysis of Packet Loss for Compressed Video: Does Burst-Length Matter?,” in Proc. of IEEE ICASSP, vol. 5, pp. 684-687, Hong Kong, Apr. 2003.
- \[24\] A. Goldsmith, “Wireless Communications,” New York, Cambridge University Press, 2005.
- \[25\] Q. Zhang, S.A. Kassam, “Finite-state Markov model for Rayleigh fading channels,” *in IEEE Trans. on Communications*, vol. 47(11), pp. 1688-1692, 1999.
- \[26\] ITU-T Rec. “H.264 : Advanced video coding for generic audiovisual services”, 2005.
- \[27\]
H.264/AVC Reference Software Version JM 8.6 http://iphome.hhi.de/suehring/tml/download/, 2006.
- \[28\] V. V. Vazirani, “Approximation Algorithms,” Springer-Verlag Berlin Heidelberg, 2003. |
0704.1833 | cs.NI | cyber | \centerfigcaptionstrue
# Analysis of the 802.11e Enhanced Distributed Channel Access Function
<sup>\*</sup>
<sup>\*</sup>\*<sup>†</sup> This work is supported by the Center for Pervasive Communications and Computing, and by National Science Foundation under Grant No. 0434928. Any opinions, findings, and conclusions or recommendations expressed in this material are those of authors and do not necessarily reflect the view of the National Science Foundation.
<sup>†</sup>
Inanc Inan, Feyza Keceli, and Ender Ayanoglu
Center for Pervasive Communications and Computing
Department of Electrical Engineering and Computer Science
The Henry Samueli School of Engineering
University of California, Irvine, 92697-2625
Email: {iinan, fkeceli, ayanoglu}@uci.edu
###### Abstract
The IEEE 802.11e standard revises the Medium Access Control (MAC) layer of the former IEEE 802.11 standard for Quality-of-Service (QoS) provision in the Wireless Local Area Networks (WLANs). The Enhanced Distributed Channel Access (EDCA) function of 802.11e defines multiple Access Categories (AC) with AC-specific Contention Window (CW) sizes, Arbitration Interframe Space (AIFS) values, and Transmit Opportunity (TXOP) limits to support MAC-level QoS and prioritization. We propose an analytical model for the EDCA function which incorporates an accurate CW, AIFS, and TXOP differentiation at any traffic load. The proposed model is also shown to capture the effect of MAC layer buffer size on the performance. Analytical and simulation results are compared to demonstrate the accuracy of the proposed approach for varying traffic loads, EDCA parameters, and MAC layer buffer space.
## I Introduction
The IEEE 802.11 standard \[[1](#bib.bib1)\] defines the Distributed Coordination Function (DCF) which provides best-effort service at the Medium Access Control (MAC) layer of the Wireless Local Area Networks (WLANs). The recently ratified IEEE 802.11e standard \[[2](#bib.bib2)\] specifies the Hybrid Coordination Function (HCF) which enables prioritized and parameterized Quality-of-Service (QoS) services at the MAC layer, on top of DCF. The HCF combines a distributed contention-based channel access mechanism, referred to as Enhanced Distributed Channel Access (EDCA), and a centralized polling-based channel access mechanism, referred to as HCF Controlled Channel Access (HCCA).
We confine our analysis to the EDCA scheme, which uses Carrier Sense Multiple Access with Collision Avoidance (CSMA/CA) and slotted Binary Exponential Backoff (BEB) mechanism as the basic access method. The EDCA defines multiple Access Categories (AC) with AC-specific Contention Window (CW) sizes, Arbitration Interframe Space (AIFS) values, and Transmit Opportunity (TXOP) limits to support MAC-level QoS and prioritization \[[2](#bib.bib2)\].
In order to assess the performance of these functions, simulations or mathematical analysis can be used. Although simulation models may capture system dynamics very closely, they lack explicit mathematical relations between the network parameters and performance measures. A number of networking functions would benefit from the insights provided by such mathematical relations. For example, analytical modeling is a more convenient way to assist embedded QoS-aware MAC scheduling and Call Admission Control (CAC) algorithms. Theoretical analysis can provide invaluable insights for QoS provisioning in the WLAN. On the other hand, analytical modeling can potentially be complex, where the effect of multiple layer network parameters makes the task of deriving a simple and accurate analytical model highly difficult. However, a set of appropriate assumptions may lead to simple yet accurate analytical models.
The majority of analytical work on the performance of 802.11e EDCA (and of 802.11 DCF) assumes that every station has always backlogged data ready to transmit in its buffer anytime (in saturation) as will be discussed in Section [III](#S3 "III Related Work ‣ Analysis of the 802.11e Enhanced Distributed Channel Access Function **footnote *† This work is supported by the Center for Pervasive Communications and Computing, and by National Science Foundation under Grant No. 0434928. Any opinions, findings, and conclusions or recommendations expressed in this material are those of authors and do not necessarily reflect the view of the National Science Foundation.†"). Analysis of the system in this state (saturation analysis) provides accurate and practical asymptotic figures. However, the saturation assumption is unlikely to be valid in practice given the fact that the demanded bandwidth for most of the Internet traffic is variable with significant idle periods. Our main contribution is an accurate EDCA analytical model which releases the saturation assumption. The model is shown to predict EDCA performance accurately for the whole traffic load range from a lightly loaded non-saturated channel to a heavily congested saturated medium for a range of traffic models.
Similarly, the majority of analytical work on the performance of 802.11e EDCA (and of 802.11 DCF) assumes constant collision probability for any transmitted packet at an arbitrary backoff slot independent of the number of retransmissions it has experienced. A complementary assumption is the constant transmission probability for any AC at an arbitrary backoff slot independent of the number of retransmissions it has experienced. As will be discussed in Section [III](#S3 "III Related Work ‣ Analysis of the 802.11e Enhanced Distributed Channel Access Function **footnote *† This work is supported by the Center for Pervasive Communications and Computing, and by National Science Foundation under Grant No. 0434928. Any opinions, findings, and conclusions or recommendations expressed in this material are those of authors and do not necessarily reflect the view of the National Science Foundation.†"), these approximations lead to accurate analysis in saturation. Our analysis shows that the slot homogeneity assumption leads to accurate performance prediction even when the saturation assumption is released.
Furthermore, the majority of analytical work on the performance of 802.11e EDCA (and of 802.11 DCF) in non-saturated conditions assumes either a very small or an infinitely large MAC layer buffer space. Our analysis removes such assumptions by incorporating the finite size MAC layer queue (interface queue between Link Layer (LL) and MAC layer) into the model. The finite size queue analysis shows the effect of MAC layer buffer space on EDCA performance which we will show to be significant.
A key contribution of this work is that the proposed analytical model incorporates all EDCA QoS parameters, CW, AIFS, and TXOP. The model also considers varying collision probabilities at different AIFS slots which is a direct result of varying number of contending stations. Comparing with simulations, we show that our model can provide accurate results for an arbitrary selection of AC-specific EDCA parameters at any load.
We present a Markov model the states of which represent the state of the backoff process and MAC buffer occupancy. To enable analysis in the Markov framework, we assume constant probability of packet arrival per state (for the sake of simplicity, Poisson arrivals). On the other hand, we have also shown that the results also hold for a range of traffic types.
## II EDCA Overview
The IEEE 802.11e EDCA is a QoS extension of IEEE 802.11 DCF. The major enhancement to support QoS is that EDCA differentiates packets using different priorities and maps them to specific ACs that are buffered in separate queues at a station. Each AC<sub>i</sub> within a station ($`0 \leq i \leq i_{m\hspace{0pt}a\hspace{0pt}x}`$, $`i_{m\hspace{0pt}a\hspace{0pt}x} = 3`$ in \[[2](#bib.bib2)\]) having its own EDCA parameters contends for the channel independently of the others. Following the convention of \[[2](#bib.bib2)\], the larger the index $`i`$ is, the higher the priority of the AC is. Levels of services are provided through different assignments of the AC specific EDCA parameters; AIFS, CW, and TXOP limits.
If there is a packet ready for transmission in the MAC queue of an AC, the EDCA function must sense the channel to be idle for a complete AIFS before it can start the transmission. The AIFS of an AC is determined by using the MAC Information Base (MIB) parameters as
<table>
<tbody>
<tr>
<td></td>
<td>AIFS=SIFS+AIFSN×Tslot,𝐴𝐼𝐹𝑆𝑆𝐼𝐹𝑆𝐴𝐼𝐹𝑆𝑁subscript𝑇𝑠𝑙𝑜𝑡\displaystyle AIFS=SIFS+AIFSN\times T_{slot},</td>
<td></td>
<td>(1)</td>
</tr>
</tbody>
</table>
where $`A\hspace{0pt}I\hspace{0pt}F\hspace{0pt}S\hspace{0pt}N`$ is the AC-specific AIFS number, $`S\hspace{0pt}I\hspace{0pt}F\hspace{0pt}S`$ is the length of the Short Interframe Space and $`T_{s\hspace{0pt}l\hspace{0pt}o\hspace{0pt}t}`$ is the duration of a time slot.
If the channel is idle when the first packet arrives at the AC queue, the packet can be directly transmitted as soon as the channel is sensed to be idle for AIFS. Otherwise, a backoff procedure is completed following the completion of AIFS before the transmission of this packet. A uniformly distributed random integer, namely a backoff value, is selected from the range $`\lbrack 0,W\rbrack`$. Should the channel be sensed busy at any time slot during AIFS or backoff, the backoff procedure is suspended at the current backoff value. The backoff resumes as soon as the channel is sensed to be idle for AIFS again. When the backoff counter reaches zero, the packet is transmitted in the following slot.
The value of $`W`$ depends on the number of retransmissions the current packet experienced. The initial value of $`W`$ is set to the AC-specific $`C\hspace{0pt}W_{m\hspace{0pt}i\hspace{0pt}n}`$. If the transmitter cannot receive an Acknowledgment (ACK) packet from the receiver in a timeout interval, the transmission is labeled as unsuccessful and the packet is scheduled for retransmission. At each unsuccessful transmission, the value of $`W`$ is doubled until the maximum AC-specific $`C\hspace{0pt}W_{m\hspace{0pt}a\hspace{0pt}x}`$ limit is reached. The value of $`W`$ is reset to the AC-specific $`C\hspace{0pt}W_{m\hspace{0pt}i\hspace{0pt}n}`$ if the transmission is successful, or the retry limit is reached thus the packet is dropped.
The higher priority ACs are assigned smaller AIFSN. Therefore, the higher priority ACs can either transmit or decrement their backoff counters while lower priority ACs are still waiting in AIFS. This results in higher priority ACs enjoying a lower average probability of collision and relatively faster progress through backoff slots. Moreover, in EDCA, the ACs with higher priority may select backoff values from a comparably smaller CW range. This approach prioritizes the access since a smaller CW value means a smaller backoff delay before the transmission.
Upon gaining the access to the medium, each AC may carry out multiple frame exchange sequences as long as the total access duration does not go over a TXOP limit. Within a TXOP, the transmissions are separated by SIFS. Multiple frame transmissions in a TXOP can reduce the overhead due to contention. A TXOP limit of zero corresponds to only one frame exchange per access.
An internal (virtual) collision within a station is handled by granting the access to the AC with the highest priority. The ACs with lower priority that suffer from a virtual collision run the collision procedure as if an outside collision has occured \[[2](#bib.bib2)\].
## III Related Work
In this section, we provide a brief summary of the theoretical DCF and EDCA function performance analysis in the literature.
The majority of previous work carries out performance analysis for asymptotical conditions assuming each station is in saturation. Three major saturation performance models have been proposed for DCF; i) assuming constant collision probability for each station, Bianchi \[[3](#bib.bib3)\] developed a simple Discrete-Time Markov Chain (DTMC) and the saturation throughput is obtained by applying regenerative analysis to a generic slot time, ii) Cali et al. \[[4](#bib.bib4)\],\[[5](#bib.bib5)\] employed renewal theory to analyze a p-persistent variant of DCF with persistence factor p derived from the CW, and iii) Tay et al. \[[6](#bib.bib6)\] instead used an average value mathematical method to model DCF backoff procedure and to calculate the average number of interruptions that the backoff timer experiences. Having the common assumption of slot homogeneity (for an arbitrary station, constant collision or transmission probability at an arbitrary slot), these models define all different renewal cycles all of which lead to accurate saturation performance analysis. Similarly, Medepalli et al. \[[7](#bib.bib7)\] provided explicit expressions for average DCF cycle time and system throughput. Pointing out another direction for future performance studies, Hui et al. \[[8](#bib.bib8)\] recently proposed the application of metamodeling techniques in order to find approximate closed-form mathematical models.
These major methods are modified by several researchers to include the extra features of the EDCA function in the saturation analysis. Xiao \[[9](#bib.bib9)\],\[[10](#bib.bib10)\] extended \[[3](#bib.bib3)\] to analyze only the CW differentiation. Kong et al. \[[11](#bib.bib11)\] took AIFS differentiation into account via a 3-dimensional DTMC. On the other hand, these EDCA extensions miss the treatment of varying collision probabilities at different AIFS slots due to varying number of contending stations. Robinson et al. \[[12](#bib.bib12)\],\[[13](#bib.bib13)\] proposed an average analysis on the collision probability for different contention zones during AIFS and employed calculated average collision probability on a 2-dimensional DTMC. Hui et al. \[[14](#bib.bib14)\],\[[15](#bib.bib15)\] unified several major approaches into one approximate average model taking into account varying collision probability in different backoff subperiods (corresponds to contention zones in \[[12](#bib.bib12)\]). Zhu et al. \[[16](#bib.bib16)\] proposed another analytical EDCA Markov model averaging the transition probabilities based on the number and the parameters of high priority flows. Inan et al. \[[17](#bib.bib17)\] proposed a simple DTMC which provides accurate treatment of AIFS and CW differentiation between the ACs for the constant transmission probability assumption. Another 3-dimensional DTMC is proposed by Tao et al. \[[18](#bib.bib18)\],\[[19](#bib.bib19)\] in which the third dimension models the state of backoff slots between successive transmission periods. In \[[18](#bib.bib18)\],\[[19](#bib.bib19)\], the fact that the number of idle slots between successive transmissions can be at most the minimum of AC-specific $`C\hspace{0pt}W_{m\hspace{0pt}a\hspace{0pt}x}`$ values is considered. Independent from \[[18](#bib.bib18)\],\[[19](#bib.bib19)\], Zhao et al. \[[20](#bib.bib20)\] had previously proposed a similar model for the heterogeneous case where each station has traffic of only one AC. Banchs et al. \[[21](#bib.bib21)\],\[[22](#bib.bib22)\] proposed another model which considers varying collision probability among different AIFS slots due to a variable number of stations. Chen et al. \[[23](#bib.bib23)\], Kuo et al. \[[24](#bib.bib24)\], and Lin et al. \[[25](#bib.bib25)\] extended \[[6](#bib.bib6)\] in order to include mean value analysis for AIFS and CW differentiation.
Although it has not yet received much attraction, the research that releases the saturation assumption basically follows two major methods; i) modeling the non-saturated behavior of DCF or EDCA function via Markov analysis, ii) employing queueing theory \[[26](#bib.bib26)\] and calculating certain quantities through average or Markov analysis. Our approach in this work falls into the first category.
Markov analysis for the non-saturated case still assumes slot homogeneity and extends \[[3](#bib.bib3)\] with necessary extra Markov states and transitions. Duffy et al. \[[27](#bib.bib27)\] and Alizadeh-Shabdiz et al. \[[28](#bib.bib28)\],\[[29](#bib.bib29)\] proposed similar extensions of \[[3](#bib.bib3)\] for non-saturated analysis of 802.11 DCF. Due to specific structure of the proposed DTMCs, these extensions assume a MAC layer buffer size of one packet. We show that this assumption may lead to significant performance prediction errors for EDCA in the case of larger buffers. Cantieni et al. \[[30](#bib.bib30)\] extended the model of \[[28](#bib.bib28)\] assuming infinitely large station buffers and the MAC queue being empty with constant probability regardless of the backoff stage the previous transmission took place. Li et al. \[[31](#bib.bib31)\] proposed an approximate model for non-saturation where only CW differentiation is considered. Engelstad et al. \[[32](#bib.bib32)\] used a DTMC model to perform delay analysis for both DCF and EDCA considering queue utilization probability as in \[[30](#bib.bib30)\]. Zaki et al. \[[33](#bib.bib33)\] proposed yet another Markov model with states that are of fixed real-time duration which cannot capture the pre-saturation DCF throughput peak.
A number of models employing queueing theory have also been developed for 802.11(e) performance analysis in non-saturated conditions. These models are assisted by independent analysis for the calculation of some quantities such as collision and transmission probabilities. Tickoo et al. \[[34](#bib.bib34)\],\[[35](#bib.bib35)\] modeled each 802.11 node as a discrete time G/G/1 queue to derive the service time distribution, but the models are based on an assumption that the saturated setting provides good approximation for certain quantities in non-saturated conditions. Chen et al. \[[36](#bib.bib36)\] employed both G/M/1 and G/G/1 queue models on top of \[[10](#bib.bib10)\] which only considers CW differentiation. Lee et al. \[[37](#bib.bib37)\] analyzed the use of M/G/1 queueing model while employing a simple non-saturated Markov model to calculate necessary quantities. Medepalli et al. \[[38](#bib.bib38)\] built upon the average cycle time derivation \[[7](#bib.bib7)\] to obtain individual queue delays using both M/G/1 and G/G/1 queueing models. Foh et al. \[[39](#bib.bib39)\] proposed a Markov framework to analyze the performance of DCF under statistical traffic. This framework models the number of contending nodes as an M/E<sub>j</sub>/1/k queue. Tantra et al. \[[40](#bib.bib40)\] extended \[[39](#bib.bib39)\] to include service differentiation in EDCA. However, such analysis is only valid for a restricted scenario where all nodes have a MAC queue size of one packet.
There are also a few studies that investigated the effect of EDCA TXOPs on 802.11e performance for a saturated scenario. Mangold et al. \[[41](#bib.bib41)\] and Suzuki et al. \[[42](#bib.bib42)\] carried out the performance analysis through simulation. The efficiency of burst transmissions with block acknowledgements is studied in \[[43](#bib.bib43)\]. Tinnirello et al. \[[44](#bib.bib44)\] also proposed different TXOP managing policies for temporal fairness provisioning. Peng et al. \[[45](#bib.bib45)\] proposed an analytical model to study the effect of burst transmissions and showed that improved service differentiation can be achieved using a novel scheme based on TXOP thresholds.
A thorough and careful literature survey shows that an EDCA analytical model which incorporates all EDCA QoS parameters, CW, AIFS, and TXOP, for any traffic load has not been designed yet.
## IV EDCA Discrete-Time Markov Chain Model
Assuming slot homogeneity, we propose a novel DTMC to model the behavior of the EDCA function of any AC at any load. The main contribution of this work is that the proposed model considers the effect of all EDCA QoS parameters (CW, AIFS, and TXOP) on the performance for the whole traffic load range from a lightly-loaded non-saturated channel to a heavily congested saturated medium. Although we assume constant probability of packet arrival per state (for the sake of simplicity, Poisson arrivals), we show that the model provides accurate performance analysis for a range of traffic types.
The state of the EDCA function of any AC at an arbitrary time $`t`$ depends on several MAC layer events that may have occured before $`t`$. We model the MAC layer state of an AC<sub>i</sub>, $`0 \leq i \leq 3`$, with a 3-dimensional Markov process, $`({s_{i}\hspace{0pt}{(t)}},{b_{i}\hspace{0pt}{(t)}},{q_{i}\hspace{0pt}{(t)}})`$. The stochastic process $`s_{i}\hspace{0pt}{(t)}`$ represents the value of the backoff stage at time $`t`$, i.e., the number of retransmissions that the packet to be transmitted currently has experienced until time $`t`$. The stochastic process $`b_{i}\hspace{0pt}{(t)}`$ represents the state of the backoff counter at time $`t`$. Up to this point, the definition of the first two dimensions follows \[[3](#bib.bib3)\] which is introduced for DCF. In order to enable the accurate non-saturated analysis considering EDCA TXOPs, we introduce another dimension which models the stochastic process $`q_{i}\hspace{0pt}{(t)}`$ denoting the number of packets buffered for transmission at the MAC layer. Moreover, as the details will be described in the sequel, in our model, $`b_{i}\hspace{0pt}{(t)}`$ does not only represent the value of the backoff counter, but also the number of transmissions carried out during the current EDCA TXOP (when the value of backoff counter is actually zero).
Using the assumption of independent and constant collision probability at an arbitrary backoff slot, the 3-dimensional process $`({s_{i}\hspace{0pt}{(t)}},{b_{i}\hspace{0pt}{(t)}},{q_{i}\hspace{0pt}{(t)}})`$ is represented as a Discrete-Time Markov Chain (DTMC) with states $`(j,k,l)`$ and index $`i`$. We define the limits on state variables as $`0 \leq j \leq {r_{i} - 1}`$, $`{- N_{i}} \leq k \leq W_{i,j}`$ and $`0 \leq l \leq {Q\hspace{0pt}S_{i}}`$. In these inequalities, we let $`r_{i}`$ be the retransmission limit of a packet of AC<sub>i</sub>; $`N_{i}`$ be the maximum number of successful packet exchange sequences of AC<sub>i</sub> that can fit into one TXOP<sub>i</sub>; $`W_{i,j} = {{2^{m\hspace{0pt}i\hspace{0pt}n\hspace{0pt}{(j,m_{i})}}\hspace{0pt}{({{C\hspace{0pt}W_{i,{m\hspace{0pt}i\hspace{0pt}n}}} + 1})}} - 1}`$ be the CW size of AC<sub>i</sub> at the backoff stage $`j`$ where $`{C\hspace{0pt}W_{i,{m\hspace{0pt}a\hspace{0pt}x}}} = {{2^{m_{i}}\hspace{0pt}{({{C\hspace{0pt}W_{i,{m\hspace{0pt}i\hspace{0pt}n}}} + 1})}} - 1}`$, $`0 \leq m_{i} < r_{i}`$; and $`Q\hspace{0pt}S_{i}`$ be the maximum number of packets that can buffered at the MAC layer, i.e., MAC queue size. Moreover, it is important to note that a couple of restrictions apply to the state indices.
- •
When there are not any buffered packets at the AC queue, the EDCA function of the corresponding AC cannot be in a retransmitting state. Therefore, if $`l = 0`$, then $`j = 0`$ should hold. Such backoff states represent the postbackoff process \[[1](#bib.bib1)\],\[[2](#bib.bib2)\], therefore called as postbackoff slots in the sequel. The postbackoff procedure ensures that the transmitting station waits at least another backoff between successive TXOPs. Note that, when $`l > 0`$ and $`k \geq 0`$, these states are named backoff slots.
- •
The states with indices $`{- N_{i}} \leq k \leq {- 1}`$ represent the negation of the number of packets that are successfully transmitted at the current TXOP rather than the value of the backoff counter (which is zero during a TXOP). For simplicity, in the design of the Markov chain, we introduced such states in the second dimension. Therefore, if $`{- N_{i}} \leq k \leq {- 1}`$, we set $`j = 0`$. As it will be clear in the sequel, the addition of these states enables EDCA TXOP analysis.
Let $`p_{c_{i}}`$ denote the average conditional probability that a packet from AC<sub>i</sub> experiences either an external or an internal collision after the EDCA function decides on the transmission. Let $`p_{n\hspace{0pt}t}\hspace{0pt}{(l',\left. T \middle| l \right.)}`$ be the probability that there are $`l'`$ packets in the MAC buffer at time $`t + T`$ given that there were $`l`$ packets at $`t`$ and no transmissions have been made during interval $`T`$. Similarly, let $`p_{s\hspace{0pt}t}\hspace{0pt}{(l',\left. T \middle| l \right.)}`$ be the probability that there are $`l'`$ packets in the MAC buffer at time $`t + T`$ given that there were $`l`$ packets at time $`t`$ and a transmission has been made during interval $`T`$. Note that since we assume Poisson arrivals, the exponential interarrival distributions are independent, and $`p_{n\hspace{0pt}t}`$ and $`p_{s\hspace{0pt}t}`$ only depend on the interval length $`T`$ and are independent of time $`t`$. Then, the nonzero state transmission probabilities of the proposed Markov model for AC<sub>i</sub>, denoted as $`P_{i}\hspace{0pt}{(j',k',\left. l' \middle| {j,k,l} \right.)}`$ adopting the same notation in \[[3](#bib.bib3)\], are calculated as follows.
1. 1\.
The backoff counter is decremented by one at the slot boundary. Note that we define the postbackoff or the backoff slot as Bianchi defines the slot time \[[3](#bib.bib3)\]. Then, for $`0 \leq j \leq {r_{i} - 1}`$, $`1 \leq k \leq W_{i,j}`$, and $`0 \leq l \leq l' \leq {Q\hspace{0pt}S_{i}}`$,
<table>
<tbody>
<tr>
<td></td>
<td>Pi(j,k−1,l′|j,k,l)=pnt(l′,Ti,bs|l).subscript𝑃𝑖𝑗𝑘1conditionalsuperscript𝑙′𝑗𝑘𝑙subscript𝑝𝑛𝑡superscript𝑙′conditionalsubscript𝑇𝑖𝑏𝑠𝑙\displaystyle P_{i}(j,k-1,l^{\prime}|j,k,l)=p_{nt}(l^{\prime},T_{i,bs}|l).</td>
<td></td>
<td>(2)</td>
</tr>
</tbody>
</table>
It is important to note that the proposed DTMC’s evolution is not real-time and the state duration varies depending on the state. The average duration of a backoff slot $`T_{i,{b\hspace{0pt}s}}`$ is calculated by ([29](#S4.E29 "In IV-A3 The state duration 𝑇_{𝑖,𝑏𝑠} and 𝑇_{𝑖,𝑏} ‣ IV-A Steady-State Solution ‣ IV EDCA Discrete-Time Markov Chain Model ‣ Analysis of the 802.11e Enhanced Distributed Channel Access Function **footnote *† This work is supported by the Center for Pervasive Communications and Computing, and by National Science Foundation under Grant No. 0434928. Any opinions, findings, and conclusions or recommendations expressed in this material are those of authors and do not necessarily reflect the view of the National Science Foundation.†")) which will be derived. Note that, in ([2](#S4.E2 "In item 1 ‣ IV EDCA Discrete-Time Markov Chain Model ‣ Analysis of the 802.11e Enhanced Distributed Channel Access Function **footnote *† This work is supported by the Center for Pervasive Communications and Computing, and by National Science Foundation under Grant No. 0434928. Any opinions, findings, and conclusions or recommendations expressed in this material are those of authors and do not necessarily reflect the view of the National Science Foundation.†")), we consider the probability of packet arrivals during $`T_{i,{b\hspace{0pt}s}}`$ (buffer size $`l'`$ after the state transition depends on this probability).
2. 2\.
We assume the transmitted packet experiences a collision with constant probability $`p_{c_{i}}`$ (slot homogeneity). In the following, note that the cases when the retry limit is reached and when the MAC buffer is full are treated separately, since the transition probabilities should follow different rules. Let $`T_{i,s}`$ and $`T_{i,c}`$ be the time spent in a successful transmission and a collision by AC<sub>i</sub> respectively which will be derived. Then, for $`0 \leq j \leq {r_{i} - 1}`$, $`0 \leq l \leq {{Q\hspace{0pt}S_{i}} - 1}`$, and $`{\max{(0,{l - 1})}} \leq l' \leq {Q\hspace{0pt}S_{i}}`$,
<table>
<tbody>
<tr>
<td></td>
<td></td>
<td>Pi(0,−1,l′|j,0,l)=(1−pci)⋅pst(l′,Ti,s|l)subscript𝑃𝑖01conditionalsuperscript𝑙′𝑗0𝑙⋅1subscript𝑝subscript𝑐𝑖subscript𝑝𝑠𝑡superscript𝑙′conditionalsubscript𝑇𝑖𝑠𝑙\displaystyle P_{i}(0,-1,l^{\prime}|j,0,l)=(1-p_{c_{i}})\cdot p_{st}(l^{\prime},T_{i,s}|l)</td>
<td></td>
<td>(3)</td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td></td>
<td>Pi(0,−1,QSi−1|j,0,QSi)=1−pci.subscript𝑃𝑖01𝑄subscript𝑆𝑖conditional1𝑗0𝑄subscript𝑆𝑖1subscript𝑝subscript𝑐𝑖\displaystyle P_{i}(0,-1,QS_{i}-1|j,0,QS_{i})=1-p_{c_{i}}.</td>
<td></td>
<td>(4)</td>
</tr>
</tbody>
</table>
For $`0 \leq j \leq {r_{i} - 2}`$, $`0 \leq k \leq W_{i,{j + 1}}`$, and $`0 \leq l \leq l' \leq {Q\hspace{0pt}S_{i}}`$,
<table>
<tbody>
<tr>
<td></td>
<td></td>
<td>Pi(j+1,k,l′|j,0,l)=pci⋅pnt(l′,Ti,c|l)Wi,j+1+1.subscript𝑃𝑖𝑗1𝑘conditionalsuperscript𝑙′𝑗0𝑙⋅subscript𝑝subscript𝑐𝑖subscript𝑝𝑛𝑡superscript𝑙′conditionalsubscript𝑇𝑖𝑐𝑙subscript𝑊𝑖𝑗11\displaystyle P_{i}(j+1,k,l^{\prime}|j,0,l)=\frac{p_{c_{i}}\cdot p_{nt}(l^{\prime},T_{i,c}|l)}{W_{i,j+1}+1}.</td>
<td></td>
<td>(5)</td>
</tr>
</tbody>
</table>
For $`0 \leq k \leq W_{i,0}`$, $`0 \leq l \leq {{Q\hspace{0pt}S_{i}} - 1}`$, and $`{\max{(0,{l - 1})}} \leq l' \leq {Q\hspace{0pt}S_{i}}`$,
<table>
<tbody>
<tr>
<td></td>
<td></td>
<td>Pi(0,k,l′|ri−1,0,l)=pciWi,0+1⋅pst(l′,Ti,s|l)subscript𝑃𝑖0𝑘conditionalsuperscript𝑙′subscript𝑟𝑖10𝑙⋅subscript𝑝subscript𝑐𝑖subscript𝑊𝑖01subscript𝑝𝑠𝑡superscript𝑙′conditionalsubscript𝑇𝑖𝑠𝑙\displaystyle P_{i}(0,k,l^{\prime}|r_{i}-1,0,l)=\frac{p_{c_{i}}}{W_{i,0}+1}\cdot p_{st}(l^{\prime},T_{i,s}|l)</td>
<td></td>
<td>(6)</td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td></td>
<td>Pi(0,k,QSi−1|ri−1,0,QSi)=pciWi,0+1subscript𝑃𝑖0𝑘𝑄subscript𝑆𝑖conditional1subscript𝑟𝑖10𝑄subscript𝑆𝑖subscript𝑝subscript𝑐𝑖subscript𝑊𝑖01\displaystyle P_{i}(0,k,QS_{i}-1|r_{i}-1,0,QS_{i})=\frac{p_{c_{i}}}{W_{i,0}+1}</td>
<td></td>
<td>(7)</td>
</tr>
</tbody>
</table>
Note that we use $`p_{n\hspace{0pt}t}`$ in ([5](#S4.E5 "In item 2 ‣ IV EDCA Discrete-Time Markov Chain Model ‣ Analysis of the 802.11e Enhanced Distributed Channel Access Function **footnote *† This work is supported by the Center for Pervasive Communications and Computing, and by National Science Foundation under Grant No. 0434928. Any opinions, findings, and conclusions or recommendations expressed in this material are those of authors and do not necessarily reflect the view of the National Science Foundation.†")) although a transmission has been made. On the other hand, the packet has collided and is still at the MAC queue for retransmission as if no transmission has occured. This is not the case in ([3](#S4.E3 "In item 2 ‣ IV EDCA Discrete-Time Markov Chain Model ‣ Analysis of the 802.11e Enhanced Distributed Channel Access Function **footnote *† This work is supported by the Center for Pervasive Communications and Computing, and by National Science Foundation under Grant No. 0434928. Any opinions, findings, and conclusions or recommendations expressed in this material are those of authors and do not necessarily reflect the view of the National Science Foundation.†")) and ([6](#S4.E6 "In item 2 ‣ IV EDCA Discrete-Time Markov Chain Model ‣ Analysis of the 802.11e Enhanced Distributed Channel Access Function **footnote *† This work is supported by the Center for Pervasive Communications and Computing, and by National Science Foundation under Grant No. 0434928. Any opinions, findings, and conclusions or recommendations expressed in this material are those of authors and do not necessarily reflect the view of the National Science Foundation.†")), since in these transitions a successful transmission or a drop occurs. When the MAC buffer is full, any arriving packet is discarded as ([4](#S4.E4 "In item 2 ‣ IV EDCA Discrete-Time Markov Chain Model ‣ Analysis of the 802.11e Enhanced Distributed Channel Access Function **footnote *† This work is supported by the Center for Pervasive Communications and Computing, and by National Science Foundation under Grant No. 0434928. Any opinions, findings, and conclusions or recommendations expressed in this material are those of authors and do not necessarily reflect the view of the National Science Foundation.†")) and ([7](#S4.E7 "In item 2 ‣ IV EDCA Discrete-Time Markov Chain Model ‣ Analysis of the 802.11e Enhanced Distributed Channel Access Function **footnote *† This work is supported by the Center for Pervasive Communications and Computing, and by National Science Foundation under Grant No. 0434928. Any opinions, findings, and conclusions or recommendations expressed in this material are those of authors and do not necessarily reflect the view of the National Science Foundation.†")) imply.
3. 3\.
Once the TXOP is started, the EDCA function may continue with as many packet SIFS-separated exchange sequences as it can fit into the TXOP duration. Let $`T_{i,{e\hspace{0pt}x\hspace{0pt}c}}`$ be the average duration of a successful packet exchange sequence for AC<sub>i</sub> which will be derived in ([24](#S4.E24 "In IV-A3 The state duration 𝑇_{𝑖,𝑏𝑠} and 𝑇_{𝑖,𝑏} ‣ IV-A Steady-State Solution ‣ IV EDCA Discrete-Time Markov Chain Model ‣ Analysis of the 802.11e Enhanced Distributed Channel Access Function **footnote *† This work is supported by the Center for Pervasive Communications and Computing, and by National Science Foundation under Grant No. 0434928. Any opinions, findings, and conclusions or recommendations expressed in this material are those of authors and do not necessarily reflect the view of the National Science Foundation.†")). Then, for $`{{- N_{i}} + 1} \leq k \leq {- 1}`$, $`1 \leq l \leq {Q\hspace{0pt}S_{i}}`$, and $`{\max{(0,{l - 1})}} \leq l' \leq {Q\hspace{0pt}S_{i}}`$,
<table>
<tbody>
<tr>
<td></td>
<td>Pi(0,k−1,l′|0,k,l)=pst(l′,Ti,exc|l).subscript𝑃𝑖0𝑘1conditionalsuperscript𝑙′0𝑘𝑙subscript𝑝𝑠𝑡superscript𝑙′conditionalsubscript𝑇𝑖𝑒𝑥𝑐𝑙\displaystyle P_{i}(0,k-1,l^{\prime}|0,k,l)=p_{st}(l^{\prime},T_{i,exc}|l).</td>
<td></td>
<td>(8)</td>
</tr>
</tbody>
</table>
When the next transmission cannot fit into the remaining TXOP, the current TXOP is immediately concluded and the unused portion of the TXOP is returned. By design, our model includes maximum number of packets that can fit into one TXOP. Then, for $`0 \leq k \leq W_{i,0}`$ and $`1 \leq l \leq {Q\hspace{0pt}S_{i}}`$,
<table>
<tbody>
<tr>
<td></td>
<td>Pi(0,k,l|0,−Ni,l)=1Wi,0+1.subscript𝑃𝑖0𝑘conditional𝑙0subscript𝑁𝑖𝑙1subscript𝑊𝑖01\displaystyle P_{i}(0,k,l|0,-N_{i},l)=\frac{1}{W_{i,0}+1}.</td>
<td></td>
<td>(9)</td>
</tr>
</tbody>
</table>
The TXOP ends when the MAC queue is empty. Then, for $`0 \leq k' \leq W_{i,0}`$ and $`{- N_{i}} \leq k \leq {- 1}`$,
<table>
<tbody>
<tr>
<td></td>
<td>Pi(0,k′,0|0,k,0)=1Wi,0+1.subscript𝑃𝑖0superscript𝑘′conditional00𝑘01subscript𝑊𝑖01\displaystyle P_{i}(0,k^{\prime},0|0,k,0)=\frac{1}{W_{i,0}+1}.</td>
<td></td>
<td>(10)</td>
</tr>
</tbody>
</table>
Note that no time passes in ([9](#S4.E9 "In item 3 ‣ IV EDCA Discrete-Time Markov Chain Model ‣ Analysis of the 802.11e Enhanced Distributed Channel Access Function **footnote *† This work is supported by the Center for Pervasive Communications and Computing, and by National Science Foundation under Grant No. 0434928. Any opinions, findings, and conclusions or recommendations expressed in this material are those of authors and do not necessarily reflect the view of the National Science Foundation.†")) and ([10](#S4.E10 "In item 3 ‣ IV EDCA Discrete-Time Markov Chain Model ‣ Analysis of the 802.11e Enhanced Distributed Channel Access Function **footnote *† This work is supported by the Center for Pervasive Communications and Computing, and by National Science Foundation under Grant No. 0434928. Any opinions, findings, and conclusions or recommendations expressed in this material are those of authors and do not necessarily reflect the view of the National Science Foundation.†")), so the definition of these states and transitions is actually not necessary for accuracy. On the other hand, they simplify the DTMC structure and symmetry.
4. 4\.
If the queue is still empty when the postbackoff counter reaches zero, the EDCA function enters the idle state until another packet arrival. Note (0,0,0) also represents the idle state. We make two assumptions; i) At most one packet may arrive during $`T_{s\hspace{0pt}l\hspace{0pt}o\hspace{0pt}t}`$ with constant probability $`\rho_{i}`$ (considering the fact that $`T_{s\hspace{0pt}l\hspace{0pt}o\hspace{0pt}t}`$ is in the order of microseconds, the probability that multiple packets can arrive in this interval is very small), ii) if the channel is idle at the slot the packet arrives at an empty queue, the transmission will be successful at AIFS completion without any backoff. The latter assumption is due to the following reason. While the probability of the channel becoming busy during AIFS or a collision occuring for the transmission at AIFS is very small at a lightly loaded scenario, the probability of a packet arrival to an empty queue is very small at a highly loaded scenario. As observed via simulations, these assumptions do not lead to any noticeable changes in the results while simplifying the Markov chain structure and symmetry. Then, for $`0 \leq k \leq W_{i,0}`$ and $`1 \leq l \leq {Q\hspace{0pt}S_{i}}`$,
<table>
<tbody>
<tr>
<td></td>
<td></td>
<td>Pi(0,0,0|0,0,0)=(1−pci)⋅(1−ρi)+pci⋅pnt(0,Ti,b|0),subscript𝑃𝑖00conditional0000⋅1subscript𝑝subscript𝑐𝑖1subscript𝜌𝑖⋅subscript𝑝subscript𝑐𝑖subscript𝑝𝑛𝑡0conditionalsubscript𝑇𝑖𝑏0\displaystyle P_{i}(0,0,0|0,0,0)=(1-p_{c_{i}})\cdot(1-\rho_{i})+p_{c_{i}}\cdot p_{nt}(0,T_{i,b}|0),</td>
<td></td>
<td>(11)</td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td></td>
<td>Pi(0,k,l|0,0,0)=pciWi,0+1⋅pnt(l,Ti,b|0),subscript𝑃𝑖0𝑘conditional𝑙000⋅subscript𝑝subscript𝑐𝑖subscript𝑊𝑖01subscript𝑝𝑛𝑡𝑙conditionalsubscript𝑇𝑖𝑏0\displaystyle P_{i}(0,k,l|0,0,0)=\frac{p_{c_{i}}}{W_{i,0}+1}\cdot p_{nt}(l,T_{i,b}|0),</td>
<td></td>
<td>(12)</td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td></td>
<td>Pi(0,−1,l|0,0,0)=(1−pci)⋅ρi⋅pnt(l,Ti,s|0).subscript𝑃𝑖01conditional𝑙000⋅1subscript𝑝subscript𝑐𝑖subscript𝜌𝑖subscript𝑝𝑛𝑡𝑙conditionalsubscript𝑇𝑖𝑠0\displaystyle P_{i}(0,-1,l|0,0,0)=(1-p_{c_{i}})\cdot\rho_{i}\cdot p_{nt}(l,T_{i,s}|0).</td>
<td></td>
<td>(13)</td>
</tr>
</tbody>
</table>
Let $`T_{i,b}`$ in ([11](#S4.E11 "In item 4 ‣ IV EDCA Discrete-Time Markov Chain Model ‣ Analysis of the 802.11e Enhanced Distributed Channel Access Function **footnote *† This work is supported by the Center for Pervasive Communications and Computing, and by National Science Foundation under Grant No. 0434928. Any opinions, findings, and conclusions or recommendations expressed in this material are those of authors and do not necessarily reflect the view of the National Science Foundation.†")) and ([12](#S4.E12 "In item 4 ‣ IV EDCA Discrete-Time Markov Chain Model ‣ Analysis of the 802.11e Enhanced Distributed Channel Access Function **footnote *† This work is supported by the Center for Pervasive Communications and Computing, and by National Science Foundation under Grant No. 0434928. Any opinions, findings, and conclusions or recommendations expressed in this material are those of authors and do not necessarily reflect the view of the National Science Foundation.†")) be the length of a backoff slot given it is not idle. Note that actually a successful transmission occurs in the state transition in ([13](#S4.E13 "In item 4 ‣ IV EDCA Discrete-Time Markov Chain Model ‣ Analysis of the 802.11e Enhanced Distributed Channel Access Function **footnote *† This work is supported by the Center for Pervasive Communications and Computing, and by National Science Foundation under Grant No. 0434928. Any opinions, findings, and conclusions or recommendations expressed in this material are those of authors and do not necessarily reflect the view of the National Science Foundation.†")). On the other hand, the transmitted packet is not reflected in the initial queue size state which is 0. Therefore, $`p_{n\hspace{0pt}t}`$ is used instead of $`p_{s\hspace{0pt}t}`$.
Parts of the proposed DTMC model are illustrated in Fig. [1](#S6.F1 "Figure 1 ‣ Analysis of the 802.11e Enhanced Distributed Channel Access Function **footnote *† This work is supported by the Center for Pervasive Communications and Computing, and by National Science Foundation under Grant No. 0434928. Any opinions, findings, and conclusions or recommendations expressed in this material are those of authors and do not necessarily reflect the view of the National Science Foundation.†") for an arbitrary AC<sub>i</sub> with $`N_{i} = 2`$. Fig. [1](#S6.F1 "Figure 1 ‣ Analysis of the 802.11e Enhanced Distributed Channel Access Function **footnote *† This work is supported by the Center for Pervasive Communications and Computing, and by National Science Foundation under Grant No. 0434928. Any opinions, findings, and conclusions or recommendations expressed in this material are those of authors and do not necessarily reflect the view of the National Science Foundation.†")(a) shows the state transitions for $`l = 0`$. Note that in Fig. [1](#S6.F1 "Figure 1 ‣ Analysis of the 802.11e Enhanced Distributed Channel Access Function **footnote *† This work is supported by the Center for Pervasive Communications and Computing, and by National Science Foundation under Grant No. 0434928. Any opinions, findings, and conclusions or recommendations expressed in this material are those of authors and do not necessarily reflect the view of the National Science Foundation.†")(a) the states with $`{- N_{i}} \leq k \leq {- 2}`$ can only be reached from the states with $`l = 1`$. Fig. [1](#S6.F1 "Figure 1 ‣ Analysis of the 802.11e Enhanced Distributed Channel Access Function **footnote *† This work is supported by the Center for Pervasive Communications and Computing, and by National Science Foundation under Grant No. 0434928. Any opinions, findings, and conclusions or recommendations expressed in this material are those of authors and do not necessarily reflect the view of the National Science Foundation.†")(b) presents the state transitions for $`0 < l < {Q\hspace{0pt}S_{i}}`$ and $`0 \leq j < r_{i}`$. Note that only the transition probabilities and the states marked with rectangles differ when $`j = {r_{i} - 1}`$ (as in ([6](#S4.E6 "In item 2 ‣ IV EDCA Discrete-Time Markov Chain Model ‣ Analysis of the 802.11e Enhanced Distributed Channel Access Function **footnote *† This work is supported by the Center for Pervasive Communications and Computing, and by National Science Foundation under Grant No. 0434928. Any opinions, findings, and conclusions or recommendations expressed in this material are those of authors and do not necessarily reflect the view of the National Science Foundation.†"))). Therefore, we do not include an extra figure for this case. Fig. [1](#S6.F1 "Figure 1 ‣ Analysis of the 802.11e Enhanced Distributed Channel Access Function **footnote *† This work is supported by the Center for Pervasive Communications and Computing, and by National Science Foundation under Grant No. 0434928. Any opinions, findings, and conclusions or recommendations expressed in this material are those of authors and do not necessarily reflect the view of the National Science Foundation.†")(c) shows the state transitions when $`l = {Q\hspace{0pt}S_{i}}`$. Note also that the states marked with rectangles differ when $`j = {r_{i} - 1}`$ (as in ([7](#S4.E7 "In item 2 ‣ IV EDCA Discrete-Time Markov Chain Model ‣ Analysis of the 802.11e Enhanced Distributed Channel Access Function **footnote *† This work is supported by the Center for Pervasive Communications and Computing, and by National Science Foundation under Grant No. 0434928. Any opinions, findings, and conclusions or recommendations expressed in this material are those of authors and do not necessarily reflect the view of the National Science Foundation.†"))). The combination of these small chains for all $`j`$, $`k`$, $`l`$ constitutes our DTMC model.
###
IV-A
Steady-State Solution
Let $`b_{i,j,k,l}`$ be the steady-state probability of the state $`(j,k,l)`$ of the proposed DTMC with index $`i`$ which can be solved using ([2](#S4.E2 "In item 1 ‣ IV EDCA Discrete-Time Markov Chain Model ‣ Analysis of the 802.11e Enhanced Distributed Channel Access Function **footnote *† This work is supported by the Center for Pervasive Communications and Computing, and by National Science Foundation under Grant No. 0434928. Any opinions, findings, and conclusions or recommendations expressed in this material are those of authors and do not necessarily reflect the view of the National Science Foundation.†"))-([13](#S4.E13 "In item 4 ‣ IV EDCA Discrete-Time Markov Chain Model ‣ Analysis of the 802.11e Enhanced Distributed Channel Access Function **footnote *† This work is supported by the Center for Pervasive Communications and Computing, and by National Science Foundation under Grant No. 0434928. Any opinions, findings, and conclusions or recommendations expressed in this material are those of authors and do not necessarily reflect the view of the National Science Foundation.†")) subject to $`{\sum_{j}{\sum_{k}{\sum_{l}b_{i,j,k,l}}}} = 1`$ (the proposed DTMC is ergodic and irreducible). Let $`\tau_{i}`$ be the probability that an AC<sub>i</sub> transmits at an arbitrary backoff or postbackoff slot
<table>
<tbody>
<tr>
<td></td>
<td>τi=(∑j=0ri−1∑l=1QSibi,j,0,l)+bi,0,0,0⋅ρi⋅(1−pci)∑j=0ri−1∑k=0Wi,j∑l=0QSibi,j,k,l.subscript𝜏𝑖superscriptsubscript𝑗0subscript𝑟𝑖1superscriptsubscript𝑙1𝑄subscript𝑆𝑖subscript𝑏𝑖𝑗0𝑙⋅subscript𝑏𝑖000subscript𝜌𝑖1subscript𝑝subscript𝑐𝑖superscriptsubscript𝑗0subscript𝑟𝑖1superscriptsubscript𝑘0subscript𝑊𝑖𝑗superscriptsubscript𝑙0𝑄subscript𝑆𝑖subscript𝑏𝑖𝑗𝑘𝑙\tau_{i}=\frac{\left(\sum_{j=0}^{r_{i}-1}\sum_{l=1}^{QS_{i}}b_{i,j,0,l}\right)+b_{i,0,0,0}\cdot\rho_{i}\cdot(1-p_{c_{i}})}{\sum_{j=0}^{r_{i}-1}\sum_{k=0}^{W_{i,j}}\sum_{l=0}^{QS_{i}}b_{i,j,k,l}}.</td>
<td></td>
<td>(14)</td>
</tr>
</tbody>
</table>
Note that $`{- N_{i}} \leq k \leq {- 1}`$ is not included in the normalization in ([14](#S4.E14 "In IV-A Steady-State Solution ‣ IV EDCA Discrete-Time Markov Chain Model ‣ Analysis of the 802.11e Enhanced Distributed Channel Access Function **footnote *† This work is supported by the Center for Pervasive Communications and Computing, and by National Science Foundation under Grant No. 0434928. Any opinions, findings, and conclusions or recommendations expressed in this material are those of authors and do not necessarily reflect the view of the National Science Foundation.†")), since these states represent a continuation in the EDCA TXOP rather than a contention for the access. The value of $`\tau_{i}`$ depends on the values of the average conditional collision probability $`p_{c_{i}}`$, the various state durations $`T_{i,{b\hspace{0pt}s}}`$, $`T_{i,b}`$, $`T_{i,s}`$ and $`T_{i,c}`$, and the conditional queue state transition probabilities $`p_{n\hspace{0pt}t}`$ and $`p_{s\hspace{0pt}t}`$.
####
IV-A1
Average conditional collision probability $`p_{c_{i}}`$
The difference in AIFS of each AC in EDCA creates the so-called contention zones as shown in Fig. [2](#S6.F2 "Figure 2 ‣ Analysis of the 802.11e Enhanced Distributed Channel Access Function **footnote *† This work is supported by the Center for Pervasive Communications and Computing, and by National Science Foundation under Grant No. 0434928. Any opinions, findings, and conclusions or recommendations expressed in this material are those of authors and do not necessarily reflect the view of the National Science Foundation.†") \[[12](#bib.bib12)\]. In each contention zone, the number of contending stations may vary. The collision probability cannot simply be assumed to be constant among all ACs.
We can define $`p_{c_{i,x}}`$ as the conditional probability that AC<sub>i</sub> experiences either an external or an internal collision given that it has observed the medium idle for $`A\hspace{0pt}I\hspace{0pt}F\hspace{0pt}S_{x}`$ and transmits in the current slot (note $`{A\hspace{0pt}I\hspace{0pt}F\hspace{0pt}S_{x}} \geq {A\hspace{0pt}I\hspace{0pt}F\hspace{0pt}S_{i}}`$ should hold). For the following, in order to be consistent with the notation of \[[2](#bib.bib2)\], we assume $`{A\hspace{0pt}I\hspace{0pt}F\hspace{0pt}S_{0}} \geq {A\hspace{0pt}I\hspace{0pt}F\hspace{0pt}S_{1}} \geq {A\hspace{0pt}I\hspace{0pt}F\hspace{0pt}S_{2}} \geq {A\hspace{0pt}I\hspace{0pt}F\hspace{0pt}S_{3}}`$. Let $`d_{i} = {{A\hspace{0pt}I\hspace{0pt}F\hspace{0pt}S_{i}} - {A\hspace{0pt}I\hspace{0pt}F\hspace{0pt}S_{3}}}`$. Also, let the total number AC<sub>i</sub> flows be $`f_{i}`$. Then, for the heterogeneous scenario in which each station has only one AC
<table>
<tbody>
<tr>
<td></td>
<td>pci,x=1−∏i′:di′≤dx(1−τi′)fi′(1−τi).subscript𝑝subscript𝑐𝑖𝑥1subscriptproduct:superscript𝑖′subscript𝑑superscript𝑖′subscript𝑑𝑥superscript1subscript𝜏superscript𝑖′subscript𝑓superscript𝑖′1subscript𝜏𝑖p_{c_{i,x}}=1-\frac{\prod\limits_{i^{\prime}:d_{i^{\prime}}\leq d_{x}}(1-\tau_{i^{\prime}})^{f_{i^{\prime}}}}{(1-\tau_{i})}.</td>
<td></td>
<td>(15)</td>
</tr>
</tbody>
</table>
When each station has multiple ACs that are active, internal collisions may occur. Then, for the scenario in which each station has all 4 ACs active
<table>
<tbody>
<tr>
<td></td>
<td>pci,x=1−∏i′:di′≤dx(1−τi′)fi′−1∏i′′>i(1−τi′′).subscript𝑝subscript𝑐𝑖𝑥1subscriptproduct:superscript𝑖′subscript𝑑superscript𝑖′subscript𝑑𝑥superscript1subscript𝜏superscript𝑖′subscript𝑓superscript𝑖′1subscriptproductsuperscript𝑖′′𝑖1subscript𝜏superscript𝑖′′p_{c_{i,x}}=1-\prod_{i^{\prime}:d_{i^{\prime}}\leq d_{x}}(1-\tau_{i^{\prime}})^{f_{i^{\prime}}-1}\prod_{i^{\prime\prime}>i}(1-\tau_{i^{\prime\prime}}).</td>
<td></td>
<td>(16)</td>
</tr>
</tbody>
</table>
Similar extensions when the number of active ACs are 2 or 3 are straightforward.
We use the Markov chain shown in Fig. [3](#S6.F3 "Figure 3 ‣ Analysis of the 802.11e Enhanced Distributed Channel Access Function **footnote *† This work is supported by the Center for Pervasive Communications and Computing, and by National Science Foundation under Grant No. 0434928. Any opinions, findings, and conclusions or recommendations expressed in this material are those of authors and do not necessarily reflect the view of the National Science Foundation.†") to find the long term occupancy of contention zones. Each state represents the $`n^{t\hspace{0pt}h}`$ backoff slot after completion of the AIFS<sub>3</sub> idle interval following a transmission period. The Markov chain model uses the fact that a backoff slot is reached if and only if no transmission occurs in the previous slot. Moreover, the number of states is limited by the maximum idle time between two successive transmissions which is $`W_{m\hspace{0pt}i\hspace{0pt}n} = {\min{({C\hspace{0pt}W_{i,{m\hspace{0pt}a\hspace{0pt}x}}})}}`$ for a saturated scenario. Although this is not the case for a non-saturated scenario, we do not change this limit. As the comparison with simulation results show, this approximation does not result in significant prediction errors. The probability that at least one transmission occurs in a backoff slot in contention zone $`x`$ is
<table>
<tbody>
<tr>
<td></td>
<td>pxtr=1−∏i′:di′≤dx(1−τi′)fi′.subscriptsuperscript𝑝𝑡𝑟𝑥1subscriptproduct:superscript𝑖′subscript𝑑superscript𝑖′subscript𝑑𝑥superscript1subscript𝜏superscript𝑖′subscript𝑓superscript𝑖′p^{tr}_{x}=1-\prod_{i^{\prime}:d_{i^{\prime}}\leq d_{x}}(1-\tau_{i^{\prime}})^{f_{i^{\prime}}}.</td>
<td></td>
<td>(17)</td>
</tr>
</tbody>
</table>
Note that the contention zones are labeled with $`x`$ regarding the indices of $`d`$. In the case of equal AIFS values, the contention zone is labeled with the index of the AC with higher priority.
Given the state transition probabilities as in Fig. [3](#S6.F3 "Figure 3 ‣ Analysis of the 802.11e Enhanced Distributed Channel Access Function **footnote *† This work is supported by the Center for Pervasive Communications and Computing, and by National Science Foundation under Grant No. 0434928. Any opinions, findings, and conclusions or recommendations expressed in this material are those of authors and do not necessarily reflect the view of the National Science Foundation.†"), the long term occupancy of the backoff slots $`b_{n}'`$ can be obtained from the steady-state solution of the Markov chain. Then, the AC-specific average collision probability $`p_{c_{i}}`$ is found by weighing zone specific collision probabilities $`p_{c_{i,x}}`$ according to the long term occupancy of contention zones (thus backoff slots)
<table>
<tbody>
<tr>
<td></td>
<td>pci=∑n=di+1Wminpci,x⋅bn′∑n=di+1Wminbn′subscript𝑝subscript𝑐𝑖superscriptsubscript𝑛subscript𝑑𝑖1subscript𝑊𝑚𝑖𝑛⋅subscript𝑝subscript𝑐𝑖𝑥subscriptsuperscript𝑏′𝑛superscriptsubscript𝑛subscript𝑑𝑖1subscript𝑊𝑚𝑖𝑛subscriptsuperscript𝑏′𝑛p_{c_{i}}=\frac{\sum_{n=d_{i}+1}^{W_{min}}p_{c_{i,x}}\cdot b^{\prime}_{n}}{\sum_{n=d_{i}+1}^{W_{min}}b^{\prime}_{n}}</td>
<td></td>
<td>(18)</td>
</tr>
</tbody>
</table>
where $`x = {\max\left( {\left. y \middle| d_{y} \right. = {\max\limits_{z}\hspace{0pt}{({\left. d_{z} \middle| d_{z} \right. \leq n})}}} \right)}`$ which shows $`x`$ is assigned the highest index value within a set of ACs that have AIFS smaller than or equal to $`n + {A\hspace{0pt}I\hspace{0pt}F\hspace{0pt}S_{3}}`$. This ensures that at backoff slot $`n`$, AC<sub>i</sub> has sensed the medium idle for AIFS<sub>x</sub>. Therefore, the calculation in ([18](#S4.E18 "In IV-A1 Average conditional collision probability 𝑝_𝑐_𝑖 ‣ IV-A Steady-State Solution ‣ IV EDCA Discrete-Time Markov Chain Model ‣ Analysis of the 802.11e Enhanced Distributed Channel Access Function **footnote *† This work is supported by the Center for Pervasive Communications and Computing, and by National Science Foundation under Grant No. 0434928. Any opinions, findings, and conclusions or recommendations expressed in this material are those of authors and do not necessarily reflect the view of the National Science Foundation.†")) fits into the definition of $`p_{c_{i,x}}`$.
Note that the average collision probability calculation in \[[12](#bib.bib12), Section IV-D\] is a special case of our calculation for two ACs.
####
IV-A2
The state duration $`T_{i,s}`$ and $`T_{i,c}`$
Let $`T_{i,p}`$ be the average payload transmission time for AC<sub>i</sub> ($`T_{i,p}`$ includes the transmission time of MAC and PHY headers), $`\delta`$ be the propagation delay, $`T_{a\hspace{0pt}c\hspace{0pt}k}`$ be the time required for acknowledgment packet (ACK) transmission. Then, for the basic access scheme, we define the time spent in a successful transmission $`T_{i,s}`$ and a collision $`T_{i,c}`$ for any AC<sub>i</sub> as
<table>
<tbody>
<tr>
<td></td>
<td>Ti,s=subscript𝑇𝑖𝑠absent\displaystyle T_{i,s}=</td>
<td>Ti,p+δ+SIFS+Tack+δ+AIFSisubscript𝑇𝑖𝑝𝛿𝑆𝐼𝐹𝑆subscript𝑇𝑎𝑐𝑘𝛿𝐴𝐼𝐹subscript𝑆𝑖\displaystyle T_{i,p}+\delta+SIFS+T_{ack}+\delta+AIFS_{i}</td>
<td></td>
<td>(19)</td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td>Ti,c=subscript𝑇𝑖𝑐absent\displaystyle T_{i,c}=</td>
<td>Ti,p∗+ACK_Timeout+AIFSisubscript𝑇𝑖superscript𝑝𝐴𝐶𝐾_𝑇𝑖𝑚𝑒𝑜𝑢𝑡𝐴𝐼𝐹subscript𝑆𝑖\displaystyle T_{i,p^{*}}+ACK\_Timeout+AIFS_{i}</td>
<td></td>
<td>(20)</td>
</tr>
</tbody>
</table>
where $`T_{i,p^{\ast}}`$ is the average transmission time of the longest packet payload involved in a collision \[[3](#bib.bib3)\]. For simplicity, we assume the packet size to be equal for any AC, then $`T_{i,p^{\ast}} = T_{i,p}`$. Being not explicitly specified in the standards, we set $`A\hspace{0pt}C\hspace{0pt}K\hspace{0pt}\_\hspace{0pt}T\hspace{0pt}i\hspace{0pt}m\hspace{0pt}e\hspace{0pt}o\hspace{0pt}u\hspace{0pt}t`$, using Extended Inter Frame Space (EIFS) as $`{E\hspace{0pt}I\hspace{0pt}F\hspace{0pt}S_{i}} - {A\hspace{0pt}I\hspace{0pt}F\hspace{0pt}S_{i}}`$.
The extensions of ([19](#S4.E19 "In IV-A2 The state duration 𝑇_{𝑖,𝑠} and 𝑇_{𝑖,𝑐} ‣ IV-A Steady-State Solution ‣ IV EDCA Discrete-Time Markov Chain Model ‣ Analysis of the 802.11e Enhanced Distributed Channel Access Function **footnote *† This work is supported by the Center for Pervasive Communications and Computing, and by National Science Foundation under Grant No. 0434928. Any opinions, findings, and conclusions or recommendations expressed in this material are those of authors and do not necessarily reflect the view of the National Science Foundation.†")) and ([20](#S4.E20 "In IV-A2 The state duration 𝑇_{𝑖,𝑠} and 𝑇_{𝑖,𝑐} ‣ IV-A Steady-State Solution ‣ IV EDCA Discrete-Time Markov Chain Model ‣ Analysis of the 802.11e Enhanced Distributed Channel Access Function **footnote *† This work is supported by the Center for Pervasive Communications and Computing, and by National Science Foundation under Grant No. 0434928. Any opinions, findings, and conclusions or recommendations expressed in this material are those of authors and do not necessarily reflect the view of the National Science Foundation.†")) for the Request-to-Send/Clear-to-Send (RTS/CTS) scheme are
<table>
<tbody>
<tr>
<td></td>
<td>Ti,s=subscript𝑇𝑖𝑠absent\displaystyle T_{i,s}=</td>
<td>Trts+δ+SIFS+Tcts+δ+SIFS+Ti,p+δ+SIFS+Tack+δ+AIFSisubscript𝑇𝑟𝑡𝑠𝛿𝑆𝐼𝐹𝑆subscript𝑇𝑐𝑡𝑠𝛿𝑆𝐼𝐹𝑆subscript𝑇𝑖𝑝𝛿𝑆𝐼𝐹𝑆subscript𝑇𝑎𝑐𝑘𝛿𝐴𝐼𝐹subscript𝑆𝑖\displaystyle T_{rts}+\delta+SIFS+T_{cts}+\delta+SIFS+T_{i,p}+\delta+SIFS+T_{ack}+\delta+AIFS_{i}</td>
<td></td>
<td>(21)</td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td>Ti,c=subscript𝑇𝑖𝑐absent\displaystyle T_{i,c}=</td>
<td>Trts+CTS_Timeout+AIFSisubscript𝑇𝑟𝑡𝑠𝐶𝑇𝑆_𝑇𝑖𝑚𝑒𝑜𝑢𝑡𝐴𝐼𝐹subscript𝑆𝑖\displaystyle T_{rts}+CTS\_Timeout+AIFS_{i}</td>
<td></td>
<td>(22)</td>
</tr>
</tbody>
</table>
where $`T_{r\hspace{0pt}t\hspace{0pt}s}`$ and $`T_{c\hspace{0pt}t\hspace{0pt}s}`$ are the time required for RTS and CTS packet transmissions respectively. Being not explicitly specified in the standards, we set $`C\hspace{0pt}T\hspace{0pt}S\hspace{0pt}\_\hspace{0pt}T\hspace{0pt}i\hspace{0pt}m\hspace{0pt}e\hspace{0pt}o\hspace{0pt}u\hspace{0pt}t`$ as we set $`A\hspace{0pt}C\hspace{0pt}K\hspace{0pt}\_\hspace{0pt}T\hspace{0pt}i\hspace{0pt}m\hspace{0pt}e\hspace{0pt}o\hspace{0pt}u\hspace{0pt}t`$.
####
IV-A3
The state duration $`T_{i,{b\hspace{0pt}s}}`$ and $`T_{i,b}`$
The average time between successive backoff counter decrements is denoted by $`T_{i,{b\hspace{0pt}s}}`$. The backoff counter decrement may be at the slot boundary of an idle backoff slot or the last slot of AIFS following an EDCA TXOP or a collision period. We start with calculating the average duration of an EDCA TXOP for AC<sub>i</sub> $`T_{i,{t\hspace{0pt}x\hspace{0pt}o\hspace{0pt}p}}`$ as
<table>
<tbody>
<tr>
<td></td>
<td>Ti,txop=∑l=0QSibi,0,−Ni,l⋅((Ni−1)⋅Ti,exc+Ti,s)+∑k=−Ni+1−1bi,0,k,0⋅((−k−1)⋅Ti,exc+Ti,s)∑k=−Ni+1−1bi,0,k,0+∑l=0QSibi,0,−Ni,lsubscript𝑇𝑖𝑡𝑥𝑜𝑝superscriptsubscript𝑙0𝑄subscript𝑆𝑖⋅subscript𝑏𝑖0subscript𝑁𝑖𝑙⋅subscript𝑁𝑖1subscript𝑇𝑖𝑒𝑥𝑐subscript𝑇𝑖𝑠superscriptsubscript𝑘subscript𝑁𝑖11⋅subscript𝑏𝑖0𝑘0⋅𝑘1subscript𝑇𝑖𝑒𝑥𝑐subscript𝑇𝑖𝑠superscriptsubscript𝑘subscript𝑁𝑖11subscript𝑏𝑖0𝑘0superscriptsubscript𝑙0𝑄subscript𝑆𝑖subscript𝑏𝑖0subscript𝑁𝑖𝑙T_{i,txop}=\frac{\sum_{l=0}^{QS_{i}}b_{i,0,-N_{i},l}\cdot((N_{i}-1)\cdot T_{i,exc}+T_{i,s})+\sum_{k=-N_{i}+1}^{-1}b_{i,0,k,0}\cdot((-k-1)\cdot T_{i,exc}+T_{i,s})}{\sum_{k=-N_{i}+1}^{-1}b_{i,0,k,0}+\sum_{l=0}^{QS_{i}}b_{i,0,-N_{i},l}}</td>
<td></td>
<td>(23)</td>
</tr>
</tbody>
</table>
where $`T_{i,{e\hspace{0pt}x\hspace{0pt}c}}`$ is defined as the duration of a successful packet exchange sequence within a TXOP. Since the packet exchanges within a TXOP are separated by SIFS rather than AIFS,
<table>
<tbody>
<tr>
<td></td>
<td>Ti,exc=Ti,s−AIFSi+SIFS,subscript𝑇𝑖𝑒𝑥𝑐subscript𝑇𝑖𝑠𝐴𝐼𝐹subscript𝑆𝑖𝑆𝐼𝐹𝑆\displaystyle T_{i,exc}=T_{i,s}-AIFS_{i}+SIFS,</td>
<td></td>
<td>(24)</td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td>Ni=max(1,⌊(TXOPi+SIFS)/Ti,exc⌋).subscript𝑁𝑖1𝑇𝑋𝑂subscript𝑃𝑖𝑆𝐼𝐹𝑆subscript𝑇𝑖𝑒𝑥𝑐\displaystyle N_{i}=\max(1,\lfloor(TXOP_{i}+SIFS)/T_{i,exc}\rfloor).</td>
<td></td>
<td>(25)</td>
</tr>
</tbody>
</table>
Given $`\tau_{i}`$ and $`f_{i}`$, simple probability theory can be used to calculate the conditional probability of no transmission ($`p_{x,i}^{i\hspace{0pt}d\hspace{0pt}l\hspace{0pt}e}`$), only one transmission from AC$`_{i'}`$ ($`p_{x,i}^{s\hspace{0pt}u\hspace{0pt}c_{i'}}`$), or at least two transmissions ($`p_{x,i}^{c\hspace{0pt}o\hspace{0pt}l}`$) at the contention zone $`x`$ given one AC<sub>i</sub> is in backoff.
<table>
<tbody>
<tr>
<td></td>
<td>px,iidle={][c]ll∏i′:di′≤dx(1−τi′)fi′,ifdi>dx∏i′:di′≤dx(1−τi′)fi′1−τi,ifdi≤dx.p_{x,i}^{idle}=\left\{\\ \begin{IEEEeqnarraybox}[]{[}][c]{ll}\prod_{i^{\prime}:d_{i^{\prime}}\leq d_{x}}(1-\tau_{i^{\prime}})^{f_{i^{\prime}}},&~{}{\rm if}~{}d_{i}>d_{x}\\ \frac{\prod_{i^{\prime}:d_{i^{\prime}}\leq d_{x}}(1-\tau_{i^{\prime}})^{f_{i^{\prime}}}}{1-\tau_{i}},&~{}{\rm if}~{}d_{i}\leq d_{x}.\end{IEEEeqnarraybox}\right.</td>
<td></td>
<td>(26)</td>
</tr>
</tbody>
</table>
<table>
<tbody>
<tr>
<td></td>
<td>px,isuci′={][c]ll0,ifdx<di′fi′τi′(1−τi′)fi′−1∏i′′:di′′≤dx(1−τi′′)fi′′,ifdi>dxanddi′≤dxfi′τi′(1−τi′)fi′−11−τi∏i′′:di′′≤dx(1−τi′′)fi′′,ifdi≤dxanddi′≤dx.p_{x,i}^{suc_{i^{\prime}}}=\left\{\\ \begin{IEEEeqnarraybox}[]{[}][c]{ll}0,&~{}{\rm if}~{}d_{x}<d_{i^{\prime}}\\ f_{i^{\prime}}\tau_{i^{\prime}}(1-\tau_{i^{\prime}})^{f_{i^{\prime}}-1}\prod_{i^{\prime\prime}:d_{i^{\prime\prime}}\leq d_{x}}(1-\tau_{i^{\prime\prime}})^{f_{i^{\prime\prime}}},&~{}{\rm if}~{}d_{i}>d_{x}~{}{\rm and}~{}d_{i^{\prime}}\leq d_{x}\\ \frac{f_{i^{\prime}}\tau_{i^{\prime}}(1-\tau_{i^{\prime}})^{f_{i^{\prime}}-1}}{1-\tau_{i}}\prod_{i^{\prime\prime}:d_{i^{\prime\prime}}\leq d_{x}}(1-\tau_{i^{\prime\prime}})^{f_{i^{\prime\prime}}},&~{}{\rm if}~{}d_{i}\leq d_{x}~{}{\rm and}~{}d_{i^{\prime}}\leq d_{x}.\end{IEEEeqnarraybox}\right.</td>
<td></td>
<td>(27)</td>
</tr>
</tbody>
</table>
<table>
<tbody>
<tr>
<td></td>
<td>px,icol=superscriptsubscript𝑝𝑥𝑖𝑐𝑜𝑙absent\displaystyle p_{x,i}^{col}=</td>
<td>1−px,iidle−∑∀i′px,isuci′1superscriptsubscript𝑝𝑥𝑖𝑖𝑑𝑙𝑒subscriptfor-allsuperscript𝑖′superscriptsubscript𝑝𝑥𝑖𝑠𝑢subscript𝑐superscript𝑖′\displaystyle 1-p_{x,i}^{idle}-\sum_{\forall i^{\prime}}p_{x,i}^{suc_{i^{\prime}}}</td>
<td></td>
<td>(28)</td>
</tr>
</tbody>
</table>
Let $`x_{i}`$ be the first contention zone in which AC<sub>i</sub> can transmit. Then,
<table>
<tbody>
<tr>
<td></td>
<td>Ti,bs=11−∑xi<x′≤3pzx′∑∀x′(px′,iidle⋅Tslot+px′,icol⋅Ti,c+∑∀i′px′,isuci′⋅Ti′,txop)⋅pzx′subscript𝑇𝑖𝑏𝑠11subscriptsubscript𝑥𝑖superscript𝑥′3subscript𝑝subscript𝑧superscript𝑥′subscriptfor-allsuperscript𝑥′⋅⋅superscriptsubscript𝑝superscript𝑥′𝑖𝑖𝑑𝑙𝑒subscript𝑇𝑠𝑙𝑜𝑡⋅superscriptsubscript𝑝superscript𝑥′𝑖𝑐𝑜𝑙subscript𝑇𝑖𝑐subscriptfor-allsuperscript𝑖′⋅superscriptsubscript𝑝superscript𝑥′𝑖𝑠𝑢subscript𝑐superscript𝑖′subscript𝑇superscript𝑖′𝑡𝑥𝑜𝑝subscript𝑝subscript𝑧superscript𝑥′\displaystyle T_{i,bs}=\frac{1}{1-\sum\limits_{x_{i}<x^{\prime}\leq 3}p_{z_{x^{\prime}}}}\sum_{\forall x^{\prime}}(p_{x^{\prime},i}^{idle}\cdot T_{slot}+p_{x^{\prime},i}^{col}\cdot T_{i,c}+\sum_{\forall i^{\prime}}p_{x^{\prime},i}^{suc_{i^{\prime}}}\cdot T_{i^{\prime},txop})\cdot p_{z_{x^{\prime}}}</td>
<td></td>
<td>(29)</td>
</tr>
</tbody>
</table>
where $`p_{z_{x}}`$ denotes the stationary distribution for a random backoff slot being in zone $`x`$. Note that, in ([29](#S4.E29 "In IV-A3 The state duration 𝑇_{𝑖,𝑏𝑠} and 𝑇_{𝑖,𝑏} ‣ IV-A Steady-State Solution ‣ IV EDCA Discrete-Time Markov Chain Model ‣ Analysis of the 802.11e Enhanced Distributed Channel Access Function **footnote *† This work is supported by the Center for Pervasive Communications and Computing, and by National Science Foundation under Grant No. 0434928. Any opinions, findings, and conclusions or recommendations expressed in this material are those of authors and do not necessarily reflect the view of the National Science Foundation.†")), the fractional term before summation accounts for the busy periods experienced before AIFS<sub>i</sub> is completed. Therefore, if we let $`d_{- 1} = W_{m\hspace{0pt}i\hspace{0pt}n}`$,
<table>
<tbody>
<tr>
<td></td>
<td>pzx=∑n=dx+1min(dx′|dx′>dx)bn′.subscript𝑝subscript𝑧𝑥superscriptsubscript𝑛subscript𝑑𝑥1𝑚𝑖𝑛subscript𝑑superscript𝑥′ketsubscript𝑑superscript𝑥′subscript𝑑𝑥superscriptsubscript𝑏𝑛′\displaystyle p_{z_{x}}=\sum_{n=d_{x}+1}^{min(d_{x^{\prime}}|d_{x^{\prime}}>d_{x})}b_{n}^{\prime}.</td>
<td></td>
<td>(30)</td>
</tr>
</tbody>
</table>
The expected duration of a backoff slot given it is busy and one AC<sub>i</sub> is in idle state is calculated as
<table>
<tbody>
<tr>
<td></td>
<td>Ti,b=∑∀x′(px′,icol1−px′,iidle⋅Ti,c+∑∀i′px′,isuci′1−px′,iidle⋅Ti′,txop)⋅pzx′.subscript𝑇𝑖𝑏subscriptfor-allsuperscript𝑥′⋅⋅superscriptsubscript𝑝superscript𝑥′𝑖𝑐𝑜𝑙1superscriptsubscript𝑝superscript𝑥′𝑖𝑖𝑑𝑙𝑒subscript𝑇𝑖𝑐subscriptfor-allsuperscript𝑖′⋅superscriptsubscript𝑝superscript𝑥′𝑖𝑠𝑢subscript𝑐superscript𝑖′1superscriptsubscript𝑝superscript𝑥′𝑖𝑖𝑑𝑙𝑒subscript𝑇superscript𝑖′𝑡𝑥𝑜𝑝subscript𝑝subscript𝑧superscript𝑥′\displaystyle T_{i,b}=\sum_{\forall x^{\prime}}\left(\frac{p_{x^{\prime},i}^{col}}{1-p_{x^{\prime},i}^{idle}}\cdot T_{i,c}+\sum_{\forall i^{\prime}}\frac{p_{x^{\prime},i}^{suc_{i^{\prime}}}}{1-p_{x^{\prime},i}^{idle}}\cdot T_{i^{\prime},txop}\right)\cdot p_{z_{x^{\prime}}}.</td>
<td></td>
<td>(31)</td>
</tr>
</tbody>
</table>
####
IV-A4
The conditional queue state transition probabilities $`p_{n\hspace{0pt}t}`$ and $`p_{s\hspace{0pt}t}`$
We assume the packets arrive at the AC queue with size $`Q\hspace{0pt}S_{i}`$ according to a Poisson process with rate $`\lambda_{i}`$ packets per second. Using the probability distribution function of the Poisson process, the probability of $`k`$ arrivals occuring in time interval $`t`$ can be calculated as
<table>
<tbody>
<tr>
<td></td>
<td>Pr(Nt,i=k)=exp−λit(λit)kk!.\displaystyle{\rm Pr}(N_{t,i}=k)=\frac{\exp^{-\lambda_{i}t}(\lambda_{i}t)^{k}}{k!}.</td>
<td></td>
<td>(32)</td>
</tr>
</tbody>
</table>
Then, $`p_{n\hspace{0pt}t}\hspace{0pt}{(l',\left. T \middle| l \right.)}`$ and $`p_{s\hspace{0pt}t}\hspace{0pt}{(l',\left. T \middle| l \right.)}`$ can be calculated as follows. Note that the finite buffer space is considered throughout calculations since the number of packets that may arrive during $`T`$ can be more than the available queue space.
<table>
<tbody>
<tr>
<td></td>
<td>pnt(l′,T|l)={][c]llPr(NT,i=l′−l),ifl′<QSi1−∑l′=lQSi−1Pr(NT,i=l′−l),ifl′=QSi.p_{nt}(l^{\prime},T|l)=\left\{\\ \begin{IEEEeqnarraybox}[]{[}][c]{ll}{\rm Pr}(N_{T,i}=l^{\prime}-l),&~{}{\rm if}~{}l^{\prime}<QS_{i}\\ 1-\sum_{l^{\prime}=l}^{QS_{i}-1}{\rm Pr}(N_{T,i}=l^{\prime}-l),&~{}{\rm if}~{}l^{\prime}=QS_{i}.\end{IEEEeqnarraybox}\right.</td>
<td></td>
<td>(33)</td>
</tr>
</tbody>
</table>
<table>
<tbody>
<tr>
<td></td>
<td>pst(l′,T|l)={][c]llPr(NT,i=l′−l+1),ifl′<QSi1−∑l′=l−1QSi−1Pr(NT,i=l′−l+1),ifl′=QSi.p_{st}(l^{\prime},T|l)=\left\{\\ \begin{IEEEeqnarraybox}[]{[}][c]{ll}{\rm Pr}(N_{T,i}=l^{\prime}-l+1),&~{}{\rm if}~{}l^{\prime}<QS_{i}\\ 1-\sum_{l^{\prime}=l-1}^{QS_{i}-1}{\rm Pr}(N_{T,i}=l^{\prime}-l+1),&~{}{\rm if}~{}l^{\prime}=QS_{i}.\end{IEEEeqnarraybox}\right.</td>
<td></td>
<td>(34)</td>
</tr>
</tbody>
</table>
Note that in ([11](#S4.E11 "In item 4 ‣ IV EDCA Discrete-Time Markov Chain Model ‣ Analysis of the 802.11e Enhanced Distributed Channel Access Function **footnote *† This work is supported by the Center for Pervasive Communications and Computing, and by National Science Foundation under Grant No. 0434928. Any opinions, findings, and conclusions or recommendations expressed in this material are those of authors and do not necessarily reflect the view of the National Science Foundation.†"))-([14](#S4.E14 "In IV-A Steady-State Solution ‣ IV EDCA Discrete-Time Markov Chain Model ‣ Analysis of the 802.11e Enhanced Distributed Channel Access Function **footnote *† This work is supported by the Center for Pervasive Communications and Computing, and by National Science Foundation under Grant No. 0434928. Any opinions, findings, and conclusions or recommendations expressed in this material are those of authors and do not necessarily reflect the view of the National Science Foundation.†")), $`\rho_{i} = {1 - {\Pr\hspace{0pt}{({N_{T_{s\hspace{0pt}l\hspace{0pt}o\hspace{0pt}t},i} = 0})}}}`$. Together with the steady-state transition probabilities, ([14](#S4.E14 "In IV-A Steady-State Solution ‣ IV EDCA Discrete-Time Markov Chain Model ‣ Analysis of the 802.11e Enhanced Distributed Channel Access Function **footnote *† This work is supported by the Center for Pervasive Communications and Computing, and by National Science Foundation under Grant No. 0434928. Any opinions, findings, and conclusions or recommendations expressed in this material are those of authors and do not necessarily reflect the view of the National Science Foundation.†"))-([34](#S4.E34 "In IV-A4 The conditional queue state transition probabilities 𝑝_{𝑛𝑡} and 𝑝_{𝑠𝑡} ‣ IV-A Steady-State Solution ‣ IV EDCA Discrete-Time Markov Chain Model ‣ Analysis of the 802.11e Enhanced Distributed Channel Access Function **footnote *† This work is supported by the Center for Pervasive Communications and Computing, and by National Science Foundation under Grant No. 0434928. Any opinions, findings, and conclusions or recommendations expressed in this material are those of authors and do not necessarily reflect the view of the National Science Foundation.†")) represent a nonlinear system which can be solved using numerical methods.
###
IV-B
Normalized Throughput Analysis
The normalized throughput of a given AC<sub>i</sub>, $`S_{i}`$, is defined as the fraction of the time occupied by the successfully transmitted information. Then,
<table>
<tbody>
<tr>
<td></td>
<td>Si=subscript𝑆𝑖absent\displaystyle S_{i}=</td>
<td>psiNi,txopTi,ppITslot+∑i′psi′Ti′,txop+(1−pI−∑i′psi′)Tcsubscript𝑝subscript𝑠𝑖subscript𝑁𝑖𝑡𝑥𝑜𝑝subscript𝑇𝑖𝑝subscript𝑝𝐼subscript𝑇𝑠𝑙𝑜𝑡subscriptsuperscript𝑖′subscript𝑝subscript𝑠superscript𝑖′subscript𝑇superscript𝑖′𝑡𝑥𝑜𝑝1subscript𝑝𝐼subscriptsuperscript𝑖′subscript𝑝subscript𝑠superscript𝑖′subscript𝑇𝑐\displaystyle\frac{p_{s_{i}}N_{i,txop}T_{i,p}}{p_{I}T_{slot}+\sum_{i^{\prime}}p_{s_{i^{\prime}}}T_{i^{\prime},txop}+(1-p_{I}-\sum_{i^{\prime}}p_{s_{i^{\prime}}})T_{c}}</td>
<td></td>
<td>(35)</td>
</tr>
</tbody>
</table>
$`p_{I}`$ is the probability of the channel being idle at a backoff slot, $`p_{s_{i}}`$ is the conditional successful transmission probability of AC<sub>i</sub> at a backoff slot, and $`N_{i,{t\hspace{0pt}x\hspace{0pt}o\hspace{0pt}p}} = {{({{T_{i,{t\hspace{0pt}x\hspace{0pt}o\hspace{0pt}p}} - {A\hspace{0pt}I\hspace{0pt}F\hspace{0pt}S_{i}}} + {S\hspace{0pt}I\hspace{0pt}F\hspace{0pt}S}})}/T_{i,{e\hspace{0pt}x\hspace{0pt}c}}}`$. Note that, we consider $`N_{i,{t\hspace{0pt}x\hspace{0pt}o\hspace{0pt}p}}`$ and $`T_{i,{t\hspace{0pt}x\hspace{0pt}o\hspace{0pt}p}}`$ in ([35](#S4.E35 "In IV-B Normalized Throughput Analysis ‣ IV EDCA Discrete-Time Markov Chain Model ‣ Analysis of the 802.11e Enhanced Distributed Channel Access Function **footnote *† This work is supported by the Center for Pervasive Communications and Computing, and by National Science Foundation under Grant No. 0434928. Any opinions, findings, and conclusions or recommendations expressed in this material are those of authors and do not necessarily reflect the view of the National Science Foundation.†")) to define the generic slot time and the time occupied by the successfully transmitted information in the case of EDCA TXOPs.
The probability of a slot being idle, $`p_{I}`$, depends on the state of previous slots. For example, conditioned on the previous slot to be busy ($`p_{B} = {1 - p_{I}}`$), $`p_{I}`$ only depends on the transmission probability of the ACs with the smallest AIFS, since others have to wait extra AIFS slots. Generalizing this to all AIFS slots, $`p_{I}`$ can be calculated as
<table>
<tbody>
<tr>
<td></td>
<td>pIsubscript𝑝𝐼\displaystyle p_{I}</td>
<td>=∑n=0WminγnpB(pI)n≅∑n=0d0−1γnpBpIn+γd0pId0absentsuperscriptsubscript𝑛0subscript𝑊𝑚𝑖𝑛subscript𝛾𝑛subscript𝑝𝐵superscriptsubscript𝑝𝐼𝑛superscriptsubscript𝑛0subscript𝑑01subscript𝛾𝑛subscript𝑝𝐵superscriptsubscript𝑝𝐼𝑛subscript𝛾subscript𝑑0superscriptsubscript𝑝𝐼subscript𝑑0\displaystyle=\sum_{n=0}^{W_{min}}\gamma_{n}p_{B}(p_{I})^{n}\cong\sum_{n=0}^{d_{0}-1}\gamma_{n}p_{B}p_{I}^{n}+\gamma_{d_{0}}p_{I}^{d_{0}}</td>
<td></td>
<td>(36)</td>
</tr>
</tbody>
</table>
where $`\gamma_{n}`$ denotes the probability of no transmission occuring at the $`{({n + 1})}^{t\hspace{0pt}h}`$ AIFS slot after $`A\hspace{0pt}I\hspace{0pt}F\hspace{0pt}S_{3}`$. Substituting $`\gamma_{n} = \gamma_{d_{0}}`$ for $`n \geq d_{0}`$, and releasing the condition on the upper limit of summation, $`W_{m\hspace{0pt}i\hspace{0pt}n}`$, to $`\infty`$, $`p_{I}`$ can be approximated as in ([36](#S4.E36 "In IV-B Normalized Throughput Analysis ‣ IV EDCA Discrete-Time Markov Chain Model ‣ Analysis of the 802.11e Enhanced Distributed Channel Access Function **footnote *† This work is supported by the Center for Pervasive Communications and Computing, and by National Science Foundation under Grant No. 0434928. Any opinions, findings, and conclusions or recommendations expressed in this material are those of authors and do not necessarily reflect the view of the National Science Foundation.†")). According to the simulation results, this approximation works well. Note that $`\gamma_{n} = {1 - p_{x}^{t\hspace{0pt}r}}`$ where $`x = {\max\left( {\left. y \middle| d_{y} \right. = {\max\limits_{z}\hspace{0pt}{({\left. d_{z} \middle| d_{z} \right. \leq n})}}} \right)}`$.
The probability of successful transmission $`p_{s_{i}}`$ is conditioned on the states of the previous slots as well. This is again because the number of stations that can contend at an arbitrary backoff slot differs depending on the number of previous consecutive idle backoff slots. Therefore, for the heterogeneous case, in which each station only has one AC, $`p_{s_{i}}`$ can be calculated as
<table>
<tbody>
<tr>
<td></td>
<td>psi=Niτi(1−τi)(∑n=di+1d0(pBpI(n−1)∏i′:0≤di′≤(n−1)(1−τi′)fi′)+(pI)d0∏∀i′(1−τi′)fi′).subscript𝑝subscript𝑠𝑖subscript𝑁𝑖subscript𝜏𝑖1subscript𝜏𝑖superscriptsubscript𝑛subscript𝑑𝑖1subscript𝑑0subscript𝑝𝐵superscriptsubscript𝑝𝐼𝑛1subscriptproduct:superscript𝑖′0subscript𝑑superscript𝑖′𝑛1superscript1subscript𝜏superscript𝑖′subscript𝑓superscript𝑖′superscriptsubscript𝑝𝐼subscript𝑑0subscriptproductfor-allsuperscript𝑖′superscript1subscript𝜏superscript𝑖′subscript𝑓superscript𝑖′\displaystyle p_{s_{i}}=\frac{N_{i}\tau_{i}}{(1-\tau_{i})}\left(\sum_{n=d_{i}+1}^{d_{0}}\left(p_{B}p_{I}^{(n-1)}\prod_{i^{\prime}:0\leq d_{i^{\prime}}\leq(n-1)}(1-\tau_{i^{\prime}})^{f_{i^{\prime}}}\right)+(p_{I})^{d_{0}}\prod_{\forall i^{\prime}}(1-\tau_{i^{\prime}})^{f_{i^{\prime}}}\right).</td>
<td></td>
<td>(37)</td>
</tr>
</tbody>
</table>
Similarly, for the scenario, in which each station has four active ACs,
<table>
<tbody>
<tr>
<td></td>
<td>psi=subscript𝑝subscript𝑠𝑖absent\displaystyle p_{s_{i}}=</td>
<td>Niτi(1−τi)(∑n=di+1d0(pBpI(n−1)∏i′:0≤di′≤(n−1)(1−τi′)fi′−1∏i′′>i(1−τi′′))\displaystyle\frac{N_{i}\tau_{i}}{(1-\tau_{i})}\left(\sum_{n=d_{i}+1}^{d_{0}}\left(p_{B}p_{I}^{(n-1)}\prod_{i^{\prime}:0\leq d_{i^{\prime}}\leq(n-1)}(1-\tau_{i^{\prime}})^{f_{i^{\prime}}-1}\prod_{i^{\prime\prime}>i}(1-\tau_{i^{\prime\prime}})\right)\right.</td>
<td></td>
<td></td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td></td>
<td>+(pI)d0∏∀i′(1−τi′)fi′−1∏i′′>i(1−τi′′)).\displaystyle+\left.(p_{I})^{d_{0}}\prod_{\forall i^{\prime}}(1-\tau_{i^{\prime}})^{f_{i^{\prime}}-1}\prod_{i^{\prime\prime}>i}(1-\tau_{i^{\prime\prime}})\right).</td>
<td></td>
<td>(38)</td>
</tr>
</tbody>
</table>
###
IV-C
Average Delay Analysis
Our goal is to find total average delay $`E\hspace{0pt}{\lbrack D_{i}\rbrack}`$ which is defined as the average time from when a packet enters the MAC layer queue of AC<sub>i</sub> until it is successfully transmitted. $`D_{i}`$ has two components; i) queueing time $`Q_{i}`$ and ii) access time $`A_{i}`$. $`Q_{i}`$ is the period that a packet waits in the queue for other packets in front to be transmitted. $`A_{i}`$ is the period a packet waits at the head of the queue until it is transmitted successfully (backoff and transmission period). We carry out a recursive calculation as in \[[11](#bib.bib11)\] to find $`E\hspace{0pt}{\lbrack A_{i}\rbrack}`$ for AC<sub>i</sub>. Then, using $`E\hspace{0pt}{\lbrack A_{i}\rbrack}`$ and $`b_{i,j,k,l}`$, we calculate $`E\hspace{0pt}{\lbrack D_{i}\rbrack}`$=$`E\hspace{0pt}{\lbrack Q_{i}\rbrack}`$+$`E\hspace{0pt}{\lbrack A_{i}\rbrack}`$. Note that, $`E\hspace{0pt}{\lbrack A_{i}\rbrack}`$ differs depending on whether the EDCA function is idle or not when the packet arrives. We will treat these cases separately. In the sequel, $`A_{i,{i\hspace{0pt}d\hspace{0pt}l\hspace{0pt}e}}`$ denotes the access delay when the EDCA function is idle at the time a packet arrives.
The recursive calculation is carried out in a bottom-to-top and left-to-right manner on the AC-specific DTMC. For the analysis, let $`A_{i}\hspace{0pt}{(j,k)}`$ denote the time delay from the current state $`(j,k,l)`$ until the packet at the head of the AC<sub>i</sub> queue is transmitted successfully ($`l \geq 1`$). The initial condition on the recursive calculation is
<table>
<tbody>
<tr>
<td></td>
<td>Ai(ri−1,0)=Ti,s.subscript𝐴𝑖subscript𝑟𝑖10subscript𝑇𝑖𝑠A_{i}(r_{i}-1,0)=T_{i,s}.</td>
<td></td>
<td>(39)</td>
</tr>
</tbody>
</table>
Recursive delay calculations for $`0 \leq j \leq {r_{i} - 1}`$ are
<table>
<tbody>
<tr>
<td></td>
<td>Ai(j,k)={Ai(j,k−1)+Ti,bs,if1≤k≤Wi,j(1−pci)Ti,s+pci(∑k′=0Wi,j+1Ai(j+1,k′)Wi,j+1+1+Ti,c),ifk=0andj≠ri−1.subscript𝐴𝑖𝑗𝑘casessubscript𝐴𝑖𝑗𝑘1subscript𝑇𝑖𝑏𝑠if1𝑘subscript𝑊𝑖𝑗1subscript𝑝subscript𝑐𝑖subscript𝑇𝑖𝑠subscript𝑝subscript𝑐𝑖superscriptsubscriptsuperscript𝑘′0subscript𝑊𝑖𝑗1subscript𝐴𝑖𝑗1superscript𝑘′subscript𝑊𝑖𝑗11subscript𝑇𝑖𝑐if𝑘0and𝑗subscript𝑟𝑖1A_{i}(j,k)=\left\{\\ \begin{array}[c]{ll}A_{i}(j,k-1)+T_{i,bs},&~{}{\rm if}~{}1\leq k\leq W_{i,j}\\ (1-p_{c_{i}})T_{i,s}+p_{c_{i}}\left(\frac{\sum_{k^{\prime}=0}^{W_{i,j+1}}A_{i}(j+1,k^{\prime})}{W_{i,j+1}+1}+T_{i,c}\right),&~{}{\rm if}~{}k=0~{}{\rm and}~{}j\neq r_{i}-1.\end{array}\right.</td>
<td></td>
<td>(40)</td>
</tr>
</tbody>
</table>
Then,
<table>
<tbody>
<tr>
<td></td>
<td>E[Ai]=∑k=0Wi,0Ai(0,k)Wi,0+1𝐸delimited-[]subscript𝐴𝑖superscriptsubscript𝑘0subscript𝑊𝑖0subscript𝐴𝑖0𝑘subscript𝑊𝑖01\displaystyle E[A_{i}]=\frac{\sum_{k=0}^{W_{i,0}}A_{i}(0,k)}{W_{i,0}+1}</td>
<td></td>
<td>(41)</td>
</tr>
</tbody>
</table>
Following the assumptions made in ([11](#S4.E11 "In item 4 ‣ IV EDCA Discrete-Time Markov Chain Model ‣ Analysis of the 802.11e Enhanced Distributed Channel Access Function **footnote *† This work is supported by the Center for Pervasive Communications and Computing, and by National Science Foundation under Grant No. 0434928. Any opinions, findings, and conclusions or recommendations expressed in this material are those of authors and do not necessarily reflect the view of the National Science Foundation.†"))-([13](#S4.E13 "In item 4 ‣ IV EDCA Discrete-Time Markov Chain Model ‣ Analysis of the 802.11e Enhanced Distributed Channel Access Function **footnote *† This work is supported by the Center for Pervasive Communications and Computing, and by National Science Foundation under Grant No. 0434928. Any opinions, findings, and conclusions or recommendations expressed in this material are those of authors and do not necessarily reflect the view of the National Science Foundation.†")) and considering the packet loss probability due to the retry limit as $`p_{l,r} = {(p_{c_{i}})}^{r_{i}}`$ (note that the delay a dropped packet experiences cannot be considered in a total delay calculation), $`E\hspace{0pt}{\lbrack A_{i,{i\hspace{0pt}d\hspace{0pt}l\hspace{0pt}e}}\rbrack}`$ can be calculated as
<table>
<tbody>
<tr>
<td></td>
<td>E[Ai,idle]=Ti,s⋅(1−pci)+(E[Ai]+Ti,b)⋅pci⋅(1−pl,r).𝐸delimited-[]subscript𝐴𝑖𝑖𝑑𝑙𝑒⋅subscript𝑇𝑖𝑠1subscript𝑝subscript𝑐𝑖⋅𝐸delimited-[]subscript𝐴𝑖subscript𝑇𝑖𝑏subscript𝑝subscript𝑐𝑖1subscript𝑝𝑙𝑟E[A_{i,idle}]=T_{i,s}\cdot(1-p_{c_{i}})+(E[A_{i}]+T_{i,b})\cdot p_{c_{i}}\cdot(1-p_{l,r}).</td>
<td></td>
<td>(42)</td>
</tr>
</tbody>
</table>
In this case, the average access delay is equal to the total average delay, i.e., $`{D_{i}\hspace{0pt}{(0,0,0)}} = {E\hspace{0pt}{\lbrack A_{i,{i\hspace{0pt}d\hspace{0pt}l\hspace{0pt}e}}\rbrack}}`$.
We perform another recursive calculation to calculate the total delay a packet experiences $`D_{i}\hspace{0pt}{(j,k,l)}`$ (given that the packet arrives while the EDCA function is at state $`(j,k,l)`$). In the calculations, we account for the remaining access delay for the packet at the head of the MAC queue and the probability that this packet may be dropped due to the retry limit.
Let $`A_{i,d}\hspace{0pt}{(j,k)}`$ be the access delay conditioned that the packet drops. $`A_{i,d}\hspace{0pt}{(j,k)}`$ can easily be calculated by modifying the recursive method of calculating $`A_{i}\hspace{0pt}{(j,k)}`$. The initial condition on this recursive calculation is
<table>
<tbody>
<tr>
<td></td>
<td>Ai,d(ri−1,0)=Ti,c.subscript𝐴𝑖𝑑subscript𝑟𝑖10subscript𝑇𝑖𝑐A_{i,d}(r_{i}-1,0)=T_{i,c}.</td>
<td></td>
<td>(43)</td>
</tr>
</tbody>
</table>
Recursive delay calculations for $`0 \leq j \leq {r_{i} - 1}`$ are
<table>
<tbody>
<tr>
<td></td>
<td>Ai,d(j,k)={][c]llAi(j,k−1)+Ti,bs,if1≤k≤Wi,j∑k′=0Wi,j+1Ai,d(j+1,k)+Ti,c,ifk=0andj≠ri−1.A_{i,d}(j,k)=\left\{\\ \begin{IEEEeqnarraybox}[]{[}][c]{ll}A_{i}(j,k-1)+T_{i,bs},&~{}{\rm if}~{}1\leq k\leq W_{i,j}\\ \sum_{k^{\prime}=0}^{W_{i,j+1}}A_{i,d}(j+1,k)+T_{i,c},&~{}{\rm if}~{}k=0~{}{\rm and}~{}j\neq r_{i}-1.\end{IEEEeqnarraybox}\right.</td>
<td></td>
<td>(44)</td>
</tr>
</tbody>
</table>
Then,
<table>
<tbody>
<tr>
<td></td>
<td>E[Ai,d]=∑k=0Wi,0Ai,d(0,k)Wi,0+1𝐸delimited-[]subscript𝐴𝑖𝑑superscriptsubscript𝑘0subscript𝑊𝑖0subscript𝐴𝑖𝑑0𝑘subscript𝑊𝑖01\displaystyle E[A_{i,d}]=\frac{\sum_{k=0}^{W_{i,0}}A_{i,d}(0,k)}{W_{i,0}+1}</td>
<td></td>
<td>(45)</td>
</tr>
</tbody>
</table>
If a packet arrives during the backoff of another packet, it is delayed at least for the remaining access time. Depending on the queue size, it may be transmitted at the current TXOP, or may be delayed till further accesses are gained. Then, for $`0 \leq j \leq {r_{i} - 1}`$, $`0 \leq k \leq W_{i,j}`$, and $`1 \leq l \leq {Q\hspace{0pt}S_{i}}`$,
<table>
<tbody>
<tr>
<td></td>
<td>Di(j,k,l)=subscript𝐷𝑖𝑗𝑘𝑙absent\displaystyle D_{i}(j,k,l)=</td>
<td>(1−pl,r)⋅(Ai(j,k)+min(Ni−1,l−1)⋅Ti,exc+Di(−1,−1,l−Ni))⋅1subscript𝑝𝑙𝑟subscript𝐴𝑖𝑗𝑘⋅subscript𝑁𝑖1𝑙1subscript𝑇𝑖𝑒𝑥𝑐subscript𝐷𝑖11𝑙subscript𝑁𝑖\displaystyle(1-p_{l,r})\cdot\left(A_{i}(j,k)+\min(N_{i}-1,l-1)\cdot T_{i,exc}+D_{i}(-1,-1,l-N_{i})\right)</td>
<td></td>
<td></td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td></td>
<td>+pl,r⋅(Ai,d(j,k)+Di(−1,−1,l−1)).⋅subscript𝑝𝑙𝑟subscript𝐴𝑖𝑑𝑗𝑘subscript𝐷𝑖11𝑙1\displaystyle+p_{l,r}\cdot\left(A_{i,d}(j,k)+D_{i}(-1,-1,l-1)\right).</td>
<td></td>
<td>(46)</td>
</tr>
</tbody>
</table>
When the packet arrives during postbackoff, the total delay is equal to the access delay. Then, for $`0 \leq k \leq W_{i,j}`$ and $`l = 0`$,
<table>
<tbody>
<tr>
<td></td>
<td>Di(j,k,l)=Ai(j,k).subscript𝐷𝑖𝑗𝑘𝑙subscript𝐴𝑖𝑗𝑘\displaystyle D_{i}(j,k,l)=A_{i}(j,k).</td>
<td></td>
<td>(47)</td>
</tr>
</tbody>
</table>
When the packet arrives during a TXOP, it may be transmitted at the current TXOP, or it may wait for further accesses. Then, for $`{{- N_{i}} + 1} \leq k \leq {- 1}`$ and $`1 \leq l \leq {Q\hspace{0pt}S_{i}}`$,
<table>
<tbody>
<tr>
<td></td>
<td>Di(j,k,l)=min(k−1,l)⋅Ti,exc+Di(−1,−1,l−k+1).subscript𝐷𝑖𝑗𝑘𝑙⋅𝑘1𝑙subscript𝑇𝑖𝑒𝑥𝑐subscript𝐷𝑖11𝑙𝑘1\displaystyle D_{i}(j,k,l)=\min(k-1,l)\cdot T_{i,exc}+D_{i}(-1,-1,l-k+1).</td>
<td></td>
<td>(48)</td>
</tr>
</tbody>
</table>
$`D_{i}\hspace{0pt}{({- 1},{- 1},l)}`$ is calculated recursively according to the value of $`l`$
<table>
<tbody>
<tr>
<td></td>
<td>Di(−1,−1,l)={0,ifl≤0E[Ai]⋅(1−pl,r),ifl=1χ,ifl>1subscript𝐷𝑖11𝑙cases0if𝑙0⋅𝐸delimited-[]subscript𝐴𝑖1subscript𝑝𝑙𝑟if𝑙1𝜒if𝑙1D_{i}(-1,-1,l)=\left\{\\ \begin{array}[c]{ll}0,&~{}{\rm if}~{}l\leq 0\\ E[A_{i}]\cdot(1-p_{l,r}),&~{}{\rm if}~{}l=1\\ \chi,&~{}{\rm if}~{}l>1\end{array}\right.</td>
<td></td>
<td>(49)</td>
</tr>
</tbody>
</table>
where
<table>
<tbody>
<tr>
<td></td>
<td>χ=𝜒absent\displaystyle\chi=</td>
<td>(1−pl,r)⋅(E[Ai]+min(Ni−1,l−1)⋅Ti,exc\displaystyle(1-p_{l,r})\cdot\left(E[A_{i}]+\min(N_{i}-1,l-1)\cdot T_{i,exc}\right.</td>
<td></td>
<td></td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td></td>
<td>+Di(−1,−1,l−Ni))+pl,r⋅(E[Ai,d]+Di(−1,−1,l−1)).\displaystyle\left.+D_{i}(-1,-1,l-N_{i})\right)+p_{l,r}\cdot\left(E[A_{i,d}]+D_{i}(-1,-1,l-1)\right).</td>
<td></td>
<td>(50)</td>
</tr>
</tbody>
</table>
Let the probability of any arriving packet seeing the EDCA function at state $`(j,k,l)`$ be $`{\overline{b}}_{i,j,k,l}`$. Since we assume independent and exponentially distributed packet interarrivals, $`{\overline{b}}_{i,j,k,l}`$ can simply be calculated by normalizing $`b_{i,j,k,l}`$ excluding the states in which no time passes, i.e., $`\forall{(j,k,l)}`$ such that $`({{0,{- N_{i}},1} \leq l \leq {Q\hspace{0pt}S_{i}}})`$ or $`({{{0,{- N_{i}}} \leq k \leq {- 1}},0})`$. Note that $`{\overline{b}}_{i,j,k,l}`$ is zero for these states
<table>
<tbody>
<tr>
<td></td>
<td>b¯i,j,k,l=bi,j,k,l1−∑l=1QSibi,0,−Ni,l−∑k=−Ni−1bi,0,k,0.subscript¯𝑏𝑖𝑗𝑘𝑙subscript𝑏𝑖𝑗𝑘𝑙1superscriptsubscript𝑙1𝑄subscript𝑆𝑖subscript𝑏𝑖0subscript𝑁𝑖𝑙superscriptsubscript𝑘subscript𝑁𝑖1subscript𝑏𝑖0𝑘0\bar{b}_{i,j,k,l}=\frac{b_{i,j,k,l}}{1-\sum_{l=1}^{QS_{i}}b_{i,0,-N_{i},l}-\sum_{k=-N_{i}}^{-1}b_{i,0,k,0}}.</td>
<td></td>
<td>(51)</td>
</tr>
</tbody>
</table>
Then, the total average delay a successful packet experiences $`E\hspace{0pt}{\lbrack D_{i}\rbrack}`$ can be calculated averaging $`D_{i}\hspace{0pt}{(j,k,l)}`$ over all possible states
<table>
<tbody>
<tr>
<td></td>
<td>E[Di]=E[Ai,idle]⋅b¯i,0,0,0+∑∀(j,k,l)/(0,0,0)Di(j,k,l)⋅b¯i,j,k,l.𝐸delimited-[]subscript𝐷𝑖⋅𝐸delimited-[]subscript𝐴𝑖𝑖𝑑𝑙𝑒subscript¯𝑏𝑖000subscriptfor-all𝑗𝑘𝑙000⋅subscript𝐷𝑖𝑗𝑘𝑙subscript¯𝑏𝑖𝑗𝑘𝑙\displaystyle E[D_{i}]=E[A_{i,idle}]\cdot\bar{b}_{i,0,0,0}+\sum_{\forall(j,k,l)/(0,0,0)}D_{i}(j,k,l)\cdot\bar{b}_{i,j,k,l}.</td>
<td></td>
<td>(52)</td>
</tr>
</tbody>
</table>
###
IV-D
Average Packet Loss Ratio
We consider two types of packet losses; i) the packet is dropped when the MAC layer retry limit is reached, ii) the packet is dropped if the MAC queue is full at the time of packet arrival. Let $`p\hspace{0pt}l\hspace{0pt}r_{i}`$ denote the average packet loss ratio for AC<sub>i</sub>. We use the steady-state probability $`b_{i,j,k,l}`$ to find the probability whether the MAC queue is full or not at the time of packet arrival. If the queue is full, the arriving packet is dropped (second term in ([53](#S4.E53 "In IV-D Average Packet Loss Ratio ‣ IV EDCA Discrete-Time Markov Chain Model ‣ Analysis of the 802.11e Enhanced Distributed Channel Access Function **footnote *† This work is supported by the Center for Pervasive Communications and Computing, and by National Science Foundation under Grant No. 0434928. Any opinions, findings, and conclusions or recommendations expressed in this material are those of authors and do not necessarily reflect the view of the National Science Foundation.†"))). Otherwise, the packet is dropped with probability $`p_{c_{i}}^{r_{i}}`$, i.e. only if the retry limit is reached (first term in ([53](#S4.E53 "In IV-D Average Packet Loss Ratio ‣ IV EDCA Discrete-Time Markov Chain Model ‣ Analysis of the 802.11e Enhanced Distributed Channel Access Function **footnote *† This work is supported by the Center for Pervasive Communications and Computing, and by National Science Foundation under Grant No. 0434928. Any opinions, findings, and conclusions or recommendations expressed in this material are those of authors and do not necessarily reflect the view of the National Science Foundation.†"))). Note that we consider packet retransmissions only due to packet collisions. Then,
<table>
<tbody>
<tr>
<td></td>
<td>plri=∑j=0ri−1∑k=0Wi,j∑l=0QSi−1bi,j,k,l⋅pciri+∑j=0ri−1∑k=0Wi,jbi,j,k,QSi.𝑝𝑙subscript𝑟𝑖superscriptsubscript𝑗0subscript𝑟𝑖1superscriptsubscript𝑘0subscript𝑊𝑖𝑗superscriptsubscript𝑙0𝑄subscript𝑆𝑖1⋅subscript𝑏𝑖𝑗𝑘𝑙superscriptsubscript𝑝subscript𝑐𝑖subscript𝑟𝑖superscriptsubscript𝑗0subscript𝑟𝑖1superscriptsubscript𝑘0subscript𝑊𝑖𝑗subscript𝑏𝑖𝑗𝑘𝑄subscript𝑆𝑖plr_{i}=\sum_{j=0}^{r_{i}-1}\sum_{k=0}^{W_{i,j}}\sum_{l=0}^{QS_{i}-1}b_{i,j,k,l}\cdot p_{c_{i}}^{r_{i}}+\sum_{j=0}^{r_{i}-1}\sum_{k=0}^{W_{i,j}}b_{i,j,k,QS_{i}}.</td>
<td></td>
<td>(53)</td>
</tr>
</tbody>
</table>
###
IV-E
Queue Size Distribution
Due to the specific structure of the proposed model, it is straightforward to calculate the MAC queue size distribution for $`A\hspace{0pt}C_{i}`$. Note that we use queue size distribution in the calculation of average packet loss ratio.
<table>
<tbody>
<tr>
<td></td>
<td>Pr(l=l′)=∑j=0ri−1∑k=0Wi,jbi,j,k,l′.Pr𝑙superscript𝑙′superscriptsubscript𝑗0subscript𝑟𝑖1superscriptsubscript𝑘0subscript𝑊𝑖𝑗subscript𝑏𝑖𝑗𝑘superscript𝑙′{\rm Pr}(l=l^{\prime})=\sum_{j=0}^{r_{i}-1}\sum_{k=0}^{W_{i,j}}b_{i,j,k,l^{\prime}}.</td>
<td></td>
<td>(54)</td>
</tr>
</tbody>
</table>
## V Numerical and Simulation Results
We validate the accuracy of the numerical results calculated via the proposed EDCA model by comparing them with the simulations results obtained from ns-2 \[[46](#bib.bib46)\]. For the simulations, we employ the IEEE 802.11e HCF MAC simulation model for ns-2.28 that we developed \[[47](#bib.bib47)\]. This module implements all the EDCA and HCCA functionalities stated in \[[2](#bib.bib2)\].
As in all work on the subject in the literature, we consider ACs that transmit fixed-size User Datagram Protocol (UDP) packets. In simulations, we consider two ACs, one high priority and one low priority. Each station runs only one traffic class. Unless otherwise stated, the packets are generated according to a Poisson process with equal rate for both ACs. We set $`{A\hspace{0pt}I\hspace{0pt}F\hspace{0pt}S\hspace{0pt}N_{1}} = 3`$, $`{A\hspace{0pt}I\hspace{0pt}F\hspace{0pt}S\hspace{0pt}N_{3}} = 2`$, $`{C\hspace{0pt}W_{1,{m\hspace{0pt}i\hspace{0pt}n}}} = 15`$, $`{C\hspace{0pt}W_{3,{m\hspace{0pt}i\hspace{0pt}n}}} = 7`$, $`m_{1} = m_{3} = 3`$, $`r_{1} = r_{3} = 7`$. For both ACs, the payload size is 1034 bytes. Again, as in most of the work on the subject, the simulation results are reported for the wireless channel which is assumed to be not prone to any errors during transmission. The errored channel case is left for future study. All the stations have 802.11g Physical Layer (PHY) using 54 Mbps and 6 Mbps as the data and basic rate respectively ($`T_{s\hspace{0pt}l\hspace{0pt}o\hspace{0pt}t} = {9\hspace{0pt}\mu\hspace{0pt}s}`$, $`{S\hspace{0pt}I\hspace{0pt}F\hspace{0pt}S} = {10\hspace{0pt}\mu\hspace{0pt}s}`$) \[[48](#bib.bib48)\]. The simulation runtime is 100 seconds.
Fig. [4](#S6.F4 "Figure 4 ‣ Analysis of the 802.11e Enhanced Distributed Channel Access Function **footnote *† This work is supported by the Center for Pervasive Communications and Computing, and by National Science Foundation under Grant No. 0434928. Any opinions, findings, and conclusions or recommendations expressed in this material are those of authors and do not necessarily reflect the view of the National Science Foundation.†") shows the differentiation of throughput for two ACs when EDCA TXOP limits of both are set to 0 (1 packet exchange per EDCA TXOP). In this scenario, there are 5 stations for both ACs and they are transmitting to an AP. The normalized throughput per AC as well as the total system throughput is plotted for increasing offered load per AC. We have carried out the analysis for maximum MAC buffer sizes of 2 packets and 10 packets. The comparison between analytical and simulation results shows that our model can accurately capture the linear relationship between throughput and offered load under low loads, the complex transition in throughput between under-loaded and saturation regimes, and the saturation throughput. Although we do not present here, considerable inaccuracy is observed if the postbackoff procedure, varying collision probability among different AIFS zones, and varying service time among different backoff stages are not modeled correctly as proposed. The results also present that the slot homogeneity assumption works accurately in a non-saturated model for throughput estimation.
The proposed model can also capture the throughput variation with respect to the size of the MAC buffer. The results reveal how significantly the size of the MAC buffer affects the throughput in the transition period from underloaded to highly loaded channel. This also shows small interface buffer assumptions of previous models \[[27](#bib.bib27)\],\[[28](#bib.bib28)\],\[[29](#bib.bib29)\],\[[40](#bib.bib40)\] can lead to considerable analytical inaccuracies. Although the total throughput for the small buffer size case has higher throughput in the transition region for the specific example, this cannot be generalized. The reason for this is that AC<sub>1</sub> suffers from low throughput for $`{Q\hspace{0pt}S_{1}} = 10`$ due to the selection of EDCA parameters, which affects the total throughput.
It is also important to note that the throughput performance does not differ significantly (around %1-%2) for buffer sizes larger than 10 packets for the given scenarios. Therefore, we do not include such cases in order not to complicate the figures. Since the complexity of the mathematical solution increases with the increasing size of the third dimension of DTMC, it may be preferable to implement the model for smaller queue sizes when the throughput performance is not expected to be affected by the selection.
Fig. [5](#S6.F5 "Figure 5 ‣ Analysis of the 802.11e Enhanced Distributed Channel Access Function **footnote *† This work is supported by the Center for Pervasive Communications and Computing, and by National Science Foundation under Grant No. 0434928. Any opinions, findings, and conclusions or recommendations expressed in this material are those of authors and do not necessarily reflect the view of the National Science Foundation.†") depicts the differentiation of throughput for two ACs when EDCA TXOP limits are set to $`1.504`$ ms and $`3.008`$ ms for high and low priority ACs respectively. For TXOP limits, we use the suggested values for voice and video ACs in \[[2](#bib.bib2)\]. It is important to note that the model works for an arbitrary selection of the TXOP limit. According to the selected TXOP limits, $`N_{1} = 5`$ and $`N_{2} = 11`$. The normalized throughput per AC as well as the total system throughput is plotted while increasing offered load per AC. We have done the analysis for maximum MAC buffer sizes of 2 packets and 10 packets. The model accurately captures the throughput for any traffic load. As expected, increasing maximum buffer size to 10 packets increases the throughput both in the transition and the saturation region. Note that when more than a packet fits into EDCA TXOPs, this decreases contention overhead which in turn increases channel utilization and throughput (comparison of Fig. [5](#S6.F5 "Figure 5 ‣ Analysis of the 802.11e Enhanced Distributed Channel Access Function **footnote *† This work is supported by the Center for Pervasive Communications and Computing, and by National Science Foundation under Grant No. 0434928. Any opinions, findings, and conclusions or recommendations expressed in this material are those of authors and do not necessarily reflect the view of the National Science Foundation.†") with Fig. [4](#S6.F4 "Figure 4 ‣ Analysis of the 802.11e Enhanced Distributed Channel Access Function **footnote *† This work is supported by the Center for Pervasive Communications and Computing, and by National Science Foundation under Grant No. 0434928. Any opinions, findings, and conclusions or recommendations expressed in this material are those of authors and do not necessarily reflect the view of the National Science Foundation.†")). Although corresponding results are not presented here, the model works accurately for higher queue sizes in the case of EDCA TXOPs as well.
Fig. [6](#S6.F6 "Figure 6 ‣ Analysis of the 802.11e Enhanced Distributed Channel Access Function **footnote *† This work is supported by the Center for Pervasive Communications and Computing, and by National Science Foundation under Grant No. 0434928. Any opinions, findings, and conclusions or recommendations expressed in this material are those of authors and do not necessarily reflect the view of the National Science Foundation.†") displays the differentiation of throughput for two ACs when packet arrival rate is fixed to $`2`$ Mbps and the station number per AC is increased. We have done the analysis for the MAC buffer size of 10 packets with EDCA TXOPs enabled. The analytical and simulation results are well in accordance. As the traffic load increases, the differentiation in throughput between the ACs is observed.
Fig. [7](#S6.F7 "Figure 7 ‣ Analysis of the 802.11e Enhanced Distributed Channel Access Function **footnote *† This work is supported by the Center for Pervasive Communications and Computing, and by National Science Foundation under Grant No. 0434928. Any opinions, findings, and conclusions or recommendations expressed in this material are those of authors and do not necessarily reflect the view of the National Science Foundation.†") shows the normalized throughput for two ACs when offered load per AC is not equal. In this scenario, we set the packet arrival rate per AC<sub>1</sub> to $`2`$ Mbps and the packet arrival rate per AC<sub>3</sub> to $`0.5`$ Mbps. The analytical and simulation results are well in accordance. As the traffic load increases, AC<sub>3</sub> maintains linear increase with respect to offered load, while AC<sub>1</sub> experiences decrease in throughput due to larger settings of AIFS and CW if the total number of stations exceeds 22.
In the design of the model, we assume constant packet arrival probability per state. The Poisson arrival process fits this definition because of the independent exponentially distributed interarrival times. We have also compared the throughput estimates obtained from the analytical model with the simulation results obtained using an On/Off traffic model in Fig. [8](#S6.F8 "Figure 8 ‣ Analysis of the 802.11e Enhanced Distributed Channel Access Function **footnote *† This work is supported by the Center for Pervasive Communications and Computing, and by National Science Foundation under Grant No. 0434928. Any opinions, findings, and conclusions or recommendations expressed in this material are those of authors and do not necessarily reflect the view of the National Science Foundation.†"). A similar study has first been made for DCF in \[[27](#bib.bib27)\]. We modeled the high priority with On/Off traffic model with exponentially distributed idle and active intervals of mean length $`1.5`$ s. In the active interval, packets are generated with Constant Bit Rate (CBR). The low priority traffic uses Poisson distributed arrivals. Note that we leave the packet size unchanged, but normalize the packet arrival rate according to the on/off pattern so that total offered load remains constant to have a fair comparison. The analytical predictions closely follow the simulation results for the given scenario. We have observed that the predictions are more sensitive if the transition region is entered with a few number of stations (5 stations per AC).
Our model also provides a very good match in terms of the throughput for CBR traffic. In Fig. [9](#S6.F9 "Figure 9 ‣ Analysis of the 802.11e Enhanced Distributed Channel Access Function **footnote *† This work is supported by the Center for Pervasive Communications and Computing, and by National Science Foundation under Grant No. 0434928. Any opinions, findings, and conclusions or recommendations expressed in this material are those of authors and do not necessarily reflect the view of the National Science Foundation.†"), we compare the throughput prediction of the proposed model with simulations using CBR traffic. The packet arrival rate is fixed to $`2`$ Mbps for both ACs and the station number per AC is increased. MAC buffer size is 10 packets and EDCA TXOPs are enabled.
Fig. [10](#S6.F10 "Figure 10 ‣ Analysis of the 802.11e Enhanced Distributed Channel Access Function **footnote *† This work is supported by the Center for Pervasive Communications and Computing, and by National Science Foundation under Grant No. 0434928. Any opinions, findings, and conclusions or recommendations expressed in this material are those of authors and do not necessarily reflect the view of the National Science Foundation.†") depicts the total average packet delay with respect to increasing traffic load per AC. We present the results for two different scenarios. In the first scenario, TXOP limits are set to 0 ms for both ACs. In the second scenario, TXOP limits are set to $`1.504`$ ms and $`3.008`$ ms for high and low priority ACs respectively. The analysis is carried out for a buffer size of 10 packets. As the results imply, the analytical results closely follow the simulation results for both scenarios. In the lightly loaded region, the delays are considerably small. The increase in the transition region is steeper when TXOP limits are 0. In the specific example, enabling TXOPs decreases the total delay where the decrease is more considerable for the low priority AC (due to selection of parameters). Since the buffer size is limited, the total average delay converges to a specific value as the load increases. Still this limit is not of interest, since the packet loss rate at this region is unpractically large. Note that this limit will be higher for larger buffers. The region of interest is the start of the transition region (between 2 Mbps and 3 Mbps for the example in Fig. [10](#S6.F10 "Figure 10 ‣ Analysis of the 802.11e Enhanced Distributed Channel Access Function **footnote *† This work is supported by the Center for Pervasive Communications and Computing, and by National Science Foundation under Grant No. 0434928. Any opinions, findings, and conclusions or recommendations expressed in this material are those of authors and do not necessarily reflect the view of the National Science Foundation.†")). On the other hand, we also display other data points to show the performance of the model for the whole load span.
Fig. [11](#S6.F11 "Figure 11 ‣ Analysis of the 802.11e Enhanced Distributed Channel Access Function **footnote *† This work is supported by the Center for Pervasive Communications and Computing, and by National Science Foundation under Grant No. 0434928. Any opinions, findings, and conclusions or recommendations expressed in this material are those of authors and do not necessarily reflect the view of the National Science Foundation.†") depicts the average packet loss ratio with respect to increasing traffic load per AC. We present the results for two different scenarios. In the first scenario, TXOP limits are set to 0 ms for both ACs. In the second scenario, TXOP limits are set to $`1.504`$ ms and $`3.008`$ ms for high and low priority ACs respectively. The analysis is carried out for a buffer size of 10 packets. As the results imply, the analytical results closely follow the simulation results for both scenarios. Although it is not presented in Fig. [11](#S6.F11 "Figure 11 ‣ Analysis of the 802.11e Enhanced Distributed Channel Access Function **footnote *† This work is supported by the Center for Pervasive Communications and Computing, and by National Science Foundation under Grant No. 0434928. Any opinions, findings, and conclusions or recommendations expressed in this material are those of authors and do not necessarily reflect the view of the National Science Foundation.†"), the packet loss ratio drops exponentially to 0 when the offered load per AC is lower than 2.5 Mbps.
The results presented in this paper fixes the AIFS and CW parameters for each AC. The results are compared for different TXOP values at varying traffic load. Therefore, the presented results can mainly indicate the effects of TXOP on the maximum throughput. The model can also be used in order to investigate the effects of AIFS and CW on the maximum throughput.
As the comparison of Fig. [4](#S6.F4 "Figure 4 ‣ Analysis of the 802.11e Enhanced Distributed Channel Access Function **footnote *† This work is supported by the Center for Pervasive Communications and Computing, and by National Science Foundation under Grant No. 0434928. Any opinions, findings, and conclusions or recommendations expressed in this material are those of authors and do not necessarily reflect the view of the National Science Foundation.†") and Fig. [5](#S6.F5 "Figure 5 ‣ Analysis of the 802.11e Enhanced Distributed Channel Access Function **footnote *† This work is supported by the Center for Pervasive Communications and Computing, and by National Science Foundation under Grant No. 0434928. Any opinions, findings, and conclusions or recommendations expressed in this material are those of authors and do not necessarily reflect the view of the National Science Foundation.†") reveals, the total throughput can be maximized with the introduction of EDCA TXOPs which enable multiple frame transmissions in one channel access (note that MAC buffer sizes for each AC should be equal to or larger than the number of packets that can fit to the AC-specific TXOP in order to efficiently utilize each TXOP gained). EDCA TXOPs decrease the channel contention overhead and the ACs can efficiently utilize the resources. Note also that the effects of EDCA TXOPs in the lightly loaded region is marginal compared to highly loaded region. This is expected since the MAC queues do not build up in the lightly loaded scenario where stations usually have just one packet to send at their access to the channel.
As Fig. [4](#S6.F4 "Figure 4 ‣ Analysis of the 802.11e Enhanced Distributed Channel Access Function **footnote *† This work is supported by the Center for Pervasive Communications and Computing, and by National Science Foundation under Grant No. 0434928. Any opinions, findings, and conclusions or recommendations expressed in this material are those of authors and do not necessarily reflect the view of the National Science Foundation.†") shows the saturation throughput is usually less than the maximum throughput that can be obtained. This is also observed for DCF in \[[3](#bib.bib3)\]. Similarly, in Fig. [6](#S6.F6 "Figure 6 ‣ Analysis of the 802.11e Enhanced Distributed Channel Access Function **footnote *† This work is supported by the Center for Pervasive Communications and Computing, and by National Science Foundation under Grant No. 0434928. Any opinions, findings, and conclusions or recommendations expressed in this material are those of authors and do not necessarily reflect the view of the National Science Foundation.†")-Fig. [9](#S6.F9 "Figure 9 ‣ Analysis of the 802.11e Enhanced Distributed Channel Access Function **footnote *† This work is supported by the Center for Pervasive Communications and Computing, and by National Science Foundation under Grant No. 0434928. Any opinions, findings, and conclusions or recommendations expressed in this material are those of authors and do not necessarily reflect the view of the National Science Foundation.†"), the total throughput slightly decreases as the total load increases. As the load in the system increases the collision overhead becomes significant which decreases the total channel utilization. On the other hand, as also discussed in \[[3](#bib.bib3)\], the point where the maximum throughput is observed is unstable in a random access system. Therefore, a good admission control algorithm should be defined to operate the system at the point right before the lightly loaded to highly loaded transition region starts.
## VI Conclusion
We have presented an accurate Markov model for analytically calculating the EDCA throughput and delay for the whole traffic load range from a lightly loaded non-saturated channel to a heavily congested saturated medium. The presented model shows the accuracy of the homogeneous slot assumption (constant collision and transmission probability at an arbitrary backoff slot) that is extensively studied in saturation scenarios for the whole traffic range. The presented model accurately captures the linear relationship between throughput and offered load under low loads and the limiting behavior of throughput at saturation.
The key contribution of this paper is that the model accounts for all of the differentiation mechanisms EDCA proposes. The analytical model can incorporate any selection of AC-specific AIFS, CW, and TXOP values for any number of ACs. The model also considers varying collision probabilities at different contention zones which provides accurate AIFS differentiation analysis. Although not presented explicitly in this paper, it is straightforward to extend the presented model for scenarios where the stations run multiple ACs (virtual collisions may take place) or RTS/CTS protection mechanism is used. The approximations made for the sake of DTMC simplicity and symmetry may also be removed easily for increased accuracy, although they are shown to be highly accurate.
We also show that the MAC buffer size affects the EDCA performance significantly between underloaded and saturation regimes (including saturation) especially when EDCA TXOPs are enabled. The presented model captures this complex transition accurately. This analysis also points out the fact that including an accurate queue treatment is vital. Incorporating MAC queue states also enables EDCA TXOP analysis so that the EDCA TXOP continuation process is modeled in considerable detail. To the authors’ knowledge this is the first demonstration of an analytic model including EDCA TXOP procedure for finite load.
It is also worth noting that our model can easily be simplified to model DCF behavior. Moreover, after modifying our model accordingly, the throughput analysis for the infrastructure WLAN where there are transmissions both in the uplink and downlink can be performed (note that in a WLAN downlink traffic load may significantly differ from uplink traffic load).
Although the Markov analysis assumes the packets are generated according to Poisson process, the comparison with simulation results shows that the throughput analysis is valid for a range of traffic types such as CBR and On/Off traffic (On/Off traffic model is a widely used model for voice and telnet traffic).
The non-existence of a closed-form solution for the Markov model limits its practical use. On the other hand, the accurate saturation throughput analysis can highlight the strengths and the shortcomings of EDCA for varying scenarios and can provide invaluable insights. The model can effectively assist EDCA parameter adaptation or a call admission control algorithm for improved QoS support in the WLAN.
## References
- \[1\] *IEEE Standard 802.11: Wireless LAN medium access control (MAC) and physical layer (PHY) specifications*, IEEE 802.11 Std., 1999.
- \[2\] *IEEE Standard 802.11: Wireless LAN medium access control (MAC) and physical layer (PHY) specifications: Medium access control (MAC) Quality of Service (QoS) Enhancements*, IEEE 802.11e Std., 2005.
- \[3\] G. Bianchi, “Performance Analysis of the IEEE 802.11 Distributed Coordination Function,” *IEEE Trans. Commun.*, pp. 535–547, March 2000.
- \[4\] F. Cali, M. Conti, and E. Gregori, “IEEE 802.11 Wireless LAN: Capacity Analysis and Protocol Enhancement,” in *Proc. IEEE Infocom ’98*, March 1998.
- \[5\] ——, “Dynamic Tuning of the IEEE 802.11 Protocol to Achieve a Theoretical Throughput Limit,” *IEEE/ACM Trans. Netw.*, pp. 785–799, December 2000.
- \[6\] J. C. Tay and K. C. Chua, “A Capacity Analysis for the IEEE 802.11 MAC Protocol,” *Wireless Netw.*, pp. 159–171, July 2001.
- \[7\] K. Medepalli and F. A. Tobagi, “Throughput Analysis of IEEE 802.11 Wireless LANs using an Average Cycle Time Approach,” in *Proc. IEEE Globecom ’05*, November 2005.
- \[8\] J. Hui and M. Devetsikiotis, “Metamodeling of Wi-Fi Performance,” in *Proc. IEEE ICC ’06*, June 2006.
- \[9\] Y. Xiao, “An Analysis for Differentiated Services in IEEE 802.11 and IEEE 802.11e Wireless LANs,” in *Proc. IEEE ICDCS ’04*, March 2004.
- \[10\] ——, “Performance Analysis of Priority Schemes for IEEE 802.11 and IEEE 802.11e Wireless LANs,” *IEEE Trans. Wireless Commun.*, pp. 1506–1515, July 2005.
- \[11\] Z. Kong, D. H. K. Tsang, B. Bensaou, and D. Gao, “Performance Analysis of the IEEE 802.11e Contention-Based Channel Access,” *IEEE J. Select. Areas Commun.*, pp. 2095–2106, December 2004.
- \[12\] J. W. Robinson and T. S. Randhawa, “Saturation Throughput Analysis of IEEE 802.11e Enhanced Distributed Coordination Function,” *IEEE J. Select. Areas Commun.*, pp. 917–928, June 2004.
- \[13\] ——, “A Practical Model for Transmission Delay of IEEE 802.11e Enhanced Distributed Channel Access,” in *Proc. IEEE PIMRC ’04*, September 2004.
- \[14\] J. Hui and M. Devetsikiotis, “Performance Analysis of IEEE 802.11e EDCA by a Unified Model,” in *Proc. IEEE Globecom ’04*, December 2004.
- \[15\] ——, “A Unified Model for the Performance Analysis of IEEE 802.11e EDCA,” *IEEE Trans. Commun.*, pp. 1498–1510, September 2005.
- \[16\] H. Zhu and I. Chlamtac, “Performance Analysis for IEEE 802.11e EDCF Service Differentiation,” *IEEE Trans. Wireless Commun.*, pp. 1779–1788, July 2005.
- \[17\] I. Inan, F. Keceli, and E. Ayanoglu, “Saturation Throughput Analysis of the 802.11e Enhanced Distributed Channel Access Function,” to appear in Proc. IEEE ICC ’07.
- \[18\] Z. Tao and S. Panwar, “An Analytical Model for the IEEE 802.11e Enhanced Distributed Coordination Function,” in *Proc. IEEE ICC ’04*, May 2004.
- \[19\] ——, “Throughput and Delay Analysis for the IEEE 802.11e Enhanced Distributed Channel Access,” *IEEE Trans. Commun.*, pp. 596–602, April 2006.
- \[20\] J. Zhao, Z. Guo, Q. Zhang, and W. Zhu, “Performance Study of MAC for Service Differentiation in IEEE 802.11,” in *Proc. IEEE Globecom ’02*, November 2002.
- \[21\] A. Banchs and L. Vollero, “A Delay Model for IEEE 802.11e EDCA,” *IEEE Commun. Lett.*, pp. 508–510, June 2005.
- \[22\] ——, “Throughput Analysis and Optimal Configuration of IEEE 802.11e EDCA,” *Comp. Netw.*, pp. 1749–1768, August 2006.
- \[23\] Y. Chen, Q.-A. Zeng, and D. P. Agrawal, “Performance Analysis of IEEE 802.11e Enhanced Distributed Coordination Function,” in *Proc. IEEE ICON ’03*, September 2003.
- \[24\] Y.-L. Kuo, C.-H. Lu, E. H.-K. Wu, G.-H. Chen, and Y.-H. Tseng, “Performance Analysis of the Enhanced Distributed Coordination Function in the IEEE 802.11e,” in *Proc. IEEE VTC ’03 - Fall*, October 2003.
- \[25\] Y. Lin and V. W. Wong, “Saturation Throughput of IEEE 802.11e EDCA Based on Mean Value Analysis,” in *Proc. IEEE WCNC ’06*, April 2006.
- \[26\] L. Kleinrock, *Queueing Systems*. John Wiley and Sons, 1975.
- \[27\] K. Duffy, D. Malone, and D. J. Leith, “Modeling the 802.11 Distributed Coordination Function in Non-Saturated Conditions,” *IEEE Commun. Lett.*, pp. 715–717, August 2005.
- \[28\] F. Alizadeh-Shabdiz and S. Subramaniam, “Analytical Models for Single-Hop and Multi-Hop Ad Hoc Networks,” in *Proc. ACM Broadnets ’04*, October 2004.
- \[29\] ——, “Analytical Models for Single-Hop and Multi-Hop Ad Hoc Networks,” *Mobile Networks and Applications*, pp. 75–90, February 2006.
- \[30\] G. R. Cantieni, Q. Ni, C. Barakat, and T. Turletti, “Performance Analysis under Finite Load and Improvements for Multirate 802.11,” *Comp. Commun.*, pp. 1095–1109, June 2005.
- \[31\] B. Li and R. Battiti, “Analysis of the IEEE 802.11 DCF with Service Differentiation Support in Non-Saturation Conditions,” in *QoFIS ’04*, September 2004.
- \[32\] P. E. Engelstad and O. N. Osterbo, “Analysis of the Total Delay of IEEE 802.11e EDCA and 802.11 DCF,” in *Proc. IEEE ICC ’06*, June 2006.
- \[33\] A. N. Zaki and M. T. El-Hadidi, “Throughput Analysis of IEEE 802.11 DCF Under Finite Load Traffic,” in *Proc. First International Symposium on Control, Communications and Signal Processing*, 2004.
- \[34\] O. Tickoo and B. Sikdar, “Queueing Analysis and Delay Mitigation in IEEE 802.11 Random Access MAC based Wireless Networks,” in *Proc. IEEE Infocom ’04*, March 2004.
- \[35\] ——, “A Queueing Model for Finite Load IEEE 802.11 Random Access MAC,” in *Proc. IEEE ICC ’04*, June 2004.
- \[36\] X. Chen, H. Zhai, X. Tian, and Y. Fang, “Supporting QoS in IEEE 802.11e Wireless LANs,” *IEEE Trans. Wireless Commun.*, pp. 2217–2227, August 2006.
- \[37\] W. Lee, C. Wang, and K. Sohraby, “On Use of Traditional M/G/1 Model for IEEE 802.11 DCF in Unsaturated Traffic Conditions,” in *Proc. IEEE WCNC ’06*, May 2006.
- \[38\] K. Medepalli and F. A. Tobagi, “System Centric and User Centric Queueing Models for IEEE 802.11 based Wireless LANs,” in *Proc. IEEE Broadnets ’05*, October 2005.
- \[39\] C. H. Foh and M. Zukerman, “A New Technique for Performance Evaluation of Random Access Protocols,” in *Proc. European Wireless ’02*, February 2002.
- \[40\] J. W. Tantra, C. H. Foh, I. Tinnirello, and G. Bianchi, “Analysis of the IEEE 802.11e EDCA Under Statistical Traffic,” in *Proc. IEEE ICC ’06*, June 2006.
- \[41\] S. Mangold, S. Choi, P. May, and G. Hiertz, “IEEE 802.11e - Fair Resource Sharing Between Overlapping Basic Service Sets,” in *Proc. IEEE PIMRC ’02*, September 2002.
- \[42\] T. Suzuki, A. Noguchi, and S. Tasaka, “Effect of TXOP-Bursting and Transmission Error on Application-Level and User-Level QoS in Audio-Video Transmission with 802.11e EDCA,” in *Proc. IEEE PIMRC ’06*, September 2006.
- \[43\] I. Tinnirello and S. Choi, “Efficiency Analysis of Burst Transmissions with Block ACK in Contention-Based 802.11e WLANs,” in *Proc. IEEE ICC ’05*, May 2005.
- \[44\] ——, “Temporal Fairness Provisioning in Multi-Rate Contention-Based 802.11e WLANs,” in *Proc. IEEE WoWMoM ’05*, June 2005.
- \[45\] F. Peng, H. M. Alnuweiri, and V. C. M. Leung, “Analysis of Burst Transmission in IEEE 802.11e Wireless LANs,” in *Proc. IEEE ICC ’06*, June 2006.
- \[46\] (2006) The Network Simulator, ns-2. \[Online\]. Available: http://www.isi.edu/nsnam/ns
- \[47\] IEEE 802.11e HCF MAC model for ns-2.28. \[Online\]. Available: http://newport.eecs.uci.edu/$`\sim`$fkeceli/ns.htm
- \[48\] *IEEE Standard 802.11: Wireless LAN medium access control (MAC) and physical layer (PHY) specifications: Further Higher Data Rate Extension in the 2.4 GHz Band*, IEEE 802.11g Std., 2003.
<table>
<tbody>
<tr>
<td>Refer to caption</td>
</tr>
<tr>
<td>(a)</td>
</tr>
<tr>
<td>Refer to caption</td>
</tr>
<tr>
<td>(b)</td>
</tr>
<tr>
<td>Refer to caption</td>
</tr>
<tr>
<td>(c)</td>
</tr>
</tbody>
</table>
Figure 1: Parts of the proposed DTMC model for $`N_{i}`$=2. The combination of these small chains for all $`j`$, $`k`$, $`l`$ constitutes the proposed DTMC model. (a) $`l = 0`$. (b) $`0 < l < {Q\hspace{0pt}S_{i}}`$. (c) $`l = {Q\hspace{0pt}S_{i}}`$. Remarks: i) the transition probabilities and the states marked with rectangles differ when $`j = {r_{i} - 1}`$ (as in ([6](#S4.E6 "In item 2 ‣ IV EDCA Discrete-Time Markov Chain Model ‣ Analysis of the 802.11e Enhanced Distributed Channel Access Function **footnote *† This work is supported by the Center for Pervasive Communications and Computing, and by National Science Foundation under Grant No. 0434928. Any opinions, findings, and conclusions or recommendations expressed in this material are those of authors and do not necessarily reflect the view of the National Science Foundation.†")) and ([7](#S4.E7 "In item 2 ‣ IV EDCA Discrete-Time Markov Chain Model ‣ Analysis of the 802.11e Enhanced Distributed Channel Access Function **footnote *† This work is supported by the Center for Pervasive Communications and Computing, and by National Science Foundation under Grant No. 0434928. Any opinions, findings, and conclusions or recommendations expressed in this material are those of authors and do not necessarily reflect the view of the National Science Foundation.†"))), ii) the limits for $`l'`$ follow the rules in ([2](#S4.E2 "In item 1 ‣ IV EDCA Discrete-Time Markov Chain Model ‣ Analysis of the 802.11e Enhanced Distributed Channel Access Function **footnote *† This work is supported by the Center for Pervasive Communications and Computing, and by National Science Foundation under Grant No. 0434928. Any opinions, findings, and conclusions or recommendations expressed in this material are those of authors and do not necessarily reflect the view of the National Science Foundation.†"))-([13](#S4.E13 "In item 4 ‣ IV EDCA Discrete-Time Markov Chain Model ‣ Analysis of the 802.11e Enhanced Distributed Channel Access Function **footnote *† This work is supported by the Center for Pervasive Communications and Computing, and by National Science Foundation under Grant No. 0434928. Any opinions, findings, and conclusions or recommendations expressed in this material are those of authors and do not necessarily reflect the view of the National Science Foundation.†")).
[Refer to caption](/html/0704.1833/assets/x4.png)
Figure 2: EDCA backoff after busy medium.
[Refer to caption](/html/0704.1833/assets/x5.png)
Figure 3: Transition through backoff slots in different contention zones for the example given in Fig.[2](#S6.F2 "Figure 2 ‣ Analysis of the 802.11e Enhanced Distributed Channel Access Function **footnote *† This work is supported by the Center for Pervasive Communications and Computing, and by National Science Foundation under Grant No. 0434928. Any opinions, findings, and conclusions or recommendations expressed in this material are those of authors and do not necessarily reflect the view of the National Science Foundation.†").
[Refer to caption](/html/0704.1833/assets/x6.png)
Figure 4: Normalized throughput prediction of the proposed model for 2 AC heterogeneous scenario with respect to increasing load per AC at each station and varying MAC buffer size in basic access mode ($`{T\hspace{0pt}X\hspace{0pt}O\hspace{0pt}P_{3}} = 0`$, $`{T\hspace{0pt}X\hspace{0pt}O\hspace{0pt}P_{1}} = 0`$). Simulation results are also added for comparison.
[Refer to caption](/html/0704.1833/assets/x7.png)
Figure 5: Normalized throughput prediction of the proposed model for 2 AC heterogeneous scenario with respect to increasing load per AC at each station and varying MAC buffer size in basic access mode ($`{T\hspace{0pt}X\hspace{0pt}O\hspace{0pt}P_{3}} = {1504\hspace{0pt}m\hspace{0pt}s}`$, $`{T\hspace{0pt}X\hspace{0pt}O\hspace{0pt}P_{1}} = {3008\hspace{0pt}m\hspace{0pt}s}`$). Simulation results are also added for comparison.
[Refer to caption](/html/0704.1833/assets/x8.png)
Figure 6: Normalized throughput prediction of the proposed model for 2 AC heterogeneous scenario with respect to increasing number of stations when MAC buffer size is 10 packets and total offered load per AC is 2 Mbps ($`{T\hspace{0pt}X\hspace{0pt}O\hspace{0pt}P_{3}} = {1504\hspace{0pt}m\hspace{0pt}s}`$, $`{T\hspace{0pt}X\hspace{0pt}O\hspace{0pt}P_{1}} = {3008\hspace{0pt}m\hspace{0pt}s}`$). Simulation results are also added for comparison.
[Refer to caption](/html/0704.1833/assets/x9.png)
Figure 7: Normalized throughput prediction of the proposed model for 2 AC heterogeneous scenario with respect to increasing number of stations when MAC buffer size is 10 packets ($`{T\hspace{0pt}X\hspace{0pt}O\hspace{0pt}P_{3}} = {1504\hspace{0pt}m\hspace{0pt}s}`$, $`{T\hspace{0pt}X\hspace{0pt}O\hspace{0pt}P_{1}} = {3008\hspace{0pt}m\hspace{0pt}s}`$). Total offered load per AC<sub>3</sub> is 0.5 Mbps while total offered load per AC<sub>3</sub> is 2 Mbps. Simulation results are also added for comparison.
[Refer to caption](/html/0704.1833/assets/x10.png)
Figure 8: Normalized throughput prediction of the proposed model for 2 AC heterogeneous scenario with respect to increasing number of stations when total offered load per AC is 0.5 Mbps ($`{T\hspace{0pt}X\hspace{0pt}O\hspace{0pt}P_{3}} = {1504\hspace{0pt}m\hspace{0pt}s}`$, $`{T\hspace{0pt}X\hspace{0pt}O\hspace{0pt}P_{1}} = {3008\hspace{0pt}m\hspace{0pt}s}`$). Simulation results are also added for the scenario when AC<sub>3</sub> uses On/Off traffic with exponentially distributed idle and active times both with mean 1.5s. $`A\hspace{0pt}C_{1}`$ uses Poisson distribution for packet arrivals.
[Refer to caption](/html/0704.1833/assets/x11.png)
Figure 9: Normalized throughput prediction of the proposed model for 2 AC heterogeneous scenario with respect to increasing number of stations when MAC buffer size is 10 packets and total offered load per AC is 2 Mbps ($`{T\hspace{0pt}X\hspace{0pt}O\hspace{0pt}P_{3}} = {1504\hspace{0pt}m\hspace{0pt}s}`$, $`{T\hspace{0pt}X\hspace{0pt}O\hspace{0pt}P_{1}} = {3008\hspace{0pt}m\hspace{0pt}s}`$). Simulation results are also added for the scenario when both AC<sub>1</sub> and AC<sub>3</sub> uses CBR traffic.
[Refer to caption](/html/0704.1833/assets/x12.png)
Figure 10: Total average delay prediction of the proposed model for 2 AC heterogeneous scenario with respect to increasing load per AC at each station. In the first scenario, TXOP limits are set to 0 ms for both ACs. In the second scenario, TXOP limits are set to $`1.504`$ ms and $`3.008`$ ms for high and low priority ACs respectively. Simulation results are also added for comparison.
[Refer to caption](/html/0704.1833/assets/x13.png)
Figure 11: Average packet loss ratio prediction of the proposed model for 2 AC heterogeneous scenario with respect to increasing load per AC at each station. In the first scenario, TXOP limits are set to 0 ms for both ACs. In the second scenario, TXOP limits are set to $`1.504`$ ms and $`3.008`$ ms for high and low priority ACs respectively. Simulation results are also added for comparison. |
0705.3503 | cs.CR | cyber | # Making Random Choices Invisible to the Scheduler
<sup>†</sup>
<sup>†</sup>thanks: This work has been partially supported by the INRIA DREI Équipe Associée PRINTEMPS and by the INRIA ARC project ProNoBiS.
Konstantinos Chatzikokolakis Catuscia Palamidessi
INRIA and LIX, École Polytechnique, Palaiseau, France {kostas,catuscia}@lix.polytechnique.fr
###### Abstract
When dealing with process calculi and automata which express both nondeterministic and probabilistic behavior, it is customary to introduce the notion of scheduler to solve the nondeterminism. It has been observed that for certain applications, notably those in security, the scheduler needs to be restricted so not to reveal the outcome of the protocol’s random choices, or otherwise the model of adversary would be too strong even for “obviously correct” protocols. We propose a process-algebraic framework in which the control on the scheduler can be specified in syntactic terms, and we show how to apply it to solve the problem mentioned above. We also consider the definition of (probabilistic) may and must preorders, and we show that they are precongruences with respect to the restricted schedulers. Furthermore, we show that all the operators of the language, except replication, distribute over probabilistic summation, which is a useful property for verification.
## 1 Introduction
Security protocols, in particular those for for anonymity and fair exchange, often use randomization to achieve their targets. Since they usually involve more than one agent, they also give rise to concurrent and interactive activities that can be best modeled by nondeterminism. Thus it is convenient to specify them using a formalism which is able to represent both *probabilistic* and *nondeterministic* behavior. Formalisms of this kind have been explored in both Automata Theory \[[1](#bib.bib1), [2](#bib.bib2), [3](#bib.bib3), [4](#bib.bib4), [5](#bib.bib5)\] and in Process Algebra \[[6](#bib.bib6), [7](#bib.bib7), [8](#bib.bib8), [9](#bib.bib9), [10](#bib.bib10), [11](#bib.bib11)\]. See also \[[12](#bib.bib12), [13](#bib.bib13)\] for comparative and more inclusive overviews.
Due to the presence of nondeterminism, in such formalisms it is not possible to define the probability of events in *absolute* terms. We need first to decide how each nondeterministic choice during the execution will be solved. This decision function is called *scheduler*. Once the scheduler is fixed, the behavior of the system (*relatively* to the given scheduler) becomes fully probabilistic and a probability measure can be defined following standard techniques.
It has been observed by several researchers that in security the notion of scheduler needs to be restricted, or otherwise any secret choice of the protocol could be revealed by making the choice of the scheduler depend on it. This issue was for instance one of the main topics of discussion at the panel of CSFW 2006. We illustrate it here with an example on anonymity. We use the standard CCS notation, plus a construct of probabilistic choice $`P +_{p}Q`$ representing a process that evolves into $`P`$ with probability $`p`$ and into $`Q`$ with probability $`1 - p`$.
The following system *Sys* consists of one receiver $`R`$ and two senders $`S,T`$ which communicate via private channels $`a,b`$ respectively. Which of the two senders is successful is decided probabilistically by $`R`$. After reception, $`R`$ sends a signal *ok*.
<table>
<tbody>
<tr>
<td></td>
<td>R=Δa.𝑜𝑘¯.0+0.5b.𝑜𝑘¯.0S=Δa¯.0T=Δb¯.0𝑆𝑦𝑠=𝛥(νa)(νb)(R|S|T)formulae-sequence𝑅Δ𝑎subscript0.5¯𝑜𝑘.0𝑏¯𝑜𝑘.0𝑆Δ¯𝑎.0𝑇Δ¯𝑏.0𝑆𝑦𝑠𝛥𝜈𝑎𝜈𝑏𝑅𝑆𝑇R\,\overset{\scriptscriptstyle\Delta}{=}\,a.\overline{\it ok}.0\;+_{0.5}\;b.\overline{\it ok}.0\;\;\;\;\;\;S\,\overset{\scriptscriptstyle\Delta}{=}\,\bar{a}.0\;\;\;\;\;\;T\,\overset{\scriptscriptstyle\Delta}{=}\,\bar{b}.0\;\;\;\;\;\;\it{Sys}\,\overset{\scriptscriptstyle\Delta}{=}\,(\nu a)(\nu b)(R\;|\;S\;|\;T)</td>
<td></td>
</tr>
</tbody>
</table>
The signal ok is not private, but since it is the same in both cases, in principle an external observer should not be able to infer from it the identity of the sender ($`S`$ or $`T`$). So the system should be anonymous. However, consider a team of two attackers $`A`$ and $`B`$ defined as
<table>
<tbody>
<tr>
<td></td>
<td>A=Δ𝑜𝑘.s¯.0B=Δ𝑜𝑘.t¯.0formulae-sequence𝐴Δ𝑜𝑘¯𝑠.0𝐵Δ𝑜𝑘¯𝑡.0A\,\overset{\scriptscriptstyle\Delta}{=}\,{\it ok}.\bar{s}.0\;\;\;\;\;\;B\,\overset{\scriptscriptstyle\Delta}{=}\,{\it ok}.\bar{t}.0</td>
<td></td>
</tr>
</tbody>
</table>
and consider the parallel composition $`{Sys}\hspace{0pt}{|A|}\hspace{0pt}B`$. We have that, under certain schedulers, the system is no longer anonymous. More precisely, a scheduler could leak the identity of the sender via the channels $`s,t`$ by forcing $`R`$ to synchronize with $`A`$ on ok if $`R`$ has chosen the first alternative, and with $`B`$ otherwise. This is because in general a scheduler can see the whole history of the computation, in particular the random choices, even those which are supposed to be private. Note that the visibility of the synchronization channels to the scheduler is not crucial for this example: we would have the same problem, for instance, if $`S`$, $`T`$ were both defined as $`\overline{a}\hspace{0pt}.0`$, $`R`$ as $`a.{\overline{ok}\hspace{0pt}.0}`$, and $`Sys`$ as $`{({\nu\hspace{0pt}a})}\hspace{0pt}{(\left. {({R +_{0.5}S})} \middle| T \right.)}`$.
The above example demonstrates that, with the standard definition of scheduler, it is not possible to represent a truly private random choice (or a truly private nondeterministic choice, for the matter) with the current probabilistic process calculi. This is a clear shortcoming when we want to use these formalisms for the specification and verification of security protocols.
There is another issue related to verification: a private choice has certain algebraic properties that would be useful in proving equivalences between processes. In fact, if the outcome of a choice remains private, then it should not matter at which point of the execution the process makes such choice, until it actually uses it. Consider for instance $`A`$ and $`B`$ defined as follows
<table>
<tbody>
<tr>
<td>A=Δa(x).([x=0]𝑜𝑘¯+0.5[x=1]𝑜𝑘¯)\begin{array}[]{r@{}l}A\overset{\scriptscriptstyle\Delta}{=}a(x).(&[x=0]\overline{\it ok}\\ &+_{0.5}\\ &[x=1]\overline{\it ok})\end{array}</td>
<td>B=Δa(x).[x=0]𝑜𝑘¯+0.5a(x).[x=1]𝑜𝑘¯𝐵Δformulae-sequence𝑎𝑥delimited-[]𝑥0¯𝑜𝑘missing-subexpressionsubscript0.5missing-subexpressionformulae-sequence𝑎𝑥delimited-[]𝑥1¯𝑜𝑘\begin{array}[]{r@{}l}B\overset{\scriptscriptstyle\Delta}{=}&a(x).[x=0]\overline{\it ok}\\ &+_{0.5}\\ &a(x).[x=1]\overline{\it ok}\\ \end{array}</td>
</tr>
</tbody>
</table>
Process $`A`$ receives a value and then decides randomly whether it will accept the value $`0`$ or $`1`$. Process $`B`$ does exactly the same thing except that the choice is performed before the reception of the value. If the random choices in $`A`$ and $`B`$ are private, intuitively we should have that $`A`$ and $`B`$ are equivalent ($`A \approx B`$). This is because it should not matter whether the choice is done before or after receiving a message, as long as the outcome of the choice is completely invisible to any other process or observer. However, consider the parallel context $`C = \left. {\overline{a}\hspace{0pt}0} \middle| {\overline{a}\hspace{0pt}1} \right.`$. Under any scheduler $`A`$ has probability at most $`1/2`$ to perform $`\overline{ok}`$. With $`B`$, on the other hand, the scheduler can choose between $`\overline{a}\hspace{0pt}0`$ and $`\overline{a}\hspace{0pt}1`$ based on the outcome of the probabilistic choice, thus making the maximum probability of $`\overline{ok}`$ equal to $`1`$. The execution trees of $`\left. A \middle| C \right.`$ and $`\left. B \middle| C \right.`$ are shown in Figure [1](#S1.F1 "Figure 1 ‣ 1 Introduction ‣ Making Random Choices Invisible to the SchedulerThis work has been partially supported by the INRIA DREI Équipe Associée PRINTEMPS and by the INRIA ARC project ProNoBiS.").
\pstree
\[treesep=2pt,levelsep=110pt,nodesep=5pt,treemode=R,edge=\nccurve,angleB=180,ncurvA=0,ncurvB=0\] \Tr$`A\hspace{0pt}{|{\overline{a}\hspace{0pt}0}|}\hspace{0pt}\overline{a}\hspace{0pt}1`$ \pstree\[levelsep=95pt\] \Tr$`\left. {({{{\lbrack{0 = 0}\rbrack}\hspace{0pt}\overline{o\hspace{0pt}k}} +_{0.5}{{\lbrack{0 = 1}\rbrack}\hspace{0pt}\overline{o\hspace{0pt}k}}})} \middle| {\overline{a}\hspace{0pt}1} \right.`$ \Tr$`\overline{o\hspace{0pt}k}`$ \ncput\rnodea1 \Tr$`0`$ \ncput\rnodeb1 \pstree\[levelsep=95pt\] \Tr$`\left. {({{{\lbrack{1 = 0}\rbrack}\hspace{0pt}\overline{o\hspace{0pt}k}} +_{0.5}{{\lbrack{1 = 1}\rbrack}\hspace{0pt}\overline{o\hspace{0pt}k}}})} \middle| {\overline{a}\hspace{0pt}0} \right.`$ \Tr$`0`$ \ncput\rnodea2 \Tr$`\overline{o\hspace{0pt}k}`$ \ncput\rnodeb2 \ncarc\[arcangle=30\]a1b1 \ncarc\[arcangle=30\]a2b2
\pstree\[treesep=2pt,levelsep=100pt,nodesep=5pt,treemode=R,edge=\nccurve,angleB=180,ncurvA=0,ncurvB=0\] \Tr$`B\hspace{0pt}{|{\overline{a}\hspace{0pt}0}|}\hspace{0pt}\overline{a}\hspace{0pt}1`$ \pstree\[levelsep=80pt\] \Tr$`{a\hspace{0pt}{(x)}}.{{\lbrack{x = 0}\rbrack}\hspace{0pt}\overline{o\hspace{0pt}k}\hspace{0pt}{|{\overline{a}\hspace{0pt}0}|}\hspace{0pt}\overline{a}\hspace{0pt}1}`$ \ncput\rnodea3 \Tr$`\overline{o\hspace{0pt}k}`$ \Tr$`0`$ \pstree\[levelsep=80pt\] \Tr$`{a\hspace{0pt}{(x)}}.{{\lbrack{x = 1}\rbrack}\hspace{0pt}\overline{o\hspace{0pt}k}\hspace{0pt}{|{\overline{a}\hspace{0pt}0}|}\hspace{0pt}\overline{a}\hspace{0pt}1}`$ \ncput\rnodeb3 \Tr$`0`$ \Tr$`\overline{o\hspace{0pt}k}`$ \ncarc\[arcangle=30\]a3b3
Figure 1: Execution trees for $`\left. A \middle| C \right.`$ and $`\left. B \middle| C \right.`$
In general when $`+_{p}`$ represents a private choice we would like to have
<table>
<tbody>
<tr>
<td></td>
<td>C[P+pQ]≈C[τ.P]+pC[τ.Q]C[P+_{p}Q]\approx C[\tau.P]+_{p}C[\tau.Q]</td>
<td></td>
<td>(1)</td>
</tr>
</tbody>
</table>
for all processes $`P,Q`$ and all contexts $`C`$ *not containing replication (or recursion)*. In the case of replication the above cannot hold since $`!{(P +_{p}Q)}`$ makes available each time the choice between $`P`$ and $`Q`$, while $`{(!\tau.P)} +_{p}{(!\tau.Q)}`$ chooses once and for all which of the two ($`P`$ or $`Q`$) should be replicated. Similarly for recursion. The reason why we need a $`\tau`$ is explained in Section [5](#S5 "5 Testing relations for CCSσ processes ‣ Making Random Choices Invisible to the SchedulerThis work has been partially supported by the INRIA DREI Équipe Associée PRINTEMPS and by the INRIA ARC project ProNoBiS.").
The algebraic property ([1](#S1.E1 "In 1 Introduction ‣ Making Random Choices Invisible to the SchedulerThis work has been partially supported by the INRIA DREI Équipe Associée PRINTEMPS and by the INRIA ARC project ProNoBiS.")) expresses in an abstract way the privacy of the probabilistic choice. Moreover, this property is also useful for the verification of security properties. The interested reader can find in \[[14](#bib.bib14)\] an example of application to a fair exchange protocol. In principle ([1](#S1.E1 "In 1 Introduction ‣ Making Random Choices Invisible to the SchedulerThis work has been partially supported by the INRIA DREI Équipe Associée PRINTEMPS and by the INRIA ARC project ProNoBiS.")) should be useful for any kind of verification in the process algebra style.
We propose a process-algebraic approach to the problem of hiding the outcome of random choices. Our framework is based on a calculus obtained by adding to CCS an internal probabilistic choice construct
<sup>1</sup>
<sup>1</sup>1We actually consider a variant of CCS where recursion is replaced by replication. The two languages are not equivalent, but we believe that the issues regarding the differences between replication and recursion are orthogonal to the topics investigated in this paper.
. This calculus, to which we refer as CCS<sub>p</sub>, is a variant of the one studied in \[[11](#bib.bib11)\], the main differences being that we use replication instead than recursion, and we lift some restrictions that were imposed in \[[11](#bib.bib11)\] to obtain a complete axiomatization. The semantics of CCS<sub>p</sub> is given in terms of Segala’s *simple probabilistic automata* \[[4](#bib.bib4), [7](#bib.bib7)\].
In order to limit the power of the scheduler, we extend CCS<sub>p</sub> with terms representing explicitly the notion of scheduler. The latter interact with the original processes via a labeling system. This will allow to specify at the syntactic level (by a suitable labeling) which choices should be visible to schedulers, and which ones should not.
### 1.1 Contribution
The main contributions of this paper are:
- •
A process calculus CCS<sub>σ</sub> in which the scheduler is represented as a process, and whose power can therefore be controlled at the syntactic level.
- •
An application of CCS<sub>σ</sub> to an extended anonymity example (the Dining Cryptographers Protocol, DCP). We also briefly outline how to extend CCS<sub>σ</sub> so to allow the definition of private nondeterministic choice, and we apply it to the DCP with nondeterministic master. To our knowledge this is the first formal treatment of the scheduling problem in DCP and the first formalization of a nondeterministic master for the (probabilistic) DCP.
- •
The adaptation of the standard notions of probabilistic testing preorders to CCS<sub>σ</sub>, and the “sanity check” that they are still precongruences with respect to all the operators except the nondeterministic sum. For the latter we have the problem that $`P`$ and $`\tau.P`$ are must equivalent, but $`Q + P`$ and $`{Q + \tau}.P`$ are not. This is typical for the CCS $`+`$: usually it does not preserve weak equivalences.
- •
The proof that, under suitable conditions on the labelings of $`C`$, $`\tau.P`$ and $`\tau.Q`$, CCS<sub>σ</sub> satisfies the property expressed by ([1](#S1.E1 "In 1 Introduction ‣ Making Random Choices Invisible to the SchedulerThis work has been partially supported by the INRIA DREI Équipe Associée PRINTEMPS and by the INRIA ARC project ProNoBiS.")), where $`\approx`$ is probabilistic testing equivalence.
### 1.2 Related work
The works that are most closely related to ours are \[[17](#bib.bib17), [18](#bib.bib18)\]. In those paper the authors consider probabilistic automata and introduce a restriction on the scheduler to the purpose of making them suitable to applications in security protocols. Their approach is based on dividing the actions of each component of the system in equivalence classes (*tasks*). The order of execution of different tasks is decided in advance by a so-called *task scheduler*. The remaining nondeterminism within a task is solved by a second scheduler, which models the standard *adversarial scheduler* of the cryptographic community. This second entity has limited knowledge about the other components: it sees only the information that they communicate during execution.
In contrast to the above approach, our definition of scheduler is based on a labeling system, and the same action can receive different labels during the execution, so our “equivalence classes” (schedulable actions with the same label) can change dynamically. However we don’t know at the moment whether this difference determines a separation in the expressive power. The main difference, anyway, is that our framework is process-algebraic and we focus on testing preorders, their congruence properties, and the conditions under which certain equivalences hold.
Another work along these lines is \[[15](#bib.bib15)\], which uses partitions on the state-space to obtain partial-information schedulers. However in that paper the authors consider a synchronous parallel composition, so the setting is rather different.
### 1.3 Plan of the paper
In the next section we briefly recall some basic notions. In Section [3](#S3 "3 A variant of CCS with explicit scheduler ‣ Making Random Choices Invisible to the SchedulerThis work has been partially supported by the INRIA DREI Équipe Associée PRINTEMPS and by the INRIA ARC project ProNoBiS.") we define a preliminary version of the language CCS<sub>σ</sub> and of the corresponding notion of scheduler. In Section [4](#S4 "4 Expressiveness of the syntactic scheduler ‣ Making Random Choices Invisible to the SchedulerThis work has been partially supported by the INRIA DREI Équipe Associée PRINTEMPS and by the INRIA ARC project ProNoBiS.") we compare our notion of scheduler with the more standard “semantic” notion, and we improve the definition of CCS<sub>σ</sub> so to retrieve the full expressive power of the semantic schedulers. In Section [5](#S5 "5 Testing relations for CCSσ processes ‣ Making Random Choices Invisible to the SchedulerThis work has been partially supported by the INRIA DREI Équipe Associée PRINTEMPS and by the INRIA ARC project ProNoBiS.") we study the probabilistic testing preorders, their compositionality properties, and the conditions under which ([1](#S1.E1 "In 1 Introduction ‣ Making Random Choices Invisible to the SchedulerThis work has been partially supported by the INRIA DREI Équipe Associée PRINTEMPS and by the INRIA ARC project ProNoBiS.")) holds. Section [6](#S6 "6 An application to security ‣ Making Random Choices Invisible to the SchedulerThis work has been partially supported by the INRIA DREI Équipe Associée PRINTEMPS and by the INRIA ARC project ProNoBiS.") presents an application to security. Section [7](#S7 "7 Conclusion and Future work ‣ Making Random Choices Invisible to the SchedulerThis work has been partially supported by the INRIA DREI Équipe Associée PRINTEMPS and by the INRIA ARC project ProNoBiS.") concludes.
## 2 Preliminaries
In this section we briefly recall some preliminary notions about the simple probabilistic automata and CCS<sub>p</sub>.
### 2.1 Simple probabilistic automata \[[4](#bib.bib4), [7](#bib.bib7)\]
A *discrete probability measure* over a set $`X`$ is a function $`\mu:{2^{X}\mapsto{\lbrack 0,1\rbrack}}`$ such that $`{\mu\hspace{0pt}{(X)}} = 1`$ and $`{\mu\hspace{0pt}{({\cup_{i}X_{i}})}} = {\sum_{i}{\mu\hspace{0pt}{(X_{i})}}}`$ where $`X_{i}`$ is a countable family of pairwise disjoint subsets of $`X`$. The set of all discrete probability measures over $`X`$ will be denoted by $`D\hspace{0pt}i\hspace{0pt}s\hspace{0pt}c\hspace{0pt}{(X)}`$. We will denote by $`{{\delta\hspace{0pt}{(x)}},x} \in X`$ (called the *Dirac measure* on $`x`$) the probability measure that assigns probability $`1`$ to $`\{ x\}`$. We will also denote by $`\sum_{i}{{\lbrack p_{i}\rbrack}\hspace{0pt}\mu_{i}}`$ the probability measure obtained as a convex sum of the measures $`\mu_{i}`$.
A *simple probabilistic automaton*
<sup>2</sup>
<sup>2</sup>2For simplicity in the following we will refer to a simple probabilistic automaton as *probabilistic automaton*. Note however that simple probabilistic automata are a subset of the probabilistic automata defined in \[[4](#bib.bib4), [5](#bib.bib5)\].
is a tuple $`(S,q,A,\mathcal{D})`$ where $`S`$ is a set of states, $`q \in S`$ is the *initial state*, $`A`$ is a set of actions and $`\mathcal{D} \subseteq {{S \times A \times D}\hspace{0pt}i\hspace{0pt}s\hspace{0pt}c\hspace{0pt}{(S)}}`$ is a *transition relation*. Intuitively, if $`{(s,a,\mu)} \in \mathcal{D}`$ then there is a transition from the state $`s`$ performing the action $`a`$ and leading to a distribution $`\mu`$ over the states of the automaton. The idea is that the choice of transition among the available ones in $`\mathcal{D}`$ is performed nondeterministically, and the choice of the target state among the ones allowed by $`\mu`$ (i.e. those states $`q`$ such that $`{\mu\hspace{0pt}{(q)}} > 0`$) is performed probabilistically.
A probabilistic automaton $`M`$ is *fully probabilistic* if from each state of $`M`$ there is at most one transition available. An execution $`\alpha`$ of a probabilistic automaton is a (possibly infinite) sequence $`s_{0}\hspace{0pt}a_{1}\hspace{0pt}s_{1}\hspace{0pt}a_{2}\hspace{0pt}s_{2}\hspace{0pt}\ldots`$ of alternating states and actions, such that $`q = s_{0}`$, and for each $`i`$ $`{(s_{i},a_{i + 1},\mu_{i})} \in \mathcal{D}`$ and $`{\mu_{i}\hspace{0pt}{(s_{i + 1})}} > 0`$ hold. We will use $`l\hspace{0pt}s\hspace{0pt}t\hspace{0pt}a\hspace{0pt}t\hspace{0pt}e\hspace{0pt}{(\alpha)}`$ to denote the last state of a finite execution $`\alpha`$, and $`e\hspace{0pt}x\hspace{0pt}e\hspace{0pt}c^{\ast}\hspace{0pt}{(M)}`$ and $`e\hspace{0pt}x\hspace{0pt}e\hspace{0pt}c\hspace{0pt}{(M)}`$ to represent the set of all the finite and of all the executions of $`M`$, respectively.
A *scheduler* of a probabilistic automaton $`M = {(S,q,A,\mathcal{D})}`$ is a function
<table>
<tbody>
<tr>
<td></td>
<td>ζ:exec∗(M)↦𝒟:𝜁maps-to𝑒𝑥𝑒superscript𝑐𝑀𝒟\zeta:exec^{*}(M)\mapsto\mathcal{D}</td>
<td></td>
</tr>
</tbody>
</table>
such that $`{\zeta\hspace{0pt}{(\alpha)}} = {(s,a,\mu)} \in \mathcal{D}`$ implies that $`s = {l\hspace{0pt}s\hspace{0pt}t\hspace{0pt}a\hspace{0pt}t\hspace{0pt}e\hspace{0pt}{(\alpha)}}`$. The idea is that a scheduler selects a transition among the ones available in $`\mathcal{D}`$ and it can base his decision on the history of the execution. The *execution tree* of $`M`$ relative to the scheduler $`\zeta`$, denoted by $`e\hspace{0pt}t\hspace{0pt}r\hspace{0pt}e\hspace{0pt}e\hspace{0pt}{(M,\zeta)}`$, is a fully probabilistic automaton $`M' = {(S',q',A',\mathcal{D}')}`$ such that $`S' \subseteq {e\hspace{0pt}x\hspace{0pt}e\hspace{0pt}c\hspace{0pt}{(M)}}`$, $`q' = q`$, $`A' = A`$, and $`{(\alpha,a,\mu')} \in \mathcal{D}'`$ if and only if $`{\zeta\hspace{0pt}{(\alpha)}} = {({l\hspace{0pt}s\hspace{0pt}t\hspace{0pt}a\hspace{0pt}t\hspace{0pt}e\hspace{0pt}{(\alpha)}},a,\mu)}`$ for some $`\mu`$ and $`{\mu'\hspace{0pt}{({\alpha\hspace{0pt}a\hspace{0pt}s})}} = {\mu\hspace{0pt}{(s)}}`$. Intuitively, $`e\hspace{0pt}t\hspace{0pt}r\hspace{0pt}e\hspace{0pt}e\hspace{0pt}{(M,\zeta)}`$ is produced by unfolding the executions of $`M`$ and resolving all deterministic choices using $`\zeta`$. Note that $`e\hspace{0pt}t\hspace{0pt}r\hspace{0pt}e\hspace{0pt}e\hspace{0pt}{(M,\zeta)}`$ is a simple
<sup>3</sup>
<sup>3</sup>3This is true because we do not consider probabilistic schedulers. If we considered such schedulers then the execution tree would no longer be a simple automaton.
and fully probabilistic automaton.
### 2.2 CCS with internal probabilistic choice
Let $`a`$ range over a countable set of *channel names*. The syntax of CCS<sub>p</sub> is the following:
<table>
<tbody>
<tr>
<td></td>
<td>α::=a|a¯|τprefixesP,Q::=processes α.P prefix |P|Q parallel |P+Q nondeterministic choice |∑ipiPi internal probabilistic choice |(νa)P restriction |!P replication |0 nil\begin{array}[t]{@{\textrm{\hspace{20pt}}}l@{}l@{\textrm{\hspace{20pt}}}l}\lx@intercol\alpha::=a\ \,|\ \,\bar{a}\ \,|\ \,\tau\hfil\lx@intercol&\textrm{{prefixes}}\\[2.0pt] \lx@intercol P,Q::=\hfil\lx@intercol&\textrm{{processes}}\\[2.0pt] \hskip 20.0pt}\lx@intercol&\alpha.P\hfil\textrm{\hskip 20.&\textrm{prefix}\\[2.0pt] \hskip 20.0pt}\lx@intercol\ \,|&P\;|\;Q\hfil\textrm{\hskip 20.&\textrm{parallel}\\[2.0pt] \hskip 20.0pt}\lx@intercol\ \,|&P+Q\hfil\textrm{\hskip 20.&\textrm{nondeterministic choice}\\[2.0pt] \hskip 20.0pt}\lx@intercol\ \,|&\textstyle{\sum_{i}\>}p_{i}P_{i}\hfil\textrm{\hskip 20.&\textrm{internal probabilistic choice}\\[2.0pt] \hskip 20.0pt}\lx@intercol\ \,|&(\nu a)P\hfil\textrm{\hskip 20.&\textrm{restriction}\\[2.0pt] \hskip 20.0pt}\lx@intercol\ \,|&!P\hfil\textrm{\hskip 20.&\textrm{replication}\\[2.0pt] \hskip 20.0pt}\lx@intercol\ \,|&0\hfil\textrm{\hskip 20.&\textrm{nil}\end{array}</td>
<td></td>
</tr>
</tbody>
</table>
$`\begin{array}{llll}
\text{ACT} & {\frac{}{\alpha.{P\hspace{0pt}\overset{\alpha}{\longrightarrow}\hspace{0pt}\delta\hspace{0pt}(P)}}\hspace{0pt}\ } & \text{RES} & \frac{{{{P\hspace{0pt}\overset{\alpha}{\longrightarrow}\hspace{0pt}\mu}\qquad\alpha} \neq a},\overline{a}}{\left( {\nu\hspace{0pt}a} \right)\hspace{0pt}P\hspace{0pt}\overset{\alpha}{\longrightarrow}\hspace{0pt}\left( {\nu\hspace{0pt}a} \right)\hspace{0pt}\mu} \\
\text{SUM1} & {\frac{P\hspace{0pt}\overset{\alpha}{\longrightarrow}\hspace{0pt}\mu}{P + {Q\hspace{0pt}\overset{\alpha}{\longrightarrow}\hspace{0pt}\mu}}\hspace{0pt}\ } & \text{PAR1} & \frac{P\hspace{0pt}\overset{\alpha}{\longrightarrow}\hspace{0pt}\mu}{P\hspace{0pt}\left| {Q\hspace{0pt}\overset{\alpha}{\longrightarrow}\hspace{0pt}\mu} \right|\hspace{0pt}Q} \\
\text{COM} & {\frac{{P\hspace{0pt}\overset{a}{\longrightarrow}\hspace{0pt}\delta\hspace{0pt}\left( P' \right)}\quad{Q\hspace{0pt}\overset{\overline{a}}{\longrightarrow}\hspace{0pt}\delta\hspace{0pt}\left( Q' \right)}}{\left. P \middle| Q\overset{\tau}{\longrightarrow}\delta\left( P' \middle| Q' \right) \right.}\hspace{0pt}\ } & \text{PROB} & \frac{}{\left. \sum{}_{i} \right.{p_{i}\hspace{0pt}P_{i}\hspace{0pt}\overset{\tau}{\longrightarrow}\hspace{0pt}{\left. \sum{}_{i} \right.{\left\lbrack p_{i} \right\rbrack\hspace{0pt}\delta\hspace{0pt}\left( P_{i} \right)}}}} \\
\text{BANG1} & {\frac{P\hspace{0pt}\overset{\alpha}{\longrightarrow}\hspace{0pt}\mu}{\left. !P\overset{\alpha}{\longrightarrow}\mu \middle| !P \right.}\hspace{0pt}\ } & \text{BANG2} & \frac{{P\hspace{0pt}\overset{a}{\longrightarrow}\hspace{0pt}\delta\hspace{0pt}\left( P_{1} \right)}\quad{P\hspace{0pt}\overset{\overline{a}}{\longrightarrow}\hspace{0pt}\delta\hspace{0pt}\left( P_{2} \right)}}{!P\overset{\tau}{\longrightarrow}\delta\left( P_{1} \middle| P_{2} \middle| !P \right)}
\end{array}`$
Figure 2: The semantics of CCS<sub>p</sub>. SUM1 and PAR1 have corresponding right rules SUM2 and PAR2, omitted for simplicity.
We will also use the notation $`P_{1} +_{p}P_{2}`$ to represent a binary sum $`\sum_{i}{p_{i}\hspace{0pt}P_{i}}`$ with $`p_{1} = p`$ and $`p_{2} = {1 - p}`$.
The semantics of a CCS<sub>p</sub> term is a probabilistic automaton defined inductively on the basis of the syntax according to the rules in Figure [2](#S2.F2 "Figure 2 ‣ 2.2 CCS with internal probabilistic choice ‣ 2 Preliminaries ‣ Making Random Choices Invisible to the SchedulerThis work has been partially supported by the INRIA DREI Équipe Associée PRINTEMPS and by the INRIA ARC project ProNoBiS."). We write $`s\hspace{0pt}\overset{a}{\longrightarrow}\hspace{0pt}\mu`$ when $`(s,a,\mu)`$ is a transition of the probabilistic automaton. We also denote by $`\left. \mu \middle| Q \right.`$ the measure $`\mu'`$ such that $`{\mu'\hspace{0pt}{(\left. P \middle| Q \right.)}} = {\mu\hspace{0pt}{(P)}}`$ for all processes $`P`$ and $`{\mu'\hspace{0pt}{(R)}} = 0`$ if $`R`$ is not of the form $`\left. P \middle| Q \right.`$. Similarly $`{{({\nu\hspace{0pt}a})}\hspace{0pt}\mu} = \mu'`$ such that $`{\mu'\hspace{0pt}{({{({\nu\hspace{0pt}a})}\hspace{0pt}P})}} = {\mu\hspace{0pt}{(P)}}`$.
A transition of the form $`P\hspace{0pt}\overset{a}{\longrightarrow}\hspace{0pt}\delta\hspace{0pt}{(P')}`$, i.e. a transition having for target a Dirac measure, corresponds to a transition of a non-probabilistic automaton (a standard labeled transition system). Thus, all the rules of CCS<sub>p</sub> imitate the ones of CCS except from PROB. The latter models the internal probabilistic choice: a silent $`\tau`$ transition is available from the sum to a measure containing all of its operands, with the corresponding probabilities.
Note that in the produced probabilistic automaton, all transitions to non-Dirac measures are silent. This is similar to the *alternating model* \[[2](#bib.bib2)\], however our case is more general because the silent and non-silent transitions are not necessarily alternated. On the other hand, with respect to to the simple probabilistic automata the fact that the probabilistic transitions are silent looks as a restriction. However, it has been proved by Bandini and Segala \[[7](#bib.bib7)\] that the simple probabilistic automata and the alternating model are essentially equivalent, so, being in the middle, our model is equivalent as well.
## 3 A variant of CCS with explicit scheduler
In this section we present a variant of CCS in which the scheduler is explicit, in the sense that it has a specific syntax and its behavior is defined by the operational semantics of the calculus. We will refer to this calculus as CCS<sub>σ</sub>. Processes in CCS<sub>σ</sub> contain labels that allow us to refer to a particular sub-process. A scheduler also behaves like a process, using however a different and much simpler syntax, and its purpose is to guide the execution of the main process using the labels that the latter provides. A *complete process* is a process running in parallel with a scheduler, and we will formally describe their interaction by defining an operational semantics for complete processes.
We will present CCS<sub>σ</sub> in an incremental way. First we define the basic calculus CCS<sub>σ</sub> which is the same as CCS<sub>p</sub> with the addition of the explicit scheduler. Then we will perform an extensions of this basic calculus by adding choice to the scheduler so to achieve its full expressive power. Finally, in Section [6](#S6 "6 An application to security ‣ Making Random Choices Invisible to the SchedulerThis work has been partially supported by the INRIA DREI Équipe Associée PRINTEMPS and by the INRIA ARC project ProNoBiS."), we outline an extension of CCS<sub>σ</sub> with a second *independent* scheduler, to the purpose of making private certain nondeterministic choices.
### 3.1 Syntax
$`\begin{matrix}
{\begin{array}{lll}
{I::=0I|\, 1I|\epsilon} & \text{label indexes} & \\
{L::=l^{I}} & \text{labels} & \\
{{P,Q}::=} & \text{processes} & \\
\ & {L:\alpha.P\ } & \text{prefix} \\
{\ |} & \left. P \middle| {Q\hspace{0pt}\ } \right. & \text{parallel} \\
{\ |} & {P + {Q\hspace{0pt}\ }} & \text{nondeterministic choice} \\
{\ |} & {l:{\sum_{i}{p_{i}\hspace{0pt}P_{i}\hspace{0pt}\ }}} & \text{internal prob. choice} \\
{\ |} & {{({\nu\hspace{0pt}a})}\hspace{0pt}P\hspace{0pt}\ } & \text{restriction} \\
{\ |} & {!P\ } & \text{replication} \\
{\ |} & {0\hspace{0pt}\ } & \text{nil}
\end{array}\mspace{39mu}} & \begin{array}{lll}
{{S,T}::=} & \text{scheduler} & \\
\ & {{\sigma\hspace{0pt}{(L)}}.{S\hspace{0pt}\ }} & \text{schedule single action} \\
{\ |} & {{\sigma\hspace{0pt}{(L,L)}}.{S\hspace{0pt}\ }} & \text{synchronization} \\
{\ |} & {0\hspace{0pt}\ } & \text{nil} \\
{CP::=P \parallel S} & \text{complete process} &
\end{array}
\end{matrix}`$
Figure 3: The syntax of the core CCS<sub>σ</sub>
Let $`a`$ range over a countable set of *channel names* and $`l`$ over a countable set of *atomic labels*. The syntax of CCS<sub>σ</sub>, shown in Figure [3](#S3.F3 "Figure 3 ‣ 3.1 Syntax ‣ 3 A variant of CCS with explicit scheduler ‣ Making Random Choices Invisible to the SchedulerThis work has been partially supported by the INRIA DREI Équipe Associée PRINTEMPS and by the INRIA ARC project ProNoBiS."), is the same as the one of CCS<sub>p</sub> except for the presence of labels. These are used to select the subprocess which “performs” a transition. Since only the operators with an initial rule can originate a transition, we only need to assign labels to the prefix and to the probabilistic sum. We use labels of the form $`l^{s}`$ where $`l`$ is an atomic label and the index $`s`$ is a finite string of $`0`$ and $`1`$, possibly empty
<sup>4</sup>
<sup>4</sup>4For simplicity we will write $`l`$ for $`l^{\epsilon}`$.
. Indexes are used to avoid multiple copies of the same label in case of replication, which occurs dynamically due to the the bang operator. As explained in the semantics, each time a process is replicated we relabel it using appropriate indexes.
A scheduler selects a sub-process for execution on the basis of its label, so we use $`{\sigma\hspace{0pt}{(l)}}.S`$ to represent a scheduler that selects the process with label $`l`$ and continues as $`S`$. In the case of synchronization we need to select two processes simultaneously, hence we need a scheduler the form $`{\sigma\hspace{0pt}{(l_{1},l_{2})}}.S`$. A complete process is a process put in parallel with a scheduler, for example $`l_{1}:a.l_{2}:b \parallel \sigma{(l_{1})}.\sigma{(l_{2})}`$. Note that for processes with an infinite execution path we need schedulers of infinite length.
### 3.2 Semantics
$`\begin{array}{llll}
\text{ACT} & {\frac{}{l:\alpha.P \parallel \sigma(l).S\overset{\alpha}{\longrightarrow}\delta(P \parallel S)}\hspace{0pt}\ } & \text{RES} & \frac{{P \parallel {{S\hspace{0pt}\overset{\alpha}{\longrightarrow}\hspace{0pt}\mu}\quad\alpha}} \neq {a,\overline{a}}}{{\left( {\nu\hspace{0pt}a} \right)\hspace{0pt}P} \parallel {S\hspace{0pt}\overset{\alpha}{\longrightarrow}\hspace{0pt}\left( {\nu\hspace{0pt}a} \right)\hspace{0pt}\mu}} \\
\text{SUM1} & {\frac{P \parallel {S\hspace{0pt}\overset{\alpha}{\longrightarrow}\hspace{0pt}\mu}}{P + {Q \parallel {S\hspace{0pt}\overset{\alpha}{\longrightarrow}\hspace{0pt}\mu}}}\hspace{0pt}\ } & \text{PAR1} & \frac{P \parallel {S\hspace{0pt}\overset{\alpha}{\longrightarrow}\hspace{0pt}\mu}}{\left. {P\hspace{0pt}\left| Q \right\|\hspace{0pt}S\hspace{0pt}\overset{\alpha}{\longrightarrow}\hspace{0pt}\mu} \middle| Q \right.} \\
\text{COM} & \frac{P \parallel \sigma\left( l_{1} \right)\overset{a}{\longrightarrow}\delta\left( P' \parallel 0 \right)\qquad Q \parallel \sigma\left( l_{2} \right)\overset{\overline{a}}{\longrightarrow}\delta\left( Q' \parallel 0 \right)}{{P\hspace{0pt}\left| Q \right\|\hspace{0pt}\sigma\hspace{0pt}\left( l_{1},l_{2} \right)}.{S\hspace{0pt}\overset{\tau}{\longrightarrow}\hspace{0pt}\delta\hspace{0pt}\left( {P'\hspace{0pt}\left| Q' \right\|\hspace{0pt}S} \right)}} & & \\
\text{BANG1} & {\frac{P \parallel {S\hspace{0pt}\overset{\alpha}{\longrightarrow}\hspace{0pt}\mu}}{\left. !P \parallel S\overset{\alpha}{\longrightarrow}\rho_{0}(\mu) \middle| \rho_{1}(!P) \right.}\hspace{0pt}\ } & \text{PROB} & \frac{}{l:\left. \sum{}_{i} \right.p_{i}P_{i} \parallel \sigma(l).S\overset{\tau}{\longrightarrow}\left. \sum{}_{i} \right.\left\lbrack p_{i} \right\rbrack\delta\left( P_{i} \parallel S \right)} \\
\text{BANG2} & \frac{P \parallel \sigma\left( l_{1} \right)\overset{a}{\longrightarrow}\delta\left( P_{1} \parallel 0 \right)\qquad P \parallel \sigma\left( l_{2} \right)\overset{\overline{a}}{\longrightarrow}\delta\left( P_{2} \parallel 0 \right)}{!P \parallel \sigma\left( l_{1},l_{2} \right).S\overset{\tau}{\longrightarrow}\delta\left( \rho_{0}\left( P_{1} \right) \middle| \rho_{10}\left( P_{2} \right) \middle| \rho_{11}(!P) \parallel S \right)} & &
\end{array}`$
Figure 4: The semantics of CCS<sub>σ</sub>. SUM1 and PAR1 have corresponding right rules SUM2 and PAR2, omitted for simplicity.
The operational semantics of the CCS<sub>σ</sub>-calculus is given in terms of probabilistic automata defined inductively on the basis of the syntax, according to the rules shown in Figure [4](#S3.F4 "Figure 4 ‣ 3.2 Semantics ‣ 3 A variant of CCS with explicit scheduler ‣ Making Random Choices Invisible to the SchedulerThis work has been partially supported by the INRIA DREI Équipe Associée PRINTEMPS and by the INRIA ARC project ProNoBiS.").
ACT is the basic communication rule. In order for $`l:\alpha.P`$ to perform $`\alpha`$, the scheduler should select this process for execution, so the scheduler needs to be of the form $`{\sigma\hspace{0pt}{(l)}}.S`$. After the execution the complete process will continue as $`P \parallel S`$. The RES rule models restriction on channel $`a`$: communication on this channel is not allowed by the restricted process. Similarly to the section [2.2](#S2.SS2 "2.2 CCS with internal probabilistic choice ‣ 2 Preliminaries ‣ Making Random Choices Invisible to the SchedulerThis work has been partially supported by the INRIA DREI Équipe Associée PRINTEMPS and by the INRIA ARC project ProNoBiS."), we denote by $`{({\nu\hspace{0pt}a})}\hspace{0pt}\mu`$ the measure $`\mu'`$ such that $`{\mu'\hspace{0pt}{({{{({\nu\hspace{0pt}a})}\hspace{0pt}P} \parallel S})}} = {\mu\hspace{0pt}{({P \parallel S})}}`$ for all processes $`P`$ and $`{\mu'\hspace{0pt}{({R \parallel S})}} = 0`$ if $`R`$ is not of the form $`{({\nu\hspace{0pt}a})}\hspace{0pt}P`$. SUM1 models nondeterministic choice. If $`P \parallel S`$ can perform a transition to $`\mu`$, which means that $`S`$ selects one of the labels of $`P`$, then $`P + {Q \parallel S}`$ will perform the same transition, i.e. the branch $`P`$ of the choice will be selected and $`Q`$ will be discarded. For example
<table>
<tbody>
<tr>
<td></td>
<td>l1:a.P+l2:b.Q∥σ(l1).S⟶𝑎δ(P∥S)l_{1}:a.P+l_{2}:b.Q\parallel\sigma(l_{1}).S\overset{a}{\longrightarrow}\delta(P\parallel S)</td>
<td></td>
</tr>
</tbody>
</table>
Note that the operands of the sum do not have labels, the labels belong to the subprocesses of $`P`$ and $`Q`$. In the case of nested choices, the scheduler must go deep and select the label of a prefix, thus resolving all the choices at once.
PAR1 has a similar behavior for parallel composition. The scheduler selects $`P`$ to perform a transition on the basis of the label. The difference is that in this case $`Q`$ is not discarded; it remains in the continuation. $`\left. \mu \middle| Q \right.`$ denotes the measure $`\mu'`$ such that $`{\mu'\hspace{0pt}{({P\hspace{0pt}{|Q\|}\hspace{0pt}S})}} = {\mu\hspace{0pt}{({P \parallel S})}}`$. COM models synchronization. If $`P \parallel {\sigma\hspace{0pt}{(l_{1})}}`$ can perform the action $`a`$ and $`Q \parallel {\sigma\hspace{0pt}{(l_{2})}}`$ can perform $`\overline{a}`$, then $`\sigma\hspace{0pt}{(l_{1},l_{2})}`$, scheduling both $`l_{1}`$ and $`l_{2}`$ at the same time, can synchronize the two. PROB models internal probabilistic choice. Note that the scheduler cannot affect the outcome of the choice, it can only schedule the choice as a whole (that’s why a probabilistic sum has a label) and the process will move to a measure containing all the operands with corresponding probabilities.
Finally, BANG1 and BANG2 model replication. The rules are the same as in CCS<sub>p</sub>, with the addition of a re-labeling operator $`\rho_{k}`$. The reason for this is that we want to avoid ending up with multiple copies of the same label as the result of replication, since this would create ambiguities in scheduling as explained in section [3.3](#S3.SS3 "3.3 Deterministic labelings ‣ 3 A variant of CCS with explicit scheduler ‣ Making Random Choices Invisible to the SchedulerThis work has been partially supported by the INRIA DREI Équipe Associée PRINTEMPS and by the INRIA ARC project ProNoBiS."). $`\rho_{k}\hspace{0pt}{(P)}`$ replaces all labels $`l^{s}`$ inside $`P`$ with $`l^{s\hspace{0pt}k}`$, and it is defined as
<table>
<tbody>
<tr>
<td></td>
<td>ρk(ls:α.P)\displaystyle\rho_{k}(l^{s}:\alpha.P)</td>
<td>=\displaystyle=</td>
<td>lsk:α.ρk(P)\displaystyle l^{sk}:\alpha.\rho_{k}(P)</td>
<td></td>
</tr>
</tbody>
</table>
and homomorphically on the other operators (for instance $`\rho_{k}{(P|Q)} = \rho_{k}{(P)}|`$ $`\rho_{k}\hspace{0pt}{(Q)}`$). We also denote by $`\rho_{k}\hspace{0pt}{(\mu)}`$ the measure $`\mu'`$ such that $`{\mu'\hspace{0pt}{({{\rho_{k}\hspace{0pt}{(P)}} \parallel S})}} = {\mu\hspace{0pt}{({P \parallel S})}}`$. Note that we relabel only the resulting process, not the continuation of the scheduler: there is no need for relabeling the scheduler since we are free to choose the continuation as we please.
Let us give an example of how BANG1 and relabeling work. Let $`P =`$ $`l_{1}:a.l_{2}:b`$. To prove a transition for $`!P \parallel \sigma{(l_{1})}.S`$ we have to prove it for $`P`$ and then relabel the resulting process:
ACT
$`l_{1}:a.l_{2}:b \parallel \sigma{(l_{1})}.S\overset{a}{\longrightarrow}\delta{(l_{2}:b \parallel S)}`$
BANG1
$`!{(l_{1}:a.l_{2}:b)} \parallel \sigma{(l_{1})}.S\overset{a}{\longrightarrow}\delta{(l_{2}^{0}:b|!{(l_{1}^{1}:a.l_{2}^{1}:b)} \parallel S)}`$
As we can see in the example, when a process $`!P`$ is activated, the spawned copy of $`P`$ is relabeled by adding $`0`$ to the index of all the labels, and $`!P`$ is relabeled by adding $`1`$. So the labels of $`\rho_{0}\hspace{0pt}{(P)}`$ and $`\rho_{1}{(!P)}`$ will be disjoint. As remarked above, the continuation $`S`$ is not relabeled, if we want to perform $`b`$ after $`a`$ then $`S`$ should start with $`\sigma\hspace{0pt}{(l_{2}^{0})}`$.
### 3.3 Deterministic labelings
The idea in CCS<sub>σ</sub> is that a *syntactic* scheduler will be able to completely solve the nondeterminism of the process, without needing to rely on a *semantic* scheduler at the level of the automaton. This means that the execution of a process in parallel with a scheduler should be fully probabilistic. To achieve this we will impose a condition on the labels that we can use in CCS<sub>σ</sub> processes. A *labeling* is an assignment of labels to the prefixes and probabilistic sums of a process. We will require all labelings to be *deterministic* in the following sense.
###### Definition 1
A labeling of a process $`P`$ is *deterministic* iff for all schedulers $`S`$ there is only one transition rule $`P \parallel {S\hspace{0pt}\overset{\alpha}{\longrightarrow}\hspace{0pt}\mu}`$ that can be applied and the labelings of all processes $`P'`$ such that $`{\mu\hspace{0pt}{(P')}} > 0`$ are also deterministic.
A labeling is *linear* iff all labels are pairwise disjoint. We can show that linear labelings are preserved by transitions, which leads to the following proposition.
###### Proposition 1
A linear labeling is deterministic.
There are labelings that are deterministic without being linear. In fact, such labelings will be the means by which we hide information from the scheduler. However, the property of being deterministic is crucial since it implies that the scheduler will resolve all the nondeterminism of the process.
###### Proposition 2
Let $`P`$ be a CCS<sub>σ</sub> process with a deterministic labeling. Then for all schedulers $`S`$, the automaton produced by $`P \parallel S`$ is fully probabilistic.
## 4 Expressiveness of the syntactic scheduler
CCS<sub>σ</sub> with deterministic labelings allows us to separate probabilities from nondeterminism in a straightforward way: a process in parallel with a scheduler behaves in a fully probabilistic way and the nondeterminism arises from the fact that we can have many different schedulers. We may now ask the question: how powerful are the syntactic schedulers wrt the semantic ones, i.e. those defined directly over the automaton?
Let $`P`$ be as CCS<sub>p</sub> process and $`P_{\sigma}`$ be the CCS<sub>σ</sub> process obtained from $`P`$ by applying a linear labeling. We say that the semantic scheduler $`\zeta`$ of $`P`$ is equivalent to the syntactic scheduler $`S`$ of $`P_{\sigma}`$, written $`\zeta \sim_{P}S`$, iff the automata
<sup>5</sup>
<sup>5</sup>5Note that with a slight abuse of notation we will use a process to denote its corresponding probabilistic automaton.
$`e\hspace{0pt}t\hspace{0pt}r\hspace{0pt}e\hspace{0pt}e\hspace{0pt}{(P,\zeta)}`$ and $`P_{\sigma} \parallel S`$ are probabilistically bisimilar in the sense of \[[5](#bib.bib5)\].
A scheduler $`S`$ is *non-blocking* for a process $`P`$ if it always schedules some transitions, except when $`P`$ itself is blocked. Since semantic schedulers are usually not allowed to block, we will restraint ourselves to non-blocking syntactic schedulers to obtain a $`1 - 1`$ correspondence. Let $`S\hspace{0pt}e\hspace{0pt}m\hspace{0pt}{(P)}`$ be the set of the semantic schedulers for the process $`P`$, and $`S\hspace{0pt}y\hspace{0pt}n\hspace{0pt}{(P_{\sigma})}`$ be the set of the non-blocking syntactic schedulers for process $`P_{\sigma}`$. The following result holds for pure CCS processes (that is, CCS<sub>p</sub> processes without probabilistic choice).
###### Proposition 3
Let $`P`$ be a pure CCS process and let $`P_{\sigma}`$ be a CCS<sub>σ</sub> process obtained by adding a linear labeling to $`P`$. Then
<table>
<tbody>
<tr>
<td></td>
<td>∀ζ∈Sem(P)∃S∈Syn(Pσ):ζ∼PSand∀S∈Syn(Pσ)∃ζ∈Sem(P):ζ∼PS:for-all𝜁𝑆𝑒𝑚𝑃𝑆𝑆𝑦𝑛subscript𝑃𝜎subscriptsimilar-to𝑃𝜁𝑆and:for-all𝑆𝑆𝑦𝑛subscript𝑃𝜎𝜁𝑆𝑒𝑚𝑃subscriptsimilar-to𝑃𝜁𝑆\begin{array}[]{l}\forall\zeta\in Sem(P)\ \exists S\in Syn(P_{\sigma}):\zeta\sim_{P}S\quad\textrm{and}\\ \forall S\in Syn(P_{\sigma})\ \exists\zeta\in Sem(P):\zeta\sim_{P}S\end{array}</td>
<td></td>
</tr>
</tbody>
</table>
### 4.1 The scheduler in the presence of probabilistic choice
In Proposition [3](#S4.Ex7 "Proposition 3 ‣ 4 Expressiveness of the syntactic scheduler ‣ Making Random Choices Invisible to the SchedulerThis work has been partially supported by the INRIA DREI Équipe Associée PRINTEMPS and by the INRIA ARC project ProNoBiS.") we considered pure CCS processes in which the execution tree has only one possible execution. Now consider a process $`P`$ containing an internal probabilistic choice. Even if we fix the scheduler, the outcome of the choice is not always the same, so $`P \parallel S`$ could produce different executions. As a consequence, the syntactic schedulers we have defined are not enough to give us back all the semantic ones. Consider the process $`P = l:{(l_{1}:a +_{p}l_{2}:b)}`$. After the probabilistic choice, either $`l_{1}:a`$ or $`l_{2}:b`$ will be available, but we cannot know which one. As a consequence, we cannot create a scheduler that selects $`a`$ or $`b`$, whatever is available. In fact, it’s not even possible to create a non-blocking scheduler at all, both $`{\sigma\hspace{0pt}{(l)}}.{\sigma\hspace{0pt}{(l_{1})}}`$ and $`{\sigma\hspace{0pt}{(l)}}.{\sigma\hspace{0pt}{(l_{2})}}`$ will block on some executions.
The problem here is that the process can make choices that are independent from the scheduler, so the latter should adapt its behavior to the outcome of these choices. To achieve this, we extend CCS<sub>σ</sub> by adding a *scheduler choice* construct. The new syntax and semantics are displayed in Figure [5](#S4.F5 "Figure 5 ‣ 4.1 The scheduler in the presence of probabilistic choice ‣ 4 Expressiveness of the syntactic scheduler ‣ Making Random Choices Invisible to the SchedulerThis work has been partially supported by the INRIA DREI Équipe Associée PRINTEMPS and by the INRIA ARC project ProNoBiS."). A scheduler can be the sum $`\sum_{i}S_{i}`$ of several schedulers, and the outcome of the probabilistic choice (in the process) will determine the one to activate. In the case where more than one could be activated at the same time, we give preference to the first one in the sum, so the scheduler still behaves in a deterministic way.
$`\begin{array}{ll}
{S::=\ldots|\sum_{i}S_{i}\mspace{54mu}} & \begin{array}{ll}
\text{TEST} & \frac{{P \parallel {{S_{i}\hspace{0pt}\overset{\alpha}{\longrightarrow}\hspace{0pt}\mu}\mspace{39mu}{\forall{j\hspace{0pt}{<{i:P}\parallel}\hspace{0pt}S_{j}}}}}\nrightarrow}{P \parallel {\left. \sum{}_{i} \right.{S_{i}\hspace{0pt}\overset{\alpha}{\longrightarrow}\hspace{0pt}\mu}}}
\end{array}
\end{array}`$
Figure 5: Adding scheduler choice to CCS<sub>σ</sub>
In our previous example, we can use the scheduler $`\sigma\hspace{0pt}{(l)}\hspace{0pt}{({{\sigma\hspace{0pt}{(l_{1})}} + {\sigma\hspace{0pt}{(l_{2})}}})}`$ which will produce $`a`$ or $`b`$ depending on the outcome of the probabilistic choice. With the scheduler choice, we can retrieve the full power of the semantic scheduler for full CCS<sub>p</sub> processes.
###### Proposition 4
Proposition [3](#S4.Ex7 "Proposition 3 ‣ 4 Expressiveness of the syntactic scheduler ‣ Making Random Choices Invisible to the SchedulerThis work has been partially supported by the INRIA DREI Équipe Associée PRINTEMPS and by the INRIA ARC project ProNoBiS.") holds for full CCS<sub>p</sub> processes if we extend schedulers with scheduler choice.
### 4.2 Using non-linear labelings
Up to now we are using only linear labelings which, as we saw, give us the whole power of semantic schedulers. However, we can construct non-linear labelings that are still deterministic, that is there is still only one transition possible at any time even though we have multiple occurrences of the same label. There are various cases of useful non-linear labelings.
###### Proposition 5
Let $`P`$,$`Q`$ be CCS<sub>σ</sub> processes with deterministic labelings (not necessarily disjoint). The following labelings are all deterministic:
<table>
<tbody>
<tr>
<td></td>
<td>l:(P+pQ):𝑙subscript𝑝𝑃𝑄\displaystyle l:(P+_{p}Q)</td>
<td></td>
<td>(2)</td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td>l1:a.P+l2:b.Q\displaystyle l_{1}:a.P+l_{2}:b.Q</td>
<td></td>
<td>(3)</td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td>(νa)(νb)(l1:a.P+l1:b.Q|l2:a¯)\displaystyle(\nu a)(\nu b)(l_{1}:a.P+l_{1}:b.Q\;|\;l_{2}:\bar{a})</td>
<td></td>
<td>(4)</td>
</tr>
</tbody>
</table>
Consider the case where $`P`$ and $`Q`$ in the above proposition share the same labels. In ([2](#S4.E2 "In Proposition 5 ‣ 4.2 Using non-linear labelings ‣ 4 Expressiveness of the syntactic scheduler ‣ Making Random Choices Invisible to the SchedulerThis work has been partially supported by the INRIA DREI Équipe Associée PRINTEMPS and by the INRIA ARC project ProNoBiS.")) the scheduler cannot select an action inside $`P,Q`$, it must select the choice itself. After the choice, only one of $`P,Q`$ will be available so there will be no ambiguity in selecting transitions. The case ([3](#S4.E3 "In Proposition 5 ‣ 4.2 Using non-linear labelings ‣ 4 Expressiveness of the syntactic scheduler ‣ Making Random Choices Invisible to the SchedulerThis work has been partially supported by the INRIA DREI Équipe Associée PRINTEMPS and by the INRIA ARC project ProNoBiS.")) is similar but with nondeterministic choice. Now the guarding prefixes must have different labels, since the scheduler should be able to resolve the choice, however after the choice only one of $`P,Q`$ will be available. Hence, again, the multiple copies of the labels do not constitute a problem. In ([4](#S4.E4 "In Proposition 5 ‣ 4.2 Using non-linear labelings ‣ 4 Expressiveness of the syntactic scheduler ‣ Making Random Choices Invisible to the SchedulerThis work has been partially supported by the INRIA DREI Équipe Associée PRINTEMPS and by the INRIA ARC project ProNoBiS.")) we allow the same label on the guarding prefixes of a nondeterministic choice. This is because the guarding channels $`a,b`$ are restricted and only one of the corresponding output actions is available ($`\overline{a}`$). As a consequence, there is no ambiguity in selecting transitions. A scheduler $`\sigma\hspace{0pt}{(l_{1},l_{2})}`$ can only perform a synchronization on $`a`$, even though $`l_{1}`$ appears twice.
However, using multiple copies of a label limits the power of the scheduler, since the labels provide information about the outcome of a probabilistic choice (and allow the scheduler to choose different strategies through the use of the scheduler choice). In fact, this is exactly the technique we will use to archive the goals described in Section [1](#S1 "1 Introduction ‣ Making Random Choices Invisible to the SchedulerThis work has been partially supported by the INRIA DREI Équipe Associée PRINTEMPS and by the INRIA ARC project ProNoBiS."). Consider for example the process:
<table>
<tbody>
<tr>
<td></td>
<td>l:(l1:a¯+pl1:a¯)|l2:a.P|l3:a.Ql:(l_{1}:\bar{a}+_{p}l_{1}:\bar{a})\;|\;l_{2}:a.P\;|\;l_{3}:a.Q</td>
<td></td>
<td>(5)</td>
</tr>
</tbody>
</table>
From Proposition [5](#Pt0.A1.EGx2 "Proposition 5 ‣ 4.2 Using non-linear labelings ‣ 4 Expressiveness of the syntactic scheduler ‣ Making Random Choices Invisible to the SchedulerThis work has been partially supported by the INRIA DREI Équipe Associée PRINTEMPS and by the INRIA ARC project ProNoBiS.")([2](#S4.E2 "In Proposition 5 ‣ 4.2 Using non-linear labelings ‣ 4 Expressiveness of the syntactic scheduler ‣ Making Random Choices Invisible to the SchedulerThis work has been partially supported by the INRIA DREI Équipe Associée PRINTEMPS and by the INRIA ARC project ProNoBiS.")) this labeling is deterministic. However, since both branches of the probabilistic sum have the same label $`l_{1}`$, the scheduler cannot resolve the choice between $`P`$ and $`Q`$ based on the outcome of the choice. There is still nondeterminism: the scheduler $`{\sigma\hspace{0pt}{(l)}}.{\sigma\hspace{0pt}{(l_{1},l_{2})}}`$ will select $`P`$ and the scheduler $`{\sigma\hspace{0pt}{(l)}}.{\sigma\hspace{0pt}{(l_{1},l_{3})}}`$ will select $`Q`$. However this selection will be independent from the outcome of the probabilistic choice.
Note that we did not impose any direct restrictions on the schedulers, we still consider all possible syntactic schedulers for the process ([5](#S4.E5 "In 4.2 Using non-linear labelings ‣ 4 Expressiveness of the syntactic scheduler ‣ Making Random Choices Invisible to the SchedulerThis work has been partially supported by the INRIA DREI Équipe Associée PRINTEMPS and by the INRIA ARC project ProNoBiS.")) above. However, having the same label twice limits the power of the syntactic schedulers with respect to the semantic ones. This approach has the advantage that the restrictions are limited to the choices with the same label. We already know that having pairwise different labels gives the full power of the semantic scheduler. So the restriction is local to the place where we, intentionally, put the same labels.
## 5 Testing relations for CCS<sub>σ</sub> processes
Testing relations \[[19](#bib.bib19)\] are a method of comparing processes by considering their interaction with the environment. A *test* is a process running in parallel with the one being tested and which can perform a distinguished action $`\omega`$ that represents success. Two processes are testing equivalent if they can pass the same tests. This idea is very useful for the analysis of security protocols, as suggested in \[[20](#bib.bib20)\], since a test can be seen as an adversary who interferes with a communication agent and declares $`\omega`$ if an attack is successful. Then two processes are testing equivalent if they are vulnerable to the same attacks.
In the probabilistic setting we take the approach of \[[13](#bib.bib13)\] which considers the exact probability of passing a test (in contrast to \[[10](#bib.bib10)\] which considers only the ability to pass a test with probability non-zero (may-testing) or one (must-testing)). This approach leads to the definition of two preorders $`\sqsubseteq_{\text{may}}`$ and $`\sqsubseteq_{\text{must}}`$. $`P \sqsubseteq_{\text{may}}Q`$ means that the if $`P`$ can pass $`O`$ then $`Q`$ can also pass $`O`$ with the same probability. $`P \sqsubseteq_{\text{must}}Q`$ means that if $`P`$ always passes $`O`$ with at least some probability then $`Q`$ always passes $`O`$ with at least the same probability.
A labeling of a process is *fresh* (with respect to a set $`\mathcal{P}`$ of processes) if it is linear and its labels do not appear in any other process in $`\mathcal{P}`$. A test $`O`$ is a CCS<sub>σ</sub> process with a fresh labeling, containing the distinguished action $`\omega`$. Let $`{Test}_{\mathcal{P}}`$ denote the set of all tests with respect to $`\mathcal{P}`$ and let $`{(\nu)}\hspace{0pt}P`$ denote the restriction on all channels of $`P`$, thus allowing only $`\tau`$ actions. We define $`p_{\omega}\hspace{0pt}{(P,S,O)}`$ to be the probability of the set of executions of the fully probabilistic automaton $`{{(\nu)}\hspace{0pt}{(\left. P \middle| O \right.)}} \parallel S`$ that contain $`\omega`$. Note that this set can be produced as a countable union of disjoint cones so its probability is well-defined.
###### Definition 2
Let $`P,Q`$ be CCS<sub>σ</sub> processes. We define must and may testing preorders as follows:
<table>
<tbody>
<tr>
<td></td>
<td>P⊑mayQ iff ∀O∈𝑇𝑒𝑠𝑡P,Q∀SP∈𝑆𝑦𝑛((ν)(P|O))∃SQ∈𝑆𝑦𝑛((ν)(Q|O)):pω(P,SP,O)≤pω(Q,SQ,O)P⊑mustQ iff ∀O∈𝑇𝑒𝑠𝑡P,Q∀SQ∈𝑆𝑦𝑛((ν)(Q|O))∃SP∈𝑆𝑦𝑛((ν)(P|O)):pω(P,SP,O)≤pω(Q,SQ,O)subscriptsquare-image-of-or-equalsmay𝑃𝑄 iff :for-all𝑂subscript𝑇𝑒𝑠𝑡𝑃𝑄for-allsubscript𝑆𝑃𝑆𝑦𝑛𝜈conditional𝑃𝑂subscript𝑆𝑄𝑆𝑦𝑛𝜈conditional𝑄𝑂absentmissing-subexpressionmissing-subexpressionsubscript𝑝𝜔𝑃subscript𝑆𝑃𝑂subscript𝑝𝜔𝑄subscript𝑆𝑄𝑂subscriptsquare-image-of-or-equalsmust𝑃𝑄 iff :for-all𝑂subscript𝑇𝑒𝑠𝑡𝑃𝑄for-allsubscript𝑆𝑄𝑆𝑦𝑛𝜈conditional𝑄𝑂subscript𝑆𝑃𝑆𝑦𝑛𝜈conditional𝑃𝑂absentmissing-subexpressionmissing-subexpressionsubscript𝑝𝜔𝑃subscript𝑆𝑃𝑂subscript𝑝𝜔𝑄subscript𝑆𝑄𝑂\begin{array}[]{lcl}P\sqsubseteq_{\textrm{{may}}}Q&\textrm{ iff }&\forall O\in\it{Test}_{P,Q}\ \forall S_{P}\in Syn((\nu)(P\;|\;O))\ \exists S_{Q}\in Syn((\nu)(Q\;|\;O)):\\[2.0pt] &&\qquad p_{\omega}(P,S_{P},O)\leq p_{\omega}(Q,S_{Q},O)\\[4.0pt] P\sqsubseteq_{\textrm{{must}}}Q&\textrm{ iff }&\forall O\in\it{Test}_{P,Q}\ \forall S_{Q}\in Syn((\nu)(Q\;|\;O))\ \exists S_{P}\in Syn((\nu)(P\;|\;O)):\\[2.0pt] &&\qquad p_{\omega}(P,S_{P},O)\leq p_{\omega}(Q,S_{Q},O)\end{array}</td>
<td></td>
</tr>
</tbody>
</table>
We also define $`\approx_{\text{may}}, \approx_{\text{must}}`$ to be the equivalences induced by $`\sqsubseteq_{\text{may}}, \sqsubseteq_{\text{must}}`$ respectively.
A context $`C`$ is a process with a hole. A preorder $`\sqsubseteq`$ is a precongruence if $`P \sqsubseteq Q`$ implies $`{C\hspace{0pt}{\lbrack P\rbrack}} \sqsubseteq {C\hspace{0pt}{\lbrack Q\rbrack}}`$ for all contexts $`C`$. May and must testing are precongruences if we restrict to contexts with fresh labelings and without occurrences of $`+`$. This result is essentially an adaptation to our framework of the analogous precongruence property in \[[3](#bib.bib3)\].
###### Proposition 6
Let $`P,Q`$ be CCS<sub>σ</sub> processes such that $`P \sqsubseteq_{\text{may}}Q`$ and let $`C`$ be a context with a fresh labeling and in which $`+`$ does not occur. Then $`{C\hspace{0pt}{\lbrack P\rbrack}} \sqsubseteq_{\text{may}}{C\hspace{0pt}{\lbrack Q\rbrack}}`$. Similarly for $`\sqsubseteq_{\text{must}}`$.
This also implies that $`\approx_{\text{may}}, \approx_{\text{must}}`$ are congruences. Note that $`P,Q`$ in the above proposition are not required to have linear labelings, $`P`$ might include multiple occurrences of the same label thus limiting the power of the schedulers $`S_{P}`$. This shows the locality of the scheduler’s restriction: some choices inside $`P`$ are hidden from the scheduler but the rest of the context is fully visible.
If we remove the freshness condition then Proposition [6](#Thmproposition6 "Proposition 6 ‣ 5 Testing relations for CCSσ processes ‣ Making Random Choices Invisible to the SchedulerThis work has been partially supported by the INRIA DREI Équipe Associée PRINTEMPS and by the INRIA ARC project ProNoBiS.") is no longer true. Let $`P = l_{1}:a.l_{2}:b`$, $`Q = l_{3}:a.l_{4}:b`$ and $`C = l:{(l_{1}:a.l_{2}:c +_{p}{\lbrack\rbrack})}`$. We have $`P \approx_{\text{may}}Q`$ but $`{C\hspace{0pt}{\lbrack P\rbrack}},{C\hspace{0pt}{\lbrack Q\rbrack}}`$ can be separated by the test $`{O = \overline{a}}.\overline{b}.\left. \omega \middle| \overline{a} \right..\overline{c}.\omega`$ (The labeling is omitted for simplicity since tests always have fresh labelings.) It is easy to see that $`C\hspace{0pt}{\lbrack Q\rbrack}`$ can pass the test with probability $`1`$ by selecting the correct branch of $`O`$ based on the outcome of the probabilistic choice. In $`C\hspace{0pt}{\lbrack P\rbrack}`$ this is not possible because of the labels $`l_{1},l_{2}`$ that are common in $`P,C`$.
We can now state the result that we announced in Section [1](#S1 "1 Introduction ‣ Making Random Choices Invisible to the SchedulerThis work has been partially supported by the INRIA DREI Équipe Associée PRINTEMPS and by the INRIA ARC project ProNoBiS.").
###### Theorem 5.1
Let $`P,Q`$ be CCS<sub>σ</sub> processes and $`C`$ a context with a fresh labeling and without occurrences of bang. Then
<table>
<tbody>
<tr>
<td></td>
<td>l:(C[l1:τ.P]+pC[l1:τ.Q])\displaystyle l:(C[l_{1}:\tau.P]+_{p}C[l_{1}:\tau.Q])</td>
<td>≈maysubscriptmay\displaystyle\approx_{\textrm{{may}}}</td>
<td>C[l:(P+pQ)]and\displaystyle C[l:(P+_{p}Q)]\quad\textrm{and}</td>
<td></td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td>l:(C[l1:τ.P]+pC[l1:τ.Q])\displaystyle l:(C[l_{1}:\tau.P]+_{p}C[l_{1}:\tau.Q])</td>
<td>≈mustsubscriptmust\displaystyle\approx_{\textrm{{must}}}</td>
<td>C[l:(P+pQ)]\displaystyle C[l:(P+_{p}Q)]</td>
<td></td>
</tr>
</tbody>
</table>
The proof is given in the appendix.
There are two crucial points in the above Theorem. The first is that the labels of the context are copied, thus the scheduler cannot distinguish between $`C{\lbrack l_{1}:\tau.P\rbrack}`$ and $`C{\lbrack l_{1}:\tau.Q\rbrack}`$ based on the labels of the context. The second is that $`P,Q`$ are protected by a $`\tau`$ action labeled by the same label $`l_{1}`$. This is to ensure that in the case of a nondeterministic sum ($`C = {R + {\lbrack\rbrack}}`$) the scheduler cannot find out whether the second operand of the choice is $`P`$ or $`Q`$ unless it commits to selecting the second operand. For example let $`R = l_{1}:{(l_{2}:a +_{0.5}0)},P = l_{3}:a,Q = 0`$. Then $`R_{1} = {{({R + a})} +_{0.1}{({R + 0})}}`$ is not testing equivalent to $`R_{2} = {R + {({a +_{0.1}0})}}`$ since they can be separated by $`{O = \overline{a}}.\omega`$ and a scheduler that resolves $`R + a`$ to $`a`$ and $`R + 0`$ to $`R`$. However, if we take $`R_{1}' = {(R + l:\tau.a)} +_{0.1}{(R + l:\tau.0)}`$ then $`R_{1}'`$ is testing equivalent to $`R_{2}`$ since the scheduler will have to resolve both branches of $`R_{1}'`$ in the same way (even though we still have non-determinism).
The problem with replication is simply the persistence of the processes. It is clear that $`!P +_{p}!Q`$ cannot be equivalent in any way to $`!{(P +_{p}Q)}`$ since the first replicates only one of $`P,Q`$ while the second replicates both. However Theorem [5.1](#Pt0.A1.EGx3 "Theorem 5.1 ‣ 5 Testing relations for CCSσ processes ‣ Making Random Choices Invisible to the SchedulerThis work has been partially supported by the INRIA DREI Équipe Associée PRINTEMPS and by the INRIA ARC project ProNoBiS.") together with Proposition [6](#Thmproposition6 "Proposition 6 ‣ 5 Testing relations for CCSσ processes ‣ Making Random Choices Invisible to the SchedulerThis work has been partially supported by the INRIA DREI Équipe Associée PRINTEMPS and by the INRIA ARC project ProNoBiS.") imply that
<table>
<tbody>
<tr>
<td></td>
<td>C′[l:(C[l1:τ.P]+pC[l1:τ.Q])]≈mayC′[C[l:(P+pQ)]]C^{\prime}[l:(C[l_{1}:\tau.P]+_{p}C[l_{1}:\tau.Q])]\approx_{\textrm{{may}}}C^{\prime}[C[l:(P+_{p}Q)]]</td>
<td></td>
<td>(6)</td>
</tr>
</tbody>
</table>
where $`C`$ is a context without bang and $`C'`$ is a context without $`+`$. The same is also true for $`\approx_{\text{must}}`$. This means that we can lift the sum towards the root of the context until we reach a bang. Intuitively we cannot move the sum outside the bang since each replicated copy must perform a different probabilistic choice with a possibly different outcome.
Theorem [5.1](#Pt0.A1.EGx3 "Theorem 5.1 ‣ 5 Testing relations for CCSσ processes ‣ Making Random Choices Invisible to the SchedulerThis work has been partially supported by the INRIA DREI Équipe Associée PRINTEMPS and by the INRIA ARC project ProNoBiS.") shows that the probabilistic choice is indeed private to the process and invisible to the scheduler. The process can perform it at any time, even in the very beginning of the execution, without making any difference to an outside observer.
## 6 An application to security
In this section we discuss an application of our framework to anonymity. In particular, we show how to specify the Dining Cryptographers protocol \[[21](#bib.bib21)\] so that it is robust to scheduler-based attacks. We first propose a method to encode *secret value passing*, which will turn out to be useful for the specification:
<table>
<tbody>
<tr>
<td></td>
<td>l:c(x).P\displaystyle l:c(x).P</td>
<td>=ΔΔ\displaystyle\overset{\scriptscriptstyle\Delta}{=}</td>
<td>∑il:cvi.P[vi/x]\displaystyle\textstyle{\sum_{i}\>}l:cv_{i}.P[v_{i}/x]</td>
<td></td>
<td>(7)</td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td>l:c¯⟨v⟩.P\displaystyle l:\bar{c}\langle v\rangle.P</td>
<td>=ΔΔ\displaystyle\overset{\scriptscriptstyle\Delta}{=}</td>
<td>l:cv¯.P\displaystyle l:\overline{cv}.P</td>
<td></td>
<td>(8)</td>
</tr>
</tbody>
</table>
This is the usual encoding of value passing in CSS except that we use the same label in all the branches of the nondeterministic sum. To ensure that the resulting labeling will be deterministic we should restrict the channels $`c\hspace{0pt}v_{i}`$ and make sure that there will be at most one output on $`c`$. We will write $`{({\nu\hspace{0pt}c})}\hspace{0pt}P`$ for $`{({\nu\hspace{0pt}c\hspace{0pt}v_{1}})}\hspace{0pt}\ldots\hspace{0pt}{({\nu\hspace{0pt}c\hspace{0pt}v_{n}})}\hspace{0pt}P`$. For example, the labeling of the following process is deterministic:
<table>
<tbody>
<tr>
<td></td>
<td>(νc)(l1:c(x).P|l:(l2:c¯⟨v1⟩+pl2:c¯⟨v2⟩))(\nu c)(l_{1}:c(x).P\;|\;l:(l_{2}:\bar{c}\langle v_{1}\rangle+_{p}l_{2}:\bar{c}\langle v_{2}\rangle))\qquad</td>
<td></td>
</tr>
</tbody>
</table>
This case is a combination of the cases ([2](#S4.E2 "In Proposition 5 ‣ 4.2 Using non-linear labelings ‣ 4 Expressiveness of the syntactic scheduler ‣ Making Random Choices Invisible to the SchedulerThis work has been partially supported by the INRIA DREI Équipe Associée PRINTEMPS and by the INRIA ARC project ProNoBiS.")) and ([4](#S4.E4 "In Proposition 5 ‣ 4.2 Using non-linear labelings ‣ 4 Expressiveness of the syntactic scheduler ‣ Making Random Choices Invisible to the SchedulerThis work has been partially supported by the INRIA DREI Équipe Associée PRINTEMPS and by the INRIA ARC project ProNoBiS.")) of Proposition [5](#Pt0.A1.EGx2 "Proposition 5 ‣ 4.2 Using non-linear labelings ‣ 4 Expressiveness of the syntactic scheduler ‣ Making Random Choices Invisible to the SchedulerThis work has been partially supported by the INRIA DREI Équipe Associée PRINTEMPS and by the INRIA ARC project ProNoBiS."). The two outputs on $`c`$ are on different branches of the probabilistic sum, so during an execution at most one of them will be available. Thus there is no ambiguity in scheduling the sum produced by $`c\hspace{0pt}{(x)}`$. The scheduler $`{\sigma\hspace{0pt}{(l)}}.{\sigma\hspace{0pt}{(l_{1},l_{2})}}`$ will perform a synchronization on $`c\hspace{0pt}v_{1}`$ or $`c\hspace{0pt}v_{2}`$, whatever is available after the probabilistic choice. In other words, using the labels we manage to hide the information about which value was transmitted to $`P`$.
### 6.1 Dining cryptographers with probabilistic master
<table>
<tbody>
<tr>
<td></td>
<td>Master𝑀𝑎𝑠𝑡𝑒𝑟\displaystyle Master</td>
<td>=ΔΔ\displaystyle\overset{\scriptscriptstyle\Delta}{=}</td>
<td>l1:∑i=02pi(m¯0⟨i== 0⟩⏟l2|m¯1⟨i== 1⟩⏟l3|m¯2⟨i== 2⟩⏟l4):subscript𝑙1superscriptsubscript𝑖02subscript𝑝𝑖subscript⏟subscript¯𝑚0delimited-⟨⟩𝑖== 0subscript𝑙2subscript⏟subscript¯𝑚1delimited-⟨⟩𝑖==1subscript𝑙3subscript⏟subscript¯𝑚2delimited-⟨⟩𝑖==2subscript𝑙4\displaystyle l_{1}:\textstyle{\sum_{i=0}^{2}\>}p_{i}(\underbrace{\bar{m}_{0}\langle i\,\textrm{{==}}\,0\rangle}_{l_{2}}\;|\;\underbrace{\bar{m}_{1}\langle i\,\textrm{{==}}\,1\rangle}_{l_{3}}\;|\;\underbrace{\bar{m}_{2}\langle i\,\textrm{{==}}\,2\rangle}_{l_{4}})</td>
<td></td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td>Crypti𝐶𝑟𝑦𝑝subscript𝑡𝑖\displaystyle Crypt_{i}</td>
<td>=ΔΔ\displaystyle\overset{\scriptscriptstyle\Delta}{=}</td>
<td>mi(pay)⏟l5,i.ci,i(coin1)⏟l6,i.ci,i⊕1(coin2)⏟l7,i.out¯i⟨pay⊗coin1⊗coin2⟩⏟l8,iformulae-sequencesubscript⏟subscript𝑚𝑖𝑝𝑎𝑦subscript𝑙5𝑖subscript⏟subscript𝑐𝑖𝑖𝑐𝑜𝑖subscript𝑛1subscript𝑙6𝑖subscript⏟subscript𝑐𝑖direct-sum𝑖1𝑐𝑜𝑖subscript𝑛2subscript𝑙7𝑖subscript⏟subscript¯𝑜𝑢𝑡𝑖delimited-⟨⟩tensor-producttensor-product𝑝𝑎𝑦𝑐𝑜𝑖subscript𝑛1𝑐𝑜𝑖subscript𝑛2subscript𝑙8𝑖\displaystyle\underbrace{m_{i}(pay)}_{l_{5,i}}.\underbrace{c_{i,i}(coin_{1})}_{l_{6,i}}.\underbrace{c_{i,i\oplus 1}(coin_{2})}_{l_{7,i}}.\underbrace{\overline{out}_{i}\langle pay\otimes coin_{1}\otimes coin_{2}\rangle}_{l_{8,i}}</td>
<td></td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td>Coini𝐶𝑜𝑖subscript𝑛𝑖\displaystyle Coin_{i}</td>
<td>=ΔΔ\displaystyle\overset{\scriptscriptstyle\Delta}{=}</td>
<td>l9,i:((c¯i,i⟨0⟩⏟l10,i|c¯i⊖1,i⟨0⟩⏟l11,i)+0.5(c¯i,i⟨1⟩⏟l10,i|c¯i⊖1,i⟨1⟩⏟l11,i)):subscript𝑙9𝑖subscript0.5conditionalsubscript⏟subscript¯𝑐𝑖𝑖delimited-⟨⟩0subscript𝑙10𝑖subscript⏟subscript¯𝑐symmetric-difference𝑖1𝑖delimited-⟨⟩0subscript𝑙11𝑖conditionalsubscript⏟subscript¯𝑐𝑖𝑖delimited-⟨⟩1subscript𝑙10𝑖subscript⏟subscript¯𝑐symmetric-difference𝑖1𝑖delimited-⟨⟩1subscript𝑙11𝑖\displaystyle l_{9,i}:((\underbrace{\bar{c}_{i,i}\langle 0\rangle}_{l_{10,i}}\;|\;\underbrace{\bar{c}_{i\ominus 1,i}\langle 0\rangle}_{l_{11,i}})+_{0.5}(\underbrace{\bar{c}_{i,i}\langle 1\rangle}_{l_{10,i}}\;|\;\underbrace{\bar{c}_{i\ominus 1,i}\langle 1\rangle}_{l_{11,i}}))</td>
<td></td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td>Prot𝑃𝑟𝑜𝑡\displaystyle Prot</td>
<td>=ΔΔ\displaystyle\overset{\scriptscriptstyle\Delta}{=}</td>
<td>(νm→)(Master|(νc→)(∏i=02Crypti|∏i=02Coini))\displaystyle(\nu\vec{m})(Master\;|\;(\nu\vec{c})(\textstyle{\prod_{i=0}^{2}\>}Crypt_{i}\;|\;\textstyle{\prod_{i=0}^{2}\>}Coin_{i}))</td>
<td></td>
</tr>
</tbody>
</table>
Figure 6: Encoding of the dining cryptographers with probabilistic master
The problem of the Dining Cryptographers is the following: Three cryptographers are dining together. At the end of the dinner, the bill has to be paid by either one of them or by another agent called the master. The master decides who will pay and then informs each of them separately whether he has to pay or not. The cryptographers would like to find out whether the payer is the master or one of them. However, in the latter case, they also wish to keep the payer anonymous.
The Dining Cryptographers Protocol (DCP) solves the above problem as follows: each cryptographer tosses a fair coin which is visible to himself and his neighbor to the right. Each cryptographer checks the two adjacent coins and, if he is not paying, announces agree if they are the same and disagree otherwise. However, the paying cryptographer will say the opposite. It can be proved that if the number of disagrees is even, then the master is paying; otherwise, one of the cryptographers is paying \[[21](#bib.bib21)\].
An external observer $`O`$ is supposed to see only the three announcements $`{\overline{o\hspace{0pt}u\hspace{0pt}t}}_{i}\hspace{0pt}{\langle\ldots\rangle}`$. As discussed in \[[22](#bib.bib22)\], DCP satisfies anonymity if we abstract from their order. If their order is observable, on the contrary, then a scheduler can reveal the identity of the payer to $`O`$ simply by forcing the payer to make his announcement first. Of course, this is possible only if the scheduler is unrestricted and can choose its strategy depending on the decision of the master (or on the results of the coins).
In our framework we can solve the problem by giving a specification of the DCP in which the choices of the master and of the coins are made invisible to the scheduler. The specification is shown in Figure [6](#S6.F6 "Figure 6 ‣ 6.1 Dining cryptographers with probabilistic master ‣ 6 An application to security ‣ Making Random Choices Invisible to the SchedulerThis work has been partially supported by the INRIA DREI Équipe Associée PRINTEMPS and by the INRIA ARC project ProNoBiS."). We use some meta-syntax for brevity: The symbols $`\oplus`$ and $`\ominus`$ represent the addition and subtraction modulo 3, while $`\otimes`$ represents the addition modulo 2 (xor). The notation $`i\hspace{0pt}\text{==}\hspace{0pt}n`$ stands for $`1`$ if $`i = n`$ and $`0`$ otherwise.
There are many sources of nondeterminism: the order of communication between the master and the cryptographers, the order of reception of the coins, and the order of the announcements. The crucial points of our specification, which make the nondeterministic choices independent from the probabilistic ones, are: (a) all communications internal to the protocol (master-cryptographers and cryptographers-coins) are done by secret value passing, and (b) in each probabilistic choice the different branches have the same labels. For example, all branches of the master contain an output on $`m_{0}`$, always labeled by $`l_{2}`$, but with different values each time.
Thanks to the above independence, the specification satisfy strong probabilistic anonymity. There are various equivalent definitions of this property, we follow here the version presented in \[[22](#bib.bib22)\]. Let $`\overset{\rightarrow}{o}`$ represent an observable (the sequence of announcements), and $`p_{S}\hspace{0pt}{(\left. \overset{\rightarrow}{o} \middle| {{\overline{m}}_{i}\hspace{0pt}{\langle 1\rangle}} \right.)}`$ represent the conditional probability, under the scheduler $`S`$, that the protocol produces $`\overset{\rightarrow}{o}`$ given that the master has selected Cryptographer $`i`$ as the payer.
###### Proposition 7 (Strong probabilistic anonymity)
The protocol in Figure [6](#S6.F6 "Figure 6 ‣ 6.1 Dining cryptographers with probabilistic master ‣ 6 An application to security ‣ Making Random Choices Invisible to the SchedulerThis work has been partially supported by the INRIA DREI Équipe Associée PRINTEMPS and by the INRIA ARC project ProNoBiS.") satisfies the following property: for all schedulers $`S`$ and all observables $`\overset{\rightarrow}{o}`$, $`{p_{S}\hspace{0pt}{(\left. \overset{\rightarrow}{o} \middle| {{\overline{m}}_{0}\hspace{0pt}{\langle 1\rangle}} \right.)}} = {p_{S}\hspace{0pt}{(\left. \overset{\rightarrow}{o} \middle| {{\overline{m}}_{1}\hspace{0pt}{\langle 1\rangle}} \right.)}} = {p_{S}\hspace{0pt}{(\left. \overset{\rightarrow}{o} \middle| {{\overline{m}}_{2}\hspace{0pt}{\langle 1\rangle}} \right.)}}`$
Note that different schedulers will produce different traces (we still have nondeterminism) but they will not depend on the choice of the master.
Some previous treatment of the DCP, including \[[22](#bib.bib22)\], had solved the problem of the leak of information due to too-powerful schedulers by simply considering as observable sets of announcements instead than sequences. Thus one could think that using a true concurrent semantics, for instance event structures, would solve the problem. We would like to remark that this is false: true concurrency would weaken the scheduler enough in the case of the DCP, but not in general. For instance, it would not help in the anonymity example in the introduction.
### 6.2 Dining cryptographers with nondeterministic master
<table>
<tbody>
<tr>
<td></td>
<td>P::=…|l:{P}CP::=P∥S,TINDEPP∥T⟶𝛼μl:{P}∥σ(l).S,T⟶𝛼μ′where μ′(P′∥S,T′)=μ(P′∥T′)\begin{array}[]{l@{\hspace{30pt}}l}\begin{array}[]{ll}P&::=\ldots\ \,|\ \,l:\{P\}\\[2.0pt] CP&::=P\parallel S,T\end{array}\hfil\hskip 30.&\begin{array}[]{ll}\textrm{INDEP}&\frac{\displaystyle P\parallel T\overset{\alpha}{\longrightarrow}\mu}{\displaystyle\begin{array}[]{c}l:\{P\}\parallel\sigma(l).S,T\overset{\alpha}{\longrightarrow}\mu^{\prime}\\ \textrm{where }\mu^{\prime}(P^{\prime}\parallel S,T^{\prime})=\mu(P^{\prime}\parallel T^{\prime})\end{array}}\end{array}\end{array}</td>
<td></td>
</tr>
</tbody>
</table>
Figure 7: Adding an “independent” scheduler to the calculus
We sketch here a method to hide also certain nondeterministic choices from the scheduler, and we show an application to the variant of the Dining Cryptographers with nondeterministic master.
First we need to extend the calculus with the concept of a second *independent* scheduler $`T`$ that we assume to solve the nondeterministic choices that we want to make transparent to the main scheduler $`S`$. The new syntax and semantics are shown in Figure [7](#S6.F7 "Figure 7 ‣ 6.2 Dining cryptographers with nondeterministic master ‣ 6 An application to security ‣ Making Random Choices Invisible to the SchedulerThis work has been partially supported by the INRIA DREI Équipe Associée PRINTEMPS and by the INRIA ARC project ProNoBiS."). $`l:{\{ P\}}`$ represents a process where the scheduling of $`P`$ is protected from the main scheduler $`S`$. The scheduler $`S`$ can “ask” $`T`$ to schedule $`P`$ by selecting the label $`l`$. Then $`T`$ resolves the nondeterminism of $`P`$ as expressed by the INDEP rule. Note that we need to adjust also the other rules of the semantics to take $`T`$ into account, but this change is straightforward. We assume that $`T`$ does not collaborate with $`S`$ so we do not need to worry about the labels in $`P`$.
To model the dining cryptographers with nondeterministic master we replace the $`M\hspace{0pt}a\hspace{0pt}s\hspace{0pt}t\hspace{0pt}e\hspace{0pt}r`$ process in Figure [6](#S6.F6 "Figure 6 ‣ 6.1 Dining cryptographers with probabilistic master ‣ 6 An application to security ‣ Making Random Choices Invisible to the SchedulerThis work has been partially supported by the INRIA DREI Équipe Associée PRINTEMPS and by the INRIA ARC project ProNoBiS.") by the following one.
<table>
<tbody>
<tr>
<td></td>
<td>Master=Δl1:{∑i=02l12,i:τ.(m¯0⟨i== 0⟩⏟l2|m¯1⟨i== 1⟩⏟l3|m¯2⟨i== 2⟩⏟l4)}:𝑀𝑎𝑠𝑡𝑒𝑟Δsubscript𝑙1conditional-setsuperscriptsubscript𝑖02subscript𝑙12𝑖formulae-sequence𝜏subscript⏟subscript¯𝑚0delimited-⟨⟩𝑖== 0subscript𝑙2subscript⏟subscript¯𝑚1delimited-⟨⟩𝑖==1subscript𝑙3subscript⏟subscript¯𝑚2delimited-⟨⟩𝑖==2subscript𝑙4Master\overset{\scriptscriptstyle\Delta}{=}l_{1}:\big{\{}\textstyle{\sum_{i=0}^{2}\>}l_{12,i}:\tau.(\underbrace{\bar{m}_{0}\langle i\,\textrm{{==}}\,0\rangle}_{l_{2}}\;|\;\underbrace{\bar{m}_{1}\langle i\,\textrm{{==}}\,1\rangle}_{l_{3}}\;|\;\underbrace{\bar{m}_{2}\langle i\,\textrm{{==}}\,2\rangle}_{l_{4}})\big{\}}</td>
<td></td>
</tr>
</tbody>
</table>
Essentially we have replaced the probabilistic choice by a *protected* nondeterministic one. Note that the labels of the operands are different but this is not a problem since this choice will be scheduled by $`T`$. Note also that after the choice we still have the same labels $`l_{2},l_{3},l_{4}`$, however the labeling is still deterministic, similarly to the case [3](#S4.E3 "In Proposition 5 ‣ 4.2 Using non-linear labelings ‣ 4 Expressiveness of the syntactic scheduler ‣ Making Random Choices Invisible to the SchedulerThis work has been partially supported by the INRIA DREI Équipe Associée PRINTEMPS and by the INRIA ARC project ProNoBiS.") of Proposition [5](#Pt0.A1.EGx2 "Proposition 5 ‣ 4.2 Using non-linear labelings ‣ 4 Expressiveness of the syntactic scheduler ‣ Making Random Choices Invisible to the SchedulerThis work has been partially supported by the INRIA DREI Équipe Associée PRINTEMPS and by the INRIA ARC project ProNoBiS.").
In case of a nondeterministic selection of the culprit, and a probabilistic anonymity protocol, the notion of strong probabilistic anonymity has not been established yet, although some possible definitions have been discussed in \[[22](#bib.bib22)\]. Our framework makes it possible to give a natural and precise definition.
###### Definition 3 (Strong probabilistic anonymity for nondeterministic selection of the culprit)
A protocol with nondeterministic selection of the culprit satisfies strong probabilistic anonymity iff for all observables $`\overset{\rightarrow}{o}`$, schedulers $`S`$, and independent schedulers $`T_{1},T_{2}`$ which select different culprits, we have: $`{p_{S,T_{1}}\hspace{0pt}{(\overset{\rightarrow}{o})}} = {p_{S,T_{2}}\hspace{0pt}{(\overset{\rightarrow}{o})}}`$.
We can prove the above property for our protocol:
###### Proposition 8
The DCP with nondeterministic selection of the culprit specified in this section satisfies strong probabilistic anonymity.
## 7 Conclusion and Future work
We have proposed a process-calculus approach to the problem of limiting the power of the scheduler so that it does not reveal the outcome of hidden random choices, and we have shown its applications to the specification of information-hiding protocols. We have also discussed a feature, namely the distributivity of certain contexts over random choices, that makes our calculus appealing for verification. Finally, we have considered the probabilistic testing preorders and shown that they are precongruences in our calculus.
Our plans for future work are in two directions: (a) we would like to investigate the possibility of giving a game-theoretic characterization of our notion of scheduler, and (b) we would like to incorporate our ideas in some existing probabilistic model checker, for instance PRISM.
#### Acknowledgments.
We would like to thank Vincent Danos for having pointed out to us an attack to the Dining Cryptographers protocol based on the order of the scheduler, which has inspired this work.
## References
- \[1\] Vardi, M.Y.: Automatic verification of probabilistic concurrent finite-state programs. In: Proc. of the 26th Annual Symp. on Foundations of Computer Science, IEEE Computer Society Press (1985) 327–338
- \[2\] Hansson, H., Jonsson, B.: A framework for reasoning about time and reliability. In: Proc. of the 10th Symposium on Real-Time Systems, IEEE Computer Society Press (1989) 102–111
- \[3\] Yi, W., Larsen, K.G.: Testing probabilistic and nondeterministic processes. In: Proc. of the 12th IFIP International Symposium on Protocol Specification, Testing and Verification, North Holland (1992)
- \[4\] Segala, R.: Modeling and Verification of Randomized Distributed Real-Time Systems. PhD thesis, Department of Electrical Engineering and Computer Science, Massachusetts Institute of Technology (1995) Available as Technical Report MIT/LCS/TR-676.
- \[5\] Segala, R., Lynch, N.: Probabilistic simulations for probabilistic processes.
Nordic Journal of Computing 2 (1995)
- \[6\] Hansson, H., Jonsson, B.: A calculus for communicating systems with time and probabitilies. In: Proc. of the 11th Symposium on Real-Time Systems, IEEE Computer Society Press (1990) 278–287
- \[7\] Bandini, E., Segala, R.: Axiomatizations for probabilistic bisimulation. In: Proc. of the 28th International Colloquium on Automata, Languages and Programming. LNCS 2076, Springer (2001) 370–381
- \[8\] Andova, S.: Probabilistic process algebra. PhD thesis, Technische Universiteit Eindhoven (2002)
- \[9\] Mislove, M., Ouaknine, J., Worrell, J.: Axioms for Probability and Nondeterminism. In: Proc. of the 10th Int. Wksh. on Expressiveness in Concurrency (EXPRESS ’03). Volume 96 of ENTCS, Elsevier (2004)
- \[10\] Palamidessi, C., Herescu, O.M.: A randomized encoding of the $`\pi`$-calculus with mixed choice.
Theoretical Computer Science 335 (2005) 373–404
<a href="http://www.lix.polytechnique.fr/~catuscia/papers/prob_enc/report.pdf" target="_blank">http://www.lix.polytechnique.fr/~catuscia/papers/prob_enc/report.pdf</a>.
- \[11\] Deng, Y., Palamidessi, C., Pang, J.: Compositional reasoning for probabilistic finite-state behaviors. In: Processes, Terms and Cycles: Steps on the Road to Infinity. LNCS 3838. Springer (2005) 309–337
<a href="http://www.lix.polytechnique.fr/~catuscia/papers/Yuxin/BookJW/par.pdf" target="_blank">http://www.lix.polytechnique.fr/~catuscia/papers/Yuxin/BookJW/par.pdf</a>.
- \[12\] Sokolova, A., Vink, E.d.: Probabilistic automata: system types, parallel composition and comparison. In: Validation of Stochastic Systems: A Guide to Current Research. LNCS 2925. Springer (2004) 1–43
- \[13\] Jonsson, B., Larsen, K.G., Yi, W.: Probabilistic extensions of process algebras. In: Handbook of Process Algebras. Elsevier (2001) 685–710
- \[14\] Chatzikokolakis, K., Palamidessi, C.: A framework for analyzing probabilistic protocols and its application to the partial secrets exchange.
Theoretical Computer Science. To appear. A short version of this paper appeared in the Proc. of the Symp. on Trustworthy Global Computing, LNCS 3705, 146-162. Springer, 2005.
<a href="http://www.lix.polytechnique.fr/~catuscia/papers/PartialSecrets/TCSreport.pdf" target="_blank">http://www.lix.polytechnique.fr/~catuscia/papers/PartialSecrets/TCSreport.pdf</a>.
- \[15\] de Alfaro, L., Henzinger, T.A., Jhala, R.: Compositional methods for probabilistic systems. In: Proceedings of CONCUR 2001. LNCS 2154, Springer (2001)
- \[16\] Mitchell, J.C., Ramanathan, A., Scedrov, A., Teague, V.: A probabilistic polynomial-time process calculus for the analysis of cryptographic protocols.
Theoretical Computer Science 353 (2006) 118–164
- \[17\] Canetti, R., Cheung, L., Kaynar, D., Liskov, M., Lynch, N., Pereira, O., Segala, R.: Task-structured probabilistic i/o automata. In: Proc. of the 8th Int. Workshop on Discrete Event Systems (WODES’06), (2006)
- \[18\] Canetti, R., Cheung, L., Kaynar, D.K., Liskov, M., Lynch, N.A., Pereira, O., Segala, R.: Time-bounded task-PIOAs: A framework for analyzing security protocols. In: Proc. of DISC ’06. LNCS 4167, Springer (2006) 238–253
- \[19\] Nicola, R.D., Hennessy, M.C.B.: Testing equivalences for processes.
Theoretical Computer Science 34 (1984) 83–133
- \[20\] Abadi, M., Gordon, A.D.: A calculus for cryptographic protocols: The spi calculus.
Information and Computation 148 (1999) 1–70
- \[21\] Chaum, D.: The dining cryptographers problem: Unconditional sender and recipient untraceability.
Journal of Cryptology 1 (1988) 65–75
- \[22\] Bhargava, M., Palamidessi, C.: Probabilistic anonymity. In: Proc. of CONCUR 2005. LNCS 3653, Springer (2005) 171–185
<a href="http://www.lix.polytechnique.fr/~catuscia/papers/Anonymity/concur.pdf" target="_blank">http://www.lix.polytechnique.fr/~catuscia/papers/Anonymity/concur.pdf</a>.
## Appendix 0.A Proofs
In this appendix we give the proof of the main technical result of our paper.
Theorem [5.1](#Pt0.A1.EGx3 "Theorem 5.1 ‣ 5 Testing relations for CCSσ processes ‣ Making Random Choices Invisible to the SchedulerThis work has been partially supported by the INRIA DREI Équipe Associée PRINTEMPS and by the INRIA ARC project ProNoBiS.") Let $`P,Q`$ be CCS<sub>σ</sub> processes and $`C`$ a context with a fresh labeling and without occurrences of bang. Then
<table>
<tbody>
<tr>
<td></td>
<td>l:(C[l0:τ.P]+pC[l0:τ.Q])\displaystyle l:(C[l_{0}:\tau.P]+_{p}C[l_{0}:\tau.Q])</td>
<td>≈maysubscriptmay\displaystyle\approx_{\textrm{{may}}}</td>
<td>C[l:(P+pQ)]and\displaystyle C[l:(P+_{p}Q)]\quad\textrm{and}</td>
<td></td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td>l:(C[l0:τ.P]+pC[l0:τ.Q])\displaystyle l:(C[l_{0}:\tau.P]+_{p}C[l_{0}:\tau.Q])</td>
<td>≈mustsubscriptmust\displaystyle\approx_{\textrm{{must}}}</td>
<td>C[l:(P+pQ)]\displaystyle C[l:(P+_{p}Q)]</td>
<td></td>
</tr>
</tbody>
</table>
###### Proof
Since we will always use the label $`l`$ for all probabilistic sum $`+_{p}`$, and $`l_{0}`$ for $`\tau.P`$ and $`\tau.Q`$, we will omit these labels to make the proof more readable. We will also denote $`({1 - p})`$ by $`\overline{p}`$.
Let $`R_{1} = C{\lbrack\tau.P\rbrack} +_{p}C{\lbrack\tau.Q\rbrack}`$ and $`R_{2} = {C\hspace{0pt}{\lbrack{P +_{p}Q}\rbrack}}`$. We will prove that for all tests $`O`$ and for all schedulers $`S_{1} \in {S\hspace{0pt}y\hspace{0pt}n\hspace{0pt}{({{(\nu)}\hspace{0pt}{(\left. R_{1} \middle| O \right.)}})}}`$ there exists $`S_{2} \in {S\hspace{0pt}y\hspace{0pt}n\hspace{0pt}{({{(\nu)}\hspace{0pt}{(\left. R_{2} \middle| O \right.)}})}}`$ such that $`{p_{\omega}\hspace{0pt}{(R_{1},S_{1},O)}} = {p_{\omega}\hspace{0pt}{(R_{2},S_{2},O)}}`$ and vice versa. This implies both $`R_{1} \approx_{\text{may}}R_{2}`$ and $`R_{1} \approx_{\text{must}}R_{2}`$.
Without loss of generality we assume that tests do not perform internal actions, but only synchronizations with the tested process. First, it is easy to see that
<table>
<tbody>
<tr>
<td></td>
<td>pω(P+pQ,σ(l).S,O)\displaystyle p_{\omega}(P+_{p}Q,\sigma(l).S,O)</td>
<td>=\displaystyle=</td>
<td>ppω(P,S,O)+p¯pω(Q,S,O)𝑝subscript𝑝𝜔𝑃𝑆𝑂¯𝑝subscript𝑝𝜔𝑄𝑆𝑂\displaystyle p\ p_{\omega}(P,S,O)+\bar{p}\ p_{\omega}(Q,S,O)</td>
<td></td>
<td>(9)</td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td>pω(l1:a.P,σ(l1,l2).S,O)\displaystyle p_{\omega}(l_{1}:a.P,\sigma(l_{1},l_{2}).S,O)</td>
<td>=\displaystyle=</td>
<td>pω(P,S,O′)subscript𝑝𝜔𝑃𝑆superscript𝑂′\displaystyle p_{\omega}(P,S,O^{\prime})</td>
<td></td>
<td>(10)</td>
</tr>
</tbody>
</table>
where $`{(\nu)}{(l_{1}:a.P|O)} \parallel \sigma{(l_{1},l_{2})}.S\overset{\tau}{\longrightarrow}\delta{({(\nu)}{(P|O' \parallel S)})}`$.
In order for the scheduler of $`R_{1}`$ to be non-blocking, it has to be of the form $`{\sigma\hspace{0pt}{(l)}}.S_{1}`$, since the only possible transition of $`R_{1}`$ is the probabilistic choice labeled by $`l`$. By ([9](#Pt0.A1.E9 "In Proof ‣ Appendix 0.A Proofs ‣ Making Random Choices Invisible to the SchedulerThis work has been partially supported by the INRIA DREI Équipe Associée PRINTEMPS and by the INRIA ARC project ProNoBiS.")) we have
<table>
<tbody>
<tr>
<td></td>
<td>pω(C[τ.P]+C[τ.Q],σ(l).S1,O)=ppω(C[τ.P],S1,O)+p¯pω(C[τ.Q],S1,O)p_{\omega}(C[\tau.P]+C[\tau.Q],\sigma(l).S_{1},O)=p\ p_{\omega}(C[\tau.P],S_{1},O)+\bar{p}\ p_{\omega}(C[\tau.Q],S_{1},O)</td>
<td></td>
</tr>
</tbody>
</table>
The proof will be by induction on the structure of $`C`$. Let $`O`$ range over tests with fresh labelings, let $`S_{1}`$ range over nonblocking schedulers for both $`C{\lbrack\tau.P\rbrack}`$ and $`C{\lbrack\tau.Q\rbrack}`$ (such that $`{\sigma\hspace{0pt}{(l)}}.S_{1}`$ is a nonblocking scheduler for $`R_{1}`$) and let $`S_{2}`$ range over nonblocking schedulers for $`R_{2}`$. The induction hypothesis is:
<table>
<tbody>
<tr>
<td></td>
<td>⇒)∀O∀S1∃S2:ppω(C[τ.P],S1,O)+p¯pω(C[τ.Q],S1,O)=pω(C[P+pQ],S2,O)and⇐)∀O∀S2∃S1:ppω(C[τ.P],S1,O)+p¯pω(C[τ.Q],S1,O)=pω(C[P+pQ],S2,O)\begin{array}[]{l}\Rightarrow)\ \forall O\ \forall S_{1}\ \exists S_{2}:\\ \qquad p\ p_{\omega}(C[\tau.P],S_{1},O)+\bar{p}\ p_{\omega}(C[\tau.Q],S_{1},O)=p_{\omega}(C[P+_{p}Q],S_{2},O)\quad\textrm{and}\\ \Leftarrow)\ \forall O\ \forall S_{2}\ \exists S_{1}:\\ \qquad p\ p_{\omega}(C[\tau.P],S_{1},O)+\bar{p}\ p_{\omega}(C[\tau.Q],S_{1},O)=p_{\omega}(C[P+_{p}Q],S_{2},O)\end{array}</td>
<td></td>
</tr>
</tbody>
</table>
We have the following cases for $`C`$:
- •
Case $`C = {\lbrack\rbrack}`$. Trivial.
- •
Case $`C = l_{1}:a.C'`$
The scheduler $`S_{1}`$ of $`C{\lbrack\tau.P\rbrack}`$ and $`C{\lbrack\tau.Q\rbrack}`$ has to be of the form $`{S_{1} = {\sigma\hspace{0pt}{(l_{1},l_{2})}}}.S_{1}'`$ where $`l_{2}`$ is the label of a $`\overline{a}`$ prefix in $`O`$ (if no such prefix exists then the case is trivial).
A scheduler of the form $`{\sigma\hspace{0pt}{(l_{1},l_{2})}}.S`$ can schedule any process of the form $`l_{1}:a.X`$ (with label $`l_{1}`$) giving the transition:
<table>
<tbody>
<tr>
<td></td>
<td>(ν)(l1:a.X|O)∥σ(l1,l2).S⟶𝜏δ((ν)(X|O′)∥S)(\nu)(l_{1}:a.X\;|\;O)\parallel\sigma(l_{1},l_{2}).S\overset{\tau}{\longrightarrow}\delta((\nu)(X\;|\;O^{\prime})\parallel S)</td>
<td></td>
</tr>
</tbody>
</table>
and producing always the same $`O'`$. The probability $`p_{\omega}`$ for these processes will be given by equation ([10](#Pt0.A1.E10 "In Proof ‣ Appendix 0.A Proofs ‣ Making Random Choices Invisible to the SchedulerThis work has been partially supported by the INRIA DREI Équipe Associée PRINTEMPS and by the INRIA ARC project ProNoBiS.")).
Thus for ($`\Rightarrow`$) we have
<table>
<tbody>
<tr>
<td></td>
<td>ppω(l1:a.C[τ.P],σ(l1,l2).S1′,O)+p¯pω(l1:a.C[τ.Q],σ(l1,l2).S1′,O) =ppω(C′[τ.P],S1′,O′)+p¯pω(C′[τ.Q],S1′,O′) (10) =pω(C′[P+pQ],S2′,O′) Ind. Hyp. =pω(l1:a.C′[P+pQ],σ(l1,l2).S2′,O) (10) =pω(R2,S2,O) \begin{array}[]{r @{\textrm{\hspace{5pt}}} cl @{\textrm{\hspace{5pt}}} l}\lx@intercol p\ p_{\omega}(l_{1}:a.C[\tau.P],\sigma(l_{1},l_{2}).S_{1}^{\prime},O)+\bar{p}\ p_{\omega}(l_{1}:a.C[\tau.Q],\sigma(l_{1},l_{2}).S_{1}^{\prime},O)\hfil\lx@intercol\\[3.0pt] \textrm{\hskip 5.0&=&p\ p_{\omega}(C^{\prime}[\tau.P],S_{1}^{\prime},O^{\prime})+\bar{p}\ p_{\omega}(C^{\prime}[\tau.Q],S_{1}^{\prime},O^{\prime})\hfil\textrm{\hskip 5.0&\textrm{(\ref{eq:pom2})}\\[3.0pt] \textrm{\hskip 5.0&=&p_{\omega}(C^{\prime}[P+_{p}Q],S_{2}^{\prime},O^{\prime})\hfil\textrm{\hskip 5.0&\textrm{Ind. Hyp.}\\[3.0pt] \textrm{\hskip 5.0&=&p_{\omega}(l_{1}:a.C^{\prime}[P+_{p}Q],\sigma(l_{1},l_{2}).S_{2}^{\prime},O)\hfil\textrm{\hskip 5.0&\textrm{(\ref{eq:pom2})}\\[3.0pt] \textrm{\hskip 5.0&=&p_{\omega}(R_{2},S_{2},O)\hfil\textrm{\hskip 5.0\end{array}}}}}}}}}</td>
<td></td>
</tr>
</tbody>
</table>
For ($`\Leftarrow`$) we can perform the above derivation in the opposite direction, given that a scheduler for $`R_{2} = l_{1}:a.C'{\lbrack P +_{p}Q\rbrack}`$ must be of the form $`{S_{2} = {\sigma\hspace{0pt}{(l_{1},l_{2})}}}.S_{2}'`$.
- •
Case $`C = \left. C' \middle| R \right.`$
Since we only consider contexts with fresh labelings, $`\left. R \middle| O \right.`$ is itself a test, and
<table>
<tbody>
<tr>
<td></td>
<td>pω(X|R,S,O)=pω(X,S,R|O)subscript𝑝𝜔conditional𝑋𝑅𝑆𝑂subscript𝑝𝜔𝑋𝑆conditional𝑅𝑂p_{\omega}(X\;|\;R,S,O)=p_{\omega}(X,S,R\;|\;O)</td>
<td></td>
<td>(11)</td>
</tr>
</tbody>
</table>
Thus for ($`\Rightarrow`$) we have
<table>
<tbody>
<tr>
<td></td>
<td>ppω(C′[τ.P]|R,S1,O)+p¯pω(C′[τ.Q]|R,S1,O) =ppω(C′[τ.P],S1,R|O)+p¯pω(C′[τ.Q],S1,R|O) (11) =pω(C′[P+pQ],S2,R|O) Ind. Hyp. =pω(C′[P+pQ]|R,S2,O) (11) =pω(R2,S2,O) \begin{array}[]{r @{\textrm{\hspace{5pt}}} cl @{\textrm{\hspace{5pt}}} l}\lx@intercol p\ p_{\omega}(C^{\prime}[\tau.P]\;|\;R,S_{1},O)+\bar{p}\ p_{\omega}(C^{\prime}[\tau.Q]\;|\;R,S_{1},O)\hfil\lx@intercol\\[3.0pt] \textrm{\hskip 5.0&=&p\ p_{\omega}(C^{\prime}[\tau.P],S_{1},R\;|\;O)+\bar{p}\ p_{\omega}(C^{\prime}[\tau.Q],S_{1},R\;|\;O)\hfil\textrm{\hskip 5.0&\textrm{(\ref{eq3})}\\[3.0pt] \textrm{\hskip 5.0&=&p_{\omega}(C^{\prime}[P+_{p}Q],S_{2},R\;|\;O)\hfil\textrm{\hskip 5.0&\textrm{Ind. Hyp.}\\[3.0pt] \textrm{\hskip 5.0&=&p_{\omega}(C^{\prime}[P+_{p}Q]\;|\;R,S_{2},O)\hfil\textrm{\hskip 5.0&\textrm{(\ref{eq3})}\\[3.0pt] \textrm{\hskip 5.0&=&p_{\omega}(R_{2},S_{2},O)\hfil\textrm{\hskip 5.0\end{array}}}}}}}}}</td>
<td></td>
</tr>
</tbody>
</table>
For ($`\Leftarrow`$) we can perform the above derivation in the opposite direction.
- •
Case $`{C = l_{1}}:{({C' +_{q}R})}`$
Since we consider only contexts with fresh labelings, the labels of $`C'`$ are disjoint from those of $`R`$, thus the scheduler of a process of the form $`l_{1}:{({{C'\hspace{0pt}{\lbrack X\rbrack}} +_{q}R})}`$ must be of the form $`{S = {\sigma\hspace{0pt}{(l_{1})}}}.{({S_{C} + S_{R}})}`$ where $`S_{C}`$ is a scheduler containing labels of $`C'\hspace{0pt}{\lbrack X\rbrack}`$ and $`S_{R}`$ a scheduler containing labels of $`R`$. Moreover
<table>
<tbody>
<tr>
<td></td>
<td colspan="3">pω(l1:(C′[X]+qR),S,O)\displaystyle p_{\omega}(l_{1}:(C^{\prime}[X]+_{q}R),S,O)</td>
<td></td>
<td rowspan="3">(12)</td>
</tr>
<tr>
<td></td>
<td></td>
<td>=\displaystyle=</td>
<td>qpω(C′[X],SC+SR,O)+q¯pω(R,SC+SR,O)𝑞subscript𝑝𝜔superscript𝐶′delimited-[]𝑋subscript𝑆𝐶subscript𝑆𝑅𝑂¯𝑞subscript𝑝𝜔𝑅subscript𝑆𝐶subscript𝑆𝑅𝑂\displaystyle q\ p_{\omega}(C^{\prime}[X],S_{C}+S_{R},O)+\bar{q}\ p_{\omega}(R,S_{C}+S_{R},O)</td>
<td></td>
</tr>
<tr>
<td></td>
<td></td>
<td>=\displaystyle=</td>
<td>qpω(C′[X],SC,O)+q¯pω(R,SR,O)𝑞subscript𝑝𝜔superscript𝐶′delimited-[]𝑋subscript𝑆𝐶𝑂¯𝑞subscript𝑝𝜔𝑅subscript𝑆𝑅𝑂\displaystyle q\ p_{\omega}(C^{\prime}[X],S_{C},O)+\bar{q}\ p_{\omega}(R,S_{R},O)</td>
<td></td>
</tr>
</tbody>
</table>
As a consequence, the scheduler $`S_{1}`$ of $`C{\lbrack\tau.P\rbrack}`$ and $`C{\lbrack\tau.Q\rbrack}`$ has to be of the form $`{S_{1} = {\sigma\hspace{0pt}{(l_{1})}}}.{({S_{C} + S_{R}})}`$.
For ($`\Rightarrow`$) we have
<table>
<tbody>
<tr>
<td></td>
<td>ppω(l1:(C′[τ.P]+qR),S1,O)+p¯pω(l1:(C′[τ.Q]+qR),S1,O) =q(ppω(C′[τ.P],SC,O)+p¯pω(C′[τ.Q],SC,O))+ q¯pω(R,SR,O) (12) =qpω(C′[P+pQ]),SC′,O)+ q¯pω(R,SR,O) Ind. Hyp. =pω(l1:(C′[P+pQ]+qR),σ(l1).(SC′+SR),O) (12) =pω(R2,S2,O) \begin{array}[]{r @{\textrm{\hspace{5pt}}} cl @{\textrm{\hspace{5pt}}} l}\lx@intercol p\ p_{\omega}(l_{1}:(C^{\prime}[\tau.P]+_{q}R),S_{1},O)+\bar{p}\ p_{\omega}(l_{1}:(C^{\prime}[\tau.Q]+_{q}R),S_{1},O)\hfil\lx@intercol\\[3.0pt] \textrm{\hskip 5.0&=&q(p\ p_{\omega}(C^{\prime}[\tau.P],S_{C},O)+\bar{p}\ p_{\omega}(C^{\prime}[\tau.Q],S_{C},O))+\hfil\textrm{\hskip 5.0\\ \textrm{\hskip 5.0&&\bar{q}\ p_{\omega}(R,S_{R},O)\hfil\textrm{\hskip 5.0&\textrm{(\ref{eq:pom3})}\\ \textrm{\hskip 5.0&=&q\ p_{\omega}(C^{\prime}[P+_{p}Q]),S_{C}^{\prime},O)+\hfil\textrm{\hskip 5.0\\ \textrm{\hskip 5.0&&\bar{q}\ p_{\omega}(R,S_{R},O)\hfil\textrm{\hskip 5.0&\textrm{Ind. Hyp.}\\ \textrm{\hskip 5.0&=&p_{\omega}(l_{1}:(C^{\prime}[P+_{p}Q]+_{q}R),\sigma(l_{1}).(S_{C}^{\prime}+S_{R}),O)\hfil\textrm{\hskip 5.0&\textrm{(\ref{eq:pom3})}\\ \textrm{\hskip 5.0&=&p_{\omega}(R_{2},S_{2},O)\hfil\textrm{\hskip 5.0\end{array}}}}}}}}}}}}}</td>
<td></td>
</tr>
</tbody>
</table>
For ($`\Leftarrow`$) we can perform the above derivation in the opposite direction, given that a scheduler for $`{R_{2} = l_{1}}:{({{C'\hspace{0pt}{\lbrack{P +_{p}Q}\rbrack}} +_{q}R})}`$ must be of the form $`{S_{2} = {\sigma\hspace{0pt}{(l_{1})}}}.{({S_{C}' + S_{R}})}`$.
- •
Case $`C = {C' + R}`$
Consider the process $`C'{\lbrack l_{0}:\tau.P\rbrack} + R`$. The scheduler $`S_{1}`$ of this process has to choose between $`C'{\lbrack l_{0}:\tau.P\rbrack}`$ and $`R`$.
There are two cases to have a transition using the SUM1, SUM2 rules.
- i\)
Either $`{(\nu)}{(C'{\lbrack l_{0}:\tau.P\rbrack} + R|O)} \parallel S_{R}\overset{\alpha}{\longrightarrow}\mu`$ such that $`{{(\nu)}\hspace{0pt}{(\left. R \middle| O \right.)}} \parallel {S_{R}\hspace{0pt}\overset{\alpha}{\longrightarrow}\hspace{0pt}\mu}`$. In this case
<table>
<tbody>
<tr>
<td></td>
<td>pω(C′[l0:τ.P]+R,SR,O)=pω(R,SR,O)p_{\omega}(C^{\prime}[l_{0}:\tau.P]+R,S_{R},O)=p_{\omega}(R,S_{R},O)</td>
<td></td>
<td>(13)</td>
</tr>
</tbody>
</table>
- ii\)
Or $`{(\nu)}{(C'{\lbrack l_{0}:\tau.P\rbrack} + R|O)} \parallel S_{C}\overset{\alpha}{\longrightarrow}\mu`$ such that $`{(\nu)}{(C'{\lbrack l_{0}:\tau.P\rbrack}|O)} \parallel S_{C}\overset{\alpha}{\longrightarrow}\mu`$. In this case
<table>
<tbody>
<tr>
<td></td>
<td>pω(C′[l0:τ.P]+R,SC,O)=pω(C′[l0:τ.P],SC,O)p_{\omega}(C^{\prime}[l_{0}:\tau.P]+R,S_{C},O)=p_{\omega}(C^{\prime}[l_{0}:\tau.P],S_{C},O)</td>
<td></td>
<td>(14)</td>
</tr>
</tbody>
</table>
Now consider the process $`C'{\lbrack l_{0}:\tau.Q\rbrack} + R`$. Since $`P`$ and $`Q`$ are behind the $`l_{1}:\tau`$ action, this process has exactly the same visible labels as $`C'{\lbrack l_{0}:\tau.P\rbrack} + R`$. Thus $`S_{R}`$ and $`S_{C}`$ will select $`R`$ and $`C'{\lbrack l_{0}:\tau.Q\rbrack}`$ respectively and the equations ([13](#Pt0.A1.E13 "In item i) ‣ 5th item ‣ Proof ‣ Appendix 0.A Proofs ‣ Making Random Choices Invisible to the SchedulerThis work has been partially supported by the INRIA DREI Équipe Associée PRINTEMPS and by the INRIA ARC project ProNoBiS.")) and ([14](#Pt0.A1.E14 "In item ii) ‣ 5th item ‣ Proof ‣ Appendix 0.A Proofs ‣ Making Random Choices Invisible to the SchedulerThis work has been partially supported by the INRIA DREI Équipe Associée PRINTEMPS and by the INRIA ARC project ProNoBiS.")) will hold.
In the case (i) ($`S = S_{R}`$) we have:
<table>
<tbody>
<tr>
<td></td>
<td>ppω(C′[τ.P]+R,SR,O)+p¯pω(C′[τ.Q]+R),SR,O) =ppω(R,SR,O)+p¯pω(R,SR,O) (13) =pω(R,SR,O) =pω(C′[P+pQ]+R,SR,O) =pω(R2,S2,O) \begin{array}[]{r @{\textrm{\hspace{5pt}}} cl @{\textrm{\hspace{5pt}}} l}\lx@intercol p\ p_{\omega}(C^{\prime}[\tau.P]+R,S_{R},O)+\bar{p}\ p_{\omega}(C^{\prime}[\tau.Q]+R),S_{R},O)\hfil\lx@intercol\\[3.0pt] \textrm{\hskip 5.0&=&p\ p_{\omega}(R,S_{R},O)+\bar{p}\ p_{\omega}(R,S_{R},O)\hfil\textrm{\hskip 5.0&\textrm{(\ref{eq4})}\\[3.0pt] \textrm{\hskip 5.0&=&p_{\omega}(R,S_{R},O)\hfil\textrm{\hskip 5.0\\[3.0pt] \textrm{\hskip 5.0&=&p_{\omega}(C^{\prime}[P+_{p}Q]+R,S_{R},O)\hfil\textrm{\hskip 5.0\\[3.0pt] \textrm{\hskip 5.0&=&p_{\omega}(R_{2},S_{2},O)\hfil\textrm{\hskip 5.0\end{array}}}}}}}}}</td>
<td></td>
</tr>
</tbody>
</table>
In the case (ii) ($`S = S_{C}`$) we have:
<table>
<tbody>
<tr>
<td></td>
<td>ppω(C′[τ.P]+R,SC,O)+p¯pω(C′[τ.Q]+R),SC,O) =ppω(C′[τ.P],SC,O)+p¯pω(C′[τ.Q],SC,O) (14) =pω(C′[P+pQ],SC′,O) Ind. Hyp. =pω(C′[P+pQ]+R,SC′,O) =pω(R2,S2,O) \begin{array}[]{r @{\textrm{\hspace{5pt}}} cl @{\textrm{\hspace{5pt}}} l}\lx@intercol p\ p_{\omega}(C^{\prime}[\tau.P]+R,S_{C},O)+\bar{p}\ p_{\omega}(C^{\prime}[\tau.Q]+R),S_{C},O)\hfil\lx@intercol\\[3.0pt] \textrm{\hskip 5.0&=&p\ p_{\omega}(C^{\prime}[\tau.P],S_{C},O)+\bar{p}\ p_{\omega}(C^{\prime}[\tau.Q],S_{C},O)\hfil\textrm{\hskip 5.0&\textrm{(\ref{eq5})}\\[3.0pt] \textrm{\hskip 5.0&=&p_{\omega}(C^{\prime}[P+_{p}Q],S_{C}^{\prime},O)\hfil\textrm{\hskip 5.0&\textrm{Ind. Hyp.}\\[3.0pt] \textrm{\hskip 5.0&=&p_{\omega}(C^{\prime}[P+_{p}Q]+R,S_{C}^{\prime},O)\hfil\textrm{\hskip 5.0\\[3.0pt] \textrm{\hskip 5.0&=&p_{\omega}(R_{2},S_{2},O)\hfil\textrm{\hskip 5.0\end{array}}}}}}}}}</td>
<td></td>
</tr>
</tbody>
</table>
For ($`\Leftarrow`$) we can perform the above derivation in the opposite direction.
- •
Case $`C = {{({\nu\hspace{0pt}a})}\hspace{0pt}C'}`$
The process $`{(\nu)}\hspace{0pt}{(\left. {{({\nu\hspace{0pt}a})}\hspace{0pt}C'\hspace{0pt}{\lbrack X\rbrack}} \middle| O \right.)}`$ has the same transitions as $`{(\nu)}\hspace{0pt}{(\left. {C'\hspace{0pt}{\lbrack X\rbrack}} \middle| {{({\nu\hspace{0pt}a})}\hspace{0pt}O} \right.)}`$. The result follows by the induction hypothesis.
$`\square`$ |
0706.2520 | cs.NI | cyber | # Analysis of Inter-Domain Traffic Correlations: Random Matrix Theory Approach
Viktoria Rojkova, Mehmed Kantardzic
###### Abstract
The traffic behavior of University of Louisville network with the interconnected backbone routers and the number of Virtual Local Area Network (VLAN) subnets is investigated using the Random Matrix Theory (RMT) approach. We employ the system of equal interval time series of traffic counts at all router to router and router to subnet connections as a representation of the inter-VLAN traffic. The cross-correlation matrix $`C`$ of the traffic rate changes between different traffic time series is calculated and tested against null-hypothesis of random interactions.
The majority of the eigenvalues $`\lambda_{i}`$ of matrix $`C`$ fall within the bounds predicted by the RMT for the eigenvalues of random correlation matrices. The distribution of eigenvalues and eigenvectors outside of the RMT bounds displays prominent and systematic deviations from the RMT predictions. Moreover, these deviations are stable in time.
The method we use provides a unique possibility to accomplish three concurrent tasks of traffic analysis. The method verifies the uncongested state of the network, by establishing the profile of random interactions. It recognizes the system-specific large-scale interactions, by establishing the profile of stable in time non-random interactions. Finally, by looking into the eigenstatistics we are able to detect and allocate anomalies of network traffic interactions.
###### Index Terms:
Network-Wide Traffic Analysis, Random Matrix Theory, Large-Scale Correlations
Department of Computer Engineering and Computer Science, University of Louisville, Louisville, KY 40292 email: {vbrozh01, mmkant01}@gwise.louisville.edu
## Categories and Subject Descriptors
C.2.3 \[Computer-Communication Networks\]: Network Operations
## General Terms
Measurement, Experimentation
## I introduction
The infrastructure, applications and protocols of the system of communicating computers and networks are constantly evolving. The traffic, which is an essence of the communication, presently is a voluminous data generated on minute-by-minute basis within multi-layered structure by different applications and according to different protocols. As a consequence, there are two general approaches in analysis of the traffic and in modeling of its healthy behavior. In the first approach, the traffic analysis considers the protocols, applications, traffic matrix and routing matrix estimates, independence of ingress and egress points and much more. The second approach treats the infrastructure between the points from which the traffic is obtained as a “black box” \[[33](#bib.bib33), [34](#bib.bib34)\].
Measuring interactions between logically and architecturally equivalent substructures of the system is a natural extension of the “black box” approach. Certain amount of work in this direction has already been done. Studies on statistical traffic flow properties revealed the “congested”, “fluid” and “transitional” regimes of the flow at a large scale \[[1](#bib.bib1), [2](#bib.bib2)\]. The observed collective behavior suggests the existence of the large-scale network-wide correlations between the network subparts. Indeed, the \[[3](#bib.bib3)\] work showed the large-scale cross-correlations between different connections of the Renater scientific network. Moreover, the analysis of correlations across all simultaneous network-wide traffic has been used in network distributed attacks detection \[[4](#bib.bib4)\].
The distributions and stability of established interactions statistics represent the characteristic features of the system and may be exploited in healthy network traffic profile creation, which is an essential part of network anomaly detection. As it is successfully demonstrated in \[[5](#bib.bib5)\], all tested traffic anomalies change the distribution of the traffic features.
Among numerous types of traffic monitoring variables, time series of traffic counts are free of applications “semantics” and thus more preferable for “black box” analysis. To extract the meaningful information about underlying interactions contained in time series, the empirical correlation matrix is a usual tool at hand. In addition, there are various classes of statistical tools, such as principal component analysis, singular value decomposition, and factor analysis, which in turn strongly rely on the validity of the correlation matrix and obtain the meaningful part of the time series. Thus, it is important to understand quantitatively the effect of noise, i.e. to separate the noisy, random interactions from meaningful ones. In addition, it is crucial to consider the finiteness of the time series in the determination of the empirical correlation, since the finite length of time series available to estimate cross correlations introduces “measurement noise” \[[19](#bib.bib19)\]. Statistically, it is also advisable to develop null-hypothesis tests in order to check the degree of statistical validity of the results obtained against cases of purely random interactions.
The methodology of random matrix theory (RMT) developed for studying the complex energy levels of heavy nuclei and is given a detailed account in \[[6](#bib.bib6), [7](#bib.bib7), [8](#bib.bib8), [9](#bib.bib9), [10](#bib.bib10), [11](#bib.bib11)\]. For our purposes this methodology comes in as a series of statistical tests run on the eigenvalues and eigenvectors of “system matrix”, which in our case is traffic time series cross-correlation matrix $`C`$ (and is Hamiltonian matrix in case of nuclei and other RMT systems \[[6](#bib.bib6), [7](#bib.bib7), [8](#bib.bib8), [9](#bib.bib9), [10](#bib.bib10), [11](#bib.bib11)\]).
In our study, we propose to investigate the network traffic as a complex system with a certain degree of mutual interactions of its constituents, i.e. single-link traffic time series, using the RMT approach. We concentrate on the large scale correlations between the time series generated by Simple Network Manage Protocol (SNMP) traffic counters at every router-router and router-VLAN subnet connection of University of Louisville backbone routers system.
The contributions of this study are as follows:
- •
We propose the application constraints free methodology of network-wide traffic time series interactions analysis. Even though in this particular study, we know in advance that VLANs represent separate broadcast domains, VLAN-router incoming traffic is a traffic intended for other VLANs and VLAN-router outgoing traffic is a routed traffic from other VLANs. Nevertheless, this information is irrelevant for our analysis and acquired only at the interpretation of the analysis results.
- •
Using the RMT, we are able to separate the random interactions from system specific interactions. The vast majority of traffic time series interact in random fashion. The time stable random interactions signify the healthy, and free of congestion traffic. The proposed analysis of eigenvector distribution allows to verify the time series content of uncongested traffic.
- •
The time stable non-random interactions provide us with information about large-scale system-specific interactions.
- •
Finally, the temporal changes in random and non-random interactions can be detected and allocated with eigenvalues and eigenvectors statistics of interactions.
The organization of this paper is as follows. Section II presents the survey of related work. We describe the RMT methodology in Section III. Section IV contains the explanation of the data analyzed. In Section V we test the eigenvalue distribution of inter-VLAN traffic time series cross-correlation matrix C against the RMT predictions. In Section VI we analyze the content of inter-VLAN traffic interactions by mean of eigenvalues and eigenvectors deviated from RMT. Section VII discusses the characteristic traffic interactions parameters of the system such as time stability of the deviating eigenvalues and eigenvectors, inverse participation ratio (IPR) of eigenvalues spectra, localization points in IPR plot, overlap matrices of the deviating eigenvectors. With series of different experiments, we demonstrate how traffic interactions anomalies can be detected and allocated in time and space using various visualization techniques on eigenvalues and eigenvectors statistics in Section VIII. We present our conclusions and prospective research steps in Section IX.
## II related work
Few works investigate the interactions of traffic time series regardless of underlying architecture of the traffic system. As it was stated in Introduction, the study of \[[3](#bib.bib3)\] showed the large-scale cross-correlations between different connections of the French scientific network Renater with 26 interconnected routers and 650 connections links. The random interactions between traffic time series of complex traffic system without the routing protocol information were established by Krbalek and Seba in \[[12](#bib.bib12)\] for transportation system in Cuernavaca (Mexico).
The urgent need for a network-wide, scalable approach to the problem of healthy network traffic profile creation is expressed in works of \[[5](#bib.bib5), [14](#bib.bib14), [15](#bib.bib15), [13](#bib.bib13), [16](#bib.bib16), [17](#bib.bib17)\]. There are several studies with the promising results, which demonstrate that the traffic anomalous events cause the temporal changes in statistical properties of traffic features. Lakhina, Crovella and Diot presented the characterization of the network-wide anomalies of the traffic flows. The authors studied three different types of traffic flows and fused the information from flow measurements taken throughout the entire network. They obtained and classified a different set of anomalies for different traffic types using the subspace method \[[14](#bib.bib14)\].
The same group of researchers extended their work in \[[5](#bib.bib5)\]. Under the new assumption that any network anomaly induces the changes in distributional aspects of packet header fields, they detected and identified large set of anomalies using the entropy measurement tool.
Hidden Markov model has been proposed to model the distribution of network-wide traffic in \[[15](#bib.bib15)\]. The observation window is used to distinguish denial of service (DoS) flooding attack mixed with the normal background traffic.
Roughan et al. combined the entire network routing and traffic data to detect the IP forwarding anomalies \[[16](#bib.bib16)\].
Huang et al., \[[17](#bib.bib17)\] used the distributed version of the Principal Component Analysis (PCA) method for centralized network-wide volume anomaly detection. A key ingredient of their framework is an analytical method based on stochastic matrix perturbation theory that balances between the accuracy of the approximate network anomaly detection and the amount of data communication over the network.
The authors of \[[13](#bib.bib13)\] found the high temporal correlation (frequently \> 0.99) between flow counts on quiescent ports (TCP/IP ports which are not in regular use) at the one of the known pre-attack, so called *reconnaissance*, anomalous behavior, vertical scan.
## III rmt methodology
The RMT was employed in the financial studies of stock correlations \[[18](#bib.bib18), [19](#bib.bib19)\], communication theory of wireless systems \[[20](#bib.bib20)\], array signal processing \[[21](#bib.bib21)\], bioinformatics studies of protein folding \[[22](#bib.bib22)\]. We are not aware of any work, except for \[[3](#bib.bib3)\], where RMT techniques were applied to the Internet traffic system.
We adopt the methodology used in works on financial time series correlations (see \[[18](#bib.bib18), [19](#bib.bib19)\] and references therein) and later in \[[3](#bib.bib3)\], which discusses cross-correlations in Internet traffic. In particular, we quantify correlations between $`N`$ traffic counts time series of $`L`$ time points, by calculating the traffic rate change of every time series $`T`$ $`i = {1,\ldots,N}`$ , over a time scale $`\Delta\hspace{0pt}t`$,
<table>
<tbody>
<tr>
<td></td>
<td>Gi(t)≡lnTi(t+Δt)−lnTi(t)subscript𝐺𝑖𝑡lnsubscript𝑇𝑖𝑡Δ𝑡lnsubscript𝑇𝑖𝑡G_{i}\left(t\right)\equiv\textrm{ln}\,T_{i}\left(t+\Delta t\right)-\textrm{ln}\,T_{i}\left(t\right)</td>
<td></td>
<td>(1)</td>
</tr>
</tbody>
</table>
where $`T{{}_{i}^{}(t)}`$ denotes the traffic rate of time series $`i`$. This measure is independent from the volume of the traffic exchange and allows capturing the subtle changes in the traffic rate \[[3](#bib.bib3)\]. The normalized traffic rate change is
<table>
<tbody>
<tr>
<td></td>
<td>gi(t)≡Gi(t)−⟨Gi(t)⟩σisubscript𝑔𝑖𝑡subscript𝐺𝑖𝑡delimited-⟨⟩subscript𝐺𝑖𝑡subscript𝜎𝑖g_{i}\left(t\right)\equiv\frac{G_{i}\left(t\right)-\left\langle G_{i}\left(t\right)\right\rangle}{\sigma_{i}}</td>
<td></td>
<td>(2)</td>
</tr>
</tbody>
</table>
where $`\sigma_{i} \equiv \sqrt{\left\langle G_{i}^{2} \right\rangle - \left\langle G_{i} \right\rangle^{2}}`$ is the standard deviation of $`G_{i}`$. The equal-time cross-correlation matrix $`C`$ can be computed as follows
<table>
<tbody>
<tr>
<td></td>
<td>Cij≡⟨gi(t)gj(t)⟩subscript𝐶𝑖𝑗delimited-⟨⟩subscript𝑔𝑖𝑡subscript𝑔𝑗𝑡C_{ij}\equiv\left\langle g_{i}\left(t\right)g_{j}\left(t\right)\right\rangle</td>
<td></td>
<td>(3)</td>
</tr>
</tbody>
</table>
The properties of the traffic interactions matrix $`C`$ have to be compared with those of a random cross-correlation matrix \[[23](#bib.bib23)\]. In matrix notation, the interaction matrix $`C`$ can be expressed as
<table>
<tbody>
<tr>
<td></td>
<td>C=1LGGT,𝐶1𝐿𝐺superscript𝐺𝑇C=\frac{1}{L}GG^{T},</td>
<td></td>
<td>(4)</td>
</tr>
</tbody>
</table>
where $`G`$ is $`N \times L`$ matrix with elements $`\left\{ g_{i\hspace{0pt}m} \equiv g_{i}(m \bigtriangleup t); \right.`$ $`{i = {1,\ldots,N}};`$ $`\left. m = 0,\ldots,L - 1 \right\},`$ and $`G^{T}`$ denotes the transpose of $`G`$. Just as was done in \[[19](#bib.bib19)\], we consider a random correlation matrix
<table>
<tbody>
<tr>
<td></td>
<td>R=1LAAT,𝑅1𝐿𝐴superscript𝐴𝑇R=\frac{1}{L}AA^{T},</td>
<td></td>
<td>(5)</td>
</tr>
</tbody>
</table>
where $`A`$ is $`N \times L`$ matrix containing $`N`$ time series of $`L`$ random elements $`a_{i\hspace{0pt}m}`$ with zero mean and unit variance, which are mutually uncorrelated as a null hypothesis.
Statistical properties of the random matrices $`R`$ have been known for years in physics literature \[[6](#bib.bib6), [10](#bib.bib10), [7](#bib.bib7), [8](#bib.bib8), [9](#bib.bib9), [11](#bib.bib11)\]. In particular, it was shown analytically \[[24](#bib.bib24)\] that, under the restriction of $`{N\rightarrow\infty},`$ $`L\rightarrow\infty`$ and providing that $`Q \equiv {L/N}`$$`\left( {> 1} \right)`$ is fixed, the probability density function $`P_{r\hspace{0pt}m}\hspace{0pt}(\lambda)`$ of eigenvalues $`\lambda`$ of the random matrix $`R`$ is given by
<table>
<tbody>
<tr>
<td></td>
<td>Prm(λ)=Q2π(λ+−λ)(λ−λ−)λsubscript𝑃𝑟𝑚𝜆𝑄2𝜋subscript𝜆𝜆𝜆subscript𝜆𝜆P_{rm}\left(\lambda\right)=\frac{Q}{2\pi}\frac{\sqrt{\left(\lambda_{+}-\lambda\right)\left(\lambda-\lambda_{-}\right)}}{\lambda}</td>
<td></td>
<td>(6)</td>
</tr>
</tbody>
</table>
where $`\lambda_{+}`$ and $`\lambda_{-}`$ are maximum and minimum eigenvalues of $`R,`$ respectively and $`\lambda_{-} \leq \lambda_{i} \leq \lambda_{+}`$. $`\lambda_{+}`$ and $`\lambda_{-}`$are given analytically by
<table>
<tbody>
<tr>
<td></td>
<td>λ±=1+1Q±21Q.subscript𝜆plus-or-minusplus-or-minus11𝑄21𝑄\lambda_{\pm}=1+\frac{1}{Q}\pm 2\sqrt{\frac{1}{Q}}.</td>
<td></td>
<td>(7)</td>
</tr>
</tbody>
</table>
Random matrices display *universal* functional forms for eigenvalues correlations which depend on the general symmetries of the matrix only. First step to test the data for such a universal properties is to find a transformation called “unfolding”, which maps the eigenvalues $`\lambda_{i}`$ to new variables, “unfolded eigenvalues” $`\xi_{i},`$ whose distribution is uniform \[[9](#bib.bib9), [10](#bib.bib10), [11](#bib.bib11)\]. Unfolding ensures that the distances between eigenvalues are expressed in units of *local* mean eigenvalues spacing \[[9](#bib.bib9)\], and thus facilitates the comparison with analytical results.
We define the cumulative distribution function of eigenvalues, which counts the number of eigenvalues in the interval $`{\lambda_{i} \leq \lambda},`$
<table>
<tbody>
<tr>
<td></td>
<td>F(λ)=N∫−∞λP(x)𝑑x,𝐹𝜆𝑁superscriptsubscript𝜆𝑃𝑥differential-d𝑥F\left(\lambda\right)=N\int_{-\infty}^{\lambda}P\left(x\right)dx,</td>
<td></td>
<td>(8)</td>
</tr>
</tbody>
</table>
where $`P\hspace{0pt}(x)`$ denotes the probability density of eigenvalues and $`N`$ is the total number of eigenvalues. The function $`F\hspace{0pt}(\lambda)`$ can be decomposed into an average and a fluctuating part,
<table>
<tbody>
<tr>
<td></td>
<td>F(λ)=Fav(λ)+Ffluc(λ),𝐹𝜆subscript𝐹𝑎𝑣𝜆subscript𝐹𝑓𝑙𝑢𝑐𝜆F\left(\lambda\right)=F_{av}\left(\lambda\right)+F_{fluc}\left(\lambda\right),</td>
<td></td>
<td>(9)</td>
</tr>
</tbody>
</table>
Since $`P_{f\hspace{0pt}l\hspace{0pt}u\hspace{0pt}c} \equiv {{{d\hspace{0pt}F_{f\hspace{0pt}l\hspace{0pt}u\hspace{0pt}c}\hspace{0pt}(\lambda)}/d}\hspace{0pt}\lambda} = 0`$ on average,
<table>
<tbody>
<tr>
<td></td>
<td>Prm(λ)≡dFav(λ)dλ,subscript𝑃𝑟𝑚𝜆𝑑subscript𝐹𝑎𝑣𝜆𝑑𝜆P_{rm}\left(\lambda\right)\equiv\frac{dF_{av}\left(\lambda\right)}{d\lambda},</td>
<td></td>
<td>(10)</td>
</tr>
</tbody>
</table>
is the averaged eigenvalues density. The dimensionless, unfolded eigenvalues are then given by
<table>
<tbody>
<tr>
<td></td>
<td>ξi≡Fav(λi).subscript𝜉𝑖subscript𝐹𝑎𝑣subscript𝜆𝑖\xi_{i}\equiv F_{av}\left(\lambda_{i}\right).</td>
<td></td>
<td>(11)</td>
</tr>
</tbody>
</table>
Three known universal properties of GOE matrices (matrices whose elements are distributed according to a Gaussian probability measure) are: (i) the distribution of nearest-neighbor eigenvalues spacing $`P_{G\hspace{0pt}O\hspace{0pt}E}\hspace{0pt}(s)`$
<table>
<tbody>
<tr>
<td></td>
<td>PGOE(s)=πs2exp(−π4s2),subscript𝑃𝐺𝑂𝐸𝑠𝜋𝑠2𝑒𝑥𝑝𝜋4superscript𝑠2P_{GOE}\left(s\right)=\frac{\pi s}{2}exp\left(-\frac{\pi}{4}s^{2}\right),</td>
<td></td>
<td>(12)</td>
</tr>
</tbody>
</table>
\(ii\) the distribution of next-nearest-neighbor eigenvalues spacing, which is according to the theorem due to \[[8](#bib.bib8)\] is identical to the distribution of nearest-neighbor spacing of Gaussian symplectic ensemble (GSE),
<table>
<tbody>
<tr>
<td></td>
<td>PGSE(s)=21836π3s4exp(−649πs2)subscript𝑃𝐺𝑆𝐸𝑠superscript218superscript36superscript𝜋3superscript𝑠4𝑒𝑥𝑝649𝜋superscript𝑠2P_{GSE}\left(s\right)=\frac{2^{18}}{3^{6}\pi^{3}}s^{4}exp\left(-\frac{64}{9\pi}s^{2}\right)</td>
<td></td>
<td>(13)</td>
</tr>
</tbody>
</table>
and finally (iii) the “number variance” statistics $`\Sigma^{2}`$, defined as the variance of the number of unfolded eigenvalues in the intervals of length $`l`$, around each $`\xi_{i}`$ \[[9](#bib.bib9), [11](#bib.bib11), [10](#bib.bib10)\].
<table>
<tbody>
<tr>
<td></td>
<td>Σ2(l)=⟨[n(ξ,l)−l]2⟩ξ,superscriptΣ2𝑙subscriptdelimited-⟨⟩superscriptdelimited-[]𝑛𝜉𝑙𝑙2𝜉\Sigma^{2}\left(l\right)=\left\langle\left[n\left(\xi,l\right)-l\right]^{2}\right\rangle_{\xi},</td>
<td></td>
<td>(14)</td>
</tr>
</tbody>
</table>
where $`n\hspace{0pt}(\xi,l)`$ is the number of the unfolded eigenvalues in the interval $`\left\lbrack {\xi - \frac{l}{2}},{\xi + \frac{l}{2}} \right\rbrack`$. The number variance is expressed as follows
<table>
<tbody>
<tr>
<td></td>
<td>Σ2(l)=l−2∫0l(l−x)Y(x)𝑑x,superscriptΣ2𝑙𝑙2superscriptsubscript0𝑙𝑙𝑥𝑌𝑥differential-d𝑥\Sigma^{2}\left(l\right)=l-2\int_{0}^{l}\left(l-x\right)Y\left(x\right)dx,</td>
<td></td>
<td>(15)</td>
</tr>
</tbody>
</table>
where $`Y\hspace{0pt}(x)`$ for the GOE case is given by \[[9](#bib.bib9)\]
<table>
<tbody>
<tr>
<td></td>
<td>Y(x)=s2(x)+dsdx∫x∞s(x′)𝑑x′,𝑌𝑥superscript𝑠2𝑥𝑑𝑠𝑑𝑥superscriptsubscript𝑥𝑠superscript𝑥′differential-dsuperscript𝑥′Y\left(x\right)=s^{2}\left(x\right)+\frac{ds}{dx}\int_{x}^{\infty}s\left(x^{\prime}\right)dx^{\prime},</td>
<td></td>
<td>(16)</td>
</tr>
</tbody>
</table>
and
<table>
<tbody>
<tr>
<td></td>
<td>s(x)=sin(πx)πx.𝑠𝑥𝑠𝑖𝑛𝜋𝑥𝜋𝑥s\left(x\right)=\frac{sin\left(\pi x\right)}{\pi x}.</td>
<td></td>
<td>(17)</td>
</tr>
</tbody>
</table>
Just as was stressed in \[[19](#bib.bib19), [18](#bib.bib18), [25](#bib.bib25)\] the overall time of observation is crucial for explaining the empirical cross-correlation coefficients. On one hand, the longer we observe the traffic the more information about the correlations we obtain and less “noise” we introduce. On the other hand, the correlations are not stationary, i.e. they can change with time. To differentiate the “random” contribution to empirical correlation coefficients from “genuine” contribution, the eigenvalues statistics of $`C`$ is contrasted with the eigenvalues statistics of a correlation matrix taken from the so called “chiral” Gaussian Orthogonal Ensemble \[[19](#bib.bib19)\]. Such an ensemble is one of the ensembles of RMT \[[25](#bib.bib25), [26](#bib.bib26)\], briefly discussed in Appendix A. A *random* cross-correlation matrix, which is a matrix filled with uncorrelated Gaussian random numbers, is supposed to represent transient uncorrelated in time network activity, that is, a completely noisy environment. In case the cross-correlation matrix $`C`$ obeys the same eigenstatistical properties as the RMT-matrix, the network traffic is equilibrated and deemed universal in a sense that every single connection interacts with the rest in a completely chaotic manner. It also means a complete absence of congestions and anomalies. Meantime, any stable in time deviations from the *universal* predictions of RMT signify system-specific, nonrandom properties of the system, providing the clues about the nature of the underlying interactions. That allows us to establish the profile of system-specific correlations.
## IV data
In this paper, we study the averaged traffic count data collected from all router-router and router-VLAN subnet connections of the University of Louisville backbone routers system. The system consists of nine interconnected multi-gigabit backbone routers, over $`200`$ Ethernet segments and over $`300`$ VLAN subnets. We collected the traffic count data for $`3`$ months, for the period from September *$`21`$, $`2006`$* to December *$`20`$, $`2006`$* from $`7`$ routers, since two routers are reserved for server farms. The overall data amounted to approximately $`18`$ GB.
The traffic count data is provided by Multi Router Traffic Grapher (MRTG) tool that reads the SNMP traffic counters. MRTG log file never grows in size due to the data consolidation algorithm: it contains records of average incoming, outgoing, max and min transfer rate in bytes per second with time intervals $`300`$ seconds, *$`30`$* minutes, $`1`$ day and $`1`$ month. We extracted $`300`$ seconds interval data for seven days. Then, we separated the incoming and outgoing traffic counts time series and considered them as independent. For $`352`$ connections we formed $`L = 2015`$ records of $`N = 704`$ time series with $`300`$ seconds interval.
We pursued the changes in the traffic rate, thus, we excluded from consideration the connections, where channel is open but the traffic is not established or there is just constant rate and equal low amount test traffic. Another reason for excluding the “empty” traffic time series is that they make the time series cross-correlation matrix unnecessary sparse. The exclusion does not influence the analysis and results. After the exclusions the number of the traffic time series became $`N = 497`$.
To calculate the traffic rate change $`G_{i}\hspace{0pt}(t)`$ we used the logarithm of the ratio of two successive counts. As it is stated earlier, $`l\hspace{0pt}o\hspace{0pt}g`$-transformation makes the ratio independent from the traffic volume and allows capturing the subtle changes in the traffic rate. We added 1 byte to all data points, to avoid manipulations with $`l\hspace{0pt}o\hspace{0pt}g\hspace{0pt}(0)`$, in cases where traffic count is equal to zero bytes. This measure did not affect the changes in the traffic rate.
## V eigenvalue distribution of cross-correlation matrix, comparison with rmt
We constructed inter-VLAN traffic cross-correlation matrix $`C`$ with number of time series $`N = 497`$ and number of observations per series $`L = 2015`$, ($`Q = 4.0625`$) so that, $`\lambda_{+} = 2.23843`$ and $`\lambda_{-} = 0.253876`$. Our first goal is to compare the eigenvalue distribution $`P\hspace{0pt}(\lambda)`$ of $`C`$ with $`P_{r\hspace{0pt}m}\hspace{0pt}(\lambda)`$ \[[23](#bib.bib23)\]. To compute eigenvalues of $`C`$ we used standard *MATLAB* function. The empirical probability distribution $`P\hspace{0pt}(\lambda)`$ is then given by the corresponding histogram. We display the resulting distribution $`P\hspace{0pt}(\lambda)`$ in Figure 1 and compare it to the probability distribution $`P_{r\hspace{0pt}m}\hspace{0pt}(\lambda)`$ taken from Eq. ([6](#S3.E6 "In III rmt methodology ‣ Analysis of Inter-Domain Traffic Correlations: Random Matrix Theory Approach")) calculated for the same value of traffic time series parameters ($`Q = 4.0625`$). The solid curve demonstrates $`P_{r\hspace{0pt}m}\hspace{0pt}(\lambda)`$ of Eq.([6](#S3.E6 "In III rmt methodology ‣ Analysis of Inter-Domain Traffic Correlations: Random Matrix Theory Approach")). The largest eigenvalue shown in inset has the value $`\lambda_{497} = 8.99`$. We zoom in the deviations from the RMT predictions on the inset to Figure 1.
[Refer to caption](/html/0706.2520/assets/x1.png)
Figure 1: Empirical probability distribution function $`P\hspace{0pt}(\lambda)`$ for the inter-VLAN traffic cross-correlations matrix $`C`$ (histogram).
We note the presence of “bulk” (RMT-like) eigenvalues which fall within the bounds \[$`\lambda_{-},`$$`\lambda_{+}`$\] for $`P_{r\hspace{0pt}m}\hspace{0pt}(\lambda)`$, and presence of the eigenvalues which lie outside of the “bulk”, representing deviations from the RMT predictions. In particular, largest eigenvalue $`\lambda_{497} = 8.99`$ for seven days period is approximately four times larger than the RMT upper bound $`\lambda_{+}`$.
The histogram for well-defined bulk agrees with $`P_{r\hspace{0pt}m}\hspace{0pt}(\lambda)`$ suggesting that the cross-correlations of matrix $`C`$ are mostly random. We observe that inter-VLAN traffic time series interact mostly in a random fashion.
Nevertheless, the agreement of empirical probability distribution $`P\hspace{0pt}(\lambda)`$ of the bulk with $`P_{r\hspace{0pt}m}\hspace{0pt}(\lambda)`$ is not sufficient to claim that the bulk of eigenvalue spectrum is random. Therefore, further RMT tests are needed \[[19](#bib.bib19)\].
To do that, we obtained the unfolded eigenvalues $`\xi_{i}`$ by following the phenomenological procedure referred to as Gaussian broadening \[[27](#bib.bib27)\], (see \[[27](#bib.bib27), [35](#bib.bib35), [19](#bib.bib19), [18](#bib.bib18)\]). The empirical cumulative distribution function of eigenvalues $`F\hspace{0pt}(\lambda)`$ agrees well with the $`F_{a\hspace{0pt}v}\hspace{0pt}(\lambda)`$ (see Figure 2), where $`\xi_{i}`$ obtained with Gaussian broadening procedure with the broadening parameter $`a = 8`$.
[Refer to caption](/html/0706.2520/assets/x2.png)
Figure 2: The empirical cumulative distribution of $`\lambda_{i}`$ and unfolded eigenvalues $`\xi_{i} \equiv {F_{a\hspace{0pt}v}\hspace{0pt}(\lambda)}`$.
The first independent RMT test is the comparison of the distribution of the nearest-neighbor unfolded eigenvalue spacing $`P_{n\hspace{0pt}n}\hspace{0pt}(s)`$, where $`s \equiv {\xi_{k + 1} - \xi_{k}}`$ with $`P_{G\hspace{0pt}O\hspace{0pt}E}\hspace{0pt}(s)`$ \[[9](#bib.bib9), [10](#bib.bib10), [11](#bib.bib11)\]. The empirical probability distribution of nearest-neighbor unfolded eigenvalues spacing $`P_{n\hspace{0pt}n}\hspace{0pt}(s)`$ and $`P_{G\hspace{0pt}O\hspace{0pt}E}\hspace{0pt}(s)`$ are presented in Figure 3. The Gaussian decay of $`P_{G\hspace{0pt}O\hspace{0pt}E}\hspace{0pt}(s)`$ for large $`s`$ suggests that $`P_{G\hspace{0pt}O\hspace{0pt}E}\hspace{0pt}(s)`$ “probes” scales only of the order of one eigenvalue spacing. The solid line represents.
[Refer to caption](/html/0706.2520/assets/x3.png)
Figure 3: Nearest-neighbor spacing distribution $`P_{n\hspace{0pt}n}\hspace{0pt}(s)`$ of unfolded eigenvalues $`\xi_{i}`$ of cross-correlation matrix $`C`$.
The agreement between empirical probability distribution $`P_{n\hspace{0pt}n}\hspace{0pt}(s)`$ and the distribution of nearest-neighbor eigenvalues spacing of the GOE matrices $`P_{G\hspace{0pt}O\hspace{0pt}E}\hspace{0pt}(s)`$ testifies that the positions of two adjacent empirical unfolded eigenvalues at the distance $`s`$ are correlated just as the eigenvalues of the GOE matrices.
Next, we took on the distribution $`P_{n\hspace{0pt}n\hspace{0pt}n}\hspace{0pt}\left( s' \right)`$ of next-nearest-neighbor spacings $`s' \equiv {\xi_{k + 2} - \xi_{k}}`$ between the unfolded eigenvalues. According to \[[8](#bib.bib8)\] this distribution should fit to the distribution of nearest-neighbor spacing of the GSE. We demonstrate this correspondence in Figure 4. The solid line shows $`P_{G\hspace{0pt}S\hspace{0pt}E}\hspace{0pt}(s)`$.
[Refer to caption](/html/0706.2520/assets/x4.png)
Figure 4: Next-nearest-neighbor eigenvalue spacing distribution $`{P_{n\hspace{0pt}n\hspace{0pt}n}\hspace{0pt}\left( s' \right)}.`$
Finally, the long-range two-point eigenvalue correlations were tested. It is known \[[9](#bib.bib9), [10](#bib.bib10), [11](#bib.bib11)\], that if eigenvalues are uncorrelated we expect the number variance to scale with $`l`$, $`\Sigma^{2} \sim l`$. Meanwhile, when the unfolded eigenvalues of $`C`$ are correlated, $`\Sigma^{2}`$ approaches constant value, revealing “spectral rigidity” \[[9](#bib.bib9), [10](#bib.bib10), [11](#bib.bib11)\]. In Figure 5, we contrasted Poissonian number variance with the one we observed, and came to the conclusion that eigenvalues belonging to the “bulk” clearly exhibit universal RMT properties. The broadening parameter $`a = 8`$ was used in Gaussian broadening procedure to unfold the eigenvalues $`\lambda_{i}`$ \[[27](#bib.bib27), [35](#bib.bib35), [19](#bib.bib19), [18](#bib.bib18)\]. The dashed line corresponds to the case of uncorrelated eigenvalues.
[Refer to caption](/html/0706.2520/assets/x5.png)
Figure 5: Number variance $`\Sigma^{2}\hspace{0pt}(l)`$ calculated from the unfolded eigenvalues $`\xi_{i}`$ of $`C`$.
These findings show that the system of inter-VLAN traffic has a *universal* part of eigenvalues spectral correlations, shared by broad class of systems, including chaotic and disordered systems, nuclei, atoms and molecules. Thus it can be concluded, that the bulk eigenvalue statistics of the inter-VLAN traffic cross-correlation matrix $`C`$ are consistent with those of real symmetric random matrix $`R`$, given by Eq. ([5](#S3.E5 "In III rmt methodology ‣ Analysis of Inter-Domain Traffic Correlations: Random Matrix Theory Approach")) \[[24](#bib.bib24)\]. Meantime, the deviations from the RMT contain the information about the system-specific correlations. The next section is entirely devoted to the analysis of the eigenvalues and eigenvectors deviating from the RMT, which signifies the meaningful inter-VLAN traffic interactions.
## VI inter-vlan traffic interactions analysis
We overview the points of interest in eigenvectors of inter-VLAN traffic cross-correlation matrix $`C`$, which are determined according to $`{C\hspace{0pt}u^{k}} = {\lambda_{k}\hspace{0pt}u^{k}}`$, where $`\lambda_{k}`$ is $`k`$-th eigenvalue. Particularly important characteristics of eigenvectors, proven to be useful in physics of disordered conductors is the inverse participation ratio (IPR) (see, for example, Ref. \[[11](#bib.bib11)\]). In such systems, the IPR being a function of an eigenstate (eigenvector) allows to judge and clarify whether the corresponding eigenstate, and therefore electron is extended or localized.
###
VI-A
Inverse participation ratio of eigenvectors components
For our purposes, it is sufficient to know that IPR quantifies the reciprocal of the number of significant components of the eigenvector. For the eigenvector $`u^{k}`$ it is defined as
<table>
<tbody>
<tr>
<td></td>
<td>Ik≡∑l=1N[ulk]4,superscript𝐼𝑘superscriptsubscript𝑙1𝑁superscriptdelimited-[]superscriptsubscript𝑢𝑙𝑘4I^{k}\equiv\sum_{l=1}^{N}\left[u_{l}^{k}\right]^{4},</td>
<td></td>
<td>(18)</td>
</tr>
</tbody>
</table>
where $`u_{l}^{k}`$, $`l = {1,\ldots,497}`$ are components of the eigenvector $`u^{k}`$. In particular, the vector with one significant component has $`I^{k} = 1`$, while vector with identical components $`u_{l}^{k} = {1/\sqrt{N}}`$ has $`I^{k} = {1/N}`$.
[Refer to caption](/html/0706.2520/assets/x6.png)
Figure 6: Inverse participation ratio as a function of eigenvalue $`\lambda`$.
Consequently, the inverse of IPR gives us a number of significant participants of the eigenvector. In Figure 6 we plot the IPR of cross-correlation matrix $`C`$ as a function of eigenvalue $`\lambda`$. The control plot is IPR of eigenvectors of random cross-correlation matrix $`R`$ of Eq. [5](#S3.E5 "In III rmt methodology ‣ Analysis of Inter-Domain Traffic Correlations: Random Matrix Theory Approach"). As we can see, eigenvectors corresponding to eigenvalues from $`0.25`$ to $`3.5`$, what is within the RMT boundaries, have IPR close to $`0`$. This means that almost all components of eigenvectors in the bulk interact in a random fashion. The number of significant components of eigenvectors deviating from the RMT is typically twenty times smaller than the one of the eigenvectors within the RMT boundaries, around twenty. For instance, IPR of eigenvector $`u^{492}`$, which corresponds to the eigenvalue $`5.9`$ in Figure 6, is $`0.05`$, i.e. twenty time series are significantly contribute to $`u^{492}`$. Another observation which we derive from Figure 6 is that the number of eigenvectors significant participants is considerably smaller at both edges of the eigenvalue spectrum. These findings resemble the results of \[[19](#bib.bib19)\], where the eigenvectors with a few participating components were referred to as *localized* vectors. The theory of *localization* is explained in the context of random band matrices, where elements independently drawn from different probability distributions \[[19](#bib.bib19)\]. These matrices despite their randomness, still contain probabilistic information. The *localization* in inter-VLAN traffic is explained as follows. The separated broadcast domains, i.e. VLANs forward traffic from one to another only through the router, reducing the routing for broadcast containment. Although the optimal VLAN deployment is to keep as much traffic as possible from traversing through the router, the bottleneck at the large number of VLANs is unavoidable.
###
VI-B
Distribution of eigenvectors components
Another target of interest is the distribution of the components $`\left\{ {{{u_{l}^{k};l} = 1},{\ldots,N}} \right\}`$ of eigenvector $`u^{k}`$ of the interactions matrix $`C`$. To calculate vectors $`u`$ we used the *MATLAB* routine again and obtained components distribution $`p\hspace{0pt}(u)`$ of the eigenvectors components. Then, we contrasted it with the RMT predictions for the eigenvector distribution $`p_{r\hspace{0pt}m}\hspace{0pt}(u)`$ of the random correlation matrix $`R`$. According to \[[11](#bib.bib11)\] $`p_{r\hspace{0pt}m}\hspace{0pt}(u)`$ has a Gaussian distribution with mean zero and unit variance, i.e.
<table>
<tbody>
<tr>
<td></td>
<td>prm(u)=12πexp(−u22).subscript𝑝𝑟𝑚𝑢12𝜋𝑒𝑥𝑝superscript𝑢22p_{rm}\left(u\right)=\frac{1}{\sqrt{2\pi}}exp\left(\frac{-u^{2}}{2}\right).</td>
<td></td>
<td>(19)</td>
</tr>
</tbody>
</table>
The weights of randomly interacting traffic counts time series, which are represented by the eigenvectors components has to be distributed normally. The results are presented in Figure 7. One can see (from Figures 7a and 7b) that $`p\hspace{0pt}(u)`$ for two $`u^{k}`$ taken from the bulk is in accord with $`p_{r\hspace{0pt}m}\hspace{0pt}(u)`$. The distribution $`p\hspace{0pt}(u)`$ corresponding to the eigenvalue $`\lambda_{i}`$, which exceeds the RMT upper bound ($`\lambda_{i} > \lambda_{+}`$), is shown in Figure 7c. The solid line shows $`p_{r\hspace{0pt}m}\hspace{0pt}(u)`$ from Eq. [19](#S6.E19 "In VI-B Distribution of eigenvectors components ‣ VI inter-vlan traffic interactions analysis ‣ Analysis of Inter-Domain Traffic Correlations: Random Matrix Theory Approach"). (c) $`p\hspace{0pt}(u)`$ for $`u^{496}`$, corresponding to the eigenvalue outside of the RMT bulk. (d) $`p\hspace{0pt}(u)`$ for $`u^{497}`$, corresponding to largest eigenvalue.
[Refer to caption](/html/0706.2520/assets/x7.png)
Figure 7: Distribution of components $`p\hspace{0pt}(u)`$ of eigenvectors corresponding to eigenvalues (a) from the middle of the bulk, i.e.$`\lambda_{-} < \lambda < \lambda_{+}`$, (b) from the bulk close to $`\lambda_{+}`$, (c) $`\lambda_{496}`$ (d) $`\lambda_{497}`$.
###
VI-C
Deviating eigenvalues and significant inter-VLAN traffic series contributing to the deviating eigenvectors.
The distribution of $`u^{497}`$, the eigenvector corresponding to the largest eigenvalue $`\lambda_{497}`$, deviates significantly from the Gaussian (as follows from Figure 7d). While Gaussian kurtosis has the value 3, the kurtosis of $`p\hspace{0pt}\left( u^{497} \right)`$ comes out to $`23.22`$. The smaller number of significant components of the eigenvector also influences the difference between Gaussian distribution and empirical distribution of eigenvector components. More than half of $`u^{497}`$components have the same sign, thus slightly shifting the $`p\hspace{0pt}(u)`$ to one side. This result suggests the existence of the common VLAN traffic intended for inter-VLAN communication that affects all of the significant participants of the eigenvector $`u^{497}`$with the same bias. We know that the number of significant components of $`u^{497}`$ is twenty two, since IPR of $`u^{497}`$is $`0.045`$. Hence, the largest eigenvector content reveals 22 traffic time series, which are affected by the same event. We obtain the time series, which affects 22 traffic time series by the following procedure. First of all, we calculate projection $`G^{497}\hspace{0pt}(t)`$ of the time series $`G_{i}\hspace{0pt}(t)`$ on the eigenvector $`u^{497}`$,
<table>
<tbody>
<tr>
<td></td>
<td>G497(t)≡∑i=1497ui497Gi(t)superscript𝐺497𝑡superscriptsubscript𝑖1497superscriptsubscript𝑢𝑖497subscript𝐺𝑖𝑡G^{497}\left(t\right)\equiv\sum_{i=1}^{497}u_{i}^{497}G_{i}\left(t\right)</td>
<td></td>
<td>(20)</td>
</tr>
</tbody>
</table>
Next, we compare $`G^{497}\hspace{0pt}(t)`$ with $`G_{i}\hspace{0pt}(t)`$, by finding the correlation coefficient $`\left\langle {\frac{G^{497}\hspace{0pt}(t)}{\sigma^{497}}\hspace{0pt}\frac{G_{i}\hspace{0pt}(t)}{\sigma_{i}}} \right\rangle`$. The Fiber Distributed Data Interface (FDDI)-VLAN internet switch at one of the routers demonstrates the largest correlation coefficient of $`0.89`$ (see Figure 8).
[Refer to caption](/html/0706.2520/assets/x8.png)
Figure 8: (a) FDDI-VLAN internet switch time series regressed against the projection $`G^{497}\hspace{0pt}(t)`$ from Eq. [20](#S6.E20 "In VI-C Deviating eigenvalues and significant inter-VLAN traffic series contributing to the deviating eigenvectors. ‣ VI inter-vlan traffic interactions analysis ‣ Analysis of Inter-Domain Traffic Correlations: Random Matrix Theory Approach"). (b) Time series defined by the eigenvector corresponding to eigenvalue within RMT bounds shows no linear dependence on $`{G^{497}\hspace{0pt}(t)}.`$
The eigenvector $`u^{497}`$ has the following content: seven most significant participants are seven FDDI-VLAN switches at the seven routers. The presence of FDDI-VLAN switch provide us with information about VLAN membership definition. FDDI is layer 2 protocol, which means that at least one of two layer 2 membership is used, port group or/and MAC address membership. The next group of significant participants comprises of VLAN traffic intended for routing and already routed traffic from different VLANs. The final group of significant participants constitutes open switches, which pick up any “leaking” traffic on the router. Usually, the “leaking” traffic is the network management traffic, a very low level traffic which spikes when queried by the management systems.
If every deviating eigenvalue notifies a particular sub-model of non-random interactions of the network, then every corresponding eigenvector presents the number of significant dimensions of sub-model. Thus, we can think of every deviating eigenvector as a representative network-wide “snapshot” of interactions within the certain dimensions.
The analysis of the significant participants of the deviating eigenvectors revealed three types of inter-VLAN traffic time series groupings. One group contains time series, which are interlinked on the router. We recognize them as, router1-VLAN_1000 traffic, router1-firewall traffic and VLAN_1000-router1 traffic. The time series, which are listed as router1-vlan_2000, router2-VLAN_2000, router3-VLAN_2000, etc., are reserved for the same service VLAN on every router and comprise another group. The content of these groups suggests the VLANs implementation, it is a mixture of infrastructural approach, where functional groups (departments, schools, etc.) are considered, and service approach, where VLAN provides a particular service (network management, firewall, etc.).
## VII stability of inter-vlan traffic interactions in time
We expect to observe the stability of inter-VLAN traffic interactions in the period of time used to compute traffic cross-correlation matrix $`C`$. The eigenvalues distribution at different time periods provides the information about the system stabilization, i.e. about the time after which the fluctuations of eigenvalues are not significant. Time periods of $`1`$ hour, $`3`$ hours and $`6`$ hours are not sufficient to gain the knowledge about the system, which is demonstrated in Figure 9a. In Figure 9b the system stabilizes after $`1`$ day period. To observe the time stability of inter-VLAN meaningful interactions we computed the “overlap matrix” of the deviating eigenvectors for the time period $`t`$ and deviating eigenvectors for the time period $`t + \tau`$, where $`{t = {60\hspace{0pt}h}},{\tau = \left\{ {0\hspace{0pt}h},{3\hspace{0pt}h},{12\hspace{0pt}h},{24\hspace{0pt}h},{36\hspace{0pt}h},{48\hspace{0pt}h} \right\}}`$.
[Refer to caption](/html/0706.2520/assets/x9.png)
Figure 9: (a) Eigenvalues distributions of traffic streams correlation matrix $`C`$ for $`1`$ hour, $`3`$ hours and $`6`$ hours time intervals. (b) Eigenvalues distributions for $`24`$ hours, $`48`$ hours and $`72`$ hours
First, we obtained matrix D from $`p = 57`$ eigenvectors, which correspond to $`p`$ eigenvalues outside of the RMT upper bound $`\lambda_{+}`$. Then we computed the “overlap matrix” $`O\hspace{0pt}(t,\tau)`$ from $`D_{A}\hspace{0pt}D_{B}^{T}`$, where $`O_{i\hspace{0pt}j}`$ is a scalar product of the eigenvector $`u^{i}`$ of period $`A`$ (starting at time $`t = t`$) with $`\text{u}^{j}`$ of period $`B`$ at the time $`t = {t + \tau}`$,
<table>
<tbody>
<tr>
<td></td>
<td>Oij(t,τ)≡∑k=1NDik(t)Dik(t+τ)subscript𝑂𝑖𝑗𝑡𝜏superscriptsubscript𝑘1𝑁subscript𝐷𝑖𝑘𝑡subscript𝐷𝑖𝑘𝑡𝜏O_{ij}\left(t,\tau\right)\equiv\sum_{k=1}^{N}D_{ik}\left(t\right)D_{ik}\left(t+\tau\right)</td>
<td></td>
<td>(21)</td>
</tr>
</tbody>
</table>
The values of $`O_{i\hspace{0pt}j}\hspace{0pt}(t,\tau)`$ elements at $`i = j`$, i.e. of diagonal elements of matrix $`O`$ will be $`1`$, if the matrix $`D\hspace{0pt}\left( {t + \tau} \right)`$ is identical to the matrix $`D\hspace{0pt}(t)`$. Clearly, the diagonal of the “overlap matrix” $`O`$ can serve as an indicator of time stability of $`p`$ eigenvectors outside of the RMT upper bound $`\lambda_{+}`$. The gray scale colormap of the “overlap matrices” $`O\hspace{0pt}\left( {{t = {60\hspace{0pt}h}},{\tau = \left\{ {0\hspace{0pt}h},{3\hspace{0pt}h},{12\hspace{0pt}h},{24\hspace{0pt}h},{36\hspace{0pt}h},{48\hspace{0pt}h} \right\}}} \right)`$ is presented in Figure 10. Black color of grayscale represents $`O_{i\hspace{0pt}j} = 1`$, white color represents $`{O_{i\hspace{0pt}j} = 0}.`$ The most stable eigenvalue is $`\lambda_{492}.`$
[Refer to caption](/html/0706.2520/assets/x10.png)
Figure 10: The grayscale of overlap matrix $`O\hspace{0pt}(t,\tau)`$ at $`t = {60\hspace{0pt}h}`$ and $`\tau = \left\{ {0\hspace{0pt}h},{3\hspace{0pt}h},{12\hspace{0pt}h},{24\hspace{0pt}h},{36\hspace{0pt}h},{48\hspace{0pt}h} \right\}`$.
At lag $`\tau = 3`$ hours the inter-VLAN interactions show the highest degree of stability. For further lags the overall stability decays. As the analysis of deviating eigenvectors content showed, the highly interacting traffic time series are time series of service based VLANs, intended for routing. Particular network services are evoked at the same time and active for the same period of time, which explains the stability and consequent decay of deviating eigenvectors of traffic interactions.
## VIII detecting anomalies of traffic interactions
We assume that the health of inter-VLAN traffic is expressed by stability of its interactions in time. Meanwhile, the temporal critical events or anomalies will cause the temporal instabilities. The “deviating” eigenvalues and eigenvectors provide us with stable in time snapshots of interactions representative of the entire network. Therefore, these eigenvectors judged on the basis of their IPR can serve as monitoring parameters of the system stability.
Among the essential anomalous events of VLAN infrastructure we can list violations in VLAN membership assignment, in address resolution protocol, in VLAN trunking protocol, router misconfiguration. The violation of membership assignment and router misconfiguration will cause the changes in the picture of random and non-random interactions of inter-VLAN traffic. To shed more light on the possibilities of anomaly detection we conducted the experiments to establish spatial-temporal traces of instabilities caused by artificial and temporal increase of the correlation in normal non-congested inter-VLAN traffic. We explored the possibility to distinguish different types of increased temporal correlations. Finally, we observed the consequences of breaking the interactions between time series, by injecting traffic counts obtained from sample of random distribution.
*Experiment 1*
We selected the traffic counts time series representing the components of the eigenvector which lies within the RMT bounds and temporarily increased the correlation between these series for three hour period. The proposed monitoring parameters show the dependence of system stability on the number of temporarily correlated time series (see Figure 11). Presented in Figure 11, left to right are (a) eigenvalue distribution of interactions with two temporarily correlated time series, (b) IPR of eigenvectors of interactions with two temporarily correlated time series, (c) the overlap matrix of deviating eigenvectors with two temporarily correlated time series. Top to bottom the layout shows these monitoring parameters when correlation is temporarily increased between 10 connections (d,e and f) and between 20 connections (g,h and i).
[Refer to caption](/html/0706.2520/assets/x11.png)
Figure 11: Eigenvalues distribution, IPR and overlap matrix of deviating eigenvectors.
One can conclude that increased temporal correlation between two time series and between ten time series does not affect system stability. Meanwhile, when the number of temporarily correlated time series reaches the number of significant participants of $`u^{497},`$ which is calculated as inverse of $`I^{497}`$and is equal to twenty two, the system becomes visibly unstable. The largest eigenvalue changes from $`10`$ in stable condition to $`12`$, the tail of inverse participation ratio plot is extended and the diagonal of “overlap matrix” disappears at twenty temporarily correlated time series.
In Figures 12 (a, b, c and d ), the temporal correlation between ten time series is traced with the matrix of sorted in decreasing order of their components deviating from RMT eigenvectors.
[Refer to caption](/html/0706.2520/assets/x12.png)
Figure 12: Sorted deviating eigenvectors with injected correlation among ten traffic time series.
The sorted in decreasing order deviating eigenvectors of $`60\hspace{0pt}h`$ of uninterrupted traffic are presented in Figure 12a. Then, after three hours of uninterrupted traffic the weights of eigenvectors components, which had zero value start changing, This is captured in Figure 12b. Same process for traffic with induced three hours correlation is captured in Figure 12c. The difference between results in Figures 12b and 12c is presented in Figure 12d. The procedure used to visualize this produces the high rate of false positive alarms.
In addition, we visualize in Figure 13 the system instability during temporal increase of correlation between twenty time series with spatial-temporal representation of eigenvector $`u^{497}`$.
[Refer to caption](/html/0706.2520/assets/x13.png)
Figure 13: (a) The weights of components of $`u^{497}`$ plotted for time period from $`36`$ to $`84`$ hours of uninterrupted traffic with 6 hours interval. (b) The weights of components of $`u^{497}`$ plotted with respect to the same time period, with induced three hours correlation. (c) The weights of components of $`u^{496}`$ plotted with respect to the same time period, with induced three hours correlation.
We used the weights of components of eigenvector $`u^{497}`$, defined for IPR computation and plotted them with respect to time $`t + \tau`$, where $`t = 36`$ hours and $`{\tau = {6\hspace{0pt}n}},`$ where $`n \in \left\{ 0,1,\ldots,7 \right\}`$. In Figure 13a the spatial-temporal pattern of $`u^{497}`$ captures precise locations of system-specific interactions of uninterrupted traffic for $`84`$ hours of observation. The abrupt change of this pattern in Figure 13b indicates the starting point of induced correlation between twenty traffic time series usually interacting in a random fashion. It turns out, that the “normal” stable pattern of eigenvector $`u^{497}`$ moves to eigenvector $`u^{496}`$, when the interruption ends. Thus, we are able to observe the end point of the induced correlations in Figure 13c, which represents weights of components of eigenvector $`u^{496}`$ plotted with respect to the same time intervals. With this setup we are able to locate the anomaly in time and space. Translated to network topological representation, the behavior of eigenvectors $`u^{497}`$and $`u^{496}`$ during our manipulations with inter-VLAN traffic may be monitored with the following graphs (see Figure 14).
[Refer to caption](/html/0706.2520/assets/x14.png)
Figure 14: Left column - behavior of $`u^{497}`$ during time period from $`48`$h to $`60`$h with $`6`$h time window, induced correlation starts at $`54`$h and lasts for $`3`$h. Right column - behavior of $`u^{496}`$ in same conditions.
*Experiment 2*
In the previous experiment we injected just one type of increased correlation among time series. Now we make two and three different types of induced correlations produce different spatial-temporal patterns on eigenvector $`u^{497}`$ components (see Figure 15). Time series for temporal increase of correlation are obtained in the same way as in Experiment 1. We temporarily increased the correlation between series by inducing elements from distributions of sine function and quadratic function, respectively for three hours.
[Refer to caption](/html/0706.2520/assets/x15.png)
Figure 15: (a) The weights of components of $`u^{497}`$ plotted for time period from $`36`$ to $`84`$ hours with $`6`$ hours interval, two different types of induced correlations. (b) The weights of components of $`u^{497}`$ plotted with respect to the same time period, three different types of induced correlations.
In Figure 15a, one type of three hours correlation is induced among ten traffic time series and another type of correlation among other ten time series. Three different types of three hours correlations are induced among twenty traffic time series in Figure 15b. The sorted in decreasing order content of significant components shows that time series tend to group according to the type of correlation they are involved in.
*Experiment 3*
Next we turn our attention to disruption of normal picture of inter-VLAN traffic interactions. This can be done by injecting the traffic from random distribution to non-randomly interacting time series for three hours. We demonstrate it by examining the eigenvalue distribution, the IPR and the deviating eigenvectors overlap matrix plotted in Figure 16.
[Refer to caption](/html/0706.2520/assets/x16.png)
Figure 16: Eigenvalues distribution, IPR and overlap matrix of deviating eigenvectors of inter-VLAN traffic cross-correlation matrix $`C`$.
After $`60`$ hours of uninterrupted traffic, we injected elements from random distribution to significant participants of $`u^{497}`$ for three hours. The largest eigenvalue increases, from $`10`$ to $`12.`$ Extended IPR tail shows the larger number of *localized* eigenvectors and we observe the dramatic break in deviating eigenvectors stability.
## IX conclusion and future work
The RMT methodology we used in this paper enables us to analyze the complex system behavior without the consideration of system constraints, type and structure. Our goal was to investigate the characteristics of day-to-day temporal dynamics of the system of interconnected routers with VLAN subnets of the University of Louisville. The type and structure of the system at hand suggests the natural interpretation of the RMT-like behavior and the RMT deviating results. The time stable random interactions signify the healthy, and free of congestion traffic. The time stable non-random interactions provide us with information about large-scale network-wide traffic interactions. The changes in the stable picture of random and non-random interactions signify the temporal traffic anomalies.
In general, the fact of sharing the universal properties of the bulk of eigenvalues spectrum of inter-VLAN traffic interactions with random matrices opens a new venue in network-wide traffic modeling. As stated in \[[19](#bib.bib19)\], in physical systems it is common to start with the model of dynamics of the system. This way, one would model the traffic time series interactions with the family of stochastic differential equations \[[28](#bib.bib28), [29](#bib.bib29)\], which describe the “instantaneous” traffic counts
<table>
<tbody>
<tr>
<td></td>
<td>gi(t)=(d/dt)lnTi(t),subscript𝑔𝑖𝑡𝑑𝑑𝑡𝑙𝑛subscript𝑇𝑖𝑡g_{i}\left(t\right)=\left(d/dt\right)lnT_{i}\left(t\right),</td>
<td></td>
<td>(22)</td>
</tr>
</tbody>
</table>
as a random walk with couplings. Then one would relate the revealed interactions to the correlated “modes” of the system.
Additional question that RMT findings raise in network-wide traffic analysis is whether the found eigenvalues spectrum correlations and *localized* eigenvectors outside of RMT bulk can add to the explanation of the fundamental property of the network traffic, such as self-similarity \[[30](#bib.bib30)\].
To summarize, we have tested the eigenvalues statistics of inter-VLAN traffic cross-correlation matrix $`C`$ against the null hypothesis of random correlation matrix. By separating the eigenvalues spectrum correlations of random matrices that are present in this system, the uncongested state of the network traffic is verified. We analyzed the stable in time system-specific correlations. The analyzed eigenvalues and eigenvectors deviating from the RMT showed the principal groups of VLAN-router switches, groups of traffic time series interlinked through the firewalls and groups of same service VLANs at every router. With straightforward experiments on the traffic time series, we demonstrated that eigenvalue distribution, IPR of eigenvectors, overlap matrix and spatial-temporal patterns of deviating eigenvectors can monitor the stability of inter-VLAN traffic interactions, detect and spot in time and space of any network-wide changes in normal traffic time series interactions.
As the reservation for the future work, we would like to investigate the behavior of delayed traffic time series cross-correlation matrix $`C_{d}`$ in the RMT terms. The importance of delay in measurement-based analysis of Internet is emphasized in \[[31](#bib.bib31)\]. To understand and quantify the effect of one time series on another at a later time, one can calculate the delay correlation matrix, where the entries are cross-correlation of one time series and another at a time delay $`\tau`$ \[[32](#bib.bib32)\]. In addition, we are interested in testing the fruitfulness of the RMT approach on the larger system of inter-domain interactions, for instance, on 5-minute averaged traffic count time series of underlying backbone circuits of Abilene backbone network.
## acknowledgment
This research was partially supported by a grant from the US Department of Treasury through a subcontract from the University of Kentucky. The authors thank Igor Rozhkov for consulting on the RMT methodology. We thank Hans Fiedler, University of Louisville network manager, for MRTG data of UofL routers system used in this study and helpful suggestions in network interpretations of our results. We are grateful to Nathan Johnson, University of Louisville super computing administrator, for providing the computing environment and space.
## References
- \[1\] K. Fukuda, PhD Thesis: A study on phase transition phenomena in internet traffic, Keio University, 1999.
- \[2\]
T. Ohira, R. Sawatari, Phase transition in a computer network traffic model, Phys. Rev. E 58, July 1998, 193-195.
- \[3\]
M. Barthelemy, B. Gondran and E. Guichard, Large scale cross-correlations in internet traffic, arXiv:cond0mat/0206185 vol 2 3 Dec 2002.
- \[4\] A. Lakhina, M. Crovella, and C. Diot, Detecting distributed attacks using network-wide flow traffic, Proceedings of FloCon 2005 Analysis Workshop, 2005.
- \[5\] A. Lakhina, M. Crovella, and C. Diot. Mining Anomalies Using Traffic Feature Distributions. Technical Report BUCS-TR-2005-002, Boston University, 2005.
- \[6\]
E.P. Wigner, On a class of analytic functions from the quantum theory of collisions, Ann. Math. 53, 36 (1951), Proc. Cambridge Philos. Soc. 47, 790 (1951).
- \[7\]
F. Dyson, Statistical theory of the energy levels of complex systems, J. Math. Phys. 3, 140 (1962).
- \[8\]
F. Dyson and M.L. Mehta, Statistical theory of the energy levels of complex systems, J. Math. Phys. 4, 701, 713 (1963).
- \[9\] M.L Mehta, Random matrices (Academic Press, Boston, 1991).
- \[10\]
T.A. Brody, J.Flores, J.B. French, P.A. Mello, A. Pandey, and S.S.M. Wong, Random-matrix physics: spectrum and strength fluctuations, Rev. Mod. Phys. 53, 385 - 479, issue 3, July 1981.
- \[11\]
T. Guhr, A. Muller-Groeling, and H.A. Weidenmuller, Random matrix theories in quantum physics: common concepts, Phys. Rep. 299, 190 (1998).
- \[12\]
M. Krbalek and P.Seba, Statistical properties of the city transport in Cuernavaca (Mexico) and random matrix theory. J. Phys. 214 (2000), 1, 91-100.
- \[13\] J. McNutt and M. De Shon, Correlation between quiescent ports in network flows, CERT network situational awareness group report, Carnegie Mellon University, September 2005.
- \[14\] A. Lakhina, M. Crovella, and C. Diot, Characterization of network-wide anomalies in traffic flows, Proceedings of the ACM/SIGCOMM Internet Measurement conference, 2004, 201-206.
- \[15\]
L. Min, Y. Shun-Zheng, A network-wide traffic anomaly detection method based on HSMM, Int. conf. on communications, circuits and system proceedings, vol 6, June 2006, 1636 - 1640.
- \[16\] M. Roughan, T. Griffin, M. Mao, A. Greenberg, and B. Freeman, Combining routing and traffic data for detection of IP forwarding anomalies, Proceedings of the joint int. conf. on Measurement and modeling of computer systems, 2004, 416 - 417.
- \[17\] L. Huang, X. Nguyen, M. Garofalakis, M. Jordan, A. Joseph and N. Taft, Distributed PCA and network anomaly detection, Technical report No. UCB/EECS-2006-99.
- \[18\]
S. Sharifi, M. Crane, A. Shamaie and H. Ruskin, Random matrix portfolio optimization: a stability approach, Physica A 335 (2004) 629-643.
- \[19\]
V. Plerou, P. Gopikrishnan, B. Rosenow, L. A. Nunes Amaral, T. Guhr, and H.E. Stanley, Random matrix theory approach to cross correlations in financial data, Phys. Rev. E, vol 65, 066126, 27 June 2002.
- \[20\]
A. Tulino and S. Verdu, Random matrix theory and wireless communications, Communications and Information theory, vol 1, issue 1, June 2004, 1 - 182.
- \[21\] D. Tse, Multiuser receivers, random matrices and free probability, Proceedings of 37th Ann. Allerton Conf., Monticello, IL, September 1999.
- \[22\]
A. Zee, Random matrix theory and RNA folding, Acta Physica Polonica B, vol 36, No 9, June 2005.
- \[23\]
L. Laloux, P. Cizeau, J.-P. Bouchaud, and M. Potters, Noise dressing of financial correlation matrices, Phys. Rev. Lett. 83, August 1999, 1467-1470.
- \[24\]
A.M. Sengupta and P.P. Mitra, Distributions of singular values for some random matrices, arXiv:cond-mat/9709283 vol 1 25 September 1997.
- \[25\] H.-J. Stockman, Quantum Chaos: an introduction, 1999.
- \[26\] J.-P. Bouchaud, Theory of financial risk and derivative pricing: from statistical physics to risk management, 1962.
- \[27\]
H. Bruus and J.-C. Angles d’Auriac, Energy level statistics of two-dimensional Hubbard model at low filling, arXiv:cond-mat/9610142 vol 1 18 October 1996.
- \[28\]
J.D. Farmer, Market Force, ecology and evolution, e-print adap-org/9812005, Int. J. Theo. Appl. fin. 3, 425, 2000.
- \[29\]
J.-P. Bouchaud, R. Cont, A Langevin approach to stock market fluctuations and crashes, European Journal of Physics, B 6, 543, 1998.
- \[30\] W.E. Leland, M.S. Taqq, W. Willinger, and D.V. Willson, On the self-similar nature of Ethernet traffic, ACM SIGCOMM, 1993, 183 - 193.
- \[31\] B. Zhang, T.S. Eugene Ng, and A. Nandi, Measurement-based analysis, modeling, and synthesis of the Internet delay space, Proceedings of the 6-th ACM SIGCOMM on Internet Measurement, 2006, 85-98.
- \[32\] K.B.K. Mayya and R.E. Amritkar, Analysis of delay correlation matrices, oai:arXiv.org:cond-mat/0601279 (2006-12-20).
- \[33\]
W.-C. Lau, S.-Q. Li, Traffic analysis in large-scale high-speed integrated networks:validation of nodal decomposition approach, INFOCOM, 1993, Proceedings of twelfth annual joint conference of the IEEE Computer and Communications Societies, vol 3, 1320-1329.
- \[34\] W.H. Allen, G.A. Marin, L.A. Rivera, Automated detection of malicious reconnaissance to enhance network security, SoutheastCon, 2005, Proceedings of IEEE, issue 8-10, April 2005, 450-454.
- \[35\]
H. Bruus and J.-C. Angles d’Auriac, The spectrum of two-dimensional Hubbard model at low filling, Europhysics letters, 35 (5), 321-326, 1999.
## Appendix A RMT
In this Appendix, we provide a short (and non-rigorous) explanation of main concepts and glossary of terms used in the RMT studies. The RMT approaches, which originated in nuclear and condensed matter physics and later became common in many branches of mathematical physics \[[25](#bib.bib25)\], have recently penetrated into econophysics, finance \[[26](#bib.bib26)\] and network traffic analysis \[[3](#bib.bib3)\].
For the statistical description of complex physical systems, such as, for example, atomic nucleus or acoustical reverberant structure, the RMT serves as guiding light when one is interested in the degree of mutual interaction of the constituents. As it turns out, the uncorrelated energy levels or acoustic eigenfrequencies would produce qualitatively different result from those obeying RMT-like correlations \[[25](#bib.bib25)\]. Therefore, real (experimentally measured) spectra can help to decide on the nature of interactions in the underlying system. To be specific, ideally, symmetric system is expected to exhibit spectral properties drastically different from the properties of generic one, and if the spectral properties are those of RMT systems, other ideas of RMT can be brought to the researcher aid.
To describe “awareness” of the structural constituents about each other, scientists in different fields use similar constructs. Physicists use Hamiltonian matrix, engineers stiffness matrix, finance and network analysts the equal-time cross-correlation matrix. Although the physical meaning of mentioned operators can be different, the eigenvalues/eigenvectors analysis seems to be a universally accepted tool. The eigenvalues have direct connection to spectrum of physical systems, while eigenvectors can be used for the description of excitation/signal/information propagation inside the system. In physics, the RMT approaches come about whenever the system of interest demonstrates certain qualitative features in their spectral behavior. For example, if one looks at nearest neighbor spacing distribution of eigenvalues and instead of Poisson law
<table>
<tbody>
<tr>
<td></td>
<td>P(s)=exp(−s),𝑃𝑠𝑠P\left(s\right)=\exp\left(-s\right),</td>
<td></td>
</tr>
</tbody>
</table>
discovers “Wigner surmise”
<table>
<tbody>
<tr>
<td></td>
<td>P(s)=π2sexp(−π2s2),𝑃𝑠𝜋2𝑠𝜋2superscript𝑠2P\left(s\right)=\frac{\pi}{2}s\exp\left(-\frac{\pi}{2}s^{2}\right),</td>
<td></td>
</tr>
</tbody>
</table>
one concludes (upon running several additional statistical tests) that apparatus of RMT can be used for the system at hand, and system matrix can be replaced by a matrix with random entries. For mathematical convenience, these entries are given Gaussian weight. The only other ingredient of this rather succinct phenomenological model is recognizing the physical situation. For example, systems with and without magnetic field and/or central symmetry are described by different matrix ensembles (that is the set of matrices) with elements distributed within distribution corresponding to the same $`\beta`$
<table>
<tbody>
<tr>
<td></td>
<td>P(β)(H)∝exp(−β4v2trH2),proportional-tosuperscript𝑃𝛽𝐻exp𝛽4superscript𝑣2𝑡𝑟superscript𝐻2P^{\left(\beta\right)}\left(H\right)\propto\textrm{exp}\left(-\frac{\beta}{4v^{2}}trH^{2}\right),</td>
<td></td>
</tr>
</tbody>
</table>
where the constant $`v`$ sets the length of the resulting eigenvalues spectrum.
The very fact that RMT can be helpful in statistical description of the broad range of systems suggests that these systems are analyzed in a certain special *universal* regime, in which physical or other laws are undermined by equilibrated and ergodic evolution. In most physical applications, a Hamiltonian matrix is rather sparse, indicating lack of interaction between different subparts of the corresponding object. However, if the universal regime is inferred from the above mentioned statistical tests, it is very beneficial to replace this single matrix with the ensemble of random matrices. Then, one can proceed with statistical analysis using matrix ensemble for calculation of statistical averages more relevant for the physical problem at hand than the statistics of eigenvalues. The latter can be mean or variance of the response to external or internal excitation. |
0706.2010 | cs.CR | cyber | # Information-theoretic security without an honest majority
Anne Broadbent and Alain Tapp
Département d’informatique et de recherche opérationnelle
Université de Montréal, C.P. 6128, Succ. Centre-Ville
Montréal (QC), H3C 3J7 Canada
[{broadbea,tappa}@iro.umontreal.ca](%7Bbroadbea,tappa%7D@iro.umontreal.ca)
(June 13, 2007)
###### Abstract
We present six multiparty protocols with information-theoretic security that tolerate an arbitrary number of corrupt participants. All protocols assume pairwise authentic private channels and a broadcast channel (in a single case, we require a simultaneous broadcast channel). We give protocols for *veto*, *vote*, *anonymous bit transmission*, *collision detection*, *notification* and *anonymous message transmission*. Not assuming an honest majority, in most cases, a single corrupt participant can make the protocol abort. All protocols achieve functionality never obtained before without the use of either computational assumptions or of an honest majority.
Keywords: multiparty computation, anonymous message transmission, election protocols, collision detection, dining cryptographers, information-theoretic security.
## 1 Introduction
In the most general case, *multiparty secure computation* enables $`n`$ participants to collaborate to compute an $`n`$-input, $`n`$-output function (one per participant). Each participant only learns his private output which, depending on the function, can be the same for each participant. Assuming that private random keys are shared between each pair of participants, we known that every function can be securely computed in the presence of an active adversary if and only if less than $`n/3`$ participants are corrupt; this fundamental result is due to Michael Ben-Or, Shafi Goldwasser and Avi Wigderson \[[BGW88](#bib.bibx2)\] and David Chaum, Claude Crépeau and Ivan Damg\arard \[[CCD88](#bib.bibx3)\]. When a broadcast channel is available, the results of Tal Rabin and Michael Ben-Or \[[RB89](#bib.bibx11)\] tell us that this proportion can be improved to $`n/2`$.
Here, we present six specific multiparty computation protocols that achieve correctness and privacy *without* any assumption on the number of corrupt participants. Naturally, we cannot always achieve the ideal functionality, for example in some cases, a single participant can make the protocol abort. This is the price to pay to tolerate an arbitrary number of corrupt participants and still provide information-theoretic privacy of the inputs.
All protocols we propose have polynomial complexity in the number of participants and the security parameter. We always assume pairwise shared private random keys between each pair of participants, which allows pairwise private authentic channels. We also assume a broadcast channel and, even though it is a strong assumption, in some cases we need the broadcast to be simultaneous \[[CGMA85](#bib.bibx6), [HM05](#bib.bibx9)\].
### 1.1 Summary of results
Our main contributions are in the areas of elections (*vote*) and anonymity (*anonymous bit transmission* and *anonymous message transmission*). Each protocol is an astute combination of basic protocols, which are also of independent interest, and that implement *parity*, *veto*, *collision detection* and *notification*.
The main ingredient for our information-theoretically secure protocols is the dining cryptographers protocol \[[Cha88](#bib.bibx8)\] (see also Section [2](#S2 "2 Parity ‣ Information-theoretic security without an honest majority")), to which we add the following simple yet powerful observation: if $`n`$ participants each hold a private bit of an $`n`$-bit string with Hamming weight of parity $`p`$, then any single participant can randomize $`p`$ by locally flipping his bit with a certain probability. It is impossible, however, for any participant to locally derandomize $`p`$. In the case of the anonymous message transmission, we also build on the dining cryptographers protocol by noting that a message that is sent can be ciphered with a one-time pad by having one participant (the receiver) broadcast a random bit. Any modification of the message can then be detected by the receiver with an *algebraic manipulation detection code* \[[CFP07](#bib.bibx4)\].
#### 1.1.1 Vote
Our *vote* protocol (Section [4](#S4 "4 Vote ‣ Information-theoretic security without an honest majority")) allows $`n`$ participants to conduct an $`m`$-candidate election. The privacy is perfect but the protocol has the drawback that if it aborts (any corrupt participant can cause an abort), the participants can still learn information that would have been available had the protocol succeeded. For this protocol, we require a simultaneous broadcast channel. It would be particularly well-suited for a small group of voters that are unwilling to trust any third party and who have no advantage in making the protocol abort.
Previous work on information-theoretically secure voting protocols include \[[CFSY96](#bib.bibx5)\], where a protocol is given in the context where many election authorities are present. To the best of our knowledge, our approach is fundamentally different from any other approaches for voting. It is the first to provide information-theoretic security without requiring or trusting any third party, while also providing ballot casting assurance (each participant is convinced that their input is correctly recorded \[[AN06](#bib.bibx1)\]) and universal verifiability (each participant is conviced that only registered voters cast ballots and that the tally is correctly computed \[[SK95](#bib.bibx12)\]).
#### 1.1.2 Anonymity
Anonymity is the power to perform a task without having to identify the participants that are involved. In the case of *anonymous message transmission*, it is simply the capacity of the sender to transmit a private message to a specific receiver of his choosing without revealing either his identity or the identity of the receiver. A number of protocols have been suggested for anonymous transmission. Many of these rely on trusted or semi-trusted third parties as well as computational assumptions (for instance, the MIX-net \[[Cha81](#bib.bibx7)\]). Here, we do not make any such assumptions. The most notable protocol for anonymous transmission in our context is the dining cryptographer’s protocol \[[Cha88](#bib.bibx8)\], which allows a single sender to anonymously broadcast a bit, and provides information-theoretical security against a passive adversary. We present the protocol in a version that implements the multiparty computation of the *parity* function in Section [2](#S2 "2 Parity ‣ Information-theoretic security without an honest majority").
The case of multiple yet honest senders in the dining cryptographer’s protocol can be solved by time slot reservation techniques, as originally noted by Chaum \[[Cha88](#bib.bibx8)\]. But nevertheless, any corrupt participant can jam the channel. Techniques offering computational security to this problem have been proposed \[[Cha88](#bib.bibx8), [WP89b](#bib.bibx14)\]. Also, computational assumptions allow the removal of the reliance on a broadcast channel \[[WP89a](#bib.bibx13)\].
In our implementation of *anonymous bit transmission* (Section [5](#S5 "5 Anonymous Bit Transmission ‣ Information-theoretic security without an honest majority")), we elegantly deal with the case of multiple senders by allowing an unlimited amount of participants to act as anonymous senders. Each anonymous sender can target any number of participants and send them each a private bit of his choice. Thus, the outcome of the protocol is, for each participant, a private list indicating how many $`0`$s and how many $`1`$s were received. The anonymity of the sender and receiver and the privacy of all transmitted bits is always perfectly achieved, but any participant can cause the protocol to abort, in which case the participants may still learn some information about their own private lists.
We need a way for all participants to find out if the protocol has succeeded. This is done with the *veto* protocol (Section [3](#S3 "3 Veto ‣ Information-theoretic security without an honest majority")), which takes as input a single bit from each participant; the output of the protocol is the logical OR of the inputs. Our implementation differs from the ideal functionality since a participant that inputs $`1`$ will learn if some other participant also input $`1`$. We make use of this deviation from the ideal functionality in further protocols.
In our *fixed role anonymous message transmission* protocol (Section [8](#S8 "8 Fixed Role Anonymous Message Transmission ‣ Information-theoretic security without an honest majority")), we present a method which allows a single sender to communicate a message of arbitrary length to a single receiver. To the best of our knowledge, this is the first protocol ever to provide perfect anonymity, message privacy and integrity. For a fixed security parameter, the anonymous message transmission is asymptotically optimal.
Our final protocol for *anonymous message transmission* (Section [9](#S9 "9 Anonymous Message Transmission ‣ Information-theoretic security without an honest majority")) allows a sender to send a message of arbitrary length to a receiver of his choosing. While any participant can cause the protocol to abort, the anonymity of the sender and receiver is always perfectly achieved. The privacy of the message is preserved except with exponentially small probability. As far as we are aware, all previous proposed protocols for this task require either computational assumptions or a majority of honest participants. The protocol deals with the case of multiple senders by first executing the *collision detection* protocol (Section [6](#S6 "6 Collision Detection ‣ Information-theoretic security without an honest majority")), in which each participant inputs a single bit. The outcome only indicates if the sum of the inputs is $`0`$, $`1`$ or more. Compared to similar protocols called *time slot reservation* \[[Cha88](#bib.bibx8), [WP89b](#bib.bibx14)\], our protocol does not leak any additional information about the number of would-be senders. The final protocol also makes use of the *notification* protocol (Section [7](#S7 "7 Notification ‣ Information-theoretic security without an honest majority")) in which each participant chooses a list of other participants that are to be notified. The output privately reveals to each participant the logical OR of his received notifications. A special case of this protocol is when a single participant notifies another single participant; this is the version used in our final protocol to enable the sender to anonymously tell to the receiver to act accordingly.
### 1.2 Common features to all protocols
All protocols presented in the following sections share some common features, which we now describe. Our protocols are given in terms of multiparty computation with inputs and outputs and involve $`n`$ participants, indexed by $`i = {1,\ldots,n}`$. In the ideal functionality, the only information that the participants learn is their output (and what can be deduced from it). *Correctness* refers to the fact that the outputs are correctly computed, while *privacy* ensures that the inputs are never revealed.
The protocols ensure correctness and privacy even in the presence of an unlimited number of misbehaving participants. Two types of such behaviour are relevant: participants who collude (they follow the protocol but pool their information in order to violate the protocol’s privacy), and participants who actively deviate from the protocol (in order to violate the protocol’s correctness or privacy). Without loss of generality, these misbehaviours are modelled by assuming a central adversary that controls some participants, rendering them *corrupt*. The adversary is either *passive* (it learns all the information held by the corrupt participants), or *active* (it takes full control of the corrupt participants). We will deal only with the most general case of active adversaries, and require them to be *static* (the set of corrupt participants does not change). A participant that is not corrupt is called *honest*. Our protocols are such that if they do not abort, there exists inputs for the corrupt participants that would lead to the same output if they were to act honestly. If a protocol aborts, the participants do not learn any more information than they could have learned in an honest execution of the protocol. The input and output description applies only to honest participants.
We assume that each pair of participants shares a private, uniformly random string that can be used to implement an authentic private channel. The participants have access to a broadcast channel and in some cases, it is simultaneous. A *broadcast* channel is an authentic broadcast channel for which the sender is confident that all participants receive the same value and the receivers know the identity of the sender. A *simultaneous broadcast* channel is a collection of broadcast channels where the input of one participant cannot depend on the input of any other participant. This could be achieved if all participants simultaneously performed a broadcast. In order to distinguish between the two types of broadcast, we sometimes call the broadcast channel a *regular* broadcast. It is not uncommon in multiparty computation to allow additional resources, even if these resources cannot be implemented with the threshold on the honest participants (the results of \[[RB89](#bib.bibx11)\] which combine a broadcast channel with $`n/2`$ honest participants being the most obvious example). Our work suggests that a simultaneous broadcast channel is an interesting primitive to study in this context.
In all protocols, the security parameter is $`s`$. Unfortunately, in many of our protocols, a single corrupt participant can cause the protocol to abort. All protocols run in polynomial time with respect to the number of participants, the security parameter and the input length. Although some of the protocols presented in this paper are efficient, our main focus here is in the *existence* of protocols for the described tasks. We leave for future work improvement of their efficiency. Finally, due to lack of space, we present only sketches of security proofs.
## 2 Parity
Protocol [1](#protocol1 "Protocol 1 ‣ 2 Parity ‣ Information-theoretic security without an honest majority") implements the *parity* function and is essentially the same as the dining cryptographers protocol \[[Cha88](#bib.bibx8)\], with the addition of a simultaneous broadcast channel. Note that if we used a broadcast channel instead, then the last participant to speak would have the unfair advantage of being able to adapt his input in order to fix the outcome of the protocol!
Protocol 1
Parity
Input: $`x_{i} \in {\{ 0,1\}}`$
Output: $`y_{i} = {x_{1} \oplus x_{2} \oplus \cdots \oplus x_{n}}`$
Broadcast type : simultaneous broadcast
Achieved functionality:
1) (Correctness) If the protocol does not abort, the output is the same as in the ideal functionality.
2) (Privacy) No adversary can learn more than the output of the ideal functionality. Each participant $`i`$ does the following:
1. 1\.
Select uniformly at random an $`n`$-bit string $`r_{i} = {r_{i}^{1}\hspace{0pt}r_{i}^{2}\hspace{0pt}\ldots\hspace{0pt}r_{i}^{n}}`$ with Hamming weight of parity $`x_{i}`$.
2. 2\.
Send $`r_{i}^{j}`$ to participant $`j`$ using the private channel; keep bit $`r_{i}^{i}`$ to yourself.
3. 3\.
Compute $`z_{i}`$, the parity of the sum of all the bits received, including $`r_{i}^{i}`$.
4. 4\.
Use the simultaneous broadcast channel to announce $`z_{i}`$.
5. 5\.
After the simultaneous broadcast is finished, compute $`y_{i} = {\bigoplus_{k = 1}^{n}z_{k}}`$. This is the outcome of the protocol. If the simultaneous broadcast fails, abort the protocol.
Correctness and privacy follows from \[[Cha88](#bib.bibx8)\]. Thus, any adversary can learn only what can be deduced from the corrupt participant’s inputs and the outcome of the protocol. Note that this means that the adversary can deduce the parity of the inputs of the other participants. We will later use the two simple observations that there is no way to cheat except by refusing to broadcast and that any value that is broadcast is consistent with a choice of valid inputs. In the following protocols, we will adapt step [4](#S2.I1.i4 "item 4 ‣ Protocol 1 ‣ 2 Parity ‣ Information-theoretic security without an honest majority") of the parity protocol to make it relevant to the scenario, this will allow us to remove the assumption of the simultaneous broadcast. We will also use the fact that if a single participant either does not broadcast, or broadcasts a random bit in step [4](#S2.I1.i4 "item 4 ‣ Protocol 1 ‣ 2 Parity ‣ Information-theoretic security without an honest majority") then the value of the output of parity is known to this participant, but is perfectly hidden to all other participants.
## 3 Veto
In this section, we build on the parity protocol to give a protocol for the secure implementation of the *veto* function, which computes the logical OR of the participant’s inputs (Protocol [2](#protocol2 "Protocol 2 ‣ 3 Veto ‣ Information-theoretic security without an honest majority")). We also discuss some important deviations from the ideal functionality.
Protocol 2
Veto
Input: $`x_{i} \in {\{ 0,1\}}`$
Output: $`y_{i} = {x_{1} \vee x_{2} \vee \cdots \vee x_{n}}`$
Broadcast type : regular broadcast
Achieved functionality:
1) (Reliability) No participant can make the protocol abort.
2) (Correctness) The outcome of the protocol equals the outcome of the ideal functionality.
3) (Privacy) Any adversary learns what it would have learned in the ideal functionally, had the corrupt participants used $`0`$ as input, but nothing more.
The $`n`$ participants agree on $`n`$ orderings such that each ordering has a different last participant.
$`\text{result}\leftarrow 0`$
For each ordering,
—-
Repeat $`s`$ times:
1. 1\.
Each participant $`i`$ sets the value of $`p_{i}`$ in the following way: if $`x_{i} = 0`$ then $`p_{i} = 0`$; otherwise, $`p_{i} = 1`$ with probability $`\frac{1}{2}`$ and $`p_{i} = 0`$ with complimentary probability.
2. 2\.
The participants execute the parity protocol with inputs $`p_{1},p_{2},{\ldots\hspace{0pt}p_{n}}`$, with the exception that the simultaneous broadcast is replaced by a regular broadcast with the participants broadcasting according to the current ordering (if any participant refuses to broadcast, set the value $`\text{result}\leftarrow 1`$). If the outcome of parity is $`1`$, then set $`\text{result}\leftarrow 1`$ .
Output the value result.
######
Lemma 3.1
.
(Reliability) No participant can make the veto protocol abort.
###### Proof.
If a participant refuses to broadcast, it is assumed that the output of the protocol is $`1`$. ∎
######
Lemma 3.2
.
(Correctness) If all participants in the veto protocol have input $`x_{i} = 0`$, then the protocol achieves the ideal functionality with probability 1. If there exists a participant with input $`x_{i} = 1`$ then the protocol is correct with probability at least $`1 - 2^{- s}`$.
###### Proof.
The correctness follows by the properties of the parity protocol, with the difference that we now have a broadcast channel instead of a simultaneous broadcast channel. The case where all inputs are 0 is trivial. Let $`x_{i} = 1`$ and suppose that the protocol is executed until the ordering in which participant $`i`$ speaks last. Then with probability at least $`1 - 2^{- s}`$, in step [2](#S3.I1.i2 "item 2 ‣ Protocol 2 ‣ 3 Veto ‣ Information-theoretic security without an honest majority") of veto, the output of the protocol will be set to $`1`$. ∎
######
Lemma 3.3
.
(Privacy) In the veto protocol, the most an adversary can learn is the information that it could have learned by assigning to all corrupt participants the input $`0`$. Additionally, this information is revealed, even to a passive adversary, with probability at least $`1 - 2^{- s}`$.
###### Proof.
This follows from the properties of the parity protocol: for a given repetition, the adversary learns the parity of the honest participants’ $`p_{i}`$’s, but nothing else. Because of the way that the $`p_{i}`$’s are chosen in step [1](#S3.I1.i1 "item 1 ‣ Protocol 2 ‣ 3 Veto ‣ Information-theoretic security without an honest majority"), if for any repetition, this parity is odd, the adversary concludes that at least one honest participant has input $`1`$, and otherwise if all repetitions yield $`0`$, then the adversary concludes that with probability at least $`1 - 2^{- s}`$, all the honest participant’s inputs are $`0`$. In all cases, this is the only information that is revealed; clearly, it is revealed to any passive adversary, except with exponentially small probability. ∎
## 4 Vote
The participants now wish to conduct an $`m`$-candidate vote. The idea of Protocol [3](#protocol3 "Protocol 3 ‣ 4 Vote ‣ Information-theoretic security without an honest majority") is simple. In the veto protocol, each participant with input $`1`$ completely randomizes his input into the parity protocol, thus randomizing the output of parity. By flipping the output of parity with probability only $`1/n`$, the probability of the outcome being odd becomes a function of the number of such flips. Using repetition, this probability can be approximated to obtain the exact number of flips with exponentially small error probability. This can be used to compute the number of votes for each candidate. Unfortunately, a corrupt participant can randomize his bit with probability higher than $`1/n`$, enabling him to vote more than once. But since a participant cannot derandomize the parity, he cannot vote less than zero times. Verifying that the sum of the votes equals $`n`$ ensures that all participants vote exactly once. Note that the protocol we present is polynomial in $`m`$ and not in the length of $`m`$.
Protocol 3
Vote
Input: $`x_{i} \in {\{ 1,\ldots,m\}}`$
Output: for $`k = 1`$ to $`m`$, $`{y\hspace{0pt}{\lbrack k\rbrack}_{i}} = {|{\{ x_{j}\mid{x_{j} = k}\}}|}`$
Broadcast type : simultaneous broadcast
Achieved functionality:
1) (Correctness) If the protocol does not abort, then there exists an input $`x_{i}`$ for each corrupt participant such that the protocol achieves the ideal functionality.
2) (Privacy) Even if the protocol aborts, no adversary can learn more that what it would have learned by setting in the ideal functionality $`x_{i} = 1`$ for all corrupt participants.
Phase A
For each candidate $`k = 1`$ to $`m`$,
For $`j = 1`$ to $`s`$,
1. 1\.
Each participant $`i`$ sets the value of $`p_{i}`$ in the following way: if $`x_{i} \neq k`$, then $`p_{i} = 0`$; otherwise, $`p_{i} = 1`$ with probability $`\frac{1}{n}`$ and $`p_{i} = 0`$ with complimentary probability.
2. 2\.
The participants execute the parity protocol to compute the *parity* of $`p_{1},p_{2},{\ldots\hspace{0pt}p_{n}}`$, but instead of broadcasting their output bit $`z_{i}`$, they store it as $`z\hspace{0pt}{\lbrack k\rbrack}_{i}^{j}`$.
Phase B
All participants simultaneously broadcast $`z\hspace{0pt}{\lbrack k\rbrack}_{i}^{j}`$ $`({j = {1,2,\ldots,s}})`$. If the simultaneous broadcast is not successful, the protocol aborts.
Phase C
To compute the tally, $`y\hspace{0pt}{\lbrack k\rbrack}_{i}`$, for each value $`k = {1\hspace{0pt}\ldots\hspace{0pt}m}`$, each participant sets: $`{p\hspace{0pt}{\lbrack k\rbrack}_{j}} = {\bigoplus_{i = 1}^{n}{z\hspace{0pt}{\lbrack k\rbrack}_{i}^{j}}}`$, $`{\sigma\hspace{0pt}{\lbrack k\rbrack}_{i}} = {\sum_{j = 1}^{s}{{p\hspace{0pt}{\lbrack k\rbrack}_{j}}/s}}`$ and if there exists an integer $`v`$ such that $`\left| {{\sigma\hspace{0pt}{\lbrack k\rbrack}_{i}} - p_{v}} \right| < \frac{1}{2\hspace{0pt}e^{2}\hspace{0pt}n}`$,
where $`p_{v} = {\frac{1}{2}\hspace{0pt}\left( \frac{n - 2}{n} \right)^{v}\hspace{0pt}\left( {\left( \frac{n}{n - 2} \right)^{v} - 1} \right)}`$, then $`{y\hspace{0pt}{\lbrack k\rbrack}_{i}} = v`$ .
If for any $`m`$, no such value $`v`$ exists, or if $`{\sum_{k = 1}^{m}{y\hspace{0pt}{\lbrack k\rbrack}_{i}}} \neq n`$, the protocol aborts.
######
Lemma 4.1
.
(Correctness) If the vote does not abort, then there exists an input for each corrupt participant such that the output of the honest participants equals the output of the ideal functionality, except with probability exponentially small in $`s`$.
###### Proof.
If all participants are honest, the correctness of the protocol is derived from the Chernoff bound as explained in the Appendix. Assume now $`t`$ corrupt participants. Since the parity protocol is perfect, the only place participant $`i`$ can deviate from the protocol is by choosing $`p_{i}`$ with an inappropriate probability. We first note that if the $`t`$ corrupt participants actually transmit the correct number of private bits in phase A and broadcast the correct number of bits in phase B, then whatever they actually send is consistent with some global probability of flipping.
We use again the fact that it is possible to randomize the parity but not to derandomize it: if the corrupt participants altogether flip with a probability not consistent with an integer number of votes, either the statistics will be inconsistent, causing the protocol to abort, or we can interpret the results as being consistent with an integer amount of votes. If they flip with a probability consistent with an integer different than $`t`$, then each $`y\hspace{0pt}{\lbrack k\rbrack}_{i}`$ will be assigned a value, but with probability exponentially close to 1, we will have $`{\sum_{k = 1}^{m}{y\hspace{0pt}{\lbrack k\rbrack}_{i}}} \neq n`$ and the protocol will abort. ∎
######
Lemma 4.2
.
(Privacy) In the vote protocol, no adversary can learn more than what it would have learned by assigning to all corrupt participants the input $`1`$ in the ideal functionality, and this even if the protocol aborts.
###### Proof.
Assume that the first $`t`$ participants are corrupt. No information is sent phase A or phase C. We thus have to concentrate on phase B where the participants broadcast their information regarding each parity. For each execution of parity, the adversary learns the parity of the honest participant’s values, $`p_{t + 1} \oplus p_{t + 2} \oplus \ldots \oplus p_{n}`$, but no information on these individual values is revealed. The adversary can thus only evaluate the probability with which the other participants have flipped the parity. But this information could be deduced from the output of the ideal functionality, for instance by fixing the corrupt participants’ inputs to $`1`$. ∎
It is important to note that the above results do not exclude the possibility of an adversary causing the protocol to abort while still learning some information as stipulated in Lemma [4.2](#S4.Thmtheorem2 "Lemma 4.2. ‣ 4 Vote ‣ Information-theoretic security without an honest majority"). This information could be used to adapt the behaviour of the adversary in a future execution of vote.
In addition to the above theorems, it follows from the use of the simultaneous broadcast channel that an adversary cannot act in a way that a corrupt participant’s vote depends an honest participant’s vote. In particular, it cannot *duplicate* an honest participant’s vote. We claim that our protocol provides ballot casting assurance and universal verifiability. This is straightforward from the fact that participants do not entrust any computation to a third party: they provide their own inputs and can verify that the final outcome is computed correctly.
## 5 Anonymous Bit Transmission
The anonymous bit transmission protocol enables a sender to privately and anonymously transmit one bit to a receiver of his choice. Protocol [4](#protocol4 "Protocol 4 ‣ 5 Anonymous Bit Transmission ‣ Information-theoretic security without an honest majority") actually deals with the usually problematic scenario of multiple *anonymous senders* in an original way: it allows an arbitrary number participants to act as anonymous senders, each one targeting any number of participants and sending them each a chosen private bit. Each participant is also simultaneously a potential *receiver*: at the end of the protocol, each participant has a private account of how many anonymous senders sent the bit $`0`$ and how many sent the bit $`1`$. Note that in our formalism for multiparty computation, the *privacy* of the inputs implies the *anonymity* of the senders and receivers.
Protocol 4
Anonymous Bit Transmission
Input: $`x_{i}^{j} \in {\{ 0,1, \perp \}}`$, $`({j = {1,2,\ldots,n}})`$
Output: $`y_{i} = {({|{\{ x_{j}^{i}\mid{x_{j}^{i} = 0}\}}|},{|{\{ x_{j}^{i}\mid{x_{j}^{i} = 1}\}}|})}`$
Broadcast type : regular broadcast
1) (Correctness) If the protocol does not abort then the output of the protocol equals the output of the ideal functionality.
2) (Privacy) The privacy is the same as in the ideal functionality.
For each participant $`j`$,
1. 1\.
Execute the vote protocol with $`m = 3`$ as modified below. The three choices are: $`0`$, $`1`$, or $`\perp`$ (*abstain*). Each participant $`i`$ chooses his input to the vote according to $`x_{i}^{j}`$, his choice of message to be sent anonymously to participant $`j`$. The vote protocol is modified such that:
1. \(a\)
The output strings are sent to participant $`j`$ through the private channel.
2. \(b\)
Participant $`j`$ computes the tally as in the vote and if this computation succeeds, he finds out how many participants sent him a $`0`$, how many sent him a $`1`$ and how many abstained. If this occurs (and the results are consistent) he sets his success bit, $`s_{j}`$ to $`0`$. If the vote aborts, he sets $`s_{j}`$ to $`1`$.
Execute the veto protocol, using as inputs the success bits $`s_{j}`$. If the output of veto is $`0`$, then the anonymous bit transmission succeeds. Otherwise, the protocol fails.
The security of the anonymous bit transmission protocol follows directly from the security of the vote and of the veto. Of course, the anonymous bit transmission also inherits the drawbacks of these protocols. More precisely we have the following:
######
Lemma 5.1
.
(Correctness) The anonymous bit transmission protocol computes the correct output, except with exponentially small probability.
###### Proof.
If the protocol does not abort, by Lemmas [3.2](#S3.Thmtheorem2 "Lemma 3.2. ‣ 3 Veto ‣ Information-theoretic security without an honest majority") and [4.1](#S4.Thmtheorem1 "Lemma 4.1. ‣ 4 Vote ‣ Information-theoretic security without an honest majority"), except with exponentially small probability, all bits are correctly transmitted. ∎
######
Lemma 5.2
.
(Privacy) In the anonymous bit transmission protocol, the privacy is the same as in the ideal functionality.
###### Proof.
Each execution of the vote protocol provides perfect privacy, even if the protocol aborts. The final veto reveals some partial information about which honest participants have been targeted by corrupt participants, but this does not compromise the privacy of the protocol.∎
In Protocol [4](#protocol4 "Protocol 4 ‣ 5 Anonymous Bit Transmission ‣ Information-theoretic security without an honest majority"), the use of the private channel in step (a) can be removed and replaced by a broadcast channel. Since participant $`j`$ does not broadcast, the messages remain private. Another modification of the protocol makes it possible to send $`m`$ possible messages instead of just two but note that the complexity is polynomial in $`m`$ and not in the length of $`m`$. The transmission of arbitrarily long strings is discussed in Sections [8](#S8 "8 Fixed Role Anonymous Message Transmission ‣ Information-theoretic security without an honest majority") and [9](#S9 "9 Anonymous Message Transmission ‣ Information-theoretic security without an honest majority").
## 6 Collision Detection
The collision detection protocol (Protocol [5](#protocol5 "Protocol 5 ‣ 6 Collision Detection ‣ Information-theoretic security without an honest majority")) enables the participants to verify whether or not there is a single sender in the group. This will be used as a procedure for the implementation of *anonymous message transmission* in Section [9](#S9 "9 Anonymous Message Transmission ‣ Information-theoretic security without an honest majority"). Ideally, a protocol to detect a collision would have as input only $`x_{i} \in {\{ 0,1\}}`$ but unfortunately we do not know how to achieve such a functionality.
Protocol 5
Collision Detection
Input: $`x_{i} \in {\{ 0,1,2\}}`$
Output: let $`r = {\sum_{i = 1}^{n}x_{i}}`$ then $`y_{i} = {\min{\{ r,2\}}}`$
Broadcast type : regular broadcast
1) (Reliability) No participant can make the protocol abort.
2) (Correctness) The output of the protocol equals the output of the ideal functionality.
3) (Privacy) An adversary cannot learn more than it could have learned by assigning to all corrupt participants the input $`0`$ in the ideal functionality.
Veto A
All participants perform the veto protocol with inputs $`\min{\{ x_{i},1\}}`$.
Veto B
If the outcome of veto A is $`0`$, skip this step. Otherwise, each participant with input 1 in veto A will set $`b_{i} = 1`$ if he detected in veto A that another participant had input $`1`$, or if $`x_{i} = 2`$. All other participants set $`b_{i} = 0`$. Then all participants perform a second veto protocol with inputs $`b_{i}`$.
Output: $`y_{i} = \begin{cases}
0 & {\text{if the
outcome of~}\text{veto~A}\text{~is~0}} \\
1 & {\text{if the outcome of~}\text{veto~A}\text{~is~1 and the outcome of~}\text{veto~B}\text{~is~0~}} \\
2 & {\text{if the outcome of~}\text{veto~A}\text{~is~1 and the outcome of~}\text{veto~B}\text{~is~1}}
\end{cases}`$
######
Lemma 6.1
.
(Reliability) No participant can make the collision detection protocol abort.
###### Proof.
This follows from the reliability of veto. ∎
######
Lemma 6.2
.
(Correctness) In the collision detection protocol, the output equals the output of the ideal functionality (except with exponentially small probability).
###### Proof.
This follows from the correctness of the veto protocol. There are only two ways a corrupt participant can deviate from the protocol. First, participant $`i`$ can set $`b_{i} = 0`$ although $`x_{i} \in {\{ 0,1\}}`$ and although in the first veto his input was 1 and a collision was detected. The outcome of veto B will still be 1 since another participant with input 1 in veto A will input 1 in veto B. This is consistent with input $`x_{i} = 1`$. Second, participant $`i`$ can set $`b_{i} = 1`$ although $`x_{i} = 0`$. If veto B is executed, then we know that another participant has input 1 in veto A. This is consistent with input $`x_{i} = 1`$. ∎
Note that we have raised a subtle deviation from the ideal protocol in the above proof: we showed how it is possible for a corrupt participant to set his input to 0 if all other participants have input $`0`$ and to $`1`$ otherwise. Fortunately, the protocol is still sufficiently good for the requirements of the following sections.
######
Lemma 6.3
.
(Privacy) In the collision detection protocol, an adversary cannot learn more than it could have learned by assigning to all corrupt participants the input $`0`$ in the ideal functionality.
###### Proof.
In each veto, an adversary can only learn whether or not there exists an honest participant with input 1. In all cases, this can be deduced from the outcome of the ideal functionality by setting the input to be $`0`$ for all corrupt participants. ∎
## 7 Notification
In the notification protocol (Protocol [6](#protocol6 "Protocol 6 ‣ 7 Notification ‣ Information-theoretic security without an honest majority")), each participant chooses a list of other participants to notify. The output privately reveals to each participant whether or not he was notified, but no information on the number or origin of such notifications is revealed.
Protocol 6
Notification
Input: $`{{\forall j} \neq i},{x_{i}^{j} \in {\{ 0,1\}}}`$
Output: $`y_{i} = {\bigvee_{j \neq i}x_{j}^{i}}`$
Broadcast type : regular broadcast
1) (Correctness) If the protocol does not abort then the output of the protocol equals the output of the ideal functionality.
2) (Privacy) The privacy is the same as in the ideal functionality.
For each participant $`i`$:
—-
Participant $`i`$ sets $`y_{i}\leftarrow 0`$.
—-
Repeat $`s`$ times:
1. 1\.
Each participant $`j \neq i`$ sets the value of $`p_{j}`$ in the following way: if $`x_{j}^{i} = 0`$ then $`p_{j} = 0`$; otherwise, $`p_{j} = 1`$ with probability $`\frac{1}{2}`$ and $`p_{i} = 0`$ with complimentary probability. Let $`p_{i} = 0`$.
2. 2\.
The participants execute the parity protocol with inputs $`p_{1},p_{2},{\ldots\hspace{0pt}p_{n}}`$, with the exception that participant $`i`$ does not broadcast his value, and the simultaneous broadcast is replaced by a regular broadcast (if any participant refuses to broadcast, abort).
3. 3\.
Participant $`i`$ computes the outcome of parity, and if it is $`1`$, $`y_{i}\leftarrow 1`$ .
######
Lemma 7.1
.
The notification protocol achieves privacy and except with exponentially small probability, the correct output is computed.
###### Proof.
Privacy and correctness are trivially deduced from properties of the parity protocol. ∎
## 8 Fixed Role Anonymous Message Transmission
In Section [5](#S5 "5 Anonymous Bit Transmission ‣ Information-theoretic security without an honest majority"), we presented an anonymous bit transmission protocol. The protocol easily generalizes to $`m`$ messages, but the complexity of the protocol becomes polynomial in $`m`$. It is not clear how to modify the protocol to transmit a string of arbitrary length, while still allowing multiple senders and receivers. However, in the context where a single sender $`S`$ is allowed, it is possible to implement a secure protocol for $`S`$ to anonymously transmit a message to a single receiver $`R`$, which we call fixed role anonymous message transmission (Protocol [7](#protocol7 "Protocol 7 ‣ 8 Fixed Role Anonymous Message Transmission ‣ Information-theoretic security without an honest majority")). If the uniqueness condition on $`S`$ and $`R`$ is not satisfied, the protocol aborts. The protocol combines the use of the parity protocol with an *algebraic manipulation detection code* \[[CFP07](#bib.bibx4)\], which we present as Theorem [8.1](#S8.Thmtheorem1 "Theorem 8.1 ([CFP07]). ‣ 8 Fixed Role Anonymous Message Transmission ‣ Information-theoretic security without an honest majority"). Due to lack of space, the encoding and decoding algorithms, $`F`$ and $`G`$, respectfully, are not repeated. For a less efficient algorithm that achieves a similar result, see \[[PSJ99](#bib.bibx10)\].
######
Theorem 8.1
(\[[CFP07](#bib.bibx4)\]).
There exists an efficient probabilistic encoding algorithm $`F:{{\{ 0,1\}}^{m}\rightarrow{\{ 0,1\}}^{m + 2{(\log{(m)} + s}}}`$ and decoding algorithm $`G:{{\{ 0,1\}}^{m + {2\hspace{0pt}{({{\log{(m)}} + s})}}}\rightarrow{\{ \perp ,{\{ 0,1\}}^{m}\}}}`$ such that for all $`w`$, $`{G\hspace{0pt}{({F\hspace{0pt}{(w)}})}} = w`$, and any combination of bit flips applied to $`w' = {F\hspace{0pt}{(w)}}`$ produces a $`w^{\operatorname{\prime\prime}}`$ such that $`{G\hspace{0pt}{(w^{\operatorname{\prime\prime}})}} = \perp`$, except with probability $`2^{- s}`$.
Protocol 7
Fixed Role Anonymous Message Transmission
Oracle: The sender $`S`$ and receiver $`R`$ know their identity
Input: $`S`$ has input $`w \in {\{ 0,1\}}^{m}`$, all other players have no input
Output: $`R`$ has output $`w`$, all other players have no output
Broadcast type : regular broadcast
1) (Correctness) If the protocol does not abort, $`R`$ obtains the correct message.
2) (Privacy) The only information that can be learned through the protocol is for $`R`$ to learn $`w`$.
3) (Oracle) If the oracle conditions are not satisfied (in the sense that more than one honest participant believes to be the sender or the receiver), the protocol will abort.
1. 1\.
$`S`$ computes $`w' = {F\hspace{0pt}{(w)}}`$
2. 2\.
The participants execute $`m + {2\hspace{0pt}{({{\log{(m)}} + s})}}`$ rounds of the parity protocol, with participants using a broadcast instead of a simultaneous broadcast and using the following inputs:
1. \(a\)
$`S`$ uses as input the bits of $`w'`$.
2. \(b\)
$`R`$ uses as input the bits of a random $`m`$-bit string, $`r`$.
3. \(c\)
All other players use $`0`$ as input for each round.
3. 3\.
Let $`d`$ be the output of the rounds of parity. $`R`$ computes $`w^{\operatorname{\prime\prime}} = {d \oplus r}`$.
4. 4\.
$`R`$ computes $`y = {G\hspace{0pt}{(w^{\operatorname{\prime\prime}})}}`$.
5. 5\.
A veto is performed: all players input 0 except $`R`$ who inputs $`1`$ if $`y = \perp`$ and $`0`$ otherwise.
If the outcome of veto is 1, the protocol aborts. Otherwise, $`R`$ sets his output to $`y`$.
######
Lemma 8.2
.
(Correctness, Privacy, Oracle) In the fixed role anonymous message transmission protocol, the probability that $`R`$ obtains as output a corrupt message is exponentially small. The protocol is perfectly private, and if the oracle conditions are not satisfied, it will abort (except with exponentially small probability).
###### Proof.
Because of the properties of parity and the fact that the receiver broadcasts a random bit, we have perfect privacy. Correctness is a direct consequence of Theorem [8.1](#S8.Thmtheorem1 "Theorem 8.1 ([CFP07]). ‣ 8 Fixed Role Anonymous Message Transmission ‣ Information-theoretic security without an honest majority"). Finally, if more than one participant acts as a sender or receiver, then again because of Theorem [8.1](#S8.Thmtheorem1 "Theorem 8.1 ([CFP07]). ‣ 8 Fixed Role Anonymous Message Transmission ‣ Information-theoretic security without an honest majority"), the message will not be faithfully transmitted and the protocol will abort in step [5](#S8.I1.i5 "item 5 ‣ Protocol 7 ‣ 8 Fixed Role Anonymous Message Transmission ‣ Information-theoretic security without an honest majority"), except with exponentially small probability. ∎
######
Theorem 8.3
.
For a fixed security parameter, the fixed role anonymous message transmission protocol is asymptotically optimal.
###### Proof.
For any protocol to preserve the anonymity of the sender and the receiver, each player must sent at least one bit to every other player for each bit of the message. In the fixed role anonymous message transmission protocol, for a fixed $`s`$, each player actually sends $`O\hspace{0pt}{(1)}`$ bits to each other player and therefore the protocol is asymptotically optimal. ∎
## 9 Anonymous Message Transmission
Our final protocol allows a sender to anonymously transmit message to a receiver of his choosing. Contrary to the fixed role anonymous message transmission protocol of Section [8](#S8 "8 Fixed Role Anonymous Message Transmission ‣ Information-theoretic security without an honest majority"), anonymous message transmission (Protocol [8](#protocol8 "Protocol 8 ‣ 9 Anonymous Message Transmission ‣ Information-theoretic security without an honest majority")) does not suppose that there is a single sender, but instead, it deals with potential collisions (or lack of any sender at all) by producing the outputs Collision or No Transmission. The only deviation from the ideal functionality in the protocol is that a single participant can force the Collision output. Note again that in this protocol, the privacy of the input implies anonymity of the sender and receiver.
Protocol 8
Anonymous Message Transmission
Input: $`x_{i} = \perp`$ or $`x_{i} = {(r,w)}`$ where $`r \in {\{ 1,\ldots,n\}}`$ and $`w \in {\{ 0,1\}}^{m}`$
Output: If $`{|{\{ x_{i}\mid{x_{i} \neq \perp}\}}|} = 0`$ then $`y_{i} = \text{No~Transmission}`$ and if $`{|{\{ x_{i}\mid{x_{i} \neq \perp}\}}|} > 1`$ then $`y_{i} = \text{Collision}`$. Otherwise let $`S`$ be such that $`x_{S} = {(r,w)}`$ then all $`y_{i} = \perp`$ except $`y_{r} = w`$.
Broadcast type : regular broadcast
1) (Correctness) The output equals the output of the ideal functionality except that a single participant can make the protocol produce the output Collision.
2) (Privacy) The privacy is the same as in the ideal functionality.
1. 1\.
The participants execute the collision detection protocol; participants who have input $`x_{i} = \perp`$ use input $`0`$ while all others use input $`1`$. If the outcome of collision detection is $`1`$, continue, otherwise output No Transmission if the output is $`0`$ and Collision if the output is $`2`$.
2. 2\.
Let the sender $`S`$ be the unique participant with $`x_{S} \neq \perp`$. The participants execute the notification protocol, with $`S`$ using input $`x_{S}^{r} = 1`$ and $`x_{S}^{j} = 0`$ otherwise. All other participants use the input bits $`0`$. Let $`R`$ be the participant who computes as output $`y_{R} = 1`$. If the notification protocol fails, abort.
3. 3\.
The participants execute the fixed role anonymous message transmission protocol.
######
Lemma 9.1
.
(Correctness) In the anonymous message transmission protocol, the output equals the output of the ideal functionality except with exponentially small probability. The only exception is that a single participant can make the protocol produce the output Collision.
###### Proof.
This follows easily from the correctness of the collision detection, notification and fixed role anonymous message transmission protocols. ∎
######
Lemma 9.2
.
(Privacy) The anonymity of the sender and receiver are perfect. If the protocol succeeds, except with exponentially small probability, participant $`r`$ is the only participant who knows $`w`$.
###### Proof.
Perfect anonymity follows from the privacy of the collision detection, notification and anonymous message transmission protocols. If the sender successfully notifies the receiver in step [2](#S9.I1.i2 "item 2 ‣ Protocol 8 ‣ 9 Anonymous Message Transmission ‣ Information-theoretic security without an honest majority"), then the privacy of $`w`$ is perfect. But with exponentially small probability, the receiver will not be correctly notified, and an adversary acting as the receiver will successfully receive the message $`w`$. ∎
## 10 Conclusion
We have given six multiparty protocols that are information-theoretically secure without any assumption on the number of honest participants. It would be interesting to see if the techniques we used can be applied to other multiparty functions or in other contexts.
Our main goal was to prove the existence of several protocols in a model that does not make use of any strong hypotheses such as computational assumptions or an honest majority. This being said, all the presented protocols are reasonably efficient: they are all polynomial in terms of communication and computational complexity and in one case, asymptotically optimal.
## Acknowledgements
The authors wish to thank Hugue Blier, Gilles Brassard, Serge Fehr and Sébastien Gambs. A. B. is supported by scholarships from the CFUW, FQRNT and NSERC. A. T. is supported by CIFAR, FQRNT, MITACS and NSERC.
## References
- \[AN06\] B. Adida and C. A. Neff. Ballot casting assurance.
In EVT ’06, Proceedings of the First Usenix/ACCURATE Electronic Voting Technology Workshop, 2006.
- \[BGW88\] M. Ben-Or, S. Goldwasser, and A. Wigderson. Completeness theorems for non-cryptographic fault-tolerant distributed computation (extended abstract).
In Proceedings of the twentieth annual ACM Symposium on Theory of Computing (STOC), pages 1–10, 1988.
- \[CCD88\] D. Chaum, C. Crépeau, and I. Damgård. Multiparty unconditionally secure protocols (extended abstract).
In Proceedings of the twentieth annual ACM Symposium on Theory of Computing (STOC), pages 11–19, 1988.
- \[CFP07\] R. Cramer, S. Fehr, and C. Padró. Combinatorial codes for detection of algebraic manipulation and their applications. Manuscript, 2007.
- \[CFSY96\] R. Cramer, M. Franklin, B. Schoenmakers, and M. Yung. Multi-authority secret-ballot elections with linear work.
In Proceedings of Eurocypt 1996, pages 72–83, 1996.
- \[CGMA85\] B. Chor, S. Goldwasser, S. Micali, and B. Awerbuch. Verifiable secret sharing and achieving simultaneity in the presence of faults.
In Proceedings of the 26th Annual Symposium on Foundations of Computer Science (FOCS), pages 383–395, 1985.
- \[Cha81\] D. Chaum. Untraceable electronic mail, return addresses, and digital pseudonyms.
Communications of the ACM, 24:84–88, 1981.
- \[Cha88\] D. Chaum. The dining cryptographers problem: Unconditional sender and recipient untraceability.
Journal of Cryptology, 1:65–75, 1988.
- \[HM05\] A. Hevia and D. Micciancio. Simultaneous broadcast revisited.
In Proceedings of the twenty-fourth annual ACM symposium on Principles of distributed computing, pages 324–333, 2005.
- \[PSJ99\] C. Padró, G. Sáez, and L. Villar J.\\lx@bibnewblockDetection of cheaters in vector space secret sharing schemes.
Designs, codes and cryptography, 16:75–85, 1999.
- \[RB89\] T. Rabin and M. Ben-Or. Verifiable secret sharing and multiparty protocols with honest majority.
In Proceedings of the twenty-first annual ACM Symposium on Theory of Computing (STOC), pages 73–85, 1989.
- \[SK95\] K. Sako and J. Kilian. Receipt-free mix-type voting scheme — a practical solution to the implementation of a voting booth.
In Proceedings of Eurocrypt 1995, pages 393–403, 1995.
- \[WP89a\] M. Waidner and B. Pfitzmann. The dining cryptographers in the disco: Unconditional sender and recipient untraceability with computationally secure serviceability.
In Proceedings of Eurocrypt ’89, page 690, 1989.
- \[WP89b\] M. Waidner and B. Pfitzmann. Unconditional sender and recipient untraceability in spite of active attacks — some remarks. Technical report, Universität Karlsruhe, 1989.
## Appendix A Proof of correctness for Protocol [3](#protocol3 "Protocol 3 ‣ 4 Vote ‣ Information-theoretic security without an honest majority")
######
Lemma A.1
.
(Correctness) If all participants are honest in the vote (Protocol [3](#protocol3 "Protocol 3 ‣ 4 Vote ‣ Information-theoretic security without an honest majority")), then the output is correct, except with probability exponentially small in $`s`$.
###### Proof.
We fix a value $`k`$ and suppose that $`v`$ participants have input $`x_{i} = k`$. Thus we need to show that in the vote, $`{y\hspace{0pt}{\lbrack k\rbrack}_{i}} = v`$, except with probability exponentially small in $`s`$.
We now give the intuition behind phase C of the vote. Let $`p_{v}`$ be the probability that $`{p\hspace{0pt}{\lbrack k\rbrack}_{j}} = {\bigoplus_{i = 1}^{n}{z\hspace{0pt}{\lbrack k\rbrack}_{i}^{j}}} = 1`$. For $`v \leq n`$, we have $`p_{0} = 0`$, $`p_{1} = \frac{1}{n}`$ and $`p_{v + 1} = {{p_{v}\hspace{0pt}\left( {1 - \frac{1}{n}} \right)} + {{({1 - p_{v}})}\hspace{0pt}\frac{1}{n}}}`$. Solving this recurrence, we get
<table>
<tbody>
<tr>
<td></td>
<td>pv=12(n−2n)v((nn−2)v−1).subscript𝑝𝑣12superscript𝑛2𝑛𝑣superscript𝑛𝑛2𝑣1p_{v}=\frac{1}{2}\left(\frac{n-2}{n}\right)^{v}\left(\left(\frac{n}{n-2}\right)^{v}-1\right)\,.</td>
<td></td>
<td>(1)</td>
</tr>
</tbody>
</table>
Thus, the idea of phase C of the vote is for the participants to approximate $`p_{v}`$ by computing $`{\sigma\hspace{0pt}{\lbrack k\rbrack}_{i}} = {\sum_{i = 1}^{s}{{p\hspace{0pt}{\lbrack k\rbrack}_{j}}/s}}`$. If the approximation is within $`\frac{1}{2\hspace{0pt}e^{2}\hspace{0pt}n}`$ of $`p_{v}`$, then the outcome is $`{y\hspace{0pt}{\lbrack k\rbrack}_{i}} = v`$. We first show that if such a $`v`$ exists, it is unique.
Clearly, for $`v < n`$, we have that $`p_{v + 1} > p_{v}`$. We also have $`{\lim_{n\rightarrow\infty}p_{n}} = {\frac{1}{2} - \frac{1}{2\hspace{0pt}e^{2}}}`$. Thus the difference between $`p_{v + 1}`$ and $`p_{v}`$ is:
<table>
<tbody>
<tr>
<td></td>
<td>pv+1−pvsubscript𝑝𝑣1subscript𝑝𝑣\displaystyle p_{v+1}-p_{v}</td>
<td>=pv(1−1n)+(1−p)1n−pvabsentsubscript𝑝𝑣11𝑛1𝑝1𝑛subscript𝑝𝑣\displaystyle=p_{v}\left(1-\frac{1}{n}\right)+(1-p)\frac{1}{n}-p_{v}</td>
<td></td>
<td>(2)</td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td></td>
<td>=1−2pvnabsent12subscript𝑝𝑣𝑛\displaystyle=\frac{1-2p_{v}}{n}</td>
<td></td>
<td>(3)</td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td></td>
<td>>1−2pnnabsent12subscript𝑝𝑛𝑛\displaystyle>\frac{1-2p_{n}}{n}</td>
<td></td>
<td>(4)</td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td></td>
<td>>1e2nabsent1superscript𝑒2𝑛\displaystyle>\frac{1}{e^{2}n}</td>
<td></td>
<td>(5)</td>
</tr>
</tbody>
</table>
Hence if such a $`v`$ exists, it is unique. We now show that except with probability exponentially small in $`s`$, the correct $`v`$ will be chosen. Let $`X = {\sum_{j = 1}^{s}{p\hspace{0pt}{\lbrack k\rbrack}_{j}}}`$ be the sum of the $`s`$ executions of parity, with $`\mu = {s\hspace{0pt}p_{v}}`$ the expected value of $`X`$. The participants have computed $`{\sigma\hspace{0pt}{\lbrack k\rbrack}_{i}} = {X/s}`$ .
By the Chernoff bound, for any $`0 < \delta \leq 1`$,
<table>
<tbody>
<tr>
<td></td>
<td>Pr[X≤(1−δ)μ]<exp(−μδ2/2)Pr𝑋1𝛿𝜇𝜇superscript𝛿22\Pr[X\leq(1-\delta)\mu]<\exp(-\mu\delta^{2}/2)</td>
<td></td>
<td>(6)</td>
</tr>
</tbody>
</table>
Let $`\delta = \frac{1}{2\hspace{0pt}e^{2}\hspace{0pt}n\hspace{0pt}p_{v}}`$. We have
<table>
<tbody>
<tr>
<td></td>
<td>Pr[X≤μ−s2e2n]<exp(−s8e4n2pv)Pr𝑋𝜇𝑠2superscript𝑒2𝑛𝑠8superscript𝑒4superscript𝑛2subscript𝑝𝑣\Pr[X\leq\mu-\frac{s}{2e^{2}n}]<\exp(-\frac{s}{8e^{4}n^{2}p_{v}})</td>
<td></td>
<td>(7)</td>
</tr>
</tbody>
</table>
and so
<table>
<tbody>
<tr>
<td></td>
<td>Pr[σ[k]i−pv≤−12e2n]<exp(−s8e4n2pv)Pr𝜎subscriptdelimited-[]𝑘𝑖subscript𝑝𝑣12superscript𝑒2𝑛𝑠8superscript𝑒4superscript𝑛2subscript𝑝𝑣\Pr[\sigma[k]_{i}-p_{v}\leq\frac{-1}{2e^{2}n}]<\exp(-\frac{s}{8e^{4}n^{2}p_{v}})</td>
<td></td>
<td>(8)</td>
</tr>
</tbody>
</table>
Similarly, still by the Chernoff bound, for any $`\delta < {{2\hspace{0pt}e} - 1}`$,
<table>
<tbody>
<tr>
<td></td>
<td>Pr[X>(1+δ)μ]<exp(−μδ2/4)Pr𝑋1𝛿𝜇𝜇superscript𝛿24\Pr[X>(1+\delta)\mu]<\exp(-\mu\delta^{2}/4)</td>
<td></td>
<td>(9)</td>
</tr>
</tbody>
</table>
Let $`\delta = \frac{1}{2\hspace{0pt}e^{2}\hspace{0pt}n\hspace{0pt}p_{v}}`$ and we get
<table>
<tbody>
<tr>
<td></td>
<td>Pr[X>μ+s2e2n]<exp(−s16e4n2pv)Pr𝑋𝜇𝑠2superscript𝑒2𝑛𝑠16superscript𝑒4superscript𝑛2subscript𝑝𝑣\Pr[X>\mu+\frac{s}{2e^{2}n}]<\exp(\frac{-s}{16e^{4}n^{2}p_{v}})</td>
<td></td>
<td>(10)</td>
</tr>
</tbody>
</table>
and so
<table>
<tbody>
<tr>
<td></td>
<td>Pr[σ[k]i−pv>12e2n]<exp(−s16e4n2pv)Pr𝜎subscriptdelimited-[]𝑘𝑖subscript𝑝𝑣12superscript𝑒2𝑛𝑠16superscript𝑒4superscript𝑛2subscript𝑝𝑣\Pr[\sigma[k]_{i}-p_{v}>\frac{1}{2e^{2}n}]<\exp(\frac{-s}{16e^{4}n^{2}p_{v}})</td>
<td></td>
<td>(11)</td>
</tr>
</tbody>
</table>
Hence the protocol produces the correct value for $`y\hspace{0pt}{\lbrack k\rbrack}_{i}`$, except with probability exponentially small in $`s`$. ∎ |
0706.1019 | cs.CR | cyber | <sup>1</sup>
<sup>1</sup>institutetext: Institute for Computing and Information Sciences,
Radboud University Nijmegen, The Netherlands.
http://www.cs.ru.nl/~{flaviog, petervr, anas}
# Probabilistic Anonymity and Admissible Schedulers
Flavio D. Garcia
Peter van Rossum
Ana Sokolova
###### Abstract
When studying safety properties of (formal) protocol models, it is customary to view the scheduler as an adversary: an entity trying to falsify the safety property. We show that in the context of security protocols, and in particular of anonymizing protocols, this gives the adversary too much power; for instance, the contents of encrypted messages and internal computations by the parties should be considered invisible to the adversary.
We restrict the class of schedulers to a class of admissible schedulers which better model adversarial behaviour. These admissible schedulers base their decision solely on the past behaviour of the system that is visible to the adversary.
Using this, we propose a definition of anonymity: for all admissible schedulers the identity of the users and the observations of the adversary are independent stochastic variables. We also develop a proof technique for typical cases that can be used to proof anonymity: a system is anonymous if it is possible to ‘exchange’ the behaviour of two users without the adversary ‘noticing’.
## 1 Introduction
Systems that include probabilities and nondeterminism are very convenient for modelling probabilistic (security) protocols. Nondeterminism is highly desirable feature for modelling implementation freedom, action of the environment, or incomplete knowledge about the system.
It is often of use to analyze probabilistic properties of such systems as for example “in 30% of the cases sending a message is followed by receiving a message” or “the system terminates successfully with probability at least 0.9”. Probabilistic anonymity \[[BP05](#bib.bibx2)\] is also such a property. In order to be able to consider such probabilistic properties we must first eliminate the nondeterminism present in the models. This is usually done by entities called schedulers or adversaries. It is common in the analysis of probabilistic systems to say that a model with nondeterminism and probability satisfies a probabilistic property if and only if it satisfies it no matter in which way the nondeterminism was resolved, i.e., for *all possible schedulers*.
On the other hand, in security protocols, adversaries or schedulers are malicious entities that try to break the security of the protocol. Therefore, allowing just any scheduler is inadmissible. We show that the well-known Chaum’s Dining Cryptographers (DC) protocol \[[Cha88](#bib.bibx3)\] is not anonymous if we allow for all possible schedulers. Since the protocol is well-known to be anonymous, this shows that for the treatment of probabilistic security properties, in particular probabilistic anonymity, the general approach to analyzing probabilistic systems does not directly fit.
We propose a solution based on restricting the class of all schedulers to a smaller class of *admissible schedulers*. Then we say that a probabilistic security property holds for a given model, if the property holds after resolving the nondeterminism under *all admissible schedulers*.
## 2 Probabilistic Automata
In this section we gather preliminary notions and results related to probabilistic automata \[[SL94](#bib.bibx10), [Seg95](#bib.bibx9)\]. Some of the formulations we borrow from \[[Sok05](#bib.bibx11)\] and \[[Che06](#bib.bibx4)\]. We shall model protocols with probabilistic automata. We start with a definition of probability distribution.
###### Definition 2.1 (Probability distribution)
A function $`\mu:{S\rightarrow{\lbrack 0,1\rbrack}}`$ is a discrete probability distribution, or distribution for short, on a set $`S`$ if $`{\sum_{x \in S}{\mu\hspace{0pt}{(x)}}} = 1`$. The set $`\left. \{{x \in S} \middle| {{\mu\hspace{0pt}{(x)}} > 0}\} \right.`$ is the support of $`\mu`$ and is denoted by $`{supp}{(\mu)}`$. By $`\mathcal{D}\hspace{0pt}{(S)}`$ we denote the set of all discrete probability distributions on the set $`S`$.
We use the simple probabilistic automata \[[SL94](#bib.bibx10), [Seg95](#bib.bibx9)\], or MDP’s \[[Bel57](#bib.bibx1)\] as models of our probabilistic processes. These models are very similar to the labelled transition systems, with the only difference that the target of each transition is a distribution over states instead of just a single next state.
###### Definition 2.2 (Probabilistic automaton)
A probabilistic automaton is a triple $`\mathcal{A} = {\langle S,A,\alpha\rangle}`$ where:
- •
$`S`$ is a set of states.
- •
$`A`$ is a set of actions or action labels.
- •
$`\alpha`$ is a transition function $`\alpha:{S\rightarrow{\mathcal{P}\hspace{0pt}{({A \times {\mathcal{D}S}})}}}`$.
A terminating state of $`\mathcal{A}`$ is a state with no outgoing transition, i.e. with $`{\alpha\hspace{0pt}{(s)}} = \varnothing`$. We might sometimes also specify an initial state $`s_{0} \in S`$ of a probabilistic automaton $`\mathcal{A}`$. We write $`s\overset{a}{\rightarrow}\mu`$ for $`{{(a,\mu)} \in {\alpha\hspace{0pt}{(s)}}},{s \in S}`$. Moreover, we write $`s\overset{a,\mu}{↝}t`$ for $`{s,t} \in S`$ whenever $`s\overset{a}{\rightarrow}\mu`$ and $`{\mu\hspace{0pt}{(t)}} > 0`$.
We will also need the notion of a fully probabilistic system.
###### Definition 2.3 (Fully Probabilistic Automaton)
A fully probabilistic automaton is a triple $`\mathcal{A} = {\langle S,A,\alpha\rangle}`$ where:
- •
$`S`$ is a set of states.
- •
$`A`$ is a set of actions or action labels.
- •
$`\alpha`$ is a transition function $`\alpha:{S\rightarrow{{\mathcal{D}{({A \times S})}} + 1}}`$.
Here $`1 = {\{ \ast \}}`$ denotes termination, i.e., if $`{\alpha\hspace{0pt}{(s)}} = \ast`$ then $`s`$ is a terminating state. It can also be understood as a zero-distribution i.e. $`{\alpha\hspace{0pt}{(s)}\hspace{0pt}{(a,t)}} = 0`$ for all $`a \in A`$ and $`t \in S`$. By $`s_{0} \in S`$ we sometimes denote an initial state of $`\mathcal{A}`$. We write $`s\rightarrow\mu`$ for $`{\mu = {\alpha\hspace{0pt}{(s)}}},{s \in S}`$. Moreover, we write $`s\overset{a}{↝}t`$ for $`{s,t} \in S`$ whenever $`s\rightarrow\mu`$ and $`{\mu\hspace{0pt}{(a,t)}} > 0`$.
A major difference between the (simple) probabilistic automata and the fully probabilistic ones is that the former can express nondeterminism. In order to reason about probabilistic properties of a model with nondeterminism we first resolve the nondeterminism with help of schedulers or adversaries – this leaves us with a fully probabilistic model whose probabilistic behaviour we can analyze. We explain this in the sequel.
###### Definition 2.4 (Paths)
A path of a *probabilistic automaton* $`\mathcal{A}`$ is a sequence
<table>
<tbody>
<tr>
<td></td>
<td>π=s0→a1,μ1s1→a2,μ2s2…𝜋subscript𝑠0superscript→subscript𝑎1subscript𝜇1subscript𝑠1superscript→subscript𝑎2subscript𝜇2subscript𝑠2…\pi=s_{0}\stackrel{{\scriptstyle a_{1},\mu_{1}}}{{\to}}s_{1}\stackrel{{\scriptstyle a_{2},\mu_{2}}}{{\to}}s_{2}\dots</td>
<td></td>
</tr>
</tbody>
</table>
where $`s_{i} \in S`$, $`a_{i} \in A`$ and $`s_{i}\overset{a_{i + 1},\mu_{i + 1}}{↝}s_{i + 1}`$.
A path of a *fully probabilistic automaton* $`\mathcal{A}`$ is a sequence
<table>
<tbody>
<tr>
<td></td>
<td>π=s0→a1s1→a2s2…𝜋subscript𝑠0superscript→subscript𝑎1subscript𝑠1superscript→subscript𝑎2subscript𝑠2…\pi=s_{0}\stackrel{{\scriptstyle a_{1}}}{{\to}}s_{1}\stackrel{{\scriptstyle a_{2}}}{{\to}}s_{2}\dots</td>
<td></td>
</tr>
</tbody>
</table>
where again $`s_{i} \in S`$, $`a_{i} \in A`$ and $`s_{i}\overset{a_{i + 1}}{↝}s_{i + 1}`$.
A path can be finite in which case it ends with a state. A path is complete if it is either infinite or finite ending in a terminating state. We let $`{last}{(\pi)}`$ denote the last state of a finite path $`\pi`$, and for arbitrary path $`{first}{(\pi)}`$ denotes its first state. A trace of a path is the sequence of actions in $`A^{\ast} \cup A^{\infty}`$ obtained by removing the states (and the distributions), hence above $`{{trace}{(\pi)}} = {a_{1}\hspace{0pt}a_{2}\hspace{0pt}\ldots}`$. The length of a finite path $`\pi`$, denoted by $`|\pi|`$ is the number of actions in its trace. Let $`{Paths}{(\mathcal{A})}`$ denote the set of all paths, $`{Paths}^{\leq \omega}{(\mathcal{A})}`$ the set of all finite paths, and $`{CPaths}{(\mathcal{A})}`$ the set of all complete paths of an automaton $`\mathcal{A}`$.
Paths are ordered by the prefix relation, which we denote by $`\leq`$.
Let $`\mathcal{A}`$ be a (fully) probabilistic automaton and let $`\pi_{i}`$ for $`i \geq 0`$ be finite paths of $`\mathcal{A}`$ all starting in the same initial state $`s_{0}`$ and such that $`\pi_{i} \leq \pi_{j}`$ for $`i \leq j`$ and $`{|\pi_{i}|} = i`$, for all $`i \geq 0`$. Then by $`\pi = {\lim_{i\rightarrow\infty}\pi_{i}}`$ we denote the infinite complete path with the property that $`\pi_{i} \leq \pi`$ for all $`i \geq 0`$.
###### Definition 2.5 (Cone)
Let $`\mathcal{A}`$ be a (fully) probabilistic automaton and let $`\pi \in {{Paths}^{\leq \omega}{(\mathcal{A})}}`$ be given. The cone generated by $`\pi`$ is the set of paths
<table>
<tbody>
<tr>
<td></td>
<td>Cπ={π′∈CPaths(𝒜)∣π≤π′}.subscript𝐶𝜋conditional-setsuperscript𝜋′CPaths𝒜𝜋superscript𝜋′C_{\pi}=\{\pi^{\prime}\in{\operatorname{{CPaths}}}(\mathcal{A})\mid\pi\leq\pi^{\prime}\}.</td>
<td></td>
</tr>
</tbody>
</table>
From now on we fix an initial state. Given a fully probabilistic automaton $`\mathcal{A}`$ with an initial state $`s_{0}`$, we can calculate the probability-value denoted by $`\mathbf{P}{(\pi)}`$ of any finite path $`\pi`$ starting in $`s_{0}`$ as follows.
<table>
<tbody>
<tr>
<td></td>
<td>𝐏(s0)𝐏subscript𝑠0\displaystyle{\operatorname{\mathbf{P}}}(s_{0})</td>
<td>=\displaystyle=</td>
<td>11\displaystyle 1</td>
<td></td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td>𝐏(π→as)𝐏superscript→𝑎𝜋𝑠\displaystyle{\operatorname{\mathbf{P}}}(\pi\stackrel{{\scriptstyle a}}{{\to}}s)</td>
<td>=\displaystyle=</td>
<td>𝐏(π)⋅μ(a,s) where last(π)→μ→⋅𝐏𝜋𝜇𝑎𝑠 where last𝜋𝜇\displaystyle{\operatorname{\mathbf{P}}}(\pi)\cdot\mu(a,s)\quad\text{~{}where~{}}\operatorname{last}(\pi)\to\mu</td>
<td></td>
</tr>
</tbody>
</table>
Let $`\Omega_{\mathcal{A}} = {{CPaths}{(\mathcal{A})}}`$ be the sample space, and let $`\mathcal{F}_{\mathcal{A}}`$ be the smallest $`\sigma`$-algebra generated by the cones. The following proposition (see \[[Seg95](#bib.bibx9), [Sok05](#bib.bibx11)\]) states that $`\mathbf{P}`$ induces a unique probability measure on $`\mathcal{F}_{\mathcal{A}}`$.
###### Proposition 1
Let $`\mathcal{A}`$ be a fully probabilistic automaton and let $`\mathbf{P}`$ denote the probability-value on paths. There exists a unique probability measure on $`\mathcal{F}_{\mathcal{A}}`$ also denoted by $`\mathbf{P}`$ such that $`{\mathbf{P}{(C_{\pi})}} = {\mathbf{P}{(\pi)}}`$ for every finite path $`\pi`$.∎
This way we are able to measure the probability of certain events described by sets of paths in an automaton with no nondeterminism. Since our models include nondeterminism, we will first resolve it by means of schedulers or adversaries. Before we define adversaries note that we can describe the set of all sub-probability distributions on a set $`S`$ by $`\mathcal{D}{({S + 1})}`$. These are functions whose sum of values on $`S`$ is not necessarily equal to 1, but it is bounded by 1.
###### Definition 2.6 (Scheduler)
A scheduler for a probabilistic automaton $`\mathcal{A}`$ is a function
<table>
<tbody>
<tr>
<td></td>
<td>ξ:Paths≤ω(𝒜)→𝒟(A×𝒟(S)+1):𝜉→superscriptPathsabsent𝜔𝒜𝒟𝐴𝒟𝑆1\xi\colon{\operatorname{{Paths}^{\leq\omega}}}(\mathcal{A})\to{\operatorname{{\mathcal{D}}}}(A\times{\operatorname{{\mathcal{D}}}}(S)+1)</td>
<td></td>
</tr>
</tbody>
</table>
satisfying $`{\xi\hspace{0pt}{(\pi)}\hspace{0pt}{(a,\mu)}} > 0`$ implies $`{{last}{(\pi)}}\overset{a}{\rightarrow}\mu`$, for each finite path $`\pi`$. By $`{Sched}{(\mathcal{A})}`$ we denote the set of all schedulers for $`\mathcal{A}`$.
Hence, a scheduler according to the previous definition imposes a probability distribution on the possible non-deterministic transitions in each state. Therefore it is randomized. It is history dependent since it takes into account the path (history) and not only the current state. It is partial since it gives a sub-probability distribution, i.e., it may halt the execution at any time.
###### Definition 2.7 (Automaton under scheduler)
A probabilistic automaton $`\mathcal{A} = {\langle S,A,\alpha\rangle}`$ together with a scheduler $`\xi`$ determine a fully probabilistic automaton
<table>
<tbody>
<tr>
<td></td>
<td>𝒜ξ=⟨Paths≤ω(𝒜),A,αξ⟩.subscript𝒜𝜉superscriptPathsabsent𝜔𝒜𝐴subscript𝛼𝜉\mathcal{A}_{\xi}=\langle{\operatorname{{Paths}^{\leq\omega}}}(\mathcal{A}),A,\alpha_{\xi}\rangle.</td>
<td></td>
</tr>
</tbody>
</table>
Its set of states are the finite paths of $`\mathcal{A}`$, its initial state is the initial state of $`\mathcal{A}`$ (seen as a path with length 1), its actions are the same as those of $`\mathcal{A}`$, and its transition function $`\alpha_{\xi}`$ is defined as follows. For any $`\pi \in {{Paths}^{\leq \omega}{(\mathcal{A})}}`$, we have $`{\alpha_{\xi}\hspace{0pt}{(\pi)}} \in {{\mathcal{D}{({A \times {{Paths}^{\leq \omega}{(\mathcal{A})}}})}} + 1}`$ as
<table>
<tbody>
<tr>
<td></td>
<td>αξ(π)(a,π′)={ξ(π)(a,μ)⋅μ(s)π′=π→a,μs0otherwisesubscript𝛼𝜉𝜋𝑎superscript𝜋′cases⋅𝜉𝜋𝑎𝜇𝜇𝑠superscript𝜋′𝜋superscript→𝑎𝜇𝑠0otherwise\alpha_{\xi}(\pi)(a,\pi^{\prime})=\left\{\begin{array}[]{ll}\xi(\pi)(a,\mu)\cdot\mu(s)&\,\,\,\pi^{\prime}=\pi\stackrel{{\scriptstyle a,\mu}}{{\to}}s\\ 0&\,\,\,\text{otherwise}\end{array}\right.</td>
<td></td>
</tr>
</tbody>
</table>
Given a probabilistic automaton $`\mathcal{A}`$ and a scheduler $`\xi`$, we denote by $`\mathbf{P}_{\xi}`$ the probability measure on sets of complete paths of the fully probabilistic automaton $`\mathcal{A}_{\xi}`$, as in Proposition [1](#Thmproposition1 "Proposition 1 ‣ 2 Probabilistic Automata ‣ Probabilistic Anonymity and Admissible Schedulers"). The corresponding $`\sigma`$-algebra generated by cones of finite paths of $`\mathcal{A}_{\xi}`$ we denote by $`\Omega_{\xi}`$. The elements of $`\Omega_{\xi}`$ are measurable sets.
By $`\Omega`$ we denote the $`\sigma`$-algebra generated by cones of finite paths of $`\mathcal{A}`$ (without fixing the scheduler!) and also call its elements measurable sets, without having a measure in mind. Actually, we will now show that any scheduler $`\xi \in {{Sched}{(\mathcal{A})}}`$ induces a measure $`\mathbf{P}_{(\xi)}`$ on a certain $`\sigma`$-algebra $`\Omega_{(\xi)}`$ of paths in $`\mathcal{A}`$ such that $`\Omega \subseteq \Omega_{(\xi)}`$. Hence, any element of $`\Omega`$ can be measured by any of these measures $`\mathbf{P}_{(\xi)}`$. We proceed with the details.
Define a function $`f:{{{Paths}^{\leq \omega}{(\mathcal{A}_{\xi})}}\rightarrow{{Paths}^{\leq \omega}{(\mathcal{A})}}}`$ by
<table>
<tbody>
<tr>
<td></td>
<td>f(π^)=last(π^)𝑓^𝜋last^𝜋f(\hat{\pi})=\operatorname{last}(\hat{\pi})</td>
<td></td>
<td>(1)</td>
</tr>
</tbody>
</table>
for any $`\hat{\pi} \in \mathcal{A}_{\xi}`$. The function $`f`$ is well-defined since states in $`\mathcal{A}_{\xi}`$ are the finite paths of $`\mathcal{A}`$. Moreover, we have the following property.
###### Lemma 1
For any $`{{\hat{\pi}}_{1},{\hat{\pi}}_{2}} \in {{Paths}^{\leq \omega}{(\mathcal{A}_{\xi})}}`$ we have
<table>
<tbody>
<tr>
<td></td>
<td>π^1≤π^2⇔f(π^1)≤f(π^2)formulae-sequencesubscript^𝜋1subscript^𝜋2iff𝑓subscript^𝜋1𝑓subscript^𝜋2\hat{\pi}_{1}\leq\hat{\pi}_{2}\quad\iff\quad f(\hat{\pi}_{1})\leq f(\hat{\pi}_{2})</td>
<td></td>
</tr>
</tbody>
</table>
where the order on the left is the prefix order in $`{Paths}^{\leq \omega}{(\mathcal{A}_{\xi})}`$ and on the right the prefix order in $`{Paths}^{\leq \omega}{(\mathcal{A})}`$.
###### Proof.
By the definition of $`\mathcal{A}_{\xi}`$ we have that for $`{\pi,\pi'} \in {{Paths}^{\leq \omega}{(\mathcal{A})}}`$ i.e. states of $`\mathcal{A}_{\xi}`$: $`\pi\overset{a}{↝}\pi'`$ if and only if $`{\xi\hspace{0pt}{(\pi)}} > 0`$ and $`{{last}{(\pi)}}\overset{a,\mu}{s}`$ in $`\mathcal{A}`$ for some $`\mu`$ and $`s`$. In other words if $`\pi\overset{a}{\rightarrow}\pi'`$, then $`\pi \leq \pi'`$ and $`{|\pi'|} = {{|\pi|} + 1}`$ i.e. $`\pi'`$ extends $`\pi`$ in one step. Therefore, if we have a path $`\pi_{0}\overset{a_{0}}{\rightarrow}\pi_{1}\overset{a_{1}}{\rightarrow}\pi_{2}\overset{a_{2}}{\rightarrow}\cdots`$ in $`\mathcal{A}_{\xi}`$ , then for all its states: if $`i \leq j`$, then $`\pi_{i} \leq \pi_{j}`$ and $`{|\pi_{j}|} = {{|\pi_{i}|} + {({j - i})}}`$. So if $`{{\hat{\pi}}_{1},{\hat{\pi}}_{2}} \in {{Paths}^{\leq \omega}{(\mathcal{A}_{\xi})}}`$ are such that $`{\hat{\pi}}_{1} \leq {\hat{\pi}}_{2}`$, then $`{last}{({\hat{\pi}}_{1})}`$ is a state in $`{\hat{\pi}}_{2}`$ and therefore we at once get $`{{last}{({\hat{\pi}}_{1})}} \leq {{last}{({\hat{\pi}}_{2})}}`$. For the opposite implication, again from the definition we notice that if a path $`\hat{\pi} \in {{Paths}^{\leq \omega}{(\mathcal{A})}}`$ contains a state $`\pi \in {{Paths}^{\leq \omega}{(\mathcal{A})}}`$, then it also contains all prefixes of $`\pi`$ as states. Hence, if $`{{last}{({\hat{\pi}}_{1})}} \leq {{last}{({\hat{\pi}}_{2})}}`$ for $`{\hat{\pi_{1}},\hat{\pi_{2}}} \in {{Paths}^{\leq \omega}{(\mathcal{A}_{\xi})}}`$, then $`{last}{({\hat{\pi}}_{1})}`$ is a state in $`{\hat{\pi}}_{2}`$ and also all its prefixes are. Since all paths start in the initial state (path), this implies that $`{\hat{\pi}}_{1} \leq {\hat{\pi}}_{2}`$. ∎∎
###### Corollary 1
The function $`f`$ defined by ([1](#S2.E1 "In 2 Probabilistic Automata ‣ Probabilistic Anonymity and Admissible Schedulers")) is injective. ∎
By Lemma [1](#Thmlemma1 "Lemma 1 ‣ 2 Probabilistic Automata ‣ Probabilistic Anonymity and Admissible Schedulers") we can extend the function $`f`$ to $`\hat{f}:{{{CPaths}{(\mathcal{A}_{\xi})}}\rightarrow{{CPaths}{(\mathcal{A})}}}`$ by
<table>
<tbody>
<tr>
<td></td>
<td>f^(π^)={f(π^)π is finitelimi→inftyf(π^i)π^i≤π,|π^i|=i^𝑓^𝜋cases𝑓^𝜋𝜋 is finitesubscript→𝑖𝑖𝑛𝑓𝑡𝑦𝑓subscript^𝜋𝑖formulae-sequencesubscript^𝜋𝑖𝜋subscript^𝜋𝑖𝑖\hat{f}(\hat{\pi})=\begin{cases}f(\hat{\pi})&\pi\text{~{}is~{}finite}\\ \lim_{i\to infty}f(\hat{\pi}_{i})&\hat{\pi}_{i}\leq\pi,\,|\hat{\pi}_{i}|=i\end{cases}</td>
<td></td>
</tr>
</tbody>
</table>
The properties from Lemma [1](#Thmlemma1 "Lemma 1 ‣ 2 Probabilistic Automata ‣ Probabilistic Anonymity and Admissible Schedulers") and Corollary [1](#Thmcorollary1 "Corollary 1 ‣ 2 Probabilistic Automata ‣ Probabilistic Anonymity and Admissible Schedulers") continue to hold for the extended function $`\hat{f}`$ as well. We will write $`f`$ for $`\hat{f}`$ as well.
Recall that $`\Omega_{\xi}`$ denotes the $`\sigma`$-algebra on which the measure $`\mathbf{P}_{\xi}`$ is defined. We now define a family of subsets $`\Omega^{\xi}`$ of $`{CPaths}{(\mathcal{A})}`$ by
<table>
<tbody>
<tr>
<td></td>
<td>Ωξ={Π∈CPaths(𝒜)∣f−1(Π)∈Ωξ.\Omega^{\xi}=\{\Pi\in{\operatorname{{CPaths}}}(\mbox{$\mathcal{A}$})\mid f^{-1}(\Pi)\in\Omega_{\xi}.</td>
<td></td>
<td>(2)</td>
</tr>
</tbody>
</table>
The following properties are instances of standard measure-theoretic results.
###### Lemma 2
The family $`\Omega^{\xi}`$ is a $`\sigma`$-algebra on $`{CPaths}{(\mathcal{A})}`$ and by
<table>
<tbody>
<tr>
<td></td>
<td>𝐏ξ(Π)=𝐏ξ(f−1(Π))superscript𝐏𝜉Πsubscript𝐏𝜉superscript𝑓1Π{\operatorname{\mathbf{P}}}^{\xi}(\Pi)={\operatorname{\mathbf{P}}}_{\xi}(f^{-1}(\Pi))</td>
<td></td>
</tr>
</tbody>
</table>
for $`\Pi \in \Omega^{\xi}`$ a measure on $`\Omega^{\xi}`$ is defined.
Recall that $`\Omega`$ denotes the $`\sigma`$-algebra on complete paths of $`\mathcal{A}`$ generated by the cones. We show that for any scheduler $`\xi`$, $`\Omega \subseteq \Omega^{\xi}`$. Hence, the measurable sets (elements of $`\Omega`$) are indeed measurable by the measure induced by any scheduler.
###### Lemma 3
For any scheduler $`\xi`$, $`\Omega \subseteq \Omega^{\xi}`$.
###### Proof.
Fix a scheduler $`\xi`$. Since $`\Omega`$ is generated by the cones it is enough to show that each cone is in $`\Omega^{\xi}`$. Let $`C_{\pi_{0},\mathcal{A}}`$ be a cone in $`{CPaths}{(\mathcal{A})}`$ generated by the finite path $`\pi_{0}`$, i.e.
<table>
<tbody>
<tr>
<td></td>
<td>Cπ0,𝒜={π∈CPaths(𝒜)∣π0≤π}.subscript𝐶subscript𝜋0𝒜conditional-set𝜋CPaths𝒜subscript𝜋0𝜋C_{\pi_{0},\mathcal{A}}=\{\pi\in{\operatorname{{CPaths}}}(\mbox{$\mathcal{A}$})\mid\pi_{0}\leq\pi\}.</td>
<td></td>
</tr>
</tbody>
</table>
We have
<table>
<tbody>
<tr>
<td></td>
<td>f^−1(Cπ0,𝒜)={∅π0∉f^(CPaths(𝒜))Cπ^0,𝒜ξf^(π^0)=π0.superscript^𝑓1subscript𝐶subscript𝜋0𝒜casessubscript𝜋0^𝑓CPaths𝒜subscript𝐶subscript^𝜋0subscript𝒜𝜉^𝑓subscript^𝜋0subscript𝜋0\hat{f}^{-1}(C_{\pi_{0},\mathcal{A}})=\begin{cases}\emptyset&\pi_{0}\not\in\hat{f}({\operatorname{{CPaths}}}(\mbox{$\mathcal{A}$}))\\ C_{\hat{\pi}_{0},\mathcal{A}_{\xi}}&\hat{f}(\hat{\pi}_{0})=\pi_{0}\end{cases}.</td>
<td></td>
</tr>
</tbody>
</table>
by Lemma [1](#Thmlemma1 "Lemma 1 ‣ 2 Probabilistic Automata ‣ Probabilistic Anonymity and Admissible Schedulers"). Indeed, let $`\pi_{0} = {\hat{f}\hspace{0pt}{({\hat{\pi}}_{0})}}`$. Then
<table>
<tbody>
<tr>
<td></td>
<td>f^−1(Cπ0,𝒜)superscript^𝑓1subscript𝐶subscript𝜋0𝒜\displaystyle\hat{f}^{-1}(C_{\pi_{0},\mathcal{A}})</td>
<td>=\displaystyle=</td>
<td>{π^∈CPaths(𝒜ξ)∣f^(π^)≥π0}conditional-set^𝜋CPathssubscript𝒜𝜉^𝑓^𝜋subscript𝜋0\displaystyle\{\hat{\pi}\in{\operatorname{{CPaths}}}(\mathcal{A}_{\xi})\mid\hat{f}(\hat{\pi})\geq\pi_{0}\}</td>
<td></td>
</tr>
<tr>
<td></td>
<td></td>
<td>=\displaystyle=</td>
<td>{π^∈CPaths(𝒜ξ)∣f^(π^)≥f^(π^0)}conditional-set^𝜋CPathssubscript𝒜𝜉^𝑓^𝜋^𝑓subscript^𝜋0\displaystyle\{\hat{\pi}\in{\operatorname{{CPaths}}}(\mathcal{A}_{\xi})\mid\hat{f}(\hat{\pi})\geq\hat{f}(\hat{\pi}_{0})\}</td>
<td></td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td>(Lem.1)Lem.1\displaystyle(\text{Lem.}~{}\ref{PrefFLem})</td>
<td>=\displaystyle=</td>
<td>{π^∈CPaths(𝒜ξ)∣π^≥π^0}conditional-set^𝜋CPathssubscript𝒜𝜉^𝜋subscript^𝜋0\displaystyle\{\hat{\pi}\in{\operatorname{{CPaths}}}(\mathcal{A}_{\xi})\mid\hat{\pi}\geq\hat{\pi}_{0}\}</td>
<td></td>
</tr>
<tr>
<td></td>
<td></td>
<td>=\displaystyle=</td>
<td>Cπ^0,𝒜ξsubscript𝐶subscript^𝜋0subscript𝒜𝜉\displaystyle C_{\hat{\pi}_{0},\mathcal{A}_{\xi}}</td>
<td></td>
</tr>
</tbody>
</table>
∎
We next define two operations on probabilistic automata used for building composed models out of basic models: parallel composition and restriction. We compose probabilistic automata in parallel in the style of the process algebra ACP. That is, asynchronously with communication function given by a semigroup operation on the set of actions. This is the most general way of composing probabilistic automata in parallel (for an overview see \[[SV04](#bib.bibx12)\]).
###### Definition 2.8 (Parallel composition)
We fix an action set $`A`$ and a communication function $`\cdot`$ on $`A`$ which is a partial commutative semigroup. Given two probabilistic automata $`\mathcal{A}_{1} = {\langle S_{1},A,\alpha_{1}\rangle}`$ and $`\mathcal{A}_{2} = {\langle S_{2},A,\alpha_{2}\rangle}`$ with actions $`A`$, their parallel composition is the probabilistic automaton $`{\mathcal{A}_{1} \parallel \mathcal{A}_{2}} = {\langle{S_{1} \times S_{2}},A,\alpha\rangle}`$ with states pairs of states of the original automata denoted by $`s_{1} \parallel s_{2}`$, the same actions, and transition function defined as follows. $`{s_{1} \parallel s_{2}}\overset{a}{\rightarrow}\mu`$ if and only if one of the following holds
- 1\.
$`s_{1}\overset{b}{\rightarrow}\mu_{1}`$ and $`s_{2}\overset{c}{\rightarrow}\mu_{2}`$ for some actions $`b`$ and $`c`$ such that $`a = {b \cdot c}`$ and
$`\mu = {\mu_{1} \cdot \mu_{2}}`$ meaning $`{\mu\hspace{0pt}{({t_{1} \parallel t_{2}})}} = {{{\mu_{1}\hspace{0pt}{(t_{1})}} \cdot \mu_{2}}\hspace{0pt}{(t_{2})}}`$.
- 2\.
$`s_{1}\overset{a}{\rightarrow}\mu'`$ where $`{\mu'\hspace{0pt}{(t_{1})}} = {\mu\hspace{0pt}{({t_{1} \parallel s_{2}})}}`$ for all states $`t_{1}`$ of the first automaton.
- 3\.
$`s_{2}\overset{a}{\rightarrow}\mu'`$ where $`{\mu'\hspace{0pt}{(t_{2})}} = {\mu\hspace{0pt}{({s_{1} \parallel t_{2}})}}`$ for all states $`t_{2}`$ of the second automaton.
Here, 1. represents a synchronous joint move of both of the automata, and 2. and 3. represent the possibilities of an asynchronous move of each of the automata. In case $`s_{1}^{0}`$ and $`s_{2}^{0}`$ are the initial states of $`\mathcal{A}_{1}`$ and $`\mathcal{A}_{2}`$, respectively, then the initial state of $`\mathcal{A}_{1} \parallel \mathcal{A}_{2}`$ is $`s_{1}^{0} \parallel s_{2}^{0}`$.
Often we will use input and output actions like $`a\hspace{0pt}?`$ and $`a!`$, respectively, in the style of CCS.In such cases we assume that the communication is defined as hand-shaking $`{{a\hspace{0pt}?} \cdot {a!}} = \tau_{a}`$ for $`\tau_{a}`$ a special invisible action.
The operation of restriction is needed to prune out some branches of a probabilistic automaton that one need not consider. For example, we will commonly use restriction to get rid of parts of a probabilistic automaton that still wait on synchronization.
###### Definition 2.9 (restriction)
Fix a subset $`I \subseteq A`$ of actions that are in the restricted set. Given an automaton $`\mathcal{A} = {\langle S,A,\alpha\rangle}`$, the automaton obtained from $`\mathcal{A}`$ by restricting the actions in $`I`$ is $`{\mathcal{R}_{I}\hspace{0pt}{(\mathcal{A})}} = {\langle S,{A \smallsetminus I},\alpha'\rangle}`$ where the transitions of $`\alpha'`$ are defined as follows: $`s\overset{a}{\rightarrow}\mu`$ in $`\mathcal{R}_{I}\hspace{0pt}{(\mathcal{A})}`$ if and only if $`s\overset{a}{\rightarrow}\mu`$ in $`\mathcal{A}`$ and $`a \notin I`$.
We now define bisimilarity - a behaviour equivalence on the states of a probabilistic automaton. For that we first need the notion of relation lifting.
###### Definition 2.10 (Relation lifting)
Let $`R`$ be an equivalence relation on the set of states $`S`$ of a probabilistic automaton $`\mathcal{A}`$. Then $`R`$ lifts to a relation $`\equiv_{R}`$ on the set $`\mathcal{D}{(S)}`$, as follows:
<table>
<tbody>
<tr>
<td></td>
<td>μ≡Rν⇔∑s∈Cμ(s)=∑s∈Cν(s)iffsubscript𝑅𝜇𝜈subscript𝑠𝐶𝜇𝑠subscript𝑠𝐶𝜈𝑠\mu\equiv_{R}\nu\iff\sum_{s\in C}\mu(s)=\sum_{s\in C}\nu(s)</td>
<td></td>
</tr>
</tbody>
</table>
for any equivalence class $`C \in {S/R}`$.
###### Definition 2.11 (Bisimulation, bisimilarity)
Let $`\mathcal{A} = {\langle\mathcal{S},\mathcal{A},\alpha\rangle}`$ be a probabilistic automaton. An equivalence $`R`$ on its set of states $`S`$ is a bisimulation if and only if whenever $`{\langle s,t\rangle} \in R`$ we have
if $`s\overset{a}{\rightarrow}\mu_{s}`$, then there exists $`\mu_{t}`$ such that $`t\overset{a}{\rightarrow}\mu_{t}`$ and $`\mu_{s} \equiv_{R}\mu_{t}`$.
Two states $`{s,t} \in S`$ are bisimilar, notation $`s \sim t`$ if they are related by some bisimulation relation $`R`$.
Note that bisimilarity $`\sim`$ is the largest bisimulation on a given probabilistic automaton $`\mathcal{A}`$.
## 3 Anonymizing Protocols
### 3.1 Dining cryptographers
The canonical example of an anonymizing protocol is Chaum’s Dining Cryptographers \[[Cha88](#bib.bibx3)\]. In Chaum’s introduction to this protocol, three cryptographers are sitting down to dine in a restaurant, when the waiter informs them that the bill has already been paid anonymously. They wonder whether one of them has paid the bill in advance, or whether the NSA has done so. Respecting each other’s right to privacy, the carry out the following protocol. Each pair of cryptographers flips a coin, invisible to the remaining cryptographer. Each cryptographer then reveals whether or not the two coins he say were equal or unequal. However, if a cryptographer is paying, he states the opposite. An even number of “equals” now indicates that the NSA is paying; an odd number that one of the cryptographers is paying.
Formally, Chaum states the result as follows. (Here we are restricting to the case with 3 cryptographers; Chaum’s version is more general.) Here, $`{\mathbb{F}}_{2}`$ is the field of two elements.
######
Theorem 3.1
Let $`K`$ be a uniformly distributed stochastic variable over $`{\mathbb{F}}_{2}^{3}`$. Let $`I`$ be a stochastic variable over $`{\mathbb{F}}_{2}^{3}`$, taking only values in $`\{{(1,0,0)},{(0,1,0)},{(0,0,1)},{(0,0,0)}\}`$. Let $`A`$ be the stochastic variable over $`{\mathbb{F}}_{2}^{3}`$ given by $`A = {({I_{1} + K_{2} + K_{3}},{K_{1} + I_{2} + K_{3}},{K_{1} + K_{2} + I_{3}})}`$. Assume that $`K`$ and $`I`$ are independent. Then
<table>
<tbody>
<tr>
<td></td>
<td>∀a∈𝔽23∀i∈{1,2,3}:ℙ[I=i]>0⟹ℙ[A=a|I=i]=14:for-all𝑎superscriptsubscript𝔽23for-all𝑖123ℙdelimited-[]𝐼𝑖0ℙdelimited-[]𝐴conditional𝑎𝐼𝑖14\forall a\in{\mathbb{F}}_{2}^{3}\;\forall i\in\{1,2,3\}:{\mathbb{P}}[I=i]>0\implies{\mathbb{P}}[A=a\;|\;I=i]=\tfrac{1}{4}</td>
<td></td>
</tr>
</tbody>
</table>
and hence
<table>
<tbody>
<tr>
<td></td>
<td>∀a∈𝔽23∀i∈{1,2,3}:ℙ[I=i]>0⟹ℙ[A=a|I=i]=ℙ[A=a].:for-all𝑎superscriptsubscript𝔽23for-all𝑖123ℙdelimited-[]𝐼𝑖0ℙdelimited-[]𝐴conditional𝑎𝐼𝑖ℙdelimited-[]𝐴𝑎\forall a\in{\mathbb{F}}_{2}^{3}\;\forall i\in\{1,2,3\}:{\mathbb{P}}[I=i]>0\implies{\mathbb{P}}[A=a\;|\;I=i]={\mathbb{P}}[A=a].</td>
<td></td>
<td>∎</td>
</tr>
</tbody>
</table>
In terms of the storyline, $`K`$ represents the coin flips, $`I`$ represents which cryptographer (if any) is paying, and $`A`$ represents the every cryptographer says.
We will now model this protocol as a probabilistic automaton. We will construct it as a parallel composition of seven components: the Master, who decides who will pay, the three cryptographers Crypt<sub>i</sub>, and the three coins Coin<sub>i</sub>. The action $`p_{i}!`$ is used by the Master to indicate to Crypt<sub>i</sub> that he should pay; the action $`n_{i}!`$ to indicate that he shouldn’t. If no cryptographer is paying, the NSA is paying, which is not explicitly modelled here. The coin Coin<sub>i</sub> is shared by Crypt<sub>i</sub> and Crypt<sub>i-1</sub> (taking the -1 modulo 3); the action $`h_{i,j}!`$ represents Coin<sub>i</sub> signalling to Crypt<sub>j</sub> that the coin was heads and similarly $`t_{i,j}!`$ signals tails. At the end, the cryptographers state whether or not the two coins they saw were equal or not by means of the actions $`a_{i}!`$ (agree) or $`d_{i}`$ (disagree).
<table>
<tbody>
<tr>
<td></td>
<td>img.svg img.svg img.svg</td>
<td></td>
</tr>
</tbody>
</table>
Now DC is the parallel composition of Master, Coin<sub>0</sub>, Coin<sub>1</sub>, Coin<sub>2</sub>, Crypt<sub>0</sub>, Crypt<sub>1</sub>, and Crypt<sub>2</sub> with all actions of the form $`p_{i}`$, $`n_{i}`$, $`h_{i,j}`$, and $`t_{i,j}`$ hidden.
Note that in Chaum’s version, there is no assumption on the probability distribution of $`I`$; in our version this is modelled by the fact that the Master makes a non-deterministic choice between the four options. Since we allow probabilistic schedulers, we later recover all possible probability distributions about who is paying, just as in the original version. Independence between the choice of the master and the coin flips ($`I`$ and $`K`$ in Chaum’s version) comes for free in the automata model: distinct probabilistic choices are always assumed to be independent.
In Section [4](#S4 "4 Anonymity for Purely Probabilistic Systems ‣ Probabilistic Anonymity and Admissible Schedulers") we formulate what it means for DC (or more general, for an anonymity automaton) to be anonymous.
### 3.2 Voting
At a very high level, a voting protocol can be seen as a blackbox that inputs the voters’ votes and outputs the result of the vote. For simplicity, assume the voters vote yes (1) or no (0), do not abstain, and that the numbers of voters is known. The result then is the number of yes-votes.
<table>
<tbody>
<tr>
<td></td>
<td>img.svg</td>
<td></td>
</tr>
</tbody>
</table>
In such a setting, it is conceivable that an observer has some a-priori knowledge about which voters are more likely to vote yes and which voters are more likely to vote no. Furthermore, there definitely is a-posteriori knowledge, since the vote result is made public. For instance, in the degenerate case where all voters vote the same way, everybody’s vote is revealed. What we expect here from the voting protocol is not that the adversary has no knowledge about the votes (since he might already have a-priori knowledge), and also not that the adversary does not gain any knowledge from observing the protocol (since the vote result is revealed), but rather that observing the protocol does not augment the adversary’s knowledge beyond learning the vote result.
For the purely probabilistic case, this notion of anonymity is formalized in Section [4](#S4 "4 Anonymity for Purely Probabilistic Systems ‣ Probabilistic Anonymity and Admissible Schedulers").
## 4 Anonymity for Purely Probabilistic Systems
This section defines anonymity systems and proposes a definition for anonymity in its simplest configuration, i.e., for purely probabilistic systems. Purely probabilistic systems are simpler because there is no need for schedulers. Throughout the following sections, this definitions will be incrementally modified towards a more general setting.
###### Definition 4.1 (Anonymity system)
Let $`M = {\langle S,\text{Act},\alpha\rangle}`$ be a fully probabilistic automaton. An *anonymity system* is a triple $`\langle M,I,{\{ A_{i}\}}_{i \in I},\text{Act}_{O}\rangle`$ where
1. 1\.
$`I`$ is the set of user identities,
2. 2\.
$`A_{i}`$ is any measurable subset of $`{CPaths}{(M)}`$ such that $`{A_{i} \cap A_{j}} = \varnothing`$ for $`i \neq j`$.
3. 3\.
$`\text{Act}_{O} \subseteq \text{Act}`$ is the set of observable actions.
4. 4\.
$`O\hspace{0pt}t\hspace{0pt}r\hspace{0pt}a\hspace{0pt}c\hspace{0pt}e\hspace{0pt}{(\pi)}`$ is the sequence of elements in $`\text{Act}_{O}`$ obtained by removing form $`{trace}{(\pi)}`$ the elements in $`\text{Act} \smallsetminus \text{Act}_{O}`$.
Define $`O`$ as the set of observations, i.e., $`O = \left. \{{{trace}{(\pi)}} \middle| {\pi \in {{Paths}{(M)}}}\} \right.`$. We also define $`A = {\bigcup_{i \in I}A_{i}}`$.
Intuitively, the $`A_{i}`$s are properties of the executions that the system is meant to hide. For example, in the case of the dining cryptographers $`A_{i}`$ would be “cryptographer $`i`$ payed”; in a voting scheme “voter $`i`$ voted for candidate $`c`$”, etc. Therefore, for the previous examples, the predicate $`A`$ would be “some of the cryptographers payed” or “the vote count” respectively.
Next, we propose a definition of anonymity for a purely probabilistic systems. We deviate from the definition proposed by Bhargava and Palamidessi \[[BP05](#bib.bibx2)\] for what we consider a more intuitive definition: We say that an anonymity system is anonymous if the probability of seeing a observation is independent of who performed the anonymous action ($`A_{i}`$), given that some anonymous action took place ($`A`$ happened). The formal definition follows.
###### Definition 4.2 (Anonymity)
A system $`\langle M,I,{\{ A_{i}\}}_{i \in I},\text{Act}_{O}\rangle`$ is said to be anonymous if
<table>
<tbody>
<tr>
<td></td>
<td>∀i∈I.∀o∈O.ℙ[π∈A]>0⟹formulae-sequencefor-all𝑖𝐼for-all𝑜𝑂ℙdelimited-[]𝜋𝐴0absent\displaystyle\forall i\in I.\forall o\in O.{\mathbb{P}}[\pi\in A]>0\implies</td>
<td>ℙ[Otrace(π)=o∧π∈Ai|π∈A]=ℙdelimited-[]𝑂𝑡𝑟𝑎𝑐𝑒𝜋𝑜𝜋conditionalsubscript𝐴𝑖𝜋𝐴absent\displaystyle{\mathbb{P}}[Otrace(\pi)=o\;\land\;\pi\in A_{i}\;|\;\pi\in A]=</td>
<td></td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td></td>
<td>ℙ[Otrace(π)=o|π∈A]ℙ[π∈Ai|π∈A].ℙdelimited-[]𝑂𝑡𝑟𝑎𝑐𝑒𝜋conditional𝑜𝜋𝐴ℙdelimited-[]𝜋conditionalsubscript𝐴𝑖𝜋𝐴\displaystyle{\mathbb{P}}[Otrace(\pi)=o\;|\;\pi\in A]\,{\mathbb{P}}[\pi\in A_{i}\;|\;\pi\in A].</td>
<td></td>
</tr>
</tbody>
</table>
In the above probabilities, $`\pi`$ is drawn from the probability space $`{Paths}{(M)}`$.
The following lemma shows that this definition is equivalent to the one proposed in Bhargava and Palamidessi \[[BP05](#bib.bibx2)\].
###### Lemma 4
A anonymity system is anonymous if and only if
<table>
<tbody>
<tr>
<td></td>
<td>∀i,j∈I.∀o∈O.(ℙ[π∈Ai]>0∧ℙ[π∈Aj]>0)⟹formulae-sequencefor-all𝑖𝑗𝐼for-all𝑜𝑂ℙdelimited-[]𝜋subscript𝐴𝑖0ℙdelimited-[]𝜋subscript𝐴𝑗0absent\displaystyle\forall i,j\in I.\forall o\in O.({\mathbb{P}}[\pi\in A_{i}]>0\land{\mathbb{P}}[\pi\in A_{j}]>0)\implies</td>
<td>ℙ[Otrace(π)=o|π∈Ai]=ℙdelimited-[]𝑂𝑡𝑟𝑎𝑐𝑒𝜋conditional𝑜𝜋subscript𝐴𝑖absent\displaystyle{\mathbb{P}}[Otrace(\pi)=o\;|\;\pi\in A_{i}]=</td>
<td></td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td></td>
<td>ℙ[Otrace(π)=o|π∈Aj]ℙdelimited-[]𝑂𝑡𝑟𝑎𝑐𝑒𝜋conditional𝑜𝜋subscript𝐴𝑗\displaystyle{\mathbb{P}}[Otrace(\pi)=o\;|\;\pi\in A_{j}]\;</td>
<td></td>
</tr>
</tbody>
</table>
###### Proof.
The only if part is trivial. For the if part we have
<table>
<tbody>
<tr>
<td></td>
<td>ℙ[\displaystyle{\mathbb{P}}[</td>
<td>Otrace(π)=o|π∈A]ℙ[π∈Ai|π∈A]\displaystyle Otrace(\pi)=o\;|\;\pi\in A]\,{\mathbb{P}}[\pi\in A_{i}\;|\;\pi\in A]</td>
<td></td>
<td></td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td></td>
<td>=ℙ[π∈Ai|π∈A]∑j∈Iℙ[Otrace(π)=o|π∈Aj∩A]ℙ[π∈Aj|π∈A]absentℙdelimited-[]𝜋conditionalsubscript𝐴𝑖𝜋𝐴subscript𝑗𝐼ℙdelimited-[]𝑂𝑡𝑟𝑎𝑐𝑒𝜋conditional𝑜𝜋subscript𝐴𝑗𝐴ℙdelimited-[]𝜋conditionalsubscript𝐴𝑗𝜋𝐴\displaystyle={\mathbb{P}}[\pi\in A_{i}\;|\;\pi\in A]\;\sum_{j\in I}{\mathbb{P}}[Otrace(\pi)=o\;|\;\pi\in A_{j}\cap A]\;{\mathbb{P}}[\pi\in A_{j}\;|\;\pi\in A]</td>
<td></td>
<td></td>
</tr>
</tbody>
<tbody>
<tr>
<td colspan="5"> (since Ai∩Aj=∅,i≠jformulae-sequencesubscript𝐴𝑖subscript𝐴𝑗𝑖𝑗A_{i}\cap A_{j}=\emptyset,i\not=j)</td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td></td>
<td>=ℙ[π∈Ai|π∈A]∑j∈Iℙ[Otrace(π)=o|π∈Aj]ℙ[π∈Aj|π∈A]absentℙdelimited-[]𝜋conditionalsubscript𝐴𝑖𝜋𝐴subscript𝑗𝐼ℙdelimited-[]𝑂𝑡𝑟𝑎𝑐𝑒𝜋conditional𝑜𝜋subscript𝐴𝑗ℙdelimited-[]𝜋conditionalsubscript𝐴𝑗𝜋𝐴\displaystyle={\mathbb{P}}[\pi\in A_{i}\;|\;\pi\in A]\;\sum_{j\in I}{\mathbb{P}}[Otrace(\pi)=o\;|\;\pi\in A_{j}]\;{\mathbb{P}}[\pi\in A_{j}\;|\;\pi\in A]</td>
<td></td>
<td></td>
</tr>
</tbody>
<tbody>
<tr>
<td colspan="5"> (by definition of π∈A𝜋𝐴\pi\in A)</td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td></td>
<td>=ℙ[π∈Ai|π∈A]ℙ[Otrace(π)=o|π∈Ai]∑j∈Iℙ[π∈Aj|π∈A]absentℙdelimited-[]𝜋conditionalsubscript𝐴𝑖𝜋𝐴ℙdelimited-[]𝑂𝑡𝑟𝑎𝑐𝑒𝜋conditional𝑜𝜋subscript𝐴𝑖subscript𝑗𝐼ℙdelimited-[]𝜋conditionalsubscript𝐴𝑗𝜋𝐴\displaystyle={\mathbb{P}}[\pi\in A_{i}\;|\;\pi\in A]\;{\mathbb{P}}[Otrace(\pi)=o\;|\;\pi\in A_{i}]\;\sum_{j\in I}{\mathbb{P}}[\pi\in A_{j}\;|\;\pi\in A]</td>
<td></td>
<td></td>
</tr>
</tbody>
<tbody>
<tr>
<td colspan="5"> (by hypothesis)</td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td></td>
<td>=ℙ[π∈Ai|π∈A]ℙ[Otrace(π)=o|π∈Ai]∑j∈Iℙ[π∈Aj]ℙ[π∈A]absentℙdelimited-[]𝜋conditionalsubscript𝐴𝑖𝜋𝐴ℙdelimited-[]𝑂𝑡𝑟𝑎𝑐𝑒𝜋conditional𝑜𝜋subscript𝐴𝑖subscript𝑗𝐼ℙdelimited-[]𝜋subscript𝐴𝑗ℙdelimited-[]𝜋𝐴\displaystyle={\mathbb{P}}[\pi\in A_{i}\;|\;\pi\in A]\;{\mathbb{P}}[Otrace(\pi)=o\;|\;\pi\in A_{i}]\;\frac{\sum_{j\in I}{\mathbb{P}}[\pi\in A_{j}]}{{\mathbb{P}}[\pi\in A]}</td>
<td></td>
<td></td>
</tr>
</tbody>
<tbody>
<tr>
<td colspan="5"> (since Aj⊆Asubscript𝐴𝑗𝐴A_{j}\subseteq A)</td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td></td>
<td>=ℙ[π∈Ai|π∈A]ℙ[Otrace(π)=o|π∈Ai]absentℙdelimited-[]𝜋conditionalsubscript𝐴𝑖𝜋𝐴ℙdelimited-[]𝑂𝑡𝑟𝑎𝑐𝑒𝜋conditional𝑜𝜋subscript𝐴𝑖\displaystyle={\mathbb{P}}[\pi\in A_{i}\;|\;\pi\in A]\;{\mathbb{P}}[Otrace(\pi)=o\;|\;\pi\in A_{i}]</td>
<td></td>
<td></td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td></td>
<td>=ℙ[π∈Ai]ℙ[π∈A]ℙ[Otrace(π)=o∧π∈Ai]ℙ[π∈Ai]absentℙdelimited-[]𝜋subscript𝐴𝑖ℙdelimited-[]𝜋𝐴ℙdelimited-[]𝑂𝑡𝑟𝑎𝑐𝑒𝜋𝑜𝜋subscript𝐴𝑖ℙdelimited-[]𝜋subscript𝐴𝑖\displaystyle=\frac{{\mathbb{P}}[\pi\in A_{i}]}{{\mathbb{P}}[\pi\in A]}\;\frac{{\mathbb{P}}[Otrace(\pi)=o\land\pi\in A_{i}]}{{\mathbb{P}}[\pi\in A_{i}]}</td>
<td></td>
<td></td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td></td>
<td>=ℙ[Otrace(π)=o∧π∈Ai|π∈A]absentℙdelimited-[]𝑂𝑡𝑟𝑎𝑐𝑒𝜋𝑜𝜋conditionalsubscript𝐴𝑖𝜋𝐴\displaystyle={\mathbb{P}}[Otrace(\pi)=o\land\pi\in A_{i}\;|\;\pi\in A]</td>
<td></td>
<td></td>
</tr>
</tbody>
<tbody>
<tr>
<td colspan="5"> (since Ai⊆Asubscript𝐴𝑖𝐴A_{i}\subseteq A)</td>
</tr>
</tbody>
</table>
which concludes the proof. ∎∎
## 5 Anonymity for Probabilistic Systems
We now try to extend the notion of anonymity to probabilistic automata that are not purely probabilistic, but that still contain some non-deterministic transitions.
One obvious try is to say that $`M`$ is anonymous if $`M_{\xi}`$ is anonymous for all schedulers $`\xi`$ of $`M`$. The following automaton $`M`$ and scheduler $`\xi`$ show that this definition would be problematic.
<table>
<tbody>
<tr>
<td></td>
<td>img.svg img.svg</td>
<td></td>
</tr>
</tbody>
</table>
Here $`a_{1}`$ and $`a_{2}`$ are invisible actions; they represent which user performed the action that was to remain hidden. The actions $`x_{1}`$ and $`x_{2}`$ are observable. Intuitively, because the adversary cannot see the messages $`a_{1}`$ and $`a_{2}`$, she cannot learn which user actually performed the hidden action. On the right hand side $`M_{\xi}`$ is shown and the branches the scheduler does not take are indicated by dotted arrows. Now $`{{\mathbb{P}}_{\xi}\hspace{0pt}{\lbrack\left. a_{1} \middle| x_{1} \right.\rbrack}} = 1`$, but $`{{\mathbb{P}}_{\xi}\hspace{0pt}{\lbrack a_{1}\rbrack}} = \frac{1}{2}`$, showing that with this particular scheduler $`M_{\xi}`$ is not anonymous.
Note that this phenomenon can easily occur as a consequence of communication non-determinism. For instance, consider the following three automata and their parallel composition in which $`c\hspace{0pt}?`$ and $`c!`$ are hidden. In this example the order of the messages $`x_{1}`$ and $`x_{2}`$ depends on a race-condition, but a scheduler can make it depend on whether $`a_{1}`$ or $`a_{2}`$ was taken. I.e., there exists a scheduler $`\xi`$ such that $`{{\mathbb{P}}_{\xi}\hspace{0pt}{\lbrack\left. {x_{1}\hspace{0pt}x_{2}} \middle| a_{1} \right.\rbrack}} = {{\mathbb{P}}_{\xi}\hspace{0pt}{\lbrack\left. {x_{2}\hspace{0pt}x_{1}} \middle| a_{2} \right.\rbrack}} = 1`$ and hence $`{{\mathbb{P}}_{\xi}\hspace{0pt}{\lbrack\left. {x_{2}\hspace{0pt}x_{1}} \middle| a_{1} \right.\rbrack}} = {{\mathbb{P}}_{\xi}\hspace{0pt}{\lbrack\left. {x_{1}\hspace{0pt}x_{2}} \middle| a_{2} \right.\rbrack}} = 0`$.
<table>
<tbody>
<tr>
<td></td>
<td>img.svg img.svg</td>
<td></td>
</tr>
</tbody>
</table>
In fact, the Dining Cryptographers example from Section [3.1](#S3.SS1 "3.1 Dining cryptographers ‣ 3 Anonymizing Protocols ‣ Probabilistic Anonymity and Admissible Schedulers") suffers from exactly the same problem. The order in which the cryptographers say $`{agree}_{i}`$ or $`{disagree}_{i}`$ is determined by the scheduler and it is possible to have a scheduler that makes the paying cryptographer, if any, go last.
In \[[BP05](#bib.bibx2)\], a system $`M`$ is called anonymous if for all schedulers $`\zeta`$, $`\xi`$, for all observables $`o`$, and for all hidden actions $`a_{i}`$, $`a_{j}`$ such that $`{{\mathbb{P}}_{\zeta}\hspace{0pt}{\lbrack a_{i}\rbrack}} > 0`$ and $`{{\mathbb{P}}_{\xi}\hspace{0pt}{\lbrack a_{j}\rbrack}} > 0`$, $`{{\mathbb{P}}_{\zeta}\hspace{0pt}{\lbrack\left. o \middle| a_{i} \right.\rbrack}} = {{\mathbb{P}}_{\xi}\hspace{0pt}{\lbrack\left. o \middle| a_{j} \right.\rbrack}}`$. This definition, of course, has the same problems as above; in the Dining Cryptographers example in \[[BP05](#bib.bibx2)\] this is solved by fixing the order in which the the cryptographers say $`{agree}_{i}`$ or $`{disagree}_{i}`$. However, also a non-deterministic choice between two otherwise anonymous systems can become non-anonymous with this definition. For instance, let $`P`$ be some anonymous system. For simplicity, assume that $`P`$ is fully probabilistic (e.g., the Dining Cryptographers with a probabilistic master and a fixed scheduler) and let $`P'`$ be a variant of $`P`$ in which the visible actions have been renamed (e.g., the actions $`{agree}_{i}`$ and $`{disagree}_{i}`$ are renamed to $`{equal}_{i}`$ and $`{unequal}_{i}`$). Now consider the probabilistic automaton $`M`$ which non-deterministically chooses between $`P`$ and $`P'`$:
<table>
<tbody>
<tr>
<td></td>
<td>img.svg</td>
<td></td>
</tr>
</tbody>
</table>
This automaton has only two schedulers: the one that chooses the left branch and then executes $`P`$ and the one that chooses the right branch and then executes $`P'`$. Let’s call these schedulers $`l`$ and $`r`$ respectively. Now pick any hidden action $`a_{i}`$ and observable $`o`$ such that $`{{\mathbb{P}}_{l}\hspace{0pt}{\lbrack\left. o \middle| a_{i} \right.\rbrack}} > 0`$. (e.g., $`o = {{agree}_{1}{{disagree}_{2}{agree}_{3}}}`$ and $`a_{i} = {pay}_{1}`$, for which $`{\mathbb{P}}{\lbrack o|{pay}_{1}\rbrack} = \frac{1}{4})`$. Then, nevertheless, $`{{\mathbb{P}}_{r}\hspace{0pt}{\lbrack\left. o \middle| a_{i} \right.\rbrack}} = 0`$, because the observation $`o`$ cannot occur in $`P'`$. So, even though intuitively this system should be anonymous, it is not so according to the definition in \[[BP05](#bib.bibx2)\].
Every time the problem is that the scheduler has access to information it shouldn’t have. When one specifies a protocol by giving a probabilistic automaton, an implementation of this protocol has to implement a scheduler as well. This is especially obvious if the non-determinism originates from communication. When we identify schedulers with adversaries, as is common, it becomes clear that the scheduler should not have access to too much information. In the next section we define a class of schedulers, called *admissible* schedulers that base their scheduling behavior on the information an adversary actually has access to: the observable history of the system.
## 6 Admissible Schedulers
As explained in the previous section, defining anonymity as a condition that should hold true for all possible schedulers is problematic. It is usual to quantify over all schedulers when showing theoretical properties of systems with both probabilities and non-determinism - for example we may say “no matter how the non-determinism is resolved, the probability of an event $`X`$ is at least $`p`$”. However, in the analysis of security protocols, for example with respect to anonymity, we would like to quantify over all possible “realistic” adversaries. These are not all possible schedulers as in our theoretic considerations since such a realistic adversary is not able to see all details of the probabilistic automaton under consideration. Hence, considering that the adversary is any scheduler enables the adversary to leak information where it normally could not. We call such schedulers interfering schedulers. This way protocols that are well-known to be anonymous turn out not to be anonymous. One such example is the dining cryptographers protocol explained above. We show that one gets a better definition of anonymity if one restricts the power of the schedulers, in a realistic way. In this section we define the type of schedulers with restricted power that we consider good enough for showing anonymity of certain protocols. We call these schedulers admissible.
Schedulers with restricted power have been treated in the literature. In general, as explained by Segala in \[[Seg95](#bib.bibx9)\], a scheduler with restricted power is given by defining two equivalences, one on the set of finite paths $`\equiv_{1}`$ and another one $`\equiv_{2}`$ on the set of possible transition, in this case $`\mathcal{D}{({A \times S})}`$. Then a scheduler $`\xi`$ is oblivious relative to $`\langle \equiv_{1}, \equiv_{2}\rangle`$ if and only if for any two paths $`\pi_{1},\pi_{2}`$ we have
<table>
<tbody>
<tr>
<td></td>
<td>π1≡1π2⟹ξ(π1)≡2ξ(π2).subscript1subscript𝜋1subscript𝜋2⟹𝜉subscript𝜋1subscript2𝜉subscript𝜋2\pi_{1}\equiv_{1}\pi_{2}\Longrightarrow\xi(\pi_{1})\equiv_{2}\xi(\pi_{2}).</td>
<td></td>
</tr>
</tbody>
</table>
### 6.1 Admissible schedulers based on bisimulation
In this section we specify $`\equiv_{1}`$ and $`\equiv_{2}`$ and obtain a class of oblivious adversaries that suits the anonymity definition.
Defined $`\equiv_{1}`$ on the set of finite paths of an automaton $`M`$ as,
<table>
<tbody>
<tr>
<td></td>
<td>π1≡1π2⇔(trace(π1)=trace(π2)∧last(π1)∼last(π2)).iffsubscript1subscript𝜋1subscript𝜋2tracesubscript𝜋1tracesubscript𝜋2lastsubscript𝜋1similar-tolastsubscript𝜋2\pi_{1}\equiv_{1}\pi_{2}\iff\big{(}\operatorname{trace}(\pi_{1})=\operatorname{trace}(\pi_{2})\land\operatorname{last}(\pi_{1})\sim\operatorname{last}(\pi_{2})\big{)}.</td>
<td></td>
</tr>
</tbody>
</table>
Recall that we defined $`\equiv_{R}`$ as the lifting of the equivalence relation $`R`$ on a set $`S`$ to an equivalence relation on $`\mathcal{D}{({A \times S})}`$. For $`\equiv_{2}`$ we take the equivalence $`\equiv_{\sim}`$ on $`\mathcal{D}{({A \times S})}`$. This is well defined since bisimilarity is an equivalence. Hence, we obtain a class of oblivious schedulers relative $`\langle \equiv_{1}, \equiv_{\sim}\rangle`$. These schedulers we call admissible.
###### Definition 6.1 (admissible scheduler)
A scheduler is admissible if for any two finite paths $`\pi_{1}`$ and $`\pi_{2}`$ we have
<table>
<tbody>
<tr>
<td></td>
<td>(trace(π1)=trace(π2)∧last(π1)∼last(π2))⟹ξ(π1)≡∼ξ(π2).⟹tracesubscript𝜋1tracesubscript𝜋2lastsubscript𝜋1similar-tolastsubscript𝜋2𝜉subscript𝜋1subscriptsimilar-to𝜉subscript𝜋2\big{(}\operatorname{trace}(\pi_{1})=\operatorname{trace}(\pi_{2})\land\operatorname{last}(\pi_{1})\sim\operatorname{last}(\pi_{2})\big{)}\Longrightarrow\xi(\pi_{1})\equiv_{\sim}\xi(\pi_{2}).</td>
<td></td>
</tr>
</tbody>
</table>
Intuitively, the definition of a admissible scheduler enforces that in cases when the schedular has observed the same history (given by the traces of the paths) and is in bisimilar states, it must schedule “the same” transitions up to bisimilarity.
### 6.2 Existence
We now show that admissible schedulers do exist. In fact, we even show that admissible history-independent schedulers exist. A scheduler $`\xi`$ is history-independent if it is completely determined by its image of paths of length 0 i.e. if for any path $`\pi`$ it holds that $`{\xi\hspace{0pt}{(\pi)}} = {\xi\hspace{0pt}{({{last}{(\pi)}})}}`$.
######
Theorem 6.2
(Existence)
There exists a admissible scheduler for every probabilistic automaton.
###### Proof.
Take a probabilistic automaton $`M`$. We first show that there exists a map $`\xi:{S\rightarrow{{\mathcal{D}{({A \times S})}} \cup {\{\bot\}}}}`$ with the property that $`{\xi\hspace{0pt}{(s)}} = \bot`$ if and only if $`s`$ terminates and for all $`{s,t} \in S`$, if $`s \sim t`$, then $`{\xi\hspace{0pt}{(s)}} \equiv_{\sim}{\xi\hspace{0pt}{(t)}}`$.
Consider the set of partial maps
<table>
<tbody>
<tr>
<td></td>
<td>Ξ={ξ:S↪𝒟(A×S)∪{⊥}|ξ(s)=⊥⇔s terminates ,s∼t⟹ξ(s)≡∼ξ(t) for s,t∈dom(ξ)}.\Xi=\left\{\xi:S\hookrightarrow{\operatorname{{\mathcal{D}}}}(A\times S)\cup\{\bot\}\left|\begin{array}[]{ll}&\xi(s)=\bot\iff s\text{~{}terminates~{}},\\ &s\sim t\Longrightarrow\xi(s)\equiv_{\sim}\xi(t)\\ &\text{~{}for~{}}s,t\in dom(\xi)\end{array}\right\}\right..</td>
<td></td>
</tr>
</tbody>
</table>
This set is not empty since the unique partial map with empty domain belongs to it. We define an order $`\leq`$ on $`\Xi`$ in a standard way by
<table>
<tbody>
<tr>
<td></td>
<td>ξ1≤ξ2⇔(dom(ξ1)⊆dom(ξ2)∧ξ2|dom(ξ1)=ξ1).iffsubscript𝜉1subscript𝜉2𝑑𝑜𝑚subscript𝜉1𝑑𝑜𝑚subscript𝜉2evaluated-atsubscript𝜉2𝑑𝑜𝑚subscript𝜉1subscript𝜉1\xi_{1}\leq\xi_{2}\iff\big{(}dom(\xi_{1})\subseteq dom(\xi_{2})\land\xi_{2}|_{dom(\xi_{1})}=\xi_{1}\big{)}.</td>
<td></td>
</tr>
</tbody>
</table>
Consider a chain $`{(\xi_{i})}_{i \in I}`$ in $`\Xi`$. Let $`\xi = {\cup_{i \in I}\xi_{i}}`$. This means that $`{d\hspace{0pt}o\hspace{0pt}m\hspace{0pt}{(\xi)}} = {\cup_{i \in I}{d\hspace{0pt}o\hspace{0pt}m\hspace{0pt}{(\xi_{i})}}}`$ and if $`x \in {d\hspace{0pt}o\hspace{0pt}m\hspace{0pt}{(\xi)}}`$, then $`{\xi\hspace{0pt}{(x)}} = {\xi_{i}\hspace{0pt}{(x)}}`$ for $`i \in I`$ such that $`x \in {d\hspace{0pt}o\hspace{0pt}m\hspace{0pt}{(\xi_{i})}}`$. Note that $`\xi`$ is well-defined since $`{(\xi_{i})}_{i \in I}`$ is a chain. Moreover, it is obvious that $`\xi_{i} \leq \xi`$ for all $`i \in I`$. We next check that $`\xi \in \Xi`$. Let $`{s,t} \in {d\hspace{0pt}o\hspace{0pt}m\hspace{0pt}{(\xi)}}`$, such that $`s \sim t`$. Then $`s \in {d\hspace{0pt}o\hspace{0pt}m\hspace{0pt}{(\xi_{i})}}`$ and $`t \in {d\hspace{0pt}o\hspace{0pt}m\hspace{0pt}{(\xi_{j})}}`$ for some $`{i,j} \in I`$ and either $`\xi_{1} \leq \xi_{2}`$ or $`\xi_{2} \leq \xi_{1}`$. Assume $`\xi_{1} \leq \xi_{2}`$. Then $`{s,t} \in {d\hspace{0pt}o\hspace{0pt}m\hspace{0pt}{(\xi_{j})}}`$ and $`\xi_{j} \in \Xi`$ so we have that $`{\xi_{j}\hspace{0pt}{(s)}} \equiv_{\sim}{\xi_{j}\hspace{0pt}{(t)}}`$ showing that $`{\xi\hspace{0pt}{(s)}} \equiv_{\sim}{\xi\hspace{0pt}{(t)}}`$ and we have established that $`\xi \in \Xi`$.
Hence, every ascending chain in $`\Xi`$ has an upper bound. By the Lemma of Zorn we conclude that $`\Xi`$ has a maximal element. Let $`\sigma`$ be such a maximal element in $`\Xi`$. We claim that $`\sigma`$ is a total map. Assume opposite, i.e., there exists $`s \in {S \smallsetminus {d\hspace{0pt}o\hspace{0pt}m\hspace{0pt}{(\sigma)}}}`$. If there exists a $`t \in {d\hspace{0pt}o\hspace{0pt}m\hspace{0pt}{(\sigma)}}`$ such that $`s \sim t`$ then we define a new partial scheduler $`\sigma'`$ as follows. If $`{\sigma\hspace{0pt}{(t)}} = \bot`$ we put $`{\sigma'\hspace{0pt}{(s)}} = \bot`$. If $`{\sigma\hspace{0pt}{(t)}} = \mu_{t}`$, then, since $`t\rightarrow\mu_{t}`$ and $`s \sim t`$, there exists $`\mu_{s}`$ such that $`s\rightarrow\mu_{s}`$ and $`\mu_{t} \equiv_{\sim}\mu_{s}`$. In this case we put $`{\sigma'\hspace{0pt}{(s)}} = \mu_{s}`$. Moreover, put $`{\sigma'\hspace{0pt}{(x)}} = {\sigma\hspace{0pt}{(x)}}`$ for $`x \in {d\hspace{0pt}o\hspace{0pt}m\hspace{0pt}{(\sigma)}}`$. Then we have $`\sigma' > \sigma`$ and $`\sigma' \in \Xi`$ contradicting the maximality of $`\sigma`$. Hence $`\sigma`$ is a total map.
Finally, we consider the (history-independent) scheduler $`\hat{\sigma}`$ induced by $`\sigma`$, i.e., defined by $`{\hat{\sigma}\hspace{0pt}{(\pi)}} = {\sigma\hspace{0pt}{({{last}{(\pi)}})}}`$ for any finite path $`\pi`$. This scheduler is admissible. Namely, given $`\pi_{1}`$ and $`\pi_{2}`$ such that $`{{trace}{(\pi_{1})}} = {{trace}{(\pi_{2})}}`$ and $`{{last}{(\pi_{1})}} \sim {{last}{(\pi_{2})}}`$ we have, since $`\sigma \in \Xi`$, that
<table>
<tbody>
<tr>
<td></td>
<td>σ^(π1)=σ(last(π1))≡∼σ(last(π2))=σ^(π2)^𝜎subscript𝜋1𝜎lastsubscript𝜋1subscriptsimilar-to𝜎lastsubscript𝜋2^𝜎subscript𝜋2\hat{\sigma}(\pi_{1})=\sigma(\operatorname{last}(\pi_{1}))\equiv_{\sim}\sigma(\operatorname{last}(\pi_{2}))=\hat{\sigma}(\pi_{2})</td>
<td></td>
</tr>
</tbody>
</table>
which completes the proof.
∎∎
We are now ready to define anonymity for probabilistic systems, the formal definition follows.
###### Definition 6.3 (Anonymity)
A system $`\langle M,I,{\{ A_{i}\}}_{i \in I},\text{Act}_{O}\rangle`$ is said to be anonymous if for all admissible schedulers $`\xi`$, for all $`i \in I`$ and for all $`o \in O`$
<table>
<tbody>
<tr>
<td></td>
<td>ℙξ[π∈A]>0⟹subscriptℙ𝜉delimited-[]𝜋𝐴0absent\displaystyle{\mathbb{P}}_{\xi}[\pi\in A]>0\implies</td>
<td>ℙξ[Otrace(π)=o∧π∈Ai|π∈A]=subscriptℙ𝜉delimited-[]𝑂𝑡𝑟𝑎𝑐𝑒𝜋𝑜𝜋conditionalsubscript𝐴𝑖𝜋𝐴absent\displaystyle{\mathbb{P}}_{\xi}[Otrace(\pi)=o\;\land\;\pi\in A_{i}\;|\;\pi\in A]=</td>
<td></td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td></td>
<td>ℙξ[Otrace(π)=o|π∈A]ℙξ[π∈Ai|π∈A].subscriptℙ𝜉delimited-[]𝑂𝑡𝑟𝑎𝑐𝑒𝜋conditional𝑜𝜋𝐴subscriptℙ𝜉delimited-[]𝜋conditionalsubscript𝐴𝑖𝜋𝐴\displaystyle{\mathbb{P}}_{\xi}[Otrace(\pi)=o\;|\;\pi\in A]\,{\mathbb{P}}_{\xi}[\pi\in A_{i}\;|\;\pi\in A].</td>
<td></td>
</tr>
</tbody>
</table>
## 7 Anonymity Examples
In the purely non-deterministic setting, anonymity of a system is often proved (or defined) as follows: take two users $`A`$ and $`B`$ and a trace in which user $`A`$ is “the culprit”. Now find a trace that looks the same to the adversary, but in which user $`B`$ is “the culprit” \[[HO03](#bib.bibx7), [GHvRP05](#bib.bibx5), [MVdV04](#bib.bibx8), [HK07](#bib.bibx6)\]. In fact, this new trace is often most easily obtained by switching the behavior of $`A`$ and $`B`$.
In this section, we make this technique explicit for anonymity in our setting, with mixed probability and non-determinism.
###### Definition 7.1
Let $`M`$ be a probabilistic automaton. A map $`\alpha:{S\rightarrow S}`$ is called an $`\text{Act}_{O}`$-automorphism if $`\alpha`$ induces an automorphism of the automation $`M_{\tau}`$, which is a copy of $`M`$ with all actions not in $`\text{Act}_{O}`$ renamed to $`\tau`$.
The following result generalized the above-mentioned proof technique that is commonly used for a purely non-deterministic setting.
######
Theorem 7.2
Consider an anonymity system $`(M,I,\text{Act}_{O})`$. Suppose that for every $`{i,j} \in I`$ there exists a $`\text{Act}_{O}`$-automorphism $`\alpha:{S\rightarrow S}`$ such that $`{\alpha\hspace{0pt}{(A_{i})}} = A_{j}`$. Then the system is anonymous.
### Anonymity of the Dining Cryptographers
We can now apply the techniques from the previous section to the Dining Cryptographers. Concretely, we show that there exists a $`\text{Act}_{O}`$-automorphism exchanging the behaviour of the Crypt<sub>1</sub> and Crypt<sub>2</sub>; by symmetry, the same holds for the other two combinations.
Consider the endomorphisms of Master and Coin<sub>2</sub> indicated in the following figure. The states in the left copy that are not explicitly mapped (by a dotted arrow) to a state in the right copy are mapped to themselves.
<table>
<tbody>
<tr>
<td></td>
<td> img.svg</td>
<td></td>
</tr>
</tbody>
</table>
<table>
<tbody>
<tr>
<td></td>
<td>img.svg</td>
<td></td>
</tr>
</tbody>
</table>
Also consider the identity endomorphism on Crypt<sub>i</sub> (for $`i = {0,1,2}`$) and on Coin<sub>i</sub> (for $`i = {0,1}`$). Taking the product of these seven endomorphisms, we obtain an endomorphism $`\alpha`$ of DC.
## References
- \[Bel57\] R. Bellman. A markovian decision process.
Journal of Mathematics and Mechanics, 6, 1957.
- \[BP05\] Mohit Bhargava and Catuscia Palamidessi. Probabilistic anonymity.
In Martín Abadi and Luca de Alfaro, editors, Concurency Theory, 16th International Conference (CONCUR ’05), volume 3653 of Lecture Notes in Computer Science, pages 171–185. Springer, 2005.
- \[Cha88\] David Chaum. The dining cryptographers problem: Unconditional sender and recipient untraceability. 1(1):65–75, 1988.
- \[Che06\] Ling Cheung.
Reconciling nondeterministic and probabilistic choices.
PhD thesis, RU Nijmegen, 2006.
- \[GHvRP05\] Flavio D. Garcia, Ichiro Hasuo, Peter van Rossum, and Wolter Pieters. Provable anonymity.
In Ralf Küsters and John Mitchell, editors, Proceedings of the 2005 ACM Workshop on Formal Methods in Security Engineering (FMSE ’05), pages 63–72. ACM, 2005.
- \[HK07\] Ichiro Hasuo and Yoshinobu Kawabe. Probabilistic anonymity via coalgebraic simulations.
In European Symposium on Programming (ESOP ’07), volume 4421 of Lecture Notes in Computer Science, pages 379–394. Springer, Berlin, 2007.
- \[HO03\] Joseph Halpern and Kevin O’Neill. Anonymity and information hiding in multiagent systems.
In 16th IEEE Computer Security Foundations Workshop (CSFW ’03), pages 75–88, 2003.
- \[MVdV04\] S. Mauw, J. Verschuren, and E.P. de Vink. A formalization of anonymity and onion routing.
In P. Samarati, P. Ryan, D. Gollmann, and R. Molva, editors, Proceedings of Esorics 2004, volume 3193 of Lecture Notes in Computer Science, pages 109–124, 2004.
- \[Seg95\] R. Segala.
Modeling and verification of randomized distributed real-time systems.
PhD thesis, MIT, 1995.
- \[SL94\] R. Segala and N.A. Lynch. Probabilistic simulations for probabilistic processes.
In Proc. Concur’94, pages 481–496. LNCS 836, 1994.
- \[Sok05\] A. Sokolova.
Coalgebraic analysis of probabilistic systems.
PhD thesis, TU Eindhoven, 2005.
- \[SV04\] A. Sokolova and E.P. de Vink. Probabilistic automata: system types, parallel composition and comparison.
In C. Baier, B.R. Haverkort, H. Hermanns, J.-P. Katoen, and M. Siegle, editors, Validation of Stochastic Systems: A Guide to Current Research, pages 1–43. LNCS 2925, 2004. |
0705.2876 | cs.CR | cyber | # An Online Algorithm for Generating Fractal Hash Chains Applied to Digital Chains of Custody
<sup>1</sup>
<sup>1</sup>1 An extended abstract of this paper is to appear in the Intelligence and Security Informatics 2007 (ISI 2007) Conference.
Phillip G. Bradford
<sup>2</sup>
<sup>2</sup>2 Department of Computer Science, The University of Alabama, 101 Houser Hall, Box 870290, Tuscaloosa, AL 35487-0290. { pgb, DanielRay } @cs.ua.edu
Daniel A. Ray
<sup>2</sup>
<sup>2</sup>2 Department of Computer Science, The University of Alabama, 101 Houser Hall, Box 870290, Tuscaloosa, AL 35487-0290. { pgb, DanielRay } @cs.ua.edu
###### Abstract
This paper gives an online algorithm for generating Jakobsson’s fractal hash chains \[[14](#bib.bib14)\]. Our new algorithm compliments Jakobsson’s fractal hash chain algorithm for preimage traversal since his algorithm assumes the entire hash chain is precomputed and a particular list of $`\lceil{\log n}\rceil`$ hash elements or pebbles are saved. Our online algorithm for hash chain traversal incrementally generates a hash chain of $`n`$ hash elements without knowledge of $`n`$ before it starts. For any $`n`$, our algorithm stores only the $`\lceil{\log n}\rceil`$ pebbles which are precisely the inputs for Jakobsson’s amortized hash chain preimage traversal algorithm. This compact representation is useful to generate, traverse, and store a number of large digital hash chains on a small and constrained device. We also give an application using both Jakobsson’s and our new algorithm applied to digital chains of custody for validating dynamically changing forensics data.
##### Keywords:
Fractal hash chains, hash chain preimage traversal, hash chain traversal, digital chains of custody, digital forensics, time-stamping.
## 1 Introduction
This paper proposes a digital forensics system for proactively capturing time-sensitive digital forensics evidence. Intuitively, we are trying to build fully digital chains of custody for digital evidence that provide forensics investigators with opportunities to enhance their data’s integrity. Perhaps this system is best applied to monitor suspects that have already been identified. In some cases, it is important to capture and validate dynamic evidentiary data. This may range from tracking a very fast virus infection to tracking more slow changes directly made by human touch. Even in monitoring slow changes to a file, capturing and validating an inappropriate email–as it is generated–is far more convincing than just getting a snapshot of it retrospectively.
Generally, there is a web-of-trust for validating classical forensic evidence. This web may include both witness and expert testimony as well as logical decuction given basic facts about a situation. The evidence in this web is held together by a chain of custody. A chain of custody is careful documentation of the evidence including details of all transfers of its possession for examination. A chain of custody is used to authenticate evidentiary exhibits as well as to verify these exhibits have not been modified.
Our proposed system also initially depends on a web-of-trust. In particular, we must trust the system administrators or even law enforcement for initiating our system under proper circumstances. This may be fleshed out in the usual ways using witnesses, experts, and logical decuction given facts of the case. We maintain a full hash chain of the dynamic evidence for the following reasons: (1) to document and sychronize the dynamic and time-specific nature of change in a system, (2) to allow repeated verifications of the data under scrutiny, and (3) in the cases where the hash elements include diverse swaths of compressed data may give extensive opportunites for validation using webs-of-trust.
Our system is based on carefully timed hash chains on small constrained tamper-resistant devices. These timed hash chains are constructed with compressed snapshots of evidence while using (suspected) one-way hash functions. Generally, the amount of data or the length of time for data capture is not known in advance for proactively monitoring forensics data. Thus, the constrained nature of the hardware, variable granularity of the data, and unknown: time and number of hash chains makes it likely that these devices cannot hold sufficient complete hash chains. Likewise, such devices cannot hold a large set of time-stamps with compressed digital finger-prints such as hashs.
It is standard forensics practice to use at least three well-known hash functions to verify data integrity in case one or more of these functions is ever compromised \[[5](#bib.bib5), [17](#bib.bib17)\]. Such integrity checks may be used for verifying a disk’s data. In our case, they may also be used to validate dynamic changes in a critical file. Alternatively, hash chains use deferred disclosure to establish time-synchronized authenticity. A hash chain can chronologically document evidence by generating a timed hash chain forward while including diverse data of interest in the hash inputs. Such evidence may be verified by traversing the hash chain backwards while supplying the correct inputs at the proper times. This may include data that may be logically deduced about data captured in the broader system itself. The backwards hash chain traversal allows repeated verification by only going back in the hash chain as far as appropriate–leaving the other concealed hash elements in the chain for subsequent deferred verification. Data may be verified by both sides in a trial, in addition to the forensic investigator. To preserve data integrity, each time data is verified, each constrained device will release another hash element generated prior to all already verified hash elements. This deferred disclosure validates the known elements in the hash chain and it may tie in to different webs-of-trust.
Another approach is to have each device hold the first and last hash element of a hash chain. Generally, a forensics verifier would start at the beginning of the hash chain and present the appropriate data to verify the entire hash chain or a large subset of the hash chain. However, such large subsets of the hash chain should include either the first or last element of the hash chain for validity and verification.
Assuming the investigator or system administrator that initiated the hash chain-based data capture is trustworthy. The data then is in the domain of the evidence clerk maintaining the chain of custody. Subsequently, the hash chain evidence may be called in to question: is this the correct data? The first challenge to this approach is if the start (or an early element) of the hash chain is discovered by an attacker, then another fictitious constrained device may be created that ‘verifies’ incorrect or modified data. In evidence storage, replacing a gun with another with different serial number or bullet grooves is hard not to miss. That is, given the first element of a trusted hash chain gives an attacker an opportunity to generate and fake the rest of the hash chain, except for the last element. The last hash element may be at the center of the contention. If there is a single end element of the hash chain that disagrees, whose do we trust? Thus, we have three hash chains that only reveal their elements on-demand by deferred validation: the trusted investigator or system administrator can validate the first or early hash elements. In fact, ideally there would be enough hash elements that they would not have to reveal the last hash element. This situation is reminicant of the motivation for Schneier and Kelsey \[[24](#bib.bib24)\]’s audit logs.
Another possibility is to determine the first time step that verifies critical data. However, authenticating the first critical data item requires the prior hash element. If it is useful to validate this ‘prior hash element,’ we must traverse the hash chain back two hash elements from the critical data, etc. This is the second challenge to storing the first and last hash elements of a hash chain on a constrained device: repeatedly validating critical data each time the data is examined through deferred disclosure requires us to traverse the hash chain backwards.
Computer forensics strives to understand the relationships between suspects and events so they can be verified by third parties such as jurors in a court of law. In such situations it is critical to establish strong credence of the validity of digital forensics evidence. Digital evidence is abstract, ephemeral, time-sensitive, compact, complex, and often encoded. In some cases, biological evidence has measurable decay characteristics that allow chronologic analysis \[[20](#bib.bib20)\]. Digital evidence does not have such measurable decay characteristics. Also, digital evidence is easily copied and copies may be readily manipulated to challenge valid evidence and diminish its credibility.
[Refer to caption](/html/0705.2876/assets/x1.png)
Figure 1: Two Critical Time Boundaries $`t_{0}`$ and $`t_{1}`$ for Evidence $`\mathcal{E}`$. Time increases from left to right and assume $`t_{1} - t_{0}`$ may be small, for instance as small as $`10^{- 6}`$ of a second.
The digital forensics system proposed here is applied to maintaining timed digital evidence. Figure [1](#S1.F1 "Figure 1 ‣ 1 Introduction ‣ An Online Algorithm for Generating Fractal Hash Chains Applied to Digital Chains of Custody11footnote 1 An extended abstract of this paper is to appear in the Intelligence and Security Informatics 2007 (ISI 2007) Conference.") illustrates a central application addressed in this paper. This challenge is the interval time-stamping problem \[[1](#bib.bib1), [11](#bib.bib11), [18](#bib.bib18), [28](#bib.bib28)\]. Consider the left-to-right time line containing the time interval $`\lbrack t_{0},t_{1}\rbrack`$. A central challenge is to demonstrate that the evidence $`\mathcal{E}`$ was in a particular user’s possession between times $`t_{0}`$ and $`t_{1}`$. The value $`t_{1} - t_{0}`$ may be very small. It is well known that when using public key systems, along with prominent and well distributed data, it is easy to show $`\mathcal{E}`$ existed after time $`t_{1}`$. However, how does one show $`\mathcal{E}`$ existed at or before time $`t_{0}`$? One approach is to divide the interval $`\left\lbrack {t_{0} - \frac{t_{1} - t_{0}}{k}},t_{1} \right\rbrack`$ into $`k + 1`$ smaller intervals each of size $`\frac{t_{1} - t_{0}}{k}`$. In each smaller interval $`\lbrack t_{0,i},t_{1,i}\rbrack`$ it may be shown that $`\mathcal{E}`$ existed after time $`t_{1,i}`$, for each $`i:{k > i \geq 1}`$ where $`t_{1,0} = t_{1}`$. Thus, we can show $`\mathcal{E}`$ existed before time $`t_{1,k} = t_{0}`$.
Throughout this paper, all logs are base 2.
### 1.1 Our Contributions
This paper proposes constrained devices for securing and validating time-sensitive (and dynamic) forensic data. It assumes tamper-resistant hardware, which may be viewed as a dependency on a trusted third party. A special online algorithm is given here to prevent a modify-and-copy attack. This online algorithm allows computer forensics specialists to maintain the verifiability of timed digital evidence.
Take an easy to compute hash function $`h`$. Assume $`h`$ is one-way \[[10](#bib.bib10)\]: So, on average it is intractable to invert $`h`$. That is, given $`y`$ where $`y\leftarrow{h\hspace{0pt}{(x)}}`$ it is on average intractable to find $`x`$ given $`y`$. Given a hash function $`h`$ where $`x\leftarrow{h\hspace{0pt}{(y)}}`$, then $`y`$ is the preimage of $`x`$.
Next is a $`n + 1`$ element hash chain:
<table>
<tbody>
<tr>
<td></td>
<td>v0⟵h(v1)v1⟵h(v2)v2⟵h(v3)⋯⟵h(vn−1)vn−1⟵h(vn)vn.\displaystyle v_{0}\ \stackrel{{\scriptstyle h(v_{1})}}{{\longleftarrow}}\ v_{1}\ \stackrel{{\scriptstyle h(v_{2})}}{{\longleftarrow}}\ v_{2}\ \stackrel{{\scriptstyle h(v_{3})}}{{\longleftarrow}}\ \ \cdots\stackrel{{\scriptstyle h(v_{n-1})}}{{\longleftarrow}}\ v_{n-1}\ \stackrel{{\scriptstyle h(v_{n})}}{{\longleftarrow}}\ v_{n}.</td>
<td></td>
<td>(1)</td>
</tr>
</tbody>
</table>
Since the hash function is one-way, this hash chain must be initially generated from right to left.
######
Definition 1
Consider a hash chain as described by Equation [1](#S1.E1 "In 1.1 Our Contributions ‣ 1 Introduction ‣ An Online Algorithm for Generating Fractal Hash Chains Applied to Digital Chains of Custody11footnote 1 An extended abstract of this paper is to appear in the Intelligence and Security Informatics 2007 (ISI 2007) Conference."). The value $`v_{n}`$ is the seed of the hash chain.
Generating the hash values forward in this order $`v_{n},v_{n - 1},\cdots,v_{0}`$ is hash chain traversal.
Computing hash elements backward in this order $`v_{0},v_{1},\cdots,v_{n}`$ is preimage traversal.
For efficient hash chain traversal see \[[13](#bib.bib13)\]. We always assume the hash function is well known. Using deferred disclosure of hash elements backwards validates prior knowledge of elements in the hash chain. In time step 0, given $`v_{0}`$, then waiting to time step $`1`$ one can verify that $`v_{0} = {h\hspace{0pt}{(v_{1})}}`$ indicating with high probability $`v_{0}`$ and $`v_{1}`$ are from the same source.
At time step $`i`$, our scheme inputs a chunk of digital evidence $`\mathcal{E}_{i}`$. Let $`c\hspace{0pt}{(\mathcal{E}_{i})}`$ be the compressed version of $`\mathcal{E}_{i}`$. Say $`c\hspace{0pt}{(\mathcal{E}_{i})}`$ contains a modest number of fixed bits, for instance 160 bits \[[26](#bib.bib26), [27](#bib.bib27)\]. For example, for the function $`c`$ we can use a technique such as Merkle-Damgård construction of a collision-resistant compression function. Then $`v_{i} \oplus {c\hspace{0pt}{(\mathcal{E}_{i})}}`$ is the input to the hash function $`h`$ giving:
<table>
<tbody>
<tr>
<td></td>
<td>vi−1subscript𝑣𝑖1\displaystyle v_{i-1}</td>
<td>←←\displaystyle\leftarrow</td>
<td>h(vi⊕c(ℰi))ℎdirect-sumsubscript𝑣𝑖𝑐subscriptℰ𝑖\displaystyle h(v_{i}\oplus c({\cal E}_{i}))</td>
<td></td>
</tr>
</tbody>
</table>
and this process is continued in a carefully timed fashion to give an entire hash chain. The function ‘$`\oplus`$’ may be either xor or concatenation.
If a hash chain is completely exposed an adversary has access to all elements of
<table>
<tbody>
<tr>
<td></td>
<td>C1subscript𝐶1\displaystyle C_{1}</td>
<td>=\displaystyle=</td>
<td>vn,vn−1,vn−2,⋯,v0subscript𝑣𝑛subscript𝑣𝑛1subscript𝑣𝑛2⋯subscript𝑣0\displaystyle v_{n},v_{n-1},v_{n-2},\cdots,v_{0}</td>
<td></td>
</tr>
</tbody>
</table>
then this hash chain’s relative and carefully timed deferred disclosure based authenticity may be easily challenged. For instance, an attacker may take $`v_{n - k}`$ and falsify the input $`\mathcal{E}_{n - k}`$ by changing it to $`\mathcal{E}_{n - k}'`$, then generate the incorrect hash chain:
<table>
<tbody>
<tr>
<td></td>
<td>C2subscript𝐶2\displaystyle C_{2}</td>
<td>=\displaystyle=</td>
<td>vn,⋯,vn−k,vn−k−1′,⋯,v0′subscript𝑣𝑛⋯subscript𝑣𝑛𝑘superscriptsubscript𝑣𝑛𝑘1′⋯superscriptsubscript𝑣0′\displaystyle v_{n},\cdots,v_{n-k},v_{n-k-1}^{\prime},\cdots,v_{0}^{\prime}</td>
<td></td>
</tr>
</tbody>
</table>
where $`v_{n - k - i}' \neq v_{n - k - i}`$, for all $`i:{{n - k} \geq i \geq 0}`$.
Now, without proofs of identity or authenticity, which chain $`C_{1}`$ or $`C_{2}`$ represents the authenticated data is not clear. Of course, proofs of identity or authenticity are only as good as the systems checking them.
Given a hash chain with hash seed $`v_{n}`$ and the last hash element $`v_{0}`$. Suppose we only save $`v_{n}`$ and $`v_{0}`$ on a small constrained device to validate the data from $`n`$ different times
<table>
<tbody>
<tr>
<td></td>
<td>ℰn−1,⋯,ℰ0.subscriptℰ𝑛1⋯subscriptℰ0\displaystyle{\cal E}_{n-1},\ \cdots,\ {\cal E}_{0}.</td>
<td></td>
</tr>
</tbody>
</table>
If $`\mathcal{E}_{i}`$ is the first piece of critical data in hash element $`v_{i}`$, then to validate $`v_{i}`$, we must know $`v_{i + 1}`$. But, the next time the same hash chain is validated, we take $`\mathcal{E}_{i + 1}`$ and validate it with $`v_{i + 2}`$. As these validation steps are repeated each time investigators weigh the evidence, we will be performing a (slow and deliberate) preimage traversal of the hash chain.
The hash chain forward traversal algorithm given in this paper extends Jakobsson’s work by augmenting his hash chain preimage backward traversal algorithm \[[14](#bib.bib14)\]. See also, \[[7](#bib.bib7)\]. Our algorithm uses hash chains to validate forensics evidence by generating the hash chain elements forward
<table>
<tbody>
<tr>
<td></td>
<td>vn,vn−1,⋯,v0subscript𝑣𝑛subscript𝑣𝑛1⋯subscript𝑣0\displaystyle v_{n},v_{n-1},\cdots,v_{0}</td>
<td></td>
</tr>
</tbody>
</table>
online but only storing $`\lceil{\log_{2}n}\rceil`$ hash elements for a hash chain of $`n`$ total hash elements. The preimages of such a hash chain may then be output using Jakobsson’s algorithm and validated by an evidence clerk. To best ensure the evidence’s credence, it is best for the clerk to only compute the backward preimage traversal as far as necessary. Then, additional deferred disclosures may be performed as needed to re-authenticate the validity of the exposed hash elements.
A difficulty is in validating a hash chain by backward preimage traversal in memory times computational complexity of less than $`O\hspace{0pt}{(n)}`$ where $`n`$ is the number of hashes required, see for example \[[14](#bib.bib14), [7](#bib.bib7), [8](#bib.bib8), [25](#bib.bib25), [19](#bib.bib19)\].
This paper gives an online algorithm for generating a forward hash chain traversal while always storing pebbles to be used for backwards preimage traversal by Jakobsson’s algorithm. The online algorithm grows a hash chain as requested to any length $`n`$, but never requires storing more than $`\lceil{\log n}\rceil`$ hash elements or pebbles, where $`n`$ hash elements have been generated so far. This is important on memory constrained devices. For any $`n`$, the $`\lceil{\log n}\rceil`$ pebbles can be directly plugged into Jakobsson’s algorithm to start backward preimage traversal for verification.
If $`n`$ is the number of hash elements stored, other recent methods applicable to emitting hash elements, double the size to $`2\hspace{0pt}n`$ to store a single additional element–the $`({n + 1})`$-st element. This is not acceptable here for several reasons: (1) our approach depends on precisely timed hash element generation and generating $`n`$ more hash elements may cause a simple and constrained device to miss data collection; and (2) the constrained devices may not have the storage to hold $`n`$ additional hash elements.
### 1.2 Previous Work
Secure audit logs applied to digital forensics were developed by Schneier and Kelsey \[[24](#bib.bib24)\]. This work assumes three machines: a trusted secure server $`\mathcal{T}`$, an small machine who secure state is untrusted $`\mathcal{U}`$ for keeping audit logs and a sometimes trusted verifier $`\mathcal{V}`$. The audit log machine $`\mathcal{U}`$ is a small constrained machine that is only occasionally securely connected to $`\mathcal{T}`$. A machine $`\mathcal{U}`$ is trusted until it is compromised. If it is compromised, then it cannot change its audit log or read audit elements before the compromise occurred. This system uses a hash chain to secure the audit logs on the untrusted machines. But, each untrusted machine deletes all but the last element in the hash chain as the audit log grows. Only $`\mathcal{T}`$ has the seed of hash chain for verification. Our online algorithm along with Jakobsson’s can be applied to variations of Schneier and Kelsey’s audit log system. This would save space and potentially allow numerous hash chains to serve as audit logs. They close their audit log with a ‘normalCloseMessage’ to prevent an adversary from extending it illicitly.
To our knowledge, all hash chain preimage traversal work to date assumes either (1) the hash chain is computed in advance, or (2) the length of the hash chain is known or a bound on the length is known in advance, or (3) given $`2^{t}`$ hash elements, the hash chain may be doubled in size to $`2^{t + 1}`$ hash elements when the $`({2^{t} + 1})`$-st element is needed, for $`t`$ a positive integer.
The first case does not seem directly applicable to our digital forensics scheme. In the last two of these cases, there may be a good deal of excess unused memory.
#### 1.2.1 Time-Stamping
Time-stamping is a significant area of research. Haber and Stornetta \[[11](#bib.bib11)\] gave time-stamping schemes using both hash chains and digital signatures. This is similar to the hash chain scheme used in this paper. The first method they give is a linking scheme using hash chains to maintain temporal integrity. All of their schemes depend on a trusted third party. The trusted third party provides a time-stamping service which applies a hash function and digitally signs it. Other related schemes and improvements were given in this paper as well as Bayer, Haber, Stornetta \[[2](#bib.bib2)\] and Haber and Stornetta \[[12](#bib.bib12)\]. Buldas, Laud, Lipmaa and Villemson \[[4](#bib.bib4)\] focus on ‘relative temporal authentication’ and give both time-stamping requirements and new algorithms suited to these requirements. Their methods require a trusted third party to provide time-stamping services. Ansper, Buldas, Saarepera, and Willemson \[[1](#bib.bib1)\] discuss linkage based protocols (i.e., hash chain based protocols) compared to hash-and-sign time-stamping protocols. Building on work of Willemson \[[28](#bib.bib28)\], Lipmaa \[[18](#bib.bib18)\] gives efficient algorithms to traverse skewed trees. Our paper uses basic time-stamping by hash chains. Though the focus is on constrained devices and digital forensics.
#### 1.2.2 Hash Chain Traversal
Our forward hash chain traversal is based on Jakobsson’s backward hash chain preimage traversal work, see also Coppersmith and Jakobsson \[[7](#bib.bib7)\]. Jakbosson gives an asymptotically optimal algorithm to compute consecutive preimages of hash chains. His algorithm requires $`\lceil{\log n}\rceil`$ storage and $`\lceil{\log n}\rceil`$ hash evaluations per hash element output. This is assuming $`O\hspace{0pt}{(n)}`$ preprocessing was used to build the hash chain of $`n`$ elements.
Coppersmith and Jakobsson \[[7](#bib.bib7)\] give an algorithm with amortized time-space product cost of about $`\frac{1}{2}\hspace{0pt}{\log^{2}n}`$ per hash chain element. This is also assuming $`O\hspace{0pt}{(n)}`$ preprocessing was used to build the hash chain of $`n`$ elements. They also give the following lower bound: Computing any element of the hash chain in the worst-case requires a time-space trade-off $`{T\hspace{0pt}S} \geq \frac{\log^{2}n}{4}`$, where $`T`$ is the number of invocations of the hash function and $`S`$ is the number of stored hash elements.
Sella \[[25](#bib.bib25)\] gives a general solution that applies $`k`$ hash function evaluations to generate any element in a hash chain, while storing $`{({k - 1})}\hspace{0pt}n^{1/{({k - 1})}}`$ total hash elements. His algorithm initially stores hash elements that are at constant intervals of distance $`{({k - 1})}\hspace{0pt}n^{1/{({k - 1})}}`$. Kim \[[15](#bib.bib15), [16](#bib.bib16)\] gives algorithms that improve Sella’s in saving up to half of the space Sella’s algorithms use while keeping the same parametric space and hash evaluation costs. This means Kim’s algorithm uses at most the same space as Jakobsson’s algorithm.
Ben-Amram and Petersen \[[3](#bib.bib3)\] give an algorithm for backing up in a singly linked list of length $`n`$ in $`O\hspace{0pt}{(n^{\epsilon})}`$ time, for any $`\epsilon > 0`$. This requires $`O\hspace{0pt}{(n)}`$ pre-processing of the linked list.
Matias and Porat \[[19](#bib.bib19)\] give list and graph traversal data structures that allow efficient backward list traversal and very efficient forward traversal. Their ‘skeleton’ data structures allow complete back traversals in $`O\hspace{0pt}{({k\hspace{0pt}n^{1/k}})}`$ amortized time given $`k`$ elements in storage. Thus, storing $`k = {O\hspace{0pt}{({\log n})}}`$ elements, their algorithm requires $`O\hspace{0pt}{({\log i})}`$ amortized element evaluations for the $`i`$-th backward traversal, where $`i \leq n`$. They bring out several interesting applications beyond hash chains. If their data structures are built for lists of $`n`$ elements, then to accommodate $`n + 1`$ elements they build a new data structure (list synopsis) for $`{2\hspace{0pt}n} + 1`$ total elements, see subsection 4.3 of the full version of \[[19](#bib.bib19)\].
### 1.3 Structure of this Paper
In the remainder of the paper we develop the ideas behind our proposed approach. Section [2](#S2 "2 Chains of Custody: Physical and Digital ‣ An Online Algorithm for Generating Fractal Hash Chains Applied to Digital Chains of Custody11footnote 1 An extended abstract of this paper is to appear in the Intelligence and Security Informatics 2007 (ISI 2007) Conference.") gives our model. Section [3](#S3 "3 Jakobsson’s Algorithm ‣ An Online Algorithm for Generating Fractal Hash Chains Applied to Digital Chains of Custody11footnote 1 An extended abstract of this paper is to appear in the Intelligence and Security Informatics 2007 (ISI 2007) Conference.") reviews Jakobsson’s algorithm in detail. We give proofs of correctness for this algorithm in the appendix. In order to prove the correctness of our algorithm we found it necessary to supply detailed proofs of Jakobsson’s algorithm (which were not given directly in Jakobsson’s paper). Section [4](#S4 "4 Online Output of Jakobsson’s Pebbles for any 𝑛 ‣ An Online Algorithm for Generating Fractal Hash Chains Applied to Digital Chains of Custody11footnote 1 An extended abstract of this paper is to appear in the Intelligence and Security Informatics 2007 (ISI 2007) Conference.") introduces the specifics of our online algorithm, discusses how it interfaces with Jakobsson’s amortized algorithm, and gives a proof of correctness. In Section [5](#S5 "5 Further Directions ‣ An Online Algorithm for Generating Fractal Hash Chains Applied to Digital Chains of Custody11footnote 1 An extended abstract of this paper is to appear in the Intelligence and Security Informatics 2007 (ISI 2007) Conference.") we give some conclusions.
## 2 Chains of Custody: Physical and Digital
Next is a formal definition of a chain of custody.
######
Definition 2
A chain of custody is a detailed account documenting the handling and access to evidence.
We quote Colquitt \[[6](#bib.bib6), Page 484\] on the purpose of a chain of custody:
> “The purpose, then, of establishing a chain of custody is to satisfy the court that it is reasonably probable that the exhibit is authentic and that no one has altered or tampered with the proffered physical exhibit.”
A chain of custody, as described in Definition [2](#Thmdefinition2 "Definition 2 ‣ 2 Chains of Custody: Physical and Digital ‣ An Online Algorithm for Generating Fractal Hash Chains Applied to Digital Chains of Custody11footnote 1 An extended abstract of this paper is to appear in the Intelligence and Security Informatics 2007 (ISI 2007) Conference."), may sometimes be referred to as a classic chain of custody. Maintaining a chain of custody is a standard practice investigators use to inextricably link the evidence that ties a crime to the suspects.
Digital evidence is often stored using a classical chain of custody. For example, documenting when a particular individual first picked up a disk drive with critical evidence, the state of the disk drive, to whom and when they transferred the drive, etc.
Digital evidence is extraordinarily easy to copy. Using standard techniques, each copy of digital evidence is easily authenticated. Physical evidence is somewhat different \[[11](#bib.bib11), [20](#bib.bib20)\]. Consider a crime committed by using a gun. Manufacturing a new $`.357`$ Magnum Ruger Blackhawk Flattop revolver with an identical ‘look,’ serial number, and bullet grooves, to one used in a crime is extraordinary work. Just finding experts to replicate such physical evidence would generally leave a substantial paper trail. Moreover, biological evidence generally also provides time frames. Thus, classic chains of custody often focus on the basic identification, authentication, uniqueness, and time. Digital data lacks such uniqueness and timing characteristics. That is, many copies may be made of digital evidence both for legitimate and illicit reasons. Furthermore, timestamps alone may not be sufficiently convincing.
######
Definition 3
A digital chain of custody is the information preserved about the data and its changes that shows specific data was in a particular state at a given date and time.
A DCoC (Digital Chain of Custody) is a small constrained device for holding, authenticating, and verifying a digital chain of custody.
Take a one-way hash function $`h`$. Suppose $`h`$ has deterministic and tightly bounded time complexity. There are some hash functions with bounded time complexity that are already in use: the RSA SecurID \[[23](#bib.bib23)\] as well as systems like TESLA \[[21](#bib.bib21), [22](#bib.bib22)\] depend on timed hash functions. This paper goes one step further advocating extremely precisely timed hash functions.
This paper is focused on several aspects of evidence maintenance that are related to time stamping, see for example \[[8](#bib.bib8)\]. One side of the challenge is to show a piece of evidence existed after a particular point in time. Using the Merkle-Damgård technique, the evidence may be combined with non-predictable, widely distributed, and time sensitive documents such as a newspaper or financial market data
<sup>\*</sup>
<sup>\*</sup>\*As financial markets become more automated and distributed, it may be feasible for the trading volume to be granulated down to minute fractions of a second. For example, in 2004 the average NYSE trading day volume per second was about $`1,500,000,{000/{({60 \cdot 60 \cdot 6.5})}}`$ or more than 64,000 shares per second. See www.nyse.com. The volume is known down to the individual share. This highly granular data may be widely distributed.
.
Similar discussions to the next definition can be found in several places, for example \[[11](#bib.bib11), [27](#bib.bib27)\].
######
Definition 4
Data that is non-predictable, widely distributed, verifiably stored, and time sensitive is socially bound data.
A critical issue is that socially bound and highly granular data are not common. In certain situations, socially bound data must be highly granular, for example in milliseconds. Thus, carefully timed hash chains may be used as a proxy for socially bound data. See also, timestamp linkages in \[[11](#bib.bib11)\].
#### 2.0.1 The Adversary
The adversary this paper assumes is either (1) an untrustworthy verifier; or (2) a general attack against known hash functions on a DCoC.
Suppose our system computes an $`n`$ element hash chain $`v_{n},\cdots,v_{0}`$, and say $`n > t`$. An untrustworthy verifier may get an element of a hash chain $`v_{t}`$ along with the files of digital evidence $`E = {\mathcal{E}_{t},\cdots,\mathcal{E}_{0}}`$. The adversary may illicitly modify the evidence to $`E' = {\mathcal{E}_{t}',\cdots,\mathcal{E}_{0}'}`$ and compute a ‘competing’ hash chain using $`v_{t}`$ and $`E'`$. Thus, adding doubt to the validity of the real evidence. Provided $`t < n`$, and the one-way hash function $`h`$ is not broken, we can validate the hash chain based on $`E`$ and not $`E'`$ by producing $`v_{t + 1}`$ and showing that $`{h\hspace{0pt}{({v_{t + 1} \oplus {c\hspace{0pt}{(\mathcal{E}_{t + 1})}}})}} = v_{t}`$.
The issue of a general breach of a hash function is primarily dealt with by following the forensics policy of having at least three different known hash functions for the data. This procedure is to ensure trust in the hash functions in case one of them is no longer trustworthy \[[5](#bib.bib5)\].
### 2.1 Overview of Our Approach
This paper assumes a model consisting of at least three constrained devices (DCoCs or cards) each having a different hash function. These constrained devices may be interfaced using a USB 2.0, for instance. In any case, these devices have a small processor as well as limited memory. They are tamper resistant.
#### 2.1.1 Data Capture
The DCoCs may each interface through separate USB 2.0 ports simultaneously to gather common snapshots of critical system states and memory. Each DCoC may be physically secured and transported independently by a different member of law enforcement. We assume these devices will be mass produced so that each DCoC has unique authentication software. (They should also have unique physical identification.) Different members of law enforcement should handle each card making it less likely that all cards may be manipulated by a single member of law enforcement. Furthermore, from physical identification, the DCoCs may become part of the classical chains of custody.
Finally, each of the DCoCs may periodically authenticate each other while capturing data. Records of these authentications can be included in the hashed values. This authentication should be zero-knowledge interactive proofs of identity such as the Feige-Fiat-Shamir protocol \[[9](#bib.bib9)\]. Such a proof of identity is important in this application domain since these proofs of identity are not transitive. Thus, a fake DCoC cannot impersonate a real one by just mimicking its proof of identity.
#### 2.1.2 Forensic Data Verification
The plaintext evidence and diversifying data has been stored as the files $`M`$ in plain sight on storage such as an optical disk. Different states or snap-shots of the evidence will be periodically concatenated into timed inputs of the hash chain. Suppose the hash computations run at a fixed and known speed on tamper resistant hardware. Thus, computing a hash chain on this special hardware can be used to certify the initial data states at specified times, see also Haber and Stornetta \[[11](#bib.bib11)\].
The verification procedure is done by an evidence clerk and consists of the following steps:
1. 1\.
Connect all three DCoCs to a secured verification machine. Initially, each DCoC is authenticated to all other DCoCs using a zero-knowledge interactive proof of identity.
2. 2\.
The DCoCs only output hash elements backwards on an as-needed basis to verify the evidence under consideration.
3. 3\.
As hash chain elements are output, the verification must continue with each hash element step. If a DCoC cannot authenticate another device, then it will alert the evidence clerk or even shut down.
4. 4\.
The hash elements from a DCoC are sent out from left to right:
<table>
<tbody>
<tr>
<td></td>
<td>hn(vn⊕c(ℰn)),hn−1(vn−1⊕c(ℰn−1)),⋯,ht(vt⊕c(ℰt)),superscriptℎ𝑛direct-sumsubscript𝑣𝑛𝑐subscriptℰ𝑛superscriptℎ𝑛1direct-sumsubscript𝑣𝑛1𝑐subscriptℰ𝑛1⋯superscriptℎ𝑡direct-sumsubscript𝑣𝑡𝑐subscriptℰ𝑡\displaystyle h^{n}(v_{n}\oplus c({\cal E}_{n})),\ h^{n-1}(v_{n-1}\oplus c({\cal E}_{n-1})),\ \cdots,\ h^{t}(v_{t}\oplus c({\cal E}_{t})),</td>
<td></td>
</tr>
</tbody>
</table>
for $`{n - 1} \geq t`$. Moreover, $`h^{n}\hspace{0pt}{({v_{n} \oplus {c\hspace{0pt}{(\mathcal{E}_{n})}}})}`$ is verified by computing
<table>
<tbody>
<tr>
<td></td>
<td>h(hn−1(vn−1⊕c(ℰn−1))⊕c(ℰn)),ℎdirect-sumsuperscriptℎ𝑛1direct-sumsubscript𝑣𝑛1𝑐subscriptℰ𝑛1𝑐subscriptℰ𝑛\displaystyle h(h^{n-1}(v_{n-1}\oplus c({\cal E}_{n-1}))\oplus c({\cal E}_{n})),</td>
<td></td>
</tr>
</tbody>
</table>
and so on.
#### 2.1.3 Putting It All Together
Given the tamper resistant hardware, just storing data with associated time stamps may quickly overflow a constrained device–especially if the time granularity is very fine. Thus, compressed representation of hash chains can be used to verify time stamps along with the authenticity of other constrained devices.
It is possible to publicly post data from a system under forensic investigation by way of a proxy server. This data may be signed by a small device and posted to a public (or private and trusted) location. Small devices may have trouble signing large amounts of data due to their constraints. Furthermore, for very fine time granularity, fast and consistent network bandwidth may not be available. In some contexts, if data is captured before a formal investigation is initiated, then it may be important to keep the suspects unaware that data is being captured as potential evidence.
A DCoC may run its hash chain until a forensic examiner or law enforcement officer carefully checks it into the physical chain of custody while noting the time in detail. This allows a forensic examiner to determine the time of the evidence by counting the exact number of hashes. A key motivator of our online algorithm is in proactive forensics the value of $`n`$ is not known in advance. Where $`n`$ is the number of time-steps recorded by the DCoCs. In the case of a digital chain of custody, it can not be known precisely how long it will take to get potential evidence to an evidence clerk. Or how long an evidence clerk will take to validate the data, etc. This precise time as well as the interval between hashes, computed by our algorithm, will have to be known in advance to fake a hash chain.
The preprocessing phase of Jakobsson’s amortized algorithm must know the value of $`n`$ in advance of when it is run \[[14](#bib.bib14)\]. Our new algorithm prepares the appropriate $`\lceil{\log n}\rceil`$ pebbles for Jakobsson’s algorithm, for any $`n`$. During hash element generation, for any $`n`$ the online algorithm never stores more than $`\lceil{\log n}\rceil`$ pebbles. Once the online algorithm has no more requests to generate new hash elements, then the amortized algorithm \[[14](#bib.bib14)\] may be immediately started.
## 3 Jakobsson’s Algorithm
This section reviews Jakobsson’s Algorithm \[[14](#bib.bib14)\]. The proofs of correctness are in the Appendix. Some of these proofs are used in the results of Section [4](#S4 "4 Online Output of Jakobsson’s Pebbles for any 𝑛 ‣ An Online Algorithm for Generating Fractal Hash Chains Applied to Digital Chains of Custody11footnote 1 An extended abstract of this paper is to appear in the Intelligence and Security Informatics 2007 (ISI 2007) Conference.").
### 3.1 A Review of Jakobsson’s Algorithm
Given a hash chain with 16 elements, Jakobsson’s algorithm is initialized as indicated in Table [1](#S3.T1 "Table 1 ‣ 3.1 A Review of Jakobsson’s Algorithm ‣ 3 Jakobsson’s Algorithm ‣ An Online Algorithm for Generating Fractal Hash Chains Applied to Digital Chains of Custody11footnote 1 An extended abstract of this paper is to appear in the Intelligence and Security Informatics 2007 (ISI 2007) Conference."). In this Table, hash element $`16`$ is the seed of the hash chain. Where the preimages are exposed sequentially in the following order: $`1,2,\cdots,16`$. In general, the position of a hash element in a hash chain ranges from $`1,\cdots,n`$ where hash element $`1`$ is the first element to be exposed in a pre-image traversal. Thus, hash element $`n`$ is the seed of the hash chain.
<table>
<thead>
<tr>
<th>Element Position</th>
<th>1</th>
<th>2</th>
<th>3</th>
<th>4</th>
<th>5</th>
<th>6</th>
<th>7</th>
<th>8</th>
<th>9</th>
<th>10</th>
<th>11</th>
<th>12</th>
<th>13</th>
<th>14</th>
<th>15</th>
<th>16</th>
</tr>
</thead>
<tbody>
<tr>
<td>Pebble Placement</td>
<td></td>
<td>×\times</td>
<td></td>
<td>×\times</td>
<td></td>
<td></td>
<td></td>
<td>×\times</td>
<td></td>
<td></td>
<td></td>
<td></td>
<td></td>
<td></td>
<td></td>
<td>×\times</td>
</tr>
</tbody>
</table>
Table 1: Initial pebble locations for Jakobsson’s algorithm
Suppose $`n = {|H|}`$ is the number of hash values in the entire hash chain. In this case, Jakobsson’s amortized algorithm only stores $`\lceil{\log n}\rceil`$ hash pebbles.
Suppose $`n = 2^{k}`$ for some integer $`k`$. Now, Jakobsson’s algorithm \[[14](#bib.bib14)\] is given:
#### 3.1.1 Jakobsson’s Setup.
Compute the entire hash chain: $`v_{0},v_{1},\cdots,v_{n}`$.
For pebble $`p\hspace{0pt}{\lbrack j\rbrack}`$, where $`j:{k = {\log_{2}n} \geq j \geq 1}`$:
<table>
<tbody>
<tr>
<td></td>
<td>p[j].𝐒𝐭𝐚𝐫𝐭_𝐈𝐧𝐜𝐫𝐞𝐦𝐞𝐧𝐭formulae-sequence𝑝delimited-[]𝑗𝐒𝐭𝐚𝐫𝐭_𝐈𝐧𝐜𝐫𝐞𝐦𝐞𝐧𝐭\displaystyle p[j].{\bf Start\_Increment}</td>
<td>←←\displaystyle\leftarrow</td>
<td>3×2j3superscript2𝑗\displaystyle 3\times 2^{j}</td>
<td></td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td>p[j].𝐃𝐞𝐬𝐭_𝐈𝐧𝐜𝐫𝐞𝐦𝐞𝐧𝐭formulae-sequence𝑝delimited-[]𝑗𝐃𝐞𝐬𝐭_𝐈𝐧𝐜𝐫𝐞𝐦𝐞𝐧𝐭\displaystyle p[j].{\bf Dest\_Increment}</td>
<td>←←\displaystyle\leftarrow</td>
<td>2j+1superscript2𝑗1\displaystyle 2^{j+1}</td>
<td></td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td>p[j].𝐏𝐨𝐬𝐢𝐭𝐢𝐨𝐧formulae-sequence𝑝delimited-[]𝑗𝐏𝐨𝐬𝐢𝐭𝐢𝐨𝐧\displaystyle p[j].{\bf Position}</td>
<td>←←\displaystyle\leftarrow</td>
<td>2jsuperscript2𝑗\displaystyle 2^{j}</td>
<td></td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td>p[j].𝐃𝐞𝐬𝐭𝐢𝐧𝐚𝐭𝐢𝐨𝐧formulae-sequence𝑝delimited-[]𝑗𝐃𝐞𝐬𝐭𝐢𝐧𝐚𝐭𝐢𝐨𝐧\displaystyle p[j].{\bf Destination}</td>
<td>←←\displaystyle\leftarrow</td>
<td>2jsuperscript2𝑗\displaystyle 2^{j}</td>
<td></td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td>p[j].𝐕𝐚𝐥𝐮𝐞formulae-sequence𝑝delimited-[]𝑗𝐕𝐚𝐥𝐮𝐞\displaystyle p[j].{\bf Value}</td>
<td>←←\displaystyle\leftarrow</td>
<td>v2jsubscript𝑣superscript2𝑗\displaystyle v_{2^{j}}</td>
<td></td>
</tr>
</tbody>
</table>
Furthermore,
<table>
<tbody>
<tr>
<td></td>
<td>𝐂𝐮𝐫𝐫𝐞𝐧𝐭.𝐏𝐨𝐬𝐢𝐭𝐢𝐨𝐧formulae-sequence𝐂𝐮𝐫𝐫𝐞𝐧𝐭𝐏𝐨𝐬𝐢𝐭𝐢𝐨𝐧\displaystyle{\bf Current.Position}</td>
<td>←←\displaystyle\leftarrow</td>
<td>00\displaystyle 0</td>
<td></td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td>𝐂𝐮𝐫𝐫𝐞𝐧𝐭.𝐕𝐚𝐥𝐮𝐞formulae-sequence𝐂𝐮𝐫𝐫𝐞𝐧𝐭𝐕𝐚𝐥𝐮𝐞\displaystyle{\bf Current.Value}</td>
<td>←←\displaystyle\leftarrow</td>
<td>v0subscript𝑣0\displaystyle v_{0}</td>
<td></td>
</tr>
</tbody>
</table>
#### 3.1.2 Jakobsson’s Main.
This algorithm in Figure [2](#S3.F2 "Figure 2 ‣ 3.1.2 Jakobsson’s Main. ‣ 3.1 A Review of Jakobsson’s Algorithm ‣ 3 Jakobsson’s Algorithm ‣ An Online Algorithm for Generating Fractal Hash Chains Applied to Digital Chains of Custody11footnote 1 An extended abstract of this paper is to appear in the Intelligence and Security Informatics 2007 (ISI 2007) Conference.") updates the pebbles and values.
<table>
<tbody>
<tr>
<td>Given an entire precomputed computed hash chain:</td>
<td></td>
<td></td>
<td></td>
</tr>
<tr>
<td></td>
<td>v1,v2,⋯,vnsubscript𝑣1subscript𝑣2⋯subscript𝑣𝑛v_{1},v_{2},\cdots,v_{n}, then compute the pebble positions & auxiliary information.</td>
<td></td>
<td></td>
</tr>
<tr>
<td>1.</td>
<td>if Current.Position=nCurrent.Position𝑛\mbox{\bf Current.Position}=n, then</td>
<td></td>
<td></td>
</tr>
<tr>
<td></td>
<td></td>
<td>Stop</td>
<td></td>
</tr>
<tr>
<td></td>
<td>else</td>
<td></td>
<td></td>
</tr>
<tr>
<td></td>
<td></td>
<td>Current.Position←1+Current.Position←Current.Position1Current.Position\mbox{\bf Current.Position}\leftarrow 1+\mbox{\bf Current.Position}</td>
<td></td>
</tr>
<tr>
<td></td>
<td>endif</td>
<td></td>
<td></td>
</tr>
<tr>
<td>2.</td>
<td>for j←1←𝑗1j\leftarrow 1 to k𝑘k do</td>
<td></td>
<td></td>
</tr>
<tr>
<td></td>
<td></td>
<td>if p[j].𝐏𝐨𝐬𝐢𝐭𝐢𝐨𝐧≠p[j].𝐃𝐞𝐬𝐭𝐢𝐧𝐚𝐭𝐢𝐨𝐧formulae-sequence𝑝delimited-[]𝑗𝐏𝐨𝐬𝐢𝐭𝐢𝐨𝐧𝑝delimited-[]𝑗𝐃𝐞𝐬𝐭𝐢𝐧𝐚𝐭𝐢𝐨𝐧p[j].{\bf Position}\neq p[j].{\bf Destination} then</td>
<td></td>
</tr>
<tr>
<td></td>
<td>2.1</td>
<td></td>
<td>p[j].𝐏𝐨𝐬𝐢𝐭𝐢𝐨𝐧←p[j].𝐏𝐨𝐬𝐢𝐭𝐢𝐨𝐧−2formulae-sequence𝑝delimited-[]𝑗←𝐏𝐨𝐬𝐢𝐭𝐢𝐨𝐧𝑝delimited-[]𝑗𝐏𝐨𝐬𝐢𝐭𝐢𝐨𝐧2p[j].{\bf Position}\leftarrow p[j].{\bf Position}-2</td>
</tr>
<tr>
<td></td>
<td>2.2</td>
<td></td>
<td>p[j].𝐕𝐚𝐥𝐮𝐞←h(h(p[j].𝐕𝐚𝐥𝐮𝐞))p[j].{\bf Value}\leftarrow h(h(p[j].{\bf Value}))</td>
</tr>
<tr>
<td></td>
<td></td>
<td>endif</td>
<td></td>
</tr>
<tr>
<td></td>
<td>endfor</td>
<td></td>
<td></td>
</tr>
<tr>
<td>3.</td>
<td>if Current.Position is odd then</td>
<td></td>
<td></td>
</tr>
<tr>
<td></td>
<td></td>
<td>output h(p[1].𝐕𝐚𝐥𝐮𝐞)h(p[1].{\bf Value})</td>
<td></td>
</tr>
<tr>
<td></td>
<td>else</td>
<td></td>
<td></td>
</tr>
<tr>
<td></td>
<td></td>
<td>output p[1].𝐕𝐚𝐥𝐮𝐞formulae-sequence𝑝delimited-[]1𝐕𝐚𝐥𝐮𝐞p[1].{\bf Value}</td>
<td></td>
</tr>
<tr>
<td></td>
<td>3.1</td>
<td>p[1].𝐏𝐨𝐬𝐢𝐭𝐢𝐨𝐧←p[1].𝐏𝐨𝐬𝐢𝐭𝐢𝐨𝐧+p[1].𝐒𝐭𝐚𝐫𝐭_𝐈𝐧𝐜𝐫𝐞𝐦𝐞𝐧𝐭p[1].{\bf Position}\ \ \ \ \ \leftarrow p[1].{\bf Position}+p[1].{\bf Start\_Increment}</td>
<td></td>
</tr>
<tr>
<td></td>
<td>3.2</td>
<td>p[1].𝐃𝐞𝐬𝐭𝐢𝐧𝐚𝐭𝐢𝐨𝐧←p[1].𝐃𝐞𝐬𝐭𝐢𝐧𝐚𝐭𝐢𝐨𝐧+p[1].𝐃𝐞𝐬𝐭_𝐈𝐧𝐜𝐫𝐞𝐦𝐞𝐧𝐭formulae-sequence𝑝delimited-[]1←𝐃𝐞𝐬𝐭𝐢𝐧𝐚𝐭𝐢𝐨𝐧𝑝delimited-[]1𝐃𝐞𝐬𝐭𝐢𝐧𝐚𝐭𝐢𝐨𝐧𝑝delimited-[]1𝐃𝐞𝐬𝐭_𝐈𝐧𝐜𝐫𝐞𝐦𝐞𝐧𝐭p[1].{\bf Destination}\leftarrow p[1].{\bf Destination}+p[1].{\bf Dest\_Increment}</td>
<td></td>
</tr>
<tr>
<td></td>
<td></td>
<td>if p[1].𝐃𝐞𝐬𝐭𝐢𝐧𝐚𝐭𝐢𝐨𝐧>nformulae-sequence𝑝delimited-[]1𝐃𝐞𝐬𝐭𝐢𝐧𝐚𝐭𝐢𝐨𝐧𝑛p[1].{\bf Destination}>n then</td>
<td></td>
</tr>
<tr>
<td></td>
<td></td>
<td></td>
<td>p[1].𝐏𝐨𝐬𝐢𝐭𝐢𝐨𝐧←+∞p[1].{\bf Position}\ \ \ \ \ \leftarrow+\infty</td>
</tr>
<tr>
<td></td>
<td></td>
<td></td>
<td>p[1].𝐃𝐞𝐬𝐭𝐢𝐧𝐚𝐭𝐢𝐨𝐧←+∞formulae-sequence𝑝delimited-[]1←𝐃𝐞𝐬𝐭𝐢𝐧𝐚𝐭𝐢𝐨𝐧p[1].{\bf Destination}\leftarrow+\infty</td>
</tr>
<tr>
<td></td>
<td></td>
<td>else</td>
<td></td>
</tr>
<tr>
<td></td>
<td></td>
<td></td>
<td>p[1].𝐕𝐚𝐥𝐮𝐞←𝐅𝐢𝐧𝐝𝐕𝐚𝐥𝐮𝐞formulae-sequence𝑝delimited-[]1←𝐕𝐚𝐥𝐮𝐞𝐅𝐢𝐧𝐝𝐕𝐚𝐥𝐮𝐞p[1].{\bf Value}\leftarrow{\bf FindValue}</td>
</tr>
<tr>
<td></td>
<td></td>
<td>endif</td>
<td></td>
</tr>
<tr>
<td></td>
<td></td>
<td>Sort Pebbles by Position</td>
<td></td>
</tr>
<tr>
<td></td>
<td>endif</td>
<td></td>
<td></td>
</tr>
</tbody>
</table>
Figure 2: Jakobsson’s Hash-Chain Pebble Position update
[Refer to caption](/html/0705.2876/assets/x2.png)
Figure 3: Jakobsson’s Algorithm: Each row from top to bottom represents the state of a hash chain traversed by Jakobsson’s algorithm. Each row represents a single hash element exposure. For each row, bold and underlined positions represent the location of pebbles where Position = Destination. Bold positions represent a temporary location for a pebble in transit only. When a pebble is reached and moved it is moved to the back of a grey area and moves to the front of the grey area or sequence of grey areas until it reaches its destination (and is underlined and in the next row).
Using Jakobsson’s amortized technique \[[14](#bib.bib14)\], a hash chain of size $`n`$ requires a total of $`k`$ pebbles where $`k = {\lceil{\log n}\rceil}`$. This amortized algorithm performs $`O\hspace{0pt}{({\log n})}`$ hash applications per hash element output. Initially, in the setup phase, the pebbles store hash elements from hash-chain positions $`2^{1},2^{2},2^{3},\ldots,2^{k}`$, respectively.
## 4 Online Output of Jakobsson’s Pebbles for any $`n`$
Jakobsson’s amortized algorithm works to conserve both stored pebbles and hash evaluations. This allows his algorithm to verify hash values on small sensors. This algorithm assumes pre-processing where all hash elements are pre-computed, perhaps by a more powerful processor \[[14](#bib.bib14)\].
Our aim with the online algorithm is to have a small constrained device that generates all requested hash elements. The online algorithm broadens the applicability of Jakobsson’s amortized algorithm. In particular, the online algorithm generates all hash elements, but only stores $`\lceil{\log n}\rceil`$ pebbles at any one time. Where $`n`$ is the total number of hash element generated so far. Every time a new hash element is generated, no additional hash evaluations must be performed.
These $`\lceil{\log n}\rceil`$ hash pebbles are positioned so that at any point the online algorithm is no longer invoked, then Jakobsson’s amortized algorithm can start to be run directly on the pebbles. Thus, we believe the amortized and online algorithms are complimentary and fit well together.
In Jakobsson’s notation each hash element keeps its index throughout the computation,
<table>
<tbody>
<tr>
<td></td>
<td>v0,⋯,vn.subscript𝑣0⋯subscript𝑣𝑛\displaystyle v_{0},\cdots,v_{n}.</td>
<td></td>
</tr>
</tbody>
</table>
This is because $`n`$ is fixed, so all hash element numbers remain the same. The pebbles in Jakobsson’s algorithm are re-numbered, according to the initial fixed hash element numbers, and sorted by their positions.
The notation for the online algorithm uses hash element index notation that changes the index values over time.
For instance, in Jakobsson’s preimage traversal algorithm, suppose the initial $`n`$ hash elements are:
<table>
<tbody>
<tr>
<td></td>
<td>u1,⋯,un.subscript𝑢1⋯subscript𝑢𝑛\displaystyle u_{1},\ \cdots,\ u_{n}.</td>
<td></td>
</tr>
</tbody>
</table>
Then, after the first element is verified, it is discarded and the following hash elements remain:
<table>
<tbody>
<tr>
<td></td>
<td>u2,⋯,un.subscript𝑢2⋯subscript𝑢𝑛\displaystyle u_{2},\ \cdots,\ u_{n}.</td>
<td></td>
</tr>
</tbody>
</table>
Alternatively, since $`n`$ is not fixed in our online algorithm, suppose the following list of elements has already been generated (with only the appropriate pebbles saved),
<table>
<tbody>
<tr>
<td></td>
<td>w1,⋯,wn.subscript𝑤1⋯subscript𝑤𝑛\displaystyle w_{1},\cdots,w_{n}.</td>
<td></td>
</tr>
</tbody>
</table>
Then, generating another hash value $`w_{0}\leftarrow{h\hspace{0pt}{(w_{1})}}`$. This gives $`n + 1`$ total hash elements, thus for simplicity all hash indices are increased by 1. This gives the following renumbering,
<table>
<tbody>
<tr>
<td></td>
<td>w1,⋯,wn+1.subscript𝑤1⋯subscript𝑤𝑛1\displaystyle w_{1},\cdots,w_{n+1}.</td>
<td></td>
</tr>
</tbody>
</table>
[Refer to caption](/html/0705.2876/assets/x3.png)
Figure 4: Online Algorithm: As before, each row represents a hash element exposure. For each row bold and underlined positions represent the location of pebbles.
Take $`v_{i}`$, then it corresponds to $`w_{j}`$, where
<table>
<tbody>
<tr>
<td></td>
<td>j𝑗\displaystyle j</td>
<td>=\displaystyle=</td>
<td>i−(2σ−𝐭𝐨𝐭𝐚𝐥𝐇𝐚𝐬𝐡𝐄𝐥𝐞𝐦𝐞𝐧𝐭𝐬),𝑖superscript2𝜎𝐭𝐨𝐭𝐚𝐥𝐇𝐚𝐬𝐡𝐄𝐥𝐞𝐦𝐞𝐧𝐭𝐬\displaystyle i-(2^{\sigma}-{\bf totalHashElements}),</td>
<td></td>
<td>(2)</td>
</tr>
</tbody>
</table>
where $`\sigma`$ is the total number of pebbles and totalHashElements is the number of hash elements constructed thus far, also $`{\mathbf{t}\mathbf{o}\mathbf{t}\mathbf{a}\mathbf{l}\mathbf{H}\mathbf{a}\mathbf{s}\mathbf{h}\mathbf{E}\mathbf{l}\mathbf{e}\mathbf{m}\mathbf{e}\mathbf{n}\mathbf{t}\mathbf{s}} \leq 2^{\sigma}`$. Note, that when $`{\mathbf{t}\mathbf{o}\mathbf{t}\mathbf{a}\mathbf{l}\mathbf{H}\mathbf{a}\mathbf{s}\mathbf{h}\mathbf{E}\mathbf{l}\mathbf{e}\mathbf{m}\mathbf{e}\mathbf{n}\mathbf{t}\mathbf{s}} = 2^{\sigma}`$, then we verify $`i = j`$ by Observation [1](#Sx1.EGx39 "Observation 1 ‣ Appendix: A Proof of Correctness for Jakobsson’s Algorithm ‣ An Online Algorithm for Generating Fractal Hash Chains Applied to Digital Chains of Custody11footnote 1 An extended abstract of this paper is to appear in the Intelligence and Security Informatics 2007 (ISI 2007) Conference.") in the appendix.
######
Fact 1
Jakobsson’s notation and our notation corresponds exactly when $`n = 2^{k}`$ for some integer $`k \geq 0`$.
The online algorithm has two distinct, non-interweavable phases: The growth phase, and the exposure phase.
### 4.1 The Growth Phase
Next is the growth phase.
#### 4.1.1 The Online Algorithm’s Setup
<table>
<tbody>
<tr>
<td></td>
<td>𝐬𝐞𝐞𝐝𝐬𝐞𝐞𝐝\displaystyle{\bf seed}</td>
<td>←←\displaystyle\leftarrow</td>
<td>the initial hash value</td>
<td></td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td>𝐭𝐨𝐭𝐚𝐥𝐇𝐚𝐬𝐡𝐄𝐥𝐞𝐦𝐞𝐧𝐭𝐬𝐭𝐨𝐭𝐚𝐥𝐇𝐚𝐬𝐡𝐄𝐥𝐞𝐦𝐞𝐧𝐭𝐬\displaystyle{\bf totalHashElements}</td>
<td>←←\displaystyle\leftarrow</td>
<td>11\displaystyle 1</td>
<td></td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td>𝐠𝐫𝐨𝐰.𝐯𝐚𝐥𝐮𝐞formulae-sequence𝐠𝐫𝐨𝐰𝐯𝐚𝐥𝐮𝐞\displaystyle{\bf grow.value}</td>
<td>←←\displaystyle\leftarrow</td>
<td>𝐬𝐞𝐞𝐝𝐬𝐞𝐞𝐝\displaystyle{\bf seed}</td>
<td></td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td>𝐠𝐫𝐨𝐰.𝐩𝐞𝐛𝐛𝐥𝐞formulae-sequence𝐠𝐫𝐨𝐰𝐩𝐞𝐛𝐛𝐥𝐞\displaystyle{\bf grow.pebble}</td>
<td>←←\displaystyle\leftarrow</td>
<td>11\displaystyle 1</td>
<td></td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td>𝐞𝐱𝐩𝐨𝐧𝐞𝐧𝐭𝐞𝐱𝐩𝐨𝐧𝐞𝐧𝐭\displaystyle{\bf exponent}</td>
<td>←←\displaystyle\leftarrow</td>
<td>11\displaystyle 1</td>
<td></td>
</tr>
</tbody>
</table>
#### 4.1.2 The Online Algorithm’s Init Pebble
<table>
<tbody>
<tr>
<td>InitializePebble(p,j,grow.value𝑝𝑗grow.valuep,j,\mbox{\bf grow.value})</td>
<td></td>
</tr>
<tr>
<td></td>
<td>p[j].𝐌𝐨𝐯𝐞_𝐈𝐧𝐜𝐫𝐞𝐦𝐞𝐧𝐭←2j+1formulae-sequence𝑝delimited-[]𝑗←𝐌𝐨𝐯𝐞_𝐈𝐧𝐜𝐫𝐞𝐦𝐞𝐧𝐭superscript2𝑗1p[j].{\bf Move\_Increment}\leftarrow 2^{j+1}</td>
</tr>
<tr>
<td></td>
<td>p[j].𝐒𝐭𝐚𝐫𝐭_𝐈𝐧𝐜𝐫𝐞𝐦𝐞𝐧𝐭←3×2jformulae-sequence𝑝delimited-[]𝑗←𝐒𝐭𝐚𝐫𝐭_𝐈𝐧𝐜𝐫𝐞𝐦𝐞𝐧𝐭3superscript2𝑗p[j].{\bf Start\_Increment}\leftarrow 3\times 2^{j}</td>
</tr>
<tr>
<td></td>
<td>p[j].𝐃𝐞𝐬𝐭_𝐈𝐧𝐜𝐫𝐞𝐦𝐞𝐧𝐭←2j+1formulae-sequence𝑝delimited-[]𝑗←𝐃𝐞𝐬𝐭_𝐈𝐧𝐜𝐫𝐞𝐦𝐞𝐧𝐭superscript2𝑗1p[j].{\bf Dest\_Increment}\leftarrow 2^{j+1}</td>
</tr>
<tr>
<td></td>
<td>p[j].𝐏𝐨𝐬𝐢𝐭𝐢𝐨𝐧←2jformulae-sequence𝑝delimited-[]𝑗←𝐏𝐨𝐬𝐢𝐭𝐢𝐨𝐧superscript2𝑗p[j].{\bf Position}\leftarrow 2^{j}</td>
</tr>
<tr>
<td></td>
<td>p[j].𝐃𝐞𝐬𝐭𝐢𝐧𝐚𝐭𝐢𝐨𝐧←2jformulae-sequence𝑝delimited-[]𝑗←𝐃𝐞𝐬𝐭𝐢𝐧𝐚𝐭𝐢𝐨𝐧superscript2𝑗p[j].{\bf Destination}\leftarrow 2^{j}</td>
</tr>
<tr>
<td></td>
<td>p[j].𝐕𝐚𝐥𝐮𝐞←𝐠𝐫𝐨𝐰.𝐯𝐚𝐥𝐮𝐞formulae-sequence𝑝delimited-[]𝑗←𝐕𝐚𝐥𝐮𝐞𝐠𝐫𝐨𝐰𝐯𝐚𝐥𝐮𝐞p[j].{\bf Value}\leftarrow{\bf grow.value}</td>
</tr>
</tbody>
</table>
Figure 5: Initialize Pebble
#### 4.1.3 The Online Algorithm’s Main
Figure [6](#S4.F6 "Figure 6 ‣ 4.1.3 The Online Algorithm’s Main ‣ 4.1 The Growth Phase ‣ 4 Online Output of Jakobsson’s Pebbles for any 𝑛 ‣ An Online Algorithm for Generating Fractal Hash Chains Applied to Digital Chains of Custody11footnote 1 An extended abstract of this paper is to appear in the Intelligence and Security Informatics 2007 (ISI 2007) Conference.") updates the pebbles and values.
<table>
<tbody>
<tr>
<td></td>
<td>while not done growing hash chain do</td>
<td></td>
<td></td>
<td></td>
<td></td>
</tr>
<tr>
<td></td>
<td></td>
<td>𝐠𝐫𝐨𝐰.𝐯𝐚𝐥𝐮𝐞←h(𝐠𝐫𝐨𝐰.𝐯𝐚𝐥𝐮𝐞){\bf grow.value}\leftarrow h({\bf grow.value})</td>
<td></td>
<td></td>
<td></td>
</tr>
<tr>
<td></td>
<td></td>
<td>𝐭𝐨𝐭𝐚𝐥𝐇𝐚𝐬𝐡𝐄𝐥𝐞𝐦𝐞𝐧𝐭𝐬←𝐭𝐨𝐭𝐚𝐥𝐇𝐚𝐬𝐡𝐄𝐥𝐞𝐦𝐞𝐧𝐭𝐬+1←𝐭𝐨𝐭𝐚𝐥𝐇𝐚𝐬𝐡𝐄𝐥𝐞𝐦𝐞𝐧𝐭𝐬𝐭𝐨𝐭𝐚𝐥𝐇𝐚𝐬𝐡𝐄𝐥𝐞𝐦𝐞𝐧𝐭𝐬1{\bf totalHashElements}\leftarrow{\bf totalHashElements}+1</td>
<td></td>
<td></td>
<td></td>
</tr>
<tr>
<td>1.</td>
<td></td>
<td>if 𝐭𝐨𝐭𝐚𝐥𝐇𝐚𝐬𝐡𝐄𝐥𝐞𝐦𝐞𝐧𝐭𝐬−1=2k for some integer k≥1𝐭𝐨𝐭𝐚𝐥𝐇𝐚𝐬𝐡𝐄𝐥𝐞𝐦𝐞𝐧𝐭𝐬1superscript2𝑘 for some integer k≥1{\bf totalHashElements}-1=2^{k}\mbox{ for some integer $k\geq 1$}, then</td>
<td></td>
<td></td>
<td></td>
</tr>
<tr>
<td></td>
<td></td>
<td></td>
<td>exponent←exponent+1←exponentexponent1\mbox{\bf exponent}\leftarrow\mbox{\bf exponent}+1</td>
<td></td>
<td></td>
</tr>
<tr>
<td></td>
<td></td>
<td></td>
<td>for all pebbles l𝑙l do</td>
<td></td>
<td></td>
</tr>
<tr>
<td></td>
<td></td>
<td></td>
<td></td>
<td>p[l].𝐏𝐨𝐬𝐢𝐭𝐢𝐨𝐧←p[l].𝐏𝐨𝐬𝐢𝐭𝐢𝐨𝐧+2exponent−1formulae-sequence𝑝delimited-[]𝑙←𝐏𝐨𝐬𝐢𝐭𝐢𝐨𝐧𝑝delimited-[]𝑙𝐏𝐨𝐬𝐢𝐭𝐢𝐨𝐧superscript2exponent1p[l].{\bf Position}\leftarrow p[l].{\bf Position}+2^{\mbox{\bf exponent}-1}</td>
<td></td>
</tr>
<tr>
<td></td>
<td></td>
<td></td>
<td>endfor</td>
<td></td>
<td></td>
</tr>
<tr>
<td></td>
<td></td>
<td></td>
<td>j←𝐠𝐫𝐨𝐰.𝐩𝐞𝐛𝐛𝐥𝐞formulae-sequence←𝑗𝐠𝐫𝐨𝐰𝐩𝐞𝐛𝐛𝐥𝐞j\leftarrow{\bf grow.pebble}</td>
<td></td>
<td></td>
</tr>
<tr>
<td></td>
<td></td>
<td></td>
<td>Create pebble p[j]𝑝delimited-[]𝑗p[j]</td>
<td></td>
<td></td>
</tr>
<tr>
<td></td>
<td></td>
<td></td>
<td>InitializePebble(p,j,grow.value𝑝𝑗grow.valuep,j,\mbox{\bf grow.value})</td>
<td></td>
<td></td>
</tr>
<tr>
<td></td>
<td></td>
<td></td>
<td>p[j].𝐝𝐢𝐬𝐭𝐚𝐧𝐜𝐞𝐟𝐫𝐨𝐦𝐒𝐞𝐞𝐝←𝐭𝐨𝐭𝐚𝐥𝐇𝐚𝐬𝐡𝐄𝐥𝐞𝐦𝐞𝐧𝐭𝐬formulae-sequence𝑝delimited-[]𝑗←𝐝𝐢𝐬𝐭𝐚𝐧𝐜𝐞𝐟𝐫𝐨𝐦𝐒𝐞𝐞𝐝𝐭𝐨𝐭𝐚𝐥𝐇𝐚𝐬𝐡𝐄𝐥𝐞𝐦𝐞𝐧𝐭𝐬p[j].{\bf distancefromSeed}\leftarrow{\bf totalHashElements}</td>
<td></td>
<td></td>
</tr>
<tr>
<td></td>
<td></td>
<td></td>
<td>𝐠𝐫𝐨𝐰.𝐩𝐞𝐛𝐛𝐥𝐞←𝐠𝐫𝐨𝐰.𝐩𝐞𝐛𝐛𝐥𝐞+1formulae-sequence𝐠𝐫𝐨𝐰←𝐩𝐞𝐛𝐛𝐥𝐞𝐠𝐫𝐨𝐰𝐩𝐞𝐛𝐛𝐥𝐞1{\bf grow.pebble}\leftarrow{\bf grow.pebble}+1</td>
<td></td>
<td></td>
</tr>
<tr>
<td>2.</td>
<td></td>
<td>else</td>
<td></td>
<td></td>
<td></td>
</tr>
<tr>
<td></td>
<td></td>
<td></td>
<td>for all pebbles l𝑙l do</td>
<td></td>
<td></td>
</tr>
<tr>
<td></td>
<td></td>
<td></td>
<td></td>
<td>t←p[l].𝐌𝐨𝐯𝐞_𝐈𝐧𝐜𝐫𝐞𝐦𝐞𝐧𝐭+p[l].𝐝𝐢𝐬𝐭𝐚𝐧𝐜𝐞𝐟𝐫𝐨𝐦𝐒𝐞𝐞𝐝+1formulae-sequence←𝑡𝑝delimited-[]𝑙𝐌𝐨𝐯𝐞_𝐈𝐧𝐜𝐫𝐞𝐦𝐞𝐧𝐭𝑝delimited-[]𝑙𝐝𝐢𝐬𝐭𝐚𝐧𝐜𝐞𝐟𝐫𝐨𝐦𝐒𝐞𝐞𝐝1t\leftarrow p[l].{\bf Move\_Increment}+p[l].{\bf distancefromSeed}+1</td>
<td></td>
</tr>
<tr>
<td>3.</td>
<td></td>
<td></td>
<td></td>
<td>if 𝐭𝐨𝐭𝐚𝐥𝐇𝐚𝐬𝐡𝐄𝐥𝐞𝐦𝐞𝐧𝐭𝐬=t𝐭𝐨𝐭𝐚𝐥𝐇𝐚𝐬𝐡𝐄𝐥𝐞𝐦𝐞𝐧𝐭𝐬𝑡{\bf totalHashElements}=t then</td>
<td></td>
</tr>
<tr>
<td></td>
<td>3.1</td>
<td></td>
<td></td>
<td></td>
<td>p[l].𝐕𝐚𝐥𝐮𝐞←𝐠𝐫𝐨𝐰.𝐯𝐚𝐥𝐮𝐞formulae-sequence𝑝delimited-[]𝑙←𝐕𝐚𝐥𝐮𝐞𝐠𝐫𝐨𝐰𝐯𝐚𝐥𝐮𝐞p[l].{\bf Value}\leftarrow{\bf grow.value}</td>
</tr>
<tr>
<td></td>
<td>3.2</td>
<td></td>
<td></td>
<td></td>
<td>p[l].𝐝𝐢𝐬𝐭𝐚𝐧𝐜𝐞𝐟𝐫𝐨𝐦𝐒𝐞𝐞𝐝←𝐭𝐨𝐭𝐚𝐥𝐇𝐚𝐬𝐡𝐄𝐥𝐞𝐦𝐞𝐧𝐭𝐬formulae-sequence𝑝delimited-[]𝑙←𝐝𝐢𝐬𝐭𝐚𝐧𝐜𝐞𝐟𝐫𝐨𝐦𝐒𝐞𝐞𝐝𝐭𝐨𝐭𝐚𝐥𝐇𝐚𝐬𝐡𝐄𝐥𝐞𝐦𝐞𝐧𝐭𝐬p[l].{\bf distancefromSeed}\leftarrow{\bf totalHashElements}</td>
</tr>
<tr>
<td></td>
<td></td>
<td></td>
<td></td>
<td></td>
<td>p[l].𝐏𝐨𝐬𝐢𝐭𝐢𝐨𝐧←p[l].𝐏𝐨𝐬𝐢𝐭𝐢𝐨𝐧−p[l].𝐌𝐨𝐯𝐞_𝐈𝐧𝐜𝐫𝐞𝐦𝐞𝐧𝐭formulae-sequence𝑝delimited-[]𝑙←𝐏𝐨𝐬𝐢𝐭𝐢𝐨𝐧𝑝delimited-[]𝑙𝐏𝐨𝐬𝐢𝐭𝐢𝐨𝐧𝑝delimited-[]𝑙𝐌𝐨𝐯𝐞_𝐈𝐧𝐜𝐫𝐞𝐦𝐞𝐧𝐭p[l].{\bf Position}\leftarrow p[l].{\bf Position}-p[l].{\bf Move\_Increment}</td>
</tr>
<tr>
<td></td>
<td></td>
<td></td>
<td></td>
<td>endif</td>
<td></td>
</tr>
<tr>
<td></td>
<td></td>
<td></td>
<td>endfor</td>
<td></td>
<td></td>
</tr>
<tr>
<td></td>
<td></td>
<td>endif</td>
<td></td>
<td></td>
<td></td>
</tr>
<tr>
<td></td>
<td>endwhile</td>
<td></td>
<td></td>
<td></td>
<td></td>
</tr>
<tr>
<td></td>
<td>j←𝐠𝐫𝐨𝐰.𝐩𝐞𝐛𝐛𝐥𝐞formulae-sequence←𝑗𝐠𝐫𝐨𝐰𝐩𝐞𝐛𝐛𝐥𝐞j\leftarrow{\bf grow.pebble}</td>
<td></td>
<td></td>
<td></td>
<td></td>
</tr>
<tr>
<td>4.</td>
<td>Create pebble p[j]𝑝delimited-[]𝑗p[j] where j←𝐠𝐫𝐨𝐰.𝐩𝐞𝐛𝐛𝐥𝐞formulae-sequence←𝑗𝐠𝐫𝐨𝐰𝐩𝐞𝐛𝐛𝐥𝐞j\leftarrow{\bf grow.pebble}</td>
<td></td>
<td></td>
<td></td>
<td></td>
</tr>
<tr>
<td></td>
<td>InitializePebble(p,j,seed𝑝𝑗seedp,j,\mbox{\bf seed})</td>
<td></td>
<td></td>
<td></td>
<td></td>
</tr>
<tr>
<td></td>
<td>p[j].𝐝𝐢𝐬𝐭𝐚𝐧𝐜𝐞𝐟𝐫𝐨𝐦𝐒𝐞𝐞𝐝←0formulae-sequence𝑝delimited-[]𝑗←𝐝𝐢𝐬𝐭𝐚𝐧𝐜𝐞𝐟𝐫𝐨𝐦𝐒𝐞𝐞𝐝0p[j].{\bf distancefromSeed}\leftarrow 0</td>
<td></td>
<td></td>
<td></td>
<td></td>
</tr>
<tr>
<td></td>
<td>Sort pebbles by Position</td>
<td></td>
<td></td>
<td></td>
<td></td>
</tr>
</tbody>
</table>
Figure 6: Hash Chain Growth Algorithm
### 4.2 The Online Algorithm’s Exposure Phase
In this phase, only the setup is different from Jakobsson’s original algorithm, because we are no longer assuming that exactly $`2^{j}`$ pebbles must be present.
#### 4.2.1 The Online Algorithm’s Setup
For $`j\leftarrow 1`$ to $`\lceil{\log n}\rceil`$ do
<table>
<tbody>
<tr>
<td></td>
<td>p[j].𝐝𝐞𝐬𝐭𝐢𝐧𝐚𝐭𝐢𝐨𝐧formulae-sequence𝑝delimited-[]𝑗𝐝𝐞𝐬𝐭𝐢𝐧𝐚𝐭𝐢𝐨𝐧\displaystyle p[j].{\bf destination}</td>
<td>←←\displaystyle\leftarrow</td>
<td>p[j].𝐩𝐨𝐬𝐢𝐭𝐢𝐨𝐧formulae-sequence𝑝delimited-[]𝑗𝐩𝐨𝐬𝐢𝐭𝐢𝐨𝐧\displaystyle p[j].{\bf position}</td>
<td></td>
</tr>
</tbody>
</table>
Furthermore, we must also initialize the next lines
<table>
<tbody>
<tr>
<td></td>
<td>𝐜𝐮𝐫𝐫𝐞𝐧𝐭.𝐩𝐨𝐬𝐢𝐭𝐢𝐨𝐧formulae-sequence𝐜𝐮𝐫𝐫𝐞𝐧𝐭𝐩𝐨𝐬𝐢𝐭𝐢𝐨𝐧\displaystyle{\bf current.position}</td>
<td>←←\displaystyle\leftarrow</td>
<td>𝐭𝐨𝐭𝐚𝐥𝐇𝐚𝐬𝐡𝐄𝐥𝐞𝐦𝐞𝐧𝐭𝐬𝐭𝐨𝐭𝐚𝐥𝐇𝐚𝐬𝐡𝐄𝐥𝐞𝐦𝐞𝐧𝐭𝐬\displaystyle{\bf totalHashElements}</td>
<td></td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td>𝐜𝐮𝐫𝐫𝐞𝐧𝐭.𝐯𝐚𝐥𝐮𝐞formulae-sequence𝐜𝐮𝐫𝐫𝐞𝐧𝐭𝐯𝐚𝐥𝐮𝐞\displaystyle{\bf current.value}</td>
<td>←←\displaystyle\leftarrow</td>
<td>𝐠𝐫𝐨𝐰.𝐯𝐚𝐥𝐮𝐞formulae-sequence𝐠𝐫𝐨𝐰𝐯𝐚𝐥𝐮𝐞\displaystyle{\bf grow.value}</td>
<td></td>
</tr>
</tbody>
</table>
The online algorithm’s Main function is performed exactly as Jakobsson’s algorithm with the exception that current.position cannot simply be set to zero during the setup phase, but rather must be determined from the value of the first pebble in the sorted order.
### 4.3 Characteristics of the Hash Chain Growth Algorithm
To prove the validity of the Hash Chain Growth Algorithm we will show: (1) at each step of the hash chain growth there are always enough pebbles, and (2) the pebbles are always properly placed such that Jackobsson’s algorithm can immediately begin to run on the stored data from the generated hash chain.
Recall, given $`n = 2^{t}`$, for some integer $`t`$, then pebbles are placed in hash element positions
<table>
<tbody>
<tr>
<td></td>
<td>t=logn,t−1=log(n/2),t−2=log(n/4),⋯,t=1.formulae-sequence𝑡𝑛formulae-sequence𝑡1𝑛2formulae-sequence𝑡2𝑛4⋯𝑡1\displaystyle t=\log n,\ \ t-1=\log(n/2),\ \ t-2=\log(n/4),\ \ \cdots,\ \ t=1.</td>
<td></td>
</tr>
</tbody>
</table>
For any initial pebble $`p\hspace{0pt}{\lbrack j\rbrack}`$ the value $`{p\hspace{0pt}{\lbrack j\rbrack}}.{\mathbf{D}\mathbf{e}\mathbf{s}\mathbf{t}\mathbf{i}\mathbf{n}\mathbf{a}\mathbf{t}\mathbf{i}\mathbf{o}\mathbf{n}}`$ never decreases in Jakobsson’s algorithm. Thus, to determine on which move back the initial pebble $`p\hspace{0pt}{\lbrack j\rbrack}`$ will be output, compute when
<table>
<tbody>
<tr>
<td></td>
<td>p[j].𝐃𝐞𝐬𝐭𝐢𝐧𝐚𝐭𝐢𝐨𝐧>nformulae-sequence𝑝delimited-[]𝑗𝐃𝐞𝐬𝐭𝐢𝐧𝐚𝐭𝐢𝐨𝐧𝑛\displaystyle p[j].{\bf Destination}>n</td>
<td></td>
</tr>
</tbody>
</table>
first occurs.
Assume $`n`$ is a power of $`2`$, and the definition of $`D\hspace{0pt}{(j,k)}`$ can be found in the appendix, and given a fixed $`j`$ solve for the minimal $`k`$ so that
<table>
<tbody>
<tr>
<td></td>
<td>2j+k⋅2j+1superscript2𝑗⋅𝑘superscript2𝑗1\displaystyle 2^{j}+k\cdot 2^{j+1}</td>
<td>>\displaystyle></td>
<td>2lognsuperscript2𝑛\displaystyle 2^{\log n}</td>
<td></td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td>2k2𝑘\displaystyle 2k</td>
<td>>\displaystyle></td>
<td>2logn−j−1.superscript2𝑛𝑗1\displaystyle 2^{\log n-j}-1.</td>
<td></td>
</tr>
</tbody>
</table>
For instance, for $`j = {\log{({n/2})}}`$, then $`k = 1`$ gives
<table>
<tbody>
<tr>
<td></td>
<td>2k=22𝑘2\displaystyle 2k=2</td>
<td>>\displaystyle></td>
<td>2logn−(logn−1)−1superscript2𝑛𝑛11\displaystyle 2^{\log n-(\log n-1)}-1</td>
<td></td>
</tr>
<tr>
<td></td>
<td></td>
<td>>\displaystyle></td>
<td>2−1= 1,\displaystyle 2-1\ \ =\ \ 1,</td>
<td></td>
</tr>
</tbody>
</table>
which holds.
######
Fact 2
Let $`n = 2^{k}`$. In Jakobsson’s algorithm, pebbles are eliminated in the reverse order of their original indices $`j = {{k - 1},\cdots,1}`$ where we consider the seed pebble at $`j = k`$ to never be discarded.
Proof: This proof follows from Jakobsson’s algorithm in Figure [2](#S3.F2 "Figure 2 ‣ 3.1.2 Jakobsson’s Main. ‣ 3.1 A Review of Jakobsson’s Algorithm ‣ 3 Jakobsson’s Algorithm ‣ An Online Algorithm for Generating Fractal Hash Chains Applied to Digital Chains of Custody11footnote 1 An extended abstract of this paper is to appear in the Intelligence and Security Informatics 2007 (ISI 2007) Conference."). By Fact [17](#Thmfact17 "Fact 17 ‣ Appendix: A Proof of Correctness for Jakobsson’s Algorithm ‣ An Online Algorithm for Generating Fractal Hash Chains Applied to Digital Chains of Custody11footnote 1 An extended abstract of this paper is to appear in the Intelligence and Security Informatics 2007 (ISI 2007) Conference.") of the appendix, it must be that only the pebble in initial position $`j = {k - 1} = {\log{({n/2})}}`$ is going to be eliminated immediately after Jakobsson’s algorithm generates and emits $`n/2`$ hash elements.
Now, suppose the hash elements are renumbered where $`1`$ represents hash element $`{n/2} + 1`$; $`2`$ represents hash element $`{n/2} + 2`$, all the way to $`n/2`$ represents hash element $`n`$. Still $`n = 2^{k}`$.
By Fact [19](#Sx1.EGx74 "Fact 19 ‣ Appendix: A Proof of Correctness for Jakobsson’s Algorithm ‣ An Online Algorithm for Generating Fractal Hash Chains Applied to Digital Chains of Custody11footnote 1 An extended abstract of this paper is to appear in the Intelligence and Security Informatics 2007 (ISI 2007) Conference.") of the appendix, with this re-numbering, the remaining pebbles will be placed at pebble positions
<table>
<tbody>
<tr>
<td></td>
<td>21,22,⋯, 2k−1,superscript21superscript22⋯superscript2𝑘1\displaystyle 2^{1},2^{2},\ \cdots,\ 2^{k-1},</td>
<td></td>
</tr>
</tbody>
</table>
where $`{n/2} = 2^{k - 1}`$. The pebble in position $`p\hspace{0pt}{\lbrack{k - 1}\rbrack}`$ is the seed of the hash chain, so it never moves. By Corollary [1](#Sx1.EGx69 "Corollary 1 ‣ Appendix: A Proof of Correctness for Jakobsson’s Algorithm ‣ An Online Algorithm for Generating Fractal Hash Chains Applied to Digital Chains of Custody11footnote 1 An extended abstract of this paper is to appear in the Intelligence and Security Informatics 2007 (ISI 2007) Conference.") also of the appendix, for each of these remaining pebbles,
<table>
<tbody>
<tr>
<td></td>
<td>p[j].𝐏𝐨𝐬𝐢𝐭𝐢𝐨𝐧formulae-sequence𝑝delimited-[]𝑗𝐏𝐨𝐬𝐢𝐭𝐢𝐨𝐧\displaystyle p[j].{\bf Position}</td>
<td>=\displaystyle=</td>
<td>p[j].𝐃𝐞𝐬𝐭𝐢𝐧𝐚𝐭𝐢𝐨𝐧.formulae-sequence𝑝delimited-[]𝑗𝐃𝐞𝐬𝐭𝐢𝐧𝐚𝐭𝐢𝐨𝐧\displaystyle p[j].{\bf Destination}.</td>
<td></td>
</tr>
</tbody>
</table>
However, for any initial pebble $`p\hspace{0pt}{\lbrack j\rbrack}`$ the value $`{p\hspace{0pt}{\lbrack j\rbrack}}.{\mathbf{D}\mathbf{e}\mathbf{s}\mathbf{t}\mathbf{i}\mathbf{n}\mathbf{a}\mathbf{t}\mathbf{i}\mathbf{o}\mathbf{n}}`$ never decreases in the algorithm of Figure [2](#S3.F2 "Figure 2 ‣ 3.1.2 Jakobsson’s Main. ‣ 3.1 A Review of Jakobsson’s Algorithm ‣ 3 Jakobsson’s Algorithm ‣ An Online Algorithm for Generating Fractal Hash Chains Applied to Digital Chains of Custody11footnote 1 An extended abstract of this paper is to appear in the Intelligence and Security Informatics 2007 (ISI 2007) Conference."). Thus, we need to determine what initial pebble $`p\hspace{0pt}{\lbrack j\rbrack}`$ is placed in position $`2^{k - 2} = {n/4}`$. This new position is one backward move beyond the $`n/2`$nd emitted hash element. Thus, the $`{\mathbf{D}\mathbf{e}\mathbf{s}\mathbf{t}}\hspace{0pt}\_\hspace{0pt}{\mathbf{I}\mathbf{n}\mathbf{c}\mathbf{r}\mathbf{e}\mathbf{m}\mathbf{e}\mathbf{n}\mathbf{t}}`$ of this pebble must be $`2^{k - 2} = {n/4}`$, so the initial pebble was $`p\hspace{0pt}{\lbrack{k - 2}\rbrack}`$. Furthermore, by Fact [17](#Thmfact17 "Fact 17 ‣ Appendix: A Proof of Correctness for Jakobsson’s Algorithm ‣ An Online Algorithm for Generating Fractal Hash Chains Applied to Digital Chains of Custody11footnote 1 An extended abstract of this paper is to appear in the Intelligence and Security Informatics 2007 (ISI 2007) Conference."), this pebble is the second pebble to be discarded.
The proof is completed by induction.
The next facts follow from the online algorithm in Figure [6](#S4.F6 "Figure 6 ‣ 4.1.3 The Online Algorithm’s Main ‣ 4.1 The Growth Phase ‣ 4 Online Output of Jakobsson’s Pebbles for any 𝑛 ‣ An Online Algorithm for Generating Fractal Hash Chains Applied to Digital Chains of Custody11footnote 1 An extended abstract of this paper is to appear in the Intelligence and Security Informatics 2007 (ISI 2007) Conference.").
######
Fact 3
Consider the online hash growth algorithm (see Figure [6](#S4.F6 "Figure 6 ‣ 4.1.3 The Online Algorithm’s Main ‣ 4.1 The Growth Phase ‣ 4 Online Output of Jakobsson’s Pebbles for any 𝑛 ‣ An Online Algorithm for Generating Fractal Hash Chains Applied to Digital Chains of Custody11footnote 1 An extended abstract of this paper is to appear in the Intelligence and Security Informatics 2007 (ISI 2007) Conference.")). A new pebble is only created when totalHashElements is such that
<table>
<tbody>
<tr>
<td></td>
<td>2k+1=𝐭𝐨𝐭𝐚𝐥𝐇𝐚𝐬𝐡𝐄𝐥𝐞𝐦𝐞𝐧𝐭𝐬,superscript2𝑘1𝐭𝐨𝐭𝐚𝐥𝐇𝐚𝐬𝐡𝐄𝐥𝐞𝐦𝐞𝐧𝐭𝐬\displaystyle 2^{k}+1={\bf totalHashElements},</td>
<td></td>
</tr>
</tbody>
</table>
for some $`k \geq 1`$.
Proof: This is a direct result of line 1 of the online hash growth algorithm.
Any time $`2^{k} = {{\mathbf{t}\mathbf{o}\mathbf{t}\mathbf{a}\mathbf{l}\mathbf{H}\mathbf{a}\mathbf{s}\mathbf{h}\mathbf{E}\mathbf{l}\mathbf{e}\mathbf{m}\mathbf{e}\mathbf{n}\mathbf{t}\mathbf{s}} - 1}`$, then between now and when $`2^{k - 1} = {{\mathbf{t}\mathbf{o}\mathbf{t}\mathbf{a}\mathbf{l}\mathbf{H}\mathbf{a}\mathbf{s}\mathbf{h}\mathbf{E}\mathbf{l}\mathbf{e}\mathbf{m}\mathbf{e}\mathbf{n}\mathbf{t}\mathbf{s}} - 1}`$, all of the pebble’s positions must be shifted by $`2^{k - 1}`$. This is done in Figure [6](#S4.F6 "Figure 6 ‣ 4.1.3 The Online Algorithm’s Main ‣ 4.1 The Growth Phase ‣ 4 Online Output of Jakobsson’s Pebbles for any 𝑛 ‣ An Online Algorithm for Generating Fractal Hash Chains Applied to Digital Chains of Custody11footnote 1 An extended abstract of this paper is to appear in the Intelligence and Security Informatics 2007 (ISI 2007) Conference.") before a new pebble is created.
######
Fact 4
Consider the online hash growth algorithm (see Figure [6](#S4.F6 "Figure 6 ‣ 4.1.3 The Online Algorithm’s Main ‣ 4.1 The Growth Phase ‣ 4 Online Output of Jakobsson’s Pebbles for any 𝑛 ‣ An Online Algorithm for Generating Fractal Hash Chains Applied to Digital Chains of Custody11footnote 1 An extended abstract of this paper is to appear in the Intelligence and Security Informatics 2007 (ISI 2007) Conference.")). Each time an element is added to the chain, totalHashElements is incremented by one.
######
Fact 5
Consider the online hash growth algorithm (see Figure [6](#S4.F6 "Figure 6 ‣ 4.1.3 The Online Algorithm’s Main ‣ 4.1 The Growth Phase ‣ 4 Online Output of Jakobsson’s Pebbles for any 𝑛 ‣ An Online Algorithm for Generating Fractal Hash Chains Applied to Digital Chains of Custody11footnote 1 An extended abstract of this paper is to appear in the Intelligence and Security Informatics 2007 (ISI 2007) Conference.")). Each time a pebble is created it is initialized as,
<table>
<tbody>
<tr>
<td></td>
<td>p[j].𝐝𝐢𝐬𝐭𝐚𝐧𝐜𝐞𝐟𝐫𝐨𝐦𝐒𝐞𝐞𝐝formulae-sequence𝑝delimited-[]𝑗𝐝𝐢𝐬𝐭𝐚𝐧𝐜𝐞𝐟𝐫𝐨𝐦𝐒𝐞𝐞𝐝\displaystyle p[j].{\bf distancefromSeed}</td>
<td>←←\displaystyle\leftarrow</td>
<td>𝐭𝐨𝐭𝐚𝐥𝐇𝐚𝐬𝐡𝐄𝐥𝐞𝐦𝐞𝐧𝐭𝐬.𝐭𝐨𝐭𝐚𝐥𝐇𝐚𝐬𝐡𝐄𝐥𝐞𝐦𝐞𝐧𝐭𝐬\displaystyle{\bf totalHashElements}.</td>
<td></td>
</tr>
</tbody>
</table>
This follows directly from section 1 of Figure [6](#S4.F6 "Figure 6 ‣ 4.1.3 The Online Algorithm’s Main ‣ 4.1 The Growth Phase ‣ 4 Online Output of Jakobsson’s Pebbles for any 𝑛 ‣ An Online Algorithm for Generating Fractal Hash Chains Applied to Digital Chains of Custody11footnote 1 An extended abstract of this paper is to appear in the Intelligence and Security Informatics 2007 (ISI 2007) Conference.").
######
Fact 6
Consider the online hash growth algorithm (see Figure [6](#S4.F6 "Figure 6 ‣ 4.1.3 The Online Algorithm’s Main ‣ 4.1 The Growth Phase ‣ 4 Online Output of Jakobsson’s Pebbles for any 𝑛 ‣ An Online Algorithm for Generating Fractal Hash Chains Applied to Digital Chains of Custody11footnote 1 An extended abstract of this paper is to appear in the Intelligence and Security Informatics 2007 (ISI 2007) Conference.")). After its initialization, the only time $`{p\hspace{0pt}{\lbrack j\rbrack}}.{\mathbf{d}\mathbf{i}\mathbf{s}\mathbf{t}\mathbf{a}\mathbf{n}\mathbf{c}\mathbf{e}\mathbf{f}\mathbf{r}\mathbf{o}\mathbf{m}\mathbf{S}\mathbf{e}\mathbf{e}\mathbf{d}}`$ changes is when pebble $`j`$ is moved and $`{p\hspace{0pt}{\lbrack j\rbrack}}.{\mathbf{d}\mathbf{i}\mathbf{s}\mathbf{t}\mathbf{a}\mathbf{n}\mathbf{c}\mathbf{e}\mathbf{f}\mathbf{r}\mathbf{o}\mathbf{m}\mathbf{S}\mathbf{e}\mathbf{e}\mathbf{d}}`$ is always updated to maintain the correct distance from pebble $`p\hspace{0pt}{\lbrack j\rbrack}`$ to the hash chain seed.
Proof: After $`{p\hspace{0pt}{\lbrack j\rbrack}}.{\mathbf{d}\mathbf{i}\mathbf{s}\mathbf{t}\mathbf{a}\mathbf{n}\mathbf{c}\mathbf{e}\mathbf{f}\mathbf{r}\mathbf{o}\mathbf{m}\mathbf{S}\mathbf{e}\mathbf{e}\mathbf{d}}`$’s initialization in section 1 of Figure [6](#S4.F6 "Figure 6 ‣ 4.1.3 The Online Algorithm’s Main ‣ 4.1 The Growth Phase ‣ 4 Online Output of Jakobsson’s Pebbles for any 𝑛 ‣ An Online Algorithm for Generating Fractal Hash Chains Applied to Digital Chains of Custody11footnote 1 An extended abstract of this paper is to appear in the Intelligence and Security Informatics 2007 (ISI 2007) Conference."), the value $`{p\hspace{0pt}{\lbrack j\rbrack}}.{\mathbf{d}\mathbf{i}\mathbf{s}\mathbf{t}\mathbf{a}\mathbf{n}\mathbf{c}\mathbf{e}\mathbf{f}\mathbf{r}\mathbf{o}\mathbf{m}\mathbf{S}\mathbf{e}\mathbf{e}\mathbf{d}}`$ only changes when the next assignment is made in line 3.2:
<table>
<tbody>
<tr>
<td></td>
<td>p[j].𝐝𝐢𝐬𝐭𝐚𝐧𝐜𝐞𝐟𝐫𝐨𝐦𝐒𝐞𝐞𝐝formulae-sequence𝑝delimited-[]𝑗𝐝𝐢𝐬𝐭𝐚𝐧𝐜𝐞𝐟𝐫𝐨𝐦𝐒𝐞𝐞𝐝\displaystyle p[j].{\bf distancefromSeed}</td>
<td>←←\displaystyle\leftarrow</td>
<td>𝐭𝐨𝐭𝐚𝐥𝐇𝐚𝐬𝐡𝐄𝐥𝐞𝐦𝐞𝐧𝐭𝐬𝐭𝐨𝐭𝐚𝐥𝐇𝐚𝐬𝐡𝐄𝐥𝐞𝐦𝐞𝐧𝐭𝐬\displaystyle{\bf totalHashElements}</td>
<td></td>
</tr>
</tbody>
</table>
and this occurs exactly when the condition in line 3 holds true since
<table>
<tbody>
<tr>
<td></td>
<td>p[j].𝐝𝐢𝐬𝐭𝐚𝐧𝐜𝐞𝐟𝐫𝐨𝐦𝐒𝐞𝐞𝐝+p[j].𝐌𝐨𝐯𝐞_𝐈𝐧𝐜𝐫𝐞𝐦𝐞𝐧𝐭formulae-sequence𝑝delimited-[]𝑗𝐝𝐢𝐬𝐭𝐚𝐧𝐜𝐞𝐟𝐫𝐨𝐦𝐒𝐞𝐞𝐝𝑝delimited-[]𝑗𝐌𝐨𝐯𝐞_𝐈𝐧𝐜𝐫𝐞𝐦𝐞𝐧𝐭\displaystyle p[j].{\bf distancefromSeed}+p[j].{\bf Move\_Increment}</td>
<td>=\displaystyle=</td>
<td>𝐭𝐨𝐭𝐚𝐥𝐇𝐚𝐬𝐡𝐄𝐥𝐞𝐦𝐞𝐧𝐭𝐬,𝐭𝐨𝐭𝐚𝐥𝐇𝐚𝐬𝐡𝐄𝐥𝐞𝐦𝐞𝐧𝐭𝐬\displaystyle{\bf totalHashElements},</td>
<td></td>
</tr>
</tbody>
</table>
completing the proof.
######
Fact 7
In order to reverse Jakobsson’s algorithm pebbles should be assigned indices in ascending order.
Proof: This holds by Fact [2](#Thmfact2 "Fact 2 ‣ 4.3 Characteristics of the Hash Chain Growth Algorithm ‣ 4 Online Output of Jakobsson’s Pebbles for any 𝑛 ‣ An Online Algorithm for Generating Fractal Hash Chains Applied to Digital Chains of Custody11footnote 1 An extended abstract of this paper is to appear in the Intelligence and Security Informatics 2007 (ISI 2007) Conference.").
######
Fact 8
For each pebble, to find the current position for the purposes of the following proof we can calculate,
<table>
<tbody>
<tr>
<td></td>
<td>p[j].𝐏𝐨𝐬𝐢𝐭𝐢𝐨𝐧=𝐭𝐨𝐭𝐚𝐥𝐇𝐚𝐬𝐡𝐄𝐥𝐞𝐦𝐞𝐧𝐭𝐬−p[j].𝐝𝐢𝐬𝐭𝐚𝐧𝐜𝐞𝐟𝐫𝐨𝐦𝐒𝐞𝐞𝐝.formulae-sequence𝑝delimited-[]𝑗𝐏𝐨𝐬𝐢𝐭𝐢𝐨𝐧𝐭𝐨𝐭𝐚𝐥𝐇𝐚𝐬𝐡𝐄𝐥𝐞𝐦𝐞𝐧𝐭𝐬𝑝delimited-[]𝑗𝐝𝐢𝐬𝐭𝐚𝐧𝐜𝐞𝐟𝐫𝐨𝐦𝐒𝐞𝐞𝐝\displaystyle p[j].{\bf Position}={\bf totalHashElements}-p[j].{\bf distancefromSeed}.</td>
<td></td>
</tr>
</tbody>
</table>
Proof: This holds because Position is a measure of how far from the front of the chain a pebble currently is. By Fact [4](#Thmfact4 "Fact 4 ‣ 4.3 Characteristics of the Hash Chain Growth Algorithm ‣ 4 Online Output of Jakobsson’s Pebbles for any 𝑛 ‣ An Online Algorithm for Generating Fractal Hash Chains Applied to Digital Chains of Custody11footnote 1 An extended abstract of this paper is to appear in the Intelligence and Security Informatics 2007 (ISI 2007) Conference.") totalHashElements is always equal to the total number of elements in the chain and by Facts [5](#Sx1.EGx27 "Fact 5 ‣ 4.3 Characteristics of the Hash Chain Growth Algorithm ‣ 4 Online Output of Jakobsson’s Pebbles for any 𝑛 ‣ An Online Algorithm for Generating Fractal Hash Chains Applied to Digital Chains of Custody11footnote 1 An extended abstract of this paper is to appear in the Intelligence and Security Informatics 2007 (ISI 2007) Conference.") and [8](#Sx1.EGx30 "Fact 8 ‣ 4.3 Characteristics of the Hash Chain Growth Algorithm ‣ 4 Online Output of Jakobsson’s Pebbles for any 𝑛 ‣ An Online Algorithm for Generating Fractal Hash Chains Applied to Digital Chains of Custody11footnote 1 An extended abstract of this paper is to appear in the Intelligence and Security Informatics 2007 (ISI 2007) Conference.") distancefromSeed is always the distance the pebble is from the seed including the seed itself. Thus, we know that we can find how far from the front of the hash chain ($`{p\hspace{0pt}{\lbrack j\rbrack}}.{\mathbf{P}\mathbf{o}\mathbf{s}\mathbf{i}\mathbf{t}\mathbf{i}\mathbf{o}\mathbf{n}}`$) a pebble is by $`{{\mathbf{t}\mathbf{o}\mathbf{t}\mathbf{a}\mathbf{l}\mathbf{H}\mathbf{a}\mathbf{s}\mathbf{h}\mathbf{E}\mathbf{l}\mathbf{e}\mathbf{m}\mathbf{e}\mathbf{n}\mathbf{t}\mathbf{s}} - {p\hspace{0pt}{\lbrack j\rbrack}}}.{\mathbf{d}\mathbf{i}\mathbf{s}\mathbf{t}\mathbf{a}\mathbf{n}\mathbf{c}\mathbf{e}\mathbf{f}\mathbf{r}\mathbf{o}\mathbf{m}\mathbf{S}\mathbf{e}\mathbf{e}\mathbf{d}}`$.
######
Fact 9
Consider the online hash growth algorithm. The only time a pebble is moved is when $`{\mathbf{t}\mathbf{o}\mathbf{t}\mathbf{a}\mathbf{l}\mathbf{H}\mathbf{a}\mathbf{s}\mathbf{h}\mathbf{E}\mathbf{l}\mathbf{e}\mathbf{m}\mathbf{e}\mathbf{n}\mathbf{t}\mathbf{s}} = t`$ where $`{t\leftarrow{p\hspace{0pt}{\lbrack j\rbrack}}}.{{{\mathbf{M}\mathbf{o}\mathbf{v}\mathbf{e}}\hspace{0pt}\_\hspace{0pt}{\mathbf{I}\mathbf{n}\mathbf{c}\mathbf{r}\mathbf{e}\mathbf{m}\mathbf{e}\mathbf{n}\mathbf{t}}} + {p\hspace{0pt}{\lbrack j\rbrack}}}.{\mathbf{d}\mathbf{i}\mathbf{s}\mathbf{t}\mathbf{a}\mathbf{n}\mathbf{c}\mathbf{e}\mathbf{f}\mathbf{r}\mathbf{o}\mathbf{m}\mathbf{S}\mathbf{e}\mathbf{e}\mathbf{d}}`$. When a pebble is moved, then it is moved to the front of the chain.
######
Fact 10
When the online hash growth algorithm is halted the final pebble is placed at the root with the seed value, and the next available pebble index.
Fact [10](#Thmfact10 "Fact 10 ‣ 4.3 Characteristics of the Hash Chain Growth Algorithm ‣ 4 Online Output of Jakobsson’s Pebbles for any 𝑛 ‣ An Online Algorithm for Generating Fractal Hash Chains Applied to Digital Chains of Custody11footnote 1 An extended abstract of this paper is to appear in the Intelligence and Security Informatics 2007 (ISI 2007) Conference.") follows immediately from the code starting at line 4 of Figure [6](#S4.F6 "Figure 6 ‣ 4.1.3 The Online Algorithm’s Main ‣ 4.1 The Growth Phase ‣ 4 Online Output of Jakobsson’s Pebbles for any 𝑛 ‣ An Online Algorithm for Generating Fractal Hash Chains Applied to Digital Chains of Custody11footnote 1 An extended abstract of this paper is to appear in the Intelligence and Security Informatics 2007 (ISI 2007) Conference."). Jakobsson assumes $`n`$ is a power of two for simplicity \[[14](#bib.bib14)\]. In the case of the online algorithm, values of $`n`$ other than powers of two are also critical.
######
Theorem 1
Suppose that the online algorithm in Figure [6](#S4.F6 "Figure 6 ‣ 4.1.3 The Online Algorithm’s Main ‣ 4.1 The Growth Phase ‣ 4 Online Output of Jakobsson’s Pebbles for any 𝑛 ‣ An Online Algorithm for Generating Fractal Hash Chains Applied to Digital Chains of Custody11footnote 1 An extended abstract of this paper is to appear in the Intelligence and Security Informatics 2007 (ISI 2007) Conference.") is halted after generating $`n`$ total hash elements. The pebbles associated with the hash chain of size $`n`$ will be stored at positions equivalent to the positions in Jakobsson’s algorithm.
Proof: The case where $`n = 2^{k}`$, the online and amortized algorithm both have pebbles in the same positions by Fact [1](#Thmfact1 "Fact 1 ‣ 4 Online Output of Jakobsson’s Pebbles for any 𝑛 ‣ An Online Algorithm for Generating Fractal Hash Chains Applied to Digital Chains of Custody11footnote 1 An extended abstract of this paper is to appear in the Intelligence and Security Informatics 2007 (ISI 2007) Conference.").
The proof follows by induction on the ranges from $`\lbrack 2^{k},2^{k + 1})`$ to $`\lbrack 2^{k + 1},2^{k + 2})`$, for $`k \geq 2`$.
Basis: Take a hash chain with $`k = 2`$, thus $`n = 2^{2} = 4`$ elements as stored by the online by the algorithm in Figure [6](#S4.F6 "Figure 6 ‣ 4.1.3 The Online Algorithm’s Main ‣ 4.1 The Growth Phase ‣ 4 Online Output of Jakobsson’s Pebbles for any 𝑛 ‣ An Online Algorithm for Generating Fractal Hash Chains Applied to Digital Chains of Custody11footnote 1 An extended abstract of this paper is to appear in the Intelligence and Security Informatics 2007 (ISI 2007) Conference.").
<table>
<tbody>
<tr>
<td></td>
<td>p[1].𝐏𝐨𝐬𝐢𝐭𝐢𝐨𝐧formulae-sequence𝑝delimited-[]1𝐏𝐨𝐬𝐢𝐭𝐢𝐨𝐧\displaystyle p[1].{\bf Position}</td>
<td>=\displaystyle=</td>
<td>22\displaystyle 2</td>
<td></td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td>p[1].𝐌𝐨𝐯𝐞_𝐈𝐧𝐜𝐫𝐞𝐦𝐞𝐧𝐭formulae-sequence𝑝delimited-[]1𝐌𝐨𝐯𝐞_𝐈𝐧𝐜𝐫𝐞𝐦𝐞𝐧𝐭\displaystyle p[1].{\bf Move\_Increment}</td>
<td>=\displaystyle=</td>
<td>4.4\displaystyle 4.</td>
<td></td>
</tr>
</tbody>
</table>
As the hash chain is being grown after $`n = 4`$, the first occurrence of CreatePebble is at $`{2^{2} + 1} = 5`$. This pebble is placed at the front of the chain by Fact [5](#Sx1.EGx27 "Fact 5 ‣ 4.3 Characteristics of the Hash Chain Growth Algorithm ‣ 4 Online Output of Jakobsson’s Pebbles for any 𝑛 ‣ An Online Algorithm for Generating Fractal Hash Chains Applied to Digital Chains of Custody11footnote 1 An extended abstract of this paper is to appear in the Intelligence and Security Informatics 2007 (ISI 2007) Conference.") and labeled $`p\hspace{0pt}{\lbrack 2\rbrack}`$ by Fact [7](#Thmfact7 "Fact 7 ‣ 4.3 Characteristics of the Hash Chain Growth Algorithm ‣ 4 Online Output of Jakobsson’s Pebbles for any 𝑛 ‣ An Online Algorithm for Generating Fractal Hash Chains Applied to Digital Chains of Custody11footnote 1 An extended abstract of this paper is to appear in the Intelligence and Security Informatics 2007 (ISI 2007) Conference."). Furthermore,
<table>
<tbody>
<tr>
<td></td>
<td>p[2].𝐏𝐨𝐬𝐢𝐭𝐢𝐨𝐧formulae-sequence𝑝delimited-[]2𝐏𝐨𝐬𝐢𝐭𝐢𝐨𝐧\displaystyle p[2].{\bf Position}</td>
<td>=\displaystyle=</td>
<td>44\displaystyle 4</td>
<td></td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td>p[2].𝐌𝐨𝐯𝐞_𝐈𝐧𝐜𝐫𝐞𝐦𝐞𝐧𝐭formulae-sequence𝑝delimited-[]2𝐌𝐨𝐯𝐞_𝐈𝐧𝐜𝐫𝐞𝐦𝐞𝐧𝐭\displaystyle p[2].{\bf Move\_Increment}</td>
<td>=\displaystyle=</td>
<td>8.8\displaystyle 8.</td>
<td></td>
</tr>
</tbody>
</table>
since $`{\mathbf{g}\mathbf{r}\mathbf{o}\mathbf{w}}.{{\mathbf{p}\mathbf{e}\mathbf{b}\mathbf{b}\mathbf{l}\mathbf{e}} = 2}`$. Suppose the hash chain growth is halted at $`n = 5`$. Due to Fact [10](#Thmfact10 "Fact 10 ‣ 4.3 Characteristics of the Hash Chain Growth Algorithm ‣ 4 Online Output of Jakobsson’s Pebbles for any 𝑛 ‣ An Online Algorithm for Generating Fractal Hash Chains Applied to Digital Chains of Custody11footnote 1 An extended abstract of this paper is to appear in the Intelligence and Security Informatics 2007 (ISI 2007) Conference.") we know that the final pebble is placed for the seed. The pebbles are sorted by their Positions.
Growing the hash chain to $`n = 6`$, prompts the online hash growth algorithm to check each pebble for $`{\mathbf{t}\mathbf{o}\mathbf{t}\mathbf{a}\mathbf{l}\mathbf{H}\mathbf{a}\mathbf{s}\mathbf{h}\mathbf{E}\mathbf{l}\mathbf{e}\mathbf{m}\mathbf{e}\mathbf{n}\mathbf{t}\mathbf{s}} = t = 6`$ where,
<table>
<tbody>
<tr>
<td></td>
<td>t𝑡\displaystyle t</td>
<td>←←\displaystyle\leftarrow</td>
<td>p[j].𝐌𝐨𝐯𝐞_𝐈𝐧𝐜𝐫𝐞𝐦𝐞𝐧𝐭+p[j].𝐝𝐢𝐬𝐭𝐚𝐧𝐜𝐞𝐟𝐫𝐨𝐦𝐒𝐞𝐞𝐝.formulae-sequence𝑝delimited-[]𝑗𝐌𝐨𝐯𝐞_𝐈𝐧𝐜𝐫𝐞𝐦𝐞𝐧𝐭𝑝delimited-[]𝑗𝐝𝐢𝐬𝐭𝐚𝐧𝐜𝐞𝐟𝐫𝐨𝐦𝐒𝐞𝐞𝐝\displaystyle p[j].{\bf Move\_Increment}+p[j].{\bf distancefromSeed}.</td>
<td></td>
</tr>
</tbody>
</table>
But, $`{\mathbf{t}\mathbf{o}\mathbf{t}\mathbf{a}\mathbf{l}\mathbf{H}\mathbf{a}\mathbf{s}\mathbf{h}\mathbf{E}\mathbf{l}\mathbf{e}\mathbf{m}\mathbf{e}\mathbf{n}\mathbf{t}\mathbf{s}} \neq t`$ for all $`j`$ and $`n = 6`$.
The algorithm finds $`t_{j}`$ where
<table>
<tbody>
<tr>
<td></td>
<td>t1subscript𝑡1\displaystyle t_{1}</td>
<td>←←\displaystyle\leftarrow</td>
<td>p[1].𝐌𝐨𝐯𝐞_𝐈𝐧𝐜𝐫𝐞𝐦𝐞𝐧𝐭+p[1].𝐝𝐢𝐬𝐭𝐚𝐧𝐜𝐞𝐟𝐫𝐨𝐦𝐒𝐞𝐞𝐝formulae-sequence𝑝delimited-[]1𝐌𝐨𝐯𝐞_𝐈𝐧𝐜𝐫𝐞𝐦𝐞𝐧𝐭𝑝delimited-[]1𝐝𝐢𝐬𝐭𝐚𝐧𝐜𝐞𝐟𝐫𝐨𝐦𝐒𝐞𝐞𝐝\displaystyle p[1].{\bf Move\_Increment}+p[1].{\bf distancefromSeed}</td>
<td></td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td>t2subscript𝑡2\displaystyle t_{2}</td>
<td>←←\displaystyle\leftarrow</td>
<td>p[2].𝐌𝐨𝐯𝐞_𝐈𝐧𝐜𝐫𝐞𝐦𝐞𝐧𝐭+p[2].𝐝𝐢𝐬𝐭𝐚𝐧𝐜𝐞𝐟𝐫𝐨𝐦𝐒𝐞𝐞𝐝.formulae-sequence𝑝delimited-[]2𝐌𝐨𝐯𝐞_𝐈𝐧𝐜𝐫𝐞𝐦𝐞𝐧𝐭𝑝delimited-[]2𝐝𝐢𝐬𝐭𝐚𝐧𝐜𝐞𝐟𝐫𝐨𝐦𝐒𝐞𝐞𝐝\displaystyle p[2].{\bf Move\_Increment}+p[2].{\bf distancefromSeed}.</td>
<td></td>
</tr>
</tbody>
</table>
At $`n = 5`$ no pebble movement occurs. At $`n = 6 = t_{1}`$ the algorithm sets
<table>
<tbody>
<tr>
<td></td>
<td>p[1].𝐝𝐢𝐬𝐭𝐚𝐧𝐜𝐞𝐟𝐫𝐨𝐦𝐒𝐞𝐞𝐝formulae-sequence𝑝delimited-[]1𝐝𝐢𝐬𝐭𝐚𝐧𝐜𝐞𝐟𝐫𝐨𝐦𝐒𝐞𝐞𝐝\displaystyle p[1].{\bf distancefromSeed}</td>
<td>=\displaystyle=</td>
<td>p[1].𝐝𝐢𝐬𝐭𝐚𝐧𝐜𝐞𝐟𝐫𝐨𝐦𝐒𝐞𝐞𝐝+p[1].𝐌𝐨𝐯𝐞_𝐈𝐧𝐜𝐫𝐞𝐦𝐞𝐧𝐭,formulae-sequence𝑝delimited-[]1𝐝𝐢𝐬𝐭𝐚𝐧𝐜𝐞𝐟𝐫𝐨𝐦𝐒𝐞𝐞𝐝𝑝delimited-[]1𝐌𝐨𝐯𝐞_𝐈𝐧𝐜𝐫𝐞𝐦𝐞𝐧𝐭\displaystyle p[1].{\bf distancefromSeed}+p[1].{\bf Move\_Increment},</td>
<td></td>
</tr>
</tbody>
</table>
such that
<table>
<tbody>
<tr>
<td></td>
<td>p[1].𝐝𝐢𝐬𝐭𝐚𝐧𝐜𝐞𝐟𝐫𝐨𝐦𝐒𝐞𝐞𝐝formulae-sequence𝑝delimited-[]1𝐝𝐢𝐬𝐭𝐚𝐧𝐜𝐞𝐟𝐫𝐨𝐦𝐒𝐞𝐞𝐝\displaystyle p[1].{\bf distancefromSeed}</td>
<td>=\displaystyle=</td>
<td>𝐭𝐨𝐭𝐚𝐥𝐇𝐚𝐬𝐡𝐄𝐥𝐞𝐦𝐞𝐧𝐭𝐬= 6.\displaystyle{\bf totalHashElements}\ \ =\ \ 6.</td>
<td></td>
</tr>
</tbody>
</table>
Suppose the hash chain growth is halted at $`n = 7`$. Again, due to Fact [10](#Thmfact10 "Fact 10 ‣ 4.3 Characteristics of the Hash Chain Growth Algorithm ‣ 4 Online Output of Jakobsson’s Pebbles for any 𝑛 ‣ An Online Algorithm for Generating Fractal Hash Chains Applied to Digital Chains of Custody11footnote 1 An extended abstract of this paper is to appear in the Intelligence and Security Informatics 2007 (ISI 2007) Conference.") we know that the final pebble is the seed. Thus, for $`n = 7`$ we have pebbles at $`{p\hspace{0pt}{\lbrack 1\rbrack}}.{{\mathbf{P}\mathbf{o}\mathbf{s}\mathbf{i}\mathbf{t}\mathbf{i}\mathbf{o}\mathbf{n}} = 2}`$, $`{p\hspace{0pt}{\lbrack 2\rbrack}}.{{\mathbf{P}\mathbf{o}\mathbf{s}\mathbf{i}\mathbf{t}\mathbf{i}\mathbf{o}\mathbf{n}} = 4}`$ and $`{p\hspace{0pt}{\lbrack 3\rbrack}}.{{\mathbf{P}\mathbf{o}\mathbf{s}\mathbf{i}\mathbf{t}\mathbf{i}\mathbf{o}\mathbf{n}} = 8}`$ by Fact [8](#Sx1.EGx30 "Fact 8 ‣ 4.3 Characteristics of the Hash Chain Growth Algorithm ‣ 4 Online Output of Jakobsson’s Pebbles for any 𝑛 ‣ An Online Algorithm for Generating Fractal Hash Chains Applied to Digital Chains of Custody11footnote 1 An extended abstract of this paper is to appear in the Intelligence and Security Informatics 2007 (ISI 2007) Conference.").
Inductive Hypothesis: The statement of this Lemma holds for hash chains from $`n = 2^{k}`$ to $`n = 2^{k + 1}`$, for some $`k \geq 2`$.
Inductive Step: Consider the case where $`n`$ is increased from $`n = 2^{k}`$ to $`n = 2^{k + 1}`$ where $`k \geq 2`$. By the inductive hypothesis we suppose the pebbles are all placed in the correct positions for $`n \leq 2^{k}`$.
By Jakobsson’s algorithm, for a chain of original size $`n`$ we know that a pebble is only discarded when the total number of remaining elements is $`n/2`$. Thus, the hash chain growth algorithm only adds new elements when the total number of pebbles is $`{({n/2})} + 1`$.
Recall, a pebble is always placed at the seed starring at line 4 of the online algorithm, this assures we always have the proper number of pebbles.
By Fact [11](#Thmfact11 "Fact 11 ‣ Appendix: A Proof of Correctness for Jakobsson’s Algorithm ‣ An Online Algorithm for Generating Fractal Hash Chains Applied to Digital Chains of Custody11footnote 1 An extended abstract of this paper is to appear in the Intelligence and Security Informatics 2007 (ISI 2007) Conference.") we know that in Jakobsson’s algorithm only pebble $`p\hspace{0pt}{\lbrack 1\rbrack}`$ is moved. Further, it is always moved a total of $`{p\hspace{0pt}{\lbrack 1\rbrack}}.{{\mathbf{D}\mathbf{e}\mathbf{s}\mathbf{t}}\hspace{0pt}\_\hspace{0pt}{\mathbf{I}\mathbf{n}\mathbf{c}\mathbf{r}\mathbf{e}\mathbf{m}\mathbf{e}\mathbf{n}\mathbf{t}}}`$ after being moved such that it reaches its destination. Since
<table>
<tbody>
<tr>
<td></td>
<td>p[i].𝐃𝐞𝐬𝐭_𝐈𝐧𝐜𝐫𝐞𝐦𝐞𝐧𝐭formulae-sequence𝑝delimited-[]𝑖𝐃𝐞𝐬𝐭_𝐈𝐧𝐜𝐫𝐞𝐦𝐞𝐧𝐭\displaystyle p[i].{\bf Dest\_Increment}</td>
<td>=\displaystyle=</td>
<td>p[i].𝐌𝐨𝐯𝐞_𝐈𝐧𝐜𝐫𝐞𝐦𝐞𝐧𝐭formulae-sequence𝑝delimited-[]𝑖𝐌𝐨𝐯𝐞_𝐈𝐧𝐜𝐫𝐞𝐦𝐞𝐧𝐭\displaystyle p[i].{\bf Move\_Increment}</td>
<td></td>
</tr>
</tbody>
</table>
and because pebbles are always moved to the front by the online algorithm by Fact [9](#Thmfact9 "Fact 9 ‣ 4.3 Characteristics of the Hash Chain Growth Algorithm ‣ 4 Online Output of Jakobsson’s Pebbles for any 𝑛 ‣ An Online Algorithm for Generating Fractal Hash Chains Applied to Digital Chains of Custody11footnote 1 An extended abstract of this paper is to appear in the Intelligence and Security Informatics 2007 (ISI 2007) Conference."), then we know that pebbles are always moved correctly.
For each $`n`$, there are always enough pebbles to support the total number of elements emitted in the online algorithm. As just shown, these pebbles are always placed correctly completing the proof.
## 5 Further Directions
It would be interesting to extend the algorithms of Sella \[[25](#bib.bib25)\] or Matias and Porat \[[19](#bib.bib19)\] to work online. That is, suppose we initially have a data structure capable of holding $`n`$ points of data, then it would be interesting to be able to generate $`n + 1`$ elements of data and expand this data structure suitably, i.e. not doubling it in size.
We have not addressed the tamper-resistant hardware. Likewise, for very small time granularity, we have not addressed clock drift and related timing issues.
## 6 Acknowledgements
Thanks to Markus Jakobsson for his comments on this work. Additionally, the referees’ comments were very helpful and insightful. Thanks to Judge Joseph Colquitt for his perspective on chains of custody and providing us with \[[6](#bib.bib6)\].
## References
- \[1\] A. Ansper, A. Buldas, M. Saarepera, and J. Willemson. Improving the availability of time-stamping services.
In 6th Australasian Conference on Information Security and Privacy (ACISP 2001), number 2119, pages 360–375. International Institute of Informatics and Systemics, Lecture Notes in Computer Science, 2001.
- \[2\] D. Bayer, S. Haber, and W. S. Stornetta. Improving the efficiency and reliability of digital time-stamping.
In Sequences’91: Methods in Communication, Security, and Computer Science, pages 329–334. Springer Verlag, 1992.
- \[3\] A. M. Ben-Amram and H. Petersen. Backing up in singly linked lists.
In Symposium on the Theory of Computing (STOC ’99), pages 780–786. ACM Press, 1999.
- \[4\] A. Buldas, P. Laud, H. Lipmaa, and J. Villemson. Time-stamping with binary linking schemes.
In Hugo Krawczyk, editor, Advances in Cryptology - CRYPTO ’98, 18th Annual International Cryptology Conference, Lecture Notes in Computer Science, number 1462, pages 486–501. Springer-Verlag, 1998.
- \[5\] E. Casey, editor.
Handbook of Computer Crime Investigation: Forensic Tools and Technology.
Academic Press, 2002.
- \[6\] J. A. Colquitt.
Alabama Law of Evidence.
The Mitchie Company–Law Publishers, Charlottesville, VA., 1990.
- \[7\] D. Coppersmith and M. Jakobsson. Almost optimal hash sequence traversal.
In The Fifth Conference on Financial Cryptography (FC ’02). Springer-Verlag, 2002.
- \[8\] M. de Mare and L. DeCoursey. A survey of the timestamping problem. Technical Report SUNYIT-2004-01, 2004. http://turing2.cs.sunyit.edu/reports/.
- \[9\] U. Feige, A. Fiat, and A. Shamir. Zero-knowledge proofs of identity.
Journal of Cryptology, 1:77–94, 1988.
- \[10\] O. Goldreich.
Foundations of Cryptography (Basic Tools).
Cambridge University Press, 2001.
- \[11\] S. Haber and W. S. Stornetta. How to time-stamp a digital document.
Journal of Cryptology, 3(2):99–111, 1991.
- \[12\] S. Haber and W. S. Stornetta. Secure names for bit-strings.
In ACM Conference on Computer and Communications Security, pages 28–35, 1997.
- \[13\] Yih-Chun Hu, Markus Jakobsson, and Adrian Perrig. Efficient constructions for one-way hash chains.
In Applied Cryptography and Network Security (ACNS), June 2005.
- \[14\] M. Jakobsson. Fractal hash sequence representation and traversal.
In 2002 IEEE International Symposium on Information Theory, pages 437–444, 2002.
- \[15\] Sung-Ryul Kim. Improved scalable hash chain traversal.
In Applied Cryptography and Network Security, Lecture Notes in Computer Science \# 2846, pages 86–95. Springer-Verlag, 2003.
- \[16\] Sung-Ryul Kim. Scalable hash chain traversal for mobile devices.
In Computational Science and Its Applications ICCSA 2005, Lecture Notes in Computer Science \# 3480, pages 359–367. Springer-Verlag, 2005.
- \[17\] W. G. Kruse II and J. G. Heiser.
Computer Forensics: Incident Response Essentials.
Addison-Wesley, 2002.
- \[18\] H. Lipmaa. On optimal hash tree traversal for interval time-stamping.
In Agnes Chan and Virgil Gligor, editors, Information Security Conference 2002, Lecture Notes in Computer Science, number 2433, pages 357–371. Springer-Verlag, 2002.
- \[19\] Y. Matias and E. Porat. Efficient pebbling for list traversal synopses.
In 30th International Colloquium on Automata, Languages and Programming (ICALP), Lecture Notes in Computer Science, number 2719, pages 918–928. Springer, 2003.
Full version: http://arxiv.org/abs/cs.DS/0306104.
- \[20\] K. Matsuura and H. Imai. Digital timestamps for dispute settlement in electronic commerce: generation, verification, and renewal. 2002.
- \[21\] A. Perrig, R. Canetti, J. D. Tygar, and D. Song. The tesla broadcast authentication protocol.
RSA CryptoBytes, 5(Summer), 2002.
- \[22\] A. Perrig and J. D. Tygar.
Secure Broadcast Communication in Wired and Wireless Networks.
Kluwer Academic Publishers, Norwell, MA, USA, 2002.
- \[23\] RSA Security Inc. www.rsasecurity.com.
- \[24\] B. Schneier and J. Kelsey. Secure audit logs to support computer forensics.
ACM Transactions on Information System Security, 2(2):159–176, 1999.
- \[25\] Y. Sella. On the computation-storage trade-offs of hash chain traversal.
In 2003 Financial Cryptography, number 2742, pages 270–285. Springer, 2003.
- \[26\] D. R. Stinson.
Cryptography, Theory and Practice.
Chapman & Hall/CRC Press, second edition edition, 2002.
- \[27\] S. Vaudenay.
A Classical Introduction to Cryptography - Applications for Communications Security.
Springer Verlag, 2006.
- \[28\] J. Willemson.
Size Efficient Interval Time Stamps.
PhD thesis, Univeristy of Tartu, 2002. http://home cyber.ee/jam/publ.html.
## Appendix: A Proof of Correctness for Jakobsson’s Algorithm
This appendix gives a correctness proof for Jakobsson’s amortized algorithm. This is done to derive the correctness of the new online algorithm given in this paper. Jakobsson’s paper \[[14](#bib.bib14)\] does not give a complete proof of correctness for his algorithm. However, our online result seems to require certain factual statements from the correctness of Jakobsson’s algorithm.
Ultimately, what we show when Jakobsson’s algorithm runs emitting $`{n/2} = 2^{k}`$ hash elements the remaining pebbles contain hash elements of hash-chain positions $`2^{1},2^{2},2^{3},\ldots,2^{k}`$ which will prove the correctness of Jakobsson’s algorithm.
Jakobsson’s work immediately gives the next observation.
######
Observation 1
Suppose a hash chain has $`n`$ hash elements and $`n = 2^{k}`$, for some integer $`k \geq 1`$ and Jakobsson’s algorithm is about to start. The $`k`$ pebbles initially contain hash elements in hash-chain positions
<table>
<tbody>
<tr>
<td></td>
<td>21,22,23,⋯,2k.superscript21superscript22superscript23⋯superscript2𝑘\displaystyle 2^{1},2^{2},2^{3},\cdots,2^{k}.</td>
<td></td>
</tr>
</tbody>
</table>
Say Jakobsson’s algorithm is about to start on a hash chain of size $`2\hspace{0pt}n`$. Then $`k + 1`$ pebbles initially contain hash elements in hash-chain positions
<table>
<tbody>
<tr>
<td></td>
<td>21,22,23,⋯,2k,2k+1.superscript21superscript22superscript23⋯superscript2𝑘superscript2𝑘1\displaystyle 2^{1},2^{2},2^{3},\cdots,2^{k},2^{k+1}.</td>
<td></td>
</tr>
</tbody>
</table>
A proof of Observation [1](#Sx1.EGx39 "Observation 1 ‣ Appendix: A Proof of Correctness for Jakobsson’s Algorithm ‣ An Online Algorithm for Generating Fractal Hash Chains Applied to Digital Chains of Custody11footnote 1 An extended abstract of this paper is to appear in the Intelligence and Security Informatics 2007 (ISI 2007) Conference.") follows directly from Jakobsson’s setup phase where the pebbles are initialized.
######
Definition 5
A pebble $`p\hspace{0pt}{\lbrack j\rbrack}`$ is moved backward if $`{p\hspace{0pt}{\lbrack j\rbrack}}.{\mathbf{P}\mathbf{o}\mathbf{s}\mathbf{i}\mathbf{t}\mathbf{i}\mathbf{o}\mathbf{n}}`$ is increased in line 3.1 of the algorithm in Figure [2](#S3.F2 "Figure 2 ‣ 3.1.2 Jakobsson’s Main. ‣ 3.1 A Review of Jakobsson’s Algorithm ‣ 3 Jakobsson’s Algorithm ‣ An Online Algorithm for Generating Fractal Hash Chains Applied to Digital Chains of Custody11footnote 1 An extended abstract of this paper is to appear in the Intelligence and Security Informatics 2007 (ISI 2007) Conference."). A pebble $`p\hspace{0pt}{\lbrack j\rbrack}`$ is moved forward if $`{p\hspace{0pt}{\lbrack j\rbrack}}.{\mathbf{P}\mathbf{o}\mathbf{s}\mathbf{i}\mathbf{t}\mathbf{i}\mathbf{o}\mathbf{n}}`$ is decreased by 2 in line 2.1 of this algorithm.
######
Fact 11
Only pebble $`p\hspace{0pt}{\lbrack 1\rbrack}`$ moves backwards in Jakobsson’s algorithm. Furthermore, it only does so when Current.position is even.
Proof: This holds since, the major else in section 3 of Figure [2](#S3.F2 "Figure 2 ‣ 3.1.2 Jakobsson’s Main. ‣ 3.1 A Review of Jakobsson’s Algorithm ‣ 3 Jakobsson’s Algorithm ‣ An Online Algorithm for Generating Fractal Hash Chains Applied to Digital Chains of Custody11footnote 1 An extended abstract of this paper is to appear in the Intelligence and Security Informatics 2007 (ISI 2007) Conference.") is the only place Jakobsson’s algorithm increases any pebble position.
######
Fact 12
A pebble $`p\hspace{0pt}{\lbrack j\rbrack}`$ moves two hash elements forward in lines 2.1 and 2.2 of Jakobsson’s algorithm only after $`p\hspace{0pt}{\lbrack j\rbrack}`$ was moved previously from the front of the chain by Fact [11](#Thmfact11 "Fact 11 ‣ Appendix: A Proof of Correctness for Jakobsson’s Algorithm ‣ An Online Algorithm for Generating Fractal Hash Chains Applied to Digital Chains of Custody11footnote 1 An extended abstract of this paper is to appear in the Intelligence and Security Informatics 2007 (ISI 2007) Conference.").
Proof: In code section 2, making $`{p\hspace{0pt}{\lbrack j\rbrack}}.{{\mathbf{P}\mathbf{o}\mathbf{s}\mathbf{i}\mathbf{t}\mathbf{i}\mathbf{o}\mathbf{n}} \neq {p\hspace{0pt}{\lbrack j\rbrack}}}.{\mathbf{D}\mathbf{e}\mathbf{s}\mathbf{t}\mathbf{i}\mathbf{n}\mathbf{a}\mathbf{t}\mathbf{i}\mathbf{o}\mathbf{n}}`$ may only be done by code in section 3 of Figure [2](#S3.F2 "Figure 2 ‣ 3.1.2 Jakobsson’s Main. ‣ 3.1 A Review of Jakobsson’s Algorithm ‣ 3 Jakobsson’s Algorithm ‣ An Online Algorithm for Generating Fractal Hash Chains Applied to Digital Chains of Custody11footnote 1 An extended abstract of this paper is to appear in the Intelligence and Security Informatics 2007 (ISI 2007) Conference.").
Recall that Jakobsson’s algorithm initially sets $`{p\hspace{0pt}{\lbrack j\rbrack}}.{{\mathbf{P}\mathbf{o}\mathbf{s}\mathbf{i}\mathbf{t}\mathbf{i}\mathbf{o}\mathbf{n}} = {p\hspace{0pt}{\lbrack j\rbrack}}}.{\mathbf{D}\mathbf{e}\mathbf{s}\mathbf{t}\mathbf{i}\mathbf{n}\mathbf{a}\mathbf{t}\mathbf{i}\mathbf{o}\mathbf{n}}`$, for all pebbles $`j:{\sigma \geq j \geq 1}`$. Where $`\sigma`$ is the number of pebbles.
######
Fact 13
For any pebble $`p\hspace{0pt}{\lbrack j\rbrack}`$, before or after any invocation of Jakobsson’s algorithm in Figure [2](#S3.F2 "Figure 2 ‣ 3.1.2 Jakobsson’s Main. ‣ 3.1 A Review of Jakobsson’s Algorithm ‣ 3 Jakobsson’s Algorithm ‣ An Online Algorithm for Generating Fractal Hash Chains Applied to Digital Chains of Custody11footnote 1 An extended abstract of this paper is to appear in the Intelligence and Security Informatics 2007 (ISI 2007) Conference."), the bound holds:
<table>
<tbody>
<tr>
<td></td>
<td>p[j].𝐏𝐨𝐬𝐢𝐭𝐢𝐨𝐧formulae-sequence𝑝delimited-[]𝑗𝐏𝐨𝐬𝐢𝐭𝐢𝐨𝐧\displaystyle p[j].{\bf Position}</td>
<td>≥\displaystyle\geq</td>
<td>p[j].𝐃𝐞𝐬𝐭𝐢𝐧𝐚𝐭𝐢𝐨𝐧,formulae-sequence𝑝delimited-[]𝑗𝐃𝐞𝐬𝐭𝐢𝐧𝐚𝐭𝐢𝐨𝐧\displaystyle p[j].{\bf Destination},</td>
<td></td>
</tr>
</tbody>
</table>
for all $`j:{\sigma \geq j \geq 1}`$ where $`{p\hspace{0pt}{\lbrack j\rbrack}}.{{\mathbf{D}\mathbf{e}\mathbf{s}\mathbf{t}\mathbf{i}\mathbf{n}\mathbf{a}\mathbf{t}\mathbf{i}\mathbf{o}\mathbf{n}} \neq {+ \infty}}`$.
Proof: For each pebble $`j \geq 1`$, the values $`{{\mathbf{S}\mathbf{t}\mathbf{a}\mathbf{r}\mathbf{t}}\hspace{0pt}\_\hspace{0pt}{\mathbf{I}\mathbf{n}\mathbf{c}\mathbf{r}\mathbf{e}\mathbf{m}\mathbf{e}\mathbf{n}\mathbf{t}}} = {3 \times 2^{j}}`$ and $`{{\mathbf{D}\mathbf{e}\mathbf{s}\mathbf{t}}\hspace{0pt}\_\hspace{0pt}{\mathbf{I}\mathbf{n}\mathbf{c}\mathbf{r}\mathbf{e}\mathbf{m}\mathbf{e}\mathbf{n}\mathbf{t}}} = 2^{j + 1}`$ are both even and never change. Initially in the setup phase
<table>
<tbody>
<tr>
<td></td>
<td>p[j].𝐃𝐞𝐬𝐭𝐢𝐧𝐚𝐭𝐢𝐨𝐧formulae-sequence𝑝delimited-[]𝑗𝐃𝐞𝐬𝐭𝐢𝐧𝐚𝐭𝐢𝐨𝐧\displaystyle p[j].{\bf Destination}</td>
<td>=\displaystyle=</td>
<td>2jsuperscript2𝑗\displaystyle 2^{j}</td>
<td></td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td>p[j].𝐏𝐨𝐬𝐢𝐭𝐢𝐨𝐧formulae-sequence𝑝delimited-[]𝑗𝐏𝐨𝐬𝐢𝐭𝐢𝐨𝐧\displaystyle p[j].{\bf Position}</td>
<td>=\displaystyle=</td>
<td>2j.superscript2𝑗\displaystyle 2^{j}.</td>
<td></td>
</tr>
</tbody>
</table>
In section 3, the bound $`{p\hspace{0pt}{\lbrack j\rbrack}}.{{\mathbf{P}\mathbf{o}\mathbf{s}\mathbf{i}\mathbf{t}\mathbf{i}\mathbf{o}\mathbf{n}} \geq {p\hspace{0pt}{\lbrack j\rbrack}}}.{\mathbf{D}\mathbf{e}\mathbf{s}\mathbf{t}\mathbf{i}\mathbf{n}\mathbf{a}\mathbf{t}\mathbf{i}\mathbf{o}\mathbf{n}}`$ is always enforced when Current.position is even, due to the values of $`{\mathbf{S}\mathbf{t}\mathbf{a}\mathbf{r}\mathbf{t}}\hspace{0pt}\_\hspace{0pt}{\mathbf{I}\mathbf{n}\mathbf{c}\mathbf{r}\mathbf{e}\mathbf{m}\mathbf{e}\mathbf{n}\mathbf{t}}`$ and $`{\mathbf{D}\mathbf{e}\mathbf{s}\mathbf{t}}\hspace{0pt}\_\hspace{0pt}{\mathbf{I}\mathbf{n}\mathbf{c}\mathbf{r}\mathbf{e}\mathbf{m}\mathbf{e}\mathbf{n}\mathbf{t}}`$ and also the assignments on lines 3.1 and 3.2,
<table>
<tbody>
<tr>
<td></td>
<td>p[j].𝐏𝐨𝐬𝐢𝐭𝐢𝐨𝐧formulae-sequence𝑝delimited-[]𝑗𝐏𝐨𝐬𝐢𝐭𝐢𝐨𝐧\displaystyle p[j].{\bf Position}</td>
<td>←←\displaystyle\leftarrow</td>
<td>p[j].𝐏𝐨𝐬𝐢𝐭𝐢𝐨𝐧+p[j].𝐒𝐭𝐚𝐫𝐭_𝐈𝐧𝐜𝐫𝐞𝐦𝐞𝐧𝐭formulae-sequence𝑝delimited-[]𝑗𝐏𝐨𝐬𝐢𝐭𝐢𝐨𝐧𝑝delimited-[]𝑗𝐒𝐭𝐚𝐫𝐭_𝐈𝐧𝐜𝐫𝐞𝐦𝐞𝐧𝐭\displaystyle p[j].{\bf Position}+p[j].{\bf Start\_Increment}</td>
<td></td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td>p[j].𝐃𝐞𝐬𝐭𝐢𝐧𝐚𝐭𝐢𝐨𝐧formulae-sequence𝑝delimited-[]𝑗𝐃𝐞𝐬𝐭𝐢𝐧𝐚𝐭𝐢𝐨𝐧\displaystyle p[j].{\bf Destination}</td>
<td>←←\displaystyle\leftarrow</td>
<td>p[j].𝐃𝐞𝐬𝐭𝐢𝐧𝐚𝐭𝐢𝐨𝐧+p[j].𝐃𝐞𝐬𝐭_𝐈𝐧𝐜𝐫𝐞𝐦𝐞𝐧𝐭.formulae-sequence𝑝delimited-[]𝑗𝐃𝐞𝐬𝐭𝐢𝐧𝐚𝐭𝐢𝐨𝐧𝑝delimited-[]𝑗𝐃𝐞𝐬𝐭_𝐈𝐧𝐜𝐫𝐞𝐦𝐞𝐧𝐭\displaystyle p[j].{\bf Destination}+p[j].{\bf Dest\_Increment}.</td>
<td></td>
</tr>
</tbody>
</table>
However, in section 2 of Figure [2](#S3.F2 "Figure 2 ‣ 3.1.2 Jakobsson’s Main. ‣ 3.1 A Review of Jakobsson’s Algorithm ‣ 3 Jakobsson’s Algorithm ‣ An Online Algorithm for Generating Fractal Hash Chains Applied to Digital Chains of Custody11footnote 1 An extended abstract of this paper is to appear in the Intelligence and Security Informatics 2007 (ISI 2007) Conference."), if $`{p\hspace{0pt}{\lbrack j\rbrack}}.{{\mathbf{D}\mathbf{e}\mathbf{s}\mathbf{t}\mathbf{i}\mathbf{n}\mathbf{a}\mathbf{t}\mathbf{i}\mathbf{o}\mathbf{n}} \neq {p\hspace{0pt}{\lbrack j\rbrack}}}.{\mathbf{P}\mathbf{o}\mathbf{s}\mathbf{i}\mathbf{t}\mathbf{i}\mathbf{o}\mathbf{n}}`$, then $`{p\hspace{0pt}{\lbrack j\rbrack}}.{\mathbf{P}\mathbf{o}\mathbf{s}\mathbf{i}\mathbf{t}\mathbf{i}\mathbf{o}\mathbf{n}}`$ is decremented as follows:
<table>
<tbody>
<tr>
<td></td>
<td>p[j].𝐏𝐨𝐬𝐢𝐭𝐢𝐨𝐧formulae-sequence𝑝delimited-[]𝑗𝐏𝐨𝐬𝐢𝐭𝐢𝐨𝐧\displaystyle p[j].{\bf Position}</td>
<td>←←\displaystyle\leftarrow</td>
<td>p[j].𝐏𝐨𝐬𝐢𝐭𝐢𝐨𝐧−2.formulae-sequence𝑝delimited-[]𝑗𝐏𝐨𝐬𝐢𝐭𝐢𝐨𝐧2\displaystyle p[j].{\bf Position}-2.</td>
<td></td>
</tr>
</tbody>
</table>
By Fact [12](#Thmfact12 "Fact 12 ‣ Appendix: A Proof of Correctness for Jakobsson’s Algorithm ‣ An Online Algorithm for Generating Fractal Hash Chains Applied to Digital Chains of Custody11footnote 1 An extended abstract of this paper is to appear in the Intelligence and Security Informatics 2007 (ISI 2007) Conference."), during a run of Jakobsson’s algorithm, both
<table>
<tbody>
<tr>
<td></td>
<td>p[1].𝐏𝐨𝐬𝐢𝐭𝐢𝐨𝐧 and p[1].𝐃𝐞𝐬𝐭𝐢𝐧𝐚𝐭𝐢𝐨𝐧formulae-sequence𝑝delimited-[]1𝐏𝐨𝐬𝐢𝐭𝐢𝐨𝐧 and 𝑝delimited-[]1𝐃𝐞𝐬𝐭𝐢𝐧𝐚𝐭𝐢𝐨𝐧\displaystyle p[1].{\bf Position}\mbox{ and }p[1].{\bf Destination}</td>
<td></td>
</tr>
</tbody>
</table>
remain even values since lines 3.1 and 3.2 add even values to even values.
Line 3.2 ensures
<table>
<tbody>
<tr>
<td></td>
<td>p[1].𝐏𝐨𝐬𝐢𝐭𝐢𝐨𝐧formulae-sequence𝑝delimited-[]1𝐏𝐨𝐬𝐢𝐭𝐢𝐨𝐧\displaystyle p[1].{\bf Position}</td>
<td>≥\displaystyle\geq</td>
<td>p[1].𝐃𝐞𝐬𝐭𝐢𝐧𝐚𝐭𝐢𝐨𝐧formulae-sequence𝑝delimited-[]1𝐃𝐞𝐬𝐭𝐢𝐧𝐚𝐭𝐢𝐨𝐧\displaystyle p[1].{\bf Destination}</td>
<td></td>
</tr>
</tbody>
</table>
and line 2.1 decrements $`{p\hspace{0pt}{\lbrack j\rbrack}}.{\mathbf{P}\mathbf{o}\mathbf{s}\mathbf{i}\mathbf{t}\mathbf{i}\mathbf{o}\mathbf{n}}`$ by 2, spanning all even numbers going down. Thus $`{p\hspace{0pt}{\lbrack j\rbrack}}.{\mathbf{P}\mathbf{o}\mathbf{s}\mathbf{i}\mathbf{t}\mathbf{i}\mathbf{o}\mathbf{n}}`$ must eventually equal $`{p\hspace{0pt}{\lbrack j\rbrack}}.{\mathbf{D}\mathbf{e}\mathbf{s}\mathbf{t}\mathbf{i}\mathbf{n}\mathbf{a}\mathbf{t}\mathbf{i}\mathbf{o}\mathbf{n}}`$, completing the proof.
######
Definition 6
The expression $`p\hspace{0pt}{\lbrack{\mathbf{j}\hspace{0pt}\overset{k}{\rightarrow}\hspace{0pt}i}\rbrack}`$ indicates the pebble initially setup in position $`j`$ is currently in position $`i`$ after $`k`$ backward moves. The expression $`p\hspace{0pt}{\lbrack{j\hspace{0pt}\overset{k}{\rightarrow}\hspace{0pt}i}\rbrack}`$ indicates this pebble was previously in position $`j`$, but $`j`$ is not necessarily the initial position of this pebble in setup.
The expression $`p\hspace{0pt}{\lbrack{j\hspace{0pt}\overset{0}{\rightarrow}\hspace{0pt}1}\rbrack}`$ indicates a pebble that was in position $`j`$ becomes $`p\hspace{0pt}{\lbrack 1\rbrack}`$ without moving backwards. That is, the pebble labeled $`p\hspace{0pt}{\lbrack j\rbrack}`$ becomes $`p\hspace{0pt}{\lbrack 1\rbrack}`$ after emission of $`j - 1`$ hash elements.
Note, $`p\hspace{0pt}{\lbrack{\mathbf{j}\hspace{0pt}\overset{k}{\rightarrow}\hspace{0pt}1}\rbrack}`$ indicates that the pebble initially labeled $`p\hspace{0pt}{\lbrack j\rbrack}`$ has become pebble $`p\hspace{0pt}{\lbrack 1\rbrack}`$ exactly $`k`$ times. Also, $`p\hspace{0pt}{\lbrack{\mathbf{1}\hspace{0pt}\overset{1}{\rightarrow}\hspace{0pt}j}\rbrack}`$ indicates that the unique pebble initially labeled $`p\hspace{0pt}{\lbrack 1\rbrack}`$ is moved backwards exactly once and eventually relabeled pebble $`p\hspace{0pt}{\lbrack j\rbrack}`$. Where $`p\hspace{0pt}{\lbrack{1\hspace{0pt}\overset{1}{\rightarrow}\hspace{0pt}j}\rbrack}`$ is a pebble that was earlier in position 1 and moved back once, but perhaps it was in another position initially at setup.
######
Definition 7
Let $`D\hspace{0pt}{(j,k)}`$ be the value of $`{p\hspace{0pt}{\lbrack{\mathbf{j}\hspace{0pt}\overset{k}{\rightarrow}\hspace{0pt}i}\rbrack}}.{\mathbf{D}\mathbf{e}\mathbf{s}\mathbf{t}\mathbf{i}\mathbf{n}\mathbf{a}\mathbf{t}\mathbf{i}\mathbf{o}\mathbf{n}}`$ immediately after $`k`$ backward moves of the pebble initially setup as pebble $`p\hspace{0pt}{\lbrack j\rbrack}`$ and became $`p\hspace{0pt}{\lbrack i\rbrack}`$, by the Algorithm in Figure [2](#S3.F2 "Figure 2 ‣ 3.1.2 Jakobsson’s Main. ‣ 3.1 A Review of Jakobsson’s Algorithm ‣ 3 Jakobsson’s Algorithm ‣ An Online Algorithm for Generating Fractal Hash Chains Applied to Digital Chains of Custody11footnote 1 An extended abstract of this paper is to appear in the Intelligence and Security Informatics 2007 (ISI 2007) Conference.").
Thus, by Fact [11](#Thmfact11 "Fact 11 ‣ Appendix: A Proof of Correctness for Jakobsson’s Algorithm ‣ An Online Algorithm for Generating Fractal Hash Chains Applied to Digital Chains of Custody11footnote 1 An extended abstract of this paper is to appear in the Intelligence and Security Informatics 2007 (ISI 2007) Conference."), each backward move is initiated when the initial $`p\hspace{0pt}{\lbrack j\rbrack}`$ has become $`p\hspace{0pt}{\lbrack 1\rbrack}`$.
Next is a bound on $`D\hspace{0pt}{(j,k)}`$, for the initial pebble $`p\hspace{0pt}{\lbrack j\rbrack}`$:
<table>
<tbody>
<tr>
<td></td>
<td>D(j,k)𝐷𝑗𝑘\displaystyle D(j,k)</td>
<td>=\displaystyle=</td>
<td>{2j+2j+1 if k=1D(j,k−1)+2j+1 otherwise. casessuperscript2𝑗superscript2𝑗1 if 𝑘1𝐷𝑗𝑘1superscript2𝑗1 otherwise. \displaystyle\left\{\begin{array}[]{ll}2^{j}+2^{j+1}&\mbox{ if }k=1\\ D(j,k-1)+2^{j+1}&\mbox{ otherwise. }\end{array}\right.</td>
<td></td>
</tr>
</tbody>
</table>
######
Definition 8
Let $`P\hspace{0pt}{(j,k)}`$ be an upper-bound on the value of $`{p\hspace{0pt}{\lbrack{\mathbf{j}\hspace{0pt}\overset{k}{\rightarrow}\hspace{0pt}i}\rbrack}}.{\mathbf{P}\mathbf{o}\mathbf{s}\mathbf{i}\mathbf{t}\mathbf{i}\mathbf{o}\mathbf{n}}`$ immediately after $`k`$ backward moves of the pebble initially setup as pebble $`p\hspace{0pt}{\lbrack j\rbrack}`$ and became $`p\hspace{0pt}{\lbrack i\rbrack}`$, by the Algorithm in Figure [2](#S3.F2 "Figure 2 ‣ 3.1.2 Jakobsson’s Main. ‣ 3.1 A Review of Jakobsson’s Algorithm ‣ 3 Jakobsson’s Algorithm ‣ An Online Algorithm for Generating Fractal Hash Chains Applied to Digital Chains of Custody11footnote 1 An extended abstract of this paper is to appear in the Intelligence and Security Informatics 2007 (ISI 2007) Conference.").
Next is a bound on $`P\hspace{0pt}{(j,k)}`$, for the initial pebble $`p\hspace{0pt}{\lbrack j\rbrack}`$:
<table>
<tbody>
<tr>
<td></td>
<td>P(j,k)𝑃𝑗𝑘\displaystyle P(j,k)</td>
<td>≤\displaystyle\leq</td>
<td>{2j+3⋅2j if k=1P(j,k−1)+3⋅2j otherwise. casessuperscript2𝑗⋅3superscript2𝑗 if 𝑘1𝑃𝑗𝑘1⋅3superscript2𝑗 otherwise. \displaystyle\left\{\begin{array}[]{ll}2^{j}+3\cdot 2^{j}&\mbox{ if }k=1\\ P(j,k-1)+3\cdot 2^{j}&\mbox{ otherwise. }\end{array}\right.</td>
<td></td>
</tr>
</tbody>
</table>
######
Fact 14
Immediately after moving pebble $`p\hspace{0pt}{\lbrack 1\rbrack}`$ backwards to $`p\hspace{0pt}{\lbrack i\rbrack}`$, the next upper bound holds:
<table>
<tbody>
<tr>
<td></td>
<td>p[𝐣→𝑘i].𝐏𝐨𝐬𝐢𝐭𝐢𝐨𝐧−p[𝐣→𝑘i].𝐃𝐞𝐬𝐭𝐢𝐧𝐚𝐭𝐢𝐨𝐧formulae-sequence𝑝delimited-[]𝐣𝑘→𝑖𝐏𝐨𝐬𝐢𝐭𝐢𝐨𝐧𝑝delimited-[]𝐣𝑘→𝑖𝐃𝐞𝐬𝐭𝐢𝐧𝐚𝐭𝐢𝐨𝐧\displaystyle p[{\bf j}\overset{k}{\rightarrow}i].{\bf Position}-p[{\bf j}\overset{k}{\rightarrow}i].{\bf Destination}</td>
<td>≤\displaystyle\leq</td>
<td>2j+k(3⋅2j)−2j−k(2⋅2j)superscript2𝑗𝑘⋅3superscript2𝑗superscript2𝑗𝑘⋅2superscript2𝑗\displaystyle 2^{j}+k\left(3\cdot 2^{j}\right)-2^{j}-k\left(2\cdot 2^{j}\right)</td>
<td></td>
</tr>
<tr>
<td></td>
<td></td>
<td>≤\displaystyle\leq</td>
<td>k⋅2j.⋅𝑘superscript2𝑗\displaystyle k\cdot 2^{j}.</td>
<td></td>
</tr>
</tbody>
</table>
Equality holds immediately after $`p\hspace{0pt}{\lbrack{\mathbf{j}\hspace{0pt}\overset{k}{\rightarrow}\hspace{0pt}i}\rbrack}`$ is placed in position $`i`$.
This fact holds by Fact [13](#Thmfact13 "Fact 13 ‣ Appendix: A Proof of Correctness for Jakobsson’s Algorithm ‣ An Online Algorithm for Generating Fractal Hash Chains Applied to Digital Chains of Custody11footnote 1 An extended abstract of this paper is to appear in the Intelligence and Security Informatics 2007 (ISI 2007) Conference.").
######
Fact 15
Consider the Algorithm in Figure [2](#S3.F2 "Figure 2 ‣ 3.1.2 Jakobsson’s Main. ‣ 3.1 A Review of Jakobsson’s Algorithm ‣ 3 Jakobsson’s Algorithm ‣ An Online Algorithm for Generating Fractal Hash Chains Applied to Digital Chains of Custody11footnote 1 An extended abstract of this paper is to appear in the Intelligence and Security Informatics 2007 (ISI 2007) Conference."). Then for every pebble $`p\hspace{0pt}{\lbrack{\mathbf{j}\hspace{0pt}\overset{k}{\rightarrow}\hspace{0pt}1}\rbrack}`$, it must be that
<table>
<tbody>
<tr>
<td></td>
<td>p[𝐣→𝑘1].𝐏𝐨𝐬𝐢𝐭𝐢𝐨𝐧formulae-sequence𝑝delimited-[]𝐣𝑘→1𝐏𝐨𝐬𝐢𝐭𝐢𝐨𝐧\displaystyle p[{\bf j}\overset{k}{\rightarrow}1].{\bf Position}</td>
<td>=\displaystyle=</td>
<td>p[𝐣→𝑘1].𝐃𝐞𝐬𝐭𝐢𝐧𝐚𝐭𝐢𝐨𝐧,formulae-sequence𝑝delimited-[]𝐣𝑘→1𝐃𝐞𝐬𝐭𝐢𝐧𝐚𝐭𝐢𝐨𝐧\displaystyle p[{\bf j}\overset{k}{\rightarrow}1].{\bf Destination},</td>
<td></td>
</tr>
</tbody>
</table>
where neither $`{p\hspace{0pt}{\lbrack{\mathbf{j}\hspace{0pt}\overset{k}{\rightarrow}\hspace{0pt}1}\rbrack}}.{{\mathbf{P}\mathbf{o}\mathbf{s}\mathbf{i}\mathbf{t}\mathbf{i}\mathbf{o}\mathbf{n}} = {+ \infty}}`$ nor $`{p\hspace{0pt}{\lbrack{\mathbf{j}\hspace{0pt}\overset{k}{\rightarrow}\hspace{0pt}1}\rbrack}}.{{\mathbf{D}\mathbf{e}\mathbf{s}\mathbf{t}\mathbf{i}\mathbf{n}\mathbf{a}\mathbf{t}\mathbf{i}\mathbf{o}\mathbf{n}} = {+ \infty}}`$.
Proof: Consider the next function that describes how $`{p\hspace{0pt}{\lbrack j\rbrack}}.{\mathbf{D}\mathbf{e}\mathbf{s}\mathbf{t}\mathbf{i}\mathbf{n}\mathbf{a}\mathbf{t}\mathbf{i}\mathbf{o}\mathbf{n}}`$ works. Note, that line 3.2 of Figure [2](#S3.F2 "Figure 2 ‣ 3.1.2 Jakobsson’s Main. ‣ 3.1 A Review of Jakobsson’s Algorithm ‣ 3 Jakobsson’s Algorithm ‣ An Online Algorithm for Generating Fractal Hash Chains Applied to Digital Chains of Custody11footnote 1 An extended abstract of this paper is to appear in the Intelligence and Security Informatics 2007 (ISI 2007) Conference.") is the only place the Destination field is changed.
Basis: It must be that
<table>
<tbody>
<tr>
<td></td>
<td>p[𝐣→11].𝐏𝐨𝐬𝐢𝐭𝐢𝐨𝐧formulae-sequence𝑝delimited-[]𝐣1→1𝐏𝐨𝐬𝐢𝐭𝐢𝐨𝐧\displaystyle p[{\bf j}\overset{1}{\rightarrow}1].{\bf Position}</td>
<td>=\displaystyle=</td>
<td>p[𝐣→11].𝐃𝐞𝐬𝐭𝐢𝐧𝐚𝐭𝐢𝐨𝐧,formulae-sequence𝑝delimited-[]𝐣1→1𝐃𝐞𝐬𝐭𝐢𝐧𝐚𝐭𝐢𝐨𝐧\displaystyle p[{\bf j}\overset{1}{\rightarrow}1].{\bf Destination},</td>
<td></td>
</tr>
</tbody>
</table>
since by the upper-bound in Fact [14](#Thmfact14 "Fact 14 ‣ Appendix: A Proof of Correctness for Jakobsson’s Algorithm ‣ An Online Algorithm for Generating Fractal Hash Chains Applied to Digital Chains of Custody11footnote 1 An extended abstract of this paper is to appear in the Intelligence and Security Informatics 2007 (ISI 2007) Conference."),
<table>
<tbody>
<tr>
<td></td>
<td>p[𝐣→11].𝐏𝐨𝐬𝐢𝐭𝐢𝐨𝐧−p[𝐣→11].𝐃𝐞𝐬𝐭𝐢𝐧𝐚𝐭𝐢𝐨𝐧formulae-sequence𝑝delimited-[]𝐣1→1𝐏𝐨𝐬𝐢𝐭𝐢𝐨𝐧𝑝delimited-[]𝐣1→1𝐃𝐞𝐬𝐭𝐢𝐧𝐚𝐭𝐢𝐨𝐧\displaystyle p[{\bf j}\overset{1}{\rightarrow}1].{\bf Position}-p[{\bf j}\overset{1}{\rightarrow}1].{\bf Destination}</td>
<td>≤\displaystyle\leq</td>
<td>2j.superscript2𝑗\displaystyle 2^{j}.</td>
<td></td>
</tr>
</tbody>
</table>
However, $`{D\hspace{0pt}{(j,1)}} = {2^{j} + 2^{j + 1}}`$ and by line 2.1 of the algorithm, we must decrement $`{p\hspace{0pt}{\lbrack{j\hspace{0pt}\overset{1}{\rightarrow}\hspace{0pt}1}\rbrack}}.{\mathbf{P}\mathbf{o}\mathbf{s}\mathbf{i}\mathbf{t}\mathbf{i}\mathbf{o}\mathbf{n}}`$ by 2 each of $`D\hspace{0pt}{(j,1)}`$ times. This is because the pebbles are sorted by Position. So immediately after the initial pebble $`p\hspace{0pt}{\lbrack j\rbrack}`$ is moved back, then its destination is $`D\hspace{0pt}{(j,1)}`$. Thus, after $`D\hspace{0pt}{(j,1)}`$ more hash elements are emitted, then this pebble will be labeled $`p\hspace{0pt}{\lbrack 1\rbrack}`$.
Thus, since $`{2\hspace{0pt}D\hspace{0pt}{(j,1)}} > 2^{j}`$, by Fact [13](#Thmfact13 "Fact 13 ‣ Appendix: A Proof of Correctness for Jakobsson’s Algorithm ‣ An Online Algorithm for Generating Fractal Hash Chains Applied to Digital Chains of Custody11footnote 1 An extended abstract of this paper is to appear in the Intelligence and Security Informatics 2007 (ISI 2007) Conference."),
<table>
<tbody>
<tr>
<td></td>
<td>p[𝐣→11].𝐏𝐨𝐬𝐢𝐭𝐢𝐨𝐧formulae-sequence𝑝delimited-[]𝐣1→1𝐏𝐨𝐬𝐢𝐭𝐢𝐨𝐧\displaystyle p[{\bf j}\overset{1}{\rightarrow}1].{\bf Position}</td>
<td>=\displaystyle=</td>
<td>p[𝐣→11].𝐃𝐞𝐬𝐭𝐢𝐧𝐚𝐭𝐢𝐨𝐧,formulae-sequence𝑝delimited-[]𝐣1→1𝐃𝐞𝐬𝐭𝐢𝐧𝐚𝐭𝐢𝐨𝐧\displaystyle p[{\bf j}\overset{1}{\rightarrow}1].{\bf Destination},</td>
<td></td>
</tr>
</tbody>
</table>
completing the basis.
Inductive Hypothesis: Suppose in $`k \geq 1`$ backward moves of the initial pebble $`p\hspace{0pt}{\lbrack j\rbrack}`$ by the Algorithm in Figure [2](#S3.F2 "Figure 2 ‣ 3.1.2 Jakobsson’s Main. ‣ 3.1 A Review of Jakobsson’s Algorithm ‣ 3 Jakobsson’s Algorithm ‣ An Online Algorithm for Generating Fractal Hash Chains Applied to Digital Chains of Custody11footnote 1 An extended abstract of this paper is to appear in the Intelligence and Security Informatics 2007 (ISI 2007) Conference."), we have
<table>
<tbody>
<tr>
<td></td>
<td>p[𝐣→𝑘1].𝐏𝐨𝐬𝐢𝐭𝐢𝐨𝐧formulae-sequence𝑝delimited-[]𝐣𝑘→1𝐏𝐨𝐬𝐢𝐭𝐢𝐨𝐧\displaystyle p[{\bf j}\overset{k}{\rightarrow}1].{\bf Position}</td>
<td>=\displaystyle=</td>
<td>p[𝐣→𝑘1].𝐃𝐞𝐬𝐭𝐢𝐧𝐚𝐭𝐢𝐨𝐧.formulae-sequence𝑝delimited-[]𝐣𝑘→1𝐃𝐞𝐬𝐭𝐢𝐧𝐚𝐭𝐢𝐨𝐧\displaystyle p[{\bf j}\overset{k}{\rightarrow}1].{\bf Destination}.</td>
<td></td>
</tr>
</tbody>
</table>
Inductive Step: For some $`k \geq 1`$, consider $`k + 1`$ backward moves of the initial pebble $`p\hspace{0pt}{\lbrack j'\rbrack}`$ by Algorithm in Figure [2](#S3.F2 "Figure 2 ‣ 3.1.2 Jakobsson’s Main. ‣ 3.1 A Review of Jakobsson’s Algorithm ‣ 3 Jakobsson’s Algorithm ‣ An Online Algorithm for Generating Fractal Hash Chains Applied to Digital Chains of Custody11footnote 1 An extended abstract of this paper is to appear in the Intelligence and Security Informatics 2007 (ISI 2007) Conference."), and this step will show:
<table>
<tbody>
<tr>
<td></td>
<td>p[𝐣′⟶k+11].𝐏𝐨𝐬𝐢𝐭𝐢𝐨𝐧formulae-sequence𝑝delimited-[]superscript𝐣′𝑘1⟶1𝐏𝐨𝐬𝐢𝐭𝐢𝐨𝐧\displaystyle p[{\bf j^{\prime}}\overset{k+1}{\longrightarrow}1].{\bf Position}</td>
<td>=\displaystyle=</td>
<td>p[𝐣′⟶k+11].𝐃𝐞𝐬𝐭𝐢𝐧𝐚𝐭𝐢𝐨𝐧.formulae-sequence𝑝delimited-[]superscript𝐣′𝑘1⟶1𝐃𝐞𝐬𝐭𝐢𝐧𝐚𝐭𝐢𝐨𝐧\displaystyle p[{\bf j^{\prime}}\overset{k+1}{\longrightarrow}1].{\bf Destination}.</td>
<td></td>
</tr>
</tbody>
</table>
Start by considering $`k + 1`$ backward moves of the initial pebble $`p\hspace{0pt}{\lbrack j'\rbrack}`$, for the sake of a contradiction suppose,
<table>
<tbody>
<tr>
<td></td>
<td>p[𝐣′→1j→𝑘1].𝐏𝐨𝐬𝐢𝐭𝐢𝐨𝐧formulae-sequence𝑝delimited-[]superscript𝐣′1→𝑗𝑘→1𝐏𝐨𝐬𝐢𝐭𝐢𝐨𝐧\displaystyle p[{\bf j^{\prime}}\overset{1}{\rightarrow}j\overset{k}{\rightarrow}1].{\bf Position}</td>
<td>≠\displaystyle\neq</td>
<td>p[𝐣′→1j→𝑘1].𝐃𝐞𝐬𝐭𝐢𝐧𝐚𝐭𝐢𝐨𝐧,formulae-sequence𝑝delimited-[]superscript𝐣′1→𝑗𝑘→1𝐃𝐞𝐬𝐭𝐢𝐧𝐚𝐭𝐢𝐨𝐧\displaystyle p[{\bf j^{\prime}}\overset{1}{\rightarrow}j\overset{k}{\rightarrow}1].{\bf Destination},</td>
<td></td>
</tr>
</tbody>
</table>
and in any case, by Fact [14](#Thmfact14 "Fact 14 ‣ Appendix: A Proof of Correctness for Jakobsson’s Algorithm ‣ An Online Algorithm for Generating Fractal Hash Chains Applied to Digital Chains of Custody11footnote 1 An extended abstract of this paper is to appear in the Intelligence and Security Informatics 2007 (ISI 2007) Conference."), the next bound holds:
<table>
<tbody>
<tr>
<td></td>
<td>p[𝐣′→1j].𝐏𝐨𝐬𝐢𝐭𝐢𝐨𝐧−p[𝐣′→1j].𝐃𝐞𝐬𝐭𝐢𝐧𝐚𝐭𝐢𝐨𝐧formulae-sequence𝑝delimited-[]superscript𝐣′1→𝑗𝐏𝐨𝐬𝐢𝐭𝐢𝐨𝐧𝑝delimited-[]superscript𝐣′1→𝑗𝐃𝐞𝐬𝐭𝐢𝐧𝐚𝐭𝐢𝐨𝐧\displaystyle p[{\bf j^{\prime}}\overset{1}{\rightarrow}j].{\bf Position}-p[{\bf j^{\prime}}\overset{1}{\rightarrow}j].{\bf Destination}</td>
<td>≤\displaystyle\leq</td>
<td>2j′.superscript2superscript𝑗′\displaystyle 2^{j^{\prime}}.</td>
<td></td>
</tr>
</tbody>
</table>
However $`{D\hspace{0pt}{(j',1)}} = {2^{j' + 1} + 2^{j'}}`$, which means after $`2^{j' + 1} + 2^{j'}`$ hash elements are emitted to take the initial pebble $`p\hspace{0pt}{\lbrack j'\rbrack}`$ to $`p\hspace{0pt}{\lbrack 1\rbrack}`$ the first time, then
<table>
<tbody>
<tr>
<td></td>
<td>p[𝐣′→1j].𝐏𝐨𝐬𝐢𝐭𝐢𝐨𝐧formulae-sequence𝑝delimited-[]superscript𝐣′1→𝑗𝐏𝐨𝐬𝐢𝐭𝐢𝐨𝐧\displaystyle p[{\bf j^{\prime}}\overset{1}{\rightarrow}j].{\bf Position}</td>
<td></td>
</tr>
</tbody>
</table>
must be decremented by 2 in each run of section 2 of Figure [2](#S3.F2 "Figure 2 ‣ 3.1.2 Jakobsson’s Main. ‣ 3.1 A Review of Jakobsson’s Algorithm ‣ 3 Jakobsson’s Algorithm ‣ An Online Algorithm for Generating Fractal Hash Chains Applied to Digital Chains of Custody11footnote 1 An extended abstract of this paper is to appear in the Intelligence and Security Informatics 2007 (ISI 2007) Conference.") decreasing $`{p\hspace{0pt}{\lbrack{\mathbf{j}'\hspace{0pt}\overset{1}{\rightarrow}\hspace{0pt}1}\rbrack}}.{\mathbf{P}\mathbf{o}\mathbf{s}\mathbf{i}\mathbf{t}\mathbf{i}\mathbf{o}\mathbf{n}}`$ by at least $`{2\hspace{0pt}D\hspace{0pt}{(j',1)}} > 2^{j'}`$.
Now, by the Inductive Hypothesis it must be that,
<table>
<tbody>
<tr>
<td></td>
<td>p[j→𝑘1].𝐏𝐨𝐬𝐢𝐭𝐢𝐨𝐧formulae-sequence𝑝delimited-[]𝑗𝑘→1𝐏𝐨𝐬𝐢𝐭𝐢𝐨𝐧\displaystyle p[j\overset{k}{\rightarrow}1].{\bf Position}</td>
<td>=\displaystyle=</td>
<td>p[j→𝑘1].𝐃𝐞𝐬𝐭𝐢𝐧𝐚𝐭𝐢𝐨𝐧.formulae-sequence𝑝delimited-[]𝑗𝑘→1𝐃𝐞𝐬𝐭𝐢𝐧𝐚𝐭𝐢𝐨𝐧\displaystyle p[j\overset{k}{\rightarrow}1].{\bf Destination}.</td>
<td></td>
</tr>
</tbody>
</table>
giving,
<table>
<tbody>
<tr>
<td></td>
<td>p[𝐣′⟶k+11].𝐏𝐨𝐬𝐢𝐭𝐢𝐨𝐧formulae-sequence𝑝delimited-[]superscript𝐣′𝑘1⟶1𝐏𝐨𝐬𝐢𝐭𝐢𝐨𝐧\displaystyle p[{\bf j^{\prime}}\overset{k+1}{\longrightarrow}1].{\bf Position}</td>
<td>=\displaystyle=</td>
<td>p[𝐣′⟶k+11].𝐃𝐞𝐬𝐭𝐢𝐧𝐚𝐭𝐢𝐨𝐧,formulae-sequence𝑝delimited-[]superscript𝐣′𝑘1⟶1𝐃𝐞𝐬𝐭𝐢𝐧𝐚𝐭𝐢𝐨𝐧\displaystyle p[{\bf j^{\prime}}\overset{k+1}{\longrightarrow}1].{\bf Destination},</td>
<td></td>
</tr>
</tbody>
</table>
indicating our supposition was incorrect, which completes the proof.
######
Fact 16
Suppose a hash chain starts with $`n = 2^{k + 1}`$ total hash elements, for some integer $`k \geq 1`$ and Jakobsson’s algorithm runs emitting $`{n/2} = 2^{k}`$ hash elements. Then the pebble initially in position $`j < k`$ is moved backwards a total of
<table>
<tbody>
<tr>
<td></td>
<td>2k−j−1superscript2𝑘𝑗1\displaystyle 2^{k-j-1}</td>
<td></td>
</tr>
</tbody>
</table>
times.
Proof: By Fact [11](#Thmfact11 "Fact 11 ‣ Appendix: A Proof of Correctness for Jakobsson’s Algorithm ‣ An Online Algorithm for Generating Fractal Hash Chains Applied to Digital Chains of Custody11footnote 1 An extended abstract of this paper is to appear in the Intelligence and Security Informatics 2007 (ISI 2007) Conference."), a pebble only moves back when it is in pebble position 1. By Fact [13](#Thmfact13 "Fact 13 ‣ Appendix: A Proof of Correctness for Jakobsson’s Algorithm ‣ An Online Algorithm for Generating Fractal Hash Chains Applied to Digital Chains of Custody11footnote 1 An extended abstract of this paper is to appear in the Intelligence and Security Informatics 2007 (ISI 2007) Conference."),
<table>
<tbody>
<tr>
<td></td>
<td>p[j].𝐏𝐨𝐬𝐢𝐭𝐢𝐨𝐧formulae-sequence𝑝delimited-[]𝑗𝐏𝐨𝐬𝐢𝐭𝐢𝐨𝐧\displaystyle p[j].{\bf Position}</td>
<td>≥\displaystyle\geq</td>
<td>p[j].𝐃𝐞𝐬𝐭𝐢𝐧𝐚𝐭𝐢𝐨𝐧,formulae-sequence𝑝delimited-[]𝑗𝐃𝐞𝐬𝐭𝐢𝐧𝐚𝐭𝐢𝐨𝐧\displaystyle p[j].{\bf Destination},</td>
<td></td>
</tr>
</tbody>
</table>
so the only concern is how $`{p\hspace{0pt}{\lbrack j\rbrack}}.{\mathbf{D}\mathbf{e}\mathbf{s}\mathbf{t}\mathbf{i}\mathbf{n}\mathbf{a}\mathbf{t}\mathbf{i}\mathbf{o}\mathbf{n}}`$ is initialized and how $`{p\hspace{0pt}{\lbrack j\rbrack}}.{\mathbf{D}\mathbf{e}\mathbf{s}\mathbf{t}\mathbf{i}\mathbf{n}\mathbf{a}\mathbf{t}\mathbf{i}\mathbf{o}\mathbf{n}}`$ changes.
After $`n/2`$ hash elements are emitted, it takes exactly
<table>
<tbody>
<tr>
<td></td>
<td>(n2−2j)1p[j].𝐃𝐞𝐬𝐭_𝐈𝐧𝐜𝐫𝐞𝐦𝐞𝐧𝐭𝑛2superscript2𝑗1formulae-sequence𝑝delimited-[]𝑗𝐃𝐞𝐬𝐭_𝐈𝐧𝐜𝐫𝐞𝐦𝐞𝐧𝐭\displaystyle\left(\frac{n}{2}-2^{j}\right)\frac{1}{p[j].{\bf Dest\_Increment}}</td>
<td>=\displaystyle=</td>
<td>(2k−2j)12j+1superscript2𝑘superscript2𝑗1superscript2𝑗1\displaystyle\left(2^{k}-2^{j}\right)\frac{1}{2^{j+1}}</td>
<td></td>
</tr>
<tr>
<td></td>
<td></td>
<td>=\displaystyle=</td>
<td>2j−k−1−12,superscript2𝑗𝑘112\displaystyle 2^{j-k-1}-\frac{1}{2},</td>
<td></td>
</tr>
</tbody>
</table>
increments of $`{p\hspace{0pt}{\lbrack j\rbrack}}.{\mathbf{D}\mathbf{e}\mathbf{s}\mathbf{t}\mathbf{i}\mathbf{n}\mathbf{a}\mathbf{t}\mathbf{i}\mathbf{o}\mathbf{n}}`$ by the time the pebble initially at position $`j`$ is disposed of.
A pebble is disposed of when $`{p\hspace{0pt}{\lbrack j\rbrack}}.{{\mathbf{D}\mathbf{e}\mathbf{s}\mathbf{t}\mathbf{i}\mathbf{n}\mathbf{a}\mathbf{t}\mathbf{i}\mathbf{o}\mathbf{n}}\leftarrow{+ \infty}}`$. In terms of Fact [16](#Thmfact16 "Fact 16 ‣ Appendix: A Proof of Correctness for Jakobsson’s Algorithm ‣ An Online Algorithm for Generating Fractal Hash Chains Applied to Digital Chains of Custody11footnote 1 An extended abstract of this paper is to appear in the Intelligence and Security Informatics 2007 (ISI 2007) Conference."), the pebble in position $`k + 1`$ is never moved back and the pebble in position $`k`$ is disposed of when it is moved back the first time.
######
Fact 17
Given an $`n = 2^{k}`$ element hash chain and take the Algorithm in Figure [2](#S3.F2 "Figure 2 ‣ 3.1.2 Jakobsson’s Main. ‣ 3.1 A Review of Jakobsson’s Algorithm ‣ 3 Jakobsson’s Algorithm ‣ An Online Algorithm for Generating Fractal Hash Chains Applied to Digital Chains of Custody11footnote 1 An extended abstract of this paper is to appear in the Intelligence and Security Informatics 2007 (ISI 2007) Conference."). The initial pebble in position $`j`$, for $`j = {k - 1} = {\log{({n/2})}}`$ is the only pebble that is disposed of by the time the first $`n/2`$ hash elements are emitted.
Proof: For each pebble, the values Start_Increment and Dest_Increment never change. Initially in setup $`{p\hspace{0pt}{\lbrack j\rbrack}}.{{\mathbf{D}\mathbf{e}\mathbf{s}\mathbf{t}\mathbf{i}\mathbf{n}\mathbf{a}\mathbf{t}\mathbf{i}\mathbf{o}\mathbf{n}}\leftarrow 2^{j}}`$, for all $`j:{{\log n} \geq j \geq 1}`$.
The value $`{p\hspace{0pt}{\lbrack j\rbrack}}.{\mathbf{D}\mathbf{e}\mathbf{s}\mathbf{t}\mathbf{i}\mathbf{n}\mathbf{a}\mathbf{t}\mathbf{i}\mathbf{o}\mathbf{n}}`$ is only increased in line 3.2 of Figure [2](#S3.F2 "Figure 2 ‣ 3.1.2 Jakobsson’s Main. ‣ 3.1 A Review of Jakobsson’s Algorithm ‣ 3 Jakobsson’s Algorithm ‣ An Online Algorithm for Generating Fractal Hash Chains Applied to Digital Chains of Custody11footnote 1 An extended abstract of this paper is to appear in the Intelligence and Security Informatics 2007 (ISI 2007) Conference."). Considering line 3.2 gives the bound
<table>
<tbody>
<tr>
<td></td>
<td>p[j].𝐃𝐞𝐬𝐭𝐢𝐧𝐚𝐭𝐢𝐨𝐧formulae-sequence𝑝delimited-[]𝑗𝐃𝐞𝐬𝐭𝐢𝐧𝐚𝐭𝐢𝐨𝐧\displaystyle p[j].{\bf Destination}</td>
<td>=\displaystyle=</td>
<td>2j+p[j].𝐃𝐞𝐬𝐭_𝐈𝐧𝐜𝐫𝐞𝐦𝐞𝐧𝐭formulae-sequencesuperscript2𝑗𝑝delimited-[]𝑗𝐃𝐞𝐬𝐭_𝐈𝐧𝐜𝐫𝐞𝐦𝐞𝐧𝐭\displaystyle 2^{j}+p[j].{\bf Dest\_Increment}</td>
<td></td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td>p[j].𝐃𝐞𝐬𝐭𝐢𝐧𝐚𝐭𝐢𝐨𝐧formulae-sequence𝑝delimited-[]𝑗𝐃𝐞𝐬𝐭𝐢𝐧𝐚𝐭𝐢𝐨𝐧\displaystyle p[j].{\bf Destination}</td>
<td>=\displaystyle=</td>
<td>2j+2j+1superscript2𝑗superscript2𝑗1\displaystyle 2^{j}+2^{j+1}</td>
<td></td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td>p[j].𝐃𝐞𝐬𝐭𝐢𝐧𝐚𝐭𝐢𝐨𝐧formulae-sequence𝑝delimited-[]𝑗𝐃𝐞𝐬𝐭𝐢𝐧𝐚𝐭𝐢𝐨𝐧\displaystyle p[j].{\bf Destination}</td>
<td>=\displaystyle=</td>
<td>3⋅2j.⋅3superscript2𝑗\displaystyle 3\cdot 2^{j}.</td>
<td></td>
</tr>
</tbody>
</table>
Pebble $`p\hspace{0pt}{\lbrack 1\rbrack}`$ is the only one that is moved back by Fact [11](#Thmfact11 "Fact 11 ‣ Appendix: A Proof of Correctness for Jakobsson’s Algorithm ‣ An Online Algorithm for Generating Fractal Hash Chains Applied to Digital Chains of Custody11footnote 1 An extended abstract of this paper is to appear in the Intelligence and Security Informatics 2007 (ISI 2007) Conference."). Further,
<table>
<tbody>
<tr>
<td></td>
<td>3⋅2j⋅3superscript2𝑗\displaystyle 3\cdot 2^{j}</td>
<td>>\displaystyle></td>
<td>n,𝑛\displaystyle n,</td>
<td></td>
</tr>
</tbody>
</table>
holds for $`j \geq {\log{({n/2})}}`$. But the only mobile pebble that satisfies this constraint is initially in position $`j = {\log{({n/2})}} = {k - 1}`$. There is a pebble $`j = {\log n}`$ at the seed of the hash chain, but this never moves.
Now, take any initial pebble that started in position $`j \leq {\log{({n/4})}}`$ then by Fact [16](#Thmfact16 "Fact 16 ‣ Appendix: A Proof of Correctness for Jakobsson’s Algorithm ‣ An Online Algorithm for Generating Fractal Hash Chains Applied to Digital Chains of Custody11footnote 1 An extended abstract of this paper is to appear in the Intelligence and Security Informatics 2007 (ISI 2007) Conference."), such a pebble makes $`2^{k - j - 1}`$ backward moves by the time $`n/2`$ hash elements are emitted. Thus, by the time the $`n/2`$-nd pebble is emitted, the initial pebble $`p\hspace{0pt}{\lbrack j\rbrack}`$ is moved from the starting position $`2^{j}`$ to position
<table>
<tbody>
<tr>
<td></td>
<td>2j+(2j+1⋅2k−j−1)superscript2𝑗⋅superscript2𝑗1superscript2𝑘𝑗1\displaystyle 2^{j}+\left(2^{j+1}\cdot 2^{k-j-1}\right)</td>
<td>=\displaystyle=</td>
<td>2j+2k,superscript2𝑗superscript2𝑘\displaystyle 2^{j}+2^{k},</td>
<td></td>
</tr>
</tbody>
</table>
and the largest $`j`$ and $`k`$ may be is $`\log{({n/4})}`$ and $`\log{({n/2})}`$, respectively, giving
<table>
<tbody>
<tr>
<td></td>
<td>2j+2ksuperscript2𝑗superscript2𝑘\displaystyle 2^{j}+2^{k}</td>
<td>≤\displaystyle\leq</td>
<td>2logn/4+2logn/2superscript2𝑛4superscript2𝑛2\displaystyle 2^{\log n/4}+2^{\log n/2}</td>
<td></td>
</tr>
<tr>
<td></td>
<td></td>
<td>≤\displaystyle\leq</td>
<td>3n4<n,\displaystyle\frac{3n}{4}\ \ <n,</td>
<td></td>
</tr>
</tbody>
</table>
completing the proof.
######
Fact 18
Suppose a hash chain starts with $`n = 2^{k + 1}`$ total hash elements, and for some integer $`k \geq 1`$, while Jakobsson’s algorithm runs emitting the first $`{n/2} = 2^{k}`$ hash elements, let $`i`$ be the first placement of $`p\hspace{0pt}{\lbrack{\mathbf{j}\rightarrow i}\rbrack}`$ between hash elements $`2^{k}`$ and $`2^{k + 1}`$. Then, Jakobsson’s algorithm will emit a total of
<table>
<tbody>
<tr>
<td></td>
<td>n2−(2k−j−1−1)2j+1−2j𝑛2superscript2𝑘𝑗11superscript2𝑗1superscript2𝑗\displaystyle\frac{n}{2}-\left(2^{k-j-1}-1\right)2^{j+1}-2^{j}</td>
<td></td>
</tr>
</tbody>
</table>
hash elements before emitting hash element $`2^{k}`$ but immediately after $`p\hspace{0pt}{\lbrack{\mathbf{j}\rightarrow i}\rbrack}`$ is placed at position $`i`$.
Proof: After Jakobsson’s algorithm emits hash element $`2^{k}`$, a total of $`n/2`$ total hash elements have already been emitted.
By Fact [16](#Thmfact16 "Fact 16 ‣ Appendix: A Proof of Correctness for Jakobsson’s Algorithm ‣ An Online Algorithm for Generating Fractal Hash Chains Applied to Digital Chains of Custody11footnote 1 An extended abstract of this paper is to appear in the Intelligence and Security Informatics 2007 (ISI 2007) Conference.") the pebble initially at position $`j < k`$ is moved backwards $`2^{k - j - 1}`$ times. That is, it appears in pebble position 1, a total of $`2^{k - j - 1}`$ times.
The last time this pebble moves backward to position $`i`$ is the first time it moves to a position behind hash position $`2^{k}`$. This gives the factor of $`({2^{k - j - 1} - 1})`$.
Finally, for initial pebble $`j`$, since Dest_Increment is always $`2^{j + 1}`$, each time we move backwards in Jakobsson’s algorithm, we move Destination back by $`2^{j + 1}`$ elements. That is, we move back a total of
<table>
<tbody>
<tr>
<td></td>
<td>(2k−j−1−1)2j+1superscript2𝑘𝑗11superscript2𝑗1\displaystyle\left(2^{k-j-1}-1\right)2^{j+1}</td>
<td></td>
</tr>
</tbody>
</table>
hash elements by the time the pebble initially labeled $`j`$ passes hash position $`2^{k}`$.
Furthermore, the initially labeled $`j`$th pebble starts at position $`2^{j}`$, so we adjust for the number of hash elements emitted it takes to traverse $`p\hspace{0pt}{\lbrack{\mathbf{j}\hspace{0pt}\overset{0}{\rightarrow}\hspace{0pt}1}\rbrack}`$.
######
Corollary 1
We know that when $`n/2`$ hash elements are emitted that all remaining pebbles will have
<table>
<tbody>
<tr>
<td></td>
<td>p[j].𝐃𝐞𝐬𝐭𝐢𝐧𝐚𝐭𝐢𝐨𝐧formulae-sequence𝑝delimited-[]𝑗𝐃𝐞𝐬𝐭𝐢𝐧𝐚𝐭𝐢𝐨𝐧\displaystyle p[j].{\bf Destination}</td>
<td>=\displaystyle=</td>
<td>p[j].𝐏𝐨𝐬𝐢𝐭𝐢𝐨𝐧.formulae-sequence𝑝delimited-[]𝑗𝐏𝐨𝐬𝐢𝐭𝐢𝐨𝐧\displaystyle p[j].{\bf Position}.</td>
<td></td>
</tr>
</tbody>
</table>
Proof: We know by Fact [18](#Thmfact18 "Fact 18 ‣ Appendix: A Proof of Correctness for Jakobsson’s Algorithm ‣ An Online Algorithm for Generating Fractal Hash Chains Applied to Digital Chains of Custody11footnote 1 An extended abstract of this paper is to appear in the Intelligence and Security Informatics 2007 (ISI 2007) Conference."), that there are
<table>
<tbody>
<tr>
<td></td>
<td>n2−(2k−j−1−1)2j+1−2j𝑛2superscript2𝑘𝑗11superscript2𝑗1superscript2𝑗\displaystyle\frac{n}{2}-\left(2^{k-j-1}-1\right)2^{j+1}-2^{j}</td>
<td></td>
</tr>
</tbody>
</table>
elements that remain to be output before a total of $`n/2`$ elements are emitted after each pebble at original pebble index $`j`$ is moved for the last time before $`n/2`$ elements are emitted.
By Jakobsson’s algorithm we know that each time a pebble at original position $`j`$ is moved, it will take,
<table>
<tbody>
<tr>
<td></td>
<td>(3⋅2j−2j+1)/2⋅3superscript2𝑗superscript2𝑗12\displaystyle(3\cdot 2^{j}-2^{j+1})/2</td>
<td></td>
</tr>
</tbody>
</table>
hash exposures for $`{p\hspace{0pt}{\lbrack j\rbrack}}.{{\mathbf{P}\mathbf{o}\mathbf{s}\mathbf{i}\mathbf{t}\mathbf{i}\mathbf{o}\mathbf{n}} = {p\hspace{0pt}{\lbrack j\rbrack}}}.{\mathbf{D}\mathbf{e}\mathbf{s}\mathbf{t}\mathbf{i}\mathbf{n}\mathbf{a}\mathbf{t}\mathbf{i}\mathbf{o}\mathbf{n}}`$.
Thus, since
<table>
<tbody>
<tr>
<td></td>
<td>n2−(2k−j−1−1)2j+1−2j>(3⋅2j−2j+1)/2𝑛2superscript2𝑘𝑗11superscript2𝑗1superscript2𝑗⋅3superscript2𝑗superscript2𝑗12\displaystyle\frac{n}{2}-\left(2^{k-j-1}-1\right)2^{j+1}-2^{j}>(3\cdot 2^{j}-2^{j+1})/2</td>
<td></td>
</tr>
</tbody>
</table>
we know that when $`n/2`$ hash elements are emitted that all remaining pebbles will have $`{p\hspace{0pt}{\lbrack j\rbrack}}.{{\mathbf{P}\mathbf{o}\mathbf{s}\mathbf{i}\mathbf{t}\mathbf{i}\mathbf{o}\mathbf{n}} = {p\hspace{0pt}{\lbrack j\rbrack}}}.{\mathbf{D}\mathbf{e}\mathbf{s}\mathbf{t}\mathbf{i}\mathbf{n}\mathbf{a}\mathbf{t}\mathbf{i}\mathbf{o}\mathbf{n}}`$. Thus, completing the proof.
The next fact, can also serve as an alternate proof-of-correctness for Jakobsson’s algorithm \[[14](#bib.bib14)\]. It is presented here, since it is needed subsequently.
First, the function $`R\hspace{0pt}{()}`$ reindexes the pebble indices from one power of two down to the next lowest power of two. When $`n/2`$ pebbles are emitted, then the total number of remaining hash elements goes from $`n = 2^{k + 1}`$ down to $`n = 2^{k}`$.
<table>
<tbody>
<tr>
<td></td>
<td>Rk(lk+1)=lk+1−2k,subscript𝑅𝑘subscript𝑙𝑘1subscript𝑙𝑘1superscript2𝑘\displaystyle R_{k}(l_{k+1})=l_{k+1}-2^{k},</td>
<td></td>
</tr>
</tbody>
</table>
where $`l_{k + 1}`$ is the hash element $`\geq 2^{k + 1}`$ contained by pebble index $`k + 1`$, and $`R_{k}\hspace{0pt}{(l_{k + 1})}`$ is the adjusted index of the hash element $`\geq 2^{k}`$.
######
Fact 19
Suppose a hash chain starts with $`n = 2^{k + 1}`$ total hash elements, for some integer $`k \geq 1`$ and Jakobsson’s algorithm runs emitting $`{n/2} = 2^{k}`$ hash elements. Then the remaining pebbles contain hash elements of hash-chain positions
<table>
<tbody>
<tr>
<td></td>
<td>21,22,23,⋯,2k.superscript21superscript22superscript23⋯superscript2𝑘\displaystyle 2^{1},2^{2},2^{3},\cdots,2^{k}.</td>
<td></td>
</tr>
</tbody>
</table>
Proof: This proof follows by induction on the algorithm in Figure [2](#S3.F2 "Figure 2 ‣ 3.1.2 Jakobsson’s Main. ‣ 3.1 A Review of Jakobsson’s Algorithm ‣ 3 Jakobsson’s Algorithm ‣ An Online Algorithm for Generating Fractal Hash Chains Applied to Digital Chains of Custody11footnote 1 An extended abstract of this paper is to appear in the Intelligence and Security Informatics 2007 (ISI 2007) Conference."). Setting $`k = 2`$ ensures sections 1,2 and 3 of Jakobsson’s algorithm will run (see Figure [2](#S3.F2 "Figure 2 ‣ 3.1.2 Jakobsson’s Main. ‣ 3.1 A Review of Jakobsson’s Algorithm ‣ 3 Jakobsson’s Algorithm ‣ An Online Algorithm for Generating Fractal Hash Chains Applied to Digital Chains of Custody11footnote 1 An extended abstract of this paper is to appear in the Intelligence and Security Informatics 2007 (ISI 2007) Conference.")). The base case will be $`k = 2`$, so $`n = 2^{3} = 8`$, exposing $`n/2`$ elements to get $`k - 1`$, $`n = 2^{k - 1}`$ total elements.
Basis: Take a hash chain with $`n = 2^{3}`$ total hash elements. Thus, $`n = 2^{k + 1}`$, so $`k = 2`$. Then the setup phase of Jakobsson’s algorithm assigns pebbles to hold hash elements in positions $`2^{1},2^{2}`$ and $`2^{3}`$.
Now, suppose $`{n/2} = 4`$ hash elements are exposed, thus the algorithm in Figure [2](#S3.F2 "Figure 2 ‣ 3.1.2 Jakobsson’s Main. ‣ 3.1 A Review of Jakobsson’s Algorithm ‣ 3 Jakobsson’s Algorithm ‣ An Online Algorithm for Generating Fractal Hash Chains Applied to Digital Chains of Custody11footnote 1 An extended abstract of this paper is to appear in the Intelligence and Security Informatics 2007 (ISI 2007) Conference.") is called 4 times.
By Fact [17](#Thmfact17 "Fact 17 ‣ Appendix: A Proof of Correctness for Jakobsson’s Algorithm ‣ An Online Algorithm for Generating Fractal Hash Chains Applied to Digital Chains of Custody11footnote 1 An extended abstract of this paper is to appear in the Intelligence and Security Informatics 2007 (ISI 2007) Conference.") only one pebble, the pebble associated with the $`2^{k}`$ position will be discarded. This leaves $`{2^{k + 1} - 1} = 2`$ pebbles remaining. One of these pebbles is associated with the root and will not change. For the remaining pebble associated with the $`2^{1}`$ position
<table>
<tbody>
<tr>
<td></td>
<td>P(j,1)𝑃𝑗1\displaystyle P(j,1)</td>
<td>=\displaystyle=</td>
<td>2j+3⋅2j=21+3⋅21=8superscript2𝑗⋅3superscript2𝑗superscript21⋅3superscript218\displaystyle 2^{j}+3\cdot 2^{j}=2^{1}+3\cdot 2^{1}=8</td>
<td></td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td>D(j,1)𝐷𝑗1\displaystyle D(j,1)</td>
<td>=\displaystyle=</td>
<td>2j+2j+1=21+22=6.superscript2𝑗superscript2𝑗1superscript21superscript226\displaystyle 2^{j}+2^{j+1}=2^{1}+2^{2}=6.</td>
<td></td>
</tr>
</tbody>
</table>
Now, $`n = 2^{k} = 4`$ elements remain. In order to put position $`l`$, which is in terms of base $`2^{k + 1}`$ hash chain, in terms of the base $`2^{k}`$ hash chain we have our two remaining pebbles at $`{R_{k = 2}\hspace{0pt}{(6_{g = 3})}} = 2`$, and $`{R_{k = 2}\hspace{0pt}{(8_{g = 3})}} = 4`$, where $`g = {k + 1}`$. Since $`k = 2`$ our remaining pebbles are at positions $`2^{1}`$ and $`2^{2}`$ or, $`2^{1}\hspace{0pt}\ldots\hspace{0pt}2^{k}`$.
Inductive Hypothesis: Suppose that for a hash chain of size $`2^{k + 1}`$ where $`k \geq 2`$ after Jakobsson’s algorithm runs emitting $`{n/2} = 2^{k}`$ hash elements that the remaining pebbles contain hash elements of hash chain positions
<table>
<tbody>
<tr>
<td></td>
<td>21,22,23,…,2k.superscript21superscript22superscript23…superscript2𝑘\displaystyle 2^{1},2^{2},2^{3},\ldots,2^{k}.</td>
<td></td>
</tr>
</tbody>
</table>
Inductive Step: Consider $`n/2`$ hash element exposures for a hash chain with $`n = 2^{k + 1}`$ total elements where $`k \geq 2`$. This step will show that the pebbles will be placed at positions equivalent to $`2^{1},2^{2},\ldots,2^{k}`$.
Again, by Fact [17](#Thmfact17 "Fact 17 ‣ Appendix: A Proof of Correctness for Jakobsson’s Algorithm ‣ An Online Algorithm for Generating Fractal Hash Chains Applied to Digital Chains of Custody11footnote 1 An extended abstract of this paper is to appear in the Intelligence and Security Informatics 2007 (ISI 2007) Conference.") only one pebble, the pebble associated with the $`2^{k}`$ position will be discarded. This leaves $`k - 1`$ pebbles remaining. One of these pebbles is associated with the root and will not change.
By Fact [15](#Thmfact15 "Fact 15 ‣ Appendix: A Proof of Correctness for Jakobsson’s Algorithm ‣ An Online Algorithm for Generating Fractal Hash Chains Applied to Digital Chains of Custody11footnote 1 An extended abstract of this paper is to appear in the Intelligence and Security Informatics 2007 (ISI 2007) Conference.") for all pebbles $`p\hspace{0pt}{\lbrack{\mathbf{j}\hspace{0pt}\overset{k}{\rightarrow}\hspace{0pt}1}\rbrack}`$, it must be that
<table>
<tbody>
<tr>
<td></td>
<td>p[𝐣→𝑘1].𝐏𝐨𝐬𝐢𝐭𝐢𝐨𝐧formulae-sequence𝑝delimited-[]𝐣𝑘→1𝐏𝐨𝐬𝐢𝐭𝐢𝐨𝐧\displaystyle p[{\bf j}\overset{k}{\rightarrow}1].{\bf Position}</td>
<td>=\displaystyle=</td>
<td>p[𝐣→𝑘1].𝐃𝐞𝐬𝐭𝐢𝐧𝐚𝐭𝐢𝐨𝐧.formulae-sequence𝑝delimited-[]𝐣𝑘→1𝐃𝐞𝐬𝐭𝐢𝐧𝐚𝐭𝐢𝐨𝐧\displaystyle p[{\bf j}\overset{k}{\rightarrow}1].{\bf Destination}.</td>
<td></td>
</tr>
</tbody>
</table>
For all remaining pebbles after $`n/2`$ hash exposures,
<table>
<tbody>
<tr>
<td></td>
<td>p[j].𝐏𝐨𝐬𝐢𝐭𝐢𝐨𝐧formulae-sequence𝑝delimited-[]𝑗𝐏𝐨𝐬𝐢𝐭𝐢𝐨𝐧\displaystyle p[j].{\bf Position}</td>
<td>=\displaystyle=</td>
<td>P(j,q),𝑃𝑗𝑞\displaystyle P(j,q),</td>
<td></td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td>p[j].𝐃𝐞𝐬𝐭𝐢𝐧𝐚𝐭𝐢𝐨𝐧formulae-sequence𝑝delimited-[]𝑗𝐃𝐞𝐬𝐭𝐢𝐧𝐚𝐭𝐢𝐨𝐧\displaystyle p[j].{\bf Destination}</td>
<td>=\displaystyle=</td>
<td>D(j,q).𝐷𝑗𝑞\displaystyle D(j,q).</td>
<td></td>
</tr>
</tbody>
</table>
By Corollary [1](#Sx1.EGx69 "Corollary 1 ‣ Appendix: A Proof of Correctness for Jakobsson’s Algorithm ‣ An Online Algorithm for Generating Fractal Hash Chains Applied to Digital Chains of Custody11footnote 1 An extended abstract of this paper is to appear in the Intelligence and Security Informatics 2007 (ISI 2007) Conference.") we know that for all remaining pebbles $`{p\hspace{0pt}{\lbrack j\rbrack}}.{{\mathbf{P}\mathbf{o}\mathbf{s}\mathbf{i}\mathbf{t}\mathbf{i}\mathbf{o}\mathbf{n}} = {p\hspace{0pt}{\lbrack j\rbrack}}}.{\mathbf{D}\mathbf{e}\mathbf{s}\mathbf{t}\mathbf{i}\mathbf{n}\mathbf{a}\mathbf{t}\mathbf{i}\mathbf{o}\mathbf{n}}`$ when $`n/2`$ elements are output and thus all pebbles are inactive.
Further, since we know that $`{p\hspace{0pt}{\lbrack j\rbrack}}.{{\mathbf{P}\mathbf{o}\mathbf{s}\mathbf{i}\mathbf{t}\mathbf{i}\mathbf{o}\mathbf{n}} = {p\hspace{0pt}{\lbrack j\rbrack}}}.{\mathbf{D}\mathbf{e}\mathbf{s}\mathbf{t}\mathbf{i}\mathbf{n}\mathbf{a}\mathbf{t}\mathbf{i}\mathbf{o}\mathbf{n}}`$ the remaining pebble positions are defined by the destination function $`D\hspace{0pt}{(j,m)}`$ as defined earlier where $`m`$ is the number of times the pebble at original position $`j`$ is moved back and
<table>
<tbody>
<tr>
<td></td>
<td>D(j,m)=2j+m⋅2j+1.𝐷𝑗𝑚superscript2𝑗⋅𝑚superscript2𝑗1\displaystyle D(j,m)=2^{j}+m\cdot 2^{j+1}.</td>
<td></td>
</tr>
</tbody>
</table>
By Fact [16](#Thmfact16 "Fact 16 ‣ Appendix: A Proof of Correctness for Jakobsson’s Algorithm ‣ An Online Algorithm for Generating Fractal Hash Chains Applied to Digital Chains of Custody11footnote 1 An extended abstract of this paper is to appear in the Intelligence and Security Informatics 2007 (ISI 2007) Conference.") $`m = 2^{k - j - 1}`$ for each pebble at initial position $`j`$ over the first $`n/2`$ hash element exposures.
Therefore,
<table>
<tbody>
<tr>
<td></td>
<td>D(j,2k−j−1)𝐷𝑗superscript2𝑘𝑗1\displaystyle D(j,2^{k-j-1})</td>
<td>=\displaystyle=</td>
<td>2j+2k−j−1⋅2j+1superscript2𝑗⋅superscript2𝑘𝑗1superscript2𝑗1\displaystyle 2^{j}+2^{k-j-1}\cdot 2^{j+1}</td>
<td></td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td>D(j,2k−j−1)𝐷𝑗superscript2𝑘𝑗1\displaystyle D(j,2^{k-j-1})</td>
<td>=\displaystyle=</td>
<td>2j+2ksuperscript2𝑗superscript2𝑘\displaystyle 2^{j}+2^{k}</td>
<td></td>
</tr>
</tbody>
</table>
given in terms of base $`n = 2^{k + 1}`$ total elements. To put this result in terms of our new hash chain $`n = 2^{k}`$ we simply apply the $`R_{k}\hspace{0pt}{(l_{k + 1})}`$ function to get
<table>
<tbody>
<tr>
<td></td>
<td>Rk(2j+2k)=2j+2k−2k=2j,subscript𝑅𝑘superscript2𝑗superscript2𝑘superscript2𝑗superscript2𝑘superscript2𝑘superscript2𝑗\displaystyle R_{k}(2^{j}+2^{k})=2^{j}+2^{k}-2^{k}=2^{j},</td>
<td></td>
</tr>
</tbody>
</table>
where $`j = {1,\cdots,k}`$.
Thus, by the inductive hypothesis it must be that when Jakobsson’s algorithm runs emitting $`{n/2} = 2^{k}`$ hash elements the remaining pebbles contain hash elements of hash-chain positions $`2^{1},2^{2},2^{3},\ldots,2^{k}`$, thus completing the proof. |
0705.2862 | cs.CR | cyber | <sup>1</sup>
<sup>1</sup>institutetext: Weizmann Institute of Science, Rehovot, Israel
# Cryptanalysis of group-based key agreement protocols using subgroup distance functions
Dima Ruinskiy
Adi Shamir
Boaz Tsaban
###### Abstract
We introduce a new approach for cryptanalysis of key agreement protocols based on noncommutative groups. This approach uses functions that estimate the distance of a group element to a given subgroup. We test it against the Shpilrain-Ushakov protocol, which is based on Thompson’s group $`F`$, and show that it can break about half the keys within a few seconds on a single PC.
Keywords Key agreement, Cryptanalysis, Thompson’s group, Shpilrain-Ushakov, Subgroup distance function
## 1 Introduction
Key agreement protocols have been the subject of extensive studies in the past 30 years. Their main task is to allow two parties (in the sequel, Alice and Bob) to agree on a common secret key over an insecure communication channel. The best known example of such a protocol is the Diffie-Hellman protocol, which uses a (commutative) cyclic group. Over the last few years, there was a lot of interest in key agreement protocols based on noncommutative groups, and much research was dedicated to analyzing these proposals and suggesting alternative ones (see, e.g., \[[1](#bib.bib1), [4](#bib.bib4), [5](#bib.bib5), [6](#bib.bib6), [7](#bib.bib7), [8](#bib.bib8), [10](#bib.bib10), [11](#bib.bib11), [12](#bib.bib12)\], and references therein).
A possible approach for attacking such systems is the *length-based cryptanalysis*, which was outlined in \[[6](#bib.bib6)\]. This approach relies on the existence of a good length function on the underlying group, i.e., a function $`\ell\hspace{0pt}{(g)}`$ that tends to grow as the number of generators multiplied to obtain $`g`$ grows. Examples of groups known to have such length functions are the *braid group $`B_{N}`$* \[[2](#bib.bib2)\] and *Thompson’s group $`F`$* \[[3](#bib.bib3)\]. For these groups, several practical realizations of length-based attacks were demonstrated \[[4](#bib.bib4), [5](#bib.bib5), self\]. These attacks can achieve good success rates, but usually only when we allow the algorithm to explore many suboptimal partial solutions, which greatly increases both the time and space complexities (see \[[5](#bib.bib5)\] for more details).
We introduce a novel approach to cryptanalysis of such key agreement protocols, which relies on the notion of *subgroup distance functions*, i.e., functions that estimate, for an element $`g \in G`$ and a subgroup $`H \leq G`$, the distance from $`g`$ to $`H`$. The motivation for these distance-based attacks is the fact that several families of public key agreement protocols suggest predefined pairs of subgroups of the main group to be used for key generation, and their security depends on the ability of the adversary to generate any elements in these subgroups, which are in some way equivalent to the originals (see \[self, [11](#bib.bib11)\]). We construct the theoretical framework for *distance-based attacks* and demonstrate its applicability using the Shpilrain-Ushakov protocol in Thompson’s group $`F`$ \[[12](#bib.bib12)\] as an example. Although it has recently been shown by Matucci \[[8](#bib.bib8)\] that the implementation of the proposed protocol in $`F`$ can be broken deterministically using a specialized attack based on the structural properties of the group, it is still an interesting test case for more generic attacks, such as the one proposed here.
The paper is organized as follows: in Section [2](#S2 "2 The Shpilrain-Ushakov Key agreement Protocol ‣ Cryptanalysis of group-based key agreement protocols using subgroup distance functions") we present the protocol in its general form. We then introduce in Section [3](#S3 "3 Subgroup distance functions ‣ Cryptanalysis of group-based key agreement protocols using subgroup distance functions") the notion of subgroup distance function and a general attack scheme based on it. Section [4](#S4 "4 Thompson’s group ‣ Cryptanalysis of group-based key agreement protocols using subgroup distance functions") describes the setting for the protocol in Thompson’s group $`F`$. In Section [5](#S5 "5 Subgroup distance functions in Thompson’s group ‣ Cryptanalysis of group-based key agreement protocols using subgroup distance functions") we introduce several subgroup distance functions in $`F`$. Section [6](#S6 "6 Experimental results ‣ Cryptanalysis of group-based key agreement protocols using subgroup distance functions") describes our experimental cryptanalytic results.
## 2 The Shpilrain-Ushakov Key agreement Protocol
The protocol below was suggested by Shpilrain and Ushakov in \[[12](#bib.bib12)\]. The authors suggested to use Thompson’s group $`F`$ for its implementation. Before we focus on that example, we’ll discuss the general case.
1. \(0\)
Alice and Bob agree (publicly) on a group $`G`$ and subgroups $`{A,B} \leq G`$, such that $`{a\hspace{0pt}b} = {b\hspace{0pt}a}`$ for each $`a \in A`$ and each $`b \in B`$.
2. 1\.
A public word $`z \in G`$ is selected.
3. 2\.
Alice selects privately at random elements $`a_{1} \in A`$ and $`b_{1} \in B`$, computes $`u_{1} = {a_{1}\hspace{0pt}z\hspace{0pt}b_{1}}`$, and sends $`u_{1}`$ to Bob.
4. 3\.
Bob selects privately at random elements $`a_{2} \in A`$ and $`b_{2} \in B`$, computes $`u_{2} = {b_{2}\hspace{0pt}z\hspace{0pt}a_{2}}`$, and sends $`u_{2}`$ to Alice.
5. 4\.
Alice computes $`K_{A} = {a_{1}\hspace{0pt}u_{2}\hspace{0pt}b_{1}} = {a_{1}\hspace{0pt}b_{2}\hspace{0pt}z\hspace{0pt}a_{2}\hspace{0pt}b_{1}}`$, whereas Bob computes $`K_{B} = {b_{2}\hspace{0pt}u_{1}\hspace{0pt}a_{2}} = {b_{2}\hspace{0pt}a_{1}\hspace{0pt}z\hspace{0pt}b_{1}\hspace{0pt}a_{2}}`$.
As $`{a_{1}\hspace{0pt}b_{2}} = {b_{2}\hspace{0pt}a_{1}}`$ and $`{a_{2}\hspace{0pt}b_{1}} = {b_{1}\hspace{0pt}a_{2}}`$, $`K_{A} = K_{B} = K`$ and so the parties share the same group element, from which a secret key can be derived.
### 2.1 Breaking the protocol
The goal of the adversary is to obtain the secret group element $`K`$ from the publicly known elements $`u_{1}`$, $`u_{2}`$ and $`z`$. For this it suffices to solve the following problem:
######
Definition 1
(Decomposition problem)
Given $`z \in G`$ and $`u = {a\hspace{0pt}z\hspace{0pt}b}`$ where $`a \in A`$ and $`b \in B`$, find some elements $`\overset{\sim}{a} \in A`$ and $`\overset{\sim}{b} \in B`$, such that $`{\overset{\sim}{a}\hspace{0pt}z\hspace{0pt}\overset{\sim}{b}} = {a\hspace{0pt}z\hspace{0pt}b}`$.
Indeed, assume that the attacker, given $`u_{1} = {a_{1}\hspace{0pt}z\hspace{0pt}b_{1}}`$, finds $`{\overset{\sim}{a}}_{1} \in A`$ and $`{\overset{\sim}{b}}_{1} \in B`$, such that $`{{\overset{\sim}{a}}_{1}\hspace{0pt}z\hspace{0pt}{\overset{\sim}{b}}_{1}} = {a_{1}\hspace{0pt}z\hspace{0pt}b_{1}}`$. Then, because $`u_{2} = {b_{2}\hspace{0pt}z\hspace{0pt}a_{2}}`$ is known, the attacker can compute
<table>
<tbody>
<tr>
<td></td>
<td>a~1u2b~1=a~1b2za2b~1=b2a~1zb~1a2=b2u1a2=KB.subscript~𝑎1subscript𝑢2subscript~𝑏1subscript~𝑎1subscript𝑏2𝑧subscript𝑎2subscript~𝑏1subscript𝑏2subscript~𝑎1𝑧subscript~𝑏1subscript𝑎2subscript𝑏2subscript𝑢1subscript𝑎2subscript𝐾𝐵\tilde{a}_{1}u_{2}\tilde{b}_{1}=\tilde{a}_{1}b_{2}za_{2}\tilde{b}_{1}=b_{2}\tilde{a}_{1}z\tilde{b}_{1}a_{2}=b_{2}u_{1}a_{2}=K_{B}\enspace.</td>
<td></td>
</tr>
</tbody>
</table>
Alternatively, the attacker can break the protocol by finding a valid decomposition of $`u_{2} = {b_{2}\hspace{0pt}z\hspace{0pt}a_{2}}`$.
For any given $`\overset{\sim}{a} \in A`$ we can compute its *complement* $`\overset{\sim}{b} = {z^{- 1}\hspace{0pt}{\overset{\sim}{a}}^{- 1}\hspace{0pt}u} = {z^{- 1}\hspace{0pt}{\overset{\sim}{a}}^{- 1}\hspace{0pt}{({a\hspace{0pt}z\hspace{0pt}b})}}`$, which guarantees that $`{\overset{\sim}{a}\hspace{0pt}z\hspace{0pt}\overset{\sim}{b}} = {a\hspace{0pt}z\hspace{0pt}b}`$. The pair $`\overset{\sim}{a},\overset{\sim}{b}`$ is a solution to this problem if, and only if, $`\overset{\sim}{b} \in B`$. A similar comment applies if we start with $`\overset{\sim}{b} \in B`$. This involves being able to solve the *group membership problem*, i.e., to determine whether $`\overset{\sim}{b} \in B`$ (or $`\overset{\sim}{a} \in A`$ in the second case).
It should be stressed that solving the decomposition problem is sufficient, but not necessary in order to cryptanalyze the system. All that is required in practice is finding some pair $`\overset{\sim}{a},\overset{\sim}{b}`$ that succeeds in decrypting the information passed between Alice and Bob. Any pair $`\overset{\sim}{a} \in A`$ and $`\overset{\sim}{b} \in B`$ will work, but there can be other pairs, which are just as good. This observation can be useful in cases where the group membership problem is difficult or in groups where the centralizers of individual elements are considerably larger than the centralizers of the subgroups (which is not the case in $`F`$, see \[[9](#bib.bib9)\]). For simplicity, in the sequel we will restrict ourselves to solutions where $`\overset{\sim}{a} \in A`$ and $`\overset{\sim}{b} \in B`$.
## 3 Subgroup distance functions
######
Definition 2
(Subgroup distance function)
Let $`G`$ be a group, $`H \leq G`$ a subgroup. A function $`d_{H}:{G\rightarrow{\mathbb{R}}^{+}}`$ is a *subgroup distance function* if it satisfies the following two axioms:
1. 1\.
Validity: $`{d_{H}\hspace{0pt}{(h)}} = 0`$ for all $`h \in H`$.
2. 2\.
Non-triviality: $`{d_{H}\hspace{0pt}{(g)}} > 0`$ for all $`g \notin H`$.
It is an *invariant* subgroup distance function if it also satisfies:
1. \(3\)
Invariance: $`{d_{H}\hspace{0pt}{({g\hspace{0pt}h})}} = {d_{H}\hspace{0pt}{({h\hspace{0pt}g})}} = {d_{H}\hspace{0pt}{(g)}}`$ for all $`g \in G`$ and $`h \in H`$.
Clearly, if it is possible to evaluate a subgroup distance function $`d_{H}`$ on all elements of $`G`$, then the membership decision problem for $`H`$ is solvable: $`{g \in H}\Leftrightarrow{{d_{H}\hspace{0pt}{(g)}} = 0}`$. Conversely, if one can solve the membership decision problem, a trivial distance function can be derived from it, e.g., $`{d_{H}\hspace{0pt}{(g)}} = {1 - {\chi_{H}\hspace{0pt}{(g)}}}`$, where $`\chi_{H}`$ is the characteristic function of $`H`$.
Obviously, this trivial distance function is not a good example. For the subgroup distance function to be useful, it has to somehow measure how close a given element $`g`$ is to $`H`$, that is, if $`{d_{H}\hspace{0pt}{(g_{1})}} < {d_{H}\hspace{0pt}{(g_{2})}}`$, then $`g_{1}`$ is closer to $`H`$ than $`g_{2}`$. This concept of “closeness” can be hard to define, and even harder to evaluate. The notion of what’s considered a good distance function may vary, depending on the subgroups and on the presentation. In the sequel we will discuss concrete examples of subgroup distance function in Thompson’s group $`F`$.
Assuming the existence of such functions, consider the following algorithm for solving the decomposition problem:
######
Algorithm 1
(Subgroup distance attack)
We are given words $`{z,{x\hspace{0pt}z\hspace{0pt}y}} \in G`$, where $`x \in X`$ and $`y \in Y`$, $`X,Y`$ are commuting subgroups of $`G`$ and $`S_{X},S_{Y}`$ are their respective (finite) generating sets. The goal it to find some $`\overset{\sim}{x} \in X`$ and $`\overset{\sim}{y} \in Y`$, such that $`{x\hspace{0pt}z\hspace{0pt}y} = {\overset{\sim}{x}\hspace{0pt}z\hspace{0pt}\overset{\sim}{y}}`$. The algorithm runs at most a predefined number of iterations $`N`$.
1. 1\.
Let $`\overset{\sim}{x}\leftarrow 1`$.
2. 2\.
For each $`g_{i} \in S_{X}^{\pm 1}`$ compute $`x_{i} = {\overset{\sim}{x}\hspace{0pt}g_{i}}`$, its complement $`y_{i} = {z^{- 1}\hspace{0pt}x_{i}^{- 1}\hspace{0pt}x\hspace{0pt}z\hspace{0pt}y}`$ and evaluate $`d_{Y}\hspace{0pt}{(y_{i})}`$. If $`{d_{Y}\hspace{0pt}{(y_{i})}} = 0`$, let $`\overset{\sim}{x} = x_{i}`$, $`\overset{\sim}{y} = y_{i}`$ and halt.
3. 3\.
Let $`j`$ be the index of the minimum $`d_{Y}\hspace{0pt}{(y_{i})}`$ (if several such $`j`$ are possible, choose one arbitrarily).
4. 4\.
If the maximal number of iterations $`N`$ has been reached, terminate. Otherwise, let $`\overset{\sim}{x}\leftarrow x_{j}`$ and return to step [2](#S3.I3.i2 "item 2 ‣ Algorithm 1 (Subgroup distance attack) ‣ 3 Subgroup distance functions ‣ Cryptanalysis of group-based key agreement protocols using subgroup distance functions").
Observe that if the algorithm halts in step [2](#S3.I3.i2 "item 2 ‣ Algorithm 1 (Subgroup distance attack) ‣ 3 Subgroup distance functions ‣ Cryptanalysis of group-based key agreement protocols using subgroup distance functions"), then the pair $`\overset{\sim}{x},\overset{\sim}{y}`$ is a solution of the decomposition problem.
Algorithm [1](#Thmalgo1 "Algorithm 1 (Subgroup distance attack) ‣ 3 Subgroup distance functions ‣ Cryptanalysis of group-based key agreement protocols using subgroup distance functions") is very similar to the length-based attacks described in \[[4](#bib.bib4), self\]. The difference is that it uses the subgroup distance function, instead of the length function to evaluate the quality of candidates. As such, any extensions applicable to the length-based algorithms (such as memory, lookahead, etc.) can be used with the distance-based attack as well. Refer to \[[5](#bib.bib5), self\] for more information.
### 3.1 Attacking the Shpilrain-Ushakov protocol
The adversary is given the common word $`z`$ and the public elements $`u_{1},u_{2}`$. These can be translated into four equations in the group:
<table>
<tbody>
<tr>
<td></td>
<td>u1=a1zb1u2=b2za2u1−1=b1−1z−1a1−1u2−1=a2−1z−1b2−1subscript𝑢1subscript𝑎1𝑧subscript𝑏1subscript𝑢2subscript𝑏2𝑧subscript𝑎2superscriptsubscript𝑢11superscriptsubscript𝑏11superscript𝑧1superscriptsubscript𝑎11superscriptsubscript𝑢21superscriptsubscript𝑎21superscript𝑧1superscriptsubscript𝑏21\begin{array}[]{rcl}u_{1}&=&a_{1}zb_{1}\\ u_{2}&=&b_{2}za_{2}\\ u_{1}^{-1}&=&b_{1}^{-1}z^{-1}a_{1}^{-1}\\ u_{2}^{-1}&=&a_{2}^{-1}z^{-1}b_{2}^{-1}\end{array}</td>
<td></td>
<td>(1)</td>
</tr>
</tbody>
</table>
Algorithm [1](#Thmalgo1 "Algorithm 1 (Subgroup distance attack) ‣ 3 Subgroup distance functions ‣ Cryptanalysis of group-based key agreement protocols using subgroup distance functions") (with or without possible extensions) can be applied to each of the four equations separately, thus attacking each of the four private elements $`a_{1},a_{2},b_{1}^{- 1},b_{2}^{- 1}`$. A single success out of the four attempts is sufficient to break the cryptosystem (see Section [2.1](#S2.SS1 "2.1 Breaking the protocol ‣ 2 The Shpilrain-Ushakov Key agreement Protocol ‣ Cryptanalysis of group-based key agreement protocols using subgroup distance functions")).
## 4 Thompson’s group
Thompson’s group $`F`$ is the infinite noncommutative group defined by the following generators and relations:
<table>
<tbody>
<tr>
<td></td>
<td>F=⟨x0,x1,x2,…|xi−1xkxi=xk+1(k>i)⟩𝐹delimited-⟨⟩subscript𝑥0subscript𝑥1subscript𝑥2conditional…superscriptsubscript𝑥𝑖1subscript𝑥𝑘subscript𝑥𝑖subscript𝑥𝑘1𝑘𝑖F=\langle\ x_{0},x_{1},x_{2},\dots\ |\ x_{i}^{-1}x_{k}x_{i}=x_{k+1}\ (k>i)\ \rangle</td>
<td></td>
<td>(2)</td>
</tr>
</tbody>
</table>
######
Remark 1
From Equation ([2](#S4.E2 "In 4 Thompson’s group ‣ Cryptanalysis of group-based key agreement protocols using subgroup distance functions")) it’s evident that the elements $`x_{0},x_{1}`$ and their inverses generate the entire group, because $`x_{k}^{\pm 1} = {x_{0}^{1 - k}\hspace{0pt}x_{1}^{\pm 1}\hspace{0pt}x_{0}^{k - 1}}`$ for every $`k \geq 2`$.
######
Definition 3
A basic generator $`x_{i}^{\pm 1}`$ of $`F`$ is called a *letter*. A generator $`x_{i}`$ is a *positive letter*. An inverse $`x_{i}^{- 1}`$ is a *negative letter*. A *word* in $`F`$ is a sequence of letters. We define $`|w|`$ as the length of the word $`w`$ , i.e., the number of letters in it.
######
Definition 4
A word $`w \in F`$ is said to be in *normal form*, if
<table>
<tbody>
<tr>
<td></td>
<td>w=xi1⋯xirxjt−1⋯xj1−1𝑤subscript𝑥subscript𝑖1⋯subscript𝑥subscript𝑖𝑟superscriptsubscript𝑥subscript𝑗𝑡1⋯superscriptsubscript𝑥subscript𝑗11w=x_{i_{1}}\cdots x_{i_{r}}x_{j_{t}}^{-1}\cdots x_{j_{1}}^{-1}</td>
<td></td>
<td>(3)</td>
</tr>
</tbody>
</table>
and the following two conditions hold:
1. (NF1)
$`i_{1} \leq \cdots \leq i_{r}`$ and $`j_{1} \leq \cdots \leq j_{t}`$
2. (NF2)
If both $`x_{i},x_{i}^{- 1}`$ occur in $`w`$, then at least one of $`x_{i + 1},x_{i + 1}^{- 1}`$ occurs too.
A word is said to be in *seminormal form* if only (NF1) holds.
While a seminormal form is not necessarily unique, a normal form is, i.e., two words represent the same group element if and only if they have the same normal form \[[3](#bib.bib3)\]. The following rewriting rules can be used to convert any word to its seminormal form \[[12](#bib.bib12)\]:
For all non-negative integers $`i < k`$:
<table>
<tbody>
<tr>
<td></td>
<td>(R1)xkxi→xixk+1(R2)xk−1xi→xixk+1−1(R3)xi−1xk→xk+1xi−1(R4)xi−1xk−1→xk+1−1xi−1𝑅1subscript𝑥𝑘subscript𝑥𝑖→subscript𝑥𝑖subscript𝑥𝑘1𝑅2superscriptsubscript𝑥𝑘1subscript𝑥𝑖→subscript𝑥𝑖superscriptsubscript𝑥𝑘11𝑅3superscriptsubscript𝑥𝑖1subscript𝑥𝑘→subscript𝑥𝑘1superscriptsubscript𝑥𝑖1𝑅4superscriptsubscript𝑥𝑖1superscriptsubscript𝑥𝑘1→superscriptsubscript𝑥𝑘11superscriptsubscript𝑥𝑖1\begin{array}[]{llcl}(R1)&x_{k}x_{i}&\rightarrow&x_{i}x_{k+1}\\ (R2)&x_{k}^{-1}x_{i}&\rightarrow&x_{i}x_{k+1}^{-1}\\ (R3)&x_{i}^{-1}x_{k}&\rightarrow&x_{k+1}x_{i}^{-1}\\ (R4)&x_{i}^{-1}x_{k}^{-1}&\rightarrow&x_{k+1}^{-1}x_{i}^{-1}\\ \end{array}</td>
<td></td>
</tr>
</tbody>
</table>
For all non-negative integers $`i`$:
<table>
<tbody>
<tr>
<td></td>
<td>(R5)xi−1xi→1𝑅5superscriptsubscript𝑥𝑖1subscript𝑥𝑖→1\begin{array}[]{llcl}(R5)&x_{i}^{-1}x_{i}\phantom{{}^{-1}}&\rightarrow&1\phantom{x_{i}x_{k+1}^{-1}}\\ \end{array}</td>
<td></td>
</tr>
</tbody>
</table>
The seminormal form can be subsequently converted to a normal form by searching for pairs of indices violating (NF2), starting from the boundary between the positive and negative parts, and applying the inverses of rewriting rules (R1) and (R4) to eliminate these pairs \[[12](#bib.bib12)\]:
Suppose that $`(x_{i_{a}},x_{j_{b}}^{- 1})`$ is a pair of letters violating (NF2) and that $`a`$ and $`b`$ are maximal with this property (i.e., there exists no violating pair $`(x_{i_{k}},x_{j_{l}}^{- 1})`$ with $`k > a`$ and $`l > b`$). Then $`i_{a} = j_{b}`$ and all indices in $`x_{i_{a + 1}}\hspace{0pt}\cdots\hspace{0pt}x_{i_{r}}\hspace{0pt}x_{j_{t}}^{- 1}\hspace{0pt}\cdots\hspace{0pt}x_{j_{b + 1}}^{- 1}`$ are higher than $`i_{a} + 1`$ (by definition of (NF2)). Applying the inverse of (R1) to $`x_{i_{a}}`$ and the inverse of (R4) to $`x_{j_{b}}^{- 1}`$ we get:
<table>
<tbody>
<tr>
<td></td>
<td>w=xi1⋯xia(xia+1⋯xirxjt−1⋯xjb+1−1)⏟cxjb−1⋯xj1→xi1⋯xia+1−1⋯xir−1(xiaxjb−1)⏟cancelxjt−1−1⋯xjb+1−1−1⋯xj1→xi1⋯xia−1(xia+1−1⋯xir−1xjt−1−1⋯xjb+1−1−1)⏟c′xjb−1⋯xj1𝑤subscript𝑥subscript𝑖1⋯subscript𝑥subscript𝑖𝑎subscript⏟subscript𝑥subscript𝑖𝑎1⋯subscript𝑥subscript𝑖𝑟superscriptsubscript𝑥subscript𝑗𝑡1⋯superscriptsubscript𝑥subscript𝑗𝑏11𝑐superscriptsubscript𝑥subscript𝑗𝑏1⋯subscript𝑥subscript𝑗1missing-subexpression→subscript𝑥subscript𝑖1⋯subscript𝑥subscript𝑖𝑎11⋯subscript𝑥subscript𝑖𝑟1subscript⏟subscript𝑥subscript𝑖𝑎superscriptsubscript𝑥subscript𝑗𝑏1𝑐𝑎𝑛𝑐𝑒𝑙superscriptsubscript𝑥subscript𝑗𝑡11⋯superscriptsubscript𝑥subscript𝑗𝑏111⋯subscript𝑥subscript𝑗1missing-subexpression→subscript𝑥subscript𝑖1⋯subscript𝑥subscript𝑖𝑎1subscript⏟subscript𝑥subscript𝑖𝑎11⋯subscript𝑥subscript𝑖𝑟1superscriptsubscript𝑥subscript𝑗𝑡11⋯superscriptsubscript𝑥subscript𝑗𝑏111superscript𝑐′subscript𝑥subscript𝑗𝑏1⋯subscript𝑥subscript𝑗1\begin{array}[]{lcl}w&=&x_{i_{1}}\cdots x_{i_{a}}\underbrace{(x_{i_{a+1}}\cdots x_{i_{r}}x_{j_{t}}^{-1}\cdots x_{j_{b+1}}^{-1})}_{c}x_{j_{b}}^{-1}\cdots x_{j_{1}}\\ &\rightarrow&x_{i_{1}}\cdots x_{i_{a+1}-1}\cdots x_{i_{r}-1}\underbrace{(x_{i_{a}}x_{j_{b}}^{-1})}_{cancel}x_{j_{t}-1}^{-1}\cdots x_{j_{b+1}-1}^{-1}\cdots x_{j_{1}}\\ &\rightarrow&x_{i_{1}}\cdots x_{i_{a-1}}\underbrace{(x_{i_{a+1}-1}\cdots x_{i_{r}-1}x_{j_{t}-1}^{-1}\cdots x_{j_{b+1}-1}^{-1})}_{c^{\prime}}x_{j_{b-1}}\cdots x_{j_{1}}\end{array}</td>
<td></td>
</tr>
</tbody>
</table>
The violating pair $`(x_{i_{a}},x_{j_{b}}^{- 1})`$ is cancelled and the subword $`c'`$ obtained from $`c`$ by index shifting contains no violating pairs (by the assumption of maximality on $`(a,b)`$). Thus, we can continue searching for bad pairs, starting from $`a - 1`$ and $`b - 1`$ down. Thus we are guaranteed to find and remove all the violating pairs and reach the normal form.
######
Definition 5
(Normal form length)
For $`w \in F`$, whose normal form is $`\hat{w}`$, define the *normal form length* as $`{\ell_{\text{NF}}\hspace{0pt}{(w)}} = {|\hat{w}|}`$.
The following lemma shows the effect multiplication by a single letter has on the normal form of the word. This result will be useful in the following sections.
######
Lemma 1
Let $`w \in F`$ and $`x = x_{t}^{\pm 1}`$ be a basic generator of $`F`$ in the presentation ([2](#S4.E2 "In 4 Thompson’s group ‣ Cryptanalysis of group-based key agreement protocols using subgroup distance functions")). Then $`{\ell_{\text{NF}}\hspace{0pt}{({x\hspace{0pt}w})}} = {{\ell_{\text{NF}}\hspace{0pt}{(w)}} \pm 1}`$ (and due to symmetry, $`\ell_{\text{NF}}{(wx)} = \ell_{\text{NF}}{(w)} \pm 1)`$.
###### Proof
We’ll concentrate on the product $`x\hspace{0pt}w`$ (obviously, the case of $`w\hspace{0pt}x`$ is similar) and observe what happens to the normal form of $`w`$ when it’s multiplied on the left by the letter $`x`$. Without loss of generality, $`w = {x_{i_{1}}\hspace{0pt}\cdots\hspace{0pt}x_{i_{k}}\hspace{0pt}x_{j_{l}}^{- 1}\hspace{0pt}\cdots\hspace{0pt}x_{j_{1}}^{- 1}}`$ is in normal form. Denote the positive and negative parts of $`w`$ by $`w_{p}`$ and $`w_{n}`$ respectively.
Assume that $`x = x_{t}`$ is a positive letter. Then $`b\hspace{0pt}w`$ is converted to a seminormal form by moving $`x`$ into its proper location, while updating its index, using repeated applications of (R1). Assuming $`m`$ applications of (R1) are necessary, the result is of the form:
<table>
<tbody>
<tr>
<td></td>
<td>bw¯=xi1⋯xim𝒙𝒕+𝒎xim+1⋯xikxjl−1⋯xj1−1,¯𝑏𝑤subscript𝑥subscript𝑖1⋯subscript𝑥subscript𝑖𝑚subscript𝒙𝒕𝒎subscript𝑥subscript𝑖𝑚1⋯subscript𝑥subscript𝑖𝑘superscriptsubscript𝑥subscript𝑗𝑙1⋯superscriptsubscript𝑥subscript𝑗11\overline{bw}=x_{i_{1}}\cdots x_{i_{m}}\bm{x_{t+m}}x_{i_{m+1}}\cdots x_{i_{k}}x_{j_{l}}^{-1}\cdots x_{j_{1}}^{-1}\enspace,</td>
<td></td>
</tr>
</tbody>
</table>
where $`i_{m} < {{t + m} - 1}`$ and $`i_{m + 1} \geq {t + m}`$.
######
Remark 2
Observe that it is not possible that $`i_{m} = {{t + m} - 1}`$, because in order to apply (R1): $`{x_{{t + m} - 1}\hspace{0pt}x_{i_{m}}}\rightarrow{x_{i_{m}}\hspace{0pt}x_{t + m}}`$, one must have $`i_{m} < {{t + m} - 1}`$.
######
Example 1
$`w = {x_{3}\hspace{0pt}x_{7}\hspace{0pt}x_{11}\hspace{0pt}x_{9}^{- 1}\hspace{0pt}x_{4}^{- 1}}`$, $`b = x_{8}`$. $`{b\hspace{0pt}w} = {{x_{8} \cdot x_{3}}\hspace{0pt}x_{7}\hspace{0pt}x_{11}\hspace{0pt}x_{9}^{- 1}\hspace{0pt}x_{4}^{- 1}}`$ is converted to $`\overline{b\hspace{0pt}w} = {x_{3}\hspace{0pt}x_{7}\hspace{0pt}\mathbf{x}_{\mathbf{1}\mathbf{0}}\hspace{0pt}x_{11}\hspace{0pt}x_{9}^{- 1}\hspace{0pt}x_{4}^{- 1}}`$, by 2 applications of (R1).
Obviously, $`\overline{b\hspace{0pt}w}`$ is a seminormal form and $`{|\overline{b\hspace{0pt}w}|} = {{|w|} + 1}`$. If $`\overline{b\hspace{0pt}w}`$ is in normal form (as in the above example), we’re done. The only situation where it’s not in normal form, is if it contains pairs violating (NF2). Since $`x_{t + m}`$ is the only letter introduced, the only violating pair can be $`(x_{t + m},x_{t + m}^{- 1})`$. This may occur, if $`w`$ contained $`x_{t + m}^{- 1}`$, but neither $`x_{t + m}`$, nor $`x_{t + m + 1}^{\pm 1}`$.
######
Example 2
$`w = {x_{3}\hspace{0pt}x_{7}\hspace{0pt}x_{11}\hspace{0pt}x_{9}^{- 1}\hspace{0pt}x_{4}^{- 1}}`$
, $`b = x_{7}`$. $`{b\hspace{0pt}w} = {{x_{7} \cdot x_{3}}\hspace{0pt}x_{7}\hspace{0pt}x_{11}\hspace{0pt}x_{9}^{- 1}\hspace{0pt}x_{4}^{- 1}}`$ is converted to $`\overline{b\hspace{0pt}w} = {x_{3}\hspace{0pt}x_{7}\hspace{0pt}\mathbf{x}_{\mathbf{9}}\hspace{0pt}x_{11}\hspace{0pt}x_{9}^{- 1}\hspace{0pt}x_{4}^{- 1}}`$. In this case $`(x_{9},x_{9}^{- 1}`$) violates (NF2). The inverse of (R1) is applied to rewrite $`{x_{9}\hspace{0pt}x_{11}}\rightarrow{x_{10}\hspace{0pt}x_{9}}`$, and $`x_{9}\hspace{0pt}x_{9}^{- 1}`$ are canceled out, yielding the (normal) word $`\hat{b\hspace{0pt}w} = {x_{3}\hspace{0pt}x_{7}\hspace{0pt}x_{10}\hspace{0pt}x_{4}^{- 1}}`$.
Whenever a situation occurs as described above, the pair $`(x_{t + m},x_{t + m}^{- 1})`$ is cancelled, according to the procedure described in Section [4](#S4 "4 Thompson’s group ‣ Cryptanalysis of group-based key agreement protocols using subgroup distance functions"). This causes all indices above $`t + m`$ to be decreased by 1. The resulting word is
<table>
<tbody>
<tr>
<td></td>
<td>bw^=xi1⋯ximxim+1−1⋯xik−1xjl−1−1⋯xjn+1−1−1xjn−1⋯xj1−1,^𝑏𝑤subscript𝑥subscript𝑖1⋯subscript𝑥subscript𝑖𝑚subscript𝑥subscript𝑖𝑚11⋯subscript𝑥subscript𝑖𝑘1superscriptsubscript𝑥subscript𝑗𝑙11⋯superscriptsubscript𝑥subscript𝑗𝑛111superscriptsubscript𝑥subscript𝑗𝑛1⋯superscriptsubscript𝑥subscript𝑗11\widehat{bw}=x_{i_{1}}\cdots x_{i_{m}}x_{i_{m+1}-1}\cdots x_{i_{k}-1}x_{j_{l}-1}^{-1}\cdots x_{j_{n+1}-1}^{-1}x_{j_{n}}^{-1}\cdots x_{j_{1}}^{-1}\enspace,</td>
<td></td>
</tr>
</tbody>
</table>
where $`i_{m} < {{t + m} - 1}`$, $`i_{m + 1} \geq {t + m + 2}`$, $`j_{n} \leq {t + m}`$ and $`j_{n + 1} \geq {t + m + 2}`$. We have $`{|\hat{b\hspace{0pt}w}|} = {{|w|} - 1}`$ and, in fact, $`\hat{b\hspace{0pt}w}`$ is in normal form. Indeed, once the pair $`(x_{t + m},x_{t + m}^{- 1})`$ is cancelled, the only new pair violating (NF2) that can be introduced is $`(x_{{t + m} - 1},x_{{t + m} - 1}^{- 1})`$, but this is not possible, because $`x_{{t + m} - 1}`$ does not appear in $`\hat{b\hspace{0pt}w}`$, due to Remark [2](#Thmremk2 "Remark 2 ‣ Proof ‣ Lemma 1 ‣ 4 Thompson’s group ‣ Cryptanalysis of group-based key agreement protocols using subgroup distance functions"). This completes the proof for positive letters.
Now, consider the case where $`x = x_{t}^{- 1}`$, a negative letter. $`b\hspace{0pt}w`$ is converted to a seminormal form by moving $`x_{t}^{- 1}`$ to the right, while updating its index, using the different rewriting rules. There are two possible outcomes:
\(1\) After $`m`$ applications of (R2) the resulting word is
<table>
<tbody>
<tr>
<td></td>
<td>bw¯=xi1⋯xim𝒙𝒕+𝒎−𝟏xim+1⋯xikxjl−1⋯xj1−1,¯𝑏𝑤subscript𝑥subscript𝑖1⋯subscript𝑥subscript𝑖𝑚superscriptsubscript𝒙𝒕𝒎1subscript𝑥subscript𝑖𝑚1⋯subscript𝑥subscript𝑖𝑘superscriptsubscript𝑥subscript𝑗𝑙1⋯superscriptsubscript𝑥subscript𝑗11\overline{bw}=x_{i_{1}}\cdots x_{i_{m}}\bm{x_{t+m}^{-1}}x_{i_{m+1}}\cdots x_{i_{k}}x_{j_{l}}^{-1}\cdots x_{j_{1}}^{-1}\enspace,</td>
<td></td>
</tr>
</tbody>
</table>
where $`i_{m + 1} = {t + m}`$, and so the pair is cancelled by applying (R5). Now, because $`i_{m} < {{t + m} - 1}`$, the elimination of the pair $`(x_{t + m},x_{t + m}^{- 1})`$ does not introduce pairs that violate (NF2), and so $`\overline{b\hspace{0pt}w}`$ is in normal form and has $`{|\overline{b\hspace{0pt}w}|} = {{|w|} - 1}`$.
######
Example 3
$`w = {x_{3}\hspace{0pt}x_{7}\hspace{0pt}x_{9}^{- 1}\hspace{0pt}x_{4}^{- 1}}`$, $`b = x_{6}^{- 1}`$. $`{b\hspace{0pt}w} = {\mathbf{x}_{\mathbf{6}}^{- \mathbf{1}}\hspace{0pt}x_{3}\hspace{0pt}x_{7}\hspace{0pt}x_{9}^{- 1}\hspace{0pt}x_{4}^{- 1}}`$ is converted to $`x_{3}\hspace{0pt}\mathbf{x}_{\mathbf{7}}^{- \mathbf{1}}\hspace{0pt}\mathbf{x}_{\mathbf{7}}\hspace{0pt}x_{9}^{- 1}`$ and the pair of inverses is cancelled out to obtain $`x_{4}^{- 1}\rightarrow{x_{3}\hspace{0pt}x_{9}^{- 1}\hspace{0pt}x_{4}^{- 1}}`$.
\(2\) $`x_{t}^{- 1}`$ is moved to its proper place among the negative letters, updating its index if necessary. This is completed through $`m`$ applications of (R2), followed by $`k - m`$ applications of (R3) and finally, $`l - n`$ applications of (R4), to obtain
<table>
<tbody>
<tr>
<td></td>
<td>bw¯=xi1⋯xim𝒙𝒕+𝒎−𝟏xim+1+1⋯xik+1xjl+1−1⋯xjn+1+1−1𝒙𝒕+𝒎−𝟏xjn−1⋯xj1−1,¯𝑏𝑤subscript𝑥subscript𝑖1⋯subscript𝑥subscript𝑖𝑚superscriptsubscript𝒙𝒕𝒎1subscript𝑥subscript𝑖𝑚11⋯subscript𝑥subscript𝑖𝑘1superscriptsubscript𝑥subscript𝑗𝑙11⋯superscriptsubscript𝑥subscript𝑗𝑛111superscriptsubscript𝒙𝒕𝒎1superscriptsubscript𝑥subscript𝑗𝑛1⋯superscriptsubscript𝑥subscript𝑗11\overline{bw}=x_{i_{1}}\cdots x_{i_{m}}\bm{x_{t+m}^{-1}}x_{i_{m+1}+1}\cdots x_{i_{k}+1}x_{j_{l}+1}^{-1}\cdots x_{j_{n+1}+1}^{-1}\bm{x_{t+m}^{-1}}x_{j_{n}}^{-1}\cdots x_{j_{1}}^{-1}\enspace,</td>
<td></td>
</tr>
</tbody>
</table>
where $`i_{m} < {{t + m} - 1}`$, $`i_{m + 1} > {t + m}`$, $`j_{n + 1} > {t + m}`$ and $`j_{n} \leq {t + m}`$. Because the letter $`x_{t + m}`$ is not present in $`\overline{b\hspace{0pt}w}`$ (otherwise the previously described situation would occur), the newly introduced letter $`x_{t + m}^{- 1}`$ cannot violate (NF2), and therefore $`\overline{b\hspace{0pt}w}`$ is in fact in normal form and $`{|\overline{b\hspace{0pt}w}|} = {{|w|} + 1}`$.
######
Example 4
$`w = {x_{3}\hspace{0pt}x_{7}\hspace{0pt}x_{9}^{- 1}\hspace{0pt}x_{4}^{- 1}}`$, $`b = x_{5}^{- 1}`$. $`{b\hspace{0pt}w} = {\mathbf{x}_{\mathbf{5}}^{- \mathbf{1}}\hspace{0pt}x_{3}\hspace{0pt}x_{7}\hspace{0pt}x_{9}^{- 1}\hspace{0pt}x_{4}^{- 1}}`$ is rewritten as: $`{\mathbf{x}_{\mathbf{6}}^{- \mathbf{1}}\hspace{0pt}x_{7}\hspace{0pt}x_{9}^{- 1}\hspace{0pt}x_{4}^{- 1}}\rightarrow{x_{3}\hspace{0pt}x_{8}\hspace{0pt}\mathbf{x}_{\mathbf{6}}^{- \mathbf{1}}\hspace{0pt}x_{9}^{- 1}\hspace{0pt}x_{4}^{- 1}}\rightarrow{x_{3}\hspace{0pt}x_{8}\hspace{0pt}x_{10}^{- 1}\hspace{0pt}\mathbf{x}_{\mathbf{6}}^{- \mathbf{1}}\hspace{0pt}x_{4}^{- 1}}`$.
This completes the proof for negative letters.
∎
### 4.1 The Shpilrain-Ushakov protocol in Thompson’s group
For a natural number $`s \geq 2`$ let $`S_{A} = {\{{x_{0}\hspace{0pt}x_{1}^{- 1}},\ldots,{x_{0}\hspace{0pt}x_{s}^{- 1}}\}}`$, $`S_{B} = {\{ x_{s + 1},\ldots,x_{2\hspace{0pt}s}\}}`$ and $`S_{W} = {\{ x_{0},\ldots,x_{s + 2}\}}`$. $`S_{W}`$ generates $`F`$ (see Remark [1](#Thmremk1 "Remark 1 ‣ 4 Thompson’s group ‣ Cryptanalysis of group-based key agreement protocols using subgroup distance functions")). Denote by $`A_{s}`$ and $`B_{s}`$ the subgroups of $`F`$ generated by $`S_{A}`$ and $`S_{B}`$, respectively.
All of the following facts are shown in \[[12](#bib.bib12)\]: $`A_{s}`$ is exactly the set of elements whose normal form is
<table>
<tbody>
<tr>
<td></td>
<td>xi1⋯ximxjm−1⋯xj1−1,subscript𝑥subscript𝑖1⋯subscript𝑥subscript𝑖𝑚superscriptsubscript𝑥subscript𝑗𝑚1⋯superscriptsubscript𝑥subscript𝑗11x_{i_{1}}\cdots x_{i_{m}}x_{j_{m}}^{-1}\cdots x_{j_{1}}^{-1}\enspace,</td>
<td></td>
</tr>
</tbody>
</table>
i.e, has positive and negative parts of the same length $`m`$, and additionally satisfies $`{i_{k} - k} < s`$ and $`{j_{k} - k} < s`$ for every $`k = {1,\ldots,m}`$. $`B_{s}`$ is the set of all elements of $`F`$ whose normal form consists only of letters with indices $`\geq {s + 1}`$. Additionally, $`A_{s}`$ and $`B_{s}`$ commute elementwise, which makes them usable for implementing the protocol in Section [2](#S2 "2 The Shpilrain-Ushakov Key agreement Protocol ‣ Cryptanalysis of group-based key agreement protocols using subgroup distance functions").
#### Key generation
Let $`s \geq 2`$ and $`L`$ be positive integers. The words $`{a_{1},a_{2}} \in A_{s}`$, $`{b_{1},b_{2}} \in B_{s}`$, and $`w \in F`$ are all chosen of normal form length $`L`$, as follows: Let $`X`$ be $`A`$, $`B`$, or $`W`$. Start with the empty word, and multiply it on the right by a generator (or inverse) selected uniformly at random from the set $`S_{X}`$. Continue this procedure until the normal form of the word has length $`L`$.
For practical and (hopefully) secure implementation of the protocol, it is suggested in \[[12](#bib.bib12)\] to use $`s \in {\{ 3,4,\ldots,8\}}`$ and $`L \in {\{ 256,258,\ldots,320\}}`$.
## 5 Subgroup distance functions in Thompson’s group
In this section we’ll suggest several natural distance functions from the subgroups $`{A_{s},B_{s}} \leq F`$ defined in Section [4.1](#S4.SS1 "4.1 The Shpilrain-Ushakov protocol in Thompson’s group ‣ 4 Thompson’s group ‣ Cryptanalysis of group-based key agreement protocols using subgroup distance functions"). These distance functions can be used to implement the attack outlined by Algorithm [1](#Thmalgo1 "Algorithm 1 (Subgroup distance attack) ‣ 3 Subgroup distance functions ‣ Cryptanalysis of group-based key agreement protocols using subgroup distance functions").
### 5.1 Distance functions from $`B_{s}`$
For $`w \in F`$ define $`P_{i}\hspace{0pt}{(w)}`$ and $`N_{i}\hspace{0pt}{(w)}`$ as the number of occurrences of $`x_{i}`$ and $`x_{i}^{- 1}`$ in the normal form $`\hat{w}`$ of $`w`$.
######
Definition 6
(Distance from $`B_{s}`$)
Let $`s \leq 2`$ be an integer. For $`w \in F`$ the distance from $`B_{s}`$ is defined as
<table>
<tbody>
<tr>
<td></td>
<td>dBs(w)=∑i=0s(Pi(w)+Ni(w))subscript𝑑subscript𝐵𝑠𝑤superscriptsubscript𝑖0𝑠subscript𝑃𝑖𝑤subscript𝑁𝑖𝑤d_{B_{s}}(w)=\sum_{i=0}^{s}{\left(P_{i}(w)+N_{i}(w)\right)}</td>
<td></td>
</tr>
</tbody>
</table>
######
Claim 1
$`d_{B_{s}}`$ is a distance function.
###### Proof
This is immediate, since an element is in $`B_{s}`$ if and only if its normal form does not contain generators with indices below $`s + 1`$ (see Section [4.1](#S4.SS1 "4.1 The Shpilrain-Ushakov protocol in Thompson’s group ‣ 4 Thompson’s group ‣ Cryptanalysis of group-based key agreement protocols using subgroup distance functions")). ∎
######
Claim 2
$`d_{B_{s}}`$ is an invariant distance function.
###### Proof
It is enough to consider only the generators of $`B_{s}`$. Indeed, if multiplication by a single generator of $`B_{s}`$ does not change the distance of a word $`w`$, neither does multiplication by a sequence of these generators.
Let $`w \in F`$. Let $`b = x_{s + \alpha}^{\pm 1}`$, where $`\alpha > 0`$. By Lemma [1](#Thmlemm1 "Lemma 1 ‣ 4 Thompson’s group ‣ Cryptanalysis of group-based key agreement protocols using subgroup distance functions"), we know that $`b`$ is either moved to its proper position (and $`{\ell_{\text{NF}}\hspace{0pt}{({b\hspace{0pt}w})}} = {{\ell_{\text{NF}}\hspace{0pt}{(w)}} + 1}`$) or it is cancelled with its inverse, either by (R5) or as part of a pair violating (NF2), in which case $`{\ell_{\text{NF}}\hspace{0pt}{({b\hspace{0pt}w})}} = {{\ell_{\text{NF}}\hspace{0pt}{(w)}} - 1}`$. The index of $`b`$ is initially above $`s`$, and may only increase when the rewriting rules are applied. Therefore, if $`b`$ is cancelled at some point, the index of its inverse is also above $`s`$. Furthermore, when pairs of elements are rewritten, the lower-indexed element is not affected, so any letters with indices $`\leq s`$ will not be affected by moving $`b`$. Finally, if $`b`$ is cancelled out due to violating (NF2), the process again only affects letters with indices higher than $`b`$’s (see the proof of Lemma [1](#Thmlemm1 "Lemma 1 ‣ 4 Thompson’s group ‣ Cryptanalysis of group-based key agreement protocols using subgroup distance functions")). In all cases, the generators with indices $`\leq s`$ are not affected at all, and so $`{d_{B_{s}}\hspace{0pt}{({b\hspace{0pt}w})}} = {d_{B_{s}}\hspace{0pt}{(w)}}`$.
∎
One can intuitively feel that $`d_{B_{s}}`$ is a natural distance function, because it counts the number of “bad” letters in $`w`$ (letters that do not belong to the subgroup $`B_{s}`$). Indeed, if $`w`$ is in normal form, $`w = {w_{p}\hspace{0pt}w_{c}\hspace{0pt}w_{n}}`$, where $`w_{p}`$ and $`w_{n}`$ are the “bad” positive and negative subwords, respectively, then $`{d_{B_{s}}\hspace{0pt}{(w)}} = {{|w_{p}|} + {|w_{n}|}}`$ and $`{w_{p}^{- 1}\hspace{0pt}w\hspace{0pt}w_{n}^{- 1}} \in B`$.
We now introduce another natural function that measures distance from $`B_{s}`$.
######
Definition 7
(Weighted distance from $`B_{s}`$)
Let $`s \leq 2`$ be an integer. For $`w \in F`$ the weighted distance from $`B_{s}`$ is defined as
<table>
<tbody>
<tr>
<td></td>
<td>dBs¯(w)=∑i=0s(s+1−i)(Pi(w^)+Ni(w^))¯subscript𝑑subscript𝐵𝑠𝑤superscriptsubscript𝑖0𝑠𝑠1𝑖subscript𝑃𝑖^𝑤subscript𝑁𝑖^𝑤\overline{d_{B_{s}}}(w)=\sum_{i=0}^{s}{(s+1-i)\left(P_{i}(\hat{w})+N_{i}(\hat{w})\right)}</td>
<td></td>
</tr>
</tbody>
</table>
$`\overline{d_{B_{s}}}`$ does not only count the “bad” letters, but assigns a score for each letter, depending on how far below $`s + 1`$ it is (in particular, $`{d_{B_{s}}\hspace{0pt}{(w)}} \leq {\overline{d_{B_{s}}}\hspace{0pt}{(w)}}`$ for all $`w \in F`$. The following claim is straightforward.
######
Claim 3
$`\overline{d_{B_{s}}}`$ is an invariant distance function.
###### Proof
The proof of Claim [2](#Thmclmm2 "Claim 2 ‣ 5.1 Distance functions from 𝐵_𝑠 ‣ 5 Subgroup distance functions in Thompson’s group ‣ Cryptanalysis of group-based key agreement protocols using subgroup distance functions") shows that multiplication by $`b`$ does not alter any letters below $`s + 1`$ in $`w`$. Therefore, the weight of each such letter is also preserved. ∎
### 5.2 Distance functions from $`A_{s}`$
We will now describe a number of natural distance functions from the subgroup $`A_{s}`$. Recall (Section [4.1](#S4.SS1 "4.1 The Shpilrain-Ushakov protocol in Thompson’s group ‣ 4 Thompson’s group ‣ Cryptanalysis of group-based key agreement protocols using subgroup distance functions")) that $`A_{s}`$ is the set of all elements in $`F`$, whose normal form is of the type $`x_{i_{1}}\hspace{0pt}\cdots\hspace{0pt}x_{i_{m}}\hspace{0pt}x_{j_{m}}^{- 1}\hspace{0pt}\cdots\hspace{0pt}x_{j_{1}}^{- 1}`$, i.e, has positive and negative parts of the same length $`m`$, and additionally satisfies $`{i_{k} - k} < s`$ and $`{j_{k} - k} < s`$ for every $`{k = {1,\ldots,m}}.`$
######
Definition 8
(Distance from $`A_{s}`$)
Let $`s \geq 2`$ be an integer. Let $`w \in F`$, such that its normal form is $`\hat{w} = {x_{i_{1}}\hspace{0pt}\cdots\hspace{0pt}x_{i_{p}}\hspace{0pt}x_{j_{n}}^{- 1}\hspace{0pt}\cdots\hspace{0pt}x_{j_{1}}^{- 1}}`$. The distance from $`A_{s}`$ is defined as
<table>
<tbody>
<tr>
<td></td>
<td>dAs(w)=|{k:ik−k≥s}|+|{l:jl−l≥s}|+|p−n|subscript𝑑subscript𝐴𝑠𝑤conditional-set𝑘subscript𝑖𝑘𝑘𝑠conditional-set𝑙subscript𝑗𝑙𝑙𝑠𝑝𝑛d_{A_{s}}(w)=\left|\{k:i_{k}-k\geq s\}\right|\ +\ \left|\{l:j_{l}-l\geq s\}\right|\ +\ \left|p-n\right|</td>
<td></td>
</tr>
</tbody>
</table>
$`d_{A_{s}}\hspace{0pt}{(w)}`$ is the number of “bad” letters in $`\hat{w}`$, i.e., letters that violate the $`A_{s}`$ property, plus the difference between the lengths of the positive or negative parts. $`d_{A_{s}}`$ is clearly a distance function. However, it is not invariant, as shown by the following example:
Similarly we can define a weighted distance function from $`A_{s}`$, which not only counts the number of bad letters, but gives a score to each such letter, based on the difference $`i_{k} - k`$ (or $`j_{k} - k`$).
######
Definition 9
(Weighted distance from $`A_{s}`$)
Let $`s \geq 2`$ be an integer. Let $`w \in F`$, such that its normal form is $`\hat{w} = {x_{i_{1}}\hspace{0pt}\cdots\hspace{0pt}x_{i_{p}}\hspace{0pt}x_{j_{n}}^{- 1}\hspace{0pt}\cdots\hspace{0pt}x_{j_{1}}^{- 1}}`$. The weighted distance from $`A_{s}`$ is defined as
<table>
<tbody>
<tr>
<td></td>
<td>dAs¯(w)=∑k=1…pik−k≥s(ik−k−s+1)+∑k=1…njk−k≥s(jk−k−s+1)+|p−n|¯subscript𝑑subscript𝐴𝑠𝑤superscriptsubscript𝑘1…𝑝subscript𝑖𝑘𝑘𝑠subscript𝑖𝑘𝑘𝑠1superscriptsubscript𝑘1…𝑛subscript𝑗𝑘𝑘𝑠subscript𝑗𝑘𝑘𝑠1𝑝𝑛\overline{d_{A_{s}}}(w)=\sum_{k=1\dots p}^{i_{k}-k\geq s}{(i_{k}-k-s+1)}\ +\sum_{k=1\dots n}^{j_{k}-k\geq s}{(j_{k}-k-s+1)}\ +\ \left|p-n\right|</td>
<td></td>
</tr>
</tbody>
</table>
For each bad letter $`x_{i_{k}}`$ or $`x_{j_{k}}^{- 1}`$, $`\overline{d_{A_{s}}}`$ adds a positive integer. As such, it’s a distance function, which is again not invariant (the example above works here too).
A somewhat different approach to defining distance from $`A_{s}`$ arises from the observation that the number of bad letters can be less important than the maximum value of the differences $`i_{k} - k`$ and $`j_{k} - k`$ across the word, which measures the size of the violation. The difference between the two distance functions roughly corresponds to the difference between the $`L_{1}`$ and $`L_{\infty}`$ norms.
Let $`\hat{w} = {x_{i_{1}}\hspace{0pt}\cdots\hspace{0pt}x_{i_{p}}\hspace{0pt}x_{j_{n}}^{- 1}\hspace{0pt}\cdots\hspace{0pt}x_{j_{1}}^{- 1}}`$. Suppose that for some integer $`k`$ we have $`{{i_{k} - k - s} + 1} = m_{p} > 0`$ and that $`m_{p}`$ is the maximum for all $`i_{k}`$. By multiplying the word by $`x_{0}^{m_{p}}`$ we shift the position for all the original positive letters of $`w`$ by $`m_{p}`$, and so all of the positive letters, including the first $`m`$ $`x_{0}`$’s have $`{i_{k} - k} < s`$. Similarly, if $`m_{n}`$ is the maximum violation in the negative subword, multiplication by $`x_{0}^{- m_{n}}`$ on the right eliminates all violations among negative letters. However, this still does not mean that the word is in $`A_{s}`$, because the positive and negative lengths may differ. Let $`\hat{w^{\prime}}`$ be the normal form obtained from $`\hat{w}`$ through multiplication by $`x_{0}^{m_{p}}`$ and $`x_{0}^{- m_{n}}`$ on the left and right, respectively. Let $`l_{p}`$ and $`l_{n}`$ be the corresponding lengths of the positive and negative parts of $`\hat{w^{\prime}}`$. If $`{l_{p} - l_{n}} > 0`$, then $`{\hat{w^{\prime}}\hspace{0pt}x_{0}^{l_{n} - l_{p}}} \in A_{s}`$. If $`{l_{p} - l_{n}} < 0`$, then $`{x_{0}^{l_{n} - l_{p}}\hspace{0pt}\hat{w^{\prime}}} \in A_{s}`$. Altogether, any word can be changed to a word in $`A_{s}`$ through multiplication by $`m_{p} + m_{n} + {|{l_{p} + l_{n}}|}`$ indices (when $`l_{p}`$ and $`l_{n}`$ are evaluated *after* multiplying by $`x_{0}^{m_{p}}`$ and $`x_{0}^{- m_{n}}`$).
This observation suggests the following distance function:
######
Definition 10
(Maximum-based distance from $`A_{s}`$)
Let $`s \geq 2`$ be an integer. Let $`w \in F`$, such that its normal form is $`\hat{w} = {x_{i_{1}}\hspace{0pt}\cdots\hspace{0pt}x_{i_{p}}\hspace{0pt}x_{j_{n}}^{- 1}\hspace{0pt}\cdots\hspace{0pt}x_{j_{1}}^{- 1}}`$. Let
<table>
<tbody>
<tr>
<td></td>
<td>mp=max({0}∪{ik−k−s+1:k=1…p})subscript𝑚𝑝0conditional-setsubscript𝑖𝑘𝑘𝑠1𝑘1…𝑝m_{p}=\max{\left(\{0\}\cup\{i_{k}-k-s+1:k=1\dots p\}\right)}</td>
<td></td>
</tr>
</tbody>
</table>
and
<table>
<tbody>
<tr>
<td></td>
<td>mn=max({0}∪{jk−k−s+1:k=1…n}).subscript𝑚𝑛0conditional-setsubscript𝑗𝑘𝑘𝑠1𝑘1…𝑛m_{n}=\max{\left(\{0\}\cup\{j_{k}-k-s+1:k=1\dots n\}\right)}\enspace.</td>
<td></td>
</tr>
</tbody>
</table>
The maximum-based distance from $`A_{s}`$ is defined as
<table>
<tbody>
<tr>
<td></td>
<td>dAsm(w)=mp+mn+|(p+mp)−(n+mn)|subscriptsuperscript𝑑𝑚subscript𝐴𝑠𝑤subscript𝑚𝑝subscript𝑚𝑛𝑝subscript𝑚𝑝𝑛subscript𝑚𝑛d^{m}_{A_{s}}(w)=m_{p}+m_{n}+\left|(p+m_{p})-(n+m_{n})\right|</td>
<td></td>
</tr>
</tbody>
</table>
For every $`w \in A_{s}`$ $`m_{p}`$, $`m_{n}`$ and $`|{p - n}|`$ are $`0`$ by definition, while for every $`w \notin A_{s}`$ at least one of them has to be positive, so the $`d_{A_{s}}^{m}`$ is a distance function. It turns out that, unlike the two previously defined distance functions, $`d_{A_{s}}^{m}`$ is also invariant.
######
Claim 4
$`d_{A_{s}}^{m}`$ is an invariant distance function.
###### Proof
As with Claim [2](#Thmclmm2 "Claim 2 ‣ 5.1 Distance functions from 𝐵_𝑠 ‣ 5 Subgroup distance functions in Thompson’s group ‣ Cryptanalysis of group-based key agreement protocols using subgroup distance functions"), it’s sufficient to prove that multiplication by a single generator of $`A_{s}`$ does not change the distance from any word $`w`$ to $`A_{s}`$. We will consider multiplications on the left by generators and their inverses. The multiplication on the right follows symmetrically.
Let $`w = {x_{i_{1}}\hspace{0pt}\cdots\hspace{0pt}x_{i_{p}}\hspace{0pt}x_{j_{n}}^{- 1}\hspace{0pt}\cdots\hspace{0pt}x_{j_{1}}^{- 1}}`$, without loss of generality, in normal form. Consider the generator $`x_{0}\hspace{0pt}x_{t}^{- 1}`$, where $`1 \leq t \leq s`$. Define $`w'`$ as the normal form of $`x_{0}\hspace{0pt}x_{t}^{- 1}\hspace{0pt}w`$. For the parameters $`p,n,m_{p},m_{n}`$ of $`w`$, denote by $`p',n',m_{p}',m_{n}'`$ their corresponding values in $`w'`$.
From Lemma [1](#Thmlemm1 "Lemma 1 ‣ 4 Thompson’s group ‣ Cryptanalysis of group-based key agreement protocols using subgroup distance functions") it follows that each of the letters $`x_{t}^{- 1}`$ and $`x_{0}`$ can either be cancelled out with the appropriate inverse, decreasing the length by 1, or placed in its appropriate location, increasing the length by 1. There is a total of 4 possible options:
\(1\) $`x_{t}^{- 1}`$ is cancelled out, but $`x_{0}`$ is not: $`w' = {x_{0}\hspace{0pt}x_{i_{1}}\hspace{0pt}\cdots\hspace{0pt}x_{i_{m}}\hspace{0pt}x_{i_{m + 2}}\hspace{0pt}\cdots\hspace{0pt}x_{i_{p}}\hspace{0pt}x_{j_{n}}^{- 1}\hspace{0pt}\cdots\hspace{0pt}x_{j_{1}}^{- 1}}`$, where $`x_{t + m}^{- 1}`$ is cancelled out with $`x_{i_{m + 1}}`$ after $`m`$ applications of (R2). It follows that $`p' = p`$, $`n' = n`$ and $`m_{n}' = m_{n}`$ (because the negative letters are unaffected). Observe also that there can be no bad letters among the first $`m`$: indeed, (R2) is applied $`m`$ times, for each $`k = {1\hspace{0pt}\ldots\hspace{0pt}m}`$ rewriting $`{x_{{t + k} - 1}^{- 1}\hspace{0pt}x_{i_{k}}}\rightarrow{x_{i_{k}}\hspace{0pt}x_{t + k}^{- 1}}`$, so necessarily $`i_{k} < {{t + k} - 1}`$ for all $`k`$, or equivalently, $`{i_{k} - k} < {t - 1} < s`$. The multiplication by $`x_{0}`$ on the left only increases their relative positions, thus decreasing $`i_{k} - k`$. Now, any possible bad letters above $`i_{m}`$ are unchanged, and neither is their relative position, so $`m_{p}' = m_{p}`$ and overall $`{d_{A_{s}}^{m}\hspace{0pt}{(w')}} = {d_{A_{s}}^{m}\hspace{0pt}{(w)}}`$.
\(2\) Both $`x_{t}^{- 1}`$ and $`x_{0}`$ are cancelled out: $`w' = {x_{i_{1} - 1}\hspace{0pt}\cdots\hspace{0pt}x_{i_{m} - 1}\hspace{0pt}x_{i_{m + 2} - 1}\hspace{0pt}\cdots\hspace{0pt}x_{i_{p} - 1}\hspace{0pt}x_{j_{n} - 1}^{- 1}\hspace{0pt}\cdots\hspace{0pt}x_{j_{q + 1} - 1}^{- 1}\hspace{0pt}x_{0}^{1 - q}}`$. Here $`p' = {p - 1}`$, $`n' = {n - 1}`$ and $`m_{n}' = m_{n}`$ because all negative letters $`x_{j_{k}}^{- 1}`$ with $`j_{k} > 0`$ had both their indices and their relative positions decreased by 1. The same thing applies to positive letters above $`i_{m}`$, which are the only positive letters that may be bad. So again, $`m_{p}' = m_{p}`$ and $`{d_{A_{s}}^{m}\hspace{0pt}{(w')}} = {d_{A_{s}}^{m}\hspace{0pt}{(w)}}`$.
\(3\) Neither $`x_{t}^{- 1}`$, nor $`x_{0}`$ are cancelled out: $`w' = {x_{0}\hspace{0pt}x_{i_{1}}\hspace{0pt}\cdots\hspace{0pt}x_{i_{m}}\hspace{0pt}x_{i_{m + 1} + 1}\hspace{0pt}\cdots\hspace{0pt}x_{i_{p} + 1}\hspace{0pt}x_{j_{n} + 1}^{- 1}\hspace{0pt}\cdots\hspace{0pt}x_{j_{q + 1} + 1}^{- 1}\hspace{0pt}{\mathbf{x}}_{{\mathbf{t}} + {\mathbf{m}}}^{- \mathbf{1}}\hspace{0pt}x_{j_{q}}^{- 1}\hspace{0pt}\cdots\hspace{0pt}x_{j_{1}}^{- 1}}`$. Here $`p' = {p + 1}`$ and $`n' = {n + 1}`$. Due to the former observation, bad positive letters may only exist beyond the first $`m`$. All these letters had their indices $`i_{k}`$ and their relative positions $`k`$ increased by 1, so the difference is preserved and $`m_{p}' = m_{p}`$. Among the negative letters, only the letters whose indices increased, also had their relative position increased, so $`j_{k} - k`$ is preserved for all the original letters of $`w`$. Hence, $`m_{n}' \geq m_{n}`$ and the only situation when it may actually increase is when the new maximum is attained at the new letter, i.e., $`m_{n}' = {{{({t + m})} - {({q + 1})} - s} + 1} > m_{n}`$. Because $`t \leq s`$, $`m \leq p`$ and $`q \leq n`$, we have $`m_{n}' \leq {p - q}`$, from which it follows that
<table>
<tbody>
<tr>
<td></td>
<td>(p′+mp′)−(n′+mn′)=(p′−n′)+(mp′−mn′)=(p+1)−(n+1)+mp−mn′≥superscript𝑝′superscriptsubscript𝑚𝑝′superscript𝑛′superscriptsubscript𝑚𝑛′superscript𝑝′superscript𝑛′superscriptsubscript𝑚𝑝′superscriptsubscript𝑚𝑛′𝑝1𝑛1subscript𝑚𝑝superscriptsubscript𝑚𝑛′absent(p^{\prime}+m_{p}^{\prime})-(n^{\prime}+m_{n}^{\prime})=(p^{\prime}-n^{\prime})+(m_{p}^{\prime}-m_{n}^{\prime})=(p+1)-(n+1)+m_{p}-m_{n}^{\prime}\geq</td>
<td></td>
</tr>
</tbody>
</table>
<table>
<tbody>
<tr>
<td></td>
<td>≥mp+(p−n)−(p−q)=mp+q−n≥0absentsubscript𝑚𝑝𝑝𝑛𝑝𝑞subscript𝑚𝑝𝑞𝑛0\geq m_{p}+(p-n)-(p-q)=m_{p}+q-n\geq 0</td>
<td></td>
</tr>
</tbody>
</table>
Assuming $`m_{n}' > m_{n}`$, it’s obvious that
<table>
<tbody>
<tr>
<td></td>
<td>(p−n)+(mp−mn)>(p′−n′)+(mp′−mn′)≥0,𝑝𝑛subscript𝑚𝑝subscript𝑚𝑛superscript𝑝′superscript𝑛′superscriptsubscript𝑚𝑝′superscriptsubscript𝑚𝑛′0(p-n)+(m_{p}-m_{n})>(p^{\prime}-n^{\prime})+(m_{p}^{\prime}-m_{n}^{\prime})\geq 0\enspace,</td>
<td></td>
</tr>
</tbody>
</table>
and so if $`m_{n}`$ increases, $`\left| {{({p + m_{p}})} - {({n + m_{n}})}} \right|`$ decreases by the same amount, and overall $`{d_{A_{s}}^{m}\hspace{0pt}{(w')}} = {d_{A_{s}}^{m}\hspace{0pt}{(w)}}`$.
\(4\) $`x_{t}^{- 1}`$ is not cancelled out, but $`x_{0}`$ is: $`w' = {x_{i_{1} - 1}\hspace{0pt}\cdots\hspace{0pt}x_{i_{m} - 1}\hspace{0pt}x_{i_{m + 1}}\hspace{0pt}\cdots\hspace{0pt}x_{i_{p}}\hspace{0pt}x_{j_{n}}^{- 1}\hspace{0pt}\cdots\hspace{0pt}x_{j_{q + 1}}^{- 1}\hspace{0pt}{\mathbf{x}}_{{{\mathbf{t}} + {\mathbf{m}}} - \mathbf{1}}^{- \mathbf{1}}\hspace{0pt}x_{j_{q} - 1}^{- 1}\hspace{0pt}\cdots\hspace{0pt}x_{j_{r + 1} - 1}^{- 1}\hspace{0pt}x_{0}^{1 - r}}`$, where $`p' = p`$, $`n' = n`$, $`m_{p}' = m_{p}`$ (because the first $`m`$ positive letters, whose indices have changed, contained no bad letters), and $`m_{n}'`$ again may only increase, if it’s attained at $`x_{{t + m} - 1}^{- 1}`$. Repeating the same calculations shows that $`{d_{A_{s}}^{m}\hspace{0pt}{(w')}} = {d_{A_{s}}^{m}\hspace{0pt}{(w)}}`$ in this case too.
Now consider the inverse $`x_{t}\hspace{0pt}x_{0}^{- 1}`$ and denote $`w' = {x_{t}\hspace{0pt}x_{0}^{- 1}\hspace{0pt}w}`$. The four possible outcomes are:
\(1\) $`x_{0}^{- 1}`$ is cancelled out, but $`x_{t}`$ is not: $`x_{0}^{- 1}`$ can only be cancelled out if $`i_{1} = 0`$, and the resulting word is: $`w' = {x_{i_{2}}\hspace{0pt}\cdots\hspace{0pt}x_{i_{m}}\hspace{0pt}{\mathbf{x}}_{{{\mathbf{t}} + {\mathbf{m}}} - \mathbf{1}}\hspace{0pt}x_{i_{m + 1}}\hspace{0pt}\cdots\hspace{0pt}x_{i_{p}}\hspace{0pt}x_{j_{n}}^{- 1}\hspace{0pt}\cdots\hspace{0pt}x_{j_{1}}^{- 1}}`$. Here $`p' = p`$, $`n' = n`$, $`m_{n}' = m_{n}`$ (negative part is not affected) and $`m_{p}' = m_{p}`$ because the letters $`x_{i_{2}}`$ to $`x_{i_{m}}`$ cannot be bad and the relative position of other positive letters has not changed.
\(2\) Both $`x_{0}^{- 1}`$ and $`x_{t}`$ are cancelled out: Assuming $`x_{t}`$ is cancelled out (due to violation of (NF2)) with $`x_{j_{q}}^{- 1}`$, $`w' = {x_{i_{2}}\hspace{0pt}\cdots\hspace{0pt}x_{i_{m}}\hspace{0pt}x_{i_{m + 1} - 1}\hspace{0pt}\cdots\hspace{0pt}x_{i_{p} - 1}\hspace{0pt}x_{j_{n} - 1}^{- 1}\hspace{0pt}\cdots\hspace{0pt}x_{j_{q + 1} - 1}^{- 1}\hspace{0pt}x_{j_{q - 1}}^{- 1}\hspace{0pt}\cdots\hspace{0pt}x_{j_{1}}^{- 1}}`$. Here $`p' = {p - 1}`$, $`n' = {n - 1}`$, $`m_{p}' = m_{p}`$, because $`x_{i_{2}}`$ to $`x_{i_{m}}`$ cannot be bad and the relative position of other positive letters has not changed, and $`m_{n}' = m_{n}`$, because the letters whose positions shifted also had their indices decreased.
\(3\) Neither $`x_{0}^{- 1}`$, nor $`x_{t}`$ are cancelled out. $`w' = {x_{i_{1} + 2}\hspace{0pt}\cdots\hspace{0pt}x_{i_{m} + 2}\hspace{0pt}{\mathbf{x}}_{{\mathbf{t}} + {\mathbf{m}}}\hspace{0pt}x_{i_{m + 1} + 1}\hspace{0pt}\cdots\hspace{0pt}x_{i_{p} + 1}\hspace{0pt}x_{j_{n} + 1}^{- 1}\hspace{0pt}\cdots\hspace{0pt}x_{j_{q} + 1}^{- 1}\hspace{0pt}{\mathbf{x}}_{\mathbf{0}}^{- {\mathbf{q}}}}`$. Here $`p' = {p + 1}`$, $`n' = {n + 1}`$, $`m_{p}' = m_{p}`$, because indices above $`i_{m}`$ grew by 1, as did their positions, and indices $`i_{1},\ldots,i_{m}`$ cannot be bad, and also $`m_{n}' = m_{n}`$, because all letters whose indices increased ($`j_{q}`$ and above) shifted in position accordingly.
\(4\) $`x_{0}^{- 1}`$ is not cancelled out, but $`x_{t}`$ is: $`w' = {x_{i_{1} + 2}\hspace{0pt}\cdots\hspace{0pt}x_{i_{m} + 2}\hspace{0pt}x_{i_{m + 1}}\hspace{0pt}\cdots\hspace{0pt}x_{i_{p}}\hspace{0pt}x_{j_{n}}^{- 1}\hspace{0pt}\cdots\hspace{0pt}x_{j_{q + 1}}^{- 1}\hspace{0pt}x_{j_{q - 1} + 1}^{- 1}\hspace{0pt}\cdots\hspace{0pt}x_{j_{r} + 1}^{- 1}\hspace{0pt}{\mathbf{x}}_{\mathbf{0}}^{- {\mathbf{r}}}}`$, the cancelled pair being $`(x_{t + m},x_{j_{q}}^{- 1})`$, where $`j_{q} = {t + m}`$. In this case, any positive letters that can be bad kept their indices and positions, the negative letters $`j_{r + 1},\ldots,j_{q - 1}`$ had their indices and positions shifted, while the letters $`j_{q + 1},\ldots,j_{n}`$ kept their indices and positions. So $`m_{p}' = m_{p}`$ and $`m_{n}' = m_{n}`$ and obviously $`p' = p`$ and $`n' = n`$.
We see that in all the possible cases, $`{d_{A_{s}}^{m}\hspace{0pt}{(w')}} = {d_{A_{s}}^{m}\hspace{0pt}{(w)}}`$. This completes the proof. ∎
## 6 Experimental results
To test the applicability of the subgroup distance functions to cryptanalysis, we tested Algorithm [1](#Thmalgo1 "Algorithm 1 (Subgroup distance attack) ‣ 3 Subgroup distance functions ‣ Cryptanalysis of group-based key agreement protocols using subgroup distance functions") against the Shpilrain-Ushakov protocol in the settings of Thompson’s group. Initially, each of the five distance functions presented in the previous section was tested separately: we generated a public element $`a\hspace{0pt}z\hspace{0pt}b`$ and tried to recover a single private element $`a`$ or $`b`$ from it. For the recovery of $`a`$, the functions $`d_{B_{s}}`$ and $`\overline{d_{B_{s}}}`$ were used to assess the quality of the complements. Similarly, for the recovery of $`b`$, we tried $`d_{A_{s}}`$, $`\overline{d_{A_{s}}}`$ and $`d_{A_{s}}^{m}`$.
For each distance function, the experiment was run at least 1000 times, each time with new, randomly generated keys, with the minimum recommended parameters of $`{s = 3},{L = 256}`$. The bound $`N = {2\hspace{0pt}L}`$ was chosen on the number of iterations, since preliminary experiments have shown that the success rates do not increase beyond that. The results are summarized in Table [1](#S6.T1 "Table 1 ‣ 6 Experimental results ‣ Cryptanalysis of group-based key agreement protocols using subgroup distance functions"). It can be seen that the distance functions $`d_{B_{s}}`$ and $`d_{A_{s}}^{m}`$ noticeably outperform the other distance functions, in recovering $`a`$ and $`b`$, respectively. The fact that $`d_{A_{s}}^{m}`$ clearly outperforms its counterparts suggests that the notion of invariance may be useful for assessing the suitability of a given distance function.
Table 1: Success rates for the different subgroup distance functions
<table>
<thead>
<tr>
<th></th>
<th>dBssubscript𝑑subscript𝐵𝑠d_{B_{s}}</th>
<th>dBs¯¯subscript𝑑subscript𝐵𝑠\phantom{\Big{(}}\overline{d_{B_{s}}}</th>
<th>dAssubscript𝑑subscript𝐴𝑠d_{A_{s}}</th>
<th>dAs¯¯subscript𝑑subscript𝐴𝑠\phantom{\Big{(}}\overline{d_{A_{s}}}</th>
<th>dAsmsubscriptsuperscript𝑑𝑚subscript𝐴𝑠d^{m}_{A_{s}}</th>
</tr>
</thead>
<tbody>
<tr>
<th>Recovery probability</th>
<th>11.7%</th>
<th>3.4%</th>
<td>3.7%</td>
<td>3.4%</td>
<td>23.3%</td>
</tr>
</tbody>
</table>
Preliminary experiments have shown that, regardless of the settings, the success probability of finding $`a_{1}`$ given $`a_{1}\hspace{0pt}z\hspace{0pt}b_{1}`$ is similar to that of finding $`a_{2}^{- 1}`$ given $`a_{2}^{- 1}\hspace{0pt}z^{- 1}\hspace{0pt}b_{2}^{- 1}`$. A similar assertion holds for $`b_{2}`$ and $`b_{1}^{- 1}`$. Therefore, in order to estimate the overall success rate against an actual instance of the cryptosystem, it’s sufficient to try to recover one of the four $`a`$’s and $`b`$’s. If we denote by $`p_{a}`$ and $`p_{b}`$ the probability of successfully recovering $`a`$ and $`b`$, respectively, and assume that all probabilities are independent, then, the expected total success rate is roughly $`1 - {{({1 - p_{a}})}^{2}\hspace{0pt}{({1 - p_{b}})}^{2}}`$ (because each instance of the protocol contains two elements of type $`a`$ and two of type $`b`$).
When the success rates of the two best distance functions, $`d_{B_{s}}`$ for $`a`$ and $`d_{A_{s}}^{m}`$ for $`b`$, are combined, the expected overall success probability, according to the above, is between 50% and 54%, which was experimentally verified. Note that this attack is very efficient, since it involves no backtracking, no lookahead, and no analysis of suboptimal partial results: it tries to peel off the generators by a greedy algorithm, which considers only locally optimal steps. Attacking each key required only a few seconds on a single PC, and it is very surprising that such a simple attack succeeds about half the time. These results are much better than those achieved by length-based attacks of similar complexity on this cryptosystem (see \[[9](#bib.bib9)\]).
It is interesting to note that possible extensions of the attack, such as memorizing many suboptimal partial solutions or using significant lookahead (which require much higher time and space complexities) have different effects on length-based and distance-based attacks. While it was shown in \[[9](#bib.bib9)\] that these extensions greatly improve the success rates of the length-based attack, experiments with the distance-based attack, with similar values of the memory and lookahead parameters, showed almost no improvement. However, the situation may be very different for other cryptosystems and other subgroup distance functions.
To further test the performance of the distance functions, several experiments were run with different values of the parameters $`(s,L)`$. We used the combination of $`d_{B_{s}}`$ and $`d_{A_{s}}^{m}`$, which was established as the best in the former experiment. Table [2](#S6.T2 "Table 2 ‣ 6 Experimental results ‣ Cryptanalysis of group-based key agreement protocols using subgroup distance functions") shows the overall success probability, for $`L \in {\{ 128,256,320,512,640,960\}}`$ and $`s \in {\{ 3,5,8\}}`$. The success rates stay remarkably consistent across different lengths for a given $`s`$, and even increasing $`s`$ does not cause a significant drop. The time complexity of the attack grows linearly with $`s`$ and roughly quadratically with $`L`$, with most of the time being spent on computing normal forms of elements in the group. For the largest parameters presented here, the attack still required under a minute in most cases. This suggests that for the Shpilrain-Ushakov cryptosystem the distance-based attack remains a viable threat, even when the security parameters $`s`$ and $`L`$ are increased beyond the original recommendations.
Table 2: Success rates for different combinations of $`(s,L)`$
<table>
<thead>
<tr>
<th></th>
<th>L=128𝐿128L=128</th>
<th>L=256𝐿256L=256</th>
<th>L=320𝐿320L=320</th>
<th>L=512𝐿512L=512</th>
<th>L=640𝐿640L=640</th>
<th>L=960𝐿960L=960</th>
</tr>
</thead>
<tbody>
<tr>
<th>s=3𝑠3s=3</th>
<td>51.7%</td>
<td>47.9%</td>
<td>55.5%</td>
<td>51.2%</td>
<td>50.4%</td>
<td>52.6%</td>
</tr>
<tr>
<th>s=5𝑠5s=5</th>
<td>46.0%</td>
<td>47.1%</td>
<td>48.4%</td>
<td>51.1%</td>
<td>48.2%</td>
<td>48.3%</td>
</tr>
<tr>
<th>s=8𝑠8s=8</th>
<td>36.2%</td>
<td>42.8%</td>
<td>41.3%</td>
<td>46.5%</td>
<td>42.4%</td>
<td>50.3%</td>
</tr>
</tbody>
</table>
## 7 Conclusion
We introduced a novel form of heuristic attacks on public key cryptosystems that are based on combinatorial group theory, using functions that estimate the distance of group elements to a given subgroup. Our results demonstrate that these distance-based attacks can achieve significantly better success rates than previously suggested length-based attacks of similar complexity, and thus they are a potential threat to any cryptosystem based on equations in a noncommutative group, which takes its elements from specific subgroups. It will be interesting to test this approach for other groups and other protocols.
## References
- \[1\]
I. Anshel, M. Anshel and D. Goldfeld, *An algebraic method for public-key cryptography*, Mathematical Research Letters 6 (1999), 287–291.
- \[2\]
E. Artin, *Theory of Braids*, Annals of Mathematics 48 (1947), 127–136.
- \[3\]
J.W. Cannon, W.J. Floyd and W.R. Parry, *Introductory notes on Richard Thompson’s groups*, L’Enseignement Mathematique (2) 42 (1996), 215–256.
- \[4\]
D. Garber, S. Kaplan, M. Teicher, B. Tsaban, and U. Vishne, *Length-based conjugacy search in the Braid group*, Contemporary Mathematics 418 (2006), 75–87.
- \[5\]
D. Garber, S. Kaplan, M. Teicher, B. Tsaban, and U. Vishne, *Probabilistic solutions of equations in the braid group*, Advances in Applied Mathematics 35 (2005), 323–334.
- \[6\] J. Hughes and A. Tannenbaum, *Length-based attacks for certain group based encryption rewriting systems*, Workshop SECI02 Sécurité de la Communication sur Internet (2002).
- \[7\]
K.H. Ko, S.J. Lee, J.H. Cheon, J.W. Han, J. Kang and C. Park, *New Public-Key Cryptosystem Using Braid Groups*, Lecture Notes in Computer Science 1880 (2000), 166–183.
- \[8\]
F. Matucci, *The Shpilrain-Ushakov Protocol for Thompson’s Group $`F`$ is always breakable*, e-print arxiv.org/math/0607184 (2006).
- \[9\]
D. Ruinskiy, A. Shamir, and B. Tsaban, *Length-based cryptanalysis: The case of Thompson’s Group*, Journal of Mathematical Cryptology 1 (2007), 359–372.
- \[10\]
V. Shpilrain, *Assessing security of some group based cryptosystems*, Contemporary Mathematics 360 (2004), 167–177.
- \[11\]
V. Shpilrain and A. Ushakov, *The conjugacy search problem in public key cryptography: unnecessary and insufficient*, Applicable Algebra in Engineering, Communication and Computing 17 (2006), 291–302.
- \[12\]
V. Shpilrain and A. Ushakov, *Thompson’s group and public key cryptography*, ACNS 2005, Lecture Notes in Computer Science 3531 (2005), 151–164. |
0706.1563 | cs.NI | cyber | INSTITUT NATIONAL DE RECHERCHE EN INFORMATIQUE ET EN AUTOMATIQUE
Optimal Choice of Threshold in
Two Level Processor Sharing
Konstantin Avrachenkov — Patrick Brown — Natalia Osipova N° 6215
June 2007
Optimal Choice of Threshold in
Two Level Processor Sharing
Konstantin Avrachenkov
<sup>†</sup>
<sup>†</sup>thanks: INRIA Sophia Antipolis, France, e-mail: K.Avrachenkov@sophia.inria.fr
, Patrick Brown
<sup>†</sup>
<sup>†</sup>thanks: France Telecom R&D, France, e-mail: Patrick.Brown@orange-ftgroup.com
, Natalia Osipova
<sup>†</sup>
<sup>†</sup>thanks: INRIA Sophia Antipolis, France, e-mail: Natalia.Osipova@sophia.inria.fr
<sup>†</sup>
<sup>†</sup>thanks: The work was supported by France Telecom R&D Grant “Modélisation et Gestion du Trafic Réseaux Internet” no. 42937433.
Thème COM — Systèmes communicants
Projet MAESTRO
Rapport de recherche n° 6215 — June 2007 — ?? pages
Abstract: We analyze the Two Level Processor Sharing (TLPS) scheduling discipline with the hyper-exponential job size distribution and with the Poisson arrival process. TLPS is a convenient model to study the benefit of the file size based differentiation in TCP/IP networks. In the case of the hyper-exponential job size distribution with two phases, we find a closed form analytic expression for the expected sojourn time and an approximation for the optimal value of the threshold that minimizes the expected sojourn time. In the case of the hyper-exponential job size distribution with more than two phases, we derive a tight upper bound for the expected sojourn time conditioned on the job size. We show that when the variance of the job size distribution increases, the gain in system performance increases and the sensitivity to the choice of the threshold near its optimal value decreases.
Key-words: Two Level Processor sharing, Hyper-Exponential distribution, Laplace transform.
Le Choix du Seuil Optimal pour La File d’Attente Munie d’une Politique à Temps Partagé Avec Deux-Niveaux
Résumé : Nous étudions la file d’attente munie d’une politique a Temps Partagé avec Deux-Niveaux "Two Level Processor Sharing" avec distribution hyper- exponentielle des temps de service et avec le processus d’arrivée Poisson. TLPS est un modèle commode pour ordonner l’accès aux ressources en fonction de la taille dans un réseau TCP/IP. Dans le cas où la distribution du temps de service est une distribution hyper- exponentielle avec deux phases, nous trouvons une expression analytique pour le temps de réponse moyen. Aussi nous trouvons une approximation de valeur de seuil optimal qui réduit au minimum le temps de réponse moyen. Dans le cas où la distribution du temps de service a plus que deux phases, nous trouvons une borne supérieure pour la fonction de temps de réponse moyen qui est conditionnée au temps de service. Nous montrons que quand la variance de la distribution des temps de service augmente, le gain dans l’exécution du système est considérable et il n’y a pas de sensibilité au choix du seuil sous optimal.
Mots-clés : La file d’attente munie d’une politique à Temps Partagées avec Deux-Niveaux, réseau TCP/IP, distribution hyper-exponentielle, Laplace transforment.
## 1 Introduction
It has been known for a long time that a clever scheduling of tasks can significantly improve system performance. For instance, Shortest Remaining Processing Time (SRPT) scheduling discipline minimizes the expected sojourn time \[[15](#bib.bib15)\]. However, SRPT requires to keep track of all jobs in the system and also requires the knowledge of the remaining processing times. These requirements are often not feasible in applications. The examples of such applications are file size based differentiation in TCP/IP networks \[[3](#bib.bib3), [9](#bib.bib9)\] and Web server request differentiation \[[10](#bib.bib10), [11](#bib.bib11)\].
The Two Level Processor Sharing (TLPS) scheduling discipline \[[12](#bib.bib12)\] helps to overcome the above mentioned requirements. It uses the differentiation of jobs according to a threshold on the attained service and gives priority to the jobs with small sizes. A detail description of the TLPS discipline is presented in the ensuing section. Of course, TLPS provides a sub-optimal mechanism in comparison with SRPT. Nevertheless, as was shown in \[[1](#bib.bib1)\], when the job size distribution has a decreasing hazard rate, the performance of TLPS with appropriate choice of threshold is very close to optimal. It turns out that the distribution of file sizes in the Internet indeed has a decreasing hazard rate and often could be modeled with a heavy-tailed distributions. It is known, that the heavy-tailed distribution could be approximated with a hyper-exponential distribution with a significant number of phases \[[5](#bib.bib5), [8](#bib.bib8)\]. Also in \[[7](#bib.bib7)\], it was shown that the hyper-exponential distribution models well the file size distribution in the Internet. Therefore, in the present work we analyze the TLPS system with hyper-exponential job size distribution.
The paper organization and main results are as follows. In Section [2](#S2 "2 Model description") we provide the model formulation, main definitions and equations. In Section [3](#S3 "3 Hyper-exponential job size distribution with two phases") we study the TLPS discipline in the case of the hyper-exponential job size distribution with two phases. It is known that the Internet connections belong to two distinct classes with very different sizes of transfer. The first class is composed of short HTTP connections and P2P signaling connections. The second class corresponds to downloads (PDF files, MP3 files, MPEG files, etc.). This fact provides motivation to consider first the hyper-exponential job size distribution with two phases.
We find an analytical expression for the expected sojourn time in the TLPS system. Then, we present the approximation of the optimal threshold which minimizes the expected sojourn time. We show that the approximated value of the threshold tends to the optimal threshold when the second moment of the job size distribution function goes to infinity.
We show that the use of the TLPS scheduling discipline can lead to 45% gain in the expected sojourn time in comparison with the standard Processor Sharing. We also show that the system performance is not too sensitive to the choice of the threshold around its optimal value.
In Section [4](#S4 "4 Hyper-exponential job size distribution with more than two phases") we analyze the TLPS discipline when the job size distribution is hyper-exponential with many phases. We provide an expression of the expected conditional sojourn time as the solution of a system of linear equations. Also we apply an iteration method to find the expression of the expected conditional sojourn time and using the resulting expression obtain an explicit and tight upper bound for the expected sojourn time function. In the experimental results we show that the relative error of the latter upper bound with respect to the expected sojourn time function is 6-7%.
We study the properties of the expected sojourn time function when the parameters of the job size distribution function are selected in a such a way that with the increasing number of phases the variance increases. We show numerically that with the increasing number of phases the relative error of the found upper bound decreases. We also show that when the variance of the job size distribution increases the gain in system performance increases and the sensitivity of the system to the selection of the optimal threshold value decreases.
We put some technical proofs in the Appendix.
## 2 Model description
### 2.1 Main definitions
We study the Two Level Processor Sharing (TLPS) scheduling discipline with the hyper-exponential job size distribution. Let us describe the model in detail.
Jobs arrive to the system according to a Poisson process with rate $`\lambda`$. We measure the job size in time units. Specifically, as the job size we define the time which would be spent by the server to treat the job if there were no other jobs in the system.
Let $`\theta`$ be a given threshold. The jobs in the system that attained a service less than $`\theta`$ are assigned to the high priority queue. If in addition there are jobs with attained service greater than $`\theta`$, such a job is separated into two parts. The first part of size $`\theta`$ is assigned to the high priority queue and the second part of size $`x - \theta`$ waits in the lower priority queue. The low priority queue is served when the high priority queue is empty. Both queues are served according to the Processor Sharing (PS) discipline.
Let us denote the job size distribution by $`F\hspace{0pt}{(x)}`$. By $`{\overline{F}\hspace{0pt}{(x)}} = {1 - {F\hspace{0pt}{(x)}}}`$ we denote the complementary distribution function. The mean job size is given by $`m = {\int_{0}^{\infty}{x\hspace{0pt}{dF}\hspace{0pt}{(x)}}}`$ and the system load is $`\rho = {\lambda\hspace{0pt}m}`$. We assume that the system is stable ($`\rho < 1`$) and is in steady state.
It is known that many important probability distributions associated with network traffic are heavy-tailed. In particular, the file size distribution in the Internet is heavy-tailed.
A distribution function has a heavy tail if $`{{e^{\epsilon\hspace{0pt}x}\hspace{0pt}{({1 - {F\hspace{0pt}{(x)}}})}}\rightarrow{\infty\hspace{0pt}\text{as}\hspace{0pt}x}\rightarrow\infty},{{\forall\epsilon} > 0}`$. The heavy-tailed distributions are not only important and prevalent, but also difficult to analyze. Often it is helpful to have the Laplace transform of the job size distribution. However, there is evidently no convenient analytic expression for the Laplace transforms of the Pareto and Weibull distributions, the most common examples of heavy-tailed distributions. In \[[5](#bib.bib5), [8](#bib.bib8)\] it was shown that it is possible to approximate heavy-tailed distributions by hyper-exponential distribution with a significant number of phases.
A hyper-exponential distribution $`F_{N}\hspace{0pt}{(x)}`$ is a convex combination of $`N`$ exponents, $`1 \leq N \leq \infty`$, namely,
<table>
<tbody>
<tr>
<td></td>
<td>FN(x)=1−∑i=1Npie−μix,μi>0,pi≥0,i=1,…,N,and∑i=1Npi=1.formulae-sequencesubscript𝐹𝑁𝑥1superscriptsubscript𝑖1𝑁subscript𝑝𝑖superscript𝑒subscript𝜇𝑖𝑥formulae-sequencesubscript𝜇𝑖0formulae-sequencesubscript𝑝𝑖0formulae-sequence𝑖1…𝑁andsuperscriptsubscript𝑖1𝑁subscript𝑝𝑖1\displaystyle F_{N}(x)=1-\sum_{i=1}^{N}p_{i}e^{-\mu_{i}x},\quad\mu_{i}>0,\,\,p_{i}\geq 0,\quad i=1,...,N,\quad\mbox{and}\quad\sum_{i=1}^{N}{p_{i}}=1.</td>
<td></td>
<td>(1)</td>
</tr>
</tbody>
</table>
In particular, we can construct a sequence of hyper-exponential distributions such that it converges to a heavy-tailed distribution \[[5](#bib.bib5)\]. For instance, if we select
<table>
<tbody>
<tr>
<td></td>
<td>pi=νiγ1,μi=ηiγ2,i=1,…,N,formulae-sequencesubscript𝑝𝑖𝜈superscript𝑖subscript𝛾1formulae-sequencesubscript𝜇𝑖𝜂superscript𝑖subscript𝛾2𝑖1…𝑁\displaystyle p_{i}=\frac{\nu}{i^{\gamma_{1}}},\quad\mu_{i}={\frac{\eta}{i^{\gamma_{2}}}},\quad i=1,...,N,</td>
<td></td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td>γ1>1,γ1−12<γ2<γ1−1,formulae-sequencesubscript𝛾11subscript𝛾112subscript𝛾2subscript𝛾11\displaystyle\gamma_{1}>1,\quad{\frac{\gamma_{1}-1}{2}}<\gamma_{2}<\gamma_{1}-1,</td>
<td></td>
</tr>
</tbody>
</table>
where $`{{\nu = {1/{\sum_{i = 1,..,N}i^{- \gamma_{1}}}}},{\eta = {{\nu/m}\hspace{0pt}{\sum_{i = {1,\ldots,N}}i^{\gamma_{2} - \gamma_{1}}}}}},`$ then the first moment of the job size distribution is finite, but the second moment is infinite when $`N\rightarrow\infty`$. Namely, the first and the second moments $`m`$ and $`d`$ for the hyper-exponential distribution are given by:
<table>
<tbody>
<tr>
<td></td>
<td>m=∫0∞x𝑑F(x)=∑i=1Npiμi,d=∫0∞x2𝑑F(x)=2∑i=1Npiμi2.formulae-sequence𝑚superscriptsubscript0𝑥differential-d𝐹𝑥superscriptsubscript𝑖1𝑁subscript𝑝𝑖subscript𝜇𝑖𝑑superscriptsubscript0superscript𝑥2differential-d𝐹𝑥2superscriptsubscript𝑖1𝑁subscript𝑝𝑖superscriptsubscript𝜇𝑖2\displaystyle m=\int_{0}^{\infty}{x\,dF(x)}=\sum_{i=1}^{N}{{\frac{p_{i}}{\mu_{i}}}},\quad d={\int_{0}^{\infty}{x^{2}\,dF(x)}}=2\sum_{i=1}^{N}{\frac{p_{i}}{\mu_{i}^{2}}}.</td>
<td></td>
<td>(2)</td>
</tr>
</tbody>
</table>
Let us denote
<table>
<tbody>
<tr>
<td></td>
<td>Fθi¯=pie−μiθ,i=1,…,N.formulae-sequence¯superscriptsubscript𝐹𝜃𝑖subscript𝑝𝑖superscript𝑒subscript𝜇𝑖𝜃𝑖1…𝑁\displaystyle\overline{F_{\theta}^{i}}=p_{i}e^{-\mu_{i}\theta},\quad i=1,...,N.</td>
<td></td>
<td>(3)</td>
</tr>
</tbody>
</table>
We note that $`{\sum_{i}\overline{F_{\theta}^{i}}} = {\overline{F}\hspace{0pt}{(\theta)}}`$. The hyper-exponential distribution has a simple Laplace transform:
<table>
<tbody>
<tr>
<td></td>
<td>LF¯(x)(s)=∑i=1Npiμis+μi.subscript𝐿¯𝐹𝑥𝑠superscriptsubscript𝑖1𝑁subscript𝑝𝑖subscript𝜇𝑖𝑠subscript𝜇𝑖\displaystyle L_{\overline{F}(x)}(s)=\sum_{i=1}^{N}{\frac{p_{i}\mu_{i}}{s+\mu_{i}}}.</td>
<td></td>
</tr>
</tbody>
</table>
We would like to note that the hyper-exponential distribution has a decreasing hazard rate. In \[[1](#bib.bib1)\] it was shown, that when a job size distribution has a decreasing hazard rate, then with the selection of the threshold the expected sojourn time of the TLPS system could be reduced in comparison to standard PS system.
Thus, in our work we use hyper-exponential distributions to represent job size distribution functions. In particular, the application of the hyper-exponential job size distribution with two phases is motivated by the fact that in the Internet connections belong to two distinct classes with very different sizes of transfer. The first class is composed of short HTTP connections and P2P signaling connections. The second class corresponds to downloads (PDF files, MP3 files, MPEG files, etc.). So, in the first part of our paper we look at the case of the hyper-exponential job size distribution with two phases and in the second part of the paper we study the case of more than two phases.
### 2.2 The expected sojourn time in TLPS system
Let us denote by $`{\overline{T}}^{T\hspace{0pt}L\hspace{0pt}P\hspace{0pt}S}\hspace{0pt}{(x)}`$ the expected conditional sojourn time in the TLPS system for a job of size $`x`$. Of course, $`{\overline{T}}^{T\hspace{0pt}L\hspace{0pt}P\hspace{0pt}S}\hspace{0pt}{(x)}`$ also depends on $`\theta`$, but for expected conditional sojourn time we only emphasize the dependence on the job size. On the other hand, we denote by $`\overline{T}\hspace{0pt}{(\theta)}`$ the overall expected sojourn time in the TLPS system. Here we emphasize the dependence on $`\theta`$ as later we shall optimize the overall expected sojourn time with respect to the threshold value.
To calculate the expected sojourn time in the TLPS system we need to calculate the time spent by a job of size $`x`$ in the first high priority queue and in the second low priority queue. For the jobs with size $`x \leq \theta`$ the system will behave as the standard PS system where the service time distribution is truncated at $`\theta`$. Let us denote by
<table>
<tbody>
<tr>
<td></td>
<td>Xθn¯=∫0θnyn−1F¯(y)𝑑y¯superscriptsubscript𝑋𝜃𝑛superscriptsubscript0𝜃𝑛superscript𝑦𝑛1¯𝐹𝑦differential-d𝑦\displaystyle{\overline{X_{\theta}^{n}}}=\int_{0}^{\theta}\!ny^{n-1}\overline{F}(y)dy</td>
<td></td>
<td>(4)</td>
</tr>
</tbody>
</table>
the $`n`$-th moment of the distribution truncated at $`\theta`$. In the following sections we will need
<table>
<tbody>
<tr>
<td></td>
<td>Xθ1¯=m−∑i=1NFθi¯μi,Xθ2¯=2∑i=1Npiμi2−2θ(m−∑i=1NFθi¯μi)−2∑i=1NFθi¯μi2.formulae-sequence¯superscriptsubscript𝑋𝜃1𝑚superscriptsubscript𝑖1𝑁¯superscriptsubscript𝐹𝜃𝑖subscript𝜇𝑖¯superscriptsubscript𝑋𝜃22superscriptsubscript𝑖1𝑁subscript𝑝𝑖superscriptsubscript𝜇𝑖22𝜃𝑚superscriptsubscript𝑖1𝑁¯superscriptsubscript𝐹𝜃𝑖subscript𝜇𝑖2superscriptsubscript𝑖1𝑁¯superscriptsubscript𝐹𝜃𝑖superscriptsubscript𝜇𝑖2\displaystyle\overline{X_{\theta}^{1}}=m-{\sum_{i=1}^{N}{\frac{\overline{F_{\theta}^{i}}}{\mu_{i}}}},\quad\overline{X_{\theta}^{2}}=2\sum_{i=1}^{N}{\frac{p_{i}}{\mu_{i}^{2}}}-2\theta\left(m-\sum_{i=1}^{N}{\frac{\overline{F_{\theta}^{i}}}{\mu_{i}}}\right)-2{\sum_{i=1}^{N}{\frac{\overline{F_{\theta}^{i}}}{\mu_{i}^{2}}}}.</td>
<td></td>
<td>(5)</td>
</tr>
</tbody>
</table>
The utilization factor for the truncated distribution is
<table>
<tbody>
<tr>
<td></td>
<td>ρθ=λXθ1¯=ρ−λ∑i=1NFθi¯μi.subscript𝜌𝜃𝜆¯superscriptsubscript𝑋𝜃1𝜌𝜆superscriptsubscript𝑖1𝑁¯superscriptsubscript𝐹𝜃𝑖subscript𝜇𝑖\displaystyle{\rho_{\theta}=\lambda\,{\overline{X_{\theta}^{1}}}}=\rho-\lambda{\sum_{i=1}^{N}{\frac{\overline{F_{\theta}^{i}}}{\mu_{i}}}}.</td>
<td></td>
<td>(6)</td>
</tr>
</tbody>
</table>
Then, the expected conditional response time is given by
<table>
<tbody>
<tr>
<td></td>
<td>T¯TLPS(x)={x1−ρθ,x∈[0,θ],W¯(θ)+θ+α(x−θ)1−ρθ,x∈(θ,∞).superscript¯𝑇𝑇𝐿𝑃𝑆𝑥cases𝑥1subscript𝜌𝜃𝑥0𝜃¯𝑊𝜃𝜃𝛼𝑥𝜃1subscript𝜌𝜃𝑥𝜃\overline{T}^{TLPS}(x)=\left\{\begin{array}[]{l l}{\displaystyle\frac{x}{1-\rho_{\theta}}},&x\in[0,\theta],\\ {\displaystyle\frac{\overline{W}(\theta)+\theta+\alpha(x-\theta)}{1-\rho_{\theta}}},&x\in(\theta,\infty).\end{array}\right.</td>
<td></td>
</tr>
</tbody>
</table>
According to \[[12](#bib.bib12)\], here $`{({{\overline{W}\hspace{0pt}{(\theta)}} + \theta})}/{({1 - \rho_{\theta}})}`$ expresses the time needed to reach the low priority queue. This time consists of the time $`\theta/{({1 - \rho_{\theta}})}`$ spent in the high priority queue, where the flow is served up to the threshold $`\theta`$, plus the time $`{\overline{W}\hspace{0pt}{(\theta)}}/{({1 - \rho_{\theta}})}`$ which is spent waiting for the high priority queue to empty. Here $`{\overline{W}\hspace{0pt}{(\theta)}} = {{\lambda\hspace{0pt}\overline{X_{\theta}^{2}}}/{({2\hspace{0pt}{({1 - \rho_{\theta}})}})}}`$.
The remaining term $`{\alpha\hspace{0pt}{({x - \theta})}}/{({1 - \rho_{\theta}})}`$ is the time spent in the low priority queue. To find $`\alpha\hspace{0pt}{(x)}`$ we can use the interpretation of the lower priority queue as a PS system with batch arrivals \[[4](#bib.bib4), [14](#bib.bib14)\]. As was shown in \[[12](#bib.bib12)\], $`{\alpha'\hspace{0pt}{(x)}} = {{{d\hspace{0pt}\alpha}/d}\hspace{0pt}x}`$ is the solution of the following integral equation
<table>
<tbody>
<tr>
<td></td>
<td>α′(x)=λn¯∫0∞α′(y)B¯(x+y)𝑑y+λn¯∫0xα′(y)B¯(x−y)𝑑y+bB¯(x)+1.superscript𝛼′𝑥𝜆¯𝑛superscriptsubscript0superscript𝛼′𝑦¯𝐵𝑥𝑦differential-d𝑦𝜆¯𝑛superscriptsubscript0𝑥superscript𝛼′𝑦¯𝐵𝑥𝑦differential-d𝑦𝑏¯𝐵𝑥1{\alpha^{\prime}(x)=\lambda\overline{n}\int_{0}^{\infty}\alpha^{\prime}(y)\overline{B}(x+y)dy+\lambda\overline{n}\int_{0}^{x}\alpha^{\prime}(y)\overline{B}(x-y)dy+b\overline{B}(x)+1}.</td>
<td></td>
<td>(7)</td>
</tr>
</tbody>
</table>
Here $`\overline{n} = {{\overline{F}\hspace{0pt}{(\theta)}}/{({1 - \rho_{\theta}})}}`$ is the average batch size, $`{\overline{B}\hspace{0pt}{(x)}} = {{{\overline{F}\hspace{0pt}{({\theta + x})}}/\overline{F}}\hspace{0pt}{(\theta)}}`$ is the complementary truncated distribution and $`b = {b\hspace{0pt}{(\theta)}} = {{2\hspace{0pt}\lambda\hspace{0pt}\overline{F}\hspace{0pt}{(\theta)}\hspace{0pt}{({{\overline{W}\hspace{0pt}{(\theta)}} + \theta})}}/{({1 - \rho_{\theta}})}}`$ is the average number of jobs that arrive to the low priority queue in addition to the tagged job.
The expected sojourn time in the system is given by the following equations:
<table>
<tbody>
<tr>
<td></td>
<td>T¯(θ)=∫0∞T¯TLPS(x)𝑑F(x),¯𝑇𝜃superscriptsubscript0superscript¯𝑇𝑇𝐿𝑃𝑆𝑥differential-d𝐹𝑥\displaystyle{\overline{T}(\theta)}=\int_{0}^{\infty}{\overline{T}}^{TLPS}(x)dF(x),</td>
<td></td>
<td></td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td>T¯(θ)=Xθ1¯+W¯(θ)F¯(θ)1−ρθ+11−ρθT¯BPS(θ),¯𝑇𝜃¯superscriptsubscript𝑋𝜃1¯𝑊𝜃¯𝐹𝜃1subscript𝜌𝜃11subscript𝜌𝜃superscript¯𝑇𝐵𝑃𝑆𝜃\displaystyle\overline{T}(\theta)=\frac{{\overline{X_{\theta}^{1}}}+\overline{W}(\theta)\overline{F}(\theta)}{1-\rho_{\theta}}+{\frac{1}{1-\rho_{\theta}}}\overline{T}^{BPS}(\theta),{}</td>
<td></td>
<td>(8)</td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td>T¯BPS(θ)=∫θ∞α(x−θ)𝑑F(x)=∫0∞α′(x)F¯(x+θ)𝑑x.superscript¯𝑇𝐵𝑃𝑆𝜃superscriptsubscript𝜃𝛼𝑥𝜃differential-d𝐹𝑥superscriptsubscript0superscript𝛼′𝑥¯𝐹𝑥𝜃differential-d𝑥\displaystyle\overline{T}^{BPS}(\theta)={\int_{\theta}^{\infty}\!\!{{\alpha(x-\theta)}}dF(x)=\int_{0}^{\infty}\!\!\alpha^{\prime}(x)\overline{F}(x+\theta)dx}.{}</td>
<td></td>
<td>(9)</td>
</tr>
</tbody>
</table>
## 3 Hyper-exponential job size distribution with two phases
### 3.1 Notations
In the first part of our work we consider the hyper-exponential job size distribution with two phases. Namely, according to (LABEL:eq:F_N) the cumulative distribution function $`F\hspace{0pt}{(x)}`$ for $`N = 2`$ is given by
<table>
<tbody>
<tr>
<td></td>
<td>F(x)=1−p1e−μ1x−p2e−μ2x,𝐹𝑥1subscript𝑝1superscript𝑒subscript𝜇1𝑥subscript𝑝2superscript𝑒subscript𝜇2𝑥\displaystyle F(x)=1-p_{1}{e^{-\mu_{1}x}}-p_{2}{e^{-\mu_{2}x}},</td>
<td></td>
</tr>
</tbody>
</table>
where $`{p_{1} + p_{2}} = 1`$ and $`{p_{1},p_{2}} > 0`$.
The mean job size $`m`$, the second moment $`d`$, the parameters $`\overline{F_{\theta}^{i}}`$, $`\overline{X_{\theta}^{1}}`$, $`\overline{X_{\theta}^{2}}`$ and $`\rho_{\theta}`$ are defined as in Section [2.1](#S2.SS1 "2.1 Main definitions ‣ 2 Model description") and Section [2.2](#S2.SS2 "2.2 The expected sojourn time in TLPS system ‣ 2 Model description") by formulas (LABEL:eq:m_d),(LABEL:eq:F_i_theta),(LABEL:eq:X1X2), (LABEL:eq:rho_theta) with $`N = 2`$.
We note that the system has four free parameters. In particular, if we fix $`\mu_{1}`$, $`\epsilon = {\mu_{2}/\mu_{1}}`$, $`m`$, and $`\rho`$, the other parameters $`\mu_{2}`$, $`p_{1}`$, $`p_{2}`$ and $`\lambda`$ will be functions of the former parameters.
### 3.2 Explicit form for the expected sojourn time
To find $`{\overline{T}}^{T\hspace{0pt}L\hspace{0pt}P\hspace{0pt}S}\hspace{0pt}{(x)}`$ we need to solve the integral equation ([7](#S2.E7 "In 2.2 The expected sojourn time in TLPS system ‣ 2 Model description")). To solve ([7](#S2.E7 "In 2.2 The expected sojourn time in TLPS system ‣ 2 Model description")) we use the Laplace transform based method described in \[[6](#bib.bib6)\].
######
Theorem 1
.
The expected sojourn time in the TLPS system with the hyper-exponential job size distribution with two phases is given by
<table>
<tbody>
<tr>
<td></td>
<td>T¯(θ)=Xθ1¯+W¯(θ)F¯(θ)1−ρθ+m−Xθ1¯1−ρ+b(θ)(μ1μ2(m−Xθ1¯)2+δρ(θ)F¯2(θ))2(1−ρ)F¯(θ)(μ1+μ2−γ(θ)F¯(θ)),¯𝑇𝜃¯superscriptsubscript𝑋𝜃1¯𝑊𝜃¯𝐹𝜃1subscript𝜌𝜃𝑚¯superscriptsubscript𝑋𝜃11𝜌𝑏𝜃subscript𝜇1subscript𝜇2superscript𝑚¯superscriptsubscript𝑋𝜃12subscript𝛿𝜌𝜃superscript¯𝐹2𝜃21𝜌¯𝐹𝜃subscript𝜇1subscript𝜇2𝛾𝜃¯𝐹𝜃{}{\overline{T}(\theta)}=\frac{\overline{X_{\theta}^{1}}+\overline{W}(\theta)\overline{F}(\theta)}{1-\rho_{\theta}}+{\frac{m-\overline{X_{\theta}^{1}}}{1-\rho}}+{\frac{b(\theta)\left({\mu_{1}\mu_{2}}(m-\overline{X_{\theta}^{1}})^{2}+{\delta_{\rho}(\theta)}{\overline{F}\,^{2}(\theta)}\right)}{2(1-\rho){\overline{F}(\theta)}{\left(\mu_{1}+\mu_{2}-{\gamma(\theta)\overline{F}(\theta)}\right)}}},</td>
<td></td>
<td>(10)</td>
</tr>
</tbody>
</table>
where $`{\delta_{\rho}\hspace{0pt}{(\theta)}} = {1 - {\gamma\hspace{0pt}{(\theta)}\hspace{0pt}{({m - \overline{X_{\theta}^{1}}})}}} = {{({1 - \rho})}/{({1 - \rho_{\theta}})}}`$ and $`{\gamma\hspace{0pt}{(\theta)}} = {\lambda/{({1 - \rho_{\theta}})}}`$.
###### Proof..
We can rewrite integral equation ([7](#S2.E7 "In 2.2 The expected sojourn time in TLPS system ‣ 2 Model description")) in the following way
<table>
<tbody>
<tr>
<td></td>
<td>α′(x)superscript𝛼′𝑥\displaystyle\alpha^{\prime}(x)</td>
<td>=\displaystyle=</td>
<td>γ(θ)∫0∞α′(y)F¯(x+y+θ)𝑑y+γ(θ)∫0xα′(y)F¯(x−y+θ)𝑑y+b(θ)B¯(x)+1,𝛾𝜃superscriptsubscript0superscript𝛼′𝑦¯𝐹𝑥𝑦𝜃differential-d𝑦𝛾𝜃superscriptsubscript0𝑥superscript𝛼′𝑦¯𝐹𝑥𝑦𝜃differential-d𝑦𝑏𝜃¯𝐵𝑥1\displaystyle\gamma(\theta)\int_{0}^{\infty}\!\!\alpha^{\prime}(y){{\overline{F}(x+y+\theta)}}dy+\gamma(\theta)\int_{0}^{x}\!\!\alpha^{\prime}(y){\overline{F}(x-y+\theta)}dy+b(\theta)\overline{B}(x)+1,</td>
<td></td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td>α′(x)superscript𝛼′𝑥\displaystyle\alpha^{\prime}(x)</td>
<td>=\displaystyle=</td>
<td>γ(θ)∑i=1,2Fθi¯e−μix∫0∞α′(y)e−μiy𝑑y+γ(θ)∫0xα′(y)F¯(x−y+θ)𝑑y+b(θ)B¯(x)+1.𝛾𝜃subscript𝑖12¯superscriptsubscript𝐹𝜃𝑖superscript𝑒subscript𝜇𝑖𝑥superscriptsubscript0superscript𝛼′𝑦superscript𝑒subscript𝜇𝑖𝑦differential-d𝑦𝛾𝜃superscriptsubscript0𝑥superscript𝛼′𝑦¯𝐹𝑥𝑦𝜃differential-d𝑦𝑏𝜃¯𝐵𝑥1\displaystyle\gamma(\theta)\sum_{i=1,2}{\overline{F_{\theta}^{i}}}e^{{-\mu_{i}x}}\int_{0}^{\infty}\!\!\alpha^{\prime}(y)e^{-\mu_{i}y}dy+\gamma(\theta)\int_{0}^{x}\!\!\alpha^{\prime}(y){\overline{F}(x-y+\theta)}dy+b(\theta)\overline{B}(x)+1.</td>
<td></td>
</tr>
</tbody>
</table>
We note that in the latter equation $`\int_{0}^{\infty}{\alpha'\hspace{0pt}{(y)}\hspace{0pt}e^{- {\mu_{i}\hspace{0pt}y}}\hspace{0pt}{dy}}`$, $`i = {1,2}`$ are the Laplace transforms of $`\alpha'\hspace{0pt}{(y)}`$ evaluated at $`{{\mu_{i},i} = 1},2`$. Denote
<table>
<tbody>
<tr>
<td></td>
<td>Li=∫0∞α′(y)e−μiy𝑑y,i=1,2.formulae-sequencesubscript𝐿𝑖superscriptsubscript0superscript𝛼′𝑦superscript𝑒subscript𝜇𝑖𝑦differential-d𝑦𝑖12\displaystyle L_{i}=\int_{0}^{\infty}\alpha^{\prime}(y){e^{-\mu_{i}y}dy},\qquad i=1,2.</td>
<td></td>
</tr>
</tbody>
</table>
Then, we have
<table>
<tbody>
<tr>
<td></td>
<td>α′(x)=γ(θ)∑i=1,2Fθi¯Lie−μix+γ(θ)∫0xα′(y)F¯(x−y+θ)𝑑y+b(θ)B¯(x)+1.superscript𝛼′𝑥𝛾𝜃subscript𝑖12¯superscriptsubscript𝐹𝜃𝑖subscript𝐿𝑖superscript𝑒subscript𝜇𝑖𝑥𝛾𝜃superscriptsubscript0𝑥superscript𝛼′𝑦¯𝐹𝑥𝑦𝜃differential-d𝑦𝑏𝜃¯𝐵𝑥1\displaystyle\alpha^{\prime}(x)=\gamma(\theta)\sum_{i=1,2}{\overline{F_{\theta}^{i}}}L_{i}e^{{-\mu_{i}x}}+\gamma(\theta)\int_{0}^{x}\alpha^{\prime}(y){\overline{F}(x-y+\theta)}dy+b(\theta)\overline{B}(x)+1.</td>
<td></td>
</tr>
</tbody>
</table>
Now taking the Laplace transform of the above equation and using the convolution property, we get
<table>
<tbody>
<tr>
<td></td>
<td></td>
<td></td>
<td>Lα′(s)=γ(θ)∑i=1,2Fθi¯Lis+μi+γ(θ)∑i=1,2Fθi¯Lα′(s)s+μi+b(θ)F¯(θ)∑i=1,2Fθi¯s+μi+1s,subscript𝐿superscript𝛼′𝑠𝛾𝜃subscript𝑖12¯superscriptsubscript𝐹𝜃𝑖subscript𝐿𝑖𝑠subscript𝜇𝑖𝛾𝜃subscript𝑖12¯superscriptsubscript𝐹𝜃𝑖subscript𝐿superscript𝛼′𝑠𝑠subscript𝜇𝑖𝑏𝜃¯𝐹𝜃subscript𝑖12¯superscriptsubscript𝐹𝜃𝑖𝑠subscript𝜇𝑖1𝑠\displaystyle L_{\alpha^{\prime}}(s)=\gamma(\theta)\sum_{i=1,2}{\frac{\overline{F_{\theta}^{i}}L_{i}}{s+\mu_{i}}}+\gamma(\theta)\sum_{i=1,2}{\frac{\overline{F_{\theta}^{i}}L_{\alpha^{\prime}}(s)}{s+\mu_{i}}}+\frac{b(\theta)}{\overline{F}(\theta)}\sum_{i=1,2}{\frac{\overline{F_{\theta}^{i}}}{s+\mu_{i}}}+{\frac{1}{s}},</td>
<td></td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td>⟹⟹\displaystyle\Longrightarrow</td>
<td></td>
<td>Lα′(s)(1−γ(θ)∑i=1,2Fθi¯s+μi)=γ(θ)∑i=1,2Fθi¯Lis+μi+b(θ)F¯(θ)∑i=1,2Fθi¯s+μi+1s.subscript𝐿superscript𝛼′𝑠1𝛾𝜃subscript𝑖12¯superscriptsubscript𝐹𝜃𝑖𝑠subscript𝜇𝑖𝛾𝜃subscript𝑖12¯superscriptsubscript𝐹𝜃𝑖subscript𝐿𝑖𝑠subscript𝜇𝑖𝑏𝜃¯𝐹𝜃subscript𝑖12¯superscriptsubscript𝐹𝜃𝑖𝑠subscript𝜇𝑖1𝑠\displaystyle L_{\alpha^{\prime}}(s)\left(1-\gamma(\theta)\sum_{i=1,2}{\frac{\overline{F_{\theta}^{i}}}{s+\mu_{i}}}\right)=\gamma(\theta)\sum_{i=1,2}{\frac{\overline{F_{\theta}^{i}}L_{i}}{s+\mu_{i}}}+\frac{b(\theta)}{\overline{F}(\theta)}\sum_{i=1,2}{\frac{\overline{F_{\theta}^{i}}}{s+\mu_{i}}}+{\frac{1}{s}}.</td>
<td></td>
</tr>
</tbody>
</table>
Here $`{L_{\alpha'}\hspace{0pt}{(s)}} = {\int_{0}^{\infty}{\alpha'\hspace{0pt}{(x)}\hspace{0pt}e^{- {s\hspace{0pt}x}}\hspace{0pt}{dx}}}`$ is the Laplace transform of $`\alpha'\hspace{0pt}{(x)}`$. Let us note that
$`{{L_{\alpha'}\hspace{0pt}{(\mu_{i})}} = L_{i}},{i = {1,2}}`$. Then, if we substitute into the above equation $`s = \mu_{1}`$ and $`s = \mu_{2}`$, we can get $`L_{1}`$ and $`L_{2}`$ as a solution of the linear system
<table>
<tbody>
<tr>
<td></td>
<td>L1subscript𝐿1\displaystyle L_{1}</td>
<td>=\displaystyle=</td>
<td>1(μ1+μ2−γ(θ)F¯(θ))δρ(θ)(b(θ)2F¯(θ)(μ2(m−Xθ1¯)+δρ(θ)F¯(θ)))+1μ1δρ(θ),1subscript𝜇1subscript𝜇2𝛾𝜃¯𝐹𝜃subscript𝛿𝜌𝜃𝑏𝜃2¯𝐹𝜃subscript𝜇2𝑚¯superscriptsubscript𝑋𝜃1subscript𝛿𝜌𝜃¯𝐹𝜃1subscript𝜇1subscript𝛿𝜌𝜃\displaystyle{\frac{1}{\left(\mu_{1}+\mu_{2}-{\gamma(\theta)\overline{F}(\theta)}\right){\delta_{\rho}(\theta)}}}\left({\frac{b(\theta)}{2\overline{F}(\theta)}}\left({\mu_{2}}(m-\overline{X_{\theta}^{1}})+{\delta_{\rho}(\theta)}{\overline{F}(\theta)}\right)\right)+{{\frac{1}{\mu_{1}{\delta_{\rho}(\theta)}}}},</td>
<td></td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td>L2subscript𝐿2\displaystyle L_{2}</td>
<td>=\displaystyle=</td>
<td>1(μ1+μ2−γ(θ)F¯(θ))δρ(θ)(b(θ)2F¯(θ)(μ1(m−Xθ1¯)+δρ(θ)F¯(θ)))+1μ2δρ(θ).1subscript𝜇1subscript𝜇2𝛾𝜃¯𝐹𝜃subscript𝛿𝜌𝜃𝑏𝜃2¯𝐹𝜃subscript𝜇1𝑚¯superscriptsubscript𝑋𝜃1subscript𝛿𝜌𝜃¯𝐹𝜃1subscript𝜇2subscript𝛿𝜌𝜃\displaystyle{\frac{1}{\left(\mu_{1}+\mu_{2}-{\gamma(\theta)\overline{F}(\theta)}\right){\delta_{\rho}(\theta)}}}\left({\frac{b(\theta)}{2{\overline{F}(\theta)}}}\left({{\mu_{1}}(m-\overline{X_{\theta}^{1}})}+{\delta_{\rho}(\theta)}{\overline{F}(\theta)}\right)\right)+{\frac{1}{\mu_{2}\delta_{\rho}(\theta)}}.</td>
<td></td>
</tr>
</tbody>
</table>
Next we need to calculate $`{\overline{T}}^{B\hspace{0pt}P\hspace{0pt}S}\hspace{0pt}{(\theta)}`$.
<table>
<tbody>
<tr>
<td></td>
<td>T¯BPS(θ)superscript¯𝑇𝐵𝑃𝑆𝜃\displaystyle\overline{T}^{BPS}(\theta)</td>
<td>=\displaystyle=</td>
<td>∫0∞α′(x)F¯(x+θ)𝑑x=∫0∞α′(x)∑i=1,2Fθi¯e−μixdx=∑i=1,2Fθi¯Li,superscriptsubscript0superscript𝛼′𝑥¯𝐹𝑥𝜃differential-d𝑥superscriptsubscript0superscript𝛼′𝑥subscript𝑖12¯superscriptsubscript𝐹𝜃𝑖superscript𝑒subscript𝜇𝑖𝑥𝑑𝑥subscript𝑖12¯superscriptsubscript𝐹𝜃𝑖subscript𝐿𝑖\displaystyle\int_{0}^{\infty}{{\alpha^{\prime}(x)}}\overline{F}(x+\theta)dx=\int_{0}^{\infty}{{\alpha^{\prime}(x)}}\sum_{i=1,2}{\overline{F_{\theta}^{i}}e^{-\mu_{i}x}}dx=\sum_{i=1,2}{{\overline{F_{\theta}^{i}}L_{i}}},</td>
<td></td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td>T¯BPS(θ)superscript¯𝑇𝐵𝑃𝑆𝜃\displaystyle\overline{T}^{BPS}(\theta)</td>
<td>=\displaystyle=</td>
<td>1−ρθ1−ρ(m−Xθ1¯+b(θ)(μ1μ2(m−Xθ1¯)2+δρ(θ)F¯2(θ))2F¯(θ)(μ1+μ2−γ(θ)F¯(θ))).1subscript𝜌𝜃1𝜌𝑚¯superscriptsubscript𝑋𝜃1𝑏𝜃subscript𝜇1subscript𝜇2superscript𝑚¯superscriptsubscript𝑋𝜃12subscript𝛿𝜌𝜃superscript¯𝐹2𝜃2¯𝐹𝜃subscript𝜇1subscript𝜇2𝛾𝜃¯𝐹𝜃\displaystyle{\frac{1-\rho_{\theta}}{1-\rho}}\left({{m-\overline{X_{\theta}^{1}}}}+{\frac{b(\theta)\left({\mu_{1}\mu_{2}}(m-\overline{X_{\theta}^{1}})^{2}+{\delta_{\rho}(\theta)}{\overline{F}\,^{2}(\theta)}\right)}{2{\overline{F}(\theta)}{\left(\mu_{1}+\mu_{2}-{\gamma(\theta)\overline{F}(\theta)}\right)}}}\right).</td>
<td></td>
</tr>
</tbody>
</table>
Finally, by ([8](#S2.E8 "In 2.2 The expected sojourn time in TLPS system ‣ 2 Model description")) we have ([10](#S3.E10 "In Theorem 1. ‣ 3.2 Explicit form for the expected sojourn time ‣ 3 Hyper-exponential job size distribution with two phases")). ∎
### 3.3 Optimal threshold approximation
We are interested in the minimization of the expected sojourn time $`\overline{T}\hspace{0pt}{(\theta)}`$ with respect to $`\theta`$. Of course, one can differentiate the exact analytic expression provided in Theorem [1](#ThmTheorem1 "Theorem 1. ‣ 3.2 Explicit form for the expected sojourn time ‣ 3 Hyper-exponential job size distribution with two phases") and set the result of the differentiation to zero. However, this will give a transcendental equation for the optimal value of the threshold.
In order to find an approximate solution of $`{{\overline{T}}'\hspace{0pt}{(\theta)}} = {{{d\hspace{0pt}\overline{T}\hspace{0pt}{(\theta)}}/d}\hspace{0pt}\theta} = 0`$, we shall approximate the derivative $`{\overline{T}}'\hspace{0pt}{(\theta)}`$ by some function $`{\overset{\sim}{\overline{T}}}'\hspace{0pt}{(\theta)}`$ and obtain a solution to $`{{\overset{\sim}{\overline{T}}}'\hspace{0pt}{({\overset{\sim}{\theta}}_{o\hspace{0pt}p\hspace{0pt}t})}} = 0`$.
Since in the Internet connections belong to two distinct classes with very different sizes of transfer (see Section [2.1](#S2.SS1 "2.1 Main definitions ‣ 2 Model description")), then to find the approximation of $`{\overline{T}}'\hspace{0pt}{(\theta)}`$ we consider a particular case when $`\mu_{2}\operatorname{<<}\mu_{1}`$. Let us introduce a small parameter $`\epsilon`$ such that
<table>
<tbody>
<tr>
<td></td>
<td>μ2=ϵμ1,ϵ→0,p1=1−ϵ(mμ1−1)1−ϵ,p2=ϵ(mμ1−1)1−ϵ.formulae-sequencesubscript𝜇2italic-ϵsubscript𝜇1formulae-sequence→italic-ϵ0formulae-sequencesubscript𝑝11italic-ϵ𝑚subscript𝜇111italic-ϵsubscript𝑝2italic-ϵ𝑚subscript𝜇111italic-ϵ\displaystyle\mu_{2}=\epsilon{\mu_{1}},\quad\epsilon\rightarrow 0,\quad p_{1}=1-{\frac{\epsilon\,\left(m{\mu_{1}}-1\right)}{1-\epsilon}},\quad p_{2}={\frac{\epsilon\,\left(m{\mu_{1}}-1\right)}{1-\epsilon}}.</td>
<td></td>
</tr>
</tbody>
</table>
We note that when $`\epsilon\rightarrow 0`$ the second moment of the job size distribution goes to infinity.
We then verify that $`{\overset{\sim}{\theta}}_{o\hspace{0pt}p\hspace{0pt}t}`$ indeed converges to the minimum of $`\overline{T}\hspace{0pt}{(\theta)}`$ when $`\epsilon\rightarrow 0`$.
######
Lemma 2
.
The following inequality holds: $`{\mu_{1}\hspace{0pt}\rho} > \lambda`$.
###### Proof..
Since $`p_{1} > 0`$ and $`p_{2} > 0`$, we have the following inequality $`{m\hspace{0pt}\mu_{1}} > 1`$. Then, $`m > \frac{1}{\mu_{1}}`$. Taking into account that $`{\lambda\hspace{0pt}m} = \rho`$ we get $`\frac{\rho}{\lambda} > \frac{1}{\mu_{1}}`$. Consequently, we have that $`{\mu_{1}\hspace{0pt}\rho} > \lambda`$.∎
######
Proposition 3
.
The derivative of $`\overline{T}\hspace{0pt}{(\theta)}`$ can be approximated by the following function:
<table>
<tbody>
<tr>
<td></td>
<td>T¯~′(θ)=−e−μ1θμ1c1+e−μ2θμ2c2,superscript~¯𝑇′𝜃superscript𝑒subscript𝜇1𝜃subscript𝜇1subscript𝑐1superscript𝑒subscript𝜇2𝜃subscript𝜇2subscript𝑐2\displaystyle{{\widetilde{\overline{T}}\!\,^{\prime}}(\theta)}=-e^{-\mu_{1}{\theta}}{\mu_{1}}{c_{1}}+e^{-\mu_{2}{\theta}}{\mu_{2}}{c_{2}},</td>
<td></td>
</tr>
</tbody>
</table>
where
<table>
<tbody>
<tr>
<td></td>
<td>c1=λ(mμ1−1)μ1(μ1−λ)(1−ρ),c2=λ(mμ1−1)(μ1−λ)2.formulae-sequencesubscript𝑐1𝜆𝑚subscript𝜇11subscript𝜇1subscript𝜇1𝜆1𝜌subscript𝑐2𝜆𝑚subscript𝜇11superscriptsubscript𝜇1𝜆2\displaystyle{}c_{1}={\frac{\lambda(m\mu_{1}-1)}{\mu_{1}(\mu_{1}-\lambda)(1-\rho)}},\quad c_{2}=\frac{\lambda(m\mu_{1}-1)}{(\mu_{1}-\lambda)^{2}}.</td>
<td></td>
<td>(11)</td>
</tr>
</tbody>
</table>
Namely,
<table>
<tbody>
<tr>
<td></td>
<td>T¯′(θ)−T¯~′(θ)=O(μ2/μ1).superscript¯𝑇′𝜃superscript~¯𝑇′𝜃𝑂subscript𝜇2subscript𝜇1\displaystyle{\overline{T}^{\prime}(\theta)}-{{\widetilde{\overline{T}}\!\,^{\prime}}(\theta)}=O(\mu_{2}/\mu_{1}).</td>
<td></td>
</tr>
</tbody>
</table>
###### Proof..
Using the analytical expression for both $`{\overline{T}}'\hspace{0pt}{(\theta)}`$ and $`{\overset{\sim}{\overline{T}}}'\hspace{0pt}{(\theta)}`$, we get the Taylor series for $`{{\overline{T}}'\hspace{0pt}{(\theta)}} - {{\overset{\sim}{\overline{T}}}'\hspace{0pt}{(\theta)}}`$ with respect to $`\epsilon`$, which shows that indeed
<table>
<tbody>
<tr>
<td></td>
<td>T¯′(θ)−T¯~′(θ)=O(ϵ).superscript¯𝑇′𝜃superscript~¯𝑇′𝜃𝑂italic-ϵ\displaystyle\overline{T}^{\prime}(\theta)-{\widetilde{\overline{T}}\!\,^{\prime}}(\theta)=O(\epsilon).</td>
<td></td>
</tr>
</tbody>
</table>
∎
Thus we have found an approximation of the derivative of $`\overline{T}\hspace{0pt}{(\theta)}`$. Now we can find an approximation of the optimal threshold by solving the equation $`{{\overset{\sim}{\overline{T}}}'\hspace{0pt}{(\theta)}} = 0`$.
######
Theorem 4
.
Let $`\theta_{o\hspace{0pt}p\hspace{0pt}t}`$ denote the optimal value of the threshold. Namely, $`\theta_{o\hspace{0pt}p\hspace{0pt}t} = {\text{arg}\hspace{0pt}{\min\overline{T}}\hspace{0pt}{(\theta)}}`$. The value $`{\overset{\sim}{\theta}}_{o\hspace{0pt}p\hspace{0pt}t}`$ given by
<table>
<tbody>
<tr>
<td></td>
<td>θ~opt=1μ1−μ2ln((μ1−λ)μ2(1−ρ))subscript~𝜃𝑜𝑝𝑡1subscript𝜇1subscript𝜇2subscript𝜇1𝜆subscript𝜇21𝜌\displaystyle\tilde{\theta}_{opt}=\frac{1}{\mu_{1}-\mu_{2}}\ln\left(\frac{(\mu_{1}-\lambda)}{\mu_{2}(1-\rho)}\right)</td>
<td></td>
</tr>
</tbody>
</table>
approximates $`\theta_{o\hspace{0pt}p\hspace{0pt}t}`$ so that $`{{\overline{T}}'\hspace{0pt}{({\overset{\sim}{\theta}}_{o\hspace{0pt}p\hspace{0pt}t})}} = {\text{o}\hspace{0pt}{({\mu_{2}/\mu_{1}})}}`$.
###### Proof..
Solving the equation
<table>
<tbody>
<tr>
<td></td>
<td>T¯~′(θ)=0,superscript~¯𝑇′𝜃0\displaystyle{{\widetilde{\overline{T}}\!\,^{\prime}}(\theta)}=0,</td>
<td></td>
</tr>
</tbody>
</table>
we get an analytic expression for the approximation of the optimal threshold:
<table>
<tbody>
<tr>
<td></td>
<td>θ~opt=−1μ1(1−ϵ)ln(ϵμ1(1−ρ)(μ1−λ))=1μ1−μ2ln((μ1−λ)μ2(1−ρ)).subscript~𝜃𝑜𝑝𝑡1subscript𝜇11italic-ϵitalic-ϵsubscript𝜇11𝜌subscript𝜇1𝜆1subscript𝜇1subscript𝜇2subscript𝜇1𝜆subscript𝜇21𝜌\displaystyle{\widetilde{\theta}_{opt}=-{\frac{1}{\mu_{1}(1-\epsilon)}}\ln\left(\epsilon\frac{\mu_{1}{(1-\rho)}}{(\mu_{1}-\lambda)}\right)={\frac{1}{\mu_{1}-\mu_{2}}}\ln\left(\frac{(\mu_{1}-\lambda)}{\mu_{2}{(1-\rho)}}\right)}.</td>
<td></td>
</tr>
</tbody>
</table>
Let us show that the above threshold approximation is greater than zero. We have to show that $`\frac{({\mu_{1} - \lambda})}{\mu_{2}\hspace{0pt}{({1 - \rho})}} > 1`$. Since $`\mu_{1} > \mu_{2}`$ and $`{\mu_{1}\hspace{0pt}\rho} > \lambda`$ (see Lemma [2](#ThmTheorem2 "Lemma 2. ‣ 3.3 Optimal threshold approximation ‣ 3 Hyper-exponential job size distribution with two phases")), we have
<table>
<tbody>
<tr>
<td></td>
<td></td>
<td></td>
<td>μ1>μ2subscript𝜇1subscript𝜇2\displaystyle\mu_{1}>{\mu_{2}}</td>
<td></td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td>⟹⟹\displaystyle\Longrightarrow</td>
<td></td>
<td>μ1(1−ρ)>μ2(1−ρ)subscript𝜇11𝜌subscript𝜇21𝜌\displaystyle\mu_{1}(1-\rho)>{\mu_{2}{(1-\rho)}}</td>
<td></td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td>⟹⟹\displaystyle\Longrightarrow</td>
<td></td>
<td>λ<μ1ρ<μ1−μ2(1−ρ)𝜆subscript𝜇1𝜌subscript𝜇1subscript𝜇21𝜌\displaystyle{\lambda}<\mu_{1}\rho<{\mu_{1}-\mu_{2}{(1-\rho)}}</td>
<td></td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td>⟹⟹\displaystyle\Longrightarrow</td>
<td></td>
<td>(μ1−λ)>μ2(1−ρ).subscript𝜇1𝜆subscript𝜇21𝜌\displaystyle{(\mu_{1}-\lambda)}>{\mu_{2}{(1-\rho)}}.</td>
<td></td>
</tr>
</tbody>
</table>
Expanding $`{\overline{T}}'\hspace{0pt}{({\overset{\sim}{\theta}}_{o\hspace{0pt}p\hspace{0pt}t})}`$ as a power series with respect to $`\epsilon`$ gives:
<table>
<tbody>
<tr>
<td></td>
<td>T¯′(θ~opt)=ϵ2(const0+const1lnϵ+const2ln2ϵ),superscript¯𝑇′subscript~𝜃𝑜𝑝𝑡superscriptitalic-ϵ2𝑐𝑜𝑛𝑠subscript𝑡0𝑐𝑜𝑛𝑠subscript𝑡1italic-ϵ𝑐𝑜𝑛𝑠subscript𝑡2superscript2italic-ϵ\displaystyle{\overline{T}^{\prime}(\widetilde{\theta}_{opt})}=\epsilon^{2}(const_{0}+const_{1}\ln{\epsilon}+const_{2}\ln^{2}{\epsilon}),</td>
<td></td>
</tr>
</tbody>
</table>
where $`{{{c\hspace{0pt}o\hspace{0pt}n\hspace{0pt}s\hspace{0pt}t_{i}},i} = 1},2`$ are some constant values
<sup>1</sup>
<sup>1</sup>1The expressions for the constants $`c\hspace{0pt}o\hspace{0pt}n\hspace{0pt}s\hspace{0pt}t_{i}`$ are cumbersome and can be found using Maple command “series”.
with respect to $`\epsilon`$. Thus,
<table>
<tbody>
<tr>
<td></td>
<td>T¯′(θ~opt)=o(ϵ)=o(μ2/μ1),superscript¯𝑇′subscript~𝜃𝑜𝑝𝑡𝑜italic-ϵ𝑜subscript𝜇2subscript𝜇1\displaystyle{{\overline{T}}^{\prime}(\widetilde{\theta}_{opt})}=o(\epsilon)=o(\mu_{2}/\mu_{1}),</td>
<td></td>
</tr>
</tbody>
</table>
which completes the proof. ∎
In the next proposition we characterize the limiting behavior of $`\overline{T}\hspace{0pt}{(\theta_{o\hspace{0pt}p\hspace{0pt}t})}`$ and $`\overline{T}\hspace{0pt}{({\overset{\sim}{\theta}}_{o\hspace{0pt}p\hspace{0pt}t})}`$ as $`\epsilon\rightarrow 0`$. In particular, we show that $`\overline{T}\hspace{0pt}{({\overset{\sim}{\theta}}_{o\hspace{0pt}p\hspace{0pt}t})}`$ tends to the exact minimum of $`\overline{T}\hspace{0pt}{(\theta)}`$ when $`\epsilon\rightarrow 0`$.
######
Proposition 5
.
<table>
<tbody>
<tr>
<td></td>
<td>limϵ→0T¯(θopt)=limϵ→0T¯(θ~opt)=m1−ρ−c1,subscript→italic-ϵ0¯𝑇subscript𝜃𝑜𝑝𝑡subscript→italic-ϵ0¯𝑇subscript~𝜃𝑜𝑝𝑡𝑚1𝜌subscript𝑐1\displaystyle{\lim_{\epsilon\rightarrow 0}{\overline{T}({\theta}_{opt})}=\lim_{\epsilon\rightarrow 0}{\overline{T}(\widetilde{\theta}_{opt})}}={\frac{m}{1-\rho}}-c_{1},</td>
<td></td>
</tr>
</tbody>
</table>
where $`c_{1}`$ is given by ([11](#S3.E11 "In Proposition 3. ‣ 3.3 Optimal threshold approximation ‣ 3 Hyper-exponential job size distribution with two phases")).
###### Proof..
We find the following limit, when $`\epsilon\rightarrow 0`$:
<table>
<tbody>
<tr>
<td></td>
<td>limϵ→0T¯(θ)=m1−ρ−λ(mμ1−1)μ1(μ1−λ)(1−ρ)+λ(mμ1−1)e−μ1θμ1(μ1−λ)(1−ρ),subscript→italic-ϵ0¯𝑇𝜃𝑚1𝜌𝜆𝑚subscript𝜇11subscript𝜇1subscript𝜇1𝜆1𝜌𝜆𝑚subscript𝜇11superscript𝑒subscript𝜇1𝜃subscript𝜇1subscript𝜇1𝜆1𝜌\displaystyle\lim_{\epsilon\rightarrow 0}{\overline{T}({\theta})}={\frac{m}{1-\rho}}-{\frac{\lambda(m\mu_{1}-1)}{\mu_{1}(\mu_{1}-\lambda)(1-\rho)}}+\frac{\lambda(m\mu_{1}-1)e^{-\mu_{1}\theta}}{\mu_{1}(\mu_{1}-\lambda)(1-\rho)},</td>
<td></td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td>limϵ→0T¯(θ)=m1−ρ−c1+c1e−μ1θ,subscript→italic-ϵ0¯𝑇𝜃𝑚1𝜌subscript𝑐1subscript𝑐1superscript𝑒subscript𝜇1𝜃\displaystyle{\lim_{\epsilon\rightarrow 0}{\overline{T}({\theta})}={\frac{m}{1-\rho}}-c_{1}+c_{1}e^{-\mu_{1}\theta}},</td>
<td></td>
</tr>
</tbody>
</table>
where $`c_{1}`$ is given by ([11](#S3.E11 "In Proposition 3. ‣ 3.3 Optimal threshold approximation ‣ 3 Hyper-exponential job size distribution with two phases")). Since the function $`\lim_{\epsilon\rightarrow 0}{\overline{T}\hspace{0pt}{(\theta)}}`$ is a decreasing function, the optimal threshold for it is $`\theta_{o\hspace{0pt}p\hspace{0pt}t} = \infty`$. Thus,
<table>
<tbody>
<tr>
<td></td>
<td>limϵ→0T¯(θopt)=limθ→∞limϵ→0T¯(θ)=m1−ρ−c1.subscript→italic-ϵ0¯𝑇subscript𝜃𝑜𝑝𝑡subscript→𝜃subscript→italic-ϵ0¯𝑇𝜃𝑚1𝜌subscript𝑐1\displaystyle{\lim_{\epsilon\rightarrow 0}{\overline{T}({\theta_{opt}})}}={\lim_{\theta\rightarrow\infty}\lim_{\epsilon\rightarrow 0}{\overline{T}({\theta})}={\frac{m}{1-\rho}}-c_{1}}.</td>
<td></td>
</tr>
</tbody>
</table>
On the other hand, we obtain
<table>
<tbody>
<tr>
<td></td>
<td>limϵ→0T¯(θ~opt)=m1−ρ−c1,subscript→italic-ϵ0¯𝑇subscript~𝜃𝑜𝑝𝑡𝑚1𝜌subscript𝑐1\displaystyle{\lim_{\epsilon\rightarrow 0}{\overline{T}({\widetilde{\theta}_{opt}})}={\frac{m}{1-\rho}}-c_{1}},</td>
<td></td>
</tr>
</tbody>
</table>
which proves the proposition. ∎
### 3.4 Experimental results
In Figure 1-2 we show the plots for the following parameters: $`\rho = {10/11}`$ (default value), $`m = {20/11}`$, $`\mu_{1} = 1`$, $`\mu_{2} = {1/10}`$, so $`\lambda = {1/2}`$ and $`\epsilon = {\mu_{2}/\mu_{1}} = {1/10}`$. Then, $`p_{1} = {10/11}`$ and $`p_{2} = {1/11}`$.
In Figure 1 we plot $`\overline{T}\hspace{0pt}{(\theta)}`$, $`{\overline{T}}^{P\hspace{0pt}S}`$ and $`\overline{T}\hspace{0pt}{({\overset{\sim}{\theta}}_{o\hspace{0pt}p\hspace{0pt}t})}`$. We note, that the expected sojourn time in the standard PS system $`{\overline{T}}^{P\hspace{0pt}S}`$ is equal to $`\overline{T}\hspace{0pt}{(0)}`$. We observe that $`\overline{T}\hspace{0pt}{({\overset{\sim}{\theta}}_{o\hspace{0pt}p\hspace{0pt}t})}`$ corresponds well to the optimum even though $`\epsilon = {1/10}`$ is not too small.
Let us now study the gain that we obtain using TLPS, by setting $`\theta = {\overset{\sim}{\theta}}_{o\hspace{0pt}p\hspace{0pt}t}`$, in comparison with the standard PS. To this end, we plot the ratio $`{g\hspace{0pt}{(\rho)}} = \frac{{\overline{T}}^{P\hspace{0pt}S} - {\overline{T}\hspace{0pt}{({\overset{\sim}{\theta}}_{o\hspace{0pt}p\hspace{0pt}t})}}}{{\overline{T}}^{P\hspace{0pt}S}}`$ in Figure 2. The gain in the system performance with TLPS in comparison with PS strongly depends on $`\rho`$, the load of the system. One can see, that the gain of the TLPS system with respect to the standard PS system goes up to 45% when the load of the system increases.
To study the sensitivity of the TLPS system with respect to $`\theta`$, we find the gain of the TLPS system with respect to the standard PS system, we plot in Figure 2 $`{g_{1}\hspace{0pt}{(\rho)}} = \frac{{\overline{T}}^{P\hspace{0pt}S} - {\overline{T}\hspace{0pt}{({\frac{3}{2}\hspace{0pt}{\overset{\sim}{\theta}}_{o\hspace{0pt}p\hspace{0pt}t}})}}}{{\overline{T}}^{P\hspace{0pt}S}}`$ and $`{g_{2}\hspace{0pt}{(\rho)}} = \frac{{\overline{T}}^{P\hspace{0pt}S} - {\overline{T}\hspace{0pt}{({\frac{1}{2}\hspace{0pt}{\overset{\sim}{\theta}}_{o\hspace{0pt}p\hspace{0pt}t}})}}}{{\overline{T}}^{P\hspace{0pt}S}}`$. Thus, even with the 50% error of the $`{\overset{\sim}{\theta}}_{o\hspace{0pt}p\hspace{0pt}t}`$ value, the system performance is close to optimal.
One can see that it is beneficial to use TLPS instead of PS in the case of heavy and moderately heavy loads. We also observe that the optimal TLPS system is not too sensitive to the choice of the threshold near its optimal value, when the job size distribution is hyper-exponential with two phases. Nevertheless, it is better to choose larger rather than smaller values of the threshold.
Figure 1: $`\overline{T}\hspace{0pt}{(\theta)}`$ - solid line, $`{\overline{T}}^{P\hspace{0pt}S}\hspace{0pt}{(\theta)}`$ - dash dot line, $`\overline{T}\hspace{0pt}{({\overset{\sim}{\theta}}_{o\hspace{0pt}p\hspace{0pt}t})}`$ - dash line
Figure 2: $`g\hspace{0pt}{(\rho)}`$ - solid line, $`g_{1}\hspace{0pt}{(\rho)}`$ - dash line, $`g_{2}\hspace{0pt}{(\rho)}`$ - dash dot line
## 4 Hyper-exponential job size distribution with more than two phases
### 4.1 Notations
In the second part of the presented work we analyze the TLPS discipline with the hyper-exponential job size distribution with more than two phases. As was shown in \[[5](#bib.bib5), [7](#bib.bib7), [8](#bib.bib8)\], the hyper-exponential distribution with a significant number of phases models well the file size distribution in the Internet. Thus, in this section as the job size distribution we take the hyper-exponential function with many phases. Namely, according to (LABEL:eq:F_N),
<table>
<tbody>
<tr>
<td></td>
<td>F(x)=1−∑i=1Npie−μix,∑i=1Npi=1,μi>0,pi≥0,i=1,…,N, 1<N≤∞.formulae-sequence𝐹𝑥1superscriptsubscript𝑖1𝑁subscript𝑝𝑖superscript𝑒subscript𝜇𝑖𝑥formulae-sequencesuperscriptsubscript𝑖1𝑁subscript𝑝𝑖1formulae-sequencesubscript𝜇𝑖0formulae-sequencesubscript𝑝𝑖0formulae-sequence𝑖1…𝑁1𝑁\displaystyle F(x)=1-\sum_{i=1}^{N}{p_{i}\,e^{-\mu_{i}\,x}},\quad{\sum_{i=1}^{N}{p_{i}}=1},\,\,\,\mu_{i}>0,\,\,p_{i}\geq 0,\,\,\,i=1,...,N,\,\,\,1<N\leq\infty.</td>
<td></td>
</tr>
</tbody>
</table>
In the following we shall write simply $`\sum_{i}`$ instead of $`\sum_{i = 1}^{N}`$.
The mean job size $`m`$, the second moment $`d`$, the parameters $`\overline{F_{\theta}^{i}}`$, $`\overline{X_{\theta}^{1}}`$, $`\overline{X_{\theta}^{2}}`$ and $`\rho_{\theta}`$ are defined as in Section [2.1](#S2.SS1 "2.1 Main definitions ‣ 2 Model description") and Section [2.2](#S2.SS2 "2.2 The expected sojourn time in TLPS system ‣ 2 Model description") by formulas (LABEL:eq:m_d),(LABEL:eq:F_i_theta),(LABEL:eq:X1X2), (LABEL:eq:rho_theta) for any $`1 \leq N \leq \infty`$. The formulas presented in Section [2.2](#S2.SS2 "2.2 The expected sojourn time in TLPS system ‣ 2 Model description") can still be used to calculate $`b\hspace{0pt}{(\theta)}`$, $`\overline{B}\hspace{0pt}{(x)}`$, $`\overline{W}\hspace{0pt}{(\theta)}`$, $`\gamma\hspace{0pt}{(\theta)}`$, $`\delta_{\rho}\hspace{0pt}{(\theta)}`$, $`{\overline{T}}^{T\hspace{0pt}L\hspace{0pt}P\hspace{0pt}S}\hspace{0pt}{(x)}`$, $`\overline{T}\hspace{0pt}{(\theta)}`$.
We shall also need the following operator notations:
<table>
<tbody>
<tr>
<td></td>
<td>Φ1(β(x))subscriptΦ1𝛽𝑥\displaystyle{\Phi_{1}(\beta(x))}</td>
<td>=\displaystyle=</td>
<td>γ(θ)∫0∞β(y)F¯(x+y+θ)𝑑y+γ(θ)∫0xβ(y)F¯(x−y+θ)𝑑y,𝛾𝜃superscriptsubscript0𝛽𝑦¯𝐹𝑥𝑦𝜃differential-d𝑦𝛾𝜃superscriptsubscript0𝑥𝛽𝑦¯𝐹𝑥𝑦𝜃differential-d𝑦\displaystyle{\gamma(\theta)\int_{0}^{\infty}\beta(y){{\overline{F}(x+y+\theta)}}dy}+{\gamma(\theta){\int_{0}^{x}\beta(y){\overline{F}(x-y+\theta)}dy}},</td>
<td></td>
<td>(12)</td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td>Φ2(β(x))subscriptΦ2𝛽𝑥\displaystyle{\Phi_{2}(\beta(x))}</td>
<td>=\displaystyle=</td>
<td>∫0∞β(y)F¯(y+θ)𝑑ysuperscriptsubscript0𝛽𝑦¯𝐹𝑦𝜃differential-d𝑦\displaystyle{\int_{0}^{\infty}\beta(y){{\overline{F}(y+\theta)}}dy}</td>
<td></td>
<td>(13)</td>
</tr>
</tbody>
</table>
for any function $`\beta\hspace{0pt}{(x)}`$. In particular, for some given constant c, we have
<table>
<tbody>
<tr>
<td></td>
<td>Φ1(c)subscriptΦ1𝑐\displaystyle{\Phi_{1}(c)}</td>
<td>=\displaystyle=</td>
<td>cγ(θ)(m−Xθ1¯)=cq,𝑐𝛾𝜃𝑚¯superscriptsubscript𝑋𝜃1𝑐𝑞\displaystyle{c\,\gamma(\theta)(m-{\overline{X_{\theta}^{1}}})=c\,q},</td>
<td></td>
<td>(14)</td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td>Φ2(c)subscriptΦ2𝑐\displaystyle{\Phi_{2}(c)}</td>
<td>=\displaystyle=</td>
<td>c(m−Xθ1¯),𝑐𝑚¯superscriptsubscript𝑋𝜃1\displaystyle{c\,(m-{\overline{X_{\theta}^{1}}})},</td>
<td></td>
<td>(15)</td>
</tr>
</tbody>
</table>
where
<table>
<tbody>
<tr>
<td></td>
<td>q=γ(θ)(m−Xθ1¯)=λ(m−Xθ1¯)1−ρθ=ρ−ρθ1−ρθ<1.𝑞𝛾𝜃𝑚¯superscriptsubscript𝑋𝜃1𝜆𝑚¯superscriptsubscript𝑋𝜃11subscript𝜌𝜃𝜌subscript𝜌𝜃1subscript𝜌𝜃1{q}={\gamma(\theta)(m-{\overline{X_{\theta}^{1}}})=\frac{\lambda(m-{\overline{X_{\theta}^{1}}})}{1-\rho_{\theta}}=\frac{\rho-\rho_{\theta}}{1-\rho_{\theta}}<1}.</td>
<td></td>
<td>(16)</td>
</tr>
</tbody>
</table>
The integral equation ([7](#S2.E7 "In 2.2 The expected sojourn time in TLPS system ‣ 2 Model description")) can now be rewritten in the form
<table>
<tbody>
<tr>
<td></td>
<td>α′(x)=Φ1(α′(y))+b(θ)F¯(θ)F¯(x+θ)+1.superscript𝛼′𝑥subscriptΦ1superscript𝛼′𝑦𝑏𝜃¯𝐹𝜃¯𝐹𝑥𝜃1{}{\alpha^{\prime}(x)=\Phi_{1}(\alpha^{\prime}(y))}{+\frac{b(\theta)}{\overline{F}(\theta)}\overline{F}(x+\theta)+1}.</td>
<td></td>
<td>(17)</td>
</tr>
</tbody>
</table>
and equation ([9](#S2.E9 "In 2.2 The expected sojourn time in TLPS system ‣ 2 Model description")) for $`{\overline{T}}^{B\hspace{0pt}P\hspace{0pt}S}\hspace{0pt}{(\theta)}`$ takes the form
<table>
<tbody>
<tr>
<td></td>
<td>T¯BPS(θ)=Φ2(α′(x)).superscript¯𝑇𝐵𝑃𝑆𝜃subscriptΦ2superscript𝛼′𝑥{}{\overline{T}^{BPS}(\theta)=\Phi_{2}(\alpha^{\prime}(x))}.</td>
<td></td>
<td>(18)</td>
</tr>
</tbody>
</table>
### 4.2 Linear system based solution
Similarly to the first part of the proof of Theorem [1](#ThmTheorem1 "Theorem 1. ‣ 3.2 Explicit form for the expected sojourn time ‣ 3 Hyper-exponential job size distribution with two phases"), we obtain the following proposition.
######
Proposition 6
.
<table>
<tbody>
<tr>
<td></td>
<td>T¯BPS(θ)=∑iFθi¯Li,superscript¯𝑇𝐵𝑃𝑆𝜃subscript𝑖¯superscriptsubscript𝐹𝜃𝑖subscript𝐿𝑖\displaystyle\overline{T}^{BPS}(\theta)=\sum_{i}\overline{F_{\theta}^{i}}L_{i},</td>
<td></td>
</tr>
</tbody>
</table>
with
<table>
<tbody>
<tr>
<td></td>
<td>Li=Li∗+1δρ(θ)μi,subscript𝐿𝑖subscriptsuperscript𝐿𝑖1subscript𝛿𝜌𝜃subscript𝜇𝑖\displaystyle{L_{i}=L^{*}_{i}+\frac{1}{\delta_{\rho}(\theta)\mu_{i}}},</td>
<td></td>
</tr>
</tbody>
</table>
where the $`L_{i}^{\ast}`$ are the solution of the linear system
<table>
<tbody>
<tr>
<td></td>
<td>Lp∗(1−γ(θ)∑iFθi¯λp+μi)=γ(θ)∑iFθi¯Li∗λp+μi+b(θ)F¯(θ)∑iFθi¯λp+μi,p=1,…,N.formulae-sequencesubscriptsuperscript𝐿𝑝1𝛾𝜃subscript𝑖¯superscriptsubscript𝐹𝜃𝑖subscript𝜆𝑝subscript𝜇𝑖𝛾𝜃subscript𝑖¯superscriptsubscript𝐹𝜃𝑖subscriptsuperscript𝐿𝑖subscript𝜆𝑝subscript𝜇𝑖𝑏𝜃¯𝐹𝜃subscript𝑖¯superscriptsubscript𝐹𝜃𝑖subscript𝜆𝑝subscript𝜇𝑖𝑝1…𝑁{}L^{*}_{p}\left(1-\gamma(\theta)\sum_{i}{\frac{\overline{F_{\theta}^{i}}}{\lambda_{p}+\mu_{i}}}\right)=\gamma(\theta)\sum_{i}{\frac{{\overline{F_{\theta}^{i}}}L^{*}_{i}}{\lambda_{p}+\mu_{i}}}+{\frac{b(\theta)}{\overline{F}(\theta)}\sum_{i}{\frac{\overline{F_{\theta}^{i}}}{\lambda_{p}+\mu_{i}}}},\quad p=1,...,N.</td>
<td></td>
<td>(19)</td>
</tr>
</tbody>
</table>
Unfortunately, the system ([19](#S4.E19 "In Proposition 6. ‣ 4.2 Linear system based solution ‣ 4 Hyper-exponential job size distribution with more than two phases")) does not seem to have a tractable finite form analytic solution. Therefore, in the ensuing subsections we proposed an alternative solution based on an operator series and construct a tight upper bound.
### 4.3 Operator series form for the expected sojourn time
Since the operator $`\Phi_{1}`$ is a contraction \[[3](#bib.bib3), [4](#bib.bib4)\], we can iterate equation ([17](#S4.E17 "In 4.1 Notations ‣ 4 Hyper-exponential job size distribution with more than two phases")) starting from some initial point $`\alpha_{0}'`$. The initial point could be simply a constant. As shown in \[[3](#bib.bib3), [4](#bib.bib4)\] the iterations will converge to the unique solution of ([17](#S4.E17 "In 4.1 Notations ‣ 4 Hyper-exponential job size distribution with more than two phases")). Specifically, we make iterations in the following way:
<table>
<tbody>
<tr>
<td></td>
<td>αn+1′(x)=Φ1(αn′(x))+b(θ)F¯(θ)F¯(x+θ)+1,n=0,1,2,…formulae-sequencesubscriptsuperscript𝛼′𝑛1𝑥subscriptΦ1subscriptsuperscript𝛼′𝑛𝑥𝑏𝜃¯𝐹𝜃¯𝐹𝑥𝜃1𝑛012…{}{\alpha^{\prime}_{n+1}(x)=\Phi_{1}(\alpha^{\prime}_{n}(x))}{+\frac{b(\theta)}{\overline{F}(\theta)}\overline{F}(x+\theta)+1},\quad n=0,1,2,...</td>
<td></td>
<td>(20)</td>
</tr>
</tbody>
</table>
At every iteration step we construct the following approximation of $`{\overline{T}}^{B\hspace{0pt}P\hspace{0pt}S}\hspace{0pt}{(\theta)}`$ according to ([18](#S4.E18 "In 4.1 Notations ‣ 4 Hyper-exponential job size distribution with more than two phases")):
<table>
<tbody>
<tr>
<td></td>
<td>T¯n+1BPS(θ)=Φ2(αn+1′(x)).subscriptsuperscript¯𝑇𝐵𝑃𝑆𝑛1𝜃subscriptΦ2subscriptsuperscript𝛼′𝑛1𝑥{}{\overline{T}^{BPS}_{n+1}(\theta)=\Phi_{2}(\alpha^{\prime}_{n+1}(x))}.</td>
<td></td>
<td>(21)</td>
</tr>
</tbody>
</table>
Using ([20](#S4.E20 "In 4.3 Operator series form for the expected sojourn time ‣ 4 Hyper-exponential job size distribution with more than two phases")) and ([21](#S4.E21 "In 4.3 Operator series form for the expected sojourn time ‣ 4 Hyper-exponential job size distribution with more than two phases")), we can construct the operator series expression for the expected sojourn time in the TLPS system.
######
Theorem 7
.
The expected sojourn time $`\overline{T}\hspace{0pt}{(\theta)}`$ in the TLPS system with the hyper-exponential job size distribution is given by
<table>
<tbody>
<tr>
<td></td>
<td>T¯(θ)=Xθ1¯+W¯(θ)F¯(θ)1−ρθ+m−Xθ1¯1−ρ+b(θ)F¯(θ)(1−ρθ)(∑i=0∞Φ2(Φ1i(F¯(x+θ)))).¯𝑇𝜃¯superscriptsubscript𝑋𝜃1¯𝑊𝜃¯𝐹𝜃1subscript𝜌𝜃𝑚¯superscriptsubscript𝑋𝜃11𝜌𝑏𝜃¯𝐹𝜃1subscript𝜌𝜃superscriptsubscript𝑖0subscriptΦ2superscriptsubscriptΦ1𝑖¯𝐹𝑥𝜃\displaystyle{\overline{T}(\theta)}=\frac{{\overline{X_{\theta}^{1}}}+\overline{W}(\theta)\overline{F}(\theta)}{1-\rho_{\theta}}+{\frac{m-{\overline{X_{\theta}^{1}}}}{1-\rho}}+{\frac{b(\theta)}{\overline{F}(\theta)(1-\rho_{\theta})}}{\left(\sum_{i=0}^{\infty}{\Phi_{2}\left(\Phi_{1}^{i}(\overline{F}(x+\theta))\right)}\right)}.</td>
<td></td>
<td>(22)</td>
</tr>
</tbody>
</table>
###### Proof..
From ([20](#S4.E20 "In 4.3 Operator series form for the expected sojourn time ‣ 4 Hyper-exponential job size distribution with more than two phases")) we have
<table>
<tbody>
<tr>
<td></td>
<td>αn′=qnα0′+∑i=1n−1qi+b(θ)F¯(θ)∑i=1n−1Φ1i(F¯(x+θ))+b(θ)F¯(θ)F¯(x+θ)+1,subscriptsuperscript𝛼′𝑛superscript𝑞𝑛subscriptsuperscript𝛼′0superscriptsubscript𝑖1𝑛1superscript𝑞𝑖𝑏𝜃¯𝐹𝜃superscriptsubscript𝑖1𝑛1superscriptsubscriptΦ1𝑖¯𝐹𝑥𝜃𝑏𝜃¯𝐹𝜃¯𝐹𝑥𝜃1\displaystyle\alpha^{\prime}_{n}=q^{n}\alpha^{\prime}_{0}+\sum_{i=1}^{n-1}{q^{i}}+\frac{b(\theta)}{\overline{F}(\theta)}\sum_{i=1}^{n-1}{\Phi_{1}^{i}(\overline{F}(x+\theta))}+\frac{b(\theta)}{\overline{F}(\theta)}\overline{F}(x+\theta)+1,</td>
<td></td>
</tr>
</tbody>
</table>
and then from ([21](#S4.E21 "In 4.3 Operator series form for the expected sojourn time ‣ 4 Hyper-exponential job size distribution with more than two phases")) and ([14](#S4.E14 "In 4.1 Notations ‣ 4 Hyper-exponential job size distribution with more than two phases")) it follows, that
<table>
<tbody>
<tr>
<td></td>
<td>T¯nBPS(θ)=(m−Xθ1¯)(qnα0′+∑i=0n−1qi)+b(θ)F¯(θ)(Φ2(∑i=0n−1Φ1i(F¯(x+θ)))).subscriptsuperscript¯𝑇𝐵𝑃𝑆𝑛𝜃𝑚¯superscriptsubscript𝑋𝜃1superscript𝑞𝑛subscriptsuperscript𝛼′0superscriptsubscript𝑖0𝑛1superscript𝑞𝑖𝑏𝜃¯𝐹𝜃subscriptΦ2superscriptsubscript𝑖0𝑛1superscriptsubscriptΦ1𝑖¯𝐹𝑥𝜃\displaystyle\overline{T}^{BPS}_{n}(\theta)=(m-{\overline{X_{\theta}^{1}}})\left(q^{n}\alpha^{\prime}_{0}+\sum_{i=0}^{n-1}{q^{i}}\right)+\frac{b(\theta)}{\overline{F}(\theta)}\left(\Phi_{2}\left(\sum_{i=0}^{n-1}{\Phi_{1}^{i}(\overline{F}(x+\theta))}\right)\right).</td>
<td></td>
</tr>
</tbody>
</table>
Using the facts (see ([16](#S4.E16 "In 4.1 Notations ‣ 4 Hyper-exponential job size distribution with more than two phases"))):
<table>
<tbody>
<tr>
<td></td>
<td>1.q<ρ<1⟹qn→0asn→∞,formulae-sequence1𝑞𝜌1⟹superscript𝑞𝑛→0as𝑛→\displaystyle 1.\,q<\rho<1\Longrightarrow q^{n}\rightarrow 0\ \mbox{as}\ n\rightarrow\infty,</td>
<td></td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td>2.∑i=0∞qi=11−q=1−ρθ1−ρ,formulae-sequence2superscriptsubscript𝑖0superscript𝑞𝑖11𝑞1subscript𝜌𝜃1𝜌\displaystyle 2.\,\sum_{i=0}^{\infty}{q^{i}}=\frac{1}{1-q}=\frac{1-\rho_{\theta}}{1-\rho},</td>
<td></td>
</tr>
</tbody>
</table>
we conclude that
<table>
<tbody>
<tr>
<td></td>
<td>T¯BPS(θ)=limn→∞T¯nBPS(θ)=(m−Xθ1¯)(1−ρθ1−ρ)+b(θ)F¯(θ)(∑i=0∞Φ2(Φ1i(F¯(x+θ)))).superscript¯𝑇𝐵𝑃𝑆𝜃subscript→𝑛subscriptsuperscript¯𝑇𝐵𝑃𝑆𝑛𝜃𝑚¯superscriptsubscript𝑋𝜃11subscript𝜌𝜃1𝜌𝑏𝜃¯𝐹𝜃superscriptsubscript𝑖0subscriptΦ2superscriptsubscriptΦ1𝑖¯𝐹𝑥𝜃\displaystyle{\overline{T}^{BPS}(\theta)}=\lim_{n\rightarrow\infty}{\overline{T}^{BPS}_{n}(\theta)={{(m-{\overline{X_{\theta}^{1}}})}}\left(\frac{1-\rho_{\theta}}{1-\rho}\right)+{\frac{b(\theta)}{\overline{F}(\theta)}}\left({\sum_{i=0}^{\infty}{\Phi_{2}\left(\Phi_{1}^{i}(\overline{F}(x+\theta))\right)}}\right)}.</td>
<td></td>
</tr>
</tbody>
</table>
Finally, using ([8](#S2.E8 "In 2.2 The expected sojourn time in TLPS system ‣ 2 Model description")) we obtain ([22](#S4.E22 "In Theorem 7. ‣ 4.3 Operator series form for the expected sojourn time ‣ 4 Hyper-exponential job size distribution with more than two phases")). ∎
The resulting formula ([22](#S4.E22 "In Theorem 7. ‣ 4.3 Operator series form for the expected sojourn time ‣ 4 Hyper-exponential job size distribution with more than two phases")) is difficult to analyze and does not have a clear analytic expression. Using this result in the next subsection we find an approximation,which is also an upper bound, of the expected sojourn time function in a more explicit form.
### 4.4 Upper bound for the expected sojourn time
Let us start with auxiliary results.
######
Lemma 8
.
For any function $`{\beta\hspace{0pt}{(x)}} \geq 0`$ with $`\beta_{j} = {\int_{0}^{\infty}{\beta\hspace{0pt}{(x)}\hspace{0pt}e^{- {\mu_{i}\hspace{0pt}x}}\hspace{0pt}{dx}}}`$,
<table>
<tbody>
<tr>
<td></td>
<td>ifd(βjμj)dμj≥0,j=1,…,N it follows, that Φ2(Φ1(β(x)))≤qΦ2(β(x)).formulae-sequenceif𝑑subscript𝛽𝑗subscript𝜇𝑗𝑑subscript𝜇𝑗0formulae-sequence𝑗1…𝑁 it follows, that subscriptΦ2subscriptΦ1𝛽𝑥𝑞subscriptΦ2𝛽𝑥\textrm{if}\quad\frac{d(\beta_{j}\mu_{j})}{d\mu_{j}}\geq 0,\quad j=1,...,N\quad\textrm{ it follows, that }\quad{\Phi_{2}\left({\Phi_{1}}({\beta}(x))\right)}\leq q\Phi_{2}\left({\beta}(x)\right).</td>
<td></td>
</tr>
</tbody>
</table>
###### Proof..
See Appendix. ∎
######
Lemma 9
.
For the TLPS system with the hyper-exponential job size distribution the following statement holds:
<table>
<tbody>
<tr>
<td></td>
<td>Φ2(Φ1(α′(x)))≤qΦ2(α′(x)).subscriptΦ2subscriptΦ1superscript𝛼′𝑥𝑞subscriptΦ2superscript𝛼′𝑥\displaystyle\Phi_{2}\left({\Phi_{1}}({\alpha}^{\prime}(x))\right)\leq q\Phi_{2}\left({\alpha}^{\prime}(x)\right).</td>
<td></td>
<td>(23)</td>
</tr>
</tbody>
</table>
###### Proof..
We define $`{\alpha_{j}' = {\int_{0}^{\infty}{\alpha'\hspace{0pt}{(x)}\hspace{0pt}e^{- {\mu_{j}\hspace{0pt}x}}\hspace{0pt}{dx}}}},{j = {1,{\ldots\hspace{0pt}N}}}`$. As was shown in \[[14](#bib.bib14)\], $`\alpha'\hspace{0pt}{(x)}`$ has the following structure:
<table>
<tbody>
<tr>
<td></td>
<td>α′(x)=a0+∑kake−bkx,a0≥0,ak≥0,bk>0,k=1,…,N.formulae-sequencesuperscript𝛼′𝑥subscript𝑎0subscript𝑘subscript𝑎𝑘superscript𝑒subscript𝑏𝑘𝑥formulae-sequencesubscript𝑎00formulae-sequencesubscript𝑎𝑘0formulae-sequencesubscript𝑏𝑘0𝑘1…𝑁\displaystyle\alpha^{\prime}(x)=a_{0}+\sum_{k}{a_{k}e^{-b_{k}x}},\quad a_{0}\geq 0,\,a_{k}\geq 0,\,b_{k}>0,\quad k=1,...,N.</td>
<td></td>
</tr>
</tbody>
</table>
Then, we have that $`{\alpha'\hspace{0pt}{(x)}} \geq 0`$ and
<table>
<tbody>
<tr>
<td></td>
<td></td>
<td></td>
<td>αj′=a0μj+∑kakbk+μj,j=1,…,N,formulae-sequencesubscriptsuperscript𝛼′𝑗subscript𝑎0subscript𝜇𝑗subscript𝑘subscript𝑎𝑘subscript𝑏𝑘subscript𝜇𝑗𝑗1…𝑁\displaystyle\alpha^{\prime}_{j}=\frac{a_{0}}{\mu_{j}}+\sum_{k}{\frac{a_{k}}{b_{k}+\mu_{j}}},\quad j=1,...,N,</td>
<td></td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td>⟹⟹\displaystyle\Longrightarrow</td>
<td></td>
<td>d(αj′μj)dμj=∑kakbk+μj−∑kakμj(bk+μj)2=∑kakbk(bk+μj)2≥0,j=1,…,N,formulae-sequence𝑑subscriptsuperscript𝛼′𝑗subscript𝜇𝑗𝑑subscript𝜇𝑗subscript𝑘subscript𝑎𝑘subscript𝑏𝑘subscript𝜇𝑗subscript𝑘subscript𝑎𝑘subscript𝜇𝑗superscriptsubscript𝑏𝑘subscript𝜇𝑗2subscript𝑘subscript𝑎𝑘subscript𝑏𝑘superscriptsubscript𝑏𝑘subscript𝜇𝑗20𝑗1…𝑁\displaystyle\frac{d(\alpha^{\prime}_{j}\mu_{j})}{d\mu_{j}}=\sum_{k}{\frac{a_{k}}{b_{k}+\mu_{j}}}-\sum_{k}{\frac{a_{k}\mu_{j}}{(b_{k}+\mu_{j})^{2}}}=\sum_{k}{\frac{a_{k}b_{k}}{(b_{k}+\mu_{j})^{2}}}\geq 0,\quad j=1,...,N,</td>
<td></td>
</tr>
</tbody>
</table>
as $`{a_{k} \geq 0},{{b_{k} > 0},{k = {1,\ldots,N}}}`$. So, then, according to Lemma [8](#ThmTheorem8 "Lemma 8. ‣ 4.4 Upper bound for the expected sojourn time ‣ 4 Hyper-exponential job size distribution with more than two phases") we have (LABEL:eq:Phi_leq_Phi). ∎
Let us state the following Theorem:
######
Theorem 10
.
An upper bound for the expected sojourn time $`\overline{T}\hspace{0pt}{(\theta)}`$ in the TLPS system with the hyper-exponential job size distribution function with many phases is given by $`\overline{\Upsilon}\hspace{0pt}{(\theta)}`$:
<table>
<tbody>
<tr>
<td></td>
<td>T¯(θ)≤Υ¯(θ)=Xθ1¯+W¯(θ)F¯(θ)1−ρθ+m−Xθ1¯1−ρ+b(θ)F¯(θ)(1−ρ)∑i,jFθi¯Fθj¯μi+μj.¯𝑇𝜃¯Υ𝜃¯superscriptsubscript𝑋𝜃1¯𝑊𝜃¯𝐹𝜃1subscript𝜌𝜃𝑚¯superscriptsubscript𝑋𝜃11𝜌𝑏𝜃¯𝐹𝜃1𝜌subscript𝑖𝑗¯superscriptsubscript𝐹𝜃𝑖¯superscriptsubscript𝐹𝜃𝑗subscript𝜇𝑖subscript𝜇𝑗\displaystyle{\overline{T}(\theta)}\leq{\overline{\Upsilon}(\theta)}=\frac{{\overline{X_{\theta}^{1}}}+\overline{W}(\theta)\overline{F}(\theta)}{1-\rho_{\theta}}+{\frac{m-{\overline{X_{\theta}^{1}}}}{1-\rho}+{\frac{b(\theta)}{\overline{F}(\theta)(1-\rho)}}}{\sum_{i,j}{\frac{\overline{F_{\theta}^{i}}\,\overline{F_{\theta}^{j}}}{\mu_{i}+\mu_{j}}}}.</td>
<td></td>
<td>(24)</td>
</tr>
</tbody>
</table>
###### Proof..
According to the recursion ([20](#S4.E20 "In 4.3 Operator series form for the expected sojourn time ‣ 4 Hyper-exponential job size distribution with more than two phases")) we have for $`\alpha_{n}'\hspace{0pt}{(x)}`$ we approximate $`\alpha'\hspace{0pt}{(x)}`$ with the function $`{\overset{\sim}{\alpha}}'\hspace{0pt}{(x)}`$, which satisfies the following equation:
<table>
<tbody>
<tr>
<td></td>
<td>α~′(x)=α~′(x)Φ1(1)+b(θ)F¯(θ)F¯(x+θ)+1.superscript~𝛼′𝑥superscript~𝛼′𝑥subscriptΦ11𝑏𝜃¯𝐹𝜃¯𝐹𝑥𝜃1\displaystyle{\widetilde{\alpha}^{\prime}(x)}={\widetilde{\alpha}^{\prime}(x)\Phi_{1}(1)+\frac{b(\theta)}{\overline{F}(\theta)}\overline{F}(x+\theta)+1}.</td>
<td></td>
</tr>
</tbody>
</table>
Then, according to ([14](#S4.E14 "In 4.1 Notations ‣ 4 Hyper-exponential job size distribution with more than two phases")) we can find the analytical expression for $`{\overset{\sim}{\alpha}}'\hspace{0pt}{(x)}`$:
<table>
<tbody>
<tr>
<td></td>
<td></td>
<td></td>
<td>α~′(x)=qα~′(x)+b(θ)F¯(θ)F¯(x+θ)+1,superscript~𝛼′𝑥𝑞superscript~𝛼′𝑥𝑏𝜃¯𝐹𝜃¯𝐹𝑥𝜃1\displaystyle{\widetilde{\alpha}^{\prime}(x)}=q{\widetilde{\alpha}^{\prime}(x)+\frac{b(\theta)}{\overline{F}(\theta)}\overline{F}(x+\theta)+1},</td>
<td></td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td>⟹⟹\displaystyle\Longrightarrow</td>
<td></td>
<td>α~′(x)=11−q(b(θ)F¯(θ)F¯(x+θ)+1).superscript~𝛼′𝑥11𝑞𝑏𝜃¯𝐹𝜃¯𝐹𝑥𝜃1\displaystyle{\widetilde{\alpha}^{\prime}(x)}=\frac{1}{1-q}\left(\frac{b(\theta)}{\overline{F}(\theta)}\overline{F}(x+\theta)+1\right).</td>
<td></td>
</tr>
</tbody>
</table>
We take $`{{\overline{\Upsilon}}^{B\hspace{0pt}P\hspace{0pt}S}\hspace{0pt}{(\theta)}} = {\Phi_{2}\hspace{0pt}{({{\overset{\sim}{\alpha}}'\hspace{0pt}{(x)}})}}`$ as an approximation for $`{{\overline{T}}^{B\hspace{0pt}P\hspace{0pt}S}\hspace{0pt}{(\theta)}} = {\Phi_{2}\hspace{0pt}{({\alpha'\hspace{0pt}{(x)}})}}`$. Then
<table>
<tbody>
<tr>
<td></td>
<td>Υ¯BPS(θ)=Φ2(α~′(x))=(m−Xθ1¯)1−q+b(θ)F¯(θ)Φ2(F¯(x+θ))=(m−Xθ1¯)1−q+b(θ)F¯(θ)∑i,jFθi¯Fθj¯μi+μj.superscript¯Υ𝐵𝑃𝑆𝜃subscriptΦ2superscript~𝛼′𝑥𝑚¯superscriptsubscript𝑋𝜃11𝑞𝑏𝜃¯𝐹𝜃subscriptΦ2¯𝐹𝑥𝜃𝑚¯superscriptsubscript𝑋𝜃11𝑞𝑏𝜃¯𝐹𝜃subscript𝑖𝑗¯superscriptsubscript𝐹𝜃𝑖¯superscriptsubscript𝐹𝜃𝑗subscript𝜇𝑖subscript𝜇𝑗\displaystyle{\overline{\Upsilon}^{BPS}(\theta)}=\Phi_{2}(\widetilde{\alpha}^{\prime}(x))=\frac{(m-\overline{X_{\theta}^{1}})}{1-q}+\frac{b(\theta)}{\overline{F}(\theta)}{\Phi_{2}(\overline{F}(x+\theta))}=\frac{(m-\overline{X_{\theta}^{1}})}{1-q}+\frac{b(\theta)}{\overline{F}(\theta)}{\sum_{i,j}{\frac{\overline{F_{\theta}^{i}}\,\overline{F_{\theta}^{j}}}{\mu_{i}+\mu_{j}}}}.</td>
<td></td>
</tr>
</tbody>
</table>
Let us prove, that
<table>
<tbody>
<tr>
<td></td>
<td>T¯BPS(θ)≤Υ¯BPS(θ),superscript¯𝑇𝐵𝑃𝑆𝜃superscript¯Υ𝐵𝑃𝑆𝜃\displaystyle\overline{T}^{BPS}(\theta)\leq\overline{\Upsilon}^{BPS}(\theta),</td>
<td></td>
</tr>
</tbody>
</table>
or equivalently
<table>
<tbody>
<tr>
<td></td>
<td>T¯BPS(θ)−Υ¯BPS(θ)=Φ2(α′(x))−Φ2(α~′(x))≤0.superscript¯𝑇𝐵𝑃𝑆𝜃superscript¯Υ𝐵𝑃𝑆𝜃subscriptΦ2superscript𝛼′𝑥subscriptΦ2superscript~𝛼′𝑥0\displaystyle\overline{T}^{BPS}(\theta)-\overline{\Upsilon}^{BPS}(\theta)=\Phi_{2}(\alpha^{\prime}(x))-\Phi_{2}(\widetilde{\alpha}^{\prime}(x))\leq 0.</td>
<td></td>
</tr>
</tbody>
</table>
Let us look at
<table>
<tbody>
<tr>
<td></td>
<td>Φ2(α′(x))−Φ2(α~′(x))=subscriptΦ2superscript𝛼′𝑥subscriptΦ2superscript~𝛼′𝑥absent\displaystyle\Phi_{2}(\alpha^{\prime}(x))-\Phi_{2}(\widetilde{\alpha}^{\prime}(x))=</td>
<td></td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td>=Φ2(Φ1(α′(x)))+Φ2(b(θ)F¯(θ)F¯(x+θ)+1)−(qΦ2(α~′(x))+Φ2(b(θ)F¯(θ)F¯(x+θ)+1))absentsubscriptΦ2subscriptΦ1superscript𝛼′𝑥subscriptΦ2𝑏𝜃¯𝐹𝜃¯𝐹𝑥𝜃1𝑞subscriptΦ2superscript~𝛼′𝑥subscriptΦ2𝑏𝜃¯𝐹𝜃¯𝐹𝑥𝜃1\displaystyle=\Phi_{2}(\Phi_{1}(\alpha^{\prime}(x)))+\Phi_{2}\left(\frac{b(\theta)}{\overline{F}(\theta)}\overline{F}(x+\theta)+1\right)-\left(q\Phi_{2}({\widetilde{\alpha}^{\prime}(x)})+\Phi_{2}\left(\frac{b(\theta)}{\overline{F}(\theta)}\overline{F}(x+\theta)+1\right)\right)</td>
<td></td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td>=Φ2(Φ1(α′(x)))−qΦ2(α′(x))+q(Φ2(α′(x))−Φ2(α~′(x)))absentsubscriptΦ2subscriptΦ1superscript𝛼′𝑥𝑞subscriptΦ2superscript𝛼′𝑥𝑞subscriptΦ2superscript𝛼′𝑥subscriptΦ2superscript~𝛼′𝑥\displaystyle=\Phi_{2}(\Phi_{1}(\alpha^{\prime}(x)))-q\Phi_{2}({{\alpha}^{\prime}(x)})+q\left(\Phi_{2}({{\alpha}^{\prime}(x)})-\Phi_{2}({\widetilde{\alpha}^{\prime}(x)})\right)</td>
<td></td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td>⟹⟹\displaystyle\Longrightarrow</td>
<td></td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td>Φ2(α′(x))−Φ2(α~′(x))=11−q(Φ2(Φ1(α′(x)))−qΦ2(α′(x))).subscriptΦ2superscript𝛼′𝑥subscriptΦ2superscript~𝛼′𝑥11𝑞subscriptΦ2subscriptΦ1superscript𝛼′𝑥𝑞subscriptΦ2superscript𝛼′𝑥\displaystyle\Phi_{2}({{\alpha}^{\prime}(x)})-\Phi_{2}({\widetilde{\alpha}^{\prime}(x)})=\frac{1}{1-q}\left(\Phi_{2}(\Phi_{1}(\alpha^{\prime}(x)))-q\Phi_{2}({{\alpha}^{\prime}(x)})\right).</td>
<td></td>
</tr>
</tbody>
</table>
And from Lemma [9](#ThmTheorem9 "Lemma 9. ‣ 4.4 Upper bound for the expected sojourn time ‣ 4 Hyper-exponential job size distribution with more than two phases") and formula ([8](#S2.E8 "In 2.2 The expected sojourn time in TLPS system ‣ 2 Model description")) we conclude that ([24](#S4.E24 "In Theorem 10. ‣ 4.4 Upper bound for the expected sojourn time ‣ 4 Hyper-exponential job size distribution with more than two phases")) is true. ∎
In this subsection we found the analytical expression of the upper bound of the expected sojourn time in the case when the job size distribution is a hyper-exponential function with many phases. In the experimental results of the following subsection we show that the obtained upper bound is also a close approximation. The analytic expression of the upper bound which we obtained is more clear and easier to analyze then the expression of the expected sojourn time. It could be used in the future research on TLPS model.
### 4.5 Experimental results
We calculate $`\overline{T}\hspace{0pt}{(\theta)}`$ and $`\overline{\Upsilon}\hspace{0pt}{(\theta)}`$ for different numbers of phases $`N`$ of the job size distribution function. We take $`N = {10,100,500,1000}`$. To calculate $`\overline{T}\hspace{0pt}{(\theta)}`$ we find the numerical solution of the system of linear equations ([19](#S4.E19 "In Proposition 6. ‣ 4.2 Linear system based solution ‣ 4 Hyper-exponential job size distribution with more than two phases")) using the Gauss method. Then using the result of Proposition [6](#ThmTheorem6 "Proposition 6. ‣ 4.2 Linear system based solution ‣ 4 Hyper-exponential job size distribution with more than two phases") we find $`\overline{T}\hspace{0pt}{(\theta)}`$. For $`\overline{\Upsilon}\hspace{0pt}{(\theta)}`$ we use equation ([24](#S4.E24 "In Theorem 10. ‣ 4.4 Upper bound for the expected sojourn time ‣ 4 Hyper-exponential job size distribution with more than two phases")).
As was mentioned in Subsection [2.1](#S2.SS1 "2.1 Main definitions ‣ 2 Model description"), by using the hyper-exponential distribution with many phases, one can approximate a heavy-tailed distribution. In our numerical experiments, we fix $`\rho`$, $`m`$, and select $`p_{i}`$ and $`\mu_{i}`$ in a such a way, that by increasing the number of phases we let the second moment $`d`$ (see (LABEL:eq:m_d)) increase as well. Here we take
<table>
<tbody>
<tr>
<td></td>
<td>ρ=10/11,λ=0.5,pi=νi2.5,μi=ηi1.2,i=1,…,N.formulae-sequence𝜌1011formulae-sequence𝜆0.5formulae-sequencesubscript𝑝𝑖𝜈superscript𝑖2.5formulae-sequencesubscript𝜇𝑖𝜂superscript𝑖1.2𝑖1…𝑁\displaystyle\rho=10/11,\,\,\lambda=0.5,\,\,p_{i}=\frac{\nu}{i^{2.5}},\,\,\mu_{i}=\frac{\eta}{i^{1.2}},\quad i=1,...,N.</td>
<td></td>
</tr>
</tbody>
</table>
In particular, we have
<table>
<tbody>
<tr>
<td></td>
<td>∑ipi=1,⟹ν=1∑ii−2.5,\displaystyle\sum_{i}{p_{i}}=1,\quad\Longrightarrow\,\,\nu=\frac{1}{\sum_{i}{i^{-2.5}}},</td>
<td></td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td>∑ipiμi=m,⟹η=νm∑ii−1.3.\displaystyle\sum_{i}{\frac{p_{i}}{\mu_{i}}}=m,\,\,\,\Longrightarrow\,\,\eta={\frac{\nu}{m}}\,{\sum_{i}{i^{-1.3}}}.</td>
<td></td>
</tr>
</tbody>
</table>
In Figure [4](#S4.F4 "Figure 4 ‣ 4.5 Experimental results ‣ 4 Hyper-exponential job size distribution with more than two phases") one can see the plots of the expected sojourn time and its upper bound as functions of $`\theta`$ when $`N`$ varies from 10 up to 1000. In Figure [4](#S4.F4 "Figure 4 ‣ 4.5 Experimental results ‣ 4 Hyper-exponential job size distribution with more than two phases") we plot the relative error of the upper bound
<table>
<tbody>
<tr>
<td></td>
<td>Δ(θ)=Υ¯(θ)−T¯(θ)T¯(θ),Δ𝜃¯Υ𝜃¯𝑇𝜃¯𝑇𝜃\displaystyle\Delta(\theta)=\frac{\overline{\Upsilon}(\theta)-\overline{T}(\theta)}{\overline{T}(\theta)},</td>
<td></td>
</tr>
</tbody>
</table>
when $`N`$ varies from 10 up to 1000. As one can see, the upper bound ([24](#S4.E24 "In Theorem 10. ‣ 4.4 Upper bound for the expected sojourn time ‣ 4 Hyper-exponential job size distribution with more than two phases")) is very tight.
We find the maximum gain of the expected sojourn time of the TLPS system with respect to the standard PS system. The gain is given by $`{g\hspace{0pt}{(\theta)}} = \frac{{\overline{T}}^{P\hspace{0pt}S} - {\overline{T}\hspace{0pt}{(\theta)}}}{{\overline{T}}^{P\hspace{0pt}S}}`$. Here $`{\overline{T}}^{P\hspace{0pt}S}`$ is an expected sojourn time in the standard PS system. Let us notice, that $`{\overline{T}}^{P\hspace{0pt}S} = {\overline{T}\hspace{0pt}{(0)}}`$.
The data and results are summarized in Table [1](#S4.T1 "Table 1 ‣ 4.5 Experimental results ‣ 4 Hyper-exponential job size distribution with more than two phases").
N $`\eta`$ d $`\theta_{o\hspace{0pt}p\hspace{0pt}t}`$ $`{\max_{\theta}g}\hspace{0pt}{(\theta)}`$ $`{\max_{\theta}\Delta}\hspace{0pt}{(\theta)}`$
10 0.95 7.20 5 32.98$`\%`$ 0.0640
100 1.26 32.28 12 45.75$`\%`$ 0.0807
500 1.40 113.31 21 49.26$`\%`$ 0.0766
1000 1.44 200.04 26 50.12$`\%`$ 0.0743
Table 1: Increasing the number of phases
With the increasing number of phases we observe that
1\. the second moment $`d`$ increases;
2\. the maximum gain $`{\max_{\theta}g}\hspace{0pt}{(\theta)}`$ in expected sojourn time in comparison with PS increases;
3\. the relative error of the upper bound $`\Delta\hspace{0pt}{(\theta)}`$ with the expected sojourn time decreases after the number of phases becomes sufficiently large;
4\. the sensitivity of the system performance with respect to the selection of the sub-optimal threshold value decreases.
Thus the TLPS system produces better and more robust performance as the variance of the job size distribution increases.
Figure 3: The expected sojourn time $`\overline{T}\hspace{0pt}{(\theta)}`$ and its upper bound $`\overline{\Upsilon}{(\theta)})`$ for $`N = 10`$, $`100`$, $`500`$, $`1000`$
Figure 4: The relative error $`{\Delta\hspace{0pt}{(\theta)}} = {{{({{\overline{T}\hspace{0pt}{(\theta)}} - {\overline{\Upsilon}\hspace{0pt}{(\theta)}}})}/\overline{T}}\hspace{0pt}{(\theta)}}`$ for $`N = 10`$, $`100`$, $`500`$, $`1000`$
## 5 Conclusion
We analyze the TLPS scheduling mechanism with the hyper-exponential job size distribution function.
In Section [3](#S3 "3 Hyper-exponential job size distribution with two phases") we analyze the system when the job size distribution function has two phases and find the analytical expressions of the expected conditional sojourn time and the expected sojourn time of the TLPS system.
Connections in the Internet belong to two distinct classes: short HTTP and P2P signaling connections and long downloads such as: PDF, MP3, and so on. Thus, according to this observation, we consider a special selection of the parameters of the job size distribution function with two phases and find the approximation of the optimal threshold, when the variance of the job size distribution goes to infinity.
We show, that the approximated value of the threshold tends to the optimal threshold, when the second moment of the distribution function goes to infinity. We found that the gain of the TLPS system compared to the standard PS system could reach 45% when the load of the system increases. Also the system is not too sensitive to the selection of the optimal value of the threshold.
In Section [4](#S4 "4 Hyper-exponential job size distribution with more than two phases") we have studied the TLPS model when the job size distribution is a hyper-exponential function with many phases. We provide an expression of the expected conditional sojourn time as a solution of the system of linear equations. Also we apply the iteration method to find the expression of the expected conditional sojourn time in the form of operator series and using the obtained expression we provide an upper bound for the expected sojourn time function. With the experimental results we show that the upper bound is very tight and could be used as an approximation of the expected sojourn time function. We show numerically, that the relative error between the upper bound and expected sojourn time function decreases when the variation of the job size distribution function increases. The obtained upper bound could be used to identify an approximation of the optimal value of the optimal threshold for TLPS system when the job size distribution is heavy-tailed.
We study the properties of the expected sojourn time function, when the parameters of the job size distribution function are selected in such a way, that it approximates a heavy-tailed distribution as the number of phases of the job size distribution increases. As the number of phases increases the gain of the TLPS system compared with the standard PS system increases and the sensitivity of the system with respect to the selection of the optimal threshold decreases.
## 6 Appendix: Proof of Lemma [8](#ThmTheorem8 "Lemma 8. ‣ 4.4 Upper bound for the expected sojourn time ‣ 4 Hyper-exponential job size distribution with more than two phases")
Let us take any function $`{\beta\hspace{0pt}{(x)}} > 0`$ and define $`{{\beta_{j} = {\int_{0}^{\infty}{\beta\hspace{0pt}{(x)}\hspace{0pt}e^{- {\mu_{j}\hspace{0pt}x}}\hspace{0pt}{dx}}}},{j = {1,\ldots,N}}}.`$ Let us show for $`{\beta\hspace{0pt}{(x)}} \geq 0`$ that if
<table>
<tbody>
<tr>
<td></td>
<td>d(βjμj)dμj≥0,j=1,…,N, then it follows that Φ2(Φ1(β(x)))≤qΦ2(β(x)).formulae-sequence𝑑subscript𝛽𝑗subscript𝜇𝑗𝑑subscript𝜇𝑗0formulae-sequence𝑗1…𝑁 then it follows that subscriptΦ2subscriptΦ1𝛽𝑥𝑞subscriptΦ2𝛽𝑥\displaystyle\frac{d({\beta}_{j}\mu_{j})}{d\mu_{j}}\geq 0,\quad j=1,...,N,\quad\textrm{ then it follows that }\quad{\Phi_{2}\left({\Phi_{1}}({\beta}(x))\right)}\leq q\Phi_{2}\left({\beta}(x)\right).</td>
<td></td>
</tr>
</tbody>
</table>
As
<table>
<tbody>
<tr>
<td></td>
<td>∫0∞∫0xβ(y)F¯(x−y+θ)F¯(x+θ)𝑑y𝑑x=∫0∞∫0∞β(y)F¯(x1+θ)F¯(x1+y+θ)𝑑x1𝑑ysuperscriptsubscript0superscriptsubscript0𝑥𝛽𝑦¯𝐹𝑥𝑦𝜃¯𝐹𝑥𝜃differential-d𝑦differential-d𝑥superscriptsubscript0superscriptsubscript0𝛽𝑦¯𝐹subscript𝑥1𝜃¯𝐹subscript𝑥1𝑦𝜃differential-dsubscript𝑥1differential-d𝑦\displaystyle{\int_{0}^{\infty}\!\!\int_{0}^{x}\!\!{\beta(y)}{\overline{F}(x-y+\theta)}{\overline{F}(x+\theta)}dydx=\int_{0}^{\infty}\!\!\int_{0}^{\infty}\!\!{\beta(y)}{\overline{F}(x_{1}+\theta)}{\overline{F}(x_{1}+y+\theta)}dx_{1}dy}</td>
<td></td>
</tr>
</tbody>
</table>
and
<table>
<tbody>
<tr>
<td></td>
<td>Φ2(Φ1(β(x)))subscriptΦ2subscriptΦ1𝛽𝑥\displaystyle\Phi_{2}(\Phi_{1}(\beta(x)))</td>
<td>=\displaystyle=</td>
<td>γ(θ)∫0∞∫0∞β(y)F¯(x+y+θ)F¯(x+θ)𝑑y𝑑x𝛾𝜃superscriptsubscript0superscriptsubscript0𝛽𝑦¯𝐹𝑥𝑦𝜃¯𝐹𝑥𝜃differential-d𝑦differential-d𝑥\displaystyle{\gamma(\theta)\int_{0}^{\infty}\!\!\int_{0}^{\infty}\beta(y){\overline{F}(x+y+\theta)}{\overline{F}(x+\theta)}dydx}</td>
<td></td>
</tr>
<tr>
<td></td>
<td></td>
<td></td>
<td>+γ(θ)∫0∞∫0xβ(y)F¯(x−y+θ)F¯(x+θ)𝑑y𝑑x,𝛾𝜃superscriptsubscript0superscriptsubscript0𝑥𝛽𝑦¯𝐹𝑥𝑦𝜃¯𝐹𝑥𝜃differential-d𝑦differential-d𝑥\displaystyle+{\gamma(\theta){\int_{0}^{\infty}\!\!\int_{0}^{x}\beta(y){\overline{F}(x-y+\theta)}{\overline{F}(x+\theta)}dydx}},</td>
<td></td>
</tr>
</tbody>
</table>
then
<table>
<tbody>
<tr>
<td></td>
<td>Φ2(Φ1(β(x)))subscriptΦ2subscriptΦ1𝛽𝑥\displaystyle{\Phi_{2}(\Phi_{1}(\beta(x)))}</td>
<td>=\displaystyle=</td>
<td>2γ(θ)∫0∞∫0∞β(x)F¯(x+θ)F¯(x+y+θ)𝑑y𝑑x=2𝛾𝜃superscriptsubscript0superscriptsubscript0𝛽𝑥¯𝐹𝑥𝜃¯𝐹𝑥𝑦𝜃differential-d𝑦differential-d𝑥absent\displaystyle 2{\gamma(\theta)\int_{0}^{\infty}\!\!\int_{0}^{\infty}{\beta(x)}{\overline{F}(x+\theta)}{\overline{F}(x+y+\theta)}dydx}=</td>
<td></td>
</tr>
<tr>
<td></td>
<td></td>
<td>=\displaystyle=</td>
<td>2γ(θ)∫0∞β(x)∑i,jFθi¯Fθj¯μi+μje−μjxdx=2γ(θ)∑i,jFθi¯Fθj¯μi+μjβj.2𝛾𝜃superscriptsubscript0𝛽𝑥subscript𝑖𝑗¯superscriptsubscript𝐹𝜃𝑖¯superscriptsubscript𝐹𝜃𝑗subscript𝜇𝑖subscript𝜇𝑗superscript𝑒subscript𝜇𝑗𝑥𝑑𝑥2𝛾𝜃subscript𝑖𝑗¯superscriptsubscript𝐹𝜃𝑖¯superscriptsubscript𝐹𝜃𝑗subscript𝜇𝑖subscript𝜇𝑗subscript𝛽𝑗\displaystyle 2{\gamma(\theta)\int_{0}^{\infty}{\beta(x)}{\sum_{i,j}{\frac{\overline{F_{\theta}^{i}}\overline{F_{\theta}^{j}}}{\mu_{i}+\mu_{j}}}{e^{-\mu_{j}x}}}dx}=2\gamma(\theta){\sum_{i,j}{\frac{\overline{F_{\theta}^{i}}\overline{F_{\theta}^{j}}}{\mu_{i}+\mu_{j}}}{\beta_{j}}}.</td>
<td></td>
</tr>
</tbody>
</table>
Also for $`\Phi_{2}\hspace{0pt}\left( {\beta\hspace{0pt}{(x)}} \right)`$, taking into account that $`q = {\gamma\hspace{0pt}{(\theta)}\hspace{0pt}{\sum_{i}\frac{\overline{F_{\theta}^{i}}}{\mu_{i}}}}`$, we obtain
<table>
<tbody>
<tr>
<td></td>
<td>qΦ2(β(x))𝑞subscriptΦ2𝛽𝑥\displaystyle q\Phi_{2}\left(\beta(x)\right)</td>
<td>=\displaystyle=</td>
<td>γ(θ)∑iFθi¯μi∑jFθj¯∫0∞β(x)e−μjx𝑑x=γ(θ)∑i,jFθi¯Fθj¯μiβj.𝛾𝜃subscript𝑖¯superscriptsubscript𝐹𝜃𝑖subscript𝜇𝑖subscript𝑗¯superscriptsubscript𝐹𝜃𝑗superscriptsubscript0𝛽𝑥superscript𝑒subscript𝜇𝑗𝑥differential-d𝑥𝛾𝜃subscript𝑖𝑗¯superscriptsubscript𝐹𝜃𝑖¯superscriptsubscript𝐹𝜃𝑗subscript𝜇𝑖subscript𝛽𝑗\displaystyle{\gamma(\theta){\sum_{i}{\frac{\overline{F_{\theta}^{i}}}{\mu_{i}}}}{\sum_{j}{\overline{F_{\theta}^{j}}}{\int_{0}^{\infty}}{\beta(x)}{e^{-\mu_{j}x}}}dx}=\gamma(\theta){\sum_{i,j}{\frac{\overline{F_{\theta}^{i}}\overline{F_{\theta}^{j}}}{\mu_{i}}}\beta_{j}}.</td>
<td></td>
</tr>
</tbody>
</table>
Thus, a sufficient condition for the inequality $`{\Phi_{2}\hspace{0pt}\left( {\Phi_{1}\hspace{0pt}{({\beta\hspace{0pt}{(x)}})}} \right)} \leq {q\hspace{0pt}\Phi_{2}\hspace{0pt}\left( {\beta\hspace{0pt}{(x)}} \right)}`$ to be satisfied is that for every pair $`i,j`$:
<table>
<tbody>
<tr>
<td></td>
<td>2μi+μjβj+2μj+μiβi≤1μiβj+1μjβi⟺−(βjμj−βiμi)(μj−μi)≤0.formulae-sequence2subscript𝜇𝑖subscript𝜇𝑗subscript𝛽𝑗2subscript𝜇𝑗subscript𝜇𝑖subscript𝛽𝑖1subscript𝜇𝑖subscript𝛽𝑗1subscript𝜇𝑗subscript𝛽𝑖⟺subscript𝛽𝑗subscript𝜇𝑗subscript𝛽𝑖subscript𝜇𝑖subscript𝜇𝑗subscript𝜇𝑖0\displaystyle{{\frac{2}{\mu_{i}+\mu_{j}}}\beta_{j}+{\frac{2}{\mu_{j}+\mu_{i}}}\beta_{i}}\leq{\frac{1}{\mu_{i}}}\beta_{j}+{\frac{1}{\mu_{j}}}\beta_{i}\quad\Longleftrightarrow\quad-(\beta_{j}\mu_{j}-\beta_{i}\mu_{i})(\mu_{j}-\mu_{i})\leq 0.</td>
<td></td>
</tr>
</tbody>
</table>
The inequality is indeed satisfied when $`\beta_{j}\hspace{0pt}\mu_{j}`$ is an increasing function of $`\mu_{j}`$. We conclude that $`{\Phi_{2}\hspace{0pt}\left( {\Phi_{1}\hspace{0pt}{({\beta\hspace{0pt}{(x)}})}} \right)} \leq {q\hspace{0pt}\Phi_{2}\hspace{0pt}\left( {\beta\hspace{0pt}{(x)}} \right)}`$, which proves Lemma [8](#ThmTheorem8 "Lemma 8. ‣ 4.4 Upper bound for the expected sojourn time ‣ 4 Hyper-exponential job size distribution with more than two phases").
## References
- \[1\] S. Aalto and U. Ayesta, “Mean delay analysis of multilevel processor sharing disciplines”, In Proceedings of IEEE INFOCOM 2006.
- \[2\] S. Aalto, U. Ayesta, and E.Nyberg, “Two-level processor sharing scheduling disciplines: mean delay analysis”, in Proceedings of ACM SIGMETRICS/Performance 2004.
- \[3\] K. Avrachenkov, U. Ayesta, P. Brown, and E. Nyberg, “Differentiation between short and long TCP flows: Predictability of the response time”, in Proceedings of IEEE INFOCOM 2004.
- \[4\] K. Avrachenkov, U. Ayesta, P. Brown, “Batch Arrival Processor-Sharing with Application to Multi-Level Processor-Sharing Scheduling”, Queuing ystems 50, pp.459-480, 2005.
- \[5\] F. Baccelli and D.R. McDonald, “A stochastic model for the rate of non-persistent TCP flows”, in Proceedings of ValueTools 2006.
- \[6\]
N. Bansal, “Analysis of the M/G/1 processor-sharing queue with bulk arrivals”, Operations Research Letters, v.31, no.5, 2003, pp.401-405.
- \[7\]
R. El Abdouni Khayari, R. Sadre, and B.R. Haverkort, “Fitting world-wide web request traces with the EM-algorithm”, Performance Evaluation, v.52, no.2-3, 2003, pp.175-191.
- \[8\]
A. Feldmann, W. Whitt, ”Fitting mixtures of exponentials to long-tail distributions to analyze network performance models”, Performance Evaluation, v.31, 1998, pp.245-258.
- \[9\]
H. Feng and V. Misra, “Mixed scheduling disciplines for network flows”, ACM SIGMETRICS Performance Evaluation Review, v.31(2), 2003, pp.36-39.
- \[10\] L. Guo and L. Matta, “Differentiated control of web traffic: A numerical analisys, In SPIE ITCOM, Boston, 2002.
- \[11\]
M. Harchol-Balter, B. Schroeder, N. Bansal, and M. Agrawal, “Size-based scheduling to improve web performance”, ACM Transactions on Computer Systems, v.21, no.2, 2003, pp.207-233.
- \[12\]
L. Kleinrock, Queueing systems, vol. 2, John Wiley and Sons, 1976.
- \[13\] Kleinrock, L., R. R. Muntz, and E. Rodemich, “The Processor-Sharing Queueing Model for Time-Shared Systems with Bulk Arrivals,” Networks Journal,1,1-13 (1971).
- \[14\] N. Osipova, “Batch Processor Sharing with Hyper-Exponential service time”, INRIA Research Report RR-6180, 2007. Available at http://hal.inria.fr/inria-00144389
- \[15\]
L.E. Schrage, “A proof of the optimality of the shortest remaining processing time discipline”, Operations Research, v.16, 1968, pp. 687-690.
###### Contents
1. [](#S1)
1 Introduction
2. [](#S2)
2 Model description
1. [](#S2.SS1 "In 2 Model description")
2.1 Main definitions
2. [](#S2.SS2 "In 2 Model description")
2.2 The expected sojourn time in TLPS system
3. [](#S3)
3 Hyper-exponential job size distribution with two phases
1. [](#S3.SS1 "In 3 Hyper-exponential job size distribution with two phases")
3.1 Notations
2. [](#S3.SS2 "In 3 Hyper-exponential job size distribution with two phases")
3.2 Explicit form for the expected sojourn time
3. [](#S3.SS3 "In 3 Hyper-exponential job size distribution with two phases")
3.3 Optimal threshold approximation
4. [](#S3.SS4 "In 3 Hyper-exponential job size distribution with two phases")
3.4 Experimental results
4. [](#S4)
4 Hyper-exponential job size distribution with more than two phases
1. [](#S4.SS1 "In 4 Hyper-exponential job size distribution with more than two phases")
4.1 Notations
2. [](#S4.SS2 "In 4 Hyper-exponential job size distribution with more than two phases")
4.2 Linear system based solution
3. [](#S4.SS3 "In 4 Hyper-exponential job size distribution with more than two phases")
4.3 Operator series form for the expected sojourn time
4. [](#S4.SS4 "In 4 Hyper-exponential job size distribution with more than two phases")
4.4 Upper bound for the expected sojourn time
5. [](#S4.SS5 "In 4 Hyper-exponential job size distribution with more than two phases")
4.5 Experimental results
5. [](#S5)
5 Conclusion
6. [](#S6)
6 Appendix: Proof of Lemma 8 |
0705.0326 | cs.NI | cyber | # Optimal Delay-Throughput Trade-offs in Mobile Ad-Hoc Networks: Hybrid Random Walk and One-Dimensional Mobility Models
<sup>†</sup>
<sup>†</sup>thanks: An earlier version of this paper appeared in the Proc. of ITA Workshop, 2007.
<sup>†</sup>
<sup>†</sup>thanks: Research presented here was supported in part by a Vodafone Fellowship and NSF grant CNS 05-19691.
Lei Ying and R. Srikant
Coordinated Science Lab
and
Department of Electrical and Computer Engineering
University of Illinois at Urbana-Champaign
{lying, rsrikant}@uiuc.edu
###### Abstract
Optimal delay-throughput trade-offs for two-dimensional i.i.d mobility models have been established in \[[23](#bib.bib23)\], where we showed that the optimal trade-offs can be achieved using rate-less codes when the required delay guarantees are sufficient large. In this paper, we extend the results to other mobility models including two-dimensional hybrid random walk model, one-dimensional i.i.d. mobility model and one-dimensional hybrid random walk model. We consider both fast mobiles and slow mobiles, and establish the optimal delay-throughput trade-offs under some conditions. Joint coding-scheduling algorithms are also proposed to achieve the optimal trade-offs.
## I Notations
The following notations are used throughout this paper, given non-negative functions $`f\hspace{0pt}{(n)}`$ and $`g\hspace{0pt}{(n)}`$:
1. \(1\)
$`{f\hspace{0pt}{(n)}} = {O\hspace{0pt}{({g\hspace{0pt}{(n)}})}}`$ means there exist positive constants $`c`$ and $`m`$ such that $`{f\hspace{0pt}{(n)}} \leq {c\hspace{0pt}g\hspace{0pt}{(n)}}`$ for all $`n \geq m`$.
2. \(2\)
$`{f\hspace{0pt}{(n)}} = {\Omega\hspace{0pt}{({g\hspace{0pt}{(n)}})}}`$ means there exist positive constants $`c`$ and $`m`$ such that $`{f\hspace{0pt}{(n)}} \geq {c\hspace{0pt}g\hspace{0pt}{(n)}}`$ for all $`{n \geq m}.`$ Namely, $`{g\hspace{0pt}{(n)}} = {O\hspace{0pt}{({f\hspace{0pt}{(n)}})}}`$.
3. \(3\)
$`{f\hspace{0pt}{(n)}} = {\Theta\hspace{0pt}{({g\hspace{0pt}{(n)}})}}`$ means that both $`{f\hspace{0pt}{(n)}} = {\Omega\hspace{0pt}{({g\hspace{0pt}{(n)}})}}`$ and $`{f\hspace{0pt}{(n)}} = {O\hspace{0pt}{({g\hspace{0pt}{(n)}})}}`$ hold.
4. \(4\)
$`{f\hspace{0pt}{(n)}} = {o\hspace{0pt}{({g\hspace{0pt}{(n)}})}}`$ means that $`{{\lim_{n\rightarrow\infty}{{{f\hspace{0pt}{(n)}}/g}\hspace{0pt}{(n)}}} = 0}.`$
5. \(5\)
$`{f\hspace{0pt}{(n)}} = {\omega\hspace{0pt}{({g\hspace{0pt}{(n)}})}}`$ means that $`{{\lim_{n\rightarrow\infty}{{{g\hspace{0pt}{(n)}}/f}\hspace{0pt}{(n)}}} = 0}.`$ Namely, $`{{g\hspace{0pt}{(n)}} = {o\hspace{0pt}{({f\hspace{0pt}{(n)}})}}}.`$
## II Introduction
Delay-throughput trade-offs in mobile ad-hoc networks have received much attention since the work of Grossglauser and Tse \[[9](#bib.bib9)\], where they showed that the throughput of ad-hoc networks can be significantly improved by exploring the node mobility. Recently the trade-off was investigated under different mobility models, which include the i.i.d. mobility \[[16](#bib.bib16), [21](#bib.bib21), [12](#bib.bib12), [23](#bib.bib23)\], one-dimensional mobility \[[3](#bib.bib3), [8](#bib.bib8)\], random walk \[[5](#bib.bib5), [6](#bib.bib6), [7](#bib.bib7), [19](#bib.bib19)\], hybrid random walk \[[19](#bib.bib19)\] and Brownian motion \[[13](#bib.bib13)\].
In \[[23](#bib.bib23)\], we demonstrated that the optimal trade-offs for two-dimensional i.i.d. mobility models can be achieved using rate-less codes when the required delay guarantees are sufficiently large. In this paper, we extend the results to the two-dimensional hybrid random walk, one-dimensional i.i.d. mobility and one-dimensional hybrid random walk models. The two-dimensional i.i.d. mobility studied in \[[23](#bib.bib23)\] only models the case where the network topology changes dramatically at each time slot. However Markovian mobility dynamics may be more realistic. Thus the two-dimensional hybrid random walk model was introduced by Sharma *et al* in \[[19](#bib.bib19)\], where the unit square is divided into $`1/S^{2}`$ small-squares, and mobiles move from the current small-square to one of its eight adjacent small-squares at the beginning of each time slot (The detailed description of the two-dimensional hybrid random walk model is presented in Section [III](#S3 "III Model ‣ Optimal Delay-Throughput Trade-offs in Mobile Ad-Hoc Networks: Hybrid Random Walk and One-Dimensional Mobility ModelsAn earlier version of this paper appeared in the Proc. of ITA Workshop, 2007.Research presented here was supported in part by a Vodafone Fellowship and NSF grant CNS 05-19691.")). Since the distance each mobile can move is at most $`{2\hspace{0pt}\sqrt{2}}/S`$ at each time slot, we can use different values of $`S`$ to model mobiles with different speeds, so this two-dimensional hybrid random walk model can be used for a wide range of scenarios. Note that the two-dimensional hybrid random walk model is the same as the two-dimensional i.i.d. mobility model when $`{S = 1}.`$ One might wonder why the results in \[[23](#bib.bib23)\] are necessary given the results in this paper. The reason is that the Markovian mobility dynamics in this paper requires a different set of tools than those in \[[23](#bib.bib23)\] and as a result, the trade-off in this paper are applicable only when $`{S = {o\hspace{0pt}{(1)}}}.`$ Thus, the results in \[[23](#bib.bib23)\] cannot be recovered from the results of this paper. We wish to comment that one of the main differences between this paper and \[[23](#bib.bib23)\] is that, the i.i.d. mobility assumption in \[[23](#bib.bib23)\] allows us to use Chernoff bounds to obtain concentration results. However, the random walk and other mobility models in this paper require the use of martingale inequalities to establish the travel patterns of the mobiles.
In this paper, we will also study one-dimensional mobility models. These models are motivated by certain types of delay-tolerant networks \[[22](#bib.bib22)\], in which a satellite sub-network is used to connect local wireless networks outside of the Internet. Since the satellites move in fixed orbits, they can be modelled as one-dimensional mobilities on a two-dimensional plane. Motivated by such a delay-tolerant network, we consider one-dimensional mobility model where $`n`$ nodes move horizontally and the other $`n`$ node move vertically. Since the node mobility is restricted to one dimension, sources have more information about the positions of destinations compared with the two-dimensional mobility models. We will see that the throughput is improved in this case; for example, under the one-dimensional i.i.d. mobility model with fast mobiles, the trade-off will be shown to be $`{\Theta\hspace{0pt}{(\sqrt[3]{D^{2}/n})}},`$ which is better than $`{\Theta\hspace{0pt}{(\sqrt{D/n})}},`$ the trade-off under the two-dimensional i.i.d. mobility model with fast mobiles. We also propose joint coding-scheduling algorithms which achieve the optimal trade-offs.
Three mobility models are included in this paper, and each model will be investigated under both the fast-mobility and slow-mobility assumptions. The detailed analysis of the two-dimensional hybrid random walk model and one-dimensional i.i.d. mobility model will be presented. The results of the one-dimensional hybrid random walk model can be obtained using the techniques used in the other two models, so the analysis is omitted in this paper for brevity. Our main results include the followings:
1. \(1\)
Two-dimensional hybrid random walk model:
1. \(i\)
Under the fast mobility assumption, it is shown that the maximum throughput per S-D pair is $`O\hspace{0pt}{(\sqrt{D/n})}`$ when $`S = {o\hspace{0pt}{(1)}}`$ and $`{D = {\omega\hspace{0pt}{({{|{\log S}|}/S^{2}})}}},`$ and Joint Coding-Scheduling Algorithm I \[[23](#bib.bib23)\] can achieve the maximum throughput when $`S = {o\hspace{0pt}{(1)}}`$ and $`D`$ is both $`\omega\hspace{0pt}{({\max{\{{{{({\log^{2}n})}\hspace{0pt}{|{\log S}|}}/S^{6}},{\sqrt[3]{n}\hspace{0pt}{\log n}}\}}})}`$ and $`{o\hspace{0pt}{({n/{\log^{2}n}})}}.`$
2. \(ii\)
Under the slow mobility assumption, it is shown that the maximum throughput per S-D pair is $`O\hspace{0pt}{(\sqrt[3]{D/n})}`$ when $`S = {o\hspace{0pt}{(1)}}`$ and $`{D = {\omega\hspace{0pt}{({{|{\log S}|}/S^{2}})}}},`$ and Joint Coding-Scheduling Algorithm II can achieve the maximum throughput when $`S = {o\hspace{0pt}{(1)}}`$ and $`D`$ is both $`\omega\hspace{0pt}{({{{({\log^{2}n})}\hspace{0pt}{|{\log S}|}}/S^{6}})}`$ and $`{o\hspace{0pt}{({n/{\log^{2}n}})}}.`$
2. \(2\)
One-dimensional i.i.d. mobility model:
1. \(i\)
Under the fast mobility assumption, it is shown that the maximum throughput per S-D is $`O\hspace{0pt}\left( \sqrt[3]{D^{2}/n} \right)`$ given delay constraint $`D.`$ Then Joint Coding-Scheduling Algorithm III is proposed to achieve the maximum throughput when $`D`$ is both $`\omega\hspace{0pt}{(\sqrt[5]{n})}`$ and $`{o\hspace{0pt}\left( {\sqrt{n}/\sqrt[{3/2}]{\log n}} \right)}.`$
2. \(ii\)
Under the slow mobility assumption, it is shown that the maximum throughput per S-D pair is $`{O\hspace{0pt}\left( \sqrt[4]{D^{2}/n} \right)}.`$ Joint Coding-Scheduling Algorithm IV is proposed to achieve the maximum throughput when $`D`$ is $`{o\hspace{0pt}\left( {\sqrt{n}/{\log^{2}n}} \right)}.`$
3. \(3\)
One-dimensional hybrid random walk model:
1. \(i\)
Under the fast mobility assumption, it is shown that the maximum throughput per S-D pair is $`O\hspace{0pt}{(\sqrt[3]{D^{2}/n})}`$ when $`S = {o\hspace{0pt}{(1)}}`$ and $`{D = {\omega\hspace{0pt}{({1/S^{2}})}}},`$ and Joint Coding-Scheduling Algorithm III can achieve the maximum throughput when $`S = {o\hspace{0pt}{(1)}}`$ and $`D`$ is both $`\omega\hspace{0pt}{({\max{\{{{{({\log^{2}n})}\hspace{0pt}{|{\log S}|}}/S^{4}},{\sqrt[5]{n}\hspace{0pt}{\log n}}\}}})}`$ and $`{o\hspace{0pt}{({\sqrt{n}/\sqrt[{3/2}]{\log n}})}}.`$
2. \(ii\)
Under the slow mobility assumption, it is shown that the maximum throughput per S-D pair is $`O\hspace{0pt}{(\sqrt[4]{D^{2}/n})}`$ when $`S = {o\hspace{0pt}{(1)}}`$ and $`{D = {\omega\hspace{0pt}{({1/S^{2}})}}},`$ and Joint Coding-Scheduling Algorithm IV can achieve the maximum throughput when $`S = {o\hspace{0pt}{(1)}}`$ and $`D`$ is both $`\omega\hspace{0pt}{({{{({\log^{2}n})}\hspace{0pt}{|{\log S}|}}/S^{4}})}`$ and $`{o\hspace{0pt}{({\sqrt{n}/{\log^{2}n}})}}.`$
Note that the optimal delay-throughput trade-off are established under some conditions on $`D.`$ When these conditions are not met, the trade-off is still unknown in general, though a trade-off of the two-dimensional hybrid random walk model with slow mobiles has been established under an assumption regarding packet replication in \[[19](#bib.bib19)\]. We also would like to mention that when the step size of the two-dimensional hybrid random walk is $`{1/\sqrt{n}},`$ our two-dimensional hybrid random walk model is identical to the random walk model studied in \[[6](#bib.bib6), [7](#bib.bib7)\], where the optimal delay-throughput trade-off has been obtained. Our results do not apply to this case since the set of allowed values for $`D`$ becomes empty in that case (see (1) (i) above).
The remainder of the paper is organized as follows: In Section [III](#S3 "III Model ‣ Optimal Delay-Throughput Trade-offs in Mobile Ad-Hoc Networks: Hybrid Random Walk and One-Dimensional Mobility ModelsAn earlier version of this paper appeared in the Proc. of ITA Workshop, 2007.Research presented here was supported in part by a Vodafone Fellowship and NSF grant CNS 05-19691."), we introduce the communication and mobility model. Then we analyze the two-dimensional hybrid random walk models in Section [IV](#S4 "IV Two-Dimensional Hybrid Random Walk Models ‣ Optimal Delay-Throughput Trade-offs in Mobile Ad-Hoc Networks: Hybrid Random Walk and One-Dimensional Mobility ModelsAn earlier version of this paper appeared in the Proc. of ITA Workshop, 2007.Research presented here was supported in part by a Vodafone Fellowship and NSF grant CNS 05-19691."), and one-dimensional i.i.d. mobility models in Section [V](#S5 "V One-Dimensional I.I.D. Mobility Models ‣ Optimal Delay-Throughput Trade-offs in Mobile Ad-Hoc Networks: Hybrid Random Walk and One-Dimensional Mobility ModelsAn earlier version of this paper appeared in the Proc. of ITA Workshop, 2007.Research presented here was supported in part by a Vodafone Fellowship and NSF grant CNS 05-19691."). The results of one-dimensional hybrid random walk model are presented in Section [VI](#S6 "VI One-Dimensional Hybrid Random Walk Model, Fast Mobiles and Slow Mobiles ‣ Optimal Delay-Throughput Trade-offs in Mobile Ad-Hoc Networks: Hybrid Random Walk and One-Dimensional Mobility ModelsAn earlier version of this paper appeared in the Proc. of ITA Workshop, 2007.Research presented here was supported in part by a Vodafone Fellowship and NSF grant CNS 05-19691."). Finally, the conclusions is given in Section [VII](#S7 "VII Conclusion ‣ Optimal Delay-Throughput Trade-offs in Mobile Ad-Hoc Networks: Hybrid Random Walk and One-Dimensional Mobility ModelsAn earlier version of this paper appeared in the Proc. of ITA Workshop, 2007.Research presented here was supported in part by a Vodafone Fellowship and NSF grant CNS 05-19691.").
## III Model
In this section, we first present the models that we use for mobility and wireless interference. Then the definitions of delay and throughput are provided.
Mobile Ad-Hoc Network Model: Consider an ad-hoc network where wireless mobile nodes are positioned in a unit square. Assume that the time is slotted, we study following three mobility models in this paper.
1. \(1\)
Two-Dimensional Random Walk Model: Consider a unit square which is further divided into $`1/S^{2}`$ squares of equal size. Each of the smaller square will be called an RW-cell (random walk cell), and indexed by $`(U^{x},U^{y})`$ where $`{{U^{x},U^{y}} \in {\{ 1,\ldots,{1/S}\}}}.`$ The unit square is assumed to be a torus, i.e., the top and bottom edges are assumed to touch each other and similarly the left and right edges also are assumed to touch other. A node which is in one RW-cell at a time slot moves to one of its eight adjacent RW-cells or stays in the same RW-cell in the next time-slot with each move being equally likely as in Figure [1](#S3.F1 "Figure 1 ‣ item (1) ‣ III Model ‣ Optimal Delay-Throughput Trade-offs in Mobile Ad-Hoc Networks: Hybrid Random Walk and One-Dimensional Mobility ModelsAn earlier version of this paper appeared in the Proc. of ITA Workshop, 2007.Research presented here was supported in part by a Vodafone Fellowship and NSF grant CNS 05-19691."). Two RW-cells are said to be adjacent if they share a common point. The node position within the RW-cell is randomly uniformly selected. There are $`n`$ S-D pairs in the network. Each node is both a source and a destination. Without loss of generality, we assume that the destination of node $`i`$ is node $`{i + 1},`$ and the destination of node $`n`$ is node $`1.`$
[Refer to caption](/html/0705.0326/assets/x1.png)
Figure 1: Two-Dimensional Random Walk Model
2. \(2\)
One-Dimensional I.I.D. Mobility Model: Our one-dimensional i.i.d. mobility model is defined as follows:
1. \(i\)
There are $`2\hspace{0pt}n`$ nodes in the network. Among them, $`n`$ nodes, named H-nodes, move horizontally; and the other $`n`$ nodes, named V-nodes, move vertically.
2. \(ii\)
Using $`(x_{i},y_{i})`$ to denote the position of node $`i.`$ If node $`i`$ is an H-node, $`y_{i}`$ is fixed and $`x_{i}`$ is a value randomly uniformly chosen from $`{\lbrack 0,1\rbrack}.`$ We also assume that H-nodes are evenly distributed vertically, so $`y_{i}`$ takes values $`{{1/n},{2/n},\ldots,1}.`$ V-nodes have similar properties.
3. \(iii\)
Assume that source and destination are the same type of nodes. Also assume that node $`i`$ is an H-node if $`i`$ is odd, and a V-node if $`i`$ is even. Further, assume that the destination of node $`i`$ is node $`{i + 2},`$ the destination of node $`{2\hspace{0pt}n} - 1`$ is node $`1,`$ and the destination of node $`2\hspace{0pt}n`$ is node $`2.`$
4. \(iv\)
The orbit distance of two H(V)-nodes is defined to be the vertical (horizontal) distance of the two nodes.
3. \(3\)
One-Dimensional Random Walk Model: Each orbit is divided into $`1/S`$ RW-intervals (random walk interval). At each time slot, a node moves into one of two adjacent RW-intervals or stays at the current RW-interval (see Figure [2](#S3.F2 "Figure 2 ‣ item (3) ‣ III Model ‣ Optimal Delay-Throughput Trade-offs in Mobile Ad-Hoc Networks: Hybrid Random Walk and One-Dimensional Mobility ModelsAn earlier version of this paper appeared in the Proc. of ITA Workshop, 2007.Research presented here was supported in part by a Vodafone Fellowship and NSF grant CNS 05-19691.")). The node position in the RW-interval is randomly, uniformly selected.
[Refer to caption](/html/0705.0326/assets/x2.png)
Figure 2: One-Dimensional Random Walk Model
Communication Model: We assume the protocol model introduced in \[[10](#bib.bib10)\] in this paper. Let $`\text{dist}\hspace{0pt}{(i,j)}`$ denote the Euclidean distance between node $`i`$ and node $`j,`$ and $`r_{i}`$ to denote the transmission radius of node $`i.`$ A transmission from node $`i`$ can be successfully received at node $`j`$ if and only if following two conditions hold:
1. \(i\)
$`{{\text{dist}\hspace{0pt}{(i,j)}} \leq r_{i}};`$
2. \(ii\)
$`{\text{dist}\hspace{0pt}{(k,j)}} \geq {{({1 + \Delta})}\hspace{0pt}\text{dist}\hspace{0pt}{(i,j)}}`$ for each node $`k \neq i`$ which transmits at the same time, where $`\Delta`$ is a protocol-specified guard-zone to prevent interference.
We further assume that at each time slot, at most $`W`$ bits can be transmitted in a successful transmission.
Time-Scale of Mobility: Two time-scales of mobility are considered in this paper.
1. \(1\)
Fast mobility: The mobility of nodes is at the same time-scale as the data transmission, so $`W`$ is a constant independent of $`n`$ and only one-hop transmissions are feasible in single time slot.
2. \(2\)
Slow mobility: The mobility of nodes is much slower than the wireless transmission, so $`{W \gg n}.`$ Under this assumption, the packet size can be scaled as $`{W/H}\hspace{0pt}{(n)}`$ for $`{H\hspace{0pt}{(n)}} = {O\hspace{0pt}{(n)}}`$ to guarantee $`H\hspace{0pt}{(n)}`$-hop transmissions are feasible in single time slot.
Delay and Throughput: We consider hard delay constraints in this paper. Given a delay constraint $`D,`$ a packet is said to be successfully delivered if the destination obtains the packet within $`D`$ time slots after it is sent out from the source.
Let $`\Lambda_{i}\hspace{0pt}{\lbrack T\rbrack}`$ denote the number of bits successfully delivered to the destination of node $`i`$ in time interval $`{\lbrack 0,T\rbrack}.`$ A throughput of $`\lambda`$ per S-D pair is said to be feasible under the delay constraint $`D`$ and loss probability constraint $`\epsilon > 0`$ if there exists $`n_{0}`$ such that for any $`{n \geq n_{0}},`$ there exists a coding/routing/scheduling algorithm with the property that each bit transmitted by a source is received at its destination with probability at least $`{1 - \epsilon},`$ and
<table>
<tbody>
<tr>
<td></td>
<td>limT→∞Pr(Λi[T]T≥λ, ∀ i)=1.subscript→𝑇PrsubscriptΛ𝑖delimited-[]𝑇𝑇𝜆 for-all 𝑖1\lim_{T\rightarrow\infty}\Pr\left(\displaystyle\frac{\Lambda_{i}[T]}{T}\geq\lambda,\hbox{ }\forall\hbox{ }i\right)=1.</td>
<td></td>
<td>(1)</td>
</tr>
</tbody>
</table>
## IV Two-Dimensional Hybrid Random Walk Models
The optimal delay-throughput trade-offs of the two-dimensional i.i.d. mobility model with fast mobiles and slow mobiles have been established in \[[23](#bib.bib23)\]. In this section, we first first extend the results to two-dimensional hybrid random walk models. We will obtain the maximum throughput for $`{D = {\omega\hspace{0pt}\left( {{|{\log S}|}/S^{2}} \right)}},`$ and then show that the maximum throughput can be achieved using the algorithms proposed in \[[23](#bib.bib23)\] under some additional constraints on $`D.`$
###
IV-A
Upper Bound
The upper bound is established under the following assumptions:
Assumption 1: Packets destined for different nodes cannot be encoded together.
Assumption 2: A new coded packet is generated right before the packet is sent out. The node generating the coded packet does not store the packet in its buffer.
Assumption 3: Once a node receives a packet (coded or uncoded), the packet is not discarded by the node till its deadline expires.
Note that Assumption $`1`$ is the only significant restriction imposed on coding/routing/scheduling schemes. Next we introduce following notations which will be used in our proof.
- •
$`{\Lambda\hspace{0pt}{\lbrack T\rbrack}}:`$ $`{{\Lambda\hspace{0pt}{\lbrack T\rbrack}} = {\sum_{i = 1}^{n}{\Lambda_{i}\hspace{0pt}{\lbrack T\rbrack}}}}.`$
- •
$`b:`$ Index of a bit stored in the network. Bit $`b`$ could be either a bit of a data packet or a bit of a coded packet.
- •
$`d_{b}:`$ The destination of bit $`b.`$
- •
$`c_{b}:`$ The node storing bit $`b.`$
- •
$`t_{b}:`$ The time slot at which bit $`b`$ is generated.
- •
$`{\overset{\sim}{L}}_{b}:`$ The minimum distance between node $`d_{b}`$ and node $`c_{b}`$ from time slot $`t_{b}`$ to time slot $`{{t_{b} + D} - 1},`$ i.e.,
<table>
<tbody>
<tr>
<td></td>
<td>L~b=mintb≤t≤tb+D−1dist(db,cb)(t).subscript~𝐿𝑏subscriptsubscript𝑡𝑏𝑡subscript𝑡𝑏𝐷1distsubscript𝑑𝑏subscript𝑐𝑏𝑡\tilde{L}_{b}=\min_{t_{b}\leq t\leq t_{b}+D-1}\hbox{dist}(d_{b},c_{b})(t).</td>
<td></td>
</tr>
</tbody>
</table>
######
Theorem 1
Consider the two-dimensional hybrid random walk model with step-size $`S = {o\hspace{0pt}{(1)}}`$ and delay constraint $`{D = {\omega\hspace{0pt}{({{|{\log S}|}/S^{2}})}}},`$ and suppose that Assumption 1-3 hold. We have following results:
1. \(1\)
For fast mobiles,
<table>
<tbody>
<tr>
<td></td>
<td>482WTΔπn(D+1)≥E[Λ[T]].482𝑊𝑇Δ𝜋𝑛𝐷1𝐸delimited-[]Λdelimited-[]𝑇\displaystyle\frac{48\sqrt{2}WT}{\Delta\sqrt{\pi}}\sqrt{n}(\sqrt{D}+1)\geq E[\Lambda[T]].</td>
<td></td>
<td>(2)</td>
</tr>
</tbody>
</table>
2. \(2\)
For slow mobiles,
<table>
<tbody>
<tr>
<td></td>
<td>893WTΔπ23/2n3/2(D3+1)≥E[Λ[T]].839𝑊𝑇32Δsuperscript𝜋232𝑛3𝐷1𝐸delimited-[]Λdelimited-[]𝑇\displaystyle\frac{8\sqrt[3]{9}WT}{\sqrt[3/2]{\Delta\pi^{2}}}\sqrt[3/2]{n}(\sqrt[3]{D}+1)\geq E[\Lambda[T]].</td>
<td></td>
<td>(3)</td>
</tr>
</tbody>
</table>
###### Proof:
Let $`N_{b}^{rw}`$ denote the number of time slots, from $`t_{b} + 1`$ to $`{t_{b} + D},`$ at which node $`c_{b}`$ and $`d_{b}`$ are in the same RW-cell or neighboring RW-cells. Then for any $`{L \in {\lbrack 0,{S/\sqrt{\pi}})}},`$ we have
<table>
<tbody>
<tr>
<td></td>
<td></td>
<td></td>
<td>Pr(L~b≤L)Prsubscript~𝐿𝑏𝐿\displaystyle\Pr\left(\tilde{L}_{b}\leq L\right)</td>
<td></td>
</tr>
<tr>
<td></td>
<td></td>
<td>=\displaystyle=</td>
<td>∑K=1DPr(L~b≤L|Nbrw=K)Pr(Nbrw=K)superscriptsubscript𝐾1𝐷Prsubscript~𝐿𝑏conditional𝐿subscriptsuperscript𝑁rw𝑏𝐾Prsubscriptsuperscript𝑁rw𝑏𝐾\displaystyle\sum_{K=1}^{D}\Pr\left(\tilde{L}_{b}\leq L|{N}^{\rm rw}_{b}=K\right)\Pr\left(N^{\rm rw}_{b}=K\right)</td>
<td></td>
</tr>
<tr>
<td></td>
<td></td>
<td>≤\displaystyle\leq</td>
<td>∑K=1D(1−(1−πL2S2)K)Pr(Nbrw=K)superscriptsubscript𝐾1𝐷1superscript1𝜋superscript𝐿2superscript𝑆2𝐾Prsubscriptsuperscript𝑁rw𝑏𝐾\displaystyle\sum_{K=1}^{D}\left(1-\left(1-\frac{\pi L^{2}}{S^{2}}\right)^{K}\right)\Pr\left(N^{\rm rw}_{b}=K\right)</td>
<td></td>
</tr>
<tr>
<td></td>
<td></td>
<td>=\displaystyle=</td>
<td>1−E[(1−πL2S2)Nbrw]1𝐸delimited-[]superscript1𝜋superscript𝐿2superscript𝑆2subscriptsuperscript𝑁rw𝑏\displaystyle 1-E\left[\left(1-\frac{\pi L^{2}}{S^{2}}\right)^{N^{\rm rw}_{b}}\right]</td>
<td></td>
</tr>
<tr>
<td></td>
<td></td>
<td>≤\displaystyle\leq</td>
<td>1−(1−πL2S2)E[Nbrw],1superscript1𝜋superscript𝐿2superscript𝑆2𝐸delimited-[]subscriptsuperscript𝑁rw𝑏\displaystyle 1-\left(1-\frac{\pi L^{2}}{S^{2}}\right)^{E\left[N^{\rm rw}_{b}\right]},</td>
<td></td>
</tr>
</tbody>
</table>
where the first inequality follows from the fact that the node position within a RW-cell is randomly uniformly selected, and the last inequality follows from the Jensen’s inequality.
Next we consider $`{E\hspace{0pt}{\lbrack N_{b}^{tw}\rbrack}}.`$ Let $`({U_{i}^{x}\hspace{0pt}{(t)}},{U_{i}^{y}\hspace{0pt}{(t)}})`$ denote the RW-cell in which node $`i`$ is at time slot $`t,`$ and $`({V_{i}^{x}\hspace{0pt}{(t)}},{V_{i}^{y}\hspace{0pt}{(t)}})`$ denote the displacement of node $`i`$ at time slot $`t,`$ i.e.,
<table>
<tbody>
<tr>
<td></td>
<td>Vix(t)={1,w.p. 130,w.p. 13−1,w.p. 13 and Viy(t)={1,w.p. 130,w.p. 13−1,w.p. 13.subscriptsuperscript𝑉𝑥𝑖𝑡cases1w.p. 130w.p. 131w.p. 13 and subscriptsuperscript𝑉𝑦𝑖𝑡cases1w.p. 130w.p. 131w.p. 13\displaystyle V^{x}_{i}(t)=\left\{\begin{array}[]{ll}1,&\hbox{w.p. }\frac{1}{3}\\ 0,&\hbox{w.p. }\frac{1}{3}\\ -1,&\hbox{w.p. }\frac{1}{3}\\ \end{array}\right.\hbox{ and }V^{y}_{i}(t)=\left\{\begin{array}[]{ll}1,&\hbox{w.p. }\frac{1}{3}\\ 0,&\hbox{w.p. }\frac{1}{3}\\ -1,&\hbox{w.p. }\frac{1}{3}\\ \end{array}\right..</td>
<td></td>
</tr>
</tbody>
</table>
It is easy to see that
<table>
<tbody>
<tr>
<td></td>
<td>Uix(t)superscriptsubscript𝑈𝑖𝑥𝑡\displaystyle U_{i}^{x}(t)</td>
<td>=\displaystyle=</td>
<td>[(Uix(0)+∑m=1t−1Vix(m))mod1S]+1;delimited-[]modulosuperscriptsubscript𝑈𝑖𝑥0superscriptsubscript𝑚1𝑡1subscriptsuperscript𝑉𝑥𝑖𝑚1𝑆1\displaystyle\left[\left(U_{i}^{x}(0)+\sum_{m=1}^{t-1}V^{x}_{i}(m)\right)\mod{\frac{1}{S}}\right]+1;</td>
<td></td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td>Uiy(t)superscriptsubscript𝑈𝑖𝑦𝑡\displaystyle U_{i}^{y}(t)</td>
<td>=\displaystyle=</td>
<td>[(Uiy(0)+∑m=1t−1Viy(m))mod1S]+1.delimited-[]modulosuperscriptsubscript𝑈𝑖𝑦0superscriptsubscript𝑚1𝑡1subscriptsuperscript𝑉𝑦𝑖𝑚1𝑆1\displaystyle\left[\left(U_{i}^{y}(0)+\sum_{m=1}^{t-1}V^{y}_{i}(m)\right)\mod{\frac{1}{S}}\right]+1.</td>
<td></td>
</tr>
</tbody>
</table>
Further, let $`({U_{i - j}^{x}\hspace{0pt}{(t)}},{U_{i - j}^{y}\hspace{0pt}{(t)}})`$ denote the relative position of node $`i`$ from node $`j,`$ i.e.,
<table>
<tbody>
<tr>
<td></td>
<td>Ui−jx(t)superscriptsubscript𝑈𝑖𝑗𝑥𝑡\displaystyle U_{i-j}^{x}(t)</td>
<td>=\displaystyle=</td>
<td>(Ui−jx(0)+∑m=1t−1V~i−jx(m))mod1S;modulosuperscriptsubscript𝑈𝑖𝑗𝑥0superscriptsubscript𝑚1𝑡1subscriptsuperscript~𝑉𝑥𝑖𝑗𝑚1𝑆\displaystyle\left(U_{i-j}^{x}(0)+\sum_{m=1}^{t-1}\tilde{V}^{x}_{i-j}(m)\right)\mod{\frac{1}{S}};</td>
<td></td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td>Ui−jy(t)superscriptsubscript𝑈𝑖𝑗𝑦𝑡\displaystyle U_{i-j}^{y}(t)</td>
<td>=\displaystyle=</td>
<td>(Ui−jy(0)+∑m=1t−1V~i−jy(m))mod1S,modulosuperscriptsubscript𝑈𝑖𝑗𝑦0superscriptsubscript𝑚1𝑡1subscriptsuperscript~𝑉𝑦𝑖𝑗𝑚1𝑆\displaystyle\left(U_{i-j}^{y}(0)+\sum_{m=1}^{t-1}\tilde{V}^{y}_{i-j}(m)\right)\mod{\frac{1}{S}},</td>
<td></td>
</tr>
</tbody>
</table>
where
<table>
<tbody>
<tr>
<td></td>
<td>V~i−jx(t)=Vix(t)−Vjx(t)={2,w.p. 191,w.p. 290,w.p. 13−1,w.p. 29−2,w.p. 19,subscriptsuperscript~𝑉𝑥𝑖𝑗𝑡subscriptsuperscript𝑉𝑥𝑖𝑡subscriptsuperscript𝑉𝑥𝑗𝑡cases2w.p. 191w.p. 290w.p. 131w.p. 292w.p. 19\displaystyle\tilde{V}^{x}_{i-j}(t)={V}^{x}_{i}(t)-{V}^{x}_{j}(t)=\left\{\begin{array}[]{ll}2,&\hbox{w.p. }\frac{1}{9}\\ 1,&\hbox{w.p. }\frac{2}{9}\\ 0,&\hbox{w.p. }\frac{1}{3}\\ -1,&\hbox{w.p. }\frac{2}{9}\\ -2,&\hbox{w.p. }\frac{1}{9}\\ \end{array}\right.,</td>
<td></td>
</tr>
</tbody>
</table>
and
<table>
<tbody>
<tr>
<td></td>
<td>V~i−jy(t)=Viy(t)−Vjy(t)={2,w.p. 191,w.p. 290,w.p. 13−1,w.p. 29−2,w.p. 19.subscriptsuperscript~𝑉𝑦𝑖𝑗𝑡subscriptsuperscript𝑉𝑦𝑖𝑡subscriptsuperscript𝑉𝑦𝑗𝑡cases2w.p. 191w.p. 290w.p. 131w.p. 292w.p. 19\displaystyle\tilde{V}^{y}_{i-j}(t)={V}^{y}_{i}(t)-{V}^{y}_{j}(t)=\left\{\begin{array}[]{ll}2,&\hbox{w.p. }\frac{1}{9}\\ 1,&\hbox{w.p. }\frac{2}{9}\\ 0,&\hbox{w.p. }\frac{1}{3}\\ -1,&\hbox{w.p. }\frac{2}{9}\\ -2,&\hbox{w.p. }\frac{1}{9}\\ \end{array}\right..</td>
<td></td>
</tr>
</tbody>
</table>
So $`({U_{i - j}^{x}\hspace{0pt}{(t)}},{U_{i - j}^{y}\hspace{0pt}{(t)}})`$ is the consequence of random walk $`({{\overset{\sim}{V}}_{i - j}^{x}\hspace{0pt}{(m)}},{{\overset{\sim}{V}}_{i - j}^{y}\hspace{0pt}{(m)}})`$ with initial position
<table>
<tbody>
<tr>
<td></td>
<td>(Ui−jx(0),Ui−jy(0))=(Uix(0)−Ujx(0),Uiy(0)−Ujy(0)).superscriptsubscript𝑈𝑖𝑗𝑥0superscriptsubscript𝑈𝑖𝑗𝑦0superscriptsubscript𝑈𝑖𝑥0superscriptsubscript𝑈𝑗𝑥0superscriptsubscript𝑈𝑖𝑦0superscriptsubscript𝑈𝑗𝑦0(U_{i-j}^{x}(0),U_{i-j}^{y}(0))=(U_{i}^{x}(0)-U_{j}^{x}(0),U_{i}^{y}(0)-U_{j}^{y}(0)).</td>
<td></td>
</tr>
</tbody>
</table>
Note that node $`c_{b}`$ and node $`d_{b}`$ are in the same RW-cell if $`{{({U_{c_{b} - d_{b}}^{x}\hspace{0pt}{(t)}},{U_{c_{b} - d_{b}}^{y}\hspace{0pt}{(t)}})} = {(0,0)}},`$ and in neighboring RW-cells if $`{{({U_{c_{b} - d_{b}}^{x}\hspace{0pt}{(t)}},{U_{c_{b} - d_{b}}^{y}\hspace{0pt}{(t)}})} \in {\{{(0,1)},{(1,0)},{(1,1)},{(0,{{1/S} - 1})},{({{1/S} - 1},0)},{({{1/S} - 1},{{1/S} - 1})}\}}}.`$ Similar to the argument in Lemma [12](#Sx2.EGx56 "Lemma 12 ‣ Appendix B: Properties of Random Walk ‣ Optimal Delay-Throughput Trade-offs in Mobile Ad-Hoc Networks: Hybrid Random Walk and One-Dimensional Mobility ModelsAn earlier version of this paper appeared in the Proc. of ITA Workshop, 2007.Research presented here was supported in part by a Vodafone Fellowship and NSF grant CNS 05-19691.") provided in Appendix B, we can conclude that for $`{D = {\omega\hspace{0pt}{({{|{\log S}|}/S^{2}})}}},`$
<table>
<tbody>
<tr>
<td></td>
<td>E[Nbrw]≤9910S2D,𝐸delimited-[]subscriptsuperscript𝑁rw𝑏9910superscript𝑆2𝐷\displaystyle E\left[N^{\rm rw}_{b}\right]\leq\frac{99}{10}S^{2}D,</td>
<td></td>
</tr>
</tbody>
</table>
which implies that
<table>
<tbody>
<tr>
<td></td>
<td>Pr(L~b≤L)Prsubscript~𝐿𝑏𝐿\displaystyle\Pr\left(\tilde{L}_{b}\leq L\right)</td>
<td>≤\displaystyle\leq</td>
<td>1−(1−πL2S2)9910S2D1superscript1𝜋superscript𝐿2superscript𝑆29910superscript𝑆2𝐷\displaystyle 1-\left(1-\frac{\pi L^{2}}{S^{2}}\right)^{\frac{99}{10}S^{2}D}</td>
<td></td>
<td rowspan="2">(7)</td>
</tr>
<tr>
<td></td>
<td></td>
<td>≤\displaystyle\leq</td>
<td>36L2D.36superscript𝐿2𝐷\displaystyle 36L^{2}D.</td>
<td></td>
</tr>
</tbody>
</table>
Based on inequality ([7](#S4.E7 "In Proof: ‣ IV-A Upper Bound ‣ IV Two-Dimensional Hybrid Random Walk Models ‣ Optimal Delay-Throughput Trade-offs in Mobile Ad-Hoc Networks: Hybrid Random Walk and One-Dimensional Mobility ModelsAn earlier version of this paper appeared in the Proc. of ITA Workshop, 2007.Research presented here was supported in part by a Vodafone Fellowship and NSF grant CNS 05-19691.")), the proof of inequality ([2](#S4.E2 "In item (1) ‣ Theorem 1 ‣ IV-A Upper Bound ‣ IV Two-Dimensional Hybrid Random Walk Models ‣ Optimal Delay-Throughput Trade-offs in Mobile Ad-Hoc Networks: Hybrid Random Walk and One-Dimensional Mobility ModelsAn earlier version of this paper appeared in the Proc. of ITA Workshop, 2007.Research presented here was supported in part by a Vodafone Fellowship and NSF grant CNS 05-19691.")) is similar to the proof of Theorem 3 of \[[23](#bib.bib23)\], and the proof of ([3](#S4.E3 "In item (2) ‣ Theorem 1 ‣ IV-A Upper Bound ‣ IV Two-Dimensional Hybrid Random Walk Models ‣ Optimal Delay-Throughput Trade-offs in Mobile Ad-Hoc Networks: Hybrid Random Walk and One-Dimensional Mobility ModelsAn earlier version of this paper appeared in the Proc. of ITA Workshop, 2007.Research presented here was supported in part by a Vodafone Fellowship and NSF grant CNS 05-19691.")) is similar to the proof of Theorem 6 of \[[23](#bib.bib23)\]. ∎
###
IV-B
Joint Coding-Scheduling Algorithms
From Theorem [1](#Thmthm1 "Theorem 1 ‣ IV-A Upper Bound ‣ IV Two-Dimensional Hybrid Random Walk Models ‣ Optimal Delay-Throughput Trade-offs in Mobile Ad-Hoc Networks: Hybrid Random Walk and One-Dimensional Mobility ModelsAn earlier version of this paper appeared in the Proc. of ITA Workshop, 2007.Research presented here was supported in part by a Vodafone Fellowship and NSF grant CNS 05-19691."), we can see that the optimal delay-throughput trade-offs of the two-dimensional hybrid random walk models are similar to the ones of the two-dimensional i.i.d. mobility models \[[23](#bib.bib23)\]. It motivates us to consider the algorithms proposed in \[[23](#bib.bib23)\]. As in \[[23](#bib.bib23)\], we define and categorize packets into four different types.
- •
Data packets: There are the uncoded data packets that have to be transmitted by the sources and received by the destinations.
- •
Coded packets: Packets generated by Raptor codes. We let $`(i,k)`$ denote the $`k^{th}`$ coded packet of node $`i.`$
- •
Duplicate packets: Each coded packet could be broadcast to other nodes to generate multiple copies, called duplicate packets. We let $`(i,k,j)`$ denote a copy of $`(i,k)`$ carried by node $`j,`$ and $`(i,k,J)`$ to denote the set of all copies of coded packet $`{(i,k)}.`$
- •
Deliverable packets: Duplicate packets that happen to be within distance $`L`$ from their destinations.
We will show that the optimal trade-offs can be achieved using Joint Coding-Scheduling Algorithm I and II presented in \[[23](#bib.bib23)\] with the following modifications:
1. \(1\)
For the fast mobility model, we use Joint Coding-Scheduling Algorithm I with the following modification: $`{2\hspace{0pt}D}/{({25\hspace{0pt}M})}`$ data packets are coded into $`D/M`$ coded packets;
2. \(2\)
For the slow mobility model, we use Joint Coding-Scheduling Algorithm II with the following modifications: $`D/7`$ data packets are coded to $`D`$ coded packets.
For the detail of the algorithms, please refer to \[[23](#bib.bib23)\].
######
Theorem 2
Consider the two-dimensional hybrid random walk models.
1. \(1\)
Fast mobility model: Suppose that $`S`$ is $`{o\hspace{0pt}{(1)}},`$ $`D`$ is both $`\omega\hspace{0pt}{({\max{\{{{\log^{2}{n\hspace{0pt}{|{\log S}|}}}/{(S^{6})}},{\sqrt[3]{n}\hspace{0pt}{\log n}}\}}})}`$ and $`{o\hspace{0pt}{({n/{({\log^{2}n})}})}},`$ and the delay constraint is $`{6\hspace{0pt}D}.`$ Then under the fast mobility model, given any $`\epsilon`$ there exists $`n_{0}`$ such that for any $`{n \geq n_{0}},`$ every data packet sent out can be recovered at the destination with probability at least $`{1 - \epsilon},`$ and
<table>
<tbody>
<tr>
<td></td>
<td>limT→∞Pr(Λi[T]T≥(W2520)(Dn) ∀i)=1.subscript→𝑇PrsubscriptΛ𝑖delimited-[]𝑇𝑇𝑊2520𝐷𝑛 for-all𝑖1\displaystyle\lim_{T\rightarrow\infty}\Pr\left(\frac{\Lambda_{i}[T]}{T}\geq\left(\frac{W}{2520}\right)\left(\sqrt{\frac{D}{n}}\right)\hbox{ }\forall i\right)=1.</td>
<td></td>
<td>(8)</td>
</tr>
</tbody>
</table>
by using the modified Joint Coding-Scheduling Algorithm I.
2. \(2\)
Slow mobility model: Suppose that $`S`$ is $`o\hspace{0pt}{(1)}`$ and $`D`$ is both $`\omega\hspace{0pt}{({{\log^{2}{n\hspace{0pt}{|{\log S}|}}}/{(S^{6})}})}`$ and $`{o\hspace{0pt}{({n/{({\log^{3}n})}})}},`$ and the delay constraints is $`{16\hspace{0pt}D}.`$ Then under the slow mobility model, given any $`\epsilon`$ there exists $`{\overset{\sim}{n}}_{0}`$ such that for any $`{n \geq {\overset{\sim}{n}}_{0}},`$ every data packet sent out can be recovered at the destination with probability at least $`{1 - \epsilon},`$ and
<table>
<tbody>
<tr>
<td></td>
<td>limT→∞Pr(Λi[T]T≥(W2242csC)(Dn3) ∀i)=1.subscript→𝑇PrsubscriptΛ𝑖delimited-[]𝑇𝑇𝑊2242subscript𝑐𝑠𝐶3𝐷𝑛 for-all𝑖1\displaystyle\lim_{T\rightarrow\infty}\Pr\left(\frac{\Lambda_{i}[T]}{T}\geq\left(\frac{W}{224\sqrt{2}c_{s}C}\right)\left(\sqrt[3]{\frac{D}{n}}\right)\hbox{ }\forall i\right)=1.</td>
<td></td>
<td>(9)</td>
</tr>
</tbody>
</table>
by using the modified Joint Coding-Scheduling Algorithm II.
###### Proof:
Let $`\overset{\sim}{A}`$ denote the area of a cell, and $`\overset{\sim}{M}\hspace{0pt}{\lbrack t\rbrack}`$ to denote the number of nodes in the cell at time slot $`t.`$ A cell is said to be a *good cell* at time $`t`$ if
<table>
<tbody>
<tr>
<td></td>
<td>910A~n+1≤M~[t]≤1110A~n.910~𝐴𝑛1~𝑀delimited-[]𝑡1110~𝐴𝑛\frac{9}{10}\tilde{A}n+1\leq\tilde{M}[t]\leq\frac{11}{10}{\tilde{A}n}.</td>
<td></td>
</tr>
</tbody>
</table>
Proof of (1): We consider one super time slot which consists of $`6\hspace{0pt}D`$ time slots, and calculate the probability that the $`{2\hspace{0pt}D}/{({25\hspace{0pt}M})}`$ data packets from node $`i`$ are fully recovered at the destination, where $`M = \sqrt{n/D}`$ is the mean number of nodes in each cell. The proof will show the following events happen with high probability.
Node distribution: All cells are good during the entire super-time-slot with high probability. Letting $`\mathcal{G}`$ denote this event, we will show
<table>
<tbody>
<tr>
<td></td>
<td>Pr(𝒢)≥1−1n2.Pr𝒢11superscript𝑛2\displaystyle\Pr\left({\cal G}\right)\geq 1-\frac{1}{n^{2}}.</td>
<td></td>
<td>(10)</td>
</tr>
</tbody>
</table>
Broadcasting: At least $`{16\hspace{0pt}D}/{({25\hspace{0pt}M})}`$ coded packets from a source are successfully duplicated after the broadcasting step with high probability, where a coded packet is said to be successfully duplicated if the packet is in at least $`{4\hspace{0pt}M}/5`$ distinct relay nodes. Letting $`A_{i}`$ denote the number of coded packets which are successfully duplicated in a super time slot, we will first show that
<table>
<tbody>
<tr>
<td></td>
<td>Pr(Ai≥16D25M|𝒢)≥1−55Dn−e−D600M.Prsubscript𝐴𝑖conditional16𝐷25𝑀𝒢155𝐷𝑛superscript𝑒𝐷600𝑀\displaystyle\Pr\left(\left.A_{i}\geq\frac{16D}{25M}\right|{\cal G}\right)\geq 1-\frac{55D}{n}-e^{-\frac{D}{600M}}.</td>
<td></td>
<td>(11)</td>
</tr>
</tbody>
</table>
Receiving: At least $`{3\hspace{0pt}D}/{({25\hspace{0pt}M})}`$ distinct coded packets from a source are delivered to its destination after the receiving step with high probability. Letting $`B_{i}`$ denote the number of distinct coded packets delivered to destination $`i + 1`$ in a super time slot, we will show
<table>
<tbody>
<tr>
<td></td>
<td>Pr(Bi≥325DM|Ai≥1625DM)≥1−2e−logD8500−e−D500MlogD.Prsubscript𝐵𝑖conditional325𝐷𝑀subscript𝐴𝑖1625𝐷𝑀12superscript𝑒𝐷8500superscript𝑒𝐷500𝑀𝐷\displaystyle\Pr\left(\left.B_{i}\geq\frac{3}{25}\frac{D}{M}\right|A_{i}\geq\frac{16}{25}\frac{D}{M}\right)\geq 1-2e^{-\frac{\log D}{8500}}-e^{-\frac{D}{500M\log D}}.</td>
<td></td>
<td>(12)</td>
</tr>
</tbody>
</table>
From inequalities ([10](#S4.E10 "In Proof: ‣ IV-B Joint Coding-Scheduling Algorithms ‣ IV Two-Dimensional Hybrid Random Walk Models ‣ Optimal Delay-Throughput Trade-offs in Mobile Ad-Hoc Networks: Hybrid Random Walk and One-Dimensional Mobility ModelsAn earlier version of this paper appeared in the Proc. of ITA Workshop, 2007.Research presented here was supported in part by a Vodafone Fellowship and NSF grant CNS 05-19691.")), ([11](#S4.E11 "In Proof: ‣ IV-B Joint Coding-Scheduling Algorithms ‣ IV Two-Dimensional Hybrid Random Walk Models ‣ Optimal Delay-Throughput Trade-offs in Mobile Ad-Hoc Networks: Hybrid Random Walk and One-Dimensional Mobility ModelsAn earlier version of this paper appeared in the Proc. of ITA Workshop, 2007.Research presented here was supported in part by a Vodafone Fellowship and NSF grant CNS 05-19691.")) and ([12](#S4.E12 "In Proof: ‣ IV-B Joint Coding-Scheduling Algorithms ‣ IV Two-Dimensional Hybrid Random Walk Models ‣ Optimal Delay-Throughput Trade-offs in Mobile Ad-Hoc Networks: Hybrid Random Walk and One-Dimensional Mobility ModelsAn earlier version of this paper appeared in the Proc. of ITA Workshop, 2007.Research presented here was supported in part by a Vodafone Fellowship and NSF grant CNS 05-19691.")), we can conclude that under the Joint Coding-Scheduling Algorithm I, at each super time slot, the $`{2\hspace{0pt}D}/{({25\hspace{0pt}M})}`$ data packets can be successfully recovered with probability at least
<table>
<tbody>
<tr>
<td></td>
<td>1−1n2−55Dn−e−D600M−2e−logD8500−e−D500MlogD.11superscript𝑛255𝐷𝑛superscript𝑒𝐷600𝑀2superscript𝑒𝐷8500superscript𝑒𝐷500𝑀𝐷1-\frac{1}{n^{2}}-\frac{55D}{n}-e^{-\frac{D}{600M}}-2e^{-\frac{\log D}{8500}}-e^{-\frac{D}{500M\log D}}.</td>
<td></td>
</tr>
</tbody>
</table>
The rest of the proof is the same as the proof of Theorem 4 in \[[23](#bib.bib23)\].
Analysis of node distribution: Since $`D = {o\hspace{0pt}\left( {n/{\log^{2}n}} \right)}`$ implies $`{M = {\omega\hspace{0pt}{({\log n})}}}.`$ Inequality ([10](#S4.E10 "In Proof: ‣ IV-B Joint Coding-Scheduling Algorithms ‣ IV Two-Dimensional Hybrid Random Walk Models ‣ Optimal Delay-Throughput Trade-offs in Mobile Ad-Hoc Networks: Hybrid Random Walk and One-Dimensional Mobility ModelsAn earlier version of this paper appeared in the Proc. of ITA Workshop, 2007.Research presented here was supported in part by a Vodafone Fellowship and NSF grant CNS 05-19691.")) can be obtained from the Chernoff bound and union bound.
Analysis of broadcasting step: Consider the broadcasting step. Note that when $`\mathcal{G}`$ occurs, node $`i`$ is selected to broadcast with probability at least $`10/{({11\hspace{0pt}M})}`$ at each time slot. Let $`\mathcal{B}_{i}\hspace{0pt}{\lbrack t\rbrack}`$ denote the event that node $`i`$ is selected to broadcast in time slot $`t.`$ From the Chernoff bound, we have
<table>
<tbody>
<tr>
<td></td>
<td>Pr(∑t=1D1ℬi[t]≥911DM|𝒢)≥1−e−D600M.Prsuperscriptsubscript𝑡1𝐷subscript1subscriptℬ𝑖delimited-[]𝑡conditional911𝐷𝑀𝒢1superscript𝑒𝐷600𝑀\displaystyle\Pr\left(\left.\sum_{t=1}^{D}1_{{\cal B}_{i}[t]}\geq\frac{9}{11}\frac{D}{M}\right|{\cal G}\right)\geq 1-e^{-\frac{D}{600M}}.</td>
<td></td>
<td>(13)</td>
</tr>
</tbody>
</table>
So node $`i`$ broadcasts $`{9\hspace{0pt}D}/{({11\hspace{0pt}M})}`$ coded packets with a high probability. Each coded packet is broadcast to $`{9\hspace{0pt}M}/10`$ relay nodes.
According to Step (2)(ii) of Joint Coding-Scheduling Algorithm I \[[23](#bib.bib23)\], each relay node keeps at most one packet for each source. Consider duplicate packet $`{(i,k,j)}.`$ It could be dropped if node $`j`$ is in the same cell as node $`i`$ and node $`i`$ is selected to broadcast. Thus, the probability that $`(i,k,j)`$ is dropped is at most
<table>
<tbody>
<tr>
<td></td>
<td>1110DMn×1091M=119Dn1110𝐷𝑀𝑛1091𝑀119𝐷𝑛\displaystyle\frac{11}{10}\frac{DM}{n}\times\frac{10}{9}\frac{1}{M}=\frac{11}{9}\frac{D}{n}</td>
<td></td>
<td>(14)</td>
</tr>
</tbody>
</table>
due to the following two facts:
1. \(a\)
Let $`\mathcal{H}_{j\hspace{0pt}i}\hspace{0pt}{\lbrack D\rbrack}`$ denote the event that node $`j`$ is in the same cell as node $`i`$ in at least one of $`D`$ consecutive time slots. Similar to ([7](#S4.E7 "In Proof: ‣ IV-A Upper Bound ‣ IV Two-Dimensional Hybrid Random Walk Models ‣ Optimal Delay-Throughput Trade-offs in Mobile Ad-Hoc Networks: Hybrid Random Walk and One-Dimensional Mobility ModelsAn earlier version of this paper appeared in the Proc. of ITA Workshop, 2007.Research presented here was supported in part by a Vodafone Fellowship and NSF grant CNS 05-19691.")), it can be shown that
<table>
<tbody>
<tr>
<td></td>
<td>Pr(ℋji[D])≤1110DMnPrsubscriptℋ𝑗𝑖delimited-[]𝐷1110𝐷𝑀𝑛\displaystyle\Pr\left({{\cal H}_{ji}[D]}\right)\leq\frac{11}{10}\frac{DM}{n}</td>
<td></td>
<td>(15)</td>
</tr>
</tbody>
</table>
under the delay constraint given in the theorem.
2. \(b\)
When $`\mathcal{G}`$ occurs, node $`i`$ is selected to broadcast with probability at most $`10/{({9\hspace{0pt}M})}`$ at each time slot.
Now suppose source $`i`$ broadcasts $`{\overset{\sim}{D}}_{i}`$ coded packets, so $`{9\hspace{0pt}M\hspace{0pt}{\overset{\sim}{D}}_{i}}/10`$ duplicate copies are generated. Let $`{\overset{\sim}{N}}_{i}^{d}`$ denote the number of duplicate packets of node $`i`$ dropped in the broadcasting step. From the Markov inequality and inequality ([14](#S4.E14 "In Proof: ‣ IV-B Joint Coding-Scheduling Algorithms ‣ IV Two-Dimensional Hybrid Random Walk Models ‣ Optimal Delay-Throughput Trade-offs in Mobile Ad-Hoc Networks: Hybrid Random Walk and One-Dimensional Mobility ModelsAn earlier version of this paper appeared in the Proc. of ITA Workshop, 2007.Research presented here was supported in part by a Vodafone Fellowship and NSF grant CNS 05-19691.")), we have
<table>
<tbody>
<tr>
<td></td>
<td></td>
<td></td>
<td>Pr(N~id≥MD~i50|𝒢,∑t=1D1ℬi[t]=D~i)Prsuperscriptsubscript~𝑁𝑖𝑑conditional𝑀subscript~𝐷𝑖50𝒢superscriptsubscript𝑡1𝐷subscript1subscriptℬ𝑖delimited-[]𝑡subscript~𝐷𝑖\displaystyle\Pr\left(\tilde{N}_{i}^{d}\geq\frac{M\tilde{D}_{i}}{50}\left|{\cal G},\sum_{t=1}^{D}1_{{\cal B}_{i}[t]}=\tilde{D}_{i}\right.\right)</td>
<td></td>
</tr>
<tr>
<td></td>
<td></td>
<td>≤\displaystyle\leq</td>
<td>E[N~id≥MD~i50|𝒢,∑t=1D1ℬi[t]=D~i]MD~i50𝐸delimited-[]superscriptsubscript~𝑁𝑖𝑑conditional𝑀subscript~𝐷𝑖50𝒢superscriptsubscript𝑡1𝐷subscript1subscriptℬ𝑖delimited-[]𝑡subscript~𝐷𝑖𝑀subscript~𝐷𝑖50\displaystyle\frac{E\left[\tilde{N}_{i}^{d}\geq\frac{M\tilde{D}_{i}}{50}\left|{\cal G},\sum_{t=1}^{D}1_{{\cal B}_{i}[t]}=\tilde{D}_{i}\right.\right]}{\frac{M\tilde{D}_{i}}{50}}</td>
<td></td>
</tr>
<tr>
<td></td>
<td></td>
<td>≤\displaystyle\leq</td>
<td>9MD~i10×11D9nMD~i509𝑀subscript~𝐷𝑖1011𝐷9𝑛𝑀subscript~𝐷𝑖50\displaystyle\frac{\frac{9M\tilde{D}_{i}}{10}\times\frac{11D}{9n}}{\frac{M\tilde{D}_{i}}{50}}</td>
<td></td>
</tr>
<tr>
<td></td>
<td></td>
<td>=\displaystyle=</td>
<td>55Dn,55𝐷𝑛\displaystyle\frac{55D}{n},</td>
<td></td>
</tr>
</tbody>
</table>
which implies
<table>
<tbody>
<tr>
<td></td>
<td>Pr(Ai≥45D~i|𝒢,∑t=1D1ℬi[t]=D~i)≥1−55DnPrsubscript𝐴𝑖conditional45subscript~𝐷𝑖𝒢superscriptsubscript𝑡1𝐷subscript1subscriptℬ𝑖delimited-[]𝑡subscript~𝐷𝑖155𝐷𝑛\displaystyle\Pr\left(A_{i}\geq\frac{4}{5}\tilde{D}_{i}\left|{\cal G},\sum_{t=1}^{D}1_{{\cal B}_{i}[t]}=\tilde{D}_{i}\right.\right)\geq 1-\frac{55D}{n}</td>
<td></td>
<td>(16)</td>
</tr>
</tbody>
</table>
since otherwise, more than $`{M\hspace{0pt}{\overset{\sim}{D}}_{i}}/50`$ duplicate packets would be dropped. Inequality ([11](#S4.E11 "In Proof: ‣ IV-B Joint Coding-Scheduling Algorithms ‣ IV Two-Dimensional Hybrid Random Walk Models ‣ Optimal Delay-Throughput Trade-offs in Mobile Ad-Hoc Networks: Hybrid Random Walk and One-Dimensional Mobility ModelsAn earlier version of this paper appeared in the Proc. of ITA Workshop, 2007.Research presented here was supported in part by a Vodafone Fellowship and NSF grant CNS 05-19691.")) follows from inequalities ([16](#S4.E16 "In Proof: ‣ IV-B Joint Coding-Scheduling Algorithms ‣ IV Two-Dimensional Hybrid Random Walk Models ‣ Optimal Delay-Throughput Trade-offs in Mobile Ad-Hoc Networks: Hybrid Random Walk and One-Dimensional Mobility ModelsAn earlier version of this paper appeared in the Proc. of ITA Workshop, 2007.Research presented here was supported in part by a Vodafone Fellowship and NSF grant CNS 05-19691.")) and ([13](#S4.E13 "In Proof: ‣ IV-B Joint Coding-Scheduling Algorithms ‣ IV Two-Dimensional Hybrid Random Walk Models ‣ Optimal Delay-Throughput Trade-offs in Mobile Ad-Hoc Networks: Hybrid Random Walk and One-Dimensional Mobility ModelsAn earlier version of this paper appeared in the Proc. of ITA Workshop, 2007.Research presented here was supported in part by a Vodafone Fellowship and NSF grant CNS 05-19691.")).
Analysis of receiving step: We group every $`{3\hspace{0pt}D}/{\log D}`$ time slots into big time slots, named as b-time-slot and indexed by $`t_{b},`$ and then divided every b-time-slot into three equal parts, indexed by $`t_{b,1},`$ $`t_{b,2}`$ and $`t_{b,3}`$ as in Figure [3](#S4.F3 "Figure 3 ‣ Proof: ‣ IV-B Joint Coding-Scheduling Algorithms ‣ IV Two-Dimensional Hybrid Random Walk Models ‣ Optimal Delay-Throughput Trade-offs in Mobile Ad-Hoc Networks: Hybrid Random Walk and One-Dimensional Mobility ModelsAn earlier version of this paper appeared in the Proc. of ITA Workshop, 2007.Research presented here was supported in part by a Vodafone Fellowship and NSF grant CNS 05-19691.").
[Refer to caption](/html/0705.0326/assets/x3.png)
Figure 3: The Division of A Super Time Slot
We first calculate the probability that coded packet $`(i,k)`$ is delivered in $`t_{b,2}.`$ Let $`\mathcal{H}_{(i,k)}\hspace{0pt}{\lbrack t_{b,2}\rbrack}`$ denote the event that at least one copy of packet $`(i,k)`$ becomes deliverable in $`t_{b,2}.`$ If $`(i,k)`$ is in at least $`{4\hspace{0pt}M}/5`$ relay nodes, we can obtain
<table>
<tbody>
<tr>
<td></td>
<td>Pr(ℋ(i,k)[tb,2])Prsubscriptℋ𝑖𝑘delimited-[]subscript𝑡𝑏2\displaystyle\Pr\left({{\cal H}_{(i,k)}\left[t_{b,2}\right]}\right)</td>
<td>≥\displaystyle\geq</td>
<td>1−((1−MnS2)4DS25logD+1n)4M51superscriptsuperscript1𝑀𝑛superscript𝑆24𝐷superscript𝑆25𝐷1𝑛4𝑀5\displaystyle 1-\left(\left(1-\frac{M}{nS^{2}}\right)^{\frac{4DS^{2}}{5\log D}}+\frac{1}{n}\right)^{\frac{4M}{5}}</td>
<td></td>
<td rowspan="2">(17)</td>
</tr>
<tr>
<td></td>
<td></td>
<td>≥\displaystyle\geq</td>
<td>351logD351𝐷\displaystyle\frac{3}{5}\frac{1}{\log D}</td>
<td></td>
</tr>
</tbody>
</table>
due to the following facts:
1. \(a\)
Given $`{{D/{\log D}} = {\omega\hspace{0pt}{({\log{{n\hspace{0pt}{|{\log S}|}}/S^{6}}})}}},`$ from Lemma [12](#Sx2.EGx56 "Lemma 12 ‣ Appendix B: Properties of Random Walk ‣ Optimal Delay-Throughput Trade-offs in Mobile Ad-Hoc Networks: Hybrid Random Walk and One-Dimensional Mobility ModelsAn earlier version of this paper appeared in the Proc. of ITA Workshop, 2007.Research presented here was supported in part by a Vodafone Fellowship and NSF grant CNS 05-19691.") provided in Appendix B, we know that with probability at least $`{1 - {1/n}},`$ two nodes are in the same RW-cell for at least $`{4\hspace{0pt}D\hspace{0pt}S^{2}}/{({5\hspace{0pt}{\log D}})}`$ time slots.
2. \(b\)
Given two nodes are in the same RW-cell, the probability that they are in the same cell is $`{M/{({n\hspace{0pt}S^{2}})}}.`$
Next note that the duration of $`t_{b,1}`$ and $`t_{b,3}`$ are of a larger order than the mixing time of the random walk (the mixing time is defined in Appendix B). From the definition of the mixing time, we have that at any time slot belonging to $`t_{b,2},`$ the nodes are almost uniformly distributed in the unit square. Let $`\mathcal{D}_{(i,k)}\hspace{0pt}{\lbrack t_{b,2}\rbrack}`$ denote the event that coded packet $`(i,k)`$ is delivered to its destination in $`t_{b,2}.`$ Following the argument used to prove inequality (13) of Theorem 4 in \[[23](#bib.bib23)\], we have
<table>
<tbody>
<tr>
<td></td>
<td>Pr(𝒟(i,k)[tb,2])≥320logD.Prsubscript𝒟𝑖𝑘delimited-[]subscript𝑡𝑏2320𝐷\displaystyle\Pr\left({{\cal D}_{(i,k)}[t_{b,2}]}\right)\geq\frac{3}{20\log D}.</td>
<td></td>
<td>(18)</td>
</tr>
</tbody>
</table>
Now let $`\mathbf{x}_{t}`$ denote the positions of the nodes at time slot $`t,`$ and
<table>
<tbody>
<tr>
<td></td>
<td>𝐗~={𝐱t}t=3(k−1)DlogD+1~𝐗subscriptsubscript𝐱𝑡𝑡3𝑘1𝐷𝐷1\tilde{\mathbf{X}}=\{{\mathbf{x}}_{t}\}_{t=\frac{3(k-1)D}{\log D}+1}</td>
<td></td>
</tr>
</tbody>
</table>
for $`{k = {1,\ldots,{5\hspace{0pt}{\log{D/3}}}}}.`$ Also let $`\mathcal{D}_{(i,k)}`$ denote the event that $`(i,k)`$ is delivered in the receiving step. It is easy to see that $`\mathcal{D}_{(i,k)}`$ occurs if $`\mathcal{D}_{(i,k)}\hspace{0pt}{\lbrack t_{b,2}\rbrack}`$ occurs for some $`{t_{b,2} \in {\{ 1,\ldots,{5\hspace{0pt}{\log{D/3}}}\}}}.`$ Note that $`{\{{\mathcal{D}_{(i,k)}\hspace{0pt}{\lbrack t_{b,2}\rbrack}}\}}_{t_{b,2}}`$ are mutually independent given $`\overset{\sim}{\mathbf{X}},`$ so from inequality ([18](#S4.E18 "In Proof: ‣ IV-B Joint Coding-Scheduling Algorithms ‣ IV Two-Dimensional Hybrid Random Walk Models ‣ Optimal Delay-Throughput Trade-offs in Mobile Ad-Hoc Networks: Hybrid Random Walk and One-Dimensional Mobility ModelsAn earlier version of this paper appeared in the Proc. of ITA Workshop, 2007.Research presented here was supported in part by a Vodafone Fellowship and NSF grant CNS 05-19691.")), we have
<table>
<tbody>
<tr>
<td></td>
<td>Pr(𝒟(i,k)|𝐗~)≥1−(1−3201logD)5logD/3≥15.Prconditionalsubscript𝒟𝑖𝑘~𝐗1superscript13201𝐷5𝐷315\Pr\left(\left.{{\cal D}_{(i,k)}}\right|\tilde{\mathbf{X}}\right)\geq 1-\left(1-\frac{3}{20}\frac{1}{\log D}\right)^{5\log D/3}\geq\frac{1}{5}.</td>
<td></td>
</tr>
</tbody>
</table>
Further since $`{B_{i} \geq {\sum_{k = 1}^{A_{i}}1_{\mathcal{D}_{(i,k)}}}},`$ we can conclude that
<table>
<tbody>
<tr>
<td></td>
<td>E[Bi|𝐗~,Ai≥1625DM]≥16125DM.𝐸delimited-[]conditionalsubscript𝐵𝑖~𝐗subscript𝐴𝑖1625𝐷𝑀16125𝐷𝑀\displaystyle E\left[B_{i}\left|\tilde{\mathbf{X}},A_{i}\geq\frac{16}{25}\frac{D}{M}\right.\right]\geq\frac{16}{125}\frac{D}{M}.</td>
<td></td>
<td>(19)</td>
</tr>
</tbody>
</table>
We next bound the number of distinct coded packets deliverable in $`t_{b}.`$ Similar to inequality ([17](#S4.E17 "In Proof: ‣ IV-B Joint Coding-Scheduling Algorithms ‣ IV Two-Dimensional Hybrid Random Walk Models ‣ Optimal Delay-Throughput Trade-offs in Mobile Ad-Hoc Networks: Hybrid Random Walk and One-Dimensional Mobility ModelsAn earlier version of this paper appeared in the Proc. of ITA Workshop, 2007.Research presented here was supported in part by a Vodafone Fellowship and NSF grant CNS 05-19691.")), we have
<table>
<tbody>
<tr>
<td></td>
<td>Pr(ℋ(i,k)[tb])≤3logD.Prsubscriptℋ𝑖𝑘delimited-[]subscript𝑡𝑏3𝐷\Pr\left({{\cal H}_{(i,k)}\left[t_{b}\right]}\right)\leq\frac{3}{\log D}.</td>
<td></td>
</tr>
</tbody>
</table>
Note that no two duplicate packets from node $`i`$ are in one relay node, so $`{\{{\mathcal{H}_{(i,k)}\hspace{0pt}{\lbrack t_{b}\rbrack}}\}}_{k}`$ are mutually independent. From the Chernoff bound, we have
<table>
<tbody>
<tr>
<td></td>
<td>Pr(∑k=1D/M1ℋ(i,k)[tb]≤165DMlogD)≥1−e−D400MlogD.Prsuperscriptsubscript𝑘1𝐷𝑀subscript1subscriptℋ𝑖𝑘delimited-[]subscript𝑡𝑏165𝐷𝑀𝐷1superscript𝑒𝐷400𝑀𝐷\Pr\left(\sum_{k=1}^{D/M}1_{{\cal H}_{(i,k)}\left[t_{b}\right]}\leq\frac{16}{5}\frac{D}{M\log D}\right)\geq 1-e^{-\frac{D}{400M\log D}}.</td>
<td></td>
</tr>
</tbody>
</table>
Let $`{\overset{\sim}{\mathcal{F}}}_{i}`$ denote the event that node $`i`$ obtains no more than $`{16\hspace{0pt}D}/{({5\hspace{0pt}M\hspace{0pt}{\log D}})}`$ coded packets at each b-time-slot in the receiving step. From the union bound, we have that for sufficiently large $`n,`$
<table>
<tbody>
<tr>
<td></td>
<td>Pr(ℱ~i)≥1−(53logD)(e−D400MlogD)≥1−e−D500MlogD.Prsubscript~ℱ𝑖153𝐷superscript𝑒𝐷400𝑀𝐷1superscript𝑒𝐷500𝑀𝐷\displaystyle\Pr\left({\tilde{{\cal F}}_{i}}\right)\geq 1-\left(\frac{5}{3}\log D\right)\left(e^{-\frac{D}{400M\log D}}\right)\geq 1-e^{-\frac{D}{500M\log D}}.</td>
<td></td>
<td>(20)</td>
</tr>
</tbody>
</table>
Now let $`B_{i}\hspace{0pt}{(\overset{\sim}{\mathbf{X}},A_{i},\mathcal{F}_{i})}`$ denote the number of distinct coded packets delivered to the destination of node $`i`$ given $`{(\overset{\sim}{\mathbf{X}},A_{i},\mathcal{F}_{i})},`$ and $`\mathbf{X}_{t_{b}}`$ denote an $`n \times {({{3\hspace{0pt}D}/{\log D}})}`$ matrix where the $`(i,t)`$ entry is the position of node $`i`$ at the $`t^{th}`$ time slot of b-time-slot $`t_{b}.`$ It is easy to see that the value of $`B_{i}\hspace{0pt}{(\overset{\sim}{\mathbf{X}},A_{i},\mathcal{F}_{i})}`$ is determined by $`{\{\mathbf{X}_{t_{b}}\}},`$ i.e., there exists a function $`f_{(\overset{\sim}{\mathbf{X}},A_{i},\mathcal{F}_{i})}`$ such that
<table>
<tbody>
<tr>
<td></td>
<td>Bi(𝐗~,Ai,ℱi)=f(𝐗~,Ai,ℱi)(𝐗1,…,𝐗5logD/3).subscript𝐵𝑖~𝐗subscript𝐴𝑖subscriptℱ𝑖subscript𝑓~𝐗subscript𝐴𝑖subscriptℱ𝑖subscript𝐗1…subscript𝐗5𝐷3B_{i}(\tilde{\mathbf{X}},A_{i},{\cal F}_{i})=f_{(\tilde{\mathbf{X}},A_{i},{\cal F}_{i})}\left(\mathbf{X}_{1},\ldots,\mathbf{X}_{5\log D/3}\right).</td>
<td></td>
</tr>
</tbody>
</table>
From the definition of $`\mathcal{F}_{i},`$ function $`f_{(\overset{\sim}{\mathbf{X}},A_{i},\mathcal{F}_{i})}`$ satisfies the following condition,
<table>
<tbody>
<tr>
<td></td>
<td>|f(𝐗~,Ai,ℱi)(𝐗1,…,𝐗tb−1,𝐗tb,𝐗tb+1,…,𝐗5logD/3)−\displaystyle\left|f_{(\tilde{\mathbf{X}},A_{i},{\cal F}_{i})}\left(\mathbf{X}_{1},\ldots,\mathbf{X}_{t_{b}-1},\mathbf{X}_{t_{b}},\mathbf{X}_{t_{b}+1},\ldots,\mathbf{X}_{5\log D/3}\right)-\right.</td>
<td></td>
<td></td>
<td></td>
<td></td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td>f(𝐗~,Ai,ℱi)(𝐗1,…,𝐗tb−1,𝐘tb,𝐗tb+1,…,𝐗5logD/3)|≤conditionalsubscript𝑓~𝐗subscript𝐴𝑖subscriptℱ𝑖subscript𝐗1…subscript𝐗subscript𝑡𝑏1subscript𝐘subscript𝑡𝑏subscript𝐗subscript𝑡𝑏1…subscript𝐗5𝐷3\displaystyle\left.f_{(\tilde{\mathbf{X}},A_{i},{\cal F}_{i})}\left(\mathbf{X}_{1},\ldots,\mathbf{X}_{t_{b}-1},\mathbf{Y}_{t_{b}},\mathbf{X}_{t_{b}+1},\ldots,\mathbf{X}_{5\log D/3}\right)\right|\leq</td>
<td>165DMlogD.165𝐷𝑀𝐷\displaystyle\frac{16}{5}\frac{D}{M\log D}.</td>
<td></td>
<td></td>
<td>(21)</td>
</tr>
</tbody>
</table>
It is easy to see that $`\{\mathbf{X}_{t_{b}}\}`$ are mutually independent given $`{(\overset{\sim}{\mathbf{X}},A_{i},{\overset{\sim}{\mathcal{F}}}_{i})}.`$ Then invoking Azuma-Hoeffding inequality provided in Appendix A, we can conclude that
<table>
<tbody>
<tr>
<td></td>
<td></td>
<td>Pr(Bi(𝐗~,Ai,ℱi)≥E[Bi|𝐗~,Ai,ℱ~i]−1125DM)Prsubscript𝐵𝑖~𝐗subscript𝐴𝑖subscriptℱ𝑖𝐸delimited-[]conditionalsubscript𝐵𝑖~𝐗subscript𝐴𝑖subscript~ℱ𝑖1125𝐷𝑀\displaystyle\Pr\left(B_{i}{(\tilde{\mathbf{X}},A_{i},{\cal F}_{i})}\geq E\left[B_{i}\left|\tilde{\mathbf{X}},A_{i},\tilde{{\cal F}}_{i}\right.\right]-\frac{1}{125}\frac{D}{M}\right)</td>
<td></td>
<td></td>
<td></td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td>≥\displaystyle\geq</td>
<td>1−2e−logD850012superscript𝑒𝐷8500\displaystyle 1-2e^{-\frac{\log D}{8500}}</td>
<td></td>
<td></td>
<td>(22)</td>
</tr>
</tbody>
</table>
holds for any $`\overset{\sim}{\mathbf{X}}`$ and $`A_{i}.`$ Inequality ([12](#S4.E12 "In Proof: ‣ IV-B Joint Coding-Scheduling Algorithms ‣ IV Two-Dimensional Hybrid Random Walk Models ‣ Optimal Delay-Throughput Trade-offs in Mobile Ad-Hoc Networks: Hybrid Random Walk and One-Dimensional Mobility ModelsAn earlier version of this paper appeared in the Proc. of ITA Workshop, 2007.Research presented here was supported in part by a Vodafone Fellowship and NSF grant CNS 05-19691.")) follows from inequalities ([19](#S4.E19 "In Proof: ‣ IV-B Joint Coding-Scheduling Algorithms ‣ IV Two-Dimensional Hybrid Random Walk Models ‣ Optimal Delay-Throughput Trade-offs in Mobile Ad-Hoc Networks: Hybrid Random Walk and One-Dimensional Mobility ModelsAn earlier version of this paper appeared in the Proc. of ITA Workshop, 2007.Research presented here was supported in part by a Vodafone Fellowship and NSF grant CNS 05-19691.")), ([20](#S4.E20 "In Proof: ‣ IV-B Joint Coding-Scheduling Algorithms ‣ IV Two-Dimensional Hybrid Random Walk Models ‣ Optimal Delay-Throughput Trade-offs in Mobile Ad-Hoc Networks: Hybrid Random Walk and One-Dimensional Mobility ModelsAn earlier version of this paper appeared in the Proc. of ITA Workshop, 2007.Research presented here was supported in part by a Vodafone Fellowship and NSF grant CNS 05-19691.")) and ([22](#S4.E22 "In Proof: ‣ IV-B Joint Coding-Scheduling Algorithms ‣ IV Two-Dimensional Hybrid Random Walk Models ‣ Optimal Delay-Throughput Trade-offs in Mobile Ad-Hoc Networks: Hybrid Random Walk and One-Dimensional Mobility ModelsAn earlier version of this paper appeared in the Proc. of ITA Workshop, 2007.Research presented here was supported in part by a Vodafone Fellowship and NSF grant CNS 05-19691.")).
Proof of (2): We consider one super time slot which consists of $`16\hspace{0pt}D`$ time slots, and calculate the probability that the $`D/7`$ data packets from node $`i`$ are fully recovered at the destination. Let $`{M_{1} = \sqrt[3]{n/D}},`$ which is the mean number of nodes in each cell at the broadcasting step. Following the analysis above, we can prove that the following events happen with high probability.
Node distribution: All cells are good during the entire super-time-slot with high probability, i.e.,
<table>
<tbody>
<tr>
<td></td>
<td>Pr(𝒢)≥1−1n2.Pr𝒢11superscript𝑛2\displaystyle\Pr\left({\cal G}\right)\geq 1-\frac{1}{n^{2}}.</td>
<td></td>
</tr>
</tbody>
</table>
Broadcasting: At least $`{4\hspace{0pt}D}/5`$ coded packets from a source are successfully duplicated after the broadcasting step with high probability. Specifically, we have
<table>
<tbody>
<tr>
<td></td>
<td>Pr(Ai≥45D|𝒢)≥1−50M12,Prsubscript𝐴𝑖conditional45𝐷𝒢150superscriptsubscript𝑀12\displaystyle\Pr\left(\left.A_{i}\geq\frac{4}{5}D\right|{\cal G}\right)\geq 1-\frac{50}{M_{1}^{2}},</td>
<td></td>
<td>(23)</td>
</tr>
</tbody>
</table>
where a coded packet is said to be successfully duplicated if it is in $`{4\hspace{0pt}M_{1}}/5`$ distinct relay nodes.
Receiving: At least $`D/6`$ distinct coded packets from a source are delivered to its destination after the receiving step with high probability. Specifically, we have
<table>
<tbody>
<tr>
<td></td>
<td>Pr(Bi≥D6|Ai≥45D)≥1−2e−logD800.Prsubscript𝐵𝑖conditional𝐷6subscript𝐴𝑖45𝐷12superscript𝑒𝐷800\displaystyle\Pr\left(\left.B_{i}\geq\frac{D}{6}\right|A_{i}\geq\frac{4}{5}D\right)\geq 1-2e^{-\frac{\log D}{800}}.</td>
<td></td>
<td>(24)</td>
</tr>
</tbody>
</table>
From inequalities ([10](#S4.E10 "In Proof: ‣ IV-B Joint Coding-Scheduling Algorithms ‣ IV Two-Dimensional Hybrid Random Walk Models ‣ Optimal Delay-Throughput Trade-offs in Mobile Ad-Hoc Networks: Hybrid Random Walk and One-Dimensional Mobility ModelsAn earlier version of this paper appeared in the Proc. of ITA Workshop, 2007.Research presented here was supported in part by a Vodafone Fellowship and NSF grant CNS 05-19691.")), ([23](#S4.E23 "In Proof: ‣ IV-B Joint Coding-Scheduling Algorithms ‣ IV Two-Dimensional Hybrid Random Walk Models ‣ Optimal Delay-Throughput Trade-offs in Mobile Ad-Hoc Networks: Hybrid Random Walk and One-Dimensional Mobility ModelsAn earlier version of this paper appeared in the Proc. of ITA Workshop, 2007.Research presented here was supported in part by a Vodafone Fellowship and NSF grant CNS 05-19691.")) and ([24](#S4.E24 "In Proof: ‣ IV-B Joint Coding-Scheduling Algorithms ‣ IV Two-Dimensional Hybrid Random Walk Models ‣ Optimal Delay-Throughput Trade-offs in Mobile Ad-Hoc Networks: Hybrid Random Walk and One-Dimensional Mobility ModelsAn earlier version of this paper appeared in the Proc. of ITA Workshop, 2007.Research presented here was supported in part by a Vodafone Fellowship and NSF grant CNS 05-19691.")), we can conclude that under the modified Joint Coding-Scheduling Algorithm II, at each super time slot, the $`D/7`$ source packets can be successfully recovered with probability at least
<table>
<tbody>
<tr>
<td></td>
<td>1−1n2−50M12−2e−logD800,11superscript𝑛250superscriptsubscript𝑀122superscript𝑒𝐷8001-\frac{1}{n^{2}}-\frac{50}{M_{1}^{2}}-2e^{-\frac{\log D}{800}},</td>
<td></td>
</tr>
</tbody>
</table>
and theorem holds. ∎
## V One-Dimensional I.I.D. Mobility Models
###
V-A
Upper Bounds
######
Theorem 3
Consider the one-dimensional i.i.d. mobility model, and assume that Assumption 1-3 hold. We have following results:
1. \(1\)
For fast mobiles,
<table>
<tbody>
<tr>
<td></td>
<td>8WT2πΔ23n3/2(D3/2+1)≥E[Λ[T]];8𝑊𝑇32𝜋superscriptΔ232𝑛32𝐷1𝐸delimited-[]Λdelimited-[]𝑇8WT\sqrt[3]{\frac{2}{\pi\Delta^{2}}}\sqrt[3/2]{n}(\sqrt[3/2]{D}+1)\geq E[\Lambda[T]];</td>
<td></td>
</tr>
</tbody>
</table>
2. \(2\)
For slow mobiles,
<table>
<tbody>
<tr>
<td></td>
<td>4WT4πΔ24n3/4(D+1)≥E[Λ[T]].4𝑊𝑇44𝜋superscriptΔ234𝑛𝐷1𝐸delimited-[]Λdelimited-[]𝑇4WT\sqrt[4]{\frac{4}{\pi\Delta^{2}}}\sqrt[3/4]{n}(\sqrt{D}+1)\geq E[\Lambda[T]].</td>
<td></td>
</tr>
</tbody>
</table>
###### Proof:
Recall that $`{\overset{\sim}{L}}_{b}`$ is the minimum distance between node $`d_{b}`$ and node $`c_{b}`$ from time slot $`t_{b}`$ to $`{{t_{b} + D} - 1}.`$ If the orbits of node $`c_{b}`$ and $`d_{b}`$ are vertical to each other, then $`{\overset{\sim}{L}}_{b} \leq L`$ holds only if at some time slot $`t,`$ node $`c_{b}`$ and $`d_{b}`$ are in the square with side length $`2\hspace{0pt}L`$ as in Figure [4](#S5.F4 "Figure 4 ‣ Proof: ‣ V-A Upper Bounds ‣ V One-Dimensional I.I.D. Mobility Models ‣ Optimal Delay-Throughput Trade-offs in Mobile Ad-Hoc Networks: Hybrid Random Walk and One-Dimensional Mobility ModelsAn earlier version of this paper appeared in the Proc. of ITA Workshop, 2007.Research presented here was supported in part by a Vodafone Fellowship and NSF grant CNS 05-19691."). In this case, we have
<table>
<tbody>
<tr>
<td></td>
<td>Pr(L~b≤L)≤1−(1−4L2)D.Prsubscript~𝐿𝑏𝐿1superscript14superscript𝐿2𝐷\Pr\left(\tilde{L}_{b}\leq L\right)\leq 1-(1-4L^{2})^{D}.</td>
<td></td>
</tr>
</tbody>
</table>
If the orbits of node $`c_{b}`$ and $`d_{b}`$ are parallel to each other, then it is easy to verify that
<table>
<tbody>
<tr>
<td></td>
<td>Pr(L~b≤L)≤1−(1−2L)D.Prsubscript~𝐿𝑏𝐿1superscript12𝐿𝐷\Pr\left(\tilde{L}_{b}\leq L\right)\leq 1-(1-2L)^{D}.</td>
<td></td>
</tr>
</tbody>
</table>
Thus for $`{L \leq {1/2}},`$ we can conclude that
<table>
<tbody>
<tr>
<td></td>
<td>Pr(L~b≤L)≤1−(1−2L)D≤2LD.Prsubscript~𝐿𝑏𝐿1superscript12𝐿𝐷2𝐿𝐷\Pr\left(\tilde{L}_{b}\leq L\right)\leq 1-(1-2L)^{D}\leq 2LD.</td>
<td></td>
</tr>
</tbody>
</table>
The rest of the proof is similar to Theorem 3 and Theorem 6 in \[[23](#bib.bib23)\].
[Refer to caption](/html/0705.0326/assets/x4.png)
Figure 4: The Two Orbits are Vertical to Each Other
∎
###
V-B
Joint Coding-Scheduling Algorithm for Fast Mobility
Choose
<table>
<tbody>
<tr>
<td></td>
<td>M=nD23.𝑀3𝑛superscript𝐷2M=\sqrt[3]{\frac{n}{D^{2}}}.</td>
<td></td>
</tr>
</tbody>
</table>
We divide the unit square into $`\sqrt{n/M}`$ horizontal rectangles, named as H-rectangles; and $`\sqrt{n/M}`$ vertical rectangles, named as V-rectangles as in Figure [5](#S5.F5 "Figure 5 ‣ V-B Joint Coding-Scheduling Algorithm for Fast Mobility ‣ V One-Dimensional I.I.D. Mobility Models ‣ Optimal Delay-Throughput Trade-offs in Mobile Ad-Hoc Networks: Hybrid Random Walk and One-Dimensional Mobility ModelsAn earlier version of this paper appeared in the Proc. of ITA Workshop, 2007.Research presented here was supported in part by a Vodafone Fellowship and NSF grant CNS 05-19691."). A packet is said to be destined to a rectangle if the orbit of its destination is contained in the rectangle.
[Refer to caption](/html/0705.0326/assets/x5.png)
Figure 5: The Unit Square is Divided into H(V)-Rectangles
The algorithms for the one-dimensional i.i.d. mobility model has four steps. The first step is the Raptor encoding. The second step is the broadcasting step. In this step, the H(V)-nodes broadcast coded packets to V(H)-nodes. The third step is the transporting step, where the V(H)-nodes transport the H(V)-packets to the H(V)-rectangles containing the orbits of corresponding destinations, and then broadcast packets to the H(V)-nodes whose orbits are contained in the rectangles. After the third step, all duplicate packets are carried by the nodes that move parallel with the destinations and their orbit distance is less than $`\sqrt{M/n}.`$ The fourth step is the receiving step, where the packets are delivered to the destinations.
Since duplicate copies are generated in both the broadcasting step and the transporting step. To distinguish them, we name the duplicate packets generated at the broadcasting step as B-duplicate packets, and the duplicate packets generated at the transporting step as T-duplicate packets. Also we say a B-duplicate packet is transportable if it is in the rectangle containing the orbit of the destination of the packet.
Consider a cell with area $`\overset{\sim}{A}`$ and use $`{\overset{\sim}{M}}^{H\hspace{0pt}{(V)}}\hspace{0pt}{\lbrack t\rbrack}`$ to denote the number of H(V)-nodes in the cell. For the one-dimensional mobility model, a cell is said to be a *good cell* at time slot $`t`$ if
<table>
<tbody>
<tr>
<td></td>
<td>910A~n+1≤M~H(V)[t]≤1110A~n.910~𝐴𝑛1superscript~𝑀𝐻𝑉delimited-[]𝑡1110~𝐴𝑛\frac{9}{10}\tilde{A}n+1\leq\tilde{M}^{H(V)}[t]\leq\frac{11}{10}\tilde{A}n.</td>
<td></td>
</tr>
</tbody>
</table>
Next we present the Joint Coding-Scheduling Algorithm III, which achieves the maximum throughput obtain in Theorem [4](#Sx2.EGx30 "Theorem 4 ‣ V-B Joint Coding-Scheduling Algorithm for Fast Mobility ‣ V One-Dimensional I.I.D. Mobility Models ‣ Optimal Delay-Throughput Trade-offs in Mobile Ad-Hoc Networks: Hybrid Random Walk and One-Dimensional Mobility ModelsAn earlier version of this paper appeared in the Proc. of ITA Workshop, 2007.Research presented here was supported in part by a Vodafone Fellowship and NSF grant CNS 05-19691."). Note that in the following algorithm, each time slot is further divide into $`C`$ mini-time slots, and each cell is guaranteed to be active in at least one of mini-time slot within each time slot.
Joint Coding-Scheduling Algorithm III: The unit square is divided into a regular lattice with $`n/M`$ cells, and the packet size is chosen to be $`{W/{({2\hspace{0pt}C})}}.`$ We group every $`7\hspace{0pt}D`$ time slots into a super time slot. At each super time slot, the nodes transmit packets as follows.
1. \(1\)
Raptor Encoding: Each source takes $`{2\hspace{0pt}D}/{({35\hspace{0pt}M})}`$ data packets, and uses the Raptor codes to generate $`D/M`$ coded packets.
2. \(2\)
Broadcasting: This step consists of $`D`$ time slots. At each time slot, the nodes execute the following tasks:
1. \(i\)
In each good cell, one H-node and one V-node are randomly selected. If the selected H(V)-node has never been in the current cell before and not already transmitted all of its $`D/M`$ coded packets, then it broadcasts a coded packet that was not previous transmitted to $`{9\hspace{0pt}M}/10`$ V(H)-nodes in the cell during the mini-time slot allocated to that cell.
2. \(ii\)
All nodes check the duplicate packets they have. If more than one B-duplicate packets are destined to the same rectangle, randomly keep one and drop the others.
3. \(3\)
Transporting: This step consists of $`D`$ time slots. At each time slot, the nodes do the following:
1. \(i\)
Suppose that node $`j`$ is a V-node ,and carries B-duplicate packet $`{(i,k,j)}.`$ Node $`j`$ broadcasts $`(i,k,j)`$ to $`{9\hspace{0pt}M}/10`$ H-nodes in the same cell if following conditions hold: (a) Node $`j`$ is in a good cell; (b) B-duplicate packet $`(i,k,j)`$ is the only transportable H-packet in the cell.
2. \(ii\)
Each node checks the T-duplicate packets it carries. If more than one T-duplicate packet has the same destination, randomly keep one and drop the others.
4. \(4\)
Receiving: This step consists of $`5\hspace{0pt}D`$ time slots. At each time slot, if there are no more than two deliverable packets in the cell, the deliverable packets are delivered to the destinations with one-hop transmissions. At the end of this step, all undelivered packets are dropped. The destinations decode the received coded packets using Raptor decoding.
######
Theorem 4
Consider Joint Coding-Scheduling Algorithm III. Suppose $`D`$ is $`\omega\hspace{0pt}{(\sqrt[5]{n})}`$ and $`{o\hspace{0pt}\left( {\sqrt{n}/\sqrt[{3/2}]{\log n}} \right)},`$ and the delay constraint is $`{7\hspace{0pt}D}.`$ Then given any $`{\epsilon > 0},`$ there exists $`n_{0}`$ such that for any $`{n \geq n_{0}},`$ every data packet sent out can be recovered at the destination with probability at least $`{1 - \epsilon},`$ and furthermore
<table>
<tbody>
<tr>
<td></td>
<td>limT→∞Pr(Λi[T]T≥(W980C)D2n3 ∀i)=1.subscript→𝑇PrsubscriptΛ𝑖delimited-[]𝑇𝑇𝑊980𝐶3superscript𝐷2𝑛 for-all𝑖1\displaystyle\lim_{T\rightarrow\infty}\Pr\left(\frac{\Lambda_{i}[T]}{T}\geq\left(\frac{W}{980C}\right)\sqrt[3]{\frac{D^{2}}{n}}\hbox{ }\forall i\right)=1.</td>
<td></td>
</tr>
</tbody>
</table>
###### Proof:
Consider one super time slot and let $`\mathcal{G}`$ denote the event that all cells are good in the super time slot. The proof will show the following events happen with high probability.
Node distribution: All cells are good during the entire super-time-slot with high probability. Specifically, it is easy to verify that
<table>
<tbody>
<tr>
<td></td>
<td>Pr(𝒢)≥1−1n2.Pr𝒢11superscript𝑛2\displaystyle\Pr\left({\cal G}\right)\geq 1-\frac{1}{n^{2}}.</td>
<td></td>
<td>(25)</td>
</tr>
</tbody>
</table>
Broadcasting: At least $`{2\hspace{0pt}D}/{({3\hspace{0pt}M})}`$ coded packets from a source are successfully duplicated after the broadcasting step with high probability, where a coded packet is said to be successfully duplicated if it has at least $`{4\hspace{0pt}M}/5`$ B-duplicate packets. Specifically, we will show
<table>
<tbody>
<tr>
<td></td>
<td>Pr(Ai≥23DM|𝒢)≥1−40M.Prsubscript𝐴𝑖conditional23𝐷𝑀𝒢140𝑀\displaystyle\Pr\left(\left.A_{i}\geq\frac{2}{3}\frac{D}{M}\right|{\cal G}\right)\geq 1-\frac{40}{M}.</td>
<td></td>
<td>(26)</td>
</tr>
</tbody>
</table>
Transporting: At least $`{9\hspace{0pt}D}/{({70\hspace{0pt}M})}`$ coded packets from a source are successfully transported after the transporting step with high probability, where a coded packet is said to be successfully transported if it has at least $`{4\hspace{0pt}M}/5`$ T-duplicate copies. Letting $`C_{i}`$ denote the number of successfully transported packets from node $`i,`$ we will show
<table>
<tbody>
<tr>
<td></td>
<td>Pr(Ci≥970DM|𝒢,Ai≥23DM)≥1−3e−D1500M−100M2.Prsubscript𝐶𝑖conditional970𝐷𝑀𝒢subscript𝐴𝑖23𝐷𝑀13superscript𝑒𝐷1500𝑀100superscript𝑀2\displaystyle\Pr\left(\left.C_{i}\geq\frac{9}{70}\frac{D}{M}\right|{\cal G},A_{i}\geq\frac{2}{3}\frac{D}{M}\right)\geq 1-3e^{-\frac{D}{1500M}}-\frac{100}{M^{2}}.</td>
<td></td>
<td>(27)</td>
</tr>
</tbody>
</table>
Receiving: At least $`{9\hspace{0pt}D}/{({140\hspace{0pt}M})}`$ distinct coded packets from a source are delivered to its destination after the receiving step. Specifically, we will show
<table>
<tbody>
<tr>
<td></td>
<td>Pr(Bi≥9140DM|Ci≥970DM)≥1−2e−D8000M.Prsubscript𝐵𝑖conditional9140𝐷𝑀subscript𝐶𝑖970𝐷𝑀12superscript𝑒𝐷8000𝑀\displaystyle\Pr\left(\left.B_{i}\geq\frac{9}{140}\frac{D}{M}\right|C_{i}\geq\frac{9}{70}\frac{D}{M}\right)\geq 1-2e^{-\frac{D}{8000M}}.</td>
<td></td>
<td>(28)</td>
</tr>
</tbody>
</table>
If the claims above hold, the probability that the $`{D/10}\hspace{0pt}M`$ data packets are fully recovered in one super time slot is at least
<table>
<tbody>
<tr>
<td></td>
<td>1−1n2−40M−3e−D1500M−100M2−2e−D8000M.11superscript𝑛240𝑀3superscript𝑒𝐷1500𝑀100superscript𝑀22superscript𝑒𝐷8000𝑀1-\frac{1}{n^{2}}-\frac{40}{M}-3e^{-\frac{D}{1500M}}-\frac{100}{M^{2}}-2e^{-\frac{D}{8000M}}.</td>
<td></td>
</tr>
</tbody>
</table>
The theorem follows from a similar argument provided in Theorem 4 of \[[23](#bib.bib23)\].
Analysis of broadcasting step: Assume that $`\mathcal{G}`$ occurs, then at each time slot, node $`i`$ is selected with probability $`{10/{({11\hspace{0pt}M})}}.`$ Note that there are $`\sqrt{n/M}`$ cells on the orbit of node $`i,`$ and node $`i`$ is uniformly randomly positioned in one of the cells. Thus, the number of coded packets broadcast by node $`i`$ is equal to the number of non-empty bins of following balls-and-bins problem.
*Balls-and-Bins Problem:* Suppose we have $`\sqrt{n/M}`$ bins and one trash can. At each time slot, we drop a ball. Each bin receives the ball with probability $`{10/{({11\hspace{0pt}\sqrt{n\hspace{0pt}M}})}},`$ and the trash can receives the ball with probability $`{1 - {10/{({11\hspace{0pt}M})}}}.`$ Repeat this $`D`$ times, i.e., $`D`$ balls are dropped.
From Lemma [9](#Thmthm9 "Lemma 9 ‣ Appendix A: Probability Results ‣ Optimal Delay-Throughput Trade-offs in Mobile Ad-Hoc Networks: Hybrid Random Walk and One-Dimensional Mobility ModelsAn earlier version of this paper appeared in the Proc. of ITA Workshop, 2007.Research presented here was supported in part by a Vodafone Fellowship and NSF grant CNS 05-19691.") provided in Appendix A, we have
<table>
<tbody>
<tr>
<td></td>
<td>Pr(∑k=1D1ℬi[t]≥9D11M|𝒢)≥1−e−3D1000M.Prsuperscriptsubscript𝑘1𝐷subscript1subscriptℬ𝑖delimited-[]𝑡conditional9𝐷11𝑀𝒢1superscript𝑒3𝐷1000𝑀\Pr\left(\left.\sum_{k=1}^{D}1_{{\cal B}_{i}[t]}\geq\frac{9D}{11M}\right|{\cal G}\right)\geq 1-e^{-\frac{3D}{1000M}}.</td>
<td></td>
</tr>
</tbody>
</table>
We say two nodes are competitive with each other if the orbits of their destinations are contained in the same rectangle, so each node has $`\sqrt{M\hspace{0pt}n} - 1`$ competitive nodes. Suppose that node $`i`$ is an H-node and node $`j`$ is a V-node. Let $`{\overset{\sim}{N}}_{i,j}^{c}\hspace{0pt}{(t)}`$ denote the number of node $`i`$’s competitive nodes in the V-rectangle containing the orbit of node $`j`$ at time slot $`t.`$ Since nodes are uniformly, randomly positioned on their orbits, from the Chernoff bound, we have
<table>
<tbody>
<tr>
<td></td>
<td>Pr(N~i,jc(t)≤1110M)≥1−e−M300.Prsuperscriptsubscript~𝑁𝑖𝑗𝑐𝑡1110𝑀1superscript𝑒𝑀300\displaystyle\Pr\left(\tilde{N}_{i,j}^{c}(t)\leq\frac{11}{10}M\right)\geq 1-e^{-\frac{M}{300}}.</td>
<td></td>
<td>(29)</td>
</tr>
</tbody>
</table>
Now consider B-duplicate packet $`(i,k,j)`$ and assume that node $`z,`$ a competitive of node $`i,`$ is in the V-rectangle containing the orbit of node $`i.`$ Then $`(i,k,j)`$ might be dropped when it is in the same cell as node $`z,`$ and node $`z`$ is selected to broadcast. The probability of this event is at most
<table>
<tbody>
<tr>
<td></td>
<td>Mn×109M.𝑀𝑛109𝑀\displaystyle\sqrt{\frac{M}{n}}\times\frac{10}{9M}.</td>
<td></td>
<td>(30)</td>
</tr>
</tbody>
</table>
From ([29](#S5.E29 "In Proof: ‣ V-B Joint Coding-Scheduling Algorithm for Fast Mobility ‣ V One-Dimensional I.I.D. Mobility Models ‣ Optimal Delay-Throughput Trade-offs in Mobile Ad-Hoc Networks: Hybrid Random Walk and One-Dimensional Mobility ModelsAn earlier version of this paper appeared in the Proc. of ITA Workshop, 2007.Research presented here was supported in part by a Vodafone Fellowship and NSF grant CNS 05-19691.")), ([30](#S5.E30 "In Proof: ‣ V-B Joint Coding-Scheduling Algorithm for Fast Mobility ‣ V One-Dimensional I.I.D. Mobility Models ‣ Optimal Delay-Throughput Trade-offs in Mobile Ad-Hoc Networks: Hybrid Random Walk and One-Dimensional Mobility ModelsAn earlier version of this paper appeared in the Proc. of ITA Workshop, 2007.Research presented here was supported in part by a Vodafone Fellowship and NSF grant CNS 05-19691.")) and the union bound, we can conclude that the probability that $`(i,k,j)`$ is dropped at time slot $`t`$ is at most
<table>
<tbody>
<tr>
<td></td>
<td>e−M300+1110M×Mn×109M=e−M300+119Mn,superscript𝑒𝑀3001110𝑀𝑀𝑛109𝑀superscript𝑒𝑀300119𝑀𝑛\displaystyle e^{-\frac{M}{300}}+\frac{11}{10}M\times\sqrt{\frac{M}{n}}\times\frac{10}{9M}=e^{-\frac{M}{300}}+\frac{11}{9}\sqrt{\frac{M}{n}},</td>
<td></td>
<td>(31)</td>
</tr>
</tbody>
</table>
which implies that the probability of $`(i,k,j)`$ dropped in the broadcasting step is at most
<table>
<tbody>
<tr>
<td></td>
<td>1−(1−e−M300−119Mn)D≤De−M300+119M1superscript1superscript𝑒𝑀300119𝑀𝑛𝐷𝐷superscript𝑒𝑀300119𝑀\displaystyle 1-\left(1-e^{-\frac{M}{300}}-\frac{11}{9}\sqrt{\frac{M}{n}}\right)^{D}\leq De^{-\frac{M}{300}}+\frac{11}{9M}</td>
<td></td>
</tr>
</tbody>
</table>
Inequality ([26](#S5.E26 "In Proof: ‣ V-B Joint Coding-Scheduling Algorithm for Fast Mobility ‣ V One-Dimensional I.I.D. Mobility Models ‣ Optimal Delay-Throughput Trade-offs in Mobile Ad-Hoc Networks: Hybrid Random Walk and One-Dimensional Mobility ModelsAn earlier version of this paper appeared in the Proc. of ITA Workshop, 2007.Research presented here was supported in part by a Vodafone Fellowship and NSF grant CNS 05-19691.")) follows from above inequality and the Markov inequality.
Analysis of transporting step: Consider an H-node $`i.`$ Let $`C_{(i,k)}`$ denote the number of B-duplicate packets which are contained in the V-rectangle where $`(i,k)`$ broadcast, and are destined to the same H-rectangle as node $`i.`$ Note the following facts:
1. \(a\)
Each node has $`\sqrt{M\hspace{0pt}n} - 1`$ competitive nodes.
2. \(b\)
Each H-node broadcasts at most $`D/M`$ coded packets. The probability that a coded packet broadcast in a specific V-rectangle is at most
<table>
<tbody>
<tr>
<td></td>
<td>DM×Mn=DnM.𝐷𝑀𝑀𝑛𝐷𝑛𝑀\frac{D}{M}\times\sqrt{\frac{M}{n}}=\frac{D}{\sqrt{nM}}.</td>
<td></td>
</tr>
</tbody>
</table>
3. \(c\)
Each broadcast generates $`{9\hspace{0pt}M}/10`$ duplicate copies.
Thus, from the Chernoff bound, we have that
<table>
<tbody>
<tr>
<td></td>
<td>Pr(C(i,k)≤DM)≥1−e−D300M,Prsubscript𝐶𝑖𝑘𝐷𝑀1superscript𝑒𝐷300𝑀\Pr\left(C_{(i,k)}\leq DM\right)\geq 1-e^{-\frac{D}{300M}},</td>
<td></td>
</tr>
</tbody>
</table>
which implies that for sufficiently large $`n,`$
<table>
<tbody>
<tr>
<td></td>
<td>Pr(C(i,k)≤DM ∀ k)≥1−DMe−D300M≥1−e−D400M.Prsubscript𝐶𝑖𝑘𝐷𝑀 for-all 𝑘1𝐷𝑀superscript𝑒𝐷300𝑀1superscript𝑒𝐷400𝑀\displaystyle\Pr\left(C_{(i,k)}\leq DM\hbox{ }\forall\hbox{ }k\right)\geq 1-\frac{D}{M}e^{-\frac{D}{300M}}\geq 1-e^{-\frac{D}{400M}}.</td>
<td></td>
<td>(32)</td>
</tr>
</tbody>
</table>
Let $`\mathcal{T}_{(i,k)}`$ denote the event that a B-duplicate packet is broadcast at time slot $`t`$ in the transporting step. If $`(i,k)`$ is successfully duplicated, i.e., there are at least $`{4\hspace{0pt}M}/5`$ B-duplicate copies of $`{(i,k)},`$ we have
<table>
<tbody>
<tr>
<td></td>
<td>Pr(𝒯(i,k)(t))≥4M5Mn(1−Mn)C(i,k)+9M10−1.Prsubscript𝒯𝑖𝑘𝑡4𝑀5𝑀𝑛superscript1𝑀𝑛subscript𝐶𝑖𝑘9𝑀101\Pr\left({\cal T}_{(i,k)}(t)\right)\geq\frac{4M}{5}\sqrt{\frac{M}{n}}\left(1-\sqrt{\frac{M}{n}}\right)^{C_{(i,k)}+\frac{9M}{10}-1}.</td>
<td></td>
</tr>
</tbody>
</table>
Further, let $`\mathcal{T}_{(i,k)}`$ denote the event that at least one copy of $`(i,k)`$ is broadcast in the transporting step. Then for sufficiently large $`n,`$ we can obtain that
<table>
<tbody>
<tr>
<td></td>
<td></td>
<td></td>
<td>Pr(𝒯(i,k)|C(i,k)≤DM)Prconditionalsubscript𝒯𝑖𝑘subscript𝐶𝑖𝑘𝐷𝑀\displaystyle\Pr\left({{\cal T}_{(i,k)}}\left|C_{(i,k)}\leq DM\right.\right)</td>
<td></td>
</tr>
<tr>
<td></td>
<td></td>
<td>≥\displaystyle\geq</td>
<td>1−(1−4M5Mn(1−Mn)DM+9M/10−1)D1superscript14𝑀5𝑀𝑛superscript1𝑀𝑛𝐷𝑀9𝑀101𝐷\displaystyle 1-\left(1-\frac{4M}{5}\sqrt{\frac{M}{n}}\left(1-\sqrt{\frac{M}{n}}\right)^{DM+9M/10-1}\right)^{D}</td>
<td></td>
</tr>
<tr>
<td></td>
<td></td>
<td>≥\displaystyle\geq</td>
<td>14.14\displaystyle\frac{1}{4}.</td>
<td></td>
</tr>
</tbody>
</table>
Let $`C_{i}^{b}`$ denote the number of distinct coded packets of node $`i`$ broadcast in the transporting step, i.e.,
<table>
<tbody>
<tr>
<td></td>
<td>Cib=∑k=1D/M1𝒯(i,k).subscriptsuperscript𝐶𝑏𝑖superscriptsubscript𝑘1𝐷𝑀subscript1subscript𝒯𝑖𝑘C^{b}_{i}=\sum_{k=1}^{D/M}1_{{\cal T}_{(i,k)}}.</td>
<td></td>
</tr>
</tbody>
</table>
Since different coded packets of node $`i`$ are broadcast in different V-rectangles, $`\{\mathcal{T}_{(i,k)}\}`$ are mutually independent. From the Chernoff bound, we have
<table>
<tbody>
<tr>
<td></td>
<td>Pr(Cib≥D7M|Ai≥2D3M,C(i,k)≤MD ∀k)≥1−2e−D900M.Prsubscriptsuperscript𝐶𝑏𝑖conditional𝐷7𝑀subscript𝐴𝑖2𝐷3𝑀subscript𝐶𝑖𝑘𝑀𝐷 for-all𝑘12superscript𝑒𝐷900𝑀\displaystyle\Pr\left(\left.C^{b}_{i}\geq\frac{D}{7M}\right|A_{i}\geq\frac{2D}{3M},C_{(i,k)}\leq MD\hbox{ }\forall k\right)\geq 1-2e^{-\frac{D}{900M}}.</td>
<td></td>
<td>(33)</td>
</tr>
</tbody>
</table>
In the transporting step, a T-duplicate copy will be dropped if the node carrying it obtains another packet destined to the same destination. Consider a T-duplicate packet $`(i,k,l)`$ carried by node $`l.`$ Note following facts:
1. \(a\)
Coded packets of node $`i`$ are broadcast in at most $`D/M`$ V-rectangles.
2. \(b\)
Each rectangle contains at most $`{9\hspace{0pt}M}/10`$ B-duplicate copies from node $`i.`$
Thus, the probability of $`(i,k,l)`$ dropped at time slot $`t`$ is at most
<table>
<tbody>
<tr>
<td></td>
<td>DMMn(1−Mn)9M10𝐷𝑀𝑀𝑛superscript1𝑀𝑛9𝑀10\frac{D}{M}\sqrt{\frac{M}{n}}\left(1-\sqrt{\frac{M}{n}}\right)^{\frac{9M}{10}}</td>
<td></td>
</tr>
</tbody>
</table>
The node mobility is independent across time, so the probability of $`(i,k,l)`$ dropped in the transporting step is at most
<table>
<tbody>
<tr>
<td></td>
<td>1−(1−DMMn(1−Mn)9M10)D≤1M2.1superscript1𝐷𝑀𝑀𝑛superscript1𝑀𝑛9𝑀10𝐷1superscript𝑀21-\left(1-\frac{D}{M}\sqrt{\frac{M}{n}}\left(1-\sqrt{\frac{M}{n}}\right)^{\frac{9M}{10}}\right)^{D}\leq\frac{1}{M^{2}}.</td>
<td></td>
</tr>
</tbody>
</table>
Let $`{\overline{N}}_{i}^{d}`$ denote the number of duplicate packets dropped in the transporting step. Note that $`{9\hspace{0pt}M\hspace{0pt}C_{i}^{b}}/10`$ T-duplicate packets are generated, and each of them has probability $`1/M^{2}`$ to be dropped. Using the Markov inequality, we have
<table>
<tbody>
<tr>
<td></td>
<td>Pr(N¯id≥MCib100)≤90M2,Prsuperscriptsubscript¯𝑁𝑖𝑑𝑀superscriptsubscript𝐶𝑖𝑏10090superscript𝑀2\Pr\left(\bar{N}_{i}^{d}\geq\frac{MC_{i}^{b}}{100}\right)\leq\frac{90}{M^{2}},</td>
<td></td>
</tr>
</tbody>
</table>
which implies
<table>
<tbody>
<tr>
<td></td>
<td>Pr(Ci≥910Cib)≥1−90M2Prsubscript𝐶𝑖910superscriptsubscript𝐶𝑖𝑏190superscript𝑀2\displaystyle\Pr\left(C_{i}\geq\frac{9}{10}C_{i}^{b}\right)\geq 1-\frac{90}{M^{2}}</td>
<td></td>
<td>(34)</td>
</tr>
</tbody>
</table>
since otherwise, more than $`{M\hspace{0pt}C_{i}^{b}}/100`$ duplicate copies are dropped. Inequality ([27](#S5.E27 "In Proof: ‣ V-B Joint Coding-Scheduling Algorithm for Fast Mobility ‣ V One-Dimensional I.I.D. Mobility Models ‣ Optimal Delay-Throughput Trade-offs in Mobile Ad-Hoc Networks: Hybrid Random Walk and One-Dimensional Mobility ModelsAn earlier version of this paper appeared in the Proc. of ITA Workshop, 2007.Research presented here was supported in part by a Vodafone Fellowship and NSF grant CNS 05-19691.")) follows form inequality ([32](#S5.E32 "In Proof: ‣ V-B Joint Coding-Scheduling Algorithm for Fast Mobility ‣ V One-Dimensional I.I.D. Mobility Models ‣ Optimal Delay-Throughput Trade-offs in Mobile Ad-Hoc Networks: Hybrid Random Walk and One-Dimensional Mobility ModelsAn earlier version of this paper appeared in the Proc. of ITA Workshop, 2007.Research presented here was supported in part by a Vodafone Fellowship and NSF grant CNS 05-19691."))-([34](#S5.E34 "In Proof: ‣ V-B Joint Coding-Scheduling Algorithm for Fast Mobility ‣ V One-Dimensional I.I.D. Mobility Models ‣ Optimal Delay-Throughput Trade-offs in Mobile Ad-Hoc Networks: Hybrid Random Walk and One-Dimensional Mobility ModelsAn earlier version of this paper appeared in the Proc. of ITA Workshop, 2007.Research presented here was supported in part by a Vodafone Fellowship and NSF grant CNS 05-19691.")).
Analysis of receiving step: The proof is similar to the proof of inequality (13) in \[[23](#bib.bib23)\]. ∎
###
V-C
Joint Coding-Scheduling Algorithm for Slow Mobility
In this subsection, we propose an algorithm which achieves the delay-throughput trade-off obtained in Theorem [3](#Thmthm3 "Theorem 3 ‣ V-A Upper Bounds ‣ V One-Dimensional I.I.D. Mobility Models ‣ Optimal Delay-Throughput Trade-offs in Mobile Ad-Hoc Networks: Hybrid Random Walk and One-Dimensional Mobility ModelsAn earlier version of this paper appeared in the Proc. of ITA Workshop, 2007.Research presented here was supported in part by a Vodafone Fellowship and NSF grant CNS 05-19691."). First choose
<table>
<tbody>
<tr>
<td></td>
<td>M1subscript𝑀1\displaystyle M_{1}</td>
<td>=\displaystyle=</td>
<td>n4D244𝑛4superscript𝐷2\displaystyle\sqrt[4]{\frac{n}{4D^{2}}}</td>
<td></td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td>M2subscript𝑀2\displaystyle M_{2}</td>
<td>=\displaystyle=</td>
<td>M12superscriptsubscript𝑀12\displaystyle M_{1}^{2}</td>
<td></td>
</tr>
</tbody>
</table>
and scale the packet size to be
<table>
<tbody>
<tr>
<td></td>
<td>W4csM1𝑊4subscript𝑐𝑠subscript𝑀1\frac{W}{4c_{s}M_{1}}</td>
<td></td>
</tr>
</tbody>
</table>
where $`c_{s}`$ is a constant independent of $`n`$ as in \[[23](#bib.bib23)\]. Further, we divide the unit square into $`\sqrt{n/M_{2}}`$ horizontal rectangles, named as H-rectangles; and $`\sqrt{n/M_{2}}`$ vertical rectangles, named V-rectangles.
Joint Coding-Scheduling Algorithm IV: We group every $`14\hspace{0pt}D`$ time slots into a super time slot. At each super time slot, the packets are coded and transmitted as follows:
1. \(1\)
Raptor Encoding: Each source takes $`D/50`$ data packets, and uses the Raptor codes to generate $`D`$ coded packets.
2. \(2\)
Broadcasting: The unit square is divided into a regular lattice with $`n/M_{1}`$ cells. This step consists of $`D`$ time slots. At each time slot, the nodes execute the following tasks:
1. \(i\)
The nodes in good cells take their turns to broadcast. If node $`i`$ is a H(V)-node and has never been in the current cell before, it randomly selects $`{9\hspace{0pt}M_{1}}/10`$ V(H)-nodes and broadcasts a coded packet to them.
2. \(ii\)
Each node checks the B-duplicate packets it carries. If there are multiple B-duplicate packets destined to the same rectangle, randomly pick one and drop the others.
3. \(3\)
Transporting: The unit square is divided into a regular lattice with $`n/M_{1}`$ cells. This step consists of $`2\hspace{0pt}D`$ time slots. At each time slot, the nodes do the following:
1. \(i\)
Suppose node $`j`$ carries duplicate packet $`{(i,k,j)},`$ which is an H-packet. If node $`j`$ is in a good cell and and $`(i,k,j)`$ is deliverable, node $`j`$ broadcasts the packet to $`{9\hspace{0pt}M_{1}}/10`$ H-nodes in the cell.
2. \(ii\)
Each node checks the T-duplicate packets it carries. if there is more than one T-duplicate packet destined to the same destination, randomly pick one and drop the others.
4. \(4\)
Receiving: The unit square is divided into a regular lattice with $`n/M_{2}`$ cells. This step consists of $`12\hspace{0pt}D`$ time slots. At each time slot, the nodes in good cells do the following at the mini-time slot allocated to their cells:
1. \(i\)
The nodes which contain deliverable packets randomly pick one deliverable packet and send a request to the corresponding destination.
2. \(ii\)
For each destination, it accepts only one request.
3. \(iii\)
The nodes whose requests are accepted transmit the deliverable packets to their destinations using the highway algorithm proposed in \[[4](#bib.bib4)\].
At the end of this step, all undelivered duplicate packets are dropped. Destinations use Raptor decoding to decode the received coded packets.
######
Theorem 5
Consider Joint Coding-Scheduling Algorithm IV. Suppose $`D`$ is both $`\omega\hspace{0pt}{(1)}`$ and $`{o\hspace{0pt}\left( {\sqrt{n}/{\log^{2}n}} \right)},`$ and the delay constraint is $`{14\hspace{0pt}D}.`$ Then given any $`{\epsilon > 0},`$ there exists $`n_{0}`$ such that for any $`{n \geq n_{0}},`$ every data packet sent out can be recovered at the destination with probability $`{1 - \epsilon},`$ and furthermore
<table>
<tbody>
<tr>
<td></td>
<td>limT→∞Pr(Λi[T]T≥(W14002csC)D2n4 ∀i)=1.subscript→𝑇PrsubscriptΛ𝑖delimited-[]𝑇𝑇𝑊14002subscript𝑐𝑠𝐶4superscript𝐷2𝑛 for-all𝑖1\displaystyle\lim_{T\rightarrow\infty}\Pr\left(\frac{\Lambda_{i}[T]}{T}\geq\left(\frac{W}{1400\sqrt{2}c_{s}C}\right)\sqrt[4]{\frac{D^{2}}{n}}\hbox{ }\forall i\right)=1.</td>
<td></td>
</tr>
</tbody>
</table>
###### Proof:
Following the analysis of Theorem [4](#Sx2.EGx30 "Theorem 4 ‣ V-B Joint Coding-Scheduling Algorithm for Fast Mobility ‣ V One-Dimensional I.I.D. Mobility Models ‣ Optimal Delay-Throughput Trade-offs in Mobile Ad-Hoc Networks: Hybrid Random Walk and One-Dimensional Mobility ModelsAn earlier version of this paper appeared in the Proc. of ITA Workshop, 2007.Research presented here was supported in part by a Vodafone Fellowship and NSF grant CNS 05-19691."), we can show the following events happen with high probability.
Node distribution: All cells are good during the entire super-time-slot with high probability, i.e.,
<table>
<tbody>
<tr>
<td></td>
<td>Pr(𝒢)≥1−1n2.Pr𝒢11superscript𝑛2\displaystyle\Pr\left({\cal G}\right)\geq 1-\frac{1}{n^{2}}.</td>
<td></td>
<td>(35)</td>
</tr>
</tbody>
</table>
Broadcasting: At least $`{3\hspace{0pt}D}/10`$ coded packets from a source are successfully duplicated after the broadcasting step with high probability, where a coded packet is said to be successfully duplicated if it has at least $`M_{1}/3`$ B-duplicate packets. Specifically, we have
<table>
<tbody>
<tr>
<td></td>
<td>Pr(Ai≥310D|𝒢)≥1−1n2.Prsubscript𝐴𝑖conditional310𝐷𝒢11superscript𝑛2\displaystyle\Pr\left(\left.A_{i}\geq\frac{3}{10}D\right|{\cal G}\right)\geq 1-\frac{1}{n^{2}}.</td>
<td></td>
<td>(36)</td>
</tr>
</tbody>
</table>
Transporting: At least $`{3\hspace{0pt}D}/40`$ coded packets from a source are successfully transported after the transporting step with high probability, where a coded packet is said to be successfully transported if it has at least $`{4\hspace{0pt}M_{1}}/5`$ T-duplicate copies. Specifically, we have
<table>
<tbody>
<tr>
<td></td>
<td>Pr(Ci≥340D|𝒢,Ai≥310D)≥1−e−D3600−180logn.Prsubscript𝐶𝑖conditional340𝐷𝒢subscript𝐴𝑖310𝐷1superscript𝑒𝐷3600180𝑛\displaystyle\Pr\left(\left.C_{i}\geq\frac{3}{40}D\right|{\cal G},A_{i}\geq\frac{3}{10}D\right)\geq 1-e^{-\frac{D}{3600}}-\frac{180}{\log n}.</td>
<td></td>
<td>(37)</td>
</tr>
</tbody>
</table>
Receiving: At least $`D/40`$ distinct coded packets from a source are delivered to its destination after the receiving step. Specifically, we have
<table>
<tbody>
<tr>
<td></td>
<td>Pr(Bi≥D40|𝒢,Ci≥340D)≥1−2e−D1000.Prsubscript𝐵𝑖conditional𝐷40𝒢subscript𝐶𝑖340𝐷12superscript𝑒𝐷1000\displaystyle\Pr\left(\left.B_{i}\geq\frac{D}{40}\right|{\cal G},C_{i}\geq\frac{3}{40}D\right)\geq 1-2e^{-\frac{D}{1000}}.</td>
<td></td>
<td>(38)</td>
</tr>
</tbody>
</table>
Thus, the probability that the $`D/50`$ data packets are fully recovered in one super time slot is at least
<table>
<tbody>
<tr>
<td></td>
<td>1−2n2−e−D3600−180logn−2e−D1000,12superscript𝑛2superscript𝑒𝐷3600180𝑛2superscript𝑒𝐷10001-\frac{2}{n^{2}}-e^{-\frac{D}{3600}}-\frac{180}{\log n}-2e^{-\frac{D}{1000}},</td>
<td></td>
</tr>
</tbody>
</table>
and theorem holds. ∎
## VI One-Dimensional Hybrid Random Walk Model, Fast Mobiles and Slow Mobiles
In this section, we present the optimal delay-throughput trade-offs of the one-dimensional hybrid random walk model. The results can be proved following the analysis of the one-dimensional i.i.d. mobility and the analysis of the two-dimensional hybrid random walk. The details are omitted here for brevity.
######
Theorem 6
Consider the one-dimensional hybrid random walk model and assume that Assumption 1-3 hold. Then for $`S = {o\hspace{0pt}{(1)}}`$ and $`{D = {\omega\hspace{0pt}{({1/S^{2}})}}},`$ We have following results:
1. \(1\)
For fast mobiles,
<table>
<tbody>
<tr>
<td></td>
<td>24WT1πΔ23n3/2(D3/2+1)≥E[Λ[T]].24𝑊𝑇31𝜋superscriptΔ232𝑛32𝐷1𝐸delimited-[]Λdelimited-[]𝑇\displaystyle 24WT\sqrt[3]{\frac{1}{\pi\Delta^{2}}}\sqrt[3/2]{n}(\sqrt[3/2]{D}+1)\geq E[\Lambda[T]].</td>
<td></td>
<td>(39)</td>
</tr>
</tbody>
</table>
When $`S = {o\hspace{0pt}{(1)}}`$ and $`D`$ is both $`\omega\hspace{0pt}{({\max{\{{{{({\log^{2}n})}\hspace{0pt}{|{\log S}|}}/S^{4}},{\sqrt[5]{n}\hspace{0pt}{\log n}}\}}})}`$ and $`{o\hspace{0pt}{({\sqrt{n}/\sqrt[{3/2}]{\log n}})}},`$ Joint Coding-Scheduling Algorithm III can be used to achieve a throughput same as ([39](#S6.E39 "In item (1) ‣ Theorem 6 ‣ VI One-Dimensional Hybrid Random Walk Model, Fast Mobiles and Slow Mobiles ‣ Optimal Delay-Throughput Trade-offs in Mobile Ad-Hoc Networks: Hybrid Random Walk and One-Dimensional Mobility ModelsAn earlier version of this paper appeared in the Proc. of ITA Workshop, 2007.Research presented here was supported in part by a Vodafone Fellowship and NSF grant CNS 05-19691.")) except for a constant factor.
2. \(2\)
For slow mobiles,
<table>
<tbody>
<tr>
<td></td>
<td>12WT1πΔ24n3/4(D+1)≥E[Λ[T]].12𝑊𝑇41𝜋superscriptΔ234𝑛𝐷1𝐸delimited-[]Λdelimited-[]𝑇\displaystyle 12WT\sqrt[4]{\frac{1}{\pi\Delta^{2}}}\sqrt[3/4]{n}(\sqrt{D}+1)\geq E[\Lambda[T]].</td>
<td></td>
<td>(40)</td>
</tr>
</tbody>
</table>
When $`S = {o\hspace{0pt}{(1)}}`$ and $`D`$ is both $`\omega\hspace{0pt}{({{{({\log^{2}n})}\hspace{0pt}{|{\log S}|}}/S^{4}})}`$ and $`{o\hspace{0pt}{({\sqrt{n}/{\log^{2}n}})}},`$ Joint Coding-Scheduling Algorithm IV can be used to achieve a throughput same as ([40](#S6.E40 "In item (2) ‣ Theorem 6 ‣ VI One-Dimensional Hybrid Random Walk Model, Fast Mobiles and Slow Mobiles ‣ Optimal Delay-Throughput Trade-offs in Mobile Ad-Hoc Networks: Hybrid Random Walk and One-Dimensional Mobility ModelsAn earlier version of this paper appeared in the Proc. of ITA Workshop, 2007.Research presented here was supported in part by a Vodafone Fellowship and NSF grant CNS 05-19691.")) except for a constant factor.
$`\square`$
## VII Conclusion
In this paper, we investigated the optimal delay-throughput trade-off of a mobile ad-hoc network under the two-dimensional hybrid random walk, one-dimensional i.i.d. mobility model and one-dimensional hybrid random walk model. The optimal trade-offs have been established under some conditions on delay $`D.`$ When these conditions are not met, the optimal trade-offs are still unknown in general. We would like to comment that the key to establishing the optimal delay-throughput trade-off is to obtain $`{P_{i,j}\hspace{0pt}{(D,L)}},`$ the probability that node $`i`$ hits node $`j`$ in one of $`D`$ consecutive time slots given a hitting distance $`L.`$ For example, under the two-dimensional hybrid random walk model, the upper bound was obtained under the condition $`D = {\omega\hspace{0pt}{({{|{\log S}|}/S^{2}})}}`$ since it was the condition under which we established an upper bound on $`P_{i,j}\hspace{0pt}{(D,L)}`$ (inequality ([7](#S4.E7 "In Proof: ‣ IV-A Upper Bound ‣ IV Two-Dimensional Hybrid Random Walk Models ‣ Optimal Delay-Throughput Trade-offs in Mobile Ad-Hoc Networks: Hybrid Random Walk and One-Dimensional Mobility ModelsAn earlier version of this paper appeared in the Proc. of ITA Workshop, 2007.Research presented here was supported in part by a Vodafone Fellowship and NSF grant CNS 05-19691."))). Further, the maximum throughput was shown to be achievable under a more restrict condition $`D = {\omega\hspace{0pt}{({{{({\log^{2}n})}\hspace{0pt}{|{\log S}|}}/S^{6}})}}`$ since it was the condition under which we established a lower bound on $`P_{i,j}\hspace{0pt}{(D,L)}`$ (inequality ([17](#S4.E17 "In Proof: ‣ IV-B Joint Coding-Scheduling Algorithms ‣ IV Two-Dimensional Hybrid Random Walk Models ‣ Optimal Delay-Throughput Trade-offs in Mobile Ad-Hoc Networks: Hybrid Random Walk and One-Dimensional Mobility ModelsAn earlier version of this paper appeared in the Proc. of ITA Workshop, 2007.Research presented here was supported in part by a Vodafone Fellowship and NSF grant CNS 05-19691."))). Thus, if we can find techniques to compute $`P_{i,j}\hspace{0pt}{(D,L)}`$ without using the restricts on $`D,`$ then the delay-throughput trade-offs can be characterized more generally. This is a topic for future research.
Acknowledgment: We thank Sichao Yang for useful discussions during the course of this work.
## References
- \[1\] S. Aeron and V. Saligrama. Wireless ad hoc networks: strategies and scaling laws for the fixed SNR regime. Preprint, 2006.
- \[2\] A. Agarwal and P. R. Kumar. Improved Capacity Bounds for Wireless Networks. In *Wireless Communications and Mobile Computing,* 4:251-261, 2004.
- \[3\] S. N. Diggavi, M. Grossglauser and D. Tse. Even one-dimensional mobility increases ad hoc wireless capacity. In *Proc. of ISIT,* Lausanne, Switzerland, July 2002.
- \[4\] M. Franceschetti, O. Dousse, D. Tse, and P. Thiran. On the throughput capacity of random wireless networks. Preprint, available at http://fleece.ucsd.edu/~massimo/.
- \[5\] A. El Gammal, J. Mammen, B. Prabhakar, and D. Shah. Throughput-delay trade-off in wireless networks. In *Proc. of IEEE INFOCOM,* 2004.
- \[6\] A. El Gammal, J. Mammen, B. Prabhakar, and D. Shah. Optimal throughput-delay scaling in wireless networks - Part I: the fluid model. In *IEEE Transactions on Information Theory,* vol. 52, no. 6, pp. 2568-2592.
- \[7\] A. El Gammal, J. Mammen, B. Prabhakar, and D. Shah. Optimal throughput-delay scaling in wireless networks - Part II: constant-size packets. In *IEEE Transactions on Information Theory,* vol. 52, no. 11, pp. 5111 - 5116.
- \[8\] J. Mammen and D. Shah. Throughput and delay in random wireless networks with restricted mobility. Preprint, available at http://www.stanford.edu/~jmammen/.
- \[9\] M. Grossglauser and D. Tse. Mobility increases the capacity of ad-hoc wireless networks. In *Proc. of IEEE INFOCOM,* 2001, pp. 1360-1369.
- \[10\] P. Gupta and P. Kumar. The capacity of wireless networks. In *IEEE Transactions on Information Theory,* vol. 46, no.2, pp. 388-404, 2000.
- \[11\] L. Lovasz. Random walks on graphs: a survey. Combinatorics, Paul Erdos in Eighty, Vol.2 Janos Bolyai Mathematical Society, Budapest, 1996, 353-398.
- \[12\] X. Lin and N. B. Shroff. The fundamental capacity-delay tradeoff in large mobile ad hoc networks. In *Third Annual Mediterranean Ad Hoc Networking Workshop,* 2004.
- \[13\] X. Lin, G. Sharma, R. R. Mazumdar and N. B. Shroff. Degenerate delay-capacity trade-offs in ad hoc networks with Brownian mobility. In *Joint Special Issue of IEEE Transactions on Information Theory and IEEE/ACM Transactions on Networking on Networking and Information Theory,* vol. 52, no. 6, pp 2777-2784, June 2006.
- \[14\] M. Luby. LT Codes. In *Proc. The 43rd Annual IEEE Symposium on Foundations of Computer Science,* pp. 271-282, November 2002.
- \[15\] M. Mitzenmacher and E. Upfal.
Probability and Computing: Ranomized Algorithms and Probabilistic Analysis.
Cambridge, 2005.
- \[16\] M.J. Neely and E. Modiano. Capacity and delay tradeoffs for ad-hoc mobile networks. In *IEEE Transactions on Information Theory,* vol. 51, No. 6, June 2005.
- \[17\] A. Ozgur, O. Leveque, and D. Tse. How does the information capacity of ad hoc networks scale? In *Proc. of Allerton Conference 2006.*
- \[18\] D. Shah and S. Shakkottai. Oblivious routing with mobile fusion centers over a sensor network. Preprint.
- \[19\] G. Sharma, R. Mazumdar, and N. Shroff. Delay and capacity trade-offs in mobile ad hoc networks: a global perspective. In *Proc. of IEEE Infocom,* 2006.
- \[20\] A. Shokrollahi. Raptor Codes. In *IEEE Transactions on Information Theory,* vol. 52, No. 6, 2006.
- \[21\] S. Toumpis and A. Goldsmith. Large wireless networks under fading, mobility, and delay constraints. In *Proc. of IEEE INFOCOM,* 2004.
- \[22\] F. Warthman. Delay-tolerant networks. Technical report, available at http://www.ipnsig.org/reports/DTN_Tutorial11.pdf
- \[23\] L. Ying, S. Yang and R. Srikant. Coding improves the optimal delay-throughput trade-offs in mobile ad-hoc networks: two-dimensional i.i.d. mobility models. Preprint.
## Appendix A: Probability Results
The following lemmas are some basic probability results.
######
Lemma 7
Let $`X_{1},\ldots,X_{n}`$ be independent $`0 - 1`$ random variables such that $`{{\sum_{i}X_{i}} = \mu}.`$ Then, the following Chernoff bounds hold
<table>
<tbody>
<tr>
<td></td>
<td>Pr(∑i=1nXi<(1−δ)μ)Prsuperscriptsubscript𝑖1𝑛subscript𝑋𝑖1𝛿𝜇\displaystyle\hbox{Pr}\left(\sum_{i=1}^{n}X_{i}<(1-\delta)\mu\right)</td>
<td>≤\displaystyle\leq</td>
<td>e−δ2μ/2;superscript𝑒superscript𝛿2𝜇2\displaystyle e^{-\delta^{2}\mu/2};</td>
<td></td>
<td>(41)</td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td>Pr(∑i=1nXi>(1+δ)μ)Prsuperscriptsubscript𝑖1𝑛subscript𝑋𝑖1𝛿𝜇\displaystyle\hbox{Pr}\left(\sum_{i=1}^{n}X_{i}>(1+\delta)\mu\right)</td>
<td>≤\displaystyle\leq</td>
<td>e−δ2μ/3.superscript𝑒superscript𝛿2𝜇3\displaystyle e^{-\delta^{2}\mu/3}.</td>
<td></td>
<td>(42)</td>
</tr>
</tbody>
</table>
###### Proof:
A detailed proof can be found in \[[15](#bib.bib15)\]. ∎
######
Lemma 8
Assume we have $`m`$ bins. At each time, choose $`h`$ bins and drop one ball in each of them. Repeat this $`n`$ times. Using $`N_{1}`$ to denote the number of bins containing at least one ball, the following inequality holds for sufficiently large $`n.`$
<table>
<tbody>
<tr>
<td></td>
<td>Pr(N1≤(1−δ)mp~1)Prsubscript𝑁11𝛿𝑚subscript~𝑝1\displaystyle\hbox{Pr}\left(N_{1}\leq(1-\delta)m\tilde{p}_{1}\right)</td>
<td>≤\displaystyle\leq</td>
<td>2e−δ2mp~1/3.2superscript𝑒superscript𝛿2𝑚subscript~𝑝13\displaystyle 2e^{-\delta^{2}m\tilde{p}_{1}/3}.</td>
<td></td>
<td>(43)</td>
</tr>
</tbody>
</table>
where $`{{\overset{\sim}{p}}_{1} = {1 - e^{- \frac{n\hspace{0pt}h}{m}}}}.`$
###### Proof:
Please refer to \[[23](#bib.bib23)\] for a detailed proof. ∎
######
Lemma 9
Suppose $`n`$ balls are independently dropped into $`m`$ bins and one trash can. After a ball is dropped, the probability in the trash can is $`{1 - p},`$ and the probability in a specific bin is $`{p/m}.`$ Using $`N_{2}`$ to denote the number of bins containing at least $`1`$ ball, the following inequality holds for sufficiently large $`n.`$
<table>
<tbody>
<tr>
<td></td>
<td>Pr(N2≤(1−δ)mp~2)Prsubscript𝑁21𝛿𝑚subscript~𝑝2\displaystyle\hbox{Pr}\left(N_{2}\leq(1-\delta)m\tilde{p}_{2}\right)</td>
<td>≤\displaystyle\leq</td>
<td>2e−δ2mp~2/3;2superscript𝑒superscript𝛿2𝑚subscript~𝑝23\displaystyle 2e^{-\delta^{2}m\tilde{p}_{2}/3};</td>
<td></td>
<td>(44)</td>
</tr>
</tbody>
</table>
where $`{{\overset{\sim}{p}}_{2} = {1 - e^{- \frac{n\hspace{0pt}p}{m}}}}.`$
###### Proof:
Please refer to \[[23](#bib.bib23)\] for a detailed proof. ∎
Next we introduce the Azuma-Hoeffding inequality.
######
Lemma 10
Suppose that $`X_{0},\ldots,X_{n}`$ are independent random variables, and there exists a constant $`c > 0`$ such that $`{f\hspace{0pt}{(\mathbf{X})}} = {f\hspace{0pt}{(X_{1},\ldots,X_{n})}}`$ satisfies the following condition for any $`i`$ and any set of values $`x_{1},\ldots,x_{n}`$ and $`y_{i}:`$
<table>
<tbody>
<tr>
<td></td>
<td>|f(x1,…,xi−1,xi,xi+1,…,xn)−\displaystyle\left|f(x_{1},\ldots,x_{i-1},x_{i},x_{i+1},\ldots,x_{n})-\right.</td>
<td></td>
<td></td>
<td></td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td>f(x1,…,xi−1,yi,xi+1,…,xn)|\displaystyle\left.f(x_{1},\ldots,x_{i-1},y_{i},x_{i+1},\ldots,x_{n})\right|</td>
<td>≤c.absent𝑐\displaystyle\leq c.</td>
<td></td>
<td></td>
</tr>
</tbody>
</table>
Then we have
<table>
<tbody>
<tr>
<td></td>
<td>Pr(|f(𝐗)−E[f(𝐗)]|≥δ)≤2e−2δ2nc2.Pr𝑓𝐗𝐸delimited-[]𝑓𝐗𝛿2superscript𝑒2superscript𝛿2𝑛superscript𝑐2\Pr\left(\left|f(\mathbf{X})-E[f(\mathbf{X})]\right|\geq\delta\right)\leq 2e^{-\frac{2\delta^{2}}{nc^{2}}}.</td>
<td></td>
</tr>
</tbody>
</table>
###### Proof:
A detailed proof can be found in \[[15](#bib.bib15)\]. ∎
## Appendix B: Properties of Random Walk
Consider following two random walks.
1. \(1\)
One dimensional random walk: A random walk on a circle with unit length and $`1/\overset{\sim}{S}`$ points. At each time slot, a node moves to one point left, one point right or doesn’t move with equal probability as in Figure [6](#Sx2.F6 "Figure 6 ‣ item (1) ‣ Appendix B: Properties of Random Walk ‣ Optimal Delay-Throughput Trade-offs in Mobile Ad-Hoc Networks: Hybrid Random Walk and One-Dimensional Mobility ModelsAn earlier version of this paper appeared in the Proc. of ITA Workshop, 2007.Research presented here was supported in part by a Vodafone Fellowship and NSF grant CNS 05-19691.").
[Refer to caption](/html/0705.0326/assets/x6.png)
Figure 6: One Dimensional Random Walk
2. \(2\)
Two dimensional random walk: A random walk on a unit torus with $`1/{\overset{\sim}{S}}^{2}`$ points. At each time slot, a node moves to one of eight neighbors or doesn’t move with equal probability as in Figure [7](#Sx2.F7 "Figure 7 ‣ Appendix B: Properties of Random Walk ‣ Optimal Delay-Throughput Trade-offs in Mobile Ad-Hoc Networks: Hybrid Random Walk and One-Dimensional Mobility ModelsAn earlier version of this paper appeared in the Proc. of ITA Workshop, 2007.Research presented here was supported in part by a Vodafone Fellowship and NSF grant CNS 05-19691.").
[Refer to caption](/html/0705.0326/assets/x7.png)
Figure 7: Two Dimensional Random Walk
We introduce following definitions.
- •
Transition matrix $`\mathbf{P}:`$ $`\mathbf{P} = \left\lbrack P_{\mathbf{i},\mathbf{j}} \right\rbrack`$ where $`P_{\mathbf{i},\mathbf{j}}`$ is the probability of moving from point $`\mathbf{i}`$ to point $`\mathbf{j}.`$
- •
Stationary distribution $`\Pi:`$ A vector which satisfies the equation $`{{\Pi\hspace{0pt}\mathbf{P}} = \Pi}.`$
- •
Hitting time $`{T_{h}\hspace{0pt}{(\mathbf{i},\mathbf{j})}}:`$ Time taken for a node to move from point $`\mathbf{i}`$ to point $`\mathbf{j}.`$
- •
Mixing time $`T_{m}:`$
<table>
<tbody>
<tr>
<td></td>
<td>Tm=inftsup𝐢∑𝐣|P𝐢𝐣t−Π𝐣|≤S~4,subscript𝑇𝑚subscriptinfimum𝑡subscriptsupremum𝐢subscript𝐣subscriptsuperscript𝑃𝑡𝐢𝐣subscriptΠ𝐣superscript~𝑆4T_{m}=\inf_{t}\sup_{\mathbf{i}}\sum_{\mathbf{j}}\left|P^{t}_{\mathbf{i}\mathbf{j}}-\Pi_{\mathbf{j}}\right|\leq{\tilde{S}}^{4},</td>
<td></td>
</tr>
</tbody>
</table>
where $`P_{\mathbf{i}\mathbf{j}}^{t}`$ is the $`{(\mathbf{i},\mathbf{j})}^{th}`$ entry of $`\mathbf{P}^{t}.`$
######
Lemma 11
For the one dimensional random walk, we have
- •
$`{{E\hspace{0pt}{\lbrack{T_{h}\hspace{0pt}{(\mathbf{i},\mathbf{j})}}\rbrack}} = {O\hspace{0pt}{({1/{\overset{\sim}{S}}^{2}})}}}.`$
- •
$`{T_{m} = {O\hspace{0pt}{({{|{\log\overset{\sim}{S}}|}/{\overset{\sim}{S}}^{2}})}}}.`$
For the two dimensional random walk, we have
- •
$`{{E\hspace{0pt}{\lbrack{T_{h}\hspace{0pt}{(\mathbf{i},\mathbf{j})}}\rbrack}} = {O\hspace{0pt}{({{|{\log\overset{\sim}{S}}|}/{\overset{\sim}{S}}^{2}})}}}.`$
- •
$`{T_{m} = {O\hspace{0pt}{({{|{\log\overset{\sim}{S}}|}/{\overset{\sim}{S}}^{2}})}}}.`$
###### Proof:
Please refer to \[[11](#bib.bib11)\] for the one dimensional random walk. The proof of the hitting time of the two dimensional random walk is presented Lemma 13 in \[[6](#bib.bib6)\], and the mixing time result holds since the two dimensional random walk can be regarded as two independent one dimensional random walks. ∎
######
Lemma 12
Let $`N_{\mathbf{i},\mathbf{j},\mathbf{k}}\hspace{0pt}{\lbrack D\rbrack}`$ denote the number of visits to point $`\mathbf{j}`$ in $`D`$ time slots starting from point $`\mathbf{i}`$ and ending at point $`\mathbf{k}.`$ If $`{D = {\omega\hspace{0pt}{({{|{\log\hat{S}}|}/\hat{S}})}}},`$ we have
<table>
<tbody>
<tr>
<td></td>
<td>910DS^≤E[N𝐢,𝐣,𝐤[D]]≤1110DS^910𝐷^𝑆𝐸delimited-[]subscript𝑁𝐢𝐣𝐤delimited-[]𝐷1110𝐷^𝑆\displaystyle\frac{9}{10}D\hat{S}\leq E[N_{\mathbf{i},\mathbf{j},\mathbf{k}}[D]]\leq\frac{11}{10}D\hat{S}</td>
<td></td>
<td>(45)</td>
</tr>
</tbody>
</table>
where $`\hat{S} = \overset{\sim}{S}`$ for the one dimensional random walk and $`\hat{S} = {\overset{\sim}{S}}^{2}`$ for the two dimensional random walk. Furthermore, if $`D = {{\kappa\hspace{0pt}\alpha\hspace{0pt}{|{\log\hat{S}}|}}/{\hat{S}}^{3}}`$ where $`{\kappa = {\omega\hspace{0pt}{(1)}}},`$ then we have
<table>
<tbody>
<tr>
<td></td>
<td>Pr(65DS^≥N𝐢,𝐣,𝐤[D]]≥45DS^)≥1−2e−α625.\displaystyle\Pr\left(\frac{6}{5}D\hat{S}\geq N_{\mathbf{i},\mathbf{j},\mathbf{k}}[D]]\geq\frac{4}{5}D\hat{S}\right)\geq 1-2e^{-\frac{\alpha}{625}}.</td>
<td></td>
<td>(46)</td>
</tr>
</tbody>
</table>
###### Proof:
First we have
<table>
<tbody>
<tr>
<td></td>
<td>Th(𝐢,𝐣)+∑l=1N𝐢,𝐣,𝐤[D]−1Thl(𝐣,𝐣)+Th(𝐣,𝐤)=D,subscript𝑇ℎ𝐢𝐣superscriptsubscript𝑙1subscript𝑁𝐢𝐣𝐤delimited-[]𝐷1subscriptsuperscript𝑇𝑙ℎ𝐣𝐣subscript𝑇ℎ𝐣𝐤𝐷T_{h}(\mathbf{i},\mathbf{j})+\sum_{l=1}^{N_{\mathbf{i},\mathbf{j},\mathbf{k}}[D]-1}T^{l}_{h}(\mathbf{j},\mathbf{j})+T_{h}(\mathbf{j},\mathbf{k})=D,</td>
<td></td>
</tr>
</tbody>
</table>
where $`T_{h}^{l}\hspace{0pt}{(\mathbf{j},\mathbf{j})}`$ is the time duration between $`l^{th}`$ visits to point $`\mathbf{j}`$ and $`{({l + 1})}^{th}`$ visits to point $`\mathbf{j}.`$ Taking the expectation on both sides, we have
<table>
<tbody>
<tr>
<td></td>
<td>E[Th(𝐢,𝐣)]+E[N𝐢,𝐣,𝐤[D]]E[Th(𝐣,𝐣)]−E[Th(𝐣,𝐣)]+E[Th(𝐣,𝐤)]=D,𝐸delimited-[]subscript𝑇ℎ𝐢𝐣𝐸delimited-[]subscript𝑁𝐢𝐣𝐤delimited-[]𝐷𝐸delimited-[]subscript𝑇ℎ𝐣𝐣𝐸delimited-[]subscript𝑇ℎ𝐣𝐣𝐸delimited-[]subscript𝑇ℎ𝐣𝐤𝐷E[T_{h}(\mathbf{i},\mathbf{j})]+E[N_{\mathbf{i},\mathbf{j},\mathbf{k}}[D]]E[T_{h}(\mathbf{j},\mathbf{j})]-E[T_{h}(\mathbf{j},\mathbf{j})]+E[T_{h}(\mathbf{j},\mathbf{k})]=D,</td>
<td></td>
</tr>
</tbody>
</table>
which implies
<table>
<tbody>
<tr>
<td></td>
<td>E[N𝐢,𝐣,𝐤[D]]=D−E[Th(𝐢,𝐣)]−E[Th(𝐣,𝐤)]+E[Th(𝐣,𝐣)]E[Th(𝐣,𝐣)].𝐸delimited-[]subscript𝑁𝐢𝐣𝐤delimited-[]𝐷𝐷𝐸delimited-[]subscript𝑇ℎ𝐢𝐣𝐸delimited-[]subscript𝑇ℎ𝐣𝐤𝐸delimited-[]subscript𝑇ℎ𝐣𝐣𝐸delimited-[]subscript𝑇ℎ𝐣𝐣E[N_{\mathbf{i},\mathbf{j},\mathbf{k}}[D]]=\frac{D-E[T_{h}(\mathbf{i},\mathbf{j})]-E[T_{h}(\mathbf{j},\mathbf{k})]+E[T_{h}(\mathbf{j},\mathbf{j})]}{E[T_{h}(\mathbf{j},\mathbf{j})]}.</td>
<td></td>
</tr>
</tbody>
</table>
Inequality ([45](#Sx2.E45 "In Lemma 12 ‣ Appendix B: Properties of Random Walk ‣ Optimal Delay-Throughput Trade-offs in Mobile Ad-Hoc Networks: Hybrid Random Walk and One-Dimensional Mobility ModelsAn earlier version of this paper appeared in the Proc. of ITA Workshop, 2007.Research presented here was supported in part by a Vodafone Fellowship and NSF grant CNS 05-19691.")) follows from the facts that $`{E\hspace{0pt}{\lbrack{T_{h}\hspace{0pt}{(\mathbf{i},\mathbf{j})}}\rbrack}} = {O\hspace{0pt}{({{|{\log\hat{S}}|}/\hat{S}})}}`$ and $`{{E\hspace{0pt}{\lbrack{T_{h}\hspace{0pt}{(\mathbf{j},\mathbf{j})}}\rbrack}} = {1/\hat{S}}}.`$
Next let $`\mathbf{x}\hspace{0pt}{\lbrack t\rbrack}`$ denote the position of the node at time slot $`t,`$ $`\overset{\sim}{\mathbf{X}}`$ denote $`\{{\mathbf{x}\hspace{0pt}{\lbrack t\rbrack}}\}`$ for $`{t = {1,{\frac{D\hspace{0pt}{\hat{S}}^{2}}{\alpha} + 1},{\frac{2\hspace{0pt}D\hspace{0pt}{\hat{S}}^{2}}{\alpha} + 1},\ldots,{D - \frac{D\hspace{0pt}{\hat{S}}^{2}}{\alpha}},D}},`$ and $`\mathbf{X}_{m}`$ denote $`\{{\mathbf{x}\hspace{0pt}{\lbrack t\rbrack}}\}`$ for $`t = {{\frac{m\hspace{0pt}D\hspace{0pt}{\hat{S}}^{2}}{\alpha} + 1},\frac{m\hspace{0pt}D\hspace{0pt}{\hat{S}}^{2}}{\alpha},\ldots,{\min\left\{ D,\frac{{({m + 1})}\hspace{0pt}D\hspace{0pt}{\hat{S}}^{2}}{\alpha} \right\}}}`$ where $`{m = {0,\ldots,{{\alpha/{\hat{S}}^{2}} - 1}}}.`$ Further, let $`N_{\mathbf{j},\overset{\sim}{\mathbf{X}}}\hspace{0pt}{\lbrack D\rbrack}`$ denote the number of visits to point $`\mathbf{j}`$ given $`\overset{\sim}{\mathbf{X}}.`$ It is easy to see that for any $`\overset{\sim}{\mathbf{X}}`$ there exists a function $`f_{\overset{\sim}{\mathbf{X}}}`$ such that
<table>
<tbody>
<tr>
<td></td>
<td>N𝐣,𝐗~[D]=f𝐗~(𝐗1,…,𝐗α/S^2−1),subscript𝑁𝐣~𝐗delimited-[]𝐷subscript𝑓~𝐗subscript𝐗1…subscript𝐗𝛼superscript^𝑆21N_{\mathbf{j},\tilde{\mathbf{X}}}[D]=f_{\tilde{\mathbf{X}}}\left(\mathbf{X}_{1},\ldots,\mathbf{X}_{\alpha/{\hat{S}}^{2}-1}\right),</td>
<td></td>
</tr>
</tbody>
</table>
where $`\{\mathbf{X}_{m}\}`$ are mutually independent given $`\overset{\sim}{\mathbf{X}}.`$ Note that $`\mathbf{X}_{m}`$ contains the position information from time slot $`{{m\hspace{0pt}D\hspace{0pt}{\hat{S}}^{2}}/\alpha} + 1`$ to time slot $`{{{({m + 1})}\hspace{0pt}D\hspace{0pt}{\hat{S}}^{2}}/\alpha},`$ so
<table>
<tbody>
<tr>
<td></td>
<td>|f𝐗~(𝐗0,…,𝐗m−1,𝐗m,𝐗m+1,…,𝐗α/S^2−1)−\displaystyle\left|f_{\tilde{\mathbf{X}}}\left(\mathbf{X}_{0},\ldots,\mathbf{X}_{m-1},\mathbf{X}_{m},\mathbf{X}_{m+1},\ldots,\mathbf{X}_{\alpha/{\hat{S}}^{2}-1}\right)-\right.</td>
<td></td>
<td></td>
<td></td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td>f𝐗~(𝐗0,…,𝐗m−1,𝐘m,𝐗m+1,…,𝐗α/S^2−1)|\displaystyle\left.f_{\tilde{\mathbf{X}}}\left(\mathbf{X}_{0},\ldots,\mathbf{X}_{m-1},\mathbf{Y}_{m},\mathbf{X}_{m+1},\ldots,\mathbf{X}_{\alpha/{\hat{S}}^{2}-1}\right)\right|</td>
<td>≤DS^2α.absent𝐷superscript^𝑆2𝛼\displaystyle\leq\frac{D{\hat{S}}^{2}}{\alpha}.</td>
<td></td>
<td></td>
</tr>
</tbody>
</table>
Next note that $`{{{D\hspace{0pt}{\hat{S}}^{2}}/\alpha} = {\omega\hspace{0pt}{({{|{\log\hat{S}}|}/\hat{S}})}}},`$ so from inequality ([45](#Sx2.E45 "In Lemma 12 ‣ Appendix B: Properties of Random Walk ‣ Optimal Delay-Throughput Trade-offs in Mobile Ad-Hoc Networks: Hybrid Random Walk and One-Dimensional Mobility ModelsAn earlier version of this paper appeared in the Proc. of ITA Workshop, 2007.Research presented here was supported in part by a Vodafone Fellowship and NSF grant CNS 05-19691.")), we can conclude that for any $`\mathbf{i},`$ $`\mathbf{j}`$ and $`\mathbf{k},`$
<table>
<tbody>
<tr>
<td></td>
<td>910DS^3α≤E[N𝐢,𝐣,𝐤[DS^2α]]≤1110DS^3α,910𝐷superscript^𝑆3𝛼𝐸delimited-[]subscript𝑁𝐢𝐣𝐤delimited-[]𝐷superscript^𝑆2𝛼1110𝐷superscript^𝑆3𝛼\frac{9}{10}\frac{D{\hat{S}}^{3}}{\alpha}\leq E\left[N_{\mathbf{i},\mathbf{j},\mathbf{k}}\left[\frac{D{\hat{S}}^{2}}{\alpha}\right]\right]\leq\frac{11}{10}\frac{D{\hat{S}}^{3}}{\alpha},</td>
<td></td>
</tr>
</tbody>
</table>
which implies that
<table>
<tbody>
<tr>
<td></td>
<td>910DS^≤E[N𝐣,𝐗~[D]]≤1110DS^910𝐷^𝑆𝐸delimited-[]subscript𝑁𝐣~𝐗delimited-[]𝐷1110𝐷^𝑆\frac{9}{10}D{\hat{S}}\leq E\left[N_{\mathbf{j},\tilde{\mathbf{\mathbf{X}}}}\left[D\right]\right]\leq\frac{11}{10}D{\hat{S}}</td>
<td></td>
</tr>
</tbody>
</table>
holds for any $`\overset{\sim}{\mathbf{X}}.`$ Then from the Azuma-Hoeffding inequality (Lemma [10](#Sx1.Ex74 "Lemma 10 ‣ Appendix A: Probability Results ‣ Optimal Delay-Throughput Trade-offs in Mobile Ad-Hoc Networks: Hybrid Random Walk and One-Dimensional Mobility ModelsAn earlier version of this paper appeared in the Proc. of ITA Workshop, 2007.Research presented here was supported in part by a Vodafone Fellowship and NSF grant CNS 05-19691.")), we have that
<table>
<tbody>
<tr>
<td></td>
<td>Pr(|N𝐣,𝐗~[D]−E[N𝐣,𝐗~[D]]|≤9100DS^)≥1−2e−α625,Prsubscript𝑁𝐣~𝐗delimited-[]𝐷𝐸delimited-[]subscript𝑁𝐣~𝐗delimited-[]𝐷9100𝐷^𝑆12superscript𝑒𝛼625\displaystyle\Pr\left(\left|N_{\mathbf{j},\tilde{\mathbf{X}}}[D]-E[N_{\mathbf{j},\tilde{\mathbf{X}}}[D]]\right|\leq\frac{9}{100}D{\hat{S}}\right)\geq 1-2e^{-\frac{\alpha}{625}},</td>
<td></td>
</tr>
</tbody>
</table>
for any $`\overset{\sim}{\mathbf{X}},`$ and inequality ([46](#Sx2.E46 "In Lemma 12 ‣ Appendix B: Properties of Random Walk ‣ Optimal Delay-Throughput Trade-offs in Mobile Ad-Hoc Networks: Hybrid Random Walk and One-Dimensional Mobility ModelsAn earlier version of this paper appeared in the Proc. of ITA Workshop, 2007.Research presented here was supported in part by a Vodafone Fellowship and NSF grant CNS 05-19691.")) holds. ∎ |
0705.0965 | cs.CR | cyber | INSTITUT NATIONAL DE RECHERCHE EN INFORMATIQUE ET EN AUTOMATIQUE
Improved Analysis of Kannan’s Shortest Lattice Vector Algorithm
Guillaume Hanrot — Damien StehléN° 6186
Mai 2007
Improved Analysis of Kannan’s Shortest Lattice Vector Algorithm
Guillaume Hanrot, Damien Stehlé
<sup>†</sup>
<sup>†</sup>thanks: CNRS and École Normale Supérieure de Lyon, LIP, 46 allée d’Italie, 69007 Lyon, France.
Thème SYM — Systèmes symboliques
Projets Arénaire et Cacao
Rapport de recherche n° 6186 — Mai 2007 — ?? pages
Abstract: The security of lattice-based cryptosystems such as NTRU, GGH and Ajtai-Dwork essentially relies upon the intractability of computing a shortest non-zero lattice vector and a closest lattice vector to a given target vector in high dimensions. The best algorithms for these tasks are due to Kannan, and, though remarkably simple, their complexity estimates have not been improved since more than twenty years. Kannan’s algorithm for solving the shortest vector problem is in particular crucial in Schnorr’s celebrated block reduction algorithm, on which are based the best known attacks against the lattice-based encryption schemes mentioned above. Understanding precisely Kannan’s algorithm is of prime importance for providing meaningful key-sizes. In this paper we improve the complexity analyses of Kannan’s algorithms and discuss the possibility of improving the underlying enumeration strategy.
Key-words: Lattice reduction, complexity analysis, lattice-based cryptosystems.
Amélioration de l’analyse de l’algorithme de Kannan pour le problème du vecteur le plus court
Résumé : La sécurité des cryptosystèmes basés sur les réseaux, tels NTRU, GGH, ou encore Ajtai-Dwork, repose essentiellement sur la difficulté à calculer un vecteur non nul le plus court, ou le plus proche d’un vecteur cible donné, en grande dimension. Les meilleurs algorithmes pour accomplir ces tâches sont dus à Kannan, et, en dépit de leur grande simplicité, l’analyse de leur complexité n’a pas été améliorée depuis plus de 20 ans. L’algorithme de Kannan pour résoudre le problème du vecteur le plus court est particulièrement critique dans le célèbre algorithme de Schnorr pour la réduction par blocs, sur lequel sont basées les meilleures attaques contre les schémas de chiffrement utilisant les réseaux mentionnées précédemment. Comprendre précisément la complexité de l’algorithme de Kannan est donc crucial pour déterminer des tailles de clé pertinentes. Dans ce travail, nous améliorons les analyses de complexité des algorithmes de Kannan, et discutons la possibilité d’améliorer la stratégie d’énumération sous-jacente.
Mots-clés : Réduction des réseaux, analyse de complexité, cryptosystèmes basés sur les réseaux
## 1 Introduction
A lattice $`L`$ is a discrete subgroup of some $`{\mathbb{R}}^{n}`$. Such an object can always be represented as the set of integer linear combinations of no more than $`n`$ vectors $`{\overset{\rightarrow}{b}}_{1},\ldots,{\overset{\rightarrow}{b}}_{d}`$. If these vectors are linearly independent, we say that they are a basis of the lattice $`L`$. The most famous algorithmic problem associated with lattices is the so-called Shortest Vector Problem (SVP). Its computational variant is to find a non-zero lattice vector of smallest Euclidean length — this length being the minimum $`\lambda\hspace{0pt}{(L)}`$ of the lattice — given a basis of the lattice. Its decisional variant is known to be NP-hard under randomised reductions \[[2](#bib.bib2)\], even if one only asks for a vector whose length is no more than $`2^{{({\log d})}^{1 - \epsilon}}`$ times the length of a shortest vector \[[12](#bib.bib12)\] (for any $`\epsilon > 0`$).
SVP is of prime importance in cryptography since a now quite large family of public-key cryptosystems rely more or less on it. The Ajtai-Dwork cryptosystem \[[4](#bib.bib4)\] relies on $`d^{c}`$-SVP for some $`c > 0`$, where $`f\hspace{0pt}{(d)}`$-SVP is the problem of finding the shortest non-zero vector in the lattice $`L`$, knowing that it is unique in the sense that any vector that is of length less than $`{{f\hspace{0pt}{(d)}} \cdot \lambda}\hspace{0pt}{(L)}`$ is parallel to it. The GGH cryptosystem \[[11](#bib.bib11)\] relies on special instances of the Closest Vector Problem (CVP), a non-homogeneous version of SVP. Finally, one strongly suspects that in NTRU \[[15](#bib.bib15)\] – the only realistic lattice-based cryptosystem nowadays, the private key can be read on the coordinates of a shortest vector of the Coppersmith-Shamir lattice \[[8](#bib.bib8)\]. The best known generic attacks on these encryption schemes are based on solving SVP. It is therefore highly important to know precisely what complexity is achievable, both in theory and practice, in particular to select meaningful key-sizes.
In practice, when one wants to obtain good approximations of the lattice minimum, one uses Schnorr’s block-based algorithms \[[23](#bib.bib23), [24](#bib.bib24)\]. These algorithms use internally either Kannan’s algorithm, or the lattice point enumeration procedure on which it relies. This is by far the most time-consuming part of these algorithms. In fact, the corresponding routine in Shoup’s NTL \[[25](#bib.bib25)\] relies on a much slower algorithm described in \[[24](#bib.bib24)\] ($`2^{O\hspace{0pt}{(d^{2})}}`$ instead of $`d^{O\hspace{0pt}{(d)}}`$). The problem is that the enumeration is performed on a basis which is not sufficiently pre-processed (only LLL-reduced). It works well in low dimension, but it can be checked that it is sub-optimal even in moderate dimensions (say 40): the efficiency gap between enumerating from an LLL-reduced basis and from an HKZ-reduced basis shows that there is much room for improving the strategy of \[[24](#bib.bib24)\] by pre-processing the basis before starting the enumeration.
Two main algorithms are known for solving SVP. The first one, which is deterministic, is based on the exhaustive enumeration of lattice points within a small convex set. It is known as Fincke-Pohst’s enumeration algorithm \[[9](#bib.bib9)\] in the algorithmic number theory community. In the cryptography community, it is known as Kannan’s algorithm \[[16](#bib.bib16)\], which is quite similar to the one of Fincke and Pohst. There are two main differences between both: firstly, in Kannan’s algorithm, a long pre-computation on the basis is performed before starting the enumeration process; secondly, Kannan enumerates points in a hyper-parallelepiped whereas Fincke and Pohst do it in an hyper-ellipsoid contained in Kannan’s hyper-parallelepiped – though it may be that Kannan chose the hyper-parallelepiped in order to simplify the complexity analysis. Kannan obtained a $`d^{d + {o\hspace{0pt}{(d)}}}`$ complexity bound (in all the complexity bounds mentioned in the introduction, there is an implicit multiplicative factor that is polynomial in the bit-size of the input). In 1985, Helfrich \[[13](#bib.bib13)\] refined Kannan’s analysis, and obtained a $`d^{{d/2} + {o\hspace{0pt}{(d)}}}`$ complexity bound. On the other hand, Ajtai, Kumar and Sivakumar \[[5](#bib.bib5)\] described a probabilistic algorithm of complexity $`2^{O\hspace{0pt}{(d)}}`$. The best exponent constant is likely to be small. Nevertheless, unless a breakthrough modification is introduced, this algorithm is bound to remain impractical even in moderate dimension since it also requires an exponential space (at least $`2^{d}`$ in dimension $`d`$). On the contrary, the deterministic algorithm of Kannan requires a polynomial space.
Our main result is to lower Helfrich’s complexity bound on Kannan’s algorithm, from $`d^{\frac{d}{2} + {o\hspace{0pt}{(d)}}} \approx d^{0.5 \cdot d}`$ to $`d^{\frac{d}{2\hspace{0pt}e} + {o\hspace{0pt}{(d)}}} \approx d^{{0.184 \cdot d} + {o\hspace{0pt}{(d)}}}`$. This may explain why Kannan’s algorithm is tractable even in moderate dimensions (higher than $`40`$). Our analysis can also be adapted to Kannan’s algorithm that solves the Closest Vector Problem: it decreases Helfrich’s complexity bound from $`d^{d + {o\hspace{0pt}{(d)}}}`$ to $`d^{{d/2} + {o\hspace{0pt}{(d)}}}`$. The complexity improvement on Kannan’s SVP algorithm directly provides better worst-case efficiency/quality trade-offs in Schnorr’s block-based algorithms \[[23](#bib.bib23), [24](#bib.bib24), [10](#bib.bib10)\].
It must be noted that if one follows our analysis step by step, the derived $`o\hspace{0pt}{(d)}`$ may be large when evaluated for some practical $`d`$: the constants hidden in the “$`o\hspace{0pt}{(d)}`$” are improvable (for some of them it may be easy, for others it is probably much harder). No effort was made to improve them, and we believe that it would have complicated the proof with irrelevant details. In fact, most of our analysis consists of estimating the number of lattice points within convex bodies, and showing that the approximation by the volume is valid. By replacing this discretization by heuristic volume estimates, one obtains very small heuristic hidden constants.
Our complexity improvement is based on a fairly simple idea. It is equivalent to generate all lattice points within a ball and to generate all integer points within an ellipsoid (consider the ellipsoid defined by the quadratic form naturally associated with the given lattice basis). Fincke and Pohst noticed that it was more efficient to work with the ellipsoid than to consider a parallelepiped containing it: indeed, when the dimension increases, the ratio of the two volumes shrinks to $`0`$ very quickly. Amazingly, in his analysis, instead of considering the ellipsoid, Kannan bounds the volume of the parallelepiped. Using rather involved technicalities, we bound the volume of the ellipsoid (in fact, the number of integer points within it). Some parts of our proof could be of independent interest. For example, we show that for any Hermite-Korkine-Zolotarev-reduced (HKZ-reduced for short) lattice basis $`({\overset{\rightarrow}{b}}_{1},\ldots,{\overset{\rightarrow}{b}}_{d})`$, and any subset $`I`$ of $`\{ 1,\ldots,d\}`$, we have:
<table>
<tbody>
<tr>
<td></td>
<td>‖b→1‖|I|∏i∈I‖b→i∗‖≤d|I|(1+logd|I|),superscriptnormsubscript→𝑏1𝐼subscriptproduct𝑖𝐼normsuperscriptsubscript→𝑏𝑖superscript𝑑𝐼1𝑑𝐼\frac{\|\vec{b}_{1}\|^{\left|I\right|}}{\prod_{i\in I}\|\vec{b}_{i}^{*}\|}\leq\sqrt{d}^{\left|I\right|\left(1+\log\frac{d}{\left|I\right|}\right)},</td>
<td></td>
</tr>
</tbody>
</table>
where $`{({\overset{\rightarrow}{b}}_{i}^{\ast})}_{i \leq d}`$ is the Gram-Schmidt orthogonalisation of the basis $`({\overset{\rightarrow}{b}}_{1},\ldots,{\overset{\rightarrow}{b}}_{d})`$. This inequality generalises the results of \[[23](#bib.bib23)\] on the quality of HKZ-reduced bases.
Road-Map of the Paper. In Section [2](#S2 "2 Background on Lattice Reduction"), we recall some basic definitions and properties on lattice reduction. Section [3](#S3 "3 Kannan’s SVP Algorithm") is devoted to the description of Kannan’s algorithm and Section [4](#S4 "4 Complexity of the Enumeration Procedure") to its complexity analysis. In Section [5](#S5 "5 CVP and Other Related Problems"), we give without much detail our sibling result on CVP, as well as very direct consequences of our result for Schnorr’s block-based algorithms.
Notation. All logarithms are natural logarithms, i.e., $`{\log{(e)}} = 1`$. Let $`\parallel \cdot \parallel`$ and $`\langle \cdot , \cdot \rangle`$ be the Euclidean norm and inner product of $`{\mathbb{R}}^{n}`$. Bold variables are vectors. We use the bit complexity model. The notation $`\mathcal{P}\hspace{0pt}{(n_{1},\ldots,n_{i})}`$ means $`{({n_{1} \cdot \ldots \cdot n_{i}})}^{c}`$ for some constant $`c > 0`$. If $`x`$ is real, we denote by $`\lfloor x\rceil`$ a closest integer to it (with any convention for making it unique) and we define the centred fractional part $`\{ x\}`$ as $`x - {\lfloor x\rceil}`$. We use the notation $`\text{frac}\hspace{0pt}{(x)}`$ to denote the classical fractional part of $`x`$, i.e., the quantity $`x - {\lfloor x\rfloor}`$. Finally, for any integers $`a`$ and $`b`$, we define $`\left. ⟦a,b \right.⟧`$ as $`{\lbrack a,b\rbrack} \cap {\mathbb{Z}}`$.
## 2 Background on Lattice Reduction
We assume the reader is familiar with the geometry of numbers and its algorithmic aspects. Complete introductions to Euclidean lattices algorithmic problems can be found in \[[20](#bib.bib20)\] and \[[22](#bib.bib22)\].
Gram-Schmidt orthogonalisation. Let $`{\overset{\rightarrow}{b}}_{1},\ldots,{\overset{\rightarrow}{b}}_{d}`$ be linearly independent vectors. Their Gram-Schmidt orthogonalisation (GSO) $`{\overset{\rightarrow}{b}}_{1}^{\ast},\ldots,{\overset{\rightarrow}{b}}_{d}^{\ast}`$ is the orthogonal family defined recursively as follows: the vector $`{\overset{\rightarrow}{b}}_{i}^{\ast}`$ is the component of the vector $`{\overset{\rightarrow}{b}}_{i}`$ which is orthogonal to the linear span of the vectors $`{\overset{\rightarrow}{b}}_{1},\ldots,{\overset{\rightarrow}{b}}_{i - 1}`$. We have $`{\overset{\rightarrow}{b}}_{i}^{\ast} = {{\overset{\rightarrow}{b}}_{i} - {\sum_{j = 1}^{i - 1}{\mu_{i,j}\hspace{0pt}{\overset{\rightarrow}{b}}_{j}^{\ast}}}}`$ where $`\mu_{i,j} = \frac{\langle{\overset{\rightarrow}{b}}_{i},{\overset{\rightarrow}{b}}_{j}^{\ast}\rangle}{\left\| {\overset{\rightarrow}{b}}_{j}^{\ast} \right\|^{2}}`$. For $`i \leq d`$ we let $`\mu_{i,i} = 1`$. Notice that the GSO family depends on the order of the vectors. If the $`{\overset{\rightarrow}{b}}_{i}`$’s are integer vectors, the $`{\overset{\rightarrow}{b}}_{i}^{\ast}`$’s and the $`\mu_{i,j}`$’s are rational.
Lattice volume. The volume of a lattice $`L`$ is defined as $`{\det{(L)}} = {\prod_{i = 1}^{d}\left\| {\overset{\rightarrow}{b}}_{i}^{\ast} \right\|}`$, where the $`{\overset{\rightarrow}{b}}_{i}`$’s are any basis of $`L`$. It does not depend on the choice of the basis of $`L`$ and can be interpreted as the geometric volume of the parallelepiped naturally spanned by the basis vectors.
Minimum and SVP. Another important lattice invariant is the minimum. The minimum $`\lambda\hspace{0pt}{(L)}`$ is the radius of the smallest closed ball centred at the origin containing at least one non-zero lattice vector. The most famous lattice problem is the shortest vector problem. We give here its computational variant: given a basis of a lattice $`L`$, find a lattice vector whose norm is exactly $`\lambda\hspace{0pt}{(L)}`$.
CVP. We give here the computational variant of the closest vector problem: given a basis of a lattice $`L`$ and a target vector in the real span of $`L`$, find a closest vector of $`L`$ to the target vector.
The volume and the minimum of a lattice cannot behave independently. Hermite \[[14](#bib.bib14)\] was the first to bound the ratio $`\frac{\lambda\hspace{0pt}{(L)}}{{({\det L})}^{1/d}}`$ as a function of the dimension only, but his bound was later on greatly improved by Minkowski in his Geometrie der Zahlen \[[21](#bib.bib21)\]. Hermite’s constant $`\gamma_{d}`$ is defined as the supremum over $`d`$ dimensional lattices $`L`$ of the ratio $`\frac{\lambda\hspace{0pt}{(L)}^{2}}{{({\det L})}^{2/d}}`$. In particular, we have $`\gamma_{d} \leq \frac{d + 4}{4}`$ (see \[[18](#bib.bib18)\]), which we will refer to as Minkowski’s theorem. Unfortunately, the proof of Minkowski’s theorem is not constructive. In practice, one often starts with a lattice basis, and tries to improve its quality. This process is called lattice reduction. The most usual ones are probably the LLL and HKZ reductions. Before defining them, we need the concept of size-reduction.
Size-reduction. A basis $`({\overset{\rightarrow}{b}}_{1},\ldots,{\overset{\rightarrow}{b}}_{d})`$ is size-reduced if its GSO family satisfies $`{|\mu_{i,j}|} \leq {1/2}`$ for all $`1 \leq j < i \leq d`$.
HKZ-reduction. A basis $`({\overset{\rightarrow}{b}}_{1},\ldots,{\overset{\rightarrow}{b}}_{d})`$ is said to be Hermite-Korkine-Zolotarev-reduced if it is size-reduced, the vector $`{\overset{\rightarrow}{b}}_{1}`$ reaches the first lattice minimum, and the projections of the $`{({\overset{\rightarrow}{b}}_{i})}_{i \geq 2}`$’s orthogonally to the vector $`{\overset{\rightarrow}{b}}_{1}`$ are an HKZ-reduced basis. The following immediately follows from this definition and Minkowski’s theorem. It is the sole property on HKZ-reduced bases that we will use:
###### Lemma 1
If $`({\overset{\rightarrow}{b}}_{1},\ldots,{\overset{\rightarrow}{b}}_{d})`$ is HKZ-reduced, then for any $`i \leq d`$, we have:
<table>
<tbody>
<tr>
<td></td>
<td>‖b→i∗‖≤d−i+54⋅(∏j≥i‖b→j∗‖)1d−i+1.normsuperscriptsubscript→𝑏𝑖⋅𝑑𝑖54superscriptsubscriptproduct𝑗𝑖normsuperscriptsubscript→𝑏𝑗1𝑑𝑖1\|\vec{b}_{i}^{*}\|\leq\sqrt{\frac{d-i+5}{4}}\cdot\left(\prod_{j\geq i}\|\vec{b}_{j}^{*}\|\right)^{\frac{1}{d-i+1}}.</td>
<td></td>
</tr>
</tbody>
</table>
HKZ-reduction is very strong, but very expensive to compute. On the contrary, LLL-reduction is fairly cheap, but an LLL-reduced basis is of much lower quality.
LLL-reduction \[[17](#bib.bib17)\]. A basis $`({\overset{\rightarrow}{b}}_{1},\ldots,{\overset{\rightarrow}{b}}_{d})`$ is LLL-reduced if it is size-reduced and if its GSO satisfies the $`({d - 1})`$ Lovász conditions: $`{\frac{3}{4} \cdot \left\| {\overset{\rightarrow}{b}}_{\kappa - 1}^{\ast} \right\|^{2}} \leq \left\| {{\overset{\rightarrow}{b}}_{\kappa}^{\ast} + {\mu_{\kappa,{\kappa - 1}}\hspace{0pt}{\overset{\rightarrow}{b}}_{\kappa - 1}^{\ast}}} \right\|^{2}`$. The LLL-reduction implies that the norms $`{\|{\overset{\rightarrow}{b}}_{1}^{\ast}\|},\ldots,{\|{\overset{\rightarrow}{b}}_{d}^{\ast}\|}`$ of the GSO vectors never drop too fast: intuitively, the vectors are not far from being orthogonal. Such bases have useful properties, like providing exponential approximations to SVP and CVP. In particular, their first vector is relatively short. More precisely:
###### Theorem 2.1 (\[[17](#bib.bib17)\])
Let $`({\overset{\rightarrow}{b}}_{1},\ldots,{\overset{\rightarrow}{b}}_{d})`$ be an LLL-reduced basis of a lattice $`L`$. Then we have $`{\|{\overset{\rightarrow}{b}}_{1}\|} \leq {2^{\frac{d - 1}{4}} \cdot {({\det L})}^{1/d}}`$. Moreover, there exists an algorithm that takes as input any set of integer vectors and outputs in deterministic polynomial time an LLL-reduced basis of the lattice they span.
In the following, we will also need the fact that if the set of vectors given as input to the LLL algorithm starts with a shortest non-zero lattice vector, then this vector is not changed during the execution of the algorithm: the output basis starts with the same vector.
## 3 Kannan’s SVP Algorithm
Kannan’s SVP algorithm \[[16](#bib.bib16)\] relies on multiple calls to the so-called short lattice points enumeration procedure. The latter aims at computing all vectors of a given lattice that are in the hyper-sphere centred in $`\overset{\rightarrow}{0}`$ and some prescribed radius. Variants of the enumeration procedure are described in \[[1](#bib.bib1)\].
### 3.1 Short Lattice Points Enumeration
Let $`({\overset{\rightarrow}{b}}_{1},\ldots,{\overset{\rightarrow}{b}}_{d})`$ be a basis of a lattice $`L \subset {\mathbb{Z}}^{n}`$ and let $`A \in {\mathbb{Z}}`$. Our goal is to find all lattice vectors $`\sum_{i = 1}^{d}{x_{i}\hspace{0pt}{\overset{\rightarrow}{b}}_{i}}`$ of squared Euclidean norm $`\leq A`$. The enumeration works as follows. Suppose that $`\left\| {\sum_{i}{x_{i}\hspace{0pt}{\overset{\rightarrow}{b}}_{i}}} \right\|^{2} \leq A`$ for some integers $`x_{i}`$’s. Then, by considering the components of the vector $`\sum_{i}{x_{i}\hspace{0pt}{\overset{\rightarrow}{b}}_{i}}`$ on each of the $`{\overset{\rightarrow}{b}}_{i}^{\ast}`$’s, we obtain:
<table>
<tbody>
<tr>
<td></td>
<td>(xd)2⋅‖b→d∗‖2⋅superscriptsubscript𝑥𝑑2superscriptnormsuperscriptsubscript→𝑏𝑑2\displaystyle\left(x_{d}\right)^{2}\cdot\|\vec{b}_{d}^{*}\|^{2}</td>
<td>≤\displaystyle\leq</td>
<td>A,𝐴\displaystyle A,</td>
<td></td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td>(xd−1+μd,d−1xd)2⋅‖b→d−1∗‖2⋅superscriptsubscript𝑥𝑑1subscript𝜇𝑑𝑑1subscript𝑥𝑑2superscriptnormsuperscriptsubscript→𝑏𝑑12\displaystyle\left(x_{d-1}+\mu_{d,d-1}x_{d}\right)^{2}\cdot\|\vec{b}_{d-1}^{*}\|^{2}</td>
<td>≤\displaystyle\leq</td>
<td>A−(xd)2⋅‖b→d∗‖2,𝐴⋅superscriptsubscript𝑥𝑑2superscriptnormsuperscriptsubscript→𝑏𝑑2\displaystyle A-\left(x_{d}\right)^{2}\cdot\|\vec{b}_{d}^{*}\|^{2},</td>
<td></td>
</tr>
<tr>
<td></td>
<td></td>
<td>……\displaystyle\ldots</td>
<td></td>
<td></td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td>(xi+∑j=i+1dμj,ixj)2⋅‖b→i∗‖2⋅superscriptsubscript𝑥𝑖superscriptsubscript𝑗𝑖1𝑑subscript𝜇𝑗𝑖subscript𝑥𝑗2superscriptnormsuperscriptsubscript→𝑏𝑖2\displaystyle\left(x_{i}+\sum_{j=i+1}^{d}\mu_{j,i}x_{j}\right)^{2}\cdot\|\vec{b}_{i}^{*}\|^{2}</td>
<td>≤\displaystyle\leq</td>
<td>A−∑j=i+1dlj,𝐴superscriptsubscript𝑗𝑖1𝑑subscript𝑙𝑗\displaystyle A-\sum_{j=i+1}^{d}l_{j},</td>
<td></td>
</tr>
<tr>
<td></td>
<td></td>
<td>……\displaystyle\ldots</td>
<td></td>
<td></td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td>(x1+∑j=2dμj,ixj)2⋅‖b→1‖2⋅superscriptsubscript𝑥1superscriptsubscript𝑗2𝑑subscript𝜇𝑗𝑖subscript𝑥𝑗2superscriptnormsubscript→𝑏12\displaystyle\left(x_{1}+\sum_{j=2}^{d}\mu_{j,i}x_{j}\right)^{2}\cdot\|\vec{b}_{1}\|^{2}</td>
<td>≤\displaystyle\leq</td>
<td>A−∑j=2dlj,𝐴superscriptsubscript𝑗2𝑑subscript𝑙𝑗\displaystyle A-\sum_{j=2}^{d}l_{j},</td>
<td></td>
</tr>
</tbody>
</table>
where $`l_{i} = {{({x_{i} + {\sum_{j > i}{x_{j}\hspace{0pt}\mu_{j,i}}}})}^{2} \cdot {\|{\overset{\rightarrow}{b}}_{i}^{\ast}\|}^{2}}`$. The algorithm of Figure [1](#S3.F1 "Figure 1 ‣ 3.1 Short Lattice Points Enumeration ‣ 3 Kannan’s SVP Algorithm") mimics the equations above. It is easy to see that the bit-cost of this algorithm is bounded by the number of loop iterations times a polynomial in the bit-size of the input. We will prove that if the input basis $`({\overset{\rightarrow}{b}}_{1},\ldots,{\overset{\rightarrow}{b}}_{d})`$ is sufficiently reduced and if $`A = {\|{\overset{\rightarrow}{b}}_{1}\|}^{2}`$, then the number of loop iterations is $`d^{\frac{d}{2\hspace{0pt}e} + {o\hspace{0pt}{(d)}}}`$.
Input: An integral lattice basis $`\left( {\overset{\rightarrow}{b}}_{1},\ldots,{\overset{\rightarrow}{b}}_{d} \right)`$, a bound $`A \in {\mathbb{Z}}`$.
Output: All vectors in $`L\hspace{0pt}\left( {\overset{\rightarrow}{b}}_{1},\ldots,{\overset{\rightarrow}{b}}_{d} \right)`$ that are of squared norm $`\leq A`$.
1\. Compute the rational $`\mu_{i,j}`$’s and $`\left\| {\overset{\rightarrow}{b}}_{i}^{\ast} \right\|^{2}`$’s.
2\. $`{\overset{\rightarrow}{x}:=\overset{\rightarrow}{0}},{{\overset{\rightarrow}{l}:=\overset{\rightarrow}{0}},{S:=\varnothing}}`$.
3\. $`i:=1`$. While $`i \leq d`$, do
4\. $`l_{i}:={\left( {x_{i} + \left. \sum{}_{j > i}{x_{j}\hspace{0pt}\mu_{j,i}} \right.} \right)^{2}\hspace{0pt}\left\| {\overset{\rightarrow}{b}}_{i}^{\ast} \right\|^{2}}`$.
5\. If $`i = 1`$ and $`{\sum_{j = 1}^{d}l_{j}} \leq A`$, then $`S:={S \cup \left\{ \overset{\rightarrow}{x} \right\}}`$, $`x_{1}:={x_{1} + 1}`$.
6\. If $`i \neq 1`$ and $`\left. \sum{}_{j \geq i}l_{j} \right. \leq A`$, then
7\. $`i:={i - 1}`$, $`x_{i}:=\left\lceil {{- \left. \sum{}_{j > i}\left( {x_{j}\hspace{0pt}\mu_{j,i}} \right) \right.} - \sqrt{\frac{A - {\left. \sum{}_{j > i} \right.l_{j}}}{\left\| {\overset{\rightarrow}{b}}_{i}^{\ast} \right\|^{2}}}} \right\rceil`$.
8\. If $`\left. \sum{}_{j \geq i}l_{j} \right. > A`$, then $`i:={i + 1}`$, $`x_{i}:={x_{i} + 1}`$.
9\. Return $`S`$.
Figure 1: The Enumeration Algorithm.
### 3.2 Solving SVP
To solve SVP, Kannan provides an algorithm that computes HKZ-reduced bases, see Figure [2](#S3.F2 "Figure 2 ‣ 3.2 Solving SVP ‣ 3 Kannan’s SVP Algorithm"). The cost of the enumeration procedure dominates the overall cost and mostly depends on the quality (i.e., the slow decrease of the $`\|{\overset{\rightarrow}{b}}_{i}^{\ast}\|`$’s) of the input basis. The main idea of Kannan’s algorithm is thus to spend a lot of time pre-computing a basis of excellent quality before calling the enumeration procedure. More precisely, it pre-computes a basis which satisfies the following definition:
###### Definition 1 (Quasi-HKZ-Reduction)
A basis $`({\overset{\rightarrow}{b}}_{1},\ldots,{\overset{\rightarrow}{b}}_{d})`$ is quasi-HKZ-reduced if it is size-reduced, if $`{\|{\overset{\rightarrow}{b}}_{2}^{\ast}\|} \geq {{\|{\overset{\rightarrow}{b}}_{1}^{\ast}\|}/2}`$ and if once projected orthogonally to $`{\overset{\rightarrow}{b}}_{1}`$, the other $`{\overset{\rightarrow}{b}}_{i}`$’s are HKZ-reduced.
Input: An integer lattice basis $`\left( {\overset{\rightarrow}{b}}_{1},\ldots,{\overset{\rightarrow}{b}}_{d} \right)`$.
Output: An HKZ-reduced basis of the same lattice.
1\. LLL-reduce the basis $`\left( {\overset{\rightarrow}{b}}_{1},\ldots,{\overset{\rightarrow}{b}}_{d} \right)`$.
2\. Do
3\. Compute the projections $`\left( {\overset{\rightarrow}{b}}_{i}' \right)_{i \geq 2}`$ of the $`{\overset{\rightarrow}{b}}_{i}`$’s orthogonally to $`{\overset{\rightarrow}{b}}_{1}`$.
4\. HKZ-reduce the $`\left( {d - 1} \right)`$-dimensional basis $`\left( {\overset{\rightarrow}{b}}_{2}',\ldots,{\overset{\rightarrow}{b}}_{d}' \right)`$.
5\. Extend the obtained $`\left( {\overset{\rightarrow}{b}}_{i}' \right)_{i \geq 2}`$’s into vectors of $`L`$ by adding to them rational
multiples of $`{\overset{\rightarrow}{b}}_{1}`$, in such a way that we have $`\left| \mu_{i,1} \right| \leq \left. 1/2 \right.`$ for any $`i > 1`$.
6\. While $`\left( {\overset{\rightarrow}{b}}_{1},\ldots,{\overset{\rightarrow}{b}}_{d} \right)`$ is not quasi-HKZ-reduced.
7\. Call the enumeration procedure to find all lattice vectors of length $`\leq \left\| {\overset{\rightarrow}{b}}_{1} \right\|`$.
Let $`{\overset{\rightarrow}{b}}_{0}`$ be a shortest non-zero vector among them.
8\. $`\left( {\overset{\rightarrow}{b}}_{1},\ldots,{\overset{\rightarrow}{b}}_{d} \right):={\text{LLL}\hspace{0pt}\left( {\overset{\rightarrow}{b}}_{0},\ldots,{\overset{\rightarrow}{b}}_{d} \right)}`$.
9\. Compute the projections $`\left( {\overset{\rightarrow}{b}}_{i}' \right)_{i \geq 2}`$’s of the $`{\overset{\rightarrow}{b}}_{i}`$’s orthogonally to the vector $`{\overset{\rightarrow}{b}}_{1}`$.
10\. HKZ-reduce the $`\left( {d - 1} \right)`$-dimensional basis $`\left( {\overset{\rightarrow}{b}}_{2}',\ldots,{\overset{\rightarrow}{b}}_{d}' \right)`$.
11\. Extend the obtained $`\left( {\overset{\rightarrow}{b}}_{i}' \right)_{i \geq 2}`$’s into vectors of $`L`$ by adding to them rational
multiples of $`{\overset{\rightarrow}{b}}_{1}`$, in such a way that we have $`\left| \mu_{i,1} \right| \leq \left. 1/2 \right.`$ for any $`i > 1`$.
Figure 2: Kannan’s SVP Algorithm.
Several comments need to be made on the algorithm of Figure [2](#S3.F2 "Figure 2 ‣ 3.2 Solving SVP ‣ 3 Kannan’s SVP Algorithm"). Steps 4 and 10 are recursive calls. Nevertheless, one should be careful because the $`{\overset{\rightarrow}{b}}_{i}'`$’s are rational vectors, whereas the input of the algorithm must be integral. One must therefore scale the vectors by a common factor. Steps 5 and 11 can be performed for example by expressing the reduced basis vectors as integer linear combinations of the initial ones, using these coefficients to recover lattice vectors and subtracting a correct multiple of the vector $`{\overset{\rightarrow}{b}}_{1}`$. In Step 7, it is alway possible to choose such a vector $`{\overset{\rightarrow}{b}}_{0}`$, since this enumeration always provides non-zero solutions (the vector $`{\overset{\rightarrow}{b}}_{1}`$ is a one of them).
### 3.3 Cost of Kannan’s SVP Solver
We recall briefly Helfrich’s complexity analysis \[[13](#bib.bib13)\] of Kannan’s algorithm and explain our complexity improvement. Let $`C\hspace{0pt}{(d,n,B)}`$ be the worst-case complexity of the algorithm of Figure [2](#S3.F2 "Figure 2 ‣ 3.2 Solving SVP ‣ 3 Kannan’s SVP Algorithm") when given as input a $`d`$-dimensional basis which is embedded in $`{\mathbb{Z}}^{n}`$ and whose coefficients are smaller than $`B`$ in absolute value. Kannan \[[16](#bib.bib16)\] and Helfrich \[[13](#bib.bib13)\] show the following properties:
- •
It computes an HKZ-reduced basis of the lattice spanned by the input vectors.
- •
All arithmetic operations performed during the execution are of cost $`\mathcal{P}\hspace{0pt}{(d,n,{\log B})}`$. This implies that the cost $`C\hspace{0pt}{(d,n,B)}`$ can be bounded by $`{{C\hspace{0pt}{(d)}} \cdot \mathcal{P}}\hspace{0pt}{({\log B},n)}`$ for some function $`C\hspace{0pt}{(d)}`$.
- •
The number of iterations of the loop of Steps 2–6 is bounded by $`{O\hspace{0pt}{(1)}} + {\log d}`$.
- •
The cost of the call to the enumeration procedure at Step 7 is bounded by $`{\mathcal{P}\hspace{0pt}{({\log B},n)}} \cdot d^{{d/2} + {o\hspace{0pt}{(d)}}}`$.
From these properties and those of the LLL algorithm as recalled in the previous section, it is easy to obtain the following equation:
<table>
<tbody>
<tr>
<td></td>
<td>C(d)≤(O(1)+logd)(C(d−1)+𝒫(d))+𝒫(d)+dd2+o(d).𝐶𝑑𝑂1𝑑𝐶𝑑1𝒫𝑑𝒫𝑑superscript𝑑𝑑2𝑜𝑑C(d)\leq(O(1)+\log d)(C(d-1)+\mathcal{P}(d))+\mathcal{P}(d)+d^{\frac{d}{2}+o(d)}.</td>
<td></td>
</tr>
</tbody>
</table>
One can then derive the bound $`{C\hspace{0pt}{(d,B,n)}} \leq {{\mathcal{P}\hspace{0pt}{({\log B},n)}} \cdot d^{\frac{d}{2} + {o\hspace{0pt}{(d)}}}}`$.
The main result of this paper is to improve this complexity upper bound to $`{\mathcal{P}\hspace{0pt}{({\log B},n)}} \cdot d^{\frac{d}{2\hspace{0pt}e} + {o\hspace{0pt}{(d)}}}`$. In fact, we show the following:
###### Theorem 3.1
Given as inputs a quasi-HKZ-reduced basis $`({\overset{\rightarrow}{b}}_{1},\ldots,{\overset{\rightarrow}{b}}_{d})`$ and $`A = {\|{\overset{\rightarrow}{b}}_{1}\|}^{2}`$, the number of loop iterations during the execution of the enumeration algorithm as described in Figure [1](#S3.F1 "Figure 1 ‣ 3.1 Short Lattice Points Enumeration ‣ 3 Kannan’s SVP Algorithm") is bounded by $`{\mathcal{P}\hspace{0pt}{({\log B})}} \cdot 2^{O\hspace{0pt}{(d)}} \cdot d^{\frac{d}{2\hspace{0pt}e}}`$, where $`B = {\max_{i}{\|{\overset{\rightarrow}{b}}_{i}\|}}`$. As a consequence, given a $`d`$-dimensional basis of $`n`$-dimensional vectors whose entries are integers with absolute values $`\leq B`$, one can compute an HKZ-reduced basis of the lattice they span in deterministic time $`{\mathcal{P}\hspace{0pt}{({\log B},n)}} \cdot d^{\frac{d}{2\hspace{0pt}e} + {o\hspace{0pt}{(d)}}}`$.
## 4 Complexity of the Enumeration Procedure
This section is devoted to proving Theorem [3.1](#S3.Thmtheorem1 "Theorem 3.1 ‣ 3.3 Cost of Kannan’s SVP Solver ‣ 3 Kannan’s SVP Algorithm").
### 4.1 From the Enumeration Procedure to Integer Points in Hyper-ellipsoids
In this subsection, we do not assume anything on the input basis $`({\overset{\rightarrow}{b}}_{1},\ldots,{\overset{\rightarrow}{b}}_{d})`$ and on the input bound $`A`$. Up to some polynomial in $`d`$ and $`\log B`$, the complexity of the enumeration procedure of Figure [1](#S3.F1 "Figure 1 ‣ 3.1 Short Lattice Points Enumeration ‣ 3 Kannan’s SVP Algorithm") is the number of loop iterations. This number of iterations is itself bounded by:
<table>
<tbody>
<tr>
<td></td>
<td>∑i=1d|{(xi,…,xd)∈ℤd−i+1,‖∑j=idxjb→j(i)‖2≤A}|,superscriptsubscript𝑖1𝑑formulae-sequencesubscript𝑥𝑖…subscript𝑥𝑑superscriptℤ𝑑𝑖1superscriptnormsuperscriptsubscript𝑗𝑖𝑑subscript𝑥𝑗superscriptsubscript→𝑏𝑗𝑖2𝐴\sum_{i=1}^{d}\left|\left\{\left(x_{i},\ldots,x_{d}\right)\in{\mathbb{Z}}^{d-i+1},\|\sum_{j=i}^{d}x_{j}\vec{b}_{j}^{(i)}\|^{2}\leq A\right\}\right|,</td>
<td></td>
</tr>
</tbody>
</table>
where $`{\overset{\rightarrow}{b}}_{j}^{(i)} = {{\overset{\rightarrow}{b}}_{j} - {\sum_{k < i}{\mu_{j,k}\hspace{0pt}{\overset{\rightarrow}{b}}_{k}^{\ast}}}}`$ is the vector $`{\overset{\rightarrow}{b}}_{j}`$ once projected orthogonally to the linear span of the vectors $`{\overset{\rightarrow}{b}}_{1},\ldots,{\overset{\rightarrow}{b}}_{i - 1}`$. Indeed, the truncated coordinate $`(x_{i},\ldots,x_{d})`$ is either a valid one, i.e., we have $`{|{\sum_{j = i}^{d}{x_{j}\hspace{0pt}{\overset{\rightarrow}{b}}_{j}^{(i)}}}\|}^{2} \leq A`$, or $`({x_{i} - 1},\ldots,x_{d})`$ is a valid one, or $`(x_{i + 1},\ldots,x_{d})`$ is a valid one. In fact, if $`(x_{i},\ldots,x_{d})`$ is a valid truncated coordinate, only two non-valid ones related to that one can possibly be considered during the execution of the algorithm: $`({x_{i} + 1},\ldots,x_{d})`$ and $`(x_{i - 1},{x_{i}\hspace{0pt}\ldots},x_{d})`$ for at most one integer $`x_{i - 1}`$.
Consider the quantity $`\left| \left\{ {{\left( x_{i},\ldots,x_{d} \right) \in {\mathbb{Z}}^{{d - i} + 1}},{{\|{\sum_{j = i}^{d}{x_{j}\hspace{0pt}{\overset{\rightarrow}{b}}_{j}^{(i)}}}\|}^{2} \leq A}} \right\} \right|`$. By applying the change of variable $`x_{j}\leftarrow x_{j} - \left\lfloor \sum_{k > j}\mu_{k,j}x_{k} \right\rceil`$, we obtain:
<table>
<tbody>
<tr>
<td></td>
<td>∑i≤d|{(xi,…,xd)∈ℤd−i+1\displaystyle\sum_{i\leq d}|\{(x_{i},\ldots,x_{d})\in{\mathbb{Z}}^{d-i+1}</td>
<td>,</td>
<td>∥∑j≥ixjb→j(i)∥2≤A}|\displaystyle\|\sum_{j\geq i}x_{j}\vec{b}_{j}^{(i)}\|^{2}\leq A\}|</td>
<td></td>
</tr>
<tr>
<td></td>
<td></td>
<td>≤\displaystyle\leq</td>
<td>∑i≤d|{(xi,…,xd)∈ℤd−i+1,∑j≥i(xj+∑k>jμk,jxk)2⋅‖b→j∗‖2≤A}|subscript𝑖𝑑formulae-sequencesubscript𝑥𝑖…subscript𝑥𝑑superscriptℤ𝑑𝑖1subscript𝑗𝑖⋅superscriptsubscript𝑥𝑗subscript𝑘𝑗subscript𝜇𝑘𝑗subscript𝑥𝑘2superscriptnormsuperscriptsubscript→𝑏𝑗2𝐴\displaystyle\sum_{i\leq d}|\{(x_{i},\ldots,x_{d})\in{\mathbb{Z}}^{d-i+1},\sum_{j\geq i}(x_{j}+\sum_{k>j}\mu_{k,j}x_{k})^{2}\cdot\|\vec{b}_{j}^{*}\|^{2}\leq A\}|</td>
<td></td>
</tr>
<tr>
<td></td>
<td></td>
<td>≤\displaystyle\leq</td>
<td>∑i≤d|{(xi,…,xd)∈ℤd−i+1,∑j≥i(xj+{∑k>jμk,jxk})2⋅‖b→j∗‖2≤A}|.subscript𝑖𝑑formulae-sequencesubscript𝑥𝑖…subscript𝑥𝑑superscriptℤ𝑑𝑖1subscript𝑗𝑖⋅superscriptsubscript𝑥𝑗subscript𝑘𝑗subscript𝜇𝑘𝑗subscript𝑥𝑘2superscriptnormsuperscriptsubscript→𝑏𝑗2𝐴\displaystyle\sum_{i\leq d}|\{(x_{i},\ldots,x_{d})\in{\mathbb{Z}}^{d-i+1},\sum_{j\geq i}(x_{j}+\{\sum_{k>j}\mu_{k,j}x_{k}\ \})^{2}\cdot\|\vec{b}_{j}^{*}\|^{2}\leq A\}|.</td>
<td></td>
</tr>
</tbody>
</table>
If $`x`$ is an integer and $`\epsilon \in {\lbrack{- {1/2}},{1/2}\rbrack}`$, then we have the relation $`{({x + \epsilon})}^{2} \geq {x^{2}/4}`$. If $`x = 0`$, this is obvious, and otherwise we use the inequality $`{|\epsilon|} \leq {1/2} \leq {{|x|}/2}`$. As a consequence, up to a polynomial factor, the complexity of the enumeration is bounded by:
<table>
<tbody>
<tr>
<td></td>
<td>∑i≤d|{(xi,…,xd)∈ℤd−i+1,∑j≥ixj2⋅‖b→j∗‖2≤4A}|.subscript𝑖𝑑formulae-sequencesubscript𝑥𝑖…subscript𝑥𝑑superscriptℤ𝑑𝑖1subscript𝑗𝑖⋅superscriptsubscript𝑥𝑗2superscriptnormsuperscriptsubscript→𝑏𝑗24𝐴\sum_{i\leq d}\left|\left\{\left(x_{i},\ldots,x_{d}\right)\in\ {\mathbb{Z}}^{d-i+1},\sum_{j\geq i}x_{j}^{2}\cdot\|\vec{b}_{j}^{*}\|^{2}\leq 4A\right\}\right|.</td>
<td></td>
</tr>
</tbody>
</table>
For any $`i \leq d`$, we define the ellipsoid $`\mathcal{E}_{i} = \left\{ {{{(y_{i},\ldots,y_{d})} \in {\mathbb{R}}^{{d - i} + 1}},{{\sum_{j \geq i}{y_{j}^{2} \cdot {\|{\overset{\rightarrow}{b}}_{j}^{\ast}\|}^{2}}} \leq {4\hspace{0pt}A}}} \right\}`$, as well as the quantity $`N_{i} = {|{\mathcal{E}_{i} \cap {\mathbb{Z}}^{{d - i} + 1}}|}`$. We want to bound the sum of the $`N_{i}`$’s. We now fix some index $`i`$. The following sequence of relations is inspired from \[[19](#bib.bib19), Lemma 1\].
<table>
<tbody>
<tr>
<td></td>
<td>Nisubscript𝑁𝑖\displaystyle N_{i}</td>
<td>=\displaystyle=</td>
<td>∑(xi,…,xd)∈ℤd−i+1𝟏ℰi(xi,…,xd)≤exp(d(1−∑j≥ixj2‖b→j∗‖24A))subscriptsubscript𝑥𝑖…subscript𝑥𝑑superscriptℤ𝑑𝑖1subscript1subscriptℰ𝑖subscript𝑥𝑖…subscript𝑥𝑑𝑑1subscript𝑗𝑖superscriptsubscript𝑥𝑗2superscriptnormsuperscriptsubscript→𝑏𝑗24𝐴\displaystyle\sum_{(x_{i},\ldots,x_{d})\in{\mathbb{Z}}^{d-i+1}}{\bf 1}_{{\mathcal{E}}_{i}}(x_{i},\ldots,x_{d})\leq\exp\left(d\left(1-\sum_{j\geq i}x_{j}^{2}\frac{\|\vec{b}_{j}^{*}\|^{2}}{4A}\right)\right)</td>
<td></td>
</tr>
<tr>
<td></td>
<td></td>
<td>≤\displaystyle\leq</td>
<td>ed⋅∏j≥i∑x∈ℤexp(−x2d‖b→i∗‖24A)=ed⋅∏j≥iΘ(d‖b→j∗‖24A),⋅superscript𝑒𝑑subscriptproduct𝑗𝑖subscript𝑥ℤsuperscript𝑥2𝑑superscriptnormsuperscriptsubscript→𝑏𝑖24𝐴⋅superscript𝑒𝑑subscriptproduct𝑗𝑖Θ𝑑superscriptnormsuperscriptsubscript→𝑏𝑗24𝐴\displaystyle e^{d}\cdot\prod_{j\geq i}\sum_{x\in{\mathbb{Z}}}\exp\left(-x^{2}\frac{d\|\vec{b}_{i}^{*}\|^{2}}{4A}\right)=e^{d}\cdot\prod_{j\geq i}\Theta\left(\frac{d\|\vec{b}_{j}^{*}\|^{2}}{4A}\right),</td>
<td></td>
</tr>
</tbody>
</table>
where $`{\Theta\hspace{0pt}{(t)}} = {\sum_{x \in {\mathbb{Z}}}{\exp{({- {t\hspace{0pt}x^{2}}})}}}`$ is defined for $`t > 0`$. Notice that $`{\Theta\hspace{0pt}{(t)}} = {1 + {2\hspace{0pt}{\sum_{x \geq 1}{\exp{({- {t\hspace{0pt}x^{2}}})}}}}} \leq {1 + {2\hspace{0pt}{\int_{0}^{\infty}{{\exp{({- {t\hspace{0pt}x^{2}}})}}\hspace{0pt}{dx}}}}} = {1 + \sqrt{\frac{\pi}{t}}}`$. Hence $`{\Theta\hspace{0pt}{(t)}} \leq \frac{1 + \sqrt{\pi}}{\sqrt{t}}`$ for $`t \leq 1`$ and $`{\Theta\hspace{0pt}{(t)}} \leq {1 + \sqrt{\pi}}`$ for $`t \geq 1`$. As a consequence, we have:
<table>
<tbody>
<tr>
<td></td>
<td>Ni≤(4e(1+π))d⋅∏j≥imax(1,Ad‖b→i∗‖).subscript𝑁𝑖⋅superscript4𝑒1𝜋𝑑subscriptproduct𝑗𝑖1𝐴𝑑normsuperscriptsubscript→𝑏𝑖N_{i}\leq(4e(1+\sqrt{\pi}))^{d}\cdot\prod_{j\geq i}\max\left(1,\frac{\sqrt{A}}{\sqrt{d}\|\vec{b}_{i}^{*}\|}\right).</td>
<td></td>
<td>(1)</td>
</tr>
</tbody>
</table>
One thus concludes that the cost of the enumeration procedure is bounded by:
<table>
<tbody>
<tr>
<td></td>
<td>𝒫(n,logA,logB)⋅2O(d)⋅maxI⊂⟦1,d⟧((A)|I|(d)|I|∏i∈I‖b→i∗‖).⋅𝒫𝑛𝐴𝐵superscript2𝑂𝑑subscript𝐼1𝑑superscript𝐴𝐼superscript𝑑𝐼subscriptproduct𝑖𝐼normsuperscriptsubscript→𝑏𝑖\mathcal{P}(n,\log A,\log B)\cdot 2^{O(d)}\cdot\max_{I\subset\left\llbracket 1,d\right\rrbracket}\left(\frac{(\sqrt{A})^{\left|I\right|}}{(\sqrt{d})^{\left|I\right|}\prod_{i\in I}\|\vec{b}_{i}^{*}\|}\right).</td>
<td></td>
</tr>
</tbody>
</table>
### 4.2 The Case of Quasi-HKZ-Reduced Bases
We know suppose that $`A = {\|{\overset{\rightarrow}{b}}_{1}\|}^{2}`$ and that the input basis $`({\overset{\rightarrow}{b}}_{1},\ldots,{\overset{\rightarrow}{b}}_{d})`$ is quasi-HKZ-reduced. Our first step is to strengthen the quasi-HKZ-reducedness hypothesis to an HKZ-reducedness hypothesis. Let $`I \subset \left. ⟦1,d \right.⟧`$. If $`1 \notin I`$, then, because of the quasi-HKZ-reducedness assumption:
<table>
<tbody>
<tr>
<td></td>
<td>‖b→1‖|I|(d)|I|∏i∈I‖b→i∗‖≤2d‖b→2∗‖|I|(d)|I|∏i∈I‖b→i∗‖.superscriptnormsubscript→𝑏1𝐼superscript𝑑𝐼subscriptproduct𝑖𝐼normsuperscriptsubscript→𝑏𝑖superscript2𝑑superscriptnormsuperscriptsubscript→𝑏2𝐼superscript𝑑𝐼subscriptproduct𝑖𝐼normsuperscriptsubscript→𝑏𝑖\frac{\|\vec{b}_{1}\|^{\left|I\right|}}{(\sqrt{d})^{\left|I\right|}\prod_{i\in I}\|\vec{b}_{i}^{*}\|}\leq 2^{d}\frac{\|\vec{b}_{2}^{*}\|^{\left|I\right|}}{(\sqrt{d})^{\left|I\right|}\prod_{i\in I}\|\vec{b}_{i}^{*}\|}.</td>
<td></td>
</tr>
</tbody>
</table>
Otherwise if $`1 \in I`$, then we have, by removing $`\|{\overset{\rightarrow}{b}}_{1}^{\ast}\|`$ from the product $`\prod_{i \in {I - {\{ 1\}}}}{\|{\overset{\rightarrow}{b}}_{i}^{\ast}\|}`$:
<table>
<tbody>
<tr>
<td></td>
<td>‖b→1‖|I|(d)|I|∏i∈I‖b→i∗‖≤2d‖b→2∗‖|I|−1(d)|I|−1∏i∈I−{1}‖b→i∗‖.superscriptnormsubscript→𝑏1𝐼superscript𝑑𝐼subscriptproduct𝑖𝐼normsuperscriptsubscript→𝑏𝑖superscript2𝑑superscriptnormsuperscriptsubscript→𝑏2𝐼1superscript𝑑𝐼1subscriptproduct𝑖𝐼1normsuperscriptsubscript→𝑏𝑖\frac{\|\vec{b}_{1}\|^{\left|I\right|}}{(\sqrt{d})^{\left|I\right|}\prod_{i\in I}\|\vec{b}_{i}^{*}\|}\leq 2^{d}\frac{\|\vec{b}_{2}^{*}\|^{\left|I\right|-1}}{(\sqrt{d})^{\left|I\right|-1}\prod_{i\in I-\{1\}}\|\vec{b}_{i}^{*}\|}.</td>
<td></td>
</tr>
</tbody>
</table>
As a consequence, in order to obtain Theorem [3.1](#S3.Thmtheorem1 "Theorem 3.1 ‣ 3.3 Cost of Kannan’s SVP Solver ‣ 3 Kannan’s SVP Algorithm"), it suffices to prove the following:
###### Theorem 4.1
Let $`{\overset{\rightarrow}{b}}_{1},\ldots,{\overset{\rightarrow}{b}}_{d}`$ be an HKZ-reduced basis. Let $`I \subset \left. ⟦1,d \right.⟧`$. Then,
<table>
<tbody>
<tr>
<td></td>
<td>‖b→1‖|I|∏i∈I‖b→i∗‖≤(d)|I|(1+logd|I|)≤(d)de+|I|.superscriptnormsubscript→𝑏1𝐼subscriptproduct𝑖𝐼normsuperscriptsubscript→𝑏𝑖superscript𝑑𝐼1𝑑𝐼superscript𝑑𝑑𝑒𝐼\frac{\|\vec{b}_{1}\|^{\left|I\right|}}{\prod_{i\in I}\|\vec{b}_{i}^{*}\|}\leq(\sqrt{d})^{\left|I\right|\left(1+\log\frac{d}{|I|}\right)}\leq(\sqrt{d})^{\frac{d}{e}+\left|I\right|}.</td>
<td></td>
</tr>
</tbody>
</table>
### 4.3 A Property on the Geometry of HKZ-Reduced Bases
In this section, we prove Theorem [4.1](#S4.Thmtheorem1 "Theorem 4.1 ‣ 4.2 The Case of Quasi-HKZ-Reduced Bases ‣ 4 Complexity of the Enumeration Procedure"), which is the last missing part to obtain the announced result. Some parts of the proof are fairly technical and have been postponed to the appendix (this is the case for the proofs of Lemmata [2](#Thmlemma2 "Lemma 2 ‣ 4.3 A Property on the Geometry of HKZ-Reduced Bases ‣ 4 Complexity of the Enumeration Procedure")–[5](#Thmlemma5 "Lemma 5 ‣ 4.3 A Property on the Geometry of HKZ-Reduced Bases ‣ 4 Complexity of the Enumeration Procedure")). As a guide, the reader should consider the typical case where $`{({\overset{\rightarrow}{b}}_{i})}_{1 \leq i \leq d}`$ is an HKZ-reduced basis for which $`{({\|{\overset{\rightarrow}{b}}_{i}^{\ast}\|})}_{i}`$ is a non-increasing sequence. In that case, the shape of the interval $`I`$ that is provided by Equation([1](#S4.E1 "In 4.1 From the Enumeration Procedure to Integer Points in Hyper-ellipsoids ‣ 4 Complexity of the Enumeration Procedure")) is much simpler: it is an interval $`\left. ⟦i,d \right.⟧`$ starting at some index $`i`$. Lemmata [4](#Thmlemma4 "Lemma 4 ‣ 4.3 A Property on the Geometry of HKZ-Reduced Bases ‣ 4 Complexity of the Enumeration Procedure") and [2](#Thmlemma2 "Lemma 2 ‣ 4.3 A Property on the Geometry of HKZ-Reduced Bases ‣ 4 Complexity of the Enumeration Procedure") (which should thus be considered as the core of the proof) and the fact that $`{x\hspace{0pt}{\log x}} \geq {- {1/e}}`$ for $`x \in {\lbrack 0,1\rbrack}`$ are sufficient to deal with such simple intervals, and thus to provide the result.
The difficulties arise when the shape of the set $`I`$ under study becomes more complicated. Though the proof is technically quite involved, the strategy itself can be summed up in a few words. We split our HKZ-reduced basis into blocks (defined by the expression of $`I`$ as a union of intervals), i.e., groups of consecutive vectors $`b_{i},b_{i + 1},\ldots,b_{j - 1}`$ such that $`{i,\ldots,{k - 1}} \notin I`$ and $`{k,\ldots,{j - 1}} \in I`$. The former vectors will be the “large ones”, and the latter the “small ones”. Over each block, Lemma [4](#Thmlemma4 "Lemma 4 ‣ 4.3 A Property on the Geometry of HKZ-Reduced Bases ‣ 4 Complexity of the Enumeration Procedure") relates the average size of the small vectors to the average size of the whole block. We consider the blocks by decreasing indices (in Lemma [6](#Thmlemma6 "Lemma 6 ‣ 4.3 A Property on the Geometry of HKZ-Reduced Bases ‣ 4 Complexity of the Enumeration Procedure")), and use an amortised analysis to combine finely the local behaviours on blocks to obtain a global bound. This recombination is extremely tight, and in order to get the desired bound we use “parts of vectors” (non-integral powers of them). This is why we need to introduce the $`\overset{\sim}{\pi}`$ (in Definition [3](#Thmdefinition3 "Definition 3 ‣ 4.3 A Property on the Geometry of HKZ-Reduced Bases ‣ 4 Complexity of the Enumeration Procedure")). A final convexity argument provided by Lemma [3](#Thmlemma3 "Lemma 3 ‣ 4.3 A Property on the Geometry of HKZ-Reduced Bases ‣ 4 Complexity of the Enumeration Procedure") gives the result.
In the sequel, $`{({\overset{\rightarrow}{b}}_{i})}_{1 \leq i \leq d}`$ is an HKZ-reduced basis of a lattice $`L`$ of dimension $`d \geq 2`$.
###### Definition 2
For any $`I \subset \left. ⟦1,d \right.⟧`$, we define $`\pi_{I} = \left( {\prod_{i \in I}{\|{\overset{\rightarrow}{b}}_{i}^{\ast}\|}} \right)^{\frac{1}{|I|}}`$. Moreover, if $`k \in \left. ⟦1,{d - 1} \right.⟧`$, we define $`{\Gamma_{d}\hspace{0pt}{(k)}} = {\prod_{i = {d - k}}^{d - 1}\gamma_{i + 1}^{\frac{1}{2\hspace{0pt}i}}}`$.
For technical purposes in the proof of Lemma [6](#Thmlemma6 "Lemma 6 ‣ 4.3 A Property on the Geometry of HKZ-Reduced Bases ‣ 4 Complexity of the Enumeration Procedure"), we also need the following definition.
###### Definition 3
If $`1 \leq a < b \leq d`$, where $`a`$ is real and $`b`$ is an integer, we define:
<table>
<tbody>
<tr>
<td></td>
<td>π~[a,b]=(‖b→⌊a⌋∗‖1−a+⌊a⌋⋅∏i=⌊a⌋+1b‖b→i∗‖)1b+1−a=(π⟦⌊a⌋,b⟧)(b+1−⌊a⌋)(1−a+⌊a⌋)b+1−a⋅(π⟦⌊a⌋+1,b⟧)(b−⌊a⌋)(a−⌊a⌋)b+1−a.subscript~𝜋𝑎𝑏superscript⋅superscriptnormsuperscriptsubscript→𝑏𝑎1𝑎𝑎superscriptsubscriptproduct𝑖𝑎1𝑏normsuperscriptsubscript→𝑏𝑖1𝑏1𝑎⋅superscriptsubscript𝜋𝑎𝑏𝑏1𝑎1𝑎𝑎𝑏1𝑎superscriptsubscript𝜋𝑎1𝑏𝑏𝑎𝑎𝑎𝑏1𝑎\tilde{\pi}_{[a,b]}=\left(\|\vec{b}_{\lfloor a\rfloor}^{*}\|^{1-a+\lfloor a\rfloor}\cdot\prod_{i=\lfloor a\rfloor+1}^{b}\|\vec{b}_{i}^{*}\|\right)^{\frac{1}{b+1-a}}=\left(\pi_{\left\llbracket\lfloor a\rfloor,b\right\rrbracket}\right)^{\frac{(b+1-\lfloor a\rfloor)(1-a+\lfloor a\rfloor)}{b+1-a}}\cdot\left(\pi_{\left\llbracket\lfloor a\rfloor+1,b\right\rrbracket}\right)^{\frac{(b-\lfloor a\rfloor)(a-\lfloor a\rfloor)}{b+1-a}}.</td>
<td></td>
</tr>
</tbody>
</table>
Note that Definition [3](#Thmdefinition3 "Definition 3 ‣ 4.3 A Property on the Geometry of HKZ-Reduced Bases ‣ 4 Complexity of the Enumeration Procedure") naturally extends Definition [2](#Thmdefinition2 "Definition 2 ‣ 4.3 A Property on the Geometry of HKZ-Reduced Bases ‣ 4 Complexity of the Enumeration Procedure"), since $`{\overset{\sim}{\pi}}_{\lbrack a,b\rbrack} = \pi_{⟦a,b⟧}`$ when $`a`$ is an integer.
We need estimates on the order of magnitude of $`\Gamma`$, and a technical lemma allowing us to recombine such estimates. Basically, the following lemma is a precise version of the identity:
<table>
<tbody>
<tr>
<td></td>
<td>logΓd(k)≈∫x=d−kdx2logxdx≈log2(d)−log2(d−k)4∼<logd2logdd−k.subscriptΓ𝑑𝑘superscriptsubscript𝑥𝑑𝑘𝑑𝑥2𝑥d𝑥superscript2𝑑superscript2𝑑𝑘4superscriptsimilar-to𝑑2𝑑𝑑𝑘\log\Gamma_{d}(k)\approx\int_{x=d-k}^{d}\frac{x}{2}\log x\,\mathrm{d}x\approx\frac{\log^{2}(d)-\log^{2}(d-k)}{4}\ {\raise-2.15277pt\hbox{$\buildrel<\over{\sim}$}}\ \frac{\log d}{2}\log\frac{d}{d-k}.</td>
<td></td>
</tr>
</tbody>
</table>
###### Lemma 2
For all $`1 \leq k < d`$, we have $`{\Gamma_{d}\hspace{0pt}{(k)}} \leq {\sqrt{d}}^{\log\frac{d}{d - k}}`$.
The following lemma derives from the convexity of the function $`x\mapsto{x\hspace{0pt}{\log x}}`$.
###### Lemma 3
Let $`\Delta \geq 1`$, and define $`{F_{\Delta}\hspace{0pt}{(k,d)}} = \Delta^{- {k\hspace{0pt}{\log\frac{k}{d}}}}`$. We have, for all integer $`t`$, for all integers $`k_{1},\ldots,k_{t}`$ and $`d_{1},\ldots,d_{t}`$ such that $`1 \leq k_{i} < d_{i}`$ for all $`i \leq t`$,
<table>
<tbody>
<tr>
<td></td>
<td>∏i≤tFΔ(ki,di)≤FΔ(∑i≤tki,∑i≤tdi).subscriptproduct𝑖𝑡subscript𝐹Δsubscript𝑘𝑖subscript𝑑𝑖subscript𝐹Δsubscript𝑖𝑡subscript𝑘𝑖subscript𝑖𝑡subscript𝑑𝑖\prod_{i\leq t}F_{\Delta}(k_{i},d_{i})\leq F_{\Delta}\left(\sum_{i\leq t}k_{i},\sum_{i\leq t}d_{i}\right).</td>
<td></td>
</tr>
</tbody>
</table>
We now give an “averaged” version of \[[23](#bib.bib23), Lemma 4\]. For completeness, we give its proof in appendix. This provides the result claimed in Theorem [4.1](#S4.Thmtheorem1 "Theorem 4.1 ‣ 4.2 The Case of Quasi-HKZ-Reduced Bases ‣ 4 Complexity of the Enumeration Procedure") for any interval $`I = \left. ⟦i,j \right.⟧`$, for any $`i \leq j \leq d`$.
###### Lemma 4
For all $`k \in \left. ⟦0,{d - 1} \right.⟧`$, we have
<table>
<tbody>
<tr>
<td></td>
<td>π⟦1,k⟧≤(Γd(k))d/k⋅π⟦k+1,d⟧andπ⟦k+1,d⟧≥(Γd(k))−1⋅(detL)1/d≥dlogd−kd(detL)1/d.formulae-sequencesubscript𝜋1𝑘⋅superscriptsubscriptΓ𝑑𝑘𝑑𝑘subscript𝜋𝑘1𝑑𝑎𝑛𝑑subscript𝜋𝑘1𝑑⋅superscriptsubscriptΓ𝑑𝑘1superscript𝐿1𝑑superscript𝑑𝑑𝑘𝑑superscript𝐿1𝑑\pi_{\left\llbracket 1,k\right\rrbracket}\leq\left(\Gamma_{d}(k)\right)^{d/k}\cdot\pi_{\left\llbracket k+1,d\right\rrbracket}\ \ and\ \ \pi_{\left\llbracket k+1,d\right\rrbracket}\geq\left(\Gamma_{d}(k)\right)^{-1}\cdot(\det L)^{1/d}\geq\sqrt{d}^{\log\frac{d-k}{d}}(\det L)^{1/d}.</td>
<td></td>
</tr>
</tbody>
</table>
The following lemma extends Lemma [4](#Thmlemma4 "Lemma 4 ‣ 4.3 A Property on the Geometry of HKZ-Reduced Bases ‣ 4 Complexity of the Enumeration Procedure") to the case where $`k`$ is not necessarily an integer. Its proof is conceptually simple, but involves rather heavy elementary calculus. It would be simpler to obtain it with a relaxation factor. The result is nevertheless worth the effort since the shape of the bound is extremely tractable in the sequel.
###### Lemma 5
If $`1 \leq x_{1} < x_{2} < d`$ are real and in $`\lbrack 1,d)`$, then $`{\overset{\sim}{\pi}}_{\lbrack x_{2},d\rbrack} \geq {{\sqrt{d}}^{\log\frac{d - x_{2}}{d - x_{1}}} \cdot {\overset{\sim}{\pi}}_{\lbrack x_{1},d\rbrack}}`$.
We prove Theorem [4.1](#S4.Thmtheorem1 "Theorem 4.1 ‣ 4.2 The Case of Quasi-HKZ-Reduced Bases ‣ 4 Complexity of the Enumeration Procedure") by induction on the number of intervals occurring in the expression of the set $`I`$ as a union of intervals. The following lemma is the induction step. This is a recombination step, where we join one block (between the indices $`1`$ and $`v`$, the “small vectors” being those between $`u + 1`$ and $`v`$) to one or more already considered blocks on its right. An important point is to ensure that the densities $`\delta_{i}`$ defined below actually decrease.
###### Lemma 6
Let $`({\overset{\rightarrow}{b}}_{1},\ldots,{\overset{\rightarrow}{b}}_{d})`$ be an HKZ-reduced basis. Let $`v \in \left. ⟦2,d \right.⟧`$, $`I \subset \left. ⟦{v + 1},d \right.⟧`$ and $`u \in \left. ⟦1,v \right.⟧`$. Assume that:
<table>
<tbody>
<tr>
<td></td>
<td>πI|I|≥∏i<t(π⟦αi+1,αi+1⟧|Ii|⋅d|Ii|logδi),superscriptsubscript𝜋𝐼𝐼subscriptproduct𝑖𝑡⋅superscriptsubscript𝜋subscript𝛼𝑖1subscript𝛼𝑖1subscript𝐼𝑖superscript𝑑subscript𝐼𝑖subscript𝛿𝑖\pi_{I}^{\left|I\right|}\geq\prod_{i<t}\left(\pi_{\left\llbracket\alpha_{i}+1,\alpha_{i+1}\right\rrbracket}^{\left|I_{i}\right|}\cdot\sqrt{d}^{\left|I_{i}\right|\log\delta_{i}}\right),</td>
<td></td>
</tr>
</tbody>
</table>
where $`{I_{i} = {I \cap \left. ⟦{\alpha_{i} + 1},\alpha_{i + 1} \right.⟧}},{\delta_{i} = \frac{\left| I_{i} \right|}{\alpha_{i + 1} - \alpha_{i}}}`$ is the density of $`I`$ in $`\left. ⟦{\alpha_{i} + 1},\alpha_{i + 1} \right.⟧`$, and the integers $`t`$ and $`\alpha_{i}`$’s, and the densities $`\delta_{i}`$ satisfy $`t \geq 1`$, $`v = \alpha_{1} < \alpha_{2} < \ldots < \alpha_{t} \leq d`$ and $`1 \geq \delta_{1} > \ldots > \delta_{t - 1} > 0`$.
Then, we have
<table>
<tbody>
<tr>
<td></td>
<td>πI′|I′|≥∏i<t′(π⟦αi′+1,αi+1′⟧|Ii′|⋅d|Ii′|logδi′),superscriptsubscript𝜋superscript𝐼′superscript𝐼′subscriptproduct𝑖superscript𝑡′⋅superscriptsubscript𝜋superscriptsubscript𝛼𝑖′1superscriptsubscript𝛼𝑖1′subscriptsuperscript𝐼′𝑖superscript𝑑subscriptsuperscript𝐼′𝑖superscriptsubscript𝛿𝑖′\pi_{I^{\prime}}^{\left|I^{\prime}\right|}\geq\prod_{i<t^{\prime}}\left(\pi_{\left\llbracket\alpha_{i}^{\prime}+1,\alpha_{i+1}^{\prime}\right\rrbracket}^{\left|I^{\prime}_{i}\right|}\cdot\sqrt{d}^{\left|I^{\prime}_{i}\right|\log\delta_{i}^{\prime}}\right),</td>
<td></td>
</tr>
</tbody>
</table>
where $`{I' = {\left. ⟦{u + 1},v \right.⟧ \cup I}},{{I_{i}' = {I' \cap \left. ⟦{\alpha_{i}' + 1},\alpha_{i + 1}' \right.⟧}},{\delta_{i}' = \frac{\left| I_{i}' \right|}{\alpha_{i + 1}' - \alpha_{i}'}}}`$ and the integers $`t'`$ and $`\alpha_{i}'`$’s, and the densities $`\delta_{i}'`$ satisfy $`t' \geq 1`$, $`0 = \alpha_{1}' < \alpha_{2}' < \ldots < \alpha_{t'}' \leq d`$ and $`1 \geq \delta_{1}' > \ldots > \delta_{t' - 1}' > 0`$.
###### Proof
Assume first that $`\frac{v - u}{v} \geq \delta_{1}`$, Then, thanks to Lemma [4](#Thmlemma4 "Lemma 4 ‣ 4.3 A Property on the Geometry of HKZ-Reduced Bases ‣ 4 Complexity of the Enumeration Procedure"),
<table>
<tbody>
<tr>
<td></td>
<td>πI′|I′|=π⟦u+1,v⟧v−u⋅πI|I|≥π⟦1,v⟧v−u⋅d(v−u)v−uv⋅πI|I|,superscriptsubscript𝜋superscript𝐼′superscript𝐼′⋅superscriptsubscript𝜋𝑢1𝑣𝑣𝑢superscriptsubscript𝜋𝐼𝐼⋅superscriptsubscript𝜋1𝑣𝑣𝑢superscript𝑑𝑣𝑢𝑣𝑢𝑣superscriptsubscript𝜋𝐼𝐼\pi_{I^{\prime}}^{\left|I^{\prime}\right|}=\pi_{\left\llbracket u+1,v\right\rrbracket}^{v-u}\cdot\pi_{I}^{\left|I\right|}\geq\pi_{\left\llbracket 1,v\right\rrbracket}^{v-u}\cdot\sqrt{d}^{(v-u)\frac{v-u}{v}}\cdot\pi_{I}^{\left|I\right|},</td>
<td></td>
</tr>
</tbody>
</table>
we are done with $`t' = {t + 1}`$, $`\alpha_{1}' = 1`$, $`\alpha_{k}' = \alpha_{k - 1}`$, $`\delta_{1}' = \frac{v - u}{v}`$, $`\delta_{k}' = \delta_{k - 1}`$.
Otherwise, we let $`\lambda_{1} > 0`$ be such that $`\frac{v - u}{v - \lambda_{1}} = \delta_{1} = \frac{{v - u} + \left| I_{1} \right|}{\alpha_{2} - \lambda_{1}}`$, where the first equality defines $`\lambda_{1}`$ and the second one follows. Note that this implies:
<table>
<tbody>
<tr>
<td></td>
<td>π~[λ1,v]v−u⋅π⟦v+1,α2⟧|I1|=π~[λ1,α2]v−u+|I1|.⋅superscriptsubscript~𝜋subscript𝜆1𝑣𝑣𝑢superscriptsubscript𝜋𝑣1subscript𝛼2subscript𝐼1superscriptsubscript~𝜋subscript𝜆1subscript𝛼2𝑣𝑢subscript𝐼1\tilde{\pi}_{[\lambda_{1},v]}^{v-u}\cdot\pi_{\left\llbracket v+1,\alpha_{2}\right\rrbracket}^{|I_{1}|}=\tilde{\pi}_{[\lambda_{1},\alpha_{2}]}^{v-u+|I_{1}|}.</td>
<td></td>
</tr>
</tbody>
</table>
Then, we have, by using Lemma [5](#Thmlemma5 "Lemma 5 ‣ 4.3 A Property on the Geometry of HKZ-Reduced Bases ‣ 4 Complexity of the Enumeration Procedure"),
<table>
<tbody>
<tr>
<td></td>
<td>πI′|I′|superscriptsubscript𝜋superscript𝐼′superscript𝐼′\displaystyle\pi_{I^{\prime}}^{\left|I^{\prime}\right|}</td>
<td>=\displaystyle=</td>
<td>π⟦u+1,v⟧v−u⋅πI|I|⋅superscriptsubscript𝜋𝑢1𝑣𝑣𝑢superscriptsubscript𝜋𝐼𝐼\displaystyle\pi_{\left\llbracket u+1,v\right\rrbracket}^{v-u}\cdot\pi_{I}^{\left|I\right|}</td>
<td></td>
</tr>
<tr>
<td></td>
<td></td>
<td>≥\displaystyle\geq</td>
<td>(π~[λ1,v]v−u⋅d(v−u)logv−uv−λ1)⋅∏i<t(π⟦αi+1,αi+1⟧|Ii|⋅d|Ii|logδi)⋅⋅superscriptsubscript~𝜋subscript𝜆1𝑣𝑣𝑢superscript𝑑𝑣𝑢𝑣𝑢𝑣subscript𝜆1subscriptproduct𝑖𝑡⋅superscriptsubscript𝜋subscript𝛼𝑖1subscript𝛼𝑖1subscript𝐼𝑖superscript𝑑subscript𝐼𝑖subscript𝛿𝑖\displaystyle\left(\tilde{\pi}_{[\lambda_{1},v]}^{v-u}\cdot\sqrt{d}^{(v-u)\log\frac{v-u}{v-\lambda_{1}}}\right)\cdot\prod_{i<t}\left(\pi_{\left\llbracket\alpha_{i}+1,\alpha_{i+1}\right\rrbracket}^{\left|I_{i}\right|}\cdot\sqrt{d}^{\left|I_{i}\right|\log\delta_{i}}\right)</td>
<td></td>
</tr>
<tr>
<td></td>
<td></td>
<td>≥\displaystyle\geq</td>
<td>(π~[λ1,v]v−u⋅π⟦v+1,α2⟧|I1|⋅d(v−u)logv−uv−λ1+|I1|⋅logδ1)⋅∏i=2t−1(π⟦αi+1,αi+1⟧|Ii|⋅d|Ii|logδi)⋅⋅superscriptsubscript~𝜋subscript𝜆1𝑣𝑣𝑢superscriptsubscript𝜋𝑣1subscript𝛼2subscript𝐼1superscript𝑑𝑣𝑢𝑣𝑢𝑣subscript𝜆1⋅subscript𝐼1subscript𝛿1superscriptsubscriptproduct𝑖2𝑡1⋅superscriptsubscript𝜋subscript𝛼𝑖1subscript𝛼𝑖1subscript𝐼𝑖superscript𝑑subscript𝐼𝑖subscript𝛿𝑖\displaystyle\left(\tilde{\pi}_{[\lambda_{1},v]}^{v-u}\cdot\pi_{\left\llbracket v+1,\alpha_{2}\right\rrbracket}^{\left|I_{1}\right|}\cdot\sqrt{d}^{(v-u)\log\frac{v-u}{v-\lambda_{1}}+\left|I_{1}\right|\cdot\log\delta_{1}}\right)\cdot\prod_{i=2}^{t-1}\left(\pi_{\left\llbracket\alpha_{i}+1,\alpha_{i+1}\right\rrbracket}^{\left|I_{i}\right|}\cdot\sqrt{d}^{\left|I_{i}\right|\log\delta_{i}}\right)</td>
<td></td>
</tr>
<tr>
<td></td>
<td></td>
<td>≥\displaystyle\geq</td>
<td>(π~[λ1,α2]v−u+|I1|⋅d(v−u+|I1|)logv−u+|I1|α2−λ1)⋅∏i=2t−1(π⟦αi+1,αi+1⟧|Ii|⋅d|Ii|logδi),⋅⋅superscriptsubscript~𝜋subscript𝜆1subscript𝛼2𝑣𝑢subscript𝐼1superscript𝑑𝑣𝑢subscript𝐼1𝑣𝑢subscript𝐼1subscript𝛼2subscript𝜆1superscriptsubscriptproduct𝑖2𝑡1⋅superscriptsubscript𝜋subscript𝛼𝑖1subscript𝛼𝑖1subscript𝐼𝑖superscript𝑑subscript𝐼𝑖subscript𝛿𝑖\displaystyle\left(\tilde{\pi}_{[\lambda_{1},\alpha_{2}]}^{v-u+\left|I_{1}\right|}\cdot\sqrt{d}^{(v-u+\left|I_{1}\right|)\log\frac{v-u+\left|I_{1}\right|}{\alpha_{2}-\lambda_{1}}}\right)\cdot\prod_{i=2}^{t-1}\left(\pi_{\left\llbracket\alpha_{i}+1,\alpha_{i+1}\right\rrbracket}^{\left|I_{i}\right|}\cdot\sqrt{d}^{\left|I_{i}\right|\log\delta_{i}}\right),</td>
<td></td>
</tr>
</tbody>
</table>
If $`\frac{{v - u} + \left| I_{1} \right|}{\alpha_{2}} > \frac{\left| I_{2} \right|}{\alpha_{3} - \alpha_{2}}`$, we conclude as in the first step, putting $`t' = t`$, $`\alpha_{1}' = 1`$, $`\alpha_{k}' = \alpha_{k}`$ for $`k \geq 2`$, $`\delta_{1}' = {{({{v - u} + {|I_{1}|}})}/\alpha_{2}}`$, $`\delta_{k}' = \delta_{k}`$ for $`k \geq 2`$. If this is not the case, we let $`\lambda_{2}`$ be such that:
<table>
<tbody>
<tr>
<td></td>
<td>v−u+|I1|α2−λ2=δ2=v−u+|I∩⟦α1+1,α3⟧|α3−λ2.𝑣𝑢subscript𝐼1subscript𝛼2subscript𝜆2subscript𝛿2𝑣𝑢𝐼subscript𝛼11subscript𝛼3subscript𝛼3subscript𝜆2\frac{v-u+\left|I_{1}\right|}{\alpha_{2}-\lambda_{2}}=\delta_{2}=\frac{v-u+\left|I\cap\left\llbracket\alpha_{1}+1,\alpha_{3}\right\rrbracket\right|}{\alpha_{3}-\lambda_{2}}.</td>
<td></td>
</tr>
</tbody>
</table>
Notice that since $`\delta_{1} = \frac{{v - u} + {|I_{1}|}}{\alpha_{2} - \lambda_{1}} > \delta_{2}`$, we have $`\lambda_{2} < \lambda_{1}`$. A similar sequence of inequalities, using Lemma [5](#Thmlemma5 "Lemma 5 ‣ 4.3 A Property on the Geometry of HKZ-Reduced Bases ‣ 4 Complexity of the Enumeration Procedure") to relate $`{\overset{\sim}{\pi}}_{\lbrack\lambda_{1},\alpha_{2}\rbrack}`$ to $`{\overset{\sim}{\pi}}_{\lbrack\lambda_{2},\alpha_{2}\rbrack}`$, leads to the following lower bound on $`\pi_{I'}^{|I'|}`$:
<table>
<tbody>
<tr>
<td></td>
<td>(π~[λ2,α3]v−u+|I∩⟦α1+1,α3⟧|⋅d(v−u+|I∩⟦α1+1,α3⟧|)logv−u+|I∩⟦α1+1,α3⟧|α3−λ2)⋅∏i=3t−1(π⟦αi+1,αi+1⟧|Ii|⋅d|Ii|logδi)⋅⋅superscriptsubscript~𝜋subscript𝜆2subscript𝛼3𝑣𝑢𝐼subscript𝛼11subscript𝛼3superscript𝑑𝑣𝑢𝐼subscript𝛼11subscript𝛼3𝑣𝑢𝐼subscript𝛼11subscript𝛼3subscript𝛼3subscript𝜆2superscriptsubscriptproduct𝑖3𝑡1⋅superscriptsubscript𝜋subscript𝛼𝑖1subscript𝛼𝑖1subscript𝐼𝑖superscript𝑑subscript𝐼𝑖subscript𝛿𝑖\left(\tilde{\pi}_{[\lambda_{2},\alpha_{3}]}^{v-u+\left|I\cap\left\llbracket\alpha_{1}+1,\alpha_{3}\right\rrbracket\right|}\cdot\sqrt{d}^{(v-u+\left|I\cap\left\llbracket\alpha_{1}+1,\alpha_{3}\right\rrbracket\right|)\log\frac{v-u+\left|I\cap\left\llbracket\alpha_{1}+1,\alpha_{3}\right\rrbracket\right|}{\alpha_{3}-\lambda_{2}}}\right)\cdot\prod_{i=3}^{t-1}\left(\pi_{\left\llbracket\alpha_{i}+1,\alpha_{i+1}\right\rrbracket}^{\left|I_{i}\right|}\cdot\sqrt{d}^{\left|I_{i}\right|\log\delta_{i}}\right)</td>
<td></td>
</tr>
</tbody>
</table>
We can proceed in the same way, constructing $`\lambda_{2} > \lambda_{3} > \ldots`$. Suppose first that the construction stops at some point. We have:
<table>
<tbody>
<tr>
<td></td>
<td>πI′|I′|superscriptsubscript𝜋superscript𝐼′superscript𝐼′\displaystyle\pi_{I^{\prime}}^{\left|I^{\prime}\right|}</td>
<td>≥\displaystyle\geq</td>
<td>(π⟦1,αk+1⟧|I′∩⟦1,αk+1⟧|⋅d|I′∩⟦1,αk+1⟧|log|I′∩⟦1,αk+1⟧|αk+1)⋅∏i=k+1t−1(π⟦αi+1,αi+1⟧|Ii|d|Ii|logδi).⋅⋅superscriptsubscript𝜋1subscript𝛼𝑘1superscript𝐼′1subscript𝛼𝑘1superscript𝑑superscript𝐼′1subscript𝛼𝑘1superscript𝐼′1subscript𝛼𝑘1subscript𝛼𝑘1superscriptsubscriptproduct𝑖𝑘1𝑡1superscriptsubscript𝜋subscript𝛼𝑖1subscript𝛼𝑖1subscript𝐼𝑖superscript𝑑subscript𝐼𝑖subscript𝛿𝑖\displaystyle\left(\pi_{\left\llbracket 1,\alpha_{k+1}\right\rrbracket}^{\left|I^{\prime}\cap\left\llbracket 1,\alpha_{k+1}\right\rrbracket\right|}\cdot\sqrt{d}^{\left|I^{\prime}\cap\left\llbracket 1,\alpha_{k+1}\right\rrbracket\right|\log\frac{\left|I^{\prime}\cap\left\llbracket 1,\alpha_{k+1}\right\rrbracket\right|}{\alpha_{k+1}}}\right)\cdot\prod_{i=k+1}^{t-1}\left(\pi_{\left\llbracket\alpha_{i}+1,\alpha_{i+1}\right\rrbracket}^{\left|I_{i}\right|}\sqrt{d}^{\left|I_{i}\right|\log\delta_{i}}\right).</td>
<td></td>
</tr>
</tbody>
</table>
We can then conclude, by putting $`t' = {{t - k} + 1}`$, $`{\alpha_{1}' = 1},{\alpha_{j}' = \alpha_{{j + k} - 1}}`$ for $`j > 1`$, $`\delta_{1}' = {{|{I' \cap \left. ⟦1,\alpha_{k + 1} \right.⟧}|}/\alpha_{k + 1}}`$, $`\delta_{j}' = \delta_{{j + k} - 1}`$ for $`j > 1`$.
Otherwise, we end up with:
<table>
<tbody>
<tr>
<td></td>
<td>πI′|I′|superscriptsubscript𝜋superscript𝐼′superscript𝐼′\displaystyle\pi_{I^{\prime}}^{\left|I^{\prime}\right|}</td>
<td>≥\displaystyle\geq</td>
<td>π~[λt−2,αt−1]|I′|⋅d|I′|log|I′∩⟦1,αt−1⟧|αt−1−λt−2,⋅superscriptsubscript~𝜋subscript𝜆𝑡2subscript𝛼𝑡1superscript𝐼′superscript𝑑superscript𝐼′superscript𝐼′1subscript𝛼𝑡1subscript𝛼𝑡1subscript𝜆𝑡2\displaystyle\tilde{\pi}_{[\lambda_{t-2},\alpha_{t-1}]}^{\left|I^{\prime}\right|}\cdot\sqrt{d}^{\left|I^{\prime}\right|\log\frac{\left|I^{\prime}\cap\left\llbracket 1,\alpha_{t-1}\right\rrbracket\right|}{\alpha_{t-1}-\lambda_{t-2}}},</td>
<td></td>
</tr>
</tbody>
</table>
to which we can apply Lemma [5](#Thmlemma5 "Lemma 5 ‣ 4.3 A Property on the Geometry of HKZ-Reduced Bases ‣ 4 Complexity of the Enumeration Procedure") to obtain $`\pi_{I'}^{|I'|} \geq {\pi_{⟦1,\alpha_{t - 1}⟧}^{|I'|} \cdot {\sqrt{d}}^{{|I'|}\hspace{0pt}{\log\frac{|{I' \cap {⟦1,\alpha_{t - 1}⟧}}|}{\alpha_{t - 1}}}}}`$, which is again in the desired form, with $`t' = 2`$, $`\alpha_{1}' = 1`$, $`\alpha_{2}' = \alpha_{t - 1}`$, $`\delta_{1}' = \frac{\left| {I' \cap \left. ⟦1,\alpha_{t - 1} \right.⟧} \right|}{\alpha_{t - 1}}`$ $`\square`$
Theorem [4.1](#S4.Thmtheorem1 "Theorem 4.1 ‣ 4.2 The Case of Quasi-HKZ-Reduced Bases ‣ 4 Complexity of the Enumeration Procedure") now follows from successive applications of Lemma [6](#Thmlemma6 "Lemma 6 ‣ 4.3 A Property on the Geometry of HKZ-Reduced Bases ‣ 4 Complexity of the Enumeration Procedure"), as follows:
Proof of Theorem [4.1](#S4.Thmtheorem1 "Theorem 4.1 ‣ 4.2 The Case of Quasi-HKZ-Reduced Bases ‣ 4 Complexity of the Enumeration Procedure"). Lemma [6](#Thmlemma6 "Lemma 6 ‣ 4.3 A Property on the Geometry of HKZ-Reduced Bases ‣ 4 Complexity of the Enumeration Procedure") gives us, by induction on the size of the considered set $`I`$, that for all $`I \subset \left. ⟦1,d \right.⟧`$, we have:
<table>
<tbody>
<tr>
<td></td>
<td>πI|I|≥∏i<t(π⟦αi+1,αi+1⟧|Ii|⋅d|Ii|logδi),superscriptsubscript𝜋𝐼𝐼subscriptproduct𝑖𝑡⋅superscriptsubscript𝜋subscript𝛼𝑖1subscript𝛼𝑖1subscript𝐼𝑖superscript𝑑subscript𝐼𝑖subscript𝛿𝑖\pi_{I}^{\left|I\right|}\geq\prod_{i<t}\left(\pi_{\left\llbracket\alpha_{i}+1,\alpha_{i+1}\right\rrbracket}^{\left|I_{i}\right|}\cdot\sqrt{d}^{\left|I_{i}\right|\log\delta_{i}}\right),</td>
<td></td>
</tr>
</tbody>
</table>
where $`I_{i} = {I \cap \left. ⟦{\alpha_{i} + 1},\alpha_{i + 1} \right.⟧}`$, and the integers $`t`$ and $`\alpha_{i}`$’s, and the densities $`\delta_{i} = \frac{\left| I_{i} \right|}{\alpha_{i + 1} - \alpha_{i}}`$ satisfy $`t \geq 1`$, $`0 = \alpha_{1} < \alpha_{2} < \ldots < \alpha_{t} \leq d`$ and $`1 \geq \delta_{1} > \ldots > \delta_{t - 1} > 0`$. By using Lemma [3](#Thmlemma3 "Lemma 3 ‣ 4.3 A Property on the Geometry of HKZ-Reduced Bases ‣ 4 Complexity of the Enumeration Procedure") with $`{\Delta:=\sqrt{d}},{k_{i}:=\left| I_{i} \right|}`$ and $`d_{i}:={\alpha_{i + 1} - \alpha_{i}}`$, we immediately obtain:
<table>
<tbody>
<tr>
<td></td>
<td>πI|I|≥(d|I|log|I|αt−α1)⋅(∏i<tπ⟦αi+1,αi+1⟧|Ii|).superscriptsubscript𝜋𝐼𝐼⋅superscript𝑑𝐼𝐼subscript𝛼𝑡subscript𝛼1subscriptproduct𝑖𝑡superscriptsubscript𝜋subscript𝛼𝑖1subscript𝛼𝑖1subscript𝐼𝑖\pi_{I}^{\left|I\right|}\geq\left(\sqrt{d}^{\left|I\right|\log\frac{\left|I\right|}{\alpha_{t}-\alpha_{1}}}\right)\cdot\left(\prod_{i<t}\pi_{\left\llbracket\alpha_{i}+1,\alpha_{i+1}\right\rrbracket}^{\left|I_{i}\right|}\right).</td>
<td></td>
</tr>
</tbody>
</table>
For convenience, we define $`\delta_{t} = 0`$. Because of the definition of the $`\alpha_{i}`$’s, we have:
<table>
<tbody>
<tr>
<td></td>
<td>∏i<tπ⟦αi+1,αi+1⟧|Ii|subscriptproduct𝑖𝑡superscriptsubscript𝜋subscript𝛼𝑖1subscript𝛼𝑖1subscript𝐼𝑖\displaystyle\prod_{i<t}\pi_{\left\llbracket\alpha_{i}+1,\alpha_{i+1}\right\rrbracket}^{\left|I_{i}\right|}</td>
<td>=\displaystyle=</td>
<td>∏i<t(π⟦αi+1,αi+1⟧αi+1−αi)δi=∏i<t∏i≤j<t(π⟦αi+1,αi+1⟧αi+1−αi)δj−δj+1subscriptproduct𝑖𝑡superscriptsuperscriptsubscript𝜋subscript𝛼𝑖1subscript𝛼𝑖1subscript𝛼𝑖1subscript𝛼𝑖subscript𝛿𝑖subscriptproduct𝑖𝑡subscriptproduct𝑖𝑗𝑡superscriptsuperscriptsubscript𝜋subscript𝛼𝑖1subscript𝛼𝑖1subscript𝛼𝑖1subscript𝛼𝑖subscript𝛿𝑗subscript𝛿𝑗1\displaystyle\prod_{i<t}\left(\pi_{\left\llbracket\alpha_{i}+1,\alpha_{i+1}\right\rrbracket}^{\alpha_{i+1}-\alpha_{i}}\right)^{\delta_{i}}=\prod_{i<t}\prod_{i\leq j<t}\left(\pi_{\left\llbracket\alpha_{i}+1,\alpha_{i+1}\right\rrbracket}^{\alpha_{i+1}-\alpha_{i}}\right)^{\delta_{j}-\delta_{j+1}}</td>
<td></td>
</tr>
<tr>
<td></td>
<td></td>
<td>=\displaystyle=</td>
<td>∏j<t(∏i≤jπ⟦αi+1,αi+1⟧αi+1−αi)δj−δj+1=∏j<t(π⟦1,αj+1⟧αj+1)δj−δj+1.subscriptproduct𝑗𝑡superscriptsubscriptproduct𝑖𝑗superscriptsubscript𝜋subscript𝛼𝑖1subscript𝛼𝑖1subscript𝛼𝑖1subscript𝛼𝑖subscript𝛿𝑗subscript𝛿𝑗1subscriptproduct𝑗𝑡superscriptsuperscriptsubscript𝜋1subscript𝛼𝑗1subscript𝛼𝑗1subscript𝛿𝑗subscript𝛿𝑗1\displaystyle\prod_{j<t}\left(\prod_{i\leq j}\pi_{\left\llbracket\alpha_{i}+1,\alpha_{i+1}\right\rrbracket}^{\alpha_{i+1}-\alpha_{i}}\right)^{\delta_{j}-\delta_{j+1}}=\prod_{j<t}\left(\pi_{\left\llbracket 1,\alpha_{j+1}\right\rrbracket}^{\alpha_{j+1}}\right)^{\delta_{j}-\delta_{j+1}}.</td>
<td></td>
</tr>
</tbody>
</table>
By using $`t - 1`$ times Minkowski’s theorem, we obtain that:
<table>
<tbody>
<tr>
<td></td>
<td>πI|I|superscriptsubscript𝜋𝐼𝐼\displaystyle\pi_{I}^{\left|I\right|}</td>
<td>≥\displaystyle\geq</td>
<td>d|I|log|I|d⋅(‖b→1‖/d)∑j<tαj+1(δj−δj+1)⋅superscript𝑑𝐼𝐼𝑑superscriptnormsubscript→𝑏1𝑑subscript𝑗𝑡subscript𝛼𝑗1subscript𝛿𝑗subscript𝛿𝑗1\displaystyle\sqrt{d}^{|I|\log\frac{|I|}{d}}\cdot(\|\vec{b}_{1}\|/\sqrt{d})^{\sum_{j<t}\alpha_{j+1}(\delta_{j}-\delta_{j+1})}</td>
<td></td>
</tr>
<tr>
<td></td>
<td></td>
<td>≥\displaystyle\geq</td>
<td>d|I|log|I|d⋅(‖b→1‖/d)∑j<t(αj+1−αj)δj⋅superscript𝑑𝐼𝐼𝑑superscriptnormsubscript→𝑏1𝑑subscript𝑗𝑡subscript𝛼𝑗1subscript𝛼𝑗subscript𝛿𝑗\displaystyle\sqrt{d}^{|I|\log\frac{|I|}{d}}\cdot(\|\vec{b}_{1}\|/\sqrt{d})^{\sum_{j<t}(\alpha_{j+1}-\alpha_{j})\delta_{j}}</td>
<td></td>
</tr>
<tr>
<td></td>
<td></td>
<td>≥\displaystyle\geq</td>
<td>d|I|(log|I|d−1)⋅‖b→1‖|I|.⋅superscript𝑑𝐼𝐼𝑑1superscriptnormsubscript→𝑏1𝐼\displaystyle\sqrt{d}^{|I|\left(\log\frac{|I|}{d}-1\right)}\cdot\|\vec{b}_{1}\|^{\left|I\right|}.</td>
<td></td>
</tr>
</tbody>
</table>
The final inequality of the theorem is just the fact that $`x\mapsto{x\hspace{0pt}{\log{({d/x})}}}`$ is maximal for $`x = {d/e}`$. $`\square`$
Note that if $`{\max I} < d`$, we can apply the result to the HKZ-reduced basis $`(b_{1},\ldots,b_{\max I})`$. In the case where $`I = {\{ i\}}`$, we recover the result of \[[23](#bib.bib23)\] that
<table>
<tbody>
<tr>
<td></td>
<td>‖bi∗‖≥(i)−logi−1⋅‖b1‖.normsuperscriptsubscript𝑏𝑖⋅superscript𝑖𝑖1normsubscript𝑏1\|b_{i}^{*}\|\geq(\sqrt{i})^{-\log i-1}\cdot\|b_{1}\|.</td>
<td></td>
<td>(2)</td>
</tr>
</tbody>
</table>
Still, our result is significantly better to what would have been obtained by combining several relations of the type of Equation ([2](#S4.E2 "In 4.3 A Property on the Geometry of HKZ-Reduced Bases ‣ 4 Complexity of the Enumeration Procedure")), when $`|I|`$ grows large. For instance, for a worst case of our analysis where $`I`$ is roughly the interval $`\lbrack{d\hspace{0pt}{({1 - {1/e}})}},d\rbrack`$, this strategy would yield a lower bound of the form $`{\| b_{1}\|}^{d/e} \cdot {\sqrt{d}}^{{({d/e})}\hspace{0pt}{\log d}}`$, which is worse than Helfrich’s analysis.
## 5 CVP and Other Related Problems
In this section, we describe what can be obtained by adapting our technique to the Closest Vector Problem and other problems related to strong lattice reduction. We only describe the proofs at a high level, since they are relatively straightforward.
In CVP, we are given a basis $`({\overset{\rightarrow}{b}}_{1},\ldots,{\overset{\rightarrow}{b}}_{d})`$ and a target vector $`\overset{\rightarrow}{t}`$, and we look for a lattice vector that is closest to $`\overset{\rightarrow}{t}`$. The first step of Kannan’s CVP algorithm is to HKZ-reduce the $`{\overset{\rightarrow}{b}}_{i}`$’s. Then one adapts the enumeration algorithm of Figure [1](#S3.F1 "Figure 1 ‣ 3.1 Short Lattice Points Enumeration ‣ 3 Kannan’s SVP Algorithm") for CVP. For the sake of simplicity, we assume that $`\|{\overset{\rightarrow}{b}}_{1}^{\ast}\|`$ is the largest of the $`\|{\overset{\rightarrow}{b}}_{i}^{\ast}\|`$’s (we refer to Kannan’s proof \[[16](#bib.bib16)\] for the general case). By using Babai’s nearest hyperplane algorithm \[[6](#bib.bib6)\], we see that there is a lattice vector $`\overset{\rightarrow}{b}`$ at distance less than $`\sqrt{d} \cdot {\|{\overset{\rightarrow}{b}}_{1}\|}`$ of the target vector $`\overset{\rightarrow}{t}`$. As a consequence, if we take $`A = {d \cdot {\|{\overset{\rightarrow}{b}}_{1}\|}}`$ in the adaptation of the enumeration procedure, we are sure to find a solution. The analysis then reduces (at the level of Equation ([1](#S4.E1 "In 4.1 From the Enumeration Procedure to Integer Points in Hyper-ellipsoids ‣ 4 Complexity of the Enumeration Procedure"))) to bound the ratio $`\frac{{\|{\overset{\rightarrow}{b}}_{1}\|}^{d}}{\prod_{i \leq d}{\|{\overset{\rightarrow}{b}}_{i}^{\ast}\|}}`$, which can be done with Minkowski’s theorem.
###### Theorem 5.1
Given a basis $`({\overset{\rightarrow}{b}}_{1},\ldots,{\overset{\rightarrow}{b}}_{d})`$ and a target vector $`\overset{\rightarrow}{t}`$, all of them in $`{\mathbb{R}}^{n}`$ and with integer coordinates whose absolute values are smaller than some $`B`$, one can find all vectors in the lattice spanned by the $`{\overset{\rightarrow}{b}}_{i}`$’s that are closest to $`\overset{\rightarrow}{t}`$ in deterministic time $`{\mathcal{P}\hspace{0pt}{({\log B},n)}} \cdot d^{{d/2} + {o\hspace{0pt}{(d)}}}`$.
The best deterministic complexity bound previously known for this problem was $`{\mathcal{P}\hspace{0pt}{({\log B},n)}} \cdot d^{d + {o\hspace{0pt}{(d)}}}`$ (see \[[13](#bib.bib13), [7](#bib.bib7)\]). Our result can also be adapted to enumerating all vectors of a lattice that are of length below a prescribed bound, which is in particular useful in the context of computing lattice theta series.
Another important consequence of our analysis is a significant worst-case bound improvement of Schnorr’s block-based strategy \[[23](#bib.bib23)\] to compute relatively short vectors. More precisely, if we take the bounds given in \[[10](#bib.bib10)\] for the quality of Schnorr’s semi-$`2\hspace{0pt}k`$ reduction and for the transference reduction, we obtain the table of Figure [3](#S5.F3 "Figure 3 ‣ 5 CVP and Other Related Problems"). Each entry of the table gives the upper bound of the quantity $`\frac{\|{\overset{\rightarrow}{b}}_{1}\|}{{({\det L})}^{1/d}}`$ which is reachable for a computational effort of $`2^{t}`$, for $`t`$ growing to infinity. To sum up, the multiplicative exponent constant is divided by $`e \approx 2.7`$. The table upper bounds can be adapted to the quantity $`\frac{\|{\overset{\rightarrow}{b}}_{1}\|}{\lambda_{1}\hspace{0pt}{(L)}}`$ by squaring them.
<table>
<tbody>
<tr>
<td></td>
<td>Semi-2k reductionTransference reductionUsing Helfrich’s complexity bound∼< 2log22dlog2tt≈20.347dlog2tt∼< 214dlog2tt≈20.250dlog2ttUsing the improved complexity bound∼< 2log22edlog2tt≈20.128dlog2tt∼< 214edlog2tt≈20.092dlog2ttmissing-subexpressionmissing-subexpressionmissing-subexpressionmissing-subexpressionSemi-2k reductionTransference reductionmissing-subexpressionmissing-subexpressionmissing-subexpressionUsing Helfrich’s complexity boundsuperscriptsimilar-toabsentsuperscript222𝑑superscript2𝑡𝑡superscript20.347𝑑superscript2𝑡𝑡superscriptsimilar-toabsentsuperscript214𝑑superscript2𝑡𝑡superscript20.250𝑑superscript2𝑡𝑡missing-subexpressionmissing-subexpressionmissing-subexpressionUsing the improved complexity boundsuperscriptsimilar-toabsentsuperscript222𝑒𝑑superscript2𝑡𝑡superscript20.128𝑑superscript2𝑡𝑡superscriptsimilar-toabsentsuperscript214𝑒𝑑superscript2𝑡𝑡superscript20.092𝑑superscript2𝑡𝑡\begin{array}[]{|c|c|c|}\hline\cr&\mbox{Semi-$2k$ reduction}&\mbox{Transference reduction}\\ \hline\cr\mbox{Using Helfrich's complexity bound}&\ {\raise-2.15277pt\hbox{$\buildrel<\over{\sim}$}}\ 2^{\frac{\log 2}{2}\frac{d\log^{2}t}{t}}\approx 2^{0.347\frac{d\log^{2}t}{t}}&\ {\raise-2.15277pt\hbox{$\buildrel<\over{\sim}$}}\ 2^{\frac{1}{4}\frac{d\log^{2}t}{t}}\approx 2^{0.250\frac{d\log^{2}t}{t}}\\ \hline\cr\mbox{Using the improved complexity bound}&\ {\raise-2.15277pt\hbox{$\buildrel<\over{\sim}$}}\ 2^{\frac{\log 2}{2e}\frac{d\log^{2}t}{t}}\approx 2^{0.128\frac{d\log^{2}t}{t}}&\ {\raise-2.15277pt\hbox{$\buildrel<\over{\sim}$}}\ 2^{\frac{1}{4e}\frac{d\log^{2}t}{t}}\approx 2^{0.092\frac{d\log^{2}t}{t}}\\ \hline\cr\end{array}\vspace*{-.2cm}</td>
<td></td>
</tr>
</tbody>
</table>
Figure 3: Worst-case bounds for block-based reduction algorithms.
Let us finish by mentioning that work under progress seems to show, by using a technique due to Ajtai \[[3](#bib.bib3)\], that our analyses are sharp, in the sense that for all $`\varepsilon > 0`$, we can build HKZ-reduced bases for which the number of steps of Kannan’s algorithm would be of the order of $`d^{d\hspace{0pt}{({\frac{1}{2\hspace{0pt}e} - \varepsilon})}}`$.
## References
- \[1\] E. Agrell, T. Eriksson, A. Vardy, and K. Zeger. Closest point search in lattices.
IEEE Transactions on Information Theory, 48(8):2201–2214, 2002.
- \[2\] M. Ajtai. The shortest vector problem in $`l_{2}`$ is NP-hard for randomized reductions (extended abstract).
In Proceedings of the 30th Symposium on the Theory of Computing (STOC 1998), pages 284–293. ACM Press, 1998.
- \[3\] M. Ajtai. The worst-case behavior of Schnorr’s algorithm approximating the shortest nonzero vector in a lattice.
In Proceedings of the 35th Symposium on the Theory of Computing (STOC 2003), pages 396–406. ACM Press, 2003.
- \[4\] M. Ajtai and C. Dwork. A public-key cryptosystem with worst-case/average-case equivalence.
In Proceedings of the 29th Symposium on the Theory of Computing (STOC 1997), pages 284–293. ACM Press, 1997.
- \[5\] M. Ajtai, R. Kumar, and D. Sivakumar. A sieve algorithm for the shortest lattice vector problem.
In Proceedings of the 33rd Symposium on the Theory of Computing (STOC 2001), pages 601–610. ACM Press, 2001.
- \[6\] L. Babai. On Lovász lattice reduction and the nearest lattice point problem.
Combinatorica, 6:1–13, 1986.
- \[7\] J. Bömer. Closest vectors, successive minima and dual-HKZ bases of lattices.
In Proceedings of the 2000 International Colloquium on Automata, Languages and Programming (ICALP 2000), volume 1853 of Lecture Notes in Computer Science, pages 248–259. Springer-Verlag, 2000.
- \[8\] D. Coppersmith and A. Shamir. Lattice attacks on NTRU.
In Proceedings of Eurocrypt 1997, volume 1233 of Lecture Notes in Computer Science, pages 52–61. Springer-Verlag, 1997.
- \[9\] U. Fincke and M. Pohst. A procedure for determining algebraic integers of given norm.
In Proceedings of EUROCAL, volume 162 of Lecture Notes in Computer Science, pages 194–202, 1983.
- \[10\] N. Gama, N. Howgrave-Graham, H. Koy, and P. Nguyen. Rankin’s constant and blockwise lattice reduction.
In Proceedings of Crypto 2006, number 4117 in Lecture Notes in Computer Science, pages 112–130. Springer-Verlag, 2006.
- \[11\] O. Goldreich, S. Goldwasser, and S. Halevi. Public-key cryptosystems from lattice reduction problems.
In Proceedings of Crypto 1997, volume 1294 of Lecture Notes in Computer Science, pages 112–131. Springer-Verlag, 1997.
- \[12\] I. Haviv and O. Regev. Tensor-based hardness of the shortest vector problem to within almost polynomial factors. Submitted.
- \[13\] B. Helfrich. Algorithms to construct Minkowski reduced and Hermite reduced lattice bases.
Theoretical Computer Science, 41:125–139, 1985.
- \[14\] C. Hermite. Extraits de lettres de M. Hermite à M. Jacobi sur différents objets de la théorie des nombres, deuxième lettre.
Journal für die reine und angewandte Mathematik, 40:279–290, 1850.
- \[15\] J. Hoffstein, J. Pipher, and J. H. Silverman. NTRU : a ring based public key cryptosystem.
In Proceedings of the 3rd Algorithmic Number Theory Symposium (ANTS III), volume 1423 of Lecture Notes in Computer Science, pages 267–288. Springer-Verlag, 1998.
- \[16\] R. Kannan. Improved algorithms for integer programming and related lattice problems.
In Proceedings of the 15th Symposium on the Theory of Computing (STOC 1983), pages 99–108. ACM Press, 1983.
- \[17\] A. K. Lenstra, H. W. Lenstra, Jr., and L. Lovász. Factoring polynomials with rational coefficients.
Mathematische Annalen, 261:513–534, 1982.
- \[18\] J. Martinet.
Perfect Lattices in Euclidean Spaces.
Springer-Verlag, 2002.
- \[19\] J. Mazo and A. Odlyzko. Lattice points in high-dimensional spheres, 1990.
- \[20\] D. Micciancio and S. Goldwasser.
Complexity of lattice problems : a cryptographic perspective.
Kluwer Academic Press, 2002.
- \[21\] H. Minkowski.
Geometrie der Zahlen.
Teubner-Verlag, 1896.
- \[22\] O. Regev.
Lecture notes of lattices in computer science, taught at the Computer Science Tel Aviv University.
Available at http://www.cs.tau.il/~odedr.
- \[23\] C. P. Schnorr. A hierarchy of polynomial lattice basis reduction algorithms.
Theoretical Computer Science, 53:201–224, 1987.
- \[24\] C. P. Schnorr and M. Euchner. Lattice basis reduction : improved practical algorithms and solving subset sum problems.
Mathematics of Programming, 66:181–199, 1994.
- \[25\] V. Shoup. NTL, Number Theory C++ Library.
Available at http://www.shoup.net/ntl/.
## Proof of Lemma [2](#Thmlemma2 "Lemma 2 ‣ 4.3 A Property on the Geometry of HKZ-Reduced Bases ‣ 4 Complexity of the Enumeration Procedure")
We prove the result by induction on $`k`$.For $`k = 1`$, the bound easily follows from $`\gamma_{d} \leq {{({d + 4})}/4}`$. Suppose now that the result holds for some $`k \in \left. ⟦1,{d - 2} \right.⟧`$, and that we want to show that it holds for $`k + 1`$. Notice that we can suppose that $`d \geq 3`$. Define $`{G_{d}\hspace{0pt}{(k)}} = {\frac{1}{2}\hspace{0pt}{\log{d\hspace{0pt}{\log\frac{d}{d - k}}}}}`$. Then for any $`\lambda > 0`$,
<table>
<tbody>
<tr>
<td></td>
<td>Gd(k+λ)−Gd(k)=−12logdlogd−k−λd−k≥12λlogdd−k.subscript𝐺𝑑𝑘𝜆subscript𝐺𝑑𝑘12𝑑𝑑𝑘𝜆𝑑𝑘12𝜆𝑑𝑑𝑘G_{d}(k+\lambda)-G_{d}(k)=-\frac{1}{2}\log d\log\frac{d-k-\lambda}{d-k}\geq\frac{1}{2}\frac{\lambda\log d}{d-k}.</td>
<td></td>
</tr>
</tbody>
</table>
Taking $`\lambda = 1`$, we see that $`{{G_{d}\hspace{0pt}{({k + 1})}} - {G_{d}\hspace{0pt}{(k)}}} \geq {\frac{1}{2}\hspace{0pt}\frac{\log d}{d - k}}`$.
From the upper bound $`\gamma_{d} \leq {{({d + 4})}/4}`$, we obtain:
<table>
<tbody>
<tr>
<td></td>
<td>logΓd(k+1)−logΓd(k)=12logγd−kd−k−1≤12log(d−k+4)/4d−k−1.subscriptΓ𝑑𝑘1subscriptΓ𝑑𝑘12subscript𝛾𝑑𝑘𝑑𝑘112𝑑𝑘44𝑑𝑘1\log\Gamma_{d}(k+1)-\log\Gamma_{d}(k)=\frac{1}{2}\frac{\log\gamma_{d-k}}{d-k-1}\leq\frac{1}{2}\frac{\log(d-k+4)/4}{d-k-1}.</td>
<td></td>
</tr>
</tbody>
</table>
Now, since the sequence $`\left( \frac{n\hspace{0pt}{\log{({{({n + 4})}/4})}}}{n - 1} \right)_{n \geq 2}`$ is increasing, we have:
<table>
<tbody>
<tr>
<td></td>
<td>(d−k)log((d−k+4)/4)d−k−1𝑑𝑘𝑑𝑘44𝑑𝑘1\displaystyle\frac{(d-k)\log((d-k+4)/4)}{d-k-1}</td>
<td>≤\displaystyle\leq</td>
<td>d−1d−2log((d+3)/4)𝑑1𝑑2𝑑34\displaystyle\frac{d-1}{d-2}\log((d+3)/4)</td>
<td></td>
</tr>
<tr>
<td></td>
<td></td>
<td>=\displaystyle=</td>
<td>logd+(d−1)log((d+3)/4)−(d−2)logdd−2𝑑𝑑1𝑑34𝑑2𝑑𝑑2\displaystyle\log d+\frac{(d-1)\log((d+3)/4)-(d-2)\log d}{d-2}</td>
<td></td>
</tr>
<tr>
<td></td>
<td></td>
<td>≤\displaystyle\leq</td>
<td>logd,𝑑\displaystyle\log d,</td>
<td></td>
</tr>
</tbody>
</table>
since the last term is a decreasing function of $`d`$, which is negative for $`d = 3`$. $`\square`$
## Proof of Lemma [3](#Thmlemma3 "Lemma 3 ‣ 4.3 A Property on the Geometry of HKZ-Reduced Bases ‣ 4 Complexity of the Enumeration Procedure")
We have $`{- {\log\hspace{0pt}{\prod_{i \leq t}\delta^{- {k_{i}\hspace{0pt}{\log\frac{k_{i}}{d_{i}}}}}}}} = {{({\log\delta})} \cdot {\sum_{i \leq t}{k_{i}\hspace{0pt}{\log\frac{k_{i}}{d_{i}}}}}}`$. Now, note that the function $`x\mapsto{x\hspace{0pt}{\log x}}`$ is convex on $`\lbrack 0,{+ \infty})`$. This means that for any $`t \geq 1`$, for any $`{a_{1},\ldots,a_{t}} > 0`$, and for any $`{\lambda_{1},\ldots,\lambda_{t}} \in {\lbrack 0,1\rbrack}`$ such that $`{\sum_{i \leq t}\lambda_{i}} = 1`$, we have:
<table>
<tbody>
<tr>
<td></td>
<td>∑i≤tλiailogai≥(∑i≤tλiai)log(∑i≤tλiai).subscript𝑖𝑡subscript𝜆𝑖subscript𝑎𝑖subscript𝑎𝑖subscript𝑖𝑡subscript𝜆𝑖subscript𝑎𝑖subscript𝑖𝑡subscript𝜆𝑖subscript𝑎𝑖\sum_{i\leq t}\lambda_{i}a_{i}\log a_{i}\geq\left(\sum_{i\leq t}\lambda_{i}a_{i}\right)\log\left(\sum_{i\leq t}\lambda_{i}a_{i}\right).</td>
<td></td>
</tr>
</tbody>
</table>
In particular, for $`\lambda_{i}:=\frac{d_{i}}{\sum_{i \leq t}d_{i}}`$ and $`a_{i}:=\frac{k_{i}}{d_{i}}`$, we get (after multiplication by $`\sum_{i \leq t}d_{i}`$):
<table>
<tbody>
<tr>
<td></td>
<td>−log∏i≤tδ−kilogkidi≥(logδ)⋅(∑i≤tki)log(∑i≤tki∑i≤tdi),subscriptproduct𝑖𝑡superscript𝛿subscript𝑘𝑖subscript𝑘𝑖subscript𝑑𝑖⋅𝛿subscript𝑖𝑡subscript𝑘𝑖subscript𝑖𝑡subscript𝑘𝑖subscript𝑖𝑡subscript𝑑𝑖-\log\prod_{i\leq t}\delta^{-k_{i}\log\frac{k_{i}}{d_{i}}}\geq(\log\delta)\cdot\left(\sum_{i\leq t}k_{i}\right)\log\left(\frac{\sum_{i\leq t}k_{i}}{\sum_{i\leq t}d_{i}}\right),</td>
<td></td>
</tr>
</tbody>
</table>
which is exactly $`- {\log\delta^{- {{({\sum_{i \leq t}k_{i}})}\hspace{0pt}{\log\frac{\sum_{i \leq t}k_{i}}{\sum_{i \leq t}d_{i}}}}}}`$. $`\square`$
## Proof of Lemma [4](#Thmlemma4 "Lemma 4 ‣ 4.3 A Property on the Geometry of HKZ-Reduced Bases ‣ 4 Complexity of the Enumeration Procedure").
###### Proof
We start with the first identity. We prove it by induction on $`k`$. For $`k = 1`$, this is Minkowski’s bound. Assume it to be true for a given $`k \leq {d - 2}`$. We are to prove that it holds for $`k + 1`$ instead of $`k`$. By applying Minkowski’s bound to the $`({d - k})`$-dimensional HKZ-reduced basis $`{\overset{\rightarrow}{b}}_{k + 1}^{\ast},\ldots,{\overset{\rightarrow}{b}}_{d}^{\ast}`$, we have:
<table>
<tbody>
<tr>
<td></td>
<td>‖b→k+1∗‖≤γd−kd−kd−k−1⋅π⟦k+2,d⟧.normsuperscriptsubscript→𝑏𝑘1⋅superscriptsubscript𝛾𝑑𝑘𝑑𝑘𝑑𝑘1subscript𝜋𝑘2𝑑\|\vec{b}_{k+1}^{*}\|\leq\sqrt{\gamma_{d-k}}^{\frac{d-k}{d-k-1}}\cdot\pi_{\left\llbracket k+2,d\right\rrbracket}.</td>
<td></td>
<td>(3)</td>
</tr>
</tbody>
</table>
We can rewrite our induction hypothesis as
<table>
<tbody>
<tr>
<td></td>
<td>π⟦1,k+1⟧k+1k⋅‖b→k+1∗‖−1k≤(Γd(k))dk⋅π⟦k+2,d⟧d−k−1d−k⋅‖b→k+1∗‖1d−k,⋅superscriptsubscript𝜋1𝑘1𝑘1𝑘superscriptnormsuperscriptsubscript→𝑏𝑘11𝑘⋅superscriptsubscriptΓ𝑑𝑘𝑑𝑘superscriptsubscript𝜋𝑘2𝑑𝑑𝑘1𝑑𝑘superscriptnormsuperscriptsubscript→𝑏𝑘11𝑑𝑘\pi_{\left\llbracket 1,k+1\right\rrbracket}^{\frac{k+1}{k}}\cdot\|\vec{b}_{k+1}^{*}\|^{-\frac{1}{k}}\leq\left(\Gamma_{d}(k)\right)^{\frac{d}{k}}\cdot\pi_{\left\llbracket k+2,d\right\rrbracket}^{\frac{d-k-1}{d-k}}\cdot\|\vec{b}_{k+1}^{*}\|^{\frac{1}{d-k}},</td>
<td></td>
</tr>
</tbody>
</table>
or, again, as
<table>
<tbody>
<tr>
<td></td>
<td>π⟦1,k+1⟧k+1k≤(Γd(k))dk⋅π⟦k+2,d⟧d−k−1d−k⋅‖b→k+1∗‖dk(d−k).superscriptsubscript𝜋1𝑘1𝑘1𝑘⋅superscriptsubscriptΓ𝑑𝑘𝑑𝑘superscriptsubscript𝜋𝑘2𝑑𝑑𝑘1𝑑𝑘superscriptnormsuperscriptsubscript→𝑏𝑘1𝑑𝑘𝑑𝑘\pi_{\left\llbracket 1,k+1\right\rrbracket}^{\frac{k+1}{k}}\leq\left(\Gamma_{d}(k)\right)^{\frac{d}{k}}\cdot\pi_{\left\llbracket k+2,d\right\rrbracket}^{\frac{d-k-1}{d-k}}\cdot\|\vec{b}_{k+1}^{*}\|^{\frac{d}{k(d-k)}}.</td>
<td></td>
</tr>
</tbody>
</table>
This gives, by using Equation ([3](#Sx3.E3 "In Proof ‣ Proof of Lemma 4.")):
<table>
<tbody>
<tr>
<td></td>
<td>π⟦1,k+1⟧k+1k≤(Γd(k))dk⋅γd−kdk(d−k−1)⋅π⟦k+2,d⟧k+1k=(Γd(k+1))dk⋅π⟦k+2,d⟧(k+1)/k.superscriptsubscript𝜋1𝑘1𝑘1𝑘⋅superscriptsubscriptΓ𝑑𝑘𝑑𝑘superscriptsubscript𝛾𝑑𝑘𝑑𝑘𝑑𝑘1superscriptsubscript𝜋𝑘2𝑑𝑘1𝑘⋅superscriptsubscriptΓ𝑑𝑘1𝑑𝑘superscriptsubscript𝜋𝑘2𝑑𝑘1𝑘\pi_{\left\llbracket 1,k+1\right\rrbracket}^{\frac{k+1}{k}}\leq\left(\Gamma_{d}(k)\right)^{\frac{d}{k}}\cdot\sqrt{\gamma_{d-k}}^{\frac{d}{k(d-k-1)}}\cdot\pi_{\left\llbracket k+2,d\right\rrbracket}^{\frac{k+1}{k}}=\left(\Gamma_{d}(k+1)\right)^{\frac{d}{k}}\cdot\pi_{\left\llbracket k+2,d\right\rrbracket}^{(k+1)/k}.</td>
<td></td>
</tr>
</tbody>
</table>
By raising this last identity to the power $`\frac{k}{k + 1}`$, we get
<table>
<tbody>
<tr>
<td></td>
<td>π⟦1,k+1⟧≤(Γd(k+1))dk+1⋅π⟦k+2,d⟧,subscript𝜋1𝑘1⋅superscriptsubscriptΓ𝑑𝑘1𝑑𝑘1subscript𝜋𝑘2𝑑\pi_{\left\llbracket 1,k+1\right\rrbracket}\leq\left(\Gamma_{d}({k+1})\right)^{\frac{d}{k+1}}\cdot\pi_{\left\llbracket k+2,d\right\rrbracket},</td>
<td></td>
</tr>
</tbody>
</table>
which, by induction, yields the first inequality.
The second inequality follows easily from the first one. Indeed, it suffices to raise the first one to the power $`k/d`$, multiply both sides by $`\left( \pi_{⟦{k + 1},d⟧} \right)^{{({d - k})}/d}`$, and use the identity $`{\det L} = {\left( \pi_{⟦1,k⟧} \right)^{k} \cdot \left( \pi_{⟦{k + 1},d⟧} \right)^{d - k}}`$.
## Proof of Lemma [5](#Thmlemma5 "Lemma 5 ‣ 4.3 A Property on the Geometry of HKZ-Reduced Bases ‣ 4 Complexity of the Enumeration Procedure").
First notice that, as a consequence of Lemma [4](#Thmlemma4 "Lemma 4 ‣ 4.3 A Property on the Geometry of HKZ-Reduced Bases ‣ 4 Complexity of the Enumeration Procedure"), we have, for $`k,l`$ integers, $`1 \leq k \leq l < d`$,
<table>
<tbody>
<tr>
<td></td>
<td>π⟦l+1,d⟧≥Γd−k(l−k)−1⋅π⟦k+1,d⟧.subscript𝜋𝑙1𝑑⋅subscriptΓ𝑑𝑘superscript𝑙𝑘1subscript𝜋𝑘1𝑑\pi_{\left\llbracket l+1,d\right\rrbracket}\geq\Gamma_{d-k}(l-k)^{-1}\cdot\pi_{\left\llbracket k+1,d\right\rrbracket}.</td>
<td></td>
<td>(4)</td>
</tr>
</tbody>
</table>
Recall that:
<table>
<tbody>
<tr>
<td></td>
<td>π~[x1,d]=(π⟦⌊x1⌋,d⟧)λ1⋅(π⟦⌊x1⌋+1,d⟧)1−λ1andπ~[x2,d]=(π⟦⌊x2⌋,d⟧)λ2⋅(π⟦⌊x2⌋+1,d⟧)1−λ2,formulae-sequencesubscript~𝜋subscript𝑥1𝑑⋅superscriptsubscript𝜋subscript𝑥1𝑑subscript𝜆1superscriptsubscript𝜋subscript𝑥11𝑑1subscript𝜆1andsubscript~𝜋subscript𝑥2𝑑⋅superscriptsubscript𝜋subscript𝑥2𝑑subscript𝜆2superscriptsubscript𝜋subscript𝑥21𝑑1subscript𝜆2\tilde{\pi}_{[x_{1},d]}=\left(\pi_{\left\llbracket\lfloor x_{1}\rfloor,d\right\rrbracket}\right)^{\lambda_{1}}\cdot\left(\pi_{\left\llbracket\lfloor x_{1}\rfloor+1,d\right\rrbracket}\right)^{1-\lambda_{1}}\ \ \mbox{and}\ \ \tilde{\pi}_{[x_{2},d]}=\left(\pi_{\left\llbracket\lfloor x_{2}\rfloor,d\right\rrbracket}\right)^{\lambda_{2}}\cdot\left(\pi_{\left\llbracket\lfloor x_{2}\rfloor+1,d\right\rrbracket}\right)^{1-\lambda_{2}},</td>
<td></td>
</tr>
</tbody>
</table>
with $`\lambda_{i} = \frac{{({{d - {\lfloor x_{i}\rfloor}} + 1})}\hspace{0pt}{({{1 - x_{i}} + {\lfloor x_{i}\rfloor}})}}{{d - x_{i}} + 1}`$ for $`i \in {\{ 1,2\}}`$. Notice that since $`x_{1} < x_{2}`$, either $`{{\lfloor x_{1}\rfloor} + 1} \leq {\lfloor x_{2}\rfloor}`$, or $`{\lfloor x_{1}\rfloor} = {\lfloor x_{2}\rfloor}`$. In the last case, since the function $`x\mapsto{{({u - x})}/{({v - x})}}`$ is decreasing when $`u < v`$ and for $`x < u`$, we must have $`\lambda_{2} < \lambda_{1}`$.
We split the proof in several cases, depending on the respective values of $`\lambda_{1}`$ and $`\lambda_{2}`$.
First case: $`\lambda_{1} \leq \lambda_{2}`$. In that case, we have $`{{\lfloor x_{1}\rfloor} + 1} \leq {\lfloor x_{2}\rfloor}`$. We define
<table>
<tbody>
<tr>
<td></td>
<td>G:=Γd−⌊x1⌋+1(⌊x2⌋−⌊x1⌋)λ1⋅Γd−⌊x1⌋(⌊x2⌋−⌊x1⌋−1)λ2−λ1⋅Γd−⌊x1⌋(⌊x2⌋−⌊x1⌋)1−λ2.assign𝐺⋅⋅subscriptΓ𝑑subscript𝑥11superscriptsubscript𝑥2subscript𝑥1subscript𝜆1subscriptΓ𝑑subscript𝑥1superscriptsubscript𝑥2subscript𝑥11subscript𝜆2subscript𝜆1subscriptΓ𝑑subscript𝑥1superscriptsubscript𝑥2subscript𝑥11subscript𝜆2G:=\Gamma_{d-\lfloor x_{1}\rfloor+1}(\lfloor x_{2}\rfloor-\lfloor x_{1}\rfloor)^{\lambda_{1}}\cdot\Gamma_{d-\lfloor x_{1}\rfloor}(\lfloor x_{2}\rfloor-\lfloor x_{1}\rfloor-1)^{\lambda_{2}-\lambda_{1}}\cdot\Gamma_{d-\lfloor x_{1}\rfloor}(\lfloor x_{2}\rfloor-\lfloor x_{1}\rfloor)^{1-\lambda_{2}}.</td>
<td></td>
</tr>
</tbody>
</table>
By using three times Equation ([4](#Sx4.E4 "In Proof of Lemma 5.")), we get:
<table>
<tbody>
<tr>
<td></td>
<td>π~[x2,d]subscript~𝜋subscript𝑥2𝑑\displaystyle\tilde{\pi}_{[x_{2},d]}</td>
<td>=\displaystyle=</td>
<td>(π⟦⌊x2⌋,d⟧)λ2⋅(π⟦⌊x2⌋+1,d⟧)1−λ2⋅superscriptsubscript𝜋subscript𝑥2𝑑subscript𝜆2superscriptsubscript𝜋subscript𝑥21𝑑1subscript𝜆2\displaystyle\left(\pi_{\left\llbracket\lfloor x_{2}\rfloor,d\right\rrbracket}\right)^{\lambda_{2}}\cdot\left(\pi_{\left\llbracket\lfloor x_{2}\rfloor+1,d\right\rrbracket}\right)^{1-\lambda_{2}}</td>
<td></td>
</tr>
<tr>
<td></td>
<td></td>
<td>≥\displaystyle\geq</td>
<td>(π⟦⌊x2⌋,d⟧)λ1⋅(π⟦⌊x2⌋,d⟧)λ2−λ1⋅(π⟦⌊x2⌋+1,d⟧)1−λ2⋅superscriptsubscript𝜋subscript𝑥2𝑑subscript𝜆1superscriptsubscript𝜋subscript𝑥2𝑑subscript𝜆2subscript𝜆1superscriptsubscript𝜋subscript𝑥21𝑑1subscript𝜆2\displaystyle\left(\pi_{\left\llbracket\lfloor x_{2}\rfloor,d\right\rrbracket}\right)^{\lambda_{1}}\cdot\left(\pi_{\left\llbracket\lfloor x_{2}\rfloor,d\right\rrbracket}\right)^{\lambda_{2}-\lambda_{1}}\cdot\left(\pi_{\left\llbracket\lfloor x_{2}\rfloor+1,d\right\rrbracket}\right)^{1-\lambda_{2}}</td>
<td></td>
</tr>
<tr>
<td></td>
<td></td>
<td>≥\displaystyle\geq</td>
<td>G−1⋅(π⟦⌊x1⌋,d⟧)λ1⋅(π⟦⌊x1⌋+1,d⟧)1−λ1.⋅superscript𝐺1superscriptsubscript𝜋subscript𝑥1𝑑subscript𝜆1superscriptsubscript𝜋subscript𝑥11𝑑1subscript𝜆1\displaystyle G^{-1}\cdot\left(\pi_{\left\llbracket\lfloor x_{1}\rfloor,d\right\rrbracket}\right)^{\lambda_{1}}\cdot\left(\pi_{\left\llbracket\lfloor x_{1}\rfloor+1,d\right\rrbracket}\right)^{1-\lambda_{1}}.</td>
<td></td>
</tr>
</tbody>
</table>
Now, Lemma [4](#Thmlemma4 "Lemma 4 ‣ 4.3 A Property on the Geometry of HKZ-Reduced Bases ‣ 4 Complexity of the Enumeration Procedure") gives that
<table>
<tbody>
<tr>
<td></td>
<td>logGlogd≤λ1logd−⌊x1⌋+1d−⌊x2⌋+1+(λ2−λ1)logd−⌊x1⌋d−⌊x2⌋+1+(1−λ2)logd−⌊x1⌋d−⌊x2⌋,𝐺𝑑subscript𝜆1𝑑subscript𝑥11𝑑subscript𝑥21subscript𝜆2subscript𝜆1𝑑subscript𝑥1𝑑subscript𝑥211subscript𝜆2𝑑subscript𝑥1𝑑subscript𝑥2\frac{\log G}{\log\sqrt{d}}\leq\lambda_{1}\log\frac{d-\lfloor x_{1}\rfloor+1}{d-\lfloor x_{2}\rfloor+1}+(\lambda_{2}-\lambda_{1})\log\frac{d-\lfloor x_{1}\rfloor}{d-\lfloor x_{2}\rfloor+1}+(1-\lambda_{2})\log\frac{d-\lfloor x_{1}\rfloor}{d-\lfloor x_{2}\rfloor},</td>
<td></td>
</tr>
</tbody>
</table>
which, by concavity of the function $`x\mapsto{\log x}`$, is at most the logarithm of
<table>
<tbody>
<tr>
<td></td>
<td>E(x1,x2):=λ1d−⌊x1⌋+1d−⌊x2⌋+1+(λ2−λ1)d−⌊x1⌋d−⌊x2⌋+1+(1−λ2)d−⌊x1⌋d−⌊x2⌋assign𝐸subscript𝑥1subscript𝑥2subscript𝜆1𝑑subscript𝑥11𝑑subscript𝑥21subscript𝜆2subscript𝜆1𝑑subscript𝑥1𝑑subscript𝑥211subscript𝜆2𝑑subscript𝑥1𝑑subscript𝑥2E(x_{1},x_{2}):=\lambda_{1}\frac{d-\lfloor x_{1}\rfloor+1}{d-\lfloor x_{2}\rfloor+1}+(\lambda_{2}-\lambda_{1})\frac{d-\lfloor x_{1}\rfloor}{d-\lfloor x_{2}\rfloor+1}+(1-\lambda_{2})\frac{d-\lfloor x_{1}\rfloor}{d-\lfloor x_{2}\rfloor}</td>
<td></td>
</tr>
</tbody>
</table>
To complete the proof of this first case, it suffices to prove that $`{E\hspace{0pt}{(x_{1},x_{2})}} \leq {{f\hspace{0pt}r\hspace{0pt}a\hspace{0pt}c\hspace{0pt}d} - {x_{1}\hspace{0pt}d} - x_{2}}`$. We have
<table>
<tbody>
<tr>
<td></td>
<td>E(x1,x2)𝐸subscript𝑥1subscript𝑥2\displaystyle E(x_{1},x_{2})</td>
<td>=\displaystyle=</td>
<td>λ1d−⌊x2⌋+1+d−⌊x1⌋d−x2+1subscript𝜆1𝑑subscript𝑥21𝑑subscript𝑥1𝑑subscript𝑥21\displaystyle\frac{\lambda_{1}}{d-\lfloor x_{2}\rfloor+1}+\frac{d-\lfloor x_{1}\rfloor}{d-x_{2}+1}</td>
<td></td>
</tr>
<tr>
<td></td>
<td></td>
<td>=\displaystyle=</td>
<td>d−x1d−x2+λ1d−⌊x2⌋+1−1−frac(x1)d−x2+1−x2−x1(d−x2)(d−x2+1),𝑑subscript𝑥1𝑑subscript𝑥2subscript𝜆1𝑑subscript𝑥211fracsubscript𝑥1𝑑subscript𝑥21subscript𝑥2subscript𝑥1𝑑subscript𝑥2𝑑subscript𝑥21\displaystyle\frac{d-x_{1}}{d-x_{2}}+\frac{\lambda_{1}}{d-\lfloor x_{2}\rfloor+1}-\frac{1-\mbox{frac}(x_{1})}{d-x_{2}+1}-\frac{x_{2}-x_{1}}{(d-x_{2})(d-x_{2}+1)},</td>
<td></td>
</tr>
<tr>
<td></td>
<td></td>
<td>≤\displaystyle\leq</td>
<td>d−x1d−x2+1d−x2+1(λ1−(1−frac(x1))−x2−x1d−x2)𝑑subscript𝑥1𝑑subscript𝑥21𝑑subscript𝑥21subscript𝜆11fracsubscript𝑥1subscript𝑥2subscript𝑥1𝑑subscript𝑥2\displaystyle\frac{d-x_{1}}{d-x_{2}}+\frac{1}{d-x_{2}+1}\left(\lambda_{1}-(1-\mbox{frac}(x_{1}))-\frac{x_{2}-x_{1}}{d-x_{2}}\right)</td>
<td></td>
</tr>
<tr>
<td></td>
<td></td>
<td>=\displaystyle=</td>
<td>d−x1d−x2+1d−x2+1((1−frac(x1))frac(x1)d−x1+1−x2−x1d−x2)𝑑subscript𝑥1𝑑subscript𝑥21𝑑subscript𝑥211fracsubscript𝑥1fracsubscript𝑥1𝑑subscript𝑥11subscript𝑥2subscript𝑥1𝑑subscript𝑥2\displaystyle\frac{d-x_{1}}{d-x_{2}}+\frac{1}{d-x_{2}+1}\left(\frac{(1-\mbox{frac}(x_{1}))\mbox{frac}(x_{1})}{d-x_{1}+1}-\frac{x_{2}-x_{1}}{d-x_{2}}\right)</td>
<td></td>
</tr>
<tr>
<td></td>
<td></td>
<td>≤\displaystyle\leq</td>
<td>d−x1d−x2+1d−x2+1(1−frac(x1)d−x2−x2−x1d−x2),𝑑subscript𝑥1𝑑subscript𝑥21𝑑subscript𝑥211fracsubscript𝑥1𝑑subscript𝑥2subscript𝑥2subscript𝑥1𝑑subscript𝑥2\displaystyle\frac{d-x_{1}}{d-x_{2}}+\frac{1}{d-x_{2}+1}\left(\frac{1-\mbox{frac}(x_{1})}{d-x_{2}}-\frac{x_{2}-x_{1}}{d-x_{2}}\right),</td>
<td></td>
</tr>
</tbody>
</table>
from which the result follows at once, since $`{\lfloor x_{1}\rfloor} < {\lfloor x_{2}\rfloor}`$ implies that $`{x_{2} - x_{1}} = {{{{\lfloor x_{2}\rfloor} - {\lfloor x_{1}\rfloor}} + {\text{frac}\hspace{0pt}{(x_{2})}}} - {\text{frac}\hspace{0pt}{(x_{1})}}} \geq {1 - {\text{frac}\hspace{0pt}{(x_{1})}}}`$.
Second case: $`\lambda_{1} > \lambda_{2}`$. Similarly, defining
<table>
<tbody>
<tr>
<td></td>
<td>H=Γd−⌊x1⌋+1(⌊x2⌋−⌊x1⌋)λ2⋅Γd−⌊x1⌋+1(⌊x2⌋−⌊x1⌋+1)λ1−λ2⋅Γd−⌊x1⌋(⌊x2⌋−⌊x1⌋)1−λ1,𝐻⋅⋅subscriptΓ𝑑subscript𝑥11superscriptsubscript𝑥2subscript𝑥1subscript𝜆2subscriptΓ𝑑subscript𝑥11superscriptsubscript𝑥2subscript𝑥11subscript𝜆1subscript𝜆2subscriptΓ𝑑subscript𝑥1superscriptsubscript𝑥2subscript𝑥11subscript𝜆1H=\Gamma_{d-\lfloor x_{1}\rfloor+1}(\lfloor x_{2}\rfloor-\lfloor x_{1}\rfloor)^{\lambda_{2}}\cdot\Gamma_{d-\lfloor x_{1}\rfloor+1}(\lfloor x_{2}\rfloor-\lfloor x_{1}\rfloor+1)^{\lambda_{1}-\lambda_{2}}\cdot\Gamma_{d-\lfloor x_{1}\rfloor}(\lfloor x_{2}\rfloor-\lfloor x_{1}\rfloor)^{1-\lambda_{1}},</td>
<td></td>
</tr>
</tbody>
</table>
we obtain
<table>
<tbody>
<tr>
<td></td>
<td>π~[x2,d]subscript~𝜋subscript𝑥2𝑑\displaystyle\tilde{\pi}_{[x_{2},d]}</td>
<td>=\displaystyle=</td>
<td>(π⟦⌊x2⌋,d⟧)λ2⋅(π⟦⌊x2⌋+1,d⟧)1−λ2⋅superscriptsubscript𝜋subscript𝑥2𝑑subscript𝜆2superscriptsubscript𝜋subscript𝑥21𝑑1subscript𝜆2\displaystyle\left(\pi_{\left\llbracket\lfloor x_{2}\rfloor,d\right\rrbracket}\right)^{\lambda_{2}}\cdot\left(\pi_{\left\llbracket{\lfloor x_{2}\rfloor}+1,d\right\rrbracket}\right)^{1-\lambda_{2}}</td>
<td></td>
</tr>
<tr>
<td></td>
<td></td>
<td>=\displaystyle=</td>
<td>(π⟦⌊x2⌋,d⟧)λ2(π⟦⌊x2⌋+1,d⟧)λ1−λ2(π⟦⌊x2⌋+1,d⟧)1−λ1superscriptsubscript𝜋subscript𝑥2𝑑subscript𝜆2superscriptsubscript𝜋subscript𝑥21𝑑subscript𝜆1subscript𝜆2superscriptsubscript𝜋subscript𝑥21𝑑1subscript𝜆1\displaystyle\left(\pi_{\left\llbracket\lfloor x_{2}\rfloor,d\right\rrbracket}\right)^{\lambda_{2}}\left(\pi_{\left\llbracket\lfloor x_{2}\rfloor+1,d\right\rrbracket}\right)^{\lambda_{1}-\lambda_{2}}\left(\pi_{\left\llbracket\lfloor x_{2}\rfloor+1,d\right\rrbracket}\right)^{1-\lambda_{1}}</td>
<td></td>
</tr>
<tr>
<td></td>
<td></td>
<td>≥\displaystyle\geq</td>
<td>H−1(pi⟦⌊x1⌋,d⟧)λ1(π⟦⌊x1⌋+1,d⟧)1−λ1.superscript𝐻1superscript𝑝subscript𝑖subscript𝑥1𝑑subscript𝜆1superscriptsubscript𝜋subscript𝑥11𝑑1subscript𝜆1\displaystyle H^{-1}\left(pi_{\left\llbracket\lfloor x_{1}\rfloor,d\right\rrbracket}\right)^{\lambda_{1}}\left(\pi_{\left\llbracket\lfloor x_{1}\rfloor+1,d\right\rrbracket}\right)^{1-\lambda_{1}}.</td>
<td></td>
</tr>
</tbody>
</table>
Lemma [4](#Thmlemma4 "Lemma 4 ‣ 4.3 A Property on the Geometry of HKZ-Reduced Bases ‣ 4 Complexity of the Enumeration Procedure") gives us that:
<table>
<tbody>
<tr>
<td></td>
<td>logHlogd≤λ2logd−⌊x1⌋+1d−⌊x2⌋+1+(λ1−λ2)logd−⌊x1⌋+1d−⌊x2⌋+(1−λ1)logd−⌊x1⌋d−⌊x2⌋.𝐻𝑑subscript𝜆2𝑑subscript𝑥11𝑑subscript𝑥21subscript𝜆1subscript𝜆2𝑑subscript𝑥11𝑑subscript𝑥21subscript𝜆1𝑑subscript𝑥1𝑑subscript𝑥2\frac{\log H}{\log\sqrt{d}}\leq\lambda_{2}\log\frac{d-\lfloor x_{1}\rfloor+1}{d-\lfloor x_{2}\rfloor+1}+(\lambda_{1}-\lambda_{2})\log\frac{d-\lfloor x_{1}\rfloor+1}{d-\lfloor x_{2}\rfloor}+(1-\lambda_{1})\log\frac{d-\lfloor x_{1}\rfloor}{d-\lfloor x_{2}\rfloor}.</td>
<td></td>
</tr>
</tbody>
</table>
By concavity of the function $`x\mapsto{\log x}`$, the right hand side is at most the logarithm of
<table>
<tbody>
<tr>
<td></td>
<td>λ2d−⌊x1⌋+1d−⌊x2⌋+1+(λ1−λ2)d−⌊x1⌋+1d−⌊x2⌋subscript𝜆2𝑑subscript𝑥11𝑑subscript𝑥21subscript𝜆1subscript𝜆2𝑑subscript𝑥11𝑑subscript𝑥2\displaystyle\lambda_{2}\frac{d-\lfloor x_{1}\rfloor+1}{d-\lfloor x_{2}\rfloor+1}+(\lambda_{1}-\lambda_{2})\frac{d-\lfloor x_{1}\rfloor+1}{d-\lfloor x_{2}\rfloor}</td>
<td>+\displaystyle+</td>
<td>(1−λ1)d−⌊x1⌋d−⌊x2⌋1subscript𝜆1𝑑subscript𝑥1𝑑subscript𝑥2\displaystyle(1-\lambda_{1})\frac{d-\lfloor x_{1}\rfloor}{d-\lfloor x_{2}\rfloor}</td>
<td></td>
</tr>
<tr>
<td></td>
<td></td>
<td>=\displaystyle=</td>
<td>E(x1,x2)+λ1−λ2(d−⌊x2⌋)(d−⌊x2⌋+1).𝐸subscript𝑥1subscript𝑥2subscript𝜆1subscript𝜆2𝑑subscript𝑥2𝑑subscript𝑥21\displaystyle E(x_{1},x_{2})+\frac{\lambda_{1}-\lambda_{2}}{(d-\lfloor x_{2}\rfloor)(d-\lfloor x_{2}\rfloor+1)}.</td>
<td></td>
</tr>
</tbody>
</table>
Hence, we just need to prove that:
<table>
<tbody>
<tr>
<td></td>
<td>E′(x1,x2):=E(x1,x2)+(λ1−λ2)(d−⌊x2⌋)(d−⌊x2⌋+1)≤d−x1d−x2.assignsuperscript𝐸′subscript𝑥1subscript𝑥2𝐸subscript𝑥1subscript𝑥2subscript𝜆1subscript𝜆2𝑑subscript𝑥2𝑑subscript𝑥21𝑑subscript𝑥1𝑑subscript𝑥2E^{\prime}(x_{1},x_{2}):=E(x_{1},x_{2})+\frac{(\lambda_{1}-\lambda_{2})}{(d-\lfloor x_{2}\rfloor)(d-\lfloor x_{2}\rfloor+1)}\leq\frac{d-x_{1}}{d-x_{2}}.</td>
<td></td>
</tr>
</tbody>
</table>
Some elementary calculus provides the equalities:
<table>
<tbody>
<tr>
<td></td>
<td>E′(x1,x2)superscript𝐸′subscript𝑥1subscript𝑥2\displaystyle E^{\prime}(x_{1},x_{2})</td>
<td>=\displaystyle=</td>
<td>d−x1d−x2+λ1d−⌊x2⌋−1−frac(x2)(d−⌊x2⌋)(d−x2+1)−1−frac(x1)d−x2+1−x2−x1(d−x2)(d−x2+1)𝑑subscript𝑥1𝑑subscript𝑥2subscript𝜆1𝑑subscript𝑥21fracsubscript𝑥2𝑑subscript𝑥2𝑑subscript𝑥211fracsubscript𝑥1𝑑subscript𝑥21subscript𝑥2subscript𝑥1𝑑subscript𝑥2𝑑subscript𝑥21\displaystyle\frac{d-x_{1}}{d-x_{2}}+\frac{\lambda_{1}}{d-\lfloor x_{2}\rfloor}-\frac{1-\mbox{frac}(x_{2})}{(d-\lfloor x_{2}\rfloor)(d-x_{2}+1)}-\frac{1-\mbox{frac}(x_{1})}{d-x_{2}+1}-\frac{x_{2}-x_{1}}{(d-x_{2})(d-x_{2}+1)}</td>
<td></td>
</tr>
<tr>
<td></td>
<td></td>
<td>=\displaystyle=</td>
<td>d−x1d−x2+λ1d−⌊x2⌋−1−frac(x1)d−x2−1−frac(x2)(d−⌊x2⌋)(d−x2+1)−x2−⌊x1⌋−1(d−x2)(d−x2+1)𝑑subscript𝑥1𝑑subscript𝑥2subscript𝜆1𝑑subscript𝑥21fracsubscript𝑥1𝑑subscript𝑥21fracsubscript𝑥2𝑑subscript𝑥2𝑑subscript𝑥21subscript𝑥2subscript𝑥11𝑑subscript𝑥2𝑑subscript𝑥21\displaystyle\frac{d-x_{1}}{d-x_{2}}+\frac{\lambda_{1}}{d-\lfloor x_{2}\rfloor}-\frac{1-\mbox{frac}(x_{1})}{d-x_{2}}-\frac{1-\mbox{frac}(x_{2})}{(d-\lfloor x_{2}\rfloor)(d-x_{2}+1)}-\frac{x_{2}-\lfloor x_{1}\rfloor-1}{(d-x_{2})(d-x_{2}+1)}</td>
<td></td>
</tr>
</tbody>
</table>
Second case, first sub-case: $`\lambda_{1} > \lambda_{2}`$, $`{\lfloor x_{1}\rfloor} < {\lfloor x_{2}\rfloor}`$. In that case,
<table>
<tbody>
<tr>
<td></td>
<td>E′(x1,x2)−d−x1d−x2superscript𝐸′subscript𝑥1subscript𝑥2𝑑subscript𝑥1𝑑subscript𝑥2\displaystyle E^{\prime}(x_{1},x_{2})-\frac{d-x_{1}}{d-x_{2}}</td>
<td>≤\displaystyle\leq</td>
<td>λ1−(1−frac(x1))d−x2−1−frac(x2)(d−⌊x2⌋)(d−x2+1)−1(d−x2)(d−x2+1)subscript𝜆11fracsubscript𝑥1𝑑subscript𝑥21fracsubscript𝑥2𝑑subscript𝑥2𝑑subscript𝑥211𝑑subscript𝑥2𝑑subscript𝑥21\displaystyle\frac{\lambda_{1}-(1-\mbox{frac}(x_{1}))}{d-x_{2}}-\frac{1-\mbox{frac}(x_{2})}{(d-\lfloor x_{2}\rfloor)(d-x_{2}+1)}-\frac{1}{(d-x_{2})(d-x_{2}+1)}</td>
<td></td>
</tr>
<tr>
<td></td>
<td></td>
<td>≤\displaystyle\leq</td>
<td>1−frac(x1)(d−x2)(d−x1+1)−1(d−x2)(d−x2+1)1fracsubscript𝑥1𝑑subscript𝑥2𝑑subscript𝑥111𝑑subscript𝑥2𝑑subscript𝑥21\displaystyle\frac{1-\mbox{frac}(x_{1})}{(d-x_{2})(d-x_{1}+1)}-\frac{1}{(d-x_{2})(d-x_{2}+1)}</td>
<td></td>
</tr>
<tr>
<td></td>
<td></td>
<td>≤\displaystyle\leq</td>
<td>00\displaystyle 0</td>
<td></td>
</tr>
</tbody>
</table>
Second case, second sub-case: $`\lambda_{1} > \lambda_{2}`$, $`{\lfloor x_{1}\rfloor} = {\lfloor x_{2}\rfloor}`$. In that case, after some rewriting which can be checked with one’s favourite computer algebra system, one finds that:
<table>
<tbody>
<tr>
<td></td>
<td>E′(x1,x2)−d−x1d−x2superscript𝐸′subscript𝑥1subscript𝑥2𝑑subscript𝑥1𝑑subscript𝑥2\displaystyle E^{\prime}(x_{1},x_{2})-\frac{d-x_{1}}{d-x_{2}}</td>
<td>=\displaystyle=</td>
<td>1(d−⌊x1⌋)(d−x2)((1−frac(x1))(x1−x2)(d−⌊x2⌋)d−x1+1−frac(x2)(λ1−λ2)d−⌊x1⌋+1)1𝑑subscript𝑥1𝑑subscript𝑥21fracsubscript𝑥1subscript𝑥1subscript𝑥2𝑑subscript𝑥2𝑑subscript𝑥11fracsubscript𝑥2subscript𝜆1subscript𝜆2𝑑subscript𝑥11\displaystyle\frac{1}{(d-\lfloor x_{1}\rfloor)(d-x_{2})}\left(\frac{(1-\mbox{frac}(x_{1}))(x_{1}-x_{2})(d-\lfloor x_{2}\rfloor)}{d-x_{1}+1}-\frac{\mbox{frac}(x_{2})(\lambda_{1}-\lambda_{2})}{d-\lfloor x_{1}\rfloor+1}\right)</td>
<td></td>
</tr>
<tr>
<td></td>
<td></td>
<td>≤\displaystyle\leq</td>
<td>0.0\displaystyle 0.</td>
<td></td>
</tr>
</tbody>
</table>
$`\square`$ |
0704.3746 | cs.NI | cyber | # Distributed Algorithms for Spectrum Allocation, Power Control, Routing, and Congestion Control in Wireless Networks
Yufang Xi and Edmund M. Yeh
Department of Electrical Engineering
Yale University
New Haven, CT 06520, USA
Email: {yufang.xi, edmund.yeh}@yale.edu
###### Abstract
We develop distributed algorithms to allocate resources in multi-hop wireless networks with the aim of minimizing total cost. In order to observe the fundamental duplexing constraint that co-located transmitters and receivers cannot operate simultaneously on the same frequency band, we first devise a spectrum allocation scheme that divides the whole spectrum into multiple sub-bands and activates conflict-free links on each sub-band. We show that the minimum number of required sub-bands grows asymptotically at a logarithmic rate with the chromatic number of network connectivity graph. A simple distributed and asynchronous algorithm is developed to feasibly activate links on the available sub-bands. Given a feasible spectrum allocation, we then design node-based distributed algorithms for optimally controlling the transmission powers on active links for each sub-band, jointly with traffic routes and user input rates in response to channel states and traffic demands. We show that under specified conditions, the algorithms asymptotically converge to the optimal operating point.
<sup>1</sup>
<sup>1</sup>footnotetext: This research is supported in part by NSF grant CNS-0626882 and AFOSR grant FA9550-06-1-0135.
## I Introduction
While offering the potential for ubiquitous and untethered communications, wireless networks typically demand more sophisticated resource management than wireline networks. Optimal resource allocation in large-scale wireless networks involves joint optimization across multiple layers as well as distributed implementation across network nodes. In this paper, we develop distributed algorithms which jointly allocate frequency spectrum, transmission powers, traffic input rates, and traffic routes on a node-by-node basis in order to minimize total cost in an interference-limited multi-hop wireless network.
While joint optimization involving power control, congestion control, and routing has been studied previously \[[1](#bib.bib1), [2](#bib.bib2), [3](#bib.bib3)\], a common shortcoming of the previous work is the failure to account for the constraint that a node cannot transmit and receive simultaneously on the same frequency band. For while some types of multi-user interference can be ameliorated by using advanced coding techniques, interference between a transmitter and a co-located receiver is very difficult to suppress due to the transmitted power being many orders of magnitude higher than the received power at the same node.
<sup>2</sup>
<sup>2</sup>2Theoretically a node is able to subtract the transmission signals generated by itself from the received signals so that the self-interference can be perfectly cancelled out. In real decoders, however, the received signals are practically irrecoverable in the face of the overwhelming transmitted signals.
The resulting constraint that a transmitter cannot be simultaneously active with a co-located receiver \[[4](#bib.bib4)\] on the same frequency band is referred to as the duplexing constraint. In practical wireless networks, the duplexing constraint appears to be quite fundamental, and thus must be observed by resource management schemes.
In this paper, we develop distributed resource allocation algorithms for wireless networks in accordance with the duplexing constraint. To accomplish this, we first devise a new spectrum allocation scheme which divides the spectrum into a sufficient number of frequency bands and activates co-located transmitters and receivers on different bands. We show that the minimum number of sub-bands needed for resolving duplexing conflicts is asymptotically logarithmic in the chromatic number
<sup>3</sup>
<sup>3</sup>3The chromatic number of a graph is the minimum number of colors with which the vertices of the graph can be colored such that adjacent vertices have different colors.
of the network connectivity graph. We provide a simple algorithm that feasibly assigns frequency bands to links in a distributed and asynchronous manner with low control overhead.
Given a conflict-free spectrum allocation, we then design a set of node-based distributed gradient projection algorithms that iteratively adjust transmission powers, traffic input rates, and traffic routes according to channel conditions and traffic demands, in order to minimize total network cost. The power control and routing algorithms we develop are frequency selective, in that for each link, the power control algorithm adjusts the transmission power on each of the link’s active sub-bands. The routing algorithm involves both inter-node routing, which specifies the allocation of incoming traffic at each node to its outgoing links, and intra-node routing, which specifies the allocation of the total traffic on a given link across its active sub-bands. Finally, we show that congestion control can be naturally incorporated by considering an equivalent routing problem on a virtual overflow link. We show that under specified conditions, the iterative algorithms converge to the optimal operating point from any initial condition.
The rest of the paper is organized as follows. In Section [II](#S2 "II Network Model and Spectrum Allocation ‣ Distributed Algorithms for Spectrum Allocation, Power Control, Routing, and Congestion Control in Wireless Networks"), we introduce the network model, discuss the duplexing constraint, and formulate the spectrum allocation problem. In Section [III](#S3 "III Spectrum Division ‣ Distributed Algorithms for Spectrum Allocation, Power Control, Routing, and Congestion Control in Wireless Networks"), we find the minimum number of sub-bands required by a feasible spectrum allocation. A distributed and asynchronous sub-band allocation algorithm is developed in Section [IV](#S4 "IV Distributed Sub-band Allocation ‣ Distributed Algorithms for Spectrum Allocation, Power Control, Routing, and Congestion Control in Wireless Networks"). In Section [V](#S5 "V Cross-layer Optimization in Multi-radio Multi-hop Wireless Networks ‣ Distributed Algorithms for Spectrum Allocation, Power Control, Routing, and Congestion Control in Wireless Networks"), we formulate the cross-layer optimization problem for networks operating on multiple sub-bands, and derive the conditions satisfied by the optimal configuration. In Section [VI](#S6 "VI Node-Based Multi-Radio Power Control and Routing Algorithms ‣ Distributed Algorithms for Spectrum Allocation, Power Control, Routing, and Congestion Control in Wireless Networks"), we present node-based gradient projection algorithms to jointly optimize power control, routing, and congestion control based on the outcome of the spectrum allocation, as well as the channel conditions and traffic demands of the network.
## II Network Model and Spectrum Allocation
Let the wireless network be modelled by a directed and connected graph $`\mathcal{G} = {(\mathcal{N},\mathcal{L})}`$, where $`\mathcal{G}`$ is referred to as the connectivity graph of the network. A node $`i \in \mathcal{N}`$ represents a wireless transceiver, and $`{(i,j)} \in \mathcal{L}`$ represents a unidirectional wireless link from node $`i`$ to $`j`$. We assume $`\mathcal{G}`$ is link-symmetric, i.e., if $`{(i,j)} \in \mathcal{L}`$, then $`{(j,i)} \in \mathcal{L}`$, and vice versa. Two nodes $`i`$ and $`j`$ are neighbors if $`{(i,j)} \in \mathcal{L}`$ (or equivalently $`{(j,i)} \in \mathcal{L}`$). Let $`\mathcal{O}_{i} \triangleq {\{ j:{{(i,j)} \in \mathcal{L}}\}}`$ be the set of $`i`$’s neighbors.
###
II-A
Duplexing Constrained Interference Model
We focus on a network model that prohibits any node from simultaneously transmitting and receiving on the same frequency band. That is, we impose the duplexing constraint on every band. Such a constraint is less stringent but more fundamental than the extensively studied primary interference constraint \[[5](#bib.bib5), [6](#bib.bib6), [7](#bib.bib7)\], where any node can transmit or receive (at any given time and on any given band) on at most one active link, and the secondary interference constraint, which further prohibits any node from transmitting when there is a neighbor receiving from another node \[[8](#bib.bib8)\]. A general approach to studying a broad class of interference constraints was presented in \[[9](#bib.bib9)\]. Indeed, duplexing constraints cannot be bypassed (at least currently) by using sophisticated coding methods, and must be observed by practical network management techniques.
Traditional network management techniques which aim to resolve various types of interference among wireless links have concentrated on scheduling in time \[[10](#bib.bib10), [11](#bib.bib11), [12](#bib.bib12), [13](#bib.bib13), [14](#bib.bib14), [15](#bib.bib15), [16](#bib.bib16), [17](#bib.bib17)\], where at any given time, only mutually-non-interfering links are activated. Scheduling, however, usually requires centralized controllers and involves high communication and computational complexity \[[7](#bib.bib7), [18](#bib.bib18)\]. Simplified distributed scheduling policies have been proposed \[[19](#bib.bib19), [8](#bib.bib8), [20](#bib.bib20), [5](#bib.bib5), [6](#bib.bib6)\] for various purposes. In general, however, the reduced implementation complexity comes at the expense of performance \[[21](#bib.bib21)\].
The difficulty in finding interference-free schedules in time leads us to seek an alternative solution. A natural approach is to consider network management in the domain of frequency instead of or in addition to time \[[22](#bib.bib22), [23](#bib.bib23), [24](#bib.bib24), [25](#bib.bib25)\]. Because communication on different frequency bands are practically non-interfering, one can think of simultaneously applying different link activation sets on non-overlapping frequency bands within the assigned spectrum. In this scheme, nodes transmit on certain bands while receiving on other bands to avoid duplexing interference. The spectrum allocation technique has an important advantage over scheduling in time: once a feasible spectrum allocation is established, the network can function relatively statically in that mode, not having to switch to another mode unless network itself changes substantially. While spectrum allocation problems have been proposed and studied in the interference graph induced by particular interference constraints, the solution using existing vertex-coloring methods \[[22](#bib.bib22), [23](#bib.bib23)\] is cumbersome and its complexity scales poorly with the size of the network. Moreover, the optimization of frequency assignment techniques has not been thoroughly investigated. In particular, the number of available frequency bands is often arbitrarily set and frequency bands are assigned to links in a heuristic manner \[[24](#bib.bib24), [25](#bib.bib25)\].
In this work, we adopt the spectrum allocation approach to resolve the fundamental duplexing conflicts for general wireless networks. In particular, we investigate two central questions: (1) what is the minimum number of frequency bands with which all co-located transmitters and receivers can be simultaneously activated subject to duplexing constraints, and (2) how can one efficiently find a feasible frequency assignment when there are enough frequency bands? We provide an exact analytical answer to question (1) and develop a distributed asynchronous algorithm which solves problem (2). Our analysis is based only the network connectivity graph. This approach requires much less storage and computation overhead than alternative methods that utilize the interference graph.
###
II-B
Spectrum Division and Sub-band Allocation
The duplexing constraint permits only a subset of the links to be activated simultaneously on each frequency band. To activate all links at the same time, it is necessary to divide the spectrum into several sub-bands and activate different subsets of conflict-free links on different sub-bands.
Suppose the network occupies a contiguous spectrum which can be partitioned into a number, say $`Q`$, of sub-bands, each of which covers a contiguous segment of the whole spectrum. Let the collection of the sub-bands be denoted by $`\mathcal{Q}`$. With a specific spectrum division in place, each link can be active on one or more of the sub-bands. If link $`(i,j)`$ uses sub-band $`q`$, we say $`(i,j)`$ is an active link on $`q`$, and $`q`$ is an active sub-band of $`(i,j)`$. Denote the subset of links that are active on a sub-band $`q`$ by $`\mathcal{L}_{q}`$, and the set of active sub-bands of $`(i,j)`$ by $`\mathcal{Q}_{i\hspace{0pt}j}`$. A spectrum allocation is given by the collection $`{\{\mathcal{L}_{q}\}}_{q \in \mathcal{Q}}`$ (or equivalently $`{\{\mathcal{Q}_{i\hspace{0pt}j}\}}_{{(i,j)} \in \mathcal{L}}`$). Note that finding a spectrum allocation involves two steps: spectrum division, which decides how many sub-bands the whole spectrum is divided into, and sub-band allocation, which determines which links are active on which sub-bands. We will address these two issues in Sections [III](#S3 "III Spectrum Division ‣ Distributed Algorithms for Spectrum Allocation, Power Control, Routing, and Congestion Control in Wireless Networks") and [IV](#S4 "IV Distributed Sub-band Allocation ‣ Distributed Algorithms for Spectrum Allocation, Power Control, Routing, and Congestion Control in Wireless Networks"), respectively.
A spectrum allocation is feasible if (i) for all $`{(i,j)} \in \mathcal{L}`$, $`\mathcal{Q}_{i\hspace{0pt}j} \neq \varnothing`$, and (ii) for all $`q \in \mathcal{Q}`$, $`\mathcal{L}_{q}`$ satisfies the duplexing constraint. Thus, any node’s outgoing and incoming links cannot be both active on the same sub-band. However, it is feasible for a node to have multiple active outgoing links or multiple active incoming links on a single sub-band.
###
II-C
Interference Graph and Number of Sub-bands
In previous studies on frequency assignment techniques, the minimum number of frequency bands is found from the interference graph induced by the specific interference constraints \[[22](#bib.bib22), [23](#bib.bib23)\]. For the duplexing constraints, the interference graph $`\overset{\sim}{\mathcal{G}}`$ is constructed as follows. Let the vertices of the interference graph $`\overset{\sim}{\mathcal{G}} = {(\mathcal{V},\mathcal{E})}`$ be the links of the network connectivity graph $`\mathcal{G} = {(\mathcal{N},\mathcal{L})}`$, i.e., $`\mathcal{V} = \mathcal{L}`$. In $`\overset{\sim}{\mathcal{G}}`$, an edge exists between two vertices (links in $`\mathcal{G}`$) if one link’s transmitter is the other link’s receiver. The interference graph, unlike the connectivity graph, is undirected.
<sup>4</sup>
<sup>4</sup>4In order to avoid confusion, we refer to the connectivity graph as having nodes $`\mathcal{N}`$ and links $`\mathcal{L}`$, and the interference graph as having vertices $`\mathcal{V}`$ and edges $`\mathcal{E}`$.
It is easy to see that a feasible spectrum allocation on $`\mathcal{G}`$ exists if and only if the number of available sub-bands is greater than or equal to the chromatic number $`\chi\hspace{0pt}{(\overset{\sim}{\mathcal{G}})}`$ of the interference graph $`\overset{\sim}{\mathcal{G}}`$. We illustrate the interference graph approach in Figure [1](#S2.F1 "Figure 1 ‣ II-C Interference Graph and Number of Sub-bands ‣ II Network Model and Spectrum Allocation ‣ Distributed Algorithms for Spectrum Allocation, Power Control, Routing, and Congestion Control in Wireless Networks") using a complete four-node network. On the left is the connectivity graph $`\mathcal{G}`$, whose induced interference graph $`\overset{\sim}{\mathcal{G}}`$ is shown in the middle. A minimal vertex-coloring on $`\overset{\sim}{\mathcal{G}}`$ using four colors $`\{ R,G,B,Y\}`$ is depicted. The right graph represents the link-coloring on $`\mathcal{G}`$ inferred by the minimal vertex-coloring on $`\overset{\sim}{\mathcal{G}}`$. The color(s) associated with each node are those assigned to the outgoing links of that node. The reason for using such a representation will be explained shortly.
[Refer to caption](/html/0704.3746/assets/x1.png)
Figure 1: The interference graph approach on the complete graph with four nodes
Although the interference graph approach can provide an answer to our problem, it has major shortcomings. Note that the size of the interference graph $`{|\mathcal{V}|} = {|\mathcal{L}|}`$ is on the order of $`{|\mathcal{N}|}^{2}`$. To compute the chromatic number of $`\overset{\sim}{\mathcal{G}}`$, $`\overset{\sim}{\mathcal{G}}`$ has to be constructed and stored at a central controller which then computes $`\chi\hspace{0pt}{(\overset{\sim}{\mathcal{G}})}`$ by finding a minimal vertex-coloring, which is itself an NP-complete problem \[[26](#bib.bib26)\]. For these reasons, the interference graph approach is not tractable for medium-and-large-scale networks. This motivates us to seek an alternative method that can find the minimum number of sub-bands directly from the connectivity graph $`\mathcal{G}`$.
## III Spectrum Division
In this section, we investigate the minimum number of sub-bands which yields a feasible spectrum allocation. Our analysis will be based entirely on the network connectivity graph $`\mathcal{G}`$. For convenience of exposition, we first transform the problem into an equivalent graph-theoretic link-coloring problem.
###
III-A
Link-Coloring Problem
For the moment, we leave the total number of sub-bands undetermined. Let each sub-band be identified by a unique color. We represent a spectrum allocation by a color assignment to links. A feasible spectrum allocation corresponds to a color assignment such that (i) all links are assigned with at least one color and (ii) for any node, no incoming link has a common color with any one of its outgoing links. We will find the minimum number of colors required for such a coloring. For any color assignment, denote the set of colors used by node $`i`$’s outgoing links by $`\mathcal{O}\hspace{0pt}\mathcal{C}_{i}`$, and the set of colors used by its incoming links by $`\mathcal{I}\hspace{0pt}\mathcal{C}_{i}`$. Since the network is connected and link-symmetric, every node has at least one outgoing and one incoming link, implying that $`\mathcal{O}\hspace{0pt}\mathcal{C}_{i}`$ and $`\mathcal{I}\hspace{0pt}\mathcal{C}_{i}`$ are both nonempty for all $`i \in \mathcal{N}`$. A color assignment is feasible if and only if $`{{\mathcal{O}\hspace{0pt}\mathcal{C}_{i}} \cap {\mathcal{I}\hspace{0pt}\mathcal{C}_{i}}} = \varnothing`$ for all $`i \in \mathcal{N}`$.
<sup>5</sup>
<sup>5</sup>5Note that we allow one link to transmit on multiple sub-bands. Hence, we do not require $`{|{\mathcal{O}\hspace{0pt}\mathcal{C}_{i}}|} \leq {|\mathcal{O}_{i}|}`$ or $`{|{\mathcal{I}\hspace{0pt}\mathcal{C}_{i}}|} \leq {|\mathcal{O}_{i}|}`$.
Note that the feasibility of a coloring scheme can be checked using the node color sets $`{\{{\mathcal{O}\hspace{0pt}\mathcal{C}_{i}},{\mathcal{I}\hspace{0pt}\mathcal{C}_{i}}\}}_{i \in \mathcal{N}}`$, regardless of the detailed assignment of colors to links. In fact, we will show that studying the node outgoing color sets $`\{{\mathcal{O}\hspace{0pt}\mathcal{C}_{i}}\}`$ alone suffices for the link coloring problem. In the following, we say a color assignment is consistent with $`\{{\mathcal{O}\hspace{0pt}\mathcal{C}_{i}}\}`$ if $`\{{\mathcal{O}\hspace{0pt}\mathcal{C}_{i}}\}`$ results from the color assignment.
######
Lemma 1
Given a graph $`\mathcal{G} = {\{\mathcal{N},\mathcal{L}\}}`$ and the nodes’ outgoing color sets $`{\{{\mathcal{O}\hspace{0pt}\mathcal{C}_{i}}\}}_{i \in \mathcal{N}}`$, there exists a consistent and feasible color assignment if and only if
\(i\) $`{{{\mathcal{O}\hspace{0pt}\mathcal{C}_{i}}\backslash\mathcal{O}}\hspace{0pt}\mathcal{C}_{j}} \neq \varnothing`$, for all $`{(i,j)} \in \mathcal{L}`$;
\(ii\) $`{\bigcup_{j \in \mathcal{O}_{i}}{{{\mathcal{O}\hspace{0pt}\mathcal{C}_{i}}\backslash\mathcal{O}}\hspace{0pt}\mathcal{C}_{j}}} = {\mathcal{O}\hspace{0pt}\mathcal{C}_{i}}`$, for all $`i \in \mathcal{N}`$.
Proof: We first prove the necessity part. Suppose first that (i) is violated, i.e., there exists a link $`(i,j)`$ such that $`{\mathcal{O}\hspace{0pt}\mathcal{C}_{i}} \subseteq {\mathcal{O}\hspace{0pt}\mathcal{C}_{j}}`$, then $`(i,j)`$ must have a color in $`\mathcal{O}\hspace{0pt}\mathcal{C}_{j}`$, meaning that some outgoing link of $`j`$ has the same color as $`(i,j)`$, which is infeasible. Next suppose that (ii) does not hold for some node $`i`$, i.e., there exists a color $`q \in {\mathcal{O}\hspace{0pt}\mathcal{C}_{i}}`$ such that $`q \in {\mathcal{O}\hspace{0pt}\mathcal{C}_{j}}`$ of every $`j \in {\mathcal{O}\hspace{0pt}\mathcal{C}_{i}}`$. Since $`q`$ is used on some outgoing link of $`i`$, say $`(i,k)`$, $`(i,k)`$ will have a color belonging to $`\mathcal{O}\hspace{0pt}\mathcal{C}_{k}`$, violating the duplexing constraint at node $`k`$.
The conditions are sufficient because given $`{\{{\mathcal{O}\hspace{0pt}\mathcal{C}_{i}}\}}_{i \in \mathcal{N}}`$ which satisfy (i)-(ii), a feasible coloring can be constructed by assigning each link $`(i,j)`$ the color(s) in $`{{\mathcal{O}\hspace{0pt}\mathcal{C}_{i}}\backslash\mathcal{O}}\hspace{0pt}\mathcal{C}_{j}`$. By (ii), this coloring is also consistent with $`{\{{\mathcal{O}\hspace{0pt}\mathcal{C}_{i}}\}}_{i \in \mathcal{N}}`$. ∎
In the following, we use the shorthand notation $`{\mathcal{O}\hspace{0pt}\mathcal{C}_{i}} \nsim {\mathcal{O}\hspace{0pt}\mathcal{C}_{j}}`$ to represent $`{{{\mathcal{O}\hspace{0pt}\mathcal{C}_{i}}\backslash\mathcal{O}}\hspace{0pt}\mathcal{C}_{j}} \neq \varnothing`$ and $`{{{\mathcal{O}\hspace{0pt}\mathcal{C}_{j}}\backslash\mathcal{O}}\hspace{0pt}\mathcal{C}_{i}} \neq \varnothing`$. A collection of node outgoing color sets $`{\{{\mathcal{O}\hspace{0pt}\mathcal{C}_{i}}\}}_{i \in \mathcal{N}}`$ is said to be feasible if it satisfies the conditions in Lemma [1](#Thmlemma1 "Lemma 1 ‣ III-A Link-Coloring Problem ‣ III Spectrum Division ‣ Distributed Algorithms for Spectrum Allocation, Power Control, Routing, and Congestion Control in Wireless Networks").
We now state the problem of finding the minimum number of sub-bands in terms of the minimal coloring problem as follows. Given a graph $`\mathcal{G} = {(\mathcal{N},\mathcal{L})}`$,
<table>
<tbody>
<tr>
<td></td>
<td>minimize</td>
<td></td>
<td>|⋃i∈𝒩𝒪𝒞i|subscript𝑖𝒩𝒪subscript𝒞𝑖\displaystyle\left|\bigcup_{i\in{\cal N}}{\cal OC}_{i}\right|</td>
<td></td>
<td>(1)</td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td>subject to</td>
<td></td>
<td>{𝒪𝒞i}i∈𝒩feasible.subscript𝒪subscript𝒞𝑖𝑖𝒩feasible\displaystyle\{{\cal OC}_{i}\}_{i\in{\cal N}}~{}\textrm{feasible}.</td>
<td></td>
<td></td>
</tr>
</tbody>
</table>
Denote the minimum number of colors by $`Q_{\mathcal{G}}`$. To solve for $`Q_{\mathcal{G}}`$, we first relax the constraints of problem ([1](#S3.E1 "In III-A Link-Coloring Problem ‣ III Spectrum Division ‣ Distributed Algorithms for Spectrum Allocation, Power Control, Routing, and Congestion Control in Wireless Networks")) by dropping the second condition in Lemma [1](#Thmlemma1 "Lemma 1 ‣ III-A Link-Coloring Problem ‣ III Spectrum Division ‣ Distributed Algorithms for Spectrum Allocation, Power Control, Routing, and Congestion Control in Wireless Networks") and consider
<table>
<tbody>
<tr>
<td></td>
<td>minimize</td>
<td></td>
<td>|⋃i∈𝒩Bi|subscript𝑖𝒩subscript𝐵𝑖\displaystyle\left|\bigcup_{i\in{\cal N}}B_{i}\right|</td>
<td></td>
<td>(2)</td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td>subject to</td>
<td></td>
<td>Bi≁Bj,∀(i,j)∈ℒ,formulae-sequencenot-similar-tosubscript𝐵𝑖subscript𝐵𝑗for-all𝑖𝑗ℒ\displaystyle B_{i}\nsim B_{j},~{}~{}\forall(i,j)\in{\cal L},</td>
<td></td>
<td></td>
</tr>
</tbody>
</table>
where $`{B_{i},i} \in \mathcal{N}`$ are any non-empty sets. The optimal solution $`Q_{\mathcal{G}}^{\ast}`$ to ([2](#S3.E2 "In III-A Link-Coloring Problem ‣ III Spectrum Division ‣ Distributed Algorithms for Spectrum Allocation, Power Control, Routing, and Congestion Control in Wireless Networks")) must be less than or equal to $`Q_{\mathcal{G}}`$. We will find a lower bound on $`Q_{\mathcal{G}}^{\ast}`$ and a matching upper bound on $`Q_{\mathcal{G}}`$, thus uniquely determining $`Q_{\mathcal{G}}`$. For lower bounding $`Q_{\mathcal{G}}^{\ast}`$, the next observation is useful. The proof is deferred to Appendix [](#A0.SS1 "-A Proof of Lemma 2 ‣ VII Conclusion ‣ Distributed Algorithms for Spectrum Allocation, Power Control, Routing, and Congestion Control in Wireless Networks")
-A
.
######
Lemma 2
Let $`B_{1},\cdots,B_{k}`$ be distinct subsets of a $`Q`$-set all of which have the same cardinality $`g`$. If $`g > {\lfloor{Q/2}\rfloor}`$, then there exist distinct subsets $`C_{1},\cdots,C_{k}`$ such that $`C_{j} \subset B_{j}`$ and $`{|C_{j}|} = {g - 1}`$, $`j = {1,\cdots,k}`$. If $`g < {\lfloor{Q/2}\rfloor}`$, then there exist distinct subsets $`C_{1},\cdots,C_{k}`$ such that $`B_{j} \subset C_{j}`$ and $`{|C_{j}|} = {g + 1}`$, $`j = {1,\cdots,k}`$.
Using Lemma [2](#Thmlemma2 "Lemma 2 ‣ III-A Link-Coloring Problem ‣ III Spectrum Division ‣ Distributed Algorithms for Spectrum Allocation, Power Control, Routing, and Congestion Control in Wireless Networks") and the fact that any two distinct subsets $`B_{i}`$, $`B_{j}`$ with the same cardinality satisfy $`B_{i} \nsim B_{j}`$, we can show the following (see the proof in Appendix [](#A0.SS2 "-B Proof of Lemma 3 ‣ VII Conclusion ‣ Distributed Algorithms for Spectrum Allocation, Power Control, Routing, and Congestion Control in Wireless Networks")
-B
).
######
Lemma 3
There exists an optimal solution $`{\{ B_{i}\}}_{i \in \mathcal{N}}`$ for ([2](#S3.E2 "In III-A Link-Coloring Problem ‣ III Spectrum Division ‣ Distributed Algorithms for Spectrum Allocation, Power Control, Routing, and Congestion Control in Wireless Networks")) such that all $`B_{i}`$, $`i \in \mathcal{N}`$, have the same cardinality.
Using Lemma [3](#Thmlemma3 "Lemma 3 ‣ III-A Link-Coloring Problem ‣ III Spectrum Division ‣ Distributed Algorithms for Spectrum Allocation, Power Control, Routing, and Congestion Control in Wireless Networks"), we can find a lower bound on $`Q_{\mathcal{G}}^{\ast}`$ as follows.
######
Lemma 4
Given a graph $`\mathcal{G} = {(\mathcal{N},\mathcal{L})}`$ with chromatic number $`\chi\hspace{0pt}{(\mathcal{G})}`$, the solution $`Q_{\mathcal{G}}^{\ast}`$ of the problem ([2](#S3.E2 "In III-A Link-Coloring Problem ‣ III Spectrum Division ‣ Distributed Algorithms for Spectrum Allocation, Power Control, Routing, and Congestion Control in Wireless Networks")) is greater than or equal to $`Q\hspace{0pt}{({\chi\hspace{0pt}{(\mathcal{G})}})}`$, where the function $`Q:{{\mathbb{Z}}_{+}\mapsto{\mathbb{Z}}_{+}}`$ is defined as
<table>
<tbody>
<tr>
<td></td>
<td>Q(N)≜min{q∈ℤ+:(q⌊q/2⌋)≥N}.≜𝑄𝑁:𝑞subscriptℤbinomial𝑞𝑞2𝑁Q(N)\triangleq\min\left\{q\in\mathbb{Z}_{+}:~{}{{q}\choose{\lfloor{q/2}\rfloor}}\geq N\right\}.</td>
<td></td>
<td>(3)</td>
</tr>
</tbody>
</table>
Proof: By Lemma [3](#Thmlemma3 "Lemma 3 ‣ III-A Link-Coloring Problem ‣ III Spectrum Division ‣ Distributed Algorithms for Spectrum Allocation, Power Control, Routing, and Congestion Control in Wireless Networks"), we can without loss of optimality consider a solution $`{\{ B_{i}\}}_{i \in \mathcal{N}}`$ of ([2](#S3.E2 "In III-A Link-Coloring Problem ‣ III Spectrum Division ‣ Distributed Algorithms for Spectrum Allocation, Power Control, Routing, and Congestion Control in Wireless Networks")) such that $`{|B_{i}|} = {|B_{j}|}`$, for all $`{i,j} \in \mathcal{N}`$. Therefore, $`B_{i} \nsim B_{j}`$ whenever $`B_{i} \neq B_{j}`$. Suppose there are $`K`$ distinct elements in $`{\{ B_{i}\}}_{i \in \mathcal{N}}`$. Because $`{|{\bigcup_{i}B_{i}}|} = Q_{\mathcal{G}}^{\ast}`$, by Sperner’s theorem \[[27](#bib.bib27)\], $`K \leq \binom{Q_{\mathcal{G}}^{\ast}}{\lfloor{Q_{\mathcal{G}}^{\ast}/2}\rfloor}`$. Moreover, since $`\{ B_{i}\}`$ satisfies the constraint in ([2](#S3.E2 "In III-A Link-Coloring Problem ‣ III Spectrum Division ‣ Distributed Algorithms for Spectrum Allocation, Power Control, Routing, and Congestion Control in Wireless Networks")), we must have $`K \geq {\chi\hspace{0pt}{(\mathcal{G})}}`$, where $`\chi\hspace{0pt}{(\mathcal{G})}`$ is the chromatic number of $`\mathcal{G}`$. Therefore, $`Q_{\mathcal{G}}^{\ast} \geq {Q\hspace{0pt}{({\chi\hspace{0pt}{(\mathcal{G})}})}}`$. ∎
Lemma [4](#Thmlemma4 "Lemma 4 ‣ III-A Link-Coloring Problem ‣ III Spectrum Division ‣ Distributed Algorithms for Spectrum Allocation, Power Control, Routing, and Congestion Control in Wireless Networks") indicates that at least $`Q\hspace{0pt}{({\chi\hspace{0pt}{(\mathcal{G})}})}`$ colors are needed to construct a feasible collection of node color sets even with the condition (ii) in Lemma [1](#Thmlemma1 "Lemma 1 ‣ III-A Link-Coloring Problem ‣ III Spectrum Division ‣ Distributed Algorithms for Spectrum Allocation, Power Control, Routing, and Congestion Control in Wireless Networks") removed. However, by assigning node color sets according to a minimal node-labelling scheme of $`\mathcal{G}`$, $`Q\hspace{0pt}{({\chi\hspace{0pt}{(\mathcal{G})}})}`$ colors are sufficient to yield a feasible configuration satisfying both conditions in Lemma [1](#Thmlemma1 "Lemma 1 ‣ III-A Link-Coloring Problem ‣ III Spectrum Division ‣ Distributed Algorithms for Spectrum Allocation, Power Control, Routing, and Congestion Control in Wireless Networks"). We can therefore conclude the following.
######
Lemma 5
Given a graph $`\mathcal{G} = {(\mathcal{N},\mathcal{L})}`$ with chromatic number $`\chi\hspace{0pt}{(\mathcal{G})}`$, the solution $`Q_{\mathcal{G}}`$ to the problem ([1](#S3.E1 "In III-A Link-Coloring Problem ‣ III Spectrum Division ‣ Distributed Algorithms for Spectrum Allocation, Power Control, Routing, and Congestion Control in Wireless Networks")) is less than or equal to $`Q\hspace{0pt}{({\chi\hspace{0pt}{(\mathcal{G})}})}`$.
Proof: For graph $`\mathcal{G}`$, $`Q\hspace{0pt}{({\chi\hspace{0pt}{(\mathcal{G})}})}`$ colors are sufficient because they can generate at least $`\chi\hspace{0pt}{(\mathcal{G})}`$ color subsets $`\{ B_{i}\}`$ with equal cardinality $`\lfloor{{Q\hspace{0pt}{({\chi\hspace{0pt}{(\mathcal{G})}})}}/2}\rfloor`$, any two of which satisfy $`B_{i} \nsim B_{j}`$. Associate any $`\chi\hspace{0pt}{(\mathcal{G})}`$ of the color subsets with $`\chi\hspace{0pt}{(\mathcal{G})}`$ labels. Assign the $`\chi\hspace{0pt}{(\mathcal{G})}`$ color subsets to nodes according to a minimal node-labelling scheme, which requires exactly $`\chi\hspace{0pt}{(\mathcal{G})}`$ labels. We then have a feasible collection of node outgoing color sets that satisfies the two conditions in Lemma [1](#Thmlemma1 "Lemma 1 ‣ III-A Link-Coloring Problem ‣ III Spectrum Division ‣ Distributed Algorithms for Spectrum Allocation, Power Control, Routing, and Congestion Control in Wireless Networks").
<sup>6</sup>
<sup>6</sup>6An arbitrary color subset assignment according to a node-labelling solution generally results in $`{\{{\mathcal{O}\hspace{0pt}\mathcal{C}_{i}}\}}_{i \in \mathcal{N}}`$ satisfying only (i) in Lemma [1](#Thmlemma1 "Lemma 1 ‣ III-A Link-Coloring Problem ‣ III Spectrum Division ‣ Distributed Algorithms for Spectrum Allocation, Power Control, Routing, and Congestion Control in Wireless Networks"). Should (ii) not hold for some $`i`$, reset $`{\mathcal{O}\hspace{0pt}\mathcal{C}_{i}}:={\bigcup_{j \in \mathcal{O}_{i}}{{{\mathcal{O}\hspace{0pt}\mathcal{C}_{i}}\backslash\mathcal{O}}\hspace{0pt}\mathcal{C}_{j}}}`$. With this modification, both (i) and (ii) in Lemma [1](#Thmlemma1 "Lemma 1 ‣ III-A Link-Coloring Problem ‣ III Spectrum Division ‣ Distributed Algorithms for Spectrum Allocation, Power Control, Routing, and Congestion Control in Wireless Networks") hold and no extra color is needed.
∎
We have from Lemmas [4](#Thmlemma4 "Lemma 4 ‣ III-A Link-Coloring Problem ‣ III Spectrum Division ‣ Distributed Algorithms for Spectrum Allocation, Power Control, Routing, and Congestion Control in Wireless Networks") and [5](#Thmlemma5 "Lemma 5 ‣ III-A Link-Coloring Problem ‣ III Spectrum Division ‣ Distributed Algorithms for Spectrum Allocation, Power Control, Routing, and Congestion Control in Wireless Networks") that
<table>
<tbody>
<tr>
<td></td>
<td>Q(χ(𝒢))≤Q𝒢∗≤Q𝒢≤Q(χ(𝒢)).𝑄𝜒𝒢superscriptsubscript𝑄𝒢subscript𝑄𝒢𝑄𝜒𝒢Q(\chi({\cal G}))\leq Q_{{\cal G}}^{*}\leq Q_{{\cal G}}\leq Q(\chi({\cal G})).</td>
<td></td>
</tr>
</tbody>
</table>
Therefore, all inequalities must hold with equality.
######
Proposition 1
Let the chromatic number of graph $`\mathcal{G}`$ be $`\chi\hspace{0pt}{(\mathcal{G})}`$. Then, the solution of ([1](#S3.E1 "In III-A Link-Coloring Problem ‣ III Spectrum Division ‣ Distributed Algorithms for Spectrum Allocation, Power Control, Routing, and Congestion Control in Wireless Networks")) is $`Q_{\mathcal{G}} = {Q\hspace{0pt}{({\chi\hspace{0pt}{(\mathcal{G})}})}}`$, where the function $`Q\hspace{0pt}{( \cdot )}`$ is given by ([3](#S3.E3 "In Lemma 4 ‣ III-A Link-Coloring Problem ‣ III Spectrum Division ‣ Distributed Algorithms for Spectrum Allocation, Power Control, Routing, and Congestion Control in Wireless Networks")).
It is easy to see that $`Q\hspace{0pt}{( \cdot )}`$ is nondecreasing. The values of $`Q\hspace{0pt}{(N)}`$ for $`N = {1,\cdots,20}`$ are plotted in Figure [2](#S3.F2 "Figure 2 ‣ III-A Link-Coloring Problem ‣ III Spectrum Division ‣ Distributed Algorithms for Spectrum Allocation, Power Control, Routing, and Congestion Control in Wireless Networks").
[Refer to caption](/html/0704.3746/assets/x2.png)
Figure 2: The values of $`Q\hspace{0pt}{(N)}`$, $`N = {1,\cdots,20}`$.
For large $`N`$, it follows from Stirling’s approximation \[[28](#bib.bib28)\]
<table>
<tbody>
<tr>
<td></td>
<td>(q⌊q/2⌋)=2πexp[qln2−12lnq−14q+o(1q2)]binomial𝑞𝑞22𝜋𝑞212𝑞14𝑞𝑜1superscript𝑞2{{q}\choose{\lfloor{q/2}\rfloor}}=\sqrt{\frac{2}{\pi}}\exp\left[q\ln 2-\frac{1}{2}\ln q-\frac{1}{4q}+o\left(\frac{1}{q^{2}}\right)\right]</td>
<td></td>
</tr>
</tbody>
</table>
that
<table>
<tbody>
<tr>
<td></td>
<td>Q(N)=Θ(logN).𝑄𝑁Θ𝑁Q(N)=\Theta(\log N).</td>
<td></td>
</tr>
</tbody>
</table>
Therefore by Proposition [1](#Thmproposition1 "Proposition 1 ‣ III-A Link-Coloring Problem ‣ III Spectrum Division ‣ Distributed Algorithms for Spectrum Allocation, Power Control, Routing, and Congestion Control in Wireless Networks"), $`Q_{\mathcal{G}}`$ grows at a *logarithmic* rate with $`\chi\hspace{0pt}{(\mathcal{G})}`$.
## IV Distributed Sub-band Allocation
###
IV-A
Spectrum Division Using Approximation of $`\chi\hspace{0pt}{(\mathcal{G})}`$
We have found that the minimum number of sub-bands depends on the chromatic number $`\chi\hspace{0pt}{(\mathcal{G})}`$. The problem of finding $`\chi\hspace{0pt}{(\mathcal{G})}`$, however, is NP-complete \[[26](#bib.bib26)\] in general. On the other hand, any upper bound on $`\chi\hspace{0pt}{(\mathcal{G})}`$ gives us a sufficient number of sub-bands. One such well-known upper bound is $`{\chi\hspace{0pt}{(\mathcal{G})}} \leq {{\Delta\hspace{0pt}{(\mathcal{G})}} + 1}`$, where $`\Delta\hspace{0pt}{(\mathcal{G})}`$ is the maximum degree of any node in $`\mathcal{G}`$. In the following, we assume that there are at least $`Q\hspace{0pt}{({{\Delta\hspace{0pt}{(\mathcal{G})}} + 1})}`$ sub-bands available.
The maximum degree $`\Delta\hspace{0pt}{(\mathcal{G})}`$ is straightforward to determine (in a distributed manner), and we assume that it is known to all nodes a priori. Notice that $`Q\hspace{0pt}{({{\Delta\hspace{0pt}{(\mathcal{G})}} + 1})}`$ is not too far from $`Q\hspace{0pt}{({\chi\hspace{0pt}{(\mathcal{G})}})}`$ in typical networks,
<sup>7</sup>
<sup>7</sup>7There are graphs, e.g. complete graphs, for which $`{\chi\hspace{0pt}{(\mathcal{G})}} \leq {{\Delta\hspace{0pt}{(\mathcal{G})}} + 1}`$ holds with equality.
since the $`Q\hspace{0pt}{( \cdot )}`$ function is piecewise flat and grows approximately as a log function. It is worth comparing the upper bound $`Q\hspace{0pt}{({{\Delta\hspace{0pt}{(\mathcal{G})}} + 1})}`$ with $`{\Delta\hspace{0pt}{(\overset{\sim}{\mathcal{G}})}} + 1`$, an upper bound of $`\chi\hspace{0pt}{(\overset{\sim}{\mathcal{G}})}`$, the chromatic number of the interference graph. Recall that vertices of $`\overset{\sim}{\mathcal{G}}`$ are (directed) links of $`\mathcal{G}`$. Any vertex $`(i,j)`$ of $`\overset{\sim}{\mathcal{G}}`$ has neighbors $`\{{{(k,l)} \in \mathcal{L}}:{{k = {j\hspace{0pt}\text{or}\hspace{0pt}l} = i},\text{but not both}}\}`$. Hence, by the assumption that $`\mathcal{G}`$ is link-symmetric, the degree of $`(i,j)`$ in $`\overset{\sim}{\mathcal{G}}`$ is given by
<table>
<tbody>
<tr>
<td></td>
<td>Δ(i,j)=|ℐi|+|𝒪j|−1=Δi+Δj−1.subscriptΔ𝑖𝑗subscriptℐ𝑖subscript𝒪𝑗1subscriptΔ𝑖subscriptΔ𝑗1\Delta_{(i,j)}=|{\cal I}_{i}|+|{\cal O}_{j}|-1=\Delta_{i}+\Delta_{j}-1.</td>
<td></td>
</tr>
</tbody>
</table>
Here, $`\Delta_{i}`$ and $`\Delta_{j}`$ are the degrees of nodes $`i`$ and $`j`$ in $`\mathcal{G}`$. Consequently, the maximum degree of $`\overset{\sim}{\mathcal{G}}`$ is
<table>
<tbody>
<tr>
<td></td>
<td>Δ(𝒢~)=max(i,j)∈ℒΔi+Δj−1.Δ~𝒢subscript𝑖𝑗ℒsubscriptΔ𝑖subscriptΔ𝑗1\Delta(\tilde{\cal G})=\max_{(i,j)\in{\cal L}}\Delta_{i}+\Delta_{j}-1.</td>
<td></td>
</tr>
</tbody>
</table>
On the other hand, $`{\Delta\hspace{0pt}{(\mathcal{G})}} = {\max_{i \in \mathcal{N}}\Delta_{i}}`$. Therefore, it is always true that
<table>
<tbody>
<tr>
<td></td>
<td>Δ(𝒢~)≥Δ(𝒢).Δ~𝒢Δ𝒢\Delta(\tilde{\cal G})\geq\Delta({\cal G}).</td>
<td></td>
</tr>
</tbody>
</table>
The equality holds only when the network topology is extremely asymmetric, e.g. the star networks. Typically, $`\Delta\hspace{0pt}{(\overset{\sim}{\mathcal{G}})}`$ can be almost twice as large as $`\Delta\hspace{0pt}{(\mathcal{G})}`$. Moreover, the $`Q\hspace{0pt}{( \cdot )}`$ function is piecewise flat and scales as $`\log_{2}{( \cdot )}`$. The difference of the two upper bounds $`{{\Delta\hspace{0pt}{(\overset{\sim}{\mathcal{G}})}} + 1} - {Q\hspace{0pt}{({{\Delta\hspace{0pt}{(\mathcal{G})}} + 1})}}`$ typically is substantial. Therefore, the connectivity graph approach developed in this work not only is more simple and straightforward but also leads to a tighter upper bound than the conventional interference graph approach.
###
IV-B
Distributed Sub-band Allocation Algorithm
With the spectrum division accomplished, we now devise an algorithm with which feasible $`\{{\mathcal{O}\hspace{0pt}\mathcal{C}_{i}}\}`$ can be determined in a distributed manner and a feasible spectrum allocation can be found. The algorithm applies to all connected and link-symmetric graphs. Assume that $`Q`$ sub-bands are available. For expositional purposes, we keep an unprocessed node set $`\mathcal{U}`$, which initially contains all the nodes. The Distributed Sub-band allocation (DSA) algorithm is iterated as follows.
Step 1. Initially $`\mathcal{U} = \mathcal{N}`$. Arbitrarily select a node $`i`$ from $`\mathcal{U}`$ and set $`\mathcal{U}:={\mathcal{U}\backslash{\{ i\}}}`$. Arbitrarily choose $`\lfloor{Q/2}\rfloor`$ sub-bands to form $`\mathcal{O}\hspace{0pt}\mathcal{C}_{i}`$, mark $`i`$ as “processed”, and go to Step 2.
Step 2. Select an arbitrary node $`i`$ from $`\mathcal{U}`$ such that $`i`$ has at least one processed neighbor (under the assumption that the network is connected, there is always such a node in $`\mathcal{U}`$ after the first node is processed) and set $`\mathcal{U}:={\mathcal{U}\backslash{\{ i\}}}`$. Node $`i`$ finds an $`\mathcal{O}\hspace{0pt}\mathcal{C}_{i}`$ with $`\lfloor{Q/2}\rfloor`$ sub-bands which is different from all $`\mathcal{O}\hspace{0pt}\mathcal{C}_{j}`$ of its processed neighbors $`j`$ (it can be shown that such an $`\mathcal{O}\hspace{0pt}\mathcal{C}_{i}`$ always exists). Moreover, to maximally avoid potential channel interference, $`i`$ selects sub-bands in increasing order of their number of occurrences
<sup>8</sup>
<sup>8</sup>8The number of occurrences can be zero.
in $`{\{{\mathcal{O}\hspace{0pt}\mathcal{C}_{j}}\}}_{j \in {\mathcal{O}_{i}\backslash\mathcal{U}}}`$ (ties are broken arbitrarily).
<sup>9</sup>
<sup>9</sup>9For instance, if the whole set of sub-bands is $`\{ R,G,B,Y\}`$, and node $`i`$ finds its two neighbors have chosen $`\{ R,G\}`$ and $`\{ R,Y\}`$, it will accordingly choose either $`\{ B,G\}`$ or $`\{ B,Y\}`$.
Mark $`i`$ as “processed”. If $`\mathcal{U} \neq \varnothing`$, repeat Step 2; otherwise go to Step 3.
Step 3. Each node $`i`$ allocates the sub-band(s) in $`\mathcal{Q}_{i\hspace{0pt}j}:={{{\mathcal{O}\hspace{0pt}\mathcal{C}_{i}}\backslash\mathcal{O}}\hspace{0pt}\mathcal{C}_{j}}`$ to each outgoing link $`(i,j)`$. The algorithm terminates.
######
Proposition 2
The set $`{\{\mathcal{Q}_{i\hspace{0pt}j}\}}_{{(i,j)} \in \mathcal{L}}`$ generated by the DSA algorithm induces a feasible spectrum allocation $`{\{\mathcal{L}_{q}\}}_{q \in \mathcal{Q}}`$.
Proof: First we show that in Step 2 of the algorithm, an appropriate $`\mathcal{O}\hspace{0pt}\mathcal{C}_{i}`$ can always be found. Because the whole set has $`Q \geq {Q\hspace{0pt}{({{\Delta\hspace{0pt}{(\mathcal{G})}} + 1})}}`$ elements, there are at least $`{\Delta\hspace{0pt}{(\mathcal{G})}} + 1`$ distinct subsets with cardinality $`\lfloor{Q/2}\rfloor`$. However, $`i`$ has at most $`\Delta\hspace{0pt}{(\mathcal{G})}`$ (processed) neighbors. Hence, there always exists at least one candidate for $`\mathcal{O}\hspace{0pt}\mathcal{C}_{i}`$.
Now we show that the $`\{{\mathcal{O}\hspace{0pt}\mathcal{C}_{i}}\}`$ obtained up to Step 3 are feasible. They satisfy (i) in Lemma [1](#Thmlemma1 "Lemma 1 ‣ III-A Link-Coloring Problem ‣ III Spectrum Division ‣ Distributed Algorithms for Spectrum Allocation, Power Control, Routing, and Congestion Control in Wireless Networks") due to the rule of successively selecting $`\mathcal{O}\hspace{0pt}\mathcal{C}_{i}`$ in Step 2. They also satisfy (ii) in Lemma [1](#Thmlemma1 "Lemma 1 ‣ III-A Link-Coloring Problem ‣ III Spectrum Division ‣ Distributed Algorithms for Spectrum Allocation, Power Control, Routing, and Congestion Control in Wireless Networks"), as proved next. Suppose on the contrary that there exists $`i \in \mathcal{N}`$ and $`q \in {\mathcal{O}\hspace{0pt}\mathcal{C}_{i}}`$ such that $`q \in {\mathcal{O}\hspace{0pt}\mathcal{C}_{j}}`$ for all $`j \in \mathcal{O}_{i}`$. If $`i`$ is not the first processed node, the fact that $`q \in {\mathcal{O}\hspace{0pt}\mathcal{C}_{i}}`$ implies that by the time $`\mathcal{O}\hspace{0pt}\mathcal{C}_{i}`$ is being determined, each of the $`Q - {\lfloor{Q/2}\rfloor}`$ sub-bands not included in $`\mathcal{O}\hspace{0pt}\mathcal{C}_{i}`$ must appear in every previously processed neighbor $`j`$’s $`\mathcal{O}\hspace{0pt}\mathcal{C}_{j}`$. Taking also $`q`$ into account, we can deduce that all those $`\mathcal{O}\hspace{0pt}\mathcal{C}_{j}`$ have at least $`{\lfloor{Q/2}\rfloor} + 1`$ elements, which is a contradiction. Even if $`i`$ is the first processed node, the next processed node $`j`$ must be $`i`$’s neighbor and $`{{\mathcal{O}\hspace{0pt}\mathcal{C}_{j}} \cup {\mathcal{O}\hspace{0pt}\mathcal{C}_{i}}} = \varnothing`$ due to the algorithm’s rule. Hence, the hypothesis cannot be true in either case.
With the feasibility of the $`\{{\mathcal{O}\hspace{0pt}\mathcal{C}_{i}}\}`$ established, it is easy to verify that the $`{\{\mathcal{Q}_{i\hspace{0pt}j}\}}_{{(i,j)} \in \mathcal{L}}`$ generated in Step 3 induces a feasible spectrum allocation $`{\{\mathcal{L}_{q}\}}_{q \in \mathcal{Q}}`$. ∎
In practice, the DSA algorithm can be implemented in a distributed fashion by nodes in the network. Specifically, after any node $`i`$ has arbitrarily set its $`\mathcal{O}\hspace{0pt}\mathcal{C}_{i}`$ at the beginning, any other node $`j`$ can determine its own $`\mathcal{O}\hspace{0pt}\mathcal{C}_{j}`$ as long as it has at least one processed neighbor and no other neighbor is being processed at the same instant. In other words, it is possible to have two non-adjacent nodes configuring their outgoing sub-bands at the same time. Since the action of a node depends only on its neighbors, node operations need not be globally coordinated or synchronized across the network. All that is required is an initialization phase that designates the first node to be processed. Thus, the sub-band allocation algorithm is distributed and asynchronous in nature.
Finally, we note that the DSA algorithm is robust to dynamic node failure and addition. Suppose that after a feasible spectrum allocation is established by the DSA algorithm, a node leaves the network. In this case, the configuration for the remaining nodes and links is still feasible. If a new node $`i`$ is to join the network, it can run Step 2 of the DSA algorithm to determine its own $`\mathcal{O}\hspace{0pt}\mathcal{C}_{i}`$ based on its neighbors’ $`{{\mathcal{O}\hspace{0pt}\mathcal{C}_{j}},j} \in \mathcal{O}_{i}`$. Then, active sub-bands for each link incident to $`i`$ can be allocated as in Step 3. As long as $`i`$ connects to at most $`\Delta\hspace{0pt}{(\mathcal{G})}`$ nodes in the old network $`\mathcal{G}`$, the algorithm yields a feasible spectrum allocation for node $`i`$, with all other nodes’ allocation unaffected.
## V Cross-layer Optimization in Multi-radio Multi-hop Wireless Networks
The algorithm in Section [IV](#S4 "IV Distributed Sub-band Allocation ‣ Distributed Algorithms for Spectrum Allocation, Power Control, Routing, and Congestion Control in Wireless Networks") generates a spectrum allocation where the active links on every sub-band satisfy the duplexing constraints. Given this feasible spectrum allocation, we now develop corresponding joint power control, routing, and congestion control algorithms which minimize total network cost, which will be specified later, given channel conditions and traffic demands.
###
V-A
Interference Limited Transmissions and Node Power Constraints
We consider networks where messages on each sub-band are coded independently, and where the receiver of a link decodes its message on an active sub-band while treating all other signals on the same sub-band as interference. We assume that the capacity $`C_{i\hspace{0pt}j}\hspace{0pt}{(q)}`$ of link $`(i,j)`$ on a sub-band $`q`$ is a function of $`x_{i\hspace{0pt}j}\hspace{0pt}{(q)}`$, the signal-to-interference-plus-noise ratio (SINR) of link $`(i,j)`$ over $`q`$. Denoting the transmission powers used by the active links on sub-band $`q`$ by $`{\{{P_{m\hspace{0pt}n}\hspace{0pt}{(q)}}\}}_{{(m,n)} \in \mathcal{L}_{q}}`$, we have
<table>
<tbody>
<tr>
<td></td>
<td>xij(q)=GijqPij(q)∑(m,n)∈ℒq(m,n)≠(i,j)GmjqPmn(q)+Njq,subscript𝑥𝑖𝑗𝑞superscriptsubscript𝐺𝑖𝑗𝑞subscript𝑃𝑖𝑗𝑞subscript𝑚𝑛subscriptℒ𝑞𝑚𝑛𝑖𝑗superscriptsubscript𝐺𝑚𝑗𝑞subscript𝑃𝑚𝑛𝑞superscriptsubscript𝑁𝑗𝑞x_{ij}(q)=\frac{G_{ij}^{q}P_{ij}(q)}{\sum_{\begin{subarray}{c}(m,n)\in{\cal L}_{q}\\ (m,n)\neq(i,j)\end{subarray}}G_{mj}^{q}P_{mn}(q)+N_{j}^{q}},</td>
<td></td>
<td>(4)</td>
</tr>
</tbody>
</table>
where $`G_{m\hspace{0pt}j}^{q}`$ is the path gain from $`m`$ to $`j`$ on sub-band $`q`$, and $`N_{j}^{q}`$ is the power of the additive noise on sub-band $`q`$ at $`j`$. Note that since the parameters $`\{ G_{i\hspace{0pt}j}^{q}\}`$ and $`\{ N_{j}^{q}\}`$ are sub-band-dependent, this framework is particularly appropriate for networks with frequency selective channels and colored noise.
Assume each node is limited by an individual power constraint $`{\overline{P}}_{i}`$, i.e.,
<table>
<tbody>
<tr>
<td></td>
<td>∑j∈𝒪i∑q∈𝒬ijPij(q)≤P¯i.subscript𝑗subscript𝒪𝑖subscript𝑞subscript𝒬𝑖𝑗subscript𝑃𝑖𝑗𝑞subscript¯𝑃𝑖\sum_{j\in{\cal O}_{i}}\sum_{q\in{\cal Q}_{ij}}P_{ij}(q)\leq\bar{P}_{i}.</td>
<td></td>
<td>(5)</td>
</tr>
</tbody>
</table>
Denote the set of power variables $`{\{{P_{i\hspace{0pt}j}\hspace{0pt}{(q)}}\}}_{{{(i,j)} \in \mathcal{L}},{q \in \mathcal{Q}_{i\hspace{0pt}j}}}`$ that satisfy ([5](#S5.E5 "In V-A Interference Limited Transmissions and Node Power Constraints ‣ V Cross-layer Optimization in Multi-radio Multi-hop Wireless Networks ‣ Distributed Algorithms for Spectrum Allocation, Power Control, Routing, and Congestion Control in Wireless Networks")) by $`\mathcal{P}`$. In Section [VI](#S6 "VI Node-Based Multi-Radio Power Control and Routing Algorithms ‣ Distributed Algorithms for Spectrum Allocation, Power Control, Routing, and Congestion Control in Wireless Networks"), we will design a set of power control algorithms that adjusts the power variables within the feasible region to minimize total network cost in conjunction with congestion control and routing.
###
V-B
Traffic Demands, Congestion Control, and Routing
Let the traffic demands for the network consist of a collection $`\mathcal{W}`$ of unicast sessions. Each (elastic) session $`w \in \mathcal{W}`$ is characterized by its fixed source-destination node pair $`({O\hspace{0pt}{(w)}},{D\hspace{0pt}{(w)}})`$ and demand rate $`{\overline{r}}_{w}`$.
<sup>10</sup>
<sup>10</sup>10We assume that user $`w`$ gains no extra utility by transmitting at a rate higher than $`{\overline{r}}_{w}`$. So without loss of optimality, $`{\overline{r}}_{w}`$ can be taken as the effective maximal incoming rate demanded by $`w`$.
We model the traffic as fluid flows. Assume congestion control is exercised at each source node. That is the source node $`O\hspace{0pt}{(w)}`$ can control the rate at which $`w`$’s traffic comes into the network. Denote the actual admitted rate of $`w`$ by $`r_{w}`$, also referred to as the end-to-end flow rate of $`w`$. Thus, the rate of the rejected traffic of $`w`$, denoted by $`F_{w}`$, is $`{\overline{r}}_{w} - r_{w}`$. After the admitted flow rates are determined by the congestion control at the source nodes, the flows entering the network are routed on (potentially) multiple paths from the source to destination. Let $`f_{i\hspace{0pt}j}\hspace{0pt}{(w)}`$ denote the rate of session $`w`$ traffic routed through link $`(i,j)`$. The session flows satisfy the flow conservation constraints, i.e.,
<table>
<tbody>
<tr>
<td></td>
<td>∑j∈𝒪ifij(w)=rw,i=O(w),formulae-sequencesubscript𝑗subscript𝒪𝑖subscript𝑓𝑖𝑗𝑤subscript𝑟𝑤𝑖𝑂𝑤\sum_{j\in{\cal O}_{i}}f_{ij}(w)=r_{w},\quad i=O(w),</td>
<td></td>
<td>(6)</td>
</tr>
</tbody>
</table>
<table>
<tbody>
<tr>
<td></td>
<td>∑j∈𝒪ifij(w)=∑k∈𝒪ifki(w)=:ti(w),i≠O(w),D(w),\sum_{j\in{\cal O}_{i}}f_{ij}(w)=\sum_{k\in{\cal O}_{i}}f_{ki}(w)=:t_{i}(w),\quad i\neq O(w),D(w),</td>
<td></td>
<td>(7)</td>
</tr>
</tbody>
</table>
and
<table>
<tbody>
<tr>
<td></td>
<td>fij(w)=0,∀j∈𝒪i,i=D(w).formulae-sequencesubscript𝑓𝑖𝑗𝑤0formulae-sequencefor-all𝑗subscript𝒪𝑖𝑖𝐷𝑤f_{ij}(w)=0,\quad\forall j\in{\cal O}_{i},\quad i=D(w).</td>
<td></td>
<td>(8)</td>
</tr>
</tbody>
</table>
Here, $`t_{i}\hspace{0pt}{(w)}`$ denotes both the total incoming and outgoing flow rates of session $`w`$ at an intermediate node $`i`$. If we consider the rejected flow $`F_{w}`$ as being routed on a virtual “overflow” link from $`O\hspace{0pt}{(w)}`$ to $`D\hspace{0pt}{(w)}`$ \[[29](#bib.bib29), [30](#bib.bib30)\], the flow conservation constraint involving both real and virtual outgoing flows from the source node is given by
<table>
<tbody>
<tr>
<td></td>
<td>∑j∈𝒪ifij(w)+Fw=r¯w.subscript𝑗subscript𝒪𝑖subscript𝑓𝑖𝑗𝑤subscript𝐹𝑤subscript¯𝑟𝑤\sum_{j\in{\cal O}_{i}}f_{ij}(w)+F_{w}=\bar{r}_{w}.</td>
<td></td>
<td>(9)</td>
</tr>
</tbody>
</table>
Thus, later on we will incorporate congestion control in a unified framework with routing. Let the set of flow vectors $`(F_{w},{({f_{i\hspace{0pt}j}\hspace{0pt}{(w)}})}_{{(i,j)} \in \mathcal{L}})`$ satisfying ([7](#S5.E7 "In V-B Traffic Demands, Congestion Control, and Routing ‣ V Cross-layer Optimization in Multi-radio Multi-hop Wireless Networks ‣ Distributed Algorithms for Spectrum Allocation, Power Control, Routing, and Congestion Control in Wireless Networks"))-([9](#S5.E9 "In V-B Traffic Demands, Congestion Control, and Routing ‣ V Cross-layer Optimization in Multi-radio Multi-hop Wireless Networks ‣ Distributed Algorithms for Spectrum Allocation, Power Control, Routing, and Congestion Control in Wireless Networks")) be denoted by $`\mathcal{F}_{w}`$.
The total flow rate on link $`(i,j)`$ is $`F_{i\hspace{0pt}j} = {\sum_{w \in \mathcal{W}}{f_{i\hspace{0pt}j}\hspace{0pt}{(w)}}}`$. To route a flow of rate $`F_{i\hspace{0pt}j}`$ from $`i`$ to $`j`$, node $`i`$ can split the traffic onto all active sub-bands of link $`(i,j)`$ and transmit them simultaneously. Let the rate of flow assigned to sub-band $`q \in \mathcal{Q}_{i\hspace{0pt}j}`$ be $`F_{i\hspace{0pt}j}\hspace{0pt}{(q)}`$. We have $`F_{i\hspace{0pt}j} = {\sum_{q \in \mathcal{Q}_{i\hspace{0pt}j}}{F_{i\hspace{0pt}j}\hspace{0pt}{(q)}}}`$. Hence,
<table>
<tbody>
<tr>
<td></td>
<td>∑q∈𝒬ijFij(q)=∑w∈𝒲fij(w).subscript𝑞subscript𝒬𝑖𝑗subscript𝐹𝑖𝑗𝑞subscript𝑤𝒲subscript𝑓𝑖𝑗𝑤\sum_{q\in{\cal Q}_{ij}}F_{ij}(q)=\sum_{w\in{\cal W}}f_{ij}(w).</td>
<td></td>
<td>(10)</td>
</tr>
</tbody>
</table>
Note that the flow on each sub-band may consist of traffic of one or more sessions. However, the traffic split conserves the total rate of any session going through link $`(i,j)`$. Therefore, at the receiver end of the link, node $`j`$ collects each session $`w`$’s traffic of rate $`f_{i\hspace{0pt}j}\hspace{0pt}{(w)}`$.
One can think of the routing scheme as a two-step process. The first step determines inter-node routing, i.e., a feasible flow vector $`{(F_{w},{({f_{i\hspace{0pt}j}\hspace{0pt}{(w)}})}_{{(i,j)} \in \mathcal{L}})} \in \mathcal{F}_{w}`$ is found for every $`w \in \mathcal{W}`$, yielding $`{(F_{i\hspace{0pt}j})}_{{(i,j)} \in \mathcal{L}}`$. The second step amounts to intra-node routing: each node $`i`$ routes the flow $`F_{i\hspace{0pt}j}`$ onto the active sub-bands $`q \in \mathcal{Q}_{i\hspace{0pt}j}`$ available from $`i`$ to $`j`$.
###
V-C
Network Cost and Optimal Resource Allocation
The cost on each (real) link $`(i,j)`$ is made up of the costs $`D_{i\hspace{0pt}j}^{q}`$ incurred on its active sub-band(s) $`q \in \mathcal{Q}_{i\hspace{0pt}j}`$. Each $`D_{i\hspace{0pt}j}^{q}`$ is a function of $`x_{i\hspace{0pt}j}\hspace{0pt}{(q)}`$, the SINR of link $`(i,j)`$ on $`q`$, and $`F_{i\hspace{0pt}j}\hspace{0pt}{(q)}`$, the flow rate on the active sub-band $`q`$ of link $`(i,j)`$. Specifically, define
<table>
<tbody>
<tr>
<td></td>
<td>Dijq=D(xij(q),Fij(q)).superscriptsubscript𝐷𝑖𝑗𝑞𝐷subscript𝑥𝑖𝑗𝑞subscript𝐹𝑖𝑗𝑞D_{ij}^{q}=D(x_{ij}(q),F_{ij}(q)).</td>
<td></td>
</tr>
</tbody>
</table>
Assume that $`D\hspace{0pt}{(x,F)}`$ is decreasing and convex in $`x`$ for fixed $`F`$, and that it is increasing and convex in $`F`$ for fixed $`x`$. To impose the hard capacity constraint $`F < {C\hspace{0pt}{(x)}}`$, where $`C\hspace{0pt}{(x)}`$ denotes the link capacity achieved by having $`{S\hspace{0pt}I\hspace{0pt}N\hspace{0pt}R} = x`$, one can define $`{D\hspace{0pt}{(x,F)}} = \infty`$ for $`F \geq {C\hspace{0pt}{(x)}}`$. An example of such cost functions is
<sup>11</sup>
<sup>11</sup>11For modelling purposes, we assume that on the transmitter side of each link there is one queue for every active sub-band of that link. Specifically, the arrival rate to the queue of link $`(i,j)`$ on sub-band $`q`$ is $`F_{i\hspace{0pt}j}\hspace{0pt}{(q)}`$, the aggregate flow routed onto $`(i,j)`$ through $`q`$. The cost function $`{D\hspace{0pt}{(x,F)}} = \frac{F}{{C\hspace{0pt}{(x)}} - F}`$ gives the expected delay in an M/M/1 queue with arrival rate $`F`$ and service capacity $`C\hspace{0pt}{(x)}`$. In our present setting, $`F`$ is the sub-band specific link flow rate and $`C\hspace{0pt}{(x)}`$ is the sub-band specific link capacity. By the Kleinrock independence approximation and Jackson’s Theorem, the $`M/M/1`$ queue is a good approximation for the behavior of individual links on active sub-bands when the system involves Poisson stream arrivals at the entry points, a densely connected network, and moderate-to-heavy traffic load \[[31](#bib.bib31), [29](#bib.bib29)\].
$`{D\hspace{0pt}{(x,F)}} = \frac{F}{{C\hspace{0pt}{(x)}} - F}`$ where $`{C\hspace{0pt}{(x)}} = {R\hspace{0pt}{\log{({K\hspace{0pt}x})}}}`$ with constants $`{R,K} > 0`$.
<sup>12</sup>
<sup>12</sup>12The constants $`R,K`$ correspond to the bandwidth and processing gain of the system. The capacity formula $`{C\hspace{0pt}{(x)}} = {R\hspace{0pt}{\log{({K\hspace{0pt}x})}}}`$ is a good approximation for systems with a high processing gain, e.g. CDMA networks.
We associate an increasing and concave utility function $`U_{w}\hspace{0pt}{(r_{w})}`$, $`0 \leq r_{w} \leq {\overline{r}}_{w}`$, to each (elastic) session $`w`$. Thus, rejecting a flow of $`F_{w}`$ from the network incurs a utility loss of $`{U_{w}\hspace{0pt}{({\overline{r}}_{w})}} - {U_{w}\hspace{0pt}{({{\overline{r}}_{w} - F_{w}})}}`$. This utility loss can also be interpreted as the cost $`{D_{w}\hspace{0pt}{(F_{w})}} \triangleq {{U_{w}\hspace{0pt}{({\overline{r}}_{w})}} - {U_{w}\hspace{0pt}{({{\overline{r}}_{w} - F_{w}})}}}`$ on a virtual overflow link \[[29](#bib.bib29)\] with flow rate $`F_{w}`$. Note that by definition, $`D_{w}\hspace{0pt}{(F_{w})}`$ is increasing and convex in $`F_{w}`$.
We solve for the feasible flow and power variables that jointly maximize the net utility, i.e., total utility minus total cost on real links. With the above observation, the maximization of net utility is equivalent to the problem of minimizing the total network cost $`E \triangleq {{\sum_{{(i,j)} \in \mathcal{L}}D_{i\hspace{0pt}j}} + {\sum_{w \in \mathcal{W}}D_{w}}}`$ consisting of costs on both real and overflow links. Formally, the Multi-Radio Minimum Cost Resource Allocation (MCRA) problem is
<table>
<tbody>
<tr>
<td></td>
<td>minimize</td>
<td></td>
<td>∑w∈𝒲Dw(Fw)+∑(i,j)∈ℒ∑q∈𝒬ijD(xij(q),Fij(q))subscript𝑤𝒲subscript𝐷𝑤subscript𝐹𝑤subscript𝑖𝑗ℒsubscript𝑞subscript𝒬𝑖𝑗𝐷subscript𝑥𝑖𝑗𝑞subscript𝐹𝑖𝑗𝑞\displaystyle\sum_{w\in{\cal W}}D_{w}(F_{w})+\sum_{(i,j)\in{\cal L}}\sum_{q\in{\cal Q}_{ij}}D(x_{ij}(q),F_{ij}(q))</td>
<td></td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td>subject to</td>
<td></td>
<td>(4)and(Pij(q))(i,j)∈ℒ,q∈𝒬ij∈𝒫,italic-(4italic-)andsubscriptsubscript𝑃𝑖𝑗𝑞formulae-sequence𝑖𝑗ℒ𝑞subscript𝒬𝑖𝑗𝒫\displaystyle\eqref{eq:SINR}~{}\textrm{and}~{}(P_{ij}(q))_{(i,j)\in{\cal L},q\in{\cal Q}_{ij}}\in{\cal P},</td>
<td></td>
</tr>
<tr>
<td></td>
<td></td>
<td></td>
<td>(10)and(Fw,(fij(w))(i,j)∈ℒ)∈ℱw,∀w∈𝒲.formulae-sequenceitalic-(10italic-)andsubscript𝐹𝑤subscriptsubscript𝑓𝑖𝑗𝑤𝑖𝑗ℒsubscriptℱ𝑤for-all𝑤𝒲\displaystyle\eqref{eq:FlowEquation}~{}\textrm{and}~{}(F_{w},(f_{ij}(w))_{(i,j)\in{\cal L}})\in{\cal F}_{w},\;\forall w\in{\cal W}.</td>
<td></td>
</tr>
</tbody>
</table>
We will devise a distributed scheme that iteratively adjusts power control, routing and congestion control on a node-by-node basis so as to find the optimal solution to the MCRA problem.
###
V-D
Control Variables and Optimality Conditions
To update transmission powers, traffic routes, and traffic input rates, we define the following power control and routing variables. These optimization variables are node-based and local in the sense that they are independently controlled at individual nodes.
####
V-D1
Power Control Variables
Let $`P_{i}\hspace{0pt}{(q)}`$ be the total power used by node $`i`$ on its active sub-band $`q`$. Let $`P_{i\hspace{0pt}j}\hspace{0pt}{(q)}`$ be the power used by node $`i`$ on its outgoing link $`(i,j)`$ over an active sub-band $`q`$. Define
<table>
<tbody>
<tr>
<td></td>
<td>ρi(q)≜Pi(q)P¯i,q∈𝒪𝒞i,formulae-sequence≜subscript𝜌𝑖𝑞subscript𝑃𝑖𝑞subscript¯𝑃𝑖𝑞𝒪subscript𝒞𝑖\rho_{i}(q)\triangleq\frac{P_{i}(q)}{\bar{P}_{i}},\quad q\in{\cal OC}_{i},</td>
<td></td>
</tr>
</tbody>
</table>
constrained by
<table>
<tbody>
<tr>
<td></td>
<td>0≤ρi(q)≤1,∑q∈𝒪𝒞iρi(q)≤1.formulae-sequence0subscript𝜌𝑖𝑞1subscript𝑞𝒪subscript𝒞𝑖subscript𝜌𝑖𝑞1\quad 0\leq\rho_{i}(q)\leq 1,\quad\sum_{q\in{\cal OC}_{i}}\rho_{i}(q)\leq 1.</td>
<td></td>
<td>(11)</td>
</tr>
</tbody>
</table>
Also define
<table>
<tbody>
<tr>
<td></td>
<td>ηij(q)≜Pij(q)Pi(q),j∈𝒪i,q∈𝒬ij,formulae-sequence≜subscript𝜂𝑖𝑗𝑞subscript𝑃𝑖𝑗𝑞subscript𝑃𝑖𝑞formulae-sequence𝑗subscript𝒪𝑖𝑞subscript𝒬𝑖𝑗\eta_{ij}(q)\triangleq\frac{P_{ij}(q)}{P_{i}(q)},\quad j\in{\cal O}_{i},q\in{\cal Q}_{ij},</td>
<td></td>
</tr>
</tbody>
</table>
subject to constraints (let $`{\eta_{i\hspace{0pt}j}\hspace{0pt}{(q)}} \equiv 0`$ for $`q \notin \mathcal{Q}_{i\hspace{0pt}j}`$)
<table>
<tbody>
<tr>
<td></td>
<td>0≤ηij(q)≤1,∑j∈𝒪iηij(q)=1.formulae-sequence0subscript𝜂𝑖𝑗𝑞1subscript𝑗subscript𝒪𝑖subscript𝜂𝑖𝑗𝑞1\quad 0\leq\eta_{ij}(q)\leq 1,\quad\sum_{j\in{\cal O}_{i}}\eta_{ij}(q)=1.</td>
<td></td>
<td>(12)</td>
</tr>
</tbody>
</table>
With the above definitions, $`x_{i\hspace{0pt}j}\hspace{0pt}{(q)}`$ can be expressed in terms of the power control variables as
<table>
<tbody>
<tr>
<td></td>
<td>GijqP¯iρi(q)ηij(q)GijqP¯iρi(q)∑n≠jηin(q)+∑m≠iGmjqP¯mρm(q)+Njq.superscriptsubscript𝐺𝑖𝑗𝑞subscript¯𝑃𝑖subscript𝜌𝑖𝑞subscript𝜂𝑖𝑗𝑞superscriptsubscript𝐺𝑖𝑗𝑞subscript¯𝑃𝑖subscript𝜌𝑖𝑞subscript𝑛𝑗subscript𝜂𝑖𝑛𝑞subscript𝑚𝑖superscriptsubscript𝐺𝑚𝑗𝑞subscript¯𝑃𝑚subscript𝜌𝑚𝑞superscriptsubscript𝑁𝑗𝑞\frac{G_{ij}^{q}\bar{P}_{i}\rho_{i}(q)\eta_{ij}(q)}{G_{ij}^{q}\bar{P}_{i}\rho_{i}(q)\sum_{n\neq j}\eta_{in}(q)+\sum_{m\neq i}G_{mj}^{q}\bar{P}_{m}\rho_{m}(q)+N_{j}^{q}}.</td>
<td></td>
<td>(13)</td>
</tr>
</tbody>
</table>
We use $`I\hspace{0pt}N_{i\hspace{0pt}j}\hspace{0pt}{(q)}`$ to denote the total interference and noise power of link $`(i,j)`$ on sub-band $`q`$, i.e. the denominator of ([13](#S5.E13 "In V-D1 Power Control Variables ‣ V-D Control Variables and Optimality Conditions ‣ V Cross-layer Optimization in Multi-radio Multi-hop Wireless Networks ‣ Distributed Algorithms for Spectrum Allocation, Power Control, Routing, and Congestion Control in Wireless Networks")).
Define $`{\mathbf{ρ}}_{i} \triangleq {({\rho_{i}\hspace{0pt}{(q)}})}_{q \in {\mathcal{O}\hspace{0pt}\mathcal{C}_{i}}}`$ and $`{\mathbf{η}}_{i} \triangleq {({\eta_{i\hspace{0pt}j}\hspace{0pt}{(q)}})}_{{j \in \mathcal{O}_{i}},{q \in \mathcal{Q}_{i\hspace{0pt}j}}}`$. We now compute the partial derivative of the network cost $`E`$ with respect to each $`\rho_{i}\hspace{0pt}{(q)}`$ and $`\eta_{i\hspace{0pt}j}\hspace{0pt}{(q)}`$. These partial derivatives are useful for characterizing the optimality conditions and for iterative adjustment of the power control variables. We have
<table>
<tbody>
<tr>
<td></td>
<td>∂E∂ηij(q)=Pi(q){∑n∈𝒪i(Dxinq)′−Ginqxin(q)INin(q)+δηij(q)},𝐸subscript𝜂𝑖𝑗𝑞subscript𝑃𝑖𝑞subscript𝑛subscript𝒪𝑖superscriptsuperscriptsubscript𝐷subscript𝑥𝑖𝑛𝑞′superscriptsubscript𝐺𝑖𝑛𝑞subscript𝑥𝑖𝑛𝑞𝐼subscript𝑁𝑖𝑛𝑞𝛿subscript𝜂𝑖𝑗𝑞\frac{\partial E}{\partial\eta_{ij}(q)}=P_{i}(q)\left\{\sum_{n\in{\cal O}_{i}}(D_{x_{in}}^{q})^{\prime}\frac{-G_{in}^{q}x_{in}(q)}{IN_{in}(q)}+\delta\eta_{ij}(q)\right\},</td>
<td></td>
<td>(14)</td>
</tr>
</tbody>
</table>
where we have used shorthand notation $`{(D_{x_{i\hspace{0pt}n}}^{q})}'`$ for $`\partial{{D\hspace{0pt}{({x_{i\hspace{0pt}n}\hspace{0pt}{(q)}},{F_{i\hspace{0pt}n}\hspace{0pt}{(q)}})}}/{\partial{x_{i\hspace{0pt}n}\hspace{0pt}{(q)}}}}`$ and defined
<table>
<tbody>
<tr>
<td></td>
<td>δηij(q)≜(Dxijq)′Gijq(1+xij(q))INij(q).≜𝛿subscript𝜂𝑖𝑗𝑞superscriptsuperscriptsubscript𝐷subscript𝑥𝑖𝑗𝑞′superscriptsubscript𝐺𝑖𝑗𝑞1subscript𝑥𝑖𝑗𝑞𝐼subscript𝑁𝑖𝑗𝑞\delta\eta_{ij}(q)\triangleq(D_{x_{ij}}^{q})^{\prime}\frac{G_{ij}^{q}(1+x_{ij}(q))}{IN_{ij}(q)}.</td>
<td></td>
<td>(15)</td>
</tr>
</tbody>
</table>
Note that $`\delta\hspace{0pt}\eta_{i\hspace{0pt}j}\hspace{0pt}{(q)}`$ as well as $`\partial{E/{\partial{\eta_{i\hspace{0pt}j}\hspace{0pt}{(q)}}}}`$ involves only local measures of $`i`$. The partial derivative of $`E`$ with respect to $`\rho_{i}\hspace{0pt}{(q)}`$ is given by
<table>
<tbody>
<tr>
<td></td>
<td>δρi(q)≜∂E∂ρi(q)=P¯i{∑n∈𝒩GinqMSGnq+∑j∈𝒪iδηij(q)ηij(q)},≜𝛿subscript𝜌𝑖𝑞𝐸subscript𝜌𝑖𝑞subscript¯𝑃𝑖subscript𝑛𝒩superscriptsubscript𝐺𝑖𝑛𝑞𝑀𝑆superscriptsubscript𝐺𝑛𝑞subscript𝑗subscript𝒪𝑖𝛿subscript𝜂𝑖𝑗𝑞subscript𝜂𝑖𝑗𝑞\delta\rho_{i}(q)\triangleq\frac{\partial E}{\partial\rho_{i}(q)}=\bar{P}_{i}\left\{\sum_{n\in{\cal N}}G_{in}^{q}MSG_{n}^{q}+\sum_{j\in{\cal O}_{i}}{\delta\eta_{ij}(q)\eta_{ij}(q)}\right\},</td>
<td></td>
<td>(16)</td>
</tr>
</tbody>
</table>
where
<table>
<tbody>
<tr>
<td></td>
<td>MSGnq≜∑m∈𝒪n(Dxmnq)′−GmnqPmn(q)INmn(q)2≜𝑀𝑆superscriptsubscript𝐺𝑛𝑞subscript𝑚subscript𝒪𝑛superscriptsuperscriptsubscript𝐷subscript𝑥𝑚𝑛𝑞′superscriptsubscript𝐺𝑚𝑛𝑞subscript𝑃𝑚𝑛𝑞𝐼subscript𝑁𝑚𝑛superscript𝑞2MSG_{n}^{q}\triangleq\sum_{m\in{\cal O}_{n}}(D_{x_{mn}}^{q})^{\prime}\frac{-G_{mn}^{q}P_{mn}(q)}{IN_{mn}(q)^{2}}</td>
<td></td>
<td>(17)</td>
</tr>
</tbody>
</table>
is the message that node $`n`$ needs to send to any other node, say $`i`$, which is active on $`q`$ in order for $`i`$ to compute $`\delta\hspace{0pt}\rho_{i}\hspace{0pt}{(q)}`$ via ([16](#S5.E16 "In V-D1 Power Control Variables ‣ V-D Control Variables and Optimality Conditions ‣ V Cross-layer Optimization in Multi-radio Multi-hop Wireless Networks ‣ Distributed Algorithms for Spectrum Allocation, Power Control, Routing, and Congestion Control in Wireless Networks")). Specifically, the message exchange works as follows.
*Power Control Message Exchange Protocol:* Let each node $`n`$ keep an up-to-date measure $`M\hspace{0pt}S\hspace{0pt}G_{n}^{q}`$ for each sub-band $`q`$, where $`M\hspace{0pt}S\hspace{0pt}G_{n}^{q}`$ is derived by assembling the measures
<table>
<tbody>
<tr>
<td></td>
<td>(Dxmnq)′−GmnqPmn(q)INmn(q)2=(Dxmnq)′−xmn(q)2GmnqPmn(q)superscriptsuperscriptsubscript𝐷subscript𝑥𝑚𝑛𝑞′superscriptsubscript𝐺𝑚𝑛𝑞subscript𝑃𝑚𝑛𝑞𝐼subscript𝑁𝑚𝑛superscript𝑞2superscriptsuperscriptsubscript𝐷subscript𝑥𝑚𝑛𝑞′subscript𝑥𝑚𝑛superscript𝑞2superscriptsubscript𝐺𝑚𝑛𝑞subscript𝑃𝑚𝑛𝑞(D_{x_{mn}}^{q})^{\prime}\frac{-G_{mn}^{q}P_{mn}(q)}{IN_{mn}(q)^{2}}=(D_{x_{mn}}^{q})^{\prime}\frac{-x_{mn}(q)^{2}}{G_{mn}^{q}P_{mn}(q)}</td>
<td></td>
</tr>
</tbody>
</table>
on all its active incoming links $`(m,n)`$ on $`q`$, and summing them up (cf. ([17](#S5.E17 "In V-D1 Power Control Variables ‣ V-D Control Variables and Optimality Conditions ‣ V Cross-layer Optimization in Multi-radio Multi-hop Wireless Networks ‣ Distributed Algorithms for Spectrum Allocation, Power Control, Routing, and Congestion Control in Wireless Networks"))). Note that $`M\hspace{0pt}S\hspace{0pt}G_{n}^{q}`$ is nonnegative and it is zero if $`n`$ has no active incoming link on $`q`$. If $`{M\hspace{0pt}S\hspace{0pt}G_{n}^{q}} > 0`$, node $`n`$ broadcasts it to the whole network via a flooding protocol. This control message generating process is illustrated by Figure [3](#S5.F3 "Figure 3 ‣ V-D1 Power Control Variables ‣ V-D Control Variables and Optimality Conditions ‣ V Cross-layer Optimization in Multi-radio Multi-hop Wireless Networks ‣ Distributed Algorithms for Spectrum Allocation, Power Control, Routing, and Congestion Control in Wireless Networks").
[Refer to caption](/html/0704.3746/assets/x3.png)
Figure 3: Power control message generation.
If $`q \in {\mathcal{O}\hspace{0pt}\mathcal{C}_{i}}`$, node $`i`$ collects $`M\hspace{0pt}S\hspace{0pt}G_{n}^{q}`$ and processes it according to the following rule. Node $`i`$ multiplies $`M\hspace{0pt}S\hspace{0pt}G_{n}^{q}`$ with the path gain $`G_{i\hspace{0pt}n}^{q}`$.
<sup>13</sup>
<sup>13</sup>13In a symmetric duplex channel, $`G_{i\hspace{0pt}n}^{q} \approx G_{n\hspace{0pt}i}^{q}`$, and node $`i`$ may use its own measurement of $`G_{n\hspace{0pt}i}^{q}`$ in the place of $`G_{i\hspace{0pt}n}^{q}`$. Otherwise, it will need channel feedback from node $`n`$ to calculate $`G_{i\hspace{0pt}n}^{q}`$.
It further adds the product to the value of local measure $`{{\delta\hspace{0pt}\eta_{i\hspace{0pt}n}\hspace{0pt}{(q)}} \cdot \eta_{i\hspace{0pt}n}}\hspace{0pt}{(q)}`$ if the origin $`n`$ of the message is a neighbor of $`i`$ and $`(i,n)`$ is active on sub-band $`q`$. Finally, node $`i`$ adds up all the processed messages, and this sum multiplied by $`{\overline{P}}_{i}`$ equals $`\delta\hspace{0pt}\rho_{i}\hspace{0pt}{(q)}`$. Note that this protocol requires only one message from each node on each sub-band $`q`$.
<sup>14</sup>
<sup>14</sup>14To be more precise, only the nodes having at least one active incoming link on a certain sub-band need to provide a message for that sub-band.
Moreover in practice, node $`i`$ can ignore the messages generated by distant nodes, because they contribute very little to $`\delta\hspace{0pt}\rho_{i}\hspace{0pt}{(q)}`$ due to the negligible multiplicative factor $`G_{i\hspace{0pt}n}^{q}`$ on $`M\hspace{0pt}S\hspace{0pt}G_{n}^{q}`$ when $`i`$ and $`n`$ are far apart (cf. ([16](#S5.E16 "In V-D1 Power Control Variables ‣ V-D Control Variables and Optimality Conditions ‣ V Cross-layer Optimization in Multi-radio Multi-hop Wireless Networks ‣ Distributed Algorithms for Spectrum Allocation, Power Control, Routing, and Congestion Control in Wireless Networks"))).
####
V-D2
Routing Variables
Routing variables were first introduced by Gallager \[[32](#bib.bib32)\] for wireline network routing problems. Here, we define routing variables in a similar fashion. In addition to inter-node routing, however, routing variables here also perform the function of congestion control and intra-node routing.
Recall that congestion control is equivalent to routing a portion of traffic demand on a virtual overflow link directly from the source to the destination. Let $`i`$ be the source node of session $`w`$, define overflow routing variable
<table>
<tbody>
<tr>
<td></td>
<td>ϕw≜Fwr¯w,≜subscriptitalic-ϕ𝑤subscript𝐹𝑤subscript¯𝑟𝑤\phi_{w}\triangleq\frac{F_{w}}{\bar{r}_{w}},</td>
<td></td>
</tr>
</tbody>
</table>
which is constrained by $`0 \leq \phi_{w} \leq 1`$. The overflow rate is then controlled by $`\phi_{w}`$ as $`F_{w} = {{\overline{r}}_{w}\hspace{0pt}\phi_{w}}`$, and the end-to-end flow rate is given by $`r_{w} = {{\overline{r}}_{w}\hspace{0pt}{({1 - \phi_{w}})}}`$. Routing variables associated with a real link $`(i,j)`$ are defined by
<table>
<tbody>
<tr>
<td></td>
<td>ϕij(w)≜fij(w)ti(w),≜subscriptitalic-ϕ𝑖𝑗𝑤subscript𝑓𝑖𝑗𝑤subscript𝑡𝑖𝑤\phi_{ij}(w)\triangleq\frac{f_{ij}(w)}{t_{i}(w)},</td>
<td></td>
</tr>
</tbody>
</table>
which gives the fraction of the incoming flow of session $`w`$ at node $`i`$ that is routed onto link $`(i,j)`$. At any node $`i`$ except the destination, $`\phi_{i\hspace{0pt}j}\hspace{0pt}{(w)}`$ of all $`j \in \mathcal{O}_{i}`$ satisfy
<table>
<tbody>
<tr>
<td></td>
<td>0≤ϕij(w)≤1and∑j∈𝒪iϕij(w)=1.formulae-sequence0subscriptitalic-ϕ𝑖𝑗𝑤1andsubscript𝑗subscript𝒪𝑖subscriptitalic-ϕ𝑖𝑗𝑤10\leq\phi_{ij}(w)\leq 1\quad\textrm{and}\quad\sum_{j\in{\cal O}_{i}}\phi_{ij}(w)=1.</td>
<td></td>
<td>(18)</td>
</tr>
</tbody>
</table>
It is easy to see that the routing variables $`\phi_{w},{\phi_{i\hspace{0pt}j}\hspace{0pt}{(w)}}`$ of all $`{(i,j)} \in \mathcal{L}`$ and $`w \in \mathcal{W}`$ uniquely determine the inter-node flow patterns $`(F_{w},{({f_{i\hspace{0pt}j}\hspace{0pt}{(w)}})}_{{(i,j)} \in \mathcal{L}})`$ of all sessions $`w`$, and hence the total flow rate on links $`F_{i\hspace{0pt}j} = {\sum_{w}{f_{i\hspace{0pt}j}\hspace{0pt}{(w)}}}`$. Now let the flow allocation on active sub-bands be specified by the intra-node routing variables defined by
<table>
<tbody>
<tr>
<td></td>
<td>μij(q)≜Fij(q)Fij,q∈𝒬ij.formulae-sequence≜subscript𝜇𝑖𝑗𝑞subscript𝐹𝑖𝑗𝑞subscript𝐹𝑖𝑗𝑞subscript𝒬𝑖𝑗\mu_{ij}(q)\triangleq\frac{F_{ij}(q)}{F_{ij}},\quad q\in{\cal Q}_{ij}.</td>
<td></td>
</tr>
</tbody>
</table>
For any link $`(i,j)`$, we must have
<table>
<tbody>
<tr>
<td></td>
<td>0≤μij(q)≤1and∑q∈𝒬ijμij(q)=1.formulae-sequence0subscript𝜇𝑖𝑗𝑞1andsubscript𝑞subscript𝒬𝑖𝑗subscript𝜇𝑖𝑗𝑞10\leq\mu_{ij}(q)\leq 1\quad\textrm{and}\quad\sum_{q\in{\cal Q}_{ij}}\mu_{ij}(q)=1.</td>
<td></td>
<td>(19)</td>
</tr>
</tbody>
</table>
The link flow rate on an active sub-band is therefore given by
<table>
<tbody>
<tr>
<td></td>
<td>Fij(q)=μij(q)∑w∈𝒲ti(w)ϕij(w),q∈𝒬ij.formulae-sequencesubscript𝐹𝑖𝑗𝑞subscript𝜇𝑖𝑗𝑞subscript𝑤𝒲subscript𝑡𝑖𝑤subscriptitalic-ϕ𝑖𝑗𝑤𝑞subscript𝒬𝑖𝑗F_{ij}(q)=\mu_{ij}(q)\sum_{w\in{\cal W}}t_{i}(w)\phi_{ij}(w),\quad q\in{\cal Q}_{ij}.</td>
<td></td>
</tr>
</tbody>
</table>
For notational brevity, denote by $`\mathbf{\phi}_{i}\hspace{0pt}{(w)}`$ the vector $`{({\phi_{i\hspace{0pt}j}\hspace{0pt}{(w)}})}_{j \in \mathcal{O}_{i}}`$ if $`i \neq {{O\hspace{0pt}{(w)}},{D\hspace{0pt}{(w)}}}`$, or $`(\phi_{w},{({\phi_{i\hspace{0pt}j}\hspace{0pt}{(w)}})}_{j \in \mathcal{O}_{i}})`$ if $`i = {O\hspace{0pt}{(w)}}`$. Also denote by $`{\mathbf{μ}}_{i}`$ the vector $`{({\mu_{i\hspace{0pt}j}\hspace{0pt}{(q)}})}_{{j \in \mathcal{O}_{i}},{q \in \mathcal{Q}_{i\hspace{0pt}j}}}`$. We illustrate the use of the above routing variables by looking at a source node $`i`$ and one of its outgoing links $`(i,j)`$ in Figures [4](#S5.F4 "Figure 4 ‣ V-D2 Routing Variables ‣ V-D Control Variables and Optimality Conditions ‣ V Cross-layer Optimization in Multi-radio Multi-hop Wireless Networks ‣ Distributed Algorithms for Spectrum Allocation, Power Control, Routing, and Congestion Control in Wireless Networks") and [5](#S5.F5 "Figure 5 ‣ V-D2 Routing Variables ‣ V-D Control Variables and Optimality Conditions ‣ V Cross-layer Optimization in Multi-radio Multi-hop Wireless Networks ‣ Distributed Algorithms for Spectrum Allocation, Power Control, Routing, and Congestion Control in Wireless Networks"). In Figure [4](#S5.F4 "Figure 4 ‣ V-D2 Routing Variables ‣ V-D Control Variables and Optimality Conditions ‣ V Cross-layer Optimization in Multi-radio Multi-hop Wireless Networks ‣ Distributed Algorithms for Spectrum Allocation, Power Control, Routing, and Congestion Control in Wireless Networks"), the overflow link is marked by a solid line with hollow arrow.
[Refer to caption](/html/0704.3746/assets/x4.png)
Figure 4: Overflow and inter-node routing at a source node.
In Figure [5](#S5.F5 "Figure 5 ‣ V-D2 Routing Variables ‣ V-D Control Variables and Optimality Conditions ‣ V Cross-layer Optimization in Multi-radio Multi-hop Wireless Networks ‣ Distributed Algorithms for Spectrum Allocation, Power Control, Routing, and Congestion Control in Wireless Networks"), we assume that link $`(i,j)`$ is active on two sub-bands $`q_{1}`$ and $`q_{2}`$, and that the inter-node flow rate from $`i`$ to $`j`$ is $`F_{i\hspace{0pt}j}`$.
[Refer to caption](/html/0704.3746/assets/x5.png)
Figure 5: Intra-node routing on link $`(i,j)`$.
We now compute the partial derivative of $`E`$ with respect to those routing variables. First we have the derivative of $`E`$ with respect to the overflow routing variables
<table>
<tbody>
<tr>
<td></td>
<td>∂E∂ϕw=r¯w(Dw′−∂E∂tw),𝐸subscriptitalic-ϕ𝑤subscript¯𝑟𝑤superscriptsubscript𝐷𝑤′𝐸subscript𝑡𝑤\frac{\partial E}{\partial\phi_{w}}=\bar{r}_{w}\left(D_{w}^{\prime}-\frac{\partial E}{\partial t_{w}}\right),</td>
<td></td>
</tr>
</tbody>
</table>
where $`\frac{\partial E}{\partial t_{w}}`$ is the marginal cost of increasing the end-to-end flow rate of session $`w`$ with all other variables held constant. This is a special case ($`i = {O\hspace{0pt}{(w)}}`$) of $`\frac{\partial E}{\partial{t_{i}\hspace{0pt}{(w)}}}`$, which is the marginal cost of increasing the incoming flow rate of $`w`$ at node $`i`$ while keeping all other variables constant. These marginal costs are computed recursively as in \[[32](#bib.bib32)\]:
<table>
<tbody>
<tr>
<td></td>
<td>∂E∂ti(w)=0,if i=D(w),formulae-sequence𝐸subscript𝑡𝑖𝑤0if 𝑖𝐷𝑤\frac{\partial E}{\partial t_{i}(w)}=0,\quad\textrm{if~{}}i=D(w),</td>
<td></td>
</tr>
</tbody>
</table>
and for $`i \neq {D\hspace{0pt}{(w)}}`$,
<table>
<tbody>
<tr>
<td></td>
<td>∂E∂ti(w)𝐸subscript𝑡𝑖𝑤\displaystyle\frac{\partial E}{\partial t_{i}(w)}</td>
<td>=\displaystyle=</td>
<td>∑j∈𝒪iϕij(w)(∑q∈𝒬ijμij(q)DFijq′+∂E∂tj(w))subscript𝑗subscript𝒪𝑖subscriptitalic-ϕ𝑖𝑗𝑤subscript𝑞subscript𝒬𝑖𝑗subscript𝜇𝑖𝑗𝑞superscriptsuperscriptsubscript𝐷subscript𝐹𝑖𝑗𝑞′𝐸subscript𝑡𝑗𝑤\displaystyle\sum_{j\in{\cal O}_{i}}\phi_{ij}(w)\left(\sum_{q\in{\cal Q}_{ij}}\mu_{ij}(q){D_{F_{ij}}^{q}}^{\prime}+\frac{\partial E}{\partial t_{j}(w)}\right)</td>
<td></td>
<td rowspan="2">(20)</td>
</tr>
<tr>
<td></td>
<td></td>
<td>=\displaystyle=</td>
<td>∑j∈𝒪iϕij(w)δϕij(w),subscript𝑗subscript𝒪𝑖subscriptitalic-ϕ𝑖𝑗𝑤𝛿subscriptitalic-ϕ𝑖𝑗𝑤\displaystyle\sum_{j\in{\cal O}_{i}}\phi_{ij}(w)\delta\phi_{ij}(w),</td>
<td></td>
</tr>
</tbody>
</table>
where we have used shorthand notation $`{}D_{F_{i\hspace{0pt}j}}^{q'}`$ for $`\partial{{D\hspace{0pt}{({x_{i\hspace{0pt}j}\hspace{0pt}{(q)}},{F_{i\hspace{0pt}j}\hspace{0pt}{(q)}})}}/{\partial{F_{i\hspace{0pt}j}\hspace{0pt}{(q)}}}}`$ and defined marginal routing cost indicators
<table>
<tbody>
<tr>
<td></td>
<td>δϕij(w)=∑q∈𝒬ijμij(q)DFijq′+∂E∂tj(w).𝛿subscriptitalic-ϕ𝑖𝑗𝑤subscript𝑞subscript𝒬𝑖𝑗subscript𝜇𝑖𝑗𝑞superscriptsuperscriptsubscript𝐷subscript𝐹𝑖𝑗𝑞′𝐸subscript𝑡𝑗𝑤\delta\phi_{ij}(w)=\sum_{q\in{\cal Q}_{ij}}\mu_{ij}(q){D_{F_{ij}}^{q}}^{\prime}+\frac{\partial E}{\partial t_{j}(w)}.</td>
<td></td>
<td>(21)</td>
</tr>
</tbody>
</table>
Using $`\delta\hspace{0pt}\phi_{i\hspace{0pt}j}\hspace{0pt}{(w)}`$, we can easily write out the partial derivative of $`E`$ in $`\phi_{i\hspace{0pt}j}\hspace{0pt}{(w)}`$ as
<table>
<tbody>
<tr>
<td></td>
<td>∂E∂ϕij(w)=ti(w)δϕij(w).𝐸subscriptitalic-ϕ𝑖𝑗𝑤subscript𝑡𝑖𝑤𝛿subscriptitalic-ϕ𝑖𝑗𝑤\frac{\partial E}{\partial\phi_{ij}(w)}=t_{i}(w)\delta\phi_{ij}(w).</td>
<td></td>
<td>(22)</td>
</tr>
</tbody>
</table>
Finally, the partial derivative of $`E`$ in $`\mu_{i\hspace{0pt}j}\hspace{0pt}{(q)}`$ is given by
<table>
<tbody>
<tr>
<td></td>
<td>∂E∂μij(q)=FijDFijq′.𝐸subscript𝜇𝑖𝑗𝑞subscript𝐹𝑖𝑗superscriptsuperscriptsubscript𝐷subscript𝐹𝑖𝑗𝑞′\frac{\partial E}{\partial\mu_{ij}(q)}=F_{ij}{D_{F_{ij}}^{q}}^{\prime}.</td>
<td></td>
<td>(23)</td>
</tr>
</tbody>
</table>
Notice that $`\frac{\partial E}{\partial{\mu_{i\hspace{0pt}j}\hspace{0pt}{(q)}}}`$ depends on local measures only. However, $`\frac{\partial E}{\partial\phi_{w}}`$ and $`\frac{\partial E}{\partial{\phi_{i\hspace{0pt}j}\hspace{0pt}{(w)}}}`$ are both tied to the value of $`\frac{\partial E}{\partial{t_{j}\hspace{0pt}{(w)}}}`$ of $`i`$’s downstream
<sup>15</sup>
<sup>15</sup>15Given a routing configuration $`{\{{\phi_{i\hspace{0pt}j}\hspace{0pt}{(w)}}\}}_{{(i,j)} \in \mathcal{L}}`$ of session $`w`$, node $`j`$ is said to be downstream to $`i`$ if there exists a path $`(i,j_{1})`$, $`(j_{1},j_{2})`$, $`\cdots`$, $`(j_{n},j)`$ such that $`\phi_{i\hspace{0pt}j_{1}}`$, $`\phi_{j_{1}\hspace{0pt}j_{2}}`$, $`\cdots`$, $`\phi_{j_{n}\hspace{0pt}j}`$ are all positive. We say $`i`$ is upstream to $`j`$ with respect to session $`w`$ if $`j`$ is downstream to $`i`$.
neighbors $`j`$, which in turn depends on $`\frac{\partial E}{\partial{t_{k}\hspace{0pt}{(w)}}}`$ of $`j`$’s downstream neighbors $`k`$. Thus, we need a sequential message passing of the marginal routing costs, from the destination upstream to the source, to permit every node to acquire the partial derivatives in its local routing variables.
*Routing Message Exchange Protocol:* In \[[32](#bib.bib32)\], the rules for propagating the marginal routing cost information are specified. In order for node $`i`$ to evaluate the terms $`\delta\hspace{0pt}\phi_{i\hspace{0pt}j}\hspace{0pt}{(w)}`$ in ([21](#S5.E21 "In V-D2 Routing Variables ‣ V-D Control Variables and Optimality Conditions ‣ V Cross-layer Optimization in Multi-radio Multi-hop Wireless Networks ‣ Distributed Algorithms for Spectrum Allocation, Power Control, Routing, and Congestion Control in Wireless Networks")), it needs to collect local measures $`{}D_{F_{i\hspace{0pt}j}}^{q'}`$ from all $`q \in \mathcal{Q}_{i\hspace{0pt}j}`$ as well as reports of marginal costs $`\partial{D/{\partial{r_{j}\hspace{0pt}{(w)}}}}`$ from its next-hop neighbors $`j \in \mathcal{O}_{i}`$, for all traversing sessions $`w \in \mathcal{W}`$. Moreover, it is responsible for calculating its own measure of marginal cost $`\frac{\partial D}{\partial{t_{i}\hspace{0pt}{(w)}}}`$ with respect to every session $`w`$ according to ([20](#S5.E20 "In V-D2 Routing Variables ‣ V-D Control Variables and Optimality Conditions ‣ V Cross-layer Optimization in Multi-radio Multi-hop Wireless Networks ‣ Distributed Algorithms for Spectrum Allocation, Power Control, Routing, and Congestion Control in Wireless Networks")), and then providing the measure to its upstream neighbors with respect to the session $`w`$. The sequential message passing terminates if and only if the routing pattern of the session contains no loops, which is guaranteed by the blocked-node-set technique developed in \[[32](#bib.bib32), [33](#bib.bib33)\].
####
V-D3
Conditions for Optimality
The power and routing configuration that solves the MCRA problem can be characterized in terms of the marginal power and routing costs as follows.
######
Theorem 1
For an instance of optimization variables $`{\{{\mathbf{ρ}}_{i},{\mathbf{η}}_{i},{\{{\mathbf{\phi}_{i}\hspace{0pt}{(w)}}\}}_{w \in \mathcal{W}},{\mathbf{μ}}_{i}\}}_{i \in \mathcal{N}}`$ to be optimal, the following conditions are necessary: For all $`i \in \mathcal{N}`$ and $`w \in \mathcal{W}`$ with $`{t_{i}\hspace{0pt}{(w)}} > 0`$,
<table>
<tbody>
<tr>
<td></td>
<td>∂E∂ti(w)=minj∈𝒪i{minq∈𝒬ij[DFijq′]+∂E∂tj(w)}.𝐸subscript𝑡𝑖𝑤subscript𝑗subscript𝒪𝑖subscript𝑞subscript𝒬𝑖𝑗superscriptsuperscriptsubscript𝐷subscript𝐹𝑖𝑗𝑞′𝐸subscript𝑡𝑗𝑤\frac{\partial E}{\partial t_{i}(w)}=\min_{j\in{\cal O}_{i}}\left\{\min_{q\in{\cal Q}_{ij}}\left[{D_{F_{ij}}^{q}}^{\prime}\right]+\frac{\partial E}{\partial t_{j}(w)}\right\}.</td>
<td></td>
<td>(24)</td>
</tr>
</tbody>
</table>
Moreover, for all $`w \in \mathcal{W}`$,
<table>
<tbody>
<tr>
<td></td>
<td>∂E∂tw{≤Dw′,if ϕw=0,=Dw′,if 0<ϕw<1,≥Dw′,if ϕw=1.𝐸subscript𝑡𝑤casesabsentsuperscriptsubscript𝐷𝑤′if subscriptitalic-ϕ𝑤0absentsuperscriptsubscript𝐷𝑤′if 0subscriptitalic-ϕ𝑤1absentsuperscriptsubscript𝐷𝑤′if subscriptitalic-ϕ𝑤1\frac{\partial E}{\partial t_{w}}\left\{\begin{array}[]{ll}\leq D_{w}^{\prime},&\textrm{if~{}}\phi_{w}=0,\\ =D_{w}^{\prime},&\textrm{if~{}}0<\phi_{w}<1,\\ \geq D_{w}^{\prime},&\textrm{if~{}}\phi_{w}=1.\end{array}\right.</td>
<td></td>
<td>(25)</td>
</tr>
</tbody>
</table>
For all $`i \in \mathcal{N}`$ and $`q \in {\mathcal{O}\hspace{0pt}\mathcal{C}_{i}}`$, there exists a $`\lambda_{i}`$ such that
<table>
<tbody>
<tr>
<td></td>
<td>δρi(q){=λi,ifρi(q)>0,≥λi,ifρi(q)=0,𝛿subscript𝜌𝑖𝑞casesabsentsubscript𝜆𝑖ifsubscript𝜌𝑖𝑞0absentsubscript𝜆𝑖ifsubscript𝜌𝑖𝑞0\delta\rho_{i}(q)\left\{\begin{array}[]{ll}=\lambda_{i},&\textrm{if}~{}\rho_{i}(q)>0,\\ \geq\lambda_{i},&\textrm{if}~{}\rho_{i}(q)=0,\end{array}\right.</td>
<td></td>
<td>(26)</td>
</tr>
</tbody>
</table>
and the constant $`\lambda_{i}`$ satisfies
<table>
<tbody>
<tr>
<td></td>
<td>λi{≥0,if ∑q∈𝒪𝒞iρi(q)=0,=0,if 0<∑q∈𝒪𝒞iρi(q)<1,≤0,if ∑q∈𝒪𝒞iρi(q)=1.subscript𝜆𝑖casesabsent0if subscript𝑞𝒪subscript𝒞𝑖subscript𝜌𝑖𝑞0absent0if 0subscript𝑞𝒪subscript𝒞𝑖subscript𝜌𝑖𝑞1absent0if subscript𝑞𝒪subscript𝒞𝑖subscript𝜌𝑖𝑞1\lambda_{i}\left\{\begin{array}[]{ll}\geq 0,&\textrm{if~{}}\sum_{q\in{\cal OC}_{i}}\rho_{i}(q)=0,\\ =0,&\textrm{if~{}}0<\sum_{q\in{\cal OC}_{i}}\rho_{i}(q)<1,\\ \leq 0,&\textrm{if~{}}\sum_{q\in{\cal OC}_{i}}\rho_{i}(q)=1.\end{array}\right.</td>
<td></td>
<td>(27)</td>
</tr>
</tbody>
</table>
Furthermore, for $`q \in {\mathcal{O}\hspace{0pt}\mathcal{C}_{i}}`$ such that $`{\rho_{i}\hspace{0pt}{(q)}} > 0`$, there exists a constant $`\gamma_{i}\hspace{0pt}{(q)}`$ such that for all $`j \in \mathcal{O}_{i}`$ with $`q \in \mathcal{Q}_{i\hspace{0pt}j}`$,
<table>
<tbody>
<tr>
<td></td>
<td>δηij(q){=γi(q),ifηij(q)>0,≥γi(q),ifηij(q)=0.𝛿subscript𝜂𝑖𝑗𝑞casesabsentsubscript𝛾𝑖𝑞ifsubscript𝜂𝑖𝑗𝑞0absentsubscript𝛾𝑖𝑞ifsubscript𝜂𝑖𝑗𝑞0\delta\eta_{ij}(q)\left\{\begin{array}[]{ll}=\gamma_{i}(q),&\textrm{if}~{}\eta_{ij}(q)>0,\\ \geq\gamma_{i}(q),&\textrm{if}~{}\eta_{ij}(q)=0.\end{array}\right.</td>
<td></td>
<td>(28)</td>
</tr>
</tbody>
</table>
We sketch the proof of Theorem [1](#Thmtheorem1 "Theorem 1 ‣ V-D3 Conditions for Optimality ‣ V-D Control Variables and Optimality Conditions ‣ V Cross-layer Optimization in Multi-radio Multi-hop Wireless Networks ‣ Distributed Algorithms for Spectrum Allocation, Power Control, Routing, and Congestion Control in Wireless Networks") in Appendix [](#A0.SS3 "-C Proof of Theorem 1 ‣ VII Conclusion ‣ Distributed Algorithms for Spectrum Allocation, Power Control, Routing, and Congestion Control in Wireless Networks")
-C
. First order conditions as in ([24](#S5.E24 "In Theorem 1 ‣ V-D3 Conditions for Optimality ‣ V-D Control Variables and Optimality Conditions ‣ V Cross-layer Optimization in Multi-radio Multi-hop Wireless Networks ‣ Distributed Algorithms for Spectrum Allocation, Power Control, Routing, and Congestion Control in Wireless Networks"))-([28](#S5.E28 "In Theorem 1 ‣ V-D3 Conditions for Optimality ‣ V-D Control Variables and Optimality Conditions ‣ V Cross-layer Optimization in Multi-radio Multi-hop Wireless Networks ‣ Distributed Algorithms for Spectrum Allocation, Power Control, Routing, and Congestion Control in Wireless Networks")) are in general only necessary for a configuration to be optimal. However, they are further sufficient if the link cost function $`D\hspace{0pt}{(x,F)}`$ has certain convexity properties as we discuss next.
Let $`x`$ generically represent the SINR of a link on one of its active sub-bands. By ([4](#S5.E4 "In V-A Interference Limited Transmissions and Node Power Constraints ‣ V Cross-layer Optimization in Multi-radio Multi-hop Wireless Networks ‣ Distributed Algorithms for Spectrum Allocation, Power Control, Routing, and Congestion Control in Wireless Networks")), $`x`$ is a function of the power variables of all active links on the same sub-band, e.g. $`{x_{i\hspace{0pt}j}\hspace{0pt}{(q)}} = {h_{i\hspace{0pt}j}^{q}\hspace{0pt}{({({P_{m\hspace{0pt}n}\hspace{0pt}{(q)}})}_{{(m,n)} \in \mathcal{L}_{q}})}}`$. Thus, $`D_{i\hspace{0pt}j}^{q}`$ is a function of the vector $`{{\mathbf{P}}\hspace{0pt}{(q)}} \triangleq {({P_{m\hspace{0pt}n}\hspace{0pt}{(q)}})}_{{(m,n)} \in \mathcal{L}_{q}}`$ and $`F_{i\hspace{0pt}j}\hspace{0pt}{(q)}`$. It turns out that a characterization of the sufficient conditions for optimality requires $`D_{i\hspace{0pt}j}^{q}`$ to be jointly convex in $`{\mathbf{P}}\hspace{0pt}{(q)}`$ and $`F_{i\hspace{0pt}j}\hspace{0pt}{(q)}`$. However, it can be shown that such a property is impossible given the assumption that $`D_{i\hspace{0pt}j}^{q}`$ is decreasing in $`x_{i\hspace{0pt}j}\hspace{0pt}{(q)}`$.
One way to remedy this problem is to work with log-power variables, first introduced in \[[1](#bib.bib1), [2](#bib.bib2)\]. For each $`{(m,n)} \in \mathcal{L}`$ and $`q \in \mathcal{Q}_{m\hspace{0pt}n}`$, define $`{S_{m\hspace{0pt}n}\hspace{0pt}{(q)}} = {{\ln P_{m\hspace{0pt}n}}\hspace{0pt}{(q)}}`$. Consider $`D_{i\hspace{0pt}j}^{q}`$ as a function of $`{{\mathbf{S}}\hspace{0pt}{(q)}} \triangleq {({S_{m\hspace{0pt}n}\hspace{0pt}{(q)}})}_{{(m,n)} \in \mathcal{L}_{q}}`$ and $`F_{i\hspace{0pt}j}\hspace{0pt}{(q)}`$. If $`D_{i\hspace{0pt}j}^{q}`$ is jointly convex in $`{\mathbf{S}}\hspace{0pt}{(q)}`$ and $`F_{i\hspace{0pt}j}\hspace{0pt}{(q)}`$, the conditions in Theorem [1](#Thmtheorem1 "Theorem 1 ‣ V-D3 Conditions for Optimality ‣ V-D Control Variables and Optimality Conditions ‣ V Cross-layer Optimization in Multi-radio Multi-hop Wireless Networks ‣ Distributed Algorithms for Spectrum Allocation, Power Control, Routing, and Congestion Control in Wireless Networks") are enough to imply the optimality of a configuration. The joint convexity property holds if and only if the link cost function $`D\hspace{0pt}{(x,F)}`$ satisfies the following.
######
Lemma 6
For all $`{(i,j)} \in \mathcal{L}`$ and $`q \in \mathcal{Q}_{i\hspace{0pt}j}`$, $`D_{i\hspace{0pt}j}^{q}`$ is jointly convex in $`\mathbf{S}\hspace{0pt}{(q)}`$ and $`F_{i\hspace{0pt}j}\hspace{0pt}{(q)}`$ if and only if the matrix
<table>
<tbody>
<tr>
<td></td>
<td>M=[Dx′′x2+Dx′xDxF′′xDxF′′xDF′′]𝑀delimited-[]superscriptsubscript𝐷𝑥′′superscript𝑥2superscriptsubscript𝐷𝑥′𝑥superscriptsubscript𝐷𝑥𝐹′′𝑥superscriptsubscript𝐷𝑥𝐹′′𝑥superscriptsubscript𝐷𝐹′′M=\left[\begin{array}[]{cc}D_{x}^{\prime\prime}x^{2}+D_{x}^{\prime}x&D_{xF}^{\prime\prime}x\\ D_{xF}^{\prime\prime}x&D_{F}^{\prime\prime}\end{array}\right]</td>
<td></td>
<td>(29)</td>
</tr>
</tbody>
</table>
is positive semidefinite for all $`(x,F)`$.
In ([29](#S5.E29 "In Lemma 6 ‣ V-D3 Conditions for Optimality ‣ V-D Control Variables and Optimality Conditions ‣ V Cross-layer Optimization in Multi-radio Multi-hop Wireless Networks ‣ Distributed Algorithms for Spectrum Allocation, Power Control, Routing, and Congestion Control in Wireless Networks")), $`D_{x}'`$ and $`D_{x}^{\operatorname{\prime\prime}}`$ denote the first and second partial derivatives of $`D\hspace{0pt}{(x,F)}`$ with respect to $`x`$, $`D_{F}^{\operatorname{\prime\prime}}`$ denotes the second partial derivative of $`D`$ with respect to $`F`$, and $`D_{x\hspace{0pt}F}^{\operatorname{\prime\prime}}`$ denotes $`\partial^{2}{D/{\partial{x\hspace{0pt}{\partial F}}}}`$. It can be shown that the cost function $`{D\hspace{0pt}{(x,F)}} = \frac{F}{{C\hspace{0pt}{(x)}} - F}`$ with $`{C\hspace{0pt}{(x)}} = {R\hspace{0pt}{\log{({K\hspace{0pt}x})}}}`$ satisfies the condition in Lemma [6](#Thmlemma6 "Lemma 6 ‣ V-D3 Conditions for Optimality ‣ V-D Control Variables and Optimality Conditions ‣ V Cross-layer Optimization in Multi-radio Multi-hop Wireless Networks ‣ Distributed Algorithms for Spectrum Allocation, Power Control, Routing, and Congestion Control in Wireless Networks"). Limited by the available space, we skip the proof of Lemma [6](#Thmlemma6 "Lemma 6 ‣ V-D3 Conditions for Optimality ‣ V-D Control Variables and Optimality Conditions ‣ V Cross-layer Optimization in Multi-radio Multi-hop Wireless Networks ‣ Distributed Algorithms for Spectrum Allocation, Power Control, Routing, and Congestion Control in Wireless Networks"). We prove in Appendix [](#A0.SS4 "-D Proof of Theorem 2 ‣ VII Conclusion ‣ Distributed Algorithms for Spectrum Allocation, Power Control, Routing, and Congestion Control in Wireless Networks")
-D
that if the condition in Lemma [6](#Thmlemma6 "Lemma 6 ‣ V-D3 Conditions for Optimality ‣ V-D Control Variables and Optimality Conditions ‣ V Cross-layer Optimization in Multi-radio Multi-hop Wireless Networks ‣ Distributed Algorithms for Spectrum Allocation, Power Control, Routing, and Congestion Control in Wireless Networks") holds, the conditions stated in the next theorem are sufficient for optimality.
######
Theorem 2
If the matrix $`M`$ in ([29](#S5.E29 "In Lemma 6 ‣ V-D3 Conditions for Optimality ‣ V-D Control Variables and Optimality Conditions ‣ V Cross-layer Optimization in Multi-radio Multi-hop Wireless Networks ‣ Distributed Algorithms for Spectrum Allocation, Power Control, Routing, and Congestion Control in Wireless Networks")) of link cost function $`D\hspace{0pt}{(x,F)}`$ is positive semidefinite, then conditions ([24](#S5.E24 "In Theorem 1 ‣ V-D3 Conditions for Optimality ‣ V-D Control Variables and Optimality Conditions ‣ V Cross-layer Optimization in Multi-radio Multi-hop Wireless Networks ‣ Distributed Algorithms for Spectrum Allocation, Power Control, Routing, and Congestion Control in Wireless Networks"))-([28](#S5.E28 "In Theorem 1 ‣ V-D3 Conditions for Optimality ‣ V-D Control Variables and Optimality Conditions ‣ V Cross-layer Optimization in Multi-radio Multi-hop Wireless Networks ‣ Distributed Algorithms for Spectrum Allocation, Power Control, Routing, and Congestion Control in Wireless Networks")) are sufficient for $`{\{{\mathbf{ρ}}_{i},{\mathbf{η}}_{i},{\{{\mathbf{\phi}_{i}\hspace{0pt}{(w)}}\}}_{w \in \mathcal{W}},{\mathbf{μ}}_{i}\}}_{i \in \mathcal{N}}`$ to be optimal if ([24](#S5.E24 "In Theorem 1 ‣ V-D3 Conditions for Optimality ‣ V-D Control Variables and Optimality Conditions ‣ V Cross-layer Optimization in Multi-radio Multi-hop Wireless Networks ‣ Distributed Algorithms for Spectrum Allocation, Power Control, Routing, and Congestion Control in Wireless Networks")) holds at every node $`i \neq {D\hspace{0pt}{(w)}}`$ whether $`{t_{i}\hspace{0pt}{(w)}} > 0`$ or not, and if ([28](#S5.E28 "In Theorem 1 ‣ V-D3 Conditions for Optimality ‣ V-D Control Variables and Optimality Conditions ‣ V Cross-layer Optimization in Multi-radio Multi-hop Wireless Networks ‣ Distributed Algorithms for Spectrum Allocation, Power Control, Routing, and Congestion Control in Wireless Networks")) holds for all $`i \in \mathcal{N}`$ and $`q \in {\mathcal{O}\hspace{0pt}\mathcal{C}_{i}}`$ whether $`{\rho_{i}\hspace{0pt}{(q)}} > 0`$ or not.
## VI Node-Based Multi-Radio Power Control and Routing Algorithms
We develop a set of scaled gradient projection algorithms \[[34](#bib.bib34)\] by which individual nodes adjust their local power control and routing variables iteratively to achieve a global configuration satisfying the optimality conditions in Theorem [2](#Thmtheorem2 "Theorem 2 ‣ V-D3 Conditions for Optimality ‣ V-D Control Variables and Optimality Conditions ‣ V Cross-layer Optimization in Multi-radio Multi-hop Wireless Networks ‣ Distributed Algorithms for Spectrum Allocation, Power Control, Routing, and Congestion Control in Wireless Networks"). Each algorithm at a node updates the appropriate vector of local optimization variables, e.g. $`{\mathbf{ρ}}_{i}`$, $`{\mathbf{η}}_{i}`$, $`\mathbf{\phi}_{i}\hspace{0pt}{(w)}`$ or $`{\mathbf{μ}}_{i}`$, such that the updated vector results in a lower cost with all other variables held constant. This is achieved by updating the old vector in the opposite gradient direction scaled by an appropriate positive definite matrix, and projecting the new vector back into the feasible set whenever it falls outside. This technique has been applied to, for example, optimal routing in wireline networks \[[33](#bib.bib33)\] and optimal power control and routing in single-radio wireless networks \[[3](#bib.bib3)\].
###
VI-A
Power Control Algorithms
We develop two sets of power control algorithms which let each node $`i`$ iteratively adjust the vector $`{({\eta_{i\hspace{0pt}j}\hspace{0pt}{(q)}})}_{j \in \mathcal{O}_{i}}`$ and the vector $`{({\rho_{i}\hspace{0pt}{(q)}})}_{q \in {\mathcal{O}\hspace{0pt}\mathcal{C}_{i}}}`$. First notice that if $`q \in {\mathcal{O}\hspace{0pt}\mathcal{C}_{i}}`$ but $`\mathcal{L}_{q}`$ contains only one of $`i`$’s outgoing links, say $`(i,j)`$, $`\eta_{i\hspace{0pt}j}\hspace{0pt}{(q)}`$ must be equal to $`1`$, and hence is not variable. Otherwise, if $`{{\mathbf{η}}_{i}\hspace{0pt}{(q)}} \triangleq {({\eta_{i\hspace{0pt}j}\hspace{0pt}{(q)}})}_{{(i,j)} \in \mathcal{L}_{q}}`$ has more than one positive element, it is updated by the following scaled gradient projection algorithm:
<table>
<tbody>
<tr>
<td></td>
<td>𝜼ik+1(q):=[𝜼ik(q)−(Qik(q))−1⋅δ𝜼ik(q)]Qik(q)+.assignsuperscriptsubscript𝜼𝑖𝑘1𝑞superscriptsubscriptdelimited-[]superscriptsubscript𝜼𝑖𝑘𝑞⋅superscriptsuperscriptsubscript𝑄𝑖𝑘𝑞1𝛿superscriptsubscript𝜼𝑖𝑘𝑞superscriptsubscript𝑄𝑖𝑘𝑞\boldsymbol{\eta}_{i}^{k+1}(q):=\left[\boldsymbol{\eta}_{i}^{k}(q)-\left(Q_{i}^{k}(q)\right)^{-1}\cdot\delta\boldsymbol{\eta}_{i}^{k}(q)\right]_{Q_{i}^{k}(q)}^{+}.</td>
<td></td>
<td>(30)</td>
</tr>
</tbody>
</table>
Here, the superscripts $`k`$, $`k + 1`$ are the iteration indices, $`Q_{i}\hspace{0pt}{(q)}`$ is a positive definite scaling matrix, $`\delta\hspace{0pt}{\mathbf{η}}_{i}\hspace{0pt}{(q)}`$ is the vector $`{({\delta\hspace{0pt}\eta_{i\hspace{0pt}j}\hspace{0pt}{(q)}})}_{{(i,j)} \in \mathcal{L}_{q}}`$, and $`{\lbrack \cdot \rbrack}_{Q_{i}^{k}\hspace{0pt}{(q)}}^{+}`$ is the projection operation onto the feasible set of $`{\mathbf{η}}_{i}\hspace{0pt}{(q)}`$ (cf. ([12](#S5.E12 "In V-D1 Power Control Variables ‣ V-D Control Variables and Optimality Conditions ‣ V Cross-layer Optimization in Multi-radio Multi-hop Wireless Networks ‣ Distributed Algorithms for Spectrum Allocation, Power Control, Routing, and Congestion Control in Wireless Networks"))) relative to the norm induced by $`Q_{i}^{k}\hspace{0pt}{(q)}`$.
<sup>16</sup>
<sup>16</sup>16In general, $`{\lbrack\overset{\sim}{\mathbf{x}}\rbrack}_{M}^{+} \triangleq \arg\min_{{\mathbf{x}} \in \mathcal{F}}{({\mathbf{x}} - \overset{\sim}{\mathbf{x}})}' \cdot M \cdot {({\mathbf{x}} - \overset{\sim}{\mathbf{x}})}`$, where $`\mathcal{F}`$ is the feasible set of $`\mathbf{x}`$.
To implement the algorithm ([30](#S6.E30 "In VI-A Power Control Algorithms ‣ VI Node-Based Multi-Radio Power Control and Routing Algorithms ‣ Distributed Algorithms for Spectrum Allocation, Power Control, Routing, and Congestion Control in Wireless Networks")), node $`i`$ needs the current value of $`\delta\hspace{0pt}{\mathbf{η}}_{i}\hspace{0pt}{(q)}`$, which by ([15](#S5.E15 "In V-D1 Power Control Variables ‣ V-D Control Variables and Optimality Conditions ‣ V Cross-layer Optimization in Multi-radio Multi-hop Wireless Networks ‣ Distributed Algorithms for Spectrum Allocation, Power Control, Routing, and Congestion Control in Wireless Networks")) is easily computable from local measures.
The algorithm used by node $`i`$ to iteratively update $`{\mathbf{ρ}}_{i}`$ is as follows:
<table>
<tbody>
<tr>
<td></td>
<td>𝝆ik+1:=[𝝆ik−(Vik)−1⋅δ𝝆ik]Vik+.assignsuperscriptsubscript𝝆𝑖𝑘1superscriptsubscriptdelimited-[]superscriptsubscript𝝆𝑖𝑘⋅superscriptsuperscriptsubscript𝑉𝑖𝑘1𝛿superscriptsubscript𝝆𝑖𝑘superscriptsubscript𝑉𝑖𝑘\boldsymbol{\rho}_{i}^{k+1}:=\left[\boldsymbol{\rho}_{i}^{k}-\left(V_{i}^{k}\right)^{-1}\cdot\delta\boldsymbol{\rho}_{i}^{k}\right]_{V_{i}^{k}}^{+}.</td>
<td></td>
<td>(31)</td>
</tr>
</tbody>
</table>
It has almost the same form as ([30](#S6.E30 "In VI-A Power Control Algorithms ‣ VI Node-Based Multi-Radio Power Control and Routing Algorithms ‣ Distributed Algorithms for Spectrum Allocation, Power Control, Routing, and Congestion Control in Wireless Networks")) except that the feasible set for the projection is defined by ([11](#S5.E11 "In V-D1 Power Control Variables ‣ V-D Control Variables and Optimality Conditions ‣ V Cross-layer Optimization in Multi-radio Multi-hop Wireless Networks ‣ Distributed Algorithms for Spectrum Allocation, Power Control, Routing, and Congestion Control in Wireless Networks")). However, notice that each component of $`{\delta\hspace{0pt}{\mathbf{ρ}}_{i}^{k}} \triangleq {({\delta\hspace{0pt}\rho_{i}\hspace{0pt}{(q)}^{k}})}_{q \in {\mathcal{O}\hspace{0pt}\mathcal{C}_{i}}}`$ depends on measures from all active links on sub-band $`q`$. Therefore, prior to implementing ([31](#S6.E31 "In VI-A Power Control Algorithms ‣ VI Node-Based Multi-Radio Power Control and Routing Algorithms ‣ Distributed Algorithms for Spectrum Allocation, Power Control, Routing, and Congestion Control in Wireless Networks")), node $`i`$ needs to collect the appropriate power control messages to determine $`\delta\hspace{0pt}{\mathbf{ρ}}_{i}^{k}`$, as described by the message exchange protocol in Section [](#S5.SS4.SSS1 "V-D1 Power Control Variables ‣ V-D Control Variables and Optimality Conditions ‣ V Cross-layer Optimization in Multi-radio Multi-hop Wireless Networks ‣ Distributed Algorithms for Spectrum Allocation, Power Control, Routing, and Congestion Control in Wireless Networks")
V-D1
.
###
VI-B
Routing Algorithms
We now present the algorithm used by a node $`i`$ to update its inter-node routing vector $`\mathbf{\phi}_{i}\hspace{0pt}{(w)}`$ for a traversing session $`w`$. An iteration of the inter-node routing algorithm has the form
<table>
<tbody>
<tr>
<td></td>
<td>ϕik+1(w):=[ϕik(w)−(Mik(w))−1⋅δϕik(w)]Mik(w)+.assignsuperscriptsubscriptbold-italic-ϕ𝑖𝑘1𝑤superscriptsubscriptdelimited-[]superscriptsubscriptbold-italic-ϕ𝑖𝑘𝑤⋅superscriptsuperscriptsubscript𝑀𝑖𝑘𝑤1𝛿superscriptsubscriptbold-italic-ϕ𝑖𝑘𝑤superscriptsubscript𝑀𝑖𝑘𝑤\boldsymbol{\phi}_{i}^{k+1}(w):=\left[\boldsymbol{\phi}_{i}^{k}(w)-(M_{i}^{k}(w))^{-1}\cdot\delta\boldsymbol{\phi}_{i}^{k}(w)\right]_{M_{i}^{k}(w)}^{+}.</td>
<td></td>
<td>(32)</td>
</tr>
</tbody>
</table>
Here, $`{\delta\hspace{0pt}\mathbf{\phi}_{i}^{k}\hspace{0pt}{(w)}} \triangleq {({\delta\hspace{0pt}\phi_{i\hspace{0pt}j}^{k}\hspace{0pt}{(w)}})}_{j \in \mathcal{O}_{i}}`$ is the vector of current marginal routing cost indicators and $`M_{i}^{k}\hspace{0pt}{(w)}`$ is a positive definite matrix used to scale the descent direction. The feasible set associated with the projection operation in ([32](#S6.E32 "In VI-B Routing Algorithms ‣ VI Node-Based Multi-Radio Power Control and Routing Algorithms ‣ Distributed Algorithms for Spectrum Allocation, Power Control, Routing, and Congestion Control in Wireless Networks")) is prescribed by the constraint ([18](#S5.E18 "In V-D2 Routing Variables ‣ V-D Control Variables and Optimality Conditions ‣ V Cross-layer Optimization in Multi-radio Multi-hop Wireless Networks ‣ Distributed Algorithms for Spectrum Allocation, Power Control, Routing, and Congestion Control in Wireless Networks")) and an additional requirement that $`{\phi_{i\hspace{0pt}j}\hspace{0pt}{(w)}} = 0`$ for all $`j \in {\mathcal{B}_{i}^{k}\hspace{0pt}{(w)}}`$, where $`\mathcal{B}_{i}^{k}\hspace{0pt}{(w)}`$ is the blocked node set of node $`i`$ relative to session $`w`$. The meaning and use of the blocked node set were briefly discussed in the routing message exchange protocol in Section [](#S5.SS4.SSS2 "V-D2 Routing Variables ‣ V-D Control Variables and Optimality Conditions ‣ V Cross-layer Optimization in Multi-radio Multi-hop Wireless Networks ‣ Distributed Algorithms for Spectrum Allocation, Power Control, Routing, and Congestion Control in Wireless Networks")
V-D2
. The exchange protocol also gives every node enough information to calculate the current $`\delta\hspace{0pt}\mathbf{\phi}_{i}\hspace{0pt}{(w)}`$ prior to each iteration of ([32](#S6.E32 "In VI-B Routing Algorithms ‣ VI Node-Based Multi-Radio Power Control and Routing Algorithms ‣ Distributed Algorithms for Spectrum Allocation, Power Control, Routing, and Congestion Control in Wireless Networks")).
For the source node $`O\hspace{0pt}{(w)}`$ of a session $`w`$, the algorithm ([32](#S6.E32 "In VI-B Routing Algorithms ‣ VI Node-Based Multi-Radio Power Control and Routing Algorithms ‣ Distributed Algorithms for Spectrum Allocation, Power Control, Routing, and Congestion Control in Wireless Networks")) is applied to the vector of routing variables associated with real outgoing links. The overflow routing variable $`\phi_{w}`$ is updated by
<table>
<tbody>
<tr>
<td></td>
<td>ϕwk+1:=[ϕwk−κwk(Dw′−∂E∂rw)]+,assignsuperscriptsubscriptitalic-ϕ𝑤𝑘1superscriptdelimited-[]superscriptsubscriptitalic-ϕ𝑤𝑘superscriptsubscript𝜅𝑤𝑘superscriptsubscript𝐷𝑤′𝐸subscript𝑟𝑤\phi_{w}^{k+1}:=\left[\phi_{w}^{k}-\kappa_{w}^{k}\left(D_{w}^{\prime}-\frac{\partial E}{\partial r_{w}}\right)\right]^{+},</td>
<td></td>
<td>(33)</td>
</tr>
</tbody>
</table>
for which the projection is onto the feasible set $`0 \leq \phi_{w} \leq 1`$. The gradient is given by subtracting $`\frac{\partial E}{\partial r_{w}}`$, which is computable after the routing message exchange, from the local measure $`D_{w}'`$.
Finally, we come to the algorithm that node $`i`$ uses to iteratively adjust the intra-node routing vector $`{\mathbf{μ}}_{i\hspace{0pt}j} \triangleq {({\mu_{i\hspace{0pt}j}\hspace{0pt}{(q)}})}_{q \in {\mathcal{O}\hspace{0pt}\mathcal{C}_{i}}}`$ applied to an outgoing link $`(i,j)`$. The intra-node routing update is iterated as
<table>
<tbody>
<tr>
<td></td>
<td>𝝁ik+1:=[𝝁ik−(Tijk)−1⋅∂𝑫Fijk]Tijk+.assignsuperscriptsubscript𝝁𝑖𝑘1superscriptsubscriptdelimited-[]superscriptsubscript𝝁𝑖𝑘⋅superscriptsuperscriptsubscript𝑇𝑖𝑗𝑘1superscriptsubscript𝑫subscript𝐹𝑖𝑗𝑘superscriptsubscript𝑇𝑖𝑗𝑘\boldsymbol{\mu}_{i}^{k+1}:=\left[\boldsymbol{\mu}_{i}^{k}-(T_{ij}^{k})^{-1}\cdot\partial\boldsymbol{D}_{F_{ij}}^{k}\right]_{T_{ij}^{k}}^{+}.</td>
<td></td>
<td>(34)</td>
</tr>
</tbody>
</table>
Here, $`\partial{\mathbf{D}}_{F_{i\hspace{0pt}j}}`$ is the vector of partial derivatives $`\partial{D_{i\hspace{0pt}j}^{q}/{\partial{F_{i\hspace{0pt}j}\hspace{0pt}{(q)}}}}`$ of all $`q \in \mathcal{Q}_{i\hspace{0pt}j}`$, which are purely local measures. The scaling matrix $`T_{i\hspace{0pt}j}^{k}`$ is positive definite. The feasible set that the projection refers to is defined by ([19](#S5.E19 "In V-D2 Routing Variables ‣ V-D Control Variables and Optimality Conditions ‣ V Cross-layer Optimization in Multi-radio Multi-hop Wireless Networks ‣ Distributed Algorithms for Spectrum Allocation, Power Control, Routing, and Congestion Control in Wireless Networks")).
###
VI-C
Convergence of Algorithms
Having described the node-based power control and routing algorithms, we now state the main convergence result in the following theorem.
######
Theorem 3
Given a feasible spectrum allocation $`{\{\mathcal{L}_{q}\}}_{q \in \mathcal{Q}}`$, let $`\{{\mathbf{η}}_{i}^{0}\}`$, $`\{{\mathbf{ρ}}_{i}^{0}\}`$, $`\{{\mathbf{\phi}_{i}\hspace{0pt}{(w)}}\}`$, $`\{{\mathbf{μ}}_{i}\}`$ be any feasible initial transmission power and routing configuration with finite cost. Then with appropriate scaling matrices, the update sequences $`{\{{\mathbf{η}}_{i}^{k}\}}_{k = 1}^{\infty}`$, $`{\{{\mathbf{ρ}}_{i}^{k}\}}_{k = 1}^{\infty}`$, $`{\{{\mathbf{\phi}_{i}^{k}\hspace{0pt}{(w)}}\}}_{k = 1}^{\infty}`$ and $`{\{{\mathbf{μ}}_{i}^{k}\}}_{k = 1}^{\infty}`$ generated by the algorithms ([30](#S6.E30 "In VI-A Power Control Algorithms ‣ VI Node-Based Multi-Radio Power Control and Routing Algorithms ‣ Distributed Algorithms for Spectrum Allocation, Power Control, Routing, and Congestion Control in Wireless Networks")), ([31](#S6.E31 "In VI-A Power Control Algorithms ‣ VI Node-Based Multi-Radio Power Control and Routing Algorithms ‣ Distributed Algorithms for Spectrum Allocation, Power Control, Routing, and Congestion Control in Wireless Networks")), ([32](#S6.E32 "In VI-B Routing Algorithms ‣ VI Node-Based Multi-Radio Power Control and Routing Algorithms ‣ Distributed Algorithms for Spectrum Allocation, Power Control, Routing, and Congestion Control in Wireless Networks")) and ([34](#S6.E34 "In VI-B Routing Algorithms ‣ VI Node-Based Multi-Radio Power Control and Routing Algorithms ‣ Distributed Algorithms for Spectrum Allocation, Power Control, Routing, and Congestion Control in Wireless Networks")) converge, i.e., $`{\mathbf{η}}_{i}^{k}\rightarrow{\mathbf{η}}_{i}^{\ast}`$, $`{\mathbf{ρ}}_{i}^{k}\rightarrow{\mathbf{ρ}}_{i}^{\ast}`$, $`{\mathbf{\phi}_{i}^{k}\hspace{0pt}{(w)}}\rightarrow{\mathbf{\phi}_{i}^{\ast}\hspace{0pt}{(w)}}`$, and $`{\mathbf{μ}}_{i}^{k}\rightarrow{\mathbf{μ}}_{i}^{\ast}`$ for all $`i \in \mathcal{N}`$ and $`w \in \mathcal{W}`$ as $`k\rightarrow\infty`$. Furthermore, the limiting configuration $`\{{\mathbf{η}}_{i}^{\ast}\}`$, $`\{{\mathbf{ρ}}_{i}^{\ast}\}`$, $`\{{\mathbf{\phi}_{i}^{\ast}\hspace{0pt}{(w)}}\}`$, $`\{{\mathbf{μ}}_{i}^{\ast}\}`$ satisfies the conditions in Theorem [2](#Thmtheorem2 "Theorem 2 ‣ V-D3 Conditions for Optimality ‣ V-D Control Variables and Optimality Conditions ‣ V Cross-layer Optimization in Multi-radio Multi-hop Wireless Networks ‣ Distributed Algorithms for Spectrum Allocation, Power Control, Routing, and Congestion Control in Wireless Networks"), and is a jointly optimal solution to the MCRA problem if the link cost function $`D\hspace{0pt}{(x,F)}`$ satisfies the condition in Lemma [6](#Thmlemma6 "Lemma 6 ‣ V-D3 Conditions for Optimality ‣ V-D Control Variables and Optimality Conditions ‣ V Cross-layer Optimization in Multi-radio Multi-hop Wireless Networks ‣ Distributed Algorithms for Spectrum Allocation, Power Control, Routing, and Congestion Control in Wireless Networks").
The proof hinges on the fact that, by using appropriate scaling matrices, every iteration of the algorithms ([30](#S6.E30 "In VI-A Power Control Algorithms ‣ VI Node-Based Multi-Radio Power Control and Routing Algorithms ‣ Distributed Algorithms for Spectrum Allocation, Power Control, Routing, and Congestion Control in Wireless Networks")), ([31](#S6.E31 "In VI-A Power Control Algorithms ‣ VI Node-Based Multi-Radio Power Control and Routing Algorithms ‣ Distributed Algorithms for Spectrum Allocation, Power Control, Routing, and Congestion Control in Wireless Networks")), ([31](#S6.E31 "In VI-A Power Control Algorithms ‣ VI Node-Based Multi-Radio Power Control and Routing Algorithms ‣ Distributed Algorithms for Spectrum Allocation, Power Control, Routing, and Congestion Control in Wireless Networks")) and ([32](#S6.E32 "In VI-B Routing Algorithms ‣ VI Node-Based Multi-Radio Power Control and Routing Algorithms ‣ Distributed Algorithms for Spectrum Allocation, Power Control, Routing, and Congestion Control in Wireless Networks")) reduces the total cost of the MCRA problem until the optimality conditions in Theorem [2](#Thmtheorem2 "Theorem 2 ‣ V-D3 Conditions for Optimality ‣ V-D Control Variables and Optimality Conditions ‣ V Cross-layer Optimization in Multi-radio Multi-hop Wireless Networks ‣ Distributed Algorithms for Spectrum Allocation, Power Control, Routing, and Congestion Control in Wireless Networks") are achieved. Finding the appropriate scaling matrices, however, is a major challenge. One approach to this problem is to choose the scaling matrices so that they upper bound the Hessian matrices with respect to the updated variables. In this way, the algorithms closely approximate Newton’s method, hence enjoying fast rate of convergence while simultaneously guaranteeing convergence from all initial conditions. This method has been successfully adopted in the power control and routing algorithms for single-radio wireless networks in \[[3](#bib.bib3)\], and can be generalized to the present context. Due to the limited space, however, we skip the details.
It is worth noting that convergence does not depend on any particular order of running the algorithms at different nodes. At any time, any node can update any set of its local optimization variables via the corresponding algorithm. All that is required for convergence is that each node iterates every algorithm until the adjusted variables have marginal costs satisfying conditions in Theorem [2](#Thmtheorem2 "Theorem 2 ‣ V-D3 Conditions for Optimality ‣ V-D Control Variables and Optimality Conditions ‣ V Cross-layer Optimization in Multi-radio Multi-hop Wireless Networks ‣ Distributed Algorithms for Spectrum Allocation, Power Control, Routing, and Congestion Control in Wireless Networks"). Finally, we note that convergence occurs from any initial configuration with finite cost. These features are crucial to the applicability of these algorithms to large networks which lack the ability of scheduling and synchronizing node operations.
Extensive simulations indicate that our algorithms are adaptive to time-varying network state, including channel fading, network topology, and traffic demand. Because every iteration of any algorithm always reduces the total cost under the current network condition, our scheme is able to constantly readjust routing and transmission powers towards the optimum that slowly shifts over time due to the network change.
## VII Conclusion
We have developed an integrated cross-layer resource allocation scheme for general wireless multi-hop networks. To satisfy the fundamental duplexing constraints, our scheme first finds a feasible spectrum allocation by (1) dividing the whole spectrum into multiple sub-bands and (2) activating conflict-free links on each sub-band. Compared with traditional scheduling in time, the spectrum allocation technique has a number of advantages in operational simplicity and amenability to distributed and asynchronous implementation. By studying an equivalent combinatorial link-coloring problem, we found that the minimum number of sub-bands required by a feasible spectrum allocation is given by a simple function of the chromatic number of the network connectivity graph. The minimum number grows asymptotically at a logarithmic rate with the chromatic number, attesting to the good scalability of the spectrum allocation technique and its robustness to network topology changes. We designed a simple distributed and asynchronous algorithm by which a feasible spectrum allocation can be constructed given enough sub-bands.
Given a feasible spectrum allocation, we developed an analytical framework and a set of node-based distributed algorithms for optimally allocating transmission powers and traffic rates on active links. Such a framework is especially suitable for the design of wireless networks with frequency selective channels. We provided the conditions that an optimal power control and routing configuration must satisfy. We then designed a set of distributed power control and routing algorithms using the scaled gradient projection method. These algorithms can be iterated at individual nodes with little control overhead. Finally, we demonstrated that the algorithms asymptotically achieve the optimal configuration regardless of the initial condition and the order of iterating different algorithms.
###
-A
Proof of Lemma [2](#Thmlemma2 "Lemma 2 ‣ III-A Link-Coloring Problem ‣ III Spectrum Division ‣ Distributed Algorithms for Spectrum Allocation, Power Control, Routing, and Congestion Control in Wireless Networks")
We provide the proof for the case $`g > {\lfloor{Q/2}\rfloor}`$. The other case $`g < {\lfloor{Q/2}\rfloor}`$ can be seen as a corollary by taking complements of all subsets involved in the first case.
First notice that since $`g > {\lfloor{Q/2}\rfloor}`$, there are $`K = \binom{Q}{g - 1}`$ distinct subsets $`C_{1},\cdots,C_{K}`$ of cardinality $`g - 1`$, where $`K \geq k = \binom{Q}{g}`$, the number of distinct subsets of cardinality $`g`$. Hence, the claim makes intuitive sense.
Consider the bipartite graph consisting of $`\{ B_{i}\}`$ and $`\{ C_{j}\}`$ where a pair of $`B_{i}`$ and $`C_{j}`$ are connected if and only if $`C_{j} \subset B_{i}`$. In this case, $`C_{j}`$ is said to be a child of $`B_{i}`$, and $`B_{i}`$ is said to be a parent of $`C_{j}`$. It is easy to see that every $`B_{i}`$ has $`g`$ children and every $`C_{j}`$ has $`{Q - g} + 1`$ parents. Now the claim in the lemma is equivalent to the existence of a complete matching (of size $`k`$) of $`\{ B_{i}\}`$ and $`\{ C_{j}\}`$ in the bipartite graph specified above. Because $`g > {\lfloor{Q/2}\rfloor}`$ implies that $`g \geq \frac{Q + 1}{2}`$, the degree of any $`B_{i}`$, which is $`g`$, is greater than or equal to $`{Q - g} + 1`$, the degree of any $`C_{j}`$. In this case, a complete matching must exist by a corollary of Hall’s theorem (cf. Corollary 13.4 of \[[35](#bib.bib35)\]).
###
-B
Proof of Lemma [3](#Thmlemma3 "Lemma 3 ‣ III-A Link-Coloring Problem ‣ III Spectrum Division ‣ Distributed Algorithms for Spectrum Allocation, Power Control, Routing, and Congestion Control in Wireless Networks")
Suppose $`{\{ B_{i}\}}_{i \in \mathcal{N}}`$ is an optimal solution and $`\left| {\bigcup_{i}B_{i}} \right| = Q`$. If $`\{ B_{i}\}`$ does not have the property in the lemma, we can always modify it to $`\{ B_{i}^{\ast}\}`$ which satisfies the property with $`Q^{\ast} = \left| {\bigcup_{i}B_{i}^{\ast}} \right|`$ being less than or equal to $`Q`$. Hence, the lemma follows. Suppose we have a feasible solution $`\mathcal{B}`$ consisting of a collection of $`N`$ subsets $`{\{ B_{i}\}}_{i = 1}^{N}`$. We show that based on $`\{ B_{i}\}`$, we can construct $`\overset{\sim}{\mathcal{B}} = {\{{\overset{\sim}{B}}_{i}\}}_{i = 1}^{N}`$ such that $`{|{\overset{\sim}{B}}_{i}|} = {\lfloor{Q/2}\rfloor}`$ for all $`i = {1,\cdots,N}`$, where $`Q = {|{\bigcup_{i}B_{i}}|}`$.
Define $`m = {\min_{i}{|B_{i}|}}`$ and $`M = {\max_{i}{|B_{i}|}}`$. If $`m < M`$, then $`m < {\lfloor{Q/2}\rfloor}`$ or $`M > {\lfloor{Q/2}\rfloor}`$ or both. If $`m < {\lfloor{Q/2}\rfloor}`$, define $`\mathcal{N}_{m} \triangleq {\{{i \in \mathcal{N}}:{{|B_{i}|} = m}\}}`$, replace each $`B_{j}`$, $`j \in \mathcal{N}_{m}`$, by subset $`B_{j}'`$ with cardinality $`m + 1`$ such that $`B_{j} \subset B_{j}'`$ for all $`j \in \mathcal{N}_{m}`$, and $`B_{j}' \neq B_{k}'`$ if and only if $`B_{j} \neq B_{k}`$. Such a replacement is possible by Lemma [2](#Thmlemma2 "Lemma 2 ‣ III-A Link-Coloring Problem ‣ III Spectrum Division ‣ Distributed Algorithms for Spectrum Allocation, Power Control, Routing, and Congestion Control in Wireless Networks"). If $`M > {\lfloor{Q/2}\rfloor}`$, define $`\mathcal{N}_{M} \triangleq {\{{i \in \mathcal{N}}:{{|B_{i}|} = M}\}}`$, replace each $`B_{j}`$, $`j \in \mathcal{N}_{M}`$, by subset $`B_{j}'`$ with cardinality $`M - 1`$ such that $`B_{j}' \subset B_{j}`$ for all $`j \in \mathcal{N}_{M}`$, and $`B_{j}' \neq B_{k}'`$ if and only if $`B_{j} \neq B_{k}`$. Such a replacement can be found also by Lemma [2](#Thmlemma2 "Lemma 2 ‣ III-A Link-Coloring Problem ‣ III Spectrum Division ‣ Distributed Algorithms for Spectrum Allocation, Power Control, Routing, and Congestion Control in Wireless Networks"). Denote by $`\mathcal{N}'`$ the subset of nodes whose $`B_{i}`$ is changed (either expanded or reduced). Because $`m < M`$, $`\mathcal{N}'`$ is always non-empty. It can be verified that the new collection of subsets $`\{{\{ B_{j}'\}}_{j \in \mathcal{N}'},{\{ B_{i}\}}_{i \notin \mathcal{N}'}\}`$ is another optimal solution of ([2](#S3.E2 "In III-A Link-Coloring Problem ‣ III Spectrum Division ‣ Distributed Algorithms for Spectrum Allocation, Power Control, Routing, and Congestion Control in Wireless Networks")). If the minimum and maximum cardinalities of the new collection of subsets are equal, we are done. Otherwise repeat the above procedure until we obtain an even collection of subsets. The iterations terminate in a finite number of steps since each iteration strictly reduces the difference between $`M`$ and $`m`$.
###
-C
Proof of Theorem [1](#Thmtheorem1 "Theorem 1 ‣ V-D3 Conditions for Optimality ‣ V-D Control Variables and Optimality Conditions ‣ V Cross-layer Optimization in Multi-radio Multi-hop Wireless Networks ‣ Distributed Algorithms for Spectrum Allocation, Power Control, Routing, and Congestion Control in Wireless Networks")
We show that whenever one of conditions ([24](#S5.E24 "In Theorem 1 ‣ V-D3 Conditions for Optimality ‣ V-D Control Variables and Optimality Conditions ‣ V Cross-layer Optimization in Multi-radio Multi-hop Wireless Networks ‣ Distributed Algorithms for Spectrum Allocation, Power Control, Routing, and Congestion Control in Wireless Networks"))-([28](#S5.E28 "In Theorem 1 ‣ V-D3 Conditions for Optimality ‣ V-D Control Variables and Optimality Conditions ‣ V Cross-layer Optimization in Multi-radio Multi-hop Wireless Networks ‣ Distributed Algorithms for Spectrum Allocation, Power Control, Routing, and Congestion Control in Wireless Networks")) is violated, the present configuration can be improved upon. We take condition ([24](#S5.E24 "In Theorem 1 ‣ V-D3 Conditions for Optimality ‣ V-D Control Variables and Optimality Conditions ‣ V Cross-layer Optimization in Multi-radio Multi-hop Wireless Networks ‣ Distributed Algorithms for Spectrum Allocation, Power Control, Routing, and Congestion Control in Wireless Networks")) as an example. Arguments for the other conditions are similar. By ([21](#S5.E21 "In V-D2 Routing Variables ‣ V-D Control Variables and Optimality Conditions ‣ V Cross-layer Optimization in Multi-radio Multi-hop Wireless Networks ‣ Distributed Algorithms for Spectrum Allocation, Power Control, Routing, and Congestion Control in Wireless Networks")) and ([20](#S5.E20 "In V-D2 Routing Variables ‣ V-D Control Variables and Optimality Conditions ‣ V Cross-layer Optimization in Multi-radio Multi-hop Wireless Networks ‣ Distributed Algorithms for Spectrum Allocation, Power Control, Routing, and Congestion Control in Wireless Networks")), we have
<table>
<tbody>
<tr>
<td></td>
<td>δϕij(w)≥minq∈𝒬ij[DFijq′]+∂E∂tj(w)𝛿subscriptitalic-ϕ𝑖𝑗𝑤subscript𝑞subscript𝒬𝑖𝑗superscriptsuperscriptsubscript𝐷subscript𝐹𝑖𝑗𝑞′𝐸subscript𝑡𝑗𝑤\delta\phi_{ij}(w)\geq\min_{q\in{\cal Q}_{ij}}\left[{D_{F_{ij}}^{q}}^{\prime}\right]+\frac{\partial E}{\partial t_{j}(w)}</td>
<td></td>
</tr>
</tbody>
</table>
and
<table>
<tbody>
<tr>
<td></td>
<td>∂E∂ti(w)≥minj∈𝒪i{minq∈𝒬ij[DFijq′]+∂E∂tj(w)}.𝐸subscript𝑡𝑖𝑤subscript𝑗subscript𝒪𝑖subscript𝑞subscript𝒬𝑖𝑗superscriptsuperscriptsubscript𝐷subscript𝐹𝑖𝑗𝑞′𝐸subscript𝑡𝑗𝑤\frac{\partial E}{\partial t_{i}(w)}\geq\min_{j\in{\cal O}_{i}}\left\{\min_{q\in{\cal Q}_{ij}}\left[{D_{F_{ij}}^{q}}^{\prime}\right]+\frac{\partial E}{\partial t_{j}(w)}\right\}.</td>
<td></td>
</tr>
</tbody>
</table>
Thus, condition ([24](#S5.E24 "In Theorem 1 ‣ V-D3 Conditions for Optimality ‣ V-D Control Variables and Optimality Conditions ‣ V Cross-layer Optimization in Multi-radio Multi-hop Wireless Networks ‣ Distributed Algorithms for Spectrum Allocation, Power Control, Routing, and Congestion Control in Wireless Networks")) holds if and only if for all $`j \in \mathcal{O}_{i}`$ such that $`{\phi_{i\hspace{0pt}j}\hspace{0pt}{(w)}} > 0`$, $`{\delta\hspace{0pt}\phi_{i\hspace{0pt}j}\hspace{0pt}{(w)}} \leq {{\min_{q \in \mathcal{Q}_{i\hspace{0pt}k}}\left\lbrack D_{F_{i\hspace{0pt}k}}^{q'} \right\rbrack} + \frac{\partial E}{\partial{t_{k}\hspace{0pt}{(w)}}}}`$ for all $`k \in \mathcal{O}_{i}`$. Suppose condition ([24](#S5.E24 "In Theorem 1 ‣ V-D3 Conditions for Optimality ‣ V-D Control Variables and Optimality Conditions ‣ V Cross-layer Optimization in Multi-radio Multi-hop Wireless Networks ‣ Distributed Algorithms for Spectrum Allocation, Power Control, Routing, and Congestion Control in Wireless Networks")) is violated at $`i`$ with $`{t_{i}\hspace{0pt}{(w)}} > 0`$, i.e., there exists $`j \in \mathcal{O}_{i}`$ such that $`{\phi_{i\hspace{0pt}j}\hspace{0pt}{(w)}} > 0`$, $`{\delta\hspace{0pt}\phi_{i\hspace{0pt}j}\hspace{0pt}{(w)}} > {{\min_{q \in \mathcal{Q}_{i\hspace{0pt}k}}\left\lbrack D_{F_{i\hspace{0pt}k}}^{q'} \right\rbrack} + \frac{\partial E}{\partial{t_{k}\hspace{0pt}{(w)}}}} \triangleq \sigma`$ for some $`k \in \mathcal{O}_{i}`$. If $`F_{i\hspace{0pt}k} > 0`$ and $`{\delta\hspace{0pt}\phi_{i\hspace{0pt}k}\hspace{0pt}{(w)}} > \sigma`$, there must exist $`q \in \mathcal{Q}_{i\hspace{0pt}k}`$ with $`{\mu_{i\hspace{0pt}k}\hspace{0pt}{(q)}} > 0`$ but $`D_{F_{i\hspace{0pt}k}}^{q'} > D_{F_{i\hspace{0pt}k}}^{v'}`$ for some other $`v \in \mathcal{Q}_{i\hspace{0pt}k}`$. Hence, the cost of the configuration can be further reduced by shifting a tiny portion from $`\mu_{i\hspace{0pt}k}\hspace{0pt}{(q)}`$ to $`\mu_{i\hspace{0pt}k}\hspace{0pt}{(v)}`$ as suggested by ([23](#S5.E23 "In V-D2 Routing Variables ‣ V-D Control Variables and Optimality Conditions ‣ V Cross-layer Optimization in Multi-radio Multi-hop Wireless Networks ‣ Distributed Algorithms for Spectrum Allocation, Power Control, Routing, and Congestion Control in Wireless Networks")). If $`{\delta\hspace{0pt}\phi_{i\hspace{0pt}k}\hspace{0pt}{(w)}} = \sigma`$, then $`{\delta\hspace{0pt}\phi_{i\hspace{0pt}j}\hspace{0pt}{(w)}} > {\delta\hspace{0pt}\phi_{i\hspace{0pt}k}\hspace{0pt}{(w)}}`$, we can reduce the cost by shifting a tiny portion from $`\phi_{i\hspace{0pt}j}\hspace{0pt}{(w)}`$ to $`\phi_{i\hspace{0pt}k}\hspace{0pt}{(w)}`$ in light of ([22](#S5.E22 "In V-D2 Routing Variables ‣ V-D Control Variables and Optimality Conditions ‣ V Cross-layer Optimization in Multi-radio Multi-hop Wireless Networks ‣ Distributed Algorithms for Spectrum Allocation, Power Control, Routing, and Congestion Control in Wireless Networks")). If $`F_{i\hspace{0pt}k} = 0`$, then we can always make $`{\delta\hspace{0pt}\phi_{i\hspace{0pt}k}\hspace{0pt}{(w)}} = \sigma`$ by setting all $`\mu_{i\hspace{0pt}k}\hspace{0pt}{(q)}`$ equal to zero but $`\mu_{i\hspace{0pt}k}\hspace{0pt}{(v)}`$ equal to one for one $`v`$ achieving the minimum in $`\sigma`$. Notice that reconfiguring $`{\{{\mu_{i\hspace{0pt}k}\hspace{0pt}{(q)}}\}}_{q \in \mathcal{Q}_{i\hspace{0pt}k}}`$ does not change the value of $`\sigma`$ since by assumption $`F_{i\hspace{0pt}k} = 0`$. Then, as before, the total cost can be further reduced by shifting a tiny portion from $`\phi_{i\hspace{0pt}j}\hspace{0pt}{(w)}`$ to $`\phi_{i\hspace{0pt}k}\hspace{0pt}{(w)}`$.
###
-D
Proof of Theorem [2](#Thmtheorem2 "Theorem 2 ‣ V-D3 Conditions for Optimality ‣ V-D Control Variables and Optimality Conditions ‣ V Cross-layer Optimization in Multi-radio Multi-hop Wireless Networks ‣ Distributed Algorithms for Spectrum Allocation, Power Control, Routing, and Congestion Control in Wireless Networks")
Suppose $`\{{\mathbf{ρ}}_{i},{\mathbf{η}}_{i},{\{{\mathbf{\phi}_{i}\hspace{0pt}{(w)}}\}}_{w \in \mathcal{W}},{\mathbf{μ}}_{i}\}`$ satisfies the conditions in Theorem [2](#Thmtheorem2 "Theorem 2 ‣ V-D3 Conditions for Optimality ‣ V-D Control Variables and Optimality Conditions ‣ V Cross-layer Optimization in Multi-radio Multi-hop Wireless Networks ‣ Distributed Algorithms for Spectrum Allocation, Power Control, Routing, and Congestion Control in Wireless Networks"). Suppose the configuration yields log-power variables $`{\{{{\mathbf{S}}\hspace{0pt}{(q)}}\}}_{q \in \mathcal{Q}}`$ and link flow variables $`{(F_{w})}_{w \in \mathcal{W}}`$, $`{({F_{i\hspace{0pt}j}\hspace{0pt}{(q)}})}_{{{(i,j)} \in \mathcal{L}},{q \in \mathcal{Q}_{i\hspace{0pt}j}}}`$. Let $`\{{\overset{\sim}{\mathbf{ρ}}}_{i},{\overset{\sim}{\mathbf{η}}}_{i},{\{{{\overset{\sim}{\mathbf{\phi}}}_{i}\hspace{0pt}{(w)}}\}}_{w \in \mathcal{W}},{\overset{\sim}{\mathbf{μ}}}_{i}\}`$ be another feasible configuration which yields $`\{{\overset{\sim}{\mathbf{S}}\hspace{0pt}{(q)}}\}`$, $`{({\overset{\sim}{F}}_{w})}_{w \in \mathcal{W}}`$, $`{({{\overset{\sim}{F}}_{i\hspace{0pt}j}\hspace{0pt}{(q)}})}_{{{(i,j)} \in \mathcal{E}},{q \in \mathcal{Q}_{i\hspace{0pt}j}}}`$. Recall that each $`D_{i\hspace{0pt}j}^{q}`$ is jointly convex in $`{{\mathbf{S}}\hspace{0pt}{(q)}} = {({S_{m\hspace{0pt}n}\hspace{0pt}{(q)}})}_{{(m,n)} \in \mathcal{L}_{q}}`$ and $`F_{i\hspace{0pt}j}\hspace{0pt}{(q)}`$ while each $`D_{w}`$ is convex in $`F_{w}`$. Moreover, the feasible sets of log-power variables and flow variables are both convex. Therefore, the cost difference under the two configurations can be bounded by
<table>
<tbody>
<tr>
<td></td>
<td></td>
<td></td>
<td>∑w∈𝒲D~w+∑(i,j)∈ℒ∑q∈𝒬ijD~ijq−∑w∈𝒲Dw−∑(i,j)∈ℒ∑q∈𝒬ijDijqsubscript𝑤𝒲subscript~𝐷𝑤subscript𝑖𝑗ℒsubscript𝑞subscript𝒬𝑖𝑗superscriptsubscript~𝐷𝑖𝑗𝑞subscript𝑤𝒲subscript𝐷𝑤subscript𝑖𝑗ℒsubscript𝑞subscript𝒬𝑖𝑗superscriptsubscript𝐷𝑖𝑗𝑞\displaystyle\sum_{w\in{\cal W}}\tilde{D}_{w}+\sum_{(i,j)\in{\cal L}}\sum_{q\in{\cal Q}_{ij}}\tilde{D}_{ij}^{q}-\sum_{w\in{\cal W}}D_{w}-\sum_{(i,j)\in{\cal L}}\sum_{q\in{\cal Q}_{ij}}D_{ij}^{q}</td>
<td></td>
<td rowspan="3">(35)</td>
</tr>
<tr>
<td></td>
<td></td>
<td>≥\displaystyle\geq</td>
<td>∑w∈𝒲Dw′(F~w−Fw)+∑(i,j)∈ℒ∑q∈𝒬ijDFijq′(F~ij(q)−Fij(q))subscript𝑤𝒲superscriptsubscript𝐷𝑤′subscript~𝐹𝑤subscript𝐹𝑤subscript𝑖𝑗ℒsubscript𝑞subscript𝒬𝑖𝑗superscriptsuperscriptsubscript𝐷subscript𝐹𝑖𝑗𝑞′subscript~𝐹𝑖𝑗𝑞subscript𝐹𝑖𝑗𝑞\displaystyle\sum_{w\in{\cal W}}D_{w}^{\prime}(\tilde{F}_{w}-F_{w})+\sum_{(i,j)\in{\cal L}}\sum_{q\in{\cal Q}_{ij}}{D_{F_{ij}}^{q}}^{\prime}(\tilde{F}_{ij}(q)-F_{ij}(q))</td>
<td></td>
</tr>
<tr>
<td></td>
<td></td>
<td></td>
<td>+∑(i,j)∈ℒ∑q∈𝒬ij∑(m,n)∈ℒq∂Dijq∂Smn(q)(S~mn(q)−Smn(q)).subscript𝑖𝑗ℒsubscript𝑞subscript𝒬𝑖𝑗subscript𝑚𝑛subscriptℒ𝑞superscriptsubscript𝐷𝑖𝑗𝑞subscript𝑆𝑚𝑛𝑞subscript~𝑆𝑚𝑛𝑞subscript𝑆𝑚𝑛𝑞\displaystyle+\sum_{(i,j)\in{\cal L}}\sum_{q\in{\cal Q}_{ij}}\sum_{(m,n)\in{\cal L}_{q}}\frac{\partial D_{ij}^{q}}{\partial S_{mn}(q)}(\tilde{S}_{mn}(q)-S_{mn}(q)).</td>
<td></td>
</tr>
</tbody>
</table>
We will show that the RHS is nonnegative. We can re-write and lower bound the first two summations by the series of equalities and inequalities on the top of the next page.
<table>
<tbody>
<tr>
<td></td>
<td></td>
<td></td>
<td>∑w∈𝒲DFw′⋅(F~w−Fw)+∑(i,j)∈ℒ∑q∈𝒬ijDFijq′⋅(F~ij(q)−Fij(q))subscript𝑤𝒲⋅superscriptsubscript𝐷subscript𝐹𝑤′subscript~𝐹𝑤subscript𝐹𝑤subscript𝑖𝑗ℒsubscript𝑞subscript𝒬𝑖𝑗⋅superscriptsuperscriptsubscript𝐷subscript𝐹𝑖𝑗𝑞′subscript~𝐹𝑖𝑗𝑞subscript𝐹𝑖𝑗𝑞\displaystyle\sum_{w\in{\cal W}}D_{F_{w}}^{\prime}\cdot(\tilde{F}_{w}-F_{w})+\sum_{(i,j)\in{\cal L}}\sum_{q\in{\cal Q}_{ij}}{D_{F_{ij}}^{q}}^{\prime}\cdot(\tilde{F}_{ij}(q)-F_{ij}(q))</td>
<td></td>
</tr>
<tr>
<td></td>
<td></td>
<td>=(a)superscript𝑎\displaystyle\stackrel{{\scriptstyle(a)}}{{=}}</td>
<td>∑w∈𝒲DFw′⋅r¯w(ϕ~w−ϕw)+∑w∈𝒲{∑(i,j)∈ℒ∑q∈𝒬ij[DFijq′⋅t~i(w)ϕ~ij(w)μ~ij(q)]−∂E∂tw(w)⋅tw(w)}subscript𝑤𝒲⋅superscriptsubscript𝐷subscript𝐹𝑤′subscript¯𝑟𝑤subscript~italic-ϕ𝑤subscriptitalic-ϕ𝑤subscript𝑤𝒲subscript𝑖𝑗ℒsubscript𝑞subscript𝒬𝑖𝑗delimited-[]⋅superscriptsuperscriptsubscript𝐷subscript𝐹𝑖𝑗𝑞′subscript~𝑡𝑖𝑤subscript~italic-ϕ𝑖𝑗𝑤subscript~𝜇𝑖𝑗𝑞⋅𝐸subscript𝑡𝑤𝑤subscript𝑡𝑤𝑤\displaystyle\sum_{w\in{\cal W}}D_{F_{w}}^{\prime}\cdot\bar{r}_{w}(\tilde{\phi}_{w}-\phi_{w})+\sum_{w\in{\cal W}}\left\{\sum_{(i,j)\in{\cal L}}\sum_{q\in{\cal Q}_{ij}}\left[{D_{F_{ij}}^{q}}^{\prime}\cdot\tilde{t}_{i}(w)\tilde{\phi}_{ij}(w)\tilde{\mu}_{ij}(q)\right]-\frac{\partial E}{\partial t_{w}(w)}\cdot t_{w}(w)\right\}</td>
<td></td>
</tr>
<tr>
<td></td>
<td></td>
<td></td>
<td>−∑w∈𝒲{∑j≠O(w),D(w)∂E∂tj(w)[t~j(w)−∑i≠D(w)t~i(w)ϕ~ij(w)]}subscript𝑤𝒲subscript𝑗𝑂𝑤𝐷𝑤𝐸subscript𝑡𝑗𝑤delimited-[]subscript~𝑡𝑗𝑤subscript𝑖𝐷𝑤subscript~𝑡𝑖𝑤subscript~italic-ϕ𝑖𝑗𝑤\displaystyle-\sum_{w\in{\cal W}}\left\{\sum_{j\neq O(w),D(w)}\frac{\partial E}{\partial t_{j}(w)}\left[\tilde{t}_{j}(w)-\sum_{i\neq D(w)}\tilde{t}_{i}(w)\tilde{\phi}_{ij}(w)\right]\right\}</td>
<td></td>
</tr>
<tr>
<td></td>
<td></td>
<td>=(b)superscript𝑏\displaystyle\stackrel{{\scriptstyle(b)}}{{=}}</td>
<td>∑w∈𝒲r¯w{[Dw′ϕ~w+(1−ϕ~w)∑j∈𝒪O(w)ϕ~ij(w)(∑q∈𝒬ijDFijq′μ~ij(q)+∂E∂tj(w))]−[Dw′ϕw+(1−ϕw)∂E∂tw(w)]}subscript𝑤𝒲subscript¯𝑟𝑤delimited-[]superscriptsubscript𝐷𝑤′subscript~italic-ϕ𝑤1subscript~italic-ϕ𝑤subscript𝑗subscript𝒪𝑂𝑤subscript~italic-ϕ𝑖𝑗𝑤subscript𝑞subscript𝒬𝑖𝑗superscriptsuperscriptsubscript𝐷subscript𝐹𝑖𝑗𝑞′subscript~𝜇𝑖𝑗𝑞𝐸subscript𝑡𝑗𝑤delimited-[]superscriptsubscript𝐷𝑤′subscriptitalic-ϕ𝑤1subscriptitalic-ϕ𝑤𝐸subscript𝑡𝑤𝑤\displaystyle\sum_{w\in{\cal W}}\bar{r}_{w}\left\{\left[D_{w}^{\prime}\tilde{\phi}_{w}+(1-\tilde{\phi}_{w})\sum_{j\in{\cal O}_{O(w)}}\tilde{\phi}_{ij}(w)\left(\sum_{q\in{\cal Q}_{ij}}{D_{F_{ij}}^{q}}^{\prime}\tilde{\mu}_{ij}(q)+\frac{\partial E}{\partial t_{j}(w)}\right)\right]-\left[D_{w}^{\prime}\phi_{w}+(1-\phi_{w})\frac{\partial E}{\partial t_{w}(w)}\right]\right\}</td>
<td></td>
</tr>
<tr>
<td></td>
<td></td>
<td></td>
<td>+∑w∈𝒲{∑i≠O(w),D(w)t~i(w)[∑j∈𝒪iϕ~ij(w)(∑q∈𝒬ijDFijq′μ~ij(q)+∂E∂tj(w))−∂E∂ti(w)]}subscript𝑤𝒲subscript𝑖𝑂𝑤𝐷𝑤subscript~𝑡𝑖𝑤delimited-[]subscript𝑗subscript𝒪𝑖subscript~italic-ϕ𝑖𝑗𝑤subscript𝑞subscript𝒬𝑖𝑗superscriptsuperscriptsubscript𝐷subscript𝐹𝑖𝑗𝑞′subscript~𝜇𝑖𝑗𝑞𝐸subscript𝑡𝑗𝑤𝐸subscript𝑡𝑖𝑤\displaystyle+\sum_{w\in{\cal W}}\left\{\sum_{i\neq O(w),D(w)}\tilde{t}_{i}(w)\left[\sum_{j\in{\cal O}_{i}}\tilde{\phi}_{ij}(w)\left(\sum_{q\in{\cal Q}_{ij}}{D_{F_{ij}}^{q}}^{\prime}\tilde{\mu}_{ij}(q)+\frac{\partial E}{\partial t_{j}(w)}\right)-\frac{\partial E}{\partial t_{i}(w)}\right]\right\}</td>
<td></td>
</tr>
<tr>
<td></td>
<td></td>
<td>≥(c)superscript𝑐\displaystyle\stackrel{{\scriptstyle(c)}}{{\geq}}</td>
<td>∑w∈𝒲r¯w{[Dw′ϕ~w+(1−ϕ~w)∂E∂tw(w)]−[Dw′ϕw+(1−ϕw)∂E∂tw(w)]}+∑w∈𝒲{∑i≠O(w),D(w)t~i(w)[∂E∂ti(w)−∂E∂ti(w)]}subscript𝑤𝒲subscript¯𝑟𝑤delimited-[]superscriptsubscript𝐷𝑤′subscript~italic-ϕ𝑤1subscript~italic-ϕ𝑤𝐸subscript𝑡𝑤𝑤delimited-[]superscriptsubscript𝐷𝑤′subscriptitalic-ϕ𝑤1subscriptitalic-ϕ𝑤𝐸subscript𝑡𝑤𝑤subscript𝑤𝒲subscript𝑖𝑂𝑤𝐷𝑤subscript~𝑡𝑖𝑤delimited-[]𝐸subscript𝑡𝑖𝑤𝐸subscript𝑡𝑖𝑤\displaystyle\sum_{w\in{\cal W}}\bar{r}_{w}\left\{\left[D_{w}^{\prime}\tilde{\phi}_{w}+(1-\tilde{\phi}_{w})\frac{\partial E}{\partial t_{w}(w)}\right]-\left[D_{w}^{\prime}\phi_{w}+(1-\phi_{w})\frac{\partial E}{\partial t_{w}(w)}\right]\right\}+\sum_{w\in{\cal W}}\left\{\sum_{i\neq O(w),D(w)}\tilde{t}_{i}(w)\left[\frac{\partial E}{\partial t_{i}(w)}-\frac{\partial E}{\partial t_{i}(w)}\right]\right\}</td>
<td></td>
</tr>
<tr>
<td></td>
<td></td>
<td>=\displaystyle=</td>
<td>∑w∈𝒲r¯w(ϕ~w−ϕw)(DFw′−∂E∂tw(w))≥0.subscript𝑤𝒲subscript¯𝑟𝑤subscript~italic-ϕ𝑤subscriptitalic-ϕ𝑤superscriptsubscript𝐷subscript𝐹𝑤′𝐸subscript𝑡𝑤𝑤0\displaystyle\sum_{w\in{\cal W}}\bar{r}_{w}(\tilde{\phi}_{w}-\phi_{w})\left(D_{F_{w}}^{\prime}-\frac{\partial E}{\partial t_{w}(w)}\right)\geq 0.</td>
<td></td>
</tr>
</tbody>
</table>
For equality (a), we used the relation that
<table>
<tbody>
<tr>
<td></td>
<td>∑w∈𝒲∂E∂tw(w)⋅tw(w)=∑(i,j)∈ℒ∑q∈𝒬ijDFijq′⋅Fij(q).subscript𝑤𝒲⋅𝐸subscript𝑡𝑤𝑤subscript𝑡𝑤𝑤subscript𝑖𝑗ℒsubscript𝑞subscript𝒬𝑖𝑗⋅superscriptsuperscriptsubscript𝐷subscript𝐹𝑖𝑗𝑞′subscript𝐹𝑖𝑗𝑞\sum_{w\in{\cal W}}\frac{\partial E}{\partial t_{w}(w)}\cdot t_{w}(w)=\sum_{(i,j)\in{\cal L}}\sum_{q\in{\cal Q}_{ij}}{D_{F_{ij}}^{q}}^{\prime}\cdot F_{ij}(q).</td>
<td></td>
</tr>
</tbody>
</table>
and appended terms $`\frac{\partial E}{\partial{t_{j}\hspace{0pt}{(w)}}}\hspace{0pt}\left\lbrack {{{\overset{\sim}{t}}_{j}\hspace{0pt}{(w)}} - {\sum_{i \neq {D\hspace{0pt}{(w)}}}{{\overset{\sim}{t}}_{i}\hspace{0pt}{(w)}\hspace{0pt}{\overset{\sim}{\phi}}_{i\hspace{0pt}j}\hspace{0pt}{(w)}}}} \right\rbrack`$ which are all equal to zero by flow conservation constraints. Equality (b) is obtained by rewriting $`r_{w}`$ and $`{\overset{\sim}{r}}_{w}`$ as $`{\overline{r}}_{w}\hspace{0pt}{({1 - \phi_{w}})}`$ and $`{\overline{r}}_{w}\hspace{0pt}{({1 - {\overset{\sim}{\phi}}_{w}})}`$, respectively, and reorganizing the terms. Condition ([24](#S5.E24 "In Theorem 1 ‣ V-D3 Conditions for Optimality ‣ V-D Control Variables and Optimality Conditions ‣ V Cross-layer Optimization in Multi-radio Multi-hop Wireless Networks ‣ Distributed Algorithms for Spectrum Allocation, Power Control, Routing, and Congestion Control in Wireless Networks")) implies that
<table>
<tbody>
<tr>
<td></td>
<td>∂E∂ti(w)≤∑j∈𝒪i∑q∈𝒬ijDFijq′μ~ij(q)+∂E∂tj(w),𝐸subscript𝑡𝑖𝑤subscript𝑗subscript𝒪𝑖subscript𝑞subscript𝒬𝑖𝑗superscriptsuperscriptsubscript𝐷subscript𝐹𝑖𝑗𝑞′subscript~𝜇𝑖𝑗𝑞𝐸subscript𝑡𝑗𝑤\frac{\partial E}{\partial t_{i}(w)}\leq\sum_{j\in{\cal O}_{i}}\sum_{q\in{\cal Q}_{ij}}{D_{F_{ij}}^{q}}^{\prime}\tilde{\mu}_{ij}(q)+\frac{\partial E}{\partial t_{j}(w)},</td>
<td></td>
</tr>
</tbody>
</table>
by which we obtain (c). The final inequality follows from condition ([25](#S5.E25 "In Theorem 1 ‣ V-D3 Conditions for Optimality ‣ V-D Control Variables and Optimality Conditions ‣ V Cross-layer Optimization in Multi-radio Multi-hop Wireless Networks ‣ Distributed Algorithms for Spectrum Allocation, Power Control, Routing, and Congestion Control in Wireless Networks")).
Next we lower bound the third summation on the RHS of ([35](#A0.E35 "In -D Proof of Theorem 2 ‣ VII Conclusion ‣ Distributed Algorithms for Spectrum Allocation, Power Control, Routing, and Congestion Control in Wireless Networks")). To begin with, we switch the roles of $`(m,n)`$ and $`(i,j)`$ in the summation to rewrite it as
<table>
<tbody>
<tr>
<td></td>
<td>∑(i,j)∈ℒ∑q∈𝒬ij∑(m,n)∈ℒq∂Dmnq∂Sij(q)(S~ij(q)−Sij(q)).subscript𝑖𝑗ℒsubscript𝑞subscript𝒬𝑖𝑗subscript𝑚𝑛subscriptℒ𝑞superscriptsubscript𝐷𝑚𝑛𝑞subscript𝑆𝑖𝑗𝑞subscript~𝑆𝑖𝑗𝑞subscript𝑆𝑖𝑗𝑞\sum_{(i,j)\in{\cal L}}\sum_{q\in{\cal Q}_{ij}}\sum_{(m,n)\in{\cal L}_{q}}\frac{\partial D_{mn}^{q}}{\partial S_{ij}(q)}(\tilde{S}_{ij}(q)-S_{ij}(q)).</td>
<td></td>
</tr>
</tbody>
</table>
The partial derivative is computed as
<table>
<tbody>
<tr>
<td></td>
<td>∂Dmnq∂Sij(q)={Dxijq′xij(q),if (i,j)=(m,n),−Dxmnq′xmn(q)GinqPij(q)INmn(q),otherwise.superscriptsubscript𝐷𝑚𝑛𝑞subscript𝑆𝑖𝑗𝑞casessuperscriptsuperscriptsubscript𝐷subscript𝑥𝑖𝑗𝑞′subscript𝑥𝑖𝑗𝑞if 𝑖𝑗𝑚𝑛superscriptsuperscriptsubscript𝐷subscript𝑥𝑚𝑛𝑞′subscript𝑥𝑚𝑛𝑞superscriptsubscript𝐺𝑖𝑛𝑞subscript𝑃𝑖𝑗𝑞𝐼subscript𝑁𝑚𝑛𝑞otherwise\frac{\partial D_{mn}^{q}}{\partial S_{ij}(q)}=\left\{\begin{array}[]{ll}{D_{x_{ij}}^{q}}^{\prime}x_{ij}(q),&\textrm{if~{}}(i,j)=(m,n),\\ -{D_{x_{mn}}^{q}}^{\prime}x_{mn}(q)\frac{G_{in}^{q}P_{ij}(q)}{IN_{mn}(q)},&\textrm{otherwise}.\end{array}\right.</td>
<td></td>
</tr>
</tbody>
</table>
Thus, we can expand the summation as shown in the second block on the next page.
<table>
<tbody>
<tr>
<td></td>
<td></td>
<td></td>
<td>∑(i,j)∈ℒ∑q∈𝒬ij∑(m,n)∈ℒq∂Dmnq∂Sij(q)(S~ij(q)−Sij(q))subscript𝑖𝑗ℒsubscript𝑞subscript𝒬𝑖𝑗subscript𝑚𝑛subscriptℒ𝑞superscriptsubscript𝐷𝑚𝑛𝑞subscript𝑆𝑖𝑗𝑞subscript~𝑆𝑖𝑗𝑞subscript𝑆𝑖𝑗𝑞\displaystyle\sum_{(i,j)\in{\cal L}}\sum_{q\in{\cal Q}_{ij}}\sum_{(m,n)\in{\cal L}_{q}}\frac{\partial D_{mn}^{q}}{\partial S_{ij}(q)}(\tilde{S}_{ij}(q)-S_{ij}(q))</td>
<td></td>
<td rowspan="5">(36)</td>
</tr>
<tr>
<td></td>
<td></td>
<td>=\displaystyle=</td>
<td>∑(i,j)∈ℒ∑q∈𝒬ij[∑(m,n)∈ℒq(m,n)≠(i,j)−Dxmnq′xmn(q)GinqPij(q)INmn(q)+Dxijq′xij(q)]lnP~ij(q)Pij(q)subscript𝑖𝑗ℒsubscript𝑞subscript𝒬𝑖𝑗delimited-[]subscript𝑚𝑛subscriptℒ𝑞𝑚𝑛𝑖𝑗superscriptsuperscriptsubscript𝐷subscript𝑥𝑚𝑛𝑞′subscript𝑥𝑚𝑛𝑞superscriptsubscript𝐺𝑖𝑛𝑞subscript𝑃𝑖𝑗𝑞𝐼subscript𝑁𝑚𝑛𝑞superscriptsuperscriptsubscript𝐷subscript𝑥𝑖𝑗𝑞′subscript𝑥𝑖𝑗𝑞subscript~𝑃𝑖𝑗𝑞subscript𝑃𝑖𝑗𝑞\displaystyle\sum_{(i,j)\in{\cal L}}\sum_{q\in{\cal Q}_{ij}}\left[\sum_{\begin{subarray}{c}(m,n)\in{\cal L}_{q}\\ (m,n)\neq(i,j)\end{subarray}}-{D_{x_{mn}}^{q}}^{\prime}x_{mn}(q)\frac{G_{in}^{q}P_{ij}(q)}{IN_{mn}(q)}+{D_{x_{ij}}^{q}}^{\prime}x_{ij}(q)\right]\ln\frac{\tilde{P}_{ij}(q)}{P_{ij}(q)}</td>
<td></td>
</tr>
<tr>
<td></td>
<td></td>
<td>=\displaystyle=</td>
<td>∑(i,j)∈ℒ∑q∈𝒬ij[∑(m,n)∈ℒq−Dxmnq′xmn(q)GinqPij(q)INmn(q)+Dxijq′xij(q)(1+xij(q))]lnP~ij(q)Pij(q)subscript𝑖𝑗ℒsubscript𝑞subscript𝒬𝑖𝑗delimited-[]subscript𝑚𝑛subscriptℒ𝑞superscriptsuperscriptsubscript𝐷subscript𝑥𝑚𝑛𝑞′subscript𝑥𝑚𝑛𝑞superscriptsubscript𝐺𝑖𝑛𝑞subscript𝑃𝑖𝑗𝑞𝐼subscript𝑁𝑚𝑛𝑞superscriptsuperscriptsubscript𝐷subscript𝑥𝑖𝑗𝑞′subscript𝑥𝑖𝑗𝑞1subscript𝑥𝑖𝑗𝑞subscript~𝑃𝑖𝑗𝑞subscript𝑃𝑖𝑗𝑞\displaystyle\sum_{(i,j)\in{\cal L}}\sum_{q\in{\cal Q}_{ij}}\left[\sum_{(m,n)\in{\cal L}_{q}}-{D_{x_{mn}}^{q}}^{\prime}x_{mn}(q)\frac{G_{in}^{q}P_{ij}(q)}{IN_{mn}(q)}+{D_{x_{ij}}^{q}}^{\prime}x_{ij}(q)(1+x_{ij}(q))\right]\ln\frac{\tilde{P}_{ij}(q)}{P_{ij}(q)}</td>
<td></td>
</tr>
<tr>
<td></td>
<td></td>
<td>=\displaystyle=</td>
<td>∑(i,j)∈ℒ∑q∈𝒬ij[∑n∈𝒩GinqMSGnq+δηij(q)]Pij(q)lnρ~i(q)η~ij(q)ρi(q)ηij(q)subscript𝑖𝑗ℒsubscript𝑞subscript𝒬𝑖𝑗delimited-[]subscript𝑛𝒩superscriptsubscript𝐺𝑖𝑛𝑞𝑀𝑆superscriptsubscript𝐺𝑛𝑞𝛿subscript𝜂𝑖𝑗𝑞subscript𝑃𝑖𝑗𝑞subscript~𝜌𝑖𝑞subscript~𝜂𝑖𝑗𝑞subscript𝜌𝑖𝑞subscript𝜂𝑖𝑗𝑞\displaystyle\sum_{(i,j)\in{\cal L}}\sum_{q\in{\cal Q}_{ij}}\left[\sum_{n\in{\cal N}}G_{in}^{q}MSG_{n}^{q}+\delta\eta_{ij}(q)\right]P_{ij}(q)\ln\frac{\tilde{\rho}_{i}(q)\tilde{\eta}_{ij}(q)}{\rho_{i}(q)\eta_{ij}(q)}</td>
<td></td>
</tr>
<tr>
<td></td>
<td></td>
<td>=\displaystyle=</td>
<td>∑i∈𝒩P¯i∑q∈𝒪𝒞iρi(q)∑j∈𝒪i(i,j)∈ℒq[∑n∈𝒩GinqMSGnq+δηij(q)]ηij(q)lnη~ij(q)ηij(q)+∑i∈𝒩∑q∈𝒪𝒞iρi(q)δρi(q)lnρ~i(q)ρi(q).subscript𝑖𝒩subscript¯𝑃𝑖subscript𝑞𝒪subscript𝒞𝑖subscript𝜌𝑖𝑞subscript𝑗subscript𝒪𝑖𝑖𝑗subscriptℒ𝑞delimited-[]subscript𝑛𝒩superscriptsubscript𝐺𝑖𝑛𝑞𝑀𝑆superscriptsubscript𝐺𝑛𝑞𝛿subscript𝜂𝑖𝑗𝑞subscript𝜂𝑖𝑗𝑞subscript~𝜂𝑖𝑗𝑞subscript𝜂𝑖𝑗𝑞subscript𝑖𝒩subscript𝑞𝒪subscript𝒞𝑖subscript𝜌𝑖𝑞𝛿subscript𝜌𝑖𝑞subscript~𝜌𝑖𝑞subscript𝜌𝑖𝑞\displaystyle\sum_{i\in{\cal N}}\bar{P}_{i}\sum_{q\in{\cal OC}_{i}}\rho_{i}(q)\sum_{\begin{subarray}{c}j\in{\cal O}_{i}\\ (i,j)\in{\cal L}_{q}\end{subarray}}\left[\sum_{n\in{\cal N}}G_{in}^{q}MSG_{n}^{q}+\delta\eta_{ij}(q)\right]\eta_{ij}(q)\ln\frac{\tilde{\eta}_{ij}(q)}{\eta_{ij}(q)}+\sum_{i\in{\cal N}}\sum_{q\in{\cal OC}_{i}}\rho_{i}(q)\delta\rho_{i}(q)\ln\frac{\tilde{\rho}_{i}(q)}{\rho_{i}(q)}.</td>
<td></td>
</tr>
</tbody>
</table>
For the first summation on the RHS of ([36](#A0.E36 "In -D Proof of Theorem 2 ‣ VII Conclusion ‣ Distributed Algorithms for Spectrum Allocation, Power Control, Routing, and Congestion Control in Wireless Networks")), we use the convention that $`{0/0} = 1`$ and $`{y/0} \geq 1`$ for all $`y \geq 0`$. As a consequence, the summand vanishes if $`{{\overset{\sim}{\eta}}_{i\hspace{0pt}j}\hspace{0pt}{(q)}} = {\eta_{i\hspace{0pt}j}\hspace{0pt}{(q)}} = 0`$. Moreover, for those $`{(i,j)} \in \mathcal{L}_{q}`$ but $`{\eta_{i\hspace{0pt}j}\hspace{0pt}{(q)}} = 0`$, we can lower bound the summand by replacing $`\delta\hspace{0pt}\eta_{i\hspace{0pt}j}\hspace{0pt}{(q)}`$ with $`\gamma_{i}\hspace{0pt}{(q)}`$ (cf. condition ([28](#S5.E28 "In Theorem 1 ‣ V-D3 Conditions for Optimality ‣ V-D Control Variables and Optimality Conditions ‣ V Cross-layer Optimization in Multi-radio Multi-hop Wireless Networks ‣ Distributed Algorithms for Spectrum Allocation, Power Control, Routing, and Congestion Control in Wireless Networks"))) to get
<table>
<tbody>
<tr>
<td></td>
<td>[∑n∈𝒩GinqMSGnq+γi(q)]ηij(q)lnη~ij(q)ηij(q).delimited-[]subscript𝑛𝒩superscriptsubscript𝐺𝑖𝑛𝑞𝑀𝑆superscriptsubscript𝐺𝑛𝑞subscript𝛾𝑖𝑞subscript𝜂𝑖𝑗𝑞subscript~𝜂𝑖𝑗𝑞subscript𝜂𝑖𝑗𝑞\left[\sum_{n\in{\cal N}}G_{in}^{q}MSG_{n}^{q}+\gamma_{i}(q)\right]\eta_{ij}(q)\ln\frac{\tilde{\eta}_{ij}(q)}{\eta_{ij}(q)}.</td>
<td></td>
</tr>
</tbody>
</table>
It is implicit from ([16](#S5.E16 "In V-D1 Power Control Variables ‣ V-D Control Variables and Optimality Conditions ‣ V Cross-layer Optimization in Multi-radio Multi-hop Wireless Networks ‣ Distributed Algorithms for Spectrum Allocation, Power Control, Routing, and Congestion Control in Wireless Networks")) and ([28](#S5.E28 "In Theorem 1 ‣ V-D3 Conditions for Optimality ‣ V-D Control Variables and Optimality Conditions ‣ V Cross-layer Optimization in Multi-radio Multi-hop Wireless Networks ‣ Distributed Algorithms for Spectrum Allocation, Power Control, Routing, and Congestion Control in Wireless Networks")) that $`{\delta\hspace{0pt}\rho_{i}\hspace{0pt}{(q)}} = {{\overline{P}}_{i}\hspace{0pt}{\lbrack{{\sum_{n \in \mathcal{N}}{G_{i\hspace{0pt}n}^{q}\hspace{0pt}M\hspace{0pt}S\hspace{0pt}G_{n}^{q}}} + {\gamma_{i}\hspace{0pt}{(q)}}}\rbrack}}`$. We thus can lower bound the first summation in ([36](#A0.E36 "In -D Proof of Theorem 2 ‣ VII Conclusion ‣ Distributed Algorithms for Spectrum Allocation, Power Control, Routing, and Congestion Control in Wireless Networks")) by
<table>
<tbody>
<tr>
<td></td>
<td></td>
<td></td>
<td>∑i∈𝒩∑q∈𝒪𝒞iρi(q)δρi(q)∑j∈𝒪i(i,j)∈ℒqηij(q)lnη~ij(q)ηij(q)subscript𝑖𝒩subscript𝑞𝒪subscript𝒞𝑖subscript𝜌𝑖𝑞𝛿subscript𝜌𝑖𝑞subscript𝑗subscript𝒪𝑖𝑖𝑗subscriptℒ𝑞subscript𝜂𝑖𝑗𝑞subscript~𝜂𝑖𝑗𝑞subscript𝜂𝑖𝑗𝑞\displaystyle\sum_{i\in{\cal N}}\sum_{q\in{\cal OC}_{i}}\rho_{i}(q)\delta\rho_{i}(q)\sum_{\begin{subarray}{c}j\in{\cal O}_{i}\\ (i,j)\in{\cal L}_{q}\end{subarray}}\eta_{ij}(q)\ln\frac{\tilde{\eta}_{ij}(q)}{\eta_{ij}(q)}</td>
<td></td>
</tr>
<tr>
<td></td>
<td></td>
<td>≥\displaystyle\geq</td>
<td>∑i∈𝒩∑q∈𝒪𝒞iρi(q)δρi(q)∑j∈𝒪i(i,j)∈ℒqηij(q)(η~ij(q)ηij(q)−1)=0.subscript𝑖𝒩subscript𝑞𝒪subscript𝒞𝑖subscript𝜌𝑖𝑞𝛿subscript𝜌𝑖𝑞subscript𝑗subscript𝒪𝑖𝑖𝑗subscriptℒ𝑞subscript𝜂𝑖𝑗𝑞subscript~𝜂𝑖𝑗𝑞subscript𝜂𝑖𝑗𝑞10\displaystyle\sum_{i\in{\cal N}}\sum_{q\in{\cal OC}_{i}}\rho_{i}(q)\delta\rho_{i}(q)\sum_{\begin{subarray}{c}j\in{\cal O}_{i}\\ (i,j)\in{\cal L}_{q}\end{subarray}}\eta_{ij}(q)\left(\frac{\tilde{\eta}_{ij}(q)}{\eta_{ij}(q)}-1\right)=0.</td>
<td></td>
</tr>
</tbody>
</table>
The inequality follows from the relation $`{\ln x} \leq {x - 1}`$ and that $`{\rho_{i}\hspace{0pt}{(q)}\hspace{0pt}\delta\hspace{0pt}\rho_{i}\hspace{0pt}{(q)}} \leq 0`$ due to conditions ([26](#S5.E26 "In Theorem 1 ‣ V-D3 Conditions for Optimality ‣ V-D Control Variables and Optimality Conditions ‣ V Cross-layer Optimization in Multi-radio Multi-hop Wireless Networks ‣ Distributed Algorithms for Spectrum Allocation, Power Control, Routing, and Congestion Control in Wireless Networks")) and ([27](#S5.E27 "In Theorem 1 ‣ V-D3 Conditions for Optimality ‣ V-D Control Variables and Optimality Conditions ‣ V Cross-layer Optimization in Multi-radio Multi-hop Wireless Networks ‣ Distributed Algorithms for Spectrum Allocation, Power Control, Routing, and Congestion Control in Wireless Networks")).
Now consider the second summation on the RHS of ([36](#A0.E36 "In -D Proof of Theorem 2 ‣ VII Conclusion ‣ Distributed Algorithms for Spectrum Allocation, Power Control, Routing, and Congestion Control in Wireless Networks")). For each $`i \in \mathcal{N}`$, let the summands with $`{\rho_{i}\hspace{0pt}{(q)}} = 0`$ be lower bounded by
<table>
<tbody>
<tr>
<td></td>
<td>ρi(q)λilnρ~i(q)ρi(q).subscript𝜌𝑖𝑞subscript𝜆𝑖subscript~𝜌𝑖𝑞subscript𝜌𝑖𝑞\rho_{i}(q)\lambda_{i}\ln\frac{\tilde{\rho}_{i}(q)}{\rho_{i}(q)}.</td>
<td></td>
</tr>
</tbody>
</table>
By condition ([26](#S5.E26 "In Theorem 1 ‣ V-D3 Conditions for Optimality ‣ V-D Control Variables and Optimality Conditions ‣ V Cross-layer Optimization in Multi-radio Multi-hop Wireless Networks ‣ Distributed Algorithms for Spectrum Allocation, Power Control, Routing, and Congestion Control in Wireless Networks")), summands with $`{\rho_{i}\hspace{0pt}{(q)}} > 0`$ are equal to
<table>
<tbody>
<tr>
<td></td>
<td>ρi(q)λilnρ~i(q)ρi(q).subscript𝜌𝑖𝑞subscript𝜆𝑖subscript~𝜌𝑖𝑞subscript𝜌𝑖𝑞\rho_{i}(q)\lambda_{i}\ln\frac{\tilde{\rho}_{i}(q)}{\rho_{i}(q)}.</td>
<td></td>
</tr>
</tbody>
</table>
Therefore, the whole summation is lower bounded by
<table>
<tbody>
<tr>
<td></td>
<td></td>
<td></td>
<td>∑i∈𝒩∑q∈𝒪𝒞iρi(q)λilnρ~i(q)ρi(q)subscript𝑖𝒩subscript𝑞𝒪subscript𝒞𝑖subscript𝜌𝑖𝑞subscript𝜆𝑖subscript~𝜌𝑖𝑞subscript𝜌𝑖𝑞\displaystyle\sum_{i\in{\cal N}}\sum_{q\in{\cal OC}_{i}}\rho_{i}(q)\lambda_{i}\ln\frac{\tilde{\rho}_{i}(q)}{\rho_{i}(q)}</td>
<td></td>
</tr>
<tr>
<td></td>
<td></td>
<td>≥\displaystyle\geq</td>
<td>∑i∈𝒩∑q∈𝒪𝒞iρi(q)λi(ρ~i(q)ρi(q)−1)subscript𝑖𝒩subscript𝑞𝒪subscript𝒞𝑖subscript𝜌𝑖𝑞subscript𝜆𝑖subscript~𝜌𝑖𝑞subscript𝜌𝑖𝑞1\displaystyle\sum_{i\in{\cal N}}\sum_{q\in{\cal OC}_{i}}\rho_{i}(q)\lambda_{i}\left(\frac{\tilde{\rho}_{i}(q)}{\rho_{i}(q)}-1\right)</td>
<td></td>
</tr>
<tr>
<td></td>
<td></td>
<td>=\displaystyle=</td>
<td>∑i∈𝒩λi(∑q∈𝒪𝒞iρ~i(q)−∑q∈𝒪𝒞iρi(q))≥0.subscript𝑖𝒩subscript𝜆𝑖subscript𝑞𝒪subscript𝒞𝑖subscript~𝜌𝑖𝑞subscript𝑞𝒪subscript𝒞𝑖subscript𝜌𝑖𝑞0\displaystyle\sum_{i\in{\cal N}}\lambda_{i}\left(\sum_{q\in{\cal OC}_{i}}\tilde{\rho}_{i}(q)-\sum_{q\in{\cal OC}_{i}}\rho_{i}(q)\right)\geq 0.</td>
<td></td>
</tr>
</tbody>
</table>
The second to last inequality is obtained by noting that $`{\rho_{i}\hspace{0pt}{(q)}\hspace{0pt}\lambda_{i}} \leq 0`$ due to condition ([27](#S5.E27 "In Theorem 1 ‣ V-D3 Conditions for Optimality ‣ V-D Control Variables and Optimality Conditions ‣ V Cross-layer Optimization in Multi-radio Multi-hop Wireless Networks ‣ Distributed Algorithms for Spectrum Allocation, Power Control, Routing, and Congestion Control in Wireless Networks")) and the identity $`{\ln x} \leq {x - 1}`$. The last inequality follows also from condition ([27](#S5.E27 "In Theorem 1 ‣ V-D3 Conditions for Optimality ‣ V-D Control Variables and Optimality Conditions ‣ V Cross-layer Optimization in Multi-radio Multi-hop Wireless Networks ‣ Distributed Algorithms for Spectrum Allocation, Power Control, Routing, and Congestion Control in Wireless Networks")).
Thus, we have shown that the cost incurred by an arbitrary feasible configuration is greater than or equal to the cost by a configuration that satisfies the conditions in Theorem [2](#Thmtheorem2 "Theorem 2 ‣ V-D3 Conditions for Optimality ‣ V-D Control Variables and Optimality Conditions ‣ V Cross-layer Optimization in Multi-radio Multi-hop Wireless Networks ‣ Distributed Algorithms for Spectrum Allocation, Power Control, Routing, and Congestion Control in Wireless Networks"). So the proof is complete.
## References
- \[1\]
M. Johansson, L. Xiao, and S. Boyd, “Simultaneous routing and power allocation in CDMA wireless data networks,” in Proceedings of IEEE International Conference on Communications, vol. 1, May 2003.
- \[2\]
M. Chiang, “To layer or not to layer: Balancing transport and physical layers in wireless multihop networks,” in Proceedings of IEEE INFOCOM 2004, vol. 4, Mar. 2004.
- \[3\]
Y. Xi and E. M. Yeh, “Optimal distributed power control and routing in wireless networks,” in Proceedings of the 2006 International Symposium on Information Theory, (Seattle, WA), July 2006.
- \[4\]
A. Goldsmith, Wireless Communications.
Cambridge University Press, 2005.
- \[5\]
L. Chen, S. H. Low, M. Chiang, and J. C. Doyle, “Jointly optimal congestion control, routing, and scheduling for wireless ad hoc networks,” in Proceedings of IEEE INFOCOM 2006, Apr. 2006.
- \[6\]
L. Bui, A. Eryilmaz, R. Srikant, and X. Wu, “Joint asynchronous congestion control and distributed scheduling for multi-hop wireless networks,” in Proceedings of IEEE INFOCOM 2006, Apr. 2006.
- \[7\]
B. Hajek and G. Sasaki, “Link scheduling in polynomial time,” IEEE Transactions on Information Theory, vol. 34, pp. 910–917, Sept. 1988.
- \[8\]
E. Modiano, D. Shah, and G. Zussman, “Maximizing throughput in wireless networks via gossiping,” in Proceedings of ACM SIGMETRICS 2006, June 2006.
- \[9\]
S. Ramanathan, “A unified framework and algorithm for (t/f/c)dma channel assignment in wireless networks,” in Proceedings of IEEE INFOCOM 1997, vol. 2, Apr. 1997.
- \[10\]
L. Tassiulas and A. Ephremides, “Stability properties of constrained queueing systems and scheduling policies for maximum throughput in multihop radio networks,” IEEE Transactions on Automatic Control, vol. 37, no. 12, pp. 1936–1948, 1992.
- \[11\]
N. McKeown, V. Anantharam, and J. Walrand, “Achieving 100% throughput in an input-queued switch,” in Proceedings of IEEE INFOCOM 1996, vol. 1, pp. 296–302, Mar. 1996.
- \[12\] D. N. Tse, “Multi-user diverstiy and proportional fairness.” US Patent 6449490.
- \[13\]
V. Subramanian and R. Agrawal, “A stochastic approximation analysis of channel condition aware wireless scheduling algorithms,” in Proceedings of INFORMS Telecommunications Conference, 2002.
- \[14\]
X. Liu, E. K. Chan, and N. B. Shroff, “Opportunistic transmission scheduling with resource-sharing constraints in wireless networks,” IEEE Journal on Selected Areas in Communications, Oct. 2001.
- \[15\]
Y. Liu and E. Knightly, “Opportunistic fair scheduling over multiple wireless channels,” in Proceedings of IEEE INFOCOM 2003, Apr. 2003.
- \[16\]
S. Borst, “User-level performance of channel-aware scheduling algorithms in wireless data networks,” in Proceedings of IEEE INFOCOM 2003, Apr. 2003.
- \[17\]
T. ElBatt and A. Ephremides, “Joint scheduling and power control for wireless ad-hoc networks,” in Proceedings of IEEE INFOCOM 2002, pp. 976–984, June 2002.
- \[18\]
E. Arikan, “Some complexity results about packet radio networks,” IEEE Transactions on Information Theory, vol. 31, pp. 910–918, 1984.
- \[19\]
M. J. Neely, E. Modiano, and C. Rohrs, “Tradeoffs in delay guarantees and computation complexity for N $`\times`$ N packet switches,” in Proceedings of the Conference on Information Sciences and Systems, (Princeton), Mar. 2002.
- \[20\]
P. Chaporkar, K. Kar, and S. Sarkar, “Throughput guarantees through maximal scheduling in wireless networks,” in Proceedings of the 2005 Allerton Conference on Communication, Control and Computing, Sept. 2005.
- \[21\]
X. Lin and N. Shroff, “The impact of imperfect scheduling on cross-layer rate control in wireless networks,” in Proceedings of IEEE INFOCOM 2005, vol. 3, pp. 1804–1814, Mar. 2005.
- \[22\]
W. K. Hale, “Frequency assignment: theory and applications,” Proceedings of the IEEE, vol. 68, no. 12, pp. 1497–1514, 1980.
- \[23\]
M. B. Cozzens and D.-I. Wang, “The general channel assignment problem,” Congressus Numerantium, vol. 41, pp. 115–129, 1984.
- \[24\]
M. Alicherry, R. Bhatia, and L. Li, “Joint channel assignment and routing for throughput optimization in multi-radio wireless mesh networks,” in Proceedings of ACM Mobicom 2005, (Cologne, Germany), pp. 58–72, 2005.
- \[25\]
M. Kodialam and T. Nandagopal, “Characterizing the capacity region in multi-radio multi-channel wireless mesh networks,” in Proceedings of the ACM Mobicom 2005, (Cologne, Germany), pp. 73–87, 2005.
- \[26\]
M. R. Garey and D. S. Johnson, Computers and intractability: A guide to the theory of NP-completeness.
W. H. Freeman and Company, 1979.
- \[27\]
E. Sperner, “Ein satz über untermegen einer endlichen menge,” Math. Zeitschr, vol. 27, pp. 544–548, 1928.
- \[28\]
W. Feller, An Introduction to Probability Theory and Its Applications.
Wiley, 3rd ed., 1968.
- \[29\]
D. Bertsekas and R. Gallager, Data Networks.
Prentice Hall, second ed., 1992.
- \[30\]
Y. Xi and E. M. Yeh, “Optimal capacity allocation, routing, and congestion control in wireless networks,” in Proceedings of the 2006 International Symposium on Information Theory, (Seattle, WA), July 2006.
- \[31\]
L. Kleinrock, Communication Net: Stochastic Message Flow and Delay.
New York: McGraw-Hill, 1964.
- \[32\]
R. Gallager, “A minimum delay routing algorithm using distributed computation,” IEEE Transactions on Communications, vol. 25, no. 1, pp. 73–85, 1977.
- \[33\]
D. Bertsekas, E. Gafni, and R. Gallager, “Second derivative algorithm for minimum delay distributed routing in networks,” IEEE Transactions on Communications, vol. 32, no. 8, pp. 911–919, 1984.
- \[34\]
D. Bertsekas, Nonlinear Programming.
Athena Scientific, second ed., 1999.
- \[35\]
R. P. Grimaldi, Discrete and Combinatorial Mathematics.
Addison Wesley, fifth ed., 2003.
<table>
<tbody>
<tr>
<td></td>
<td>Yufang Xi (S’06) received the B. Sci degree in Electronic Engineering from Tsinghua University, Beijing, China, in 2003, and the M. Sci and M. Phil degrees in Electrical Engineering from Yale University, New Haven, CT, in 2005 and 2006, respectively. He is currently a Ph.D. candidate at the Department of Electrical Engineering at Yale. His research interests include cross-layer optimization and distributed resource allocation algorithms for wireless networks.</td>
</tr>
</tbody>
</table>
<table>
<tbody>
<tr>
<td></td>
<td>Edmund M. Yeh received his B.S. in Electrical Engineering with Distinction from Stanford University in 1994, M.Phil in Engineering from the University of Cambridge in 1995, and Ph.D. in Electrical Engineering and Computer Science from MIT in 2001. Since July 2001, he has been on the faculty at Yale University, New Haven, Connecticut, where he is currently an Associate Professor of Electrical Engineering and Computer Science. Dr. Yeh is a recipient of the Army Research Office (ARO) Young Investigator Program (YIP) Award (2003), the Winston Churchill Scholarship (1994), the National Science Foundation and Office of Naval Research Fellowships (1994) for graduate study, the Frederick E. Terman Award from Stanford University (1994) and the Barry M. Goldwater Scholarship from the United States Congress (1993). Dr. Yeh is a member of Phi Beta Kappa, Tau Beta Pi, and IEEE. He has been visiting faculty at MIT, Princeton University, University of California at Berkeley, and Swiss Federal Institute of Technology, Lausanne.</td>
</tr>
</tbody>
</table> |
0705.0315 | cs.NI | cyber | # WDM and Directed Star Arboricity
Omid Amini <sup>,1</sup>
CNRS-DMA, École Normale Supérieure, Paris, France. This work was done while this author was PhD student at École Polytechnique and Projet Mascotte, INRIA Sophia-Antipolis, France. omid.amini@m4x.org
Frédéric Havet <sup>,1,2</sup>
Projet Mascotte, CNRS/INRIA/UNSA, INRIA Sophia-Antipolis, France. fhavet@sophia.inria.fr
Florian Huc <sup>,1</sup>
Université de Genève, Switzerland. This work was done while this author was PhD student in Projet Mascotte, INRIA Sophia-Antipolis, France. florian.huc@unige.ch
Stéphan Thomassé <sup>,2</sup>
LIRMM, Montpellier, France. thomasse@lirmm.fr
<sup>1</sup> These authors were partially supported by the European project AEOLUS.
<sup>2</sup> Supported by the ANR Blanc AGAPE.
###### Abstract
A digraph is $`m`$-labelled if every arc is labelled by an integer in $`\{ 1,\ldots,m\}`$. Motivated by wavelength assignment for multicasts in optical networks, we introduce and study $`n`$-fibre colourings of labelled digraphs. These are colourings of the arcs of $`D`$ such that at each vertex $`v`$, and for each colour $`\alpha`$, $`{{i\hspace{0pt}n\hspace{0pt}{(v,\alpha)}} + {o\hspace{0pt}u\hspace{0pt}t\hspace{0pt}{(v,\alpha)}}} \leq n`$ with $`i\hspace{0pt}n\hspace{0pt}{(v,\alpha)}`$ the number of arcs coloured $`\alpha`$ entering $`v`$ and $`o\hspace{0pt}u\hspace{0pt}t\hspace{0pt}{(v,\alpha)}`$ the number of labels $`l`$ such that there is at least one arc of label $`l`$ leaving $`v`$ and coloured with $`\alpha`$. The problem is to find the minimum number of colours $`\lambda_{n}\hspace{0pt}{(D)}`$ such that the $`m`$-labelled digraph $`D`$ has an $`n`$-fibre colouring. In the particular case when $`D`$ is $`1`$-labelled, $`\lambda_{1}\hspace{0pt}{(D)}`$ is called the directed star arboricity of $`D`$, and is denoted by $`d\hspace{0pt}s\hspace{0pt}t\hspace{0pt}{(D)}`$. We first show that $`{d\hspace{0pt}s\hspace{0pt}t\hspace{0pt}{(D)}} \leq {{2\hspace{0pt}\Delta^{-}\hspace{0pt}{(D)}} + 1}`$, and conjecture that if $`{\Delta^{-}\hspace{0pt}{(D)}} \geq 2`$, then $`{d\hspace{0pt}s\hspace{0pt}t\hspace{0pt}{(D)}} \leq {2\hspace{0pt}\Delta^{-}\hspace{0pt}{(D)}}`$. We also prove that for a subcubic digraph $`D`$, then $`{d\hspace{0pt}s\hspace{0pt}t\hspace{0pt}{(D)}} \leq 3`$, and that if $`{{\Delta^{+}\hspace{0pt}{(D)}},{\Delta^{-}\hspace{0pt}{(D)}}} \leq 2`$, then $`{d\hspace{0pt}s\hspace{0pt}t\hspace{0pt}{(D)}} \leq 4`$. Finally, we study $`{\lambda_{n}\hspace{0pt}{(m,k)}} = {\max{\{{\left. {\lambda_{n}\hspace{0pt}{(D)}} \middle| {D\hspace{0pt}{{\text{~is~}m\text{-labelled}}\text{~and~}}\hspace{0pt}\Delta^{-}\hspace{0pt}{(D)}} \right. \leq k}\}}}`$. We show that if $`m \geq n`$, then $`\left\lceil {{\frac{m}{n}\hspace{0pt}\left\lceil \frac{k}{n} \right\rceil} + \frac{k}{n}} \right\rceil \leq {\lambda_{n}\hspace{0pt}{(m,k)}} \leq {\left\lceil {{\frac{m}{n}\hspace{0pt}\left\lceil \frac{k}{n} \right\rceil} + \frac{k}{n}} \right\rceil + {C\hspace{0pt}\frac{m^{2}\hspace{0pt}{\log k}}{n}}}`$ for some constant $`C`$. We conjecture that the lower bound should be the right value of $`\lambda_{n}\hspace{0pt}{(m,k)}`$.
## 1 Introduction
The origin of this paper is the study of wavelength assignment for multicasts in star networks. We are given a star network in which a central node is connected by optical fibres to a set of nodes $`V`$. The nodes of $`V`$ communicates together using a technology called WDM (wavelength-division multiplexing), which allows to send different signals at the same time through the same fibre but on different wavelengths. The central node or hub is an all-optical transmitter which can redirect a signal arriving from a node on a particular wavelength to some (one or more) of the other nodes on the same wavelength. It means that the central node is able to duplicate a message incoming on a wavelength to different fibres without changing its wavelength. Therefore if a node $`v`$ sends a multicast to a set of nodes $`S\hspace{0pt}{(v)}`$, $`v`$ should send the message to the central node on a set of wavelengths so that the central node redirect it to each node of $`S\hspace{0pt}{(v)}`$ using one of these wavelengths. The aim is to minimise the total number of used wavelengths. We refer to Brandt and Gonzalez \[[4](#bib.bib4)\] for a more complete description of the model and for some partial results. In what follows, we will briefly explain the main contributions of this paper.
We first study the basic case when there is a unique fibre between the central node and each node of $`V`$ and each vertex $`v`$ sends a unique multicast $`M\hspace{0pt}{(v)}`$ to a set $`S\hspace{0pt}{(v)}`$ of nodes. In this case, the problem becomes equivalent to directed star colouring: let $`D`$ be the digraph with vertex set $`V`$ such that the outneighbourhood of a vertex $`v`$ is $`S\hspace{0pt}{(v)}`$. We note that $`D`$ is a digraph and not a multidigraph, i.e., there are no parallel arcs in $`D`$, as $`S\hspace{0pt}{(v)}`$ is a set. The problem is then to find the smallest $`k`$ such that there exists a mapping $`\phi:{{A\hspace{0pt}{(D)}}\rightarrow{\{ 1,\ldots,k\}}}`$ satisfying the following two conditions:
- $`(i)`$
For all pair of arcs $`u\hspace{0pt}v`$ and $`v\hspace{0pt}w`$, $`{\phi\hspace{0pt}{({u\hspace{0pt}v})}} \neq {\phi\hspace{0pt}{({v\hspace{0pt}w})}}`$;
- $`({i\hspace{0pt}i})`$
For all pair of arcs $`u\hspace{0pt}v`$ and $`u'\hspace{0pt}v`$, $`{\phi\hspace{0pt}{({u\hspace{0pt}v})}} \neq {\phi\hspace{0pt}{({u'\hspace{0pt}v})}}`$.
Such a mapping is called directed star $`k`$-colouring. The directed star arboricity of a digraph $`D`$, denoted by $`d\hspace{0pt}s\hspace{0pt}t\hspace{0pt}{(D)}`$, is the minimum integer $`k`$ such that there exists a directed star $`k`$-colouring. This notion has been introduced by Guiduli in \[[6](#bib.bib6)\] and is an analog of the star arboricity defined by Algor and Alon in \[[1](#bib.bib1)\].
The indegree of a vertex $`v`$, $`d^{-}\hspace{0pt}{(v)}`$, corresponds to the number of multicasts that $`v`$ receives. A sensible assumption on the model is that a node receives a bounded number of multicasts. Hence, Brandt and Gonzalez \[[4](#bib.bib4)\] studied the directed star arboricity of a digraph $`D`$ with regards to its maximum indegree. The maximum indegree of a digraph $`D`$, denoted by $`\Delta^{-}\hspace{0pt}{(D)}`$ or simply $`\Delta^{-}`$ when $`D`$ is clearly understood from the context, is $`\max{\{{\left. {d^{-}\hspace{0pt}{(v)}} \middle| v \right. \in {V\hspace{0pt}{(D)}}}\}}`$. Brandt and Gonzalez showed that $`{d\hspace{0pt}s\hspace{0pt}t\hspace{0pt}{(D)}} \leq {\lceil{{5\hspace{0pt}\Delta^{-}}/2}\rceil}`$. This upper bound is tight if $`\Delta^{-} = 1`$, because odd circuits have directed star arboricity three. However, as we will show in Section [2](#S2 "2 Directed Star Arboricity of Digraphs with Bounded Indegrees ‣ WDM and Directed Star Arboricity"), the upper bound can be improved for larger values of $`\Delta^{-}`$.
######
Theorem 1
Every digraph $`D`$ satisfies $`{d\hspace{0pt}s\hspace{0pt}t\hspace{0pt}{(D)}} \leq {{2\hspace{0pt}\Delta^{-}} + 1}`$.
We conjecture that
######
Conjecture 2
Every digraph $`D`$ with maximum indegree $`\Delta^{-} \geq 2`$ satisfies $`{d\hspace{0pt}s\hspace{0pt}t\hspace{0pt}{(D)}} \leq {2\hspace{0pt}\Delta^{-}}`$.
This conjecture would be tight as Brandt \[[3](#bib.bib3)\] showed that for every $`\Delta^{-}`$, there is an acyclic digraph $`D_{\Delta^{-}}`$ with maximum indegree $`\Delta^{-}`$ and $`{d\hspace{0pt}s\hspace{0pt}t\hspace{0pt}{(D_{\Delta^{-}})}} = {2\hspace{0pt}\Delta^{-}}`$. His construction is the special case for $`n = m = 1`$ of the construction given in Proposition [17](#Thmtheorem17 "Proposition 17 ‣ 5 Multiple Fibres ‣ WDM and Directed Star Arboricity"). We settle Conjecture [2](#Thmtheorem2 "Conjecture 2 ‣ 1 Introduction ‣ WDM and Directed Star Arboricity") for acyclic digraphs in Section [2](#S2 "2 Directed Star Arboricity of Digraphs with Bounded Indegrees ‣ WDM and Directed Star Arboricity"). So combined with Brand’s construction, $`2\hspace{0pt}\Delta^{-}`$ is the best bound we can expect for acyclic digraphs.
######
Remark 3
Let us note at this point that we restrict ourselves to simple digraphs, i.e., we allow circuits of length two but multiple arcs are not permitted. When multiple arcs are allowed, all the bounds above do not hold. Indeed, given an integer $`\Delta^{-}`$, the multidigraph $`T_{\Delta^{-}}`$ with three vertices $`u`$, $`v`$ and $`w`$, and $`\Delta^{-}`$ parallel arcs to each of $`u\hspace{0pt}v`$, $`v\hspace{0pt}w`$ and $`w\hspace{0pt}u`$ satisfies $`{d\hspace{0pt}s\hspace{0pt}t\hspace{0pt}{(T_{\Delta^{-}})}} = {3\hspace{0pt}\Delta^{-}}`$. Moreover, this example is extremal since every multidigraph satisfies $`{d\hspace{0pt}s\hspace{0pt}t\hspace{0pt}{(D)}} \leq {3\hspace{0pt}\Delta^{-}}`$. This can be shown by induction: pick a vertex $`v`$ with outdegree at most its indegree. (Such a vertex exists since $`{\sum_{u \in {V\hspace{0pt}{(D)}}}{d^{+}\hspace{0pt}{(u)}}} = {\sum_{u \in {V\hspace{0pt}{(D)}}}{d^{-}\hspace{0pt}{(u)}}}`$.) If $`v`$ has no inneighbour, then $`v`$ is isolated, and we can remove $`v`$ and apply induction. Otherwise, we consider any arc $`u\hspace{0pt}v`$. The colour of $`u\hspace{0pt}v`$ must be different from the colours of the $`d^{-}\hspace{0pt}{(u)}`$ arcs entering $`u`$, the $`d^{+}\hspace{0pt}{(v)}`$ arcs leaving $`v`$, and the $`{d^{-}\hspace{0pt}{(v)}} - 1`$ other arcs entering $`v`$, so at most $`{3\hspace{0pt}\Delta^{-}} - 1`$ arcs in total. Hence, we may remove the arc $`u\hspace{0pt}v`$, apply induction to obtain a colouring of $`D \smallsetminus {u\hspace{0pt}v}`$. Extending this colouring to $`u\hspace{0pt}v`$, we obtain a directed star colouring of $`D`$ with at most $`3\hspace{0pt}\Delta^{-}`$ colours.
Note that to prove Conjecture [2](#Thmtheorem2 "Conjecture 2 ‣ 1 Introduction ‣ WDM and Directed Star Arboricity"), it will be enough to consider the two cases $`\Delta^{-} = 2`$ and $`\Delta^{-} = 3`$. To see this, let $`D`$ be a digraph with maximum indegree $`\Delta^{-} \geq 2`$ and $`k = {\lfloor{\Delta^{-}/2}\rfloor}`$. For every vertex $`v`$, let $`({N_{1}^{-}\hspace{0pt}{(v)}},{N_{2}^{-}\hspace{0pt}{(v)}},\ldots,{N_{k}^{-}\hspace{0pt}{(v)}})`$ be a partition of $`N^{-}\hspace{0pt}{(v)}`$ such that $`{|{N_{i}^{-}\hspace{0pt}{(v)}}|} \leq 2`$ for all $`1 \leq i \leq {k - 1}`$ and $`{|{N_{k}^{-}\hspace{0pt}{(v)}}|} \leq 2`$ if $`\Delta^{-}`$ is even and $`{|{N_{k}^{-}\hspace{0pt}{(v)}}|} \leq 3`$ if $`\Delta^{-}`$ is odd. Then the digraph $`D_{i}`$ with vertex set $`V\hspace{0pt}{(D)}`$ and such that $`{N_{D_{i}}^{-}\hspace{0pt}{(v)}} = {N_{i}^{-}\hspace{0pt}{(v)}}`$ for every vertex $`v \in {V\hspace{0pt}{(D)}}`$, has maximum indegree at most two except if $`i = k`$ and $`\Delta^{-}`$ is odd, in which case $`D_{k}`$ has maximum indegree at most three. If Conjecture [2](#Thmtheorem2 "Conjecture 2 ‣ 1 Introduction ‣ WDM and Directed Star Arboricity") holds for every $`D_{i}`$ then it would also hold for $`D`$.
We next consider the directed star arboricity of a digraph with bounded maximum degree. The degree of a vertex $`v`$ is $`{d\hspace{0pt}{(v)}} = {{d^{-}\hspace{0pt}{(v)}} + {d^{+}\hspace{0pt}{(v)}}}`$. This corresponds to the degree of the vertex in the underlying multigraph. (We have edges with multiplicity two in the underlying multigraph each time there is a circuit of length two in the digraph.) The maximum degree of a digraph $`D`$, denoted by $`\Delta\hspace{0pt}{(D)}`$, or simply $`\Delta`$ when $`D`$ is clearly understood from the context, is $`\max{\{{d\hspace{0pt}{(v)}},{v \in {V\hspace{0pt}{(D)}}}\}}`$. Let us denote by $`\mu\hspace{0pt}{(G)}`$, the maximum multiplicity of an edge in a multigraph. By Vizing’s theorem \[[11](#bib.bib11)\], one can colour the edges of a multigraph with $`{\Delta\hspace{0pt}{(G)}} + {\mu\hspace{0pt}{(G)}}`$ colours so that two edges have different colours if they are incident. Since the multigraph underlying a digraph has maximum multiplicity at most two, for any digraph $`D`$, $`{d\hspace{0pt}s\hspace{0pt}t\hspace{0pt}{(D)}} \leq {\Delta + 2}`$. We conjecture the following:
######
Conjecture 4
Let $`D`$ be a digraph with maximum degree $`\Delta \geq 3`$. Then $`{d\hspace{0pt}s\hspace{0pt}t\hspace{0pt}{(D)}} \leq \Delta`$.
This conjecture would be tight since every digraph with $`\Delta = \Delta^{-}`$ has directed star arboricity at least $`\Delta`$. In Section [3](#S3 "3 Directed Star Arboricity of Subcubic Digraphs ‣ WDM and Directed Star Arboricity"), we prove that Conjecture [4](#Thmtheorem4 "Conjecture 4 ‣ 1 Introduction ‣ WDM and Directed Star Arboricity") holds when $`\Delta = 3`$.
######
Theorem 5
Every subcubic digraph has directed star arboricity at most three.
A first step towards Conjectures [2](#Thmtheorem2 "Conjecture 2 ‣ 1 Introduction ‣ WDM and Directed Star Arboricity") and [4](#Thmtheorem4 "Conjecture 4 ‣ 1 Introduction ‣ WDM and Directed Star Arboricity") would be to prove the following weaker statement.
######
Conjecture 6
Let $`k \geq 2`$ and $`D`$ be a digraph. If $`{\max{(\Delta^{-},\Delta^{+})}} \leq k`$ then $`{d\hspace{0pt}s\hspace{0pt}t\hspace{0pt}{(D)}} \leq {2\hspace{0pt}k}`$.
This conjecture holds and is far from being tight for large values of $`k`$. Indeed Guiduli \[[6](#bib.bib6)\] showed that if $`{\max{(\Delta^{-},\Delta^{+})}} \leq k`$, then $`{d\hspace{0pt}s\hspace{0pt}t\hspace{0pt}{(D)}} \leq {k + {20\hspace{0pt}{\log k}} + 84}`$. Guiduli’s proof is based on the fact that, when both out- and indegrees are bounded, the colour of an arc depends on the colour of few other arcs. This bounded dependency allows the use of the Lovász Local Lemma. This idea was first used by Algor and Alon \[[1](#bib.bib1)\] for the star arboricity of undirected graphs. We also note that Guiduli’s result is (almost) tight since there are digraphs $`D`$ with $`{\max{(\Delta^{-},\Delta^{+})}} \leq k`$ and $`{d\hspace{0pt}s\hspace{0pt}t\hspace{0pt}{(D)}} \geq {k + {\Omega\hspace{0pt}{({\log k})}}}`$ (see \[[6](#bib.bib6)\]).
As for Conjecture [2](#Thmtheorem2 "Conjecture 2 ‣ 1 Introduction ‣ WDM and Directed Star Arboricity"), it is quite straightforward to check that it is sufficient to prove Conjecture [6](#Thmtheorem6 "Conjecture 6 ‣ 1 Introduction ‣ WDM and Directed Star Arboricity") for $`k = 2`$ and $`k = 3`$. In Section [4](#S4 "4 Directed Star Arboricity of Digraphs with Maximum In- and Outdegree Two ‣ WDM and Directed Star Arboricity"), we prove that Conjecture [6](#Thmtheorem6 "Conjecture 6 ‣ 1 Introduction ‣ WDM and Directed Star Arboricity") holds for $`k = 2`$. By the above remark, this implies that Conjecture [6](#Thmtheorem6 "Conjecture 6 ‣ 1 Introduction ‣ WDM and Directed Star Arboricity") holds for all even values of $`k`$.
######
Theorem 7
Let $`D`$ be a digraph. If $`\Delta^{-} \leq 2`$ and $`\Delta^{+} \leq 2`$, then $`{d\hspace{0pt}s\hspace{0pt}t\hspace{0pt}{(D)}} \leq 4`$. In particular, *Conjecture [6](#Thmtheorem6 "Conjecture 6 ‣ 1 Introduction ‣ WDM and Directed Star Arboricity")* holds for all even values of $`k`$.
Next, we study the more general and more realistic problem in which every vertex of $`V`$ is connected to the hub by $`n`$ optical fibres. Moreover each node may send several multicasts. We note $`{M_{1}\hspace{0pt}{(v)}},\ldots,{M_{s\hspace{0pt}{(v)}}\hspace{0pt}{(v)}}`$ the $`s\hspace{0pt}{(v)}`$ multicasts that node $`v`$ sends. For $`1 \leq i \leq {s\hspace{0pt}{(v)}}`$, the set of nodes to which the multicast $`M_{i}\hspace{0pt}{(v)}`$ is sent is denoted by $`S_{i}\hspace{0pt}{(v)}`$. The problem is still to find the minimum number of wavelengths used considering that all fibres are identical.
We model this as a problem on labelled digraphs: We construct a multidigraph $`D`$ on vertex set $`V`$. For each multicast $`{M_{i}\hspace{0pt}{(v)}} = {(v,{S_{i}\hspace{0pt}{(v)}})}`$, $`v \in V`$, $`1 \leq i \leq {s\hspace{0pt}{(v)}}`$, we add the set of arcs $`{A_{i}\hspace{0pt}{(v)}} = {\{{{{v\hspace{0pt}w},w} \in {S_{i}\hspace{0pt}{(v)}}}\}}`$ with label $`i`$. The label of an arc $`\overset{\rightarrow}{a}`$ is denoted by $`l\hspace{0pt}{(\overset{\rightarrow}{a})}`$. Thus for every ordered pair $`(u,v)`$ of vertices and label $`i`$ there is at most one arc $`u\hspace{0pt}v`$ labelled by $`i`$. If each vertex sends at most $`m`$ multicasts, there are at most $`m`$ labels on the arcs. Such a digraph is said to be $`m`$-labelled. One wishes to find an $`n`$-fibre wavelength assignment of $`D`$, that is a mapping $`\Phi:{{A\hspace{0pt}{(D)}}\rightarrow{\Lambda \times {\{ 1,\ldots,n\}} \times {\{ 1,{\ldots\hspace{0pt}n}\}}}}`$ in which every arc $`u\hspace{0pt}v`$ is associated a triple $`({\lambda\hspace{0pt}{({u\hspace{0pt}v})}},{f^{+}\hspace{0pt}{({u\hspace{0pt}v})}},{f^{-}\hspace{0pt}{({u\hspace{0pt}v})}})`$ such that :
- $`(i)`$
For each pair of arcs $`u\hspace{0pt}v`$ and $`v\hspace{0pt}w`$, $`{({\lambda\hspace{0pt}{({u\hspace{0pt}v})}},{f^{-}\hspace{0pt}{({u\hspace{0pt}v})}})} \neq {({\lambda\hspace{0pt}{({v\hspace{0pt}w})}},{f^{+}\hspace{0pt}{({v\hspace{0pt}w})}})}`$;
- $`({i\hspace{0pt}i})`$
For each pair of arcs $`u\hspace{0pt}v`$ and $`u'\hspace{0pt}v`$, $`{({\lambda\hspace{0pt}{({u\hspace{0pt}v})}},{f^{-}\hspace{0pt}{({u\hspace{0pt}v})}})} \neq {({\lambda\hspace{0pt}{({u'\hspace{0pt}v})}},{f^{-}\hspace{0pt}{({u'\hspace{0pt}w})}})}`$;
- $`({i\hspace{0pt}i\hspace{0pt}i})`$
For each pair of arcs $`v\hspace{0pt}w`$ and $`v\hspace{0pt}w'`$, if $`{l\hspace{0pt}{({v\hspace{0pt}w})}} \neq {l\hspace{0pt}{({v\hspace{0pt}w'})}}`$, then $`{({\lambda\hspace{0pt}{({v\hspace{0pt}w})}},{f^{+}\hspace{0pt}{({v\hspace{0pt}w})}})} \neq {({\lambda\hspace{0pt}{({v\hspace{0pt}w'})}},{f^{+}\hspace{0pt}{({v\hspace{0pt}w'})}})}`$.
Here $`\Lambda`$ is the set of available wavelengths, $`\lambda\hspace{0pt}{({u\hspace{0pt}v})}`$ corresponds to the wavelength of $`u\hspace{0pt}v`$, and $`f^{+}\hspace{0pt}{({u\hspace{0pt}v})}`$ and $`f^{-}\hspace{0pt}{({u\hspace{0pt}v})}`$ are the fibres used in $`u`$ and $`v`$, respectively. We can describe the above equations as follows:
- •
Condition $`(i)`$ corresponds to the requirement that an arc entering $`v`$ and an arc leaving $`v`$ should have either different wavelengths or different fibres;
- •
Condition $`({i\hspace{0pt}i})`$ corresponds to the requirement that two arcs entering $`v`$ should have either different wavelengths or different fibres; and finally
- •
Condition $`({i\hspace{0pt}i\hspace{0pt}i})`$ corresponds to the requirement that two arcs leaving $`v`$ with different labels have either different wavelengths or different fibres.
The problem is to find the minimum cardinality $`\lambda_{n}\hspace{0pt}{(D)}`$ of $`\Lambda`$ such that there exists an $`n`$-fibre wavelength assignment of $`D`$.
The crucial part of an $`n`$-fibre wavelength assignment is the function $`\lambda`$ which assigns colours (wavelengths) to the arcs. It must be an $`n`$-fibre colouring, that is a function $`\phi:{{A\hspace{0pt}{(D)}}\rightarrow\Lambda}`$, such that at each vertex $`v`$, for each colour $`\omega \in \Lambda`$, $`{{i\hspace{0pt}n\hspace{0pt}{(v,\omega)}} + {o\hspace{0pt}u\hspace{0pt}t\hspace{0pt}{(v,\omega)}}} \leq n`$ where $`i\hspace{0pt}n\hspace{0pt}{(v,\omega)}`$ denotes the number of arcs coloured by $`\omega`$ entering $`v`$ and $`o\hspace{0pt}u\hspace{0pt}t\hspace{0pt}{(v,\omega)}`$ denotes the number of labels $`l`$ such that there exists an arc leaving $`v`$ coloured by $`\omega`$. Once we have an $`n`$-fibre colouring, one can easily find a suitable wavelength assignment. For every vertex $`v`$ and every colour $`\omega`$, this is done by assigning a different fibre to each arc of colour $`\omega`$ entering $`v`$, and to each set of arcs of colour $`\omega`$ of the same label that leave $`v`$. We conclude that $`\lambda_{n}\hspace{0pt}{(D)}`$ is the minimum number of colours such that there exists an $`n`$-fibre colouring.
We are particularly interested in $`{\lambda_{n}\hspace{0pt}{(m,k)}} = {\max{\{{\left. {\lambda_{n}\hspace{0pt}{(D)}} \middle| {D\hspace{0pt}{{\text{~is~}m\text{-labelled}}\text{~and~}}\hspace{0pt}\Delta^{-}\hspace{0pt}{(D)}} \right. \leq k}\}}}`$, that is the maximum number of wavelengths that may be necessary if there are $`n`$ fibres, and each node sends at most $`m`$ multicasts and receives at most $`k`$ multicasts. In particular, $`{\lambda_{1}\hspace{0pt}{(1,k)}} = {\max{\{{\left. {d\hspace{0pt}s\hspace{0pt}t\hspace{0pt}{(D)}} \middle| {\Delta^{-}\hspace{0pt}{(D)}} \right. \leq k}\}}}`$. (So our above mentioned results show that $`{2\hspace{0pt}k} \leq {\lambda_{1}\hspace{0pt}{(1,k)}} \leq {{2\hspace{0pt}k} + 1}`$.) Brandt and Gonzalez showed that for $`n \geq 2`$ we have $`{\lambda_{n}\hspace{0pt}{(1,k)}} \leq \left\lceil \frac{k}{n - 1} \right\rceil`$. In Section [5](#S5 "5 Multiple Fibres ‣ WDM and Directed Star Arboricity"), we study the case when $`n \geq 2`$ and $`m \geq 2`$. We show in Proposition [17](#Thmtheorem17 "Proposition 17 ‣ 5 Multiple Fibres ‣ WDM and Directed Star Arboricity") and Theorem [24](#Thmtheorem24 "Theorem 24 ‣ 5 Multiple Fibres ‣ WDM and Directed Star Arboricity") that
<table>
<tbody>
<tr>
<td></td>
<td>if m≥n then⌈mn⌈kn⌉+kn⌉≤λn(m,k)≤⌈mn⌈kn⌉+kn⌉+Cm2logkn for some constant C.if m≥n then𝑚𝑛𝑘𝑛𝑘𝑛subscript𝜆𝑛𝑚𝑘𝑚𝑛𝑘𝑛𝑘𝑛𝐶superscript𝑚2𝑘𝑛 for some constant C.\mbox{if $m\geq n$ then}\ \ \ \left\lceil\frac{m}{n}\left\lceil\frac{k}{n}\right\rceil+\frac{k}{n}\right\rceil\>\leq\>\lambda_{n}(m,k)\>\leq\>\left\lceil\frac{m}{n}\left\lceil\frac{k}{n}\right\rceil+\frac{k}{n}\right\rceil+C\frac{m^{2}\log k}{n}\mbox{\ \ \ for some constant $C$.}</td>
<td></td>
</tr>
</tbody>
</table>
We conjecture that the lower bound is the right value of $`\lambda_{n}\hspace{0pt}{(m,k)}`$ when $`m \geq n`$. We also show in Proposition [17](#Thmtheorem17 "Proposition 17 ‣ 5 Multiple Fibres ‣ WDM and Directed Star Arboricity") and Proposition [25](#Thmtheorem25 "Proposition 25 ‣ 5 Multiple Fibres ‣ WDM and Directed Star Arboricity") that
<table>
<tbody>
<tr>
<td></td>
<td>if m<n, then ⌈mn⌈kn⌉+kn⌉≤λn(m,k)≤⌈kn−m⌉.if m<n, then 𝑚𝑛𝑘𝑛𝑘𝑛subscript𝜆𝑛𝑚𝑘𝑘𝑛𝑚\mbox{if $m<n$, then \hskip 14.22636pt}\left\lceil\frac{m}{n}\left\lceil\frac{k}{n}\right\rceil+\frac{k}{n}\right\rceil\leq\lambda_{n}(m,k)\leq\left\lceil\frac{k}{n-m}\right\rceil.</td>
<td></td>
</tr>
</tbody>
</table>
The lower bound generalises Brandt and Gonzalez \[[4](#bib.bib4)\] results which established this inequality in the particular cases when $`k \leq 2`$, $`m \leq 2`$ and $`k = m`$. The digraphs used to show this lower bound are all acyclic. We show that if $`m \geq n`$ then this lower bound is tight for acyclic digraphs. Moreover the above mentioned digraphs have large outdegree. Generalising the result of Guiduli \[[6](#bib.bib6)\], we show that for an $`m`$-labelled digraph $`D`$ with both in- and outdegree bounded by $`k`$ only few colours are needed when $`m \geq n`$:
<table>
<tbody>
<tr>
<td></td>
<td>λn(D)≤kn+C′m2logkn for some constant C′.subscript𝜆𝑛𝐷𝑘𝑛superscript𝐶′superscript𝑚2𝑘𝑛 for some constant C′.\lambda_{n}(D)\leq\frac{k}{n}+C^{\prime}\frac{m^{2}\log k}{n}\mbox{\ \ \ \ \ for some constant $C^{\prime}$.}</td>
<td></td>
</tr>
</tbody>
</table>
Finally, in Section [6](#S6 "6 Concluding Remarks ‣ WDM and Directed Star Arboricity"), we consider the complexity of finding the directed star arboricity of a digraph, and prove that, unsurprisingly, this is an $`\mathcal{N}\hspace{0pt}\mathcal{P}`$-hard problem. More precisely, we show that determining the directed star arboricity of a digraph with in- and outdegree at most two is $`\mathcal{N}\hspace{0pt}\mathcal{P}`$-complete. We then give a very short proof of a theorem of Pinlou and Sopena \[[9](#bib.bib9)\], showing that acircuitic directed star arboricity of subcubic graphs is at most four (see Section [6](#S6 "6 Concluding Remarks ‣ WDM and Directed Star Arboricity") for the definitions).
## 2 Directed Star Arboricity of Digraphs with Bounded Indegrees
In this section, we give the proof of Theorem [1](#Thmtheorem1 "Theorem 1 ‣ 1 Introduction ‣ WDM and Directed Star Arboricity") and settle Conjecture [2](#Thmtheorem2 "Conjecture 2 ‣ 1 Introduction ‣ WDM and Directed Star Arboricity") for acyclic digraphs.
An arborescence is a connected digraph in which every vertex has indegree one except one, called root, which has indegree zero. A forest is the disjoint union of arborescences. A star is an arborescence in which the root dominates all the other vertices. A galaxy is a forest of stars. Clearly, every colour class of a directed star colouring is a galaxy. Hence, the directed star arboricity of a digraph $`D`$ is the minimum number of galaxies into which $`A\hspace{0pt}{(D)}`$ may be partitioned.
It is easy to see that a forest has directed star arboricity at most two. Hence, an idea to prove Conjecture [2](#Thmtheorem2 "Conjecture 2 ‣ 1 Introduction ‣ WDM and Directed Star Arboricity") would be to show that every digraph has an arc-partition into $`\Delta^{-}`$ forests. However this statement is false. Indeed a theorem of Frank \[[5](#bib.bib5)\] (see also Chapter 53 of \[[10](#bib.bib10)\]) characterises all digraphs which have an arc-partition into $`k`$ forests. Let $`D = {(V,A)}`$. For any $`U \subset V`$, the digraph induced by the vertices of $`U`$ is denoted $`D\hspace{0pt}{\lbrack U\rbrack}`$.
######
Theorem 8
(A. Frank)
A digraph $`D = {(V,A)}`$ has an arc-partition into $`k`$ forests if and only if $`{\Delta^{-}\hspace{0pt}{(D)}} \leq k`$ and for every $`U \subset V`$, the digraph $`D\hspace{0pt}{\lbrack U\rbrack}`$ has at most $`k\hspace{0pt}{({{|U|} - 1})}`$ arcs.
This theorem implies that every digraph $`D`$ has an arc-partition into $`\Delta^{-} + 1`$ forests. Indeed for any $`U \subset V`$, $`{\Delta^{-}\hspace{0pt}{({D\hspace{0pt}{\lbrack U\rbrack}})}} \leq {\min{\{\Delta^{-},{{|U|} - 1}\}}}`$, so $`D\hspace{0pt}{\lbrack U\rbrack}`$ has at most $`{{\min{\{\Delta^{-},{{|U|} - 1}\}}} \times {|U|}} \leq {{({\Delta^{-} + 1})}\hspace{0pt}{({{|U|} - 1})}}`$ arcs. Hence, every digraph has directed star arboricity at most $`{2\hspace{0pt}\Delta^{-}} + 2`$.
######
Corollary 9
Every digraph $`D`$ satisfies $`{d\hspace{0pt}s\hspace{0pt}t\hspace{0pt}{(D)}} \leq {{2\hspace{0pt}\Delta^{-}} + 2}`$.
Theorem [1](#Thmtheorem1 "Theorem 1 ‣ 1 Introduction ‣ WDM and Directed Star Arboricity") states that $`{d\hspace{0pt}s\hspace{0pt}t\hspace{0pt}{(D)}} \leq {{2\hspace{0pt}\Delta^{-}} + 1}`$. The idea to prove this theorem is to show that every digraph has an arc-partition into $`\Delta^{-}`$ forests and a galaxy $`G`$. To do so, we prove a stronger result, Lemma [10](#Thmtheorem10 "Lemma 10 ‣ 2 Directed Star Arboricity of Digraphs with Bounded Indegrees ‣ WDM and Directed Star Arboricity") below.
We need some extra definitions. A sink is a vertex with outdegree $`0`$. A source is a vertex with indegree $`0`$. A multidigraph $`D`$ will be called $`k`$-nice if $`\Delta^{-} \leq k`$, and if the tails of parallel arcs, if any, are sources. A $`k`$-decomposition of $`D`$ is an arc-partition into $`k`$ forests and a galaxy $`G`$ such that every source of $`D`$ is isolated in $`G`$. Let $`u`$ be a vertex of $`D`$. A $`k`$-decomposition of $`D`$ is $`u`$-suitable if no arc of $`G`$ has head $`u`$.
######
Lemma 10
Let $`u`$ be a vertex of a $`k`$-nice multidigraph $`D`$. Then $`D`$ has a $`u`$-suitable $`k`$-decomposition.
Proof. We proceed by induction on $`n + k`$ by considering (strong) connectivity of $`D`$:
- •
If $`D`$ is not connected as graph, we apply induction on every component.
- •
If $`D`$ is strongly connected, every vertex has indegree at least one. (Recall that there are no parallel arcs.) Let $`v`$ be an outneighbour of $`u`$. There exists a spanning arborescence $`T`$ with root $`v`$ which contains all the arcs with tail $`v`$. Let $`D'`$ be the digraph obtained from $`D`$ by removing the arcs of $`T`$ and $`v`$. Observe that $`D'`$ is $`({k - 1})`$-nice. By induction, it has a $`u`$-suitable $`({k - 1})`$-decomposition $`(F_{1},\ldots,F_{k - 1},G)`$. Note that each $`F_{i}`$, for $`1 \leq i \leq {k - 1}`$, $`T`$ and $`G`$ contain all the arcs of $`D`$ except those with head $`v`$. By construction, $`G' = {G \cup {u\hspace{0pt}v}}`$ is a galaxy since no arc of $`G`$ has head $`u`$. Let $`u_{1},\ldots,u_{l - 1}`$ be the inneighbours of $`v`$ distinct from $`u`$, where $`l \leq k`$. Let $`F_{i}' = {F_{i} \cup {u_{i}\hspace{0pt}v}}`$, for all $`1 \leq i \leq {l - 1}`$. Each $`F_{i}'`$ is a forest, so $`(F_{1},\ldots,F_{k - 1},T,G')`$ is a $`u`$-suitable $`k`$-decomposition of $`D`$.
- •
In the only remaining case, $`D`$ is connected but not strongly connected. We consider a terminal strongly connected component $`D_{1}`$ of $`D`$. Set $`D_{2} = {D \smallsetminus D_{1}}`$. Let $`u_{1}`$ and $`u_{2}`$ be two vertices of $`D_{1}`$ and $`D_{2}`$, respectively, such that $`u`$ is one of them.
If $`D_{2}`$ has a unique vertex $`v`$ (thus $`u_{2} = v`$), since $`D`$ is connected and $`D_{1}`$ is strong, there exists a spanning arborescence $`T`$ of $`D`$ with root $`v`$. Now $`D' = {D \smallsetminus {A\hspace{0pt}{(T)}}}`$ is a $`({k - 1})`$-nice multidigraph, so by induction it has a $`u`$-suitable $`({k - 1})`$-decomposition. Adding $`T`$ to this decomposition, we obtain a $`u`$-suitable $`k`$-decomposition. If $`D_{2}`$ has more than one vertex, it admits a $`u_{2}`$-suitable $`k`$-decomposition $`(F_{1}^{2},\ldots,F_{k}^{2},G^{2})`$, by induction. Moreover the digraph $`D_{1}'`$ obtained by contracting $`D_{2}`$ to a single vertex $`v`$ is $`k`$-nice and so has a $`u_{1}`$-suitable $`k`$-decomposition $`(F_{1}^{1},\ldots,F_{k}^{1},G^{1})`$. Moreover, since $`v`$ is a source, it is isolated in $`G^{1}`$. Hence $`G = {G^{1} \cup G^{2}}`$ is a galaxy. We now let $`F_{i}`$ be the union of $`F_{i}^{1}`$ and $`F_{i}^{2}`$ by replacing the arcs of $`F_{i}^{1}`$ with tail $`v`$ by the corresponding arcs in $`D`$. Then $`(F_{1},\ldots,F_{k},G)`$ is a $`k`$-decomposition of $`D`$ which is suitable for both $`u_{1}`$ and $`u_{2}`$.
$`\square`$
Theorem [1](#Thmtheorem1 "Theorem 1 ‣ 1 Introduction ‣ WDM and Directed Star Arboricity") is an immediate consequence of Lemma [10](#Thmtheorem10 "Lemma 10 ‣ 2 Directed Star Arboricity of Digraphs with Bounded Indegrees ‣ WDM and Directed Star Arboricity").
### 2.1 Acyclic Digraphs
It is not very hard to show that $`{d\hspace{0pt}s\hspace{0pt}t\hspace{0pt}{(D)}} \leq {2\hspace{0pt}\Delta^{-}}`$ when $`D`$ is acyclic, but we will prove this result in a more constrained way. For $`n \leq p`$, a cyclic $`n`$-interval of $`\{ 1,2,\ldots,p\}`$ is a set of $`n`$ consecutive numbers modulo $`p`$. Now for the directed star colouring, we will insist that for every vertex $`v`$, the (distinct) colours used to colour the arcs with head $`v`$ are chosen in a cyclic $`k`$-interval of $`\{ 1,2,\ldots,{2\hspace{0pt}k}\}`$. Thus, the number of possible sets of colours used to colour the entering arcs of a vertex $`v`$ drastically falls from $`\binom{2\hspace{0pt}k}{d^{-}\hspace{0pt}{(v)}}`$ when every set is a priori possible, to at most $`{2\hspace{0pt}k} \times \binom{k}{d^{-}\hspace{0pt}{(v)}}`$. Note that having consecutive colours on the arcs entering a vertex corresponds to having consecutive wavelengths on the link between the corresponding node and the central one. This may of importance for issues related to grooming in optical networks. For details about grooming, we refer the reader to the two comprehensive surveys \[[7](#bib.bib7), [8](#bib.bib8)\].
######
Theorem 11
Let $`D`$ be an acyclic digraph with maximum indegree $`k`$. Then $`D`$ admits a directed star $`2\hspace{0pt}k`$-colouring such that for every vertex, the colours assigned to its entering arcs are included in a cyclic $`k`$-interval of $`\{ 1,2,\ldots,{2\hspace{0pt}k}\}`$.
To prove this theorem, we first state and prove the following result on sets of distinct representatives.
######
Lemma 12
Let $`I_{1},\ldots,I_{k}`$ be $`k`$ non necessarily distinct cyclic $`k`$-intervals of $`\{ 1,2,\ldots,{2\hspace{0pt}k}\}`$. Then $`I_{1},\ldots,I_{k}`$ admit a set of distinct representatives forming a cyclic $`k`$-interval.
Proof. We consider $`I_{1},\ldots,I_{k}`$ as a set of $`p`$ distinct cyclic $`k`$-intervals $`I_{1},\ldots,I_{p}`$ with respective multiplicity $`m_{1},\ldots,m_{p}`$ such that $`{\sum_{i = 1}^{p}m_{i}} = k`$. Such a system will be denoted by $`({(I_{1},m_{1})},\ldots,{(I_{p},m_{p})})`$. We shall prove the existence of a cyclic $`k`$-interval $`J`$, such that $`J`$ can be partitioned into $`p`$ subsets $`J_{i}`$, $`1 \leq i \leq p`$, such that $`{|J_{i}|} = m_{i}`$ and $`J_{i} \subset I_{i}`$. This proves the lemma (by associating distinct elements of $`J_{i}`$ to each copy of $`I_{i}`$).
We proceed by induction on $`p`$. The result holds trivially for $`p = 1`$. We have to deal with two cases:
- •
There exist $`i`$ and $`j`$ such that $`{|{I_{j} \smallsetminus I_{i}}|} = {|{I_{i} \smallsetminus I_{j}}|} \leq {\max{(m_{i},m_{j})}}`$.
Suppose without loss of generality that $`i < j`$ and $`m_{i} \geq m_{j}`$. We apply the induction hypothesis to $`({(I_{1},m_{1})},\cdots,{(I_{i},{m_{i} + m_{j}})},\cdots,{(I_{j - 1},m_{j - 1})},{(I_{j + 1},m_{j + 1})},\cdots,{(I_{p},m_{p})})`$, in order to find a cyclic interval $`J'`$, such that $`J'`$ admits a partition into subsets $`J_{r}'`$, such that for any $`r`$ different from $`i`$ and $`j`$, the set $`J_{r}' \subset I_{r}`$ is a subset of size $`m_{r}`$, and $`J_{i}' \subset I_{i}`$ is of size $`m_{i} + m_{j}`$. We now partition $`J_{i}'`$ into two sets $`J_{i}`$ and $`J_{j}`$ with respective size $`m_{i}`$ and $`m_{j}`$, in such a way that $`{{({I_{i} \smallsetminus I_{j}})} \cap J_{i}'} \subseteq J_{i}`$. Remark that this is possible precisely because of our assumption $`{|{I_{j} \smallsetminus I_{i}}|} = {|{I_{i} \smallsetminus I_{j}}|} \leq m_{i}`$. Since $`J_{i} \subset I_{i}`$ and $`J_{j} \subset I_{j}`$, this refined partition of $`J'`$ is the desired one.
- •
For any $`i,j`$ we have $`{|{I_{j} \smallsetminus I_{i}}|} = {|{I_{i} \smallsetminus I_{j}}|} \geq {{\max{(m_{i},m_{j})}} + 1}`$.
Each $`I_{i}`$ intersects exactly $`{2\hspace{0pt}m_{i}} - 1`$ other cyclic $`k`$-intervals on less than $`m_{i}`$ elements. Since there are $`2\hspace{0pt}k`$ cyclic $`k`$-intervals in total and $`{\sum_{i = 1}^{p}{({{2\hspace{0pt}m_{i}} - 1})}} = {{2\hspace{0pt}k} - p} < {2\hspace{0pt}k}`$, we conclude the existence of a cyclic $`k`$-interval $`J`$ which intersects each $`I_{i}`$ in an interval of size at least $`m_{i}`$.
Let us prove that one can partition $`J`$ in the desired way. By Hall’s matching theorem, it suffices to prove that for every subset $`\mathcal{I}`$ of $`\{ 1,\ldots,p\}`$, we have $`{|{{\bigcup_{i \in \mathcal{I}}I_{i}} \cap J}|} \geq {\sum_{i \in \mathcal{I}}m_{i}}`$.
Suppose for the sake of a contradiction that a subset $`\mathcal{I}`$ of $`\{ 1,\ldots,p\}`$ violates this inequality. Such a subset will be called contracting. Without loss of generality, we assume that $`\mathcal{I}`$ is a contracting set with minimum cardinality and that $`\mathcal{I} = {\{ 1,\ldots,q\}}`$. Observe that by the choice of $`J`$, we have $`q \geq 2`$. The set $`K:={{\bigcup_{i \in \mathcal{I}}I_{i}} \cap J}`$ consists of one or two intervals of $`J`$, each containing one extremity of $`J`$. By the minimality of $`\mathcal{I}`$, $`K`$ must be a single interval (if not, one would take $`\mathcal{I}_{1}`$ (resp. $`\mathcal{I}_{2}`$), all the elements of $`\mathcal{I}`$ which contains the first (resp. the second) extremity of $`J`$. Then one of $`\mathcal{I}_{1}`$ or $`\mathcal{I}_{2}`$ would be contracting). Thus, one of the two extremities of $`J`$ is in every $`I_{i}`$, $`i \in \mathcal{I}`$. Without loss of generality, we may assume that $`{({I_{1} \cap J})} \subset {({I_{2} \cap J})} \subset \cdots \subset {({I_{q} \cap J})}`$. Now, for every $`2 \leq i \leq q`$, $`{|{I_{i} \smallsetminus I_{i - 1}}|} = {|{{({I_{i} \cap J})} \smallsetminus {({I_{i - 1} \cap J})}}|} \geq {{\max{(m_{i},m_{i - 1})}} + 1} \geq {m_{i} + 1}`$. But $`{|{{\bigcup_{i \in \mathcal{I}}I_{i}} \cap J}|} = {{|{({I_{1} \cap J})}|} + {\sum_{i = 2}^{q}{|{{({I_{i} \cap J})} \smallsetminus {({I_{i - 1} \cap J})}}|}}}`$. So $`{|{{\bigcup_{i \in \mathcal{I}}I_{i}} \cap J}|} \geq {{{\sum_{i = 1}^{q}m_{i}} + q} - 1}`$, which is a contradiction.
$`\square`$
Proof of Theorem [11](#Thmtheorem11 "Theorem 11 ‣ 2.1 Acyclic Digraphs ‣ 2 Directed Star Arboricity of Digraphs with Bounded Indegrees ‣ WDM and Directed Star Arboricity"). By induction on the number of vertices, the result being trivial if $`D`$ has one vertex. Suppose now that $`D`$ has at least two vertices. Then $`D`$ has a sink $`x`$. By the induction hypothesis, $`D \smallsetminus x`$ has a directed star $`2\hspace{0pt}k`$-colouring $`c`$ such that for every vertex, the colours assigned to its entering arcs are included in a cyclic $`k`$-interval. Let $`v_{1},v_{2},\ldots,v_{l}`$ be the inneighbours of $`x`$ in $`D`$, where $`l \leq k`$ because $`{\Delta^{-}\hspace{0pt}{(D)}} \leq k`$. For each $`1 \leq i \leq l`$, let $`I_{i}'`$ be a cyclic $`k`$-interval which contains all the colours of the arcs with head $`v_{i}`$. We set $`I_{i} = {{\{ 1,\ldots,{2\hspace{0pt}k}\}} \smallsetminus I_{i}'}`$. Clearly, $`I_{i}`$ is a cyclic $`k`$-interval and the arc $`v_{i}\hspace{0pt}x`$ can be coloured by any element of $`I_{i}`$. By Lemma [12](#Thmtheorem12 "Lemma 12 ‣ 2.1 Acyclic Digraphs ‣ 2 Directed Star Arboricity of Digraphs with Bounded Indegrees ‣ WDM and Directed Star Arboricity"), $`I_{1},\ldots,I_{l}`$ have a set of distinct representatives included in a cyclic $`2\hspace{0pt}k`$-interval $`J`$. Hence assigning $`J`$ to $`x`$, and colouring the arc $`v_{i}\hspace{0pt}x`$ by the representative of $`I_{i}`$ gives a directed star $`2\hspace{0pt}k`$-colouring of $`D`$. $`\square`$
Theorem [11](#Thmtheorem11 "Theorem 11 ‣ 2.1 Acyclic Digraphs ‣ 2 Directed Star Arboricity of Digraphs with Bounded Indegrees ‣ WDM and Directed Star Arboricity") is tight : Brandt \[[3](#bib.bib3)\] showed that for every $`k`$, there is an acyclic digraph such that $`{\Delta^{-}\hspace{0pt}{(D_{k})}} = k`$ and $`{d\hspace{0pt}s\hspace{0pt}t\hspace{0pt}{(D_{k})}} = {2\hspace{0pt}k}`$. His construction is the special case for $`n = m = 1`$ of the construction given in Proposition [17](#Thmtheorem17 "Proposition 17 ‣ 5 Multiple Fibres ‣ WDM and Directed Star Arboricity").
## 3 Directed Star Arboricity of Subcubic Digraphs
Recall that a subcubic digraph is a graph with degree at most three. In this section, we give the proof of Theorem [5](#Thmtheorem5 "Theorem 5 ‣ 1 Introduction ‣ WDM and Directed Star Arboricity") which states that the directed star arboricity of a subcubic digraph is at most three.
To do so, we need to establish some preliminary lemmas which will enable us to extend a partial directed star colouring into a directed star colouring of the whole digraph. To state these lemmas, we need the following definition. Let $`D = {(V,A)}`$ be a digraph and $`S`$ be a subset of $`V \cup A`$. Suppose that each element $`x`$ of $`S`$ is assigned a list $`L\hspace{0pt}{(x)}`$. A colouring $`c`$ of $`S`$ is an $`L`$-colouring if $`{c\hspace{0pt}{(x)}} \in {L\hspace{0pt}{(x)}}`$ for every $`x \in S`$.
######
Lemma 13
Let $`C`$ be a circuit in which every vertex $`v`$ receives a list $`L\hspace{0pt}{(v)}`$ of two colours among $`\{ 1,2,3\}`$ and each arc $`\overset{\rightarrow}{a}`$ receives the list $`{L\hspace{0pt}{(\overset{\rightarrow}{a})}} = {\{ 1,2,3\}}`$. The following two statements are equivalent:
- •
There is no $`L`$-colouring $`c`$ of the arcs and vertices such that $`{c\hspace{0pt}{(x)}} \neq {c\hspace{0pt}{({x\hspace{0pt}y})}}`$, $`{c\hspace{0pt}{(y)}} \neq {c\hspace{0pt}{({x\hspace{0pt}y})}}`$, and $`{c\hspace{0pt}{({x\hspace{0pt}y})}} \neq {c\hspace{0pt}{({y\hspace{0pt}z})}}`$, for all arcs $`x\hspace{0pt}y`$ and $`y\hspace{0pt}z`$.
- •
$`C`$ is an odd circuit and all the vertices have the same list.
Proof. Assume first that every vertex is assigned the same list, say $`\{ 1,2\}`$. If $`C`$ is odd, it is a simple matter to check that we can not find the desired colouring. Indeed, among two consecutive arcs, one has to be coloured $`3`$. If $`C`$ is even, we colour the vertices by $`1`$ and the arcs alternately by $`2`$ and $`3`$.
Now assume that $`C = {x_{1}\hspace{0pt}x_{2}\hspace{0pt}\ldots\hspace{0pt}x_{k}\hspace{0pt}x_{1}}`$ and $`x_{1}`$ and $`x_{2}`$ are assigned different lists. Say $`{L\hspace{0pt}{(x_{1})}} = {\{ 1,2\}}`$ and $`{L\hspace{0pt}{(x_{2})}} = {\{ 2,3\}}`$. We colour the arc $`x_{1}\hspace{0pt}x_{2}`$ by $`3`$, the vertex $`x_{2}`$ by $`2`$ and the arc $`x_{2}\hspace{0pt}x_{3}`$ by $`1`$. Then we colour $`x_{3}`$, $`x_{3}\hspace{0pt}x_{4}`$, …, $`x_{k}`$ greedily. It remains to colour $`x_{k}\hspace{0pt}x_{1}`$ and $`x_{1}`$. Two cases may happen: If we can colour $`x_{k}\hspace{0pt}x_{1}`$ by 1 or 2, we do it and colour $`x_{1}`$ by $`2`$ or $`1`$ respectively. Otherwise the set of colours assigned to $`x_{k}`$ and $`x_{k - 1}\hspace{0pt}x_{k}`$ is $`\{ 1,2\}`$. Hence, we colour $`x_{k}\hspace{0pt}x_{1}`$ with $`3`$, $`x_{1}`$ by $`1`$, and recolour $`x_{1}\hspace{0pt}x_{2}`$ by $`2`$ and $`x_{2}`$ by $`3`$. $`\square`$
######
Lemma 14
Let $`D`$ be a subcubic digraph with no vertex of outdegree two and indegree one. Suppose that every arc $`\overset{\rightarrow}{a}`$ has a list of colours $`{L\hspace{0pt}{(\overset{\rightarrow}{a})}} \subset {\{ 1,2,3\}}`$ such that:
- •
If the head of $`\overset{\rightarrow}{a}`$ is a sink $`s`$ (in which case, $`\overset{\rightarrow}{a}`$ will be called a final arc), $`{|{L\hspace{0pt}{(\overset{\rightarrow}{a})}}|} \geq {d^{-}\hspace{0pt}{(s)}}`$.
- •
If $`\overset{\rightarrow}{a}`$ is not a final arc and the tail of $`\overset{\rightarrow}{a}`$ is a source (in which case, $`\overset{\rightarrow}{a}`$ will be called an initial arc), $`{|{L\hspace{0pt}{(\overset{\rightarrow}{a})}}|} \geq 2`$.
- •
In all the other cases, $`{|{L\hspace{0pt}{(\overset{\rightarrow}{a})}}|} = 3`$.
In addition, assume that the followings hold:
- •
If a vertex is the head of at least two initial arcs $`\overset{\rightarrow}{a}`$ and $`\overset{\rightarrow}{b}`$, the union of the lists of colours $`L\hspace{0pt}{(\overset{\rightarrow}{a})}`$ and $`L\hspace{0pt}{(\overset{\rightarrow}{b})}`$ contains all the three colours.
- •
If all the vertices of an odd circuit are the tails of initial arcs, the union of the lists of colours of these initial arcs contains all the three colours.
Then $`D`$ has a directed star $`L`$-colouring.
Proof. We colour the graph inductively. Consider a terminal strong component $`C`$ of $`D`$. Since $`D`$ has no vertex with indegree one and outdegree two, $`C`$ induces either a singleton or a circuit.
- 1\)
Assume that $`C`$ is a singleton $`v`$ which is the head of a unique arc $`\overset{\rightarrow}{a} = {u\hspace{0pt}v}`$. If $`u`$ has indegree zero, we colour $`\overset{\rightarrow}{a}`$ with a colour of its list. If $`u`$ has indegree one, and thus total degree two, we colour $`\overset{\rightarrow}{a}`$ by the colour of its list and remove this colour from the list of the arc with head $`u`$. If $`u`$ is the head of $`\overset{\rightarrow}{e}`$ and $`\overset{\rightarrow}{f}`$, observe that $`L\hspace{0pt}{(\overset{\rightarrow}{e})}`$ and $`L\hspace{0pt}{(\overset{\rightarrow}{f})}`$ have at least two colours and their union have all the three colours. To conclude, we colour $`\overset{\rightarrow}{a}`$ with a colour in its list, remove this colour from $`L\hspace{0pt}{(\overset{\rightarrow}{e})}`$ and $`L\hspace{0pt}{(\overset{\rightarrow}{f})}`$, remove $`\overset{\rightarrow}{a}`$, and split $`u`$ into two vertices, one with head $`\overset{\rightarrow}{e}`$ and the other with head $`\overset{\rightarrow}{f}`$. Now, we choose different colours for the arcs $`\overset{\rightarrow}{e}`$ and $`\overset{\rightarrow}{f}`$ in their respective lists to form the new list $`L\hspace{0pt}{(\overset{\rightarrow}{e})}`$ and $`L\hspace{0pt}{(\overset{\rightarrow}{f})}`$.
- 2\)
Assume that $`C`$ is a singleton $`v`$ which is the head of several arcs, including $`\overset{\rightarrow}{a} = {u\hspace{0pt}v}`$. In this case, we reduce $`L\hspace{0pt}{(\overset{\rightarrow}{a})}`$ to a single colour, remove this colour from the other arcs with head $`v`$ and split $`v`$ into $`v_{1}`$, which becomes the head of $`\overset{\rightarrow}{a}`$, and $`v_{2}`$ which becomes the head of the other arcs.
- 3\)
Assume that $`C`$ is a circuit. Every arc entering $`C`$ has a list of at least two colours. We can apply Lemma [13](#Thmtheorem13 "Lemma 13 ‣ 3 Directed Star Arboricity of Subcubic Digraphs ‣ WDM and Directed Star Arboricity") to conclude.
$`\square`$
Proof of Theorem [5](#Thmtheorem5 "Theorem 5 ‣ 1 Introduction ‣ WDM and Directed Star Arboricity"). Assume for the sake of a contradiction that the digraph $`D`$ has directed star arboricity more than three and is minimum for this property with respect to the number of arcs. Observe that $`D`$ has no source, otherwise we simply delete it with all its incident arcs, apply induction and extend the colouring. This is possible since arcs leaving from a source can be coloured arbitrarily. Let $`D_{1}`$ be the subdigraph of $`D`$ induced by the vertices of indegree at most $`1`$. We denote by $`D_{2}`$ the digraph induced by the other vertices, and by $`\lbrack D_{i},D_{j}\rbrack`$ the set of arcs with tail in $`D_{i}`$ and head in $`D_{j}`$. We claim that $`D_{1}`$ contains no even circuit. If not, we simply remove the arcs of this even circuit, apply induction. We can extend the colouring to the arcs of the even circuit since every arc of the circuit has two colours available.
A critical set of vertices of $`D_{2}`$ is either a vertex of $`D_{2}`$ with indegree at least two in $`D_{1}`$, or an odd circuit of $`D_{2}`$ having all its inneighbours in $`D_{1}`$. Observe that critical sets are disjoint. For every critical set $`S`$, we select two arcs entering $`S`$ from $`D_{1}`$, called selected arcs of $`S`$.
Let $`D'`$ be digraph induced by the arc set $`A' = {{A\hspace{0pt}{(D_{1})}} \cup {\lbrack D_{2},D_{1}\rbrack}}`$. We now define a conflict graph on the arcs of $`D'`$ in the following way:
- •
Two arcs $`x\hspace{0pt}y`$, $`y\hspace{0pt}v`$ of $`D'`$ are in conflict, called normal conflict at $`y`$.
- •
Two arcs $`x\hspace{0pt}y`$, $`u\hspace{0pt}v`$ of $`D'`$ are also in conflict if there exists two selected arcs of the same set $`S`$ with tails $`y`$ and $`v`$. These conflicts are called selected conflicts at $`y`$ and $`v`$.
Let us analyse the structure of the conflict graph. Observe first that an arc is in conflict with three arcs : one normal conflict at its tail and at most two (normal or selected) at its head. We claim that there is no $`K_{4}`$ in the conflict graph. For the sake of a contradiction, suppose there is one. This means that there are four arcs $`\overset{\rightarrow}{a},\overset{\rightarrow}{b},\overset{\rightarrow}{c}`$ and $`\overset{\rightarrow}{d}`$ pairwise in conflict. Since each of these arcs have degree three in $`K_{4}`$, each of these arcs should have a normal conflict at its tail, and so the digraph induced by these four arcs contains a circuit. This circuit cannot be of even length (two or four) so it has to be of length three. It follows that the four arcs $`\overset{\rightarrow}{a},\overset{\rightarrow}{b},\overset{\rightarrow}{c}`$ and $`\overset{\rightarrow}{d}`$ are as in Figure [1](#S3.F1 "Figure 1 ‣ 3 Directed Star Arboricity of Subcubic Digraphs ‣ WDM and Directed Star Arboricity") below (modulo a permutation of the labels). Let $`D^{\ast}`$ be the digraph obtained from $`D`$ by removing the arcs $`\overset{\rightarrow}{a},\overset{\rightarrow}{b},\overset{\rightarrow}{c},\overset{\rightarrow}{d}`$ and their four incident vertices. By minimality of $`D`$, $`D^{\ast}`$ admits a directed star $`3`$-colouring which can be extended to $`D`$ as depicted below depending if the two leaving arcs are coloured the same or differently. This proves the claim.
[Refer to caption](/html/0705.0315/assets/x1.png)
Figure 1: A $`K_{4}`$ in the conflict graph and the two ways of extending the colouring.
Brooks Theorem asserts that every subcubic graph without $`K_{4}`$ is 3-colourable. So the conflict graph admits a $`3`$-colouring $`c`$. This gives a colouring of the arcs of $`D'`$. Let $`D^{\operatorname{\prime\prime}}`$ be the digraph obtained from $`D`$, and let $`L`$ be the list-assignment on the arcs of $`D^{\operatorname{\prime\prime}}`$ defined simultaneously as follow:
- •
Remove the arcs of $`D_{1}`$ from $`D`$,
- •
Assign to each arc of $`\lbrack D_{2},D_{1}\rbrack`$ the singleton list containing the colour it has in $`D'`$,
- •
For each arc $`u\hspace{0pt}v`$ of $`\lbrack D_{1},D_{2}\rbrack`$, there is a unique arc $`t\hspace{0pt}u`$ in $`A\hspace{0pt}{(D')}`$. Assign to $`u\hspace{0pt}v`$ the list $`{L\hspace{0pt}{({u\hspace{0pt}v})}} = {{\{ 1,2,3\}} \smallsetminus {c\hspace{0pt}{({t\hspace{0pt}u})}}}`$.
- •
Assign the list $`\{ 1,2,3\}`$ to the other arcs.
- •
If there are vertices with indegree one and outdegree two (they were in $`D_{1}`$), split each of them into one source of degree two and a sink of degree one.
Note that there is a trivial one-to-one correspondence between $`A\hspace{0pt}{(D^{\operatorname{\prime\prime}})}`$ and $`{A\hspace{0pt}{(D)}} \smallsetminus {A\hspace{0pt}{(D')}}`$. By the definition of the conflict graph and $`D^{\operatorname{\prime\prime}}`$, one can easily check that $`D^{\operatorname{\prime\prime}}`$ and $`L`$ satisfies the condition of Lemma [14](#Thmtheorem14 "Lemma 14 ‣ 3 Directed Star Arboricity of Subcubic Digraphs ‣ WDM and Directed Star Arboricity"). Hence $`D^{\operatorname{\prime\prime}}`$ admits a directed star $`L`$-colouring which union with $`c`$ is a directed star $`3`$-colouring of $`D`$, a contradiction. The proof of Theorem [5](#Thmtheorem5 "Theorem 5 ‣ 1 Introduction ‣ WDM and Directed Star Arboricity") is now complete. $`\square`$
## 4 Directed Star Arboricity of Digraphs with Maximum In- and Outdegree Two
The goal of this section is to prove Theorem [7](#Thmtheorem7 "Theorem 7 ‣ 1 Introduction ‣ WDM and Directed Star Arboricity"): Every digraph with outdegree and indegree at most two has directed star arboricity at most four. However, the class of digraphs with in- and outdegree at most two is certainly not an easy class with respect to directed star arboricity, as we will show in Section [6.1](#S6.SS1 "6.1 Complexity ‣ 6 Concluding Remarks ‣ WDM and Directed Star Arboricity").
In order to prove Theorem [7](#Thmtheorem7 "Theorem 7 ‣ 1 Introduction ‣ WDM and Directed Star Arboricity"), it suffices to show that $`D`$ contains a galaxy $`G`$ which spans all the vertices of degree four. Indeed, if this is true, then $`D' = {D - {A\hspace{0pt}{(G)}}}`$ has maximum degree at most $`3`$ and by Theorem [5](#Thmtheorem5 "Theorem 5 ‣ 1 Introduction ‣ WDM and Directed Star Arboricity"), $`{d\hspace{0pt}s\hspace{0pt}t\hspace{0pt}{(D')}} \leq 3`$. So $`{d\hspace{0pt}s\hspace{0pt}t\hspace{0pt}{(D)}} \leq 4`$. Hence Theorem [7](#Thmtheorem7 "Theorem 7 ‣ 1 Introduction ‣ WDM and Directed Star Arboricity") is directly implied by the following lemma:
######
Lemma 15
Let $`D`$ be a digraph with maximum indegree and outdegree two. Then $`D`$ contains a galaxy which spans the set of vertices with degree four.
To prove this lemma, we need some preliminaries.
Let $`V`$ be a set. An ordered digraph on $`V`$ is a pair $`( \leq ,D)`$ where:
- •
$`\leq`$ is a partial order on $`V`$;
- •
$`D`$ is a digraph with vertex set $`V`$;
- •
$`D`$ contains the Hasse diagram of $`\leq`$. I.e., when $`x \leq y \leq z`$ implies $`x = y`$ or $`y = z`$, then $`x\hspace{0pt}z`$ is an arc of $`D`$;
- •
If $`x\hspace{0pt}y`$ is an arc of $`D`$, the vertices $`x,y`$ are $`\leq`$-comparable.
The arcs $`x\hspace{0pt}y`$ of $`D`$ thus belong to two different types: the forward arcs, when $`x \leq y`$, and the backward arcs, when $`y \leq x`$.
######
Lemma 16
Let $`( \leq ,D)`$ be an ordered digraph on $`V`$. Assume that every vertex is the tail of at most one backward arc and at most two forward arcs, and that the indegree of every vertex of $`D`$ is at least two, except possibly one vertex $`x`$ with indegree one. Then $`D`$ contains two arcs $`\gamma\hspace{0pt}\alpha`$ and $`\beta\hspace{0pt}\lambda`$ such that $`\alpha \leq \beta \leq \gamma`$, $`\beta \leq \lambda`$ and $`\gamma \nleq \lambda`$, all four vertices being distinct except possibly $`\alpha = \beta`$.
Proof. For the sake of a contradiction, let us consider a counterexample with minimum $`|V|`$.
An interval is a subset $`I`$ of $`V`$ which has a minimum $`m`$ and a maximum $`M`$ such that $`I = {\{ z:{m \leq z \leq M}\}}`$. An interval $`I`$ is good if every arc with tail in $`I`$ and head outside $`I`$ has tail $`M`$ and every backward arc in $`I`$ has tail $`M`$.
Let $`I`$ be an interval of $`D`$. The digraph $`D/I`$ obtained from $`D`$ by contracting $`I`$ is the digraph with vertex set $`{({V \smallsetminus I})} \cup {\{ v_{I}\}}`$ such that $`x\hspace{0pt}y`$ is an arc if and only either $`v_{I} \notin {\{ x,y\}}`$ and $`{x\hspace{0pt}y} \in {A\hspace{0pt}{(D)}}`$, or $`x = v_{I}`$ and there exists $`x_{I} \in I`$ such that $`{x_{I}\hspace{0pt}y} \in {A\hspace{0pt}{(D)}}`$, or $`y = v_{I}`$ and there exists $`y_{I} \in I`$ such that $`{x\hspace{0pt}y_{I}} \in {A\hspace{0pt}{(D)}}`$.
Similarly, the binary relation $`\leq_{/I}`$ obtained from $`\leq`$ by contracting $`I`$ is the binary relation on $`{({V \smallsetminus I})} \cup {\{ v_{I}\}}`$ such that $`x \leq_{/I}y`$ if and only if either $`v_{I} \notin {\{ x,y\}}`$ and $`x \leq y`$, or $`x = v_{I}`$ and there exists $`x_{I} \in I`$ such that $`x_{I} \leq y`$, or $`y = v_{I}`$ and there exists $`y_{I} \in I`$ such that $`x \leq y_{I}`$. We claim that if $`I`$ is good then $`\leq_{/I}`$ is a partial order. Indeed suppose it is not, then there are two elements $`u`$ and $`t`$ such that $`u \leq_{/I}v_{I}`$, $`v_{I} \leq t`$ and $`u \nleq_{/I}t`$. Then $`M \nleq t`$. Let $`\alpha \in I`$ be the maximal element of $`I`$ such that $`\alpha \leq t`$, $`\lambda`$ be a successor of $`\alpha`$ in $`I`$, and $`\gamma`$ a successor of $`\alpha`$ not in $`I`$ (it exists as $`t \notin I`$ and $`\alpha \leq t`$ and maximal in $`I`$ with this property). Then $`\lambda`$ and $`\gamma`$ are incomparable, and $`\alpha\hspace{0pt}\gamma`$ and $`\alpha\hspace{0pt}\lambda`$ are in the Hasse diagram of $`\leq`$. Because $`I`$ is good, it follows that $`\gamma\hspace{0pt}\alpha`$ and $`\alpha\hspace{0pt}\lambda`$ are arcs of $`D`$, which is impossible as $`D`$ is supposed to be a counterexample.
Hence, if $`I`$ is a good interval, then $`( \leq_{/I},{D/I})`$ is an ordered digraph. Note that if $`x \leq_{/I}v_{I}`$, then $`x \leq M`$ with $`M`$ the maximum of $`I`$. The crucial point is that if $`I`$ a good interval of $`D`$ for which the conclusion of Lemma [16](#Thmtheorem16 "Lemma 16 ‣ 4 Directed Star Arboricity of Digraphs with Maximum In- and Outdegree Two ‣ WDM and Directed Star Arboricity") holds for $`( \leq_{/I},{D/I})`$, then it holds for $`( \leq ,D)`$. Indeed, suppose there exists two arcs $`\gamma\hspace{0pt}\alpha`$ and $`\beta\hspace{0pt}\lambda`$ of $`D/I`$ such that $`\alpha \leq_{/I}\beta \leq_{/I}\gamma`$, $`\beta \leq_{/I}\lambda`$, and $`\gamma \nleq_{/I}\lambda`$. Note that since $`I`$ is good, we have $`v_{I} \neq \gamma`$. Let $`M`$ be the maximum of $`I`$.
If $`v_{I} \notin {\{\alpha,\beta,\gamma,\lambda\}}`$, then $`\gamma\hspace{0pt}\alpha`$ and $`\beta\hspace{0pt}\lambda`$ gives the conclusion for $`D`$.
If $`v_{I} = \alpha`$, then $`\gamma\hspace{0pt}M`$ is an arc. Let us show that $`M \leq \beta`$. Indeed, let $`x`$ be a maximal vertex in $`I`$ such that $`x \leq \beta`$ and let $`y`$ be a minimal vertex such that $`x \leq y \leq \beta`$. Since the Hasse diagram of $`\leq`$ is included in $`D`$, $`x\hspace{0pt}y`$ is an arc, and so $`x = M`$ (since $`I`$ is good). Thus $`\gamma\hspace{0pt}M`$ and $`\beta\hspace{0pt}\lambda`$ are the desired arcs.
If $`v_{I} = \beta`$, then $`M\hspace{0pt}\lambda`$ is an arc and $`\alpha \leq M`$, so $`\gamma\hspace{0pt}\alpha`$ and $`M\hspace{0pt}\lambda`$ are the desired arcs.
If $`v_{I} = \lambda`$, then there exists $`\lambda_{I} \in I`$ such that $`\beta\hspace{0pt}\lambda_{I}`$, so $`\gamma\hspace{0pt}\alpha`$ and $`\beta\hspace{0pt}\lambda_{I}`$ are the desired arcs.
Hence to get a contradiction, it is sufficient to find a good interval $`I`$ such that $`( \leq_{/I},{D/I})`$ satisfies the hypotheses of Lemma [16](#Thmtheorem16 "Lemma 16 ‣ 4 Directed Star Arboricity of Digraphs with Maximum In- and Outdegree Two ‣ WDM and Directed Star Arboricity").
Observe that there are at least two backward arcs. Indeed, if there are two minimal elements for $`\leq`$, there are at least three backward arcs entering these vertices (since one of them can be $`x`$). And if there is a unique minimum $`m`$, by letting $`m'`$ minimal in $`V \smallsetminus m`$, at least two arcs are entering $`m,m'`$.
Let $`M`$ be a vertex which is the tail of a backward arc and which is minimal for $`\leq`$ for this property. Since two arcs cannot have the same tail, $`M`$ is not the maximum of $`\leq`$ (if any). Let $`M\hspace{0pt}m`$ be the backward arc with tail $`M`$.
We claim that the interval $`J`$ with minimum $`m`$ and maximum $`M`$ is good. Indeed, by the definition of $`M`$, no backward arc has its tail in $`J \smallsetminus {\{ M\}}`$. Moreover, any forward arc $`\beta\hspace{0pt}\lambda`$ with its tail in $`J \smallsetminus {\{ M\}}`$ and its head outside $`J`$ would give our conclusion (with $`\alpha = m`$ and $`\gamma = M`$), a contradiction.
Now consider a good interval $`I`$ with maximum $`M`$ which is maximal with respect to inclusion. We claim that if $`x \in I`$, then there is at least one arc entering $`I`$, and if $`x \notin I`$, there are at least two arcs entering $`I`$ with different tails.
Call $`m_{1}`$ the minimum of $`I`$ and $`m_{2}`$ any minimal element of $`I \smallsetminus m_{1}`$. First assume that $`x`$ is in $`I`$. There are at least three arcs with heads $`m_{1}`$ or $`m_{2}`$. One of them is $`m_{1}\hspace{0pt}m_{2}`$, one of them can be with tail $`M`$, but there is still one left with tail not in $`I`$. Now assume that $`x`$ is not in $`I`$. There are at least two arcs with heads $`m_{1}`$ or $`m_{2}`$ and tails not in $`I`$. If the tails are different, we are done. If the tails are the same, say $`v`$, observe that $`v\hspace{0pt}m_{1}`$ and $`v\hspace{0pt}m_{2}`$ are both backward or both forward (otherwise $`v`$ would be in $`I`$). Since both cannot be backward, both $`v\hspace{0pt}m_{1}`$ and $`v\hspace{0pt}m_{2}`$ are forward. Hence the interval with minimum $`v`$ and maximum $`M`$ is a good interval, contradicting the maximality of $`I`$. This proves the claim.
This in turn implies that $`( \leq_{/I},{D/I})`$ satisfies the hypotheses of Lemma [16](#Thmtheorem16 "Lemma 16 ‣ 4 Directed Star Arboricity of Digraphs with Maximum In- and Outdegree Two ‣ WDM and Directed Star Arboricity"), yielding a contradiction. $`\square`$
Proof of Lemma [15](#Thmtheorem15 "Lemma 15 ‣ 4 Directed Star Arboricity of Digraphs with Maximum In- and Outdegree Two ‣ WDM and Directed Star Arboricity"). Let $`G`$ be a galaxy of $`D`$ which spans a maximum number of vertices of degree four. Suppose for the sake of a contradiction that some vertex $`x`$ with degree four is not spanned.
An alternating path is an oriented path ending at $`x`$, starting by an arc of $`G`$, and alternating with arcs of $`G`$ and arcs of $`{A\hspace{0pt}{(D)}} \smallsetminus {A\hspace{0pt}{(G)}}`$. We denote by $`\mathcal{A}`$ the set of arcs of $`G`$ which belong to an alternating path.
######
Claim 1
Every arc of $`\mathcal{A}`$ is a component of $`G`$.
Proof. Indeed, if $`u\hspace{0pt}v`$ belongs to $`\mathcal{A}`$, it starts some alternating path $`P`$. Thus, if $`u`$ has outdegree more than one in $`G`$, the digraph with the set of arcs $`A\hspace{0pt}{(G)}\hspace{0pt}\bigtriangleup\hspace{0pt}A\hspace{0pt}{(P)}`$ is a galaxy and spans $`{V\hspace{0pt}{(G)}} \cup x`$. $`\square`$
######
Claim 2
There is no circuits alternating arcs of $`\mathcal{A}`$ and arcs of $`{A\hspace{0pt}{(D)}} \smallsetminus \mathcal{A}`$.
Proof. Assume that there is such a circuit $`C`$. Consider a shortest alternating path $`P`$ starting with some arc of $`\mathcal{A}`$ in $`C`$. Now the digraph with arcs $`A\hspace{0pt}{(G)}\hspace{0pt}\bigtriangleup\hspace{0pt}{({{A\hspace{0pt}{(P)}} \cup {A\hspace{0pt}{(C)}}})}`$ is a galaxy which spans $`{V\hspace{0pt}{(G)}} \cup x`$, contradicting the maximality of $`G`$. $`\square`$
We now endow $`\mathcal{A} \cup x`$ with a partial order structure by letting $`a \leq b`$ if there exists an alternating path starting at $`a`$ and ending at $`b`$. The fact that this relation is a partial order relies on Claim [2](#Thmclaim2 "Claim 2 ‣ 4 Directed Star Arboricity of Digraphs with Maximum In- and Outdegree Two ‣ WDM and Directed Star Arboricity"). Observe that $`x`$ is the maximum of this order.
We also construct a digraph $`\mathcal{D}`$ on vertex set $`\mathcal{A} \cup x`$ with all arcs $`{u\hspace{0pt}v}\rightarrow{s\hspace{0pt}t}`$ such that $`u\hspace{0pt}s`$ or $`v\hspace{0pt}s`$ is an arc of $`D`$ (and $`{u\hspace{0pt}v}\rightarrow x`$ such that $`u\hspace{0pt}x`$ or $`v\hspace{0pt}x`$ is an arc of $`D`$).
######
Claim 3
The pair $`(\mathcal{D}, \leq )`$ is an ordered digraph. Moreover an arc of $`\mathcal{A}`$ is the tail of at most one backward arc and two forward arcs, and $`x`$ is the tail of at most two backward arcs.
Proof. The fact that the Hasse diagram of $`\leq`$ is contained in $`\mathcal{D}`$ follows from the fact that if $`{u\hspace{0pt}v} \leq {s\hspace{0pt}t}`$ belongs to the Hasse diagram of $`\leq`$, there is an alternating path starting by $`u\hspace{0pt}v\hspace{0pt}s\hspace{0pt}t`$, in particular, the arc $`v\hspace{0pt}s`$ belongs to $`D`$, and thus $`{u\hspace{0pt}v}\rightarrow{s\hspace{0pt}t}`$ in $`\mathcal{D}`$.
Suppose that $`{u\hspace{0pt}v}\rightarrow{s\hspace{0pt}t}`$, then $`v\hspace{0pt}s`$ or $`u\hspace{0pt}s`$ is an arc of $`D`$. If $`v\hspace{0pt}s`$ is an arc, because there is no alternating circuit, $`s\hspace{0pt}t`$ follows $`u\hspace{0pt}v`$ on some alternating path, and so $`{u\hspace{0pt}v} \leq {s\hspace{0pt}t}`$. In this case, $`{u\hspace{0pt}v}\rightarrow{s\hspace{0pt}t}`$ is forward.
If $`u\hspace{0pt}s`$ is an arc of $`D`$, we claim that $`{s\hspace{0pt}t} \leq {u\hspace{0pt}v}`$. Indeed, if an alternating path $`P`$ starting at $`s\hspace{0pt}t`$ does not contain $`u\hspace{0pt}v`$, the galaxy with arcs $`{({A\hspace{0pt}{(G)}\hspace{0pt}\bigtriangleup\hspace{0pt}A\hspace{0pt}{(P)}})} \cup {\{{u\hspace{0pt}s}\}}`$ spans $`{V\hspace{0pt}{(G)}} \cup x`$, contradicting the maximality of $`G`$. In this case, $`{u\hspace{0pt}v}\rightarrow{s\hspace{0pt}t}`$ is backward.
It follows that an arc $`u\hspace{0pt}v`$ of $`\mathcal{A}`$ is the tail of at most one backward arc (since this arc and $`u\hspace{0pt}v`$ are the two arcs leaving $`u`$ in $`D`$), and $`u\hspace{0pt}v`$ is the tail of at most two forward arcs (since $`v`$ has outdegree at most two). Furthermore, since $`x`$ has outdegree at most two, it follows that $`x`$ is the tail of at most two backward arcs. $`\square`$
######
Claim 4
The indegree of every vertex of $`\mathcal{D}`$ is two.
Proof. Let $`u\hspace{0pt}v`$ be a vertex of $`\mathcal{D}`$ which starts an alternating path $`P`$. If $`u`$ has indegree less than two, and thus does not belong to the set of vertices of degree four, the galaxy with arcs $`A\hspace{0pt}{(G)}\hspace{0pt}\bigtriangleup\hspace{0pt}A\hspace{0pt}{(P)}`$ spans more vertices of degree four than $`G`$, a contradiction. Let $`s`$ and $`t`$ be the two inneighbours of $`u`$ in $`D`$. An element of $`\mathcal{A} \cup x`$ should contain $`s`$, since otherwise, the galaxy with arcs $`{({A\hspace{0pt}{(G)}\hspace{0pt}\bigtriangleup\hspace{0pt}A\hspace{0pt}{(P)}})} \cup {\{{s\hspace{0pt}u}\}}`$ spans $`{V\hspace{0pt}{(G)}} \cup x`$ and contradicts the maximality of $`G`$. Similarly an element of $`\mathcal{A} \cup x`$ contains $`t`$.
Observe that the same element of $`\mathcal{A} \cup x`$ cannot contain both $`s`$ and $`t`$ (either the arc $`s\hspace{0pt}t`$ or the arc $`t\hspace{0pt}s`$), otherwise the arcs $`s\hspace{0pt}u`$ and $`t\hspace{0pt}u`$ would be both backward or forward, which is impossible. $`\square`$
At this stage, in order to apply Lemma [16](#Thmtheorem16 "Lemma 16 ‣ 4 Directed Star Arboricity of Digraphs with Maximum In- and Outdegree Two ‣ WDM and Directed Star Arboricity"), we just need to insure that the backward outdegree of every vertex is at most one. Since the only element of $`\mathcal{D}`$ which is the tail of two backward arcs is $`x`$, we simply delete any of these two backward arcs. The indegree of a vertex of $`\mathcal{D}`$ decreases by one but we are still fulfilling the hypothesis of Lemma [16](#Thmtheorem16 "Lemma 16 ‣ 4 Directed Star Arboricity of Digraphs with Maximum In- and Outdegree Two ‣ WDM and Directed Star Arboricity").
Hence according to this lemma, $`\mathcal{D}`$ contains two arcs $`\gamma\hspace{0pt}\alpha`$ and $`\beta\hspace{0pt}\lambda`$ such that $`\alpha \leq \beta \leq \gamma`$, $`\beta \leq \lambda`$ and $`\gamma \nleq \lambda`$. Recall that $`\alpha,\beta,\gamma,\lambda`$ are elements of $`\mathcal{A} \cup x`$. In particular, there is an alternating path $`P`$ containing $`\alpha,\beta,\lambda`$ (in this order) which does not contain $`\gamma`$. Setting $`\alpha = {\alpha_{1}\hspace{0pt}\alpha_{2}}`$ and $`\gamma = {\gamma_{1}\hspace{0pt}\gamma_{2}}`$, note that the backward arc $`\gamma\hspace{0pt}\alpha`$ corresponds to the arc $`\gamma_{1}\hspace{0pt}\alpha_{1}`$ in $`D`$. We reach a contradiction by considering the galaxy with arcs $`{({A\hspace{0pt}{(G)}\hspace{0pt}\bigtriangleup\hspace{0pt}A\hspace{0pt}{(P)}})} \cup {\{{\gamma_{1}\hspace{0pt}\alpha_{1}}\}}`$ which spans $`{V\hspace{0pt}{(D')}} \cup x`$. The proof of Lemma [15](#Thmtheorem15 "Lemma 15 ‣ 4 Directed Star Arboricity of Digraphs with Maximum In- and Outdegree Two ‣ WDM and Directed Star Arboricity") is now complete. $`\square`$
## 5 Multiple Fibres
In this section we consider the general problem with $`n \geq 2`$ fibres, and give lower and upper bounds on $`\lambda_{n}\hspace{0pt}{(m,k)}`$. Let us start by proving a lower bound on $`\lambda_{n}\hspace{0pt}{(m,k)}`$.
######
Proposition 17
For all $`{m,n,k} \in {\mathbb{N}}`$ , we have $`{\lambda_{n}\hspace{0pt}{(m,k)}} \geq \left\lceil {{\frac{m}{n}\hspace{0pt}\left\lceil \frac{k}{n} \right\rceil} + \frac{k}{n}} \right\rceil`$
Proof. Consider the following $`m`$-labelled digraph $`G_{n,m,k}`$ with vertex set $`X \sqcup Y \sqcup Z`$ such that :
- •
$`{|X|} = k`$, $`{|Y|} = {k\hspace{0pt}2^{{({m + 1})}\hspace{0pt}k}}`$ and $`{|Z|} = {m\hspace{0pt}\binom{|Y|}{k}}`$.
- •
For any $`x \in X`$ and $`y \in Y`$, there is an arc $`x\hspace{0pt}y`$ (of whatever label).
- •
For every set $`S`$ of $`k`$ vertices of $`Y`$ and any integer $`1 \leq i \leq m`$, there is a vertex $`z_{S}^{i}`$ in $`Z`$ which is dominated by all the vertices of $`S`$ via arcs labelled $`i`$.
Suppose there exists an $`n`$-fibre colouring of $`G_{n,m,k}`$ with $`c < \left\lceil {{\frac{m}{n}\hspace{0pt}\left\lceil \frac{k}{n} \right\rceil} + \frac{k}{n}} \right\rceil`$ colours. For $`y \in Y`$ and $`1 \leq i \leq m`$, let $`C_{i}\hspace{0pt}{(y)}`$ be the set of colours assigned to the arcs labelled $`i`$ leaving $`y`$. For $`0 \leq j \leq n`$, let $`P_{j}`$ be the set of colours used on $`j`$ arcs entering $`y`$ (and necessarily with two different fibres). Then $`{\sum_{j = 0}^{n}{j\hspace{0pt}{|P_{j}|}}} = k`$ as $`k`$ arcs enter $`y`$. Moreover $`{\sum_{j = 0}^{n}{|P_{j}|}} = c`$, since $`(P_{0},P_{1},\ldots,P_{n})`$ is a partition of the set of colours. Now each colour of $`P_{j}`$ may appear in at most $`n - j`$ of the $`C_{i}\hspace{0pt}{(y)}`$, so
<table>
<tbody>
<tr>
<td></td>
<td>∑i=1m|Ci(y)|≤∑j=0n(n−j)|Pj|=n∑j=0n|Pj|−∑j=0nj|Pj|=cn−k.superscriptsubscript𝑖1𝑚subscript𝐶𝑖𝑦superscriptsubscript𝑗0𝑛𝑛𝑗subscript𝑃𝑗𝑛superscriptsubscript𝑗0𝑛subscript𝑃𝑗superscriptsubscript𝑗0𝑛𝑗subscript𝑃𝑗𝑐𝑛𝑘\sum_{i=1}^{m}|C_{i}(y)|\leq\sum_{j=0}^{n}(n-j)|P_{j}|=n\sum_{j=0}^{n}|P_{j}|-\sum_{j=0}^{n}j|P_{j}|=cn-k.</td>
<td></td>
</tr>
</tbody>
</table>
Because $`{|Y|} > {{({k - 1})}\hspace{0pt}2^{c\hspace{0pt}m}}`$, there is a set $`S`$ of $`k`$ vertices $`y`$ of $`Y`$ having the same $`m`$-tuple $`{({C_{1}\hspace{0pt}{(y)}},\ldots,{C_{m}\hspace{0pt}{(y)}})} = {(C_{1},\ldots,C_{m})}`$. Without loss of generality, we may assume $`{|C_{1}|} = {\min{\{{\left. {|C_{i}|} \middle| 1 \right. \leq i \leq m}\}}}`$. Hence $`{|C_{1}|} \leq \frac{{c\hspace{0pt}n} - k}{m}`$. But the vertex $`z_{S}^{1}`$ has indegree $`k`$, so $`{|C_{1}|} \geq \frac{k}{n}`$. Since $`|C_{1}|`$ is an integer, we have $`\left\lfloor \frac{{c\hspace{0pt}n} - k}{m} \right\rfloor \geq {|C_{1}|} \geq \left\lceil \frac{k}{n} \right\rceil`$. So $`c \geq {{\frac{m}{n}\hspace{0pt}\left\lceil \frac{k}{n} \right\rceil} + \frac{k}{n}}`$ . Since $`c`$ is an integer, we get $`c \geq \left\lceil {{\frac{m}{n}\hspace{0pt}\left\lceil \frac{k}{n} \right\rceil} + \frac{k}{n}} \right\rceil`$, a contradiction. $`\square`$
Note that the graph $`G_{n,m,k}`$ is acyclic. The following lemma shows that, if $`m \geq n`$, one cannot expect better lower bounds by considering acyclic digraphs. Indeed $`G_{n,m,k}`$ is the $`m`$-labelled acyclic digraph with indegree at most $`k`$ for which an $`n`$-fibre colouring requires the more colours.
######
Lemma 18
Let $`D`$ be an acyclic $`m`$-labelled digraph with $`\Delta^{-} \leq k`$. If $`m \geq n`$, then $`{\lambda_{n}\hspace{0pt}{(D)}} \leq \left\lceil {{\frac{m}{n}\hspace{0pt}\left\lceil \frac{k}{n} \right\rceil} + \frac{k}{n}} \right\rceil`$.
Proof. Since $`D`$ is acyclic, its vertex set admits an ordering $`(v_{1},v_{2},\ldots,v_{p})`$ such that if $`v_{j}\hspace{0pt}v_{j'}`$ is an arc, then $`j < j'`$.
By induction on $`q`$, we shall find an $`n`$-fibre colouring of $`D\hspace{0pt}{\lbrack{\{ v_{1},\ldots,v_{q}\}}\rbrack}`$ together with sets $`C_{i}\hspace{0pt}{(v_{r})}`$ of $`\lceil\frac{k}{n}\rceil`$ (potential) colours, for $`1 \leq i \leq m`$ and $`1 \leq r \leq q`$, such that assigning a colour in $`C_{i}\hspace{0pt}{(v_{r})}`$ to an arc labelled $`i`$ leaving $`v_{r}`$ (in the future) will fulfil the condition of an $`n`$-fibre colouring at $`v_{r}`$.
Starting the process is easy. We may let $`C_{i}\hspace{0pt}{(v_{1})}`$’s to be any family of $`\lceil\frac{k}{n}\rceil`$-sets such that a colour appears in at most $`n`$ of them.
Suppose now that we have an $`n`$-fibre colouring of $`D\hspace{0pt}{\lbrack{\{ v_{1},\ldots,v_{q - 1}\}}\rbrack}`$, and that, for any $`1 \leq i \leq m`$ and $`1 \leq r \leq {q - 1}`$, the set $`C_{i}\hspace{0pt}{(v_{r})}`$ is already determined. Let us colour the arcs entering $`v_{q}`$. Each of these arcs $`v_{r}\hspace{0pt}v_{q}`$ may be assigned one of the $`\lceil\frac{k}{n}\rceil`$ colours of $`C_{l\hspace{0pt}{({v_{r}\hspace{0pt}v_{q}})}}\hspace{0pt}{(v_{r})}`$. Since a colour may be assigned to $`n`$ arcs (using different fibres) entering $`v_{q}`$, one can assign a colour and a fibre to each such arc. It remains to determine the sets $`C_{i}\hspace{0pt}{(v_{q})}`$, $`1 \leq i \leq m`$.
For $`0 \leq j \leq n`$, let $`P_{j}`$ be the set of colours assigned to $`j`$ arcs entering $`v_{q}`$. Let $`N = {\sum_{i = 0}^{n}{{({n - j})}\hspace{0pt}{|P_{j}|}}}`$ and $`(c_{1},c_{2},\ldots,c_{N})`$ be a sequence of colours such that each colour of $`P_{j}`$ appears exactly $`n - j`$ times and consecutively. For $`1 \leq i \leq m`$, set $`{C_{i}\hspace{0pt}{(v_{q})}} = \left. \{ c_{a} \middle| {a \equiv {i\operatorname{mod}m}}\} \right.`$. As $`n \leq m`$, a colour appears at most once in each $`C_{i}\hspace{0pt}{(v_{q})}`$. Moreover, $`N = {{n\hspace{0pt}\left\lceil {{\frac{m}{n}\hspace{0pt}\left\lceil \frac{k}{n} \right\rceil} + \frac{k}{n}} \right\rceil} - k} \geq {m\hspace{0pt}\left\lceil \frac{k}{n} \right\rceil}`$. So for $`1 \leq i \leq m`$, $`{|{C_{i}\hspace{0pt}{(v_{q})}}|} \geq \left\lceil \frac{k}{n} \right\rceil`$. $`\square`$
Lemma [18](#Thmtheorem18 "Lemma 18 ‣ 5 Multiple Fibres ‣ WDM and Directed Star Arboricity") shows that the lower bound of Proposition [17](#Thmtheorem17 "Proposition 17 ‣ 5 Multiple Fibres ‣ WDM and Directed Star Arboricity") is tight for acyclic digraphs. In fact, we conjecture that the lower bound remains tight for digraphs in general:
######
Conjecture 19
$`{\lambda_{n}\hspace{0pt}{(m,k)}} = \left\lceil {{\frac{m}{n}\hspace{0pt}\left\lceil \frac{k}{n} \right\rceil} + \frac{k}{n}} \right\rceil`$
We now establish an upper bound on $`\lambda_{n}\hspace{0pt}{(m,k)}`$ for general digraphs. Note that the graphs $`G_{n,m,k}`$ requires lots of colours but have very large outdegree. We first give an upper bound on $`\lambda_{n}\hspace{0pt}{(D)}`$ for $`m`$-labelled digraphs with bounded in- and outdegree. In this case, on can show that only ”few” colours are needed. This is derived from the following theorem of Guiduli.
######
Theorem 20
(Guiduli \[[6](#bib.bib6)\])
If $`{\Delta^{-},\Delta^{+}} \leq k`$, then $`{d\hspace{0pt}s\hspace{0pt}t\hspace{0pt}{(D)}} \leq {k + {20\hspace{0pt}{\log k}} + 84}`$. Moreover, $`D`$ admits a directed star colouring with $`k + {20\hspace{0pt}{\log k}} + 84`$ colours such that for each vertex $`v`$, there are at most $`{10\hspace{0pt}{\log k}} + 42`$ colours assigned to its leaving arcs.
As we will show below, Guiduli’s Theorem can be extended to the following statement for $`m`$-labelled digraphs.
######
Theorem 21
Let $`{f\hspace{0pt}{(n,m,k)}} = \left\lceil \frac{k + {{({{10\hspace{0pt}m^{2}} + 5})}\hspace{0pt}{\log k}} + {80\hspace{0pt}m^{2}} + m + 21}{n} \right\rceil`$ and let $`D`$ be an $`m`$-labelled digraph with $`{\Delta^{-},\Delta^{+}} \leq k`$. Then $`{\lambda_{n}\hspace{0pt}{(D)}} \leq {f\hspace{0pt}{(n,m,k)}}`$. Moreover, $`D`$ admits an $`n`$-fibre colouring with $`f\hspace{0pt}{(n,m,k)}`$ colours such that for each vertex $`v`$ and each label $`l`$, the number of colours assigned to the arcs labelled $`l`$ and leaving $`v`$ is at most $`{g\hspace{0pt}{(m,k)}} = \left\lceil {{{({{10\hspace{0pt}m} + 5})}\hspace{0pt}{\log k}} + {40\hspace{0pt}m} + 21} \right\rceil`$.
As one can notice, Theorem [21](#Thmtheorem21 "Theorem 21 ‣ 5 Multiple Fibres ‣ WDM and Directed Star Arboricity") in the case $`n = m = 1`$ is slightly better than Theorem [20](#Thmtheorem20 "Theorem 20 (Guiduli [6]) ‣ 5 Multiple Fibres ‣ WDM and Directed Star Arboricity") (for $`{\Delta^{-},\Delta^{+}} \leq k`$, Theorem [21](#Thmtheorem21 "Theorem 21 ‣ 5 Multiple Fibres ‣ WDM and Directed Star Arboricity") gives $`{d\hspace{0pt}s\hspace{0pt}t\hspace{0pt}{(D)}} \leq {k + {15\hspace{0pt}{\log k}} + 102}`$). But this is superficial and is only due to the upper bound given in Lemma [1](#Thmtheorem1 "Theorem 1 ‣ 1 Introduction ‣ WDM and Directed Star Arboricity"), which is better than the upper bound $`3\hspace{0pt}\Delta`$ used by Guiduli. Indeed, the methods are identical.
We recall the following definition: given a family of sets $`\mathcal{F} = {({{A_{i},i} \in I})}`$, a transversal of $`\mathcal{F}`$ is a family of distinct elements $`({{t_{i},i} \in I})`$ with $`t_{i} \in A_{i}`$ for all $`i \in I`$.
######
Lemma 22
Let $`D`$ be an $`m`$-labelled digraph with $`\Delta^{-} \leq k`$. Suppose that for each vertex $`v`$, there are $`m`$ disjoint lists $`L_{v}^{1},\ldots,L_{v}^{m}`$ of $`c`$ colours each being a subset of $`\{ 1,\ldots,{k + c}\}`$. If for each vertex $`v`$, the family $`\left\{ L_{y}^{i} \middle| {{y\hspace{0pt}x} \in {E\hspace{0pt}{(D)}\hspace{0pt}{\ {\text{and~}{y\hspace{0pt}x}\text{~is labelled~}i}}}} \right\}`$ has a transversal, then there is a $`1`$-fibre colouring of $`D`$ with $`k + {{({{2\hspace{0pt}m^{2}} + 1})}\hspace{0pt}c} + m`$ colours such that for each vertex $`v`$ and each label $`l`$, at most $`{{({{2\hspace{0pt}m} + 1})}\hspace{0pt}c} + 1`$ colours are assigned to arcs labelled $`l`$ that leave $`v`$.
Proof. Using the transversal to colour the entering arcs at each vertex, we obtain a colouring with few conflicts. Indeed there is no conflict between arcs entering a same vertex. So the only possible conflicts are between an arc entering a vertex $`v`$ and an arc leaving $`v`$. Since arcs leaving $`v`$ use at most $`m\hspace{0pt}c`$ colours (those of $`L_{v}^{1} \cup \ldots \cup L_{v}^{m}`$), there are at most $`m\hspace{0pt}c`$ arcs entering $`v`$ having the same colour as an arc leaving $`c`$. Removing such entering arcs for every vertex $`v`$, we obtain a digraph $`D'`$ for which the colouring with the $`k + c`$ colours is a $`1`$-fibre colouring. We now want to colour the arcs of $`D - D'`$ with few extra colours. Consider a label $`1 \leq l \leq m`$ and let $`D_{l}'`$ be the digraph induced by the arcs of $`D - D'`$ labelled $`l`$. Then $`D_{l}'`$ has indegree at most $`m\hspace{0pt}c`$. By Theorem [1](#Thmtheorem1 "Theorem 1 ‣ 1 Introduction ‣ WDM and Directed Star Arboricity"), we can partition $`D_{l}'`$ in $`{2\hspace{0pt}m}.{c + 1}`$ star forests. Thus $`D`$ can be $`1`$-fibre coloured with $`k + c + {m\hspace{0pt}{({{2\hspace{0pt}m\hspace{0pt}c} + 1})}}`$ colours. Moreover, in the above described colouring, arcs labelled $`l`$ which leave a vertex $`v`$ have a colour in $`L_{v}^{l}`$ or corresponding to one of the $`{2\hspace{0pt}m\hspace{0pt}c} + 1`$ star forests of $`D_{l}'`$. So at most $`{{({{2\hspace{0pt}m} + 1})}\hspace{0pt}c} + 1`$ colours are assigned to arcs labelled $`l`$ leaving $`v`$. $`\square`$
We will also need the following theorem.
######
Theorem 23
(Alon, McDiarmid and Reed \[[2](#bib.bib2)\])
Let $`k`$ and $`c`$ be positive integers with $`k \geq c \geq {{5\hspace{0pt}{\log k}} + 20}`$. Choose independent random subsets $`S_{1},\ldots,S_{k}`$ of $`X = {\{ 1,\ldots,{k + c}\}}`$ as follows. For each $`i`$, choose $`S_{i}`$ by performing $`c`$ independent uniform samplings from $`X`$. Then the probability that $`S_{1},\ldots,S_{k}`$ do not have a transversal is at most $`k^{3 - \frac{c}{2}}`$
Proof of Theorem [21](#Thmtheorem21 "Theorem 21 ‣ 5 Multiple Fibres ‣ WDM and Directed Star Arboricity"). It suffices to prove the result for $`n = 1`$. Indeed we can extend a $`1`$-fibre colouring satisfying the conditions of the theorem into an $`n`$-fibre colouring satisfying the conditions by replacing all the colours $`{q\hspace{0pt}n} + r`$ with $`1 \leq r \leq n`$ by the colour $`q + 1`$ on fibre $`r`$.
Let $`c = \left\lceil {{5\hspace{0pt}{\log k}} + 20} \right\rceil`$. We can assume $`k \geq {m\hspace{0pt}c}`$. For all vertices $`x`$, select $`m\hspace{0pt}c`$ different ordered elements $`e_{1},e_{2},\cdots,e_{m\hspace{0pt}c}`$ independently and uniformly. For all $`1 \leq i \leq m`$, let $`L_{x}^{i} = {\{ e_{{c\hspace{0pt}i} + 1},\cdots,e_{c\hspace{0pt}{({i + 1})}}\}}`$. Each set has the same distribution a set of $`c`$ elements chosen uniformly and independently.
Let $`A_{x}`$ be the event that the family $`\left\{ L_{y}^{i} \middle| {{y\hspace{0pt}x} \in {E\hspace{0pt}{(D)}\hspace{0pt}{\text{~and~}{y\hspace{0pt}x}\text{~is labelled~}i}}} \right\}`$ fails to have a transversal. By Theorem [23](#Thmtheorem23 "Theorem 23 (Alon, McDiarmid and Reed [2]) ‣ 5 Multiple Fibres ‣ WDM and Directed Star Arboricity"), $`{P\hspace{0pt}{(A_{x})}} \leq k^{3 - {c/2}}`$. Furthermore, the event $`A_{x}`$ is independent of all $`A_{y}`$ for which there is no vertex $`z`$ such that both $`z\hspace{0pt}x`$ and $`z\hspace{0pt}y`$ are in $`E\hspace{0pt}{(D)}`$. It follows that the dependency graph for these events has degree at most $`k^{2}`$, and so we can apply Lovász Local Lemma to obtain that there exists a family of lists satisfying conditions of Lemma [22](#Thmtheorem22 "Lemma 22 ‣ 5 Multiple Fibres ‣ WDM and Directed Star Arboricity"). This lemma gives the desired colouring. $`\square`$
For general digraphs, when we do not have $`{\Delta^{-},\Delta^{+}} \leq k`$, we may use the following trick to obtain an upper bound. Any digraph $`D`$ may be decomposed into an acyclic digraph $`D_{a}`$ and an Eulerian digraph $`D_{e}`$ (i.e., in $`D_{e}`$, for every vertex $`v`$, $`{d_{D_{e}}^{-}\hspace{0pt}{(V)}} = {d_{D_{e}}^{+}\hspace{0pt}{(v)}}`$). (To see this, consider an Eulerian subdigraph $`D_{e}`$ of $`D`$ which has a maximum number of arcs. Then the digraph $`D_{a} = {D - D_{e}}`$ is necessarily acyclic.) Hence by Lemma [18](#Thmtheorem18 "Lemma 18 ‣ 5 Multiple Fibres ‣ WDM and Directed Star Arboricity") (applied to $`D_{a}`$) and Theorem [21](#Thmtheorem21 "Theorem 21 ‣ 5 Multiple Fibres ‣ WDM and Directed Star Arboricity") (applied to $`D_{e}`$), we have
<table>
<tbody>
<tr>
<td></td>
<td>if m≥n, then λn(D)≤⌈mn⌈kn⌉+kn⌉+f(n,m,k),formulae-sequenceif 𝑚𝑛 then subscript𝜆𝑛𝐷𝑚𝑛𝑘𝑛𝑘𝑛𝑓𝑛𝑚𝑘\mbox{if }m\geq n,\mbox{ then }\lambda_{n}(D)\leq\left\lceil\frac{m}{n}\left\lceil\frac{k}{n}\right\rceil+\frac{k}{n}\right\rceil+f(n,m,k),</td>
<td></td>
</tr>
</tbody>
</table>
for $`f\hspace{0pt}{(n,m,k)}`$ the function given in Theorem [21](#Thmtheorem21 "Theorem 21 ‣ 5 Multiple Fibres ‣ WDM and Directed Star Arboricity"). But, as we will show now, it is possible to lessen this bound by roughly $`\frac{k}{n}`$.
######
Theorem 24
If $`m \geq n`$, then
<table>
<tbody>
<tr>
<td></td>
<td>λn(m,k)≤⌈mn⌈kn⌉+kn⌉+2m⌈(10m+5)logk+40m+21⌉n.subscript𝜆𝑛𝑚𝑘𝑚𝑛𝑘𝑛𝑘𝑛2𝑚10𝑚5𝑘40𝑚21𝑛\lambda_{n}(m,k)\leq\left\lceil\frac{m}{n}\left\lceil\frac{k}{n}\right\rceil+\frac{k}{n}\right\rceil+2m\frac{\left\lceil(10m+5)\log k+40m+21\right\rceil}{n}.</td>
<td></td>
</tr>
</tbody>
</table>
Proof. Let $`D`$ be an $`m`$-labelled digraph with $`{\Delta^{-}\hspace{0pt}{(D)}} \leq k`$. Consider a decomposition of $`D`$ into an acyclic digraph $`D_{a}`$ and an Eulerian digraph $`D_{e}`$. We first apply Theorem [21](#Thmtheorem21 "Theorem 21 ‣ 5 Multiple Fibres ‣ WDM and Directed Star Arboricity") to find an $`n`$-fibre colouring of the arcs of $`D_{e}`$ with $`f\hspace{0pt}{(n,m,k)}`$ colours such that, in addition, at most $`g\hspace{0pt}{(m,k)}`$ colours are assigned to the arcs leaving each vertex.
We shall extend the $`n`$-fibre colouring of $`D_{e}`$ to the arcs of $`D_{a}`$ in a way similar to the proof of Lemma [18](#Thmtheorem18 "Lemma 18 ‣ 5 Multiple Fibres ‣ WDM and Directed Star Arboricity"). I.e., we will assign to each vertex $`v`$, sets $`C_{i}\hspace{0pt}{(v)}`$, $`1 \leq i \leq m`$ of $`\lceil{\frac{k}{n} + {m\hspace{0pt}g\hspace{0pt}{(m,k)}}}\rceil`$ colours such that an arc labelled $`i`$ leaving $`v`$ will be labelled using a colour in $`C_{i}\hspace{0pt}{(v)}`$.
Let $`(v_{1},\ldots,v_{n})`$ be an ordering of the vertices of $`A`$ such that if $`v_{j}\hspace{0pt}v_{j'}`$ is an arc then $`j < j'`$. We start to build the $`C_{i}\hspace{0pt}{(v_{1})}`$ with the colours assigned to the leaving arcs of $`v_{1}`$ labelled $`i`$. The vertex $`v_{1}`$ has at most $`k`$ entering arcs. Each of them forbid one type (colour, fibre). In the colouring of $`D_{e}`$ induced by Theorem [21](#Thmtheorem21 "Theorem 21 ‣ 5 Multiple Fibres ‣ WDM and Directed Star Arboricity"), there are at most $`m\hspace{0pt}g\hspace{0pt}{(m,k)}`$ types assigned to the arcs leaving $`v_{1}`$. So there are at least $`{{\left\lceil \frac{m\hspace{0pt}k}{n^{2}} \right\rceil + \left\lceil \frac{k}{n} \right\rceil + {2\hspace{0pt}m\hspace{0pt}\frac{g\hspace{0pt}{(m,k)}}{n}}} - k - {m\hspace{0pt}g\hspace{0pt}{(m,k)}}} \geq {{m\hspace{0pt}\left\lceil \frac{k}{n} \right\rceil} + {m\hspace{0pt}g\hspace{0pt}{(m,k)}}}`$ types unused at vertex $`v_{1}`$. Since $`m \geq n`$, we can partition these types into $`m`$ sets of size at least $`\frac{k}{n}`$ such that no two types having the same colour are in the same set. These sets are the $`C_{i}\hspace{0pt}{(v_{1})}`$.
Suppose that the sets have been defined for $`v_{1}`$ up to $`v_{q - 1}`$, and that all the arcs $`v_{i}\hspace{0pt}v_{j}`$ for $`i < j < q`$ have been assigned a colour. We now give a colour to each arc $`v_{i}\hspace{0pt}v_{q}`$ for $`i < q`$.
There are $`k_{e}`$ arcs entering $`v_{q}`$ in $`D_{e}`$ which are already coloured. So it remains to give a colour to $`k_{a} \leq {k - k_{e}}`$ arcs. Each uncoloured arc may be assigned a colour in a list of size at least $`\left\lceil {\frac{k}{n} + {m\hspace{0pt}g\hspace{0pt}{(m,k)}}} \right\rceil`$. This gives a choice between $`n\hspace{0pt}\left\lceil {\frac{k}{n} + {m\hspace{0pt}g\hspace{0pt}{(m,k)}}} \right\rceil`$ different types. $`k_{e}`$ types are forbidden by the entering arcs in $`D_{e}`$ while at most $`m\hspace{0pt}g\hspace{0pt}{(m,k)}`$ types are forbidden by the leaving arcs in $`D_{e}`$. Hence, it remains at least $`{{n\hspace{0pt}\left\lceil {\frac{k}{n} + {m\hspace{0pt}g\hspace{0pt}{(m,k)}}} \right\rceil} - k_{e} - {m\hspace{0pt}g\hspace{0pt}{(m,k)}}} \geq k_{a}`$ types for the entering arcs of $`D_{a}`$. So one can assign distinct available colours to each of the $`k_{a}`$ arcs entering $`v_{q}`$. We then build the $`C_{i}\hspace{0pt}{(v_{q})}`$ as we did for $`v_{1}`$.
Once this process is finished, we obtain an $`n`$-fibre colouring of $`D`$ using $`\left\lceil \frac{m\hspace{0pt}k}{n^{2}} \right\rceil + \left\lceil \frac{k}{n} \right\rceil + {2\hspace{0pt}m\hspace{0pt}\frac{g\hspace{0pt}{(m,k)}}{n}}`$ colours.
$`\square`$
Theorem [24](#Thmtheorem24 "Theorem 24 ‣ 5 Multiple Fibres ‣ WDM and Directed Star Arboricity") gives an upper bound on $`\lambda_{n}\hspace{0pt}{(m,k)}`$ when $`m \geq n`$. We now give an upper bound for the case $`m < n`$.
######
Proposition 25
If $`m < n`$ then $`{\lambda_{n}\hspace{0pt}{(m,k)}} \leq \left\lceil \frac{k}{n - m} \right\rceil`$.
Proof. Let $`D`$ be an $`m`$-labelled digraph with $`\Delta^{-} \leq k`$. We should show the existence of a proper $`n`$-fibre colouring with $`\left\lceil \frac{k}{n - m} \right\rceil`$. For each vertex $`v`$, we give to each of its entering arcs a colour such that none of the colours is used more than $`n - m`$ times. This is possible since there are at most $`k \leq {{({n - m})}\hspace{0pt}\left\lceil \frac{k}{n - m} \right\rceil}`$ arcs entering $`v`$. So we now have $`{i\hspace{0pt}n\hspace{0pt}{(v,\lambda)}} \leq {n - m}`$. Moreover each arc $`v\hspace{0pt}w`$ is given a colour by $`w`$. Since $`D`$ is $`m`$-labelled, a colour $`\lambda`$ can be used to colour an arc of at most $`m`$ different labels, i.e., $`{o\hspace{0pt}u\hspace{0pt}t\hspace{0pt}{(v,\lambda)}} \leq m`$. Consequently $`{{i\hspace{0pt}n\hspace{0pt}{(v,\lambda)}} + {o\hspace{0pt}u\hspace{0pt}t\hspace{0pt}{(v,\lambda)}}} \leq n`$. This gives a proper $`n`$-fibre colouring. $`\square`$
## 6 Concluding Remarks
One question arising naturally from the previous sections is the complexity of calculating $`\lambda_{n}\hspace{0pt}{(D)}`$ for an $`m`$-labelled digraph $`D`$. As we will show in the first subsection, unsurprisingly, this problem is $`\mathcal{N}\hspace{0pt}\mathcal{P}`$-hard even for the simpler problem of directed star arboricity and even for restricted class of digraphs of in- and outdegree bounded by two. We end this section by showing how a similar approach to the one in Section [3](#S3 "3 Directed Star Arboricity of Subcubic Digraphs ‣ WDM and Directed Star Arboricity") allows us to give a very short proof of a recent result of Pinlou and Sopena \[[9](#bib.bib9)\].
### 6.1 Complexity
The digraphs with directed star arboricity one are the galaxies, so one can decide in polynomial time if $`{d\hspace{0pt}s\hspace{0pt}t\hspace{0pt}{(D)}} = 1`$. Deciding whether $`{d\hspace{0pt}s\hspace{0pt}t\hspace{0pt}{(D)}} = 2`$ or not is also easy since we just have to check that the conflict graph (with vertex set the arcs of $`D`$, two distinct arcs $`{x\hspace{0pt}y},{u\hspace{0pt}v}`$ being in conflict when $`y = u`$ or $`y = v`$) is bipartite. However for larger values, as expected, it is $`\mathcal{N}\hspace{0pt}\mathcal{P}`$-complete to decide if a digraph has directed star arboricity at most $`k`$. This is illustrated by the next result:
######
Theorem 26
The following decision problem is $`\mathcal{N}\hspace{0pt}\mathcal{P}`$-complete:
*Instance: A digraph $`D`$ with $`{\Delta^{+}\hspace{0pt}{(D)}} \leq 2`$ and $`{\Delta^{-}\hspace{0pt}{(D)}} \leq 2`$.*
*Question: Is $`d\hspace{0pt}s\hspace{0pt}t\hspace{0pt}{(D)}`$ at most $`3`$?*
Proof. The proof is by a reduction from $`3`$-edge-colouring of $`3`$-regular graphs, which is known to be $`\mathcal{N}\hspace{0pt}\mathcal{P}`$-complete.
Let $`G`$ be a $`3`$-regular graph. It is easy to see that $`G`$ admits an orientation $`D`$ such that every vertex has in- and outdegree at least one (i.e., $`D`$ does not have neither sink nor source).
Let $`D'`$ be the digraph obtained from $`D`$ by replacing every vertex with indegree one and outdegree two by the subgraph $`H`$ depicted in Figure [2](#S6.F2 "Figure 2 ‣ 6.1 Complexity ‣ 6 Concluding Remarks ‣ WDM and Directed Star Arboricity") which has also one entering arc (namely $`\overset{\rightarrow}{a}`$) and two leaving arcs ($`\overset{\rightarrow}{b}`$ and $`\overset{\rightarrow}{c}`$).
[Refer to caption](/html/0705.0315/assets/x2.png)
Figure 2: The graph $`H`$ and one of its directed star $`3`$-colouring
It is quite easy to check that in any directed star $`3`$-colouring of $`H`$, the three arcs $`\overset{\rightarrow}{a}`$, $`\overset{\rightarrow}{b}`$ and $`\overset{\rightarrow}{c}`$ get different colours. Moreover, if these three arcs are precoloured with three different colours, we can extend this to a directed star $`3`$-colouring of $`H`$. Such a colouring with $`\overset{\rightarrow}{a}`$ coloured $`1`$, $`\overset{\rightarrow}{b}`$ coloured $`2`$ and $`\overset{\rightarrow}{c}`$ coloured $`3`$ is given in Figure [2](#S6.F2 "Figure 2 ‣ 6.1 Complexity ‣ 6 Concluding Remarks ‣ WDM and Directed Star Arboricity"). Furthermore, in a directed star $`3`$-colouring, a vertex with indegree two and outdegree one must have its three incident arcs coloured differently. So $`{d\hspace{0pt}s\hspace{0pt}t\hspace{0pt}{(D')}} = 3`$ if and only if $`G`$ is $`3`$-edge colourable. $`\square`$
### 6.2 Acircuitic Directed Star Arboricity
A directed star colouring is acircuitic if there is no bicoloured circuits, i.e., circuits for which only two colours appear on its arcs. The acircuitic directed star arboricity of a digraph $`D`$ is the minimum number $`k`$ of colours such that there exists an acircuitic directed star $`k`$-colouring of $`D`$.
In this last section, we give a short alternative proof of the following theorem.
######
Theorem 27
(Pinlou and Sopena \[[9](#bib.bib9)\])
Every subcubic oriented graph has acircuitic directed star arboricity at most $`4`$.
Indeed, it is possible to apply our Theorem [5](#Thmtheorem5 "Theorem 5 ‣ 1 Introduction ‣ WDM and Directed Star Arboricity") directly to derive this theorem. However, there is a shorter proof using the following lemma.
######
Lemma 28
Let $`D`$ be an acyclic subcubic digraph. Let $`L`$ be a list-assignment on the arcs of $`D`$ such that for every arc $`u\hspace{0pt}v`$, $`{|{L\hspace{0pt}{({u\hspace{0pt}v})}}|} \geq {d\hspace{0pt}{(v)}}`$. Then $`D`$ admits a directed star $`L`$-colouring.
Proof. We prove the result by induction on the number of arcs of $`D`$, the result holds trivially if $`D`$ has no arcs.
Since $`D`$ is acyclic, it has an arc $`x\hspace{0pt}y`$ with $`y`$ a sink. Let $`\omega`$ be a colour in $`L\hspace{0pt}{({x\hspace{0pt}y})}`$. For any arc $`\overset{\rightarrow}{a}`$ distinct from $`x\hspace{0pt}y`$, set $`{L'\hspace{0pt}{(\overset{\rightarrow}{a})}} = {{L\hspace{0pt}{(\overset{\rightarrow}{a})}} \smallsetminus {\{\omega\}}}`$ if $`\overset{\rightarrow}{a}`$ is incident to $`x\hspace{0pt}y`$ (and thus has head in $`\{ x,y\}`$ since $`y`$ is a sink), and $`{L'\hspace{0pt}{(\overset{\rightarrow}{a})}} = {L\hspace{0pt}{(\overset{\rightarrow}{a})}}`$ otherwise. Then in $`D' = {D - {x\hspace{0pt}y}}`$, we have $`{|{L'\hspace{0pt}{({u\hspace{0pt}v})}}|} \geq {d\hspace{0pt}{(v)}}`$ for any arc $`{u\hspace{0pt}v} \neq {x\hspace{0pt}y}`$. Hence, by induction hypothesis, $`D'`$ admits a directed star $`L'`$-colouring that can be extended to a directed star $`L`$-colouring of $`D`$ by colouring $`x\hspace{0pt}y`$ with $`\omega`$. $`\square`$
Proof of Theorem [27](#Thmtheorem27 "Theorem 27 (Pinlou and Sopena [9]) ‣ 6.2 Acircuitic Directed Star Arboricity ‣ 6 Concluding Remarks ‣ WDM and Directed Star Arboricity").
Let $`V_{1}`$ be the set of vertices of outdegree at most one and $`V_{2} = {V \smallsetminus V_{1}}`$. Every vertex of $`V_{2}`$ has outdegree at least two (and so indegree at most one).
Let $`M`$ be the set of arcs with tail in $`V_{1}`$ and head in $`V_{2}`$. We colour all the arcs of $`M`$ with colour $`4`$. Moreover for every circuit $`C`$ in $`D\hspace{0pt}{\lbrack V_{1}\rbrack}`$ or in $`D\hspace{0pt}{\lbrack V_{2}\rbrack}`$, we choose an arc $`\overset{\rightarrow}{a}\hspace{0pt}{(C)}`$ and colour it by $`4`$. Note that, by definition of $`V_{1}`$ and $`V_{2}`$, the arc $`\overset{\rightarrow}{a}\hspace{0pt}{(C)}`$ is not incident to any arc of $`M`$, and in addition, $`C`$ is the unique circuit containing $`\overset{\rightarrow}{a}\hspace{0pt}{(C)}`$. Let us denote by $`M_{4}`$ the set of all arcs coloured by $`4`$. It is easily seen that $`M_{4}`$ is a matching and $`D - M_{4}`$ is acyclic.
We shall now find a directed star colouring of $`D - M_{4}`$ with colours $`\{ 1,2,3\}`$ that does not create any bicoloured circuit. In any colouring of the arcs, if such a circuit existed, $`4`$ would be one of its colour because $`D - M_{4}`$ is acyclic, and moreover, all the arcs of this circuit coloured by $`4`$ would be in $`M`$, because each arc in $`M_{4} \smallsetminus M`$ is in a unique circuit and this unique circuit has a unique arc coloured by $`4`$. Hence we just have to be careful when dealing with arcs in the digraph induced by the endvertices of the arcs of $`M`$.
Let us denote the arcs of $`M`$ by $`x_{i}\hspace{0pt}y_{i}`$, $`1 \leq i \leq p`$, and set $`X = {\{{{x_{i},1} \leq i \leq p}\}}`$ and $`Y = {\{{{y_{i},1} \leq i \leq p}\}}`$ (we have then $`x_{i} \in V_{1}`$ and $`y_{i} \in V_{2}`$). Let $`E'`$ be the set of arcs with tail in $`Y`$ and head in $`X`$. Let $`H`$ be the graph with vertex set $`E'`$ such that an arc $`y_{i}\hspace{0pt}x_{j}`$ is adjacent to an arc $`y_{k}\hspace{0pt}x_{l}`$ if
- $`(a)`$
Either $`k = l`$,
- $`(b)`$
Or $`j = k`$ and $`i > j`$ and $`l > j`$.
Since a vertex of $`X`$ has indegree at most two and a vertex of $`Y`$ has outdegree at most two, $`H`$ has maximum degree three. Moreover $`H`$ contains no $`K_{4}`$, because two arcs of $`E'`$ with same tail $`y_{k}`$ are not adjacent in $`H`$. Hence, by Brooks Theorem, $`H`$ has a vertex-colouring with colours $`\{ 1,2,3\}`$, and this colouring corresponds to a colouring $`c`$ of the arcs of $`E'`$. Since $`(a)`$ is satisfied, $`c`$ is a directed star colouring. Moreover, this colouring creates no bicoloured circuits. Indeed, a circuit contains a subpath $`y_{i}\hspace{0pt}x_{j}\hspace{0pt}y_{j}\hspace{0pt}x_{l}`$, with $`i > j`$ and $`k > j`$, whose three arcs are coloured differently by $`(b)`$.
Let $`D' = {D - {({M_{4} \cup E'})}}`$. For any arc $`u\hspace{0pt}v`$ in $`D'`$, let $`{L\hspace{0pt}{({u\hspace{0pt}v})}} = {{\{ 1,2,3\}} \smallsetminus \left. \{{c\hspace{0pt}{({w\hspace{0pt}v})}} \middle| {{w\hspace{0pt}v} \in E'}\} \right.}`$. The set $`L\hspace{0pt}{({u\hspace{0pt}v})}`$ is the set of colours in $`\{ 1,2,3\}`$ that may be assigned to $`u\hspace{0pt}v`$ without creating any conflict with the already coloured arcs. The digraph $`D'`$ is acyclic and $`{|{L\hspace{0pt}{({u\hspace{0pt}v})}}|} \geq {d\hspace{0pt}{(v)}}`$, so by Lemma [28](#Thmtheorem28 "Lemma 28 ‣ 6.2 Acircuitic Directed Star Arboricity ‣ 6 Concluding Remarks ‣ WDM and Directed Star Arboricity"), it admits a directed star $`L`$-colouring. We infer that $`D`$ has an acircuitic directed star colouring with colours in $`\{ 1,2,3,4\}`$ and the theorem follows. In addition, we note that in this colouring, the arcs coloured by $`4`$ form a matching. $`\square`$
Acknowledgement. The authors are grateful to J-C. Bermond for bringing out their attention to the paper by Brandt and Gonzalez. They thank an anonymous referee for the remarks which helped them to improve the presentation of the paper.
## References
- \[1\]
I. Algor and N. Alon, The star arboricity of graphs, Discrete Mathematics 75 (1989), 11–22.
- \[2\]
N. Alon, C. McDiarmid, B. Reed, Star arboricity, Combinatorica 12 (1992), 375–380.
- \[3\]
R. Brandt, Multicasting using WDM in Multifiber Optical Star Networks, Thesis, UCSB, 2003.
- \[4\]
R. Brandt and T. F. Gonzalez, Wavelength assignment in multifiber optical star networks under the multicasting communication mode, Journal of Interconnection Networks 6 (2005), 383–405.
- \[5\]
A. Frank, Covering branchings, Acta Scientiarum Mathematicarum (Szeged) 41 (1979), 77–81.
- \[6\]
B. Guiduli, On incidence coloring and star arboricity of graphs, Discrete Mathematics 163 (1997), 275–278.
- \[7\] A. Lardies, R. Gupta and R. Patterson. Traffic grooming in a multi-layer network.
Optical Networks Magazine 2, (2001), 91–99.
- \[8\] E. Modiano and P.J. Lin. Traffic grooming in WDM networks.
IEEE Communications Magazine, 39(7), (2001), 124–129.
- \[9\]
A. Pinlou and E. Sopena, The acircuitic directed star arboricity of subcubic graph is at most four, Discrete Mathematics 306(24) (2006), 3281–3289.
- \[10\]
A. Schrijver, Combinatorial optimization. Polyhedra and efficiency, Algorithms and Combinatorics 24, Springer-Verlag, Berlin, 2003.
- \[11\] V. G. Vizing. On an estimate of the chromatic class of a p-graph.
Metody Diskret. Analyz. 3 (1964), 25–30. |
0707.3670 | cs.NI | cyber | # Understanding the Characteristics of Internet Short Video Sharing: YouTube as a Case Study
Xu Cheng
School of Computing Science
Simon Fraser University
Burnaby, BC, Canada
Email: xuc@cs.sfu.ca
Cameron Dale
School of Computing Science
Simon Fraser University
Burnaby, BC, Canada
Email: camerond@cs.sfu.ca
Jiangchuan Liu
School of Computing Science
Simon Fraser University
Burnaby, BC, Canada
Email: jcliu@cs.sfu.ca
###### Abstract
Established in 2005, YouTube has become the most successful Internet site providing a new generation of short video sharing service. Today, YouTube alone comprises approximately 20% of all HTTP traffic, or nearly 10% of all traffic on the Internet. Understanding the features of YouTube and similar video sharing sites is thus crucial to their sustainable development and to network traffic engineering.
In this paper, using traces crawled in a 3-month period, we present an in-depth and systematic measurement study on the characteristics of YouTube videos. We find that YouTube videos have noticeably different statistics compared to traditional streaming videos, ranging from length and access pattern, to their active life span, ratings, and comments. The series of datasets also allows us to identify the growth trend of this fast evolving Internet site in various aspects, which has seldom been explored before.
We also look closely at the social networking aspect of YouTube, as this is a key driving force toward its success. In particular, we find that the links to related videos generated by uploaders’ choices form a small-world network. This suggests that the videos have strong correlations with each other, and creates opportunities for developing novel caching or peer-to-peer distribution schemes to efficiently deliver videos to end users.
## I Introduction
The recent two years have witnessed an explosion of networked video sharing as a new killer Internet application. The most successful site, YouTube, now features over 40 million videos and enjoys 20 million visitors each month \[[1](#bib.bib1)\]. The success of similar sites like GoogleVideo, YahooVideo, MySpace, ClipShack, and VSocial, and the recent expensive acquisition of YouTube by Google, further confirm the mass market interest. Their great achievement lies in the combination of the content-rich videos and, equally or even more importantly, the establishment of a social network. These sites have created a video village on the web, where anyone can be a star, from lip-synching teenage girls to skateboarding dogs. With no doubt, they are changing the content distribution landscape and even the popular culture \[[2](#bib.bib2)\].
Established in 2005, YouTube is one of the fastest-growing websites, and has become the 4th most accessed site in the Internet. It has a significant impact on the Internet traffic distribution, and itself is suffering from severe scalability constraints. Understanding the features of YouTube and similar video sharing sites is crucial to network traffic engineering and to sustainable development of this new generation of service.
In this paper, we present an in-depth and systematic measurement study on the characteristics of YouTube videos. We have crawled the YouTube site for a 3-month period in early 2007, and have obtained 27 datasets totaling 2,676,388 videos. This constitutes a significant portion of the entire YouTube video repository, and because most of these videos are accessible from the YouTube homepage in less than 10 clicks, they are generally active and thus representative for measuring the repository. Using this collection of datasets, we find that YouTube videos have noticeably different statistics from traditional streaming videos, in aspects from video length and access pattern, to life span. There are also new features that have not been examined by previous measurement studies, for example, the ratings and comments. In addition, the series of datasets also allows us to identify the growth trend of this fast evolving Internet site in various aspects, which has seldom been explored before.
We also look closely at the social networking aspect of YouTube, as this is a key driving force toward the success of YouTube and similar sites. In particular, we find that the links to related videos generated by uploader’s choices form a small-world network. This suggests that the videos have strong correlations with each other, and creates opportunities for developing novel caching or peer-to-peer distribution schemes to efficiently deliver videos to end users.
The rest of the paper is organized as follows. Section [II](#S2 "II Background and Related Work ‣ Understanding the Characteristics of Internet Short Video Sharing: YouTube as a Case Study") presents some background information and other related work. Section [III](#S3 "III Methodology of Measurement ‣ Understanding the Characteristics of Internet Short Video Sharing: YouTube as a Case Study") describes our method of gathering information about YouTube videos, which is analyzed generally in Section [IV](#S4 "IV Characteristics of YouTube Video ‣ Understanding the Characteristics of Internet Short Video Sharing: YouTube as a Case Study"), while the social networking aspects are analyzed separately in Section [V](#S5 "V The Social Network in YouTube ‣ Understanding the Characteristics of Internet Short Video Sharing: YouTube as a Case Study"). Section [VI](#S6 "VI Further Discussions ‣ Understanding the Characteristics of Internet Short Video Sharing: YouTube as a Case Study") discusses the implications of the results, and suggests ways that the YouTube service could be improved. Finally, Section [VII](#S7 "VII Conclusion ‣ Understanding the Characteristics of Internet Short Video Sharing: YouTube as a Case Study") concludes the paper.
## II Background and Related Work
###
II-A
Internet Video Sharing
Online videos existed long before YouTube entered the scene. However, uploading videos, managing, sharing and watching them was very cumbersome due to a lack of an easy-to-use integrated platform. More importantly, the videos distributed by traditional media servers and peer-to-peer file downloads like BitTorrent were standalone units of content. Each single video was not connected to other related video clips, for example other episodes of a show that the user had just watched. Also, there was very little in the way of content reviews or ratings.
The new generation of video sharing sites, YouTube and its competitors, overcame these problems. They allow content suppliers to upload video effortlessly, automatically converting from many different formats, and to tag uploaded videos with keywords. Users can easily share videos by mailing links to them, or embedding them on web pages or in blogs. Users can also rate and comment on videos, bringing new social aspects to the viewing of videos. Consequently, popular videos can rise to the top in a very organic fashion.
The social network existing in YouTube further enables communities and groups. Videos are no longer independent from each other, and neither are users. This has substantially contributed to the success of YouTube and similar sites.
###
II-B
Workload Measurement of Traditional Media Servers
There has been a significant research effort into understanding the workloads of traditional media servers, looking at, for example, the video popularity and access locality \[[3](#bib.bib3), [4](#bib.bib4), [5](#bib.bib5), [6](#bib.bib6)\]. The different aspects of media and web objects, and those of live and stored video streams have also been compared \[[7](#bib.bib7), [8](#bib.bib8)\]. We have found that, while sharing similar features, many of the video statistics of these traditional media servers are quite different from YouTube; for example, the video length distribution and life span. More importantly, these traditional studies lack a social network among the videos.
The most similar work to ours is the very recent study by Huang et. al. \[[9](#bib.bib9)\]. They analyzed a 9-month trace of *MSN Video*, Microsoft’s VoD service, examining the user behavior and popularity distribution of videos. This analysis led to a *peer-assisted VoD* design for reducing the server’s bandwidth costs. The difference to our work is that MSN Video is a more traditional video service, with much fewer videos, most of which are longer than all YouTube videos. MSN Video also has no listings of related videos or user information, and thus no social networking aspect.
## III Methodology of Measurement
In this paper, we focus on the access patterns and social networks present in YouTube. To this end, we have crawled the YouTube site for a 3-month period and obtained information on its videos through a combination of the YouTube API and scrapes of YouTube video web pages. The results offer a series of representative partial snapshots of the YouTube video repository as well as its changing trends.
###
III-A
Video Format and Meta-data
YouTube’s video playback technology is based on Macromedia’s Flash Player and uses the Sorenson Spark H.263 video codec with pixel dimensions of 320 by 240 and 25 frames per second. This technology allows YouTube to display videos with quality comparable to more established video playback technologies (such as Windows Media Player, Realplayer or Apple’s Quicktime Player). YouTube accepts uploaded videos in WMV, AVI, MOV and MPEG formats, which are converted into .FLV (Adobe Flash Video) format after uploading \[[10](#bib.bib10)\]. It has been recognized that the use of a uniform easily-playable format has been a key in the success of YouTube.
There are many ways that YouTube’s service differs from a traditional media server. YouTube’s FLV videos are not streamed to the user, but are instead downloaded over a normal HTTP connection. They are also not rate controlled to the playback rate of the video but are sent at the maximum rate that the server and user can accomplish, and there is no user interactivity from the server’s point of view (except for possibly stopping the download). In order to fast forward the user must wait for that part of the video to download, and pausing the playback does not pause the download.
YouTube randomly assigns each video a distinct 64-bit number, which is represented in base 64 by an 11-digit ID composed of 0-9, a-z, A-Z, -, and \_. Each video contains the following intuitive meta-data: user who uploaded it, date when it was uploaded, category, length, number of views, number of ratings, number of comments, and a list of “related videos”. The related videos are links to other videos that have a similar title, description, or tags, all of which are chosen by the uploader. A video can have hundreds of related videos, but the webpage only shows at most 20 at once, so we limit our scrape to these top 20 related videos. A typical example of the meta-data is shown in Table [I](#S3.T1 "TABLE I ‣ III-A Video Format and Meta-data ‣ III Methodology of Measurement ‣ Understanding the Characteristics of Internet Short Video Sharing: YouTube as a Case Study").
<table>
<tbody>
<tr>
<th>ID</th>
<td>2AYAY2TLves</td>
</tr>
<tr>
<th>Uploader</th>
<td>GrimSanto</td>
</tr>
<tr>
<th>Added Date</th>
<td>May 19, 2007</td>
</tr>
<tr>
<th>Category</th>
<td>Gadgets & Games</td>
</tr>
<tr>
<th>Video Length</th>
<td>268 seconds</td>
</tr>
<tr>
<th>Number of Views</th>
<td>185,615</td>
</tr>
<tr>
<th>Number of Ratings</th>
<td>546</td>
</tr>
<tr>
<th>Number of Comments</th>
<td>588</td>
</tr>
<tr>
<th>Related Videos</th>
<td>aUXoekeDIW8,</td>
</tr>
<tr>
<th></th>
<td>Sog2k6s7xVQ, …</td>
</tr>
</tbody>
</table>
TABLE I: Meta-data of a YouTube Video
###
III-B
YouTube Crawler
We consider all the YouTube videos to form a directed graph, where each video is a node in the graph. If video $`b`$ is in the related video list (first 20 only) of video $`a`$, then there is a directed edge from $`a`$ to $`b`$. Our crawler uses a breadth-first search to find videos in the graph. We define the initial set of 0-depth video IDs, which the crawler reads in to a queue at the beginning of the crawl. When processing each video, it checks the list of related videos and adds any new ones to the queue. The crawler is single-threaded to avoid being suspected of a network attack.
Given a video ID, the crawler first extracts information from the YouTube API, which contains all the meta-data except date added, category, and related videos. The crawler then scrapes the video’s webpage to obtain the remaining information.
Our first crawl was on February 22nd, 2007, and started with the initial set of videos from the list of “Recently Featured”, “Most Viewed”, “Top Rated” and “Most Discussed”, for “Today”, “This Week”, “This Month” and ”All Time”, which totalled 189 unique videos on that day. The crawl went to more than four depths (the fifth was not completed), finding approximately 750 thousand videos in about five days.
In the following weeks we ran the the crawler every two to three days, each time defining the initial set of videos from the list of “Most Viewed”, “Top Rated”, and “Most Discussed”, for “Today” and “This Week”, which is about 200 to 300 videos. On average, the crawl finds 80 thousand videos each time in less than 10 hours.
To study the growth trend of the video popularity, we also use the crawler to update the statistics of some previously found videos. For this crawl we only retrieve the number of views for relatively new videos (uploaded after February 15th, 2007). This crawl is performed once a week from March 5th to April 16th 2007, which results in seven datasets.
We also separately crawled the file size and bit-rate information. To get the file size, the crawler retrieves the response information from the server when requesting to download the video file and extracts the information on the size of the download. Some videos also have the bit-rate embedded in the FLV video meta-data, which the crawler extracts after downloading the beginning of the video file.
Finally, we have also collected some information about YouTube users. The crawler retrieves information on the number of uploaded videos and friends of each user from the YouTube API, for a total of more than 1 million users.
## IV Characteristics of YouTube Video
From the first crawling on February 22nd, 2007, to the end of April, 2007, we have obtained 27 datasets totaling 2,676,388 videos. This constitutes a significant portion of the entire YouTube video repository (there are an estimated 42.5 million videos on YouTube \[[11](#bib.bib11)\]). Also, because most of these videos can be accessed from the YouTube homepage in less than 10 clicks, they are generally active and thus representative for measuring characteristics of the repository.
In the measurements, some characteristics are static and can be measured once from the entire dataset: e.g. category, length, and date added. Some characteristics are dynamic and can change from dataset to dataset: e.g. number of views, ratings, and comments. We consider this dynamic information to be static over a single crawl. Later, the updated number of views information will be used to measure the growth trend and life span of videos.
###
IV-A
Video Category
<table>
<thead>
<tr>
<th>Category</th>
<th>Count</th>
<th>Percentage</th>
</tr>
</thead>
<tbody>
<tr>
<td>Autos & Vehicles</td>
<td>66878</td>
<td>2.5%</td>
</tr>
<tr>
<td>Comedy</td>
<td>323814</td>
<td>12.1%</td>
</tr>
<tr>
<td>Entertainment</td>
<td>475821</td>
<td>17.8%</td>
</tr>
<tr>
<td>Film & Animation</td>
<td>225817</td>
<td>8.4%</td>
</tr>
<tr>
<td>Gadgets & Games</td>
<td>196026</td>
<td>7.3%</td>
</tr>
<tr>
<td>Howto & DIY</td>
<td>53291</td>
<td>2.0%</td>
</tr>
<tr>
<td>Music</td>
<td>613754</td>
<td>22.9%</td>
</tr>
<tr>
<td>News & Politics</td>
<td>116153</td>
<td>4.3%</td>
</tr>
<tr>
<td>People & Blogs</td>
<td>199014</td>
<td>7.4%</td>
</tr>
<tr>
<td>Pets & Animals</td>
<td>50092</td>
<td>1.9%</td>
</tr>
<tr>
<td>Sports</td>
<td>258375</td>
<td>9.7%</td>
</tr>
<tr>
<td>Travel & Places</td>
<td>58678</td>
<td>2.2%</td>
</tr>
<tr>
<td>Unavailable</td>
<td>24068</td>
<td>0.9%</td>
</tr>
<tr>
<td>Removed</td>
<td>14607</td>
<td>0.5%</td>
</tr>
</tbody>
</table>
TABLE II: List of YouTube video categories
[Refer to caption](/html/0707.3670/assets/x1.png)
Figure 1: Distribution of YouTube Videos’ Categories
One of 12 categories is selected by the user when uploading the video. Table [II](#S4.T2 "TABLE II ‣ IV-A Video Category ‣ IV Characteristics of YouTube Video ‣ Understanding the Characteristics of Internet Short Video Sharing: YouTube as a Case Study") lists the number and percentage of all the categories, which is also shown graphically in Figure [1](#S4.F1 "Figure 1 ‣ IV-A Video Category ‣ IV Characteristics of YouTube Video ‣ Understanding the Characteristics of Internet Short Video Sharing: YouTube as a Case Study"). In our entire dataset we can see that the distribution is highly skewed: the most popular category is Music, at about 22.9%; the second is Entertainment, at about 17.8%; and the third is Comedy, at about 12.1%.
In the table, we also list two other categories. “Unavailable” are videos set to private, or videos that have been flagged as inappropriate video, which the crawler can only get information for from the YouTube API. “Removed” are videos that have been deleted by the uploader, or by a YouTube moderator (due to the violation of the terms of use), but still are linked to by other videos.
[Refer to caption](/html/0707.3670/assets/x2.png)
Figure 2: Distribution of Video Length
[Refer to caption](/html/0707.3670/assets/x3.png)
Figure 3: Video Length for the 4 Most Popular Categories
[Refer to caption](/html/0707.3670/assets/x4.png)
Figure 4: Video Bit-rate
###
IV-B
Video Length
The length of YouTube videos is the biggest difference from traditional media content servers. Whereas most traditional servers contain a small to medium number of long videos, typically 0.5-2 hour movies (e.g. HPLabs Media Server \[[3](#bib.bib3)\]), YouTube is mostly comprised of videos that are short clips.
In our entire dataset, 97.8% of the videos’ lengths are within 600 seconds, and 99.1% are within 700 seconds. This is mainly due to the limit of 10 minutes imposed by YouTube on regular users uploads. We do find videos longer than this limit though, as the limit was only established in March, 2006, and also the YouTube Director Program allows a small group of authorized users to upload videos longer than 10 minutes \[[12](#bib.bib12)\].
Figure [4](#S4.F4 "Figure 4 ‣ IV-A Video Category ‣ IV Characteristics of YouTube Video ‣ Understanding the Characteristics of Internet Short Video Sharing: YouTube as a Case Study") shows the histogram of YouTube videos’ lengths within 700 seconds, which exhibits three peaks. The first peak is within one minute, and contains more than 20% of the videos, which shows that YouTube is primarily a site for very short videos. The second peak is between 3 and 4 minutes, and contains about 16.7% of the videos. This peak is mainly caused by the large number of videos in the “Music” category. Music is the most popular category for YouTube, and the typical length of a music video is often within this range, as shown in Figure [4](#S4.F4 "Figure 4 ‣ IV-A Video Category ‣ IV Characteristics of YouTube Video ‣ Understanding the Characteristics of Internet Short Video Sharing: YouTube as a Case Study"). The third peak is near the maximum of 10 minutes, and is caused by the limit on the length of uploaded videos. This encourages some users to circumvent the length restriction by dividing long videos into several parts, each being near the limit of 10 minutes.
We find that the the length histogram can be fit by an aggregate of four normal distributions, whose parameters are shown in Table [III](#S4.T3 "TABLE III ‣ IV-B Video Length ‣ IV Characteristics of YouTube Video ‣ Understanding the Characteristics of Internet Short Video Sharing: YouTube as a Case Study"). The location parameter $`\mu`$ determines the mean, the scale parameter $`\sigma`$ determines the width, and the ratio $`r`$ shows the weight of the four curves in the aggregated distribution. The first three columns in the table correspond to the three peaks of the distribution, while the last column represents the rest of the data.
<table>
<tbody>
<tr>
<th>Parameter</th>
<td>Peak 1</td>
<td>Peak 2</td>
<td>Peak 3</td>
<td>The Rest</td>
</tr>
<tr>
<th>μ𝜇\mu</th>
<td>16</td>
<td>208</td>
<td>583</td>
<td>295</td>
</tr>
<tr>
<th>σ𝜎\sigma</th>
<td>62</td>
<td>58</td>
<td>16</td>
<td>172</td>
</tr>
<tr>
<th>r𝑟r</th>
<td>48.6%</td>
<td>26.2%</td>
<td>2.7%</td>
<td>22.5%</td>
</tr>
</tbody>
</table>
TABLE III: Parameters for the aggregated normal distribution
Figure [4](#S4.F4 "Figure 4 ‣ IV-A Video Category ‣ IV Characteristics of YouTube Video ‣ Understanding the Characteristics of Internet Short Video Sharing: YouTube as a Case Study") shows the video length histograms for the top four most popular categories. We can see “Music” videos have a very large peak between three and four minutes, and “Entertainment” videos have a similar (though smaller) peak. In comparison, “Comedy” and “Sports” videos have more videos within two minutes, probably corresponding to “highlight” type clips. We also used an aggregated normal distribution to get the fits for the four length distributions.
###
IV-C
File Size and Bit-rate
Using the video IDs from the normal crawl on April 10th 2007 (about 200 thousand videos), we retrieved the file size of nearly 190 thousand videos. In our crawled data, 98.8% of the videos are less than 30MB. Not surprisingly, we find that the distribution of video sizes is very similar to the distribution of video lengths. We calculate an average video file size to be about 8.4 MBytes. Considering there are over 42.5 million YouTube videos, the total disk space required to store all the videos is more than 357 terabytes! Smart storage management is thus quite demanding for such a ultra-huge and still growing site, which we will discuss in more detail in Section [VI](#S6 "VI Further Discussions ‣ Understanding the Characteristics of Internet Short Video Sharing: YouTube as a Case Study").
We found that 87.6% of the videos we crawled contained FLV meta-data specifying the video’s bit-rate in the beginning of the file, indicating that they are Constant-Bit-Rate (CBR). For the rest of the videos that do not contain this meta-data (probably Variable-Bit-Rate, or VBR, videos), we calculate an average bit-rate from the file size and its length.
In Figure [4](#S4.F4 "Figure 4 ‣ IV-A Video Category ‣ IV Characteristics of YouTube Video ‣ Understanding the Characteristics of Internet Short Video Sharing: YouTube as a Case Study"), the videos’ bit-rate has three clear peaks. Most videos have a bit-rate around 330 kbps, with two other peaks at around 285 kbps and 200 kbps. This implies that YouTube videos have a moderate bit-rate that balances the quality and the bandwidth.
[Refer to caption](/html/0707.3670/assets/x5.png)
Figure 5: Date Added
[Refer to caption](/html/0707.3670/assets/x6.png)
Figure 6: Rank of Views
[Refer to caption](/html/0707.3670/assets/x7.png)
Figure 7: Rank of Ratings and Comments
###
IV-D
Date Added – Growth Trend of Uploading
During our crawl we record the date that each video was uploaded, so we can study the growth trend of YouTube. Figure [7](#S4.F7 "Figure 7 ‣ IV-C File Size and Bit-rate ‣ IV Characteristics of YouTube Video ‣ Understanding the Characteristics of Internet Short Video Sharing: YouTube as a Case Study") shows the number of new videos added every two weeks in our entire crawled dataset.
February 15th, 2005 is the day that YouTube was established. Our first crawl was on February 22nd, 2007, thus we can get the early videos only if they are still very popular videos or are linked to by other videos we crawled. We can see there is a slow start, the earliest video we crawled was uploaded on April 27th, 2005. After 6 months from YouTube’s establishment, the number of uploaded videos increases steeply. We use a power law curve to fit this trend.
In the dataset we collected, the number of uploaded videos decreases linearly and steeply starting in March, 2007. However, this does not imply that the uploading rate of YouTube videos has suddenly decreased. The reason is that many recently uploaded videos have not been so popular, and are probably not listed in other videos related videos’ list. Since few videos have linked to those new videos, they are not likely to be found by our crawler. Nevertheless, as those videos become popular or get linked to by others, our crawler may find them and get their information. Comparing the entire dataset to the first and largest dataset, which was crawled on February 22nd, we also see the same trend.
###
IV-E
Views, Ratings – User Access Pattern
The number of views a video has had is the most important characteristic we measured, as it reflects the popularity and access patterns of the videos. Because this property is changing over time, we cannot use the entire dataset that combines all the data together. Therefore we use a single dataset from April 3rd, 2007, containing more than 100 thousand videos, which is considered to be relatively static.
Figure [7](#S4.F7 "Figure 7 ‣ IV-C File Size and Bit-rate ‣ IV Characteristics of YouTube Video ‣ Understanding the Characteristics of Internet Short Video Sharing: YouTube as a Case Study") shows the number of views as a function of the rank of the video by its number of views. Though the plot has a long tail on the linear scale, it does NOT follow a Zipf distribution, which should be a straight line on a log-log scale. This is consistent with some previous observations \[[6](#bib.bib6), [3](#bib.bib3), [4](#bib.bib4), [5](#bib.bib5)\] that also found that video accesses on a media server does not follow Zipf’s law. We can see in the figure, the beginning of the curve is linear on a log-log scale, but the tail (after the $`2 \times 10^{3}`$ video) decreases tremendously, indicating there are not so many less popular videos as Zipf’s law predicts. This result seems consistent with some results \[[6](#bib.bib6)\], but differs from others \[[3](#bib.bib3), [4](#bib.bib4), [5](#bib.bib5)\] in which the curve is skewed from linear from beginning to end. Their results indicate that the popular videos are also not as popular as Zipf’s law predicts, which is not the case in our experiment.
To fit the skewed curve, some use a generalized Zipf-like distribution \[[3](#bib.bib3)\], while others use a concatenation of two Zipf-like distributions \[[5](#bib.bib5)\]. Because our curve is different, we attempted to use three different distributions: Weibull, Gamma and Zipf. We find that Weibull and Gamma distributions both fit better than Zipf, due to the drop-off in the tail (in log-log scale) that they have.
Figure [7](#S4.F7 "Figure 7 ‣ IV-C File Size and Bit-rate ‣ IV Characteristics of YouTube Video ‣ Understanding the Characteristics of Internet Short Video Sharing: YouTube as a Case Study") plots the number of ratings against the rank of the video by the number of ratings, and similarly for the number of comments. The two both have the same distribution, and are very similar to the plot of the number of views in Figure [7](#S4.F7 "Figure 7 ‣ IV-C File Size and Bit-rate ‣ IV Characteristics of YouTube Video ‣ Understanding the Characteristics of Internet Short Video Sharing: YouTube as a Case Study"), yet the tails of the two do not drop so quickly compared to that of number of views.
[Refer to caption](/html/0707.3670/assets/x8.png)
Figure 8: Growth Trend Factor
[Refer to caption](/html/0707.3670/assets/x9.png)
Figure 9: Estimated Active Life Span (t=10%)
[Refer to caption](/html/0707.3670/assets/x10.png)
Figure 10: Rank of User Friends
###
IV-F
Growth Trend of Number of Views and Active Life Span
Comparing the popularity of YouTube videos, we find that some are very popular (their number of views grows very fast), while others are not. Also, after a certain period of time, some videos are almost never watched.
Starting on March 5th, 2007, we updated the number of views statistic of relatively new videos (uploaded after February 15th, 2007) every week for seven weeks. To be sure the growth trend will be properly modelled, we eliminate any videos that have been removed and so do not have the full seven data points, resulting in a dataset size totaling approximately 43 thousand videos.
We have found that the growth trend can be modeled better by a power law than a linear fit. Therefore, a video’s growth trend can be increasing (if the power is greater than 1), growing relatively constantly (power near 1), or slowing in growth (power less than 1). The trend depends on the exponent factor used in the power law, which we call the growth trend factor $`p`$. We define the views count after $`x`$ weeks as
<table>
<tbody>
<tr>
<td></td>
<td>v(x)=v0×(x+μ)pμp𝑣𝑥subscript𝑣0superscript𝑥𝜇𝑝superscript𝜇𝑝v(x)=v_{0}\times\frac{(x+\mu)^{p}}{\mu^{p}}</td>
<td></td>
<td>(1)</td>
</tr>
</tbody>
</table>
where $`\mu`$ is the number of weeks before March 5th that the video has been uploaded, $`x`$ indicates the week of the crawled data (from 0 to 6), and $`v_{0}`$ is the number of views the video had on March 5th.
We modelled the 43 thousand videos using equation [1](#S4.E1 "In IV-F Growth Trend of Number of Views and Active Life Span ‣ IV Characteristics of YouTube Video ‣ Understanding the Characteristics of Internet Short Video Sharing: YouTube as a Case Study") to get the distribution of growth trend factors $`p`$, which is shown in Figure [10](#S4.F10 "Figure 10 ‣ IV-E Views, Ratings – User Access Pattern ‣ IV Characteristics of YouTube Video ‣ Understanding the Characteristics of Internet Short Video Sharing: YouTube as a Case Study"). Over 70% of the videos have a growth trend factor that is less than 1, indicating that most videos grow in popularity more slowly as time passes.
Since YouTube has no policy on removing videos after a period of time or when their popularity declines, the life span of a YouTube video is almost infinite. However, when the video’s popularity grows more and more slowly, the popularity growth curve will become horizontal. Since it will almost stop growing after some time, we will define that as the video’s active life span. From this active life span, we can extract some characteristics of the temporal locality of videos.
If a video’s number of views increases by a factor less than $`t`$ from the previous week, we define the video’s active life span to be over. We prefer this relative comparison to an absolute comparison, since we are only concerned with the shape of the curve instead of the scale.
For each video that has a growth trend factor $`p`$ less than 1, we can compute its active life span $`l`$ from
<table>
<tbody>
<tr>
<td></td>
<td>v(l)v(l−1)−1=t𝑣𝑙𝑣𝑙11𝑡\frac{v(l)}{v(l-1)}-1=t</td>
<td></td>
</tr>
</tbody>
</table>
which can be solved for the active life span
<table>
<tbody>
<tr>
<td></td>
<td>l=11+tp−1+1−μ𝑙1𝑝1𝑡11𝜇l=\frac{1}{\sqrt[p]{1+t}-1}+1-\mu</td>
<td></td>
<td>(2)</td>
</tr>
</tbody>
</table>
Thus we see that the active life span is dependent on the growth trend factor $`p`$ and the number of weeks the video has been on YouTube, but does not depend on the number of views the video had at the start of the experiment.
Figure [10](#S4.F10 "Figure 10 ‣ IV-E Views, Ratings – User Access Pattern ‣ IV Characteristics of YouTube Video ‣ Understanding the Characteristics of Internet Short Video Sharing: YouTube as a Case Study") shows the probability density function (PDF) for the active life span of the approximately 30 thousand videos (with $`p`$ less than 1), for a life span factor of $`t = {10\%}`$. The solid line is the Pareto distribution fit to the data, which fits very well, and results in a parameter $`k`$ of 1.06. From looking at multiple fits with various values of $`t`$, we find that they all result in the same parameter $`k`$, the only difference is the location of the line.
Since we do not have the server logs of YouTube, we cannot accurately measure the characteristic of temporal locality, which would show whether recently accessed videos are likely to be accessed in the near future. However, the active life span gives us another way to view the temporal locality of YouTube videos. Figure [10](#S4.F10 "Figure 10 ‣ IV-E Views, Ratings – User Access Pattern ‣ IV Characteristics of YouTube Video ‣ Understanding the Characteristics of Internet Short Video Sharing: YouTube as a Case Study") implies that most videos have a short active life span, which means the videos have been watched frequently in a short span of time. Then, after the video’s active life span is complete, fewer and fewer people will access them.
This characteristic has good implications for web caching and server storage. We can design a predictor to predict the active life span using our active life span model from equation [2](#S4.E2 "In IV-F Growth Trend of Number of Views and Active Life Span ‣ IV Characteristics of YouTube Video ‣ Understanding the Characteristics of Internet Short Video Sharing: YouTube as a Case Study"). The predictor can help a proxy or server to make more intelligent decisions, such as when to drop a video from the cache. We will discuss this in more detail in Section [VI](#S6 "VI Further Discussions ‣ Understanding the Characteristics of Internet Short Video Sharing: YouTube as a Case Study").
## V The Social Network in YouTube
YouTube is a prominent social media application: there are communities and groups in YouTube, there are statistics and awards for videos and personal channels. Videos are no longer independent from each other, and neither are the users. It is therefore important to understand the social network characteristics of YouTube. We next examine the social network among YouTube users and videos, which is a very unique and interesting aspect of this kind of video sharing sites, as compared to traditional media services.
###
V-A
User Friends and Upload
We have examined the relations among the YouTube users. From the crawl of user information we performed on May 28th 2007, we can extract two characteristics of YouTube users: the number of friends and the number of uploaded videos. We did this for the more than 1 million users found by our crawler in all the crawls performed before this one.
Figure [10](#S4.F10 "Figure 10 ‣ IV-E Views, Ratings – User Access Pattern ‣ IV Characteristics of YouTube Video ‣ Understanding the Characteristics of Internet Short Video Sharing: YouTube as a Case Study") shows the number of friends each user has, compared with the rank of the user by the number of friends. Compared with previous plots, it is much closer to linear on a log-log scale, though we still use the same three distributions to get the best fit. Interestingly, in over 1 million users’ data, we found that 58% of the user’s have no friends. We believe that this is partially because YouTube is still quite young, with more connections to be established between its users.
We also plotted the number of uploaded videos each user has, compared with the rank of the user by number of uploads. It is very similar to the previous plots of the number of views and friends, and so we omit it for brevity.
###
V-B
Small-World Networks
[Refer to caption](/html/0707.3670/assets/x11.png)
Figure 11: Dataset Size
[Refer to caption](/html/0707.3670/assets/x12.png)
Figure 12: Clustering Coefficient
[Refer to caption](/html/0707.3670/assets/x13.png)
Figure 13: Characteristic Path Length
Small-world network phenomenon is probably the most interesting characteristic for social networks. It has been found in various real-world situations: URL links in the Web \[[13](#bib.bib13)\], Gnutella’s search overlay topology \[[14](#bib.bib14)\], and Freenet’s file distribution network \[[15](#bib.bib15)\].
The concept of a small-world was first introduced by Milgram \[[16](#bib.bib16)\] to refer to the the principle that people are linked to all others by short chains of acquaintances (popularly known as *six degrees of separation*). This formulation was used by Watts and Strogatz to describe networks that are neither completely random, nor completely regular, but possess characteristics of both \[[17](#bib.bib17)\]. They introduce a measure of one of these characteristics, the cliquishness of a typical neighborhood, as the *clustering coefficient* of the graph. They define a small-world graph as one in which the clustering coefficient is still large, as in regular graphs, but the measure of the average distance between nodes (the *characteristic path length*) is small, as in random graphs.
Given the network as a graph $`G = {(V,E)}`$, the clustering coefficient $`C_{i}`$ of a node $`i \in V`$ is the proportion of all the possible edges between neighbors of the node that actually exist in the graph. The clustering coefficient of the graph $`C\hspace{0pt}{(G)}`$ is then the average of the clustering coefficients of all nodes in the graph. The characteristic path length $`d_{i}`$ of a node $`i \in V`$ is the average of the minimum number of hops it takes to reach all other nodes in $`V`$ from node $`i`$. The characteristic path length of the graph $`D\hspace{0pt}{(G)}`$ is then the average of the characteristic path lengths of all nodes in the graph.
###
V-C
The Small-World in YouTube
We measured the graph topology for all the YouTube data gathered, by using the related links in YouTube pages to form directed edges in a video graph for each dataset. Videos that have no outgoing or no incoming links are removed from the analysis. In addition, a combined dataset consisting of all the crawled data integrated into one set is also created. Since not all of YouTube is crawled, the resulting graphs are not strongly connected, making it difficult to calculate the characteristic path length. Therefore, we also use the Largest strongly Connected Component (LCC) of each graph for the measurements. Every crawled dataset therefore results in 2 graphs, plus 2 more graphs for the combined dataset.
For comparison, we also generate random graphs that are strongly connected. Each of the random graphs has the same number of nodes and average node degree of the strongly connected component of the crawled dataset, and is also limited to a maximum node out-degree of 20, similar to the crawled datasets. The only exception is the combined dataset of all the crawled data, which was too large to generate a comparable random graph for.
Some graphs use the dataset size for the x-axis values, so that we can see trends as the dataset size increases. This is very informative, as we are not mapping the entire YouTube website, but only a portion of it. Therefore, some extrapolation as the dataset size increases will be needed to draw insights into the graph formed by all of the YouTube videos.
Figure [13](#S5.F13 "Figure 13 ‣ V-B Small-World Networks ‣ V The Social Network in YouTube ‣ Understanding the Characteristics of Internet Short Video Sharing: YouTube as a Case Study") shows the dataset sizes and the date they were created on. It also has the strongly connected component size and the random graph size, both of which are very close to the total dataset size for the larger datasets. The combined dataset is also shown, and is given the most recent date. By far the largest crawled dataset is the first one, crawled on Feb 22.
Figure [13](#S5.F13 "Figure 13 ‣ V-B Small-World Networks ‣ V The Social Network in YouTube ‣ Understanding the Characteristics of Internet Short Video Sharing: YouTube as a Case Study") shows the average clustering coefficient for the entire graph, as a function of the size of the dataset. The clustering coefficient is quite high in most cases, especially in comparison to the random graphs. There is a noticeable drop in the clustering coefficient for the largest datasets, showing that there is some inverse dependence on the size of the graph, which is common for some small-world networks \[[18](#bib.bib18)\].
Figure [13](#S5.F13 "Figure 13 ‣ V-B Small-World Networks ‣ V The Social Network in YouTube ‣ Understanding the Characteristics of Internet Short Video Sharing: YouTube as a Case Study") shows the characteristic path length for each of the datasets’ graphs. There are two factors influencing the shape of the graph. As the dataset size increases, the maximum possible diameter increases, which is seen in the smallest datasets. Once the dataset reaches a size of a few thousand nodes, the diameter starts to decrease as the small-world nature of the graph becomes evident. For the largest datasets, the average diameter is only slightly larger than the diameter of a random graph, which is quite good considering the still large clustering coefficient of these datasets.
The network formed by YouTube’s related videos list has definite small-world characteristics. The clustering coefficient is very large compared to a similar sized random graph, while the characteristic path length of the larger datasets are approaching the short path lengths measured in the random graphs. This finding is expected, due to the user-generated nature of the tags, title and description of the videos that is used by YouTube to find related ones.
These results are similar to other real-world user-generated graphs that exist, yet their parameters can be quite different. For example, the graph formed by URL links in the world wide web exhibits a much longer characteristic path length of 18.59 \[[13](#bib.bib13)\]. This could possibly be due to the larger number of nodes ($`8 \times 10^{8}`$ in the web), but it may also indicate that the YouTube network of videos is a much closer group.
[Refer to caption](/html/0707.3670/assets/x14.png)
Figure 14: Cache Size to Hit-ratio
[Refer to caption](/html/0707.3670/assets/x15.png)
Figure 15: Accuracy to Update Threshold
[Refer to caption](/html/0707.3670/assets/x16.png)
Figure 16: Views to Neighbor Views Correlation
## VI Further Discussions
A very recent study shows that YouTube alone has comprised approximately 20% of all HTTP traffic, or nearly 10% of all traffic on the Internet, with a nearly 20% growth rate per month \[[19](#bib.bib19), [20](#bib.bib20)\]. Assuming the network traffic cost is \$10/Mbps, the estimated YouTube transit expenses is currently more than \$2 million per month. This high and rising expense for network traffic is probably one of the reasons YouTube was sold to Google.
According to Alexa \[[21](#bib.bib21)\], the current speed of YouTube has become “Very Slow” and is considered slower than 81% of the surveyed sites. This situation is only getting worse. Scalability is no doubt the biggest challenge that YouTube faces, particularly considering that websites such as YouTube survive by attracting more users. In this section, we briefly discuss the implications of our measurement results toward improving the scalability of YouTube.
###
VI-A
Implications on Proxy Caching and Storage Management
Caching frequently used data at proxies close to clients is an effective way to save backbone bandwidth and prevent users from experiencing excessive access delays. Numerous algorithms have been developed for caching web objects or streaming videos. While we believe that YouTube will benefit from proxy caching \[[22](#bib.bib22)\], three distinct features call for novel cache designs. First, the number of YouTube videos (42.5 million \[[3](#bib.bib3)\]) is orders of magnitude higher than that of traditional video streams (e.g. HPC: 2999, HPL: 412 \[[11](#bib.bib11)\]). The size of YouTube videos is also much smaller than a traditional video (98.8% are less than 30MB in YouTube versus a typical MPEG-1 movie of 700MB). Finally, the view frequencies of YouTube videos do not well fit a Zipf distribution, which has important implications on web caching \[[23](#bib.bib23)\].
Considering these factors, full-object caching for web or segment caching for streaming video are not practical solutions for YouTube. Prefix caching \[[24](#bib.bib24)\] is probably the best choice. Assume for each video, the proxy will cache a 5 second initial clip, i.e. about 200KB of the video. Given the Gamma distribution of view frequency suggested by our measurements, we plot the hit-ratio as a function of the cache size in Figure [16](#S5.F16 "Figure 16 ‣ V-C The Small-World in YouTube ‣ V The Social Network in YouTube ‣ Understanding the Characteristics of Internet Short Video Sharing: YouTube as a Case Study"), assuming that the cache space is devoted only to the most popular videos. To achieve a 60% hit-ratio, the proxy would require about 1 GByte of disk space for the current YouTube video repository, and nearly 8 GByte for a 95% hit-ratio. Such demand on disk space is acceptable for today’s proxy servers.
Given the constant evolution of YouTube’s video repository, a remaining critical issue is when to release the space for a cached prefix. We found in Section [](#S4.SS6 "IV-F Growth Trend of Number of Views and Active Life Span ‣ IV Characteristics of YouTube Video ‣ Understanding the Characteristics of Internet Short Video Sharing: YouTube as a Case Study")
IV-F
that the active life span of YouTube videos follows a Pareto distribution, implying that most videos are popular during a relatively short span of time. Therefore, a predictor can be developed to forecast the active life span of a video. With the predictor, the proxy can decide which videos have already passed their life span, and replace it if the cache space is insufficient.
The life span predictor can also facilitate disk space management on the YouTube server. Currently, videos on a YouTube server will not be removed by the operator unless they violate the terms of service. With a daily 65,000 new videos introduced, the server storage will soon become a problem. A hierarchical storage structure can be built with videos passing their active life span being moved to slower and cheaper storage media. From our 30 thousand videos dataset (Section [](#S4.SS6 "IV-F Growth Trend of Number of Views and Active Life Span ‣ IV Characteristics of YouTube Video ‣ Understanding the Characteristics of Internet Short Video Sharing: YouTube as a Case Study")
IV-F
), we calculate the predictor accuracy from the number of videos that have an active life span (according to equation [2](#S4.E2 "In IV-F Growth Trend of Number of Views and Active Life Span ‣ IV Characteristics of YouTube Video ‣ Understanding the Characteristics of Internet Short Video Sharing: YouTube as a Case Study")) less than an update threshold divided by the total number of videos, which is plotted in Figure [16](#S5.F16 "Figure 16 ‣ V-C The Small-World in YouTube ‣ V The Social Network in YouTube ‣ Understanding the Characteristics of Internet Short Video Sharing: YouTube as a Case Study"). This result facilitates the determination of an update threshold for the predictor with a given accuracy.
The cache efficiency can be further improved by exploring the small-world characteristic of the related video links (see Section [](#S5.SS3 "V-C The Small-World in YouTube ‣ V The Social Network in YouTube ‣ Understanding the Characteristics of Internet Short Video Sharing: YouTube as a Case Study")
V-C
). That is, if a group of videos have a tight relation, then a user is likely to watch another video in the group after finishing the first one. This expectation is confirmed by Figure [16](#S5.F16 "Figure 16 ‣ V-C The Small-World in YouTube ‣ V The Social Network in YouTube ‣ Understanding the Characteristics of Internet Short Video Sharing: YouTube as a Case Study"), which shows a clear correlation between the number of views and the average of the neighbors’ number of views. Once a video is played and cached, the prefixes of its directly related videos can also be prefetched and cached, if the cache space allows. We have evaluated the effectiveness of this prefetching strategy, which shows that the resultant hit-ratio is almost the same as that of always caching the most popular videos, and yet its communication overhead is significantly lower because it does not have to keep track of the most popular videos list.
###
VI-B
Can Peer-to-Peer Save YouTube?
Short video sharing and peer-to-peer streaming have been widely cited as two key driving forces to Internet video distribution, yet their development remains largely separated. The peer-to-peer technology has been quite successful in supporting large-scale live video streaming (e.g. TV programs like PPLive and CoolStreaming) and even on-demand streaming (e.g. GridMedia). Since each peer contributes its bandwidth to serve others, a peer-to-peer overlay scales extremely well with larger user bases. YouTube and similar sites still use the traditional client-server architecture, restricting their scalability.
Unfortunately, our YouTube measurement results suggest that using peer-to-peer delivery for YouTube could be quite challenging. In particular, the length of a YouTube video is quite short (many are shorter than the typical connection time in a peer-to-peer overlay), and a user often quickly loads another video when finishing a previous one, so the overlay will suffer from an extremely high churn rate. Moreover, there are a huge number of videos, so the peer-to-peer overlays will appear very small.
<sup>1</sup>
<sup>1</sup>1A very recent study on MSN Video \[[9](#bib.bib9)\] has suggested a *peer-assisted VoD*. We notice however that the statistics for MSN Video are quite different from YouTube, and the technique has yet to be substantially revised for YouTube.
Our social network finding again could be exploited by considering a group of related videos as a single large video, with each video in the group being a portion of the large one. Therefore the overlay would be much larger and more stable. Although a user may only watch one video from the group, it can download the other portions of the large video from the server when there is enough bandwidth and space, and upload those downloaded portions to other clients who are interested in them. This behavior can significantly reduce the bandwidth consumption from the server and greatly increase the scalability of the system.
Finally, note that another benefit of using a peer-to-peer model is to avoid single-point of failures and enhance data availability. While this is in general attractive, it is worth noting that timely removing of videos that violate the terms of use (e.g., copyright-protected or illegal content, referred to by the “Removed” category in Section [](#S4.SS1 "IV-A Video Category ‣ IV Characteristics of YouTube Video ‣ Understanding the Characteristics of Internet Short Video Sharing: YouTube as a Case Study")
IV-A
) have constantly been one of the most annoying issues for YouTube and similar sites. Peer-to-peer delivery will clearly make the situation even worse, which must be well-addressed before we shift such sites to the peer-to-peer communication paradigm.
## VII Conclusion
This paper has presented a detailed investigation of the characteristics of YouTube, the most popular Internet short video sharing site to date. Through examining the massive amounts of data collected in a 3-month period, we have demonstrated that, while sharing certain similar features with traditional video repositories, YouTube exhibits many unique characteristics, especially in length distribution, access pattern, and growth trend. These characteristics introduce novel challenges and opportunities for optimizing the performance of short video sharing services.
We have also investigated the social network among YouTube videos, which is probably its most unique and interesting aspect, and has substantially contributed to the success of this new generation of service. We have found that the networks of related videos, which are chosen based on user-generated content, have both small-world characteristics of a short characteristic path length linking any two videos, and a large clustering coefficient indicating the grouping of videos. We have suggested that these features can be exploited to facilitate the design of novel caching or peer-to-peer strategies for short video sharing.
## References
- \[1\]
YouTube serves up 100 million videos a day online. \[Online\]. Available: http://www.usatoday.com/tech/news/2006-07-16-youtube-views_x.htm
- \[2\]
YouTube video-sharing site is changing popular culture. \[Online\]. Available: http://www.kcrw.com/news/programs/ww/ww061122youtube_video-sharin
- \[3\] W. Tang, Y. Fu, L. Cherkasova, and A. Vahdat, “Long-term Streaming Media Server Workload Analysis and Modeling,” HP Labs, Tech. Rep., 2003.
- \[4\] J. M. Almeida, J. Krueger, D. L. Eager, and M. K. Vernon, “Analysis of Educational Media Server Workloads,” in *Proc. NOSSDAV’01*.
- \[5\] S. Acharya, B. Smith, and P. Parnes, “Characterizing User Access To Videos On The World Wide Web,” in *Proc. SPIE’00*.
- \[6\] H. Yu, D. Zheng, B. Y. Zhao, and W. Zheng, “Understanding User Behavior in Large-Scale Video-on-Demand Systems,” in *EuroSys’06*.
- \[7\] M. Chesire, A. Wolman, G. M. Voelker, and H. M. Levy, “Measurement and Analysis of a Streaming-Media Workload,” in *Proc. USITS’01*.
- \[8\] E. Veloso, V. Almeida, W. Meira, A. Bestavros, and S. Jin, “A Hierarchical Characterization of a Live Streaming Media Workload,” in *Proc. SIGCOMM’02*.
- \[9\] C. Huang, J. Li, and K. W. Ross, “Can Internet Video-on-Demand be Profitable?” in *SIGCOMM’07*, 2007.
- \[10\]
YouTube: Video Format (from Wikipedia). \[Online\]. Available: http://en.wikipedia.org/wiki/Youtube#Video_format
- \[11\]
Google frames a video search engine. \[Online\]. Available: http://googlesystem.blogspot.com/2007/06/google-videos-new-frame.html
- \[12\]
YouTube Blog. \[Online\]. Available: http://youtube.com/blog
- \[13\] R. Albert, H. Jeong, and A. Barabasi, “The Diameter of the World Wide Web,” *Nature*, vol. 401, pp. 130–131, 1999.
- \[14\] G. Liu, M. Hu, B. Fang, and H. Zhang, “Measurement and Modeling of Large-Scale Peer-to-Peer Storage System,” *Lecture Notes in Computer Science*, pp. 270–277, 2004.
- \[15\] T. Hong, “Performance,” in *Peer-to-Peer: Harnessing the Power of Disruptive Technologies*, A. Oram, Ed. Sebastopol, CA, USA: O’Reilly & Associates, Inc., 2001, ch. 14, pp. 203–241.
- \[16\] S. Milgram, “The Small World Problem,” *Psychology Today*, vol. 2, no. 1, pp. 60–67, 1967.
- \[17\] D. Watts and S. Strogatz, “Collective Dynamics of ‘Small-World’ Networks,” *Nature*, vol. 393, no. 6684, pp. 409–10, 1998.
- \[18\] E. Ravasz and A.-L. Barabási, “Hierarchical organization in complex networks,” *Physical Review E*, vol. 67, no. 2, p. 026112, Feb 2003.
- \[19\]
Ellacoya data shows web traffic overtakes Peer-to-Peer (P2P) as largest percentage of bandwidth on the network. \[Online\]. Available: http://www.ellacoya.com/news/pdf/2007/NXTcommEllacoyaMediaAlert.pdf
- \[20\] C. Corbett, *Peering of Video*, 2006.
- \[21\]
Alexa. \[Online\]. Available: http://www.alexa.com
- \[22\] J. Liu and J. Xu, “Proxy Caching for Media Streaming over the Internet,” *IEEE Communications Magazine*, vol. 42, no. 8, pp. 88–94, 2004.
- \[23\] L. Breslau, P. Cao, L. Fan, G. Phillips, and S. Shenker, “Web Caching and Zipf-like Distributions: Evidence and Implications,” in *INFOCOM’99*.
- \[24\] S. Sen, J. Rexford, and D. F. Towsley, “Proxy Prefix Caching for Multimedia Streams,” in *INFOCOM’99*. |
0706.2888 | cs.CR | cyber | # Variations on Kak’s Three Stage Quantum Cryptography Protocol
James Harold Thomas
Department of Electrical and Computer Engineering
Louisiana State University
Baton Rouge, Louisiana, 70803, USA
###### Abstract.
This paper introduces a variation on Kak’s three-stage quanutm key distribution protocol which allows for defence against the man in the middle attack. In addition, we introduce a new protocol, which also offers similar resiliance against such an attack.
## 1. Introduction
In the relatively new field of quantum information science, there have been limited numbers of true breakthroughs. The quantum computer remains a machine on paper only, for a working realization has proved elusive. For one of the most powerful tools that the quantum computer would provide, namely the reduction of the factorization problem to polynomial time by Shor’s algorithm, the most successful implementation to date has been to factor the number 15 into 3 times 5 \[1\].
The quantum computer might remain unrealized for years to come. Indeed, Kak has suggested that the current quantum circuit model for quantum computers is fundamentally flawed and new models must be developed to tackle the problem \[2\].
While there have been limited strides in some areas of quantum information science, one area in particular has produced realizable solutions in the field of cryptography. Quantum key distribution protocols have been successfully implemented and have produced commercially available products.
In this paper, we will discuss a new protocol proposed by Kak called the “Three Stage Protocol.” To Kak’s protocol, we will introduce a modification which allows for greater security against man in the middle attacks. In addition, we introduce a new single stage protocol which similarly allows for security against such attacks.
## 2. Quantum Key Distribution
The usefulness of quantum key distribution lies in the properties of the qubit, the quantum unit of information. Since a qubit is an object representing a quantum superposition state, the qubit cannot be copied. This is commonly called the no-cloning theorem \[3\]. This property ensures that during qubit data transmission, it is impossible for an evesedropper (Eve) to simply make copies of the qubits being sent, and thus manipulate these copied qubits to obtain the message. This useful property allows quantum data transmission to be used effectively in key distribution protocols as shown in \[3\].
When a private key can be transmitted securly along a quantum channel, then secure classical communication between the two parties (Alice and Bob) can be achieved using the private key (Figure 1). One private key cryptosystem in use today is the Vernam cipher, also called the one time pad. According to Nielsen and Chuang, the security of the private key used in the Vernam cipher is sometimes ensured by transmitting it via such low-tech solutions as clandestine meetings or trusted couriers. The need for better transmission protocols is obvious.
[Refer to caption](/html/0706.2888/assets/x1.png)
Figure 1. Quantum Key Distribution
## 3. Kak’s Three-Stage Protocol
In \[4\], S. Kak proposed a new quantum key distribution protocol based on secret unitary transformations (Figure 2). His protocol, like BB84, has three stages, but unlike BB84, it remains quantum across all three stages. In the first stage, Alice manipulates the message $`X`$, which is simply one of two orthogonal stages (e.g. $`{\alpha\hspace{0pt}{|0\rangle}} + {\beta\hspace{0pt}{|1\rangle}}`$ and $`{\beta\hspace{0pt}{|0\rangle}} - {\alpha\hspace{0pt}{|1\rangle}}`$) by means of a unitary transformation $`U_{A}`$, known only to her. Bob receives the new state, and in the second stage, applies his own secret transformation $`U_{B}`$, which is both a unitary transformation, and one that commutes with $`U_{A}`$, and sends the result back to Alice. In the third stage, Alice applies the Hermitian conjugate of her transformation, $`U_{A}^{\dagger}`$, and sends the result back to Bob. Since $`{U_{A}^{\dagger}\hspace{0pt}U_{B}\hspace{0pt}U_{A}\hspace{0pt}{(X)}} = {U_{B}\hspace{0pt}{(X)}}`$, Bob simply applies $`U_{B}^{\dagger}`$ and obtains the previously unknown state, $`X`$.
[Refer to caption](/html/0706.2888/assets/x2.png)
Figure 2. Kak’s Three-Stage Protocol
## 4. The Man in the Middle Attack
The Suceptability of both BB84 and Kak’s three-stage protocol to man in the middle attacks has been documented \[e.g. 5,6\], and various methods to counter these attacks have been proposed \[e.g. 7,8\]. In such an attack, the eavesdropper, Eve, can attempt to thwart the communication between Alice and Bob in one of the following ways (Figure 3).
- \(1\)
Eve receives the message from Alice by impersonating Bob. Eve then decodes Alice’s message, and, now impersonating Alice, duplicates this message to Bob. In this scenerio, both Eve and Bob obtain the secret message.
- \(2\)
Eve impersonates Bob and decodes Alice’s message as in scenerio 1, but instead of relaying the actual message to Bob, Eve relays a different message of her own choosing. In this scenerio, only Eve obtains the secret message.
- \(3\)
Eve impersonates Bob, but is not able to decode Alice’s message. Instead, she impersonates Alice and sends her own message to Bob. In this scenerio, communication between Alice and Bob is blocked, but no secret message is comprimised.
[Refer to caption](/html/0706.2888/assets/x3.png)
Figure 3. Man in the Middle Attack
## 5. A Variation on Kak’s Three Stage Protocol
In Kak’s paper \[4\], he suggests using secret real valued orthogonal transformations to encrypt the qubits. Under orthogonal transformations of the same form (see below), the selection of the angles $`\theta`$ and $`\phi`$ by Alice and Bob respectively does not affect the outcome of the protocol. Furthermore, both Alice and Bob do not need to know what each other’s angle selection is.
The reason that a man in the middle attack can be carried out is that when $`U_{A}`$ is assumed to be real valued, it is very easy for Eve to find another unitary transformation, $`U_{E}`$, which commutes with $`U_{A}`$. This is the underlying assumption by both Perkins \[5\], and Basuchowdhuri \[7\]. Indeed, for a 2x2 real valued unitary transformation (i.e. an orthogonal transformation), there is a limitation on its form.
Consider a 2x2 transformation:
<table>
<tbody>
<tr>
<td></td>
<td>𝐔𝐀=[wxyz],w,x,y,z∈ℜformulae-sequencesubscript𝐔𝐀matrix𝑤𝑥𝑦𝑧𝑤𝑥𝑦𝑧\mathbf{U_{A}}=\begin{bmatrix}w&x\\ y&z\end{bmatrix},w,x,y,z\in\Re</td>
<td></td>
</tr>
</tbody>
</table>
Then, for $`U_{A}`$ to be orthogonal (unitary), $`{U_{A}\hspace{0pt}U_{A}} = I`$. This gives rise to the following equations:
<table>
<tbody>
<tr>
<td></td>
<td>{w2+y2=1x2+z2=1wx+yz=0casessuperscript𝑤2superscript𝑦21otherwisesuperscript𝑥2superscript𝑧21otherwise𝑤𝑥𝑦𝑧0otherwise\begin{cases}w^{2}+y^{2}=1\\ x^{2}+z^{2}=1\\ wx+yz=0\end{cases}</td>
<td></td>
</tr>
</tbody>
</table>
These equations are satisfied only when $`U_{A}`$ has one of the two following forms:
<table>
<tbody>
<tr>
<td></td>
<td>𝐔𝟏(θ)=[cos(θ)−sin(θ)sin(θ)cos(θ)],subscript𝐔1𝜃matrix𝑐𝑜𝑠𝜃𝑠𝑖𝑛𝜃𝑠𝑖𝑛𝜃𝑐𝑜𝑠𝜃\mathbf{U_{1}(\theta)}=\begin{bmatrix}cos(\theta)&-sin(\theta)\\ sin(\theta)&cos(\theta)\end{bmatrix},</td>
<td></td>
</tr>
</tbody>
</table>
a rotation, as Kak proposed, or
<table>
<tbody>
<tr>
<td></td>
<td>𝐔𝟐(θ)=[cos(θ)sin(θ)sin(θ)−cos(θ)],subscript𝐔2𝜃matrix𝑐𝑜𝑠𝜃𝑠𝑖𝑛𝜃𝑠𝑖𝑛𝜃𝑐𝑜𝑠𝜃\\ \mathbf{U_{2}(\theta)}=\begin{bmatrix}cos(\theta)&sin(\theta)\\ sin(\theta)&-cos(\theta)\end{bmatrix},</td>
<td></td>
</tr>
</tbody>
</table>
a reflection across the line $`\frac{\theta}{2}`$.
The three-stage protocol demands that while Bob doesn’t know the value of $`\theta`$ that Alice is using, he must know which of the two above forms of $`U_{A}`$ that Alice chooses. The reason for this is that while $`U_{1}\hspace{0pt}{(\theta)}`$ commutes with $`U_{1}\hspace{0pt}{(\phi)}`$ for any $`\theta`$ and $`\phi`$, and $`U_{2}\hspace{0pt}{(\theta)}`$ commutes with $`U_{2}\hspace{0pt}{(\phi)}`$ for any $`\theta`$ and $`\phi`$, $`U_{1}\hspace{0pt}{(\theta)}`$ does not commute with $`U_{2}\hspace{0pt}{(\phi)}`$ in general. So we will consider the choice of the form of $`U_{A}`$ to be public information. Once this information is known, Bob simply needs to choose his own angle $`\phi`$, and his transformation $`U_{B}`$ will be of the same form as $`U_{A}`$.
<table>
<tbody>
<tr>
<td></td>
<td>UB={U1(ϕ),UA=U1(θ)U2(ϕ),UA=U2(θ).subscript𝑈𝐵casessubscript𝑈1italic-ϕsubscript𝑈𝐴subscript𝑈1𝜃otherwisesubscript𝑈2italic-ϕsubscript𝑈𝐴subscript𝑈2𝜃otherwiseU_{B}=\begin{cases}U_{1}(\phi),U_{A}=U_{1}(\theta)\\ U_{2}(\phi),U_{A}=U_{2}(\theta)\end{cases}.</td>
<td></td>
</tr>
</tbody>
</table>
It is easy to see that when the same form is used, $`U_{A}`$ and $`U_{B}`$ commute (i.e. $`{U_{A}\hspace{0pt}U_{B}} = {U_{B}\hspace{0pt}U_{A}}`$).
For Form 1:
<table>
<tbody>
<tr>
<td></td>
<td>UAUB=[cos(θ)−sin(θ)sin(θ)cos(θ)][cos(ϕ)−sin(ϕ)sin(ϕ)cos(ϕ)]=subscript𝑈𝐴subscript𝑈𝐵matrix𝑐𝑜𝑠𝜃𝑠𝑖𝑛𝜃𝑠𝑖𝑛𝜃𝑐𝑜𝑠𝜃matrix𝑐𝑜𝑠italic-ϕ𝑠𝑖𝑛italic-ϕ𝑠𝑖𝑛italic-ϕ𝑐𝑜𝑠italic-ϕabsentU_{A}U_{B}=\begin{bmatrix}cos(\theta)&-sin(\theta)\\ sin(\theta)&cos(\theta)\end{bmatrix}\begin{bmatrix}cos(\phi)&-sin(\phi)\\ sin(\phi)&cos(\phi)\end{bmatrix}=\\</td>
<td></td>
</tr>
</tbody>
</table>
<table>
<tbody>
<tr>
<td></td>
<td>[cos(θ)cos(ϕ)−sin(θ)sin(ϕ)−cos(θ)sin(ϕ)−cos(ϕ)sin(θ)cos(ϕ)sin(θ)+cos(θ)sin(ϕ)−sin(θ)sin(ϕ)+cos(θ)cos(ϕ)]matrix𝑐𝑜𝑠𝜃𝑐𝑜𝑠italic-ϕ𝑠𝑖𝑛𝜃𝑠𝑖𝑛italic-ϕ𝑐𝑜𝑠𝜃𝑠𝑖𝑛italic-ϕ𝑐𝑜𝑠italic-ϕ𝑠𝑖𝑛𝜃𝑐𝑜𝑠italic-ϕ𝑠𝑖𝑛𝜃𝑐𝑜𝑠𝜃𝑠𝑖𝑛italic-ϕ𝑠𝑖𝑛𝜃𝑠𝑖𝑛italic-ϕ𝑐𝑜𝑠𝜃𝑐𝑜𝑠italic-ϕ\begin{bmatrix}cos(\theta)cos(\phi)-sin(\theta)sin(\phi)&-cos(\theta)sin(\phi)-cos(\phi)sin(\theta)\\ cos(\phi)sin(\theta)+cos(\theta)sin(\phi)&-sin(\theta)sin(\phi)+cos(\theta)cos(\phi)\end{bmatrix}</td>
<td></td>
</tr>
</tbody>
</table>
Applying trigonometric identities,
<table>
<tbody>
<tr>
<td></td>
<td>UAUB=[cos(θ+ϕ)−sin(θ+ϕ)sin(θ+ϕ)cos(θ+ϕ)].subscript𝑈𝐴subscript𝑈𝐵matrix𝑐𝑜𝑠𝜃italic-ϕ𝑠𝑖𝑛𝜃italic-ϕ𝑠𝑖𝑛𝜃italic-ϕ𝑐𝑜𝑠𝜃italic-ϕU_{A}U_{B}=\begin{bmatrix}cos(\theta+\phi)&-sin(\theta+\phi)\\ sin(\theta+\phi)&cos(\theta+\phi)\end{bmatrix}.</td>
<td></td>
</tr>
</tbody>
</table>
Since $`U_{A}`$ and $`U_{B}`$ have the same form, it is clear that they commute.
For form 2:
<table>
<tbody>
<tr>
<td></td>
<td>UAUB=[cos(θ)sin(θ)sin(θ)−cos(θ)][cos(ϕ)sin(ϕ)sin(ϕ)−cos(ϕ)]=subscript𝑈𝐴subscript𝑈𝐵matrix𝑐𝑜𝑠𝜃𝑠𝑖𝑛𝜃𝑠𝑖𝑛𝜃𝑐𝑜𝑠𝜃matrix𝑐𝑜𝑠italic-ϕ𝑠𝑖𝑛italic-ϕ𝑠𝑖𝑛italic-ϕ𝑐𝑜𝑠italic-ϕabsentU_{A}U_{B}=\begin{bmatrix}cos(\theta)&sin(\theta)\\ sin(\theta)&-cos(\theta)\end{bmatrix}\begin{bmatrix}cos(\phi)&sin(\phi)\\ sin(\phi)&-cos(\phi)\end{bmatrix}=\\</td>
<td></td>
</tr>
</tbody>
</table>
<table>
<tbody>
<tr>
<td></td>
<td>[cos(θ)cos(ϕ)+sin(θ)sin(ϕ)cos(θ)sin(ϕ)−cos(ϕ)sin(θ)cos(ϕ)sin(θ)−cos(θ)sin(ϕ)sin(θ)sin(ϕ)+cos(θ)cos(ϕ)].matrix𝑐𝑜𝑠𝜃𝑐𝑜𝑠italic-ϕ𝑠𝑖𝑛𝜃𝑠𝑖𝑛italic-ϕ𝑐𝑜𝑠𝜃𝑠𝑖𝑛italic-ϕ𝑐𝑜𝑠italic-ϕ𝑠𝑖𝑛𝜃𝑐𝑜𝑠italic-ϕ𝑠𝑖𝑛𝜃𝑐𝑜𝑠𝜃𝑠𝑖𝑛italic-ϕ𝑠𝑖𝑛𝜃𝑠𝑖𝑛italic-ϕ𝑐𝑜𝑠𝜃𝑐𝑜𝑠italic-ϕ\begin{bmatrix}cos(\theta)cos(\phi)+sin(\theta)sin(\phi)&cos(\theta)sin(\phi)-cos(\phi)sin(\theta)\\ cos(\phi)sin(\theta)-cos(\theta)sin(\phi)&sin(\theta)sin(\phi)+cos(\theta)cos(\phi)\end{bmatrix}.</td>
<td></td>
</tr>
</tbody>
</table>
Again, applying trigonometric identities,
<table>
<tbody>
<tr>
<td></td>
<td>UAUB=[cos(θ+ϕ)sin(ϕ−θ)sin(θ−ϕ)cos(θ+ϕ)]subscript𝑈𝐴subscript𝑈𝐵matrix𝑐𝑜𝑠𝜃italic-ϕ𝑠𝑖𝑛italic-ϕ𝜃𝑠𝑖𝑛𝜃italic-ϕ𝑐𝑜𝑠𝜃italic-ϕU_{A}U_{B}=\begin{bmatrix}cos(\theta+\phi)&sin(\phi-\theta)\\ sin(\theta-\phi)&cos(\theta+\phi)\end{bmatrix}</td>
<td></td>
</tr>
</tbody>
</table>
Since $`U_{A}`$ and $`U_{B}`$ always commute given the same form, then for Eve to impersonate Bob and obtain Alice’s secret message, she only needs to select any angle $`\psi`$ and use it in her own transformation $`U_{E}`$ where $`U_{E}`$ is of the same form as $`U_{A}`$. Using $`U_{E}`$, Eve can obtain the secret state $`X`$ in the exact same way that Bob can obtain it. In addition, Eve can relay a message to Bob using her own $`U_{E}`$ transformation. Since Bob doesn’t know what Alice’s $`U_{A}`$ transformation is, Eve’s $`U_{E}`$ is a valid substitution.
Suppose now instead of a orthogonal (i.e. unitary and real value) transformation, Alice chooses a more general complex valued unitary transformation,
<table>
<tbody>
<tr>
<td></td>
<td>𝐔𝐀(θ)=12[eiθe−iθieiθ−ie−iθ],θ∈[0,2π).formulae-sequencesubscript𝐔𝐀𝜃12matrixsuperscript𝑒𝑖𝜃superscript𝑒𝑖𝜃𝑖superscript𝑒𝑖𝜃𝑖superscript𝑒𝑖𝜃𝜃02𝜋\mathbf{U_{A}(\theta)}=\frac{1}{\sqrt{2}}\begin{bmatrix}e^{i\theta}&e^{-i\theta}\\ ie^{i\theta}&-ie^{-i\theta}\end{bmatrix},\theta\in[0,2\pi).</td>
<td></td>
</tr>
</tbody>
</table>
When Alice chooses a $`U_{A}`$ of this form, it is more difficult for Bob to find another transform, $`U_{B}`$ which commutes. We see that when
<table>
<tbody>
<tr>
<td></td>
<td>𝐔𝐁(ϕ)=12[eiϕe−iϕieiϕ−ie−iϕ],θ∈[0,2π),formulae-sequencesubscript𝐔𝐁italic-ϕ12matrixsuperscript𝑒𝑖italic-ϕsuperscript𝑒𝑖italic-ϕ𝑖superscript𝑒𝑖italic-ϕ𝑖superscript𝑒𝑖italic-ϕ𝜃02𝜋\mathbf{U_{B}(\phi)}=\frac{1}{\sqrt{2}}\begin{bmatrix}e^{i\phi}&e^{-i\phi}\\ ie^{i\phi}&-ie^{-i\phi}\end{bmatrix},\theta\in[0,2\pi),</td>
<td></td>
</tr>
</tbody>
</table>
then
<table>
<tbody>
<tr>
<td></td>
<td>𝐔𝐀𝐔𝐁=[ei(θ+ϕ)+iei(ϕ−θ)ei(θ−ϕ)−ie−i(θ+ϕ)iei(θ+ϕ)+ei(ϕ−θ)iei(θ−ϕ)−e−i(θ+ϕ)]subscript𝐔𝐀subscript𝐔𝐁matrixsuperscript𝑒𝑖𝜃italic-ϕ𝑖superscript𝑒𝑖italic-ϕ𝜃superscript𝑒𝑖𝜃italic-ϕ𝑖superscript𝑒𝑖𝜃italic-ϕ𝑖superscript𝑒𝑖𝜃italic-ϕsuperscript𝑒𝑖italic-ϕ𝜃𝑖superscript𝑒𝑖𝜃italic-ϕsuperscript𝑒𝑖𝜃italic-ϕ\mathbf{U_{A}U_{B}}=\begin{bmatrix}e^{i(\theta+\phi)}+ie^{i(\phi-\theta)}&e^{i(\theta-\phi)}-ie^{-i(\theta+\phi)}\\ ie^{i(\theta+\phi)}+e^{i(\phi-\theta)}&ie^{i(\theta-\phi)}-e^{-i(\theta+\phi)}\end{bmatrix}</td>
<td></td>
</tr>
</tbody>
</table>
and
<table>
<tbody>
<tr>
<td></td>
<td>𝐔𝐁𝐔𝐀=[ei(ϕ+θ)+iei(θ−ϕ)ei(ϕ−θ)−ie−i(ϕ+θ)iei(ϕ+θ)+ei(θ−ϕ)iei(ϕ−θ)−e−i(ϕ+θ)].subscript𝐔𝐁subscript𝐔𝐀matrixsuperscript𝑒𝑖italic-ϕ𝜃𝑖superscript𝑒𝑖𝜃italic-ϕsuperscript𝑒𝑖italic-ϕ𝜃𝑖superscript𝑒𝑖italic-ϕ𝜃𝑖superscript𝑒𝑖italic-ϕ𝜃superscript𝑒𝑖𝜃italic-ϕ𝑖superscript𝑒𝑖italic-ϕ𝜃superscript𝑒𝑖italic-ϕ𝜃\mathbf{U_{B}U_{A}}=\begin{bmatrix}e^{i(\phi+\theta)}+ie^{i(\theta-\phi)}&e^{i(\phi-\theta)}-ie^{-i(\phi+\theta)}\\ ie^{i(\phi+\theta)}+e^{i(\theta-\phi)}&ie^{i(\phi-\theta)}-e^{-i(\phi+\theta)}\end{bmatrix}.</td>
<td></td>
</tr>
</tbody>
</table>
These transformations commute only when $`\phi = {\theta + \pi}`$ (or any 2$`\pi`$ multiple). For Bob to decode Alice’s message, he must have more information than simply the form of her transformation. He must know also the value of $`\theta`$ that she has chosen.
While this might seem to be a hindrence to the protocol, it allows for much greater security against a man in the middle attack. Eve attempts to intercept Alice’s message to Bob by choosing a $`U_{E}`$ to impersonate Bob’s $`U_{B}`$. As we saw earlier, when $`U_{A}`$ is real valued, Eve can simply pick any angle $`\psi`$ and generate a transformation $`U_{E}`$ that commutes. But with a complex valued $`U_{A}`$, Eve cannot guarantee a commuting transformation without knowing the value of $`\theta`$. Consider Eve’s choice of a $`\psi`$ without knowledge of $`\theta`$. Then, $`{U_{A}^{\dagger}\hspace{0pt}U_{E}\hspace{0pt}U_{A}} \neq U_{E}`$.
## 6. The Single Stage Quantum Cryptography Protocol
When, as in the variation to Kak’s protocol described above, Bob knows the value of $`\theta`$ that Alice has chosen for her transformation $`U_{A}`$ (assuming as above that the form of $`U_{A}`$ is public information), then he has full knowledge of $`U_{A}`$. In this situation, Alice and Bob can forego the second two stages of the protocol and let Bob perform the transform $`U_{A}^{\dagger}`$ to obtain the unknown state $`X`$ (Figure 4). We have simply, $`{U_{A}^{\dagger}\hspace{0pt}U_{A}\hspace{0pt}{(X)}} = X`$. In this situation, there is no need for $`U_{A}`$ to be complex valued. We can have, as Kak proposed in \[4\],
<table>
<tbody>
<tr>
<td></td>
<td>UA=[cos(θ)−sin(θ)sin(θ)cos(θ)]subscript𝑈𝐴matrix𝑐𝑜𝑠𝜃𝑠𝑖𝑛𝜃𝑠𝑖𝑛𝜃𝑐𝑜𝑠𝜃U_{A}=\begin{bmatrix}cos(\theta)&-sin(\theta)\\ sin(\theta)&cos(\theta)\end{bmatrix}</td>
<td></td>
</tr>
</tbody>
</table>
So for Eve to intercept the message $`X`$ and properly decode it, she would have to know the value of $`\theta`$. The strength of this protocol is dependent on keeping the value of $`\theta`$ a secret known only to Alice and Bob. We enhance the security of our protocol by allowing for $`\theta`$ to change, which blocks any attempt by Eve at a statistical analysis of the qubits. We assume that before secure transmission may begin, there is some other secure protocol that Alice may use to transmit her initial value of $`\theta`$ to Bob. One example is Perkins’ protocol which uses trusted certificates \[6\].
Suppose we restrict $`\theta`$ to the upper half plane of the unit circle. After $`l`$ qubits are successfully transmitted from Alice to Bob, the qubits $`l + 1`$ to $`l + k`$ will be used to obtain the new value of $`\theta`$. The $`k`$ data bits selected by Alice for these $`k`$ qubits will represent an integer $`N`$ such that if $`b_{n}`$ is the $`n`$th bit transmitted ($`b_{n} \in {\{ 0,1\}}`$), then $`N = {b_{l + 1} + {2\hspace{0pt}b_{l + 2}} + {4\hspace{0pt}b_{l + 3}} + \ldots + {2^{k}\hspace{0pt}b_{l + k}}}`$. When these four qubits are received by Bob and decoded, Alice and Bob adjust their transformations $`U_{A}`$ and $`U_{A}^{\dagger}`$ respectively such that
<table>
<tbody>
<tr>
<td></td>
<td>UA=[cos(θN)−sin(θN)sin(θN)cos(θN)],θN=Nπ2k.formulae-sequencesubscript𝑈𝐴matrix𝑐𝑜𝑠subscript𝜃𝑁𝑠𝑖𝑛subscript𝜃𝑁𝑠𝑖𝑛subscript𝜃𝑁𝑐𝑜𝑠subscript𝜃𝑁subscript𝜃𝑁𝑁𝜋superscript2𝑘U_{A}=\begin{bmatrix}cos(\theta_{N})&-sin(\theta_{N})\\ sin(\theta_{N})&cos(\theta_{N})\end{bmatrix},\theta_{N}=\frac{N\pi}{2^{k}}.</td>
<td></td>
</tr>
</tbody>
</table>
After this, Alice transmits $`l`$ more qubits to Bob before again changing the value of $`\theta`$. In this fashion, any attempt by Eve to obtain the value of $`\theta`$ with no prior knowledge would be extremely difficult.
[Refer to caption](/html/0706.2888/assets/x4.png)
Figure 4. The Single Stage Quantum Cryptography Protocol
[Refer to caption](/html/0706.2888/assets/x5.png)
Figure 5. Framing on the Single Stage Protocol
## Acknowledgements
The author thanks the Louisiana Board of Regents, BoRSF, under agreement NASA/LEQSF(2005-2010)-LaSPACE and NASA/LaSPACE under grant NNG05GH22H for support during this project.
## References
- \[1\] L.M.K. Vandersypen, M. Steffen, G. Breyta, C.S. Yannoni, M.H. Sherwood, I.L. Chuang, Experimental realization of Shor’s quantum factoring algorithm using nuclear magnetic resonance. Nature 414, 883-887 (20 Dec 2001). arXiv:quant-ph/0112176v1
- \[2\] S. Kak, Are Quantum Computing Models Realistic? ACM Ubiquity, 7 (11): 1-9, 2006. arXiv:quant-ph/0110040v5
- \[3\] M.A. Nielsen, I.L. Chuang, Quantum Computation and Quantum Information. Cambridge University Press, 2000.
- \[4\] S. Kak, A Three-Stage Quantum Cryptography Protocol. Foundations of Physics Letters 19 (2006), 293-296. arXiv:quant-ph/0503027v2
- \[5\] G. Gilbert, M. Hamrick (MITRE), Constraints on Eavesdropping on the BB84 Protocol. arXiv:quant-ph/0106034v2
- \[6\] W. Perkins, Trusted Certificates in Quantum Cryptography. arXiv:cs/0603046v1 \[cs.CR\]
- \[7\] K. Svozil, Feasibility of the interlock protocol against man-in-the-middle attacks on quantum cryptography. International Journal of Quantum Information, Vol. 3, No. 4 (2005) 649-654. arXiv:quant-ph/0501062v4
- \[8\] P. Basuchowdhuri, Classical Authentication Aided Three-Stage Quantum Protocol. arXiv:cs/0605083v1 |
0706.4175 | cs.NI | cyber | INSTITUT NATIONAL DE RECHERCHE EN INFORMATIQUE ET EN AUTOMATIQUE
Heuristics for Network Coding in Wireless Networks
Song Yean Cho, Cédric Adjih, Philippe JacquetN° 6238
Juin 2007
Heuristics for Network Coding in Wireless Networks
Song Yean Cho, Cédric Adjih, Philippe Jacquet
Thème COM — Systèmes communicants
Projet HIPERCOM
Rapport de recherche n° 6238 — Juin 2007 — ?? pages
Abstract:
Multicast is a central challenge for emerging multi-hop wireless architectures such as wireless mesh networks, because of its substantial cost in terms of bandwidth.
In this report, we study one specific case of multicast: broadcasting, sending data from one source to all nodes, in a multi-hop wireless network. The broadcast we focus on is based on network coding, a promising avenue for reducing cost; previous work of ours showed that the performance of network coding with simple heuristics is asymptotically optimal: each transmission is beneficial to nearly every receiver. This is for homogenous and large networks of the plan.
But for small, sparse or for inhomogeneous networks, some additional heuristics are required. This report proposes such additional new heuristics (for selecting rates) for broadcasting with network coding. Our heuristics are intended to use only simple local topology information. We detail the logic of the heuristics, and with experimental results, we illustrate the behavior of the heuristics, and demonstrate their excellent performance.
Key-words: wireless, broadcast, network coding, hypergraph, min-cut, heuristics, rate selection, subgraph selection
Heuristiques pour le codage de réseau dans les réseaux sans fil
Résumé : Le multicast est un enjeu central pour les nouvelles architectures de réseaux sans fil, tels que les réseaux maillés “mesh”, à cause de son coût substanciel en termes de bande passante.
Dans ce rapport, nous étudions une méthode de multicast spécificique: la diffusion à tout le réseau, dans un réseau sans fil multi-sauts. Plus précisement, la diffusion que nous considérons est fondée sur l’utilisation du codage de réseau, une méthode prometteuse pour réduire le coût. Nos précédents travaux ont montré que la performance du codage de réseaux avec une heuristique simple était assymptotiquement optimale : chaque transmission est utile à presque tous les destinataires. C’est vrai pour de grands réseaux homogènes du plan Euclidien.
Mais pour des réseaux plus petits, ou inhomogènes, des heuristiques supplémentaires sont requises. Ce rapport propose de telles heuristiques (de choix de débit) pour la diffusion par codage de réseau. Nos heuristiques ont pour objectif d’utiliser seulement des informations simples concernant la topologie locale. Nous détaillons la logique des heuristiques, et par des résultats expérimentaux, nous illustrons le comportement de ces heuristiques, et mettons en évidence leur excellente performance.
Mots-clés : sans fil, diffusion, codage de réseau, hypergraph, coupe minimale, heuristique, sélection de débit
## 1 Introduction
The confirmed success of wireless networks has made wireless communication ubiquitous. One of the predicted use of wireless networks is multicast, which could be used for multimedia content diffusion, video conference, software distribution, and a number of other applications. However, in multi-hop networks such as wireless mesh networks, multicast transmissions incur a substantial cost by simple virtue of requiring relaying over several forwarders, in order to cover every destination. For this reason efficient techniques for multicasting are of prime interest. One of them is the recently proposed method, network coding. Network coding was introduced by the seminal work of \[[1](#bib.bib1)\] as a new paradigm where intermediate nodes mix information from different flows (different bits or different packets). We use network coding specifically for broadcasting rather then general multicasting in wireless multi-hop networks.
The problem that we are addressing is *efficient broadcast*:
- •
Broadcast packets from one source to all nodes, with the minimum number of transmissions.
Without network coding, finding the optimal broadcasting is an NP complete problem \[[2](#bib.bib2)\], but a number of heuristics exist for efficient broadcasting such as MPR-flooding \[[3](#bib.bib3)\] or techniques based on connected dominating sets \[[4](#bib.bib4), [5](#bib.bib5)\]. But with network coding, the optimal broadcasting can be found in polynomial time. Finding an optimal solution (subgraph selection) consists in finding the coding nodes and their optimal rates \[[6](#bib.bib6), [7](#bib.bib7), [8](#bib.bib8)\]. This problem can be formulated as a linear program, which can be solved in polynomial time \[[11](#bib.bib11), [10](#bib.bib10)\], and possibly in a distributed fashion \[[10](#bib.bib10)\].
However, we adopt a different, even simpler, approach: previous work \[[12](#bib.bib12), [13](#bib.bib13)\] has shown that a simple heuristic for selecting rates could achieve asymptotically the optimal efficiency for homogeneous large and dense wireless networks of the plane — and also that, noticeably, it would outperform methods without network coding. This is asymptotically true for homogeneous networks, but the heuristic needs adjustments for less homogeneous, smaller or sparser networks. The adjustment is the topic of this report. Our key contributions are the following:
- •
We propose an improved heuristic for rate selection, inspired by \[[14](#bib.bib14)\]. It requires only local topology information: knowledge of two-hop neighbors.
- •
We empirically study its performance on representative graphs with different densities and different sizes. We investigate and explain the variation of the performance, and also compare it to other techniques (including without using network coding).
The rest of the paper is organized as follows: section [2](#S2 "2 Definitions") details the network model and definitions, section [2.6](#S2.SS6 "2.6 The Proposed Heuristic: IR-MS ‣ 2 Definitions") describes the heuristics, section [3](#S3 "3 Experimental Results") analyzes performance with experimental results and section [4](#S4 "4 Conclusion") concludes.
## 2 Definitions
### 2.1 Network Model
In this report, we study the problem of broadcasting from one source to all nodes. We will focus on getting a preliminary idea of the performance of our heuristics for wireless networks. Hence, we will assume an ideal wireless model in this report (and realistic models would be subject of future work): lossless wireless transmissions without collisions or interferences. We also assume that each node of the network is operating below its maximum transmission capacity.
In an idealized model, multi-hop wireless networks can be modeled as *unit disk graphs* of the plane, where two nodes are neighbors if their distance is lower than a fixed radio range as seen in Fig. [1](#S2.F1 "Figure 1 ‣ 2.1 Network Model ‣ 2 Definitions"). In addition, in wireless networks, the *wireless broadcast advantage* is used: each transmission is overheard by several nodes. As a result the graph is in reality a *(unit disk) hypergraph*. Precisely, we consider the following networks: unit disk \[hyper\]graphs where nodes are either distributed randomly ([Fig. 1.1](#S2.F1.sf1 "In Figure 1 ‣ 2.1 Network Model ‣ 2 Definitions")) or more regularly organized in a lattice ([Fig. 1.2](#S2.F1.sf2 "In Figure 1 ‣ 2.1 Network Model ‣ 2 Definitions")). In addition, in both cases, we also consider their variants where the network is a torus, with wrap-around connections in both the $`x`$ and $`y`$ directions as on [Fig. 1.3](#S2.F1.sf3 "In Figure 1 ‣ 2.1 Network Model ‣ 2 Definitions").
[Refer to caption](/html/0706.4175/assets/x1.png)
Fig. 1.1
random unit disk graph
[Refer to caption](/html/0706.4175/assets/x2.png)
Fig. 1.2
lattice graph
[Refer to caption](/html/0706.4175/assets/x3.png)
Fig. 1.3
on a torus
Figure 1: Network Models
### 2.2 Notation
We will consider the following items in each instance of the several types of graphes in section [2](#S2 "2 Definitions"), and use the following notation in the rest of the report:
- •
Nodes: $`\mathcal{V}`$, set of vertices (nodes) of the graph
- •
Number of nodes: $`N`$
- •
Expected/average number of neighbors: $`M`$
- •
Hyperedge : $`H_{v}`$, $`H_{v} \subset \mathcal{V}`$ is the subset of nodes which are reached by one transmission of node $`v`$ (the neighbors of $`v`$)
- •
Rate: Each node $`v`$ retransmits coded packets with a fixed rate $`C_{v}`$.
- •
Cost per broadcast: $`E_{cost}`$; defined is section [2.3](#S2.SS3 "2.3 Performance ‣ 2 Definitions")
- •
Optimal cost per broadcast : $`E_{optimal}`$ (section [2.3](#S2.SS3 "2.3 Performance ‣ 2 Definitions"))
- •
Min-cut of the source $`s`$, for broadcast to the entire network: $`C_{\min}\hspace{0pt}{(s)}`$; see section [2.4](#S2.SS4 "2.4 Achievable Broadcast Rate: Min-cut ‣ 2 Definitions")
### 2.3 Performance
Because we focus on broadcasting, our approach uses a simplified subgraph selection, where every node is a coding node. The solution for the simplified subgraph consists of defining the rate of each node – by the heuristic.
To measure the performance of the heuristics, we use the number of transmissions per broadcast as the metric for the cost. We consider:
- •
the number of retransmissions from every node per unit time (directly given by selected rate).
- •
the number of packets successfully broadcasted from the source to the entire network per unit time; it is the achievable broadcast rate.
By dividing the number of retransmissions by the number of packets successfully broadcasted, we obtain our metric for the cost per broadcast and denote it $`E_{cost}`$. For reference, we will also use the cost of the optimal solution, $`E_{optimal}`$, obtained by solving the linear program as presented in \[[10](#bib.bib10)\].
The number of packets successfully broadcasted per the unit time , i.e. the maximum achievable broadcast rate is computed as the *min-cut* from the source to every destination in the network considered as a hypergraph as defined in section [2.4](#S2.SS4 "2.4 Achievable Broadcast Rate: Min-cut ‣ 2 Definitions").
### 2.4 Achievable Broadcast Rate: Min-cut
A central result of the performance of network coding in wireless networks gives the maximum broadcast (more generally: multicast) rate for a source. It is given by the *min-cut* from the source to each individual destination of the networks, viewed as a hypergraph \[[15](#bib.bib15)\].
Let us consider the source $`s`$, and one of the broadcast destinations $`t \in \mathcal{V}`$. The definition of a *$`s`$-$`t`$ cut* is: a partition of the set of vertices $`V`$ in two sets $`S`$, $`T`$ such as $`s \in S`$ and $`t \in T`$. Let $`Q\hspace{0pt}{(s,t)}`$ be the set of such *$`s`$-$`t`$ cuts*: $`{(S,T)} \in {Q\hspace{0pt}{(s,t)}}`$.
We denote $`\Delta\hspace{0pt}S`$, the set of nodes of $`S`$ which are neighbors of at least one node of $`T`$ ; the *capacity of the cut* $`C\hspace{0pt}{(S)}`$ is defined as the maximum rate between the nodes in $`S`$ and the nodes in $`T`$:
<table>
<tbody>
<tr>
<td></td>
<td>C(S)≜∑v∈ΔSCv≜𝐶𝑆subscript𝑣Δ𝑆subscript𝐶𝑣C(S)\triangleq\sum_{v\in\Delta{}S}C_{v}</td>
<td></td>
<td>(1)</td>
</tr>
</tbody>
</table>
The *min-cut* between $`s`$ and $`t`$ is the cut of $`Q\hspace{0pt}{(s,t)}`$ with the minimum capacity. Let us denote $`C_{\min}\hspace{0pt}{(s,t)}`$ its capacity. From \[[9](#bib.bib9), [16](#bib.bib16), [15](#bib.bib15)\], the maximum broadcast capacity is given by the minimum of capacity of the min-cut of every destination $`C_{\min}\hspace{0pt}{(s)}`$, with:
<table>
<tbody>
<tr>
<td></td>
<td>Cmin(s,t)≜min(S,T)∈Q(s,t)C(S)andCmin(s)≜mint∈𝒱∖{s}Cmin(s,t)formulae-sequence≜subscript𝐶min𝑠𝑡subscript𝑆𝑇𝑄𝑠𝑡𝐶𝑆and≜subscript𝐶min𝑠subscript𝑡𝒱𝑠subscript𝐶min𝑠𝑡C_{\mathrm{min}}(s,t)\triangleq\min_{(S,T)\in Q(s,t)}C(S)\leavevmode\nobreak\ \leavevmode\nobreak\ \leavevmode\nobreak\ \mathrm{and}\leavevmode\nobreak\ \leavevmode\nobreak\ \leavevmode\nobreak\ C_{\mathrm{min}}(s)\triangleq\min_{t\in\mathcal{V}\setminus\{s\}}C_{\mathrm{min}}(s,t)</td>
<td></td>
</tr>
</tbody>
</table>
Once computed, the performance given by the global min-cut $`C_{\min}\hspace{0pt}{(s)}`$ can be achieved by different coding methods (precisely or asymptotically).
### 2.5 Previous Heuristic
In \[[12](#bib.bib12), [13](#bib.bib13)\] we proposed a simple heuristic, where most nodes have the same rate except the source and some nodes near the edge of the network (*exceptional nodes*): IREN/IRON (*Increased Rate for Exceptional Nodes, Identical Rate for Other Nodes*).
The simple heuristic achieved asymptotically near-optimal efficiency for unit disk graphs in large dense homogenous networks . In other words, the simple heuristic enables for every transmission to bring *innovative* information to almost every receiver.
Let us reproduce the logic of the heuristic:
1. 1\.
Assume that the every node has an identical retransmission rate. Assume it is $`1`$, arbitrarily, e.g. one packet per second.
2. 2\.
Then every node with $`M`$ neighbors can receive $`M`$ coded packets per second. Assume that nearly all of them innovative.
3. 3\.
Then the source should inject at least $`M`$ packets per a second.
4. 4\.
An issue is the nodes of near the border , because they have less neighbors - so in order to be safe, their rate is set to $`M`$ as well (IREN).
Within this framework, it was proven that the achievable broadcast capacity, the min-cut, would be $`M`$ for lattice graphs, and asymptotically $`M`$ for dense random unit-disk graphs. For large lattice or random unit-disk graphs, the cost per broadcast would also converge to the optimal.However the result is asymptotical; for a practical instance of a graph, the cost of the “increased rate” of the border nodes could be considerable, and unnecessary high.
An immediate way to alleviate the cost is to omit step [4](#S2.I3.i4 "item 4 ‣ 2.5 Previous Heuristic ‣ 2 Definitions") of the reasoning entirely, ie, ignoring the issue of border, thus using the following heuristics:
- •
IRON only (Identical rate for other nodes): every node retransmits with rate $`1`$, except for the source, which transmits with rate $`M`$.
For illustrative purposes, we show the maximum achievable broadcast rate, for two instances of the networks with the rates selected by “IRON only” on Fig. [2](#S2.F2 "Figure 2 ‣ 2.5 Previous Heuristic ‣ 2 Definitions").
[Refer to caption](/html/0706.4175/assets/x4.png)
Fig. 2.1
min-cut with IRON only, on a lattice
[Refer to caption](/html/0706.4175/assets/x5.png)
Fig. 2.2
min-cut with IRON only, on a random unit disk
Figure 2: Min-cut with IRON only
The Fig. [2](#S2.F2 "Figure 2 ‣ 2.5 Previous Heuristic ‣ 2 Definitions") represents the min-cut $`C_{\min}\hspace{0pt}{(s,t)}`$ of each node $`t`$ with respect to its position when the total number of nodes is $`N = 400`$. Both $`x`$ and $`y`$ axis represent the position of each node on the plane; the value on the vertical $`z`$ axis represents min-cut
<sup>1</sup>
<sup>1</sup>1for the randomly generated unit disk graph, there is interpolation
. On a lattice, [Fig. 2.1](#S2.F2.sf1 "In Figure 2 ‣ 2.5 Previous Heuristic ‣ 2 Definitions"), every node has $`M = 4`$ neighbors, and the source is in the middle of the network. In the random unit disk graph in [Fig. 2.2](#S2.F2.sf2 "In Figure 2 ‣ 2.5 Previous Heuristic ‣ 2 Definitions") we select the node with the most number of neighbors as the source, and generate graphs with density (the average number of nodes in a range) equal to $`M = 20`$.
The min-cut was computed with the software library implementing the maxflow computation algorithm from \[[17](#bib.bib17)\] (and one additional layer to model a directed hypergraph as a directed graph). The optimizations for tree reuse from \[[18](#bib.bib18)\] were also used.
As seen inFig. [2](#S2.F2 "Figure 2 ‣ 2.5 Previous Heuristic ‣ 2 Definitions"), the nodes near the border have a min-cut much lower than nodes in the middle of the network ($`4`$ compared to $`2`$ for [Fig. 2.1](#S2.F2.sf1 "In Figure 2 ‣ 2.5 Previous Heuristic ‣ 2 Definitions")). From section [2.4](#S2.SS4 "2.4 Achievable Broadcast Rate: Min-cut ‣ 2 Definitions"), recall that maximum achievable broadcast rate from the source, $`C_{\min}\hspace{0pt}{(s)}`$, is the minimum of the min-cuts $`C_{\min}\hspace{0pt}{(s,t)}`$ to each destination $`t`$. Hence these nodes near the border become the bottleneck and dictate a bound for the source rate much lower than what other nodes could have achieved. Also in random unit-disk graph [Fig. 2.1](#S2.F2.sf1 "In Figure 2 ‣ 2.5 Previous Heuristic ‣ 2 Definitions"), notice the same phenomenom occurs on the borders but also that the irregularity of number of neighbors results in irregularity of the min-cut (and hence potential lower minimum min-cut).
These are reasons of inefficiency of IRON.
### 2.6 The Proposed Heuristic: IR-MS
As illustrated with the simple heuristics in section [2.5](#S2.SS5 "2.5 Previous Heuristic ‣ 2 Definitions"), the lack of neighbors lets border nodes and some other nodes receive less than $`M`$ packets per second and these nodes cause a decrease of the performance. We name these nodes *starving nodes*
To alleviate the bottleneck from the starving nodes, their neighbors compensate their starving by increasing the transmission rates with the following heuristic, inspired by \[[14](#bib.bib14)\]:
- •
IR-MS (Increased Rate for the Most Starving node): the rate of a node $`v`$ is set to $`C_{v}`$, with:
$`C_{v}`$ = $`\frac{M}{m\hspace{0pt}i\hspace{0pt}n_{u \in H_{v}}\hspace{0pt}{({|H_{u}|})}}`$, where $`H_{w}`$ is the set of neighbors of $`w`$, and $`M`$ is the source rate.
This rate selection adjusts the step [2](#S2.I3.i2 "item 2 ‣ 2.5 Previous Heuristic ‣ 2 Definitions") of the reasoning presented in section [2.5](#S2.SS5 "2.5 Previous Heuristic ‣ 2 Definitions"), to the fact that some nodes have less than $`M`$ neighbors. When IR-MS, if a node $`v`$ has less than $`{|H_{v}|} < M`$ neighbors, it would still receive a packet rate $`\geq \frac{M}{|H_{v}|}`$ from each of its $`|H_{v}|`$ neighbors, and hence an overall rate at least equal to $`M`$. Notice that this does not necessarily result in a min-cut $`{C_{\min}\hspace{0pt}{(s,v)}} \geq M`$.
[Refer to caption](/html/0706.4175/assets/x6.png)
Fig. 3.1
min-cut with IR-MS, on a lattice
[Refer to caption](/html/0706.4175/assets/x7.png)
Fig. 3.2
min-cut with IR-MS, on a random unit disk
Figure 3: Min-cut with IR-MS
The Fig. [3](#S2.F3 "Figure 3 ‣ 2.6 The Proposed Heuristic: IR-MS ‣ 2 Definitions") represents the value of the min-cut with the rate selection of IR-MS, with the same topologies and same parameters as for IRON only, in previous Fig. [2](#S2.F2 "Figure 2 ‣ 2.5 Previous Heuristic ‣ 2 Definitions"). It appears that for these topologies, the min-cut has the value which is much closer to the targeted value $`M`$. The increased rates of IR-MS at neighbors of bottleneck may bring additional costs. However this additinal cost is relatively much lower than the increase of min-cut, thus the performance is overally improved. We will present more systematic experiments on the performance of the heuristic IR-MS in the next section.
## 3 Experimental Results
### 3.1 Efficiency on Different Types of Networks
We first evaluate the efficiency of the heuristic IR-MS, with the relative cost w.r.t. the optimal:$`E_{{rel} - {eff}} \triangleq \frac{E_{optimal}}{E_{cost}}`$
One reference point is the approximative upper bound in \[[13](#bib.bib13)\], the achievable performance without network coding, which translates into: $`E_{{bound} - {rel} - {eff}}^{({{no} - {coding}})} \approx {0.609\hspace{0pt}\ldots}`$.
[Refer to caption](/html/0706.4175/assets/x8.png)
Fig. 4.1
micut: cumulative distribution
[Refer to caption](/html/0706.4175/assets/x9.png)
Fig. 4.2
mincut vs distance from the border
[Refer to caption](/html/0706.4175/assets/x10.png)
Fig. 4.3
cost vs avg neighbor num on square N=200
[Refer to caption](/html/0706.4175/assets/x11.png)
Fig. 4.4
cost vs total node num on square M=20
Figure 4:
For comparison purposes, we evaluated the IR-MS heuristic on instances of lattice unit disk graphs and random unit disk graphs, both with and without torus effect — four variants in total. Our parameters are the following: number of nodes $`N = 196`$, the avg. number of neighbors is successively $`4,12,28,48,80`$.
[Refer to caption](/html/0706.4175/assets/x12.png)
Figure 5: Relative cost (efficiency compared to optimal)
Fig. [5](#S3.F5 "Figure 5 ‣ 3.1 Efficiency on Different Types of Networks ‣ 3 Experimental Results") represents the efficiency obtained for different cases (average of 10 results). The central result is the appreciable performance of IR-MS compared to the achievable performance without network coding: it mostly outperforms the upper bound without network coding ($`0.609`$).
Now consider the four different types of networks:
$`\bullet`$ First type, the most regular graphs: lattice unit disk graphs on torus. IR-MS closely approaches optimality ($`> 0.95`$). This is because every node has the same number of neighbors and the rate adjustment is not needed.
$`\bullet`$ Second type, lattice unit disk graphs without torus: the nodes near the border have less neighbors than others are the issue. IR-MS successfully solves the border issue, show in next section, approaching the targeted maximum broadcast rate, $`\frac{\min - {cut}}{M} = 1`$. However efficiency decreases as density increases, because of an increase of the cost of nodes near the border.
$`\bullet`$ Third type, random unit disk graph on torus. No border effect here, but IR-MS has to overcome the effects of non-homogeneity, which is done convincingly.
$`\bullet`$ Fourth type, genuine random unit disk graph. The performance is acceptable, but on low density, performance becomes lower, because IR-MS does not fully achieve the maximum broadcast rate, $`M`$.
### 3.2 Distribution of the Min-cut
In this section, further results for case of the random unit disk graphs are provided: these graphs were the ones with the most problematic performance.
Again, the minimum of the min-cuts $`C_{\min}\hspace{0pt}{(s,t)}`$ (for all $`t \in \mathcal{V}`$) decides the overall maximum broadcast rate of the source. Hence good performance is achieved when the distribution of these min-cuts is tighter. Deeper insight is gained by analyzing the cumulative distribution of the min-cut of each node, for the random graph $`N = 400`$ previously studied: it is displayed on [Fig. 4.1](#S3.F4.sf1 "In 3.1 Efficiency on Different Types of Networks ‣ 3 Experimental Results"). As evidenced, without IR-MS, the distribution of the min-cut is wider, but with IR-MS, the distribution is closer to $`M`$ (with a peak for $`M = 20`$, the targetted min-cut). Still, there is some room for improvement, because a few nodes have a min-cut around $`15`$.
The [Fig. 4.2](#S3.F4.sf2 "In 3.1 Efficiency on Different Types of Networks ‣ 3 Experimental Results") provides additional information about the value of the min-cut depending on the position of the nodes, for IR-MS and IRON respectively. The distance of each nodes to the border of the network is computed, and statistics are made for the min-cut of nodes with same distance to the border. The [Fig. 4.1](#S3.F4.sf1 "In 3.1 Efficiency on Different Types of Networks ‣ 3 Experimental Results") gives the average min-cut and the minimum min-cut for nodes at a given distance from the border.
It evidences that the border effect is key, with a lower min-cut when the node is near the border: again, we see that IR-MS improves the minimum min-cut but does not always achieve the target min-cut$`= M`$.
### 3.3 Random Unit Disk Graphs $`N`$, $`M`$
The [Fig. 4.3](#S3.F4.sf3 "In 3.1 Efficiency on Different Types of Networks ‣ 3 Experimental Results") and [Fig. 4.4](#S3.F4.sf4 "In 3.1 Efficiency on Different Types of Networks ‣ 3 Experimental Results"), show different perpectives on the performance. The performance measured on these graphs, is $`E_{cost}`$, the number of retransmissions per broadcast packet.
First, different algorithms are compared:
$`\bullet`$ Network coding with IRON.
$`\bullet`$ Network coding with IR-MS.
$`\bullet`$ MPR-based dominating sets from \[[5](#bib.bib5)\] (with performance close to MultiPoint-Relays (MPR)-based flooding of \[[3](#bib.bib3)\]). It is representative of the performance of algorithms, with only local topology information, without network coding.
$`\bullet`$ Connected dominating set from \[[4](#bib.bib4)\] (efficient variant of a greedy algorithm). It is representative of the performance of centralized connected dominating sets algorithms (and to some extent, representative of what could be achieved without network coding).
$`\bullet`$ for [Fig. 4.3](#S3.F4.sf3 "In 3.1 Efficiency on Different Types of Networks ‣ 3 Experimental Results"), the optimal solution with network coding.
[Fig. 4.3](#S3.F4.sf3 "In 3.1 Efficiency on Different Types of Networks ‣ 3 Experimental Results") displays the performance of each algorithm when the density of the network increases.
One result is that the performance of the optimal network coding is about $`\frac{1}{3}`$ better than the performance of the connected dominating set from \[[4](#bib.bib4)\], without network coding.
Then it appears that IR-MS is also close to this efficient dominating set, which is already an interesting result. The gap between IR-MS and the optimal with network coding however indicates that some improvements are possible.
Last, MPR-based dominating sets, which use only local topology information are have the lowest performance: this indicates that network coding, with IR-MS, would be useful in practice, if a fully distributed solution could be designed (with distributed min-cut computation \[[19](#bib.bib19)\]).
The [Fig. 4.4](#S3.F4.sf4 "In 3.1 Efficiency on Different Types of Networks ‣ 3 Experimental Results") shows the results for increasing density ; they are essentially similar.
In appendix [A](#A1 "Appendix A Additional Results"), additional results are given on Fig. [7](#A1.F7 "Figure 7 ‣ Appendix A Additional Results"), for higher densities ($`N = 400`$ on Fig. [Fig. 7.1](#A1.F7.sf1 "In Figure 7 ‣ Appendix A Additional Results"), and $`N = 800`$ on Fig. [Fig. 7.2](#A1.F7.sf2 "In Figure 7 ‣ Appendix A Additional Results")), and for torus (torus with $`N = 200`$ on Fig. [Fig. 7.3](#A1.F7.sf3 "In Figure 7 ‣ Appendix A Additional Results") and torus with $`N = 400`$ on Fig. [Fig. 7.4](#A1.F7.sf4 "In Figure 7 ‣ Appendix A Additional Results")). As one can see, the results are even better on a torus, because there is no border effect.
### 3.4 Difficulties for Distributed Rate Selection
[Refer to caption](/html/0706.4175/assets/x13.png)
Figure 6: Example of cut
Finally, the Fig. [6](#S3.F6 "Figure 6 ‣ 3.4 Difficulties for Distributed Rate Selection ‣ 3 Experimental Results") gives an example of an instance of a random unit disk graph, where the performance of IR-MS was found to be low. The cut ($`S`$/$`T`$ corresponding to the min-cut) for a node at the bottom of left corner is represented: the green dots are the only nodes which are connecting the part in the corner to the rest of the network. With IR-MS, the green nodes do not have extremely starving nodes as neighbors, and hence increase their rate little. But data from the source is only transmitted through these green nodes to the sets of white dots, hence for this reason their rate should be greater. This example perfectly illustrates the difficulties found in sparse networks: notice how coordination between the green nodes would require multi-hop communication to detect the issue.
## 4 Conclusion
We proposed and experimentally studied a heuristic for efficient broadcasting with network coding only using static local information: one hop or two hop neighbors. We showed excellent performance of this rate selection, and detailed reasons for variations of performance.Future work includes the use of dynamic information for the heuristics in complement of the static local topology information.
## References
- \[1\] R. Ahlswede, N. Cai, S.-Y. R. Li and R. W. Yeung, *“Network Information Flow”*, IEEE Trans. Inf. Theory, vol. 46, no.4, pp. 1204-1216, Jul. 2000
- \[2\] M. Cagalij, J.-P. Hubaux, and C.Enz, *“Minimun-energy broadcast in all wireless networks: NP-completeness and distribution issues,”* in ACM/IEEE Mobicom 2002
- \[3\] T. Clausen and P. Jacquet (eds.), C. Adjih, A. Laouiti, P. Minet, P. Mühlethaler, A. Qayyum, L. Viennot *“Optimized Link State Routing Protocol (OLSR)”*, IETF RFC 3626, October 2003.
- \[4\] S. Guha and S. Khuller, *“Approximation algorithms for connected dominating sets”*, European Symposium on Algorithms, 1996.
- \[5\] C. Adjih, Ph. Jacquet, and Laurent Viennot, *“Computing Connected Dominated Sets with Multipoint Relays”*, Ad Hoc & Sensor Wireless Networks, 2005.
- \[6\] D. S. Lun, M. Médard, R. Koetter, and M. Effros, *“Further Results on Coding for Reliable Communication over Packet Networks”* International Symposium on Information Theory (ISIT 2005), Sept. 2005
- \[7\] T. Ho, R. Koetter, M. Médard, D. Karger and M. Effros, *“The Benefits of Coding over Routing in a Randomized Setting”*, International Symposium on Information Theory (ISIT 2003), Jun. 2003
- \[8\] D. S. Lun, M. Médard, R. Koetter, and M. Effros, *“On coding for reliable communication over packet networks”*, Tech. Report \#2741, MIT LIDS, Jan. 2007
- \[9\] Gowaikar, Radhika and Dana, Amir F. and Palanki, Ravi and Hassibi, Babak and Effros, Michelle *“On the capacity of wireless erasure networks”*, IEEE International Symposium on Information Theory (ISIT’04), Chicago, IL, 27 June-2 July 2004
- \[10\] D. S. Lun, N. Ratnakar, M. Médard, R. Koetter, D. R. Karger, T. Ho, E. Ahmed, and F. Zhao, *“Minimum-Cost Multicast over Coded Packet Networks”*, IEEE/ACM Trans. Netw., Jun. 2006
- \[11\] Y. Wu, P. A. Chou, and S.-Y. Kung, *“Minimum-energy multicast in mobile ad hoc networks using network coding”*, IEEE Trans. Commun., vol. 53, no. 11, pp. 1906-1918, Nov. 2005
- \[12\] C. Adjih, S. Y. Cho and P. Jacquet, *“Near Optimal Broadcast with Network Coding in Large Sensor Networks”*, accepted at Workshop Information Theory for Sensor Networks, Sante Fe, Jun. 2007
- \[13\] C. Adjih, S. Y. Cho and P. Jacquet,*“Near Optimal Broadcast with Network Coding in Large Homogeneous Networks”*,INRIA Research Report RR-6188, May 2007
- \[14\] C. Fragouli, J. Widmer, and J.-Y. L. Boudec, *“A Network Coding Approach to Energy Efficient Broadcasting”*, INFOCOM 2006, Apr. 2006
- \[15\] S.Deb, M.Effros, T.Ho, D.Karger, R.Koetter, D.Lun, M.Médard, N.Ratnakar, *“Network Coding for Wireless Applications: A Brief Tutorial”*, IWWAN, May 2005.
- \[16\] A. Dana, R. Gowaikar, R. Palanki, B. Hassibi, and M. Effros, *“Capacity of Wireless Erasure Networks”*, IEEE Trans. Inf. Theory, vol. 52, no.3, pp. 789-804, Mar. 2006
- \[17\] Y. Boykov and V. Kolmogorov, *“An Experimental Comparison of Min-Cut/Max-Flow Algorithms for Energy Minimization in Vision”*, IEEE Transactions on Pattern Analysis and Machine Intelligence (PAMI), September 2004
- \[18\] Pushmeet Kohli and Philip H.S. Torr, *“Efficiently Solving Dynamic Markov Random Fields Using Graph Cuts.”*, International Conference on Computer Vision (ICCV), 2005
- \[19\] Z. Li and B. Li, *”Efficient and Distributed Computation of Maximum Multicast Rates”*, INFOCOM 2005, March. 2005
APPENDIX
## Appendix A Additional Results
[Refer to caption](/html/0706.4175/assets/x14.png)
Fig. 7.1
cost vs avg neighbor num on square N=400
[Refer to caption](/html/0706.4175/assets/x15.png)
Fig. 7.2
cost vs total node num on square N=800
[Refer to caption](/html/0706.4175/assets/x16.png)
Fig. 7.3
cost vs avg neighbor num on square torus N=200
[Refer to caption](/html/0706.4175/assets/x17.png)
Fig. 7.4
cost vs total node num on square M=20
Figure 7: Additional results for higher density and for torus square |
0707.1083 | cs.NI | cyber | # Delayed Correlations in Inter-Domain Network Traffic
Viktoria Rojkova, Mehmed Kantardzic
###### Abstract
To observe the evolution of network traffic correlations we analyze the eigenvalue spectra and eigenvectors statistics of delayed correlation matrices of network traffic counts time series. Delayed correlation matrix $`D\hspace{0pt}(\tau)`$ is composed of the correlations between one variable in the multivariable time series and another at a time delay $`\tau`$. We determined, that inverse participation ratio (IPR) of eigenvectors of $`D\hspace{0pt}(\tau)`$ deviates substantially from that of eigenvectors of the equal time correlation matrix. The largest eigenvalue $`\lambda_{m\hspace{0pt}a\hspace{0pt}x}`$ of $`D\hspace{0pt}(\tau)`$ and the corresponding IPR oscillate with two characteristic periods of $`3\hspace{0pt}\tau`$ and $`6\hspace{0pt}\tau`$. Found delayed correlations between network time series fits well into the long range dependence (LRD) property of the network traffic.
Injecting the random traffic counts between non-randomly correlated time series, we were able to break the picture of periodicity of $`\lambda_{m\hspace{0pt}a\hspace{0pt}x}`$. In addition, we investigated influence of the periodic injections on both largest eigenvalue and the IPR, and addressed relevance of these indicators for the LRD and self-similarity of the network traffic.
###### Index Terms:
Network-Wide Traffic Analysis, Random Correlations, Time-Lagged Correlations, Long Range Dependence
Department of Computer Engineering and Computer Science, University of Louisville, Louisville, KY 40292 email: {vbrozh01, mmkant01}@louisville.edu
## Categories and Subject Descriptors
C.2.3 \[Computer-Communication Networks\]: Network Operations
## General Terms
Measurement, Experimentation
## I introduction
The cross-correlation matrix is one of the tools of time series analysis, used in studies of the underlying interactions between network structural constituents. The network traffic system is often seen as a collection of times series with traffic bytes per time interval or number of packets per time interval at a single link. Within this collection, temporal correlations between network structural constituents and consequently patterns of collective behavior might be present.
To study such patterns one can employ the equal-time cross-correlation matrix of traffic time series at all network links (see, for example \[[1](#bib.bib1), [2](#bib.bib2)\]). Further statistical description of the awareness is commonly done via eigen-decomposition of cross-correlation matrix. Detailed studies showed that the major portion of eigenvalues of cross-correlation matrix fall into the theoretically predicted boundaries of eigenvalues spectrum of random matrices \[[1](#bib.bib1), [2](#bib.bib2)\]. This portion satisfies the so called universal properties described by the random matrix theory (RMT) \[[4](#bib.bib4)\]. It also recognizes the uncongested state of the traffic, where each router is able to communicate with any other router and with each subnet under its service. The part of the spectrum deviating from the RMT boundaries represents stable in time non-random correlations between network traffic time series \[[1](#bib.bib1), [2](#bib.bib2)\].
Even though, such distinction provides a valuable insight into the meaning of empirical data, the equal-time cross-correlation matrix on its own may not be sufficient for understanding of the effect of inter-domain correlations at different times. In fact, the analysis of correlation as a function of time lags has already been used in econometric time series systems. For example, the analysis of stock returns portfolio showed the asymmetric lead-lag relationship between stock returns; high-volume stocks lead the low-volume stocks \[[7](#bib.bib7), [8](#bib.bib8)\]. This finding is attributed to information adjustment asymmetry \[[6](#bib.bib6)\]. The uncongested balanced state of the network traffic implies the "symmetric" information/traffic flow exchange. Thus, to control the congestion level of network traffic systems, it is crucial to recognize the collective behavior or correlation patterns between network traffic time series, and to observe their evolution in time.
In order to trace the evolution of correlation pattern between traffic time series the equal-time cross-correlation matrix is replaced with time-lagged correlation matrix \[[3](#bib.bib3)\]. The obtained matrix of delayed correlations $`D\hspace{0pt}(\tau)`$ can be subjected to eigen-analysis just like its equal-time counterpart. In contrast to equal-time correlation matrices which have a real eigenvalue spectrum, the spectrum of $`D\hspace{0pt}(\tau)`$ is complex since matrices of these types are asymmetric. While the general properties of complex spectrum of $`D\hspace{0pt}(\tau)`$ is unknown so far, results of symmetrized version of lagged correlation matrices have been reported recently \[[3](#bib.bib3), [5](#bib.bib5)\].
In this paper, we concentrated on time-lagged correlations between time series generated by Simple Network Manage Protocol (SNMP) traffic counters of University of Louisville backbone routers system. We established that time-lagged correlations between traffic time series sustain for up to $`100\hspace{0pt}\tau`$, where $`\tau = 300`$ sec. Moreover, the largest and smallest eigenvalues $`\lambda_{m\hspace{0pt}a\hspace{0pt}x}`$ and $`\lambda_{m\hspace{0pt}i\hspace{0pt}n}`$ of $`D\hspace{0pt}(\tau)`$ and the inverse participation ratio (IPR) of the eigenvector corresponding to $`\lambda_{m\hspace{0pt}a\hspace{0pt}x}`$ were found to oscillate with two characteristic periods of $`3\hspace{0pt}\tau`$ and $`6\hspace{0pt}\tau`$.
In addition, within the content of the eigenvector corresponding to $`\lambda_{m\hspace{0pt}a\hspace{0pt}x}`$ we determined the index of a single time series which is driving the $`\lambda_{m\hspace{0pt}a\hspace{0pt}x}`$ oscillation. The IPR of all eigenvectors of $`D\hspace{0pt}(\tau)`$ shows the single time series contributor in the eigenvector which at the equal-time correlations belonged to the random part of the spectrum. Since the theoretical prediction for the spectrum of time-lagged correlation matrices is unknown, we keep the terminology of the spectrum derived from eigen-decomposition of equal-time correlation matrices. Thus, *random* in our text refers to the part of the spectrum within the boundaries predicted by the RMT for equal-time correlation matrix. *Non-random* denotes the part of the spectrum outside of the RMT predictions.
Finally, we found that the injection of random traffic counts between time series which interact *non-randomly* destroys the oscillatory picture of $`\lambda_{m\hspace{0pt}a\hspace{0pt}x}`$ and $`\lambda_{m\hspace{0pt}i\hspace{0pt}n}`$ and that of the corresponding IPRs. Meanwhile, the injection of traffic counts with smooth (periodic) time dependence between *randomly* interacting traffic time series with respect to characteristic periods of $`\lambda_{m\hspace{0pt}a\hspace{0pt}x}`$, $`\lambda_{m\hspace{0pt}i\hspace{0pt}n}`$ and IPR reveals new periodicity (of $`4\hspace{0pt}\tau`$). The findings of eigenvalue spectrum analysis and experimental results suggest the asymmetric long lasting relationship between traffic time series.
Network traffic analysis had undergone the evolution from considering the network traffic time series as an outcome of Poisson and memory-less processes to recognizing the long range dependencies and self-similarity of the traffic. The statistics of eigenvalue spectrum and IPR of eigenvectors of time-lagged correlation matrices provide essential dimensional reduction in the investigation of long-ranged dependence (LRD) of the network traffic. Before starting the expensive process of Hurst parameter estimation, one can attempt to find just a few indicators of LRD, self-similarity or deviations from thereof.
The rest of the paper is organized as follows. We introduce time-lagged correlation matrix and its eigensystem in Section II. Section III is devoted to numerical analysis of time-lagged eigensystem. Then, in Section IV by experimenting with the content of the original time series we addressed LRD and self-similarity of lagged correlations through proposed indicators. Conclusions and discussion are given in Section V.
## II time-lagged correlation matrix of network traffic time series
The starting point of our discussion is averaged traffic count data collected from router-router and router-VLAN subnet connections of the University of Louisville backbone routers system. The same data set was used in \[[1](#bib.bib1)\] in the context of equal-time correlation matrix analysis. Below we recap the details relevant to the construction of lagged correlation matrix, relegating the information on the network and the way data was processed to \[[1](#bib.bib1)\]. We will be dealing with total of $`L = 2015`$ records of $`N = 497`$ time series averaged over $`300`$ seconds, where incoming and outgoing traffic generate independent time series.
In order to define traffic rate change $`G_{i}\hspace{0pt}(t)`$ we used the logarithm of the ratio of two successive counts upon calculating the traffic rate change of time series $`T_{i}`$, $`i = {1,\ldots,N}`$ , over time $`\Delta\hspace{0pt}t`$,
<table>
<tbody>
<tr>
<td></td>
<td>Gi(t)≡lnTi(t+Δt)−lnTi(t),subscript𝐺𝑖𝑡lnsubscript𝑇𝑖𝑡Δ𝑡lnsubscript𝑇𝑖𝑡G_{i}\left(t\right)\equiv\textrm{ln}\,T_{i}\left(t+\Delta t\right)-\textrm{ln}\,T_{i}\left(t\right),</td>
<td></td>
</tr>
</tbody>
</table>
\[[2](#bib.bib2)\]. Then we introduce normalization, according to $`{g_{i}\hspace{0pt}(t)} = {\left( {{G_{i}\hspace{0pt}(t)} - \left\langle {G_{i}\hspace{0pt}(t)} \right\rangle} \right)/\sqrt{\left\langle G_{i}^{2} \right\rangle - \left\langle G_{i} \right\rangle^{2}}}`$ and built the time-lagged correlation matrix $`D\hspace{0pt}(\tau)`$ as follows \[[3](#bib.bib3)\]
<table>
<tbody>
<tr>
<td></td>
<td>Dij(τ)subscript𝐷𝑖𝑗𝜏\displaystyle D_{ij}\left(\tau\right)</td>
<td>≡Sym⟨gi(t)gj(t+τ)⟩absentSymdelimited-⟨⟩subscript𝑔𝑖𝑡subscript𝑔𝑗𝑡𝜏\displaystyle\equiv\textrm{{Sym}}\left\langle g_{i}\left(t\right)g_{j}\left(t+\tau\right)\right\rangle</td>
<td></td>
<td></td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td></td>
<td>=12L∑t=0t=L(gi(t)gj(t+τ)+gj(t)gi(t+τ)).absent12𝐿superscriptsubscript𝑡0𝑡𝐿subscript𝑔𝑖𝑡subscript𝑔𝑗𝑡𝜏subscript𝑔𝑗𝑡subscript𝑔𝑖𝑡𝜏\displaystyle=\frac{1}{2L}\sum_{t=0}^{t=L}\left(g_{i}\left(t\right)g_{j}\left(t+\tau\right)+g_{j}\left(t\right)g_{i}\left(t+\tau\right)\right).</td>
<td></td>
<td>(1)</td>
</tr>
</tbody>
</table>
Here the sole purpose of symmetrization is the restriction of the eigenvalues and eigenvectors to real values. In principle, the numerical experiments we run below can be repeated for the eigensystem of non-symmetric correlation matrix. Studies of the latter are already in progress in a different setting (see, for example, \[[6](#bib.bib6)\]).
###
II-A
Eigenvalues and eigenvectors
Next we proceed with defining the indicators we focus on in what follows. First of all, the eigenproblem for our cross-correlation matrix is time dependent
<table>
<tbody>
<tr>
<td></td>
<td>D(τ)uk(τ)=λk(τ)uk(τ),𝐷𝜏subscript𝑢𝑘𝜏subscript𝜆𝑘𝜏subscript𝑢𝑘𝜏D\left(\tau\right)u_{k}\left(\tau\right)=\lambda_{k}\left(\tau\right)u_{k}\left(\tau\right),</td>
<td></td>
<td>(2)</td>
</tr>
</tbody>
</table>
where $`\lambda_{k}`$ is $`k`$-th eigenvalue, corresponding to $`k`$th eigenvector. In other words, the eigensystem $`\left\{ {\lambda_{k}\hspace{0pt}(\tau)},{u_{k}\hspace{0pt}(\tau)} \right\}`$ is defined for each increment of delay time. As opposed to same time eigensystem $`\left\{ {\lambda_{k}\hspace{0pt}(0)},{u_{k}\hspace{0pt}(0)} \right\}`$, our eigensystem does not characterize presence or lack of organization (localization) in the system at a given time. Instead it can serve as a measure of back-in-time (or forward in time, depending on prospective) correlation within the network structure.
Furthermore, the RMT picture of the eigenvalues and eigenvectors, in which spectrum is split into three parts is no longer valid. By three parts we understand the RMT part, which behave universally at the center of the eigenvalue spectrum, and “left” and “right” parts (lying to the left and to the right from predicted RMT bounds) which exhibit non-universal features \[[1](#bib.bib1)\]. Although, for very small $`\tau`$ this subdivision is clearly still accurate, we can expect, transient behavior of $`\left\{ {\lambda_{k}\hspace{0pt}(\tau)},{u_{k}\hspace{0pt}(\tau)} \right\}`$ to reveal new, otherwise undetectable correlations within the network. Hence, we found it convenient to keep track of quantitative and qualitative changes in eigensystem using left/random/right terminological distinction.
On the other hand, observing the entire system of $`497`$ eigenvalues and the same amount of eigenvectors can be quite tedious task. One needs to come up with more concise indicators of network behavior. Selecting efficient indicators can help in defining “normal” state of the system, a task quite challenging on its own, and in predicting structural reaction to the external or internal disruption. Indicators can be chosen based on the experiments described below, or on their advanced variations. The candidates are those eigenvalues, that are most receptive to a particular probe. As far as eigenvectors are concerned we decided to test corresponding IPRs.
###
II-B
Inverse participation ratio
Given the eigenvector $`u_{k}\hspace{0pt}(\tau)`$ the IPR is computed according to
<table>
<tbody>
<tr>
<td></td>
<td>Ik(τ)≡∑l=1N[ukl(τ)]4,subscript𝐼𝑘𝜏superscriptsubscript𝑙1𝑁superscriptdelimited-[]superscriptsubscript𝑢𝑘𝑙𝜏4I_{k}\left(\tau\right)\equiv\sum_{l=1}^{N}\left[u_{k}^{l}\left(\tau\right)\right]^{4},</td>
<td></td>
<td>(3)</td>
</tr>
</tbody>
</table>
with $`u_{k}^{l}`$, $`l = {1,\ldots,497}`$ stands for components of the $`k`$th eigenvector \[[4](#bib.bib4)\]. The IPR is quite indicative in terms of signaling the number of significant $`u_{k}^{l}`$, i.e. “contributors” to the eigenvector of interest. For example, if we have reasons to expect absence of correlations between routers input into the experimental data, $`I_{k}\hspace{0pt}(0)`$ should have its value around $`1/\sqrt{N}`$. Indeed, the eigenvector is normalized, thus $`{\sum_{l = 1}^{N}\left\lbrack {u_{k}^{l}\hspace{0pt}(\tau)} \right\rbrack^{2}} = 1`$. It has $`N`$ components, and they are all roughly the same in magnitude (otherwise correlations must be present). Therefore, $`u_{k}^{l} \simeq {1/\sqrt{N}}`$, and $`{I_{k}\hspace{0pt}(0)} \simeq {1/N}`$. Note, that since $`N`$ is typically much greater than $`1`$, any finite value of IPR signals *localization* in inter-VLAN traffic \[[1](#bib.bib1)\]. For non-zero values of $`\tau`$, IPRs acquire more general meaning in a sense that routers which interact heavily at time $`t`$ may loose their “bond” at time $`t + \tau`$, while these not knowing about one another at time $`t`$ may have significant mutual contribution at time $`t + \tau`$. Other more complex possibilities can be perceived via $`I_{k}\hspace{0pt}(\tau)`$ as well.
## III Eigenanalysis for time-lagged correlation matrix
###
III-A
Stroboscopic sequence for eigensystem
Upon building the cross-correlation matrix $`D\hspace{0pt}(\tau)`$ with the help of Eq. ([1](#S2.E1 "In II time-lagged correlation matrix of network traffic time series ‣ Delayed Correlations in Inter-Domain Network Traffic")) we perform eigen-decomposition (see Eq. ([2](#S2.E2 "In II-A Eigenvalues and eigenvectors ‣ II time-lagged correlation matrix of network traffic time series ‣ Delayed Correlations in Inter-Domain Network Traffic"))) numerically, using standard *MATLAB* routine. We look at the result of calculation of eigenvalues for all delay times $`\tau`$. A noticeable spike for very small values of delay time is expected, notwithstanding the position in spectrum. However, our increments in $`\tau`$ ($`= {300\hspace{0pt}s\hspace{0pt}e\hspace{0pt}c}`$) may not be small enough to observe it. For the remainder of observation the result has to uncover the way system constituents communicate with themselves and their neighbors on a long run.
[Refer to caption](/html/0707.1083/assets/x1.png)
Figure 1: (a) left, (b) random, and (c) right parts of the eigenvalue spectrum as obtained from actual data. Same graphs are presented in (d), (e), (f) respectively, after noise-like injections are made.
[Refer to caption](/html/0707.1083/assets/x2.png)
Figure 2: (a) $`I\hspace{0pt}(0)`$ versus position in spectrum. Stroboscopic representation of IPRs corresponding to (b) first $`10\hspace{0pt}\tau`$ (c) second $`10\hspace{0pt}\tau`$; (d)-(f) are the same representations upon noise-like injections.
In Fig. [1](#S3.F1 "Figure 1 ‣ III-A Stroboscopic sequence for eigensystem ‣ III Eigenanalysis for time-lagged correlation matrix ‣ Delayed Correlations in Inter-Domain Network Traffic")(a)-(c) we illustrate how left/random/right structure of the spectrum evolves with $`\tau`$. As it turns out, “randomness” and “regularity” find their new interpretations in the context of system remembering itself after $`\tau`$ has elapsed. With the exception of a few located at the right and left edges of the spectrum, most eigenvalues fall very close to each other numerically. To make it more transparent we plotted their $`\tau`$-dependence using different offset values (these values are the same within each part). Only ten eigenvalues are offseted in each case and plotted versus time. The lowest eigenvalue was excluded from consideration here and throughout the paper due to it secular behavior in $`\tau`$.
At a glance, non-edge eigenvalues Fig. [1](#S3.F1 "Figure 1 ‣ III-A Stroboscopic sequence for eigensystem ‣ III Eigenanalysis for time-lagged correlation matrix ‣ Delayed Correlations in Inter-Domain Network Traffic")(b) with the exception of an expected spike at small $`\tau`$ does not seem to represent any process. Such a lack of forward in time correlation is not completely surprising, as the eigenvalues from middle part of the spectrum were referred as RMT-like. It follows, that the random interactions between traffic time series are time delay invariant. In other words, random spectrum of eigenvalues is an indicator of self-similarity \[[12](#bib.bib12)\]. Meantime, the eigenvalues at the edges (Figs. [1](#S3.F1 "Figure 1 ‣ III-A Stroboscopic sequence for eigensystem ‣ III Eigenanalysis for time-lagged correlation matrix ‣ Delayed Correlations in Inter-Domain Network Traffic")(a) and (c)) represent a quasiperiodic process, distinguishing themselves from their "peers", the eigenvalues belonging to regular part of the spectrum for $`\tau = 0`$ quantitatively, and both qualitatively and quantitatively from the eigenvalues belonging to the RMT-part for $`\tau = 0`$. They scale with delay time and clearly exhibit long time dependence \[[12](#bib.bib12)\]. Therefore, it makes sense to look further into the properties of edge eigenvalues, especially into the properties of those with relatively high absolute values. The actual values might be used as a measure of delayed time correlations. Having located potential indicators we move ahead with the search for others.
A remarkable property of the IPR for equal time cross-correlation matrix (Fig. [2](#S3.F2 "Figure 2 ‣ III-A Stroboscopic sequence for eigensystem ‣ III Eigenanalysis for time-lagged correlation matrix ‣ Delayed Correlations in Inter-Domain Network Traffic")(a)) is its consistently low, order $`1/N`$, value for the major part of the spectrum. This segment in Fig. [2](#S3.F2 "Figure 2 ‣ III-A Stroboscopic sequence for eigensystem ‣ III Eigenanalysis for time-lagged correlation matrix ‣ Delayed Correlations in Inter-Domain Network Traffic")(a) is known to obey the RMT \[[1](#bib.bib1)\]. To the left and to the right from this segment there is a strong evidence of regular, non-random behavior. Now, if the first $`20`$ instances are considered as in Figs. [2](#S3.F2 "Figure 2 ‣ III-A Stroboscopic sequence for eigensystem ‣ III Eigenanalysis for time-lagged correlation matrix ‣ Delayed Correlations in Inter-Domain Network Traffic")(b) and (c), where IPRs offseted by an arbitrary amount for transparency and plotted versus the eigenvalue position, the situation looks drastically different. The peak located close to the center of the spectrum signifies the presence of previously undetected correlations, and the lead-lag relationship between time series.
Close examination of Figs. [2](#S3.F2 "Figure 2 ‣ III-A Stroboscopic sequence for eigensystem ‣ III Eigenanalysis for time-lagged correlation matrix ‣ Delayed Correlations in Inter-Domain Network Traffic")(b) and (c) shows, that initially, the high IPR has changing support in the spectrum. Furthermore, as explained in Section 2.2, peak value tells us, that about four time series drive the correlation pattern. Later on, the peak "settles down" and establishes itself around median eigenvalue position (Fig. [2](#S3.F2 "Figure 2 ‣ III-A Stroboscopic sequence for eigensystem ‣ III Eigenanalysis for time-lagged correlation matrix ‣ Delayed Correlations in Inter-Domain Network Traffic")(b)). The meaning of this and other two peaks differs from that of the IPR peaks in Fig. [2](#S3.F2 "Figure 2 ‣ III-A Stroboscopic sequence for eigensystem ‣ III Eigenanalysis for time-lagged correlation matrix ‣ Delayed Correlations in Inter-Domain Network Traffic")(a). The increase in IPR computed from the time delayed matrix $`D\hspace{0pt}(\tau)`$ indicate correlations between system’s behavior at a given time and its stroboscopic image after $`\tau`$, rather than correlations within the spectrum. In addition, it provides reasonable way of tracking down the sources of lead-lag behavior. Thus, the observed features make IPR a good candidate for an indicator of the network congestion state. Note also significant change in height of the central peak.
###
III-B
Frequency domain analysis
[Refer to caption](/html/0707.1083/assets/x3.png)
Figure 3: Eigenvalues number (a) $`2`$, (b) $`257`$, and (c) $`497`$, plotted with respect to time and their respective Fourier spectra ((d) through (f)).
[Refer to caption](/html/0707.1083/assets/x4.png)
Figure 4: IPRs for eigenvalues number (a) $`2`$, (b) $`257`$, and (c) $`497`$, plotted with respect to time and their respective Fourier spectra ((d) through (f)).
The next step in getting more quantitative on long memory processes in network traffic is to analyze transient behavior of eigenvalues and IPRs of matrix $`D\hspace{0pt}(\tau)`$ in detail. Since quasiperiodic behavior is present in the majority of quantities of interest we focus on their frequency content. The standard way to proceed is to transform $`\lambda_{i}\hspace{0pt}(\tau)`$ into frequency domain using fast Fourier transform. In a sense, we construct a spectra of the spectrum. The same operation is performed on respective IPRs. We take fast Fourier transforms for all of the functions at hand, and then, take the square of their absolute value. The result is referred to as a power spectrum. There should be no confusion, as graphs of power always accompany the corresponding time domain quantity.
In Fig. 3 we display representative eigenvalue dynamics. Once again, random $`\lambda_{i}`$ (Fig. 3(b)) does not exhibit anything remarkable, compared to its regular counterparts. The latter resemble each other, reflecting a symmetry of the spectra induced by symmetrizing procedure (Eq. ([1](#S2.E1 "In II time-lagged correlation matrix of network traffic time series ‣ Delayed Correlations in Inter-Domain Network Traffic"))). For now, we can talk about them in parallel.
Aside from a substantial low and high frequency contribution, which could have already been guessed from Figs. 3(a) and (c), we discover two strong contributions from frequencies corresponding to oscillations with time periods $`15`$ and $`30`$ minutes respectively (cf. Figs. 3(d) and (f)). This is in evident contrast to the situation with power of a random eigenvalue. Such an eigenvalue has equal (negligibly small) contribution from the entire range of frequencies. The existence of these two characteristic frequencies suggests a natural way assessing the current state of the inter-domain network traffic. In fact, it might be possible to use these as the LRD quantifier estimators \[[12](#bib.bib12)\] in the future.
## IV experiments with altering actual network traffic
###
IV-A
Noise-like injections
Next, we investigated consequences of modifying the time-lagged correlations between time series. We have already known the time series contributing the most to the correlation pattern \[[1](#bib.bib1)\]. All of them can be linked to eigenvalues which fall in what we term here as the right segment of eigenvalue spectrum. In these series we replaced the original traffic counts with counts obtained by random number generator for a certain period of time. Then, we constructed matrix $`D\hspace{0pt}(\tau)`$ for all hundred increments and repeated manipulations described in previous Section 3.1. The results are shown in Fig. [1](#S3.F1 "Figure 1 ‣ III-A Stroboscopic sequence for eigensystem ‣ III Eigenanalysis for time-lagged correlation matrix ‣ Delayed Correlations in Inter-Domain Network Traffic")(d) through (f).
The eigenvalues belonging to middle of the spectrum are completely unaffected, i.e. they are still time delay invariant. Clearly, our manipulations with the traffic are not disturbing the self-similar nature of delayed correlations. However, edge eigenvalues loose time scales present in their original transient behavior (see Fig. [1](#S3.F1 "Figure 1 ‣ III-A Stroboscopic sequence for eigensystem ‣ III Eigenanalysis for time-lagged correlation matrix ‣ Delayed Correlations in Inter-Domain Network Traffic") (d) through (f)). In other words, the LRD gets destroyed. Effect on the IPR (Figs. [2](#S3.F2 "Figure 2 ‣ III-A Stroboscopic sequence for eigensystem ‣ III Eigenanalysis for time-lagged correlation matrix ‣ Delayed Correlations in Inter-Domain Network Traffic")(d) and (f).) is less noticeable but is still there, while for the random segment it is absent. The result of random counts injections can be summarized as presence of randomly positioned of small peaks superimposed on the original IPR picture. Indeed, in Figs. [2](#S3.F2 "Figure 2 ‣ III-A Stroboscopic sequence for eigensystem ‣ III Eigenanalysis for time-lagged correlation matrix ‣ Delayed Correlations in Inter-Domain Network Traffic")(a) and (c) small peaks are very infrequent and unstable in time, unlike these in Figs. [2](#S3.F2 "Figure 2 ‣ III-A Stroboscopic sequence for eigensystem ‣ III Eigenanalysis for time-lagged correlation matrix ‣ Delayed Correlations in Inter-Domain Network Traffic")(d) and (f).
[Refer to caption](/html/0707.1083/assets/x5.png)
Figure 5: Eigenvalues number (a) $`2`$, (b) $`257`$, and (c) $`497`$, plotted with respect to time and their respective Fourier spectra ((d) through (f)) after noise-like sample was injected.
[Refer to caption](/html/0707.1083/assets/x6.png)
Figure 6: IPRs for eigenvalues number (a) $`2`$, (b) $`257`$, and (c) $`497`$, plotted with respect to time and their respective Fourier spectra ((d) through (f)) after noise-like sample was injected.
The above outcome calls for a more close look into eigenvalues and IPRs for the system experienced the noisy injections into the time series, which are believed to have major contribution to the overall traffic in router network. We present three eigenvalues considered in Section 3.2 as the functions of time delay together with their respected power spectra. As can be concluded from Fig. 5(a) and (c), the time dependence looses its LRD structure. It is backed up by the fact that a lot more frequencies contribute to power spectra upon random injection. Middle part of the spectra also undergoes certain transformation, but is still scale-free Fig. 5(b), as actual values of power are small relative to the power corresponding to edge eigenvalues. The quantitative changes are also in place for both edge eigenvalues. The effect can be judged based on comparison of the tallest peaks in Figs. 5(d) and (f) to their counterparts in Figs. 3(d) and (f).
Similar conclusions can be derived for the IPR as we take a look at Fig. 6(d) and (f) and compare the outcome of our experiment with the graphs in Fig. 4(d) and (f).
###
IV-B
Periodic in time injections
[Refer to caption](/html/0707.1083/assets/x7.png)
Figure 7: Eigenvalues number $`2`$, $`257`$, and $`497`$: The results of the injections with $`2.5\hspace{0pt}m\hspace{0pt}i\hspace{0pt}n`$ period.
[Refer to caption](/html/0707.1083/assets/x8.png)
Figure 8: Eigenvalues number $`2`$, $`257`$, and $`497`$: The results of the injections with $`15\hspace{0pt}m\hspace{0pt}i\hspace{0pt}n`$ period.
A logical continuation of the above experiment is the injection of an artificial traffic counts which possess regularity into actual experimental data. This time, however, we perform the replacements for the time series which can be traced back to the eigenvalues falling into the random segment. Time series for this replacement were chosen at random. Other possibilities can also be considered, but since random segment was much less susceptible to the previous experiment, the above choice seems natural.
[Refer to caption](/html/0707.1083/assets/x9.png)
Figure 9: Eigenvalues number $`2`$, $`257`$, and $`497`$: The results of the injections with $`20\hspace{0pt}m\hspace{0pt}i\hspace{0pt}n`$ period.
[Refer to caption](/html/0707.1083/assets/x10.png)
Figure 10: Eigenvalues number $`2`$, $`257`$, and $`497`$: The results of the injections with $`30\hspace{0pt}m\hspace{0pt}i\hspace{0pt}n`$ period.
We choose four injections to be cosinusoidal, having periods of $`2.5`$; $`15`$; $`20`$ and $`30\hspace{0pt}m\hspace{0pt}i\hspace{0pt}n`$ and repeating the same manipulations as in the first experiment discussed in Section 4.1. The results are fairly sound. Although the random part of the eigenvalue spectrum is again unaltered, the “reaction” of left and right parts is both qualitative and quantitative. When cosinusoidal sample with period much smaller than both characteristic periods ($`15`$ and $`30\hspace{0pt}m\hspace{0pt}i\hspace{0pt}n`$), power spectra in Figs. 7(d) and (f) are not significantly changed. Two characteristic periods are still present, and yet certain narrow frequency range gets suppressed (note anti-peak between the main two). Notice slight asymmetry in the way smallest and largest eigenvalues react to the injection. We should add, that we observed essentially the same picture for the injections with period of $`5`$ and $`10`$ minutes (both not matching, but commensurate with characteristic periods).
Now, we turn to Fig. 8, where the cosinusoidal replacement with period $`15\hspace{0pt}m\hspace{0pt}i\hspace{0pt}n`$ of actual traffic counts leads to the dramatic change in power spectrum. We observe enhancement of the peak corresponding to period of $`15\hspace{0pt}m\hspace{0pt}i\hspace{0pt}n`$, which can be termed as resonance phenomenon (Figs. 8(d) and (f)). The very same plots show the suppression of peaks corresponding to the other characteristic period of $`30\hspace{0pt}m\hspace{0pt}i\hspace{0pt}n`$. Similar resonant effect is achieved when the period of injection is changed to $`20\hspace{0pt}m\hspace{0pt}i\hspace{0pt}n`$ (see Fig. 9). This time, both peaks are gone, while the new characteristic period is detected in Fig. 9(f). It is approximately equal to the period of injection. Finally, for the experiment, in which period of the injection was chosen to be $`30\hspace{0pt}m\hspace{0pt}i\hspace{0pt}n`$, i.e. matching to another characteristic period, we obtained yet another result supporting previous conclusions.
In this case, however, the resonance phenomenon is slightly more difficult to establish. From the results displayed in Fig. 10. We see, that relative contribution to power spectrum is now changed for two main peaks. Before the experiment was performed, the higher harmonic (smaller period) dominated by quite a few orders of magnitude. After running the experiment, this is still the case for the spectrum of largest eigenvalue, but now the difference is marginal (see Fig. 10(d)). At the same time for the left most eigenvalue, we determined, that lower harmonic (period, matching the period of injection) now contributes the most, as can be checked in Fig. 10(d). Two power spectra for the edge eigenvalues are no longer symmetric and the contributions from some ranges of frequencies are again strongly diminished. As for the random eigenvalue considered Fig. 10(b) and (e), no impact has been recorded, just as in all other cases.
## V Discussion
Long range or time dependent processes which show significant correlations across large time scales were first discovered in network traffic over a decade ago. Since then, LRD was found and studied intensively in various aspects of network behavior. LRD is a manifestation of self-similarity of the process, meaning that the behavior of the process is space and time scale invariant. Leland and colleagues performed first rigorous statistical analysis of self-similar characteristics in Local Area Network (LAN) traffic \[[15](#bib.bib15)\]. They showed that the aggregated Ethernet traffic is not smoothing out with accordance to Poisson model, it is time scale invariant. In this framework the traditional Poisson or memory-less models of network traffic became inadequate. Since high variability across different time scales produces high congestion level, the impact of the self-similar traffic models on queuing performance is considerable \[[9](#bib.bib9)\].
However, the identification of self-similarity origin and estimation of LRD in network traffic are far from straightforward. One of the effective procedures to quantify LRD is to calculate the value of Hurst parameter. Even though the Hurst parameter is a scalar it cannot be calculated directly, it can be only estimated. There are several methods to estimate the Hurst parameter and sometimes they produce the conflicting results. Ineptly, it is not clear which method provides the most accurate estimation \[[11](#bib.bib11), [10](#bib.bib10)\].
In this paper, we proposed the LRD and self-similarity indicators of delayed correlations in network traffic. We demonstrated, that the time delay invariant behavior of non-edge eigenvalues of $`D\hspace{0pt}(\tau)`$ reflects the self-similar nature of delayed correlations. Meanwhile, the scaling with time of edge eigenvalues or their lagged-time dependence is an exhibition of self-similarity of delay correlations.
In addition, we established that the IPR for eigenvectors of lagged correlations are concise parameters of realistic model for network congestion pattern. As was shown in \[[1](#bib.bib1)\], IPR for $`D\hspace{0pt}(0)`$ contains two localization trends in network interactions, i.e. two regions in spectrum, traced back to a small number of time series, which create the bottleneck at the routers. It is noteworthy, that the IPR for $`{D\hspace{0pt}(\tau)},`$ where $`\tau > 0`$, reveals the third localization trend, which has different origin. The significantly increased and time delay invariant IPR around the median eigenvalue indicates presence of lead-lag relationship between time series.
With experiments altering the original traffic time series several distinctive effects has been uncovered. First of all, we demonstrated that tempering with time series has no effect on self-similar transient behavior of eigenvalues and IPRs, located in the middle segment of the spectrum. By contrast, both stochastic and periodic injections into the right (non-random) and middle (random) segments respectively yielded dramatic changes in chosen indicators. In particular, we recorded the destructive effect of random noise on otherwise simplistic double-peaked power spectra.
One of the main results we obtained from periodic injection experiments was presence of resonance phenomenon. When the period of injection coincides with one of the characteristic time scales of the network (i.e. oscillation periods of edge eigenvalues) the corresponding spectral peak gets enhanced. The Fourier transform peak, corresponding to the other scale gets suppressed and sometimes even annihilated. Finally, injection with the period much less than both scales has little effect on Fourier spectra, while period of the same order in magnitude rearranges the original spectra completely.
The above described time-lagged correlational analysis has a broad area of applications, where delayed correlations between system substructures are essential. For instance, it can be applied to electro-physiological time series of brain response \[[13](#bib.bib13)\], earthquake relocations \[[14](#bib.bib14)\], financial portfolios \[[3](#bib.bib3), [6](#bib.bib6)\], and atmospheric data \[[3](#bib.bib3)\]. To support this assertion we point out that edge eigenvalues of $`D\hspace{0pt}(\tau)`$ behave almost identically to these of atmospheric data, while the delay eigenvalues of the stock market data act just like the eigenvalues, we termed random \[[3](#bib.bib3)\].
## acknowledgment
This research was partially supported by a grant from the US Department of Treasury through a subcontract from the University of Kentucky.
## References
- \[1\] V. Rojkova and M. Kantardzic, Analysis of inter-domain traffic correlations: random ratrix rheory approach, arXiv.org:cs/0706.2520 (2007-06-21).
- \[2\]
M. Barthelemy, B. Gondran and E. Guichard, Large scale cross-correlations in internet traffic, arXiv:cond0mat/0206185 vol 2 3 Dec 2002.
- \[3\] K.B.K. Mayya and R.E. Amritkar, Analysis of delay correlation matrices, oai:arXiv.org:cond-mat/0601279 (2006-12-20).
- \[4\]
T. Guhr, A. Muller-Groeling, and H.A. Weidenmuller, Random matrix theories in quantum physics: common concepts, Phys. Rep. 299, 190 (1998).
- \[5\] Z. Burda, A. Jarosz, J. Jurkiewicz, M.A. Nowak, G. Papp, I. Zahed, Applying free random variables to random matrix analysis of financial data, cond-mat/0603024 (2006).
- \[6\]
C. Biely and S. Thurner, Random matrix ensemble of time-lagged correlation matrices: derivation of eigenvalue spectra and analysis of financial time-series, arXiv:physics/0609053 vol 1 7 Sep 2006.
- \[7\]
A.W. Lo and A.C MacKinlay, When are contrarian profits due to stock market over-reaction?, Review of Financial Studies, 3, 175-206 (1990).
- \[8\]
T. Chordia and B. Swaminathan, Trading volume and cross-autocorrelations in stock returns, Journal of Finance, 55, 913-936 (2000).
- \[9\]
A. Erramilli, O. Narayan, and W. Willinger, Experimental queuing analysis with long-range dependent packet traffic, IEEE/ACM Trans. Networking, vol 4, 2, 1996, 209-223.
- \[10\] M. Krunz, On the limitations of the variance-time test for inference of long-range dependence, IEEE INFOCOM, 2001, 1254-1260.
- \[11\] S. Molnar and T.D. Dang, Pitfalls in long range dependence testing and estimation. GLOBECOM, 2000.
- \[12\] T. Karagiannis, M. Molle, and M. Faloutsos, Long-range dependence, ten years of Internet traffic modeling, IEEE Internet Computing, Oct 2004.
- \[13\]
J. Kwapien, S. Drozdz, and A.A. Ioannides, Temporal correlations versus noise in the correlation matrix formalism: an example of the brain auditory response, arXiv:cond-mat/0002175, vol 1, 11 Feb 2000.
- \[14\]
W.-X. Du, C. H. Thurber, and D. Eberhart-Phillips, Earthquake relocation using cross-correlation time delay estimates verified with the bispectrum method, Bulletin of the Seismological Society of America, June 2004, vol 94, 3, 856-866.
- \[15\]
W.E. Leland et al., On the self-similar nature of Ethernet traffic, IEEE/ACM Trans. Networking, vol 2, 1, 1994, 1-15. |
0708.1579 | cs.NI | cyber | # Homogeneous temporal activity patterns in a large online communication space
\nameAndreas Kaltenbrunner \emailandreas.kaltenbrunner@upf.edu
\nameVicenç Gómez
\nameAyman Moghnieh
\nameRodrigo Meza
\nameJosep Blat
\nameVicente López
\addrDepartament de les Tecnologies de la Informació i les comunicacions
Universitat Pompeu Fabra
Passeig de Circumval·lació 8, 08003 Barcelona, Spain
Barcelona Media Centre d’Innovació
Ocata 1, 08003 Barcelona, Spain
###### Abstract
The many-to-many social communication activity on the popular technology-news website Slashdot has been studied. We have concentrated on the dynamics of message production without considering semantic relations and have found regular temporal patterns in the reaction time of the community to a news-post as well as in single user behavior. The statistics of these activities follow log-normal distributions. Daily and weekly oscillatory cycles, which cause slight variations of this simple behavior, are identified. A superposition of two log-normal distributions can account for these variations. The findings are remarkable since the distribution of the number of comments per users, which is also analyzed, indicates a great amount of heterogeneity in the community. The reader may find surprising that only a few parameters allow a detailed description, or even prediction, of social many-to-many information exchange in this kind of popular public spaces.
Keywords: Social interaction, information diffusion, log-normal activity, heavy tails, Slashdot
## 1 Introduction
Nowadays, an important part of human activity leaves electronic traces in form of server logs, e-mails, loan registers, credit card transactions, blogs, etc. This huge amount of generated data allows to observe human behavior and communication patterns at nearly no cost on a scale and dimension which would have been impossible some decades ago. A considerable number of studies have emerged in recent years using some part of these data to investigate the time patterns of human activity. The studied temporal events are rather diverse and reach from directory listings and file transfers (FTP requests) (Paxson and Floyd,, [1995](#bib.bib25)), job submissions on a supercomputer (Kleban and Clearwater,, [2003](#bib.bib16)), arrival times of consecutive printing-job submissions (Harder and Paczuski,, [2006](#bib.bib11)) over trades in bond (Mainardi et al.,, [2000](#bib.bib20)) or currency futures (Masoliver et al.,, [2003](#bib.bib22)) to messages in Internet chat systems (Dewes et al.,, [2003](#bib.bib6)), online games (Henderson and Bhatti,, [2001](#bib.bib12)), page downloads on a news site (Dezsö et al.,, [2006](#bib.bib7)) and e-mails (Johansen,, [2004](#bib.bib13)). A common characteristic of these studies is that the observed probability distributions for the waiting or inter-event times are heavy tailed. In other words, if the response time ever exceeds a large value, then it is likely to exceed any larger value as well (Sigman,, [1999](#bib.bib27)). A recent study (Barabási,, [2005](#bib.bib4)) tries to explain this behavior under the assumption that these heavy tailed distributions can be well approximated by a power-law or at least by a power-law with an exponential cut-off (Newman,, [2005](#bib.bib24)). The cited study presents a model which seems to explain the distribution of e-mail response times and has been used later to account for the inter-event times of web-browsing, library loans, trade transactions and correspondence patterns of letters (Vázquez et al.,, [2006](#bib.bib29)). However, the hypothesis of a power-law distribution is not generally accepted, at least in case of e-mail response times. Stouffer et al., ([2006](#bib.bib28)) claim that the data can be much better fitted with either a log-normal (LN) distribution (Limpert et al.,, [2001](#bib.bib19)) or the superposition of two LN. This debate has been repeated across many areas of science for decades, as noticed by Mitzenmacher, ([2003](#bib.bib23)).
To the authors’ knowledge no study of this type has been performed on systems where social interaction occurs in a more complex manner than just person to person (one-to-one) communication. We think it is valuable to analyze the temporal patterns of the many-to-many social interaction on a technology-related news-website which supports user participation. We have chosen Slashdot
<sup>1</sup>
<sup>1</sup>1http://www.slashdot.org
, a popular website for people interested in reading and discussing about technology and its ramifications. It gave name to the “Slashdot effect” (Adler,, [1999](#bib.bib1)), a huge influx of traffic to a hosted link during a short period of time, causing it to slow down or even to temporarily collapse.
Slashdot was created at the end of 1997 and has ever since metamorphosed into a website that hosts a large interactive community capable of influencing public perceptions and awareness on the topics addressed. Its role can be metaphorically compared to that of commercial malls in developed markets, or hubs in intricate large networks. The site’s interaction consists of short-story posts that often carry fresh news and links to sources of information with more details. These posts incite many readers to comment on them and provoke discussions that may trail for hours or even days. Most of the commentators register and comment under their nicknames, although a considerable amount participates anonymously.
Although Slashdot allows users to express their opinion freely, moderation and meta-moderation mechanisms are employed to judge comments and enable readers to filter them by quality. The moderation system was analyzed by Lampe and Resnick, ([2004](#bib.bib17)) who concluded that it upholds the quality of discussions by discouraging spam and offensive comments, marking a difference between Slashdot and regular discussion forums. This high quality social interaction has prompted several socio-analytical studies about Slashdot. Poor, ([2005](#bib.bib26)) and Baoill, ([2000](#bib.bib3)) have both conducted independent inquiries on the extent to which the site represents an online public sphere as defined by Habermas, ([1989](#bib.bib10)).
Given that a great amount of users with different interests and motivations participates in discussions about very different topics, one would expect to observe a high degree of heterogeneity on a site like Slashdot. However, what if the posts and comments were analyzed just as imprints of an occurring information exchange, with no regard to semantic aspects? Is there a homogeneous behavior pattern underlying heterogeneity? To answer these and related questions we collected and studied one year’s worth of interchanged messages along with the associated meta-data from Slashdot. We show here that the temporal patterns of the comments provoked by a post are very similar, indicating that homogeneity is the rule not the exception. The temporal patterns of the social activity fit accurately log-normal distributions, thus giving empirical evidence of our hypothesis and establishing a link with previous studies where social interaction occurs in a simpler way.
Finally, our analysis allows more insight into questions such as: is there a time-scale common to all discussions, or are they scale-free? What does incite a user to write a comment, is it the relevance of the topic, or maybe just the hour of the day? Can we predict the amount of activity a post will trigger already some minutes after it has been written? Which type of applications can we devise on the basis of using these conclusions?
The rest of the article is organized as follows: In section [2](#S2 "2 Methods ‣ Homogeneous temporal activity patterns in a large online communication space") we briefly explain the process of data acquisition. We then present the results in section [3](#S3 "3 Results ‣ Homogeneous temporal activity patterns in a large online communication space") providing first an overview of the global activity and then explaining our analysis in detail. We finish the paper with section [4](#S4 "4 Discussion ‣ Homogeneous temporal activity patterns in a large online communication space") where we discuss the results.
## 2 Methods
In this section we explain the methods used to crawl and analyze Slashdot. The crawled
<sup>2</sup>
<sup>2</sup>2Software used: wget, Perl scripts, and Tidy on a GNU/Linux, Ubuntu 6.0.6 OS.
data correspond to posts and comments published between August $`26`$th, $`2005`$ and August $`31`$th, $`2006`$. We divided the crawling process into two stages. The first stage included crawling the main HTML (posts) and first level comments and the second stage covered all additional comment pages. Crawling all the data took $`4.5`$ days and produced approximately $`4.54`$ GB of data. Post-processing caused by the presence of duplicated comments was necessary (due to an error of representation on the website). Although a high amount of information was extracted from the raw HTML (sub-domains, title, topics, hierarchical relations between comments) we concentrated only on a minimal amount of information: type of contribution (either post or comment), its identifier, author’s identifier and time-stamp or date of publishing. The selected information was extracted to XML-files and imported into Matlab where the statistical analysis was performed. Table [1](#S2.T1 "Table 1 ‣ 2 Methods ‣ Homogeneous temporal activity patterns in a large online communication space") shows the main quantities of the crawling and the extracted data.
Table 1: Main quantities of crawling and retrieved data.
<table>
<thead>
<tr>
<th>Period covered</th>
<th>262626-888-050505 −- 313131-888-060606</th>
</tr>
</thead>
<tbody>
<tr>
<th>Time needed for crawling</th>
<td>4.54.54.5 days</td>
</tr>
<tr>
<th>Amount of data mined</th>
<td>4.544.544.54 GB</td>
</tr>
<tr>
<th>Posts</th>
<td>100161001610016</td>
</tr>
<tr>
<th>Comments</th>
<td>207508520750852075085</td>
</tr>
<tr>
<th>Commentators</th>
<td>936369363693636</td>
</tr>
<tr>
<th>Anonymous comments</th>
<td>18.6%percent18.618.6\%</td>
</tr>
</tbody>
</table>
The time-stamps of post and comments can be obtained from Slashdot with minute-precision and corresponded to the EDT time zone ($`=`$ GMT$`- 4`$ hours). They allow to calculate the following two quantities:
The Post-Comment-Interval (PCI) stands for the difference between the time-stamps of a comment and its corresponding post.
The Inter-Comment-Interval (ICI) refers to the difference between the time-stamps of two consecutive comments of the same user (no matter what post he/she comments on).
## 3 Results
In this section we first give an overview of the global activity looking at the data on different temporal scales and analyzing some relations between variables of interest. We then focus on the activity provoked by single posts and analyze the behavior of single users, concentrating on the most active ones.
### 3.1 Global cyclic activity
As previously explained, comments can be considered as reactions triggered by the publishing of posts. This difference in nature between both types of contributions justifies a separate analysis of their dynamics.
[Refer to caption](/html/0708.1579/assets/x1.png)
[Refer to caption](/html/0708.1579/assets/x2.png)
Figure 1: (a) Weekly and (b) daily activity cycles.
Figure [1](#S3.F1 "Figure 1 ‣ 3.1 Global cyclic activity ‣ 3 Results ‣ Homogeneous temporal activity patterns in a large online communication space") shows (normalized) mean activity and standard deviations of both posts and comments. It illustrates patterns in agreement with the social activity outside the public sphere. Figure [1](#S3.F1 "Figure 1 ‣ 3.1 Global cyclic activity ‣ 3 Results ‣ Homogeneous temporal activity patterns in a large online communication space")a shows regular, steady activity during working days which slows down during weekends. This weekly cycle is interleaved by daily oscillations illustrated in Figure [1](#S3.F1 "Figure 1 ‣ 3.1 Global cyclic activity ‣ 3 Results ‣ Homogeneous temporal activity patterns in a large online communication space")b. The daily activity cycle reaches its maximum at $`1`$pm approximately and its minimum during the night between $`3`$am and $`4`$am. Although Slashdot is open to public access around the world, we see that its activity profile is clearly biased towards the American time-schedule.
Interestingly, although post activity shows more fluctuations and higher standard deviations than comment activity, there is little discrepancy between their mean temporal profiles. This difference in the deviations is not surprising given the greater number of comments (see Table [1](#S2.T1 "Table 1 ‣ 2 Methods ‣ Homogeneous temporal activity patterns in a large online communication space")). We notice that the standard deviations of the daily post- and commenting activities also show similar cyclic behavior (Figure [1](#S3.F1 "Figure 1 ‣ 3.1 Global cyclic activity ‣ 3 Results ‣ Homogeneous temporal activity patterns in a large online communication space")b).
[Refer to caption](/html/0708.1579/assets/x3.png)
Figure 2: Histogram of the number of comments per post (inset shows the corresponding cdf).
### 3.2 Post-induced activity
In this section we analyze the activity (comments) a post induces on the site. The histogram of Figure [2](#S3.F2 "Figure 2 ‣ 3.1 Global cyclic activity ‣ 3 Results ‣ Homogeneous temporal activity patterns in a large online communication space") gives an idea of the number of comments the posts receive. Note that half of the posts provoke more than $`160`$ comments and some of them even trigger more than $`1000`$. To analyze the time-distribution of these comments we study their post-comment intervals (PCIs).
[Refer to caption](/html/0708.1579/assets/x4.png)
Figure 3: LN-approximation (dashed lines) of the PCI-distribution (solid lines and bars) of a post which received $`1341`$ comments. (a) Comments per minutes (bin-with$`= 2`$ for better visualization) for the first $`200`$ minutes after the post has been published. (b) Same as (a) in logarithmic scale. (c) The cumulative distribution of the data shown in (a). Inset shows a zoom on the first $`2000`$ minutes. (d) Same as (c) in logarithmic scale.
#### 3.2.1 Analysis of the activity generated by a single post
We are especially interested in the resulting probability distribution of all the PCIs of a certain post. This distribution reveals us the probability for a post to receive a comment $`t`$ minutes after it has been published. Figures [3](#S3.F3 "Figure 3 ‣ 3.2 Post-induced activity ‣ 3 Results ‣ Homogeneous temporal activity patterns in a large online communication space")a and [3](#S3.F3 "Figure 3 ‣ 3.2 Post-induced activity ‣ 3 Results ‣ Homogeneous temporal activity patterns in a large online communication space")b show this distribution for a post which provoked $`1341`$ comments. Although there are some important fluctuations, the characteristic shape of the probability density function (pdf) resembles a LN-distribution. This becomes even clearer if the cumulative probability distribution (cdf) is observed, since there the fluctuations of the pdf are averaged out. Figures [3](#S3.F3 "Figure 3 ‣ 3.2 Post-induced activity ‣ 3 Results ‣ Homogeneous temporal activity patterns in a large online communication space")c and [3](#S3.F3 "Figure 3 ‣ 3.2 Post-induced activity ‣ 3 Results ‣ Homogeneous temporal activity patterns in a large online communication space")d show a good fit of the PCI-cdf of the data with the cdf of the LN-distribution. To quantify the quality of the fit we have used a normalized error measure $`\epsilon`$ based on the $`\ell^{1}`$-norm (see Appendix [B](#S2a "B Error Measure ϵ ‣ Homogeneous temporal activity patterns in a large online communication space")). For the post shown in Figure [3](#S3.F3 "Figure 3 ‣ 3.2 Post-induced activity ‣ 3 Results ‣ Homogeneous temporal activity patterns in a large online communication space") we obtain $`\epsilon = 0.007`$, meaning that the average error is below $`1\%`$.
[Refer to caption](/html/0708.1579/assets/x5.png)
[Refer to caption](/html/0708.1579/assets/x6.png)
[Refer to caption](/html/0708.1579/assets/x7.png)
Figure 4: LN-approximation of the PCI-distribution of $`3`$ different posts.
The PCI-cdf of three more posts can be observed in Figure [4](#S3.F4 "Figure 4 ‣ 3.2.1 Analysis of the activity generated by a single post ‣ 3.2 Post-induced activity ‣ 3 Results ‣ Homogeneous temporal activity patterns in a large online communication space"). The top two sub-figures show good fits, indicating that the PCI is well approximated even for a small number of comments. However, the fit is not that accurate for all posts. E.g. the comments of the post shown in Figure [4](#S3.F4 "Figure 4 ‣ 3.2.1 Analysis of the activity generated by a single post ‣ 3.2 Post-induced activity ‣ 3 Results ‣ Homogeneous temporal activity patterns in a large online communication space") (bottom) start to show considerable different behavior from the expected LN-approximation about $`3`$ hours after its publication. The activity is lower than predicted, but starts to increase again at about $`6`$am in the morning the next day. At around $`8`$:$`30`$pm it increases further to recover the lost activity during the night. More such oscillations of activity can be observed during the following days. The time-spans of variations in activity coincide quite exactly with the average daily activity cycle shown in Figure [1](#S3.F1 "Figure 1 ‣ 3.1 Global cyclic activity ‣ 3 Results ‣ Homogeneous temporal activity patterns in a large online communication space")b. We analyze this coincidence further in the next section.
[Refer to caption](/html/0708.1579/assets/x8.png)
[Refer to caption](/html/0708.1579/assets/x9.png)
Figure 5: (a) Errors $`\epsilon`$ of the LN-approximation of the PCI-cdf (bin-width $`= 10^{- 3}`$). Inset shows the corresponding cdf. (b) Dependence of mean and median of the approximation error $`\epsilon`$ on the hour the post is published.
#### 3.2.2 Approximation quality
With the LN shape of the PCI-distribution identified, we focus on the quality of this approximation in general. We therefore calculate the error measure $`\epsilon`$ of the fit for all posts which received comments. The resulting distribution of $`\epsilon`$ can be seen in Figure [5](#S3.F5 "Figure 5 ‣ 3.2.1 Analysis of the activity generated by a single post ‣ 3.2 Post-induced activity ‣ 3 Results ‣ Homogeneous temporal activity patterns in a large online communication space")a. For $`87`$% of the posts the approximation error $`\epsilon`$ is lower than $`0.05`$, and for $`29`$% of them lower than $`0.02`$.
If we take a closer look at the data, we notice a dependence of $`\epsilon`$ on the publishing-hour of a post (Figure [5](#S3.F5 "Figure 5 ‣ 3.2.1 Analysis of the activity generated by a single post ‣ 3.2 Post-induced activity ‣ 3 Results ‣ Homogeneous temporal activity patterns in a large online communication space")b). The best fit is reached when the post is published between $`6`$am and $`11`$am. Then the mean error increases successively until $`11`$pm to stay high during the night and recover again in the early morning. This behavior can be understood looking at the daily activity cycle (Figure [1](#S3.F1 "Figure 1 ‣ 3.1 Global cyclic activity ‣ 3 Results ‣ Homogeneous temporal activity patterns in a large online communication space")b). The less time the community has to comment on a post during the time-window of high activity, the greater is the need to comment on it the next time the high activity phase is reached, and hence the expected LN behavior is altered. Figure [4](#S3.F4 "Figure 4 ‣ 3.2.1 Analysis of the activity generated by a single post ‣ 3.2 Post-induced activity ‣ 3 Results ‣ Homogeneous temporal activity patterns in a large online communication space") (bottom) gives an example of such a late post (published at $`10`$:$`35`$pm).
#### 3.2.3 Approximation with double log-normal distributions
We approximate the data as well with a double log-normal distribution (DLN), i.e. a superposition of two LN-distributions (See appendix [A](#S1a "A Log-normal and double log-normal distributions ‣ Homogeneous temporal activity patterns in a large online communication space")). To find their parameters and especially their mixing coefficient, we use maximum likelihood estimation (Stouffer et al.,, [2006](#bib.bib28); DeGroot and Schervish,, [2002](#bib.bib5)). The DLN should lead to better results in general and reduce the dependency on the circadian rhythm since it represents two waves of activity: one starting when the post is published and another being caused by the next increase of activity in the circadian cycle.
[Refer to caption](/html/0708.1579/assets/x10.png)
Figure 6: Comparison of LN and DLN-approximations (dashed-dotted lines) of the PCI-distribution (solid lines and bars) of a post which received $`1567`$ comments. The DLN-distribution is a superposition of LN<sub>1</sub> and LN<sub>2</sub>, which in the above figure are rescaled according to the coefficient $`c`$ of the DLN. Rest of legend as in Figure [3](#S3.F3 "Figure 3 ‣ 3.2 Post-induced activity ‣ 3 Results ‣ Homogeneous temporal activity patterns in a large online communication space")
.
An example of this behavior is shown in Figure [6](#S3.F6 "Figure 6 ‣ 3.2.3 Approximation with double log-normal distributions ‣ 3.2 Post-induced activity ‣ 3 Results ‣ Homogeneous temporal activity patterns in a large online communication space") where we compare LN and DLN-approxima·tion of the same post as used in Figure [4](#S3.F4 "Figure 4 ‣ 3.2.1 Analysis of the activity generated by a single post ‣ 3.2 Post-induced activity ‣ 3 Results ‣ Homogeneous temporal activity patterns in a large online communication space") (bottom). The red and blue lines indicate the two log-normals whose superposition results in a DLN (gray, dashed-dotted), which clearly outperforms the previous LN (black, dashed) approach. The error $`\epsilon`$ decreases from $`0.031`$ to $`0.009`$ and the approximation is much closer to the cdf of the data (black continuous line in Figures [6](#S3.F6 "Figure 6 ‣ 3.2.3 Approximation with double log-normal distributions ‣ 3.2 Post-induced activity ‣ 3 Results ‣ Homogeneous temporal activity patterns in a large online communication space")c and [6](#S3.F6 "Figure 6 ‣ 3.2.3 Approximation with double log-normal distributions ‣ 3.2 Post-induced activity ‣ 3 Results ‣ Homogeneous temporal activity patterns in a large online communication space")d). We notice that the first $`10`$ hours of activity are well approximated by a single LN-distribution (red line). Then the activity increases due to the high phase of the circadian cycle (compare also with the labels of Figure [4](#S3.F4 "Figure 4 ‣ 3.2.1 Analysis of the activity generated by a single post ‣ 3.2 Post-induced activity ‣ 3 Results ‣ Homogeneous temporal activity patterns in a large online communication space") bottom). The second LN distribution (blue line) accounts for this increase and therefore the DLN-approximation reflects the first bump in the PCI-cdf and fits well the data.
[Refer to caption](/html/0708.1579/assets/x11.png)
[Refer to caption](/html/0708.1579/assets/x12.png)
Figure 7: (a) Errors $`\epsilon`$ of the DLN-approximation of the PCI-cdf (bin-width $`= 10^{- 3}`$). Inset shows the corresponding cdf. (b) Dependence of mean and median of the approximation error $`\epsilon`$ on the hour the post is published.
To quantify the overall performance of a DLN-fit we apply it on all posts and plot the distribution of its approximation error $`\epsilon`$ in Figure [7](#S3.F7 "Figure 7 ‣ 3.2.3 Approximation with double log-normal distributions ‣ 3.2 Post-induced activity ‣ 3 Results ‣ Homogeneous temporal activity patterns in a large online communication space")a. The inset compares the error-cdfs of DLN (continuous) and LN-approach (dashed-dotted). We notice a significant improvement of the approximation quality. For example, the error of the DLN-fits is below $`0.02`$ for more than $`80\%`$ of the posts compared to only $`29\%`$ in the case of LN-approximations. Figure [7](#S3.F7 "Figure 7 ‣ 3.2.3 Approximation with double log-normal distributions ‣ 3.2 Post-induced activity ‣ 3 Results ‣ Homogeneous temporal activity patterns in a large online communication space")b shows only a minor dependency of the quality of the DLN-fits on the publishing hour of the post (compare with Figure [5](#S3.F5 "Figure 5 ‣ 3.2.1 Analysis of the activity generated by a single post ‣ 3.2 Post-induced activity ‣ 3 Results ‣ Homogeneous temporal activity patterns in a large online communication space")b), which allows us to conclude that the DLN-distributions accounts for the major part of the aberration of the log-normal behavior caused by the circadian cycle.
#### 3.2.4 Approximation parameters
For the cases where a LN-distribution leads to good results we can describe the activity triggered by a post with only two parameters: the median
<sup>3</sup>
<sup>3</sup>3Note that the median coincides with the geometric mean for a log-normally distributed random variable.
and the geometric standard deviation $`\sigma_{g}`$ of the PCI-pdf, commonly used to compare log-normally distributed quantities (Limpert et al.,, [2001](#bib.bib19)). The median and $`\sigma_{g}`$ relate to the parameters of the LN-distribution in the following way.
<table>
<tbody>
<tr>
<td></td>
<td>median</td>
<td>=exp(μ),absent𝜇\displaystyle=\exp(\mu)~{},</td>
<td>σgsubscript𝜎𝑔\displaystyle\sigma_{g}</td>
<td>=exp(σ).absent𝜎\displaystyle=\exp(\sigma).</td>
<td></td>
<td>(1)</td>
</tr>
</tbody>
</table>
[Refer to caption](/html/0708.1579/assets/x13.png)
[Refer to caption](/html/0708.1579/assets/x14.png)
Figure 8: (a) Histograms of the estimates of medians (bin-width $`= 10`$) and geometric standard deviations (inset, bin-width $`= 0.1`$) of the PCI-distributions. (b) Parameters of LN and DLN-approximations. Bin-width=$`0.1`$ for $`\mu`$ and $`\sigma`$, $`0.01`$ for $`c`$ (inset).
Figure [8](#S3.F8 "Figure 8 ‣ 3.2.4 Approximation parameters ‣ 3.2 Post-induced activity ‣ 3 Results ‣ Homogeneous temporal activity patterns in a large online communication space")a shows the distribution of these quantities for all posts
<sup>4</sup>
<sup>4</sup>4Instead of calculating $`\sigma_{g}`$ directly from the data as in a previous version of this study ([Kaltenbrunner et al., 2007b,](#bib.bib15) ) we used equation ([1](#S3.E1 "In 3.2.4 Approximation parameters ‣ 3.2 Post-induced activity ‣ 3 Results ‣ Homogeneous temporal activity patterns in a large online communication space")) and the estimates of $`\sigma`$, which led to different results. Compare also with Limpert et al., ([2001](#bib.bib19)).
. The inset shows the distribution of $`\sigma_{g}`$, which is centered around $`4.5`$ and has a standard deviation of $`0.91`$. The median of the post-induced activity on the other hand shows more variations, but is rather short (for $`50`$% of the posts it is below $`2.5`$ hours, for $`90`$% below $`6`$ hours) compared to the maximum PCI (approx. $`12`$ days). We can thus conclude that although the total activity a post generates covers a large time interval, the major part of the activity happens within the first few hours after the post’s publication.
If we use a DLN-distribution to approximate the data we need five parameters. Their distributions together with those of the parameters $`\sigma`$ and $`\mu`$ of the LN-approximation are displayed in Figure [8](#S3.F8 "Figure 8 ‣ 3.2.4 Approximation parameters ‣ 3.2 Post-induced activity ‣ 3 Results ‣ Homogeneous temporal activity patterns in a large online communication space")b. For better visualization we choose a stair plot instead of a bar-graph. Clearly the regions of $`\mu_{1}`$ (continuous line with circles) and $`\sigma_{1}`$ (continuous line) are very similar to those of the parameters of LN-approximations (dashed-dotted lines), indicating that the first one of the two log-normal distributions used to generate the DLN is similar to the LN-approximations. The parameters $`\mu_{2}`$ and $`\sigma_{2}`$, on the other hand, show an interesting bimodal behavior. One of the two peaks of the distribution falls within the regions of $`\mu_{1}`$ or $`\sigma_{1}`$ respectively. Those cases correspond to posts for which the two superposed log-normal distributions are very similar and the data fits well already a single LN-distribution. The second peak in the $`\mu_{2}`$-distribution represents those posts which provoke a second wave of activity due to the circadian cycle. In those cases the parameter $`\sigma_{2}`$ is usually smaller than $`\sigma_{1}`$. The inset of Figure [8](#S3.F8 "Figure 8 ‣ 3.2.4 Approximation parameters ‣ 3.2 Post-induced activity ‣ 3 Results ‣ Homogeneous temporal activity patterns in a large online communication space")b shows the mixing parameter $`c`$, which is nearly uniformly distributed although values in $`\lbrack 0.7,1\rbrack`$ are slightly more likely than lower ones. We sorted the parameters to ensure a value of $`c \geq 0.5`$.
### 3.3 User dynamics
In this section we analyze the activity on Slashdot taking the authorship of the comments into account. We first study the distribution of activity among all the users participating in the debates and then focus on the temporal activity patterns of single users.
#### 3.3.1 Global user activity
The activity of all users is best illustrated by the distribution of the number of comments per user. It is shown in double-logarithmic scale in Figure [9](#S3.F9 "Figure 9 ‣ 3.3.1 Global user activity ‣ 3.3 User dynamics ‣ 3 Results ‣ Homogeneous temporal activity patterns in a large online communication space")a. The obtained distribution follows quite closely a straight line, suggesting a power-law probability distribution governing this relation. We note that $`53\%`$ of the users write $`3`$ or less comments whereas only $`93`$ users ($`0.1\%`$) write more than $`1000`$ comments. Indeed, after applying linear regression as in other studies (Faloutsos et al.,, [1999](#bib.bib8); Albert et al.,, [1999](#bib.bib2)) we obtain a quite large correlation coefficient $`R^{2} = {- 0.97}`$ for an exponent of $`\gamma = {- 1.79}`$.
[Refer to caption](/html/0708.1579/assets/x15.png)
[Refer to caption](/html/0708.1579/assets/x16.png)
Figure 9: (a) Histogram of the number of comments per user and (b) and its corresponding cdf.
However, if we apply rigorous statistical analysis as proposed by Goldstein et al., ([2004](#bib.bib9)) the picture changes. First, we estimate the power-law exponent computing the less biased maximum likelihood estimator (MLE). The resulting exponent $`\gamma = {- 1.5}`$ differs significantly from the previous one and is illustrated in Figure [9](#S3.F9 "Figure 9 ‣ 3.3.1 Global user activity ‣ 3.3 User dynamics ‣ 3 Results ‣ Homogeneous temporal activity patterns in a large online communication space") (dashed-line). Although Figure [9](#S3.F9 "Figure 9 ‣ 3.3.1 Global user activity ‣ 3.3 User dynamics ‣ 3 Results ‣ Homogeneous temporal activity patterns in a large online communication space")a tempts one to accept the power-law hypothesis, the cdf shown in Figure [9](#S3.F9 "Figure 9 ‣ 3.3.1 Global user activity ‣ 3.3 User dynamics ‣ 3 Results ‣ Homogeneous temporal activity patterns in a large online communication space")b discards it. It is thus not surprising that the Kolmogorov-Smirnov test forces us to reject the power-law hypothesis with statistical significance at the $`0.1\%`$ level.
As an alternative hypothesis to describe the data we propose a truncated LN probability distribution, shown in Figure [9](#S3.F9 "Figure 9 ‣ 3.3.1 Global user activity ‣ 3.3 User dynamics ‣ 3 Results ‣ Homogeneous temporal activity patterns in a large online communication space") as grey-solid-line. Its parameters are found using the MLE. Clearly, the fit is better using this hypothesis. We remark that in many studies some data points (considered outliers) are discarded to improve the power-law fit. Here, in contrast, the truncated LN-approximation can characterize the entire data-set.
#### 3.3.2 Single user dynamics
After characterizing the user activity at a general level, we investigate the temporal behavior patterns of single users . The analysis concentrates on the two most active users (to protect their privacy we call them user$`1`$ and user$`2`$). Table [2](#S3.T2 "Table 2 ‣ 3.3.2 Single user dynamics ‣ 3.3 User dynamics ‣ 3 Results ‣ Homogeneous temporal activity patterns in a large online communication space") shows the number of commented posts and the total number of comments these two users published during the time-span covered by our data.
Table 2: Contributions of the two most active users.
<table>
<thead>
<tr>
<th></th>
<th>user1</th>
<th>user2</th>
</tr>
</thead>
<tbody>
<tr>
<th>commented posts</th>
<td>118911891189</td>
<td>130613061306</td>
</tr>
<tr>
<th>comments</th>
<td>364236423642</td>
<td>335033503350</td>
</tr>
</tbody>
</table>
We focus on the distribution of the PCIs of all of their comments as well as on their inter-comment-interval (ICI) distribution, i.e. the time-difference between two comments of the same user.
We approximate the PCI-cdf (gray lines in Figure [10](#S3.F10 "Figure 10 ‣ 3.3.2 Single user dynamics ‣ 3.3 User dynamics ‣ 3 Results ‣ Homogeneous temporal activity patterns in a large online communication space")a) also with LN (dashed and dashed-dotted lines) and DLN-distributions (blue and red lines with box and circle markers). The quality of the LN-fit is worse than in the case of the post-induced comment activity, but the DLN-distribution is a good explanation of the data with a small approximation error $`\epsilon`$. Again we notice a clear dependence of the quality of the fit on the activity cycle (shown in the insets of Figure [10](#S3.F10 "Figure 10 ‣ 3.3.2 Single user dynamics ‣ 3.3 User dynamics ‣ 3 Results ‣ Homogeneous temporal activity patterns in a large online communication space")a). The approximation is much better for user1, whose daily and especially weekly activity cycles are much more balanced than those of user2. The activity of the latter user concentrates almost exclusively on the working hours from Monday to Friday. Hence his PCI-distribution shows a clear decrease after $`8`$ but increases again after $`16`$ hours. This increase is less pronounced if only the first comment to a post is considered (data not shown), indicating that the user frequently rechecks the posts he commented the day before to participate again in an ongoing discussion.
[Refer to caption](/html/0708.1579/assets/x17.png)
[Refer to caption](/html/0708.1579/assets/x18.png)
Figure 10: Activity patterns of the two most active users: (a) PCI-distributions, insets shows daily and weekly activity cycles. (b) Distribution of the inter-comment intervals (ICI) compared with the whole population (dashed line).
The same effect can be observed in their ICIs, which are illustrated in Figure [10](#S3.F10 "Figure 10 ‣ 3.3.2 Single user dynamics ‣ 3.3 User dynamics ‣ 3 Results ‣ Homogeneous temporal activity patterns in a large online communication space")b. There the cdf (inset of Figure [10](#S3.F10 "Figure 10 ‣ 3.3.2 Single user dynamics ‣ 3.3 User dynamics ‣ 3 Results ‣ Homogeneous temporal activity patterns in a large online communication space")b) of user$`2`$ shows an even more pronounced increase around an ICI of $`16`$ hours. We further observe that the ICI-pdf peaks for both users as well as for the whole population at $`3`$ minutes. This is probably caused by an anti-troll filter (Malda,, [2002](#bib.bib21)), which should prevent a user from commenting more than once within $`120`$ seconds. The medians of the ICI-distributions of user1 and user2 are rather short ($`11`$ and $`7`$ minutes respectively) compared to the median of the whole population (about $`17`$ hours), indicating that the two users engage in discussions frequently during their activity phase.
## 4 Discussion
The special architecture of the technology-related news website Slashdot allowed us to analyze the temporal communication patterns of an online society without considering semantic aspects. The site activity is driven by news-posts which provoke communication activity in the form of comments.
Despite the great amount of users participating in the discussions, close to $`10^{5}`$ in the data we have studied, and the diversity of themes (games, politics, science, books, etc.) some simple patterns can be identified, which repeat themselves over and over again. One of these patterns appears in the shape of the distribution of time differences between a post and its comments (the PCIs). It can be well approximated by a log-normal distribution (Figures [3](#S3.F3 "Figure 3 ‣ 3.2 Post-induced activity ‣ 3 Results ‣ Homogeneous temporal activity patterns in a large online communication space") and [4](#S3.F4 "Figure 4 ‣ 3.2.1 Analysis of the activity generated by a single post ‣ 3.2 Post-induced activity ‣ 3 Results ‣ Homogeneous temporal activity patterns in a large online communication space")) for most of the posts. The only remarkable deviations from these approximations are caused by oscillatory daily and weekly activity patterns (Figure [1](#S3.F1 "Figure 1 ‣ 3.1 Global cyclic activity ‣ 3 Results ‣ Homogeneous temporal activity patterns in a large online communication space")), which become less noticeable if a post is published early in the morning (Figure [5](#S3.F5 "Figure 5 ‣ 3.2.1 Analysis of the activity generated by a single post ‣ 3.2 Post-induced activity ‣ 3 Results ‣ Homogeneous temporal activity patterns in a large online communication space")a). A significant improvement of the approximation can be achieved using a superposition of two log-normal distributions. Such a double log-normal accounts for the first oscillation caused by the circadian cycle. It can be interpreted as two independent waves of activity, one starting directly after a post has been published, and the second at the next increase of activity due to the circadian rhythm. Although more such oscillations may occur during the life-time of a post, their amplitude is low compared to the first one, suggesting that a combination of more than two LN-distributions would only increase the complexity of parameter-finding (via MLE) without improving significantly the approximation quality. Nevertheless, a combination of a DLN-distribution with an oscillatory function emulating the circadian cycle leads to slightly better results ([Kaltenbrunner et al., 2007a,](#bib.bib14) ), without affecting the complexity of MLE.
In single user behavior an akin pattern appears in the PCI-distribution of all of the comments a user writes to several posts (Figure [10](#S3.F10 "Figure 10 ‣ 3.3.2 Single user dynamics ‣ 3.3 User dynamics ‣ 3 Results ‣ Homogeneous temporal activity patterns in a large online communication space")a). Again deviations are caused by the circadian cycle. Another interesting pattern can be observed analyzing the ICI of single-users, i.e. the time-span between two consecutive comments of a certain user. In the case of the two most active users (Figure [10](#S3.F10 "Figure 10 ‣ 3.3.2 Single user dynamics ‣ 3.3 User dynamics ‣ 3 Results ‣ Homogeneous temporal activity patterns in a large online communication space")b) the ICI-distributions are very similar, which further supports our hypothesis of the existence of homogeneous temporal patterns on Slashdot.
We would expect that the time-spans between publishing and reading of a post also follow log-normal patterns. This could be easily verified checking the server logs of Slashdot or access-times of an external homepage linked by a Slashdot post. Such a study has been performed to show the Slashdot effect (Adler,, [1999](#bib.bib1)), but the scale of the data presented does not allow to draw significant conclusions. Further investigation is needed to verify this claim.
Log-normal temporal patterns similar to those described above were found in person-to-person communication by Stouffer et al., ([2006](#bib.bib28)), who investigated the waiting and inter-event times of an e-mail activity dataset. A second coincidence between their study and our findings is that the number of comments (or e-mails in their case) can be well approximated by the same distribution (a truncated log-normal in this case). The temporal patterns of the e-mail data were previously claimed to show power-law behavior, which would be explained by a queuing model (Barabási,, [2005](#bib.bib4)). Although this model might allow insight into other types of human activity (Vázquez et al.,, [2006](#bib.bib29)) it is not able to account for the observed log-normal behavior patterns. We hope therefore to encourage further research towards a theoretical understanding of the underlying phenomena responsible for this apparently quite general human behavior pattern.
The medians (Figure [8](#S3.F8 "Figure 8 ‣ 3.2.4 Approximation parameters ‣ 3.2 Post-induced activity ‣ 3 Results ‣ Homogeneous temporal activity patterns in a large online communication space")) of the PCI-distributions are very small compared to the overall duration of the activity provoked by a post. Although the posts might be available for commenting for more than $`10`$ days, the first few hours decide whether they will become highly debated or just receive some sporadic comments. We would therefore expect that the simplicity of the approximation together with the high initial activity should make an accurate prediction of the expected user behavior feasible at an early phase after a post has been put online. The accuracy of such forecasting methods is subject of current research ([Kaltenbrunner et al., 2007a,](#bib.bib14) ).
An early characterization of the activity triggered by a post could be applied, for instance, on dynamic pricing or placing of online advertisements or on the improvement of online marketing. The success of a campaign might be predicted already after a short time-period, thus allowing an early adaptation of the strategy of information diffusion. In this context the viral marketing concept (Leskovec et al.,, [2006](#bib.bib18)), which relies on personal communication might be the most promising field.
In our opinion, the regular communication activity patterns described in this work may be relevant in two aspects. The first, simpler one, is related to applications where a better understanding of information trade in the web translates easily into a better description, and even quantification, of Internet audience. But a second, more complex, aspect is related to the human “communicative” behavior uncovered at present time: Internet based communication capabilities. We face a new, large scale, all-to-all public space in which a novel kind of social behavior arises, a scenario that we do not yet fully understand. However, we should not forget that the new activity is being largely recorded and the data can be available for research. The work presented in this contribution is a good example of how those data can be collected and analyzed to give, at least, a quantitative description of the behavior. This is a first step towards a more ambitious target: to develop “ab initio” models for the population dynamics of message interchange, which is also the goal of our current research.
Acknowledgments
This work has been partially funded by Càtedra Telefónica de Producció Multimèdia de la Universitat Pompeu Fabra.
## A Log-normal and double log-normal distributions
The following two probability distributions have been used in this article:
A log-normal (LN) distribution, which has the following probability density function (pdf):
<table>
<tbody>
<tr>
<td></td>
<td>fLN(t;μ,σ)subscript𝑓𝐿𝑁𝑡𝜇𝜎\displaystyle f_{LN}(t;\mu,\sigma)</td>
<td>=\displaystyle=</td>
<td>1tσ2πexp(−(ln(t)−μ)22σ2)1𝑡𝜎2𝜋superscript𝑡𝜇22superscript𝜎2\displaystyle\frac{1}{t\sigma\sqrt{2\pi}}\exp\left(\frac{-(\ln(t)-\mu)^{2}}{2\sigma^{2}}\right)</td>
<td></td>
<td>(2)</td>
</tr>
</tbody>
</table>
and its cumulative distribution function (cdf) is given by:
<table>
<tbody>
<tr>
<td></td>
<td>FLN(t;μ,σ)subscript𝐹𝐿𝑁𝑡𝜇𝜎\displaystyle F_{LN}(t;\mu,\sigma)</td>
<td>=\displaystyle=</td>
<td>12+12erf(ln(t)−μ2σ),1212erf𝑡𝜇2𝜎\displaystyle\frac{1}{2}+\frac{1}{2}\mbox{erf}\left(\frac{\ln(t)-\mu}{\sqrt{2}\sigma}\right),</td>
<td></td>
<td>(3)</td>
</tr>
</tbody>
</table>
where erf(x) is the Gauss error function being defined as
<table>
<tbody>
<tr>
<td></td>
<td>erf(x)=2π∫0xexp(−u2)𝑑u.erf𝑥2𝜋superscriptsubscript0𝑥superscript𝑢2differential-d𝑢\displaystyle\mbox{erf}(x)=\frac{2}{\sqrt{\pi}}\int_{0}^{x}\exp(-u^{2})du.</td>
<td></td>
<td>(4)</td>
</tr>
</tbody>
</table>
And a double log-normal (DLN) distribution, which is a superposition of two independent LN-distributions and has the following pdf:
<table>
<tbody>
<tr>
<td></td>
<td>fDLN(t;θ)subscript𝑓𝐷𝐿𝑁𝑡𝜃\displaystyle f_{DLN}(t;\theta)</td>
<td>=\displaystyle=</td>
<td>cfLN(t;μ1,σ1)+(1−c)fLN(t;μ2,σ2)𝑐subscript𝑓𝐿𝑁𝑡subscript𝜇1subscript𝜎11𝑐subscript𝑓𝐿𝑁𝑡subscript𝜇2subscript𝜎2\displaystyle cf_{LN}(t;\mu_{1},\sigma_{1})+(1-c)f_{LN}(t;\mu_{2},\sigma_{2})</td>
<td></td>
<td>(5)</td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td>where θwhere 𝜃\displaystyle\mbox{where }\theta</td>
<td>=\displaystyle=</td>
<td>(μ1,σ1,c,μ2,σ2).subscript𝜇1subscript𝜎1𝑐subscript𝜇2subscript𝜎2\displaystyle(\mu_{1},\sigma_{1},c,\mu_{2},\sigma_{2}).</td>
<td></td>
<td></td>
</tr>
</tbody>
</table>
The corresponding cdf can be easily derived from equations ([3](#S1.E3 "In A Log-normal and double log-normal distributions ‣ Homogeneous temporal activity patterns in a large online communication space")) and ([5](#S1.E5 "In A Log-normal and double log-normal distributions ‣ Homogeneous temporal activity patterns in a large online communication space")).
## B Error Measure $`\epsilon`$
We use the following distance measure to calculate the error of the approximations. The distance between approximation and data is only calculated for the time-bins (i.e. minutes) where a post actually receives a comment to avoid a distortion of the error measure by the periods with low comment activity.
######
Definition 1
Let $`\mathbb{T}`$ be the set of time-bins where a post receives at least one comment and $`T`$ its cardinality. We define then the approximation error $`\epsilon`$ of a function $`f\hspace{0pt}{(t)}`$ approximating $`g\hspace{0pt}{(t)}`$ (both defined for all $`t \in {\mathbb{T}}`$) as the normalized $`\ell^{1}`$-norm of $`{f\hspace{0pt}{(t)}} - {g\hspace{0pt}{(t)}}`$:
<table>
<tbody>
<tr>
<td></td>
<td>ϵ=∑t∈𝕋|f(t)−g(t)|T.italic-ϵsubscript𝑡𝕋𝑓𝑡𝑔𝑡𝑇\epsilon=\sum_{t\in\mathbb{T}}\frac{|f(t)-g(t)|}{T}.</td>
<td></td>
<td>(6)</td>
</tr>
</tbody>
</table>
If $`f\hspace{0pt}{(t)}`$ and $`g\hspace{0pt}{(t)}`$ are cumulative probability density functions (i.e. $`0 \leq {f\hspace{0pt}{(t)}} \leq 1`$ and $`0 \leq {g\hspace{0pt}{(t)}} \leq 1`$), it follows that $`0 \leq \epsilon \leq 1`$.
## References
- Adler, (1999) Adler, S. (1999). The Slashdot effect, an analysis of three Internet publications. http://ssadler.phy.bnl.gov/adler/SDE/SlashDotEffect.html.
- Albert et al., (1999) Albert, R., Jeong, H., and Barabási, A.-L. (1999). Internet: Diameter of the World-Wide Web.
Nature, 401:130, arXiv:cond-mat/9907038.
- Baoill, (2000) Baoill, A. Ó. (2000). Slashdot and the public sphere.
First Monday, 5(9).
- Barabási, (2005) Barabási, A.-L. (2005). The origin of bursts and heavy tails in human dynamics.
Nature, 435:207–211, arXiv:cond-mat/0505371.
- DeGroot and Schervish, (2002) DeGroot, M. H. and Schervish, M. J. (2002).
Probability and Statistics.
Addison-Wesley, New York, 3rd edition.
- Dewes et al., (2003) Dewes, C., Wichmann, A., and Feldmann, A. (2003). An analysis of Internet chat systems.
In IMC ’03: Proceedings of the 3rd ACM SIGCOMM conference on Internet measurement, pages 51–64, New York, NY, USA. ACM Press.
- Dezsö et al., (2006) Dezsö, Z., Almaas, E., Lukács, A., Rácz, B., Szakadát, I., and Barabási, A.-L. (2006). Dynamics of information access on the web.
Physical Review E, 73(6):066132, arXiv:physics/0505087.
- Faloutsos et al., (1999) Faloutsos, M., Faloutsos, P., and Faloutsos, C. (1999). On power-law relationships of the Internet topology.
In SIGCOMM ’99: Proceedings of the conference on Applications, technologies, architectures, and protocols for computer communication, pages 251–262, New York, NY, USA. ACM Press.
- Goldstein et al., (2004) Goldstein, M. L., Morris, S. A., and Yen, G. G. (2004). Problems with fitting to the power-law distribution.
The European Physical Journal B, 41(2):255–258, arXiv:cond-mat/0402322.
- Habermas, (1989) Habermas, J. (1962/1989).
The Structural Transformation of the Public Sphere: Inquiry into a Category of Bourgeois Society.
Cambridge, MA: MIT Press.
- Harder and Paczuski, (2006) Harder, U. and Paczuski, M. (2006). Correlated dynamics in human printing behavior.
Physica A: Statistical Mechanics and its Applications, 361(1):329–336.
- Henderson and Bhatti, (2001) Henderson, T. and Bhatti, S. (2001). Modelling user behaviour in networked games.
In MULTIMEDIA ’01: Proceedings of the 9th ACM International Conference on Multimedia, pages 212–220, New York, NY, USA. ACM Press.
- Johansen, (2004) Johansen, A. (2004). Probing human response times.
Physica A: Statistical Mechanics and its Applications, 338(1–2):286–291, arXiv:cond-mat/0305079.
- \(14\) Kaltenbrunner, A., Gómez, V., and López, V. (2007a). Description and prediction of Slashdot activity.
In Proceedings of the 5th Latin American Web Congress (LA-WEB 2007), Santiago de Chile. IEEE Computer Society.
- \(15\) Kaltenbrunner, A., Gómez, V., Moghnieh, A., Meza, R., Blat, J., and López, V. (2007b). Homogeneous temporal activity patterns in a large online communication space.
In Proceedings of the BIS 2007 Workshop on Social Aspects of the Web (SAW 2007). Poznan, Poland.
- Kleban and Clearwater, (2003) Kleban, S. D. and Clearwater, S. H. (2003). Hierarchical dynamics, interarrival times, and performance.
In SC ’03: Proceedings of the 2003 ACM/IEEE conference on Supercomputing, page 28, Washington, DC, USA. IEEE Computer Society.
- Lampe and Resnick, (2004) Lampe, C. and Resnick, P. (2004). Slash(dot) and burn: Distributed moderation in a large online conversation space.
In CHI ’04: Proceedings of the SIGCHI conference on Human factors in computing systems, pages 543–550, New York, NY, USA. ACM Press.
- Leskovec et al., (2006) Leskovec, J., Adamic, L. A., and Huberman, B. A. (2006). The dynamics of viral marketing.
In EC ’06: Proceedings of the 7th ACM conference on Electronic commerce, pages 228–237, New York, NY, USA. ACM Press, arXiv:physics/0509039.
- Limpert et al., (2001) Limpert, E., Stahel, W. A., and Abbt, M. (2001). Log-normal distributions across the sciences: Keys and clues.
Bioscience, 51(5):341–352(12).
- Mainardi et al., (2000) Mainardi, F., Raberto, M., Gorenflo, R., and Scalas, E. (2000). Fractional calculus and continuous-time finance II: the waiting-time distribution.
Physica A: Statistical Mechanics and its Applications, 287(3):468–481, arXiv:cond-mat/0006454.
- Malda, (2002) Malda, R. (2002). Slashdot FAQ: Comments and Moderation. http://slashdot.org/faq/com-mod.shtml.
- Masoliver et al., (2003) Masoliver, J., Montero, M., and Weiss, G. H. (2003). Continuous-time random-walk model for financial distributions.
Physical Review E, 67(2):021112, arXiv:cond-mat/0210513.
- Mitzenmacher, (2003) Mitzenmacher, M. (2003). A brief history of generative models for power law and lognormal distributions.
Internet Mathematics, 1(2):226–251.
- Newman, (2005) Newman, M. E. J. (2005). Power laws, pareto distributions and zipf’s law.
Contemporary Physics, 46:323–351, arXiv:cond-mat/0412004.
- Paxson and Floyd, (1995) Paxson, V. and Floyd, S. (1995). Wide area traffic: The failure of Poisson modeling.
IEEE-ACM Transactions On Networking, 3(3):226–244.
- Poor, (2005) Poor, N. (2005). Mechanisms of an online public sphere: The web site Slashdot.
Journal of Computer-Mediated Communication, 10(2).
- Sigman, (1999) Sigman, K. (1999). Appendix: A primer on heavy-tailed distributions.
Queueing Systems: Theory and Applications, 33(1–3):261–275.
- Stouffer et al., (2006) Stouffer, D. B., Malmgren, R. D., and Amaral, L. A. N. (2006). Log-normal statistics in e-mail communication patterns, arXiv:physics/0605027.
- Vázquez et al., (2006) Vázquez, A., Oliveira, J. G., Dezso, Z., Goh, K. I., Kondor, I., and Barabási, A.-L. (2006). Modeling bursts and heavy tails in human dynamics.
Physical Review E, 73:036127, arXiv:physics/0510117. |
0707.2293 | cs.NI | cyber | # Worm Epidemics in Wireless Adhoc Networks
Maziar Nekovee
BT Research, Polaris 134, Adastral Park, Martlesham, Suffolk IP5 3RE, UK
and
Centre for Computational Science, University College London, 20 Gordon Street, London WC1H 0AJ, UK
###### Abstract
A dramatic increase in the number of computing devices with wireless communication capability has resulted in the emergence of a new class of computer worms which specifically target such devices. The most striking feature of these worms is that they do not require Internet connectivity for their propagation but can spread directly from device to device using a short-range radio communication technology, such as WiFi or Bluetooth. In this paper we develop a new model for epidemic spreading of these worms and investigate their spreading in wireless adhoc networks via extensive Monte Carlo simulations. Our studies show that the threshold behaviour and dynamics of worm epidemics in these networks are greatly affected by a combination of spatial and temporal correlations which characterise these networks, and are significantly different from the previously studied epidemics in the Internet.
###### pacs:
89.75.Hc,05.70.Jk,87.19.Xx, 89.75.Fb
## I INTRODUCTION
Worms are self-replicating computer viruses which can propagate through computer networks without any human intervention \[[1](#bib.bib1), [2](#bib.bib2), [3](#bib.bib3)\]. Cyber attacks by this type of viruses present one of the most dangerous threats to the security and integrity of computer and telecommunications networks. The Code Red \[[5](#bib.bib5), [4](#bib.bib4)\] and Nimda \[[4](#bib.bib4)\] worms, for example, infected hundreds of thousands of computers at alarming speeds and the resulting worm epidemics cost both the public and the private sector a great deal of money. The last few years have seen the emergence of a new type of worms which specifically targets portable computing devices, such as smartphones and laptops. The novel feature of these worms is that they do not necessarily require Internet connectivity for their propagation. They can spread directly from device to device using a short-range wireless communication technology, such as WiFi or Bluetooth \[[4](#bib.bib4), [6](#bib.bib6), [7](#bib.bib7), [8](#bib.bib8)\], creating in their wake an adhoc contact network along which they propagate. The first computer worm written specially for wireless devices was detected in 2003 and within three years the number of such viruses soared from one to more than $`300`$ \[[8](#bib.bib8)\]. With wireless networks becoming increasingly popular, many security experts predict that these networks will soon be a main target of attacks by worms and other type of malware. \[[8](#bib.bib8)\].
Worm and virus attacks on the Internet have been the subject of extensive empirical, theoretical and simulation studies \[[1](#bib.bib1), [9](#bib.bib9), [10](#bib.bib10), [11](#bib.bib11), [12](#bib.bib12)\]. These studies have greatly contributed to our understanding of the impact of network topology on the properties of virus spreading \[[9](#bib.bib9), [10](#bib.bib10)\] and have inspired the design of more effective immunisation strategies to prevent and combat Internet epidemics \[[11](#bib.bib11), [12](#bib.bib12)\]. Investigation of virus spreading in wireless networks in general and worms in particular is, however, at its infancy, and there has been very limited studies which address this problem \[[7](#bib.bib7), [13](#bib.bib13)\].
In this paper we develop a new model for the spreading of worms in Wi-Fi-based wireless adhoc networks and investigate the properties of worm epidemics in these networks via extensive Monte Carlo simulations. Wireless adhoc networks \[[14](#bib.bib14), [15](#bib.bib15), [16](#bib.bib16), [16](#bib.bib16), [17](#bib.bib17), [18](#bib.bib18)\] are distributed networks which can be formed on the fly by WiFi-equipped devices, such as laptops and smartphones. Nodes in these networks communicate directly with each other and can route data packets wirelessly, either among themselves or to the nearest Internet accesspoint. Adhoc technology has important applications in the provisioning of ubiquitous wireless Internet access, disaster relief operations and wireless sensor networks. From the perspective of complex network theory \[[20](#bib.bib20), [22](#bib.bib22), [21](#bib.bib21), [23](#bib.bib23)\] the study of these networks is important as their topology provides a clear-cut example of spatial networks \[[24](#bib.bib24)\]. Spatial networks are embedded in a metric space where interactions between the nodes is a function of their spatial distance \[[24](#bib.bib24), [25](#bib.bib25)\]. Despite their relevance to many real-life phenomena the properties of these networks are much less studied than abstract graphs.
Our Monte Carlo simulations show that epidemic spreading in wireless adhoc networks is significantly different from the previously studied epidemics in the Internet. The initial growth of the epidemic is significantly slower than the exponential growth observed for worm spreading in the Internet, and the epidemic prevalence exhibits a density-dependent critical threshold which is higher than the value predicted by the mean-field theory. We show that these differences are due to strong spatial and temporal correlations which characterise these networks. Our study also reveals the presence of a self-throttling effect in the spreading of worms in wireless networks which greatly slows down the speed of worm invasion in these networks.
The rest of this paper is organised as follows. In section II we describe our models of network topology, data communication mechanism, and worm spreading in wireless adhoc networks. In section III we present and discuss results of our Monte Carlo simulations studies of epidemics in these networks for a range of device densities and infection rates. We close this paper in section IV with conclusions.
## II Models
### II.1 Network model
We consider a collection of nodes distributed in a two dimensional plane which communicate using short-range radio transmissions. The received radio signal strength at a device $`j`$ resulting from a transmission by a device $`i`$ decays with the distance between the sender and the receiver due to a combination of free-space attenuation and fading effects. Phenomenologically this effect is described using the so-called pathloss model \[[26](#bib.bib26)\] which states that the mean value of the signal power at a receiving device $`j`$ is related to the signal power of the transmitting node $`i`$ via the following equation:
<table>
<tbody>
<tr>
<td></td>
<td>Pij=Picrijα.superscript𝑃𝑖𝑗superscript𝑃𝑖𝑐superscriptsubscript𝑟𝑖𝑗𝛼P^{ij}=\frac{P^{i}}{cr_{ij}^{\alpha}}.</td>
<td></td>
<td>(1)</td>
</tr>
</tbody>
</table>
In the above equation $`r_{i\hspace{0pt}j}`$ is the Euclidean distance between node $`i`$ and node $`j`$, $`P^{i}`$ and $`P^{i\hspace{0pt}j}`$ are the transmit power and the received power, respectively, and $`c`$ is a constant whose precise value depends on a number of factors including the transmission frequency. For free space propagation $`\alpha = 2`$, but depending on the specific indoor/outdoor propagation scenario it is found empirically that $`\alpha`$ can vary typically between $`2`$ and $`5`$. A data transmission by node $`i`$ is correctly received at node $`j`$, i.e. $`i`$ can establish a communication link with $`j`$, provided that:
<table>
<tbody>
<tr>
<td></td>
<td>Pijν=Pi/crijαν≥βth.superscript𝑃𝑖𝑗𝜈superscript𝑃𝑖𝑐superscriptsubscript𝑟𝑖𝑗𝛼𝜈subscript𝛽𝑡ℎ\frac{P^{ij}}{\nu}=\frac{P^{i}/cr_{ij}^{\alpha}}{\nu}\geq\beta_{th}.</td>
<td></td>
<td>(2)</td>
</tr>
</tbody>
</table>
In the above equation $`\beta_{t\hspace{0pt}h}`$ is an attenuation threshold and $`\nu`$ is the noise level at node $`j`$.
Condition (2) translates into a maximum transmission range for node $`i`$:
<table>
<tbody>
<tr>
<td></td>
<td>rti=(Picβthν)1/α,superscriptsubscript𝑟𝑡𝑖superscriptsuperscript𝑃𝑖𝑐subscript𝛽𝑡ℎ𝜈1𝛼r_{t}^{i}=\left(\frac{P^{i}}{c\beta_{th}\nu}\right)^{1/\alpha},</td>
<td></td>
<td>(3)</td>
</tr>
</tbody>
</table>
such that each device can establish wireless links with only those devices within a circle of radius $`r_{t}^{i}`$. A communication graph is then constructed by creating an edge between node $`i`$ and all other nodes in the plane that are within the transmission range of $`i`$, and repeating this procedure for all nodes in the network. In general wireless devices may use different transmit powers such that the existence of a wireless link from $`i`$ to $`j`$ does not imply that a link from $`j`$ to $`i`$ also exists. Consequently the resulting communication graph is directed. Assuming, however, that all devices use the same transmit power $`P`$, and a corresponding transmission range $`r_{t}`$, the topology of the resulting network can be described as a two dimensional random geometric graph (RGG) \[[27](#bib.bib27), [28](#bib.bib28)\]. Random geometric graphs have been used extensively in the study of continuum percolation and more recently for modelling wireless adhoc networks \[[15](#bib.bib15), [16](#bib.bib16), [17](#bib.bib17), [18](#bib.bib18), [19](#bib.bib19)\]. Like Erdős-Rényi random graphs (RG) \[[30](#bib.bib30)\], these graphs have a Poisson degree distribution, $`P\hspace{0pt}{(k)}`$, which peaks at an average value $`\langle k\rangle`$ and shows small fluctuations around $`\langle k\rangle`$. However, other properties of a RGG are radically different from a Erdős-Rényi random graph. Most notably, these networks are characterised by a large cluster coefficient, $`C = 0.59`$, which is a purely geometric quantity independent of both node density and $`\langle k\rangle`$ \[[28](#bib.bib28), [15](#bib.bib15)\]. Furthermore, it has been shown numerically that the critical connectivity in these networks is at $`{\langle k\rangle} = 4.52`$ \[[28](#bib.bib28)\], which is much higher than the well-known $`{\langle k\rangle} = 1`$ value in RG.
### II.2 Medium access control
In WiFi networks access to the available frequency channels is controlled by a coordination mechanism called the Medium Access Control (MAC) \[[29](#bib.bib29)\]. The function of the MAC is to ensure interference-free wireless transmissions of data packets in the network. This is achieved by scheduling in time the transmissions of nearby devices in such a way that devices whose radio transmissions may interfere with each other do not get access to the wireless channel at the same time. The presence of the MAC introduces novel spatio-temporal correlations in the dynamics of data communications in these networks which are absent in the Internet communications.
The MAC protocol used by WiFi-based wireless devices follows the IEEE 802.11 standard \[[29](#bib.bib29)\], which specifies a set of rules that enable nearby devices coordinate their transmissions in a distributed manner. The IEEE 802.11 MAC is a highly complex protocol and we do not attempt to fully model this protocol. Instead we focus on the most relevant aspect of this protocol, the so-called listen-before-talk (LBT) rule. This rule dictates that each device should check the occupancy of the wireless medium before starting a data transmission and refrain from transmitting if it senses that the medium is busy. The precise implementation of the LBT algorithm will be discussed in section II.D.
### II.3 Worm propagation model
Worms are stand-alone computer viruses which use networks for their spreading among computing devices. Consequently, computer worms can propagate automatically from device to device, in contrast to other types of virus which require some form of user involvement for their spreading.
Several previous studies have analysed and modelled the propagation of computer worms on the Internet. Most contemporary Internet worms work as follows \[[4](#bib.bib4)\]. When a computer worm is fired into the Internet, it scans the IP (Internet Protocol) addresses and sends a probe to infect the corresponding machines. When a vulnerable machine becomes infected by such a probe, it begins running the worm and tries to infect other machines. A patch, which repairs the security holes of the machine, is used to defend against worms. When an infected or vulnerable machine is patched against a worm, it becomes immune to that worm. There are several different scanning mechanisms that worms deploy. Two main mechanisms are random scanning and local subnet scanning. In random scanning an infected computer scans the entire IP address space and selects its targets randomly from this space. In local scanning the worm scans the nearby targets (e.g. machines on the same subnet) with a higher probability. Many recent worms, such as Code Red v2 have used localised scanning.
The above mechanisms require that both the infected and the vulnerable nodes are connected to the Internet and rely on the IP routing mechanism for worm delivery. However, it is well-known that point-to-point routing of data packets in wireless adhoc networks could be problematic due to the highly dynamic nature of these networks. A much more robust mechanism for disseminating packets in such networks is by multihop forwarding in which a packet propagates in the network by broadcast radio transmissions from device to device, without the need for any routing mechanism or Internet connectivity. This mechanism shows interesting analogies with the way airborne diseases spread in populations and has been exploited in a recent worm attack on Bluetooth-enabled smartphones \[[8](#bib.bib8)\]. We assume therefore that worms targeting these networks will utilise multihop broadcasts as their primary method of propagation. With respect to an attacking worm, we assume that nodes in the network to be in one of the following three states: vulnerable, infected, or immune. Infected nodes try to transmit the worm to their neighbours at every possible opportunity. Vulnerable nodes can become infected at a rate $`\lambda`$ when they receive a transmission containing a copy of the worm from an infected neighbour. Finally, infected nodes get patched and become immune to the worm at a rate $`\delta`$. We denote by $`S\hspace{0pt}{(t)}`$, $`I\hspace{0pt}{(t)}`$ and $`R\hspace{0pt}{(t)}`$ the population of vulnerable, infected and immune nodes, respectively.
### II.4 Implementation
In our simulations we have implemented the above model of worm spreading in wireless adhoc networks in the following way. At each timestep of simulations we create a randomly ordered list of the infected nodes in the network at that timestep. The first node on the list then gets access to the wireless channel and is allowed to transmit the worm. All other infected nodes that are within the transmission range of this node are eliminated from the list as their transmission may cause interference to that node, and is therefore blocked by the LBT rule. This procedure is repeated for the remaining nodes until the list is reduced to a set of non-interfering infected nodes which can transmit the worm at that same timestep.
Subsequently, all infected nodes which are on the above list go through a broadcast round in which they transmit the worm to their neighbours. Finally, all infected nodes (i.e. both those who were able to transmit the worm and those whose transmissions were blocked by the MAC protocol) go through a patching round in which they may become immune with probability $`\delta`$.
## III Simulation studies
We simulated the propagation of worms in wireless adhoc networks comprising $`N`$ devices spread in a $`L^{2} = {1000 \times 1000}`$ $`m^{2}`$ area. The transmission range of all devices was set at $`50`$ m, which is somewhere between the typical minimum ($`30`$ m) and maximum ($`100`$ m) range of the WiFi systems. In order to investigate the impact of device density we performed our simulations for a range of densities, corresponding to $`N = {4000,6000,8000,10000,20000}`$.
For a given density, nodes were distributed randomly and uniformly in the simulation cell. The resulting RGG networks were constructed following the prescription of Sec. II. A, and periodic boundary conditions were used in order to reduce finite-size effects. We verified numerically that all the networks considered were connected, and their degree distributions were well-described by the Poisson distribution:
<table>
<tbody>
<tr>
<td></td>
<td>P(k)=e−⟨k⟩⟨k⟩kk!,𝑃𝑘superscript𝑒delimited-⟨⟩𝑘superscriptdelimited-⟨⟩𝑘𝑘𝑘P(k)=e^{-\langle k\rangle}\frac{\langle k\rangle^{k}}{k!},</td>
<td></td>
<td>(4)</td>
</tr>
</tbody>
</table>
with the average degree,$`\langle k\rangle`$, given by:
<table>
<tbody>
<tr>
<td></td>
<td>⟨k⟩=πrt2ρ,delimited-⟨⟩𝑘𝜋superscriptsubscript𝑟𝑡2𝜌\langle k\rangle=\pi r_{t}^{2}\rho,</td>
<td></td>
<td>(5)</td>
</tr>
</tbody>
</table>
where $`\rho = {N/L^{2}}`$ is the device density.
The spreading dynamics was simulated on top of the above networks using Monte Carlo simulations. Each Monte Carlo run starts by infecting a single randomly chosen node and proceeds following the rules described in sections II.C and II.D until the epidemic dies out (i.e. no infected node is left in the network). We typically average our results over $`500`$ Monte Carlo runs. Furthermore, the results were also averaged over simulations starting from at least $`5`$ different initial infected seeds. Since the time scale of the epidemic spreading depends only on the ratio $`\lambda/\delta`$, rather than $`\lambda`$ and $`\delta`$ separately, without the loss of generality we set the patching rate at $`\delta = 1`$ and performed our simulations for a range of values of the infection rate, $`\lambda`$.
In order to investigate the impact of the MAC on worm epidemics all simulations were performed both in the presence and in the absence of this mechanism. The latter case corresponds to an idealised scenario where nearby devices can communicate with each other without causing harmful interference, for example by using non-overlapping frequency channels \[[34](#bib.bib34)\], and maps the dynamics of worm spreading onto the standard susceptible-infected-removed (SIR) epidemic model. Finally, as a point of reference, we also performed simulation studies of the SIR model on a set of Erdős-Rényi random graphs which were constructed such that their degree distribution virtually coincided with that of our RGG networks. In the following we shall refer to our full simulations as RGG+MAC while simulations in the absence of MAC will be labelled as RGG and those performed on random graphs as RG. For future reference we note that the SIR epidemic on random graphs roughly mimics the spread of Internet worms via random scanning, and is well-described by the mean-field theory.
[Refer to caption](/html/0707.2293/assets/x1.png)
Figure 1: The epidemic prevalence, $`R_{\infty}`$, is shown as a function of the infection rate $`\lambda`$ for a wireless adhoc network consisting of $`N = 10000`$ nodes, both in the absence and presence of the MAC mechanism. Also shown is the result for a random graph network with the same number of nodes and the same degree distribution as the wireless adhoc network.
[Refer to caption](/html/0707.2293/assets/x2.png)
Figure 2: The epidemic prevalence, $`R_{\infty}`$, as a function of the infection rate $`\lambda`$ is shown for wireless adhoc networks comprising $`N = {6000,8000,10000,20000}`$ nodes, respectively. Results are shown for simulations performed in the absence of the MAC mechanism.
[Refer to caption](/html/0707.2293/assets/x3.png)
Figure 3: Collapse plots of $`R_{\infty}`$ vs. $`\kappa = {\lambda\hspace{0pt}{\langle k\rangle}}`$ is shown for wireless adhoc networks comprising $`N = {6000,8000,10000,20000}`$ nodes, both in the presence and absence of the MAC mechanism.
### III.1 Prevalence and epidemic threshold
A key quantity in the study of epidemics in networks is the epidemic prevalence. For the SIR-type epidemics this quantity is defined as $`R_{\infty} = {\lim_{t\rightarrow\infty}{{R\hspace{0pt}{(t)}}/N}}`$ \[[31](#bib.bib31)\]. In Fig. 1 prevalence as a function of $`\lambda`$ is shown as obtained from our simulations of, respectively, RG, RGG and RGG+MAC networks comprising $`N = 10000`$ nodes. It can be seen that in all these networks $`R_{\infty}`$ exhibits a critical threshold $`\lambda_{c}`$ below which a worm cannot spread in the network and above which it infects a finite fraction of the nodes. However, the epidemic threshold corresponding to RGG is at a considerably higher value than that of RG, despite the fact that the degree distributions of these two networks are identical. Furthermore, it can be seen that the inclusion of the MAC mechanism results in an increase in the value of the epidemic threshold in RGG, shifting the position of $`\lambda_{c}`$ even further away from the RG value. Our computed epidemic thresholds for the above networks are $`\lambda = {0.0140,0.0210,0.0265}`$ for RG, RGG and RGG+MAC, respectively. The mean-field theory, which in the infinite system size limit becomes exact for the RG network, predicts an epidemic threshold at $`\lambda_{c} = {1/{\langle k\rangle}}`$ \[[31](#bib.bib31), [32](#bib.bib32)\] , yielding $`\lambda_{c} = 0.0127`$ for the above networks. This is in good agreement with our Monte Carlo result for the RG network, indicating that the above differences between $`\lambda_{c}`$ in RG, RGG and RGG+MAC are not due to finite-size effects (or statistical fluctuations) but are caused by a combination of topological and dynamic correlations in our wireless networks, which are absent in RG.
Qualitatively, we can understand the above results by noting that due to spatial ordering in random geometric graphs, the epidemic state of a node in a RGG is strongly correlated to the state of its neighbours. These correlations reduce the so-called reproductive rate of the epidemic, the average number of new infections that can be produced by an infective node, below the value predicted by the mean-field approximation, hence increasing $`\lambda_{c}`$ above the mean-field value. The presence of the MAC introduces additional temporal correlations between the transmission times of adjacent infective nodes which further reduce the reproductive rate, hence further increasing $`\lambda_{c}`$ above the mean-field value. It can be seen from Fig.1 that this latter mechanism not only affects $`\lambda_{c}`$ but also results in a significant reduction in the epidemic prevalence.
Next we investigate the impact of node density, $`\rho`$, on the behaviour of epidemic prevalence in our networks. In Fig. 2 we plot $`R_{\infty}`$ as function of $`\lambda`$ for different device densities. Results are shown only for RGG but they show a similar behaviour for RGG+MAC. It can be seen that for all densities considered the prevalence shows a critical behaviour. However, the position of $`\lambda_{c}`$ decreases monotonically with increasing density, i.e. the worm epidemic is more successful in invading the network when the density of devices is high and less so when density is low. In order to better understand the density dependent behaviour of $`R_{\infty}`$ we plot in Fig. 3 this quantity as function of $`\kappa = {\lambda\hspace{0pt}{\langle k\rangle}} = {\lambda\hspace{0pt}\pi\hspace{0pt}\rho\hspace{0pt}r_{t}^{2}}`$. It can be seen that for RGG there is a good collapse of curves in an extended region around the threshold. This indicates that in this region the prevalence of the SIR model on RGG is well-described by the scaling relation $`{R_{\infty}\hspace{0pt}{(\lambda,\rho)}} = {f\hspace{0pt}{(\kappa)}}`$ (For RGG+MAC this scaling holds only approximately). In particular, we find that the epidemic threshold itself can be written as:
<table>
<tbody>
<tr>
<td></td>
<td>λc=κc⟨k⟩=κcπρrt2,subscript𝜆𝑐subscript𝜅𝑐delimited-⟨⟩𝑘subscript𝜅𝑐𝜋𝜌superscriptsubscript𝑟𝑡2\lambda_{c}=\frac{\kappa_{c}}{\langle k\rangle}=\frac{\kappa_{c}}{\pi\rho r_{t}^{2}},</td>
<td></td>
<td>(6)</td>
</tr>
</tbody>
</table>
with $`\kappa_{c} = 1.50`$, a correction to the mean-field model resulting from spatial correlations in RGG. Since the cluster coefficient, $`C`$, is a measure of correlations in a network, we note that the value of $`\kappa`$ is in fact very close to $`1/C`$ indicating that the departure from the mean-field model is possibly controlled by this quantity.
Next, we investigate the dependence of the epidemic prevalence on device density. In Fig. 4 this quantity is plotted as a function of $`N`$ and for $`\lambda = 0.1`$, both in the presence of MAC and when this mechanism is switched off. It can be seen that in both cases $`R_{\infty}`$ increases monotonically with increasing node density. Furthermore, for all values of $`N`$ the curve corresponding to RGG+MAC lies below that of RGG. However, the gap between the two curves decreases as $`N`$ is increased, indicating that the impact of MAC on $`R_{\infty}`$ becomes less significant at high densities.
### III.2 Spreading dynamics
Finally, we discuss the propagation dynamics of worms in our networks. Fig. 5 displays, as an example, time evolution of the total fraction of infected nodes, $`{I\hspace{0pt}{(t)}}/N`$, in the $`N = 10000`$ node network and for $`\lambda = 0.1`$. For comparison we have also plotted the result obtained for the corresponding random graph network. As can be seen from Fig. 5 worm spreading on the RGG network takes place at a much slower pace than on the RG network. In particular, the initial growth of the epidemic on RGG is much slower than the exponential growth seen for RG, which is a hallmark of mean-field models \[[33](#bib.bib33)\] and is also observed in epidemics in the Internet \[[2](#bib.bib2)\].
The above slow growth of the epidemic on the RGG is a purely topological effect, which we attribute to a combination of spatial correlations and high clustering in this network. As can be seen from Fig. 5 switching on the MAC protocol in RGG slows down the epidemic on this network even further. This effect, which we call self-throttling, is caused by temporal correlations in the spreading dynamics introduced by the MAC and has also been observed in a recent study of IP-based worm spreading in mobile adhoc networks \[[7](#bib.bib7)\]. It results because adjacent infective devices compete with each other in accessing the shared wireless medium, hence effectively blocking each other’s broadcasts and slowing down the overall progress of the epidemic. In Fig. 6 we display our result for $`{R\hspace{0pt}{(t)}}/N`$ in the $`N = 10000`$ networks, which further demonstrates the great impact of spatial and temporal correlations on the spreading process in our wireless networks.
Next we discuss the impact of device density on the speed of worm propagation. As an indicator of the speed we use the position of the peak in $`{I\hspace{0pt}{(t)}}/N`$, which we call $`T_{m\hspace{0pt}a\hspace{0pt}x}`$. This quantity is plotted in Fig. 7 as a function of $`N`$, both in the absence and presence of MAC. It can be seen that both curves show a monotonic decrease with increasing node density. Furthermore, we see that the self-throttling mechanism is most effective in slowing down the epidemic at the lowest density, where we observe a $`\sim {40\%}`$ increase in $`T_{m\hspace{0pt}a\hspace{0pt}x}`$ when MAC is switched on. As density increases the gap between the two curves becomes initially smaller before settling down at higher densities.
[Refer to caption](/html/0707.2293/assets/x4.png)
Figure 4: The epidemic prevalence in a wireless adhoc network is plotted as a function of the number of devices in the network. Results are shown of simulations performed in the absence (circles) and presence of the MAC (squares).
[Refer to caption](/html/0707.2293/assets/x5.png)
Figure 5: Time evolution of the fraction of infective devices, $`{I\hspace{0pt}{(t)}}/N`$, is shown for networks consisting of $`N = 10000`$ nodes. Results of simulations are shown both in the absence (squares) and presence (triangles) of MAC. Also shown are the results for the corresponding random graph network (circles).
[Refer to caption](/html/0707.2293/assets/x6.png)
Figure 6: Time evolution of the fraction of immunised devices, $`{R\hspace{0pt}{(t)}}/N`$, is shown for the same networks as in Fig. 5.
[Refer to caption](/html/0707.2293/assets/x7.png)
Figure 7: The position of the epidemic peak, $`T_{m\hspace{0pt}a\hspace{0pt}x}`$, in RGG networks is plotted as function of $`N`$. Results are shown when the MAC is switched off (circles) and when it is switched on (squares).
## IV CONCLUSIONS
In this paper we introduced a model for the propagation of a new class of computer worms which specifically target wireless computing devices. Using extensive Monte Carlo simulations we investigated the epidemic spreading of such worms in WiFi-based wireless adhoc networks. We incorporated the spatial topology of these networks via a random geometric graph model, and also took into account the impact of the Medium Access Control on wireless data communications in these networks.
Our studies show that worm epidemics in wireless adhoc networks are greatly different from the previously studied epidemics in the Internet. The epidemic threshold was found to be density-dependent and for all densities considered significantly higher than the value predicted by the mean-field theory. Furthermore, the initial growth of the epidemic was found to be significantly slower than the exponential growth observed in Internet epidemics and predicted by the mean-field theory. We showed that these differences were due to a combination of spatial and temporal correlations which are inherent to wireless data networks. Our study also revealed the presence of a self-throttling mechanism which results from a competition between adjacent infected devices for access to the shared wireless medium. This mechanism greatly reduces the speed of worm propagation and the risk of large-scale worm epidemics in these networks.
An understanding of the propagation characteristics of worm attacks on wireless networks is of great importance for the design of effective detection and prevention strategies for these networks. The work presented in this paper is a first step in this direction and, we hope, will inspire future empirical and theoretical investigations. From the perspective of complex network theory, our work presents an extensive study of epidemic spreading in random geometric graphs, and highlights the important role that spatial correlations play in dynamic processes on these and other spatial networks.
###### Acknowledgements.
M.N. acknowledges support from the Royal Society through an Industry Fellowship and thanks the Centre for Computational Science at UCL for hospitality.
## References
- \[1\] S. Stantiford, V. Paxton and N. Weaver, Proc. of the 11th USENIX Security Symposium (Security ’02), 2000.
- \[2\] T. M. Chen and J-M Robert, IEEE Computer, pp 48-53, June 2004.
- \[3\] R. Pastor-Satorras and A. Vespignani,
Evolution and Structure of the Internet: A Statistical Physics Approach.
Cambridge University Press, Cambridge, UK, 2004.
- \[4\]
P. Szor, The Art of Computer Virus Research and Defense, Symantec Press, 2006.
- \[5\] D. Moore and C. Shannon, Code-Red: A case study on the spread of victims of Internet worms, in Proceedings of the 2002 SIGCOMM Internet Measurement Workshop, Marseilles, France, pp. 273-248, 2003.
- \[6\]
N. Levitt, IEEE Computer 38, No. 4, 20-23, 2005; D. Dagon, T. Martin and T. Starner, IEEE Pervasive Computing 3, 11-15, 2004.
- \[7\] R. G. Cole, Army Science Conference 2004, Orlando, FL, 2004.
- \[8\] For a recent review see M. Hypponen, Scientific American, 70–77, November 2006.
- \[9\] R. Pastor-Satorras and A. Vespignani,
Phys. Rev. Lett., 86, 3200 (2001).
- \[10\] M. E. J. Newman, S. Forrest, and J. Balthrop,
Phys. Rev. E, 66, 035101(R) (2002).
- \[11\]
J. Goldenberg, Y. Shavitt, E. Shir, S. Solomon, Nature Phys. 1, 184-188 (2005).
- \[12\]
J. Balthrop, S. Forrest, M. E. Newman, M. M. Williamson, Science 304, No. 5670, 527-529 (2004).
- \[13\] M. Nekovee, Proceedings of IEEE VTC Spring 2006, Melbourne, Australia, 2006.
- \[14\]
R. Hekmat, Adhoc networks: Fundamental properties and network topologies, Springer, 2006.
- \[15\]
I. Glauche, W. Krause, R. Sollacher, and M. Geiner, Physica A 325, 577-600 (2002).
- \[16\]
W. Kraus, I. Glauche, R. Sollacher, and M. Geiner, Physica A 338, 633-658 (2004).
- \[17\]
I. Glauche W. Krause, R. Sollacher,and M. Geiner, Physica A 341, 677-701 (2004).
- \[18\]
W. Krause, J. Scholz, and M. Geiner, Physica A 361, 707-723 (2006).
- \[19\]
R. Sosperda Alfonso, On Random Geometric Graphs: Structure and Epidemics, MPhil Thesis, ICTP, 2005.
- \[20\] R. Albert and A-L Barabási,
Rev. Mod. Phys., 74, 47 (2002).
- \[21\] S. N. Dorogovtsev and J. F F. Mendes,
Adv. Phys., 51 1079-1187 (2002).
- \[22\] M. E. J. Newman.
SIAM Rev., 45, 167 (2003).
- \[23\] S. Boccaletti, V. Latora, Y. Moreno, M. Chavez, and D.-U. Hwang, Physics Reports 424, 2006.
- \[24\]
C. Heramnn, M.Barthélemy, and P. Provero, Phys. Rev. E 68, 026128 (2003).
- \[25\]
D. ben-Avraham, A. F. Rozenfeld, R. Cohen, and S. Havlin, Physica A 330, 107-116 (2003).
- \[26\]
T. Rappaport, Wireless Communications, Principle and Parctice, Prentice-Hall, 2000.
- \[27\]
M. Penrose, Random Geometric Graphs, Oxford University Press, Oxford, 2003.
- \[28\]
J. Dall and M. Christensen, Phys. Rev. E 66, 016121 (2002).
- \[29\]
M. S. Gast, 802.11 Wireless Networks, Second Edition, O’Reily, 2005; W. Stalling, Wireless Communications Networks, Prentice Hall, 2005.
- \[30\]
B. Bollobas, Modern Graph Theory, Springer, New York, 1998.
- \[31\] Y. Moreno, R. Pastor-Satorras, and A. Vespignani,
Eur. Phys. J. B, 63, 521 (2002).
- \[32\]
M. Boguna, R. Pastor-Satorras, and A. Vespignani, Lecture Notes in Physics 625, 127-147 (2003).
- \[33\]
M. Bathélemy, A. Barrat, R. Pastor-Satoras, and A. Vespignani, Phys. Rev. Lett. 92, 178701 (2004).
- \[34\] This might be the case, for example, in worm attacks on Bluetooth networks. |
0706.3146 | cs.CR | cyber | # WiFi Epidemiology: Can Your Neighbors’ Router Make Yours Sick?
Hao Hu<sup>1,2</sup>, Steven Myers<sup>2,∗</sup>, Vittoria Colizza<sup>2,3</sup>, Alessandro Vespignani<sup>2,3</sup>
<sup>1</sup>Department of Physics, Indiana University, Bloomington, IN 47405, USA
<sup>2</sup>School of Informatics, Indiana University, Bloomington, IN 47406, USA
<sup>3</sup>Complex Networks Lagrange Laboratory, Institute for Scientific Interchange (ISI), Torino 10133, Italy
<sup>∗</sup>e-mail: samyers@indiana.edu
In densely populated urban areas WiFi routers form a tightly interconnected proximity network that can be exploited as a substrate for the spreading of malware able to launch massive fraudulent attack and affect entire urban areas WiFi networks. In this paper we consider several scenarios for the deployment of malware that spreads solely over the wireless channel of major urban areas in the US. We develop an epidemiological model that takes into consideration prevalent security flaws on these routers. The spread of such a contagion is simulated on real-world data for geo-referenced wireless routers. We uncover a major weakness of WiFi networks in that most of the simulated scenarios show tens of thousands of routers infected in as little time as two weeks, with the majority of the infections occurring in the first 24 to 48 hours. We indicate possible containment and prevention measure to limit the eventual harm of such an attack.
Correspondence and requests for material should be addressed to S.M. (samyers@indiana.edu) or A.V. (alexv@indiana.edu).
The use of WiFi routers is becoming close to mainstream in the US and Europe, with 8.4% and 7.9% of *all* such respective households having deployed such routers by 2006 <sup>[1](#bib.bib1)</sup>, and a WiFi market expected to grow quickly in the next few years as more new digital home devices are being shipped with WiFi technology. In 2006 alone 200 million WiFi chipsets were shipped worldwide, representing nearly half of the 500 million cumulative total <sup>[2](#bib.bib2)</sup>. As the WiFi deployment becomes more and more pervasive, the larger is the risk that massive attacks exploiting the WiFi security weaknesses could affect large numbers of users.
Recent years have witnessed a change in the designers behind a malware attack and in their motivations, corresponding to the ever increasing sophistication needed to bypass newly developed security technologies. Malware creators have shifted from programmer enthusiasts attempting to get peer credit from the “hacker” community, to organized crime engaging in fraud and money laundering through different forms of online crime. In this context WiFi routers represent valuable targets when compared to the PC’s that malware traditionally infect, as they have several differing properties representing strong incentives for the attacks. They are the perfect platform to launch a number of fraudulent attacks <sup>[3](#bib.bib3), [4](#bib.bib4), [5](#bib.bib5), [6](#bib.bib6)</sup> that previous security technologies have reasonably assumed were unlikely <sup>[7](#bib.bib7)</sup>. Unlike PCs, they tend to be always on and connected to the Internet, and currently there is no software aimed at specifically detecting or preventing their infection. Further, as routers need to be within relatively close proximity to each other to communicate wirelessly, an attack can now take advantage of the increasing density of WiFi routers in urban areas that creates large geographical networks where the malware can propagate undisturbed. Indeed, many WiFi security threats have been downplayed based on the belief that the physical proximity needed for the potential attack to occur would represent an obstacle for attackers. The presence nowadays of large ad-hoc networks of routers make these vulnerabilities considerably more risky than previously believed.
Here we assess for the first time the vulnerability of WiFi networks of different US cities by simulating the wireless propagation of malware, a malicious worm spreading directly from wireless router to wireless router. We construct an epidemiological model that takes into account several widely known and prevalent weaknesses in commonly deployed WiFi routers’ security <sup>[3](#bib.bib3), [8](#bib.bib8)</sup>, (e.g., default and poor password selection and cracks in the WEP cryptographic protocol <sup>[9](#bib.bib9)</sup>). The WiFi proximity networks over which the attack is simulated are obtained from real-world geographic location data for wireless routers. The infection scenarios obtained for a variety of US urban areas are troublesome in that the infection of a small number of routers in most of these cities can lead to the infection of tens of thousands of routers in a week, with most of the infection occurring in the first 24 hours. We address quantitatively the behavior of the spreading process and we provide specific suggestions to minimize the WiFi network weakness and mitigate an eventual attack.
Results and Discussion
WiFi networks. WiFi routers, even if generally deployed without a global organizing principle, define a self-organized proximity communication network. Indeed, any two routers which are in the range of each other’s WiFi signal can exchange information and may define an ad-hoc communication network. These networks belong to the class of *spatial* or *geometric networks* in that nodes are embedded in a metric space and the interaction between two nodes strongly depends on the range of their spatial interaction <sup>[10](#bib.bib10), [11](#bib.bib11), [12](#bib.bib12), [13](#bib.bib13)</sup>.
In this perspective, one might wonder if the actual deployment of WiFi routers is sufficient at the moment to generate large connected networks spanning sizeable geographic areas. This problem, equivalent to the percolation of giant connected component in graph theory<sup>[14](#bib.bib14), [15](#bib.bib15)</sup>, is however constrained by the urban area’s topology and demographic distribution dictating the geographical locations of WiFi routers. Here we consider WiFi networks as obtained from the public worldwide database of the Wireless Geographic Logging Engine (WiGLE) website <sup>[16](#bib.bib16)</sup>. The database collects data on the worldwide geographic location of wireless routers and counts more than $`10`$ million unique networks on just under 600 million observations <sup>[17](#bib.bib17)</sup>, providing good coverage of the wireless networks in the United States and in North Central Europe. The data provide a wealth of information that include, among other things, the routers’ geographic locations (expressed in latitude $`L\hspace{0pt}A\hspace{0pt}T`$ and longitude $`L\hspace{0pt}O\hspace{0pt}N`$) and their encryption statuses. In particular, we focused on the wireless data extracted from seven urban areas or regions within the United States – Chicago, Boston, New York City, San Francisco Bay Area, Seattle, and Northern and Southern Indiana. Starting from the set of vertices corresponding to geo-referenced routers in a given region, we construct the proximity network <sup>[10](#bib.bib10), [11](#bib.bib11), [12](#bib.bib12), [13](#bib.bib13)</sup> by drawing an edge between any two routers $`i`$ and $`j`$ located at $`{\overset{\rightarrow}{p}}_{i} = {({L\hspace{0pt}O\hspace{0pt}N_{i}},{L\hspace{0pt}A\hspace{0pt}T_{i}})}`$ and $`{\overset{\rightarrow}{p}}_{j} = {({L\hspace{0pt}O\hspace{0pt}N_{j}},{L\hspace{0pt}A\hspace{0pt}T_{j}})}`$, respectively, whose geographical distance $`d\hspace{0pt}{({\overset{\rightarrow}{p}}_{i},{\overset{\rightarrow}{p}}_{j})}`$ is smaller than the maximum interaction radius $`R_{i\hspace{0pt}n\hspace{0pt}t}`$ (i.e., $`{d\hspace{0pt}{({\overset{\rightarrow}{p}}_{i},{\overset{\rightarrow}{p}}_{j})}} \leq R_{i\hspace{0pt}n\hspace{0pt}t}`$), as shown in Figure [1](#S0.F1 "Figure 1 ‣ WiFi Epidemiology: Can Your Neighbors’ Router Make Yours Sick?")A. In the WiFi networks, the maximum interaction radius $`R_{i\hspace{0pt}n\hspace{0pt}t}`$ strongly depends on the local environment of any specific router. In practice, $`R_{i\hspace{0pt}n\hspace{0pt}t}`$ ranges from 15m for a closed office with poor transmission to approximately 100m outdoors <sup>[18](#bib.bib18)</sup>. For simplicity, we assume that $`R_{i\hspace{0pt}n\hspace{0pt}t}`$ is constant, independent of the actual location of a given router, and we consider four different values of the maximum interaction radius — $`R_{i\hspace{0pt}n\hspace{0pt}t} \in {\{\text{15m, 30m, 45m, 100m}\}}`$ — analyzing the resulting networks for each of the seven regions under study. A more detailed account of the network construction procedure and the filtering methods used to minimize potential biases introduced by the data collection mechanisms are described in the Materials and Method section.
In Figure [1](#S0.F1 "Figure 1 ‣ WiFi Epidemiology: Can Your Neighbors’ Router Make Yours Sick?")B we report an illustration of the giant component of the network obtained in the Chicago area for different values of $`R_{i\hspace{0pt}n\hspace{0pt}t}`$. It is possible to observe that despite the clear geographical embedding and the city constraints, a large network of more than 48,000 routers spans the downtown area for $`R_{i\hspace{0pt}n\hspace{0pt}t}`$ set to 45 meters. The degree distributions of the giant components, reported in Figure [1](#S0.F1 "Figure 1 ‣ WiFi Epidemiology: Can Your Neighbors’ Router Make Yours Sick?")C, are characterized by an exponential decrease <sup>[12](#bib.bib12)</sup> with a cutoff clearly increasing with the interaction radius, since a larger range increases the number $`k`$ of nodes found within the signal area. Very similar properties are observed in all the networks analyzed. It is important to stress that the metric space embedding exerts a strong preventative force on the small-world behavior of the WiFi networks, since the limited WiFi interaction rules out the possibility of long range connections.
Infecting a Router. The infection of a susceptible router occurs when the malware of an already infected router is able to interface with the susceptible’s administrative interface over the wireless channel. Two main technologies aim at preventing such infection through i) the use of encrypted and authenticated wireless channel communication through the WEP and WPA cryptographic protocols, and ii) the use of a standard password for access control. The encryption should provide a higher level of security, as it needs to be bypassed before a potential attacker could attempt to enter the router’s password. Most users do not currently employ their routers encryption capabilities – indeed the encryption rates in the considered cities vary from 21% to 40% of the population. For the purposes of this work we assume that WPA is not vulnerable to attack
<sup>1</sup>
<sup>1</sup>1This is not completely accurate, see supplementary discussion for a more in depth discussion.
, and therefore any router that uses it is considered *immune* to the worm. Because of cryptographic flaws in WEP, this protocol can always be broken given that the attacker has access to enough encrypted communication. This can be achieved by waiting for the router to be used by legitimate clients, or by deploying more advanced active attacks. Bypassing WEP encryption is therefore feasible and only requires a given amount of time.
Once the malware has bypassed any cryptographic protocol and established a communication channel, it may then attempt to bypass the password. A large percentage of users do not change their password from the default established by the router manufacturer, and these passwords are easily obtainable. For legal reasons it is difficult to measure exactly what this percentage is, so here we use as a proxy the percentage of users who do not change their routers default SSID. For all the other routers, we assume that $`25\%`$ of them can have the password guessed with 65,000 login attempts, based on the evidence provided by security studies <sup>[19](#bib.bib19)</sup> which showed that $`\approx {25\%}`$ of all users passwords are contained in a dictionary of 65,000 words. We then pessimistically assume, based on previous worms, that another $`11\%`$ of passwords are contained in a larger library of approximately a million words <sup>[20](#bib.bib20)</sup>. No backoff mechanism exists on the routers that prevents systematic dictionary attacks. In case the password is not found in either dictionary, the attack cannot proceed. Alternatively, if the password has been overcome, the attacker can upload the worm’s code into the router’s firmware, a process that typically takes just a few minutes. In the Material and Methods section we report a list of the typical time scales related to each step of the attack strategy.
Construction of the epidemic model. The construction of the wireless router network defines the population and the related connectivity pattern over which the epidemic will spread. In order to describe the dynamical evolution of the epidemic (i.e., the number of infected routers in the population as a function of time) we use a framework analogous to epidemic modeling that assumes that each individual (i.e. each router) in the population is in a given class depending on the stage of the infection <sup>[21](#bib.bib21)</sup>. Generally, the basic modeling approaches consider three classes of individuals: susceptible (those who can contract the infection), infectious (those who contracted the infection and are contagious), and recovered (those who recovered or are immune from the disease and cannot be infected). In our case the heterogeneity of the WiFi router population in terms of security attributes calls for an enlarged scheme that takes into account the differences in the users’ security settings. We consider three basic levels of security and identify the corresponding classes: routers with no encryption, which are obviously the most exposed to the attack, are mapped into a first type of susceptible class $`S`$; routers with WEP encryption, which provides a certain level of protection that can be eventually overcome with enough time, are mapped into a second type of susceptible class denoted $`S_{WEP}`$; routers with WPA encryption, which are assumed to resist any type of attacks, correspond to the removed class $`R`$. This classification however needs to be refined to take into account the password settings of the users that range from a default password to weak or strong passwords and finally to non-crackable passwords. For this reason, we can think of the non-encrypted class $`S`$ as being subdivided into four subclasses. First, we distinguish between the routers with default password $`S_{nopass}`$ and the ones with password $`S_{p\hspace{0pt}a\hspace{0pt}s\hspace{0pt}s\hspace{0pt}1}`$. The latter contains routers with all sorts of passwords that undergo the first stage of the attack which employs the smaller dictionary. If this strategy fails, the routers are then classified as $`S_{p\hspace{0pt}a\hspace{0pt}s\hspace{0pt}s\hspace{0pt}2}`$ and undergo the attack which employs the larger dictionary. Finally, if the password is unbreakable, the router is classified as $`R_{h\hspace{0pt}i\hspace{0pt}d\hspace{0pt}d\hspace{0pt}e\hspace{0pt}n}`$. The last class represents routers whose password cannot be bypassed. However, their immune condition is *hidden* in that it is known only to the attacker who failed in the attempt, while for all the others the router appears in the susceptible class as it was in its original state. This allows us to model the unsuccessful attack attempts of other routers in the dynamics. WEP encrypted routers have the same properties in terms of password, but the password relevance starts only when the WEP encryption has been broken on the router. At this stage of the attack it can be considered to be in the non-encrypted state, and therefore no subclasses of $`S_{WEP}`$ have to be defined. In addition to the above classes, the model includes the infected class ($`I`$) with those routers which have been infected by the malware and have the ability to spread it to other routers.
The model dynamics are specified by the transition rates among different classes for routers under attack. Transitions will occur only if a router is attacked and can be described as a reaction process. For instance the infection of a non-encrypted router with no password is represented by the process $`{S_{nopass} + I}\rightarrow{2\hspace{0pt}I}`$. The transition rates are all expressed as the inverse of the average time needed to complete the attack. In the above case the average time of the infection process is $`\tau = 5`$ minutes and the corresponding rate $`\beta`$ for the transition $`{S_{nopass} + I}\rightarrow{2\hspace{0pt}I}`$ is $`\beta = \tau^{- 1}`$. Similarly the time scale $`\tau_{WEP}`$ needed to break a WEP encryption will define the rate $`\beta_{WEP}`$ ruling the transition from the $`S_{WEP}`$ to the non-encrypted class. In the Materials and Methods section we report in detail all the transition processes and the associated rates defining the epidemic processes.
One of the most common approaches to the study of epidemic processes is to use deterministic differential equations based on the assumption that individuals mix homogeneously in the population, each of them potentially in contact with every other <sup>[21](#bib.bib21)</sup>. In our case, the static non-mobile nature of wireless routers and their geographical embedding make this assumption completely inadequate, showing the need to study the epidemic dynamics by explicitly considering the underlying contact pattern <sup>[22](#bib.bib22), [23](#bib.bib23), [24](#bib.bib24), [25](#bib.bib25), [26](#bib.bib26)</sup>. For this reason, we rely on numerical simulations obtained by using an individual-based modeling strategy. At each time step the stochastic disease dynamics is applied to each router by considering the actual state of the router and those of its neighbors as defined by the actual connectivity pattern of the network. It is then possible to measure the evolution of the number of infected individuals and keep track of the epidemic progression at the level of single routers. In addition, given the stochastic nature of the model, different initial conditions can be used to obtain different evolution scenarios.
As multiple seed attacks are likely we report simulations with initial conditions set with $`5`$ infected routers randomly distributed within the population under study. Single seed attacks and different number of initial seeds have similar effects. The initial state of each router is directly given by the real WiFi data or is obtained from estimates based on real data, as detailed in the Material and Methods section. Finally, for each scenario we perform averages over 100 realizations.
Spreading of synthetic epidemics. According to the simulation procedure outlined above we study the behavior of synthetic epidemics in the seven urban areas we used to characterize the properties of WiFi router networks. The urban areas considered are quite diverse in that they range from a relatively small college town as West Lafayette (Indiana) to big metropolis such as New York city and Chicago. In each urban area we focus on the giant component of the network obtained with a given $`R_{i\hspace{0pt}n\hspace{0pt}t}`$ that may vary consistently in size.
Here we report the results for a typical epidemic spreading scenario in which the time scales of the processes are chosen according to their average estimates. The best and worst case scenarios could also be obtained by considering the combination of parameters that maximize and minimize the rate of success of each attack process, respectively. The networks used as substrate are obtained in the intermediate interaction range of 45m.
The four snapshots of Figure [2](#S0.F2 "Figure 2 ‣ WiFi Epidemiology: Can Your Neighbors’ Router Make Yours Sick?") provide an illustration of the evolution of a synthetic epidemic in the Chicago area; shown in red are the routers which are progressively infected by malware. The striking observation is that the malware rapidly propagates on the WiFi network in the first few hours, taking control of about 37% of the routers after two weeks from the infection of the first router. The quantitative evidence of the potential impact of the epidemic is reported in Figure [3](#S0.F3 "Figure 3 ‣ WiFi Epidemiology: Can Your Neighbors’ Router Make Yours Sick?")A-B, where the average profile of the density of infected routers is reported for all the urban areas considered in the numerical experiment, together with the corresponding fluctuations. While it is possible to notice a considerable difference among the various urban areas, in all cases we observe a sharp rise of the epidemic within the first couple of days and then a slower increase, which after two weeks leaves about 10% to 55% of the routers in the giant component controlled by malware. The similar time scale in the rise of the epidemic in different urban areas is not surprising as it is mainly determined by the time scale of the specific attacks considered in the malware spreading model. In general the sharp rise of the epidemic in its early stages is due to the non-encrypted routers which are infected in a very short time. The slower progression at later stages is instead due to the progressive infection of WEP routers whose attack time scale is about one order of magnitude longer. Single realization results clearly show the effect of the interplay of different time scales involved in the spreading phenomenon.
A more complicated issue is understanding the different attack (infection) rates that the epidemic attains in different urban area networks. The pervasiveness of the epidemic can be seen as a percolation effect on the WiFi network <sup>[27](#bib.bib27), [28](#bib.bib28)</sup>. The WPA encrypted routers and those with unbreakable passwords represent obstacles to the percolation process and define an effective percolation probability that has to be compared with the intrinsic percolation threshold of the network<sup>[28](#bib.bib28), [29](#bib.bib29), [30](#bib.bib30)</sup>. The larger the effective percolation probability with respect to the threshold, the larger the final density of infected routers. On the other hand, the epidemic thresholds of the networks are not easy to estimate because they are embedded in the particular geometries of the cities’ geographies. In random networks, large average degree and large degree fluctuations favor the spreading of epidemics and tend to reduce the network percolation threshold <sup>[26](#bib.bib26), [31](#bib.bib31)</sup>. Figure [3](#S0.F3 "Figure 3 ‣ WiFi Epidemiology: Can Your Neighbors’ Router Make Yours Sick?")C-D shows an appreciable statistical correlation between the attack rate and these quantities. On the other hand, there are many other network features that affect the percolation properties of the networks. First, the cities have different fractions of encrypted routers. While these fraction are not extremely dissimilar, it is clear that given the non-linear effect close to the percolation threshold, small differences may lead to large difference in the final attack rate. For instance, San Francisco, with the largest fraction of encrypted routers corresponding to about 40% of the population, exhibits the smallest attack rate amongst all the urban areas considered. Second, the geometrical constraints imposed by the urban area geography may have a large impact on the percolation threshold, which can be rather sensitive to the local graph topology. For instance, network layouts with one dimensional bottlenecks or locally very sparse connectivity may consistently lower the attack rate by sealing part of the network, and thus protecting it from the epidemic. Indeed, a few WPA routers at key bottlenecks can make entire subnetworks of the giant component impenetrable to the malware.
The present results offer general quantitative conclusions on the impact and threat offered by the WiFi malware spreading in different areas, whereas the impact of specific geographical properties of each urban area on the epidemic pattern will be the object of further studies.
Conclusions
Based on this work, we note that there is a real concern about the wireless spread of WiFi based malware. This suggests that action needs to be taken to detect and prevent such outbreaks, as well as more thoughtful planning for the security of future wireless devices, so that such scenarios do not occur or worsen with future technology. For instance, given the increasing popularity of 802.11n, with its increased wireless communications range, the possibility for larger infections to occur is heightened, due to the larger connected components that will emerge. Further, many devices such as printers, and DVR systems now ship with 802.11 radios in them, and if they are programmable, as many are, they become vulnerable in manners similar to routers. Lastly, it is highly likely that we will only see the proliferation of more wireless standards as time goes by, and all of these standards should consider the possibility of such epidemics.
There are two preventive actions that can be easily considered to successfully reduce the rates of infection. First, force users to change default passwords, and secondly the adoption of WPA, the cryptographic protocol meant to replace WEP that does not share its cryptographic weaknesses. Unfortunately, the dangers of poorly chosen user passwords have been widely publicized for at least two decades now, and there has been little evidence of a change in the public’s behavior. In addition, there are many barriers to public adoption of WPA on wireless routers. The use of only one device in the home that does not support WPA, but that does support the more widely implemented WEP, is sufficient to encourage people to use WEP at home. However, unlike the more traditional realm of internet malware, the lack of a small world contagion graph implies that small increases in the deployment of WPA or strong passwords can significantly reduce the size of a contagion graph’s largest connected component, significantly limiting the impact of such malware. In future work we plan a detailed study of the percolation threshold of the giant component as a function of the proportion of WPA immune nodes in order to provide quantitative estimates for the systems’ immunization thresholds.
Materials and Methods
WiFi Data and Networks. WiFi data is downloaded from the WiGLE website <sup>[16](#bib.bib16)</sup> for seven urban areas in the US and is processed in order to eliminate potential biases introduced by data collection. Records that appear as *probe* in their type classification are removed from the dataset since they correspond to wireless signals originating from non-routers. Such records represent a very small percentage of the total number in every city considered. For example, in the Chicago urban area there were 4433 *probe* records, corresponding to 3.7% of the total.
A preliminary spatial analysis of the data for each urban area reveals the presence of sets of several WiFi routers sharing an identical geographic location. In order to avoid biases due to overrepresentation of records, we checked for unique BSSID (i.e., MAC address) and assume that each of these locations could contain at most $`n`$ overlapping routers, where $`n`$ was fixed to 20 to provide a realistic scenario, such as a building with several hotspots. For the Chicago urban area, this procedure led to the elimination of 3194 records, which represent 2.7% of the total number of WiFi routers.
More importantly, we adopt a randomized procedure to redefine the position of each router in a circle of radius $`R_{r\hspace{0pt}a\hspace{0pt}n}`$ centered on the GPS coordinates provided by the original data. This procedure is applied to approximate the actual location of each router which would be otherwise localized along city streets, due to an artifact of the wardriving data collection method <sup>[16](#bib.bib16), [32](#bib.bib32)</sup>. The newly randomized positions for the set of routers completely determine the connectivity pattern of the spatial WiFi network and its giant component substrate for the epidemic simulation. Results presented here are obtained as 5 averages over several randomization procedures. Table [1](#S0.T1 "Table 1 ‣ WiFi Epidemiology: Can Your Neighbors’ Router Make Yours Sick?") reports the main topological indicators of the giant components of each urban area extracted from the WiFi network built assuming $`R_{i\hspace{0pt}n\hspace{0pt}t} = 45`$m.
Epidemic model. Figure [4](#S0.F4 "Figure 4 ‣ WiFi Epidemiology: Can Your Neighbors’ Router Make Yours Sick?") shows the flow diagram of the transmission model.
Initial conditions set the number of routers belonging to each of the following compartments: $`S_{n\hspace{0pt}o\hspace{0pt}p\hspace{0pt}a\hspace{0pt}s\hspace{0pt}s}`$ (routers with no encryption and default password), $`S_{p\hspace{0pt}a\hspace{0pt}s\hspace{0pt}s\hspace{0pt}1}`$ (routers with no encryption and user set password), $`S_{WEP}`$ (routers with WEP encryption), and $`R`$ (routers with WPA encryption, here considered immune). The classes $`S_{p\hspace{0pt}a\hspace{0pt}s\hspace{0pt}s\hspace{0pt}2}`$ and $`R_{h\hspace{0pt}i\hspace{0pt}d\hspace{0pt}d\hspace{0pt}e\hspace{0pt}n}`$ are void at the beginning of the simulations since they represent following stages of the infection dynamics. Encrypted routers are identified from original data, and the fraction of $`R`$ out of the total number of encrypted routers is assumed to be $`30\%`$, in agreement with estimates on real world WPA usage. Analogously, we assume that the non-encrypted routers are distributed according to the following proportions: 50% in class $`S_{n\hspace{0pt}o\hspace{0pt}p\hspace{0pt}a\hspace{0pt}s\hspace{0pt}s}`$ and 50% in class $`S_{p\hspace{0pt}a\hspace{0pt}s\hspace{0pt}s\hspace{0pt}1}`$.
The infection dynamics proceeds as follows. A router with no encryption enters the infectious class with unitary rate $`\beta`$ if attacked. The attack to a router in class $`S_{p\hspace{0pt}a\hspace{0pt}s\hspace{0pt}s\hspace{0pt}1}`$ is characterized by a transition rate $`\beta_{1}`$ and has two possible outcomes: with probability $`({1 - p_{1}})`$ the router is infected and enters $`I`$, whereas with probability $`p_{1}`$ it enters $`S_{p\hspace{0pt}a\hspace{0pt}s\hspace{0pt}s\hspace{0pt}2}`$ since the attacker is not able to overcome the password and the infection attempt requires additional time and resources. Once in class $`S_{p\hspace{0pt}a\hspace{0pt}s\hspace{0pt}s\hspace{0pt}2}`$, it can become infectious with probability $`({1 - p_{2}})`$ if the attack is successful, or otherwise the router enters $`R_{h\hspace{0pt}i\hspace{0pt}d\hspace{0pt}d\hspace{0pt}e\hspace{0pt}n}`$ with probability $`p_{2}`$ because the password has not been bypassed. This process occurs with a transition rate $`\beta_{2}`$. WEP encrypted routers follow the same dynamics once the encryption is broken and they enter $`S_{p\hspace{0pt}a\hspace{0pt}s\hspace{0pt}s\hspace{0pt}1}`$ with transition rate $`\beta_{W\hspace{0pt}E\hspace{0pt}P}`$.
The numerical simulations consider the discrete nature of the individuals and progress in discrete time steps. We assume that the attacker will target the router, among its neighbors, with the lowest security settings. In addition, we do not allow simultaneous attacks, so that each infected router will choose its next target only among those routers which are not already under attack. Once an attack has started, the attacker will keep trying to bypass the security setting of the same target until the attempt is finally successful or not. In both cases, the attacker will then move to another target. The simulation’s unitary time step is defined by the shortest time scale among all processes involved, i.e. the time $`\tau`$ needed to complete an attack to a non-encrypted router with no password. This automatically defines as unitary the transition rate $`\beta`$ associated to the reaction $`{S_{nopass} + I}\rightarrow{2\hspace{0pt}I}`$, given $`{\beta\hspace{0pt}\tau} = 1`$. Typical time scales for the other processes are: $`\tau_{1} =`$ 6 – 15 minutes to bypass a password in the smaller dictionary, $`\tau_{2} =`$ 400 – 1000 minutes to bypass a password in the larger dictionary, $`\tau_{W\hspace{0pt}E\hspace{0pt}P} =`$ 2880 – 5760 minutes to crack the WEP encryption. The corresponding transition rates can be analogously defined as probabilities expressed in terms of their ratio with $`\beta`$ that defines the unitary rate.
Simulations run for 4,032 time steps, corresponding to 20,160 minutes (i.e., 2 weeks). At each time step we measure the global attack rate defined as the number of infectious $`I\hspace{0pt}{(t)}`$ at time $`t`$ over the total population of the network discounted by the number of recovered, $`N - R`$. In this way, we can take into account for the differences of the encryption percentages observed in different urban areas. Given the stochasticity of the process, the numerical simulations are repeated 100 times changing initial conditions and assuming different configurations of the randomized network. Average values and corresponding confidence intervals are also measured.
Acknowledgments
H.H. thanks the Institute for Scientific Interchange in Turin for its hospitality during the time this work was completed. A.V. is partially supported by the NSF award IIS-0513650.
## References
- 1
Mercer, D. (2006) Home Network Adoption: Wi-Fi Emerges As Mass Market Phenomenon (Market Report).
http://www.strategyanalytics.net/default.aspx?mod=ReportAbstract Viewer&%a0=2909.
- 2
Telecoms Europe website. (2007) Wi-fi chipset shipments reach 200m in 2006 – report.
http://www.telecomseurope.net/article.php?id_article=3549.
- 3
Stamm, S., Ramzan, Z., and Jakobsson, M. (2006) Drive-by pharming (Technical Report 641, Indiana University).
- 4
Ollmann, G. (2006) The pharming guide (Technical report, Next Generation Security Software Ltd.).
- 5
Jakobsson, M. and Myers, S. editors. (2007) Phishing and Countermeasures: Undertanding the Increasing Problem of Electronic Identity Theft (Wiley).
- 6
Soghoian, C. (2007) A deceit-augmented man in the middle attack against Bank of America’s SiteKey service.
http://paranoia.dubfire.net/2007/04/deceit-augmented-man-in-middle-attack.html.
- 7
Myers, S. and Stamm, S. (2007) Trawler phishing (Submitted to Computer and Communications Security Conference).
- 8
Tsow, A., Jakobsson, M., Yang, L. and Wetzel, S. (2006) Journal of Digital Forensic Practice, 1, Special Issue 3.
- 9
Bittau, A., Handley, M. and Lackey, J. (2006) SP ’06: Proceedings of the 2006 IEEE Symposium on Security and Privacy (S&P’06) (Washington, DC, USA. IEEE Computer Society), 386 – 400
- 10
Dall, J. and Christensen, M. (2002) Phys. Rev. E, 66, 016121.
- 11
Nemeth, G. and Vattay, G. (2003) Phys. Rev. E, 67, 036110.
- 12 Herrmann, C., Barthélemy, M. and Provero, P. (2003)
Phys. Rev. E, 68, 026128.
- 13
Helmy, A. (2003) IEEE Comm. Lett., 7, 490 – 492.
- 14
Molloy, M. and Reed B. (1995) Random Struct Alg., 6, 161 – 179.
- 15
Bollobas, B. and Riordan, O. (2006) Percolation (Cambridge University Press).
- 16 WiGLE: Wireless Geographic Logging Engine. http://www.wigle.net/.
- 17 WiGLE General Stats, first accessed on Febrary, 2007.
http://www.wigle.net/gps/gps/main/stats/.
- 18
Gast, M. (2005) 802.11 Wireless Networks: The Definitive Guide, Second Edition (O’Reilly).
- 19
Klein, D. V. (1990) Proceedings of the second USENIX Workshop on Security, 5 – 14.
- 20
Jeff, J., Alan, Y., Ross, B. and Alasdair, A. (2000) The memorability and security of passwords – some empirical results (Technical Report No. 500, Computer Laboratory, University of Cambridge).
http://www.ftp.cl.cam.ac.uk/ftp/users/rja14/tr500.pdf
- 21
Anderson, R. M. and May, R. M. (1992) Infectious Diseases of Humans: Dynamics and Control (Oxford Univ. Press, Oxford).
- 22
Watts, D. J. and Strogatz, S. H. (1998) Nature, 393, 440 – 442.
- 23
Barabási, A. L. and Albert, R. (1999) Science, 286, 509 – 512.
- 24 Keeling, M. J. (1999)
Proc. R. Soc. Lond. B, 266, 859 – 867.
- 25
Moore, C. and Newman, M. E. J. (2000) Phys. Rev. E, 61, 5678 – 5682.
- 26
Pastor-Satorras, R. and Vespignani, A. (2001) Phys. Rev. Lett., 86, 3200 – 3203.
- 27
Grassberger, P. (1983) Math. Biosci., 63, 157 –172.
- 28
Ben-Avraham, B. and Havlin, S. (2000) Diffusion and Reactions in Fractals and Disordered Systems (Cambridge University Press).
- 29
Cohen, R., Erez, K., Ben-Avraham, D. and Havlin, S. (2000) Phys. Rev. Lett., 85, 4626 – 4628.
- 30
Callaway, D. S., Newman, M. E. J., Strogatz, S. H. and Watts, D. J. (2000) Phys. Rev. Lett., 85, 5468 – 5471.
- 31 Lloyd, A. L. and May, R. M. (2001)
Science, 292,1316 – 1317.
- 32 Wardriving. http://www.wardriving.com/.
Table and Figure captions
Table 1. Properties of the giant components of the WiFi networks for $`R_{i\hspace{0pt}n\hspace{0pt}t} =`$45m: size of the giant component $`N`$, percentage of encrypted routers $`f_{e\hspace{0pt}n\hspace{0pt}c\hspace{0pt}r}`$, maximum degree $`k_{m\hspace{0pt}a\hspace{0pt}x}`$, average degree $`\langle k\rangle`$, degree fluctuations $`{\langle k\rangle}^{2}/{\langle k\rangle}`$. The results presented are obtained as averages over 5 different randomization procedures to redefine the location of each router.
Figure 1. (A): Construction of WiFi networks. Given two routers $`i`$ and $`j`$ located at $`{\overset{\rightarrow}{p}}_{i} = {({L\hspace{0pt}O\hspace{0pt}N_{i}},{L\hspace{0pt}A\hspace{0pt}T_{i}})}`$ and $`{\overset{\rightarrow}{p}}_{j} = {({L\hspace{0pt}O\hspace{0pt}N_{j}},{L\hspace{0pt}A\hspace{0pt}T_{j}})}`$, we place an edge between them if their distance $`d\hspace{0pt}{({\overset{\rightarrow}{p}}_{i},{\overset{\rightarrow}{p}}_{j})}`$ is smaller than the maximum interaction radius $`R_{i\hspace{0pt}n\hspace{0pt}t}`$. (B): Map representation of the giant components of the WiFi network in the Chicago area as obtained with different values of $`R_{i\hspace{0pt}n\hspace{0pt}t}`$. (C): The degree distribution for different values of the interaction radius $`R_{i\hspace{0pt}n\hspace{0pt}t}`$ show an exponential decay and a cutoff which depends on $`R_{i\hspace{0pt}n\hspace{0pt}t}`$. The result is obtained as averages over 5 different randomization procedures to redefine the location of each router.
Figure 2. Illustration of the spread of a wireless worm through Chicago in several time slices. In this series, the result is based on 1 randomization procedure for the location of each router and the maximum interaction radius $`R_{i\hspace{0pt}n\hspace{0pt}t}`$ is set to 45 meters.
Figure 3. (A): Attack rate versus time for the giant component of all the seven urban areas, keeping $`R_{i\hspace{0pt}n\hspace{0pt}t} = {45\hspace{0pt}m}`$. (B): The average and 90% C. I. for three prototypical cases. (C): The correlation between the final attack rate and average degree. (D): The correlation between the final attack rate and degree fluctuations.
Figure 4. Compartmental Flows for the Epidemic Model.
Table 1:
<table>
<tbody>
<tr>
<td>City</td>
<td>N𝑁N</td>
<td>fencrsubscript𝑓𝑒𝑛𝑐𝑟f_{encr}</td>
<td>kmaxsubscript𝑘𝑚𝑎𝑥k_{max}</td>
<td>⟨k⟩delimited-⟨⟩𝑘\langle k\rangle</td>
<td>⟨k⟩2⟨k⟩superscriptdelimited-⟨⟩𝑘2delimited-⟨⟩𝑘\frac{\langle k\rangle^{2}}{\langle k\rangle}</td>
<td></td>
</tr>
<tr>
<td>NYC</td>
<td>36807</td>
<td>25.8%</td>
<td>109</td>
<td>23.9</td>
<td>31.3</td>
<td></td>
</tr>
<tr>
<td>Boston</td>
<td>15899</td>
<td>21.7%</td>
<td>116</td>
<td>21.3</td>
<td>35.7</td>
<td></td>
</tr>
<tr>
<td>Seattle</td>
<td>3013</td>
<td>26.0%</td>
<td>49</td>
<td>11.9</td>
<td>16.7</td>
<td></td>
</tr>
<tr>
<td>Chicago</td>
<td>50084</td>
<td>33.7%</td>
<td>154</td>
<td>20.3</td>
<td>33.6</td>
<td></td>
</tr>
<tr>
<td>N.IN</td>
<td>2629</td>
<td>24.3%</td>
<td>87</td>
<td>18.4</td>
<td>29.9</td>
<td></td>
</tr>
<tr>
<td>S.IN</td>
<td>998</td>
<td>11.0%</td>
<td>32</td>
<td>10.4</td>
<td>14.1</td>
<td></td>
</tr>
<tr>
<td>SF Bay</td>
<td>3106</td>
<td>40.1%</td>
<td>31</td>
<td>9.6</td>
<td>12.8</td>
<td></td>
</tr>
</tbody>
</table>
Table 1.
[Refer to caption](/html/0706.3146/assets/GC_properties.png)
Figure 1.
Figure 1:
[Refer to caption](/html/0706.3146/assets/maps_snapshots.png)
Figure 2:
Figure 2.
[Refer to caption](/html/0706.3146/assets/attack_rate.png)
Figure 3:
Figure 3.
[Refer to caption](/html/0706.3146/assets/compartment_flow.png)
Figure 4:
Figure 4. |
0706.2748 | cs.OS | cyber | INSTITUT NATIONAL DE RECHERCHE EN INFORMATIQUE ET EN AUTOMATIQUE
A Survey of Unix Init Schemes
Yvan Royon — Stéphane FrénotN° 0338
June 2007
A Survey of Unix Init Schemes
Yvan Royon, Stéphane Frénot
Thème COM — Systèmes communicants
Projet Ares
Rapport technique n° 0338 — June 2007 — ?? pages
Abstract:
In most modern operating systems, init (as in “initialization”) is the program launched by the kernel at boot time. It runs as a daemon and typically has PID 1. Init is responsible for spawning all other processes and scavenging zombies. It is also responsible for reboot and shutdown operations.
This document describes existing solutions that implement the init process and/or init scripts in Unix-like systems. These solutions range from the legacy and still-in-use BSD and SystemV schemes, to recent and promising schemes from Ubuntu, Apple, Sun and independent developers. Our goal is to highlight their focus and compare their sets of features.
Key-words: init process, init scheme
Etat de l’art des processus init
Résumé :
Dans la plupart des systèmes d’exploitation modernes, init (comme “initialisation”) est l’application lancée par le noyau au démarrage. Il s’agit d’un démon qui a le PID 1. init a pour tâches de lancer les autres processus et de nettoyer les processus zombies. Il est aussi responsable du démarrage et redémarrage du système.
Ce document décrit les implantations existantes du processus init et des scripts init pour variantes d’Unix. Ces implantations vont des solutions BSD et SystemV, pionnières et toujours utilisées, à celles plus récentes et prometteuses venant de Ubuntu, Apple, Sun, ainsi que de développeurs indépendants. Notre but ici est d’analyser les objectifs de ces implantations, et de comparer leurs fonctionnalités.
Mots-clés : processus init
## The Boot Process
What happens between the moment a computer is switched on and the moment a user can log in?
Let us begin by a quick and simplified overview of the boot process in most modern operating systems.
1. 1\.
The computer is switched on or rebooted;
2. 2\.
The BIOS, stored on the motherboard, is executed from a known location (flash memory);
3. 3\.
The BIOS determines the boot device (local disc, removable media, PXE from the network);
4. 4\.
The boot device has a special, known sector called the Master Boot Record, which contains the stage 1 bootloader. The stage 1 bootloader is loaded into RAM, and points to the stage 2 bootloader. The latter can be several sectors long, and is stored on the boot device in a location unknown to the BIOS and subject to change;
5. 5\.
The stage 2 bootloader (e.g., LILO or GRUB for Linux) uncompresses and launches the kernel;
6. 6\.
The kernel runs a single application. For Linux, the default is /sbin/init, and can be modified in LILO or GRUB by passing the init=/path/to/executable parameter to the kernel.
Usually, init launches a few getty processes for virtual terminals, a few daemons (cron, dhcp, acpi), the X server, etc.
On Unix-like systems, init (as in “initialization”) is the program launched by the kernel at boot time. It runs as a daemon and typically has PID 1. init is responsible for spawning all other processes and scavenging zombies. It is also responsible for reboot and shutdown operations.
## A Bit of History
The first widespread init is the simple solution from BSD. Its main competitor is the more featureful but more complex System V init.
The BSD init process simply launches a list of processes. It basically reads a configuration file (/etc/rc) where all programs to launch at startup are statically listed. If the administrator wants to modify this list, she must edit the file manually, which is clearly error-prone and has potentially disastrous effects. This is a very simple and lightweight scheme, but it is completely static.
System V, on the other hand, chose to be more flexible. It introduces the notion of *runlevel*, which can be viewed as multiple and specialized BSD rc files. A runlevel has an identifier and a purpose. Runlevel 1 boots into single mode, runlevel 2 boots into multi-user mode, etc. Only a fixed number of eight runlevels are available. What each runlevel does is specified in the /etc/inittab file.
A runlevel is a *software configuration* of the system which allows only a selected group of applications to exist. When init is requested to change the runlevel, it kills all applications that are not listed in the new runlevel, and starts all unstarted applications in the new runlevel.
The second great improvement in SysVinit is the notion of rc script. Each application contained in a runlevel comes with a wrapper script; it is an executable that accepts start and stop as parameters. This allows to launch and kill each application using the same syntax, regardless of how they are implemented and how they handle parameters internally.
Most often, the combination of rc scripts with runlevels makes use of symbolic links. A runlevel is a directory that contains symlinks named with a numbered prefix. When entering a runlevel, symlinks are accessed alphabetically, in the static order defined by their number. This allows to reuse the same rc script in multiple runlevels.
## Evolution of Init Schemes
Today, most GNU/Linux distributions use a derivative of System V’s init. Even NetBSD went from the BSD init to a SystemV-like init. However, there is room for lots of improvements, and lots of proposals exist on several topics.
The first topic is the order in which rc scripts are started. In 2002, the need(8) scheme proposed to drop statically, manually ordered lists in favour of dependencies. Each rc script lists the services it needs (e.g., the DHCP daemon needs a network connection), and init computes which rc scripts must be started, in what order. All init schemes except BSD, System V and three minor others have a way to express dependencies.
The second topic is the enhancement of configuration management. For instance, Gentoo rc, initNG and runit identify runlevels using a name instead of a number. This breaks the limitation on the number of runlevels, and allows to dedicate a runlevel to a context or a particular usage (e.g., ”AC mode” and ”battery mode” on a laptop). Similarly, upstart, cinit and Sun SMF introduce a notion of profile (or ”mode” for einit).
Another topic is life cycle management. There are two main solutions when monitoring the life cycle of applications. The first one is when using rc scripts: init (or a separate rc application) use the /var directory to store which daemon has been started, with which PID. If no process with such PID is running, then the daemon has been killed. This is a reactive life cycle management, since there is a delay between the death of the daemon and init or rc noticing it. This scheme is used in System V, Gentoo rc, simpleinit, initNG, minit, cinit, monit and the LSB specifications. We must also note that rc scripts may differ for each distribution and init scheme: they handle /var differently. The other way to monitor the life cycle of daemons is not to use rc scripts. A ”universal father” process forks and execs daemon processes directly. The pro is that it is a proactive life cycle management: POSIX signals are used whenever a daemon dies. The con is that the daemon must handle itself portability (the environment variables and the parameters it uses), instead of delegating it to an external script. Sun SMF, launchd, upstart and einit use such a ”universal father”. runit and svscan use a different ”father” process for each daemon.
Init schemes allow a number of operations to instrument the life cycle of applications. Solutions that use father processes allow mostly what POSIX signals handle: start, stop, and a hook to execute when a daemon dies unexpectedly. This hook can be used to log errors or to automatically restart (”respawn”) the faulty daemon. Solutions that use rc scripts may allow a restart operation on a script (Gentoo rc, monit, NetBSD rc.d, LSB specifications), or respawn (initNG, minit, jinit, cinit), in addition to start and stop (System V and its derivatives).
Yet another topic for init enhancement is the tools that ease development and usage. System V uses telinit to pass commands to init. Debian introduced start-stop-daemon, that handles PID files and such. Gentoo uses rc-update to create and modify runlevels. In Sun SMF, svccfg switches profiles. Most init schemes implement their own specific tools.
Finally, the great number of existing init schemes is explained by their different focus. Some focus on startup speed (initNG parallelizes startup scripts). Some focus on implementation size (minit, cinit), in order to be usable on resource-constrained devices. Others focus on one of the topics listed above (dependencies, runlevels, life cycle management and instrumentation, tools). But the most interesting schemes are the fairly new ones backed up by Apple, Sun and Ubuntu.
Apple’s launchd aims to replace all programs that start applications: xinetd, crond, anacron and init. Daemons can be started, respawned, etc, but also scheduled.
Sun’s SMF provides extensive failure detection and diagnostics. It allows to run services in degraded or maintenance mode. The goal is to ease the maintenance of the system (presumably a big server) and to enhance uptime.
Ubuntu’s upstart aims to replace the same tools as launchd, but also to unify devices management and power management. It transforms anything into events and triggers: a new device appearing is an event, the screen going into sleep mode is an event, etc. Actions (e.g., mount a filesystem) can be triggered when specific events are thrown. Daemons and dependencies are handled the same way: the DHCP daemon cannot start until a ”network is ready” event has been thrown.
## Technical Summaries
The remainder of this document are technical summaries of all init-related schemes that the authors know of. Each summary gives a short description on the init scheme, its goals and/or its origin. It then lists features in terms of configuration, life cycle, dependency and service management.
###### Contents
1. [](#S1)
1 BSD init
2. [](#S2)
2 System V init
3. [](#S3)
3 The need(8) Concept
4. [](#S4)
4 Gentoo rc
5. [](#S5)
5 InitNG
6. [](#S6)
6 runit
7. [](#S7)
7 Apple launchd
8. [](#S8)
8 Ubuntu upstart
9. [](#S9)
9 eINIT
10. [](#S10)
10 Pardus
11. [](#S11)
11 LSB init Specifications
12. [](#S12)
12 Daemontools: svscan
13. [](#S13)
13 minit
14. [](#S14)
14 jinit
15. [](#S15)
15 cinit
16. [](#S16)
16 monit
17. [](#S17)
17 depinit
18. [](#S18)
18 NetBSD rc.d
19. [](#S19)
19 Solaris SMF
20. [](#S20)
20 Other Contenders
## 1 BSD init
### 1.1 Home Page
Documentation for recent versions of the BSD-style init:
<a href="http://www.freebsd.org/doc/en_US.ISO8859-1/books/handbook/boot-init.html" target="_blank">http://www.freebsd.org/doc/en_US.ISO8859-1/books/handbook/boot-init.html</a>
The source code for FreeBSD and OpenBSD can be found at :
<a href="http://www.freebsd.org/cgi/cvsweb.cgi/src/sbin/init/" target="_blank">http://www.freebsd.org/cgi/cvsweb.cgi/src/sbin/init/</a>
<a href="http://www.openbsd.org/cgi-bin/cvsweb/src/sbin/init/" target="_blank">http://www.openbsd.org/cgi-bin/cvsweb/src/sbin/init/</a>
### 1.2 About
/sbin/init runs a master boot script (/etc/rc) which orchestrates the whole boot procedure. There may be a few additional rc.\* scripts. init launches them before going in multi-user mode. Because the boot process is sequential and has very few scripts, it is simple and fast.
However, if a new package needs to register itself into the boot process, it needs to edit one of the existing script files. This is very dangerous: if the package installer makes a mistake, the whole startup process will fail.
### 1.3 Configuration Management
BSD init only handles one or a few boot scripts when switching from single-user mode to multi-user mode, and with halt / reboot / shutdown operations. There are no profiles or runlevels.
### 1.4 Life Cycle Management
None.
### 1.5 Dependency Resolution
No dependencies; only an ordered, linear startup defined in /etc/rc.
### 1.6 Service Management Model
None.
### 1.7 Service Management Tools
None. Unix tools such as ps or signal are not part of init.
Communication with the outside world is achieved through signals.
## 2 System V init
### 2.1 Home Page
<a href="http://freshmeat.net/projects/sysvinit/" target="_blank">http://freshmeat.net/projects/sysvinit/</a>
### 2.2 About
From the Freshmeat page:
Init’s primary role is to create processes from a script stored in the /etc/inittab file. This package also contains well known and used utilities like reboot, shutdown, killall, poweroff, telinit, sulogin, wall, etc.
### 2.3 Configuration Management
SysVinit introduced the notion of runlevel. From the man page:
A runlevel is a software configuration of the system which allows only a selected group of processes to exist. The processes spawned by init for each of these runlevels are defined in the /etc/inittab file. Init can be in one of eight runlevels: 0-6 and S. The runlevel is changed by having a privileged user run telinit, which sends appropriate signals to init, telling it which runlevel to change to.
Runlevels 0, 1, and 6 are reserved. Runlevel 0 is used to halt the system, runlevel 6 is used to reboot, and runlevel 1 is used to get the system down into single user mode. Runlevel S is used before entering runlevel 1 (single-user mode). Runlevels 2-5 are multi-user modes, and are different for each Unix / Linux distribution.
When entering a new runlevel (at system boot the runlevel is undefined, sometimes described as “N”), SysVinit executes a master script (/etc/init.d/rc for instance). This script executes all stop-scripts from the old runlevel, then all start-scripts from the new runlevel. Those scripts are found either in /etc/init.d, /etc/rc.d or /sbin/init.d. Each runlevel has a directory: /etc/rc2.d, /etc/rc6.d, etc. Those directories contain links to the real scripts. SysVinit detects whether to stop or to start a script by its name: start-scripts begin with a “S”, stop-scripts with a “K” (as in “kill”). When starting the start-scripts, they are called with “start” as the first and only parameter. When starting stop-scripts, they are called with “stop” as first parameter. This is how a configuration for runlevel 2 could look like:
scice% ls -l /etc/rc2.d
total 0
lrwxr-xr-x 1 root root 18 Feb 20 2004 S10sysklogd -\> ../init.d/sysklogd
lrwxr-xr-x 1 root root 15 Feb 20 2004 S11klogd -\> ../init.d/klogd
lrwxr-xr-x 1 root root 13 Feb 20 2004 S14ppp -\> ../init.d/ppp
lrwxrwxrwx 1 root root 15 Feb 24 2004 S15bind9 -\> ../init.d/bind9
lrwxrwxrwx 1 root root 18 Feb 20 2004 S18quotarpc -\> ../init.d/quotarpc
\[…\]
scice% ls -l /etc/rc6.d
lrwxrwxrwx 1 root root 15 Aug 31 16:22 /etc/rc6.d/K19aumix -\> ../init.d/aumix
lrwxrwxrwx 1 root root 15 Apr 19 09:55 /etc/rc6.d/K19samba -\> ../init.d/samba
lrwxrwxrwx 1 root root 13 Feb 25 2004 /etc/rc6.d/K20gpm -\> ../init.d/gpm
\[…\]
(Source: [nico.schotteli.us](nico.schotteli.us))
The number behind the S or the K is the priority. The lower the priority, the earlier the script is started.
Each service has its own init-script, which sometimes also accepts “restart” or “reload”. This way, runlevels are modular.
### 2.4 Life Cycle Management
The /var/log/wtmp binary file records all user logins and logouts. It is maintained by init, login and getty.
The /var/run/utmp binary file holds information on users and processes currently using the system. Not all applications use it, so it may be incomplete.
Also, service scripts in /etc/init.d can be called during runtime from the command line (not only at startup).
SysVinit can dump its state and exec a later version of itself, so no state is lost between upgrades (source: jinit’s page).
### 2.5 Dependency Resolution
None. Services are started in alphabetical order, hence the priority numbers in the script names. We just hope the administrator made no mistakes…
### 2.6 Service Management Model
Wrapper scripts are called with a start or stop parameter.
### 2.7 Service Management Tools
telinit switches the current runlevel.
## 3 The need(8) Concept
### 3.1 Home Page
<a href="http://www.atnf.csiro.au/people/rgooch/linux/boot-scripts/" target="_blank">http://www.atnf.csiro.au/people/rgooch/linux/boot-scripts/</a>
### 3.2 About
Richard Gooch wrote this whitepaper in 2002. He argues that BSD init is not scalable (modifications are risky), and that SysVinit is too complex and ugly.
Gooch proposes to remove all master scripts and numbering schemes. init runs all scripts in /etc/init.d, in random (?) order. The scripts themselves express their dependencies with the need program. This ensures that a correct startup order is automatically observed. need also ensures that all services are run only once.
The concept of virtual service names is also introduced. For instance, sendmail and qmail both provide the mta service. Which implementation is used is not important, so other services simply need “mta”. This is done with the provide program.
These ideas are implemented in simpleinit.
### 3.3 Configuration Management
A runlevel is represented by a script, e.g. /sbin/init.d/runlevel.3, which will need an unordered set of services, as well as e.g. runlevel.2. This makes it easy to increment and decrement the runlevel.
However, this does not support runlevels that bypass these precedence relations.
### 3.4 Life Cycle Management
Same as SysVinit.
### 3.5 Dependency Resolution
Simplistic dependency expression: need serviceXX.
It is unclear how we ensure that provide, called in service wrapper scripts, is run before dependency resolution.
### 3.6 Service Management Model
Same as SysVinit.
### 3.7 Service Management Tools
None.
## 4 Gentoo rc
### 4.1 Home Page
<a href="http://www.gentoo.org/doc/en/handbook/handbook-x86.xml?part=2&chap=4" target="_blank">http://www.gentoo.org/doc/en/handbook/handbook-x86.xml?part=2&chap=4</a>
### 4.2 About
Gentoo uses SysVinit with modified scripts and tools. Scripts are not handled directly by init, but by Gentoo’s /sbin/rc.
Runlevels have names instead of numbers; this removes the limit on their number, and allows specialized runlevels such as “unplugged” for a laptop. Additional tools allow to create, delete, modify runlevels from the command line.
A global option allows to start scripts in parallel, for a performance gain (more visible on multi-processor or multi-core machines).
Like in SysVinit, after init is invoked as the last step of the kernel boot sequence, it reads /etc/inittab. First, the sysinit entry is run:
si::sysinit:/sbin/rc sysinit
rc is called and mounts all filesystems. Second, the bootwait entry is run:
rc::bootwait:/sbin/rc boot
Here, rc switches the runlevel to boot. All links in the /etc/runlevels/boot directory correspond to services that always must be run at startup: checkroot, checkfs, clock, etc. Lastly, the initdefault entry is run:
id:3:initdefault:
l3:3:wait:/sbin/rc default
Here, initdefault indicates that the default runlevel is 3 (as in the old SysVinit scheme), and runlevel 3 says that it will wait until /sbin/rc has finished switching to runlevel default.
Once rc has finished, init launches virtual consoles as usual, and each console runs a tty with login (through agetty).
c1:12345:respawn:/sbin/agetty 38400 tty1 linux
Note that when an agetty dies, init respawns it.
Gentoo’s rc program is currently being rewritten in C (baselayout package version 2, see <a href="http://roy.marples.name/node/293" target="_blank">http://roy.marples.name/node/293</a>). The goal is to improve performance and not to need bash any more.
### 4.3 Configuration Management
Runlevels are no longer named rc\[0-6S\]. Instead, they have an alphanumerical name. Each runlevel has a corresponding directory in /etc/runlevels, which contains symbolic links to scripts in /etc/init.d.
The rc-update command is used to create, delete and modify runlevels. Each time it is used, a service tree is recomputed. This tree contains the list of services to start and their order, once all dependencies have been computed and resolved.
The rc-status command shows the running status for each service in a runlevel.
The rc command switches to a runlevel passed in argument. Switching to a runlevel means comparing the list of running services with the runlevel’s service tree. If a service from the service tree is not running, it is started. If a service is running, but is neither in the service tree nor a dependency of a service in the service tree, it is stopped. This does not apply when leaving runlevels sysinit and boot.
### 4.4 Life Cycle Management
Each service has an init script particular to the Gentoo distribution. Such a script at least understands start, stop, restart, pause, zap, status, ineed, iuse, needsme, usesme and broken parameters. Life-cycle related parameters are: start and stop, restart (with possibly extra things to do between the stop and the re-start), zap (to force the status to “stopped” in case it is wrongly reported as “stopped”) and status (to query the running status).
Once a script is successfully started, its status is set to “started”. This status will be set to “stopped” if the script is stopped or zapped, and will be re-evaluated if the script is status’ed. Re-evaluation is done by checking the pid file in /var/run and the daemon file in /var/lib/init.d. This means that if a service’s process is manually killed, the service’s status stays “running” until a stop, zap or status command is issued.
### 4.5 Dependency Resolution
Gentoo init scripts respond to the following parameters: ineed, iuse, needsme, usesme and broken. The first two list the hard and soft dependencies of the scripts, the next two list the scripts that are hard- and soft-dependent on this script, and the last one lists broken dependencies (unresolved, “need”-type).
Inside the script, the following metadata tags can be used: need, use, before, after, provide. need are hard dependencies: the current script cannot run without them. use are soft dependencies: the current script uses them if they are present in the current runlevel or in boot, but can still run without them. before and after specify a startup order, regardless of dependencies. Finally, provide indicates virtual service names that the current scripts fulfills (like in Gooch’s need-based init). All these tags are used when a script is manually invoked, and when a runlevel is modified.
Network dependencies are a special case. A service will need a network interface to be up (any interface other than net.lo). This dependency is written: need net.
### 4.6 Service Management Model
All services must have a Gentoo-specific wrapper script (an additional effort for the packager). Configurations and services can be managed through good CLI tools.
For each service, settings (preferencess, configuration) are separated from the wrapper script. Scripts are in /etc/init.d/\<script\>, while settings are in /etc/conf.d/\<script\>.
The main problem is that rc-status does not show the exact state of services (cf. killed processes not seamlessly taken into account).
### 4.7 Service Management Tools
rc-status, rc-config, rc-update: see above.
/sbin/rc: switch runlevel (sys-apps/baselayout package).
depscan.sh: rebuild the dependency tree.
## 5 InitNG
### 5.1 Home Page
<a href="http://www.initng.org/wiki" target="_blank">http://www.initng.org/wiki</a>
### 5.2 About
InitNG’s goal is to start services in parallel, for a reduced startup time.
As a side note, the tool’s syntax is reversed compared to System V: instead of calling \<script\> start, InitNG is invoked via ngc start \<script\>. It makes more sense to call a single tool to manage life cycles than to call multiple scripts that duplicate similar behaviour.
The configuration files support lots of fancy features, such as respawn mode, suid, niceness, output redirect, arguments passing, environment variables passing, etc.
There is not much information on the wiki, except that it is supposed to be faster than SysVinit.
### 5.3 Configuration Management
ng-update manages configurations similarly to Gentoo’s rc-update.
The wiki does not provide more information.
### 5.4 Life Cycle Management
Respawn mode for daemons. Otherwise, similar to SysVinit.
### 5.5 Dependency Resolution
need and use dependencies, similar to Gentoo init.
Network access is a special dependency, as in Gentoo init.
### 5.6 Service Management Model
Every command goes through an InitNG program, instead of directly invoking rc shell scripts.
### 5.7 Service Management Tools
The ngc tool is used to start / stop / restart services.
The ng-update tool manages runlevels.
## 6 runit
### 6.1 Home Page
<a href="http://smarden.org/runit/" target="_blank">http://smarden.org/runit/</a>
### 6.2 About
From the home page: runit is a cross-platform Unix init scheme with service supervision. Running and shutting down is done in three stages:
1. 1\.
Stage 1:
runit starts /etc/runit/1 and waits for it to terminate. The system’s one time initialization tasks are done here. /etc/runit/1 has full control over /dev/console to be able to start an emergency shell in case the one time initialization tasks fail.
2. 2\.
Stage 2:
runit starts /etc/runit/2 which should not return until the system is going to halt or reboot; if it crashes, it will be restarted. Normally, /etc/runit/2 runs runsvdir.
3. 3\.
Stage 3:
If runit is told to halt or reboot the system, or Stage 2 returns without errors, it terminates Stage 2 if it is running, and runs /etc/runit/3. The systems tasks to shutdown and halt or reboot are done here.
runit is “optimized for reliability and small size”. If compiled and statically linked with dietlibc, it produces an $`\approx`$ 8.5k executable.
### 6.3 Configuration Management
Runlevels are handled through the runsvdir and runsvchdir programs.
A runlevel is a directory in /etc/runit/runsvdir/. Symlinks indicate which are the current and previous runlevel. A runlevel is launched by runsvdir. Runlevel switching is performed by runsvchdir, which switches directories for runsvdir.
### 6.4 Life Cycle Management
Each of a runlevel’s subdirectories is a service. For each service, runsvdir launches a runsv, which starts and monitors this specific service. A service contains a ./run and an optional ./finish script. Its state is maintained in a file by runsv. A named pipe allows to pass commands to runsv, e.g. to start, stop, run once or in respawn mode, pause (SIGSTOP), etc.
The sv program sends commands to one or all runsvs.
### 6.5 Dependency Resolution
Expression of dependencies is poor. If service A depends on service B, A’s run script must explicitly invoke sv to ensure that B is running, e.g.:
\#!/bin/sh
sv start B \|\| exit 1
exec A
There are no optional dependencies, no virtual service names, and no tools to explore dependencies.
### 6.6 Service Management Model
One management process is launched for each service. A single runlevel process is launched to manage all these management processes.
Commands are passed through named pipes.
### 6.7 Service Management Tools
sv: pass life cycle-related commands from the CLI.
runsvchdir: switch runlevel.
chpst (change process state): set limits for a service or for a user (depends on the limit, not really explicit). e.g. \# open files per process, \# processes per user, memory per process.
## 7 Apple launchd
### 7.1 Home Page
<a href="http://developer.apple.com/macosx/launchd.html" target="_blank">http://developer.apple.com/macosx/launchd.html</a>
### 7.2 About
Starting from Mac OS X 10.4, launchd aims to replace init, the old mach_init, cron and xinetd. It starts daemons, shuts them down and restarts them when a request comes (inetd behaviour), runs programs at a specified time or later (anacron-like behaviour).
Each job (service) has an XML property file, replacing the different syntaxes from crontab, inittab etc. This includes program arguments, I/O redirect, setuid, chroot, etc.
Dependencies are automatically discovered (?). Services are run in parallel.
launchd can be controlled from a single utility, launchctl.
The source code has been relicensed under the ASL in august 2006.
### 7.3 Configuration Management
Runlevels are not supported on Mac OS (no order maintained, no selective startup of services). Directories of jobs can be used to launch a set of jobs together.
### 7.4 Life Cycle Management
Jobs can be started once, started in respawn mode, started in launch-on-demand mode, started based on a schedule, stopped. How a job is started is decided by its property file.
As usual, respawn mode requires that the service does not daemonize itself (fork+exec, fork+exit).
### 7.5 Dependency Resolution
Dependencies are a bit unusual and lack documentation. The Uses, Requires and Provides keywords have been dropped. Instead, services watch for changes on a file or watch for IPCs. The advantage is that we ensure dependencies are not only launched, but are also in a usable state, already initialized and ready to communicate. The drawback is that it is totally intrusive and quite obscure.
Special dependencies can be specified, also in the source code (not as metadata), but using specific APIs: disk volume mounted, network available, kernel module present (else it is launched), user Y is logged in.
### 7.6 Service Management Model
Everything life cycle-related goes through launchd, instead of using various tools (cron, inet). Every command is issued using launchctl, which also has a prompt mode.
Instances of launchd can be run for a specific user. This allows non-root users to schedule jobs.
No configuration management is provided, nor virtual service names.
### 7.7 Service Management Tools
launchctl: takes commands and passes them to launchd through a mutex.
## 8 Ubuntu upstart
### 8.1 Home Page
<a href="http://upstart.ubuntu.com/" target="_blank">http://upstart.ubuntu.com/</a>
### 8.2 About
upstart aims to replace init, cron, anacron, at and inet. It also aims to interact with udev, acpi and module-init-tools (modprobe, etc). The basic idea is that every action is triggered by an event.
Tasks (or jobs, or services) begin in the “waiting” state. Upon reception of a task-dependent set of events, the task is run and goes in the “running” state. Upon reception of another task-dependent set of events, it goes to the “dead” state (which should be “stopping”), cleans up, then returns to the “waiting” state.
An event can be triggered manually, by ACPI, by udev, by upstart, etc. Dependencies are also expressed as events: upstart sends an event when a task goes “running”; its dependencies simply wait for this event. Everything task-related is configured in /etc/event.d.
Security around events and event senders is unclear.
The project has been started before launchd went open source, and shares some similarities (mainly inet and cron capabilities).
Both code and documentation are quickly progressing, so some limitations presented here may already have been fixed.
### 8.3 Configuration Management
A notion of profile is intended, as a replacement for runlevels. It is still unspecified and unimplemented. Runlevels are still present as dependencies: “start on runlevel-2”.
upstart launches all tasks present in /etc/event.d; the order is defined by dependencies (expressed as events).
### 8.4 Life Cycle Management
Tasks can be waiting, running or dying. A set of events triggers state switching. Actions can be performed at each state change (start script, stop script, anything).
### 8.5 Dependency Resolution
Dependencies are expressed as e.g. “start on event1”. Event1 can be that a service has begun to start (“on starting s1”), has finished to stop (“on stopped s2”), etc. This unusual approach gives a lot of flexibility: it allows virtual service names, optional dependencies, and a lot more.
### 8.6 Service Management Model
The event system is very powerful and flexible. However, it might be a bit complex to get the whole picture of what happens on the system.
### 8.7 Service Management Tools
start, stop and status commands are provided to manage tasks.
initctl is used to throw events from the command line (using a mutex).
## 9 eINIT
### 9.1 Home Page
<a href="http://einit.sourceforge.net/" target="_blank">http://einit.sourceforge.net/</a>
### 9.2 About
eINIT aims to be a dependency-based init, written entirely in C, suitable for embedded systems. It wants to get rid of rc scripts, to have flexible dependencies, to be modular (e.g. with audio output for the blind), and to support respawn mode.
### 9.3 Configuration Management
eINIT supports runlevels called modes. A mode is described by an XML file. It can depend on other modes (like Gentoo’s boot runlevel), i.e. it comes in addition to the mode it depends on. A mode says which services must be enabled (started), disabled (stopped), and reset (restarted). A mode also defines orders of precedence: “try to start service A, and start service B if it fails”. Finally, there is the notion of service groups, e.g. all network interfaces, or all graphical devices. Dependencies can be set on one, all or part of the service group.
Variables can be set globally and per-mode. (It is unclear what the use case is)
### 9.4 Life Cycle Management
Services are either started once, started in respawn mode, or stopped.
Dependencies are expressed inside runlevels (cf. above), not in service wrapper scripts. Services are described in the mode’s XML file, with attributes such as “enable” and “disable” that indicate what to run when the service is started and stopped.
### 9.5 Dependency Resolution
Other attributes are “provides” and “requires” for dependencies. This allows for virtual service names. This also means that the XML description must be duplicated between modes, unless we create a mode dependency.
### 9.6 Service Management Model
Most management activities can be seen as switching runlevels.
eINIT supports enabling/disabling eINIT modules (e.g. TTYs, shell commands, daemons), but not services themselves. This is a much coarser granularity.
### 9.7 Service Management Tools
einit-control passes Commands through a Unix socket.
## 10 Pardus
### 10.1 Home Page
<a href="http://www.pardus.org.tr/eng/projeler/comar/SpeedingUpLinuxWithPardus.html" target="_blank">http://www.pardus.org.tr/eng/projeler/comar/SpeedingUpLinuxWithPardus.html</a>
### 10.2 About
Pardus is a GNU/Linux distribution from Turkey. Two subprojects are within our scope: Mudur and Çomar.
Mudur is the init subsystem, which replaces the rc part of init. It is called by init via the /etc/inittab file. Mudur performs runlevel switching tasks, according to the usual System V runlevels (boot, single, reboot, etc). Mudur is written in Python.
Çomar is the ”configuration management system” which manages settings and preferencesi, and handles the way to start applications. It needs modified rc scripts, as do most of its competitors. According to the wiki, it ”handles parallel execution, access control, profile management and remote management”. Çomar is written in C, and rc scripts are written in Python.
The main objective is to gain speed at startup, while being clean and hiding complexity to the user.
### 10.3 Configuration Management
Runlevels have the same meaning as in System V. For the sysinit runlevel, Mudur mounts pseudo-filesystems, starts udev, loads kernel modules, etc. For other runlevels, Mudur calls Çomar, which in turn handles rc scripts.
### 10.4 Life Cycle Management
rc scripts accept parameters such as start, stop, status, configure. The start and stop directives make use of Debian’s start-stop-daemon.
### 10.5 Dependency Resolution
An rc script calls the start methods of the other scripts it depends on.
There are no virtual service names.
### 10.6 Service Management Model
Command-line tools allow to call methods on rc scripts via Çomar.
### 10.7 Service Management Tools
Not presented.
## 11 LSB init Specifications
### 11.1 Home Page
<a href="http://www.freestandards.org/spec/refspecs/LSB_1.3.0/gLSB/gLSB/sysinit.html" target="_blank">http://www.freestandards.org/spec/refspecs/LSB_1.3.0/gLSB/gLSB/sysinit.html</a>
### 11.2 About
LSB, the Linux Standards Base, defines what init scripts should conform to.
This is an opinion, but the LSB seems more keen on standardizing what RedHat does than on being innovative.
### 11.3 Configuration Management
LSB specifies the following runlevels:
0: halt
1: single user mode
2: multiuser with no network services exported
3: normal/full multiuser
4: reserved for local use, default is normal/full multiuser
5: multiuser with xdm or equivalent
6: reboot
This forbids “profiles” and useful runlevel switching as seen in previous init schemes (e.g. Gentoo).
### 11.4 Life Cycle Management
Init files provided by LSB applications shall accept one argument, which states the action to perform:
start: start the service
stop: stop the service
restart: stop and restart the service if the service is already running, otherwise start the service
reload: cause the configuration of the service to be reloaded without actually stopping and restarting the service \[optional\]
force-reload: cause the configuration to be reloaded if the service supports this, otherwise restart the service
status: print the current status of the service
There is no respawn mode.
### 11.5 Dependency Resolution
LSB-compliant applications init scripts (but not system init scripts) must adopt this syntax for dependency declaration:
\# Provides: boot_facility_1 \[ boot_facility_2 …\]
\# Required-Start: boot_facility_1 \[boot_facility_2 …\]
\# Required-Stop: boot_facility_1 \[boot_facility_2 …\]
\# Default-Start: run_level_1 \[ run_level_2 …\]
\# Default-Stop: run_level_1 \[ run_level_2 …\]
\# Short-Description: short_description
\# Description: multiline_description
Required-stop are applications that must still run before this one stops.
The “Default-Start” and “Default-Stop” headers define which runlevels should by default run the script with a start or stop argument, respectively, to start or stop the services controlled by the init script.
“Provides” allows to have virtual service names.
### 11.6 Service Management Model
Out of scope.
### 11.7 Service Management Tools
Out of scope.
## 12 Daemontools: svscan
### 12.1 Home Page
<a href="http://code.dogmap.org/svscan-1/" target="_blank">http://code.dogmap.org/svscan-1/</a>
<a href="http://cr.yp.to/daemontools.html" target="_blank">http://cr.yp.to/daemontools.html</a>
### 12.2 About
From the daemontools page: daemontools is a collection of tools for managing UNIX services. The svscan program can be used as an init replacement.
### 12.3 Configuration Management
Configurations / runlevels are supported as directories of run scripts.
svscan starts and monitors a collection of services. svscan starts one supervise process for each subdirectory of the current directory. Every five seconds, svscan checks for subdirectories again. If it sees a new subdirectory, it starts a new supervise process.
Runlevel switching is not really specified.
### 12.4 Life Cycle Management
supervise monitors a service. It starts the service and restarts it if it dies. To set up a new service, all supervise needs is a directory with a run script that runs the service.
supervise maintains status information in a binary format inside the directory
\<service\>/supervise
, which must be writable by supervise. The status information can be read using svstat.
There is no stop script.
### 12.5 Dependency Resolution
No dependencies.
### 12.6 Service Management Model
Put run scripts in a directory; daemontools will run the service and monitor it.
It is simpler than SysVinit, but lacks features (stop command, runlevel switching).
### 12.7 Service Management Tools
svc is used to send signals to a service.
svstat prints the status of services monitored by supervise.
multilog saves error messages to one or more logs. It optionally timestamps each line and, for each log, includes or excludes lines matching specified patterns.
## 13 minit
### 13.1 Home Page
<a href="http://www.fefe.de/minit/" target="_blank">http://www.fefe.de/minit/</a>
### 13.2 About
Minit aims to be a very small init. It can start a service, restart it, start it when another is finished, and take dependencies into account.
### 13.3 Configuration Management
None.
### 13.4 Life Cycle Management
Reads /var/run/\<service\>.pid and checks if the process has died.
### 13.5 Dependency Resolution
Yes, but lacks documentation.
### 13.6 Service Management Model
None.
### 13.7 Service Management Tools
msvc: similar to svc from daemontools.
pidfilehack: reads /var/run/\<service\>.pid.
## 14 jinit
### 14.1 Home Page
<a href="http://john.fremlin.de/programs/linux/jinit/" target="_blank">http://john.fremlin.de/programs/linux/jinit/</a>
### 14.2 About
jinit is based on an extended version of Richard Gooch’s need(8) scheme for init. Currently there is no provide(8) functionality.
### 14.3 Configuration Management
None? (or at least documentation is lacking).
### 14.4 Life Cycle Management
Supports respawn mode. Apart from that, same as SysVinit.
If it receives a fatal error signal like SIGSEGV it will fork a child to dump core, and exec itself (with the “panic” option so no bootscripts are run). That means it should never bring the system down unnecessarily.
### 14.5 Dependency Resolution
Same as the need(8) scheme.
### 14.6 Service Management Model
Same as SysVinit.
### 14.7 Service Management Tools
None. Communicates over SysV message queues.
## 15 cinit
### 15.1 Home Page
<a href="http://linux.schottelius.org/cinit/" target="_blank">http://linux.schottelius.org/cinit/</a>
### 15.2 About
cinit claims to be a fast, small and simple init with support for profiles. It supports parallel startup, mandatory and optional dependencies, and does seem simple to use.
Size is $`\approx`$ 50KiB when statically linked against uclibc.
### 15.3 Configuration Management
“Profiles” are directories of services. They are put in /etc/cinit.
Choosing the profile at startup is supported, but profile switching is not documented.
### 15.4 Life Cycle Management
A service is a subfolder in a configuration (or a link to another folder). Everything is done through files or symlinks: ./on and ./off indicate the program to run when the service goes on/off. ./on.params and ./on.env contain parameters and environment variables to pass to the program.
### 15.5 Dependency Resolution
Dependencies are listed in ./wants (optional) and ./needs (mandatory).
Virtual service names are not supported. The author suggests using symlinks, e.g. cron -\> dcron.
### 15.6 Service Management Model
None.
### 15.7 Service Management Tools
A bunch of executables:
cinit.add.dependency - add a dependency to a service
cinit.add.getty - add a new getty
cinit.create.empty.service - create an empty service
cinit.reboot - reboot in /bin/sh
cinit.remove.getty - remove a getty service
cinit.respawn.off - switch respawing off
cinit.respawn.on - switch respawing on
cinit.shutdown - shutdown in /bin/sh
## 16 monit
### 16.1 Home Page
<a href="http://www.tildeslash.com/monit/index.php" target="_blank">http://www.tildeslash.com/monit/index.php</a>
### 16.2 About
monit manages and monitors processes, files, directories and devices. It does not replace init: instead, it acts as a service manager, and can reuse rc scripts.
monit can start a process which is not running, restart a process if it does not respond, and stop a process if it uses too much resources. It can monitor files, directories and devices for changes on timestamp, checksum or size. All checks are performed at a polling interval time.
Management activities are exported through HTTP(S). Monit is configured via the /etc/monitrc file.
### 16.3 Configuration Management
Only supports groups of services which can be started / stopped together.
### 16.4 Life Cycle Management
A service can be (re)started and stopped. Its state is checked using pid files. This means that wrapper scripts may have to be written.
### 16.5 Dependency Resolution
The syntax is simplistic: depends on serviceX (in file /etc/monitrc).
There are no virtual service names.
### 16.6 Service Management Model
Periodically check a resource (process, file…). Upon changes, actions are triggered. For example:
check process apache with pidfile /var/run/httpd.pid
start program = “/etc/init.d/httpd start”
stop program = “/etc/init.d/httpd stop”
if failed host www.tildeslash.com port 80 then restart
depends on apache_bin
### 16.7 Service Management Tools
HTTP interface.
## 17 depinit
### 17.1 Home Page
<a href="http://www.nezumi.plus.com/depinit/" target="_blank">http://www.nezumi.plus.com/depinit/</a>
### 17.2 About
From the home page: ”Depinit is an alternative init program that can handle parallel execution, dependencies, true roll-back, pipelines, improved signaling and unmounting filesystems on shutdown. It incorporates ideas from sysvinit, simpleinit, daemontools and make. At present, it is a bit experimental, and requires good knowledge of the initialisation process to set up.”
While depinit boasts more features than sysvinit, it is quite complex to setup and to use. It also has trouble working with the login program.
### 17.3 Configuration Management
A runlevel is a directory in /etc/depinit. How it operates and how it can be switched is unclear.
### 17.4 Life Cycle Management
Each service has its own directory in /etc/depinit, which will contain shell scripts named start, stop, source, filter, dest and/or daemon. Their purpose is either obvious (start and stop describe shell commands that will be called to start and stop the service), or unexplained.
### 17.5 Dependency Resolution
Dependencies are listed in a depend file within each service’s directory. Each line in the file corresponds to another directory in /etc/depinit.
### 17.6 Service Management Model
All interactions with depinit are performed through the depctl command-line tool. Daemons are started in respawn mode.
### 17.7 Service Management Tools
depinit: can either replace init (and rc), or be started with a PID \> 1 and manage a subset of services.
depctl: start/stop a service, shutdown/reboot the system, send a signal to a daemon.
These tools communicate via signals.
## 18 NetBSD rc.d
### 18.1 Home Page
<a href="http://ezine.daemonnews.org/200108/rcdsystem.html" target="_blank">http://ezine.daemonnews.org/200108/rcdsystem.html</a>
<a href="http://www.netbsd.org/guide/en/chap-rc.html" target="_blank">http://www.netbsd.org/guide/en/chap-rc.html</a>
<a href="http://cvsweb.netbsd.org/bsdweb.cgi/src/sbin/init/" target="_blank">http://cvsweb.netbsd.org/bsdweb.cgi/src/sbin/init/</a>
### 18.2 About
From the home page:
As of NetBSD 1.5, the startup of the system changed to using rc-scripts for controlling services, similar to SysVinit, but without runlevels.
### 18.3 Configuration Management
None.
### 18.4 Life Cycle Management
Service wrapper scripts understand start, stop, restart and kill.
They are as limited as SysVinit’s. There is no service monitoring facility.
### 18.5 Dependency Resolution
Dependencies are expressed in rc scripts as four keywords. REQUIRE is a mandatory dependency; PROVIDE allows virtual service names; BEFORE’s explanation fell through a spatiotemporal void; KEYWORD is a tag used to include or exclude the script from ordering, e.g. with values ’nostart’ or ’shutdown’.
Ordering is done via the rcorder command.
### 18.6 Service Management Model
None. Start scripts, hope they run, optionally check a pid file.
### 18.7 Service Management Tools
None.
## 19 Solaris SMF
### 19.1 Home Page
<a href="http://www.sun.com/bigadmin/content/selfheal/smf-quickstart.html" target="_blank">http://www.sun.com/bigadmin/content/selfheal/smf-quickstart.html</a>
### 19.2 About
SMF (Service Management Facility) is a part of SUN Solaris 10 (historically a System V variant). It targets big production servers. From their page:
SMF has added the notion of the relationship between a service, its processes, and another service that is responsible for restarting the service, to the Solaris kernel. SMF understands whether a service process failed as the result of an administrator error, failure of a dependent service, software bug, or underlying hardware failure. SMF informs the appropriate restarter, which decides either to disable the service by placing it in maintenance mode because it appears to be defective, or to restart it automatically.
SMF still supports inittab, inetd and rc scripts, but aims at replacing them.
SMF services are started in parallel, in a dependency-based order.
Users can be given the rights to manage services and their state, without being root. Still, there are no per-user services.
Alas, the source code is closed.
### 19.3 Configuration Management
Runlevels are replaced by milestones. A specific set of services must be enabled (started) before a milestone is reached (incremental runlevels). Milestones are: single-user, multi-user, multi-user-server, all, none.
The notion of profile also exist: it is an XML file which lists SMF services, and whether each should be enabled or disabled.
### 19.4 Life Cycle Management
SMF-enabled services can be permanently enabled / disabled (persists on reboot). When enabled, they can be online (started), offline (stopped), uninitialized (config has not been read yet), degraded (runs at limited capacity), and maintenance (stopped, needs an action from the administrator).
Possible commands are: permanent or temporary enable / disable, start, stop, restart, refresh. Refresh means reloading a service’s config file, which is usually triggered by SIGHUP on daemons (SIGHUP should kill non-daemon processes).
### 19.5 Dependency Resolution
A dependency is formed as an URI. It can be set on services, devices, network configuration, kernel configuration, milestone (runlevel), and be optional or mandatory. Shortened example from SUN’s site:
% svcs -l network/smtp:sendmail
dependency require_all/refresh file://localhost/etc/mail/sendmail.cf (-)
dependency optional_all/none svc:/system/system-log (online)
dependency require_all/refresh svc:/system/identity:domain (online)
dependency require_all/refresh svc:/milestone/name-services (online)
dependency require_all/none svc:/network/service (online)
dependency require_all/none svc:/system/filesystem/local (online)
### 19.6 Service Management Model
SMF manages services, rc scripts, inet daemons, configurations in an (almost) uniform way. The goal is to obtain the best uptime possible, with as little human intervention as possible. The other big goal is to provide detailed information on failures to the administrator, in the case he must intervene.
There are lots and lots of features, with probably lots of complicated code: even the kernel is modified. This solution is ideal for servers, but not suitable for small systems.
### 19.7 Service Management Tools
Amongst others:
scvs queries the system state, including old rc scripts.
svcadm passes commands to SMF services and their restarters.
svccfg switches profiles.
## 20 Other Contenders
The following projects are considered less relevant. Most are dead and/or lack documentation.
### 20.1 DMD
Daemons-Managing Daemon is a stalled project from the Free Software Foundation. Documentation is non-existent.
<a href="http://directory.fsf.org/GNU/DMD.html" target="_blank">http://directory.fsf.org/GNU/DMD.html</a>
### 20.2 twsinit
twsinit is a minimal init replacement (about 8 KB) in x86 assembler. Documentation is non-existent.
<a href="http://www.energymech.net/users/proton/" target="_blank">http://www.energymech.net/users/proton/</a>
### 20.3 Busybox
BusyBox combines tiny versions of many common UNIX utilities (GNU fileutils, shellutils, textutils) into a single small executable. It contains a small init system, which is a stripped down SysVinit scheme. It can read an inittab file, but does not support runlevels.
The project is alive.
<a href="http://www.busybox.net/" target="_blank">http://www.busybox.net/</a>
### 20.4 Fedora New Init System
Fedora had planned to implement its own init replacement. Its goals were a bit confused and mixed: better startup speed, LSB compliance, respawn, logging, tools. The project is stalled / dead.
<a href="http://fedoraproject.org/wiki/FCNewInit" target="_blank">http://fedoraproject.org/wiki/FCNewInit</a>
### 20.5 serel
Serel aims to speed up the boot sequence using parallel startup. The project is stalled.
<a href="http://www.fastboot.org/" target="_blank">http://www.fastboot.org/</a>
<table>
<thead>
<tr>
<th></th>
<th>Focus</th>
<th>Configuration Management</th>
<th>Life Cycle Instrumentation</th>
<th>Life Cycle Monitoring</th>
<th>Expression of Dependencies</th>
<th>License</th>
</tr>
</thead>
<tbody>
<tr>
<td>BSD init</td>
<td>-</td>
<td>No</td>
<td>No</td>
<td>No</td>
<td>Static ordering</td>
<td>BSD</td>
</tr>
<tr>
<td>SysVinit</td>
<td>-</td>
<td>Numbered runlevels</td>
<td>start, stop</td>
<td>Rc scripts (/var)</td>
<td>Static ordering</td>
<td>GPL</td>
</tr>
<tr>
<td>simpleinit</td>
<td>Have dependencies</td>
<td>No</td>
<td>start, stop</td>
<td>Rc scripts (/var)</td>
<td>need</td>
<td>GPL</td>
</tr>
<tr>
<td>Gentoo rc</td>
<td>Better runlevels and dependencies</td>
<td>Named runlevels</td>
<td>start, stop, restart</td>
<td>Rc scripts (/var)</td>
<td>need, use, before, after, provide</td>
<td>GPL</td>
</tr>
<tr>
<td>InitNG</td>
<td>Speed</td>
<td>Named runlevels</td>
<td>start, stop, respawn</td>
<td>Process (/var)</td>
<td>need, use</td>
<td>GPL</td>
</tr>
<tr>
<td>runit</td>
<td>Life Cycle Management</td>
<td>Named runlevels</td>
<td>POSIX signals</td>
<td>Father processes</td>
<td>manual scripting</td>
<td>BSD</td>
</tr>
<tr>
<td>launchd</td>
<td>Lots of features</td>
<td>No (Groups of jobs)</td>
<td>start, stop, inet, cron, respawn</td>
<td>Father process</td>
<td>In code: watch files & IPCs</td>
<td>ASL</td>
</tr>
<tr>
<td>upstart</td>
<td>Event-based</td>
<td>Profiles</td>
<td>Events</td>
<td>Father process</td>
<td>stopping, stopped, starting, started, any event</td>
<td>GPL</td>
</tr>
<tr>
<td>eINIT</td>
<td>Runlevels</td>
<td>Modes (services, dependencies, vars)</td>
<td>start, stop, respawn</td>
<td>Father process</td>
<td>provides, requires</td>
<td>BSD</td>
</tr>
<tr>
<td>Pardus</td>
<td>Speed</td>
<td>Numbered runlevels</td>
<td>start, stop</td>
<td>Rc scripts (/var)</td>
<td>need</td>
<td>GPL</td>
</tr>
<tr>
<td>LSB</td>
<td>Standardize dependencies</td>
<td>Numbered runlevels</td>
<td>start, stop, restart</td>
<td>Rc scripts</td>
<td>provides, required-to-start, required-to-stop</td>
<td>-</td>
</tr>
<tr>
<td>svscan</td>
<td>Service Management</td>
<td>Directories of run scripts</td>
<td>start, respawn</td>
<td>Father processes</td>
<td>No</td>
<td>freedist</td>
</tr>
<tr>
<td>minit</td>
<td>Size</td>
<td>No</td>
<td> respawn</td>
<td>/var</td>
<td>?</td>
<td>GPL</td>
</tr>
<tr>
<td>jinit</td>
<td>need(8) deps</td>
<td>No</td>
<td> respawn</td>
<td>/var ?</td>
<td>need</td>
<td>GPL</td>
</tr>
<tr>
<td>cinit</td>
<td>Size, runlevels</td>
<td>Unswitchable profiles</td>
<td>respawn</td>
<td>Scripts (/var)</td>
<td>wants, needs</td>
<td>GPL</td>
</tr>
<tr>
<td>monit</td>
<td>Service Management</td>
<td>No (Groups of jobs)</td>
<td>start, stop, restart, quotas</td>
<td>Scripts (/var)</td>
<td>depends</td>
<td>GPL</td>
</tr>
<tr>
<td>NetBSD rc.d</td>
<td>Have rc scripts</td>
<td>No</td>
<td>start, stop, restart</td>
<td>Rc scripts</td>
<td>require, provide, before</td>
<td>BSD</td>
</tr>
<tr>
<td>SMF</td>
<td>Heavy Management</td>
<td>Milestones and profiles</td>
<td>start, stop, restart + persist + self-healing</td>
<td>Father Process / kernel</td>
<td>require, optional on URIs (svc, file…)</td>
<td>Closed source</td>
</tr>
</tbody>
</table> |
0708.1413 | cs.NI | cyber | # On Optimal Turbo Decoding of Wideband MIMO-OFDM Systems Under Imperfect Channel State Information
Sajad Sadough<sup>†∗</sup> and Pierre Duhamel<sup>∗</sup>
<sup>†</sup>Ecole Nationale Supérieure de Techniques Avancées, 75015 Paris, France
<sup>∗</sup> Laboratoire des Signaux et Systèmes, CNRS/Supélec, F-91192 Gif-sur-Yvette, France
Email:{sadough, pierre.duhamel}@lss.supelec.fr
###### Abstract
We consider the decoding of bit interleaved coded modulation (BICM) applied to both multiband and MIMO OFDM systems for typical scenarios where only a noisy (possibly very bad) estimate of the channel is provided by sending a limited number of pilot symbols. First, by using a Bayesian framework involving the channel a posteriori density, we adopt a practical decoding metric that is robust to the presence of channel estimation errors. Then this metric is used in the demapping part of BICM multiband and MIMO OFDM receivers. We also compare our results with the performance of a mismatched decoder that replaces the channel by its estimate in the decoding metric. Numerical results over both realistic UWB and theoretical Rayleigh fading channels show that the proposed method provides significant gain in terms of bit error rate compared to the classical mismatched detector, without introducing any additional complexity.
## I Introduction
Ultra-Wide-Band (UWB) is defined as any wireless transmission scheme that occupies a bandwidth of more than 25 % of its center frequency or greater than 500 MHz over the 3.1-10.6 GHz frequency band \[[1](#bib.bib1)\]. Multiband Orthogonal Frequency division multiplexing (MB-OFDM) \[[2](#bib.bib2)\] is a spectrally efficient technique proposed for high data rate, short range UWB applications. This approach uses a conventional OFDM system, combined with bit interleaved coded modulation (BICM) and frequency hopping for improved diversity and multiple access.
It is well known that multiple-input multiple-output (MIMO) antennas systems is a promising technique for high-speed, spectrally efficient and reliable wireless communications. However, as higher data rates lead to wideband communications, the underlying MIMO channels exhibit strong frequency selectivity. By using orthogonal frequency division multiplexing (OFDM) and applying a proper cyclic prefix (CP), the frequency selective channels are transformed to an equivalent set of frequency-flat subchannels. These considerations motivate the combination of MIMO and OFDM, referred to as MIMO-OFDM, as a promising technology for the future generation of wideband wireless systems \[[3](#bib.bib3)\].
In both MB-OFDM and MIMO-OFDM systems, a typical scenario occurs when the channel is changing so slowly that it is considered time invariant during the transmission of an entire frame. In such situations, channel estimation is performed by sending training symbols (pilots) transmitted at the beginning of the information frame while the rest of the frame is decoded based on the estimated channel. Due to the limited number of pilots, the estimate of the channel is imperfect and the receiver has only access to this noisy channel estimate. However, the receiver/decoder metric for any coherent detector, requires knowledge of the exact channel.
A standard sub-optimal technique, known as mismatched ML decoding, consists in replacing the exact channel by its estimate in the receiver metric. Hence, the resulting decoding metric is not adapted to the presence of channel estimation errors (CEE). Although, this scheme is not optimal under imperfect channel estimation, it has been extensively adopted for performance evaluation of single an multi-carrier MIMO systems \[[4](#bib.bib4)\]. Basically, the transmitter and the receiver strive to construct codes for ensuring reliable communication with a quality of service (QoS), no matter what degree of channel estimation accuracy arises during the transmission. The QoS requirements stand for achieving target rates with small error probability even with very bad channel estimates. Thus the presence of CEE, arises the following important question: what type of practical encoder/decoder can achieve the best performance under imperfect channel estimation ?
As an alternative to the aforementioned mismatch scenario, in \[[5](#bib.bib5)\] the authors proposed a different decoding metric in the case of space-time decoding of MIMO channels. In this paper we see that the metric of \[[5](#bib.bib5)\] can be derived as the average of the likelihood that would be used if the channel is perfectly known, over all realizations of the channel uncertainty which mitigates the impact of CEE on the decoding performance and provides a robust design. The averaging of this metric is performed in the Bayesian framework provided a posteriori pdf of the perfect channel conditioned on its estimate that characterizes the channel estimation process and matches well the channel knowledge available at the receiver. Based on that metric, we formulate our decoding rule for BICM MB-OFDM and MIMO-OFDM.
The outline of this paper is as follows. In Section [II](#S2 "II TRANSMISSION MODEL ‣ On Optimal Turbo Decoding of Wideband MIMO-OFDM Systems Under Imperfect Channel State Information") we describe the system model for MB-OFDM and MIMO-OFDM transmission over a frequency selective fading channel. Section [III](#S3 "III Pilot Based Channel Estimation ‣ On Optimal Turbo Decoding of Wideband MIMO-OFDM Systems Under Imperfect Channel State Information") presents the pilot assisted channel estimation: we specify the statistics of the CEE and then calculate the posterior distribution of the perfect channel conditioned on the estimated channel. This posterior distribution is used in section [IV](#S4 "IV Maximum-Likelihood Detection in the Presence of Channel Estimation Errors ‣ On Optimal Turbo Decoding of Wideband MIMO-OFDM Systems Under Imperfect Channel State Information") to formulate the improved ML decoding metric in the presence of imperfect channel state information at the receiver (CSIR). In section [V](#S5 "V Iterative Decoding of BICM MIMO-OFDM Based On Imperfect CSIR ‣ On Optimal Turbo Decoding of Wideband MIMO-OFDM Systems Under Imperfect Channel State Information"), we use the general modified metric for soft decoding BICM MB-OFDM and MIMO-OFDM systems. Section [VI](#S6 "VI Simulation results ‣ On Optimal Turbo Decoding of Wideband MIMO-OFDM Systems Under Imperfect Channel State Information") illustrates via simulations the performance of the proposed receiver over both realistic UWB channel environments and uncorrelated Rayleigh fading channels and section [VII](#S7 "VII Conclusion ‣ On Optimal Turbo Decoding of Wideband MIMO-OFDM Systems Under Imperfect Channel State Information") concludes the paper.
Notational conventions are as follows. $`\mathbf{I}_{N}`$ represents an ($`N \times N`$) identity matrix; $`{\mathbb{E}}_{\mathbf{x}}{\lbrack.\rbrack}`$ refers to expectation with respect to $`\mathbf{x}`$; $`|.|`$ and $`\parallel . \parallel`$ and $`{Tr}{(.)}`$ denote matrix determinant, Frobenius norm and matrix trace respectively; $`{(.)}^{T}`$ and $`{(.)}^{\mathcal{H}}`$ denote vector transpose and Hermitian transpose, respectively.
## II TRANSMISSION MODEL
###
II-A
Multiband OFDM
A MB-OFDM system divides the spectrum between 3.1 to 10.6 GHz into several non-overlapping subbands each one occupying approximately 500 MHZ of bandwidth \[[2](#bib.bib2)\]. Information is transmitted using OFDM modulation over one of the subbands in a particular time-slot. The transmitter architecture is depicted in Fig. [1](#S2.F1 "Figure 1 ‣ II-A Multiband OFDM ‣ II TRANSMISSION MODEL ‣ On Optimal Turbo Decoding of Wideband MIMO-OFDM Systems Under Imperfect Channel State Information"). The binary sequence is convolutionally encoded and then interleaved by a random interleaver. The interleaved bits are gathered in subsequences of $`B`$ bits $`d_{k}^{1},\ldots,d_{k}^{B}`$ and mapped to complex M-QAM ($`M = 2^{B}`$) symbols $`s_{k}`$ with average energy $`E_{s} = {{\mathbb{E}}\hspace{0pt}{\lbrack{|s_{k}|}^{2}\rbrack}}`$. At a particular time-slot, a time-frequency code (TFC) selects the center frequency of the subband over which the OFDM symbol is sent. The TFC is used not only to provide frequency diversity but also to distinguish between multiple users. We assume an OFDM transmission with $`M`$ subcarriers, through a frequency selective multipath fading channel, described in discrete-time baseband equivalent form by the taps $`{\{ h_{l}\}}_{l = 0}^{L}`$. At the receiver, after removing the cyclic prefix (CP) and performing fast Fourier transform (FFT), a received OFDM symbol over a given subband can be written as
<table>
<tbody>
<tr>
<td></td>
<td>𝒚=𝐃hf𝒔+𝒛,𝒚subscript𝐃subscriptℎ𝑓𝒔𝒛{\bm{y}}=\mathbf{D}_{h_{f}}\,{\bm{s}}+{\bm{z}},</td>
<td></td>
<td>(1)</td>
</tr>
</tbody>
</table>
where $`({M \times 1})`$ vectors $`\mathbf{y}`$ and $`\mathbf{s}`$ denote received and transmitted symbols, respectively; the noise block $`\mathbf{z}`$ is assumed to be a circularly symmetric complex Gaussian random vector with distribution $`{\mathbf{z}} \sim {\mathcal{C}\hspace{0pt}\mathcal{N}\hspace{0pt}{(\mathbf{0},{\sigma_{z}^{2}\hspace{0pt}\mathbf{I}_{M}})}}`$; and $`\mathbf{D}_{h_{f}}`$ is a diagonal matrix with diagonal elements $`{\mathbf{h}}_{f} = {\lbrack{h_{f}\hspace{0pt}{(0)}},\ldots,{h_{f}\hspace{0pt}{({M - 1})}}\rbrack}^{T}`$, where $`{h_{f}\hspace{0pt}{(k)}} = {\sum_{l = 0}^{L}{h_{l}\hspace{0pt}e^{- {{j\hspace{0pt}2\hspace{0pt}\pi\hspace{0pt}k\hspace{0pt}l}/M}}}}`$.
\psfrag{A}{Binary Data}\psfrag{B}{{\parbox{60.00009pt}{\centering\$R=1/2\$ Convolutional Encoder\\add@centering}}}\psfrag{C}{{\parbox{60.00009pt}{\centering Bit\\ interleaver\\add@centering}}}\psfrag{D}{\parbox{60.00009pt}{\centering M-QAM\\ Mapping\\add@centering}}\psfrag{E}{\parbox{60.00009pt}{\centering IFFT\\ CP \\ GI\\ Addition\\add@centering}}\psfrag{F}{\parbox{60.00009pt}{\centering DAC\\add@centering}}\psfrag{G}{\parbox{60.00009pt}{\centering\$\exp\left(j2\pi f\_{c}t\right)\$\\add@centering}}\psfrag{H}{{TFC: Subband Selection}}\includegraphics\[width=260.17464pt,height=142.26378pt\]{TX_chainII}
Figure 1: TX architecture of the multiband OFDM system.
###
II-B
MIMO-OFDM
We consider a single-user MIMO-OFDM communication system over a memoryless frequency selective Rayleigh fading channel. The system consists of $`M_{T}`$ transmit and $`M_{R}`$ receive antennas ($`M_{R} \geq M_{T}`$), and $`M`$ is the total number of subcarriers. Fig. [2](#S2.F2 "Figure 2 ‣ II-B MIMO-OFDM ‣ II TRANSMISSION MODEL ‣ On Optimal Turbo Decoding of Wideband MIMO-OFDM Systems Under Imperfect Channel State Information") depicts the BICM coding scheme used at the transmitter. The binary data sequence $`\mathbf{b}`$ are encoded by a non-recursive non-systematic convolutional (NRNSC) code before being interleaved by a quasi-random interleaver. The output bits $`\mathbf{d}`$ are multiplexed to $`M_{T}`$ substreams and mapped to complex $`\overset{\sim}{M}`$-QAM symbols before being modulated by the OFDM modulator and transmitted through $`M_{T}`$ antennas.
[Refer to caption](/html/0708.1413/assets/x1.png)
Figure 2: Block diagram of MIMO-OFDM transmission scheme.
Let $`\mathbf{s}`$ be the $`{M\hspace{0pt}M_{T}} \times 1`$ vector containing the OFDM symbols transmitted simultaneously over $`M_{T}`$ antennas. The symbols are assumed to be independent identically distributed (i.i.d.) with zero mean and unit covariance matrix $`\mathbf{\Sigma}_{s} = {{\mathbb{E}}\hspace{0pt}{\lbrack{{\mathbf{s}}\hspace{0pt}{\mathbf{s}}^{\mathcal{H}}}\rbrack}} = {\mathbb{I}}_{M \times M_{T}}`$. Assuming an invariant channel over a frame of $`L`$ symbols, the received vector $`\mathbf{y}`$ at a given time index $`l`$ (omitted for brevity) can be written as
<table>
<tbody>
<tr>
<td></td>
<td>𝒚=𝐇𝒔+𝒛𝒚𝐇𝒔𝒛{\bm{y}}={\mathbf{H}}\,{\bm{s}}+{\bm{z}}</td>
<td></td>
<td>(2)</td>
</tr>
</tbody>
</table>
where $`\mathbf{H}`$ is a $`{{M\hspace{0pt}M_{R}} \times M}\hspace{0pt}M_{T}`$ block diagonal channel matrix containing the frequency response of the MIMO channels and the noise vector $`\mathbf{z}`$ is assumed to be a zero-mean circularly symmetric complex Gaussian (ZMCSCG) random vector with covariance matrix $`\mathbf{\Sigma}_{z} \triangleq {{\mathbb{E}}\hspace{0pt}{({{\mathbf{z}}\hspace{0pt}{\mathbf{z}}^{\mathcal{H}}})}} = {\sigma_{z}^{2}\hspace{0pt}{\mathbb{I}}_{M \times M_{R}}}`$. We assume that for each frame, a different realization of $`\mathbf{H}`$ independent of both $`\mathbf{s}`$ and $`\mathbf{z}`$ is drawn and remains constant during this frame. The MIMO-OFDM channel can be decoupled into $`M`$ frequency flat MIMO channels by exploiting the block diagonal structure in ([2](#S2.E2 "In II-B MIMO-OFDM ‣ II TRANSMISSION MODEL ‣ On Optimal Turbo Decoding of Wideband MIMO-OFDM Systems Under Imperfect Channel State Information")) which can be rewritten as a set of $`M`$ equations that contains only one subcarrier each
<table>
<tbody>
<tr>
<td></td>
<td>𝒚k=𝐇k𝒔k+𝒛kk=1,…,M,formulae-sequencesubscript𝒚𝑘subscript𝐇𝑘subscript𝒔𝑘subscript𝒛𝑘𝑘1…𝑀{\bm{y}}_{k}={\mathbf{H}}_{k}\,{\bm{s}}_{k}+{\bm{z}}_{k}\;\;\;\;k=1,...,M,</td>
<td></td>
<td>(3)</td>
</tr>
</tbody>
</table>
where $`\mathbf{H} = {{diag}\hspace{0pt}{\lbrack{\mathbf{H}_{1}\hspace{0pt}\mathbf{H}_{2}\hspace{0pt}\ldots\hspace{0pt}\mathbf{H}_{M}}\rbrack}}`$, $`{\mathbf{y}}^{T} = {\lbrack{{\mathbf{y}}_{1}^{T}\hspace{0pt}\ldots\hspace{0pt}{\mathbf{y}}_{M}^{T}}\rbrack}`$, $`{\mathbf{s}}^{T} = {\lbrack{{\mathbf{s}}_{1}^{T}\hspace{0pt}\ldots\hspace{0pt}{\mathbf{s}}_{M}^{T}}\rbrack}`$ and $`{\mathbf{z}}^{T} = {\lbrack{{\mathbf{z}}_{1}^{T}\hspace{0pt}\ldots\hspace{0pt}{\mathbf{z}}_{M}^{T}}\rbrack}`$. The architecture of ([3](#S2.E3 "In II-B MIMO-OFDM ‣ II TRANSMISSION MODEL ‣ On Optimal Turbo Decoding of Wideband MIMO-OFDM Systems Under Imperfect Channel State Information")) constitutes the basis for the study in this paper.
Hereafter, without loss of generality, we adopt to the more general MIMO-OFDM model of equation ([2](#S2.E2 "In II-B MIMO-OFDM ‣ II TRANSMISSION MODEL ‣ On Optimal Turbo Decoding of Wideband MIMO-OFDM Systems Under Imperfect Channel State Information")), where the MB-OFDM case can be deduced by setting $`M_{T} = M_{R} = 1`$.
## III Pilot Based Channel Estimation
In practical situations, the receiver has only access to a noisy estimate of the channel that differs from the true channel. Under the assumption of a time-invariant channel over the entire transmitted frame, channel estimation is usually performed on the basis of known training (pilot) symbols, transmitted at the beginning of each frame.
We consider the estimation of channel matrix $`\mathbf{H}_{k}`$ via the transmission of $`N`$ training vectors $`{\mathbf{s}}_{_{T,i}}`$, ($`i = {1,\ldots,N}`$). According to ([2](#S2.E2 "In II-B MIMO-OFDM ‣ II TRANSMISSION MODEL ‣ On Optimal Turbo Decoding of Wideband MIMO-OFDM Systems Under Imperfect Channel State Information")), when pilot symbols are transmitted we receive
<table>
<tbody>
<tr>
<td></td>
<td>𝐘T=𝐇k𝐒T+𝐙Tsubscript𝐘𝑇subscript𝐇𝑘subscript𝐒𝑇subscript𝐙𝑇\mathbf{Y}_{T}={\mathbf{H}}_{k}\,\mathbf{S}_{T}+\mathbf{Z}_{T}</td>
<td></td>
<td>(4)</td>
</tr>
</tbody>
</table>
where each column of the $`M_{T} \times N`$ matrix $`\mathbf{S}_{T} = {\lbrack{{\mathbf{s}}_{_{T,1}}\hspace{0pt}{|\ldots|}\hspace{0pt}{\mathbf{s}}_{_{T,N}}}\rbrack}`$ contains one pilot symbol and the noise $`\mathbf{Z}_{T}`$ has the same distribution as the noise $`{\mathbf{z}}_{k}`$. The average energy of the training symbols is $`P_{T} = {\frac{1}{N\hspace{0pt}M_{T}}\hspace{0pt}{Tr}\hspace{0pt}\left( {\mathbf{S}_{T}\hspace{0pt}\mathbf{S}_{T}^{\mathcal{H}}} \right)}`$. The ML estimate of $`\mathbf{H}_{k}`$ is obtained by minimizing $`{\|{\mathbf{Y}_{T} - {\mathbf{H}_{k}\hspace{0pt}\mathbf{S}_{T}}}\|}^{2}`$ with respect to $`\mathbf{H}_{k}`$. We have
<table>
<tbody>
<tr>
<td></td>
<td>𝐇^kML=𝐘T𝐒Tℋ(𝐒T𝐒Tℋ)−1=𝐇k+ℰsuperscriptsubscript^𝐇𝑘MLsubscript𝐘𝑇superscriptsubscript𝐒𝑇ℋsuperscriptsubscript𝐒𝑇superscriptsubscript𝐒𝑇ℋ1subscript𝐇𝑘ℰ\widehat{{\mathbf{H}}}_{k}^{\rm ML}=\mathbf{Y}_{T}\,\mathbf{S}_{T}^{\mathcal{H}}\,(\mathbf{S}_{T}\mathbf{S}_{T}^{\mathcal{H}})^{-1}={\mathbf{H}}_{k}+\mathbf{\mathcal{E}}</td>
<td></td>
<td>(5)</td>
</tr>
</tbody>
</table>
where $`\mathcal{E} = {\mathbf{Z}_{T}\hspace{0pt}\mathbf{S}_{T}^{\mathcal{H}}\hspace{0pt}{({\mathbf{S}_{T}\hspace{0pt}\mathbf{S}_{T}^{\mathcal{H}}})}^{- 1}}`$ denotes the estimation error matrix. When the training sequence is orthogonal ($`{\mathbf{S}_{T}\hspace{0pt}\mathbf{S}_{T}^{\mathcal{H}}} = {N\hspace{0pt}P_{T}\hspace{0pt}{\mathbb{I}}_{M_{T}}}`$), the $`j`$-th row $`\mathcal{E}_{j}`$ of the estimation error matrix reduces to a white noise vector with covariance matrix $`\mathbf{\Sigma}_{\mathcal{E},j} = {{\mathbb{E}}\hspace{0pt}\left\lbrack {{\mathcal{E}}_{j}^{\mathcal{H}}\hspace{0pt}{\mathcal{E}}_{j}} \right\rbrack} = {\sigma_{\mathcal{E},k}^{2}\hspace{0pt}{\mathbb{I}}_{M_{T}}}`$, where $`\sigma_{\mathcal{E},k}^{2} = {S\hspace{0pt}N\hspace{0pt}R_{T}^{- 1}} \triangleq \frac{N\hspace{0pt}P_{T}}{\sigma_{z}^{2}}`$. By assuming that the channel matrix $`\mathbf{H}_{k}`$ has the prior distribution $`\mathbf{H}_{k} \sim {\mathcal{C}\hspace{0pt}\mathcal{N}\hspace{0pt}{(\mathbf{0},{{\mathbb{I}}_{M_{T}} \otimes \mathbf{\Sigma}_{H,k}})}}`$ and choosing an orthogonal training sequence, we can derive the posterior distribution of the perfect channel conditioned on the estimated channel as
<table>
<tbody>
<tr>
<td></td>
<td>f(𝐇k|𝐇^kML)=𝒞𝒩(𝚺Δ𝐇^ML,𝕀MT⊗𝚺Δ𝚺ℰ)𝑓conditionalsubscript𝐇𝑘superscriptsubscript^𝐇𝑘ML𝒞𝒩subscript𝚺Δsubscript^𝐇MLtensor-productsubscript𝕀subscript𝑀𝑇subscript𝚺Δsubscript𝚺ℰf({\mathbf{H}}_{k}|\widehat{{\mathbf{H}}}_{k}^{\rm ML})=\mathcal{CN}(\mathbf{\Sigma}_{\Delta}\widehat{{\mathbf{H}}}_{\rm ML},\,{\mathbb{I}}_{M_{T}}\otimes\mathbf{\Sigma}_{\Delta}\mathbf{\Sigma}_{\mathcal{E}})</td>
<td></td>
<td>(6)</td>
</tr>
</tbody>
</table>
where $`\mathbf{\Sigma}_{\Delta} = {\mathbf{\Sigma}_{H,k}\hspace{0pt}{({\mathbf{\Sigma}_{\mathcal{E}} + \mathbf{\Sigma}_{H,k}})}^{- 1}} = {\delta\hspace{0pt}{\mathbb{I}}_{M_{R}}}`$ and $`\delta = \frac{S\hspace{0pt}N\hspace{0pt}R_{T}\hspace{0pt}\sigma_{h}^{2}}{({{S\hspace{0pt}N\hspace{0pt}R_{T}\hspace{0pt}\sigma_{h}^{2}} + 1})}`$.
The availability of the estimation error distribution constitutes an interesting feature of pilot assisted channel estimation that we used to derive the posterior distribution ([6](#S3.E6 "In III Pilot Based Channel Estimation ‣ On Optimal Turbo Decoding of Wideband MIMO-OFDM Systems Under Imperfect Channel State Information")). This distribution is exploited in the next section, in the formulation of a modified metric for improving the detection performance under imperfect channel estimation.
## IV Maximum-Likelihood Detection in the Presence of Channel Estimation Errors
###
IV-A
Mismatched ML Detection
It is well known that under i.i.d. Gaussian noise, detecting $`{\mathbf{s}}_{k}`$ is given by maximizing the likelihood function $`W\hspace{0pt}{(\left. {\mathbf{y}}_{k} \middle| {{\mathbf{s}}_{k},\mathbf{H}_{k}} \right.)}`$ which is equivalent to minimizing the Euclidean distance $`\mathcal{D}_{ML}`$
<table>
<tbody>
<tr>
<td></td>
<td>𝒔^kML(𝐇k)=argmin𝒔k∈ℂMT×1{𝒟ML(𝒔k,𝒚k,𝐇k)},superscriptsubscript^𝒔𝑘MLsubscript𝐇𝑘subscriptargminsubscript𝒔𝑘superscriptℂsubscript𝑀𝑇1subscript𝒟MLsubscript𝒔𝑘subscript𝒚𝑘subscript𝐇𝑘\hat{{\bm{s}}}_{k}^{\rm ML}({\mathbf{H}}_{k})=\mathop{\mathrm{arg\,min}}_{{\bm{s}}_{k}\in\,\mathbb{C}^{M_{T}\times 1}}\,\big{\{}\,\mathcal{D}_{\rm ML}({\bm{s}}_{k},{\bm{y}}_{k},{\mathbf{H}}_{k})\big{\}},</td>
<td></td>
<td>(7)</td>
</tr>
</tbody>
</table>
where $`{\mathcal{D}_{ML}\hspace{0pt}{({\mathbf{s}}_{k},{\mathbf{y}}_{k},\mathbf{H}_{k})}} \triangleq {- {{\ln W}\hspace{0pt}{(\left. {\mathbf{y}}_{k} \middle| {{\mathbf{s}}_{k},\mathbf{H}_{k}} \right.)}}} \propto {\|{{\mathbf{y}}_{k} - {\mathbf{H}_{k}\hspace{0pt}{\mathbf{s}}_{k}}}\|}^{2}`$. Since the above detection rule requires the knowledge of the perfect channel matrix $`\mathbf{H}_{k}`$, one sub-optimal approach, referred to as mismatched detection, consists in replacing the exact channel by its estimate in the receiver metric as
<table>
<tbody>
<tr>
<td></td>
<td>𝒔^kML(𝐇^k)=argmin𝒔k∈ℂMT×1{‖𝒚k−𝐇k𝒔k‖2}|𝐇k=𝐇^k\hat{{\bm{s}}}_{k}^{\rm ML}(\widehat{{\mathbf{H}}}_{k})=\mathop{\mathrm{arg\,min}}_{{\bm{s}}_{k}\in\,\mathbb{C}^{M_{T}\times 1}}\,\big{\{}\,\|{\bm{y}}_{k}-{\mathbf{H}}_{k}{\bm{s}}_{k}\|^{2}\big{\}}_{\big{|}_{{\mathbf{H}}_{k}=\widehat{{\mathbf{H}}}_{k}}}</td>
<td></td>
<td>(8)</td>
</tr>
</tbody>
</table>
###
IV-B
Improved ML Detection for imperfect CSIR
The Bayesian framework introduced previously let us to define a new likelihood function $`\overset{\sim}{W}\hspace{0pt}{(\left. {\mathbf{y}}_{k} \middle| {{\hat{\mathbf{H}}}_{k},{\mathbf{s}}_{k}} \right.)}`$ by averaging the likelihood that would be used if the channel were perfectly known ($`W\hspace{0pt}{(\left. {\mathbf{y}}_{k} \middle| {H_{k},{\mathbf{s}}_{k}} \right.)}`$), over all realizations of the perfect channel for a given estimated channel state (the posterior distribution ([6](#S3.E6 "In III Pilot Based Channel Estimation ‣ On Optimal Turbo Decoding of Wideband MIMO-OFDM Systems Under Imperfect Channel State Information"))). This yields
<table>
<tbody>
<tr>
<td></td>
<td>W~(𝒚k|𝐇^k,𝒔k)~𝑊conditionalsubscript𝒚𝑘subscript^𝐇𝑘subscript𝒔𝑘\displaystyle{\widetilde{W}}({\bm{y}}_{k}|\widehat{{\mathbf{H}}}_{k},{\bm{s}}_{k})</td>
<td>=𝔼𝐇k|𝐇^k[W(𝒚k|𝐇k,𝒔k)|𝐇^k]absentsubscript𝔼conditionalsubscript𝐇𝑘subscript^𝐇𝑘delimited-[]conditional𝑊conditionalsubscript𝒚𝑘subscript𝐇𝑘subscript𝒔𝑘subscript^𝐇𝑘\displaystyle={\mathbb{E}}_{{\mathbf{H}}_{k}|\widehat{{\mathbf{H}}}_{k}}\big{[}{W}({\bm{y}}_{k}|{\mathbf{H}}_{k},{\bm{s}}_{k})\big{|}\,\widehat{{\mathbf{H}}}_{k}\big{]}</td>
<td></td>
<td></td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td></td>
<td>=∫𝐇k∈ℂMR×MTW(𝒚k|𝐇k,𝒔k)f(𝐇k|𝐇^k)d𝐇kabsentsubscriptsubscript𝐇𝑘superscriptℂsubscript𝑀𝑅subscript𝑀𝑇𝑊conditionalsubscript𝒚𝑘subscript𝐇𝑘subscript𝒔𝑘𝑓conditionalsubscript𝐇𝑘subscript^𝐇𝑘differential-dsubscript𝐇𝑘\displaystyle=\int_{{\mathbf{H}}_{k}\in\mathbb{C}^{M_{R}\times M_{T}}}{W}({\bm{y}}_{k}|{\mathbf{H}}_{k},{\bm{s}}_{k})\;f({\mathbf{H}}_{k}|\widehat{{\mathbf{H}}}_{k})\;\;\;{\rm d}{\mathbf{H}}_{k}</td>
<td></td>
<td>(9)</td>
</tr>
</tbody>
</table>
Since both $`W\hspace{0pt}{(\left. {\mathbf{y}}_{k} \middle| {\mathbf{H}_{k},{\mathbf{s}}_{k}} \right.)}`$ and $`f\hspace{0pt}{(\left. \mathbf{H}_{k} \middle| {\hat{\mathbf{H}}}_{k} \right.)}`$ are Gaussian densities, it is easy to show that $`{\overset{\sim}{W}\hspace{0pt}{(\left. {\mathbf{y}}_{k} \middle| {{\hat{\mathbf{H}}}_{k},{\mathbf{s}}_{k}} \right.)}} = {\mathcal{C}\hspace{0pt}\mathcal{N}\hspace{0pt}{({\mathbf{μ}}_{_{\mathcal{M}}},\mathbf{\Sigma}_{_{\mathcal{M}}})}}`$ with \[[6](#bib.bib6)\]
<table>
<tbody>
<tr>
<td></td>
<td>{𝝁ℳ=δ𝐇^k𝒔k,𝚺ℳ=𝚺z+δ𝚺ℰ‖𝒔k‖2.casessubscript𝝁ℳabsent𝛿subscript^𝐇𝑘subscript𝒔𝑘subscript𝚺ℳabsentsubscript𝚺𝑧𝛿subscript𝚺ℰsuperscriptnormsubscript𝒔𝑘2\left\{\begin{array}[]{ll}\bm{\mu}_{{}_{\mathcal{M}}}&=\delta\,\widehat{{\mathbf{H}}}_{k}\,{\bm{s}}_{k},\\ \mathbf{\Sigma}_{{}_{\mathcal{M}}}&=\mathbf{\Sigma}_{z}+\delta\,\mathbf{\Sigma}_{\mathcal{E}}\,\|{\bm{s}}_{k}\|^{2}.\end{array}\right.</td>
<td></td>
<td>(10)</td>
</tr>
</tbody>
</table>
Now the ML estimate of $`{\mathbf{s}}_{k}`$ can be formulated as
<table>
<tbody>
<tr>
<td></td>
<td>𝒔^kℳ(𝐇^k)=argmin𝒔k∈ℂMT×1{𝒟ℳ(𝒔k,𝒚k,𝐇k)},superscriptsubscript^𝒔𝑘ℳsubscript^𝐇𝑘subscriptargminsubscript𝒔𝑘superscriptℂsubscript𝑀𝑇1subscript𝒟ℳsubscript𝒔𝑘subscript𝒚𝑘subscript𝐇𝑘\hat{{\bm{s}}}_{k}^{\mathcal{M}}(\widehat{{\mathbf{H}}}_{k})=\mathop{\mathrm{arg\,min}}_{{\bm{s}}_{k}\in\,\mathbb{C}^{M_{T}\times 1}}\,\big{\{}\,\mathcal{D}_{\mathcal{M}}({\bm{s}}_{k},{\bm{y}}_{k},{\mathbf{H}}_{k})\big{\}},</td>
<td></td>
<td>(11)</td>
</tr>
</tbody>
</table>
with
<table>
<tbody>
<tr>
<td></td>
<td></td>
<td>𝒟ℳ(𝒔k,𝒚k,𝐇^k)≜−lnW~(𝒚k|𝒔k,𝐇^k)≜subscript𝒟ℳsubscript𝒔𝑘subscript𝒚𝑘subscript^𝐇𝑘~𝑊conditionalsubscript𝒚𝑘subscript𝒔𝑘subscript^𝐇𝑘\displaystyle\mathcal{D}_{{}_{\mathcal{M}}}({\bm{s}}_{k},{\bm{y}}_{k},\widehat{{\mathbf{H}}}_{k})\triangleq-\ln{\widetilde{W}}({\bm{y}}_{k}|{\bm{s}}_{k},\widehat{{\mathbf{H}}}_{k})</td>
<td></td>
<td></td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td></td>
<td>=MRlnπ(σz2+δσℰ2‖𝒔k‖2)+‖𝒚k−δ𝐇^k𝒔k‖2σz2+δσℰ2‖𝒔k‖2absentsubscript𝑀𝑅𝜋subscriptsuperscript𝜎2𝑧𝛿subscriptsuperscript𝜎2ℰsuperscriptnormsubscript𝒔𝑘2superscriptnormsubscript𝒚𝑘𝛿subscript^𝐇𝑘subscript𝒔𝑘2subscriptsuperscript𝜎2𝑧𝛿subscriptsuperscript𝜎2ℰsuperscriptnormsubscript𝒔𝑘2\displaystyle=M_{R}\ln\,\pi(\sigma^{2}_{z}+\delta\,\sigma^{2}_{\mathcal{E}}\,\|{\bm{s}}_{k}\|^{2})+\frac{\|{\bm{y}}_{k}-\delta\,\widehat{{\mathbf{H}}}_{k}\,{\bm{s}}_{k}\|^{2}}{\sigma^{2}_{z}+\delta\,\sigma^{2}_{\mathcal{E}}\,\|{\bm{s}}_{k}\|^{2}}</td>
<td></td>
<td>(12)</td>
</tr>
</tbody>
</table>
being the new ML decision metric under CEE.
We note that when exact channel is available ($`{\hat{\mathbf{H}}}_{k} = \mathbf{H}_{k}`$), the posterior expectation of ([](#S4.Ex1 "IV-B Improved ML Detection for imperfect CSIR ‣ IV Maximum-Likelihood Detection in the Presence of Channel Estimation Errors ‣ On Optimal Turbo Decoding of Wideband MIMO-OFDM Systems Under Imperfect Channel State Information")
IV-B
) becomes equivalent to replacing $`\mathbf{H}_{k}`$ by $`{\hat{\mathbf{H}}}_{k}`$ in $`W\hspace{0pt}{(\left. {\mathbf{y}}_{k} \middle| {\mathbf{H}_{k},{\mathbf{s}}_{k}} \right.)}`$ and consequently the the two metrics $`\mathcal{D}_{_{\mathcal{M}}}`$ and $`\mathcal{D}_{ML}`$ coincides. Under near perfect CSIR, occurred either when $`\sigma_{\mathcal{E}}^{2}\rightarrow 0`$ or when $`N\rightarrow\infty`$, we have $`\delta\rightarrow 1`$, $`{\delta\hspace{0pt}\sigma_{\mathcal{E}}^{2}}\rightarrow 0`$, $`{\mathbf{μ}}_{_{\mathcal{M}}}\rightarrow{{\hat{\mathbf{H}}}_{k}\hspace{0pt}{\mathbf{s}}}`$ and $`\mathbf{\Sigma}_{_{\mathcal{M}}}\rightarrow{\sigma_{z}^{2}\hspace{0pt}{\mathbb{I}}_{M_{R}}}`$. Consequently, the improved metric $`\mathcal{D}_{_{\mathcal{M}}}`$ behaves similarly to the classical Euclidean distance metric $`\mathcal{D}_{ML}`$.
[Refer to caption](/html/0708.1413/assets/x2.png)
Figure 3: Block diagram of MIMO-OFDM BICM receiver.
## V Iterative Decoding of BICM MIMO-OFDM Based On Imperfect CSIR
As a practical application of the general decoding metric ([](#S4.Ex2 "IV-B Improved ML Detection for imperfect CSIR ‣ IV Maximum-Likelihood Detection in the Presence of Channel Estimation Errors ‣ On Optimal Turbo Decoding of Wideband MIMO-OFDM Systems Under Imperfect Channel State Information")
IV-B
) proposed in \[[5](#bib.bib5)\], we consider soft iterative decoding of BICM MIMO-OFDM under imperfect CSIR. This problem has been addressed in \[[7](#bib.bib7)\] under the assumption of perfect CSIR. Here, without going into the details, we extend the results of \[[6](#bib.bib6)\] to MIMO-OFDM block fading channels estimated by a finite number of training symbols.
As shown in Fig. [3](#S4.F3 "Figure 3 ‣ IV-B Improved ML Detection for imperfect CSIR ‣ IV Maximum-Likelihood Detection in the Presence of Channel Estimation Errors ‣ On Optimal Turbo Decoding of Wideband MIMO-OFDM Systems Under Imperfect Channel State Information"), the BICM receiver consists of a bunch of demodulator/demapper, a de-interleaver and a soft-input-soft-output (SISO) decoder. Let $`d_{k}^{j,m}`$ be the $`m`$-th coded and interleaved bit ($`m = {1,2,\ldots,{\log_{2}\overset{\sim}{M}}}`$) of the constellation symbol $`{\mathbf{s}}_{k}`$ at the the $`i`$-th transmit antenna and the $`k`$-th subcarrier. We denote by $`L\hspace{0pt}{(d_{k}^{j,m})}`$ the coded log-likelihood ratio (LLR) value of the bit $`d_{k}^{j,m}`$. At each decoding iteration, the LLR values conditioned on the CSIR are given by
<table>
<tbody>
<tr>
<td></td>
<td>L(dkj,m)=logPdem(dkj,m=1)|𝒚k,𝐇k)Pdem(dkj,m=0|𝒚k,𝐇k).L(d_{k}^{j,m})=\log\frac{P_{\rm dem}(d_{k}^{j,m}=1)|{\bm{y}}_{k},{\mathbf{H}}_{k})}{P_{\rm dem}(d_{k}^{j,m}=0|{\bm{y}}_{k},{\mathbf{H}}_{k})}.</td>
<td></td>
<td>(13)</td>
</tr>
</tbody>
</table>
We have (see \[[6](#bib.bib6)\] and references therein)
<table>
<tbody>
<tr>
<td></td>
<td>L(dkj,m)=log∑𝒔k:dkj,m=1exp{−𝒟ML(𝒔k,𝒚k,𝐇k)}∏i≠ji=1BPdec+(dki,m)∑𝒔k:dkj,m=0exp{−𝒟ML(𝒔k,𝒚k,𝐇k)}∏i≠ji=1BPdec−(dki,m)𝐿superscriptsubscript𝑑𝑘𝑗𝑚subscript:subscript𝒔𝑘superscriptsubscript𝑑𝑘𝑗𝑚1subscript𝒟MLsubscript𝒔𝑘subscript𝒚𝑘subscript𝐇𝑘superscriptsubscriptproductsuperscript𝑖𝑗𝑖1𝐵subscriptsuperscript𝑃decsuperscriptsubscript𝑑𝑘𝑖𝑚subscript:subscript𝒔𝑘superscriptsubscript𝑑𝑘𝑗𝑚0subscript𝒟MLsubscript𝒔𝑘subscript𝒚𝑘subscript𝐇𝑘superscriptsubscriptproductsuperscript𝑖𝑗𝑖1𝐵subscriptsuperscript𝑃decsuperscriptsubscript𝑑𝑘𝑖𝑚L(d_{k}^{j,m})=\log\frac{\sum\limits_{{\bm{s}}_{k}:\,d_{k}^{j,m}=1}\exp\big{\{}-\mathcal{D}_{{\rm ML}}({\bm{s}}_{k},{\bm{y}}_{k},{\mathbf{H}}_{k})\big{\}}\prod\limits_{\stackrel{{\scriptstyle i=1}}{{i\neq j}}}^{B}P^{+}_{\rm dec}\big{(}d_{k}^{i,m}\big{)}}{\sum\limits_{{\bm{s}}_{k}:\,d_{k}^{j,m}=0}\exp\big{\{}-\mathcal{D}_{{\rm ML}}({\bm{s}}_{k},{\bm{y}}_{k},{\mathbf{H}}_{k})\big{\}}\prod\limits_{\stackrel{{\scriptstyle i=1}}{{i\neq j}}}^{B}P^{-}_{\rm dec}\big{(}d_{k}^{i,m}\big{)}}</td>
<td></td>
<td>(14)</td>
</tr>
</tbody>
</table>
where $`P_{dec}^{+}\hspace{0pt}{(d_{k}^{i,m})}`$ and $`P_{dec}^{-}\hspace{0pt}{(d_{k}^{i,m})}`$ are extrinsic information coming from the SISO decoder.
Notice that the metric $`\mathcal{D}_{ML}\hspace{0pt}{({\mathbf{s}}_{k},{\mathbf{y}}_{k},\mathbf{H}_{k})}`$ involved in ([14](#S5.E14 "In V Iterative Decoding of BICM MIMO-OFDM Based On Imperfect CSIR ‣ On Optimal Turbo Decoding of Wideband MIMO-OFDM Systems Under Imperfect Channel State Information")) requires the knowledge of the perfect channel $`\mathbf{H}_{k}`$ of which the receiver has solely an estimate $`{\hat{\mathbf{H}}}_{k}`$. Contrary to the mismatch approach that replaces the perfect channel by its estimate, we propose to use the improved decoding metric $`\mathcal{D}_{_{\mathcal{M}}}\hspace{0pt}{({\mathbf{s}}_{k},{\mathbf{y}}_{k},{\hat{\mathbf{H}}}_{k})}`$ in ([14](#S5.E14 "In V Iterative Decoding of BICM MIMO-OFDM Based On Imperfect CSIR ‣ On Optimal Turbo Decoding of Wideband MIMO-OFDM Systems Under Imperfect Channel State Information")), in order to derive a new demaping rule adapted to the imperfect channel available at the receiver. The decoder accepts the LLRs of all coded bits and employs the well known forward-backward algorithm \[[8](#bib.bib8)\] to compute the LLRs of information bits, which are used for the decision.
## VI Simulation results
Simulations have been carried out in the context of IEEE802.15 wireless PAN \[[9](#bib.bib9)\]: $`M = 100`$ data subcarriers along with 32 CP samples compose one OFDM symbol. Information bits are encoded by a rate $`R = {1/2}`$ convolutional encoder with constraint length 3 defined in octal form by (5,7). The interleaver is a pseudo-random one operating over the entire frame and the mapping is 16-QAM. Throughout the simulations, each frame is assumed to consists of one OFDM symbol with 100 subcarriers belonging to a 16-QAM constellation with Gray or set partion (SP) labeling. The interleaver is a pseudo-random one operating over the entire frame with size $`M \cdot M_{T} \cdot {\log_{2}{(\overset{\sim}{M})}}`$ bits. For each transmitted frame, a different realization of the realistic UWB channel model specified in \[[10](#bib.bib10)\] or Rayleigh distributed channel has been drawn and remains constant during the whole frame. Besides, it is assumed that the average pilot symbol energy is equal to the average data symbol energy. Moreover, the number of decoding iterations are set to either 3 or 4.
Figure [4](#Sx1.F4 "Figure 4 ‣ On Optimal Turbo Decoding of Wideband MIMO-OFDM Systems Under Imperfect Channel State Information") depicts the bit error rate (BER) performance gain that is obtained by decoding BICM MB-OFDM with the modified ML decoder in the context of LOS CM1 channel, where the channel estimation is performed by sending different training sequences with lengths $`N = {\{ 1,2,8\}}`$ per frame. Similar plots are shown in figure [5](#Sx1.F5 "Figure 5 ‣ On Optimal Turbo Decoding of Wideband MIMO-OFDM Systems Under Imperfect Channel State Information") with SP labeling. It can be noticed that for both Gray and SP labeling, the proposed decoder outperforms the mismatched decoder especially when few numbers of pilot symbols are dedicated for channel estimation. Note that the modified decoder with $`N = 1`$ pilot performs very close to the mismatched decoder with $`N = 2`$ pilots. For comparison, results obtained with theoretical Rayleigh fading channel are illustrated in figure [6](#Sx1.F6 "Figure 6 ‣ On Optimal Turbo Decoding of Wideband MIMO-OFDM Systems Under Imperfect Channel State Information"). It can be observed that for $`N = 2`$, the SNR to obtain a BER of $`10^{- 3}`$ is reduced by about $`1.5`$ dB if the modified ML decoding is used instead of the mismatched approach.
Figure [7](#Sx1.F7 "Figure 7 ‣ On Optimal Turbo Decoding of Wideband MIMO-OFDM Systems Under Imperfect Channel State Information") shows the BER performance versus the training sequence length $`N`$ at a fixed $`{E\hspace{0pt}b}/N_{0}`$ of $`12`$ dB for the CM1 channel with 16-QAM and Gray labeling. This allows to evaluate the number of training sequence necessary to achieve a certain BER. We observe that at $`{{E\hspace{0pt}b}/N_{0}} = 12`$ dB, the modified ML decoder requires $`N = 9`$ pilot symbols per frame to achieve a BER of $`10^{- 4}`$ while the mismatched decoder attains this BER for $`N = 12`$ pilot symbols. Besides, this has been outlined in Section IV, for large training sequence lengths ($`N \geq 12`$), both decoders have close performance.
Figure [8](#Sx1.F8 "Figure 8 ‣ On Optimal Turbo Decoding of Wideband MIMO-OFDM Systems Under Imperfect Channel State Information") depicts the BER performance over a $`2 \times 2`$ MIMO-OFDM channel estimated by $`N \in {\{ 2,4,8\}}`$ pilot symbols per frame. As observed, the increase in the required $`E_{b}/N_{0}`$ caused by CEE is an important effect of imperfect CSIR in the case of mismatched ML decoding. The figure shows that the SNR to obtain a BER of $`10^{- 5}`$ with $`N = 2`$ pilots is reduced by about $`1.5`$ dB if the improved decoder is used instead of the mismatched decoder. We also notice that the performance loss of the mismatched receiver with respect to the derived receiver becomes insignificant for $`N \geq 8`$. This can be explained from the expression of the metric ([](#S4.Ex2 "IV-B Improved ML Detection for imperfect CSIR ‣ IV Maximum-Likelihood Detection in the Presence of Channel Estimation Errors ‣ On Optimal Turbo Decoding of Wideband MIMO-OFDM Systems Under Imperfect Channel State Information")
IV-B
), where we note that by increasing the number of pilot symbols, this expression tends to the classical Euclidean distance metric. However, the proposed decoder outperforms the mismatched decoder especially when few numbers of pilot symbols are dedicated for channel estimation.
## VII Conclusion
This paper studied the problem of ML reception when the receiver has only access to a noisy estimate of the channel in the case of pilot assisted channel estimation. By using the statistics of the estimation error, we adopted a modified ML criterion that is expressed in terms of the estimated channel coefficient. This modified metric let us to formulate appropriate branch metrics for decoding BICM MB-OFDM with imperfect channel knowledge. Simulation results conducted over realistic UWB channels, indicate that mismatched decoding is quite sub-optimal for short training sequence and confirmed the adequacy of the adopted decoding rule in the presence of channel estimation errors. This was obtained without introducing any additional complexity.
## Acknowledgment
The authors would like to acknowledge discussions with Pablo Piantanida.
## References
- \[1\] FCC, “First report and order, revision of part 15 of the commission’s rules regarding ultra-wideband transmission systems,” Tech. Rep., Feb. 2004.
- \[2\] A. Batra, J. Balakrishnan, G. R. Aiello, J. R. Foerster, and A. Dabak, “Design of multiband OFDM system for realistic UWB channel environments,”
IEEE Transaction on Microwave Theory and Techniques, vol. 52, pp. 2123–2138, september 2004.
- \[3\] H. Bolcskei, D. Gesbert, and A. J. Paulraj, “On the capacity of OFDM-based spatial multiplexing systems,”
IEEE Trans. Comun., pp. 225–234, Feb. 2002.
- \[4\] P. Garg, R. K. Mallik, and H. M. Gupta, “Performance analysis of space-time coding with imperfect channel estimation,”
IEEE Trans. Wireless Commun., vol. 4, pp. 257–265, Jan. 2005.
- \[5\] Giorgio Taricco and Ezio Biglieri, “Space-time decoding with imperfect channel estimation,”
IEEE Transaction on Wireless Communications, vol. 4, no. 4, pp. 1874, 1888 2005.
- \[6\] S. Sadough, P. Piantanida, and P. Duhamel, “Achievable outage rates with improved decoding of BICM Multiband OFDM under channel estimation errors,”
in Asilomar Conf. on Signals, systems and computers, Oct. 2006.
- \[7\] J. J. Boutros, F. Boixadera, and C. Lamy, “Bit-interleaved coded modulations for multiple-input multiple-output channels,”
in Int. Symp. on Spread Spectrum Techniques and Applications, Sept. 2000, pp. 123–126.
- \[8\] L. Bahl, J. Cocke, F. Jelinek, and J. Raviv, “Optimal decoding of linear codes for minimizing symbol error rate,”
IEEE Trans. on Inf. Theory, pp. 284–287, March 1974.
- \[9\] Anuj Batra, Jaiganesh Balakrishnan, and Anad Dabak, “Multiband OFDM physical layer proposal for IEEE 802.15 task group 3a,” Tech. Rep., IEEE, july 2003.
- \[10\] J. Foerster, “Channel modeling sub-committee report final,” Tech. Rep., IEEE802.15-02/490, 2003.
[Refer to caption](/html/0708.1413/assets/x3.png)
Figure 4: BER performance of the proposed decoder over the LOS UWB channel CM1 for various training sequence lengths, 16-QAM with Gray labeling.
[Refer to caption](/html/0708.1413/assets/x4.png)
Figure 5: BER performance of the proposed decoder over the LOS UWB channel CM1 for various training sequence lengths, 16-QAM with Set Partition labeling.
[Refer to caption](/html/0708.1413/assets/x5.png)
Figure 6: BER performance of the proposed decoder over Rayleigh fading channel for various training sequence lengths, 16-QAM with Gray labeling.
[Refer to caption](/html/0708.1413/assets/x6.png)
Figure 7: BER versus the number of pilot symbols at $`{E_{b}/N_{0}} = 12`$ dB over the LOS channel CM1.
[Refer to caption](/html/0708.1413/assets/x7.png)
Figure 8: BER performance of the proposed and mismatched decoders in the case of a $`2 \times 2`$ MIMO-OFDM Rayleigh fading channel with $`M = 100`$ subcarriers for training sequence length $`N \in {\{ 2,4,8\}}`$. |
0708.3166 | cs.NI | cyber | # Web Server Benchmark Application WiiBench using Erlang/OTP R11 and Fedora-Core Linux 5.0
Achmad B. Mutiara\memberRegular Member
\authorinfoA. B. Mutiara is with the Graduate Program in Information System, Gunadarma University, Depok 16424, Indonesia. E-mail: amutiara@staff.gunadarma.ac.id
and T.A. Sabastian \authorinfoT. A. Sabastian is with the Department of Informatics Engineering, Gunadarma University, Depok 16424, Indonesia
###### Abstract
As the web grows and the amount of traffics on the web server increase, problems related to performance begin to appear. Some of the problems, such as the number of users that can access the server simultaneously, the number of requests that can be handled by the server per second (requests per second) to bandwidth consumption and hardware utilization like memories and CPU. To give better quality of service (QoS), web hosting providers and also the system administrators and network administrators who manage the server need a benchmark application to measure the capabilities of their servers. Later, the application intends to work under Linux/Unix – like platforms and built using Erlang/OTP R11 as a concurrent oriented language under Fedora Core Linux 5.0. WiiBench is divided into two main parts, the controller section and the launcher section. Controller is the core of the application. It has several duties, such as read the benchmark scenario file, configure the program based on the scenario, initialize the launcher section, gather the benchmark results from local and remote Erlang node where the launcher runs and write them in a log file (later the log file will be used to generate a report page for the sysadmin). Controller also has function as a timer which act as timing for user inters arrival to the server. Launcher generates a number of users based on the scenario, initialize them and start the benchmark by sending requests to the web server. The clients also gather the benchmark result and send them to the controller.
###### keywords:
Erlang, QoS, Network Management, Concurrent Programming, Distribution
## 1 Introduction
\PARstart
In the last two decades, human necessities in fast and accurate information create a lot of innovations in information technology, one of them is the internet. Since TCP/IP released to public in 1982 and World Wide Web (WWW) introduced in 1991, internet has became a popular media to access and publish information. The easy to use web mechanisms make people easy to search and publish information on the internet. The web service later grows to many aspects, such as entertainment, education, scientific research and many more.
To access the web on the internet, we need a certain server than can provide user access on the web pages. This server is called web server or HTTP server and has a main duty to serve user access to web pages contents, either static or dynamic.
As the web grows and the amount of traffics on the web server increase, problems related to performance begin to appear. Some of the problems, such as the number of users that can access the server simultaneously, the number of requests that can be handled by the server per second (requests per second) to bandwidth consumption and hardware utilization like memories and CPU.
To give better quality of service (QoS) , web hosting providers and also the system administrators and network administrators who manage the server need a benchmark application to measure the capabilities of their servers. Later, the application intends to work under Linux/Unix – like platforms and built using Erlang/OTP R11 as a concurrent oriented language under Fedora Core Linux 5.0.
Base on the above descriptions, there are some problems than can be summarized, such as:
1. 1\.
To give better Quality of Service, web hosting provider and also system administrators and network administrators who manage the web server need a benchmark application to measure the capabilities/ performances of their servers.
2. 2\.
The benchmark application is intended to be use by the network administrators and system administrators who work under Linux/Unix – like systems.
3. 3\.
The application is made by utilizing the concurrent capability of Erlang programming language under Linux operating system.
## 2 Theories
### 2.1 Web Server
The term web server can mean one of two things \[2\]:
1. 1\.
A computer or a number of computers which responsible for accepting HTTP requests from clients, which are known as web browsers, and serving them web pages, either static or dynamic pages.
2. 2\.
A computer program that provides the functionality described in the first sense of the term.
Web server also works based on several standards, such as \[2\]: HTTP response to HTTP Request, Logging, Configurability, Authentication, Handling Static and Dynamic Contents, Modular Support, Virtual Hosts
### 2.2 Erlang/OTP
Erlang is a concurrent programming language with a functional core. By this we mean that the most important property of the language is that it is concurrent and that secondly, the sequential part of the language is a functional programming language. Concurrent means that the language has focus on how to makes multiple executions threads to run and do computational work together. In Erlang, these execution threads are called processes. The sequential sub-set of the language expresses what happens from the point it time where a process receives a message to the point in time when it emits a message.
The early version of Erlang was developed by Ericsson Computer Science Laboratory in 1985. During that time, Ericsson couldn’t find an appropriate language that has high performance in concurrency especially for telecommunication applications programming (for switching, trunking, etc), so they developed their own language. OTP stands for Open Telecom Platform, OTP was developed by Ericsson Telecom AB for programming next generation switches and many Ericsson products are based on OTP. OTP includes the entire Erlang development system together with a set of libraries written in Erlang and other languages. OTP was originally designed for writing telecoms application but has proved equally useful for a wide range of non-telecom that have concurrent, distributed, and also fault tolerant applications. In 1998 Ericsson released Erlang and the OTP libraries as open source. Now, Erlang/OTP has reached the R11 version.
## 3 Why We Use Erlang ?
The simple answer to the question above that is we need concurrency in the benchmark application. The application must be able to generate multiple users to do some stress tests to the web server.
But, there are some good features in Erlang, and even the other languages don’t have these features. Some of these Erlang features are described below:
1. 1\.
In Erlang processes are light weight.
2. 2\.
Not only are Erlang processes light-weight, but also we can create many hundreds of thousands of such processes without noticeably degrading the performance of the system (unless of course they are all doing something at the same time)\[5\].
3. 3\.
In Erlang, processes share no data and the only way in which they can exchange data is by explicit message passing. “dangling” pointers are very difficult to program in the presence of hardware failures - we took the easy way out, by disallowing all such data structures. \[4\]
4. 4\.
In Erlang, processes scheduling operation is done by its own virtual machine, so Erlang didn’t inherit the underlying operating system processes scheduling.
5. 5\.
Real time. Erlang is intended for programming soft real-time systems where response times in the order of milliseconds are required.\[4\]
6. 6\.
Continuous operation.
7. 7\.
Automatic Memory management.
8. 8\.
Distribution.
### 3.1 Concurrent and Distributed Erlang
Concurrent in Erlang involves processes creation and deletion. In order to create a new process in Erlang, we use BIF (Built In Function) spawn/3 : spawn(modulename,functionname,argumentlists)
or
pid_variabe=spawn(modulename,functionname,
argumentlists)
The illustration of process creation can be seen in figure [1](#S3.F1 "Figure 1 ‣ 3.1 Concurrent and Distributed Erlang ‣ 3 Why We Use Erlang ? ‣ Web Server Benchmark Application WiiBench using Erlang/OTP R11 and Fedora-Core Linux 5.0").
[Refer to caption](/html/0708.3166/assets/image1.jpg)
Figure 1: Process Creation Illustration
A process which no longer need by the system will be automatically shutdown/delete by the virtual machine (Erlang Runtime System/ERTS). Meanwhile, the message parsing mechanism can be done by these codes :
<table>
<thead>
<tr>
<th>Pid ! Message,</th>
</tr>
</thead>
<tbody>
<tr>
<td>…..</td>
</tr>
<tr>
<td>receive</td>
</tr>
<tr>
<td>Message1 ->></td>
</tr>
<tr>
<td>Actions1;</td>
</tr>
<tr>
<td>Message2 ->></td>
</tr>
<tr>
<td>Actions2;</td>
</tr>
<tr>
<td>…</td>
</tr>
<tr>
<td>after Time ->></td>
</tr>
<tr>
<td>TimeOutActions</td>
</tr>
<tr>
<td>end</td>
</tr>
</tbody>
</table>
The illustration for message parsing can be seen in figure [2](#S3.F2 "Figure 2 ‣ 3.1 Concurrent and Distributed Erlang ‣ 3 Why We Use Erlang ? ‣ Web Server Benchmark Application WiiBench using Erlang/OTP R11 and Fedora-Core Linux 5.0").
[Refer to caption](/html/0708.3166/assets/image2.jpg)
Figure 2: Message Parsing Illustration
Besides the example that has been shown above, in Erlang we can also use Behaviour in OTP Design Principles to create, delete and do message parsing between processes. A distributed Erlang system is a number of Erlang Runtime System (we called them nodes) that communicated each other by using message parsing with pid (process indentifier) through TCP/IP sockets transparently. A node must be given a name before it can communicate each other. The name is either a long name or a short name like the examples below:
```
\$erl -name dilbert[long name]
(dilbert@uab.ericsson.se)1$>$
@erl-sname dilbert[short name]
(dilbert@uab)1$>$
```
A simple distribution in Erlang can be done by these codes:
<table>
<thead>
<tr>
<th>…</th>
</tr>
</thead>
<tbody>
<tr>
<td>Pid = spawn(Fun@Node)</td>
</tr>
<tr>
<td>…</td>
</tr>
<tr>
<td>alive(Node)</td>
</tr>
<tr>
<td>…</td>
</tr>
<tr>
<td>not_alive(Node)</td>
</tr>
<tr>
<td></td>
</tr>
</tbody>
</table>
### 3.2 Benchmark
The term benchmark can be described as:
1. 1\.
A group of parameters in which products (software or hardware) can be measured the performance according to these parameters.
2. 2\.
A computer program designed to measure the performance of software or hardware according to certain parameters.
3. 3\.
A group of performance criteria that must be complied by software or hardware.
Web server benchmark means that a benchmark activity is made to the web server to measure its performance based on several parameters and using certain computer program to do this activity.
## 4 Design
### 4.1 WiiBench Concepts
WiiBench is divided into two main parts, the controller section and the launcher section.
Controller is the core of the application. It has several duties, such as read the benchmark scenario file, configure the program based on the scenario, initialize the launcher section, gather the benchmark results from local and remote Erlang node where the launcher runs and write them in a log file (later the log file will be used to generate a report page for the sysadmin). Controller also has function as a timer which act as timing for user inters arrival to the server.
Launcher generates a number of user based on the scenario, initialize them and start the benchmark by sending requests to the web server. The clients also gather the benchmark result and send them to the controller.
The illustration for WiiBench Concepts can be seen in figure [3](#S4.F3 "Figure 3 ‣ 4.1 WiiBench Concepts ‣ 4 Design ‣ Web Server Benchmark Application WiiBench using Erlang/OTP R11 and Fedora-Core Linux 5.0")
Several parameters that can be measured by this application are:
1. 1\.
Number of Requests per second.
2. 2\.
Simultaneous users that can be served by the server (per second).
3. 3\.
Time that needs to connect for the client so it can be connected to server.
4. 4\.
Number of user that can be served during the duration of benchmark.
5. 5\.
Time that needs for a user so it can receive a full page of document/web page according to the request.
6. 6\.
Time that needs to complete a session (a group of requests) as described in the scenario.
7. 7\.
Network throughput.
8. 8\.
HTTP Status (200, 404).
The application also has a report generator that written in PERL and using GNUPLOT to generate graphs based on the benchmark results from the log file.
[Refer to caption](/html/0708.3166/assets/image3.jpg)
Figure 3: Illustration for WiiBench Concepts
### 4.2 The Scenario Files
The benchmark scenario file is written using XML and consists of several sections:
1. 1\.
The server section, where the user (sysadmin) describe the IP address of the server that he/she wants to benchmark.
2. 2\.
The client section, in this section, user can write the IP addresses where the launcher section starts and generate a numbers of clients.
3. 3\.
Inters arrival phase and benchmark duration section.
4. 4\.
The simulated user agents (web browser) section.
5. 5\.
The session and request section.
Each of the section describes above can be modified according to user necessity.
## 5 Implementation
### 5.1 Hardware and Software Specifications
WiiBench is implemented by using a simple topology consists of two computers and a server across Local Area Network in Kapuk Valley, Margonda, Depok. The two computers using an Intel Celeron Processor (1.8 GHz and 2.28 GHz) and running Linux operating System (SuSE 10.0 and Slackware Linux 11.0), Open SSH v2, Erlang/OTP R11, PERL v5.8, and also BASH Shell. Both of them also connected to the TCP/IP network and within Kapuk-Valley domain (hostname mobile and posen-asik). The server using an Intel Pentium II 400 MHz processor with 768 megabytes SDRAM memories and running Slackware Linux 11.0 with Apache 2.0.55 web server.
### 5.2 Testing and Implementation Process
In this implementation process, we’re going to make a benchmark scenario with these parameters:
1. 1\.
Total clients to generate: 600 clients.
2. 2\.
Benchmark duration: 10 minutes.
3. 3\.
Client inters arrival phase: 1 second.
4. 4\.
300 clients will be generating in host mobile, meanwhile the other 300 clients will be generate in host posen-asik.
Before we run the program, we must generate a pair of authentication key (public and private) for passwordless authentication using SHH in order to get the Erlang nodes to communicate each other.
\[root@mobile~\]#ssh-keygen -t rsa
Enter file in which to save the key (/root/.ssh/id_rsa):
Enter passphrase (empty for no passphrase):
Enter same passphrase again:
Your identification has been saved in /root/.ssh/id_rsa.
Your public key has been saved in /root/.ssh/id_rsa.pub.
The key fingerprint is:
ec:30:2c:c9:e0:0a:92:48:3c:e5:5a:f3:7c:69:d8:92 root@mobile.myownlinux.org
\[root@mobile~\]#scp /root/.ssh/id_rsa.pub root@posen-asik:/root/.ssh/
\[root@posen-asik\]#echo .ssh/id_rsa.pub $`>`$$`>`$ .ssh/authorized_keys
After the process above, we can start the benchmark process by executing the shell script to initialize and start the application.
\[root@mobile~\]#/usr/local/bin/wii start
Several results from the important parameters to examine by the sysadmin are listed in the table below.
Table 4.1 Benchmark Results
<table>
<thead>
<tr>
<th>Parameters</th>
<th>Result</th>
</tr>
</thead>
<tbody>
<tr>
<td>Request per Second</td>
<td>3.4 requests per second</td>
</tr>
<tr>
<td>Connection per Second</td>
<td>1.8 connections per second</td>
</tr>
<tr>
<td>Page loaded per second</td>
<td>1.8 pages per second</td>
</tr>
<tr>
<td>Total user served</td>
<td>595 of 600 users</td>
</tr>
</tbody>
</table>
## 6 Conclusions
After our studies we show that in Erlang:
1. 1\.
Processes are light weight. Not only are Erlang processes light-weight, but also we can create many hundreds of thousands of such processes without noticeably degrading the performance of the system (unless of course they are all doing something at the same time),
2. 2\.
Processes share no data and the only way in which they can exchange data is by explicit message passing. Erlang message never contain pointers to data and since there is no concept of shared data, each process must work on a copy of the data that it needs. All synchronization is performed by exchanging messages.
3. 3\.
Processes scheduling operation is done by its own virtual machine.
4. 4\.
Processes are real time. Erlang is intended for programming soft real-time systems where response times in the order of milliseconds are required.
5. 5\.
Processes code has continuous operation. Erlang has primitives which allow code to be replaced in a running system and allow old and new versions of code to execute at the same time.
6. 6\.
Processes operation use automatic memory management. Memory is allocated automatically when required, and deallocated when no longer used.
7. 7\.
All interaction between processes is by asynchronous message passing. Distributed systems can easily be built.
By examine the results based on the important parameters; we hope that the syadmin and netadmin can make fine tuning to their server.
## Acknowledgment
The authors would like to thank Gunadarma University’s Research Council
## References
- \[1\]
Anonymous, Panduan Lengkap Pengembangan Jaringan Linux, Penerbit Andi, Yogyakarta.
- \[2\]
Anonymous, Web Server, http://en.wikipedia.org/wiki/Web_server (05 September 2006).
- \[3\]
Anonymous, An Erlang Course, http://www.erlang.org/download/course.pdf, (05 September 2006).
- \[4\]
Armstrong, Joe. et al., Concurrent Programming in Erlang, Prentice Hall, New Jersey.
- \[5\] Armstrong, Joe, Concurrency Oriented Programming in Erlang, Swedish Institute of Computer Science.
- \[6\]
Cisco System, Cisco Networking Academy Program – CCNA Modules – 640-801, http://cisco.netacad.net, Washington.
- \[7\]
Dudy Rudianto, Desain dan Implementasi Sistem Operasi Linux, Elex Media Komputindo, Jakarta, 2002.
- \[8\]
Dudy Rudianto, PERL Untuk Pemula, Elex Media Komputindo, Jakarta.
- \[9\]
Erricson Tellecomunications Systems Laboratories Team, Erlang Reference Manual, Erricsson Inc., Sweden.
- \[10\]
Erricson Tellecomunications Systems Laboratories Team, Getting Started With Erlang, Erricsson Inc., Sweden.
- \[11\]
Hedqvist, Pedka, A Parallel and Multhitreaded Erlang Implementation, Uppsala University, Sweden.
{biography}
Achmad B. Mutiara (M’05) was born in Jakrata, in 1967. He received the B.Sc. degree in Informatics Engineering from Gunadarma University, Depok, in 1991, and the M.S. degree and Ph.D degree in computational material physics from Goettingen University, Germany, in 1997 and 2000.
{biography}
T.A. Sabastian He received the B.Sc. degree in Informatics Engineering from Gunadarma University, Depok, in 2007. |
0707.3717 | cs.NI | cyber | INSTITUT NATIONAL DE RECHERCHE EN INFORMATIQUE ET EN AUTOMATIQUE
GCP: Gossip-based Code Propagation for Large-scale Mobile Wireless Sensor Networks
Yann Busnel — Marin Bertier — Eric Fleury — Anne-Marie KermarrecN° 6251
Juin 2007
GCP: Gossip-based Code Propagation for Large-scale Mobile Wireless Sensor Networks
Yann Busnel
<sup>†</sup>
<sup>†</sup>thanks: IRISA / UR1 - ENS Cachan
, Marin Bertier
<sup>†</sup>
<sup>†</sup>thanks: IRISA / INSA Rennes
, Eric Fleury
<sup>†</sup>
<sup>†</sup>thanks: CITI / INSA Lyon
, Anne-Marie Kermarrec
<sup>†</sup>
<sup>†</sup>thanks: IRISA / INRIA Rennes
Thème NUM — Systèmes numériques
Projets Asap et Ares
Rapport de recherche n° 6251 — Juin 2007 — ?? pages
Abstract: Wireless sensor networks (WSN) have recently received an increasing interest. They are now expected to be deployed for long periods of time, thus requiring software updates. Updating the software code automatically on a huge number of sensors is a tremendous task, as ”by hand” updates can obviously not be considered, especially when all participating sensors are embedded on mobile entities.
In this paper, we investigate an approach to automatically update software in mobile sensor-based application when no localization mechanism is available. We leverage the peer-to-peer cooperation paradigm to achieve a good trade-off between reliability and scalability of code propagation. More specifically, we present the design and evaluation of GCP (*G*ossip-based Code Propagation), a distributed software update algorithm for mobile wireless sensor networks. GCP relies on two different mechanisms (piggy-backing and forwarding control) to improve significantly the load balance without sacrificing on the propagation speed. We compare GCP against traditional dissemination approaches. Simulation results based on both synthetic and realistic workloads show that GCP achieves a good convergence speed while balancing the load evenly between sensors.
Key-words: Wireless sensor network, mobile computing, large scale, diffusion, software update, peer-to-peer algorithm, simulation
GCP: Mise à jour épidémique de logiciels pour les réseaux de capteurs mobiles, larges échelles
Résumé : GCP est un systeme de mise à jour de code automatique pour réseaux de capteurs mobiles, utilisant le concept de diffusion épidémique développés dans le cadre de réseaux filaires.
Ce rapport présente la conception et l’évaluation de GCP (Gossip-based Code Propagation), protocole proposé dans le cadre de ces travaux de recherches.
Celui-ci est évalué par comparaison avec des algorithmes traditionnels de dissemination de données. Les résultats de simulations sont fondés à la fois sur des traces réelles et générées, permettant de montrer l’efficacité de GCP tant dans la vitesse de propagation que dans l’équilibrage des charges sur le réseau.
Mots-clés : réseaux de capteurs, informatique mobile, large échelle, diffusion, mise à jour de code, systèmes pair-à-pair, simulation
## 1 Introduction
Recently, compact devices, called micro-electro mechanical systems (MEMS), have appeared. Such devices combine small size, low cost, adaptability, low power consumption, large scale and self-organization. Equipped with wireless communication capability, such appliances (called mote node or
sensor
<sup>1</sup>
<sup>1</sup>
1
We will called these devices sensor in the remaining of the paper
) together form a wireless sensor network (WSN). Due to their tiny size, sensors possess slim resources in term of memory, CPU, energy, *e*tc. \[[1](#bib.bib1), [16](#bib.bib16)\].
The increasing interest in WSNs is fundamentally due to their reliability, accuracy, flexibility, cost effectiveness and ease of deployment characteristics. Such WSNs can be deployed for monitoring purposes for example. For example we can cite Ecosystem monitoring, Military (battlefield surveillance, enemy tracking, …), Biomedical and health monitoring (cancer detector, artificial retina, organ monitor, …), Home (childhood education, smart home/office environment, …).
Sensors may be deployed both in static and dynamic environments. They are usually deployed for a long period of time, during which the software may require updates. While efficient solutions to software update may be deployed in fixed WSNs, this is far more complex when sensors are embedded on mobile entities such as people. In this paper, we consider this latter setting, namely a WSN deployed over a group of people.
Given the potentially large number of participating sensors in WSN and their limited resources, it is crucial to use fully decentralized solutions and to balance the load as evenly as possible between participating sensors. In that context, we investigate the use of the P2P communication paradigm which turns out to be a relevant candidate in this context. Considering similarities between these two systems, Section [2](#S2 "2 Applying P2P algorithms to mobile WSN") investigates the relevance of adapting the P2P paradigm to mobile WSN. Using epidemic-based dissemination, we introduce a greedy protocol (GCP: Gossip-based Code Propagation) balancing the dissemination load without increasing diffusion time. GCP relies on *piggy-backing* to save up energy and *forwarding control* to balance the load among the nodes.
Code propagation (or reprogramming service) has a lot in common with broadcast and data dissemination \[[5](#bib.bib5), [8](#bib.bib8)\] with an additional main-constraint. In broadcast, each message sent before a node arrival can be ignored by this node. In software update protocols, each new node in the network has to be informed of the existence of the software’s latest version as soon as possible to be operational.
This paper is organized as follows. Section [2](#S2 "2 Applying P2P algorithms to mobile WSN") presents the P2P paradigm and more specifically the epidemic principle. Section [3](#S3 "3 GCP: Introducing control in flooding as a miracle drug to mobile WSNs") introduces the GCP algorithm and alternative approaches. We compared GCP with classical approaches by simulation and depict the results in Section [4](#S4 "4 Simulation results"). Finally, Section [5](#S5 "5 Related works") introduces a short state of art of the different domains cited above before concluding in Section [6](#S6 "6 Conclusion").
## 2 Applying P2P algorithms to mobile WSN
Classical peer-to-peer systems are composed of millions of Personal Computers connected together by a wired network and as opposed to sensor networks are not limited by node capacity, communication range or system size (see Table [1](#S2.T1 "Table 1 ‣ 2 Applying P2P algorithms to mobile WSN")). In this section we promote the idea that sensor networks and peer-to-peer systems are similar enough so that P2P solutions can be seriously considered in the context of WSNs.
<table>
<tbody>
<tr>
<td></td>
<td>Peer to peer systems</td>
<td>Sensor networks</td>
</tr>
<tr>
<td colspan="3">Similarities</td>
</tr>
<tr>
<td>System size</td>
<td>Millions</td>
<td>Thousands</td>
</tr>
<tr>
<td>Dynamicity</td>
<td>Connection/disconnection</td>
<td>Mobility</td>
</tr>
<tr>
<td></td>
<td>Failure</td>
<td>Failure (low power)</td>
</tr>
<tr>
<td colspan="3">Differences</td>
</tr>
<tr>
<td>Resource</td>
<td>Plentiful</td>
<td>Tiny</td>
</tr>
<tr>
<td>Potential neighbourhood</td>
<td>Chosen among the whole system</td>
<td>Impose</td>
</tr>
<tr>
<td>Connectivity</td>
<td>Persistent</td>
<td>Temporary</td>
</tr>
</tbody>
</table>
Table 1: Similarity and divergence between P2P systems and WSNs
### 2.1 Peer-to-Peer vs. mobile WSN
On one hand, sensors and personal computers have incomparable resources: PCs have large resources in terms of CPU, storage while sensors are very limited; sensors have strong energy constraints, limiting their capability to communicate. On the other hand, due to the node’s resources compared to the size of the system, no entity is able to manage the entire network in both systems. Every application designed for both networks requires a strong cooperation between entities to be able to manage the network and to take advantage of the entire network. Peer to peer solutions heavily rely on such collaboration.
The peer to peer communication paradigm has been clearly identified as a key to scalability in wired systems. In a P2P system, each node may act both as a client and a server, and knows only few other nodes. Each node is logically connected to a subset of participating nodes forming a logical overlay over the physical network. With this local knowledge, the resource aggregation and the load
<sup>2</sup>
<sup>2</sup>2The load is composed of forwarding messages, storing data, *etc*.
are evenly balanced between all peers in the systems. Central points of failures disappear as well as associated performance bottlenecks \[[3](#bib.bib3), [21](#bib.bib21)\].
In a sensor network, a node is able to communicate only with a subset of the network within its communication range and has to opportunity to “choose”its neighbours. In addition, in a mobile WSN, the neighbourhood of a node changes according to its mobility pattern. With the exception of the sensor’s energy constraint, the other differences between this two systems which have a direct impact on the algorithm behaviour is the multicast advantage of the wireless medium. When a sensor node sends a message, this message can be received by every node in its direct neighbourhood while in a wired network a message is received only by the nodes which are explicitly designated in the message.
If solutions designed for a large scale wired network can not be applied directly in a sensor network, the peer-to-peer paradigm is implicitly used in sensor networks.
### 2.2 Epidemic algorithms
Epidemic or gossip-based communication is well-known to provide a simple scalable efficient and reliable way to disseminate information \[[10](#bib.bib10)\]. Epidemic protocols are based on continuous information exchange between nodes. Periodically, each node in the system chooses randomly a node in its neighbourhood to exchange information about itself or its neighbourhood. One of the key results is that in a random graph, if the node’s view (set of knowing nodes, i.e. the local view of the system) is constraint to $`c = {O\hspace{0pt}{({\log{(n)}})}}`$, where $`n`$ is the size of the network, it assures that every node in the system receive a broadcast message with a probability of $`e^{- e^{- c}}`$.
Based on unstructured P2P overlay
<sup>3</sup>
<sup>3</sup>3The logical layer on top of the physical one is not constrained to a define structure.
, gossip-based protocols can be successfully applied in WSNs. Recently, several approaches based on gossip have been proposed in the context of WSNs \[[12](#bib.bib12), [13](#bib.bib13), [19](#bib.bib19), [20](#bib.bib20)\].
The objective of this paper is to adapt such an approach to achieve efficient and reliable software dissemination in mobile WSN.
## 3 GCP: Introducing control in flooding as a miracle drug to mobile WSNs
In mobile wireless sensor networks, routing and broadcasting is a challenging task due to the network dynamicity. To the best of our knowledge, existing approaches do not deal with diffusion persistence (*cf*. Section [5](#S5 "5 Related works")).
In the following, we consider a distributed system consisting of a finite set of mobile sensor nodes which are not aware of their geographic location. The network may not be connected at any time as at a time $`t`$, a node can only communicate with nodes in its communication range. However, we consider that over an application duration, given that there are an infinite number of paths between two nodes, the network is eventually connected. In order to discover its neighbourhood, each node periodically
<sup>4</sup>
<sup>4</sup>4This period is a parameter of the system
broadcasts locally Hello messages, called beacon.
### 3.1 GCP design
Flooding paradigm is a simple way to disseminate informations. Used commonly in the network area, it consists in forwarding to everyone known a new received information. Rather than applying classical flooding algorithms having an ideal speed propagation at the price of a high energy consumption, we use the epidemic communication paradigm, proposed in the context of P2P systems. To this end, GCP is inspired from the flooding paradigm enhanced through*Piggy-Backing* and *Forwarding Control*.
#### Piggy-Backing mechanism
In order to avoid unnecessary software transmissions, nodes have to be aware of the software versions hold by their neighbours. To this end, each node simply piggy-backs its own version number into beacon messages.
#### Forwarding Control mechanism
In order to balance the load among node and increase the overall lifetime of a system, each node sends its current software version a limited number of times. To this end, each node owns a given number of tokens, which value is a system parameter. Sending a software update is worth a token. When a node has spent all its tokens, it is not allowed to send this version of the software. This number of tokens is associated to each version. Upon receiving a new version of the software, the number of tokens is set to the default initial value. This mimics the behaviour of an epidemic protocol, where each node sends a predefined number of time a message (typically $`\log{(N)}`$, $`N`$ being the size of the system) \[[10](#bib.bib10)\]. Likewise, the default value of the number of tokens can be set according the order of magnitude of WSN size.
### 3.2 GCP algorithm
[Refer to caption](/html/0707.3717/assets/x1.png)
Figure 1: Summarized behaviour of GCP
Figure [1](#S3.F1 "Figure 1 ‣ 3.2 GCP algorithm ‣ 3 GCP: Introducing control in flooding as a miracle drug to mobile WSNs") represents the three possible different cases and the GCP behaviour.
Each node in the transmission range of $`a`$ receives the beacon (In Figure [1](#S3.F1 "Figure 1 ‣ 3.2 GCP algorithm ‣ 3 GCP: Introducing control in flooding as a miracle drug to mobile WSNs"): nodes $`b`$, $`c`$ and $`d`$; node $`e`$ is out of range). A beacon message received by node $`b`$ is processed as follows:
1. 2a.
If $`b`$ owns the same version as $`a`$ ($`v_{a} = v_{b}`$) due to the piggy-backing mechanism or $`{t\hspace{0pt}o\hspace{0pt}k\hspace{0pt}e\hspace{0pt}n_{b}} = 0`$ due to the forwarding control mechanism, then no action is required.
2. 2b.
If $`b`$ owns a more recent version than $`a`$ ($`v_{a} < v_{b}`$) due to the piggy-backing mechanism, and, if $`b`$ still holds some tokens ($`{t\hspace{0pt}o\hspace{0pt}k\hspace{0pt}e\hspace{0pt}n_{b}} > 0`$) due to the forwarding control mechanism, it sends its own version to $`a`$ thus consuming a token ($`t\hspace{0pt}o\hspace{0pt}k\hspace{0pt}e\hspace{0pt}n_{b}`$–). Note that if other nodes, within the transmission range, holding an older version than $`b`$’s, they leverage the software update and update their own version (“free update”).
3. 2c.
If $`a`$ owns a version newer than $`b`$ ($`v_{a} > v_{b}`$) due to the piggy-backing mechanism, the node $`b`$ sends immediately a beacon message in order to request a software update from $`a`$ while $`a`$ is still in its radio range.
### 3.3 Alternative algorithms
In order to assess the efficiency of GCP, we compare it against three other protocols, directly derived from wired networks. We briefly present those protocols in this section.
The Flooding Protocol (FP) Each time a node receives a beacon from another one, it sends its own version of the software, whether the node needs it or not. This algorithm obviously leads to load unbalance and does not take into account energy consumption. This algorithm is presented here because it provides good software propagation speed.
The Forwarding Control Protocol (FCP) This algorithm is an enhancement of the flooding protocol, using the forwarding control mechanism.
The Piggy-Backing Protocol (PBP) This last algorithm is an enhancement of the flooding algorithm, with the piggy-backing mechanism.
### 3.4 Protocols in Pseudo-Code
Thereafter, we introduce a formalized version of these protocols.
- •
$`v\hspace{0pt}e\hspace{0pt}r\hspace{0pt}s\hspace{0pt}i\hspace{0pt}o\hspace{0pt}n`$ and $`v\hspace{0pt}e\hspace{0pt}r\hspace{0pt}s\hspace{0pt}i\hspace{0pt}o\hspace{0pt}n_{r}`$ represent respectively the local and the remote version number of the software;
- •
$`s\hspace{0pt}o\hspace{0pt}f\hspace{0pt}t\hspace{0pt}w\hspace{0pt}a\hspace{0pt}r\hspace{0pt}e`$ and $`s\hspace{0pt}o\hspace{0pt}f\hspace{0pt}t\hspace{0pt}w\hspace{0pt}a\hspace{0pt}r\hspace{0pt}e_{r}`$ represent respectively the local and the remote binary of the software;
- •
$`t\hspace{0pt}o\hspace{0pt}k\hspace{0pt}e\hspace{0pt}n`$ represents the remaining number of tokens available on the local node;
- •
$`i\hspace{0pt}n\hspace{0pt}i\hspace{0pt}t\hspace{0pt}i\hspace{0pt}a\hspace{0pt}l\hspace{0pt}N\hspace{0pt}u\hspace{0pt}m\hspace{0pt}b\hspace{0pt}e\hspace{0pt}r\hspace{0pt}O\hspace{0pt}f\hspace{0pt}T\hspace{0pt}o\hspace{0pt}k\hspace{0pt}e\hspace{0pt}n\hspace{0pt}s`$ represents the initial number of tokens available on the local node.
FP is presented in Figure [2](#S3.F2 "Figure 2 ‣ 3.4 Protocols in Pseudo-Code ‣ 3 GCP: Introducing control in flooding as a miracle drug to mobile WSNs").
ReceiveBeacon$`(\text{nil})`$
1
$`\text{send~}\hspace{0pt}{({s\hspace{0pt}o\hspace{0pt}f\hspace{0pt}t\hspace{0pt}w\hspace{0pt}a\hspace{0pt}r\hspace{0pt}e},{v\hspace{0pt}e\hspace{0pt}r\hspace{0pt}s\hspace{0pt}i\hspace{0pt}o\hspace{0pt}n})}`$
ReceiveSoftware$`({s\hspace{0pt}o\hspace{0pt}f\hspace{0pt}t\hspace{0pt}w\hspace{0pt}a\hspace{0pt}r\hspace{0pt}e_{r}},{v\hspace{0pt}e\hspace{0pt}r\hspace{0pt}s\hspace{0pt}i\hspace{0pt}o\hspace{0pt}n_{r}})`$
1
if
$`{v\hspace{0pt}e\hspace{0pt}r\hspace{0pt}s\hspace{0pt}i\hspace{0pt}o\hspace{0pt}n_{r}} > {v\hspace{0pt}e\hspace{0pt}r\hspace{0pt}s\hspace{0pt}i\hspace{0pt}o\hspace{0pt}n}`$
2
then
3
$`\text{updateCode~}\hspace{0pt}{({s\hspace{0pt}o\hspace{0pt}f\hspace{0pt}t\hspace{0pt}w\hspace{0pt}a\hspace{0pt}r\hspace{0pt}e_{r}})}`$
4
$`{v\hspace{0pt}e\hspace{0pt}r\hspace{0pt}s\hspace{0pt}i\hspace{0pt}o\hspace{0pt}n}\leftarrow{v\hspace{0pt}e\hspace{0pt}r\hspace{0pt}s\hspace{0pt}i\hspace{0pt}o\hspace{0pt}n_{r}}`$
Figure 2: Flooding algorithm
The GCP algorithm is presented in Figure [3](#S3.F3 "Figure 3 ‣ 3.4 Protocols in Pseudo-Code ‣ 3 GCP: Introducing control in flooding as a miracle drug to mobile WSNs"). PBP has the same pseudo code without lines 2, 5 in ReceiveBeacon and without line 5 in ReceiveSoftware. FCP has the same pseudo code without lines 1, 6, 7 and 8 in ReceiveBeacon.
Each time a node receives a beacon message, it compares its own version number with the remote one. If it possesses a newer version, it sends its code to its neighbourhood. Otherwise, if it owns an older version, it sends immediately a beacon, to pull the newer version from the remote node.
When a node receives a software version, it checks this version number. If this version is a newer one, it replaces its own version by the new one and reinitializes its token number to be able to forward it.
ReceiveBeacon$`({v\hspace{0pt}e\hspace{0pt}r\hspace{0pt}s\hspace{0pt}i\hspace{0pt}o\hspace{0pt}n_{r}})`$
1
if
$`{v\hspace{0pt}e\hspace{0pt}r\hspace{0pt}s\hspace{0pt}i\hspace{0pt}o\hspace{0pt}n_{r}} < {v\hspace{0pt}e\hspace{0pt}r\hspace{0pt}s\hspace{0pt}i\hspace{0pt}o\hspace{0pt}n}`$
2
$`{\text{~and~}\hspace{0pt}t\hspace{0pt}o\hspace{0pt}k\hspace{0pt}e\hspace{0pt}n} > 0`$
3
then
4
$`\text{send~}\hspace{0pt}{({s\hspace{0pt}o\hspace{0pt}f\hspace{0pt}t\hspace{0pt}w\hspace{0pt}a\hspace{0pt}r\hspace{0pt}e},{v\hspace{0pt}e\hspace{0pt}r\hspace{0pt}s\hspace{0pt}i\hspace{0pt}o\hspace{0pt}n})}`$
5
$`{t\hspace{0pt}o\hspace{0pt}k\hspace{0pt}e\hspace{0pt}n}\leftarrow{{t\hspace{0pt}o\hspace{0pt}k\hspace{0pt}e\hspace{0pt}n} - 1}`$
6
else
if
$`{v\hspace{0pt}e\hspace{0pt}r\hspace{0pt}s\hspace{0pt}i\hspace{0pt}o\hspace{0pt}n_{r}} > {v\hspace{0pt}e\hspace{0pt}r\hspace{0pt}s\hspace{0pt}i\hspace{0pt}o\hspace{0pt}n}`$
7
then
8
$`\text{send~}\hspace{0pt}{Beacon}\hspace{0pt}{({v\hspace{0pt}e\hspace{0pt}r\hspace{0pt}s\hspace{0pt}i\hspace{0pt}o\hspace{0pt}n})}`$
ReceiveSoftware$`({s\hspace{0pt}o\hspace{0pt}f\hspace{0pt}t\hspace{0pt}w\hspace{0pt}a\hspace{0pt}r\hspace{0pt}e_{r}},{v\hspace{0pt}e\hspace{0pt}r\hspace{0pt}s\hspace{0pt}i\hspace{0pt}o\hspace{0pt}n_{r}})`$
1
if
$`{v\hspace{0pt}e\hspace{0pt}r\hspace{0pt}s\hspace{0pt}i\hspace{0pt}o\hspace{0pt}n_{r}} > {v\hspace{0pt}e\hspace{0pt}r\hspace{0pt}s\hspace{0pt}i\hspace{0pt}o\hspace{0pt}n}`$
2
then
3
$`\text{updateCode~}\hspace{0pt}{({s\hspace{0pt}o\hspace{0pt}f\hspace{0pt}t\hspace{0pt}w\hspace{0pt}a\hspace{0pt}r\hspace{0pt}e_{r}})}`$
4
$`{v\hspace{0pt}e\hspace{0pt}r\hspace{0pt}s\hspace{0pt}i\hspace{0pt}o\hspace{0pt}n}\leftarrow{v\hspace{0pt}e\hspace{0pt}r\hspace{0pt}s\hspace{0pt}i\hspace{0pt}o\hspace{0pt}n_{r}}`$
5
$`{t\hspace{0pt}o\hspace{0pt}k\hspace{0pt}e\hspace{0pt}n}\leftarrow{i\hspace{0pt}n\hspace{0pt}i\hspace{0pt}t\hspace{0pt}i\hspace{0pt}a\hspace{0pt}l\hspace{0pt}N\hspace{0pt}u\hspace{0pt}m\hspace{0pt}b\hspace{0pt}e\hspace{0pt}r\hspace{0pt}O\hspace{0pt}f\hspace{0pt}T\hspace{0pt}o\hspace{0pt}k\hspace{0pt}e\hspace{0pt}n\hspace{0pt}s}`$
Figure 3: Gossip-based Code Propagation algorithm
### 3.5 Theoretical analysis
We compare these protocols along two metrics: the software propagation speed and the number of versions sent by each node. This last metric reflects the load balance in the system. Some notations are needed and listed in the following paragraph:
- •
$`n_{v}`$ represents the number of upgrades during the WSN lifetime (how many times the software need to be updated);
- •
$`t`$ represents the number of tokens available on each node;
- •
$`n_{n\hspace{0pt}h}`$ represents the average size of a node neighbourhood;
- •
$`d`$ represents the duration of the experimentation;
- •
$`p_{b}`$ represents the period of beacon emission;
- •
$`n_{s}`$ represents the network size in terms of number of nodes;
For all of the following theoretical results, each equation presents the upper bound average values. To obtain more precise results, we may take into account the diameter size, the topology and mobility model of the WSN considered.
#### Load balancing
For the load balancing, the analysis extracts the quantity of large message sent by each node, as software binary sending here. The following equations present the average number of software sent by one node during the whole deployment of the WSN for each algorithm:
<table>
<tbody>
<tr>
<td></td>
<td>Flooding algorithm:</td>
<td>∼similar-to\displaystyle\sim</td>
<td>dpb×nnh𝑑subscript𝑝𝑏subscript𝑛𝑛ℎ\displaystyle\frac{d}{p_{b}}\times n_{nh}</td>
<td></td>
<td>(1)</td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td>FCP algorithm:(lbFCP)FCP algorithm:𝑙subscript𝑏𝐹𝐶𝑃\displaystyle\textrm{FCP algorithm:}(lb_{FCP})</td>
<td>∼similar-to\displaystyle\sim</td>
<td>(nv+1)×tsubscript𝑛𝑣1𝑡\displaystyle(n_{v}+1)\times t</td>
<td></td>
<td>(2)</td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td>PBP algorithm:(lbPBP)PBP algorithm:𝑙subscript𝑏𝑃𝐵𝑃\displaystyle\textrm{PBP algorithm:}(lb_{PBP})</td>
<td>∼similar-to\displaystyle\sim</td>
<td>nv×(ns−1)subscript𝑛𝑣𝑛𝑠1\displaystyle n_{v}\times(ns-1)</td>
<td></td>
<td>(3)</td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td>GCP algorithm:(lbGCP)GCP algorithm:𝑙subscript𝑏𝐺𝐶𝑃\displaystyle\textrm{GCP algorithm:}(lb_{GCP})</td>
<td>∼similar-to\displaystyle\sim</td>
<td>nv×tsubscript𝑛𝑣𝑡\displaystyle n_{v}\times t</td>
<td></td>
<td>(4)</td>
</tr>
</tbody>
</table>
In the flooding algorithm, at each beacon reception, a node sends its own version. Equation ([1](#S3.E1 "In Load balancing ‣ 3.5 Theoretical analysis ‣ 3 GCP: Introducing control in flooding as a miracle drug to mobile WSNs")) illustrates that the number of time a node sends its version is equivalent to the number of beacon sent by each node multiplied by the average number of nodes in the neighbourhood.
Both the FCP and GCP algorithms have a bound on the quantity of time a node can send its software according to $`n_{v}`$ and $`t`$. For FCP, as the version number is not communicated to neighbourhood, the first version is sent as a new one by each node. Moreover, in the FCP algorithm, tokens may be spent unnecessarily, as opposed to GCP consuming only necessary tokens thanks to the version number contained in beacon messages. We assume that $`n_{s} \gg t`$ to optimize balancing for token rule using algorithms in a large scale environment.
We can conclude that in the worst configuration:
<table>
<tbody>
<tr>
<td></td>
<td>lbFP≫lbPBP≫lbFCP>lbGCPmuch-greater-than𝑙subscript𝑏𝐹𝑃𝑙subscript𝑏𝑃𝐵𝑃much-greater-than𝑙subscript𝑏𝐹𝐶𝑃𝑙subscript𝑏𝐺𝐶𝑃lb_{FP}\gg lb_{PBP}\gg lb_{FCP}>lb_{GCP}</td>
<td></td>
</tr>
</tbody>
</table>
and in most cases:
<table>
<tbody>
<tr>
<td></td>
<td>lbFP≫lbPBP>lbFCP>lbGCPmuch-greater-than𝑙subscript𝑏𝐹𝑃𝑙subscript𝑏𝑃𝐵𝑃𝑙subscript𝑏𝐹𝐶𝑃𝑙subscript𝑏𝐺𝐶𝑃lb_{FP}\gg lb_{PBP}>lb_{FCP}>lb_{GCP}</td>
<td></td>
</tr>
</tbody>
</table>
#### Propagation speed
Considering the propagation speed, the analysis extracts the propagation of a new software in the WSN. We consider the software update propagation by dividing time into software forwarding step.
We start by comparing Flooding with GCP and PBP. Let $`t`$ be the current step of propagation. Consider that at this step, the network is in the same state (same nodes own the newest version of the protocol, same positions of nodes and future moving, *etc*.) Obviously, using flooding software is equivalent to forward the software to all nodes, regardless of their need for it. At the opposite, in GCP and PBP, a node sends the newest software version only to nodes needing it. They exhibit a similar propagation speed. GCP and PBP however provides software updates “for free”. Effectively, if a node is located in the transmission range of the newest software, it will update its software without requesting it explicitly. By recurrence on $`t`$, Flooding algorithm propagation speed ($`p\hspace{0pt}s_{F\hspace{0pt}P}`$) is greater or equal to GCP and PBP ones (respectively $`p\hspace{0pt}s_{G\hspace{0pt}C\hspace{0pt}P}`$ and $`p\hspace{0pt}s_{P\hspace{0pt}B\hspace{0pt}P}`$).
Considering PBP and GCP, they are equivalent in the most common case. But, PBP may have slightly better speed propagation in case of the meeting set is unbalanced (one node is transmit the newest version most of the time *i.e.* this node is consuming a large part of its power). The GCP propagation speed can be slowed down as this potential node does not have enough token. However, this case does not balance the load among the network with PBP, which is one of our objective.
We are now considering GCP and FCP algorithms. By choosing an appropriate initial number of tokens, the propagation speed of these two algorithms can approach the ideal one. Furthermore, most of the time, using FCP algorithm could waste a non-negligible number of tokens by sending unnecessarily the software. The FCP algorithm propagation speed ($`p\hspace{0pt}s_{F\hspace{0pt}C\hspace{0pt}P}`$) can often be strictly lower of the GCP one.
So, we can conclude that in the worst configuration:
<table>
<tbody>
<tr>
<td></td>
<td>spFP≥spPBP≥spGCP≥spFCP𝑠subscript𝑝𝐹𝑃𝑠subscript𝑝𝑃𝐵𝑃𝑠subscript𝑝𝐺𝐶𝑃𝑠subscript𝑝𝐹𝐶𝑃sp_{FP}\geq sp_{PBP}\geq sp_{GCP}\geq sp_{FCP}</td>
<td></td>
</tr>
</tbody>
</table>
and most of the case:
<table>
<tbody>
<tr>
<td></td>
<td>spFP≥spPBP∼spGCP≥spFCP𝑠subscript𝑝𝐹𝑃𝑠subscript𝑝𝑃𝐵𝑃similar-to𝑠subscript𝑝𝐺𝐶𝑃𝑠subscript𝑝𝐹𝐶𝑃sp_{FP}\geq sp_{PBP}\sim sp_{GCP}\geq sp_{FCP}</td>
<td></td>
</tr>
</tbody>
</table>
### 3.6 Summary
Figure [4](#S3.F4 "Figure 4 ‣ 3.6 Summary ‣ 3 GCP: Introducing control in flooding as a miracle drug to mobile WSNs") presents the relationship between each of the previous proposed algorithm. It summarizes how to transform one algorithm to obtain another one.
<table>
<tbody>
<tr>
<td></td>
<td></td>
<td>FCP algorithm</td>
<td></td>
<td></td>
</tr>
<tr>
<td></td>
<td>EGG GGCRemove version in beaconAdd version in beaconsubscriptsuperscriptEGG GGCAdd version in beaconRemove version in beacon\mathop{\vbox{\hbox{\ooalign{\raise 1.37775pt\hbox{$\hbox to94.28914pt{$\mathord{\textrm{\char 69\relax}}\mkern-6.0mu\cleaders\hbox{$\mkern-2.0mu\mathord{\textrm{\char 71\relax}}\mkern-2.0mu$}\hfill\mkern-6.0mu\mathord{\textrm{\char 71\relax}}$\hskip 0.40002pt}$}\crcr$\hbox to94.28914pt{\hskip 0.40002pt$\mathord{\textrm{\char 71\relax}}\mkern-6.0mu\cleaders\hbox{$\mkern-2.0mu\mathord{\textrm{\char 71\relax}}\mkern-2.0mu$}\hfill\mkern-6.0mu\mathord{\textrm{\char 67\relax}}$}$}}}}\limits^{\hbox{Add version in beacon}}_{\hbox{Remove version in beacon}}</td>
<td></td>
<td>EGG GGCRemove token numberAdd token numbersubscriptsuperscriptEGG GGCAdd token numberRemove token number\mathop{\vbox{\hbox{\ooalign{\raise 1.37775pt\hbox{$\hbox to81.8002pt{$\mathord{\textrm{\char 69\relax}}\mkern-6.0mu\cleaders\hbox{$\mkern-2.0mu\mathord{\textrm{\char 71\relax}}\mkern-2.0mu$}\hfill\mkern-6.0mu\mathord{\textrm{\char 71\relax}}$\hskip 0.40002pt}$}\crcr$\hbox to81.8002pt{\hskip 0.40002pt$\mathord{\textrm{\char 71\relax}}\mkern-6.0mu\cleaders\hbox{$\mkern-2.0mu\mathord{\textrm{\char 71\relax}}\mkern-2.0mu$}\hfill\mkern-6.0mu\mathord{\textrm{\char 67\relax}}$}$}}}}\limits^{\hbox{Add token number}}_{\hbox{Remove token number}}</td>
<td></td>
</tr>
<tr>
<td>GCP algorithm</td>
<td></td>
<td></td>
<td></td>
<td>Flooding algorithm</td>
</tr>
<tr>
<td></td>
<td>EGG GGCRemove token numberAdd token numbersubscriptsuperscriptEGG GGCAdd token numberRemove token number\mathop{\vbox{\hbox{\ooalign{\raise 1.37775pt\hbox{$\hbox to81.8002pt{$\mathord{\textrm{\char 69\relax}}\mkern-6.0mu\cleaders\hbox{$\mkern-2.0mu\mathord{\textrm{\char 71\relax}}\mkern-2.0mu$}\hfill\mkern-6.0mu\mathord{\textrm{\char 71\relax}}$\hskip 0.40002pt}$}\crcr$\hbox to81.8002pt{\hskip 0.40002pt$\mathord{\textrm{\char 71\relax}}\mkern-6.0mu\cleaders\hbox{$\mkern-2.0mu\mathord{\textrm{\char 71\relax}}\mkern-2.0mu$}\hfill\mkern-6.0mu\mathord{\textrm{\char 67\relax}}$}$}}}}\limits^{\hbox{Add token number}}_{\hbox{Remove token number}}</td>
<td></td>
<td>EGG GGCRemove version in beaconAdd version in beaconsubscriptsuperscriptEGG GGCAdd version in beaconRemove version in beacon\mathop{\vbox{\hbox{\ooalign{\raise 1.37775pt\hbox{$\hbox to94.28914pt{$\mathord{\textrm{\char 69\relax}}\mkern-6.0mu\cleaders\hbox{$\mkern-2.0mu\mathord{\textrm{\char 71\relax}}\mkern-2.0mu$}\hfill\mkern-6.0mu\mathord{\textrm{\char 71\relax}}$\hskip 0.40002pt}$}\crcr$\hbox to94.28914pt{\hskip 0.40002pt$\mathord{\textrm{\char 71\relax}}\mkern-6.0mu\cleaders\hbox{$\mkern-2.0mu\mathord{\textrm{\char 71\relax}}\mkern-2.0mu$}\hfill\mkern-6.0mu\mathord{\textrm{\char 67\relax}}$}$}}}}\limits^{\hbox{Add version in beacon}}_{\hbox{Remove version in beacon}}</td>
<td></td>
</tr>
<tr>
<td></td>
<td></td>
<td>PBP algorithm</td>
<td></td>
<td></td>
</tr>
</tbody>
</table>
Figure 4: Algorithms dependence
Table [2](#S3.T2 "Table 2 ‣ 3.6 Summary ‣ 3 GCP: Introducing control in flooding as a miracle drug to mobile WSNs") summarizes the theoretical characteristics of the studied protocols. The additional local information consists in the token value included in GCP and FCP, and the beacon one is the version number included in GCP and PBP. The two latest columns are some assumptions according to load balancing and propagation speed efficiency.
<table>
<tbody>
<tr>
<td></td>
<td>No local</td>
<td>No beacon</td>
<td>Load</td>
<td>Propagation</td>
</tr>
<tr>
<td></td>
<td>additional</td>
<td>additional</td>
<td>balancing</td>
<td>speed</td>
</tr>
<tr>
<td></td>
<td>information</td>
<td>information</td>
<td></td>
<td></td>
</tr>
<tr>
<td>Flooding Algorithm</td>
<td>+</td>
<td>+</td>
<td>– –</td>
<td>+ +</td>
</tr>
<tr>
<td>FCP Algorithm</td>
<td>–</td>
<td>+</td>
<td>+</td>
<td>–</td>
</tr>
<tr>
<td>PBP Algorithm</td>
<td>+</td>
<td>–</td>
<td>–</td>
<td>+ +</td>
</tr>
<tr>
<td>GCP Algorithm</td>
<td>–</td>
<td>–</td>
<td>+ +</td>
<td>+ +</td>
</tr>
</tbody>
</table>
Table 2: Algorithms comparative statement
Brown and Sreenan \[[2](#bib.bib2)\] proposed a model to compare software update algorithms in WSNs According to this model, GCP provides the following capabilities:
Propagation capabilities
Advertise
(of the existence of a new version) is provided by the piggy-backing of the version number in each beacon message;
Transfer/send
actions are provided by the MAC layer of the node;
Listen
is provided by receiving beacon message with their additional information;
Decide
is provided by the comparison of version number.
Activation capabilities
Verify
is provided by sending a MD5 signature before sending the new software version (if the software appears to be corrupted, the node resends a beacon to request another software copy);
Transfer/send
is provided by the MAC layer of the node;
As most of this kind of algorithm, GCP does not provide generation capabilities and high-level activation capabilities. It is possible to put GCP in the Deluge \[[7](#bib.bib7)\] equivalence class as it is providing the same capabilities according to the previous cite model \[[2](#bib.bib2)\].
<table>
<tbody>
<tr>
<td>[Uncaptioned image] Figure 5: Several cases of transmission range.</td>
<td>[Uncaptioned image] Figure 6: Sensor transmission probability according to r=3𝑟3r=3, R=5𝑅5R=5, and Pmin=0.3subscript𝑃𝑚𝑖𝑛0.3P_{min}=0.3.</td>
</tr>
</tbody>
</table>
## 4 Simulation results
### 4.1 System model
We assume that nodes can communicate only by 1-hop broadcast with nodes in their transmission range, with no collision. For a node $`S`$, we distinguish two ranges of transmission,$`r`$ and $`R`$, where $`0 < r < R`$ and $`{\mathcal{S}_{S}\hspace{0pt}{(r)}} \subset {\mathcal{S}_{S}\hspace{0pt}{(R)}}`$
<sup>5</sup>
<sup>5</sup>5$`\mathcal{S}_{S}\hspace{0pt}{(x)}`$ is the sphere notation with center $`S`$ and radius $`x`$.
(cf. Figure [6](#S3.F6 "Figure 6 ‣ 3.6 Summary ‣ 3 GCP: Introducing control in flooding as a miracle drug to mobile WSNs")). $`r`$ represents the radius where the transmission range is uniform, and thus messages sent by nodes separated by less than $`r`$ are always received (Node $`A`$ in Figure [6](#S3.F6 "Figure 6 ‣ 3.6 Summary ‣ 3 GCP: Introducing control in flooding as a miracle drug to mobile WSNs")). The second range $`R`$ represents the radius where transmission range may be not uniform. No nodes separated by more than $`R`$ can receive each other transmission (Node $`B`$ in Figure [6](#S3.F6 "Figure 6 ‣ 3.6 Summary ‣ 3 GCP: Introducing control in flooding as a miracle drug to mobile WSNs")). Thus, nodes separated by a distance between $`r`$ and $`R`$ may or not receive each other transmitted messages according to the Equation [5](#S4.E5 "In 4.1 System model ‣ 4 Simulation results") (Node $`C`$ and $`D`$ in Figure [6](#S3.F6 "Figure 6 ‣ 3.6 Summary ‣ 3 GCP: Introducing control in flooding as a miracle drug to mobile WSNs"), where the filled shape represented the transmission area of $`S`$). In Equation [5](#S4.E5 "In 4.1 System model ‣ 4 Simulation results"), $`P_{m\hspace{0pt}i\hspace{0pt}n}`$ is the transmission’s lower bound probability parameter for two nodes separated by $`R`$. We consider that sensor nodes have equal communication ranges. Nodes have a transmission probability defined as follows, plot in Figure [6](#S3.F6 "Figure 6 ‣ 3.6 Summary ‣ 3 GCP: Introducing control in flooding as a miracle drug to mobile WSNs"):
<table>
<tbody>
<tr>
<td></td>
<td>{1 if d<rPmin−R−dR−r⋅(R−dR−r−5)⋅1−Pmin4 if r<d<R0 if d>Rcases1 if 𝑑𝑟subscript𝑃𝑚𝑖𝑛⋅𝑅𝑑𝑅𝑟𝑅𝑑𝑅𝑟51subscript𝑃𝑚𝑖𝑛4 if 𝑟𝑑𝑅0 if 𝑑𝑅\left\{\begin{array}[]{ll}1&\textrm{ if }d<r\\ P_{min}-\sqrt{\frac{R-d}{R-r}}\cdot\left(\frac{R-d}{R-r}-5\right)\cdot\frac{1-P_{min}}{4}&\textrm{ if }r<d<R\\ 0&\textrm{ if }d>R\\ \end{array}\right.</td>
<td></td>
<td>(5)</td>
</tr>
</tbody>
</table>
In our model, we consider mobile sensors. This mobility can be treated among different mobility models. In order to compare our results with other ones in the literature, we choose, for the synthetic workload, the widely used random way point mobility model \[[9](#bib.bib9)\].
### 4.2 Simulation setup
#### Simulator
In order to evaluate GCP, we developed SeNSim, a software implemented for mobile wireless sensor-based applications’ simulation. SeNSim is a Java software which allows the creation of mobile wireless sensor networks and analyses information dissemination under different mobility and failures scenarios. The simulator also allows the evaluation of the characteristics related to this protocol under different mobility, failures, and stimulus scenarios.
In order to simulate large scale sensor networks scenarios during a long period of time, the designed simulator is based on a discrete-event system. This software is composed of two different parts: (1) generation of synthetic workloads and (2) mobile WSN simulation.
<table>
<thead>
<tr>
<th> Refer to caption</th>
<th> Refer to caption</th>
</tr>
</thead>
<tbody>
<tr>
<td>(a) Clustering scenario</td>
<td>(b) Socialized clustering scenario</td>
</tr>
</tbody>
</table>
Figure 7: Examples of synthetic scenarios with four clusters
#### Workloads
We evaluate different scenarios with the same set of workloads for comparison purposes. By running different algorithms on a same persistent trace, we obtain a fair comparison between each solution introduced in Section [3](#S3 "3 GCP: Introducing control in flooding as a miracle drug to mobile WSNs").
We explore the performance of GCP in various scenarios. Such scenarios exhibit different clustering and mobility patterns. Eight synthetic and one realistic scenarios have been simulated:
#### Synthetic
- •
1 cluster scenario – One large cluster composed of 2,000 sensor nodes in a $`{{250\hspace{0pt}m} \times 250}\hspace{0pt}m`$ area.
- •
1 sparse cluster scenario – The large big cluster in a wide area ($`1,{{100\hspace{0pt}m} \times 1},{100\hspace{0pt}m}`$).
- •
2 cluster scenario – Two 1,000 nodes’ clusters in two $`{{800\hspace{0pt}m} \times 800}\hspace{0pt}m`$ area with a $`{{100\hspace{0pt}m} \times 100}\hspace{0pt}m`$ intersection
- •
2 socializing cluster scenario – Two separated 950 nodes’ clusters in two $`{{800\hspace{0pt}m} \times 800}\hspace{0pt}m`$ area. These two clusters can’t communicate with each other. We put 100 transmitters which moved in the whole area ($`2,{{000\hspace{0pt}m} \times 2},{000\hspace{0pt}m}`$) to assure the connectivity.
- •
4 cluster scenario – Four separated 500 nodes’ clusters in four $`{{550\hspace{0pt}m} \times 550}\hspace{0pt}m`$ areas. Connectivity is assured by the common border. These areas are set as in Figure [7](#S4.F7 "Figure 7 ‣ Simulator ‣ 4.2 Simulation setup ‣ 4 Simulation results")(a).
- •
4 socializing cluster scenario – Four separated 475 nodes’ clusters in four $`{{550\hspace{0pt}m} \times 550}\hspace{0pt}m`$ areas. These four clusters can’t communicate with each other. We put 100 transmitters which moved in the whole area ($`1,{{300\hspace{0pt}m} \times 1},{300\hspace{0pt}m}`$) to assure the connectivity. These areas are set as in Figure [7](#S4.F7 "Figure 7 ‣ Simulator ‣ 4.2 Simulation setup ‣ 4 Simulation results")(b), where transmitters are represented by a filled pentagon.
- •
9 cluster scenario – Nine separated 250 nodes’ clusters in four $`{{400\hspace{0pt}m} \times 400}\hspace{0pt}m`$ areas. Connectivity is assured by the common border.
- •
9 socializing cluster scenario– Nine separated 240 nodes’ clusters in four $`{{400\hspace{0pt}m} \times 400}\hspace{0pt}m`$ area. These nine clusters can’t communicate with each other. We put 90 transmitters which moved in the whole area ($`1,{{500\hspace{0pt}m} \times 1},{500\hspace{0pt}m}`$) to assure the connectivity.
#### Realistic
- •
MIT Campus scenario – In order to evaluate the performance of GCP with a realistic movement behaviour, we used the mit/reality data set \[[6](#bib.bib6)\] from CRAWDAD. This data set provides captured communication, proximity, location and activity information from 100 subjects at MIT over the course of the 2004-2005 academic year.
We have used the following parameters:
- •
The time is discretized by millisecond.
- •
Each sensor node:
- –
is initially randomly placed inside its defined area for synthetic workloads.
- –
sends a beacon periodically every 100 ms (common period encountered in the literature).
- –
has the transmission ranges set as follows: $`r = {3\hspace{0pt}m}`$ and $`R = {5\hspace{0pt}m}`$ and a minimum transmission probability inside $`R`$, $`P_{m\hspace{0pt}i\hspace{0pt}n} = 0.3`$.
- –
is mobile, following a Random Way Point strategy for synthetic workloads, with a maximum pause time of $`100\hspace{0pt}m\hspace{0pt}s`$, each movement duration between $`100`$ and $`500\hspace{0pt}m\hspace{0pt}s`$, with a speed included between $`{0.8\hspace{0pt}m}.s^{- 1}`$ ($`{2.88\hspace{0pt}k\hspace{0pt}m}.h^{- 1}`$) and $`{2\hspace{0pt}m}.s^{- 1}`$ ($`{7.2\hspace{0pt}k\hspace{0pt}m}.h^{- 1}`$) (equivalent to human walking speed). Every movement is bounded by the defined area. The border rules are defined as each node bounce back according to the bisector of incidence angle.
- –
has a number of tokens, for each code propagation, according to simulation configuration describes below.
- •
Simulation last around $`50,{000\hspace{0pt}m\hspace{0pt}s}`$.
- •
A new version is sent to a sensor picked up at random after $`1\hspace{0pt}s`$ of simulation.
In order to compare the efficiency of the four algorithms, we compared them along the following metric, previously introduced Section [3.5](#S3.SS5 "3.5 Theoretical analysis ‣ 3 GCP: Introducing control in flooding as a miracle drug to mobile WSNs"):
Code propagation speed
We observe the number of nodes owning the newest version of the software during all the simulation, and plot these values in Section [4.3](#S4.SS3 "4.3 Convergence speed ‣ 4 Simulation results") for each scenario and algorithms.
Load balancing
At the simulation termination, we extract from each node the number of times it sends the software. Results are depicted in Section [4.4](#S4.SS4 "4.4 Load Balancing ‣ 4 Simulation results") for each scenario and algorithms.
### 4.3 Convergence speed
For each simulation scenario, we have followed the propagation advancement, and extracted at each time, the number of sensors owning the newest version of the software.
Figures [8](#S4.F8 "Figure 8 ‣ 4.3 Convergence speed ‣ 4 Simulation results"), [9](#S4.F9 "Figure 9 ‣ 4.3 Convergence speed ‣ 4 Simulation results") and [10](#S4.F10 "Figure 10 ‣ 4.3 Convergence speed ‣ 4 Simulation results") present the results according to time, for three scenarios ordered as above:
1. 1\.
9 clusters scenario organized in the same way than in Figure [7](#S4.F7 "Figure 7 ‣ Simulator ‣ 4.2 Simulation setup ‣ 4 Simulation results")(a);
2. 2\.
9 socializing clusters scenario organized in the same way than in Figure [7](#S4.F7 "Figure 7 ‣ Simulator ‣ 4.2 Simulation setup ‣ 4 Simulation results")(b);
3. 3\.
the MIT campus realistic workload scenario.
For each one, we have plotted the code propagation speed for GCP, FP, PBP and FCP. Each synthetic scenario has approximately the same propagation behaviour as the 9 socializing clusters scenario (cf. Figure [9](#S4.F9 "Figure 9 ‣ 4.3 Convergence speed ‣ 4 Simulation results")). Due to the space constrain, we have represented the results for only two synthetic scenarios. The 9 clusters scenario is presented here to illustrate the fact that GCP outperforms the FCP algorithm.
In the flooding algorithm, each time a sensor node meets another one, it sends its own software. As explained before, the flooding code propagation speed can be considered as the ideal transmission speed and is taken as reference thereafter. For the two algorithms with token rules (Forwarding control mechanism), we have plotted the results obtained by using $`k`$ tokens a node, where $`k`$ is respectively equal to 2, 3 and 5. We do not represent results for more than 5 tokens as in the case majority, GCP tends to approach the ideal reference by using only 5 tokens (Each synthetic simulation system counts around 2,000 sensors, so for all $`k \sim {\log{(2000)}} = 3.3`$ \[[10](#bib.bib10)\]).
[Refer to caption](/html/0707.3717/assets/x6.png)
Figure 8: Convergence speed for the largest clustered scenario
[Refer to caption](/html/0707.3717/assets/x7.png)
Figure 9: Convergence speed for the largest socialized scenario
[Refer to caption](/html/0707.3717/assets/x8.png)
Figure 10: Convergence speed for realistic scenario
Regardless of the number of tokens chosen, in each scenario, GCP outperforms FCP as far as propagation speed is concerned. Taken flooding algorithm as ideal reference, Figure [8](#S4.F8 "Figure 8 ‣ 4.3 Convergence speed ‣ 4 Simulation results") shows different inflection points. These are due to the software transmission from a cluster to another. This is clearly denoted in Figure [8](#S4.F8 "Figure 8 ‣ 4.3 Convergence speed ‣ 4 Simulation results"): when almost half of sensors nodes have picked up the software’s newest version, there is a period during which the newest version are moving from a cluster to the other.
Figure [10](#S4.F10 "Figure 10 ‣ 4.3 Convergence speed ‣ 4 Simulation results") presents the propagation speed during the realistic scenario. It is interesting to observe that FCP algorithm is not efficient in this scenario. In real life, some people are together most of the time. As nodes used FCP algorithm, they are not aware from the remote nodes’ version, and may spend all their tokens for the same node. GCP with a small number of tokens (2 and 3 here for instance) remains better than FCP with a large number of tokens. With only 5 tokens per node, GCP is almost as efficient as the flooding algorithm with respect to propagation speed.
For each scenario, PBP is significantly slower than the flooding one and is always equivalent to GCP with 5 tokens per node.
Simulation results show the propagation speed efficiency of GCP according to the FCP algorithm for the same number of tokens and to the flooding algorithm as ideal reference. We have measured as well the network load balancing, presented in the next subsection.
### 4.4 Load Balancing
[Refer to caption](/html/0707.3717/assets/x9.png)
Figure 11: Load balancing for the largest socialized scenario
In order to estimate the benefit of GCP in term of load balancing, we have collected for each simulation (scenarios and algorithms), the number of sent software messages by node. As each simulation have the same network load behaviour, Figure [11](#S4.F11 "Figure 11 ‣ 4.4 Load Balancing ‣ 4 Simulation results") presents the results for the 9 socializing clusters scenario. For each number of message sent (represents by the X-axes), we have represent for all simulations the number of nodes which have sent exactly this number of time its software. As the flooding consumes much more message than the three other algorithms (PBP, FCP and GCP), the network load with flooding is represent in the upper-right corner of Figure [11](#S4.F11 "Figure 11 ‣ 4.4 Load Balancing ‣ 4 Simulation results").
For each scenario, the benefit of GCP or FCP compared to the flooding algorithm is clear. When considering the number of software binary sent, the two other algorithms save between 82 % and 93 % of messages for FCP and more than 98 % for GCP for a 50 seconds simulation only. As we have presented above, the number of software sent message increase according linearly to time with the flooding algorithm.
By focusing only on the three other algorithms, main plots of Figure [11](#S4.F11 "Figure 11 ‣ 4.4 Load Balancing ‣ 4 Simulation results") show the benefits of GCP according to PBP and FCP. In fact, as FCP is not aware of the remote node’s version, the local node sends its own version as long as it still possesses tokens. That implies the number of software sending messages with FCP algorithm is almost constant, corresponding of two times the number of tokens owns by each node ($`k`$ tokens for the first version plus $`k`$ tokens for the newer version: in these simulations, the software is updated only once).
With GCP and PBP, the current local version number is sent in the beacon message (Piggy-backing mechanism). So, nodes will not send the first version. They only send the newest version, and only if the beacon sender does not owned the latest version. The total number of software send messages in the network is almost equivalent to the number of participating nodes. Moreover, as every node in the transmission range of a sending node receive the sending version, freely for the sender, the network load benefits using GCP is decrease all the more. Comparing GCP with 5 tokens a node and PBP, which have the same propagation speed, we observe that near 5 % of the nodes have send the software more that 5 times, contrary to GCP where nodes have consumed at most 5 tokens.
We do not represent the load extracted from simulation on the realistic trace because it shares the same aspect as the synthetic ones.
## 5 Related works
This section presents various works in two related axes namely: data dissemination (or broadcasting) and software update (or reprogramming / code propagation). Among those, we focus on the approaches relying on gossip-based algorithms.
### 5.1 Dissemination in WSN
Reliable data dissemination shares some goals and assumptions with code propagation. Reliable broadcasting consist in a one-to-all message dissemination (one entry point in the system sends some information to all participating nodes in the network).
Vollset and al. \[[18](#bib.bib18)\] propose a classification of reliable broadcast in two classes: deterministic and probabilistic protocols. Deterministic protocols attempt to enforce hard reliability guarantees, as probabilistic protocols provide guaranteed delivery with a certain probability. This paper concludes that deterministic approach tends to have bad tradeoffs necessary between reliability and scalability/mobility, instead of probabilistic protocols, which do not provide deterministic delivery guarantees. Due to the constraints present in Section [1](#S1 "1 Introduction"), we focus on probabilistic protocols to obtain a reliable efficiency in scalability and mobility models.
Some study used geographical information to increase broadcast efficiency (load balancing, propagation speed, …). In \[[17](#bib.bib17)\], several degrees of local information has introduced in broadcast protocols. Subramanian and al. present three degrees of knowledge: no geographic or state information, coarse geographic information about the origin of the broadcast and no geographic information, but remember previously received messages. Authors conclude that local information have a not negligible role in broadcasting in WSN. As our contribution, several works tends to increase the broadcast efficiency without using locality. \[[8](#bib.bib8), [14](#bib.bib14), [15](#bib.bib15)\]
Another probabilistic broadcast is based on the gossip-based model \[[12](#bib.bib12), [19](#bib.bib19)\]. Gossip-based protocol is a promising model for WSN as presented in Section [1](#S1 "1 Introduction"). Gossip as a general technique has been used to solve several problems (data management, failure detection, …). In the end of this subsection, we present different gossip-based dissemination protocols for WSNs.
In \[[12](#bib.bib12)\], a reliable broadcast service based on the gossip model is presented. Using probabilistic forwarding, gossip is used to self-adapt the probability of information send for any topology. But, in most case, mobility is not taken into account.
Wang and al. \[[19](#bib.bib19)\] proposed a reliable broadcast protocol for mobile wireless sensor network. By using clustering technique and gossiping, this protocol has high delivery ratio and low end-to-end delay. Another noteworthy work is a reliable multicast protocol for mobile ad-hoc network (MANET), called Anonymous gossip \[[4](#bib.bib4)\]. Based on reliable gossip-based multicast for wired networks, gossip is used to obtain some information about message which node has not yet received. In these two last references, authors are studying MANETs, which have more resources than sensor nodes.
Our works is a transversal axis of these last studies. We are taking into account mobility and power consumption of WSN by using gossip-based paradigm.
### 5.2 Code propagation in WSN
Recently, gossip-based model has been used in code propagation service \[[13](#bib.bib13), [20](#bib.bib20)\]. Trickle \[[13](#bib.bib13)\] is a code propagation algorithm using a ”polite gossip” policy. This algorithm permits to propagate and maintain code updates in WSN. It regulates the transmission by gossiping code meta-data. If the information is up-to-date, the receivers stay quiet. Otherwise, if the information is issued from an older version, the gossiper can be brought up to date, and similarly.
Other works propose several protocols to apply code propagation. Deluge \[[7](#bib.bib7)\] assure to reprogramming the network by a reliable data dissemination protocol. Authors argue that Deluge can characterize its overall performance. Notably, they assume that it may be difficult to significantly improve the transmission rate obtained by Deluge. Kulkarni and Wang proposed MNP \[[11](#bib.bib11)\], a multi-hop reprogramming service, by splitting code into several segments. Using pipelining and sleep cycle, this algorithm also guarantees that, in a neighbourhood, there is at most one source transmitting the program at a time. Recently, these authors present Gappa \[[20](#bib.bib20)\], an extension of MNP. Using an Unmanned Ariel Vehicle (UAV), this algorithm can communicate parts of the code to a subset of sensor nodes on a multiple channel at once. The protocol ensures that at any time, there is at most one sensor transmitting on a given frequency.
In the best of our knowledge, none of work treats about code propagation in *mobile* WSN.
## 6 Conclusion
In this paper, we have proposed a software update algorithm for mobile wireless sensor networks. As the best of our knowledge, tackle code propagation in mobile WSN has not been done before. Based on leverage works on epidemic protocols and on similarities and differences between P2P systems and mobile WSNs, Gossip-based Code Propagation algorithm tends to outperforms classical dissemination algorithms, with only a small overhead by adding little extra information on sensor nodes and in beacon messages.
We have exposed the benefit of GCP on several simulation scenarios, compared as three other dissemination algorithms: one ideal in speed convergence but with a large number of software send messages and, therefore, very high power consumption, another one based on forwarding control and a last one based on piggy-backing message.
For each of these algorithms, GCP obtains an important profit accordingly to the little overhead information. With a clearly load balance through the network, GCP can disseminate the new software with almost the same propagation speed than the ideal one.
One of these work perspectives consists to include the Delay Tolerant Networks (DTN) paradigm into GCP in order to take into account and optimize the free receptions of the software due to the omnidirectionnal wireless transmission.
## References
- \[1\] I. F. Akyildiz, W. Su, Y. Sankarasubramaniam, and E. Cayirci. A survey on sensor networks.
IEEE Communications Magazine, 40(8):102–114, 2002.
- \[2\] S. Brown and C. J. Sreenan. A new model for updating software in wireless sensor networks.
IEEE Network, pages 42–47, November/December 2006.
- \[3\] M. Castro, P. Druschel, Y. C. Hu, and A. Rowstron. Proximity neighbor selection in tree-based structured peer-to-peer overlays. Technical Report MSR-TR-2003-52, Microsoft Reasearch, Cambridge, UK, 2003.
- \[4\] R. Chandra, V. Ramasubramanian, and K. Birman. Anonymous gossip: Improving multicast reliability in ad-hoc networks.
In International Conference on Distributed Computing Systems (ICDCS 2001), Phoenix, Arizona, April 2001.
- \[5\] K. Chen and Y. Jian. Survey on peer to peer data dissemination in manet. Computer Information Science Engineering Department, November 2005.
- \[6\] N. Eagle and A. S. Pentland. CRAWDAD data set mit/reality (v. 2005-07-01). Downloaded from http://crawdad.cs.dartmouth.edu/mit/reality, July 2005.
- \[7\] J. W. Hui and D. Culler. The dynamic behavior of a data dissemination protocol for network programming at scale.
In Proceedings of the 2nd international conference on Embedded networked sensor systems (SenSys’04), pages 81–94, New York, NY, USA, 2004. ACM Press.
- \[8\] C. Intanagonwiwat, R. Govindan, D. Estrin, J. Heidemann, and F. Silva. Directed diffusion for wireless sensor networking.
IEEE/ACM Transactions Networks, 11(1):2–16, February 2003.
- \[9\] D. B. Johnson and D. A. Maltz. Dynamic source routing in ad hoc wireless networks.
Mobile Computing, 353:153–181, 1996.
- \[10\] A.-M. Kermarrec, L. Massoulié, and A. J. Ganesh. Probabilistic reliable dissemination in large-scale systems.
IEEE Transactions on Parallel and Distributed Systems, 14(3), March 2003.
- \[11\] S. S. Kulkarni and L. Wang. Mnp: multihop network reprogramming service for sensor networks.
In the 25th International Conference on Distributed Computing System (ICDCS’05), pages 7–16, Columbus, OH, USA, June 2005.
- \[12\] P. Kyasanur, R. R. Choudhury, and I. Gupta. Smart gossip: An adaptive gossip-based broadcasting service for sensor networks.
In The Third IEEE International Conference on Mobile Ad-hoc and Sensor Systems (MASS 2006), Vancouver, Canada, October 2006.
- \[13\] P. Levis, N. Patel, S. Shenker, and D. Culler. Trickle: A self-regulating algorithm for code propagation and maintenance in wireless sensor networks.
In First Symposium on Network Systems Design and Implementation (NSDI), march 2004.
- \[14\] N. Li and J. C. Hou. Blmst: A scalable, power-efficient broadcast algorithm for wireless networks.
In QSHINE ’04: Proceedings of the First International Conference on Quality of Service in Heterogeneous Wired/Wireless Networks (QSHINE’04), pages 44–51, Washington, DC, USA, 2004. IEEE Computer Society.
- \[15\] N. Mitton, A. Busson, and E. Fleury. Broadcasting in self-organizing wireless multi-hop network. Research report RR-5487, INRIA, February 2005.
- \[16\] P. Rentala, R. Musunuri, S. Gandham, and U. Saxena. Survey on sensor networks.
In Proceedings of International Conference on Mobile Computing and Networking, 2001.
- \[17\] S. Subramanian, S. Shakkottai, and A. Arapostathis. Broadcasting in sensor networks: The role of local information.
In IEEE, editor, IEEE Infocom, 2006.
- \[18\] E. Vollset and P. Ezhilchelvan. A survey of reliable broadcast protocols for mobile ad-hoc networks. Technical Report CS-TR-792, University of Newcastle upon Tyne, 2003.
- \[19\] G. Wang, D. Lu, W. Jia, and J. Cao. Reliable gossip-based broadcast protocol in mobile ad hoc networks.
In Mobile Ad-hoc and Sensor Networks: First International Conference, MSN 2005, pages 207–218, Wuhan, China, December 2005.
- \[20\] L. Wang and S. S. Kulkarni. Gappa: Gossip based multi-channel reprogramming for sensor networks.
In Second IEEE International Conference in Distributed Computing in Sensor Systems (DCOSS 2006), pages 119–134, 2006.
- \[21\] B. Xu and O. Wolfson. Data management in mobile peer-to-peer networks.
In S. V. L. N. in Computer Science, editor, Proc. of the 2nd International Workshop on Databases, Information Systems, and Peer-to-Peer Computing (DBISP2P’04), pages 1–15, Toronto, Canada, August 2004.
###### Contents
1. [](#S1)
1 Introduction
2. [](#S2)
2 Applying P2P algorithms to mobile WSN
1. [](#S2.SS1 "In 2 Applying P2P algorithms to mobile WSN")
2.1 Peer-to-Peer vs. mobile WSN
2. [](#S2.SS2 "In 2 Applying P2P algorithms to mobile WSN")
2.2 Epidemic algorithms
3. [](#S3)
3 GCP : Control flooding for mobile WSNs
1. [](#S3.SS1 "In 3 GCP: Introducing control in flooding as a miracle drug to mobile WSNs")
3.1 GCP design
2. [](#S3.SS2 "In 3 GCP: Introducing control in flooding as a miracle drug to mobile WSNs")
3.2 GCP algorithm
3. [](#S3.SS3 "In 3 GCP: Introducing control in flooding as a miracle drug to mobile WSNs")
3.3 Alternative algorithms
4. [](#S3.SS4 "In 3 GCP: Introducing control in flooding as a miracle drug to mobile WSNs")
3.4 Protocols in Pseudo-Code
5. [](#S3.SS5 "In 3 GCP: Introducing control in flooding as a miracle drug to mobile WSNs")
3.5 Theoretical analysis
6. [](#S3.SS6 "In 3 GCP: Introducing control in flooding as a miracle drug to mobile WSNs")
3.6 Summary
4. [](#S4)
4 Simulation results
1. [](#S4.SS1 "In 4 Simulation results")
4.1 System model
2. [](#S4.SS2 "In 4 Simulation results")
4.2 Simulation setup
3. [](#S4.SS3 "In 4 Simulation results")
4.3 Convergence speed
4. [](#S4.SS4 "In 4 Simulation results")
4.4 Load Balancing
5. [](#S5)
5 Related works
1. [](#S5.SS1 "In 5 Related works")
5.1 Dissemination in WSN
2. [](#S5.SS2 "In 5 Related works")
5.2 Code propagation in WSN
6. [](#S6)
6 Conclusion |
0708.1414 | cs.NI | cyber | # Wavelet Based Semi-blind Channel Estimation For Multiband OFDM
Sajad Sadough<sup>†∗</sup>, Mahieddine M. Ichir<sup>∗</sup>, Emmanuel Jaffrot<sup>†</sup> and Pierre Duhamel<sup>∗</sup>
<sup>†</sup> UEI, ENSTA, 32 boulevard Victor, 75015 Paris, France
<sup>∗</sup> Laboratoire des Signaux et Systèmes, CNRS-Supélec,
Plateau de Moulon, 91190 Gif–sur–Yvette, France
Email: sajad.sadough@ensta.fr
###### Abstract
This paper introduces an expectation-maximization (EM) algorithm within a wavelet domain Bayesian framework for semi-blind channel estimation of multiband OFDM based UWB communications. A prior distribution is chosen for the wavelet coefficients of the unknown channel impulse response in order to model a sparseness property of the wavelet representation. This prior yields, in maximum a posteriori estimation, a thresholding rule within the EM algorithm. We particularly focus on reducing the number of estimated parameters by iteratively discarding “unsignificant” wavelet coefficients from the estimation process. Simulation results using UWB channels issued from both models and measurements show that under sparsity conditions, the proposed algorithm outperforms pilot based channel estimation in terms of mean square error and bit error rate and enhances the estimation accuracy with less computational complexity than traditional semi-blind methods.
## I Introduction
A UWB radio signal is defined as any signal whose bandwidth is larger than 20% of its center frequency or greater than 500 MHz \[[1](#bib.bib1)\]. In recent years, UWB system design has experienced a shift from the traditional “single-band” radio that occupies the whole 7.5 GHz allocated spectrum to a “multiband” design approach \[[2](#bib.bib2)\]. That consists in dividing the available UWB spectrum into several subbands, each one occupying approximately 500 MHz.
Multiband Orthogonal Frequency Division Multiplexing (MB-OFDM) \[[3](#bib.bib3)\] is a strong candidate for multiband UWB which enables high data rate UWB transmission to inherit all the strength of OFDM that has already been shown for wireless communications (ADSL, DVB, 802.11a, 802.16.a, etc.). This approach uses a conventional coded OFDM system \[[4](#bib.bib4)\] together with bit interleaved coded modulation (BICM) and frequency hopping over different subbands to improve diversity and to enable multiple access.
Basic receivers proposed for MB-OFDM \[[3](#bib.bib3)\], estimate the channel by using pilots (known training symbols) transmitted at the beginning of the information frame, implicitly assuming a time invariant channel within a single frame. Thus, for an accurate channel acquisition, one must send several pilot patterns resulting in a significant loss in spectral efficiency.
Recent works \[[5](#bib.bib5), [6](#bib.bib6)\] have reported promising results on the combination of channel estimation and data decoding process by using the Expectation-Maximization (EM) algorithm \[[7](#bib.bib7)\] . Though the latter scheme outperforms pilot based receivers, it has a higher complexity that may be of a critical concern for its practical implementations. This complexity is mainly dominated by the number of estimated parameters for channel updating and the decoding algorithm within each iteration.
In this work, we consider a semi-blind joint channel estimation and data detection scheme based on the EM algorithm, with the objective of minimizing the number of estimated parameters and enhancing the estimation accuracy. This is achieved by expressing the unknown channel impulse response (CIR) in terms of its discrete wavelet series, which has been shown to provide a parsimonious representation \[[8](#bib.bib8), [9](#bib.bib9)\]. Thus, we choose a particular prior distribution for the channel wavelet coefficients that renders the maximum a posteriori (MAP) channel estimation equivalent to a hard thresholding rule at each iteration of the EM algorithm. The latter is then exploited to reduce the estimator computational load by discarding “unsignificant” wavelet coefficients from the estimation process. Moreover, since the probability of encoded bits are involved in the EM computation, we naturally combine the iterative process of channel estimation with the decoding operation of encoded data.
This paper is organized as follows. Section [II](#S2 "II System model and wavelet domain problem formulation ‣ Wavelet Based Semi-blind Channel Estimation For Multiband OFDM") introduces MB-OFDM and its wavelet domain channel estimation observation model. In section [III](#S3 "III The EM-MAP algorithm for wavelet domain channel estimation ‣ Wavelet Based Semi-blind Channel Estimation For Multiband OFDM"), we first describe a MAP version of the EM algorithm for channel estimation and then show how the number of estimated parameters can be reduced through the EM iterations. The combination of the channel estimation part with the decoding operation and implementation issues are also discussed. Section [V](#S5 "V Simulation results ‣ Wavelet Based Semi-blind Channel Estimation For Multiband OFDM") illustrates, via simulations, the performance of the proposed receiver over a realistic UWB channel environment and section [VI](#S6 "VI Conclusion ‣ Wavelet Based Semi-blind Channel Estimation For Multiband OFDM") concludes the paper.
Notational conventions are as follows: $`\mathcal{D}_{\mathbf{x}}`$ is a diagonal matrix with diagonal elements $`\mathbf{x} = {\lbrack x_{1},\ldots,x_{N}\rbrack}^{T}`$, $`{\mathbb{E}}_{\mathbf{x}}{\lbrack.\rbrack}`$ refers to expectation with respect to $`\mathbf{x}`$, $`\mathbf{I}_{N}`$ denotes an $`({N \times N})`$ identity matrix; $`\parallel . \parallel`$, $`{(.)}^{\ast}`$, $`{(.)}^{T}`$ and $`{(.)}^{\mathcal{H}}`$ denote Frobenious norm, matrix or vector conjugation, transpose and Hermitian transpose, respectively.
## II System model and wavelet domain problem formulation
MB-OFDM system divides the spectrum between 3.1 to 10.6 GHz into several non-overlapping subbands each one occupying 528 MHz of bandwidth \[[3](#bib.bib3)\]. The transmitter architecture for the MB-OFDM system is very similar to that of a conventional wireless OFDM system. The main difference is that MB-OFDM system uses a time-frequency code (TFC) to select the center frequency of different subbands which is used not only to provide frequency diversity but also to distinguish between multiple users (see figures [1](#S2.F1 "Figure 1 ‣ II System model and wavelet domain problem formulation ‣ Wavelet Based Semi-blind Channel Estimation For Multiband OFDM") and [2](#S2.F2 "Figure 2 ‣ II System model and wavelet domain problem formulation ‣ Wavelet Based Semi-blind Channel Estimation For Multiband OFDM")). Here, we consider MB-OFDM in its basic mode ie. employing the three first subbands.
<table>
<tbody>
<tr>
<td>3 lower subbands {casesmissing-subexpressionmissing-subexpressionmissing-subexpressionmissing-subexpressionmissing-subexpressionmissing-subexpressionmissing-subexpression\left\{\begin{array}[]{c}\\ \\ \\ \\ \\ \\ \\ \\ \end{array}\right.</td>
<td>\psfrag{B}{\scriptsize\raisebox{-7.00002pt}{Band \# 1}}\psfrag{C}{\scriptsize\raisebox{-7.00002pt}{Band \# 2}}\psfrag{D}{\scriptsize Band \# 3}\psfrag{1}{{\scriptsize 3168}}\psfrag{2}{\scriptsize 3696}\psfrag{3}{\scriptsize 4224}\psfrag{4}{\scriptsize 4752}\psfrag{E}{\scriptsize 9.5 ns}\psfrag{F}{ \parbox{59.50012pt}{\centering\scriptsize Guard Interval for\\ TX/RX switching\@add@centering}}\psfrag{G}{\scriptsize 312.5 ns}\psfrag{I}{\scriptsize 60.6 ns}\psfrag{A}{\scriptsize Time (ns)}\psfrag{H}{\rotatebox{90.0}{ Frequency (MHz)}}\includegraphics[width=173.44534pt,height=113.81102pt]{TFC_eII}</td>
</tr>
</tbody>
</table>
Figure 1: Example of time-frequency coding for the multiband OFDM system: TFC={1, 3, 2, 1, 3, 2, …}.
\psfrag{A}{{\small Binary Data}}\psfrag{B}{{\parbox{55.00008pt}{\centering{\small Punctured Convolutional Encoder}\\add@centering}}}\psfrag{C}{{\parbox{60.00009pt}{\centering{\small Bit\\ Interleaver}\\add@centering}}}\psfrag{D}{\parbox{60.00009pt}{\centering{\small QPSK\\ Mapping}\\add@centering}}\psfrag{E}{\parbox{60.00009pt}{\centering{\small IFFT\\ CP \\ GI\\ Addition}\\add@centering}}\psfrag{F}{\parbox{60.00009pt}{\centering{\small DAC}\\add@centering}}\psfrag{G}{\parbox{60.00009pt}{\centering\$\exp\left(j2\pi f\_{c}t\right)\$\\add@centering}}\psfrag{H}{{{\small TFC: Subband Selection}}}\includegraphics\[width=195.12767pt,height=113.81102pt\]{TX_chainII}
Figure 2: TX architecture of the multiband OFDM system.
We consider the multiband OFDM transmission of figure [2](#S2.F2 "Figure 2 ‣ II System model and wavelet domain problem formulation ‣ Wavelet Based Semi-blind Channel Estimation For Multiband OFDM") using $`N`$ data subcarriers. At the receiver, assuming a cyclic prefix (CP) longer than the channel maximum delay spread and perfect synchronization, OFDM converts a frequency selective channel into $`N`$ parallel flat fading subchannels \[[4](#bib.bib4)\] for each subband as
<table>
<tbody>
<tr>
<td></td>
<td>𝐲i,n=𝒟𝐬i,n𝐡i,n+𝐳i,ni∈{1,2,3},n=1,…,Nsymformulae-sequencesubscript𝐲𝑖𝑛subscript𝒟subscript𝐬𝑖𝑛subscript𝐡𝑖𝑛subscript𝐳𝑖𝑛formulae-sequence𝑖123𝑛1…subscriptNsym\mathbf{y}_{i,n}=\mathcal{D}_{{\mathbf{s}_{i,n}}}\,\mathbf{h}_{i,n}+\mathbf{z}_{i,n}\;\;\;\;i\in\{1,2,3\},\;\;\;n=1,\ldots,{\rm N_{sym}}</td>
<td></td>
<td>(1)</td>
</tr>
</tbody>
</table>
where $`({1 \times N})`$ vectors $`\mathbf{y}_{i,n}`$, $`\mathbf{s}_{i,n}`$ and $`\mathbf{h}_{i,n}`$ denote received and transmitted symbols, and the channel frequency response respectively; the noise block $`\mathbf{z}_{i,n}`$ is assumed to be a zero mean white complex Gaussian noise with distribution $`\mathcal{C}\hspace{0pt}\mathcal{N}\hspace{0pt}{(\mathbf{0},{\sigma^{2}\hspace{0pt}\mathbf{I}_{N}})}`$ ; $`i`$ is the subband index and $`n`$ refers to the OFDM symbol index inside the frame. The observation model corresponding to all three subbands can be written in frequency domain as
<table>
<tbody>
<tr>
<td></td>
<td>𝐘m=𝒟𝐒m𝐇m+𝐙mm=1,…,Msymformulae-sequencesubscript𝐘𝑚subscript𝒟subscript𝐒𝑚subscript𝐇𝑚subscript𝐙𝑚𝑚1…subscriptMsym\mathbf{Y}_{m}=\mathcal{D}_{{}_{\mathbf{S}_{m}}}\,\mathbf{H}_{m}+\mathbf{Z}_{m}\;\;\;\;m=1,\ldots,{\rm M_{sym}}</td>
<td></td>
<td>(2)</td>
</tr>
</tbody>
</table>
where $`\mathbf{Y}_{m} = {\lbrack\mathbf{y}_{1,n},\mathbf{y}_{2,n},\mathbf{y}_{3,n}\rbrack}^{T}`$, $`\mathbf{S}_{m} = {\lbrack\mathbf{s}_{1,n},\mathbf{s}_{2,n},\mathbf{s}_{3,n}\rbrack}^{T}`$, $`\mathbf{H}_{m} = {\lbrack\mathbf{h}_{1,n},\mathbf{h}_{2,n},\mathbf{h}_{3,n}\rbrack}^{T}`$ and $`\mathbf{Z}_{m} = {\lbrack\mathbf{z}_{1,n},\mathbf{z}_{2,n},\mathbf{z}_{3,n}\rbrack}^{T}`$ are ($`M \times 1`$) vectors, with $`M = {3\hspace{0pt}N}`$ and $`M_{sym} = {N_{sym}/3}`$. In the remainder, unless otherwise mentioned, we will not write the time index $`m`$ for notational convenience.
In order to take advantage of the wavelet based estimation, the channel impulse response is expressed in terms of its orthogonal discrete wavelet coefficients. Let $`\mathbf{F}_{M,L}`$ be the truncated fast Fourier transform (FFT) matrix constructed from the ($`M \times M`$) FFT matrix by keeping the first $`L`$ columns where $`L`$ is the length of the CIR over a group of three subbands. We define $`\mathbf{W}`$ as the $`({L \times L})`$ orthogonal discrete wavelet transform (ODWT) matrix. The unknown channel can be expressed as $`\mathbf{H} = {\mathbf{F}_{M,L}\hspace{0pt}\mathbf{W}^{\mathcal{H}}\hspace{0pt}\mathbf{g}}`$, where $`\mathbf{g}`$ is the $`({L \times 1})`$ vector of the CIR wavelet coefficients. The Observation model [2](#S2.E2 "In II System model and wavelet domain problem formulation ‣ Wavelet Based Semi-blind Channel Estimation For Multiband OFDM") is rewritten as
<table>
<tbody>
<tr>
<td></td>
<td>𝐘=𝒟𝐒𝐓𝐠+𝐙𝐘subscript𝒟𝐒𝐓𝐠𝐙\mathbf{Y}=\mathcal{D}_{{}_{\mathbf{S}}}\,\mathbf{T}\,\mathbf{g}+\mathbf{Z}</td>
<td></td>
<td>(3)</td>
</tr>
</tbody>
</table>
where $`\mathbf{T} = {\mathbf{F}_{M,L}\hspace{0pt}\mathbf{W}^{\mathcal{H}}}`$.
Although at the transmitter, the channel is practically used by slices of 528 MHz bandwidth that corresponds to one of the subbands, at the receiver side we gather three received OFDM symbols for estimating the wavelet coefficients of the CIR, taken over all of the subbands (1.584 GHz bandwidth). This is motivated by the fact that estimating the channel over a wider bandwidth leads to a sparser wavelet representation. Besides, this approach simplifies the receiver architecture since there is no need to change the central frequency for down converting different subbands.
## III The EM-MAP algorithm for wavelet domain channel estimation
The EM algorithm proposed in this section is able to integrate the advantages of wavelet based estimation via the prior choosen for channel wavelet coefficients. Next, we see how the MAP estimator leads to a thresholding procedure which is used for reducing the number of estimated coefficients at each iteration of the EM algorithm.
###
III-A
An equivalent model and the EM principle
Our first step consists in decomposing the AWGN in ([3](#S2.E3 "In II System model and wavelet domain problem formulation ‣ Wavelet Based Semi-blind Channel Estimation For Multiband OFDM")) into the sum of two different Gaussian noise terms as
<table>
<tbody>
<tr>
<td></td>
<td>𝐙=𝒟𝐒𝐙1+𝐙2𝐙subscript𝒟𝐒subscript𝐙1subscript𝐙2\mathbf{Z}=\mathcal{D}_{{}_{\mathbf{S}}}\mathbf{Z}_{1}+\mathbf{Z}_{2}</td>
<td></td>
<td>(4)</td>
</tr>
</tbody>
</table>
where $`\mathbf{Z}_{1}`$ and $`\mathbf{Z}_{2}`$ are $`({M \times 1})`$ independent Gaussian noise vectors such that $`{p\hspace{0pt}{(\mathbf{Z}_{1})}} = {\mathcal{C}\hspace{0pt}\mathcal{N}\hspace{0pt}{(\mathbf{0},{\alpha^{2}\hspace{0pt}\mathbf{I}_{M}})}}`$ and $`{p\hspace{0pt}{(\mathbf{Z}_{2})}} = {\mathcal{C}\hspace{0pt}\mathcal{N}\hspace{0pt}{(\mathbf{0},{{\sigma^{2}\hspace{0pt}\mathbf{I}_{M}} - {\alpha^{2}\hspace{0pt}\mathcal{D}_{_{\mathbf{S}}}\hspace{0pt}\mathcal{D}_{_{\mathbf{S}}}^{\mathcal{H}}}})}}`$. Since we are using normalized QPSK symbols, $`{\mathcal{D}_{_{\mathbf{S}}}\hspace{0pt}\mathcal{D}_{_{\mathbf{S}}}^{\mathcal{H}}} = \mathbf{I}_{M}`$ and the covariance matrix of $`\mathbf{Z}_{2}`$ reduces to $`\mathbf{\Sigma}_{_{2}} = {{({\sigma^{2} - \alpha^{2}})}\hspace{0pt}\mathbf{I}_{M}}`$. We define the positive design parameter $`\rho \triangleq {\alpha^{2}/\sigma^{2}}`$, $`({0 < \rho \leq 1})`$ and notice that setting $`\rho = 1`$ leads to $`\mathbf{Z}_{2} = 0`$ which is equivalent to working with the initial model ([3](#S2.E3 "In II System model and wavelet domain problem formulation ‣ Wavelet Based Semi-blind Channel Estimation For Multiband OFDM")). However, for $`0 < \rho < 1`$, the above noise decomposition allows the introduction of a hidden channel vector $`\overset{\sim}{\mathbf{H}}`$ defined as
<table>
<tbody>
<tr>
<td></td>
<td>{𝐇~=𝐓𝐠+𝐙1𝐘=𝒟𝐒𝐇~+𝐙2.cases~𝐇absent𝐓𝐠subscript𝐙1𝐘absentsubscript𝒟𝐒~𝐇subscript𝐙2\left\{\begin{array}[]{ll}\widetilde{\mathbf{H}}&=\mathbf{T}\,\mathbf{g}+\mathbf{Z}_{1}\\ \mathbf{Y}&=\mathcal{D}_{{}_{\mathbf{S}}}\,\widetilde{\mathbf{H}}+\mathbf{Z}_{2}.\end{array}\right.</td>
<td></td>
<td>(5)</td>
</tr>
</tbody>
</table>
The hidden vector $`\overset{\sim}{\mathbf{H}}`$ provides a direct relation between true and estimated wavelet coefficients corrupted by an additive white Gaussian noise, allowing the two-stage observation model ([5](#S3.E5 "In III-A An equivalent model and the EM principle ‣ III The EM-MAP algorithm for wavelet domain channel estimation ‣ Wavelet Based Semi-blind Channel Estimation For Multiband OFDM")) which is equivalent to ([3](#S2.E3 "In II System model and wavelet domain problem formulation ‣ Wavelet Based Semi-blind Channel Estimation For Multiband OFDM")). However, the difference with a standard denoising problem is that $`\mathbf{S}`$ and $`\overset{\sim}{\mathbf{H}}`$ are unknown. Hence, the observation model has missing datas and hidden variables and the MAP solution of $`\mathbf{g}`$ has no closed form. In such situations, the EM algorithm \[[7](#bib.bib7)\] is often used to maximize the expectation of the posterior distribution over all possible missing and hidden variables.
Let $`\mathbf{X} = {\{\mathbf{Y},\mathbf{S},\overset{\sim}{\mathbf{H}}\}}`$ be the complete data set in the EM algorithm terminology. Note that the observation set $`\mathbf{Y}`$ determines only a subset of the space $`\mathcal{X}`$ of which $`\mathbf{X}`$ is an outcome. We search $`\mathbf{g}`$ that maximizes $`{\log p}\hspace{0pt}{(\left. \mathbf{g} \middle| \mathbf{X} \right.)}`$. After initialization by a short pilot sequence at the beginning of the frame, the EM algorithm alternates between the following two steps (until some stopping criterion) to produce a sequence of estimates $`\{{{{\mathbf{g}^{(t)},t} = 0},{1,\ldots,t_{\max}}}\}`$.
- •
Expectation Step (E-step): The conditional expectation of the complete log-likelihood given the observed vector and the current estimate $`\mathbf{g}^{(t)}`$ is calculated. This quantity is called the auxiliary or $`Q`$-function
<table>
<tbody>
<tr>
<td></td>
<td>Q(𝐠,𝐠(t))=𝔼𝐒,𝐇~[logp(𝐘,𝐒,𝐇~|𝐠)|𝐲,𝐠(t)]𝑄𝐠superscript𝐠𝑡subscript𝔼𝐒~𝐇delimited-[]conditional𝑝𝐘𝐒conditional~𝐇𝐠𝐲superscript𝐠𝑡Q\big{(}\mathbf{g},\mathbf{g}^{(t)}\big{)}=\mathbb{E}_{\mathbf{S},\widetilde{\mathbf{H}}}\Big{[}\log p(\mathbf{Y},\mathbf{S},\widetilde{\mathbf{H}}|\mathbf{g})\Big{|}\mathbf{y},\mathbf{g}^{(t)}\Big{]}</td>
<td></td>
<td>(6)</td>
</tr>
</tbody>
</table>
- •
Maximization Step (M-step): The estimated parameter is updated according to
<table>
<tbody>
<tr>
<td></td>
<td>𝐠(t+1)=argmax𝐠{Q(𝐠,𝐠(t))+logπ(𝐠)}superscript𝐠𝑡1subscriptargmax𝐠𝑄𝐠superscript𝐠𝑡𝜋𝐠{\mathbf{g}}^{(t+1)}=\displaystyle\mathop{\mathrm{arg\,max}}_{\mathbf{g}}\left\{{Q\big{(}\mathbf{g},{\mathbf{g}}^{(t)}\big{)}+\log\pi(\mathbf{g})}\right\}</td>
<td></td>
<td>(7)</td>
</tr>
</tbody>
</table>
where $`\pi\hspace{0pt}{(\mathbf{g})}`$ is a prior distribution for the wavelet coefficients. Next, we derive the specific formulas of each step, according to ([5](#S3.E5 "In III-A An equivalent model and the EM principle ‣ III The EM-MAP algorithm for wavelet domain channel estimation ‣ Wavelet Based Semi-blind Channel Estimation For Multiband OFDM")).
###
III-B
E-step: Computation of the $`Q`$-function
The complete likelihood is
<table>
<tbody>
<tr>
<td></td>
<td>p(𝐘,𝐒,𝐇~|𝐠)=p(𝐘|𝐒,𝐇~,𝐠)p(𝐒|𝐇~,𝐠)p(𝐇~|𝐠).𝑝𝐘𝐒conditional~𝐇𝐠𝑝conditional𝐘𝐒~𝐇𝐠𝑝conditional𝐒~𝐇𝐠𝑝conditional~𝐇𝐠p(\mathbf{Y},\mathbf{S},\widetilde{\mathbf{H}}|\mathbf{g})=p(\mathbf{Y}|\mathbf{S},\widetilde{\mathbf{H}},\mathbf{g})\,p(\mathbf{S}|\widetilde{\mathbf{H}},\mathbf{g})\,p(\widetilde{\mathbf{H}}|\mathbf{g}).</td>
<td></td>
</tr>
</tbody>
</table>
According to ([5](#S3.E5 "In III-A An equivalent model and the EM principle ‣ III The EM-MAP algorithm for wavelet domain channel estimation ‣ Wavelet Based Semi-blind Channel Estimation For Multiband OFDM")), conditioned on $`\overset{\sim}{\mathbf{H}}`$, $`\mathbf{Y}`$ is independent of $`\mathbf{g}`$. Furthermore, $`\mathbf{S}`$ which results from coding and interleaving of bit sequence is independent of $`\overset{\sim}{\mathbf{H}}`$ and $`\mathbf{g}`$. Since $`\mathbf{Z}_{1}`$ is a complex white Gaussian noise, the complete log-likelihood can be simplified to
<table>
<tbody>
<tr>
<td></td>
<td>logp(𝐘,𝐒,𝐇~|𝐠)𝑝𝐘𝐒conditional~𝐇𝐠\displaystyle\log p(\mathbf{Y},\mathbf{S},\widetilde{\mathbf{H}}|\mathbf{g})</td>
<td>=\displaystyle=</td>
<td>log[p(𝐘|𝐒,𝐇~)p(𝐒)p(𝐇~|𝐠)]𝑝conditional𝐘𝐒~𝐇𝑝𝐒𝑝conditional~𝐇𝐠\displaystyle\log\big{[}p(\mathbf{Y}|\mathbf{S},\widetilde{\mathbf{H}})\;p(\mathbf{S})\;p(\widetilde{\mathbf{H}}|\mathbf{g})\big{]}</td>
<td></td>
</tr>
<tr>
<td></td>
<td></td>
<td>=\displaystyle=</td>
<td>logp(𝐇~|𝐠)+cst.𝑝conditional~𝐇𝐠cst.\displaystyle\log p(\widetilde{\mathbf{H}}|\mathbf{g})\;+\;\textrm{cst.}</td>
<td></td>
</tr>
<tr>
<td></td>
<td></td>
<td>=\displaystyle=</td>
<td>−𝐠ℋ𝐓ℋ𝐓𝐠−2𝐠ℋ𝐓ℋ𝐇~α2+cst.superscript𝐠ℋsuperscript𝐓ℋ𝐓𝐠2superscript𝐠ℋsuperscript𝐓ℋ~𝐇superscript𝛼2cst.\displaystyle-\frac{\mathbf{g}^{\mathcal{H}}\mathbf{T}^{\mathcal{H}}\mathbf{T}\mathbf{g}-2\mathbf{g}^{\mathcal{H}}\mathbf{T}^{\mathcal{H}}\widetilde{\mathbf{H}}}{{\alpha}^{2}}\;+\;\textrm{cst.}</td>
<td></td>
</tr>
</tbody>
</table>
where cst. are different constant terms that do not depend on $`\mathbf{g}`$. According to ([6](#S3.E6 "In 1st item ‣ III-A An equivalent model and the EM principle ‣ III The EM-MAP algorithm for wavelet domain channel estimation ‣ Wavelet Based Semi-blind Channel Estimation For Multiband OFDM")) we have
<table>
<tbody>
<tr>
<td></td>
<td>Q(𝐠,𝐠(t))𝑄𝐠superscript𝐠𝑡\displaystyle Q\big{(}\mathbf{g},{\mathbf{g}}^{(t)}\big{)}</td>
<td>=𝔼𝐒,𝐇~[−𝐠ℋ𝐓ℋ𝐓𝐠−2𝐠ℋ𝐓ℋ𝐇~α2+cst.|𝐘,𝐠(t)]absentsubscript𝔼𝐒~𝐇delimited-[]superscript𝐠ℋsuperscript𝐓ℋ𝐓𝐠2superscript𝐠ℋsuperscript𝐓ℋ~𝐇superscript𝛼2conditionalcst.𝐘superscript𝐠𝑡\displaystyle=\mathbb{E}_{\mathbf{S},\widetilde{\mathbf{H}}}\Big{[}-\frac{\mathbf{g}^{\mathcal{H}}\mathbf{T}^{\mathcal{H}}\mathbf{T}\mathbf{g}-2\mathbf{g}^{\mathcal{H}}\mathbf{T}^{\mathcal{H}}\widetilde{\mathbf{H}}}{{\alpha}^{2}}+\textrm{cst.}\,\Big{|}\mathbf{Y},\mathbf{g}^{(t)}\Big{]}</td>
<td></td>
<td></td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td></td>
<td>=−‖⟨𝐇~(t)⟩−𝐓𝐠‖2α2+cst.absentsuperscriptnormdelimited-⟨⟩superscript~𝐇𝑡𝐓𝐠2superscript𝛼2cst.\displaystyle=-\frac{{\|\,\langle\widetilde{\mathbf{H}}^{(t)}\rangle-\mathbf{T}\mathbf{g}\,\|}^{2}}{{\alpha}^{2}}+\textrm{cst.}</td>
<td></td>
<td>(9)</td>
</tr>
</tbody>
</table>
where $`{\langle{\overset{\sim}{\mathbf{H}}}^{(t)}\rangle} \triangleq {{\mathbb{E}}_{\mathbf{S},\overset{\sim}{\mathbf{H}}}\hspace{0pt}{\lbrack\left. \overset{\sim}{\mathbf{H}} \middle| {\mathbf{Y},\mathbf{g}^{(t)}} \right.\rbrack}}`$.
From ([](#S3.Ex5 "III-B E-step: Computation of the Q-function ‣ III The EM-MAP algorithm for wavelet domain channel estimation ‣ Wavelet Based Semi-blind Channel Estimation For Multiband OFDM")
III-B
), it is obvious that the E-step involves only the computation of $`\langle{\overset{\sim}{\mathbf{H}}}^{(t)}\rangle`$, we have
<table>
<tbody>
<tr>
<td></td>
<td>⟨𝐇~(t)⟩delimited-⟨⟩superscript~𝐇𝑡\displaystyle\langle\widetilde{\mathbf{H}}^{(t)}\rangle</td>
<td>=∑𝐒∈𝒞(∫𝐇~∈ℋ𝐇~p(𝐇~|𝐘,𝐠(t))d𝐇~)p(𝐒|𝐘,𝐠(t))absentsubscript𝐒𝒞subscript~𝐇ℋ~𝐇𝑝conditional~𝐇𝐘superscript𝐠𝑡d~𝐇𝑝conditional𝐒𝐘superscript𝐠𝑡\displaystyle=\sum_{\mathbf{S}\in\mathscr{C}}{\bigg{(}\int_{\widetilde{\mathbf{H}}\in\mathscr{H}}\widetilde{\mathbf{H}}\quad p(\widetilde{\mathbf{H}}|\mathbf{Y},\mathbf{g}^{(t)})\,\mathrm{d}\widetilde{\mathbf{H}}\bigg{)}}p(\mathbf{S}|\mathbf{Y},\mathbf{g}^{(t)})</td>
<td></td>
<td>(10)</td>
</tr>
</tbody>
</table>
where the last equation results from the independence between $`\mathbf{S}`$ and $`\overset{\sim}{\mathbf{H}}`$ belonging respectively to the sets $`\mathcal{C}`$ and $`\mathcal{H}`$ which contain all of their possible values.
In order to evaluate $`\langle{\overset{\sim}{\mathbf{H}}}^{(t)}\rangle`$, we first have to evaluate the conditional mean $`{\mathbf{μ}}_{\overset{\sim}{\mathbf{H}}}^{(t)}`$ of $`\overset{\sim}{\mathbf{H}}`$ as
<table>
<tbody>
<tr>
<td></td>
<td>𝝁𝐇~(t)=∫𝐇~∈ℋ𝐇~p(𝐇~|𝐘,𝐠(t))d𝐇~superscriptsubscript𝝁~𝐇𝑡subscript~𝐇ℋ~𝐇𝑝conditional~𝐇𝐘superscript𝐠𝑡d~𝐇\boldsymbol{\mu}_{\widetilde{\mathbf{H}}}^{(t)}=\int_{\widetilde{\mathbf{H}}\in\mathscr{H}}\widetilde{\mathbf{H}}\quad p(\widetilde{\mathbf{H}}|\mathbf{Y},\mathbf{g}^{(t)})\,\mathrm{d}\widetilde{\mathbf{H}}</td>
<td></td>
<td>(11)</td>
</tr>
</tbody>
</table>
Since both $`p\hspace{0pt}{(\left. \mathbf{Y} \middle| \overset{\sim}{\mathbf{H}} \right.)}`$ and $`p\hspace{0pt}{(\left. \overset{\sim}{\mathbf{H}} \middle| \mathbf{g}^{(t)} \right.)}`$ are Gaussian densities, $`{p\hspace{0pt}{(\left. \overset{\sim}{\mathbf{H}} \middle| {\mathbf{Y},\mathbf{g}^{(t)}} \right.)}} \propto {p\hspace{0pt}{(\left. \mathbf{Y} \middle| \overset{\sim}{\mathbf{H}} \right.)}\hspace{0pt}p\hspace{0pt}{(\left. \overset{\sim}{\mathbf{H}} \middle| \mathbf{g}^{(t)} \right.)}}`$ is also Gaussian. By standard manipulation of Gaussian densities, we obtain
<table>
<tbody>
<tr>
<td></td>
<td>𝝁h(t)=𝐓𝐠(t)+ρ𝒟𝐒ℋ(𝐘−𝒟𝐒𝐓𝐠(t)).superscriptsubscript𝝁ℎ𝑡superscript𝐓𝐠𝑡𝜌superscriptsubscript𝒟𝐒ℋ𝐘subscript𝒟𝐒superscript𝐓𝐠𝑡\boldsymbol{\mu}_{h}^{(t)}=\mathbf{T}\mathbf{g}^{(t)}+\rho\,\mathcal{D}_{{}_{\mathbf{S}}}^{\mathcal{H}}\Big{(}\mathbf{Y}-\mathcal{D}_{{}_{\mathbf{S}}}\mathbf{T}\mathbf{g}^{(t)}\Big{)}.</td>
<td></td>
<td>(12)</td>
</tr>
</tbody>
</table>
By using ([12](#S3.E12 "In III-B E-step: Computation of the Q-function ‣ III The EM-MAP algorithm for wavelet domain channel estimation ‣ Wavelet Based Semi-blind Channel Estimation For Multiband OFDM")) in ([10](#S3.E10 "In III-B E-step: Computation of the Q-function ‣ III The EM-MAP algorithm for wavelet domain channel estimation ‣ Wavelet Based Semi-blind Channel Estimation For Multiband OFDM")) and after some simplifications we get
<table>
<tbody>
<tr>
<td></td>
<td>⟨𝐇~(t)⟩=(1−ρ)𝐓𝐠(t)+ρ𝒟¯𝐒ℋ𝐘delimited-⟨⟩superscript~𝐇𝑡1𝜌superscript𝐓𝐠𝑡𝜌subscriptsuperscript¯𝒟ℋ𝐒𝐘\langle\widetilde{\mathbf{H}}^{(t)}\rangle=(1-\rho)\,\mathbf{T}{\mathbf{g}}^{(t)}+\rho\,\overline{\mathcal{D}}^{\mathcal{H}}_{{}_{\mathbf{S}}}\mathbf{Y}</td>
<td></td>
<td>(13)</td>
</tr>
</tbody>
</table>
where $`{\overline{\mathcal{D}}}_{_{\mathbf{S}}} = {\sum_{\mathbf{s} \in \mathcal{C}}{\mathcal{D}_{_{\mathbf{S}}}\hspace{0pt}p\hspace{0pt}{(\left. \mathbf{S} \middle| {\mathbf{Y},\mathbf{g}^{(t)}} \right.)}}}`$.
The E-step is then completed by inserting $`\langle{\overset{\sim}{\mathbf{H}}}^{(t)}\rangle`$ into $`Q\hspace{0pt}{(\mathbf{g},\mathbf{g}^{(t)})}`$, equation ([](#S3.Ex5 "III-B E-step: Computation of the Q-function ‣ III The EM-MAP algorithm for wavelet domain channel estimation ‣ Wavelet Based Semi-blind Channel Estimation For Multiband OFDM")
III-B
).
###
III-C
M-step: Wavelet Based MAP Estimation
In this step the estimate of the parameter $`\mathbf{g}`$ is updated as given in ([7](#S3.E7 "In 2nd item ‣ III-A An equivalent model and the EM principle ‣ III The EM-MAP algorithm for wavelet domain channel estimation ‣ Wavelet Based Semi-blind Channel Estimation For Multiband OFDM")) where $`Q\hspace{0pt}{({\mathbf{θ}},{\mathbf{θ}}^{(t)})}`$ is given by ([](#S3.Ex5 "III-B E-step: Computation of the Q-function ‣ III The EM-MAP algorithm for wavelet domain channel estimation ‣ Wavelet Based Semi-blind Channel Estimation For Multiband OFDM")
III-B
)
<table>
<tbody>
<tr>
<td></td>
<td>𝐠(t+1)=argmax𝐠{−‖⟨𝐇~(t)⟩−𝐓𝐠‖2α2+logπ(𝐠)}.superscript𝐠𝑡1subscriptargmax𝐠superscriptnormdelimited-⟨⟩superscript~𝐇𝑡𝐓𝐠2superscript𝛼2𝜋𝐠{\mathbf{g}}^{(t+1)}=\displaystyle\mathop{\mathrm{arg\,max}}_{\mathbf{g}}\left\{{-\frac{{\|\,\langle\widetilde{\mathbf{H}}^{(t)}\rangle-\mathbf{T}\mathbf{g}\,\|}^{2}}{{\alpha}^{2}}+\log\pi(\mathbf{g})}\right\}.</td>
<td></td>
<td>(14)</td>
</tr>
</tbody>
</table>
Due to the orthonormality of both Fourier and wavelet transforms, $`{\mathbf{T}^{\mathcal{H}}\hspace{0pt}\mathbf{T}} = \mathbf{I}_{L}`$ and we can replace $`{\|{{\langle{\overset{\sim}{\mathbf{H}}}^{(t)}\rangle} - {\mathbf{T}\mathbf{g}}}\|}^{2}`$ by $`{\|{{\overset{\sim}{\mathbf{g}}}^{(t)} - \mathbf{g}}\|}^{2}`$, where
<table>
<tbody>
<tr>
<td></td>
<td>𝐠~(t)superscript~𝐠𝑡\displaystyle\widetilde{\mathbf{g}}^{(t)}</td>
<td>=𝐓ℋ⟨𝐇~(t)⟩absentsuperscript𝐓ℋdelimited-⟨⟩superscript~𝐇𝑡\displaystyle={\mathbf{T}}^{\mathcal{H}}\langle\widetilde{\mathbf{H}}^{(t)}\rangle</td>
<td></td>
<td></td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td></td>
<td>=(1−ρ)𝐠(t)+ρ(𝒟¯𝐒𝐓)ℋ𝐘absent1𝜌superscript𝐠𝑡𝜌superscriptsubscript¯𝒟𝐒𝐓ℋ𝐘\displaystyle=(1-\rho)\,{\mathbf{g}}^{(t)}+\rho\,(\overline{\mathcal{D}}_{{}_{\mathbf{S}}}\mathbf{T})^{\mathcal{H}}\mathbf{Y}</td>
<td></td>
<td>(15)</td>
</tr>
</tbody>
</table>
The M-step can be written as
<table>
<tbody>
<tr>
<td></td>
<td>𝐠(t+1)=argmax𝐠{−‖𝐠~(t)−𝐠‖2α2+logπ(𝐠)}.superscript𝐠𝑡1subscriptargmax𝐠superscriptnormsuperscript~𝐠𝑡𝐠2superscript𝛼2𝜋𝐠\mathbf{g}^{(t+1)}=\displaystyle\mathop{\mathrm{arg\,max}}_{\mathbf{g}}\left\{{-\frac{{\|\,\widetilde{\mathbf{g}}^{(t)}-\mathbf{g}\,\|}^{2}}{\alpha^{2}}+\log\pi(\mathbf{g})}\right\}.</td>
<td></td>
<td>(16)</td>
</tr>
</tbody>
</table>
Actually $`\mathbf{g}^{({t + 1})}`$ in ([14](#S3.E14 "In III-C M-step: Wavelet Based MAP Estimation ‣ III The EM-MAP algorithm for wavelet domain channel estimation ‣ Wavelet Based Semi-blind Channel Estimation For Multiband OFDM")) is no more than the MAP estimate of $`\mathbf{g}`$ from the observation model
<table>
<tbody>
<tr>
<td></td>
<td>𝐠~(t)=𝐠+𝐙1′superscript~𝐠𝑡𝐠subscriptsuperscript𝐙′1\widetilde{\mathbf{g}}^{(t)}=\mathbf{g}+\mathbf{Z}^{\prime}_{1}</td>
<td></td>
<td>(17)</td>
</tr>
</tbody>
</table>
where $`\mathbf{Z}_{1}' = {\mathbf{T}^{\mathcal{H}}\hspace{0pt}\mathbf{Z}_{1}} \sim {\mathcal{C}\hspace{0pt}\mathcal{N}\hspace{0pt}{(\mathbf{0},{\alpha^{2}\hspace{0pt}\mathbf{I}_{L}})}}`$. From the Bayes theorem, the posterior distribution of $`\mathbf{g}`$ is given by
<table>
<tbody>
<tr>
<td></td>
<td>p(𝐠|𝐠~(t))∝p(𝐠~(t)|𝐠)π(𝐠)proportional-to𝑝conditional𝐠superscript~𝐠𝑡𝑝conditionalsuperscript~𝐠𝑡𝐠𝜋𝐠p\left(\mathbf{g}|\widetilde{\mathbf{g}}^{(t)}\right)\;\propto\;p\left(\widetilde{\mathbf{g}}^{(t)}|\mathbf{g}\right)\pi\left(\mathbf{g}\right)</td>
<td></td>
<td>(18)</td>
</tr>
</tbody>
</table>
where $`p\hspace{0pt}{(\left. {\overset{\sim}{\mathbf{g}}}^{(t)} \middle| \mathbf{g} \right.)}`$ is the Gaussian likelihood, $`\overset{\sim}{\mathbf{g}} \sim {\mathcal{C}\hspace{0pt}\mathcal{N}\hspace{0pt}{(\mathbf{g},{\alpha^{2}\hspace{0pt}\mathbf{I}_{L}})}}`$. In this approach, we adopt the Bernoulli-Gaussian prior distribution $`\pi\hspace{0pt}(\mathbf{g})`$ for the wavelet coefficients $`\mathbf{g}`$ of the unknown CIR described by
<table>
<tbody>
<tr>
<td></td>
<td>π(gj)=λδ(gj)+(1−λ)𝒞𝒩gj(0,τ2)𝜋subscript𝑔𝑗𝜆𝛿subscript𝑔𝑗1𝜆𝒞subscript𝒩subscript𝑔𝑗0superscript𝜏2\pi\left(g_{j}\right)=\lambda\;\delta(g_{j})+(1-\lambda)\;\mathcal{CN}_{g_{j}}\left(0,\tau^{2}\right)</td>
<td></td>
<td>(19)</td>
</tr>
</tbody>
</table>
for $`j = {1,\ldots,L}`$, which allows us to model a sparseness property of UWB channels in wavelet domain. This amounts considering that the wavelet coefficients have a probability $`\lambda`$ to be zero and a probability $`1 - \lambda`$ to be distributed as $`\mathcal{C}\hspace{0pt}\mathcal{N}\hspace{0pt}{(0,\tau^{2})}`$. In order to deal with that particular model, we introduce an additional state variable (or indicator) $`\beta_{j} \in {\{ 0,1\}}`$ such that we can express this prior conditionally as
<table>
<tbody>
<tr>
<td></td>
<td>{(gj|βj=0)∼δ(gj)with probabilityλ,(gj|βj=1)∼𝒞𝒩gj(0,τ2)with probability 1−λ.casesconditionalsubscript𝑔𝑗subscript𝛽𝑗0similar-to𝛿subscript𝑔𝑗with probability𝜆missing-subexpressionmissing-subexpressionmissing-subexpressionmissing-subexpressionmissing-subexpressionmissing-subexpressionconditionalsubscript𝑔𝑗subscript𝛽𝑗1similar-to𝒞subscript𝒩subscript𝑔𝑗0superscript𝜏2with probability1𝜆missing-subexpression\left\{\begin{array}[]{lllll}(g_{j}|\beta_{j}=0)&\sim&\delta(g_{j})&\textrm{with probability}\;\lambda,\\ \\ (g_{j}|\beta_{j}=1)&\sim&\mathcal{CN}_{g_{j}}\left(0,\tau^{2}\right)&\textrm{with probability}\;1-\lambda.\\ \end{array}\right.</td>
<td></td>
<td>(20)</td>
</tr>
</tbody>
</table>
This prior model, conditionally on that state variable, leads to a Gaussian posterior for $`g_{j}`$ which makes the estimation explicit; from the direct observation model $`{\overset{\sim}{g}}_{j}^{(t)} = {g_{j} + Z_{1,j}'}`$, we can express these posterior probabilities of $`\beta_{j}`$ as
<table>
<tbody>
<tr>
<td></td>
<td>p(βj=0|g~j(t))=λ𝒩(0,α2)/cp(βj=1|g~j(t))=(1−λ)𝒩(0,α2+τ2)/c𝑝subscript𝛽𝑗conditional0superscriptsubscript~𝑔𝑗𝑡𝜆𝒩0superscript𝛼2𝑐𝑝subscript𝛽𝑗conditional1superscriptsubscript~𝑔𝑗𝑡1𝜆𝒩0superscript𝛼2superscript𝜏2𝑐\begin{array}[]{lll}p\left(\beta_{j}=0|\widetilde{g}_{j}^{(t)}\right)&=&\lambda\;\mathcal{N}\left(0,\alpha^{2}\right)/c\\ p\left(\beta_{j}=1|\widetilde{g}_{j}^{(t)}\right)&=&(1-\lambda)\;\mathcal{N}\left(0,\alpha^{2}+\tau^{2}\right)/c\end{array}</td>
<td></td>
<td>(21)</td>
</tr>
</tbody>
</table>
where the constant $`c = {{\lambda\hspace{0pt}\mathcal{N}\hspace{0pt}\left( 0,\alpha^{2} \right)} + {{({1 - \lambda})}\hspace{0pt}\mathcal{N}\hspace{0pt}\left( 0,{\alpha^{2} + \tau^{2}} \right)}}`$. From this set of equations, we easily notice that the indicator variable $`\beta_{j}`$ allows us to discriminate between the noise coefficients (for $`\beta_{j} = 0`$) and the effective channel wavelet coefficients (for $`\beta_{j} = 1`$), eventually corrupted by noise. The indicator variables $`\beta_{j}`$ are estimated, in the MAP sense, by
<table>
<tbody>
<tr>
<td></td>
<td>βj(t+1)={0,if p(βj=0|g~j(t))≥0.51,elsewhere.superscriptsubscript𝛽𝑗𝑡1cases0if 𝑝subscript𝛽𝑗conditional0superscriptsubscript~𝑔𝑗𝑡0.5missing-subexpressionmissing-subexpression1elsewhere\beta_{j}^{(t+1)}=\left\{\begin{array}[]{ll}0,&\textrm{if }\;p\left(\beta_{j}=0|\widetilde{g}_{j}^{(t)}\right)\geq 0.5\\ \\ 1,&\textrm{elsewhere}.\end{array}\right.</td>
<td></td>
<td>(22)</td>
</tr>
</tbody>
</table>
Therefore, the MAP estimates of the channel wavelet coefficients are obtained by a simple denoising/thresholding rule as
<table>
<tbody>
<tr>
<td></td>
<td>gj(t+1)={0,if βj(t+1)=0τ2α2+τ2g~j(t+1),if βj(t+1)=1superscriptsubscript𝑔𝑗𝑡1cases0if superscriptsubscript𝛽𝑗𝑡10superscript𝜏2superscript𝛼2superscript𝜏2superscriptsubscript~𝑔𝑗𝑡1if superscriptsubscript𝛽𝑗𝑡11g_{j}^{(t+1)}=\left\{\begin{array}[]{ll}0,&\textrm{if }\;\beta_{j}^{(t+1)}=0\\ \displaystyle\frac{\tau^{2}}{\alpha^{2}+\tau^{2}}\,\,\widetilde{g}_{j}^{(t+1)},&\textrm{if }\;\beta_{j}^{(t+1)}=1\end{array}\right.</td>
<td></td>
<td>(23)</td>
</tr>
</tbody>
</table>
####
III-C1
$`\tau`$ and $`\lambda`$ updating
The prior parameters $`\tau`$ and $`\lambda`$ stand respectively for the (significant)-wavelet coefficients energy and unsignificant coefficient probability. The update rules for these two parameters are MAP based rules derived from assigning conjugate priors to these parameters \[[10](#bib.bib10)\]:
<table>
<tbody>
<tr>
<td></td>
<td>λ^=(L~−1/2)/(L−1),τ^2=η/(L−L~)^𝜆~𝐿12𝐿1superscript^𝜏2𝜂𝐿~𝐿\begin{array}[]{lll}\hat{\lambda}&=&{(\widetilde{L}-1/2)}/{(L-1)},\\ \hat{\tau}^{2}&=&{\eta}/{(L-\widetilde{L})}\end{array}</td>
<td></td>
<td>(24)</td>
</tr>
</tbody>
</table>
where $`\overset{\sim}{L} = {\text{Card}\hspace{0pt}\left. \{ j \middle| {\beta_{j} = 0}\} \right.}`$ and $`\eta = {\sum_{\beta_{j} = 1}\left| g_{j}^{({t + 1})} \right|^{2}}`$; Card{.} denoting the set cardinality.
####
III-C2
Reduction of the number of estimated parameters
The thresholding procedure derived in this section, provides an easy framework for reducing the number of estimated coefficients. This can be done by discarding at each iteration, the elements of $`\mathbf{g}^{({t + 1})}`$ that are replaced by zero in ([23](#S3.E23 "In III-C M-step: Wavelet Based MAP Estimation ‣ III The EM-MAP algorithm for wavelet domain channel estimation ‣ Wavelet Based Semi-blind Channel Estimation For Multiband OFDM")). The underlying assumption is as follows: whenever the estimator attributes an unknown wavelet coefficient to noise (replace it by zero), this coefficient will always be considered as noise and so will not be estimated in future iterations.
This operation is shown in figure [3](#S4.F3 "Figure 3 ‣ IV decoding method and implementation issues ‣ Wavelet Based Semi-blind Channel Estimation For Multiband OFDM") and can be modeled as:
<table>
<tbody>
<tr>
<td></td>
<td>𝐠tr(t+1)=Θ(𝐠(t+1)),𝐓tr=Ξ(𝐓)formulae-sequencesubscriptsuperscript𝐠𝑡1trΘsuperscript𝐠𝑡1subscript𝐓trΞ𝐓\mathbf{g}^{(t+1)}_{\mathrm{tr}}=\Theta\big{(}\,{\mathbf{g}}^{(t+1)}\,\big{)},\quad\mathbf{T}_{\mathrm{tr}}=\Xi\big{(}\,\mathbf{T}\,\big{)}</td>
<td></td>
<td>(25)</td>
</tr>
</tbody>
</table>
where the truncation operator $`\Theta{(.)}`$ gathers in $`\mathbf{g}_{tr}^{({t + 1})}`$ the components of $`\mathbf{g}^{({t + 1})}`$ that must be kept and the operator $`\Xi{(.)}`$ constructs $`\mathbf{T}_{tr}`$ from $`\mathbf{T}`$ by keeping the rows corresponding to kept indexes. During the first iteration ($`t = 0`$), the algorithm does not perform any truncation and the EM algorithm estimates all coefficients. However, after each M-step, the number of unknown parameters to be estimated in the next iteration is reduced according ([25](#S3.E25 "In III-C2 Reduction of the number of estimated parameters ‣ III-C M-step: Wavelet Based MAP Estimation ‣ III The EM-MAP algorithm for wavelet domain channel estimation ‣ Wavelet Based Semi-blind Channel Estimation For Multiband OFDM")) by using $`\mathbf{g}_{tr}^{({t + 1})}`$ and $`\mathbf{T}_{tr}`$ in the update formula of the E-step ([13](#S3.E13 "In III-B E-step: Computation of the Q-function ‣ III The EM-MAP algorithm for wavelet domain channel estimation ‣ Wavelet Based Semi-blind Channel Estimation For Multiband OFDM")).
## IV decoding method and implementation issues
According to equation ([10](#S3.E10 "In III-B E-step: Computation of the Q-function ‣ III The EM-MAP algorithm for wavelet domain channel estimation ‣ Wavelet Based Semi-blind Channel Estimation For Multiband OFDM")), we make use of the information on transmitted symbols, obtained from the decoder, to update the channel estimate at each iteration. Besides, the decoder requires an estimate of the channel in order to provide the probability of encoded bits. Hence, the semi-blind channel estimation algorithm is naturally combined with the process of data decoding. The a posteriori probability of the unknown symbol $`S_{k}`$, $`p\hspace{0pt}{(\left. S_{k} \middle| {Y_{k},{\hat{H}}_{k}^{(t)}} \right.)}`$, is calculated using the a posteriori probabilities provided by the decoder at the end of the $`t`$-th iteration as
<table>
<tbody>
<tr>
<td></td>
<td>p(Sk|Yk,H^)=∏i=1BPdec(ck,i)𝑝conditionalsubscript𝑆𝑘subscript𝑌𝑘^𝐻superscriptsubscriptproduct𝑖1𝐵subscript𝑃decsubscript𝑐𝑘𝑖p(S_{k}|Y_{k},\widehat{H})=\prod_{i=1}^{B}P_{\rm dec}(c_{k,i})</td>
<td></td>
<td>(26)</td>
</tr>
</tbody>
</table>
where $`P_{dec}\hspace{0pt}{(c_{k,i})}`$ is the a posteriori probability corresponding to the $`i`$-th bit of $`S_{k}`$, $`c_{k,i}`$. At the first iteration, where no a priori information is available on bits $`c_{k,i}`$, $`P_{dec}\hspace{0pt}{(c_{k,i})}`$ are set to $`1/2`$.
\psfrag{A}{\$\mathbf{\Theta}\$}\psfrag{B}{\parbox{95.00014pt}{\centering\small{EM-MAP Estimation\\ of Channel Wavelet\\ Coefficients}\\add@centering}}\psfrag{I}{\$\mathbf{g}^{(t)}\$}\psfrag{H}{\\\\\$\mathbf{g}\_{\rm tr}^{(t)}\$}\psfrag{C}{\parbox{75.0001pt}{\centering{\small{Soft Demapping\\ \\\\ SISO Decoder}}\\add@centering}}\psfrag{K}{\scriptsize\$P\_{\rm dec}^{(t)}(c\_{k,i})\$}\psfrag{L}{\scriptsize Uncoded Bits Probabilities}\psfrag{M}{\scriptsize\quad\quad at last iteration}\psfrag{D}{\scriptsize\\\\ Observed Frame}\psfrag{E}{\scriptsize\$\mathbf{g}^{(0)}\$}\psfrag{F}{\scriptsize\$P\_{\rm dec}^{(0)}(c\_{k,i})=.5\$}\psfrag{G}{ \parbox{28.00006pt}{\centering\scriptsize Observed\\ Frame\\add@centering}}\psfrag{J}{ \parbox{28.00006pt}{\centering\scriptsize Decoded\\ Bits\\add@centering}}\includegraphics\[width=199.4681pt,height=119.50148pt\]{EM_MAP_RXII}
Figure 3: EM-MAP channel estimation combined with the decoding process.
Among several possible ways to practically implement a joint channel estimation and decoding receiver, we adopt the following global procedure (see figure [3](#S4.F3 "Figure 3 ‣ IV decoding method and implementation issues ‣ Wavelet Based Semi-blind Channel Estimation For Multiband OFDM")).
- •
Initialization ($`t = 0`$)
- –
Set all probabilities of coded bits $`P_{dec}\hspace{0pt}{(c_{k,i})}`$ to $`1/2`$ and derive $`p\hspace{0pt}{(\left. \mathbf{S} \middle| {\mathbf{Y},\mathbf{g}^{(0)}} \right.)}`$ according to ([26](#S4.E26 "In IV decoding method and implementation issues ‣ Wavelet Based Semi-blind Channel Estimation For Multiband OFDM")).
- –
Initialize the unknown vector $`\mathbf{g}`$ by $`\mathbf{g}^{(0)}`$ obtained from pilot symbols.
- •
for $`t = {1,\ldots,t_{\max}}`$
- –
Use the current estimate $`\mathbf{g}^{(t)}`$ to calculate $`\mathbf{g}^{({t + 1})}`$ according to ([23](#S3.E23 "In III-C M-step: Wavelet Based MAP Estimation ‣ III The EM-MAP algorithm for wavelet domain channel estimation ‣ Wavelet Based Semi-blind Channel Estimation For Multiband OFDM")).
- –
Discard the wavelet coefficients that are replaced by zero for the next iteration by evaluating $`\mathbf{g}_{tr}^{(t)}`$ and $`\mathbf{T}_{tr}`$ from ([25](#S3.E25 "In III-C2 Reduction of the number of estimated parameters ‣ III-C M-step: Wavelet Based MAP Estimation ‣ III The EM-MAP algorithm for wavelet domain channel estimation ‣ Wavelet Based Semi-blind Channel Estimation For Multiband OFDM")).
- –
if $`t \neq t_{\max}`$: Use the current estimate $`\mathbf{g}_{tr}^{(t)}`$ to update the probability of encoded bits $`P_{dec}\hspace{0pt}{(c_{k,i})}`$ and derive $`p\hspace{0pt}{(\left. \mathbf{S} \middle| {\mathbf{Y},\mathbf{g}^{(t)}} \right.)}`$ from ([26](#S4.E26 "In IV decoding method and implementation issues ‣ Wavelet Based Semi-blind Channel Estimation For Multiband OFDM")).
else: Decode the information data by thresholding the uncoded bit probabilities with 1/2.
## V Simulation results
In this section we present a comparative performance study of the proposed EM-MAP algorithm. The binary information data are encoded by a non-recursive non-systematic convolutional encoder with rate $`R = {1/2}`$ and constraint length 3. Each frame has a payload of 1 KB along with 3 pilot symbols at the beginning for initializing the channel of each subband. The interleaver is random and operates over the entire frame. Among different wavelet families, “symmetric” wavelet basis functions \[[11](#bib.bib11)\] providing the sparser representation \[[9](#bib.bib9)\] have been considered.Unless otherwise mentioned, the curves are obtained after $`t_{\max} = 4`$ iterations.
First, a sparse channel model where only 20 wavelet coefficients out of total 96 have non zero values, is considered. The second channel, referred to as Corridor, is a line of sight (LOS) scenario issued from realistic UWB indoor channel measurements \[[12](#bib.bib12)\] where the receive and transmit antennas are located in a corridor separated by 9 meters.
Performance comparison is made with two pilot-only based approach using ML and minimum mean square error (MMSE) channel estimation, referred to as pilot-ML and pilot-MMSE. We also compare the proposed algorithm with two semi-blind channel estimation based on the EM algorithm, called respectively EM-Freq and EM-Wav. The first approach, consists of estimating the channel over all of the three subbands, using the model ([3](#S2.E3 "In II System model and wavelet domain problem formulation ‣ Wavelet Based Semi-blind Channel Estimation For Multiband OFDM")), similar to \[[5](#bib.bib5)\] while the second scheme is a wavelet domain EM based estimation of the channel where the prior model is set to have a uniform distribution.
Figure [4](#S5.F4 "Figure 4 ‣ V Simulation results ‣ Wavelet Based Semi-blind Channel Estimation For Multiband OFDM") depicts the mean square error (MSE) between true and estimated channel as a function of $`E_{b}/N_{0}`$. It can be noticed that, although the pilot-MMSE approach improves the estimation accuracy for low SNR values, the performance of pilot based channel estimation methods are very far from the family of semi-blind methods. Comparing the wavelet domain semi-blind approach (EM-Wav) and the frequency domain approach (EM-freq), shows that significant gain is achieved by the former method. As shown, the best performance is achieved by the EM-MAP method. We see that by using EM-MAP, a gain of almost 4 dB in SNR is achieved at MSE=$`2 \times 10^{- 3}`$, as compared to the EM-Wav method. This clearly shows the adequacy of the EM-MAP method for the case where the unknown channel has few non zero wavelet coefficients, which is in perfect agreement with the prior model.
Figure [5](#S5.F5 "Figure 5 ‣ V Simulation results ‣ Wavelet Based Semi-blind Channel Estimation For Multiband OFDM") shows the BER results along with the BER for the case of perfect channel state information (CSI). It can be seen that at a BER of $`10^{- 3}`$, the pilot-ML and the EM-Freq approaches are respectively $`3.9`$ and $`2`$ dB of SNR far from the BER obtained with the perfect channel. Furthermore, the performance of the Pilot-MMSE approach is not shown since it was very close to that of Pilot-ML. Also, we observe that wavelet based semi-blind methods perform closely to the perfect CSI case. For example, at BER=$`10^{- 4}`$, the EM-MAP and EM-Wav method have respectively about 0.2 dB and 0.5 dB of SNR degradation from the performance obtained with perfect CSI.
[Refer to caption](/html/0708.1414/assets/x1.png)
Figure 4: Mean square error between the true and estimated coefficients for the sparse channel model.
[Refer to caption](/html/0708.1414/assets/x2.png)
Figure 5: BER performance of the EM-MAP method over the sparse channel model.
We now evaluate the performance of EM-MAP by considering the Corridor channel. Figure [6](#S5.F6 "Figure 6 ‣ V Simulation results ‣ Wavelet Based Semi-blind Channel Estimation For Multiband OFDM"), shows that wavelet based methods again outperforms pilot based and EM-Freq methods in terms of MSE and BER. However, the EM-MAP performance is now comparable to that of EM-Wav method. This can be explained by noting that when the channel is not sparse, small values are attributed to $`\lambda`$ by the algorithm (see ([24](#S3.E24 "In III-C1 𝜏 and 𝜆 updating ‣ III-C M-step: Wavelet Based MAP Estimation ‣ III The EM-MAP algorithm for wavelet domain channel estimation ‣ Wavelet Based Semi-blind Channel Estimation For Multiband OFDM"))). This leads to a gaussian prior model with a large variance compared to the noise variance, which can be approximated with a uniform prior. As a results, the prior becomes less informative and the EM-MAP performs close to EM-Wav, as shown in figures [6](#S5.F6 "Figure 6 ‣ V Simulation results ‣ Wavelet Based Semi-blind Channel Estimation For Multiband OFDM"). Thus, the proposed EM-MAP algorithm is able to adapt its prior model parameters for each propagation environment.
[Refer to caption](/html/0708.1414/assets/x3.png)
Figure 6: Mean square error between the true and estimated coefficients over the Corridor channel.
Figure [7](#S5.F7 "Figure 7 ‣ V Simulation results ‣ Wavelet Based Semi-blind Channel Estimation For Multiband OFDM") shows the average number of estimated parameters versus the iteration number different channel scenarios. As observed, the EM-MAP approach tends to reduce significantly the number of estimated parameters. This can be seen for the sparse channel where the number of estimated parameters is reduced up to 20 parameters at the fifth iteration. Furthermore, under non-sparse Corridor channel, the figure shows that EM-MAP method is preferred to EM-Wav, due to its lower computational load.
[Refer to caption](/html/0708.1414/assets/x4.png)
Figure 7: Reduction of the number of estimated parameters through iterations, $`{{{E\hspace{0pt}b}/N}\hspace{0pt}0} = 8`$ dB.
## VI Conclusion
This paper proposed a semi-blind MAP channel estimation algorithm that integrates the advantages of wavelet based estimation. The investigated method naturally combines the EM iterations with the decoding process. We derived an equivalent data model for the multiband OFDM system involving the channel over all 3 subbands expressed in the wavelet domain. By choosing a Bernoulli-Gaussian prior distribution for the channel wavelet coefficients, the MAP estimator yields a thresholding procedure at the M-step of the EM algorithm which we used to reduce the number of estimated coefficients. With only few iterations, the EM-MAP method provides significant reduction in the number of estimated parameters and outperforms all considered pilot based and semi-blind methods.
## References
- \[1\] FCC, “First report and order, revision of part 15 of the commission’s rules regarding ultra-wideband transmission systems,” Tech. Rep., Feb. 2004.
- \[2\] S. Roy, J. R. Foerster, V. S. Somayazulu, and D. G. Leeper, “Ultrawideband radio design: The promise of high-speed, short range wireless conectivity,”
Proceedings of the IEEE, vol. 92, February 2004.
- \[3\] A. Batra, J. Balakrishnan, G. R. Aiello, J. R. Foerster, and A. Dabak, “Design of multiband OFDM system for realistic UWB channel environments,”
IEEE transaction on microwave theory and techniques, vol. 52, pp. 2123–2138, september 2004.
- \[4\] R. Prasad,
OFDM for Wireless Communications Systems,
Artech House Publishers, september 2004.
- \[5\] X. Ma, H. Kobayashi, and S. C. Schawrtz, “EM-based channel estimation algorithms for OFDM,”
EURASIP Journal on Applied Signal Processing, vol. 10, pp. 1460–1477, 2004.
- \[6\] T. Y. Al-Naffouri, A. Bahai, and A. Paulraj, “Semi-blind channel identification and equalization in OFDM: An expectation-maximization approach,”
in in Proc. IEEE Vehicular Technology Conference, 2002, pp. 13–17.
- \[7\] G. J. McLachlan and T. Krishnan,
The EM Algorithm and Extensions,
Wiley Series in probability and Statistics, Wiley, 1997.
- \[8\] I. M. Johnstone and B. W. Silverman, “Neddles and straw in haystacks: Empirical bayes estimates of possibly sparse sequences,”
Annals of Statistics, vol. 32, pp. 1594–1649, 2004.
- \[9\] S. Sadough and E. Jaffrot, “A wavelet packet based model for an ultra-wideband propagation channel,”
in Proc. ECPS 2005, Brest, France, March 15-18. 2005.
- \[10\] J. N. Bernardo and A. F. M. Smith,
Bayesian Theory,
Wiley ans Sons, Chichester, England, 1994.
- \[11\] S. Mallat,
A Wavelet Tour of Signal Processing,
Academic Press, 1999.
- \[12\] S. Bories, A. Sibille, and C. Roblin, “UWB indoor channel measurement,”
in Proc. IWAT, March 2005, pp. 466–469. |
0708.1411 | cs.NI | cyber | # ACHIEVABLE OUTAGE RATES WITH IMPROVED DECODING OF BICM MULTIBAND OFDM UNDER CHANNEL ESTIMATION ERRORS
###### Abstract
We consider the decoding of bit interleaved coded modulation (BICM) applied to multiband OFDM for practical scenarios where only a noisy (possibly very bad) estimate of the channel is available at the receiver. First, a decoding metric based on the channel a posteriori probability density, conditioned on the channel estimate is derived and used for decoding BICM multiband OFDM. Then, we characterize the limits of reliable information rates in terms of the maximal achievable outage rates associated to the proposed metric. We also compare our results with the outage rates of a system using a theoretical decoder. Our results are useful for designing a communication system where a prescribed quality of service (QoS), in terms of achievable target rates with small error probability, must be satisfied even in the presence of imperfect channel estimation. Numerical results over both realistic UWB and theoretical Rayleigh fading channels show that the proposed method provides significant gain in terms of BER and outage rates compared to the classical mismatched detector, without introducing any additional complexity.
## 1 Introduction
Ultra-Wide-Band (UWB) is defined as any wireless transmission scheme that occupies a bandwidth of more than 25 % of its center frequency or greater than 500 MHz over the 3.1-10.6 GHz frequency band \[[1](#bib.bib1)\]. Multiband Orthogonal Frequency division multiplexing (MB-OFDM) \[[2](#bib.bib2)\] is a spectrally efficient technique proposed for high data rate, short range UWB applications. This approach uses a conventional OFDM system, combined with bit interleaved coded modulation (BICM) and frequency hopping for improved diversity and multiple access. In MB-OFDM, the channel is assumed changing so slowly that it is considered time invariant during the transmission of an entire frame. Channel estimation is performed by one known symbol (pilot) transmitted at the beginning of the information frame while the rest of the frame is decoded based on the estimated channel. Due to the limited number of pilots, the estimate of the channel is imperfect and the receiver has only access to this noisy channel estimate. However, the receiver/decoder metric for any maximum-likelihood (ML) based detector, requires knowledge of the exact channel.
A standard sub-optimal technique, known as mismatched ML decoding, consists in replacing the exact channel by its estimate in the receiver metric. Hence, the resulting decoding metric is not adapted to the presence of channel estimation errors (CEE). This practical scenario, with mismatched decoding, leads to the following important questions. Firstly, in presence of CEE with a given amount of training, what are the limits of reliable transmission (the capacity). Secondly, what type of encoder/decoder is necessary to transmit reliable information close to the performance limits. The first problem has been recently addressed in \[[3](#bib.bib3)\], characterizing the maximal reliable information rate, by using the notion of estimation induced outage capacity. Unfortunately, the theoretical encoder/decoder used to achieve this capacity can not be implemented on practical communication systems. Besides, the mismatched ML decoding has been showed to be largely suboptimal for the considered class of channels. Basically, the transmitter and the receiver strive to construct codes for ensuring reliable communication with a quality of service (QoS), no matter what degree of channel estimation accuracy arises during the transmission. The QoS requirements stand for achieving target rates with small error probability even with very bad channel estimates. In this paper, we propose a practical decoding metric for the aforementioned mismatch scenario. This metric uses the estimated channel and the a posteriori pdf characterizing the channel estimation process, which matches well the channel knowledge available at the receiver. Based on the derived metric, we formulate our decoding rule for BICM MB-OFDM. Interestingly, the present metric coincides with that derived for space-time decoding from independent results in \[[4](#bib.bib4)\]. In order to determine the limits of reliable information rates associated to the proposed metric, we use the complementary results obtained in \[[5](#bib.bib5)\]. This allows us to compare the maximal supported rate associated to our metric with that of the classical mismatched ML and the theoretical decoder. Our results are relevant for communication systems where a prescribed quality of service (given by the outage probability) must be ensured even in the presence of imperfect channel estimation.
The outline of this paper is as follows. In Section <a href="#S2" style="font-size:90%;" title="2 TRANSMISSION MODEL ‣ ACHIEVABLE OUTAGE RATES WITH IMPROVED DECODING OF BICM MULTIBAND OFDM UNDER CHANNEL ESTIMATION ERRORS">2</a> we describe the system model for MB-OFDM transmission over a frequency selective fading channel. Section <a href="#S3" style="font-size:90%;" title="3 PILOT ASSISTED CHANNEL ESTIMATION ‣ ACHIEVABLE OUTAGE RATES WITH IMPROVED DECODING OF BICM MULTIBAND OFDM UNDER CHANNEL ESTIMATION ERRORS">3</a> presents the pilot assisted channel estimation: we specify the statistics of the CEE and then calculate the posterior distribution of the perfect channel conditioned on the estimated channel. This posterior distribution is used in section <a href="#S4" style="font-size:90%;" title="4 Maximum-likelihood detection metric in the presence of channel estimation errors ‣ ACHIEVABLE OUTAGE RATES WITH IMPROVED DECODING OF BICM MULTIBAND OFDM UNDER CHANNEL ESTIMATION ERRORS">4</a> to derive the ML decoding metric in the presence of imperfect channel state information at the receiver (CSIR). In section <a href="#S5" style="font-size:90%;" title="5 Decoding of BICM-OFDM based on imperfect channel estimation ‣ ACHIEVABLE OUTAGE RATES WITH IMPROVED DECODING OF BICM MULTIBAND OFDM UNDER CHANNEL ESTIMATION ERRORS">5</a>, we use the general modified metric for soft decoding BICM MB-OFDM systems. In section <a href="#S6" style="font-size:90%;" title="6 Achievable Outage Rates Associated to the Proposed Metric ‣ ACHIEVABLE OUTAGE RATES WITH IMPROVED DECODING OF BICM MULTIBAND OFDM UNDER CHANNEL ESTIMATION ERRORS">6</a>, we derive the achievable outage rates of a receiver using the proposed metric. Section <a href="#S7" style="font-size:90%;" title="7 Simulation results ‣ ACHIEVABLE OUTAGE RATES WITH IMPROVED DECODING OF BICM MULTIBAND OFDM UNDER CHANNEL ESTIMATION ERRORS">7</a> illustrates via simulations the performance of the proposed receiver in realistic UWB channel environments and section <a href="#S8" style="font-size:90%;" title="8 Conclusion ‣ ACHIEVABLE OUTAGE RATES WITH IMPROVED DECODING OF BICM MULTIBAND OFDM UNDER CHANNEL ESTIMATION ERRORS">8</a> concludes the paper.
Notational conventions are as follows : upper case bold symbols denote vectors, $`\mathbf{I}_{N}`$ represents an ($`N \times N`$) identity matrix; $`\left( . \right)^{T}`$ and $`\left( . \right)^{\mathcal{H}}`$ denote vector transpose and Hermitian transpose, respectively.
## 2 TRANSMISSION MODEL
An MB-OFDM system divides the spectrum between 3.1 to 10.6 GHz into several non-overlapping subbands each one occupying approximately 500 MHZ of bandwidth \[[2](#bib.bib2)\]. Information is transmitted using OFDM modulation over one of the subbands in a particular time-slot. The transmitter architecture is depicted in figure <a href="#S2.F1" style="font-size:90%;" title="Figure 1 ‣ 2 TRANSMISSION MODEL ‣ ACHIEVABLE OUTAGE RATES WITH IMPROVED DECODING OF BICM MULTIBAND OFDM UNDER CHANNEL ESTIMATION ERRORS">1</a>: each bit $`b_{i}`$ is convolutionally encoded into two bits $`c_{i}^{0}`$ and $`c_{i}^{1}`$ which are interleaved in order to break the error bursts. The interleaved bits are gathered in subsequences of $`B`$ bits $`d_{k}^{1},\ldots,d_{k}^{B}`$ and mapped to complex M-QAM
<sup>1</sup>
<sup>1</sup>1Here, 16-QAM mapping is used instead of QPSK, proposed in \[[6](#bib.bib6)\].
($`M = 2^{B}`$) symbols $`S_{k}`$ with average energy $`E_{s} = {{\mathbb{E}}\hspace{0pt}\left\lbrack \left| S_{k} \right|^{2} \right\rbrack}`$. At a particular time-slot, a time-frequency code (TFC) selects the center frequency of the subband over which the OFDM symbol is sent. The TFC is used not only to provide frequency diversity but also to distinguish between multiple users.
At the receiver, assuming a cyclic prefix (CP) longer than the channel maximum delay spread and after performing fast Fourier transform (FFT), OFDM converts the the channel into $`M`$ parallel Rayleigh distributed flat fading subchannels, where $`M`$ is the number of subcarriers. The baseband equivalent observation model over a given subband can be written as
<table>
<tbody>
<tr>
<td></td>
<td>𝐘=𝐃𝐇𝐒+𝐙,𝐘subscript𝐃𝐇𝐒𝐙{\mathbf{Y}}=\mathbf{D}_{{}_{\mathbf{H}}}\,{\mathbf{S}}+{\mathbf{Z}},</td>
<td></td>
<td>(1)</td>
</tr>
</tbody>
</table>
where $`\left( {M \times 1} \right)`$ vectors $`\mathbf{Y}`$ and $`\mathbf{S}`$ denote received and transmitted symbols, respectively; the noise block $`\mathbf{Z}`$ is assumed to be a circularly symmetric complex Gaussian random vector with distribution $`\mathbf{Z} \sim {\mathcal{C}\hspace{0pt}\mathcal{N}\hspace{0pt}\left( \mathbf{0},{\sigma_{z}^{2}\hspace{0pt}\mathbf{I}_{M}} \right)}`$; and $`\mathbf{D}_{_{\mathbf{H}}}`$ is a diagonal matrix with diagonal elements equal to $`\mathbf{H} = \left\lbrack H_{1},\ldots,H_{N} \right\rbrack^{T}`$, where $`\mathbf{H}`$ is the vector of channel FFT coefficients.
\psfrag{A}{Binary Data}\psfrag{B}{{\parbox{54.00009pt}{\centering\$R=1/2\$ Convolutional Encoder\\add@centering}}}\psfrag{C}{{\parbox{54.00009pt}{\centering Bit\\ interleaver\\add@centering}}}\psfrag{D}{\parbox{54.00009pt}{\centering M-QAM\\ Mapping\\add@centering}}\psfrag{E}{\parbox{54.00009pt}{\centering IFFT\\ CP \\ GI\\ Addition\\add@centering}}\psfrag{F}{\parbox{54.00009pt}{\centering DAC\\add@centering}}\psfrag{G}{\parbox{54.00009pt}{\centering\$\exp\left(j2\pi f\_{c}t\right)\$\\add@centering}}\psfrag{H}{{TFC: Subband Selection}}\includegraphics\[width=227.90501pt,height=113.81102pt\]{TXchainII}
Fig. 1:
TX architecture of the multiband OFDM system.
## 3 PILOT ASSISTED CHANNEL ESTIMATION
Under the assumption of time-invariant channel characteristics over the entire transmitted frame, channel estimation is usually performed by transmitting known training symbols at the beginning of each frame.
Assume that we are interested to estimate the $`k`$-th, $`k = \left\{ 1,\ldots, \right.`$
$`\left. M \right\}`$, fade coefficient $`H_{k}`$ via the transmission of $`N`$ pilot symbols and let $`\mathbf{S}_{_{T}} = \left\lbrack {\overset{\sim}{S}}_{1},\ldots,{\overset{\sim}{S}}_{N} \right\rbrack^{T}`$ be the ($`N \times 1`$) vector of transmitted training sequence. We will assume constant modulus training symbols for the $`k`$-th subcarrier, i.e., $`\left| {\overset{\sim}{S}}_{1} \right|^{2} = \ldots = \left| {\overset{\sim}{S}}_{N} \right|^{2} = P_{T,k}`$. From (<a href="#S2.E1" style="font-size:90%;" title="In 2 TRANSMISSION MODEL ‣ ACHIEVABLE OUTAGE RATES WITH IMPROVED DECODING OF BICM MULTIBAND OFDM UNDER CHANNEL ESTIMATION ERRORS">1</a>) we have
<table>
<tbody>
<tr>
<td></td>
<td>𝐘T=𝐃𝐇k𝐒T+𝐙T,subscript𝐘𝑇subscript𝐃subscript𝐇𝑘subscript𝐒𝑇subscript𝐙𝑇{\mathbf{Y}}_{{}_{T}}=\mathbf{D}_{{}_{{\mathbf{H}}_{k}}}\,{\mathbf{S}}_{{}_{T}}+{\mathbf{Z}}_{{}_{T}},</td>
<td></td>
<td>(2)</td>
</tr>
</tbody>
</table>
where $`\mathbf{D}_{_{\mathbf{H}_{k}}}`$ is a diagonal matrix with diagonal elements equal to the $`\left( {N \times 1} \right)`$ vector $`\mathbf{H}_{k} = \left\lbrack H_{k},\ldots,H_{k} \right\rbrack^{T}`$ and $`\mathbf{Z}_{_{T}}`$ is a noise vector with the same distribution as $`\mathbf{Z}`$ affecting the transmission of pilot symbols. The ML estimate of $`H_{k}`$ is obtained by maximizing the likelihood function $`p\hspace{0pt}\left( \mathbf{Y}_{_{T}} \middle| {\mathbf{H}_{k},\mathbf{S}_{_{T}}} \right)`$. After some standard calculus we have
<table>
<tbody>
<tr>
<td></td>
<td>H^kML=𝐒Tℋ𝐘TNPT,k=Hk+ℰk,k=1,…,Mformulae-sequencesubscriptsuperscript^𝐻ML𝑘superscriptsubscript𝐒𝑇ℋsubscript𝐘𝑇𝑁subscript𝑃𝑇𝑘subscript𝐻𝑘subscriptℰ𝑘𝑘1…𝑀{{\widehat{H}}}^{\mathrm{ML}}_{k}=\frac{{\mathbf{S}}_{{}_{T}}^{\mathcal{H}}\,{\mathbf{Y}}_{{}_{T}}}{N\,P_{T,k}}=H_{k}+\mathcal{E}_{k},\quad\quad k=1,\ldots,M</td>
<td></td>
<td>(3)</td>
</tr>
</tbody>
</table>
where $`\mathcal{E}_{k} \sim {\mathcal{C}\hspace{0pt}\mathcal{N}\hspace{0pt}\left( 0,\sigma_{\mathcal{E}_{k}}^{2} \right)}`$ is the estimation error. From equations (<a href="#S3.E3" style="font-size:90%;" title="In 3 PILOT ASSISTED CHANNEL ESTIMATION ‣ ACHIEVABLE OUTAGE RATES WITH IMPROVED DECODING OF BICM MULTIBAND OFDM UNDER CHANNEL ESTIMATION ERRORS">3</a>), it is obvious that $`\sigma_{\mathcal{E}_{k}}^{2} = {SNR}_{T,k}^{- 1}`$, where $`{SNR}_{T,k} \triangleq \frac{N\hspace{0pt}P_{T,k}}{\sigma_{z}^{2}}`$.
Assuming that $`H_{k}`$s are i.i.d. and distributed as $`\mathcal{C}\hspace{0pt}\mathcal{N}\hspace{0pt}\left( 0,\sigma_{h}^{2} \right)`$, the posterior distribution of the perfect channel conditioned on the estimated channel coefficient is given by $`{p\hspace{0pt}\left( H_{k} \middle| {\hat{H}}_{k} \right)} \propto {p\hspace{0pt}\left( {\hat{H}}_{k} \middle| H_{k} \right)\hspace{0pt}p\hspace{0pt}\left( H_{k} \right)}`$. After standard manipulation of Gaussian densities we obtain
<table>
<tbody>
<tr>
<td></td>
<td>p(Hk|H^k)=𝒞𝒩(ρH^k,ρσz2NPT,k),𝑝conditionalsubscript𝐻𝑘subscript^𝐻𝑘𝒞𝒩𝜌subscript^𝐻𝑘𝜌subscriptsuperscript𝜎2𝑧𝑁subscript𝑃𝑇𝑘p(H_{k}|{\widehat{H}}_{k})=\,\mathcal{CN}\left(\,\rho{\widehat{H}}_{k}\,,\frac{\rho\,{\sigma}^{2}_{z}}{NP_{T,k}}\right),</td>
<td></td>
<td>(4)</td>
</tr>
</tbody>
</table>
where $`\rho_{k} \triangleq \left. \sigma_{h}^{2}/\left( {\sigma_{h}^{2} + \sigma_{\mathcal{E}_{k}}^{2}} \right) \right.`$.
The availability of the estimation error distribution constitutes an interesting feature of pilot assisted channel estimation that we used to derive the posterior distribution (<a href="#S3.E4" style="font-size:90%;" title="In 3 PILOT ASSISTED CHANNEL ESTIMATION ‣ ACHIEVABLE OUTAGE RATES WITH IMPROVED DECODING OF BICM MULTIBAND OFDM UNDER CHANNEL ESTIMATION ERRORS">4</a>). Next, we see how this additional information is exploited in a modified ML metric for improving the detection process.
## 4 Maximum-likelihood detection metric in the presence of channel estimation errors
### 4.1 Mismatched ML metric
The observation model (<a href="#S2.E1" style="font-size:90%;" title="In 2 TRANSMISSION MODEL ‣ ACHIEVABLE OUTAGE RATES WITH IMPROVED DECODING OF BICM MULTIBAND OFDM UNDER CHANNEL ESTIMATION ERRORS">1</a>) can be rewritten in a component-wise form as
<table>
<tbody>
<tr>
<td></td>
<td>Yk=HkSk+Zk,k=1,…,Mformulae-sequencesubscript𝑌𝑘subscript𝐻𝑘subscript𝑆𝑘subscript𝑍𝑘𝑘1…𝑀Y_{k}=H_{k}S_{k}+Z_{k},\quad\quad k=1,\ldots,M</td>
<td></td>
<td>(5)</td>
</tr>
</tbody>
</table>
It is well known that the ML estimate of $`S_{k}`$ under the assumption of Gaussian noise is given by maximizing $`p\hspace{0pt}\left( Y_{k} \middle| {H_{k},S_{k}} \right)`$ which is equivalent to minimizing the euclidean distance
<table>
<tbody>
<tr>
<td></td>
<td>S^kML(Hk)=argminSk∈𝒮{|Yk−HkSk|2},subscriptsuperscript^𝑆ML𝑘subscript𝐻𝑘subscriptargminsubscript𝑆𝑘𝒮superscriptsubscript𝑌𝑘subscript𝐻𝑘subscript𝑆𝑘2{\widehat{S}}^{\mathrm{ML}}_{k}(H_{k})=\mathop{\mathrm{arg\,min}}_{S_{k}\in\,{\mathscr{S}}}\,\big{\{}\,|Y_{k}-H_{k}S_{k}|^{2}\,\big{\}},</td>
<td></td>
<td>(6)</td>
</tr>
</tbody>
</table>
where the set $`\mathcal{S}`$ contains all of the possible discrete values in the constellation that $`S_{k}`$ can take. Note that the above metric depends on the realization of the perfect channel $`H_{k}`$ and is optimum when perfect CSIR is available. However, in a real communication system, the receiver has only access to an imperfect estimated version of the channel. In the mismatched decoder, the estimated channel $`{\hat{H}}_{k}`$ is used instead of $`H_{k}`$ in the ML metric of (<a href="#S4.E6" style="font-size:90%;" title="In 4.1 Mismatched ML metric ‣ 4 Maximum-likelihood detection metric in the presence of channel estimation errors ‣ ACHIEVABLE OUTAGE RATES WITH IMPROVED DECODING OF BICM MULTIBAND OFDM UNDER CHANNEL ESTIMATION ERRORS">6</a>) as
<table>
<tbody>
<tr>
<td></td>
<td>S^kML(H^k)=argminSk∈𝒮{|Yk−HkSk|2}|Hk=H^k,{\widehat{S}}^{\mathrm{ML}}_{k}({\widehat{H}}_{k})=\mathop{\mathrm{arg\,min}}_{S_{k}\in\,{\mathscr{S}}}\,\big{\{}\,|Y_{k}-H_{k}S_{k}|^{2}\,\big{\}}_{{\big{|}_{H_{k}={\widehat{H}}_{k}}}},</td>
<td></td>
<td>(7)</td>
</tr>
</tbody>
</table>
which leads to a sub-optimal solution due to the mismatch introduced by the CEE. Next, we will see how the derived posterior channel distribution (<a href="#S3.E4" style="font-size:90%;" title="In 3 PILOT ASSISTED CHANNEL ESTIMATION ‣ ACHIEVABLE OUTAGE RATES WITH IMPROVED DECODING OF BICM MULTIBAND OFDM UNDER CHANNEL ESTIMATION ERRORS">4</a>), can be used in a modified ML metric that takes into account the available imperfect CSI.
### 4.2 Modified ML metric for imperfect CSIR
Since in practice the receiver knows solely the imperfect channel, it is relevant to express the likelihood criterion in terms of the estimated fade coefficient. For this purpose, we consider the problem of detecting the transmitted symbol $`S_{k}`$ by using the imperfect channel estimate $`{\hat{H}}_{k} = {\hat{H}}_{k,0}`$ in an ML criterion as
<table>
<tbody>
<tr>
<td></td>
<td>S^kℳ(H^k,0)=argminSk∈𝒮𝒟ℳ(Yk,Sk|H^k=H^k,0),subscriptsuperscript^𝑆ℳ𝑘subscript^𝐻𝑘0subscriptargminsubscript𝑆𝑘𝒮subscript𝒟ℳsubscript𝑌𝑘conditionalsubscript𝑆𝑘subscript^𝐻𝑘subscript^𝐻𝑘0{\widehat{S}}^{\mathcal{M}}_{k}({\widehat{H}}_{k,0})=\mathop{\mathrm{arg\,min}}_{S_{k}\in\,{\mathscr{S}}}\,\mathcal{D}_{{}_{\mathcal{M}}}(Y_{k},S_{k}|{\widehat{H}}_{k}={\widehat{H}}_{k,0}),</td>
<td></td>
<td>(8)</td>
</tr>
</tbody>
</table>
where $`{\mathcal{D}_{_{\mathcal{M}}}\hspace{0pt}\left( Y_{k},\left. S_{k} \middle| {\hat{H}}_{k} \right. \right)} \triangleq {- {{\ln p}\hspace{0pt}\left( Y_{k} \middle| {{\hat{H}}_{k},S_{k}} \right)}}`$. The pdf $`p\hspace{0pt}\left( Y_{k} \middle| {{\hat{H}}_{k},S_{k}} \right)`$ can be calculated as
<table>
<tbody>
<tr>
<td></td>
<td>p(Yk|H^k,Sk)=∫Hkp(Yk,Hk|H^k,Sk)dHk,𝑝conditionalsubscript𝑌𝑘subscript^𝐻𝑘subscript𝑆𝑘subscriptsubscript𝐻𝑘𝑝subscript𝑌𝑘conditionalsubscript𝐻𝑘subscript^𝐻𝑘subscript𝑆𝑘differential-dsubscript𝐻𝑘p(Y_{k}|{\widehat{H}}_{k},S_{k})=\int_{H_{k}}p(Y_{k},H_{k}|{\widehat{H}}_{k},S_{k})\;\;\mathrm{d}H_{k},</td>
<td></td>
<td>(9)</td>
</tr>
</tbody>
</table>
The joint likelihood function $`p\hspace{0pt}\left( Y_{k},\left. H_{k} \middle| {{\hat{H}}_{k},S_{k}} \right. \right)`$ is related to known probabilities as
<table>
<tbody>
<tr>
<td></td>
<td>p(Yk,Hk|H^k,Sk)=p(Yk|Hk,Sk)p(Hk|H^k),𝑝subscript𝑌𝑘conditionalsubscript𝐻𝑘subscript^𝐻𝑘subscript𝑆𝑘𝑝conditionalsubscript𝑌𝑘subscript𝐻𝑘subscript𝑆𝑘𝑝conditionalsubscript𝐻𝑘subscript^𝐻𝑘\displaystyle p(Y_{k},H_{k}|{\widehat{H}}_{k},S_{k})=p(Y_{k}|H_{k},S_{k})p(H_{k}|{\widehat{H}}_{k}),</td>
<td></td>
<td>(10)</td>
</tr>
</tbody>
</table>
where the last equation results from the independence between $`S_{k}`$ and $`\left( H_{k},{\hat{H}}_{k} \right)`$. It is clear from (<a href="#S4.E9" style="font-size:90%;" title="In 4.2 Modified ML metric for imperfect CSIR ‣ 4 Maximum-likelihood detection metric in the presence of channel estimation errors ‣ ACHIEVABLE OUTAGE RATES WITH IMPROVED DECODING OF BICM MULTIBAND OFDM UNDER CHANNEL ESTIMATION ERRORS">9</a>) and (<a href="#S4.E10" style="font-size:90%;" title="In 4.2 Modified ML metric for imperfect CSIR ‣ 4 Maximum-likelihood detection metric in the presence of channel estimation errors ‣ ACHIEVABLE OUTAGE RATES WITH IMPROVED DECODING OF BICM MULTIBAND OFDM UNDER CHANNEL ESTIMATION ERRORS">10</a>) that $`p\hspace{0pt}\left( Y_{k} \middle| {{\hat{H}}_{k},S_{k}} \right)`$ is obtained by the following conditional expectation
<table>
<tbody>
<tr>
<td></td>
<td>p(Yk|H^k,Sk)=𝔼Hk|H^k[p(Yk|Hk,Sk)|H^k],𝑝conditionalsubscript𝑌𝑘subscript^𝐻𝑘subscript𝑆𝑘subscript𝔼conditionalsubscript𝐻𝑘subscript^𝐻𝑘delimited-[]conditional𝑝conditionalsubscript𝑌𝑘subscript𝐻𝑘subscript𝑆𝑘subscript^𝐻𝑘p(Y_{k}|{\widehat{H}}_{k},S_{k})=\mathbb{E}_{H_{k}|{\widehat{H}}_{k}}\big{[}\,p(Y_{k}|H_{k},S_{k})\big{|}{\widehat{H}}_{k}\,\big{]},</td>
<td></td>
<td>(11)</td>
</tr>
</tbody>
</table>
evaluated over the posterior distribution of the exact channel given its estimate (equation (<a href="#S3.E4" style="font-size:90%;" title="In 3 PILOT ASSISTED CHANNEL ESTIMATION ‣ ACHIEVABLE OUTAGE RATES WITH IMPROVED DECODING OF BICM MULTIBAND OFDM UNDER CHANNEL ESTIMATION ERRORS">4</a>)). Since $`p\hspace{0pt}\left( H_{k} \middle| {\hat{H}}_{k} \right)`$ and $`p\hspace{0pt}\left( Y_{k} \middle| {H_{k},S_{k}} \right)`$ are Gaussian densities, their product remains Gaussian and it is easy to verify that $`{p\hspace{0pt}\left( Y_{k} \middle| {{\hat{H}}_{k},S_{k}} \right)} = {\mathcal{C}\hspace{0pt}\mathcal{N}\hspace{0pt}\left( \mu_{_{\mathcal{M}}},\sigma_{_{\mathcal{M}}}^{2} \right)}`$ where $`\mu_{_{\mathcal{M}}} = {\rho\hspace{0pt}{\hat{H}}_{k}\hspace{0pt}S_{k}}`$ and $`\sigma_{_{\mathcal{M}}}^{2} = {\sigma_{z}^{2} + {\left( {1 - \rho} \right)\hspace{0pt}\left| S_{k} \right|^{2}}}`$.
According to (<a href="#S4.E8" style="font-size:90%;" title="In 4.2 Modified ML metric for imperfect CSIR ‣ 4 Maximum-likelihood detection metric in the presence of channel estimation errors ‣ ACHIEVABLE OUTAGE RATES WITH IMPROVED DECODING OF BICM MULTIBAND OFDM UNDER CHANNEL ESTIMATION ERRORS">8</a>), the decision metric to be minimized for ML detection at the receiver is easily seen to be
<table>
<tbody>
<tr>
<td></td>
<td>𝒟ℳ(Sk,Yk|H^k)=ln(σz2+(1−ρ)|Sk|2)+|Yk−ρH^kSk|2σz2+(1−ρ)|Sk|2.subscript𝒟ℳsubscript𝑆𝑘conditionalsubscript𝑌𝑘subscript^𝐻𝑘subscriptsuperscript𝜎2𝑧1𝜌superscriptsubscript𝑆𝑘2superscriptsubscript𝑌𝑘𝜌subscript^𝐻𝑘subscript𝑆𝑘2subscriptsuperscript𝜎2𝑧1𝜌superscriptsubscript𝑆𝑘2\mathcal{D}_{{}_{\mathcal{M}}}(S_{k},Y_{k}|{\widehat{H}}_{k})=\ln\,({\sigma^{2}_{z}}+(1-\rho)|S_{k}|^{2})+\frac{|Y_{k}-\rho{\widehat{H}}_{k}S_{k}|^{2}}{{\sigma^{2}_{z}}+(1-\rho)|S_{k}|^{2}}.</td>
<td></td>
<td>(12)</td>
</tr>
</tbody>
</table>
Let us first consider the ideal situation when exact channel knowledge is available. In this case $`p\hspace{0pt}\left( H_{k} \middle| {\hat{H}}_{k} \right)`$ reduces to $`\delta\hspace{0pt}\left( {H_{k} - {\hat{H}}_{k}} \right)`$ and the averaging $`{\mathbb{E}}_{_{H_{k}}}\left( . \right)`$ of (<a href="#S4.E11" style="font-size:90%;" title="In 4.2 Modified ML metric for imperfect CSIR ‣ 4 Maximum-likelihood detection metric in the presence of channel estimation errors ‣ ACHIEVABLE OUTAGE RATES WITH IMPROVED DECODING OF BICM MULTIBAND OFDM UNDER CHANNEL ESTIMATION ERRORS">11</a>) becomes equivalent to replacing $`H_{k}`$ by $`{\hat{H}}_{k}`$ in $`p\hspace{0pt}\left( Y_{k} \middle| {H_{k},S_{k}} \right)`$. Similarly, we note that with near perfect channel (i.e. $`\sigma_{\mathcal{E}}^{2}\rightarrow 0`$ or $`N\rightarrow\infty`$), $`\rho\rightarrow 1`$ and we have $`{{\mathcal{D}_{_{\mathcal{M}}}\hspace{0pt}\left( Y_{k},\left. S_{k} \middle| {\hat{H}}_{k} \right. \right)}\rightarrow\left| {Y_{k} - {{\hat{H}}_{k}\hspace{0pt}S_{k}}} \right|^{2}},`$ which is the euclidean distance metric for perfect CSIR. However, in the presence of CEE, the modified metric evaluates the average of the pdf $`p\hspace{0pt}\left( Y_{k} \middle| {H_{k},S_{k}} \right)`$ (that would be used if the exact channel were known) over the true channel $`H_{k}`$ using the conditional expectation in (<a href="#S4.E11" style="font-size:90%;" title="In 4.2 Modified ML metric for imperfect CSIR ‣ 4 Maximum-likelihood detection metric in the presence of channel estimation errors ‣ ACHIEVABLE OUTAGE RATES WITH IMPROVED DECODING OF BICM MULTIBAND OFDM UNDER CHANNEL ESTIMATION ERRORS">11</a>).
While we have focused on OFDM, the metric (<a href="#S4.E12" style="font-size:90%;" title="In 4.2 Modified ML metric for imperfect CSIR ‣ 4 Maximum-likelihood detection metric in the presence of channel estimation errors ‣ ACHIEVABLE OUTAGE RATES WITH IMPROVED DECODING OF BICM MULTIBAND OFDM UNDER CHANNEL ESTIMATION ERRORS">12</a>) can also be applied for ML decoding/detection of single carrier systems with frequency selective fading. In the following, we use the above modified metric in the decoding process of BICM MB-OFDM.
## 5 Decoding of BICM-OFDM based on imperfect channel estimation
The problem of decoding BICM-OFDM with large size constellations has been addressed in \[[7](#bib.bib7)\] under the assumption of perfect CSIR. Here we consider the same problem in the case where the decoder has an imperfect estimate of the channel state. We denote by $`\mathbf{S}_{1}^{\tau} = \left\lbrack \mathbf{S}_{1}^{T},\ldots,\mathbf{S}_{\tau}^{T} \right\rbrack`$ the sequence of transmitted complex symbols and by $`\mathbf{Y}_{1}^{\tau} = \left\lbrack \mathbf{Y}_{1}^{T},\ldots,\mathbf{Y}_{\tau}^{T} \right\rbrack`$ the corresponding sequence of received symbols where $`\tau`$ is number of OFDM symbols in a frame. Based on the transmission model of figure <a href="#S2.F1" style="font-size:90%;" title="Figure 1 ‣ 2 TRANSMISSION MODEL ‣ ACHIEVABLE OUTAGE RATES WITH IMPROVED DECODING OF BICM MULTIBAND OFDM UNDER CHANNEL ESTIMATION ERRORS">1</a>, the sequence of transmitted bits $`\mathbf{b}`$ has to be estimated in the ML sense so as to maximize the likelihood function $`p\hspace{0pt}\left( \mathbf{Y}_{1}^{\tau} \middle| \mathbf{b} \right)`$. In practice, due to the one to one correspondence between the information bits and the interleaved codeword $`\mathbf{d}`$, one look for the ML estimate $`\hat{\mathbf{d}}`$ of the interleaved codeword as
<table>
<tbody>
<tr>
<td></td>
<td>𝐝^ℳ=argmax𝐝∈𝒟p(𝐘1τ|𝐇^1τ,𝐒1τ).superscript^𝐝ℳsubscriptargmax𝐝𝒟𝑝conditionalsuperscriptsubscript𝐘1𝜏superscriptsubscript^𝐇1𝜏superscriptsubscript𝐒1𝜏\widehat{\mathbf{d}}^{\mathcal{M}}=\mathop{\mathrm{arg\,max}}_{\mathbf{d}\in\,{\mathscr{D}}}p({\mathbf{Y}}_{1}^{\tau}\big{|}\widehat{{\mathbf{H}}}_{1}^{\tau},{\mathbf{S}}_{1}^{\tau}).</td>
<td></td>
<td>(13)</td>
</tr>
</tbody>
</table>
By assuming that the received symbols are independent and the channel noise is white, (<a href="#S5.E13" style="font-size:90%;" title="In 5 Decoding of BICM-OFDM based on imperfect channel estimation ‣ ACHIEVABLE OUTAGE RATES WITH IMPROVED DECODING OF BICM MULTIBAND OFDM UNDER CHANNEL ESTIMATION ERRORS">13</a>) is equivalent to (see \[[7](#bib.bib7)\])
<table>
<tbody>
<tr>
<td></td>
<td>𝐝^ℳ≈argmax𝐝∈𝒟∏k∏l=1Bp(dkl|H^k,Yk),superscript^𝐝ℳsubscriptargmax𝐝𝒟subscriptproduct𝑘superscriptsubscriptproduct𝑙1𝐵𝑝conditionalsuperscriptsubscript𝑑𝑘𝑙subscript^𝐻𝑘subscript𝑌𝑘\widehat{\mathbf{d}}^{\mathcal{M}}\approx\mathop{\mathrm{arg\,max}}_{\mathbf{d}\in\,{\mathscr{D}}}\prod_{k}\prod_{l=1}^{B}p(d_{k}^{l}|{\widehat{H}}_{k},Y_{k}),</td>
<td></td>
<td>(14)</td>
</tr>
</tbody>
</table>
where the bit metrics $`p\hspace{0pt}\left( d_{k}^{l} \middle| {{\hat{H}}_{k},Y_{k}} \right)`$ are given by
<table>
<tbody>
<tr>
<td></td>
<td>p(dkl|H^k,Yk)𝑝conditionalsuperscriptsubscript𝑑𝑘𝑙subscript^𝐻𝑘subscript𝑌𝑘\displaystyle p(d_{k}^{l}|{\widehat{H}}_{k},Y_{k})</td>
<td>∝∑dkn∈{0,1}∀l≠np(Yk|H^k,Sk(dk1,…,dkB)),proportional-toabsentsubscriptsuperscriptsubscript𝑑𝑘𝑛01for-all𝑙𝑛𝑝conditionalsubscript𝑌𝑘subscript^𝐻𝑘subscript𝑆𝑘superscriptsubscript𝑑𝑘1…superscriptsubscript𝑑𝑘𝐵\displaystyle\propto\sum_{\begin{subarray}{l}d_{k}^{n}\in\{0,1\}\\ \forall l\neq n\end{subarray}}p(Y_{k}|{\widehat{H}}_{k},S_{k}(d_{k}^{1},\ldots,d_{k}^{B})),</td>
<td></td>
<td></td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td></td>
<td>∝∑dkn∈{0,1}∀l≠n1πσℳ2exp(−|Yk−μℳ|2σℳ2),proportional-toabsentsubscriptsuperscriptsubscript𝑑𝑘𝑛01for-all𝑙𝑛1𝜋subscriptsuperscript𝜎2ℳsuperscriptsubscript𝑌𝑘subscript𝜇ℳ2subscriptsuperscript𝜎2ℳ\displaystyle\propto\sum_{\begin{subarray}{l}d_{k}^{n}\in\{0,1\}\\ \forall l\neq n\end{subarray}}\frac{1}{\pi\sigma^{2}_{{}_{\mathcal{M}}}}\,\exp\big{(}-\frac{|Y_{k}-\mu_{{}_{\mathcal{M}}}|^{2}}{\sigma^{2}_{{}_{\mathcal{M}}}}\big{)},</td>
<td></td>
<td>(15)</td>
</tr>
</tbody>
</table>
and $`p\hspace{0pt}\left( Y_{k} \middle| {{\hat{H}}_{k},S_{k}} \right)`$ is specified from (<a href="#S4.E11" style="font-size:90%;" title="In 4.2 Modified ML metric for imperfect CSIR ‣ 4 Maximum-likelihood detection metric in the presence of channel estimation errors ‣ ACHIEVABLE OUTAGE RATES WITH IMPROVED DECODING OF BICM MULTIBAND OFDM UNDER CHANNEL ESTIMATION ERRORS">11</a>). We emphasize that the above decoding rule is conditioned on the imperfect estimated channel sequence $`{\hat{\mathbf{H}}}_{1}^{\tau}`$ available at the receiver. Information bits are estimated in the trellis decoder part from the above bit metrics by applying either non-iterative or turbo procedures (see figure <a href="#S5.F2" style="font-size:90%;" title="Figure 2 ‣ 5 Decoding of BICM-OFDM based on imperfect channel estimation ‣ ACHIEVABLE OUTAGE RATES WITH IMPROVED DECODING OF BICM MULTIBAND OFDM UNDER CHANNEL ESTIMATION ERRORS">2</a>).
[Refer to caption](/html/0708.1411/assets/x1.png)
Fig. 2:
BICM iterative demodulation with CEE.
## 6 Achievable Outage Rates Associated to the Proposed Metric
We now derive the achievable rates $`C_{_{\mathcal{M}}}`$ of a receiver using the modified metric $`\mathcal{D}_{_{\mathcal{M}}}`$ of (<a href="#S4.E12" style="font-size:90%;" title="In 4.2 Modified ML metric for imperfect CSIR ‣ 4 Maximum-likelihood detection metric in the presence of channel estimation errors ‣ ACHIEVABLE OUTAGE RATES WITH IMPROVED DECODING OF BICM MULTIBAND OFDM UNDER CHANNEL ESTIMATION ERRORS">12</a>). This is done by using the following theorem proposed in \[[5](#bib.bib5)\] for a discrete-time memoryless channel $`{W\hspace{0pt}\left( \mathbf{y} \middle| {\mathbf{s},\mathbf{H}} \right)} = {\prod_{k = 1}^{M}{\mathcal{C}\hspace{0pt}\mathcal{N}\hspace{0pt}\left( {H_{k}\hspace{0pt}s_{k}},\sigma_{z}^{2} \right)}}`$, where $`\mathbf{y} = \left( y_{1},\ldots,y_{M} \right)`$, $`\mathbf{s} = \left( s_{1},\ldots,s_{M} \right)`$ with decoding metrics $`\mathcal{D}_{_{\mathcal{M}}}\hspace{0pt}\left( y_{k},\left. s_{k} \middle| {\hat{H}}_{k} \right. \right)`$:
######
Theorem 1
.
<table>
<tbody>
<tr>
<td></td>
<td>Cℳ=supp(s)inff(y|s)∈ℱD(f∘PS∥PSPY),subscript𝐶ℳsubscriptsupremum𝑝𝑠subscriptinfimum𝑓conditional𝑦𝑠ℱ𝐷conditional𝑓subscript𝑃𝑆subscript𝑃𝑆subscript𝑃𝑌C_{{}_{\mathcal{M}}}=\sup_{p(s)}\;\inf_{f(y|s)\in\mathcal{F}}D(f\!\circ\!P_{S}\|P_{S}P_{Y}),</td>
<td></td>
<td>(16)</td>
</tr>
</tbody>
</table>
where the relative entropy functional $`{D\hspace{0pt}{({{f \circ P_{S}} \parallel {P_{S}\hspace{0pt}P_{Y}}})}} =`$
<table>
<tbody>
<tr>
<td></td>
<td>∫∫PS(𝐬)f(𝐲|𝐬)logf(𝐲|𝐬)∫PS′(s′)f(𝐲|𝐬′)d𝐬′d𝐲d𝐬,subscript𝑃𝑆𝐬𝑓conditional𝐲𝐬𝑓conditional𝐲𝐬subscript𝑃superscript𝑆′superscript𝑠′𝑓conditional𝐲superscript𝐬′dsuperscript𝐬′d𝐲d𝐬\int\int P_{S}(\mathbf{s})f(\mathbf{y}|\mathbf{s})\log\frac{f(\mathbf{y}|\mathbf{s})}{\int P_{S^{\prime}}(s^{\prime})f(\mathbf{y}|\mathbf{s}^{\prime})\textrm{d}\mathbf{s}^{\prime}}\textrm{d}\mathbf{y}\textrm{d}\mathbf{s},</td>
<td></td>
<td>(17)</td>
</tr>
</tbody>
</table>
and $`\mathcal{F}`$ denotes the set of all possible channels $`f\hspace{0pt}{(\left. \mathbf{y} \middle| \mathbf{s} \right.)}`$ that satisfy:
<table>
<tbody>
<tr>
<td></td>
<td></td>
<td>c1:𝔼𝐬[f(𝐲|𝐬)]=𝔼𝐬[W(𝐲|𝐬,𝐇)],a.s.:subscriptc1subscript𝔼𝐬delimited-[]𝑓conditional𝐲𝐬subscript𝔼𝐬delimited-[]𝑊conditional𝐲𝐬𝐇a.s.\displaystyle\mathrm{c_{1}}:\;\mathbb{E}_{\mathbf{s}}\big{[}f(\mathbf{y}|\mathbf{s})\big{]}=\mathbb{E}_{\mathbf{s}}\big{[}W(\mathbf{y}|\mathbf{s},\mathbf{H})\big{]},\;\;\;\;\;\;\textrm{a.s.}\vspace{-2mm}</td>
<td></td>
<td>(18)</td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td></td>
<td>c2:𝔼𝐬[𝔼f(𝐲|𝐬)[𝒟(sk,yk|H^k)]]≤𝔼𝐬[𝔼W(𝐲|𝐬)[𝒟(sk,yk|H^k)]],:subscriptc2subscript𝔼𝐬delimited-[]subscript𝔼𝑓conditional𝐲𝐬delimited-[]𝒟subscript𝑠𝑘conditionalsubscript𝑦𝑘subscript^𝐻𝑘subscript𝔼𝐬delimited-[]subscript𝔼𝑊conditional𝐲𝐬delimited-[]𝒟subscript𝑠𝑘conditionalsubscript𝑦𝑘subscript^𝐻𝑘\displaystyle\mathrm{c_{2}}:\;\mathbb{E}_{\mathbf{s}}\,\bigg{[}\mathbb{E}_{{}_{f(\mathbf{y}|\mathbf{s})}}\big{[}\mathcal{D}(s_{k},y_{k}|\widehat{H}_{k})\big{]}\bigg{]}\leq\mathbb{E}_{\mathbf{s}}\,\bigg{[}\mathbb{E}_{{}_{W(\mathbf{y}|\mathbf{s})}}\big{[}\mathcal{D}(s_{k},y_{k}|\widehat{H}_{k})\big{]}\bigg{]},\vspace{-2mm}</td>
<td></td>
<td>(19)</td>
</tr>
</tbody>
</table>
for every $`k = {1,\ldots,M}`$.
We assume Gaussian i.i.d. inputs $`s_{k} \sim {\mathcal{C}\hspace{0pt}\mathcal{N}\hspace{0pt}(0,P)}`$. Furthermore, since we aim to minimize the mutual information in (<a href="#S6.E16" style="font-size:90%;" title="In Theorem 1. ‣ 6 Achievable Outage Rates Associated to the Proposed Metric ‣ ACHIEVABLE OUTAGE RATES WITH IMPROVED DECODING OF BICM MULTIBAND OFDM UNDER CHANNEL ESTIMATION ERRORS">16</a>), we choose the Gaussian distribution $`{f\hspace{0pt}\left( \mathbf{y} \middle| {\mathbf{s},\underset{¯}{\mu}} \right)} = {\prod_{k = 1}^{M}{\mathcal{C}\hspace{0pt}\mathcal{N}\hspace{0pt}\left( {\mu_{k}\hspace{0pt}s_{k}},\sigma_{k}^{2} \right)}}`$ with $`\underset{¯}{\mu} = \left( \mu_{1},\ldots,\mu_{M} \right)`$. It is easy to see from constraint $`c_{1}`$ that the variances of the optimal channel distribution $`f\hspace{0pt}\left( \mathbf{y} \middle| {\mathbf{s},\underset{¯}{\mu}} \right)`$ are given by
<table>
<tbody>
<tr>
<td></td>
<td>σk2=σz2+P(|Hk|2−|μk|2),k=1,…,M.formulae-sequencesubscriptsuperscript𝜎2𝑘subscriptsuperscript𝜎2𝑧𝑃superscriptsubscript𝐻𝑘2superscriptsubscript𝜇𝑘2𝑘1…𝑀\sigma^{2}_{k}=\sigma^{2}_{z}+P(|H_{k}|^{2}-|\mu_{k}|^{2}),\,\,\,\,k=1,\dots,M.</td>
<td></td>
<td>(20)</td>
</tr>
</tbody>
</table>
Then, for any given channel $`\mathbf{H}`$ and its estimate $`\hat{\mathbf{H}}`$, the resulting achievable rates for the proposed metric are obtained as
<table>
<tbody>
<tr>
<td></td>
<td>Cℳ(𝐇,𝐇^)=infμ¯:μk∈𝒱(Hk,H^k)∑k=1Mlog2(1+P|μk|2σk2),subscript𝐶ℳ𝐇^𝐇subscriptinfimum:¯𝜇subscript𝜇𝑘𝒱subscript𝐻𝑘subscript^𝐻𝑘superscriptsubscript𝑘1𝑀subscript21𝑃superscriptsubscript𝜇𝑘2subscriptsuperscript𝜎2𝑘C_{{}_{\mathcal{M}}}(\mathbf{H},\widehat{\mathbf{H}})=\!\!\!\!\!\!\!\!\!\!\inf_{\underline{\mu}:\,\,\mu_{k}\in\mathcal{V}(H_{k},\widehat{H}_{k})}\sum\limits_{k=1}^{M}\log_{2}\left(1+\frac{P\;|\mu_{k}|^{2}}{\sigma^{2}_{k}}\right),</td>
<td></td>
<td>(21)</td>
</tr>
</tbody>
</table>
where $`\mathcal{V}\hspace{0pt}\left( H_{k},{\hat{H}}_{k} \right)`$ are the constraint sets $`c_{2}`$ corresponding to the metric $`\mathcal{D}_{\mathcal{M}}`$, derived from the inequality constraint (<a href="#S6.E19" style="font-size:90%;" title="In Theorem 1. ‣ 6 Achievable Outage Rates Associated to the Proposed Metric ‣ ACHIEVABLE OUTAGE RATES WITH IMPROVED DECODING OF BICM MULTIBAND OFDM UNDER CHANNEL ESTIMATION ERRORS">19</a>) as
<table>
<tbody>
<tr>
<td></td>
<td>𝒱(Hk,H^k)={μk∈ℂ:|Hk−akH^k|2≤|μk−aH^k|2},𝒱subscript𝐻𝑘subscript^𝐻𝑘conditional-setsubscript𝜇𝑘ℂsuperscriptsubscript𝐻𝑘subscript𝑎𝑘subscript^𝐻𝑘2superscriptsubscript𝜇𝑘𝑎subscript^𝐻𝑘2\mathcal{V}(H_{k},\widehat{H}_{k})=\big{\{}\mu_{k}\in\mathbb{C}:|H_{k}-a_{k}\widehat{H}_{k}|^{2}\leq|\mu_{k}-a\widehat{H}_{k}|^{2}\big{\}},</td>
<td></td>
<td>(22)</td>
</tr>
</tbody>
</table>
where $`a_{k} = \left. {\rho_{k}\hspace{0pt}\left( {{\lambda_{k}\hspace{0pt}\sigma_{z}^{2}} - {P\hspace{0pt}\left( {1 - \rho_{k}} \right)}} \right)}/\left( {{\lambda_{k}\hspace{0pt}\sigma_{z}^{2}} - {P\hspace{0pt}\left( {1 - \rho_{k}} \right)\hspace{0pt}\left( {1 - \lambda_{k}} \right)}} \right) \right.`$;
$`\lambda_{k} = {{\exp\left( \frac{\sigma_{z}^{2}}{P\hspace{0pt}\left( {1 - \rho_{k}} \right)} \right)}\hspace{0pt}E_{1}\hspace{0pt}\left( \frac{\sigma_{z}^{2}}{P\hspace{0pt}\left( {1 - \rho_{k}} \right)} \right)}`$ and $`{E_{1}\hspace{0pt}(x)} \triangleq {\int_{x}^{+ \infty}{\frac{\exp\left( {- u} \right)}{u}\hspace{0pt}{du}}}`$ is defined as the exponential integral.
We note that (<a href="#S6.E21" style="font-size:90%;" title="In 6 Achievable Outage Rates Associated to the Proposed Metric ‣ ACHIEVABLE OUTAGE RATES WITH IMPROVED DECODING OF BICM MULTIBAND OFDM UNDER CHANNEL ESTIMATION ERRORS">21</a>) is an increasing function of $`\left\| \underset{¯}{\mu} \right\|^{2}`$. Consequently, the infimum is obtained for $`\mu_{k} = \mu_{{opt},k}`$ resulting from minimizing $`\left\| \underset{¯}{\mu} \right\|^{2}`$ constrained by the equality in the constraint set (<a href="#S6.E22" style="font-size:90%;" title="In 6 Achievable Outage Rates Associated to the Proposed Metric ‣ ACHIEVABLE OUTAGE RATES WITH IMPROVED DECODING OF BICM MULTIBAND OFDM UNDER CHANNEL ESTIMATION ERRORS">22</a>). Finally, by using Lagrange multipliers we obtain
<table>
<tbody>
<tr>
<td></td>
<td>μopt,k=ηℳ,kH^k,k=1,…,Mformulae-sequencesubscript𝜇opt𝑘subscript𝜂ℳ𝑘subscript^𝐻𝑘𝑘1…𝑀\mu_{\mathrm{opt},k}=\eta_{{}_{\mathcal{M},k}}\widehat{H}_{k},\,\;\;\,k=1,\dots,M</td>
<td></td>
<td>(23)</td>
</tr>
</tbody>
</table>
where
<table>
<tbody>
<tr>
<td></td>
<td>ηℳ,k={ak−|Hk−akH^k||H^k|,ifak≥0ak+|Hk−akH^k||H^k|,ifak<0.subscript𝜂ℳ𝑘casessubscript𝑎𝑘subscript𝐻𝑘subscript𝑎𝑘subscript^𝐻𝑘subscript^𝐻𝑘ifsubscript𝑎𝑘0subscript𝑎𝑘subscript𝐻𝑘subscript𝑎𝑘subscript^𝐻𝑘subscript^𝐻𝑘ifsubscript𝑎𝑘0\eta_{{}_{\mathcal{M},k}}=\left\{\begin{array}[]{ll}a_{k}-\displaystyle{\frac{|H_{k}-a_{k}\widehat{H}_{k}|}{|\widehat{H}_{k}|}},&\mathrm{if}\,\,a_{k}\geq 0\\ a_{k}+\displaystyle{\frac{|H_{k}-a_{k}\widehat{H}_{k}|}{|\widehat{H}_{k}|}},&\mathrm{if}\,\,a_{k}<0.\end{array}\right.</td>
<td></td>
<td>(24)</td>
</tr>
</tbody>
</table>
The achievable rates associated to the metric $`\mathcal{D}_{_{\mathcal{M}}}`$ are given by
<table>
<tbody>
<tr>
<td></td>
<td>Cℳ(𝐇,𝐇^)=∑k=1Mlog2(1+Pηℳ,k2|H^k|2σz2+P(|Hk|2−ηℳ,k2|H^k|2)).subscript𝐶ℳ𝐇^𝐇superscriptsubscript𝑘1𝑀subscript21𝑃superscriptsubscript𝜂ℳ𝑘2superscriptsubscript^𝐻𝑘2subscriptsuperscript𝜎2𝑧𝑃superscriptsubscript𝐻𝑘2superscriptsubscript𝜂ℳ𝑘2superscriptsubscript^𝐻𝑘2C_{{}_{\mathcal{M}}}(\mathbf{H},\widehat{\mathbf{H}})=\sum\limits_{k=1}^{M}\log_{2}\left(1+\frac{P\;\eta_{{}_{\mathcal{M},k}}^{2}|\widehat{H}_{k}|^{2}}{\sigma^{2}_{z}+P(|H_{k}|^{2}-\eta_{{}_{\mathcal{M},k}}^{2}|\widehat{H}_{k}|^{2})}\right).</td>
<td></td>
<td>(25)</td>
</tr>
</tbody>
</table>
Then, the associated outage probability for an outage rate $`R \geq 0`$ is defined as
<table>
<tbody>
<tr>
<td></td>
<td>Pℳout(R,𝐇^)=Pr𝐇|𝐇^(𝐇∈Λℳ(R,𝐇^)|𝐇^),superscriptsubscript𝑃ℳout𝑅^𝐇subscriptPrconditional𝐇^𝐇𝐇conditionalsubscriptΛℳ𝑅^𝐇^𝐇P_{{}_{\mathcal{M}}}^{\mathrm{out}}(R,\widehat{\mathbf{H}})=\mathrm{Pr}_{\mathbf{H}|\widehat{\mathbf{H}}}\big{(}\mathbf{H}\in\Lambda_{{}_{\mathcal{M}}}(R,\widehat{\mathbf{H}})|\widehat{\mathbf{H}}\big{)},</td>
<td></td>
</tr>
</tbody>
</table>
with $`{\Lambda_{_{\mathcal{M}}}\hspace{0pt}\left( R,\hat{\mathbf{H}} \right)} = \left\{ \mathbf{H}:{{C_{_{\mathcal{M}}}\hspace{0pt}\left( \mathbf{H},\hat{\mathbf{H}} \right)} < R} \right\}`$ and the maximal outage rate for an outage probability $`\gamma`$ is
<table>
<tbody>
<tr>
<td></td>
<td>Cℳout(γ,𝐇^)=supR{R≥0:Pℳout(R,𝐇^)≤γ}.superscriptsubscript𝐶ℳout𝛾^𝐇subscriptsupremum𝑅conditional-set𝑅0superscriptsubscript𝑃ℳout𝑅^𝐇𝛾C_{{}_{\mathcal{M}}}^{\mathrm{out}}(\gamma,\widehat{\mathbf{H}})=\sup_{R}\big{\{}R\geq 0:P_{{}_{\mathcal{M}}}^{\mathrm{out}}(R,\widehat{\mathbf{H}})\leq\gamma\big{\}}.</td>
<td></td>
<td>(26)</td>
</tr>
</tbody>
</table>
Since this outage rate still depends on the random channel estimates $`\hat{\mathbf{H}}`$, we will consider the average rates over all channel estimates
<table>
<tbody>
<tr>
<td></td>
<td>C¯ℳout(γ)=𝔼𝐇^[Cℳout(γ,𝐇^)].superscriptsubscript¯𝐶ℳout𝛾subscript𝔼^𝐇delimited-[]superscriptsubscript𝐶ℳout𝛾^𝐇\overline{C}_{{}_{\mathcal{M}}}^{\;\mathrm{out}}(\gamma)=\mathbb{E}_{\widehat{\mathbf{H}}}\big{[}C_{{}_{\mathcal{M}}}^{\mathrm{out}}(\gamma,\widehat{\mathbf{H}})\big{]}.</td>
<td></td>
<td>(27)</td>
</tr>
</tbody>
</table>
The achievable rates (<a href="#S6.E27" style="font-size:90%;" title="In 6 Achievable Outage Rates Associated to the Proposed Metric ‣ ACHIEVABLE OUTAGE RATES WITH IMPROVED DECODING OF BICM MULTIBAND OFDM UNDER CHANNEL ESTIMATION ERRORS">27</a>) are upper bounded by the estimation-induced outage capacity, the maximal outage rates (i.e. maximizing over all possible transmitter-receiver pairs), derived in \[[3](#bib.bib3)\] by using a theoretical decoder. In our case, this capacity is given by
<table>
<tbody>
<tr>
<td></td>
<td>C¯𝒢out(γ)=𝔼𝐇^[C𝒢out(γ,𝐇^)],superscriptsubscript¯𝐶𝒢out𝛾subscript𝔼^𝐇delimited-[]superscriptsubscript𝐶𝒢out𝛾^𝐇\overline{C}_{{}_{\mathcal{G}}}^{\;\mathrm{out}}(\gamma)=\mathbb{E}_{\widehat{\mathbf{H}}}\big{[}C_{{}_{\mathcal{G}}}^{\mathrm{out}}(\gamma,\widehat{\mathbf{H}})\big{]},</td>
<td></td>
<td>(28)</td>
</tr>
</tbody>
</table>
where the outage rates $`C_{_{\mathcal{G}}}^{out}\hspace{0pt}\left( \gamma,\hat{\mathbf{H}} \right)`$ are computed using
<table>
<tbody>
<tr>
<td></td>
<td>C𝒢(𝐇,𝐇^)=∑k=1Mlog2(1+P|Hk|2σz2).subscript𝐶𝒢𝐇^𝐇superscriptsubscript𝑘1𝑀subscript21𝑃superscriptsubscript𝐻𝑘2subscriptsuperscript𝜎2𝑧C_{{}_{\mathcal{G}}}(\mathbf{H},\widehat{\mathbf{H}})=\sum\limits_{k=1}^{M}\log_{2}\left(1+\frac{P\;|H_{k}|^{2}}{\sigma^{2}_{z}}\right).</td>
<td></td>
<td>(29)</td>
</tr>
</tbody>
</table>
[Refer to caption](/html/0708.1411/assets/x2.png)
Fig. 3:
BER performance of the proposed decoder over the LOS UWB channel CM1 for various training sequence lengths.
[Refer to caption](/html/0708.1411/assets/x3.png)
Fig. 4:
BER performance of the proposed decoder over Rayleigh fading channel for various training sequence lengths.
## 7 Simulation results
Simulations have been carried out in the context of IEEE802.15 wireless PAN \[[6](#bib.bib6)\]: $`M = 100`$ data subcarriers along with 32 CP samples compose one OFDM symbol. Information bits are encoded by a rate $`R = \left. 1/2 \right.`$ convolutional encoder with constraint length 3 defined in octal form by (5,7). The interleaver is a pseudo-random one operating over the entire frame and the mapping is 16-QAM. For each frame that contains 100 OFDM symbols, a different realization of the realistic UWB channel model specified in \[[8](#bib.bib8)\] has been drawn and supposed constant during the whole frame. Besides, the same average energy is allocated to training symbols and data, i.e., $`E_{_{T}} = {{\mathbb{E}}\hspace{0pt}\left( \left| {\overset{\sim}{S}}_{k} \right|^{2} \right)} = E_{s}`$.
Figure <a href="#S6.F3" style="font-size:90%;" title="Figure 3 ‣ 6 Achievable Outage Rates Associated to the Proposed Metric ‣ ACHIEVABLE OUTAGE RATES WITH IMPROVED DECODING OF BICM MULTIBAND OFDM UNDER CHANNEL ESTIMATION ERRORS">3</a> depicts the bit error rate (BER) performance gain that is obtained by decoding BICM MB-OFDM with the modified ML decoder in the context of LOS CM1 channel, where the channel estimation is performed by sending different training sequences with lengths $`N = \left\{ 1,2,8 \right\}`$ per frame. It can be noticed that the proposed decoder outperforms the mismatched decoder especially when few numbers of pilot symbols are dedicated for channel estimation. Note that the modified decoder with $`N = 1`$ pilot performs very close to the mismatched decoder with $`N = 2`$ pilots. For comparison, results obtained with theoretical Rayleigh fading channel are illustrated in figure <a href="#S6.F4" style="font-size:90%;" title="Figure 4 ‣ 6 Achievable Outage Rates Associated to the Proposed Metric ‣ ACHIEVABLE OUTAGE RATES WITH IMPROVED DECODING OF BICM MULTIBAND OFDM UNDER CHANNEL ESTIMATION ERRORS">4</a>. It can be observed that for $`N = 2`$, the SNR to obtain a BER of $`10^{- 3}`$ is reduced by about $`1.5`$ dB if the modified ML decoding is used instead of the mismatched approach.
Figure <a href="#S8.F5" style="font-size:90%;" title="Figure 5 ‣ ACHIEVABLE OUTAGE RATES WITH IMPROVED DECODING OF BICM MULTIBAND OFDM UNDER CHANNEL ESTIMATION ERRORS">5</a> shows the BER performance versus the training sequence length $`N`$ at a fixed $`\left. {E\hspace{0pt}b}/N_{0} \right.`$ of $`12`$ dB for the CM1 channel. This allows to evaluate the number of training sequence necessary to achieve a certain BER. We observe that at $`\left. {E\hspace{0pt}b}/N_{0} \right. = 12`$ dB, the modified ML decoder requires $`N = 9`$ pilot symbols per frame to achieve a BER of $`10^{- 4}`$ while the mismatched decoder attains this BER for $`N = 12`$ pilot symbols. Besides, this has been outlined in Section IV, for large training sequence lengths ($`N \geq 12`$), both decoders have close performance.
Figure <a href="#S8.F6" style="font-size:90%;" title="Figure 6 ‣ ACHIEVABLE OUTAGE RATES WITH IMPROVED DECODING OF BICM MULTIBAND OFDM UNDER CHANNEL ESTIMATION ERRORS">6</a> shows the expected achievable rates (<a href="#S6.E27" style="font-size:90%;" title="In 6 Achievable Outage Rates Associated to the Proposed Metric ‣ ACHIEVABLE OUTAGE RATES WITH IMPROVED DECODING OF BICM MULTIBAND OFDM UNDER CHANNEL ESTIMATION ERRORS">27</a>) corresponding to one OFDM symbol transmission with $`M = 16`$ data subcarriers (in bits per channel use) versus the SNR, for an outage probability $`\gamma = 0.01`$. For comparison, we show the expected outage rates of mismatched ML decoding and that obtained in the case of the perfect theoretical encoder/decoder (<a href="#S6.E28" style="font-size:90%;" title="In 6 Achievable Outage Rates Associated to the Proposed Metric ‣ ACHIEVABLE OUTAGE RATES WITH IMPROVED DECODING OF BICM MULTIBAND OFDM UNDER CHANNEL ESTIMATION ERRORS">28</a>). It can be observed that at a mean outage rate of 4 bits, the modified metric reduces the SNR by about 1 dB. Indeed, we observe that the achievable rates using the proposed decoding metric are close (only $`2`$ dB of SNR) to those obtained by using a theoretical decoder.
## 8 Conclusion
This paper studied the problem of ML reception when the receiver has only access to a noisy estimate of the channel in the case of pilot assisted channel estimation. By using the statistics of the estimation error, we proposed a modified ML criterion that is expressed in terms of the estimated channel coefficient. This modified metric let us to formulate appropriate branch metrics for decoding BICM MB-OFDM with imperfect channel knowledge. Achievable outage rates associated to this modified metric have also been derived. Comparison in terms of BER and achievable outage rates was made with a system whose receiver uses mismatched ML decoding and a perfect theoretical decoder. Simulation results conducted over realistic UWB channels, indicate that mismatched decoding is quite sub-optimal for short training sequence and confirmed the adequacy of the proposed decoding rule in the presence of channel estimation errors. This was obtained without introducing any additional complexity.
## References
- \[1\]
FCC,
“First report and order, revision of part 15 of the commission’s rules regarding ultra-wideband transmission systems,”
Tech. Rep., Feb. 2004.
- \[2\]
A. Batra, J. Balakrishnan, G. R. Aiello, J. R. Foerster, and A. Dabak,
“Design of multiband OFDM system for realistic UWB channel environments,”
IEEE Transaction on Microwave Theory and Techniques, vol. 52, pp. 2123–2138, september 2004.
- \[3\]
Pablo Piantanida, Gerald Matz, and Pierre Duhamel,
“Estimation-induced outage capacity of Ricean channels,”
in Proc. SPAWC, July 2006.
- \[4\]
Giorgio Taricco and Ezio Biglieri,
“Space-time decoding with imperfect channel estimation,”
IEEE Transaction on Wireless Communications, vol. 4, no. 4, pp. 1874, 1888 2005.
- \[5\]
Neri Merhav, Gideon Kaplan, Amos Lapidoth, and Sholmo Shamai,
“On information rates for mismatched decoders,”
IEEE Transactions on Information Theory, vol. 40, pp. 1953–1967, November 1994.
- \[6\]
Anuj Batra, Jaiganesh Balakrishnan, and Anad Dabak,
“Multiband OFDM physical layer proposal for IEEE 802.15 task group 3a,”
Tech. Rep., IEEE, july 2003.
- \[7\]
P. Magniez, B. Muquet, P. Duhamel, V. Buzenac, and M. Decourville,
“Optimal decoding of bit-interleaved modulations: Theoretical aspects and practical algorithms,”
in Proc Int. Symp. on Turbo Codes and related topics, Sep. 2004, pp. 169–172.
- \[8\]
J. Foerster,
“Channel modeling sub-committee report final,”
Tech. Rep., IEEE802.15-02/490, 2003.
[Refer to caption](/html/0708.1411/assets/x4.png)
Fig. 5:
BER versus the number of pilot symbols at $`{E_{b}/N_{0}} = 12`$ dB over the LOS channel CM1.
[Refer to caption](/html/0708.1411/assets/x5.png)
Fig. 6:
Expected outage rates of OFDM transmission with $`M = 100`$ subcarriers versus SNR $`({N = 1})`$. |
0708.2309 | cs.NI | cyber | # On Compact Routing for the Internet
Dmitri Krioukov
kc claffy
CAIDA <dima@caida.org> CAIDA <kc@caida.org>
Kevin Fall
Arthur Brady
Intel Research Berkeley <kfall@intel.com> Tufts University <abrady@cs.tufts.edu>
###### Abstract
The Internet’s routing system is facing stresses due to its poor fundamental scaling properties. Compact routing is a research field that studies fundamental limits of routing scalability and designs algorithms that try to meet these limits. In particular, compact routing research shows that shortest-path routing, forming a core of traditional routing algorithms, cannot guarantee routing table (RT) sizes that on all network topologies grow slower than linearly as functions of the network size. However, there are plenty of compact routing schemes that relax the shortest-path requirement and allow for improved, sublinear RT size scaling that is mathematically provable for all static network topologies. In particular, there exist compact routing schemes designed for grids, trees, and Internet-like topologies that offer RT sizes that scale logarithmically with the network size.
In this paper, we demonstrate that in view of recent results in compact routing research, such logarithmic scaling on Internet-like topologies is fundamentally impossible in the presence of topology dynamics or topology-independent (flat) addressing. We use analytic arguments to show that the number of routing control messages per topology change cannot scale better than linearly on Internet-like topologies. We also employ simulations to confirm that logarithmic RT size scaling gets broken by topology-independent addressing, a cornerstone of popular locator-identifier split proposals aiming at improving routing scaling in the presence of network topology dynamics or host mobility. These pessimistic findings lead us to the conclusion that a fundamental re-examination of assumptions behind routing models and abstractions is needed in order to find a routing architecture that would be able to scale “indefinitely.”
###### category:
C.2.2 Network Protocols Routing protocols
###### category:
G.2.2 Graph Theory Graph algorithms, Network problems
###### category:
C.2.1 Network Architecture and Design Network topology
###### keywords:
Compact routing, Internet routing, routing scalability
<sup>†</sup>
<sup>†</sup>terms: Algorithms, Design, Theory
## 1 Introduction
Despite prevailing concerns among network operators and developers that the current Internet interdomain routing system will not scale to meet the needs of the 21st century global Internet, networking research has not yet produced a new routing architecture with satisfactory flexibility and mathematically provable scalability characteristics. We believe that we will soon be faced with a critical architectural inflection point posing a significant challenge to scaling the size of the future Internet. The last critical point was reached when the Internet’s routing system adopted strong address aggregation using CIDR to handle address scaling in the mid 90’s. While CIDR was an extremely effective tactic, most experts agree that the growth behavior of the routing table in the last decade confirms that the type of address management required by CIDR will not suffice to meet future Internet routing needs, and that a fundamental top-to-bottom reexamination of the routing architecture may be required \[[1](#bib.bib1)\].
In this paper we set to work on such a reexamination.
<sup>1</sup>
<sup>1</sup>1We clarify up front that scalability is only one of several problems of the current Internet routing architecture. Other problems include security, isolation, configuration control, etc. See \[[2](#bib.bib2)\] for a long list of future routing architecture requirements.
As we proceed, we first identify, in Section [2](#S2 "2 Internet Routing Today ‣ On Compact Routing for the Internet"), the fundamental causes of Internet routing scalability problems, and how they imply the need for dramatically more efficient routing algorithms with rigorously proven worst-case performance guarantees. Such algorithms exist—they are known collectively as compact routing schemes. They do not and, generally, can not guarantee routing along shortest paths—they stretch them. In Section [3](#S3 "3 Routing stretch ‣ On Compact Routing for the Internet"), we juxtapose the formal notion of stretch (or hop stretch) with better-known forms of path inflation present in conventional networking. There is a fundamental trade-off between stretch and routing table (RT) size, and we show in Section [4](#S4 "4 Hierarchical routing, addressing, and their alternatives ‣ On Compact Routing for the Internet") why hierarchical routing and addressing fall short in finding an optimal balance point of this trade-off for topologies of interest. We also outline, in the same section, known alternative addressing techniques that compact routing schemes use. We analyze the key ideas behind these schemes in Section [5](#S5 "5 Universal compact routing ‣ On Compact Routing for the Internet"). We focus on universal schemes in this section and shift our attention to the schemes specialized for Internet-like topologies in Section [6](#S6 "6 Compact routing on scale-free graphs ‣ On Compact Routing for the Internet"). All these schemes are static. They ignore any communication overhead their implementations might require. We use analytic arguments in Section [7](#S7 "7 Dynamic routing on scale-free graphs ‣ On Compact Routing for the Internet") to directly prove that recent results in compact routing research render fundamentally unscalable the communication costs of any routing algorithm applied to Internet-like topologies. Despite these findings, “locator-identifier split” (LIS) proposals are becoming popular these days, as they are widely believed to scale better in the presence of network topology dynamics and host mobility. In Section [8](#S8 "8 Locator-identifier split (LIS) ‣ On Compact Routing for the Internet") we establish that LIS is a form of name-independent compact routing, discussed in Section [5](#S5 "5 Universal compact routing ‣ On Compact Routing for the Internet") and using topology-unaware (flat) addressing. We show that contrary to common belief, the scaling properties of such routing can only be worse compared to the topology-aware, name-dependent case. We employ simulations to demonstrate that both the RT sizes and stretch induced by the best-performing name-independent schemes are significantly worse compared to their name-dependent counterparts. We conclude and pose open problems in Section [9](#S9 "9 Conclusion ‣ On Compact Routing for the Internet").
## 2 Internet Routing Today
There are two main related problems with the Internet’s current routing architecture. First, it is not flexible, due primarily to the need to assign node addresses based on topological location. In order to achieve highly efficient address aggregation and relatively small routing tables, a CIDR network must have two characteristics: an essentially tree-like graph structure, and address assignment that follows this structure. Node mobility and end site multi-homing fundamentally break this model. But even recommended Internet traffic engineering techniques, e.g., load balancing, require de-aggregation of address blocks that cause the routing table to grow even if the network itself is not growing.
The second problem with the Internet’s routing architecture is how the interdomain routing algorithm (path vector, for BGP) and its protocol implementation scale with the size of the network. Poor scaling of a routing algorithm expresses itself in terms of rapid, e.g., linear, rates of growth of the routing table (RT) size. Poor scaling of RT sizes exacerbates convergence problems (convergence time, churn, instabilities, etc.). Not only is the communication overhead of BGP known to be exponential \[[3](#bib.bib3)\], but the BGP RT size also appears to grow exponentially \[[4](#bib.bib4)\].
The immediate causes of Internet RT growth have been extensively analyzed \[[1](#bib.bib1), [5](#bib.bib5), [6](#bib.bib6), [7](#bib.bib7), [8](#bib.bib8)\]; they reflect voluntary business relationships occurring at increasing density, traffic engineering techniques to improve robustness or reduce operational costs, and address allocation policies developed by IANA and the regional address registries—all resulting in various forms of de-aggregation, which is the primary cause of super-linear growth of the BGP RT size. Several studies, proposals, and even new routing architectures seek to mitigate these effects. Prominent efforts in this area include \[[9](#bib.bib9), [10](#bib.bib10), [11](#bib.bib11), [12](#bib.bib12), [13](#bib.bib13)\] and most recently \[[14](#bib.bib14), [15](#bib.bib15), [16](#bib.bib16)\]. However, these approaches do not rigorously provide non-trivial scaling guarantees, and many of them represent only short-term fixes since they address the symptoms rather than the root of the problem. More precisely, most of these approaches try to find shortest paths in the topology, leading to RT sizes that cannot scale better than linearly \[[17](#bib.bib17)\]. We are interested in algorithms that scale better than that. Ideally, we want to achieve logarithmic scaling
<sup>2</sup>
<sup>2</sup>2To be more precise, we mean polylogarithmic scaling, but we abuse the terminology and do not differentiate between logarithmic and polylogarithmic growth in this paper.
that comes close to satisfying our desire for “infinite scalability” in a future routing architecture \[[2](#bib.bib2)\].
As a representative example of the efforts mentioned above, we consider the idea of routing on autonomous system (AS) numbers, wherein AS numbers replace IP prefixes in the role of interdomain routing addresses. Using the coarser granularity of AS numbers is a natural idea, and is now a common part of many proposals \[[10](#bib.bib10), [12](#bib.bib12), [13](#bib.bib13), [14](#bib.bib14), [15](#bib.bib15)\]. At first glance, routing on AS numbers looks like a solution to the RT size problem. Indeed, such an approach could immediately reduce the RT size by an order of magnitude, since there are on the order of $`10^{5}`$ IP prefixes and $`10^{4}`$ AS numbers in the global RT today. However, one can easily overlook that it is just a one-time constant-factor reduction, and not a change of the scaling behavior. As such, it is not a solution to the core problem, but rather a temporary relief measure. Indeed, if algorithms routing on AS numbers are still shortest-path routing algorithms, then they still cannot produce RT sizes that grow slower than linearly with the network size measured, in this case, by the total number of ASs in the Internet. To make things even worse, this number currently grows faster than the total number of IP prefixes \[[4](#bib.bib4), [7](#bib.bib7)\], and if the trend continues after a hypothetical adoption of routing on AS numbers, then the actual rate of growth of RT sizes can only be worse than today. Furthermore, routing on AS numbers requires some form of global directory to map IP addresses or prefixes to the AS numbers they are associated with.
So far we have argued that inflexibility of the current routing architecture precludes long-term scalability. An ideal routing architecture would allow for assigning node addresses independent of topology, and would employ routing algorithms that can truly scale. No proposal thus far has simultaneously solved both objectives; address aggregation remains the only generally accepted method for effectively limiting RT size growth, yet operational and business imperatives require intentional violation of this tenet. Something has to give.
In order to find sustainable solutions to the global Internet routing problem, we need to investigate routing scalability at its most fundamental level. For interdomain routing, the framing parameters are the performance guarantees of routing algorithms operating on graphs with topologies similar to the Internet AS-level topology, and which can support topology-independent addressing of nodes. In particular, we seek rigorous lower and upper bounds on the RT size used in supporting all routing features, i.e., mapping node names to their locations, and reaching those locations. Compact routing research establishes these bounds and constructs algorithms (or schemes) that try to meet them. One of the central themes in compact routing is the investigation of the inevitable trade-off between RT sizes and routing stretch.
## 3 Routing stretch
The stretch of a routing algorithm is defined as the worst-case path-length increase factor relative to shortest paths. More specifically, for every pair of nodes in all graphs in the set of graphs the algorithm can operate on, we find the ratio of the length of the route taken by the algorithm to the length of the shortest available path between the same pair of nodes. The maximum of this ratio among all node pairs in all the graphs in the set is the algorithm’s stretch. The average stretch is the average of this ratio across all source-destination pairs in a given graph or set of graphs.
We emphasize that this notion of stretch has nothing in common with known forms of “path inflation” in contemporary Internet routing. Path inflation in today’s Internet is not due to the stretch of the underlying routing algorithm, but rather due to intra- and inter-domain routing policies, and various incongruities across multiple levels (geographical, router, AS, etc.) of network topology abstraction \[[18](#bib.bib18)\].
We are not aware of any widely-used routing protocol that produces other than stretch-1 (shortest) paths by default. Link-state (LS) in OSPF or ISIS, distance-vector (DV) in RIP or (E)IGRP, LS/DV hybrid in LVA \[[19](#bib.bib19)\], and path-vector (PV) in BGP, are all forms of trivial shortest-path routing (stretch-1). Consider two examples: BGP and OSPF. Non-trivial policy configurations generally prevent BGP from routing along the actual shortest path, but without policies the BGP route selection process would always select the shortest paths in the AS-level graph. Similarly, non-trivial area configurations prevent OSPF from routing along the shortest paths, but routing inside an area is always along the shortest path in the weighted router-level graph.
The above examples seem straightforward, but we identify them to eliminate any possible confusion between stretch of a routing algorithm and inflation of paths produced by a routing protocol subjected to policy or configuration constraints. Clarifying the difference allows us to formulate both necessary and sufficient conditions for the stretch of a routing algorithm proposed for use in the Internet.
The sufficient routing stretch requirement is that the routing algorithm underlying any realistic interdomain routing protocol must be of stretch-1. To understand the reasoning behind this requirement, consider a stretch$`>`$1 routing algorithm and apply it to a complete graph, which has all nodes directly connected to all other nodes so that all shortest paths are of length 1. Since this algorithm does not guarantee finding all shortest paths, RTs at some nodes might lack entries corresponding to their own directly-connected neighbors. Put in different terms, in order to shrink RTs, the algorithm might have to force some nodes to delete information about some of their directly-attached interfaces.
Of course, the Internet interdomain topology is not a full mesh, and it might turn out that the same routing algorithm that fails to be stretch-1 on a full mesh does in fact find all shortest paths to each node’s neighbors in realistic Internet-like topologies. We thus might also pursue the explicit task of finding a non-universal routing algorithm that is applicable only to Internet-like topologies. We do not require this algorithm be stretch-1, but it must be of stretch-1 on paths of length 1 leading to nodes’ neighbors. In other words, the necessary routing stretch requirement is that the routing algorithm underlying any realistic interdomain routing protocol must be of stretch-1 on paths of length 1 in realistic Internet-like topologies.
These requirements are purely practical: if a link exists between a pair of autonomous systems, it must be made available for carrying traffic. If our “sophisticated” routing algorithm removes routing information about such links from routing tables, then network operators will feel compelled to manually reinsert it, which counters the scalability objective of reducing the RT size and requires manual intervention which we would in general like to avoid. We will refer to this (undesirable) administrative re-adding of information about shortest paths to nodes’ neighbors as manual reinsertion.
The absence of an immediate neighbor in a node’s RT appears to contradict common sense, but one can verify that this situation can occur with the algorithms we discuss in Section [5](#S5 "5 Universal compact routing ‣ On Compact Routing for the Internet"). We have not had to deal with this effect in practice yet because the routing algorithms in operation today are all stretch-1. Stretch-1 algorithms require complete knowledge of the network topology (or at least of distances between nodes in it) and maintenance of an individualized optimal next hop for every possible destination in the network. This means $`\Omega\hspace{0pt}{(n)}`$ entries are required in the RTs of every node in the worst case. In order to shrink RTs below this size, we must prepare to lose topological information, and consequently lose optimal next hops for some destinations. Balancing this fundamental and unavoidable trade-off between stretch of a routing algorithm and the sizes of RTs it produces is at the center of compact routing research today, but the first formalization of this trade-off was introduced in early work on hierarchical routing and addressing.
## 4 Hierarchical routing, addressing, and their alternatives
In 1977 Kleinrock and Kamoun published their pioneering paper on hierarchical routing \[[20](#bib.bib20)\]. They showed how hierarchical clustering could be used with appropriate node addresses to produce highly scalable routing tables. This technique is the basis of the popular hierarchical routing approaches used today in practice, e.g., in the form of CIDR in the interdomain routing case, or usage of OSPF/ISIS areas in the intradomain routing case. The key idea is to group (or aggregate) nearby nodes into clusters, clusters into super-clusters, and so on, in a bottom-up fashion. Determining these clusters can be accomplished through a hierarchical clustering algorithm, the performance of which is key to good routing performance.
Hierarchical aggregation and addressing offer substantial RT size reduction by abstracting out unnecessary topological details about remote portions of the network: nodes in one (super-)cluster need to keep only one RT entry for all nodes in another (super-)cluster. Almost all proposals for future Internet routing architectures trying to address the RT size problem are based, explicitly or implicitly, on this concept of hierarchical routing. In fact, no other concept behind RT size reduction has been systematically considered in the networking literature. Unfortunately, hierarchical routing is simply not a good candidate for interdomain routing, despite its current widespread use and reputation as realistically scalable, and we show why in the rest of this section.
In the same paper \[[20](#bib.bib20)\], Kleinrock and Kamoun were the first to analyze the stretch/RT size trade-off. They showed that the routing stretch produced by the hierarchical approach is satisfactory only for topologies with average shortest path hop-length (distance) $`\overline{d}\hspace{0pt}{(n)}`$ that grows quickly with network size $`n`$, which means it performs well only for graphs where large distances between nodes prevail. In fact, they assumed that the average distance grows polynomially with the network size, $`{\overline{d}\hspace{0pt}{(n)}} \sim n^{\nu}`$ with $`\nu > 0`$. Such graphs have many remote nodes, i.e., nodes at long hop distances from each other, which allows for efficient aggregation of topology details without substantial path length increase.
### 4.1 Grids
The simplest example of graphs with average distance $`{\overline{d}\hspace{0pt}{(n)}} \sim n^{\nu}`$ are $`\delta`$-dimensional grids. The exponent $`\nu`$ is $`1/\delta`$ in this case, and node aggregation works naturally: clusters are composed from nodes located at small Manhattan distances from each other, while super-clusters contain neighboring clusters, and so on.
We emphasize that aggregation is not the only technique to achieve logarithmic RT scaling for grids. We could alternatively embed a $`\delta`$-dimensional grid into a $`\delta`$-dimensional Euclidean space, and address nodes by their coordinates in the space. With such an embedding, nodes would not need to keep any RT information except for their own addresses, since knowing their own coordinates and the coordinates of the destination allows for shortest-path routing according to the Euclidean, i.e., Manhattan, distance. The RT size thus scales as the address size, i.e., logarithmically.
Logarithmic scaling extends from $`\delta`$-dimensional grids to graphs with doubling dimension $`\delta`$ \[[21](#bib.bib21)\]. The doubling dimension of a graph, or more generally, of a metric space, is the smallest $`\delta`$ such that every ball of radius $`2\hspace{0pt}r`$ can be covered by $`2^{\delta}`$ balls of radius $`r`$. Informally, such graphs are “perturbed versions” of $`\delta`$-dimensional grids. A key idea behind efficient addressing in such graphs allowing for logarithmic RT size growth is to introduce distance scales \[[22](#bib.bib22), [23](#bib.bib23)\] or maps of node coordinates. The address of a node becomes a collection of its coordinates, which are the distances from the node to carefully constructed subsets of nodes that form a “frame of reference.” In other words, knowing the distances to these frame-of-reference subsets, we can uniquely identify a node’s location in the graph. One can show then that in order to bring stretch arbitrarily close to 1, we need only a logarithmic number of such subsets, which guarantees logarithmic RT sizes.
### 4.2 Trees
A strikingly similar situation occurs with routing on regular, i.e., $`b`$-ary, trees. They also allow for efficient aggregation: combine leaves of a single parent into a single cluster, while super-clusters contain clusters with a common parent of the parent, and so on.
Yet again we find that aggregation is not the only path toward logarithmic RT scaling on such trees. For example, we can use node addressing induced by a depth-first-search (DFS) node numbering. Given a DFS-based addressing convention, any node can infer its exact location on the tree from its address. In particular, a node knows what nodes are its children or descendants of which one of its children, and what nodes are not its descendants.
As with grids, logarithmic scaling extends from regular $`b`$-ary trees to a wider class of similar graphs—to arbitrary trees in this case \[[24](#bib.bib24), [25](#bib.bib25)\]. The main idea behind efficient addressing schemes is still to use DFS numbering, but we can no longer rely only on the node DFS ID to identify its location on a tree. It turns out however that it is possible to encode the path from the root of a tree to a node in such a way that this encoding preserves the logarithmic upper bound. If we now make the sequence of nodes lying on the path to the root a part of the node address, then one can check that routing along shortest paths becomes trivially possible.
### 4.3 Grids and trees
In both trees and grids, shortest-path routing is possible given only the addresses of the current and destination nodes. The RT size in either case scales logarithmically. This scaling performance can be extended beyond regular grids and trees to much wider classes of graphs, but not with hierarchical routing. The simplest example is routing on star graphs. These graphs are trees, but no hierarchical aggregation is possible on them. Logarithmic scaling in such cases is possible due only to specialized addressing techniques that leverage topological peculiarities of a graph’s structure to succinctly encapsulate the location of a node into its address. Of course, these addressing techniques work trivially for regular grids and trees as well.
Recent progress in this area indicates that the efficiency of these types of addressing schemes is directly related to existence of low-distortion embeddings of finite metric spaces induced by shortest-path distances in grids and trees into normed spaces of low dimension \[[26](#bib.bib26)\]. Low dimensions of host spaces guarantee small RT sizes, while low distortion of embeddings corresponds to small stretch. Grids embed in flat Euclidean spaces, but trees require spaces of negative curvature \[[27](#bib.bib27), [28](#bib.bib28)\].
### 4.4 Scale-free graphs
According to the best available data \[[29](#bib.bib29)\], the Internet topology is neither a tree nor a grid. Rather, it is scale-free. To avoid terminology disputes, by scale-free we simply mean networks with heavy-tail, e.g., power-law, node degree distributions and strong clustering, i.e., large numbers of triangular subgraphs.
<sup>3</sup>
<sup>3</sup>3There is an unfortunate terminological collision here: clustering means both grouping nodes into areas and presence of triangles in a graph.
The latter characteristic implies that such graphs are not trees at all. The former, on the other hand, means that in theory, the average distance in them grows with the network size much more slowly than Kleinrock and Kamoun \[[20](#bib.bib20)\] assumed: at most $`{\overline{d}\hspace{0pt}{(n)}} \sim {\log n}`$ \[[30](#bib.bib30)\]. In practice, the average hop distance between autonomous systems in the Internet stays virtually constant or even decreases due to increasing inter-AS connectivity driven by economic or performance considerations \[[31](#bib.bib31), [7](#bib.bib7), [4](#bib.bib4)\]. According to multiple data sources, the average AS-hop distance in the current Internet is between 3.1 and 3.7, with more than 80% of AS pairs being 2-4 hops away from each other \[[29](#bib.bib29)\]. Topologies with small average distances are also called small worlds. There are essentially no remote nodes in small-world networks; all nodes are close to each other.
These characteristics are extremely bad news for hierarchical routing approaches, because the effectiveness of hierarchical network partitioning and aggregation depends either on the abundance of remote notes or on strong regularity of $`b`$-ary tree structure. None of these properties is present in scale-free graphs. In short, we are faced with an unsettling reality: one cannot hope to find a way to efficiently apply hierarchical, aggregation-based routing to Internet-like topologies.
These general arguments do not preclude the possibility of applying some generic hierarchical routing algorithm to the Internet. But they do imply that such an application cannot be efficient and we must prepare to see high stretch. Simple analytical estimates \[[32](#bib.bib32)\] show that applying hierarchical routing to an Internet AS-level topology incurs a $`\sim`$15-times path length increase, which, although alarming enough by itself, would also lead to a substantial RT size surge caused by manual reinsertion.
If we accept our argument in Section [3](#S3 "3 Routing stretch ‣ On Compact Routing for the Internet") that any truly scalable Internet routing algorithm must have stretch as close as possible to 1, we must also accept the fact that hierarchical routing will not meet our needs. Since all previous Internet interdomain routing proposals are heavily based on hierarchical routing, we recognize that while deploying one of them might offer short-term relief, none of them would genuinely scale to meet future demands.
## 5 Universal compact routing
Compact routing research establishes fundamental limits for routing scalability, and constructs concrete routing algorithms (or schemes) that try to meet these limits. Formally, a routing scheme is said to be compact if it produces logarithmic address and header sizes, a sublinear RT size, and stretch bounded by a constant, as opposed to the growing stretch more typical of hierarchical routing.
<sup>4</sup>
<sup>4</sup>4For example, one can verify that on an $`n`$-node full mesh, \[[20](#bib.bib20)\] produces stretch growing to infinity as $`\Theta\hspace{0pt}{({\log n})}`$.
A routing scheme is universal if it works correctly and satisfies promised scaling bounds on all graphs. If it does so only on some specific graph classes, it is called specialized.
We can express fundamental limits of routing scalability using lower bounds on RT size versus stretch. For example, we can fix the maximum stretch and establish the minimum RT size for this stretch. Such a task usually involves constructing a concrete worst-case graph, enumerating RT size requirements of all possible routing schemes, and demonstrating that no routing scheme, applied to this worst-case graph, produces an RT smaller than the minimum \[[33](#bib.bib33)\]. As soon as we have constructed a concrete routing scheme, we can establish its upper bounds, i.e., the maximum RT size and stretch it produces across all graphs. If a scheme’s upper bound equals the theoretical lower bound, the scheme is said to be optimal.
The lower bounds for universal stretch-1 (shortest-path) routing are somewhat pessimistic. We first notice that we can construct RTs at each node by storing, for every destination node, the ID of the outgoing port on the shortest path to the destination. The number of destinations is $`n - 1`$, the maximum number of ports a node can have (equal to maximum possible node degree) is also $`n - 1`$, and thus a maximum $`O\hspace{0pt}{({n\hspace{0pt}{\log n}})}`$ bits of memory is required for an RT, i.e., $`O\hspace{0pt}{({n\hspace{0pt}{\log n}})}`$ is its upper bound. This construction is called trivial shortest-path routing, and all deployed LS-, DV-, or PV-based routing protocols implement it. However, Gavoille and Pérennès \[[17](#bib.bib17)\] showed that the lower bound of universal stretch-1 routing is also $`\Omega\hspace{0pt}{({n\hspace{0pt}{\log n}})}`$, i.e., there is no shortest path routing scheme that for all nodes in all graphs, guarantees RTs smaller than the RTs of trivial shortest-path routing. In other words, shortest-path routing is incompressible. In order to decrease the maximum RT size, we therefore must allow maximum stretch to increase above 1.
Gavoille and Genegler \[[33](#bib.bib33)\] showed that for any stretch strictly below 3, the local space lower bound is nearly the same as for stretch-1 routing, $`\Omega\hspace{0pt}{(n)}`$, meaning that no universal stretch$`<`$3 routing scheme can guarantee sublinear RT sizes: the minimum value of maximum stretch allowing sublinear RT sizes is 3. Thorup and Zwick \[[34](#bib.bib34)\] proved that any routing scheme with stretch strictly below 5 cannot guarantee space smaller than $`\Omega\hspace{0pt}{(n^{1/2})}`$. It is widely believed (although not yet proven) that stretch below $`{2\hspace{0pt}k} - 1`$ imposes an RT size lower bound of $`\Omega\hspace{0pt}{(n^{1/k})}`$ for all positive integers $`k`$.
Besides being universal or specialized, another important classification of routing schemes is whether they support name independence. Name-dependent schemes embed some topological information in node addresses (or labels) which thus cannot be arbitrary. We analyze a few classic examples of such addressing in Section [4](#S4 "4 Hierarchical routing, addressing, and their alternatives ‣ On Compact Routing for the Internet"). Name-independent schemes work on graphs with arbitrary, e.g., flat, node labels. As discussed in Section [2](#S2 "2 Internet Routing Today ‣ On Compact Routing for the Internet"), scalable name-independent schemes are highly desirable for future Internet routing. The distinction between these two cases is directly related to a popular discussion theme in the networking community, in which terms such as “node name” or “node identifier” essentially refer to the name-independent, topologically agnostic label, while terms “node address” or “node locator” usually imply a topologically informative node label, i.e., the name-dependent label. Somewhat surprisingly, the lower and upper bounds for both name-dependent and name-independent universal routing are essentially the same.
### 5.1 Name-dependent routing
Kleinrock and Kamoun’s hierarchical approach \[[20](#bib.bib20)\] was essentially the first name-dependent routing scheme. Besides problems discussed in Section [4](#S4 "4 Hierarchical routing, addressing, and their alternatives ‣ On Compact Routing for the Internet"), their scheme provided no algorithm to construct a required partitioning for a given network. In subsequent works, first Kleinrock himself, then Perlman, and more recently Awerbuch, Peleg and others, invested much effort trying to find efficient network partitioning algorithms needed for correct operation of hierarchical routing.
It became evident that hierarchical routing approaches were intrinsically suboptimal when Cowen delivered her compact routing scheme \[[35](#bib.bib35)\]. The scheme required no network partitioning, i.e., it was non-hierarchical, and it was fairly simple compared to numerous hierarchical routing algorithms accumulated by that time. It guaranteed the maximum stretch 3 and RT size of $`\overset{\sim}{O}\hspace{0pt}{(n^{2/3})}`$. Thorup and Zwick \[[25](#bib.bib25)\] (TZ) soon improved on Cowen’s RT size upper bound, bringing it to $`\overset{\sim}{O}\hspace{0pt}{(n^{1/2})}`$ while maintaining stretch of 3. The TZ scheme was thus the first nearly optimal universal stretch-3 routing scheme because its RT size upper bound was nearly equal to the existing lower bound, up to logarithmic factors implied by ‘ $`\overset{\sim}{}`$ ’ in the $`\overset{\sim}{O}`$ notation.
The TZ scheme is heavily based on the Cowen scheme. Both schemes first preprocess the graph and select a set of landmarks in it. In the Cowen scheme, this selection is based on results for dominating set construction, while the TZ scheme selects landmarks by means of randomized techniques. The landmark set sizes in the Cowen and TZ cases are $`\overset{\sim}{O}\hspace{0pt}{(n^{2/3})}`$ and $`\overset{\sim}{O}\hspace{0pt}{(n^{1/2})}`$, and that is essentially the only difference between the two schemes. The most non-trivial part in both the algorithms is to guarantee a proper balance between the sizes of landmark sets and clusters. Clusters are defined for every node $`v`$ in a graph, and $`v`$’s cluster is the set of nodes that are closer to $`v`$ than to their corresponding closest landmarks. The TZ scheme, for example, guarantees that not only the landmark set size, but also the cluster size is upper-bounded by $`\overset{\sim}{O}\hspace{0pt}{(n^{1/2})}`$. The schemes then assign to all nodes $`v`$ the addresses, or labels, consisting of the following three parts: $`v`$’s original identifier, the identifier of the landmark $`L\hspace{0pt}{(v)}`$ closest to $`v`$, and the identifier of the interface at $`L\hspace{0pt}{(v)}`$ that lies on the shortest path from $`L\hspace{0pt}{(v)}`$ to $`v`$. The RT at $`v`$ consists of the next hops along shortest paths to all landmarks and all nodes in $`v`$’s cluster. The RT size is thus $`\overset{\sim}{O}\hspace{0pt}{(n^{1/2})}`$. Node $`v`$ forwards a packet along the shortest path if the destination is either a landmark or a node in $`v`$’s cluster. If $`v`$ is the closest landmark for the destination, $`v`$ forwards the packet along the shortest path using the interface identifier in the address in the packet header. Otherwise, $`v`$ forwards the packet to the landmark closest to the destination—its identifier is in the packet header and the next hop on the shortest-path route to it is in $`v`$’s RT. It is a trivial exercise to check that the worst-case stretch experienced by such a routing scheme is 3.
### 5.2 Name-independent routing
Abraham et al. \[[36](#bib.bib36)\] improved on previous results by Arias et al. \[[37](#bib.bib37)\] and delivered the first nearly optimal universal name-independent routing scheme, which had the same upper bounds as the name-dependent TZ scheme—maximum stretch of 3 and $`\overset{\sim}{O}\hspace{0pt}{(n^{1/2})}`$ RT sizes.
All name-independent schemes known to us consist of the following two parts:
- •
a name-dependent compact routing scheme operating underneath the name-independent one; and
- •
dictionary tables forming an efficiently-distributed
database containing information necessary to translate nodes’ name-independent addresses (flat identifiers) to name-dependent addresses (locators).
An important property of the dictionary tables is that they must be “everywhere dense” in the topology, so that any node $`v`$ can resolve all identifiers to locators by consulting dictionary tables of nodes located in $`v`$’s neighborhood only. Obviously, node $`v`$ has to keep all routes to nodes in its neighborhood. Routing proceeds as follows: upon receiving a packet destined to identifier $`X`$, node $`v`$ forwards a packet to node $`u`$ in its neighborhood that according to the dictionary table construction rules (dependent on the routing scheme), contains an identifier-to-locator mapping for $`X`$. Upon reaching $`u`$, the packet learns the locator $`x`$ for $`X`$ and proceeds to $`x`$ as dictated by the underlying name-dependent scheme.
We emphasize that identifier-to-locator mapping is an integral part of the routing process, and all performance bounds apply to the combined process of mapping and forwarding. Specifically, name-independent RTs include both the RTs of underlying name-dependent schemes and the dictionary tables, whereas total stretch accounts for travel to a name-resolution node and then from it to the destination along a name-dependent, potentially stretched route. This is in contrast to the separation of DNS name resolution and IP routing in the Internet, where DNS is essentially an application running on top of IP and requiring routing to work in the first place. It is therefore remarkable that there exist both name-dependent and name-independent universal stretch-3 compact routing schemes with the same $`\overset{\sim}{O}\hspace{0pt}{(n^{1/2})}`$ RT size upper bound.
## 6 Compact routing on scale-free graphs
One might be alarmed by the previous section, as it seems to conflict with Section [3](#S3 "3 Routing stretch ‣ On Compact Routing for the Internet") where we show that because of practical reasons, stretch must be close to 1. The combination of shortest-path routing incompressibility and the impossibility of stretch$`<`$3 routing with sublinear RT sizes seem to imply that scalable interdomain routing is impossible. Recall, however, that the stretch values mentioned in Section [5](#S5 "5 Universal compact routing ‣ On Compact Routing for the Internet") refer to the maximum stretch across all paths in all graphs, meaning that there exists some worst-case graph
<sup>5</sup>
<sup>5</sup>5See \[[33](#bib.bib33)\] for an example of these worst-case graphs.
on which this maximum is achieved. More importantly, all schemes mentioned in Section [5](#S5 "5 Universal compact routing ‣ On Compact Routing for the Internet") are universal. Since they can work on all graphs, the worst-case graph may not be, and usually is not, “Internet-like.” We can consequently expect the average stretch of these schemes on scale-free graphs be lower than 3. But how close is it to 1?
At first glance, it seems unlikely that it could be too close to 1, since we saw in Section [4.4](#S4.SS4 "4.4 Scale-free graphs ‣ 4 Hierarchical routing, addressing, and their alternatives ‣ On Compact Routing for the Internet") that the stretch of hierarchical routing is high on Internet-like graphs, and there is no apparent reason to believe that compact routing would be drastically better. In \[[32](#bib.bib32)\] however, the authors showed that the average performance of the first optimal stretch-3 scheme (TZ) \[[25](#bib.bib25)\] on scale-free topologies is much better than its worst case. While TZ’s upper bounds are around 2200 for RT size and 3 for stretch, the average RT size was found to be around 50 entries for the whole AS-level Internet, and the average stretch was only 1.1! These numbers are striking evidence that, on the same topology and under the same assumptions, compact routing has dramatically superior scaling characteristics compared to hierarchical routing \[[20](#bib.bib20)\], with its stretch of 15 and RTs of unbounded size, and compared to simple routing on AS numbers (e.g., \[[15](#bib.bib15)\]), with RT sizes of the order of $`10^{4}`$.
The observation in \[[32](#bib.bib32)\] that scale-free graphs yield essentially the best possible performance of the TZ scheme compared to all other graphs was the first indication that scale-free graphs are optimally structured for high-performance routing. This indication inspired several research groups to work on construction of specialized routing schemes \[[38](#bib.bib38), [39](#bib.bib39)\] designed to utilize structural peculiarities of scale-free graphs in order to improve routing performance guarantees. The schemes in \[[38](#bib.bib38), [39](#bib.bib39)\] achieve logarithmic RT scaling and infinitesimally small stretch on scale-free graphs. Both schemes are based on the observation that although scale-free graphs are not trees, trees cover scale-free graphs without incurring much stretch.
As a representative example, the scheme by Brady and Cowen \[[38](#bib.bib38)\] (“BC scheme”) grows one shortest-path tree rooted at the highest-degree node and a small fixed number of additional trees to cover edges in the fringe of a power-law graph. Routing is then confined to this tree collection according to the schemes discussed in Section [4.2](#S4.SS2 "4.2 Trees ‣ 4 Hierarchical routing, addressing, and their alternatives ‣ On Compact Routing for the Internet"). Since their RT sizes scale logarithmically, the BC scheme achieves logarithmic scaling. Specifically, it guarantees a maximum RT size of $`O\hspace{0pt}{({\log^{2}n})}`$. The $`O\hspace{0pt}{({\log^{2}n})}`$ scaling means that even if the graph size is $`2^{128}`$, i.e., if all IPv6-capable devices independently participate in global routing, then the maximum routing table size in such an Internet would still contain not more than $`128^{2} \sim {16,000}`$ entries! The corresponding limit for IPv4 is $`\sim {1,000}`$.
We thus see that even though hierarchical routing approaches based on address aggregation cannot work efficiently on Internet-like topologies (Section [4](#S4 "4 Hierarchical routing, addressing, and their alternatives ‣ On Compact Routing for the Internet")), more modern routing algorithms achieve essentially “infinite” scalability on scale-free graphs, if they are static. Can there exist any routing schemes that scale as well on dynamic networks?
## 7 Dynamic routing on scale-free graphs
One of the most limiting assumptions behind the routing algorithms we have considered thus far is that they assume that the network is effectively static. More precisely, it is not that the network is actually static per se, but that the routing algorithms require a full view of the graph representing the network topology at any given instant in time. Any topology change yields a new graph, on which the algorithm might have to perform its calculations from scratch. This behavior is characteristic of currently-deployed Internet routing algorithms as well.
The most basic parameter measuring the dynamic performance of routing algorithms is the communication cost—the number of control messages needed to converge after a topology change. Communication costs are as important as RT size or stretch. They are associated with nodes requiring properly updated information about the network topology in order to route. Deployed routing protocols, for example, require all nodes to have a coherent full view of the network topology or of the distances in it. Indeed, link-state algorithms, typically used for intra-domain routing, require each node to have a complete view of all links in the network, and each node performs a shortest-path computation on essentially a static graph. Distance- or path-vector algorithms need each node to know distances or paths to all other nodes in the network.
In order to achieve such coherent full views of the network topology, timely update messages, a.k.a. routing updates, are employed. Updates requiring recalculation of routing tables can lead to delay, instabilities, churn, and other complications. For example, there are BGP oscillation scenarios with infinite communication costs \[[40](#bib.bib40), [41](#bib.bib41)\], and even during more typical BGP behavior, they are exponential \[[3](#bib.bib3)\]. High communication costs and resulting long convergence times are both a critical problem and an absolutely inevitable implication of growth of the current routing architecture \[[1](#bib.bib1)\].
Using fairly general arguments, Afek et al. \[[42](#bib.bib42)\] showed that the communication cost cannot scale better than $`\Omega\hspace{0pt}{(n)}`$. Their results applied to universal schemes, so we could still hope that schemes specialized for scale-free graphs would behave better, as the $`\Omega\hspace{0pt}{(n)}`$ bound could be attained only on some marginal worst-case graphs. After seventeen years of little progress on this problem, Korman and Peleg have recently tried to improve this pessimistic lower bound in \[[43](#bib.bib43)\]. Unfortunately for Internet routing, their results are even more pessimistic than Afek’s since we use them, combined with simple analytic estimates, to demonstrate in the Appendix that the communication cost lower bound for scale-free graphs is $`\overset{\sim}{\Omega}\hspace{0pt}{(n)}`$. This finding means that as far as communication costs are concerned, Internet-like graphs belong to the class of worst-case graphs, across all possible network topologies!
## 8 Locator-identifier split (LIS)
Despite all the findings mentioned above, a popular proposed approach to Internet routing scalability and flexibility is to separate the identifier and locator of a node in the network. In an architecture where one label identifies a node and a different label indicates its location, topology changes will only change the locators, which are assumed to follow topology and allow for aggressive aggregation. Other functions (access control lists, application and transport layer identification of endpoints) can use the identifiers. This would support significantly enhanced flexibility of the Internet routing system, but does it really address the scalability concerns?
In light of the discussion in Sections [4.4](#S4.SS4 "4.4 Scale-free graphs ‣ 4 Hierarchical routing, addressing, and their alternatives ‣ On Compact Routing for the Internet"), [5.2](#S5.SS2 "5.2 Name-independent routing ‣ 5 Universal compact routing ‣ On Compact Routing for the Internet"), and [7](#S7 "7 Dynamic routing on scale-free graphs ‣ On Compact Routing for the Internet"), we must admit that LIS cannot really improve routing scalability, for at least two reasons:
- •
Since LIS approaches are based, explicitly or implicitly, on aggressive aggregation of locators, they cannot improve scaling of RT sizes because such aggregation is impossible on scale-free topologies;
- •
Even if LIS approaches did not use aggressive aggregation, they still could not improve scaling, because in addition to maintaining and updating RTs of locators as before, the networks must also maintain and update a (distributed) database of identifier-to-locator mappings.
To ground these claims more firmly, we first observe that name-independent routing from Section [5.2](#S5.SS2 "5.2 Name-independent routing ‣ 5 Universal compact routing ‣ On Compact Routing for the Internet") naturally implements LIS. In fact, the only conceptual difference between name-independent routing and LIS is that the former is more flexible than the latter, since it does not assume that locators, i.e., name-dependent addresses, are amenable to aggressive aggregation. It is obvious, however, that the average scaling characteristics of name-independent schemes cannot be better than name-dependent ones, since the name independent schemes are essentially name dependent schemes plus something else, i.e., plus dictionary tables and identifier-to-locator name-resolution, which incurs both RT size increase and stretch, cf. Section [5.2](#S5.SS2 "5.2 Name-independent routing ‣ 5 Universal compact routing ‣ On Compact Routing for the Internet").
To experimentally confirm this reasoning, we implemented the best-performing name-dependent and name-independent schemes. Specifically, among name-dependent schemes, we chose the universal optimal stretch-3 TZ scheme \[[25](#bib.bib25)\] and the BC scheme \[[38](#bib.bib38)\] designed for scale-free graphs. We also tested the TZ/BC hybrid scheme, which simply selects the lower-stretch path by running both schemes in parallel. For the name-independent case, we tested the optimal universal Abraham scheme \[[36](#bib.bib36)\]. We then applied all schemes to AS-level Internet topologies measured by skitter \[[44](#bib.bib44)\] and DIMES \[[45](#bib.bib45)\], and calculated the resulting average RT size and stretch. Table [1](#S8.T1 "Table 1 ‣ 8 Locator-identifier split (LIS) ‣ On Compact Routing for the Internet") and Figure [1](#S8.F1 "Figure 1 ‣ 8 Locator-identifier split (LIS) ‣ On Compact Routing for the Internet") show the outcome of these experiments, revealing that all characteristics of the name-independent routing are significantly worse than their name-dependent counterparts. By no means do these observations conflict with our statements in Section [5](#S5 "5 Universal compact routing ‣ On Compact Routing for the Internet") that the worst-case scaling of name-dependent and name-independent universal routing schemes is the same. They do show, however, that name-independent scaling on Internet topologies is much worse on average.
Table 1: The average stretch and routing table sizes (in entries and bits) that the best performing name-dependent (BC, TZ, and their hybrid TZ/BC) and name-independent (Abraham) schemes produce on the skitter and DIMES topologies.
<table>
<thead>
<tr>
<th>Scheme</th>
<th colspan="2">skitter (9204 nodes)</th>
<th colspan="2">DIMES (13931 nodes)</th>
</tr>
</thead>
<tbody>
<tr>
<th>TZ</th>
<td>62 entries (1629 bits),</td>
<td>1.08 stretch</td>
<td>69 entries (1887 bits),</td>
<td>1.13 stretch</td>
</tr>
<tr>
<th>BC</th>
<td>22 entries (1025 bits),</td>
<td>1.06 stretch</td>
<td>22 entries (1103 bits),</td>
<td>1.03 stretch</td>
</tr>
<tr>
<th>TZ/BC hybrid</th>
<td>84 entries (2654 bits),</td>
<td>1.02 stretch</td>
<td>91 entries (2990 bits),</td>
<td>1.01 stretch</td>
</tr>
<tr>
<th>Abraham</th>
<td>3251 entries (240427 bits),</td>
<td>1.35 stretch</td>
<td>4145 entries (326781 bits),</td>
<td>1.45 stretch</td>
</tr>
</tbody>
</table>
[Refer to caption](/html/0708.2309/assets/x1.png)
[Refer to caption](/html/0708.2309/assets/x2.png)
[Refer to caption](/html/0708.2309/assets/x3.png)
[Refer to caption](/html/0708.2309/assets/x4.png)
Figure 1: The distributions of stretch (top row) and routing table size (bottom row) that the hybrid TZ/BC (left column) and Abraham (right column) schemes produce on the skitter and DIMES topologies.
Furthermore, our findings are consistent with the scaling properties of name-independent routing on trees. We mention trees here because tree routing from Section [4.2](#S4.SS2 "4.2 Trees ‣ 4 Hierarchical routing, addressing, and their alternatives ‣ On Compact Routing for the Internet") forms the core of routing schemes specialized for Internet topologies in Section [6](#S6 "6 Compact routing on scale-free graphs ‣ On Compact Routing for the Internet"). Unfortunately, as found in \[[46](#bib.bib46), [47](#bib.bib47)\], name-independent routing on trees cannot scale better than on general graphs. In other words, trees are the worst-case graphs for universal name-independent routing.
<sup>6</sup>
<sup>6</sup>6More specifically, Abraham et al. \[[47](#bib.bib47)\] explicitly use stars to prove the name-independent lower bounds.
The reason for that is simple: trees cannot guarantee that nodes’ neighborhoods, i.e., balls of small radii that store dictionary tables, comprise only small portions of the graph. Examples of graphs that satisfy this property are grids or their perturbations from Section [4.1](#S4.SS1 "4.1 Grids ‣ 4 Hierarchical routing, addressing, and their alternatives ‣ On Compact Routing for the Internet"), i.e., graphs of low doubling dimension. The average distance in these graphs grows quickly with the graph size, the relative size of nodes’ neighborhoods can thus be made small, and as a consequence, the RT sizes of name-independent routing scale logarithmically on such graphs \[[21](#bib.bib21)\]. We summarize these scaling characteristics in Table [2](#S8.T2 "Table 2 ‣ 8 Locator-identifier split (LIS) ‣ On Compact Routing for the Internet").
Table 2:
Routing table (RT) size scaling for name-dependent and name-independent routing on different graph classes. Both “Logarithmic” and “Polynomial” tags in this table have specific meanings. “Logarithmic” means that RT sizes scale as $`\Omega\hspace{0pt}{({\log n})}`$ even for shortest-path (stretch-1) routing. “Polynomial” means $`\Omega\hspace{0pt}{(n)}`$ for stretch-1 routing, $`\Omega\hspace{0pt}{(\sqrt{n})}`$ for stretch-3 routing, and more generally, $`\Omega\hspace{0pt}{(n^{1/k})}`$ for routing with stretch of $`{2\hspace{0pt}k} - 1`$, $`k = {1,2,\ldots}`$.
<table>
<thead>
<tr>
<th>Graph class</th>
<th>Name-dependent routing</th>
<th>Name-independent routing</th>
</tr>
</thead>
<tbody>
<tr>
<td>General graphs</td>
<td>Polynomial</td>
<td>Polynomial</td>
</tr>
<tr>
<td>Trees</td>
<td>Logarithmic</td>
<td>Polynomial</td>
</tr>
<tr>
<td>Grids</td>
<td>Logarithmic</td>
<td>Logarithmic</td>
</tr>
</tbody>
</table>
## 9 Conclusion
Compact routing is a research area that aims at solving the following two classes of problems: 1) identify the fundamental scaling limits (lower bounds) of routing on graphs; and 2) construct routing algorithms that meet those limits, i.e., whose upper bounds are equal to the corresponding lower bounds. In other words, compact routing reveals the most fundamental properties of routing scalability, refined from layers of complexity associated with routing in practice.
As such, compact routing algorithms offer remarkably better scaling than we observe in the Internet routing today. In particular, existing interdomain routing exhibits exponentially growing routing table sizes, while those of compact routing algorithms designed for Internet-like topologies scale logarithmically.
Logarithmic scaling is essentially the only scaling behavior that satisfies the requirement for a future routing architecture to “scale indefinitely” \[[2](#bib.bib2)\]. Unfortunately, only static and name-dependent, i.e., topology-aware, routing exhibit logarithmic scaling on observed Internet topologies.
If name-independence is a requirement, i.e., if we need to route on topology-unaware flat identifiers, then routing tables cannot scale better than polynomially on Internet-like topologies. The lower bound is $`\Omega\hspace{0pt}{(n)}`$ for shortest-path routing, $`\Omega\hspace{0pt}{(\sqrt{n})}`$ for stretch-3 routing, and more generally, $`\Omega\hspace{0pt}{(n^{1/k})}`$ for routing with stretch of $`{2\hspace{0pt}k} - 1`$, $`k = {1,2,\ldots}`$.
Worse than that, the communication cost, i.e., the number of routing control messages per topology change, cannot grow slower than linearly on Internet-like topologies. Of course, this scaling is still much better than the exponential communication costs of deployed routing protocols, but linear scaling can definitely not be considered as satisfying the requirement of “infinite scalability.”
We are at an apparent impasse. Routing schemes specialized for static Internet-like graphs scale essentially indefinitely, but in adding the associated updates required to handle dynamic graphs, we can no longer guarantee the required scaling behavior. We must face the reality that scalable routing which uses topology update messages to dynamically react to topology changes is not possible in principle, which explains, in part, the lack of significant progress in search of truly scalable routing. We thus conclude that in order to find approaches that would lead us to required routing scalability, we need some radically new ideas that would allow us to construct convergence-free, “updateless” routing requiring no full view of network topologies. It is widely known and popularized that having no full view of social network topologies, humans can still efficiently route messages through them \[[48](#bib.bib48)\]. Whether we can devise routing protocols that would do the same for the Internet is an open question.
## 10 Acknowledgements
This work is supported in part by NSF CNS-0434996.
## References
- \[1\] D. Meyer, L. Zhang, and K. Fall (Eds.). Report from the IAB workshop on routing and addressing. IRTF, Internet Draft, 2007.
http://www.ietf.org/internet-drafts/draft-iab-raws-report-02.txt.
- \[2\] A. Doria, E. Davies, and F. Kastenholz (Edts.). Requirements for inter-domain routing. IRTF, Internet Draft, 2006.
- \[3\] C. Labovitz, A. Ahuja, A. Bose, and F. Jahanian. Delayed Internet routing convergence.
Transactions on Networking, 9(3), 2001.
- \[4\]
BGP Table Data, http://bgp.potaroo.net/.
- \[5\] G. Huston. Scaling inter-domain routing—a view forward.
The Internet Protocol Journal, 4(4), 2001.
- \[6\] G. Huston. Commentary on inter-domain routing in the Internet. IETF, RFC 3221, 2001.
- \[7\] A. Broido, E. Nemeth, and kc claffy. Internet expansion, refinement, and churn.
European Transactions on Telecommunications, 13(1):33–51, 2002.
- \[8\] H. Narayan, R. Govindan, and G. Varghese. The impact of address allocation and routing on the structure and implementation of routing tables.
In SIGCOMM, 2003.
- \[9\] Beyond BGP.
http://www.beyondbgp.net/.
- \[10\] P. Verkaik, A. Broido, kc claffy, R. Gao, Y. Hyun, and R. van der Pol. Beyond CIDR aggregation. Technical Report TR-2004-1, CAIDA, 2004.
- \[11\] Internet Engineering Task Force (IETF). Site multihoming in IPv6. Working Group.
- \[12\] F. Kastenholz. ISLAY: A new routing and addressing architecture. IRTF, Internet Draft, 2002.
- \[13\] I. Castineyra, N. Chiappa, and M. Steenstrup. The nimrod routing architecture. IETF, RFC 1992, 1996.
- \[14\] R. Gummadi, R. Govindan, N. Kothari, B. Karp, Y.-J. Kim, and S. Shenker. Reduced state routing in the Internet.
In HotNets, 2004.
- \[15\] L. Subramanian, M. Caesar, C. T. Ee, M. Handley, M. Mao, S. Shenker, and I. Stoica. HLP: A next generation inter-domain routing protocol.
In SIGCOMM, 2005.
- \[16\] M. Caesar, T. Condie, J. Kannan, K. Lakshminarayanan, I. Stoica, and S. Shenker. ROFL: Routing on flat labels.
In SIGCOMM, 2006.
- \[17\] C. Gavoille and S. Pérennès. Memory requirement for routing in distributed networks.
In PODC, 1996.
- \[18\] N. Spring, R. Mahajan, and T. Anderson. Quantifying the causes of path inflation.
In SIGCOMM, 2003.
- \[19\] J. Behrens and J. J. Garcia-Luna-Aceves. Distributed, scalable routing based on link-state vectors.
In SIGCOMM, 1994.
- \[20\] L. Kleinrock and F. Kamoun. Hierarchical routing for large networks: Performance evaluation and optimization.
Computer Networks, 1:155–174, 1977.
- \[21\] I. Abraham, C. Gavoille, A. Goldberg, and D. Malkhi. Routing in networks with low doubling dimension.
In ICDCS, 2006.
- \[22\] P. Indyk and J. Matoušek.
Handbook of Discrete and Computational Geometry, chapter 8, Low-Distortion Embeddings of Finite Metric Spaces.
Chapman & Hall/CRC, Boca Raton, 2004.
- \[23\] I. Abraham, Y. Bartal, T.-H. H. Chan, K. Dhamdhere, A. Gupta, J. Kleinberg, O. Neiman, and A. Slivkins. Metric embeddings with relaxed guarantees.
In FOCS, 2005.
- \[24\] P. Fraigniaud and C. Gavoille. Routing in trees.
In ICALP, 2001.
- \[25\] M. Thorup and U. Zwick. Compact routing schemes.
In SPAA, 2001.
- \[26\] J. Matoušek.
Lectures on Discrete Geometry, chapter 15, Embedding Finite Metric Spaces into Normed Spaces.
Springer, New York, 2002.
- \[27\] R. Krauthgamer and J. R. Lee. Algorithms on negatively curved spaces.
In FOCS, 2006.
- \[28\] R. Kleinberg. Geographic routing using hyperbolic space.
In INFOCOM, 2007.
- \[29\] P. Mahadevan, D. Krioukov, M. Fomenkov, B. Huffaker, X. Dimitropoulos, kc claffy, and A. Vahdat. The Internet AS-level topology: Three data sources and one definitive metric.
Computer Communication Review, 36(1), 2006.
- \[30\] F. Chung and L. Lu. The average distance in a random graph with given expected degrees.
Internet Mathematics, 1(1):91–114, 2003.
- \[31\] J. Leskovec, J. Kleinberg, and C. Faloutsos. Graphs over time: Densification laws, shrinking diameters and possible explanations.
In KDD, 2005.
- \[32\] D. Krioukov, K. Fall, and X. Yang. Compact routing on Internet-like graphs.
In INFOCOM, 2004.
- \[33\] C. Gavoille and M. Genegler. Space-efficiency for routing schemes of stretch factor three.
Journal of Parallel and Distributed Computing, 61(5):679–687, 2001.
- \[34\] M. Thorup and U. Zwick. Approximate distance oracles.
In STOC, 2001.
- \[35\] L. Cowen. Compact routing with minimum stretch.
Journal of Algorithms, 38(1):170–183, 2001.
- \[36\] I. Abraham, C. Gavoille, D. Malkhi, N. Nisan, and M. Thorup. Compact name-independent routing with minimum stretch.
In SPAA, 2004.
- \[37\] M. Arias, L. Cowen, K. A. Laing, R. Rajaraman, and O. Taka. Compact routing with name independence.
In SPAA, 2003.
- \[38\] A. Brady and L. Cowen. Compact routing on power-law graphs with additive stretch.
In ALENEX, 2006.
- \[39\] S. Carmi, R. Cohen, and D. Dolev. Searching complex networks efficiently with minimal information.
Europhysics Letters, 74:1102–1108, 2006.
- \[40\] T. Griffin and G. Wilfong. An analysis of BGP convergence properties.
In SIGCOMM, 1999.
- \[41\] T. Griffin and G. Wilfong. Analysis of the MED oscillation problem in BGP.
In ICNP, 2002.
- \[42\] Y. Afek, E. Gafni, and M. Ricklin. Upper and lower bounds for routing schemes in dynamic networks.
In FOCS, 1989.
- \[43\] A. Korman and D. Peleg. Dynamic routing schemes for general graphs.
In ICALP, 2006.
- \[44\] CAIDA. Macroscopic topology AS adjacencies.
http://www.caida.org/tools/measurement/skitter/as_adjacencies.xml.
- \[45\] The DIMES project.
http://www.netdimes.org/.
- \[46\] K. A. Laing and R. Rajaraman. A space lower bound for name-independent compact routing in trees.
In SPAA, 2005.
- \[47\] I. Abraham, C. Gavoille, and D. Malkhi. On space-stretch trade-offs: Lower bounds.
In SPAA, 2006.
- \[48\] S. Milgram. The small world problem.
Psychology Today, 1:61–67, 1967.
- \[49\] S. N. Dorogovtsev, J. F. F. Mendes, and A. N. Samukhin. Metric structure of random networks.
Nuclear Physics B, 653(3):307–422, 2003.
We demonstrate that the lower bound for communication costs in random scale-free graphs of size $`n`$ is $`\overset{\sim}{\Omega}\hspace{0pt}{(n)}`$. The average distance $`\overline{d}`$ in such graphs grows logarithmically with $`n`$, $`\overline{d} \sim {\log n}`$ \[[30](#bib.bib30), [49](#bib.bib49)\], while the width (standard deviation) of the distance distribution approaches zero in the large graph limit \[[49](#bib.bib49)\]. The latter property implies that the distance distribution $`d_{x}\hspace{0pt}{(r)}`$ for every node $`x`$, i.e., the number of nodes located at distance $`r`$ from $`x`$ divided by $`n`$, approaches a delta-function as $`n\rightarrow\infty`$:
<table>
<tbody>
<tr>
<td></td>
<td>dx(r)→n→∞{1if r=d¯∼logn,0otherwise.→𝑛absent→subscript𝑑𝑥𝑟cases1if r=d¯∼logn,0otherwise.d_{x}(r)\xrightarrow[n\to\infty]{}\begin{cases}1&\text{if \; $r=\bar{d}\sim\log n$,}\\ 0&\text{otherwise.}\end{cases}</td>
<td></td>
<td>(1)</td>
</tr>
</tbody>
</table>
Korman and Peleg \[[43](#bib.bib43)\] define the local density as
<table>
<tbody>
<tr>
<td></td>
<td>D=maxx,r|Bx(r)|2r,𝐷subscript𝑥𝑟subscript𝐵𝑥𝑟2𝑟D=\max_{x,r}\frac{|B_{x}(r)|}{2r},</td>
<td></td>
<td>(2)</td>
</tr>
</tbody>
</table>
where $`B_{x}\hspace{0pt}{(r)}`$ is the ball of radius $`r`$ around node $`x`$, i.e., the set of nodes located within distance at most $`r`$ from $`x`$, and $`|{B_{x}\hspace{0pt}{(r)}}|`$ is the size of this set. The maximum is taken over all nodes $`x`$ in the graph and all radii $`r`$. Observing that $`{|{B_{x}\hspace{0pt}{(r)}}|} = {n\hspace{0pt}{\int_{0}^{r}{d_{x}\hspace{0pt}{(\rho)}\hspace{0pt}{d\rho}}}}`$ and using eq. ([1](#A0.E1 "In On Compact Routing for the Internet")), we immediately obtain that $`|{B_{x}\hspace{0pt}{(r)}}|`$ in large scale-free graphs is given by the Heaviside function:
<table>
<tbody>
<tr>
<td></td>
<td>|Bx(r)|→n→∞{nif r⩾d¯∼logn,0otherwise.→𝑛absent→subscript𝐵𝑥𝑟cases𝑛if r⩾d¯∼logn,0otherwise.|B_{x}(r)|\xrightarrow[n\to\infty]{}\begin{cases}n&\text{if \; $r\geqslant\bar{d}\sim\log n$,}\\ 0&\text{otherwise.}\end{cases}</td>
<td></td>
<td>(3)</td>
</tr>
</tbody>
</table>
Substitution of this expression into eq. ([2](#A0.E2 "In On Compact Routing for the Internet")) yields
<table>
<tbody>
<tr>
<td></td>
<td>D→n→∞nlogn.→𝑛absent→𝐷𝑛𝑛D\xrightarrow[n\to\infty]{}\frac{n}{\log n}.</td>
<td></td>
<td>(4)</td>
</tr>
</tbody>
</table>
The last expression and the demonstration in \[[43](#bib.bib43)\] that the communication cost lower bound for graphs with local density $`D`$ is $`\Omega\hspace{0pt}{(D)}`$ complete the proof. |
0708.0975 | cs.NI | cyber | ###### Abstract
We study efficient broadcasting for wireless sensor networks, with network coding. We address this issue for homogeneous sensor networks in the plane. Our results are based on a simple principle (IREN/IRON), which sets the same rate on most of the nodes (wireless links) of the network. With this rate selection, we give a value of the maximum achievable broadcast rate of the source: our central result is a proof of the value of the *min-cut* for such networks, viewed as *hypergraphs*. Our metric for efficiency is the number of transmissions necessary to transmit one packet from the source to every destination: we show that IREN/IRON achieves near optimality for large networks; that is, asymptotically, nearly every transmission brings new information from the source to the receiver. As a consequence, network coding asymptotically outperforms any scheme that does not use network coding.
\normallatexbib
\articletitle
Near Optimal Broadcast
with Network Coding
in Large Sensor Networks
## Introduction
Seminal work in \[[1](#bib.bib1)\] has introduced the idea of network coding, whereby intermediate nodes are mixing information from different flows (different bits or different packets).
One logical domain of application is *wireless sensor networks*. Indeed, for wireless networks, a generalization of the results in \[[1](#bib.bib1)\] exists: when the capacity of the links are known and fixed, the maximal broadcast rate of the source can be computed, as shown in \[[3](#bib.bib3)\]. Essentially, for one source, it is the min-cut of the network from the source to the destinations, as for wired networks \[[1](#bib.bib1)\], but considering *hypergraphs* rather than graphs. This is true whether the rate and the capacity are expressed in bits per second or packets per second \[[4](#bib.bib4)\].
However, in wireless sensor networks, a primary constraint is not necessarily the capacity of the wireless links: because of the limited battery of each node, the limiting factor is the cost of wireless transmissions. Hence a different focus is *energy-efficiency*, rather than the maximum achievable broadcast rate:
> $`\bullet`$ Given one source, minimize the total number of transmissions used to achieve the broadcast to destination nodes.
The problem is no longer related to the capacity, because the same transmissions can be streched in time, with an identical cost. However, one can still imagine using network coding, where each node repeats combinations of packets with an average interval between transmissions: this defines the rate of the node, and the rate is an unknown.
The problem of energy-efficiency is to compute a set of transmission rates for each node, with minimal cost. With network coding, the problem turns out to be solvable in polynomial time: for the stated problem, \[[5](#bib.bib5), [6](#bib.bib6)\] describe methods to find the optimal transmission rate of each node with a linear program. However, this does not necessarily provide direct insight about the optimal rates and their associated optimal cost: those are obtained by solving the linear program on instances of networks.
For large-scale sensor networks, one assumption could be that the nodes are distributed in a homogeneous way, and a question would be: “Is there a simple near-optimal rate selection ?” Considering the results of min-cut estimates for random graphs \[[7](#bib.bib7), [8](#bib.bib8), [9](#bib.bib9)\], one intuition is that most nodes have similar neighborhood; hence the performance, when setting an identical rate for each node, deserves to be explored. This is the starting point of this paper, and we will focus on homogeneous networks, which can be modeled as unit disk graphs:
1. 1\.
We introduce a simple rate principle where most nodes have the same transmission rate: IREN/IRON principle (Increased Rate for Exceptional Nodes, Identical Rate for Other Nodes).
2. 2\.
We give a proof for the min-cut for some lattice graphs (modeled as hypergraphs). It is also an intermediate step for the following:
3. 3\.
We deduce an estimate of the min-cut for unit disk hypergraphs.
4. 4\.
We show that this simple rate selection achieves “near optimal performance” in some classes of homogeneous networks, based on min-cut computation — and may outperforms any scheme that is not using network coding.
The rest of this paper is organized as follows: section [1](#S1 "1 Network Model and Related Work") details the network model and related work; section [2](#S2 "2 Main Results") describes the main results; section [3](#S3 "3 Proofs of the Min-Cut") gives proofs of the min-cut; and section [4](#S4 "4 Conclusion") concludes.
## 1 Network Model and Related Work
In this article, we study the problem of broadcasting from one source to all nodes. We will assume an ideal wireless model, wireless transmissions without loss, collisions or interferences and that each node of the network is operating well below its maximum transmission capacity.
Our focus is on large-scale wireless sensor networks. Such networks have been modeled as *unit disk graphs* \[[12](#bib.bib12)\] of the plane, where two nodes are neighbors whenever their distance is lower than a fixed radio range; see Fig. [1(a)](#S1.F1.sf1 "In Figure 1 ‣ 1 Network Model and Related Work") for the principle of unit disk graphs.
[Refer to caption](/html/0708.0975/assets/x1.png)
\(a\)
Unit disk graph
[Refer to caption](/html/0708.0975/assets/x2.png)
\(b\)
Lattice
[Refer to caption](/html/0708.0975/assets/x3.png)
\(c\)
Unit disk range $`R`$ for a lattice
Figure 1: Network Models
Precisely, the sensor networks considered will be:
- •
Random unit disk graphs with nodes uniformly distributed (Fig. [1(a)](#S1.F1.sf1 "In Figure 1 ‣ 1 Network Model and Related Work"))
- •
Unit disk graphs with nodes organized on a lattice (Fig. [1(b)](#S1.F1.sf2 "In Figure 1 ‣ 1 Network Model and Related Work")).
An important assumption is that the *wireless broadcast advantage* is used: each transmission is overheard by several nodes. As a result the graph is in reality a *(unit disk) hypergraph*.
### 1.1 Related Work
In general, specifying the network coding protocol is reduced to specifying the transmission rates for each node \[[10](#bib.bib10)\]. Once the optimal rates are computed, the performance can be asymptotically achieved with *distributed random linear coding*, for instance \[[11](#bib.bib11), [4](#bib.bib4)\]. This article is in the spirit of \[[13](#bib.bib13)\], which starts with exhibiting an energy-efficient algorithm for simple networks. The central element for computing the performance is the estimation of the min-cut of the network and we are inspired by the existing techniques and results surrouding the expected value of the min-cut on some classes of networks: for instance, \[[7](#bib.bib7)\] explored the capacity of networks where a source is two hops from the destination, through a network of relay nodes; \[[8](#bib.bib8)\] studied some classes of random geometric graphs. Recently, \[[9](#bib.bib9)\] gave bounds of the min-cut of dual radio networks.
### 1.2 Network Coding
In the network coding literature, several results are for multicast, and, in this section, they are quoted as such. They apply to the topic of this article, broadcast, since broadcast is a special case of multicast.
A central result for network coding in wireless networks gives the maximum multicast rate for a source. The capacity is given by the min-cut from the source to each individual destination of the network, viewed as a hypergraph \[[3](#bib.bib3), [6](#bib.bib6)\].A precise description includes:
$`\bullet`$ Nodes: $`\mathcal{V} = {\{ v_{i},i = 1,\ldots N\}}`$, set of nodes of the hypergraph
$`\bullet`$ Hyperarc: $`h_{v} = {(v,H_{v})}`$, where $`H_{v} \subset V`$ is the subset of nodes that are reached by one transmission of node $`v`$(neighbors).
$`\bullet`$ Rate: Each node $`v`$ emits on the hyperarc $`(v,H_{v})`$ with rate $`C_{v}`$.
Let us consider the source $`s`$, and one of the multicast destinations $`t \in \mathcal{V}`$. The definition of an *$`s`$-$`t`$ cut* is: a partition of the set of nodes $`V`$ in two sets $`S`$, $`T`$ such as $`s \in S`$ and $`t \in T`$. Let $`Q\hspace{0pt}{(s,t)}`$ be the set of such *$`s`$-$`t`$ cuts*: $`{(S,T)} \in {Q\hspace{0pt}{(s,t)}}`$.
We denote $`\Delta\hspace{0pt}S`$, the set of nodes of $`S`$ that are neighbors of at least one node of $`T`$; the *capacity of the cut* $`C\hspace{0pt}{(S)}`$ is defined as the maximum rate between the nodes in $`S`$ and the nodes in $`T`$:
<table>
<tbody>
<tr>
<td></td>
<td>ΔS≜{v∈S:Hv∩T≠∅}andC(S)≜∑v∈ΔSCvformulae-sequence≜Δ𝑆conditional-set𝑣𝑆subscript𝐻𝑣𝑇and≜𝐶𝑆subscript𝑣Δ𝑆subscript𝐶𝑣\Delta{}S\triangleq\{v\in S:H_{v}\cap T\neq\emptyset\}\leavevmode\nobreak\ \leavevmode\nobreak\ \leavevmode\nobreak\ \mathrm{and}\leavevmode\nobreak\ \leavevmode\nobreak\ \leavevmode\nobreak\ C(S)\triangleq\sum_{v\in\Delta{}S}C_{v}</td>
<td></td>
<td>(1)</td>
</tr>
</tbody>
</table>
The *min-cut* between $`s`$ and $`t`$ is the cut of $`Q\hspace{0pt}{(s,t)}`$ with the minimum capacity. Let us denote $`C_{\min}\hspace{0pt}{(s,t)}`$ as its capacity. From \[[3](#bib.bib3), [6](#bib.bib6)\], the maximum multicast capacity is given by the minimum of capacity of the min-cut of every destination, $`C_{\min}\hspace{0pt}{(s)}`$, with:
<table>
<tbody>
<tr>
<td></td>
<td>Cmin(s,t)≜min(S,T)∈Q(s,t)C(S)andCmin(s)≜mint∈𝒱∖{s}Cmin(s,t)formulae-sequence≜subscript𝐶min𝑠𝑡subscript𝑆𝑇𝑄𝑠𝑡𝐶𝑆and≜subscript𝐶min𝑠subscript𝑡𝒱𝑠subscript𝐶min𝑠𝑡C_{\mathrm{min}}(s,t)\triangleq\min_{(S,T)\in Q(s,t)}C(S)\leavevmode\nobreak\ \leavevmode\nobreak\ \leavevmode\nobreak\ \mathrm{and}\leavevmode\nobreak\ \leavevmode\nobreak\ \leavevmode\nobreak\ C_{\mathrm{min}}(s)\triangleq\min_{t\in\mathcal{V}\setminus\{s\}}C_{\mathrm{min}}(s,t)</td>
<td></td>
<td>(2)</td>
</tr>
</tbody>
</table>
## 2 Main Results
### 2.1 Overview
As described in the introduction, our approach is to choose an intuitive transmission rate for each node: essentially, the same rate for most nodes, as described in section [2.3](#S2.SS3 "2.3 Rate Selection with IREN/IRON ‣ 2 Main Results"). Then, we determine the maximum broadcast rate that can be achieved to transmit from the source to every node in the network as the min-cut of the hypergraph, for both random and lattice graphs in section [2.4](#S2.SS4 "2.4 Performance: Min-Cut (Achievable Broadcast Rate) ‣ 2 Main Results"). And finally, from the expression of the cost in section [2.5](#S2.SS5 "2.5 Performance: Transmission Cost per Broadcast ‣ 2 Main Results"), we deduce asymptotic optimality (section [2.6](#S2.SS6 "2.6 Near Optimal Performance for Large Networks ‣ 2 Main Results")).
### 2.2 Further Definitions
Consider a network inside a square area $`G`$ of edge length $`L`$, such as the one on Fig. [1(a)](#S1.F1.sf1 "In Figure 1 ‣ 1 Network Model and Related Work").
- •
The radio range of the network is $`\rho`$.
- •
For a lattice, we denote $`R`$ the set of neighbors of the origin node $`(0,0)`$, as represented on Fig. [1(c)](#S1.F1.sf3 "In Figure 1 ‣ 1 Network Model and Related Work"): $`R \triangleq {\{{{(x,y)} \in {\mathbb{Z}}^{2}}:{{x^{2} + y^{2}} \leq \rho^{2}}\}}`$
- •
Let $`M`$ be the “expected” number of neighbors of one node. For a lattice, it is $`M = {{|R|} - 1}`$. For a random disk unit graph with $`N`$ nodes, $`M`$ is related to the density $`\mu = \frac{N}{L^{2}}`$ and range as follows: $`M = {\pi\hspace{0pt}\rho^{2}\hspace{0pt}\mu} = {\pi\hspace{0pt}\rho^{2}\hspace{0pt}\frac{N}{L^{2}}}`$.
We define the *border area* as the area of fixed width $`W > \rho`$ near the edge of that square, and *border nodes* as the nodes in that area. Hence, the area $`L \times L`$ of $`G`$ is partitioned into:
- •
$`\Delta\hspace{0pt}G`$, the border, with area $`A_{\Delta\hspace{0pt}G} = {4\hspace{0pt}W\hspace{0pt}{({L - W})}}`$
- •
$`G_{i}`$, the “interior” $`G_{i} \triangleq {G \smallsetminus {\Delta\hspace{0pt}G}}`$, with area $`A_{G_{i}} = {({L - {2\hspace{0pt}W}})}^{2}`$
### 2.3 Rate Selection with IREN/IRON
The principle IREN/IRON sets the following transmission rates:
- •
IREN (Increased Rate for Exceptional Nodes): the rate of transmission is set to $`M`$, for the source node and all the border nodes (the “exceptional” nodes).
- •
IRON (Identical Rate for Other Nodes): every other node transmits with rate $`1`$.
### 2.4 Performance: Min-Cut (Achievable Broadcast Rate)
The essence of our main result is the following property proved in section [3.1](#S3.SS1 "3.1 Proof for Lattice Graphs ‣ 3 Proofs of the Min-Cut"), Th. [3.4](#S3.Thmtheorem4 "Theorem 3.4. ‣ 3.1.3 Value of the Min-cut 𝐶ₘᵢₙ(𝑠) ‣ 3.1 Proof for Lattice Graphs ‣ 3 Proofs of the Min-Cut"):
######
Property 1
With the rate selection IREN/IRON, the min-cut of a lattice graph is equal to $`C_{\min} = M`$ (with $`M = {{|R|} - 1}`$).
For random unit disk graphs, by mapping the points to an imaginary lattice graph (*embedded lattice*) as an intermediary step, we are able to find bounds of the capacity of random unit disk graphs. This turns out to be much in the spirit of \[[9](#bib.bib9)\]. This is used to deduce an asymptotic result for unit disk graphs, proven in section [3.2](#S3.SS2 "3.2 Proof of the Value of Min-Cut for Unit Disk Graphs ‣ 3 Proofs of the Min-Cut"), Th. [3.11](#S3.Thmtheorem11 "Theorem 3.11. ‣ 3.2.5 Asymptotic Values of the Min-Cut of Unit-Disk Graphs ‣ 3.2 Proof of the Value of Min-Cut for Unit Disk Graphs ‣ 3 Proofs of the Min-Cut"):
######
Property 2
Assume a fixed range. For a sequence of random unit disk graphs $`(\mathcal{V}_{i})`$, with sources $`s_{i}`$, with size $`L\rightarrow\infty`$ and with a density $`M\rightarrow\infty`$ such as $`M = L^{\theta}`$, for any fixed $`\theta > 0`$, we have the following convergence in probability: $`\frac{C_{\min}\hspace{0pt}{(s)}}{M}\overset{p}{\rightarrow}\, 1`$.
### 2.5 Performance: Transmission Cost per Broadcast
Recall that the metric for cost is the number of (packet) transmissions per a (packet) broadcast from the source to the entire network. Let us denote $`E_{cost}`$ as this “transmissions per broadcast.”
This cost of broadcasting with IREN/IRON rate selection can be equivalently computed from the rates as the ratio of the number of transmissions per unit time to the number of packets broadcast into the network per unit time. Then $`E_{cost}`$ is deduced from the min-cut $`C_{\min}`$, the areas $`A_{\Delta\hspace{0pt}G},A_{G_{i}}`$, the associated node rates and the node density $`\mu`$. For fixed $`W`$, $`{M,L}\rightarrow\infty`$: $`E_{cost} = {\frac{1}{C_{\min}}\hspace{0pt}\mu\hspace{0pt}L^{2}\hspace{0pt}\left( {1 + {O\hspace{0pt}{(\frac{1}{L})}} + {\frac{4\hspace{0pt}M\hspace{0pt}W}{L}\hspace{0pt}{({1 + {O\hspace{0pt}{(\frac{1}{L})}}})}}} \right)}`$.
For random unit disk graphs, $`E_{cost}`$ is an expected value, and $`\mu = \frac{N}{L^{2}}`$. For a lattice, $`\mu = 1`$.
### 2.6 Near Optimal Performance for Large Networks
Sections [2.4](#S2.SS4 "2.4 Performance: Min-Cut (Achievable Broadcast Rate) ‣ 2 Main Results") and [2.5](#S2.SS5 "2.5 Performance: Transmission Cost per Broadcast ‣ 2 Main Results") gave the performance and cost with the
IREN/IRON principle. The optimal cost is not easily computed, and in this section an indirect route is chosen, by using a bound.
Assume that every node has at most $`M_{\max}`$ neighbors: one single transmission can provide information to $`M_{\max}`$ nodes at most. Hence, in order to broadcast one packet to all $`N`$ nodes, at least $`E_{bound} = \frac{N}{M_{\max}}`$ transmissions are necessary.
W.r.t. this bound, let the relative cost be: $`E_{{rel} - {cost}} = \frac{E_{cost}}{E_{bound}} \geq 1`$.
We will prove that $`E_{{rel} - {cost}}\rightarrow 1`$ for the following networks:
#### 2.6.1 Lattice Graphs
For lattices, $`W`$ and the neighborhood $`R`$ are kept fixed (hence also $`M = {{|R|} - 1}`$) and only the size $`L`$ of the network increases to infinity. The number of nodes is $`N = L^{2}`$. The maximum number of neighbors $`M_{\max}`$ is exactly $`M_{\max} = M`$.
From section [2.5](#S2.SS5 "2.5 Performance: Transmission Cost per Broadcast ‣ 2 Main Results") and from Property [1](#Thmproperty1 "Property 1 ‣ 2.4 Performance: Min-Cut (Achievable Broadcast Rate) ‣ 2 Main Results"), we have:
$`E_{{rel} - {cost}} = {E_{cost}\hspace{0pt}\frac{M_{\max}}{N}} = \left( {1 + {O\hspace{0pt}{(\frac{1}{L})}} + {\frac{4\hspace{0pt}M\hspace{0pt}W}{L}\hspace{0pt}{({1 + {O\hspace{0pt}{(\frac{1}{L})}}})}}} \right) = {1 + {O\hspace{0pt}{(\frac{1}{L})}}}`$.
#### 2.6.2 Random Unit Disk Graphs
For random unit disk graphs, first notice that an increase of the density $`M`$ does not improve the relative cost $`E_{{rel} - {cost}}`$. Now consider a sequence of random graphs, as in Property [2](#Thmproperty2 "Property 2 ‣ 2.4 Performance: Min-Cut (Achievable Broadcast Rate) ‣ 2 Main Results"), with fixed $`\rho`$, fixed $`W`$, and size $`L\rightarrow\infty`$ and with a density $`M\rightarrow\infty`$ such as $`M = L^{\theta}`$, for some arbitrary fixed $`\theta > 0`$, with the additional constraint that $`\theta < 1`$. We have:
$`E_{{rel} - {cost}} = {E_{cost}\hspace{0pt}\frac{M_{\max}}{N}} = {\frac{M}{C_{\min}}\hspace{0pt}\frac{M_{\max}}{M}\hspace{0pt}\frac{\mu\hspace{0pt}L^{2}}{N}\hspace{0pt}\left( {1 + {O\hspace{0pt}{(\frac{1}{L})}} + {\frac{4\hspace{0pt}M\hspace{0pt}W}{L}\hspace{0pt}{({1 + {O\hspace{0pt}{(\frac{1}{L})}}})}}} \right)}`$.
Each of part of the product converges toward $`1`$, either surely, or in probability: using Property [2](#Thmproperty2 "Property 2 ‣ 2.4 Performance: Min-Cut (Achievable Broadcast Rate) ‣ 2 Main Results"), we have the convergence of $`\frac{C_{\min}}{M}\overset{p}{\rightarrow}\, 1`$, when $`L\rightarrow\infty`$ and similarly with Th. [3.11](#S3.Thmtheorem11 "Theorem 3.11. ‣ 3.2.5 Asymptotic Values of the Min-Cut of Unit-Disk Graphs ‣ 3.2 Proof of the Value of Min-Cut for Unit Disk Graphs ‣ 3 Proofs of the Min-Cut") we have $`\frac{M_{\max}}{M}\overset{p}{\rightarrow}\, 1`$. By definition $`N = {\mu\hspace{0pt}L^{2}}`$. Finally, $`M = L^{\theta}`$ for $`\theta < 1`$ implies that $`\frac{4\hspace{0pt}M\hspace{0pt}W}{L}\rightarrow 0`$.
As a result we have: $`E_{{rel} - {cost}}\overset{p}{\rightarrow}\, 1`$ in probability, when $`L\rightarrow\infty`$
#### 2.6.3 Random Unit Disk Graphs without Network Coding
In order to compare the results that are obtained when network coding is not used, one can reuse the argument of \[[13](#bib.bib13)\]. Consider the broadcasting of one packet. Consider one node of the network that has repeated the packets. It must have received the transmission from another connected neighbor. In a unit disk graph, these two connected neighbors share a neighborhood area at least equal to $`{({\frac{2\hspace{0pt}\pi}{3} - \frac{\sqrt{3}}{2}})}\hspace{0pt}\rho^{2}`$, and every node lying within that area will receive duplicate of the packets. Considering this inefficiency, for dense unit disk graphs one can deduce the following bound: $`E_{{rel} - {cost}}^{({{no} - {coding}})} \geq \frac{6\hspace{0pt}\pi}{{2\hspace{0pt}\pi} + {3\hspace{0pt}\sqrt{3}}}`$ . Notice that $`\frac{6\hspace{0pt}\pi}{{2\hspace{0pt}\pi} + {3\hspace{0pt}\sqrt{3}}} \approx {1.6420\hspace{0pt}\ldots} > 1`$.
#### 2.6.4 Near Optimality
The asymptotic optimality is a consequence of the convergence of the cost bound $`E_{{rel} - {cost}}`$ toward $`1`$. This indirect proof is in fact a stronger statement than optimality of the rate selection in terms of energy-efficiency: it exhibits the fact that asymptotically (nearly) all the transmissions will be *innovative* for the receivers. Note that it is not the case in general for a given instance of a hypergraph. It evidences the following remarkable fact for the large homogeneous networks considered: network coding may be achieving not only optimal efficiency, but also, asymptotically, perfect efficiency — achieving the information-theoretic bound for each transmission.
Notice that traditional broadcast methods without network coding (such as the ones based on connected dominating sets) cannot achieve this efficiency, since their lower bound is $`1.642`$.
## 3 Proofs of the Min-Cut
### 3.1 Proof for Lattice Graphs
#### 3.1.1 Preliminaries
Let $`\Gamma`$ be full, *integer lattice* in $`n`$-dimensional space; it is the set $`{\mathbb{Z}}^{n}`$, where the lattice points are $`n`$-tuples of integers.
For lattice graphs, only points on the full lattice are relevant; therefore in this section, the notations $`\mathcal{L},\mathcal{L}_{i},{\Delta\hspace{0pt}\mathcal{L}}`$ will be used, for the parts of the full lattice $`\Gamma`$ that are in $`G,G_{i},{\Delta\hspace{0pt}G}`$ respectively.
The proof is based on the use of the Minkowski addition, and a specific property of discrete geometry ([3](#S3.E3 "In Property 3 ‣ 3.1.1 Preliminaries ‣ 3.1 Proof for Lattice Graphs ‣ 3 Proofs of the Min-Cut")) below. The Minkowski addition is a classical way to express the neighborhood of one area (for instance, see \[[14](#bib.bib14)\] and the figure 3(a), and figure 4 of that reference).
Given two sets $`A`$ and $`B`$ of $`{\mathbb{R}}^{n}`$, the Minkowski sum of the two sets $`A \oplus B`$ is defined as the set of all vector sums generated by all pairs of points in $`A`$ and $`B`$, respectively: $`{A \oplus B} \triangleq {\{{a + b}:{{a \in A},{b \in B}}\}}`$
Then the set of neighbors $`\mathcal{N}\hspace{0pt}{(t)}`$ of one node $`t`$, with $`t`$ itself, is:
$`{{\mathcal{N}\hspace{0pt}{(t)}} \cup {\{ t\}}} = {{\{ t\}} \oplus R}`$
This extends to the neighborhood of a set of points. For Minkowski sums on the lattice $`\Gamma`$, there exist variants of the *Brunn-Minkowski inequality*, including the following one \[[15](#bib.bib15)\]:
######
Property 3
For two subsets $`A,B`$ of the integer lattice $`{\mathbb{Z}}^{n}`$,
<table>
<tbody>
<tr>
<td></td>
<td>|A⊕B|≥|A|+|B|−1direct-sum𝐴𝐵𝐴𝐵1|A\oplus B|\geq|A|+|B|-1</td>
<td></td>
<td>(3)</td>
</tr>
</tbody>
</table>
where $`|X|`$ represents the number of elements of a subset $`X`$ of $`{\mathbb{Z}}^{n}`$
#### 3.1.2 Bound on the capacity of one cut $`C\hspace{0pt}{(S)}`$
Consider a lattice $`\mathcal{L}`$ and a source $`s`$. Let $`C\hspace{0pt}{(S)}`$ be the capacity of an $`s`$-$`t`$ cut $`{S,T} \in {Q\hspace{0pt}{(s,t)}}`$.
######
Lemma 3.1
.
$`{C\hspace{0pt}{(S)}} \geq {|{\Delta\hspace{0pt}S}|}`$ (with $`\Delta\hspace{0pt}S`$ defined in ([1](#S1.E1 "In 1.2 Network Coding ‣ 1 Network Model and Related Work")))
######
Proof 3.2
.
$`C_{v} \geq 1`$ with IREN/IRON and with ([1](#S1.E1 "In 1.2 Network Coding ‣ 1 Network Model and Related Work")), $`{C\hspace{0pt}{(S)}} = {\sum_{v \in {\Delta\hspace{0pt}S}}C_{v}}`$
######
Theorem 3.3
.
The capacity of one cut $`C\hspace{0pt}{(S)}`$ is such that: $`{C\hspace{0pt}{(S)}} \geq M`$
*Proof*: There are three possible cases, either the set $`T`$ has no common nodes with the border $`\Delta\hspace{0pt}\mathcal{L}`$, or $`T`$ includes all nodes of $`\Delta\hspace{0pt}\mathcal{L}`$, or finally $`T`$ includes only part of nodes in the border area.
First case, $`{T \cap {\Delta\hspace{0pt}\mathcal{L}}} = \varnothing`$:
We know that $`{T \oplus R} \subset \mathcal{L}`$, hence we can effectively write the neighbors of nodes in $`T`$ as a Minkowski addition (without getting points in $`\Gamma`$ but out of $`\mathcal{L}`$): $`{\Delta\hspace{0pt}T} \triangleq {{({T \oplus R})} \smallsetminus T}`$
It follows that: $`{|{\Delta\hspace{0pt}T}|} \geq {{|{T \oplus R}|} - {|T|}}`$
Now the inequality ([3](#S3.E3 "In Property 3 ‣ 3.1.1 Preliminaries ‣ 3.1 Proof for Lattice Graphs ‣ 3 Proofs of the Min-Cut")) can be used: $`{|{T \oplus R}|} \geq {{{|T|} + {|R|}} - 1}`$
Hence we get: $`{|{\Delta\hspace{0pt}T}|} \geq {{{|T|} + {|R|}} - 1 - {|T|}}`$, and therefore:
<table>
<tbody>
<tr>
<td></td>
<td>|ΔT|≥|R|−1Δ𝑇𝑅1|\Delta T|\geq|R|-1</td>
<td></td>
<td>(4)</td>
</tr>
</tbody>
</table>
Recall that $`S`$ and $`T`$ form a partition of $`\mathcal{L}`$ ; and since $`\Delta\hspace{0pt}T`$ is a subset of $`\mathcal{L}`$, by definition without any point of of $`T`$, we have $`{\Delta\hspace{0pt}T} \subset S`$. Hence actually $`{\Delta\hspace{0pt}T} \subset {\Delta\hspace{0pt}S}`$ (with the definition of $`\Delta\hspace{0pt}S`$ in ([1](#S1.E1 "In 1.2 Network Coding ‣ 1 Network Model and Related Work"))). We can combine this fact with lemma [3.1](#S3.Thmtheorem1 "Lemma 3.1. ‣ 3.1.2 Bound on the capacity of one cut 𝐶(𝑆) ‣ 3.1 Proof for Lattice Graphs ‣ 3 Proofs of the Min-Cut") and ([4](#S3.E4 "In 3.1.2 Bound on the capacity of one cut 𝐶(𝑆) ‣ 3.1 Proof for Lattice Graphs ‣ 3 Proofs of the Min-Cut")), to get:
$`{|{C\hspace{0pt}{(S)}}|} \geq {{|R|} - 1}`$ and the Th. [3.3](#S3.Thmtheorem3 "Theorem 3.3. ‣ 3.1.2 Bound on the capacity of one cut 𝐶(𝑆) ‣ 3.1 Proof for Lattice Graphs ‣ 3 Proofs of the Min-Cut") is proved for the first case.
The second case is similar, but considering the source, while the third case uses the fact that a path can be found in the border between any two border nodes \[[17](#bib.bib17)\]. $`\square`$
#### 3.1.3 Value of the Min-cut $`C_{\min}\hspace{0pt}{(s)}`$
The results of the previous section immediately result in a property on the capacity of every $`s`$-$`t`$ min-cut:
######
Theorem 3.4
.
For any $`t \in \mathcal{L}`$ different from the source $`s`$:
$`{C_{\min}\hspace{0pt}{(s,t)}} = M`$ ; and as a result: $`{C_{\min}\hspace{0pt}{(s)}} = M`$
######
Proof 3.5
.
From Th. [3.3](#S3.Thmtheorem3 "Theorem 3.3. ‣ 3.1.2 Bound on the capacity of one cut 𝐶(𝑆) ‣ 3.1 Proof for Lattice Graphs ‣ 3 Proofs of the Min-Cut"), we have the capacity of every $`s - t`$ cut $`S/T`$ verifies: $`{C\hspace{0pt}{(S)}} \geq M`$. Hence $`{C_{\min}\hspace{0pt}{(s,t)}} \geq M`$
Conversely let us consider a specific cut, $`S_{s} = {\{ s\}}`$ and $`T_{s} = {\mathcal{L} \smallsetminus {\{ s\}}}`$. Obviously $`s`$ has at least one neighbor hence $`{\Delta\hspace{0pt}S} = {\{ s\}}`$. The capacity of the cut is $`{C\hspace{0pt}{(S_{s})}} = {\sum_{v \in {\Delta\hspace{0pt}S}}C_{v}} = C_{s} = M`$ and thus $`{C_{\min}\hspace{0pt}{(s,t)}} \leq M`$, and the theorem follows. $`\square`$
### 3.2 Proof of the Value of Min-Cut for Unit Disk Graphs
In this section, we will prove a probabilistic result on the min-cut, in the case of random unit disk graphs, using an virtual “embedded” lattice. The unit graph will be denoted $`\mathcal{V}`$, whereas for the embedded lattice the notation of section [3](#S3 "3 Proofs of the Min-Cut") is used: $`\mathcal{L}`$ (along with $`\Delta\hspace{0pt}\mathcal{L}`$ and $`\mathcal{L}_{i}`$).
#### 3.2.1 Embedded Lattice
Given the square area $`L \times L`$, we start with fitting a *rescaled* lattice inside it, with a scaling factor $`r`$. Precisely, it is the intersection of square $`G`$ and the set $`\{{({r\hspace{0pt}x},{r\hspace{0pt}y})}:{{(x,y)} \in {\mathbb{Z}}^{2}}\}`$.
We will map the points of $`G`$ to the closest point of the rescaled lattice $`\mathcal{L}`$: let us denote $`\lambda\hspace{0pt}{(x)}`$, the application that transforms a point $`u`$ of the Euclidian space $`{\mathbb{R}}^{2}`$ to its closest point of $`\mathcal{L}`$. Formally, for $`u = {(x,y)} \in {\mathbb{Z}}^{2}`$, $`{\lambda\hspace{0pt}{(x)}} \triangleq {({r\hspace{0pt}{\lfloor{\frac{x}{r} + \frac{1}{2}}\rfloor}},{r\hspace{0pt}{\lfloor{\frac{y}{r} + \frac{1}{2}}\rfloor}})}`$
For $`u \in \mathcal{L}`$, $`\lambda^{- 1}\hspace{0pt}{(u)}`$ is the set of nodes of $`\mathcal{V}`$ that are mapped to $`u`$. The area of $`{\mathbb{R}}^{2}`$ that is mapped to a same point of the lattice, is a square $`r \times r`$ around that point.
Let $`u`$ be a point of the lattice $`\mathcal{L}`$, and let denote the $`m\hspace{0pt}{(u)}`$ the number of points of $`\mathcal{V}`$ that are mapped to $`u`$ with $`g`$ (they are in the square around $`u`$ ; and $`{m\hspace{0pt}{(u)}} \triangleq {|{\lambda^{- 1}\hspace{0pt}{(u)}}|}`$). $`m\hspace{0pt}{(u)}`$ is a random variable.
Let us denote: $`m_{\min} \triangleq {{\min_{u \in \mathcal{L}}m}\hspace{0pt}{(u)}\hspace{0pt}{and}\hspace{0pt}m_{\max}} \triangleq {{\max_{u \in \mathcal{L}}m}\hspace{0pt}{(u)}}`$
#### 3.2.2 Neighborhood of the Embedded Lattice
We start by defining the neighborhood $`R`$ for the embedded lattice. We choose $`R\hspace{0pt}{(r)}`$ to be the points of the lattice inside a disk of radius $`\rho - {2\hspace{0pt}r}`$.
######
Lemma 3.6
.
Let us consider two nodes of $`u`$, $`v`$ of $`\mathcal{V}`$ that are mapped on the lattice $`\mathcal{L}`$ to $`u_{\mathcal{L}}`$ and $`v_{\mathcal{L}}`$ respectively:
$`\bullet`$ if $`u_{\mathcal{L}}`$ and $`v_{\mathcal{L}}`$ are neighbors on the lattice, them $`u`$ and $`v`$ are neighbors on the graph $`\mathcal{V}`$
This results from triangle inequalities on the distances.
#### 3.2.3 Relationship between the Capacities of the Cuts of the Embedded Lattice and the Random Disk Unit Graph
Let us consider one source $`s \in \mathcal{V}`$, one destination $`t \in \mathcal{V}`$ and the capacity of any $`S/T`$ cut. Every node of $`S`$ and $`T`$ is then mapped to the nearest point of the embedded lattice. For the source, we denote: $`s_{\mathcal{L}} = {\lambda\hspace{0pt}{(s)}}`$.
An *induced cut* of the embedded lattice is constructed as follows:
- •
The border area width $`W_{\mathcal{L}}`$ is selected so as to be the greatest integer multiple of $`r`$ which is smaller than $`W`$ ; and $`r < {W - \rho}`$
- •
For any point of the lattice $`v_{\mathcal{L}} \in \mathcal{L}`$, the rate $`C_{v_{\mathcal{L}}}^{(\mathcal{L})}`$ is set according to IREN/IRON on the lattice: $`C_{v_{\mathcal{L}}}^{(\mathcal{L})} = {{|{R\hspace{0pt}{(r)}}|} - 1}`$ when $`v_{\mathcal{L}}`$ is within the border area of width $`W_{\mathcal{L}}`$, and $`C_{v_{\mathcal{L}}}^{(\mathcal{L})} = 1`$ otherwise.
- •
$`S_{\mathcal{L}}`$ is the set including the point $`s_{\mathcal{L}}`$, and the points of the lattice $`\mathcal{L}`$ such as only nodes of $`S`$ are mapped to them:
$`S_{\mathcal{L}} \triangleq {{\{ s_{\mathcal{L}}\}} \cup {\{ u_{\mathcal{L}}:{{\lambda^{- 1}\hspace{0pt}{(u_{\mathcal{L}})}} \subset S}\}}}`$
- •
$`T_{\mathcal{L}}`$ is the set of the rest of points of $`\mathcal{L}`$.
Note that $`t \in T_{\mathcal{L}}`$ ; that all the points of the lattice, to which both points from $`S`$ and $`T`$ are mapped, those points are in $`T_{\mathcal{L}}`$ ; and that the points to which no points are mapped are in $`S_{\mathcal{L}}`$: $`S_{\mathcal{L}}/T_{\mathcal{L}}`$ is indeed a partition and a $`s_{\mathcal{L}} - t_{\mathcal{L}}`$ cut.
######
Lemma 3.7
.
The capacity $`C\hspace{0pt}{(S)}`$ of the cut $`S/T`$ and the capacity of the induced cut $`C^{(\mathcal{L})}\hspace{0pt}{(S_{\mathcal{L}})}`$ verify: $`{C\hspace{0pt}{(S)}} \geq {m_{\min}\hspace{0pt}C^{(\mathcal{L})}\hspace{0pt}{(S_{\mathcal{L}})}}`$
This comes from the fact that neighborhood on the lattice implies neighborhood in $`\mathcal{V}`$ (lemma [3.6](#S3.Thmtheorem6 "Lemma 3.6. ‣ 3.2.2 Neighborhood of the Embedded Lattice ‣ 3.2 Proof of the Value of Min-Cut for Unit Disk Graphs ‣ 3 Proofs of the Min-Cut")), and then an inclusion is proved between the $`\Delta\hspace{0pt}S_{\mathcal{L}}`$ of the capacity of cut of the lattice from ([2](#S1.E2 "In 1.2 Network Coding ‣ 1 Network Model and Related Work")) and the $`\Delta\hspace{0pt}S`$ of the cut $`S/T`$ \[[17](#bib.bib17)\].
######
Theorem 3.8
.
The min-cut $`C_{\min}\hspace{0pt}{(s)}`$ of the graph $`\mathcal{V}`$, verifies:
$`{C_{\min}\hspace{0pt}{(s)}} \geq {m_{\min}\hspace{0pt}{({{|{R\hspace{0pt}{(r)}}|} - 1})}}`$
######
Proof 3.9
.
From lemma [3.7](#S3.Thmtheorem7 "Lemma 3.7. ‣ 3.2.3 Relationship between the Capacities of the Cuts of the Embedded Lattice and the Random Disk Unit Graph ‣ 3.2 Proof of the Value of Min-Cut for Unit Disk Graphs ‣ 3 Proofs of the Min-Cut"), any cut $`C\hspace{0pt}{(S)}`$ is lower bounded by $`m_{\min}\hspace{0pt}C^{(\mathcal{L})}\hspace{0pt}{(S_{\mathcal{L}})}`$. Since $`C^{(\mathcal{L})}\hspace{0pt}{(S_{\mathcal{L}})}`$ is the capacity of a cut of a lattice with IREN/IRON, Th. [3.4](#S3.Thmtheorem4 "Theorem 3.4. ‣ 3.1.3 Value of the Min-cut 𝐶ₘᵢₙ(𝑠) ‣ 3.1 Proof for Lattice Graphs ‣ 3 Proofs of the Min-Cut") also indicates that: $`{C^{(\mathcal{L})}\hspace{0pt}{(S_{\mathcal{L}})}} \geq {C_{\min}^{(\mathcal{L})}\hspace{0pt}{(s_{\mathcal{L}})}} = {{|{R\hspace{0pt}{(r)}}|} - 1}`$. Hence the lower bound $`m_{\min}\hspace{0pt}{({{|{R\hspace{0pt}{(r)}}|} - 1})}`$ for any $`C\hspace{0pt}{(S)}`$, and therefore for the min-cut $`C_{\min}\hspace{0pt}{(s)}`$.
#### 3.2.4 Nodes of $`\mathcal{V}`$ Mapped to One Lattice Point
In Th. [3.8](#S3.Thmtheorem8 "Theorem 3.8. ‣ 3.2.3 Relationship between the Capacities of the Cuts of the Embedded Lattice and the Random Disk Unit Graph ‣ 3.2 Proof of the Value of Min-Cut for Unit Disk Graphs ‣ 3 Proofs of the Min-Cut"), $`m_{\min}`$ plays a central part. Let us start with $`m\hspace{0pt}{(u_{\mathcal{L}})}`$: it is actually a random variable that is the sum of $`N`$ Bernoulli trials. With a Chernoff tail bound \[[16](#bib.bib16)\], we get, for $`\delta \in \rbrack 0,1\lbrack`$:
$`{P\hspace{0pt}r\hspace{0pt}{\lbrack{{m\hspace{0pt}{(u_{\mathcal{L}})}} \leq {{({1 - \delta})}\hspace{0pt}E\hspace{0pt}{\lbrack{m\hspace{0pt}{(u_{\mathcal{L}})}}\rbrack}}}\rbrack}} \leq {\exp{({- \frac{E\hspace{0pt}{\lbrack{m\hspace{0pt}{(u_{\mathcal{L}})}}\rbrack}\hspace{0pt}\delta^{2}}{2}})}}`$
A bound on $`m_{\min}`$ is deduced from the fact that it is the minimum of $`m\hspace{0pt}{(u)}`$ and from the fact that for two events $`A`$ and $`B`$,
$`Pr{\lbrack A}`$ or $`B\rbrack \leq Pr\lbrack A\rbrack + Pr\lbrack B\rbrack`$:
######
Theorem 3.10
.
$`{P\hspace{0pt}r\hspace{0pt}{\lbrack{m_{\min} \leq {{({1 - \delta})}\hspace{0pt}\mu\hspace{0pt}r^{2}}}\rbrack}} \leq {\exp\left( {{({\log\frac{L^{2}}{r^{2}}})}\hspace{0pt}{({1 - \frac{\mu\hspace{0pt}r^{2}\hspace{0pt}\delta^{2}}{2\hspace{0pt}{\log\frac{L^{2}}{r^{2}}}}})}} \right)}`$
#### 3.2.5 Asymptotic Values of the Min-Cut of Unit-Disk Graphs
######
Theorem 3.11
.
For a sequence of random unit disk graphs and associated sources $`({{\mathcal{V}_{i},s_{i}} \in \mathcal{V}_{i}})`$, with fixed radio range $`\rho`$, fixed border area width $`W`$, with a size $`L_{i}\rightarrow\infty`$, and a density $`M = L^{\theta}`$ with fixed $`\theta > 0`$, we have the following limit of the min-cut $`C_{\min}\hspace{0pt}{(s_{i})}`$:
$`\frac{C_{\min}\hspace{0pt}{(s_{i})}}{M}\overset{p}{\rightarrow}\, 1{in}{probability}.{Additionally}:\frac{M_{\max}}{M}\overset{p}{\rightarrow}\, 1`$
######
Proof 3.12
.
Notice that $`{P\hspace{0pt}r\hspace{0pt}{\lbrack{m_{\min} > {\mu\hspace{0pt}r^{2}}}\rbrack}} = 0`$, because otherwise we would have a minimum of some values $`m\hspace{0pt}{(u_{\mathcal{L}})}`$ greater than their average. Starting from Th. [3.10](#S3.Thmtheorem10 "Theorem 3.10. ‣ 3.2.4 Nodes of 𝒱 Mapped to One Lattice Point ‣ 3.2 Proof of the Value of Min-Cut for Unit Disk Graphs ‣ 3 Proofs of the Min-Cut"), several variables appear: $`L`$, $`\mu`$, $`\delta`$, and $`r`$. Assume that $`\rho`$ is fixed and that $`\mu = {K\hspace{0pt}L^{\theta}}`$ for some fixed $`\theta > 0`$ and $`K > 0`$. Then we propose the following settings: $`\delta = L^{- \frac{\theta}{8}}`$ ; $`r = L^{- \frac{\theta}{8}}`$
As a result, from Th. [3.10](#S3.Thmtheorem10 "Theorem 3.10. ‣ 3.2.4 Nodes of 𝒱 Mapped to One Lattice Point ‣ 3.2 Proof of the Value of Min-Cut for Unit Disk Graphs ‣ 3 Proofs of the Min-Cut") we have: $`\frac{m_{\min}}{\mu\hspace{0pt}r^{2}}\overset{p}{\rightarrow}\, 1`$. Now Th. [3.8](#S3.Thmtheorem8 "Theorem 3.8. ‣ 3.2.3 Relationship between the Capacities of the Cuts of the Embedded Lattice and the Random Disk Unit Graph ‣ 3.2 Proof of the Value of Min-Cut for Unit Disk Graphs ‣ 3 Proofs of the Min-Cut") gives:
$`{C_{\min}\hspace{0pt}{(s)}} \geq {m_{\min}\hspace{0pt}{({{|{R\hspace{0pt}{(r)}}|} - 1})}}`$. Hence: $`\frac{C_{\min}\hspace{0pt}{(s)}}{M} \geq {\frac{m_{\min}}{\mu\hspace{0pt}r^{2}}\hspace{0pt}\frac{\mu}{M}\hspace{0pt}r^{2}\hspace{0pt}{({{|{R\hspace{0pt}{(r)}}|} - 1})}}`$
From the fact that $`{|{R\hspace{0pt}{(r)}}|} = {{\pi\hspace{0pt}{(\frac{\rho}{r})}^{2}} + {O\hspace{0pt}{(\frac{1}{r})}}}`$ \[[17](#bib.bib17)\], and that $`M = {\pi\hspace{0pt}\rho^{2}\hspace{0pt}\mu}`$, we get the lower bound $`1`$ for the limit of $`\frac{C_{\min}}{M}`$. The upper bound comes indirectly from the fact that $`\frac{M_{\max}}{M}\overset{p}{\rightarrow}\, 1`$ \[[17](#bib.bib17)\]. $`\square`$
## 4 Conclusion
We have presented a simple rate selection for network coding for large sensor networks. We computed the broadcast performance from the min-cut with networks modeled as hypergraphs. The central result is that selecting nearly the same rate for all nodes achieves asymptotic optimality for the homogeneous networks that are presented, when the size of the networks becomes larger. This can be translated into this remarkable property: nearly every transmission becomes innovative for the receivers. As a result, it was shown that network coding would asymptotically outperform any method that does not use network coding. We believe that the results presented here are a first step for a simple but efficient rate selection in wireless sensor networks in the plane. Future research work will determine how to adapt the rate selection for smaller and less homogeneous networks.
{chapthebibliography}
1
## References
- \[1\] R. Ahlswede, N. Cai, S.-Y. R. Li and R. W. Yeung, *“Network Information Flow”*, IEEE Trans. on Information Theory, vol. 46, no.4, pp. 1204-1216, Jul. 2000
- \[2\]
- \[3\] A. Dana, R. Gowaikar, R. Palanki, B. Hassibi, and M. Effros, *“Capacity of Wireless Erasure Networks”*, IEEE Trans. on Information Theory, vol. 52, no.3, pp. 789-804, Mar. 2006
- \[4\] D. S. Lun, M. Médard, R. Koetter, and M. Effros, *“On coding for reliable communication over packet networks”*, Technical Report \#2741, MIT LIDS, Jan. 2007
- \[5\] Y. Wu, P. A. Chou, and S.-Y. Kung, *“Minimum-energy multicast in mobile ad hoc networks using network coding”*, IEEE Trans. Commun., vol. 53, no. 11, pp. 1906-1918, Nov. 2005
- \[6\] D. S. Lun, N. Ratnakar, M. Médard, R. Koetter, D. R. Karger, T. Ho, E. Ahmed, and F. Zhao, *“Minimum-Cost Multicast over Coded Packet Networks”*, IEEE/ACM Trans. Netw., vol. 52, no. 6, pp 2608-2623, Jun. 2006
- \[7\] A. Ramamoorthy, J. Shi, and R. D. Wesel, *“On the Capacity of Network Coding for Random Networks”*, IEEE Trans. on Information Theory, Vol. 51 No. 8, pp. 2878-2885, Aug. 2005
- \[8\] S. A. Aly, V. Kapoor, J. Meng, and A. Klappenecker, *“Bounds on the Network Coding Capacity for Wireless Random Networks”*, Third Workshop on Network Coding, Theory, and Applications (Netcoding07), Jan. 2007
- \[9\] R. A. Costa and J. Barros. *“Dual Radio Networks: Capacity and Connectivity”*, Spatial Stochastic Models in Wireless Networks (SpaSWiN 2007), Apr. 2007
- \[10\] D. S. Lun, M. Médard, R. Koetter, and M. Effros, *“Further Results on Coding for Reliable Communication over Packet Networks”* International Symposium on Information Theory (ISIT 2005), Sept. 2005
- \[11\] T. Ho, R. Koetter, M. Médard, D. Karger and M. Effros, *“The Benefits of Coding over Routing in a Randomized Setting”*, International Symposium on Information Theory (ISIT 2003), Jun. 2003
- \[12\] B. Clark, C. Colbourn, and D. Johnson, *“Unit disk graphs”*, Discrete Mathematics, Vol. 86, Issues 1-3, Dec. 1990
- \[13\] C. Fragouli, J. Widmer, and J.-Y. L. Boudec, *“A Network Coding Approach to Energy Efficient Broadcasting”*, Proceedings of INFOCOM 2006, Apr. 2006
- \[14\] I.K. Lee, M.S. Kim, G. Elber, *“Polynomial/Rational Approximation of Minkowski Sum Boundary Curves”*, Graphical Models and Image Processing, Vol. 69, No. 2, pp 136-165, Mar. 1998
- \[15\] R. J. Gardner, P. Gronchi, *“A Brunn-Minkowski inequality for the integer lattice”*, Trans. Amer. Math. Soc., 353 (2001), 3995-4042
- \[16\] P. Barbe, M. Ledoux, *Probabilité* Editions Espaces 34, Belin, 1998.
- \[17\] C. Adjih, S. Y. Cho, P. Jacquet, *“Near Optimal Broadcast with Network Coding in Large Homogeneous Networks”*, `http://hal.inria.fr/inria-00145231/en/`, INRIA Research Report, May 2007 |
0708.1416 | cs.NI | cyber | # MIMO-OFDM Optimal Decoding And Achievable Information Rates Under Imperfect Channel Estimation
###### Abstract
Optimal decoding of bit interleaved coded modulation (BICM) MIMO-OFDM where an imperfect channel estimate is available at the receiver is investigated. First, by using a Bayesian approach involving the channel a posteriori density, we derive a practical decoding metric for general memoryless channels that is robust to the presence of channel estimation errors. Then, we evaluate the outage rates achieved by a decoder that uses our proposed metric. The performance of the proposed decoder is compared to the classical mismatched decoder and a theoretical decoder defined as the best decoder in the presence of imperfect channel estimation. Numerical results over Rayleigh block fading MIMO-OFDM channels show that the proposed decoder outperforms mismatched decoding in terms of bit error rate and outage capacity without introducing any additional complexity.
## 1 Introduction
Multiple-input multiple-output (MIMO) antennas systems is a promising technique for high-speed, spectrally efficient and reliable wireless communications. However, as higher data rates lead to wideband communications, the underlying MIMO channels exhibit strong frequency selectivity. By using orthogonal frequency division multiplexing (OFDM) and applying a proper cyclic prefix (CP), the frequency selective channels are transformed to an equivalent set of frequency-flat subchannels. These considerations motivate the combination of MIMO and OFDM, referred to as MIMO-OFDM, as a promising technology for the future generation of wireless systems \[[1](#bib.bib1)\].
It is well known that reliable coherent data detection is not possible unless an accurate channel state information (CSI) is available at the receiver. A typical scenario for wireless communication systems assumes the channel changing so slowly that it can be considered time invariant during the transmission of an entire frame. In such situations, pilot symbol assisted modulation (PSAM) has been shown to be an effective solution for obtaining CSI. It consists of inserting known training symbols (pilot) at the beginning of the information frame from which the receiver estimates the channel before decoding the rest of the frame based on that channel estimate. However, in practical systems, due to the finite number of pilot symbols and noise, the receiver has only access to an imperfect (possibly very bad) estimate of the channel.
In order to deal with imperfect CSI, one sub-optimal technique, known as mismatched decoding, consists in replacing the exact channel by its estimate in the receiver metric. Although, this scheme is not adapted to the presence of channel estimation errors (CEE), it has been extensively adopted for performance evaluation of single an multi-carrier MIMO systems \[[2](#bib.bib2)\]. However, adopting a mismatched decoding approach arises two important questions: i) what are the limits of achievable information rates and how close it is to the rates obtained by a theoretical decoder in the presence of CEE ? and ii) what type of practical encoder/decoder can achieve the best performance under imperfect channel estimation ? The first problem has been addressed in \[[3](#bib.bib3)\] where mismatched decoding has been shown to be largely suboptimal compared to the capacity provided by a theoretical decoder. Since the theoretical encoder/decoder can not be implemented on practical communication systems, we have recently addressed in \[[4](#bib.bib4)\] the second question in the case of multicarrier systems by deriving a practical decoding metric adapted to imperfect channel estimation and its associated achievable rates.
In this work, we address the above problems for more general wideband block fading MIMO-OFDM channels estimated by a finite number of training sequence. In our approach we exploit an interesting feature of PSAM, the availability of CEE statistics, in order to derive the a posteriori pdf of the perfect channel conditioned on its estimate. The latter let us to formulate a new channel likelihood as a function of the estimated channel and derive our improved decoding rule adapted to the presence of CEE. Interestingly, the present metric coincides with that proposed in \[[5](#bib.bib5)\] for MIMO space-time decoding. Furthermore, we evaluate the capacity associated to the improved and mismatched decoders by using the results obtained in \[[6](#bib.bib6)\]. Although in this work we will focus on wideband MIMO-OFDM systems working in a bit interleaved coded modulation (BICM) framework, our results can be applied to general memoryless channels.
The outline of this paper is as follows. In Section <a href="#S2" style="font-size:90%;" title="2 System Model and Channel Estimation ‣ MIMO-OFDM Optimal Decoding And Achievable Information Rates Under Imperfect Channel Estimation">2</a> we describe the MIMO-OFDM channel and its pilot based estimation. We also calculate the posterior distribution of the perfect channel conditioned on its estimate. This posterior distribution is used in section <a href="#S3" style="font-size:90%;" title="3 Maximum-Likelihood Detection in the Presence of Channel Estimation Errors ‣ MIMO-OFDM Optimal Decoding And Achievable Information Rates Under Imperfect Channel Estimation">3</a> to derive the improved decoding metric in the presence of imperfect channel state information at the receiver (CSIR). Section <a href="#S4" style="font-size:90%;" title="4 Iterative Decoding of BICM MIMO-OFDM Based On Imperfect CSIR ‣ MIMO-OFDM Optimal Decoding And Achievable Information Rates Under Imperfect Channel Estimation">4</a>, provides the application of the improved metric for iterative decoding of BICM MIMO-OFDM. In section <a href="#S5" style="font-size:90%;" title="5 Achievable Rates of MIMO-OFDM Under Channel Estimation Errors ‣ MIMO-OFDM Optimal Decoding And Achievable Information Rates Under Imperfect Channel Estimation">5</a>, we evaluate the achievable outage rates of a receiver using the proposed metric. Section <a href="#S6" style="font-size:90%;" title="6 Numerical Results ‣ MIMO-OFDM Optimal Decoding And Achievable Information Rates Under Imperfect Channel Estimation">6</a> illustrates via simulations, a comparative performance study of the proposed decoder and section <a href="#S7" style="font-size:90%;" title="7 Conclusion ‣ MIMO-OFDM Optimal Decoding And Achievable Information Rates Under Imperfect Channel Estimation">7</a> concludes the paper.
Notational conventions are as follows. Upper and lower case bold symbols are used to denote matrices and vectors, respectively; $`{\mathbb{I}}_{N}`$ represents an $`N \times N`$ identity matrix; $`{\mathbb{E}}_{\mathbf{x}}\left\lbrack . \right\rbrack`$ refers to expectation with respect to $`\mathbf{x}`$; $`\left| . \right|`$ and $`\parallel . \parallel`$ and $`{Tr}\left( . \right)`$ denote matrix determinant, Frobenius norm and matrix trace respectively; $`\left( . \right)^{T}`$ and $`\left( . \right)^{\mathcal{H}}`$ denote vector transpose and Hermitian transpose, respectively.
## 2 System Model and Channel Estimation
We consider a single-user MIMO-OFDM communication system over a memoryless frequency selective Rayleigh fading channel. The system consists of $`M_{T}`$ transmit and $`M_{R}`$ receive antennas ($`M_{R} \geq M_{T}`$), and $`M`$ is the total number of subcarriers. Fig. <a href="#S2.F1" style="font-size:90%;" title="Figure 1 ‣ 2 System Model and Channel Estimation ‣ MIMO-OFDM Optimal Decoding And Achievable Information Rates Under Imperfect Channel Estimation">1</a> depicts the BICM coding scheme used at the transmitter. The binary data sequence $`\mathbf{b}`$ are encoded by a non-recursive non-systematic convolutional (NRNSC) code before being interleaved by a quasi-random interleaver. The output bits $`\mathbf{d}`$ are multiplexed to $`M_{T}`$ substreams and mapped to complex $`M_{c}`$-QAM symbols before being modulated by the OFDM modulator and transmitted through $`M_{T}`$ antennas.
[Refer to caption](/html/0708.1416/assets/x1.png)
Fig. 1:
Block diagram of MIMO-OFDM transmission scheme.
Let $`\mathbf{s}`$ be the $`{M\hspace{0pt}M_{T}} \times 1`$ vector containing the OFDM symbols transmitted simultaneously over $`M_{T}`$ antennas. The symbols are assumed to be independent identically distributed (i.i.d.) with zero mean and unit covariance matrix $`\mathbf{\Sigma}_{s} = {{\mathbb{E}}\hspace{0pt}\left\lbrack {{\mathbf{s}}\hspace{0pt}{\mathbf{s}}^{\mathcal{H}}} \right\rbrack} = {\mathbb{I}}_{M \times M_{T}}`$. Assuming an invariant channel over a frame of $`L`$ symbols, the received vector $`\mathbf{y}`$ at a given time index $`l`$ (omitted for brevity) can be written as
<table>
<tbody>
<tr>
<td></td>
<td>𝒚=𝐇𝒔+𝒛𝒚𝐇𝒔𝒛{\bm{y}}={\mathbf{H}}\,{\bm{s}}+{\bm{z}}</td>
<td></td>
<td>(1)</td>
</tr>
</tbody>
</table>
where $`\mathbf{H}`$ is a $`{{M\hspace{0pt}M_{R}} \times M}\hspace{0pt}M_{T}`$ block diagonal channel matrix containing the frequency response of the MIMO channels and the noise vector $`\mathbf{z}`$ is assumed to be a zero-mean circularly symmetric complex Gaussian (ZMCSCG) random vector with covariance matrix $`\mathbf{\Sigma}_{z} \triangleq {{\mathbb{E}}\hspace{0pt}\left( {{\mathbf{z}}\hspace{0pt}{\mathbf{z}}^{\mathcal{H}}} \right)} = {\sigma_{z}^{2}\hspace{0pt}{\mathbb{I}}_{M \times M_{R}}}`$. We assume that for each frame, a different realization of $`\mathbf{H}`$ independent of both $`\mathbf{s}`$ and $`\mathbf{z}`$ is drawn and remains constant during this frame. The MIMO-OFDM channel can be decoupled into $`M`$ frequency flat MIMO channels by exploiting the block diagonal structure in (<a href="#S2.E1" style="font-size:90%;" title="In 2 System Model and Channel Estimation ‣ MIMO-OFDM Optimal Decoding And Achievable Information Rates Under Imperfect Channel Estimation">1</a>) which can be rewritten as a set of $`M`$ equations that contains only one subcarrier each
<table>
<tbody>
<tr>
<td></td>
<td>𝒚k=𝐇k𝒔k+𝒛kk=1,…,M,formulae-sequencesubscript𝒚𝑘subscript𝐇𝑘subscript𝒔𝑘subscript𝒛𝑘𝑘1…𝑀{\bm{y}}_{k}={\mathbf{H}}_{k}\,{\bm{s}}_{k}+{\bm{z}}_{k}\;\;\;\;k=1,...,M,</td>
<td></td>
<td>(2)</td>
</tr>
</tbody>
</table>
where $`\mathbf{H} = {{diag}\hspace{0pt}\left\lbrack {\mathbf{H}_{1}\hspace{0pt}\mathbf{H}_{2}\hspace{0pt}\ldots\hspace{0pt}\mathbf{H}_{M}} \right\rbrack}`$, $`{\mathbf{y}}^{T} = \left\lbrack {{\mathbf{y}}_{1}^{T}\hspace{0pt}\ldots\hspace{0pt}{\mathbf{y}}_{M}^{T}} \right\rbrack`$, $`{\mathbf{s}}^{T} = \left\lbrack {{\mathbf{s}}_{1}^{T}\hspace{0pt}\ldots\hspace{0pt}{\mathbf{s}}_{M}^{T}} \right\rbrack`$ and $`{\mathbf{z}}^{T} = \left\lbrack {{\mathbf{z}}_{1}^{T}\hspace{0pt}\ldots\hspace{0pt}{\mathbf{z}}_{M}^{T}} \right\rbrack`$. The architecture of (<a href="#S2.E2" style="font-size:90%;" title="In 2 System Model and Channel Estimation ‣ MIMO-OFDM Optimal Decoding And Achievable Information Rates Under Imperfect Channel Estimation">2</a>) constitutes the basis for the study in this paper.
Pilot Based Channel Estimation: We consider the estimation of channel matrix $`\mathbf{H}_{k}`$ via the transmission of $`N`$ training vectors $`{\mathbf{s}}_{_{T,i}}`$, ($`i = {1,\ldots,N}`$). According to (<a href="#S2.E2" style="font-size:90%;" title="In 2 System Model and Channel Estimation ‣ MIMO-OFDM Optimal Decoding And Achievable Information Rates Under Imperfect Channel Estimation">2</a>), when pilot symbols are transmitted we receive
<table>
<tbody>
<tr>
<td></td>
<td>𝐘T=𝐇k𝐒T+𝐙Tsubscript𝐘𝑇subscript𝐇𝑘subscript𝐒𝑇subscript𝐙𝑇\mathbf{Y}_{T}={\mathbf{H}}_{k}\,\mathbf{S}_{T}+\mathbf{Z}_{T}</td>
<td></td>
<td>(3)</td>
</tr>
</tbody>
</table>
where each column of the $`M_{T} \times N`$ matrix $`\mathbf{S}_{T} = \left\lbrack {{\mathbf{s}}_{_{T,1}}\hspace{0pt}|\ldots|\hspace{0pt}{\mathbf{s}}_{_{T,N}}} \right\rbrack`$ contains one pilot symbol and the noise $`\mathbf{Z}_{T}`$ has the same distribution as the noise $`{\mathbf{z}}_{k}`$. The average energy of the training symbols is $`P_{T} = {\frac{1}{N\hspace{0pt}M_{T}}\hspace{0pt}{Tr}\hspace{0pt}\left( {\mathbf{S}_{T}\hspace{0pt}\mathbf{S}_{T}^{\mathcal{H}}} \right)}`$. The ML estimate of $`\mathbf{H}_{k}`$ is obtained by minimizing $`\left\| {\mathbf{Y}_{T} - {\mathbf{H}_{k}\hspace{0pt}\mathbf{S}_{T}}} \right\|^{2}`$ with respect to $`\mathbf{H}_{k}`$. We have
<table>
<tbody>
<tr>
<td></td>
<td>𝐇^kML=𝐘T𝐒Tℋ(𝐒T𝐒Tℋ)−1=𝐇k+ℰsuperscriptsubscript^𝐇𝑘MLsubscript𝐘𝑇superscriptsubscript𝐒𝑇ℋsuperscriptsubscript𝐒𝑇superscriptsubscript𝐒𝑇ℋ1subscript𝐇𝑘ℰ\widehat{{\mathbf{H}}}_{k}^{\rm ML}=\mathbf{Y}_{T}\,\mathbf{S}_{T}^{\mathcal{H}}\,(\mathbf{S}_{T}\mathbf{S}_{T}^{\mathcal{H}})^{-1}={\mathbf{H}}_{k}+\mathbf{\mathcal{E}}</td>
<td></td>
<td>(4)</td>
</tr>
</tbody>
</table>
where $`\mathcal{E} = {\mathbf{Z}_{T}\hspace{0pt}\mathbf{S}_{T}^{\mathcal{H}}\hspace{0pt}\left( {\mathbf{S}_{T}\hspace{0pt}\mathbf{S}_{T}^{\mathcal{H}}} \right)^{- 1}}`$ denotes the estimation error matrix. When the training sequence is orthogonal ($`{\mathbf{S}_{T}\hspace{0pt}\mathbf{S}_{T}^{\mathcal{H}}} = {N\hspace{0pt}P_{T}\hspace{0pt}{\mathbb{I}}_{M_{T}}}`$), the $`j`$-th column $`\mathcal{E}_{j}`$ of the estimation error matrix reduces to a white noise vector with covariance matrix $`\mathbf{\Sigma}_{\mathcal{E}} = {{\mathbb{E}}\hspace{0pt}\left\lbrack {{\mathcal{E}}_{j}\hspace{0pt}{\mathcal{E}}_{j}^{\mathcal{H}}} \right\rbrack} = {\sigma_{\mathcal{E},k}^{2}\hspace{0pt}{\mathbb{I}}_{M_{R}}}`$ (for all $`j`$), where $`\sigma_{\mathcal{E},k}^{2} = \frac{\sigma_{z}^{2}}{N\hspace{0pt}P_{T}}`$. By assuming that the channel matrix $`\mathbf{H}_{k}`$ has the prior distribution $`\mathbf{H}_{k} \sim {\mathcal{C}\hspace{0pt}\mathcal{N}\hspace{0pt}\left( \mathbf{0},{{\mathbb{I}}_{M_{T}} \otimes \mathbf{\Sigma}_{H,k}} \right)}`$, ($`\mathbf{\Sigma}_{H,k} = {\sigma_{h}^{2}\hspace{0pt}{\mathbb{I}}_{M_{R}}}`$) and choosing an orthogonal training sequence, we can derive the posterior distribution of the perfect channel conditioned on the estimated channel as
<table>
<tbody>
<tr>
<td></td>
<td>f(𝐇k|𝐇^kML)=𝒞𝒩(𝚺Δ𝐇^ML,𝕀MT⊗𝚺Δ𝚺ℰ)𝑓conditionalsubscript𝐇𝑘superscriptsubscript^𝐇𝑘ML𝒞𝒩subscript𝚺Δsubscript^𝐇MLtensor-productsubscript𝕀subscript𝑀𝑇subscript𝚺Δsubscript𝚺ℰf({\mathbf{H}}_{k}|\widehat{{\mathbf{H}}}_{k}^{\rm ML})=\mathcal{CN}(\mathbf{\Sigma}_{\Delta}\widehat{{\mathbf{H}}}_{\rm ML},\,{\mathbb{I}}_{M_{T}}\otimes\mathbf{\Sigma}_{\Delta}\mathbf{\Sigma}_{\mathcal{E}})</td>
<td></td>
<td>(5)</td>
</tr>
</tbody>
</table>
where $`\mathbf{\Sigma}_{\Delta} = {\mathbf{\Sigma}_{H,k}\hspace{0pt}\left( {\mathbf{\Sigma}_{\mathcal{E}} + \mathbf{\Sigma}_{H,k}} \right)^{- 1}} = {\delta\hspace{0pt}{\mathbb{I}}_{M_{R}}}`$ and $`\delta = \frac{\sigma_{h}^{2}}{\sigma_{h}^{2} + \sigma_{\mathcal{E},k}^{2}}`$.
The availability of the estimation error distribution constitutes an interesting feature of pilot assisted channel estimation that we used to derive the posterior distribution (<a href="#S2.E5" style="font-size:90%;" title="In 2 System Model and Channel Estimation ‣ MIMO-OFDM Optimal Decoding And Achievable Information Rates Under Imperfect Channel Estimation">5</a>). This distribution is exploited in the next section, in the derivation of a new metric for improving the detection performance under imperfect channel estimation.
## 3 Maximum-Likelihood Detection in the Presence of Channel Estimation Errors
### 3.1 Mismatched ML Detection
It is well known that under i.i.d. Gaussian noise, detecting $`{\mathbf{s}}_{k}`$ is given by maximizing the likelihood function $`W\hspace{0pt}\left( {\mathbf{y}}_{k} \middle| {{\mathbf{s}}_{k},\mathbf{H}_{k}} \right)`$ which is equivalent to minimizing the Euclidean distance $`\mathcal{D}_{ML}`$
<table>
<tbody>
<tr>
<td></td>
<td>𝒔^kML(𝐇k)=argmin𝒔k∈ℂMT×1{𝒟ML(𝒔k,𝒚k,𝐇k)},superscriptsubscript^𝒔𝑘MLsubscript𝐇𝑘subscriptargminsubscript𝒔𝑘superscriptℂsubscript𝑀𝑇1subscript𝒟MLsubscript𝒔𝑘subscript𝒚𝑘subscript𝐇𝑘\hat{{\bm{s}}}_{k}^{\rm ML}({\mathbf{H}}_{k})=\mathop{\mathrm{arg\,min}}_{{\bm{s}}_{k}\in\,\mathbb{C}^{M_{T}\times 1}}\,\big{\{}\,\mathcal{D}_{\rm ML}({\bm{s}}_{k},{\bm{y}}_{k},{\mathbf{H}}_{k})\big{\}},</td>
<td></td>
<td>(6)</td>
</tr>
</tbody>
</table>
where $`{\mathcal{D}_{ML}\hspace{0pt}\left( {\mathbf{s}}_{k},{\mathbf{y}}_{k},\mathbf{H}_{k} \right)} \triangleq {- {{\ln W}\hspace{0pt}\left( {\mathbf{y}}_{k} \middle| {{\mathbf{s}}_{k},\mathbf{H}_{k}} \right)}} \propto \left\| {{\mathbf{y}}_{k} - {\mathbf{H}_{k}\hspace{0pt}{\mathbf{s}}_{k}}} \right\|^{2}`$. Since the above detection rule requires the knowledge of the perfect channel matrix $`\mathbf{H}_{k}`$, one sub-optimal approach, referred to as mismatched detection, consists in replacing the exact channel by its estimate in the receiver metric as
<table>
<tbody>
<tr>
<td></td>
<td>𝒔^kML(𝐇^k)=argmin𝒔k∈ℂMT×1{‖𝒚k−𝐇k𝒔k‖2}|𝐇k=𝐇^k\hat{{\bm{s}}}_{k}^{\rm ML}(\widehat{{\mathbf{H}}}_{k})=\mathop{\mathrm{arg\,min}}_{{\bm{s}}_{k}\in\,\mathbb{C}^{M_{T}\times 1}}\,\big{\{}\,\|{\bm{y}}_{k}-{\mathbf{H}}_{k}{\bm{s}}_{k}\|^{2}\big{\}}_{\big{|}_{{\mathbf{H}}_{k}=\widehat{{\mathbf{H}}}_{k}}}</td>
<td></td>
<td>(7)</td>
</tr>
</tbody>
</table>
### 3.2 Improved ML Detection for imperfect CSIR
The Bayesian framework introduced previously let us to define a new likelihood function $`\overset{\sim}{W}\hspace{0pt}\left( {\mathbf{y}}_{k} \middle| {{\hat{\mathbf{H}}}_{k},{\mathbf{s}}_{k}} \right)`$ by averaging the likelihood that would be used if the channel were perfectly known ($`W\hspace{0pt}\left( {\mathbf{y}}_{k} \middle| {H_{k},{\mathbf{s}}_{k}} \right)`$), over all realizations of the perfect channel for a given estimated channel state (the posterior distribution (<a href="#S2.E5" style="font-size:90%;" title="In 2 System Model and Channel Estimation ‣ MIMO-OFDM Optimal Decoding And Achievable Information Rates Under Imperfect Channel Estimation">5</a>)). This yields
<table>
<tbody>
<tr>
<td></td>
<td>W~(𝒚k|𝐇^k,𝒔k)~𝑊conditionalsubscript𝒚𝑘subscript^𝐇𝑘subscript𝒔𝑘\displaystyle{\widetilde{W}}({\bm{y}}_{k}|\widehat{{\mathbf{H}}}_{k},{\bm{s}}_{k})</td>
<td>=𝔼𝐇k|𝐇^k[W(𝒚k|𝐇k,𝒔k)|𝐇^k]absentsubscript𝔼conditionalsubscript𝐇𝑘subscript^𝐇𝑘delimited-[]conditional𝑊conditionalsubscript𝒚𝑘subscript𝐇𝑘subscript𝒔𝑘subscript^𝐇𝑘\displaystyle={\mathbb{E}}_{{\mathbf{H}}_{k}|\widehat{{\mathbf{H}}}_{k}}\big{[}{W}({\bm{y}}_{k}|{\mathbf{H}}_{k},{\bm{s}}_{k})\big{|}\,\widehat{{\mathbf{H}}}_{k}\big{]}</td>
<td></td>
<td></td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td></td>
<td>=∫𝐇k∈ℂMR×MTW(𝒚k|𝐇k,𝒔k)f(𝐇k|𝐇^k)d𝐇k.absentsubscriptsubscript𝐇𝑘superscriptℂsubscript𝑀𝑅subscript𝑀𝑇𝑊conditionalsubscript𝒚𝑘subscript𝐇𝑘subscript𝒔𝑘𝑓conditionalsubscript𝐇𝑘subscript^𝐇𝑘differential-dsubscript𝐇𝑘\displaystyle=\int_{{\mathbf{H}}_{k}\in\mathbb{C}^{M_{R}\times M_{T}}}{W}({\bm{y}}_{k}|{\mathbf{H}}_{k},{\bm{s}}_{k})\;f({\mathbf{H}}_{k}|\widehat{{\mathbf{H}}}_{k})\;\;\;{\rm d}{\mathbf{H}}_{k}.</td>
<td></td>
<td>(8)</td>
</tr>
</tbody>
</table>
Since both $`W\hspace{0pt}\left( {\mathbf{y}}_{k} \middle| {\mathbf{H}_{k},{\mathbf{s}}_{k}} \right)`$ and $`f\hspace{0pt}\left( \mathbf{H}_{k} \middle| {\hat{\mathbf{H}}}_{k} \right)`$ are gaussian densities, it is easy to show that $`{\overset{\sim}{W}\hspace{0pt}\left( {\mathbf{y}}_{k} \middle| {{\hat{\mathbf{H}}}_{k},{\mathbf{s}}_{k}} \right)} = {\mathcal{C}\hspace{0pt}\mathcal{N}\hspace{0pt}\left( {\mathbf{μ}}_{_{\mathcal{M}}},\mathbf{\Sigma}_{_{\mathcal{M}}} \right)}`$ with \[[4](#bib.bib4)\]
<table>
<tbody>
<tr>
<td></td>
<td>{𝝁ℳ=δ𝐇^k𝒔k,𝚺ℳ=𝚺z+δ𝚺ℰ‖𝒔k‖2.casessubscript𝝁ℳabsent𝛿subscript^𝐇𝑘subscript𝒔𝑘subscript𝚺ℳabsentsubscript𝚺𝑧𝛿subscript𝚺ℰsuperscriptnormsubscript𝒔𝑘2\left\{\begin{array}[]{ll}\bm{\mu}_{{}_{\mathcal{M}}}&=\delta\,\widehat{{\mathbf{H}}}_{k}\,{\bm{s}}_{k},\\ \mathbf{\Sigma}_{{}_{\mathcal{M}}}&=\mathbf{\Sigma}_{z}+\delta\,\mathbf{\Sigma}_{\mathcal{E}}\,\|{\bm{s}}_{k}\|^{2}.\end{array}\right.</td>
<td></td>
<td>(9)</td>
</tr>
</tbody>
</table>
Now the ML estimate of $`{\mathbf{s}}_{k}`$ can be formulated as
<table>
<tbody>
<tr>
<td></td>
<td>𝒔^kℳ(𝐇^k)=argmin𝒔k∈ℂMT×1{𝒟ℳ(𝒔k,𝒚k,𝐇k)},superscriptsubscript^𝒔𝑘ℳsubscript^𝐇𝑘subscriptargminsubscript𝒔𝑘superscriptℂsubscript𝑀𝑇1subscript𝒟ℳsubscript𝒔𝑘subscript𝒚𝑘subscript𝐇𝑘\hat{{\bm{s}}}_{k}^{\mathcal{M}}(\widehat{{\mathbf{H}}}_{k})=\mathop{\mathrm{arg\,min}}_{{\bm{s}}_{k}\in\,\mathbb{C}^{M_{T}\times 1}}\,\big{\{}\,\mathcal{D}_{\mathcal{M}}({\bm{s}}_{k},{\bm{y}}_{k},{\mathbf{H}}_{k})\big{\}},</td>
<td></td>
<td>(10)</td>
</tr>
</tbody>
</table>
with
<table>
<tbody>
<tr>
<td></td>
<td></td>
<td>𝒟ℳ(𝒔k,𝒚k,𝐇^k)≜−lnW~(𝒚k|𝒔k,𝐇^k)≜subscript𝒟ℳsubscript𝒔𝑘subscript𝒚𝑘subscript^𝐇𝑘~𝑊conditionalsubscript𝒚𝑘subscript𝒔𝑘subscript^𝐇𝑘\displaystyle\mathcal{D}_{{}_{\mathcal{M}}}({\bm{s}}_{k},{\bm{y}}_{k},\widehat{{\mathbf{H}}}_{k})\triangleq-\ln{\widetilde{W}}({\bm{y}}_{k}|{\bm{s}}_{k},\widehat{{\mathbf{H}}}_{k})</td>
<td></td>
<td></td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td></td>
<td>=MRlnπ(σz2+δσℰ2‖𝒔k‖2)+‖𝒚k−δ𝐇^k𝒔k‖2σz2+δσℰ2‖𝒔k‖2absentsubscript𝑀𝑅𝜋subscriptsuperscript𝜎2𝑧𝛿subscriptsuperscript𝜎2ℰsuperscriptnormsubscript𝒔𝑘2superscriptnormsubscript𝒚𝑘𝛿subscript^𝐇𝑘subscript𝒔𝑘2subscriptsuperscript𝜎2𝑧𝛿subscriptsuperscript𝜎2ℰsuperscriptnormsubscript𝒔𝑘2\displaystyle=M_{R}\ln\,\pi(\sigma^{2}_{z}+\delta\,\sigma^{2}_{\mathcal{E}}\,\|{\bm{s}}_{k}\|^{2})+\frac{\|{\bm{y}}_{k}-\delta\,\widehat{{\mathbf{H}}}_{k}\,{\bm{s}}_{k}\|^{2}}{\sigma^{2}_{z}+\delta\,\sigma^{2}_{\mathcal{E}}\,\|{\bm{s}}_{k}\|^{2}}</td>
<td></td>
<td>(11)</td>
</tr>
</tbody>
</table>
being the new ML decision metric under CEE.
We note that when exact channel is available ($`{\hat{\mathbf{H}}}_{k} = \mathbf{H}_{k}`$), the posterior expectation of (<a href="#S3.Ex1" style="font-size:90%;" title="3.2 Improved ML Detection for imperfect CSIR ‣ 3 Maximum-Likelihood Detection in the Presence of Channel Estimation Errors ‣ MIMO-OFDM Optimal Decoding And Achievable Information Rates Under Imperfect Channel Estimation">3.2</a>) becomes equivalent to replacing $`\mathbf{H}_{k}`$ by $`{\hat{\mathbf{H}}}_{k}`$ in $`W\hspace{0pt}\left( {\mathbf{y}}_{k} \middle| {\mathbf{H}_{k},{\mathbf{s}}_{k}} \right)`$ and consequently the the two metrics $`\mathcal{D}_{_{\mathcal{M}}}`$ and $`\mathcal{D}_{ML}`$ coincides. Under near perfect CSIR, occurred either when $`\sigma_{\mathcal{E}}^{2}\rightarrow 0`$ or when $`N\rightarrow\infty`$, we have $`\delta\rightarrow 1`$, $`{\delta\hspace{0pt}\sigma_{\mathcal{E}}^{2}}\rightarrow 0`$, $`{\mathbf{μ}}_{_{\mathcal{M}}}\rightarrow{{\hat{\mathbf{H}}}_{k}\hspace{0pt}{\mathbf{s}}}`$ and $`\mathbf{\Sigma}_{_{\mathcal{M}}}\rightarrow{\sigma_{z}^{2}\hspace{0pt}{\mathbb{I}}_{M_{R}}}`$. Consequently, the improved metric $`\mathcal{D}_{_{\mathcal{M}}}`$ behaves similarly to the classical Euclidean distance metric $`\mathcal{D}_{ML}`$.
[Refer to caption](/html/0708.1416/assets/x2.png)
Fig. 2:
Block diagram of MIMO-OFDM BICM receiver.
## 4 Iterative Decoding of BICM MIMO-OFDM Based On Imperfect CSIR
As a practical application of the general decoding metric (<a href="#S3.Ex2" style="font-size:90%;" title="3.2 Improved ML Detection for imperfect CSIR ‣ 3 Maximum-Likelihood Detection in the Presence of Channel Estimation Errors ‣ MIMO-OFDM Optimal Decoding And Achievable Information Rates Under Imperfect Channel Estimation">3.2</a>), we consider soft iterative decoding of BICM MIMO-OFDM under imperfect CSIR. This problem has been addressed in \[[7](#bib.bib7)\] under the assumption of perfect CSIR. Here, without going into the details, we extend the results of \[[4](#bib.bib4)\] to MIMO-OFDM block fading channels estimated by a finite number of training symbols.
As shown in Fig. <a href="#S3.F2" style="font-size:90%;" title="Figure 2 ‣ 3.2 Improved ML Detection for imperfect CSIR ‣ 3 Maximum-Likelihood Detection in the Presence of Channel Estimation Errors ‣ MIMO-OFDM Optimal Decoding And Achievable Information Rates Under Imperfect Channel Estimation">2</a>, the BICM receiver consists of a bunch of demodulator/demapper, a de-interleaver and a soft-input-soft-output (SISO) decoder. Let $`d_{k}^{i,m}`$ be the $`m`$-th coded and interleaved bit ($`{m = {1,2,\ldots}},{B = {\log_{2}M_{c}}}`$) of the constellation symbol $`{\mathbf{s}}_{k}`$ at the the $`i`$-th transmit antenna and the $`k`$-th subcarrier. We denote by $`L\hspace{0pt}\left( d_{k}^{i,m} \right)`$ the coded log-likelihood ratio (LLR) value of the bit $`d_{k}^{i,m}`$. At each decoding iteration, the LLR values conditioned on the CSIR are given by
<table>
<tbody>
<tr>
<td></td>
<td>L(dki,m)=logPdem(dki,m=1)|𝒚k,𝐇k)Pdem(dki,m=0|𝒚k,𝐇k).L(d_{k}^{i,m})=\log\frac{P_{\rm dem}(d_{k}^{i,m}=1)|{\bm{y}}_{k},{\mathbf{H}}_{k})}{P_{\rm dem}(d_{k}^{i,m}=0|{\bm{y}}_{k},{\mathbf{H}}_{k})}.</td>
<td></td>
<td>(12)</td>
</tr>
</tbody>
</table>
We have (see \[[4](#bib.bib4)\] and references therein)
<table>
<tbody>
<tr>
<td></td>
<td>L(dki,m)=log∑𝒔k:dki,m=1W(𝒚k|𝒔k,𝐇k)∏n≠mn=1BMTPdec1(dki,n)∑𝒔k:dki,m=0W(𝒚k|𝒔k,𝐇k)∏n≠mn=1BMTPdec0(dki,n)𝐿superscriptsubscript𝑑𝑘𝑖𝑚subscript:subscript𝒔𝑘superscriptsubscript𝑑𝑘𝑖𝑚1𝑊conditionalsubscript𝒚𝑘subscript𝒔𝑘subscript𝐇𝑘superscriptsubscriptproductsuperscript𝑛𝑚𝑛1𝐵subscript𝑀𝑇subscriptsuperscript𝑃1decsuperscriptsubscript𝑑𝑘𝑖𝑛subscript:subscript𝒔𝑘superscriptsubscript𝑑𝑘𝑖𝑚0𝑊conditionalsubscript𝒚𝑘subscript𝒔𝑘subscript𝐇𝑘superscriptsubscriptproductsuperscript𝑛𝑚𝑛1𝐵subscript𝑀𝑇subscriptsuperscript𝑃0decsuperscriptsubscript𝑑𝑘𝑖𝑛L(d_{k}^{i,m})=\log\frac{\sum\limits_{{\bm{s}}_{k}:d_{k}^{i,m}=1}{W}({\bm{y}}_{k}|{\bm{s}}_{k},{\mathbf{H}}_{k})\prod\limits_{\stackrel{{\scriptstyle n=1}}{{n\neq m}}}^{BM_{T}}P^{1}_{\rm dec}(d_{k}^{i,n})}{\sum\limits_{{\bm{s}}_{k}:d_{k}^{i,m}=0}{W}({\bm{y}}_{k}|{\bm{s}}_{k},{\mathbf{H}}_{k})\prod\limits_{\stackrel{{\scriptstyle n=1}}{{n\neq m}}}^{BM_{T}}P^{0}_{\rm dec}(d_{k}^{i,n})}\vspace{-2mm}</td>
<td></td>
<td>(13)</td>
</tr>
</tbody>
</table>
where $`P_{dec}^{1}\hspace{0pt}\left( d_{k}^{i,m} \right)`$ and $`P_{dec}^{0}\hspace{0pt}\left( d_{k}^{i,m} \right)`$ are extrinsic information coming from the SISO decoder.
Notice that the channel likelihood $`W\hspace{0pt}\left( {\mathbf{y}}_{k} \middle| {{\mathbf{s}}_{k},\mathbf{H}_{k}} \right)`$ involved in (<a href="#S4.E13" style="font-size:90%;" title="In 4 Iterative Decoding of BICM MIMO-OFDM Based On Imperfect CSIR ‣ MIMO-OFDM Optimal Decoding And Achievable Information Rates Under Imperfect Channel Estimation">13</a>) is conditioned on the perfect channel $`\mathbf{H}_{k}`$ of which the receiver has solely an estimate $`{\hat{\mathbf{H}}}_{k}`$. In order to enhance the detection performance under imperfect CSIR, we propose an improved decoder that uses $`\overset{\sim}{W}\hspace{0pt}\left( {\mathbf{y}}_{k} \middle| {{\mathbf{s}}_{k},{\hat{\mathbf{H}}}_{k}} \right)`$ of (<a href="#S3.Ex1" style="font-size:90%;" title="3.2 Improved ML Detection for imperfect CSIR ‣ 3 Maximum-Likelihood Detection in the Presence of Channel Estimation Errors ‣ MIMO-OFDM Optimal Decoding And Achievable Information Rates Under Imperfect Channel Estimation">3.2</a>) instead of $`W\hspace{0pt}\left( {\mathbf{y}}_{k} \middle| {{\mathbf{s}}_{k},{\hat{\mathbf{H}}}_{k}} \right)`$ (mismatched approach), for the derivation of the coded LLRs of (<a href="#S4.E13" style="font-size:90%;" title="In 4 Iterative Decoding of BICM MIMO-OFDM Based On Imperfect CSIR ‣ MIMO-OFDM Optimal Decoding And Achievable Information Rates Under Imperfect Channel Estimation">13</a>). The decoder accepts the LLRs of all coded bits and employs the well known forward-backward algorithm \[[8](#bib.bib8)\] to compute the LLRs of information bits, which are used for the decision.
## 5 Achievable Rates of MIMO-OFDM Under Channel Estimation Errors
In this section we provide the instantaneous achievable information rates $`C_{\mathcal{M}}`$ and $`C_{ML}`$ associated to a receiver using the decoding rules of (<a href="#S3.E10" style="font-size:90%;" title="In 3.2 Improved ML Detection for imperfect CSIR ‣ 3 Maximum-Likelihood Detection in the Presence of Channel Estimation Errors ‣ MIMO-OFDM Optimal Decoding And Achievable Information Rates Under Imperfect Channel Estimation">10</a>) (improved) and (<a href="#S3.E7" style="font-size:90%;" title="In 3.1 Mismatched ML Detection ‣ 3 Maximum-Likelihood Detection in the Presence of Channel Estimation Errors ‣ MIMO-OFDM Optimal Decoding And Achievable Information Rates Under Imperfect Channel Estimation">7</a>) (mismatched), respectively. We consider a MIMO-OFDM channel $`{W\hspace{0pt}\left( {\mathbf{y}} \middle| {{\mathbf{s}},\mathbf{H}} \right)} = {\prod_{k = 1}^{M}{\mathcal{C}\hspace{0pt}\mathcal{N}\hspace{0pt}\left( {\mathbf{H}_{k}\hspace{0pt}{\mathbf{s}}_{k}},\mathbf{\Sigma}_{{\mathbf{z}}_{k}} \right)}}`$, with $`\mathbf{\Sigma}_{{\mathbf{z}}_{k}} = {\sigma_{z,k}^{2}\hspace{0pt}{\mathbb{I}}_{M_{R}}}`$. Furthermore, we assume that the transmitter does not disposes of the channel estimate to perform power control. Thus, equal power is allocated to symbols $`{\mathbf{s}}_{k}`$ by setting $`\mathbf{\Sigma}_{{\mathbf{s}}_{k}} = {\overline{P}\hspace{0pt}{\mathbb{I}}_{M_{T}}}`$.
### 5.1 Achievable Rates Associated to the Improved Decoder
The capacity of a general mismatched decoder assuming Gaussian i.i.d inputs $`{\mathbf{s}} \sim {\mathcal{C}\hspace{0pt}\mathcal{N}\hspace{0pt}\left( \mathbf{0},\mathbf{\Sigma}_{s} \right)}`$ is given in \[[6](#bib.bib6)\] by minimizing the mutual information $`I\hspace{0pt}(S;Y)`$ defined for MIMO-OFDM channels as \[[1](#bib.bib1)\]
<table>
<tbody>
<tr>
<td></td>
<td>I(S;Y)≜1M∑k=1Mlog2det(𝕀MR+Υk𝚺𝒔kΥkℋ𝚺k−1)≜𝐼𝑆𝑌1𝑀superscriptsubscript𝑘1𝑀subscript2subscript𝕀subscript𝑀𝑅subscriptΥ𝑘subscript𝚺subscript𝒔𝑘superscriptsubscriptΥ𝑘ℋsuperscriptsubscript𝚺𝑘1I(S;Y)\triangleq\frac{1}{M}\sum_{k=1}^{M}\log_{2}\det\left({\mathbb{I}}_{M_{R}}+\Upsilon_{k}\mathbf{\Sigma}_{{\bm{s}}_{k}}\Upsilon_{k}^{\mathcal{H}}\mathbf{\Sigma}_{k}^{-1}\right)</td>
<td></td>
<td>(14)</td>
</tr>
</tbody>
</table>
where the minimization is done over all general channels $`{V\hspace{0pt}\left( {\mathbf{y}} \middle| {{\mathbf{s}},\Upsilon} \right)} = {\prod_{k = 1}^{M}{\mathcal{C}\hspace{0pt}\mathcal{N}\hspace{0pt}\left( {\Upsilon_{k}\hspace{0pt}{\mathbf{s}}_{k}},\mathbf{\Sigma}_{k} \right)}}`$. Since in the present work the decoding is performed on a single subcarrier basis, the minimization can be done independently for each subcarrier according to its corresponding channel and estimate pair $`\left( \mathbf{H}_{k},{\hat{\mathbf{H}}}_{k} \right)`$. Thus, we have to find the optimal $`\Upsilon_{k} \in {\mathbb{C}}^{M_{R} \times M_{T}}`$ and covariance matrices $`\Sigma_{k} = {\sigma_{k}^{2}\hspace{0pt}{\mathbb{I}}_{M_{R}}}`$ so as to minimize $`I\hspace{0pt}(S;Y)`$ under the set of constraints \[[6](#bib.bib6)\]
$`\left( c_{1} \right):{{{Tr}\hspace{0pt}\left( {{\mathbb{E}}_{\mathbf{s}}\hspace{0pt}\left\lbrack {V\hspace{0pt}\left( \mathbf{y} \middle| {\mathbf{s},\Upsilon} \right)} \right\rbrack} \right)} = {{Tr}\hspace{0pt}\left( {{\mathbb{E}}_{\mathbf{s}}\hspace{0pt}\left\lbrack {W\hspace{0pt}\left( \mathbf{y} \middle| {\mathbf{s},\mathbf{H}} \right)} \right\rbrack} \right)}}`$,
$`\left( c_{2,k} \right):{{{\mathbb{E}}_{{\mathbf{s}}_{k}}\hspace{0pt}\left\lbrack {{\mathbb{E}}_{V}\hspace{0pt}\left\lbrack {\mathcal{D}_{\mathcal{M}}\hspace{0pt}\left( {\mathbf{s}}_{k},\left. {\mathbf{y}}_{k} \middle| {\hat{\mathbf{H}}}_{k} \right. \right)} \right\rbrack} \right\rbrack} \leq {{\mathbb{E}}_{{\mathbf{s}}_{k}}\hspace{0pt}\left\lbrack {{\mathbb{E}}_{W}\hspace{0pt}\left\lbrack {\mathcal{D}_{\mathcal{M}}\hspace{0pt}\left( {\mathbf{s}}_{k},\left. {\mathbf{y}}_{k} \middle| {\hat{\mathbf{H}}}_{k} \right. \right)} \right\rbrack} \right\rbrack}}`$,
for $`k = {1,\ldots,M}`$. Applying our channel model (<a href="#S2.E1" style="font-size:90%;" title="In 2 System Model and Channel Estimation ‣ MIMO-OFDM Optimal Decoding And Achievable Information Rates Under Imperfect Channel Estimation">1</a>), and after some calculus, the above set of two constraints are obtained as
<table>
<tbody>
<tr>
<td></td>
<td>(c1):det(Υ𝚺𝒔Υℋ+𝚺)=det(𝐇𝚺𝒔𝐇ℋ+𝚺𝒛):subscriptc1Υsubscript𝚺𝒔superscriptΥℋ𝚺𝐇subscript𝚺𝒔superscript𝐇ℋsubscript𝚺𝒛\displaystyle\mathrm{(c_{1})}:\det\left(\Upsilon\mathbf{\Sigma}_{\bm{s}}\Upsilon^{\mathcal{H}}+\mathbf{\Sigma}\right)=\det\left({\mathbf{H}}\mathbf{\Sigma}_{\bm{s}}{\mathbf{H}}^{\mathcal{H}}+\mathbf{\Sigma}_{{\bm{z}}}\right)</td>
<td></td>
<td>(15)</td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td>(c2,k):‖Υk+ak𝐇^k‖2≤‖𝐇k+ak𝐇^k‖2+Cstk,:subscriptc2ksuperscriptnormsubscriptΥ𝑘subscript𝑎𝑘subscript^𝐇𝑘2superscriptnormsubscript𝐇𝑘subscript𝑎𝑘subscript^𝐇𝑘2subscriptCst𝑘\displaystyle\mathrm{(c_{2,k})}:\|\Upsilon_{k}+a_{k}\widehat{{\mathbf{H}}}_{k}\|^{2}\leq\|{\mathbf{H}}_{k}+a_{k}\widehat{{\mathbf{H}}}_{k}\|^{2}+\textrm{Cst}_{k},</td>
<td></td>
<td>(16)</td>
</tr>
</tbody>
</table>
where
$`a_{k} = {\delta_{k}\hspace{0pt}\left( {{\delta_{k}\hspace{0pt}\sigma_{\mathcal{E},k}^{2}\hspace{0pt}\overline{P}} - {\lambda_{n,k}\hspace{0pt}\sigma_{z,k}^{2}}} \right)\hspace{0pt}\left\lbrack {{{M_{T}\hspace{0pt}\delta\hspace{0pt}\sigma_{\mathcal{E},k}^{2}\hspace{0pt}\lambda_{n,k}\hspace{0pt}\overline{P}} + {\lambda_{n,k}\hspace{0pt}\sigma_{z,k}^{2}}} - {\delta_{k}\hspace{0pt}\sigma_{\mathcal{E},k}^{2}\hspace{0pt}\overline{P}}} \right\rbrack^{- 1}}`$; $`\text{Cst}_{k} = {M_{T}\hspace{0pt}\lambda_{n,k}\hspace{0pt}\left\lbrack {{\left\| \mathbf{H}_{k} \right\|^{2} - \left\| \Upsilon_{k} \right\|^{2}} + {\frac{1}{\overline{P}}\hspace{0pt}\left( {{{Tr}\hspace{0pt}\left( \mathbf{\Sigma}_{{\mathbf{z}}_{k}} \right)} - {{Tr}\hspace{0pt}\left( \mathbf{\Sigma}_{k} \right)}} \right)}} \right\rbrack\hspace{0pt}\left\lbrack {1 - {\frac{\sigma_{z,k}^{2}}{\delta_{k}\hspace{0pt}\overline{P}\hspace{0pt}\sigma_{\mathcal{E},k}^{2}}\hspace{0pt}\lambda_{n,k}} - {M_{T}\hspace{0pt}\lambda_{n,k}}} \right\rbrack^{- 1}}`$, and
$`\lambda_{n,k} = {\left( \frac{\sigma_{z,k}^{2}}{\delta_{k}\hspace{0pt}\overline{P}\hspace{0pt}\sigma_{\mathcal{E},k}^{2}} \right)^{n}\hspace{0pt}{\exp\left( \frac{\sigma_{Z,k}^{2}}{\delta_{k}\hspace{0pt}\overline{P}\hspace{0pt}\sigma_{\mathcal{E},k}^{2}} \right)}\hspace{0pt}\Gamma\hspace{0pt}\left( {- n},\frac{\sigma_{Z,k}^{2}}{\delta_{k}\hspace{0pt}\overline{P}\hspace{0pt}\sigma_{\mathcal{E},k}^{2}} \right)}`$,
with $`n = {M_{T} - 1}`$, and
<table>
<tbody>
<tr>
<td></td>
<td>Γ(−n,t)=(−1)nn![Γ(0,t)−exp(−t)∑i=0n−1(−1)ii!ti+1],Γ𝑛𝑡superscript1𝑛𝑛delimited-[]Γ0𝑡𝑡superscriptsubscript𝑖0𝑛1superscript1𝑖𝑖superscript𝑡𝑖1\Gamma(-n,t)=\frac{(-1)^{n}}{n!}\Big{[}\Gamma(0,t)-\exp(-t)\sum\limits_{i=0}^{n-1}(-1)^{i}\frac{i!}{t^{i+1}}\Big{]},\vspace{-1mm}</td>
<td></td>
</tr>
</tbody>
</table>
where $`\Gamma\hspace{0pt}(0,t)`$ is the exponential integral function.
We consider the singular value decomposition (SVD) of $`\mathbf{H}_{k} = {\mathbf{U}_{k}\hspace{0pt}\Lambda_{k}\hspace{0pt}\mathbf{V}_{k}^{\mathcal{H}}}`$ with $`\Lambda_{k} = {\text{diag}\hspace{0pt}\left( \lambda_{k,1},\ldots,\lambda_{k,M_{R}} \right)}`$. Let $`D_{\mu,k}`$ be a diagonal matrix such that $`D_{\mu,k} = {\mathbf{U}_{k}^{\mathcal{H}}\hspace{0pt}\Upsilon_{k}\hspace{0pt}\mathbf{V}_{k}}`$, whose diagonal values are given by the vector $`{\underset{¯}{\mu}}_{k} = \left( \mu_{k,1},\ldots,\mu_{k,M_{R}} \right)^{T}`$. We define $`{\overset{\sim}{\mathbf{H}}}_{k}^{\mathcal{H}} = {\mathbf{V}_{k}^{\mathcal{H}}\hspace{0pt}{\hat{\mathbf{H}}}_{k}^{\mathcal{H}}\hspace{0pt}\mathbf{U}_{k}}`$ and the vector $`{\overset{\sim}{\mathbf{h}}}_{k}^{\mathcal{H}} = {\text{diag}\hspace{0pt}\left( {\overset{\sim}{\mathbf{H}}}_{k}^{\mathcal{H}} \right)^{T}}`$ resulting of its diagonal and let $`b_{k} = {\left\| {\mathbf{H}_{k} + {a_{k}\hspace{0pt}{\hat{\mathbf{H}}}_{k}}} \right\|^{2} - {a_{k}^{2}\hspace{0pt}\left( {\left\| {\overset{\sim}{\mathbf{H}}}_{k} \right\|^{2} - \left\| {\overset{\sim}{\mathbf{h}}}_{k} \right\|^{2}} \right)}}`$. Using the above definitions, it is easy to reformulate (<a href="#S5.E14" style="font-size:90%;" title="In 5.1 Achievable Rates Associated to the Improved Decoder ‣ 5 Achievable Rates of MIMO-OFDM Under Channel Estimation Errors ‣ MIMO-OFDM Optimal Decoding And Achievable Information Rates Under Imperfect Channel Estimation">14</a>) as
<table>
<tbody>
<tr>
<td></td>
<td>Cℳ(𝐇,𝐇^)={minμ¯1M∑k=1M∑i=1MRlog2(1+P¯|μk,i|2σ2(μk¯)),subject to‖μ¯k+a𝐡~𝐤‖2≤bk.subscript𝐶ℳ𝐇^𝐇casessubscript¯𝜇1𝑀superscriptsubscript𝑘1𝑀superscriptsubscript𝑖1subscript𝑀𝑅subscript21¯𝑃superscriptsubscript𝜇𝑘𝑖2superscript𝜎2¯subscript𝜇𝑘missing-subexpressionsubject tosuperscriptnormsubscript¯𝜇𝑘𝑎subscript~𝐡𝐤2subscript𝑏𝑘missing-subexpressionC_{\mathcal{M}}({\mathbf{H}},\widehat{{\mathbf{H}}})=\left\{\begin{array}[]{ll}\min\limits_{\underline{\mu}}\,\,\,\,\,\,\frac{1}{M}\sum\limits_{k=1}^{M}\sum\limits_{i=1}^{M_{R}}\log_{2}\left(1+\displaystyle{\frac{\bar{P}|\mu_{k,i}|^{2}}{\sigma^{2}(\underline{\mu_{k}})}}\right),\\ \textrm{subject to}\,\,\,\,\,\,\|\underline{\mu}_{k}+a\mathbf{\tilde{h}_{k}}\|^{2}\leq b_{k}.\end{array}\right.</td>
<td></td>
<td>(17)</td>
</tr>
</tbody>
</table>
According to (<a href="#S5.E17" style="font-size:90%;" title="In 5.1 Achievable Rates Associated to the Improved Decoder ‣ 5 Achievable Rates of MIMO-OFDM Under Channel Estimation Errors ‣ MIMO-OFDM Optimal Decoding And Achievable Information Rates Under Imperfect Channel Estimation">17</a>), for each subcarrier $`k = {1,\ldots,M}`$, the achievable rates associated to the metric $`\mathcal{D}_{\mathcal{M}}`$ of (<a href="#S3.E10" style="font-size:90%;" title="In 3.2 Improved ML Detection for imperfect CSIR ‣ 3 Maximum-Likelihood Detection in the Presence of Channel Estimation Errors ‣ MIMO-OFDM Optimal Decoding And Achievable Information Rates Under Imperfect Channel Estimation">10</a>), can be obtained by using standard Lagrange method. We have
<table>
<tbody>
<tr>
<td></td>
<td>Cℳ(𝐇,𝐇^)=1M∑k=1Mlog2det(𝕀MR+P¯Υopt,kΥopt,kℋσk2(μ¯ℳ,kopt)),subscript𝐶ℳ𝐇^𝐇1𝑀superscriptsubscript𝑘1𝑀subscript2detsubscript𝕀subscript𝑀𝑅¯𝑃subscriptΥopt𝑘superscriptsubscriptΥopt𝑘ℋsuperscriptsubscript𝜎𝑘2subscriptsuperscript¯𝜇optℳ𝑘C_{\mathcal{M}}({\mathbf{H}},\widehat{{\mathbf{H}}})=\frac{1}{M}\sum\limits_{k=1}^{M}\log_{2}\textrm{det}\left(\mathbb{I}_{M_{R}}+\frac{\bar{P}\Upsilon_{\textrm{opt},k}\Upsilon_{\textrm{opt},k}^{\mathcal{H}}}{\sigma_{k}^{2}(\underline{\mu}^{\textrm{opt}}_{\mathcal{M},k})}\right),</td>
<td></td>
<td>(18)</td>
</tr>
</tbody>
</table>
where $`{\sigma_{k}^{2}\hspace{0pt}\left( {\underset{¯}{\mu}}_{\mathcal{M},k}^{\text{opt}} \right)} = {{\frac{\overline{P}}{M_{T}}\hspace{0pt}\left( {\left\| \Lambda_{k} \right\|^{2} - \left\| D_{\mu,k} \right\|^{2}} \right)} + \sigma_{z,k}^{2}}`$,
$`\Upsilon_{\text{opt},k} = {\mathbf{U}_{k}\hspace{0pt}D_{\mu_{\mathcal{M},k}^{\text{opt}}}\hspace{0pt}\mathbf{V}_{k}^{\mathcal{H}}}`$ and the vector containing the diagonal elements of the optimal solution $`D_{\mu_{\mathcal{M},k}^{\text{opt}}}`$ is given by
<table>
<tbody>
<tr>
<td></td>
<td>μ¯ℳ,kopt={(bk‖𝐡~k‖−|ak|)𝐡~kif bk≥0,0¯otherwise.subscriptsuperscript¯𝜇optℳ𝑘casessubscript𝑏𝑘normsubscript~𝐡𝑘subscript𝑎𝑘subscript~𝐡𝑘if bk≥0,¯0otherwise\underline{\mu}^{\textrm{opt}}_{\mathcal{M},k}=\left\{\begin{array}[]{ll}\displaystyle{\left(\frac{\sqrt{b_{k}}}{\|\mathbf{\tilde{h}}_{k}\|}-|a_{k}|\right)\mathbf{\widetilde{h}}_{k}}&\,\,\textrm{if $b_{k}\geq 0$,}\\ \underline{0}&\,\,\textrm{otherwise}.\end{array}\right.</td>
<td></td>
<td>(19)</td>
</tr>
</tbody>
</table>
Similarly, we can compute the achievable rates associated to the mismatched ML decoder (<a href="#S3.E7" style="font-size:90%;" title="In 3.1 Mismatched ML Detection ‣ 3 Maximum-Likelihood Detection in the Presence of Channel Estimation Errors ‣ MIMO-OFDM Optimal Decoding And Achievable Information Rates Under Imperfect Channel Estimation">7</a>). This is given by replacing the vector $`{\underset{¯}{\mu}}_{\mathcal{M},k}^{\text{opt}}`$ by
<table>
<tbody>
<tr>
<td></td>
<td>μ¯ML,kopt=ℝe{Tr(Λkℋ𝐡~k)}‖𝐡~k‖2𝐡~k,subscriptsuperscript¯𝜇optML𝑘ℝ𝑒TrsuperscriptsubscriptΛ𝑘ℋsubscript~𝐡𝑘superscriptnormsubscript~𝐡𝑘2subscript~𝐡𝑘\underline{\mu}^{\textrm{opt}}_{\textrm{ML},k}=\frac{\mathbb{R}e\{{\rm Tr}(\Lambda_{k}^{\mathcal{H}}\tilde{\mathbf{h}}_{k})\}}{\|\tilde{\mathbf{h}}_{k}\|^{2}}\tilde{\mathbf{h}}_{k},</td>
<td></td>
<td>(20)</td>
</tr>
</tbody>
</table>
in equation (<a href="#S5.E18" style="font-size:90%;" title="In 5.1 Achievable Rates Associated to the Improved Decoder ‣ 5 Achievable Rates of MIMO-OFDM Under Channel Estimation Errors ‣ MIMO-OFDM Optimal Decoding And Achievable Information Rates Under Imperfect Channel Estimation">18</a>).
### 5.2 Outage Rates Evaluation
Given any pair of matrices $`\left( \mathbf{H},\hat{\mathbf{H}} \right)`$, the expression (<a href="#S5.E18" style="font-size:90%;" title="In 5.1 Achievable Rates Associated to the Improved Decoder ‣ 5 Achievable Rates of MIMO-OFDM Under Channel Estimation Errors ‣ MIMO-OFDM Optimal Decoding And Achievable Information Rates Under Imperfect Channel Estimation">18</a>) provides the instantaneous achievable rates associated to a receiver using the decoding rule (<a href="#S3.E10" style="font-size:90%;" title="In 3.2 Improved ML Detection for imperfect CSIR ‣ 3 Maximum-Likelihood Detection in the Presence of Channel Estimation Errors ‣ MIMO-OFDM Optimal Decoding And Achievable Information Rates Under Imperfect Channel Estimation">10</a>). The outage probability associated to an outage rate $`R \geq 0`$ is defined as
<table>
<tbody>
<tr>
<td></td>
<td>Pℳout(R,𝐇^)=Pr𝐇|𝐇^(𝐇∈Λℳ(R,𝐇^)|𝐇^),superscriptsubscript𝑃ℳout𝑅^𝐇subscriptPrconditional𝐇^𝐇𝐇conditionalsubscriptΛℳ𝑅^𝐇^𝐇P_{{}_{\mathcal{M}}}^{\mathrm{out}}(R,\widehat{\mathbf{H}})=\mathrm{Pr}_{\mathbf{H}|\widehat{\mathbf{H}}}\big{(}\mathbf{H}\in\Lambda_{{}_{\mathcal{M}}}(R,\widehat{\mathbf{H}})|\widehat{\mathbf{H}}\big{)},</td>
<td></td>
</tr>
</tbody>
</table>
with $`{\Lambda_{_{\mathcal{M}}}\hspace{0pt}\left( R,\hat{\mathbf{H}} \right)} = \left\{ \mathbf{H}:{{C_{_{\mathcal{M}}}\hspace{0pt}\left( \mathbf{H},\hat{\mathbf{H}} \right)} < R} \right\}`$. Using this, the outage rate of the improved decoder for an outage probability $`\gamma`$ is
<table>
<tbody>
<tr>
<td></td>
<td>Cℳout(γ,𝐇^)=supR{R≥0:Pℳout(R,𝐇^)≤γ}superscriptsubscript𝐶ℳout𝛾^𝐇subscriptsupremum𝑅conditional-set𝑅0superscriptsubscript𝑃ℳout𝑅^𝐇𝛾C_{\mathcal{M}}^{\mathrm{out}}(\gamma,\widehat{\mathbf{H}})=\sup_{R}\big{\{}R\geq 0:P_{{}_{\mathcal{M}}}^{\mathrm{out}}(R,\widehat{\mathbf{H}})\leq\gamma\big{\}}</td>
<td></td>
<td>(21)</td>
</tr>
</tbody>
</table>
and $`C_{\text{ML}}^{out}\hspace{0pt}\left( \gamma,\hat{\mathbf{H}} \right)`$ (for mismatched ML decoder), is given by replacing $`\Lambda_{_{\mathcal{M}}}\hspace{0pt}\left( R,\hat{\mathbf{H}} \right)`$ by $`{\Lambda_{\text{ML}}\hspace{0pt}\left( R,\hat{\mathbf{H}} \right)} = \left\{ \mathbf{H}:{{C_{\text{ML}}\hspace{0pt}\left( \mathbf{H},\hat{\mathbf{H}} \right)} < R} \right\}`$ in equation (<a href="#S5.E21" style="font-size:90%;" title="In 5.2 Outage Rates Evaluation ‣ 5 Achievable Rates of MIMO-OFDM Under Channel Estimation Errors ‣ MIMO-OFDM Optimal Decoding And Achievable Information Rates Under Imperfect Channel Estimation">21</a>). Since these outage rates still depend on the random channel estimate $`\hat{\mathbf{H}}`$, we will consider the expected outage rates over all channel estimates as
<table>
<tbody>
<tr>
<td></td>
<td>C¯ℳout(γ)=𝔼𝐇^[Cℳout(γ,𝐇^)].superscriptsubscript¯𝐶ℳout𝛾subscript𝔼^𝐇delimited-[]superscriptsubscript𝐶ℳout𝛾^𝐇\overline{C}_{\mathcal{M}}^{\;\mathrm{out}}(\gamma)=\mathbb{E}_{\widehat{\mathbf{H}}}\big{[}C_{{}_{\mathcal{M}}}^{\mathrm{out}}(\gamma,\widehat{\mathbf{H}})\big{]}.</td>
<td></td>
<td>(22)</td>
</tr>
</tbody>
</table>
The achievable rates (<a href="#S5.E22" style="font-size:90%;" title="In 5.2 Outage Rates Evaluation ‣ 5 Achievable Rates of MIMO-OFDM Under Channel Estimation Errors ‣ MIMO-OFDM Optimal Decoding And Achievable Information Rates Under Imperfect Channel Estimation">22</a>) are upper bounded by the outage rates provided by a theoretical decoder (i.e. the best decoder in the presence of CEE). In our case, this capacity is given by
<table>
<tbody>
<tr>
<td></td>
<td>C¯𝒢out(γ)=𝔼𝐇^[C𝒢out(γ,𝐇^)],superscriptsubscript¯𝐶𝒢out𝛾subscript𝔼^𝐇delimited-[]superscriptsubscript𝐶𝒢out𝛾^𝐇\overline{C}_{{}_{\mathcal{G}}}^{\;\mathrm{out}}(\gamma)=\mathbb{E}_{\widehat{\mathbf{H}}}\big{[}C_{{}_{\mathcal{G}}}^{\mathrm{out}}(\gamma,\widehat{\mathbf{H}})\big{]},</td>
<td></td>
<td>(23)</td>
</tr>
</tbody>
</table>
where the outage rates $`C_{_{\mathcal{G}}}^{out}\hspace{0pt}\left( \gamma,\hat{\mathbf{H}} \right)`$ are computed using
<table>
<tbody>
<tr>
<td></td>
<td>C𝒢(𝐇,𝐇^)=1M∑k=1Mlog2det(𝕀MR+P¯𝐇k𝐇kℋσz2),subscript𝐶𝒢𝐇^𝐇1𝑀superscriptsubscript𝑘1𝑀subscript2subscript𝕀subscript𝑀𝑅¯𝑃subscript𝐇𝑘superscriptsubscript𝐇𝑘ℋsubscriptsuperscript𝜎2𝑧C_{{}_{\mathcal{G}}}(\mathbf{H},\widehat{\mathbf{H}})=\frac{1}{M}\sum\limits_{k=1}^{M}\log_{2}\det\left({\mathbb{I}}_{M_{R}}+\frac{\bar{P}\;{\mathbf{H}}_{k}{\mathbf{H}}_{k}^{\mathcal{H}}}{\sigma^{2}_{z}}\right),</td>
<td></td>
<td>(24)</td>
</tr>
</tbody>
</table>
and $`\mathbf{H}_{k}`$ are random channels drawn from the posterior distribution of (<a href="#S2.E5" style="font-size:90%;" title="In 2 System Model and Channel Estimation ‣ MIMO-OFDM Optimal Decoding And Achievable Information Rates Under Imperfect Channel Estimation">5</a>).
## 6 Numerical Results
Next, the performance of the improved decoder is measured in terms of bit error rate (BER) and achievable outage rates. The binary information data are encoded by a rate $`\left. 1/2 \right.`$ NRNSC code with constraint length 3 defined in octal form by (5,7). Throughout the simulations, each frame is assumed to consists of one OFDM symbol with 50 subcarriers belonging to a 16-QAM constellation with Gray labeling. The interleaver is a pseudo-random one operating over the entire frame with size $`M \cdot M_{T} \cdot {\log_{2}\left( M_{c} \right)}`$ bits. For each transmitted frame, a different realization of a Rayleigh distributed channel has been drawn and remains constant during the whole frame. Besides, it is assumed that the average pilot symbol energy is equal to the average data symbol energy. Moreover, the number of decoding iterations are set to 4.
Fig. <a href="#S7.F3" style="font-size:90%;" title="Figure 3 ‣ MIMO-OFDM Optimal Decoding And Achievable Information Rates Under Imperfect Channel Estimation">3</a> depicts the BER performance over a $`2 \times 2`$ MIMO-OFDM channel estimated by $`N \in \left\{ 2,4,8 \right\}`$ pilot symbols per frame. As observed, the increase in the required $`\left. E_{b}/N_{0} \right.`$ caused by CEE is an important effect of imperfect CSIR in the case of mismatched ML decoding. The figure shows that the SNR to obtain a BER of $`10^{- 5}`$ with $`N = 2`$ pilots is reduced by about $`1.3`$ dB if the improved decoder is used instead of the mismatched decoder. We also notice that the performance loss of the mismatched receiver with respect to the derived receiver becomes insignificant for $`N \geq 8`$. This can be explained from the expression of the metric (<a href="#S3.Ex2" style="font-size:90%;" title="3.2 Improved ML Detection for imperfect CSIR ‣ 3 Maximum-Likelihood Detection in the Presence of Channel Estimation Errors ‣ MIMO-OFDM Optimal Decoding And Achievable Information Rates Under Imperfect Channel Estimation">3.2</a>), where we note that by increasing the number of pilot symbols, this expression tends to the classical Euclidean distance metric. However, the proposed decoder outperforms the mismatched decoder especially when few numbers of pilot symbols are dedicated for channel estimation.
Fig. <a href="#S7.F4" style="font-size:90%;" title="Figure 4 ‣ MIMO-OFDM Optimal Decoding And Achievable Information Rates Under Imperfect Channel Estimation">4</a> shows the expected outage rates (in bits per channel use) corresponding to the transmission of one OFDM symbol with $`M = 16`$ subcarriers and $`M_{T} = M_{R} = 2`$ antennas, achieved by adopting mismatched ML decoding and the improved decoder (expression (<a href="#S5.E22" style="font-size:90%;" title="In 5.2 Outage Rates Evaluation ‣ 5 Achievable Rates of MIMO-OFDM Under Channel Estimation Errors ‣ MIMO-OFDM Optimal Decoding And Achievable Information Rates Under Imperfect Channel Estimation">22</a>)). For comparison, we also display the upper bounds on these achievable outage rates (expression (<a href="#S5.E23" style="font-size:90%;" title="In 5.2 Outage Rates Evaluation ‣ 5 Achievable Rates of MIMO-OFDM Under Channel Estimation Errors ‣ MIMO-OFDM Optimal Decoding And Achievable Information Rates Under Imperfect Channel Estimation">23</a>)) and the ergodic capacity. $`N = 2`$ pilot symbols are sent per frame for CSIR acquisition and the outage probability has been fixed to $`\gamma = 0.01`$. At a mean outage rate of $`8`$ bits, we note that the achievable rate of the mismatched decoder is about $`5`$ dB of SNR far from the capacity provided by the theoretical decoder. As observed, by using the improved decoder, higher rates are obtained for any SNR values and the aforementioned SNR gap is reduced by about $`1.8`$ dB.
Similarly, Fig. <a href="#S7.F5" style="font-size:90%;" title="Figure 5 ‣ MIMO-OFDM Optimal Decoding And Achievable Information Rates Under Imperfect Channel Estimation">5</a> compares the outage rates in the case of a $`4 \times 4`$ MIMO-OFDM channel estimated by $`N = 4`$ training symbols. Again, it can be observed that the proposed decoder achieves higher rates than the mismatched decoder. However, we note that the increase in the SNR (at a given mean outage rate) induced by using the mismatched decoder rather than the improved decoder, is less than that obtained for a $`2 \times 2`$ MIMO-OFDM channel (see Fig. <a href="#S7.F4" style="font-size:90%;" title="Figure 4 ‣ MIMO-OFDM Optimal Decoding And Achievable Information Rates Under Imperfect Channel Estimation">4</a>). This can be explained by noting that when $`M_{T} = 4`$, we must send $`N \geq M_{T} = 4`$ pilot symbols per frame which yields a more accurate estimate of the channel and consequently bring closer the performance of the two decoders. This observation is in consistence with those presented in \[[2](#bib.bib2)\] where it is reported that the performance degradation due to imperfect channel estimation can be reduced by increasing the number of antennas.
## 7 Conclusion
A Bayesian approach for the design of a decoding method that is robust to channel estimation errors was presented. The robustness of our method comes from the averaging of the decoding rule, that would be used if the channel were perfectly known, over all channel estimation errors. We also derived the expression of the achievable rates associated to our proposed decoder and compared it to that provided by the classical mismatched decoding approach. As a practical application, the proposed decoder was exploited for iterative BICM decoding of MIMO-OFDM under imperfect channel knowledge. Simulation results showed that, without introducing any additional complexity, the proposed decoder outperforms the classical mismatched approach in terms of BER and achievable outage rates for short training sequences. Although our proposed method outperforms mismatched decoding, the derivation of a practical decoder achieving the maximum outage rate under imperfect channel estimation is still an open problem.
## References
- \[1\]
H. Bolcskei, D. Gesbert, and A. J. Paulraj,
“On the capacity of OFDM-based spatial multiplexing systems,”
IEEE Trans. Comun., pp. 225–234, Feb. 2002.
- \[2\]
P. Garg, R. K. Mallik, and H. M. Gupta,
“Performance analysis of space-time coding with imperfect channel estimation,”
IEEE Trans. Wireless Commun., vol. 4, pp. 257–265, Jan. 2005.
- \[3\]
Pablo Piantanida, Gerald Matz, and Pierre Duhamel,
“Estimation-induced outage capacity of Ricean channels,”
in Proc. SPAWC, July 2006.
- \[4\]
S. Sadough, P. Piantanida, and P. Duhamel,
“Achievable outage rates with improved decoding of BICM Multiband OFDM under channel estimation errors,”
in Asilomar Conf. on Signals, systems and computers, Oct. 2006.
- \[5\]
Giorgio Taricco and Ezio Biglieri,
“Space-time decoding with imperfect channel estimation,”
IEEE Transaction on Wireless Communications, vol. 4, no. 4, pp. 1874, 1888 2005.
- \[6\]
Neri Merhav, Gideon Kaplan, Amos Lapidoth, and Sholmo Shamai,
“On information rates for mismatched decoders,”
IEEE Transactions on Information Theory, vol. 40, pp. 1953–1967, November 1994.
- \[7\]
J. J. Boutros, F. Boixadera, and C. Lamy,
“Bit-interleaved coded modulations for multiple-input multiple-output channels,”
in Int. Symp. on Spread Spectrum Techniques and Applications, Sept. 2000, pp. 123–126.
- \[8\]
L. Bahl, J. Cocke, F. Jelinek, and J. Raviv,
“Optimal decoding of linear codes for minimizing symbol error rate,”
IEEE Trans. on Inf. Theory, pp. 284–287, March 1974.
[Refer to caption](/html/0708.1416/assets/x3.png)
Fig. 3:
BER performance of the proposed and mismatched decoders in the case of a $`2 \times 2`$ MIMO-OFDM Rayleigh fading channel with $`M = 50`$ subcarriers for training sequence length $`N \in {\{ 2,4,8\}}`$.
[Refer to caption](/html/0708.1416/assets/x4.png)
Fig. 4:
Expected outage rates of a $`2 \times 2`$ MIMO-OFDM system with $`M = 16`$ subcarriers versus SNR for $`N = 2`$ pilots.
[Refer to caption](/html/0708.1416/assets/x5.png)
Fig. 5:
Expected outage rates of a $`4 \times 4`$ MIMO-OFDM system with $`M = 16`$ subcarriers versus SNR for $`N = 4`$ pilots. |
0709.0965 | cs.NI | cyber | # The Induced Bounded-Degree Subgraph Problem and Stream Control in MIMO Networks
Harish Sethu
Department of Electrical and Computer Engineering
Drexel University, Philadelphia, PA 19104-2875
Email: sethu@drexel.edu
###### Abstract
In this report, we consider maximal solutions to the induced bounded-degree subgraph problem and relate it to issues concerning stream control in multiple-input multiple-output (MIMO) networks. We present a new distributed algorithm that completes in logarithmic time with high probability and is guaranteed to complete in linear time. We conclude the report with simulation results that address the effectiveness of stream control and the relative impact of receiver overloading and flexible interference suppression.
<sup>1</sup>
<sup>1</sup>footnotetext: This work was supported in part by NSF Award CNS-0322797.
## I Introduction
Multi-hop wireless sensor networks and mobile ad hoc networks have been suggested for a variety of applications in both military and civilian environments. In such networks, the aggregate traffic-carrying capacity is determined by the number of concurrent transmissions that can be supported without interference from other transmissions. If we consider each transmission link in a network as a vertex in a graph with an edge connecting each pair of vertices corresponding to interfering links, the problem of finding the maximum aggregate capacity reduces to the problem of finding the largest independent set in the graph \[[1](#bib.bib1), [2](#bib.bib2)\]. In this report, we pose a generalized version of the independent set problem inspired by the technology of multiple-input multiple-output (MIMO) enabled by smart antennas \[[3](#bib.bib3)\].
In a MIMO link, multiple antenna elements at both ends of a communication link allow simultaneous transmission of multiple data streams, thus achieving a higher capacity without increased bandwidth or power requirements. With $`q`$ degrees of freedom (number of antenna elements), a MIMO receiver can isolate and decipher up to $`q`$ streams that reach the receiver. A MIMO link, however, may choose only the strongest streams (channel modes) and transmit on fewer than $`q`$ streams. Such stream control, illustrated in Fig. [1](#S2.F1 "Figure 1 ‣ II Problem Statement ‣ The Induced Bounded-Degree Subgraph Problem and Stream Control in MIMO Networks"), is known to allow multiple interfering links operating simultaneously to achieve a better overall throughput than if each link transmits on all available streams \[[4](#bib.bib4), [5](#bib.bib5)\]. As identified in \[[6](#bib.bib6)\], there are at least two aspects related to stream control that influence, in opposite directions, the potential capacity of a MIMO network. The phenomenon of receiver overloading offsets the gains from stream control because an active receiver in a contention zone reduces the number of potential transmitters in all the contention zones that it belongs to. This can reduce the ability to exploit spatial diversity since more links are active when stream control is employed and therefore, more active receivers may be overloaded. On the other hand, stream control allows flexible interference suppression because the number of degrees of freedom required to suppress interference at a node may be smaller than the number of interfering streams. For example, it may be possible to decode two desired streams using up two degrees of freedom while suppressing two other undesirable interfering streams using up only one additional degree of freedom, thus using a total of only three (instead of four) degrees of freedom.
In this report, we seek a graph-theoretic approach to gain some insight into the relative influence of receiver overloading and flexible interference suppression on the effectiveness of stream control in a MIMO network. We relate the issue to the problem of finding a maximal induced bounded-degree subgraph of the contention graph of a MIMO network.
## II Problem Statement
[Refer to caption](/html/0709.0965/assets/x1.png)
Figure 1: Two interfering MIMO links, each comprised of two streams, can achieve a higher data rate if both links transmit simultaneously on their stronger streams instead of taking turns and transmitting during different slots on all streams. Solid line indicates an active stream while a dashed line indicates an unused stream.
Consider a MIMO network in which each node possesses $`q`$ antenna elements (therefore, $`q`$ degrees of freedom) and is capable of transmitting on $`q`$ different streams to a neighboring node. Consider a graph $`G = {(V,E)}`$ in which each vertex represents a potential stream in a MIMO link. Two streams that interfere with each other are connected by an edge in the graph. Thus, the streams comprising a single MIMO link make up a $`q`$-vertex clique. Fig. [2](#S3.F2 "Figure 2 ‣ III The Distributed Algorithm ‣ The Induced Bounded-Degree Subgraph Problem and Stream Control in MIMO Networks") shows a simple MIMO network with five physical nodes. Assuming two streams per MIMO link, Fig. [2](#S3.F2 "Figure 2 ‣ III The Distributed Algorithm ‣ The Induced Bounded-Degree Subgraph Problem and Stream Control in MIMO Networks") shows the corresponding graph representing the interference pattern between the different streams. Our goal is to maximize the number of active streams while ensuring that there are no more than a certain number, say $`k`$, of streams interfering with any given active stream. For example, in the case in which it takes exactly one degree of freedom at an active receiver to suppress an undesired interfering stream (i.e., no flexible interference suppression), an active stream should have no more than $`k = {q - 1}`$ other streams interfering at the intended receiver. The problem now becomes that of finding the largest subset of vertices $`V' \subset V`$ such that for any vertex $`v \in V'`$, there exist no more than $`k`$ other vertices in $`V'`$ connected to $`v`$ by an edge in $`G`$. Thus, this becomes the maximum induced bounded-degree-$`k`$ subgraph (IBDS$`(k)`$) problem which is known to be NP-complete for general graphs \[[7](#bib.bib7)\], although linear time solutions have been discovered for certain types of graphs \[[8](#bib.bib8)\]. The well-known maximum independent set problem is an instance of maximum IBDS($`k`$) for $`k = 0`$. We hasten to add that maximum IBDS($`k`$) is different from another better-known and well-studied generalization of the maximum independent set problem, the $`k`$-independent set problem (related to the $`k`$-coloring of graphs) \[[9](#bib.bib9)\].
We define a maximal IBDS($`k`$) as an induced bounded-degree-$`k`$ subgraph to which no additional vertex can be added. We also define two simple variations of the maximal IBDS($`k`$) problem for relevance in practical MIMO networks. Streams corresponding to the same MIMO link (i.e., between the same pair of nodes) are said to belong to the same family. Two potential streams with at least one physical node in common are said to belong to the same superfamily. In MIMO networks, practical considerations often impose that a node does not use two streams corresponding to different MIMO links simultaneously. We define the maximal IBDS-R$`(k)`$ problem as the maximal IBDS($`k`$) problem with the additional restriction that for any two vertices $`{v_{1},v_{2}} \in V'`$, if $`v_{1}`$ and $`v_{2}`$ belong to the same superfamily, then they belong to the same family. A second variation is motivated by the fact that exploiting flexible interference suppression requires that each MIMO link operate only a subset (say, no more than $`g`$) of the maximum number of streams possible and employ the remaining available degrees of freedom for interference suppression. We define the maximal IBDS-R$`(k,g)`$ problem as the maximal IBDS-R$`(k)`$ problem with the restriction that no more than $`g`$ vertices from the same family are selected into the maximal subgraph.
## III The Distributed Algorithm
[Refer to caption](/html/0709.0965/assets/x2.png)
[Refer to caption](/html/0709.0965/assets/x3.png)
Figure 2: Example illustrating (a) A simple network of MIMO nodes, and (b) the corresponding graph on which the distributed induced bounded-degree subgraph algorithm is run.
Fig. [3](#S3.F3 "Figure 3 ‣ III The Distributed Algorithm ‣ The Induced Bounded-Degree Subgraph Problem and Stream Control in MIMO Networks") presents a pseudo-code description of one round of DML-IBDS($`k`$), a new distributed randomized algorithm for the maximal IBDS($`k`$) problem (DML stands for distributed maximal). Each vertex $`v`$ holds four variables, $`r\hspace{0pt}{(v)}`$, $`b\hspace{0pt}{(v)}`$, $`\hat{b}\hspace{0pt}{(v)}`$ and $`a\hspace{0pt}{(v)}`$. In the beginning before the first round, each vertex $`v`$ computes a unique random integer $`r\hspace{0pt}{(v)}`$. As in most distributed algorithms for networks, we assume that the vertices possess unique ids and therefore, the uniqueness of the random numbers is readily ensured by appending the vertex id to the generated pseudo-random number. $`r\hspace{0pt}{(v)}`$ is updated in subsequent rounds and is not necessarily a unique number except at the beginning of the first round. $`b\hspace{0pt}{(v)}`$ is initially set to 0 and indicates whether or not vertex $`v`$, in some round, has $`r\hspace{0pt}{(v)}`$ smaller than $`r\hspace{0pt}{(u)}`$ for all neighbors $`u`$ (this is what qualifies vertex $`v`$ for inclusion in the subgraph). $`\hat{b}`$ is initially set to $`- 1`$. When $`\hat{b}`$ is 1, it indicates that the vertex is chosen for inclusion in the subgraph and when it is 0, it indicates that the vertex has been eliminated from consideration. $`a\hspace{0pt}{(v)}`$ is initially set to $`k + 1`$ and represents the maximum number of additional vertices that may be chosen for the subgraph out of the set comprising of $`v`$ and its neighbors.
01:
Initialization (at vertex $`v`$):
02:
$`{r\hspace{0pt}(v)} =`$ pseudo-random integer
03:
$`{b\hspace{0pt}(v)} = 0`$
04:
$`{\hat{b}\hspace{0pt}(v)} = {- 1}`$
05:
$`{a\hspace{0pt}(v)} = {k + 1}`$
06:
One round of DML-IBDS($`k`$) (at vertex $`v`$):
07:
if $`{a\hspace{0pt}(v)} > 0`$:
08:
Send $`r\hspace{0pt}(v)`$ to neighbors
09:
Receive $`r\hspace{0pt}(u)`$ from each neighbor $`u`$
10:
if $`{r\hspace{0pt}(v)} \leq {r\hspace{0pt}(u)}`$ for every neighbor $`u`$:
11:
$`{b\hspace{0pt}(v)} = 1`$
12:
Send $`b\hspace{0pt}(v)`$ to neighbors
13:
Receive $`b\hspace{0pt}(u)`$ from each neighbor $`u`$
14:
if $`{b\hspace{0pt}(v)} = 1`$:
15:
if $`\nexists\hspace{0pt}u`$ such that $`{b\hspace{0pt}(u)} = 0`$ and $`{r\hspace{0pt}(u)} = {r\hspace{0pt}(v)}`$:
16:
$`{r\hspace{0pt}(v)} = {\min\left\{ {r\hspace{0pt}(u)\hspace{0pt}\left| {r\hspace{0pt}(u)}> \right.\hspace{0pt}r\hspace{0pt}(v)} \right\}}`$
17:
if $`{\hat{b}\hspace{0pt}(v)} = {- 1}`$ and $`{b\hspace{0pt}(v)} = 1`$:
18:
$`{\hat{b}\hspace{0pt}(v)} = 1`$
19:
if $`{b\hspace{0pt}(u)} = 1`$ for every neighbor $`u`$
20:
Halt
21:
Decrement $`a\hspace{0pt}(v)`$
22:
Decrement $`a\hspace{0pt}(u)`$ at each neighbor $`u`$
23:
if $`{a\hspace{0pt}(v)} \leq 0`$:
24:
if $`{\hat{b}\hspace{0pt}(v)} = 1`$:
25:
for each neighbor $`u`$:
26:
$`{a\hspace{0pt}(u)} = 0`$
27:
if $`{\hat{b}\hspace{0pt}(u)} = {- 1}`$
28:
$`{\hat{b}\hspace{0pt}(u)} = 0`$
29:
Halt
Figure 3: Pseudo-code description of DML-IBDS($`k`$). A vertex $`v`$ is selected into the subgraph if $`\hat{b} = 1`$.
Each vertex $`v`$ sends its random number, $`r\hspace{0pt}{(v)}`$, to its neighbors and receives numbers from its neighbors (lines 08–09). If none of $`v`$’s neighbors carries a smaller value of $`r`$, then vertex $`v`$ sets its $`b\hspace{0pt}{(v)}`$ to 1 (lines 10–11). Neighbors communicate their value of $`b`$ to each other (lines 12–13). A vertex that has its $`b`$ value set to 1 will now consider updating its $`r`$ value. If $`v`$ has no neighbors with the same $`r`$ value as itself or if all neighbors of $`v`$ with the same $`r`$ value are already chosen for inclusion into the subgraph (i.e., $`b = 1`$), then $`v`$ updates its $`r`$ value to the smallest among the received values but larger than $`r\hspace{0pt}{(v)}`$ (lines 15–16). This step ensures that no more than one of the neighbors of a vertex with $`b = 1`$ will be selected into the subgraph in the same round (thus guaranteeing that $`a\hspace{0pt}{(v)}`$ does not suddenly reach below zero; a vertex $`v`$ with $`{b\hspace{0pt}{(v)}} = 1`$ that does not yet have $`{a\hspace{0pt}{(v)}} = 0`$ will always possess an $`r\hspace{0pt}{(v)}`$ equal to the smallest of its neighbors, and therefore, only its smallest neighbor will potentially be chosen into the subgraph.)
######
Theorem 1
The DML-IBDS($`k`$) algorithm finds a maximal subgraph with degree bound $`k`$ in $`O\hspace{0pt}{(n)}`$ rounds where $`n`$ is the number of vertices in the graph.
###### Proof:
At the beginning of a round, consider the vertex $`v`$ with the smallest value of $`r\hspace{0pt}{(v)}`$ such that $`{\hat{b}\hspace{0pt}{(v)}} = {- 1}`$. By lines 15–16, any vertex $`u`$ with $`{\hat{b}\hspace{0pt}{(u)}} = 1`$ that is still participating in the rounds will have $`r\hspace{0pt}{(u)}`$ no smaller than $`r\hspace{0pt}{(v)}`$. Therefore, vertex $`v`$ will have $`\hat{b}`$ set to 1 by the end of the round. Thus, in each round, at least one additional vertex has its $`\hat{b}`$ set to one, and therefore, the algorithm completes in $`O\hspace{0pt}{(n)}`$ rounds.
A vertex $`v`$ reduces $`a\hspace{0pt}{(v)}`$ by one whenever itself or one of its neighbors is chosen into the subgraph (lines 21–22). However, $`a\hspace{0pt}{(v)}`$ is never allowed to go below zero for vertices that have $`\hat{b}`$ set to 1 (i.e., vertices corresponding to streams chosen for the subgraph). Since a vertex halts when $`{a\hspace{0pt}{(v)}} = 0`$ (line 29), the induced subgraph will satisfy the desired degree bound. Also, it is a maximal subgraph because a vertex halts only when $`{a\hspace{0pt}{(v)}} \leq 0`$ (line 29) or when the vertex along with all its neighbors has been chosen for inclusion into the subgraph (line 20). ∎
In real networks, after a topology control algorithm such as RNG has been executed, each physical node would have a bounded number of MIMO links. The corresponding graph of interference pattern among the streams, therefore, has a degree bound as well. The probability that any given vertex $`v`$ ends up with an $`r\hspace{0pt}{(v)}`$ no smaller than that of any of its neighbors is less than or equal to $`1/{({1 + {d\hspace{0pt}e\hspace{0pt}g_{G}\hspace{0pt}{(v)}}})}`$. Given a bound on the degree, each node has a probability greater than some constant of having its $`a\hspace{0pt}{(v)}`$ reduced by one. Since each of the $`n`$ vertices begins with $`a = k`$ degrees of freedom, even though the worst-case number of rounds is $`O\hspace{0pt}{(n)}`$, the actual number of rounds is $`O\hspace{0pt}{({\log{n\hspace{0pt}k}})}`$ with high probability (i.e., with probability greater than or equal to $`1 - {1/{({n\hspace{0pt}k})}^{c}}`$ for any $`n \geq n_{0}`$, $`k > k_{0}`$ for some positive constants $`c,k_{0}`$ and $`n_{0}`$).
Distributed solutions to variations of maximal IBDS($`k`$) described in the previous section are only trivially different and are denoted in this report by DML-IBDS-R$`(k)`$ and DML-IBDS-R$`(k,g)`$.
##
IV
Results and Concluding Remarks
[Refer to caption](/html/0709.0965/assets/x4.png)
Figure 4:
Size of the subgraph with DML-IBDS($`k`$) and DML-IBDS-R$`(k)`$ plotted against $`k + 1`$. Note that in the absence of flexible interference suppression, $`k + 1`$ may be thought of as the number of antenna elements per node.
Fig. [4](#S4.F4 "Figure 4 ‣ IV Results and Concluding Remarks ‣ The Induced Bounded-Degree Subgraph Problem and Stream Control in MIMO Networks") presents simulation results on the performance of DML-IBDS($`k`$) and DML-IBDS-R$`(k)`$ as measured by the number of vertices in the degree-bounded subgraph that is generated. The networks used in our experiments consist of nodes located randomly in a unit square area and with the IBDS algorithms applied after the application of the DRNG (Directed Relative Neighborhood Graph) topology control algorithm \[[10](#bib.bib10)\]. We choose the DRNG protocol because it is one that can be employed in a multipath environment where a node cannot readily deduce the distance and direction of its neighbors. Each data point in this report represents an average of over 25 different randomly generated networks.
As can be seen from Fig. [4](#S4.F4 "Figure 4 ‣ IV Results and Concluding Remarks ‣ The Induced Bounded-Degree Subgraph Problem and Stream Control in MIMO Networks"), DML-IBDS-R$`(k)`$ performs slightly better than that for the unrestricted version, DML-IBDS($`k`$), indicating that the negative impact of receiver overloading is reduced when the active streams at a node all belong to the same MIMO link. This suggests that the synchronization necessary to allow simultaneous use of streams to different neighbors at a node may not yield a higher aggregate throughput unless the stream control gains are high (i.e., if the strongest channel modes are significantly stronger than the weaker channel modes).
Fig. [4](#S4.F4 "Figure 4 ‣ IV Results and Concluding Remarks ‣ The Induced Bounded-Degree Subgraph Problem and Stream Control in MIMO Networks") also shows that the number of streams used (size of induced subgraph) does not increase linearly as the number of antenna elements increases. This is obviously sub-optimal to a trivial solution that involves no stream control: generate a maximal independent set, such as by running DML-IBDS(0), and at all the MIMO links chosen to transmit on a stream just allow all the streams to transmit. Since practical MAC algorithms do not usually approach the capacity corresponding to a maximal independent set, this sublinear increase in the number of active streams does not necessarily suggest that stream control is not desirable in practice. In fact, as shown in \[[6](#bib.bib6)\], stream control does improve performance in real contexts. Our results, however, shed light on the significant negative effect of receiver overloading and help us quantify the desired stream gains needed to offset the effect. For example, in a 500-node network in the absence of flexible interference suppression, the number of streams with one antenna element per node is 115 (subgraph size yielded by DML-IBDS-R$`(0)`$) while that with two antenna elements is 189 (subgraph size yielded by DML-IBDS-R$`(1)`$). This implies that for stream control to increase capacity in a network with two antenna elements per node, in the absence of flexible interference suppression, the average gain due to stream control should be about 22% (2$`\times`$$`{115/189} - 1`$).
[Refer to caption](/html/0709.0965/assets/x5.png)
Figure 5:
Size of subgraph produced by DML-IBDS-R$`(k,g)`$, i.e., when each MIMO link is allowed no more than $`g`$ active streams and no more than $`k + 1`$ total streams (desirable plus undesirable) interfere at any receiver.
Fig. [5](#S4.F5 "Figure 5 ‣ IV Results and Concluding Remarks ‣ The Induced Bounded-Degree Subgraph Problem and Stream Control in MIMO Networks") shows us the gains achievable through flexible interference suppression and permit a quantitative comparison to the negative impact of receiver overloading. The subgraph size with DML-IBDS-R$`(k,g)`$ indicates the number of active streams in the network when the number of active streams per MIMO link is bounded by $`g`$ and the total number of streams (both desirable and the undesirable interfering ones) arriving at a receiver is bounded by $`k + 1`$. For example, for 500-node networks, Figs. [4](#S4.F4 "Figure 4 ‣ IV Results and Concluding Remarks ‣ The Induced Bounded-Degree Subgraph Problem and Stream Control in MIMO Networks") and [5](#S4.F5 "Figure 5 ‣ IV Results and Concluding Remarks ‣ The Induced Bounded-Degree Subgraph Problem and Stream Control in MIMO Networks") show that the subgraph size obtained with DML-IBDS-R$`(1)`$ lies somewhere between those obtained by DML-IBDS-R$`(4,1)`$ and DML-IBDS-R$`(5,1)`$. If we have two antenna elements per node and if only one stream ($`g = 1`$) is used per MIMO link in order to exploit flexible interference suppression, the remaining one available degree of freedom has to be sufficient to suppress up to five additional undesirable streams to exceed the subgraph size that can be obtained by allowing both streams to operate whenever possible. These results suggest that, unless each available degree of freedom can be used to suppress several undesirable interfering streams, it is better to use any available degrees of freedom to active streams instead of flexible interference suppression. These insights offered by our algorithmic results can be used to design better MAC protocols for MIMO networks.
Acknowledgments
The author wishes to thank Kapil R. Dandekar for introducing him to MIMO networks. This work was supported in part by NSF Award CNS-0322797.
## References
- \[1\]
D. Peleg, *Distributed Computing: A Locality-Sensitive Approach*. SIAM, 2000.
- \[2\]
H. Balakrishnan, C. L. Barrett, V. S. A. Kumar, M. V. Marathe, and S. Thite, “The distance-2 matching problem and its relationship to the MAC-layer capacity of ad hoc wireless network,” *IEEE Journal on Selected Areas in Communications*, vol. 22, no. 6, pp. 1069–1079, August 2004.
- \[3\]
S. Haykin and M. Moher, *Modern Wireless Communications*. Prentice Hall, 2005.
- \[4\]
M. F. Demirkol and M. A. Ingram, “Stream control in networks with interfering MIMO links,” in *Proceedings of the IEEE Wireless Communications and Networking Conference*, 2003, pp. 343–348.
- \[5\]
R. S. Blum, J. H. Winters, and N. R. Sollenberger, “On the capacity of cellular systems with MIMO,” *IEEE Communication Letters*, vol. 6, no. 6, pp. 242–244, June 2002.
- \[6\]
K. Sundaresan, R. Sivakumar, M. A. Ingram, and T.-Y. Chang, “Medium access control in ad hoc networks with MIMO links: Optimization considerations and algorithms,” *IEEE Transactions on Mobile Computing*, vol. 3, no. 4, pp. 350–365, October-December 2004.
- \[7\]
J. A. Telle, “Complexity of domination-type problems in graphs,” *Nordic Journal of Computing*, vol. 1, no. 1, pp. 157–171, Spring 1994.
- \[8\]
H. L. Bodlaender and B. de Fluiter, “Reduction algorithms for constructing solutions in graphs with small treewidth,” in *Computing and Combinatorics*, ser. Lecture Notes in Computer Science, vol. 1090. Springer, 1996, pp. 199–208.
- \[9\]
D. B. West, *Introduction to Graph Theory*, 2nd ed. Prentice Hall, 2001.
- \[10\]
N. Li and J. C. Hou, “Localized topology control algorithms for heterogeneous wireless networks,” *IEEE/ACM Transactions on Networking*, vol. 13, no. 6, pp. 1313–1324, December 2005. |
0709.2252 | cs.NI | cyber | # Otiy: Locators Tracking Nodes
Mathias Boc
Anne Fladenmuller
and Marcelo Dias de Amorim
This work has been partially supported by the IST European project WIP under contract 27402 and by the RNRT project Airnet under contract 01205. Laboratoire d’Informatique de Paris 6 (LIP6/CNRS) Université Pierre et Marie Curie – Paris 6 <%7Bboc,fladenmu,amorim%7D@rp.lip6.fr>
\(2007\)
###### Abstract
We propose Otiy, a node-centric location service that limits the impact of location updates generate by mobile nodes in IEEE 802.11-based wireless mesh networks. Existing location services use node identifiers to determine the locator (aka anchor) that is responsible for keeping track of a node’s location. Such a strategy can be inefficient because: (i) identifiers give no clue on the node’s mobility and (ii) locators can be far from the source/destination shortest path, which increases both location delays and bandwidth consumption. To solve these issues, Otiy introduces a new strategy that identifies nodes to play the role of locators based on the likelihood of a destination to be close to these nodes – i.e., locators are identified depending on the mobility pattern of nodes. Otiy relies on the cyclic mobility patterns of nodes and creates a slotted agenda composed of a set of predicted locations, defined according to the past and present patterns of mobility. Correspondent nodes fetch this agenda only once and use it as a reference for identifying which locators are responsible for the node at different points in time. Over a period of about one year, the weekly proportion of nodes having at least 50% of exact location predictions is in average about 75%. This proportion increases by 10% when nodes also consider their closeness to the locator from only what they know about the network.
<sup>†</sup>
<sup>†</sup>conference: CoNext ’2007 Columbia, NY USA
## 1 Introduction
Promoting node mobility in self-organizing wireless networks implies setting up an efficient location management scheme. Indeed, the degree of node mobility impacts the amount of location updates disseminated throughout the network; this problem becomes even more critical in dense areas, where the profusion of signaling messages induces serious contentions and penalizes the overall performance of applications.
There are different techniques to implement location services. In flooding-based approaches, no proactive decision is made; when a source wants to communicate with a destination, it floods the entire network with a lookup message. Although simple, such an approach is very resource consuming and thus inappropriate for wireless networks. In order to solve this problem, some solutions propose to use location anchors – when a source wants to communicate with a destination, it must first ask the anchor about the current location of the node. Anchor-based architectures can be implemented in a centralized or distributed manner \[[2](#bib.bib2), [7](#bib.bib7), [13](#bib.bib13), [15](#bib.bib15)\]. Nevertheless, these solutions share a common pitfall: they have no control on the location of the anchor. This means that a source may have to send a lookup message to a far-away anchor for a destination that is possibly nearby. The consequences are twofold: (i) lookup phase may experience large delays and (ii) lookup messages may travel long distances, reducing the overall capacity of the network. The consequences are the same for the destination node which also have to update its current location to a possibly far-away responsible (or assigned) anchor.
To address the abovementioned problems, we propose
Otiy
<sup>1</sup>
<sup>1</sup>
1
Oti’y is a Creole word meaning “where can I find him?”
, a two-tiered location service that relies on the fact that most nodes present cyclic mobility patterns. Indeed, many works have shown that, in many situations, nodes do show mobility characteristics that can be quite well predicted based on the node’s mobility history \[[1](#bib.bib1), [5](#bib.bib5), [8](#bib.bib8), [16](#bib.bib16), [17](#bib.bib17)\].
Otiy benefits from mobility prediction by decoupling the location service into two tiers as follows:
1. 1\.
Global service. This service determines the anchor of a node, in a similar way to traditional solutions. The difference here is that anchors in Otiy do not store the current location of nodes; instead, it returns an agenda containing for each period of the day the location server (or locator, cf., next bullet) which is most probably the closest to the mobile node. This agenda is available for a reasonable amount of time and thus can be stored by the communicating nodes to prevent frequent access to the global service.
2. 2\.
Local service. Locators are points in the infrastructure that effectively know at a given time where mobile nodes are. They respond to lookup requests and inform about the current position of mobile nodes. To determine which of its locators is currently in charge of storing its location, a mobile node refers to its agenda. As a locator is chosen by a mobile node because of its high probability of being close to it, location updates remain localized.
The idea behind this system is to use the global service from time to time and the local service most of the time – and thus reduce the overhead found in flat solutions. More specifically, an agenda contains a list of pairs $`({{\mathtt{T}\mathtt{i}\mathtt{m}\mathtt{e}}\hspace{0pt}\_\hspace{0pt}{\mathtt{s}\mathtt{l}\mathtt{o}\mathtt{t}}},{\mathtt{L}\mathtt{o}\mathtt{c}\mathtt{a}\mathtt{t}\mathtt{o}\mathtt{r}})`$, where for the time period $`{\mathtt{T}\mathtt{i}\mathtt{m}\mathtt{e}}\hspace{0pt}\_\hspace{0pt}{\mathtt{s}\mathtt{l}\mathtt{o}\mathtt{t}}`$ the location information of the mobile node will be managed by $`\mathtt{L}\mathtt{o}\mathtt{c}\mathtt{a}\mathtt{t}\mathtt{o}\mathtt{r}`$. This agenda remains valid for a certain period, after which it must be renewed at the corresponding $`a\hspace{0pt}n\hspace{0pt}c\hspace{0pt}h\hspace{0pt}o\hspace{0pt}r`$. A source willing to communicate with a destination first fetches the agenda of the destination at its corresponding anchor. Then, during the validity period of the agenda – which can last for weeks depending on the prediction accuracy – the source only contacts locators to obtain the exact location of the destination. The generation of a pertinent agenda to identify the best locator (closest to the mobile node) for each time slot is thus at the heart of Otiy. We argue that mobile nodes don’t necessarily have a positioning system and should therefore determine their mobility pattern based only on topological information, meaning the logs of attachment to the access points.
Although the basic concepts of Otiy can be generalized to different types of self-organizing wireless networks, in this paper we focus on the context of wireless mesh networks (WMN) composed of IEEE 802.11 access points (mesh routers). In this case, mobility is defined as a sequence of access points a node associates to along time. We will see later in this paper that defining exact mobility patterns in a wireless mesh network is a complex task. For instance, misinterpretations of mobility may happen mainly due to ping-pong effects, which are oscillations of associations/disassociations to nearby mesh routers due to changes in medium conditions. In order to address this problem, we introduce a self-organizing scheme that groups nearby mesh routers into clusters from each node standpoint. This clustering scheme masks ping-pong effects and allows reducing the space of possible locations.
We evaluate Otiy under a large population of nodes using real traces of mobility in a campus scenario \[[9](#bib.bib9)\]. We find that nodes, although heterogeneous in nature, do show cyclic behaviors according to their own rules. The analysis covers periods ranging from one month to more than one year and presents results in function of the number of active nodes and the different periods animating the campus (e.g., holidays, school periods, weekends). From the observations, the weekly proportion of nodes having at least 50% of exact location predictions is in average about 75%. This proportion increase by 10% if nodes also consider known properties on the deployment area (e.g., buildings, offices, paths).
In Section [2](#S2 "2 Otiy’s design ‣ Otiy: Locators Tracking Nodes"), we present the rationale for a node-centric approach and introduce Otiy’s architecture. In Section [3.2](#S3.SS2 "3.2 Individual-based clustering ‣ 3 Persistence of Nodes’ behavior ‣ Otiy: Locators Tracking Nodes"), we present our algorithm to readapt the association logs and to elect the appropriate locators. In Section [3.3](#S3.SS3 "3.3 Persistence of cyclic behaviors ‣ 3 Persistence of Nodes’ behavior ‣ Otiy: Locators Tracking Nodes"), we analyze the cyclicity and the persistence of the nodes’ behaviors upon which is based Otiy. In Section [4](#S4 "4 Evaluation ‣ Otiy: Locators Tracking Nodes"), we evaluate the agendas accuracy according to the wireless data traces from the Dartmouth campus. We delay our discussion of related work until Section [5](#S5 "5 Related Work ‣ Otiy: Locators Tracking Nodes") in order to have enough context to make the necessary connections. We finally present some conclusion in Section [6](#S6 "6 Conclusion ‣ Otiy: Locators Tracking Nodes").
## 2 Otiy’s design
Otiy introduces a different approach for distributed location services in wireless mesh networks. In this section we first motivate our proposal and then present Otiy’s architecture and operation. For lack of space, many details are omitted.
### 2.1 Rationale
We consider wireless mesh networks with following characteristics: (i) resources are scarce (wireless medium), (ii) the backbone is static, (iii) nodes are mobile, and (iv) the backbone can be highly dense (e.g., for over-provisioning) in localized hotspot areas. In such a context, which is expected to happen in many situations, the networking architecture becomes fully dependent on an efficient location service. An “efficient” location service should have at least the following characteristics:
1. 1\.
Location updates do not interfere much with the network operations.
2. 2\.
Location signaling messages have low latency.
3. 3\.
Location information is accurate.
This paper provides a response to these three requirements. We base our reasoning on the possibility of having persistent location information (i.e., with long validity duration) – location updates to the anchor (called location dissemination in the rest of the paper) become then more spaced in time, which reduces the amount of propagated signaling messages. This is the response to requirement #1 above.
To provide persistent location information, we take as a premise that different nodes have different levels of mobility. Many studies in the literature have shown that there is a large part of predictability in how nodes move at an AP \[[8](#bib.bib8), [12](#bib.bib12), [4](#bib.bib4)\] granularity. Location services that are based on these studies use prediction to precisely identify the APs to which nodes will be associated with. Nevertheless, the presence of the ping-pong effect and the development of cognitive radio can have major impact on the efficiency of these approaches.
In Otiy, we take a different approach. Instead of trying to obtain the exact location of a node, we use the node’s past and present mobility pattern to distribute locators throughout the network. With high probability (as shown in Section [3.3](#S3.SS3 "3.3 Persistence of cyclic behaviors ‣ 3 Persistence of Nodes’ behavior ‣ Otiy: Locators Tracking Nodes")), locator nodes are placed close to the current location of the node and so, close to the shortest path between sources and destinations. This responds to both requirements #2 and #3.
### 2.2 The reasons for node-centric mobility
We have identified four aspects related to the mobility of nodes:
- •
Periods of activity. Depending on the device type and the necessity to being connected, nodes can show completely different periods of activity, ranging from diurnal activity to sporadic connections only on week-ends.
- •
Mobility coverage. We refer to “mobility coverage” as the number of different APs visited during a single session. We observe that some nodes are aware of the roaming capabilities offered by their underlying network and take advantage of them, while others remain static or have only a nomadic behavior.
- •
Home location characterization. It is not always possible to identify for each node a home location (an area where a node spends more than 50% of its association time \[[6](#bib.bib6)\]). This parameter gives an indication on the regularity of the associations of the nodes to a specific location.
- •
Number of visited areas. This parameter is an extended view of the mobility coverage; the number of visited areas accounts for all visits of a node for the entire observed period (multiple sessions). Mobile or not, the disparities between nodes on this point are deep. The interest of visiting different APs in the network has clear relationship with the social interest of visiting different areas in the environment.
Interests, constraints, and motivations behind the patterns of mobility are manifold and have different impact on the complexity of network management. Furthermore, each node has its own mobility pattern, which itself varies in time. For these reasons, we advocate that, for a location service to be efficient, it must manage mobility at a node scale.
Otiy goes a bit further and proposes that nodes self-profile their behaviors. Such a node-centric approach consists in making nodes themselves log their sequence of associations together with timestamps and SSIDs. We show in the following how Otiy makes use of such information.
### 2.3 Agenda of locators
The key element of our proposal is the agenda of locators. The image of the agenda is important and contributes to highlight the relationship we want to give between a precise location and the considered time.
We delay our discussion of cyclicity of mobility until Section [3.3](#S3.SS3 "3.3 Persistence of cyclic behaviors ‣ 3 Persistence of Nodes’ behavior ‣ Otiy: Locators Tracking Nodes") in order to avoid interrupting our reasoning. In this way, we ask the reader to just assume for the moment that nodes have cyclic mobility.
#### 2.3.1 Description
We define the agenda for node $`n`$ as table $`A_{n}`$. It is composed of $`N`$ columns defining the days of cycle and $`M`$ rows that give the granularity of the estimation we give to the node’s mobility. W.l.g., in this paper we consider a cycle of one week and a granularity of one hour (i.e., $`N = 7`$ and $`M = 24`$). The values of $`N`$ and $`M`$ will be motivated in Section [3.3](#S3.SS3 "3.3 Persistence of cyclic behaviors ‣ 3 Persistence of Nodes’ behavior ‣ Otiy: Locators Tracking Nodes")). We also assign a validity period for the agenda, which is a multiple of $`N`$ representing the number of full cycles the agenda will remain unchanged (no updates will be provided).
To each time slot $`a_{i\hspace{0pt}j}`$ (with $`0 \leq i < N`$ and $`0 \leq j < M`$) of an equal duration $`D = {\lceil\frac{24}{M}\rceil}`$ hour(s), is assigned one locator node. The choice of this locator depends on the mobility of the node observed during the same time slot in the past and is typically a mesh router network address.
We store the history of mobility pattern of node $`n`$ in a set of tables $`H_{n}^{k}`$ (of same dimensions as $`A_{n}`$), where $`k \geq 0`$ indicates the “age” of the table. We define $`H_{n}^{0}`$ as the table containing the current mobility of the node, $`H_{n}^{1}`$ the mobility of the precedent week and so on until $`k = k_{\max}`$ ($`k_{\max}`$ typically varies between 2 and 4). In each time slot $`h_{i,j}^{k}`$, the node records the location of the area at which the node spent most of its time during slot $`j`$ for the day $`i`$, $`k`$ weeks before, according to the mobility log file. In this way, we define a notion of prevalence of a particular area for a given time slot. Furthermore, for each $`h_{i,j}^{0}`$ the node records the duration of association $`d\hspace{0pt}{(h_{i,j}^{0})}`$ (where $`0 < {d\hspace{0pt}{(h_{i,j}^{0})}} \leq D`$) of the selected location.
[Refer to caption](/html/0709.2252/assets/x1.png)
Figure 1: Example of a simplified agenda of associations.
At the end of each validity period, the node generates a new agenda based on its mobility history (as on Fig. [1](#S2.F1 "Figure 1 ‣ 2.3.1 Description ‣ 2.3 Agenda of locators ‣ 2 Otiy’s design ‣ Otiy: Locators Tracking Nodes")). The resulting locator $`a_{i,j}`$ will depends on the oldness he have in the network, the necessary mobility history needed to provide an accurate location $`k_{\max}`$ (studied in Section [4](#S4 "4 Evaluation ‣ Otiy: Locators Tracking Nodes")), and the duration of association $`d\hspace{0pt}{(h_{i,j}^{0})}`$.
We can now define the locator $`a_{i,j}`$ as:
<table>
<tbody>
<tr>
<td></td>
<td>ai,j={hi,j0ifd(hi,j0)=Dmaxoccur(hi,j0,…,hi,jkmax)otherwise,subscript𝑎𝑖𝑗casessuperscriptsubscriptℎ𝑖𝑗0if𝑑superscriptsubscriptℎ𝑖𝑗0𝐷maxoccursuperscriptsubscriptℎ𝑖𝑗0…superscriptsubscriptℎ𝑖𝑗subscript𝑘𝑚𝑎𝑥otherwisea_{i,j}=\left\{\begin{array}[]{ll}h_{i,j}^{0}&\qquad\mathrm{if}\quad d(h_{i,j}^{0})=D\\ \text{maxoccur}(h_{i,j}^{0},...,h_{i,j}^{k_{max}})&\qquad\mathrm{otherwise,}\\ \end{array}\right.</td>
<td></td>
<td>(1)</td>
</tr>
</tbody>
</table>
In the case where $`{d\hspace{0pt}{(h_{i,j}^{0})}} = D`$, we assume that the duration of association at this particular area has been greater than $`D`$. If in the node’s history it has always been the same area, then giving a higher weight to the latest observation does not improves the quality of the estimation. If it is a different area, we conclude that it indicates a deep change in the behavior of the node and then we select this locator for the agenda.
If $`{d\hspace{0pt}{(h_{i,j}^{0})}} < D`$, we choose the locator which has the maximum number of occurrences in the mobility history for the same time slot (if equal, we choose the most recent locator). In this case, we can not judge on the persistence that $`h_{i,j}^{0}`$ will have in the future; in this way, we give more weight to the node’s habits for this time slot.
######
Scheme 1
To avoid holes in the agenda, we extend the locator of the preceding non empty slots to cover an empty time slot.
#### 2.3.2 Bootstrap
For the first association in the network, the node $`n`$ generates an agenda $`A_{n}`$ where all $`a_{i,j}`$ are set to the location of the first visited mesh router. This agenda has a validity period until the end of the cycle (in our case the end of the week). At the end of this period, the new generated agenda will have logically the values given by $`H_{n}^{0}`$. After a longer period, the choice of the locators become more accurate because the node disposes of a larger history.
######
Scheme 2
The bootstrap procedure is executed for the first association in the network and also after long periods of inactivity. In this latter case, we observe that the behaviors of the nodes often become completely different (e.g., change of home location and different mobility pattern). The minimal duration of inactivity before a restart of the bootstrap is discussed in Section [4](#S4 "4 Evaluation ‣ Otiy: Locators Tracking Nodes").
### 2.4 Location updates
[Refer to caption](/html/0709.2252/assets/x2.png)
Figure 2: Mechanisms for location dissemination and updates.
We make a logical distinction between updating the generated agenda (called “location dissemination”) and updating the current location to the corresponding locator in a given time slot (called “location correction”). The term “correction” is related to the idea that we hope that the locator will be the point where the node will be directly associated with.
#### 2.4.1 Agenda dissemination
As shown in Fig. [2](#S2.F2 "Figure 2 ‣ 2.4 Location updates ‣ 2 Otiy’s design ‣ Otiy: Locators Tracking Nodes"), after having generated the agenda of locators, the node has two ways to disseminate it (dashed lines). The first is with the basic primitive of sending (“PUT”) a copy of the agenda to the responsible anchor. The agenda will thus be available for every new contact. The second is through a new primitive “SHARE”. The node, in a P2P way, can share its agenda or the agenda of a known contact with others peers. We added it for two main objectives: (i) to support viral dissemination of this agenda in order to push decentralization to its limits, and (ii) to give a community dimension to this agenda. Nodes in the same social and/or physical community will be able to exchange between them the agenda of a particular server/device or of a common friend.
<sup>2</sup>
<sup>2</sup>2Although this is an important optimization aspect of Otiy, we do not address it in details here.
#### 2.4.2 Location correction
There are two situations where location corrections occur (semi-dashed line in Fig. [2](#S2.F2 "Figure 2 ‣ 2.4 Location updates ‣ 2 Otiy’s design ‣ Otiy: Locators Tracking Nodes")):
- •
If at $`h_{i,j}^{0}`$ the node is still in the area of $`a_{i,{j - 1}}`$ (or $`a_{{i - 1},M}`$ if $`j = 0`$) or on its way to $`a_{i,j}`$. In this cases the node registers its current location to the right locator (given by the agenda).
- •
If at $`h_{i,j}^{0}`$ the node is in the same region than $`a_{i,j}`$ but not directly associated with the locator.
We make this distinction between the two situations because, for the latter one, one could setup a local mobility management system in order to limit the number of location corrections. Location corrections are made pro-actively and are used by the locator to return the current location of the node when required by correspondent nodes.
## 3 Persistence of Nodes’ behavior
In Otiy each node creates its own agenda, based on simple information collected from its preceding associations and movements. In order to validate the concept of agenda, we first need in this section to validate the assumption that nodes’ behaviors are persistent and mostly periodic by nature. We use the node movements collected on the wireless access network of Dartmouth campus \[[10](#bib.bib10)\] to study the perception nodes can get of their own mobility.
### 3.1 Retrieving behaviors from data logs
#### 3.1.1 Experimental Data Set
The data set we use represent three years (2001-04-11 to 2004-06-30) of collected information about all the wireless cards connecting to the wireless access network of Dartmouth campus \[[9](#bib.bib9)\]. The campus is composed of 188 buildings covered by 566 official APs on 200 acres and about 5,500 students.
To better understand nodes’ mobility characteristics, we focus our analysis on the movement files accompanying these data traces \[[10](#bib.bib10)\]. These files detail the associations and disconnections periods of each anonymized wireless adapter to any of the APs. A disconnection is recorded either as the result of disassociation requests or after 20 minutes of inactivity.
Our analysis is based on a four-week period, from the 5th of January 2004 until the end of the 1st of February of the same year. The choice of a four-week period is to smooth the impact of punctual changes in behaviors. Although this period length might seem too large or too short to capture a complete behavior for a certain number of observed nodes, we noticed that this is a good compromise to study the persistence of behaviors.
The choice of the month of January 2004 has been made for three main reasons: (i) to avoid the reported bugs in the collection of Syslog events, (ii) because a large number of the devices were active during this month (just after Christmas holidays), and (iii) the year 2004 showed more mature nodes.
<sup>3</sup>
<sup>3</sup>3We consider as “mature nodes” the nodes that know the network well enough so that they make use of this knowledge and change their displacements in function \[[6](#bib.bib6)\].
#### 3.1.2 Identification of movements and positions
The comprehension of node mobility is a tough problem when it relies on raw measurement data. The wireless nature of the network with all the variations and their consequences, as well as the density of the APs in the environment, is reflected as variations in the observed topology. We can cite at least four types of events that cause these variations:
1. 1\.
Ping-pong effect. It refers to the succession of associations-disassociations between two ore more APs. It is caused by the closeness of the signal to noise ratio (SNR) of neighboring APs and/or the aggressiveness of the wireless card.
2. 2\.
Localized network problems. If for some technical reasons, one AP becomes disabled for a certain amount of time, the node probably associates with another neighboring AP.
3. 3\.
Physical micro-variations. The physical mobility of nodes is frequently very localized (about a few meters). In topological dense areas, these micro-variations result in highly variable association patterns.
4. 4\.
Erroneous reproducibility. There is a probability that the same physical movement results in different association patterns.
Concerning this latter point, the repetition of movements can create junctures between those different patterns and create sectors of micro-mobility which can be detected with a topological standpoint. The need of an algorithm to recognize these junctures is then required to better understand the real objectives of the observed movements. To do so, we introduce a clustering algorithm that uses roaming events contained in the data logs to help each node identifying nearby APs from their mobility point of view. Otiy relies on these clusters to provide more accurate predictions of associations.
### 3.2 Individual-based clustering
The goal of the clustering algorithm we propose in the following is to identify “places” of association that hide areas of micro-variations (and thus reduce useless location updates). The methodology is decoupled in two parts: the collection of network associations and the clustering procedure.
#### 3.2.1 Collection of network associations
The collection of network events is assured by two data structures. For each node, the relationship with the $`M`$ APs in the network are represented through the roaming matrix $`\mathbf{R} = {M \times M}`$, where each element $`r_{i\hspace{0pt}j} \in \mathbf{R}`$ informs about the number of cumulated roaming events from AP $`i`$ to AP $`j`$. The cumulated roaming events imply that the relationship between two subsets of APs can appear after independent sessions.
The second data structure is also created in a per-node basis. It is an $`M`$-row table that stores general information about each AP, such as the total number of associations, the cumulated association duration, and the average association duration.
#### 3.2.2 The clustering algorithm
We now define the terms which will be used in the algorithm:
- •
Link. There is a “link” $`l_{i\hspace{0pt}j}`$ between two APs $`i`$ and $`j`$ if there are bi-directional roaming events between these two APs ($`r_{i\hspace{0pt}j} \neq 0`$ and $`r_{j\hspace{0pt}i} \neq 0`$).
- •
Cost of a link. The cost of a link or “distance” between two APs $`i`$ and $`j`$ is equal to $`r_{i\hspace{0pt}j} + r_{j\hspace{0pt}i}`$. We thus define the cost of a link $`l_{i\hspace{0pt}j}`$ as $`c_{i\hspace{0pt}j} = c_{j\hspace{0pt}i} = {r_{i\hspace{0pt}j} + r_{j\hspace{0pt}i}}`$.
- •
Cluster: A “cluster” is a group (possibly unitary) of APs. Each AP belongs to only one cluster. Two APs are eligible to be merged in the same cluster if there exists a link between them.
- •
Weight of a cluster. The weight $`w_{i}`$ of a cluster $`c_{i}`$ is the value of the maximum link cost within the cluster.
We start with a graph whose vertices represent the APs of the network. An edge exists between two vertices if there is a link (as defined above) between the corresponding APs. In order to limit the variations of intra-cluster link costs, we define a threshold $`k`$ (with $`0 \leq k \leq 1`$).
At the beginning, each AP $`i`$ becomes a cluster $`c_{i}`$ of size 1 and weight $`w_{i} = 0`$. We consider first the link with the highest value in the graph. If two APs $`i`$ (cluster $`c_{i}`$) and $`j`$ (cluster $`c_{j}`$) can merge together (i.e., $`c_{i\hspace{0pt}j} \geq {k \times {\max{\{ w_{i},w_{j}\}}}}`$), then the weight of the resulting cluster is equal to the highest value of the links within the cluster, i.e., $`w_{c_{i} \cup c_{j}} = {\max{\{ w_{i},w_{j}\}}}`$.
[Refer to caption](/html/0709.2252/assets/x3.png)
Figure 3: Clustering mechanism according to the number of roaming events between access points (with $`k = 0.5`$).
We repeat the clustering process until there are no more links to be considered. An example of a resulting clustered graph representation is illustrated in Fig. [3](#S3.F3 "Figure 3 ‣ 3.2.2 The clustering algorithm ‣ 3.2 Individual-based clustering ‣ 3 Persistence of Nodes’ behavior ‣ Otiy: Locators Tracking Nodes"). One can notice that the costs $`l_{D\hspace{0pt}F}`$ and $`l_{D\hspace{0pt}G}`$ are not sufficient to merge both clusters. The same happens with $`l_{A\hspace{0pt}C}`$, $`l_{C\hspace{0pt}E}`$, and $`l_{E\hspace{0pt}G}`$. Such an approach serves to differentiate paths from locations where an node stays longer.
We can now define that within a cluster, an eligible locator will be the mesh router with the greater cumulated association duration to interpret the likelihood to be associated on a particular AP.
#### 3.2.3 Resulting properties
[Refer to caption](/html/0709.2252/assets/x4.png)
Figure 4: CDFs of the completeness level of the resulting embedded graphs and the gathering level.
We study the variations of two important properties of the resulting embedded graphs generated through our clustering algorithm among the patterns of mobility of the nodes (see Fig. [4](#S3.F4 "Figure 4 ‣ 3.2.3 Resulting properties ‣ 3.2 Individual-based clustering ‣ 3 Persistence of Nodes’ behavior ‣ Otiy: Locators Tracking Nodes")). We ran the clustering algorithm on 4 weeks by using $`k = 0.5`$ on 4,766 active nodes.
The gathering level. We define as “gathering level” the ratio of the number of generated clusters on the total number of visited APs. For 68% of the nodes, the ratio is greater than $`0.5`$ and strictly inferior to 1, while for 23% the ratio is equal to $`1`$. The clustering algorithm thus performs well as it gathers mostly APs involved in ping-pong effects and high micro-mobility highlighting the different places of association.
The completeness level. We define as “completeness level” the ratio of the number of edges between the clusters on the number of edges required to obtain a complete graph. A graph of $`n`$ vertices is complete when there is an edge between every pairs of distinct vertices. This represents $`\frac{n\hspace{0pt}{({n - 1})}}{2}`$ edges. With this ratio, we have an idea of the closeness of the different clusters according to the mobility of the nodes. 22% of the nodes have a ratio equal to $`0`$. They do not have inter-clusters mobility and/or have only one generated cluster. In contrast only 10% have a ratio equal to $`1`$ which represents a limited coverage area of mobility. Finally 58% of the nodes have a ratio comprised between $`\rbrack 0,32\rbrack`$. They have inter-cluster mobility but not completely connected graphs.
### 3.3 Persistence of cyclic behaviors
Otiy relies on notions of persistence of cyclic behaviors and preferential time of attachment to specific areas. In this section, we analyze patterns of association with the above-created clusters. To do so, we rewrite each movement file by replacing APs’ identifiers by their corresponding clusters’ identifiers and by aggregating the timestamps of consecutive associations in the same cluster.
#### 3.3.1 Cyclic time-related (re)association behavior
Individuals which have habits in an environment, show cyclic patterns of mobility at the topological level. We asset this aspect through the analysis of the re-associations at prevalent locations in an hourly basis by making distinction between the days of the week.
For this analysis, we enlarge the observed period to eight weeks to be sure to not be in a situation where the nodes show optimum cyclic behaviors. We define a location (cluster) as prevalent for a specific time slot if the cumulated duration of association in this location is greater than in the others locations visited during the same time slot. For a specific time slot we take into account only the nodes which have been associated the preceding week of activity in the same time slot. Finally, each day are defined through 24 time slots of one hour each.
[Refer to caption](/html/0709.2252/assets/x5.png)
Figure 5: Percentage of nodes which have been associated to the same prevalent location of the week before during the same time slot.
In the Fig. [5](#S3.F5 "Figure 5 ‣ 3.3.1 Cyclic time-related (re)association behavior ‣ 3.3 Persistence of cyclic behaviors ‣ 3 Persistence of Nodes’ behavior ‣ Otiy: Locators Tracking Nodes"), we analyze the persistence of the prevalence of a predicted location for each time slot by making the distinction between the days. This persistence is analyzed, each time, through two consecutive weeks of activity. We plot the percentage of nodes for which the analysis of the time slots have been possible.
From this figure we can make three main observations: (i) from weeks to weeks, the percentage of correct predictions stays approximately high and stable between 75 and 95%. (ii) The differences between the days are not strong and denote that making the distinction between the days does not affect the accuracy of the predictions. (iii) The percentage of nodes which have correct predictions is minimal in diurnal periods between (in average 75%) and maximal the nights (in average 90%). This is because we have more observed nodes during diurnal periods than in nights. The increasing number of nodes and their high activity in diurnal periods decreases sensibly the results.
However, this figure does not give information on the correctness prediction of the intra-day sequence of prevalent locations as we consider each time slot independently. We thus consider that through the scheme 1 (the expanding of location prevalence in the following empty slot) and changes in pattern of mobility, that the accuracy of the predictions will slightly decrease.
#### 3.3.2 Persistent periods of activity
As mentioned in Section [2.2](#S2.SS2 "2.2 The reasons for node-centric mobility ‣ 2 Otiy’s design ‣ Otiy: Locators Tracking Nodes"), the period of activity is a node-related parameter. In this subsection, we analyze the differences among the nodes and the persistence of this parameter.
In the following, we provide results about nodes which present a certain consistency in their activity in the network. For instance, results about mobility patterns every Monday are based on the nodes who were active the four Monday of the observed period. In the same way, weekly results are based on nodes who were active each day of at least one week. The constraint about the consistency in the activity is mostly motivated by the necessity to compare patterns of mobility on a same plan, and to be able to analyze persistence in behaviors.
[Refer to caption](/html/0709.2252/assets/x6.png)
[Refer to caption](/html/0709.2252/assets/x7.png)
Figure 6: Different types of periods of activity. It shows the results for two nodes on Monday of each week of our observed period.
The Fig. [6](#S3.F6 "Figure 6 ‣ 3.3.2 Persistent periods of activity ‣ 3.3 Persistence of cyclic behaviors ‣ 3 Persistence of Nodes’ behavior ‣ Otiy: Locators Tracking Nodes") presents the periods of activity of two different nodes for the same day (Monday) of the four weeks. In this figure, we analyze the behavior of the nodes within one particular day. While for the node at the left we can observe regularity in the period of non activity (between 9h and 17h), the periods of activity of the node at the right are completely different. The main important aspect is in the regularity and so the persistence of the period of activity of the node at the left. For the node at the right, the cumulated periods of activity represents approximately an activity all the day. This is an entirely different behavior which let suppose that the node can have access to the network at any time.
If these two patterns of activity can be classified in two different categories, it should exist two more different categories in this classification: (i) nodes active along the day and (ii) nodes active only during work hours. Under this classification, it is different ways for different needs which can dictate the behavior of the nodes in the access of the network. This supposes persistence at short and average terms of the behaviors. However, we can not conclude on this intra-day behavior without comparing with the periods of activity the others days of the week.
[Refer to caption](/html/0709.2252/assets/x8.png)
Figure 7: For each day, the CDF of the length of time slots where the nodes have never been active.
The Fig. [7](#S3.F7 "Figure 7 ‣ 3.3.2 Persistent periods of activity ‣ 3.3 Persistence of cyclic behaviors ‣ 3 Persistence of Nodes’ behavior ‣ Otiy: Locators Tracking Nodes") represents, for each day, the CDF of the maximum consecutive duration of non activity in the network. For each day, the periods of activity of the four weeks are cumulated to give the periods of non activity. Then the durations of non activity observed correspond to periods where the nodes have never been active in the network. The first observation is that around 40% of nodes are active all the day for every week days and 48% to 70% the week-end. The second observation is that for 60% of nodes the maximum consecutive period of non activity is nearly the same for each day of the week days except for the Fridays. 40% of these nodes is, for 12 hours and less, absent from the network.
It is important to note that the durations of non activity are clearly different between the week days and the week-end. The nodes which are subject to constraints (social or not) the week days are more free to access the network differently the week-end.
To summarize, the nodes can have cyclic periods of activity which can be for a majority of them persistent. These behaviors are not directly correlated from one day to another. It is thus required to take each day independently.
## 4 Evaluation
As explained previously, a locator is assigned to each time slot. To generate an agenda, each mobile node first determines for each time slot what is the prevalent cluster. At the end of the validity period of its agenda, it schedules the new locators’ positions. For each time slot, the locator is positioned in the prevalent cluster. To evaluate the accuracy of our predicted agenda, we need to check for each time slot whether the mobile node was close to its locator. Therefore we verify if the mobile node has visited its locator’s cluster during the given time period.
### 4.1 Methodology
For this evaluation we enlarge the period of analysis. It starts now from the 6th of January 2003 (timestamp 1041829200) to finish the end of the 29th of February 2004 (timestamp 1078117200). We choose this period length to see how the learning step can improve the accuracy of our predictions after each school holidays and with the arrival of new nodes.
To evaluate the pertinence of our agenda, we check whether the mobile node has visited its locator’s cluster during a time slot. If we find a match, we consider the mobile node is effectively close to its locator. We perform this comparison for each time slot, each time a mobile node was connected to the network at least once during a time slot. We do not consider time slots where a node is inactive. Similarly, we do not include in our study the reliability of the first bootstrapped agenda, as it does not reflect an observed mobility pattern. The creation of the first agenda is detailed in Section [2.3.2](#S2.SS3.SSS2 "2.3.2 Bootstrap ‣ 2.3 Agenda of locators ‣ 2 Otiy’s design ‣ Otiy: Locators Tracking Nodes").
For each time slot, we can thus define a good or a bad prediction. We evaluate the accuracy of the agenda as the ratio of bad predictions ($`N\hspace{0pt}b\hspace{0pt}B\hspace{0pt}a\hspace{0pt}d\hspace{0pt}P\hspace{0pt}r\hspace{0pt}e\hspace{0pt}d\hspace{0pt}i\hspace{0pt}c\hspace{0pt}t\hspace{0pt}i\hspace{0pt}o\hspace{0pt}n\hspace{0pt}s`$) over the total number of considered time slot ($`T\hspace{0pt}o\hspace{0pt}t\hspace{0pt}a\hspace{0pt}l\hspace{0pt}N\hspace{0pt}b\hspace{0pt}S\hspace{0pt}l\hspace{0pt}o\hspace{0pt}t\hspace{0pt}s`$). We define the error of prediction in the agenda as $`A_{e\hspace{0pt}r}`$ by:
<table>
<tbody>
<tr>
<td></td>
<td>Aer=NbBadPredictions∗100TotalNbSlots.subscript𝐴𝑒𝑟𝑁𝑏𝐵𝑎𝑑𝑃𝑟𝑒𝑑𝑖𝑐𝑡𝑖𝑜𝑛𝑠100𝑇𝑜𝑡𝑎𝑙𝑁𝑏𝑆𝑙𝑜𝑡𝑠A_{er}=\frac{NbBadPredictions*100}{TotalNbSlots}.</td>
<td></td>
</tr>
</tbody>
</table>
W.l.g, we choose to update, each week and for each node the clusters by taking into account the recent patterns of mobility but also all the preceding history of patterns of mobility. The update of the clusters is made after the agenda evaluation and we make sure that the preceding locator predictions are still correct even if the eligible locators for the clusters have changed.
### 4.2 Evaluation of the $`A_{e\hspace{0pt}r}`$
[Refer to caption](/html/0709.2252/assets/x9.png)
Figure 8: Histogram of the proportion of nodes according to their $`A_{e\hspace{0pt}r}`$ for each week.
For each week in Fig. [8](#S4.F8 "Figure 8 ‣ 4.2 Evaluation of the 𝐴_{𝑒𝑟} ‣ 4 Evaluation ‣ Otiy: Locators Tracking Nodes"), we plot the histogram of the proportion of nodes which have an $`A_{e\hspace{0pt}r}`$ value of 0% (perfect prediction), less or equal to 25%, 50%, 75% and 100%. We used a minimum of one week and up to two weeks of mobility history to predict the value of the agenda.
The first observation is the large proportion of nodes having an accurate agenda for the all week, $`A_{e\hspace{0pt}r} = {0\%}`$. Peaks of this phenomenon can be observed at summer breaks or during Christmas holiday. This certainly results from a static behavior of nodes. This proportion increases as the number of active node decreases.
We also observed that only around 40% of the average proportion of nodes have at least 25% of incorrect predictions in their agenda. Depending on the observed week, this proportion of nodes varies between 20 and 50%. An average of 15% of the proportion of nodes have more than 75% of wrong predictions in their agenda and approximately 10% have between 50% and 75% of bad accuracy.
It is important to note that if a node is not active during a time slot, a default locator is set in the agenda, thus repeating the value of the previous locator as explained in Section [2.3.1](#S2.SS3.SSS1 "2.3.1 Description ‣ 2.3 Agenda of locators ‣ 2 Otiy’s design ‣ Otiy: Locators Tracking Nodes"). A bad prediction can be easily explained by a lack of information about a node mobility pattern. This phenomenon can be observed in the Fig. [8](#S4.F8 "Figure 8 ‣ 4.2 Evaluation of the 𝐴_{𝑒𝑟} ‣ 4 Evaluation ‣ Otiy: Locators Tracking Nodes"), when the proportion of active nodes increases and remains stable (marked as a line on the plot). We can note that the persistence of the patterns of mobility then improves the accuracy of the generated agenda. This accuracy decreases logically at the beginning and at the end of the different holidays periods as mobility constraints are relaxed during holidays and new nodes join the network.
In facts, the reasons of most of the wrong matches are simple: (i) often, nodes active only two weeks in the observed period do not develop any cyclic pattern of mobility and thus have an high $`A_{e\hspace{0pt}r}`$ value. (ii) More rarely, places which certainly constitute part of a path between two significant areas are assigned to time slots (through scheme 1). Finally, (iii) overlapping between prevalent places for one node occurs most of the time between 9am and 3pm.
These first results appear extremely promising for Otiy, our location scheme, as it give in much cases an agenda that can allow the optimization of the localization process.
### 4.3 Impact of the history length
[Refer to caption](/html/0709.2252/assets/x10.png)
Figure 9: Impact of history length used to generate the agenda on the number of correct matches ($`A_{e\hspace{0pt}r} \leq {50\%}`$).
In the Fig. [9](#S4.F9 "Figure 9 ‣ 4.3 Impact of the history length ‣ 4 Evaluation ‣ Otiy: Locators Tracking Nodes"), we analyze the contribution of the length of mobility history used to create the agenda on the percentage of correct matches. Here, we plotted only the case where there are up to 50% bad matches in the agenda prediction.
We get the higher proportion of nodes having at least 50% of the time a good estimation of their agenda prediction for a mobility history of one week. This means that recent mobility patterns are sufficient to estimate an accurate agenda. Nevertheless, the duration of activity during the current and preceding weeks also has a clear impact on the $`A_{e\hspace{0pt}r}`$.
A convergence between different length of history can be observed after as a result of more stable behavior of nodes. But using long mobility history, one can note that the agendas appears more sensitive to changes of mobility patterns. On the figure (week 11, weeks 20 to 23, and weeks 32 to 35), we clearly see that the accuracy of the agendas drop roughly at the beginning and the end of the school holidays period.
In order to get an agenda reactive to the changes of mobility pattern but also be able to gather enough data about nodes’ behavior to get the right locator estimate, we choose in the rest of our tests to use at a two weeks mobility history.
### 4.4 Comparison between the different days
[Refer to caption](/html/0709.2252/assets/x11.png)
Figure 10: Proportion of the nodes which have an $`A_{e\hspace{0pt}r} \leq {50\%}`$ for each day.
It is interesting to check whether the mobility behaviors affect differently the percentage of incorrect predictions depending particular week days. In the Fig. [10](#S4.F10 "Figure 10 ‣ 4.4 Comparison between the different days ‣ 4 Evaluation ‣ Otiy: Locators Tracking Nodes") we compare the proportion of nodes which have an $`A_{e\hspace{0pt}r} \leq {50\%}`$ for each day of the week.
The first observation, is that the proportion of nodes with the same accuracy in their agenda remains approximately the same whatever day is considered (between 70 and 90%). However, the variations of proportion are more pronounced for week-ends and Mondays. These days are more sensitive to changes in nodes’ behavior at the beginning and at the end of the holidays. The rest of the time the pertinence of the agenda seems more important for these three days.
We envision taking advantage of this phenomenon to better detect changes in behaviors and to more accurately restart the bootstrapping procedures.
### 4.5 Evaluation of the physical distance
[Refer to caption](/html/0709.2252/assets/x12.png)
Figure 11: Histogram of the proportion of nodes according to their $`A_{e\hspace{0pt}r}`$ values for each week. In this figure, we consider also a correct prediction if the node has visited at least one cluster connected (at one hop) to the locator.
As mentioned before, the topology of the network is not provided with the data traces. As a consequence we can not determine the topological path length(s) between the visited areas during a time slots and the locator’s position.
In order to have an idea of how close the nodes have been from the locator’ cluster, we use the graph from which is built the clusters. In the Fig. [11](#S4.F11 "Figure 11 ‣ 4.5 Evaluation of the physical distance ‣ 4 Evaluation ‣ Otiy: Locators Tracking Nodes"), we evaluate $`A_{e\hspace{0pt}r}`$ considering a prediction as correct, if during a time slot, a node has visited the locator’ cluster until a cluster one hop away.
Comparing the results with Fig. [8](#S4.F8 "Figure 8 ‣ 4.2 Evaluation of the 𝐴_{𝑒𝑟} ‣ 4 Evaluation ‣ Otiy: Locators Tracking Nodes"), we observe that the number of nodes with an $`A_{e\hspace{0pt}r}`$ value of up to 50% of their agenda increases roughly by 10% for each week. The number of nodes which have made more than 75% of incorrect predictions is reduced to, in average, 5% per week.
These results show, that for the majority of the nodes and most of the time we successfully provide a locator close to the current physical location of the nodes. We therefore can provide an agenda that satisfactory reflects nodes mobility.
## 5 Related Work
The problematic of reducing the amount of generated location updates in a wireless network has been well studied in PCS and is very close to our work. Even if the proposals are not really adapted to IEEE 802.11 wireless mesh networks, we will relate the different concepts of the most relevant work.
Tabbane is one of the firsts to have introduced the node (mobility) profiling to improve location management in PCS. In \[[14](#bib.bib14)\] the profiling is operated by the network and shared with the node’s subscriber identity module SIM. Thanks to this profiling, whatever the period of time $`\lbrack t_{i},t_{j})`$ the system can find a list of areas where the nodes could be. This list of areas is decreasingly ordered by the probabilities of being in the different areas. One probability is dependent of a function associated with and has several parameters such as the time, the pattern of mobility, the last location, the weather etc. Until the node is in one of these areas it does not update its location. When the system needs to locate him, it asks sequentially the different areas within the list. Two notions are shared with our approach: (i) the node profiling although in Otiy it is the nodes which make their self-profiling. (ii) The relation with the time. However, in Otiy the period of time are predefined and only one area (anchor) is assigned to each time slot.
Chuon et al. in \[[3](#bib.bib3)\] by calculating the prevalence of the daily different visited cells create an node profile graph (IPG) per node and through their monitoring by the network. This graph is composed by vertices (cells) where the normalized probability of visit on N-days is greater than a specified value (comprised between $`0`$ and $`1`$) and the connectors between these cells (also called anchors). Furthermore, the vertices in the IPG are classified by decreasing order of probability of visit. Until the node stays in the set of vertices in his IPG it does not have to updates its location. To locate it in its IPG, the network pages it in the decreasing order of probability of the anchors as in \[[11](#bib.bib11)\]. We differ from this approach in several ways. We do not cumulate daily patterns of mobility to see a prevalent mobility graph. Rather, we distinguish each daily pattern of mobility of the week to capture the different constraints and interests which are dependent of the days. Although they used the division in time slots to prove that their IPG captures well the diurnal mobility of the nodes, they do not used these time slots to make the relationship, that we make, between a particular area and the time. From this point, our proposal is clearly different of the methods used to updates the locations and to page the nodes.
In \[[16](#bib.bib16)\], Wu et al. mine the mobility behavior nodely (operated by the nodes) from long term mobility history. From this information they evaluate the time-varying probability of the different nodely-defined regions. The prevalence of a region on the time is defined through a cost model. Finally, they obtain a vector $`< time,area >`$ of mobility which will define the region to be paged in function of the time. The location updates and paging schemes are approximately the same than the afore-presented proposals. Here, the prevalence of a particular area on the time is more flexible and accurate than ours. However, the complexity of the algorithm is more important. Furthermore, the length of the vector of mobility can be a lot more important than a division in time slots of equal duration.
Finally in IEEE 802.11 wireless networks, Ghosh et al. have profiling the nodes associations sequences by making the difference between the days in \[[5](#bib.bib5)\]. The result is several patterns of mobility such as Weekend Profile, Home Profile, etc. Here, the areas are called “sociological hub” and are defined at a building granularity. With this set of profiles they are capable to determine which mobility profile the node follow currently (from the firsts associations) and/or define a window of day with attributed mobility profiles. This approach is sensibly different of Otiy but here, we recognize the necessity to differentiate the patterns of mobility which could be different from day to day. Furthermore, our significant areas (they “hub”) are determined by the local micro-mobility of each node and are not based on any geographical information (e.g., building segmentation).
## 6 Conclusion
We introduced Otiy, a node-centric architecture to control the propagation of the location updates in networks of mobile nodes. To limit the impact of location updates, each node creates an agenda of locators, which are access points at which the node will probably get associated with according to the past and present patterns of mobility. The idea behind this approach is to make location update overhead more localized.
The keys principles we achieve are:
- •
A fully decentralized location service.
- •
Short path lengths for location updates.
- •
Shared knowledge on absolute time locations. At anytime, both the node and its interlocutor(s) know the location of the locator.
To achieve our goals, we proposed to gather nearby APs in terms of node roaming events in order to create areas (clusters) of micro-mobility. Our clustering algorithm does not use any geographical information and is based only on the sequence of associations/disassociations. This drastically limits the propagation of updates due to artifacts such as ping-pong effects.
Using around one year of mobility traces, we have in average 75% of the nodes that have at least 50% of their time slots which match exactly the right predictions. These results encourage us to keep improving Otiy and evaluate it using real testbeds.
## References
- \[1\] M. Balazinska and P. Castro. Characterizing mobility and network usage in a corporate wireless local-area network.
In Proc. of MobiSys 2003, pages 303–316, San Francisco, CA, May 2003.
- \[2\] C. Cheng, H. Lemberg, S. Philip, E. van den Berg, and T. Zhang. Slalom: A scalable location management scheme for large mobile ad-hoc networks.
Wireless Communications and Networking Conference, 2:574–578, 2002.
- \[3\] C. Chuon, S. Guha, and A. K. M. M. Hossain. Individual profile graphs for location management in pcs networks.
Wireless Networks, Communications and Mobile Computing, 2005 International Conference on, 1:187–192, June 2005.
- \[4\] J.-M. François and G. Leduc. AP and MN-centric mobility prediction: A comparative study based on wireless traces.
In Proc. of IFIP Networking 2007, pages 322–332, Atlanta, GA, USA, May 2007.
- \[5\] J. Ghosh, M. J. Beal, H. Q. Ngo, and C. Qiao. On profiling mobility and predicting locations of campus-wide wireless users.
In Proc. of the second international workshop on Multi-hop ad hoc networks: from theory to reality, Florence, Italy.
- \[6\] T. Henderson, D. Kotz, and I. Abyzov. The changing usage of a mature campus-wide wireless network.
In Proc. of the Tenth Annual International Conference on Mobile Computing and Networking (MobiCom), pages 187–201, Philadelphia, PA, USA, 2004.
- \[7\] J. P. Hubaux, T. Gross, J. Y. L. Boudec, and M. Vetterli. Towards self-organized mobile ad hoc networks: the Terminodes project.
IEEE Communications Magazine, 31(1):118–124, 2001.
- \[8\] G. L. Jean-Marc François and S. Martin. Learning movement patterns in mobile networks: a generic method.
In European Wireless 2004, Barcelona, Spain.
- \[9\] D. Kotz, T. Henderson, and I. Abyzov. CRAWDAD trace dartmouth/campus/syslog/01 04 (v. 2004-12-18), dec 2004.
- \[10\] D. Kotz, T. Henderson, and I. Abyzov. CRAWDAD trace set dartmouth/campus/movement (v. 2005-03-08), March 2005.
- \[11\] G. P. Pollini and C.-L. I. A profile-based location strategy and its performance.
IEEE Journal on Selected Areas in Communications, 15(8):1415–1424, 1997.
- \[12\] L. Song, D. Kotz, R. Jain, and X. He. Evaluating location predictors with extensive wi-fi mobility data.
In Proc. of the 23rd Annual Joint Conference of the IEEE Computer and Communications Societies (INFOCOM), volume 2, pages 1414–1424, 2004.
- \[13\] I. Stojmenovic and B. Vukojevic. A routing strategy and quorum based location update scheme for ad hoc wireless networks, 1999.
- \[14\] S. Tabbane. An alternative strategy for location tracking.
IEEE Journal on Selected Areas in Communications, 13, Issue 5:880–892, June 1995.
- \[15\] S.-C. M. Woo and S. Singh. Scalable routing protocol for ad hoc networks.
Wireless Networks, 7(5):513–529, 2001.
- \[16\] H.-K. Wu, M.-H. Jin, and J.-T. Horng. Personal paging area design based on mobiles moving behaviors.
In INFOCOM, pages 21–30, 2001.
- \[17\] J. Zhang and L. Gruenwald. Spatial and temporal aware, trajectory mobility profile based.
In Proc. of The International Database and Expert Systems Applications (DEXA) Workshop on Mobile Database and Distributed Systems, pages 716–720, September 2002. |
0708.0660 | cs.NI | cyber | Network synchronizability analysis: the theory of subgraphs and complementary graphs
<sup>\*</sup>
<sup>\*</sup>
\*
This work was supported by the National Science Foundation of China under grant 60674093 and the City University of Hong Kong under the Research Enhancement Scheme and SRG grant 7002134.
Zhisheng Duan<sup>1</sup>, Chao Liu<sup>1</sup>, and Guanrong Chen<sup>1,2</sup>
<sup>1</sup>State Key Laboratory for Turbulence and Complex Systems, Department of Mechanics and Aerospace Engineering, College of Engineering, Peking University, Beijing 100871, P. R. China
<sup>2</sup> Department of Electronic Engineering, City University of Hong Kong, Hong Kong
Email: duanzs@pku.edu.cn, Fax: (8610)62765037
Abstract. In this paper, subgraphs and complementary graphs are used to analyze the network synchronizability. Some sharp and attainable bounds are provided for the eigenratio of the network structural matrix, which characterizes the network synchronizability, especially when the network’s corresponding graph has cycles, chains, bipartite graphs or product graphs as its subgraphs.
Keywords. Complex network, Subgraph, Complementary graph, Synchronizability.
## 1 Introduction
Synchronization of complex networks has attracted increasing attention from many scientists, for its important applications in various areas of physical and biological sciences \[[3](#bib.bib3), [4](#bib.bib4), [6](#bib.bib6), [9](#bib.bib9), [10](#bib.bib10), [24](#bib.bib24), [26](#bib.bib26)\]. Synchronization is a ubiquitous phenomenon especially in social and biological networks where, quite often, it is desirable, e.g., in the consensus of multi-agent activities, while in some other cases it is undesirable, e.g., in traffic congestion \[[6](#bib.bib6), [16](#bib.bib16), [18](#bib.bib18), [19](#bib.bib19), [22](#bib.bib22), [23](#bib.bib23), [25](#bib.bib25)\]. The connection structure of a network plays an important role in its synchronization. Of particular interest in this concern is how the synchronizability depends on various structural parameters of the network, such as average distance, clustering coefficient, coupling strength, degree distribution and weight distribution, among others. Some interesting results have been established for such important problems based on the notions of master stability function and synchronized region \[[3](#bib.bib3), [5](#bib.bib5), [11](#bib.bib11), [12](#bib.bib12), [13](#bib.bib13), [16](#bib.bib16), [21](#bib.bib21), [29](#bib.bib29)\]. Some relationships between synchronizability and structural parameters of complex networks have also been reported, e.g., smaller average network distance does not necessarily mean better synchronizability \[[19](#bib.bib19)\], therefore the betweenness centrality was proposed as a good indicator for synchronizability \[[10](#bib.bib10)\], and two networks with the same degree sequence can have different synchronizabilities \[[27](#bib.bib27)\]. Some optimizing rules were established for the network synchronizability \[[6](#bib.bib6)\]. Moreover, how the network statistical properties influence the spectrum of the structural matrix was analyzed and reported in \[[2](#bib.bib2)\]. And the effects of graph operations on the network synchronizability were studied in \[[1](#bib.bib1)\]. Last but not least, complementary graphs were used to analyze the network synchronizability in \[[7](#bib.bib7)\]. Motivated by all these research works, this paper attempts to further establish some sharp upper bounds for the network synchronizability based on the theory of subgraphs and complementary graphs.
Consider a dynamical network consisting of $`N`$ coupled identical nodes, with each node being an $`n`$-dimensional dynamical system, described by
<table>
<tbody>
<tr>
<td></td>
<td>x˙i=f(xi)−c∑j=1NaijH(xj),i=1,2,⋯,N,formulae-sequencesubscript˙𝑥𝑖𝑓subscript𝑥𝑖𝑐superscriptsubscript𝑗1𝑁subscript𝑎𝑖𝑗𝐻subscript𝑥𝑗𝑖12⋯𝑁\dot{x}_{i}=f(x_{i})-c\sum_{j=1}^{N}a_{ij}H(x_{j}),\;i=1,2,\cdots,N,</td>
<td></td>
<td>(1)</td>
</tr>
</tbody>
</table>
where $`x_{i} = {(x_{i\hspace{0pt}1},x_{i\hspace{0pt}2},\cdots,x_{i\hspace{0pt}n})} \in {\mathbb{R}}^{n}`$ is the state vector of node $`i`$, $`{f\hspace{0pt}{( \cdot )}}:{{\mathbb{R}}^{n}\rightarrow{\mathbb{R}}^{n}}`$ is a smooth vector-valued function, constant $`c > 0`$ represents the coupling strength, $`{H\hspace{0pt}{( \cdot )}}:{{\mathbb{R}}^{n}\rightarrow{\mathbb{R}}^{n}}`$ is called the inner linking function, and $`A = {(a_{i\hspace{0pt}j})}_{N \times N}`$ is called the outer coupling matrix or structural matrix, which represents the coupling configuration of the entire network. This paper only considers the case that the network is diffusively connected, i.e., $`A`$ is irreducible and its entries satisfy $`{{a_{i\hspace{0pt}i} = {- {\sum_{{j = 1},{j \neq i}}^{N}a_{i\hspace{0pt}j}}}},{i = {1,2,\cdots,N}}}.`$ Further, suppose that, if there is an edge between node $`i`$ and node $`j`$, then $`a_{i\hspace{0pt}j} = a_{j\hspace{0pt}i} = {- 1}`$, i.e., $`A`$ is a Laplacian matrix corresponding to network ([1](#S1.E1 "In 1 Introduction")). Therefore, $`0`$ is an eigenvalue of $`A`$ with multiplicity 1, and all the other eigenvalues of $`A`$ are strictly positive, which are denoted by
<table>
<tbody>
<tr>
<td></td>
<td>0=λ1<λ2≤λ3≤⋯≤λN=λmax.0subscript𝜆1subscript𝜆2subscript𝜆3⋯subscript𝜆𝑁subscript𝜆𝑚𝑎𝑥0=\lambda_{1}<\lambda_{2}\leq\lambda_{3}\leq\cdots\leq\lambda_{N}=\lambda_{max}.</td>
<td></td>
<td>(2)</td>
</tr>
</tbody>
</table>
The dynamical network ([1](#S1.E1 "In 1 Introduction")) is said to achieve (asymptotical) synchronization if $`{{{x_{1}\hspace{0pt}{(t)}}\rightarrow{x_{2}\hspace{0pt}{(t)}}\rightarrow\cdots\rightarrow{x_{N}\hspace{0pt}{(t)}}\rightarrow{s\hspace{0pt}{(t)}}},{{\text{as}\hspace{0pt}t}\rightarrow\infty}},`$ where, because of the diffusive coupling configuration, the synchronous state $`{s\hspace{0pt}{(t)}} \in {\mathbb{R}}^{n}`$ is a solution of an individual node, i.e., $`{\overset{˙}{s}\hspace{0pt}{(t)}} = {f\hspace{0pt}{({s\hspace{0pt}{(t)}})}}`$.
It is well known \[[3](#bib.bib3)\] that the eigenratio $`{r\hspace{0pt}{(A)}} = \frac{\lambda_{2}}{\lambda_{N}}`$ of network structural matrix $`A`$ characterizes the network synchronizability: the larger the $`r\hspace{0pt}{(A)}`$, the better the synchronizability. In this paper, $`r\hspace{0pt}{(A)}`$ will be used as a synchronizability index for the networks with structural matrix $`A`$. The relationships between $`r\hspace{0pt}{(A)}`$ and network structural characteristics such as average distance, node betweenness, degree distribution, clustering coefficient, etc. have been studied \[[7](#bib.bib7), [2](#bib.bib2), [27](#bib.bib27), [28](#bib.bib28)\]. A well-known upper bound for the eigenratio $`r\hspace{0pt}{(A)}`$ in graph theory is $`{r\hspace{0pt}{(A)}} \leq \frac{d_{m\hspace{0pt}i\hspace{0pt}n}}{d_{m\hspace{0pt}a\hspace{0pt}x}}`$, where $`d_{m\hspace{0pt}i\hspace{0pt}n}`$ and $`d_{m\hspace{0pt}a\hspace{0pt}x}`$ denote the smallest and largest degrees of the corresponding graph to the network \[[27](#bib.bib27)\]. To the best of our knowledge, besides this coarse bound, there are no sharp upper bounds available in the literature today. Some sharp upper bounds for $`r\hspace{0pt}{(A)}`$ are derived in this paper by using graph-theoretical methods.
The rest of this paper is organized as follows. In Section 2, some necessary preliminaries in graph theory are provided. In Section 3, for a given graph $`G`$, some upper bounds are derived for the eigenratio $`r\hspace{0pt}{(G)}`$ when $`G`$ has cycles as its subgraphs. In Section 4, some lower bounds for the largest eigenvalue of a graph are presented when $`G`$ has other subgraphs, such as chains, bipartite graphs and product graphs. The paper is concluded by the last section.
Throughout this paper, for any given undirected graph $`G`$, eigenvalues of $`G`$ mean eigenvalues of its corresponding Laplacian matrix. Notations for graphs and their corresponding Laplacian matrices are not differentiated, and networks and their corresponding graphs are not distinguished, unless otherwise indicated.
## 2 Preliminaries
For a given graph $`G`$, let $`\mathcal{V}`$ and $`\mathcal{E}`$ denote the sets of nodes and edges of $`G`$, respectively. A graph $`G_{1}`$ is called an induced subgraph of $`G`$, if the node set $`\mathcal{V}_{1}`$ of $`G_{1}`$ is a subset of $`\mathcal{V}`$ and the edges of $`G_{1}`$ are all edges among nodes $`\mathcal{V}_{1}`$ in $`\mathcal{E}`$. The complementary graph of $`G`$, denoted by $`G^{c}`$, is the graph containing all the nodes of $`G`$ and all the edges that are not in $`G`$. In this paper, subgraphs and complementary graphs are used to discuss the network synchronizability. For this purpose, the following lemmas are needed.
Lemma 1 \[[14](#bib.bib14), [15](#bib.bib15)\] For any given graph $`G`$ of size $`N`$, its nonzero eigenvalues indexed as in ([2](#S1.E2 "In 1 Introduction")) grow monotonically with the number of added edges; that is, for any added edge $`e`$, $`{\lambda_{i}\hspace{0pt}{({G + e})}} \geq {\lambda_{i}\hspace{0pt}{(G)}}`$, $`i = {1,\cdots,N}`$.
Lemma 2 \[[14](#bib.bib14)\] For any given connected graph $`G`$ of size $`N`$, its largest eigenvalue $`\lambda_{N}`$ satisfies $`\lambda_{N} \geq {d_{m\hspace{0pt}a\hspace{0pt}x} + 1}`$, with equality if and only if $`d_{m\hspace{0pt}a\hspace{0pt}x} = {N - 1}`$, where $`d_{m\hspace{0pt}a\hspace{0pt}x}`$ is the maximum degree of $`G`$.
Lemma 3 \[[8](#bib.bib8), [27](#bib.bib27)\] For any cycle $`C_{N}`$ at $`N`$ ($`\geq 4`$) nodes, its eigenvalues are given by $`\mu_{1},\cdots,\mu_{N}`$ (not necessarily ordered) with $`\mu_{1} = 0`$ and
<table>
<tbody>
<tr>
<td></td>
<td>μk+1=3−sin(3kπN)sin(kπN),k=1,⋯,N−1.formulae-sequencesubscript𝜇𝑘133𝑘𝜋𝑁𝑘𝜋𝑁𝑘1⋯𝑁1\mu_{k+1}=3-\frac{\sin(\frac{3k\pi}{N})}{\sin(\frac{k\pi}{N})},\,\,k=1,\cdots,N-1.</td>
<td></td>
</tr>
</tbody>
</table>
By Lemma 3, one knows that, for any cycle $`C_{N}`$ at $`N`$ ($`\geq 4`$) nodes, if $`N`$ is even, its largest eigenvalue is 4; if $`N`$ is odd, its largest eigenvalue is $`{{\lambda_{N}\hspace{0pt}{(C_{N})}} = {3 - \frac{\sin{(\frac{3\hspace{0pt}{({N - 1})}\hspace{0pt}\pi}{2\hspace{0pt}N})}}{\sin{(\frac{{({N - 1})}\hspace{0pt}\pi}{2\hspace{0pt}N})}}}}.`$
Obviously, for cycles with odd numbers of nodes, the largest eigenvalue converges to 4 as the number of its lengths tends to $`+ \infty`$. For eigenvalues of graphs and complementary graphs, the following lemma is useful (see \[[14](#bib.bib14), [15](#bib.bib15)\] and references therein).
Lemma 4 For any given graph $`G`$ of size $`N`$, the following statements hold:
\(i\) $`\lambda_{N}\hspace{0pt}{(G)}`$, the largest eigenvalue of $`G`$, satisfies $`{{\lambda_{N}\hspace{0pt}{(G)}} \leq N}.`$
\(ii\) $`{\lambda_{N}\hspace{0pt}{(G)}} = N`$ if and only if $`G^{c}`$ is disconnected.
(iii) If $`G^{c}`$ is disconnected and has (exactly) $`q`$ connected components, then the multiplicity of $`{\lambda_{N}\hspace{0pt}{(G)}} = N`$ is $`q - 1`$.
\(iv\) $`{{{\lambda_{i}\hspace{0pt}{(G^{c})}} + {\lambda_{{N - i} + 2}\hspace{0pt}{(G)}}} = N},{2 \leq i \leq N}`$.
For a given graph, generally its largest eigenvalue is easier to compute than the smallest nonzero eigenvalue. Hence, by Lemma 4, one can obtain the smallest nonzero eigenvalue by computing the largest eigenvalue of its complementary graph.
Corollary 1 For any given graph $`G`$ of size $`N`$, if its second smallest eigenvalue equals its smallest node degree, i.e., $`{\lambda_{2}\hspace{0pt}{(G)}} = {d_{m\hspace{0pt}i\hspace{0pt}n}\hspace{0pt}{(G)}}`$, then $`G`$ or $`G^{c}`$ is disconnected; if $`{\lambda_{2}\hspace{0pt}{(G)}} > {d_{m\hspace{0pt}i\hspace{0pt}n}\hspace{0pt}{(G)}}`$, then $`G`$ is a complete graph; if both $`G`$ and $`G^{c}`$ are connected, then $`{{\lambda_{2}\hspace{0pt}{(G)}} < {d_{m\hspace{0pt}i\hspace{0pt}n}\hspace{0pt}{(G)}}}.`$
Proof If $`G^{c}`$ is connected, then $`{\lambda_{m\hspace{0pt}a\hspace{0pt}x}\hspace{0pt}{(G^{c})}} = {N - {\lambda_{2}\hspace{0pt}{(G)}}} = {N - {d_{m\hspace{0pt}i\hspace{0pt}n}\hspace{0pt}{(G_{1})}}} = {{d_{m\hspace{0pt}a\hspace{0pt}x}\hspace{0pt}{(G^{c})}} + 1}`$. By Lemma 2, $`{d_{m\hspace{0pt}a\hspace{0pt}x}\hspace{0pt}{(G^{c})}} = {N - 1}`$, so $`G`$ is disconnected. Further, if $`{\lambda_{2}\hspace{0pt}{(G)}} > {d_{m\hspace{0pt}i\hspace{0pt}n}\hspace{0pt}{(G)}}`$, by Lemma 4, $`{\lambda_{N}\hspace{0pt}{(G^{c})}} < {d_{m\hspace{0pt}a\hspace{0pt}x} + 1}`$. Combining with Lemma 2, $`G^{c}`$ can only have isolated nodes, i.e., $`G`$ is a complete graph. The third statement holds obviously. $`\square`$
Combining Lemmas 1 and 2 with Corollary 1, one can easily get the following result.
Corollary 2 For any given connected graph $`G`$ and every its induced subgraph $`G_{1}`$, one has $`{\lambda_{m\hspace{0pt}a\hspace{0pt}x}\hspace{0pt}{(G)}} \geq {\lambda_{m\hspace{0pt}a\hspace{0pt}x}\hspace{0pt}{(G_{1})}}`$, so the synchronizability index of $`G`$ satisfies $`{{r\hspace{0pt}{(G)}} < \frac{d_{m\hspace{0pt}i\hspace{0pt}n}\hspace{0pt}{(G)}}{\lambda_{m\hspace{0pt}a\hspace{0pt}x}\hspace{0pt}{(G_{1})}}};`$ if both $`G`$ and $`G^{c}`$ are connected, then $`{{r\hspace{0pt}{(G)}} < \frac{d_{m\hspace{0pt}i\hspace{0pt}n}}{d_{m\hspace{0pt}a\hspace{0pt}x} + 1}}.`$
Since subgraphs have less nodes, this corollary is useful when a graph $`G`$ contains some canonical graphs whose largest eigenvalues can be easily obtained as subgraphs (see the section below for further discussion).
Corollary 3 For a given graph $`G`$, if the largest eigenvalue of $`G^{c}`$ is $`\lambda_{m\hspace{0pt}a\hspace{0pt}x} = {{d_{m\hspace{0pt}a\hspace{0pt}x}\hspace{0pt}{(G^{c})}} + \alpha}`$, then $`{\lambda_{2}\hspace{0pt}{(G)}} = {{{d_{m\hspace{0pt}i\hspace{0pt}n}\hspace{0pt}{(G)}} + 1} - \alpha}`$. Consequently, the synchronizability index of $`G`$ satisfies $`{{r\hspace{0pt}{(G)}} = \frac{{{d_{m\hspace{0pt}i\hspace{0pt}n}\hspace{0pt}{(G)}} + 1} - \alpha}{\lambda_{m\hspace{0pt}a\hspace{0pt}x}\hspace{0pt}{(G)}} \leq \frac{{{d_{m\hspace{0pt}i\hspace{0pt}n}\hspace{0pt}{(G)}} + 1} - \alpha}{{d_{m\hspace{0pt}a\hspace{0pt}x}\hspace{0pt}{(G)}} + 1}}.`$
By Lemma 2, generally $`\alpha \geq 1`$, so the bound in Corollary 3 is better than the one in Corollary 2.
## 3 Graphs having cycles as subgraphs
In this section, an even cycle means a cycle with an even number of nodes and an odd cycle means a cycle with an odd number of nodes. By the discussion in the above section, one can get the following results.
Lemma 5 For any even cycle, $`- 2`$ is an eigenvalue of its adjacency matrix. For any odd cycle with $`n_{1}`$ nodes, $`{- 1} + \frac{\sin{(\frac{3\hspace{0pt}{({n_{1} - 1})}\hspace{0pt}\pi}{2\hspace{0pt}n_{1}})}}{\sin{(\frac{{({n_{1} - 1})}\hspace{0pt}\pi}{2\hspace{0pt}n_{1}})}}`$ is an eigenvalue of its adjacency matrix.
Proof Lemma 3 directly leads to the result. $`\square`$
Theorem 1 For any given graph $`G`$, suppose $`G_{1}`$ is its induced subgraph composing of all nodes of $`G`$ with the maximum degree $`d_{m\hspace{0pt}a\hspace{0pt}x}\hspace{0pt}{(G)}`$, and $`\mathcal{G}_{1}`$ is the induced subgraph of $`G^{c}`$ composing of all nodes of $`G^{c}`$ with the maximum degree $`d_{m\hspace{0pt}a\hspace{0pt}x}\hspace{0pt}{(G^{c})}`$. Then, the following statements hold:
\(i\) If both $`G_{1}`$ and $`\mathcal{G}_{1}`$ have even cycles as induced subgraphs, then $`{\lambda_{m\hspace{0pt}a\hspace{0pt}x}\hspace{0pt}{(G)}} \geq {{d_{m\hspace{0pt}a\hspace{0pt}x}\hspace{0pt}{(G)}} + 2}`$ and $`{\lambda_{m\hspace{0pt}a\hspace{0pt}x}\hspace{0pt}{(G^{c})}} \geq {{d_{m\hspace{0pt}a\hspace{0pt}x}\hspace{0pt}{(G^{c})}} + 2}`$. Consequently, the synchronizability index of $`G`$ satisfies $`{r\hspace{0pt}{(G)}} \leq \frac{{d_{m\hspace{0pt}i\hspace{0pt}n}\hspace{0pt}{(G)}} - 1}{{d_{m\hspace{0pt}a\hspace{0pt}x}\hspace{0pt}{(G)}} + 2}`$.
\(ii\) If both $`G_{1}`$ and $`\mathcal{G}_{1}`$ have odd cycles as induced subgraphs, and if the longest odd cycle of $`G_{1}`$ has $`n_{1}`$ nodes and the longest odd cycle of $`\mathcal{G}_{1}`$ has $`n_{2}`$ nodes, then $`{\lambda_{m\hspace{0pt}a\hspace{0pt}x}\hspace{0pt}{(G)}} \geq {{{d_{m\hspace{0pt}a\hspace{0pt}x}\hspace{0pt}{(G)}} + 1} - \frac{\sin{(\frac{3\hspace{0pt}{({n_{1} - 1})}\hspace{0pt}\pi}{2\hspace{0pt}n_{1}})}}{\sin{(\frac{{({n_{1} - 1})}\hspace{0pt}\pi}{2\hspace{0pt}n_{1}})}}}`$ and $`{\lambda_{m\hspace{0pt}a\hspace{0pt}x}\hspace{0pt}{(G^{c})}} \geq {{{d_{m\hspace{0pt}a\hspace{0pt}x}\hspace{0pt}{(G^{c})}} + 1} - \frac{\sin{(\frac{3\hspace{0pt}{({n_{2} - 1})}\hspace{0pt}\pi}{2\hspace{0pt}n_{2}})}}{\sin{(\frac{{({n_{2} - 1})}\hspace{0pt}\pi}{2\hspace{0pt}n_{2}})}}}`$. Consequently, the synchronizability index of $`G`$ satisfies $`{r\hspace{0pt}{(G)}} \leq \frac{{d_{m\hspace{0pt}i\hspace{0pt}n}\hspace{0pt}{(G)}} + \frac{\sin{(\frac{3\hspace{0pt}{({n_{2} - 1})}\hspace{0pt}\pi}{2\hspace{0pt}n_{2}})}}{\sin{(\frac{{({n_{2} - 1})}\hspace{0pt}\pi}{2\hspace{0pt}n_{2}})}}}{{{d_{m\hspace{0pt}a\hspace{0pt}x}\hspace{0pt}{(G)}} + 1} - \frac{\sin{(\frac{3\hspace{0pt}{({n_{1} - 1})}\hspace{0pt}\pi}{2\hspace{0pt}n_{1}})}}{\sin{(\frac{{({n_{1} - 1})}\hspace{0pt}\pi}{2\hspace{0pt}n_{1}})}}}`$.
\(iii\) If $`G_{1}`$ has an even cycle as an induced subgraph, and $`\mathcal{G}_{1}`$ has odd cycles as induced subgraphs with the longest odd cycle having $`n_{2}`$ nodes, then $`{\lambda_{m\hspace{0pt}a\hspace{0pt}x}\hspace{0pt}{(G)}} \geq {{d_{m\hspace{0pt}a\hspace{0pt}x}\hspace{0pt}{(G)}} + 2}`$ and $`{\lambda_{m\hspace{0pt}a\hspace{0pt}x}\hspace{0pt}{(G^{c})}} \geq {{{d_{m\hspace{0pt}a\hspace{0pt}x}\hspace{0pt}{(G^{c})}} + 1} - \frac{\sin{(\frac{3\hspace{0pt}{({n_{2} - 1})}\hspace{0pt}\pi}{2\hspace{0pt}n_{2}})}}{\sin{(\frac{{({n_{2} - 1})}\hspace{0pt}\pi}{2\hspace{0pt}n_{2}})}}}`$. Consequently, the synchronizability index of $`G`$ satisfies
$`{r\hspace{0pt}{(G)}} \leq \frac{{d_{m\hspace{0pt}i\hspace{0pt}n}\hspace{0pt}{(G)}} + \frac{\sin{(\frac{3\hspace{0pt}{({n_{2} - 1})}\hspace{0pt}\pi}{2\hspace{0pt}n_{2}})}}{\sin{(\frac{{({n_{2} - 1})}\hspace{0pt}\pi}{2\hspace{0pt}n_{2}})}}}{{d_{m\hspace{0pt}a\hspace{0pt}x}\hspace{0pt}{(G)}} + 2}`$.
\(iv\) If $`G_{1}`$ has odd cycles as induced subgraphs with the longest odd cycle having $`n_{1}`$ nodes and $`\mathcal{G}_{1}`$ has an even cycle as an induced subgraph, then $`{\lambda_{m\hspace{0pt}a\hspace{0pt}x}\hspace{0pt}{(G)}} \geq {{{d_{m\hspace{0pt}a\hspace{0pt}x}\hspace{0pt}{(G)}} + 1} - \frac{\sin{(\frac{3\hspace{0pt}{({n_{1} - 1})}\hspace{0pt}\pi}{2\hspace{0pt}n_{1}})}}{\sin{(\frac{{({n_{1} - 1})}\hspace{0pt}\pi}{2\hspace{0pt}n_{1}})}}}`$ and $`{\lambda_{m\hspace{0pt}a\hspace{0pt}x}\hspace{0pt}{(G^{c})}} \geq {{d_{m\hspace{0pt}a\hspace{0pt}x}\hspace{0pt}{(G^{c})}} + 2}`$. Consequently, the synchronizability index of $`G`$ satisfies
$`{r\hspace{0pt}{(G)}} \leq \frac{{d_{m\hspace{0pt}i\hspace{0pt}n}\hspace{0pt}{(G)}} - 1}{{{d_{m\hspace{0pt}a\hspace{0pt}x}\hspace{0pt}{(G)}} + 1} - \frac{\sin{(\frac{3\hspace{0pt}{({n_{1} - 1})}\hspace{0pt}\pi}{2\hspace{0pt}n_{1}})}}{\sin{(\frac{{({n_{1} - 1})}\hspace{0pt}\pi}{2\hspace{0pt}n_{1}})}}}`$.
Proof (i) By Lemmas 3 and 5, for any even cycle, its largest eigenvalue is 4 and -2 is an eigenvalue of its adjacency matrix. Let $`L_{1}`$ be the sub-matrix of the Laplacian matrix of $`G`$ related to all the nodes in $`G_{1}`$, and $`A\hspace{0pt}{(G_{1})}`$ be the adjacency matrix of $`G_{1}`$. Then, $`A\hspace{0pt}{(G_{1})}`$ contains a sub-matrix which is the adjacency matrix of the corresponding even cycle. Therefore, one has
<table>
<tbody>
<tr>
<td></td>
<td>(dmax(G)+2)I−L1=2I+A(G1)≯0.subscript𝑑𝑚𝑎𝑥𝐺2𝐼subscript𝐿12𝐼𝐴subscript𝐺1not-greater-than0(d_{max}(G)+2)I-L_{1}=2I+A(G_{1})\not>0.</td>
<td></td>
</tr>
</tbody>
</table>
In fact, the sub-matrix of $`{2\hspace{0pt}I} + {A\hspace{0pt}{(G_{1})}}`$ corresponding to the even cycle has a zero eigenvalue. This implies that the largest eigenvalue of $`G_{1}`$ is larger than or equal to $`d_{m\hspace{0pt}a\hspace{0pt}x} + 2`$. Then, by Lemma 1, $`{\lambda_{m\hspace{0pt}a\hspace{0pt}x}\hspace{0pt}{(G)}} \geq {d_{m\hspace{0pt}a\hspace{0pt}x} + 2}`$. Similarly, $`{\lambda_{m\hspace{0pt}a\hspace{0pt}x}\hspace{0pt}{(G^{c})}} \geq {d_{m\hspace{0pt}a\hspace{0pt}x} + 2}`$.
Further, suppose the number of nodes of $`G`$ is $`N`$. Then, the smallest nonzero eigenvalue of $`G`$ is $`\lambda_{2} = {N - {\lambda_{m\hspace{0pt}a\hspace{0pt}x}\hspace{0pt}{(G^{c})}}}`$ and the minimum node degree of $`G`$ is $`{d_{m\hspace{0pt}i\hspace{0pt}n}\hspace{0pt}{(G)}} = {N - 1 - {d_{m\hspace{0pt}a\hspace{0pt}x}\hspace{0pt}{(G^{c})}}}`$. If $`{\lambda_{m\hspace{0pt}a\hspace{0pt}x}\hspace{0pt}{(G^{c})}} \geq {{d_{m\hspace{0pt}a\hspace{0pt}x}\hspace{0pt}{(G^{c})}} + 2}`$, then $`\lambda_{2} \leq {N - {d_{m\hspace{0pt}a\hspace{0pt}x}\hspace{0pt}{(G^{c})}} - 2} = {{d_{m\hspace{0pt}i\hspace{0pt}n}\hspace{0pt}{(G)}} - 1}`$. Therefore, Theorem 1 (i) obviously holds.
(ii) Follows easily from (i) and Lemma 5.
\(iii\) and (iv) can be similarly proved. $`\square`$
Remark 1 For a given graph $`G`$, clearly, the bounds given in Theorem 1 for the synchronizability index $`r\hspace{0pt}{(G)}`$ are better than the general bound $`{r\hspace{0pt}{(G)}} \leq \frac{d_{m\hspace{0pt}i\hspace{0pt}n}}{d_{m\hspace{0pt}a\hspace{0pt}x}}`$. And the existence of cycles can be easily tested by drawing graphs. Since the networks with good synchronizabilities always have homogeneous degree distributions \[[6](#bib.bib6), [10](#bib.bib10)\], Theorem 1 is especially useful for estimating the synchronizabilities of homogeneous networks.
Remark 2 The smallest even cycle is cycle $`C_{4}`$, and its complementary graph $`C_{4}^{c}`$ has two separated edges. $`C_{4}`$ and $`C_{4}^{c}`$ are very important in graph theory \[[15](#bib.bib15)\]. A graph has a $`C_{4}`$ as an induced subgraph if and only if $`G^{c}`$ has $`C_{4}^{c}`$ as an induced subgraph.
Remark 3 The cycle $`C_{3}`$ is the smallest cycle. If a graph $`G`$ contains only $`C_{3}`$ cycles, its largest eigenvalue satisfies $`{{\lambda_{m\hspace{0pt}a\hspace{0pt}x}\hspace{0pt}{(G)}} \geq {d_{m\hspace{0pt}a\hspace{0pt}x} + 1}}.`$
Example 1 Consider cycle $`C_{5}`$ in Fig. 1. Its complementary graph is also $`C_{5}`$ (see Fig. 2). Testing the eigenvalues of $`C_{5}`$ and its synchronizability, one finds that they attain the exact bounds in Theorem 1 (ii).
$`\Leftrightarrow`$
Fig. 1 Graph $`C_{5}`$ Fig. 2 Graph $`C_{5}^{c}`$
Example 2 Consider graph $`\Gamma_{1}`$ in Fig. 3. Its complementary graph is $`C_{6}`$ (see Fig. 4). Testing the eigenvalues of $`\Gamma_{1}`$ and its synchronizability, one finds that they attain the exact bounds in Theorem 1 (i).
$`\Leftrightarrow`$
Fig. 3 Graph $`\Gamma_{1}`$ Fig. 4 Graph $`\Gamma_{1}^{c} = C_{6}`$
## 4 Applications of other subgraphs
According to the discussion in the above section, cycles are very important for estimating the network synchronizability. In this section, consider the estimation of the largest eigenvalue of a given graph by its subgraphs. First, the following lemma for adjacency matrices is needed.
Lemma 6 \[[17](#bib.bib17)\] Given a graph $`G`$, the smallest eigenvalue of its adjacency matrix $`A\hspace{0pt}{(G)}`$ satisfies $`{{\lambda_{m\hspace{0pt}i\hspace{0pt}n}\hspace{0pt}{({A\hspace{0pt}{(G)}})}} \leq {{d_{m\hspace{0pt}a\hspace{0pt}x}\hspace{0pt}{(G)}} - {\lambda_{m\hspace{0pt}a\hspace{0pt}x}\hspace{0pt}{(G)}}}}.`$
By this lemma, one can get the following result.
Theorem 2 For a given graph $`G`$, let $`H`$ be a subgraph of $`G`$ containing all nodes of $`G`$ with the same node degree $`d`$. Let $`H_{1}`$ be a subgraph of $`H`$. Then, the largest eigenvalue of $`G`$ satisfies $`{{\lambda_{m\hspace{0pt}a\hspace{0pt}x}\hspace{0pt}{(G)}} \geq {{d + {\lambda_{m\hspace{0pt}a\hspace{0pt}x}\hspace{0pt}{(H_{1})}}} - {d_{m\hspace{0pt}a\hspace{0pt}x}\hspace{0pt}{(H_{1})}}}}.`$
Proof It follows from Lemma 6 and the proof of Theorem 1. $`\square`$
Remark 4 In the above section, cycles as subgraphs have the same node degrees. So the smallest eigenvalues of the adjacency matrices of cycles can be exactly computed. In Theorem 2, for a general subgraph $`H_{1}`$, its node degrees are not necessarily the same. In this case, Lemma 6 is very useful for estimating the largest eigenvalues.
In what follows, consider some canonical subgraphs which can be used to estimate the largest eigenvalues of the original graphs. All the results can be combined with the results in the above section to obtain upper bounds for the network synchronizability.
### 4.1 Graphs having chains as subgraphs
A chain is a nearest-neighbor coupled graph without an edge between the first and the last nodes. For a given chain $`P_{N}`$ with $`N`$ nodes, it is well known that its largest and smallest nonzero eigenvalues are $`2\hspace{0pt}{({1 + {\cos{(\frac{\pi}{N})}}})}`$ and $`2\hspace{0pt}{({1 - {\cos{(\frac{\pi}{N})}}})}`$, respectively \[[8](#bib.bib8), [14](#bib.bib14), [27](#bib.bib27)\]. Obviously, the largest eigenvalue of a chain converges to $`4`$ as its length tends to $`+ \infty`$. According to the above discussion, one can establish the following result.
Theorem 3 For a given graph $`G`$, let $`G_{1}`$ be a subgraph of $`G`$ containing all nodes of $`G`$ with the maximum degree. Suppose $`G_{1}`$ contains a chain $`P_{k}`$ with $`k`$ nodes as its subgraph. Then, the largest eigenvalue of $`G`$ satisfies $`{{\lambda_{m\hspace{0pt}a\hspace{0pt}x}\hspace{0pt}{(G)}} \geq {d_{m\hspace{0pt}a\hspace{0pt}x} + {2\hspace{0pt}{\cos{(\frac{\pi}{N})}}}}}.`$
Proof Theorems 1 and 2 directly lead to the conclusion. $`\square`$
Note that chain $`P_{4}`$ is a very important graph in graph theory \[[15](#bib.bib15)\]. The complementary graph of $`P_{4}`$ is still a $`P_{4}`$ graph. Therefore, a graph $`G`$ has $`P_{4}`$ as a subgraph if and only if $`G^{c}`$ has $`P_{4}`$ as a subgraph.
### 4.2 Graphs having bipartite graphs as subgraphs
Let $`G_{1}\hspace{0pt}{(\mathcal{V}_{1},\mathcal{E}_{1})}`$ and $`G_{2}\hspace{0pt}{(\mathcal{V}_{2},\mathcal{E}_{2})}`$ be two graphs on disjoint sets of $`n_{1}`$ and $`n_{2}`$ nodes, respectively, where $`\mathcal{V}_{i}`$ and $`\mathcal{E}_{i}`$ are the corresponding sets of nodes and edges. Their disjoint union $`G_{1} + G_{2}`$ is the graph $`{{G_{1} + G_{2}} = {({\mathcal{V}_{1}\hspace{0pt}{\bigcup\mathcal{V}_{2}}},{\mathcal{E}_{1}\hspace{0pt}{\bigcup\mathcal{E}_{2}}})}}.`$ Then, a bipartite graph $`G_{1} \ast G_{2}`$ generated by $`G_{1}`$ and $`G_{2}`$ is a new graph composing of $`G_{1} + G_{2}`$ and the new edges connecting each node of $`G_{1}`$ to every node of $`G_{2}`$, as detailed in \[[1](#bib.bib1), [15](#bib.bib15), [20](#bib.bib20)\]. It is well known that the largest eigenvalue of the bipartite graph $`G_{1} \ast G_{2}`$ is $`{{\lambda_{m\hspace{0pt}a\hspace{0pt}x}\hspace{0pt}{({G_{1} \ast G_{2}})}} = {n_{1} + n_{2}}},`$ i.e., the sum of the numbers of nodes of $`G_{1}`$ and $`G_{2}`$. In fact, the complementary graphs of bipartite graphs are always disconnected.
By Lemma 6, for a bipartite graph $`G_{1} \ast G_{2}`$, as discussed above, one can has $`{{\lambda_{m\hspace{0pt}i\hspace{0pt}n}\hspace{0pt}{({A\hspace{0pt}{({G_{1} \ast G_{2}})}})}} \leq {{d_{m\hspace{0pt}a\hspace{0pt}x}\hspace{0pt}{({G_{1} \ast G_{2}})}} - n_{1} - n_{2}}}.`$
Theorem 4 For a given graph $`G`$, let $`H`$ be a subgraph of $`G`$ containing all nodes of $`G`$ with the same degree $`d`$. Suppose $`H`$ contains a bipartite subgraph $`H_{1} \ast H_{2}`$, and the numbers of nodes of $`H_{1}`$ and $`H_{2}`$ are $`n_{1}`$ and $`n_{2}`$, respectively. Then, the largest eigenvalue of $`G`$ satisfies $`{{\lambda_{m\hspace{0pt}a\hspace{0pt}x}\hspace{0pt}{(G)}} \geq {{d + n_{1} + n_{2}} - {d_{m\hspace{0pt}a\hspace{0pt}x}\hspace{0pt}{({H_{1} \ast H_{2}})}}}}.`$
Proof This is a direct result of Theorem 2 for bipartite subgraphs. $`\square`$
Example 3 Consider graph $`\Gamma_{2}`$ in Fig. 5. Obviously, $`\Gamma_{2}`$ has a bipartite graph $`H`$ as its subgraph which is composed of all nodes with degree 6 from $`\Gamma_{2}`$. The largest eigenvalue of this bipartite graph is 8. So, by Theorem 4, $`{{\lambda_{m\hspace{0pt}a\hspace{0pt}x}\hspace{0pt}{(\Gamma_{2})}} \geq 9}.`$ On the other hand, the maximum degree of the complementary graph $`\Gamma_{2}^{c}`$ of $`\Gamma_{2}`$ is 8. And the nodes with degree 8 in $`\Gamma_{2}^{c}`$ form a cycle $`C_{4}`$. By Theorem 1, the smallest nonzero eigenvalue of $`\Gamma_{2}`$ satisfies $`{{\lambda_{2}\hspace{0pt}{(\Gamma_{2})}} \leq {{d_{m\hspace{0pt}i\hspace{0pt}n}\hspace{0pt}{(\Gamma_{2})}} - 1} = 2}.`$ So, $`{{r\hspace{0pt}{(\Gamma_{2})}} \leq \frac{2}{9}}.`$ Simply computing the Laplacian eigenvalues of $`\Gamma_{2}`$, one has $`\lambda_{2} = 1.7251`$ and $`{\lambda_{m\hspace{0pt}a\hspace{0pt}x} = 9.2749}.`$ Consequently, $`{{r\hspace{0pt}{(\Gamma_{2})}} = \frac{1.7251}{9.2749} \approx 0.176}.`$ Therefore, the theorems in this paper successfully give the upper integer of the largest eigenvalue and the lower integer of the smallest nonzero eigenvalue of $`\Gamma_{2}`$.
Fig. 5 Graph $`\Gamma_{2}`$
### 4.3 Graphs having product graphs as subgraphs
Let $`G_{1}\hspace{0pt}{(\mathcal{V}_{1},\mathcal{E}_{1})}`$ and $`G_{2}\hspace{0pt}{(\mathcal{V}_{2},\mathcal{E}_{2})}`$ be two nonempty graphs. Their product graph $`G_{1} \times G_{2}`$ is a graph with node set $`\mathcal{V}_{1} \times \mathcal{V}_{2}`$, and $`{(x_{1},x_{2})}\hspace{0pt}{(y_{1},y_{1})}`$ is an edge in $`\mathcal{E}\hspace{0pt}{({G_{1} \times G_{2}})}`$ if and only if either $`x_{2} = y_{2}`$ with $`{x_{1}\hspace{0pt}y_{1}} \in {\mathcal{E}\hspace{0pt}{(G_{1})}}`$ or $`x_{1} = y_{1}`$ with $`{x_{2}\hspace{0pt}y_{2}} \in {\mathcal{E}\hspace{0pt}{(G_{2})}}`$. One may view $`G_{1} \times G_{2}`$ as the graph obtained from $`G_{1}`$ by replacing each of its nodes with a copy of $`G_{2}`$ and each of its edges with the number of nodes of $`G_{2}`$ edges joining the corresponding nodes of $`G_{2}`$ in two copies, as detailed in \[[1](#bib.bib1), [15](#bib.bib15)\].
It is well known that the largest eigenvalue of the product graph $`G_{1} \times G_{2}`$ is $`{{\lambda_{m\hspace{0pt}a\hspace{0pt}x}\hspace{0pt}{({G_{1} \times G_{2}})}} = {{\lambda_{m\hspace{0pt}a\hspace{0pt}x}\hspace{0pt}{(G_{1})}} + {\lambda_{m\hspace{0pt}a\hspace{0pt}x}\hspace{0pt}{(G_{2})}}}}.`$
Theorem 5 For a given graph $`G`$, let $`H`$ be a subgraph of $`G`$ containing all nodes of $`G`$ with the same node degree $`d`$. Suppose $`H`$ contains a product graph $`H_{1} \times H_{2}`$ as its subgraph. Then, the largest eigenvalue of $`G`$ satisfies $`{{\lambda_{m\hspace{0pt}a\hspace{0pt}x}\hspace{0pt}{(G)}} \geq {{d + {\lambda_{m\hspace{0pt}a\hspace{0pt}x}\hspace{0pt}{(H_{1})}} + {\lambda_{m\hspace{0pt}a\hspace{0pt}x}\hspace{0pt}{(H_{2})}}} - {d_{m\hspace{0pt}a\hspace{0pt}x}\hspace{0pt}{({H_{1} \times H_{2}})}}}}.`$
Example 4 Consider graph $`\Gamma_{3}`$ in Fig. 6. Obviously, all nodes of $`\Gamma_{3}`$ have degree 4. And $`\Gamma_{3}`$ has a product graph $`C_{4} \times P_{3}`$ (nodes 1 to 12) as its subgraph, where $`P_{3}`$ denotes a chain with three nodes. The largest eigenvalue of this product subgraph is 7. So, by Theorem 5 or Lemma 1, $`{{\lambda_{m\hspace{0pt}a\hspace{0pt}x}\hspace{0pt}{(\Gamma_{3})}} \geq 7}.`$ On the other hand, the complementary graph $`\Gamma_{3}^{c}`$ of $`\Gamma_{3}`$ has a bipartite graph as its subgraph which is composed of nodes 1 to 4 and nodes 9 to 12. By Theorem 4, similarly to Example 3, the largest eigenvalue of $`\Gamma_{3}^{c}`$ satisfies $`{\lambda\hspace{0pt}{(\Gamma_{3}^{c})}} \geq {{d_{m\hspace{0pt}a\hspace{0pt}x}\hspace{0pt}{(\Gamma_{3}^{c})}} + 3}`$. Thus, by Corollary 3, the smallest nonzero eigenvalue of $`\Gamma_{3}`$ satisfies $`{{\lambda_{2}\hspace{0pt}{(\Gamma_{3})}} \leq {{d_{m\hspace{0pt}i\hspace{0pt}n}\hspace{0pt}{(\Gamma_{2})}} - 2} = 2}.`$ So, $`{{r\hspace{0pt}{(\Gamma_{3})}} \leq \frac{2}{7}}.`$ Simply computing the Laplacian eigenvalues of $`\Gamma_{3}`$, one has $`\lambda_{2} = 1.2679`$ and $`{\lambda_{m\hspace{0pt}a\hspace{0pt}x} = 7.4142}.`$ Consequently, $`{{r\hspace{0pt}{(\Gamma_{2})}} = \frac{1.2679}{7.4142} \approx 0.171}.`$ Similarly to Example 3, the corresponding theorems proved in this paper successfully give the upper integer of the largest eigenvalue and the lower integer of the smallest nonzero eigenvalue of $`\Gamma_{3}`$.
Fig. 6 Graph $`\Gamma_{3}`$
### 4.4 The maximum disconnected subgraph
Given a graph $`G`$ of size $`N`$, suppose $`G_{1}`$ is an induced subgraph of $`G`$ of size $`n_{1}`$, and $`G_{1}`$ is disconnected. $`G_{1}`$ is called a maximum disconnected subgraph, if the node number of any other disconnected subgraph of $`G`$ is less than or equal to $`n_{1}`$.
Theorem 5 For a given connected graph $`G`$ of size $`N`$, if the node number of its maximum disconnected subgraph is $`n_{1}`$, then the smallest nonzero eigenvalue of $`G`$ satisfies $`\lambda_{2} \leq {N - n_{1}}`$. Consequently, $`{{r\hspace{0pt}{(G)}} \leq \frac{N - n_{1}}{{d_{m\hspace{0pt}a\hspace{0pt}x}\hspace{0pt}{(G)}} + 1}}.`$
Proof Suppose $`G_{1}`$ is a maximum disconnected subgraph of size $`n_{1}`$. Then, by Lemma 4, $`{\lambda_{m\hspace{0pt}a\hspace{0pt}x}\hspace{0pt}{(G_{1}^{c})}} \geq n_{1}`$. Further, $`{\lambda_{m\hspace{0pt}a\hspace{0pt}x}\hspace{0pt}{(G^{c})}} \geq n_{1}`$, so $`{\lambda_{2}\hspace{0pt}{(G)}} \leq {N - n_{1}}`$. $`\square`$
Example 5 Consider graph $`\Gamma_{4}`$ in Fig. 7. By deleting node 3 or 6, one can verify that the node number of its maximum disconnected subgraph is 7. So, $`{\lambda_{2}\hspace{0pt}{(\Gamma_{4})}} \leq 1`$. Combining Lemma 2 and Theorem 6, one has $`{{r\hspace{0pt}{(\Gamma_{4})}} < \frac{1}{5}}.`$ It is well known that graphs as in Fig. 7 have large node and edge betweenness centralities therefore have bad synchronizabilities. Theorem 6, based on the theory of subgraphs and complementary graphs, gives an explanation why such graphs indeed have bad synchronizabilities.
Fig. 7 Graph $`\Gamma_{4}`$
Remark 5 From Examples 3 to 5, it can be seen that actually one does not need to draw out the complementary graphs; the properties of complementary graphs can be analyzed using the original graphs.
## 5 Conclusion
In this paper, some relatively tight upper bounds have been obtained for the network synchronizability based on the theory of subgraphs and complementary graphs. Especially, some sharp bounds are given for the eigenratioes and the largest eigenvalues when a graph has some canonical subgraphs, such as cycles, chains, bipartite graphs and product graphs. Considering that the networks with good synchronizabilities typically have homogeneous degree distributions, the results obtained in this paper are particularly useful for estimating the synchronizabilities of homogeneous networks. And for a given network, its corresponding graph as well as its complementary graph and subgraphs can be easily obtained. This paper shows that better understanding and helpful manipulation of subgraphs and complementary graphs are very useful for enhancing the network synchronizability. Therefore, the graph-theoretical method provided in this paper is deemed important in the study of network synchronization problems.
## References
- \[1\]
F. M. Atay, T. Bıyıkoğlu. Graph operations and synchronization of complex networks, Phys. Rev. E, 72: 016217, 2005.
- \[2\]
F. M. Atay, T. Bıyıkoğlu, J. Jost. Network synchronization: Spectral versus statistical properties, Physica D, 224: 35-41, 2006.
- \[3\]
M. Barahona, L. M. Pecora. Synchronization in small-world systems, Phys. Rev. Lett., 89(5): 054101, 2002.
- \[4\]
I. V. Belykh, E. Lange, M. Hasler. Synchronization of bursting neurons: what matters in the network topology, Phys. Rev. Lett., 94: 188101, 2005.
- \[5\]
S. Boccaletti, V. Latora, Y. Moreno, M. Chavez, D. U. Hwang. Complex networks: structure and dynamics, Physics Reports, 424: 175-308, 2006.
- \[6\]
L. Donetti, P. I. Hurtado, M. A. Muñoz. Entangled networks, synchronization, and optimal network topology, Phys. Rev. Lett., 95: 188701, 2005.
- \[7\]
Z. S. Duan, G. Chen, L. Huang. Analysis and control of network synchronizability, arXiv:0706.2901v1, 2007.
- \[8\]
C. Godsil, G. Royle. Algebric Graph Theory, Springer-Verlag, New York, 2001.
- \[9\]
P. Holme, B. J. Kim. Vertex overload breakdown in evolving networks. Phys. Rev. E, 65: 066109, 2002.
- \[10\]
H. Hong, B. J. Kim, M. Y. Choi, H. Park. Factors that predict better synchronizability on complex networks, Phys. Rev. E, 69: 067105, 2004.
- \[11\]
L. Kocarev, P. Amato. Synchronization in power-law networks, Chaos, 15: 024101, 2005.
- \[12\]
W. Lu, T. Chen. Synchronization analysis of linearly coupled networks of discrete time systems, Physica D, 198: 148-168, 2004.
- \[13\]
J. H. Lü, X. H. Yu, G. Chen, D. Z. Cheng. Characterizing the synchronizability of small-world dynamical networks, IEEE. Trans. Circuits Syst.-I , 51(4), 787-796, 2004.
- \[14\]
R. Merris. Laplacian matrices of graph: a survey, Linear Algebra and its Applications, 197-198: 143-176, 1994.
- \[15\]
R. Merris. Laplacian graph eigenvectors, Linear Algebra and its Applications, 278: 221-236, 1998.
- \[16\]
A. E. Motter, C. S. Zhou, J. Kurths. Enhancing complex-network synchronization, Europhysics Letters, 69(3): 334-340, 2005.
- \[17\]
V. Nikiforov. Eigenvalues and extremal degrees of graphs, Linear Algebra and its Applications, 419: 735-738, 2006.
- \[18\]
T. Nishikawa, A. E. Motter. Maximum peformance at minimum cost in network synchronization, Physica D, 224: 77-89, 2006.
- \[19\]
T. Nishikawa, A. E. Motter, Y. C. Lai, F. C. Hoppensteadt. Heterogeneity in oscillator netwroks: Are smaller worlds easier to synchronize? Phys. Rev. Lett., 91(1): 014101, 2003.
- \[20\]
Y. L. Pan. Sharp upper bounds for the Laplacian graph eigenvalues, Linear Algebra and its Applications, 355: 287-295, 2002.
- \[21\]
L. M. Pecora, T. L. Carroll. Master stability functions for synchronized coupled systems, Phys. Rev. Lett., 80(10): 2109-2112, 1998.
- \[22\]
F. Sorrentino, M. di Bernardo, F. Garofalo, G. Chen. Controllability of complex networks via pinning, Phys. Rev. E, 75: 046103, 2007.
- \[23\]
A. Stefański, P. Perlikowski, T. Kapitaniak. Ragged synchronizability of coupled oscillators, Phys. Rev. E, 75: 016210, 2007.
- \[24\]
X. Wang, Y. C. Lai, C. H. Lai. Synchronization is hierchical cluster networks, arXiv. nlin. CD/0612057v1, 2006.
- \[25\]
X. F. Wang, G. Chen. Synchronization in scale-free dynamical networks: robustness and fragility, IEEE Trans. Circuits Syst-I, 49(1): 54-62, 2002.
- \[26\]
D. J. Watts, S. H. Strogatz. Collective dynamics of ‘small-world’ networks, Nature, 393(6684): 440-442, 1998.
- \[27\]
C. W. Wu. Synchronizability of networks of chaotic systems coupled via a graph with a prescribed degree sequence, Phys. Lett. A, 346: 281-287, 2005.
- \[28\]
M. Zhao, T. Zhou, B. H. Wang, G. Yan, H. J. Yang, W. J. Bai. Relations between average distance, heterogeneity and network synchronizability, Physica A, 371(2): 773-780, 2006.
- \[29\]
C. S. Zhou, J. Kurths. Dynamical weights and enhanced synchronization in adaptive complex networks, Phys. Rev. Lett., 96(16): 164102, 2006. |
0708.3014 | cs.CR | cyber | <sup>1</sup>
<sup>1</sup>institutetext: University of Zurich, Switzerland
<sup>1</sup>
<sup>1</sup>email: elisa.gorla@math.unizh.ch
<sup>2</sup>
<sup>2</sup>institutetext: Heinz Nixdorf Institute, University of Paderborn, Germany
<sup>2</sup>
<sup>2</sup>email: puttmann@hni.upb.de
<sup>3</sup>
<sup>3</sup>institutetext: Ruhr University, Bochum, Germany
<sup>3</sup>
<sup>3</sup>email: jamshid@crypto.rub.de
# Explicit formulas for efficient multiplication in $`{\mathbb{F}}_{3^{6\hspace{0pt}m}}`$
Elisa Gorla 11
Christoph Puttmann 22
Jamshid Shokrollahi 33
###### Abstract
Efficient computation of the Tate pairing is an important part of pairing-based cryptography. Recently with the introduction of the Duursma-Lee method special attention has been given to the fields of characteristic $`3`$. Especially multiplication in $`{\mathbb{F}}_{3^{6\hspace{0pt}m}}`$, where $`m`$ is prime, is an important operation in the above method. In this paper we propose a new method to reduce the number of $`{\mathbb{F}}_{3^{m}}`$-multiplications for multiplication in $`{\mathbb{F}}_{3^{6\hspace{0pt}m}}`$ from $`18`$ in recent implementations to $`15`$. The method is based on the fast Fourier transform and its explicit formulas are given. The execution times of our software implementations for $`{\mathbb{F}}_{3^{6\hspace{0pt}m}}`$ show the efficiency of our results.
Keywords: Finite field arithmetic, fast Fourier transform, Lagrange interpolation, Tate pairing computation
## 1 Introduction
Efficient multiplication in finite fields is a central task in the implementation of most public key cryptosystems. A great amount of work has been devoted to this topic (see \[[1](#bib.bib1)\] or \[[2](#bib.bib2)\] for a comprehensive list). The two types of finite fields which are mostly used in cryptographic standards are binary finite fields of type $`{\mathbb{F}}_{2^{m}}`$ and prime fields of type $`{\mathbb{F}}_{p}`$, where $`p`$ is a prime (cf. \[[3](#bib.bib3)\]). Efforts to efficiently fit finite field arithmetic into commercial processors resulted into applications of medium characteristic finite fields like those reported in \[[4](#bib.bib4)\] and \[[5](#bib.bib5)\]. Medium characteristic finite fields are fields of type $`{\mathbb{F}}_{p^{m}}`$, where $`p`$ is a prime slightly smaller than the word size of the processor, and has a special form that simplifies the modular reduction. Mersenne prime numbers constitute an example of primes which are used in this context. The security parameter is given by the length of the binary representations of the field elements, and the extension degree $`m`$ is selected appropriately. Due to security considerations, the extension degree for fields of characteristic $`2`$ or medium characteristic is usually chosen to be prime.
With the introduction of the method of Duursma and Lee for the computation of the Tate pairing (see \[[6](#bib.bib6)\]), fields of type $`{\mathbb{F}}_{3^{m}}`$ for $`m`$ prime have attracted special attention. Computing the Tate pairing on elliptic curves defined over $`{\mathbb{F}}_{3^{m}}`$ requires computations both in $`{\mathbb{F}}_{3^{m}}`$ and in $`{\mathbb{F}}_{3^{6\hspace{0pt}m}}`$. In the paper \[[7](#bib.bib7)\], calculations are implemented using the tower of extensions
<table>
<tbody>
<tr>
<td></td>
<td>𝔽3m⊂𝔽32m⊂𝔽36m.subscript𝔽superscript3𝑚subscript𝔽superscript32𝑚subscript𝔽superscript36𝑚\mathbb{F}_{3^{m}}\subset\mathbb{F}_{3^{2m}}\subset\mathbb{F}_{3^{6m}}.</td>
<td></td>
</tr>
</tbody>
</table>
Multiplications in $`{\mathbb{F}}_{3^{2\hspace{0pt}m}}`$ and $`{\mathbb{F}}_{3^{6\hspace{0pt}m}}`$ are done using $`3`$ and $`6`$ multiplications, respectively. This requires a total $`18`$ multiplications in $`{\mathbb{F}}_{3^{m}}`$. In this paper we make use of the same extension tower, using $`3`$ multiplications in $`{\mathbb{F}}_{3^{m}}`$ to multiply elements in $`{\mathbb{F}}_{3^{2\hspace{0pt}m}}`$. Since we represent the elements of $`{\mathbb{F}}_{3^{6\hspace{0pt}m}}`$ as polynomials with coefficients in $`{\mathbb{F}}_{3^{2\hspace{0pt}m}}`$, we can use Lagrange interpolation to perform the multiplication. This requires only $`5`$ multiplications in $`{\mathbb{F}}_{3^{2\hspace{0pt}m}}`$, thus reducing the total number of $`{\mathbb{F}}_{3^{m}}`$ multiplications from $`18`$ to $`15`$. The method that we propose has a slightly increased number of additions in comparison to the Karatsuba method. Notice however that for $`m > 90`$ (which is the range used in the cryptographic applications) a multiplication in $`{\mathbb{F}}_{3^{m}}`$ requires many more resources than an addition, therefore the overall resource consumption is reduced, as also shown by the results of our software experiments shown in Section [4](#S4 "4 Efficient implementation ‣ Explicit formulas for efficient multiplication in 𝔽_3^{6𝑚}").
In comparison to the classical multiplication method, the Karatsuba method (see \[[8](#bib.bib8)\], \[[9](#bib.bib9)\], and \[[7](#bib.bib7)\]) reduces the number of multiplications while introducing extra additions. Since the cost of addition grows linearly in the length of the polynomials, when the degree of the field extension gets larger multiplication will be more expensive than addition. Hence the above tradeoff makes sense. The negligibility of the cost of addition compared to that of multiplication has gone so far that the theory of multiplicative complexity of bilinear maps, especially polynomial multiplication, takes into account only the number of variable multiplications (see e.g. \[[10](#bib.bib10)\] and \[[11](#bib.bib11)\]). Obviously this theoretical model is of practical interest only when the number of additions and the costs of scalar multiplications can be kept small. A famous result in the theory of multiplicative complexity establishes a lower bound of $`{2\hspace{0pt}n} + 1`$ for the number of variable multiplications needed for the computation of the product of two polynomials of degree at most $`n`$. This lower bound can be achieved only when the field contains enough elements (see \[[12](#bib.bib12)\] or \[[13](#bib.bib13)\]). The proof of the theorem uses Lagrange evaluation-interpolation, which is also at the core of our approach. This is similar to the short polynomial multiplication (convolution) methods for complex or real numbers in \[[14](#bib.bib14)\]. In order for this method to be especially efficient, the points at which evaluation and interpolations are done are selected as primitive $`({{2\hspace{0pt}n} + 1})`$st roots of unity. In a field of type $`{\mathbb{F}}_{3^{2\hspace{0pt}m}}`$, fifth roots of unity do not exist for odd $`m`$. We overcome this problem by using fourth roots of unity instead. Notice that a primitive fourth root of unity always exist in a field of type $`{\mathbb{F}}_{3^{2\hspace{0pt}m}}.`$ We use an extra point to compute the fifth coefficient of the product. An advantage of using a primitive fourth root of unity is that the corresponding interpolation matrix will be a $`4 \times 4`$ DFT matrix, and the evaluations and interpolations can be computed using radix-$`2`$ FFT techniques (see \[[15](#bib.bib15)\] or \[[16](#bib.bib16)\]) to save some further number of additions and scalar multiplications. The current work can be considered as the continuation of that in \[[17](#bib.bib17)\] for combination of the linear-time multiplication methods with the classical or Karatsuba ones to achieve efficient polynomial multiplication formulas.
Our work is organized as follows. Section [2](#S2 "2 Multiplication using evaluation and interpolation ‣ Explicit formulas for efficient multiplication in 𝔽_3^{6𝑚}") is devoted to explaining how evaluation-interpolation can be used in general to produce short polynomial multiplication methods. In Section [3](#S3 "3 Multiplication using roots of unity ‣ Explicit formulas for efficient multiplication in 𝔽_3^{6𝑚}") we show how to apply this method to our special case, and produce explicit formulas for multiplication of polynomials of degree at most $`2`$ over $`{\mathbb{F}}_{3^{2\hspace{0pt}m}}`$. In Section [4](#S4 "4 Efficient implementation ‣ Explicit formulas for efficient multiplication in 𝔽_3^{6𝑚}") we fine-tune our method using FFT techniques, and give timing results of software implementations and also explicit multiplication formulas. Section [5](#S5 "5 Other applications of the proposed method ‣ Explicit formulas for efficient multiplication in 𝔽_3^{6𝑚}") shows how our results can be used in conjunction with the method of Duursma-Lee for computing the Tate pairing on some elliptic and hyperelliptic curves. Section [6](#S6 "6 Conclusion ‣ Explicit formulas for efficient multiplication in 𝔽_3^{6𝑚}") contains some final remarks and conclusions.
## 2 Multiplication using evaluation and interpolation
We now explain the Lagrange evaluation-interpolation for polynomials with coefficients in $`{\mathbb{F}}_{p^{m}}`$. Throughout this section $`m`$ is not assumed to be prime (in the next section we will replace $`m`$ by $`2\hspace{0pt}m`$). Let
<table>
<tbody>
<tr>
<td></td>
<td>a(z)=a0+a1z+⋯+anzn∈𝔽pm[z]𝑎𝑧subscript𝑎0subscript𝑎1𝑧⋯subscript𝑎𝑛superscript𝑧𝑛subscript𝔽superscript𝑝𝑚delimited-[]𝑧\displaystyle a(z)=a_{0}+a_{1}z+\cdots+a_{n}z^{n}\in\mathbb{F}_{p^{m}}[z]</td>
<td></td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td>b(z)=a0+a1z+⋯+anzn∈𝔽pm[z]𝑏𝑧subscript𝑎0subscript𝑎1𝑧⋯subscript𝑎𝑛superscript𝑧𝑛subscript𝔽superscript𝑝𝑚delimited-[]𝑧\displaystyle b(z)=a_{0}+a_{1}z+\cdots+a_{n}z^{n}\in\mathbb{F}_{p^{m}}[z]</td>
<td></td>
</tr>
</tbody>
</table>
be given such that
<table>
<tbody>
<tr>
<td></td>
<td>pm>2n.superscript𝑝𝑚2𝑛p^{m}>2n.</td>
<td></td>
<td>(1)</td>
</tr>
</tbody>
</table>
We represent the product of the two polynomials by
<table>
<tbody>
<tr>
<td></td>
<td>c(z)=a(z)b(z)=c0+c1z+⋯+c2nz2n𝑐𝑧𝑎𝑧𝑏𝑧subscript𝑐0subscript𝑐1𝑧⋯subscript𝑐2𝑛superscript𝑧2𝑛c(z)=a(z)b(z)=c_{0}+c_{1}z+\cdots+c_{2n}z^{2n}</td>
<td></td>
</tr>
</tbody>
</table>
and let $`e = {(e_{0},\cdots,e_{2\hspace{0pt}n})} \in {\mathbb{F}}_{p^{m}}^{{2\hspace{0pt}n} + 1}`$ be a vector with $`{2\hspace{0pt}n} + 1`$ distinct entries. Evaluation at these points is given by the map $`\phi_{e}`$
<table>
<tbody>
<tr>
<td></td>
<td>ϕe:𝔽pm[z]→𝔽pm2n+1:subscriptitalic-ϕ𝑒→subscript𝔽superscript𝑝𝑚delimited-[]𝑧superscriptsubscript𝔽superscript𝑝𝑚2𝑛1\displaystyle\phi_{e}:\mathbb{F}_{p^{m}}[z]\rightarrow\mathbb{F}_{p^{m}}^{2n+1}</td>
<td></td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td>ϕe(f)=(f(e0),⋯,f(e2n)).subscriptitalic-ϕ𝑒𝑓𝑓subscript𝑒0⋯𝑓subscript𝑒2𝑛\displaystyle\phi_{e}(f)=(f(e_{0}),\cdots,f(e_{2n})).</td>
<td></td>
</tr>
</tbody>
</table>
Let $`{A,B,C} \in {\mathbb{F}}_{p^{m}}^{{2\hspace{0pt}n} + 1}`$ denote the vectors $`(a_{0},\cdots,a_{n},0,\cdots,0)`$, $`(b_{0},\cdots,b_{n},0,\cdots,0)`$, and $`(c_{0},\cdots,c_{2\hspace{0pt}n})`$, respectively. Using the above notation we have
<table>
<tbody>
<tr>
<td></td>
<td>ϕe(a)=VeAT,ϕe(b)=VeBT,andϕe(c)=VeCT,formulae-sequencesubscriptitalic-ϕ𝑒𝑎subscript𝑉𝑒superscript𝐴𝑇formulae-sequencesubscriptitalic-ϕ𝑒𝑏subscript𝑉𝑒superscript𝐵𝑇andsubscriptitalic-ϕ𝑒𝑐subscript𝑉𝑒superscript𝐶𝑇\phi_{e}(a)=V_{e}A^{T},\;\;\;\phi_{e}(b)=V_{e}B^{T},\;\;\;\mbox{and}\;\;\;\phi_{e}(c)=V_{e}C^{T},</td>
<td></td>
</tr>
</tbody>
</table>
where $`V_{e}`$ is the Vandermonde matrix
<table>
<tbody>
<tr>
<td></td>
<td>Ve=(1e0⋯e02n1e1⋯e12n⋮⋮⋱⋮1e2n⋯e2n2n).subscript𝑉𝑒1subscript𝑒0⋯superscriptsubscript𝑒02𝑛1subscript𝑒1⋯superscriptsubscript𝑒12𝑛⋮⋮⋱⋮1subscript𝑒2𝑛⋯superscriptsubscript𝑒2𝑛2𝑛V_{e}=\left(\begin{array}[]{cccc}1&e_{0}&\cdots&e_{0}^{2n}\\ 1&e_{1}&\cdots&e_{1}^{2n}\\ \vdots&\vdots&\ddots&\vdots\\ 1&e_{2n}&\cdots&e_{2n}^{2n}\end{array}\right).</td>
<td></td>
</tr>
</tbody>
</table>
The $`{2\hspace{0pt}n} + 1`$ coefficients of the product $`{c\hspace{0pt}{(z)}} = {{{a\hspace{0pt}{(z)}} \cdot b}\hspace{0pt}{(z)}}`$ can be computed using interpolation applied to the evaluations of $`c\hspace{0pt}{(z)}`$ at the chosen $`{2\hspace{0pt}n} + 1`$ (distinct) points of $`{\mathbb{F}}_{p^{m}}`$. These evaluations can be computed by multiplying the evaluations of $`a\hspace{0pt}{(z)}`$ and $`b\hspace{0pt}{(z)}`$ at these points. This can be formally written as
<table>
<tbody>
<tr>
<td></td>
<td>ϕe(c)=ϕe(a)∗ϕe(b)subscriptitalic-ϕ𝑒𝑐subscriptitalic-ϕ𝑒𝑎subscriptitalic-ϕ𝑒𝑏\phi_{e}(c)=\phi_{e}(a)*\phi_{e}(b)</td>
<td></td>
</tr>
</tbody>
</table>
where we denote componentwise multiplication of vectors by $`\ast`$. Equivalently, if we let $`W_{e}`$ be the inverse of the matrix $`V_{e}`$, we have that
<table>
<tbody>
<tr>
<td></td>
<td>CT=We(ϕe(a)∗ϕe(b))superscript𝐶𝑇subscript𝑊𝑒subscriptitalic-ϕ𝑒𝑎subscriptitalic-ϕ𝑒𝑏C^{T}=W_{e}(\phi_{e}(a)*\phi_{e}(b))</td>
<td></td>
</tr>
</tbody>
</table>
which allows us to compute the vector $`C`$, whose entries are the coefficients of the polynomial $`c\hspace{0pt}{(z)}`$.
When condition ([1](#S2.E1 "In 2 Multiplication using evaluation and interpolation ‣ Explicit formulas for efficient multiplication in 𝔽_3^{6𝑚}")) is satisfied, the polynomial multiplication methods constructed in this way have the smallest multiplicative complexity, i.e. the number of variable multiplications in $`{\mathbb{F}}_{p^{m}}`$ achieves the lower bound $`{2\hspace{0pt}n} + 1`$ (see \[[12](#bib.bib12)\]). Indeed ([1](#S2.E1 "In 2 Multiplication using evaluation and interpolation ‣ Explicit formulas for efficient multiplication in 𝔽_3^{6𝑚}")) can be relaxed to hold even for $`p^{m} = {2\hspace{0pt}n}`$. In this case, a virtual element $`\infty`$ is added to the finite field. This corresponds to the fact that the leading coefficient of the product is the product of the leading coefficients of the factors.
Application of this method to practical situations is not straightforward, since the number of additions increases and eventually dominates the reduction in the number of multiplications. In order for this method to be efficient, $`n`$ must be much smaller than $`p^{m}`$. An instance of this occurs when computing in extensions of medium size primes (see e.g. \[[13](#bib.bib13)\]). The case of small values of $`p`$ is more complicated, even for small values of $`n`$. We recall that in this case the entries of the matrix $`V_{e}`$ are in $`{\mathbb{F}}_{p^{m}}`$ and are generally represented as polynomials of length $`m - 1`$ over $`{\mathbb{F}}_{p}`$. For multiplication of $`V_{e}`$ by vectors to be efficient, the entries of this matrix must be chosen to be sparse. However, this gives no control on the sparsity of the entries of $`W_{e}`$. Indeed one requirement for the entries of $`W_{e}`$, in the basis $`\mathcal{B}`$, to be sparse is that the inverse of the determinant of $`V_{e}`$, namely
<table>
<tbody>
<tr>
<td></td>
<td>∏0≤i,j≤2n,i≠j(ei−ej)subscriptproductformulae-sequence0𝑖formulae-sequence𝑗2𝑛𝑖𝑗subscript𝑒𝑖subscript𝑒𝑗\prod_{0\leq i,j\leq 2n,i\neq j}(e_{i}-e_{j})</td>
<td></td>
</tr>
</tbody>
</table>
has a sparse representation in $`\mathcal{B}`$. We are not aware of any method which can be used here. On the other hand, it is known that if the $`e_{i}`$’s are the elements of the geometric progression $`\omega^{i}`$, $`0 \leq i \leq {2\hspace{0pt}n}`$, and $`\omega`$ is a $`({{2\hspace{0pt}n} + 1})`$st primitive root of unity, then the inverse $`W_{e}`$ equals $`1/{({{2\hspace{0pt}n} + 1})}`$ times the Vandermonde matrix whose $`e_{i}`$’s are the elements of the geometric progression of $`\omega^{- 1}`$ (see \[[2](#bib.bib2)\]). We denote these two matrices by $`V_{\omega}`$ and $`V_{\omega^{- 1}}`$, respectively. The above fact suggests that choosing powers of roots of unity as interpolation points should enable us to control the sparsity of the entries of the corresponding Vandermonde matrix. Roots of unity are used in different contexts for multiplication of polynomials, e.g. in the FFT (see \[[2](#bib.bib2)\]) or for the construction of short multiplication methods in \[[14](#bib.bib14)\]. In the next section we discuss how to use fourth roots of unity to compute multiplication in $`{\mathbb{F}}_{p^{6\hspace{0pt}m}}`$, using only $`5`$ multiplications in $`{\mathbb{F}}_{3^{2\hspace{0pt}m}}`$.
## 3 Multiplication using roots of unity
Elements of $`{\mathbb{F}}_{3^{6\hspace{0pt}m}}`$ can be represented as polynomials of degree at most $`2`$ over $`{\mathbb{F}}_{3^{2\hspace{0pt}m}}`$. Therefore, their product is given by a polynomial of degree at most $`4`$ with coefficients in $`{\mathbb{F}}_{3^{2\hspace{0pt}m}}`$. In order to use the classical evaluation-interpolation method we would need a primitive fifth root of unity. This would require $`3^{2\hspace{0pt}m} - 1`$ to be a multiple of $`5`$, and this is never the case unless $`m`$ is even (recall that cryptographic applications require $`m`$ to be prime). However using the relation
<table>
<tbody>
<tr>
<td></td>
<td>c4=a2b2subscript𝑐4subscript𝑎2subscript𝑏2c_{4}=a_{2}b_{2}</td>
<td></td>
<td>(2)</td>
</tr>
</tbody>
</table>
we can compute the coefficients of $`c\hspace{0pt}{(x)}`$ via
<table>
<tbody>
<tr>
<td></td>
<td>(11111ωω2ω31ω21ω21ω3ω2ω)(c0c1c2c3)=(a(1)b(1)−c4a(ω)b(ω)−c4a(ω2)b(ω2)−c4a(ω3)b(ω3)−c4)11111𝜔superscript𝜔2superscript𝜔31superscript𝜔21superscript𝜔21superscript𝜔3superscript𝜔2𝜔subscript𝑐0subscript𝑐1subscript𝑐2subscript𝑐3𝑎1𝑏1subscript𝑐4𝑎𝜔𝑏𝜔subscript𝑐4𝑎superscript𝜔2𝑏superscript𝜔2subscript𝑐4𝑎superscript𝜔3𝑏superscript𝜔3subscript𝑐4\left(\begin{array}[]{cccc}1&1&1&1\\ 1&\omega&\omega^{2}&\omega^{3}\\ 1&\omega^{2}&1&\omega^{2}\\ 1&\omega^{3}&\omega^{2}&\omega\end{array}\right)\left(\begin{array}[]{c}c_{0}\\ c_{1}\\ c_{2}\\ c_{3}\end{array}\right)=\left(\begin{array}[]{c}a(1)b(1)-c_{4}\\ a(\omega)b(\omega)-c_{4}\\ a(\omega^{2})b(\omega^{2})-c_{4}\\ a(\omega^{3})b(\omega^{3})-c_{4}\end{array}\right)</td>
<td></td>
<td>(3)</td>
</tr>
</tbody>
</table>
where $`\omega`$ is a fourth root of unity. Now we apply ([2](#S3.E2 "In 3 Multiplication using roots of unity ‣ Explicit formulas for efficient multiplication in 𝔽_3^{6𝑚}")) and ([3](#S3.E3 "In 3 Multiplication using roots of unity ‣ Explicit formulas for efficient multiplication in 𝔽_3^{6𝑚}")) to find explicit formulas for multiplying two polynomials of degree at most $`2`$ over $`{\mathbb{F}}_{3^{2\hspace{0pt}m}}`$, where $`m > 2`$ is a prime.
We follow the tower representation of \[[7](#bib.bib7)\], i.e.
<table>
<tbody>
<tr>
<td></td>
<td>𝔽3m≅𝔽3[x]/(f(x))𝔽32m≅𝔽3m[y]/(y2+1)subscript𝔽superscript3𝑚subscript𝔽3delimited-[]𝑥𝑓𝑥subscript𝔽superscript32𝑚subscript𝔽superscript3𝑚delimited-[]𝑦superscript𝑦21\begin{array}[]{l}\mathbb{F}_{3^{m}}\cong\mathbb{F}_{{3}}[x]/(f(x))\\ \mathbb{F}_{3^{2m}}\cong\mathbb{F}_{3^{m}}[y]/(y^{2}+1)\\ \end{array}</td>
<td></td>
<td>(4)</td>
</tr>
</tbody>
</table>
where $`{f\hspace{0pt}{(x)}} \in {{\mathbb{F}}_{3}\hspace{0pt}{\lbrack x\rbrack}}`$ is an irreducible polynomial of degree $`m`$. Denote by $`s`$ the equivalence class of $`y`$. Note that for odd $`m > 2`$, $`4\operatorname{|\not{}}{3^{m} - 1}`$ and hence $`y^{2} + 1`$ is irreducible over $`{\mathbb{F}}_{3^{m}}`$ since the roots of $`y^{2} + 1`$ are fourth roots of unity. Let
<table>
<tbody>
<tr>
<td></td>
<td>a(z)=a0+a1z+a2z2,b(z)=b0+b1z+b2z2formulae-sequence𝑎𝑧subscript𝑎0subscript𝑎1𝑧subscript𝑎2superscript𝑧2𝑏𝑧subscript𝑏0subscript𝑏1𝑧subscript𝑏2superscript𝑧2a(z)=a_{0}+a_{1}z+a_{2}z^{2},\;\;\;\;b(z)=b_{0}+b_{1}z+b_{2}z^{2}</td>
<td></td>
<td>(5)</td>
</tr>
</tbody>
</table>
be polynomials in $`{{\mathbb{F}}_{p^{3^{2\hspace{0pt}m}}}\hspace{0pt}{\lbrack z\rbrack}^{\leq 2}}.`$ Our goal is computing the coefficients of the polynomial
<table>
<tbody>
<tr>
<td></td>
<td>c(z)=a(z)b(z)=c0+c1z+⋯c4z4.𝑐𝑧𝑎𝑧𝑏𝑧subscript𝑐0subscript𝑐1𝑧⋯subscript𝑐4superscript𝑧4c(z)=a(z)b(z)=c_{0}+c_{1}z+\cdots c_{4}z^{4}.</td>
<td></td>
</tr>
</tbody>
</table>
Evaluation of $`a\hspace{0pt}{(z)}`$ and $`b\hspace{0pt}{(z)}`$ at $`{(1,s,s^{2},s^{3})} = {(1,s,{- 1},{- s})}`$ can be done by multiplying the Vandermonde matrix of powers of $`s`$
<table>
<tbody>
<tr>
<td></td>
<td>Vs=(11111s−1−s1−11−11−s−1s)subscript𝑉𝑠11111𝑠1𝑠11111𝑠1𝑠V_{s}=\left(\begin{array}[]{cccc}1&1&1&1\\ 1&s&-1&-s\\ 1&-1&1&-1\\ 1&-s&-1&s\end{array}\right)</td>
<td></td>
<td>(6)</td>
</tr>
</tbody>
</table>
by the vectors $`{(a_{0},a_{1},a_{2},0)}^{T}`$ and $`{(b_{0},b_{1},b_{2},0)}^{T}`$, respectively. This yields the vectors
<table>
<tbody>
<tr>
<td></td>
<td>ϕe(a)=(a0+a1+a2a0+sa1−a2a0−a1+a2a0−sa1−a2)andϕe(b)=(b0+b1+b2b0+sb1−b2b0−b1+b2b0−sb1−b2).formulae-sequencesubscriptitalic-ϕ𝑒𝑎subscript𝑎0subscript𝑎1subscript𝑎2subscript𝑎0𝑠subscript𝑎1subscript𝑎2subscript𝑎0subscript𝑎1subscript𝑎2subscript𝑎0𝑠subscript𝑎1subscript𝑎2andsubscriptitalic-ϕ𝑒𝑏subscript𝑏0subscript𝑏1subscript𝑏2subscript𝑏0𝑠subscript𝑏1subscript𝑏2subscript𝑏0subscript𝑏1subscript𝑏2subscript𝑏0𝑠subscript𝑏1subscript𝑏2\phi_{e}(a)=\left(\begin{array}[]{c}a_{0}+a_{1}+a_{2}\\ a_{0}+sa_{1}-a_{2}\\ a_{0}-a_{1}+a_{2}\\ a_{0}-sa_{1}-a_{2}\end{array}\right)\;\;\;\;\mbox{and}\;\;\;\;\phi_{e}(b)=\left(\begin{array}[]{c}b_{0}+b_{1}+b_{2}\\ b_{0}+sb_{1}-b_{2}\\ b_{0}-b_{1}+b_{2}\\ b_{0}-sb_{1}-b_{2}\end{array}\right).</td>
<td></td>
</tr>
</tbody>
</table>
Let $`{\phi_{e}\hspace{0pt}{(c)}} = {{{\phi_{e}\hspace{0pt}{(a)}} \ast \phi_{e}}\hspace{0pt}{(b)}}`$ be the componentwise product of $`\phi_{e}\hspace{0pt}{(a)}`$ and $`\phi_{e}\hspace{0pt}{(b)}`$
<table>
<tbody>
<tr>
<td></td>
<td>ϕe(c)=(P0P1P2P3)=((a0+a1+a2)(b0+b1+b2)(a0+sa1−a2)(b0+sb1−b2)(a0−a1+a2)(b0−b1+b2)(a0−sa1−a2)(b0−sb1−b2)).subscriptitalic-ϕ𝑒𝑐subscript𝑃0subscript𝑃1subscript𝑃2subscript𝑃3subscript𝑎0subscript𝑎1subscript𝑎2subscript𝑏0subscript𝑏1subscript𝑏2subscript𝑎0𝑠subscript𝑎1subscript𝑎2subscript𝑏0𝑠subscript𝑏1subscript𝑏2subscript𝑎0subscript𝑎1subscript𝑎2subscript𝑏0subscript𝑏1subscript𝑏2subscript𝑎0𝑠subscript𝑎1subscript𝑎2subscript𝑏0𝑠subscript𝑏1subscript𝑏2\phi_{e}(c)=\left(\begin{array}[]{c}P_{0}\\ P_{1}\\ P_{2}\\ P_{3}\end{array}\right)=\left(\begin{array}[]{c}(a_{0}+a_{1}+a_{2})(b_{0}+b_{1}+b_{2})\\ (a_{0}+sa_{1}-a_{2})(b_{0}+sb_{1}-b_{2})\\ (a_{0}-a_{1}+a_{2})(b_{0}-b_{1}+b_{2})\\ (a_{0}-sa_{1}-a_{2})(b_{0}-sb_{1}-b_{2})\end{array}\right).</td>
<td></td>
</tr>
</tbody>
</table>
Using ([2](#S3.E2 "In 3 Multiplication using roots of unity ‣ Explicit formulas for efficient multiplication in 𝔽_3^{6𝑚}")) and ([3](#S3.E3 "In 3 Multiplication using roots of unity ‣ Explicit formulas for efficient multiplication in 𝔽_3^{6𝑚}")) we get
<table>
<tbody>
<tr>
<td></td>
<td>(c0c1c2c3)=Ws(P0−P4P1−P4P2−P4P3−P4),subscript𝑐0subscript𝑐1subscript𝑐2subscript𝑐3subscript𝑊𝑠subscript𝑃0subscript𝑃4subscript𝑃1subscript𝑃4subscript𝑃2subscript𝑃4subscript𝑃3subscript𝑃4\left(\begin{array}[]{c}c_{0}\\ c_{1}\\ c_{2}\\ c_{3}\end{array}\right)=W_{s}\left(\begin{array}[]{c}P_{0}-P_{4}\\ P_{1}-P_{4}\\ P_{2}-P_{4}\\ P_{3}-P_{4}\end{array}\right),</td>
<td></td>
</tr>
</tbody>
</table>
where $`P_{4} = {a_{2}\hspace{0pt}b_{2}}`$ and
<table>
<tbody>
<tr>
<td></td>
<td>Ws=Vs−1=(11111−s−1s1−11−11s−1−s)subscript𝑊𝑠superscriptsubscript𝑉𝑠111111𝑠1𝑠11111𝑠1𝑠W_{s}=V_{s}^{-1}=\left(\begin{array}[]{cccc}1&1&1&1\\ 1&-s&-1&s\\ 1&-1&1&-1\\ 1&s&-1&-s\end{array}\right)</td>
<td></td>
<td>(7)</td>
</tr>
</tbody>
</table>
Thus the explicit formulas for the coefficients of the product are
<table>
<tbody>
<tr>
<td></td>
<td>c0=P0+P1+P2+P3−P4c1=P0−sP1−P2+sP3c2=P0−P1+P2−P3c3=P0+sP1−P2−sP3c4=P4.subscript𝑐0subscript𝑃0subscript𝑃1subscript𝑃2subscript𝑃3subscript𝑃4subscript𝑐1subscript𝑃0𝑠subscript𝑃1subscript𝑃2𝑠subscript𝑃3subscript𝑐2subscript𝑃0subscript𝑃1subscript𝑃2subscript𝑃3subscript𝑐3subscript𝑃0𝑠subscript𝑃1subscript𝑃2𝑠subscript𝑃3subscript𝑐4subscript𝑃4\begin{array}[]{l}c_{0}=P_{0}+P_{1}+P_{2}+P_{3}-P_{4}\\ c_{1}=P_{0}-sP_{1}-P_{2}+sP_{3}\\ c_{2}=P_{0}-P_{1}+P_{2}-P_{3}\\ c_{3}=P_{0}+sP_{1}-P_{2}-sP_{3}\\ c_{4}=P_{4}.\end{array}</td>
<td></td>
<td>(8)</td>
</tr>
</tbody>
</table>
## 4 Efficient implementation
We owe the efficiency of our method to the Cooley-Tukey factorization of the DFT matrix (\[[15](#bib.bib15)\]). The matrices $`V_{s}`$ and $`W_{s}`$ in ([6](#S3.E6 "In 3 Multiplication using roots of unity ‣ Explicit formulas for efficient multiplication in 𝔽_3^{6𝑚}")) and ([7](#S3.E7 "In 3 Multiplication using roots of unity ‣ Explicit formulas for efficient multiplication in 𝔽_3^{6𝑚}")) are not sparse, but they are the DFT matrices of the fourth roots of unity $`s`$ and $`s^{3}`$, respectively. Hence they can be factored as a product of two sparse matrices as shown in ([21](#S4.E21 "In 4 Efficient implementation ‣ Explicit formulas for efficient multiplication in 𝔽_3^{6𝑚}")) and ([34](#S4.E34 "In 4 Efficient implementation ‣ Explicit formulas for efficient multiplication in 𝔽_3^{6𝑚}")).
<table>
<tbody>
<tr>
<td></td>
<td>Vs=(11111s−1−s1−11−11−s−1s)=subscript𝑉𝑠11111𝑠1𝑠11111𝑠1𝑠absent\displaystyle V_{s}=\left(\begin{array}[]{cccc}1&1&1&1\\ 1&s&-1&-s\\ 1&-1&1&-1\\ 1&-s&-1&s\end{array}\right)=</td>
<td>(1100001s1−100001−s)(1010010110−10010−1),1100001𝑠1100001𝑠1010010110100101\displaystyle\left(\begin{array}[]{cccc}1&1&0&0\\ 0&0&1&s\\ 1&-1&0&0\\ 0&0&1&-s\end{array}\right)\left(\begin{array}[]{cccc}1&0&1&0\\ 0&1&0&1\\ 1&0&-1&0\\ 0&1&0&-1\end{array}\right),</td>
<td></td>
<td>(21)</td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td>Ws=(11111−s−1s1−11−11s−1−s)=subscript𝑊𝑠11111𝑠1𝑠11111𝑠1𝑠absent\displaystyle W_{s}=\left(\begin{array}[]{cccc}1&1&1&1\\ 1&-s&-1&s\\ 1&-1&1&-1\\ 1&s&-1&-s\end{array}\right)=</td>
<td>(1100001−s1−100001s)(1010010110−10010−1).1100001𝑠1100001𝑠1010010110100101\displaystyle\left(\begin{array}[]{cccc}1&1&0&0\\ 0&0&1&-s\\ 1&-1&0&0\\ 0&0&1&s\\ \end{array}\right)\left(\begin{array}[]{cccc}1&0&1&0\\ 0&1&0&1\\ 1&0&-1&0\\ 0&1&0&-1\end{array}\right).</td>
<td></td>
<td>(34)</td>
</tr>
</tbody>
</table>
The factorizations in ([21](#S4.E21 "In 4 Efficient implementation ‣ Explicit formulas for efficient multiplication in 𝔽_3^{6𝑚}")) and ([34](#S4.E34 "In 4 Efficient implementation ‣ Explicit formulas for efficient multiplication in 𝔽_3^{6𝑚}")) allow us to efficiently compute the product of the matrices $`V_{s}`$ and $`W_{s}`$ with vectors. Notice also that the product of an element $`\omega = {{u\hspace{0pt}s} + v} \in {{\mathbb{F}}_{3^{m}}\hspace{0pt}{\lbrack s\rbrack}^{\leq 1}} \cong {\mathbb{F}}_{3^{2\hspace{0pt}m}}`$ with $`s`$ equals $`{v\hspace{0pt}s} - u`$. Hence multiplying by $`s`$ an element of $`{\mathbb{F}}_{3^{2\hspace{0pt}m}}`$ is not more expensive than a change of sign.
Notice that in alternative to the Vandermonde matrix corresponding to $`s`$ we could use the matrix
<table>
<tbody>
<tr>
<td></td>
<td>(100011111−11−11s−1−s)1000111111111𝑠1𝑠\left(\begin{array}[]{cccc}1&0&0&0\\ 1&1&1&1\\ 1&-1&1&-1\\ 1&s&-1&-s\end{array}\right)</td>
<td></td>
</tr>
</tbody>
</table>
whose inverse is
<table>
<tbody>
<tr>
<td></td>
<td>(1000s1−s−1−ss−1−1−10−s1+s−1+s−s).1000𝑠1𝑠1𝑠𝑠1110𝑠1𝑠1𝑠𝑠\left(\begin{array}[]{cccc}1&0&0&0\\ s&1-s&-1-s&s\\ -1&-1&-1&0\\ -s&1+s&-1+s&-s\end{array}\right).</td>
<td></td>
</tr>
</tbody>
</table>
Obviously the latter matrices are sparse but since they do not possess any special structure up to our knowledge, multiplying them by vectors is more expensive than multiplying $`V_{s}`$ and $`W_{s}`$.
Multiplying elements in the field $`{\mathbb{F}}_{3^{6 \cdot 97}}`$ is required in the Tate pairing computation on the group of $`{\mathbb{F}}_{3^{97}}`$-rational points of the elliptic curves
<table>
<tbody>
<tr>
<td></td>
<td>Ed:y2=x3−x+dd∈{−1,1}:subscript𝐸𝑑superscript𝑦2superscript𝑥3𝑥𝑑𝑑11E_{d}:y^{2}=x^{3}-x+d\;\;\;d\in\{-1,1\}</td>
<td></td>
</tr>
</tbody>
</table>
defined over $`{\mathbb{F}}_{3}`$. An efficient algorithm for the computation of the Tate pairings on these curves is discussed in \[[6](#bib.bib6)\].
We have implemented the multiplication over $`{\mathbb{F}}_{3^{6 \cdot 97}}`$ using the Karatsuba method, the Montgomery method from \[[18](#bib.bib18)\], and our proposed method on a PC with an AMD Athlon 64 processor 3500+. The processor was running at 2.20 GHz and we have used the NTL library (see \[[19](#bib.bib19)\]) for multiplication in $`{\mathbb{F}}_{3^{97}}`$. Please note that although we have chosen $`m = 97`$ for benchmarking purposes, these methods can be applied to any odd $`m > 2`$ as mentioned in Section [3](#S3 "3 Multiplication using roots of unity ‣ Explicit formulas for efficient multiplication in 𝔽_3^{6𝑚}").
<table>
<thead>
<tr>
<th>Multiplication method</th>
<th>Elapsed time (ms)</th>
</tr>
</thead>
<tbody>
<tr>
<th>Karatsuba method</th>
<td>1.6981.6981.698</td>
</tr>
<tr>
<th>Montgomery method</th>
<td>1.6051.6051.605</td>
</tr>
<tr>
<th>Proposed method</th>
<td>1.4511.4511.451</td>
</tr>
</tbody>
</table>
Table 1: Comparison of the execution times of the Karatsuba and Montgomery multipliers with the proposed method for $`{\mathbb{F}}_{3^{6\hspace{0pt}m}}`$.
The execution times are shown in Table [1](#S4.T1 "Table 1 ‣ 4 Efficient implementation ‣ Explicit formulas for efficient multiplication in 𝔽_3^{6𝑚}"). For the Karatsuba and the proposed methods we have used the tower of extensions
<table>
<tbody>
<tr>
<td></td>
<td>𝔽397⊂𝔽32⋅97⊂𝔽36⋅97,subscript𝔽superscript397subscript𝔽superscript3⋅297subscript𝔽superscript3⋅697\mathbb{F}_{3^{97}}\subset\mathbb{F}_{3^{2\cdot 97}}\subset\mathbb{F}_{3^{6\cdot 97}},</td>
<td></td>
</tr>
</tbody>
</table>
where
<table>
<tbody>
<tr>
<td></td>
<td>𝔽397subscript𝔽superscript397\displaystyle\mathbb{F}_{3^{97}}</td>
<td>≅𝔽3[x]/(x97+x16+2)absentsubscript𝔽3delimited-[]𝑥superscript𝑥97superscript𝑥162\displaystyle\cong\mathbb{F}_{{3}}[x]/(x^{97}+x^{16}+2)</td>
<td></td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td>𝔽32⋅97subscript𝔽superscript3⋅297\displaystyle\mathbb{F}_{3^{2\cdot 97}}</td>
<td>≅𝔽397[y]/(y2+1)absentsubscript𝔽superscript397delimited-[]𝑦superscript𝑦21\displaystyle\cong\mathbb{F}_{3^{97}}[y]/(y^{2}+1)</td>
<td></td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td>𝔽36⋅97subscript𝔽superscript3⋅697\displaystyle\mathbb{F}_{3^{6\cdot 97}}</td>
<td>≅𝔽32⋅97[z]/(z3−z−1),absentsubscript𝔽superscript3⋅297delimited-[]𝑧superscript𝑧3𝑧1\displaystyle\cong\mathbb{F}_{3^{2\cdot 97}}[z]/(z^{3}-z-1),</td>
<td></td>
</tr>
</tbody>
</table>
whereas for the Montgomery method the representation
<table>
<tbody>
<tr>
<td></td>
<td>𝔽36⋅97≅𝔽397[y]/(y6+y−1)subscript𝔽superscript3⋅697subscript𝔽superscript397delimited-[]𝑦superscript𝑦6𝑦1\mathbb{F}_{3^{6\cdot 97}}\cong\mathbb{F}_{3^{97}}[y]/(y^{6}+y-1)</td>
<td></td>
</tr>
</tbody>
</table>
has been used. Our implementations show that the new method is almost $`14`$% faster than the Karatsuba and $`10`$% faster than the Montgomery method, which is almost the ratio of saved multiplications. This provides further evidence for the fact that the number of multiplications in $`{\mathbb{F}}_{3^{97}}`$ is a good indicator of the performance of the method for $`{\mathbb{F}}_{3^{6 \cdot 97}}`$.
Our multiplications are based on the following formulas. Let $`{\alpha,\beta} \in {\mathbb{F}}_{3^{6 \cdot m}}`$ be given as:
<table>
<tbody>
<tr>
<td></td>
<td>α=a0+a1s+a2r+a3rs+a4r2+a5r2s,𝛼subscript𝑎0subscript𝑎1𝑠subscript𝑎2𝑟subscript𝑎3𝑟𝑠subscript𝑎4superscript𝑟2subscript𝑎5superscript𝑟2𝑠\displaystyle\alpha=a_{0}+a_{1}s+a_{2}r+a_{3}rs+a_{4}r^{2}+a_{5}r^{2}s,</td>
<td></td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td>β=b0+b1s+b2r+b3rs+b4r2+b5r2s,𝛽subscript𝑏0subscript𝑏1𝑠subscript𝑏2𝑟subscript𝑏3𝑟𝑠subscript𝑏4superscript𝑟2subscript𝑏5superscript𝑟2𝑠\displaystyle\beta=b_{0}+b_{1}s+b_{2}r+b_{3}rs+b_{4}r^{2}+b_{5}r^{2}s,</td>
<td></td>
</tr>
</tbody>
</table>
where $`{a_{0},\cdots,b_{5}} \in {\mathbb{F}}_{3^{m}}`$ and $`s \in F_{3}^{2 \cdot m}`$ , $`r \in {\mathbb{F}}_{3}^{6 \cdot m}`$ are roots of $`y^{2} + 1`$ and $`z^{3} - z - 1`$, respectively. Let their product $`\gamma = {\alpha\hspace{0pt}\beta} \in {\mathbb{F}}_{3^{6 \cdot m}}`$ be
<table>
<tbody>
<tr>
<td></td>
<td>γ=c0+c1s+c2r+c3rs+c4r2+c5r2s.𝛾subscript𝑐0subscript𝑐1𝑠subscript𝑐2𝑟subscript𝑐3𝑟𝑠subscript𝑐4superscript𝑟2subscript𝑐5superscript𝑟2𝑠\gamma=c_{0}+c_{1}s+c_{2}r+c_{3}rs+c_{4}r^{2}+c_{5}r^{2}s.</td>
<td></td>
</tr>
</tbody>
</table>
The coefficients $`c_{i}`$, for $`0 \leq i \leq 5`$ are computed using:
<table>
<tbody>
<tr>
<td></td>
<td>P0=(a0+a2+a4)(b0+b2+b4)P1=(a0+a1+a2+a3+a4+a5)(b0+b1+b2+b3+b4+b5)P2=(a1+a3+a5)(b1+b3+b5)P3=(a0−a3−a4)(b0−b3−b4)P4=(a0+a1+a2−a3−a4−a5)(b0+b1+b2−b3−b4−b5)P5=(a1+a2−a5)(b1+b2−b5)P6=(a0−a2+a4)(b0−b2+b4)P7=(a0+a1−a2−a3+a4+a5)(b0+b1−b2−b3+b4+b5)P8=(a1−a3+a5)(b1−b3+b5)P9=(a0−a3−a4)(b0+b3−b4)P10=(a0+a1−a2+a3−a4−a5)(b0+b1−b2+b3−b4−b5)P11=(a1−a2−a5)(b1−b2−b5)P12=a4b4P13=(a4+a5)(b4+b5)P14=a5b5c0=−P0+P2−P3−P4+P10+P11−P12+P14;c1=P0−P1+P2+P4+P5+P9+P10+P12−P13+P14c2=−P0+P2+P6−P8+P12−P14c3=P0−P1+P2−P6+P7−P8−P12+P13−P14c4=P0−P2−P3+P5+P6−P8−P9+P11+P12−P14c5=−P0+P1−P2+P3−P4+P5−P6+P7−P8+P9−P10+P11−P12+P13−P14subscript𝑃0subscript𝑎0subscript𝑎2subscript𝑎4subscript𝑏0subscript𝑏2subscript𝑏4subscript𝑃1subscript𝑎0subscript𝑎1subscript𝑎2subscript𝑎3subscript𝑎4subscript𝑎5subscript𝑏0subscript𝑏1subscript𝑏2subscript𝑏3subscript𝑏4subscript𝑏5subscript𝑃2subscript𝑎1subscript𝑎3subscript𝑎5subscript𝑏1subscript𝑏3subscript𝑏5subscript𝑃3subscript𝑎0subscript𝑎3subscript𝑎4subscript𝑏0subscript𝑏3subscript𝑏4subscript𝑃4subscript𝑎0subscript𝑎1subscript𝑎2subscript𝑎3subscript𝑎4subscript𝑎5subscript𝑏0subscript𝑏1subscript𝑏2subscript𝑏3subscript𝑏4subscript𝑏5subscript𝑃5subscript𝑎1subscript𝑎2subscript𝑎5subscript𝑏1subscript𝑏2subscript𝑏5subscript𝑃6subscript𝑎0subscript𝑎2subscript𝑎4subscript𝑏0subscript𝑏2subscript𝑏4subscript𝑃7subscript𝑎0subscript𝑎1subscript𝑎2subscript𝑎3subscript𝑎4subscript𝑎5subscript𝑏0subscript𝑏1subscript𝑏2subscript𝑏3subscript𝑏4subscript𝑏5subscript𝑃8subscript𝑎1subscript𝑎3subscript𝑎5subscript𝑏1subscript𝑏3subscript𝑏5subscript𝑃9subscript𝑎0subscript𝑎3subscript𝑎4subscript𝑏0subscript𝑏3subscript𝑏4subscript𝑃10subscript𝑎0subscript𝑎1subscript𝑎2subscript𝑎3subscript𝑎4subscript𝑎5subscript𝑏0subscript𝑏1subscript𝑏2subscript𝑏3subscript𝑏4subscript𝑏5subscript𝑃11subscript𝑎1subscript𝑎2subscript𝑎5subscript𝑏1subscript𝑏2subscript𝑏5subscript𝑃12subscript𝑎4subscript𝑏4subscript𝑃13subscript𝑎4subscript𝑎5subscript𝑏4subscript𝑏5subscript𝑃14subscript𝑎5subscript𝑏5subscript𝑐0subscript𝑃0subscript𝑃2subscript𝑃3subscript𝑃4subscript𝑃10subscript𝑃11subscript𝑃12subscript𝑃14subscript𝑐1subscript𝑃0subscript𝑃1subscript𝑃2subscript𝑃4subscript𝑃5subscript𝑃9subscript𝑃10subscript𝑃12subscript𝑃13subscript𝑃14subscript𝑐2subscript𝑃0subscript𝑃2subscript𝑃6subscript𝑃8subscript𝑃12subscript𝑃14subscript𝑐3subscript𝑃0subscript𝑃1subscript𝑃2subscript𝑃6subscript𝑃7subscript𝑃8subscript𝑃12subscript𝑃13subscript𝑃14subscript𝑐4subscript𝑃0subscript𝑃2subscript𝑃3subscript𝑃5subscript𝑃6subscript𝑃8subscript𝑃9subscript𝑃11subscript𝑃12subscript𝑃14subscript𝑐5subscript𝑃0subscript𝑃1subscript𝑃2subscript𝑃3subscript𝑃4subscript𝑃5subscript𝑃6subscript𝑃7subscript𝑃8subscript𝑃9limit-fromsubscript𝑃10subscript𝑃11subscript𝑃12subscript𝑃13subscript𝑃14\begin{array}[]{l}P_{0}=(a_{0}+a_{2}+a_{4})(b_{0}+b_{2}+b_{4})\\ P_{1}=(a_{0}+a_{1}+a_{2}+a_{3}+a_{4}+a_{5})(b_{0}+b_{1}+b_{2}+b_{3}+b_{4}+b_{5})\\ P_{2}=(a_{1}+a_{3}+a_{5})(b_{1}+b_{3}+b_{5})\\ P_{3}=(a_{0}-a_{3}-a_{4})(b_{0}-b_{3}-b_{4})\\ P_{4}=(a_{0}+a_{1}+a_{2}-a_{3}-a_{4}-a_{5})(b_{0}+b_{1}+b_{2}-b_{3}-b_{4}-b_{5})\\ P_{5}=(a_{1}+a_{2}-a_{5})(b_{1}+b_{2}-b_{5})\\ P_{6}=(a_{0}-a_{2}+a_{4})(b_{0}-b_{2}+b_{4})\\ P_{7}=(a_{0}+a_{1}-a_{2}-a_{3}+a_{4}+a_{5})(b_{0}+b_{1}-b_{2}-b_{3}+b_{4}+b_{5})\\ P_{8}=(a_{1}-a_{3}+a_{5})(b_{1}-b_{3}+b_{5})\\ P_{9}=(a_{0}-a_{3}-a_{4})(b_{0}+b_{3}-b_{4})\\ P_{10}=(a_{0}+a_{1}-a_{2}+a_{3}-a_{4}-a_{5})(b_{0}+b_{1}-b_{2}+b_{3}-b_{4}-b_{5})\\ P_{11}=(a_{1}-a_{2}-a_{5})(b_{1}-b_{2}-b_{5})\\ P_{12}=a_{4}b_{4}\\ P_{13}=(a_{4}+a_{5})(b_{4}+b_{5})\\ P_{14}=a_{5}b_{5}\\ c_{0}=-P_{0}+P_{2}-P_{3}-P_{4}+P_{10}+P_{11}-P_{12}+P_{14};\\ c_{1}=P_{0}-P_{1}+P_{2}+P_{4}+P_{5}+P_{9}+P_{10}+P_{12}-P_{13}+P_{14}\\ c_{2}=-P_{0}+P_{2}+P_{6}-P_{8}+P_{12}-P_{14}\\ c_{3}=P_{0}-P_{1}+P_{2}-P_{6}+P_{7}-P_{8}-P_{12}+P_{13}-P_{14}\\ c_{4}=P_{0}-P_{2}-P_{3}+P_{5}+P_{6}-P_{8}-P_{9}+P_{11}+P_{12}-P_{14}\\ c_{5}=-P_{0}+P_{1}-P_{2}+P_{3}-P_{4}+P_{5}-P_{6}+P_{7}-P_{8}+P_{9}-P_{10}+\\ P_{11}-P_{12}+P_{13}-P_{14}\\ \end{array}</td>
<td></td>
</tr>
</tbody>
</table>
## 5 Other applications of the proposed method
Consider the family of hyperelliptic curves
<table>
<tbody>
<tr>
<td></td>
<td>Cd:y2=xp−x+dd∈{−1,1}:subscript𝐶𝑑superscript𝑦2superscript𝑥𝑝𝑥𝑑𝑑11C_{d}:\;y^{2}=x^{p}-x+d\;\;\;d\in\{-1,1\}</td>
<td></td>
<td>(35)</td>
</tr>
</tbody>
</table>
defined over $`{\mathbb{F}}_{p}`$, for $`p = 3`$ mod. $`4`$. Let $`m`$ be such that $`{({2\hspace{0pt}p},m)} = 1`$ (in practice $`m`$ will often be prime), and consider the $`{\mathbb{F}}_{p^{m}}`$-rational points of the Jacobian of $`C_{d}`$. An efficient implementation of the Tate pairing on these groups is given by Duursma and Lee in \[[6](#bib.bib6)\] and \[[20](#bib.bib20)\], where they extend analogous results of Barreto et. al. and of Galbraith et. al. for the case $`p = 3`$. Notice that this family of curves includes the elliptic curves $`E_{d}`$ that we mentioned in the last section. In the aforementioned papers it is also shown that the curve $`C_{d}`$ has embedding degree $`2\hspace{0pt}p`$. In order to compute the Tate pairing on this curve, one works with the tower of field extensions
<table>
<tbody>
<tr>
<td></td>
<td>𝔽pm⊂𝔽p2m⊂𝔽p2pmsubscript𝔽superscript𝑝𝑚subscript𝔽superscript𝑝2𝑚subscript𝔽superscript𝑝2𝑝𝑚\mathbb{F}_{p^{m}}\subset\mathbb{F}_{p^{2m}}\subset\mathbb{F}_{p^{2pm}}</td>
<td></td>
</tr>
</tbody>
</table>
where the fields are represented as
<table>
<tbody>
<tr>
<td></td>
<td>𝔽p2m≅𝔽pm[y]/(y2+1)and𝔽p2pm≅𝔽p2m[z]/(zp−z+2d).subscript𝔽superscript𝑝2𝑚subscript𝔽superscript𝑝𝑚delimited-[]𝑦superscript𝑦21andsubscript𝔽superscript𝑝2𝑝𝑚subscript𝔽superscript𝑝2𝑚delimited-[]𝑧superscript𝑧𝑝𝑧2𝑑\mathbb{F}_{p^{2m}}\cong\mathbb{F}_{p^{m}}[y]/(y^{2}+1)\;\;\;\mbox{and}\;\;\;\mathbb{F}_{p^{2pm}}\cong\mathbb{F}_{p^{2m}}[z]/(z^{p}-z+2d).</td>
<td></td>
</tr>
</tbody>
</table>
Let $`{{a\hspace{0pt}{(z)}},{b\hspace{0pt}{(z)}}} \in {{\mathbb{F}}_{p^{2\hspace{0pt}p\hspace{0pt}m}}\hspace{0pt}{\lbrack z\rbrack}^{\leq {p - 1}}}`$,
<table>
<tbody>
<tr>
<td></td>
<td>a(z)=a0+a1z+…+ap−1zp−1,𝑎𝑧subscript𝑎0subscript𝑎1𝑧…subscript𝑎𝑝1superscript𝑧𝑝1a(z)=a_{0}+a_{1}z+\ldots+a_{p-1}z^{p-1},</td>
<td></td>
</tr>
</tbody>
</table>
<table>
<tbody>
<tr>
<td></td>
<td>b(z)=b0+b1z+…+bp−1zp−1.𝑏𝑧subscript𝑏0subscript𝑏1𝑧…subscript𝑏𝑝1superscript𝑧𝑝1b(z)=b_{0}+b_{1}z+\ldots+b_{p-1}z^{p-1}.</td>
<td></td>
</tr>
</tbody>
</table>
Then $`{c\hspace{0pt}{(z)}} = {a\hspace{0pt}{(z)}\hspace{0pt}b\hspace{0pt}{(z)}}`$ has $`{2\hspace{0pt}p} - 1`$ coefficients, two of which can be computed as
<table>
<tbody>
<tr>
<td></td>
<td>c0=a0b0andc2(p−1)=a2(p−1)b2(p−1).subscript𝑐0subscript𝑎0subscript𝑏0andsubscript𝑐2𝑝1subscript𝑎2𝑝1subscript𝑏2𝑝1c_{0}=a_{0}b_{0}\;\;\;\mbox{and}\;\;\;c_{2(p-1)}=a_{2(p-1)}b_{2(p-1)}.</td>
<td></td>
</tr>
</tbody>
</table>
In order to determine the remaining $`{2\hspace{0pt}p} - 3`$ coefficients, we can write a Vandermonde matrix with entries in $`{\mathbb{F}}_{p^{2\hspace{0pt}m}}^{\ast}`$ using, e.g., the elements
<table>
<tbody>
<tr>
<td></td>
<td>1,2,…,p−1,±s,…,±p−32s,p−12s.12…𝑝1plus-or-minus𝑠…plus-or-minus𝑝32𝑠𝑝12𝑠1,2,\ldots,p-1,\pm s,\ldots,\pm\frac{p-3}{2}s,\frac{p-1}{2}s.</td>
<td></td>
</tr>
</tbody>
</table>
Another option is writing a Vandermonde matrix using a primitive $`2\hspace{0pt}{({p - 1})}`$-st root of unity combined with the relation:
<table>
<tbody>
<tr>
<td></td>
<td>c2(p−1)=a2(p−1)b2(p−1).subscript𝑐2𝑝1subscript𝑎2𝑝1subscript𝑏2𝑝1c_{2(p-1)}=a_{2(p-1)}b_{2(p-1)}.</td>
<td></td>
</tr>
</tbody>
</table>
Notice that there is an element of order $`2\hspace{0pt}{({p - 1})}`$ in $`{\mathbb{F}}_{p^{2}}`$, since $`\left. {2\hspace{0pt}{({p - 1})}} \middle| {p^{2} - 1} \right.`$. If $`a`$ is a primitive element in $`{\mathbb{F}}_{p^{2}}`$, then $`\omega = a^{{({p + 1})}/2}`$ is a primitive $`2\hspace{0pt}{({p - 1})}`$st root of unity.
## 6 Conclusion
In this paper we derived new formulas for multiplication in $`{\mathbb{F}}_{3^{6\hspace{0pt}m}}`$, which use only $`15`$ multiplications in $`{\mathbb{F}}_{3^{m}}`$. Being able to efficiently multiply elements in $`{\mathbb{F}}_{3^{6\hspace{0pt}m}}`$ is a central task for the computation of the Tate pairing on elliptic and hyperelliptic curves. Our method is based on the fast Fourier transform, slightly modified to be adapted to the finite fields that we work on. Our software experiments show that this method is at least $`10`$% faster than other proposed methods in the literature. We have also discussed use of these ideas in conjunction with the general methods of Duursma-Lee for Tate pairing computations on elliptic and hyperelliptic curves.
## Acknowledgement
The research described in this paper was funded in part by the Swiss National Science Foundation, registered there under grant number 107887, and by the German Research Foundation (Deutsche Forschungsgemeinschaft DFG) under project RU 477/8. We thank also the reviewers for their precise comments.
## References
- \[1\] Knuth, D.E.: The Art of Computer Programming, vol. 2, Seminumerical Algorithms. 3rd edn. Addison-Wesley, Reading MA (1998) First edition 1969.
- \[2\] von zur Gathen, J., Gerhard, J.: Modern Computer Algebra. Second edn. Cambridge University Press, Cambridge, UK (2003) First edition 1999.
- \[3\] U.S. Department of Commerce / National Institute of Standards and Technology: Digital Signature Standard (DSS). (2000) Federal Information Processings Standards Publication 186-2.
- \[4\] Bailey, D.V., Paar, C.: Optimal extension fields for fast arithmetic in public-key algorithms. In Krawczyk, H., ed.: Advances in Cryptology: Proceedings of CRYPTO ’98, Santa Barbara CA. Number 1462 in Lecture Notes in Computer Science, Springer-Verlag (1998) 472–485
- \[5\] Avanzi, R.M., Mihăilescu, P.: Generic efficient arithmetic algorithms for PAFFs (processor adequate finite fields) and related algebraic structures (extended abstract). In: Selected Areas in Cryptography (SAC 2003), Springer-Verlag (2003) 320–334
- \[6\] Duursma, I., Lee, H.: (Tate-pairing implementations for tripartite key agreement)
- \[7\] Kerins, T., Marnane, W.P., Popovici, E.M., Barreto, P.S.L.M.: Efficient hardware for the tate pairing calculation in characteristic three. In: Cryptographic Hardware and Embedded Systems, CHES2005. Number 3659 in Lecture Notes in Computer Science, Springer-Verlag (2005) 412–426
- \[8\] Karatsuba, A., Ofman, Y.: Multiplication of multidigit numbers on automata.
Soviet Physics–Doklady 7 (1963) 595–596 translated from Doklady Akademii Nauk SSSR, Vol. 145, No. 2, pp. 293–294, July, 1962.
- \[9\] Paar, C.: Efficient VLSI Architectures for Bit-Parallel Computation in Galois Fields. PhD thesis, Institute for Experimental Mathematics, University of Essen, Essen, Germany (1994)
- \[10\] Lempel, A., Winograd, S.: A new approach to error-correcting codes.
IEEE Transactions on Information Theory IT-23 (1977) 503–508
- \[11\] Winograd, S.: Arithmetic Complexity of computations. Volume 33. SIAM, Philadelphia (1980)
- \[12\] Bürgisser, P., Clausen, M., Shokrollahi, M.A.: Algebraic Complexity Theory. Number 315 in Grundlehren der mathematischen Wissenschaften. Springer-Verlag (1997)
- \[13\] Bajard, J.C., Imbert, L., Negre, C.: Arithmetic operations in finite fields of medium prime characteristic using the lagrange representation.
IEEE Transactions on Computers 55 (2006) 1167–1177
- \[14\] Blahut, R.E.: Fast Algorithms for Digital Signal Processing. Addison-Wesley, Reading MA (1985)
- \[15\] Cooley, J.W., Tukey, J.W.: An algorithm for the machine computation of the complex fourier series.
Mathematics of Computation 19 (1965) 297–31
- \[16\] Loan, C.V.: Computational Frameworks for the Fast Fourier Transform. Society for Industrial and Applied Mathematics (siam), Philadelphia (1992)
- \[17\] von zur Gathen, J., Shokrollahi, J.: Efficient FPGA-based Karatsuba multipliers for polynomials over $`{\mathbb{F}}_{2}`$. In Preneel, B., Tavares, S., eds.: Selected Areas in Cryptography (SAC 2005). Number 3897 in Lecture Notes in Computer Science, Kingston, ON, Canada, Springer-Verlag (2005) 359–369
- \[18\] Montgomery, P.L.: Five, Six, and seven-Term Karatsuba-Like Formulae.
IEEE Transactions on Computers 54 (2005) 362–369
- \[19\] Shoup, V.:
(NTL: A library for doing number theory, http://www.shoup.net/ntl)
- \[20\] Duursma, I., Lee, H.S.: Tate pairing implementation for hyperelliptic curves $`y^{2} = {{x^{p} - x} + d}`$. In: Advances in cryptology—ASIACRYPT 2003. Volume 2894 of Lecture Notes in Comput. Sci. Springer, Berlin (2003) 111–123 |
0707.1820 | cs.NI | cyber | # Understanding the Properties of the BitTorrent Overlay
Anwar Al-Hamra, Arnaud Legout, and Chadi Barakat
INRIA
Sophia Antipolis, France
Email: {anwar.al_hamra, arnaud.legout, chadi.barakat}@sophia.inria.fr
###### Abstract
In this paper, we conduct extensive simulations to understand the properties of the overlay generated by BitTorrent. We start by analyzing how the overlay properties impact the efficiency of BitTorrent. We focus on the average peer set size (i.e., average number of neighbors), the time for a peer to reach its maximum peer set size, and the diameter of the overlay. In particular, we show that the later a peer arrives in a torrent, the longer it takes to reach its maximum peer set size. Then, we evaluate the impact of the maximum peer set size, the maximum number of outgoing connections per peer, and the number of NATed peers on the overlay properties. We show that BitTorrent generates a robust overlay, but that this overlay is not a random graph. In particular, the connectivity of a peer to its neighbors depends on its arriving order in the torrent. We also show that a large number of NATed peers significantly compromise the robustness of the overlay to attacks. Finally, we evaluate the impact of peer exchange on the overlay properties, and we show that it generates a chain-like overlay with a large diameter, which will adversely impact the efficiency of large torrents.
## I Introduction
Recently, Peer-to-Peer (P2P) networks have emerged as an attractive architecture for content sharing over the Internet. By leveraging the available resources at the peers, P2P networks have the potential to scale to a large number of peers. Nowadays, P2P networks support a variety of applications, for instance, file sharing (e.g., BitTorrent, Emule), audio conferencing (e.g., Skype), or video conferencing (e.g., End System Multicast \[[1](#bib.bib1)\]). Among all existing P2P applications, file sharing is still the most popular one. A study in 2004 by the Digital Music Weblog magazine \[[2](#bib.bib2)\] states that P2P file sharing is responsible for $`70 - {80\%}`$ of the overall European Internet traffic. And among the many P2P file sharing protocols, BitTorrent \[[3](#bib.bib3)\] is the most popular one. Alone, BitTorrent generates more than half of the P2P traffic \[[4](#bib.bib4)\].
Invented by Bram Cohen, BitTorrent \[[5](#bib.bib5)\] targets distributing efficiently large files, split into multiple pieces, in case of a massive and sudden demand. The popularity of BitTorrent comes from its efficiency ensured by its peer and piece selection strategies. The peer selection strategy aims at enforcing the cooperation between peers while the piece selection strategy tends to maximize the variety of pieces available among those peers. The great success of BitTorrent has attracted the curiosity of the research community and several papers have appeared on this subject. Thanks to this research effort, we now have a better idea on the strengths and weaknesses of the protocol \[[6](#bib.bib6), [7](#bib.bib7), [8](#bib.bib8), [9](#bib.bib9)\]. We also have a clear idea on the peers’ behavior (i.e., arrival and departure processes), and on the quality of service they experience \[[10](#bib.bib10), [11](#bib.bib11), [12](#bib.bib12)\]. But so far, little effort has been spent to understand the properties of the distribution overlay generated by BitTorrent. As already showed by Urvoy et al. \[[13](#bib.bib13)\], the time to distribute a file in BitTorrent is directly influenced by the overlay topology. For example, it is reasonable to believe that BitTorrent performs better on a full mesh overlay than on a chain one. In addition, as compared to a chain, a full mesh overlay makes BitTorrent more robust to peers’ departures and overlay partitions.
We conduct in this paper extensive simulations to isolate the main properties of the overlay generated by BitTorrent. Our contributions are summarized as follows.
- •
We first evaluate the impact of the overlay properties on the BitTorrent efficiency. We show that a large peer set increases the efficiency of BitTorrent, and that a small diameter is a necessary, but not sufficient, condition for this efficiency. We also show that the time for a peer to reach its maximum peer set size depends on the size of the torrent it joins. The larger the torrent when a peer joins it, the longer the time for this peer to reach its maximum peer set size.
- •
We then study the properties of the overlay generated by BitTorrent. We show that BitTorrent generates a graph with with a small diameter. However, this graph is not random and the average peer set size is significantly lower than the maximum possible peer set size. We also show that this overlay is robust to attacks and to churn.
- •
We show that the properties of the overlay are not significantly impacted by the torrent size, and that a peer set size of 80 is a sensible choice. However, a larger peer set size increases the efficiency of the protocol at the expense of a higher overhead on each peer. We also explain why a maximum number of outgoing connections set to half the maximum peer set size is a good choice, and we show that a large fraction of NATed peers decreases significantly the robustness of the overlay to attacks.
- •
Finally, we evaluate the impact of peer exchange on the overlay properties. Whereas peer exchange allows peers to reach fast their maximum peer set size, it builds a chain-like overlay with a large diameter.
The closest work to ours is the one done by Urvoy et al. \[[13](#bib.bib13)\]. The authors focus on two parameters, the maximum peer set size and the maximum number of outgoing connections. As a result, they show that these two parameters influence the distribution speed of the content and the properties of the overlay.
In this paper, we go further and we provide an analysis that highlights the relation between the overlay properties and the performance of BitTorrent. We also present an in-depth study that characterizes the properties of the BitTorrent overlay. Finally, we show how the overlay properties change as we vary the different system parameters. These parameters include, in addition to the maximum peer set size and maximum number of outgoing connections, the torrent size (i.e., number of peers), the percentage of NATed peers, and the peer exchange extension protocol.
The rest of this paper is organized as follows. In Section [II](#S2 "II Overview of BitTorrent ‣ Understanding the Properties of the BitTorrent Overlay") we give a brief overview of BitTorrent. In Section [III](#S3 "III Simulation Methodology ‣ Understanding the Properties of the BitTorrent Overlay") we describe our methodology and we give results in Sections [IV](#S4 "IV Impact of the Overlay on BitTorrent’s efficiency ‣ Understanding the Properties of the BitTorrent Overlay") and [V](#S5 "V Characterizing the Properties of BitTorrent’s Overlay ‣ Understanding the Properties of the BitTorrent Overlay"). In Section [VI](#S6 "VI Impact of Peer Exchange on the Overlay ‣ Understanding the Properties of the BitTorrent Overlay") we discuss the impact of peer exchange on the overlay and we conclude the work with Section [VII](#S7 "VII Discussion ‣ Understanding the Properties of the BitTorrent Overlay").
## II Overview of BitTorrent
BitTorrent is a P2P file distribution protocol with a focus on scalable and efficient content replication. In particular, BitTorrent capitalizes on the upload capacity of each peer in order to increase the global system capacity as the number of peers increases. This section introduces the terminology used in this paper and gives a short overview of BitTorrent.
###
II-A
Terminology
The terminology used in the BitTorrent community is not standardized. For the sake of clarity, we define here the terms used throughout this paper.
Torrent: A torrent is a set of peers cooperating to share the same content using the BitTorrent protocol.
Tracker: The tracker is a central component that stores the IP addresses of all peers in the torrent. The tracker is used as a rendez-vous point in order to allow new peers to discover existing ones. The tracker also maintains statistics on the torrent. Each peer periodically (typically every 30 minutes) report, for instance, the amount of bytes it has uploaded and downloaded since it joined the torrent.
Leecher and Seed. A peer can be in one of two states: the leecher state, when it is still downloading pieces of the content, and the seed state, when it has all the pieces and is sharing them with others.
Peer Set: Each peer maintains a list of other peers to which it has open TCP connections. We call this list the peer set. This is also known as the neighbor set.
Neighbor: A neighbor of peer $`P`$ is a peer in $`P`$’s peer set.
Maximum Peer Set Size: Each peer cannot have a peer set larger than the maximum peer set size. This is a configuration parameter of the protocol.
Average Peer Set Size: The average peer set size is the sum of the peer set size of each peer in the torrent divided by the number of peers in that torrent.
Maximum Number of Outgoing Connections: Each peer has a limitation on the number of outgoing connections it can establish. This is a configuration parameter of the protocol.
Pieces and Blocks: A file transferred using BitTorrent is split into pieces, and each piece is split into multiple blocks. Blocks are the transmission unit in the network, and peers can only share complete pieces with others. A typical piece size is equal to 512 kBytes, and the block size is equal to 16 kBytes.
Official BitTorrent Client: The official BitTorrent client \[[5](#bib.bib5)\], also known as Mainline client, was initially developed by Bram Cohen and is now maintained by the company he founded.
###
II-B
BitTorrent Overview
Prior to distribution, the content is divided into multiple pieces, and each piece into multiple blocks. A metainfo file is then created by the content provider. This metainfo file, also called a torrent file, contains all the information necessary to download the content and includes the number of pieces, SHA-1 hashes for all the pieces that are used to verify the integrity of the received data, and the IP address and port number of the tracker.
To join a torrent, a peer $`P`$ retrieves the metainfo file out of band, usually from a well-known website, and contacts the tracker that responds with an initial peer set of randomly selected peers, possibly including both seeds and leechers. This initial peer set is augmented later by peers connecting directly to this new peer. Such peers are aware of the new peer by receiving its IP address from the tracker. If ever the peer set size of a peer falls below a given threshold, it re-contacts the tracker to obtain additional peers.
Once $`P`$ has received its initial peer set from the tracker, it starts contacting peers in this set and requesting different pieces of the content. BitTorrent uses specific peer and piece selection strategies to decide with which peers to reciprocate pieces, and which pieces to ask to those peers. The piece selection strategy is called the local rarest first algorithm, and the peer selection strategy is called the choking algorithm. We describe briefly those strategies in the following.
Local rarest first algorithm: Each peer maintains a list of the number of copies of each piece that peers in its peer set have. It uses this information to define a rarest pieces set, which contains the indices of all the pieces with the least number of copies. This set is updated every time a neighbor in the peer set acquires a new piece, and each time a peer joins or leaves the peer set. The rarest pieces set is consulted for the selection of the next piece to download.
Choking algorithm: A peer uses the choking algorithm to decide which peers to exchange data with. The choking algorithm is different when the peer is a leecher or a seed. We only describe here the choking algorithm for leechers. The algorithm gives preference to those peers who upload data at high rates. Once per rechoke period, typically set to ten seconds, a peer re-calculates the data receiving rates from all peers in its peer set. It then selects the fastest ones, a fixed number of them, and uploads only to those for the duration of the period. We say that a peer unchokes the fastest uploaders via a regular unchoke, and chokes all the rest. In addition, it unchokes a randomly selected peer via an optimistic unchoke. The rational is to discover the capacity of new peers, and to give a chance to peers with no piece to start reciprocating. Peers that do not contribute should not be able to attain high download rates, since such peers will be choked by others. Thus, free-riders, i.e., peers that never upload, are penalized. The algorithm does not prevent all free-riding \[[14](#bib.bib14), [8](#bib.bib8)\], but it performs well in a variety of circumstances \[[7](#bib.bib7)\]. Interested readers can refer to Sections 2.3.1 and 2.3.2 in \[[6](#bib.bib6)\] for a detailed description of the choking algorithm for leechers and seeds.
## III Simulation Methodology
To evaluate the properties of the overlay distribution, we have developed a simulator that captures the evolution of the overlay over time, as peers join and leave. We present here the methodology used and in particular the use of simulations over experiments in Section [](#S3.SS3 "III-C Simulations vs. Experiments ‣ III Simulation Methodology ‣ Understanding the Properties of the BitTorrent Overlay")
III-C
.
###
III-A
Parameters Used in the Simulations
BitTorrent has the following parameters to adjust the overlay topology: (1) the maximum peer set size, (2) the maximum number of outgoing connections, (3) the minimum number of neighbors before re-contacting the tracker, and (4) the number of peers returned by the tracker. The default value of those four parameters can be different depending on the version of BitTorrent. For example, the maximum peer set size was recently changed in the mainline client \[[5](#bib.bib5)\] from $`80`$ to $`200`$. Our study shows how these parameters influence the properties of the overlay and the efficiency of BitTorrent.
Another parameter that can have an impact on the overlay properties is the percentage of NATed peers. We will evaluate how this parameter influences the overlay properties. Note that a NATed peer refers to a peer behind a NAT or a firewall.
###
III-B
Simulation Details
Our Simulator, that we made public \[[15](#bib.bib15)\], was developed in MATLAB. We have simulated the tracker protocol as it is implemented in the BitTorrent mainline client 4.0.2. In the following, we give the details of our simulator. First of all, the tracker keeps two lists of peers, $`L_{n\hspace{0pt}a\hspace{0pt}t\hspace{0pt}e\hspace{0pt}d}`$ for NATed peers and $`L_{{n\hspace{0pt}o\hspace{0pt}t} - {n\hspace{0pt}a\hspace{0pt}t\hspace{0pt}e\hspace{0pt}d}}`$ for non NATed ones. Assume that the percentage of NATed peers in the torrent is $`X\%`$. Thus, when a new peer $`P_{i}`$ joins the torrent, it is considered NATed with a probability of $`X\%`$. Then, $`P_{i}`$ contacts the tracker, which in turn returns the IP addresses of up to $`\sigma`$ (e.g., $`50`$) non NATed existing peers (if there are any). These $`\sigma`$ IP addresses are selected at random from the $`L_{{n\hspace{0pt}o\hspace{0pt}t} - {n\hspace{0pt}a\hspace{0pt}t\hspace{0pt}e\hspace{0pt}d}}`$ list. Then, the tracker adds $`P_{i}`$ to $`L_{n\hspace{0pt}a\hspace{0pt}t\hspace{0pt}e\hspace{0pt}d}`$ if $`P_{i}`$ is NATed or to $`L_{{n\hspace{0pt}o\hspace{0pt}t} - {n\hspace{0pt}a\hspace{0pt}t\hspace{0pt}e\hspace{0pt}d}}`$ otherwise.
When $`P_{i}`$ receives the list of peers from the tracker, it stores them in a list called $`L_{t\hspace{0pt}r\hspace{0pt}a\hspace{0pt}c\hspace{0pt}k\hspace{0pt}e\hspace{0pt}r}^{P_{i}}`$. Then, $`P_{i}`$ starts initiating connections to those peers sequentially. When $`P_{i}`$ initiates a connection to peer $`P_{j}`$, $`P_{i}`$ removes $`P_{j}`$ from $`L_{t\hspace{0pt}r\hspace{0pt}a\hspace{0pt}c\hspace{0pt}k\hspace{0pt}e\hspace{0pt}r}^{P_{i}}`$. When a peer $`P_{j}`$ receives a connection request from peer $`P_{i}`$, $`P_{j}`$ will accept this connection only if its peer set size is less than the maximum peer set size. In this case, $`P_{i}`$ adds $`P_{j}`$ to its list of neighbors $`L_{n\hspace{0pt}e\hspace{0pt}i\hspace{0pt}g\hspace{0pt}h\hspace{0pt}b\hspace{0pt}o\hspace{0pt}r\hspace{0pt}s}^{P_{i}}`$. $`P_{j}`$ also adds $`P_{i}`$ to $`L_{n\hspace{0pt}e\hspace{0pt}i\hspace{0pt}g\hspace{0pt}h\hspace{0pt}b\hspace{0pt}o\hspace{0pt}r\hspace{0pt}s}^{P_{j}}`$. Note that, in practice, peer $`P_{i}`$ would initiate TCP connections to the peers in its $`L_{t\hspace{0pt}r\hspace{0pt}a\hspace{0pt}c\hspace{0pt}k\hspace{0pt}e\hspace{0pt}r}^{P_{i}}`$. In our simulator, establishing a connection between peers $`P_{i}`$ and $`P_{j}`$ results in adding $`P_{i}`$ to the list of neighbors of $`P_{j}`$ $`L_{n\hspace{0pt}e\hspace{0pt}i\hspace{0pt}g\hspace{0pt}h\hspace{0pt}b\hspace{0pt}o\hspace{0pt}r\hspace{0pt}s}^{P_{j}}`$ and also adding $`P_{j}`$ to the list of neighbors of $`P_{i}`$ $`L_{n\hspace{0pt}e\hspace{0pt}i\hspace{0pt}g\hspace{0pt}h\hspace{0pt}b\hspace{0pt}o\hspace{0pt}r\hspace{0pt}s}^{P_{i}}`$. This is reasonable because our goal is to reproduce the topology properties of the overlay and no data exchange is simulated over the links between peers.
After the connection has been accepted or refused by $`P_{j}`$, $`P_{i}`$ initiates a new connections to the next peer in $`L_{t\hspace{0pt}r\hspace{0pt}a\hspace{0pt}c\hspace{0pt}k\hspace{0pt}e\hspace{0pt}r}^{P_{i}}`$. Peer $`P_{i}`$ keeps on contacting the peers it discovered from the tracker until (1) it reaches its maximum number of outgoing connections, or (2) $`L_{t\hspace{0pt}r\hspace{0pt}a\hspace{0pt}c\hspace{0pt}k\hspace{0pt}e\hspace{0pt}r}^{P_{i}}`$ becomes empty.
Assume that $`P_{i}`$ and $`P_{j}`$ are neighbors. When $`P_{j}`$ leaves the torrent, $`P_{i}`$ removes $`P_{j}`$ from its list of neighbors $`L_{n\hspace{0pt}e\hspace{0pt}i\hspace{0pt}g\hspace{0pt}h\hspace{0pt}b\hspace{0pt}o\hspace{0pt}r\hspace{0pt}s}^{P_{i}}`$. The tracker also removes $`P_{j}`$ from $`L_{n\hspace{0pt}a\hspace{0pt}t\hspace{0pt}e\hspace{0pt}d}`$ or $`L_{{n\hspace{0pt}o\hspace{0pt}t} - {n\hspace{0pt}a\hspace{0pt}t\hspace{0pt}e\hspace{0pt}d}}`$ depending on whether $`P_{j}`$ was NATed or not. In addition, $`P_{i}`$ will try to replace the neighbor it lost. For this purpose, $`P_{i}`$ checks whether the number of connections it has initiated to its actual neighbors is less than the maximum number of outgoing connections. If this is the case, $`P_{i}`$ checks whether it still knows about other peers in the torrent, i.e., if $`L_{t\hspace{0pt}r\hspace{0pt}a\hspace{0pt}c\hspace{0pt}k\hspace{0pt}e\hspace{0pt}r}^{P_{i}}`$ is not empty. If this is the case, it contacts them sequentially until either (1) one of them accepts the initiated connection or (2) $`L_{t\hspace{0pt}r\hspace{0pt}a\hspace{0pt}c\hspace{0pt}k\hspace{0pt}e\hspace{0pt}r}^{P_{i}}`$ becomes empty.
Whenever the peer set size of $`P_{i}`$ falls below a given threshold (typically $`20`$), it recontacts the tracker and asks for more peers. We set the minimum interval time between two requests to the tracker to $`300`$ simulated seconds.
Finally, each peer contacts the tracker once every $`30`$ minutes to indicate that it is still present in the network. If no report is received from a peer within $`30 - 45`$ minutes, the tracker considers that the peer has left and deletes it from $`L_{n\hspace{0pt}a\hspace{0pt}t\hspace{0pt}e\hspace{0pt}d}`$ or $`L_{{n\hspace{0pt}o\hspace{0pt}t} - {n\hspace{0pt}a\hspace{0pt}t\hspace{0pt}e\hspace{0pt}d}}`$.
Our simulator mimics the real overlay topology construction in BitTorrent and therefore, we believe that our conclusions will hold true for real torrents.
###
III-C
Simulations vs. Experiments
There are three reasons that motivated us to perform simulations and not experiments. First, in BitTorrent, we cannot use solutions that rely on a crawler to infer the topology properties as already done in the context of Gnutella \[[16](#bib.bib16)\]. The reasons is that BitTorrent does not offer distributed mechanisms for peers discovery or data lookup. Thus, there is no way to make a BitTorrent peer give information about its neighbors.
Second, we cannot take advantage of existing traces collected at various trackers. The reason is that a peer never sends information to the tracker concerning its connectivity with other peers, e.g., its list or number of neighbor.
Third, we can experimentally create our own controlled torrents. However, in order to give significant results, we need torrents of moderate size with more than $`1000`$ peers, which is not easy to obtain. In addition, as we have much less flexibility with real experiments, and as we are only concerned by the overlay construction (except in Section [](#S4.SS1 "IV-A Impact of the Average Peer Set Size ‣ IV Impact of the Overlay on BitTorrent’s efficiency ‣ Understanding the Properties of the BitTorrent Overlay")
IV-A
) which is far easier to simulate than the data exchange protocol, we decided to run simulations instead of experiments. We validate our simulator on a small torrent of $`420`$ peers in Section [](#S5.SS2 "V-B Validation with Experiments ‣ V Characterizing the Properties of BitTorrent’s Overlay ‣ Understanding the Properties of the BitTorrent Overlay")
V-B
.
###
III-D
Arrival Distribution of Peers
[Refer to caption](/html/0707.1820/assets/x1.png)
Figure 1: The evolution of our torrent size over time. The overall number of peers that join this torrent from the beginning to the end is equal to $`1867`$, and the maximum number of simultaneous peers is about $`1250`$. The rate at which peers join the torrent decreases exponentially with time.
We assume that peers’ arrivals follow an exponential distribution, i.e., the rate at which peers join the torrent decreases exponentially with time. More precisely, we split the simulated time into slots. Each slot represents $`10`$ minutes of simulated time. The first slot of time refers to the first $`10`$ simulated minutes. More formally, slot $`i`$ is defined as the simulated time elapsed between the moment $`t = {{({i - 1})} \cdot 10}`$ minutes and the moment $`t = {i \cdot 10}`$ minutes. Then, within each slot of time $`i`$, the number of new peers that join the torrent is computed as
<table>
<tbody>
<tr>
<td></td>
<td>arrivalsatsloti𝑎𝑟𝑟𝑖𝑣𝑎𝑙𝑠𝑎𝑡𝑠𝑙𝑜𝑡𝑖\displaystyle arrivals~{}at~{}slot~{}i</td>
<td>=\displaystyle=</td>
<td>1000⋅exp−0.7⋅(i−1)ifi≤4⋅1000superscript⋅0.7𝑖1𝑖𝑓𝑖4\displaystyle 1000\cdot\exp^{-0.7\cdot(i-1)}~{}if~{}i\leq 4</td>
<td></td>
</tr>
<tr>
<td></td>
<td></td>
<td>=\displaystyle=</td>
<td>0ifi>40𝑖𝑓𝑖4\displaystyle 0~{}if~{}i>4</td>
<td></td>
</tr>
</tbody>
</table>
Each peer stays on-line for a random amount of time uniformly distributed between $`10`$ and $`20`$ simulated minutes. Under this assumption on the arrival process, $`1000`$ peers will arrive during the first $`10`$ minutes of the simulations, $`497`$ peers during the second $`10`$ minutes, $`247`$ peers during the third $`10`$ minutes, $`123`$ peers during the fourth $`10`$ minutes. Note that no peer will arrive after the first $`40`$ minutes of the simulation. As a result, we have more arrivals than departures during the first two time slots. In contrast, starting from the third time slot, the departure rate becomes higher than the arrival rate. The torrent size that results from these arrivals and from the lifetime distribution described above corresponds to a typical torrent size evolution \[[10](#bib.bib10), [11](#bib.bib11)\]. The overall number of peers that join this torrent from the beginning to the end is equal to $`1867`$, and the maximum number of simultaneous peers is about $`1250`$.
Even if this torrent is of moderate size, we will show later that it allows us to gain important insights on the properties of the overlay. Moreover, we will explain how we can extrapolate our results to larger torrents. Note that, the lifetime of our torrent is of $`70`$ simulated minutes and the average lifetime of a peer is $`15`$ minutes. One may wonder whether this is realistic as BitTorrent is mostly used to download large files. Typically, the lifetime of a BitTorrent’s peer is of several hours and the torrent’s lifetime ranges from several hours to several months. However, we are interested only in the construction of the overlay and not in the data exchange. Thus, we only need to see how the overlay adapts dynamically to the arrival and departure of peers, which is ensured by the arrival distribution we consider. As a result, considering torrents and peers with larger lifetimes will not give any new insights. It will only increase the run time of the simulations.
###
III-E
Metrics
We consider $`4`$ different metrics to evaluate the overlay properties in this paper. Those metrics are discussed below.
Average peer set size: The peer set size is critical to the efficiency of BitTorrent. Indeed, the peer set size impacts the piece and peer selection strategies, which are at the core of the BitTorrent efficiency.
The piece selection strategy aims at creating a high diversity of pieces among peers. The rational is to guarantee that each peer can always find a piece it needs at any other peer. This way, the peer selection strategy can choose any peer in order to maximize the efficiency of the system, without being biased by the piece availability on those peers. However, this piece selection strategy is based on a version of rarest first with local knowledge. Whereas with global rarest first each peer replicates pieces that are globally the rarest, with local rarest first each peer replicates pieces that are the rarest in its peer set. Therefore, the peer set size is critical to the efficiency of local rarest first. The larger the peer set, the closer local rarest first will be to global rarest first.
The peer selection strategy aims at encouraging high peer reciprocation by favoring peers who contribute. Recently, Legout et al. \[[7](#bib.bib7)\] showed that peers tend to unchoke more frequently other peers with similar upload speeds, since those are the peers that can reciprocate with high enough rates. Thus, the larger the peer set, the higher the probability that a peer will find peers with similar upload capacity and the more efficient the choking algorithm. We confirm this analysis with simulations in Section [](#S4.SS1 "IV-A Impact of the Average Peer Set Size ‣ IV Impact of the Overlay on BitTorrent’s efficiency ‣ Understanding the Properties of the BitTorrent Overlay")
IV-A
.
Speed to converge to the maximum peer set: As we will show later, a large peer set helps the peer to progress fast in the download of the file. Thus, it is important to investigate how long a peer takes in order to reach its maximum peer set size.
Diameter of the overlay distribution: A short diameter is essential to provide a fast distribution of pieces. In Section [](#S4.SS3 "IV-C Impact of the Diameter of the Distribution Overlay ‣ IV Impact of the Overlay on BitTorrent’s efficiency ‣ Understanding the Properties of the BitTorrent Overlay")
IV-C
, we develop a simple analysis to support this claim.
Robustness of the overlay to attacks and high churn rate: P2P networks represent a dynamic environment where peers can join and leave the torrent at any time. As a result, it is important to know whether the overlay generated by BitTorrent is robust to high churn rate. In addition, P2P overlays may be subject to attacks that target to partition the overlay. In Section [V](#S5 "V Characterizing the Properties of BitTorrent’s Overlay ‣ Understanding the Properties of the BitTorrent Overlay"), we explain how we simulate churn rates and attacks.
## IV Impact of the Overlay on BitTorrent’s efficiency
In this section, we investigate the impact of the overlay structure on the efficiency of BitTorrent. First, we evaluate the impact of the peer set size on BitTorrent efficiency. Second, we analyze the convergence speed of peers toward their maximum peer set size. Finally, we develop a simple model that highlights the relation between the diameter and the distribution speed of pieces. Note that the robustness of the overlay will be studied in Section [V](#S5 "V Characterizing the Properties of BitTorrent’s Overlay ‣ Understanding the Properties of the BitTorrent Overlay") through simulations.
###
IV-A
Impact of the Average Peer Set Size
In this section, we simulate the exchange of pieces in BitTorrent in order to understand the influence of the average peer set size on the efficiency of the protocol. Our simulator runs in rounds where each round corresponds to 10 simulated seconds, which is the typical duration between two calls of the choking algorithm in BitTorrent. Every 10 seconds, we scan all peers one after the other. For each peer, we apply the choking algorithm to identify the set of peers it is actively exchanging data with. Then, we apply the piece selection strategy to discover which pieces to upload to each peer chosen by the choking algorithm. The choking algorithm is implemented as explained in Section 2.3.2 in \[[6](#bib.bib6)\]. We consider that bottlenecks are at the access links of the peers. We do not consider network congestion, propagation delays, and network failures.
We generate three overlays each with $`1000`$ peers and a diameter of $`2`$. They only differ in their peer set size. The first overlay has a peer set size of $`50`$, the second one has a peer set size of $`100`$, and the third one has a peer set of $`150`$. We now explain how to construct an overlay with $`1000`$ peers, with a diameter of $`2`$ and a peer set size of $`50`$. The same methodology is used to construct the two other overlays. We apply this algorithm for each peer sequentially starting with $`P_{1}`$. For each peer $`P_{i}`$, we connect it to other peers randomly selected from the set of peers $`\{ P_{1},\ldots,P_{1000}\}`$. The following two conditions should never be violated. First, no peer is allowed to have more than $`50`$ neighbors. Second, a peer cannot be its own neighbor. Note that there is no guarantee that each node will have exactly $`50`$ neighbors. Yet, our results show that very few peers have less than $`50`$ neighbors. Then, we select one source at random to distribute a file of $`100`$MB, which is split into $`100`$ pieces. We assume that all peers join the torrent at its beginning and stay until the end of the simulation. Each peer has a download capacity of 1 Mbit/s, and the upload capacity is randomly selected, with a uniform distribution, between $`160`$ kbit/s and $`350`$ kbit/s. This homogeneous download capacity of peers is reasonable as the replication speed of pieces is usually limited by the upload capacity of peers. Our goal is to study the evolution of the total number of pieces received by each peer in the torrent.
[Refer to caption](/html/0707.1820/assets/x2.png)
Figure 2: Impact of the peer set size on the BitTorrent efficiency. These results are the average over ten independent runs. The larger the peer set size, the faster the replication of pieces.
Our results confirm previous ones showed by Urvoy et al. \[[13](#bib.bib13)\] and show that BitTorrent replicates pieces faster with a larger peer set. Indeed, Fig. [2](#S4.F2 "Figure 2 ‣ IV-A Impact of the Average Peer Set Size ‣ IV Impact of the Overlay on BitTorrent’s efficiency ‣ Understanding the Properties of the BitTorrent Overlay") shows that, as we increase the peer set size from $`50`$ to $`100`$ (respectively from $`100`$ to $`150`$), the replication speed of pieces improves by $`12\%`$ (respectively by $`5\%`$).
In summary, a larger peer set improves the speed of piece replication. However, this is at the expense of an additional load on each peer that has to maintain a larger number of TCP connections and has to handle an additional signaling overhead per connection. Keep in mind that, in the following, and while evaluating the overlay properties, there will be no data exchange between peers. We now only focus on the evolution of the overlay as peers join and leave the torrent.
###
IV-B
Analysis of the Convergence Speed
A BitTorrent client usually needs time to reach its maximum peer set size. In this section, we show that this is a structural problem in BitTorrent and that the convergence speed depends on the torrent size and on the arrival rate of peers. We consider in our analysis a maximum peer set size of $`80`$, a maximum number of outgoing connections of $`40`$, and $`50`$ peers returned by the tracker. We have chosen fixed values for the sake of clarity, and it is straight forward to extend our analysis to other parameter sets.
When a new peer $`P_{i}`$ joins the torrent, it receives from the tracker the IP addresses of 50 peers chosen at random among all peers in the torrent. Then, $`P_{i}`$ connects to at most 40 out of these 50 peers. To complete its peer set and have $`80`$ neighbors, $`P_{i}`$ keeps on cumulating new connections received from the peers that arrive after it. One can easily derive on average how long a peer needs to wait until it completes its peer set.
We assume that the number of peers in the torrent is $`N_{s}`$ when $`P_{i}`$ arrives. We also assume that $`P_{i}`$ has succeeded to initiate $`40`$ outgoing connections and still misses $`40`$ incoming connections in order to reach its maximum peer set of $`80`$ connections. Therefore, the probability that peer $`P_{i}`$ receives a new connection from a new peer $`P_{j}`$ joining the torrent is $`\frac{40}{N_{s}}`$. Thus, the number of peers $`K`$ that should arrive after peer $`P_{i}`$ in order for this peer to cumulate $`40`$ incoming connections is on average given by:
<table>
<tbody>
<tr>
<td></td>
<td>1=∑n=Ns+1Ns+K1n1superscriptsubscript𝑛subscript𝑁𝑠1subscript𝑁𝑠𝐾1𝑛1=\sum_{n=N_{s}+1}^{N_{s}+K}\frac{1}{n}</td>
<td></td>
<td>(2)</td>
</tr>
</tbody>
</table>
Fig. [3](#S4.F3 "Figure 3 ‣ IV-B Analysis of the Convergence Speed ‣ IV Impact of the Overlay on BitTorrent’s efficiency ‣ Understanding the Properties of the BitTorrent Overlay") shows $`K`$ as a function of $`N_{s}`$ as obtained from Eq. ([2](#S4.E2 "In IV-B Analysis of the Convergence Speed ‣ IV Impact of the Overlay on BitTorrent’s efficiency ‣ Understanding the Properties of the BitTorrent Overlay")). We see that the time for a peer to reach its maximum peer set size increases linearly with the torrent size $`N_{s}`$. For example, peer $`P_{100}`$ should wait the arrival of $`173`$ peers after it in order to receive $`40`$ incoming connections, peer $`P_{1000}`$ should wait for $`1720`$ peers and peer peer $`P_{10000}`$ for $`17184`$ peers.
[Refer to caption](/html/0707.1820/assets/x3.png)
Figure 3: Convergence speed of peers toward their maximum peer set size. The number of peers that should arrive after peer $`P_{i}`$ so that $`P_{i}`$ receives $`40`$ incoming connections depends linearly on the size of the torrent when $`P_{i}`$ has joined the torrent.
This linear dependency can be further shown through the following approximation obtained from Eq. [3](#S4.E3 "In IV-B Analysis of the Convergence Speed ‣ IV Impact of the Overlay on BitTorrent’s efficiency ‣ Understanding the Properties of the BitTorrent Overlay"):
<table>
<tbody>
<tr>
<td></td>
<td>K∼(e−1)⋅Nasimilar-to𝐾⋅𝑒1subscript𝑁𝑎K\sim(e-1)\cdot N_{a}</td>
<td></td>
<td>(3)</td>
</tr>
</tbody>
</table>
The error generated by this approximation is low even for very small torrent sizes. For example, we obtain an error of $`0.6\%`$ for a torrent of $`100`$ peers and an error of $`0.09\%`$ for a torrent of $`1000`$ peers.
In summary, the larger the torrent when a peer joins it, the longer this peer will wait to reach its maximum peer set size.
###
IV-C
Impact of the Diameter of the Distribution Overlay
Yang et al. \[[17](#bib.bib17)\] shows that the service capacity of P2P protocols scales exponentially with the number of peers in the torrent. In this section, we apply their analysis to show the impact of the diameter on the capacity of service of P2P protocols.
We consider a torrent with $`N = 2^{k}`$ peers. We assume that all peers have the same upload and download capacity $`b`$. Moreover, we assume that all peers join the system at time $`t = 0`$ and stay until the file is distributed to all peers. The unit of time is $`T = \frac{C_{s}}{b}`$, where $`C_{s}`$ is the content size. Each peer downloads the content from a single peer at a time. A peer can start uploading when it receives entirely the content. Then, it can upload to a single peer at a time. We finally assume that the file is initially available only at the source $`S`$.
At time $`t = 0`$, $`S`$ starts serving the file to peer $`P_{1}`$. At time $`t = T`$, $`P_{1}`$ receives completely the file and starts serving it to peer $`P_{2}`$. At the same time, $`S`$ schedules a new copy of the file to peer $`P_{3}`$. After $`l.T`$ units of time, $`2^{l}`$ peers have entirely the file. As a result, the number of sources, thus the capacity of service, scales exponentially with time. However, this means that, at any time $`i \cdot T`$, the $`2^{i}`$ sources of the file should find $`2^{i}`$ other peers that have not yet received the content. In other words, each of the $`2^{i}`$ sources must have a direct connection in the overlay to a different peer that does not have yet the content. Whereas it is likely that such a condition is verified most of the time for an overlay with a small diameter, it is not clear what happens when the diameter is large.
[Refer to caption](/html/0707.1820/assets/x4.png)
Figure 4: An example of a chain-like overlay. For clarity, we consider only 3 chains that expand in parallel. At each level in this overlay, we have a cluster that includes three peers fully interconnected with each other. The distribution time of the file increases linearly with the number of clusters in the overlay.
Now, we evaluate how the capacity of service scales on the chain-like overlay shown in Fig. [4](#S4.F4 "Figure 4 ‣ IV-C Impact of the Diameter of the Distribution Overlay ‣ IV Impact of the Overlay on BitTorrent’s efficiency ‣ Understanding the Properties of the BitTorrent Overlay"). The overlay includes multiple levels. At each level, we have a cluster that includes $`2^{m}`$ peers. Each peer is connected to all peers in its cluster. In addition, it maintains one connection to the two clusters that surround its own cluster. The source is connected only to the peers in the first cluster. For this overlay, at time $`t = 0`$, the source serves the file to peer $`P_{1}`$ in the first cluster. At time $`t = T`$, after receiving the entire file, $`P_{1}`$ starts serving it to the peer it knows in the second cluster while $`S`$ schedules a new copy of the file to peer $`P_{2}`$, in the first cluster. At time $`{t = 2}.T`$, $`P_{1}`$ does not know any other peer in the second cluster and therefore, it will serve the file to a new peer in its own cluster. We can easily verify that the intra-cluster capacity of service, i.e., the capacity of service inside a cluster once it has at least once source of the content, increases exponentially with time. However, the inter-cluster capacity of service, i.e., the time for each cluster to have at least one source of the content, increases linearly with time. In other words, once the file is served by the source $`S`$, it needs $`p.T`$ units of time to reach the cluster $`P`$. Thus, for a chain with $`N_{c\hspace{0pt}l\hspace{0pt}u\hspace{0pt}s\hspace{0pt}t\hspace{0pt}e\hspace{0pt}r\hspace{0pt}s}`$ of size $`2^{m}`$ each, the service time of the file is $`{({N_{c\hspace{0pt}l\hspace{0pt}u\hspace{0pt}s\hspace{0pt}t\hspace{0pt}e\hspace{0pt}r\hspace{0pt}s} + m})}.T`$, where $`N_{c\hspace{0pt}l\hspace{0pt}u\hspace{0pt}s\hspace{0pt}t\hspace{0pt}e\hspace{0pt}r\hspace{0pt}s}.T`$ units of time are needed to reach the $`N^{t\hspace{0pt}h}`$ cluster in the overlay and $`m.T`$ units of time are needed to duplicate the file over the $`2^{m}`$ peers inside the last cluster. As a result, a chain-like overlay fails to keep peers busy all the time and, consequently, the distribution time of the file increases linearly with the number of clusters in the system.
In summary, a short diameter is necessary to have a capacity of service that scales exponentially with time. However, this condition is not sufficient. For instance, a star overlay where all peers are connected to the same peer in the center has a diameter of 2. In this topology, the peer at the center of the star is a bottleneck, which leads to a poor capacity of service. Therefore, when analyzing the properties of an overlay, the diameter should be interpreted along with the shape of the overlay before making any conclusion.
## V Characterizing the Properties of BitTorrent’s Overlay
In Section [IV](#S4 "IV Impact of the Overlay on BitTorrent’s efficiency ‣ Understanding the Properties of the BitTorrent Overlay") we have evaluated the impact of the overlay properties on BitTorrent efficiency. In this section, we conduct extensive simulations in order to study the properties of the overlay generated by BitTorrent. We first describe the properties of the overlay for a default set of parameters. Then, we vary some of them in order to identify their impact on the overlay properties. For each simulation, we evaluate the overlay properties according to the four metrics introduced in Section [](#S3.SS5 "III-E Metrics ‣ III Simulation Methodology ‣ Understanding the Properties of the BitTorrent Overlay")
III-E
.
###
V-A
Initial Scenario
For this initial scenario, we consider a set of parameters used by default in several BitTorrent clients (e.g., mainline 4.x.y \[[5](#bib.bib5)\]). We set the maximum peer set size to $`80`$, the maximum number of outgoing connections to $`40`$, the number of peers returned by the tracker to $`50`$, and the minimum number of peers to $`20`$. We also consider a torrent with $`1867`$ peers, as described in Section [](#S3.SS4 "III-D Arrival Distribution of Peers ‣ III Simulation Methodology ‣ Understanding the Properties of the BitTorrent Overlay")
III-D
, and we assume that no peer is NATed. We study the case of NATed peers in Section [](#S5.SS6 "V-F Impact of the Number of NATed Peers ‣ V Characterizing the Properties of BitTorrent’s Overlay ‣ Understanding the Properties of the BitTorrent Overlay")
V-F
.
[Refer to caption](/html/0707.1820/assets/x5.png)
Figure 5: The average peer set size over time for the initial scenario, averaged over ten independent runs. At a given time $`t`$, the average peer set size is the sum of the peer set size at time $`t`$ of each peer in the torrent divided by the number of peers in that torrent. The error bars indicate the minimum and maximum. The average peer set size is lower than the maximum peer set size.
Average peer set size: As we can see in Fig. [5](#S5.F5 "Figure 5 ‣ V-A Initial Scenario ‣ V Characterizing the Properties of BitTorrent’s Overlay ‣ Understanding the Properties of the BitTorrent Overlay"), BitTorrent generates an average peer set size that is low compared to the maximum peer set size targeted. For example, the average peer set size does not exceed 65 while the maximum peer set size is set to 80. To explain this low average peer set size we focus on the convergence speed.
[Refer to caption](/html/0707.1820/assets/x6.png)
Figure 6: The peer set size as a function of the peer id at time $`t = 10`$ minutes, averaged over ten independent runs. The error bars indicate the minimum and maximum. Peers are ordered according to their arriving time, the smaller the index the earlier the arrival time. The later a peer joins the torrent, the smaller its peer set size.
Convergence speed: In Section [](#S4.SS2 "IV-B Analysis of the Convergence Speed ‣ IV Impact of the Overlay on BitTorrent’s efficiency ‣ Understanding the Properties of the BitTorrent Overlay")
IV-B
, we have seen that the time for a peer to reach its maximum peer set size depends on the torrent size at the moment of its arrival and on the arrival rate of new peers. In Fig. [6](#S5.F6 "Figure 6 ‣ V-A Initial Scenario ‣ V Characterizing the Properties of BitTorrent’s Overlay ‣ Understanding the Properties of the BitTorrent Overlay"), we depict the distribution of the peer set size after $`10`$ simulated minutes over the first $`1000`$ peers in the torrent. As we can see, the later a peer joins the torrent, the smaller its peer set size. More precisely, the peers that join the torrent earlier reach their maximum peer set size. In contrast, the peers that arrive later do not cumulate enough incoming connections to saturate their peer set, and therefore, their peer set size is around the maximum number of outgoing connections of $`40`$. As a result, the average peer set size is only $`65`$. In Fig. [6](#S5.F6 "Figure 6 ‣ V-A Initial Scenario ‣ V Characterizing the Properties of BitTorrent’s Overlay ‣ Understanding the Properties of the BitTorrent Overlay"), one would expect peer $`P_{1000}`$ to have a peer set size of $`40`$ and not $`30`$. In fact, among the $`50`$ peers returned by the tracker at random, only $`30`$ had a peer set size lower than 80. This is also the reason behind the oscillations in this figure.
[Refer to caption](/html/0707.1820/assets/x7.png)
Figure 7: The diameter of the overlay over time for the initial scenario, averaged over ten independent runs. The error bars indicate the minimum and maximum. The average diameter is lower than 4 during the entire simulations.
Diameter of the overlay: BitTorrent generates an overlay with a short diameter. As we can observe in Fig. [7](#S5.F7 "Figure 7 ‣ V-A Initial Scenario ‣ V Characterizing the Properties of BitTorrent’s Overlay ‣ Understanding the Properties of the BitTorrent Overlay"), the average diameter of the overlay is between $`2`$ and $`4`$ most of the time
<sup>1</sup>
<sup>1</sup>1In this paper, we compute the diameter of the overlay as the longest shortest path between $`1000`$ peers of the overlay selected at random. This method allows us to obtain very good approximation of the diameter and speeds up the run time of the simulator.
. However, at the end time of the torrent, the overlay may get partitioned. If we look closely at the right part of Fig. [7](#S5.F7 "Figure 7 ‣ V-A Initial Scenario ‣ V Characterizing the Properties of BitTorrent’s Overlay ‣ Understanding the Properties of the BitTorrent Overlay"), we notice that the minimum value of the diameter goes to zero, a value we use to indicate partitions. Actually, after a massive departure of peers, we may obtain many small partitions each of tens of peers. The partitions are due to the minimum number of neighbors $`\delta`$ a peer should reach before recontacting the tracker for new peers. Indeed, to minimize the interaction between peers and the tracker, a peer asks the tracker for more peers only if its number of neighbors falls below $`\delta`$. Therefore, as long as the peer set size is larger than $`\delta`$, to recover from a decrease in its peer set size, a peer has to wait for new incoming connections from newly arriving peers, which does not happen toward the end of the torrent. That is why the torrent does not merge again. To prevent such a behavior, one needs to assign a high value to $`\delta`$. The value of $`\delta`$ is then a trade-off between having a connected overlay at the end of the torrent, and a high load at the tracker.
In our analysis in Section [](#S4.SS3 "IV-C Impact of the Diameter of the Distribution Overlay ‣ IV Impact of the Overlay on BitTorrent’s efficiency ‣ Understanding the Properties of the BitTorrent Overlay")
IV-C
, we show that a short diameter is a necessary, but not sufficient, condition for an efficient distribution of the file. We draw in Fig. [8](#S5.F8 "Figure 8 ‣ V-A Initial Scenario ‣ V Characterizing the Properties of BitTorrent’s Overlay ‣ Understanding the Properties of the BitTorrent Overlay") the shape of the overlay generated by BitTorrent.
[Refer to caption](/html/0707.1820/assets/x8.png)
Figure 8: The connectivity matrix of the BitTorrent’s overlay after 10 minutes, for a single run. We only show the first $`1000`$ peers. A dot at (i,j) means that i and j are neighbors. We have a bottleneck of $`1640`$, which refers to the number of connections between the first $`80`$ peers and the rest of the torrent. BitTorrent does not generate a random graph.
We see that BitTorrent does not generate a random overlay, and that the overlay has a specific geometry. Indeed, Fig. [8](#S5.F8 "Figure 8 ‣ V-A Initial Scenario ‣ V Characterizing the Properties of BitTorrent’s Overlay ‣ Understanding the Properties of the BitTorrent Overlay") shows a clustering among peers that arrive first. For example, peer $`P_{25}`$ is connected only to the first hundred peers. The reason is that when $`P_{25}`$ arrives, it connects to all the $`24`$ peers already existing in the torrent. Then, $`P_{25}`$ waits for new arrivals in order to complete the 56 peers it still needs to saturate its peer set. According to Eq. ([2](#S4.E2 "In IV-B Analysis of the Convergence Speed ‣ IV Impact of the Overlay on BitTorrent’s efficiency ‣ Understanding the Properties of the BitTorrent Overlay")), these missing connections can be fulfilled after the arrival of 75 peers on average, $`P_{26},\ldots,P_{100}`$. Similarly, when peer $`P_{200}`$ arrives, it establishes up to 40 outgoing connections. However, $`P_{200}`$ needs to wait the arrival of a large number of peers in order to complete its 40 incoming connections. This explains why, as compared to $`P_{25}`$, the neighbors of $`P_{200}`$ are selected from a larger set of peers (i.e., between $`P_{60}`$ and $`P_{600}`$). Even though we have this clustering phenomena, the overlay does not include bottlenecks. Actually, the number of connections between the first $`80`$ peers and the rest of the network is equal to $`1640`$. Therefore, the BitTorrent has the potential to allow a fast expansion of the pieces.
[Refer to caption](/html/0707.1820/assets/x9.png)
Figure 9: The number of partitions in the overlay as a function of the percentage of departed peers, averaged over ten independent runs. In addition to the mean, we also draw the min and max of our results. The overlay is robust to attacks and churn.
Robustness to attacks and churn: We investigate the robustness of the overlay to a massive departure of peers, which can be due to an attack or a high churn rate. We now consider how we simulate the attack scenario. First, we consider the overlay topology shown in Fig. [8](#S5.F8 "Figure 8 ‣ V-A Initial Scenario ‣ V Characterizing the Properties of BitTorrent’s Overlay ‣ Understanding the Properties of the BitTorrent Overlay"), which represents a snapshot of the topology at time $`t = 10`$. Then, we force the most connected peers to leave. For example, assume that we want to evaluate the robustness of the overlay to attacks after the departure of $`30\%`$ of the peers. In this case, we identify the $`30\%`$ most connected peers in the overlay and we disconnect them from the overlay. Forcing a peer to leave means that we remove all connections between this peer and the rest of the torrent. Once these peers are disconnected, we check whether the overlay becomes partitioned, i.e., includes more than one partition. By varying the percentage of peers that we force to leave, we are able to explore the robustness limits of the BitTorrent overlay.
To simulate a churn rate, we proceed similarly as for the case of an attack. The only difference is that, the peers that we force to leave are selected randomly instead of the most connected ones.
Fig. [9](#S5.F9 "Figure 9 ‣ V-A Initial Scenario ‣ V Characterizing the Properties of BitTorrent’s Overlay ‣ Understanding the Properties of the BitTorrent Overlay") shows that BitTorrent’s overlay is robust to attacks and churn. Indeed, the overlay stays connected, i.e., there is a single partition, when up to $`80\%`$ of the peers leave due to an attack or to churn rate. When more than $`85\%`$ of the peers leave the torrent, partitions appear. However, there is one major partition that includes most of the peers and a few others with one peer each.
For example, when $`95\%`$ of the peers leave the torrent due to an attack, the result of $`1`$ run produced $`18`$ partitions. More precisely, we had $`1`$ partition that included $`23`$ peers, $`2`$ partitions that included each $`7`$ peers, $`2`$ other partitions with $`2`$ peers each, and $`13`$ partitions with $`1`$ single peer each. Similarly, when $`95\%`$ of the peers leave the torrent due to a high churn rate, the overlay was split into $`4`$ partitions, $`1`$ partition with $`43`$ peers, $`1`$ partition with $`5`$ peers, and $`2`$ partitions with $`1`$ peer each.
In summary, we have seen that the average peer set size is significantly lower than the maximum peer set size, and the peer set size of a peer depends on its arriving time in the torrent. In addition, BitTorrent generates an overlay with a short diameter, but this overlay is not a random graph. However, the overlay is robust to attacks and churn.
###
V-B
Validation with Experiments
To validate our simulation results, we have run real experiments using the mainline client 4.0.2 and its tracker implementation that we described in details in Section [](#S3.SS2 "III-B Simulation Details ‣ III Simulation Methodology ‣ Understanding the Properties of the BitTorrent Overlay")
III-B
. In Fig. [10(a)](#S5.F10.sf1 "In Figure 10 ‣ V-B Validation with Experiments ‣ V Characterizing the Properties of BitTorrent’s Overlay ‣ Understanding the Properties of the BitTorrent Overlay") and Fig. [10(b)](#S5.F10.sf2 "In Figure 10 ‣ V-B Validation with Experiments ‣ V Characterizing the Properties of BitTorrent’s Overlay ‣ Understanding the Properties of the BitTorrent Overlay") we draw the connectivity matrix of the overlay as obtained from experiments and simulations respectively. The connectivity matrix is computed after the arrival of $`420`$ peers to the network. As we can see from these two figures, the real experiments and the simulations show similar properties of the overlay. As a result, our simulator produces accurate results, and as compared to real experiments, it offers much more flexibility and allows us to consider larger torrents. Therefore, in the following, we will give results only for simulations.
[Refer to caption](/html/0707.1820/assets/x10.png)
\(a\)
Experimental results
[Refer to caption](/html/0707.1820/assets/x11.png)
\(b\)
Simulation results
Figure 10: The connectivity matrix of the BitTorrent’s overlay after the arrival of $`420`$ peers, for a single run as obtained from real experiments and simulations. A dot at (i,j) means that i and j are neighbors. Real experiments and simulations show similar properties of the overlay.
In the next sections, we will investigate how these results are influenced by (1) the size of the torrent, (2) the maximum peer set size, (3) the maximum number of outgoing connections, and (4) the percentage of NATed peers. We will also discuss in Section [VI](#S6 "VI Impact of Peer Exchange on the Overlay ‣ Understanding the Properties of the BitTorrent Overlay") how peer exchange impacts those results.
###
V-C
Varying the Number of Peers
The torrent for the initial scenario is of moderate size. Indeed, it consists of $`1867`$ peers and of a maximum of $`1200`$ simultaneous peers. To validate how the results of Section [](#S5.SS1 "V-A Initial Scenario ‣ V Characterizing the Properties of BitTorrent’s Overlay ‣ Understanding the Properties of the BitTorrent Overlay")
V-A
are impacted by larger torrents, we have considered two other torrent sizes: $`5598`$ and $`9329`$ peers. For the largest torrent, the maximum number of simultaneous peers is $`6282`$.
[Refer to caption](/html/0707.1820/assets/x12.png)
Figure 11: The average peer set size over time for different torrent sizes, averaged over ten independent runs. The average peer set size is independent of the torrent size.
Average peer set size: The average peer set size does not depends on the torrent size. Indeed, Fig. [11](#S5.F11 "Figure 11 ‣ V-C Varying the Number of Peers ‣ V Characterizing the Properties of BitTorrent’s Overlay ‣ Understanding the Properties of the BitTorrent Overlay") shows that the average peer set size is roughly the same for the three torrent sizes. For example, after 10 minutes, when the number of simultaneous peers is $`1200`$ for the smallest torrent and $`6282`$ for the largest torrent, the average peer set size is the same for the three torrents. The evolution of the peer set size for the three torrents is similar to the once presented in Fig. [6](#S5.F6 "Figure 6 ‣ V-A Initial Scenario ‣ V Characterizing the Properties of BitTorrent’s Overlay ‣ Understanding the Properties of the BitTorrent Overlay"). The peer set size decreases from $`80`$ for the peers that join the torrent early to around $`40`$ for the peers that arrive toward the end.
Convergence speed: According to section [](#S4.SS2 "IV-B Analysis of the Convergence Speed ‣ IV Impact of the Overlay on BitTorrent’s efficiency ‣ Understanding the Properties of the BitTorrent Overlay")
IV-B
, the convergence speed of the peer set decreases when the torrent size increases as shown in Fig. [3](#S4.F3 "Figure 3 ‣ IV-B Analysis of the Convergence Speed ‣ IV Impact of the Overlay on BitTorrent’s efficiency ‣ Understanding the Properties of the BitTorrent Overlay").
[Refer to caption](/html/0707.1820/assets/x13.png)
Figure 12: The diameter over time for different torrent sizes, averaged over ten independent runs. The diameter increases slowly with the torrent size.
Diameter of the overlay: The diameter of the overlay increases slowly with the torrent size. As we can observe in Fig. [12](#S5.F12 "Figure 12 ‣ V-C Varying the Number of Peers ‣ V Characterizing the Properties of BitTorrent’s Overlay ‣ Understanding the Properties of the BitTorrent Overlay"), after 10 minutes, the diameter of the overlay is $`4`$ for the smallest torrent and $`5.5`$ for the largest torrent. The connectivity matrix presents the same characteristics as in Fig. [8](#S5.F8 "Figure 8 ‣ V-A Initial Scenario ‣ V Characterizing the Properties of BitTorrent’s Overlay ‣ Understanding the Properties of the BitTorrent Overlay") for the three torrents.
[Refer to caption](/html/0707.1820/assets/x14.png)
Figure 13: The robustness of the overlay under churn and attacks, averaged over ten different experiments. The overlay shows similar robustness for the different torrent sizes.
Robustness to attacks and churn: Fig. [13](#S5.F13 "Figure 13 ‣ V-C Varying the Number of Peers ‣ V Characterizing the Properties of BitTorrent’s Overlay ‣ Understanding the Properties of the BitTorrent Overlay") shows that the robustness of the overlay is independent of the torrent size. In particular, the three overlays stay connected for up to $`85\%`$ of peer departures. Then, the overlay is partitioned with a single large partition and several partitions with a few peers. To show the similarity at the robustness level for the different torrent sizes, we now analyze the number of partitions that we obtain with the attack scenario after the departure of $`90\%`$ of the most connected peers. Our results show that the torrent of $`1867`$ peers becomes partitioned into $`4`$ partitions with one partition of $`96`$ peers and three others of one single peer each. Similarly, the torrent of $`5598`$ peers becomes partitioned into $`5`$ partitions with one partition of $`294`$ peers and four others of one single peer each. Finally, the torrent of $`9329`$ peers becomes partitioned into $`9`$ partitions with one partition of $`498`$ peers, another partition of two peers, and seven other partitions of one peer each. We found the same tendency for the churn scenario.
In summary, we have investigated the impact of the torrent size on the properties of the overlay formed by BitTorrent. We have found that the results obtained for the initial scenario still hold for larger torrents. Therefore, and in order to reduce significantly the run time of these simulations, in the following, we will focus on a torrent with $`1867`$ peers.
###
V-D
Impact of the Maximum Peer Set Size
The maximum peer set size is usually set to $`80`$. However, some clients choose higher values of this parameter, e.g., mainline 5.x \[[5](#bib.bib5)\] has a maximum peer set size set to $`200`$. In this section, we evaluate the impact of this parameter on the properties of the overlay.
We run simulations with a maximum peer set size $`\Delta`$ varying from $`20`$ to $`200`$. For each value of $`\Delta`$, we set the maximum number of outgoing connections $`O_{m\hspace{0pt}a\hspace{0pt}x}`$ to $`\frac{\Delta}{2}`$, the number of peers returned by the tracker to $`\frac{\Delta + O_{m\hspace{0pt}a\hspace{0pt}x}}{2}`$, and the minimum number of neighbors $`\delta`$ to $`20`$. Then, we evaluate the overlay after $`10`$ simulated minutes, because it is the time at which the number of simultaneous peers in the torrent reaches its upper bound of $`1200`$ peers.
Note that there is no specific rule to set the value of the number of peers returned by the tracker $`\sigma`$ when we change the maximum peer set size $`\Delta`$ and maximum number of outgoing connections $`O_{m\hspace{0pt}a\hspace{0pt}x}`$. Intuitively, $`\sigma`$ should be larger than $`O_{m\hspace{0pt}a\hspace{0pt}x}`$. The reason is that each peer $`P_{i}`$ seeks to initiate $`O_{m\hspace{0pt}a\hspace{0pt}x}`$ connections, which is only possible if the tracker provides $`P_{i}`$ with the addresses of at least $`O_{m\hspace{0pt}a\hspace{0pt}x}`$ other peers in the torrent. Yet, the value of $`\sigma`$ should not be much larger than $`O_{m\hspace{0pt}a\hspace{0pt}x}`$ in order not to increase the load on the tracker. We tried several values of $`\sigma`$ between $`O_{m\hspace{0pt}a\hspace{0pt}x}`$ and $`\Delta`$, but we obtained similar results. Therefore, we decided to set $`\sigma`$ to $`\frac{\Delta + O_{m\hspace{0pt}a\hspace{0pt}x}}{2}`$.
[Refer to caption](/html/0707.1820/assets/x15.png)
Figure 14: The average peer set size as a function of the maximum peer set size, averaged over ten independent runs. The error bars indicate the minimum and maximum. The average peer set size is a linear function of the maximum peer set size.
Average peer set size: The average peer set size increases linearly with the maximum peer set size $`\Delta`$. Indeed, Fig. [14](#S5.F14 "Figure 14 ‣ V-D Impact of the Maximum Peer Set Size ‣ V Characterizing the Properties of BitTorrent’s Overlay ‣ Understanding the Properties of the BitTorrent Overlay") shows that the average peer set size is roughly equal to $`\frac{2 \cdot \Delta}{3}`$. For instance, for a maximum peer set size of $`100`$, the average peer set size is $`65`$.
We found this linear trend in all our simulations, but the slope depends on the instant at which we perform the measurements.
Convergence speed: We extend the analysis in Section [](#S4.SS2 "IV-B Analysis of the Convergence Speed ‣ IV Impact of the Overlay on BitTorrent’s efficiency ‣ Understanding the Properties of the BitTorrent Overlay")
IV-B
by considering a variable maximum peer set size $`\Delta`$ and a maximum number of outgoing connections $`O_{m\hspace{0pt}a\hspace{0pt}x}`$. We rewrite Eq. ([2](#S4.E2 "In IV-B Analysis of the Convergence Speed ‣ IV Impact of the Overlay on BitTorrent’s efficiency ‣ Understanding the Properties of the BitTorrent Overlay")) as follows:
<table>
<tbody>
<tr>
<td></td>
<td>Δ−OmaxΔsubscript𝑂𝑚𝑎𝑥\displaystyle\Delta-O_{max}</td>
<td>=\displaystyle=</td>
<td>∑n=Ns+1Ns+KOmaxnsuperscriptsubscript𝑛subscript𝑁𝑠1subscript𝑁𝑠𝐾subscript𝑂𝑚𝑎𝑥𝑛\displaystyle\sum_{n=N_{s}+1}^{N_{s}+K}\frac{O_{max}}{n}</td>
<td></td>
<td></td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td>ΔOmaxΔsubscript𝑂𝑚𝑎𝑥\displaystyle\frac{\Delta}{O_{max}}</td>
<td>=\displaystyle=</td>
<td>1+∑n=Ns+1Ns+K1n1superscriptsubscript𝑛subscript𝑁𝑠1subscript𝑁𝑠𝐾1𝑛\displaystyle 1+\sum_{n=N_{s}+1}^{N_{s}+K}\frac{1}{n}</td>
<td></td>
<td>(4)</td>
</tr>
</tbody>
</table>
where $`\Delta - O_{m\hspace{0pt}a\hspace{0pt}x}`$ is the number of missing connections, i.e., the number of incoming connections the peer is still waiting for, assuming that the peer has succeeded to initiate $`O_{m\hspace{0pt}a\hspace{0pt}x}`$ outgoing connections. $`\frac{O_{m\hspace{0pt}a\hspace{0pt}x}}{n}`$ is the probability that a peer receives a new incoming connection from a new peer given the number of peers $`n`$ in the torrent. Recall that $`N_{s}`$ represents the torrent size at the moment of the arrival of peer $`P_{i}`$, and $`K`$ is the average number of peers that should arrive after peer $`P_{i}`$ in order for this peer to complete its $`\Delta - O_{m\hspace{0pt}a\hspace{0pt}x}`$ missing connections. We assume that no peer leaves the system. When $`O_{m\hspace{0pt}a\hspace{0pt}x} = \frac{\Delta}{2}`$ Eq. ([](#S5.Ex2 "V-D Impact of the Maximum Peer Set Size ‣ V Characterizing the Properties of BitTorrent’s Overlay ‣ Understanding the Properties of the BitTorrent Overlay")
V-D
) is equivalent to $`1 = {\sum_{n = {N_{s} + 1}}^{N_{s} + K}\frac{1}{n}}`$, which is Eq. ([2](#S4.E2 "In IV-B Analysis of the Convergence Speed ‣ IV Impact of the Overlay on BitTorrent’s efficiency ‣ Understanding the Properties of the BitTorrent Overlay")). Therefore, when $`O_{m\hspace{0pt}a\hspace{0pt}x} = \frac{\Delta}{2}`$, the convergence speed is independent of the value of $`\Delta`$.
[Refer to caption](/html/0707.1820/assets/x16.png)
Figure 15: The diameter of the overlay as a function of the maximum peer set size, averaged over ten independent runs. The error bars indicate the minimum and maximum. The diameter decreases slowly after a peer set size of $`100`$.
Diameter of the overlay: The diameter of the overlay decreases slowly with the maximum peer set size $`\Delta`$ as shown in Fig. [15](#S5.F15 "Figure 15 ‣ V-D Impact of the Maximum Peer Set Size ‣ V Characterizing the Properties of BitTorrent’s Overlay ‣ Understanding the Properties of the BitTorrent Overlay"). The diameter is $`6.5`$ when $`\Delta`$ is $`20`$, $`5.5`$ when $`\Delta`$ is $`40`$, $`4.5`$ when $`\Delta`$ is $`80`$, and $`3`$ when $`\Delta`$ is $`200`$. However, in contrast to the average peer set size, there is no clear trend that can be used to predict the diameter as a function of the maximum peer set size.
[Refer to caption](/html/0707.1820/assets/x17.png)
Figure 16: The robustness of the overlay under churn and attacks, averaged over ten different experiments. The robustness of the overlay increases with the maximum peer set size.
Robustness to attacks and churn: The robustness of the overlay increases with the maximum peer set size $`\Delta`$ as shown in Fig [16](#S5.F16 "Figure 16 ‣ V-D Impact of the Maximum Peer Set Size ‣ V Characterizing the Properties of BitTorrent’s Overlay ‣ Understanding the Properties of the BitTorrent Overlay"). For example, when we set $`\Delta`$ to $`40`$, $`80`$, and $`200`$, the overlay is not partitioned for up to respectively $`65\%`$, $`80\%`$, and $`95\%`$ of peers that leave. There is no discernible distinction between leaves due to churn or attacks.
However, if we carefully look at Fig. [16](#S5.F16 "Figure 16 ‣ V-D Impact of the Maximum Peer Set Size ‣ V Characterizing the Properties of BitTorrent’s Overlay ‣ Understanding the Properties of the BitTorrent Overlay"), we can see that the attack scenario produces partitions when the percentage of departed peers is at $`45\%`$. To understand this behavior, we plot in Fig. [17](#S5.F17 "Figure 17 ‣ V-D Impact of the Maximum Peer Set Size ‣ V Characterizing the Properties of BitTorrent’s Overlay ‣ Understanding the Properties of the BitTorrent Overlay") the connectivity matrix of the overlay after attacking $`45\%`$ of the peers. Fig. [17](#S5.F17 "Figure 17 ‣ V-D Impact of the Maximum Peer Set Size ‣ V Characterizing the Properties of BitTorrent’s Overlay ‣ Understanding the Properties of the BitTorrent Overlay") shows that there are around $`12`$ peers that are disconnected from the rest of the torrent.
From our previous results, we know that the peers that arrive at the beginning of the torrent are the most connected ones and highly connected among each other (see Fig. [8](#S5.F8 "Figure 8 ‣ V-A Initial Scenario ‣ V Characterizing the Properties of BitTorrent’s Overlay ‣ Understanding the Properties of the BitTorrent Overlay")). In addition, these first peers are the most concerned ones by the attack. Recall that the attack scenario forces the most connected peers to leave the torrent. Thus, with a $`45\%`$ of departed peers, only very few of those first peers will remain present in the torrent after the attack. These few peers will have a very few neighbors and will become disconnected from other peers. However, if we consider a percentage of departed peers of $`50\%`$ instead of $`45\%`$, more peers will leave the torrent and those $`12`$ peers will disappear. This means that the torrent will be connected again. In contrast, if we consider a percentage of departed peers of $`40\%`$ instead of $`45\%`$, there will be more peers present in the torrent after the attack, which helps the torrent to remain connected.
In summary, the maximum peer set size does not have a major impact on the properties of the overlay, as long as the maximum peer set size is large enough to have a small diameter. In our simulations, we do not see a major difference in the overlay properties between a maximum peer set size of $`80`$ and $`200`$. However, as the maximum peer set size increases linearly the average peer set size, it also increases the speed of replication of the pieces (according to Section [](#S4.SS1 "IV-A Impact of the Average Peer Set Size ‣ IV Impact of the Overlay on BitTorrent’s efficiency ‣ Understanding the Properties of the BitTorrent Overlay")
IV-A
). Therefore, the main reason to increase the maximum peer set size is to improve the speed of replication. But, there is a tradeoff, as a larger maximum peer set size increases the load on each client due to the larger number of TCP connections to maintain and due to the signaling overhead per connection.
[Refer to caption](/html/0707.1820/assets/x18.png)
Figure 17: The connectivity matrix of the overlay after forcing $`45\%`$ of the most connected peers to leave the torrent, for a single run. A dot at (i,j) means that i and j are neighbors. The first peers are disconnected from the rest of torrent.
###
V-E
Impact of the Maximum Number of Outgoing Connections
The maximum number of outgoing connections $`O_{m\hspace{0pt}a\hspace{0pt}x}`$ is critical to the properties of the overlay. Indeed, when $`O_{m\hspace{0pt}a\hspace{0pt}x}`$ is close to the maximum peer set size $`\Delta`$, the peer set size will converge fast to $`\Delta`$, but new peers will find few peers with available incoming connections, hence a larger diameter.
In this section, we evaluate the impact of $`O_{m\hspace{0pt}a\hspace{0pt}x}`$ on the overlay properties. For the simulations, we set $`\Delta`$ to $`80`$, the minimum number of neighbors to $`20`$, and we vary $`O_{m\hspace{0pt}a\hspace{0pt}x}`$ from $`5`$ to $`80`$ with a step of $`5`$. For each value of $`\Delta`$ and $`O_{m\hspace{0pt}a\hspace{0pt}x}`$, the number of peers returned by the tracker is equal to $`\frac{\Delta + O_{m\hspace{0pt}a\hspace{0pt}x}}{2}`$.
[Refer to caption](/html/0707.1820/assets/x19.png)
Figure 18: The average peer set size as a function of the maximum number of outgoing connections, averaged over ten independent runs. The error bars indicate the minimum and maximum. The average peer set size increases slowly when the number of maximum number of outgoing connections is larger than $`\frac{\Delta}{2}`$.
Average peer set size: Fig. [18](#S5.F18 "Figure 18 ‣ V-E Impact of the Maximum Number of Outgoing Connections ‣ V Characterizing the Properties of BitTorrent’s Overlay ‣ Understanding the Properties of the BitTorrent Overlay") shows the evolution of the average peer set size as a function of the maximum number of outgoing connections. We see that the average peer set size increases fast with $`O_{m\hspace{0pt}a\hspace{0pt}x}`$ when $`O_{m\hspace{0pt}a\hspace{0pt}x}`$ is smaller than $`\frac{\Delta}{2}`$, and it increases slowly with $`O_{m\hspace{0pt}a\hspace{0pt}x}`$ when $`O_{m\hspace{0pt}a\hspace{0pt}x}`$ is larger than $`\frac{\Delta}{2}`$. We notice that, a small $`O_{m\hspace{0pt}a\hspace{0pt}x}`$ leads to a small average peer set size. For example, when $`O_{m\hspace{0pt}a\hspace{0pt}x}`$ is equal to $`5`$, the average peer set size is around $`10`$.
[Refer to caption](/html/0707.1820/assets/x20.png)
Figure 19: The peer set size vs. the peer id, at time $`t = 10`$ minutes, averaged over ten independent simulations. The peer set size increases slowly when the number of maximum number of outgoing connections is larger than $`\frac{\Delta}{2}`$.
Convergence speed: In Fig. [19](#S5.F19 "Figure 19 ‣ V-E Impact of the Maximum Number of Outgoing Connections ‣ V Characterizing the Properties of BitTorrent’s Overlay ‣ Understanding the Properties of the BitTorrent Overlay") we plot the peer set size of each peer at time $`t = 10`$ minutes as a function of the peer id for three values of the maximum number of outgoing connections. These outdegree distributions of peers reflect the convergence speed of peers towards their maximum peer set size. As we can see from Fig. [19](#S5.F19 "Figure 19 ‣ V-E Impact of the Maximum Number of Outgoing Connections ‣ V Characterizing the Properties of BitTorrent’s Overlay ‣ Understanding the Properties of the BitTorrent Overlay"), the peer set size of the different peers improves with $`O_{m\hspace{0pt}a\hspace{0pt}x}`$ when $`O_{m\hspace{0pt}a\hspace{0pt}x}`$ is smaller than $`\frac{\Delta}{2}`$, and it increases slowly with $`O_{m\hspace{0pt}a\hspace{0pt}x}`$ when $`O_{m\hspace{0pt}a\hspace{0pt}x}`$ is larger than $`\frac{\Delta}{2}`$. For example, after the arrival of the first $`1000`$ peers $`P_{1},\ldots,P_{1000}`$, peer $`P_{500}`$ has a peer set size of $`34`$ when $`O_{m\hspace{0pt}a\hspace{0pt}x}`$ is at $`20`$ and a peer set size of $`70`$ when $`O_{m\hspace{0pt}a\hspace{0pt}x}`$ is at $`40`$. However, when we increase $`O_{m\hspace{0pt}a\hspace{0pt}x}`$ from $`40`$ to $`60`$, the peer set size of $`P_{500}`$ increases from $`70`$ to $`76`$ only. This result means that, when we increase $`O_{m\hspace{0pt}a\hspace{0pt}x}`$ beyond $`\frac{\Delta}{2}`$, the convergence speed of a peer towards its maximum peer set increases slowly. We explain this conclusion as follows. According to Eq. ([](#S5.Ex2 "V-D Impact of the Maximum Peer Set Size ‣ V Characterizing the Properties of BitTorrent’s Overlay ‣ Understanding the Properties of the BitTorrent Overlay")
V-D
), when $`O_{m\hspace{0pt}a\hspace{0pt}x}`$ increases, $`K`$ decreases, where $`K`$ is the number of peers that should arrive after a peer $`P_{i}`$, so that $`P_{i}`$ reaches its maximum peer set size. Indeed, when $`O_{m\hspace{0pt}a\hspace{0pt}x}`$ increases, the probability that a peer receives incoming connections from new peers increases too. However, to derive Eq. ([](#S5.Ex2 "V-D Impact of the Maximum Peer Set Size ‣ V Characterizing the Properties of BitTorrent’s Overlay ‣ Understanding the Properties of the BitTorrent Overlay")
V-D
), we assumed that a peer succeeds to establish all its allowed $`O_{m\hspace{0pt}a\hspace{0pt}x}`$ outgoing connections and that the number of connections it misses is $`\Delta - O_{m\hspace{0pt}a\hspace{0pt}x}`$. This is the most optimistic case, and it is not true when $`O_{m\hspace{0pt}a\hspace{0pt}x}`$ is larger than $`\frac{\Delta}{2}`$. Indeed, in that case, peers that arrive at the beginning of the torrent are able to establish a lot of connections among themselves and reach fast their maximum peer set size. However, those peers leave few rooms for incoming connections, as $`O_{m\hspace{0pt}a\hspace{0pt}x}`$ is close to the maximum peer set size $`\Delta`$. Therefore, peers that join later the torrent will not be able to establish $`O_{m\hspace{0pt}a\hspace{0pt}x}`$ outgoing connections, which results in a larger number than $`\Delta - O_{m\hspace{0pt}a\hspace{0pt}x}`$ of missing connections. As a consequence, the increase in the probability that a peer is selected by new arriving peers is compensated by the increase in the number of missing connections.
[Refer to caption](/html/0707.1820/assets/x21.png)
\(a\)
maximum number of outgoing connections of $`70`$
[Refer to caption](/html/0707.1820/assets/x22.png)
\(b\)
maximum number of outgoing connections of $`80`$
Figure 20: The connectivity matrix of the BitTorrent overlay after 10 minutes for a maximum number of outgoing connections of $`70`$ and $`80`$, for a single run. A dot at (i,j) means that i and j are neighbors. When we increase $`O_{m\hspace{0pt}a\hspace{0pt}x}`$ beyond $`\frac{\Delta}{2}`$, the connectivity matrix becomes more narrow around peer index 80.
[Refer to caption](/html/0707.1820/assets/x23.png)
Figure 21: The diameter of the overlay as a function of the maximum number of outgoing connections, averaged over ten independent runs. The error bars indicate the minimum and maximum. Taking a maximum number of outgoing connections $`O_{m\hspace{0pt}a\hspace{0pt}x}`$ larger than $`\frac{\Delta}{2}`$ increases the diameter of the overlay.
Diameter of the overlay: Taking a maximum number of outgoing connections $`O_{m\hspace{0pt}a\hspace{0pt}x}`$ larger than $`\frac{\Delta}{2}`$ increases the diameter of the overlay. As we can see in Fig. [21](#S5.F21 "Figure 21 ‣ V-E Impact of the Maximum Number of Outgoing Connections ‣ V Characterizing the Properties of BitTorrent’s Overlay ‣ Understanding the Properties of the BitTorrent Overlay"), for $`O_{m\hspace{0pt}a\hspace{0pt}x}`$ equal to $`40`$ (respectively $`70`$), the diameter of the overlay is equal to $`4`$ (respectively $`5.5`$). When $`O_{m\hspace{0pt}a\hspace{0pt}x}`$ is equal to $`\Delta`$, i.e., 80, the overlay is partitioned. If we focus on the connectivity matrix of the overlay, we observe how the overlay gets partitioned into two partitions. Indeed, Fig. [20(a)](#S5.F20.sf1 "In Figure 20 ‣ V-E Impact of the Maximum Number of Outgoing Connections ‣ V Characterizing the Properties of BitTorrent’s Overlay ‣ Understanding the Properties of the BitTorrent Overlay") shows that when $`O_{m\hspace{0pt}a\hspace{0pt}x}`$ is equal to 70, the connectivity matrix becomes narrow around peer index 80. This results in the first $`80`$ peers in the torrent being highly connected among themselves with $`3115`$ connections, and poorly connected with the rest of the torrent with $`170`$ connections. When $`O_{m\hspace{0pt}a\hspace{0pt}x}`$ is equal to 80, the first 80 peers become disconnected from the rest of the torrent. This might be a major issue if the source of the torrent is among those 80 peers, which is the regular case.
[Refer to caption](/html/0707.1820/assets/x24.png)
Figure 22: The robustness of the overlay under churn and attacks, averaged over ten different experiments. The robustness of the overlay increases slightly with the maximum peer set size.
Robustness to attacks and churn: Fig. [22](#S5.F22 "Figure 22 ‣ V-E Impact of the Maximum Number of Outgoing Connections ‣ V Characterizing the Properties of BitTorrent’s Overlay ‣ Understanding the Properties of the BitTorrent Overlay") draws the robustness of the overlay with a maximum number of outgoing connections $`O_{m\hspace{0pt}a\hspace{0pt}x}`$ set to $`20`$, $`40`$, and $`60`$. We observe that large values of $`O_{m\hspace{0pt}a\hspace{0pt}x}`$ make the overlay slightly more robust to attacks and churn. For example, in case of an attack, when setting $`O_{m\hspace{0pt}a\hspace{0pt}x}`$ to $`20`$, the partitions appear after the departure of $`60\%`$ of the peers. In contrast, when setting $`O_{m\hspace{0pt}a\hspace{0pt}x}`$ to $`40`$ or $`60`$, the partitions appear after the departure of $`85\%`$ of the peers.
In Fig. [22](#S5.F22 "Figure 22 ‣ V-E Impact of the Maximum Number of Outgoing Connections ‣ V Characterizing the Properties of BitTorrent’s Overlay ‣ Understanding the Properties of the BitTorrent Overlay"), we can also show that, when we set $`O_{m\hspace{0pt}a\hspace{0pt}x}`$ of $`20`$, the number of partitions decreases at the end of the curve. The reason is that, when we force $`90\%`$ or more of the most connected peers to leave the network, a $`O_{m\hspace{0pt}a\hspace{0pt}x}`$ of $`20`$ will produce a lot of partitions with a very few peers each. Thus, increasing the number of departing peers removes those “one single peer partitions”.
Note that, when $`O_{m\hspace{0pt}a\hspace{0pt}x}`$ is set to $`40`$ (respectively to $`60`$), the attack scenario produces partitions when the percentage of departed peers is at $`45\%`$ (respectively $`50\%`$). The reason is that, among the most connected peers, the remaining peers are very few to stay connected with the rest of the torrent.
In summary, setting the maximum number of outgoing connections $`O_{m\hspace{0pt}a\hspace{0pt}x}`$ to $`\frac{\Delta}{2}`$ is a good tradeoff between the average peer set size and the diameter of the overlay.
###
V-F
Impact of the Number of NATed Peers
In this section, we discuss the impact of the percentage of NATed peers on the overlay properties. When a peer is behind a NAT, it cannot receive incoming connections from other peers in the torrent. However, it can initiate outgoing connections to non NATed ones. For the simulations, we set the maximum peer set size to $`80`$, the maximum number of outgoing connections to $`40`$, the minimum number of neighbors $`\delta`$ to $`20`$, and the number of returned peers by the tracker to $`50`$. Then, we vary the percentage of NATed peers from $`0\%`$ to $`90\%`$ with steps of $`10`$.
[Refer to caption](/html/0707.1820/assets/x25.png)
Figure 23: The average peer set size vs. the percentage of NATed peers, averaged over $`10`$ independent runs. The error bars indicate the minimum and maximum. The average peer set decreases linearly with the percentage of NATed peers.
Average peer set size: BitTorrent mitigates very efficiently the impact of the NATed peers on the overlay. For example, we see in Fig. [23](#S5.F23 "Figure 23 ‣ V-F Impact of the Number of NATed Peers ‣ V Characterizing the Properties of BitTorrent’s Overlay ‣ Understanding the Properties of the BitTorrent Overlay") that as we increase the percentage of NATed peers from $`0`$ to $`30\%`$, the average peer set size is reduced from $`65`$ to $`55`$. Indeed, the average peer set size decreases slowly with the percentage of NATed peers. However, the slope of the curve becomes much sharper when the percentage of NATed peers exceeds $`60\%`$.
Convergence speed: The convergence speed that we derive in Eq. ([2](#S4.E2 "In IV-B Analysis of the Convergence Speed ‣ IV Impact of the Overlay on BitTorrent’s efficiency ‣ Understanding the Properties of the BitTorrent Overlay")) holds for non NATed peers. When a peer is NATed, it will establish at most $`O_{m\hspace{0pt}a\hspace{0pt}x}`$ outgoing connections, which is the higher bound for its maximum peer set size.
[Refer to caption](/html/0707.1820/assets/x26.png)
Figure 24: The diameter of the overlay vs. the percentage of NATed peers, averaged over $`10`$ independent runs. The error bars indicate the minimum and maximum. The diameter increases slowly with the number of NATed peers.
Diameter of the overlay: NATed peers do not make the diameter significantly larger. For example, as shown in Fig. [24](#S5.F24 "Figure 24 ‣ V-F Impact of the Number of NATed Peers ‣ V Characterizing the Properties of BitTorrent’s Overlay ‣ Understanding the Properties of the BitTorrent Overlay"), when $`10\%`$ (respectively $`80\%`$ of the peers are NATed, the average value of the diameter is at $`4.5`$ (respectively $`5.5`$). However, in extreme cases where $`90\%`$ of the peers are NATed, the diameter can reach $`9`$.
Note that, NATed peers may cause partitions. For example, assume that peer $`P_{i}`$ is NATed. It may happen that, out of the peers returned by the tracker to $`P_{i}`$, no one has room for more connections. As a result, $`P_{i}`$ will not be able to establish any outgoing connections. In addition, and because it is NATed, $`P_{i}`$ cannot receive connections from other peers. As a result, peer $`P_{i}`$ will be isolated alone (disconnected from the rest of the torrent) until it contacts the tracker again and discovers more peers. This behavior becomes more common as the percentage of NATed peers increases.
[Refer to caption](/html/0707.1820/assets/x27.png)
Figure 25: Robustness of the overlay with different percentage of NATed peers. The robustness to churn does not depend on the percentage of NATed peers, but the overlay is not robust to attacks when there are NATed peers.
Robustness to attacks and churn: The robustness of the overlay to high churn rates does not depend on the presence of NATed peers. Indeed, Fig. [25](#S5.F25 "Figure 25 ‣ V-F Impact of the Number of NATed Peers ‣ V Characterizing the Properties of BitTorrent’s Overlay ‣ Understanding the Properties of the BitTorrent Overlay") shows that the overlay stays connected when up to $`80\%`$ of peers leave the torrent. However, the overlay is not robust to attacks when there is a large number of NATed peers. Indeed, Fig. [25](#S5.F25 "Figure 25 ‣ V-F Impact of the Number of NATed Peers ‣ V Characterizing the Properties of BitTorrent’s Overlay ‣ Understanding the Properties of the BitTorrent Overlay") shows that when there are $`50\%`$ of NATed peers, the overlay starts to be partitioned when $`25\%`$ of the peers leave due to an attack. We see that the number of partitions decreases for a large percentage of departing peers, because as there are many small partitions, increasing the number of departing peers removes those partitions.
In summary, NATed peers decrease significantly the robustness of the overlay to attacks.
## VI Impact of Peer Exchange on the Overlay
We have seen that BitTorrent generates overlays with a short diameter that are robust to churn and attacks. However, the time for a peer to reach its maximum peer set size depends on the torrent size and peer arrival rate. One way to reach faster the maximum peer set size is to increase the number of requests to the tracker in order to discover more peers and establish more connections. However, such requests increase the load on the tracker, whereas the tracker is known to have scarce resources \[[18](#bib.bib18), [19](#bib.bib19)\].
In Mai $`2005`$, Azureus 2.3.0.0 \[[20](#bib.bib20)\] introduced a new feature, namely peer exchange (PEX), where neighbors periodically exchange their list of neighbors. For example, assume that peers $`P_{i}`$ and $`P_{j}`$ are neighbors. Then, every minute, $`P_{i}`$ sends its list of neighbors to $`P_{j}`$ and vice versa. As a result, each peer knows its neighbors and the neighbors of its neighbors. The intuition behind PEX is that peers will be able to discover fast a lot of peers and consequently achieve a larger peer set size.
Note that the results that we have given in previous sections are for the case of BitTorrent without PEX, e.g., the official BitTorrent client \[[5](#bib.bib5)\]. In this section, we extend our work and analyze how PEX impacts the overlay topology of BitTorrent. PEX is becoming very popular and, in addition to Azureus, it is now implemented in several other P2P clients including KTorrent, libtorrent, $`\mu`$Torrent, or BitComet, but with incompatible implementations. To the best of our knowledge, the impact of PEX has never been discussed previously.
###
VI-A
Simulating PEX
To evaluate the impact of PEX on the overlay topology, we added this feature to our simulator exactly as it is implemented in Azureus. Concerning the communications between peers and the tracker, all what we described in Section [](#S3.SS2 "III-B Simulation Details ‣ III Simulation Methodology ‣ Understanding the Properties of the BitTorrent Overlay")
III-B
is still valid. That is, the tracker keeps two lists, $`L_{n\hspace{0pt}a\hspace{0pt}t\hspace{0pt}e\hspace{0pt}d}`$ and $`L_{{n\hspace{0pt}o\hspace{0pt}t} - {n\hspace{0pt}a\hspace{0pt}t\hspace{0pt}e\hspace{0pt}d}}`$ peers. And, when a peer $`P_{i}`$ joins the torrent, it gets from the tracker up to $`\sigma`$ peers randomly selected from $`L_{{n\hspace{0pt}o\hspace{0pt}t} - {n\hspace{0pt}a\hspace{0pt}t\hspace{0pt}e\hspace{0pt}d}}`$. Then, $`P_{i}`$ stores those IP addresses in its $`L_{t\hspace{0pt}r\hspace{0pt}a\hspace{0pt}c\hspace{0pt}k\hspace{0pt}e\hspace{0pt}r}^{P_{i}}`$ list and initiates sequentially up to $`O_{m\hspace{0pt}a\hspace{0pt}x}`$ connections to those peers. Moreover, $`P_{i}`$ will be added at the tracker to either $`L_{n\hspace{0pt}a\hspace{0pt}t\hspace{0pt}e\hspace{0pt}d}`$ or $`L_{{n\hspace{0pt}o\hspace{0pt}t} - {n\hspace{0pt}a\hspace{0pt}t\hspace{0pt}e\hspace{0pt}d}}`$.
We now explain the modifications that we made in our simulator. Assume that, at time $`t = 0`$, a connection has been established between two peers $`P_{i}`$ and $`P_{j}`$. Just after, $`P_{i}`$ sends its list of neighbors to $`P_{j}`$ and vice versa. Then, every $`1`$ simulated minute, $`P_{i}`$ and $`P_{j}`$ repeat this exchange process.
Assume now that, after performing PEX with its neighbor $`P_{j}`$, $`P_{i}`$ discovers peer $`P_{k}`$. Then, $`P_{i}`$ checks whether (1) it already has a connection with $`P_{k}`$ or (2) it already knows $`P_{k}`$ from the tracker. If none of these two conditions holds true, then $`P_{i}`$ adds $`P_{k}`$ to the list $`L_{p\hspace{0pt}e\hspace{0pt}x}^{P_{i}}`$ of peers it discovered through PEX. Note that, $`P_{i}`$ may receive the IP address of $`P_{k}`$ from many neighbors. In this case, $`P_{k}`$ will appear only once in $`L_{p\hspace{0pt}e\hspace{0pt}x}^{P_{i}}`$.
Note that, when establishing connections, peers discovered from the tracker are given more priority. For example, assume that peer $`P_{i}`$ decides to initiate a new connection, which can be due to the departure of one of its neighbors or after discovering new peers. In this case, $`P_{i}`$ contacts first the peers it has discovered from the tracker. If none of those peers accepts the connection request, $`P_{i}`$ contacts the peers that it discovered through PEX.
###
VI-B
Analysis of PEX
We implement PEX in our simulator and run simulations with the following parameters. We set the maximum peer set size to $`80`$, the maximum number of outgoing connections to $`40`$, the minimum number of neighbors to $`20`$, and the number of peers returned by the tracker to $`50`$. However, due to the gossiping messages between peers, the PEX feature makes our simulator very slow. In order to save time, we run simulations for a torrent of $`1000`$ peers that arrive to the torrent within the first $`60`$ simulated minutes according to Eq. [](#S3.Ex1 "III-D Arrival Distribution of Peers ‣ III Simulation Methodology ‣ Understanding the Properties of the BitTorrent Overlay")
III-D
. The departure of peers is scheduled during the next simulated $`60`$ minutes, and it follows a random uniform distribution. For example, if peer $`A`$ arrives at time $`t = 30`$, it will leave the network at a random time uniformly selected between $`t = 60`$ and $`t = 120`$. Still, this torrent allows us to understand how the overlay is constructed with PEX and how it evolves as peers join and leave.
[Refer to caption](/html/0707.1820/assets/x28.png)
Figure 26: Evolution of the outdegree of peers with time with PEX, averaged over ten independent runs. The error bars indicate the minimum and maximum. PEX produces an average peer set size that is very close to the maximum peer set size.
Average peer set size: PEX meets its intended goal and permits peers to be at their maximum peer set size most of the time as shown in Fig. [26](#S6.F26 "Figure 26 ‣ VI-B Analysis of PEX ‣ VI Impact of Peer Exchange on the Overlay ‣ Understanding the Properties of the BitTorrent Overlay"). Moreover, PEX prevents the average peer set size from decreasing in case of a massive departure of peers. Indeed, when a peer loses a connection due to the departure of a neighbor, it can replace it by a new connection to one of its neighbors’ neighbors. As a result, the average peer set size stays at its maximum value of $`80`$ as long as there are $`80`$ peers in the torrent.
Convergence speed: Each peer reaches its maximum peer set size within a few gossiping period (that we set to one minute in our simulations).
[Refer to caption](/html/0707.1820/assets/x29.png)
Figure 27: Evolution of the overlay diameter with time with PEX, averaged over ten independent runs. We plot the mean, max, and min values. PEX generates overlays with very large diameter.
[Refer to caption](/html/0707.1820/assets/x30.png)
Figure 28: The topology evolution with PEX with 9 peers. Connections established with the tracker information are shown with solid lines, and connection established with PEX are shown with dashed lines.
Diameter of the overlay: The increase in the average peer set size comes at the expense of a larger diameter. As we see in Fig. [27](#S6.F27 "Figure 27 ‣ VI-B Analysis of PEX ‣ VI Impact of Peer Exchange on the Overlay ‣ Understanding the Properties of the BitTorrent Overlay"), the maximum value of the diameter reaches $`20`$ when the number of peers in the torrent is $`1000`$.
To explain why PEX produces such a long diameter, we plot in Fig. [28](#S6.F28 "Figure 28 ‣ VI-B Analysis of PEX ‣ VI Impact of Peer Exchange on the Overlay ‣ Understanding the Properties of the BitTorrent Overlay") the evolution of a torrent with 9 peers ($`P_{1},\ldots,P_{9}`$) that arrive sequentially, one every 1 unit of time. In this example, we set the maximum peer set size to 4, the maximum number of outgoing connections to 2, and the number of peers returned by the tracker to 2. At time $`t = 0`$, there is only peer $`P_{1}`$. At time $`t = 2`$, $`P_{2}`$ joins the torrent and connects to $`P_{1}`$. At time $`t = 3`$, $`P_{3}`$ arrives and connects to $`P_{1}`$ and $`P_{2}`$. Then, $`P_{4}`$ arrives and connects to two existing peers selected at random, say $`P_{1}`$ and $`P_{2}`$. At the end of time $`t = 3`$, PEX has not yet been used. At time $`t = 4`$, $`P_{5}`$ arrives and connects to $`P_{3}`$ and $`P_{4}`$, which in turn tell $`P_{1}`$ and $`P_{2}`$ about this new neighbor. At this time, $`P_{1}`$ and $`P_{2}`$ each has a room for one more outgoing connection and both connect to $`P_{5}`$. Thus, at time $`t = 4`$, only peers $`P_{3}`$ and $`P_{4}`$ can accept new incoming connections, i.e., $`P_{1}`$, $`P_{2}`$, and $`P_{5}`$ have already reached the maximum number of connections. At time $`t = 5`$, peer $`P_{6}`$ joins the torrent and gets from the tracker the addresses of $`P_{1}`$ and $`P_{4}`$. However, $`P_{6}`$ can only initiate a connection to $`P_{4}`$, as $`P_{5}`$ has already reached its maximum peer set size. Peer $`P_{7}`$ arrives at time $`t = 6`$, gets the addresses of $`P_{3}`$ and $`P_{5}`$ and initiates one connection to $`P_{3}`$. Then, $`P_{8}`$ joins the torrent at time $`t = 7`$, obtains the addresses of $`P_{2}`$ and $`P_{6}`$ and initiates only one connection to $`P_{6}`$. At time $`t = 8`$, peer $`P_{9}`$ arrives and gets the addresses of $`P_{6}`$ and $`P_{7}`$. Given that these two peers have not reached their maximum peer set size, $`P_{9}`$ succeeds to connect to both of them. Afterward, $`P_{6}`$ tells its neighbor $`P_{8}`$ about $`P_{9}`$, thus a new connection is initiated from $`P_{8}`$ to $`P_{9}`$. Then, $`P_{9}`$ tells $`P_{7}`$ about $`P_{8}`$ and a new connection is initiated from $`P_{7}`$ to $`P_{8}`$.
As we can see, PEX tries to maximize the number of outgoing connections at each peer. Peers keep on gossiping and whenever they discover new peers, they establish new connections if they still have room for. However, the disadvantage is that peers that arrive at the beginning establish a lot of connections among each other and leave only a few free connections for the peers that arrive afterward. In the example that we consider here, $`P_{1},\ldots,P_{5}`$ are highly interconnected and they leave only two connections for next peers. Similarly, when peers $`P_{6},\ldots,P_{9}`$ arrive, they connect to the overlay at peers $`P_{3}`$ and $`P_{4}`$ and then interconnect strongly with each other.
[Refer to caption](/html/0707.1820/assets/x31.png)
Figure 29: The connectivity matrix of the overlay distribution with PEX after $`59`$ minutes, for a single run. A dot at (i,j) means that i and j are neighbors. At $`t = 59`$, almost all the $`1000`$ peers have joined the network and no peer has left yet. We observe a chain-like overlay.
As a result, PEX leads to a clustering phenomena, where each cluster contains approximately a number of peers close to the maximum peer set size. Each cluster exhibits a high intra-cluster connectivity and a poor inter-cluster connectivity with the cluster that arrives just before and to the one that arrives just after. To confirm our analysis, we draw in Fig. [29](#S6.F29 "Figure 29 ‣ VI-B Analysis of PEX ‣ VI Impact of Peer Exchange on the Overlay ‣ Understanding the Properties of the BitTorrent Overlay") a snapshot of the connectivity matrix of the overlay distribution after $`60`$ minutes when the first $`1000`$ peers have arrived in the torrent. The clustering phenomena appears clearly in the figure, which explains the large diameter of the overlay. As we explained in Section [](#S4.SS3 "IV-C Impact of the Diameter of the Distribution Overlay ‣ IV Impact of the Overlay on BitTorrent’s efficiency ‣ Understanding the Properties of the BitTorrent Overlay")
IV-C
, such a chain-like overlay constraints the distribution time in the system to be a linear function of the number of clusters. As compared to the overlay generate by the tracker only, this chain-like overlay becomes less efficient when the number of clusters becomes larger than the number of pieces. Typical files distributed using BitTorrent includes an average of $`1000`$ pieces. In this case, this chain-like overlay will become inefficient when the number of clusters is larger than $`1000`$, i.e., the number of peers is larger than $`800.000`$ peers. Current torrents are much smaller and they rarely exceed 100.000 peers and therefore, no one has noticed yet the negative impact of the PEX on the download time of files.
Let us now go back to Fig. [27](#S6.F27 "Figure 27 ‣ VI-B Analysis of PEX ‣ VI Impact of Peer Exchange on the Overlay ‣ Understanding the Properties of the BitTorrent Overlay"). If we carefully look at this figure, at time $`t = 60`$ minutes, the average value of the diameter drops from $`18`$ to $`11`$. Actually, at $`t = 60`$ minutes, peers start leaving the torrent. At the same time, the last peers to join the torrent also arrive at $`t = 60`$ minutes. Those departed peers will allow the arriving ones to connect at different levels of the chain and not only at the tail, which consequently reduces the diameter of the overlay.
To better explain this behavior, consider the overlay shown in Fig. [29](#S6.F29 "Figure 29 ‣ VI-B Analysis of PEX ‣ VI Impact of Peer Exchange on the Overlay ‣ Understanding the Properties of the BitTorrent Overlay"). In this chain-like overlay, all peers are at their maximum peer set size except those that are at the tail. Assume now that peers $`{P_{1},\ldots,P_{15}},`$ leave the torrent. Assume that, at the same time, $`{P_{985},\ldots,P_{1000}},`$ join the torrent. In this case, the departed peers $`{P_{1},\ldots,P_{15}},`$ will leave rooms for incoming connections inside the first cluster, i.e., at the head of the chain. Thus, the arriving peers $`{P_{985},\ldots,P_{1000}},`$ will be able to establish connections to the tail as well as the head of the chain. As a result, the head and tail of the chain become connected and the diameter of the overlay drops by half. Still, the diameter remains very high when compared to an overlay generated only by the tracker.
[Refer to caption](/html/0707.1820/assets/x32.png)
\(a\)
Impact of attack on the PEX overlay
[Refer to caption](/html/0707.1820/assets/x33.png)
\(b\)
Impact of churn on the PEX overlay
Figure 30: Robustness of the chain-like overlay generated by PEX. The PEX overlay can be easily partitioned by simply attacking a peer, its neighbors, and its neighbors of neighbors.
Robustness to attacks and churn: Surprisingly, the overlay produced with PEX is robust to churn rate and to the attack that targets the most connect peers. As shown in Fig. [30(a)](#S6.F30.sf1 "In Figure 30 ‣ VI-B Analysis of PEX ‣ VI Impact of Peer Exchange on the Overlay ‣ Understanding the Properties of the BitTorrent Overlay") and Fig. [30(b)](#S6.F30.sf2 "In Figure 30 ‣ VI-B Analysis of PEX ‣ VI Impact of Peer Exchange on the Overlay ‣ Understanding the Properties of the BitTorrent Overlay"), the overlay stays connected with up to $`80\%`$ of the peers leaving the torrent.
As we can see in Fig. [30(a)](#S6.F30.sf1 "In Figure 30 ‣ VI-B Analysis of PEX ‣ VI Impact of Peer Exchange on the Overlay ‣ Understanding the Properties of the BitTorrent Overlay"), up to $`80\%`$ of departed peers, the attack scenario produces a maximum number of partitions of $`2`$. Actually, out of the ten runs that we performed, we obtained partitions with only $`1`$ run. In particular, we obtained one major partition and a second one with only $`1`$ peer. For example, for $`80\%`$ of departed rate, we obtained $`1`$ partition that includes $`167`$ peers and a second partition with only $`1`$ peer. Same conclusions apply on the churn scenario.
Even though PEX shows good robustness to the attack that we have been using so far, this chain-like overlay can be easily partitioned by using more sophisticated attacks that target a peer, its neighbors, and its neighbors of neighbors.
In summary, even if PEX significantly decreases the time for a peer to reach its maximum peer set size, it creates a chain-like overlay that is not robust against partitions and whose diameter is large. This large diameter will lead to a long download time of files when the number of simultaneous peers is large. We plan to evaluate how much this overlay impacts the efficiency of the transfer when compared to an overlay created only by the tracker.
## VII Discussion
###
VII-A
Summary of our Contributions
We have conducted a large set of simulations to investigate the properties of the overlay formed by BitTorrent. Below is a list of our main contributions.
- •
First, we have analyzed the relation between the overlay properties and the performance of BitTorrent. In particular, we have shown that a large peer set size increases the efficiency of BitTorrent, and that a small diameter is a necessary, but not sufficient, condition for this efficiency.
- •
Second, we have shown for the first time that the overlay generated by BitTorrent is not a random graph, as it is commonly believed. The connectivity of a peer with neighbors in the torrent is highly biased by its arriving order in the torrent. Whereas it is beyond the scope of this study to evaluate the robustness of the overlay structure to elaborated attacks, i.e., attacks that do not only focus on the most connected peers, it is an interesting area for future research. In particular, it is critical to understand such issues when a public service is to be built on top of BitTorrent.
- •
Third, we have evaluated the impact of the maximum peer set size and of the maximum number of outgoing connections. Whereas there are several magic numbers in BitTorrent, we have identified that the maximum peer set size is a tradeoff between efficiency and peers overhead, and we have explained why the maximum number of outgoing connections must be set to half of the maximum peer set size.
- •
Finally, we have identified two potentially significant problems in the overlay, which deserve further investigations. We have shown that a large number of NATed peers decrease significantly the robustness of the overlay to attacks, and we have shown that peer exchange creates a chain-like overlay that might adversely impact the efficiency of BitTorrent.
In conclusion, we expect this study to shed light on the impact of the overlay structure on BitTorrent efficiency, and to foster further researches in that direction.
###
VII-B
Future Work
Our future work will progress along two avenues.
- •
Mitigate the impact of NATed peers on the robustness of the overlay. Actually, with its current implementation, BitTorrent produces an overlay where non-NATed peers have a higher connectivity than NATed ones. As a result, one can create partitions by attacking the non-NATed peers, which are the most connected ones. One possible solution to this problem is to allow NATed peers to initiate more connections than non-NATed ones. For example, one can imagine that the tracker reports the number of NATed peers to new peers so that they can weight their maximum number of outgoing connections. Our goal is to still have a highly connected graph, but without peers with significantly more connections. The intuition behind this solution is that the robustness of the overlay would improve. This solution, and in particular how to weight the maximum number of outgoing connections, will be subject to further investigation.
- •
Extend peer exchange in order to still converge fast to the maximum peer set size while maintaining a low diameter overlay. Indeed, with the current implementation of peer exchange, peers converge fast to their maximum peer set size, but only peers that are at the tail of the overlay chain have rooms for incoming connections. As a result, new arriving peers can only connect to the tail of the overlay chain. We are investigating possible solutions to this problem whose main goal is to add randomness in the overlay generated with peer exchange.
One solution is to allow peers coming from the tracker to preempt connections of peers discovered with peer exchange. For example, assume that peer $`P_{i}`$ has reached its maximum peer set size and amongst its neighbors, there is $`P_{j}`$ that it discovered with peer exchange. Assume now that $`P_{k}`$ joins the torrent and receives the IP address of $`P_{i}`$ from the tracker. If $`P_{k}`$ initiates a connection to $`P_{i}`$, $`P_{i}`$ will accept this connection and drop its connection to $`P_{j}`$.
Another solution is to add randomness during the construction of the overlay. For instance, instead of collecting a list of neighbors of its neighbors, which creates locality in the graph construction, a peer $`P_{i}`$ can ask neighbors to randomly selected peers. The rational is to discover and create connections to peers that are far from $`P_{i}`$ in the overlay. The choice of the random function to discover those peers is critical and currently under investigation.
## Acknowledgments
We would like to thank Olivier Chalouhi, Alon Rohter, and Paul Gardner from Azureus Inc. for their information on peer exchange.
## References
- \[1\] Y. H. Chu, S. G. Rao, and H. Zhang, “A case for end system multicast,” in *Proc. of ACM SIGMETRICS*, Santa Clara, CA, USA, June 2000, pp. 1–12.
- \[2\] B. Hill, “P2P: 70-80 percent of all euronet traffic,” The Digital Music Weblog, May 2004, http://digitalmusic.weblogsinc.com/.
- \[3\] B. Cohen, “Incentives to build robustness in Bittorrent,” in *Proc. of the Workshop on Economics of Peer-to-Peer Systems*, Berkeley, CA, USA, June 2003, http://bitconjurer.org/BitTorrent.
- \[4\] “BitTorrent beats Kazaa, accounts for 53% of P2P traffic,” AlwaysOn Site, July 2004, http://www.alwayson-network.com/index.php.
- \[5\] “Bittorrent, inc,” http://www.bittorrent.com.
- \[6\] A. Legout, G. Urvoy-Keller, and P. Michiardi, “Rarest first and choke algorithms are enough,” in *Proc. of IMC*, Rio de Janeiro, Brazil, October 2006.
- \[7\] A. Legout, N. Liogkas, E. Kohler, and L. Zhang, “Clustering and sharing incentives in bittorrent systems,” in *Proc. of SIGMETRICS*, San Diego, CA, USA, June 2007.
- \[8\] T. Locher, P. Moor, S. Schmid, and R.Wattenhofer, “Free riding in bittorrent is cheap,” in *Proc. of HotNets-V*, Irvine, CA, USA, November 2006.
- \[9\] Y. Tian, D. Wu, and K.-W. Ng, “Modeling, analysis and improvement for bittorrent-like file sharing networks,” in *Proc. of INFOCOM*, Barcelona, Spain, April 2006.
- \[10\] M. Izal, G. Urvoy-Keller, E. W. Biersack, P. A. Felber, A. A. Hamra, and L. Garcés-Erice, “Dissecting bittorrent: Five months in a torrent’s lifetime,” in *Proc. of PAM*, Juan-les-Pins, France, April 2004.
- \[11\] L. Guo, S. Chen, Z. Xiao, E. Tan, X. Ding, and X. Zhang, “Measurement, analysis, and modeling of bittorrent-like systems,” in *Proc. of IMC*, New Orleans, LA, USA, October 2005.
- \[12\] D. Qiu and R. Srikant, “Modeling and performance analysis of bittorrent-like peer-to-peer networks,” in *Proc. of SIGCOMM*, Portland, Oregon, USA, August 2004.
- \[13\] G. Urvoy-Keller and P. Michiardi, “Impact of inner parameters and overlay structure on the performance of bittorrent,” in *Proc. of Global Internet Symposium*, Barcelona, Spain, April 2006.
- \[14\] N. Liogkas, R. Nelson, E. Kohler, and L. Zhang, “Exploiting bittorrent for fun (but not profit),” in *Proc. of IPTPS*, Santa Barbara, CA, USA, February 2006.
- \[15\] Http://planete.inria.fr/software/BitSim/.
- \[16\] D. Stutzbach and R. Rejaie, “Characterizing the two-tier gnutella topology,” in *Proc. of SIGMETRICS ’05*, Banff, Alberta, Canada, June 2005.
- \[17\] X. Yang and G. de Veciana, “Service capacity of peer-to-peer networks,” in *Proc. of INFOCOM*, Hong-Kong, China, March 2004.
- \[18\] G. Neglia, G. Reina, H. Zhang, D. Towsley, A. Venkataramani, and J. Danaher, “Availability in bittorrent systems,” in *Proc. of INFOCOM*, Anchorage, Alaska, USA, May 2007.
- \[19\] J. Pouwelse, P. Garbacki, D. Epema, and H. Sips, “The bittorrent p2p file-sharing system: Measurements and analysis,” in *Proc. of IPTPS*, Ithaca, NY, USA, February 2005.
- \[20\] “Azureus,” http://sourceforge.net/projects/azureus/. |
0706.3812 | cs.CR | cyber | INSTITUT NATIONAL DE RECHERCHE EN INFORMATIQUE ET EN AUTOMATIQUE
Java Components Vulnerabilities
-
An Experimental Classification
Targeted at the OSGi Platform
Pierre Parrend — Stéphane Frénot N° 6231
June 2007
Java Components Vulnerabilities
-
An Experimental Classification
Targeted at the OSGi Platform
<sup>†</sup>
<sup>†</sup>thanks: This Work is partialy founded by Muse IST Project n°026442.
Pierre Parrend, Stéphane Frénot
Thème COM — Systèmes communicants
Projet ARES
Rapport de recherche n° 6231 — June 2007 — ?? pages
Abstract: The OSGi Platform finds a growing interest in two different applications domains: embedded systems, and applications servers. However, the security properties of this platform are hardly studied, which is likely to hinder its use in production systems. This is all the more important that the dynamic aspect of OSGi-based applications, that can be extended at runtime, make them vulnerable to malicious code injection.
We therefore perform a systematic audit of the OSGi platform so as to build a vulnerability catalog that intends to reference OSGi Vulnerabilities originating in the Core Specification, and in behaviors related to the use of the Java language. Implementation of Standard Services are not considered.
To support this audit, a Semi-formal Vulnerability Pattern is defined, that enables to uniquely characterize fundamental properties for each vulnerability, to include verbose description in the pattern, to reference known security protections, and to track the implementation status of the proof-of-concept OSGi Bundles that exploit the vulnerability.
Based on the analysis of the catalog, a robust OSGi Platform is built, and recommendations are made to enhance the OSGi Specifications.
Key-words: OSGi<sup>tm</sup> Platform, Security, Dependability, Java, Hardened Execution Platform, Vulnerability Catalog
Vulnerabilités des Composants Java
\-
Une Classification Expérimentale
Dans le Cadre de la Plate-forme OSGi
Résumé : La plate-forme d’exécution OSGi rencontre un intérêt grandissant dans deux domaines d’applications différents: les systèmes embarqués, et les serveurs d’applications. Cependant, les propriétés de cette plate-forme relatives à la sécurité ne sont que très peu étudiées, ce qui peut fortement freiner son adoption dans les systèmes industriels. Ceci est d’autant plus critique que la possibilité d’extension dynamique à l’exécution offerte par la plate-forme OSGi rend celle-ci vulnérable à l’injection de code malicieux.
Nous effectuons un audit de l’environnement d’exécution OSGi, afin de créer un catalogue de vulnérabilités. Nous cherchons à référencer les vulnérabilités causées par la spécification ‘Core’, ou bien par la machine virtuelle Java sous-jacente. Les autres éléments définis par OSGi, comme les services standards, ne sont pas considérés.
Afin de mener à bien cet audit, nous définissons un Pattern de Vulnérabilité semi-formel, qui permet de décrire les caractéristiques des vulnérabilités de manière unique, de donner des informations complémentaires, de référencer les protections existantes, et d’identifier le status de l’implémentation des Bundles OSGi de test qui démontrent chaque vulnérabilité.
A partir de cette analyse, un plate-forme OSGi robuste est construite, et des recommandations pour les spécifications OSGi sont données.
Mots-clés : Plate-forme OSGi<sup>tm</sup>, Sécurité, Java, Plate-forme d’exécution renforcée, Catalogue de Vulnérabilités
###### Contents
1. [](#S1)
1 Introduction
2. [](#S2)
2 Characterization of Vulnerabilities in Component-based Systems
1. [](#S2.SS1 "In 2 Characterization of Vulnerabilities in Component-based Systems")
2.1 Definitions
2. [](#S2.SS2 "In 2 Characterization of Vulnerabilities in Component-based Systems")
2.2 From Databases to Top-Vulnerability Lists
3. [](#S2.SS3 "In 2 Characterization of Vulnerabilities in Component-based Systems")
2.3 Vulnerability Patterns
3. [](#S3)
3 The Semi-formal Software Vulnerability Pattern
1. [](#S3.SS1 "In 3 The Semi-formal Software Vulnerability Pattern")
3.1 The Structure of the Semi-formal Vulnerability Pattern
2. [](#S3.SS2 "In 3 The Semi-formal Software Vulnerability Pattern")
3.2 Vulnerability Taxonomies for OSGi-based Systems
3. [](#S3.SS3 "In 3 The Semi-formal Software Vulnerability Pattern")
3.3 A Vulnerability Example: ‘Management Utility Freezing - Infinite Loop’
4. [](#S4)
4 Requirements for secure OSGi Systems
1. [](#S4.SS1 "In 4 Requirements for secure OSGi Systems")
4.1 Catalog Analysis
2. [](#S4.SS2 "In 4 Requirements for secure OSGi Systems")
4.2 Requirements for a Hardened OSGi Platform
3. [](#S4.SS3 "In 4 Requirements for secure OSGi Systems")
4.3 Recommendations for a Hardened Execution Environment
5. [](#S5)
5 Conclusions
6. [](#A1)
A The OSGi platform
1. [](#A1.SS1 "In Appendix A The OSGi platform")
A.1 Overview
2. [](#A1.SS2 "In Appendix A The OSGi platform")
A.2 The Bundles
3. [](#A1.SS3 "In Appendix A The OSGi platform")
A.3 Interactions between Bundles
7. [](#A2)
B Vulnerabilities List
1. [](#A2.SS1 "In Appendix B Vulnerabilities List")
B.1 The Lindqvist Classification
2. [](#A2.SS2 "In Appendix B Vulnerabilities List")
B.2 Common Weaknesses Enumeration (CWE)
3. [](#A2.SS3 "In Appendix B Vulnerabilities List")
B.3 Nineteen Dealy Sins
4. [](#A2.SS4 "In Appendix B Vulnerabilities List")
B.4 OWASP Top Ten
5. [](#A2.SS5 "In Appendix B Vulnerabilities List")
B.5 Seven Kingdoms
8. [](#A3)
C Formal Expression of the Vulnerability Pattern
9. [](#A4)
D Vulnerability Catalog
1. [](#A4.SS1 "In Appendix D Vulnerability Catalog")
D.1 Bundle Archive
1. [](#A4.SS1.SSS1 "In D.1 Bundle Archive ‣ Appendix D Vulnerability Catalog")
D.1.1 Invalid Digital Signature Validation
2. [](#A4.SS1.SSS2 "In D.1 Bundle Archive ‣ Appendix D Vulnerability Catalog")
D.1.2 Big Component Installer
3. [](#A4.SS1.SSS3 "In D.1 Bundle Archive ‣ Appendix D Vulnerability Catalog")
D.1.3 Decompression Bomb
2. [](#A4.SS2 "In Appendix D Vulnerability Catalog")
D.2 Bundle Manifest
1. [](#A4.SS2.SSS1 "In D.2 Bundle Manifest ‣ Appendix D Vulnerability Catalog")
D.2.1 Duplicate Package Import
2. [](#A4.SS2.SSS2 "In D.2 Bundle Manifest ‣ Appendix D Vulnerability Catalog")
D.2.2 Excessive Size of Manifest File
3. [](#A4.SS2.SSS3 "In D.2 Bundle Manifest ‣ Appendix D Vulnerability Catalog")
D.2.3 Erroneous values of Manifest attributes
3. [](#A4.SS3 "In Appendix D Vulnerability Catalog")
D.3 Bundle Activator
1. [](#A4.SS3.SSS1 "In D.3 Bundle Activator ‣ Appendix D Vulnerability Catalog")
D.3.1 Management Utility Freezing - Infinite Loop
2. [](#A4.SS3.SSS2 "In D.3 Bundle Activator ‣ Appendix D Vulnerability Catalog")
D.3.2 Management Utility Freezing - Thread Hanging
4. [](#A4.SS4 "In Appendix D Vulnerability Catalog")
D.4 Bundle Code - Native
1. [](#A4.SS4.SSS1 "In D.4 Bundle Code - Native ‣ Appendix D Vulnerability Catalog")
D.4.1 Runtime.exec.kill
2. [](#A4.SS4.SSS2 "In D.4 Bundle Code - Native ‣ Appendix D Vulnerability Catalog")
D.4.2 CPU Load Injection
5. [](#A4.SS5 "In Appendix D Vulnerability Catalog")
D.5 Bundle Code - Java
1. [](#A4.SS5.SSS1 "In D.5 Bundle Code - Java ‣ Appendix D Vulnerability Catalog")
D.5.1 System.exit
2. [](#A4.SS5.SSS2 "In D.5 Bundle Code - Java ‣ Appendix D Vulnerability Catalog")
D.5.2 Runtime.halt
3. [](#A4.SS5.SSS3 "In D.5 Bundle Code - Java ‣ Appendix D Vulnerability Catalog")
D.5.3 Recursive Thread Creation
4. [](#A4.SS5.SSS4 "In D.5 Bundle Code - Java ‣ Appendix D Vulnerability Catalog")
D.5.4 Hanging Thread
5. [](#A4.SS5.SSS5 "In D.5 Bundle Code - Java ‣ Appendix D Vulnerability Catalog")
D.5.5 Sleeping Bundle
6. [](#A4.SS5.SSS6 "In D.5 Bundle Code - Java ‣ Appendix D Vulnerability Catalog")
D.5.6 Big File Creator
7. [](#A4.SS5.SSS7 "In D.5 Bundle Code - Java ‣ Appendix D Vulnerability Catalog")
D.5.7 Code Observer
8. [](#A4.SS5.SSS8 "In D.5 Bundle Code - Java ‣ Appendix D Vulnerability Catalog")
D.5.8 Component Data Modifier
9. [](#A4.SS5.SSS9 "In D.5 Bundle Code - Java ‣ Appendix D Vulnerability Catalog")
D.5.9 Hidden Method Launcher
10. [](#A4.SS5.SSS10 "In D.5 Bundle Code - Java ‣ Appendix D Vulnerability Catalog")
D.5.10 Memory Load Injection
11. [](#A4.SS5.SSS11 "In D.5 Bundle Code - Java ‣ Appendix D Vulnerability Catalog")
D.5.11 Stand Alone Infinite Loop
12. [](#A4.SS5.SSS12 "In D.5 Bundle Code - Java ‣ Appendix D Vulnerability Catalog")
D.5.12 Infinite Loop in Method Call
13. [](#A4.SS5.SSS13 "In D.5 Bundle Code - Java ‣ Appendix D Vulnerability Catalog")
D.5.13 Exponential Object Creation
6. [](#A4.SS6 "In Appendix D Vulnerability Catalog")
D.6 Bundle Code - OSGi APi
1. [](#A4.SS6.SSS1 "In D.6 Bundle Code - OSGi APi ‣ Appendix D Vulnerability Catalog")
D.6.1 Launch a Hidden Bundle
2. [](#A4.SS6.SSS2 "In D.6 Bundle Code - OSGi APi ‣ Appendix D Vulnerability Catalog")
D.6.2 Pirat Bundle Manager
3. [](#A4.SS6.SSS3 "In D.6 Bundle Code - OSGi APi ‣ Appendix D Vulnerability Catalog")
D.6.3 Zombie Data
4. [](#A4.SS6.SSS4 "In D.6 Bundle Code - OSGi APi ‣ Appendix D Vulnerability Catalog")
D.6.4 Cycle Between Services
5. [](#A4.SS6.SSS5 "In D.6 Bundle Code - OSGi APi ‣ Appendix D Vulnerability Catalog")
D.6.5 Numerous Service Registration
6. [](#A4.SS6.SSS6 "In D.6 Bundle Code - OSGi APi ‣ Appendix D Vulnerability Catalog")
D.6.6 Freezing Numerous Service Registration
7. [](#A4.SS7 "In Appendix D Vulnerability Catalog")
D.7 Bundle Fragments
1. [](#A4.SS7.SSS1 "In D.7 Bundle Fragments ‣ Appendix D Vulnerability Catalog")
D.7.1 Execute Hidden Classes
2. [](#A4.SS7.SSS2 "In D.7 Bundle Fragments ‣ Appendix D Vulnerability Catalog")
D.7.2 Fragment Substitution
3. [](#A4.SS7.SSS3 "In D.7 Bundle Fragments ‣ Appendix D Vulnerability Catalog")
D.7.3 Access Protected Package through split Packages
10. [](#A5)
E Attack Implementations
1. [](#A5.SS1 "In Appendix E Attack Implementations")
E.1 Infinite Loops
1. [](#A5.SS1.SSS1 "In E.1 Infinite Loops ‣ Appendix E Attack Implementations")
E.1.1 First Option
2. [](#A5.SS1.SSS2 "In E.1 Infinite Loops ‣ Appendix E Attack Implementations")
E.1.2 Second Option
3. [](#A5.SS1.SSS3 "In E.1 Infinite Loops ‣ Appendix E Attack Implementations")
E.1.3 Third Option
4. [](#A5.SS1.SSS4 "In E.1 Infinite Loops ‣ Appendix E Attack Implementations")
E.1.4 Fourth Option
5. [](#A5.SS1.SSS5 "In E.1 Infinite Loops ‣ Appendix E Attack Implementations")
E.1.5 Fifth Option
6. [](#A5.SS1.SSS6 "In E.1 Infinite Loops ‣ Appendix E Attack Implementations")
E.1.6 Other Implementations
2. [](#A5.SS2 "In Appendix E Attack Implementations")
E.2 Hanging Thread
1. [](#A5.SS2.SSS1 "In E.2 Hanging Thread ‣ Appendix E Attack Implementations")
E.2.1 First Option
2. [](#A5.SS2.SSS2 "In E.2 Hanging Thread ‣ Appendix E Attack Implementations")
E.2.2 Second Option
3. [](#A5.SS2.SSS3 "In E.2 Hanging Thread ‣ Appendix E Attack Implementations")
E.2.3 Other Implementations
11. [](#A6)
F XML2Tex Documentation Generator
###### List of Figures
1. [](#S3.F1 "Figure 1In 3 The Semi-formal Software Vulnerability Pattern")
1 Potential Locations of malicious Code in a Bundle
2. [](#S3.F2 "Figure 2In 3 The Semi-formal Software Vulnerability Pattern")
2 Vulnerability Sources in an OSGi-based System
3. [](#S3.F3 "Figure 3In 3 The Semi-formal Software Vulnerability Pattern")
3 Potential Targets of Attacks against an OSGi Platform
4. [](#S3.F4 "Figure 4In 3 The Semi-formal Software Vulnerability Pattern")
4 Consequences of the Vulnerabilities of the OSGi Platform
5. [](#S3.F5 "Figure 5In 3 The Semi-formal Software Vulnerability Pattern")
5 Introduction Time for the identified flaws
6. [](#S3.F6 "Figure 6In 3 The Semi-formal Software Vulnerability Pattern")
6 Exploit Time for the identified Flaws
7. [](#S4.F7 "Figure 7In 4 Requirements for secure OSGi Systems")
7 Entities that are Source of the vulnerabilities
8. [](#S4.F8 "Figure 8In 4 Requirements for secure OSGi Systems")
8 Functions that prove to be dangerous in the context of an OSGi Platform
9. [](#S4.F9 "Figure 9In 4 Requirements for secure OSGi Systems")
9 Flaws in an OSGi Execution Environment
10. [](#S4.F10 "Figure 10In 4 Requirements for secure OSGi Systems")
10 Targets of Attacks against an OSGi Execution environment
11. [](#S4.F11 "Figure 11In 4 Requirements for secure OSGi Systems")
11 Actual Protection Mechanisms
12. [](#S4.F12 "Figure 12In 4 Requirements for secure OSGi Systems")
12 Cardinality for each potential Security Mechanisms for the OSGi platform
13. [](#A1.F13 "Figure 13In Appendix A The OSGi platform")
13 Overview of an OSGi Platform
14. [](#A1.F14 "Figure 14In Appendix A The OSGi platform")
14 Intern Structure of an OSGi bundle
15. [](#A1.F15 "Figure 15In Appendix A The OSGi platform")
15 Life Cycle of an OSGi Bundles inside the platform
16. [](#A1.F16 "Figure 16In Appendix A The OSGi platform")
16 Interaction Mechanisms between the OSGi Bundles
17. [](#A6.F17 "Figure 17")
17 XML2Tex Process
###### List of Tables
1. [](#S4.T1 "Table 1In 4 Requirements for secure OSGi Systems")
1 Vulnerabilities for the main Open Source OSGi Platforms
## 1 Introduction
The OSGi Platform, which enables multi-application management over Java Virtual Machines, is currently seeing a dramatic increase in its application domains. First targeted at embedded systems such as multimedia automotive devices, it has since widespread in the world of applications, with the Eclipse Integrated Development Environment, and then to application servers, such as IBM Websphere 6.1, or recent development with JBoss. Sun is envisioning to integrate it in the Sun JVM, and several Java Specification Request (JSR) work groups have been set up on the subject
<sup>1</sup>
<sup>1</sup>1http://jcp.org/en/jsr/detail?id=277,http://jcp.org/en/jsr/detail?id=291
.
However, target systems are likely to be highly networked ones, and security implications have so far been mostly overlooked. Actually, the runtime extensibility of applications that is supported by the OSGi platform open a brand new attack vector: code can be installed on the fly, and no mechanism currently guarantees that this code is not malicious. As OSGi-based systems move from Open-Source projects and closed embedded devices toward large-scale systems, this weakness can turn into a major vulnerability, unless security implications are better understood. We therefore perform in this study a systematic analysis of vulnerabilities that are implied by OSGi bundles, and propose adequate counter-measures.
Up to now, Two complementary mechanisms are used to enforce security in the context of OSGi-based systems. The first mechanism is bundle digital signature \[[OSG05](#bib.bibx24), [PF06](#bib.bibx25)\], which guarantees that only bundles from trusted issuers are installed. This trust requirement forces the issuer to publish only safe bundles, since he will liable for any incident caused by the code he provides. The second mechanism is based on Java permissions, that enable to switch on or off some attack-prone features of the Java Virtual Machine. These mechanisms are mostly insufficient to guarantee that systems are safe. First, knowing the identity of a bundle issuer does not give guarantees related to the quality of its bundles. Secondly, most implementations do not have a proper implementation of the digital signature mechanism: they rely on the JVM built-in verification mechanism, which is not compliant with OSGi specifications \[[PF07](#bib.bibx26)\]. And, lastly, Java permissions can not be considered as a panacea, since they are usually not dynamic, and have a great cost in term of functionality, but also in term of performance.
New methods and new security mechanisms therefore need to be defined to provide hardened OSGi Platforms. We present in this report our contribution to this problem, by addressing several requirements. A method for analyzing the security properties of the OSGi Platform is defined. It is based on a catalog of vulnerabilities, and can therefore be completed when further knowledge relative to OSGi Vulnerabilities is gathered. Based on the analysis of this catalog, OSGi specific vulnerabilities are identified, and a prototype is built to show the security mechanisms that can be used. Recommendations for an evolution of the specification of the OSGi Core platform are proposed to enable the OSGi Community to take advantage of this work.
This research report is organized as follows. Works related to vulnerability characterization and analysis are presented in Section [2](#S2 "2 Characterization of Vulnerabilities in Component-based Systems"). A definition of our Software Vulnerability Pattern is given in Section [3](#S3 "3 The Semi-formal Software Vulnerability Pattern"): it characterized the properties of an OSGi system that need to be listed so as to support vulnerability analysis. The analysis of the vulnerability catalog is then provided, and recommendation for building a hardened OSGi Platform is given in Section [4](#S4 "4 Requirements for secure OSGi Systems").
Complementary informations are to be found in the Appendices. In particular, a presentation of the OSGi Platform is given in Appendix [A](#A1 "Appendix A The OSGi platform"); the formal expression of the Vulnerability Pattern we defined is given in Appendix [C](#A3 "Appendix C Formal Expression of the Vulnerability Pattern"); the vulnerability catalog in given in its integrality in Appendix [D](#A4 "Appendix D Vulnerability Catalog"); and the specific implementations of attacks based on the identified vulnerabilities are given in Appendix [E](#A5 "Appendix E Attack Implementations").
## 2 Characterization of Vulnerabilities in Component-based Systems
The classification of the security - and vulnerability - properties of systems is necessary to comprehend their weaknesses and to make them more robust. We present here the effort that have been done to establish a precise knowledge of what vulnerabilities are, how to analyse them, and how to take advantage from them to improve the computing systems.
First, the terms that are used to characterize vulnerabilities are defined. Next, the disclosure mechanisms for software flaws are presented. And Vulnerability Patterns that support vulnerability analysis are given.
### 2.1 Definitions
The classification of security properties is based on the distinction between attack, vulnerability, and flaw. The related malicious actions can be prevented by the use of security protections, or countermeasures. The definitions of these terms follow.
Security:
the concurrent existence of a) availability for authorized users only, b) confidentiality, and c) integrity \[[AJB00](#bib.bibx2)\].
Attacks:
actions that attempt to defeat the expected security status of a system.
Software vulnerability:
an instance of an error in the specification, development, or configuration of software such that its execution can violate the security policy \[[Krs98](#bib.bibx16)\].
Software Flaw:
a flaw is a characteristic of a software system that builds, when put together with other flaws, a vulnerability. The more generic term of WIFF (Weaknesses, Idiosyncrasies, Faults, Flaws) is also used \[[MCJ05](#bib.bibx20)\].
Security Protection,
or Mitigations, or Countermeasures or Avoidance strategies: mechanisms and techniques that control the access of executing programs to stored information \[[SS73](#bib.bibx29)\] or to other programs.
Now that the necessary terms are defined, disclosure mechanisms are presented.
### 2.2 From Databases to Top-Vulnerability Lists
Vulnerability disclosure aims at providing users and designers informations that enable them to track the security status of their systems. Two main approaches exist: first, vulnerabilities of widespread applications are published in Reference Vulnerability Information (RVI) Databases so as to centralize this information; secondly, these vulnerabilities are classified according to Top-Vulnerability Lists, that support a comprehensive views of potential weaknesses.
##### Reference Vulnerability Information (RVI) Databases
Catalogs, Lists and Taxonomies are the favorite vector for expressing the vulnerabilities that are identified in software systems. The approach varies according to the target of the vulnerability identification work. Extensive databases are meant to maintain up to date references on known software vulnerabilities, so as to force the system vendor to patch the error before hackers can exploit it. Taxonomies are particularly used in research works, which foster to improve the knowledge relative to the vulnerabilities. Their goal is to develop tools based on this taxonomies. The drawback of these systematic approaches - catalog and taxonomies - is that they are not easy to remember, and are thus of limited usefulness for developers or code auditor. Several Top Vulnerability Lists have been proposed to solve this problem, and provide the software professionals with convenient practical data.
The main existing references are vulnerability databases. They are also known under the denomination of Refined Vulnerability Information (RVI) sources. Two main types of RVI exists: the vulnerability mailing lists, and the vulnerability databases.
The main mailing lists are the following:
- •
Bugtraq, 1993 onwards (see http://msgs.securepoint.com/bugtraq/),
- •
Vulnwatch, 2002 onwards (see http://www.vulnwatch.org/),
- •
Full Disclosure, 2002 onwards (see among others http://seclists.org/).
The reference vulnerability databases are the following. They are meant to publish and maintain reference lists of identified vulnerabilities.
- •
the CERT (Computer Emergency Response Team) Database. It is based on the Common Language for Security Incidents \[[HL98](#bib.bibx14)\]
<sup>2</sup>
<sup>2</sup>2http://www.cert.org/
.
- •
the CVE (Common Vulnerabilities and Exposures) Database
<sup>3</sup>
<sup>3</sup>3http://cve.mitre.org/
.
- •
the CWE (Common Weaknesses Enumeration) Database. It is bounded with the CWE, and aims at tracking weaknesses and flaws that have not yet turned out to be exploitable for attackers
<sup>4</sup>
<sup>4</sup>4http://cwe.mitre.org/index.html
.
- •
the CIAC (Computer Incident Advisory Capability) Database
<sup>5</sup>
<sup>5</sup>5http://www.ciac.org/ciac/index.html
.
- •
the OSVDB, Open Source Vulnerability Database
<sup>6</sup>
<sup>6</sup>6http://osvdb.org/
. It is centered at Open Source Products.
Complementary Refined Vulnerability Informations Sources are the following organizations: SecuriTeam
<sup>7</sup>
<sup>7</sup>7http://www.securiteam.com/
, Packet Storm Security
<sup>8</sup>
<sup>8</sup>8http://packetstormsecurity.nl/
, the French Security Incident Response Team
<sup>9</sup>
<sup>9</sup>9http://www.frsirt.com/
, ISS X-Force
<sup>10</sup>
<sup>10</sup>10http://xforce.iss.net/xforce/alerts
, Secunia, and SecurityFocus.
The limitations of the RVIs is that they follow no stable policy, which makes comparison between sources and between the item of a given sources difficult \[[Chr06](#bib.bibx10)\].
##### Top-Vulnerability Lists
Since catalogs are not that easy to remember, and therefore to put into practice, several ‘Top N’ lists have been defined. The motivation for such lists is the recurrent drawbacks of other approaches: vulnerability catalogs do not provide a useful overview of the identified vulnerabilities \[[Chr06](#bib.bibx10)\].
Therefore, an alternative approach has been developed: to publish lists of prevalent attack categories. Their goal is to be used as reminder for developer or security analysts \[[McG06](#bib.bibx19)\], and to serve as reference for software product characterization, through integration into security-based code assessment tools \[[MCJ05](#bib.bibx20)\]. The most important of these vulnerability lists are presented.
One classification of computer security Intrusions is given by Lindqvist \[[LJ97](#bib.bibx18)\] (see appendix [B.1](#A2.SS1 "B.1 The Lindqvist Classification ‣ Appendix B Vulnerabilities List")). It contains external and hardware misuse, and several software misuse cases: bypassing intended control, active and passive misuse of resources, preparation for other misuse cases.
The Plover classification
<sup>11</sup>
<sup>11</sup>11http://cve.mitre.org/docs/plover/
is an example of rationalization of Vulnerability catalogs to support analysis. It is based on the MITRE CVE Database, and contains 300 specific entries that reflect 1400 vulnerabilities identified in the CVE database. Its goal is to suppress redundancy from the original database, so as to enable scientific analysis, *e.g.* using statistical approaches \[[Chr05](#bib.bibx9)\].
The Nineteen Deadly Sins of software systems are defined by Michael Howard, from Microsoft \[[HLV05](#bib.bibx15)\] (see appendix [B.3](#A2.SS3 "B.3 Nineteen Dealy Sins ‣ Appendix B Vulnerabilities List")). They describe the most common vulnerabilities that are to be found in enterprise information systems. They concern Web based systems, as well as the architecture of the information systems and the technologies involved.
The Open Web Application Security Project (OWASP) maintains a TOP 10 of Web Applications vulnerabilities
<sup>12</sup>
<sup>12</sup>12http://www.owasp.org/index.php/OWASP_Top_Ten_Project
(see the appendix [B.4](#A2.SS4 "B.4 OWASP Top Ten ‣ Appendix B Vulnerabilities List")). It concerns input validation, data storage, as well as configuration and error management. Another consortium for Web Application security enforcement, the WASC (Web Application Security Consortium), provides its own threat classification
<sup>13</sup>
<sup>13</sup>13http://www.webappsec.org/projects/threat/
.)
A convenient vulnerability list is provided by Gary McGraw, through the Seven Kingdoms of software vulnerabilities \[[McG06](#bib.bibx19)\] \[[TCM05](#bib.bibx31)\]. The number 7 is chosen to be easily remembered, and each entry is completed with Phyla *i.e.* precise example of the broader categories that are defined by the Kingdoms. The kingdoms are the following: Input Validation and representation, API abuse, Security Features, Time and state, error handling, code quality, encapsulation + environment (see the appendix [B.5](#A2.SS5 "B.5 Seven Kingdoms ‣ Appendix B Vulnerabilities List")). This classification is targeted at enterprise information systems.
The publication of newly discovered vulnerabilities and of Top-Lists helps the practitioner stay informed of the actual security risks of the system they use, but they provide little support for systematic analysis. Vulnerability Patterns must be defined to formalize vulnerability informations.
### 2.3 Vulnerability Patterns
The descriptive spirit of Design Pattern \[[Ale77](#bib.bibx3)\], \[[GHJV94](#bib.bibx12)\], \[[MM97](#bib.bibx22)\], is well suited for application in the security fields, where the question of organization and exploitation of the knowledge is central to the protection of systems - and not straightforward, if one judges from the various approaches that are used. Two types of patterns are defined in the security domain: Attack Patterns, and Vulnerability Patterns.
Attack Patterns represent potential attacks against a system. They model the preconditions, process and postconditions of the attack. They can be combined with attack trees, so as to automate the identification of attacks that are actually build from simpler atomic attacks \[[MEL01](#bib.bibx21)\]. An extensive presentation of the application of attack pattern is given in the book by Markus Schumacher \[[Sch03](#bib.bibx27)\]. The use of Attack Patterns together with with software architecture description to identify vulnerabilities is described by Gegick \[[GW05](#bib.bibx13)\]. The limitation of this approach is that the attacks must be modelized, but the system must also be, which makes this approach impractical, and often not realistic based on the actual knowledge that is available on systems.
The Vulnerability Patterns are used in the catalog of vulnerabilities. They often contain a limited number of information that are meant to identify the vulnerability, but also to not make it easily reproduceable without a reasonable amount of effort, to prevent lazy hackers to exploit the vulnerability databases as a source of ready-to-exploit attack references.
We list here the most wide-spread Vulnerability Patterns, along with the attribute they contain:
- •
Rocky Heckman pattern
<sup>14</sup>
<sup>14</sup>14http://www.rockyh.net/
: Name, type, subtype, AKA, description, more information;
- •
CERT (Computer Emergency Response Team) pattern: name, date, source, systems affected, overview, description, qualitative impact, solution, references;
- •
CVE
<sup>15</sup>
<sup>15</sup>15http://cve.mitre.org/
(Common Vulnerability and Exposures) pattern: name, description, status, reference(s);
- •
CIAC
<sup>16</sup>
<sup>16</sup>16http://www.ciac.org/ciac/index.html
(US Department of Energy) pattern: identifier, name, problem description, platform, damage, solution, vulnerability assessment, references.
These Vulnerability Patterns are quite simple ones. They have an informative goal, but do not intend as other patterns do at supporting the reproduction of the vulnerability with a minimum of effort. This approach makes sense relative to their use context - making users and administrators aware of the existence of the flaws - but are not sufficient to support detailed analysis of the related vulnerabilities.
So as to support the automation of the security process, and to make vulnerability analysis possible, it is necessary to put constraints on the Vulnerability Patterns. This is performed through the definition of taxonomies, which provide a fine grain description of the properties of each vulnerability.
Each taxonomy should verify the properties of a valid taxonomy, as defined by \[[Krs98](#bib.bibx16)\] and \[[HL98](#bib.bibx14)\]. These properties are the following: objectivity, determinism, repeatability, specificity (disjunction), observability.
The seminal work on vulnerability taxonomy has been performed by Abbott \[[ACD<sup>+</sup>75](#bib.bibx1)\] and Bisbey \[[BH78](#bib.bibx7)\]. The flaws are classified by type of error (such as incomplete Parameter validation). This approach turns out not to support deterministic decisions, since one flaw can often be classified in several categories according to the context. To solve this problem, Landwehr \[[LBMC94](#bib.bibx17)\] defines three fundamental types of taxonomies for vulnerabilities: classification by genesis of the vulnerability, by time of introduction, and by location (or source).
Moreover, vulnerabilities should be considered according to specific constraints or assumptions, since there existence most of the time depends on the properties of the environment \[[Krs98](#bib.bibx16)\]. This assumptions make it necessary to rely on a well defined system model. For generic computing systems, such a model is proposed by the Process/Object Model \[[BAT06](#bib.bibx4)\]. This requirement makes it impossible for generic purpose databases to rely on specific taxonomies. For instance, the Common Vulnerability Enumeration \[[BCHM99](#bib.bibx5)\] project has given up the use of taxonomies. An explicit system model must thus be available to support vulnerability taxonomies, and therefore the possibility of security automation or analysis.
Extensive discussions of vulnerability taxonomies can be found in \[[Krs98](#bib.bibx16)\], \[[SH05](#bib.bibx28)\], \[[WKP05](#bib.bibx32)\]. The CWE (Common Weaknesses Enumeration) Project maintains a web page with additional references, and a graphical representation of each taxonomy
<sup>17</sup>
<sup>17</sup>17http://cwe.mitre.org/about/sources.html
.
In this section, fundamental concepts of vulnerability analysis have been introduced: definitions have been given to provide a firm basis to work on, and the existing works in the domain of vulnerability analysis have been presented. This work concerns Vulnerability properties, which are often presented under the form of a taxonomy, and Vulnerability Patterns, which gather the information concerning several properties in a formalized way.
Existing Properties and Patterns are not sufficient to describe the vulnerabilities of an OSGi Platform, for several reason: first, they do not take explicitly into account the presence of a virtual machine; secondly, they are usually targeted at monolithic systems, whereas OSGi provides a high degree of modularity through the bundles and the dependency resolution. We therefore first need to define the properties of interest for an OSGi-based System, as well as a suitable Pattern, before the actual vulnerabilities of the platform can be analyzed.
## 3 The Semi-formal Software Vulnerability Pattern
The goal of this study is to identify and to characterize the vulnerabilities of the OSGi platform, which is introduced in the Appendix [A](#A1 "Appendix A The OSGi platform"). This characterization is to be done with a set of specific properties, and organized in a semi-formal Vulnerability Pattern. Existing references are not sufficient to describe the vulnerabilities of the OSGi Platform: neither virtualization nor componentization, that are provided in the context of OSGi, are taken into account. Moreover, we want our Vulnerability Pattern to provide us with enough information to patch them or build suitable security mechanism, which is not the case in the literature.
The properties of interested that are tracked are taken from existing software security taxonomies. We add a new entry, the ‘Consequence Description’, that aims at evaluating the seriousness of the vulnerability. The Pattern is made up of four parts:
- •
a Reference, for rapid consultation,
- •
a Description part, for additional and potentially more verbose information,
- •
an Implementation part, to identify the test conditions of the vulnerability,
- •
a Protection part, because the objective of identifying the vulnerability is to be able to patch them.
Our experimental process is the following. First, known flaws that can affect Java code \[[Blo01](#bib.bibx8), [BG05](#bib.bibx6)\] have been identified, and their impact on an OSGi Platform has been tested. Secondly, potentially dangerous mechanisms, such as native code execution, have been selected from related projects. The third source of information in our quest for vulnerabilities of OSGi bundles is the specifications of the elements that make up an OSGi platform: the Java Virtual Machine, and the OSGi platform itself. Several Java API let the code access to the Virtual Machine itself (*e.g.* the System or Runtime API), or are known to cause the execution hang (Threads). The behavior of the OSGi platform in the presence of malformed or malicious bundles is not specified. We therefore review the various entities that make up this execution environment: the format of the bundle meta-data (Manifest File), the registration of services, the bundle management and fragment functionalities. For each potential vulnerability, we implemented a malicious bundle. This makes possible to validate the hypothesis, and to identify the conditions for each attack. When protections against these attacks exist, they are validated through experiment. The attack bundles are tested in the four main Open Source implementations of OSGi, Felix, Knopflerfish, and Equinox, and Concierge.
We focus on the behavior of the core of the considered execution environment, which comprises the JVM and the OSGi platform. We therefore do not consider the management tools for Java systems, such as JMX, or JVM TI. JMX enables to manage a JVM though code executed inside it. JVM TI is a C library that makes full control over the JVM possible through a third party program, which can then access the available threads, provides an extensive debugging of the platform, and control the JVM state. Secondly, the OSGi bundles communicate through services they publish inside the framework. According to the type of data they handle, these services can be subject to specific vulnerabilities. A list of Service-Level vulnerabilities is given by the Findbugs reference list (http://findbugs.sourceforge.net/bugDescriptions.html, ‘Malicious Code Vulnerability’ category). Lastly, the OSGi specification defines a bunch of standard services (HTTP, device, service wiring, UPnP services, etc.). We do not consider these services either.
A Vulnerability Pattern is defined to normalize the information gathered relative to each vulnerability (see section [3.1](#S3.SS1 "3.1 The Structure of the Semi-formal Vulnerability Pattern ‣ 3 The Semi-formal Software Vulnerability Pattern")). The taxonomies for each properties of interest are given and explained in section [3.2](#S3.SS2 "3.2 Vulnerability Taxonomies for OSGi-based Systems ‣ 3 The Semi-formal Software Vulnerability Pattern"). An example is given in section [3.3](#S3.SS3 "3.3 A Vulnerability Example: ‘Management Utility Freezing - Infinite Loop’ ‣ 3 The Semi-formal Software Vulnerability Pattern") to highlight the information provided by the Vulnerability Pattern.
### 3.1 The Structure of the Semi-formal Vulnerability Pattern
The characteristics of interest to describe the vulnerabilities of a software system need to be gathered in a coherent set that contains all the informations that are useful to understand and prevent these vulnerabilities. We therefore define a ‘Semi-formal Vulnerability Pattern’ that is similar to the ‘Attack Patterns’ \[[MEL01](#bib.bibx21)\]. On the opposite of this latter, the Vulnerability Pattern is centered around the identified vulnerability, so as to make their correction easy. Existing vulnerability patterns, which are presented in the section [2](#S2 "2 Characterization of Vulnerabilities in Component-based Systems"), can not be reused as-is, since they provide not enough details for our purpose.
The Vulnerability Pattern is compound of following informations. Its formal expression is given in the Apendix [C](#A3 "Appendix C Formal Expression of the Vulnerability Pattern").
<table>
<tbody>
<tr>
<td>Vulnerability Reference • Vulnerability Name: The descriptive name of the vulnerability • Identifier: a unique identifier for each vulnerability. In our catalog, the identifier is built out of the catalog identifier, the abbreviation of the source entity, and the number ID of the vulnerability in the catalog for the related source entity. • Origin: The bibliographic reference of the vulnerability. • Location of Exploit Code: Where the code that performs the attack is located in the malicious Bundle (see Figure 1). • Source: the entity in the execution platform that is the source of the vulnerability, along with the exact flaw or functionality causing it. • Target: the target of the attack that can be performed through the vulnerability, i.e. the victim of the attack (see figure 3). • Consequence Type: the type of consequence of an attack exploiting this vulnerability (see figure 4). • Time of Introduction: the Life Cycle phase where the vulnerability is introduced. Corrective measures can be taken at this time. Security measures can be taken in subsequent phases so as to prevent the exploitation of the vulnerability (see figure 5). • Time of Exploit: the life-cycle phase where the vulnerability can be exploited (see figure 6). This is the last phase where security measures can be undertaken.</td>
</tr>
</tbody>
</table>
<table>
<tbody>
<tr>
<td>Vulnerability Description • Description: a description of the attack • Preconditions: properties if the systems that must be true so as to make the exploitation of the vulnerability possible. • Attack Process: description of the process of exploitation of the vulnerability. • Consequence Description: more information relative to the consequences of an attack using this vulnerability. • See also: other vulnerabilities based on similar attack sources. Vulnerability Implementation • Code Reference: the reference of the implementation code (i.e. the name of the malicious OSGi bundle.) • Concerned OSGi Profile: the OSGi profile(s) where this vulnerability exists. • Date: the date of the creation of the Vulnerability Pattern (for reference) • Test Coverage: the percentage of the known implementations of the vulnerability that have been implemented in a test bundle. The identified implementations for the main attacks are given in the Appendix E. • Tested on: the OSGi Platform Implementations for which this vulnerability have been tested. Protection • Existing mechanisms: available protections to prevent this vulnerability from being exploited. • Life-cycle enforcement point: the life-cycle phase where the protection mechanisms must be enforced. • Potential mechanisms: protections that could be developed so as to prevent this vulnerability from being exploited. • Attack Prevention: the measures that can be taken to prevent an attack based on this vulnerability to be fulfilled, even if it is launched. • Reaction: the correction action that can be taken to recover from a successful attack.</td>
</tr>
</tbody>
</table>
### 3.2 Vulnerability Taxonomies for OSGi-based Systems
Before analyzing each vulnerability, it is necessary to identify the properties of interest that need to be characterized. Moreover, the potential values for each property should be identified, and build a properly defined taxonomy. Following aspects need to be considered: the reference of the attack bundle implementation that takes advantage of the vulnerability, the life-cycle characteristics of the vulnerability, so as to know when the vulnerability is introduced, and when it is exploited, and the existing and potential security mechanisms.
The properties we selected to be included in the vulnerability pattern are thus the following:
- •
the reference of the vulnerability pattern (to identify the code, and the condition of the experiments),
- •
the location of the malicious code,
- •
the source of the flaw(s) in the system (and the specific flaw(s) and/or the dangerous functionality(ies)),
- •
the target of attacks based on the vulnerability,
- •
the consequences of the related attack, and
- •
the time of introduction of the vulnerability,
- •
the time of exploitation of the vulnerability,
- •
the exiting protections against this attack,
- •
the potential protections against this attack.
The goal of these properties is to make the information explanatory, predictive \[[Krs98](#bib.bibx16)\], but also useful \[[WKP05](#bib.bibx32)\]. Explanatory, because the vulnerability should be intuitively understood by the persons who consult the vulnerability catalog we propose, even with little previous knowledge of the OSGi Platform. Predictive, because the potential values of each characteristic should cover the whole field of possible options, or to explain why some are not concerned. Useful, because the objective of a vulnerability catalog is to highlight the security requirements of the platform under study. The conclusions of the analysis is presented in the section [4](#S4 "4 Requirements for secure OSGi Systems").
For each property, a taxonomy is defined, that contains the values this property can take. Two approaches are used to define this taxonomy. It is either defined a priori, *i.e.* before the catalog is completed, or a posteriori, with data that are identified during the experiments.
We now present the taxonomy for each of the properties of interest.
The first two properties of interest are the Location of the malicious code and the source of the vulnerability. The location concerns the place in the attack bundle where the attack ‘payload’ is located. The source indicates which entity in the execution environment is responsible for the vulnerability, *i.e.* for the system behavior that opens the door to the attack.
Figure [1](#S3.F1 "Figure 1 ‣ 3.2 Vulnerability Taxonomies for OSGi-based Systems ‣ 3 The Semi-formal Software Vulnerability Pattern") shows the potential locations of malicious code inside a malevolent bundle. The malicious code can be located in the archive structure (such as archive oversize, or a decompression bomb), in the manifest (such as duplicate imports, which make the installation abord), or in the bundle Activator (is this latter is hanging). It can also be located in the applicative code of the bundle, being native code, Java code, the Java APIs, or the OSGi API. The malicious code can also be located in fragment, which are specific bundle types.
[Refer to caption](/html/0706.3812/assets/x1.png)
Figure 1: Potential Locations of malicious Code in a Bundle
The actual sources of vulnerabilities match the different Layers that are defined by the OSGi Specification, along with the Bundle Repository client which enables installation from remote bundles, and various code properties such as Services, the JVM APIs, or the algorithmic properties of the programs. They are shown in Figure [2](#S3.F2 "Figure 2 ‣ 3.2 Vulnerability Taxonomies for OSGi-based Systems ‣ 3 The Semi-formal Software Vulnerability Pattern").
[Refer to caption](/html/0706.3812/assets/x2.png)
Figure 2: Vulnerability Sources in an OSGi-based System
Figure [3](#S3.F3 "Figure 3 ‣ 3.2 Vulnerability Taxonomies for OSGi-based Systems ‣ 3 The Semi-formal Software Vulnerability Pattern") shows the potential targets of attacks against an OSGi Platform. These targets can be either the whole platform, or specific OSGi Elements. Attacks against the whole Platform can for instance result in complete unavailability if this latter. The victim OSGi Elements can be the Platform Management Utility, which makes it possible for the user to control the life-cycle of bundles (the activator can hang), the bundle itself (which can be started or stopped), Services (which can suffer from cycles) or packages (for instance, static data that is by default not accessible could be modified through Bundle Fragments).
[Refer to caption](/html/0706.3812/assets/x3.png)
Figure 3: Potential Targets of Attacks against an OSGi Platform
Figure [4](#S3.F4 "Figure 4 ‣ 3.2 Vulnerability Taxonomies for OSGi-based Systems ‣ 3 The Semi-formal Software Vulnerability Pattern") shows the potential consequences of an attack against an OSGi Platform. Three types of consequences are identified: Unavailability, Performance Breakdown, and Undue Access. Unavailability can be cause by stopping the platform; Performance Breakdown can be the result of an infinite loop; and Undue Access can be performed through Fragments or through Reflection over Services.
[Refer to caption](/html/0706.3812/assets/x4.png)
Figure 4: Consequences of the Vulnerabilities of the OSGi Platform
Figure [5](#S3.F5 "Figure 5 ‣ 3.2 Vulnerability Taxonomies for OSGi-based Systems ‣ 3 The Semi-formal Software Vulnerability Pattern") shows the actual introduction time of the vulnerabilities. The introduction time can be as early as the design and implementation of the platform (when the flaw originates in the platform), or be the development time, the generation of the Meta-data of the bundles, the digital signature of the bundle, the installation, or even the publication and resolution of services.
[Refer to caption](/html/0706.3812/assets/x5.png)
Figure 5: Introduction Time for the identified flaws
The taxonomy for Time of Exploit of the vulnerability is represented in figure [6](#S3.F6 "Figure 6 ‣ 3.2 Vulnerability Taxonomies for OSGi-based Systems ‣ 3 The Semi-formal Software Vulnerability Pattern"). This time of exploit concerns necessarily the Life-Cycle steps inside the execution platform. They can therefore be: the download, installation, Bundle start (if the vulnerability is present in the bundle activator) or execution time (either through service call, or through use of exported packages).
[Refer to caption](/html/0706.3812/assets/x6.png)
Figure 6: Exploit Time for the identified Flaws
The existing protections against attacks based on the identified vulnerability are the following: only runtime execution permissions, either at the JVM level or at the OSGi Platform level, are currently available to protect an OSGi Platform from hackers. We propose our own implementation of the OSGi Bundle Digital Signature validation process, which is part of the OSGi Security Layer.
- •
Java Permissions,
- •
OSGi Permissions (in particular AdminPermission),
- •
SFelix implementation of the Bundle Digital Signature Validation.
The properties of interest that characterize a vulnerability have been presented. Next paragraph gives the full Vulnerability Pattern that is based on these properties, and adapted for better comprehension.
### 3.3 A Vulnerability Example: ‘Management Utility Freezing - Infinite Loop’
So as to highlight the role of the defined Vulnerability Pattern, we now present an example of vulnerability; the ‘Management Utility Freezing - Infinite Loop’ vulnerability. The whole vulnerability catalog is given in the Appendix [D](#A4 "Appendix D Vulnerability Catalog"). This vulnerability consists in the presence of an infinite Loop in the activator of a given Bundle, which causes the platform management tool (often an OSGi shell) to freeze. The presence of infinite loops as a vulnerability is given by Blooch is ‘Java Puzzlers - Traps, Pitfalls and Corner Cases’, puzzlers 26 to 33 \[[BG05](#bib.bibx6)\]. The matching Pattern is first given, and then explained.
##### The Vulnerability Pattern
##### Vulnerability Reference
- •
Vulnerability Name: Management Utility Freezing - Infinite Loop
- •
Extends: Infinite Loop in Method Call
- •
Identifier: Mb.osgi.4
- •
Origin: Ares research project ‘malicious-bundle’
- •
Location of Exploit Code: Bundle Activator
- •
Source: OSGi Platform - Life-Cycle Layer (No safe Bundle Start)
- •
Target: OSGi Element - Platform Management Utility
- •
Consequence Type: Performance Breakdown; Unavailability
- •
Introduction Time: Development
- •
Exploit Time: Bundle Start
##### Vulnerability Description
- •
Description: An infinite loop is executed in the Bundle Activator
- •
Preconditions: -
- •
Attack Process: An infinite loop is executed in the Bundle Activator
- •
Consequence Description: Block the OSGi Management entity (the felix, equinox or knopflerfish shell; when launched in the KF graphical interface, the shell remain available but the GUI is frozen). Because of the infinite loop, most CPU resource is consumed
- •
See Also: CPU Load Injection, Infinite Loop in Method Call, Stand Alone Infinite Loop, Hanging Thread
##### Protection
- •
Existing Mechanisms: -
- •
Enforcement Point: -
- •
Potential Mechanisms: Code static Analysis ; Resource Control and Isolation - CPU ; OSGi Platform Modification - Bundle Startup Process (launch the bundle activator in a separate thread to prevent startup hanging)
- •
Attack Prevention: -
- •
Reaction: -
##### Vulnerability Implementation
- •
Code Reference: Fr.inria.ares.infiniteloopinmethodcall-0.1.jar
- •
OSGi Profile: J2SE-1.5
- •
Date: 2006-08-24
- •
Test Coverage: 10%
- •
Known Vulnerable Platforms: Felix; Equinox; Knopflerfish; Concierge
- •
Known Robust Platforms: SFelix
##### Details of the Vulnerability Pattern
The ‘Management Utility Freezing - Infinite Loop’ is referenced under the identifier ‘mb.osgi.4’, which means ‘malicious bundle catalog - originates in the OSGi Platform itself - number 4. This vulnerability is an extension of the ‘Infinite Loop in Method Call’ one. It has been identified in the frame of the research project ‘Malicious Bundles’ of the INRIA Ares Team..
The location of the malicious code that performs the attack is the Bundle Activator. Its source is the Life-Cycle Layer of the OSGi Platform, which is not robust against such a vulnerability. Its target is the Platform Management Utility, which can be either the OSGi shell or a graphical interface such as the Knopflerfish GUI. This vulnerability has a two-fold consequence: the method does not return, so that the caller also freezes; and the infinite loop consumes most of the available CPU, which causes the existing services to suffer from a serious performance breakdown.
This vulnerability is introduced during development, and exploited at bundle start time.
Related Vulnerability Patterns are ‘Management Utility Freezing - Hanging Thread’, that also targets the Management Utility, ‘Infinite Loop in Method Call’, ‘CPU Load Injection’, ‘Stand-alone Infinite Loop’ that have the same consequence of performance breakdown, and the ‘Hanging Thread’, that also freezes the calling thread.
No specific protection currently exists. Two potential solutions have been identified. The first consists in launching every Bundle Activator in a new Thread, so as not to block the caller if the activator hangs. The second solution would enable to prevent invalid algorithms to be executed: static code analysis techniques such as Proof Carrying Code or similar approaches \[[Nec97](#bib.bibx23)\] can provide formal proves of code wellformedness.
Its reference implementation is available in the OSGi bundle named ‘fr.inria.ares.infinite- loopinmethodcall-0.1.jar’, referenced the 2006-08-24. The test coverage is 10 %, since ten types of infinite loops have been identified (see the appendix [E](#A5 "Appendix E Attack Implementations")), and only one has been implemented. The test bundle have been tested on the following implementations of the OSGi platform: Felix, Equinox, Knopflerfish, and Concierge. The only robust Platform is our SFelix Platform, which is a prototype meant to enhance to current Felix implementation.
This example highlights the informations that can be found in each Vulnerability Patterns. The information related to the other vulnerabilities is given under the form of patterns, to provide a quick overview of the characteristics, and to make analysis possible. The catalog of the vulnerability patterns is presented in the section [D](#A4 "Appendix D Vulnerability Catalog"). The section [4](#S4 "4 Requirements for secure OSGi Systems") presents the analysis of this catalog, and the security requirements can can be deduced from it.
## 4 Requirements for secure OSGi Systems
The analysis of the Vulnerability Patterns we presented in the catalog provides guidelines for programming secure OSGi-based systems. The objective here is two fold. First, the weaknesses of the OSGi Platform are to be identified, so as to provide a framework for the evolution of its specification. Secondly, these weaknesses are to be made available in a developer-compliant way, so that programmers can refer to them to verify that their system do not open the way to known attacks: the *Seven Deadly Sins* of the OSGi R4 Platform are therefore defined.
These guidelines are - off course - based on the catalog at the moment of its publication, and can therefore evolve in the future, when new vulnerabilities will be discovered, or when new part of OSGi systems will be considered. Actually, the management tools such as JVMTI, and the OSGi standard services are not considered, and can be the source for new vulnerabilities.
This section is organized as follows. Subsection [4.1](#S4.SS1 "4.1 Catalog Analysis ‣ 4 Requirements for secure OSGi Systems") presents the analysis of the catalog through statistics related to the significant properties of the vulnerabilities. Subsection [4.2](#S4.SS2 "4.2 Requirements for a Hardened OSGi Platform ‣ 4 Requirements for secure OSGi Systems") presents the Security Requirements for a hardened OSGi Platform. Lastly, Subsection [4.3](#S4.SS3 "4.3 Recommendations for a Hardened Execution Environment ‣ 4 Requirements for secure OSGi Systems") gives a series of recommendation for the OSGi Specification, in order to make the platform more robust.
### 4.1 Catalog Analysis
The analysis of the identified Vulnerability Patterns provides quantitative data relative to these vulnerabilities. This subsection provides a summary of the significant properties that characterize a vulnerability in an OSGi Execution Environment. We use the term OSGi Execution Environment to describe an execution platform running OSGi on top of a JVM. This denomination highlights the fact that not all vulnerabilities are bound to the OSGi specification itself, but can also originate in other parts of the system.
First, quantitative results relative to the vulnerability sources, functions, flaws, as well as the identified attack targets are given. Next, a summary of the vulnerabilities for each tested OSGi Platform implementation is presented.
Figure [7](#S4.F7 "Figure 7 ‣ 4.1 Catalog Analysis ‣ 4 Requirements for secure OSGi Systems") shows the source entity of the identified vulnerabilities. The most important source is the Java API, which causes the bigger part of the vulnerabilities. Next, the Application Code properties, the OSGi Life-Cycle Layer and the OSGi Module Layer also generate an important number of vulnerabilities. Next comes the Java Runtime API, which is particularly sensitive, and the OSGi Service Layer. The Operating System and the Bundle Repository Client must also be considered as potential source of vulnerabilities, even though their impact is more marginal.
[Refer to caption](/html/0706.3812/assets/x7.png)
Figure 7: Entities that are Source of the vulnerabilities
Figure [8](#S4.F8 "Figure 8 ‣ 4.1 Catalog Analysis ‣ 4 Requirements for secure OSGi Systems") shows the cardinality of the identified dangerous functions. First come the OSGi Bundle Facility, and the Java APIs Reflection, ClassLoader, and Thread. Next, the bundle management, the Java File API and the opportunity of executing native code open the way to abuses. Several other functions prove to be dangerous: the Runtime.halt() and the System.exit() methods, the lack of control on method parameters, and the kill utility at the OS level which can be used to shut the platform down.
[Refer to caption](/html/0706.3812/assets/x8.png)
Figure 8: Functions that prove to be dangerous in the context of an OSGi Platform
Figure [9](#S4.F9 "Figure 9 ‣ 4.1 Catalog Analysis ‣ 4 Requirements for secure OSGi Systems") shows the cardinality of the flaws of a Java-based execution environment with the OSGi Platform. The most important flaw is the lack of algorithm safety in the Java language. Next come several properties of the OSGi platforms, such as the lack of safe-default bundle meta-data handling during the dependency resolution phase, the lack of control on the service registration process, and the lack of robustness of the bundle start mechanism, which heavily relies on the validity of the bundle activators. Several other punctual flaws have been identified: data of uninstalled bundles is often kept on the disk space, being not accessible, no dependency control is performed at the service level, the process of Digital Signature validation is sometimes not compliant with the OSGi R4 Specifications, and the bundle archive is never checked for size or validity, which provides no protection against decompression bombs or large files, in particular in resource-constraint environments.
[Refer to caption](/html/0706.3812/assets/x9.png)
Figure 9: Flaws in an OSGi Execution Environment
Figure [10](#S4.F10 "Figure 10 ‣ 4.1 Catalog Analysis ‣ 4 Requirements for secure OSGi Systems") shows the target of attacks against an OSGi execution environment. The entity that is the first target of the identified attacks is the platform. This means that most of the identified attack can easily prevent all services on the platform the be executed in a satisfactorily manner. OSGi specific elements, such as packages, Bundles or services are other frequent targets. Lastly, the Platform Management Utility can also be targeted, which would not prevent the platform to provide existing services, but would prevent any evolution of these services - as well as the removal of the malicious bundle.
[Refer to caption](/html/0706.3812/assets/x10.png)
Figure 10: Targets of Attacks against an OSGi Execution environment
The table [1](#S4.T1 "Table 1 ‣ 4.1 Catalog Analysis ‣ 4 Requirements for secure OSGi Systems") shows a summary of the properties of the OSGi Platform implementations under study. The considered platforms are the main Open Source Projects: Felix
<sup>18</sup>
<sup>18</sup>18http://cwiki.apache.org/FELIX/index.html
, Knopflerfish
<sup>19</sup>
<sup>19</sup>19http://www.knopflerfish.org/
, Eclipse Equinox
<sup>20</sup>
<sup>20</sup>20http://www.eclipse.org/equinox/
, and Concierge
<sup>21</sup>
<sup>21</sup>21http://concierge.sourceforge.net/
. For comparison, we also provide the data related to SFelix, which is the Hardened OSGi Platform we develop. It is based on the Felix Platform 0.8.0.
Most Open Source OSGi Platforms are very fragile regarding the set of vulnerability we identified. Equinox proves to have slightly better results than the other ones. SFelix currently does not intend to provide protections against all the identified vulnerabilities, but only to provide a first enhancement of current implementations.
<table>
<tbody>
<tr>
<td>Vulnerability</td>
<td>Felix</td>
<td>Knopflerfish</td>
<td>Equinox</td>
<td>Concierge</td>
<td>SFelix</td>
<td>Any with Java Permissions</td>
</tr>
<tr>
<td>Exponential Object Creation</td>
<td>V</td>
<td>V</td>
<td>V</td>
<td>V</td>
<td>V</td>
<td>-</td>
</tr>
<tr>
<td>Excessive Size of Manifest File</td>
<td>V</td>
<td>V</td>
<td>R</td>
<td>V</td>
<td>R</td>
<td>-</td>
</tr>
<tr>
<td>Access Protected Package through split Packages</td>
<td>V</td>
<td>V</td>
<td>V</td>
<td>-</td>
<td>V</td>
<td>R</td>
</tr>
<tr>
<td>Freezing Numerous Service Registration</td>
<td>-</td>
<td>-</td>
<td>-</td>
<td>V</td>
<td>-</td>
<td>-</td>
</tr>
<tr>
<td>Big File Creator</td>
<td>V</td>
<td>V</td>
<td>V</td>
<td>V</td>
<td>V</td>
<td>R</td>
</tr>
<tr>
<td>Management Utility Freezing - Thread Hanging</td>
<td>V</td>
<td>V</td>
<td>V</td>
<td>V</td>
<td>R</td>
<td>-</td>
</tr>
<tr>
<td>Erroneous values of Manifest attributes</td>
<td>V</td>
<td>V</td>
<td>V</td>
<td>V</td>
<td>V</td>
<td>-</td>
</tr>
<tr>
<td>Invalid Digital Signature Validation</td>
<td>V</td>
<td>-</td>
<td>-</td>
<td>-</td>
<td>R</td>
<td>-</td>
</tr>
<tr>
<td>Cycle Between Services</td>
<td>V</td>
<td>V</td>
<td>V</td>
<td>V</td>
<td>V</td>
<td>-</td>
</tr>
<tr>
<td>Hanging Thread</td>
<td>V</td>
<td>V</td>
<td>V</td>
<td>V</td>
<td>V</td>
<td>-</td>
</tr>
<tr>
<td>Management Utility Freezing - Infinite Loop</td>
<td>V</td>
<td>V</td>
<td>V</td>
<td>V</td>
<td>R</td>
<td>-</td>
</tr>
<tr>
<td>Fragment Substitution</td>
<td>V</td>
<td>V</td>
<td>V</td>
<td>-</td>
<td>V</td>
<td>R</td>
</tr>
<tr>
<td>Numerous Service Registration</td>
<td>V</td>
<td>V</td>
<td>V</td>
<td>V</td>
<td>R</td>
<td>-</td>
</tr>
<tr>
<td>Sleeping Bundle</td>
<td>V</td>
<td>V</td>
<td>V</td>
<td>V</td>
<td>V</td>
<td>-</td>
</tr>
<tr>
<td>Code Observer</td>
<td>V</td>
<td>V</td>
<td>V</td>
<td>-</td>
<td>V</td>
<td>R</td>
</tr>
<tr>
<td>Recursive Thread Creation</td>
<td>V</td>
<td>V</td>
<td>V</td>
<td>V</td>
<td>V</td>
<td>-</td>
</tr>
<tr>
<td>Duplicate Package Import</td>
<td>V</td>
<td>V</td>
<td>R</td>
<td>R</td>
<td>R</td>
<td>-</td>
</tr>
<tr>
<td>Memory Load Injection</td>
<td>V</td>
<td>V</td>
<td>V</td>
<td>V</td>
<td>V</td>
<td>-</td>
</tr>
<tr>
<td>Infinite Loop in Method Call</td>
<td>V</td>
<td>V</td>
<td>V</td>
<td>V</td>
<td>V</td>
<td>-</td>
</tr>
<tr>
<td>Runtime.halt</td>
<td>V</td>
<td>V</td>
<td>V</td>
<td>V</td>
<td>V</td>
<td>R</td>
</tr>
<tr>
<td>CPU Load Injection</td>
<td>V</td>
<td>V</td>
<td>V</td>
<td>V</td>
<td>V</td>
<td>R</td>
</tr>
<tr>
<td>System.exit</td>
<td>V</td>
<td>V</td>
<td>V</td>
<td>V</td>
<td>V</td>
<td>R</td>
</tr>
<tr>
<td>Runtime.exec.kill</td>
<td>V</td>
<td>V</td>
<td>V</td>
<td>V</td>
<td>V</td>
<td>R</td>
</tr>
<tr>
<td>Component Data Modifier</td>
<td>V</td>
<td>V</td>
<td>V</td>
<td>V</td>
<td>V</td>
<td>R</td>
</tr>
<tr>
<td>Stand Alone Infinite Loop</td>
<td>V</td>
<td>V</td>
<td>V</td>
<td>V</td>
<td>V</td>
<td>-</td>
</tr>
<tr>
<td>Execute Hidden Classes</td>
<td>V</td>
<td>V</td>
<td>V</td>
<td>-</td>
<td>V</td>
<td>R</td>
</tr>
<tr>
<td>Pirat Bundle Manager</td>
<td>V</td>
<td>V</td>
<td>V</td>
<td>V</td>
<td>V</td>
<td>R</td>
</tr>
<tr>
<td>Big Component Installer</td>
<td>-</td>
<td>-</td>
<td>-</td>
<td>-</td>
<td>R</td>
<td>-</td>
</tr>
<tr>
<td>Launch a Hidden Bundle</td>
<td>V</td>
<td>V</td>
<td>V</td>
<td>V</td>
<td>V</td>
<td>R</td>
</tr>
<tr>
<td>Zombie Data</td>
<td>V</td>
<td>R</td>
<td>R</td>
<td>V</td>
<td>R</td>
<td>-</td>
</tr>
<tr>
<td>Decompression Bomb</td>
<td>-</td>
<td>-</td>
<td>-</td>
<td>-</td>
<td>-</td>
<td>-</td>
</tr>
<tr>
<td>Hidden Method Launcher</td>
<td>V</td>
<td>V</td>
<td>V</td>
<td>V</td>
<td>V</td>
<td>R</td>
</tr>
<tr>
<td colspan="7">V: Platform is Vulnerable; R: Platform is Robust; - : not relevant</td>
</tr>
</tbody>
</table>
Table 1: Vulnerabilities for the main Open Source OSGi Platforms
The result of our analysis have been presented for the properties that characterize vulnerabilities of an OSGi Platform: Location of the malicious payload in Bundles, Vulnerability Source, Flaws and dangerous Functions, as well as the identified Attack Targets. It is now possible to identify the requirement for a Hardened OSGi Platform.
### 4.2 Requirements for a Hardened OSGi Platform
The requirement for a Hardened OSGi Platform can be deduced from the actual and potential security mechanisms. The objective is to highlight the security mechanisms that need to be better exploited (for the existing ones), and the ones that need to be developed (for the potential ones). Priorities can be set according to the type of target and consequences of the attacks: it is in any case worth preventing an attack that makes the whole platform unavailable, but it may be less important to prevent attacks that provoke only the unavailability of the malicious bundle itself.
Figure [11](#S4.F11 "Figure 11 ‣ 4.2 Requirements for a Hardened OSGi Platform ‣ 4 Requirements for secure OSGi Systems") shows the cardinality of the actual protection mechanisms. Most of the vulnerabilities can be prevented by Java Permissions. However, an important number of them currently have no associated protections. The OSGi Admin Permission and the SFelix implementation of the OSGi Security Layer (Digital Signature Validation part) account each for one vulnerability.
[Refer to caption](/html/0706.3812/assets/x11.png)
Figure 11: Actual Protection Mechanisms
The Figure [12](#S4.F12 "Figure 12 ‣ 4.2 Requirements for a Hardened OSGi Platform ‣ 4 Requirements for secure OSGi Systems") shows the potential protections identified to protect an OSGi Platform against the considered attacks. This potential protections are off course only set as hypotheses: as long as no implementation is available, it is not possible to assert that no special case or hard-to-track false positives and negatives dot not occur if the propose technique is used. The most promising approach seems to be static code analysis, that would help track both dangerous calls without heavyweight permissions and unsafe algorithms. The OSGi Platform itself would take benefit of several minor modifications: better handling of ill-formed meta-data, safe startup process for bundles, better control of service publication. Some of these mechanisms have been experienced in the SFelix Platform, and prove to be easy to implement. Also, resource control and isolation mechanisms (CPU, Memory, disk space) would make the support of multi-processes safer.
[Refer to caption](/html/0706.3812/assets/x12.png)
Figure 12: Cardinality for each potential Security Mechanisms for the OSGi platform
The potential protection mechanism represent the element that are worth an important development effort. However, they do not show the relative priority of the security mechanisms. Urgent security mechanisms are the ones that prevent attacks with serious consequences - for instance platform unavailability - to be performed, or the ones that are required to make the use of existing mechanisms efficient and convenient for developers. Consequently, the priority is to be set on following protection mechanisms:
- •
Protection of Attacks targeted at the whole Platform (see Figure [10](#S4.F10 "Figure 10 ‣ 4.1 Catalog Analysis ‣ 4 Requirements for secure OSGi Systems")), and that impair the availability or the performance of all executed bundles simultaneously,
- •
Protection against silent attacks: classical access control mechanisms are required inside the OSGi Platform, to support mutually untrustful bundles,
- •
tools are required to take advantages of existing mechanisms: for instance, Permission are supported, but currently extremely unconvenient to set and manage.
We presented the requirement for developing a Hardened OSGi Platform by identifying the best promising potential security mechanisms as well as the most urgent tools for preventing serious attacks, or taking advantage of existing protections. However, developers require ready-to-use guidelines to take advantage of the knowledge we gathered in these study: in the absence of available tools, they have to take care by themselves that the code they produce is safe from the known vulnerabilities.
### 4.3 Recommendations for a Hardened Execution Environment
##### Hardening the Specifications of the OSGi Platform
Based on the identified vulnerability of the OSGi Platform, we propose following recommendation for an enhanced OSGi Platform. These recommendation do not pretend to solve every identified problems, but intend to make the community aware of the easy changes that can be made to the OSGi Specification so as to prevent avoidable flaws.
These recommendations are validated by the Platform SFelix version 0.2, which is a robust extension to the Felix 0.8.0 implementation of the OSGi Platform.
Following improvement to the OSGi Release 4 Specification should be made:
- •
Bundle Installation Process: a maximum storage size for bundle archives is set. Alternatively, a maximum storage size for all data stored on the local disk is set (*Bundle Archives and files created by the bundles*); *OSGi R4 par. 4.3.3*.
- •
Bundle Uninstallation Process: remove the data on the local bundle filesystem when a bundle is uninstalled (and not when the platform is stopped); *OSGi R4 par. 4.3.8*.
- •
Bundle Signature Validation Process: the digital signature must be checked at installed time. It must not rely on the Java built-in validation mechanism, since this latter is not compliant with the OSGi R4 Specifications \[[PF07](#bib.bibx26)\]; *OSGi R4, Paragraph 2.3*.
- •
Bundle Dependency Resolution Process: do not reject duplicate imports. just ignore them; *OSGi R4 par. 3.5.4*.
- •
Bundle Start Process: launch the Bundle Activator in a separate thread; *OSGi R4 par. 4.3.5*.
- •
OSGi Service Registration: set a Platform Property that explicitly limits the number of registered services (default could be 50); *OSGi R4 par. 5.2.3*.
- •
Bundle Download: when a bundle download facility is available, the total size of the bundles to installed should be checked immediately after the dependency resolution process. The bundles should be installed only if the required storage is available.
To support this modifications of the OSGi R4 specifications, following changes have been applied to the API:
- •
In the Class BundleContext, a method ‘getAvailableStorage()’ is defined,
- •
A property ‘osgi.storage.max’ is defined, that is set in the property configuration file of the OSGi framework.
- •
In the class org.osgi.service.obr.Resource, a method ‘getSize()’ is defined. This method relies on the ‘size’ entry of the bundle meta-data representation (usually a XML file).
In addition to these simple enhancement, more research work is required in order to define proper solution to the identified vulnerabilities. The most important ones are the following:
- •
Static Code Analysis for Java,
- •
Convenient Permission Management for Java and OSGi,
- •
Resource isolation in component systems,
- •
Mandatory Service Management.
Through this study, we identified both technical requirements for enhancement of the OSGi R4 Specifications, and necessary research work that is necessary to protect the OSGi Platform.
##### Hardening the Specifications of the Java Virtual Machine
Some safety requirements have also been identified at the Virtual Machine Level.
The flaws that have been identified in the Sun Java Virtual Machine version 1.6 are the following:
- •
the Java Permission ‘exitVM’ appears not to be effective,
- •
the presence of a manifest with a huge size in a loaded Jar file introduces a dramatic slowdown of the JVM when the archive Manifest is extracted. Our implementation shows that a simple patch can correct this matter of fact.
A flaw has also been identified in the Gnu Classpath, which is an open source implementation of the Java classes. Gnu Classpath is used in conjunction with the JamVM Virtual Machine and targets resource-limited devices:
- •
the presence of a manifest with a huge size in a loaded Jar file introduces a dramatic slowdown of the JVM when the corresponding JarFile Object is created, even though the Manifest stays unused.
Requirements for programming secure OSGi Systems have been identified. First, a hardened version of the OSGi Platform is needed to prevent most of the identified vulnerabilities to be exploited. However, since such a platform will take time to develop and validate, a pragmatic approach is to be taken. First, tools should be developed to ease the management of current security mechanisms such as Java Permissions, which are currently not adapted to dynamic systems. Secondly, developers need to keep on mind what the OSGi vulnerabilities are: this is made possible by the *Seven Deadly Sins* of the OSGi R4 Platform.
## 5 Conclusions
The objective of our study is to improve the dependability level of the OSGi platform, as well as the knowledge that is available relative to the vulnerabilities of the OSGi Platform. This improvement is achieved through four complementary contributions. First, we define a method for analyzing the security status of software systems, based on a specific Software Vulnerability Pattern. Secondly, we provide a vulnerability catalog that identified a set of vulnerabilities, and the key properties for understanding - and preventing - them. Thirdly, we developed a hardened OSGi Platform, SFelix v0.2
<sup>22</sup>
<sup>22</sup>22http://sfelix.gforge.inria.fr/
, that provides proof of concept protection mechanisms. And we issue a set of recommendations for the OSGi Specifications that integrate these protection mechanisms.
Our study is centered on the OSGi Core specification, and does not take into account several mechanisms that are - or can be - often used together with OSGi platforms. In particular, management facilities, such as JVMTI, and with less impact JMX have not been studied. OSGi standard services are neither been considered, and service engineering questions have been neglected. These three elements will require further work, and will likely enrich our vulnerability catalog.
A side-effect achievement of our study is to precisely identify the requirements in term of research and development, so as to provide OSGi platform that are actually robust, and not just partially hardened. Static Code analysis seem to be very promising, but suffers from significant theoretical limitation, especially in the world of Object-Oriented Languages. Convenient permission management, and proper resource isolation in Java multi-application systems are also a strong need on the road toward OSGi security.
The present study provides a pragmatic approach to software security concerns, targeted at the world of OSGi Platforms. It present an important step toward a better understanding of OSGi-related security, and help practitioners implement safer system by providing a hardened OSGi prototype, SFelix v0.2. An important research effort is still required to provide an OSGi platform which security mechanisms can be said to be complete.
## References
- \[ACD<sup>+</sup>75\] R. P. Abbott, J. S. Chin, J. E. Donnelley, W. L. Konigsford, S. Tokubo, and D. A. Webb. Security analysis and enhancements of computer operating systems. Technical report, NATIONAL BUREAU OF STANDARDS WASHINGTONDC INST FOR COMPUTER SCIENCES AND TECHNOLOGY, December 1975.
- \[AJB00\] A.Avizienis, J.C.Laprie, and B.Randell. Fundamental concepts of dependability. Technical Report No00493, LAAS (Toulouse, France), 2000. 3rd Information Survivability Workshop (ISW’2000), Boston (USA), 24-26 Octobre 2000, pp.7-12.
- \[Ale77\] Christopher Alexander.
A Pattern Language.
Oxford University Press, 1977.
- \[BAT06\] Anil Bazaz, James D. Arthur, and Joseph G. Tront. Modeling security vulnerabilities: A constraints and assumptions perspective.
In 2nd IEEE International Symposium on Dependable, Autonomic and Secure Computing (DASC’06), 2006.
- \[BCHM99\] David W. Baker, Steven M. Christey, William H. Hill, and David E. Mann. The development of a common enumeration of vulnerabilities and exposures.
In Second International Workshop on Recent Advances in Intrusion Detection, 1999.
- \[BG05\] Joshua Bloch and Neal Gafter.
Java Puzzlers - Traps, Pitfalls and Corner Cases.
Pearson Education, June 2005.
- \[BH78\] Richard Bisbey and Dennis Hollingworth. Protection analysis: Final report. Technical Report ARPA ORDER NO. 2223, ISI/SR-78-13, Information Sciences Institute, University of Southern California, May 1978.
- \[Blo01\] Joshua Bloch.
Effective Java Programming Language Guide.
Addison-Wesley Professional, 2001.
- \[Chr05\] Steve Christey. The preliminary list of vulnerability examples for researchers (plover).
In NIST Workshop Defining the State of the Art of Software Security Tools, Gaithersburg, MD, August 2005.
- \[Chr06\] Steven M. Christey. Open letter on the interpretation of "vulnerability statistics". Bugtraq, Full-Disclosure Mailing list, January 2006.
- \[CO05\] D. Crocker and P. Overell. Augmented bnf for syntax specifications: Abnf. IETF RfC 4234, October 2005.
- \[GHJV94\] Erich Gamma, Richard Helm, Ralph Johnson, and John M. Vlissides.
Design Patterns: Elements of Reusable Object-Oriented Software.
Addison-Wesley Professional Computing Series. Addison Wesley Professional., 1994.
- \[GW05\] Michael Gegick and Laurie Williams. Matching attack patterns to security vulnerabilities in software-intensive system designs.
ACM SIGSOFT Software Engineering Notes, 30(4), July 2005.
- \[HL98\] John D. Howard and Thomas A. Longstaff. A common language for computer security incidents. Technical Report SAND98-8667, Sandia National Laboratories, USA, October 1998.
- \[HLV05\] Michael Howard, David LeBlanc, and John Viega.
19 Deadly Sins of Software Security.
McGraw-Hill Osborne Media, July 2005.
- \[Krs98\] Ivan Victor Krsul.
SOFTWARE VULNERABILITY ANALYSIS.
PhD thesis, Purdue University, May 1998.
- \[LBMC94\] Carl E. Landwehr, Alan R. Bull, John P. McDermott, and William S. Choi. A taxonomy of computer program security flaws, with examples.
In ACM Computing Surveys, volume 26, pages 211–254, September 1994.
- \[LJ97\] Ulf Lindqvist and Erland Jonsson. How to systematically classify computer security intrusions.
In IEEE Symposium on Security and Privacy, pages 154–163, May 1997.
- \[McG06\] Gary McGraw.
Software Security - Building Security In.
Pearson Education, January 2006.
- \[MCJ05\] Robert A. Martin, Steven M. Christey, and Joe Jarzombek. The case for common flaw enumeration.
In NIST Workshop on "Software Security Assurance Tools, Techniques, and Methods", Long Beach, CA., USA, November 2005.
- \[MEL01\] Andrew P. Moore, Robert J. Ellison, and Richard C. Linger. Attack modeling for information security and survivability. Technical Report CMU/SEI-2001-TN-001, CMU/SEI, March 2001.
- \[MM97\] Thomas J. Mowbray and Raphael C. Malveau.
Corba Design Patterns.
John Wiley & Sons, January 1997.
- \[Nec97\] George C. Necula. Proof-carrying code.
In Conference Record of POPL ’97: The 24th ACM SIGPLAN-SIGACT Symposium on Principles of Programming Languages, pages 106–119, Paris, France, jan 1997.
- \[OSG05\] OSGI Alliance. Osgi service platform, core specification release 4. Draft, 07 2005.
- \[PF06\] Pierre Parrend and Stephane Frenot. Secure component deployment in the osgi(tm) release 4 platform. Technical Report RT-0323, INRIA, June 2006.
- \[PF07\] Pierre Parrend and Stephane Frenot. Supporting the secure deployment of osgi bundles.
In First IEEE WoWMoM Workshop on Adaptive and DependAble Mission- and BUsiness-critical mobile Systems, Helsinki, Finland, June 2007.
- \[Sch03\] Markus Schumacher.
Security Engineering with Patterns.
Springer Verlag, 2003. LNCS n 2754.
- \[SH05\] Robert C. Seacord and Allen Householder. A structured approach to classifying security vulnerabilities. Technical Report CMU/SEI-2005-TN-003, Carnegie Mellon University - Software Engineering Institute, January 2005.
- \[SS73\] Jerome H. Saltzer and Michael D. Schroeder. The protection of information in computer systems.
In Fourth ACM Symposium on Operating System Principles, October 1973.
- \[Sun03\] Sun Microsystems, Inc. Jar file specification. Sun Java Specifications, 2003.
- \[TCM05\] Katrina Tsipenyuk, Brian Chess, and Gary McGraw. Seven pernicious kingdoms: A taxonomy of software security errors.
IEEE Security & Privacy, 3(6):81–84, November/December 2005.
- \[WKP05\] Sam Weber, Paul A. Karger, and Amit Paradkar. A software flaw taxonomy: Aiming tools at security.
In Software Engineering at Secure Systems - Building Trustworthy Applications, June 2005.
## Appendix A The OSGi platform
The OSGi Platform
<sup>23</sup>
<sup>23</sup>23http://www.osgi.org/
\[[OSG05](#bib.bibx24)\] is a componentization layer to the Java Virtual Machine. It supports the runtime extension of Java-based application through a modular approach: the applications are parted into ‘bundles’, that can be loaded, installed and managed independently from each other.
In this section, we present first an overview of the OSGi Platform, then the core concept of OSGi: the bundles and their Life Cycle, and the possible interactions between bundles.
### A.1 Overview
The OSGi Platform has been developed so as to support extensible Java-based systems in resource-constraint systems, such as automotive and mobile environments. It has since then spread into the world of Integrated Development Applications (in particular with Eclipse), and into applicative servers (IBM Websphere 6.1, Harmony, Cocoon, Directory…).
It runs as an overlay to the Java Virtual Machine (JVM). The figure [13](#A1.F13 "Figure 13 ‣ A.1 Overview ‣ Appendix A The OSGi platform") shows the overview of an OSGi-based system, with the Operating System (OS), the JVM, the platform itself, and the bundles it contains.
[Refer to caption](/html/0706.3812/assets/x13.png)
Figure 13: Overview of an OSGi Platform
Three main concepts sustain the OSGi platform: the platform, the bundle, and the interoperability between the bundles. The Platform manages the applications. The bundles are the unit of deployment and execution. The interoperability between the bundles is achieved at the class level (access to packages from other bundles) and at the service level (access to services registered by other bundles).
### A.2 The Bundles
An OSGi bundle is a Jar file \[[Sun03](#bib.bibx30)\] which is enhanced by specific meta-data. The typical structure of a bundle is shown in the figure [14](#A1.F14 "Figure 14 ‣ A.2 The Bundles ‣ Appendix A The OSGi platform"). The META-INF/MANIFEST.MF file contains the necessary OSGi meta-data: the bundle reference name (the ‘symbolic name’), its version, the dependencies and the provided resources. Some packages are exported, *i.e.* accessible from other bundles inside the platform. The activator is used by the platform as an initial bootstrap when the bundle is started. Packages can be exported. Services can be registered, so as to be available for other bundles.
[Refer to caption](/html/0706.3812/assets/x14.png)
Figure 14: Intern Structure of an OSGi bundle
Each bundle has a restricted view on the OSGi platform: the OSGi Context, which is transmitted to the bundle activator at start time. This context reference is needed to publish and look-up for services. It also supports the access to the management functionalities of the platform.
The OSGi bundles can also access the Operating System of the machine it is running on through native libraries. This possibility is not specific to the OSGi environment, since it relies on the Java Runtime API, but it allows the bundles to break their isolation.
The Life Cycle of a bundle inside the OSGi Platform is defined as follows. The bundle must first be installed. When it is required to start, the package-level dependencies with other bundles are resolved. When all dependencies are resolved, the bundle activator is launched: the start() method is called, and the related code is executed. Typically, these operations consist in configuration and publication of services. The bundle is then in the ‘started’ state. Updating, stopping and uninstalling build the last possible operations for bundle management The figure [15](#A1.F15 "Figure 15 ‣ A.2 The Bundles ‣ Appendix A The OSGi platform") shows the Life Cycle of a bundle inside a OSGi Platform.
[Refer to caption](/html/0706.3812/assets/x15.png)
Figure 15: Life Cycle of an OSGi Bundles inside the platform
### A.3 Interactions between Bundles
The interactions between the bundles are done through two complementary mechanisms: the package export/import and the service registration lookup facility. These mechanisms are shown in the figure [16](#A1.F16 "Figure 16 ‣ A.3 Interactions between Bundles ‣ Appendix A The OSGi platform").
[Refer to caption](/html/0706.3812/assets/x16.png)
Figure 16: Interaction Mechanisms between the OSGi Bundles
The publication and lookup of services are performed through the BundleContext reference that each bundle receives ar startup time. During the publication process, the advertising bundles registers a service by publishing a Java interface it is implementing, and by providing a class implementing this interface. The lookup is performed by the client bundle, which gets the service from the BundleContext and uses it as a standard Java object.
## Appendix B Vulnerabilities List
The most common Vulnerability Lists presented in the section [2.2](#S2.SS2 "2.2 From Databases to Top-Vulnerability Lists ‣ 2 Characterization of Vulnerabilities in Component-based Systems") are given here.
### B.1 The Lindqvist Classification
The computer security intrusions identified by Lindqvist \[[LJ97](#bib.bibx18)\] are the following:
- •
external misuse (not technical),
- •
hardware misuse,
- •
masquerading,
- •
setting up subsequent misuse,
- •
bypassing intended controls,
- •
active misuse of resource,
- •
passive misuse of resource,
- •
misuse resulting from inaction,
- •
use of an indirect aid in committing other misuse.
### B.2 Common Weaknesses Enumeration (CWE)
The categories defined in the Common Weaknesses Enumeration \[[MCJ05](#bib.bibx20)\] are the following:
- •
Buffer overflows, format strings, etc. \[BUFF\];
- •
Structure and Validity Problems;\[SVM\];
- •
Special Elements \[SPEC\];
- •
Common Special Element Manipulations\[SPECM\];
- •
Technology-Specific Special Elements\[SPECTS\];
- •
Pathname Traversal and Equivalence Errors \[PATH\];
- •
Channel and Path Errors \[CP\];
- •
Information Management Errors \[INFO\];
- •
Race Conditions \[RACE\];
- •
Permissions, Privileges, and ACLs \[PPA\];
- •
Handler Errors \[HAND\];
- •
User Interface Errors \[UI\];
- •
Interaction Errors \[INT\];
- •
Initialization and Cleanup Errors \[INIT\];
- •
Resource Management Errors \[RES\];
- •
Numeric Errors \[NUM\];
- •
Authentication Error \[AUTHENT\];
- •
Cryptographic errors \[CRYPTO\];
- •
Randomness and Predictability \[RAND\];
- •
Code Evaluation and Injection \[CODE\];
- •
Error Conditions, Return Values, Status Codes \[ERS\];
- •
Insufficient Verification of Data \[VER\];
- •
Modification of Assumed-Immutable Data \[MAID\];
- •
Product-Embedded Malicious Code \[MAL\];
- •
Common Attack Mitigation Failures \[ATTMIT\];
- •
Containment errors (container errors) \[CONT\];
- •
Miscellaneous WIFFs \[MISC\].
### B.3 Nineteen Dealy Sins
The 19 Deadly Sins defined by Howard \[[HLV05](#bib.bibx15)\] are the following:
- •
buffer overflows,
- •
command injection,
- •
Cross-site scripting (XSS),
- •
format string problems,
- •
integer range error,
- •
SQL injection,
- •
trusting network address information,
- •
failing to protect network traffic,
- •
failing to store and protect data,
- •
failing to use cryptographically strong random numbers,
- •
improper file access,
- •
improper use of SSL,
- •
use of weak password-based systems,
- •
unauthenticated key exchange,
- •
signal race condition,
- •
use of ’magic’ URLs and hidden forms,
- •
failure to handle errors,
- •
poor usability,
- •
information leakage.
### B.4 OWASP Top Ten
The OWASP Top Ten Vulnerability list for 2007 is the following
<sup>24</sup>
<sup>24</sup>24http://www.owasp.org/index.php/Top_10_2007-WIKI-FORMAT-TEST
:
- •
Cross Site Scripting (XSS)
- •
Injection Flaws
- •
Malicious File Execution
- •
Insecure Direct Object Reference
- •
Cross Site Request Forgery (CSRF)
- •
Information Leakage and Improper Error Handling
- •
Broken Authentication and Session Management
- •
Insecure Cryptographic Storage
- •
Insecure Communications
- •
Failure to Restrict URL Access
### B.5 Seven Kingdoms
The Seven Kingdoms defined by Gary McGraw \[[McG06](#bib.bibx19)\] are the following. Note that each Kingdom contains a certain number of Phyla, that help give more precise hints so as the actual vulnerabilities.
- •
Input Validation and representation,
- •
API abuse,
- •
Security Features,
- •
Time and state,
- •
error handling,
- •
code quality,
- •
encapsulation
- •
\+ environment
## Appendix C Formal Expression of the Vulnerability Pattern
This section presents the Vulnerability Pattern in the Augmented Backus Naur Form (BNF) \[[CO05](#bib.bibx11)\].
The current grammar is not meant to be closed: it reflects the knowledge relative to the considered vulnerabilities at a given time. It can be extended with additional attribute values.
The catalog of the OSGi Malicious Bundles is referred as the ‘mb’ catalog.
##### Vulnerability Reference
- •
VULNERABILITY_NAME ::= text
- •
IDENTIFIER ::= CATALOG_ID.SRC_REF.ID
with:
CATALOG_ID ::= mb
SRC_REF ::= archive\|java\|native\|osgi
ID ::= (0-9)\*
- •
ORIGIN ::= text
- •
LOCATION ::= Bundle ( Archive \| Manifest \| Activator \| Fragment ) \| Application Code - ( Native Code \| Java ( Code \| API ) \| OSGi API )
- •
SOURCE ::= (ENTITY ( FUNCTIONNALITY \| FLAW ;)+;)+
with ENTITY ::= OS \| JVM - ( Runtime API \| APIs )\| OSGi Platform - (( Module \| Life-Cycle \| Service ) Layer \| Bundle Repository Client )\| Application Code
FUNCTIONNALITY ::= Kill utility \| Value of Method Parameters \| ( System.exit \| Runtime.halt ) method \| Native Code Execution \| Thread API \| Reflection API \| ClassLoader API \| File API \| Java Archive \| Bundle Management \| Bundle Fragments and:
FLAW ::= No Algorithm Safety - ( Java \| Native Code )\| Non OSGi R4-compliant Digital Signature Validation in the JVM \| No Verification of Bundle Archive Validity \| No Check of Size of Loaded Bundles \| No Check of Size of stored Data \| No safe Bundle Start \| No Removal of Uninstalled Bundle Data \| Bundle Meta-data Handling - No Safe-Default \| Uncontrolled Service Registration \| Architecture of the Application - No Validation of Service Dependency
- •
TARGET ::= Platform \| OSGi Element - ( Platform Management Utility \| Bundle \| Service\|Package )
- •
CONSEQUENCE_TYPE ::= ( Unavailability \| Performance Breakdown \| Undue Access )( - ( Platform \| Service \| Package )(, ( Platform \| Service \| Package ))\*)?
- •
INTRODUCTION_TIME ::= Platform Design or Implementation \| Development \| Bundle Meta-data Generation \| Bundle Digital Signature \| Installation \| Service Publication or Resolution
- •
EXPLOIT_TIME ::= Download \| Installation \| Bundle Start \| Execution
##### Vulnerability Description
- •
DESCRIPTION ::= text
- •
PRECONDITIONS ::= text
- •
ATTACK_PROCESS ::= text
- •
CONSEQUENCE_DESCRIPTION ::= text
- •
SEE_ALSO ::= VULNERABILITY_NAME (, VULNERABILITY_NAME)\*
##### Vulnerability Implementation
- •
CODE_REFERENCE ::= FILE_NAME
with FILE_NAME the name of a file, as defined by Unix File Names
- •
OSGI_PROFILE ::= CDC-1.0/Foundation-1.0 \| OSGi/Minimum-1.1 \| JRE-1.1 \| J2SE-1.2 \| J2SE-1.3 \| J2SE-1.4 \| J2SE-1.5 \| J2SE-1.6 \| PersonalJava-1.1 \| PersonalJava-1.2 \| CDC-1.0/PersonalBasis-1.0 \| CDC-1.0/PersonalJava-1.0
- •
DATE ::= MONTH.DAY.YEAR
with MONTH ::= (1-12), DAY ::= (1-31), YEAR ::= (0-3000)
- •
TEST_COVERAGE ::= (0-100) %
- •
TESTED_ON ::= Oscar \| Felix \| Knopflerfish \| Equinox
##### Protection
- •
EXISTING_MECHANISMS ::= Java Permissions \| OSGi AdminPermission \| SFelix OSGi Security Layer \| -
- •
ENFORCEMENT_POINT ::= Platform startup \| Bundle Installation \| -
- •
POTENTIAL_MECHANISMS ::= (POTENTIAL_MECHANISM_NAME (POTENTIAL_MECHANISM_DESCR)?)+ with POTENTIAL_MECHANISM_NAME ::= Code static Analysis \| OSGi Platform Modification - ( Bundle Startup Process \| Installation Meta-data Handling \| Service Publication )\| Bundle size control before download \| Service-level dependency validation \| Resource Control and Isolation - ( CPU \| Memory \| Disk Space )\| Access Control - FileSystem \| Miscellaneous \| - and POTENTIAL_MECHANISM_DESCR ::= text
- •
ATTACK_PREVENTION ::= Stop a ill-behaving thread \| -
- •
REACTION ::= Uninstall the malicious bundle \| Erase files \| Stop the system process \| Restart the platform \| -
## Appendix D Vulnerability Catalog
### D.1 Bundle Archive
#### D.1.1 Invalid Digital Signature Validation
##### Vulnerability Reference
- •
Vulnerability Name: Invalid Digital Signature Validation
- •
Identifier: Mb.archive.1
- •
Origin: Ares research project ‘malicious-bundle’
- •
Location of Exploit Code: Bundle Archive
- •
Source: OSGi Platform - Life-Cycle Layer (Non OSGi R4-compliant Digital Signature Validation in the JVM)
- •
Target: Platform
- •
Consequence Type: Undue Access
- •
Introduction Time: Bundle Digital Signature
- •
Exploit Time: Installation
##### Vulnerability Description
- •
Description: A bundle which signature is NOT compliant to the OSGi R4 Digital Signature is installed on the platform
- •
Preconditions: No Digital Signature Validation, or Digital Signature Validation Process that relies on the Java JarFile API to perform the validation of the digital signature. The bundle signature must be non OSGi R4 compliant in one of the following ways: resources have been removed from the archive; resources have been added; the first resources in the archive are NOT the Manifest File, the Signature File and the Signature Block file in this order (see \[[PF06](#bib.bibx25)\]).
- •
Attack Process: Install a bundle with an invalid digital signature (see preconditions)
- •
Consequence Description: -
- •
See Also: -
##### Protection
- •
Existing Mechanisms: SFelix OSGi Security Layer
- •
Enforcement Point: Bundle Installation
- •
Potential Mechanisms: -
- •
Attack Prevention: -
- •
Reaction: Uninstall the malicious bundle
##### Vulnerability Implementation
- •
Code Reference: Bindex-resourceRemoved-1.0.jar, bindex-resourcesAdded-1.0.jar, bindex-unvalidResourceOrder-1.0.jar
- •
OSGi Profile: J2SE-1.5
- •
Date: 2007-04-25
- •
Test Coverage: 100%
- •
Known Vulnerable Platforms: Felix
- •
Known Robust Platforms: SFelix
#### D.1.2 Big Component Installer
##### Vulnerability Reference
- •
Vulnerability Name: Big Component Installer
- •
Identifier: Mb.archive.2
- •
Origin: Ares research project ‘malicious-bundle’
- •
Location of Exploit Code: Bundle Archive
- •
Source: OSGi Platform - Bundle Repository Client (No Check of Size of Loaded Bundles)
- •
Target: Platform
- •
Consequence Type: Performance Breakdown
- •
Introduction Time: Platform Design or Implementation
- •
Exploit Time: Execution
##### Vulnerability Description
- •
Description: Remote installation of a bundle which size is of similar to the available device memory
- •
Preconditions: OSGi platform running on a memory limited device
- •
Attack Process: -
- •
Consequence Description: Little memory is available for subsequent operations
- •
See Also: Big File Creator
##### Protection
- •
Existing Mechanisms: OSGi AdminPermission
- •
Enforcement Point: -
- •
Potential Mechanisms: Bundle size control before download
- •
Attack Prevention: -
- •
Reaction: -
##### Vulnerability Implementation
- •
Code Reference: -
- •
OSGi Profile: J2SE-1.5
- •
Date: 2007-02-20
- •
Test Coverage: 00%
#### D.1.3 Decompression Bomb
##### Vulnerability Reference
- •
Vulnerability Name: Decompression Bomb
- •
Identifier: Mb.archive.3
- •
Origin: Ares research project ‘malicious-bundle’
- •
Location of Exploit Code: Bundle Archive
- •
Source: OSGi Platform - Life-Cycle Layer (No Verification of Bundle Archive Validity)
- •
Target: Platform
- •
Consequence Type: Performance Breakdown
- •
Introduction Time: Development
- •
Exploit Time: Execution
##### Vulnerability Description
- •
Description: The Bundle Archive is a decompression Bomb (either a huge file made of identical bytes, or a recursive archive)
- •
Preconditions: -
- •
Attack Process: Provide a Bundle Archive that is a decompression Bomb for installation (on a OBR, etc.)
- •
Consequence Description: Important consumption of CPU or memory.
- •
See Also: -
##### Protection
- •
Existing Mechanisms: -
- •
Enforcement Point: -
- •
Potential Mechanisms: OSGi Platform Modification - Bundle Startup Process (Check that the Bundle is not a Decompression Bomb archive)
- •
Attack Prevention: -
- •
Reaction: -
##### Vulnerability Implementation
- •
Code Reference: Fr.inria.ares.decompressionbomb-0.1.jar
- •
OSGi Profile: J2SE-1.5
- •
Date: 2007-04-20
- •
Test Coverage: 50%
### D.2 Bundle Manifest
#### D.2.1 Duplicate Package Import
##### Vulnerability Reference
- •
Vulnerability Name: Duplicate Package Import
- •
Identifier: Mb.osgi.1
- •
Origin: Ares research project ‘malicious-bundle’
- •
Location of Exploit Code: Bundle Manifest
- •
Source: OSGi Platform - Module Layer (Bundle Meta-data Handling - No Safe-Default)
- •
Target: OSGi Element - Bundle
- •
Consequence Type: Unavailability
- •
Introduction Time: Bundle Meta-data Generation
- •
Exploit Time: Installation
##### Vulnerability Description
- •
Description: A package is imported twice (or more) according to manifest attribute ’Import-Package’. In the Felix and Knopflerfish OSGi implementations, the bundle can not be installed
- •
Preconditions: -
- •
Attack Process: -
- •
Consequence Description: -
- •
See Also: Excessive Size of Manifest File, Unvalid Activator Meta-data, Erroneous values of Manifest attributes, Insufficient User Meta-data
##### Protection
- •
Existing Mechanisms: -
- •
Enforcement Point: -
- •
Potential Mechanisms: OSGi Platform Modification - Installation Meta-data Handling (ignore the repeated imports during OSGi metadata analysis)
- •
Attack Prevention: -
- •
Reaction: -
##### Vulnerability Implementation
- •
Code Reference: Fr.inria.ares.duplicateimport-0.1.ja
- •
OSGi Profile: J2SE-1.5
- •
Date: 2006-10-28
- •
Test Coverage: 100%
- •
Known Vulnerable Platforms: Felix; Knopflerfish
- •
Known Robust Platforms: Equinox; Concierge; SFelix
#### D.2.2 Excessive Size of Manifest File
##### Vulnerability Reference
- •
Vulnerability Name: Excessive Size of Manifest File
- •
Identifier: Mb.osgi.2
- •
Origin: Ares research project ‘malicious-bundle’
- •
Location of Exploit Code: Bundle Manifest
- •
Source: OSGi Platform - Module Layer (Bundle Meta-data Handling - No Safe-Default)
- •
Target: OSGi Element - Bundle
- •
Consequence Type: Unavailability
- •
Introduction Time: Bundle Meta-data Generation
- •
Exploit Time: Installation
##### Vulnerability Description
- •
Description: A bundle with a huge number of (similar) package imports (more than 1 Mbyte)
- •
Preconditions: -
- •
Attack Process: Insert a big number of imports in the manifest file of the bundle
- •
Consequence Description: In the Felix and Knopflerfish implementations, the launcher process takes a long time (several minutes) to parse the metadata file
- •
See Also: Duplicate Package Import, Unvalid Activator Meta-data, Erroneous values of Manifest attributes, Insufficient User Meta-data
##### Protection
- •
Existing Mechanisms: -
- •
Enforcement Point: -
- •
Potential Mechanisms: OSGi Platform Modification - Installation Meta-data Handling (check the size of manifest before the installation; more generally, check the format of the manifest size before the installation)
- •
Attack Prevention: -
- •
Reaction: -
##### Vulnerability Implementation
- •
Code Reference: Fr.inria.ares.hugemanifest-0.1.jar
- •
OSGi Profile: J2SE-1.5
- •
Date: 2006-10-28
- •
Test Coverage: 100%
- •
Known Vulnerable Platforms: Felix; Knopflerfish; Concierge
- •
Known Robust Platforms: SFelix; Equinox
#### D.2.3 Erroneous values of Manifest attributes
##### Vulnerability Reference
- •
Vulnerability Name: Erroneous values of Manifest attributes
- •
Identifier: Mb.osgi.3
- •
Origin: Ares research project ‘malicious-bundle’
- •
Location of Exploit Code: Bundle Manifest
- •
Source: OSGi Platform - Module Layer (Bundle Meta-data Handling - No Safe-Default)
- •
Target: OSGi Element - Bundle
- •
Consequence Type: Unavailability
- •
Introduction Time: Bundle Meta-data Generation
- •
Exploit Time: Installation
##### Vulnerability Description
- •
Description: A bundle that provides false meta-data, in this example an non existent bundle update location
- •
Preconditions: -
- •
Attack Process: Set a false value for a given meta-data entry
- •
Consequence Description: The actions that rely on the meta-data can not be executed (here, no update possible)
- •
See Also: Duplicate Import, Excessive Size of Manifest File
##### Protection
- •
Existing Mechanisms: -
- •
Enforcement Point: -
- •
Potential Mechanisms: OSGi Platform Modification - Installation Meta-data Handling (check the format of the manifest size before the installation, and provide failsafe default)
- •
Attack Prevention: -
- •
Reaction: -
##### Vulnerability Implementation
- •
Code Reference: Fr.inria.ares.malformedupdatelocation-0.1.jar
- •
OSGi Profile: J2SE-1.5
- •
Date: 2006-10-28
- •
Test Coverage: 10%
- •
Known Vulnerable Platforms: Felix; Equinox; Knopflerfish; Concierge; SFelix
### D.3 Bundle Activator
#### D.3.1 Management Utility Freezing - Infinite Loop
##### Vulnerability Reference
- •
Vulnerability Name: Management Utility Freezing - Infinite Loop
- •
Extends: Infinite Loop in Method Call
- •
Identifier: Mb.osgi.4
- •
Origin: Ares research project ‘malicious-bundle’
- •
Location of Exploit Code: Bundle Activator
- •
Source: OSGi Platform - Life-Cycle Layer (No safe Bundle Start)
- •
Target: OSGi Element - Platform Management Utility
- •
Consequence Type: Performance Breakdown; Unavailability
- •
Introduction Time: Development
- •
Exploit Time: Bundle Start
##### Vulnerability Description
- •
Description: An infinite loop is executed in the Bundle Activator
- •
Preconditions: -
- •
Attack Process: An infinite loop is executed in the Bundle Activator
- •
Consequence Description: Block the OSGi Management entity (the felix, equinox or knopflerfish shell; when launched in the KF graphical interface, the shell remain available but the GUI is frozen). Because of the infinite loop, most CPU resource is consumed
- •
See Also: CPU Load Injection, Infinite Loop in Method Call, Stand Alone Infinite Loop, Hanging Thread
##### Protection
- •
Existing Mechanisms: -
- •
Enforcement Point: -
- •
Potential Mechanisms: Code static Analysis ; Resource Control and Isolation - CPU ; OSGi Platform Modification - Bundle Startup Process (launch the bundle activator in a separate thread to prevent startup hanging)
- •
Attack Prevention: -
- •
Reaction: -
##### Vulnerability Implementation
- •
Code Reference: Fr.inria.ares.infiniteloopinmethodcall-0.1.jar
- •
OSGi Profile: J2SE-1.5
- •
Date: 2006-08-24
- •
Test Coverage: 10%
- •
Known Vulnerable Platforms: Felix; Equinox; Knopflerfish; Concierge
- •
Known Robust Platforms: SFelix
#### D.3.2 Management Utility Freezing - Thread Hanging
##### Vulnerability Reference
- •
Vulnerability Name: Management Utility Freezing - Thread Hanging
- •
Extends: Hanging Thread
- •
Identifier: Mb.osgi.5
- •
Origin: Ares research project ‘malicious-bundle’
- •
Location of Exploit Code: Bundle Activator
- •
Source: OSGi Platform - Life-Cycle Layer (No safe Bundle Start)
- •
Target: OSGi Element - Platform Management Utility
- •
Consequence Type: Unavailability
- •
Introduction Time: Development
- •
Exploit Time: Bundle Start
##### Vulnerability Description
- •
Description: A hanging thread in the Bundle Activator makes the management utility freeze
- •
Preconditions: -
- •
Attack Process: -
- •
Consequence Description: Block the OSGi Management entity (the felix, equinox or knopflerfish shell; when launched in the KF graphical interface, the shell remain available but the GUI is frozen).
- •
See Also: Management Utility Freezing - Infinite Loop, Hanging Thread
##### Protection
- •
Existing Mechanisms: -
- •
Enforcement Point: -
- •
Potential Mechanisms: OSGi Platform Modification - Bundle Startup Process (launch the bundle activator in a separate thread); Code static Analysis
- •
Attack Prevention: -
- •
Reaction: -
##### Vulnerability Implementation
- •
Code Reference: Fr.inria.ares.hangingthread-0.1.jar, fr.inria.ares.hangingthread2-0.1.jar
- •
OSGi Profile: J2SE-1.5
- •
Date: 2006-08-24
- •
Test Coverage: 20%
- •
Known Vulnerable Platforms: Felix; Equinox; Knopflerfish; Concierge
- •
Known Robust Platforms: SFelix
### D.4 Bundle Code - Native
#### D.4.1 Runtime.exec.kill
##### Vulnerability Reference
- •
Vulnerability Name: Runtime.exec.kill
- •
Identifier: Mb.native.1
- •
Origin: Ares research project ‘malicious-bundle’
- •
Location of Exploit Code: Application Code - Native Code
- •
Source: OS (Kill utility); JVM - Runtime API (Native Code Execution)
- •
Target: Platform
- •
Consequence Type: Unavailability
- •
Introduction Time: Development
- •
Exploit Time: Execution
##### Vulnerability Description
- •
Description: A bundle that stops the execution platform through an OS call
- •
Preconditions: No SecurityManager, or FilePermission ‘execute’ on the required utilities (kill, ps, grep, cut)
- •
Attack Process: Kill the OS process which corresponds to the execution platform; this process is identified as far it is the parent process of the process in which the malicious script is executed
- •
Consequence Description: The shutdown hooks of the platforms arer executed
- •
See Also: System.exit, Runtime.halt, Recursive Thread Creation
##### Protection
- •
Existing Mechanisms: Java Permissions
- •
Enforcement Point: Platform startup
- •
Potential Mechanisms: Code static Analysis
- •
Attack Prevention: -
- •
Reaction: Restart the platform
##### Vulnerability Implementation
- •
Code Reference: Fr.inria.ares.runtime_exec_kill-0.1.jar
- •
OSGi Profile: J2SE-1.5
- •
Date: 2006-08-21
- •
Test Coverage: 100%
- •
Known Vulnerable Platforms: Felix; Equinox; Knopflerfish; Concierge; SFelix
#### D.4.2 CPU Load Injection
##### Vulnerability Reference
- •
Vulnerability Name: CPU Load Injection
- •
Identifier: Mb.native.2
- •
Origin: MOSGI, Ares research project
- •
Location of Exploit Code: Application Code - Native Code
- •
Source: Application Code (No Algorithm Safety - Native Code); JVM - Runtime API (Native Code Execution)
- •
Target: Platform
- •
Consequence Type: Unavailability
- •
Introduction Time: Development
- •
Exploit Time: Execution
##### Vulnerability Description
- •
Description: A malicious bundle that consumes 80% of the host CPU
- •
Preconditions: No SecurityManager, or RuntimePermission ‘loadLibrary’
- •
Attack Process: Execute a C call that consume CPU by switching between CPU-intensive calculation and sleep time, according to the specified ratio
- •
Consequence Description: Most of the available CPU of the system is consumed artificially
- •
See Also: Memory Load Injection, Ramping Memory Load Injection, Infinite Loop, Stand-alone Infinite Loop
##### Protection
- •
Existing Mechanisms: Java Permissions
- •
Enforcement Point: Platform startup
- •
Potential Mechanisms: Miscellaneous (extension of the Java-Level security mechanisms to the native code)
- •
Attack Prevention: -
- •
Reaction: Uninstall the malicious bundle
##### Vulnerability Implementation
- •
Code Reference: Fr.inria.ares.cpuloadinjector-0.1.jar
- •
OSGi Profile: J2SE-1.5
- •
Date: 2006-08-24
- •
Test Coverage: 00%
- •
Known Vulnerable Platforms: Felix; Equinox; Knopflerfish; Concierge; SFelix
### D.5 Bundle Code - Java
#### D.5.1 System.exit
##### Vulnerability Reference
- •
Vulnerability Name: System.exit
- •
Identifier: Mb.java.1
- •
Origin: Ares research project ‘malicious-bundle’
- •
Location of Exploit Code: Application Code - Java API
- •
Source: JVM - Runtime API (System.exit method)
- •
Target: Platform
- •
Consequence Type: Unavailability
- •
Introduction Time: Development
- •
Exploit Time: Execution
##### Vulnerability Description
- •
Description: A bundle that stops the platform by calling ’System.exit(0)’
- •
Preconditions: No SecurityManager, or presence of the RuntimePermission ‘exitVM’
- •
Attack Process: -
- •
Consequence Description: -
- •
See Also: Runtime.halt, Exec.Kill, Recursive Thread Creation
##### Protection
- •
Existing Mechanisms: Java Permissions
- •
Enforcement Point: Platform startup
- •
Potential Mechanisms: Code static Analysis
- •
Attack Prevention: -
- •
Reaction: Restart the platform
##### Vulnerability Implementation
- •
Code Reference: Fr.inria.ares.system_exit-0.1.jar
- •
OSGi Profile: J2SE-1.5
- •
Date: 2006-08-11
- •
Test Coverage: 100%
- •
Known Vulnerable Platforms: Felix; Equinox; Knopflerfish; Concierge; SFelix
#### D.5.2 Runtime.halt
##### Vulnerability Reference
- •
Vulnerability Name: Runtime.halt
- •
Identifier: Mb.java.2
- •
Origin: Ares research project ‘malicious-bundle’
- •
Location of Exploit Code: Application Code - Java API
- •
Source: JVM - Runtime API (Runtime.halt method)
- •
Target: Platform
- •
Consequence Type: Unavailability
- •
Introduction Time: Development
- •
Exploit Time: Execution
##### Vulnerability Description
- •
Description: A bundle that stops the platform by calling ’Runtime.getRuntime.halt(0)’
- •
Preconditions: No SecurityManager, or RuntimePermission ‘exitVM’
- •
Attack Process: -
- •
Consequence Description: The shutdown hooks are by-passed
- •
See Also: System.exit, Exec.Kill, Recursive Thread Creation
##### Protection
- •
Existing Mechanisms: Java Permissions
- •
Enforcement Point: Platform startup
- •
Potential Mechanisms: Code static Analysis
- •
Attack Prevention: -
- •
Reaction: Restart the platform
##### Vulnerability Implementation
- •
Code Reference: Fr.inria.ares.runtime_halt-0.1.jar
- •
OSGi Profile: J2SE-1.5
- •
Date: 2006-08-11
- •
Test Coverage: 100%
- •
Known Vulnerable Platforms: Felix; Equinox; Knopflerfish; Concierge; SFelix
#### D.5.3 Recursive Thread Creation
##### Vulnerability Reference
- •
Vulnerability Name: Recursive Thread Creation
- •
Identifier: Mb.java.3
- •
Origin: Ares research project ‘malicious-bundle’
- •
Location of Exploit Code: Application Code - Java API
- •
Source: JVM - APIs (Thread API); Application Code (No Algorithm Safety - Java)
- •
Target: Platform
- •
Consequence Type: Unavailability
- •
Introduction Time: Development
- •
Exploit Time: Execution
##### Vulnerability Description
- •
Description: The execution platform is brought to crash by the creation of an exponential number of threads
- •
Preconditions: -
- •
Attack Process: Each thread created by the attack bundle creates three other threads, and contains a relatively small payload (a pdf file). An excessive number of StackOverflowErrors causes an OutOfMemoryError
- •
Consequence Description: -
- •
See Also: System.exit, Runtime.halt, Exec.kill, Exponential Object Creation
##### Protection
- •
Existing Mechanisms: -
- •
Enforcement Point: -
- •
Potential Mechanisms: -
- •
Attack Prevention: Stop the ill-behaving thread
- •
Reaction: Restart the platform
##### Vulnerability Implementation
- •
Code Reference: Fr.inria.ares.exponentialthreadnumber-0.1.jar
- •
OSGi Profile: J2SE-1.5
- •
Date: 2006-08-21
- •
Test Coverage: 50%
- •
Known Vulnerable Platforms: Felix; Equinox; Knopflerfish; Concierge; SFelix
#### D.5.4 Hanging Thread
##### Vulnerability Reference
- •
Vulnerability Name: Hanging Thread
- •
Identifier: Mb.java.4
- •
Origin: Java puzzlers 77, 85 \[[BG05](#bib.bibx6)\]
- •
Location of Exploit Code: Application Code - Java API
- •
Source: JVM - APIs (Thread API); Application Code (Value of Method Parameters)
- •
Target: OSGi Element - Service; OSGi Element - Package
- •
Consequence Type: Unavailability
- •
Introduction Time: Development
- •
Exploit Time: Execution
##### Vulnerability Description
- •
Description: Thread that make the calling entity hang (service, or package)
- •
Preconditions: -
- •
Attack Process: Use the Thread.sleep call with a large sleep duration to make the execution hang
- •
Consequence Description: If the sleep call is performed in a synchronized block, the SIG_KILL (Ctrl+C) signal is caught by the platform
- •
See Also: Infinite Startup Loop
##### Protection
- •
Existing Mechanisms: -
- •
Enforcement Point: -
- •
Potential Mechanisms: Code static Analysis
- •
Attack Prevention: -
- •
Reaction: -
##### Vulnerability Implementation
- •
Code Reference: Fr.inria.ares.hangingthread-0.1.jar, fr.inria.ares.hangingthread2-0.1.jar
- •
OSGi Profile: J2SE-1.5
- •
Date: 2006-08-28
- •
Test Coverage: 20%
- •
Known Vulnerable Platforms: Felix; Equinox; Knopflerfish; Concierge; SFelix
#### D.5.5 Sleeping Bundle
##### Vulnerability Reference
- •
Vulnerability Name: Sleeping Bundle
- •
Identifier: Mb.java.5
- •
Origin: Ares research project ‘malicious-bundle’
- •
Location of Exploit Code: Application Code - Java API
- •
Source: JVM - APIs (Thread API)
- •
Target: OSGi Element - Service; OSGi Element - Package
- •
Consequence Type: Performance Breakdown
- •
Introduction Time: Development
- •
Exploit Time: Execution
##### Vulnerability Description
- •
Description: A malicious bundle that goes to sleep during a specified amount of time before having finished its job (experience time is 50 sec.)
- •
Preconditions: -
- •
Attack Process: -
- •
Consequence Description: -
- •
See Also: Hanging Thread
##### Protection
- •
Existing Mechanisms: -
- •
Enforcement Point: -
- •
Potential Mechanisms: Code static Analysis
- •
Attack Prevention: -
- •
Reaction: -
##### Vulnerability Implementation
- •
Code Reference: Fr.inria.ares.sleepingbundle-0.1.jar
- •
OSGi Profile: J2SE-1.5
- •
Date: 2006-08-28
- •
Test Coverage: 100%
- •
Known Vulnerable Platforms: Felix; Equinox; Knopflerfish; Concierge; SFelix
#### D.5.6 Big File Creator
##### Vulnerability Reference
- •
Vulnerability Name: Big File Creator
- •
Identifier: Mb.java.6
- •
Origin: Ares research project ‘malicious-bundle’
- •
Location of Exploit Code: Application Code - Java API
- •
Source: JVM - APIs (File API)
- •
Target: Platform
- •
Consequence Type: Performance Breakdown
- •
Introduction Time: Development
- •
Exploit Time: Execution
##### Vulnerability Description
- •
Description: A malicious bundle that create a big (relative to available resources) files to consume disk memory space
- •
Preconditions: No SecurityManager, or FilePermission ‘write’ to the malicious bundl
- •
Attack Process: -
- •
Consequence Description: -
- •
See Also: Big Bundle Installer
##### Protection
- •
Existing Mechanisms: Java Permissions
- •
Enforcement Point: Platform startup
- •
Potential Mechanisms: Resource Control and Isolation - Disk Space (per user/bundle); Access Control - FileSystem (Limit the access to the FileSystem to the data directory of the bundle; control the size of the data created through the BundleContext)
- •
Attack Prevention: -
- •
Reaction: Erase files
##### Vulnerability Implementation
- •
Code Reference: Fr.inria.ares.bigfilecreator-0.1.jar
- •
OSGi Profile: J2SE-1.5
- •
Date: 2006-08-29
- •
Test Coverage: 00%
- •
Known Vulnerable Platforms: Felix; Equinox; Knopflerfish; Concierge; SFelix
#### D.5.7 Code Observer
##### Vulnerability Reference
- •
Vulnerability Name: Code Observer
- •
Identifier: Mb.java.7
- •
Origin: Ares research project ‘malicious-bundle’
- •
Location of Exploit Code: Application Code - Java API
- •
Source: JVM - APIs (Reflection API; ClassLoader API)
- •
Target: OSGi Element - Package; OSGi Element - Service
- •
Consequence Type: Undue Access
- •
Introduction Time: Development
- •
Exploit Time: Execution
##### Vulnerability Description
- •
Description: A component that observes the content of another one
- •
Preconditions: No SecurityManager, or ReflectPermission
- •
Attack Process: Use of the reflection API and the ClassLoader API
- •
Consequence Description: Observation of the implementation of the published packages and services, the classes that are aggregated to these packages and services, and classes which name are known (or guessed)
- •
See Also: Component Data Modifier, Hidden Method Launcher
##### Protection
- •
Existing Mechanisms: Java Permissions
- •
Enforcement Point: Platform startup
- •
Potential Mechanisms: -
- •
Attack Prevention: -
- •
Reaction: -
##### Vulnerability Implementation
- •
Code Reference: Privateclassspy/fr.inria.ares.serviceabuser-0.1.jar
- •
OSGi Profile: J2SE-1.5
- •
Date: 2007-02-12
- •
Test Coverage: 100%
- •
Known Vulnerable Platforms: Felix; Equinox; Knopflerfish; SFelix
#### D.5.8 Component Data Modifier
##### Vulnerability Reference
- •
Vulnerability Name: Component Data Modifier
- •
Extends: Code Observer
- •
Identifier: Mb.java.8
- •
Origin: Ares research project ‘malicious-bundle’
- •
Location of Exploit Code: Application Code - Java API
- •
Source: JVM - APIs (Reflection API; ClassLoader API)
- •
Target: OSGi Element - Package; OSGi Element - Service
- •
Consequence Type: Undue Access
- •
Introduction Time: Development
- •
Exploit Time: Execution
##### Vulnerability Description
- •
Description: A bundle that modifies the data (i.e. the value of the attributes of the classes) of another one
- •
Preconditions: No SecurityManager or ReflectPermission set; the name of the non-exported modified class must be known beforehand, and contain a public static field.
- •
Attack Process: Use of the reflection API and the ClassLoader API to access and modify the value of attributes
- •
Consequence Description: Modification of the public fields of the objects that are accessible from another bundle (service implementations, or objects that are attributes of these service implementations, or objects that are attributes of these latter objects, …), or of the public static (non final) fields of classes which name is known
- •
See Also: Code Observer, Hidden Method Launcher
##### Protection
- •
Existing Mechanisms: Java Permissions
- •
Enforcement Point: Platform startup
- •
Potential Mechanisms: -
- •
Attack Prevention: -
- •
Reaction: -
##### Vulnerability Implementation
- •
Code Reference: Privateclassmanipulator/fr.inria.ares.serviceabuser-0.1.jar
- •
OSGi Profile: J2SE-1.5
- •
Date: 2007-02-12
- •
Test Coverage: 100%
- •
Known Vulnerable Platforms: Felix; Equinox; Knopflerfish; Concierge; SFelix
#### D.5.9 Hidden Method Launcher
##### Vulnerability Reference
- •
Vulnerability Name: Hidden Method Launcher
- •
Extends: Code Observer
- •
Identifier: Mb.java.9
- •
Origin: Ares research project ‘malicious-bundle’
- •
Location of Exploit Code: Application Code - Java API
- •
Source: JVM - APIs (Reflection API; ClassLoader API)
- •
Target: OSGi Element - Package
- •
Consequence Type: Undue Access
- •
Introduction Time: Development
- •
Exploit Time: Execution
##### Vulnerability Description
- •
Description: A bundle that executes methods from classes that are not exported of provided as service. All classes that are referenced (directly or indirectly) as class attributes can be accessed. Only public methods can be invoked
- •
Preconditions: No SecurityManager, or ReflectPermission set
- •
Attack Process: Use of the reflection API and the ClassLoader API to access and executes methods in classes that are not exported by the bundle
- •
Consequence Description: -
- •
See Also: Code Observer, Component Data Modifier
##### Protection
- •
Existing Mechanisms: Java Permissions
- •
Enforcement Point: Platform startup
- •
Potential Mechanisms: -
- •
Attack Prevention: -
- •
Reaction: -
##### Vulnerability Implementation
- •
Code Reference: Hiddenclassexecutor/fr.inria.ares.serviceabuser-0.1.jar
- •
OSGi Profile: J2SE-1.5
- •
Date: 2007-02-12
- •
Test Coverage: 100%
- •
Known Vulnerable Platforms: Felix; Equinox; Knopflerfish; Concierge; SFelix
#### D.5.10 Memory Load Injection
##### Vulnerability Reference
- •
Vulnerability Name: Memory Load Injection
- •
Identifier: Mb.java.10
- •
Origin: MOSGI Ares Research Project (OSGi Platform Monitoring)
- •
Location of Exploit Code: Application Code - Java API
- •
Source: Application Code (No Algorithm Safety - Java)
- •
Target: Platform
- •
Consequence Type: Performance Breakdown
- •
Introduction Time: Development
- •
Exploit Time: Execution
##### Vulnerability Description
- •
Description: A malicious bundle that consumes most of available memory (61,65 MB in the example)
- •
Preconditions: -
- •
Attack Process: Store a huge amount of data in a byte array
- •
Consequence Description: Only a limited memory space is available for the execution of programs
- •
See Also: Ramping Memory Load Injection, CPU Load Injection
##### Protection
- •
Existing Mechanisms: -
- •
Enforcement Point: -
- •
Potential Mechanisms: Code static Analysis ; Resource Control and Isolation - Memory
- •
Attack Prevention: -
- •
Reaction: -
##### Vulnerability Implementation
- •
Code Reference: Fr.inria.ares.memloadinjector-0.1.jar
- •
OSGi Profile: J2SE-1.5
- •
Date: 2006-08-24
- •
Test Coverage: 100%
- •
Known Vulnerable Platforms: Felix; Equinox; Knopflerfish; Concierge; SFelix
#### D.5.11 Stand Alone Infinite Loop
##### Vulnerability Reference
- •
Vulnerability Name: Stand Alone Infinite Loop
- •
Identifier: Mb.java.11
- •
Origin: Ares research project ‘malicious-bundle’
- •
Location of Exploit Code: Application Code - Java Code
- •
Source: Application Code (No Algorithm Safety - Java)
- •
Target: Platform
- •
Consequence Type: Performance Breakdown
- •
Introduction Time: Development
- •
Exploit Time: Execution
##### Vulnerability Description
- •
Description: A void loop in a lonesome thread that consumes much of the available CPU
- •
Preconditions: -
- •
Attack Process: Infinite loop launched in an independent thread
- •
Consequence Description: -
- •
See Also: Infinite Startup Loop, CPU Load Injection
##### Protection
- •
Existing Mechanisms: -
- •
Enforcement Point: -
- •
Potential Mechanisms: Code static Analysis ; Resource Control and Isolation - CPU
- •
Attack Prevention: -
- •
Reaction: -
##### Vulnerability Implementation
- •
Code Reference: Fr.inria.ares.standaloneloop-0.1.jar
- •
OSGi Profile: J2SE-1.5
- •
Date: 2006-09-22
- •
Test Coverage: 10%
- •
Known Vulnerable Platforms: Felix; Equinox; Knopflerfish; Concierge; SFelix
#### D.5.12 Infinite Loop in Method Call
##### Vulnerability Reference
- •
Vulnerability Name: Infinite Loop in Method Call
- •
Identifier: Mb.java.12
- •
Origin: Java puzzlers 26 to 33 \[[BG05](#bib.bibx6)\]
- •
Location of Exploit Code: Application Code - Java Code
- •
Source: Application Code (No Algorithm Safety - Java)
- •
Target: Platform; OSGi Element - Service; OSGi Element - Package
- •
Consequence Type: Performance Breakdown - Platform; Unavailability - Service, Package
- •
Introduction Time: Development
- •
Exploit Time: Execution
##### Vulnerability Description
- •
Description: An infinite loop is executed in a method call (at class use, package use)
- •
Preconditions: -
- •
Attack Process: An infinite loop is executed in a method call
- •
Consequence Description: Block the calling entity (the calling class or service. Because of the infinite loop, most CPU resource is consumed
- •
See Also: CPU Load Injection, Stand-alone Infinite Loop, Hanging Thread
##### Protection
- •
Existing Mechanisms: -
- •
Enforcement Point: -
- •
Potential Mechanisms: Code static Analysis ; Resource Control and Isolation - CPU
- •
Attack Prevention: -
- •
Reaction: -
##### Vulnerability Implementation
- •
Code Reference: Fr.inria.ares.infiniteloopinmethodcall-0.1.jar
- •
OSGi Profile: J2SE-1.5
- •
Date: 2006-08-24
- •
Test Coverage: 10%
- •
Known Vulnerable Platforms: Felix; Equinox; Knopflerfish; Concierge; SFelix
#### D.5.13 Exponential Object Creation
##### Vulnerability Reference
- •
Vulnerability Name: Exponential Object Creation
- •
Identifier: Mb.java.13
- •
Origin: Ares research project ‘malicious-bundle’
- •
Location of Exploit Code: Application Code - Java Code
- •
Source: Application Code (No Algorithm Safety - Java)
- •
Target: OSGi Element - Service; OSGi Element - Package
- •
Consequence Type: Unavailability
- •
Introduction Time: Development
- •
Exploit Time: Execution
##### Vulnerability Description
- •
Description: Objects are created in a exponential way
- •
Preconditions: -
- •
Attack Process: A given object create in its constructor 3 instances of object of the same class
- •
Consequence Description: The method call aborts with a ’StackOverflowError’
- •
See Also: Recursive Thread Creation
##### Protection
- •
Existing Mechanisms: -
- •
Enforcement Point: -
- •
Potential Mechanisms: Code static Analysis ; Resource Control and Isolation - Memory
- •
Attack Prevention: -
- •
Reaction: -
##### Vulnerability Implementation
- •
Code Reference: Fr.inria.ares.exponentialobjectcreation-0.1
- •
OSGi Profile: J2SE-1.5
- •
Date: 2007-01-09
- •
Test Coverage: 50%
- •
Known Vulnerable Platforms: Felix; Equinox; Knopflerfish; Concierge; SFelix
### D.6 Bundle Code - OSGi APi
#### D.6.1 Launch a Hidden Bundle
##### Vulnerability Reference
- •
Vulnerability Name: Launch a Hidden Bundle
- •
Identifier: Mb.osgi.6
- •
Origin: Ares research project ‘malicious-bundle’
- •
Location of Exploit Code: Application Code - OSGi API
- •
Source: OSGi Platform - Life-Cycle Layer (Bundle Management); JVM - APIs (File API)
- •
Target: Platform
- •
Consequence Type: Undue Access
- •
Introduction Time: Development
- •
Exploit Time: Execution
##### Vulnerability Description
- •
Description: A bundle that launches another bundle it contains (the contained bundle could be masqued as with a ’MyFile.java’ file name)
- •
Preconditions: No SecurityManager, or OSGi PermissionAdmin and FilePermission ‘write’ for the malicious bundle
- •
Attack Process: A bundle creates a new bundle on the file system, and launches i
- •
Consequence Description: A non foreseen bundle is installed. If install time checking exists (such as digital signature), it passes through the verification process
- •
See Also: Pirat Bundle Manager
##### Protection
- •
Existing Mechanisms: Java Permissions
- •
Enforcement Point: Platform startup
- •
Potential Mechanisms: -
- •
Attack Prevention: -
- •
Reaction: Uninstall the malicious bundle
##### Vulnerability Implementation
- •
Code Reference: Fr.inria.ares.silentloader-0.1.jar, fr.inria.ares.silentloader.concierge-0.1.jar (without swing)
- •
OSGi Profile: J2SE-1.5
- •
Date: 2006-10-28
- •
Test Coverage: 100%
- •
Known Vulnerable Platforms: Felix; Equinox; Knopflerfish; Concierge; SFelix
#### D.6.2 Pirat Bundle Manager
##### Vulnerability Reference
- •
Vulnerability Name: Pirat Bundle Manager
- •
Identifier: Mb.osgi.7
- •
Origin: Ares research project ‘malicious-bundle’
- •
Location of Exploit Code: Application Code - OSGi API
- •
Source: OSGi Platform - Life-Cycle Layer (Bundle Management)
- •
Target: OSGi Element - Bundle
- •
Consequence Type: Undue Access
- •
Introduction Time: Development
- •
Exploit Time: Execution
##### Vulnerability Description
- •
Description: A bundle that manages others without being requested to do so(here: stops and starts the victim bundle)
- •
Preconditions: No SecurityManager, or OSGi Permission Admin
- •
Attack Process: The pirat bundle accesses to the bundle context, and then to its victim bundle
- •
Consequence Description: -
- •
See Also: Launch Hidden Bundle
##### Protection
- •
Existing Mechanisms: Java Permissions
- •
Enforcement Point: Platform startup
- •
Potential Mechanisms: -
- •
Attack Prevention: -
- •
Reaction: -
##### Vulnerability Implementation
- •
Code Reference: Fr.inria.ares.piratbundlemanager-0.1.jar, fr.inria.ares.piratbundlemanager.concierge-0.1.jar (no swing)
- •
OSGi Profile: J2SE-1.5
- •
Date: 2006-10-30
- •
Test Coverage: 40%
- •
Known Vulnerable Platforms: Felix; Equinox; Knopflerfish; Concierge; SFelix
#### D.6.3 Zombie Data
##### Vulnerability Reference
- •
Vulnerability Name: Zombie Data
- •
Identifier: Mb.osgi.8
- •
Origin: Ares research project ‘malicious-bundle’
- •
Location of Exploit Code: Application Code - OSGi API
- •
Source: OSGi Platform - Life-Cycle Layer (No Removal of Uninstalled Bundle Data)
- •
Target: Platform
- •
Consequence Type: Performance Breakdown
- •
Introduction Time: Development
- •
Exploit Time: Execution
##### Vulnerability Description
- •
Description: Data stored in the local OSGi data store are not deleted when the related bundle is uninstalled. It thus becomes unavailable and consumes disks space (especially on resource constraint devices)
- •
Preconditions: No SecuriyManager, or FilePermission set
- •
Attack Process: -
- •
Consequence Description: -
- •
See Also: Big File Creator
##### Protection
- •
Existing Mechanisms: Java Permissions
- •
Enforcement Point: Platform startup
- •
Potential Mechanisms: OSGi Platform Modification - Bundle Uninstall Process (Delete Bundle Data when Bundles are uninstalled)
- •
Attack Prevention: -
- •
Reaction: Erase files
##### Vulnerability Implementation
- •
Code Reference: Fr.inria.ares.bigfilecreator-0.1.jar
- •
OSGi Profile: J2SE-1.5
- •
Date: 2007-04-20
- •
Test Coverage: 100%
- •
Known Vulnerable Platforms: Felix; Concierge
- •
Known Robust Platforms: Equinox; Knopflerfish; SFelix
#### D.6.4 Cycle Between Services
##### Vulnerability Reference
- •
Vulnerability Name: Cycle Between Services
- •
Identifier: Mb.osgi.9
- •
Origin: Ares research project ‘malicious-bundle’
- •
Location of Exploit Code: Application Code - OSGi API
- •
Source: OSGi Platform - Service Layer (Architecture of the Application - No Validation of Service Dependency)
- •
Target: OSGi Element - Service; OSGi Element - Package
- •
Consequence Type: Unavailability
- •
Introduction Time: Service Publication or Resolution
- •
Exploit Time: Execution
##### Vulnerability Description
- •
Description: A cycle exists in the services call
- •
Preconditions: -
- •
Attack Process: Service 1 calls service 2, which calls service 1. The attack can be implemented as a fake ’service 2’, which calls service 1 back instead of return properly from the method that service 1 called
- •
Consequence Description: ‘java.lang.StackOverflowError’, service 1 can not be executed
- •
See Also: -
##### Protection
- •
Existing Mechanisms: -
- •
Enforcement Point: -
- •
Potential Mechanisms: Service-level dependency validation
- •
Attack Prevention: -
- •
Reaction: -
##### Vulnerability Implementation
- •
Code Reference: Fr.inria.ares.clientserver1-0.1.jar, fr.inria.ares.clientserver2-0.1.jar
- •
OSGi Profile: J2SE-1.5
- •
Date: 2006-10-28
- •
Test Coverage: 100%
- •
Known Vulnerable Platforms: Felix; Equinox; Knopflerfish; Concierge; SFelix
#### D.6.5 Numerous Service Registration
##### Vulnerability Reference
- •
Vulnerability Name: Numerous Service Registration
- •
Identifier: Mb.osgi.10
- •
Origin: Ares research project ‘malicious-bundle’
- •
Location of Exploit Code: Application Code - OSGi API
- •
Source: OSGi Platform - Service Layer (Uncontrolled Service Registration)
- •
Target: OSGi Element - Bundle; OSGi Element - Platform Management Utility
- •
Consequence Type: Performance Breakdown
- •
Introduction Time: Development
- •
Exploit Time: Execution
##### Vulnerability Description
- •
Description: Registration of a high number of (possibly identical) services through an loop
- •
Preconditions: No SecurityManager, or OSGi ServicePermission
- •
Attack Process: Publish a given service in a(n) (e.g. infinite) loop
- •
Consequence Description: Important duration of bundle stop
- •
See Also: -
##### Protection
- •
Existing Mechanisms: Java Permissions
- •
Enforcement Point: Platform startup
- •
Potential Mechanisms: OSGi Platform Modification - Service Publication (limitation of the number of services published in the framework)
- •
Attack Prevention: -
- •
Reaction: -
##### Vulnerability Implementation
- •
Code Reference: Fr.inria.ares.numerousservices-0.1.jar
- •
OSGi Profile: J2SE-1.5
- •
Date: 2007-01-09
- •
Test Coverage: 100%
- •
Known Vulnerable Platforms: Felix; Equinox; Knopflerfish; Concierge
- •
Known Robust Platforms: SFelix
#### D.6.6 Freezing Numerous Service Registration
##### Vulnerability Reference
- •
Vulnerability Name: Freezing Numerous Service Registration
- •
Identifier: Mb.osgi.11
- •
Origin: Ares research project ‘malicious-bundle’
- •
Location of Exploit Code: Application Code - OSGi API
- •
Source: OSGi Platform - Service Layer (Uncontrolled Service Registration)
- •
Target: OSGi Element - Bundle; OSGi Element - Platform Management Utility
- •
Consequence Type: Performance Breakdown
- •
Introduction Time: Development
- •
Exploit Time: Execution
##### Vulnerability Description
- •
Description: Registration of a high number of (possibly identical) services through an loop, in the Concierge OSGi Platform implementation
- •
Preconditions: No SecurityManager, or OSGi ServicePermission; execution in the Concierge OSGi Platform
- •
Attack Process: Publish a given service in a(n) (e.g. infinite) loop
- •
Consequence Description: The Platform almost totally freeze. OutOfMemoryErrors are reported very frequently when the shell is used or when bundles perform actions.
- •
See Also: -
##### Protection
- •
Existing Mechanisms: Java Permissions
- •
Enforcement Point: Platform startup
- •
Potential Mechanisms: OSGi Platform Modification - Service Publication (limitation of the number of services published in the framework)
- •
Attack Prevention: -
- •
Reaction: -
##### Vulnerability Implementation
- •
Code Reference: Fr.inria.ares.numerousservices-0.1.jar
- •
OSGi Profile: J2SE-1.5
- •
Date: 2007-04-20
- •
Test Coverage: 100%
- •
Known Vulnerable Platforms: Concierge
### D.7 Bundle Fragments
#### D.7.1 Execute Hidden Classes
##### Vulnerability Reference
- •
Vulnerability Name: Execute Hidden Classes
- •
Identifier: Mb.osgi.12
- •
Origin: Ares research project ‘malicious-bundle’
- •
Location of Exploit Code: Bundle Fragment
- •
Source: OSGi Platform - Module Layer (Bundle Fragments)
- •
Target: OSGi Element - Package
- •
Consequence Type: Undue Access
- •
Introduction Time: Development
- •
Exploit Time: Execution
##### Vulnerability Description
- •
Description: A fragment bundle exports a pckage that is not made visible by the host. Other bundles can then execute the classes in this package
- •
Preconditions: No SecurityManager, or BundlePermission, ‘HOST’ set to the host, and BundlePermission, ‘FRAGMENT’, set to the malicious fragment
- •
Attack Process: -
- •
Consequence Description: Modification of static attributes, publication of hidden data or execution of secret procedure; Concierge does not support fragment, and does therefore not contains this vulnerability
- •
See Also: Fragment Substitution, Access Protected Package through split Packages
##### Protection
- •
Existing Mechanisms: Java Permissions
- •
Enforcement Point: Platform startup
- •
Potential Mechanisms: -
- •
Attack Prevention: -
- •
Reaction: -
##### Vulnerability Implementation
- •
Code Reference: Usehiddenclass/packageexportfragment.jar
+ usehiddenclass/fr.inria.ares.fragmentaccomplice-0.1.jar
- •
OSGi Profile: J2SE-1.5
- •
Date: 2007-02-14
- •
Test Coverage: 100%
- •
Known Vulnerable Platforms: Felix; Equinox; Knopflerfish; SFelix
#### D.7.2 Fragment Substitution
##### Vulnerability Reference
- •
Vulnerability Name: Fragment Substitution
- •
Identifier: Mb.osgi.13
- •
Origin: Ares research project ‘malicious-bundle’
- •
Location of Exploit Code: Bundle Fragment
- •
Source: OSGi Platform - Module Layer (Bundle Fragments)
- •
Target: OSGi Element - Bundle
- •
Consequence Type: Undue Access
- •
Introduction Time: Development
- •
Exploit Time: Execution
##### Vulnerability Description
- •
Description: A specific fragment bundle is replace by another, which provides the same classes but with malicious implementation
- •
Preconditions: No SecurityManager, or BundlePermission ‘HOST’ and ‘FRAGMENT’ set to the required bundles, and OSGi AdminPermission set to the substitutor bundle
- •
Attack Process: A malicious bundles uninstalls the current fragment, and install a malicious one (that is embedded in it) instead
- •
Consequence Description: The host bundle executes a false implementations of classes provided by a fragment; Concierge does not support fragment, and does therefore not contains this vulnerability
- •
See Also: Launch Hidden Bundle, Pirat Bundle Manager, Execute Hidden Class, Access Protected Package through split Packages
##### Protection
- •
Existing Mechanisms: Java Permissions
- •
Enforcement Point: Platform startup
- •
Potential Mechanisms: -
- •
Attack Prevention: -
- •
Reaction: -
##### Vulnerability Implementation
- •
Code Reference: Fragmentprovidespackagestohost/fr.inria.ares.fragmentsubstitutor (has an embedded fragmentprovidespackagestohost/testfragmentclone bundle)
- •
OSGi Profile: J2SE-1.5
- •
Date: 2007-02-16
- •
Test Coverage: 100%
- •
Known Vulnerable Platforms: Felix; Equinox; Knopflerfish; SFelix
#### D.7.3 Access Protected Package through split Packages
##### Vulnerability Reference
- •
Vulnerability Name: Access Protected Package through split Packages
- •
Identifier: Mb.osgi.14
- •
Origin: Ares research project ‘malicious-bundle’
- •
Location of Exploit Code: Bundle Fragment
- •
Source: OSGi Platform - Module Layer (Bundle Fragments)
- •
Target: OSGi Element - Package
- •
Consequence Type: Undue Access
- •
Introduction Time: Development
- •
Exploit Time: Execution
##### Vulnerability Description
- •
Description: A package is built in the fragment, that have the same name than a package in the host. All package-protected classes and methods can then be accessed from the fragment, and through a proxy exported in the framework
- •
Preconditions: No SecurityManager, or BundlePermission ‘HOST’ and ‘FRAGMENT’ set to the suitable bundles
- •
Attack Process: -
- •
Consequence Description: Concierge does not support fragment, and does therefore not contains this vulnerability
- •
See Also: Execute Hidden Class, Fragment Substitution
##### Protection
- •
Existing Mechanisms: Java Permissions
- •
Enforcement Point: Platform startup
- •
Potential Mechanisms: Miscellaneous (the Java Archive defined ’seal’ keyword in the Manifest File does not prevent the package to be completed by a split package in the fragment. It probably should)
- •
Attack Prevention: -
- •
Reaction: -
##### Vulnerability Implementation
- •
Code Reference: Fragmentsplitpackage/fr.inria.ares.testhostbundle-0.1.jar +
fragmentsplitpackage/testfragment-0.1.jar +
fragmentsplitpackage/fr.inria.ares.fragmentclient-0.1.jar
- •
OSGi Profile: J2SE-1.5
- •
Date: 2007-02-17
- •
Test Coverage: 100%
- •
Known Vulnerable Platforms: Felix; Equinox; Knopflerfish; SFelix
## Appendix E Attack Implementations
The implementations of the presented attacks are given. Two types of attacks can be performed through several different implementations.
### E.1 Infinite Loops
The various implementations of Infinites Loops in Java are given. They match the Vulnerability mb.java.8 and mb.java.12 in our catalog.
#### E.1.1 First Option
```
boolean condition==true;
While(condition);
```
#### E.1.2 Second Option
This implementation is given in the Java Puzzler \#26 \[[BG05](#bib.bibx6)\].
```
public static final int END = Integer.MAX\_VALUE;
public static final int START = END - 100; // or any other start value
for (int i = START; i <= END; i++);
```
#### E.1.3 Third Option
This implementation is given in the Java Puzzler \#27 \[[BG05](#bib.bibx6)\].
```
int i = 0;
while (-1 << i != 0)
{i++;}
```
#### E.1.4 Fourth Option
This implementation is given in the Java Puzzler \#28 \[[BG05](#bib.bibx6)\].
```
double i = 1.0/0.0; //can also be set to Double.POSITIVE\_INFINITY,
//or a sufficiently big number (such as 1.0e17 or bigger)
while (i == i + 1);
```
#### E.1.5 Fifth Option
This implementation is given in the Java Puzzler \#45 \[[BG05](#bib.bibx6)\].
```
public static void main(String[] args) {
workHard();
}
private static void workHard() {
try {
workHard();
} finally {
workHard();
}
}
```
#### E.1.6 Other Implementations
For further implementation of the infinite loop, you can also refer to
- •
the Java Puzzler \#29 (double i = Double.NaN;while(i! =i);),
- •
the Java Puzzler \#30 (String i = “a”; while (i != i + 0)), puzzler 31 ( short i = -1; while (i != 0) i \>\>\>= 1;),
- •
the Java Puzzler \#32 (Integer i = new Integer(0); Integer j = new Integer(0); while (i \<= j && j \<= i && i != j);),
- •
the Java Puzzler \#33 (int i = Integer.MIN_VALUE; while (i != 0 && i == -i);(//or long i = Long.MIN_VALUE), for() loop with unsuitable modification of the variables that intervene in the stop condition).
### E.2 Hanging Thread
The various implementations of Hanging Threads in Java are given. They match the Vulnerability mb.java.9 in our catalog.
#### E.2.1 First Option
This implementation is given in the Java Puzzler \#77 \[[BG05](#bib.bibx6)\].
```
import java.util.Timer;
import java.util.TimerTask;
public class Stopper extends Thread{
private volatile boolean quittingTime = false;
public void run() {
while (!quittingTime)
pretendToWork();
System.out.println(‘‘Beer is good’’);
}
private void pretendToWork() {
try {
Thread.sleep(300); // Sleeping on the job?
} catch (InterruptedException ex) { }
}
// It’s quitting time, wait for worker - Called by good boss
synchronized void quit() throws InterruptedException {
quittingTime = true;
join();
}
// Rescind quitting time - Called by evil boss
synchronized void keepWorking() {
quittingTime = false;
}
public void hang(){
System.out.println(‘‘HangingThread Stopper ready’’
¯¯+‘‘to behave badly’’);
try
{
final Stopper worker = new Stopper();
worker.start();
Timer t = new Timer(true); // Daemon thread
t.schedule(new TimerTask() {
public void run() { worker.keepWorking(); }
}, 500);
Thread.sleep(400);
worker.quit();
}
catch( InterruptedException e)
{
e.printStackTrace();
}
}
}
```
#### E.2.2 Second Option
This implementation is given in the Java Puzzler \#85 \[[BG05](#bib.bibx6)\].
```
static {
Thread t = new Thread(new Runnable() {
public void run() {
initialized = true;
}
});
t.start();
try {
t.join();
} catch(InterruptedException e) {
throw new AssertionError(e);
}
}
```
#### E.2.3 Other Implementations
For other implementations, see the JLint Manual
<sup>25</sup>
<sup>25</sup>25http://artho.com/jlint/manual.html
for ‘deadlock errors’ (2 ocurrences).
## Appendix F XML2Tex Documentation Generator
To ease the validation of the Vulnerability Patterns and the generation of the catalog, we developed a small tool based on XML technologies, XML2Tex. The overall process of the XML2tex Documentation Generation process is given in Figure [17](#A6.F17 "Figure 17 ‣ Appendix F XML2Tex Documentation Generator").
Figure [17](#A6.F17 "Figure 17 ‣ Appendix F XML2Tex Documentation Generator") presents the overview of the XML2tex Documentation Generation process.
[Refer to caption](/html/0706.3812/assets/x17.png)
Figure 17: XML2Tex Process
First, the vulnerability pattern is checked against a reference XML Schema (XSD). Secondly, its is transformed into an XML report through XSL transformations. An XML Report is a specific XML file which is easily mappable to documentation: its contains in particular a title, paragraphs, and so on. Thirdly, the XML report is transformed into a Tex file through an Ad-Hoc parser named Report2Tex.
The validity of the Vulnerability Patterns of the catalog is guaranteed by their validation against the formal expression of the Pattern given in the appendix [C](#A3 "Appendix C Formal Expression of the Vulnerability Pattern"). The well-formedness of the XML Report and of the Tex file are ensures by the successive parsers. |
0708.3022 | cs.CR | cyber | # Efficient FPGA-based multipliers for $`{\mathbb{F}}_{3^{97}}`$ and $`{\mathbb{F}}_{3^{6 \cdot 97}}`$
###### Abstract
In this work we present a new structure for multiplication in finite fields. This structure is based on a digit-level LFSR (Linear Feedback Shift Register) multiplier in which the area of digit-multipliers are reduced using the Karatsuba method. We compare our results with the other works of the literature for $`{\mathbb{F}}_{3^{97}}`$. We also propose new formulas for multiplication in $`{\mathbb{F}}_{3^{6 \cdot 97}}`$. These new formulas reduce the number of $`{\mathbb{F}}_{3^{97}}`$-multiplications from $`18`$ to $`15`$. The finite fields $`{\mathbb{F}}_{3^{97}}`$ and $`{\mathbb{F}}_{3^{6 \cdot 97}}`$ are important fields for pairing based cryptography.
Keywords: finite field multiplication, FPGA, pairing based cryptography.
## 1 Introduction
Efficient multiplication in finite fields is a central task in the implementation of most public key cryptosystems. A great amount of work has been devoted to this topic (see \[[1](#bib.bib1)\] or \[[2](#bib.bib2)\] for a comprehensive list). The two types of finite fields which are mostly used in cryptographic standards are binary finite fields of type $`{\mathbb{F}}_{2^{m}}`$ and prime fields of type $`{\mathbb{F}}_{p}`$, where $`p`$ is a prime (cf. \[[3](#bib.bib3)\]). Efforts to efficiently fit finite field arithmetic into commercial processors resulted into applications of medium characteristic finite fields like those reported in \[[4](#bib.bib4)\] and \[[5](#bib.bib5)\]. Medium characteristic finite fields are fields of type $`{\mathbb{F}}_{p^{m}}`$, where $`p`$ is a prime slightly smaller than the word size of the processor, and has a special form that simplifies the modular reduction. Mersenne prime numbers constitute an example of primes which are used in this context. The security parameter is given by the length of the binary representations of the field elements, and the extension degree $`m`$ is selected appropriately. Due to security considerations, the extension degree for fields of characteristic $`2`$ or medium characteristic is usually chosen to be prime.
With the introduction of the method of Duursma and Lee for the computation of the Tate pairing (cf. \[[6](#bib.bib6)\]), fields of type $`{\mathbb{F}}_{3^{m}}`$ for $`m`$ prime have attracted special attention. Computing the Tate pairing on elliptic curves defined over $`{\mathbb{F}}_{3^{m}}`$ requires computations both in $`{\mathbb{F}}_{3^{m}}`$ and in $`{\mathbb{F}}_{3^{6\hspace{0pt}m}}`$. In \[[7](#bib.bib7)\] calculations are implemented using the tower of extensions
<table>
<tbody>
<tr>
<td></td>
<td>𝔽3m⊂𝔽32m⊂𝔽36msubscript𝔽superscript3𝑚subscript𝔽superscript32𝑚subscript𝔽superscript36𝑚\mathbb{F}_{3^{m}}\subset\mathbb{F}_{3^{2m}}\subset\mathbb{F}_{3^{6m}}</td>
<td></td>
</tr>
</tbody>
</table>
and the inherent parallelism of multiplication in extension fields is used to accelerate the operations. Hardware designs and especially FPGA-based ones are suitable platforms for parallel implementation of algorithms. In that work multiplications in the first and the second field extensions are computed via $`3`$ and $`6`$ multiplications in the ground fields, respectively, requiring $`18`$ multiplications in $`{\mathbb{F}}_{3^{97}}`$.
In our current work, which is mostly based on \[[7](#bib.bib7)\], on the one hand, we use asymptotically fast methods to improve the performance of multiplication in $`{\mathbb{F}}_{3^{97}}`$, and on the other hand, we propose new multiplication formulas to speedup multiplication in $`{\mathbb{F}}_{3^{6 \cdot 97}}`$. Using the new formulas, multiplication in $`{\mathbb{F}}_{3^{6 \cdot 97}}`$ is done with only $`15`$ multiplications instead of $`18`$. We use the same extension tower, using $`3`$ multiplications in $`{\mathbb{F}}_{3^{97}}`$ to multiply elements in $`{\mathbb{F}}_{3^{2 \cdot 97}}`$, but only $`5`$ multiplications in $`{\mathbb{F}}_{3^{2 \cdot 97}}`$ for $`{\mathbb{F}}_{3^{6 \cdot 97}}`$. Our proposed method has a slightly increased number of additions in comparison to the Karatsuba method. Notice however that a multiplication in $`{\mathbb{F}}_{3^{97}}`$ requires many more resources than an addition, therefore the overall resource consumption will be reduced. The details of our method to generate the new formulas have been omitted to limit the complexity and diversity of materials in this paper, and have been submitted as another paper for CHES 2007.
A consistent amount of work has been done on hardware-based multiplication in finite fields, especially those of characteristic $`3`$. The authors of \[[8](#bib.bib8)\] propose a least significant digit-element (LSDE) multiplier for $`{\mathbb{F}}_{3^{m}}`$. This multiplier divides the input polynomials into digits of length D. Whereas the digits of one input polynomial are processed in parallel, the digits of the other input polynomial are handled serially. Then the result is reduced modulo the irreducible polynomial. The same structure has also been used in \[[7](#bib.bib7)\] for multiplication in $`{\mathbb{F}}_{3^{97}}`$. Our multiplier, on the other hand, is based on the digit-serial implementation of LFSR (Linear Feedback Shift Register) multiplier which is widely used in the literature (see \[[9](#bib.bib9)\] or \[[10](#bib.bib10)\]), and performs the modular reduction during the multiplication. The first contribution of our current work is the application of the Karatsuba multiplier inside the digit-multipliers, which results in smaller area for these multipliers. Our results demonstrate the efficiency of this design compared to other works. The second contribution is the application of a method using only $`5`$ multiplications in $`{\mathbb{F}}_{3^{2 \cdot 97}}`$ for multiplication in $`{\mathbb{F}}_{3^{6 \cdot 97}}`$. This results in an area-saving of almost $`17\%`$ compared to the Karatsuba method which is used in \[[7](#bib.bib7)\].
Our work is organized as follows. Section [2](#S2 "2 Multiplication in 𝔽_3⁹⁷ ‣ Efficient FPGA-based multipliers for 𝔽_3⁹⁷ and 𝔽_3^{6⋅97}") is devoted to the general structure of our multiplier for $`{\mathbb{F}}_{3^{97}}`$. In Section [3](#S3 "3 The Karatsuba method ‣ Efficient FPGA-based multipliers for 𝔽_3⁹⁷ and 𝔽_3^{6⋅97}") we describe some improvements on the traditional LFSR multiplier and compare our results with other works from the literature. In Section [4](#S4 "4 Multiplication in 𝔽_3^{6⋅97} ‣ Efficient FPGA-based multipliers for 𝔽_3⁹⁷ and 𝔽_3^{6⋅97}") the new formulas for $`{\mathbb{F}}_{3^{6 \cdot 97}}`$ together with suggestions for a new multiplier are presented, and Section [5](#S5 "5 Conclusion ‣ Efficient FPGA-based multipliers for 𝔽_3⁹⁷ and 𝔽_3^{6⋅97}") concludes the paper.
## 2 Multiplication in $`{\mathbb{F}}_{3^{97}}`$
The finite field $`{\mathbb{F}}_{3^{97}}`$ can be represented as a vector space over $`{\mathbb{F}}_{3}`$. In this representation, elements of $`{\mathbb{F}}_{3^{97}}`$ are vectors of length $`97`$ over $`{\mathbb{F}}_{3}`$. Addition of elements is computed by adding corresponding vectors. Multiplication is more complicated, and depends on the selected basis for $`{\mathbb{F}}_{3^{97}}`$. There are two popular bases which are used often in the literature, namely polynomial and normal bases. A polynomial basis is generally more suitable for multiplication, hence we choose this basis in our work.
In the polynomial basis, elements of $`{\mathbb{F}}_{3^{97}}`$ are represented as polynomials of degree at most $`96`$ over $`{\mathbb{F}}_{3}`$. Two elements are added by adding of the corresponding polynomials. Multiplication is based on polynomial multiplication followed by reduction modulo the irreducible polynomial, which generates the polynomial basis. In our case the irreducible polynomial, which we denote by $`f\hspace{0pt}{(x)}`$, is
<table>
<tbody>
<tr>
<td></td>
<td>x97+x16+2.superscript𝑥97superscript𝑥162x^{97}+x^{16}+2.</td>
<td></td>
<td>(1)</td>
</tr>
</tbody>
</table>
In the next sections we show the details of polynomial arithmetic in our designs.
### 2.1 Arithmetic in $`{\mathbb{F}}_{3}`$
The element $`a \in {\mathbb{F}}_{3}`$ is represented using the vector $`(a_{1},a_{0})`$ of two bits such that the elements $`0`$, $`1`$, and $`2`$ are $`(0,0)`$, $`(0,1)`$, $`(1,0)`$, respectively. In this representation the operations addition, multiplication, and negation (multiplication by $`2`$) are done, as shown in \[[11](#bib.bib11)\], using Equations [2](#S2.E2 "In 2.1 Arithmetic in 𝔽₃ ‣ 2 Multiplication in 𝔽_3⁹⁷ ‣ Efficient FPGA-based multipliers for 𝔽_3⁹⁷ and 𝔽_3^{6⋅97}"), [3](#S2.E3 "In 2.1 Arithmetic in 𝔽₃ ‣ 2 Multiplication in 𝔽_3⁹⁷ ‣ Efficient FPGA-based multipliers for 𝔽_3⁹⁷ and 𝔽_3^{6⋅97}"), and [4](#S2.E4 "In 2.1 Arithmetic in 𝔽₃ ‣ 2 Multiplication in 𝔽_3⁹⁷ ‣ Efficient FPGA-based multipliers for 𝔽_3⁹⁷ and 𝔽_3^{6⋅97}"), respectively.
<table>
<tbody>
<tr>
<td></td>
<td></td>
<td>(a1,a0)+(b1,b0)=((a0∨b0)⊕t,(a1∨b1)⊕t),subscript𝑎1subscript𝑎0subscript𝑏1subscript𝑏0direct-sumsubscript𝑎0subscript𝑏0𝑡direct-sumsubscript𝑎1subscript𝑏1𝑡\displaystyle(a_{1},a_{0})+(b_{1},b_{0})=((a_{0}\vee b_{0})\oplus t,(a_{1}\vee b_{1})\oplus t),</td>
<td></td>
<td>(2)</td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td></td>
<td>where t=(a0∨b1)⊕(a1∨b0)where 𝑡direct-sumsubscript𝑎0subscript𝑏1subscript𝑎1subscript𝑏0\displaystyle\text{where }t=(a_{0}\vee b_{1})\oplus(a_{1}\vee b_{0})</td>
<td></td>
<td></td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td></td>
<td>(a1,a0)⋅(b1,b0)=((a1∧b0)∨(a0∧b1),\displaystyle(a_{1},a_{0})\cdot(b_{1},b_{0})=((a_{1}\wedge b_{0})\vee(a_{0}\wedge b_{1}),</td>
<td></td>
<td>(3)</td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td></td>
<td>(a0∧b0)∨(a1∧b1),subscript𝑎0subscript𝑏0subscript𝑎1subscript𝑏1\displaystyle(a_{0}\wedge b_{0})\vee(a_{1}\wedge b_{1}),</td>
<td></td>
<td></td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td></td>
<td>−(a1,a0)=(a0,a1).subscript𝑎1subscript𝑎0subscript𝑎0subscript𝑎1\displaystyle-(a_{1},a_{0})=(a_{0},a_{1}).</td>
<td></td>
<td>(4)</td>
</tr>
</tbody>
</table>
The implementation of Equations [2](#S2.E2 "In 2.1 Arithmetic in 𝔽₃ ‣ 2 Multiplication in 𝔽_3⁹⁷ ‣ Efficient FPGA-based multipliers for 𝔽_3⁹⁷ and 𝔽_3^{6⋅97}") and [3](#S2.E3 "In 2.1 Arithmetic in 𝔽₃ ‣ 2 Multiplication in 𝔽_3⁹⁷ ‣ Efficient FPGA-based multipliers for 𝔽_3⁹⁷ and 𝔽_3^{6⋅97}") is done using 2 LUTs in the FPGA, whereas ([4](#S2.E4 "In 2.1 Arithmetic in 𝔽₃ ‣ 2 Multiplication in 𝔽_3⁹⁷ ‣ Efficient FPGA-based multipliers for 𝔽_3⁹⁷ and 𝔽_3^{6⋅97}")) is only a permutation of bits.
### 2.2 Structure of the multiplier for $`{\mathbb{F}}_{3^{97}}`$
The structure of a digit-level LFSR multiplier is shown in Figure [1](#S2.F1 "Figure 1 ‣ 2.2 Structure of the multiplier for 𝔽_3⁹⁷ ‣ 2 Multiplication in 𝔽_3⁹⁷ ‣ Efficient FPGA-based multipliers for 𝔽_3⁹⁷ and 𝔽_3^{6⋅97}"). In this figure the two input polynomials $`a\hspace{0pt}{(x)}`$, and $`b\hspace{0pt}{(x)}`$ are loaded into registers $`A`$ and $`B`$, respectively, and divided into digits of length $`D`$. In each clock cycle the most significant digit of $`B`$ is multiplied by the words of $`A`$, through digit-multipliers specified by M, and added to the content of the register in the feedback circuit. Inputs to the digit multipliers are two polynomials of degree $`D - 1`$ in $`x`$. The product is a polynomial of degree $`2\hspace{0pt}{({D - 1})}`$. Powers $`x^{D}`$ to $`x^{2\hspace{0pt}{({D - 1})}}`$ of each multiplier must be added to the powers $`x^{0}`$ to $`x^{D - 2}`$ of the next multiplier. This is done by the overlap circuit. In each clock cycle the register $`B`$ and LFSR are shifted by $`D`$ bits to the right. Shifting LFSR to right is equivalent to multiplication by $`x^{D}`$ which generates the powers $`x^{97}`$ to $`x^{96 + D}`$. These powers are reduced modulo $`f\hspace{0pt}{(x)}`$ of ([1](#S2.E1 "In 2 Multiplication in 𝔽_3⁹⁷ ‣ Efficient FPGA-based multipliers for 𝔽_3⁹⁷ and 𝔽_3^{6⋅97}")) using the feedback circuit. The name Linear Feedback Shift Register descends from these feedback structures. For more information about the digit-level LFSR multiplier and its costs for classical methods see \[[10](#bib.bib10)\]. In the next section we discuss our improvements to the traditional LFSR multiplier.
[Refer to caption](/html/0708.3022/assets/x1.png)
Fig. 1:
Structure of a digit-level LFSR multiplier
## 3 The Karatsuba method
In this section we use asymptotically fast methods to reduce the size of digit-multipliers. We use a similar approach to \[[12](#bib.bib12)\] and combine the classical and the Karatsuba methods to build small digit-multipliers. Two linear polynomials $`{a_{1}\hspace{0pt}x} + a_{0}`$ and $`{b_{1}\hspace{0pt}x} + b_{0}`$ are multiplied classically using the formula
<table>
<tbody>
<tr>
<td></td>
<td>a1b1x2+(a1b0+a0b1)x+a0b0subscript𝑎1subscript𝑏1superscript𝑥2subscript𝑎1subscript𝑏0subscript𝑎0subscript𝑏1𝑥subscript𝑎0subscript𝑏0a_{1}b_{1}x^{2}+(a_{1}b_{0}+a_{0}b_{1})x+a_{0}b_{0}</td>
<td></td>
<td>(5)</td>
</tr>
</tbody>
</table>
with $`4`$ multiplications and $`1`$ addition. The same product can also be computed via
<table>
<tbody>
<tr>
<td></td>
<td>a1b1x2+((a1+a0)(b1+b0)−a1b1−a0b0)x+a0b0.subscript𝑎1subscript𝑏1superscript𝑥2subscript𝑎1subscript𝑎0subscript𝑏1subscript𝑏0subscript𝑎1subscript𝑏1subscript𝑎0subscript𝑏0𝑥subscript𝑎0subscript𝑏0a_{1}b_{1}x^{2}+((a_{1}+a_{0})(b_{1}+b_{0})-a_{1}b_{1}-a_{0}b_{0})x+a_{0}b_{0}.</td>
<td></td>
<td>(6)</td>
</tr>
</tbody>
</table>
The new formula is called the Karatsuba method (see \[[13](#bib.bib13)\]). It requires $`7`$ operations instead of $`5`$, but only $`3`$ multiplications, and uses fewer resources when the coefficients $`{a_{0},a_{1}},`$ $`b_{0},b_{1}`$ are replaced by polynomials. The classical method for multiplication of two polynomials of degree $`n - 1`$ requires $`O\hspace{0pt}{(n^{2})}`$ operations. Recursive application of the Karatsuba method reduces the cost of a multiplication to $`O\hspace{0pt}{(n^{1.59})}`$ operations. We represent the classical multiplication of two polynomials of degree $`n - 1`$ by $`\mathcal{C}_{n}`$ and the method of ([6](#S3.E6 "In 3 The Karatsuba method ‣ Efficient FPGA-based multipliers for 𝔽_3⁹⁷ and 𝔽_3^{6⋅97}")) by $`\mathcal{K}`$. The methods $`\mathcal{C}_{n}`$ for $`n \in {\mathbb{N}}`$, and $`\mathcal{K}`$ constitute a set of polynomial multiplication methods. We call this set $`\mathsf{T}`$. Using the elements of $`\mathsf{T}`$ we define the set of recursive multiplication methods $`\mathsf{T}^{\ast}`$ which contains the elements of $`\mathsf{T}`$ and all recursive combinations of elements of $`\mathsf{T}^{\ast}`$. The recursive combination of the two methods $`\mathcal{M}`$ and $`\mathcal{N}`$, for polynomials of lengths $`m`$ and $`n`$, respectively, is the multiplication method $`\mathcal{M}\hspace{0pt}\mathcal{N}`$ for polynomials of length $`m\hspace{0pt}n`$. Let
<table>
<tbody>
<tr>
<td></td>
<td>a(x)=amn−1xmn−1+⋯+a0, and𝑎𝑥subscript𝑎𝑚𝑛1superscript𝑥𝑚𝑛1⋯subscript𝑎0 and\displaystyle a(x)=a_{mn-1}x^{mn-1}+\cdots+a_{0},\text{ and}</td>
<td></td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td>b(x)=bmn−1xmn−1+⋯+b0𝑏𝑥subscript𝑏𝑚𝑛1superscript𝑥𝑚𝑛1⋯subscript𝑏0\displaystyle b(x)=b_{mn-1}x^{mn-1}+\cdots+b_{0}</td>
<td></td>
</tr>
</tbody>
</table>
be given polynomials. In order to apply $`\mathcal{M}\hspace{0pt}\mathcal{N}`$, we write these polynomials as
<table>
<tbody>
<tr>
<td></td>
<td>a(x)=Am−1Xm−1+⋯+A0, and𝑎𝑥subscript𝐴𝑚1superscript𝑋𝑚1⋯subscript𝐴0 and\displaystyle a(x)=A_{m-1}X^{m-1}+\cdots+A_{0},\text{ and}</td>
<td></td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td>b(x)=Bm−1Xm−1+⋯+B0,𝑏𝑥subscript𝐵𝑚1superscript𝑋𝑚1⋯subscript𝐵0\displaystyle b(x)=B_{m-1}X^{m-1}+\cdots+B_{0},</td>
<td></td>
</tr>
</tbody>
</table>
where $`X = x^{n}`$ and $`A_{0},{\cdots\hspace{0pt}A_{m - 1}},B_{0},{\cdots\hspace{0pt}B_{m - 1}}`$ are polynomials of degree $`n - 1`$. If the polynomials $`A_{i}`$ and $`B_{i}`$ were coefficients, the two polynomials $`a\hspace{0pt}{(x)}`$ and $`b\hspace{0pt}{(x)}`$ would be multiplied using $`\mathcal{M}`$. The product using the method $`\mathcal{M}\hspace{0pt}\mathcal{N}`$ consists of several multiplications of the polynomials $`A_{i}`$ and $`B_{i}`$, which are performed using $`\mathcal{N}`$. We implement the digit-multipliers using the elements of $`\mathsf{T}^{\ast}`$ to reduce their size. Our approach is similar to \[[12](#bib.bib12)\].
In Table [1](#S3.T1 "Table 1 ‣ 3 The Karatsuba method ‣ Efficient FPGA-based multipliers for 𝔽_3⁹⁷ and 𝔽_3^{6⋅97}") we show the results of implementing $`{\mathbb{F}}_{3^{97}}`$ multipliers on a XC2VP20-6FF896 FPGA. In this table the first column is the digit-size $`D`$. In a digit-level multiplier with digit-size $`D`$, inputs are preceded by enough zeros so that their length becomes a multiple of $`D`$. Hence it is natural to choose a value of $`D`$ such that the difference $`{\lceil{m/D}\rceil} - {m/D}`$ is as small as possible. Our values for $`D`$ are selected using this criteria and hence differ from other standard values like multiples of $`4`$ in other works (see \[[8](#bib.bib8)\] and \[[7](#bib.bib7)\]). The string in the second column shows the recursive combination of the Karatsuba and classical methods which is applied. It is important to notice that the method $`\mathcal{K}\hspace{0pt}\mathcal{C}_{2}`$, which we used for polynomials of degree $`6`$, applies to polynomials of length $`7`$. Therefore, we add a zero in front of the polynomial and then remove all the gates containing an operation with the coefficients which are known to be zero. Hence this multiplier requires fewer resources than a complete $`\mathcal{K}\hspace{0pt}\mathcal{C}_{2}`$. This point distinguishes our approach from that in \[[12](#bib.bib12)\]. In the third, fourth, and fifth columns are the number of slices, maximum working frequency of the multiplier, and the required clock cycles for our designs.
Table 1:
Timing and area costs of digit-level LFSR multipliers in $`{\mathbb{F}}_{3^{97}}`$ for different values of digit-size $`D`$
<table>
<thead>
<tr>
<th>D𝐷D</th>
<th>Multiplication</th>
<th># of slices</th>
<th>Maximum</th>
<th># of clock</th>
</tr>
<tr>
<th></th>
<th></th>
<th></th>
<th>frequency (MHz)</th>
<th>cycles = ⌈97/D⌉97𝐷\lceil 97/D\rceil</th>
</tr>
</thead>
<tbody>
<tr>
<td>111</td>
<td>−-</td>
<td>327327327</td>
<td>300300300</td>
<td>979797</td>
</tr>
<tr>
<td>222</td>
<td>𝒞2subscript𝒞2\mathcal{C}_{2}</td>
<td>800800800</td>
<td>174174174</td>
<td>494949</td>
</tr>
<tr>
<td>444</td>
<td>𝒞4subscript𝒞4\mathcal{C}_{4}</td>
<td>171617161716</td>
<td>125125125</td>
<td>252525</td>
</tr>
<tr>
<td>777</td>
<td>𝒦𝒞4𝒦subscript𝒞4\mathcal{K}\mathcal{C}_{4}</td>
<td>295429542954</td>
<td>111111111</td>
<td>141414</td>
</tr>
<tr>
<td>141414</td>
<td>𝒦𝒦𝒞4𝒦𝒦subscript𝒞4\mathcal{K}\mathcal{K}\mathcal{C}_{4}</td>
<td>400640064006</td>
<td>727272</td>
<td>777</td>
</tr>
</tbody>
</table>
The results of comparing our results with those in \[[7](#bib.bib7)\] are shown in Figure [2](#S3.F2 "Figure 2 ‣ 3 The Karatsuba method ‣ Efficient FPGA-based multipliers for 𝔽_3⁹⁷ and 𝔽_3^{6⋅97}"). Different digit-levels result in different circuits, which we compare with respect to both time and area. Area is the number of slices, whereas time is the product of clock cycles and minimum period. Both designs are on the same technology, but the speed grade of the FPGA in \[[7](#bib.bib7)\] is not available. As it is shown, our designs have better area-time performance. These improvements result, on the one hand, by using asymptotically faster methods, and on the other hand, by integrating the modular reduction stage into the LFSR. When a small digit-serial multiplier is used even the small size of a modular reduction must be taken into account.
[img.svg](data:image...)
Fig. 2:
Time vs. area comparisons of our multipliers with those in \[[7](#bib.bib7)\]
## 4 Multiplication in $`{\mathbb{F}}_{3^{6 \cdot 97}}`$
Multiplication in $`{\mathbb{F}}_{3^{6 \cdot 97}}`$ is done in the same way as in \[[7](#bib.bib7)\], as a tower of extensions of degrees $`2`$ and $`3`$, i.e.
<table>
<tbody>
<tr>
<td></td>
<td>𝔽397≅𝔽3/(x97+x16+2)𝔽32⋅97≅𝔽397/(y2+1)𝔽36⋅97≅𝔽32⋅97/(z3−z−1).subscript𝔽superscript397subscript𝔽3superscript𝑥97superscript𝑥162subscript𝔽superscript3⋅297subscript𝔽superscript397superscript𝑦21subscript𝔽superscript3⋅697subscript𝔽superscript3⋅297superscript𝑧3𝑧1\begin{array}[]{lcl}\mathbb{F}_{3^{97}}&\cong&\mathbb{F}_{3}/(x^{97}+x^{16}+2)\\ \mathbb{F}_{3^{2\cdot 97}}&\cong&\mathbb{F}_{3^{97}}/(y^{2}+1)\\ \mathbb{F}_{3^{6\cdot 97}}&\cong&\mathbb{F}_{3^{2\cdot 97}}/(z^{3}-z-1).\\ \end{array}</td>
<td></td>
</tr>
</tbody>
</table>
The elements of $`{\mathbb{F}}_{3^{2 \cdot 97}}`$ are polynomials of degree $`1`$ in $`s`$ over $`{\mathbb{F}}_{3^{97}}`$, for $`s`$ a root of $`y^{2} + 1`$ in $`{\mathbb{F}}_{3^{2 \cdot 97}}`$. The polynomials are multiplied by applying ([6](#S3.E6 "In 3 The Karatsuba method ‣ Efficient FPGA-based multipliers for 𝔽_3⁹⁷ and 𝔽_3^{6⋅97}")) and then reduced modulo $`s^{2} + 1`$. The elements of $`{\mathbb{F}}_{3^{6 \cdot 97}}`$ are polynomials of degree $`3`$ in $`r`$, a root of $`z^{3} - z - 1`$ in $`{\mathbb{F}}_{3^{6 \cdot 97}}`$. They are multiplied using the formulas ([7](#S4.E7 "In 4 Multiplication in 𝔽_3^{6⋅97} ‣ Efficient FPGA-based multipliers for 𝔽_3⁹⁷ and 𝔽_3^{6⋅97}")) and then reduced modulo $`r^{3} - r - 1`$.
<table>
<tbody>
<tr>
<td></td>
<td>(a0+a1r+a2r2)(b0+b1r+b2r2)=c0+c1r+c2r2+c3r3+c4r4,whereP0=(a0+a1+a2)(b0+b1+b2)P1=(a0+sa1−a2)(b0+sb1−b2)P2=(a0−a1+a2)(b0−b1+b2)P3=(a0−sa1−a2)(b0−sb1−b2)P4=a2b2, andc0=P0+P1+P2+P3−P4c1=P0−sP1−P2+sP3c2=P0−P1+P2−P3c3=P0+sP1−P2−sP3c4=P4,subscript𝑎0subscript𝑎1𝑟subscript𝑎2superscript𝑟2subscript𝑏0subscript𝑏1𝑟subscript𝑏2superscript𝑟2absentsubscript𝑐0subscript𝑐1𝑟subscript𝑐2superscript𝑟2subscript𝑐3superscript𝑟3subscript𝑐4superscript𝑟4wheresubscript𝑃0subscript𝑎0subscript𝑎1subscript𝑎2subscript𝑏0subscript𝑏1subscript𝑏2subscript𝑃1subscript𝑎0𝑠subscript𝑎1subscript𝑎2subscript𝑏0𝑠subscript𝑏1subscript𝑏2subscript𝑃2subscript𝑎0subscript𝑎1subscript𝑎2subscript𝑏0subscript𝑏1subscript𝑏2subscript𝑃3subscript𝑎0𝑠subscript𝑎1subscript𝑎2subscript𝑏0𝑠subscript𝑏1subscript𝑏2subscript𝑃4subscript𝑎2subscript𝑏2 andsubscript𝑐0subscript𝑃0subscript𝑃1subscript𝑃2subscript𝑃3subscript𝑃4subscript𝑐1subscript𝑃0𝑠subscript𝑃1subscript𝑃2𝑠subscript𝑃3subscript𝑐2subscript𝑃0subscript𝑃1subscript𝑃2subscript𝑃3subscript𝑐3subscript𝑃0𝑠subscript𝑃1subscript𝑃2𝑠subscript𝑃3subscript𝑐4subscript𝑃4\begin{array}[]{l}(a_{0}+a_{1}r+a_{2}r^{2})(b_{0}+b_{1}r+b_{2}r^{2})=\\ c_{0}+c_{1}r+c_{2}r^{2}+c_{3}r^{3}+c_{4}r^{4},\text{where}\\ P_{0}=(a_{0}+a_{1}+a_{2})(b_{0}+b_{1}+b_{2})\\ P_{1}=(a_{0}+sa_{1}-a_{2})(b_{0}+sb_{1}-b_{2})\\ P_{2}=(a_{0}-a_{1}+a_{2})(b_{0}-b_{1}+b_{2})\\ P_{3}=(a_{0}-sa_{1}-a_{2})(b_{0}-sb_{1}-b_{2})\\ P_{4}=a_{2}b_{2},\text{ and}\\ c_{0}=P_{0}+P_{1}+P_{2}+P_{3}-P_{4}\\ c_{1}=P_{0}-sP_{1}-P_{2}+sP_{3}\\ c_{2}=P_{0}-P_{1}+P_{2}-P_{3}\\ c_{3}=P_{0}+sP_{1}-P_{2}-sP_{3}\\ c_{4}=P_{4},\end{array}</td>
<td></td>
<td>(7)</td>
</tr>
</tbody>
</table>
Combining ([6](#S3.E6 "In 3 The Karatsuba method ‣ Efficient FPGA-based multipliers for 𝔽_3⁹⁷ and 𝔽_3^{6⋅97}")), ([7](#S4.E7 "In 4 Multiplication in 𝔽_3^{6⋅97} ‣ Efficient FPGA-based multipliers for 𝔽_3⁹⁷ and 𝔽_3^{6⋅97}")) we have the following theorem.
######
Theorem 1
Let $`{\alpha,\beta} \in {\mathbb{F}}_{3^{6 \cdot 97}}`$ be given as:
<table>
<tbody>
<tr>
<td></td>
<td>α=𝛼absent\displaystyle\alpha=</td>
<td>a0+a1s+a2r+a3rs+a4r2+a5r2ssubscript𝑎0subscript𝑎1𝑠subscript𝑎2𝑟subscript𝑎3𝑟𝑠subscript𝑎4superscript𝑟2subscript𝑎5superscript𝑟2𝑠\displaystyle a_{0}+a_{1}s+a_{2}r+a_{3}rs+a_{4}r^{2}+a_{5}r^{2}s</td>
<td></td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td>β=𝛽absent\displaystyle\beta=</td>
<td>b0+b1s+b2r+b3rs+b4r2+b5r2s.subscript𝑏0subscript𝑏1𝑠subscript𝑏2𝑟subscript𝑏3𝑟𝑠subscript𝑏4superscript𝑟2subscript𝑏5superscript𝑟2𝑠\displaystyle b_{0}+b_{1}s+b_{2}r+b_{3}rs+b_{4}r^{2}+b_{5}r^{2}s.</td>
<td></td>
</tr>
</tbody>
</table>
Let further their product $`\gamma = {\alpha\hspace{0pt}\beta} \in {\mathbb{F}}_{3^{6 \cdot 97}}`$ be
<table>
<tbody>
<tr>
<td></td>
<td>γ=c0+c1s+c2r+c3rs+c4r2+c5r2s.𝛾subscript𝑐0subscript𝑐1𝑠subscript𝑐2𝑟subscript𝑐3𝑟𝑠subscript𝑐4superscript𝑟2subscript𝑐5superscript𝑟2𝑠\displaystyle\gamma=c_{0}+c_{1}s+c_{2}r+c_{3}rs+c_{4}r^{2}+c_{5}r^{2}s.</td>
<td></td>
</tr>
</tbody>
</table>
Then the coefficients $`c_{0}\hspace{0pt}\cdots\hspace{0pt}c_{5}`$ of the product can be computed using only $`15`$ multiplications in $`{\mathbb{F}}_{3^{97}}`$.
Closed-form formulas for this multiplication are shown in Appendix [A](#A1 "Appendix A Multiplication formulas for 𝔽_3^{6⋅97} ‣ Efficient FPGA-based multipliers for 𝔽_3⁹⁷ and 𝔽_3^{6⋅97}"). Scalar multiplications are particularly simple using these formulas. Scalar multiplications are multiplications by $`- 1`$, $`s`$, and $`- s`$. Negation of coefficients and consequently of polynomials is only a permutation of bits, as seen in Section [2](#S2 "2 Multiplication in 𝔽_3⁹⁷ ‣ Efficient FPGA-based multipliers for 𝔽_3⁹⁷ and 𝔽_3^{6⋅97}"). Indeed multiplication of an element in $`{\mathbb{F}}_{3^{2 \cdot 97}}`$ by $`s`$ is a permutation, too. Let $`\alpha = {{a_{1}\hspace{0pt}s} + a_{0}} \in {\mathbb{F}}_{3^{2 \cdot 97}}`$, then
<table>
<tbody>
<tr>
<td></td>
<td>sα=a1s2+a0smods2+1=a0s−a1.𝑠𝛼modulosubscript𝑎1superscript𝑠2subscript𝑎0𝑠superscript𝑠21subscript𝑎0𝑠subscript𝑎1s\alpha=a_{1}s^{2}+a_{0}s\mod s^{2}+1=a_{0}s-a_{1}.</td>
<td></td>
</tr>
</tbody>
</table>
All of the $`{\mathbb{F}}_{3^{97}}`$-multiplications can be done in parallel. This property allows designers to implement as many of these multipliers as possible, according to their time-area constraints. On the other hand, these multipliers are used for other computations such as point addition and doubling on elliptic curves for pairing-based cryptography. Reading and writing intermediate values into register files in such applications is time-consuming. To solve this problem we propose a new multiplier which is shown in Figure [3](#S4.F3 "Figure 3 ‣ 4 Multiplication in 𝔽_3^{6⋅97} ‣ Efficient FPGA-based multipliers for 𝔽_3⁹⁷ and 𝔽_3^{6⋅97}"). The new multiplier consists of three pipeline stages, namely, input, multiplication, and output. During the time of each multiplication in $`{\mathbb{F}}_{3^{97}}`$, the input stage loads the coefficients $`a_{i}`$ and $`b_{i}`$ from memory for the next multiplication, and computes the linear combinations in ([8](#A1.E8 "In Appendix A Multiplication formulas for 𝔽_3^{6⋅97} ‣ Efficient FPGA-based multipliers for 𝔽_3⁹⁷ and 𝔽_3^{6⋅97}")) to compute $`P_{i}`$s. In this time the output stage adds the last computed product $`P_{i}`$ to memory variables according to ([9](#A1.E9 "In Appendix A Multiplication formulas for 𝔽_3^{6⋅97} ‣ Efficient FPGA-based multipliers for 𝔽_3⁹⁷ and 𝔽_3^{6⋅97}")). In this structure the hatched multiplexers can select either one of their inputs or the sum of the inputs. In this way all possible multiples of input polynomials can be selected and added to the accumulators.
[img.svg](data:image...)
Fig. 3:
The proposed structure block for implementing the formulas of Appendix [A](#A1 "Appendix A Multiplication formulas for 𝔽_3^{6⋅97} ‣ Efficient FPGA-based multipliers for 𝔽_3⁹⁷ and 𝔽_3^{6⋅97}")
## 5 Conclusion
In this paper we proposed a new structure for multiplication in $`{\mathbb{F}}_{3^{97}}`$. This structure is based on digit-level LFSR multipliers, where the area of digit-multipliers are reduced using the Karatsuba method. Another advantage of this approach is performing the modular reduction during the multiplication. Our synthesis results showed the performance improvement compared to other designs in the literature. We have also presented new formulas for multiplication in $`{\mathbb{F}}_{3^{6 \cdot 97}}`$ using only $`15`$ multiplications in $`{\mathbb{F}}_{3^{97}}`$. When the Karatsuba method is applied 18 multiplications are required. Furthermore, we have introduced a feasible hardware structure for realizing our proposed formulas. Our formulas are for the case that $`{\mathbb{F}}_{3^{6 \cdot 97}}`$ is constructed from $`{\mathbb{F}}_{3^{2 \cdot 97}}`$ using the irreducible polynomial $`z^{3} - z - 1`$. In case that the finite field is constructed using $`{z^{3} - z} + 1`$, the formulas require slight modifications.
## References
- \[1\]
D. E. Knuth, *The Art of Computer Programming, vol. 2, Seminumerical Algorithms*, 3rd ed. Reading MA: Addison-Wesley, 1998, first edition 1969.
- \[2\]
J. von zur Gathen and J. Gerhard, *Modern Computer Algebra*, 2nd ed. Cambridge, UK: Cambridge University Press, 2003, first edition 1999.
- \[3\]
*Digital Signature Standard (DSS)*, U.S. Department of Commerce / National Institute of Standards and Technology, January 2000, federal Information Processings Standards Publication 186-2. \[Online\]. Available: http://csrc.nist.gov/publications/fips/fips186-2/fips186-2.pdf
- \[4\]
D. V. Bailey and C. Paar, “Optimal extension fields for fast arithmetic in public-key algorithms,” in *Advances in Cryptology: Proceedings of CRYPTO ’98, Santa Barbara CA*, ser. Lecture Notes in Computer Science, H. Krawczyk, Ed., no. 1462. Springer-Verlag, 1998, pp. 472–485.
- \[5\]
R. M. Avanzi and P. Mihăilescu, “Generic efficient arithmetic algorithms for PAFFs (processor adequate finite fields) and related algebraic structures (extended abstract),” in *Selected Areas in Cryptography (SAC 2003)*. Springer-Verlag, 2003, pp. 320–334.
- \[6\]
I. Duursma and H. Lee, “Tate-pairing implementations for tripartite key agreement.” \[Online\]. Available: citeseer.ist.psu.edu/duursma03tatepairing.html
- \[7\]
T. Kerins, W. P. Marnane, E. M. Popovici, and P. S. L. M. Barreto, “Efficient hardware for the tate pairing calculation in characteristic three,” in *Cryptographic Hardware and Embedded Systems, CHES2005*, ser. Lecture Notes in Computer Science, J. R. Rao and B. Sunar, Eds., vol. 3659. Springer-Verlag, 2005, pp. 412–426.
- \[8\]
G. Bertoni, J. Guajardo, S. Kumar, G. Orlando, C. Paar, and T. Wollinger, “Efficient $`G\hspace{0pt}F\hspace{0pt}\left( p^{m} \right)`$ arithmetic architectures for cryptographic applications,” in *Topics in cryptology, CT-RSA 2003: the Cryptographers’ Track at the RSA Conference 2003, San Francisco, CA, USA, April 13–17, 2003: Proceedings*, ser. Lecture Notes in Computer Science, M. Joye, Ed., vol. 2612. Springer-Verlag, 2003, pp. 158–175.
- \[9\]
R. J. McEliece, *Finite Fields for Computer Scientists and Engineers*. Kluwer Academic Publishers, 1987.
- \[10\]
J. Shokrollahi, “Efficient implementation of elliptic curve cryptography on fpgas,” Ph.D. dissertation, Bonn University, Bonn, December 2006. \[Online\]. Available: http://hss.ulb.uni-bonn.de/diss_online/math_nat_fak/2007/shokrollahi_jamshid/ index.htm
- \[11\]
R. Granger, D. Page, and M. Stam, “Hardware and software normal basis arithmetic for pairing-based cryptogaphy in characteristic three,” *IEEE Transactions on Computers*, vol. 54, no. 7, pp. 852–860, 2005. \[Online\]. Available: http://
- \[12\]
J. von zur Gathen and J. Shokrollahi, “Fast arithmetic for polynomials over $`{\mathbb{F}}_{2}`$ in hardware,” in *IEEE Information Theory Workshop (2006)*. Punta del Este, Uruguay: IEEE, March 2006, pp. 107–111.
- \[13\]
A. Karatsuba and Y. Ofman, “Multiplication of multidigit numbers on automata,” *Soviet Physics–Doklady*, vol. 7, no. 7, pp. 595–596, January 1963, translated from Doklady Akademii Nauk SSSR, Vol. 145, No. 2, pp. 293–294, July, 1962.
## Appendix A Multiplication formulas for $`{\mathbb{F}}_{3^{6 \cdot 97}}`$
Let $`{\alpha,\beta} \in {\mathbb{F}}_{3^{6 \cdot 97}}`$ be given as:
<table>
<tbody>
<tr>
<td></td>
<td>α=𝛼absent\displaystyle\alpha=</td>
<td>a0+a1s+a2r+a3rs+a4r2+a5r2s,subscript𝑎0subscript𝑎1𝑠subscript𝑎2𝑟subscript𝑎3𝑟𝑠subscript𝑎4superscript𝑟2subscript𝑎5superscript𝑟2𝑠\displaystyle a_{0}+a_{1}s+a_{2}r+a_{3}rs+a_{4}r^{2}+a_{5}r^{2}s,</td>
<td></td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td>β=𝛽absent\displaystyle\beta=</td>
<td>b0+b1s+b2r+b3rs+b4r2+b5r2s,subscript𝑏0subscript𝑏1𝑠subscript𝑏2𝑟subscript𝑏3𝑟𝑠subscript𝑏4superscript𝑟2subscript𝑏5superscript𝑟2𝑠\displaystyle b_{0}+b_{1}s+b_{2}r+b_{3}rs+b_{4}r^{2}+b_{5}r^{2}s,</td>
<td></td>
</tr>
</tbody>
</table>
where $`{a_{0},\cdots,b_{5}} \in {\mathbb{F}}_{3^{97}}`$ and $`s \in {\mathbb{F}}_{3^{2 \cdot 97}}`$, $`r \in {\mathbb{F}}_{3^{6 \cdot 97}}`$ are roots of $`y^{2} + 1`$ and $`z^{3} - z - 1`$, respectively. Let their product $`\gamma = {\alpha\hspace{0pt}\beta} \in {\mathbb{F}}_{3^{6 \cdot 97}}`$ be
<table>
<tbody>
<tr>
<td></td>
<td>γ=c0+c1s+c2r+c3rs+c4r2+c5r2s.𝛾subscript𝑐0subscript𝑐1𝑠subscript𝑐2𝑟subscript𝑐3𝑟𝑠subscript𝑐4superscript𝑟2subscript𝑐5superscript𝑟2𝑠\displaystyle\gamma=c_{0}+c_{1}s+c_{2}r+c_{3}rs+c_{4}r^{2}+c_{5}r^{2}s.</td>
<td></td>
</tr>
</tbody>
</table>
Then the coefficients $`{c_{0}\hspace{0pt}\cdots\hspace{0pt}c_{5}} \in {\mathbb{F}}_{3^{97}}`$ of the product can be computed using the following formulas.
<table>
<tbody>
<tr>
<td></td>
<td>P0=(a0+a2+a4)(b0+b2+b4)P1=(a0+a1+a2+a3+a4+a5)(b0+b1+b2+b3+b4+b5)P2=(a1+a3+a5)(b1+b3+b5)P3=(a0+sa2−a4)(b0+sb2−b4)P4=(a0+a1+sa2+sa3−a4−a5)(b0+b1+sb2+sb3−b4−b5)P5=(a1+sa3−a5)(b1+sb3−b5)P6=(a0−a2+a4)(b0−b2+b4)P7=(a0+a1−a2−a3+a4+a5)(b0+b1−b2−b3+b4+b5)P8=(a1−a3+a5)(b1−b3+b5)P9=(a0−sa2−a4)(b0−sb2−b4)P10=(a0+a1−sa2−sa3−a4−a5)(b0+b1−sb2−sb3−b4−b5)P11=(a1−sa3−a5)(b1−sb3−b5)P12=a4b4P13=(a4+a5)(b4+b5)P14=a5b5subscript𝑃0absentsubscript𝑎0subscript𝑎2subscript𝑎4𝑏0𝑏2𝑏4subscript𝑃1absentsubscript𝑎0subscript𝑎1subscript𝑎2subscript𝑎3subscript𝑎4subscript𝑎5missing-subexpressionsubscript𝑏0subscript𝑏1subscript𝑏2subscript𝑏3subscript𝑏4subscript𝑏5subscript𝑃2absentsubscript𝑎1subscript𝑎3subscript𝑎5subscript𝑏1subscript𝑏3subscript𝑏5subscript𝑃3absentsubscript𝑎0𝑠subscript𝑎2subscript𝑎4subscript𝑏0𝑠subscript𝑏2subscript𝑏4subscript𝑃4absentsubscript𝑎0subscript𝑎1𝑠subscript𝑎2𝑠subscript𝑎3subscript𝑎4subscript𝑎5missing-subexpressionsubscript𝑏0subscript𝑏1𝑠subscript𝑏2𝑠subscript𝑏3subscript𝑏4subscript𝑏5subscript𝑃5absentsubscript𝑎1𝑠subscript𝑎3subscript𝑎5subscript𝑏1𝑠subscript𝑏3subscript𝑏5subscript𝑃6absentsubscript𝑎0subscript𝑎2subscript𝑎4subscript𝑏0subscript𝑏2subscript𝑏4subscript𝑃7absentsubscript𝑎0subscript𝑎1subscript𝑎2subscript𝑎3subscript𝑎4subscript𝑎5missing-subexpressionsubscript𝑏0subscript𝑏1subscript𝑏2subscript𝑏3subscript𝑏4subscript𝑏5subscript𝑃8absentsubscript𝑎1subscript𝑎3subscript𝑎5subscript𝑏1subscript𝑏3subscript𝑏5subscript𝑃9absentsubscript𝑎0𝑠subscript𝑎2subscript𝑎4subscript𝑏0𝑠subscript𝑏2subscript𝑏4subscript𝑃10absentsubscript𝑎0subscript𝑎1𝑠subscript𝑎2𝑠subscript𝑎3subscript𝑎4subscript𝑎5missing-subexpressionsubscript𝑏0subscript𝑏1𝑠subscript𝑏2𝑠subscript𝑏3subscript𝑏4subscript𝑏5subscript𝑃11absentsubscript𝑎1𝑠subscript𝑎3subscript𝑎5subscript𝑏1𝑠subscript𝑏3subscript𝑏5subscript𝑃12absentsubscript𝑎4subscript𝑏4subscript𝑃13absentsubscript𝑎4subscript𝑎5subscript𝑏4subscript𝑏5subscript𝑃14absentsubscript𝑎5subscript𝑏5\begin{array}[]{ll}P_{0}=&(a_{0}+a_{2}+a_{4})(b0+b2+b4)\\ P_{1}=&(a_{0}+a_{1}+a_{2}+a_{3}+a_{4}+a_{5})\\ &(b_{0}+b_{1}+b_{2}+b_{3}+b_{4}+b_{5})\\ P_{2}=&(a_{1}+a_{3}+a_{5})(b_{1}+b_{3}+b_{5})\\ P_{3}=&(a_{0}+sa_{2}-a_{4})(b_{0}+sb_{2}-b_{4})\\ P_{4}=&(a_{0}+a_{1}+sa_{2}+sa_{3}-a_{4}-a_{5})\\ &(b_{0}+b_{1}+sb_{2}+sb_{3}-b_{4}-b_{5})\\ P_{5}=&(a_{1}+sa_{3}-a_{5})(b_{1}+sb_{3}-b_{5})\\ P_{6}=&(a_{0}-a_{2}+a_{4})(b_{0}-b_{2}+b_{4})\\ P_{7}=&(a_{0}+a_{1}-a_{2}-a_{3}+a_{4}+a_{5})\\ &(b_{0}+b_{1}-b_{2}-b_{3}+b_{4}+b_{5})\\ P_{8}=&(a_{1}-a_{3}+a_{5})(b_{1}-b_{3}+b_{5})\\ P_{9}=&(a_{0}-sa_{2}-a_{4})(b_{0}-sb_{2}-b_{4})\\ P_{10}=&(a_{0}+a_{1}-sa_{2}-sa_{3}-a_{4}-a_{5})\\ &(b_{0}+b_{1}-sb_{2}-sb_{3}-b_{4}-b_{5})\\ P_{11}=&(a_{1}-sa_{3}-a_{5})(b_{1}-sb_{3}-b_{5})\\ P_{12}=&a_{4}b_{4}\\ P_{13}=&(a_{4}+a_{5})(b_{4}+b_{5})\\ P_{14}=&a_{5}b_{5}\\ \end{array}</td>
<td></td>
<td>(8)</td>
</tr>
</tbody>
</table>
<table>
<tbody>
<tr>
<td></td>
<td>c0=−P0+P2+(s+1)P3−(s+1)P5−(s−1)P9+(s−1)P11−P12+P14c1=P0−P1+P2−(s+1)P3+(s+1)P4−(s+1)P5+(s−1)P9−(s−1)P10+(s−1)P11−P12−P13+P14c2=−P0+P2+P6−P8+P12−P14c3=P0−P1+P2−P6+P7−P8−P12+P13−P14c4=P0−P2−P3+P5+P6−P8−P9+P11+P12−P14c5=P0+P1−P2+P3−P4+P5−P6+P7−P8+P9−P10+P11−P12+P13−P14subscript𝑐0absentsubscript𝑃0subscript𝑃2𝑠1subscript𝑃3limit-from𝑠1subscript𝑃5missing-subexpression𝑠1subscript𝑃9𝑠1subscript𝑃11subscript𝑃12subscript𝑃14subscript𝑐1absentsubscript𝑃0subscript𝑃1subscript𝑃2𝑠1subscript𝑃3limit-from𝑠1subscript𝑃4missing-subexpression𝑠1subscript𝑃5𝑠1subscript𝑃9limit-from𝑠1subscript𝑃10missing-subexpression𝑠1subscript𝑃11subscript𝑃12subscript𝑃13subscript𝑃14subscript𝑐2absentsubscript𝑃0subscript𝑃2subscript𝑃6subscript𝑃8subscript𝑃12subscript𝑃14subscript𝑐3absentsubscript𝑃0subscript𝑃1subscript𝑃2subscript𝑃6subscript𝑃7subscript𝑃8subscript𝑃12missing-subexpressionsubscript𝑃13subscript𝑃14subscript𝑐4absentsubscript𝑃0subscript𝑃2subscript𝑃3subscript𝑃5subscript𝑃6subscript𝑃8subscript𝑃9limit-fromsubscript𝑃11missing-subexpressionsubscript𝑃12subscript𝑃14subscript𝑐5absentsubscript𝑃0subscript𝑃1subscript𝑃2subscript𝑃3subscript𝑃4subscript𝑃5subscript𝑃6limit-fromsubscript𝑃7missing-subexpressionsubscript𝑃8subscript𝑃9subscript𝑃10subscript𝑃11subscript𝑃12subscript𝑃13subscript𝑃14\begin{array}[]{ll}c_{0}=&-P_{0}+P_{2}+(s+1)P_{3}-(s+1)P_{5}-\\ &(s-1)P_{9}+(s-1)P_{11}-P_{12}+P_{14}\\ c_{1}=&P_{0}-P_{1}+P_{2}-(s+1)P_{3}+(s+1)P_{4}-\\ &(s+1)P_{5}+(s-1)P_{9}-(s-1)P_{10}+\\ &(s-1)P_{11}-P_{12}-P_{13}+P_{14}\\ c_{2}=&-P_{0}+P_{2}+P_{6}-P_{8}+P_{12}-P_{14}\\ c_{3}=&P_{0}-P_{1}+P_{2}-P_{6}+P_{7}-P_{8}-P_{12}\\ &+P_{13}-P_{14}\\ c_{4}=&P_{0}-P_{2}-P_{3}+P_{5}+P_{6}-P_{8}-P_{9}+P_{11}+\\ &P_{12}-P_{14}\\ c_{5}=&P_{0}+P_{1}-P_{2}+P_{3}-P_{4}+P_{5}-P_{6}+P_{7}-\\ &P_{8}+P_{9}-P_{10}+P_{11}-P_{12}+P_{13}-P_{14}\end{array}</td>
<td></td>
<td>(9)</td>
</tr>
</tbody>
</table> |
0710.1484 | cs.NI | cyber | # The structure and modeling results of the parallel spatial switching system
Denis Kutuzov ([d_kutuzov@mail.ru](d_kutuzov@mail.ru)), Alexey Osovsky
(April 20, 2007)
###### Abstract
Problems of the switching parallel system designing provided spatial switching of packets from random time are discussed. Results of modeling of switching system as systems of mass service are resulted.
Index Terms: parallel switching, information systems, system of mass service, packet switching, communication network.
Information Systems Dept., Astrakhan State University, 20a, Tatischeva Str., Astrakhan, 414056, Russia, (<a href="http://www.aspu.ru/" target="_blank">http://www.aspu.ru/</a>)
The increase of rate of switching system elements, as well as other devices occurs basically due to advantages in technological base. At the same time another decision is possible. It is perfection of algorithms and methods of switching and creation a new switching decisions, not dependent on technological base.
Such decision for switching systems may become a parallel process method which would allow to increase transfer rate and to decrease losses at processing packets.
The parallel switching system \[[1](#bib.bib1)\] provides a parallel spatial switching of the information packets at the random moments of time and consists of the following blocks (see Fig. [1](#S0.F1 "Figure 1 ‣ The structure and modeling results of the parallel spatial switching system")): block of synchronization of the input packet with the moment of identification (1), elements of switching (2) and blocks of tag generation for the target lines (3).
[Refer to caption](/html/0710.1484/assets/1.jpg)
Figure 1: The structure of the parallel spatial switching system
Operation of system is carried out with the following algorithm: a requirement of connection establishment is transferred on the system input. Each requirement represents the information packet having headline (tag) which identifies a target line of the switching system which it is necessary to send the given packet. As requirements of the establishment connections (packets) arrived on the system inputs at the random time, that is asynchronously, and identification v is possible only at the certain moments of time, the requirement (packet) should be delayed up to the moment when the tag comparison and to start switching is possible. Such delay is carried out by blocks of packet synchronization with the moment of identification and may be various for the packets which have arrived at die different time. Some requirements (packets), expecting the beginning of identification represent a pack.
Identification represents process of bit-by-bit tag comparison for connection lines (it contains in a packet), and the target line tags which generated blocks 3 (Fig. [1](#S0.F1 "Figure 1 ‣ The structure and modeling results of the parallel spatial switching system")). At coincidence of the target line tag with heading tag in the packet, the appropriate switching element establishes connection for transfer of packet on the system output. After transfer of the packet the connection should be destroyed. All functions of the conflict resolution, for example, at attempt of two and more input lines to establish connection with one target line are assigned to switching elements 2 (Fig. [1](#S0.F1 "Figure 1 ‣ The structure and modeling results of the parallel spatial switching system")).
Diagrams of loss probabilities of packets (P) from average (modeling) time between packet receiving are shown in Fig. [2](#S0.F2 "Figure 2 ‣ The structure and modeling results of the parallel spatial switching system").
[Refer to caption](/html/0710.1484/assets/2.jpg)
Figure 2: The modeling results of the parallel spatial switching system
It is possible to make a conclusion on reduction of loss probability in comparison with the system with consecutive switching by results of modeling of the parallel switching system \[[2](#bib.bib2)\]. The diagram 1 in Fig. [2](#S0.F2 "Figure 2 ‣ The structure and modeling results of the parallel spatial switching system") corresponds to parallel switching system, the diagram 2 is consecutive one.
## References
- \[1\] D.V. Kutuzov, “Parallel adjustment of matrix switching system at dynamical requirements” // Engineering and technology, 2005. No. 2(8), pp. 57–59.
- \[2\] A.P. Lunev, I.Yu. Petrova, D.V. Kutuzov, A.V. Osovsky, “Simulation of matrix switchboards”. The patent regisrtation No. 2005611003, the legal owner Astrakhan State University. |
0706.3305 | cs.CR | cyber | # A simple generalization of the ElGamal cryptosystem to non-abelian groups II
Ayan Mahalanobis
Indian Institute of Science Education and Research Pune, Pashan Pune-411021, India <ayanm@iiserpune.ac.in>
###### Abstract.
This is a study of the MOR cryptosystem using the special linear group over finite fields. The automorphism group of the special linear group is analyzed for this purpose. At our current state of knowledge, I show that this MOR cryptosystem has better security than the ElGamal cryptosystem over finite fields.
###### Key words and phrases:
MOR cryptosystem, special linear groups, the discrete logarithm problem
###### 2010 Mathematics Subject Classification:
94A60, 20G40
## 1. Introduction
The MOR cryptosystem is a generalization of the ElGamal cryptosystem, where the *discrete logarithm problem* works in the automorphism group of a group $`G`$, instead of the group $`G`$ itself. The framework for the MOR cryptosystem was first proposed in Crypto2001 by Paeng et al. \[[13](#bib.bib13)\]. Mahalanobis \[[10](#bib.bib10)\] used the group of unitriangular matrices for the MOR cryptosystem. That effort was successful, the MOR cryptosystem over the group of unitriangular matrices over $`{\mathbb{F}}_{q}`$ is as secure as the ElGamal cryptosystem over the finite field $`{\mathbb{F}}_{q}`$.
In this paper we study the MOR cryptosystem over SL$`(d,q)`$. If we assume, that the only way to break the proposed MOR cryptosystem, is to solve the discrete logarithm problem in the automorphism group; then it follows that the proposed MOR cryptosystem is as secure as the ElGamal cryptosystem over $`{\mathbb{F}}_{q^{d}}`$.
This is a major improvement. This MOR cryptosystem works with matrices of degree $`d`$ over $`{\mathbb{F}}_{q}`$. To encrypt(decrypt) a plaintext(ciphertext) one works over the field $`{\mathbb{F}}_{q}`$. To break this cryptosystem, one has to solve a discrete logarithm problem in $`{\mathbb{F}}_{q^{d}}`$. Even for a small positive integer $`d`$, this provides us with a considerable security advantage.
There are some challenges in the implementation of this cryptosystem. Implementing matrix multiplication is hard. Though we have not reached the optimum speed for that \[[4](#bib.bib4)\], it might always stay harder than multiplication in a finite field. So one needs to find an optimum strategy to present the automorphisms and the underlying group for the MOR cryptosystem, see Section 8 for more details. The key-size for this MOR cryptosystem is big, compared with the conventional ElGamal cryptoystem.
At the end, I provide parameters for the proposed MOR cryptosystem. I suspect that the parameters are too conservative and the degree of the matrix is unnecessarily big. The overly conservative estimates are to show that for chosen parameters, the MOR cryptosystem is almost as secure as the ElGamal cryptosystem over elliptic curves using fields of same size; the golden standard in public key cryptography. For most practical purposes, the degree of the matrix can be chosen smaller. However the key-size for this MOR cryptosystem is larger than that of the ElGamal over elliptic curves.
## 2. The MOR cryptosystem
This section contains a bare-bone description of the MOR cryptosystem \[[13](#bib.bib13)\], see also \[[12](#bib.bib12)\]. A description and a critical analysis of the MOR cryptosystem is also in \[[10](#bib.bib10)\] and the references there.
### 2.1. Description of the MOR cryptosystem
Let $`G = {\langle g_{1},g_{2},\ldots,g_{\tau}\rangle}`$, $`\tau \in {\mathbb{N}}`$ be a finite group and $`\phi`$ a non-trivial (public) automorphism of $`G`$. Alice’s keys are as follows:
Private Key:
$`m`$, $`m \in {\mathbb{N}}`$.
Public Key:
$`\left\{ {\phi\hspace{0pt}{(g_{i})}} \right\}_{i = 1}^{\tau}`$ and $`\left\{ {\phi^{m}\hspace{0pt}{(g_{i})}} \right\}_{i = 1}^{\tau}`$.
##### Encryption
a:
To send a message (plaintext) $`a \in G`$ Bob computes $`\phi^{r}`$ and $`\phi^{m\hspace{0pt}r}`$ for a random $`r \in {\mathbb{N}}`$.
b:
The ciphertext is $`\left( \left\{ {\phi^{r}\hspace{0pt}{(g_{i})}} \right\}_{i = 1}^{\tau},{\phi^{m\hspace{0pt}r}\hspace{0pt}{(a)}} \right)`$.
##### Decryption
a:
Alice knows $`m`$, so if she receives the ciphertext $`\left( \phi^{r},{\phi^{m\hspace{0pt}r}\hspace{0pt}{(a)}} \right)`$, she computes $`\phi^{m\hspace{0pt}r}`$ from $`\phi^{r}`$ and then $`\phi^{- {m\hspace{0pt}r}}`$ and then computes $`a`$ from $`\phi^{m\hspace{0pt}r}\hspace{0pt}{(a)}`$.
If Alice has the information necessary to find the order of the automorphism $`\phi`$, then she can use the identity $`\phi^{t - 1} = \phi^{- 1}`$ whenever $`\phi^{t} = 1`$ to compute $`\phi^{- {m\hspace{0pt}r}}`$ . Also, she can find the order of some subgroup in which $`\phi`$ belongs and use the same identity. However, the smaller the subgroup, more efficient the decryption algorithm.
## 3. The unimodular group of degree $`d`$ over $`{\mathbb{F}}_{q}`$
The group SL$`(d,q)`$ is the set of all matrices of degree $`d`$ with determinant $`1`$. It is well known that SL$`(d,q)`$ is a normal subgroup of GL$`(d,q)`$ the group of non-singular matrices of degree $`d`$ over $`{\mathbb{F}}_{q}`$. In this article I consider $`{\mathbb{F}}_{q}`$ to be a finite extension of the prime field $`{\mathbb{Z}}_{p}`$ of degree $`\gamma`$ where $`\gamma \geq 1`$.
######
Definition 1
.
For distinct ordered pair $`(i,j)`$, define a matrix unit $`e_{i,j}`$ as a matrix of degree $`d`$, such that, all entries in $`e_{i,j}`$ are 0, except the intersection of the i<sup>th</sup> row and the j<sup>th</sup> column; which is 1 (the identity in the field $`{\mathbb{F}}_{q}`$). Matrices of the form $`1 + {\lambda\hspace{0pt}e_{i,j}}`$, $`\lambda \in {\mathbb{F}}_{q}^{\times}`$ and $`i \neq j`$ are called the elementary matrices or elementary transvections. Here $`1`$ is the identity matrix of degree $`d`$. I shall abuse the notation a little bit and use $`1`$ for the identity of the field and the matrix group simultaneously.
It is known that the group SL$`(d,q)`$ is generated by elementary transvections \[[14](#bib.bib14), Theorem 8.8\]. The fundamental relations between the elementary transvections are the relations in the field and the ones stated below:
<table>
<tbody>
<tr>
<td>(4)</td>
<td></td>
<td></td>
<td>[1+λei,j,1+μek,l]={1+λμei,lifj=k,i≠l1−λμek,jifi=l,j≠k1otherwise1𝜆subscript𝑒𝑖𝑗1𝜇subscript𝑒𝑘𝑙cases1𝜆𝜇subscript𝑒𝑖𝑙ifformulae-sequence𝑗𝑘𝑖𝑙1𝜆𝜇subscript𝑒𝑘𝑗ifformulae-sequence𝑖𝑙𝑗𝑘1otherwisemissing-subexpression\displaystyle[1+\lambda e_{i,j},1+\mu e_{k,l}]=\left\{\begin{array}[]{ccc}1+\lambda\mu e_{i,l}&\text{if}&j=k,\;\;i\neq l\\ 1-\lambda\mu e_{k,j}&\text{if}&i=l,\;\;j\neq k\\ 1&\text{otherwise}&\end{array}\right.</td>
<td></td>
</tr>
</tbody>
</table>
<table>
<tbody>
<tr>
<td>(5)</td>
<td></td>
<td></td>
<td>(1+λei,j)(1+μei,j)=1+(λ+μ)ei,j1𝜆subscript𝑒𝑖𝑗1𝜇subscript𝑒𝑖𝑗1𝜆𝜇subscript𝑒𝑖𝑗\displaystyle\left(1+\lambda e_{i,j}\right)\left(1+\mu e_{i,j}\right)=1+\left(\lambda+\mu\right)e_{i,j}</td>
<td></td>
</tr>
</tbody>
<tbody>
<tr>
<td>(6)</td>
<td></td>
<td></td>
<td>(1+λei,j)−1=(1−λei,j)superscript1𝜆subscript𝑒𝑖𝑗11𝜆subscript𝑒𝑖𝑗\displaystyle\left(1+\lambda e_{i,j}\right)^{-1}=\left(1-\lambda e_{i,j}\right)</td>
<td></td>
</tr>
</tbody>
<tbody>
<tr>
<td>(7)</td>
<td></td>
<td></td>
<td>(1+λei,j)k=1+kλei,jk∈ℕsuperscript1𝜆subscript𝑒𝑖𝑗𝑘1𝑘𝜆subscript𝑒𝑖𝑗𝑘ℕ\displaystyle\left(1+\lambda e_{i,j}\right)^{k}=1+k\lambda e_{i,j}\;\;\;k\in\mathbb{N}</td>
<td></td>
</tr>
</tbody>
</table>
where $`{\lambda,\mu} \in {\mathbb{F}}_{q}`$.
## 4. Automorphisms of the unimodular group over $`{\mathbb{F}}_{q}`$
It is well known that the automorphisms of SL$`(d,q)`$ are generated by the following \[[3](#bib.bib3), [5](#bib.bib5), [17](#bib.bib17)\]:
Diagonal Automorphism:
This is conjugation by a non-scalar diagonal matrix. Notice that: since all diagonal matrices are not of determinant 1, the diagonal matrices are often not in SL$`(d,q)`$. So a diagonal automorphism is not always an inner automorphism.
Inner Automorphism:
This is the most well known automorphism of a non-abelian group $`G`$, defined by $`x\mapsto{g^{- 1}\hspace{0pt}x\hspace{0pt}g}`$ for $`g \in G`$.
Graph Automorphism:
The graph automorphism induces the map $`A\mapsto\left( A^{- 1} \right)^{T}`$, $`A \in {\text{SL}\hspace{0pt}{(d,q)}}`$. Clearly, graph automorphisms are involutions, i.e., of order two and are not inner automorphisms.
Field Automorphism:
This automorphisms is the action of a field automorphism of the underlying field to the individual entries of a matrix.
In this section, I am interested in a special class of inner automorphisms, “the permutation automorphisms”. For a permutation automorphism the conjugator $`g`$ in the inner automorphism is a permutation matrix. It is well known that for a permutation matrix $`P`$, det$`{(P)} = {\pm 1}`$ and $`P^{- 1} = P^{T}`$. The permutation matrix is constructed by taking the identity matrix $`1`$ and then exchanging the rows based on some permutation $`\alpha`$. If the permutation $`\alpha`$ is even then the determinant of $`P`$ is $`1`$ otherwise it is $`- 1`$. Note that if the determinant is $`- 1`$, then conjugation by that permutation matrix is not an inner automorphism; but it is close to being one and I will treat it like an inner automorphism in this paper.
#### 4.0.1. Effect of a permutation automorphism on an elementary transvections
If $`A`$ is an elementary transvection, i.e., $`A = {1 + {\lambda\hspace{0pt}e_{i,j}}}`$ and $`P`$ be a permutation matrix, then $`{P^{- 1}\hspace{0pt}A\hspace{0pt}P} = {1 + {\lambda\hspace{0pt}e_{{\alpha^{- 1}\hspace{0pt}{(i)}},{\alpha^{- 1}\hspace{0pt}{(j)}}}}}`$.
#### 4.0.2. Effect of a diagonal automorphism on an elementary transvection
Let $`D = {\lbrack w_{1},w_{2},\ldots,w_{d}\rbrack}`$ be a diagonal matrix. If $`A = {1 + {\lambda\hspace{0pt}e_{i,j}}}`$ then $`{D^{- 1}\hspace{0pt}A\hspace{0pt}D} = {1 + {{({w_{i}^{- 1}\hspace{0pt}\lambda\hspace{0pt}w_{j}})}\hspace{0pt}e_{i,j}}}`$. Let us fix a $`(i,j)`$ such that $`{1 \leq i},{j \leq d}`$, then look at the *root subgroup* $`\langle{1 + {\lambda\hspace{0pt}e_{i,j}}}\rangle`$, $`\lambda \in {\mathbb{F}}_{q}`$ and $`i \neq j`$. This subgroup is clearly isomorphic to $`{\mathbb{F}}_{q}^{+}`$.
Assume for a moment that I am using the MOR cryptosystem as described in Section [2.1](#S2.SS1 "2.1. Description of the MOR cryptosystem ‣ 2. The MOR cryptosystem ‣ A simple generalization of the ElGamal cryptosystem to non-abelian groups II") with $`G`$ as the root subgroup defined above and $`\phi`$ as a diagonal automorphism. Then clearly for some $`k \in {\mathbb{F}}_{q}^{\times}`$.
<table>
<tbody>
<tr>
<td></td>
<td>ϕ::italic-ϕabsent\displaystyle\phi:</td>
<td>1+ei,j↦maps-to1subscript𝑒𝑖𝑗absent\displaystyle 1+e_{i,j}\mapsto</td>
<td>1+kei,j1𝑘subscript𝑒𝑖𝑗\displaystyle 1+ke_{i,j}</td>
<td></td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td>ϕm::superscriptitalic-ϕ𝑚absent\displaystyle\phi^{m}:</td>
<td>1+ei,j↦maps-to1subscript𝑒𝑖𝑗absent\displaystyle 1+e_{i,j}\mapsto</td>
<td>1+kmei,j.1superscript𝑘𝑚subscript𝑒𝑖𝑗\displaystyle 1+k^{m}e_{i,j}.</td>
<td></td>
</tr>
</tbody>
</table>
Clearly we see that this MOR cryptosystem is equivalent to the ElGamal cryptosystem over finite fields. Since SL$`(d,q)`$ is generated by elementary transvections, I claim that using the diagonal automorphisms of the special linear groups over finite fields, the MOR cryptosystem is identical to the ElGamal cryptosystem over finite fields. It is reasonable to assume that there are other automorphisms, composition of which with the diagonal automorphisms will provide us with better security.
#### 4.0.3. The effect of the graph automorphism on an elementary transvection
It is easy to see from the definition of the graph automorphism that if $`A = {1 + {\lambda\hspace{0pt}e_{i,j}}}`$ then $`\left( A^{- 1} \right)^{T} = {1 - {\lambda\hspace{0pt}e_{j,i}}}`$.
#### 4.0.4. The effect of field automorphisms on an elementary transvections
It is well known that the field automorphisms form a cyclic group generated by the Frobenius automorphism of the field $`{\mathbb{F}}_{q}`$, given by $`\lambda\mapsto\lambda^{p}`$, where $`p`$ is the characteristic of the field $`{\mathbb{F}}_{q}`$. Then the action of field automorphism on an elementary transvection is $`{1 + {\lambda\hspace{0pt}e_{i,j}}}\mapsto{1 + {\lambda^{p^{s}}\hspace{0pt}e_{i,j}}}`$ where $`1 \leq s < \gamma`$.
## 5. MOR with monomial automorphisms
Assume for a moment that I am only using the composition of a diagonal and an inner automorphism of SL$`(d,q)`$, i.e., I am using MOR (Section [2.1](#S2.SS1 "2.1. Description of the MOR cryptosystem ‣ 2. The MOR cryptosystem ‣ A simple generalization of the ElGamal cryptosystem to non-abelian groups II")) where $`\phi = {\phi_{1} \circ \phi_{2}}`$ where $`\phi_{1}`$ is a diagonal automorphism and $`\phi_{2}`$ is a permutation automorphism. Then clearly $`\phi`$ is a monomial automorphism, conjugation by a monomial matrix. The diagonal automorphism $`\phi_{1}`$ changes $`1 + e_{i,j}`$ to $`1 + {\lambda_{i,j}\hspace{0pt}e_{i,j}}`$ for some $`\lambda_{i,j} \in {\mathbb{F}}_{q}^{\times}`$. Note that the $`\lambda_{i,j}`$ depends on the diagonal automorphism and once the diagonal automorphism is fixed $`\lambda_{i,j}`$ is also fixed for a particular $`(i,j)`$. The permutation automorphism $`\phi_{2}`$ changes $`1 + {\lambda_{i,j}\hspace{0pt}e_{i,j}}`$ to $`1 + {\lambda_{i,j}\hspace{0pt}e_{{\beta\hspace{0pt}{(i)}},{\beta\hspace{0pt}{(j)}}}}`$ where $`\beta = \alpha^{- 1}`$. Here $`\alpha`$ is the permutation that gives rise to the permutation matrix $`P`$, used in the permutation automorphism.
I now look at the action of the exponentiation of the automorphism $`\phi = {\phi_{1} \circ \phi_{2}}`$ on the elementary transvection $`1 + e_{i,j}`$. Notice that if
<table>
<tbody>
<tr>
<td>(8)</td>
<td></td>
<td>ϕ:1+ei,j→diagonal1+λi,jei,j→permutation1+λi,jeβ(i),β(j),:italic-ϕcommutative-diagram1subscript𝑒𝑖𝑗superscript→diagonal1subscript𝜆𝑖𝑗subscript𝑒𝑖𝑗superscript→permutation1subscript𝜆𝑖𝑗subscript𝑒𝛽𝑖𝛽𝑗\phi:\begin{CD}1+e_{i,j}@>{\text{diagonal}}>{}>1+\lambda_{i,j}e_{i,j}@>{\text{permutation}}>{}>1+\lambda_{i,j}e_{\beta(i),\beta(j)},\end{CD}</td>
<td></td>
</tr>
</tbody>
</table>
then
<table>
<tbody>
<tr>
<td>(9)</td>
<td></td>
<td>ϕm:1+ei,j→1+∏l=1mλβl(i)βl(j)eβm(i),βm(j):superscriptitalic-ϕ𝑚commutative-diagram1subscript𝑒𝑖𝑗→1superscriptsubscriptproduct𝑙1𝑚subscript𝜆superscript𝛽𝑙𝑖superscript𝛽𝑙𝑗subscript𝑒superscript𝛽𝑚𝑖superscript𝛽𝑚𝑗\phi^{m}:\begin{CD}1+e_{i,j}@>{}>{}>1+\prod\limits_{l=1}^{m}\lambda_{\beta^{l}(i)\beta^{l}(j)}e_{\beta^{m}(i),\beta^{m}(j)}\end{CD}</td>
<td></td>
</tr>
</tbody>
</table>
Now let us assume that the order of $`\beta`$, $`\circ {(\beta)} = \nu`$ then
<table>
<tbody>
<tr>
<td></td>
<td>ϕν:1+ei,j↦1+∏l=1νλβl(i)βl(j)ei,j.:superscriptitalic-ϕ𝜈maps-to1subscript𝑒𝑖𝑗1superscriptsubscriptproduct𝑙1𝜈subscript𝜆superscript𝛽𝑙𝑖superscript𝛽𝑙𝑗subscript𝑒𝑖𝑗\phi^{\nu}:1+e_{i,j}\mapsto 1+\prod\limits_{l=1}^{\nu}\lambda_{\beta^{l}(i)\beta^{l}(j)}e_{i,j}.</td>
<td></td>
</tr>
</tbody>
</table>
This shows that a cycle is formed and if $`\nu < m`$, then this reduces the discrete logarithm problem in $`\langle\phi\rangle`$ to a discrete logarithm problem in the finite field $`{\mathbb{F}}_{q}`$. Though it is well known that in the symmetric group $`S_{n}`$, acting on $`n`$ points, one can get elements with very high order. In our problem I am actually interested in the length of the orbit formed by the action of a cyclic subgroup of $`S_{n}`$, generated by $`\beta`$, on the set of distinct ordered pair of $`\{ 1,2,\ldots,n\}`$. It is known that these orbits are quite small.
Since the permutation $`\beta`$ is easy to find from the public information $`\phi`$ and $`\phi^{m}`$, unless the degree of the matrix $`d`$ is astronomically big, we do not have any chance for a MOR cryptosystem which is more secure than that of the ElGamal cryptosystem over finite fields.
Since the conjugacy problem is easy in GL$`(d,q)`$, from the public information of $`\phi_{1}`$ and $`\phi_{2}`$ one can compute the conjugator monomial matrices for $`\phi_{1}`$ and $`\phi_{2}`$ modulo an element of the center of GL$`(d,q)`$. I shall come back to this topic later (Section 7.2) in more details.
## 6. Structure of the automorphism group of SL$`(d,q)`$
Let us start with a well known theorem describing the structure of the automorphism group of SL$`(d,q)`$. Let $`\mathcal{A}`$ be the group of automorphisms generated by the diagonal and the inner automorphisms and $`\mathcal{B}`$ be the group generated by the graph and the field automorphisms. Recall that the center of the group GL$`(d,q)`$ is the set of all scalar matrices $`\lambda\hspace{0pt}1`$ where $`\lambda \in {\mathbb{F}}_{q}^{\times}`$ and $`1`$ is the identity matrix of degree $`d`$. I shall denote the center of GL$`(d,q)`$ by $`Z`$ and *the projective general linear group* $`\frac{\text{GL}\hspace{0pt}{(d,q)}}{Z}`$ by PGL$`(d,q)`$.
A brief *warning* about the notation. To increase readability of the text, from now on, the image of $`a`$ under $`f`$ will be denoted either by $`a^{f}`$ or by $`f\hspace{0pt}{(a)}`$. Also, I denote the conjugation of $`X`$ by $`A`$ as $`X^{A}`$.
######
Theorem 6.1
.
The group $`\mathcal{A}`$ is isomorphic to PGL$`(d,q)`$ and Aut$`\left( {\text{SL}\hspace{0pt}{(d,q)}} \right)`$ is a semidirect product of $`\mathcal{A}`$ with $`\mathcal{B}`$.
###### Proof.
From \[[2](#bib.bib2), Theorem 2.12\] we know that any element in GL$`(d,q)`$ is generated by the set consisting of all invertible diagonal matrices and all transvections. Then we can define a map $`ϝ:{{\text{GL}\hspace{0pt}{(d,q)}}\rightarrow\mathcal{A}}`$ defined by $`ϝ\hspace{0pt}{(A)}`$ maps $`X\mapsto X^{A}`$, clearly $`ϝ`$ is an epimorphism and Ker$`{(ϝ)} = Z`$. From first isomorphism theorem we have that PGL$`{(d,q)} \cong \mathcal{A}`$.
We are left to show that Aut$`\left( {\text{SL}\hspace{0pt}{(d,q)}} \right)`$ is a semidirect product of $`\mathcal{A}`$ with $`\mathcal{B}`$. To prove this we need to show that $`\mathcal{A}`$ is a normal subgroup of $`\text{Aut}\hspace{0pt}\left( {\text{SL}\hspace{0pt}{(d,q)}} \right)`$ and Aut$`\left( {\text{SL}\hspace{0pt}{(d,q)}} \right) = {\mathcal{A}\hspace{0pt}\mathcal{B}}`$. Notice that any $`f \in \mathcal{B}`$ is an automorphism of GL$`(d,q)`$. With this in mind we see that for $`A \in {\text{GL}\hspace{0pt}{(d,q)}}`$ and $`X \in {\text{SL}\hspace{0pt}{(d,q)}}`$
<table>
<tbody>
<tr>
<td></td>
<td>XfAf−1=f(A−1f−1(X)A)=f(A)−1Xf(A)=Xf(A).superscript𝑋𝑓𝐴superscript𝑓1𝑓superscript𝐴1superscript𝑓1𝑋𝐴𝑓superscript𝐴1𝑋𝑓𝐴superscript𝑋𝑓𝐴X^{fAf^{-1}}=f\left(A^{-1}f^{-1}(X)A\right)=f(A)^{-1}Xf(A)=X^{f(A)}.</td>
<td></td>
</tr>
</tbody>
</table>
This proves that $`\mathcal{A}`$ is a normal subgroup of Aut$`\left( {\text{SL}\hspace{0pt}{(d,q)}} \right)`$. Now notice that for any $`f \in \mathcal{B}`$, $`{A^{- 1}\hspace{0pt}X^{f}\hspace{0pt}A} = \left( {{(A^{- 1})}^{f^{- 1}}\hspace{0pt}X\hspace{0pt}A^{f^{- 1}}} \right)^{f}`$, where $`A \in {\text{GL}\hspace{0pt}{(d,q)}}`$. This proves that we can move elements of $`\mathcal{B}`$ to the right of the product of automorphisms. This proves our theorem. ∎
Now notice that the order of $`\mathcal{A}`$ is actually big, it is $`q^{\frac{d\hspace{0pt}{({d - 1})}}{2}}\hspace{0pt}{({q^{d} - 1})}\hspace{0pt}\cdots\hspace{0pt}{({q - 2})}`$ but the order of $`\mathcal{B}`$ is small. The group $`\mathcal{B}`$ is the direct product of the graph and field automorphisms. The order of $`\mathcal{B}`$ is $`2\hspace{0pt}\gamma`$, where $`\gamma`$ is the degree for the extension $`{\mathbb{F}}_{q}`$ over the prime subfield. Let $`\gamma_{1} = {2\hspace{0pt}\gamma}`$.
Let $`\phi`$ and $`\phi^{m}`$ be as in Section [2.1](#S2.SS1 "2.1. Description of the MOR cryptosystem ‣ 2. The MOR cryptosystem ‣ A simple generalization of the ElGamal cryptosystem to non-abelian groups II"), then from the previous theorem $`\phi = {A\hspace{0pt}\psi_{1}}`$ and $`\phi^{m} = {A'\hspace{0pt}\psi_{2}}`$, where $`{A,A'} \in \mathcal{A}`$ and $`{\psi_{1},\psi_{2}} \in \mathcal{B}`$. I shall consider $`A \in \mathcal{A}`$ as the conjugator as well, this is clearly the case because $`\mathcal{A} \cong {\text{PGL}\hspace{0pt}{(d,q)}}`$.
Now if $`\phi = {A\hspace{0pt}\psi_{1}}`$, then $`\phi^{m} = {A\hspace{0pt}A^{\psi_{1}}\hspace{0pt}\cdots\hspace{0pt}A^{\psi_{1}^{m - 2}}\hspace{0pt}A^{\psi_{1}^{m - 1}}\hspace{0pt}\psi_{1}^{m}}`$. In this case
$`{A\hspace{0pt}A^{\psi_{1}}\hspace{0pt}\cdots\hspace{0pt}A^{\psi_{1}^{m - 2}}\hspace{0pt}A^{\psi_{1}^{m - 1}}} \in \mathcal{A}`$ and $`\psi_{1}^{m} \in \mathcal{B}`$.
Now if $`\gamma_{1} < m`$ and since the order of $`\psi_{1}`$ divides $`\gamma_{1}`$, there are $`r_{1}`$ and $`r_{2}`$ such that $`{m - 1} = {{k_{1}\hspace{0pt}\gamma_{1}} + r_{1}}`$, where $`0 \leq r_{1} < \gamma_{1}`$ and $`r_{2} = {m\operatorname{mod}\gamma_{1}}`$. Then $`{A\hspace{0pt}A^{\psi_{1}}\hspace{0pt}\cdots\hspace{0pt}A^{\psi_{1}^{m - 1}}\hspace{0pt}\psi_{1}^{m}} = {A_{1}^{k_{1}}\hspace{0pt}A\hspace{0pt}A^{\psi_{1}}\hspace{0pt}\cdots\hspace{0pt}A^{\psi_{1}^{r_{1}}}\hspace{0pt}\psi_{1}^{r_{2}}}`$, where $`A_{1} = {A\hspace{0pt}A^{\psi_{1}}\hspace{0pt}\cdots\hspace{0pt}A^{\psi_{1}^{\gamma_{1} - 1}}}`$. From the information of $`\phi`$ and $`\phi^{m}`$ we then have the information of $`\psi_{1}`$ and $`\psi_{1}^{r_{2}}`$. For all practical purposes of implementing this cryptosystem, the degree of the field extension cannot be too large, in this case one can do an exhaustive search on the cosets of $`\mathcal{A}`$ and find out $`\psi_{1}`$ and $`\psi_{1}^{r_{2}}`$ and do another exhaustive search to solve the discrete logarithm problem in $`\psi_{1}`$ and find the $`r_{2}`$. The information of $`r_{2}`$ gives us a vital information about the secret key $`m`$. This is clearly unacceptable. So the only way out of this situation is not to use automorphisms from $`\mathcal{B}`$.
Then for $`X \in {\text{SL}\hspace{0pt}{(d,q)}}`$ the automorphisms $`\phi`$ and $`\phi^{m}`$ as in Section [2.1](#S2.SS1 "2.1. Description of the MOR cryptosystem ‣ 2. The MOR cryptosystem ‣ A simple generalization of the ElGamal cryptosystem to non-abelian groups II") is given by
<table>
<tbody>
<tr>
<td>(10)</td>
<td></td>
<td>ϕ(X)italic-ϕ𝑋\displaystyle\phi(X)</td>
<td>=A−1XAabsentsuperscript𝐴1𝑋𝐴\displaystyle=A^{-1}XA</td>
<td>for someA∈GL(d,q)for some𝐴GL𝑑𝑞\displaystyle\text{for some}\;\;A\in\text{GL}(d,q)</td>
<td></td>
</tr>
</tbody>
<tbody>
<tr>
<td>(11)</td>
<td></td>
<td>ϕm(X)superscriptitalic-ϕ𝑚𝑋\displaystyle\phi^{m}(X)</td>
<td>=A′−1XA′absentsuperscript𝐴′1𝑋superscript𝐴′\displaystyle=A^{\prime{-1}}XA^{\prime}</td>
<td>for someA′∈GL(d,q)for somesuperscript𝐴′GL𝑑𝑞\displaystyle\text{for some}\;\;A^{\prime}\in\text{GL}(d,q)</td>
<td></td>
</tr>
</tbody>
</table>
Now notice, in the description of the MOR protocol, we presented the automorphisms as action on generators and furthermore a set of generators for SL$`(d,q)`$ are the elementary transvections.
In this case from the public information of $`\phi`$ and $`\phi^{m}`$ one can find a candidate for $`A`$ and $`A'`$. This problem is known to be easy in GL$`(d,q)`$ and is often refereed to as *the special conjugacy problem* \[[12](#bib.bib12), [13](#bib.bib13)\]. However, notice that $`A`$ and $`A'`$ are not unique. For example, if $`A`$ and $`A'`$ satisfy the above equations then so will $`A\hspace{0pt}z`$ and $`A'\hspace{0pt}z'`$ for any $`{z,z'} \in Z`$, see Section [7.1.1](#S7.SS1.SSS1 "7.1.1. Recovering the conjugator up to a scalar multiple ‣ 7.1. Inner automorphisms as matrices ‣ 7. Security of the proposed MOR cryptosystem ‣ A simple generalization of the ElGamal cryptosystem to non-abelian groups II").
We just saw that the only way to build a secure MOR cryptosystem using SL$`(d,q)`$ is using automorphisms from $`\mathcal{A}`$. Henceforth, whenever we are talking about the MOR cryptosystem, we are using the group SL$`(d,q)`$ and the automorphisms from $`\mathcal{A}`$.
## 7. Security of the proposed MOR cryptosystem
This paper is primarily focused on the discrete logarithm problem in the automorphism group of a non-abelian group. There are two kinds of attack on the discrete logarithm problem over finite fields. One is the generic attack; this attack uses a *black box* group algorithm and the other is an *index calculus* attack.
Since the black box group algorithms work in any group, they will work in the automorphism group too, see \[[9](#bib.bib9), Theorem 1\]. We have no way to prevent that. On the other hand, these generic attacks are of exponential time complexity and so is of the least concern.
The biggest computational threat to any cryptosystem using the discrete logarithm problem is a subexponential attack like the index calculus attack \[[15](#bib.bib15)\]. It is often argued \[[8](#bib.bib8), [16](#bib.bib16)\] that there is no index calculus algorithm for most elliptic curve cryptosystems that has subexponential time complexity. This fact is often presented to promote elliptic curve cryptosystem over a finite field cryptosystem \[[8](#bib.bib8)\]. So, the best we can hope from the present MOR cryptosystem is that there is no index calculus attack or the index calculus attack becomes exponential.
### 7.1. Inner automorphisms as matrices
As it turns out the only way that a secure MOR cryptosystem might work for the unimodular group is through conjugation of matrices, i.e., automorphisms from $`\mathcal{A}`$. This MOR cryptosystem can be seen as working with inner automorphisms of GL$`(d,q)`$. It is well known that the inner automorphisms work linearly on the $`d^{2}`$-dimensional algebra of matrices of degree $`d`$ over $`{\mathbb{F}}_{q}`$. For a fixed basis, any linear operator on a vector space can be represented as a matrix. So, the discrete logarithm problem on $`\langle\phi\rangle`$ (Section [2.1](#S2.SS1 "2.1. Description of the MOR cryptosystem ‣ 2. The MOR cryptosystem ‣ A simple generalization of the ElGamal cryptosystem to non-abelian groups II")) is now reduced to the discrete logarithm problem in GL$`(d^{2},q)`$. The question is, how easy is it to solve this discrete logarithm problem?
The best algorithm for solving the discrete logarithm problem in GL$`(k,q)`$ was given by Menezes and Wu \[[11](#bib.bib11)\]. In this case, the authors show that for $`{X,Y} \in {\text{GL}\hspace{0pt}{(k,q)}}`$, such that, $`X^{l} = Y`$, $`l \in {\mathbb{N}}`$; we can solve the discrete logarithm problem, if $`\chi\hspace{0pt}{(x)}`$ the characteristic polynomial of $`X`$, factors into irreducible polynomials of small degree. If the characteristic polynomial is irreducible then the discrete logarithm problem in $`\langle X\rangle`$ reduces to the discrete logarithm problem in $`{\mathbb{F}}_{q^{k}}`$.
In our case we are working in GL$`(d^{2},q)`$. So the characteristic polynomial has degree $`d^{2}`$. It is easy to see that if the characteristic polynomial is irreducible then the extension of the lowest degree in which the characteristic polynomial will split is $`{\mathbb{F}}_{q^{d^{2}}}`$. However this is not the case, since $`{\phi\hspace{0pt}{(1)}} = 1`$, $`1`$ is an eigenvalue of $`\phi`$ and so the best we can hope for is $`{\mathbb{F}}_{{q\hspace{0pt}d^{2}} - 1}`$.
#### 7.1.1. Recovering the conjugator up to a scalar multiple
Let $`{\phi\hspace{0pt}{(X)}} = {A^{- 1}\hspace{0pt}X\hspace{0pt}A}`$, where $`A \in {\text{GL}\hspace{0pt}{(d,q)}}`$. Since $`\phi`$ is linear, if we take $`X = {1 + e_{i\hspace{0pt}j}}`$, $`i \neq j`$; then $`{\phi\hspace{0pt}{(X)}} = {A^{- 1}\hspace{0pt}X\hspace{0pt}A} = {1 + {A^{- 1}\hspace{0pt}e_{i\hspace{0pt}j}\hspace{0pt}A}}`$. Now if we look at $`e_{i\hspace{0pt}j}\hspace{0pt}A`$ closely, then $`e_{i\hspace{0pt}j}\hspace{0pt}A`$ is a matrix where the $`j`$<sup>th</sup> row of $`A`$ is the $`i`$<sup>th</sup> row of $`e_{i\hspace{0pt}j}\hspace{0pt}A`$, and the rest all zeros. Since $`A`$ is non-singular, all the contents of any row can not be all zeros. From this it follows that the matrix $`A^{- 1}\hspace{0pt}e_{i\hspace{0pt}j}\hspace{0pt}A`$ consists of $`d`$ columns, each of which is a constant multiple of the $`i`$<sup>th</sup> column of $`A^{- 1}`$. One of these columns must be nonzero. Now consider $`{A^{- 1}\hspace{0pt}\left( {1 + e_{i,{i + 1}}} \right)\hspace{0pt}A} - 1`$ for $`i = {1,2,\ldots,{d - 1}}`$, and let each $`I_{i}`$ be a corresponding nonzero column. Then construct a $`d \times d`$ matrix, whose first $`d - 1`$ columns are the columns $`I_{1},I_{2},\ldots,I_{d - 1}`$ and the $`d`$<sup>th</sup> column being a nonzero column of $`{A^{- 1}\hspace{0pt}\left( {1 + e_{n\hspace{0pt}1}} \right)\hspace{0pt}A} - 1`$. Then we end up with a matrix $`N = {A^{- 1}\hspace{0pt}D}`$, where $`D = {\lbrack w_{1},w_{2},\ldots,w_{d}\rbrack}`$ is a diagonal matrix. Since $`N`$ is known, we have found $`A^{- 1}`$ up to a diagonal matrix.
It is obvious that $`{N^{- 1}\hspace{0pt}\phi\hspace{0pt}{(X)}\hspace{0pt}N} = {D^{- 1}\hspace{0pt}X\hspace{0pt}D}`$ and hence $`{{N^{- 1}\hspace{0pt}\left( {1 + e_{i\hspace{0pt}j}} \right)\hspace{0pt}N} - 1} = {w_{i}^{- 1}\hspace{0pt}w_{j}\hspace{0pt}e_{i\hspace{0pt}j}}`$. Then by taking $`j = {1,2,\ldots,d}`$ and $`i = 1`$, we can find $`{w_{2}^{- 1}\hspace{0pt}w_{1}},{w_{3}^{- 1}\hspace{0pt}w_{1}},\ldots,{w_{d}^{- 1}\hspace{0pt}w_{1}}`$, and form the diagonal matrix $`D' = {\lbrack 1,{w_{1}\hspace{0pt}w_{2}^{- 1}},{w_{1}\hspace{0pt}w_{3}^{- 1}},\ldots,{w_{1}\hspace{0pt}w_{d}^{- 1}}\rbrack}`$. It is easy to see now that $`N\hspace{0pt}D'`$ is $`A^{- 1}\hspace{0pt}w_{1}`$ and we have found $`A`$ up to a scalar multiple.
It is not hard to convince oneself that once $`A`$ is found up to a scalar multiple from $`\phi`$, in most cases the discrete logarithm problem in $`\phi`$ turns out to be a discrete logarithm problem in $`A`$. If one recovers $`A\hspace{0pt}c_{1}`$ and $`A^{m}\hspace{0pt}c_{2}`$ from $`\phi`$ and $`\phi^{m}`$, where $`{c_{1},c_{2}} \in {\mathbb{F}}_{q}^{\times}`$, then compute $`{({A\hspace{0pt}c_{1}})}^{q - 1} = A^{q - 1}`$ and $`{({A^{m}\hspace{0pt}c_{2}})}^{q - 1} = A^{m\hspace{0pt}{({q - 1})}}`$ and solve the corresponding discrete logarithm problem. From Menezes-Wu \[[11](#bib.bib11)\] it is clear that this discrete logarithm problem can have a worst case complexity of that of a discrete logarithm problem in $`{\mathbb{F}}_{q^{d}}`$.
We can stop this attack by taking $`A`$ to be of order $`q - 1`$. But, if the eigenvalues of $`A`$ are $`\mu_{1},\mu_{2},\ldots,\mu_{d}`$, then the eigenvalues of $`A^{m}`$ are $`\mu_{1}^{m},\mu_{2}^{m},\ldots,\mu_{d}^{m}`$. On the other hand the eigenvalues of $`c\hspace{0pt}A`$ are $`{c\hspace{0pt}\mu_{1}},{c\hspace{0pt}\mu_{2}},\ldots,{c\hspace{0pt}\mu_{d}}`$, $`c \in {\mathbb{F}}_{q}`$. When one recovers $`c_{1}\hspace{0pt}A`$ and $`c_{2}\hspace{0pt}A^{m}`$, $`{c_{1},c_{2}} \in {\mathbb{F}}_{q}`$ one recovers $`{c_{1}\hspace{0pt}\mu_{1}},{c_{1}\hspace{0pt}\mu_{2}},\ldots,{c_{1}\hspace{0pt}\mu_{d}}`$ and $`{c_{2}\hspace{0pt}\mu_{1}^{m}},{c_{2}\hspace{0pt}\mu_{2}^{m}},\ldots,{c_{2}\hspace{0pt}\mu_{d}^{m}}`$. Then one can compute $`\frac{\mu_{i}}{\mu_{j}}`$ and $`\left( \frac{\mu_{i}}{\mu_{j}} \right)^{m}`$, by taking quotients. Notice that these quotients belong to $`{\mathbb{F}}_{q^{d}}`$. However since there is no unique way to order the eigenvalues, one might not be able to match a quotient with its power. Then we might have to deal with several quotients to get to the right $`m`$. However, for all practical applications the size of the matrix $`d`$ is small and so this search is not going to cost much; on top of that one can do this in parallel. So it is resonable to claim at this stage that the discrete logarithm problem in $`\phi`$ is reduced to a discrete logarithm problem in $`{\mathbb{F}}_{q^{d}}`$.
The expected asymptotic complexity of the index calculus algorithm in $`{\mathbb{F}}_{q^{k}}`$ is $`\exp\left( {{({c + {o\hspace{0pt}{(1)}}})}\hspace{0pt}{({\log q^{k}})}^{\frac{1}{3}}\hspace{0pt}{({\log{\log q^{k}}})}^{\frac{2}{3}}} \right)`$ , where $`c`$ is a constant, see \[[15](#bib.bib15)\] and \[[8](#bib.bib8), Section 4\]. If the degree of the extension, $`k`$, is greater than $`\log^{2}q`$ then the asymptotic time complexity of the index calculus algorithm is exponential. In our case that means, if $`d > {\log^{2}q}`$ then the asymptotic complexity of the index calculus algorithm becomes exponential.
If we choose $`d \geq {\log^{2}q}`$ then this MOR cryptosystems becomes as secure as the ElGamal over the elliptic curve groups, because the index calculus algorithm is exponential; otherwise we can not guarantee. But on the other hand in the proposed MOR cryptosystem encryption and decryption works on $`{\mathbb{F}}_{q}`$ and breaking the cryptosystem depends on solving a discrete logarithm problem on $`{\mathbb{F}}_{q^{d}}`$. Since, implementing the index calculus attack becomes harder as the field gets bigger, it is clear that if we take $`d \ll {\log^{2}q}`$, then the MOR cryptosystem is much more secure than the ElGamal cryptosystem over $`{\mathbb{F}}_{q}`$.
## 8. Implementation of this MOR cryptosystem
The cryptosystem we have in mind is the MOR cryptosystem (Section [2.1](#S2.SS1 "2.1. Description of the MOR cryptosystem ‣ 2. The MOR cryptosystem ‣ A simple generalization of the ElGamal cryptosystem to non-abelian groups II")), the non-abelian group is SL$`(d,q)`$ and the automorphisms are the automorphisms from $`\mathcal{A}`$. In this implementation the most important thing will be the presentation of $`\phi`$ and $`\phi^{m}`$. We decided earlier that the presentation will be the action of the automorphisms on a set of generators $`\{ g_{1},g_{2},\ldots,g_{\tau}\}`$. Now we can write $`\phi\hspace{0pt}{(g_{i})}`$ as a word in the generators $`g_{1},g_{2},\ldots,g_{\tau}`$ or we can write the product of the generators as a matrix. We choose the later, there are two reasons for that:
:
This will contain the growth in the length of the word, especially while computing the powers of $`\phi`$. That will stop any length based attack.
:
This will add to the diffusion.
The set of generators for SL$`(d,q)`$ that we have in mind is the elementary transvections. It is easy to go back and forth as words in elementary transvections and matrices using row reduction.
A big question is how to compute large powers of $`\phi`$ efficiently? This is not the principal object of study for this paper and we will be brief on this topic.
Since a set of generators are elementary transvections, computing the power of $`\phi`$ can be done using only words in elementary transvections and the image of the automorphism on these elementary transvections. This can be done very efficiently. However, we have decided to write $`\phi^{m}\hspace{0pt}{(g_{i})}`$ as matrices. So, while computing the power of $`\phi`$, one might have to go back and forth between words and matrices. The objective of this exercise is to reduce the amount of matrix multiplication, in computing the power of $`\phi`$. Also, one can use the relations among the elementary transvections to shorten the length of the word. There are quite a few options available.
We explore one such option in more details. Assume that we are computing the $`\phi^{m}`$ using the *square and multiply* algorithm \[[18](#bib.bib18), Algorithm 5.5\]. In this algorithm one needs to multiply two group elements, in our case it is composing two automorphisms. So, we need to find out the worst-case complexity for multiplying two automorphisms. I further assume that the automorphism is given as the image of $`\left( {1 + e_{i,j}} \right)`$, $`i \neq j`$, $`{i,j} \in {\{ 1,2,\ldots,d\}}`$, the image is one $`d \times d`$ matrix. So for sake of notational convenience I assume that we are squaring $`\phi`$, where $`\phi`$ is given by the action on elementary transvections. As is customary we assume that the field addition is free and we count the number of field multiplications necessary to do the computation.
Let’s start with the matrix $`M`$ such that $`M = {\phi\hspace{0pt}\left( {1 + e_{i,j}} \right)}`$, I shall use row operations to write $`M`$ as product of elementary transvections. We count each row operation as $`d`$ field multiplications and there are utmost $`d^{2}`$ row operation. So in the worst case after $`d^{3}`$ many field multiplication we have written $`M`$ as a product of elementary transvection. At most there are $`d^{2}`$ many elementary transvections in the product
<sup>1</sup>
<sup>1</sup>1Some small examples computed by the author using GAP \[[6](#bib.bib6)\] suggests that in practice this number is much smaller.
.
Using the relation in Equation [5](#S3.E5 "In 3. The unimodular group of degree 𝑑 over 𝔽_𝑞 ‣ A simple generalization of the ElGamal cryptosystem to non-abelian groups II"), we split each transvection into product of elementary transvections over the prime subfield. So now there are $`\gamma\hspace{0pt}d^{2}`$ elementary transvections over the prime subfield, for each of which the image under $`\phi`$ is known. Then the image under $`\phi`$ is computed and then and then there are $`{({p - 1})}\hspace{0pt}\gamma\hspace{0pt}d^{4}`$ elementary transvection. The question is how to compute the matrix corresponding to that? I propose the following:
There are utmost $`{({p - 1})}\hspace{0pt}\gamma\hspace{0pt}d^{4}`$ elementary transvections in the product of $`\phi\hspace{0pt}{(M)}`$. Make $`d`$ equally spaced partition of the product of $`\phi\hspace{0pt}{(M)}`$. Then each one of these partitions can have utmost $`{({p - 1})}\hspace{0pt}\gamma\hspace{0pt}d^{3}`$ elementary transvections. Now we multiply the $`{({p - 1})}\hspace{0pt}\gamma\hspace{0pt}d^{3}`$ elementary transvections to get $`d`$ many matrices and them multiply these $`d`$ many matrices to get the final matrix corresponding to $`\phi^{2}\hspace{0pt}\left( {1 + e_{i,j}} \right)`$. Now we multiply the $`{({p - 1})}\hspace{0pt}\gamma\hspace{0pt}d^{3}`$ elementary transvections linearly, one after the other, and use the relations in Equations [4](#S3.E4 "In 3. The unimodular group of degree 𝑑 over 𝔽_𝑞 ‣ A simple generalization of the ElGamal cryptosystem to non-abelian groups II") and [5](#S3.E5 "In 3. The unimodular group of degree 𝑑 over 𝔽_𝑞 ‣ A simple generalization of the ElGamal cryptosystem to non-abelian groups II"). Notice that one of the components in this multiplication is an elementary transvection. So every multiplication can take utmost $`d`$ many field multiplication. So the total complexity of multiplying $`{({p - 1})}\hspace{0pt}\gamma\hspace{0pt}d^{3}`$ many elementary transvections is $`{({p - 1})}\hspace{0pt}\gamma\hspace{0pt}d^{4}`$. Since different partitions can be multiplied in parallel, we assume that the worst-case complexity is $`{({p - 1})}\hspace{0pt}\gamma\hspace{0pt}d^{4}`$ field multiplications.
Now we have to multiply the $`d`$ many matrices thus obtained. We assume that we use a straight line program to compute the product. Assuming that matrix multiplication can be done in $`d^{3}`$ field multiplication, we see that this also requires $`d^{4}`$ field multiplications. Since we can compute $`\phi^{2}\hspace{0pt}\left( {1 + e_{i,j}} \right)`$ in parallel for different $`i`$ and $`j`$, we claim that we can multiply two automorphisms with worst-case complexity $`{{({p - 1})}\hspace{0pt}\gamma\hspace{0pt}d^{4}} + d^{4}`$ field multiplications.
### 8.1. Parameters for the cryptosystem
We realized that if the conjugator $`A`$ in $`\phi`$ (Equation [10](#S6.E10 "In 6. Structure of the automorphism group of SL(𝑑,𝑞) ‣ A simple generalization of the ElGamal cryptosystem to non-abelian groups II")) is a monomial matrix then that prevents the formation of a discrete logarithm problem in the $`\lambda`$ of an elementary transvection $`1 + {\lambda\hspace{0pt}e_{i,j}}`$. We need the inner automorphism so that the attack due to small cycle size of the permutation in the monomial matrix can be avoided. So we have to take the automorphism as conjugation by $`A \in {\text{GL}\hspace{0pt}{(d,q)}}`$.
The size of $`d`$ and $`q`$ is an open question. With the limited amount of knowledge we have about this cryptosystem, we can only make a preliminary attempt to encourage further research. The current standard for security in the public key cryptography is 80-bit security. This means that the best known attack to the cryptosystem should take at least $`2^{80}`$ steps.
The best known attack on the discrete logarithm problem in the matrices $`A`$ and $`A'`$ (Equations [10](#S6.E10 "In 6. Structure of the automorphism group of SL(𝑑,𝑞) ‣ A simple generalization of the ElGamal cryptosystem to non-abelian groups II") and [11](#S6.E11 "In 6. Structure of the automorphism group of SL(𝑑,𝑞) ‣ A simple generalization of the ElGamal cryptosystem to non-abelian groups II")) is the generic *square root* attack. So we have to ensure that to find $`m`$ from $`A`$ and $`A'`$ one needs at least $`2^{80}`$ steps. For an attack algorithm we assume that computing in $`{\mathbb{F}}_{q}`$ and in GL$`(d,q)`$ takes the same amount of time. If we assume that the order of the matrix $`A`$ is the same as the order of the field
<sup>2</sup>
<sup>2</sup>2The size of the field is motivated by the use of similar field in elliptic curve cryptography. For elliptic curves, the choice depends on the fact that the size of the group of rational points on an elliptic curve is roughly the size of the field. In our case, there are matrices of high order in GL$`(d,q)`$. So the field can be chosen smaller, depending on the matrix we choose to use.
, then the order of the field should be around $`2^{160}`$. So there are two choices for $`q`$, take $`q`$ to be a prime of the order $`2^{160}`$, i.e., a 160 bit prime; or take $`{\mathbb{F}}_{q} = {\mathbb{F}}_{2^{160}}`$.
A similar situation arises with the discrete logarithm problem over the group of an elliptic curve over a finite field. The MOV attack reduces the discrete logarithm problem in the group of the elliptic curve over $`{\mathbb{F}}_{q}`$ to a discrete logarithm problem in $`{\mathbb{F}}_{q^{k}}^{\times}`$ for some positive integer $`k`$. This is of concern in the implementation of the elliptic curve cryptosystem, because if $`k`$ is too small then there is an subexponential attack on the elliptic curve discrete logarithm problem. On the other hand, the size of the elliptic curve group is almost as big as the field. To prevent the square root attack the size of the field has to be considerably higher. Once you assume that the field is of appropriate size $`(2^{160})`$, small $`k`$ provides adequate security. Our case is quite similar.
Koblitz et al. \[[8](#bib.bib8), Section 5.2\] mentions that in practice $`k \approx 20`$ is enough for security. If we buy their argument, then it would seem that one can choose $`d`$ to be a around $`20`$. We suspect that one might be able to go even smaller. In our MOR cryptosystem, Menezes-Wu algorithm reduces the discrete logarithm problem in $`{\mathbb{F}}_{q^{d}}`$.
So we propose $`d = 19`$, and $`q`$ is as described earlier. Then we see that if $`q = 2^{160}`$, then we are talking about a discrete logarithm problem in $`{\mathbb{F}}_{2^{3040}}`$. This clearly surpasses every standard for discrete logarithm problem over finite fields. At this size of the field, it does not matter if the index-calculus is exponential or sub-exponential. It is simply not doable.
### 8.2. Generators for the cryptosystem
The question I raise in this section is, are their better generators than the elementary transvections in SL$`(d,q)`$? We saw that if we use the elementary transvections for a prime field, then one needs $`({d^{2} - d})`$ elementary transvections and $`{({d^{2} - d})}\hspace{0pt}\gamma`$ elementary transvections for $`{\mathbb{F}}_{q}`$ where $`q = p^{\gamma}`$.
This is one of the major problems in the implementation of this cryptosystem. We now try to solve this problem for SL$`(d,p)`$, where $`p`$ is a prime. In this MOR cryptosystem (Section [2.1](#S2.SS1 "2.1. Description of the MOR cryptosystem ‣ 2. The MOR cryptosystem ‣ A simple generalization of the ElGamal cryptosystem to non-abelian groups II")), generators play a major role. There are some properties of the generators that help. Two of them are:
i:
There should be an efficient algorithm to solve the word problem in these generators.
ii:
The less the number of generators of the group, the better is the cryptosystem.
Albert and Thompson \[[1](#bib.bib1)\] provides us with two generators for SL$`(d,q)`$. They are
<table>
<tbody>
<tr>
<td></td>
<td>C=1+αed−1,2+ed,1C1𝛼subscript𝑒𝑑12subscript𝑒𝑑1\displaystyle\text{C}=1+\alpha e_{d-1,2}+e_{d,1}</td>
<td></td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td>D=(−1)d(e1,2−e2,3+∑i=3dei,i+1)Dsuperscript1𝑑subscript𝑒12subscript𝑒23superscriptsubscript𝑖3𝑑subscript𝑒𝑖𝑖1\displaystyle\text{D}=(-1)^{d}\left(e_{1,2}-e_{2,3}+\sum\limits_{i=3}^{d}e_{i,i+1}\right)</td>
<td></td>
</tr>
</tbody>
</table>
where $`\alpha`$ is a primitive element of $`{\mathbb{F}}_{q}`$. It is clear from the proof of \[[1](#bib.bib1), Lemma 1\] that to solve the word problem in these generators one has to solve the discrete logarithm problem in $`{\mathbb{F}}_{q}`$. This is clearly not useful for our cause. So we adapt the generators and extend it to show that for these generators one can compute the elementary transvections. Since the number of generators is $`2`$, this gives us an advantage for the presentation of the public key and the ciphertext over elementary transvections. However, I know of no efficient algorithm to solve the word problem in these generators. If we can find one such algorithm then it can be argued that this cryptosystem would become more economical(efficient).
I now prove a theorem which is an adaptation of \[[1](#bib.bib1), Lemma 1\]. I use the convention used by Albert and Thomson,
<table>
<tbody>
<tr>
<td></td>
<td>ei,j=ed+i,j=ei,d+j.subscript𝑒𝑖𝑗subscript𝑒𝑑𝑖𝑗subscript𝑒𝑖𝑑𝑗e_{i,j}=e_{d+i,j}=e_{i,d+j}.</td>
<td></td>
</tr>
</tbody>
</table>
The proof of this lemma is practically identical with the proof of \[[1](#bib.bib1), Lemma 1\]. I include a short proof for the convenience of the reader and some of the formulas we produce in the proof are useful for implementation.
######
Theorem 8.1
.
Let
<table>
<tbody>
<tr>
<td></td>
<td>C=1+ed−1,2+ed,1andD=(−1)d(e1,2−e2,3+∑i=3dei,i+1)𝐶1subscript𝑒𝑑12subscript𝑒𝑑1and𝐷superscript1𝑑subscript𝑒12subscript𝑒23superscriptsubscript𝑖3𝑑subscript𝑒𝑖𝑖1C=1+e_{d-1,2}+e_{d,1}\;\;\;\text{and}\;\;\;D=(-1)^{d}\left(e_{1,2}-e_{2,3}+\sum\limits_{i=3}^{d}e_{i,i+1}\right)</td>
<td></td>
</tr>
</tbody>
</table>
be elements of SL$`(d,p)`$ where $`d \geq 5`$. Then $`C`$ and $`D`$ generates SL$`(d,p)`$.
###### Proof.
Let $`G_{0}`$ be the subgroup of SL$`(d,p)`$ generated by $`C`$ and $`D`$. I will now write down a few formulas, which follow from direct computation. For $`2 \leq k \leq {d - 2}`$ we have
<table>
<tbody>
<tr>
<td>(12)</td>
<td></td>
<td></td>
<td>D−1=superscript𝐷1absent\displaystyle D^{-1}=</td>
<td>(−1)d(e2,1−e3,2+∑i=3dei+1,i)superscript1𝑑subscript𝑒21subscript𝑒32superscriptsubscript𝑖3𝑑subscript𝑒𝑖1𝑖\displaystyle(-1)^{d}\left(e_{2,1}-e_{3,2}+\sum\limits_{i=3}^{d}e_{i+1,i}\right)</td>
<td></td>
</tr>
</tbody>
<tbody>
<tr>
<td>(13)</td>
<td></td>
<td></td>
<td>C1=subscript𝐶1absent\displaystyle C_{1}=</td>
<td>D−1CD=1−ed,3+e1,2superscript𝐷1𝐶𝐷1subscript𝑒𝑑3subscript𝑒12\displaystyle D^{-1}CD=1-e_{d,3}+e_{1,2}</td>
<td></td>
</tr>
</tbody>
</table>
<table>
<tbody>
<tr>
<td>(14)</td>
<td></td>
<td></td>
<td>CC1C−1C1−1=𝐶subscript𝐶1superscript𝐶1superscriptsubscript𝐶11absent\displaystyle CC_{1}C^{-1}C_{1}^{-1}=</td>
<td>1+ed,21subscript𝑒𝑑2\displaystyle 1+e_{d,2}</td>
<td></td>
</tr>
</tbody>
</table>
<table>
<tbody>
<tr>
<td>(15)</td>
<td></td>
<td>Dk=(−1)dk(−e1,1+k−e2,2+k+∑i=3dei,i+k)superscript𝐷𝑘superscript1𝑑𝑘subscript𝑒11𝑘subscript𝑒22𝑘superscriptsubscript𝑖3𝑑subscript𝑒𝑖𝑖𝑘\displaystyle D^{k}=(-1)^{dk}\left(-e_{1,1+k}-e_{2,2+k}+\sum\limits_{i=3}^{d}e_{i,i+k}\right)</td>
<td></td>
</tr>
</tbody>
<tbody>
<tr>
<td>(16)</td>
<td></td>
<td>D−k=(−1)dk(−e1+k,1−e2+k,2+∑i=3dei+k,i)superscript𝐷𝑘superscript1𝑑𝑘subscript𝑒1𝑘1subscript𝑒2𝑘2superscriptsubscript𝑖3𝑑subscript𝑒𝑖𝑘𝑖\displaystyle D^{-k}=(-1)^{dk}\left(-e_{1+k,1}-e_{2+k,2}+\sum\limits_{i=3}^{d}e_{i+k,i}\right)</td>
<td></td>
</tr>
</tbody>
</table>
<table>
<tbody>
<tr>
<td>(17)</td>
<td></td>
<td>Ck=D−kCDk=1−ek−1,k+2−ek,k+1subscript𝐶𝑘superscript𝐷𝑘𝐶superscript𝐷𝑘1subscript𝑒𝑘1𝑘2subscript𝑒𝑘𝑘1\displaystyle C_{k}=D^{-k}CD^{k}=1-e_{k-1,k+2}-e_{k,k+1}</td>
<td></td>
</tr>
</tbody>
<tbody>
<tr>
<td>(18)</td>
<td></td>
<td>Ck−1=1+ek−1,k+2+ek,k+1superscriptsubscript𝐶𝑘11subscript𝑒𝑘1𝑘2subscript𝑒𝑘𝑘1\displaystyle C_{k}^{-1}=1+e_{k-1,k+2}+e_{k,k+1}</td>
<td></td>
</tr>
</tbody>
<tbody>
<tr>
<td>(19)</td>
<td></td>
<td>(1+ed,k)Ck(1−ed,k)Ck−1=1−ed,k+11subscript𝑒𝑑𝑘subscript𝐶𝑘1subscript𝑒𝑑𝑘superscriptsubscript𝐶𝑘11subscript𝑒𝑑𝑘1\displaystyle\left(1+e_{d,k}\right)C_{k}\left(1-e_{d,k}\right)C_{k}^{-1}=1-e_{d,k+1}</td>
<td></td>
</tr>
</tbody>
</table>
From Equation (11) we see that $`1 + e_{d,2}`$ belongs to $`G_{0}`$ and then we use mathematical induction on $`k`$ and Equation (16) proves that $`{1 + e_{d,k}} \in G_{0}`$ for $`k = {2,\ldots,{d - 1}}`$. Also $`{D^{- 2}\hspace{0pt}\left( {1 + e_{d,{d - 1}}} \right)\hspace{0pt}D^{2}} = {1 + e_{2,1}} \in G_{0}`$. Furthermore $`\left\lbrack {1 + e_{d,2}},{1 + e_{2,1}} \right\rbrack = {1 + e_{d,1}}`$. This proves that $`{1 + e_{d,k}} \in G_{0}`$ for $`k = {1,2,\ldots,{d - 1}}`$. Then we can use the relations in SL$`(d,p)`$ to prove that $`{1 + e_{i,j}} \in G_{0}`$ for $`{i,j} \in {\{ 1,2,\ldots,d\}}`$ and $`i \neq j`$. This proves the theorem. ∎
The proof of the theorem is constructive. It gives us a way to compute the elementary transvections from these generators of Albert and Thomson; one can use them effectively to publish the public key. There will be some precomputation involved to change the action of $`\phi`$ from these generators to elementary transvections.
## 9. Conclusions
This paper studies the MOR cryptosystem for the special linear group over finite fields. Cryptography is primarily driven by applicability. So it is natural to ask, how efficiently can one implement this MOR cryptosystem? How secure is the cryptosystem? I talked in details on both these issues in Sections 8 and 7 respectively. These are often hard questions to answer from a preliminary investigation. The worst case complexity is often far off from the actual cost of computation and security in itself is a very elusive concept. We now offer some realistic expectations on the computational cost of this MOR cryptosystem when $`q = 2^{\gamma}`$.
From the small experiments we did, it seems reasonable to assume that a randomly chosen element of SL$`(d,q)`$ is generated by approximately $`d`$ elementary transvections, not $`d^{2}`$ elementary transvections. This story is also corroborated by the proof of the previous theorem, where we show that SL$`(d,p)`$ is generated by all transvections of the form $`1 + e_{d,k}`$, $`k = {1,2,\ldots,{d - 1}}`$ and by Humphries \[[7](#bib.bib7)\].
Then we need to compute the image of these $`d`$ elementary transvections under the automorphism $`\phi`$. For that we need to split each elementary transvections into product of elementary transvections over the ground field using Equation [5](#S3.E5 "In 3. The unimodular group of degree 𝑑 over 𝔽_𝑞 ‣ A simple generalization of the ElGamal cryptosystem to non-abelian groups II"). Then in the worst case we now have $`\gamma\hspace{0pt}d`$ elementary transvections. But since in any random binary string of length $`\gamma`$ on average there are utmost $`\frac{\gamma}{2}`$ ones. So a more realistic expectation of the number of transvections is $`\frac{\gamma}{2}\hspace{0pt}d`$. Using the same expectation as before the image of these transvections under $`\phi`$ will be a string of $`\frac{\gamma}{2}\hspace{0pt}d^{2}`$ elementary transvections. Now if we use a straight line program, i.e., use the elementary transvections to multiply the one next to it to form the matrix, then the worst case complexity will be $`\frac{\gamma}{2}\hspace{0pt}d^{3}`$ field multiplication. However, in reality that complexity will be something like $`\frac{\gamma}{2}\hspace{0pt}d^{\lambda}`$ where $`2 < \lambda \leq 3`$. So it is safe to assume that in practice $`\lambda`$ will be around $`2.5`$.
With all this understanding we can say that if $`q`$ is a field of characteristic $`2`$ and degree $`\gamma`$, then composition of two automorphisms require around
<table>
<tbody>
<tr>
<td></td>
<td>d2+γ2d2.5superscript𝑑2𝛾2superscript𝑑2.5d^{2}+\dfrac{\gamma}{2}d^{2.5}</td>
<td></td>
</tr>
</tbody>
</table>
field multiplications. If we were working with a finite field $`{\mathbb{F}}_{q^{d}}`$, then the naive product of two non-zero field element costs around $`d^{2}`$ field multiplications. We are quite close to that.
Lastly, I recommend that the plaintext should be an elementary transvection. It is known that trace and determinant is invariant under matrix conjugation. So the trace or the determinant can give out information about the plaintext. However, if it is an elementary transvection, then the trace is always $`d`$ and the determinant $`1`$.
## References
- \[1\]
A.A. Albert and John Thompson, *Two-element genration of the projective unimodular group*, Illionois Journal of Mathematics 3 (1959), 421–439.
- \[2\]
J.L. Alperin and Rowen B. Bell, *Groups and Representations*, Springer, 1995.
- \[3\]
Roger W. Carter, *Simple groups of Lie type*, John Willey & Sons, 1989.
- \[4\]
Don Coppersmith and Shmuel Winograd, *Matrix multiplication via arithmatic progression*, Proceedings of the nineteenth annual ACM conference on Theory of Computing, 1987, pp. 1–6.
- \[5\]
Jean Dieudonné and Loo-Keng Hua, *On the automorphisms of the classical groups*, Memoirs of the American Mathematical Society (1951), no. 2.
- \[6\]
The GAP Group, *GAP – Groups, Algorithms, and Programming, Version 4.4.10*, 2007.
- \[7\]
Stephen P. Humphries, *Generation of special linear groups by transvections*, Journal of Algebra 99 (1986), 480–495.
- \[8\]
Neal Koblitz, Alfred Menezes, and Scott Vanstone, *The state of elliptic curve cryptography*, Designs, Codes and Cryptogrpahy 19 (2000), 173–193.
- \[9\]
In-Sok Lee, Woo-Hwan Kim, Daesung Kwon, Sangil Nahm, Nam-Soek Kwak, and Yoo-Jin Baek, *On the security of MOR public key cryptosystem*, Asiacrypt 2004 (P.J.Lee, ed.), LNCS, no. 3329, Springer-Verlag, 2004, pp. 387–400.
- \[10\]
Ayan Mahalanobis, *A simple generalization of ElGamal cryptosystem to non-abelian groups*, Communication in Algebra 36 (2008), no. 10, 3878–3889.
- \[11\]
Alfred Menezes and Yi-Hong Wu, *The discrete logarithm problem in GL$`(n,q)`$*, Ars Combinatorica 47 (1997), 23–32.
- \[12\]
Seong-Hun Paeng, *On the security of cryptosystem using the automorphism groups*, Information Processing Letters 88 (2003), 293–298.
- \[13\]
Seong-Hun Paeng, Kil-Chan Ha, Jae Heon Kim, Seongtaek Chee, and Choonsik Park, *New public key cryptosystem using finite non-abelian groups*, Crypto 2001 (J. Kilian, ed.), LNCS, vol. 2139, Springer-Verlag, 2001, pp. 470–485.
- \[14\]
Joseph J. Rotman, *An introduction to the theory of groups*, 4 ed., Springer-Velag, 1994.
- \[15\]
Oliver Schirokauer, Damian Weber, and Thomas Denny, *Discrete logarithm: the effectiveness of the index calculus method*, Algorithmic number theory (Talence, 1996), LNCS, vol. 1122, 1996, pp. 337–361.
- \[16\]
Joseph Silverman and Joe Suzuki, *Elliptic curve discrete logarithms and the index calculus*, Asiacrypt’98 (K. Ohra and D. Pei, eds.), LNCS, vol. 1514, 1998, pp. 110–125.
- \[17\]
Robert Steinberg, *Automorphisms of finite linear groups*, Canadian Journal of Mathematics 12 (1960), 606–615.
- \[18\]
Douglas Stinson, *Cryptography theory and practice*, third ed., Chapman & Hall/CRC, 2006. |
0708.0522 | cs.NI | cyber | INSTITUT NATIONAL DE RECHERCHE EN INFORMATIQUE ET EN AUTOMATIQUE
Quasi-stationary distributions
as centrality measures of reducible graphs
Konstantin Avrachenkov — Vivek Borkar — Danil Nemirovsky N° 6263
August 2007
Quasi-stationary distributions
as centrality measures of reducible graphs
Konstantin Avrachenkov
<sup>†</sup>
<sup>†</sup>thanks: INRIA Sophia Antipolis, K.Avrachenkov@sophia.inria.fr
, Vivek Borkar
<sup>†</sup>
<sup>†</sup>thanks: Tata Institute of Fundamental Research, India, E-mail: borkar@tifr.res.in
, Danil Nemirovsky
<sup>†</sup>
<sup>†</sup>thanks: INRIA Sophia Antipolis, France and St. Petersburg State University, Russia, E-mail: danil.nemirovsky@sophia.inria.fr
Thème COM — Systèmes communicants
Projets MAESTRO
Rapport de recherche n° 6263 — August 2007 — ?? pages
<sup>0</sup>
<sup>0</sup>footnotetext: This research was supported by RIAM INRIA-Canon grant, European research project Bionets, and by CEFIPRA grant no-2900-IT.
Abstract: Random walk can be used as a centrality measure of a directed graph. However, if the graph is reducible the random walk will be absorbed in some subset of nodes and will never visit the rest of the graph. In Google PageRank the problem was solved by introduction of uniform random jumps with some probability. Up to the present, there is no clear criterion for the choice this parameter. We propose to use parameter-free centrality measure which is based on the notion of quasi-stationary distribution. Specifically we suggest four quasi-stationary based centrality measures, analyze them and conclude that they produce approximately the same ranking. The new centrality measures can be applied in spam detection to detect “link farms” and in image search to find photo albums.
Key-words: centrality measure, directed graph, quasi-stationary distribution, PageRank, Web graph, link farm
Distributions quasi-stationnaires
comme les mesures de centralité pour des graphes réductible
Résumé : Une marche au hasard peut être utilisée comme mesure de centralité d’un graphe orienté. Cependant, si le graphe est réductible la marche au hasard sera absorbée dans un quelque sous-ensemble de noeuds et ne visitera jamais le reste du graphe. Dans Google PageRank, le problème a été résolu par l’introduction des sauts aléatoires uniformes avec une certaine probabilité. Jusqu’à présent, il n’y a aucun critère clair pour le choix de ce paramètre. Nous proposons d’utiliser la mesure de centralité sans paramètre qui est basée sur la notion de la distribution quasi-stationnaire. Nous analysons les quatre mesures et concluons qu’elles produisent presque le même classement de noeuds. Les nouvelles mesures de centralité peuvent être appliquées dans le context de la détection de spam pour détecter les “link farms” et dans le context de la recherche d’image pour trouver des albums photo.
Mots-clés : mesure de centralité, marche au hasard, graphe orienté, distribution quasi-stationnaire, PageRank, graphe du Web, link farm
## 1 Introduction
Random walk can be used as a centrality measure of a directed graph. An example of random walk based centrality measures is PageRank \[[21](#bib.bib21)\] used by search engine Google. PageRank is used by Google to sort the relevant answers to user’s query. We shall follow the formal definition of PageRank from \[[18](#bib.bib18)\]. Denote by $`n`$ the total number of pages on the Web and define the $`n \times n`$ hyperlink matrix $`P`$ such that
<table>
<tbody>
<tr>
<td></td>
<td>pij={1/di,if page i links to j,1/n,if page i is dangling,0,otherwise,subscript𝑝𝑖𝑗cases1subscript𝑑𝑖if page i links to j1𝑛if page i is dangling0otherwisep_{ij}=\left\{\begin{array}[]{ll}1/d_{i},&\mbox{if page $i$ links to $j$},\\ 1/n,&\mbox{if page $i$ is dangling},\\ 0,&\mbox{otherwise},\end{array}\right.</td>
<td></td>
<td>(1)</td>
</tr>
</tbody>
</table>
for $`{{i,j} = 1},{\ldots,n}`$, where $`d_{i}`$ is the number of outgoing links from page $`i`$. A page with no outgoing links is called dangling. We note that according to ([1](#S1.E1 "In 1 Introduction")) there exist artificial links to all pages from a dangling node. In order to make the hyperlink graph connected, it is assumed that at each step, with some probability $`c`$, a random surfer goes to an arbitrary Web page sampled from the uniform distribution. Thus, the PageRank is defined as a stationary distribution of a Markov chain whose state space is the set of all Web pages, and the transition matrix is
<table>
<tbody>
<tr>
<td></td>
<td>G=cP+(1−c)(1/n)E,𝐺𝑐𝑃1𝑐1𝑛𝐸G=cP+(1-c)(1/n)E,</td>
<td></td>
</tr>
</tbody>
</table>
where $`E`$ is a matrix whose all entries are equal to one, and $`c \in {(0,1)}`$ is a probability of following a hyperlink. The constant $`c`$ is often referred to as a damping factor. The Google matrix $`G`$ is stochastic, aperiodic, and irreducible, so the PageRank vector $`\pi`$ is the unique solution of the system
<table>
<tbody>
<tr>
<td></td>
<td>πG=π,π𝟏=1,formulae-sequence𝜋𝐺𝜋𝜋11\pi G=\pi,\quad\pi{\bf 1}=1,</td>
<td></td>
</tr>
</tbody>
</table>
where $`\mathbf{1}`$ is a column vector of ones.
Even though in a number of recent works, see e.g., \[[5](#bib.bib5), [6](#bib.bib6), [8](#bib.bib8)\], the choice of the damping factor $`c`$ has been discussed, there is still no clear criterion for the choice of its value. The goal of the present work is to explore parameter-free centrality measures.
In \[[5](#bib.bib5), [7](#bib.bib7), [15](#bib.bib15)\] the authors have studied the graph structure of the Web. In particular, in \[[7](#bib.bib7), [15](#bib.bib15)\] it was shown that the Web Graph can be divided into three principle components: the Giant Strongly Connected Component, to which we simply refer as SCC component, the IN component and the OUT component. The SCC component is the largest strongly connected component in the Web Graph. In fact, it is larger than the second largest strongly connected component by several orders of magnitude. Following hyperlinks one can come from the IN component to the SCC component but it is not possible to return back. Then, from the SCC component one can come to the OUT component and it is not possible to return to SCC from the OUT component. In \[[7](#bib.bib7), [15](#bib.bib15)\] the analysis of the structure of the Web was made assuming that dangling nodes have no outgoing links. However, according to ([1](#S1.E1 "In 1 Introduction")) there is a probability to jump from a dangling node to an arbitrary node. This can be viewed as a link between the nodes and we call such a link the artificial link. As was shown in \[[5](#bib.bib5)\], these artificial links significantly change the graph structure of the Web. In particular, the artificial links of dangling nodes in the OUT component connect some parts of the OUT component with IN and SCC components. Thus, the size of the Giant Strongly Connected Component increases further. If the artificial links from dangling nodes are taken into account, it is shown in \[[5](#bib.bib5)\] that the Web Graph can be divided in two disjoint components: Extended Strongly Connected Component (ESCC) and Pure OUT (POUT) component. The POUT component is small in size but if the damping factor $`c`$ is chosen equal to one, the random walk absorbs with probability one into POUT. We note that nearly all important pages are in ESCC. We also note that even if the damping factor is chosen close to one, the random walk can spend a significant amount of time in ESCC before the absorption. Therefore, for ranking Web pages from ESCC we suggest to use the quasi-stationary distributions \[[9](#bib.bib9), [22](#bib.bib22)\].
It turns out that there are several versions of quasi-stationary distribution. Here we study four versions of the quasi-stationary distribution. Our main conclusion is that the rankings provided by them are very similar. Therefore, one can chose a version of stationary distribution which is easier for computation.
The paper is organized as follows: In the next Section [2](#S2 "2 Quasi-stationary distributions as centrality measures") we discuss different notions of quasi-stationarity, the relation among them, and the relation between the quasi-stationary distribution and PageRank. Then, in Section [3](#S3 "3 Numerical experiments and Applications") we present the results of numerical experiments on Web Graph which confirm our theoretical findings and suggest the application of quasi-stationarity based centrality measures to link spam detection and image search. Some technical results we place in the Appendix.
## 2 Quasi-stationary distributions as centrality measures
As noted in \[[5](#bib.bib5)\], by renumbering the nodes the transition matrix $`P`$ can be transformed to the following form
<table>
<tbody>
<tr>
<td></td>
<td>P=[Q0RT],𝑃delimited-[]𝑄0𝑅𝑇P=\left[\begin{array}[]{cc}Q&0\\ R&T\end{array}\right],</td>
<td></td>
</tr>
</tbody>
</table>
where the block $`T`$ corresponds to the ESCC, the block $`Q`$ corresponds to the part of the OUT component without dangling nodes and their predecessors, and the block $`R`$ corresponds to the transitions from ESCC to the nodes in block $`Q`$. We refer to the set of nodes in the block $`Q`$ as POUT component.
The POUT component is small in size but if the damping factor $`c`$ is chosen equal to one, the random walk absorbs with probability one into POUT. We are mostly interested in the nodes in the ESCC component. Denote by $`\pi_{Q}`$ a part of the PageRank vector corresponding to the POUT component and denote by $`\pi_{T}`$ a part of the PageRank vector corresponding to the ESCC component. Using the following formula \[[20](#bib.bib20)\]
<table>
<tbody>
<tr>
<td></td>
<td>π(c)=1−cn𝟏T[I−cP]−1,𝜋𝑐1𝑐𝑛superscript1𝑇superscriptdelimited-[]𝐼𝑐𝑃1\pi(c)=\frac{1-c}{n}{\bf 1}^{T}[I-cP]^{-1},</td>
<td></td>
</tr>
</tbody>
</table>
we conclude that
<table>
<tbody>
<tr>
<td></td>
<td>πT(c)=1−cn𝟏T[I−cT]−1,subscript𝜋𝑇𝑐1𝑐𝑛superscript1𝑇superscriptdelimited-[]𝐼𝑐𝑇1\pi_{T}(c)=\frac{1-c}{n}{\bf 1}^{T}[I-cT]^{-1},</td>
<td></td>
</tr>
</tbody>
</table>
where $`\mathbf{1}`$ is a vector of ones of appropriate dimension.
Let us define
<table>
<tbody>
<tr>
<td></td>
<td>π^T(c)=πT(c)‖πT(c)‖1.subscript^𝜋𝑇𝑐subscript𝜋𝑇𝑐subscriptnormsubscript𝜋𝑇𝑐1\hat{\pi}_{T}(c)=\frac{\pi_{T}(c)}{||\pi_{T}(c)||_{1}}.</td>
<td></td>
</tr>
</tbody>
</table>
Since the matrix $`T`$ is substochastic, we have the next result.
######
Proposition 1
The following limit exists
<table>
<tbody>
<tr>
<td></td>
<td>π^T(1)=limc→1πT(c)‖πT(c)‖1=𝟏T[I−T]−1𝟏T[I−T]−1𝟏,subscript^𝜋𝑇1subscript→𝑐1subscript𝜋𝑇𝑐subscriptnormsubscript𝜋𝑇𝑐1superscript1𝑇superscriptdelimited-[]𝐼𝑇1superscript1𝑇superscriptdelimited-[]𝐼𝑇11\hat{\pi}_{T}(1)=\lim_{c\to 1}\frac{\pi_{T}(c)}{||\pi_{T}(c)||_{1}}=\frac{{\bf 1}^{T}[I-T]^{-1}}{{\bf 1}^{T}[I-T]^{-1}{\bf 1}},</td>
<td></td>
</tr>
</tbody>
</table>
and the ranking of pages in ESCC provided by the PageRank vector converges to the ranking provided by $`{\hat{\pi}}_{T}\hspace{0pt}{(1)}`$ as the damping factor goes to one. Moreover, these two rankings coincide for all values of $`c`$ above some value $`c^{\ast}`$.
Next we denote $`{\hat{\pi}}_{T}\hspace{0pt}{(1)}`$ simply by $`{\hat{\pi}}_{T}`$. Following \[[9](#bib.bib9), [12](#bib.bib12)\] we shall call the vector $`{\hat{\pi}}_{T}`$ pseudo-stationary distribution. The $`i^{\text{th}}`$ component of $`{\hat{\pi}}_{T}`$ can be interpreted as a fraction of time the random walk (with $`c = 1`$) spends in node $`i`$ prior to absorption. We recall that the random walk as defined in Introduction starts from the uniform distribution. If the random walk were initiated from another distribution, the pseudo-stationary distribution would change.
Denote by $`\overline{T}`$ the hyperlink matrix associated with ESCC when the links leading outside of ESCC are neglected. Clearly, we have
<table>
<tbody>
<tr>
<td></td>
<td>T¯ij=Tij[T𝟏]i,subscript¯𝑇𝑖𝑗subscript𝑇𝑖𝑗subscriptdelimited-[]𝑇1𝑖\bar{T}_{ij}=\frac{T_{ij}}{[T{\bf 1}]_{i}},</td>
<td></td>
</tr>
</tbody>
</table>
where $`{\lbrack{T\hspace{0pt}\mathbf{1}}\rbrack}_{i}`$ denotes the $`i^{\text{th}}`$ component of vector $`T\hspace{0pt}\mathbf{1}`$. In other words, $`{\lbrack{T\hspace{0pt}\mathbf{1}}\rbrack}_{i}`$ is the sum of elements in row $`i`$ of matrix $`T`$. The $`{\overline{T}}_{i\hspace{0pt}j}`$ entry of the matrix $`\overline{T}`$ can be considered as a conditional probability to jump from the node $`i`$ to the node $`j`$ under the condition that random walk does not leave ESCC at the jump. Let $`{\overline{\pi}}_{T}`$ be a stationary distribution of $`\overline{T}`$.
Let us now consider the substochastic matrix $`T`$ as a perturbation of stochastic matrix $`\overline{T}`$. We introduce the perturbation term
<table>
<tbody>
<tr>
<td></td>
<td>εD=T¯−T,𝜀𝐷¯𝑇𝑇\varepsilon D=\bar{T}-T,</td>
<td></td>
</tr>
</tbody>
</table>
where the parameter $`\varepsilon`$ is the perturbation parameter, which is typically small. The following result holds.
######
Proposition 2
The vector $`{\hat{\pi}}_{T}`$ is close to $`{\overline{\pi}}_{T}`$. Namely,
<table>
<tbody>
<tr>
<td></td>
<td>π^T=π¯T−π¯T1nT(π¯TεD𝟏)𝟏TX0𝟏+𝟏TX01nT(π¯TεD𝟏)+o(ε),subscript^𝜋𝑇subscript¯𝜋𝑇subscript¯𝜋𝑇1subscript𝑛𝑇subscript¯𝜋𝑇𝜀𝐷1superscript1𝑇subscript𝑋01superscript1𝑇subscript𝑋01subscript𝑛𝑇subscript¯𝜋𝑇𝜀𝐷1𝑜𝜀\hat{\pi}_{T}=\bar{\pi}_{T}-\bar{\pi}_{T}\frac{1}{n_{T}}(\bar{\pi}_{T}\varepsilon D{\bf 1}){\bf 1}^{T}X_{0}{\bf 1}+{\bf 1}^{T}X_{0}\frac{1}{n_{T}}(\bar{\pi}_{T}\varepsilon D{\bf 1})+o(\varepsilon),</td>
<td></td>
<td>(2)</td>
</tr>
</tbody>
</table>
where $`n_{T}`$ is the number of nodes in ESCC and $`X_{0}`$ is given in Lemma [1](#Thmlemma1 "Lemma 1") from the Appendix.
Proof: We substitute $`T = {\overline{T} - {\varepsilon\hspace{0pt}D}}`$ into $`{\lbrack{I - T}\rbrack}^{- 1}`$ and use Lemma [1](#Thmlemma1 "Lemma 1"), to get
<table>
<tbody>
<tr>
<td></td>
<td>[I−T]−1=1π¯εD𝟏𝟏π¯+X0+O(ε).superscriptdelimited-[]𝐼𝑇11¯𝜋𝜀𝐷11¯𝜋subscript𝑋0𝑂𝜀[I-T]^{-1}=\frac{1}{\bar{\pi}\varepsilon D{\bf 1}}{\bf 1}\bar{\pi}+X_{0}+O(\varepsilon).</td>
<td></td>
</tr>
</tbody>
</table>
Using the above expression, we can write
<table>
<tbody>
<tr>
<td></td>
<td>π^T=𝟏T[I−T]−1𝟏T[I−T]−1𝟏=1π¯TεD𝟏nTπ¯T+𝟏TX0+O(ε)1π¯TεD𝟏nT+𝟏TX0𝟏+O(ε)=π¯T+1nT(π¯TεD𝟏)𝟏TX0+o(ε)1+1nT(π¯TεD𝟏)𝟏TX0𝟏+o(ε)subscript^𝜋𝑇superscript1𝑇superscriptdelimited-[]𝐼𝑇1superscript1𝑇superscriptdelimited-[]𝐼𝑇111subscript¯𝜋𝑇𝜀𝐷1subscript𝑛𝑇subscript¯𝜋𝑇superscript1𝑇subscript𝑋0𝑂𝜀1subscript¯𝜋𝑇𝜀𝐷1subscript𝑛𝑇superscript1𝑇subscript𝑋01𝑂𝜀subscript¯𝜋𝑇1subscript𝑛𝑇subscript¯𝜋𝑇𝜀𝐷1superscript1𝑇subscript𝑋0𝑜𝜀11subscript𝑛𝑇subscript¯𝜋𝑇𝜀𝐷1superscript1𝑇subscript𝑋01𝑜𝜀\hat{\pi}_{T}=\frac{{\bf 1}^{T}[I-T]^{-1}}{{\bf 1}^{T}[I-T]^{-1}{\bf 1}}=\frac{\frac{1}{\bar{\pi}_{T}\varepsilon D{\bf 1}}n_{T}\bar{\pi}_{T}+{\bf 1}^{T}X_{0}+O(\varepsilon)}{\frac{1}{\bar{\pi}_{T}\varepsilon D{\bf 1}}n_{T}+{\bf 1}^{T}X_{0}{\bf 1}+O(\varepsilon)}=\frac{\bar{\pi}_{T}+\frac{1}{n_{T}}(\bar{\pi}_{T}\varepsilon D{\bf 1}){\bf 1}^{T}X_{0}+o(\varepsilon)}{1+\frac{1}{n_{T}}(\bar{\pi}_{T}\varepsilon D{\bf 1}){\bf 1}^{T}X_{0}{\bf 1}+o(\varepsilon)}</td>
<td></td>
</tr>
</tbody>
</table>
<table>
<tbody>
<tr>
<td></td>
<td>=(π¯T+1nT(π¯TεD𝟏)𝟏TX0+o(ε))(1−1nT(π¯TεD𝟏)𝟏TX0𝟏+o(ε))absentsubscript¯𝜋𝑇1subscript𝑛𝑇subscript¯𝜋𝑇𝜀𝐷1superscript1𝑇subscript𝑋0𝑜𝜀11subscript𝑛𝑇subscript¯𝜋𝑇𝜀𝐷1superscript1𝑇subscript𝑋01𝑜𝜀=\left(\bar{\pi}_{T}+\frac{1}{n_{T}}(\bar{\pi}_{T}\varepsilon D{\bf 1}){\bf 1}^{T}X_{0}+o(\varepsilon)\right)\left(1-\frac{1}{n_{T}}(\bar{\pi}_{T}\varepsilon D{\bf 1}){\bf 1}^{T}X_{0}{\bf 1}+o(\varepsilon)\right)</td>
<td></td>
</tr>
</tbody>
</table>
<table>
<tbody>
<tr>
<td></td>
<td>=π¯T−π¯T1nT(π¯TεD𝟏)𝟏TX0𝟏+𝟏TX01nT(π¯TεD𝟏)+o(ε).absentsubscript¯𝜋𝑇subscript¯𝜋𝑇1subscript𝑛𝑇subscript¯𝜋𝑇𝜀𝐷1superscript1𝑇subscript𝑋01superscript1𝑇subscript𝑋01subscript𝑛𝑇subscript¯𝜋𝑇𝜀𝐷1𝑜𝜀=\bar{\pi}_{T}-\bar{\pi}_{T}\frac{1}{n_{T}}(\bar{\pi}_{T}\varepsilon D{\bf 1}){\bf 1}^{T}X_{0}{\bf 1}+{\bf 1}^{T}X_{0}\frac{1}{n_{T}}(\bar{\pi}_{T}\varepsilon D{\bf 1})+o(\varepsilon).</td>
<td></td>
</tr>
</tbody>
</table>
$`\square`$
Since $`{{R\hspace{0pt}\mathbf{1}} + {T\hspace{0pt}\mathbf{1}}} = \mathbf{1}`$ and $`{\overline{T}\hspace{0pt}\mathbf{1}} = \mathbf{1}`$, in lieu of $`{\overline{\pi}}_{T}\hspace{0pt}\varepsilon\hspace{0pt}D\hspace{0pt}\mathbf{1}`$ we can write $`{\overline{\pi}}_{T}\hspace{0pt}R\hspace{0pt}\mathbf{1}`$. The latter expression has a clear probabilistic interpretation. It is a probability to exit ESCC in one step starting from the distribution $`{\overline{\pi}}_{T}`$. Later we shall demonstrate that this probability is indeed small. We note that not only $`{\overline{\pi}}_{T}\hspace{0pt}R\hspace{0pt}\mathbf{1}`$ is small but also the factor $`1/n_{T}`$ is small, as the number of states in ESCC is large.
In the next Proposition [3](#Thmproposition3 "Proposition 3 ‣ 2 Quasi-stationary distributions as centrality measures") we provide alternative expression for the first order terms of $`{\hat{\pi}}_{T}`$.
######
Proposition 3
<table>
<tbody>
<tr>
<td></td>
<td>π^T=π¯T−επ¯TDH+ε𝟏T1nT(π¯TD𝟏)H+o(ε).subscript^𝜋𝑇subscript¯𝜋𝑇𝜀subscript¯𝜋𝑇𝐷𝐻𝜀superscript1𝑇1subscript𝑛𝑇subscript¯𝜋𝑇𝐷1𝐻𝑜𝜀\hat{\pi}_{T}=\bar{\pi}_{T}-\varepsilon\bar{\pi}_{T}DH+\varepsilon{\bf 1}^{T}\frac{1}{n_{T}}(\bar{\pi}_{T}D{\bf 1})H+o(\varepsilon).</td>
<td></td>
</tr>
</tbody>
</table>
Proof: Let us consider $`{\hat{\pi}}_{T}`$ as power series:
<table>
<tbody>
<tr>
<td></td>
<td>π^T=π^T(0)+επ^T(1)+ε2π^T(2)+….subscript^𝜋𝑇superscriptsubscript^𝜋𝑇0𝜀superscriptsubscript^𝜋𝑇1superscript𝜀2superscriptsubscript^𝜋𝑇2…\displaystyle\hat{\pi}_{T}=\hat{\pi}_{T}^{(0)}+\varepsilon\hat{\pi}_{T}^{(1)}+\varepsilon^{2}\hat{\pi}_{T}^{(2)}+\ldots.</td>
<td></td>
</tr>
</tbody>
</table>
From ([2](#S2.E2 "In Proposition 2 ‣ 2 Quasi-stationary distributions as centrality measures")) we obtain
<table>
<tbody>
<tr>
<td></td>
<td>π^Tsubscript^𝜋𝑇\displaystyle\hat{\pi}_{T}</td>
<td>=\displaystyle=</td>
<td>π¯T−π¯T1nT(π¯TεD𝟏)𝟏TX0𝟏+𝟏TX01nT(π¯TεD𝟏)+o(ε)=subscript¯𝜋𝑇subscript¯𝜋𝑇1subscript𝑛𝑇subscript¯𝜋𝑇𝜀𝐷1superscript1𝑇subscript𝑋01superscript1𝑇subscript𝑋01subscript𝑛𝑇subscript¯𝜋𝑇𝜀𝐷1𝑜𝜀absent\displaystyle\bar{\pi}_{T}-\bar{\pi}_{T}\frac{1}{n_{T}}(\bar{\pi}_{T}\varepsilon D{\bf 1}){\bf 1}^{T}X_{0}{\bf 1}+{\bf 1}^{T}X_{0}\frac{1}{n_{T}}(\bar{\pi}_{T}\varepsilon D{\bf 1})+o(\varepsilon)=</td>
<td></td>
</tr>
<tr>
<td></td>
<td></td>
<td>=\displaystyle=</td>
<td>π¯T+ε(𝟏TX01nT(π¯TD𝟏)−π¯T1nT(π¯TD𝟏)𝟏TX0𝟏)+o(ε),subscript¯𝜋𝑇𝜀superscript1𝑇subscript𝑋01subscript𝑛𝑇subscript¯𝜋𝑇𝐷1subscript¯𝜋𝑇1subscript𝑛𝑇subscript¯𝜋𝑇𝐷1superscript1𝑇subscript𝑋01𝑜𝜀\displaystyle\bar{\pi}_{T}+\varepsilon\left({\bf 1}^{T}X_{0}\frac{1}{n_{T}}(\bar{\pi}_{T}D{\bf 1})-\bar{\pi}_{T}\frac{1}{n_{T}}(\bar{\pi}_{T}D{\bf 1}){\bf 1}^{T}X_{0}{\bf 1}\right)+o(\varepsilon),</td>
<td></td>
</tr>
</tbody>
</table>
and hence
<table>
<tbody>
<tr>
<td></td>
<td>π^T(1)=𝟏TX01nT(π¯TD𝟏)−π¯T1nT(π¯TD𝟏)𝟏TX0𝟏,superscriptsubscript^𝜋𝑇1superscript1𝑇subscript𝑋01subscript𝑛𝑇subscript¯𝜋𝑇𝐷1subscript¯𝜋𝑇1subscript𝑛𝑇subscript¯𝜋𝑇𝐷1superscript1𝑇subscript𝑋01\hat{\pi}_{T}^{(1)}={\bf 1}^{T}X_{0}\frac{1}{n_{T}}(\bar{\pi}_{T}D{\bf 1})-\bar{\pi}_{T}\frac{1}{n_{T}}(\bar{\pi}_{T}D{\bf 1}){\bf 1}^{T}X_{0}{\bf 1},</td>
<td></td>
<td>(3)</td>
</tr>
</tbody>
</table>
where $`X_{0}`$ is given by ([30](#S4.E30 "In Lemma 1")). Before substituting ([30](#S4.E30 "In Lemma 1")) into ([3](#S2.E3 "In 2 Quasi-stationary distributions as centrality measures")) let us make transformations
<table>
<tbody>
<tr>
<td></td>
<td>X0subscript𝑋0\displaystyle X_{0}</td>
<td>=\displaystyle=</td>
<td>(I−X−1D)H(I−DX−1)=𝐼subscript𝑋1𝐷𝐻𝐼𝐷subscript𝑋1absent\displaystyle(I-X_{-1}D)H(I-DX_{-1})=</td>
<td></td>
</tr>
<tr>
<td></td>
<td></td>
<td>=\displaystyle=</td>
<td>H−HDX−1−X−1DH+X−1DHDX−1,𝐻𝐻𝐷subscript𝑋1subscript𝑋1𝐷𝐻subscript𝑋1𝐷𝐻𝐷subscript𝑋1\displaystyle H-HDX_{-1}-X_{-1}DH+X_{-1}DHDX_{-1},</td>
<td></td>
</tr>
</tbody>
</table>
where $`X_{- 1}`$ is defined by ([29](#S4.E29 "In Lemma 1")). Pre-multiplying $`X_{0}`$ by $`\mathbf{1}^{T}`$, we obtain
<table>
<tbody>
<tr>
<td></td>
<td>𝟏TX0superscript1𝑇subscript𝑋0\displaystyle{\bf 1}^{T}X_{0}</td>
<td>=\displaystyle=</td>
<td>𝟏TH−π¯T(𝟏THD𝟏)(π¯TD𝟏)−1−nTπ¯T(π¯TD𝟏)−1DH+superscript1𝑇𝐻subscript¯𝜋𝑇superscript1𝑇𝐻𝐷1superscriptsubscript¯𝜋𝑇𝐷11limit-fromsubscript𝑛𝑇subscript¯𝜋𝑇superscriptsubscript¯𝜋𝑇𝐷11𝐷𝐻\displaystyle{\bf 1}^{T}H-\bar{\pi}_{T}({\bf 1}^{T}HD{\bf 1})(\bar{\pi}_{T}D{\bf 1})^{-1}-n_{T}\bar{\pi}_{T}(\bar{\pi}_{T}D{\bf 1})^{-1}DH+</td>
<td></td>
</tr>
<tr>
<td></td>
<td></td>
<td>+\displaystyle+</td>
<td>nTπ¯TDHD𝟏π¯T(π¯TD𝟏)−2.subscript𝑛𝑇subscript¯𝜋𝑇𝐷𝐻𝐷1subscript¯𝜋𝑇superscriptsubscript¯𝜋𝑇𝐷12\displaystyle n_{T}\bar{\pi}_{T}DHD{\bf 1}\bar{\pi}_{T}(\bar{\pi}_{T}D{\bf 1})^{-2}.</td>
<td></td>
</tr>
</tbody>
</table>
Post-multiplying $`X_{0}`$ by $`\mathbf{1}`$, we obtain
<table>
<tbody>
<tr>
<td></td>
<td>X0𝟏subscript𝑋01\displaystyle X_{0}{\bf 1}</td>
<td>=\displaystyle=</td>
<td>X−1DHDX−1𝟏−HDX−1𝟏subscript𝑋1𝐷𝐻𝐷subscript𝑋11𝐻𝐷subscript𝑋11\displaystyle X_{-1}DHDX_{-1}{\bf 1}-HDX_{-1}{\bf 1}</td>
<td></td>
</tr>
</tbody>
</table>
and hence
<table>
<tbody>
<tr>
<td></td>
<td>𝟏TX0𝟏superscript1𝑇subscript𝑋01\displaystyle{\bf 1}^{T}X_{0}{\bf 1}</td>
<td>=\displaystyle=</td>
<td>nTπ¯TDHD𝟏(π¯TD𝟏)−2−𝟏THD𝟏(π¯TD𝟏)−1.subscript𝑛𝑇subscript¯𝜋𝑇𝐷𝐻𝐷1superscriptsubscript¯𝜋𝑇𝐷12superscript1𝑇𝐻𝐷1superscriptsubscript¯𝜋𝑇𝐷11\displaystyle n_{T}\bar{\pi}_{T}DHD{\bf 1}(\bar{\pi}_{T}D{\bf 1})^{-2}-{\bf 1}^{T}HD{\bf 1}(\bar{\pi}_{T}D{\bf 1})^{-1}.</td>
<td></td>
<td>(5)</td>
</tr>
</tbody>
</table>
Substituting ([5](#S2.E5 "In 2 Quasi-stationary distributions as centrality measures")) and ([2](#S2.Ex20 "2 Quasi-stationary distributions as centrality measures")) into ([3](#S2.E3 "In 2 Quasi-stationary distributions as centrality measures")), we get
<table>
<tbody>
<tr>
<td></td>
<td>π^T(1)superscriptsubscript^𝜋𝑇1\displaystyle\hat{\pi}_{T}^{(1)}</td>
<td>=\displaystyle=</td>
<td>𝟏TX01nT(π¯TD𝟏)−π¯T1nT(π¯TD𝟏)𝟏TX0𝟏=superscript1𝑇subscript𝑋01subscript𝑛𝑇subscript¯𝜋𝑇𝐷1subscript¯𝜋𝑇1subscript𝑛𝑇subscript¯𝜋𝑇𝐷1superscript1𝑇subscript𝑋01absent\displaystyle{\bf 1}^{T}X_{0}\frac{1}{n_{T}}(\bar{\pi}_{T}D{\bf 1})-\bar{\pi}_{T}\frac{1}{n_{T}}(\bar{\pi}_{T}D{\bf 1}){\bf 1}^{T}X_{0}{\bf 1}=</td>
<td></td>
</tr>
<tr>
<td></td>
<td></td>
<td>=\displaystyle=</td>
<td>𝟏TH1nT(π¯TD𝟏)−1nTπ¯T𝟏THD𝟏−π¯TDH+superscript1𝑇𝐻1subscript𝑛𝑇subscript¯𝜋𝑇𝐷11subscript𝑛𝑇subscript¯𝜋𝑇subscript1𝑇𝐻𝐷1limit-fromsubscript¯𝜋𝑇𝐷𝐻\displaystyle{\bf 1}^{T}H\frac{1}{n_{T}}(\bar{\pi}_{T}D{\bf 1})-\frac{1}{n_{T}}\bar{\pi}_{T}{\bf 1}_{T}HD{\bf 1}-\bar{\pi}_{T}DH+</td>
<td></td>
</tr>
<tr>
<td></td>
<td></td>
<td>+\displaystyle+</td>
<td>π¯T(π¯TDHD𝟏)(π¯TD𝟏)−1−π¯T(π¯TDHD𝟏)(π¯TD𝟏)−1+1nTπ¯T𝟏THD𝟏=subscript¯𝜋𝑇subscript¯𝜋𝑇𝐷𝐻𝐷1superscriptsubscript¯𝜋𝑇𝐷11subscript¯𝜋𝑇subscript¯𝜋𝑇𝐷𝐻𝐷1superscriptsubscript¯𝜋𝑇𝐷111subscript𝑛𝑇subscript¯𝜋𝑇subscript1𝑇𝐻𝐷1absent\displaystyle\bar{\pi}_{T}(\bar{\pi}_{T}DHD{\bf 1})(\bar{\pi}_{T}D{\bf 1})^{-1}-\bar{\pi}_{T}(\bar{\pi}_{T}DHD{\bf 1})(\bar{\pi}_{T}D{\bf 1})^{-1}+\frac{1}{n_{T}}\bar{\pi}_{T}{\bf 1}_{T}HD{\bf 1}=</td>
<td></td>
</tr>
<tr>
<td></td>
<td></td>
<td>=\displaystyle=</td>
<td>𝟏TH1nT(π¯TD𝟏)−π¯TDH.superscript1𝑇𝐻1subscript𝑛𝑇subscript¯𝜋𝑇𝐷1subscript¯𝜋𝑇𝐷𝐻\displaystyle{\bf 1}^{T}H\frac{1}{n_{T}}(\bar{\pi}_{T}D{\bf 1})-\bar{\pi}_{T}DH.</td>
<td></td>
</tr>
</tbody>
</table>
Thus, we have
<table>
<tbody>
<tr>
<td></td>
<td>π^T(1)superscriptsubscript^𝜋𝑇1\displaystyle\hat{\pi}_{T}^{(1)}</td>
<td>=\displaystyle=</td>
<td>𝟏TH1nT(π¯TD𝟏)−π~T.superscript1𝑇𝐻1subscript𝑛𝑇subscript¯𝜋𝑇𝐷1subscript~𝜋𝑇\displaystyle{\bf 1}^{T}H\frac{1}{n_{T}}(\bar{\pi}_{T}D{\bf 1})-\tilde{\pi}_{T}.</td>
<td></td>
</tr>
</tbody>
</table>
$`\square`$
Next, we consider a quasi-stationary distribution \[[9](#bib.bib9), [22](#bib.bib22)\] defined by equation
<table>
<tbody>
<tr>
<td></td>
<td>π~TT=λ1π~T,subscript~𝜋𝑇𝑇subscript𝜆1subscript~𝜋𝑇\tilde{\pi}_{T}T=\lambda_{1}\tilde{\pi}_{T},</td>
<td></td>
<td>(6)</td>
</tr>
</tbody>
</table>
and the normalization condition
<table>
<tbody>
<tr>
<td></td>
<td>π~T𝟏=1,subscript~𝜋𝑇11\tilde{\pi}_{T}{\bf 1}=1,</td>
<td></td>
<td>(7)</td>
</tr>
</tbody>
</table>
where $`\lambda_{1}`$ is the Perron-Frobenius eigenvalue of matrix $`T`$. The quasi-stationary distribution can be interpreted as a proper initial distribution on the non-absorbing states (states in ESCC) which is such that the distribution of the random walk, conditioned on the non-absorption prior time $`t`$, is independent of $`t`$ \[[11](#bib.bib11)\]. As in the analysis of the pseudo-stationary distribution, we take the matrix $`T`$ in the form of perturbation $`T = {\overline{T} - {\varepsilon\hspace{0pt}D}}`$.
######
Proposition 4
The vector $`{\overset{\sim}{\pi}}_{T}`$ is close to the vector $`{\overline{\pi}}_{T}`$. Namely,
<table>
<tbody>
<tr>
<td></td>
<td>π~T=π¯T−επ¯TDH+o(ε).subscript~𝜋𝑇subscript¯𝜋𝑇𝜀subscript¯𝜋𝑇𝐷𝐻𝑜𝜀\tilde{\pi}_{T}=\bar{\pi}_{T}-\varepsilon\bar{\pi}_{T}DH+o(\varepsilon).</td>
<td></td>
</tr>
</tbody>
</table>
Proof: We look for the quasi-stationary distribution and the Perron-Frobenius eigenvalue in the form of power series
<table>
<tbody>
<tr>
<td></td>
<td>π~T=π~T(0)+επ~T(1)+ε2π~T(2)+…,subscript~𝜋𝑇superscriptsubscript~𝜋𝑇0𝜀superscriptsubscript~𝜋𝑇1superscript𝜀2superscriptsubscript~𝜋𝑇2…\tilde{\pi}_{T}=\tilde{\pi}_{T}^{(0)}+\varepsilon\tilde{\pi}_{T}^{(1)}+\varepsilon^{2}\tilde{\pi}_{T}^{(2)}+\ldots,</td>
<td></td>
<td>(8)</td>
</tr>
</tbody>
</table>
<table>
<tbody>
<tr>
<td></td>
<td>λ1=1+ελ1(1)+ε2λ1(2)+….subscript𝜆11𝜀superscriptsubscript𝜆11superscript𝜀2superscriptsubscript𝜆12…\lambda_{1}=1+\varepsilon\lambda_{1}^{(1)}+\varepsilon^{2}\lambda_{1}^{(2)}+\ldots.</td>
<td></td>
</tr>
</tbody>
</table>
Substituting $`T = {\overline{T} - {\varepsilon\hspace{0pt}D}}`$ and the above series into ([6](#S2.E6 "In 2 Quasi-stationary distributions as centrality measures")), and equating terms with the same powers of $`\varepsilon`$, we obtain
<table>
<tbody>
<tr>
<td></td>
<td>π~T(0)T¯=π~T(0),superscriptsubscript~𝜋𝑇0¯𝑇superscriptsubscript~𝜋𝑇0\tilde{\pi}_{T}^{(0)}\bar{T}=\tilde{\pi}_{T}^{(0)},</td>
<td></td>
<td>(9)</td>
</tr>
</tbody>
</table>
<table>
<tbody>
<tr>
<td></td>
<td>π~T(1)T¯−π~T(0)D=1π~T(1)+λ1(1)π~T(0),superscriptsubscript~𝜋𝑇1¯𝑇superscriptsubscript~𝜋𝑇0𝐷1superscriptsubscript~𝜋𝑇1superscriptsubscript𝜆11superscriptsubscript~𝜋𝑇0\tilde{\pi}_{T}^{(1)}\bar{T}-\tilde{\pi}_{T}^{(0)}D=1\tilde{\pi}_{T}^{(1)}+\lambda_{1}^{(1)}\tilde{\pi}_{T}^{(0)},</td>
<td></td>
<td>(10)</td>
</tr>
</tbody>
</table>
Substituting ([8](#S2.E8 "In 2 Quasi-stationary distributions as centrality measures")) into the normalization condition ([7](#S2.E7 "In 2 Quasi-stationary distributions as centrality measures")), we get
<table>
<tbody>
<tr>
<td></td>
<td>π~T(0)𝟏=1,superscriptsubscript~𝜋𝑇011\tilde{\pi}_{T}^{(0)}{\bf 1}=1,</td>
<td></td>
<td>(11)</td>
</tr>
</tbody>
</table>
<table>
<tbody>
<tr>
<td></td>
<td>π~T(1)𝟏=0.superscriptsubscript~𝜋𝑇110\tilde{\pi}_{T}^{(1)}{\bf 1}=0.</td>
<td></td>
<td>(12)</td>
</tr>
</tbody>
</table>
From ([9](#S2.E9 "In 2 Quasi-stationary distributions as centrality measures")) and ([11](#S2.E11 "In 2 Quasi-stationary distributions as centrality measures")) we conclude that $`{\overset{\sim}{\pi}}_{T}^{(0)} = {\overline{\pi}}_{T}`$. Thus, the equation ([10](#S2.E10 "In 2 Quasi-stationary distributions as centrality measures")) takes the form
<table>
<tbody>
<tr>
<td></td>
<td>π~T(1)T¯−π¯TD=1π~T(1)+λ1(1)π¯T.superscriptsubscript~𝜋𝑇1¯𝑇subscript¯𝜋𝑇𝐷1superscriptsubscript~𝜋𝑇1superscriptsubscript𝜆11subscript¯𝜋𝑇\tilde{\pi}_{T}^{(1)}\bar{T}-\bar{\pi}_{T}D=1\tilde{\pi}_{T}^{(1)}+\lambda_{1}^{(1)}\bar{\pi}_{T}.</td>
<td></td>
</tr>
</tbody>
</table>
Post-multiplying this equation by $`\mathbf{1}`$, we get
<table>
<tbody>
<tr>
<td></td>
<td>π~T(1)T¯𝟏−π¯TD𝟏=1π~T(1)𝟏+λ1(1)π¯T𝟏.superscriptsubscript~𝜋𝑇1¯𝑇1subscript¯𝜋𝑇𝐷11superscriptsubscript~𝜋𝑇11superscriptsubscript𝜆11subscript¯𝜋𝑇1\tilde{\pi}_{T}^{(1)}\bar{T}{\bf 1}-\bar{\pi}_{T}D{\bf 1}=1\tilde{\pi}_{T}^{(1)}{\bf 1}+\lambda_{1}^{(1)}\bar{\pi}_{T}{\bf 1}.</td>
<td></td>
</tr>
</tbody>
</table>
Now using $`{\overline{T}\hspace{0pt}\mathbf{1}} = \mathbf{1}`$, ([11](#S2.E11 "In 2 Quasi-stationary distributions as centrality measures")) and ([12](#S2.E12 "In 2 Quasi-stationary distributions as centrality measures")), we conclude that
<table>
<tbody>
<tr>
<td></td>
<td>λ1(1)=−π¯TD𝟏,superscriptsubscript𝜆11subscript¯𝜋𝑇𝐷1\lambda_{1}^{(1)}=-\bar{\pi}_{T}D{\bf 1},</td>
<td></td>
</tr>
</tbody>
</table>
and, consequently,
<table>
<tbody>
<tr>
<td></td>
<td>λ1=1−επ¯TD𝟏+o(ε).subscript𝜆11𝜀subscript¯𝜋𝑇𝐷1𝑜𝜀\lambda_{1}=1-\varepsilon\bar{\pi}_{T}D{\bf 1}+o(\varepsilon).</td>
<td></td>
<td>(13)</td>
</tr>
</tbody>
</table>
Now the equation ([10](#S2.E10 "In 2 Quasi-stationary distributions as centrality measures")) can be rewritten as follows:
<table>
<tbody>
<tr>
<td></td>
<td>π~T(1)[I−T¯]=π¯T[(π¯TD𝟏)I−D].superscriptsubscript~𝜋𝑇1delimited-[]𝐼¯𝑇subscript¯𝜋𝑇delimited-[]subscript¯𝜋𝑇𝐷1𝐼𝐷\tilde{\pi}_{T}^{(1)}[I-\bar{T}]=\bar{\pi}_{T}[(\bar{\pi}_{T}D{\bf 1})I-D].</td>
<td></td>
</tr>
</tbody>
</table>
Its general solution is given by
<table>
<tbody>
<tr>
<td></td>
<td>π~T(1)=νπ¯T+π¯T[(π¯TD𝟏)I−D]H,superscriptsubscript~𝜋𝑇1𝜈subscript¯𝜋𝑇subscript¯𝜋𝑇delimited-[]subscript¯𝜋𝑇𝐷1𝐼𝐷𝐻\tilde{\pi}_{T}^{(1)}=\nu\bar{\pi}_{T}+\bar{\pi}_{T}[(\bar{\pi}_{T}D{\bf 1})I-D]H,</td>
<td></td>
</tr>
</tbody>
</table>
where $`\nu`$ is some constant. To find constant $`\nu`$, we substitute the above general solution into condition ([12](#S2.E12 "In 2 Quasi-stationary distributions as centrality measures")).
<table>
<tbody>
<tr>
<td></td>
<td>π~T(1)𝟏=νπ¯T𝟏+π¯T[(π¯TD𝟏)I−D]H𝟏=0.superscriptsubscript~𝜋𝑇11𝜈subscript¯𝜋𝑇1subscript¯𝜋𝑇delimited-[]subscript¯𝜋𝑇𝐷1𝐼𝐷𝐻10\tilde{\pi}_{T}^{(1)}{\bf 1}=\nu\bar{\pi}_{T}{\bf 1}+\bar{\pi}_{T}[(\bar{\pi}_{T}D{\bf 1})I-D]H{\bf 1}=0.</td>
<td></td>
</tr>
</tbody>
</table>
Since $`{{\overline{\pi}}_{T}\hspace{0pt}\mathbf{1}} = 1`$ and $`{H\hspace{0pt}\mathbf{1}} = 0`$, we get $`\nu = 0`$. Consequently, we have
<table>
<tbody>
<tr>
<td></td>
<td>π~T(1)=π¯T[(π¯TD𝟏)I−D]H=(π¯TD𝟏)π¯TH−π¯TDH=−π¯TDH.superscriptsubscript~𝜋𝑇1subscript¯𝜋𝑇delimited-[]subscript¯𝜋𝑇𝐷1𝐼𝐷𝐻subscript¯𝜋𝑇𝐷1subscript¯𝜋𝑇𝐻subscript¯𝜋𝑇𝐷𝐻subscript¯𝜋𝑇𝐷𝐻\tilde{\pi}_{T}^{(1)}=\bar{\pi}_{T}[(\bar{\pi}_{T}D{\bf 1})I-D]H=(\bar{\pi}_{T}D{\bf 1})\bar{\pi}_{T}H-\bar{\pi}_{T}DH=-\bar{\pi}_{T}DH.</td>
<td></td>
</tr>
</tbody>
</table>
In the above, we have used the fact that $`{{\overline{\pi}}_{T}\hspace{0pt}H} = 0`$. This completes the proof.
$`\square`$
Since $`\lambda_{1}`$ is very close to one, we conclude from ([13](#S2.E13 "In 2 Quasi-stationary distributions as centrality measures")) and the equality $`{\varepsilon\hspace{0pt}{\overline{\pi}}_{T}\hspace{0pt}D\hspace{0pt}\mathbf{1}} = {{\overline{\pi}}_{T}\hspace{0pt}R\hspace{0pt}\mathbf{1}}`$ that indeed $`{\overline{\pi}}_{T}\hspace{0pt}R\hspace{0pt}\mathbf{1}`$ is typically very small.
There is also a simple relation between $`\lambda_{1}`$ and $`{\overset{\sim}{\pi}}_{T}`$.
######
Proposition 5
The Perron-Frobenius eigenvalue $`\lambda_{1}`$ of matrix $`T`$ is given by
<table>
<tbody>
<tr>
<td></td>
<td>λ1=1−π~TR𝟏.subscript𝜆11subscript~𝜋𝑇𝑅1\lambda_{1}=1-\tilde{\pi}_{T}R{\bf 1}.</td>
<td></td>
<td>(14)</td>
</tr>
</tbody>
</table>
Proof: Post-multiplying the equation ([6](#S2.E6 "In 2 Quasi-stationary distributions as centrality measures")) by $`\mathbf{1}`$, we obtain
<table>
<tbody>
<tr>
<td></td>
<td>λ1=π~TT𝟏.subscript𝜆1subscript~𝜋𝑇𝑇1\lambda_{1}=\tilde{\pi}_{T}T{\bf 1}.</td>
<td></td>
</tr>
</tbody>
</table>
Then, using the fact that $`{T\hspace{0pt}\mathbf{1}} = {\mathbf{1} - {R\hspace{0pt}\mathbf{1}}}`$ we derive the formula ([14](#S2.E14 "In Proposition 5 ‣ 2 Quasi-stationary distributions as centrality measures")).
$`\square`$
Proposition [5](#Thmproposition5 "Proposition 5 ‣ 2 Quasi-stationary distributions as centrality measures") indicates that if $`\lambda_{1}`$ is close to one then $`{\overset{\sim}{\pi}}_{T}\hspace{0pt}R\hspace{0pt}\mathbf{1}`$ is small.
As we mentioned above the $`{\overline{T}}_{i\hspace{0pt}j}`$ entry of the matrix $`\overline{T}`$ can be considered as a conditional probability to jump from the node $`i`$ to the node $`j`$ under the condition that random walk does not leave ESCC at the jump.
Let us consider the situation when the random walk stays inside ESCC after some finite number of jumps. The probability of such an event can be expressed as follows:
<table>
<tbody>
<tr>
<td></td>
<td>P(X1=j|X0=i∧⋀m=1NXm∈S),𝑃subscript𝑋1conditional𝑗subscript𝑋0𝑖superscriptsubscript𝑚1𝑁subscript𝑋𝑚𝑆P\left(X_{1}=j|X_{0}=i\wedge\bigwedge_{m=1}^{N}X_{m}\in S\right),</td>
<td></td>
</tr>
</tbody>
</table>
where ESCC is denoted by $`S`$ for the sake of shortening notation and $`N`$ is the number of jumps during which the random walk stays in ESCC.
Let us denote by $`T_{i\hspace{0pt}j}^{(N)}`$ the element of $`T^{N}`$ (the $`N^{\text{th}}`$ power of T) and by $`T_{i}^{(N)}`$ the $`i^{\text{th}}`$ row of the matrix $`T^{N}`$. Then
<table>
<tbody>
<tr>
<td></td>
<td>Ti(N)=(TN)i=(TTN−1)i=TiTN−1.subscriptsuperscript𝑇𝑁𝑖subscriptsuperscript𝑇𝑁𝑖subscript𝑇superscript𝑇𝑁1𝑖subscript𝑇𝑖superscript𝑇𝑁1T^{(N)}_{i}=(T^{N})_{i}=(TT^{N-1})_{i}=T_{i}T^{N-1}.</td>
<td></td>
</tr>
</tbody>
</table>
######
Proposition 6
<table>
<tbody>
<tr>
<td></td>
<td>P(X1=j|X0=i∧⋀m=1NXm∈S)=TijTj(N−1)𝟏Ti(N)𝟏.𝑃subscript𝑋1conditional𝑗subscript𝑋0𝑖superscriptsubscript𝑚1𝑁subscript𝑋𝑚𝑆subscript𝑇𝑖𝑗subscriptsuperscript𝑇𝑁1𝑗1subscriptsuperscript𝑇𝑁𝑖1P\left(X_{1}=j|X_{0}=i\wedge\bigwedge_{m=1}^{N}X_{m}\in S\right)=\frac{T_{ij}T^{(N-1)}_{j}{\bf 1}}{T^{(N)}_{i}{\bf 1}}.</td>
<td></td>
<td>(15)</td>
</tr>
</tbody>
</table>
Proof: see Appendix.
Then, if we denote
<table>
<tbody>
<tr>
<td></td>
<td>Tˇij(N)=P(X1=j|X0=i∧⋀m=1NXm∈S),superscriptsubscriptˇ𝑇𝑖𝑗𝑁𝑃subscript𝑋1conditional𝑗subscript𝑋0𝑖superscriptsubscript𝑚1𝑁subscript𝑋𝑚𝑆\check{T}_{ij}^{(N)}=P\left(X_{1}=j|X_{0}=i\wedge\bigwedge_{m=1}^{N}X_{m}\in S\right),</td>
<td></td>
</tr>
</tbody>
</table>
we will be able to find stationary distributions of $`{\check{T}}_{i\hspace{0pt}j}^{(N)}`$, which can be viewed as generalization of $`{\overline{\pi}}_{T}`$. Let us now consider the limiting case, when $`N`$ goes to infinity.
Before we continue let us analyze the principle right eigenvector $`u`$ of the matrix $`T`$:
<table>
<tbody>
<tr>
<td></td>
<td>Tu=λ1u,𝑇𝑢subscript𝜆1𝑢\displaystyle Tu=\lambda_{1}u,</td>
<td></td>
<td>(16)</td>
</tr>
</tbody>
</table>
where $`\lambda_{1}`$ is as in the previous section, the Perron-Frobenius eigenvalue.
The vector $`u`$ can be normalized in different ways. Let us define the main normalization for $`u`$ as
<table>
<tbody>
<tr>
<td></td>
<td>𝟏Tu=nT.superscript1𝑇𝑢subscript𝑛𝑇{\bf 1}^{T}u=n_{T}.</td>
<td></td>
</tr>
</tbody>
</table>
Let us also define $`\overline{u}`$ as
<table>
<tbody>
<tr>
<td></td>
<td>u¯=uπ¯Tu, so that π¯Tu¯=1,¯𝑢𝑢subscript¯𝜋𝑇𝑢, so that subscript¯𝜋𝑇¯𝑢1\bar{u}=\frac{u}{\bar{\pi}_{T}u}\mbox{, so that }\bar{\pi}_{T}\bar{u}=1,</td>
<td></td>
<td>(17)</td>
</tr>
</tbody>
</table>
and
<table>
<tbody>
<tr>
<td></td>
<td>u~=uπ~Tu, so that π~Tu~=1.~𝑢𝑢subscript~𝜋𝑇𝑢, so that subscript~𝜋𝑇~𝑢1\tilde{u}=\frac{u}{\tilde{\pi}_{T}u}\mbox{, so that }\tilde{\pi}_{T}\tilde{u}=1.</td>
<td></td>
<td>(18)</td>
</tr>
</tbody>
</table>
######
Proposition 7
The vector $`\overline{u}`$ is close to the vector $`\mathbf{1}`$. Namely,
<table>
<tbody>
<tr>
<td></td>
<td>u¯=𝟏−εHD𝟏+o(ε).¯𝑢1𝜀𝐻𝐷1𝑜𝜀\bar{u}={\bf 1}-\varepsilon HD{\bf 1}+o(\varepsilon).</td>
<td></td>
</tr>
</tbody>
</table>
Proof: We look for the right eigenvector and the Perron-Frobenius eigenvalue in the form of power series
<table>
<tbody>
<tr>
<td></td>
<td>u¯=u¯(0)+εu¯(1)+ε2u¯(2)+….¯𝑢superscript¯𝑢0𝜀superscript¯𝑢1superscript𝜀2superscript¯𝑢2…\bar{u}=\bar{u}^{(0)}+\varepsilon\bar{u}^{(1)}+\varepsilon^{2}\bar{u}^{(2)}+\ldots.</td>
<td></td>
<td>(19)</td>
</tr>
</tbody>
</table>
<table>
<tbody>
<tr>
<td></td>
<td>λ1=1+ελ1(1)+ε2λ1(2)+….subscript𝜆11𝜀superscriptsubscript𝜆11superscript𝜀2superscriptsubscript𝜆12…\lambda_{1}=1+\varepsilon\lambda_{1}^{(1)}+\varepsilon^{2}\lambda_{1}^{(2)}+\ldots.</td>
<td></td>
</tr>
</tbody>
</table>
Substituting $`T = {\overline{T} - {\varepsilon\hspace{0pt}D}}`$ and the above series into ([16](#S2.E16 "In 2 Quasi-stationary distributions as centrality measures")), and equating terms with the same powers of $`\varepsilon`$, we obtain
<table>
<tbody>
<tr>
<td></td>
<td>T¯u¯(0)=u¯(0),¯𝑇superscript¯𝑢0superscript¯𝑢0\bar{T}\bar{u}^{(0)}=\bar{u}^{(0)},</td>
<td></td>
<td>(20)</td>
</tr>
</tbody>
</table>
<table>
<tbody>
<tr>
<td></td>
<td>T¯u¯(1)−Du¯(0)=u¯(1)+λ1(1)u¯(0).¯𝑇superscript¯𝑢1𝐷superscript¯𝑢0superscript¯𝑢1superscriptsubscript𝜆11superscript¯𝑢0\bar{T}\bar{u}^{(1)}-D\bar{u}^{(0)}=\bar{u}^{(1)}+\lambda_{1}^{(1)}\bar{u}^{(0)}.</td>
<td></td>
<td>(21)</td>
</tr>
</tbody>
</table>
Substituting ([19](#S2.E19 "In 2 Quasi-stationary distributions as centrality measures")) into the normalization condition ([17](#S2.E17 "In 2 Quasi-stationary distributions as centrality measures")), we obtain
<table>
<tbody>
<tr>
<td></td>
<td>π¯Tu¯(0)=1,subscript¯𝜋𝑇superscript¯𝑢01\bar{\pi}_{T}\bar{u}^{(0)}=1,</td>
<td></td>
<td>(22)</td>
</tr>
</tbody>
</table>
<table>
<tbody>
<tr>
<td></td>
<td>π¯Tu¯(1)=0.subscript¯𝜋𝑇superscript¯𝑢10\bar{\pi}_{T}\bar{u}^{(1)}=0.</td>
<td></td>
<td>(23)</td>
</tr>
</tbody>
</table>
From ([20](#S2.E20 "In 2 Quasi-stationary distributions as centrality measures")) and ([22](#S2.E22 "In 2 Quasi-stationary distributions as centrality measures")) we conclude that $`{\overline{u}}^{(0)} = \mathbf{1}`$. Thus, the equation ([21](#S2.E21 "In 2 Quasi-stationary distributions as centrality measures")) takes the form
<table>
<tbody>
<tr>
<td></td>
<td>T¯u¯(1)−D𝟏=u¯(1)+λ1(1)𝟏.¯𝑇superscript¯𝑢1𝐷1superscript¯𝑢1superscriptsubscript𝜆111\bar{T}\bar{u}^{(1)}-D{\bf 1}=\bar{u}^{(1)}+\lambda_{1}^{(1)}{\bf 1}.</td>
<td></td>
</tr>
</tbody>
</table>
Pre-multiplying this equation by $`{\overline{\pi}}_{T}`$, we get
<table>
<tbody>
<tr>
<td></td>
<td>π¯Tu¯(1)−π¯TD𝟏=π¯Tu¯(1)+π¯Tλ1(1)𝟏.subscript¯𝜋𝑇superscript¯𝑢1subscript¯𝜋𝑇𝐷1subscript¯𝜋𝑇superscript¯𝑢1subscript¯𝜋𝑇superscriptsubscript𝜆111\bar{\pi}_{T}\bar{u}^{(1)}-\bar{\pi}_{T}D{\bf 1}=\bar{\pi}_{T}\bar{u}^{(1)}+\bar{\pi}_{T}\lambda_{1}^{(1)}{\bf 1}.</td>
<td></td>
</tr>
</tbody>
</table>
Now using $`{\overline{T}\hspace{0pt}\mathbf{1}} = \mathbf{1}`$, ([22](#S2.E22 "In 2 Quasi-stationary distributions as centrality measures")) and ([23](#S2.E23 "In 2 Quasi-stationary distributions as centrality measures")), we conclude that
<table>
<tbody>
<tr>
<td></td>
<td>λ1(1)=−π¯TD𝟏,superscriptsubscript𝜆11subscript¯𝜋𝑇𝐷1\lambda_{1}^{(1)}=-\bar{\pi}_{T}D{\bf 1},</td>
<td></td>
</tr>
</tbody>
</table>
and, consequently,
<table>
<tbody>
<tr>
<td></td>
<td>λ1=1−επ¯TD𝟏+o(ε).subscript𝜆11𝜀subscript¯𝜋𝑇𝐷1𝑜𝜀\lambda_{1}=1-\varepsilon\bar{\pi}_{T}D{\bf 1}+o(\varepsilon).</td>
<td></td>
</tr>
</tbody>
</table>
Now the equation ([21](#S2.E21 "In 2 Quasi-stationary distributions as centrality measures")) can be rewritten as follows:
<table>
<tbody>
<tr>
<td></td>
<td>[I−T¯]u¯(1)=[(π¯TD𝟏)I−D]𝟏.delimited-[]𝐼¯𝑇superscript¯𝑢1delimited-[]subscript¯𝜋𝑇𝐷1𝐼𝐷1\left[I-\bar{T}\right]\bar{u}^{(1)}=\left[\left(\bar{\pi}_{T}D{\bf 1}\right)I-D\right]{\bf 1}.</td>
<td></td>
</tr>
</tbody>
</table>
Its general solution is given by
<table>
<tbody>
<tr>
<td></td>
<td>u¯(1)=ν𝟏+H[(π¯TD𝟏)I−D]𝟏,superscript¯𝑢1𝜈1𝐻delimited-[]subscript¯𝜋𝑇𝐷1𝐼𝐷1\bar{u}^{(1)}=\nu{\bf 1}+H\left[\left(\bar{\pi}_{T}D{\bf 1}\right)I-D\right]{\bf 1},</td>
<td></td>
</tr>
</tbody>
</table>
where $`\nu`$ is some constant. To find constant $`\nu`$, we substitute the above general solution into condition ([23](#S2.E23 "In 2 Quasi-stationary distributions as centrality measures")).
<table>
<tbody>
<tr>
<td></td>
<td>π¯Tu¯(1)=νπ¯T𝟏+π¯TH[(π¯TD𝟏)I−D]𝟏.subscript¯𝜋𝑇superscript¯𝑢1𝜈subscript¯𝜋𝑇1subscript¯𝜋𝑇𝐻delimited-[]subscript¯𝜋𝑇𝐷1𝐼𝐷1\bar{\pi}_{T}\bar{u}^{(1)}=\nu\bar{\pi}_{T}{\bf 1}+\bar{\pi}_{T}H\left[\left(\bar{\pi}_{T}D{\bf 1}\right)I-D\right]{\bf 1}.</td>
<td></td>
</tr>
</tbody>
</table>
Since $`{{\overline{\pi}}_{T}\hspace{0pt}\mathbf{1}} = 1`$ and $`{{\overline{\pi}}_{T}\hspace{0pt}H} = 0`$, we get $`\nu = 0`$. Consequently, we have
<table>
<tbody>
<tr>
<td></td>
<td>u¯(1)=−HD𝟏.superscript¯𝑢1𝐻𝐷1\bar{u}^{(1)}=-HD{\bf 1}.</td>
<td></td>
</tr>
</tbody>
</table>
In the above, we have used the fact that $`{H\hspace{0pt}\mathbf{1}} = 0`$. This completes the proof.
$`\square`$
We note that the elements of the vector $`\overset{\sim}{u}`$ can be calculated by the power iteration method.
######
Proposition 8
The following convergence takes place
<table>
<tbody>
<tr>
<td></td>
<td>u~i=limn→∞TiTn−1eλ1n,subscript~𝑢𝑖subscript→𝑛subscript𝑇𝑖superscript𝑇𝑛1𝑒superscriptsubscript𝜆1𝑛\tilde{u}_{i}=\lim_{n\to\infty}\frac{T_{i}T^{n-1}e}{\lambda_{1}^{n}},</td>
<td></td>
<td>(24)</td>
</tr>
</tbody>
</table>
where $`T_{i}`$ is the $`i^{\text{th}}`$ row of the matrix $`T`$.
Proof:
<table>
<tbody>
<tr>
<td></td>
<td>u~i(1)subscriptsuperscript~𝑢1𝑖\displaystyle\tilde{u}^{(1)}_{i}</td>
<td>=\displaystyle=</td>
<td>Tieπ~TTe=Tieλ1,subscript𝑇𝑖𝑒subscript~𝜋𝑇𝑇𝑒subscript𝑇𝑖𝑒subscript𝜆1\displaystyle\frac{T_{i}e}{\tilde{\pi}_{T}Te}=\frac{T_{i}e}{\lambda_{1}},</td>
<td></td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td>u~i(2)subscriptsuperscript~𝑢2𝑖\displaystyle\tilde{u}^{(2)}_{i}</td>
<td>=\displaystyle=</td>
<td>Tiu~(1)π~TTu~(1)=TiTeλ1λ1=TiTeλ12,subscript𝑇𝑖superscript~𝑢1subscript~𝜋𝑇𝑇superscript~𝑢1subscript𝑇𝑖𝑇𝑒subscript𝜆1subscript𝜆1subscript𝑇𝑖𝑇𝑒superscriptsubscript𝜆12\displaystyle\frac{T_{i}\tilde{u}^{(1)}}{\tilde{\pi}_{T}T\tilde{u}^{(1)}}=\frac{T_{i}\frac{Te}{\lambda_{1}}}{\lambda_{1}}=\frac{T_{i}Te}{\lambda_{1}^{2}},</td>
<td></td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td>u~i(3)subscriptsuperscript~𝑢3𝑖\displaystyle\tilde{u}^{(3)}_{i}</td>
<td>=\displaystyle=</td>
<td>Tiu~(2)π~TTu~(2)=TiT2eλ13,subscript𝑇𝑖superscript~𝑢2subscript~𝜋𝑇𝑇superscript~𝑢2subscript𝑇𝑖superscript𝑇2𝑒superscriptsubscript𝜆13\displaystyle\frac{T_{i}\tilde{u}^{(2)}}{\tilde{\pi}_{T}T\tilde{u}^{(2)}}=\frac{T_{i}T^{2}e}{\lambda_{1}^{3}},</td>
<td></td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td>⋮⋮\displaystyle\vdots</td>
<td></td>
<td></td>
<td></td>
</tr>
</tbody>
</table>
$`\square`$
Let us consider the twisted kernel $`\check{T}`$ defined by
<table>
<tbody>
<tr>
<td></td>
<td>Tˇij=Tijujλ1ui.subscriptˇ𝑇𝑖𝑗subscript𝑇𝑖𝑗subscript𝑢𝑗subscript𝜆1subscript𝑢𝑖\check{T}_{ij}=\frac{T_{ij}u_{j}}{\lambda_{1}u_{i}}.</td>
<td></td>
</tr>
</tbody>
</table>
As one can see the twisted kernel does not depend on the normalization of $`u`$. Hence, we can take any normalization.
######
Proposition 9
The twisted kernel is a limit of ([15](#S2.E15 "In Proposition 6 ‣ 2 Quasi-stationary distributions as centrality measures")) as $`N`$ goes to infinity, that is
<table>
<tbody>
<tr>
<td></td>
<td>Tˇij=limN→∞TijTj(N−1)𝟏Ti(N)𝟏.subscriptˇ𝑇𝑖𝑗subscript→𝑁subscript𝑇𝑖𝑗subscriptsuperscript𝑇𝑁1𝑗1subscriptsuperscript𝑇𝑁𝑖1\check{T}_{ij}=\lim_{N\to\infty}\frac{T_{ij}T^{(N-1)}_{j}{\bf 1}}{T^{(N)}_{i}{\bf 1}}.</td>
<td></td>
</tr>
</tbody>
</table>
Proof:
<table>
<tbody>
<tr>
<td></td>
<td>TijTj(N−1)𝟏Ti(N)𝟏=TijTjTN−2𝟏TiTN−1𝟏=Tijλ1TjTN−2𝟏λ1N−1TiTN−1𝟏λ1N.subscript𝑇𝑖𝑗subscriptsuperscript𝑇𝑁1𝑗1subscriptsuperscript𝑇𝑁𝑖1subscript𝑇𝑖𝑗subscript𝑇𝑗superscript𝑇𝑁21subscript𝑇𝑖superscript𝑇𝑁11subscript𝑇𝑖𝑗subscript𝜆1subscript𝑇𝑗superscript𝑇𝑁21superscriptsubscript𝜆1𝑁1subscript𝑇𝑖superscript𝑇𝑁11superscriptsubscript𝜆1𝑁\displaystyle\frac{T_{ij}T^{(N-1)}_{j}{\bf 1}}{T^{(N)}_{i}{\bf 1}}=T_{ij}\frac{T_{j}T^{N-2}{\bf 1}}{T_{i}T^{N-1}{\bf 1}}=\frac{T_{ij}}{\lambda_{1}}\frac{\frac{T_{j}T^{N-2}{\bf 1}}{\lambda_{1}^{N-1}}}{\frac{T_{i}T^{N-1}{\bf 1}}{\lambda_{1}^{N}}}.</td>
<td></td>
</tr>
</tbody>
</table>
<table>
<tbody>
<tr>
<td></td>
<td>limN→∞TijTj(N−1)𝟏Ti(N)𝟏=Tijλ1limN→∞TjTN−2𝟏λ1N−1TiTN−1𝟏λ1N=Tijλ1limN→∞TjTN−2𝟏λ1N−1limN→∞TiTN−1𝟏λ1N.subscript→𝑁subscript𝑇𝑖𝑗subscriptsuperscript𝑇𝑁1𝑗1subscriptsuperscript𝑇𝑁𝑖1subscript𝑇𝑖𝑗subscript𝜆1subscript→𝑁subscript𝑇𝑗superscript𝑇𝑁21superscriptsubscript𝜆1𝑁1subscript𝑇𝑖superscript𝑇𝑁11superscriptsubscript𝜆1𝑁subscript𝑇𝑖𝑗subscript𝜆1subscript→𝑁subscript𝑇𝑗superscript𝑇𝑁21superscriptsubscript𝜆1𝑁1subscript→𝑁subscript𝑇𝑖superscript𝑇𝑁11superscriptsubscript𝜆1𝑁\displaystyle\lim_{N\to\infty}\frac{T_{ij}T^{(N-1)}_{j}{\bf 1}}{T^{(N)}_{i}{\bf 1}}=\frac{T_{ij}}{\lambda_{1}}\lim_{N\to\infty}\frac{\frac{T_{j}T^{N-2}{\bf 1}}{\lambda_{1}^{N-1}}}{\frac{T_{i}T^{N-1}{\bf 1}}{\lambda_{1}^{N}}}=\frac{T_{ij}}{\lambda_{1}}\frac{\lim_{N\to\infty}\frac{T_{j}T^{N-2}{\bf 1}}{\lambda_{1}^{N-1}}}{\lim_{N\to\infty}\frac{T_{i}T^{N-1}{\bf 1}}{\lambda_{1}^{N}}}.</td>
<td></td>
</tr>
</tbody>
</table>
Using ([24](#S2.E24 "In Proposition 8 ‣ 2 Quasi-stationary distributions as centrality measures")), we can write
<table>
<tbody>
<tr>
<td></td>
<td>limN→∞TijTj(N−1)𝟏Ti(N)𝟏=Tiju~jλ1u~i.subscript→𝑁subscript𝑇𝑖𝑗subscriptsuperscript𝑇𝑁1𝑗1subscriptsuperscript𝑇𝑁𝑖1subscript𝑇𝑖𝑗subscript~𝑢𝑗subscript𝜆1subscript~𝑢𝑖\displaystyle\lim_{N\to\infty}\frac{T_{ij}T^{(N-1)}_{j}{\bf 1}}{T^{(N)}_{i}{\bf 1}}=\frac{T_{ij}\tilde{u}_{j}}{\lambda_{1}\tilde{u}_{i}}.</td>
<td></td>
</tr>
</tbody>
</table>
After renormalization, we obtain
<table>
<tbody>
<tr>
<td></td>
<td>limN→∞TijTj(N−1)𝟏Ti(N)𝟏=Tijujλ1ui.subscript→𝑁subscript𝑇𝑖𝑗subscriptsuperscript𝑇𝑁1𝑗1subscriptsuperscript𝑇𝑁𝑖1subscript𝑇𝑖𝑗subscript𝑢𝑗subscript𝜆1subscript𝑢𝑖\displaystyle\lim_{N\to\infty}\frac{T_{ij}T^{(N-1)}_{j}{\bf 1}}{T^{(N)}_{i}{\bf 1}}=\frac{T_{ij}u_{j}}{\lambda_{1}u_{i}}.</td>
<td></td>
</tr>
</tbody>
</table>
$`\square`$
The twisted kernel plays an important role in multiplicative ergodic theory and large deviations for Markov chains, see, e.g., \[[14](#bib.bib14)\]. The matrix $`\check{T}`$ is clearly a transition probability kernel, i.e., $`{{\check{T}}_{i\hspace{0pt}j} \geq {{0\hspace{0pt}{\forall i}},j}},`$ and $`{\sum_{j}{\check{T}}_{i\hspace{0pt}j}} = {1\hspace{0pt}{\forall i}}`$. Also, it is irreducible if there exists an path $`i\rightarrow j`$ under $`T`$ for all $`i,j`$, which we assume to be the case. In particular, it will have a unique stationary distribution $`{\check{\pi}}_{T}`$ associated with it:
<table>
<tbody>
<tr>
<td></td>
<td>πˇT=πˇTTˇ,subscriptˇ𝜋𝑇subscriptˇ𝜋𝑇ˇ𝑇\displaystyle\check{\pi}_{T}=\check{\pi}_{T}\check{T},</td>
<td></td>
<td>(25)</td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td>πˇT𝟏=1.subscriptˇ𝜋𝑇11\displaystyle\check{\pi}_{T}{\bf 1}=1.</td>
<td></td>
<td>(26)</td>
</tr>
</tbody>
</table>
If we assume aperiodicity in addition, $`{\check{T}}_{i\hspace{0pt}j}`$ can be given the interpretation of the probability of transition from $`i`$ to $`j`$ in the ESCC for the chain, conditioned on the fact that it never leaves the ESCC. Thus, $`{\check{\pi}}_{T}`$ qualifies as an alternative definition of a quasi-stationary distribution.
######
Proposition 10
The following expression for $`{\check{\pi}}_{T}`$ holds:
<table>
<tbody>
<tr>
<td></td>
<td>πˇT=π~Tiu~i.subscriptˇ𝜋𝑇subscript~𝜋𝑇𝑖subscript~𝑢𝑖\check{\pi}_{T}=\tilde{\pi}_{Ti}\tilde{u}_{i}.</td>
<td></td>
<td>(27)</td>
</tr>
</tbody>
</table>
Proof: The normalization condition ([26](#S2.E26 "In 2 Quasi-stationary distributions as centrality measures")) is satisfied due to ([18](#S2.E18 "In 2 Quasi-stationary distributions as centrality measures")). Let us show that ([25](#S2.E25 "In 2 Quasi-stationary distributions as centrality measures")) holds as well, i.e.
<table>
<tbody>
<tr>
<td></td>
<td>πˇTj=∑i=1nTπˇTiTˇij,subscriptˇ𝜋𝑇𝑗superscriptsubscript𝑖1subscript𝑛𝑇subscriptˇ𝜋𝑇𝑖subscriptˇ𝑇𝑖𝑗\displaystyle\check{\pi}_{Tj}=\sum_{i=1}^{n_{T}}\check{\pi}_{Ti}\check{T}_{ij},</td>
<td></td>
</tr>
</tbody>
</table>
where $`n_{T}`$ is the dimension of $`{\check{\pi}}_{T}`$. And for the right hand side of ([27](#S2.E27 "In Proposition 10 ‣ 2 Quasi-stationary distributions as centrality measures")) we have
<table>
<tbody>
<tr>
<td></td>
<td>∑i=1nTπ~Tiu~iTˇij=∑i=1nTπ~Tiu~iTiju~jλ1u~i=∑i=1nTπ~Tiu~iTiju~jλ1u~i=u~jλ1λ1π~Tj=π~Tju~j.superscriptsubscript𝑖1subscript𝑛𝑇subscript~𝜋𝑇𝑖subscript~𝑢𝑖subscriptˇ𝑇𝑖𝑗superscriptsubscript𝑖1subscript𝑛𝑇subscript~𝜋𝑇𝑖subscript~𝑢𝑖subscript𝑇𝑖𝑗subscript~𝑢𝑗subscript𝜆1subscript~𝑢𝑖superscriptsubscript𝑖1subscript𝑛𝑇subscript~𝜋𝑇𝑖subscript~𝑢𝑖subscript𝑇𝑖𝑗subscript~𝑢𝑗subscript𝜆1subscript~𝑢𝑖subscript~𝑢𝑗subscript𝜆1subscript𝜆1subscript~𝜋𝑇𝑗subscript~𝜋𝑇𝑗subscript~𝑢𝑗\displaystyle\sum_{i=1}^{n_{T}}\tilde{\pi}_{Ti}\tilde{u}_{i}\check{T}_{ij}=\sum_{i=1}^{n_{T}}\tilde{\pi}_{Ti}\tilde{u}_{i}\frac{T_{ij}\tilde{u}_{j}}{\lambda_{1}\tilde{u}_{i}}=\sum_{i=1}^{n_{T}}\tilde{\pi}_{Ti}\tilde{u}_{i}\frac{T_{ij}\tilde{u}_{j}}{\lambda_{1}\tilde{u}_{i}}=\frac{\tilde{u}_{j}}{\lambda_{1}}\lambda_{1}\tilde{\pi}_{Tj}=\tilde{\pi}_{Tj}\tilde{u}_{j}.</td>
<td></td>
</tr>
</tbody>
</table>
$`\square`$
This suggests that $`{\check{\pi}}_{T\hspace{0pt}i}`$, or equivalently $`{\overset{\sim}{\pi}}_{T\hspace{0pt}i}\hspace{0pt}{\overset{\sim}{u}}_{i}`$, may be used as another alternative centrality measure. Since the substochastic matrix $`T`$ is close to stochastic, the vector $`u`$ will be very close to $`\mathbf{1}`$. Consequently, the vector $`{\check{\pi}}_{T}`$ will be close to $`{\overset{\sim}{\pi}}_{T}`$ and to $`\overline{\pi}`$ as well. This shows that in the case when the matrix $`T`$ is close to the stochastic matrix all the alternative definitions of quasi-stationary distribution are quite close to each other. And then, from Proposition [1](#Thmproposition1 "Proposition 1 ‣ 2 Quasi-stationary distributions as centrality measures"), we conclude that the PageRank ranking converges to the quasi-stationarity based ranking as the damping factor goes to one.
## 3 Numerical experiments and Applications
For our numerical experiments we have used the Web site of INRIA (http://www.inria.fr). It is a typical Web site with about 300 000 Web pages and 2 200 000 hyperlinks. Since the Web has a fractal structure \[[10](#bib.bib10)\], we expect that our dataset is sufficiently representative. Accordingly, datasets of similar or even smaller sizes have been extensively used in experimental studies of novel algorithms for PageRank computation \[[1](#bib.bib1), [16](#bib.bib16), [17](#bib.bib17)\]. To collect the Web graph data, we construct our own Web crawler which works with the Oracle database. The crawler consists of two parts: the first part is realized in Java and is responsible for downloading pages from the Internet, parsing the pages, and inserting their hyperlinks into the database; the second part is written in PL/SQL and is responsible for the data management. For detailed description of the crawler reader is referred to \[[3](#bib.bib3)\].
As was shown in \[[7](#bib.bib7), [15](#bib.bib15)\], a Web graph has three major distinct components: IN, OUT and SCC. However, if one takes into account the artificial links from the dangling nodes, a Web graph has two major distinct components: POUT and ESCC \[[5](#bib.bib5)\]. In our experiments we consider the artificial links from the dangling nodes and compute $`{\overline{\pi}}_{T}`$, $`{\overset{\sim}{\pi}}_{T}`$, $`{\hat{\pi}}_{T}`$, and $`{\check{\pi}}_{T}`$ with 5 digits precision. We provide the statistics for the INRIA Web site in Table 1.
$`I\hspace{0pt}N\hspace{0pt}R\hspace{0pt}I\hspace{0pt}A`$
Total size 318585
Number of nodes in SCC 154142
Number of nodes in IN 0
Number of nodes in OUT 164443
Number of nodes in ESCC 300682
Number of nodes in POUT 17903
Number of SCCs in OUT 1148
Number of SCCs in POUT 631
Table 1: Component sizes in INRIA dataset
For each pair of these vectors we calculated Kendall Tau metric (see Table 2). The Kendall Tau metric shows how two rankings are different in terms of the number of swaps which are needed to transform one ranking to the other. The Kendall Tau metric has the value of one if two rankings are identical and minus one if one ranking is the inverse of the other.
$`{\overline{\pi}}_{T}`$ $`{\overset{\sim}{\pi}}_{T}`$ $`{\hat{\pi}}_{T}`$ $`{\check{\pi}}_{T}`$
$`{\overline{\pi}}_{T}`$ $`1.0`$ $`0.99390`$ $`0.99498`$ $`0.98228`$
$`{\overset{\sim}{\pi}}_{T}`$ $`1.0`$ $`0.99770`$ $`0.98786`$
$`{\hat{\pi}}_{T}`$ $`1.0`$ $`0.98597`$
$`{\check{\pi}}_{T}`$ $`1.0`$
Table 2: Kendall Tau comparison
In our case, the Kendall Tau metrics for all the pairs is very close to one. Thus, we can conclude that all four quasi-stationarity based centrality measures produce very similar rankings.
We have also analyzed the Kendall Tau metric between $`{\overset{\sim}{\pi}}_{T}`$ and PageRank of ESCC as a function of damping factor (see Figure [1](#S3.F1 "Figure 1 ‣ 3 Numerical experiments and Applications")). As $`c`$ goes to one, the Kendall Tau approaches one. This is in agreement with Proposition [1](#Thmproposition1 "Proposition 1 ‣ 2 Quasi-stationary distributions as centrality measures").
[Refer to caption](/html/0708.0522/assets/x1.png)
Figure 1: The Kendall Tau metric between $`{\overset{\sim}{\pi}}_{T}`$ and PageRank of ESCC as a function of the damping factor.
Finally, we would like to note that in the case of quasi-stationarity based centrality measures the first ranking places were occupied by the sites with the internal structure depicted in Figure [2](#S3.F2 "Figure 2 ‣ 3 Numerical experiments and Applications"). Therefore, we suggest to use the quasi-stationarity based centrality measures to detect “link farms” and to discover photo albums. It turns out that the quasi-stationarity based centrality measures highlights the sites with structure as in Figure [2](#S3.F2 "Figure 2 ‣ 3 Numerical experiments and Applications") but at the same time the relative ranking of the other sites provided by the standard PageRank with $`c = 0.85`$ is preserved. To illustrate this fact, we give in Table 3 rankings of some sites under different centrality measures. Even though the absolute value of ranking is changing, the relative ranking among these sites is the same for all centrality measures. This indicates that the quasi-stationarity based centrality measures help to discover “link farms” and photo albums and at the same time the ranking of sites of the other type stays consistent with the standard PageRank ranking.
[Refer to caption](/html/0708.0522/assets/x2.png)
Figure 2: The album like Web site structure
$`\pi_{T}\hspace{0pt}{(0.85)}`$ $`{\overline{\pi}}_{T}`$ $`{\overset{\sim}{\pi}}_{T}`$ $`{\hat{\pi}}_{T}`$ $`{\check{\pi}}_{T}`$
$`http://www.inria.fr/`$ $`1`$ $`31`$ $`189`$ $`105`$ $`200`$
$`http://www.loria.fr/`$ $`13`$ $`310`$ $`1605`$ $`356`$ $`1633`$
$`http://www.irisa.fr/`$ $`16`$ $`432`$ $`1696`$ $`460`$ $`757`$
$`http://www - sop.inria.fr/`$ $`30`$ $`508`$ $`1825`$ $`532`$ $`1819`$
$`http://www - rocq.inria.fr/`$ $`74`$ $`1333`$ $`2099`$ $`1408`$ $`2158`$
$`http://www - futurs.inria.fr/`$ $`102`$ $`2201`$ $`2360`$ $`2206`$ $`2404`$
Table 3: Examples of sites’ rankings
## 4 Conclusion
In the paper we have proposed centrality measures which can be applied to a reducible graph to avoid the absorbtion problem. In Google PageRank the problem was solved by introduction of uniform random jumps with some probability. Up to the present, there is no clear criterion for the choice this parameter. In the paper we have suggested four quasi-stationarity based parameter-free centrality measures, analyzed them and concluded that they produce approximately the same ranking. Therefore, in practice it is sufficient to compute only one quasi-stationarity based centrality measure. All our theoretical results are confirmed by numerical experiments. The numerical experiments have also showed that the new centrality measures can be applied in spam detection to detect “link farms” and in image search to find photo albums.
## References
- \[1\]
S. Abiteboul, M. Preda, and G. Cobena, “Adaptive on-line page importance computation”, in Proceedings of the 12 International World Wide Web Conference, Budapest, 2003.
- \[2\]
K. Avrachenkov, Analytic Perturbation Theory and its Applications, PhD thesis, University of South Australia, 1999.
- \[3\]
K. Avrachenkov, D. Nemirovsky, and N. Osipova. “Web Graph Analyzer Tool”. In Proceedings of the IEEE ValueTools conference, 2006.
- \[4\]
K. Avrachenkov, M. Haviv and P.G. Howlett, “Inversion of analytic matrix functions that are singular at the origin”, SIAM Journal on Matrix Analysis and Applications, v. 22(4), pp.1175-1189, 2001.
- \[5\]
K. Avrachenkov, N. Litvak and K.S. Pham, “A singular perturbation approach for choosing PageRank damping factor”, preprint, available at http://arxiv.org/abs/math.PR/0612079, 2006.
- \[6\]
P. Boldi, M. Santini, and S. Vigna, “PageRank as a function of the damping factor”, in Proceedings of the 14 World Wide Web Conference, New York, 2005.
- \[7\]
A. Broder, R. Kumar, F. Maghoul, P. Raghavan, S. Rajagopalan, R. Stata, A. Tomkins and J. Wiener, “Graph structure in the Web”, Computer Networks, v. 33, pp.309-320, 2000.
- \[8\]
P. Chen, H. Xie, S. Maslov, and S. Redner, “Finding scientific gems with Google’s PageRank algorithm”, Journal of Informetrics, v.1, pp.8–15, 2007.
- \[9\]
J. N. Darroch and E. Seneta, “On Quasi-Stationary Distributions in Absorbing Discrete-Time Finite Markov Chains”, Journal of Applied Probability, v. 2(1), pp.88-100, 1965.
- \[10\]
S. Dill, R. Kumar, K. McCurley, S. Rajagopalan, D. Sivakumar, and A. Tomkins, “Self-similarity in the Web”, ACM Trans. Internet Technol., 2 (2002), pp. 205–223.
- \[11\]
E.A. van Doorn, “Quasi-stationary distributions and convergence to quasi-stationarity of birth-death processes”, Advances in Applied Probability, v. 23(4), pp. 683-700, 1991.
- \[12\]
W.J. Ewens, “The diffusion equation and pseudo-distribution in genetics”, J.R. Statist. Soc. B, v. 25, pp. 405-412, 1963.
- \[13\]
J. Kleinberg, “Authoritative sources in a hyperlinked environment”, Journal of ACM, v. 46, pp.604-632, 1999.
- \[14\]
I. Kontoyiannis, and S. P. Meyn, “Spectral theory and limit theorems for geometrically ergodic Markov processes”, Ann. Appl. Probab., v. 13, no. 1, pp. 304362, 2003.
- \[15\]
R. Kumar, P. Raghavan, S. Rajagopalan, D. Sivakumar, A. Tompkins and E. Upfal, “The Web as a graph”, PODS’00: Proceedings of the nineteenth ACM SIGMOD-SIGACT-SIGART Symposium on Principles of Database Systems, pp. 1-10, 2000.
- \[16\]
A. N. Langville and C. D. Meyer, “Deeper Inside PageRank”, Internet Math., 1 (2004), pp. 335–400; also available online at http://www4.ncsu.edu/$`_{^{\sim}}`$anlangvi/.
- \[17\]
A. N. Langville and C. D. Meyer, “Updating PageRank with iterative aggregation”, in Proceedings of the 13th World Wide Web Conference, New York, 2004.
- \[18\]
A.N. Langville and C.D. Meyer, “Google’s PageRank and Beyond: The Science of Search Engine Rankings”, Princeton University Press, 2006.
- \[19\]
R. Lempel and S. Moran, “The stochastic approach for link-structure analysis (SALSA) and the TKC effect”, Computer Networks, v. 33, pp. 387-401, 2000.
- \[20\]
C.D. Moler and K.A. Moler, “Numerical Computing with MATLAB”, SIAM , 2003.
- \[21\] L. Page, S. Brin, R. Motwani and T. Winograd, “The pagerank citation ranking: Bringing order to the web”, Stanford Technical Report, 1998.
- \[22\]
E. Seneta, “Non-negative matrices and Markov chains”, Springer, 1973.
Appendix
Here we present a couple of important auxiliary results.
######
Lemma 1
Let $`\overline{T}`$ be an irreducible stochastic matrix. And let $`{T\hspace{0pt}{(\varepsilon)}} = {\overline{T} - {\varepsilon\hspace{0pt}D}}`$ be a perturbation of $`\overline{T}`$ such that $`T\hspace{0pt}{(\varepsilon)}`$ is substochastic matrix. Then, for sufficiently small $`\varepsilon`$ the following Laurent series expansion holds
<table>
<tbody>
<tr>
<td></td>
<td>[I−T(ε)]−1=1εX−1+X0+εX1+…,superscriptdelimited-[]𝐼𝑇𝜀11𝜀subscript𝑋1subscript𝑋0𝜀subscript𝑋1…[I-T(\varepsilon)]^{-1}=\frac{1}{\varepsilon}X_{-1}+X_{0}+\varepsilon X_{1}+\ldots,</td>
<td></td>
<td>(28)</td>
</tr>
</tbody>
</table>
with
<table>
<tbody>
<tr>
<td></td>
<td>X−1=1π¯D𝟏𝟏π¯,subscript𝑋11¯𝜋𝐷11¯𝜋X_{-1}=\frac{1}{\bar{\pi}D{\bf 1}}{\bf 1}\bar{\pi},</td>
<td></td>
<td>(29)</td>
</tr>
</tbody>
</table>
<table>
<tbody>
<tr>
<td></td>
<td>X0=(I−X−1D)H(I−DX−1),subscript𝑋0𝐼subscript𝑋1𝐷𝐻𝐼𝐷subscript𝑋1X_{0}=(I-X_{-1}D)H(I-DX_{-1}),</td>
<td></td>
<td>(30)</td>
</tr>
</tbody>
</table>
where $`\overline{\pi}`$ is the stationary distribution of $`\overline{T}`$ and $`H = {{({{I - \overline{T}} + {\mathbf{1}\hspace{0pt}\overline{\pi}}})}^{- 1} - {\mathbf{1}\hspace{0pt}\overline{\pi}}}`$ is the deviation matrix.
Proof: The proof of this result is based on the approach developed in \[[2](#bib.bib2), [4](#bib.bib4)\]. The existence of the Laurent series ([28](#S4.E28 "In Lemma 1")) is a particular case of more general results of \[[4](#bib.bib4)\]. To calculate the terms of the Laurent series, let us equate the terms with the same powers of $`\varepsilon`$ in the following identity
<table>
<tbody>
<tr>
<td></td>
<td>(I−T¯+εD)(1εX−1+X0+εX1+…)=I,𝐼¯𝑇𝜀𝐷1𝜀subscript𝑋1subscript𝑋0𝜀subscript𝑋1…𝐼(I-\bar{T}+\varepsilon D)(\frac{1}{\varepsilon}X_{-1}+X_{0}+\varepsilon X_{1}+\ldots)=I,</td>
<td></td>
</tr>
</tbody>
</table>
which results in
<table>
<tbody>
<tr>
<td></td>
<td>(I−T¯)X−1=0,𝐼¯𝑇subscript𝑋10(I-\bar{T})X_{-1}=0,</td>
<td></td>
<td>(31)</td>
</tr>
</tbody>
</table>
<table>
<tbody>
<tr>
<td></td>
<td>(I−T¯)X0+DX−1=I,𝐼¯𝑇subscript𝑋0𝐷subscript𝑋1𝐼(I-\bar{T})X_{0}+DX_{-1}=I,</td>
<td></td>
<td>(32)</td>
</tr>
</tbody>
</table>
<table>
<tbody>
<tr>
<td></td>
<td>(I−T¯)X1+DX0=0.𝐼¯𝑇subscript𝑋1𝐷subscript𝑋00(I-\bar{T})X_{1}+DX_{0}=0.</td>
<td></td>
<td>(33)</td>
</tr>
</tbody>
</table>
From equation ([31](#S4.E31)) we conclude that
<table>
<tbody>
<tr>
<td></td>
<td>X−1=𝟏μ−1,subscript𝑋11subscript𝜇1X_{-1}={\bf 1}\mu_{-1},</td>
<td></td>
<td>(34)</td>
</tr>
</tbody>
</table>
where $`\mu_{- 1}`$ is some vector. We find this vector from the condition that the equation ([32](#S4.E32)) has a solution. In particular, equation ([32](#S4.E32)) has a solution if and only if
<table>
<tbody>
<tr>
<td></td>
<td>π¯(I−DX−1)=0.¯𝜋𝐼𝐷subscript𝑋10\bar{\pi}(I-DX_{-1})=0.</td>
<td></td>
</tr>
</tbody>
</table>
By substituting into the above equation the expression ([34](#S4.E34)), we obtain
<table>
<tbody>
<tr>
<td></td>
<td>π¯−π¯D𝟏μ−1=0,¯𝜋¯𝜋𝐷1subscript𝜇10\bar{\pi}-\bar{\pi}D{\bf 1}\mu_{-1}=0,</td>
<td></td>
</tr>
</tbody>
</table>
and, consequently,
<table>
<tbody>
<tr>
<td></td>
<td>μ−1=1π¯D𝟏π¯,subscript𝜇11¯𝜋𝐷1¯𝜋\mu_{-1}=\frac{1}{\bar{\pi}D{\bf 1}}\bar{\pi},</td>
<td></td>
</tr>
</tbody>
</table>
which together with ([34](#S4.E34)) gives ([29](#S4.E29 "In Lemma 1")).
Since the deviation matrix $`H`$ is a Moore-Penrose generalized inverse of $`I - \overline{T}`$, the general solution of equation ([32](#S4.E32)) with respect to $`X_{0}`$ is given by
<table>
<tbody>
<tr>
<td></td>
<td>X0=H(I−DX−1)+𝟏μ0,subscript𝑋0𝐻𝐼𝐷subscript𝑋11subscript𝜇0X_{0}=H(I-DX_{-1})+{\bf 1}\mu_{0},</td>
<td></td>
<td>(35)</td>
</tr>
</tbody>
</table>
where $`\mu_{0}`$ is some vector. The vector $`\mu_{0}`$ can be found from the condition that the equation ([33](#S4.E33)) has a solution. In particular, equation ([33](#S4.E33)) has a solution if and only if
<table>
<tbody>
<tr>
<td></td>
<td>π¯DX0=0.¯𝜋𝐷subscript𝑋00\bar{\pi}DX_{0}=0.</td>
<td></td>
</tr>
</tbody>
</table>
By substituting into the above equation the expression for the general solution ([35](#S4.E35)), we obtain
<table>
<tbody>
<tr>
<td></td>
<td>π¯DH(I−DX−1)+π¯D𝟏μ0=0.¯𝜋𝐷𝐻𝐼𝐷subscript𝑋1¯𝜋𝐷1subscript𝜇00\bar{\pi}DH(I-DX_{-1})+\bar{\pi}D{\bf 1}\mu_{0}=0.</td>
<td></td>
</tr>
</tbody>
</table>
Consequently, we have
<table>
<tbody>
<tr>
<td></td>
<td>μ0=−1π¯D𝟏π¯DH(I−DX−1)subscript𝜇01¯𝜋𝐷1¯𝜋𝐷𝐻𝐼𝐷subscript𝑋1\mu_{0}=-\frac{1}{\bar{\pi}D{\bf 1}}\bar{\pi}DH(I-DX_{-1})</td>
<td></td>
</tr>
</tbody>
</table>
and we obtain ([30](#S4.E30 "In Lemma 1")).
$`\square`$
######
Proposition 11
<table>
<tbody>
<tr>
<td></td>
<td>P(X1=j|X0=i∧⋀m=1NXm∈S)=TijTj(N−1)𝟏Ti(N)𝟏𝑃subscript𝑋1conditional𝑗subscript𝑋0𝑖superscriptsubscript𝑚1𝑁subscript𝑋𝑚𝑆subscript𝑇𝑖𝑗subscriptsuperscript𝑇𝑁1𝑗1subscriptsuperscript𝑇𝑁𝑖1P\left(X_{1}=j|X_{0}=i\wedge\bigwedge_{m=1}^{N}X_{m}\in S\right)=\frac{T_{ij}T^{(N-1)}_{j}{\bf 1}}{T^{(N)}_{i}{\bf 1}}</td>
<td></td>
</tr>
</tbody>
</table>
Proof:
<table>
<tbody>
<tr>
<td></td>
<td></td>
<td></td>
<td>P(X1=j|X0=i∧⋀m=1NXm∈S)=𝑃subscript𝑋1conditional𝑗subscript𝑋0𝑖superscriptsubscript𝑚1𝑁subscript𝑋𝑚𝑆absent\displaystyle P\left(X_{1}=j|X_{0}=i\wedge\bigwedge_{m=1}^{N}X_{m}\in S\right)=</td>
<td></td>
</tr>
<tr>
<td></td>
<td></td>
<td>=\displaystyle=</td>
<td>P(X0=i∧X1=j∧⋀m=2NXm∈S)P(X0=i∧⋀m=1NXm∈S)𝑃subscript𝑋0𝑖subscript𝑋1𝑗superscriptsubscript𝑚2𝑁subscript𝑋𝑚𝑆𝑃subscript𝑋0𝑖superscriptsubscript𝑚1𝑁subscript𝑋𝑚𝑆\displaystyle\frac{P\left(X_{0}=i\wedge X_{1}=j\wedge\bigwedge_{m=2}^{N}X_{m}\in S\right)}{P\left(X_{0}=i\wedge\bigwedge_{m=1}^{N}X_{m}\in S\right)}</td>
<td></td>
</tr>
</tbody>
</table>
Denominator:
<table>
<tbody>
<tr>
<td></td>
<td></td>
<td></td>
<td>P(X0=i∧⋀m=1NXm∈S)=𝑃subscript𝑋0𝑖superscriptsubscript𝑚1𝑁subscript𝑋𝑚𝑆absent\displaystyle P\left(X_{0}=i\wedge\bigwedge_{m=1}^{N}X_{m}\in S\right)=</td>
<td></td>
</tr>
<tr>
<td></td>
<td></td>
<td>=\displaystyle=</td>
<td>P(X0=i∧⋀m=1N⋁km∈SXm=km)=𝑃subscript𝑋0𝑖superscriptsubscript𝑚1𝑁subscriptsubscript𝑘𝑚𝑆subscript𝑋𝑚subscript𝑘𝑚absent\displaystyle P\left(X_{0}=i\wedge\bigwedge_{m=1}^{N}\bigvee_{k_{m}\in S}X_{m}=k_{m}\right)=</td>
<td></td>
</tr>
<tr>
<td></td>
<td></td>
<td>=\displaystyle=</td>
<td>P(X0=i∧⋁k1∈SX1=k1∧⋀m=2N⋁km∈SXm=km)=𝑃subscript𝑋0𝑖subscriptsubscript𝑘1𝑆subscript𝑋1subscript𝑘1superscriptsubscript𝑚2𝑁subscriptsubscript𝑘𝑚𝑆subscript𝑋𝑚subscript𝑘𝑚absent\displaystyle P\left(X_{0}=i\wedge\bigvee_{k_{1}\in S}X_{1}=k_{1}\wedge\bigwedge_{m=2}^{N}\bigvee_{k_{m}\in S}X_{m}=k_{m}\right)=</td>
<td></td>
</tr>
<tr>
<td></td>
<td></td>
<td>=\displaystyle=</td>
<td>P(X0=i)∑k1∈SP(X1=k1∧⋀m=2N⋁km∈SXm=km)=𝑃subscript𝑋0𝑖subscriptsubscript𝑘1𝑆𝑃subscript𝑋1subscript𝑘1superscriptsubscript𝑚2𝑁subscriptsubscript𝑘𝑚𝑆subscript𝑋𝑚subscript𝑘𝑚absent\displaystyle P\left(X_{0}=i\right)\sum_{k_{1}\in S}P\left(X_{1}=k_{1}\wedge\bigwedge_{m=2}^{N}\bigvee_{k_{m}\in S}X_{m}=k_{m}\right)=</td>
<td></td>
</tr>
<tr>
<td></td>
<td></td>
<td>=\displaystyle=</td>
<td>P(X0=i)∑k1∈SP(X1=k1|X0=i)P(⋀m=2N⋁km∈SXm=km|X1=k1)=𝑃subscript𝑋0𝑖subscriptsubscript𝑘1𝑆𝑃subscript𝑋1conditionalsubscript𝑘1subscript𝑋0𝑖𝑃superscriptsubscript𝑚2𝑁subscriptsubscript𝑘𝑚𝑆subscript𝑋𝑚conditionalsubscript𝑘𝑚subscript𝑋1subscript𝑘1absent\displaystyle P\left(X_{0}=i\right)\sum_{k_{1}\in S}P\left(X_{1}=k_{1}|X_{0}=i\right)P\left(\bigwedge_{m=2}^{N}\bigvee_{k_{m}\in S}X_{m}=k_{m}|X_{1}=k_{1}\right)=</td>
<td></td>
</tr>
<tr>
<td></td>
<td></td>
<td>=\displaystyle=</td>
<td>P(X0=i)∑k1∈SP(X1=k1|X0=i)P(⋁k2∈SX2=k2∧⋀m=3N⋁km∈SXm=km|X1=k1)=𝑃subscript𝑋0𝑖subscriptsubscript𝑘1𝑆𝑃subscript𝑋1conditionalsubscript𝑘1subscript𝑋0𝑖𝑃subscriptsubscript𝑘2𝑆subscript𝑋2subscript𝑘2superscriptsubscript𝑚3𝑁subscriptsubscript𝑘𝑚𝑆subscript𝑋𝑚conditionalsubscript𝑘𝑚subscript𝑋1subscript𝑘1absent\displaystyle P\left(X_{0}=i\right)\sum_{k_{1}\in S}P\left(X_{1}=k_{1}|X_{0}=i\right)P\left(\bigvee_{k_{2}\in S}X_{2}=k_{2}\wedge\bigwedge_{m=3}^{N}\bigvee_{k_{m}\in S}X_{m}=k_{m}|X_{1}=k_{1}\right)=</td>
<td></td>
</tr>
<tr>
<td></td>
<td></td>
<td>=\displaystyle=</td>
<td>P(X0=i)∑k1∈SP(X1=k1|X0=i)∑k2∈SP(X2=k2∧⋀m=3N⋁km∈SXm=km|X1=k1)=𝑃subscript𝑋0𝑖subscriptsubscript𝑘1𝑆𝑃subscript𝑋1conditionalsubscript𝑘1subscript𝑋0𝑖subscriptsubscript𝑘2𝑆𝑃subscript𝑋2subscript𝑘2superscriptsubscript𝑚3𝑁subscriptsubscript𝑘𝑚𝑆subscript𝑋𝑚conditionalsubscript𝑘𝑚subscript𝑋1subscript𝑘1absent\displaystyle P\left(X_{0}=i\right)\sum_{k_{1}\in S}P\left(X_{1}=k_{1}|X_{0}=i\right)\sum_{k_{2}\in S}P\left(X_{2}=k_{2}\wedge\bigwedge_{m=3}^{N}\bigvee_{k_{m}\in S}X_{m}=k_{m}|X_{1}=k_{1}\right)=</td>
<td></td>
</tr>
<tr>
<td></td>
<td></td>
<td>=\displaystyle=</td>
<td>P(X0=i)∑k1∈SP(X1=k1|X0=i)𝑃subscript𝑋0𝑖subscriptsubscript𝑘1𝑆𝑃subscript𝑋1conditionalsubscript𝑘1subscript𝑋0𝑖\displaystyle P\left(X_{0}=i\right)\sum_{k_{1}\in S}P\left(X_{1}=k_{1}|X_{0}=i\right)</td>
<td></td>
</tr>
<tr>
<td></td>
<td></td>
<td></td>
<td>∑k2∈SP(⋀m=3N⋁km∈SXm=km|X2=k2∧X1=k1)P(X2=k2|X1=k1)=subscriptsubscript𝑘2𝑆𝑃superscriptsubscript𝑚3𝑁subscriptsubscript𝑘𝑚𝑆subscript𝑋𝑚conditionalsubscript𝑘𝑚subscript𝑋2subscript𝑘2subscript𝑋1subscript𝑘1𝑃subscript𝑋2conditionalsubscript𝑘2subscript𝑋1subscript𝑘1absent\displaystyle\sum_{k_{2}\in S}P\left(\bigwedge_{m=3}^{N}\bigvee_{k_{m}\in S}X_{m}=k_{m}|X_{2}=k_{2}\wedge X_{1}=k_{1}\right)P\left(X_{2}=k_{2}|X_{1}=k_{1}\right)=</td>
<td></td>
</tr>
<tr>
<td></td>
<td></td>
<td>=\displaystyle=</td>
<td>P(X0=i)∑k1∈SP(X1=k1|X0=i)𝑃subscript𝑋0𝑖subscriptsubscript𝑘1𝑆𝑃subscript𝑋1conditionalsubscript𝑘1subscript𝑋0𝑖\displaystyle P\left(X_{0}=i\right)\sum_{k_{1}\in S}P\left(X_{1}=k_{1}|X_{0}=i\right)</td>
<td></td>
</tr>
<tr>
<td></td>
<td></td>
<td></td>
<td>∑k2∈SP(⋀m=3N⋁km∈SXm=km|X2=k2)P(X2=k2|X1=k1)=subscriptsubscript𝑘2𝑆𝑃superscriptsubscript𝑚3𝑁subscriptsubscript𝑘𝑚𝑆subscript𝑋𝑚conditionalsubscript𝑘𝑚subscript𝑋2subscript𝑘2𝑃subscript𝑋2conditionalsubscript𝑘2subscript𝑋1subscript𝑘1absent\displaystyle\sum_{k_{2}\in S}P\left(\bigwedge_{m=3}^{N}\bigvee_{k_{m}\in S}X_{m}=k_{m}|X_{2}=k_{2}\right)P\left(X_{2}=k_{2}|X_{1}=k_{1}\right)=</td>
<td></td>
</tr>
<tr>
<td></td>
<td></td>
<td>=\displaystyle=</td>
<td>P(X0=i)∑k1∈SP(X1=k1|X0=i)𝑃subscript𝑋0𝑖subscriptsubscript𝑘1𝑆𝑃subscript𝑋1conditionalsubscript𝑘1subscript𝑋0𝑖\displaystyle P\left(X_{0}=i\right)\sum_{k_{1}\in S}P\left(X_{1}=k_{1}|X_{0}=i\right)</td>
<td></td>
</tr>
<tr>
<td></td>
<td></td>
<td></td>
<td>∑k2∈SP(⋀m=3N⋁km∈SXm=km|X2=k2)P(X2=k2|X1=k1)=subscriptsubscript𝑘2𝑆𝑃superscriptsubscript𝑚3𝑁subscriptsubscript𝑘𝑚𝑆subscript𝑋𝑚conditionalsubscript𝑘𝑚subscript𝑋2subscript𝑘2𝑃subscript𝑋2conditionalsubscript𝑘2subscript𝑋1subscript𝑘1absent\displaystyle\sum_{k_{2}\in S}P\left(\bigwedge_{m=3}^{N}\bigvee_{k_{m}\in S}X_{m}=k_{m}|X_{2}=k_{2}\right)P\left(X_{2}=k_{2}|X_{1}=k_{1}\right)=</td>
<td></td>
</tr>
<tr>
<td></td>
<td></td>
<td>=\displaystyle=</td>
<td>P(X0=i)∑k2∈SP(⋀m=3N⋁km∈SXm=km|X2=k2)𝑃subscript𝑋0𝑖subscriptsubscript𝑘2𝑆𝑃superscriptsubscript𝑚3𝑁subscriptsubscript𝑘𝑚𝑆subscript𝑋𝑚conditionalsubscript𝑘𝑚subscript𝑋2subscript𝑘2\displaystyle P\left(X_{0}=i\right)\sum_{k_{2}\in S}P\left(\bigwedge_{m=3}^{N}\bigvee_{k_{m}\in S}X_{m}=k_{m}|X_{2}=k_{2}\right)</td>
<td></td>
</tr>
<tr>
<td></td>
<td></td>
<td></td>
<td>∑k1∈SP(X2=k2|X1=k1)P(X1=k1|X0=i)=subscriptsubscript𝑘1𝑆𝑃subscript𝑋2conditionalsubscript𝑘2subscript𝑋1subscript𝑘1𝑃subscript𝑋1conditionalsubscript𝑘1subscript𝑋0𝑖absent\displaystyle\sum_{k_{1}\in S}P\left(X_{2}=k_{2}|X_{1}=k_{1}\right)P\left(X_{1}=k_{1}|X_{0}=i\right)=</td>
<td></td>
</tr>
<tr>
<td></td>
<td></td>
<td>=\displaystyle=</td>
<td>P(X0=i)∑k2∈SP(⋀m=3N⋁km∈SXm=km|X2=k2)P(X2=k2|X0=i)=𝑃subscript𝑋0𝑖subscriptsubscript𝑘2𝑆𝑃superscriptsubscript𝑚3𝑁subscriptsubscript𝑘𝑚𝑆subscript𝑋𝑚conditionalsubscript𝑘𝑚subscript𝑋2subscript𝑘2𝑃subscript𝑋2conditionalsubscript𝑘2subscript𝑋0𝑖absent\displaystyle P\left(X_{0}=i\right)\sum_{k_{2}\in S}P\left(\bigwedge_{m=3}^{N}\bigvee_{k_{m}\in S}X_{m}=k_{m}|X_{2}=k_{2}\right)P\left(X_{2}=k_{2}|X_{0}=i\right)=</td>
<td></td>
</tr>
<tr>
<td></td>
<td></td>
<td>=\displaystyle=</td>
<td>P(X0=i)∑k2∈SP(⋁k3∈SX3=k3∧⋀m=4N⋁km∈SXm=km|X2=k2)P(X2=k2|X0=i)=𝑃subscript𝑋0𝑖subscriptsubscript𝑘2𝑆𝑃subscriptsubscript𝑘3𝑆subscript𝑋3subscript𝑘3superscriptsubscript𝑚4𝑁subscriptsubscript𝑘𝑚𝑆subscript𝑋𝑚conditionalsubscript𝑘𝑚subscript𝑋2subscript𝑘2𝑃subscript𝑋2conditionalsubscript𝑘2subscript𝑋0𝑖absent\displaystyle P\left(X_{0}=i\right)\sum_{k_{2}\in S}P\left(\bigvee_{k_{3}\in S}X_{3}=k_{3}\wedge\bigwedge_{m=4}^{N}\bigvee_{k_{m}\in S}X_{m}=k_{m}|X_{2}=k_{2}\right)P\left(X_{2}=k_{2}|X_{0}=i\right)=</td>
<td></td>
</tr>
<tr>
<td></td>
<td></td>
<td>=\displaystyle=</td>
<td>P(X0=i)∑k2∈S∑k3∈SP(X3=k3∧⋀m=4N⋁km∈SXm=km|X2=k2)P(X2=k2|X0=i)=𝑃subscript𝑋0𝑖subscriptsubscript𝑘2𝑆subscriptsubscript𝑘3𝑆𝑃subscript𝑋3subscript𝑘3superscriptsubscript𝑚4𝑁subscriptsubscript𝑘𝑚𝑆subscript𝑋𝑚conditionalsubscript𝑘𝑚subscript𝑋2subscript𝑘2𝑃subscript𝑋2conditionalsubscript𝑘2subscript𝑋0𝑖absent\displaystyle P\left(X_{0}=i\right)\sum_{k_{2}\in S}\sum_{k_{3}\in S}P\left(X_{3}=k_{3}\wedge\bigwedge_{m=4}^{N}\bigvee_{k_{m}\in S}X_{m}=k_{m}|X_{2}=k_{2}\right)P\left(X_{2}=k_{2}|X_{0}=i\right)=</td>
<td></td>
</tr>
<tr>
<td></td>
<td></td>
<td>=\displaystyle=</td>
<td>P(X0=i)∑k3∈S∑k2∈SP(⋀m=4N⋁km∈SXm=km|X3=k3∧X2=k2)𝑃subscript𝑋0𝑖subscriptsubscript𝑘3𝑆subscriptsubscript𝑘2𝑆𝑃superscriptsubscript𝑚4𝑁subscriptsubscript𝑘𝑚𝑆subscript𝑋𝑚conditionalsubscript𝑘𝑚subscript𝑋3subscript𝑘3subscript𝑋2subscript𝑘2\displaystyle P\left(X_{0}=i\right)\sum_{k_{3}\in S}\sum_{k_{2}\in S}P\left(\bigwedge_{m=4}^{N}\bigvee_{k_{m}\in S}X_{m}=k_{m}|X_{3}=k_{3}\wedge X_{2}=k_{2}\right)</td>
<td></td>
</tr>
<tr>
<td></td>
<td></td>
<td></td>
<td>P(X3=k3|X2=k2)P(X2=k2|X0=i)=𝑃subscript𝑋3conditionalsubscript𝑘3subscript𝑋2subscript𝑘2𝑃subscript𝑋2conditionalsubscript𝑘2subscript𝑋0𝑖absent\displaystyle P\left(X_{3}=k_{3}|X_{2}=k_{2}\right)P\left(X_{2}=k_{2}|X_{0}=i\right)=</td>
<td></td>
</tr>
<tr>
<td></td>
<td></td>
<td>=\displaystyle=</td>
<td>P(X0=i)∑k3∈S∑k2∈SP(⋀m=4N⋁km∈SXm=km|X3=k3)𝑃subscript𝑋0𝑖subscriptsubscript𝑘3𝑆subscriptsubscript𝑘2𝑆𝑃superscriptsubscript𝑚4𝑁subscriptsubscript𝑘𝑚𝑆subscript𝑋𝑚conditionalsubscript𝑘𝑚subscript𝑋3subscript𝑘3\displaystyle P\left(X_{0}=i\right)\sum_{k_{3}\in S}\sum_{k_{2}\in S}P\left(\bigwedge_{m=4}^{N}\bigvee_{k_{m}\in S}X_{m}=k_{m}|X_{3}=k_{3}\right)</td>
<td></td>
</tr>
<tr>
<td></td>
<td></td>
<td></td>
<td>P(X3=k3|X2=k2)P(X2=k2|X0=i)=𝑃subscript𝑋3conditionalsubscript𝑘3subscript𝑋2subscript𝑘2𝑃subscript𝑋2conditionalsubscript𝑘2subscript𝑋0𝑖absent\displaystyle P\left(X_{3}=k_{3}|X_{2}=k_{2}\right)P\left(X_{2}=k_{2}|X_{0}=i\right)=</td>
<td></td>
</tr>
<tr>
<td></td>
<td></td>
<td>=\displaystyle=</td>
<td>P(X0=i)∑k3∈SP(⋀m=4N⋁km∈SXm=km|X3=k3)𝑃subscript𝑋0𝑖subscriptsubscript𝑘3𝑆𝑃superscriptsubscript𝑚4𝑁subscriptsubscript𝑘𝑚𝑆subscript𝑋𝑚conditionalsubscript𝑘𝑚subscript𝑋3subscript𝑘3\displaystyle P\left(X_{0}=i\right)\sum_{k_{3}\in S}P\left(\bigwedge_{m=4}^{N}\bigvee_{k_{m}\in S}X_{m}=k_{m}|X_{3}=k_{3}\right)</td>
<td></td>
</tr>
<tr>
<td></td>
<td></td>
<td></td>
<td>∑k2∈SP(X3=k3|X2=k2)P(X2=k2|X0=i)=subscriptsubscript𝑘2𝑆𝑃subscript𝑋3conditionalsubscript𝑘3subscript𝑋2subscript𝑘2𝑃subscript𝑋2conditionalsubscript𝑘2subscript𝑋0𝑖absent\displaystyle\sum_{k_{2}\in S}P\left(X_{3}=k_{3}|X_{2}=k_{2}\right)P\left(X_{2}=k_{2}|X_{0}=i\right)=</td>
<td></td>
</tr>
<tr>
<td></td>
<td></td>
<td>=\displaystyle=</td>
<td>P(X0=i)∑k3∈SP(⋀m=4N⋁km∈SXm=km|X3=k3)P(X3=k3|X0=i)=…𝑃subscript𝑋0𝑖subscriptsubscript𝑘3𝑆𝑃superscriptsubscript𝑚4𝑁subscriptsubscript𝑘𝑚𝑆subscript𝑋𝑚conditionalsubscript𝑘𝑚subscript𝑋3subscript𝑘3𝑃subscript𝑋3conditionalsubscript𝑘3subscript𝑋0𝑖…\displaystyle P\left(X_{0}=i\right)\sum_{k_{3}\in S}P\left(\bigwedge_{m=4}^{N}\bigvee_{k_{m}\in S}X_{m}=k_{m}|X_{3}=k_{3}\right)P\left(X_{3}=k_{3}|X_{0}=i\right)=\ldots</td>
<td></td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td>……\displaystyle\ldots</td>
<td>=\displaystyle=</td>
<td>P(X0=i)∑kN−2∈SP(⋀m=N−1N⋁km∈SXm=km|XN−2=kN−2)P(XN−2=kN−2|X0=i)=𝑃subscript𝑋0𝑖subscriptsubscript𝑘𝑁2𝑆𝑃superscriptsubscript𝑚𝑁1𝑁subscriptsubscript𝑘𝑚𝑆subscript𝑋𝑚conditionalsubscript𝑘𝑚subscript𝑋𝑁2subscript𝑘𝑁2𝑃subscript𝑋𝑁2conditionalsubscript𝑘𝑁2subscript𝑋0𝑖absent\displaystyle P\left(X_{0}=i\right)\sum_{k_{N-2}\in S}P\left(\bigwedge_{m=N-1}^{N}\bigvee_{k_{m}\in S}X_{m}=k_{m}|X_{N-2}=k_{N-2}\right)P\left(X_{N-2}=k_{N-2}|X_{0}=i\right)=</td>
<td></td>
</tr>
<tr>
<td></td>
<td></td>
<td>=\displaystyle=</td>
<td>P(X0=i)∑kN−2∈SP(⋁kN−1∈SXN−1=kN−1∧⋁kN∈SXN=kN|XN−2=kN−2)𝑃subscript𝑋0𝑖subscriptsubscript𝑘𝑁2𝑆𝑃subscriptsubscript𝑘𝑁1𝑆subscript𝑋𝑁1subscript𝑘𝑁1subscriptsubscript𝑘𝑁𝑆subscript𝑋𝑁conditionalsubscript𝑘𝑁subscript𝑋𝑁2subscript𝑘𝑁2\displaystyle P\left(X_{0}=i\right)\sum_{k_{N-2}\in S}P\left(\bigvee_{k_{N-1}\in S}X_{N-1}=k_{N-1}\wedge\bigvee_{k_{N}\in S}X_{N}=k_{N}|X_{N-2}=k_{N-2}\right)</td>
<td></td>
</tr>
<tr>
<td></td>
<td></td>
<td></td>
<td>P(XN−2=kN−2|X0=i)=𝑃subscript𝑋𝑁2conditionalsubscript𝑘𝑁2subscript𝑋0𝑖absent\displaystyle P\left(X_{N-2}=k_{N-2}|X_{0}=i\right)=</td>
<td></td>
</tr>
<tr>
<td></td>
<td></td>
<td>=\displaystyle=</td>
<td>P(X0=i)∑kN−2∈S∑kN−1∈SP(XN−1=kN−1∧⋁kN∈SXN=kN|XN−2=kN−2)𝑃subscript𝑋0𝑖subscriptsubscript𝑘𝑁2𝑆subscriptsubscript𝑘𝑁1𝑆𝑃subscript𝑋𝑁1subscript𝑘𝑁1subscriptsubscript𝑘𝑁𝑆subscript𝑋𝑁conditionalsubscript𝑘𝑁subscript𝑋𝑁2subscript𝑘𝑁2\displaystyle P\left(X_{0}=i\right)\sum_{k_{N-2}\in S}\sum_{k_{N-1}\in S}P\left(X_{N-1}=k_{N-1}\wedge\bigvee_{k_{N}\in S}X_{N}=k_{N}|X_{N-2}=k_{N-2}\right)</td>
<td></td>
</tr>
<tr>
<td></td>
<td></td>
<td></td>
<td>P(XN−2=kN−2|X0=i)=𝑃subscript𝑋𝑁2conditionalsubscript𝑘𝑁2subscript𝑋0𝑖absent\displaystyle P\left(X_{N-2}=k_{N-2}|X_{0}=i\right)=</td>
<td></td>
</tr>
<tr>
<td></td>
<td></td>
<td>=\displaystyle=</td>
<td>P(X0=i)∑kN−2∈S∑kN−1∈SP(⋁kN∈SXN=kN|XN−1=kN−1∧XN−2=kN−2)𝑃subscript𝑋0𝑖subscriptsubscript𝑘𝑁2𝑆subscriptsubscript𝑘𝑁1𝑆𝑃subscriptsubscript𝑘𝑁𝑆subscript𝑋𝑁conditionalsubscript𝑘𝑁subscript𝑋𝑁1subscript𝑘𝑁1subscript𝑋𝑁2subscript𝑘𝑁2\displaystyle P\left(X_{0}=i\right)\sum_{k_{N-2}\in S}\sum_{k_{N-1}\in S}P\left(\bigvee_{k_{N}\in S}X_{N}=k_{N}|X_{N-1}=k_{N-1}\wedge X_{N-2}=k_{N-2}\right)</td>
<td></td>
</tr>
<tr>
<td></td>
<td></td>
<td></td>
<td>P(XN−1=kN−1|XN−2=kN−2)P(XN−2=kN−2|X0=i)=𝑃subscript𝑋𝑁1conditionalsubscript𝑘𝑁1subscript𝑋𝑁2subscript𝑘𝑁2𝑃subscript𝑋𝑁2conditionalsubscript𝑘𝑁2subscript𝑋0𝑖absent\displaystyle P\left(X_{N-1}=k_{N-1}|X_{N-2}=k_{N-2}\right)P\left(X_{N-2}=k_{N-2}|X_{0}=i\right)=</td>
<td></td>
</tr>
<tr>
<td></td>
<td></td>
<td>=\displaystyle=</td>
<td>P(X0=i)∑kN−2∈S∑kN−1∈SP(⋁kN∈SXN=kN|XN−1=kN−1)𝑃subscript𝑋0𝑖subscriptsubscript𝑘𝑁2𝑆subscriptsubscript𝑘𝑁1𝑆𝑃subscriptsubscript𝑘𝑁𝑆subscript𝑋𝑁conditionalsubscript𝑘𝑁subscript𝑋𝑁1subscript𝑘𝑁1\displaystyle P\left(X_{0}=i\right)\sum_{k_{N-2}\in S}\sum_{k_{N-1}\in S}P\left(\bigvee_{k_{N}\in S}X_{N}=k_{N}|X_{N-1}=k_{N-1}\right)</td>
<td></td>
</tr>
<tr>
<td></td>
<td></td>
<td></td>
<td>P(XN−1=kN−1|XN−2=kN−2)P(XN−2=kN−2|X0=i)=𝑃subscript𝑋𝑁1conditionalsubscript𝑘𝑁1subscript𝑋𝑁2subscript𝑘𝑁2𝑃subscript𝑋𝑁2conditionalsubscript𝑘𝑁2subscript𝑋0𝑖absent\displaystyle P\left(X_{N-1}=k_{N-1}|X_{N-2}=k_{N-2}\right)P\left(X_{N-2}=k_{N-2}|X_{0}=i\right)=</td>
<td></td>
</tr>
<tr>
<td></td>
<td></td>
<td>=\displaystyle=</td>
<td>P(X0=i)∑kN−1∈SP(⋁kN∈SXN=kN|XN−1=kN−1)𝑃subscript𝑋0𝑖subscriptsubscript𝑘𝑁1𝑆𝑃subscriptsubscript𝑘𝑁𝑆subscript𝑋𝑁conditionalsubscript𝑘𝑁subscript𝑋𝑁1subscript𝑘𝑁1\displaystyle P\left(X_{0}=i\right)\sum_{k_{N-1}\in S}P\left(\bigvee_{k_{N}\in S}X_{N}=k_{N}|X_{N-1}=k_{N-1}\right)</td>
<td></td>
</tr>
<tr>
<td></td>
<td></td>
<td></td>
<td>∑kN−2∈SP(XN−1=kN−1|XN−2=kN−2)P(XN−2=kN−2|X0=i)=subscriptsubscript𝑘𝑁2𝑆𝑃subscript𝑋𝑁1conditionalsubscript𝑘𝑁1subscript𝑋𝑁2subscript𝑘𝑁2𝑃subscript𝑋𝑁2conditionalsubscript𝑘𝑁2subscript𝑋0𝑖absent\displaystyle\sum_{k_{N-2}\in S}P\left(X_{N-1}=k_{N-1}|X_{N-2}=k_{N-2}\right)P\left(X_{N-2}=k_{N-2}|X_{0}=i\right)=</td>
<td></td>
</tr>
<tr>
<td></td>
<td></td>
<td>=\displaystyle=</td>
<td>P(X0=i)∑kN−1∈SP(⋁kN∈SXN=kN|XN−1=kN−1)P(XN−1=kN−1|X0=i)=𝑃subscript𝑋0𝑖subscriptsubscript𝑘𝑁1𝑆𝑃subscriptsubscript𝑘𝑁𝑆subscript𝑋𝑁conditionalsubscript𝑘𝑁subscript𝑋𝑁1subscript𝑘𝑁1𝑃subscript𝑋𝑁1conditionalsubscript𝑘𝑁1subscript𝑋0𝑖absent\displaystyle P\left(X_{0}=i\right)\sum_{k_{N-1}\in S}P\left(\bigvee_{k_{N}\in S}X_{N}=k_{N}|X_{N-1}=k_{N-1}\right)P\left(X_{N-1}=k_{N-1}|X_{0}=i\right)=</td>
<td></td>
</tr>
<tr>
<td></td>
<td></td>
<td>=\displaystyle=</td>
<td>P(X0=i)∑kN−1∈S∑kN∈SP(XN=kN|XN−1=kN−1)P(XN−1=kN−1|X0=i)=𝑃subscript𝑋0𝑖subscriptsubscript𝑘𝑁1𝑆subscriptsubscript𝑘𝑁𝑆𝑃subscript𝑋𝑁conditionalsubscript𝑘𝑁subscript𝑋𝑁1subscript𝑘𝑁1𝑃subscript𝑋𝑁1conditionalsubscript𝑘𝑁1subscript𝑋0𝑖absent\displaystyle P\left(X_{0}=i\right)\sum_{k_{N-1}\in S}\sum_{k_{N}\in S}P\left(X_{N}=k_{N}|X_{N-1}=k_{N-1}\right)P\left(X_{N-1}=k_{N-1}|X_{0}=i\right)=</td>
<td></td>
</tr>
<tr>
<td></td>
<td></td>
<td>=\displaystyle=</td>
<td>P(X0=i)∑kN∈S∑kN−1∈SP(XN=kN|XN−1=kN−1)P(XN−1=kN−1|X0=i)=𝑃subscript𝑋0𝑖subscriptsubscript𝑘𝑁𝑆subscriptsubscript𝑘𝑁1𝑆𝑃subscript𝑋𝑁conditionalsubscript𝑘𝑁subscript𝑋𝑁1subscript𝑘𝑁1𝑃subscript𝑋𝑁1conditionalsubscript𝑘𝑁1subscript𝑋0𝑖absent\displaystyle P\left(X_{0}=i\right)\sum_{k_{N}\in S}\sum_{k_{N-1}\in S}P\left(X_{N}=k_{N}|X_{N-1}=k_{N-1}\right)P\left(X_{N-1}=k_{N-1}|X_{0}=i\right)=</td>
<td></td>
</tr>
<tr>
<td></td>
<td></td>
<td>=\displaystyle=</td>
<td>P(X0=i)∑kN∈SP(XN=kN|X0=i)=𝑃subscript𝑋0𝑖subscriptsubscript𝑘𝑁𝑆𝑃subscript𝑋𝑁conditionalsubscript𝑘𝑁subscript𝑋0𝑖absent\displaystyle P\left(X_{0}=i\right)\sum_{k_{N}\in S}P\left(X_{N}=k_{N}|X_{0}=i\right)=</td>
<td></td>
</tr>
<tr>
<td></td>
<td></td>
<td>=\displaystyle=</td>
<td>∑kN=1nTTikN(N)P(X0=i)=superscriptsubscriptsubscript𝑘𝑁1subscript𝑛𝑇subscriptsuperscript𝑇𝑁𝑖subscript𝑘𝑁𝑃subscript𝑋0𝑖absent\displaystyle\sum_{k_{N}=1}^{n_{T}}T^{(N)}_{ik_{N}}P\left(X_{0}=i\right)=</td>
<td></td>
</tr>
<tr>
<td></td>
<td></td>
<td>=\displaystyle=</td>
<td>Ti(N)𝟏P(X0=i)=subscriptsuperscript𝑇𝑁𝑖1𝑃subscript𝑋0𝑖absent\displaystyle T^{(N)}_{i}{\bf 1}P\left(X_{0}=i\right)=</td>
<td></td>
</tr>
</tbody>
</table>
<table>
<tbody>
<tr>
<td></td>
<td>P(X0=i∧⋀m=1NXm∈S)=Ti(N)𝟏P(X0=i)𝑃subscript𝑋0𝑖superscriptsubscript𝑚1𝑁subscript𝑋𝑚𝑆subscriptsuperscript𝑇𝑁𝑖1𝑃subscript𝑋0𝑖P\left(X_{0}=i\wedge\bigwedge_{m=1}^{N}X_{m}\in S\right)=T^{(N)}_{i}{\bf 1}P\left(X_{0}=i\right)</td>
<td></td>
</tr>
</tbody>
</table>
Numerator:
<table>
<tbody>
<tr>
<td></td>
<td></td>
<td></td>
<td>P(X0=i∧X1=j∧⋀m=2NXm∈S)=𝑃subscript𝑋0𝑖subscript𝑋1𝑗superscriptsubscript𝑚2𝑁subscript𝑋𝑚𝑆absent\displaystyle P\left(X_{0}=i\wedge X_{1}=j\wedge\bigwedge_{m=2}^{N}X_{m}\in S\right)=</td>
<td></td>
</tr>
<tr>
<td></td>
<td></td>
<td>=\displaystyle=</td>
<td>P(⋀m=2N⋁km∈SXm=km)P(X1=j|X0=i)P(X0=i)=𝑃superscriptsubscript𝑚2𝑁subscriptsubscript𝑘𝑚𝑆subscript𝑋𝑚subscript𝑘𝑚𝑃subscript𝑋1conditional𝑗subscript𝑋0𝑖𝑃subscript𝑋0𝑖absent\displaystyle P\left(\bigwedge_{m=2}^{N}\bigvee_{k_{m}\in S}X_{m}=k_{m}\right)P\left(X_{1}=j|X_{0}=i\right)P\left(X_{0}=i\right)=</td>
<td></td>
</tr>
<tr>
<td></td>
<td></td>
<td>=\displaystyle=</td>
<td>TijTj(N−1)𝟏P(X0=i)=subscript𝑇𝑖𝑗subscriptsuperscript𝑇𝑁1𝑗1𝑃subscript𝑋0𝑖absent\displaystyle T_{ij}T^{(N-1)}_{j}{\bf 1}P\left(X_{0}=i\right)=</td>
<td></td>
</tr>
</tbody>
</table>
<table>
<tbody>
<tr>
<td></td>
<td>P(X0=i∧X1=j∧⋀m=2NXm∈S)=TijTj(N−1)𝟏P(X0=i)𝑃subscript𝑋0𝑖subscript𝑋1𝑗superscriptsubscript𝑚2𝑁subscript𝑋𝑚𝑆subscript𝑇𝑖𝑗subscriptsuperscript𝑇𝑁1𝑗1𝑃subscript𝑋0𝑖P\left(X_{0}=i\wedge X_{1}=j\wedge\bigwedge_{m=2}^{N}X_{m}\in S\right)=T_{ij}T^{(N-1)}_{j}{\bf 1}P\left(X_{0}=i\right)</td>
<td></td>
</tr>
</tbody>
</table>
$`\square`$
###### Contents
1. [](#S1)
1 Introduction
2. [](#S2)
2 Quasi-stationary distributions as centrality measures
3. [](#S3)
3 Numerical experiments and Applications
4. [](#S4)
4 Conclusion |
0709.1080 | cs.CR | cyber | # On the Protocol Composition Logic PCL
<sup>†</sup>
<sup>†</sup>thanks: This work was supported by the Hasler Foundation, ManCom project 2071.
Manuscript
Cas Cremers
Department of Computer Science, ETH Zurich, 8092 Zurich, Switzerland
cas.cremers@inf.ethz.ch
###### Abstract
A recent development in formal security protocol analysis is the Protocol Composition Logic (PCL). We identify a number of problems with this logic as well as with extensions of the logic, as defined in \[[DDMP05](#bib.bibx13), [HSD<sup>+</sup>05](#bib.bibx20), [He05](#bib.bibx19), [Dat05](#bib.bibx8), [Der06](#bib.bibx16), [DDMR07](#bib.bibx14)\]. The identified problems imply strong restrictions on the scope of PCL, and imply that some currently claimed PCL proofs cannot be proven within the logic, or make use of unsound axioms. Where possible, we propose solutions for these problems.
## 1 Introduction
Formally establishing properties of security protocols has been investigated extensively during the last twenty years, but a definite model and a corresponding method for security protocol analysis has remained elusive thus far.
The most successful approaches to security protocol analysis have been focused on tools for finding attacks, which are often based on bounded model checking or constraint solving, e.g. \[[Low97a](#bib.bibx21), [ABB<sup>+</sup>05](#bib.bibx1)\]. When such tools find an attack, one can easily verify manually whether or not the attack actually exists on the protocol. Some tools even allow for unbounded verification of protocols, e.g. \[[Bla01](#bib.bibx3), [Cre06](#bib.bibx7)\]. If no attack is found with such tools, correctness of the protocol follows. However, this provides little insight into why a protocol is correct.
An alternative approach is to develop a logic for reasoning about security protocols. When a protocol is proven correct in such a logic, the derivation steps can provide insight into the mechanisms that make the protocol work. Despite the obvious promise of such an approach, several attempts have failed, most notably the BAN logic from \[[BAN90](#bib.bibx2)\]. One of the stumbling points seems to be to provide a logic that is sound with respect to the complicated semantics of security protocol execution in the presence of an active intruder, and is able to provide concise formal proofs.
A recent attempt to develop such a logic is the Protocol Composition Logic (PCL) from e.g. \[[DDMR07](#bib.bibx14)\]. This logic has evolved from a protocol model to express protocol composition and refinement, into a model with an associated logic that can be used to formally prove security properties of protocols \[[DDMP05](#bib.bibx13), [Dat05](#bib.bibx8), [Der06](#bib.bibx16)\]. Variants of PCL have been applied to many case studies and offer several interesting features. For example, one can reason about security protocols without explicitly reasoning about the (complex) intruder by means of a special kind of invariant reasoning captured by the honesty rule. This kind of reasoning also allows the protocol logic to deal with protocol composition and refinement, where proofs can be reused. PCL has been extended with several features in further work, such as an extension for hash functions in \[[HSD<sup>+</sup>05](#bib.bibx20)\] that was used for a modular correctness proof of IEEE 802.11i and TLS.
In this paper, we identify a number of problems with PCL as defined in \[[DDMP05](#bib.bibx13), [Dat05](#bib.bibx8), [Der06](#bib.bibx16), [DDMR07](#bib.bibx14), [HSD<sup>+</sup>05](#bib.bibx20), [He05](#bib.bibx19)\]. They have implications for the scope of PCL, a number of claimed formal proofs, and several extensions to the base model. In particular, we show that in contrast with the claims in e.g. the introduction of \[[DDMR07](#bib.bibx14)\], PCL as defined in \[[DDMP05](#bib.bibx13), [Dat05](#bib.bibx8), [Der06](#bib.bibx16), [DDMR07](#bib.bibx14)\] cannot be used to prove common authentication properties of protocols that do not include signatures. We show that a number of claimed proofs in PCL cannot be correct because (a) there is no way to establish preceding actions in a thread, and (b) there is no way to express type restrictions in PCL. With respect to existing PCL extensions, we identify two problems: the Diffie-Hellman extension from \[[DDMP05](#bib.bibx13), [Dat05](#bib.bibx8), [Der06](#bib.bibx16), [DDMR07](#bib.bibx14)\] does not correctly capture the algebraic behaviour of Diffie-Hellman-like protocols, and the extension for hash functions from \[[HSD<sup>+</sup>05](#bib.bibx20), [He05](#bib.bibx19)\] is not sound. Some of these problems can be resolved by minor modifications to PCL, but other problems require further investigation. Our observations suggest that it is at least required to make changes to existing axioms, to introduce new axioms, and to add a mechanism for a type system.
The purpose of this paper is to identify some of the challenges that need to be addressed in order to make a logic like PCL work. We hope it will contribute to the improvement of PCL, and will lead to a better understanding of some of the pitfalls of designing a compact and usable formal logic for security protocols.
The scope of this paper. The presentation of this paper is inherently difficult, not least because there are a number of different papers on PCL, which vary in notation and technical details. Many ideas were already present in precursors of PCL, e.g. \[[DMP01](#bib.bibx17), [DMP03](#bib.bibx18)\], but these variants use different concepts than later versions of PCL. These early variants in \[[DMP01](#bib.bibx17), [DMP03](#bib.bibx18)\] have no notion of thread (a.k.a. process, run, or role instance), and events are bound to agents. More recent versions of PCL bind events to threads of agents, and therefore distinguish between several threads of the same agent. PCL versions of the latter type can be found in \[[DDMP03b](#bib.bibx10), [DDMP03a](#bib.bibx9), [DDMP04b](#bib.bibx12), [DDMP04a](#bib.bibx11)\]. Subsequently, \[[DDMP03b](#bib.bibx10), [DDMP03a](#bib.bibx9), [DDMP04b](#bib.bibx12), [DDMP04a](#bib.bibx11)\] have been claimed to be either subsumed, or revised and extended, by more recent works \[[DDMP05](#bib.bibx13), [Dat05](#bib.bibx8), [Der06](#bib.bibx16), [DDMR07](#bib.bibx14)\]. Hence we choose here to focus on \[[DDMP05](#bib.bibx13), [Dat05](#bib.bibx8), [Der06](#bib.bibx16), [DDMR07](#bib.bibx14)\], which contain similar descriptions of PCL. Throughout this paper we write basic PCL to refer to \[[DDMP05](#bib.bibx13), [Dat05](#bib.bibx8), [Der06](#bib.bibx16), [DDMR07](#bib.bibx14)\]. The publications on basic PCL describe the fundamental part of PCL that focusses on authentication. In general, the comments in this paper apply to basic PCL. The comments in Section [4.2](#S4.SS2 "4.2 Keyed hashes ‣ 4 Problems with PCL extensions ‣ On the Protocol Composition Logic PCLThis work was supported by the Hasler Foundation, ManCom project 2071. Manuscript") apply only to the extensions found in \[[HSD<sup>+</sup>05](#bib.bibx20), [He05](#bib.bibx19)\]. Our comments here do not cover the recent extensions to basic PCL for the analysis of secrecy, as found in \[[RDD<sup>+</sup>06](#bib.bibx23)\], nor the computational variants of PCL, as found in e.g. \[[DDMW06](#bib.bibx15)\].
Syntax and page references. In order to pinpoint our observations to specific formulas, we select a specific version of PCL to refer to. We have chosen the most recent description of PCL from 2007 as found in \[[DDMR07](#bib.bibx14)\]. In particular, we will use \[[DDMR07](#bib.bibx14)\] as a reference for the syntax of PCL formulas, and to provide specific page references. Hence we use \[[DDMR07](#bib.bibx14)\] as the reference paper to present the problems with basic PCL from the papers \[[DDMP05](#bib.bibx13), [Dat05](#bib.bibx8), [Der06](#bib.bibx16), [DDMR07](#bib.bibx14)\].
For the technical details, in particular the formulas, we assume the reader is at least somewhat familiar with one of the papers from \[[DDMP05](#bib.bibx13), [Dat05](#bib.bibx8), [Der06](#bib.bibx16), [DDMR07](#bib.bibx14)\] or \[[He05](#bib.bibx19), [HSD<sup>+</sup>05](#bib.bibx20)\]. However, the main points should be clear to readers familiar with formal security protocol analysis.
The remainder of the paper is structured in the following way. We start off by recalling some PCL notation and concepts in Section [2](#S2 "2 Preliminaries ‣ On the Protocol Composition Logic PCLThis work was supported by the Hasler Foundation, ManCom project 2071. Manuscript"). Then, in Section [3](#S3 "3 Problems with basic PCL ‣ On the Protocol Composition Logic PCLThis work was supported by the Hasler Foundation, ManCom project 2071. Manuscript") we discuss problems with the basic definition of PCL. In Section [4](#S4 "4 Problems with PCL extensions ‣ On the Protocol Composition Logic PCLThis work was supported by the Hasler Foundation, ManCom project 2071. Manuscript") we identify two problems with existing PCL extensions. We conclude in Section [5](#S5 "5 Conclusions ‣ On the Protocol Composition Logic PCLThis work was supported by the Hasler Foundation, ManCom project 2071. Manuscript").
##### Acknowledgements.
The author would like to thank David Basin, Anupam Datta, Felix Klaedtke, Sjouke Mauw, Simon Meier, John C. Mitchell, Arnab Roy and the anonymous referees for useful discussions and feedback on earlier versions of this paper.
## 2 Preliminaries
The purpose of this section is to recall some PCL notions that are required to interpret the forthcoming sections. We use the syntax from \[[DDMR07](#bib.bibx14)\]. This partial summary of PCL is incomplete, and we encourage the reader to use the original papers for reference.
The structure of PCL is as follows: first notation is introduced to define terms, which in turn are used to define protocols. For such protocols, an execution model is defined, assigning to each protocol a set of possible execution histories, called runs. Then, a protocol logic is defined in order to reason about (sets of) runs of a protocol. This logic is proven sound with respect to the execution model. This means that if one proves a property in terms of the protocol logic, such as $`\text{Receive}\hspace{0pt}{(\ldots,m)}`$, then a similar property should hold for the corresponding set of runs in the execution model, such as “$`{\text{receive~}\hspace{0pt}\ldots},m`$ has occurred in the protocol run”. We touch upon these elements below.
Terms. A term system is introduced that contains constants (nonces, names, keys, etc.), variables, and compound terms such as tuples, encryptions, and signatures.
Protocols. In PCL, a protocol $`Q`$ is defined as a set of roles. Each role $`\rho \in Q`$ is defined as a list of actions. Examples of possible actions can be found in the actions column of Table [1](#S2.T1 "Table 1 ‣ 2 Preliminaries ‣ On the Protocol Composition Logic PCLThis work was supported by the Hasler Foundation, ManCom project 2071. Manuscript"), and correspond respectively to sending terms, receiving terms, generating fresh terms, encryption, decryption, and signature verification. Each role is partitioned into a set of basic sequences. A basic sequence $`{BS}_{i}`$ of a role $`\rho`$ is a contiguous subsequence of $`\rho`$ that starts with either the first action of $`\rho`$ or a receive action, and of which the other actions are not receive actions. The basic sequences intuitively represent the idea that agents only block at receive actions, and execute all other actions immediately, allowing one to regard basic sequences as atomic actions in some respects. The notion of basic sequences therefore roughly corresponds to the notion of step-compression in other models, e.g. \[[BMV03](#bib.bibx4)\]. The basic sequences of a protocol are defined as the union of the basic sequences of each of its roles, and hence each action of a role of a protocol occurs in exactly one of its basic sequences.
Observe that although some versions of PCL have two distinct actions receive (for reading a term from the network without parsing) and match (for parsing a term and blocking if it is not as expected), these two actions are in many proofs collapsed together into a single receive action that receives and matches terms at the same time.
Execution model. An execution model is defined for protocols (in terms of cords). Actions are executed by agents, with names like $`\hat{X},\hat{Y}`$. A subset of these agents, defined as $`\text{HONEST}\hspace{0pt}{(C)}`$ (where $`C`$ is the initial configuration of the system), are called the honest agents, and execute their protocol roles as expected. The agents may execute each protocol role any number of times; each such instance of a role is called a thread (in other formalisms, this notion is known as a strand, or a process, or a run). Threads are usually denoted by symbols such as $`X,Y,Z`$: a sequence of actions $`P`$ executed within a thread $`X`$ is written as $`{\lbrack P\rbrack}_{X}`$. The notation $`\hat{X}`$ is often
<sup>1</sup>
<sup>1</sup>1The hat notation $`\hat{}`$ is used in at least two different interpretations in both \[[DDMP05](#bib.bibx13)\] and \[[DDMR07](#bib.bibx14)\]. In some cases $`X`$ is used to denote a particular thread, and $`\hat{X}`$ is then interpreted as the agent that executes that thread, as e.g. can be seen in the usage of $`Y`$ in the SEC axiom \[[DDMR07](#bib.bibx14), page 327\], replicated here on page [3.1](#S3.SS1.SSS0.Px1 "Other comments regarding non-signature protocols. ‣ 3.1 Authentication properties only provable for signature protocols ‣ 3 Problems with basic PCL ‣ On the Protocol Composition Logic PCLThis work was supported by the Hasler Foundation, ManCom project 2071. Manuscript"). Thus, in one interpretation $`\hat{}`$ can be regarded as a function from a thread to an agent. However, in other cases $`\hat{X}`$ is used to denote a particular agent, and $`X`$ is then interpreted as “any thread executed by the agent $`\hat{X}`$”, as in the honesty rule HON \[[DDMR07](#bib.bibx14), page 329\], and the VER axiom \[[DDMR07](#bib.bibx14), pp. 329–330\], replicated here on page [3.1](#S3.SS1 "3.1 Authentication properties only provable for signature protocols ‣ 3 Problems with basic PCL ‣ On the Protocol Composition Logic PCLThis work was supported by the Hasler Foundation, ManCom project 2071. Manuscript"). In this second interpretation, $`X`$ is a variable that ranges over all threads executed by the agent $`\hat{X}`$, and $`\hat{X}`$ effectively expresses a domain restriction on $`X`$.
used to denote the agent executing the thread $`X`$. Informally, if the agent $`\hat{X}`$ is honest, $`{\lbrack P\rbrack}_{X}`$ is a sequence of actions from a protocol role. The agents that are not part of $`\text{HONEST}\hspace{0pt}{(C)}`$ can execute so-called intruder roles, in line with the common Dolev-Yao intruder model. In the context of this execution model, the protocol description $`Q`$ gives rise to a set of runs denoted by $`\text{Runs}\hspace{0pt}{(Q)}`$. A run is typically denoted by $`R`$, and corresponds to a sequence of actions as executed by threads. A run represents a possible execution history of the system.
<table>
<thead>
<tr>
<th>Action</th>
<th>Associated term</th>
<th>Predicate</th>
</tr>
<tr>
<th></th>
<th>structure</th>
<th>(in thread X𝑋X)</th>
</tr>
</thead>
<tbody>
<tr>
<td>send X^,Y^,msend ^𝑋^𝑌𝑚\mbox{{\tt send\;}}\hat{X},\hat{Y},m</td>
<td></td>
<td>Send(X,m)Send𝑋𝑚\mbox{{\sf Send}}(X,m)</td>
</tr>
<tr>
<td>receive X^,Y^,mreceive ^𝑋^𝑌𝑚\mbox{{\tt receive\;}}\hat{X},\hat{Y},m</td>
<td></td>
<td>Receive(X,m)Receive𝑋𝑚\mbox{{\sf Receive}}(X,m)</td>
</tr>
<tr>
<td>new xnew 𝑥\mbox{{\tt new\;}}x</td>
<td></td>
<td>Gen(X,x)Gen𝑋𝑥\mbox{{\sf Gen}}(X,x)</td>
</tr>
<tr>
<td>enc m,Kenc 𝑚𝐾\mbox{{\tt enc\;}}m,K</td>
<td>enct=ENCK{|m|}𝑒𝑛𝑐𝑡subscriptENC𝐾𝑚enct=\mbox{{\it ENC}}_{K}\{\mkern-4.0mu\lvert\,{m}\,\rvert\mkern-4.0mu\}</td>
<td>Encrypt(X,enct)Encrypt𝑋𝑒𝑛𝑐𝑡\mbox{{\sf Encrypt}}(X,enct)</td>
</tr>
<tr>
<td>dec enct,Kdec 𝑒𝑛𝑐𝑡𝐾\mbox{{\tt dec\;}}enct,K</td>
<td>enct=ENCK{|m|}𝑒𝑛𝑐𝑡subscriptENC𝐾𝑚enct=\mbox{{\it ENC}}_{K}\{\mkern-4.0mu\lvert\,{m}\,\rvert\mkern-4.0mu\}</td>
<td>Decrypt(X,enct)Decrypt𝑋𝑒𝑛𝑐𝑡\mbox{{\sf Decrypt}}(X,enct)</td>
</tr>
<tr>
<td>verify sigt,m,Kverify 𝑠𝑖𝑔𝑡𝑚𝐾\mbox{{\tt verify\;}}sigt,m,K</td>
<td>sigt=SIGK{|m|}𝑠𝑖𝑔𝑡subscriptSIG𝐾𝑚sigt=\mbox{{\it SIG}}_{K}\{\mkern-4.0mu\lvert\,{m}\,\rvert\mkern-4.0mu\}</td>
<td>Verify(X,sigt)Verify𝑋𝑠𝑖𝑔𝑡\mbox{{\sf Verify}}(X,sigt)</td>
</tr>
<tr>
<td></td>
<td></td>
<td>Has(X,m)Has𝑋𝑚\mbox{{\sf Has}}(X,m)</td>
</tr>
<tr>
<td></td>
<td></td>
<td>Honest(X^)Honest^𝑋\mbox{{\sf Honest}}(\hat{X})</td>
</tr>
<tr>
<td></td>
<td></td>
<td>Contains(bigt,smallt)Contains𝑏𝑖𝑔𝑡𝑠𝑚𝑎𝑙𝑙𝑡\mbox{{\sf Contains}}(bigt,smallt)</td>
</tr>
</tbody>
</table>
Table 1: Some examples of PCL actions, action predicates, terms, and their relations. Here $`\hat{X},\hat{Y}`$ denote agents, $`m,x,{e\hspace{0pt}n\hspace{0pt}c\hspace{0pt}t},{s\hspace{0pt}i\hspace{0pt}g\hspace{0pt}t},{b\hspace{0pt}i\hspace{0pt}g\hspace{0pt}t},{s\hspace{0pt}m\hspace{0pt}a\hspace{0pt}l\hspace{0pt}l\hspace{0pt}t}`$ denote terms, and $`X`$ denotes a thread.
Protocol logic. For each of the actions that can occur in a run, a corresponding action predicate is defined. Some examples are given in the right column of Table [1](#S2.T1 "Table 1 ‣ 2 Preliminaries ‣ On the Protocol Composition Logic PCLThis work was supported by the Hasler Foundation, ManCom project 2071. Manuscript"). The protocol logic is extended with logical connectives, a predicate to reason about the ordering of actions in a run ($`<`$), as well as predicates to reason about the knowledge of threads.
One of the predicates of the logic is the Honest predicate, which is closely related to the set $`\text{HONEST}\hspace{0pt}{(C)}`$. For a run $`R`$, $`\text{Honest}\hspace{0pt}{(\hat{X})}`$ is used to denote that $`\hat{X} \in {\text{HONEST}\hspace{0pt}{(C)}}`$ and “all threads of $`\hat{X}`$ are in a ‘pausing’ state in $`R`$” \[[DDMR07](#bib.bibx14), page 325\]. This is a technicality
<sup>2</sup>
<sup>2</sup>2The Honest predicate serves within the honesty rule as an encoding of the atomic nature of basic sequences.
that is needed to make the honesty rule (which is not described here) work.
One predicate of interest for this paper is the Contains predicate, which is used to reason about the relation between two terms. In particular, $`\text{Contains}\hspace{0pt}{(t_{1},t_{2})}`$ is defined by means of the subterm relation $`\subseteq`$ in \[[DDMR07](#bib.bibx14), page 323\], stating that $`t_{1}`$ contains $`t_{2}`$ as a subterm. The subterm relation $`\subseteq`$ is never formally defined. Here we assume that the subterm relation is defined syntactically
<sup>3</sup>
<sup>3</sup>3Observe that if we would alternatively assume that the subterm relation involves only tuple projection, as seems to be suggested by the CON axiom \[[DDMP05](#bib.bibx13), page 444\], i.e. $`t \subseteq t' \equiv {(t = t')} \vee {(\exists t^{\operatorname{\prime\prime}}.t' = {(t,t^{\operatorname{\prime\prime}})} \vee t' = {(t^{\operatorname{\prime\prime}},t)})}`$, then e.g. the P2 and VER axioms in \[[DDMP05](#bib.bibx13), [DDMR07](#bib.bibx14)\] are unsound, because respectively the fresh value and the signature might be sent as part of an encrypted term, and decrypted by an agent that knows the key.
. In particular, we assume that a signature such as $`\text{SIG}_{\hat{X}}\hspace{0pt}{\{{|m|}\}}`$ contains $`m`$, i.e. $`\text{Contains}\hspace{0pt}{({\text{SIG}_{\hat{X}}\hspace{0pt}{\{{|m|}\}}},m)}`$ holds. Note that this assumption only plays a role in the construction of a particular counterexample in Section [3.3](#S3.SS3 "3.3 No formal type system ‣ 3 Problems with basic PCL ‣ On the Protocol Composition Logic PCLThis work was supported by the Hasler Foundation, ManCom project 2071. Manuscript"), and does not influence any of our general observations.
Using the protocol logic, one aims to establish properties of all runs of a protocol. A run $`R`$ of a protocol $`Q`$ that satisfies a property $`\phi`$ is denoted as $`{Q,R} \models \phi`$. If all runs of a protocol $`Q`$ satisfy $`\phi`$, we write $`Q \models \phi`$. If a formula $`\phi`$ is provable using the logic by using any PCL axioms except the honesty rule, we write $`\vdash \phi`$, which expresses that $`\phi`$ holds for all protocols. For formulas provable using the axioms and the honesty rule for a protocol $`Q`$, we write $`\vdash_{Q}\phi`$, which expresses that $`\phi`$ holds for the protocol $`Q`$.
In the remainder of this paper we use the following convention: All formulas that are numbered are ours. All unnumbered formulas, including the named axioms, are copied from PCL papers, in which case the source paper and page number is given.
## 3 Problems with basic PCL
In this section, we address two problems with basic PCL as defined in \[[DDMP05](#bib.bibx13), [Dat05](#bib.bibx8), [Der06](#bib.bibx16), [DDMR07](#bib.bibx14)\]. First, we identify in Section [3.1](#S3.SS1 "3.1 Authentication properties only provable for signature protocols ‣ 3 Problems with basic PCL ‣ On the Protocol Composition Logic PCLThis work was supported by the Hasler Foundation, ManCom project 2071. Manuscript") a strong restriction on the scope of basic PCL. Then, we identify two missing proof mechanisms that seem to be necessary to prove simple protocols correct in Sections [3.2](#S3.SS2 "3.2 No means to establish preceding actions in a thread ‣ 3 Problems with basic PCL ‣ On the Protocol Composition Logic PCLThis work was supported by the Hasler Foundation, ManCom project 2071. Manuscript") and [3.3](#S3.SS3 "3.3 No formal type system ‣ 3 Problems with basic PCL ‣ On the Protocol Composition Logic PCLThis work was supported by the Hasler Foundation, ManCom project 2071. Manuscript").
### 3.1 Authentication properties only provable for signature protocols
In basic PCL \[[DDMP05](#bib.bibx13), [Dat05](#bib.bibx8), [Der06](#bib.bibx16), [DDMR07](#bib.bibx14)\], the possibility of proving authentication properties is limited to protocols that use signatures.
Proving an authentication property such as aliveness or agreement requires a proof of existence of actions of another agent. In basic PCL, there is only one axiom that allows for a conclusion of that type. The precondition of this axiom can only be met by protocols that use signatures. As a consequence, if a protocol does not use signatures, the existence of a communication partner cannot be proven within the logic.
This is surprising, as PCL was “initially designed as a logic of authentication” \[[DDMR07](#bib.bibx14), page 313\], and it is stated in the abstract of the paper that “PCL \[…\] has been applied to a number of industry standards including SSL/TLS, IEEE 802.11i \[…\]” \[[DDMR07](#bib.bibx14), page 311, abstract\]. These protocols consist of many subprotocols that do not rely only on signatures, but also rely on symmetric key cryptography and symmetric MACs.
The problem occurs for all authentication properties that imply the existence of a communication partner. This includes aliveness or any form of agreement from \[[Low97b](#bib.bibx22)\], matching conversations from \[[BR94](#bib.bibx5)\], or synchronisation from \[[CMdV06](#bib.bibx6)\]. The matching conversations property is the authentication property used within basic PCL \[[DDMR07](#bib.bibx14), page 331\]. All these properties are of the form:
<table>
<tbody>
<tr>
<td></td>
<td>ϕ(X)⊃∃Y.ψ(Y)formulae-sequence𝑌italic-ϕ𝑋𝜓𝑌\phi(X)\;\supset\;\exists Y.\psi(Y)</td>
<td></td>
<td>(1)</td>
</tr>
</tbody>
</table>
where typically, $`\phi\hspace{0pt}{(X)}`$ denotes that thread $`X`$ executes a role that authenticates another role. Such a property states that if a thread $`X`$ executes actions of a certain role (e.g. all actions of an initiator role), then there exists a thread $`Y`$ that has executed some actions of another role (e.g. the first two actions of a responder role), and possibly some further condition holds. This is captured in $`\psi\hspace{0pt}{(Y)}`$. Note that both the weak and strong authentication properties in the example of \[[DDMR07](#bib.bibx14), page 331\] belong to this class of authentication properties.
In order to prove such a property, it is required to prove the existence of a thread. Examination of all axioms of the logic reveals that only one axiom allows for the conclusion of the existence of a thread, which is the VER axiom \[[DDMR07](#bib.bibx14), page 327\]:
<table>
<tbody>
<tr>
<td></td>
<td>VERHonest(X^)∧Verify(Y,SIGX^{|x|})∧X^≠Y^⊃∃X.Send(X,m)∧Contains(m,SIGY^{|x|})formulae-sequenceVERHonest^𝑋Verify𝑌subscriptSIG^𝑋𝑥^𝑋^𝑌superset-of𝑋Send𝑋𝑚Contains𝑚subscriptSIG^𝑌𝑥\mbox{{\bf VER}}\qquad\mbox{{\sf Honest}}(\hat{X})\land\mbox{{\sf Verify}}(Y,\mbox{{\it SIG}}_{\hat{X}}\{\mkern-4.0mu\lvert\,{x}\,\rvert\mkern-4.0mu\})\land\hat{X}\not=\hat{Y}\supset\\ \exists X.\mbox{{\sf Send}}(X,m)\land\mbox{{\sf Contains}}(m,\mbox{{\it SIG}}_{\hat{Y}}\{\mkern-4.0mu\lvert\,{x}\,\rvert\mkern-4.0mu\})</td>
<td></td>
</tr>
</tbody>
</table>
Because this axiom has the signature verification predicate Verify in the precondition, it can only be used for signature protocols. Hence there is no way to prove the existence of another thread for protocols without signatures.
##### Other comments regarding non-signature protocols.
Whilst introducing new axioms for establishing thread existence for non-signature protocols is non-trivial, there is a related problem with the inconsistent use of symmetric/asymmetric cryptography. In basic PCL, there is only one type of encryption operator, and only a single key set, in the sense that the rules for encryption and decryption do not distinguish between different types of keys. This suggests that either only symmetric, or only asymmetric encryption is supported.
The definitions of the reaction rules \[[DDMR07](#bib.bibx14), page 318\], the action formulas \[[DDMR07](#bib.bibx14), page 324\] and the Has predicate \[[DDMR07](#bib.bibx14), pp.324–325\] all indicate that one encrypts and decrypts a message using the same key, e.g.:
<table>
<tbody>
<tr>
<td></td>
<td>Hasi+1(A,ENCK{|m|})if Hasi(A,m) and Hasi(A,K)Hasi+1(A,m)if Hasi(A,ENCK{|m|}) and Hasi(A,K)subscriptHas𝑖1𝐴subscriptENC𝐾𝑚if subscriptHas𝑖𝐴𝑚subscript and Has𝑖𝐴𝐾subscriptHas𝑖1𝐴𝑚if subscriptHas𝑖𝐴subscriptENC𝐾𝑚subscript and Has𝑖𝐴𝐾\begin{array}[]{ll}\mbox{{\sf Has}}_{i+1}(A,\mbox{{\it ENC}}_{K}\{\mkern-4.0mu\lvert\,{m}\,\rvert\mkern-4.0mu\})&\mbox{if }\;\mbox{{\sf Has}}_{i}(A,m)\mbox{ and }\mbox{{\sf Has}}_{i}(A,K)\\ \mbox{{\sf Has}}_{i+1}(A,m)&\mbox{if }\;\mbox{{\sf Has}}_{i}(A,\mbox{{\it ENC}}_{K}\{\mkern-4.0mu\lvert\,{m}\,\rvert\mkern-4.0mu\})\mbox{ and }\mbox{{\sf Has}}_{i}(A,K)\end{array}</td>
<td></td>
</tr>
</tbody>
</table>
The same assumption, that one encrypts and decrypts with the same key, can be found in axiom AR3 \[[DDMR07](#bib.bibx14), page 327\]. Without explaining the full details of the notation of this axiom, we want to point out that the key $`K`$ is used as the key to decrypt a message encrypted with $`K`$:
<table>
<tbody>
<tr>
<td></td>
<td>AR3a(x)[y:=dec x,K]Xa(ENCK{|y|})AR3a𝑥subscriptdelimited-[]assign𝑦dec 𝑥𝐾𝑋asubscriptENC𝐾𝑦\begin{array}[]{ll}\mbox{{\bf AR3}}&\mbox{{\sf a}}(x)\;[y:=\mbox{{\tt dec\;}}x,K]_{X}\;\mbox{{\sf a}}(\mbox{{\it ENC}}_{K}\{\mkern-4.0mu\lvert\,{y}\,\rvert\mkern-4.0mu\})\end{array}</td>
<td></td>
</tr>
</tbody>
</table>
However, the idea that symmetric encryption is intended seems to be contradicted by the SEC axiom \[[DDMR07](#bib.bibx14), page 327\], which states that there is only one agent that can decrypt a message encrypted with a key, along the lines of asymmetric encryption:
<table>
<tbody>
<tr>
<td></td>
<td>SECHonest(X^)∧Decrypt(Y,ENCX^{|x|})⊃(Y^=X^)^𝑌^𝑋SECHonest^𝑋Decrypt𝑌subscriptENC^𝑋𝑥\;\;\;\mbox{{\bf SEC}}\qquad\mbox{{\sf Honest}}(\hat{X})\land\mbox{{\sf Decrypt}}(Y,\mbox{{\it ENC}}_{\hat{X}}\{\mkern-4.0mu\lvert\,{x}\,\rvert\mkern-4.0mu\})\supset(\hat{Y}=\hat{X})</td>
<td></td>
</tr>
</tbody>
</table>
Combined with the definition of the Has predicate, one is lead to conclude that the only one who can create the encrypted message that occurs in the SEC axiom, is the agent that can decrypt it.
#### 3.1.1 Implications
In order to prove any authentication property of the form of ([1](#S3.E1 "In 3.1 Authentication properties only provable for signature protocols ‣ 3 Problems with basic PCL ‣ On the Protocol Composition Logic PCLThis work was supported by the Hasler Foundation, ManCom project 2071. Manuscript")) for a protocol that does not use signatures, one needs to introduce consistent machinery for symmetric and asymmetric cryptography and several new axioms.
In basic PCL, protocols like Needham-Schroeder-Lowe
<sup>4</sup>
<sup>4</sup>4Observe that in a precursor of PCL in \[[DMP03](#bib.bibx18)\] there is a proof of Needham-Schroeder-Lowe, but this logic has only asymmetric cryptography, and has no notion of threads or processes.
and many key establishment protocols cannot be modeled, and even if they could be, no authentication proofs could be given for them. Similarly, it is impossible to use basic PCL as defined in \[[DDMP05](#bib.bibx13), [Dat05](#bib.bibx8), [Der06](#bib.bibx16), [DDMR07](#bib.bibx14)\] to prove the authentication properties of the protocols “SSL/TLS, IEEE 802.11i” \[[DDMR07](#bib.bibx14), page 311, abstract\]. In order to prove authentication of these protocols, one is required to introduce additional axioms.
#### 3.1.2 Resolving the problem
The problem can be split into three subparts. First, symmetric and asymmetric cryptography must be dealt with consistently. Second, basic PCL must be extended with axioms that enable proving authentication properties (existence of a thread) for symmetric cryptography. Third, resolving the problem for public-key encryption protocols (which includes many key agreement protocols, and the well-known Needham-Schroeder-Lowe protocol) requires the introduction of additional theory. We first address the two easier problems before turning to public-key encryption.
##### Dealing consistently with symmetric and asymmetric cryptography.
The first requirement for resolving this problem is to split symmetric and asymmetric encryption either by having two types of encrypt/decrypt actions as in e.g. \[[ABB<sup>+</sup>05](#bib.bibx1)\], or by splitting the key set as in e.g. \[[Cre06](#bib.bibx7)\]. Either choice impacts the action sets, the action formulas, the Has predicate, and requires the addition of further ARx axioms and an alternative for the SEC axiom. Most of the additions are trivial, but introduce additional complexity into the logic and possibly also the execution model.
##### Proving authentication for protocols using symmetric encryption or keyed hashes.
The axiom for signature protocols deals with the simplest possible case of authentication by cryptography: the signature of an honest agent can be used immediately to conclude that there exists a thread executed by this agent. If that agent differs from any agent executing the currently known threads, this implies the existence of another thread. This conclusion is based on the fact that only one agent has the private key needed to perform the signature.
For symmetric encryption and keyed hashes there is usually a notion of symmetry: in most cases, two agents share a symmetric key. Thus, if a symmetric key $`K`$ shared by two honest agents $`\hat{X}`$ and $`\hat{Y}`$, occurs in a message of a thread $`X`$, there are two candidates for the agent who created the message. If we can exclude that the message was generated in thread $`X`$ executed by $`\hat{X}`$, we can derive the existence of another thread $`Y`$ that must have created it.
The authors of PCL have explored a variant of this route for e.g. the extensions needed for the keyed-hash based protocols in \[[HSD<sup>+</sup>05](#bib.bibx20)\]. We discuss the merits of those extensions in detail in Section [4.2](#S4.SS2 "4.2 Keyed hashes ‣ 4 Problems with PCL extensions ‣ On the Protocol Composition Logic PCLThis work was supported by the Hasler Foundation, ManCom project 2071. Manuscript").
##### Proving authentication for public-key encryption protocols.
If we assume that basic PCL is extended to consistently deal with public-key encryption, the authentication properties still cannot be proven, as we lack an axiom for establishing thread existence as pointed out previously. Contrary to signatures or symmetric-key cryptography, there is no easy solution with an axiom based on honest agents only, because the public keys are also known to the intruder. If a message encrypted with the public key of an honest agent $`\hat{X}`$ occurs, no conclusion can be drawn about the sender or creator of the message. Thus we must also consider the possibility that the message was sent by the adversary.
A first step towards a solution would be to introduce an axiom in the protocol logic, along the lines of Lemma 2.3 \[[DDMR07](#bib.bibx14), page 321\] of the execution model. In fact, such an axiom was present in a precursor of basic PCL. In \[[DMP03](#bib.bibx18), page 701\] one can find axiom AR1 which, when recast in the basic PCL model, would read
<table>
<tbody>
<tr>
<td></td>
<td>AR1[receive t]X∃Z.Send(Z,t)(This axiom is not in basic PCL)formulae-sequenceAR1subscriptdelimited-[]receive 𝑡𝑋𝑍Send𝑍𝑡(This axiom is not in basic PCL)\mbox{{\bf AR1}}\quad[\mbox{{\tt receive\;}}t]_{X}\exists Z.\mbox{{\sf Send}}(Z,t)\qquad\mbox{(This axiom is not in basic PCL)}</td>
<td></td>
<td>(2)</td>
</tr>
</tbody>
</table>
A second axiom that might be adapted for proving these properties is the SRC axiom in \[[DMP03](#bib.bibx18), page 703\].
Unfortunately, for either of these axioms, authentication proofs would require either further machinery to prove that $`Z`$ is executed by an honest agent, or explicit reasoning about the intruder, as the sender thread $`Z`$ need not be executed by an honest agent. This type of reasoning is not supported by basic PCL, and would require a significant amount of additional machinery.
### 3.2 No means to establish preceding actions in a thread
By the definition of the PCL execution model, threads of honest agents start at the beginning of a role description, and always execute the actions in the order given by the role. Thus, if one can establish that a thread of an honest agent has executed the $`n`$th action of a role, it should also be possible to conclude a similar result about preceding actions within the logic: in particular, one should be able to conclude that the preceding $`n - 1`$ actions have also occurred in the same thread. However, there is no mechanism in the logic of basic PCL \[[DDMP05](#bib.bibx13), [Dat05](#bib.bibx8), [Der06](#bib.bibx16), [DDMR07](#bib.bibx14)\] to draw such a conclusion.
Note that one can use the honesty rule to prove from, e.g., a send action that a receive action must have occurred previously, but only if both actions occur within the same basic sequence. However, if one wants to prove that given a basic sequence, an action has occurred from another basic sequence, there are no rules to enable this type of reasoning.
#### 3.2.1 Implications
The consequence of not having a means of establishing preceding actions is that some claimed proofs do not seem to be correct. For example, observe the initiator role of the $`Q_{CR}`$ protocol \[[DDMR07](#bib.bibx14), page 315\]. In order to prove invariants for the protocol using the honesty rule, one has to prove that the invariants hold for all basic sequences of the protocol. The initiator role consists of two basic sequences $`{BS}_{1}`$ and $`{BS}_{2}`$ \[[DDMR07](#bib.bibx14), page 335\]:
<table>
<tbody>
<tr>
<td></td>
<td>BS1≡[\displaystyle\mathrm{BS}_{1}\equiv[\;</td>
<td>new m;send X^,Y^,m;]X\displaystyle\mbox{{\tt new\;}}m;\;\mbox{{\tt send\;}}\hat{X},\hat{Y},m;\;]_{X}</td>
<td></td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td>BS2≡[\displaystyle\mathrm{BS}_{2}\equiv[\;</td>
<td>receive Y^,X^,y,s;verify s,(y,m,X^),Y^;r:=sign (y,m,Y^),X^;send X^,Y^,r;]X\displaystyle\mbox{{\tt receive\;}}\hat{Y},\hat{X},y,s;\;\mbox{{\tt verify\;}}s,(y,m,\hat{X}),\hat{Y};\;r:=\mbox{{\tt sign\;}}(y,m,\hat{Y}),\hat{X};\;\mbox{{\tt send\;}}\hat{X},\hat{Y},r;\;]_{X}</td>
<td></td>
</tr>
</tbody>
</table>
In this protocol, the basic sequence $`{BS}_{1}`$ defines $`m`$ as a generated value, which determines the semantics of $`m`$ in $`{BS}_{2}`$. Observe that there is no way to tell if $`m`$ was received or generated from just investigating $`{BS}_{2}`$.
In order to show how this makes certain proofs impossible in basic PCL, consider a protocol $`Q'`$, that contains a role $`\rho`$ consisting of the basic sequences $`{BS}_{1}'`$ followed by $`{BS}_{2}`$, where $`{BS}_{1}'`$ is defined as:
<table>
<tbody>
<tr>
<td></td>
<td>BS1′≡[receive ENCK{|m|};send X^,Y^,m;]X\mathrm{BS}_{1}^{\prime}\equiv[\mbox{{\tt receive\;}}\mbox{{\it ENC}}_{K}\{\mkern-4.0mu\lvert\,{m}\,\rvert\mkern-4.0mu\};\mbox{{\tt send\;}}\hat{X},\hat{Y},m;]_{X}</td>
<td></td>
<td>(3)</td>
</tr>
</tbody>
</table>
Thus the protocol $`Q'`$ shares the basic sequence $`{BS}_{2}`$ with $`Q_{CR}`$, but in $`Q'`$, $`m`$ is a received valued as opposed to a generated value.
Now, in order to prove invariant $`\gamma_{1}`$ \[[DDMR07](#bib.bibx14), page 334\] for basic sequence $`{BS}_{2}`$, we must prove that $`m`$ is either generated, or received earlier as the main message, by the same thread that sends out the signature:
<table>
<tbody>
<tr>
<td></td>
<td>γ1≡(Send(Y,t)∧Contains(t,SIGY^{|y,m,X^|}))⊃(Gen(Y,m)∨(Receive(Y,(X^,Y^,m))<Send(Y,(Y^,X^,y,SIGY^{|y,m,X^|}))))subscript𝛾1Send𝑌𝑡Contains𝑡subscriptSIG^𝑌𝑦𝑚^𝑋superset-ofGen𝑌𝑚Receive𝑌^𝑋^𝑌𝑚Send𝑌^𝑌^𝑋𝑦subscriptSIG^𝑌𝑦𝑚^𝑋\gamma_{1}\equiv\bigl{(}\mbox{{\sf Send}}(Y,t)\land\mbox{{\sf Contains}}(t,\mbox{{\it SIG}}_{\hat{Y}}\{\mkern-4.0mu\lvert\,{y,m,\hat{X}}\,\rvert\mkern-4.0mu\})\bigr{)}\supset\\ \Bigl{(}\mbox{{\sf Gen}}(Y,m)\lor\bigl{(}\mbox{{\sf Receive}}(Y,(\hat{X},\hat{Y},m))<\mbox{{\sf Send}}(Y,(\hat{Y},\hat{X},y,\mbox{{\it SIG}}_{\hat{Y}}\{\mkern-4.0mu\lvert\,{y,m,\hat{X}}\,\rvert\mkern-4.0mu\}))\bigr{)}\Bigr{)}</td>
<td></td>
</tr>
</tbody>
</table>
In order to prove this invariant with respect to basic sequence $`{BS}_{2}`$, one needs to prove that $`m`$ is generated in the thread $`Y`$. However, there is no mechanism to reason about any preceding actions, and thus it has to be dealt with in the same way for both protocols $`Q_{CR}`$ and $`Q'`$. One must therefore also consider the case $`m`$ was first received, and not generated, in a previous basic sequence as part of a bigger term, as happens in $`Q'`$. In fact, the invariant is false for protocol $`Q'`$.
Thus, in order to prove the invariant for $`Q_{CR}`$, we must be able to distinguish between the different semantics of $`{BS}_{2}`$ within the context of $`Q_{CR}`$ and $`Q'`$. Such reasoning is not supported by basic PCL.
Without the ability to reason about preceding actions, the protocol descriptions are effectively cut up into independent basic sequences, which can then only be dealt with as separate protocols. This is evident from the formulation of the honesty rule \[[DDMR07](#bib.bibx14), page 329\], in which the reference to the protocol $`Q`$ is only used for the definition of the basic sequences. Thus one is required to create much stronger proofs than would be strictly necessary. Put differently, in basic PCL one has to reason with an over approximation of protocol execution, in which basic sequences occur in no fixed order in a thread.
This phenomenon can also be observed in the following example. Let $`P`$ be a protocol, and let $`P_{1}`$ be the protocol defined as $`P`$ extended with a role consisting of the basic sequences $`{BS};{BS}';{BS}^{\operatorname{\prime\prime}}`$. Similarly, let $`P_{2}`$ be defined as $`P`$ extended with a role $`{BS}^{\operatorname{\prime\prime}};{BS}';{BS}`$, i.e., a role with the same basic sequences as the role in $`P_{1}`$, but composed in a different order. Then, any invariant $`\gamma`$ that can be proven for protocol $`P_{1}`$ using the honesty rule, must also hold for $`P_{2}`$. In fact, the proof is identical. Conversely, any invariant that holds for $`P_{1}`$ but not for $`P_{2}`$, cannot be proven using the honesty rule. This puts a strong restriction on the type of invariants provable in basic PCL, because the structure (and hence the properties) of both protocols can be very different.
Ultimately, this problem seems to be a side effect of the weak link between the protocol description $`Q`$ and the actions of honest agents in a run $`R \in {\text{Runs}\hspace{0pt}{(Q)}}`$ in the protocol logic. In basic PCL, the only way to make the link between a protocol description and the actions of honest agents, is by means of defining an invariant and then proving it with the honesty rule. This can be seen from inspecting the protocol logic rules: the only occurrences of a protocol name (e.g. $`Q`$) are in the honesty rule and the composition rules. As the honesty rule only reasons about isolated basic sequences of the protocol, the relations among the basic sequences of the protocol are inevitably lost.
#### 3.2.2 Resolving the problem
Based on the semantics of PCL, it should be possible to extend the logic with a “precedence” inference rule, that would allow one to infer that given the $`n`$th action of a role $`\rho`$ occurs in a thread $`X`$, also the $`({n - 1})`$th action of the same role must have previously occurred in the same thread. Such an inference rule would allow for establishing preceding actions that must have occurred from the existence of other actions. This is particularly useful for proving invariants for basic sequences.
In general one could extend basic PCL with additional mechanisms to enable reasoning about the relation between protocol descriptions and the actions of honest agents in a run. However, the current weak link between the two has one major advantage: it eases compositional proofs. For example, because no constraints are put on any other (i.e. non-protocol) actions of honest agents, the sequential composition of protocols allows for re-use of proofs of invariants for basic sequences. Thus, one must be careful when introducing such links and introduce only links between the protocol and the actions of honest agent, such that the links are invariant under e.g. sequential composition. A “precedence” inference rule of the type sketched above should meet this condition.
### 3.3 No formal type system
It is mentioned that in PCL, “We assume enough typing to distinguish the keys $`K`$ from the principals $`\hat{A}`$, the nonces $`n`$ and so on.” \[[DDMR07](#bib.bibx14), page 316\]. However, there are no constructs in PCL that allow for formal reasoning about the type of variables.
As a consequence, some claimed proofs of invariants are not correct within the logic, as they are only true under type assumptions, that cannot be expressed in PCL.
#### 3.3.1 Implications
Many protocols have properties in a typed model, that do not hold for an untyped model. In particular, some protocols are correct in a typed model, but are vulnerable to so-called type flaw attacks in an untyped model. Such attacks exploit for example that an agent could mistake an encrypted session key for an encrypted nonce, and sends the key unencrypted in a next step, revealing it to the intruder. It is therefore often easier to prove properties for the typed model, but this requires that the logic supports drawing conclusions based on types. Currently, this is not possible in PCL.
As an example, we show that invariant $`\gamma_{1}`$ from \[[DDMR07](#bib.bibx14), page 334\] (reproduced in this paper on page [3.2.1](#S3.SS2.SSS1 "3.2.1 Implications ‣ 3.2 No means to establish preceding actions in a thread ‣ 3 Problems with basic PCL ‣ On the Protocol Composition Logic PCLThis work was supported by the Hasler Foundation, ManCom project 2071. Manuscript")) is false when variables are untyped. The invariant $`\gamma_{1}`$ states that if a thread (of an honest agent) sends a message that contains a signature term $`S`$, with a subterm $`m`$, then either
1. 1\.
$`m`$ was generated in the thread $`Y`$, or
2. 2\.
the thread $`Y`$ executes a receive of $`m`$, and later a send of the message tuple $`(y,S)`$.
Now consider the basic sequence $`{BS}_{3}`$ from \[[DDMR07](#bib.bibx14), page 335\]:
<table>
<tbody>
<tr>
<td></td>
<td>BS3subscriptBS3\displaystyle\mathrm{BS}_{3}</td>
<td>≡[receive X^,Y^,x;new n;r:=sign (n,x,X^),Y^;send Y^,X^,n,r;]Y\displaystyle\equiv[\;\mbox{{\tt receive\;}}\hat{X},\hat{Y},x;\;\mbox{{\tt new\;}}n;\;r:=\mbox{{\tt sign\;}}(n,x,\hat{X}),\hat{Y};\;\;\,\mbox{{\tt send\;}}\hat{Y},\hat{X},n,r;\;]_{Y}</td>
<td></td>
</tr>
</tbody>
</table>
This basic sequence corresponds to the responder receiving an unknown value, supposedly a nonce, and sending back a signed version that includes a freshly generated nonce.
In order to show the invariant is false, consider a thread $`Y'`$ that is executed by agent $`\hat{Y}`$. If we assume that $`x = {\text{SIG}_{\hat{Y}}\hspace{0pt}{\{{|y,m,\hat{X}|}\}}}`$, where $`m`$ is generated by thread $`Y'`$, the invariant is violated: we have that $`x = {\text{SIG}_{\hat{Y}}\hspace{0pt}{\{{|y,m,\hat{X}|}\}}} \supset {\text{Contains}\hspace{0pt}{(x,m)}}`$ and by substitution in $`{BS}_{3}`$ and the Contains predicate, we find that $`\text{Contains}\hspace{0pt}{(r,m)}`$. As a result, the message sent at the end of $`{BS}_{3}`$ of a thread $`Y`$ ($`Y' \neq Y`$) will contain $`m`$. However, $`m`$ is neither generated by this thread nor is it the exact term that was received. Thus the invariant $`\gamma_{1}`$ does not hold for basic sequence $`{BS}_{3}`$. Hence the example authentication proof of the $`CR`$ protocol in \[[DDMR07](#bib.bibx14)\] is incorrect.
#### 3.3.2 Resolving the problem
The model can be extended with typing information for the variables, and an axiom could be introduced that captures the fact that variables are only instantiated by their typing information. This would not introduce much new machinery, but requires additional reasoning steps in most of the proofs.
## 4 Problems with PCL extensions
In this Section we discuss two additional mechanisms for PCL, in particular the extension for Diffie-Hellman exponentials as found in \[[DDMP05](#bib.bibx13), [Dat05](#bib.bibx8), [Der06](#bib.bibx16), [DDMR07](#bib.bibx14)\] and the extension for hash functions from \[[He05](#bib.bibx19), [HSD<sup>+</sup>05](#bib.bibx20)\].
### 4.1 Diffie-Hellman exponentials
In basic PCL \[[DDMP05](#bib.bibx13), [Dat05](#bib.bibx8), [Der06](#bib.bibx16), [DDMR07](#bib.bibx14)\], axioms are provided for reasoning about Diffie-Hellman (DH) exponentials. To that end, the logic is extended with four additional axioms and some changes are made to the language and execution model.
#### 4.1.1 Capturing Diffie-Hellman behaviour
Below we recall the three elements of the DH extension (execution model, logic, proof system) and discuss their implications.
##### DH extension of the language and execution model.
The programming language and execution model are extended \[[DDMR07](#bib.bibx14), page 354\] with constructs $`g\hspace{0pt}{(a)}`$ and $`h\hspace{0pt}{(a,b)}`$, representing respectively $`g^{a}\operatorname{mod}p`$ and $`{({g^{a}\operatorname{mod}p})}^{b}\operatorname{mod}p`$. In the PCL papers, the mod $`p`$ and brackets are usually omitted, resulting in the abbreviations $`g^{a}`$ and $`g^{a\hspace{0pt}b}`$.
This extension does not reflect the actual algebraic properties of the exponential. For Diffie-Hellman, the crucial property is that $`g^{a\hspace{0pt}b} = g^{b\hspace{0pt}a}`$. This equation must be included in the execution model: if it is not, the following sequence of actions (denoted by their action predicates), which is perfectly valid DH-type behaviour, does not correspond to a valid execution in PCL:
<table>
<tbody>
<tr>
<td></td>
<td>Send(X,h(a,b))<Receive(Y,h(b,a))Send𝑋ℎ𝑎𝑏Receive𝑌ℎ𝑏𝑎\mbox{{\sf Send}}(X,h(a,b))<\mbox{{\sf Receive}}(Y,h(b,a))</td>
<td></td>
<td>(4)</td>
</tr>
</tbody>
</table>
Just extending the Has predicate in the logic is not sufficient, as the equivalence still has no counterpart in the execution model of PCL.
##### DH extension of the protocol logic.
In basic PCL, the predicate $`\text{Fresh}\hspace{0pt}{(X,x)}`$ holds for any $`x`$ generated by thread $`X`$ (captured by $`\text{Gen}\hspace{0pt}{(X,x)}`$), as long as $`x`$ is not sent as part of another term. The protocol logic is extended with an additional rule for the Fresh predicate, stating that $`\text{Fresh}\hspace{0pt}{(X,{g\hspace{0pt}{(x)}})}`$ is true if and only if $`\text{Fresh}\hspace{0pt}{(X,x)}`$ is true, and Has is extended in \[[DDMR07](#bib.bibx14), page 354\] by
<table>
<tbody>
<tr>
<td></td>
<td>Has(X,a)∧Has(X,g(b))⊃Has(X,h(a,b))∧Has(X,h(b,a))Has𝑋ℎ𝑎𝑏Has𝑋ℎ𝑏𝑎Has𝑋𝑎Has𝑋𝑔𝑏\mbox{{\sf Has}}(X,a)\land\mbox{{\sf Has}}(X,g(b))\supset\mbox{{\sf Has}}(X,h(a,b))\land\mbox{{\sf Has}}(X,h(b,a))</td>
<td></td>
</tr>
</tbody>
</table>
This rule is intended to capture the Diffie-Hellman equivalence relation. It is not sufficient, even with the addition of further axioms, as we show below.
##### DH extension of the proof system.
The proof system is extended in \[[DDMR07](#bib.bibx14), page 354, Table A.1\] by a definition and four new axioms. We reproduce them here:
<table>
<tbody>
<tr>
<td></td>
<td>Define Computes (DH)Computes(X,gab)≡((Has(X,a)∧Has(X,gb))∨(Has(X,b)∧Has(X,ga)))Define Computes (DH)Computes𝑋superscript𝑔𝑎𝑏Has𝑋𝑎Has𝑋superscript𝑔𝑏Has𝑋𝑏Has𝑋superscript𝑔𝑎\mbox{Define {\bf Computes (DH)}}\qquad\\ \mbox{{\sf Computes}}(X,g^{ab})\equiv\Bigl{(}\bigl{(}\mbox{{\sf Has}}(X,a)\land\mbox{{\sf Has}}(X,g^{b})\bigr{)}\lor\bigl{(}\mbox{{\sf Has}}(X,b)\land\mbox{{\sf Has}}(X,g^{a})\bigr{)}\Bigr{)}</td>
<td></td>
</tr>
</tbody>
</table>
This definition captures the intuition that a thread can compute the value of $`g^{a\hspace{0pt}b}`$ if and only if it has the required components. Note that for a thread that just has $`g^{a\hspace{0pt}b}`$ but not $`a`$ or $`b`$, this predicate is false, and therefore we have that $`\text{Has}\hspace{0pt}{(X,g^{a\hspace{0pt}b})}`$ does not imply $`\text{Computes}\hspace{0pt}{(X,g^{a\hspace{0pt}b})}`$.
Note that the definition of Computes suggests that $`g^{a\hspace{0pt}b} \neq g^{b\hspace{0pt}a}`$, because of the explicit listing of both sides of the disjunction. If the equality would hold, one could just define $`{\text{Computes}\hspace{0pt}{(X,g^{a\hspace{0pt}b})}} \equiv {{\text{Has}\hspace{0pt}{(X,a)}} \land {\text{Has}\hspace{0pt}{(X,g^{b})}}}`$ to achieve the same result.
Using the definition of Computes, the first axiom is given as:
<table>
<tbody>
<tr>
<td></td>
<td>DH1Computes(X,gab)⊃Has(X,gab)Has𝑋superscript𝑔𝑎𝑏DH1Computes𝑋superscript𝑔𝑎𝑏\mbox{\bf DH1}\qquad\mbox{{\sf Computes}}(X,g^{ab})\supset\mbox{{\sf Has}}(X,g^{ab})</td>
<td></td>
</tr>
</tbody>
</table>
and corresponds to the extension of the Has predicate, which can be seen by unfolding the definition of Computes. The second axiom states
<table>
<tbody>
<tr>
<td></td>
<td>DH2Has(X,gab)⊃(Computes(X,gab)∨∃m.(Receive(X,m)∧Contains(m,gab)))\mbox{\bf DH2}\qquad\mbox{{\sf Has}}(X,g^{ab})\supset\Bigl{(}\mbox{{\sf Computes}}(X,g^{ab})\lor\exists m.\bigl{(}\mbox{{\sf Receive}}(X,m)\land\mbox{{\sf Contains}}(m,g^{ab})\bigr{)}\Bigr{)}</td>
<td></td>
</tr>
</tbody>
</table>
This axiom captures the notion that one can only possess such a term $`g^{a\hspace{0pt}b}`$ by computing it from its parts, or receiving it. Effectively this restricts any agent (including the intruder) from knowing a term of the form $`g^{a\hspace{0pt}b}`$ at the start of each run, but it also excludes that $`g^{a\hspace{0pt}b}`$ is used as a parameter for a protocol.
For completeness, we also give the remaining DH axioms
<table>
<tbody>
<tr>
<td></td>
<td>DH3</td>
<td>(Receive(X,m)∧Contains(m,gab))⊃absentReceive𝑋𝑚Contains𝑚superscript𝑔𝑎𝑏\displaystyle\bigl{(}\mbox{{\sf Receive}}(X,m)\land\mbox{{\sf Contains}}(m,g^{ab})\bigr{)}\supset</td>
<td></td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td></td>
<td>∃Y,m′.(Computes(Y,gab)∧Send(Y,m′)∧Contains(m′,gab))formulae-sequence𝑌superscript𝑚′Computes𝑌superscript𝑔𝑎𝑏Send𝑌superscript𝑚′Containssuperscript𝑚′superscript𝑔𝑎𝑏\displaystyle\qquad\exists Y,m^{\prime}.\bigl{(}\mbox{{\sf Computes}}(Y,g^{ab})\land\mbox{{\sf Send}}(Y,m^{\prime})\land\mbox{{\sf Contains}}(m^{\prime},g^{ab})\bigr{)}\qquad\qquad\qquad\qquad</td>
<td></td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td>DH4</td>
<td>Fresh(X,a)⊃Fresh(X,ga)Fresh𝑋superscript𝑔𝑎Fresh𝑋𝑎\displaystyle\mbox{{\sf Fresh}}(X,a)\supset\mbox{{\sf Fresh}}(X,g^{a})</td>
<td></td>
</tr>
</tbody>
</table>
which capture the assumption that $`g^{a\hspace{0pt}b}`$ must have been either computed or received (but was not given as a parameter). The fourth axiom echoes the extension of the Fresh predicate.
Summarizing, even with the extension of the Has predicate and the additional axioms, the behaviour of the equivalence for Diffie-Hellman is not captured in PCL, for two reasons. First, with respect to the execution model, the sequence of actions represented by Formula ([4](#S4.E4 "In DH extension of the language and execution model. ‣ 4.1.1 Capturing Diffie-Hellman behaviour ‣ 4.1 Diffie-Hellman exponentials ‣ 4 Problems with PCL extensions ‣ On the Protocol Composition Logic PCLThis work was supported by the Hasler Foundation, ManCom project 2071. Manuscript")) cannot be enabled in the execution model by just changing the protocol logic. Second, we observe that given $`\text{Has}\hspace{0pt}{(X,{h\hspace{0pt}{(a,b)}})}`$, the protocol logic does not allow us to conclude that $`\text{Has}\hspace{0pt}{(X,{h\hspace{0pt}{(b,a)}})}`$ as one would expect (especially in the case where $`\neg{\text{Computes}\hspace{0pt}{(X,{h\hspace{0pt}{(a,b)}})}}`$). Hence, the essential equivalence for any Diffie-Hellman type protocol is currently not captured in the execution model, nor in the protocol logic.
#### 4.1.2 Implications
The Diffie-Hellman extension does not correctly capture the algebraic properties of Diffie-Hellman exponentials. As a result, certain possible behaviours of DH-like protocols are not considered within the logic and its execution model. The extension is therefore not a faithful representation of DH-like protocols.
#### 4.1.3 Resolving the problem
The essential feature to be captured is the equality $`g^{a\hspace{0pt}b} = g^{b\hspace{0pt}a}`$. If this is introduced at the term level, e.g. by having the equality $`{h\hspace{0pt}{(a,b)}} = {h\hspace{0pt}{(b,a)}}`$ in the term system, this solves at the same time the problem in the execution model, protocol logic and proof system. Some further modifications to the axioms are required, and the current proofs have to be modified to take the term equality into account.
### 4.2 Keyed hashes
In \[[He05](#bib.bibx19)\] and \[[HSD<sup>+</sup>05](#bib.bibx20)\] basic PCL is extended and applied to a case study of protocols that do not rely solely on signatures. As observed previously, such an application requires additional axioms. In particular, to prove authentication, we need to introduce an axiom that allows us to conclude the existence of a thread.
Appendix A of \[[HSD<sup>+</sup>05](#bib.bibx20)\] gives such additional axioms and definitions for keyed hash functions: one definition and four hashing axioms. We reproduce them here for convenience.
<table>
<tbody>
<tr>
<td></td>
<td>Define Computes (HASH)Computes(X,HASHK(a))≡Has(X,K)∧Has(X,a)Define Computes (HASH)Computes𝑋subscriptHASH𝐾𝑎Has𝑋𝐾Has𝑋𝑎\mbox{Define {\bf Computes (HASH)}}\qquad\mbox{{\sf Computes}}(X,\mbox{{\it HASH}}_{K}(a))\equiv\mbox{{\sf Has}}(X,K)\land\mbox{{\sf Has}}(X,a)</td>
<td></td>
</tr>
</tbody>
</table>
As we will see in the following, the intention of this predicate seems to be to express which thread computed the hashed value from its components. However, this intuition is not correctly captured by the definition: a typical use pattern of a keyed hash would be to provide an integrity check for a message $`m`$, as in
<table>
<tbody>
<tr>
<td></td>
<td>X^→Y^:m,HASHK(m) ,:→^𝑋^𝑌𝑚subscriptHASH𝐾𝑚 ,\hat{X}\rightarrow\hat{Y}\;:\;m,\mbox{{\it HASH}}_{K}(m)\mbox{ ,}</td>
<td></td>
<td>(5)</td>
</tr>
</tbody>
</table>
where $`K`$ is a key shared between $`\hat{X}`$ and $`\hat{Y}`$. In the typical use pattern, the $`\text{HASH}_{K}\hspace{0pt}{(m)}`$ is received by an agent which at that point can construct the message $`m`$ himself. In this use case, the message hash is received with (or after) the message, and is then used to verify the integrity of the message.
Assume we have that $`m`$ (or some subterm of it) is freshly generated by $`X`$, and we have that $`\text{Computes}\hspace{0pt}{(X,{\text{HASH}_{K}\hspace{0pt}{(m)}})}`$ holds. However, we can also prove that once a thread $`Y`$ of the recipient $`\hat{Y}`$ receives the message, $`Y`$ not only has $`K`$, but also $`m`$. Thus the predicate $`\text{Computes}\hspace{0pt}{(Y,{\text{HASH}_{K}\hspace{0pt}{(m)}})}`$ holds as well. Put differently, we have that for a typical use case, $`\text{Computes}\hspace{0pt}{(\ldots)}`$ holds for both the thread who creates as well as the thread who receives. In particular, this is also true for the protocols under consideration in \[[HSD<sup>+</sup>05](#bib.bibx20)\]. The protocols in the paper assume typing information (no hash mistaken for a nonce) and receive the hashed values at a point where they know the hash components.
Let $`K`$ be a key shared between $`\hat{X}`$ and $`\hat{Y}`$. Then, we can prove
<sup>5</sup>
<sup>5</sup>5Note that this cannot be proven using only basic PCL, but it can be proven using PCL combined with the meta-reasoning required to capture the assumption that variables are typed, as pointed out in Section [3.3](#S3.SS3 "3.3 No formal type system ‣ 3 Problems with basic PCL ‣ On the Protocol Composition Logic PCLThis work was supported by the Hasler Foundation, ManCom project 2071. Manuscript").
the following property (invariant) for the protocols in \[[HSD<sup>+</sup>05](#bib.bibx20)\]:
<table>
<tbody>
<tr>
<td></td>
<td>Honest(X^)∧Has(X,HASHK(m))⊃Computes(X,HASHK(m))Computes𝑋subscriptHASH𝐾𝑚Honest^𝑋Has𝑋subscriptHASH𝐾𝑚\mbox{{\sf Honest}}(\hat{X})\land\mbox{{\sf Has}}(X,\mbox{{\it HASH}}_{K}(m))\supset\mbox{{\sf Computes}}(X,\mbox{{\it HASH}}_{K}(m))</td>
<td></td>
<td>(6)</td>
</tr>
</tbody>
</table>
Intuitively one can see that an honest agent either creates the hash by having the components, or receives it. In this last case, where an honest agent receives the hash, the message integrity is verified using the hash, which is possible only because the recipient also has both the key $`K`$ and the message $`m`$. We will use this result in our discussion of the HASH4 axiom later on.
The first axiom in \[[HSD<sup>+</sup>05](#bib.bibx20), Appendix A\] that uses the definition of $`\text{Computes}\hspace{0pt}{()}`$ is the following:
<table>
<tbody>
<tr>
<td></td>
<td>HASH1Computes(X,HASHK(x))⊃Has(X,x)∧Has(X,K)Has𝑋𝑥Has𝑋𝐾HASH1Computes𝑋subscriptHASH𝐾𝑥\mbox{\bf HASH1}\qquad\mbox{{\sf Computes}}(X,\mbox{{\it HASH}}_{K}(x))\supset\mbox{{\sf Has}}(X,x)\land\mbox{{\sf Has}}(X,K)</td>
<td></td>
</tr>
</tbody>
</table>
If we unfold the definition of Computes in this axiom, it reduces to $`\phi \supset \phi`$. The second hash axiom states the following:
<table>
<tbody>
<tr>
<td></td>
<td>HASH2Computes(X,HASHK(x))⊃Has(X,HASHK(x))Has𝑋subscriptHASH𝐾𝑥HASH2Computes𝑋subscriptHASH𝐾𝑥\mbox{\bf HASH2}\qquad\mbox{{\sf Computes}}(X,\mbox{{\it HASH}}_{K}(x))\supset\mbox{{\sf Has}}(X,\mbox{{\it HASH}}_{K}(x))</td>
<td></td>
</tr>
</tbody>
</table>
This axiom informally states that whoever computes the hash value also has it. If we again unfold the definition of Computes, we can see that this works as an extension of the closure properties of the Has predicate (as defined in \[[DDMR07](#bib.bibx14)\]). Effectively, it introduces a new term structure $`f_{x}\hspace{0pt}{(y)}`$ to the PCL syntax, and expresses the one-way property of the hash function.
The third hash function axiom is
<table>
<tbody>
<tr>
<td></td>
<td>HASH3Receive(X,HASHK(x))⊃∃Y.Computes(Y,HASHK(x))∧Send(Y,HASHK(x)) .formulae-sequence𝑌HASH3Receive𝑋subscriptHASH𝐾𝑥Computes𝑌subscriptHASH𝐾𝑥Send𝑌subscriptHASH𝐾𝑥 .\mbox{\bf HASH3}\quad\mbox{{\sf Receive}}(X,\mbox{{\it HASH}}_{K}(x))\supset\exists Y.\mbox{{\sf Computes}}(Y,\mbox{{\it HASH}}_{K}(x))\land\mbox{{\sf Send}}(Y,\mbox{{\it HASH}}_{K}(x))\mbox{ .}</td>
<td></td>
</tr>
</tbody>
</table>
This axiom is not sound. Consider the following protocol in Alice and Bob style notation, where $`m`$ is a freshly generated nonce of the initiator, and $`K'`$ is a symmetrical key shared by the initiator and responder:
<table>
<tbody>
<tr>
<td></td>
<td>Init→Resp:ENCK′{|HASHK(m)|}Resp→Init:HASHK(m):→𝐼𝑛𝑖𝑡𝑅𝑒𝑠𝑝→subscriptENCsuperscript𝐾′subscriptHASH𝐾𝑚𝑅𝑒𝑠𝑝𝐼𝑛𝑖𝑡:subscriptHASH𝐾𝑚\begin{split}Init\rightarrow Resp&\;:\;\mbox{{\it ENC}}_{K^{\prime}}\{\mkern-4.0mu\lvert\,{\mbox{{\it HASH}}_{K}(m)}\,\rvert\mkern-4.0mu\}\\ Resp\rightarrow Init&\;:\;\mbox{{\it HASH}}_{K}(m)\end{split}</td>
<td></td>
<td>(7)</td>
</tr>
</tbody>
</table>
In a normal run of this protocol, the initiator sends the hash as part of a bigger term, but does not send $`m`$. Thus the responder cannot compute the hash itself, but simply decrypts the message, and sends the hash back. Thus, the preconditions are fulfilled by an initiator thread of this protocol, but the postcondition does not hold: only the initiator thread can compute it, but it does not send it out in the required form. (Observe that contrary to the protocols in \[[HSD<sup>+</sup>05](#bib.bibx20)\], this protocol in ([7](#S4.E7 "In 4.2 Keyed hashes ‣ 4 Problems with PCL extensions ‣ On the Protocol Composition Logic PCLThis work was supported by the Hasler Foundation, ManCom project 2071. Manuscript")) does not satisfy the typical usage pattern, and therefore Formula ([6](#S4.E6 "In 4.2 Keyed hashes ‣ 4 Problems with PCL extensions ‣ On the Protocol Composition Logic PCLThis work was supported by the Hasler Foundation, ManCom project 2071. Manuscript")) does not hold here.)
The fourth axiom is
<table>
<tbody>
<tr>
<td></td>
<td>HASH4Has(X,HASHK(x))⊃Computes(X,HASHK(x))∨∃Y,m.Computes(Y,HASHK(x))∧Send(Y,m)∧Contains(m,HASHK(x)) .formulae-sequenceComputes𝑋subscriptHASH𝐾𝑥𝑌HASH4Has𝑋subscriptHASH𝐾𝑥𝑚Computes𝑌subscriptHASH𝐾𝑥Send𝑌𝑚Contains𝑚subscriptHASH𝐾𝑥 .\mbox{\bf HASH4}\quad\mbox{{\sf Has}}(X,\mbox{{\it HASH}}_{K}(x))\supset\mbox{{\sf Computes}}(X,\mbox{{\it HASH}}_{K}(x))\lor{}\\ \exists Y,m.\;\mbox{{\sf Computes}}(Y,\mbox{{\it HASH}}_{K}(x))\land\mbox{{\sf Send}}(Y,m)\land\mbox{{\sf Contains}}(m,\mbox{{\it HASH}}_{K}(x))\mbox{ .}</td>
<td></td>
</tr>
</tbody>
</table>
This axiom seems to express: if someone has the hash value, she computed the hash herself, or somebody computed it and sent it (possibly as a subterm).
With a suitable restriction on the intruder knowledge (the intruder should not know some $`\text{HASH}_{K}\hspace{0pt}{(x)}`$ initially without knowing the components) this axiom can be proven sound. However, there is a problem with the applications of the axiom in the proofs of \[[DDMR07](#bib.bibx14)\]. Each time this axiom is applied in proofs in the paper, one assumes honesty of $`\hat{X}`$ and $`\hat{Y}`$, and that $`K`$ is the key shared between them. Thus we can rewrite the application of axiom HASH4 using Formula ([6](#S4.E6 "In 4.2 Keyed hashes ‣ 4 Problems with PCL extensions ‣ On the Protocol Composition Logic PCLThis work was supported by the Hasler Foundation, ManCom project 2071. Manuscript")), as in the following. First we unfold the definition of the axiom
<table>
<tbody>
<tr>
<td></td>
<td>Honest(X^)∧𝐇𝐀𝐒𝐇𝟒=Honest(X^)∧(Has(X,HASHK(x))⊃Computes(X,HASHK(x))∨Φ) ,Honest^𝑋𝐇𝐀𝐒𝐇𝟒Honest^𝑋Computes𝑋subscriptHASH𝐾𝑥ΦHas𝑋subscriptHASH𝐾𝑥 ,\mbox{{\sf Honest}}(\hat{X})\land{\bf HASH4}=\\ \mbox{{\sf Honest}}(\hat{X})\land\bigl{(}\mbox{{\sf Has}}(X,\mbox{{\it HASH}}_{K}(x))\supset\mbox{{\sf Computes}}(X,\mbox{{\it HASH}}_{K}(x))\lor\Phi\bigr{)}\mbox{ ,}</td>
<td></td>
<td>(8)</td>
</tr>
</tbody>
</table>
where $`\Phi`$ is used to denote the right-hand disjunct of the axiom HASH4, starting from the existential quantifier. Now we use Formula ([6](#S4.E6 "In 4.2 Keyed hashes ‣ 4 Problems with PCL extensions ‣ On the Protocol Composition Logic PCLThis work was supported by the Hasler Foundation, ManCom project 2071. Manuscript")) to get
<table>
<tbody>
<tr>
<td></td>
<td>Honest(X^)∧𝐇𝐀𝐒𝐇𝟒=Honest(X^) .Honest^𝑋𝐇𝐀𝐒𝐇𝟒Honest^𝑋 .\mbox{{\sf Honest}}(\hat{X})\land{\bf HASH4}=\mbox{{\sf Honest}}(\hat{X})\mbox{ .}</td>
<td></td>
<td>(9)</td>
</tr>
</tbody>
</table>
In other words, if $`\hat{X}`$ is assumed to be honest when applying axiom HASH4, the left-hand disjunct of the conclusion is always true (rendering the right-hand disjunct inconsequential). Because the right-hand disjunct of the conclusion contains the required thread existence, but is rendered useless, the axiom cannot be used here for proving authentication properties, based on our observations in Section [3.1](#S3.SS1 "3.1 Authentication properties only provable for signature protocols ‣ 3 Problems with basic PCL ‣ On the Protocol Composition Logic PCLThis work was supported by the Hasler Foundation, ManCom project 2071. Manuscript").
#### 4.2.1 Implications
The authentication proofs of the four-way handshake and group key handshake protocols in \[[HSD<sup>+</sup>05](#bib.bibx20), [He05](#bib.bibx19)\] are not correct in their current form. The reason for this is that these protocols do not contain signatures, and based on the observations in Section [3.1](#S3.SS1 "3.1 Authentication properties only provable for signature protocols ‣ 3 Problems with basic PCL ‣ On the Protocol Composition Logic PCLThis work was supported by the Hasler Foundation, ManCom project 2071. Manuscript"), any authentication proofs for such protocols must rely on newly introduced axioms. In this case, the only candidates are the axioms HASH3 and HASH4. As shown above, HASH3 is not sound, and for the protocols in \[[HSD<sup>+</sup>05](#bib.bibx20), [He05](#bib.bibx19)\], HASH4 cannot be used to prove the existence of a thread. Hence the authentication proofs of the handshakes are incorrect, and therefore also the compositional proof of authentication of IEEE 802.11i is incorrect.
#### 4.2.2 Resolving the problem
The properties of a keyed hash function like the ones occurring here are similar to the properties of symmetric-key cryptography. Thus, once sufficient reasoning infrastructure is in place for symmetric-key cryptography, one can devise a straightforward definition of a non-keyed hash function. Combining these two elements should lead to a natural definition and extension of the logic for keyed hashes.
Alternatively, it might be possible to reinstate rules that were used in older works, as e.g. found in \[[DDMP04a](#bib.bibx11), page 15\], assuming that their soundness can be proven in the current semantics.
## 5 Conclusions
In this paper, we have first shown that basic PCL as defined in \[[DDMP05](#bib.bibx13), [Dat05](#bib.bibx8), [Der06](#bib.bibx16), [DDMR07](#bib.bibx14)\] cannot be used to prove authentication properties of protocols without signatures. We consider this to be a strong limitation on the scope of basic PCL. Next, we have pointed out two reasoning tools that are missing in basic PCL: reasoning about preceding actions of a role within a thread, and the lack of a formal type system. With respect to PCL extensions, we have shown that the PCL Diffie-Hellman extension from \[[DDMP05](#bib.bibx13), [Dat05](#bib.bibx8), [Der06](#bib.bibx16), [DDMR07](#bib.bibx14)\] does not capture the algebraic behaviour of Diffie-Hellman protocols correctly in the execution model and protocol logic. Finally, we have shown that the extension for protocols with keyed hash functions in \[[HSD<sup>+</sup>05](#bib.bibx20), [He05](#bib.bibx19)\] is not sound.
It follows that some currently claimed PCL proofs cannot be proven within the basic logic as defined in \[[DDMP05](#bib.bibx13), [Dat05](#bib.bibx8), [Der06](#bib.bibx16), [DDMR07](#bib.bibx14)\], or make use of unsound axioms. These proofs include the authentication proofs of the $`CR`$ protocol from \[[DDMR07](#bib.bibx14), [DDMP05](#bib.bibx13)\] and the SSL/TLS and IEEE 802.11i protocols from \[[He05](#bib.bibx19), [HSD<sup>+</sup>05](#bib.bibx20)\].
In the papers on PCL the proofs of the invariants are often not given. This lack of proof details for the invariants is surprising, as the invariants themselves are the difficult part. Furthermore, many of the proofs seem impossible to be completed in PCL without resorting to meta-reasoning.
Some of the problems identified here can be solved easily, but for some problems more work is required. For example, straightforward solutions include adding a formal type system and adding a mechanism to reason about earlier actions. Other problems, such as establishing thread existence of honest agents for protocols based on public-key cryptography, do not seem to be solvable by straightforward fixes, and suggest that more extensive modifications to PCL are required.
It remains to be seen whether formal proofs in such a modified version of PCL can be concise.
## References
- \[ABB<sup>+</sup>05\] A. Armando, D. Basin, Y. Boichut, Y. Chevalier, L. Compagna, L. Cuellar, P.H. Drielsma, P. Heám, O. Kouchnarenko, J. Mantovani, S. Mödersheim, D. von Oheimb, M. Rusinowitch, J. Santiago, M. Turuani, L. Viganò, and L. Vigneron. The AVISPA tool for the automated validation of internet security protocols and applications.
volume 3576 of Lecture Notes in Computer Science, pages 281–285. Springer-Verlag, 2005.
- \[BAN90\] M. Burrows, M. Abadi, and R. Needham. A logic of authentication.
ACM Transactions on Computer Systems, 8(1):18–36, 1990.
- \[Bla01\] B. Blanchet. An efficient cryptographic protocol verifier based on Prolog rules.
In Proc. 14th IEEE Computer Security Foundations Workshop (CSFW), pages 82–96. IEEE Computer Society, 2001.
- \[BMV03\] D. Basin, S. Mödersheim, and L. Viganò. An on-the-fly model-checker for security protocol analysis.
In ESORICS, volume 2808 of Lecture Notes in Computer Science, pages 253–270. Springer-Verlag, 2003.
- \[BR94\] M. Bellare and P. Rogaway. Entity authentication and key distribution.
In CRYPTO ’93: Proceedings of the 13th annual international cryptology conference on Advances in cryptology, pages 232–249, New York, NY, USA, 1994. Springer-Verlag.
- \[CMdV06\] C.J.F. Cremers, S. Mauw, and E.P. de Vink. Injective synchronisation: an extension of the authentication hierarchy.
Theoretical Computer Science, 2006.
- \[Cre06\] C.J.F. Cremers.
Scyther - Semantics and Verification of Security Protocols.
PhD thesis, Computer Science Department, Eindhoven University of Technology, November 2006.
- \[Dat05\] A. Datta.
Security Analysis of Network Protocols: Compositional Reasoning and Complexity-theoretic Foundations.
PhD thesis, Computer Science Department, Stanford University, September 2005.
- \[DDMP03a\] A. Datta, A. Derek, J. C. Mitchell, and D. Pavlovic. Secure protocol composition (extended abstract).
In FMSE ’03: Proceedings of the 2003 ACM workshop on Formal methods in security engineering, pages 11–23, New York, NY, USA, 2003. ACM.
- \[DDMP03b\] A. Datta, A. Derek, J.C. Mitchell, and D. Pavlovic. A derivation system for security protocols and its logical formalization.
In Proc. 16th IEEE Computer Security Foundations Workshop (CSFW), pages 109–125. IEEE Computer Society, 2003.
- \[DDMP04a\] A. Datta, A. Derek, J.C. Mitchell, and D. Pavlovic. Abstraction and refinement in protocol derivation.
In Proc. 17th IEEE Computer Security Foundations Workshop (CSFW), pages 30–45, Washington, DC, USA, June 2004. IEEE Computer Society.
- \[DDMP04b\] A. Datta, A. Derek, J.C. Mitchell, and D. Pavlovic. Secure protocol composition.
Electron. Notes Theor. Comput. Sci., 83, 2004.
Proceedings of 19th Annual Conference on Mathematical Foundations of Programming Semantics.
- \[DDMP05\] A. Datta, A. Derek, J.C. Mitchell, and D. Pavlovic. A derivation system and compositional logic for security protocols.
Journal of Computer Security, 13(3):423–482, 2005.
- \[DDMR07\] A. Datta, A. Derek, J.C. Mitchell, and A. Roy. Protocol Composition Logic (PCL).
Electron. Notes Theor. Comput. Sci., 172:311–358, 2007.
Computation, Meaning, and Logic: Articles dedicated to Gordon Plotkin. Editors: L. Cardelli, M. Fiore, and G. Winskel.
- \[DDMW06\] A. Datta, A. Derek, J.C. Mitchell, and B. Warinschi. Computationally sound compositional logic for key exchange protocols.
Proc. 19th IEEE Computer Security Foundations Workshop (CSFW), 0:321–334, 2006.
- \[Der06\] A. Derek.
Formal Analysis of Security Protocols: Protocol Composition Logic.
PhD thesis, Computer Science Department, Stanford University, December 2006.
- \[DMP01\] N.A. Durgin, J.C. Mitchell, and D. Pavlovic. A compositional logic for protocol correctness.
In Proc. 14th IEEE Computer Security Foundations Workshop (CSFW), pages 241–272, 2001.
- \[DMP03\] N. Durgin, J.C. Mitchell, and D. Pavlovic. A compositional logic for proving security properties of protocols.
Journal of Computer Security, 11:667–721, 2003.
- \[He05\] C. He.
Analysis of Security Protocols for Wireless Networks.
PhD thesis, Department of Electrical Engineering, Stanford University, December 2005.
- \[HSD<sup>+</sup>05\] C. He, M. Sundararajan, A. Datta, A. Derek, and J.C. Mitchell. A modular correctness proof of IEEE 802.11i and TLS.
In CCS ’05: Proceedings of the 12th ACM conference on Computer and communications security, pages 2–15. ACM Press, 2005.
- \[Low97a\] G. Lowe. Casper: A compiler for the analysis of security protocols.
In Proc. 10th IEEE Computer Security Foundations Workshop (CSFW), pages 18–30. IEEE Computer Society, 1997.
- \[Low97b\] G. Lowe. A hierarchy of authentication specifications.
In Proc. 10th IEEE Computer Security Foundations Workshop (CSFW), pages 31–44. IEEE Computer Society, 1997.
- \[RDD<sup>+</sup>06\] A. Roy, A. Datta, A. Derek, J. C. Mitchell, and J.-P. Seifert. Secrecy analysis in protocol composition logic.
In M. Okada and I. Satoh, editors, Proceedings of 11th Annual Asian Computing Science Conference, Lecture Notes in Computer Science, December 2006.
Preliminary version. |
0709.3420 | cs.NI | cyber | # Co-Betweenness: A Pairwise Notion of Centrality
Eric D. Kolaczyk
Dept. of Mathematics and Statistics
Boston University
Boston, MA, USA
David B. Chua
Boston, MA
Marc Barthélemy
CEA-Centre d’Etudes de Bruyères-le-Châtel, Département de Physique Théorique et Appliquée BP12, 91680 Bruyères-Le-Châtel, France
###### Abstract
Betweenness centrality is a metric that seeks to quantify a sense of the importance of a vertex in a network graph in terms of its ‘control’ on the distribution of information along geodesic paths throughout that network. This quantity however does not capture how different vertices participate together in such control. In order to allow for the uncovering of finer details in this regard, we introduce here an extension of betweenness centrality to pairs of vertices, which we term co-betweenness, that provides the basis for quantifying various analogous pairwise notions of importance and control. More specifically, we motivate and define a precise notion of co-betweenness, we present an efficient algorithm for its computation, extending the algorithm of [brandes01](#bib.bib1) in a natural manner, and we illustrate the utilization of this co-betweenness on a handful of different communication networks. From these real-world examples, we show that the co-betweenness allows one to identify certain vertices which are not the most central vertices but which, nevertheless, act as important actors in the relaying and dispatching of information in the network.
## I Introduction
In social network analysis, the problem of determining the importance of actors in a network has been studied for a long time (see, for example, [Wasserman:1994](#bib.bib2) ). It is in this context that the concept of the centrality of a vertex in a network emerged. There are numerous measures that have been proposed to numerically quantify centrality which differ both in the nature of the underlying notion of vertex importance that they seek to capture, and in the manner in which that notion is encoded through some functional of the network graph. See [borgatti.everett](#bib.bib3) , for example, for a recent review and categorization of centrality measures.
Paths – as the routes by which flows (e.g., of information or commodities) travel over a network – are fundamental to the functioning of many networks. Therefore, not surprisingly, a number of centrality measures quantity importance with respect to the sharing of paths in the network. One popular measure is *betweenness centrality*. First introduced in its modern form by [freeman77](#bib.bib4) , the betweenness centrality is essentially a measure of how many geodesic (ie., shortest) paths run over a given vertex. In other words, in a social network for example, the betweenness centrality measures the extent to which an actor “lies between” other individuals in the network, with respect to the network path structure. As such, it is a measure of the control that actor has over the distribution of information in the network.
The betweenness centrality – as with all other centrality measures of which we are aware – is defined specifically with respect to a single given vertex. In particular, vertex centralities produce an ordering of the vertices in terms of their individual importance, but do not provide insight into the manner in which vertices act together in the spread of information across the network. Insight of this kind can be important in presenting an appropriately more nuanced view of the roles of the different vertices, beyond their individual importance. A first natural extension of the idea of centrality in this manner is to pairs of vertices.
In this paper, we introduce such an extension, which we term the *co-betweenness centrality*, or simply the co-betweenness. The co-betweenness of two vertices is essentially a measure of how many geodesic paths are shared by the vertices, and as such provides us with a sense of the interplay of vertices across the network. For example, the co-betweenness alone quantifies the extent to which pairs of vertices jointly control the distribution of information in the network. Alternatively, a standardized version of co-betweenness produces a well-defined measure of correlation between flows over the two vertices. Finally, an alternative normalization quantifies the extent to which one vertex controls the distribution of information to another vertex.
This paper is organized as follows. In Section [II](#S2 "II Background ‣ Co-Betweenness: A Pairwise Notion of Centrality"), we briefly review necessary technical background. In Section [III](#S3 "III Co-Betweenness ‣ Co-Betweenness: A Pairwise Notion of Centrality"), we provide a precise definition for the co-betweenness and related measures, and motivate each in the context of an Internet communication network. An algorithm for the efficient computation of co-betweenness, for all pairs of vertices in a network, is sketched in Section [IV](#S4 "IV Computation of Co-Betweenness ‣ Co-Betweenness: A Pairwise Notion of Centrality"), and its properties are discussed. In Section [V](#S5 "V Additional Illustrations ‣ Co-Betweenness: A Pairwise Notion of Centrality"), we further illustrate our measures using two social networks whose ties are reflective of communication. Some additional discussion is provided in Section [VI](#S6 "VI Discussion ‣ Co-Betweenness: A Pairwise Notion of Centrality"). Finally, a formal description of our algorithm, as well as pseudo-code, may be found in the appendix.
## II Background
Let $`\mathcal{G} = {(\mathcal{V},\mathcal{E})}`$ denote an undirected, connected network graph with $`n_{v}`$ vertices in $`\mathcal{V}`$ and $`n_{e}`$ edges in $`\mathcal{E}`$. A *walk* on $`\mathcal{G}`$, from a vertex $`v_{0}`$ to another vertex $`v_{\ell}`$, is an alternating sequence of vertices and edges, say $`\{ v_{0},e_{1},v_{1},\ldots,v_{\ell - 1},e_{\ell},v_{\ell}\}`$, where the endpoints of $`e_{i}`$ are $`\{ v_{i - 1},v_{i}\}`$. The length of this walk is said to be $`\ell`$. A *trail* is a walk without repeated edges, and a *path*, a trail without repeated vertices. A shortest path between two vertices $`{u,v} \in \mathcal{V}`$ is a path between $`u`$ and $`v`$ whose length $`\ell`$ is a minimum. Such a path is also called a geodesic and its length, the geodesic distance between $`u`$ and $`v`$. In the case that the graph $`\mathcal{G}`$ is weighted i.e., there is a collection of edge weights $`{\{ w_{e}\}}_{e \in \mathcal{E}}`$, where $`w_{e} \geq 0`$, shortest paths may be instead defined as paths for which the total sum of edge weights is a minimum. In the material that follows, we will restrict our exposition primarily to the case of unweighted graphs, but extensions to weighted graphs are straightforward. For additional background of this type, see, for example, the textbook [Clark:1991](#bib.bib5) .
Let $`\sigma_{s\hspace{0pt}t}`$ denote the total number of shortest paths that connect vertices $`s`$ and $`t`$ (with $`\sigma_{s\hspace{0pt}s} \equiv 1`$), and let $`\sigma_{s\hspace{0pt}t}\hspace{0pt}{(v)}`$ denote the number of shortest paths between $`s`$ and $`t`$ that also run over vertex $`v`$. Then we define the betweenness centrality of a vertex $`v`$ as a weighted sum of the number of paths through $`v`$,
<table>
<tbody>
<tr>
<td></td>
<td>B(v)=∑s,t∈𝒱∖{v}σst(v)σst.𝐵𝑣subscript𝑠𝑡𝒱𝑣subscript𝜎𝑠𝑡𝑣subscript𝜎𝑠𝑡.B(v)=\sum_{s,t\in\mathcal{V}\setminus\{v\}}\frac{\sigma_{st}(v)}{\sigma_{st}}\text{.}</td>
<td></td>
<td>(1)</td>
</tr>
</tbody>
</table>
Note that this definition excludes the shortest paths that start or end at $`v`$. However, in a connected graph we will have $`{\sigma_{s\hspace{0pt}t}\hspace{0pt}{(v)}} = \sigma_{s\hspace{0pt}t}`$ whenever $`s = v`$ or $`t = v`$, so the exclusion amounts to removing a constant term that would otherwise be present in the betweenness centrality of every vertex.
As an illustration, which we will use throughout this section and the next, consider the network in Figure [1](#S2.F1 "Figure 1 ‣ II Background ‣ Co-Betweenness: A Pairwise Notion of Centrality").
[Refer to caption](/html/0709.3420/assets/x1.png)
Figure 1: Graph representation of the physical topology of the Abilene network. Nodes represent regional network aggregation points (so-called ‘Points-of-Presence’ or PoP’s), and are labeled according to their metropolitan area, while the edges represent systems of optical transportation technologies and routing devices.
This is the Abilene network, an Internet network that is part of the Internet2 project address0, a research project devoted to development of the ‘next generation’ Internet. It serves as a so-called ‘backbone’ network for universities and research labs across the United States, in a manner analogous to the federal highway system of roads. We use this network for illustration because, as a technological communication network, the notions of connectivity, information, flows, and paths are all explicit and physical, and hence facilitate our initial discussion of betweenness and co-betweenness. Later, in Section [V](#S5 "V Additional Illustrations ‣ Co-Betweenness: A Pairwise Notion of Centrality"), we will illustrate further with two communication networks from the social network literature.
The information traversing this network takes the form of so-called ‘packets’, and the packets flow between origins and destinations on this network along paths strictly determined according to a set of underlying routing protocols (Technically, the Abilene network is more accurately described by a directed graph. But, given the fact that routing is typically symmetric in this network, we follow the Internet2 convention of displaying Abilene using an undirected graph.). A reasonable first approximation of the routing of information in this network is with respect to a set of unique shortest paths. In this case, the betweenness $`B\hspace{0pt}{(v)}`$ of any given vertex $`v \in \mathcal{V}`$ will be exactly equal to the number of shortest paths through $`v`$. The vertices in Figure [1](#S2.F1 "Figure 1 ‣ II Background ‣ Co-Betweenness: A Pairwise Notion of Centrality") correspond to metropolitan regions, and have been laid out roughly with respect to their true geographical locations. Intuitively and according to earlier work on centrality in spatial networks [Barrat:2005](#bib.bib7) , one might suspect that vertices near the central portion of the network, such as Denver or Indianapolis, have larger betweenness, being likely forced to support most of the flows of communication between east and west. We will see in Section [III](#S3 "III Co-Betweenness ‣ Co-Betweenness: A Pairwise Notion of Centrality") that such is indeed the case.
Until recently, standard algorithms for computing betweenness centralities $`B\hspace{0pt}{(v)}`$ for all vertices in a network had $`O\hspace{0pt}{(n_{v}^{3})}`$ running times, which was a stumbling block to their application in large-scale network analyses. Faster algorithms now exist, such as those introduced in [brandes01](#bib.bib1) , which have running time of $`O\hspace{0pt}{({n_{v}\hspace{0pt}n_{e}})}`$ on unweighted networks and $`O\hspace{0pt}{({{n_{v}\hspace{0pt}n_{e}} + {n_{v}^{2}\hspace{0pt}{\log n_{v}}}})}`$ on weighted networks, with an $`O\hspace{0pt}{({n_{v} + n_{e}})}`$ space requirement. These improvements derive from exploiting a clever recursive relation for the partial sums $`\sum_{t \in \mathcal{V}}{{\sigma_{s,t}\hspace{0pt}{(v)}}/\sigma_{s,t}}`$. As we will see, the need for efficient algorithms is even more important in the case of the co-betweenness, and we will make similar usage of recursions in developing an efficient algorithm for computing this quantity.
## III Co-Betweenness
We extend the concept of vertex betweenness centrality to pairs of vertices $`u`$ and $`v`$ by letting $`\sigma_{s\hspace{0pt}t}\hspace{0pt}{(u,v)}`$ denote the number of shortest paths between vertices $`s`$ and $`t`$ that pass through both $`u`$ and $`v`$, and defining the vertex co-betweenness as
<table>
<tbody>
<tr>
<td></td>
<td>𝒞(u,v)=∑s,t∈𝒱∖{u,v}σst(u,v)σst .𝒞𝑢𝑣subscript𝑠𝑡𝒱𝑢𝑣subscript𝜎𝑠𝑡𝑢𝑣subscript𝜎𝑠𝑡 .\mathcal{C}(u,v)=\sum_{\begin{subarray}{c}s,t\in\mathcal{V}\setminus\{u,v\}\end{subarray}}\frac{\sigma_{st}(u,v)}{\sigma_{st}}\text{\;.}</td>
<td></td>
<td>(2)</td>
</tr>
</tbody>
</table>
Thus co-betweenness gives us a measure of the number of shortest paths that run through both vertices $`u`$ and $`v`$.
To gain some insight into the relation between betweenness and co-betweenness, consider the following statistical perspective. Recall the Abilene network described in the previous section, and suppose that $`x_{s,t}`$ is a measure of the information (i.e., Internet packets) flowing between vertices $`s`$ and $`t`$ in the network. Similarly, let $`y_{v}`$ be the total information flowing through vertex $`v`$. Next, define $`\mathbf{x}`$ to be the $`n_{p} \times 1`$ vector of values $`x_{s,t}`$, where $`n_{p}`$ is the total number of pairs of vertices exchanging information, and $`\mathbf{y}`$, to be the $`n_{v} \times 1`$ vector of values $`y_{v}`$. A common expression modeling the relation between these two quantities is simply $`\mathbf{y} = {R\hspace{0pt}\mathbf{x}}`$, where $`R`$ is an $`n_{v} \times n_{p}`$ matrix (i.e., the so-called ‘routing matrix’) of $`0`$’s and $`1`$’s, indicating through which vertices each given routed path goes.
Now if $`\mathbf{x}`$ is considered as a random variable, with uncorrelated elements, then its covariance matrix is simply equal to the $`n_{p} \times n_{p}`$ identity matrix. The elements of $`\mathbf{y}`$, however, will be correlated, and their covariance matrix takes the form $`\Omega = {R\hspace{0pt}R^{T}}`$, by virtue of the linear relation between $`\mathbf{y}`$ and $`\mathbf{x}`$. Importantly, note that the diagonal elements of $`\Omega`$ are the betweenness’ $`B\hspace{0pt}{(v)}`$. Furthermore, the off-diagonal elements are the co-betweenness’ $`\mathcal{C}\hspace{0pt}{(u,v)}`$. When shortest paths are not unique, the same results hold if the matrix $`R`$ is expanded so that each shortest path between a pair of vertices $`s`$ and $`t`$ is afforded a separate column, and the non-zero entries of each such column has the value $`\sigma_{s,t}^{- 1}`$, rather than $`1`$. In this case, $`R`$ may be interpreted as a stochastic routing matrix.
To illustrate, in Figure [2](#S3.F2 "Figure 2 ‣ III Co-Betweenness ‣ Co-Betweenness: A Pairwise Notion of Centrality"), we show a network graph representation of the matrix $`\Omega`$ for the Abilene network.
[Refer to caption](/html/0709.3420/assets/x2.png)
Figure 2: Graph representation of the betweenness and co-betweenness values for the Abilene network. Vertices are in proportion to their betweeness. The width of each link is drawn in proportion to the co-betweenness of the two vertices incident to it.
The vertices are again placed roughly with respect to their actual geographic location, but are now drawn in proportion to their betweenness. Edges between pairs of vertices now represent non-zero co-betweenness for the pair, and are also drawn with a thickness in proportion to their value. A number of interesting features are evident from this graph. First, we see that, as surmised earlier, the more centrally located vertices tend to have the largest betweenness values. And it is these vertices that typically are involved with the larger co-betweenness values. Since the paths going through both a vertex $`s`$ and a vertex $`t`$ are a subset of the paths going through either one or the other, this tendancy for large co-betweenness to associate with large betweenness should not be a surprise. Also note that the co-betweenness values tend to be smaller between vertices separated by a larger geographical distance, which again seems intuitive.
Somewhat more surprising perhaps, however, is the manner in which the network becomes disconnected. The Seattle vertex is now isolated, as there are no paths that route through that vertex – only to and from. Additionally, the vertices Houston, Atlanta, and Washington now form a separate component in this graph, indicating that information is routed on paths running through both the first two and the last two, but not through all three, and also not through any of these and some other vertex. Overall, one gets the impression of information being routed primarily over paths along the upper portion of the network in Figure [1](#S2.F1 "Figure 1 ‣ II Background ‣ Co-Betweenness: A Pairwise Notion of Centrality"). A similar observation has been made in [chua05:kriging](#bib.bib8) , using different techniques.
While the raw co-betweenness values appear to be quite informative, one can imagine contexts in which it would be useful to compare co-betweenness’ across pairs of vertices in a manner that adjusts for the unequal betweenness of the participating vertices. The value
<table>
<tbody>
<tr>
<td></td>
<td>𝒞corr(u,v)=𝒞(u,v)B(u)B(v)superscript𝒞𝑐𝑜𝑟𝑟𝑢𝑣𝒞𝑢𝑣𝐵𝑢𝐵𝑣\mathcal{C}^{corr}(u,v)=\frac{\mathcal{C}(u,v)}{\sqrt{B(u)B(v)}}</td>
<td></td>
<td>(3)</td>
</tr>
</tbody>
</table>
is a natural candidate for a standardized version of the co-betweeness in ([2](#S3.E2 "In III Co-Betweenness ‣ Co-Betweenness: A Pairwise Notion of Centrality")), being simply the corresponding entry of the correlation matrix deriving from $`\Omega = {R\hspace{0pt}R^{T}}`$.
Figure [3](#S3.F3 "Figure 3 ‣ III Co-Betweenness ‣ Co-Betweenness: A Pairwise Notion of Centrality") shows a network graph representation of the quantities in $`\mathcal{C}^{c\hspace{0pt}o\hspace{0pt}r\hspace{0pt}r}`$ for the Abilene network, with edges again drawn in proportion to the values and vertices now naturally all drawn to be the same size.
[Refer to caption](/html/0709.3420/assets/x3.png)
Figure 3: Graph representation of the standardized co-betweenness values $`\mathcal{C}^{c\hspace{0pt}o\hspace{0pt}r\hspace{0pt}r}`$ for the Abilene network. Vertices are all drawn with equal size. Edge width is drawn in proportion to the standardized co-betweenness of the two vertices indicent to it.
Much of this network looks like that in Figure [2](#S3.F2 "Figure 2 ‣ III Co-Betweenness ‣ Co-Betweenness: A Pairwise Notion of Centrality"). The one notable exception is that the magnitude of the values between the three vertices in the lower subgraph component are now of a similar order to most of the other values in the other component. This fact may be interpreted as indicating that among themselves, adjusting for the lower levels of information flowing through this part of the network, these vertices are as strongly ‘correlated’ as many of the others.
The co-betweenness may also be used to define a directed notion of the strength of pairwise relationships. Let
<table>
<tbody>
<tr>
<td></td>
<td>𝒞(u|v)=𝒞(u,v)B(v)𝒞conditional𝑢𝑣𝒞𝑢𝑣𝐵𝑣\mathcal{C}(u|v)=\frac{\mathcal{C}(u,v)}{B(v)}</td>
<td></td>
<td>(4)</td>
</tr>
</tbody>
</table>
denote the relative proportion of shortest paths through $`v`$ that also go through $`u`$. This quantity may be interpreted as a measure of the control that vertex $`v`$ has over the information that passes through vertex $`u`$. Alternatively, under uniqueness of shortest paths, if from among the set of shortest paths through $`v`$ one is chosen uniformly at random, the value $`\mathcal{C}\hspace{0pt}{(\left. u \middle| v \right.)}`$ is the probabilty that the chosen path will also go through $`u`$. We call $`\mathcal{C}\hspace{0pt}{(\left. u \middle| v \right.)}`$ the conditional betweenness of $`u`$, given $`v`$. Note that, in general, $`{\mathcal{C}\hspace{0pt}{(\left. u \middle| v \right.)}} \neq {\mathcal{C}\hspace{0pt}{(\left. v \middle| u \right.)}}`$.
Figure [4](#S3.F4 "Figure 4 ‣ III Co-Betweenness ‣ Co-Betweenness: A Pairwise Notion of Centrality") shows a graph representation of the values $`\mathcal{C}\hspace{0pt}{(\left. u \middle| v \right.)}`$ for the Abilene network.
[Refer to caption](/html/0709.3420/assets/x4.png)
Figure 4: Directed graph representation of the conditional betweenness values $`\mathcal{C}\hspace{0pt}{(\left. u \middle| v \right.)}`$ (given by Eq. (4)) for the Abilene network. Edges are drawn with width in proportion to their value of $`\mathcal{C}\hspace{0pt}{(\left. u \middle| v \right.)}`$ and indicate how one vertex (at the head) controls the flow of information through another (at the tail).
Due to the asymmetry of these values in $`u`$ and $`v`$, arcs are used, rather than edges, with an arc from $`v`$ to $`u`$ corresponding to $`C\hspace{0pt}{(\left. u \middle| v \right.)}`$. The thickness of the arcs is proportional to these values, and is therefore indicative of the control exercised on the vertex at the tail by the vertex at the head. For improved visualization, we have used a simple circular layout for the vertices. Examination of this figure shows symmetry in the relationships between some pairs of vertices, but a strong asymmetry between most others. For example, vertices like Indianapolis, which were seen previously to have a large betweenness, clearly exercise a strong degree of control over almost any other vertices with which they share paths. More interestingly, note that certain vertices that are neighbors in the original Abilene network have more symmetric relationships than others. The conditional betweenness’ for Atlanta and Washington, DC, are fairly similar in magnitude, while those for Los Angeles and Sunnyvale are quite dissimilar, with the latter evidently exercising a noticeably greater degree of control over the former.
## IV Computation of Co-Betweenness
We discuss here the calculation of the co-betweenness values $`\mathcal{C}\hspace{0pt}{(u,v)}`$ in ([2](#S3.E2 "In III Co-Betweenness ‣ Co-Betweenness: A Pairwise Notion of Centrality")), for all pairs $`(u,v)`$, from which the other quantities in ([3](#S3.E3 "In III Co-Betweenness ‣ Co-Betweenness: A Pairwise Notion of Centrality")) and ([4](#S3.E4 "In III Co-Betweenness ‣ Co-Betweenness: A Pairwise Notion of Centrality")) follow trivially. At a first glance, it would appear that an algorithm of $`O\hspace{0pt}{(n_{v}^{4})}`$ running time is necessary, given that the number of vertex pairs grows as the square of the number of vertices. Such an implementation would render the notion of co-betweenness infeasible to implement in any but network graphs of relatively modest size. However, exploiting ideas similar to those underlying the algorithms of [brandes01](#bib.bib1) for calculating the betweenness’ $`B\hspace{0pt}{(v)}`$, a decidedly more efficient implementation may be obtained, as we now describe briefly. Details may be found in the appendix.
Our algorithm for computing co-betweenness involves a three-stage procedure for each vertex $`v \in \mathcal{V}`$. In the first stage, we perform a breadth-first traversal of the network graph $`\mathcal{G}`$, to quickly compute intermediary quantities such as $`\sigma_{s\hspace{0pt}v}`$, the number of shortest paths from a source $`s`$ to each other vertex $`v`$ in the network; in the process we form a directed acyclic graph that contains all shortest paths leading from vertex $`s`$. In the second stage, we iterate through each vertex in order of decreasing distance from $`s`$ and compute a score $`\delta_{s}\hspace{0pt}{(v)}`$ for each vertex that is related to its contribution to the co-betweenness. These contributions are then aggregated in a depth-first traversal of the directed acyclic graph, which is carried out in the third and final stage.
In order to compute the number of shortest paths $`\sigma_{s\hspace{0pt}v}`$ in the first stage, we note that the number of shortest paths from $`s`$ to a vertex $`v`$ is the sum of all shortest paths to each parent of $`v`$ in the directed acyclic graph rooted at $`s`$, namely,
<table>
<tbody>
<tr>
<td></td>
<td>σsv=∑t∈ps(v)σst.subscript𝜎𝑠𝑣subscript𝑡subscript𝑝𝑠𝑣subscript𝜎𝑠𝑡.\sigma_{sv}=\sum_{t\in p_{s}(v)}\sigma_{st}\text{.}</td>
<td></td>
<td>(5)</td>
</tr>
</tbody>
</table>
In the case of an undirected graph, this can be computed in the course of a breadth-first search with a running time of $`O\hspace{0pt}{(n_{e})}`$.
In the second stage, we compute $`\delta_{s}\hspace{0pt}{(v)}`$ using the recursive relation established in Theorem 6 of [brandes01](#bib.bib1) ,
<table>
<tbody>
<tr>
<td></td>
<td>δs(v)=∑w∈cs(w)σsvσsw(1+δs(w)) ,subscript𝛿𝑠𝑣subscript𝑤subscript𝑐𝑠𝑤subscript𝜎𝑠𝑣subscript𝜎𝑠𝑤1subscript𝛿𝑠𝑤 ,\delta_{s}(v)=\sum_{w\in c_{s}(w)}\frac{\sigma_{sv}}{\sigma_{sw}}\;(1+\delta_{s}(w))\text{\;,}</td>
<td></td>
<td>(6)</td>
</tr>
</tbody>
</table>
where $`c_{s}\hspace{0pt}{(v)}`$ denotes the set of child vertices of $`v`$ in the directed acyclic graph rooted at $`s`$.
Finally, in the third stage, we compute the co-betweennesses by interpreting the relation
<table>
<tbody>
<tr>
<td></td>
<td>𝒞(u,v)=∑s∈𝒱∖{u,v}δs(v)σsvσsv(u)𝒞𝑢𝑣subscript𝑠𝒱𝑢𝑣subscript𝛿𝑠𝑣subscript𝜎𝑠𝑣subscript𝜎𝑠𝑣𝑢\mathcal{C}(u,v)=\sum_{s\in\mathcal{V}\setminus\{u,v\}}\frac{\delta_{s}(v)}{\sigma_{sv}}\;\sigma_{sv}(u)</td>
<td></td>
<td>(7)</td>
</tr>
</tbody>
</table>
as assigning a contribution of $`\frac{\delta_{s}\hspace{0pt}{(v)}}{\sigma_{s\hspace{0pt}v}}`$ to $`C\hspace{0pt}{(u,v)}`$ for each of the $`\sigma_{s\hspace{0pt}v}\hspace{0pt}{(u)}`$ shortest paths to $`v`$ that run through $`u`$. We accumulate these contributions at each step of the depth-first traversal when we visit a vertex $`v`$ by adding $`\frac{\delta_{s}\hspace{0pt}{(v)}}{\sigma_{s\hspace{0pt}v}}`$ to $`\mathcal{C}\hspace{0pt}{(u,v)}`$ for every ancestor $`u`$ of the current vertex $`v`$.
Our proposed algorithms exploit recursions analogous to those of [brandes01](#bib.bib1) to produce run-times that are in the worst case $`O\hspace{0pt}{(n_{v}^{3})}`$, but in empirical studies were found to vary like $`O\hspace{0pt}{({{n_{v}\hspace{0pt}n_{e}} + {n_{v}^{2 + p}\hspace{0pt}{\log n_{v}}}})}`$ in general, or $`O\hspace{0pt}{({n_{v}^{2 + p}\hspace{0pt}{\log n_{v}}})}`$ in the case of sparse graphs. Here $`p`$ is related to the total number of shortest paths in the network and seems to lie comfortably between $`0.1`$ and $`0.5`$ in our experience. In the case of unique shortest paths, it may be shown rigorously that the running time reduces to $`O\hspace{0pt}{({{n_{v}\hspace{0pt}n_{e}} + {n_{v}^{2}\hspace{0pt}{\log n_{v}}}})}`$, and $`O\hspace{0pt}{({n_{v}^{2}\hspace{0pt}{\log n_{v}}})}`$ if the network is sparse as well as ‘small-world’ (i.e., with diameter of size $`O\hspace{0pt}{({\log n_{v}})}`$). See the appendix for details.
## V Additional Illustrations
We provide in this section additional illustration of the use of co-betweenness, based on two other networks graphs. Both graphs originally derive from social network analyses in which one goal was to understand the flow of certain information among actors.
### V.1 Michael’s Strike Network
Our first illustration involves the strike dataset of [michael](#bib.bib9) , which is also analyzed in detail in Chapter 7 of [pajek.book](#bib.bib10) . New management took over at a forest products manufacturing facility, and this management team proposed certain changes to the compensation package of the workers. The changes were not accepted by the workers, and a strike ensued, which was then followed by a halt in negotiations. At the request of management, who felt that the information about their proposed changes was not being communicated adequately, an outside consultant analyzed the communication structure among $`24`$ relevant actors.
The social network graph in Figure [5](#S5.F5 "Figure 5 ‣ V.1 Michael’s Strike Network ‣ V Additional Illustrations ‣ Co-Betweenness: A Pairwise Notion of Centrality") represents the communication structure among these actors, with an edge between two actors indicating that they communicated at some minimally sufficient level of frequency about the strike.
[Refer to caption](/html/0709.3420/assets/x5.png)
Figure 5: Original strike-group communication network of [michael](#bib.bib9) . Three subgroups are represented in this network: younger, Spanish-speaking employees (black vertices), younger, English-speaking employees (gray vertices), and older, English-speaking employees (white vertices). The two union negotiators, Sam and Wendle, are indicated by asterix’ next to their names. Edges indicate that the two incident actors communicated at some minimally sufficient level of frequency about the strike.
Three subgroups are present in the network: younger, Spanish-speaking employees (black vertices), younger, English-speaking employees (gray vertices), and older, English-speaking employees (white vertices). In addition, the two union negotiators, Sam and Wendle, are indicated by asterix’ next to their names. It is these last two that were responsible for explaining the details of the proposed changes to the employees. When the structure of this network was revealed, two additional actors – Bob and Norm – were approached, had the changes explained to them, which they then discussed with their colleagues, and within two days the employees requested that their union representatives re-open negotiations. The strike was resolved soon thereafter.
That such a result could follow by targeting Bob and Norm is not entirely surprising, from the perspective of the network structure. Both are cut-vertices (i.e., their removal would disconnect the network), and are incident to edges serving as bridges (i.e., their removal similarly would disconnect the network) from their respective groups to at least one of the other groups.
Co-betweenness provides a useful alternative characterization, one which explicitly emphasizes the patterns of communication in the network, as shown in Figure [6](#S5.F6 "Figure 6 ‣ V.1 Michael’s Strike Network ‣ V Additional Illustrations ‣ Co-Betweenness: A Pairwise Notion of Centrality").
[Refer to caption](/html/0709.3420/assets/x6.png)
Figure 6: Co-betweenness for the strike-group communication network. Actors located apart from the network, in the corners, are isolated under this representation, as they have zero betweenness and hence no co-betweenness with any other actors. (Note: Isolated vertices are drawn to have unit diameter, and not in proportion to their (zero) betweenness.)
As with Figure [2](#S3.F2 "Figure 2 ‣ III Co-Betweenness ‣ Co-Betweenness: A Pairwise Notion of Centrality"), vertices (now arranged in a circular layout) are drawn in proportion to their betweenness, and edges, to their co-betweenness. Bob and Norm clearly have the largest betweenness values, followed by Alejandro, who we remark also is a cut-vertex, but incident to a bridge to a smaller subnetwork than the other two (i.e., four younger Spanish-speakers, in comparison to nine younger English-speakers and 11 older English-speakers, for Bob and Norm, respectively). The importance of these three actors on the communication process is evident from the distinct triangle formed by their large co-betweenness values. Note that for the two union representatives, the co-betweenness values suggest that Sam also plays a non-trivial role in facilitating communication, but that Wendle is not well-situated in this regard. In fact, Wendle is not even connected to the main component of the graph, since his betweenness is zero (as is also true for six other actors).
A plot of the standardized co-betweenness $`\mathcal{C}^{c\hspace{0pt}o\hspace{0pt}r\hspace{0pt}r}`$ shows similar patterns overall, and we have therefore not included it here. The conditional betweenness $`\mathcal{C}\hspace{0pt}{(\left. u \middle| v \right.)}`$ for this network primarily shows most of the actors with large arcs pointing to Bob and Norm, and much smaller arcs pointing the opposite direction. This pattern further confirms the influence that these two actors can have on the other actors in the communication process. However, there are also some interesting asymmetrical relationships among the actors with smaller parts. For example, consider Figure [7](#S5.F7 "Figure 7 ‣ V.1 Michael’s Strike Network ‣ V Additional Illustrations ‣ Co-Betweenness: A Pairwise Notion of Centrality"), which shows the conditional betweenness among the older English-speaking employees.
[Refer to caption](/html/0709.3420/assets/x7.png)
Figure 7: Conditional co-betweenness for the older English-speaking actors in the strike-group communication network.
Ultrecht, for example, clearly has potential for a large amount of control on the communication of information passing through Russ, and similarly, Karl, on that through John.
### V.2 Zachary’s Karate Club Network
Our second illustration uses the karate club dataset of [zachary77](#bib.bib11) . Over the course of a couple of years in the 1970s, Zachary collected information from the members of a university karate club, including the number of situations (both inside and outside of the club) in which interactions occurred between members. During the course of this study, there was a dispute between the club’s administrator and the principal karate instructor. As a result, the club eventually split into two smaller clubs of approximately equal size—one centered around the administrator and the other centered around the instructor.
Figure [8](#S5.F8 "Figure 8 ‣ V.2 Zachary’s Karate Club Network ‣ V Additional Illustrations ‣ Co-Betweenness: A Pairwise Notion of Centrality") displays the network of social interactions between club members.
[Refer to caption](/html/0709.3420/assets/x8.png)
Figure 8: Karate club network of [zachary77](#bib.bib11) . The gray vertices represent members of one of the two smaller clubs and the white vertices represent members who went to the other club. The edges are drawn with a width proportional to the number of situations in which the two members interacted.
The gray vertices represent members of one of the two smaller clubs and the white vertices represent members who went to the other club. The edges are drawn with a width proportional to the number of situations in which the two members interacted. The graph clearly shows that the original club was already polarized into two groups centered about actors 1 and 34, who were the key players in the dispute that split the club in two.
The co-betweenness for this network is shown in Figure [9](#S5.F9 "Figure 9 ‣ V.2 Zachary’s Karate Club Network ‣ V Additional Illustrations ‣ Co-Betweenness: A Pairwise Notion of Centrality").
[Refer to caption](/html/0709.3420/assets/x9.png)
Figure 9: Co-betweenness for the karate club network. Actors in the upper-left and lower-right corners, separated from the connected component, are isolated due to zero betweenness. The two actors in the lower right-hand corner (i.e., $`a\hspace{0pt}5`$ and $`a\hspace{0pt}11`$) have non-zero betweenness, but are bridges, in the sense that they only serve to connect to other vertices, and hence have zero co-betweenness. (Note: The vertices for actors with zero betweenness are drawn to have unit diameter, for purposes of visibility.)
As in Figure [8](#S5.F8 "Figure 8 ‣ V.2 Zachary’s Karate Club Network ‣ V Additional Illustrations ‣ Co-Betweenness: A Pairwise Notion of Centrality"), the layout is done using an energy minimization algorithm. Again, as in our other examples, the co-betweenness entries are dominated by a handful of larger values. As might be expected, actors 1 and 34, who were at the center of the dispute, have the largest betweenness centralities and are also involved in the largest co-betweenness’. More interesting, however, is the fact that these two actors have a large co-betweenness with each other – despite not being directly connected in the original network graph. This indicates that they are nevertheless involved in connecting a large number of other pairs – probably through key intermediaries such as actors 3 and 32. These latter two actors, while certainly not cut-vertices, nevertheless seem to operate like conduits between the two groups, quite likely due to their direct ties to both actor 1 and either of actors 33 and 34, the latter of which are both central to the group of white vertices. The co-betweenness for actors 1 and 32 is in fact the largest in the entire network.
Also of potential interest are the 14 vertices that are isolated from the network in the co-betweenness representation. Some of these vertices, such as actor 8, have strong social interactions with certain other actors (i.e., with actors 1, 2, 3 and 4), but evidently play a peripheral role in the communication patterns of the network, as evidenced by their lack of betweenness. Alternatively, there are the vertices like those representing actors 5 and 11, who have some betweenness centrality but nonetheless find themselves cut off from the connected component in the co-betweenness graph. An examination of the definition of the co-betweenness tells us that such vertices must be bridge-vertices, in the sense that they only serve to connect pairs of other vertices, *i.e.*, they only occur in the middle of paths of length two.
## VI Discussion
We introduced in this paper the notion of co-betweenness as a natural and interpretable metric for quantifying the interplay between pairs of vertices in a network graph. As we discussed in different real world examples, this quantity has several interesting features. In particular, unlike the usual betweenness centrality which orders the vertices according to their importance in the information flow on the network, the co-betweenness gives additional information about the flow structure and the correlations between different actors. Using this quantity, we were able to identify vertices which are not the most central ones, but which however play a very important role in relaying the information and which therefore appear as crucial vertices in the control of the information flow.
In principle, of course, one could continue to define higher-order analogues, involving three or more vertices at a time. But the computational requirements associated with calculating such analogues would soon become burdensome. In the case of triplets of vertices, one can expect algorithms analogous to those presented here to scale no better than $`O\hspace{0pt}{(n_{v}^{3})}`$. Additionally, we remark that, in keeping with the statistics analogy made in Section [III](#S3 "III Co-Betweenness ‣ Co-Betweenness: A Pairwise Notion of Centrality"), it is likely that the pairwise ‘correlations’ picked up by co-betweenness captures to a large extent the more important elements of vertex interplay in the network, with respect to shortest paths.
Following the tendancies in the statistical physics literature on complex networks [Barabasi](#bib.bib12) ; [Vespignani](#bib.bib13) , it can be of some interest to explore the statistical properties of co-betweenness in large-scale networks. Some work in this direction may be found in [Chua.thesis](#bib.bib14) , where co-betweenness and functions thereof were examined in the context of standard network graph models. The most striking properties discovered were certain basic scaling relations with distance between vertices.
On a final note, we point out that, while our discussion here has been focused on co-betweenness for pairs of vertices in unweighted graphs, we have also developed the analogous quantities and algorithms for vertex co-betweenness on weighted graphs and for edge co-betweenness on unweighted and weighted graphs. Also see [chua05:kriging](#bib.bib8) , where a result is given relating edge betweenness to the eigen-values of the matrix edge-betweenness ‘covariance’ matrix, defined in analogy to the matrix $`\Omega`$ in Section [III](#S3 "III Co-Betweenness ‣ Co-Betweenness: A Pairwise Notion of Centrality").
This appendix contains details specific to the proposed algorithm for computing co-betweeness, including a derivation of key expressions, a rough analysis of algorithmic complexity. The pseudo-codes can be found at the address [address1](#bib.bib15) . Actual software implementing our algorithm, written in the Matlab software enviroment, is available at [address2](#bib.bib16) .
## Appendix A Derivation of Key Expressions
Central to our algorithm are the expressions in ([6](#S4.E6 "In IV Computation of Co-Betweenness ‣ Co-Betweenness: A Pairwise Notion of Centrality")) and ([7](#S4.E7 "In IV Computation of Co-Betweenness ‣ Co-Betweenness: A Pairwise Notion of Centrality")), the derivations for which we present here. Before doing so, however, we need to introduce some definitions and relations. First note that a simple combinatorial argument will show that
<table>
<tbody>
<tr>
<td></td>
<td>σst(v)={σsvσvtif d(s,t)=d(s,v)+d(v,t),0otherwise,subscript𝜎𝑠𝑡𝑣casessubscript𝜎𝑠𝑣subscript𝜎𝑣𝑡if d(s,t)=d(s,v)+d(v,t),0otherwise,\sigma_{st}(v)=\begin{cases}\sigma_{sv}\,\sigma_{vt}&\text{if $d(s,t)=d(s,v)+d(v,t)$,}\\ 0&\text{otherwise,}\end{cases}</td>
<td></td>
<td>(8)</td>
</tr>
</tbody>
</table>
and
<table>
<tbody>
<tr>
<td></td>
<td>σst(u,v)={σsuσuvσvtif d(s,t)=d(s,u),+d(u,v)+d(v,t),σsvσvuσutif d(s,t)=d(s,v),+d(v,u)+d(u,t),0otherwise.subscript𝜎𝑠𝑡𝑢𝑣casessubscript𝜎𝑠𝑢subscript𝜎𝑢𝑣subscript𝜎𝑣𝑡if d(s,t)=d(s,u),otherwise𝑑𝑢𝑣𝑑𝑣𝑡subscript𝜎𝑠𝑣subscript𝜎𝑣𝑢subscript𝜎𝑢𝑡if d(s,t)=d(s,v),otherwise𝑑𝑣𝑢𝑑𝑢𝑡0otherwise.\sigma_{st}(u,v)=\begin{cases}\sigma_{su}\,\sigma_{uv}\,\sigma_{vt}&\text{if $d(s,t)=d(s,u)$,}\\ &+d(u,v)+d(v,t),\\ \sigma_{sv}\,\sigma_{vu}\,\sigma_{ut}&\text{if $d(s,t)=d(s,v)$,}\\ &+d(v,u)+d(u,t),\\ 0&\text{otherwise.}\end{cases}</td>
<td></td>
<td>(9)</td>
</tr>
</tbody>
</table>
For the the sake of notational simplicity, we will assume, without loss of generality, that
<table>
<tbody>
<tr>
<td></td>
<td>d(s,u)≤d(s,v).𝑑𝑠𝑢𝑑𝑠𝑣d(s,u)\leq d(s,v).</td>
<td></td>
<td>(10)</td>
</tr>
</tbody>
</table>
for the remainder of this discussion.
The remaining quantities we need to introduce are notions of the path-dependency of vertices. In the spirit of [brandes01](#bib.bib1) , we define the “dependency” of vertices $`s`$ and $`t`$ on the vertex pair $`(u,v)`$ as
<table>
<tbody>
<tr>
<td></td>
<td>δst(u,v)=σst(u,v)σst,subscript𝛿𝑠𝑡𝑢𝑣subscript𝜎𝑠𝑡𝑢𝑣subscript𝜎𝑠𝑡,\delta_{st}(u,v)=\frac{\sigma_{st}(u,v)}{\sigma_{st}}\text{,}</td>
<td></td>
<td>(11)</td>
</tr>
</tbody>
</table>
and we define the dependency of $`s`$ alone on the pair of vertices $`(u,v)`$ as
<table>
<tbody>
<tr>
<td></td>
<td>δs(u,v)=∑t∈𝒱∖{u,v}δst(u,v)=∑t∈𝒱∖{u,v}σst(u,v)σst..subscript𝛿𝑠𝑢𝑣subscript𝑡𝒱𝑢𝑣subscript𝛿𝑠𝑡𝑢𝑣subscript𝑡𝒱𝑢𝑣subscript𝜎𝑠𝑡𝑢𝑣subscript𝜎𝑠𝑡..\delta_{s}(u,v)=\sum_{t\in\mathcal{V}\setminus\{u,v\}}\delta_{st}(u,v)=\sum_{t\in\mathcal{V}\setminus\{u,v\}}\frac{\sigma_{st}(u,v)}{\sigma_{st}}\text{.}\text{.}</td>
<td></td>
<td>(12)</td>
</tr>
</tbody>
</table>
Similarly, we define the pair-wise dependency of $`s`$ and $`t`$ on a single vertex $`v`$ as
<table>
<tbody>
<tr>
<td></td>
<td>δst(v)=σst(v)σst,subscript𝛿𝑠𝑡𝑣subscript𝜎𝑠𝑡𝑣subscript𝜎𝑠𝑡,\delta_{st}(v)=\frac{\sigma_{st}(v)}{\sigma_{st}}\text{,}</td>
<td></td>
<td>(13)</td>
</tr>
</tbody>
</table>
and the dependency of $`s`$ alone on $`v`$ as
<table>
<tbody>
<tr>
<td></td>
<td>δs(v)=∑t∈𝒱∖{v}δst(v)=∑t∈𝒱∖{v}σst(v)σst.subscript𝛿𝑠𝑣subscript𝑡𝒱𝑣subscript𝛿𝑠𝑡𝑣subscript𝑡𝒱𝑣subscript𝜎𝑠𝑡𝑣subscript𝜎𝑠𝑡.\delta_{s}(v)=\sum_{t\in\mathcal{V}\setminus\{v\}}\delta_{st}(v)=\sum_{t\in\mathcal{V}\setminus\{v\}}\frac{\sigma_{st}(v)}{\sigma_{st}}\text{.}</td>
<td></td>
<td>(14)</td>
</tr>
</tbody>
</table>
Note that unlike [brandes01](#bib.bib1) , we exclude $`t = v`$ from the sum in ([14](#A1.E14 "In Appendix A Derivation of Key Expressions ‣ Co-Betweenness: A Pairwise Notion of Centrality")). Two relations that follow immediately from these definitions, combined with ([8](#A1.E8 "In Appendix A Derivation of Key Expressions ‣ Co-Betweenness: A Pairwise Notion of Centrality")) and ([9](#A1.E9 "In Appendix A Derivation of Key Expressions ‣ Co-Betweenness: A Pairwise Notion of Centrality")), are
<table>
<tbody>
<tr>
<td></td>
<td>σst(u,v)subscript𝜎𝑠𝑡𝑢𝑣\displaystyle\sigma_{st}(u,v)</td>
<td>=\displaystyle=</td>
<td>σsuσuvσvtsubscript𝜎𝑠𝑢subscript𝜎𝑢𝑣subscript𝜎𝑣𝑡\displaystyle\sigma_{su}\,\sigma_{uv}\,\sigma_{vt}</td>
<td></td>
<td rowspan="4">(15)</td>
</tr>
<tr>
<td></td>
<td></td>
<td>=\displaystyle=</td>
<td>σsv(u)σvtsubscript𝜎𝑠𝑣𝑢subscript𝜎𝑣𝑡\displaystyle\sigma_{sv}(u)\,\sigma_{vt}</td>
<td></td>
</tr>
<tr>
<td></td>
<td></td>
<td>=\displaystyle=</td>
<td>σsv(u)σsvσsvσvtsubscript𝜎𝑠𝑣𝑢subscript𝜎𝑠𝑣subscript𝜎𝑠𝑣subscript𝜎𝑣𝑡\displaystyle\frac{\sigma_{sv}(u)}{\sigma_{sv}}\;\sigma_{sv}\,\sigma_{vt}</td>
<td></td>
</tr>
<tr>
<td></td>
<td></td>
<td>=\displaystyle=</td>
<td>δsv(u)σst(v),subscript𝛿𝑠𝑣𝑢subscript𝜎𝑠𝑡𝑣,\displaystyle\delta_{sv}(u)\,\sigma_{st}(v)\text{,}</td>
<td></td>
</tr>
</tbody>
</table>
and
<table>
<tbody>
<tr>
<td></td>
<td>δst(u,v)=σst(u,v)σst=δsv(u)σst(v)σst=δsv(u)δst(v).subscript𝛿𝑠𝑡𝑢𝑣subscript𝜎𝑠𝑡𝑢𝑣subscript𝜎𝑠𝑡subscript𝛿𝑠𝑣𝑢subscript𝜎𝑠𝑡𝑣subscript𝜎𝑠𝑡subscript𝛿𝑠𝑣𝑢subscript𝛿𝑠𝑡𝑣.\delta_{st}(u,v)=\frac{\sigma_{st}(u,v)}{\sigma_{st}}=\frac{\delta_{sv}(u)\,\sigma_{st}(v)}{\sigma_{st}}=\delta_{sv}(u)\,\delta_{st}(v)\text{.}</td>
<td></td>
<td>(16)</td>
</tr>
</tbody>
</table>
These two relations allow us to show that
<table>
<tbody>
<tr>
<td></td>
<td>δs(u,v)subscript𝛿𝑠𝑢𝑣\displaystyle\delta_{s}(u,v)</td>
<td>=∑t∈𝒱∖{u,v}δst(u,v)absentsubscript𝑡𝒱𝑢𝑣subscript𝛿𝑠𝑡𝑢𝑣\displaystyle=\sum_{t\in\mathcal{V}\setminus\{u,v\}}\delta_{st}(u,v)</td>
<td colspan="3"></td>
<td>(17)</td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td></td>
<td>=∑t∈𝒱∖{u,v}δsv(u)δst(v)absentsubscript𝑡𝒱𝑢𝑣subscript𝛿𝑠𝑣𝑢subscript𝛿𝑠𝑡𝑣\displaystyle=\sum_{t\in\mathcal{V}\setminus\{u,v\}}\delta_{sv}(u)\,\delta_{st}(v)</td>
<td></td>
<td>by (16)</td>
<td></td>
<td>(18)</td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td></td>
<td>=δsv(u)δs(v)absentsubscript𝛿𝑠𝑣𝑢subscript𝛿𝑠𝑣\displaystyle=\delta_{sv}(u)\,\delta_{s}(v)</td>
<td colspan="3"></td>
<td>(19)</td>
</tr>
</tbody>
</table>
since $`{\delta_{s\hspace{0pt}u}\hspace{0pt}{(v)}} = 0`$ by ([10](#A1.E10 "In Appendix A Derivation of Key Expressions ‣ Co-Betweenness: A Pairwise Notion of Centrality")) and using Eq. ([15](#A1.E15 "In Appendix A Derivation of Key Expressions ‣ Co-Betweenness: A Pairwise Notion of Centrality")), we obtain
<table>
<tbody>
<tr>
<td></td>
<td>δs(u,v)=δs(v)σsvσsv(u)subscript𝛿𝑠𝑢𝑣subscript𝛿𝑠𝑣subscript𝜎𝑠𝑣subscript𝜎𝑠𝑣𝑢\delta_{s}(u,v)=\frac{\delta_{s}(v)}{\sigma_{sv}}\;\sigma_{sv}(u)</td>
<td></td>
<td>(20)</td>
</tr>
</tbody>
</table>
We use this result to re-express the co-betweenness defined in ([2](#S3.E2 "In III Co-Betweenness ‣ Co-Betweenness: A Pairwise Notion of Centrality")) as
<table>
<tbody>
<tr>
<td></td>
<td>𝒞(u,v)𝒞𝑢𝑣\displaystyle\mathcal{C}(u,v)</td>
<td>=∑s,t∈𝒱∖{u,v}δst(u,v)absentsubscript𝑠𝑡𝒱𝑢𝑣subscript𝛿𝑠𝑡𝑢𝑣\displaystyle=\sum_{s,t\in\mathcal{V}\setminus\{u,v\}}\delta_{st}(u,v)</td>
<td></td>
<td>(21)</td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td></td>
<td>=∑s∈𝒱∖{u,v}(∑t∈𝒱∖{u,v}δst(u,v))absentsubscript𝑠𝒱𝑢𝑣subscript𝑡𝒱𝑢𝑣subscript𝛿𝑠𝑡𝑢𝑣\displaystyle=\sum_{s\in\mathcal{V}\setminus\{u,v\}}\left(\sum_{t\in\mathcal{V}\setminus\{u,v\}}\delta_{st}(u,v)\right)</td>
<td></td>
<td>(22)</td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td></td>
<td>=∑s∈𝒱∖{u,v}δs(u,v)absentsubscript𝑠𝒱𝑢𝑣subscript𝛿𝑠𝑢𝑣\displaystyle=\sum_{s\in\mathcal{V}\setminus\{u,v\}}\delta_{s}(u,v)</td>
<td></td>
<td>(23)</td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td></td>
<td>=∑s∈𝒱∖{u,v}δs(v)σsvσsv(u).absentsubscript𝑠𝒱𝑢𝑣subscript𝛿𝑠𝑣subscript𝜎𝑠𝑣subscript𝜎𝑠𝑣𝑢.\displaystyle=\sum_{s\in\mathcal{V}\setminus\{u,v\}}\frac{\delta_{s}(v)}{\sigma_{sv}}\;\sigma_{sv}(u)\text{.}</td>
<td></td>
<td>(24)</td>
</tr>
</tbody>
</table>
Lastly, to establish the recursive relation in ([6](#S4.E6 "In IV Computation of Co-Betweenness ‣ Co-Betweenness: A Pairwise Notion of Centrality")), note that for a child vertex $`w \in {c_{s}\hspace{0pt}{(v)}}`$ every path to $`v`$ gives rise to exactly one path to $`w`$ by following the edge $`(v,w)`$. This means that
<table>
<tbody>
<tr>
<td></td>
<td>σsw(v)=σsv for w∈cs(v),subscript𝜎𝑠𝑤𝑣subscript𝜎𝑠𝑣 for w∈cs(v),\sigma_{sw}(v)=\sigma_{sv}\text{\quad for $w\in c_{s}(v)$,}</td>
<td></td>
<td>(25)</td>
</tr>
</tbody>
</table>
and that
<table>
<tbody>
<tr>
<td></td>
<td>δsw(v)=σsw(v)σsw=σsvσsw for w∈cs(v).subscript𝛿𝑠𝑤𝑣subscript𝜎𝑠𝑤𝑣subscript𝜎𝑠𝑤subscript𝜎𝑠𝑣subscript𝜎𝑠𝑤 for w∈cs(v).\delta_{sw}(v)=\frac{\sigma_{sw}(v)}{\sigma_{sw}}=\frac{\sigma_{sv}}{\sigma_{sw}}\text{\quad for $w\in c_{s}(v)$.}</td>
<td></td>
<td>(26)</td>
</tr>
</tbody>
</table>
Also note that for $`t = w`$ we have
<table>
<tbody>
<tr>
<td></td>
<td>δst(w)=1.subscript𝛿𝑠𝑡𝑤1.\delta_{st}(w)=1\text{.}</td>
<td></td>
<td>(27)</td>
</tr>
</tbody>
</table>
This allows us to decompose $`\delta_{s}\hspace{0pt}{(v)}`$ in essentially the same manner as [brandes01](#bib.bib1) , namely,
<table>
<tbody>
<tr>
<td></td>
<td>δs(v)subscript𝛿𝑠𝑣\displaystyle\delta_{s}(v)</td>
<td>=∑t∈𝒱∖{v}δst(v)absentsubscript𝑡𝒱𝑣subscript𝛿𝑠𝑡𝑣\displaystyle=\sum_{t\in\mathcal{V}\setminus\{v\}}\delta_{st}(v)</td>
<td colspan="3"></td>
<td>(28)</td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td></td>
<td>=∑t∈𝒱∖{v}∑w∈cs(v)δst(v,w)absentsubscript𝑡𝒱𝑣subscript𝑤subscript𝑐𝑠𝑣subscript𝛿𝑠𝑡𝑣𝑤\displaystyle=\sum_{t\in\mathcal{V}\setminus\{v\}}\;\sum_{w\in c_{s}(v)}\delta_{st}(v,w)</td>
<td colspan="3"></td>
<td>(29)</td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td></td>
<td>=∑w∈cs(v)∑t∈𝒱∖{v}δst(v,w)absentsubscript𝑤subscript𝑐𝑠𝑣subscript𝑡𝒱𝑣subscript𝛿𝑠𝑡𝑣𝑤\displaystyle=\sum_{w\in c_{s}(v)}\;\sum_{t\in\mathcal{V}\setminus\{v\}}\delta_{st}(v,w)</td>
<td colspan="3"></td>
<td>(30)</td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td></td>
<td>=∑w∈cs(v)∑t∈𝒱∖{v}δsw(v)δst(w)absentsubscript𝑤subscript𝑐𝑠𝑣subscript𝑡𝒱𝑣subscript𝛿𝑠𝑤𝑣subscript𝛿𝑠𝑡𝑤\displaystyle=\sum_{w\in c_{s}(v)}\;\sum_{t\in\mathcal{V}\setminus\{v\}}\delta_{sw}(v)\,\delta_{st}(w)</td>
<td></td>
<td>by (16)</td>
<td></td>
<td>(31)</td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td></td>
<td>=∑w∈cs(v)σsvσsw(1+∑t∈𝒱∖{v,w}δst(w))absentsubscript𝑤subscript𝑐𝑠𝑣subscript𝜎𝑠𝑣subscript𝜎𝑠𝑤1subscript𝑡𝒱𝑣𝑤subscript𝛿𝑠𝑡𝑤\displaystyle=\sum_{w\in c_{s}(v)}\frac{\sigma_{sv}}{\sigma_{sw}}\left(1+\sum_{t\in\mathcal{V}\setminus\{v,w\}}\delta_{st}(w)\right)</td>
<td colspan="3"></td>
<td>(32)</td>
</tr>
</tbody>
</table>
Using ([26](#A1.E26 "In Appendix A Derivation of Key Expressions ‣ Co-Betweenness: A Pairwise Notion of Centrality")) and ([27](#A1.E27 "In Appendix A Derivation of Key Expressions ‣ Co-Betweenness: A Pairwise Notion of Centrality")), we then obtain
<table>
<tbody>
<tr>
<td></td>
<td>δs(v)=∑w∈cs(v)σsvσsw(1+δs(w)).subscript𝛿𝑠𝑣subscript𝑤subscript𝑐𝑠𝑣subscript𝜎𝑠𝑣subscript𝜎𝑠𝑤1subscript𝛿𝑠𝑤.\delta_{s}(v)=\sum_{w\in c_{s}(v)}\frac{\sigma_{sv}}{\sigma_{sw}}\;(1+\delta_{s}(w))\text{.}</td>
<td></td>
<td>(33)</td>
</tr>
</tbody>
</table>
Where the last equality is due to the fact that since $`w`$ is a child of $`v`$ we have $`{\sigma_{s\hspace{0pt}v}\hspace{0pt}{(w)}} = 0`$ and thus $`{\delta_{s\hspace{0pt}v}\hspace{0pt}{(w)}} = 0`$.
## Appendix B Algorithmic Complexity
Standard breadth-first search results put the running time for the first stage of our algorithm at $`O\hspace{0pt}{(n_{e})}`$, and since we touch each edge at most twice when we compute the dependency scores $`\delta_{s}\hspace{0pt}{(v)}`$, the running time for the second stage is also $`O\hspace{0pt}{(n_{e})}`$. Since we repeat each stage for each vertex in the network, the first two stages have a running time of $`O\hspace{0pt}{({n_{v}\hspace{0pt}n_{e}})}`$. The running time for the depth-first traversal, that occurs during the third stage, depends on the number and length of all shortest paths in the network. Overall, we visit every shortest path once and compute a co-betweenness contribution for each edge of every shortest path. For ‘small-world’ networks i.e., networks with an $`O\hspace{0pt}{({\log n_{v}})}`$ diameter, we must compute $`O\hspace{0pt}{({\sigma \cdot {\log n_{v}}})}`$ contributions, where $`\sigma = {\sum_{{u,v} \in \mathcal{V}}\sigma_{u\hspace{0pt}v}}`$ is the total number of shortest paths in the network. So the overall running time for the algorithm is $`O\hspace{0pt}{({{n_{v}\hspace{0pt}n_{e}} + {\sigma\hspace{0pt}{\log n_{v}}}})}`$. Empirical evidence suggests that the upper bound for the average $`\frac{1}{|\mathcal{V}|}\hspace{0pt}{\sum_{u \in \mathcal{V}}\sigma_{u\hspace{0pt}v}}`$ ranges from $`n_{v}^{0.19}`$ to $`n_{v}^{0.32}`$ for common random graph models, and at worst has been seen to reach $`n_{v}^{0.62}`$ in the case of a network of airports. (In the latter case, there were extreme fluctuations in $`\frac{1}{|\mathcal{V}|}\hspace{0pt}{\sum_{u \in \mathcal{V}}\sigma_{u\hspace{0pt}v}}`$ so the total number of shortest paths, $`\sigma`$, might be much smaller than $`n_{v}\hspace{0pt}{({n_{v} - 1})}`$ times this upper bound.) This suggests a running time of $`O\hspace{0pt}{({{n_{v}\hspace{0pt}n_{e}} + {n_{v}^{2 + p}\hspace{0pt}{\log n_{v}}}})}`$, though it is an open question to show this rigorously. In the case of sparse networks, where $`n_{e} \sim n_{v}`$, this reduces to a running time of $`O\hspace{0pt}{({n_{v}^{2 + p}\hspace{0pt}{\log n_{v}}})}`$.
## References
- \(1\)
U. Brandes, Journal of Mathematical Sociology 25, 163 (2001).
- \(2\)
S. Wasserman and K. Faust, Social Network Analysis: Methods and applications, Cambridge University Press (1994).
- \(3\)
S.P. Borgatti, M.G. Everett, Social Networks 28, 466-484 (2006).
- \(4\)
L.C. Freeman, Sociometry 40, 35-41 (1977).
- \(5\)
J. Clark and D.A. Holton, A first look at graph theory, World Scientific (1991).
- \(6\)
http://www.internet2.edu/
- \(7\) A. Barrat, M. Barthélemy, A. Vespignani, J. Stat. Mech. (2005) P05003.
- \(8\)
D.B. Chua, E.D. Kolaczyk, M. Crovella, IEEE Journal on Selected Areas in Communications, Special issue on ‘Sampling the Internet’, 24, 2263-2272 (2006).
- \(9\)
J.H. Michael, Forest Products Journal 47, 41-45 (1997).
- \(10\)
W. de Nooy, A. Mrvar, V. Batagelj, Exploratory Social Network Analysis with Pajek, Cambridge University Press (Cambridge, UK, 2005).
- \(11\)
W. Zachary, Journal of Anthropological Research 33, 452-473 (1977).
- \(12\)
R. Albert and A.-L. Barabási, Rev. Mod. Phys. 74, 47 (2000).
- \(13\)
R. Pastor-Satorras and A. Vespignani, Evolution and structure of the Internet: A statistical physics approach (Cambridge University Press, Cambridge, 2003).
- \(14\) D.B. Chua, PhD thesis (2007).
- \(15\)
http://math.bu.edu/people/kolaczyk/pubs/ChuaThesis/
- \(16\)
http://math.bu.edu/people/kolaczyk/software.html |
0710.0021 | cs.CR | cyber | New families of cryptographic systems
Maryna NESTERENKO <sup>†§</sup>, Jiri PATERA <sup>†</sup> and Dmytro ZHAVROTSKYI <sup>‡</sup>
<sup>†</sup> Centre de Recherches Mathématiques , Université de Montréal, 6128, succursale Centre-ville,
Montréal, QC, H3C3J7, Canada
E-mail: patera@crm.umontreal.ca
<sup>§</sup> Institute of Mathematics of NAS of Ukraine, 3 Tereshchenkivs’ka Str., Kyiv-4, 01601 Ukraine
E-mail: maryna@imath.kiev.ua
<sup>‡</sup> Interzvyazok, 15 Novokostyantynivska str., Kyiv, 04080, Ukraine
E-mail: dima_zh@isv.com.ua
###### Abstract
A symmetric encryption method based on properties of quasicrystals is proposed. The advantages of the cipher are strict aperiodicity and everywhere discontinuous property as well as the speed of computation, simplicity of implementation and a straightforward possibility of extending the method to encryption of higher dimensional data.
## 1 Introduction
Here we address the ‘classical’ cryptographic problem, i.e. protection of transmitted and stored data from divulgence and distortion. Such problem frequently arise when data creation or/and data usage are disjoint in time or/and space. There exist many approaches to this problem in modern cryptography, which can be divided into three main groups in accordance with their key primitives, i.e., unkeyed, symmetryc-key and public-key encodings, see \[[2](#bib.bib2), [4](#bib.bib4), [9](#bib.bib9)\].
In the present work we propose a new generic encoding procedure based on the aperiodic point sets, called quasicrystals in the physics literature and model sets in the mathematics literature \[[1](#bib.bib1), [3](#bib.bib3)\]. Generally speaking, such an algorithm is a symmetric stream cipher endowed with strict aperiodicity (no periodic subsets) and everywhere discontinuous properties.
The paper starts by preliminaries and general encoding idea, which is followed by the precise statement of pertinent mathematical ingredients, Sections II and III. In Section [4](#S4 "4 Application to image coding") applications of the proposed cipher to encoding of bitmap pictures are presented and discussed. The advantages of the approach are in the speed of computation and a straightforward possibility of extending the method to encryption of 3 and higher dimensional data. Possible range of private keys and some other applications of quasicrystals are discussed in Conclusions.
## 2 Preliminaries and encryption procedure
We call cut-and-project quasicrystal a discrete deterministic aperiodic point set $`\Lambda`$, in a finite-dimensional real Euclidean space. In many ways such quasicrystals resemble lattices in all but the translation symmetry. Sometimes they are even called aperiodic lattices or aperiodic crystals. In this work we make use of two remarkable properties of the quasicrystals: (i) no periodic subsets of any kind are contained in a quasicrystal, and (ii) the discontinuity of the ‘star map’ between a quasicrystal $`\Lambda`$ and its ‘acceptance window’ $`\Omega`$.
General idea of the proposed approach is to take given digital data, set a one-to-one correspondence between information bites and integer numbers $`N`$, and then to map points of $`N`$ on a quasicrystal fragment $`\Lambda`$ (see the first two lines of Fig. [1](#S2.F1 "Figure 1 ‣ 2 Preliminaries and encryption procedure")). At this stage the data is mapped to the quasicrystal points. It is followed by the application of the star map to the quasicrystal points (see the second and the third lines of Fig. [1](#S2.F1 "Figure 1 ‣ 2 Preliminaries and encryption procedure")), which is a crucial step of the encryption. As a result, we get the quasicrystal points, carrying our data, ‘tossed up’ in the acceptance window $`\Omega`$. It remains to map the content of $`\Omega`$ on a desirable lattice for the output (see the last two lines of Fig. [1](#S2.F1 "Figure 1 ‣ 2 Preliminaries and encryption procedure")).
Decryption proceeds in the opposite direction.
The set of private keys in our encryption method is infinite and is discussed in Conclusions. In addition, one may choose to use any common substitution scheme.
Generation of quasicrystal points is very flexible. It can either be done as a single stream or split into several blocks. Moreover, calculation of the points of each block is easily amenable to parallel computing.
Using quasicrystals, all calculations can be carried out in integers. That guarantees absolute accuracy of coding and decoding procedures, assures the stability of the algorithm.
[Refer to caption](/html/0710.0021/assets/x1.png)
Figure 1: Scheme of the simplest quasicrystal-based cipher
## 3 Mathematical ingredients
### 3.1 Theoretical background
Theoretical background of the work can be found in \[[1](#bib.bib1), [3](#bib.bib3), [5](#bib.bib5), [7](#bib.bib7), [8](#bib.bib8), [10](#bib.bib10)\].
Our aim is to consider two-dimensional data. We start with a choice of two straight lines $`V_{1}:{y = {\tau\hspace{0pt}x}}`$ and $`V_{2}:{y = {\tau'\hspace{0pt}x}}`$ on a square lattice $`{\mathbb{Z}}_{2}`$ (see Fig. [2](#S3.F2 "Figure 2 ‣ 3.1 Theoretical background ‣ 3 Mathematical ingredients")),
[Refer to caption](/html/0710.0021/assets/x2.png)
Figure 2: Construction of one-dimensional quasicrystal
where $`\tau`$ and $`\tau'`$ are two irrational numbers to be specified later. The simplest choice is $`\tau = {\frac{1}{2}\hspace{0pt}{({1 + \sqrt{5}})}}`$ and $`\tau' = {\frac{1}{2}\hspace{0pt}{({1 - \sqrt{5}})}}`$. One of the lines, say $`V_{1}`$, plays the role of the quasicrystal space onto which a ‘cut’ of the lattice $`{\mathbb{Z}}_{2}`$ is projected. In the other direction $`V_{2}`$, we choose a finite interval $`\Omega`$. Point of the lattice $`{\mathbb{Z}}_{2}`$ becomes a point of the quasicrystal, after projection on $`V_{1}`$, provided its projection on $`V_{2}`$ falls within $`\Omega`$, i.e. it is within the ‘cut’.
We denote by $`{\mathbb{Z}}\hspace{0pt}{\lbrack\tau\rbrack}`$ all the real numbers of the form $`({a + {b\hspace{0pt}\tau}})`$, with integers $`a`$ and $`b`$.
The star map between a quasicrystal point $`({a + {b\hspace{0pt}\tau}})`$ and the point $`({a + {b\hspace{0pt}\tau'}})`$ of $`\Omega`$ is given by
<table>
<tbody>
<tr>
<td></td>
<td>⋆:ℤ[τ]\star\ \colon\quad\mathbb{Z}[\tau] ⟶⟶\longrightarrow ℤ[τ′]ℤdelimited-[]superscript𝜏′\mathbb{Z}[\tau^{\prime}], (a+bτ)𝑎𝑏𝜏(a+b\tau) ↦maps-to\mapsto (a+bτ′)𝑎𝑏superscript𝜏′(a+b\tau^{\prime}), a,b∈ℤ𝑎𝑏ℤa,b\in\mathbb{Z}.</td>
<td></td>
</tr>
</tbody>
</table>
There is an unlimited number of choices of the irrational pairs to use in our construction. For example, one may choose as $`\tau`$ and $`\tau'`$ the solutions of the two infinite series of the quadratic equations $`{x^{2} = {{m\hspace{0pt}x} + 1}},{m = {1,2,\ldots}}`$ and $`{x^{2} = {{m\hspace{0pt}x} - 1}},{m = {3,4,\ldots}}`$.
The one-dimensional quasicrystal or cut and project set $`\Lambda_{\tau}\hspace{0pt}{(\Omega)}`$ is described as follows,
######
Definition 1
.
Let $`\Omega`$ be a finite interval and $`\tau`$, $`\tau'`$ be irrational numbers, then
<table>
<tbody>
<tr>
<td></td>
<td>Λτ(Ω)={a+bτ∣a,b∈ℤ,a+bτ′∈Ω},subscriptΛ𝜏Ωconditional-set𝑎𝑏𝜏formulae-sequence𝑎𝑏ℤ𝑎𝑏superscript𝜏′Ω\displaystyle\Lambda_{\tau}(\Omega)=\{a+b\tau\mid a,b\in\mathbb{Z},\ a+b\tau^{\prime}\in\Omega\},</td>
<td></td>
</tr>
</tbody>
</table>
where $`\Omega`$ is called the acceptance window of $`\Lambda_{\tau}\hspace{0pt}{(\Omega)}`$.
Let us underline, that $`\Lambda_{\tau}\hspace{0pt}{(\Omega)}`$ defined here is an infinite, uniformly dense, and uniformly discrete point set, while its window $`\Omega`$ is a finite interval densely covered by the star map of the points of $`\Lambda_{\tau}\hspace{0pt}{(\Omega)}`$.
From the definition of $`\Lambda_{\tau}\hspace{0pt}{(\Omega)}`$ a number of properties of $`\Lambda_{\tau}\hspace{0pt}{(\Omega)}`$ can be shown \[[3](#bib.bib3), [10](#bib.bib10)\]. In particular, the aperiodicity, discontinuity of the star map and the existence of only two or three different distances between adjacent points of $`\Lambda_{\tau}\hspace{0pt}{(\Omega)}`$. More precisely, we recall the following theorems.
######
Theorem 1
.
Let $`\Omega`$ be a semi-closed interval. For every $`\Lambda_{\tau}\hspace{0pt}{(\Omega)}`$ there exist positive numbers $`{\Delta_{1},\Delta_{2}} \in {{\mathbb{Z}}\hspace{0pt}{\lbrack\tau\rbrack}}`$, such that the distances between adjacent points in $`\Lambda_{\tau}\hspace{0pt}{(\Omega)}`$ take values in $`\{\Delta_{1},\Delta_{2},{\Delta_{1} + \Delta_{2}}\}`$. The distances depend only on $`\tau,\tau'`$ and length of the acceptance window.
More can be said about the distances between adjacent points, namely
<table>
<tbody>
<tr>
<td></td>
<td>{xn+1−xn|n∈ℤ}={{Δ1,Δ2,Δ1+Δ2},whenΔ1⋆−Δ2⋆>|Ω|{Δ1,Δ2},whenΔ1⋆−Δ2⋆=|Ω|.conditional-setsubscript𝑥𝑛1subscript𝑥𝑛𝑛ℤcasessubscriptΔ1subscriptΔ2subscriptΔ1subscriptΔ2whensuperscriptsubscriptΔ1⋆superscriptsubscriptΔ2⋆ΩsubscriptΔ1subscriptΔ2whensuperscriptsubscriptΔ1⋆superscriptsubscriptΔ2⋆Ω\displaystyle\{x_{n+1}-x_{n}|n\in\mathbb{Z}\}=\left\{\begin{array}[]{l}\{\Delta_{1},\ \Delta_{2},\ \Delta_{1}+\Delta_{2}\},\\ \text{when}\quad\Delta_{1}^{\star}-\Delta_{2}^{\star}>|\Omega|\\[4.30554pt] \{\Delta_{1},\ \Delta_{2}\},\\ \text{when}\quad\Delta_{1}^{\star}-\Delta_{2}^{\star}=|\Omega|.\end{array}\right.</td>
<td></td>
</tr>
</tbody>
</table>
One has a general scaling property of quasicrystals:
######
Theorem 2
.
For every irrational numbers $`\tau,\tau'`$, $`\tau \neq \tau'`$ and a bounded interval $`\Omega`$, there exist $`{\overset{\sim}{\tau}}' \in {({- 1},0)}`$, $`\overset{\sim}{\tau} > 0`$, $`s \in {\mathbb{R}}`$ and $`\overset{\sim}{\Omega}`$ (satisfying $`{\max{\{{1 + {\overset{\sim}{\tau}}'},{- {\overset{\sim}{\tau}}'}\}}} < {|\overset{\sim}{\Omega}|} \leq 1`$), such that
<table>
<tbody>
<tr>
<td></td>
<td>Λτ(Ω)=sΛτ~(Ω~).subscriptΛ𝜏Ω𝑠subscriptΛ~𝜏~Ω\displaystyle\Lambda_{\tau}(\Omega)=s\Lambda_{\tilde{\tau}}(\tilde{\Omega}).</td>
<td></td>
</tr>
</tbody>
</table>
Moreover, the distances between adjacent points in $`\Lambda_{\overset{\sim}{\tau}}\hspace{0pt}{(\overset{\sim}{\Omega})}`$ are equal to $`\overset{\sim}{\tau}`$, $`1 + \overset{\sim}{\tau}`$, $`1 + {2\hspace{0pt}\overset{\sim}{\tau}}`$, when $`{|\overset{\sim}{\Omega}|} \neq 1`$ and to $`\overset{\sim}{\tau}`$, $`1 + \overset{\sim}{\tau}`$, when $`{|\overset{\sim}{\Omega}|} = 1`$.
Finally we underline that the 1-dimensional quasicrystal $`\Lambda_{\tau}\hspace{0pt}{(\Omega)}`$ is an infinite discrete aperiodic set of points uniquely determined by the choice of $`\tau`$, $`\tau'`$ and of acceptance window $`\Omega = {\lbrack c,{c + d})}`$, where $`c`$ and $`d`$ are any real numbers. Due to the theorem 2, without loss of generality, we can consider $`\tau' \in {({- 1},0)}`$, $`\tau > 0`$, $`c = 0`$ and $`d \in {({\max{\{{1 + \tau'},{- \tau'}\}}},1\rbrack}`$.
In higher dimensions, it is adequate for our purposes to take for the quasicrystals a straightforward concatenation of one-dimensional ones in pairwise orthogonal directions.
### 3.2 Computational aspects
First let us introduce two specific examples of one-dimensional quasicrystals $`\Lambda_{\tau}\hspace{0pt}{(\Omega)}`$.
######
Example 1
.
Take $`\tau = \frac{1 + \sqrt{5}}{2}`$, $`\tau' = \frac{1 - \sqrt{5}}{2}`$, the two roots of the quadratic equation $`x^{2} = {x + 1}`$. Then $`{\tau + \tau'} = 1`$, $`{\tau\hspace{0pt}\tau'} = {- 1}`$. Choosing $`\Omega = {\lbrack 0,d)}`$, $`d = \frac{\tau}{2}`$ the conditions of Theorem 2 are satisfied: $`\tau' \approx {- 0.62} \in {({- 1},0)}`$ and $`\tau \approx 1.62 > 0`$.
Due to Theorem 2, the distances between adjacent points of the quasicrystal $`\Lambda_{\tau}\hspace{0pt}{({\lbrack 0,d)})}`$ are $`\Delta_{1} = \tau`$, $`\Delta_{2} = {1 + \tau}`$ and $`\Delta_{3} = {1 + {2\hspace{0pt}\tau}}`$.
As the seed point of the quasicrystal $`\Lambda_{\tau}\hspace{0pt}{({\lbrack 0,d)})}`$, we can choose any point $`x = {a + {b\hspace{0pt}\tau}}`$, such that $`x' = {a + {b\hspace{0pt}\tau'}} \in {\lbrack 0,d)}`$. For our example we put $`x = 0`$.
Unlike Definition 1, which is not constructive, here we have the information for fast generation of quasicrystal points. We can move right or left from the seed point by adding or subtracting one-by-one of the distances $`\Delta_{1}`$, $`\Delta_{2}`$ or $`\Delta_{3}`$.
Suppose we have already established, that $`x`$ is a point of $`\Lambda_{\tau}\hspace{0pt}{({\lbrack 0,d)})}`$, i.e. $`x' \in {\lbrack 0,d)}`$. The point adjacent to $`x`$ from the right is one of the three
<table>
<tbody>
<tr>
<td></td>
<td>x+τ,x+1+τorx+1+2τ.𝑥𝜏𝑥1𝜏or𝑥12𝜏\displaystyle x+\tau,\quad x+1+\tau\quad\text{or}\quad x+1+2\tau.</td>
<td></td>
</tr>
</tbody>
</table>
In order to decide which one is the case, we verify one-by-one the inclusions $`{x + \tau'} \in {\lbrack 0,d)}`$, $`{x + 1 + \tau'} \in {\lbrack 0,d)}`$ or $`{x + 1 + {2\hspace{0pt}\tau'}} \in {\lbrack 0,d)}`$. The first confirmed inclusion determines the new quasicrystal point.
In such a way the following two subsets of the quasicrystal $`\Lambda_{\tau}\hspace{0pt}{({\lbrack 0,d)})}`$ were obtained
<table>
<tbody>
<tr>
<td></td>
<td>{…,−1−2τ,−τ, 0, 1+τ, 2+2τ, 2+3τ, 3+4τ,…}↔︎↔︎…12𝜏𝜏 01𝜏22𝜏23𝜏34𝜏…absent\displaystyle\{\ldots,-1\!-\!2\tau,\ -\tau,\ 0,\ 1\!+\!\tau,\ 2\!+\!2\tau,\ 2\!+\!3\tau,\ 3\!+\!4\tau,\ldots\}\leftrightarrow</td>
<td></td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td>↔︎{…,Δ2,Δ1,Δ2,Δ2,Δ1,Δ2,…};↔︎absent…subscriptΔ2subscriptΔ1subscriptΔ2subscriptΔ2subscriptΔ1subscriptΔ2…\displaystyle\qquad\qquad\qquad\leftrightarrow\{\ldots,\ \Delta_{2},\ \Delta_{1},\ \Delta_{2},\ \Delta_{2},\ \Delta_{1},\ \Delta_{2},\ \ldots\};</td>
<td></td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td>{…,−τ, 0, 1+τ, 2+3τ, 3+4τ,…}↔︎↔︎…𝜏 01𝜏23𝜏34𝜏…absent\displaystyle\!\{\ldots,-\tau,\ 0,\ 1+\tau,\ 2+3\tau,\ 3+4\tau,\ldots\}\!\leftrightarrow</td>
<td></td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td>↔︎{…,Δ1,Δ2,Δ3,Δ2,…}.↔︎absent…subscriptΔ1subscriptΔ2subscriptΔ3subscriptΔ2…\displaystyle\qquad\qquad\qquad\leftrightarrow\{\ldots,\Delta_{1},\ \Delta_{2},\ \Delta_{3},\ \Delta_{2},\ldots\}.</td>
<td></td>
</tr>
</tbody>
</table>
Here the second line of each example contains the corresponding sequence of distances between quasicrystal points. The two quasicrystals differ by the sequence in which the three distances $`\Delta_{1}`$, $`\Delta_{2}`$ and $`\Delta_{3}`$ were tried in the construction.
The first case is was shown as the second line on Fig. [1](#S2.F1 "Figure 1 ‣ 2 Preliminaries and encryption procedure").
######
Example 2
.
Next quasicrystal $`\Lambda_{\tau}\hspace{0pt}{(\Omega)}`$ is built using the irrationalities $`\tau = {1 + \sqrt{2}}`$, $`\tau' = {1 - \sqrt{2}}`$, roots of the quadratic equation $`x^{2} = {{2\hspace{0pt}x} + 1}`$.
The conditions of Theorem 2 are again satisfied:$`\tau' \approx {- 0.42} \in {({- 1},0)}`$ and $`\tau \approx 2.42 > 0`$. Putting $`\Omega = {\lbrack 0,1)}`$, there are only two distances $`\Delta_{1} = \tau`$ and $`\Delta_{2} = {1 + \tau}`$ between adjacent points of the quasicrystal $`\Lambda_{\tau}\hspace{0pt}{({\lbrack 0,1)})}`$.
Using the zero point as the seed point, we obtain the cut and project set
<table>
<tbody>
<tr>
<td></td>
<td>{…,−1−3τ,−2τ,−τ,0,1+τ,1+2τ,2+3τ,2+4τ,3+5τ,…}…13𝜏2𝜏𝜏01𝜏12𝜏23𝜏24𝜏35𝜏…\displaystyle\{\ldots,\!-\!1\!-\!3\tau,\!-\!2\tau,\!-\!\tau,0,1\!+\!\tau,1\!+\!2\tau,2\!+\!3\tau,2\!+\!4\tau,3\!+\!5\tau,\ldots\}</td>
<td></td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td>↔︎{…,Δ2,Δ1,Δ1,Δ2,Δ1,Δ2,Δ1,Δ2,…}.↔︎absent…subscriptΔ2subscriptΔ1subscriptΔ1subscriptΔ2subscriptΔ1subscriptΔ2subscriptΔ1subscriptΔ2…\displaystyle\qquad\leftrightarrow\{\ldots,\ \Delta_{2},\ \Delta_{1},\ \Delta_{1},\ \Delta_{2},\ \Delta_{1},\ \Delta_{2},\ \Delta_{1},\ \Delta_{2},\ \ldots\}.</td>
<td></td>
</tr>
</tbody>
</table>
It may happen that some point $`x'`$ is arbitrary close to an end point of $`\Omega`$. Then it would be a time demanding computational task to decide whether $`x' \in \Omega`$ or $`x' \notin \Omega`$. Such a difficulty is simply avoided by disqualifying from our consideration any point $`x'`$ which comes closer to the boundary of $`\Omega`$ than the distance $`\varepsilon`$ agreed in advance.
The real numbers of the form $`a + {b\hspace{0pt}\tau}`$ can be understood as given by two integer components $`(a,b)`$. Since our operations are only addition or subtraction, the arithmetics of such numbers is elementary introduced. Consequently all transformations are performed with absolute precision.
The transformations between the original and encrypted data go through the several stages. At each stage we need only the previous point. Therefore we have only minimal requirements to RAM.
## 4 Application to image coding
In this section fragments of quasicrystal point sets are used for encryption/decryption of 2-dimensional digital data.
Assuming that the data is given on the rectangular grid containing final number $`n_{1}`$, $`n_{2}`$ of points in each direction, we deal with final fragments $`\Lambda^{(i)}\hspace{0pt}{(\Omega_{i})}`$, $`i = {1,2}`$ of the quasicrystal $`{\Lambda\hspace{0pt}{(\Omega)}} = {{{\Lambda^{(1)}\hspace{0pt}{(\Omega_{1})}} \otimes \Lambda^{(2)}}\hspace{0pt}{(\Omega_{2})}}`$. Number of points in each fragment is given by digital data size $`n_{i} = {|{\Lambda^{(i)}\hspace{0pt}{(\Omega_{i})}}|}`$.
Each $`\Omega_{i}`$ is an interval on real axis in mutually orthogonal directions. They can be chosen for both fragments independently.
Our main tool is the one-to-one map (star map) between the finite point sets $`\Lambda_{i}\hspace{0pt}{(\Omega_{i})}`$ and its image in $`\Omega_{i}`$
<table>
<tbody>
<tr>
<td></td>
<td>Λi(Ωi)⟷Ωi,i=1,2.formulae-sequencesubscriptΛ𝑖subscriptΩ𝑖⟷subscriptΩ𝑖𝑖12\displaystyle\Lambda_{i}(\Omega_{i})\quad\longleftrightarrow\quad\Omega_{i},\qquad i=1,2\,.</td>
<td></td>
</tr>
</tbody>
</table>
Note, that roles of $`\Lambda_{i}\hspace{0pt}{(\Omega_{i})}`$ and its image in $`\Omega_{i}`$ can be interchanged, because the fragments $`\Lambda_{i}\hspace{0pt}{(\Omega_{i})}`$ are finite sets of points.
Fast generation of $`\Lambda_{i}\hspace{0pt}{(\Omega_{i})}`$ was described in previous section. It requires that one is given the values of $`n_{1}`$, $`n_{2}`$, $`\tau_{i}`$, $`\tau_{i}'`$ and $`\Omega_{i}`$, and two seed points of $`\Lambda_{i}\hspace{0pt}{(\Omega_{i})}`$, $`i = {1,2}`$.
Suppose one is given a bitmap picture $`P\hspace{0pt}{(L)}`$, sampled on the points $`(x,y)`$ of a fragment $`L`$ of a rectangular lattice.
Then we treat independently each coordinate of the data using one fragment of the quasicrystal as illustrated in Fig. [1](#S2.F1 "Figure 1 ‣ 2 Preliminaries and encryption procedure").
### 4.1 Basic encryption algorithm
- •
Construction of two different 1-dimensional quasicrystals $`\Lambda^{(1)}`$ and $`\Lambda^{(2)}`$ of sizes $`n_{1}`$ and $`n_{2}`$ appropriate for the given data $`P\hspace{0pt}{(L)}`$. (Each fragment, as well as its acceptance window–interval can be situated anywhere on the corresponding real axes.)
- •
Matching one-by-one the sequence of coordinates $`x`$ with points of $`\Lambda^{(1)}`$, similarly matching the coordinates $`y`$ with points of $`\Lambda^{(2)}`$. (In this way $`x`$ becomes point $`x_{1} + {x_{2}\hspace{0pt}\tau}`$ of the $`\Lambda^{(1)}`$ and $`y`$ becomes point $`y_{1} + {y_{2}\hspace{0pt}\tau}`$ of the $`\Lambda^{(2)}`$, where $`x_{1}`$, $`x_{2}`$, $`y_{1}`$, $`y_{2}`$ are integers.)
- •
The main step of the encoding is the star map of $`\Lambda^{(1)} \otimes \Lambda^{(2)}`$ into the corresponding window $`\Omega_{1} \otimes \Omega_{2}`$
<table>
<tbody>
<tr>
<td></td>
<td>Λ(1)(Ω1)⟶Ω1⟶superscriptΛ1subscriptΩ1subscriptΩ1\Lambda^{(1)}(\Omega_{1})\longrightarrow\Omega_{1}, Λ(2)(Ω2)⟶Ω2⟶superscriptΛ2subscriptΩ2subscriptΩ2\Lambda^{(2)}(\Omega_{2})\longrightarrow\Omega_{2}; x1+x2τ1↦x1+x2τ1′maps-tosubscript𝑥1subscript𝑥2subscript𝜏1subscript𝑥1subscript𝑥2superscriptsubscript𝜏1′x_{1}+x_{2}\tau_{1}\mapsto x_{1}+x_{2}\tau_{1}^{\prime}, y1+y2τ2↦y1+y2τ2′maps-tosubscript𝑦1subscript𝑦2subscript𝜏2subscript𝑦1subscript𝑦2superscriptsubscript𝜏2′y_{1}+y_{2}\tau_{2}\mapsto y_{1}+y_{2}\tau_{2}^{\prime}.</td>
<td></td>
</tr>
</tbody>
</table>
(The sequence of the original data coordinates is dramatically changed due to the discontinuity of the star map.)
- •
It remains to map the points of encrypted quasicrystal fragments, which are now in $`\Omega_{1}`$ and $`\Omega_{2}`$, to a sequence of integers along $`x`$ and $`y`$ directions.
The present algorithm permutes the data points without changing the value of the data at each point. It is straightforward to consider the possible values of the data function as another discrete point set and to apply the same algorithm to it. Effectively we are thus dealing with a three-dimensional problem, where the data function takes only two values at 3-dimensional points.
All the maps we have used in the encryption algorithm are one-to-one, therefore they can be used in the inverse order for the decryption. For the same reason, nothing prevent us from iterative using our algorithm.
In certain types of data the encryption algorithm leaves residue of the orthogonal directions, see Fig. [3](#S4.F3 "Figure 3 ‣ 4.1 Basic encryption algorithm ‣ 4 Application to image coding")b. It can be avoided in a many different ways. One of them was mentioned above, when the problem was viewed as a three-dimensional one. Another possibility is illustrated on Fig. [3](#S4.F3 "Figure 3 ‣ 4.1 Basic encryption algorithm ‣ 4 Application to image coding")c. The two-dimensional data is taken as an ordered one-dimensional sequence (modification 2). The particularly fast possibility is to introduce the additional cyclic permutations of the points (modification 1), see Fig. [3](#S4.F3 "Figure 3 ‣ 4.1 Basic encryption algorithm ‣ 4 Application to image coding")d.
[Refer to caption](/html/0710.0021/assets/x3.png)
Figure 3: Application of the algorithms proposed in Section [4](#S4 "4 Application to image coding"): a – initial image; b – main algorithm; c – second modification; d – first modification
The computation time of our algorithms is compared in Table 1. Computations were carried out by Pentium M processor 1.70 GHz with 1 Gb RAM. As the test image we took the 512x512-pixel 8 bit grayscale picture Lenna (see Figs. [4](#A2.F4 "Figure 4 ‣ Appendix B Application examples") and [5](#A2.F5 "Figure 5 ‣ Appendix B Application examples")). But the second modification was tested on 200x200-pixel 8 bit grayscale picture Lenna (Fig. 5). Note, that the program does not cut the images into blocks for processing.
Table 1. Program working time.
<table>
<tbody>
<tr>
<td></td>
<td>Main algorithm</td>
<td>Modification 1</td>
<td>Modification 2</td>
</tr>
<tr>
<td>time</td>
<td>3 seconds</td>
<td>6 seconds</td>
<td>192 seconds</td>
</tr>
</tbody>
</table>
Program realizing these algorithms can be freely downloaded from http://132.204.90.210/maryna/ and listing of the main coding algorithm is adduced in Appendix [A](#A1 "Appendix A Program listing").
## 5 Conclusions
In the paper one of the simplest versions of the quasicrystal-based symmetric cipher is described. Even in this form it appears adequate for many applications. Its complicated key is described below. It is based on strongly aperiodic sets, includes everywhere discontinuous ‘star map’, and it is absolutely stable, since it operates with integer numbers only. Similarly, damaging a few data points does not affect the rest of the encryption/decryption process.
Furthermore coordinates of the data can be divided into several non-overlapping intervals, then contents of all intervals could be encrypted in parallel.
### 5.1 Cryptovariables
Application of the algorithm requires that the following variables (cryptovariables of the private key) are fixed
<table>
<tbody>
<tr>
<td></td>
<td>τ_τ′_a_b_c_d_sml_X_ε_I,where𝜏_superscript𝜏′_𝑎_𝑏_𝑐_𝑑_𝑠𝑚𝑙_𝑋_𝜀_𝐼where\displaystyle\tau\_\tau^{\prime}\_a\_b\_c\_d\_sml\_X\_\varepsilon\_I,\quad\text{where}</td>
<td></td>
</tr>
</tbody>
</table>
$`\tau`$ and $`\tau'`$ are irrational numbers, solutions of one of equations of the two infinite series of the quadratic equations $`x^{2} = {{m\hspace{0pt}x} + 1}`$, $`m = {1,2,\ldots}`$ or $`x^{2} = {{m\hspace{0pt}x} - 1}`$, $`m = {3,4,\ldots}`$;
$`a`$ and $`b`$ are integers, fixing a seed point of the quasicrystal;
$`c`$ and $`d`$ are arbitrary real numbers, fixing the position of the acceptance window $`\Omega`$;
$`X`$ takes values {‘+’, ‘-’}, defining direction of quasicrystal construction relative to the seed point;
$`s\hspace{0pt}m\hspace{0pt}l`$ sequence of distances of adjacent points during quasicrystal generation;
$`\varepsilon`$ is the maximal distance from the end points of $`\Omega`$ at which the star map points are still considered;
$`I`$ is the positive integer number specifying number of iterations of the certain quasicrystal cipher.
More elaborate encryption system can be built by exploiting properties of 1 or 2-dimensional quasicrystals. Using the present algorithm in several iterations, one may choose new quasicrystal for each iteration, or new position of the quasicrystal fragment, or different size, position of the acceptance window, etc.
Rich scaling properties of the quasicrystal could be involved into encryption scheme \[[3](#bib.bib3)\].
Genuinely two-dimensional quasicrystals can be used for the encryption, some examples of the possibilities are found in \[[12](#bib.bib12)\]. The number of distinct quasicrystals of this kind is unlimited.
One can easily build one-dimensional quasicrystal by projecting cuts of n-dimensional latices on the direction of the quasicrystal.
Intriguing appears to be the possibility to read the initial lattice data as an ordered stream of points and map them into single one-dimensional quasicrystal, perform the star map and reorder transform points into encrypted data. There are also other applications of the quasicrystals e.g. for generation of random numbers \[[6](#bib.bib6)\], even some cryptographic ones \[[11](#bib.bib11)\].
## Acknowledgment
One of us (M.N.) is grateful for the hospitality extended to her at the Center de recherche mathématique, Université de Montréal during the work on this project.
We are grateful for partial support of the work by the Natural Science and Engineering Research Council of Canada, MITACS, Lockheet Martin of Canada and MIND Research Institute of California.
## References
- \[1\]
L. Chen, R. Moody and J. Patera, Non-crystallographic root systems. Quasicrystals and discrete geometry (Toronto, ON, 1995), 135–178, Fields Inst. Monogr., 10, Amer. Math. Soc., Providence, RI, 1998.
- \[2\]
W. Diffie and M. Hellman, New Directions in Cryptography, IEEE Transactions on Information Theory, IT-22 (1976), 644–654.
- \[3\]
J.P. Gazeau, Z. Masakova and E. Pelantova, Nested quasicrystalline discretisations of the line, Physics and number theory 10 (2006), 79–131.
- \[4\]
O. Goldreich, Foundations of Cryptography, V.1: Basic Tools, Cambridge University Press, 2001.
- \[5\]
L.S. Guimond, Z. Masakova and E. Pelantova, Combinatorial properties of infinite words associated with cut-and-project sequences, J. Theor. Nombres Bordeaux 15 (2003), 697–725.
- \[6\]
L.S. Guimond, Jan Patera and Jiri Patera, Statistics and implementation of aperiodic pseudorandom number generators, Applied Numerical Mathematics 46 (2003), 295–318.
- \[7\]
Z. Masakova, J. Patera and E. Pelantova, Inflation centres of the cut and project quasicrystals, J. Phys. A 31 (1998), 1443–1453.
- \[8\]
Z. Masakova, J. Patera and E. Pelantova, Minimal distances in quasicrystals, J. Phys. A 31 (1998), 1539–1552.
- \[9\]
A.J. Menezes, P.C. van Oorschot and S.A. Vanstone, Handbook of Applied Cryptography, CRC Press, 2001.
- \[10\]
R. Moody and J. Patera, Quasicrystals and icosians, J. Phys. A 26 (1993), 2829–2853.
- \[11\]
Z. Masáková, J. Patera, E. Pelantová, Aperiodic Encryption Method, US patent, September 14, 2004, US 6,792,108 B1
- \[12\]
J. Patera and M. Nesterenko, Quasicrystals in cryptography, E. Kranakis, E. Haroutunian and E. Shahbazian, editors, Aspects of Network and Information Security, Proceedings of NATO Advanced Studies Institute on Network Security and Intrusion Detection, held in Nork, Yerevan, Armenia, October 01-12, 2006, IOS Press, 2007, to appear.
## Appendix A Program listing
Realization of the main algorithm on Delphi 6.0:
unit Unit1;
interface
uses
Windows, Messages, SysUtils, Variants, Classes, Graphics, Controls, Forms,
Dialogs, StdCtrls;
type
TNode = Record
a : integer;
b : integer;
rs, rb : real;
ns, nb : integer;
end;
TSequence = Record
n : integer;
Nodes : array of TNode;
a,b : integer;
s : string;
end;
TForm1 = class(TForm)
Button1 : TButton;
Button2: TButton;
procedure Button1Click(Sender: TObject);
procedure FormCreate(Sender: TObject);
procedure Button2Click(Sender: TObject);
private
Private declarations
public
Public declarations
end;
var
Form1 : TForm1;
RootS, RootB : real;
implementation
{\$R \*.dfm}
function CalcSmall (a, b: real) : real;
begin
CalcSmall := a + b \* RootS;
end;
function CalcBig (a, b: real) : real;
begin
a := a + 1;
b := - b;
CalcBig := CalcSmall (a, b);
CalcBig := a + b \* RootB;
end;
function CheckShort (a, b: integer; var Node : Tnode; dir: integer) : boolean;
var tempB : real;
res : boolean;
begin
a := a - dir;
b := b + dir;
tempB := CalcBig (a,b);
res := ((tempB $`<`$ 0) and (tempB $`<`$ 2));
If res then begin
node.a := a;
node.b := b;
node.rb := tempB;
end;
CheckShort := res;
end; //CheckShort (a, b: integer; var node : Tnode; dir: integer) : boolean;
function CheckMedium (a, b: integer; var node : Tnode; dir: integer) : boolean;
var tempB : real;
res : boolean;
begin
a := a + dir;
tempB := CalcBig (a,b);
res := ((tempB $`>`$ 0) and (tempB $`<`$ 2));
If res then begin
node.a := a;
node.b := b;
node.rb := tempB;
end;
CheckMedium := res;
end; //CheckMedium (a, b: integer; var node : Tnode; dir: integer) : boolean;
function CheckLong (a, b: integer; var node : Tnode; dir: integer) : boolean;
var tempB : real;
res : boolean;
begin
b := b + dir;
tempB := CalcBig (a,b);
res := ((tempB $`>`$ 0) and (tempB $`<`$ 2));
If res then begin
node.a := a;
node.b := b;
node.rb := tempB;
end;
CheckLong := res;
end; //CheckLong (a, b: integer; var node : Tnode; dir: integer) : boolean;
function FillSequence (var seq : TSequence; a, b, n, dir: integer): string;
var res : string;
aPrev, bPrev : integer;
i : integer;
begin
res := ”;
seq.n := n;
seq.a := a;
seq.b := b;
SetLength (seq.Nodes, n);
seq.Nodes\[0\].a := a;
seq.Nodes\[0\].b := b;
seq.Nodes\[0\].rb := CalcBig (a, b);
seq.Nodes\[0\].rs := CalcSmall (a, b);
seq.Nodes\[0\].ns := 0;
for i := 1 to n - 1 do begin
aPrev := seq.Nodes \[i-1\].a;
bPrev := seq.Nodes \[i-1\].b;
seq.Nodes \[i\].ns := i;
if dir = 1 then begin
if CheckShort (aPrev, bPrev, seq.Nodes \[i\], dir) then
res := res + ’s’
else if CheckMedium (aPrev, bPrev, seq.Nodes \[i\], dir) then
res := res + ’m’
else if CheckLong (aPrev, bPrev, seq.Nodes \[i\], dir) then
res := res + ’l’
end else if CheckLong (aPrev, bPrev, seq.Nodes \[i\], dir) then
res := res + ’l’
else if CheckMedium (aPrev, bPrev, seq.Nodes \[i\], dir) then
res := res + ’m’
else if CheckShort (aPrev, bPrev, seq.Nodes \[i\], dir) then
res := res + ’s’;
seq.Nodes\[i\].rs := CalcSmall (seq.Nodes\[i\].a, seq.Nodes\[i\].b);
end;
seq.s := res;
FillSequence := res;
end; //FillSequence (var seq : TSequence; a, b, n, dir: integer): string;
Procedure SaveSequence (seq : TSequence; fname : string);
var f : TextFile;
var i : integer;
begin
AssignFile (f, fname);
rewrite (f);
writeln (f, seq.n,’ : ’, seq.s);
for i:= 0 to seq.n - 1 do
writeln (f, seq.Nodes\[i\].ns, ’, ’, seq.Nodes\[i\].nb, ’, ’, seq.Nodes\[i\].a, ’, ’, seq.Nodes\[i\].b, ’, ’, seq.Nodes\[i\].rs, ’, ’, seq.Nodes\[i\].rb);
Flush (f);
CloseFile (f);
end; //SaveSequence (seq : TSequence; fname : string);
Procedure SortSequence (var seq : TSequence; dir: integer);
var node : TNode;
i,j : integer;
flag : integer;
begin
for i := seq.n-2 downto 0 do
for j:= 0 to i do begin
flag := 0;
if dir = 1 then begin
if seq.Nodes\[j\].rb $`>`$ seq.Nodes\[j + 1\].rb then
flag := 1;
end else
if seq.Nodes\[j\].rs $`>`$ seq.Nodes\[j + 1\].rs then
flag := 1;
if flag = 1 then begin
node := seq.Nodes\[j + 1\];
seq.Nodes\[j + 1\] := seq.Nodes\[j\];
seq.Nodes\[j\] := node;
end;
end;
if dir = 1 then
for i := 0 to seq.n do
seq.Nodes\[i\].nb := i;
end; //SorteSequence (var seq : TSequence; dir: integer)
Procedure CodeImg (dir: integer);
var fname1, fname2 : string;
Bitmap1, Bitmap2 : TBitmap;
SeqX, SeqY : TSequence;
i,j : integer;
CurWidth, CurHeight : integer;
strX, strY : string;
begin
if dir = 1 then begin
fname1 := ’in.bmp’;
fname2 := ’out.bmp’;
end else begin
fname1 := ’out.bmp’;
fname2 := ’back.bmp’;
end;
Bitmap1 := TBitmap.Create;
Bitmap1.LoadFromFile (fname1);
Bitmap2 := TBitmap.Create;
CurWidth := Bitmap1.Width;
CurHeight := Bitmap1.Height;
Bitmap2.Width := CurWidth;
Bitmap2.Height := CurHeight;
StrX := FillSequence (SeqX, 1, 0, CurWidth, 1);
StrY := FillSequence (SeqY, 1, 0, CurHeight, 1);
SortSequence (SeqX, 1);
SortSequence (SeqY, 1);
for i:= 0 to CurWidth - 1 do
for j:= 0 to CurHeight - 1 do
if dir = 1 then
Bitmap2.Canvas.Pixels \[i,j\] := Bitmap1.Canvas.Pixels \[SeqX.Nodes \[i\].ns, SeqY.Nodes \[j\].ns\]
else
Bitmap2.Canvas.Pixels \[SeqX.Nodes \[i\].ns,SeqY.Nodes \[j\].ns\] := Bitmap1.Canvas.Pixels \[i,j\];
Bitmap2.SaveToFile (fname2);
Bitmap1.Free;
Bitmap2.Free;
if dir = 1 then begin
SortSequence (SeqX, -1);
SortSequence (SeqY, -1);
SaveSequence (SeqX, ’x.txt’);
SaveSequence (SeqY, ’y.txt’);
end;
end; //CodeImg (dir: integer)
procedure TForm1.Button1Click(Sender: TObject);
begin
Button1.Enabled := False;
Button2.Enabled := False;
CodeImg (1);
Button1.Enabled := True;
Button2.Enabled := True;
end;
procedure TForm1.FormCreate(Sender: TObject);
begin
RootS := (1 + sqrt (5)) / 2;
RootB := (1 - sqrt (5)) / 2;
end;
procedure TForm1.Button2Click(Sender: TObject);
begin
Button1.Enabled := False;
Button2.Enabled := False;
CodeImg (-1);
Button1.Enabled := True;
Button2.Enabled := True;
end;
end.
## Appendix B Application examples
In this section we will show several examples of the application of the main algorithm, its modification 1, and modification 2, as well as its iterations several times. Let us emphasize that the data was encrypted as a single block. Computing time is ranging between two seconds (one iteration) and nine seconds (five iterations) for the main algorithm.
[Refer to caption](/html/0710.0021/assets/x4.png)
Figure 4: Main algorithm coding without iterations
[Refer to caption](/html/0710.0021/assets/x5.png)
Figure 5: Modification 1 of the algorithm without iterations
[Refer to caption](/html/0710.0021/assets/x6.png)
Figure 6: Modification 2 of the algorithm without iterations
[Refer to caption](/html/0710.0021/assets/x7.png)
Figure 7: Main algorithm coding without iterations
[Refer to caption](/html/0710.0021/assets/x8.png)
Figure 8: Modification 1 of the algorithm without iterations |
0708.3879 | cs.NI | cyber | # Graph Annotations in Modeling Complex Network Topologies
Xenofontas Dimitropoulos
Dmitri Krioukov
IBM Research, Zürich <xed@zurich.ibm.com> CAIDA/UCSD <dima@caida.org>
Amin Vahdat
George Riley
UCSD <vahdat@cs.ucsd.edu> Georgia Tech <riley@ece.gatech.edu>
###### Abstract
The coarsest approximation of the structure of a complex network, such as the Internet, is a simple undirected unweighted graph. This approximation, however, loses too much detail. In reality, objects represented by vertices and edges in such a graph possess some non-trivial internal structure that varies across and differentiates among distinct types of links or nodes. In this work, we abstract such additional information as network annotations. We introduce a network topology modeling framework that treats annotations as an extended correlation profile of a network. Assuming we have this profile measured for a given network, we present an algorithm to rescale it in order to construct networks of varying size that still reproduce the original measured annotation profile.
Using this methodology, we accurately capture the network properties essential for realistic simulations of network applications and protocols, or any other simulations involving complex network topologies, including modeling and simulation of network evolution. We apply our approach to the Autonomous System (AS) topology of the Internet annotated with business relationships between ASs. This topology captures the large-scale structure of the Internet. In depth understanding of this structure and tools to model it are cornerstones of research on future Internet architectures and designs. We find that our techniques are able to accurately capture the structure of annotation correlations within this topology, thus reproducing a number of its important properties in synthetically-generated random graphs.
###### category:
C.2.1 Network Architecture and Design Network topology
###### category:
C.2.5 Local and Wide-Area Networks Internet
###### category:
G.3 Probability and Statistics Distribution functions, multivariate statistics, correlation and regression analysis
###### category:
G.2.2 Graph Theory Network problems
###### keywords:
Annotations, AS relationships, complex networks, topology
<sup>†</sup>
<sup>†</sup>terms: Measurement, Design, Theory
## 1 Introduction
Simulations of new network protocols and architectures are pointless without realistic models of network structure and evolution. Performance of routing \[[26](#bib.bib26)\], multicast \[[34](#bib.bib34)\], and other protocols depends crucially on network topology. Simulations of these protocols with inaccurate topology models can thus result in misleading outcomes.
Inaccuracies associated with representing complex network topologies as simple undirected unweighted graphs come not only from potential sampling biases in topology measurements \[[27](#bib.bib27), [10](#bib.bib10), [11](#bib.bib11)\], but also from neglecting link and node annotations. By annotations we mean various types of links and nodes that abstract their intrinsic structural and functional differences to a certain degree. For example, consider the Internet topology at the Autonomous System (AS) level. Here, link annotations may represent different business relationship between ASs, e.g., customer-to-provider, peer-to-peer, etc. \[[13](#bib.bib13)\], while node annotations may represent different types of ASs, e.g., large or small Internet Service Providers (ISPs), exchange points, universities, customer enterprises, etc. \[[15](#bib.bib15)\]. In router-level Internet topologies, link annotations can be different transmission speeds, latencies, packet loss rates, etc. One can also differentiate between distinct types of links and nodes in other networks, such as social, biological, or transportation networks. In many cases, simply reproducing the structure of a given network is insufficient; we must also understand and reproduce domain-specific annotations.
We propose network annotations as a general framework to provide the next level of detail describing the “microscopic” structure of links and nodes. Clearly, since links and nodes are constituents of a global network, increasing description accuracy at the “microscopic” level will also increase overall accuracy at the “macroscopic” level as well. That is, including appropriate per-node or per-link annotations will allow us to capture and reproduce more accurately a variety of important global graph properties. In the AS topology case, for example, instead of considering only shortest paths, we will be able to study the structure of paths that respect constraints imposed by routing policies and AS business relationships.
Higher accuracy in approximating network structure is desirable not only for studying applications and protocols that depend on such structure, but also for modeling network evolution. For example, realistic Internet AS topology growth models should be based on economic realities of the Internet since AS links are nothing but reflections of AS contractual relationships, i.e., results of business decisions made by organizations that the corresponding ASs represent. Therefore, economy-based AS topology models naturally produce links annotated with AS relationships. AS relationship annotations are thus intrinsic to such models.
Network annotations should also be useful for researchers studying only those networks that preserve some domain-specific constraints, thus avoiding “too random” networks that violate these constraints. Examples of such “technological” constraints for router topologies include maximum node degree limits, specific relationships between node degree and centrality, etc. \[[28](#bib.bib28)\]. In this context, we note that any node or link attributes, including their degrees and centrality, are forms of annotations. Therefore, one can use the network annotation framework to introduce domain-specific or any other constraints to work with network topologies narrowed down to a specific class. We also note that the network annotation framework is sufficiently general to include directed and weighted networks as partial cases, since both link directions and weights are forms of annotations.
After reviewing, in Section [2](#S2 "2 Related work ‣ Graph Annotations in Modeling Complex Network Topologies"), past work on network topology modeling and generation, which largely ignores annotations, we make the following contributions in this paper:
- •
In Section [3](#S3 "3 AS relationships and why they matter ‣ Graph Annotations in Modeling Complex Network Topologies"), we demonstrate the importance of network annotations using the specific example of AS business relationships in the Internet.
- •
In Section [4](#S4 "4 Network topology annotations ‣ Graph Annotations in Modeling Complex Network Topologies"), we introduce a general network annotation formalism and apply it to the Internet AS topology annotated with AS business relationships.
- •
In Section [5](#S5 "5 Generating annotated AS graphs ‣ Graph Annotations in Modeling Complex Network Topologies"), we formulate a general methodology and specific algorithms to: i) rescale the annotation correlation profile of the observed AS topology to arbitrary network sizes; and ii) construct synthetic networks reproducing the rescaled annotation profiles. While we discuss our graph rescaling and construction techniques in the specific context of AS topologies, these techniques are generic and can be used for generating synthetic annotated networks that model other complex systems.
- •
In Section [6](#S6 "6 Evaluation ‣ Graph Annotations in Modeling Complex Network Topologies"), we evaluate the properties of the resulting synthetic AS topologies and show that they recreate the annotation correlations observed in real annotated AS topologies as well as other important properties directly related to common metrics used in simulation and performance evaluation studies.
We conclude by outlining some implications and directions for future work in Section [7](#S7 "7 Conclusions ‣ Graph Annotations in Modeling Complex Network Topologies").
## 2 Related work
A large number of works have focused on modeling Internet topologies and on developing realistic topology generators. Waxman \[[44](#bib.bib44)\] introduced the first topology generator that became widely known. The Waxman generator was based on the classical (Erdős-Rényi) random graph model \[[18](#bib.bib18)\]. After it became evident that observed networks have little in common with classical random graphs, new generators like GT-ITM \[[46](#bib.bib46)\] and Tiers \[[17](#bib.bib17)\] tried to mimic the perceived hierarchical network structure and were consequently called structural. In 1999, Faloutsos et al. \[[19](#bib.bib19)\] discovered that the degree distributions of router- and AS-level topologies of the Internet followed a power law. Structural generators failed to reproduce the observed power laws. This failure led to a number of subsequent works trying to resolve the problem.
The existing topology models capable of reproducing power laws can be roughly divided into the following two classes: causality-aware and causality-oblivious. The first class includes the Barabási-Albert (BA) \[[2](#bib.bib2)\] preferential attachment model, the Highly Optimized Tolerance (HOT) model \[[6](#bib.bib6)\], and their derivatives. The BRITE \[[32](#bib.bib32)\] topology generator belongs to this class, as it employs preferential attachment mechanisms to generate synthetic Internet topologies. The models in this class grow a network by incrementally adding nodes and links to a graph based on a formalized network evolution process. One can show that both BA and HOT growth mechanisms produce power laws.
On the other hand, the causality-oblivious approaches try to match a given (power-law) degree distribution without accounting for different forces that might have driven evolution of a network to its currently observed state. The models in this class include random graphs with given expected \[[9](#bib.bib9)\] and exact \[[1](#bib.bib1)\] degree sequences, Markov graph rewiring models \[[31](#bib.bib31), [23](#bib.bib23)\], and the Inet \[[45](#bib.bib45)\] topology generator. Recent work by Mahadevan et al. introduced the $`d\hspace{0pt}K`$-series \[[29](#bib.bib29)\] extending this class of models to account for node degree correlations of arbitrary order. Whereas the $`d\hspace{0pt}K`$-series provides a set of increasingly accurate descriptions of network topologies represented as graphs, network annotations are another, independent and “orthogonal” to $`d\hspace{0pt}K`$-series, way to provide more accurate and complete information about actual complex systems that these graphs represent.
Frank and Strauss first formally introduced the annotated (colored) random Markov graphs in \[[20](#bib.bib20)\]. In their definition, every edge is colored by one of $`T`$ colors. More recently, Söderberg suggested a slightly different definition \[[42](#bib.bib42)\], where every half-edge, i.e., stub, is colored by one of $`T`$ colors. Every edge is thus characterized by a pair of colors. This definition is very generic. It includes uncolored and standard colored \[[20](#bib.bib20)\] random graphs, random vertex-colored graphs \[[40](#bib.bib40)\]
<sup>1</sup>
<sup>1</sup>1Random graphs with colored nodes are a partial case of random graphs with hidden variables \[[3](#bib.bib3)\].
, and random directed graphs \[[5](#bib.bib5)\] as partial cases. Söderberg considers some analytic properties of the ensemble of these random colored graphs in \[[41](#bib.bib41)\]. In \[[43](#bib.bib43)\], he observes strong similarities between random graphs colored by $`T`$ colors and random Feynman graphs representing a perturbative description of a $`T`$-dimensional system from quantum or statistical mechanics.
Recent works on annotation techniques specific to AS graphs include \[[12](#bib.bib12)\] and \[[8](#bib.bib8)\]. The GHITLE \[[12](#bib.bib12)\] topology generator produces AS topologies with c2p and p2p annotations based on simple design heuristics and user-controlled parameters. The work by Chang et al. \[[8](#bib.bib8)\] describes a topology evolution framework that models ASs’ decision criteria in establishing c2p and p2p relationships. Our methodology is different in that it explores the orthogonal, causality-oblivious approach to modeling link annotations. Its main advantage is that it is applicable to modeling any type of complex networks.
## 3 AS relationships and why they matter
In this section, we introduce our specific example of network annotations—AS relationships. We first describe what AS relationships represent and then discuss the results of simple simulation experiments showing why preserving AS relationship information is important.
AS relationships are annotations of links of the Internet AS-level topology. They represent business agreements between pairs of AS neighbors. There are three major types of AS relationships: 1) customer-to-provider (c2p), connecting customer and provider ASs; 2) peer-to-peer (p2p), connecting two peer ASs; and 3) sibling-to-sibling (s2s), connecting two sibling ASs. This classification stems from the following BGP route export policies, dictated by business agreements between ASs:
- •
exporting routes to a provider or a peer, an AS advertises its local routes and routes received from its customer ASs only;
- •
exporting routes to a customer or a sibling, an AS advertises all its routes, i.e., its local routes and routes received from all its AS neighbors.
Even though there are only two distinct export policies, they lead to the three different AS relationship types when combined in an asymmetric (c2p) or symmetric (p2p or s2s) manner.
If all ASs strictly adhere to these export policies, then one can easily check \[[21](#bib.bib21)\] that every AS path must be of the following valley-free or valid pattern: zero or more c2p links, followed by zero or one p2p links, followed by zero or more p2c links, where by ‘p2c’ links we mean c2p links in the direction from the provider to the customer.
Routing policies reflect business agreements and economic incentives. For this reason, they are deemed more important than quality of service and other criteria. As a result, suboptimal routing and inflated AS paths often occur. Gao and Wang \[[22](#bib.bib22)\] used BGP data to measure the extent of AS path inflation in the Internet. They found that at least 45% of the AS paths observed in BGP data are inflated by at least one AS hop and that AS paths can be inflated by as long as 9 AS hops.
[Refer to caption](/html/0708.3879/assets/x1.png)
Figure 1: Example AS topology annotated with AS relationships. The dotted lines represent shortest paths between ASs 4, 6 and 8 to AS 2. The dashed lines represent policy compliant paths from the same sources to the same destination.
Taking into account such inflation effects is important for meaningful and realistic simulation studies. For example, consider the AS topology in Figure [1](#S3.F1 "Figure 1 ‣ 3 AS relationships and why they matter ‣ Graph Annotations in Modeling Complex Network Topologies"), which is a small part of the real (measured) AS topology annotated with AS relationships inferred using heuristics in \[[14](#bib.bib14)\]. Directed links represent c2p relationships that point towards the provider and undirected links represent p2p relationships. If we ignore AS relationships then the shortest paths from ASs 4, 6, and 8 to AS 2 are shown with dotted lines. On the other hand, if we account for AS relationships these paths are no longer valid. In particular, the path 4$`\rightarrow`$3$`\rightarrow`$2 transverses two p2p links; the path 6$`\rightarrow`$3$`\rightarrow`$2 transverses a p2c link followed by a p2p link; and the path 8$`\rightarrow`$1$`\rightarrow`$2 transverses a c2p link after having gone through a p2c link. As all these paths are not valid, they are not used in practice. The paths actually used are the policy compliant paths marked with dashed lines.
In other words, the first effect of taking AS relationships into account is that paths become longer than the corresponding shortest paths. From a performance perspective, longer paths can affect metrics such as end-to-end (e2e) delay, server response time, jitter, convergence time, and others.
To illustrate this effect, we simulated the topology in Figure [1](#S3.F1 "Figure 1 ‣ 3 AS relationships and why they matter ‣ Graph Annotations in Modeling Complex Network Topologies") using BGP++ \[[16](#bib.bib16)\]. We used a single router per AS and configured appropriate export rules between ASs according to the guidelines discussed above. We set the delay of each link to 10 milliseconds and the bandwidth to 400kbps. Then, we configured exponential on/off traffic sources at ASs 4, 6 and 8 that send traffic to AS 2 at a rate of 500kbps. We run the simulation for 120 seconds; for the first 100 seconds we waited for routers to converge
<sup>2</sup>
<sup>2</sup>2Typically routers take much less than 100 seconds to converge, but to be conservative we used a longer period.
and at the 100th second we started the traffic sources. We then measured the e2e delay between the sources and the destination with AS relationships disabled and enabled.
In Figure [2](#S3.F2 "Figure 2 ‣ 3 AS relationships and why they matter ‣ Graph Annotations in Modeling Complex Network Topologies") we depict the cumulative distribution function (CDF) of the e2e delays for the both cases. We first notice that the CDF with AS relationships enabled shifts to the right, which means that there is a significant increase in the e2e delay. In particular, the average e2e delay with AS relationships enabled is 0.853 seconds, whereas without AS relationships it drops to 0.389 seconds. Besides the decrease in the e2e delay, we see that the CDF with AS relationships is much smoother than the other CDF, which exhibits a step-wise increase. The reason for that difference is that in the former case we have more flows sharing multiple queues and, consequently, more diverse queue dynamics, while in the latter case, almost all paths are disjoint, leading to mostly fixed e2e delays. The observed difference signifies that the e2e delay with AS relationships enabled exhibits a much higher variability compared to the case with AS relationships ignored. This difference in variability is likely to affect other performance metrics like jitter and router buffer occupancy.
[Refer to caption](/html/0708.3879/assets/x2.png)
Figure 2: CDF of e2e delay between traffic sources and destination.
Table 1: Total number of paths for each AS with AS relationships enabled and AS relationships disabled.
<table>
<thead>
<tr>
<th colspan="2">AS number</th>
<th>1</th>
<th>2</th>
<th>3</th>
<th>4</th>
<th>5</th>
<th>6</th>
<th>7</th>
<th>8</th>
</tr>
</thead>
<tbody>
<tr>
<td rowspan="2">Number of paths</td>
<td>AS relationships disabled</td>
<td>12</td>
<td>13</td>
<td>16</td>
<td>15</td>
<td>13</td>
<td>15</td>
<td>15</td>
<td>13</td>
</tr>
<tr>
<td>AS relationships enabled</td>
<td>12</td>
<td>9</td>
<td>10</td>
<td>8</td>
<td>8</td>
<td>7</td>
<td>9</td>
<td>6</td>
</tr>
</tbody>
</table>
Another consequence of policy-constrained routing is that ASs have fewer alternative AS paths. For example, in Figure [1](#S3.F1 "Figure 1 ‣ 3 AS relationships and why they matter ‣ Graph Annotations in Modeling Complex Network Topologies") when ignoring AS relationships AS 7 has three (one through each neighbor) disjoint paths to reach destination 2. One the other hand, with AS relationships enabled, AS 7 has only one possible path through AS 5, since the other two paths are not valid. In Table [1](#S3.T1 "Table 1 ‣ 3 AS relationships and why they matter ‣ Graph Annotations in Modeling Complex Network Topologies"), we show the total number of paths we found in the BGP tables of the eight ASs in our simulations. The consistent decrease in the number of paths when AS relationships are enabled highlights that ignoring AS relationships increases the path diversity of the ASs in a simulation. Path diversity is an important property related to network robustness, vulnerability to attacks, links and router failures, load balancing, multi-path routing, convergence of routing protocols, and others.
Yet another effect of policy routing is different distribution of load on ASs and AS links. Indeed, due to the smaller number of available AS paths, compared to shortest path routing, some nodes and links are likely to experience greater traffic load. For example, in Figure [1](#S3.F1 "Figure 1 ‣ 3 AS relationships and why they matter ‣ Graph Annotations in Modeling Complex Network Topologies") the dashed paths share the links from AS 7 to AS 5. On the other hand, when assuming shortest path routing the three paths are mostly disjoint: only one link, the link between AS 3 and AS 2, is shared by two flows. Thus, AS links and nodes will receive greater load, compared to the case with AS relationships ignored. Higher load is likely to produce more packet loss, increased delay, congestion, router failures, and other undesirable effects. In Table [2](#S3.T2 "Table 2 ‣ 3 AS relationships and why they matter ‣ Graph Annotations in Modeling Complex Network Topologies") we list the average bandwidth in our simulations for each of the three flows with and without AS relationships enabled. We find that because of the increased load on the links between AS 7 and AS 5 the average bandwidth of the three flows decreases substantially.
Table 2: Average bandwidth per flow with AS relationships enabled or disabled.
<table>
<thead>
<tr>
<th colspan="2">Flow</th>
<th>4 →→\rightarrow 2</th>
<th>6 →→\rightarrow 2</th>
<th>8 →→\rightarrow 2</th>
</tr>
</thead>
<tbody>
<tr>
<td rowspan="2">Bandwidth (Kbps)</td>
<td>AS relationships disabled</td>
<td>202</td>
<td>196</td>
<td>397</td>
</tr>
<tr>
<td>AS relationships enabled</td>
<td>113</td>
<td>164</td>
<td>121</td>
</tr>
</tbody>
</table>
To summarize this section, we have provided three examples showing that ignoring AS relationship annotations leads to inaccuracies, which make the corresponding properties look “better” than they are in reality. Indeed, if AS relationships are ignored, then:
- •
paths are shorter than in reality;
- •
path diversity is larger than in reality; and
- •
traffic load is lower than in reality.
## 4 Network topology annotations
In this section we first introduce our general formalism to annotate network topologies. We then show how this formalism applies to our example of the AS-level Internet topology annotated with AS relationships.
### 4.1 General formalism
Our general formalism is close to random colored graph definitions from \[[42](#bib.bib42)\] and borrows parts of the convenient $`d\hspace{0pt}K`$-series terminology from \[[29](#bib.bib29)\].
We define the annotated network as a graph $`G\hspace{0pt}{(V,E)}`$, $`{|V|} = n`$ and $`{|E|} = m`$, such that all $`2\hspace{0pt}m`$ edge-ends (stubs) of all $`m`$ edges in $`E`$ are of one of several colors $`c`$, $`c = {1\hspace{0pt}\ldots\hspace{0pt}C}`$, where $`C`$ is the total number of stub colors. We also allow for node annotations by an independent set of node colors $`\theta = {1\hspace{0pt}\ldots\hspace{0pt}\Theta}`$. We do not use node annotations in this paper and we do not include them in the expressions below in order to keep them clearer. It is however trivial to add node annotations to these expressions.
Compared to the non-annotated case when the node degree is fully specified by an integer value $`k`$ of the number of stubs attached to the node, we now have to list the numbers of attached stubs of each color to fully describe the node degree. Instead of scalar $`k`$, we thus have the node degree vector
<table>
<tbody>
<tr>
<td></td>
<td>𝐤=(k1,…,kC),𝐤subscript𝑘1…subscript𝑘𝐶{\mathbf{k}}=(k_{1},\ldots,k_{C}),</td>
<td></td>
</tr>
</tbody>
</table>
which has $`C`$ components $`k_{c}`$, each specifying the number of $`c`$-colored stubs attached to the node, cf. the left side of Figure [3](#S4.F3 "Figure 3 ‣ 4.1 General formalism ‣ 4 Network topology annotations ‣ Graph Annotations in Modeling Complex Network Topologies"). The $`L^{1}`$-norm of this vector yields the node degree with annotations ignored,
<table>
<tbody>
<tr>
<td></td>
<td>k=|𝐤|1=∑c=1Ckc.𝑘subscript𝐤1superscriptsubscript𝑐1𝐶subscript𝑘𝑐k=|{\mathbf{k}}|_{1}=\sum_{c=1}^{C}k_{c}.</td>
<td></td>
<td>(1)</td>
</tr>
</tbody>
</table>
[Refer to caption](/html/0708.3879/assets/x3.png)
Figure 3: The $`1\hspace{0pt}K`$- and $`2\hspace{0pt}K`$-annotations. Three different stub colors are represented by dashed (color red), dash-dotted (color green), and dash-double-dotted (color blue) lines.
The number $`n\hspace{0pt}{(\mathbf{k})}`$ of nodes of degree $`\mathbf{k}`$ defines the node degree distribution,
<table>
<tbody>
<tr>
<td></td>
<td>n(𝐤)n→n→∞P(𝐤),→𝑛absent→𝑛𝐤𝑛𝑃𝐤\frac{n({\mathbf{k}})}{n}\xrightarrow[n\to\infty]{}P({\mathbf{k}}),</td>
<td></td>
<td>(2)</td>
</tr>
</tbody>
</table>
in the large-graph limit. We can think of $`n\hspace{0pt}{(\mathbf{k})}`$ as a non-normalized form of $`P\hspace{0pt}{(\mathbf{k})}`$. From the statistical perspective, the $`n\hspace{0pt}{(\mathbf{k})}`$ ($`P\hspace{0pt}{(\mathbf{k})}`$) distribution is a multivariate distribution. Its C marginal distributions are the distributions of node degrees of each color $`c`$:
<table>
<tbody>
<tr>
<td></td>
<td>n(kc)=∑𝐤′|kc′=kcn(𝐤′),𝑛subscript𝑘𝑐subscriptconditionalsuperscript𝐤′subscriptsuperscript𝑘′𝑐subscript𝑘𝑐𝑛superscript𝐤′n(k_{c})=\sum_{{\mathbf{k}}^{\prime}\;\big{|}\;k^{\prime}_{c}=k_{c}}n({\mathbf{k}}^{\prime}),</td>
<td></td>
<td>(3)</td>
</tr>
</tbody>
</table>
where the summation is over all vectors $`\mathbf{k}'`$ such that their $`c`$’s component is equal to $`k_{c}`$. The degree distribution $`n\hspace{0pt}{(\mathbf{k})}`$ thus represents per-node correlations of degrees of different colors. Following the terminology in \[[29](#bib.bib29)\], we call the node degree distribution $`n\hspace{0pt}{(\mathbf{k})}`$ ($`P\hspace{0pt}{(\mathbf{k})}`$) the $`1\hspace{0pt}K`$-annotated distribution.
We then define the $`2\hspace{0pt}K`$-annotated distribution as correlations of annotated degrees of connected nodes, or simply as the number of edges that have stub of color $`c`$ connected to a node of degree $`\mathbf{k}`$ and the other stub of color $`c'`$ connected to a node of degree $`\mathbf{k}'`$, $`n\hspace{0pt}{(c,\mathbf{k};c',\mathbf{k}')}`$. See the right side of Figure [3](#S4.F3 "Figure 3 ‣ 4.1 General formalism ‣ 4 Network topology annotations ‣ Graph Annotations in Modeling Complex Network Topologies") for illustration.
As in the non-annotated case, the $`2\hspace{0pt}K`$-distribution yields a more exhaustive statistics about the annotated network topology and fully defines the $`1\hspace{0pt}K`$-distribution. To see that, we introduce the following notations:
<table>
<tbody>
<tr>
<td></td>
<td>𝐤~~𝐤\displaystyle\tilde{{\mathbf{k}}}</td>
<td>=\displaystyle=</td>
<td>(c,𝐤),𝑐𝐤\displaystyle(c,{\mathbf{k}}),</td>
<td></td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td>𝐤~′superscript~𝐤′\displaystyle\tilde{{\mathbf{k}}}^{\prime}</td>
<td>=\displaystyle=</td>
<td>(c′,𝐤′),superscript𝑐′superscript𝐤′\displaystyle(c^{\prime},{\mathbf{k}}^{\prime}),</td>
<td></td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td>μ(c,c′)𝜇𝑐superscript𝑐′\displaystyle\mu(c,c^{\prime})</td>
<td>=\displaystyle=</td>
<td>1+δ(c,c′),1𝛿𝑐superscript𝑐′\displaystyle 1+\delta(c,c^{\prime}),</td>
<td></td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td>μ(𝐤,𝐤′)𝜇𝐤superscript𝐤′\displaystyle\mu({\mathbf{k}},{\mathbf{k}}^{\prime})</td>
<td>=\displaystyle=</td>
<td>1+δ(𝐤,𝐤′),1𝛿𝐤superscript𝐤′\displaystyle 1+\delta({\mathbf{k}},{\mathbf{k}}^{\prime}),</td>
<td></td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td>μ(𝐤~,𝐤~′)𝜇~𝐤superscript~𝐤′\displaystyle\mu(\tilde{{\mathbf{k}}},\tilde{{\mathbf{k}}}^{\prime})</td>
<td>=\displaystyle=</td>
<td>1+δ(𝐤~,𝐤~′),1𝛿~𝐤superscript~𝐤′\displaystyle 1+\delta(\tilde{{\mathbf{k}}},\tilde{{\mathbf{k}}}^{\prime}),</td>
<td></td>
</tr>
</tbody>
</table>
where $`\delta\hspace{0pt}{(x,x')}`$ is the standard Kronecker delta:
<table>
<tbody>
<tr>
<td></td>
<td>δ(x,x′)={1if x=x′,0otherwise,𝛿𝑥superscript𝑥′cases1if x=x′,0otherwise,\delta(x,x^{\prime})=\begin{cases}1&\text{if \; $x=x^{\prime}$,}\\ 0&\text{otherwise,}\end{cases}</td>
<td></td>
</tr>
</tbody>
</table>
and $`x`$ is either $`c`$, $`\mathbf{k}`$, or $`\overset{\sim}{\mathbf{k}}`$. With these notations, one can easily check that the normalized $`2\hspace{0pt}K`$-annotated distribution is
<table>
<tbody>
<tr>
<td></td>
<td>P(𝐤~,𝐤~′)=n(𝐤~,𝐤~′)μ(𝐤~,𝐤~′)/(2m),𝑃~𝐤superscript~𝐤′𝑛~𝐤superscript~𝐤′𝜇~𝐤superscript~𝐤′2𝑚P(\tilde{{\mathbf{k}}},\tilde{{\mathbf{k}}}^{\prime})=n(\tilde{{\mathbf{k}}},\tilde{{\mathbf{k}}}^{\prime})\mu(\tilde{{\mathbf{k}}},\tilde{{\mathbf{k}}}^{\prime})/(2m),</td>
<td></td>
<td>(4)</td>
</tr>
</tbody>
</table>
the number of edges of any pair of colors connecting nodes of degrees $`\mathbf{k}`$ and $`\mathbf{k}'`$ is
<table>
<tbody>
<tr>
<td></td>
<td>n(𝐤,𝐤′)=∑c,c′n(𝐤~,𝐤~′)μ(𝐤~,𝐤~′)/μ(𝐤,𝐤′),𝑛𝐤superscript𝐤′subscript𝑐superscript𝑐′𝑛~𝐤superscript~𝐤′𝜇~𝐤superscript~𝐤′𝜇𝐤superscript𝐤′n({\mathbf{k}},{\mathbf{k}}^{\prime})=\sum_{c,c^{\prime}}n(\tilde{{\mathbf{k}}},\tilde{{\mathbf{k}}}^{\prime})\mu(\tilde{{\mathbf{k}}},\tilde{{\mathbf{k}}}^{\prime})/\mu({\mathbf{k}},{\mathbf{k}}^{\prime}),</td>
<td></td>
</tr>
</tbody>
</table>
the normalized form of this distributions is
<table>
<tbody>
<tr>
<td></td>
<td>P(𝐤,𝐤′)=n(𝐤,𝐤′)μ(𝐤,𝐤′)/(2m),𝑃𝐤superscript𝐤′𝑛𝐤superscript𝐤′𝜇𝐤superscript𝐤′2𝑚P({\mathbf{k}},{\mathbf{k}}^{\prime})=n({\mathbf{k}},{\mathbf{k}}^{\prime})\mu({\mathbf{k}},{\mathbf{k}}^{\prime})/(2m),</td>
<td></td>
</tr>
</tbody>
</table>
and the $`1\hspace{0pt}K`$-distribution is given by
<table>
<tbody>
<tr>
<td></td>
<td>n(𝐤)𝑛𝐤\displaystyle n({\mathbf{k}})</td>
<td>=\displaystyle=</td>
<td>∑𝐤′n(𝐤,𝐤′)μ(𝐤,𝐤′)/k,subscriptsuperscript𝐤′𝑛𝐤superscript𝐤′𝜇𝐤superscript𝐤′𝑘\displaystyle\sum_{{\mathbf{k}}^{\prime}}n({\mathbf{k}},{\mathbf{k}}^{\prime})\mu({\mathbf{k}},{\mathbf{k}}^{\prime})/k,</td>
<td></td>
<td>(5)</td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td>P(𝐤)𝑃𝐤\displaystyle P({\mathbf{k}})</td>
<td>=\displaystyle=</td>
<td>k¯k∑𝐤′P(𝐤,𝐤′),¯𝑘𝑘subscriptsuperscript𝐤′𝑃𝐤superscript𝐤′\displaystyle\frac{\bar{k}}{k}\sum_{{\mathbf{k}}^{\prime}}P({\mathbf{k}},{\mathbf{k}}^{\prime}),</td>
<td></td>
<td>(6)</td>
</tr>
</tbody>
</table>
where $`\overline{k} = {{2\hspace{0pt}m}/n}`$ is the average degree. The last two expressions show how one can find the $`1\hspace{0pt}K`$-annotated distribution given the $`2\hspace{0pt}K`$-annotated distribution, and they look exactly the same as in the non-annotated case \[[29](#bib.bib29)\], except that we have vectors $`\mathbf{k},\mathbf{k}'`$ instead of scalars $`k,k'`$.
The $`d\hspace{0pt}K`$-annotated distributions with $`d > 2`$ \[[29](#bib.bib29)\] can be defined in a similar way.
### 4.2 The AS relationship annotations
In the specific case of the AS-level Internet topology that interests us in this paper, we have just three colors: customer, provider, and peer.
<sup>3</sup>
<sup>3</sup>3We ignore sibling relationships, since they typically account for a very small fraction of the total number of edges. As found in \[[13](#bib.bib13)\], the number of s2s edges is only 0.46% of the total number of edges in the AS-level Internet.
We assign the following numeric values to represent these three colors:
<table>
<tbody>
<tr>
<td></td>
<td>c={1customer,2provider,3peer.𝑐cases1customer2provider3peerc=\begin{cases}1&\;\text{customer},\\ 2&\;\text{provider},\\ 3&\;\text{peer}.\end{cases}</td>
<td></td>
</tr>
</tbody>
</table>
These three stub annotations come under the following two constraints defining the only two types of edges that we have: 1) c2p edges: if one stub of an edge is customer, then the other stub of the same edge is provider, and vice versa; and 2) p2p edges: if one stub of an edge is peer, then the other stub of the same edges is also peer. The c2p edges are thus asymmetric, i.e., a generalization of directed edges, while the p2p edges are symmetric, i.e., a generalization of bi-directed or undirected edges.
While the $`2\hspace{0pt}K`$-annotated distribution $`n\hspace{0pt}{(\overset{\sim}{\mathbf{k}},{\overset{\sim}{\mathbf{k}}}')}`$ contains the most exhaustive information about the network topology, it has too many (seven) independent arguments. As a result, the full $`2\hspace{0pt}K`$-annotated statistics is extremely sparse, which makes it difficult to model and reproduce directly. We thus have to find some summary statistics of $`n\hspace{0pt}{(\overset{\sim}{\mathbf{k}},{\overset{\sim}{\mathbf{k}}}')}`$ that we can model in practice. For each concrete complex network type, these summary statistics might be different. Given measurement data for a specific complex network, one would usually have to start with identifying a meaningful set of summary statistics of the $`2\hspace{0pt}K`$-annotated distribution, and then proceed from there. At the same time, we believe that as soon as the $`2\hspace{0pt}K`$-annotated distribution fully defines an observed complex network, i.e., the network is $`2\hspace{0pt}K`$-annotated-random \[[29](#bib.bib29)\], one can generally use the set of summary statistics that we found necessary and sufficient to reproduce in order to model correctly the Internet AS topology. In the rest of this section, we list these statistics and describe the specific meanings that they have in the AS topology case.
Degree distribution (DD). This statistics is the traditional non-annotated degree distribution $`n\hspace{0pt}{(k)}`$, where $`k`$ is as in eq. ([1](#S4.E1 "In 4.1 General formalism ‣ 4 Network topology annotations ‣ Graph Annotations in Modeling Complex Network Topologies")). The DD tells us how many ASs of each total degree $`k`$ are in the network.
Annotation distributions (ADs). The DD of an AS topology does not convey any information about the AS relationships. The initial step to account for this information is to reproduce the distributions of ASs with specific numbers of attached customer, provider, or peer stubs. These annotation distributions (ADs) are the marginal distributions $`n\hspace{0pt}{(k_{c})}`$, $`c = {1,2,3}`$, of the $`1\hspace{0pt}K`$-annotated distribution. They are given by eq. ([3](#S4.E3 "In 4.1 General formalism ‣ 4 Network topology annotations ‣ Graph Annotations in Modeling Complex Network Topologies")). If $`k_{1}`$ ($`k_{2}`$) customer (provider) stubs attach to an AS, then this AS has exactly $`k_{1}`$ ($`k_{2}`$) providers (customers), since the c2p edges are asymmetric. Consequently, the ADs $`n\hspace{0pt}{(k_{1})}`$ and $`n\hspace{0pt}{(k_{2})}`$ tell us how many ASs with the specific numbers of providers and, respectively, customers the network has. Since the p2p edges are symmetric, the AD $`n\hspace{0pt}{(k_{3})}`$ is the distribution of ASs with specific numbers of peers.
Annotated degree distribution (ADD). The ADs do not tell us anything about the correlations among annotated degrees of the same node, i.e., how many customers, providers, and, simultaneously, peers a specific AS has. Correlations of this type are fully described by the $`1\hspace{0pt}K`$-annotated distribution in eq. ([2](#S4.E2 "In 4.1 General formalism ‣ 4 Network topology annotations ‣ Graph Annotations in Modeling Complex Network Topologies")), which we also call the annotated degree distribution (ADD). These correlations are present in the Internet. For example, large tier-1 ISPs typically have a large number of customers, i.e., large $`k_{2}`$, no providers, i.e., zero $`k_{1}`$, and a small number of peers, i.e., small $`k_{3}`$. On the other hand, medium-size ISPs tend to have a small set of customers, several peers, and few providers. Ignoring the ADD while generating synthetic graphs can lead to artifacts like high-degree nodes with many providers—a property obviously absent in the real Internet.
Joint degree distributions (JDDs). While the ADD contains the full information about degree correlations “at nodes,” it does not tell us anything about degree correlations “across links,” while the latter type of correlations is also characteristic for the Internet. For example, large tier-1 ISPs typically have p2p relationships with other tier-1 ISPs, not with much smaller ISPs, while small ISPs have p2p links with other small ISPs. In other words, p2p links usually connect ASs of similar degrees, i.e., $`k \sim k'`$. Similarly, c2p links tend to connect low-degree customers to high-degree providers, i.e., $`k \ll k'`$. If we ignore these correlations, we can synthesize graphs with inaccuracies like p2p links connecting ASs of drastically dissimilar degrees. To reproduce these correlations, we work with the following summary statistics of the $`2\hspace{0pt}K`$-annotated distribution in eq. ([4](#S4.E4 "In 4.1 General formalism ‣ 4 Network topology annotations ‣ Graph Annotations in Modeling Complex Network Topologies")):
<table>
<tbody>
<tr>
<td></td>
<td>nc2p(k,k′)subscript𝑛𝑐2𝑝𝑘superscript𝑘′\displaystyle n_{c2p}(k,k^{\prime})</td>
<td>=\displaystyle=</td>
<td>∑𝐤,𝐤′||𝐤|1=k,|𝐤′|1=k′n(1,𝐤;2,𝐤′),\displaystyle\sum_{{\mathbf{k}},{\mathbf{k}}^{\prime}\;\big{|}\;|{\mathbf{k}}|_{1}=k,\,|{\mathbf{k}}^{\prime}|_{1}=k^{\prime}}n(1,{\mathbf{k}};2,{\mathbf{k}}^{\prime}),</td>
<td></td>
<td>(7)</td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td>np2p(k,k′)subscript𝑛𝑝2𝑝𝑘superscript𝑘′\displaystyle n_{p2p}(k,k^{\prime})</td>
<td>=\displaystyle=</td>
<td>∑𝐤,𝐤′||𝐤|1=k,|𝐤′|1=k′n(3,𝐤;3,𝐤′),\displaystyle\sum_{{\mathbf{k}},{\mathbf{k}}^{\prime}\;\big{|}\;|{\mathbf{k}}|_{1}=k,\,|{\mathbf{k}}^{\prime}|_{1}=k^{\prime}}n(3,{\mathbf{k}};3,{\mathbf{k}}^{\prime}),</td>
<td></td>
<td>(8)</td>
</tr>
</tbody>
</table>
where the summation is over such vectors $`\mathbf{k}`$ and $`\mathbf{k}'`$ that their $`L^{1}`$-norms are $`k`$ and $`k'`$ respectively. The first expression gives the number of c2p links that have their customer stub attached to a node of total degree $`k`$ and provider stub attached to a node of total degree $`k'`$. The second expression is the number of p2p links between nodes of total degrees $`k`$ and $`k'`$. In other words, these two objects are the joint degree distributions (JDDs) for the c2p and p2p links.
In summary, we work with the four types of distributions, i.e., DD, ADs, ADD, and JDDs, that allow two types of classification:
1. 1\.
Univariate vs. multivariate distributions:
1. \(a\)
Univariate. The ADs and DD are distribution of only one random variable.
2. \(b\)
Multivariate. The ADD and JDDs are joint distribution of three and two random variables. The marginal distribution of these variables are the ADs, cf. eq. ([3](#S4.E3 "In 4.1 General formalism ‣ 4 Network topology annotations ‣ Graph Annotations in Modeling Complex Network Topologies")), and DD, cf. eqs. ([7](#S4.E7 "In 4.2 The AS relationship annotations ‣ 4 Network topology annotations ‣ Graph Annotations in Modeling Complex Network Topologies"),[8](#S4.E8 "In 4.2 The AS relationship annotations ‣ 4 Network topology annotations ‣ Graph Annotations in Modeling Complex Network Topologies")), respectively.
2. 2\.
$`\mathbf{1}\hspace{0pt}\mathbf{K}`$- vs. $`\mathbf{2}\hspace{0pt}\mathbf{K}`$-summary statistics:
1. \(a\)
$`1\hspace{0pt}K`$-derived. The DD, ADs, and ADD are fully defined by the $`1\hspace{0pt}K`$-annotated distribution: that is, we do not need to know the $`2\hspace{0pt}K`$-annotated distribution to calculate the distributions in this class.
2. \(b\)
$`2\hspace{0pt}K`$-derived. The JDDs are fully defined only by the $`2\hspace{0pt}K`$-annotated distribution. Note that it also defines the $`1\hspace{0pt}K`$-annotated distribution via eqs. ([5](#S4.E5 "In 4.1 General formalism ‣ 4 Network topology annotations ‣ Graph Annotations in Modeling Complex Network Topologies"),[6](#S4.E6 "In 4.1 General formalism ‣ 4 Network topology annotations ‣ Graph Annotations in Modeling Complex Network Topologies")).
## 5 Generating annotated AS graphs
In this section we describe how we generate synthetic annotated AS graphs of arbitrary sizes. We want our synthetic graphs to reproduce as many important properties of the original measured topology as possible. For this purpose we decide to explicitly model and reproduce the summary statistics of the $`2\hspace{0pt}K`$-annotated distribution from Section [4.2](#S4.SS2 "4.2 The AS relationship annotations ‣ 4 Network topology annotations ‣ Graph Annotations in Modeling Complex Network Topologies"), because \[[29](#bib.bib29)\] showed that by reproducing $`2\hspace{0pt}K`$-distributions, one automatically captures a long list of other important properties of AS topologies. In other words, the task of generating synthetic annotated topologies becomes equivalent to the task of generating random annotated graphs that reproduce the summary statistics of the $`2\hspace{0pt}K`$-annotated distribution of the measured AS topology.
We wish to be able to generate synthetic topologies of different sizes, but the $`2\hspace{0pt}K`$-summary statistics defined in Section [4.2](#S4.SS2 "4.2 The AS relationship annotations ‣ 4 Network topology annotations ‣ Graph Annotations in Modeling Complex Network Topologies") are all bound to a specific graph size. Therefore, in order to generate arbitrarily-sized graphs, we need first to rescale the $`2\hspace{0pt}K`$-summary statistics from the original to target graph sizes. We say that an empirical distribution is rescaled with respect to another empirical distribution, if the both distributions are defined by two different finite collections of random numbers drawn from the same continuous distribution. For example, the distributions of node scalar (or vector) degrees in two different graphs are rescaled with respect to each other if these degrees are drawn from the same continuous univariate (or multivariate) probability distribution. We say that a $`2\hspace{0pt}K`$-annotated graph is rescaled with respect to another $`2\hspace{0pt}K`$-annotated graph, if all the $`2\hspace{0pt}K`$-summary statistics of the first graphs are rescaled with respect to the corresponding $`2\hspace{0pt}K`$-summary statistics of the second graph. This definition of rescaling is equivalent to assuming that for each summary statistic, the same distribution function describes the ensemble of empirical distributions of the statistic in past, present, and future Internet topologies. In other words, we assume that the $`2\hspace{0pt}K`$-annotated correlation profile of the Internet AS topology is an invariant of its evolution. This assumption is realistic, as discussed, for example, in \[[36](#bib.bib36)\], where it is shown that the non-annotated $`1\hspace{0pt}K`$- and $`2\hspace{0pt}K`$-distributions of the Internet AS topology have stayed approximately the same during all the years (more than a decade) of the existing data time span.
To illustrate what we mean by rescaling, consider the empirical distribution of peer degrees, i.e., the AD $`n\hspace{0pt}{(k_{3})}`$, in the measured AS topology annotated with AS relationships in Figure [4(a)](#S5.F4.sf1 "In Figure 4 ‣ 5 Generating annotated AS graphs ‣ Graph Annotations in Modeling Complex Network Topologies"). The figure shows the empirical complementary cumulative distribution function (CCDF) for peer-degrees of 19,036 nodes, i.e., 19,036 numbers of peer stubs attached to a node, and the largest such number is 448. The continuous probability distribution of Figure [4(b)](#S5.F4.sf2 "In Figure 4 ‣ 5 Generating annotated AS graphs ‣ Graph Annotations in Modeling Complex Network Topologies") approximates the empirical distribution in Figure [4(a)](#S5.F4.sf1 "In Figure 4 ‣ 5 Generating annotated AS graphs ‣ Graph Annotations in Modeling Complex Network Topologies"). Figures [4(c)](#S5.F4.sf3 "In Figure 4 ‣ 5 Generating annotated AS graphs ‣ Graph Annotations in Modeling Complex Network Topologies"), [4(d)](#S5.F4.sf4 "In Figure 4 ‣ 5 Generating annotated AS graphs ‣ Graph Annotations in Modeling Complex Network Topologies"), and [4(e)](#S5.F4.sf5 "In Figure 4 ‣ 5 Generating annotated AS graphs ‣ Graph Annotations in Modeling Complex Network Topologies") show the CCDFs of three collections of 5,000, 20,000, and 50,000 random numbers drawn from the probability distribution in Figure [4(b)](#S5.F4.sf2 "In Figure 4 ‣ 5 Generating annotated AS graphs ‣ Graph Annotations in Modeling Complex Network Topologies"). According to our definition of rescaling, the distributions in Figures [4(c)](#S5.F4.sf3 "In Figure 4 ‣ 5 Generating annotated AS graphs ‣ Graph Annotations in Modeling Complex Network Topologies"), [4(d)](#S5.F4.sf4 "In Figure 4 ‣ 5 Generating annotated AS graphs ‣ Graph Annotations in Modeling Complex Network Topologies"), and [4(e)](#S5.F4.sf5 "In Figure 4 ‣ 5 Generating annotated AS graphs ‣ Graph Annotations in Modeling Complex Network Topologies") are rescaled with respect to the distribution of Figure [4(a)](#S5.F4.sf1 "In Figure 4 ‣ 5 Generating annotated AS graphs ‣ Graph Annotations in Modeling Complex Network Topologies"). We see that all the empirical distributions have the same overall shape, but differ in the total number of samples and in the maximum values within these sample collections. Distributions with larger maximums correspond, as expected, to bigger collections of samples.
[Refer to caption](/html/0708.3879/assets/x4.png)
\(a\)
Original distribution of 19,036 peer-degrees in the measured AS topology.
[Refer to caption](/html/0708.3879/assets/x5.png)
\(b\)
Continuous distribution approximating the original one.
[Refer to caption](/html/0708.3879/assets/x6.png)
\(c\)
Rescaled distribution of 5,000 samples.
[Refer to caption](/html/0708.3879/assets/x7.png)
\(d\)
Rescaled distribution of 20,000 samples.
[Refer to caption](/html/0708.3879/assets/x8.png)
\(e\)
Rescaled distribution of 50,000 samples.
Figure 4: Rescaling an empirical distribution. The distributions in the three bottom figures are rescaled with respect to the empirical distribution in the first figure. The distribution in the second figure is a continuous approximation of the distribution in the first figure and is used to generate the rescaled distributions in the bottom figures. For each discrete distribution, we show its maximum in the top-right corners of the plots.
### 5.1 Overview of the approach
We now move to describing the details of our approach, which consists of the following three major phases:
1. 1\.
Extraction.
We first extract the empirical $`2\hspace{0pt}K`$-summary distributions from available AS topology measurement data. We annotate links of the AS topology extracted from this data using existing AS relationship inference heuristics. This extraction step is conceptually simplest. On its output, we obtain the extracted $`2\hspace{0pt}K`$-summary distributions that are all bound to the size of the measured AS graph.
2. 2\.
Rescaling.
1. \(a\)
We use the extracted empirical distributions to find their continuous approximations. Referring to our example in Figure [4](#S5.F4 "Figure 4 ‣ 5 Generating annotated AS graphs ‣ Graph Annotations in Modeling Complex Network Topologies"), this step corresponds to computing the continuous probability distribution in Figure [4(b)](#S5.F4.sf2 "In Figure 4 ‣ 5 Generating annotated AS graphs ‣ Graph Annotations in Modeling Complex Network Topologies") based on the empirical distribution in Figure [4(a)](#S5.F4.sf1 "In Figure 4 ‣ 5 Generating annotated AS graphs ‣ Graph Annotations in Modeling Complex Network Topologies").
2. \(b\)
We then use the computed probability distributions to rescale the empirical distributions obtained at the extraction step. We generate a desired, target number of random scalar or vector degree samples drawn from the corresponding probability distributions. The generated degree samples have empirical distributions that are rescaled with respect to the corresponding empirical distributions of the measured topology. Referring to our example in Figure [4](#S5.F4 "Figure 4 ‣ 5 Generating annotated AS graphs ‣ Graph Annotations in Modeling Complex Network Topologies"), this step corresponds to generating the rescaled empirical distributions in Figures [4(c)](#S5.F4.sf3 "In Figure 4 ‣ 5 Generating annotated AS graphs ‣ Graph Annotations in Modeling Complex Network Topologies"), [4(d)](#S5.F4.sf4 "In Figure 4 ‣ 5 Generating annotated AS graphs ‣ Graph Annotations in Modeling Complex Network Topologies"), and [4(e)](#S5.F4.sf5 "In Figure 4 ‣ 5 Generating annotated AS graphs ‣ Graph Annotations in Modeling Complex Network Topologies") based on the probability distribution in Figure [4(b)](#S5.F4.sf2 "In Figure 4 ‣ 5 Generating annotated AS graphs ‣ Graph Annotations in Modeling Complex Network Topologies").
3. 3\.
Construction.
Finally, we develop algorithms to generate synthetic graphs that have their $`2\hspace{0pt}K`$-summary distributions equal to given distributions, i.e., to the corresponding distributions obtained at the previous step. The generated graphs thus reproduce the rescaled replicas of the $`2\hspace{0pt}K`$-annotated distribution of the original topology, but they are “maximally random” in all other respects.
In the rest of this section, we describe each of these phases in detail.
### 5.2 Extraction
We extract the AS topology from the RouteViews \[[38](#bib.bib38)\] data, performing some standard data cleaning, such as ignoring private AS numbers, AS sets, etc. \[[30](#bib.bib30)\] The resulting AS graph is initially non-annotated. To annotate it, we infer c2p and p2p relationships for AS links using the heuristics in \[[13](#bib.bib13)\]. We thus obtain the real Internet AS topology annotated with c2p and p2p relationships. Given this annotated topology, we straightforwardly calculate all the empirical $`2\hspace{0pt}K`$-summary distributions that we have defined in Section [4.2](#S4.SS2 "4.2 The AS relationship annotations ‣ 4 Network topology annotations ‣ Graph Annotations in Modeling Complex Network Topologies").
While the extraction phase is conceptually and technically the simplest phase of the overall approach, it is its basis. Therefore the quality of the input Internet topology data is a natural concern. This data is known to exhibit a variety of vagaries, e.g., due to sampling biases \[[27](#bib.bib27), [10](#bib.bib10), [11](#bib.bib11)\]. However, our approach is oblivious with respect to data quality. It takes any available data, extracts the described statistics from it, and reproduces them, properly rescaled, in random synthetic graphs. A given input topology data set thus defines an ensemble of random graphs generated by our method. By construction, all graphs in this ensemble reproduce the described set of annotated distributions. In addition, in Section [6](#S6 "6 Evaluation ‣ Graph Annotations in Modeling Complex Network Topologies"), we perform sensitivity analysis in order to see the strength of fluctuations of these and other basic graph metrics within an ensemble. The quality of these graph ensembles, in terms of how veraciously they reflect reality, will improve as the quality of available topology data improves in the future. In this paper, we simply illustrate our approach with the currently available topology data. The RouteViews \[[38](#bib.bib38)\] is just one of very few sources of such data \[[30](#bib.bib30)\]. We select it because it appears to be the most frequently cited Internet topology data source.
### 5.3 Rescaling
Our rescaling approach differs for univariate and multivariate distributions.
#### 5.3.1 Rescaling univariate distributions
We recall from the end of Section [4.2](#S4.SS2 "4.2 The AS relationship annotations ‣ 4 Network topology annotations ‣ Graph Annotations in Modeling Complex Network Topologies") that we have the following two types of univariate distributions: the ADs and the DD. Here we describe how we rescale ADs. We note that we do not have to rescale the DD the same way. The reason is that our approach to rescaling the ADD, which we discuss below in Section [5.3.2](#S5.SS3.SSS2 "5.3.2 Rescaling multivariate distributions ‣ 5.3 Rescaling ‣ 5 Generating annotated AS graphs ‣ Graph Annotations in Modeling Complex Network Topologies"), automatically takes care of rescaling the DD, since the ADD is the distribution of degree vectors and the DD is the distribution of the $`L^{1}`$-norms of these vectors, cf. eq. ([1](#S4.E1 "In 4.1 General formalism ‣ 4 Network topology annotations ‣ Graph Annotations in Modeling Complex Network Topologies")).
The first problem we face trying to compute a continuous approximation for a given finite empirical distribution is that we have to not only interpolate between points of the empirical distribution, but also extrapolate above its maximum value. For example, if we want to construct a synthetic graph bigger than the original, then we expect its maximum degree to be larger than the maximum degree in the original graph. Therefore we have to properly extrapolate the observed degree distribution beyond the observed maximum degree.
We solve this problem by fitting the univariate empirical distributions with smoothing splines. Spline smoothing is a non-parametric estimator of an unknown function represented by a collection of empirical data points. Spline smoothing produces a smooth curve passing through or near the data points. For example, the curve in Figure [4(b)](#S5.F4.sf2 "In Figure 4 ‣ 5 Generating annotated AS graphs ‣ Graph Annotations in Modeling Complex Network Topologies") is a smooth spline of the empirical distribution of Figure [4(a)](#S5.F4.sf1 "In Figure 4 ‣ 5 Generating annotated AS graphs ‣ Graph Annotations in Modeling Complex Network Topologies"). Spline smoothing can also extrapolate the shape of an empirical function beyond the original data range.
Another reason to select spline smoothing is that it comes useful for fitting distributions that do not closely follow regular shapes, e.g., “clean” power laws. The ADs of the Internet topology do not necessarily have such regular shapes. For example, the distribution of the number of peers, i.e., the AD $`n\hspace{0pt}{(k_{3})}`$, has a complex shape that we found impossible to fit with any single-parametric distribution.
Among available implementations of spline smoothing techniques, we select the one in the smooth.spline method of the R project \[[37](#bib.bib37)\], a popular statistical computing package. The specific details of this technique are in \[[7](#bib.bib7)\].
We can approximate with splines either the CDFs or CCDFs
<sup>4</sup>
<sup>4</sup>4Recall that the CCDF of CDF $`F\hspace{0pt}{(x)}`$ is $`1 - {F\hspace{0pt}{(x)}}`$.
of the ADs obtained at the extraction step. We chose to fit the CCDFs rather than the CDFs because the former better capture the shapes of high-degree tails of our heavy-tailed ADs.
Another important detail is that we can define an empirical CCDF to be either a left- or right-continuous step function \[[24](#bib.bib24)\]. Usually an empirical CCDF at some point $`x`$ is defined as the fraction of samples with values strictly larger than $`x`$, which means that the distribution is right-continuous and that the probability of a value larger than the observed maximum value is zero, whereas the probability of a value smaller than the observed minimum value is unspecified. For degree distributions, we know that the probability of a degree smaller than zero is zero,
<sup>5</sup>
<sup>5</sup>5For a given node, some but not all the degrees $`k_{1}`$, $`k_{2}`$, and $`k_{3}`$ can be equal to zero.
but we do not know the probability of a degree larger than the maximum observed degree. For this reason, we decide to fit the left-continuous variants of empirical CCDFs, i.e., we define a CCDF at some point $`x`$ as the fraction of samples with values greater than or equal to $`x`$.
Having the original ADs fitted with splines and assuming that our target graph size is $`N`$, we finally use the standard, inverse-CDF method to produce $`N`$ random numbers that follow the continuous distributions given by the splines. Recall that the inverse-CDF method is based on the observation that if the CDF of $`N`$ random numbers $`x_{j}`$, $`j = {1\hspace{0pt}\ldots\hspace{0pt}N}`$, closely follows some function $`F\hspace{0pt}{(x)}`$, then the distribution of numbers $`y_{j} = {F\hspace{0pt}{(x_{j})}}`$ is approximately uniform in the interval $`\lbrack 0,1\rbrack`$. As its name suggests, the inverse-CDF inverts this observation and operates as follows \[[25](#bib.bib25)\]: given a target CDF $`F\hspace{0pt}{(x)}`$ and a target size $`N`$ of a collection of random samples, the method first generates $`N`$ random numbers $`y_{j}`$ uniformly distributed in $`\lbrack 0,1\rbrack`$ and then outputs numbers $`x_{j} = {F^{- 1}\hspace{0pt}{(y_{j})}}`$, where $`F^{- 1}\hspace{0pt}{(y)}`$ is the inverse of CDF $`F\hspace{0pt}{(x)}`$, i.e., $`{F^{- 1}\hspace{0pt}{({F\hspace{0pt}{(x)}})}} = x`$. The CDF of numbers $`x_{j}`$ closely follows $`F\hspace{0pt}{(x)}`$. Figures [4(c)](#S5.F4.sf3 "In Figure 4 ‣ 5 Generating annotated AS graphs ‣ Graph Annotations in Modeling Complex Network Topologies"), [4(d)](#S5.F4.sf4 "In Figure 4 ‣ 5 Generating annotated AS graphs ‣ Graph Annotations in Modeling Complex Network Topologies"), and [4(e)](#S5.F4.sf5 "In Figure 4 ‣ 5 Generating annotated AS graphs ‣ Graph Annotations in Modeling Complex Network Topologies") show random numbers generated this way. These numbers follow the distribution in Figure [4(b)](#S5.F4.sf2 "In Figure 4 ‣ 5 Generating annotated AS graphs ‣ Graph Annotations in Modeling Complex Network Topologies"). To compute values of inverse CDFs on $`N`$ random numbers uniformly distributed in $`\lbrack 0,1\rbrack`$ in our case, we use the predict.smooth.spline method of the R project. Since the random numbers produced in this way are not, in general, integers, we convert them to integer degree values using the floor function. We have to use the floor and not the ceiling function because we work with left- rather than right-continuous distributions.
The outcome of the described process is three sets of $`N`$ random numbers that represent $`N`$ customer degrees $`d_{1}^{j}`$, $`N`$ provider degrees $`d_{2}^{j}`$, and $`N`$ peer degrees $`d_{3}^{j}`$ of nodes in the target graph, $`j = {1\hspace{0pt}\ldots\hspace{0pt}N}`$. We denote the CDFs of these random numbers by $`D_{1}\hspace{0pt}{(d_{1})}`$, $`D_{2}\hspace{0pt}{(d_{2})}`$, and $`D_{3}\hspace{0pt}{(d_{3})}`$ respectively. By construction, these distributions are properly rescaled versions of the customer-, provider- and peer-annotation distributions (ADs) in the measured AS topology.
#### 5.3.2 Rescaling multivariate distributions
Rescaling multivariate distributions is not as simple as rescaling univariate distributions. Our approach for rescaling univariate distributions is not practically applicable to rescaling multivariate distributions because it is difficult to fit distributions that have many variates and complex shapes. To rescale multivariate distributions, we use copulas \[[33](#bib.bib33)\], which are a statistical tool for quantifying correlations between several random variables. Compared to other well-known correlation metrics, such as Pearson’s coefficient, copulas give not a single scalar value but a function of several arguments that fully describes complex, fine-grained details of the structure of correlations among the variables, i.e., their correlation profile.
According to Sklar’s theorem \[[39](#bib.bib39)\], any $`p`$-dimensional multivariate CDF $`F`$ of $`p`$ random variables $`\mathbf{k} = {(k_{1},\ldots,k_{p})}`$ can be written in the following form:
<table>
<tbody>
<tr>
<td></td>
<td>F(𝐤)=H(𝐮),𝐹𝐤𝐻𝐮F({\mathbf{k}})=H({\mathbf{u}}),</td>
<td></td>
<td>(9)</td>
</tr>
</tbody>
</table>
where $`\mathbf{u}`$ is the $`p`$-dimensional vector composed of the $`F`$’s marginal CDFs $`F_{m}\hspace{0pt}{(k_{m})}`$, $`m = {1,\ldots,p}`$:
<table>
<tbody>
<tr>
<td></td>
<td>Fm(km)subscript𝐹𝑚subscript𝑘𝑚\displaystyle F_{m}(k_{m})</td>
<td>=\displaystyle=</td>
<td>F(∞,…,∞,km,∞,…,∞),𝐹…subscript𝑘𝑚…\displaystyle F(\infty,\ldots,\infty,k_{m},\infty,\ldots,\infty),</td>
<td></td>
<td>(10)</td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td>𝐮𝐮\displaystyle{\mathbf{u}}</td>
<td>=\displaystyle=</td>
<td>(F1(k1),…,Fp(kp)).subscript𝐹1subscript𝑘1…subscript𝐹𝑝subscript𝑘𝑝\displaystyle(F_{1}(k_{1}),\ldots,F_{p}(k_{p})).</td>
<td></td>
<td>(11)</td>
</tr>
</tbody>
</table>
The function $`H`$ is called a copula and each of its marginal distributions is uniform in $`\lbrack 0,1\rbrack`$.
Copulas play a critical role at the following two steps in our approach for rescaling multivariate distributions. First, they allow us to split a multivariate distribution of the original, measured topology into two parts: the first part consists of the marginal distributions $`F_{m}`$, while the second part is their correlation profile, i.e., copula $`H`$. These two parts are independent. Therefore, we can independently rescale the marginal distributions and the correlation profile. This property tremendously simplifies the rescaling process. The marginals are univariate distributions that we rescale as in Section [5.3.1](#S5.SS3.SSS1 "5.3.1 Rescaling univariate distributions ‣ 5.3 Rescaling ‣ 5 Generating annotated AS graphs ‣ Graph Annotations in Modeling Complex Network Topologies"), while this section contains the details of how we rescale the correlation profile. We use copulas the second time to merge together rescaled marginals and their correlation profile to yield a rescaled multivariate distribution in its final form.
In Figure [5](#S5.F5 "Figure 5 ‣ 5.3.2 Rescaling multivariate distributions ‣ 5.3 Rescaling ‣ 5 Generating annotated AS graphs ‣ Graph Annotations in Modeling Complex Network Topologies") we present a high-level overview of our approaches for rescaling univariate and multivariate distributions. To rescale an original empirical univariate distribution, we first approximate it with splines and then use these splines to generate random numbers. We split the process of rescaling an original empirical multivariate distribution into two independent rescaling sub-processes, i.e., rescaling the marginals and their copula. We rescale the marginals as any other univariate distributions. To rescale the copula, we re-sample measured correlation data as we describe below in this section. At the end of multivariate rescaling, we merge the rescaled marginals with the rescaled copula to yield a rescaled multivariate distribution. One can see from Figure [5](#S5.F5 "Figure 5 ‣ 5.3.2 Rescaling multivariate distributions ‣ 5.3 Rescaling ‣ 5 Generating annotated AS graphs ‣ Graph Annotations in Modeling Complex Network Topologies") that multivariate rescaling is a “superset,” in terms of actions involved, of univariate rescaling. The following three steps summarize the high-level description of our multivariate rescaling approach:
1. 1\.
extract and rescale the univariate marginals of a multivariate distribution as described in Section [5.3.1](#S5.SS3.SSS1 "5.3.1 Rescaling univariate distributions ‣ 5.3 Rescaling ‣ 5 Generating annotated AS graphs ‣ Graph Annotations in Modeling Complex Network Topologies") (boxes (a), (b), and (c) in Figure [5](#S5.F5 "Figure 5 ‣ 5.3.2 Rescaling multivariate distributions ‣ 5.3 Rescaling ‣ 5 Generating annotated AS graphs ‣ Graph Annotations in Modeling Complex Network Topologies"));
2. 2\.
extract and rescale the copula of the multivariate distribution (boxes (d) and (e) in Figure [5](#S5.F5 "Figure 5 ‣ 5.3.2 Rescaling multivariate distributions ‣ 5.3 Rescaling ‣ 5 Generating annotated AS graphs ‣ Graph Annotations in Modeling Complex Network Topologies")); and
3. 3\.
merge the rescaled marginals and copula yielding a rescaled multivariate distribution (box (f) in Figure [5](#S5.F5 "Figure 5 ‣ 5.3.2 Rescaling multivariate distributions ‣ 5.3 Rescaling ‣ 5 Generating annotated AS graphs ‣ Graph Annotations in Modeling Complex Network Topologies")).
In the rest of this section, we provide the low-level details for the last two steps, using the ADD multivariate distribution as an example.
[Refer to caption](/html/0708.3879/assets/x9.png)
Figure 5: Overview of rescaling univariate and multivariate distributions.
At Step 2, we compute a rescaled ADD copula as follows. The collected AS topology has $`n`$ nodes, and for each node $`i`$, $`i = {1\hspace{0pt}\ldots\hspace{0pt}n}`$, we record its degree vector $`\mathbf{k}_{i} = {(k_{1}^{i},k_{2}^{i},k_{3}^{i})}`$ producing an $`n`$-sized set of degree triplets. We then perform statistical simulation on this set to produce another set of a desired size that has the same correlations as the measured data $`\mathbf{k}_{i}`$. Specifically, we re-sample, uniformly at random and with replacement, $`N`$ degree triplets from the set of vectors $`\mathbf{k}_{i}`$, where $`N`$ is the target size of our synthetic topology. We thus obtain an $`N`$-sized set of random triplets $`\mathbf{k}_{j}`$, $`j = {1\hspace{0pt}\ldots\hspace{0pt}N}`$, and we denote their joint CDF by $`F\hspace{0pt}{(\mathbf{k})}`$. By construction, the empirical distribution of triplets $`\mathbf{k}_{j}`$ has the same correlation profile as original triplets $`\mathbf{k}_{i}`$. This procedure corresponds for box (d) in Figure [5](#S5.F5 "Figure 5 ‣ 5.3.2 Rescaling multivariate distributions ‣ 5.3 Rescaling ‣ 5 Generating annotated AS graphs ‣ Graph Annotations in Modeling Complex Network Topologies").
Next, see box (e) in Figure [5](#S5.F5 "Figure 5 ‣ 5.3.2 Rescaling multivariate distributions ‣ 5.3 Rescaling ‣ 5 Generating annotated AS graphs ‣ Graph Annotations in Modeling Complex Network Topologies"), we compute the empirical copula of distribution $`F\hspace{0pt}{(\mathbf{k})}`$. By definition, the copula of $`F\hspace{0pt}{(\mathbf{k})}`$ is simply the joint distribution of vectors $`\mathbf{u}`$ in eqs. ([9](#S5.E9 "In 5.3.2 Rescaling multivariate distributions ‣ 5.3 Rescaling ‣ 5 Generating annotated AS graphs ‣ Graph Annotations in Modeling Complex Network Topologies"),[11](#S5.E11 "In 5.3.2 Rescaling multivariate distributions ‣ 5.3 Rescaling ‣ 5 Generating annotated AS graphs ‣ Graph Annotations in Modeling Complex Network Topologies")). Therefore, we first compute the marginal CDFs $`F_{1}\hspace{0pt}{(k_{1})}`$, $`F_{2}\hspace{0pt}{(k_{2})}`$, and $`F_{3}\hspace{0pt}{(k_{3})}`$ as CDFs of the first, second, and third components of vectors $`\mathbf{k}_{j}`$:
<table>
<tbody>
<tr>
<td></td>
<td>umj=Fm(kmj)=rmj/N,m=1,2,3,formulae-sequencesuperscriptsubscript𝑢𝑚𝑗subscript𝐹𝑚superscriptsubscript𝑘𝑚𝑗superscriptsubscript𝑟𝑚𝑗𝑁𝑚123u_{m}^{j}=F_{m}(k_{m}^{j})=r_{m}^{j}/N,\quad m=1,2,3,</td>
<td></td>
<td>(12)</td>
</tr>
</tbody>
</table>
where $`r_{m}^{j}`$ is the rank (position number) of value $`k_{m}^{j}`$ in the $`N`$-sized list of values $`k_{m}`$ sorted in the non-decreasing order. Random triplets $`\mathbf{u}_{j} = {({F_{1}\hspace{0pt}{(k_{1}^{j})}},{F_{2}\hspace{0pt}{(k_{2}^{j})}},{F_{3}\hspace{0pt}{(k_{3}^{j})}})}`$ are uniformly distributed in the cube $`{\lbrack 0,1\rbrack}^{3}`$, and their joint CDF $`H\hspace{0pt}{(\mathbf{u})}`$, $`\mathbf{u} = {({F_{1}\hspace{0pt}{(k_{1})}},{F_{2}\hspace{0pt}{(k_{2})}},{F_{3}\hspace{0pt}{(k_{3})}})}`$, is the empirical copula for distribution $`F\hspace{0pt}{(\mathbf{k})}`$, cf. eq. ([9](#S5.E9 "In 5.3.2 Rescaling multivariate distributions ‣ 5.3 Rescaling ‣ 5 Generating annotated AS graphs ‣ Graph Annotations in Modeling Complex Network Topologies")), that describes the correlations among $`k_{1}`$, $`k_{2}`$, and $`k_{3}`$.
At Step 3, box (f) in Figure [5](#S5.F5 "Figure 5 ‣ 5.3.2 Rescaling multivariate distributions ‣ 5.3 Rescaling ‣ 5 Generating annotated AS graphs ‣ Graph Annotations in Modeling Complex Network Topologies"), we merge the rescaled marginals $`D_{m}\hspace{0pt}{(d_{m})}`$, $`m = {1,2,3}`$, from Section [5.3.1](#S5.SS3.SSS1 "5.3.1 Rescaling univariate distributions ‣ 5.3 Rescaling ‣ 5 Generating annotated AS graphs ‣ Graph Annotations in Modeling Complex Network Topologies") and copula $`H\hspace{0pt}{(\mathbf{u})}`$ by computing the target graph degree triplets $`\mathbf{q}_{j} = {(q_{1}^{j},q_{2}^{j},q_{3}^{j})}`$, $`j = {1,\ldots,N}`$, as
<table>
<tbody>
<tr>
<td></td>
<td>qmj=Dm−1(umj),superscriptsubscript𝑞𝑚𝑗superscriptsubscript𝐷𝑚1superscriptsubscript𝑢𝑚𝑗q_{m}^{j}=D_{m}^{-1}(u_{m}^{j}),</td>
<td></td>
<td>(13)</td>
</tr>
</tbody>
</table>
where $`D_{m}^{- 1}`$ are inverse CDFs of $`D_{m}`$ from Section [5.3.1](#S5.SS3.SSS1 "5.3.1 Rescaling univariate distributions ‣ 5.3 Rescaling ‣ 5 Generating annotated AS graphs ‣ Graph Annotations in Modeling Complex Network Topologies"). By construction, the correlation profile of annotation-degree vectors $`\mathbf{q}_{j}`$ is the same as of the ADD in the original topology, while the distributions of their components $`q_{m}^{j}`$ are rescaled ADs.
Algorithm [1](#algorithm1 "In 5.3.2 Rescaling multivariate distributions ‣ 5.3 Rescaling ‣ 5 Generating annotated AS graphs ‣ Graph Annotations in Modeling Complex Network Topologies") lists the described low-level details of our multivariate rescaling, using the ADD as an example.
Input:
Degree vectors $`\mathbf{k}_{i} = {(k_{1}^{i},k_{2}^{i},k_{3}^{i})}`$, $`i = {1\hspace{0pt}\ldots\hspace{0pt}n}`$, of the measured topology;
Input:
Size $`N`$ of the target synthetic topology.
// Step 1: AD rescaling
forall *$`m = {1,2,3}`$*
do
Let $`k_{m}`$ be the list of the $`m^{\text{th}}`$ component values of vectors $`\mathbf{k}_{i}`$;
Approximate distribution $`k_{m}`$ by a smoothing spline $`S_{m}`$;
Sample $`N`$ numbers $`d_{m}^{j}`$, $`j = {1\hspace{0pt}\ldots\hspace{0pt}N}`$, with probability distribution given by $`S_{m}`$;
Let $`D_{m}\hspace{0pt}{(d_{m})}`$ be the CDF of $`d_{m}`$. end
// Step 2: copula rescaling
Re-sample $`N`$ degree triplets $`\mathbf{k}_{j}`$ from the set of $`\mathbf{k}_{i}`$;
forall *$`m = {1,2,3}`$*
do
Let $`k_{m}`$ be the list of the $`m^{\text{th}}`$ component values of vectors $`\mathbf{k}_{j}`$;
Sort list $`k_{m}`$ in the non-degreasing order of values;
forall *$`j = {1\hspace{0pt}\ldots\hspace{0pt}N}`$*
do
Let $`r_{m}^{j}`$ be the position number of value $`k_{m}^{j}`$ in the sorted list;
$`u_{m}^{j} = {r_{m}^{j}/N}`$. end end
// Step 3: merge rescaled ADs and the ADD copula
forall *$`m = {1,2,3}`$*
do
forall *$`j = {1\hspace{0pt}\ldots\hspace{0pt}N}`$*
do
$`q_{m}^{j} = {D_{m}^{- 1}\hspace{0pt}{(u_{m}^{j})}}`$.
end
end
Output:
Degree vectors $`\mathbf{q}_{j} = {(q_{1}^{j},q_{2}^{j},q_{3}^{j})}`$, $`j = {1\hspace{0pt}\ldots\hspace{0pt}N}`$, of the synthetic topology.
Algorithm 1
Rescaling ADD
We conclude our discussion of rescaling with the following remark. Recall from the end of Section [4.2](#S4.SS2 "4.2 The AS relationship annotations ‣ 4 Network topology annotations ‣ Graph Annotations in Modeling Complex Network Topologies") that we have the following two types of multivariate statistics: the ADD and the JDDs. As illustrated in Figure [5](#S5.F5 "Figure 5 ‣ 5.3.2 Rescaling multivariate distributions ‣ 5.3 Rescaling ‣ 5 Generating annotated AS graphs ‣ Graph Annotations in Modeling Complex Network Topologies"), we rescale the ADD using all the three steps described in this section. For rescaling a JDD, it is not necessary to separately rescale its marginals, i.e., to use the first step of the described rescaling process, since the marginals of JDDs are distributions of scalar degrees that we automatically rescale during the ADD rescaling. To rescale a JDD, we execute only the second step of the described rescaling process to obtain the rescaled empirical JDD copula. We then use this copula to determine proper placement of edges in the final synthetic graph that we construct. In other words, the last, third step of our multivariate rescaling process applied to JDDs takes place during the graph construction phase, which we describe next.
### 5.4 Construction
We describe the $`1\hspace{0pt}K`$- and $`2\hspace{0pt}K`$-annotated random graph constructors that are both generalizations of the well-known configuration or pseudograph approach in the terminology of \[[29](#bib.bib29)\]. The $`1\hspace{0pt}K`$-constructor requires only the rescaled ADD, while the $`2\hspace{0pt}K`$-constructor needs also the rescaled JDD copulas.
#### 5.4.1 Constructing $`1\hspace{0pt}K`$-annotated random graphs
Using the rescaled degree vectors $`\mathbf{q}_{j}`$, $`j = {1\hspace{0pt}\ldots\hspace{0pt}N}`$, we construct $`1\hspace{0pt}K`$-annotated random graphs using the following algorithm:
1. 1\.
for each vector $`\mathbf{q}_{j} = {(q_{1}^{j},q_{2}^{j},q_{3}^{j})}`$, prepare a node with $`q_{1}^{j}`$ customer stubs, $`q_{2}^{j}`$ provider stubs, and $`q_{3}^{j}`$ peer stubs;
2. 2\.
randomly select pairs of either customer-and-provider or peer-and-peer stubs, and connect (match) them together to form c2p or p2p links;
3. 3\.
remove unmatched stubs, multiple edges between the same pair of nodes (loops), links with both ends connected to the same node (self-loops), and extract the largest connected component.
The last step deals with the known problem of the pseudograph approach. As its name suggests, it does not necessarily produce simple connected graphs. In general, it generates pseudographs, i.e., graphs with (self-)loops, consisting of several connected components. The size of the largest connected component is usually comparable with the total pseudograph size, while all others are small. Extraction of this largest connected component and removal of all (self-)loops
<sup>6</sup>
<sup>6</sup>6Self-loops are removed, while multiple edges between the same pair of nodes are mapped to a single edge between the two nodes.
alters the target degree distributions. Therefore, the resulting simple connected graph has a slightly different ADD than the one on the algorithm input.
Annotations alleviate this problem since they introduce a series of additional constraints. For example, in the non-annotated case, loops tend to form between high-degree ASs, simply because these ASs have a lot of stubs attached to them after step 1 of the algorithm. In the annotated case, the number of such loops is smaller because most stubs attached to high-degree ASs are annotated as provider stubs that can be matched only with customer stubs attached mostly to low-degree ASs.
Still, the $`1\hspace{0pt}K`$-annotated random graphs are not perfect as, for example, p2p links might end up connecting nodes with drastically dissimilar degree, cf. the JDD discussion in Section [4.2](#S4.SS2 "4.2 The AS relationship annotations ‣ 4 Network topology annotations ‣ Graph Annotations in Modeling Complex Network Topologies"). The $`2\hspace{0pt}K`$-annotated random graphs do not have this problem.
#### 5.4.2 Constructing $`2\hspace{0pt}K`$-annotated random graphs
Earlier work \[[29](#bib.bib29)\] extends the pseudograph approach to non-annotated $`2\hspace{0pt}K`$-distributions. We extend it even further for the $`2\hspace{0pt}K`$-annotated case in the following algorithm:
1. 1\.
for each vector $`\mathbf{q}_{j} = {(q_{1}^{j},q_{2}^{j},q_{3}^{j})}`$, prepare a node with $`q_{1}^{j}`$ customer stubs, $`q_{2}^{j}`$ provider stubs, $`q_{3}^{j}`$ peer stubs, and total degree $`q^{j} = {|\mathbf{q}_{j}|}_{1}`$;
2. 2\.
determine the total numbers $`n_{c\hspace{0pt}2\hspace{0pt}p}`$ and $`n_{p\hspace{0pt}2\hspace{0pt}p}`$ of c2p and p2p edges in the target graph as the maximum possible number of customer-and-provider and peer-and-peer stubs that can be matched within the stub collection $`\mathbf{q}_{j}`$;
3. 3\.
rescale the c2p and p2p JDD copulas
<sup>7</sup>
<sup>7</sup>7See the remark at the end of Section [5.3.2](#S5.SS3.SSS2 "5.3.2 Rescaling multivariate distributions ‣ 5.3 Rescaling ‣ 5 Generating annotated AS graphs ‣ Graph Annotations in Modeling Complex Network Topologies").
to target sizes of $`n_{c\hspace{0pt}2\hspace{0pt}p}`$ and $`n_{p\hspace{0pt}2\hspace{0pt}p}`$ degree pairs $`(q,q')`$ corresponding to c2p and p2p edges between nodes of total degrees $`q`$ and $`q'`$ in the target graph;
4. 4\.
for each c2p (or p2p) degree pair $`(q,q')`$ select randomly a customer (or peer) stub attached to a node of degree $`q`$ and a provider (or peer) stub attached to a node of degree $`q'`$ and form a c2p (or p2p) edge;
5. 5\.
use the procedure described below to rewire (self-)loops;
6. 6\.
remove unmatched stubs, remaining (self-)loops, and extract the largest connected component.
The following rewiring procedure reduces the number of edges removed from the final graph. For each edge involved in a (self-)loop between nodes of degrees $`q_{1}`$ and $`q_{2}`$, we randomly select two non-adjacent nodes of degrees $`q_{1}`$ and $`q_{2}`$ and move the edge to these nodes. This procedure retains a large number of edges that would, otherwise, be removed from the graph. In theory, this procedure may skew the original $`2\hspace{0pt}K`$-summary statistics. In practice, however, it alters these statistics negligibly.
The resulting graph has both the ADD and JDDs approximately the same as those obtained after rescaling. Minor discrepancies are due to the last step of the algorithm, but the number of (self-)loops and small connected components are even smaller than in the $`1\hspace{0pt}K`$-annotated case. The reason for these improvements is yet additional structural constraints, compared with the $`1\hspace{0pt}K`$-annotated case. For example, the JDD-induced constraints force the algorithm to create only one link between a pair of high-degree nodes, or no links between a pair of nodes of degree 1, thus avoiding creation of many connected components composed of such node pairs. The original graph does not have such links, and the rescaled JDDs preserve these structural properties, thus improving the resulting graph quality.
## 6 Evaluation
In this section, we present results of evaluation of our $`2\hspace{0pt}K`$-annotated graph generation method. We also evaluated the $`1\hspace{0pt}K`$-annotated generator and found that, as expected, it produced less accurate graphs with defects such as those mentioned in Section [4.2](#S4.SS2 "4.2 The AS relationship annotations ‣ 4 Network topology annotations ‣ Graph Annotations in Modeling Complex Network Topologies"), e.g., with p2p links connecting ASs of dissimilar degrees, etc.
Experiments. To evaluate the accuracy of our $`2\hspace{0pt}K`$-annotated generator, we want to compare graphs it produces with the measured annotated Internet AS graph from Section [5.2](#S5.SS2 "5.2 Extraction ‣ 5 Generating annotated AS graphs ‣ Graph Annotations in Modeling Complex Network Topologies"). To simplify comparisons, we select one, most representative graph from a set of 50 random synthetic graphs. We select this most representative graph as follows. We first look for a simple graph metric that exhibits high variability across the generated graphs. One such metric is the maximum degree. The expected maximum degree in an $`n`$-node graph with a power-law degree distribution $`{P\hspace{0pt}{(k)}} \sim k^{- \gamma}`$ is $`k_{m\hspace{0pt}a\hspace{0pt}x} \approx n^{1/{({\gamma - 1})}}`$ \[[4](#bib.bib4)\]. Exponent $`\gamma`$ is approximately $`2.1`$ for the Internet AS topology. This value of $`\gamma`$ stays constant as the Internet grows, and it implies almost linear scaling of the maximum degree since $`{1/{({\gamma - 1})}} \approx 0.9`$, which is consistent with scaling of maximum degree in historical Internet topologies \[[36](#bib.bib36)\]. For these reasons, our most representative graph is the one with its maximum degree closest to its expected value, across all the generated graphs.
In addition, we evaluate the variance of important graph metrics described below, across ensembles of random graphs that we generate. Studying the variance properties of a graph generation technique is essential for estimating structural differences between equal-sized random graphs generated by the model, and for gaining insight on how such differences affect performance evaluation experiments. The variance properties of a graph generation technique is associated with the following tradeoff. On the one hand, variance should be small so that generated graphs closely match the observed topology. On the other hand, though, random graphs should not all be identical or almost identical, because if they do not exhibit sufficient structural diversity, then they have little value for performance evaluation studies. In our experiments, we compute and report the variance of important graph metrics in sets of 50 equal-sized random graphs.
Metrics. Since it is practically impossible to compare graphs over every existing graph metric, we select a set of metrics that were found particularly important in the Internet topology literature. These metrics include the degree distributions that we deal with in previous sections, assortativity coefficient, distance distribution, and spectrum. The assortativity coefficient is essentially the Pearson correlation coefficient of the joint degree distribution (JDD). Its positive (negative) values indicate that degrees of connected nodes are positively (negatively) correlated, meaning that nodes with similar (dissimilar) degrees interconnect with higher probabilities. The distance distribution is the distribution of lengths of the shortest paths in a graph, which we compute both with and without constraints imposed by annotations (routing policies). The spectrum of a graph is the set of the eigenvalues of its Laplacian $`L`$. The Laplacian’s matrix elements $`L_{i\hspace{0pt}j}`$ are $`- {1/{({k_{i}\hspace{0pt}k_{j}})}^{1/2}}`$ if there is an edge between node $`i`$ of degree $`k_{i}`$ and node $`j`$ of degree $`k_{j}`$; 1 if $`i = j`$; and 0 otherwise. Among the $`n`$ eigenvalues of $`L`$, the smallest non-zero and largest eigenvalues are most interesting, since they provide tight bounds to a number of important network properties. For more details on these and other metrics, and why they are important, see \[[30](#bib.bib30)\].
Results. In Figure [6](#S6.F6 "Figure 6 ‣ 6 Evaluation ‣ Graph Annotations in Modeling Complex Network Topologies") we plot the ADs of the measured AS topology and of the most representative synthetic graph of the equal size. We observe that the distributions of the customer, provider, and peer degrees in the synthetic graph are very close to the corresponding distributions in the measured topology. The close match demonstrates that: 1) spline-smoothing accurately models complex ADs of real Internet topologies, 2) random number generation yields empirical distributions that follow the modeled distributions, and 3) rewiring and removal of (self-)loops do not introduce any significant artifacts. It is, of course, expected that our generator accurately reproduces ADs, as they are part of the $`2\hspace{0pt}K`$-summary statistics we explicitly model. We also confirm that synthetic graphs, also as expected, closely reproduce all the other summary statistics that we explicitly model: the DD, ADD, and JDDs of the synthetic graph are very close to the originals. We do not show the corresponding plots for brevity.
[Refer to caption](/html/0708.3879/assets/x10.png)
Figure 6: CCDFs of the number of customer, provider, and peer stubs in the synthetic versus measured AS topology.
In Figures [7(a)](#S6.F7.sf1 "In Figure 7 ‣ 6 Evaluation ‣ Graph Annotations in Modeling Complex Network Topologies") and [7(b)](#S6.F7.sf2 "In Figure 7 ‣ 6 Evaluation ‣ Graph Annotations in Modeling Complex Network Topologies") we compare the distance distributions of the measured and equal-sized synthetic topology ignoring and accounting for annotation-induced, i.e., routing policy, constraints. In the former case, we calculate lengths of the standard shortest paths between nodes in a graph as if the graph was non-annotated. In the latter case, we find lengths of shortest valid, i.e., valley-free, paths defined in Section [3](#S3 "3 AS relationships and why they matter ‣ Graph Annotations in Modeling Complex Network Topologies"). In both the non-annotated and annotated cases, we observe that the distance distribution in the synthetic graph closely matches the distance distribution in the measured topology, even though we have not explicitly modeled or tried to reproduce the distance distributions. We also observe that the distance distributions in the non-annotated and annotated cases are different, meaning that annotations in the synthetic graph properly filter realistic, policy-constrained paths from the set of all possible path in the non-annotated case.
[Refer to caption](/html/0708.3879/assets/x11.png)
\(a\)
Shortest paths.
[Refer to caption](/html/0708.3879/assets/x12.png)
\(b\)
Shortest valid paths.
Figure 7: Distance distributions in synthetic and measured AS topologies.
In Table [3](#S6.T3 "Table 3 ‣ 6 Evaluation ‣ Graph Annotations in Modeling Complex Network Topologies") we compare the measured topology with synthetic graphs of different sizes over a set of important scalar metrics, including those we do not explicitly model or try to reproduce, e.g., the eigenvalues of the Laplacian, etc. We compute these metrics for five synthetic graphs of sizes $`5,000`$, $`10,000`$, $`30,000`$, and $`19,036`$ nodes, the last size being equal to the size of the original topology. The first three metrics are the number of (c2p or p2p) edges in a graph. We observe that the number of such edges grows almost linearly with the number of nodes. This observation is consistent with that the average degree in historical Internet topologies stays almost constant \[[36](#bib.bib36)\]. The fourth metric is the maximum degree. As expected, the maximum degree grows with the size of the graph slightly slower than linearly. The next five metrics in the table describe properties that have stayed relatively constant in historical Internet topologies. These properties have small variations in the synthetic graphs as well.
Table 3: Scalar metrics of synthetic and collected graphs. Note that smallest eigenvalues are positive, but some may round to zero.
<table>
<tbody>
<tr>
<th></th>
<td colspan="5">Number of nodes</td>
</tr>
<tr>
<th rowspan="2">Metric</th>
<td>19036</td>
<td>5000</td>
<td>10000</td>
<td>19036</td>
<td>30000</td>
</tr>
<tr>
<td>(measured)</td>
<td>(synthetic)</td>
<td>(synthetic)</td>
<td>(synthetic)</td>
<td>(synthetic)</td>
</tr>
<tr>
<th>Number of edges</th>
<td>40115</td>
<td>10179</td>
<td>20730</td>
<td>39595</td>
<td>62853</td>
</tr>
<tr>
<th>Number of c2p edges</th>
<td>36188</td>
<td>9409</td>
<td>18917</td>
<td>36146</td>
<td>56870</td>
</tr>
<tr>
<th>Number of p2p edges</th>
<td>3927</td>
<td>770</td>
<td>1813</td>
<td>3448</td>
<td>5983</td>
</tr>
<tr>
<th>Maximum degree</th>
<td>2384</td>
<td>1014</td>
<td>1492</td>
<td>2385</td>
<td>3461</td>
</tr>
<tr>
<th>Average degree</th>
<td>4.21</td>
<td>4.07</td>
<td>4.15</td>
<td>4.16</td>
<td>4.19</td>
</tr>
<tr>
<th>Assortativity coefficient</th>
<td>-0.20</td>
<td>-0.30</td>
<td>-0.24</td>
<td>-0.25</td>
<td>-0.18</td>
</tr>
<tr>
<th>Largest eigenvalue of Laplacian</th>
<td>1.97</td>
<td>1.85</td>
<td>1.88</td>
<td>1.91</td>
<td>1.92</td>
</tr>
<tr>
<th>Smallest eigenvalue of Laplacian</th>
<td>0.03</td>
<td>0.00</td>
<td>0.12</td>
<td>0.09</td>
<td>0.00</td>
</tr>
<tr>
<th>Average distance</th>
<td>3.76</td>
<td>3.26</td>
<td>3.52</td>
<td>3.57</td>
<td>3.75</td>
</tr>
</tbody>
</table>
Next we investigate the benefit of modeling the ADs and ADD in addition to the DD and JDDs. Previous work \[[29](#bib.bib29)\] shows that modeling DD and JDD is sufficient for capturing and reliably reproducing most important non-annotated graph metrics. The main value of modeling ADs and ADD is that our generated synthetic graphs are properly annotated. We saw in Section [3](#S3 "3 AS relationships and why they matter ‣ Graph Annotations in Modeling Complex Network Topologies") that the Internet topology annotations are important. Here we provide another evidence that they are non-trivial. Specifically, in Figures [8(a)](#S6.F8.sf1 "In Figure 8 ‣ 6 Evaluation ‣ Graph Annotations in Modeling Complex Network Topologies"), [8(b)](#S6.F8.sf2 "In Figure 8 ‣ 6 Evaluation ‣ Graph Annotations in Modeling Complex Network Topologies"), and [8(c)](#S6.F8.sf3 "In Figure 8 ‣ 6 Evaluation ‣ Graph Annotations in Modeling Complex Network Topologies") we plot the total degrees of ASs in the measured AS topology versus their annotation degrees: the number of customers $`k_{1}`$, providers $`k_{2}`$, and peers $`k_{3}`$, respectively. We observe that a given total degree can correspond to a wide range of different values of $`k_{1}`$, $`k_{2}`$, and $`k_{3}`$. The JDD provides information only on total node degrees and on their correlations, whereas it is completely agnostic to annotation degrees. On the other hand, the ADs and ADD capture the distribution of annotation degrees and the correlations between annotation degrees, respectively. Therefore, the JDD alone is in principle incapable of capturing topology annotations, while the benefit of modeling ADs and ADD lies in reproducing realistic annotations in generated graphs.
[Refer to caption](/html/0708.3879/assets/x13.png)
\(a\)
Total degree vs. number of customers.
[Refer to caption](/html/0708.3879/assets/x14.png)
\(b\)
Total degree vs. number of providers.
[Refer to caption](/html/0708.3879/assets/x15.png)
\(c\)
Total degree vs. number of peers.
Figure 8: Scatterplots demonstrating the diversity of annotation degrees and that total node degrees are agnostic with respect to annotations.
To quantify the variance properties of randomly generated graphs, we compute the standard deviation of our metrics across sets of 50 random graphs. We construct 4 sets with topologies of 5,000, 10,000, 19,036, and 30,000 nodes, a total of 200 random graphs. Among our evaluation metrics, we do not compute the eigenvalues of the Laplacian and the assortativity coefficient, since they require prohibitively long computation times for 200 graphs. In Table [4](#S6.T4 "Table 4 ‣ 6 Evaluation ‣ Graph Annotations in Modeling Complex Network Topologies"), we show the standard deviation and mean value of the remaining metrics. The maximum degree exhibits the highest standard deviation (with respect to the mean) taking values between 376 and 471 for graphs of different size. The high variance of the maximum degree is expected, since the degree distribution of Internet topologies is highly skewed. On the other hand, the remaining metrics in Table [4](#S6.T4 "Table 4 ‣ 6 Evaluation ‣ Graph Annotations in Modeling Complex Network Topologies") exhibit low variance. These metrics reflect aggregate graph properties and can be modeled as a sum of many i.i.d. random variables. Therefore, according to the central limit theorem, their distribution is approximately normal and their variance is consequently smaller than the variance of the maximum degree.
An important difference between the graph generation method described in this study and the graph generation methods described in our previous work \[[29](#bib.bib29)\] is that the former exhibits higher variance. The two methods are conceptually similar in generating synthetic graphs that reproduce the correlation profile of an observed topology—albeit \[[29](#bib.bib29)\] does not consider annotations. They differ in that our previous techniques directly use the degree distribution or correlations of an observed topology to generate new similar topologies. On the other hand, the present work first models the degree correlations of a topology and then uses random number generators to produce synthetic degree distributions fed into final graph constructors. In simpler words, our present technique induces more randomness by means of the synthetic generation of degree correlations based on the correlation profile extracted from the real topology. The two approaches are complementary and together provide a wider range of options for generating synthetic topologies with desired variance characteristics.
Table 4: Variance of graph metrics across sets of equal-sized random graphs.
<table>
<tbody>
<tr>
<th rowspan="2">graph metric</th>
<th colspan="4">std. deviation / mean</th>
</tr>
<tr>
<th>5000</th>
<th>10000</th>
<th>19036</th>
<th>30000</th>
</tr>
<tr>
<td>nodes</td>
<td>nodes</td>
<td>nodes</td>
<td>nodes</td>
<td></td>
</tr>
<tr>
<td>c2p edges</td>
<td>399/17,905</td>
<td>267/36,876</td>
<td>349/71,055</td>
<td>410/112,549</td>
</tr>
<tr>
<td>p2p edges</td>
<td>100/1,238</td>
<td>203/2,980</td>
<td>396/6,412</td>
<td>541/12,863</td>
</tr>
<tr>
<td>max degree</td>
<td>387/1,618</td>
<td>417/2,090</td>
<td>471/2,335</td>
<td>376/2,599</td>
</tr>
<tr>
<td>av. degree</td>
<td>0.08/3.83</td>
<td>0.03/3.99</td>
<td>0.02/4.07</td>
<td>0.02/4.11</td>
</tr>
<tr>
<td>av. distance</td>
<td>0.13/3.16</td>
<td>0.09/3.40</td>
<td>0.10/3.61</td>
<td>0.06/3.77</td>
</tr>
</tbody>
</table>
Overall, our evaluation results show that:
- •
$`2\hspace{0pt}K`$-annotated random graphs generated with our approach faithfully reproduce a number of important properties of Internet topologies;
- •
rescaled graphs exhibit the expected behavior according to a number of definitive graph metrics, i.e., these metrics are either properly-rescaled or stay relatively stable as the size of synthetic graphs varies;
- •
the profile of correlations between annotation and total degrees is diverse; and
- •
random graphs generated with our method exhibit small variance, although higher than in our previous work \[[29](#bib.bib29)\].
## 7 Conclusions
In this work, we have focused on the problem of generating synthetic annotated graphs that model real complex networks. Our techniques are likely to have many applications not only in networking, but also in other disciplines where annotated graphs are used to abstract and represent network structure. For example, two groups have recently contacted us to discuss our techniques as they were searching for tools to generate synthetic, semantic-rich, i.e., annotated, networks for their simulation studies. The first group works on modeling the European powerline networks, while the second is in brain and neural network research. Other networks to which our techniques are immediately applicable include the router-level Internet, WWW, networks of critical resources dependencies, as well as many types of social and biological networks, such as regulatory pathways \[[35](#bib.bib35)\].
A number of open problems remain. In particular, our techniques construct synthetic versions of real topologies available from measurement projects. However, it is well-known that in many cases, the outcome of measurements does not accurately represent a real complete topology. In fact, there might exist inherent limitations in measuring certain network topologies with 100% accuracy. A venue for further research is the development of prediction techniques that extrapolate what we can presently measure in order to predict what we can not measure.
Another substantial problem is the difficulties in validating results of topology inference studies. For example, in the specific context of Internet topologies, validation is hard because of the unwillingness of service providers to release data on their infrastructure, network design, configuration, and performance. On the other hand, validation of any research result is a cornerstone to its reliability and utility. Therefore we believe it is imperative to focus on new validation techniques that would combine the limited ground truth data available today with convincing testbed or simulation experiments.
## 8 Acknowledgments
This work was supported in part by NSF CNS-0434996 and CNS-0722070, by DHS N66001-08-C-2029, and by Cisco Systems.
## References
- \[1\] W. Aiello, F. Chung, and L. Lu. A random graph model for massive graphs.
In STOC, 2000.
- \[2\] R. Albert and A.-L. Barabási. Topology of evolving networks: Local events and universality.
Phys Rev Lett, 85(24):5234–5237, 2000.
- \[3\] M. Boguñá and R. Pastor-Satorras. Class of correlated random networks with hidden variables.
Phys Rev E, 68:036112, 2003.
- \[4\] M. Boguñá, R. Pastor-Satorras, and A. Vespignani. Cut-offs and finite size effects in scale-free networks.
Eur Phys J B, 38:205–209, 2004.
- \[5\] M. Boguñá and M. A. Serrano. Generalized percolation in random directed networks.
Phys Rev E, 72:016106, 2005.
- \[6\] J. M. Carlson and J. Doyle. Highly optimized tolerance: A mechanism for power-laws in designed systems.
Phys Rev E, 60:1412–1427, 1999.
- \[7\] J. M. Chambers and T. J. Hastie.
Statistical Models in S.
Wadsworth and Brooks/Cole, Pacific Grove, California, 1992.
- \[8\] H. Chang, S. Jamin, and W. Willinger. To peer or not to peer: Modeling the evolution of the Internet’s AS-level topology.
In INFOCOM, 2006.
- \[9\] F. Chung and L. Lu. Connected components in random graphs with given degree sequences.
Ann Comb, 6:125–145, 2002.
- \[10\] A. Clauset and C. Moore. Accuracy and scaling phenomena in Internet mapping.
Phys Rev Lett, 94:018701, 2005.
- \[11\] L. Dall’Asta, I. Alvarez-Hamelin, A. Barrat, A. Vázquez, and A. Vespignani. Exploring networks with traceroute-like probes: Theory and simulations.
Theor Comput Sci, Complex Networks, 355(1):6–24, 2006.
- \[12\] C. de Launois. GHITLE: Generator of Hierarchical Internet Topologies using LEvels.
http://ghitle.info.ucl.ac.be/.
- \[13\] X. Dimitropoulos, D. Krioukov, M. Fomenkov, B. Huffaker, Y. Hyun, kc claffy, and G. Riley. AS relationships: Inference and validation.
Comput Commun Rev, 37(1), 2007.
- \[14\] X. Dimitropoulos, D. Krioukov, B. Huffaker, kc claffy, and G. Riley. Inferring AS relationships: Dead end or lively beginning?
In WEA, 2005.
- \[15\] X. Dimitropoulos, D. Krioukov, G. Riley, and kc claffy. Revealing the Autonomous System taxonomy: The machine learning approach.
In PAM, 2006.
- \[16\] X. Dimitropoulos and G. Riley. Creating realistic BGP models.
In MASCOTS, 2003.
- \[17\] M. Doar. A better model for generating test networks.
In GLOBECOM, 1996.
- \[18\] P. Erdős and A. Rényi. On random graphs.
Publ Math, 6:290–297, 1959.
- \[19\] M. Faloutsos, P. Faloutsos, and C. Faloutsos. On power-law relationships of the Internet topology.
In SIGCOMM, 1999.
- \[20\] O. Frank and D. Strauss. Markov graphs.
J Am Stat Assoc, 81(395):832–842, 1986.
- \[21\] L. Gao. On inferring Autonomous System relationships in the Internet.
IEEE ACM T Network, 9(6), 2001.
- \[22\] L. Gao and F. Wang. The extent of AS path inflation by routing policies.
In Global Internet, 2002.
- \[23\] C. Gkantsidis, M. Mihail, and E. Zegura. The Markov chain simulation method for generating connected power law random graphs.
In ALENEX, 2003.
- \[24\] E. Hewitt.
Real and Abstract Analysis: A Modern Treatment of the Theory of Functions of a Real Variable.
Springer Verlag, 1975.
- \[25\] W. Hörmann, J. Leydold, and G. Derflinger.
Automatic Nonuniform Random Variate Generation.
Springer-Verlag, Berlin, 2004.
- \[26\] D. Krioukov, kc claffy, K. Fall, and A. Brady. On compact routing for the Internet.
Comput Commun Rev, 37(3):41–52, 2007.
- \[27\] A. Lakhina, J. Byers, M. Crovella, and P. Xie. Sampling biases in IP topology measurements.
In INFOCOM, 2003.
- \[28\] L. Li, D. Alderson, W. Willinger, and J. Doyle. A first-principles approach to understanding the Internet s router-level topology.
In SIGCOMM, 2004.
- \[29\] P. Mahadevan, D. Krioukov, K. Fall, and A. Vahdat. Systematic topology analysis and generation using degree correlations.
In SIGCOMM, 2006.
- \[30\] P. Mahadevan, D. Krioukov, M. Fomenkov, B. Huffaker, X. Dimitropoulos, kc claffy, and A. Vahdat. The Internet AS-level topology: Three data sources and one definitive metric.
Comput Commun Rev, 36(1):17–26, 2006.
- \[31\] S. Maslov, K. Sneppen, and A. Zaliznyak. Detection of topological patterns in complex networks: Correlation profile of the Internet.
Physica A, 333:529–540, 2004.
- \[32\] A. Medina, A. Lakhina, I. Matta, and J. Byers. BRITE: An approach to universal topology generation.
In MASCOTS, 2001.
- \[33\] R. B. Nelson. An introduction to copulas.
Springer-Verlag Lecture Notes in Statistics, 139:216, 1999.
- \[34\] C. Palmer and G. Steffan. Generating network topologies that obey power laws.
In GLOBECOM, November 2000.
- \[35\] J. Pandey, M. Koyuturk, Y. Kim, W. Szpankowski, S. Subramaniam, and A. Grama. Functional annotation of regulatory pathways.
Bioinformatics, 23(13):i377–i386, 2007.
- \[36\] R. Pastor-Satorras and A. Vespignani.
Evolution and Structure of the Internet: A Statistical Physics Approach.
Cambridge University Press, Cambridge, 2004.
- \[37\] The R Project for Statistical Computing.
http://www.r-project.org/.
- \[38\] University of Oregon RouteViews Project.
http://www.routeviews.org/.
- \[39\] A. Sklar. Fonctions de répartition à $`n`$ dimensions et leures marges.
Publications de l’Institut de Statistique de L’Université de Paris, 8:229 231, 1959.
- \[40\] B. Söderberg. A general formalism for inhomogeneous random graphs.
Phys Rev E, 66:066121, 2002.
- \[41\] B. Söderberg. Properties of random graphs with hidden color.
Phys Rev E, 68:026107, 2003.
- \[42\] B. Söderberg. Random graphs with hidden color.
Phys Rev E, 68:015102(R), 2003.
- \[43\] B. Söderberg. Random Feynman graphs.
In AIP Conf Procs, volume 776, pages 118–132, 2005.
- \[44\] B. M. Waxman. Routing of multipoint connections.
JSAC, 1988.
- \[45\] J. Winick and S. Jamin. Inet-3.0: Internet topology generator. Technical Report UM-CSE-TR-456-02, University of Michigan, 2002.
- \[46\] E. W. Zegura, K. Calvert, and S. Bhattacharjee. How to model an internetwork.
In INFOCOM, 1996. |
0707.3936 | cs.NI | cyber | INSTITUT NATIONAL DE RECHERCHE EN INFORMATIQUE ET EN AUTOMATIQUE
Closed form solutions for symmetric
water filling games
Eitan Altman — Konstantin Avrachenkov — Andrey Garnaev N° 6254
July 2007
Closed form solutions for symmetric
water filling games
Eitan Altman
<sup>†</sup>
<sup>†</sup>thanks: INRIA Sophia Antipolis, Altman@sophia.inria.fr
, Konstantin Avrachenkov
<sup>†</sup>
<sup>†</sup>thanks: INRIA Sophia Antipolis, K.Avrachenkov@sophia.inria.fr
, Andrey Garnaev
<sup>†</sup>
<sup>†</sup>thanks: St.Petersburg State University, agarnaev@rambler.ru
Thème COM — Systèmes communicants
Projets MAESTRO
Rapport de recherche n° 6254 — July 2007 — ?? pages
<sup>0</sup>
<sup>0</sup>footnotetext: This work was supported by BioNets European project and by joint RFBR and NNSF Grant no.06-01-39005.
Abstract: We study power control in optimization and game frameworks. In the optimization framework there is a single decision maker who assigns network resources and in the game framework users share the network resources according to Nash equilibrium. The solution of these problems is based on so-called water-filling technique, which in turn uses bisection method for solution of non-linear equations for Lagrange multiplies. Here we provide a closed form solution to the water-filling problem, which allows us to solve it in a finite number of operations. Also, we produce a closed form solution for the Nash equilibrium in symmetric Gaussian interference game with an arbitrary number of users. Even though the game is symmetric, there is an intrinsic hierarchical structure induced by the quantity of the resources available to the users. We use this hierarchical structure to perform a successive reduction of the game. In addition, to its mathematical beauty, the explicit solution allows one to study limiting cases when the crosstalk coefficient is either small or large. We provide an alternative simple proof of the convergence of the Iterative Water Filling Algorithm. Furthermore, it turns out that the convergence of Iterative Water Filling Algorithm slows down when the crosstalk coefficient is large. Using the closed form solution, we can avoid this problem. Finally, we compare the non-cooperative approach with the cooperative approach and show that the non-cooperative approach results in a more fair resource distribution.
Key-words: wireless networks, power control, symmetric water-filling game, Nash equilibrium, price of anarchy
Solution analytique des jeux de water-filling symétriques
Résumé : Nous étudions le contrôle de puissance dans le cadre de l’optimisation et dans celui de la théorie des jeux. Dans le premier, il y a un seul agent qui assigne les ressources du réseau tandis que dans le deuxième, les utilisateurs se partagent les ressources du réseau selon l’équilibre de Nash. La solution de ces problèmes est basée sur la méthode du water-filling. On calcule des multiplicateurs de Lagrange en utilisant une méthode de bisection pour resoudre des équations non linéaires. Nous fournissons ici une solution analytique au problème du water-filling, qui nous permet de le résoudre en un nombre fini d’opérations. En outre, nous produisons une solution analytique de l’équilibre de Nash dans le cadre de la théorie des jeux. Nous étudions un jeu symétrique en terme d’interférence avec un nombre arbitraire d’utilisateurs. Quoique le jeu soit symétrique, il y a une structure hiérarchique induite par la quantité des ressources disponibles pour les utilisateurs. Nous utilisons cette structure pour effectuer une réduction successive du jeu. En plus de son éléguence mathématique, la solution analytique permet d’étudier des cas limites quand le coefficient d’interférence est petit ou grand. Nous fournissons une preuve simple de la convergence de l’algorithme itératif de water-filling (l’algorithme de meilleur réponse). Il s’avère que la convergence de l’agorithme est ralentie quand le coefficient d’interférence est proche de l’unité. En utilisant la solution analytique, nous pouvons éviter ce problème. Aussi, nous comparons l’approche non coopérative à l’approche coopérative et montrons que l’approche non coopérative fournit une distribution des ressources plus équitable.
Mots-clés : réseaux sans fils, contrôle de puissance, jeu water-filling symétrique, équilibre de Nash, coût de l’anarchie
## 1 Introduction
In wireless networks and DSL access networks the total available power for signal transmission has to be distributed among several resources. In the context of wireless networks, the resources may correspond to frequency bands (e.g. as in OFDM), or they may correspond to capacity available at different time slots. In the context of DSL access networks, the resources correspond to available frequency tones. This spectrum of problems can be considered in either optimization scenario or game scenario. The optimization scenario leads to “Water Filling Optimization Problem” \[[3](#bib.bib3), [6](#bib.bib6), [14](#bib.bib14)\] and the game scenario leads to “Water Filling Game” or “Gaussian Interference Game” \[[8](#bib.bib8), [11](#bib.bib11), [12](#bib.bib12), [15](#bib.bib15)\]. In the optimization scenario, one needs to maximize a concave function (Shannon capacity) subject to power constraints. The Lagrange multiplier corresponding to the power constraint is determined by a non-linear equation. In the previous works \[[3](#bib.bib3), [6](#bib.bib6), [14](#bib.bib14)\], it was suggested to find the Lagrange multiplier by means of a bisection algorithm, where comes the name “Water Filling Problem”. Here we show that the Lagrange multiplier and hence the optimal solution of the water filling problem can be found in explicit form with a finite number of operations. In the multiuser context, one can view the problem in either cooperative or non-cooperative setting. If a centralized controller wants to maximize the sum of all users’ rates, the controller will face a non-convex optimization problem with many local maxima \[[13](#bib.bib13)\]. On the other hand, in the non-cooperative setting, the power allocation problem becomes a game problem where each user perceives the signals of the other users as interference and maximizes a concave function of the noise to interference ratio. A natural approach in the non-cooperative setting is the application of the Iterative Water Filling Algorithm (IWFA) \[[16](#bib.bib16)\]. Recently, the authors of \[[10](#bib.bib10)\] proved the convergence of IWFA under fairly general conditions. In the present work we study the case of symmetric water filling game. There is an intrinsic hierarchical structure induced by the quantity of the resources available to the users. We use this hierarchical structure to perform a successive reduction of the game, which allows us to find Nash equilibrium in explicit form. In addition, to its mathematical beauty, the explicit solution allows one to find the Nash equilibrium in water filling game in a finite number of operations and to study limiting cases when the crosstalk coefficient is either small or large. As a by-product, we obtain an alternative simple proof of the convergence of the Iterative Water Filling Algorithm. Furthermore, it turns out that the convergence of IWFA slows down when the crosstalk coefficient is large. Using the closed form solution, we can avoid this problem. Finally, we compare the non-cooperative approach with the cooperative approach and conclude that the cost of anarchy is small in the case of small crosstalk coefficients and that the the decentralized solution is better than the centralized one with respect to fairness. Applications that can mostly benefit from decentralized non-cooperative power control are ad-hoc and sensor networks with no predefined base stations \[[4](#bib.bib4), [9](#bib.bib9), [7](#bib.bib7)\]. An interested reader can find more references on non-cooperative power control in \[[2](#bib.bib2), [8](#bib.bib8)\]. We would like to mention that the water filling problem and jamming games with transmission costs have been analyzed in \[[1](#bib.bib1)\].
The paper is organized as follows: In Section [2](#S2 "2 Single decision maker") we recall the single decision maker setup of the water filling optimization problem and provide its explicit solution. Then in Sections [3](#S3 "3 Symmetric water filling game")-[7](#S7 "7 Closed form solution for 𝐿 person game") we formulate multiuser symmetric water filling game and characterize its Nash equilibrium, also we give an alternative simple proof of the convergence of the iterative water filling algorithm and suggest the explicit form of the users’ strategy in the Nash equilibrium. In Section [8](#S8 "8 Numerical examples") we confirm our finding with the help of numerical examples and compare the decentralized approach with the centralized one.
## 2 Single decision maker
First let us consider the power allocation problem in the case of a single decision maker. The single decision maker (also called “user” or “transmitter”) wants to send information using $`n`$ independent resources so that to maximize the Shannon capacity. We further assume that resource $`i`$ has a “weight” of $`\pi_{i}`$.
Possible interpretations:
- \(i\)
The resources may correspond to capacity available at different time slots; we assume that there is a varying environment whose state changes among a finite set of states $`i \in {\lbrack 1,n\rbrack}`$, according to some ergodic stochastic process with stationary distribution $`{\{\pi_{i}\}}_{i = 1}^{n}`$. We assume that the user has perfect knowledge of the environment state at the beginning of each time slot.
- \(ii\)
The resources may correspond to frequency bands (e.g. as in OFDM) where one should assign different power levels for different sub-carriers \[[14](#bib.bib14)\]. In that case we may take $`\pi_{i} = {1/n}`$ for all $`i`$.
The strategy of user is $`T = {(T_{1},\ldots,T_{n})}`$ with $`{\sum_{i = 1}^{n}{\pi_{i}\hspace{0pt}T_{i}}} = \overline{T}`$, $`{T_{i} \geq 0},{\pi_{i} > 0}`$ for $`i \in {\lbrack 1,n\rbrack}`$ and $`\overline{T} > 0`$. As the payoff to user we take the Shannon capacity
<table>
<tbody>
<tr>
<td></td>
<td>v(T)=∑i=1nπiln(1+Ti/Ni0),𝑣𝑇superscriptsubscript𝑖1𝑛subscript𝜋𝑖1subscript𝑇𝑖subscriptsuperscript𝑁0𝑖v(T)=\sum_{i=1}^{n}\pi_{i}\ln\left(1+T_{i}/N^{0}_{i}\right),</td>
<td></td>
</tr>
</tbody>
</table>
where $`N_{i}^{0} > 0`$ is the noise level in the sub-carrier $`i`$.
We would like to emphasize that this generalized description of the water-filling problem can be used for power allocation in time as well as power allocation in space-frequency. Following the standard water-filling approach \[[3](#bib.bib3), [6](#bib.bib6), [14](#bib.bib14)\] we have the following result.
######
Theorem 1
Let $`{T_{i}\hspace{0pt}{(\omega)}} = \left\lbrack {{1/\omega} - N_{i}^{0}} \right\rbrack_{+}`$ for $`i \in {\lbrack 1,n\rbrack}`$ and $`{H_{T}\hspace{0pt}{(\omega)}} = {\sum_{i = 1}^{n}{\pi_{i}\hspace{0pt}T_{i}\hspace{0pt}{(\omega)}}}`$. Then $`{T\hspace{0pt}{(\omega^{\ast})}} = {({T_{1}\hspace{0pt}{(\omega^{\ast})}},\ldots,{T_{n}\hspace{0pt}{(\omega^{\ast})}})}`$ is the unique optimal strategy and its payoff is $`v\hspace{0pt}{({T\hspace{0pt}{(\omega^{\ast})}})}`$ where $`\omega^{\ast}`$ is the unique root of the equation
<table>
<tbody>
<tr>
<td></td>
<td>H(ω)=T¯.𝐻𝜔¯𝑇H(\omega)=\bar{T}.</td>
<td></td>
<td>(1)</td>
</tr>
</tbody>
</table>
In the previous studies of the water-filling problems it was suggested to use numerical (e.g., bisection) method to solve the equation ([1](#S2.E1 "In Theorem 1 ‣ 2 Single decision maker")). Here we propose an explicit form approach for its solution.
Without loss of generality we can assume that
<table>
<tbody>
<tr>
<td></td>
<td>1/N10≥1/N20≥…≥1/Nn0.1subscriptsuperscript𝑁011subscriptsuperscript𝑁02…1subscriptsuperscript𝑁0𝑛1/N^{0}_{1}\geq 1/N^{0}_{2}\geq\ldots\geq 1/N^{0}_{n}.</td>
<td></td>
<td>(2)</td>
</tr>
</tbody>
</table>
Then, since $`H\hspace{0pt}{( \cdot )}`$ is decreasing, we have the following result:
######
Theorem 2
The solution of the water-filling optimization problem is given by
<table>
<tbody>
<tr>
<td></td>
<td>Ti∗={(T¯+∑t=1kπt(Nt0−Ni0))/(∑t=1kπt),i≤k,0,i>k,superscriptsubscript𝑇𝑖cases¯𝑇superscriptsubscript𝑡1𝑘subscript𝜋𝑡subscriptsuperscript𝑁0𝑡subscriptsuperscript𝑁0𝑖superscriptsubscript𝑡1𝑘subscript𝜋𝑡𝑖𝑘0𝑖𝑘T_{i}^{*}=\begin{cases}\Bigl{(}\displaystyle\bar{T}+\sum_{t=1}^{k}\pi_{t}(N^{0}_{t}-N^{0}_{i})\Bigr{)}\Bigl{/}\Bigl{(}\displaystyle\sum_{t=1}^{k}\pi_{t}\Bigr{)},&i\leq k,\\ 0,&i>k,\end{cases}</td>
<td></td>
</tr>
</tbody>
</table>
where $`k`$ can be found from the following condition:
<table>
<tbody>
<tr>
<td></td>
<td>φk<T¯≤φk+1,subscript𝜑𝑘¯𝑇subscript𝜑𝑘1\varphi_{k}<\bar{T}\leq\varphi_{k+1},</td>
<td></td>
</tr>
</tbody>
</table>
where
<table>
<tbody>
<tr>
<td></td>
<td>φt=∑i=1tπi(Nt0−Ni0) for t∈[1,n].subscript𝜑𝑡superscriptsubscript𝑖1𝑡subscript𝜋𝑖subscriptsuperscript𝑁0𝑡subscriptsuperscript𝑁0𝑖 for 𝑡1𝑛\varphi_{t}=\sum_{i=1}^{t}\pi_{i}(N^{0}_{t}-N^{0}_{i})\mbox{ for }t\in[1,n].</td>
<td></td>
</tr>
</tbody>
</table>
Thus, contrary to the numerical (bisection) approach, in order to find an optimal resource allocation we need to execute only a finite number of operations.
## 3 Symmetric water filling game
Let us now consider a multi-user scenario. Specifically, we consider $`L`$ users who try to send information through $`n`$ resources so that to maximize their transmission rates. The strategy of user $`j`$ is $`T^{j} = {(T_{1}^{j},\ldots,T_{n}^{j})}`$ subject to
<table>
<tbody>
<tr>
<td></td>
<td>∑i=1nπiTij=Tj¯,superscriptsubscript𝑖1𝑛subscript𝜋𝑖subscriptsuperscript𝑇𝑗𝑖¯superscript𝑇𝑗\sum_{i=1}^{n}\pi_{i}T^{j}_{i}=\bar{T^{j}},</td>
<td></td>
<td>(3)</td>
</tr>
</tbody>
</table>
where $`\overline{T^{j}} > 0`$ for $`j \in {\lbrack 1,L\rbrack}`$. The element $`T_{i}^{j}`$ is the power level used by transmitter $`j`$ when the environment is in state $`i`$. The payoff to user $`j`$ is given as follows:
<table>
<tbody>
<tr>
<td></td>
<td>vj(T1,…,TL)=∑i=1nπiln(1+αijTijNi0+gi∑k≠jαikTik),superscript𝑣𝑗superscript𝑇1…superscript𝑇𝐿superscriptsubscript𝑖1𝑛subscript𝜋𝑖1subscriptsuperscript𝛼𝑗𝑖subscriptsuperscript𝑇𝑗𝑖subscriptsuperscript𝑁0𝑖subscript𝑔𝑖subscript𝑘𝑗subscriptsuperscript𝛼𝑘𝑖subscriptsuperscript𝑇𝑘𝑖v^{j}(T^{1},\ldots,T^{L})=\sum_{i=1}^{n}\pi_{i}\ln\left(1+\frac{\alpha^{j}_{i}T^{j}_{i}}{N^{0}_{i}+g_{i}\sum_{k\not=j}\alpha^{k}_{i}T^{k}_{i}}\right),</td>
<td></td>
</tr>
</tbody>
</table>
where $`N_{i}^{0}`$ is the noise level and $`g_{i} \in {(0,1)}`$ and $`\alpha_{i}^{j}`$ are fading channel gains of user $`j`$ when the environment is in state $`i`$. These payoffs correspond to Shannon capacities. The constraint ([3](#S3.E3 "In 3 Symmetric water filling game")) corresponds to the average power consumption constraint. This is an instance of the Water Filling or Gaussian Interference Game \[[8](#bib.bib8), [11](#bib.bib11), [12](#bib.bib12), [15](#bib.bib15), [16](#bib.bib16)\]. In the important particular cases of OFDM wireless network and DSL access network, $`{\pi_{i} = {1/n}},{i = {1,\ldots,n}}`$.
We will look for a Nash Equilibrium (NE) of this problem. The strategies $`T^{1 \ast}`$,…,$`T^{L \ast}`$ constitute a NE, if for any strategies $`T^{1}`$,…,$`T^{L}`$ the following inequalities hold:
<table>
<tbody>
<tr>
<td></td>
<td>v1(T1,T2∗,…,TL∗)≤v1(T1∗,T2∗…,TL∗),⋯vL(T1∗,…,T(L−1)∗,TL)≤vL(T1∗,…,T(L−1)∗,TL∗).formulae-sequencesuperscript𝑣1superscript𝑇1superscript𝑇2…superscript𝑇𝐿superscript𝑣1superscript𝑇1superscript𝑇2…superscript𝑇𝐿⋯superscript𝑣𝐿superscript𝑇1…superscript𝑇𝐿1superscript𝑇𝐿superscript𝑣𝐿superscript𝑇1…superscript𝑇𝐿1superscript𝑇𝐿\begin{split}v^{1}(T^{1},T^{2*},\ldots,T^{L*})&\leq v^{1}(T^{1*},T^{2*}\ldots,T^{L*}),\\ &\cdots\\ v^{L}(T^{1*},\ldots,T^{(L-1)*},T^{L})&\leq v^{L}(T^{1*},\ldots,T^{(L-1)*},T^{L*}).\end{split}</td>
<td></td>
</tr>
</tbody>
</table>
For finding NE of such game usually the following numerical algorithm is applied. First, a strategy of $`L - 1`$ users (say, user 2,…, $`L`$) are fixed. Then, the best reply of user 1 is found solving the Water Filling optimization problem. Then, the best reply of user 2 on these strategies of the users is found solving the optimization problem and so on. It is possible to prove that under some assumption on fading channel gains this sequence of the strategies converge to a NE \[[10](#bib.bib10)\].
In this work we restrict ourselves to the case of symmetric game with equal crosstalk coefficients. This situation can for example correspond to the scenario when the users are situated at about the same distance from the base station. Namely, we assume that $`\alpha_{i}^{1} = \ldots = \alpha_{i}^{L}`$ and $`g_{i} = g`$ for $`i \in {(0,1)}`$. So, in our case the payoffs to users are given as follows
<table>
<tbody>
<tr>
<td></td>
<td>vj(T1,…,TL)=∑i=1nπiln(1+TijNi0+g∑k≠jTik),superscript𝑣𝑗superscript𝑇1…superscript𝑇𝐿superscriptsubscript𝑖1𝑛subscript𝜋𝑖1subscriptsuperscript𝑇𝑗𝑖subscriptsuperscript𝑁0𝑖𝑔subscript𝑘𝑗subscriptsuperscript𝑇𝑘𝑖v^{j}(T^{1},\ldots,T^{L})=\sum_{i=1}^{n}\pi_{i}\ln\left(1+\frac{T^{j}_{i}}{N^{0}_{i}+g\sum_{k\not=j}T^{k}_{i}}\right),</td>
<td></td>
</tr>
</tbody>
</table>
where $`N_{i}^{0} = {N^{0}/\alpha_{i}}`$, $`i \in {\lbrack 1,n\rbrack}`$ and without loss of generality we can assume that the channels are arranged in such a way that the inequalities ([2](#S2.E2 "In 2 Single decision maker")) hold. We would like to emphasize that the dependance of $`N_{i}^{0}`$ on $`i`$ allows us to model an environment with varying transmission conditions.
For this problem we propose a new algorithm of finding the NE. The algorithm is based on closed form expressions and hence it requires only a finite number of operations. Also, explaining this algorithm we will prove that the game has the unique NE under assumption that $`g \in {(0,1)}`$.
Since $`v_{j}`$ is concave on $`T^{j}`$, the Kuhn-Tucker Theorem implies the following theorem.
######
Theorem 3
$`(T^{1 \ast},\ldots,T^{L \ast})`$ is a Nash equilibrium if and only if there are non-negative $`\omega^{j}`$, $`j \in {\lbrack 1,L\rbrack}`$ (Lagrange multipliers) such that
<table>
<tbody>
<tr>
<td></td>
<td>∂∂Tijvj(T1∗,…,TL∗)=1Tij∗+Ni0+g∑k≠jTik∗{=ωjfor Tij∗>0,≤ωjfor Tij∗=0.subscriptsuperscript𝑇𝑗𝑖superscript𝑣𝑗superscript𝑇1…superscript𝑇𝐿1subscriptsuperscript𝑇𝑗𝑖subscriptsuperscript𝑁0𝑖𝑔subscript𝑘𝑗subscriptsuperscript𝑇𝑘𝑖casesabsentsuperscript𝜔𝑗for subscriptsuperscript𝑇𝑗𝑖0absentsuperscript𝜔𝑗for subscriptsuperscript𝑇𝑗𝑖0\begin{split}\frac{\partial}{\partial T^{j}_{i}}v^{j}(T^{1*},\ldots,T^{L*})&=\frac{\displaystyle 1}{\displaystyle T^{j*}_{i}+N^{0}_{i}+g\sum_{k\not=j}T^{k*}_{i}}\\ &\begin{cases}=\omega^{j}&\text{for }T^{j*}_{i}>0,\\ \leq\omega^{j}&\text{for }T^{j*}_{i}=0.\end{cases}\end{split}</td>
<td></td>
<td>(4)</td>
</tr>
</tbody>
</table>
It is clear that all $`\omega^{j}`$ are positive.
The assumption that $`g < 1`$ is crucial for uniqueness of equilibrium as it is shown in the following proposition.
######
Proposition 1
For $`g = 1`$ the symmetric water filling game has infinite number (continuum) of Nash equilibria.
Proof. Suppose that $`(T^{1 \ast},\ldots,T^{L \ast})`$ is a Nash equilibrium. Then, by Theorem [3](#ThmTheorem3 "Theorem 3 ‣ 3 Symmetric water filling game"), there are non-negative $`\omega^{j}`$, $`j \in {\lbrack 1,L\rbrack}`$ such that
<table>
<tbody>
<tr>
<td></td>
<td>1/(Ni0+∑k=1LTik∗){=ωjfor Tij∗>0,≤ωjfor Tij∗=0.1subscriptsuperscript𝑁0𝑖superscriptsubscript𝑘1𝐿subscriptsuperscript𝑇𝑘𝑖casesabsentsuperscript𝜔𝑗for subscriptsuperscript𝑇𝑗𝑖0absentsuperscript𝜔𝑗for subscriptsuperscript𝑇𝑗𝑖01\bigl{/}\bigl{(}N^{0}_{i}+\sum_{k=1}^{L}T^{k*}_{i}\bigr{)}\begin{cases}=\omega^{j}&\text{for }T^{j*}_{i}>0,\\ \leq\omega^{j}&\text{for }T^{j*}_{i}=0.\end{cases}</td>
<td></td>
</tr>
</tbody>
</table>
Thus, $`\omega^{1} = \ldots = \omega^{L} = \omega`$. So, $`T_{i}^{1 \ast}`$, …, $`T_{i}^{L \ast}`$, $`i \in {\lbrack 1,n\rbrack}`$ have to be any non-negative such that
<table>
<tbody>
<tr>
<td></td>
<td>∑k=1LTik∗=πi[1/ω−Ni0]+,superscriptsubscript𝑘1𝐿subscriptsuperscript𝑇𝑘𝑖subscript𝜋𝑖subscriptdelimited-[]1𝜔subscriptsuperscript𝑁0𝑖\sum_{k=1}^{L}T^{k*}_{i}=\pi_{i}[1/\omega-N^{0}_{i}]_{+},</td>
<td></td>
</tr>
</tbody>
</table>
and
<table>
<tbody>
<tr>
<td></td>
<td>∑i=1nπiTik∗=T¯k for k∈[1,L],superscriptsubscript𝑖1𝑛subscript𝜋𝑖subscriptsuperscript𝑇𝑘𝑖superscript¯𝑇𝑘 for 𝑘1𝐿\sum_{i=1}^{n}\pi_{i}T^{k*}_{i}=\bar{T}^{k}\mbox{ for }k\in[1,L],</td>
<td></td>
</tr>
</tbody>
</table>
where $`\omega`$ is the unique positive root of the equation
<table>
<tbody>
<tr>
<td></td>
<td>∑i=1n[1/ω−Ni0]+=∑k=1LT¯k.superscriptsubscript𝑖1𝑛subscriptdelimited-[]1𝜔subscriptsuperscript𝑁0𝑖superscriptsubscript𝑘1𝐿superscript¯𝑇𝑘\sum_{i=1}^{n}[1/\omega-N^{0}_{i}]_{+}=\sum_{k=1}^{L}\bar{T}^{k}.</td>
<td></td>
</tr>
</tbody>
</table>
It is clear that there are infinite number of such strategies. For example, if $`T_{i}^{a \ast}`$ and $`T_{i}^{b \ast}`$, $`i \in {\lbrack 1,n\rbrack}`$ ($`a \neq b`$) is the one of them and $`{T_{k}^{a \ast},T_{k}^{b \ast}} > 0`$ and $`{T_{k}^{a \ast},T_{m}^{b \ast}} >`$ for some $`k`$ and $`m`$. Then, it is clear that the following strategies for any small enough positive $`\epsilon`$ are also optimal:
<table>
<tbody>
<tr>
<td></td>
<td>T~ia∗={Tia∗for i≠k,m,Tia∗+ϵfor i=k,Tia∗−ϵπk/πmfor i=m,subscriptsuperscript~𝑇𝑎𝑖casessubscriptsuperscript𝑇𝑎𝑖for 𝑖𝑘𝑚subscriptsuperscript𝑇𝑎𝑖italic-ϵfor 𝑖𝑘subscriptsuperscript𝑇𝑎𝑖italic-ϵsubscript𝜋𝑘subscript𝜋𝑚for 𝑖𝑚\tilde{T}^{a*}_{i}=\begin{cases}{T}^{a*}_{i}&\text{for }i\not=k,m,\\ {T}^{a*}_{i}+\epsilon&\text{for }i=k,\\ {T}^{a*}_{i}-\epsilon\pi_{k}/\pi_{m}&\text{for }i=m,\end{cases}</td>
<td></td>
</tr>
</tbody>
</table>
<table>
<tbody>
<tr>
<td></td>
<td>T~ib∗={Tib∗for i≠k,m,Tib∗−ϵfor i=k,Tib∗+ϵπk/πmfor i=m.subscriptsuperscript~𝑇𝑏𝑖casessubscriptsuperscript𝑇𝑏𝑖for 𝑖𝑘𝑚subscriptsuperscript𝑇𝑏𝑖italic-ϵfor 𝑖𝑘subscriptsuperscript𝑇𝑏𝑖italic-ϵsubscript𝜋𝑘subscript𝜋𝑚for 𝑖𝑚\tilde{T}^{b*}_{i}=\begin{cases}{T}^{b*}_{i}&\text{for }i\not=k,m,\\ {T}^{b*}_{i}-\epsilon&\text{for }i=k,\\ {T}^{b*}_{i}+\epsilon\pi_{k}/\pi_{m}&\text{for }i=m.\end{cases}</td>
<td></td>
</tr>
</tbody>
</table>
This completes the proof of Proposition [1](#ThmProposition1 "Proposition 1 ‣ 3 Symmetric water filling game").
## 4 A recursive approach to the symmetric water filling game
Let $`\omega^{1}`$,…, $`\omega^{L}`$ be some parameters which in the future will act as Lagrangian multiplies. Using these parameters we introduce some auxiliary notations. Assume that these parameters are arranged as follows (this assumption does not reduce the generality of our forthcoming conclusions):
<table>
<tbody>
<tr>
<td></td>
<td>ω1≤…≤ωL.superscript𝜔1…superscript𝜔𝐿\omega^{1}\leq\ldots\leq\omega^{L}.</td>
<td></td>
<td>(5)</td>
</tr>
</tbody>
</table>
Also denote
<table>
<tbody>
<tr>
<td></td>
<td>ω¯=(ω1,…,ωL).¯𝜔superscript𝜔1…superscript𝜔𝐿\bar{\omega}=(\omega^{1},\ldots,\omega^{L}).</td>
<td></td>
</tr>
</tbody>
</table>
Introduce the following auxiliary sequence:
<table>
<tbody>
<tr>
<td></td>
<td>tr=11−g(1+(r−1)gωr−g∑j=1r1ωj) for r∈[1,L].superscript𝑡𝑟11𝑔1𝑟1𝑔superscript𝜔𝑟𝑔superscriptsubscript𝑗1𝑟1superscript𝜔𝑗 for 𝑟1𝐿t^{r}=\frac{1}{1-g}\left(\frac{1+(r-1)g}{\omega^{r}}-g\sum_{j=1}^{r}\frac{1}{\omega^{j}}\right)\mbox{ for }r\in[1,L].</td>
<td></td>
</tr>
</tbody>
</table>
It is clear that by ([5](#S4.E5 "In 4 A recursive approach to the symmetric water filling game"))
<table>
<tbody>
<tr>
<td></td>
<td>tr+1=1+(r−1)g1−g(1ωr+1−1ωr)+tr≤tr.superscript𝑡𝑟11𝑟1𝑔1𝑔1superscript𝜔𝑟11superscript𝜔𝑟superscript𝑡𝑟superscript𝑡𝑟t^{r+1}=\frac{1+(r-1)g}{1-g}\left(\frac{1}{\omega^{r+1}}-\frac{1}{\omega^{r}}\right)+t^{r}\leq t^{r}.</td>
<td></td>
</tr>
</tbody>
</table>
Thus,
<table>
<tbody>
<tr>
<td></td>
<td>tL≤tL−1≤…≤t1,superscript𝑡𝐿superscript𝑡𝐿1…superscript𝑡1t^{L}\leq t^{L-1}\leq\ldots\leq t^{1},</td>
<td></td>
</tr>
</tbody>
</table>
and
<table>
<tbody>
<tr>
<td></td>
<td>1ωr+1−1ωr=1−g1+(r−1)g(tr+1−tr).1superscript𝜔𝑟11superscript𝜔𝑟1𝑔1𝑟1𝑔superscript𝑡𝑟1superscript𝑡𝑟\begin{split}\frac{1}{\omega^{r+1}}-\frac{1}{\omega^{r}}=\frac{1-g}{1+(r-1)g}(t^{r+1}-t^{r}).\end{split}</td>
<td></td>
<td>(6)</td>
</tr>
</tbody>
</table>
Hence, for $`j \in {\lbrack{k + 1},L\rbrack}`$ we have:
<table>
<tbody>
<tr>
<td></td>
<td>1ωk−1ωj=∑r=kj−11−g1+(r−1)g(tr−tr+1).1superscript𝜔𝑘1superscript𝜔𝑗superscriptsubscript𝑟𝑘𝑗11𝑔1𝑟1𝑔superscript𝑡𝑟superscript𝑡𝑟1\begin{split}\frac{1}{\omega^{k}}-\frac{1}{\omega^{j}}=\sum_{r=k}^{j-1}\frac{1-g}{1+(r-1)g}(t^{r}-t^{r+1}).\end{split}</td>
<td></td>
<td>(7)</td>
</tr>
</tbody>
</table>
Then, sequences $`\{\omega^{r}\}`$ and $`\{ t^{r}\}`$ has the following recurrent relations:
<table>
<tbody>
<tr>
<td></td>
<td>1ω1=t1,1ω2=(1−g)t2+gt1,1ωr+1=1−g1+(r−1)gtr+1+∑j=2r(1−g)g(1+(j−1)g)(1+(j−2)g)tj+t1,\begin{split}\frac{1}{\omega^{1}}&=t^{1},\quad\frac{1}{\omega^{2}}=(1-g)t^{2}+gt^{1},\\ \frac{1}{\omega^{r+1}}&=\frac{1-g}{1+(r-1)g}t^{r+1}+\sum^{r}_{j=2}\frac{(1-g)g}{(1+(j-1)g)(1+(j-2)g)}t^{j}+t^{1},\end{split}</td>
<td></td>
<td>(8)</td>
</tr>
</tbody>
</table>
where $`r \geq 1`$. If we know sequence $`\{ t^{r}\}`$ we can restore sequence $`\{\omega^{r}\}`$. Thus, these two sequences are equivalent.
Introduce one more auxiliary sequence as follows:
<table>
<tbody>
<tr>
<td></td>
<td>τrk=11−g(1+(L−1−r+k)gωk−g∑j=1L−r+k1ωj),subscriptsuperscript𝜏𝑘𝑟11𝑔1𝐿1𝑟𝑘𝑔superscript𝜔𝑘𝑔superscriptsubscript𝑗1𝐿𝑟𝑘1superscript𝜔𝑗\tau^{k}_{r}=\frac{1}{1-g}\left(\frac{1+(L-1-r+k)g}{\omega^{k}}-g\sum_{j=1}^{L-r+k}\frac{1}{\omega^{j}}\right),</td>
<td></td>
</tr>
</tbody>
</table>
where $`{r \in {\lbrack k,L\rbrack}},{k \in {\lbrack 1,L\rbrack}}`$. There is a simple relation between sequences $`\{\omega^{k}\}`$ and $`\{ t^{k}\}`$ and $`\{\tau_{r}^{k}\}`$:
<table>
<tbody>
<tr>
<td></td>
<td>τLk=tk,subscriptsuperscript𝜏𝑘𝐿superscript𝑡𝑘\tau^{k}_{L}=t^{k},</td>
<td></td>
<td>(9)</td>
</tr>
</tbody>
</table>
and
<table>
<tbody>
<tr>
<td></td>
<td>τrk=1+(L−1−r+k)g1−g(1ωk−1ωL−r+k)+tL−r+k.subscriptsuperscript𝜏𝑘𝑟1𝐿1𝑟𝑘𝑔1𝑔1superscript𝜔𝑘1superscript𝜔𝐿𝑟𝑘superscript𝑡𝐿𝑟𝑘\tau^{k}_{r}=\frac{1+(L-1-r+k)g}{1-g}\left(\frac{1}{\omega^{k}}-\frac{1}{\omega^{L-r+k}}\right)+t^{L-r+k}.</td>
<td></td>
<td>(10)</td>
</tr>
</tbody>
</table>
So, by ([7](#S4.E7 "In 4 A recursive approach to the symmetric water filling game")), collecting terms which depends on $`t^{k}`$ we obtain
<table>
<tbody>
<tr>
<td></td>
<td>τrk=bk,rtk+Ak,r,subscriptsuperscript𝜏𝑘𝑟superscript𝑏𝑘𝑟superscript𝑡𝑘superscript𝐴𝑘𝑟\tau^{k}_{r}=b^{k,r}t^{k}+A^{k,r},</td>
<td></td>
<td>(11)</td>
</tr>
</tbody>
</table>
where
<table>
<tbody>
<tr>
<td></td>
<td>bk,r=1+(L−1−r+k)g1+(k−1)g,superscript𝑏𝑘𝑟1𝐿1𝑟𝑘𝑔1𝑘1𝑔b^{k,r}=\frac{1+(L-1-r+k)g}{1+(k-1)g},</td>
<td></td>
</tr>
</tbody>
</table>
and
<table>
<tbody>
<tr>
<td></td>
<td>Ak,r=g∑j=k+1L−r+k−11+(L−1−r+k)g(1+(j−1)g)(1+jg)tj−g(1+(L−2−r+k)gtL−r+k.\begin{split}A^{k,r}&=g\sum_{j=k+1}^{L-r+k-1}\frac{1+(L-1-r+k)g}{(1+(j-1)g)(1+jg)}t^{j}-\frac{g}{(1+(L-2-r+k)g}t^{L-r+k}.\end{split}</td>
<td></td>
</tr>
</tbody>
</table>
Thus, $`A^{k,r}`$ depends only on $`\{ t^{j}\}`$ with $`j > k`$.
Finally introduce the following notation:
\(a\) for $`N_{i}^{0} < t^{L}`$
<table>
<tbody>
<tr>
<td></td>
<td>Tik(ω¯)=11+(L−1)g(τkk−Ni0),subscriptsuperscript𝑇𝑘𝑖¯𝜔11𝐿1𝑔subscriptsuperscript𝜏𝑘𝑘subscriptsuperscript𝑁0𝑖T^{k}_{i}(\bar{\omega})=\frac{1}{1+(L-1)g}(\tau^{k}_{k}-N^{0}_{i}),</td>
<td></td>
</tr>
</tbody>
</table>
\(b\) $`t^{{L + k + 1} - r} \leq N_{i}^{0} < t^{{L + k} - r}`$ where $`r \in {\lbrack{k + 1},L\rbrack}`$
<table>
<tbody>
<tr>
<td></td>
<td>Tik(ω¯)=11+(L−1−r+k)g(τrk−Ni0),subscriptsuperscript𝑇𝑘𝑖¯𝜔11𝐿1𝑟𝑘𝑔subscriptsuperscript𝜏𝑘𝑟subscriptsuperscript𝑁0𝑖T^{k}_{i}(\bar{\omega})=\frac{1}{1+(L-1-r+k)g}(\tau^{k}_{r}-N^{0}_{i}),</td>
<td></td>
</tr>
</tbody>
</table>
\(c\) for $`t^{k} \leq N_{i}^{0}`$
<table>
<tbody>
<tr>
<td></td>
<td>Tik(ω¯)=0.subscriptsuperscript𝑇𝑘𝑖¯𝜔0T^{k}_{i}(\bar{\omega})=0.</td>
<td></td>
</tr>
</tbody>
</table>
For others combinations of relations between $`\omega^{j}`$, $`j \in {\lbrack 1,L\rbrack}`$, $`T_{i}^{k}`$ are defined by symmetry. By Theorem [3](#ThmTheorem3 "Theorem 3 ‣ 3 Symmetric water filling game") we have the following result.
######
Theorem 4
Each Nash equilibrium is of the form $`({T^{1}\hspace{0pt}{(\overline{\omega})}},\ldots,{T^{L}\hspace{0pt}{(\overline{\omega})}})`$.
The next lemma provides a nice relation between $`L`$ and $`L - 1`$ person games which shows that the introduction of a new user into the game leads to a bigger competition for the better quality channels meanwhile users prefer to keep the old structure of their strategies for worse quality channels.
######
Lemma 1
Let $`({T^{1,L}\hspace{0pt}{(\omega_{1},\ldots,\omega_{L})}},\ldots,{T^{L,L}\hspace{0pt}{(\omega_{1},\ldots,\omega_{L})}})`$ given by Theorem [4](#ThmTheorem4 "Theorem 4 ‣ 4 A recursive approach to the symmetric water filling game") (here we added the second super-script index in the notation of the strategies in order to emphasize that the strategies depend on the number of users). Then, we have
<table>
<tbody>
<tr>
<td></td>
<td>Tik,L(ω1,…,ωL)={τkk−Ni01+(L−1)gfor Ni0<tL,Tik,L−1(ω1,…,ωL−1)for tL≤Ni0,subscriptsuperscript𝑇𝑘𝐿𝑖subscript𝜔1…subscript𝜔𝐿casessubscriptsuperscript𝜏𝑘𝑘subscriptsuperscript𝑁0𝑖1𝐿1𝑔for subscriptsuperscript𝑁0𝑖superscript𝑡𝐿subscriptsuperscript𝑇𝑘𝐿1𝑖subscript𝜔1…subscript𝜔𝐿1for superscript𝑡𝐿subscriptsuperscript𝑁0𝑖T^{k,L}_{i}(\omega_{1},\ldots,\omega_{L})=\begin{cases}\frac{\displaystyle\tau^{k}_{k}-N^{0}_{i}}{\displaystyle 1+(L-1)g}&\text{for }N^{0}_{i}<t^{L},\\ T^{k,L-1}_{i}(\omega_{1},\ldots,\omega_{L-1})&\text{for }t^{L}\leq N^{0}_{i},\end{cases}</td>
<td></td>
</tr>
</tbody>
</table>
where $`k \in {\lbrack 1,{L - 1}\rbrack}`$ and
<table>
<tbody>
<tr>
<td></td>
<td>TiL,L(ω1,…,ωL)={tL−Ni01+(L−1)gfor Ni0<tL,0for tL≤Ni0.subscriptsuperscript𝑇𝐿𝐿𝑖subscript𝜔1…subscript𝜔𝐿casessuperscript𝑡𝐿subscriptsuperscript𝑁0𝑖1𝐿1𝑔for subscriptsuperscript𝑁0𝑖superscript𝑡𝐿0for superscript𝑡𝐿subscriptsuperscript𝑁0𝑖T^{L,L}_{i}(\omega_{1},\ldots,\omega_{L})=\begin{cases}\frac{\displaystyle t^{L}-N^{0}_{i}}{\displaystyle 1+(L-1)g}&\text{for }N^{0}_{i}<t^{L},\\ 0&\text{for }t^{L}\leq N^{0}_{i}.\end{cases}</td>
<td></td>
</tr>
</tbody>
</table>
## 5 A water-filling algorithm
In this section we describe a version of the water-filling algorithm for finding the NE and supply a simple proof of its convergence based on some monotonicity properties.
Let
<table>
<tbody>
<tr>
<td></td>
<td>Hk(ω¯)=∑i=1nπiTik(ω¯) for k∈[1,L].superscript𝐻𝑘¯𝜔superscriptsubscript𝑖1𝑛subscript𝜋𝑖subscriptsuperscript𝑇𝑘𝑖¯𝜔 for 𝑘1𝐿H^{k}(\bar{\omega})=\sum_{i=1}^{n}\pi_{i}T^{k}_{i}(\bar{\omega})\mbox{ for }k\in[1,L].</td>
<td></td>
</tr>
</tbody>
</table>
To find a NE we have to find $`\overline{\omega}`$ such that
<table>
<tbody>
<tr>
<td></td>
<td>Hk(ω¯)=T¯k for k∈[1,L].superscript𝐻𝑘¯𝜔superscript¯𝑇𝑘 for 𝑘1𝐿H^{k}(\bar{\omega})=\bar{T}^{k}\mbox{ for }k\in[1,L].</td>
<td></td>
<td>(12)</td>
</tr>
</tbody>
</table>
It is clear that $`H^{k}\hspace{0pt}{(\overline{\omega})}`$ has the following properties, collected in the next Lemma, which follow directly from the explicit formulas of the NE.
######
Lemma 2
\(i\) $`H^{k}\hspace{0pt}{(\overline{\omega})}`$ is nonnegative and continuous, (ii) $`H^{k}\hspace{0pt}{(\overline{\omega})}`$ is decreasing on $`\omega^{k}`$, (iii) $`{H^{k}\hspace{0pt}{(\overline{\omega})}}\rightarrow\infty`$ for $`\omega^{k}\rightarrow 0`$, (iv) $`{H^{k}\hspace{0pt}{(\overline{\omega})}} = 0`$ for enough big $`\omega^{k}`$, say for $`\omega^{k} \geq {1/N_{1}^{0}}`$, (v) $`H^{k}\hspace{0pt}{(\overline{\omega})}`$ is non-increasing by $`\omega^{j}`$ where $`j \neq k`$.
This properties give a simple proof of the convergence of the following iterative water filling algorithm for finding the NE.
Let $`\omega_{0}^{k}`$ for all $`k \in {\lbrack 1,L\rbrack}`$ be such that $`{H^{k}\hspace{0pt}{({\overline{\omega}}_{0})}} = 0`$, for example $`\omega_{0}^{k} = {1/N_{1}^{0}}`$. Let $`\omega_{1}^{k} = \omega_{0}^{k}`$ for all $`k \in {\lbrack 2,L\rbrack}`$ and define $`\omega_{1}^{1}`$ such that $`{H^{1}\hspace{0pt}{({\overline{\omega}}_{1})}} = {\overline{T}}^{1}`$. Such $`\omega_{1}^{1}`$ exists by Lemma [2](#ThmLemma2 "Lemma 2 ‣ 5 A water-filling algorithm")(i)-(iii). Then, by Lemma [2](#ThmLemma2 "Lemma 2 ‣ 5 A water-filling algorithm")(i),(v) $`{H^{k}\hspace{0pt}{({\overline{\omega}}_{0})}} = 0`$ for $`k \in {\lbrack 2,L\rbrack}`$. Let $`\omega_{2}^{k} = \omega_{1}^{k}`$ for all $`k \neq 2`$ and define $`\omega_{2}^{2}`$ such that $`{H^{2}\hspace{0pt}{({\overline{\omega}}_{2})}} = {\overline{T}}^{2}`$. Then, by Lemma [2](#ThmLemma2 "Lemma 2 ‣ 5 A water-filling algorithm")(i),(v) $`{H^{k}\hspace{0pt}{({\overline{\omega}}_{0})}} = 0`$ for $`k > 2`$ and $`{H^{k}\hspace{0pt}{({\overline{\omega}}_{0})}} \leq {\overline{T}}^{k}`$ for $`k = 1`$ and so on. Let $`\omega_{L}^{k} = \omega_{L - 1}^{k}`$ for all $`k \neq L`$ and define $`\omega_{L}^{L}`$ such that $`{H^{L}\hspace{0pt}{({\overline{\omega}}_{L})}} = {\overline{T}}^{L}`$. Then, by Lemma [2](#ThmLemma2 "Lemma 2 ‣ 5 A water-filling algorithm")(i),(v) $`{H^{k}\hspace{0pt}{({\overline{\omega}}_{L})}} \leq {\overline{T}}^{k}`$ for $`k \neq L`$ and so on. So we have non-increasing positive sequence $`\omega^{k}`$. Thus, it converges to an $`{\overline{\omega}}_{\ast}`$ which produces a NE.
## 6 Existence and uniqueness of the Nash equilibrium
In this section we will prove existence and uniqueness of the Nash equilibrium for $`L`$ person symmetric water-filling game. Our proof will have constructive character which allows us to produce an effective algorithm for finding the equilibrium strategies.
First note that there is a monotonous dependence between the resources the users can apply and Lagrange multipliers.
######
Lemma 3
Let $`({T^{1}\hspace{0pt}{(\overline{\omega})}},\ldots,{T^{L}\hspace{0pt}{(\overline{\omega})}})`$ be a Nash equilibrium. If
<table>
<tbody>
<tr>
<td></td>
<td>T¯1≥…≥T¯Lsuperscript¯𝑇1…superscript¯𝑇𝐿\bar{T}^{1}\geq\ldots\geq\bar{T}^{L}</td>
<td></td>
<td>(13)</td>
</tr>
</tbody>
</table>
then ([5](#S4.E5 "In 4 A recursive approach to the symmetric water filling game")) holds.
Proof. The result immediately follows from the following monotonicity property implied by explicit formulas of the Nash equilibrium, namely, if $`\omega^{i} < \omega^{j}`$ then $`{H^{i}\hspace{0pt}{(\overline{\omega})}} > {H^{j}\hspace{0pt}{(\overline{\omega})}}`$.
Without loss of generality we can assume that ([13](#S6.E13 "In Lemma 3 ‣ 6 Existence and uniqueness of the Nash equilibrium")) holds. Thus, by Lemma [3](#ThmLemma3 "Lemma 3 ‣ 6 Existence and uniqueness of the Nash equilibrium"), ([5](#S4.E5 "In 4 A recursive approach to the symmetric water filling game")) also holds.
Let $`\overline{\omega}`$ be the positive solution of ([12](#S5.E12 "In 5 A water-filling algorithm")). Then, by Lemma [3](#ThmLemma3 "Lemma 3 ‣ 6 Existence and uniqueness of the Nash equilibrium"), the relation ([5](#S4.E5 "In 4 A recursive approach to the symmetric water filling game")) holds. To find $`\overline{\omega}`$ we have to solve the system of non-linear equations ([12](#S5.E12 "In 5 A water-filling algorithm")). It is quite bulky system and it looks hard to solve. We will not solve it directly. What we will do we express $`\omega^{1}`$,…$`\omega^{L}`$ by $`t^{1}`$, …, $`t^{L}`$, substitute these expression into ([12](#S5.E12 "In 5 A water-filling algorithm")). The transformed system will have a triangular form, namely
<table>
<tbody>
<tr>
<td></td>
<td>H~L(tL)=T¯L,H~L−1(tL−1,tL)=T¯L−1,⋯H~1(t1,…,tL−1,tL)=T¯1.formulae-sequencesuperscript~𝐻𝐿superscript𝑡𝐿superscript¯𝑇𝐿formulae-sequencesuperscript~𝐻𝐿1superscript𝑡𝐿1superscript𝑡𝐿superscript¯𝑇𝐿1⋯superscript~𝐻1superscript𝑡1…superscript𝑡𝐿1superscript𝑡𝐿superscript¯𝑇1\begin{split}&\tilde{H}^{L}(t^{L})=\bar{T}^{L},\\ &\tilde{H}^{L-1}(t^{L-1},t^{L})=\bar{T}^{L-1},\\ &\cdots\\ &\tilde{H}^{1}(t^{1},\ldots,t^{L-1},t^{L})=\bar{T}^{1}.\\ \end{split}</td>
<td></td>
<td>(14)</td>
</tr>
</tbody>
</table>
The last system, because of monotonicity properties of $`{\overset{\sim}{H}}^{k}`$ on $`t^{k}`$, can be easily solved. Now we can move on to construction of $`{\overset{\sim}{H}}^{L}\hspace{0pt}{(t^{L})}`$, …, $`{\overset{\sim}{H}}^{1}\hspace{0pt}{(t^{1},\ldots,t^{L - 1},t^{L})}`$. First we will construct $`{\overset{\sim}{H}}^{L}\hspace{0pt}{(t^{L})}`$ and find the optimal $`t^{L}`$. Note that,
<table>
<tbody>
<tr>
<td></td>
<td>HL(ω¯)=∑Ni0<tLπiTiL(ω¯)==11+(L−1)g∑Ni0<tLπi(τLL−Ni0)=11+(L−1)g∑Ni0<tLπi(tL−Ni0)=H~L(tL).superscript𝐻𝐿¯𝜔subscriptsubscriptsuperscript𝑁0𝑖superscript𝑡𝐿subscript𝜋𝑖subscriptsuperscript𝑇𝐿𝑖¯𝜔11𝐿1𝑔subscriptsubscriptsuperscript𝑁0𝑖superscript𝑡𝐿subscript𝜋𝑖subscriptsuperscript𝜏𝐿𝐿subscriptsuperscript𝑁0𝑖11𝐿1𝑔subscriptsubscriptsuperscript𝑁0𝑖superscript𝑡𝐿subscript𝜋𝑖superscript𝑡𝐿subscriptsuperscript𝑁0𝑖superscript~𝐻𝐿superscript𝑡𝐿\begin{split}H^{L}(\bar{\omega})&=\sum_{N^{0}_{i}<t^{L}}\pi_{i}T^{L}_{i}(\bar{\omega})=\\ &=\frac{1}{1+(L-1)g}\sum_{N^{0}_{i}<t^{L}}\pi_{i}(\tau^{L}_{L}-N^{0}_{i})\\ &=\frac{1}{1+(L-1)g}\sum_{N^{0}_{i}<t^{L}}\pi_{i}(t^{L}-N^{0}_{i})=\tilde{H}^{L}(t^{L}).\end{split}</td>
<td></td>
</tr>
</tbody>
</table>
It is clear that $`{\overset{\sim}{H}}^{L}\hspace{0pt}{( \cdot )}`$ is continuous in $`(0,\infty)`$, $`{{\overset{\sim}{H}}^{L}\hspace{0pt}{(\tau)}} = 0`$ for $`\tau \leq N_{1}^{0}`$, $`{{\overset{\sim}{H}}^{L}\hspace{0pt}{({+ \infty})}} = {+ \infty}`$ and $`{\overset{\sim}{H}}^{L}\hspace{0pt}{( \cdot )}`$ is strictly increasing in $`(N_{1}^{0},\infty)`$. Then, there is the unique positive $`t_{\ast}^{L}`$ such that
<table>
<tbody>
<tr>
<td></td>
<td>H~L(t∗L)=TL¯.superscript~𝐻𝐿subscriptsuperscript𝑡𝐿¯superscript𝑇𝐿\tilde{H}^{L}(t^{L}_{*})=\bar{T^{L}}.</td>
<td></td>
<td>(15)</td>
</tr>
</tbody>
</table>
Now we move on to construction of $`{\overset{\sim}{H}}^{L - 1}\hspace{0pt}{(t^{L - 1},t^{L})}`$ and finding the optimal $`t^{L - 1}`$. Note that $`\tau_{L}^{L - 1} = t^{L - 1}`$ and by ([7](#S4.E7 "In 4 A recursive approach to the symmetric water filling game")) and ([10](#S4.E10 "In 4 A recursive approach to the symmetric water filling game")), we have
<table>
<tbody>
<tr>
<td></td>
<td>τL−1L−1=τLL−1+g1−g(1ωL−1−1ωL)=tL−1+g1+(L−2)g(tL−1−tL)=1+(L−1)g1+(L−2)gtL−1−g1+(L−2)gtL.subscriptsuperscript𝜏𝐿1𝐿1subscriptsuperscript𝜏𝐿1𝐿𝑔1𝑔1superscript𝜔𝐿11superscript𝜔𝐿superscript𝑡𝐿1𝑔1𝐿2𝑔superscript𝑡𝐿1superscript𝑡𝐿1𝐿1𝑔1𝐿2𝑔superscript𝑡𝐿1𝑔1𝐿2𝑔superscript𝑡𝐿\begin{split}&\tau^{L-1}_{L-1}=\tau^{L-1}_{L}+\frac{g}{1-g}\left(\frac{1}{\omega^{L-1}}-\frac{1}{\omega^{L}}\right)\\ &=t^{L-1}+\frac{g}{1+(L-2)g}(t^{L-1}-t^{L})\\ &=\frac{1+(L-1)g}{1+(L-2)g}t^{L-1}-\frac{g}{1+(L-2)g}t^{L}.\end{split}</td>
<td></td>
</tr>
</tbody>
</table>
Thus,
<table>
<tbody>
<tr>
<td></td>
<td>HL−1(ω¯)=∑Ni0<tLπiTiL−1(ω¯)+∑tL≤Ni0<tL−1πiTiL−1(ω¯)==11+(L−1)g∑Ni0<tLπi(τL−1L−1−Ni0)+11+(L−2)g∑tL≤Ni0<tL−1πi(τLL−1−Ni0)=11+(L−1)g∑Ni0<tLπi(1+(L−1)g1+(L−2)gtL−1−g1+(L−2)gtL−Ni0)+11+(L−2)g∑tL≤Ni0<tL−1πi(tL−1−Ni0)=H~L−1(tL−1,tL).superscript𝐻𝐿1¯𝜔subscriptsubscriptsuperscript𝑁0𝑖superscript𝑡𝐿subscript𝜋𝑖subscriptsuperscript𝑇𝐿1𝑖¯𝜔subscriptsuperscript𝑡𝐿subscriptsuperscript𝑁0𝑖superscript𝑡𝐿1subscript𝜋𝑖subscriptsuperscript𝑇𝐿1𝑖¯𝜔11𝐿1𝑔subscriptsubscriptsuperscript𝑁0𝑖superscript𝑡𝐿subscript𝜋𝑖subscriptsuperscript𝜏𝐿1𝐿1subscriptsuperscript𝑁0𝑖11𝐿2𝑔subscriptsuperscript𝑡𝐿subscriptsuperscript𝑁0𝑖superscript𝑡𝐿1subscript𝜋𝑖subscriptsuperscript𝜏𝐿1𝐿subscriptsuperscript𝑁0𝑖11𝐿1𝑔subscriptsubscriptsuperscript𝑁0𝑖superscript𝑡𝐿subscript𝜋𝑖1𝐿1𝑔1𝐿2𝑔superscript𝑡𝐿1𝑔1𝐿2𝑔superscript𝑡𝐿subscriptsuperscript𝑁0𝑖11𝐿2𝑔subscriptsuperscript𝑡𝐿subscriptsuperscript𝑁0𝑖superscript𝑡𝐿1subscript𝜋𝑖superscript𝑡𝐿1subscriptsuperscript𝑁0𝑖superscript~𝐻𝐿1superscript𝑡𝐿1superscript𝑡𝐿\begin{split}&H^{L-1}(\bar{\omega})=\sum_{N^{0}_{i}<t^{L}}\pi_{i}T^{L-1}_{i}(\bar{\omega})+\sum_{t^{L}\leq N^{0}_{i}<t^{L-1}}\pi_{i}T^{L-1}_{i}(\bar{\omega})=\\ &=\frac{1}{1+(L-1)g}\sum_{N^{0}_{i}<t^{L}}\pi_{i}(\tau^{L-1}_{L-1}-N^{0}_{i})+\frac{1}{1+(L-2)g}\sum_{t^{L}\leq N^{0}_{i}<t^{L-1}}\pi_{i}(\tau^{L-1}_{L}-N^{0}_{i})\\ &=\frac{1}{1+(L-1)g}\sum_{N^{0}_{i}<t^{L}}\pi_{i}\Bigl{(}\frac{1+(L-1)g}{1+(L-2)g}t^{L-1}-\frac{g}{1+(L-2)g}t^{L}-N^{0}_{i}\Bigr{)}\\ &+\frac{1}{1+(L-2)g}\sum_{t^{L}\leq N^{0}_{i}<t^{L-1}}\pi_{i}(t^{L-1}-N^{0}_{i})\\ &=\tilde{H}^{L-1}(t^{L-1},t^{L}).\end{split}</td>
<td></td>
</tr>
</tbody>
</table>
It is clear that $`{\overset{\sim}{H}}^{L - 1}\hspace{0pt}{( \cdot ,t_{\ast}^{L})}`$ is continuous and increasing in $`(t_{\ast}^{L},\infty)`$, $`{{\overset{\sim}{H}}^{L - 1}\hspace{0pt}{(\infty,t_{\ast}^{L})}} = {+ \infty}`$ and $`{{\overset{\sim}{H}}^{L - 1}\hspace{0pt}{(t_{\ast}^{L},t_{\ast}^{L})}} = {{\overset{\sim}{H}}^{L}\hspace{0pt}{(t_{\ast}^{L})}} = \overline{T^{L}} \leq \overline{T^{L-1}}`$. So, there is the unique positive $`t_{\ast}^{L - 1}`$ such that
<table>
<tbody>
<tr>
<td></td>
<td>H~L−1(t∗L−1,t∗L)=TL¯.superscript~𝐻𝐿1subscriptsuperscript𝑡𝐿1subscriptsuperscript𝑡𝐿¯superscript𝑇𝐿\tilde{H}^{L-1}(t^{L-1}_{*},t^{L}_{*})=\bar{T^{L}}.</td>
<td></td>
<td>(16)</td>
</tr>
</tbody>
</table>
Next we construct $`{\overset{\sim}{H}}^{k}\hspace{0pt}{(t^{k},\ldots,t^{L - 1},t^{L})}`$ and find the optimal $`t^{k}`$ where $`k \in {\lbrack 1,{L - 2}\rbrack}`$. By ([9](#S4.E9 "In 4 A recursive approach to the symmetric water filling game")) and ([9](#S4.E9 "In 4 A recursive approach to the symmetric water filling game")), we have
<table>
<tbody>
<tr>
<td></td>
<td>Hk(ω¯)=∑Ni0<tLπiTik+∑r=k+1L∑tL+k+1−r≤Ni0<tL+k−rπiTik=11+(L−1)g∑Ni0<tLπi(τLk−Ni0)+∑r=k+1L∑tL+k+1−r≤Ni0<tL+k−rπi(τrk−Ni0)1+(L−1−r+k)g=11+(L−1)g∑Ni0<tLπi(bk,ktk+Ak,k−Ni0)+∑r=k+1L∑tL+k+1−r≤Ni0<tL+k−rπi(bk,rtk+Ak,r−Ni0)1+(L−1−r+k)g=H~k(tk,tk+1,…,tL).superscript𝐻𝑘¯𝜔subscriptsubscriptsuperscript𝑁0𝑖superscript𝑡𝐿subscript𝜋𝑖subscriptsuperscript𝑇𝑘𝑖superscriptsubscript𝑟𝑘1𝐿subscriptsuperscript𝑡𝐿𝑘1𝑟subscriptsuperscript𝑁0𝑖superscript𝑡𝐿𝑘𝑟subscript𝜋𝑖subscriptsuperscript𝑇𝑘𝑖11𝐿1𝑔subscriptsubscriptsuperscript𝑁0𝑖superscript𝑡𝐿subscript𝜋𝑖subscriptsuperscript𝜏𝑘𝐿subscriptsuperscript𝑁0𝑖superscriptsubscript𝑟𝑘1𝐿subscriptsuperscript𝑡𝐿𝑘1𝑟subscriptsuperscript𝑁0𝑖superscript𝑡𝐿𝑘𝑟subscript𝜋𝑖subscriptsuperscript𝜏𝑘𝑟subscriptsuperscript𝑁0𝑖1𝐿1𝑟𝑘𝑔11𝐿1𝑔subscriptsubscriptsuperscript𝑁0𝑖superscript𝑡𝐿subscript𝜋𝑖superscript𝑏𝑘𝑘superscript𝑡𝑘superscript𝐴𝑘𝑘subscriptsuperscript𝑁0𝑖superscriptsubscript𝑟𝑘1𝐿subscriptsuperscript𝑡𝐿𝑘1𝑟subscriptsuperscript𝑁0𝑖superscript𝑡𝐿𝑘𝑟subscript𝜋𝑖superscript𝑏𝑘𝑟superscript𝑡𝑘superscript𝐴𝑘𝑟subscriptsuperscript𝑁0𝑖1𝐿1𝑟𝑘𝑔superscript~𝐻𝑘superscript𝑡𝑘superscript𝑡𝑘1…superscript𝑡𝐿\begin{split}&H^{k}(\bar{\omega})=\sum_{N^{0}_{i}<t^{L}}\pi_{i}T^{k}_{i}+\sum_{r=k+1}^{L}\sum_{t^{L+k+1-r}\leq N^{0}_{i}<t^{L+k-r}}\pi_{i}T^{k}_{i}\\ &=\frac{1}{1+(L-1)g}\sum_{N^{0}_{i}<t^{L}}\pi_{i}(\tau^{k}_{L}-N^{0}_{i})+\sum_{r=k+1}^{L}\sum_{t^{L+k+1-r}\leq N^{0}_{i}<t^{L+k-r}}\frac{\pi_{i}(\tau^{k}_{r}-N^{0}_{i})}{1+(L-1-r+k)g}\\ &=\frac{1}{1+(L-1)g}\sum_{N^{0}_{i}<t^{L}}\pi_{i}(b^{k,k}t^{k}+A^{k,k}-N^{0}_{i})\\ &+\sum_{r=k+1}^{L}\sum_{t^{L+k+1-r}\leq N^{0}_{i}<t^{L+k-r}}\frac{\pi_{i}(b^{k,r}t^{k}+A^{k,r}-N^{0}_{i})}{1+(L-1-r+k)g}\\ &=\tilde{H}^{k}(t^{k},t^{k+1},\ldots,t^{L}).\end{split}</td>
<td></td>
</tr>
</tbody>
</table>
It is clear that $`{\overset{\sim}{H}}^{k}\hspace{0pt}{( \cdot ,t_{\ast}^{k + 1},\ldots,t_{\ast}^{L})}`$ is continuous and increasing in $`(t_{\ast}^{k + 1},\infty)`$, $`{{\overset{\sim}{H}}^{k}\hspace{0pt}{(\infty,t_{\ast}^{k + 1},\ldots,t_{\ast}^{L})}} = {+ \infty}`$ and by Lemma [1](#ThmLemma1 "Lemma 1 ‣ 4 A recursive approach to the symmetric water filling game") $`{{\overset{\sim}{H}}^{k}\hspace{0pt}{(t_{\ast}^{k + 1},t_{\ast}^{k + 1},\ldots,t_{\ast}^{L})}} = {{\overset{\sim}{H}}^{k + 1}\hspace{0pt}{(t_{\ast}^{k + 1},\ldots,t_{\ast}^{L})}} = {\overline{T}}^{k + 1} \leq {\overline{T}}^{k}`$. So, there is the unique positive $`t_{\ast}^{k}`$ such that
<table>
<tbody>
<tr>
<td></td>
<td>H~k(t∗k,t∗k+1,…,t∗L)=Tk¯.superscript~𝐻𝑘subscriptsuperscript𝑡𝑘subscriptsuperscript𝑡𝑘1…subscriptsuperscript𝑡𝐿¯superscript𝑇𝑘\tilde{H}^{k}(t^{k}_{*},t^{k+1}_{*},\ldots,t^{L}_{*})=\bar{T^{k}}.</td>
<td></td>
<td>(17)</td>
</tr>
</tbody>
</table>
Thus, we have proved the following result:
######
Lemma 4
Solution of the system ([12](#S5.E12 "In 5 A water-filling algorithm")) is equivalent to solution of the triangular system (LABEL:eqnFFFFFF). This system has the unique solution which can be found sequentially from $`t^{L}`$ down to $`t^{1}`$, applying either the bisection method or the explicit scheme suggested in Section II. The optimal Lagrangian multipliers can be reconstructed from $`\{ t^{r}\}`$ by ([8](#S4.E8 "In 4 A recursive approach to the symmetric water filling game")).
Finally we also have the following result:
######
Theorem 5
The symmetric water filling game has the unique Nash equilibrium $`({T^{1}\hspace{0pt}{({\overline{\omega}}_{\ast})}},\ldots,{T^{L}\hspace{0pt}{({\overline{\omega}}_{\ast})}})`$, where $`{\overline{\omega}}_{\ast}`$ is given by Lemma [4](#ThmLemma4 "Lemma 4 ‣ 6 Existence and uniqueness of the Nash equilibrium").
Note that although the payoffs have symmetric form, the equilibrium strategies, because of triangular form of system (LABEL:eqnFFFFFF), have hierarchical structure induced by difference in power levels available to the users. Namely, the user who has to transmit with smaller average power consumption, in our case it is user $`L`$, acts first. He assigns his optimal strategies as if there is no other users at all but taking into account the total number of users and fading channels gains. Then, the turn to act is given to user $`L - 1`$. He takes into account only the behavior of the user $`L`$ with smaller average power consumption than he has, the total number of users and fading channels gains and so on. The last user who constructs the equilibrium strategy is user 1 with the largest available power resource.
## 7 Closed form solution for $`L`$ person game
In this section for the case of $`L`$ users we show how Theorem [5](#ThmTheorem5 "Theorem 5 ‣ 6 Existence and uniqueness of the Nash equilibrium") and Lemma [4](#ThmLemma4 "Lemma 4 ‣ 6 Existence and uniqueness of the Nash equilibrium") can be used to construct NE in closed form.
Assume that $`{\overline{T}}^{1} > \ldots > {\overline{T}}^{L}`$. We will construct the optimal strategies $`T^{L \ast}`$, …, $`T^{1 \ast}`$ sequentially.
Step for construction of $`T^{L \ast}`$. Since $`{\overset{\sim}{H}}^{L}\hspace{0pt}{( \cdot )}`$ is strictly increasing we can find an integer $`k_{L}`$ such that
<table>
<tbody>
<tr>
<td></td>
<td>H~L(NkL0)<T¯L≤H~L(NkL+10).superscript~𝐻𝐿subscriptsuperscript𝑁0superscript𝑘𝐿superscript¯𝑇𝐿superscript~𝐻𝐿subscriptsuperscript𝑁0superscript𝑘𝐿1\tilde{H}^{L}(N^{0}_{k^{L}})<\bar{T}^{L}\leq\tilde{H}^{L}(N^{0}_{k^{L}+1}).</td>
<td></td>
</tr>
</tbody>
</table>
or from the following equivalent conditions:
<table>
<tbody>
<tr>
<td></td>
<td>φkLL<T¯L≤φkL+1L,subscriptsuperscript𝜑𝐿superscript𝑘𝐿superscript¯𝑇𝐿subscriptsuperscript𝜑𝐿superscript𝑘𝐿1\varphi^{L}_{k^{L}}<\bar{T}^{L}\leq\varphi^{L}_{k^{L}+1},</td>
<td></td>
</tr>
</tbody>
</table>
where
<table>
<tbody>
<tr>
<td></td>
<td>φkL=11+(L−1)g∑i=1kπi(Nk0−Ni0),subscriptsuperscript𝜑𝐿𝑘11𝐿1𝑔superscriptsubscript𝑖1𝑘subscript𝜋𝑖subscriptsuperscript𝑁0𝑘subscriptsuperscript𝑁0𝑖\varphi^{L}_{k}=\frac{1}{1+(L-1)g}\sum_{i=1}^{k}\pi_{i}(N^{0}_{k}-N^{0}_{i}),</td>
<td></td>
</tr>
</tbody>
</table>
for $`k \leq n`$, and $`\varphi_{n + 1}^{L} = \infty`$. Then, since $`{{\overset{\sim}{H}}^{L}\hspace{0pt}{(t_{\ast}^{L})}} = {\overline{T}}^{L}`$, we have that
<table>
<tbody>
<tr>
<td></td>
<td>t∗L=(1+(L−1)g)T¯L+∑i=1kLπiNi0∑i=1kLπi.subscriptsuperscript𝑡𝐿1𝐿1𝑔superscript¯𝑇𝐿subscriptsuperscriptsubscript𝑘𝐿𝑖1subscript𝜋𝑖subscriptsuperscript𝑁0𝑖subscriptsuperscriptsubscript𝑘𝐿𝑖1subscript𝜋𝑖t^{L}_{*}=\frac{(1+(L-1)g)\bar{T}^{L}+\sum^{k_{L}}_{i=1}\pi_{i}N^{0}_{i}}{\sum^{k_{L}}_{i=1}\pi_{i}}.</td>
<td></td>
</tr>
</tbody>
</table>
Thus, the optimal strategy of user $`L`$ is given as follows
<table>
<tbody>
<tr>
<td></td>
<td>TiL∗={11+(L−1)g(t∗L−Ni0)if i∈[1,kL],0if i∈[kL+1,n].subscriptsuperscript𝑇𝐿𝑖cases11𝐿1𝑔subscriptsuperscript𝑡𝐿subscriptsuperscript𝑁0𝑖if 𝑖1superscript𝑘𝐿0if 𝑖superscript𝑘𝐿1𝑛T^{L*}_{i}=\begin{cases}\frac{\displaystyle 1}{\displaystyle 1+(L-1)g}(t^{L}_{*}-N^{0}_{i})&\text{if }i\in[1,k^{L}],\\ 0&\text{if }i\in[k^{L}+1,n].\end{cases}</td>
<td></td>
</tr>
</tbody>
</table>
Step for construction of $`T^{{({L - 1})} \ast}`$. Since $`t_{\ast}^{L - 1}`$ is the root of the equation $`{{\overset{\sim}{H}}^{L - 1}\hspace{0pt}{( \cdot ,t_{\ast}^{L})}} = {\overline{T}}^{L - 1}`$ there is $`k^{L - 1}`$ such that $`k^{L - 1} \geq k^{L}`$ and $`N_{k^{L - 1} + 1}^{0} \geq t_{\ast}^{L - 1} > N_{k^{L - 1}}^{0}`$. Thus,
<table>
<tbody>
<tr>
<td></td>
<td>t∗L−1=(T¯L−1+11+(L−2)g∑i=kL+1kL−1πiNi0+11+(L−1)g∑i=1kLπi(gtL∗1+(L−2)g+Ni0))/(11+(L−2)g∑i=1kL−1πi).subscriptsuperscript𝑡𝐿1superscript¯𝑇𝐿111𝐿2𝑔superscriptsubscript𝑖superscript𝑘𝐿1superscript𝑘𝐿1subscript𝜋𝑖subscriptsuperscript𝑁0𝑖11𝐿1𝑔superscriptsubscript𝑖1superscript𝑘𝐿subscript𝜋𝑖𝑔subscriptsuperscript𝑡𝐿1𝐿2𝑔subscriptsuperscript𝑁0𝑖11𝐿2𝑔superscriptsubscript𝑖1superscript𝑘𝐿1subscript𝜋𝑖\begin{split}t^{L-1}_{*}&=\Bigl{(}\bar{T}^{L-1}+\frac{1}{1+(L-2)g}\sum_{i={k^{L}+1}}^{k^{L-1}}\pi_{i}N^{0}_{i}\\ &+\frac{1}{1+(L-1)g}\sum_{i=1}^{k^{L}}\pi_{i}(\frac{gt^{*}_{L}}{1+(L-2)g}+N^{0}_{i})\Bigr{)}\\ &\Bigr{/}\Bigl{(}\frac{1}{1+(L-2)g}\sum_{i=1}^{k^{L-1}}\pi_{i}\Bigr{)}.\end{split}</td>
<td></td>
</tr>
</tbody>
</table>
Here and bellow we assume that $`{\sum_{x}^{y}1} = 0`$ for $`y < x`$. So, $`k^{L - 1} \geq k^{L}`$ can be found as follows:
\(i\) $`k^{L - 1} = k^{L}`$ if $`{\overline{T}}^{L - 1} \leq \varphi_{k^{L - 1} + 1}^{L - 1}`$,
\(ii\) otherwise $`k^{L - 1}`$ is given by the condition:
<table>
<tbody>
<tr>
<td></td>
<td>φkL−1L−1<T¯L−1≤φkL−1+1L−1,subscriptsuperscript𝜑𝐿1superscript𝑘𝐿1superscript¯𝑇𝐿1subscriptsuperscript𝜑𝐿1superscript𝑘𝐿11\varphi^{L-1}_{k^{L-1}}<\bar{T}^{L-1}\leq\varphi^{L-1}_{k^{L-1}+1},</td>
<td></td>
</tr>
</tbody>
</table>
where
<table>
<tbody>
<tr>
<td></td>
<td>φkL−1=∑i=kL+1kπi1+(L−2)g(Nk0−Ni0)+∑i=1kLπi1+(L−1)g×(1+(L−1)g1+(L−2)gNk0−Ni0−g1+gt∗L−1),subscriptsuperscript𝜑𝐿1𝑘superscriptsubscript𝑖superscript𝑘𝐿1𝑘subscript𝜋𝑖1𝐿2𝑔subscriptsuperscript𝑁0𝑘subscriptsuperscript𝑁0𝑖superscriptsubscript𝑖1superscript𝑘𝐿subscript𝜋𝑖1𝐿1𝑔1𝐿1𝑔1𝐿2𝑔subscriptsuperscript𝑁0𝑘subscriptsuperscript𝑁0𝑖𝑔1𝑔subscriptsuperscript𝑡𝐿1\begin{split}\varphi^{L-1}_{k}&=\sum_{i=k^{L}+1}^{k}\frac{\pi_{i}}{1+(L-2)g}(N^{0}_{k}-N^{0}_{i})\\ &+\sum_{i=1}^{k^{L}}\frac{\pi_{i}}{1+(L-1)g}\\ &\times\left(\frac{1+(L-1)g}{1+(L-2)g}N^{0}_{k}-N^{0}_{i}-\frac{g}{1+g}t^{L-1}_{*}\right),\end{split}</td>
<td></td>
</tr>
</tbody>
</table>
for $`k \in {\lbrack{k^{L - 1} + 1},n\rbrack}`$ and $`\varphi_{n + 1}^{L - 1} = \infty`$.
Thus, the optimal strategy $`T^{{({L - 1})} \ast}`$ of user $`L - 1`$ is given by
<table>
<tbody>
<tr>
<td></td>
<td>Ti(L−1)∗={t∗L−11+(L−2)g−g1+gt∗L+Ni01+(L−1)g,i∈[1,kL],11+(L−2)g(t∗L−1−Ni0),i∈[kL+1,kL−1],0,i∈[kL−1+1,n].subscriptsuperscript𝑇𝐿1𝑖casessubscriptsuperscript𝑡𝐿11𝐿2𝑔otherwise𝑔1𝑔subscriptsuperscript𝑡𝐿subscriptsuperscript𝑁0𝑖1𝐿1𝑔𝑖1superscript𝑘𝐿11𝐿2𝑔subscriptsuperscript𝑡𝐿1subscriptsuperscript𝑁0𝑖𝑖superscript𝑘𝐿1superscript𝑘𝐿10𝑖superscript𝑘𝐿11𝑛T^{(L-1)*}_{i}=\begin{cases}\frac{\displaystyle t^{L-1}_{*}}{\displaystyle 1+(L-2)g}\\ -\frac{\displaystyle\frac{g}{1+g}t^{L}_{*}+N^{0}_{i}}{\displaystyle 1+(L-1)g},&i\in[1,k^{L}],\\ \frac{\displaystyle 1}{\displaystyle 1+(L-2)g}(t^{L-1}_{*}-N^{0}_{i}),&i\in[k^{L}+1,k^{L-1}],\\ 0,&i\in[k^{L-1}+1,n].\end{cases}</td>
<td></td>
</tr>
</tbody>
</table>
Step for construction of $`T^{M \ast}`$ where $`M < L`$. We have already constructed $`T^{L \ast}`$, …, $`T^{{({M + 1})} \ast}`$ and now we are going to construct $`T^{M \ast}`$. Since $`t_{\ast}^{M}`$ is the root of the equation $`{{\overset{\sim}{H}}^{M}\hspace{0pt}{( \cdot ,t_{\ast}^{M + 1},\ldots,t_{\ast}^{L})}} = {\overline{T}}^{M}`$ there is $`k^{M}`$ such that $`k^{M} \geq k^{M + 1}`$ and $`N_{k^{M} + 1}^{0} \geq t_{\ast}^{M} > N_{k^{M}}^{0}`$. Thus,
<table>
<tbody>
<tr>
<td></td>
<td>t∗M=(T¯M+11+(L−1)g∑i=1kMπi(Ak,k−Ni0)+∑r=M+1L∑i=kp+1kp−1πi(Ap,r−Ni0)1+(L−1−r+p)g)/(11+(M−1)g∑i=1kMπi).subscriptsuperscript𝑡𝑀superscript¯𝑇𝑀11𝐿1𝑔superscriptsubscript𝑖1subscript𝑘𝑀subscript𝜋𝑖superscript𝐴𝑘𝑘subscriptsuperscript𝑁0𝑖superscriptsubscript𝑟𝑀1𝐿superscriptsubscript𝑖superscript𝑘𝑝1superscript𝑘𝑝1subscript𝜋𝑖superscript𝐴𝑝𝑟subscriptsuperscript𝑁0𝑖1𝐿1𝑟𝑝𝑔11𝑀1𝑔superscriptsubscript𝑖1superscript𝑘𝑀subscript𝜋𝑖\begin{split}t^{M}_{*}&=\Bigl{(}\bar{T}^{M}+\frac{1}{1+(L-1)g}\sum_{i=1}^{k_{M}}\pi_{i}(A^{k,k}-N^{0}_{i})\\ &+\sum_{r=M+1}^{L}\sum_{i=k^{p}+1}^{k^{p-1}}\frac{\pi_{i}(A^{p,r}-N^{0}_{i})}{1+(L-1-r+p)g})\\ &\Bigr{/}\Bigl{(}\frac{1}{1+(M-1)g}\sum_{i=1}^{k^{M}}\pi_{i}\Bigr{)}.\end{split}</td>
<td></td>
</tr>
</tbody>
</table>
So, $`k^{M} \geq k^{M + 1}`$ can be found as follows:
\(i\) $`k^{M} = k^{M + 1}`$ if $`{\overline{T}}^{M} \leq \varphi_{k^{M} + 1}^{M}`$,
\(ii\) otherwise $`k^{M}`$ is given by the condition:
<table>
<tbody>
<tr>
<td></td>
<td>φkMM<T¯M≤φkM+1Msubscriptsuperscript𝜑𝑀superscript𝑘𝑀superscript¯𝑇𝑀subscriptsuperscript𝜑𝑀superscript𝑘𝑀1\varphi^{M}_{k^{M}}<\bar{T}^{M}\leq\varphi^{M}_{k^{M}+1}</td>
<td></td>
</tr>
</tbody>
</table>
where
<table>
<tbody>
<tr>
<td></td>
<td>φkM=11+(L−1)g∑i=1kπi(bk,kNk0+Ak,k−Ni0)+∑r=M+1L∑i=kp+1kp−1πi(bp,rNk0+Ap,r−Ni0)1+(L−1−r+p)g.subscriptsuperscript𝜑𝑀𝑘11𝐿1𝑔superscriptsubscript𝑖1𝑘subscript𝜋𝑖superscript𝑏𝑘𝑘subscriptsuperscript𝑁0𝑘superscript𝐴𝑘𝑘subscriptsuperscript𝑁0𝑖superscriptsubscript𝑟𝑀1𝐿superscriptsubscript𝑖superscript𝑘𝑝1superscript𝑘𝑝1subscript𝜋𝑖superscript𝑏𝑝𝑟subscriptsuperscript𝑁0𝑘superscript𝐴𝑝𝑟subscriptsuperscript𝑁0𝑖1𝐿1𝑟𝑝𝑔\begin{split}\varphi^{M}_{k}&=\frac{1}{1+(L-1)g}\sum_{i=1}^{k}\pi_{i}(b^{k,k}N^{0}_{k}+A^{k,k}-N^{0}_{i})\\ &+\sum_{r=M+1}^{L}\sum_{i={k^{p}+1}}^{k^{p-1}}\frac{\pi_{i}(b^{p,r}N^{0}_{k}+A^{p,r}-N^{0}_{i})}{1+(L-1-r+p)g}.\end{split}</td>
<td></td>
</tr>
</tbody>
</table>
Thus, the optimal strategy of user $`L`$ is given as follows
<table>
<tbody>
<tr>
<td></td>
<td>TiM∗={τMM−Ni01+(L−1)g,i∈[1,kL],τrM−Ni01+(L−1−r+M)g,i∈[kr+1,kr−1],r∈[M+1,L]0,i∈[kM+1,n].subscriptsuperscript𝑇𝑀𝑖casessubscriptsuperscript𝜏𝑀𝑀subscriptsuperscript𝑁0𝑖1𝐿1𝑔𝑖1superscript𝑘𝐿subscriptsuperscript𝜏𝑀𝑟subscriptsuperscript𝑁0𝑖1𝐿1𝑟𝑀𝑔𝑖superscript𝑘𝑟1superscript𝑘𝑟1otherwise𝑟𝑀1𝐿0𝑖superscript𝑘𝑀1𝑛T^{M*}_{i}=\begin{cases}\frac{\displaystyle\tau^{M}_{M}-N^{0}_{i}}{\displaystyle 1+(L-1)g},&i\in[1,k^{L}],\\ \frac{\displaystyle\tau^{M}_{r}-N^{0}_{i}}{\displaystyle 1+(L-1-r+M)g},&i\in[k^{r}+1,k^{r-1}],\\ &r\in[M+1,L]\\ 0,&i\in[k^{M}+1,n].\end{cases}</td>
<td></td>
</tr>
</tbody>
</table>
In particular for two and three person games ($`L = 2`$ and $`L = 3`$) we have the following results.
######
Theorem 6
Let $`{\overline{T}}_{1} > {\overline{T}}_{2}`$. Then, the Nash equilibrium strategies are given by
<table>
<tbody>
<tr>
<td></td>
<td>Ti1∗={t∗1−gt∗2+Ni01+gif i∈[1,k2],t∗1−Ni0if i∈[k2+1,k1],0if i∈[k1+1,n],Ti2∗={11+g(t∗2−Ni0)if i∈[1,k2],0if i∈[k2+1,n],subscriptsuperscript𝑇1𝑖casessubscriptsuperscript𝑡1𝑔subscriptsuperscript𝑡2subscriptsuperscript𝑁0𝑖1𝑔if 𝑖1superscript𝑘2subscriptsuperscript𝑡1subscriptsuperscript𝑁0𝑖if 𝑖superscript𝑘21superscript𝑘10if 𝑖superscript𝑘11𝑛subscriptsuperscript𝑇2𝑖cases11𝑔subscriptsuperscript𝑡2subscriptsuperscript𝑁0𝑖if 𝑖1superscript𝑘20if 𝑖superscript𝑘21𝑛\begin{split}T^{1*}_{i}&=\begin{cases}t^{1}_{*}-\frac{\displaystyle gt^{2}_{*}+N^{0}_{i}}{\displaystyle 1+g}&\text{if }i\in[1,k^{2}],\\ t^{1}_{*}-N^{0}_{i}&\text{if }i\in[k^{2}+1,k^{1}],\\ 0&\text{if }i\in[k^{1}+1,n],\end{cases}\\ T^{2*}_{i}&=\begin{cases}\frac{\displaystyle 1}{\displaystyle 1+g}(t^{2}_{*}-N^{0}_{i})&\text{if }i\in[1,k^{2}],\\ 0&\text{if }i\in[k^{2}+1,n],\end{cases}\\ \end{split}</td>
<td></td>
</tr>
</tbody>
</table>
where
\(a\) $`k^{2}`$, $`t_{\ast}^{2}`$ are given by
<table>
<tbody>
<tr>
<td></td>
<td>t∗2=(1+g)T¯2+∑i=1k2πiNi0∑i=1k2πi,subscriptsuperscript𝑡21𝑔superscript¯𝑇2subscriptsuperscriptsubscript𝑘2𝑖1subscript𝜋𝑖subscriptsuperscript𝑁0𝑖subscriptsuperscriptsubscript𝑘2𝑖1subscript𝜋𝑖t^{2}_{*}=\frac{(1+g)\bar{T}^{2}+\sum^{k_{2}}_{i=1}\pi_{i}N^{0}_{i}}{\sum^{k_{2}}_{i=1}\pi_{i}},</td>
<td></td>
</tr>
</tbody>
</table>
$`k_{2}`$ can be found from the condition
<table>
<tbody>
<tr>
<td></td>
<td>φk22<T¯2≤φk2+12,subscriptsuperscript𝜑2superscript𝑘2superscript¯𝑇2subscriptsuperscript𝜑2superscript𝑘21\varphi^{2}_{k^{2}}<\bar{T}^{2}\leq\varphi^{2}_{k^{2}+1},</td>
<td></td>
</tr>
</tbody>
</table>
where
<table>
<tbody>
<tr>
<td></td>
<td>φk2=11+g∑i=1kπi(Nk0−Ni0),subscriptsuperscript𝜑2𝑘11𝑔superscriptsubscript𝑖1𝑘subscript𝜋𝑖subscriptsuperscript𝑁0𝑘subscriptsuperscript𝑁0𝑖\varphi^{2}_{k}=\frac{1}{1+g}\sum_{i=1}^{k}\pi_{i}(N^{0}_{k}-N^{0}_{i}),</td>
<td></td>
</tr>
</tbody>
</table>
for $`k \leq n`$, and $`\varphi_{n + 1}^{2} = \infty`$,
\(b\) $`k^{1}`$ and $`t_{\ast}^{1}`$ are given by
<table>
<tbody>
<tr>
<td></td>
<td>t∗1=T¯1+∑i=k2+1k1πiNi0+11+g∑i=1k2πi(gt2∗+Ni0)∑i=1k1πi,subscriptsuperscript𝑡1superscript¯𝑇1superscriptsubscript𝑖superscript𝑘21superscript𝑘1subscript𝜋𝑖subscriptsuperscript𝑁0𝑖11𝑔superscriptsubscript𝑖1superscript𝑘2subscript𝜋𝑖𝑔subscriptsuperscript𝑡2subscriptsuperscript𝑁0𝑖superscriptsubscript𝑖1superscript𝑘1subscript𝜋𝑖t^{1}_{*}=\frac{\displaystyle\bar{T}^{1}+\sum_{i={k^{2}+1}}^{k^{1}}\pi_{i}N^{0}_{i}+\frac{1}{1+g}\sum_{i=1}^{k^{2}}\pi_{i}(gt^{*}_{2}+N^{0}_{i})}{\displaystyle\sum_{i=1}^{k^{1}}\pi_{i}},</td>
<td></td>
</tr>
</tbody>
</table>
$`k^{1} \geq k^{2}`$ can be found as follows:
\(i\) $`k^{1} = k^{2}`$ if $`{\overline{T}}^{1} \leq \varphi_{k^{2} + 1}^{1}`$
\(ii\) otherwise $`k^{1}`$ is given by the condition:
<table>
<tbody>
<tr>
<td></td>
<td>φk11<T¯1≤φk1+11,subscriptsuperscript𝜑1superscript𝑘1superscript¯𝑇1subscriptsuperscript𝜑1superscript𝑘11\varphi^{1}_{k^{1}}<\bar{T}^{1}\leq\varphi^{1}_{k^{1}+1},</td>
<td></td>
</tr>
</tbody>
</table>
where
<table>
<tbody>
<tr>
<td></td>
<td>φk1=∑i=k2+1kπi(Nk0−Ni0)+11+g∑i=1k2πi((1+g)Nk0−Ni0−gt∗2)subscriptsuperscript𝜑1𝑘superscriptsubscript𝑖superscript𝑘21𝑘subscript𝜋𝑖subscriptsuperscript𝑁0𝑘subscriptsuperscript𝑁0𝑖11𝑔superscriptsubscript𝑖1superscript𝑘2subscript𝜋𝑖1𝑔subscriptsuperscript𝑁0𝑘subscriptsuperscript𝑁0𝑖𝑔subscriptsuperscript𝑡2\begin{split}\varphi^{1}_{k}&=\sum_{i=k^{2}+1}^{k}\pi_{i}(N^{0}_{k}-N^{0}_{i})\\ &+\frac{1}{1+g}\sum_{i=1}^{k^{2}}\pi_{i}\left((1+g)N^{0}_{k}-N^{0}_{i}-gt^{2}_{*}\right)\end{split}</td>
<td></td>
</tr>
</tbody>
</table>
for $`k \in {\lbrack{k_{2} + 1},n\rbrack}`$, and $`\varphi_{n + 1}^{1} = \infty`$.
######
Theorem 7
Let $`{\overline{T}}_{1} > {\overline{T}}_{2} > {\overline{T}}_{3}`$. Then, the Nash equilibrium strategies are given by
<table>
<tbody>
<tr>
<td></td>
<td>Ti1∗={t∗1−gt∗21+g−gt∗31+g+Ni01+2gif i∈[1,k3],t∗1−gt∗2+Ni01+gif i∈[k3+1,k2],t∗1−Ni0if i∈[k2+1,k1],0if i∈[k1+1,n],Ti2∗={t∗21+g−g1+gt∗3+Ni01+2gif i∈[1,k3],11+g(t∗2−Ni0)if i∈[k3+1,k2],0if i∈[k2+1,n],Ti3∗={11+2g(t∗3−Ni0)if i∈[1,k3],0if i∈[k3+1,n],subscriptsuperscript𝑇1𝑖casessubscriptsuperscript𝑡1𝑔subscriptsuperscript𝑡21𝑔𝑔subscriptsuperscript𝑡31𝑔subscriptsuperscript𝑁0𝑖12𝑔if 𝑖1superscript𝑘3subscriptsuperscript𝑡1𝑔subscriptsuperscript𝑡2subscriptsuperscript𝑁0𝑖1𝑔if 𝑖superscript𝑘31superscript𝑘2subscriptsuperscript𝑡1subscriptsuperscript𝑁0𝑖if 𝑖superscript𝑘21superscript𝑘10if 𝑖superscript𝑘11𝑛subscriptsuperscript𝑇2𝑖casessubscriptsuperscript𝑡21𝑔𝑔1𝑔subscriptsuperscript𝑡3subscriptsuperscript𝑁0𝑖12𝑔if 𝑖1superscript𝑘311𝑔subscriptsuperscript𝑡2subscriptsuperscript𝑁0𝑖if 𝑖superscript𝑘31superscript𝑘20if 𝑖superscript𝑘21𝑛subscriptsuperscript𝑇3𝑖cases112𝑔subscriptsuperscript𝑡3subscriptsuperscript𝑁0𝑖if 𝑖1superscript𝑘30if 𝑖superscript𝑘31𝑛\begin{split}T^{1*}_{i}&=\begin{cases}t^{1}_{*}-\frac{\displaystyle gt^{2}_{*}}{\displaystyle 1+g}-\frac{\displaystyle\frac{gt^{3}_{*}}{1+g}+N^{0}_{i}}{\displaystyle 1+2g}&\text{if }i\in[1,k^{3}],\\ t^{1}_{*}-\frac{\displaystyle gt^{2}_{*}+N^{0}_{i}}{\displaystyle 1+g}&\text{if }i\in[k^{3}+1,k^{2}],\\ t^{1}_{*}-N^{0}_{i}&\text{if }i\in[k^{2}+1,k^{1}],\\ 0&\text{if }i\in[k^{1}+1,n],\end{cases}\\ T^{2*}_{i}&=\begin{cases}\frac{\displaystyle t^{2}_{*}}{\displaystyle 1+g}-\frac{\displaystyle\frac{g}{1+g}t^{3}_{*}+N^{0}_{i}}{\displaystyle 1+2g}&\text{if }i\in[1,k^{3}],\\ \frac{\displaystyle 1}{\displaystyle 1+g}(t^{2}_{*}-N^{0}_{i})&\text{if }i\in[k^{3}+1,k^{2}],\\ 0&\text{if }i\in[k^{2}+1,n],\end{cases}\\ T^{3*}_{i}&=\begin{cases}\frac{\displaystyle 1}{\displaystyle 1+2g}(t^{3}_{*}-N^{0}_{i})&\text{if }i\in[1,k^{3}],\\ 0&\text{if }i\in[k^{3}+1,n],\end{cases}\\ \end{split}</td>
<td></td>
</tr>
</tbody>
</table>
where
\(a\) $`k^{3}`$, $`t_{\ast}^{3}`$ are given by
<table>
<tbody>
<tr>
<td></td>
<td>t∗3=((1+2g)T¯3+∑i=1k3πiNi0)/(∑i=1k3πi),subscriptsuperscript𝑡312𝑔superscript¯𝑇3subscriptsuperscriptsubscript𝑘3𝑖1subscript𝜋𝑖subscriptsuperscript𝑁0𝑖subscriptsuperscriptsubscript𝑘3𝑖1subscript𝜋𝑖t^{3}_{*}=((1+2g)\bar{T}^{3}+\sum^{k_{3}}_{i=1}\pi_{i}N^{0}_{i})/(\sum^{k_{3}}_{i=1}\pi_{i}),</td>
<td></td>
</tr>
</tbody>
</table>
<table>
<tbody>
<tr>
<td></td>
<td>φk33<T¯3≤φk3+13,subscriptsuperscript𝜑3superscript𝑘3superscript¯𝑇3subscriptsuperscript𝜑3superscript𝑘31\varphi^{3}_{k^{3}}<\bar{T}^{3}\leq\varphi^{3}_{k^{3}+1},</td>
<td></td>
</tr>
</tbody>
</table>
and
<table>
<tbody>
<tr>
<td></td>
<td>φk3=11+2g∑i=1kπi(Nk0−Ni0),subscriptsuperscript𝜑3𝑘112𝑔superscriptsubscript𝑖1𝑘subscript𝜋𝑖subscriptsuperscript𝑁0𝑘subscriptsuperscript𝑁0𝑖\varphi^{3}_{k}=\frac{1}{1+2g}\sum_{i=1}^{k}\pi_{i}(N^{0}_{k}-N^{0}_{i}),</td>
<td></td>
</tr>
</tbody>
</table>
for $`k \leq n`$, and $`\varphi_{n + 1}^{3} = \infty`$,
\(b\) $`k^{2}`$, $`t_{\ast}^{2}`$ are given by
<table>
<tbody>
<tr>
<td></td>
<td>t∗2=(T¯2+11+g∑i=k3+1k2πiNi0+11+2g∑i=1k3πi(gt∗31+g+Ni0))/(11+g∑i=1k2πi),subscriptsuperscript𝑡2superscript¯𝑇211𝑔superscriptsubscript𝑖superscript𝑘31superscript𝑘2subscript𝜋𝑖subscriptsuperscript𝑁0𝑖112𝑔superscriptsubscript𝑖1superscript𝑘3subscript𝜋𝑖𝑔subscriptsuperscript𝑡31𝑔subscriptsuperscript𝑁0𝑖11𝑔superscriptsubscript𝑖1superscript𝑘2subscript𝜋𝑖\begin{split}t^{2}_{*}&=\Bigl{(}\bar{T}^{2}+\frac{1}{1+g}\sum_{i={k^{3}+1}}^{k^{2}}\pi_{i}N^{0}_{i}+\frac{1}{1+2g}\sum_{i=1}^{k^{3}}\pi_{i}(\frac{gt^{3}_{*}}{1+g}+N^{0}_{i})\Bigr{)}\Bigr{/}\Bigl{(}\frac{1}{1+g}\sum_{i=1}^{k^{2}}\pi_{i}\Bigr{)},\end{split}</td>
<td></td>
</tr>
</tbody>
</table>
\(i\) $`k^{2} = k^{3}`$ if $`{\overline{T}}^{2} \leq \varphi_{k^{3} + 1}^{2}`$,
\(ii\) otherwise $`k^{2}`$ is given by the condition:
<table>
<tbody>
<tr>
<td></td>
<td>φk22<T¯2≤φk2+12subscriptsuperscript𝜑2superscript𝑘2superscript¯𝑇2subscriptsuperscript𝜑2superscript𝑘21\varphi^{2}_{k^{2}}<\bar{T}^{2}\leq\varphi^{2}_{k^{2}+1}</td>
<td></td>
</tr>
</tbody>
</table>
and
<table>
<tbody>
<tr>
<td></td>
<td>φk2=∑i=k3+1kπi1+g(Nk0−Ni0)+∑i=1k3πi(11+gNk0−Ni0+gt∗3/(1+g)1+2g).subscriptsuperscript𝜑2𝑘superscriptsubscript𝑖superscript𝑘31𝑘subscript𝜋𝑖1𝑔subscriptsuperscript𝑁0𝑘subscriptsuperscript𝑁0𝑖superscriptsubscript𝑖1superscript𝑘3subscript𝜋𝑖11𝑔subscriptsuperscript𝑁0𝑘subscriptsuperscript𝑁0𝑖𝑔subscriptsuperscript𝑡31𝑔12𝑔\begin{split}\varphi^{2}_{k}&=\sum_{i=k^{3}+1}^{k}\frac{\pi_{i}}{1+g}(N^{0}_{k}-N^{0}_{i})+\sum_{i=1}^{k^{3}}\pi_{i}\left(\frac{1}{1+g}N^{0}_{k}-\frac{N^{0}_{i}+gt^{3}_{*}/(1+g)}{1+2g}\right).\end{split}</td>
<td></td>
</tr>
</tbody>
</table>
for $`k \in {\lbrack{k^{3} + 1},n\rbrack}`$ and $`\varphi_{n + 1}^{2} = \infty`$
\(c\) $`k^{1}`$, $`t_{\ast}^{1}`$ are given by
<table>
<tbody>
<tr>
<td></td>
<td>t∗1=(T¯1+∑i=k2+1k1πiNi0+∑i=k3+1k2πigt∗2+Ni01+g+∑i=1k3πi(gt∗21+g+gt∗31+g+Ni01+2g))/∑i=1k1πi.subscriptsuperscript𝑡1superscript¯𝑇1superscriptsubscript𝑖superscript𝑘21superscript𝑘1subscript𝜋𝑖subscriptsuperscript𝑁0𝑖superscriptsubscript𝑖superscript𝑘31superscript𝑘2subscript𝜋𝑖𝑔subscriptsuperscript𝑡2subscriptsuperscript𝑁0𝑖1𝑔superscriptsubscript𝑖1superscript𝑘3subscript𝜋𝑖𝑔subscriptsuperscript𝑡21𝑔𝑔subscriptsuperscript𝑡31𝑔subscriptsuperscript𝑁0𝑖12𝑔superscriptsubscript𝑖1superscript𝑘1subscript𝜋𝑖\begin{split}t^{1}_{*}=&\Bigl{(}\displaystyle\bar{T}^{1}+\sum_{i={k^{2}+1}}^{k^{1}}\pi_{i}N^{0}_{i}+\sum_{i=k^{3}+1}^{k^{2}}\pi_{i}\frac{\displaystyle gt^{2}_{*}+N^{0}_{i}}{\displaystyle 1+g}+\sum_{i=1}^{k^{3}}\pi_{i}\Bigl{(}\frac{\displaystyle gt^{2}_{*}}{\displaystyle 1+g}+\frac{\displaystyle\frac{gt^{3}_{*}}{1+g}+N^{0}_{i}}{\displaystyle 1+2g}\Bigr{)}\Bigr{)}\Bigr{/}\sum_{i=1}^{k^{1}}\pi_{i}.\end{split}</td>
<td></td>
</tr>
</tbody>
</table>
So, $`k^{1} \geq k^{2}`$ can be found as follows:
\(i\) $`k^{1} = k^{2}`$ if $`{\overline{T}}^{1} \leq \varphi_{k^{2} + 1}^{1}`$,
\(ii\) otherwise $`k^{1}`$ is given by the condition:
<table>
<tbody>
<tr>
<td></td>
<td>φk11<T¯1≤φk1+11subscriptsuperscript𝜑1superscript𝑘1superscript¯𝑇1subscriptsuperscript𝜑1superscript𝑘11\varphi^{1}_{k^{1}}<\bar{T}^{1}\leq\varphi^{1}_{k^{1}+1}</td>
<td></td>
</tr>
</tbody>
</table>
where
<table>
<tbody>
<tr>
<td></td>
<td>φk1=∑i=k2+1kπi(Nk0−Ni0)+∑i=k3+1k2πi(Nk0−gt∗2+Ni01+g)+∑i=1k3πi(Nk0−gt∗21+g−gt∗31+g+Ni01+2g).subscriptsuperscript𝜑1𝑘superscriptsubscript𝑖superscript𝑘21𝑘subscript𝜋𝑖subscriptsuperscript𝑁0𝑘subscriptsuperscript𝑁0𝑖superscriptsubscript𝑖superscript𝑘31superscript𝑘2subscript𝜋𝑖subscriptsuperscript𝑁0𝑘𝑔subscriptsuperscript𝑡2subscriptsuperscript𝑁0𝑖1𝑔superscriptsubscript𝑖1superscript𝑘3subscript𝜋𝑖subscriptsuperscript𝑁0𝑘𝑔subscriptsuperscript𝑡21𝑔𝑔subscriptsuperscript𝑡31𝑔subscriptsuperscript𝑁0𝑖12𝑔\begin{split}\varphi^{1}_{k}&=\sum_{i={k^{2}+1}}^{k}\pi_{i}(N^{0}_{k}-N^{0}_{i})+\sum_{i=k^{3}+1}^{k^{2}}\pi_{i}\bigl{(}N^{0}_{k}-\frac{\displaystyle gt^{2}_{*}+N^{0}_{i}}{\displaystyle 1+g}\bigr{)}+\sum_{i=1}^{k^{3}}\pi_{i}\Bigl{(}N^{0}_{k}-\frac{\displaystyle gt^{2}_{*}}{\displaystyle 1+g}-\frac{\displaystyle\frac{gt^{3}_{*}}{1+g}+N^{0}_{i}}{\displaystyle 1+2g}\Bigr{)}.\end{split}</td>
<td></td>
</tr>
</tbody>
</table>
## 8 Numerical examples
Let us demonstrate the closed form approach by numerical examples. Take $`n = 5`$, $`N_{i}^{0} = \kappa^{i - 1}`$, $`\kappa = 1.7`$, $`\pi_{i} = {1/5}`$ for $`i \in {\lbrack 1,5\rbrack}`$. We consider the cases 1, 2 and 3 users scenari.
Single user scenario. Let $`\overline{T} = 5`$. Then, by Theorem [2](#ThmTheorem2 "Theorem 2 ‣ 2 Single decision maker") as the first step we calculate $`\varphi_{t}`$ for $`t \in {\lbrack 1,5\rbrack}`$. In our case we get (0, 0.14, 0.616, 1.8298, 4.58108). Thus, we have $`k = 5`$ and the optimal water-filling strategy is $`T^{\ast} = {(7.771,7.071,5.881,3.858,0.419)}`$ with payoff 1.11.
Two users scenario. Let also $`g = 0.9`$, $`{\overline{T}}^{1} = 5`$, $`{\overline{T}}^{2} = 1`$. Then, by Theorem [6](#ThmTheorem6 "Theorem 6 ‣ 7 Closed form solution for 𝐿 person game") as the first step we calculate $`\varphi_{t}^{2}`$ for $`t \in {\lbrack 1,5\rbrack}`$. In our case we get (0, 0.074, 0.324, 0.963, 2.411). Thus, $`k^{2} = 4`$ and $`t_{\ast}^{2} = 5.001`$. Then we calculate $`\varphi_{t}^{1}`$ for $`t = 5`$. In our case we get 6.994052. Thus, $`k^{1} = 4`$ and $`t_{\ast}^{1} = 0.010`$. Therefore, we have the following equilibrium strategies $`T^{1 \ast} = {(7.106,6.737,6.111,5.046,0)}`$ and $`T^{2 \ast} = {(2.106,1.737,1,111,0.0462,0)}`$ with payoffs 0.801 and 0.116, respectively.
Three users scenario. Let us introduce the third player with the average power constraint $`{\overline{T}}^{3} = 0.5`$. Then, by Theorem [7](#ThmTheorem7 "Theorem 7 ‣ 7 Closed form solution for 𝐿 person game") we can find that $`T^{1 \ast} = {(6.419,6.169,5.744,4.900,1.769)}`$, $`T^{2 \ast} = {(1.861,1.611,1.186,0.342,0)}`$ and $`T^{3 \ast} = {(1.142,0.892,0.467,0,0)}`$ are equilibrium strategies with payoffs 0.728, 0.113 and 0.055, respectively.
The equilibrium strategies of all three cases are shown in Figure [1](#S8.F1.fig1 "Figure 1 ‣ 8 Numerical examples"). When a new user comes into competition, it leads to a bigger rivalry for using good quality channels and it results in the situation when bad quality channels turn out to become more attractive for users than they were when there were smaller number of users.
Figure 1: Optimal strategies for 1, 2 and 3 user games
We have run IWFA, which produced the same values for the optimal strategies and payoffs. However, we have observed that the convergence of IWFA is slow when $`g \approx 1`$. In Figure [2](#S8.F2.fig1 "Figure 2 ‣ 8 Numerical examples"), for the two users scenario, we have plotted the total error in strategies $`{\|{T_{k}^{1} - T^{1 \ast}}\|}_{2} + {\|{T_{k}^{2} - T^{2 \ast}}\|}_{2}`$, where $`T_{k}^{i}`$ are the strategies produced by IWFA on the $`k`$-th iteration and $`T^{i \ast}`$ are the Nash equilibrium strategies. Our approach instantaneously finds the Nash equilibrium for all values of $`g`$. Also, it is interesting to note that by Theorems [6](#ThmTheorem6 "Theorem 6 ‣ 7 Closed form solution for 𝐿 person game") and [7](#ThmTheorem7 "Theorem 7 ‣ 7 Closed form solution for 𝐿 person game") the quantity of channels as well as the channels themselves used by weaker user (with smaller resources) is independent from the behavior of the stronger user (with larger resources). Of course, each user allocates his/her resources among the channels taking into account the opponent behavior.
In Figures [3](#S8.F3.fig1 "Figure 3 ‣ 8 Numerical examples") and [4](#S8.F4.fig1 "Figure 4 ‣ 8 Numerical examples"), we compare the non-cooperative approach with the cooperative approach. Specifically, we compare the transmission rates and their sum under Nash equilibrium strategies and under strategies obtained from the centralized optimization of the sum of users’ rates. The main conclusions are: the cost of anarchy is nearly zero for $`g \in {\lbrack 0,{1/4}\rbrack}`$ and then it grows up to 22% when $`g`$ grows from $`1/4`$ to $`1`$; the user with more resources gains significantly more from the centralized optimization. Hence, the non-cooperative approach results in a more fair resource distribution. In Figure [4](#S8.F4.fig1 "Figure 4 ‣ 8 Numerical examples") we plot the total transmission rate under Nash equilibrium strategies and under strategies obtained from the centralized optimization for the cases of 2 and 3 users. As expected the introduction of a new user increases the cost of anarchy. Furthermore, in the case of the centralized optimization with the introduction of a new user the total rate increases, and on contrary in the game setting the total rate decreases.
Figure 2: Convergence of IWFA
Figure 3: Centralized Optimization vs. Game
Figure 4: The effect of a new user
## 9 Conclusion
We have considered power control for wireless networks in optimization and game frameworks. Closed form solutions for the water filling optimization problem and $`L`$ users symmetric water filling games have been provided. Namely, now one can calculate optimal/equilibrium strategies with a finite number of arithmetic operations. This was possible due to the intrinsic hierarchical structure induced by the quantity of the resources available to the users. We have also provided a simple alternative proof of convergence for a version of iterative water filling algorithm. It had been known before that the iterative water filling algorithm converges very slow when the crosstalk coefficient is close to one. For our closed form approach possible proximity of the crosstalk coefficient to one is not a problem. We have shown that when the crosstalk coefficient is equal to one, there is a continuum of Nash equilibria. Finally, we have demonstrated that the price of anarchy is small when the crosstalk coefficient is small and that the decentralized solution is better than the centralized one with respect to fairness.
## References
- \[1\]
E. Altman, K. Avrachenkov, A. Garnaev, “A jamming game in wireless networks with transmission cost”. in Proc. of NET-COOP 2007. Lecture Notes in Computer Science, v.4465, pp.1-12, 2007.
- \[2\]
E. Altman, K. Avrachenkov, G. Miller and B. Prabhu, “Discrete power control: cooperative and non-cooperative optimization”, in Proceedings of IEEE INFOCOM 2007. An extended version is available as INRIA Research Report no.5818.
- \[3\]
T. Cover and J. Thomas, Elements of Information Theory, Wiley, 1991.
- \[4\] W. R. Heinzelman, A. Chandrakasan, and H. Balakrishnan, “Energy-efficient communication protocol for wireless microsensor networks,” in *Proc. of the 33rd Annual Hawaii International Conference on System Sciences*, v.2, Jan. 2000.
- \[5\]
A. Garnaev, Search Games and Other Applications of Game Theory, Springer, 2000.
- \[6\]
A.J. Goldsmith and P.P. Varaiya, “Capacity of fading channels with channel side information”, IEEE Trans. Information Theory, v.43(6), pp.1986-1992, 1997.
- \[7\] T. J. Kwon and M. Gerla, “Clustering with power control,” in *Proc. IEEE Military Communications Conference (MILCOM’99)*, v.2, Atlantic City, NJ, USA, 1999, pp.1424–1428.
- \[8\]
L. Lai and H. El Gamal, “The water-filling game in fading multiple access channels”, submitted to IEEE Trans. Information Theory, 2005.
- \[9\] C. R. Lin and M. Gerla, “Adaptive clustering for mobile wireless networks,” *IEEE JSAC*, v.15, no.7, pp.1265–1275, 1997.
- \[10\]
Z.-Q. Luo and J.-S. Pang, “Analysis of iterative waterfilling algorithm for multiuser power control in digital subscriber lines”, EURASIP Journal on Applied Signal Processing, 2006.
- \[11\]
O. Popescu and C. Rose, “Water filling may not good neighbors make”, in Proceedings of GLOBECOM 2003, v.3, pp.1766–1770, 2003.
- \[12\]
D.C. Popescu, O. Popescu and C. Rose, “Interference avoidance versus iterative water filling in multiaccess vector channels”, in Proceedings of IEEE VTC 2004 Fall, v.3, pp.2058–2062, 2004.
- \[13\]
K.B. Song, S.T. Chung, G. Ginis and J.M. Cioffi, “Dynamic spectrum management for next-generation DSL systems, IEEE Communications Magazine, v.40, pp.101–109, 2002.
- \[14\]
D. Tse and P. Viswanath, Fundamentals of Wireless Communication, Cambridge University Press, 2005.
- \[15\]
W. Yu, Competition and cooperation in multi-user communication environements, PhD Thesis, Stanford University, June 2002.
- \[16\]
W. Yu, G. Ginis and J.M. Cioffi, “Distributed multiuser power control for digital subscriber lines”, IEEE JSAC, v.20, pp.1105–1115, 2002.
###### Contents
1. [](#S1)
1 Introduction
2. [](#S2)
2 Single decision maker
3. [](#S3)
3 Symmetric water filling game
4. [](#S4)
4 A recursive approach to the symmetric water filling game
5. [](#S5)
5 A water-filling algorithm
6. [](#S6)
6 Existence and uniqueness of the Nash equilibrium
7. [](#S7)
7 Closed form solution for $`L`$ person game
8. [](#S8)
8 Numerical examples
9. [](#S9)
9 Conclusion |
0708.2571 | cs.CR | cyber | On the Security of the Cha-Ko-Lee-Han-Cheon Braid Group Public-key Cryptosystem
Milton M. Chowdhury
1\. Abstract
We show that a number of cryptographic protocols using non-commutative semigroups including the Cha-Ko-Lee-Han-Cheon braid group cryptosystem have security based on the MSCSP. We give two algorithms to solve the DP using the MSCSP.
2\. Introduction
At the CRYPTO 2000 conference the seminal KLCHKP (Ko-Lee-Cheon-Han-Kang-Park) braid group public-key cryptosystem was published see \[[2](#bib.bib2)\]. An updated version of the KLCHKP cryptosystem which is the CKLHC (Cha-Ko-Lee-Han-Cheon) braid group cryptosystem was introduced at ASIACRYPT 2001 conference \[[10](#bib.bib10)\] the claim of the authors was the updated cryptosystem is based on the DH-DP (Diffie-Hellman Decomposition Problem). We show that the KLCHKP and CKLHC cryptosystems are based on the MSCSP and it has been assumed for several years the security of these cryptosystems are based on the DH-CP and DH-DP respectively, we also show the related cryptosystems may be based on the MSCSP and hence give a new way to break the KLCHKP and CKLHC cryptosystems and the related cryptosystems for some parameters. It has been shown there is a linear algebraic attack on the KLCHKP and CKLHC cryptosystems but our attack is more practical.
3\. Hard Problems in Non-Abelian Groups
Definition-The MSCSP (multiple simultaneous conjugacy search problem) \[[3](#bib.bib3)\] is find elements $`g \in G`$ such that $`y_{i} = {g\hspace{0pt}x_{i}\hspace{0pt}g^{- 1}}`$, given the publicly known information: $`G`$ is a group, $`{x_{i},y_{i}} \in G`$ with $`{x_{i},y_{i}} = {a\hspace{0pt}x_{i}\hspace{0pt}a^{- 1}}`$, $`{1 \leq i \leq u},`$ with the secret element $`{a \in G}.`$
Definition-The CSP \[[3](#bib.bib3)\] can be defined as the MSCSP with $`u = 1`$.
Notation-We refer an example of the MSCSP as $`({(x_{1},x_{2},\ldots,x_{u})},{(y_{1},y_{2},\ldots,y_{u})})`$ with solution $`g.`$
The DP (Decomposition Problem) \[[6](#bib.bib6)\] is defined as follows.
Public Information: $`G`$ is a semigroup, $`A`$ is a subset of $`G`$. $`{x,y} \in G`$ with $`y = {a\hspace{0pt}x\hspace{0pt}b}`$.
Secret information: $`{a,b} \in A`$.
Objective: find elements $`{f,g} \in A`$ such that $`{f\hspace{0pt}x\hspace{0pt}g} = y`$.
The definition of the $`D\hspace{0pt}P`$ above generalises the definition of a less general version of the $`D\hspace{0pt}P`$ given in \[[8](#bib.bib8)\], \[[3](#bib.bib3)\] and \[[7](#bib.bib7)\]. The less general version only differs from the above definition of $`D\hspace{0pt}P`$ because $`G`$ is a group and $`A`$ is a subgroup. In our notation in all of this paper we omit the binary operation $`\ast`$ when writing products so for example $`f \ast x \ast g`$ is understood to mean $`f\hspace{0pt}x\hspace{0pt}g`$. We require that $`\ast`$ is efficiently computable.
The CSP (Conjugacy Search Problem) \[[1](#bib.bib1)\], \[[3](#bib.bib3)\] is defined as follows.
Public Information: $`G`$ is a group. $`{x,y} \in G`$ with $`y = {f^{- 1}\hspace{0pt}x\hspace{0pt}f}`$.
Secret Information: $`f \in G`$.
Objective: find an element $`g \in G`$ such that $`{g^{- 1}\hspace{0pt}x\hspace{0pt}g} = y`$.
Notation-We refer an example of the CSP as $`(x,y)`$ with solution $`g.`$
The DH-DP (Diffie-Hellman Decomposition Problem) \[[8](#bib.bib8)\], \[[3](#bib.bib3)\] is defined as follows.
Public Information: $`G`$ is a group. $`A,B`$ are subgroups of $`G`$ with $`{\lbrack A,B\rbrack} = 1`$. $`{x,y_{a},y_{b}} \in G`$ with $`{y_{a} = {a\hspace{0pt}x\hspace{0pt}b}},`$ $`y_{b} = {c\hspace{0pt}x\hspace{0pt}d}`$.
Secret Information: $`{a,b} \in A`$, $`{c,d} \in B`$.
Objective: find the element $`cy_{a}d{( = ay_{b}b = acxbd)}.`$
The DH-CP (Diffie-Hellman Conjugacy Problem) is the specialisation of the DH-DP \[[8](#bib.bib8)\] with $`a = b^{- 1}`$ and $`c = d^{- 1}`$.
We now re-define the DP and DH-DP above as used in our key agreement protocol given in \[[12](#bib.bib12)\]. In the rest of this paper below the DP and DH-DP will mean their re-definitions.
The re-definition of the DP is as follows.
Public Information: $`G`$ is a semigroup. $`A`$, $`B`$ are subsets of $`G`$. $`{x,y} \in G`$ with $`y = {a\hspace{0pt}x\hspace{0pt}b}`$.
Secret Information: $`a \in A`$, $`b \in B`$.
Objective: find elements $`f \in A`$, $`g \in B`$ such that $`{f\hspace{0pt}x\hspace{0pt}g} = y`$.
The re-definition of the DH-DP is as follows.
Public information: $`G`$ is a semigroup. $`A,B,C,D`$ are subsets of $`G`$. $`{x,y_{a},y_{b}} \in G`$ with $`y_{a} = {a\hspace{0pt}x\hspace{0pt}b}`$, $`y_{b} = {c\hspace{0pt}x\hspace{0pt}d}`$.
Secret Information: $`a \in A`$, $`b \in B`$, $`c \in C`$, $`d \in D`$.$`{{\lbrack A,C\rbrack} = 1},{{\lbrack B,D\rbrack} = 1}`$
Objective: find the element $`c\hspace{0pt}y_{a}\hspace{0pt}d`$ $`({= {a\hspace{0pt}y_{b}\hspace{0pt}b} = {a\hspace{0pt}c\hspace{0pt}x\hspace{0pt}b\hspace{0pt}d}})`$.
The EDL problem is to decide if the discrete logarithm of two elements in an abelian group are the same \[[9](#bib.bib9)\]. The EDL type problem is as follows \[[9](#bib.bib9)\].
Public information: $`G`$ is a group. $`{a,b,y_{a},y_{b}} \in G`$ with $`y_{a} = {u\hspace{0pt}a\hspace{0pt}v}`$, $`y_{b} = {w\hspace{0pt}b\hspace{0pt}x}`$.
Secret Information: $`{u,v,w},`$ $`x \in G`$.
Objective: Decide if $`{{F_{a}\hspace{0pt}{(y_{a})}} \cap {F_{b}\hspace{0pt}{(y_{b})}}} \neq \varnothing`$. Where $`F_{\beta}{(\alpha)} = {\{{(a,b)} \in B_{n} \times B_{n}}`$ :$`\alpha = a\beta b\}.`$
We redefine the EDL problem more generally as follows.
Public information: $`G`$ is a Semigroup. $`A,B,C,D`$ are subsets of $`G.`$ $`{a,b,y_{a},y_{b}} \in G`$ with $`y_{a} = {u\hspace{0pt}a\hspace{0pt}v}`$, $`y_{b} = {w\hspace{0pt}b\hspace{0pt}x}`$.
Secret Information: $`u \in A`$, $`v \in B`$, $`w \in C`$, $`x \in D`$.
Objective: Decide if $`{{F_{a}\hspace{0pt}{(y_{a})}} \cap {F_{b}\hspace{0pt}{(y_{b})}}} \neq \varnothing`$. Where $`F_{\beta}{(\alpha)} = {\{{(a,b)} \in B_{n} \times B_{n}}`$ :$`\alpha = a\beta b\}.`$
4\. Key Agreement Protocol Using Non-Commutative Semigroups
In \[[12](#bib.bib12)\] we introduced a key agreement protocol and a variant of it which we briefly describe below.
- •
Phase 0. Initial setup
i) $`G`$ is chosen and is publicly known.
A first method to select the parameters is to select publicly known subsets or privately known subsets $`L_{A}`$, $`L_{B}`$, $`R_{A}`$, $`R_{B}`$ and $`Z`$ of $`G`$ are chosen for which either property a) below is true or property b) below is true. Let $`z \in Z`$ with $`z`$ the publicly known element which is the value of $`x`$ in the definition of the DH-DP used in the example of the DH-DP in our new authentication scheme.
Following \[[7](#bib.bib7)\] let $`g \in G`$ for $`G`$ a group, $`C_{G}\hspace{0pt}{(g)}`$ is the centraliser of $`g`$, we sketch the modifications to the authentication scheme (and these apply to the key agreement protocol described below) to give two further methods to select the subgroups as follows. Publicly known subsets or privately known $`L_{A}`$, $`L_{B}`$, $`R_{A}`$, $`R_{B}`$ and $`Z`$ of $`G`$ are chosen for which either property a) below is true or property b) below for the second and third methods below.
The second method to select the subgroups is $`A`$ chooses $`{(a_{1},a_{2})} \in {G \times G}`$ and publishes the subgroups as a set of generators of the centralisers $`{{L_{B},R_{B},L_{B}} \subseteq {C_{G}\hspace{0pt}{(a_{1})}}},{{R_{B} \subseteq {C_{G}\hspace{0pt}{(a_{2})}}},{L_{B} = {\{\alpha_{1},\ldots,\alpha_{k}\}}}}`$ etc. $`B`$ chooses $`{(b_{1},b_{2})} \in {L_{B} \times R_{B}}`$, and hence can compute $`x`$ below etc. Following \[[7](#bib.bib7)\] there is no explicit indication of where to select $`a_{1}`$ and/or $`a_{2}`$ from. Hence before attempting a length based attack in this case the attacker has to compute the centraliser of $`L_{B},R_{B}`$.
So a third method to select the subgroups is
$`A`$ chooses $`{L_{A} = G},{a_{1} \in G}`$, and publishes $`{L_{B} \subseteq {C_{G}\hspace{0pt}{(a_{1})}}},{L_{B} = {\{\alpha_{1},\ldots,\alpha_{k}\}}}`$,
$`B`$ chooses $`{L_{B} = G},{b_{2} \in G}`$, and publishes $`{R_{A} \subseteq {C_{G}\hspace{0pt}{(b_{2})}}},{R_{A} = {\{\beta_{1},\ldots,\beta_{k}\}}}`$,
Hence $`A`$ chooses $`{(a_{1},a_{2})} \in {{G \times C_{G}}\hspace{0pt}{(b_{2})}}`$ and publishes the subgroup(s) as a set of generators of the centralisers $`B`$ chooses $`{(b_{1},b_{2})} \in {{C_{G}\hspace{0pt}{(a_{1})}} \times G}`$, and hence can compute $`x`$ etc. Again there is no explicit indication of where to select $`a_{1}`$ and/or $`b_{2}`$ from. Hence before attempting a length based attack in this case the attacker has to compute the centraliser of $`L_{B}`$ and/or $`R_{B}`$.
a\) If $`z \neq e`$ we require the following conditions
<table>
<tbody>
<tr>
<td></td>
<td>[LA,LB]=1,[RA,RB]=1,[LB,Z]≠1,[LA,Z]≠1,[RB,Z]≠1,[RA,Z]≠1,[LA,RA]≠1,[LB,RB]≠1.subscript𝐿𝐴subscript𝐿𝐵1subscript𝑅𝐴subscript𝑅𝐵1subscript𝐿𝐵𝑍1subscript𝐿𝐴𝑍1subscript𝑅𝐵𝑍1subscript𝑅𝐴𝑍1subscript𝐿𝐴subscript𝑅𝐴1subscript𝐿𝐵subscript𝑅𝐵1\begin{array}[t]{ll}[L_{A},L_{B}]=1,&[R_{A},R_{B}]=1,\\ [L_{B},Z]\neq 1,&[L_{A},Z]\neq 1,\\ [R_{B},Z]\neq 1,&[R_{A},Z]\neq 1,\\ [L_{A},R_{A}]\neq 1,&[L_{B},R_{B}]\neq 1.\end{array}</td>
<td></td>
<td>(2)</td>
</tr>
</tbody>
</table>
All the above conditions for $`z \neq e`$ can arise by generalising from properties of subgroups used in the SDG or CKLHC schemes for example the second and third conditions in (2) arise from the observations that in general $`{{\lbrack{L\hspace{0pt}B_{n}},B_{n}\rbrack} \neq 1},{{\lbrack{L\hspace{0pt}B_{n}},{U\hspace{0pt}B_{n}}\rbrack} = 1}`$.
b) If $`z = e`$ we require the following conditions
<table>
<tbody>
<tr>
<td></td>
<td>[LA,LB]=1,[RA,RB]=1,[LA,RA]≠1,[LB,RB]≠1,[LB,RA]≠1,[LA,RB]≠1.subscript𝐿𝐴subscript𝐿𝐵1subscript𝑅𝐴subscript𝑅𝐵1subscript𝐿𝐴subscript𝑅𝐴1subscript𝐿𝐵subscript𝑅𝐵1subscript𝐿𝐵subscript𝑅𝐴1subscript𝐿𝐴subscript𝑅𝐵1\begin{array}[t]{ll}[L_{A},L_{B}]=1,&[R_{A},R_{B}]=1,\\ [L_{A},R_{A}]\neq 1,&[L_{B},R_{B}]\neq 1,\\ [L_{B},R_{A}]\neq 1,&[L_{A},R_{B}]\neq 1.\end{array}</td>
<td></td>
<td>(3)</td>
</tr>
</tbody>
</table>
.
- •
Choose $`z \in B_{n}`$.
ii) A(lice) chooses a secret braid $`a_{1} \in L_{A}`$, $`a_{2} \in R_{A}`$, her private key; she publishes $`K_{A} = {a_{1}\hspace{0pt}z\hspace{0pt}a_{2}}`$; the pair $`(w,K_{A})`$ is the public key.
i) B(ob) chooses a secret braid $`b_{1} \in L_{B}`$, $`b_{2} \in R_{B}`$, her private key; she publishes $`K_{B} = {b_{1}\hspace{0pt}z\hspace{0pt}b_{2}}`$; the pair $`(w,K_{B})`$ is the public key.
iii) A and B can compute the common shared secret key $`\kappa`$ as $`\kappa = {a_{1}\hspace{0pt}K_{B}\hspace{0pt}a_{2}}`$ and $`\kappa = {b_{1}\hspace{0pt}K_{A}\hspace{0pt}b_{2}}`$ respectively.
i) Choose a public $`w`$
$`h`$ is a fixed collision-free hash function from braids to sequences of 0’s and 1’s or, possibly, to braids, for which this choice for $`h`$. Again the above protocol is considered with the commutativity conditions 2 or 3. Note the braids $`K_{A}`$ and $`K_{B}`$ are rewritten for example a normal form to make the protocol secure. Full detail are given in \[[12](#bib.bib12)\]. It was shown in \[[12](#bib.bib12)\] that the above key agreement protocol is a generalisation of the key protocols given in \[[10](#bib.bib10)\],\[[2](#bib.bib2)\],\[[6](#bib.bib6)\],\[[7](#bib.bib7)\].
5\. The Diffie-Hellman Decomposition Problem is Equivalent to the
Multiple Simultaneous Conjugacy Search Problem
In this section we will show that the DH-DP is equivalent to the MSCSP in our key agreement protocol in section 2.1 hence showing the key exchange protocols given in \[[10](#bib.bib10)\],\[[2](#bib.bib2)\],\[[11](#bib.bib11)\] are based on the MSCSP and the key exchange protocols in \[[6](#bib.bib6)\],\[[7](#bib.bib7)\] may be based on the MSCSP. In the braid group there are various algorithms MSCSP can be solved with non-negligible probability such as length based algorithms or the algorithm using ultra summit sets given in \[[4](#bib.bib4)\] so for braid group implementations the algorithms we show are based on the MSCSP should not be used. Our result also applies to the variant key exchange protocol and variant authentication scheme given in \[[12](#bib.bib12)\].
We now introduce the concept of a CE (conjugacy extractor) function which we build our attack upon.
Notation-We define a CE (conjugacy extractor) function to be a function that on input of information from an user and information transmitted in a cryptographic protocol gives as its output a conjugacy equation, by conjugacy equation we mean an instantiation of the CSP. We denote $`i`$ CE functions as CE<sub>i</sub>, or CE if there is just one function involved.
Theorem 1
Solving the DH-DP is equivalent solving the MSCSP assuming $`y_{a}`$ and/or $`y_{b}`$ are invertible elements in the DH-DP.
Proof
The proof is for when considering commutativity conditions 2 or 3 in the generalised protocol above, when condition 3 are used then $`x = e`$. Firstly define the CE for a protocol based on the DH-DP as follows
<table>
<tbody>
<tr>
<td></td>
<td>CE1(RI,ya)𝐶subscript𝐸1subscript𝑅𝐼subscript𝑦𝑎\displaystyle CE_{1}(R_{I},y_{a})</td>
<td>=\displaystyle=</td>
<td>yaRIya−1=axbRIb−1x−1a−1=axRIx−1a−1,RI∈Dformulae-sequencesubscript𝑦𝑎subscript𝑅𝐼superscriptsubscript𝑦𝑎1𝑎𝑥𝑏subscript𝑅𝐼superscript𝑏1superscript𝑥1superscript𝑎1𝑎𝑥subscript𝑅𝐼superscript𝑥1superscript𝑎1subscript𝑅𝐼𝐷\displaystyle y_{a}R_{I}y_{a}^{-1}=axbR_{I}b^{-1}x^{-1}a^{-1}=axR_{I}x^{-1}a^{-1},R_{I}\in D</td>
<td></td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td>CE2(SI,yb)𝐶subscript𝐸2subscript𝑆𝐼subscript𝑦𝑏\displaystyle CE_{2}(S_{I},y_{b})</td>
<td>=\displaystyle=</td>
<td>ybSIyb−1=cxdSId−1x−1c−1=cxSIx−1c−1,SI∈Bformulae-sequencesubscript𝑦𝑏subscript𝑆𝐼superscriptsubscript𝑦𝑏1𝑐𝑥𝑑subscript𝑆𝐼superscript𝑑1superscript𝑥1superscript𝑐1𝑐𝑥subscript𝑆𝐼superscript𝑥1superscript𝑐1subscript𝑆𝐼𝐵\displaystyle y_{b}S_{I}y_{b}^{-1}=cxdS_{I}d^{-1}x^{-1}c^{-1}=cxS_{I}x^{-1}c^{-1},S_{I}\in B</td>
<td></td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td>CE3(RI′,ya)𝐶subscript𝐸3superscriptsubscript𝑅𝐼′subscript𝑦𝑎\displaystyle CE_{3}(R_{I}^{\prime},y_{a})</td>
<td>=\displaystyle=</td>
<td>ya−1RI′ya=b−1x−1a−1RI′axb=b−1x−1RI′xb,RI′∈Cformulae-sequencesuperscriptsubscript𝑦𝑎1superscriptsubscript𝑅𝐼′subscript𝑦𝑎superscript𝑏1superscript𝑥1superscript𝑎1superscriptsubscript𝑅𝐼′𝑎𝑥𝑏superscript𝑏1superscript𝑥1superscriptsubscript𝑅𝐼′𝑥𝑏superscriptsubscript𝑅𝐼′𝐶\displaystyle y_{a}^{-1}R_{I}^{\prime}y_{a}=b^{-1}x^{-1}a^{-1}R_{I}^{\prime}axb=b^{-1}x^{-1}R_{I}^{\prime}xb,R_{I}^{\prime}\in C</td>
<td></td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td>CE4(SI′,yb)𝐶subscript𝐸4superscriptsubscript𝑆𝐼′subscript𝑦𝑏\displaystyle CE_{4}(S_{I}^{{}^{\prime}},y_{b})</td>
<td>=\displaystyle=</td>
<td>yb−1SI′yb=d−1x−1c−1SI′cxd=d−1x−1SI′xd,SI′∈Aformulae-sequencesuperscriptsubscript𝑦𝑏1superscriptsubscript𝑆𝐼′subscript𝑦𝑏superscript𝑑1superscript𝑥1superscript𝑐1superscriptsubscript𝑆𝐼′𝑐𝑥𝑑superscript𝑑1superscript𝑥1superscriptsubscript𝑆𝐼′𝑥𝑑superscriptsubscript𝑆𝐼′𝐴\displaystyle y_{b}^{-1}S_{I}^{\prime}y_{b}=d^{-1}x^{-1}c^{-1}S_{I}^{\prime}cxd=d^{-1}x^{-1}S_{I}^{\prime}xd,S_{I}^{{}^{\prime}}\in A</td>
<td></td>
</tr>
</tbody>
</table>
$`R_{I}`$ is chosen from the subset that commutes with the subset that the secret $`a`$ in $`y_{a}`$ is chosen from. $`S_{I}`$ is chosen from the subset that commutes with the subset that the secret $`b`$ in $`y_{b}`$ is chosen from etc. Since all the parameters are known to compute the $`C\hspace{0pt}E_{1}`$,…,$`C\hspace{0pt}E_{4}`$ are easily computable. Note it is sufficient for one $`C\hspace{0pt}E`$ to exist to prove the theorem but we may want to compute more than one $`C\hspace{0pt}E`$ because their difficulty may vary, for example one of the $`C\hspace{0pt}E`$ of the protocols \[[7](#bib.bib7)\] can be used in a known length attack. Obviously $`R_{I}`$ (in general) does not commute with $`x`$ (similarly $`R_{I}`$ in general does not commute with $`a`$ when conditions 3 are used) as this would mean an attacker could easily recover the common secret key. So for $`1 \leq I \leq u`$ this shows that the key agreement protocol in \[[2](#bib.bib2)\] is based on the MSCSP for the secret in either $`a\hspace{0pt}x`$ or the MSCSP in the secret $`c\hspace{0pt}x`$. So $`a`$ or $`c`$ can be found by right multiplying by $`x^{- 1}`$ which is publicly known. Hence the protocols in \[[10](#bib.bib10)\], \[[2](#bib.bib2)\],\[[6](#bib.bib6)\],\[[7](#bib.bib7)\],\[[11](#bib.bib11)\] are based on an example of the MSCSP as
<table>
<tbody>
<tr>
<td></td>
<td>((R1,R2,…,Ru),(CE1(R1,ya),CE1(R2,ya),…,CE1(Ru,ya)))with solution ax.subscript𝑅1subscript𝑅2…subscript𝑅𝑢𝐶subscript𝐸1subscript𝑅1subscript𝑦𝑎𝐶subscript𝐸1subscript𝑅2subscript𝑦𝑎…𝐶subscript𝐸1subscript𝑅𝑢subscript𝑦𝑎with solution 𝑎𝑥\displaystyle((R_{1},R_{2},...,R_{u}),(CE_{1}(R_{1},y_{a}),CE_{1}(R_{2},y_{a}),...,CE_{1}(R_{u},y_{a})))\text{with solution }ax.</td>
<td></td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td>((S1,S2,…,Su),(CE2(S1,yb),CE2(S2,yb),…,CE2(Su,yb)))with solution cx.subscript𝑆1subscript𝑆2…subscript𝑆𝑢𝐶subscript𝐸2subscript𝑆1subscript𝑦𝑏𝐶subscript𝐸2subscript𝑆2subscript𝑦𝑏…𝐶subscript𝐸2subscript𝑆𝑢subscript𝑦𝑏with solution 𝑐𝑥\displaystyle((S_{1},S_{2},...,S_{u}),(CE_{2}(S_{1},y_{b}),CE_{2}(S_{2},y_{b}),...,CE_{2}(S_{u},y_{b})))\text{with solution }cx.</td>
<td></td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td>((R1′,R2′,…,Ru′),(CE1(R1′,ya),CE1(R2′,ya),…,CE1(Ru′,ya)))with solution b−1x−1.superscriptsubscript𝑅1′superscriptsubscript𝑅2′…superscriptsubscript𝑅𝑢′𝐶subscript𝐸1superscriptsubscript𝑅1′subscript𝑦𝑎𝐶subscript𝐸1superscriptsubscript𝑅2′subscript𝑦𝑎…𝐶subscript𝐸1superscriptsubscript𝑅𝑢′subscript𝑦𝑎with solution superscript𝑏1superscript𝑥1\displaystyle((R_{1}^{\prime},R_{2}^{\prime},...,R_{u}^{\prime}),(CE_{1}(R_{1}^{\prime},y_{a}),CE_{1}(R_{2}^{\prime},y_{a}),...,CE_{1}(R_{u}^{\prime},y_{a})))\text{with solution }b^{-1}x^{-1}.</td>
<td></td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td>((S1′,S2′,…,Su′),(CE2(S1′,yb),CE2(S2′,yb),…,CE2(Su′,yb)))with solution d−1x−1.superscriptsubscript𝑆1′superscriptsubscript𝑆2′…superscriptsubscript𝑆𝑢′𝐶subscript𝐸2superscriptsubscript𝑆1′subscript𝑦𝑏𝐶subscript𝐸2superscriptsubscript𝑆2′subscript𝑦𝑏…𝐶subscript𝐸2superscriptsubscript𝑆𝑢′subscript𝑦𝑏with solution superscript𝑑1superscript𝑥1\displaystyle((S_{1}^{\prime},S_{2}^{\prime},...,S_{u}^{\prime}),(CE_{2}(S_{1}^{\prime},y_{b}),CE_{2}(S_{2}^{\prime},y_{b}),...,CE_{2}(S_{u}^{\prime},y_{b})))\text{with solution }d^{-1}x^{-1}.</td>
<td></td>
</tr>
</tbody>
</table>
We now give applications of our theorem for the protocols \[[10](#bib.bib10)\], \[[2](#bib.bib2)\],\[[6](#bib.bib6)\],\[[7](#bib.bib7)\],\[[11](#bib.bib11)\] there are algorithm that solve the MSCSP with non-negligible probability such as a length attack \[[5](#bib.bib5)\], so a length attack may be used for the protocol \[[10](#bib.bib10)\],\[[2](#bib.bib2)\],\[[6](#bib.bib6)\], \[[7](#bib.bib7)\]. Following the notation in \[[7](#bib.bib7)\] , where $`G`$ is a group so the security of the protocol in \[[7](#bib.bib7)\] is always based on the MSCSP (because we know the generators for the elements $`a_{2}`$ and $`b_{1}`$ we can use a length attack so disproving the claim in \[[7](#bib.bib7)\] ), we have
<table>
<tbody>
<tr>
<td></td>
<td>CE1(AI,PA)𝐶subscript𝐸1subscript𝐴𝐼subscript𝑃𝐴\displaystyle CE_{1}(A_{I},P_{A})</td>
<td>=\displaystyle=</td>
<td>PA−1AIPA=a2−1w−1a1−1AIa1wa2=a2w−1AIwa2,superscriptsubscript𝑃𝐴1subscript𝐴𝐼subscript𝑃𝐴superscriptsubscript𝑎21superscript𝑤1superscriptsubscript𝑎11subscript𝐴𝐼subscript𝑎1𝑤subscript𝑎2subscript𝑎2superscript𝑤1subscript𝐴𝐼𝑤subscript𝑎2\displaystyle P_{A}^{-1}A_{I}P_{A}=a_{2}^{-1}w^{-1}a_{1}^{-1}A_{I}a_{1}wa_{2}=a_{2}w^{-1}A_{I}wa_{2},</td>
<td></td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td>AIsubscript𝐴𝐼\displaystyle A_{I}</td>
<td>∈\displaystyle\in</td>
<td>A , A is a subgroup of CG(a1)𝐴 , 𝐴 is a subgroup of subscript𝐶𝐺subscript𝑎1\displaystyle A\text{ , }A\text{ is a subgroup of }C_{G}(a_{1})</td>
<td></td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td>CE2(BI,PB)𝐶subscript𝐸2subscript𝐵𝐼subscript𝑃𝐵\displaystyle CE_{2}(B_{I},P_{B})</td>
<td>=\displaystyle=</td>
<td>PBBIPB−1=b1wb2BIb2−1w−1b1−1=b1wBIw−1b1−1,subscript𝑃𝐵subscript𝐵𝐼superscriptsubscript𝑃𝐵1subscript𝑏1𝑤subscript𝑏2subscript𝐵𝐼superscriptsubscript𝑏21superscript𝑤1superscriptsubscript𝑏11subscript𝑏1𝑤subscript𝐵𝐼superscript𝑤1superscriptsubscript𝑏11\displaystyle P_{B}B_{I}P_{B}^{-1}=b_{1}wb_{2}B_{I}b_{2}^{-1}w^{-1}b_{1}^{-1}=b_{1}wB_{I}w^{-1}b_{1}^{-1},</td>
<td></td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td>RIsubscript𝑅𝐼\displaystyle R_{I}</td>
<td>∈\displaystyle\in</td>
<td>B, B is a subgroup of CG(b2)𝐵, 𝐵 is a subgroup of subscript𝐶𝐺subscript𝑏2\displaystyle B\text{, }B\text{ is a subgroup of }C_{G}(b_{2})</td>
<td></td>
</tr>
</tbody>
</table>
if it may be easy for a some sets $`\{ g_{1},g_{2},\ldots,g_{k}\}`$ of the elements of $`G`$ to compute a part of or all (if $`G`$ is the braid group there are algorithms that will compute a large part of the centraliser)
<table>
<tbody>
<tr>
<td></td>
<td>C(g1,…,gk)=C(g1)∩…C(gk)𝐶subscript𝑔1…subscript𝑔𝑘𝐶subscript𝑔1…𝐶subscript𝑔𝑘C(g_{1},...,g_{k})=C(g_{1})\cap...C(g_{k})</td>
<td></td>
</tr>
</tbody>
</table>
then we can compute the following
<table>
<tbody>
<tr>
<td></td>
<td>CE3(CI,PA)𝐶subscript𝐸3subscript𝐶𝐼subscript𝑃𝐴\displaystyle CE_{3}(C_{I},P_{A})</td>
<td>=\displaystyle=</td>
<td>PACIPA−1=a1wa2CIa2−1w−1a1−1=a1wAIw−1a1−1,subscript𝑃𝐴subscript𝐶𝐼superscriptsubscript𝑃𝐴1subscript𝑎1𝑤subscript𝑎2subscript𝐶𝐼superscriptsubscript𝑎21superscript𝑤1superscriptsubscript𝑎11subscript𝑎1𝑤subscript𝐴𝐼superscript𝑤1superscriptsubscript𝑎11\displaystyle P_{A}C_{I}P_{A}^{-1}=a_{1}wa_{2}C_{I}a_{2}^{-1}w^{-1}a_{1}^{-1}=a_{1}wA_{I}w^{-1}a_{1}^{-1},</td>
<td></td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td>CIsubscript𝐶𝐼\displaystyle C_{I}</td>
<td>∈\displaystyle\in</td>
<td>C,C a subgroup of CG(b2)𝐶,𝐶 a subgroup of subscript𝐶𝐺subscript𝑏2\displaystyle C\text{,}C\text{ a subgroup of }C_{G}(b_{2})</td>
<td></td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td>CE4(DI,PB)𝐶subscript𝐸4subscript𝐷𝐼subscript𝑃𝐵\displaystyle CE_{4}(D_{I},P_{B})</td>
<td>=\displaystyle=</td>
<td>PB−1DIPB=b2−1w−1b1−1DIb1wb2=b2−1w−1BIwb2,superscriptsubscript𝑃𝐵1subscript𝐷𝐼subscript𝑃𝐵superscriptsubscript𝑏21superscript𝑤1superscriptsubscript𝑏11subscript𝐷𝐼subscript𝑏1𝑤subscript𝑏2superscriptsubscript𝑏21superscript𝑤1subscript𝐵𝐼𝑤subscript𝑏2\displaystyle P_{B}^{-1}D_{I}P_{B}=b_{2}^{-1}w^{-1}b_{1}^{-1}D_{I}b_{1}wb_{2}=b_{2}^{-1}w^{-1}B_{I}wb_{2},</td>
<td></td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td>RIsubscript𝑅𝐼\displaystyle R_{I}</td>
<td>∈\displaystyle\in</td>
<td>D,D a subgroup of CG(a1)𝐷𝐷 a subgroup of subscript𝐶𝐺subscript𝑎1\displaystyle D,D\text{ a subgroup of }C_{G}(a_{1})</td>
<td></td>
</tr>
</tbody>
</table>
Following the notation in \[[6](#bib.bib6)\] let the elements transmitted by Alice and Bob be invertible then
<table>
<tbody>
<tr>
<td></td>
<td>CE1(EI)𝐶subscript𝐸1subscript𝐸𝐼\displaystyle CE_{1}(E_{I})</td>
<td>=\displaystyle=</td>
<td>b1−1w−1a1−1EIa1wb1=b1w−1EIwb1,EI∈Bformulae-sequencesuperscriptsubscript𝑏11superscript𝑤1superscriptsubscript𝑎11subscript𝐸𝐼subscript𝑎1𝑤subscript𝑏1subscript𝑏1superscript𝑤1subscript𝐸𝐼𝑤subscript𝑏1subscript𝐸𝐼𝐵\displaystyle b_{1}^{-1}w^{-1}a_{1}^{-1}E_{I}a_{1}wb_{1}=b_{1}w^{-1}E_{I}wb_{1},E_{I}\in B</td>
<td></td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td>CE2(FI)𝐶subscript𝐸2subscript𝐹𝐼\displaystyle CE_{2}(F_{I})</td>
<td>=\displaystyle=</td>
<td>b2wa2FIa2−1w−1b2−1=b1wFIw−1b1−1,BI∈Aformulae-sequencesubscript𝑏2𝑤subscript𝑎2subscript𝐹𝐼superscriptsubscript𝑎21superscript𝑤1superscriptsubscript𝑏21subscript𝑏1𝑤subscript𝐹𝐼superscript𝑤1superscriptsubscript𝑏11subscript𝐵𝐼𝐴\displaystyle b_{2}wa_{2}F_{I}a_{2}^{-1}w^{-1}b_{2}^{-1}=b_{1}wF_{I}w^{-1}b_{1}^{-1},B_{I}\in A</td>
<td></td>
</tr>
</tbody>
</table>
and so we can use this equation in a length attack. It may be asked how much better is a length attack using the conjugacy equations above for \[[7](#bib.bib7)\],\[[6](#bib.bib6)\] (and related protocols) compared to the known length attacks on \[[7](#bib.bib7)\],\[[6](#bib.bib6)\], and if the above equations can be used to improve the known length attacks (for example we may try using one and/or both equations above they can be used to decide peeling off the correct generator in combination with the algorithm that decides to peel of generators in an existing attack, an example would be if the above existing algorithm is unable (i.e. pick at random) to decide which is the correct generator to peel then peeling from $`C\hspace{0pt}E_{1}`$ and /or $`C\hspace{0pt}E_{2}`$ may be used to decide the correct generator). Following the notation in \[[11](#bib.bib11)\] we have
<table>
<tbody>
<tr>
<td></td>
<td>CE1(GI,c)=cGIc−1=arbsGIb−sa−r=arbαa−r,GI=bα, for some α chosen by attackerformulae-sequence𝐶subscript𝐸1subscript𝐺𝐼𝑐𝑐subscript𝐺𝐼superscript𝑐1superscript𝑎𝑟superscript𝑏𝑠subscript𝐺𝐼superscript𝑏𝑠superscript𝑎𝑟superscript𝑎𝑟superscript𝑏𝛼superscript𝑎𝑟subscript𝐺𝐼superscript𝑏𝛼, for some 𝛼 chosen by attackerCE_{1}(G_{I},c)=cG_{I}c^{-1}=a^{r}b^{s}G_{I}b^{-s}a^{-r}=a^{r}b^{\alpha}a^{-r},G_{I}=b^{\alpha}\text{, for some }\alpha\text{ chosen by attacker}</td>
<td></td>
</tr>
</tbody>
</table>
or using the suggestion of using the element $`e`$ in \[[11](#bib.bib11)\] we have
<table>
<tbody>
<tr>
<td></td>
<td>CE2(HI,c)𝐶subscript𝐸2subscript𝐻𝐼𝑐\displaystyle CE_{2}(H_{I},c)</td>
<td>=\displaystyle=</td>
<td>cGIc−1=arebsHIb−se−1a−r=ar(ebαe−1)a−r,HI=bα,formulae-sequence𝑐subscript𝐺𝐼superscript𝑐1superscript𝑎𝑟𝑒superscript𝑏𝑠subscript𝐻𝐼superscript𝑏𝑠superscript𝑒1superscript𝑎𝑟superscript𝑎𝑟𝑒superscript𝑏𝛼superscript𝑒1superscript𝑎𝑟subscript𝐻𝐼superscript𝑏𝛼,\displaystyle cG_{I}c^{-1}=a^{r}eb^{s}H_{I}b^{-s}e^{-1}a^{-r}=a^{r}(eb^{\alpha}e^{-1})a^{-r},H_{I}=b^{\alpha}\text{,}</td>
<td></td>
</tr>
<tr>
<td></td>
<td></td>
<td></td>
<td>for some α chosen by attacker.for some 𝛼 chosen by attacker.\displaystyle\text{for some }\alpha\text{ chosen by attacker.}</td>
<td></td>
</tr>
</tbody>
</table>
Following the notation in \[[2](#bib.bib2)\] we have
<table>
<tbody>
<tr>
<td></td>
<td>CE1(KI,y1)𝐶subscript𝐸1subscript𝐾𝐼subscript𝑦1\displaystyle CE_{1}(K_{I},y_{1})</td>
<td>=\displaystyle=</td>
<td>y1KIy1−1=axa−1KIax−1a−1=axKIx−1a−1,KI∈RBrformulae-sequencesubscript𝑦1subscript𝐾𝐼superscriptsubscript𝑦11𝑎𝑥superscript𝑎1subscript𝐾𝐼𝑎superscript𝑥1superscript𝑎1𝑎𝑥subscript𝐾𝐼superscript𝑥1superscript𝑎1subscript𝐾𝐼𝑅subscript𝐵𝑟\displaystyle y_{1}K_{I}y_{1}^{-1}=axa^{-1}K_{I}ax^{-1}a^{-1}=axK_{I}x^{-1}a^{-1},K_{I}\in RB_{r}</td>
<td></td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td>CE2(LI,y2)𝐶subscript𝐸2subscript𝐿𝐼subscript𝑦2\displaystyle CE_{2}(L_{I},y_{2})</td>
<td>=\displaystyle=</td>
<td>y2LIy2−1=bxb−1LIbx−1b−1=bxLIx−1b−1,LI∈LBlformulae-sequencesubscript𝑦2subscript𝐿𝐼superscriptsubscript𝑦21𝑏𝑥superscript𝑏1subscript𝐿𝐼𝑏superscript𝑥1superscript𝑏1𝑏𝑥subscript𝐿𝐼superscript𝑥1superscript𝑏1subscript𝐿𝐼𝐿subscript𝐵𝑙\displaystyle y_{2}L_{I}y_{2}^{-1}=bxb^{-1}L_{I}bx^{-1}b^{-1}=bxL_{I}x^{-1}b^{-1},L_{I}\in LB_{l}</td>
<td></td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td>CE3(MI,y1)𝐶subscript𝐸3subscript𝑀𝐼subscript𝑦1\displaystyle CE_{3}(M_{I},y_{1})</td>
<td>=\displaystyle=</td>
<td>y1−1MIy1=ax−1a−1MIaxa−1=ax−1MIxa−1,KI∈RBrformulae-sequencesuperscriptsubscript𝑦11subscript𝑀𝐼subscript𝑦1𝑎superscript𝑥1superscript𝑎1subscript𝑀𝐼𝑎𝑥superscript𝑎1𝑎superscript𝑥1subscript𝑀𝐼𝑥superscript𝑎1subscript𝐾𝐼𝑅subscript𝐵𝑟\displaystyle y_{1}^{-1}M_{I}y_{1}=ax^{-1}a^{-1}M_{I}axa^{-1}=ax^{-1}M_{I}xa^{-1},K_{I}\in RB_{r}</td>
<td></td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td>CE4(NI,y2)𝐶subscript𝐸4subscript𝑁𝐼subscript𝑦2\displaystyle CE_{4}(N_{I},y_{2})</td>
<td>=\displaystyle=</td>
<td>y2−1LIy2=bx−1b−1NIbxb−1=bx−1NIxb−1,LI∈LBlformulae-sequencesuperscriptsubscript𝑦21subscript𝐿𝐼subscript𝑦2𝑏superscript𝑥1superscript𝑏1subscript𝑁𝐼𝑏𝑥superscript𝑏1𝑏superscript𝑥1subscript𝑁𝐼𝑥superscript𝑏1subscript𝐿𝐼𝐿subscript𝐵𝑙\displaystyle y_{2}^{-1}L_{I}y_{2}=bx^{-1}b^{-1}N_{I}bxb^{-1}=bx^{-1}N_{I}xb^{-1},L_{I}\in LB_{l}</td>
<td></td>
</tr>
</tbody>
</table>
Following the notation in \[[13](#bib.bib13)\] we have
<table>
<tbody>
<tr>
<td></td>
<td></td>
<td></td>
<td>CE1(TI,wi)=wiTIwi−1=yi−1viyi−1KIyivi−1yi−1−1=yi−1viKIvi−1yi−1−1𝐶subscript𝐸1subscript𝑇𝐼subscript𝑤𝑖subscript𝑤𝑖subscript𝑇𝐼superscriptsubscript𝑤𝑖1subscript𝑦𝑖1subscript𝑣𝑖superscriptsubscript𝑦𝑖1subscript𝐾𝐼subscript𝑦𝑖superscriptsubscript𝑣𝑖1superscriptsubscript𝑦𝑖11subscript𝑦𝑖1subscript𝑣𝑖subscript𝐾𝐼superscriptsubscript𝑣𝑖1superscriptsubscript𝑦𝑖11\displaystyle CE_{1}(T_{I},w_{i})=w_{i}T_{I}w_{i}^{-1}=y_{i-1}v_{i}y_{i}^{-1}K_{I}y_{i}v_{i}^{-1}y_{i-1}^{-1}=y_{i-1}v_{i}K_{I}v_{i}^{-1}y_{i-1}^{-1}</td>
<td></td>
</tr>
<tr>
<td></td>
<td></td>
<td></td>
<td>CE1(TI,wi−1)=wi−1TIwi=yivi−1yi−1−1KIyi−1viyi−1=yivi−1KIviyi−1𝐶subscript𝐸1subscript𝑇𝐼superscriptsubscript𝑤𝑖1superscriptsubscript𝑤𝑖1subscript𝑇𝐼subscript𝑤𝑖subscript𝑦𝑖superscriptsubscript𝑣𝑖1superscriptsubscript𝑦𝑖11subscript𝐾𝐼subscript𝑦𝑖1subscript𝑣𝑖superscriptsubscript𝑦𝑖1subscript𝑦𝑖superscriptsubscript𝑣𝑖1subscript𝐾𝐼subscript𝑣𝑖superscriptsubscript𝑦𝑖1\displaystyle CE_{1}(T_{I},w_{i}^{-1})=w_{i}^{-1}T_{I}w_{i}=y_{i}v_{i}^{-1}y_{i-1}^{-1}K_{I}y_{i-1}v_{i}y_{i}^{-1}=y_{i}v_{i}^{-1}K_{I}v_{i}y_{i}^{-1}</td>
<td></td>
</tr>
<tr>
<td></td>
<td></td>
<td></td>
<td>KI is chosen from the subgroup that generates the elements xjsubscript𝐾𝐼 is chosen from the subgroup that generates the elements subscript𝑥𝑗\displaystyle K_{I}\text{ is chosen from the subgroup that generates the elements }x_{j}</td>
<td></td>
</tr>
<tr>
<td></td>
<td></td>
<td></td>
<td>UI is chosen from the subgroup that generates the elements yJsubscript𝑈𝐼 is chosen from the subgroup that generates the elements subscript𝑦𝐽\displaystyle U_{I}\text{ is chosen from the subgroup that generates the elements }y_{J}</td>
<td></td>
</tr>
<tr>
<td></td>
<td></td>
<td></td>
<td>So the secrets y1,…,yk can be recovered and hence vi=yi−1−1wiyiSo the secrets subscript𝑦1…subscript𝑦𝑘 can be recovered and hence subscript𝑣𝑖superscriptsubscript𝑦𝑖11subscript𝑤𝑖subscript𝑦𝑖\displaystyle\text{So the secrets }y_{1},...,y_{k}\text{ can be recovered and hence }v_{i}=y_{i-1}^{-1}w_{i}y_{i}</td>
<td></td>
</tr>
<tr>
<td></td>
<td></td>
<td></td>
<td>CE2(UI,w)=wUIw−1=x0v1x1v2…vkxkUIxk−1vk−1…v2−1x1−1v1−1x0−1𝐶subscript𝐸2subscript𝑈𝐼𝑤𝑤subscript𝑈𝐼superscript𝑤1subscript𝑥0subscript𝑣1subscript𝑥1subscript𝑣2…subscript𝑣𝑘subscript𝑥𝑘subscript𝑈𝐼superscriptsubscript𝑥𝑘1superscriptsubscript𝑣𝑘1…superscriptsubscript𝑣21superscriptsubscript𝑥11superscriptsubscript𝑣11superscriptsubscript𝑥01\displaystyle CE_{2}(U_{I},w)=wU_{I}w^{-1}=x_{0}v_{1}x_{1}v_{2}...v_{k}x_{k}U_{I}x_{k}^{-1}v_{k}^{-1}...v_{2}^{-1}x_{1}^{-1}v_{1}^{-1}x_{0}^{-1}</td>
<td></td>
</tr>
<tr>
<td></td>
<td></td>
<td>=\displaystyle=</td>
<td>x0v1x1v2…vkUIvk−1…v2−1x1−1v1−1x0−1subscript𝑥0subscript𝑣1subscript𝑥1subscript𝑣2…subscript𝑣𝑘subscript𝑈𝐼superscriptsubscript𝑣𝑘1…superscriptsubscript𝑣21superscriptsubscript𝑥11superscriptsubscript𝑣11superscriptsubscript𝑥01\displaystyle x_{0}v_{1}x_{1}v_{2}...v_{k}U_{I}v_{k}^{-1}...v_{2}^{-1}x_{1}^{-1}v_{1}^{-1}x_{0}^{-1}</td>
<td></td>
</tr>
<tr>
<td></td>
<td></td>
<td></td>
<td>recovering x0v1x1v2…vk gives xk=(x0v1x1v2…vk)−1w similarlyrecovering subscript𝑥0subscript𝑣1subscript𝑥1subscript𝑣2…subscript𝑣𝑘 gives subscript𝑥𝑘superscriptsubscript𝑥0subscript𝑣1subscript𝑥1subscript𝑣2…subscript𝑣𝑘1𝑤 similarly\displaystyle\text{recovering }x_{0}v_{1}x_{1}v_{2}...v_{k}\text{ gives }x_{k}=(x_{0}v_{1}x_{1}v_{2}...v_{k})^{-1}w\text{ similarly}</td>
<td></td>
</tr>
<tr>
<td></td>
<td></td>
<td></td>
<td>CE2(UI,w−1)=w−1UIw gives x0. Because all vi can be recovered𝐶subscript𝐸2subscript𝑈𝐼superscript𝑤1superscript𝑤1subscript𝑈𝐼𝑤 gives subscript𝑥0. Because all subscript𝑣𝑖 can be recovered\displaystyle CE_{2}(U_{I},w^{-1})=w^{-1}U_{I}w\text{ gives }x_{0}\text{. Because all }v_{i}\text{ can be recovered}</td>
<td></td>
</tr>
<tr>
<td></td>
<td></td>
<td></td>
<td>so similarly repeating the attack above using</td>
<td></td>
</tr>
<tr>
<td></td>
<td></td>
<td></td>
<td>CE2(UI,(x0v1)−1wxk−1vk−1) for x1,xk−1 and similarly by repeating𝐶subscript𝐸2subscript𝑈𝐼superscriptsubscript𝑥0subscript𝑣11𝑤superscriptsubscript𝑥𝑘1superscriptsubscript𝑣𝑘1 for subscript𝑥1,subscript𝑥𝑘1 and similarly by repeating\displaystyle CE_{2}(U_{I},(x_{0}v_{1})^{-1}wx_{k}^{-1}v_{k}^{-1})\text{ for }x_{1}\text{,}x_{k-1}\text{ and similarly by repeating}</td>
<td></td>
</tr>
<tr>
<td></td>
<td></td>
<td></td>
<td>again all the xi can be recovered.again all the subscript𝑥𝑖 can be recovered.\displaystyle\text{again all the }x_{i}\text{ can be recovered. }</td>
<td></td>
</tr>
</tbody>
</table>
Following the notation in \[[10](#bib.bib10)\] we have
<table>
<tbody>
<tr>
<td></td>
<td>CE1(PI,c1)𝐶subscript𝐸1subscript𝑃𝐼subscript𝑐1\displaystyle CE_{1}(P_{I},c_{1})</td>
<td>=\displaystyle=</td>
<td>c1PIc1−1=a1xa2PIa2−1x−1a1−1=a1xPIx−1a2−1,PI∈UBrformulae-sequencesubscript𝑐1subscript𝑃𝐼superscriptsubscript𝑐11subscript𝑎1𝑥subscript𝑎2subscript𝑃𝐼superscriptsubscript𝑎21superscript𝑥1superscriptsubscript𝑎11subscript𝑎1𝑥subscript𝑃𝐼superscript𝑥1superscriptsubscript𝑎21subscript𝑃𝐼𝑈subscript𝐵𝑟\displaystyle c_{1}P_{I}c_{1}^{-1}=a_{1}xa_{2}P_{I}a_{2}^{-1}x^{-1}a_{1}^{-1}=a_{1}xP_{I}x^{-1}a_{2}^{-1},P_{I}\in UB_{r}</td>
<td></td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td>CE1(QI,y1)𝐶subscript𝐸1subscript𝑄𝐼subscript𝑦1\displaystyle CE_{1}(Q_{I},y_{1})</td>
<td>=\displaystyle=</td>
<td>c2QIc2−1=b1xb2QIb2−1x−1b1−1=b1xQIx−1b1−1,KI∈LBrformulae-sequencesubscript𝑐2subscript𝑄𝐼superscriptsubscript𝑐21subscript𝑏1𝑥subscript𝑏2subscript𝑄𝐼superscriptsubscript𝑏21superscript𝑥1superscriptsubscript𝑏11subscript𝑏1𝑥subscript𝑄𝐼superscript𝑥1superscriptsubscript𝑏11subscript𝐾𝐼𝐿subscript𝐵𝑟\displaystyle c_{2}Q_{I}c_{2}^{-1}=b_{1}xb_{2}Q_{I}b_{2}^{-1}x^{-1}b_{1}^{-1}=b_{1}xQ_{I}x^{-1}b_{1}^{-1},K_{I}\in LB_{r}</td>
<td></td>
</tr>
</tbody>
</table>
The CKLHC protocol in \[[10](#bib.bib10)\] was introduced as an improvement of the KLCHKP protocol which it is a generalisation/modification of but we have shown the CKLHC does not improve the security of the KLCHKP protocol as they are both based on the MSCSP. This means using the KLCHKP and CKLHC cryptosystems is no more secure than using the AAG (Anshel-Anshel-Goldfeld) scheme \[[1](#bib.bib1)\] in the connection that they can all be broken using by solving the MSCSP. Hence this means there is no need to use the CKLHC cryptosystem any longer. The theorem implies the Turing reduction of the DH-DP to the MSCSP (MSCSP $`\propto_{T}`$ DH-DP) for the case when the above DH-DP have related solutions, clearly a conjugacy extractor can be feasibly computed- that is in polynomial time and polynomial space (for parameters used in the CKLHC cryptosystem) using a finite number of group operations. If the conjugacy extractor is not computable in polynomial time and polynomial space in the connection of breaking a cryptographic protocol then the above protocol may be secure from an attack by solving the MSCSP, we may consider a generalisation of the MSCSP in the above cryptographic protocol where $`G`$ is a semigroup instead of a group. Note if we have one $`C\hspace{0pt}E`$ then we exactly have the Turing reduction of the DP to the MSCSP (MSCSP $`\propto_{T}`$DP) , and hence the Turing reduction of the DH-DP to the MSCSP (MSCSP $`\propto_{T}`$ DH-DP).
In \[[14](#bib.bib14)\] an authentication scheme is given based on the problem of shifted conjugacy search problem (SCSP). It is not stated in \[[14](#bib.bib14)\] not to select $`r`$ (the random value used in the commitment $`x = {r \ast p} = {{{{r \cdot d}\hspace{0pt}p} \cdot \sigma_{1} \cdot d}\hspace{0pt}r^{- 1}}`$) from a publicly known subgroup. Then an attack is as follows.
1\. Suppose $`r \in R`$ where $`R = {\{\alpha_{1},\ldots,\alpha_{k}\}}`$ is a publicly known subgroup of $`B_{n}`$. In this step it is required the attacker just needs to find one element that commutes with $`r`$ and not al least with $`{d\hspace{0pt}p} \cdot \sigma_{1}`$ (using a chosen algorithm by the attacker) to show that SCSP can be reduced to solving the CSP. The attacker picks a subgroup of $`R`$ given by the generators $`g_{1},\ldots,g_{k}`$. Then the attacker computes all of or a large part of
<table>
<tbody>
<tr>
<td></td>
<td>N=C(α1,…,αk)=C(α1)∩…C(αk).𝑁𝐶subscript𝛼1…subscript𝛼𝑘𝐶subscript𝛼1…𝐶subscript𝛼𝑘.N=C(\alpha_{1},...,\alpha_{k})=C(\alpha_{1})\cap...C(\alpha_{k})\text{.}</td>
<td></td>
</tr>
</tbody>
</table>
2\. Then
<table>
<tbody>
<tr>
<td></td>
<td>CE1(NI,r∗p)=(r⋅dp⋅σ1⋅dr−1)−1NI(r⋅dp⋅σ1⋅dr−1)=𝐶subscript𝐸1subscript𝑁𝐼∗𝑟𝑝superscript⋅⋅𝑟𝑑𝑝subscript𝜎1𝑑superscript𝑟11subscript𝑁𝐼⋅⋅𝑟𝑑𝑝subscript𝜎1𝑑superscript𝑟1absent\displaystyle CE_{1}(N_{I},r\ast p)=(r\cdot dp\cdot\sigma_{1}\cdot dr^{-1})^{-1}N_{I}(r\cdot dp\cdot\sigma_{1}\cdot dr^{-1})=</td>
<td></td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td>dr⋅σ1−1⋅dp−1⋅r−1⋅NI⋅(r⋅dp⋅σ1⋅dr−1)=dr⋅σ1−1⋅dp−1NIdp⋅σ1⋅dr−1⋅⋅𝑑𝑟superscriptsubscript𝜎11𝑑superscript𝑝1superscript𝑟1subscript𝑁𝐼⋅⋅𝑟𝑑𝑝subscript𝜎1𝑑superscript𝑟1⋅⋅𝑑𝑟superscriptsubscript𝜎11𝑑superscript𝑝1subscript𝑁𝐼𝑑𝑝subscript𝜎1𝑑superscript𝑟1\displaystyle dr\cdot\sigma_{1}^{-1}\cdot dp^{-1}\cdot r^{-1}\cdot N_{I}\cdot(r\cdot dp\cdot\sigma_{1}\cdot dr^{-1})=dr\cdot\sigma_{1}^{-1}\cdot dp^{-1}N_{I}dp\cdot\sigma_{1}\cdot dr^{-1}</td>
<td></td>
</tr>
</tbody>
</table>
will be true.$`N_{I} \in N`$,$`1 \leq I \leq M`$. The the protocol can be based on the MSCSP with
<table>
<tbody>
<tr>
<td></td>
<td>((N1,N2,…,NM),(CE(N1,r∗p),CE(N2,r∗p),…,CE(NM,r∗p)))subscript𝑁1subscript𝑁2…subscript𝑁𝑀𝐶𝐸subscript𝑁1∗𝑟𝑝𝐶𝐸subscript𝑁2∗𝑟𝑝…𝐶𝐸subscript𝑁𝑀∗𝑟𝑝\displaystyle((N_{1},N_{2},...,N_{M}),(CE(N_{1},r\ast p),CE(N_{2},r\ast p),...,CE(N_{M},r\ast p)))\text{ }</td>
<td></td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td>with solution (dr⋅σ1−1⋅dp−1,dp⋅σ1⋅dr−1), O=dp⋅σ1⋅dr−1.with solution ⋅𝑑𝑟superscriptsubscript𝜎11𝑑superscript𝑝1⋅𝑑𝑝subscript𝜎1𝑑superscript𝑟1, 𝑂⋅𝑑𝑝subscript𝜎1𝑑superscript𝑟1\displaystyle\text{with solution }(dr\cdot\sigma_{1}^{-1}\cdot dp^{-1},dp\cdot\sigma_{1}\cdot dr^{-1})\text{, }O=dp\cdot\sigma_{1}\cdot dr^{-1}.</td>
<td></td>
</tr>
</tbody>
</table>
and $`r`$ can be found by computing $`{({{\sigma_{1}^{- 1} \cdot d}\hspace{0pt}p^{- 1}\hspace{0pt}O})}^{- 1} = r`$, there is a similar attack if $`s`$ (Alice’s private key) is chosen from a subgroup that is publicly known. Note the similar attack with $`N_{I}`$ commuting $`{r \cdot d}\hspace{0pt}p`$ would mean the SCSP is just the CSP (no extra computation using $`d`$ is required).
As a variant of the above algorithm an attacker may try to compute an element $`N_{I}^{'} \in {C\hspace{0pt}{(L)}}`$ then it may be possible to use $`N_{I}^{'}`$ instead of $`N_{I}`$ in the attack above where $`L = {r \ast p}`$ or $`L = {r \ast p^{'}}`$, so in this variant knowledge of $`s`$ $`(L = p^{'}`$ here) or $`r`$ being chosen from a subgroup is not required. A different second CE on the authentication scheme in \[[14](#bib.bib14)\] or the SCSP, is suppose in a general case we have a pair of examples of the SCSP that have the same secret element $`{x = {r \ast p}},{x^{'} = {r \ast p^{'}}}`$ (the notation here follows \[[14](#bib.bib14)\] with the secret element $`r`$) then
<table>
<tbody>
<tr>
<td></td>
<td>CE2(x,x−′1)\displaystyle CE_{2}(x,x^{{}^{\prime}-1})</td>
<td>=\displaystyle=</td>
<td></td>
<td></td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td>CE2(r∗p,r∗p′)𝐶subscript𝐸2∗𝑟𝑝∗𝑟superscript𝑝′\displaystyle CE_{2}(r\ast p,r\ast p^{{}^{\prime}})</td>
<td>=\displaystyle=</td>
<td>x⋅x−′1\displaystyle x\cdot x^{{}^{\prime}-1}</td>
<td></td>
</tr>
<tr>
<td></td>
<td></td>
<td>=\displaystyle=</td>
<td>(r⋅dp⋅σ1⋅dr−1)⋅(r⋅dp′⋅σ1⋅dr−1)−1⋅⋅⋅𝑟𝑑𝑝subscript𝜎1𝑑superscript𝑟1superscript⋅⋅𝑟𝑑superscript𝑝′subscript𝜎1𝑑superscript𝑟11\displaystyle(r\cdot dp\cdot\sigma_{1}\cdot dr^{-1})\cdot(r\cdot dp^{{}^{\prime}}\cdot\sigma_{1}\cdot dr^{-1})^{-1}</td>
<td></td>
</tr>
<tr>
<td></td>
<td></td>
<td>=\displaystyle=</td>
<td>(r⋅dp⋅σ1⋅dr−1)⋅dr⋅σ1−1⋅dp−′1⋅r−1\displaystyle(r\cdot dp\cdot\sigma_{1}\cdot dr^{-1})\cdot dr\cdot\sigma_{1}^{-1}\cdot dp^{{}^{\prime}-1}\cdot r^{-1}</td>
<td></td>
</tr>
<tr>
<td></td>
<td></td>
<td>=\displaystyle=</td>
<td>r⋅dp⋅dp−′1⋅r−1\displaystyle r\cdot dp\cdot dp^{{}^{\prime}-1}\cdot r^{-1}</td>
<td></td>
</tr>
</tbody>
</table>
so the secret $`r`$ can be found by solving the CSP pair ($`{{{d\hspace{0pt}p} \cdot d}\hspace{0pt}p^{{{}_{}' -}1}},{C\hspace{0pt}E\hspace{0pt}{({r \ast p},{r \ast p^{'}})}}`$) for $`r`$, there is a similar CE for $`d\hspace{0pt}r`$ (use $`x^{- 1} \cdot x^{'}`$ instead of $`x \cdot x^{{{}_{}' -}1}`$ and then $`d\hspace{0pt}r`$ is transformed to $`r`$ etc). Then the attack waits until $`b = 1`$ is so in this case Alice send to Bob $`{r \ast s} = {{{{r \cdot d}\hspace{0pt}s} \cdot \sigma_{1} \cdot d}\hspace{0pt}r^{- 1}}`$ hence the attacker computes the private key $`s = {{{r^{- 1} \cdot {({r \ast s})} \cdot d}\hspace{0pt}r} \cdot \sigma_{1}^{- 1}}`$. We note the above attacks is can be used to answer question 2.6 in \[[14](#bib.bib14)\] (with $`y = p`$ in th CSP). We note our attack can be used to solve the shifted conjugacy decision problem. Our results suggest that CE functions should be hard to compute hence semigroups may be considered because then the theorem 1 may be false because the elements $`y_{a}`$ and/or $`y_{b}`$ are not invertible. This suggestion applies to any hard problem such as the EDL problem below. Note in the algorithms in CE computations it may be centraliser element(s) (call these $`p_{i}`$) that multiply the secret(s) cancel out and it can be shown these factors in the centraliser are efficiently computable, for example one way to do this is if $`p_{i}`$ is a power of the fundamental braid then we can estimate a power of the fundamental braid from the public elements (for example using a length function) and so recover a $`p_{i}`$, or instead find this power by brute force.
We now consider a problem related to the SCSP/CSP which is given a semigroup $`G`$, and publicly known functions $`u:{G\rightarrow{G_{1},v}}:`$ $`{G_{2}\rightarrow{G,w}}:{G_{3}\rightarrow G}`$, $`G_{1},G_{2},G_{3}`$ are subsets of $`G`$, and publicly the publicly known elements $`{y_{i} = {u\hspace{0pt}{(r)}\hspace{0pt}v\hspace{0pt}{(p_{i})}\hspace{0pt}w\hspace{0pt}{(r^{- 1})}}},`$ $`{v\hspace{0pt}{(p_{i})}},`$ $`1 \leq i \leq n`$ find if the element $`r`$. We observe that the problem generalises the twisted conjugacy problem \[[15](#bib.bib15)\] and the doubly twisted conjugacy problem \[[15](#bib.bib15)\], e.g. with $`i = {1,u}`$ an endomorphism, $`v,w`$ the identity map we recover the twisted conjugacy problem, $`G`$ a group, $`{G_{1} = G},{{G_{2} = G},{G_{3} = G}}`$; so we refer to the above problem as the GTCP (generalised twisted conjugacy problem) which we now describe solutions for. Now consider the GTCP with $`i > 1`$, select a pair $`i,j`$ with $`{v\hspace{0pt}{(p_{i})}} \neq {v\hspace{0pt}{(p_{j})}}`$,$`{1 \leq i},{j \leq n}`$ we have the conjugacy extractor
<table>
<tbody>
<tr>
<td></td>
<td>CE1(yi,yj)𝐶subscript𝐸1subscript𝑦𝑖subscript𝑦𝑗\displaystyle CE_{1}(y_{i},y_{j})</td>
<td>=\displaystyle=</td>
<td>yi⋅yj−1⋅subscript𝑦𝑖superscriptsubscript𝑦𝑗1\displaystyle y_{i}\cdot y_{j}^{-1}</td>
<td></td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td>CE1(u(r)v(pi)w(r−1),u(r)v(pj)w(r−1))𝐶subscript𝐸1𝑢𝑟𝑣subscript𝑝𝑖𝑤superscript𝑟1𝑢𝑟𝑣subscript𝑝𝑗𝑤superscript𝑟1\displaystyle CE_{1}(u(r)v(p_{i})w(r^{-1}),u(r)v(p_{j})w(r^{-1}))</td>
<td>=\displaystyle=</td>
<td>u(r)v(pi)w(r−1)⋅(u(r)v(pj)w(r−1))−1⋅𝑢𝑟𝑣subscript𝑝𝑖𝑤superscript𝑟1superscript𝑢𝑟𝑣subscript𝑝𝑗𝑤superscript𝑟11\displaystyle u(r)v(p_{i})w(r^{-1})\cdot(u(r)v(p_{j})w(r^{-1}))^{-1}</td>
<td></td>
</tr>
<tr>
<td></td>
<td></td>
<td>=\displaystyle=</td>
<td>u(r)v(pi)w(r−1)⋅w(r−1)−1v(pj)−1u(r)−1⋅𝑢𝑟𝑣subscript𝑝𝑖𝑤superscript𝑟1𝑤superscriptsuperscript𝑟11𝑣superscriptsubscript𝑝𝑗1𝑢superscript𝑟1\displaystyle u(r)v(p_{i})w(r^{-1})\cdot w(r^{-1})^{-1}v(p_{j})^{-1}u(r)^{-1}</td>
<td></td>
</tr>
<tr>
<td></td>
<td></td>
<td>=\displaystyle=</td>
<td>u(r)v(pi)v(pj)−1u(r)−1𝑢𝑟𝑣subscript𝑝𝑖𝑣superscriptsubscript𝑝𝑗1𝑢superscript𝑟1\displaystyle u(r)v(p_{i})v(p_{j})^{-1}u(r)^{-1}</td>
<td></td>
</tr>
</tbody>
</table>
so here we solve the CSP pair $`(v{(p_{i})}v{(p_{j})}^{- 1}`$,$`CE_{1}{(y_{i},y_{j})}).`$ Once $`u\hspace{0pt}{(r)}`$ is obtained (attempt to) use the inverse of $`u`$ to get $`r`$.
<table>
<tbody>
<tr>
<td></td>
<td>CE2(yi,yj)𝐶subscript𝐸2subscript𝑦𝑖subscript𝑦𝑗\displaystyle CE_{2}(y_{i},y_{j})</td>
<td>=\displaystyle=</td>
<td>yj−1⋅yi⋅superscriptsubscript𝑦𝑗1subscript𝑦𝑖\displaystyle y_{j}^{-1}\cdot y_{i}</td>
<td></td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td>CE2(u(r)v(pi)w(r−1),u(r)v(pj)w(r−1))𝐶subscript𝐸2𝑢𝑟𝑣subscript𝑝𝑖𝑤superscript𝑟1𝑢𝑟𝑣subscript𝑝𝑗𝑤superscript𝑟1\displaystyle CE_{2}(u(r)v(p_{i})w(r^{-1}),u(r)v(p_{j})w(r^{-1}))</td>
<td>=\displaystyle=</td>
<td>(u(r)v(pj)w(r−1))−1⋅u(r)v(pi)w(r−1)⋅superscript𝑢𝑟𝑣subscript𝑝𝑗𝑤superscript𝑟11𝑢𝑟𝑣subscript𝑝𝑖𝑤superscript𝑟1\displaystyle(u(r)v(p_{j})w(r^{-1}))^{-1}\cdot u(r)v(p_{i})w(r^{-1})</td>
<td></td>
</tr>
<tr>
<td></td>
<td></td>
<td>=\displaystyle=</td>
<td>w(r−1)−1v(pj)−1u(r)−1⋅u(r)v(pi)w(r−1)⋅𝑤superscriptsuperscript𝑟11𝑣superscriptsubscript𝑝𝑗1𝑢superscript𝑟1𝑢𝑟𝑣subscript𝑝𝑖𝑤superscript𝑟1\displaystyle w(r^{-1})^{-1}v(p_{j})^{-1}u(r)^{-1}\cdot u(r)v(p_{i})w(r^{-1})</td>
<td></td>
</tr>
<tr>
<td></td>
<td></td>
<td>=\displaystyle=</td>
<td>w(r−1)−1v(pj)−1v(pi)w(r−1)𝑤superscriptsuperscript𝑟11𝑣superscriptsubscript𝑝𝑗1𝑣subscript𝑝𝑖𝑤superscript𝑟1\displaystyle w(r^{-1})^{-1}v(p_{j})^{-1}v(p_{i})w(r^{-1})</td>
<td></td>
</tr>
</tbody>
</table>
so here we solve the CSP pair $`(v{(p_{j})}^{- 1}v{(p_{i})}`$,$`CE_{2}{(y_{i},y_{j})}).`$ Once $`w\hspace{0pt}{(r)}`$ is obtained use the inverse of $`w`$ to get $`r`$. Note for one of the above $`C\hspace{0pt}E`$ functions for the twisted conjugacy problem the problem is just the conjugacy search problem. The above can be repeated for different $`i,j`$ to get a reduction to the MSCSP. Another method to solve the GTCP as follows.
1\. Suppose $`{G_{1} = {\{\alpha_{1},\ldots,\alpha_{k}\}}},{G_{3} = {\{\beta_{1},\ldots,\beta_{l}\}}}`$ are publicly known. In this step it is required that one element that commutes with $`r`$ is to be found. Pick subgroups of $`G_{1},G_{3}`$ given by the generators $`{g_{1},\ldots,g_{L},h_{1},\ldots,h_{L}}.`$ Then compute all of or a large part of
<table>
<tbody>
<tr>
<td></td>
<td>N𝑁\displaystyle N</td>
<td>=\displaystyle=</td>
<td>C(α1,…,αk)=C(α1)∩…C(αk).𝐶subscript𝛼1…subscript𝛼𝑘𝐶subscript𝛼1…𝐶subscript𝛼𝑘.\displaystyle C(\alpha_{1},...,\alpha_{k})=C(\alpha_{1})\cap...C(\alpha_{k})\text{.}</td>
<td></td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td>M𝑀\displaystyle M</td>
<td>=\displaystyle=</td>
<td>C(β1,…,βk)=C(β1)∩…C(βl).𝐶subscript𝛽1…subscript𝛽𝑘𝐶subscript𝛽1…𝐶subscript𝛽𝑙.\displaystyle C(\beta_{1},...,\beta_{k})=C(\beta_{1})\cap...C(\beta_{l})\text{.}</td>
<td></td>
</tr>
</tbody>
</table>
2\. Then
<table>
<tbody>
<tr>
<td></td>
<td>CE3(MI,yi)=u(r)v(pi)w(r−1)MI(u(r)v(pi)w(r−1))−1=𝐶subscript𝐸3subscript𝑀𝐼subscript𝑦𝑖𝑢𝑟𝑣subscript𝑝𝑖𝑤superscript𝑟1subscript𝑀𝐼superscript𝑢𝑟𝑣subscript𝑝𝑖𝑤superscript𝑟11absent\displaystyle CE_{3}(M_{I},y_{i})=u(r)v(p_{i})w(r^{-1})M_{I}(u(r)v(p_{i})w(r^{-1}))^{-1}=</td>
<td></td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td>u(r)v(pi)w(r−1)MIw(r−1)−1v(pi)−1u(r)−1=𝑢𝑟𝑣subscript𝑝𝑖𝑤superscript𝑟1subscript𝑀𝐼𝑤superscriptsuperscript𝑟11𝑣superscriptsubscript𝑝𝑖1𝑢superscript𝑟1absent\displaystyle u(r)v(p_{i})w(r^{-1})M_{I}w(r^{-1})^{-1}v(p_{i})^{-1}u(r)^{-1}=</td>
<td></td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td>u(r)v(pi)MIv(pi)−1u(r)−1𝑢𝑟𝑣subscript𝑝𝑖subscript𝑀𝐼𝑣superscriptsubscript𝑝𝑖1𝑢superscript𝑟1\displaystyle u(r)v(p_{i})M_{I}v(p_{i})^{-1}u(r)^{-1}</td>
<td></td>
</tr>
</tbody>
</table>
and
<table>
<tbody>
<tr>
<td></td>
<td>CE4(NI,yi)=(u(r)v(pi)w(r−1))−1NJu(r)v(pi)w(r−1)=𝐶subscript𝐸4subscript𝑁𝐼subscript𝑦𝑖superscript𝑢𝑟𝑣subscript𝑝𝑖𝑤superscript𝑟11subscript𝑁𝐽𝑢𝑟𝑣subscript𝑝𝑖𝑤superscript𝑟1absent\displaystyle CE_{4}(N_{I},y_{i})=(u(r)v(p_{i})w(r^{-1}))^{-1}N_{J}u(r)v(p_{i})w(r^{-1})=</td>
<td></td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td>w(r−1)−1v(pj)−1NJv(pi)w(r−1))\displaystyle w(r^{-1})^{-1}v(p_{j})^{-1}N_{J}v(p_{i})w(r^{-1}))</td>
<td></td>
</tr>
</tbody>
</table>
$`{{N_{J} \in N},{{M_{I} \in M},{1 \leq I \leq m}}},{1 \leq J \leq n}`$. Hence using $`{C\hspace{0pt}E_{3}\hspace{0pt}{(M_{I},y_{i})}},{C\hspace{0pt}E_{4}\hspace{0pt}{(N_{I},y_{i})}}`$: the GTCP has a solution respectively in the MSCSPs with
<table>
<tbody>
<tr>
<td></td>
<td>((N1,N2,…,Nn),(CE4(N1,yj1),CE4(N2,yj2),…,CE4(Nn,yjn)))subscript𝑁1subscript𝑁2…subscript𝑁𝑛𝐶subscript𝐸4subscript𝑁1subscript𝑦subscript𝑗1𝐶subscript𝐸4subscript𝑁2subscript𝑦subscript𝑗2…𝐶subscript𝐸4subscript𝑁𝑛subscript𝑦subscript𝑗𝑛\displaystyle((N_{1},N_{2},...,N_{n}),(CE_{4}(N_{1},y_{j_{1}}),CE_{4}(N_{2},y_{j_{2}}),...,CE_{4}(N_{n},y_{j_{n}})))\text{ }</td>
<td></td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td>with solution w(r−1)−1v(pj)−1;with solution 𝑤superscriptsuperscript𝑟11𝑣superscriptsubscript𝑝𝑗1;\displaystyle\text{with solution }w(r^{-1})^{-1}v(p_{j})^{-1}\text{;}</td>
<td></td>
</tr>
</tbody>
</table>
<table>
<tbody>
<tr>
<td></td>
<td>((M1,M2,…,Mm),(CE3(M1,yi1),CE3(M3,yi2),…,CE3(Mm,yim)))subscript𝑀1subscript𝑀2…subscript𝑀𝑚𝐶subscript𝐸3subscript𝑀1subscript𝑦subscript𝑖1𝐶subscript𝐸3subscript𝑀3subscript𝑦subscript𝑖2…𝐶subscript𝐸3subscript𝑀𝑚subscript𝑦subscript𝑖𝑚\displaystyle((M_{1},M_{2},...,M_{m}),(CE_{3}(M_{1},y_{i_{1}}),CE_{3}(M_{3},y_{i_{2}}),...,CE_{3}(M_{m},y_{i_{m}})))\text{ }</td>
<td></td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td>with solution u(r)v(pi);with solution 𝑢𝑟𝑣subscript𝑝𝑖;\displaystyle\text{with solution }u(r)v(p_{i})\text{;}</td>
<td></td>
</tr>
</tbody>
</table>
; from $`w\hspace{0pt}{(r^{- 1})}^{- 1}\hspace{0pt}v\hspace{0pt}{(p_{j})}^{- 1}`$, $`u\hspace{0pt}{(r)}\hspace{0pt}v\hspace{0pt}{(p_{i})}`$ we can obtain $`r`$ by using a right multiplication and using the inverses of $`w,u`$; this show the twisted conjugacy problem can be deterministically reduced to the MSCSP. We observe the algorithm in section 7 below can be used to attempt to solve the twisted or doubly twisted conjugacy problem with or without using $`u`$ or $`v`$. Observe once we have found a solutions to the twisted conjugacy problem, SCSP this means we can solve the decision version of the twisted conjugacy problem and SCSP.
6\. A Solution for the EDL type problem in Non-commutative Semigroups.
The EDL braid type problem was proposed in \[[9](#bib.bib9)\] where it is assumed to be hard. Following the notation in our definition above of the EDL we have the following theorem.
Theorem 2
Given (the DP equations) $`y_{a} = {u\hspace{0pt}a\hspace{0pt}v}`$, $`y_{b} = {w\hspace{0pt}b\hspace{0pt}x}`$ in the EDL is sometimes equivalent to solving the CSP if $`y_{a}`$,$`y_{b}`$ are both invertible elements.
Proof
We will solve this problem by solving a system of DP equations for certain values of the secrets so our proof can be used to solve a system of DP equations for example in showing above the shifted conjugacy based protocol is based on the CSP. Assume $`w = u`$ and $`x = v`$,$`a \neq b`$, then $`y_{a} = {u\hspace{0pt}a\hspace{0pt}v}`$, $`y_{b} = {u\hspace{0pt}b\hspace{0pt}v}`$, and so the $`C\hspace{0pt}E`$ functions give
<table>
<tbody>
<tr>
<td></td>
<td>yayb−1subscript𝑦𝑎superscriptsubscript𝑦𝑏1\displaystyle y_{a}y_{b}^{-1}</td>
<td>=\displaystyle=</td>
<td>uab−1u−1𝑢𝑎superscript𝑏1superscript𝑢1\displaystyle uab^{-1}u^{-1}</td>
<td></td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td>ya−1ybsuperscriptsubscript𝑦𝑎1subscript𝑦𝑏\displaystyle y_{a}^{-1}y_{b}</td>
<td>=\displaystyle=</td>
<td>v−1a−1bvsuperscript𝑣1superscript𝑎1𝑏𝑣\displaystyle v^{-1}a^{-1}bv</td>
<td></td>
</tr>
</tbody>
</table>
and so we can solve the CSP pairs for $`({a\hspace{0pt}b^{- 1}},{y_{a}\hspace{0pt}y_{b}^{- 1}})`$, and $`({a^{- 1}\hspace{0pt}b},{y_{a}^{- 1}\hspace{0pt}y_{b}})`$ for $`g_{1} = u`$ and $`g_{2} = v`$ respectively using an algorithm for the CSP. The the verification (can be done efficiently using the algorithms for the word problem when $`G`$ is the braid group e.g. see \[[10](#bib.bib10)\]) $`y_{a} = {g_{1}\hspace{0pt}a\hspace{0pt}g_{2}}`$, $`y_{b} = {g_{1}\hspace{0pt}b\hspace{0pt}g_{2}}`$ will be true by the above assumption, hence we have shown the EDL to be true in this case. For some examples of the CSP there are fast algorithm for it for example see \[[4](#bib.bib4)\] hence the assumption in \[[9](#bib.bib9)\] that the EDL is hard is not always true. If we know the generators of the subgroups A and B then we may use a length based algorithm to recover the secret element with non-negligible probability.
We re-define again the EDL problem more generally as follows
Public information: $`G`$ is a Semigroup.$`A,B,C,D`$ are subsets of $`G.`$ $`{a_{i},b_{i},x_{i},y_{i}} \in G`$ with $`y_{i} = {a_{i}\hspace{0pt}x_{i}\hspace{0pt}b_{i}}`$, $`1 \leq i \leq m`$
Secret Information: $`a_{i} \in A_{i}`$, $`b_{i} \in B_{i}`$, ($`A_{i}`$ and $`B_{i}`$ are subgroups).
Objective: Decide if $`{{F_{x_{1}}\hspace{0pt}{(y_{_{1}})}} \cap {F_{x_{2}}\hspace{0pt}{(y_{_{2}})}\hspace{0pt}\ldots} \cap {F_{x_{r}}\hspace{0pt}{(y_{_{r}})}}} \neq \varnothing`$.. Where $`F_{\beta}{(\alpha)} = {\{{(a,b)} \in B_{n} \times B_{n}}`$ :$`\alpha = a\beta b\}.`$
Theorem 3
Given the generalised EDL above is sometimes equivalent to solving the CSP. The generalised EDL may be partially solved in the connection a subset integers $`t_{1},t_{2},\ldots,t_{r}`$ in $`\lbrack 1,m\rbrack`$ we can decide if $`{{F_{x_{t_{1}}}\hspace{0pt}{(y_{_{t_{1}}})}} \cap {F_{x_{t_{2}}}\hspace{0pt}{(y_{_{t_{2}}})}\hspace{0pt}\ldots} \cap {F_{x_{t_{r}}}\hspace{0pt}{(y_{_{t_{r}}})}}} \neq \varnothing`$.
Proof
Again we will solve this problem by solving the a system of DP equations for certain values of the secrets so our proof can be used to solve a system of DP equations. Assume $`a_{i} = a_{j}`$ and $`b_{i} = b_{j}`$ for all $`{i,j} \in {\lbrack 1,m\rbrack}`$ and $`i \neq j`$ then $`y_{i} = {a_{i}\hspace{0pt}x_{i}\hspace{0pt}b_{i}}`$, $`y_{j} = {a_{j}\hspace{0pt}x_{j}\hspace{0pt}b_{j}}`$, and so the CE functions give
<table>
<tbody>
<tr>
<td></td>
<td>yiyj−1subscript𝑦𝑖superscriptsubscript𝑦𝑗1\displaystyle y_{i}y_{j}^{-1}</td>
<td>=\displaystyle=</td>
<td>aixixj−1ai−1subscript𝑎𝑖subscript𝑥𝑖superscriptsubscript𝑥𝑗1superscriptsubscript𝑎𝑖1\displaystyle a_{i}x_{i}x_{j}^{-1}a_{i}^{-1}</td>
<td></td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td>yi−1yjsuperscriptsubscript𝑦𝑖1subscript𝑦𝑗\displaystyle y_{i}^{-1}y_{j}</td>
<td>=\displaystyle=</td>
<td>bi−1xi−1xjbisuperscriptsubscript𝑏𝑖1superscriptsubscript𝑥𝑖1subscript𝑥𝑗subscript𝑏𝑖\displaystyle b_{i}^{-1}x_{i}^{-1}x_{j}b_{i}</td>
<td></td>
</tr>
</tbody>
</table>
and so we can solve the CSP pairs $`({x_{i}\hspace{0pt}x_{j}^{- 1}},{y_{i}\hspace{0pt}y_{j}^{- 1}})`$, and $`({x_{i}^{- 1}\hspace{0pt}x_{j}},{y_{i}^{- 1}\hspace{0pt}y_{j}})`$ for the solutions $`g_{1} = a_{i}`$ and $`g_{2} = b_{i}`$ respectively using an algorithm for the CSP. We can get more conjugacy extractor functions by choosing different values for $`i`$ and $`j`$. The the verification (can be done efficiently using the algorithms for the word problem when $`G`$ is the braid group e.g. see \[[10](#bib.bib10)\]) $`y_{a} =^{?}{g_{1}\hspace{0pt}x_{a}\hspace{0pt}g_{2}}`$, $`y_{b} =^{?}{g_{1}\hspace{0pt}x_{b}\hspace{0pt}g_{2}}`$ will be true by the above assumption for all $`{(a,b)} = {(i,j)}`$, hence we have shown the EDL to be true in this case.
If we know the generators of the subgroups $`A`$ and $`B`$ then we may use a length based algorithm to recover the secret element with non-negligible probability.
The EDL can be partially solved if it is true that the assumption $`a_{i} = a_{j}`$ and $`b_{i} = b_{j}`$ for at least two integers $`i`$ and $`j`$, $`{i,j} \in {\lbrack 1,m\rbrack}`$ and $`{i \neq j}.`$ Then the proof that the EDL can be partially solved is the same as above except there are fewer choices for $`i`$ and $`j`$.
7\. Second Algorithm using CE Functions
Given (the DP equation) $`u = {x\hspace{0pt}a\hspace{0pt}z}`$. This attack reveals partial information about the secret $`z`$ or totally recover $`z`$. This attack is a generalisation of our attack on the DP by using a MSCSP.
1\. The attacker picks elements $`S_{I}`$ according to some criteria relating to commutativity, for example elements $`S_{I}`$ may be picked randomly or $`S_{I}`$ may be composed of a few Artin generators as these may commute to some degree with $`z`$.
2\. Then for $`1 \leq I \leq M`$ for a sequence of integers $`T_{I}`$
<table>
<tbody>
<tr>
<td></td>
<td>CEI(SI,u)=uSIu−1=xazSIz−1a−1x−1=xaz¯SIz¯−1a−1x−1𝐶subscript𝐸𝐼subscript𝑆𝐼𝑢𝑢subscript𝑆𝐼superscript𝑢1𝑥𝑎𝑧subscript𝑆𝐼superscript𝑧1superscript𝑎1superscript𝑥1𝑥𝑎¯𝑧subscript𝑆𝐼superscript¯𝑧1superscript𝑎1superscript𝑥1CE_{I}(S_{I},u)=uS_{I}u^{-1}=xazS_{I}z^{-1}a^{-1}x^{-1}=xa\overline{z}S_{I}\overline{z}^{-1}a^{-1}x^{-1}</td>
<td></td>
</tr>
</tbody>
</table>
where (with probability $`\rho)`$ $`\overline{z}`$ is a partial factor of $`z`$ for some $`I`$ this means $`z = {z_{T_{I}}\hspace{0pt}{\overline{z}}_{T_{I}}}`$.
3\. We solve for each $`I`$ the CSP $`(S_{I},{C\hspace{0pt}E_{I}\hspace{0pt}{(S_{I},u)}})`$ for the solution $`xa\overline{z})`$ and hence compute $`z_{T_{I}} = {({{({x\hspace{0pt}a\hspace{0pt}{\overline{z}}^{- 1}})}^{- 1}\hspace{0pt}x\hspace{0pt}a\hspace{0pt}z})}^{- 1}`$. Note if $`S_{I}`$ is selected from the centraliser of $`z`$ then we can use the MSCSP at this step (so this shows DP is Turing reducible to MSCSP).
4\. We now find (in some way) $`z`$ using the information $`(S_{I},{x\hspace{0pt}a\hspace{0pt}\overline{z}},z_{T_{I}})`$ and the other information used in the protocol. One of the simplest choices to implement this step is trying to find $`\overline{z}`$ for each $`I`$ by brute force and hence possibly recover $`z`$.
A variant of the above attack is after $`z_{T_{I}}`$ is recovered is to repeat at the attack (at least once) by iterating with $`u\hspace{0pt}z_{T_{I}}^{- 1}`$ instead of $`u`$ (and obviously all other values may be different) so in this way we may be able to find a bigger factor of $`z`$. It may be true (with some probability $`\rho_{2}`$) that $`\overline{z}`$ contains a partial factor of $`a`$ which means the CSP is solved to give $`{\overline{z}}_{T_{I}}\hspace{0pt}a_{T_{I}}`$ where $`a_{T_{I}}`$ is some partial factor of $`a`$. Then the simplest choice at this step to recover $`z`$ is to find $`a_{T_{I}}`$ by brute force and use $`a_{T_{I}}`$ to recover $`z.`$
Conclusion
We have shown the protocols \[[10](#bib.bib10)\],\[[2](#bib.bib2)\],\[[6](#bib.bib6)\],\[[7](#bib.bib7)\],\[[11](#bib.bib11)\],\[[13](#bib.bib13)\],\[[14](#bib.bib14)\] can have security based on the MSCSP. We have shown the DP and DH-DP can be solved by the MSCSP. Our theorem 1 implies that the CKLHC cryptosystem and related cryptosystems are MSCSP based so are no more secure than using the AAG protocol \[[1](#bib.bib1)\]. Our theorem 1 implies that semigroups should (for $`G`$) be used for the protocol in \[[12](#bib.bib12)\] to be secure so not based on the MSCSP. We should not use the CKLHC protocol in \[[10](#bib.bib10)\] or related protocols (which are suggested to be used in braid groups) compared to using the AAG protocol as it is no more secure than using the AAG in the connection they are based on the MSCSP.
## References
- \[1\] I. Anshel, M. Anshel, D. Goldfeld. An Algebraic Method for Public-key Cryptography, Mathematical Research Letters, 6, 1999, pp. 287-292.
- \[2\] K.K. Ko, S.J. Lee, J.H. Cheon, J.W. Han, J.S. Kang, C. Park. New public-key cryptosystem using braid groups. CRYPTO 2000. LNCS, 1880, 2000, pp. 166-183.
- \[3\] K. H. Ko, Tutorial on Braid Cryptosystems 3, PKC 2001, Korea, February 13-15, 2001. Available at www.ipkc.org/pre_conf/pkc2001/PKCtp_ko.ps
- \[4\] V. Gebhardt. A New approach to the conjugacy problem in Garside groups. J Algebra 292 (1) , 2005, pp. 282-302.
- \[5\] D.Garber, S. Kaplan, M. Teicher, B. Tsaban, U. Vishne. Length-Based Conjugacy Search in the Braid Group, available at http://arxiv.org/abs/math.GR/0209267
- \[6\] V. Shpilrain and A. Ushakov, ”Thompson’s group and public key cryptography”, LNCS, 3531, Springer-Verlag, 2005, pp. 151-164.
- \[7\] V. Shpilrain and A. Ushakov, A new key exchange protocol based on the decomposition problem, available at http://eprint.iacr.org/2005/447.pdf (Accessed 20/05/2006)
- \[8\]
J. Cheon and B. Jun, ”A Polynomial Time Algorithm for the Braid Diffie Hellman Conjugacy Problem”, Proceedings of Crypto 2003, LNCS, 2729, Springer-Verlag , 2003, pp. 212-215.
- \[9\] T. Thomas and A. K. Lal, Group Signature Schemes Using Braid Groups, arXiv:cs.CR/0602063, 2006
- \[10\] J.C. Cha, K.H. Ko, S.J. Lee, J.W. Han, J.H. Cheon, An efficient implementation of braid groups, Advances in cryptology: Proceedings of ASIACRYPT 2001,Lecture Notes in Computer Science, Springer-Verlag, 2248, pp.144-156, 2001
- \[11\] E. Stickel, A New Method for Exchanging Secrets, Proceedings of the Third International Conference on Information Technology and Applications, 2005
- \[12\] M. M. Chowdhury, Key Agreement and Authentication Scheme Using Non-Commutative Semigroups, ArXiv preprint, August 2007
- \[13\] Tang Xue-Ming, Hong Fan, Cui Guo-Hua, A Public Key Encryption Algorithm on Braid Groups, Journal of Software, Vol 18, No. 3,March 2007, pp.722-729
- \[14\] P. Dehornoy, Using Shifted Conjugacy in braid-based cryptography, arXiv preprint, 2006
- \[15\] P. Staecker, Computing Twisted Conjugacy Classes for Free groups using nilpotent quotients, 2007 |
0710.5006 | cs.NI | cyber | # CANE: The Content Addressed Network Environment
Paul Gardner-Stephen (gardners@infoeng.flinders.edu.au)
###### Abstract
The fragmented nature and asymmetry of local and remote file access and network access, combined with the current lack of robust authenticity and privacy, hamstrings the current internet. The collection of disjoint and often ad-hoc technologies currently in use are at least partially responsible for the magnitude and potency of the plagues besetting the information economy, of which spam and email borne virii are canonical examples. The proposed replacement for the internet, Internet Protocol Version 6 (IPv6\[[13](#bib.bib13)\]), does little to tackle these underlying issues, instead concentrating on addressing the technical issues of a decade ago.
This paper introduces CANE, a Content Addressed Network Environment, and compares it against current internet and related technologies. Specifically, CANE presents a simple computing environment in which location is abstracted away in favour of identity, and trust is explicitly defined. Identity is cryptographically verified and yet remains pervasively open in nature. It is argued that this approach is capable of being generalised such that file storage and network access can be unified and subsequently combined with human interfaces to result in a Unified Theory of Access, which addresses many of the significant problems besetting the internet community of the early 21st century.
## 1 Introduction
What are the real problems with the internet today? Is the internet user base most acutely pressed by the impending exhaustion of the IPv4\[[1](#bib.bib1)\] address space? Is it perhaps that the current Internet Protocol Version 4 (IPv4) has trouble with sending more than a few giga-bytes in a single session? Maybe it is that 65,536 ports per host just isn’t enough? Or is it that getting allocated an IP address and finding the grass roots network services you require on a new network, such as DNS, is too hard?
While these are problems with the current internet, they are not impassible road blocks in the progress of the so called “information society” or “information economy”. These problems have been mitigated by new technologies such as network address translation and PPP. Indeed, IPv6 primarily targets technical issues which have already been mitigated. Although these issues may remain important, they are no longer the most pressing issues facing the internet today. Thus, there is no universally compelling advantage or reason to make the transition from IPv4 to IPv6, explaining why IPv6 is still not in extensive use a decade after its release.
If it is not the technical issues addressed by IPv6 that are the bugbear of the internet then this logically implies that there are other problems which need to be identified and addressed; IPv6 fails to address the real problems facing the internet in the 21st century.
The strict traditional definition of the internet as the global inter-network of computers is too narrow and technical to be effective here. I suggest that the a more useful definition of the internet is the global inter-network of agents (human or artificial) communicating using computers. This difference highlights the reality that the outstanding problems of the current internet lie in the interaction of agents (human or artificial), not computers. The internet should make it simple for people using computers to do what they want to do. If they want to access and modify data, then they should not be hampered by their location or that of the data. They may also want to do this with appropriate security and privacy. All other uses of the internet (according to our definition of it) are derivatives of this. Consider the following two examples: (a) publishing web pages; and (b) email.
The first example, publishing web pages (in its simplest form) can be modelled as providing globally unrestricted read-only access to a collection of files. In the current internet this is implemented using a special program (the web server) to make the pages available, and another special program (the web browser) to read the pages.
The second example, email, can be modelled as writing a file you have created (the message) into a well known writable location in the recipients storage (their email inbox). In the current internet this is implemented using a special program (an email composing and sending program) which communicates to another special program (the email delivery server) using a special protocol (usually SMTP), which places the message into a special kind of storage (the recipients email inbox, which usually exists outside the name space of the recipients regular file storage). The recipient then uses a special program (an email reading program, which is probably integrated into the email composing and sending program) to talk to another special program (the email access server) using another special protocol (usually POP or IMAP) to retrieve the messages and display them to the user.
Now consider how these examples might look if there was a protocol which made it possible to access remotely stored data as though it were locally stored, including read and write access, as well as the enforcement of user stipulated permissions.
The first example, publishing web pages would be accomplished by making a collection of files globally readable. The pages could be accessed by the reader using the same tools they use to explore and access files stored on their own computer. Thus the result is a simpler and more consistent architecture.
The second example, email, would be accomplished by writing a file to a well known location (their email inbox) in the recipients storage. However, we can go further with email by creating additional delivery locations (perhaps sub-directories of the well known location), and placing restrictive permissions on them. You could create a “friends” folder that only accepts email messages from a list of known senders. The email address notation could be refined to take this into account, e.g., by allowing email addresses such as paul/friends@infoeng.flinders.edu.au. While Spammers could still deliver messages to paul@infoeng.flinders.edu.au, only people I specify could deliver messages to paul/friends@infoeng.flinders.edu.au.
In short, by changing the internet from a computer centric collection of specialised programs and protocols to a simplified entity centric model it is possible, not merely to remove complexity, but to add substantial value to the internet. The remainder of this paper visits, in more detail, the issues touched on in this section, namely the real problems with the internet, and the need for an alternative (Section [2](#S2 "2 The Real Problems With The Internet ‣ CANE: The Content Addressed Network Environment")). Such an alternative is proposed and is described in two parts: (1) simplifying data access (Section [3](#S3 "3 The Content Addressed Environment ‣ CANE: The Content Addressed Network Environment")); and (2) making the internet entity centric instead of computer centric (Section [4](#S4 "4 Agent Identification and Security ‣ CANE: The Content Addressed Network Environment")).
## 2 The Real Problems With The Internet
Having defined the internet as a essentially a horde of agents acting on a plethora of computers, it becomes clear how IPv4 and IPv6 are addressing superficial problems when compared with those present in the rich and socially involved internet as we have defined it. What follows is a list of the real problems with the internet.
### 2.1 Trojan Everything
By far the biggest problem on the internet is that of trust. We cannot trust that a packet, file or email message has really come from the agent that it claims, or that it has not been subverted or intercepted in transit. It is the lack of trust which makes the internet a dangerous place to do business, work or play. This environment of fear has serious ramifications for business\[[5](#bib.bib5)\]. For justification of this premise look no further than the current emphasis on security and security products, such as virus scanners, spam filters, digitally signed device drivers, cryptographic privacy and authenticity protocols and products. The list is almost endless, and yet practically every one of them is actually just a salve for the symptoms of the lack of trust endemic to the internet. If trust could be assured, blights such as spam could be wiped out almost instantly; it would no longer be possible to hide behind open relays or on compromised computers. The source of the spam or phishing attempt would be revealed in every delivered message. Virii and mal-ware are similarly the trojan version of software. If an effective mechanism existed for ascertaining the ultimate source of a given piece of software, then no one would ever run malicious software in the first place.
### 2.2 Remote Data Access and Manipulation
If the biggest problem with the internet is trusting what you get, then the next biggest is actually getting what you want. The web is good for providing read only access to public data. It is also fair for granting read only access to a restricted audience. However, it is woeful at providing read/write access with anything like the ease and control of such access to a locally resident file system. There are a whole suite of protocols which attempt to provide such access with varying focii and success, such as WebDAV\[[2](#bib.bib2)\], SFTP\[[12](#bib.bib12)\], NFS\[[9](#bib.bib9)\], SMB/CIFS\[[6](#bib.bib6)\], CODA\[[11](#bib.bib11)\], rsync\[[15](#bib.bib15)\], Klik\[[8](#bib.bib8)\] and ZeroInstall\[[7](#bib.bib7)\]. Their sheer quantity and diversity proclaim that there is some way to go before, from the users perspective, remotely resident data is equivalent in all ways to locally resident data.
### 2.3 Software Installation and Maintenance
Assuming that you have got what you want, the next big problem is getting what you’ve got to work. This discussion will focus on applications, because if you can get applications right, this all but guarantees solutions for documents and data files. The crux of this problem is that downloading a bunch of software applications, including all their dependencies, and getting them to run on on the same computer at the same time is often difficult or impossible. To demonstrate some of the complexities present, consider the not uncommon case where you want two versions of the same application - and their incompatible dependencies - on the one computer. This is a surprisingly common situation on, for example, web servers where the web server, database and scripting language versions all have to be matched. Even more common is the desire to run the same application on multiple computers with different processor architectures and/or operating systems. Conflicts aside, managing software installation without administrator or super-user privileges is generally problematic, despite the fact that there is no fundamental reason to demand such access to accomplish the installation. To summarise, while appearing benign, software management is actually one of the most painful activities regularly required to be performed on a computer today. This is especially true when combined with the related ogre of Operating System Rot, where many operating systems require complete re-installation in order to restore functionality if too much software has been installed and uninstalled. A number of technologies have been devised to deal with the manifold problems of software management, including ZeroInstall\[[7](#bib.bib7)\], Klik\[[8](#bib.bib8)\], Java, AppDirs, ROX, numerous \[Un\]Installers, BSD Ports, and countless other packaging systems. The degree of success of these approaches varies.
### 2.4 Backup, Restore and Archiving
This is the problem of getting back something that you had in the past. The goals of effective data storage archiving and recovery are problematic. It is not just the question of can we get a file back from a backup tape made a week ago. In the ideal world, people would like to be able to undo file system changes just like they undo typing in a word processor. Recovering deleted, corrupted, replaced, subverted or edited data are significant, though perhaps subconscious, goals of computer users. Consider web server administrators who would like to be able to rapidly reverse the changes made by a group of hackers who have defaced the corporate web site. Episodic backups with lengthly off-line recovery processes are no longer satisfactory. This is evidenced by the scrabbling among backup vendors to attain true continuous and on line backup and recovery. The reality is the products sold as continuous are almost invariably frequent episodic rather than truly continuous. These systems are also notoriously expensive, in part because the backup vendors know that people desire and are willing to pay for continuous access to the history of their data. This significance of this issue has also been appreciated in various research environments, such as Plan9\[[10](#bib.bib10)\], where the authors caution the reader not to underestimate the productivity gains of having instant and on line access to (in their case) nightly backups of all data.
### 2.5 Congestion Instead of Cooperation
Finally, the internet has congestion problems which have little to do with the available bandwidth. Numerous researchers have pointed out that the internet has sufficient bandwidth to supply our needs, but that it is often poorly used. As an example of this phenomena, consider the manner in which the internet behaves when many users attempt to download the same resource: Pathological congestion occurs instead of cooperation. The very activity which is so clearly shareable, the downloading of the same file, is that which instead clogs the internet with a plurality of simultaneous and near identical data streams. By way of solutions to this, we have mirrors which suffer from synchronisation and non-transparency issues. That is the end user may be aware that they are not dealing with the original source, and may be required to employ alternative access methods. IP Multicast has the right idea by sending the data only once. However it is poorly suited to downloads which commence at different times. Finally, the recent explosion of peer-to-peer (P2P) protocols such as Bit Torrent, while possibly lessening congestion at the back bone tend to make the edge congestion much worse.
### 2.6 The Spatial Addressing Paradigm Is Flawed
These problems exist today because IPv4 is based on spacial addressing, i.e., it addresses data based on its location. Moreover, these problems will continue to exist if the internet does eventually transition to IPv6, because data will still be addressed by location. It is true that many partial solutions exist, and in the case of certain problems quite good solutions. It is also true that when IPv4 was developed, that spacial addressing was the most effective paradigm given the limited speed and memory of computers at the time. However, there is no single coordinated solution based on spacial addressing which is able to address the problems, at their cause rather than merely symptomatically, with the internet that have been described in this section. A paradigm shift is required: The current paradigm of spatial addressing must stand aside. Yet we must address by some criterion. The remainder of this paper considers the benefits of a content addressing paradigm, and briefly explores how the paradigm could operate in practise.
## 3 The Content Addressed Environment
The content addressing paradigm addresses by the identity of data rather than its storage location. That is, stop asking for that thing stored in some disk block or on some server on the other side of the world - which is really an indirection - and instead ask for the thing itself. Cryptographic hashes turn out to be well suited to fulfilling this concept of identity
<sup>1</sup>
<sup>1</sup>1This is of course assuming that random cryptographic hashes exist. MD5 and more recently SHA-1 have come under increasing fire in the form of discoveries of cryptographic weaknesses in them \[[14](#bib.bib14)\]. For the purposes of this paper we assume that a sufficiently sound cryptographic hash algorithm does exist.
. They are designed to produce a unique hash value, or identity, for any given input. The hash has the important property of being much shorter than the data it identifies. This simple difference in addressing has profound impact on the way computers and networks of computers can operate. This section will discuss the immediate benefits that straight forwardly arise from such a change in methodology.
### 3.1 Unification of Remote and Local Data Storage
The first consequence is that we can all but remove the distinction between file system and network resident storage. Present file systems specify data by where it is stored on the disk. This implies that the data must be stored on the disk to be accessible. Implicit then is that accessing network resident data requires a different protocol: Network resident storage becomes second class compared to local storage. If instead we specify data by its identity, then the data can be sourced from anywhere, local or not. An additional advantage of the hash which is not present in traditional location addressed storage in that it allows us to verify that the data we have read is correct by verifying that it produces the expected hash
<sup>2</sup>
<sup>2</sup>2Sun’s new ZFS file system is a notable exception to this, in that is stores cryptographic hashes of data as well as their storage location. However ZFS does not make full advantage of the opportunities afforded by true content addressing.
: If even a single bit is altered the hash will be completely different. If the hashes are long enough and are truly random, we have enough address space to be confident in discriminating between all data in the world. A good quality hash with length of perhaps 512 bits should be adequate in this role: If $`2^{40}`$ computers generated $`2^{40}`$ data items per second for the next four billion years, this would be only $`2^{40} \times 2^{40} \times 2^{32} \times 365.24 \times 24 \times 60 \times 60`$ items, which is less than $`2^{140}`$. Using a 512 bit randomly distributed hash this would put the probability of a hash collision, i.e., a situation where two data would generate the same hash and so be indistinguishable to the system at less than $`10^{- 70}`$. In comparison the probability of winning a typical national lottery is around $`10^{- 7}`$. This means that it is $`10^{63}`$ (the number one followed by 63 zeros) times more probable that you will win your national lottery at least once in your life time than a hash collision will occur in the next four billion years. This does of course assume that the hash algorithm is sound, and that collisions cannot be easily coerced. A full discussion of this issue is outside of the scope of this paper.
### 3.2 Efficient On Line Continuous Backup and Storage
The true potential of a content addressed file system comes when it is combined with persistent data storage, such as proposed in Plan9. Unlike Plan9 however, a content addressed data store does not require explicit backups, and is not constrained to episodic access to old data. This is because the hash of a given version of a file or directory is a permanent and immutable identifier: If you change a single bit in the file it will be identified by a new hash, and the old hash will still identify the old version. All that needs to be done to provide on line access to historical data is to put the hash of the old version of a file somewhere convenient whenever the file is updated. One simple approach to this problem is to make use of the well known meta-directories present in every subdirectory. Most computer systems already have the “.” and “..” directories which refer to the same directory and the parent directory. In a location addressed file system this is easy to achieve - these directories just point to the file block concerned. However, there is a circularity problem when this approach is attempted in a content addressed file system. This occurs because the hash of the current directory which is available when you write a directory listing is the hash from before the write occurs, not after. Therefore the “.” meta-directory will actually point to the previous version of the directory. This is illustrated in Figure [1](#S3.F1 "Figure 1 ‣ 3.2 Efficient On Line Continuous Backup and Storage ‣ 3 The Content Addressed Environment ‣ CANE: The Content Addressed Network Environment") where we want to write a “.” entry into the directory listing on the left, so we record its hash 0x7654 against the name “.”. But as a result the hash of the directory has changed to 0xb219, leaving “.” pointing to the old version. Similar effects occur if a file changes and cascade all the way to the top of the file tree - always leaving the previous hashes associated with the “.” directories as a trail of bread crumbs leading back in time. This scheme can be further developed by augmenting the “.” directory as a pointer to the current directory with a new meta-directory, perhaps “…” which points to an automatically created and maintained directory which contains pointers to all versions of the directory and named with their date of modification. This avoids having to construct exceptionally long paths to get to ancient versions of a file. In practice this means that we can use a path such as “Documents/…/14July2005@14:23.17.1/File.doc” instead of “Documents/./././././././././././././././././././././././File.doc” to refer to the last good copy of some important file.
[Refer to caption](/html/0710.5006/assets/x1.png)
Figure 1: Using “.” Meta-Directory to Facilitate Access to Old File Versions
We have as a natural consequence of content addressing, and with practically no effort at all, achieved on line access to truly continuous backups. Provided the final backing data store is persistent, and grows faster than data usage, then there is no limit to the depth of the backups which can be accessed. The total storage space required to introduce such a regime is perhaps surprisingly smaller than what might first be expected. This is because we are addressing by content and not location, and therefore duplicate instances of the same data does not result in duplicated storage. This applies not only to spatial duplication where a file exists in more than one place at the same time, but also to temporal duplication where a file may be large but may take several recurring forms over time. If the content addressing is performed at a suitable block level, then this mitigation of redundant storage applies to the content within a file. A file full of all zeroes will always require only one data block. A new version of a file with only one modified data block will only require one new data block. A file modified and then reverted to a previous version will require no new data blocks. Identical code and data pages can be shared not only between instances of the same library or application, but among any combination of libraries or applications that have the same binary content. In short, there are a number of mechanisms by which content addressing, complete with on line storage of all previous versions of all files may actually require less storage than traditional file systems, and yet be able to make more effective use of available storage facilities.
Content addressed storage systems lend themselves naturally to hierarchical storage systems. This is of great relevance because combining the on line continuous backup capability of a content addressed system with a well designed, and preferably replicated, hierarchical storage system obviates the need for any other backup system. Better yet, the entire capacity that was once dedicated to massively redundant backups (consider for a moment the number of copies of common operating system files that would be in the typical enterprise backup archive) can now effectively be used as primary storage. In most organisations this would involve from tera-bytes to exa-bytes of additional capacity: A well designed and implemented backup regime will usually contain at least twice, and often many more times, the capacity of the system it is protecting. This capacity is well priced in terms of \$/giga-byte, but is effectively latent to the organisation. Because of the independence of the storage in a content addressed environment it is trivial to continuously add more capacity in order to accommodate demand. This significantly simplifies the data and backup management issues which the computer industry faces today.
A few final notes about the capabilities of a file system which automatically records its entire edit history are worth mentioning. One is that if a computer system is compromised or subverted it becomes trivial to revert it to the last known good state by traversing the unalterable file system history. Interestingly it is also trivial to regain any logs or other records which intruders or saboteurs might seek to alter or remove in order to conceal their activities. The second interesting feature is that of version control and management of race conditions. Since the file system will record each event in a race condition where two people try to simultaneously write to the same file, it is no longer possible to lose data in such a situation. It may even be possible to develop the file system to the point where it can detect that a race condition has occurred, and either flag the files or perhaps undertake automatic resolution of any conflicts. Finally, the immutable nature of storage in this context may have ramifications from a civil liberties perspective, and also with regard to various national privacy laws which can require the irrevocable destruction of data.
### 3.3 Fat Recursive AppDirs and Fast Copies
It has already been mentioned that content addressing makes it space efficient to store redundant copies of either entire files or portions of them, and that this efficiency is an automatic byproduct. This, when combined with the unification of local and remote storage, offers a compelling opportunity to radically simplify application management. AppDirs exist in several operating systems today, most notably NeXT and Apple’s OS-X, but increasingly in various forms, such as the ROX desktop on UNIX and UNIX-like systems generally. AppDirs simplify application installation by storing all the files which are required for an application to run in a single directory. The directory can be placed anywhere in the file system and still function. They can therefore be installed and uninstalled by unprivileged users without contributing to operating system rot. In practice AppDirs are hamstrung by their incomplete encapsulation: Most applications depend on other applications or libraries, which may or may not also be available as AppDirs. Whichever the case, the point remains that AppDirs still often have external dependencies. In theory it is possible to correct this by including every dependency of a given application inside its AppDir. However this is grossly inefficient when you have multiple applications sharing the same dependencies. In contrast, Content addressing however completely removes this inefficiency through the singular storage of redundant data. Hence it becomes feasible to produce recursive AppDirs. Recursive AppDirs are AppDirs which include their entire dependency tree, or more precisely, the hash, and hence address, of each dependency. Because every reference to a specific dependency will be identical, duplicated dependencies (both within and among applications) are stored only once. Retrieval of such large AppDirs over the network also benefits from the properties of content addressed storage in that any portions of the AppDir which are already stored locally do not need to be fetched via the network.
The concept of recursive AppDirs can be broadened to efficiently accommodate cross platform compatibility. This is accomplished by the natural property of a hierarchical and network connected content addressed storage system to only fetch data on demand. Therefore a Fat Recursive AppDir could contain the complete dependency tree for your favourite application on seven versions of Linux, Microsoft Windows NT, 95,98,2k,XP and 2003, five BSDs, four processor architectures, three byte orders, two CPUs and a hand held PDA - and yet thanks to lazy fetching it would take no longer to install or run than a lean AppDir with only the compilation you require. For applications which are not suitable for implementation in Java, yet can be ported to multiple operating systems this offers a near perfect solution to a broad range of software management and operating system stability problems, without demanding convergence on a single operating system or processor architecture.
### 3.4 Cachable Network Traffic
Because data is addressed by its identity rather than location, it is possible for the network to cache recently seen data. If another agent requests the same data it can safely reply with the cached copy, because it is provably asking for the cached data. This has the potential to profoundly improve network performance in the face of congestion caused by repeated access to identical data, e.g., when the new version of a popular software distribution is released and simultaneously downloaded by huge numbers of people. On the current internet this results in the network equivalent of grid lock. In contrast on a content addressed network the data would be cached in major routers and requests for the data would be satisfied by the first router contacted which had cached the data. In this way data moves from its point of origin, with its potential bandwidth bottleneck, to a myriad of points on the edge of the network where it is wanted. There are two consequences of this which are significant besides generally speeding up the internet: First, it becomes exceedingly difficult to take a site off the internet via a DoS or DDoS attack. This is because after a short while the data is being fed from many points instead of just the point under attack. The second effect is that because of this caching effect it is no longer necessary for web servers serving popular static content to require large and expensive internet connections. The average citizen using a dial up modem could serve up static content which could be accessed by millions of people simultaneously.
### 3.5 File Mapped Human Interfaces
The non-redundancy and manner in which change in a content addressed file system can be detected by monitoring file and directory hashes can be harnessed to facilitate the production of efficient new technologies which address a variety of problems. One example of this is human interfaces. In this context, human interfaces are defined as the mechanisms which a program uses to communicate with a human. This includes principally the traditional video display, keyboard and mouse. Presently programming these devices, and the video display in particular, is performed in a myriad different ways on different systems; for evidence of this consider that the three predominant operating systems of today each use radically different display libraries and protocols. We have the Windows’ GDI, UNIX/Linux’s X11 and Apple’s Aqua. What is interesting is that each of these is attempting to provide the same thing: a mechanism for programmers to describe to the operating system how to display certain data and to receive feedback from the user. Even if constrained to the WIMP
<sup>3</sup>
<sup>3</sup>3Windows, Icons, Menus and Pointer.
paradigm, the programming mechanisms required differ greatly, and are often surprisingly indirect. Ultimately the programmer attempts to describe the scene to be displayed to the computer. Such scenes are hierarchical compositions of a relatively few basic elements, such as window, text, button, image and so on. This type of description can be readily described by a file system, where directories represent branches and files the leaf nodes. If this is done, then it will be possible to unify not only the file system and network, but also the most significant remaining input and output channels.
If a graphical display is described as a series of nested directories and files, representing components such as windows, images and text, it is mandatory that it be efficient to transform this into its visual form. If a content addressed file system is used to represent a display, it becomes easy to cache the rendered version of the display using the hash of its representation as the key. Moreover, this process can be performed recursively, so that a display that contains identical components need render each unique component only once; subsequent renderings are performed rapidly by copying them from the cache, even within the same display. This caching of displays and their components makes it efficient to update a changing display, as only the changed components require rendering. Further, if a display or component changes to a different, but previously rendered one, it can again be rendered by copying from the cache. Because drawing the display in this way is performed by recursively rendered from the top level down, the caching can occur at any level of composition, and automatically re-uses the highest level of composition possible in any given circumstance. As an example, moving a window containing a complex composition would result in only the window frame altering (changing coordinates), while all the internal compositions would remain unchanged. Therefore the contents of the window would be redrawn from the cache rather than rendered again. Further efficiencies can be realised by collapsing small files into their parent directory to produce “light weight files” (LWF). The use of LWFs reduces the number of block retrievals required to obtain the necessary scene data.
Input and event handling is achieved using the same process and caching algorithm as for rendering, except that the product is an event map rather than a rendered display. The event map is simply the sub-division and appropriation of the visual real estate to the elements and event receptors in the scene. Event receptors are simply files in the directory hierarchy which exist solely for the display engine to deposit events in order to communicate them to the program. Display efficiency and functionality can be enhanced by the addition of a simple scripting language which allows selected events to be handled completely by the client side. Theoretically any event which does not require the retrieval of new display information can be dealt with in this way. A list of events suited to local processing includes selection of radio buttons, menu fade in, fade out, pop up and pop down, tool tip display and retraction, scrolling and zooming of panes in response to scroll bar activity. Importantly, these can be processed locally are also those where the introduction of latency has the largest degrading effect on interactivity. Indeed, it would be possible to have the file hierarchy describe not only the visual components of the scene, but also application state and program, facilitating the flexible combination of local and remote computation.
Such a rich and precise description of a scene, combined with its clear delineation between interface description and visualised representation would make this approach well suited as a replacement to HTML. The user interfaces would be richer and more interactive than those produced with AJAX, and yet have significantly reduced complexity.
Finally, the clear division of the description and rendering layers also stand to make accessibility and access by restricted devices (such as small embedded devices) much easier to implement. Essentially any format of representation could be implemented because the all of the necessary information is present in the scene description. The standard graphical WIMP renderer could be replaced with a speech based renderer for the visually impaired, or a text based renderer for low bandwidth applications, or in an instance of fact following fiction, it would be possible to create a glyph based rendered like the one in the Matrix movies.
-
## 4 Agent Identification and Security
### 4.1 Public Key Addressing
The previous section has considered some of the benefits which may result from adopting a content addressing paradigm for general purpose computing. By applying a similar philosophy to network identification and authentication it can be conceived that what agents really want with network identification is to be able to select the agent or agents they wish to communicate with, and then to be able to communicate with them securely and confidently. In such a framework the concept of each computer rather than agent having a digital address is an indirection at best. Rather we want each person (natural or otherwise) to have their own address. Ideally it should not be feasible to impersonate another person’s digital address. A natural choice for the digital address is the public key from a public/private cryptographic key pair.
Public/private cryptography works on the basis of creating two binary patterns. One of which allows the encryption of data, and the other decryption. The encrypting key can be made public with little risk of any third party being able to decrypt anything which is protected by the encrypting key. Further, in many schemes it is possible to use the private key to sign data such that is cannot be modified without breaking the signature. These are the necessary ingredients for making secure and confidential bidirectional communications. With the benefit of public/private cryptography a person can generate a key pair and release the public key to the world. This allows any other person to encrypt data such that only the first person can decrypt it. Conversely, any data which the first person sends to any other person can be signed with their private key to “prove” that they were the originator. This approach is already in use widely, for example in the Secure Sockets Layer (SSL\[[4](#bib.bib4)\]) which is used to protect web traffic using the HTTPS\[[3](#bib.bib3)\] protocol.
### 4.2 No Need For Central Authorities
An Achilles heel in the traditional application of asymmetric cryptography is the highly centralised and at times problematic certificate authority structure. Certificates must be obtained before you can vouch for your own identity, but this is not necessary. For example, when you telephone your financial institution the only assumption they make is that your telephone call is not being listened to or subject to hijacking by malicious third parties; the line is assumed to be reasonably secure and confidential. Banks do not (generally) require your telephone number and name to be listed together in some centrally administrated repository. Rather they use the confidentiality and security offered by the phone line as a platform for asking you a series of questions which do verify your identity (postal address, mother’s maiden name, etc). This same policy can be applied to network identification; let users create any secure and private connection and determine the trust relationships as they go. This elides the requirement for a central certifying authority for most users, and any large institutional organisations, such as banks, which still prefer to obtain certificates from a trusted authority would be able to do so. The net result of this is that end users can join a network which authenticates and addresses by public cryptographic key simply by generating a random key pair and keeping their private key private. There is no longer need for DHCP, centralised IP address allocation or similar schemes. Using key pairs large enough to offer medium to long term security, perhaps 2048 bits, results in a massive address space such that all molecules in the solar system could have individual addresses with practically zero probability of address collision. From a civil liberties and global equality perspective the removal of the need for central authorities and the need to authorise people joining the internet is a tremendous gain. While it is true that you can be certain of the party you are communicating in such an environment, anonymity can be obtained as required by generating a new temporary key pair for the necessary duration, after which the incognito identity can be completely discarded.
### 4.3 Mobility, Trust and Security
The concept of addressing by public key rather than by the computer you are connected to has several synergies with a unified computing environment as introduced in this paper. First, it provides a global authentication mechanism for the content addressed storage environment. That is public keys can be used in the place of machine local user and group ids (groups are just virtual people). This makes it possible to truly make local and remote storage equivalent, as fine grained security policies can be equally applied to both. If group membership is represented by holding a certificate from the group vouching for membership then we actually gain some valuable semantics currently not readily available. Certificates can have begin and end time points. This means it would be possible to grant membership to a group for only a limited period of time, after which the membership implicitly and automatically is revoked. The only requirement is that any server which stores your data must honour the permissions set on files it stores. Files can be signed with the private key of the creator to authenticate the permissions and restrictions applied. File servers can then demand that a validly signed file block be presented as evidence that a client is entitled to the content of a file. This also has the result of all files and applications being implicitly signed by someone. Recall that changing a single bit in a file will change its hash, which will then change the hash of the directory containing it, cascading up the entire directory tree. Therefore it is impossible to add, remove, modify or alter the permissions of even a single file in an entire application without it being noticed. Each user can choose which agents they will consider as trustworthy software publishers. By placing these checks into the operating system virii and mal-ware can be controlled or even eradicated. Yet at the same time this form of trusted computing is open and decentralised in nature while still offering the attractive fruit of the closed platforms.
Having established true and secure global authentication, network mobility comes for free. We have already described how using public keys as network addresses creates the framework for secure and authenticated communication. This means that trust is based on explicit trust relationships rather than location in the network topology. Since trust and security are independent of location, virtual private networks and firewalls are both completely obviated. An agent may relocate anywhere in the network without affecting any trust relationships. Conversely, no person may attempt to acquire trust relationships by moving (or pretending to move) closer on the network. This framework of explicit rather than implicit formation of trust relationships simply makes good security sense.
### 4.4 Protecting Identity and Privacy at Home and Away
The lynch pin in this authentication framework is the effective protection of each person’s private key. Storing the private key on a conventional computer is possible, and may be justifiable in many circumstances. However a more robust solution is required in order to fully capitalise on the opportunity for transparent global mobility. This is a particularly difficult problem as without the proprietary Trustworthy Computing platforms it is not reasonable to trust any device that you do not have adequate control over. In fact in many cases it is not even wise to trust a computer which you do control as it may still be subverted. To tackle this, some additional hardware device is required which can be fully trusted and controlled by the owner. If such a device contained the cryptographic keys and signing hardware it would be possible to operate a computer in connection with the device such that the computer asks the device to sign outgoing packets and to decrypt incoming packets along with any other procedure which requires the use of the private key. This makes the software required on such an identity token relatively simple and clearly defined. The token would generate the key pair internally, and be designed in such a way that it cannot reveal the private key.
The use of an identity token provides a convenient manner in which to realise convenient global mobility. Note that by using a physical token with perhaps a simple pin number or biometric “password” which is programmed into the device itself that two factor authentication is instituted, sufficient for most organisations security requirements. In addition, the token can theoretically be inserted into any reasonably trusted computer. The token provides your public key as your digital identity, and signs any requests for data you may want. The computer therefore has all that it requires in order gain access to your data on your behalf. Owing to the unification of file, network and graphical user interfaces it is trivial to display your current computer session to you anywhere in the world. With the use of Fat Recursive AppDirs it is also straight forward to run applications locally on the computer you have connected to - without requiring administrator rights on that computer.
It is not always wise to trust a computer which you do not control. Therefore for completely secure global roaming the identity token could be supplemented with a simple note book type computer. This computer would require only a very simple operating system which can communicate with the content addressed network. Its sole job would be to provide rendering of a user session and accept keyboard and mouse input from the user: It is a thin terminal which can be used anywhere in the world. By controlling the user interface, particularly the keyboard and mouse, and encrypting and authenticating all traffic the security loop is closed: in no place can user input or server output be intercepted. This approach would make the production cost and mass of such a unit very low compared to a fully fledged mobile computer, yet would still offer a rich user interface. Finally, it would be possible to create a hybrid approach where the identity token sports a USB port to accept keyboard and mouse and then on its secure platform encrypts the user input for transmission.
## 5 Conclusions and Future Directions
We have introduced CANE, a content addressed network environment, which we have argued can address the major deficits of the current internet environment:
- \*
Trojaned correspondence, communication and identity are rectified by addressing people using public key cryptographic keys.
- \*
Remote data access and manipulation is made a first class citizen through the combination of content addressing and the secure global identification attained through using public-private cryptographic identities.
- \*
Software installation and management is made trivial through Fat Recursive AppDirs which are in turn made efficient by content addressing and first class access to remote data.
- \*
Truly on line backup and self service access to historical data, i.e., on line restoration of backups, and long term archiving of data are presented as natural consequences of a well developed content addressed file store.
- \*
Network congestion in the face of redundant data access is shown to be turned into cooperation by the reusable data packets in a content addressed network.
In addition the number of protocols and technologies required to implement the desirable features of the internet is dramatically reduced. Virus scanners, firewalls, backup and restore software, and network file protocols are among the many made redundant. Yet at the same time the inherent security is increased to a level appropriate given the hostile nature of the internet. Moreover, as is with any protocol, CANE could be tunnelled over the existing IPv4 to provide an feasible transition strategy.
This paper has described the concepts rather than the implementation of a content addressed network environment. Implementation is the next logical step, and is currently being pursued along with consideration of the problems surrounding the selection of an appropriate hash algorithm. Perhaps the largest challenge to be addressed is that of routing. By decentralising network addressing and introducing a vast non-geographically bound address space the task of routers is made difficult. Edge routing is relatively trivial, however backbone routers may require giga-bytes of memory and hardware accelerated cryptography in order to function effectively. This might be solved by leveraging the existing IPv4 infrastructure to provide not only carriage, but also “land marks” to facilitate efficient routing of the non-geographically bound CANE addresses. Collaborations are invited in this and any other area of the refinement and realisation of a content addressed network environment.
## References
- Agency \[1981\] Defense Advanced Research Projects Agency. Rfc 791 - internet protocol. *Internet Request For Comments*, 1981.
- E. James Whitehead and Goland \[1999\] Jr. E. James Whitehead and Yaron Goland. Webdav: a network protocol for remote collaborative authoring on the web. In *Proceedings of the Sixth European conference on Computer supported cooperative work*, pages 291–310, Norwell, MA, USA, 1999. Kluwer Academic Publishers. ISBN 0-7923-5948-8.
- E.Rescorla and Schiffman \[1996\] E.Rescorla and A. Schiffman. The secure hypertext transfer protocol. Technical report, 1996.
- Freier et al. \[1995\] Alan Freier, Philip Karlton, and Paul Kocher. The ssl protocol version 3. Technical report, 1995. URL <a href="http://home.netscape.com/eng/ssl3/ssl-toc.html" target="_blank">http://home.netscape.com/eng/ssl3/ssl-toc.html</a>.
- \[5\] Sara Gates. http://research.sun.com/minds/2005-1201/. Technical report, Sun Microsystems.
- Leach and Perry \[1996\] Paul Leach and Dan Perry. Cifs: A common internet file system. Technical report, Microsoft Corporation, 1996. URL <a href="http://www.microsoft.com/mind/1196/cifs.asp" target="_blank">http://www.microsoft.com/mind/1196/cifs.asp</a>.
- \[7\] Thomas Leonard. The zero install system. Technical report, http://zero-install.sourceforge.net/.
- \[8\] Klik Maintainers. Klik philosphical articles. Technical report, http://klik.atekon.de/docs/?page=Philosophical
- Microsystems \[1989\] Sun Microsystems. Rfc 1094 - nfs: Network file system protocol specification. *Internet Request for Comments*, 1989.
- Pike et al. \[1990\] R. Pike, D. Presotto, K. Thompson, and H. Trickey. Plan 9 from Bell Labs. *EUUG Newsletter*, 10(3):2–11, Autumn 1990. ISSN 1011-4211.
- Satyanarayanan et al. \[1990\] M. Satyanarayanan, James J. Kistler, Puneet Kumar, Maria E. Okasaki, Ellen H. Siegel, and David C. Steere. Coda: A highly available file system for a distributed workstation environment. *IEEE Transactions on Computers*, 39(4):447–459, 1990. URL [citeseer.ist.psu.edu/satyanarayanan90coda.html](citeseer.ist.psu.edu/satyanarayanan90coda.html).
- \[12\] IETF Secure Shell (secsh) Working Group. Secure shell (secsh) charter. Technical report, 1999. URL <a href="http://www.ietf.org/html.charters/secsh-charter.html" target="_blank">http://www.ietf.org/html.charters/secsh-charter.html</a>.
- Society \[1998\] The Internet Society. Rfc 2460 - internet protocol, version 6 (ipv6) specification. *Internet Request For Comments*, 1998.
- Stevens et al. \[2006\] Marc Stevens, Arjen Lenstra, and Benne de Weger. Target collisions for md5 and colliding x.509 certificates for different identities. Cryptology ePrint Archive, Report 2006/360, 2006. URL <a href="http://eprint.iacr.org/" target="_blank">http://eprint.iacr.org/</a>.
- Tridgell and Mackerras \[1998\] Andrew Tridgell and Paul Mackerras. The rsync algorithm. Technical report, Australian National University, 1998. URL <a href="http://rsync.samba.org/tech_report/" target="_blank">http://rsync.samba.org/tech_report/</a>. |
0709.2016 | cs.NI | cyber | # Distribution of PageRank Mass Among Principle Components of the Web
Konstantin Avrachenkov INRIA Sophia Antipolis, 2004, Route des Lucioles, 06902, France, E-mail: k.avrachenkov@sophia.inria.fr
Nelly Litvak University of Twente, Dept. of Applied Mathematics, P.O. Box 217, 7500AE Enschede, The Netherlands, E-mail: n.litvak@ewi.utwente.nl
Kim Son Pham St.Petersburg State University,35, University Prospect, 198504, Peterhof, St.Petersburg, Russia, E-mail: sonsecure@yahoo.com.sg
(September 2007)
###### Abstract
We study the PageRank mass of principal components in a bow-tie Web Graph, as a function of the damping factor $`c`$. Using a singular perturbation approach, we show that the PageRank share of IN and SCC components remains high even for very large values of the damping factor, in spite of the fact that it drops to zero when $`c\rightarrow 1`$. However, a detailed study of the OUT component reveals the presence “dead-ends” (small groups of pages linking only to each other) that receive an unfairly high ranking when $`c`$ is close to one. We argue that this problem can be mitigated by choosing $`c`$ as small as 1/2.
## 1 Introduction
The link-based ranking schemes such as PageRank \[[1](#bib.bib1)\], HITS \[[2](#bib.bib2)\], and SALSA \[[3](#bib.bib3)\] have been successfully used in search engines to provide adequate importance measures for Web pages. In the present work we restrict ourselves to the analysis of the PageRank criterion and use the following definition of PageRank from \[[4](#bib.bib4)\]. Denote by $`n`$ the total number of pages on the Web and define the $`n \times n`$ hyper-link matrix $`W`$ as follows:
<table>
<tbody>
<tr>
<td></td>
<td>wij={1/di,if page i links to j,1/n,if page i is dangling,0,otherwise,subscript𝑤𝑖𝑗cases1subscript𝑑𝑖if page i links to j1𝑛if page i is dangling0otherwisew_{ij}=\left\{\begin{array}[]{ll}1/d_{i},&\mbox{if page $i$ links to $j$},\\ 1/n,&\mbox{if page $i$ is dangling},\\ 0,&\mbox{otherwise},\end{array}\right.</td>
<td></td>
<td>(1)</td>
</tr>
</tbody>
</table>
for $`{{i,j} = 1},{\ldots,n}`$, where $`d_{i}`$ is the number of outgoing links from page $`i`$. A page is called dangling if it does not have outgoing links. The PageRank is defined as a stationary distribution of a Markov chain whose state space is the set of all Web pages, and the transition matrix is
<table>
<tbody>
<tr>
<td></td>
<td>G=cW+(1−c)(1/n)𝟏T𝟏.𝐺𝑐𝑊1𝑐1𝑛superscript1𝑇1G=cW+(1-c)(1/n){\bf 1}^{T}{\bf 1}.</td>
<td></td>
<td>(2)</td>
</tr>
</tbody>
</table>
Here and throughout the paper we use the symbol $`\mathbf{1}`$ for a column vector of ones having by default an appropriate dimension. In ([2](#S1.E2 "In 1 Introduction ‣ Distribution of PageRank Mass Among Principle Components of the Web")), $`\mathbf{1}^{T}\hspace{0pt}\mathbf{1}`$ is a matrix whose all entries are equal to one, and $`c \in {(0,1)}`$ is the parameter known as a damping factor. Let $`\pi`$ be the PageRank vector. Then by definition, $`{\pi\hspace{0pt}G} = \pi`$, and $`{\|\pi\|} = {\pi\hspace{0pt}\mathbf{1}} = 1`$, where we write $`\|\mathbf{x}\|`$ for the $`L_{1}`$-norm of vector $`\mathbf{x}`$.
The damping factor $`c`$ is a crucial parameter in the PageRank definition. It regulates the level of the uniform noise introduced to the system. Based on the publicly available information Google originally used $`c = 0.85`$, which appears to be a reasonable compromise between the true reflection of the Web structure and numerical efficiency (see \[[5](#bib.bib5)\] for more detail). However, it was mentioned in \[[6](#bib.bib6)\] that the value of $`c`$ too close to one results into distorted ranking of important pages. This phenomenon was also independently observed in \[[7](#bib.bib7)\]. Moreover, with smaller $`c`$, the PageRank is more rebust, that is, one can bound the influence of outgoing links of a page (or a small group of pages) on the PageRank of other groups \[[8](#bib.bib8)\] and on its own PageRank \[[7](#bib.bib7)\].
In this paper we explore the idea of relating the choice of $`c`$ to specific properties of the Web structure. In papers \[[9](#bib.bib9), [10](#bib.bib10)\] the authors have shown that the Web graph can be divided into three principle components. The Giant Strongly Connected Component (SCC) contains a large group of pages all having a hyper-link path to each other. The pages in the IN (OUT) component have a path to (from) the SCC, but not back. Furthermore, the SCC component is larger than the second largest strongly connected component by several orders of magnitude.
In Section [3](#S3 "3 The structure of the hyper-link transition matrix ‣ Distribution of PageRank Mass Among Principle Components of the Web") we consider a Markov walk governed by the hyperlink matrix $`W`$ and explicitly describe the limiting behavior of the PageRank vector as $`c\rightarrow 1`$. We experimentally study the OUT component in more detail to discover a so-called Pure OUT component (the OUT component without dangling nodes and their predecessors) and show that Pure OUT contains a number of small sub-SCC’s, or dead-ends, that absorb the total PageRank mass when $`c = 1`$. In Section [4](#S4 "4 PageRank mass of IN+SCC ‣ Distribution of PageRank Mass Among Principle Components of the Web") we apply the singular perturbation theory \[[11](#bib.bib11), [12](#bib.bib12), [13](#bib.bib13), [14](#bib.bib14)\] to analyze the shape of the PageRank of IN+SCC as a function of $`c`$. The dangling nodes turn out to play an unexpectedly important role in the qualitative behavior of this function. Our analytical and experimental results suggest that the PageRank mass of IN+SCC is sustained on a high level for quite large values of $`c`$, in spite of the fact that it drops to zero as $`c\rightarrow 1`$. Further, in Section [5](#S5 "5 PageRank mass of ESCC ‣ Distribution of PageRank Mass Among Principle Components of the Web") we show that the total PageRank mass of Pure OUT component increases with $`c`$. We argue that $`c = 0.85`$ results in an inadequately high ranking for Pure OUT pages and we present an argument for choosing $`c`$ as small as 1/2. We confirm our theoretical argument by experiments with log files. We would like to mention that the value $`c = {1/2}`$ was also used in \[[15](#bib.bib15)\] to find gems in scientific citations. This choice was justified intuitively by stating that researchers may check references in cited papers but on average they hardly go deeper than two levels. Nowadays, when search engines work really fast, this argument also applies to Web search. Indeed, it is easier for the user to refine a query and receive a proper page in fraction of seconds than to look for this page by clicking on hyper-links. Therefore, we may assume that a surfer searching for a page, on average, does not go deeper than two clicks.
The body of the paper contains main ideas and results. The necessary information from the perturbation theory and the proofs are given in Appendix.
## 2 Datasets
We have collected two Web graphs, which we denote by INRIA and FrMathInfo. The Web graph INRIA was taken from the site of INRIA, the French Research Institute of Informatics and Automatics. The seed for the INRIA collection was Web page www.inria.fr. It is a typical large Web site with around 300.000 pages and 2 millions hyper-links. We have collected all pages belonging to INRIA. The Web graph FrMathInfo was crawled with the initial seeds of 50 mathematics and informatics laboratories of France, taken from Google Directory. The crawl was executed by Breadth First Search of depth 6. The FrMathInfo Web graph contains around 700.000 pages and 8 millions hyper-links. Because of the fractal structure of the Web \[[16](#bib.bib16)\] we expect our datasets to be enough representative.
The link structure of the two Web graphs is stored in Oracle database. We could store the adjacency lists in RAM to speed up the computation of PageRank and other quantities of interest. This enables us to make more iterations, which is extremely important when the damping factor $`c`$ is close to one. Our PageRank computation program consumes about one hour to make 500 iterations for the FrMathInfo dataset and about half an hour for the INRIA dataset for the same number of iterations. Our algorithms for discovering the structures of the Web graph are based on Breadth First Search and Depth First Search methods, which are linear in the sum of number of nodes and links.
## 3 The structure of the hyper-link transition matrix
With the bow-tie Web structure \[[9](#bib.bib9), [10](#bib.bib10)\] in mind, we would like to analyze a stationary distribution of a Markov random walk governed by the hyper-link transition matrix $`W`$ given by ([1](#S1.E1 "In 1 Introduction ‣ Distribution of PageRank Mass Among Principle Components of the Web")). Such random walk follows an outgoing link chosen uniformly at random, and dangling nodes are assumed to have links to all pages in the Web. We note that the methods presented below can be easily extended to the case of personalized PageRank \[[17](#bib.bib17)\], when after a visit to a dangling node, the next page is sampled from some prescribed distribution.
Obviously, the graph induced by $`W`$ has a much higher connectivity than the original Web graph. In particular, if the random walk can move from a dangling node to an arbitrary node with the uniform distribution, then the Giant SCC component increases further in size. We refer to this new strongly connected component as the Extended Strongly Connected Component (ESCC). Due to the artificial links from the dangling nodes, the SCC component and IN component are now inter-connected and are parts of the ESCC. Furthermore, if there are dangling nodes in the OUT component, then these nodes together with all their predecessors become a part of the ESCC.
In the mini-example in Figure [2](#S3.F2 "Figure 2 ‣ 3 The structure of the hyper-link transition matrix ‣ Distribution of PageRank Mass Among Principle Components of the Web"), node 0 represents the IN component, nodes from 1 to 3 form the SCC component, and the rest of the nodes, nodes from 4 to 11, are in the OUT component. Node 5 is a dangling node, thus, artificial links go from the dangling node 5 to all other nodes. After addition of the artificial links, all nodes from 0 to 5 form the ESCC.
[Refer to caption](/html/0709.2016/assets/x1.png)
Figure 1: Example of a graph
<table>
<thead>
<tr>
<th>#</th>
<th>INRIA𝐼𝑁𝑅𝐼𝐴INRIA</th>
<th>FrMathInfo𝐹𝑟𝑀𝑎𝑡ℎ𝐼𝑛𝑓𝑜FrMathInfo</th>
</tr>
</thead>
<tbody>
<tr>
<th>total nodes</th>
<td>318585</td>
<td>764119</td>
</tr>
<tr>
<th>nodes in SCC</th>
<td>154142</td>
<td>333175</td>
</tr>
<tr>
<th>nodes in IN</th>
<td>0</td>
<td>0</td>
</tr>
<tr>
<th>nodes in OUT</th>
<td>164443</td>
<td>430944</td>
</tr>
<tr>
<th>nodes in ESCC</th>
<td>300682</td>
<td>760016</td>
</tr>
<tr>
<th>nodes in Pure OUT</th>
<td>17903</td>
<td>4103</td>
</tr>
<tr>
<th>SCCs in OUT</th>
<td>1148</td>
<td>1382</td>
</tr>
<tr>
<th>SCCs in Pure Out</th>
<td>631</td>
<td>379</td>
</tr>
</tbody>
</table>
Figure 2: Component sizes in INRIA and FrMathInfo datasets
In the Markov chain induced by the matrix $`W`$, all states from ESCC are transient, that is, with probability 1, the Markov chain eventually leaves this set of states and never returns back. The stationary probability of all these states is zero. The part of the OUT component without dangling nodes and their predecessors forms a block that we refer to as a Pure OUT component. In Figure [2](#S3.F2 "Figure 2 ‣ 3 The structure of the hyper-link transition matrix ‣ Distribution of PageRank Mass Among Principle Components of the Web") the Pure OUT component consists of nodes from 6 to 11. Typically, the Pure OUT component is much smaller than the Extended SCC. However, this is the set where the total stationary probability mass is concentrated. The sizes of all components for our two datasets are given in Figure [2](#S3.F2 "Figure 2 ‣ 3 The structure of the hyper-link transition matrix ‣ Distribution of PageRank Mass Among Principle Components of the Web"). Here the size of the IN components is zero because in the Web crawl we used the Breadth First Search method and we started from important pages in the Giant SCC. For the purposes of the present research it does not make any difference since we always consider IN and SCC together.
Let us now analyze the structure of the Pure OUT component in more detail. It turns out that inside Pure OUT there are many disjoint strongly connected components. All states in these sub-SCC’s (or, “dead-ends”) are recurrent, that is, the Markov chain started from any of these states always returns back to it. In particular, we have observed that there are many dead-ends of size 2 and 3. The Pure OUT component also contains transient states that eventually bring the random walk into one of the dead-ends. For simplicity, we add these states to the giant transient ESCC component.
Now, by appropriate renumbering of the states, we can refine the matrix $`W`$ by subdividing all states into one giant transient block and a number of small recurrent blocks as follows:
<table>
<tbody>
<tr>
<td></td>
<td>W=[Q100⋱0Qm0R1⋯RmT]dead-end (recurrent)⋯dead-end (recurrent)ESCC+[transient states in Pure OUT] (transient)𝑊delimited-[]subscript𝑄1missing-subexpression00missing-subexpression⋱missing-subexpressionmissing-subexpression0missing-subexpressionsubscript𝑄𝑚0subscript𝑅1⋯subscript𝑅𝑚𝑇dead-end (recurrent)⋯dead-end (recurrent)ESCC+[transient states in Pure OUT] (transient)W=\left[\begin{array}[]{cccc}Q_{1}&&0&0\\ &\ddots&&\\ 0&&Q_{m}&0\\ {R}_{1}&\cdots&{R}_{m}&{T}\end{array}\right]\;\begin{array}[]{l}\mbox{dead-end (recurrent)}\\ \cdots\phantom{\ddots}\\ \mbox{dead-end (recurrent)}\\ \mbox{ESCC+[transient states in Pure OUT] (transient)}\end{array}</td>
<td></td>
<td>(3)</td>
</tr>
</tbody>
</table>
Here for $`i = {1,\ldots,m}`$, a block $`Q_{i}`$ corresponds to transitions inside the $`i`$-th recurrent block, and a block $`R_{i}`$ contains transition probabilities from transient states to the $`i`$-th recurrent block. Block $`T`$ corresponds to transitions between the transient states. For instance, in example of the graph from Figure [2](#S3.F2 "Figure 2 ‣ 3 The structure of the hyper-link transition matrix ‣ Distribution of PageRank Mass Among Principle Components of the Web"), the nodes 8 and 9 correspond to block $`Q_{1}`$, nodes 10 and 11 correspond to block $`Q_{2}`$, and all other nodes belong to block $`T`$.
We would like to emphasis that the recurrent blocks here are really small, constituting altogether about 5% for INRIA and about 0.5% for FrMathInfo. We believe that for larger data sets, this percentage will be even less. By far most important part of the pages is contained in the ESCC, which constitutes the major part of the giant transient block.
Next, we note that if $`c < 1`$, then all states in the Markov chain induced by the Google matrix $`G`$ are recurrent, which automatically implies that they all have positive stationary probabilities. However, if $`c = 1`$, the majority of pages turn into transient states with stationary probability zero. Hence, the random walk governed by the Google transition matrix ([2](#S1.E2 "In 1 Introduction ‣ Distribution of PageRank Mass Among Principle Components of the Web")) is in fact a singularly perturbed Markov chain. Informally, by singular perturbation we mean relatively small changes in elements of the matrix, that lead to altered connectivity and stationary behavior of the chain. Using the results of the singular perturbation theory (see e.g., \[[11](#bib.bib11), [12](#bib.bib12), [13](#bib.bib13), [14](#bib.bib14)\]), in the next proposition we characterize explicitly the limiting PageRank vector as $`c\rightarrow 1`$ (see Appendix [A.2](#Sx2.SS2 "A.2 Proofs ‣ Appendix ‣ Distribution of PageRank Mass Among Principle Components of the Web") for the proof).
######
Proposition 1
Let $`\pi_{\text{OUT},i}`$ be a stationary distribution of the Markov chain governed by $`Q_{i}`$, $`i = {1,\ldots,m}`$. Then, we have
<table>
<tbody>
<tr>
<td></td>
<td>limc→1π(c)=[πOUT,1⋯πOUT,m 0],subscript→𝑐1𝜋𝑐delimited-[]subscript𝜋OUT1⋯subscript𝜋OUT𝑚 0\lim_{c\to 1}\pi(c)=\left[{\pi}_{{\mbox{\tiny OUT}},1}\ \cdots\ {\pi}_{{\mbox{\tiny OUT}},m}\ {\bf 0}\right],</td>
<td></td>
</tr>
</tbody>
</table>
where
<table>
<tbody>
<tr>
<td></td>
<td>πOUT,i=(# nodes in block Qin+1n𝟏T[I−T]−1Ri𝟏)π¯OUT,isubscript𝜋OUT𝑖# nodes in block Qi𝑛1𝑛superscript1𝑇superscriptdelimited-[]𝐼𝑇1subscript𝑅𝑖1subscript¯𝜋OUT𝑖{\pi}_{{\mbox{\tiny OUT}},i}=\left(\frac{\mbox{\# nodes in block $Q_{i}$}}{n}+\frac{1}{n}{\bf 1}^{T}[I-{T}]^{-1}{R}_{i}{\bf 1}\right)\bar{\pi}_{{\mbox{\tiny OUT}},i}</td>
<td></td>
<td>(4)</td>
</tr>
</tbody>
</table>
for $`i = {1,\ldots,m}`$, $`I`$ is the identity matrix, and $`\mathbf{0}`$ is a row vector of zeros that correspond to stationary probabilities of the states in the transient block.
The second term inside the brackets in formula ([4](#S3.E4 "In Proposition 1 ‣ 3 The structure of the hyper-link transition matrix ‣ Distribution of PageRank Mass Among Principle Components of the Web")) corresponds to the PageRank mass received by a dead-end from the Extended SCC. If $`c`$ is close to one, then this contribution can outweight by far the fair share of the PageRank, whereas the PageRank mass of the giant transient block decreases to zero. How large is the neighborhood of one where the ranking is skewed towards the Pure OUT? Is the value $`c = 0.85`$ already too large? We will address these questions in the remainder of the paper. In the next section we analyze the PageRank mass IN+SCC component, which is an important part of the transient block.
## 4 PageRank mass of IN+SCC
In Figure 3 we depict the PageRank mass of the giant component IN+SCC, as a function of the damping factor, for FrMathInfo.
[Refer to caption](/html/0709.2016/assets/x2.png)
Figure 3: The PageRank mass of IN+SCC as a function of $`c`$.
Here we see a typical behavior also observed for several pages in the mini-web from \[[6](#bib.bib6)\]: the PageRank first grows with $`c`$ and then decreases to zero. In our case, the PageRank mass of IN+SCC drops drastically starting from some value $`c`$ close to one. We can explain this phenomenon by highlighting the role of the dangling nodes.
We start the analysis by subdividing the Web graph sample into three subsets of nodes: IN+SCC, OUT, and the set of dangling nodes DN. We assume that no dangling node originates from OUT. This simplifies the derivation but does not change our conclusions. Then the Web hyper-link matrix $`W`$ in ([1](#S1.E1 "In 1 Introduction ‣ Distribution of PageRank Mass Among Principle Components of the Web")) can be written in the form
<table>
<tbody>
<tr>
<td></td>
<td>W=[Q00RPS1n𝟏𝟏T1n𝟏𝟏T1n𝟏𝟏T]OUTIN+SCC ,DN𝑊delimited-[]𝑄00𝑅𝑃𝑆1𝑛superscript11𝑇1𝑛superscript11𝑇1𝑛superscript11𝑇OUTIN+SCC ,DNW=\left[\begin{array}[]{ccc}Q&0&0\\ R&P&S\\ \frac{1}{n}{\bf 1}{\bf 1}^{T}&\frac{1}{n}{\bf 1}{\bf 1}^{T}&\frac{1}{n}{\bf 1}{\bf 1}^{T}\end{array}\right]\;\begin{array}[]{l}\mbox{OUT}\\ \mbox{IN+SCC\;\;,}\\ \mbox{DN}\end{array}</td>
<td></td>
</tr>
</tbody>
</table>
where the block $`Q`$ corresponds to the hyper-links inside the OUT component, the block $`R`$ corresponds to the hyper-links from IN+SCC to OUT, the block $`P`$ corresponds to the hyper-links inside the IN+SCC component, and the block $`S`$ corresponds to the hyper-links from SCC to dangling nodes. In the above, $`n`$ is the total number of pages in the Web graph sample, and the blocks $`\mathbf{1}\mathbf{1}^{T}`$ are the matrices of ones adjusted to appropriate dimensions.
Dividing the PageRank vector in segments corresponding to the blocks OUT, IN+SCC and DN,
<table>
<tbody>
<tr>
<td></td>
<td>π=[πOUTπIN+SCCπDN],𝜋delimited-[]subscript𝜋OUTsubscript𝜋IN+SCCsubscript𝜋DN\pi=[\pi_{\mbox{\tiny OUT}}\,\pi_{\mbox{\tiny IN+SCC}}\,\pi_{\mbox{\tiny DN}}],</td>
<td></td>
</tr>
</tbody>
</table>
we can rewrite the well-known formula (see e.g. \[[18](#bib.bib18)\])
<table>
<tbody>
<tr>
<td></td>
<td>π=1−cn𝟏T[I−cP]−1𝜋1𝑐𝑛superscript1𝑇superscriptdelimited-[]𝐼𝑐𝑃1\pi=\frac{1-c}{n}{\bf 1}^{T}[I-cP]^{-1}</td>
<td></td>
<td>(5)</td>
</tr>
</tbody>
</table>
as a system of three linear equations:
<table>
<tbody>
<tr>
<td></td>
<td>πOUT[I−cQ]−πIN+SCCcR−cnπDN𝟏𝟏T=1−cn𝟏T,subscript𝜋OUTdelimited-[]𝐼𝑐𝑄subscript𝜋IN+SCC𝑐𝑅𝑐𝑛subscript𝜋DNsuperscript11𝑇1𝑐𝑛superscript1𝑇\displaystyle\pi_{\mbox{\tiny OUT}}[I-cQ]-\pi_{\mbox{\tiny IN+SCC}}cR-\frac{c}{n}\pi_{\mbox{\tiny DN}}{\bf 1}{\bf 1}^{T}=\frac{1-c}{n}{\bf 1}^{T},</td>
<td></td>
<td>(6)</td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td>πIN+SCC[I−cP]−cnπDN𝟏𝟏T=1−cn𝟏T,subscript𝜋IN+SCCdelimited-[]𝐼𝑐𝑃𝑐𝑛subscript𝜋DNsuperscript11𝑇1𝑐𝑛superscript1𝑇\displaystyle\pi_{\mbox{\tiny IN+SCC}}[I-cP]-\frac{c}{n}\pi_{\mbox{\tiny DN}}{\bf 1}{\bf 1}^{T}=\frac{1-c}{n}{\bf 1}^{T},</td>
<td></td>
<td>(7)</td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td>−πIN+SCCcS+πDN−cnπDN𝟏𝟏T=1−cn𝟏T.subscript𝜋IN+SCC𝑐𝑆subscript𝜋DN𝑐𝑛subscript𝜋DNsuperscript11𝑇1𝑐𝑛superscript1𝑇\displaystyle-\pi_{\mbox{\tiny IN+SCC}}cS+\pi_{\mbox{\tiny DN}}-\frac{c}{n}\pi_{\mbox{\tiny DN}}{\bf 1}{\bf 1}^{T}=\frac{1-c}{n}{\bf 1}^{T}.</td>
<td></td>
<td>(8)</td>
</tr>
</tbody>
</table>
Solving ([6](#S4.E6 "In 4 PageRank mass of IN+SCC ‣ Distribution of PageRank Mass Among Principle Components of the Web")–[8](#S4.E8 "In 4 PageRank mass of IN+SCC ‣ Distribution of PageRank Mass Among Principle Components of the Web")) for $`\pi_{\text{IN+SCC}}`$ we obtain
<table>
<tbody>
<tr>
<td></td>
<td>πIN+SCC(c)=(1−c)α1−cβ𝐮IN+SCC[I−cP−c2α1−cβS𝟏𝐮IN+SCC]−1,subscript𝜋IN+SCC𝑐1𝑐𝛼1𝑐𝛽subscript𝐮IN+SCCsuperscriptdelimited-[]𝐼𝑐𝑃superscript𝑐2𝛼1𝑐𝛽𝑆1subscript𝐮IN+SCC1\pi_{\mbox{\tiny IN+SCC}}(c)=\frac{(1-c)\alpha}{1-c\beta}{{\bf u}_{\mbox{\tiny IN+SCC}}}\left[I-cP-\frac{c^{2}\alpha}{1-c\beta}S{\bf 1}{{\bf u}_{\mbox{\tiny IN+SCC}}}\right]^{-1},</td>
<td></td>
<td>(9)</td>
</tr>
</tbody>
</table>
where
<table>
<tbody>
<tr>
<td></td>
<td>α=|IN+SCC|/nandβ=|DN|/nformulae-sequence𝛼𝐼𝑁𝑆𝐶𝐶𝑛and𝛽𝐷𝑁𝑛\alpha=|IN+SCC|/{n}\quad\mbox{and}\quad\beta=|DN|/{n}</td>
<td></td>
</tr>
</tbody>
</table>
are the fractions of nodes in IN+SCC and DN, respectively, and $`\mathbf{u}_{\text{IN+SCC}} = {{|{{I\hspace{0pt}N} + {S\hspace{0pt}C\hspace{0pt}C}}|}^{- 1}\hspace{0pt}\mathbf{1}^{T}}`$ is a uniform probability row-vector of dimension $`|{{I\hspace{0pt}N} + {S\hspace{0pt}C\hspace{0pt}C}}|`$. The detailed derivation of ([9](#S4.E9 "In 4 PageRank mass of IN+SCC ‣ Distribution of PageRank Mass Among Principle Components of the Web")) can be found in Appendix [A.2](#Sx2.SS2 "A.2 Proofs ‣ Appendix ‣ Distribution of PageRank Mass Among Principle Components of the Web").
Now, define
<table>
<tbody>
<tr>
<td></td>
<td>k(c)=(1−c)α1−cβ,andU(c)=P+cα1−cβS𝟏𝐮IN+SCC.formulae-sequence𝑘𝑐1𝑐𝛼1𝑐𝛽and𝑈𝑐𝑃𝑐𝛼1𝑐𝛽𝑆1subscript𝐮IN+SCCk(c)=\frac{(1-c)\alpha}{1-c\beta},\quad\mbox{and}\quad U(c)=P+\frac{c\alpha}{1-c\beta}S{\bf 1}{{\bf u}_{\mbox{\tiny IN+SCC}}}.</td>
<td></td>
<td>(10)</td>
</tr>
</tbody>
</table>
Then the derivative of $`\pi_{\text{IN+SCC}}\hspace{0pt}{(c)}`$ with respect to $`c`$ is given by
<table>
<tbody>
<tr>
<td></td>
<td>πIN+SCC′(c)=𝐮IN+SCC{k′(c)I+k(c)[I−cU(c)]−1(cU(c))′}[I−cU(c)]−1,subscriptsuperscript𝜋′IN+SCC𝑐subscript𝐮IN+SCCsuperscript𝑘′𝑐𝐼𝑘𝑐superscriptdelimited-[]𝐼𝑐𝑈𝑐1superscript𝑐𝑈𝑐′superscriptdelimited-[]𝐼𝑐𝑈𝑐1\displaystyle\pi^{\prime}_{\mbox{\tiny IN+SCC}}(c)={{\bf u}_{\mbox{\tiny IN+SCC}}}\left\{k^{\prime}(c)I+k(c)[I-cU(c)]^{-1}(cU(c))^{\prime}\right\}[I-cU(c)]^{-1},</td>
<td></td>
<td>(11)</td>
</tr>
</tbody>
</table>
where using ([10](#S4.E10 "In 4 PageRank mass of IN+SCC ‣ Distribution of PageRank Mass Among Principle Components of the Web")) after simple calculations we get
<table>
<tbody>
<tr>
<td></td>
<td>k′(c)=−(1−β)α(1−cβ)2,(cU(c))′=U(c)+cα(1−cβ)2S𝟏𝐮IN+SCC.formulae-sequencesuperscript𝑘′𝑐1𝛽𝛼superscript1𝑐𝛽2superscript𝑐𝑈𝑐′𝑈𝑐𝑐𝛼superscript1𝑐𝛽2𝑆1subscript𝐮IN+SCCk^{\prime}(c)=-\frac{(1-\beta)\alpha}{(1-c\beta)^{2}},\quad(cU(c))^{\prime}=U(c)+\frac{c\alpha}{(1-c\beta)^{2}}S{\bf 1}{{\bf u}_{\mbox{\tiny IN+SCC}}}.</td>
<td></td>
</tr>
</tbody>
</table>
Let us consider the point $`c = 0`$. Using ([11](#S4.E11 "In 4 PageRank mass of IN+SCC ‣ Distribution of PageRank Mass Among Principle Components of the Web")), we obtain
<table>
<tbody>
<tr>
<td></td>
<td>πIN+SCC′(0)subscriptsuperscript𝜋′IN+SCC0\displaystyle\pi^{\prime}_{\mbox{\tiny IN+SCC}}(0)</td>
<td>=\displaystyle=</td>
<td>−α(1−β)𝐮IN+SCC+α𝐮IN+SCCP.𝛼1𝛽subscript𝐮IN+SCC𝛼subscript𝐮IN+SCC𝑃\displaystyle-\alpha(1-\beta){{\bf u}_{\mbox{\tiny IN+SCC}}}+\alpha{{\bf u}_{\mbox{\tiny IN+SCC}}}P.</td>
<td></td>
<td>(12)</td>
</tr>
</tbody>
</table>
One can see from the above equation that the PageRank of pages in IN+SCC with many incoming links will increase as $`c`$ increases from zero, which explains the graphs presented in \[[6](#bib.bib6)\].
Next, let us analyze the total mass of the IN+SCC component. From ([12](#S4.E12 "In 4 PageRank mass of IN+SCC ‣ Distribution of PageRank Mass Among Principle Components of the Web")) we obtain
<table>
<tbody>
<tr>
<td></td>
<td>‖πIN+SCC′(0)‖=−α(1−β)𝐮IN+SCC+α𝐮IN+SCCP𝟏=α(−1+β+p1),normsubscriptsuperscript𝜋′IN+SCC0𝛼1𝛽subscript𝐮IN+SCC𝛼subscript𝐮IN+SCC𝑃1𝛼1𝛽subscript𝑝1||\pi^{\prime}_{\mbox{\tiny IN+SCC}}(0)||=-\alpha(1-\beta){{\bf u}_{\mbox{\tiny IN+SCC}}}+\alpha{{\bf u}_{\mbox{\tiny IN+SCC}}}P{\bf 1}=\alpha(-1+\beta+p_{1}),</td>
<td></td>
</tr>
</tbody>
</table>
where $`p_{1} = {\mathbf{u}_{\text{IN+SCC}}\hspace{0pt}P\hspace{0pt}\mathbf{1}}`$ is the probability that a random walk on the hyperlink matrix stays in IN+SCC for one step if the initial distribution is uniform over IN+SCC. If $`{1 - \beta} < p_{1}`$ then the derivative at 0 is positive. Since dangling nodes typically constitute more than 25% of the graph \[[19](#bib.bib19)\], and $`p_{1}`$ is usually close to one, the condition $`{1 - \beta} < p_{1}`$ seems to be comfortably satisfied in Web samples. Thus, the total PageRank of the IN+SCC increases in $`c`$ when $`c`$ is small. Note by the way that if $`\beta = 0`$ then $`\|{\pi_{\text{IN+SCC}}\hspace{0pt}{(c)}}\|`$ is strictly decreasing in $`c`$. Hence, surprisingly, the presence of dangling nodes qualitatively changes the behavior of the IN+SCC PageRank mass.
Now let us consider the point $`c = 1`$. Again using ([11](#S4.E11 "In 4 PageRank mass of IN+SCC ‣ Distribution of PageRank Mass Among Principle Components of the Web")), we obtain
<table>
<tbody>
<tr>
<td></td>
<td>πIN+SCC′(1)=−α1−β𝐮IN+SCC[I−P−α1−βS𝟏𝐮IN+SCC]−1.subscriptsuperscript𝜋′IN+SCC1𝛼1𝛽subscript𝐮IN+SCCsuperscriptdelimited-[]𝐼𝑃𝛼1𝛽𝑆1subscript𝐮IN+SCC1\pi^{\prime}_{\mbox{\tiny IN+SCC}}(1)=-\frac{\alpha}{1-\beta}{{\bf u}_{\mbox{\tiny IN+SCC}}}[I-P-\frac{\alpha}{1-\beta}S{\bf 1}{{\bf u}_{\mbox{\tiny IN+SCC}}}]^{-1}.</td>
<td></td>
<td>(13)</td>
</tr>
</tbody>
</table>
Note that the matrix in the square braces is close to singular. Denote by $`\overline{P}`$ the hyper-link matrix of IN+SCC when the outer links are neglected. Then, $`\overline{P}`$ is an irreducible stochastic matrix. Denote its stationary distribution by $`{\overline{\pi}}_{\text{IN+SCC}}`$. Then we can apply Lemma [A.1](#Thmlemma1 "Lemma A.1 ‣ A.1 Results from Singular Perturbation Theory ‣ Appendix ‣ Distribution of PageRank Mass Among Principle Components of the Web") from the singular perturbation theory to ([13](#S4.E13 "In 4 PageRank mass of IN+SCC ‣ Distribution of PageRank Mass Among Principle Components of the Web")) by taking
<table>
<tbody>
<tr>
<td></td>
<td>A=P¯,εC=P¯−P−α1−βS𝟏𝐮IN+SCC,formulae-sequence𝐴¯𝑃𝜀𝐶¯𝑃𝑃𝛼1𝛽𝑆1subscript𝐮IN+SCCA=\bar{P},\quad\varepsilon C=\bar{P}-P-\frac{\alpha}{1-\beta}S{\bf 1}{{\bf u}_{\mbox{\tiny IN+SCC}}},</td>
<td></td>
</tr>
</tbody>
</table>
and noting that $`{{\varepsilon\hspace{0pt}C\hspace{0pt}\mathbf{1}} = {{R\hspace{0pt}\mathbf{1}} + {{({1 - \alpha - \beta})}\hspace{0pt}{({1 - \beta})}^{- 1}\hspace{0pt}S\hspace{0pt}\mathbf{1}}}}.`$ Combining all terms together and using $`{{\overline{\pi}}_{\text{IN+SCC}}\hspace{0pt}\mathbf{1}} = {\|{\overline{\pi}}_{\text{IN+SCC}}\|} = 1`$ and $`{\mathbf{u}_{\text{IN+SCC}}\hspace{0pt}\mathbf{1}} = {\|\mathbf{u}_{\text{IN+SCC}}\|} = 1`$, from ([A.1](#Sx2.E1 "In Lemma A.1 ‣ A.1 Results from Singular Perturbation Theory ‣ Appendix ‣ Distribution of PageRank Mass Among Principle Components of the Web")) we obtain
<table>
<tbody>
<tr>
<td></td>
<td>‖πIN+SCC′(1)‖normsubscriptsuperscript𝜋′IN+SCC1\displaystyle||\pi^{\prime}_{\mbox{\tiny IN+SCC}}(1)||</td>
<td>≈−α1−β1π¯IN+SCCR𝟏+1−β−α1−βπ¯IN+SCCS𝟏.absent𝛼1𝛽1subscript¯𝜋IN+SCC𝑅11𝛽𝛼1𝛽subscript¯𝜋IN+SCC𝑆1\displaystyle\approx-\frac{\alpha}{1-\beta}\frac{1}{\bar{\pi}_{\mbox{\tiny IN+SCC}}R{\bf 1}+\frac{1-\beta-\alpha}{1-\beta}\bar{\pi}_{\mbox{\tiny IN+SCC}}S{\bf 1}}.</td>
<td></td>
</tr>
</tbody>
</table>
It is expected that the value of $`{{\overline{\pi}}_{\text{IN+SCC}}\hspace{0pt}R\hspace{0pt}\mathbf{1}} + {\frac{1 - \beta - \alpha}{1 - \beta}\hspace{0pt}{\overline{\pi}}_{\text{IN+SCC}}\hspace{0pt}S\hspace{0pt}\mathbf{1}}`$ is typically small (indeed, in our dataset $`I\hspace{0pt}N\hspace{0pt}R\hspace{0pt}I\hspace{0pt}A`$, the value is 0.022), and hence the mass
$`\|{\pi_{\text{IN+SCC}}\hspace{0pt}{(c)}}\|`$ decreases very fast as $`c`$ approaches one.
Having described the behavior of the PageRank mass $`\|{\pi_{\text{IN+SCC}}\hspace{0pt}{(c)}}\|`$ at the boundary points $`c = 0`$ and $`c = 1`$, now we would like to show that there is at most one extremum on $`(0,1)`$. It is sufficient to prove that if $`{\|{\pi_{\text{IN+SCC}}'\hspace{0pt}{(c_{0})}}\|} \leq 0`$ for some $`c_{0} \in {(0,1)}`$ then $`{\|{\pi_{\text{IN+SCC}}'\hspace{0pt}{(c)}}\|} \leq 0`$ for all $`c > c_{0}`$. To this end, we apply the Sherman-Morrison formula to ([9](#S4.E9 "In 4 PageRank mass of IN+SCC ‣ Distribution of PageRank Mass Among Principle Components of the Web")), which yields
<table>
<tbody>
<tr>
<td></td>
<td>πIN+SCC(c)=π~IN+SCC(c)+c2α1−cβ𝐮IN+SCC[I−cP]−1S𝟏1+c2α1−cβ𝐮IN+SCC[I−cP]−1S𝟏π~IN+SCC(c),subscript𝜋IN+SCC𝑐subscript~𝜋IN+SCC𝑐superscript𝑐2𝛼1𝑐𝛽subscript𝐮IN+SCCsuperscriptdelimited-[]𝐼𝑐𝑃1𝑆11superscript𝑐2𝛼1𝑐𝛽subscript𝐮IN+SCCsuperscriptdelimited-[]𝐼𝑐𝑃1𝑆1subscript~𝜋IN+SCC𝑐\pi_{\mbox{\tiny IN+SCC}}(c)=\tilde{\pi}_{\mbox{\tiny IN+SCC}}(c)+\frac{\frac{c^{2}\alpha}{1-c\beta}{{\bf u}_{\mbox{\tiny IN+SCC}}}[I-cP]^{-1}S{\bf 1}}{1+\frac{c^{2}\alpha}{1-c\beta}{{\bf u}_{\mbox{\tiny IN+SCC}}}[I-cP]^{-1}S{\bf 1}}\,\tilde{\pi}_{\mbox{\tiny IN+SCC}}(c),</td>
<td></td>
<td>(14)</td>
</tr>
</tbody>
</table>
where
<table>
<tbody>
<tr>
<td></td>
<td>π~IN+SCC(c)=(1−c)α1−cβ𝐮IN+SCC[I−cP]−1.subscript~𝜋IN+SCC𝑐1𝑐𝛼1𝑐𝛽subscript𝐮IN+SCCsuperscriptdelimited-[]𝐼𝑐𝑃1\tilde{\pi}_{{\mbox{\tiny IN+SCC}}}(c)=\frac{(1-c)\alpha}{1-c\beta}{{\bf u}_{\mbox{\tiny IN+SCC}}}[I-cP]^{-1}.</td>
<td></td>
<td>(15)</td>
</tr>
</tbody>
</table>
represents the main term in the right-hand side of ([14](#S4.E14 "In 4 PageRank mass of IN+SCC ‣ Distribution of PageRank Mass Among Principle Components of the Web")). (The second summand in ([14](#S4.E14 "In 4 PageRank mass of IN+SCC ‣ Distribution of PageRank Mass Among Principle Components of the Web")) is about 10% of the total sum for the $`I\hspace{0pt}N\hspace{0pt}R\hspace{0pt}I\hspace{0pt}A`$ dataset for $`c = 0.85`$.) Now the behavior of $`\pi_{\text{IN+SCC}}\hspace{0pt}{(c)}`$ in Figure 3 can be explained by means of the next proposition (see Appendix [A.2](#Sx2.SS2 "A.2 Proofs ‣ Appendix ‣ Distribution of PageRank Mass Among Principle Components of the Web") for the proof).
######
Proposition 2
The term $`\|{{\overset{\sim}{\pi}}_{\text{IN+SCC}}\hspace{0pt}{(c)}}\|`$ given by ([15](#S4.E15 "In 4 PageRank mass of IN+SCC ‣ Distribution of PageRank Mass Among Principle Components of the Web")) has exactly one local maximum at some $`c_{0} \in {\lbrack 0,1\rbrack}`$. Moreover, $`{\|{{\overset{\sim}{\pi}}_{\text{IN+SCC}}^{\operatorname{\prime\prime}}\hspace{0pt}{(c)}}\|} < 0`$ for $`c \in {(c_{0},1\rbrack}`$.
We conclude that $`\|{{\overset{\sim}{\pi}}_{\text{IN+SCC}}\hspace{0pt}{(c)}}\|`$ is decreasing and concave for $`c \in {\lbrack c_{0},1\rbrack}`$, where $`{\|{{\overset{\sim}{\pi}}_{\text{IN+SCC}}'\hspace{0pt}{(c_{0})}}\|} = 0`$. This is exactly the behavior we observe in the experiments. The analysis and experiments suggest that $`c_{0}`$ is definitely larger than 0.85 and actually is quite close to one. Thus, one may want to choose large $`c`$ in order to maximize the PageRank mass of IN+SCC. However, in the next section we will indicate important drawbacks of this choice.
## 5 PageRank mass of ESCC
Let us now consider the PageRank mass of the Extended SCC component (ESCC) described in Section [3](#S3 "3 The structure of the hyper-link transition matrix ‣ Distribution of PageRank Mass Among Principle Components of the Web"), as a function of $`c \in {\lbrack 0,1\rbrack}`$. Subdividing the PageRank vector in the blocks $`\pi = {\lbrack{\pi_{\text{PureOUT}}\hspace{0pt}\pi_{\text{ESCC}}}\rbrack}`$, from ([5](#S4.E5 "In 4 PageRank mass of IN+SCC ‣ Distribution of PageRank Mass Among Principle Components of the Web")) we obtain
<table>
<tbody>
<tr>
<td></td>
<td>‖πESCC(c)‖normsubscript𝜋ESCC𝑐\displaystyle||\pi_{\mbox{\tiny ESCC}}(c)||</td>
<td>=(1−c)γ𝐮ESCC[I−cT]−1𝟏,absent1𝑐𝛾subscript𝐮ESCCsuperscriptdelimited-[]𝐼𝑐𝑇11\displaystyle=(1-c)\gamma{{\bf u}_{\mbox{\tiny ESCC}}}[I-cT]^{-1}{\bf 1},</td>
<td></td>
<td>(16)</td>
</tr>
</tbody>
</table>
where $`T`$ represents the transition probabilitites inside the ESCC block, $`\gamma = {{|{E\hspace{0pt}S\hspace{0pt}C\hspace{0pt}C}|}/n}`$, and $`\mathbf{u}_{\text{ESCC}}`$ is a uniform probability row-vector over ESCC. Clearly, we have $`{\|{\pi_{\text{ESCC}}\hspace{0pt}{(0)}}\|} = \gamma`$ and $`{\|{\pi_{\text{ESCC}}\hspace{0pt}{(1)}}\|} = 0`$. Furthermore, by taking derivatives we easily show that $`\|{\pi_{\text{ESCC}}\hspace{0pt}{(c)}}\|`$ is a concave decreasing function. In the next proposition (proved in the Appendix), we derive a series of bounds for $`\|{\pi_{\text{ESCC}}\hspace{0pt}{(c)}}\|`$.
######
Proposition 3
Let $`\lambda_{1}`$ be the Perron-Frobenius eigenvalue of $`T`$, and let $`p_{1} = {\mathbf{u}_{\text{ESCC}}\hspace{0pt}T\hspace{0pt}\mathbf{1}}`$ be the probability that the random walk started from a randomly chosen state in ESCC, stays in ESCC for one step.
- \(i\)
If $`p_{1} < \lambda_{1}`$ then
<table>
<tbody>
<tr>
<td></td>
<td>‖πESCC(c)‖<γ(1−c)1−cλ1,c∈(0,1).formulae-sequencenormsubscript𝜋ESCC𝑐𝛾1𝑐1𝑐subscript𝜆1𝑐01||\pi_{\mbox{\tiny ESCC}}(c)||<\frac{\gamma(1-c)}{1-c\lambda_{1}},\quad c\in(0,1).</td>
<td></td>
<td>(17)</td>
</tr>
</tbody>
</table>
- \(ii\)
If $`{1/{({1 - p_{1}})}} < {\mathbf{u}_{\text{ESCC}}\hspace{0pt}{\lbrack{I - T}\rbrack}^{- 1}\hspace{0pt}\mathbf{1}}`$ then
<table>
<tbody>
<tr>
<td></td>
<td>‖πESCC(c)‖>γ(1−c)1−cp1,c∈(0,1).formulae-sequencenormsubscript𝜋ESCC𝑐𝛾1𝑐1𝑐subscript𝑝1𝑐01||\pi_{\mbox{\tiny ESCC}}(c)||>\frac{\gamma(1-c)}{1-cp_{1}},\quad c\in(0,1).</td>
<td></td>
<td>(18)</td>
</tr>
</tbody>
</table>
The condition $`p_{1} < \lambda_{1}`$ has a clear intuitive interpretation. Let $`{\hat{\pi}}_{\text{ESCC}}`$ be the probability-normed left Perron-Frobenius eigenvector of $`T`$. Then $`{\hat{\pi}}_{\text{ESCC}}`$, also known as a quasi-stationary distribution of $`T`$, is the limiting probability distribution of the Markov chain given that the random walk never leaves the block $`T`$ (see e.g. \[[20](#bib.bib20)\]). Since $`{{\hat{\pi}}_{\text{ESCC}}\hspace{0pt}T} = \lambda_{1}`$, the condition $`p_{1} < \lambda_{1}`$ means that the chance to stay in ESCC for one step in the quasi-stationary regime is higher than starting from the uniform distribution $`\mathbf{u}_{\text{ESCC}}`$. Although $`p_{1} < \lambda_{1}`$ does not hold in general, one may expect that it should hold for transition matrices describing large entangled graphs since quasi-stationary distribution should favor states, from which the chance to leave ESCC is lower.
Both conditions of Proposition [3](#Thmproposition3 "Proposition 3 ‣ 5 PageRank mass of ESCC ‣ Distribution of PageRank Mass Among Principle Components of the Web") are satisfied in our experiments. With the help of the derived bounds we conclude that $`\|{\pi_{\text{ESCC}}\hspace{0pt}{(c)}}\|`$ decreases very slowly for small and moderate values of $`c`$, and it decreases extremely fast when $`c`$ becomes close to 1. This typical behavior is clearly seen in Figure [4](#S5.F4 "Figure 4 ‣ 5 PageRank mass of ESCC ‣ Distribution of PageRank Mass Among Principle Components of the Web"), where $`\|{\pi_{\text{ESCC}}\hspace{0pt}{(c)}}\|`$ is plotted with a solid line.
[Refer to caption](/html/0709.2016/assets/x3.png)
[Refer to caption](/html/0709.2016/assets/x4.png)
Figure 4: PageRank mass of ESCC and bounds, INRIA (left) and FrMathInfo (right)
The bounds are plotted in Figure [4](#S5.F4 "Figure 4 ‣ 5 PageRank mass of ESCC ‣ Distribution of PageRank Mass Among Principle Components of the Web") with dashed lines. For the INRIA dataset we have $`p_{1} = 0.97557`$, $`\lambda_{1} = 0.99954`$, and for the FrMathInfo dataset we have $`p_{1} = 0.99659`$, $`\lambda_{1} = 0.99937`$.
From the above we conclude that the PageRank mass of ESCC is smaller than $`\gamma`$ for any value $`c > 0`$. On contrary, the PageRank mass of Pure OUT increases in $`c`$ beyond its “fair share” $`\delta = {{|{P\hspace{0pt}u\hspace{0pt}r\hspace{0pt}e\hspace{0pt}O\hspace{0pt}U\hspace{0pt}T}|}/n}`$. With $`c = 0.85`$, the PageRank mass of the Pure OUT component in the INRIA dataset is equal to $`1.95\hspace{0pt}\delta`$. In the FrMathInfo dataset, the unfairness is even more pronounced: the PageRank mass of the Pure OUT component is equal to $`3.44\hspace{0pt}\delta`$. This gives users an incentive to create dead-ends: groups of pages that link only to each other. Clearly, this can be mitigated by choosing a smaller damping factor. Below we propose one way to determine an “optimal” value of $`c`$.
Let $`\mathbf{v}`$ be some probability vector over ESCC. We would like to choose $`c = c^{\ast}`$ that satisfies the condition
<table>
<tbody>
<tr>
<td></td>
<td>‖πESCC(c)‖=‖𝐯T‖,normsubscript𝜋ESCC𝑐norm𝐯𝑇||\pi_{\mbox{\tiny ESCC}}(c)||=||{\bf v}T||,</td>
<td></td>
<td>(19)</td>
</tr>
</tbody>
</table>
that is, starting from $`\mathbf{v}`$, the probability mass preserved in ESCC after one step should be equal to the PageRank of ESCC. One can think for instance of the following three reasonable choices of $`\mathbf{v}`$: 1) $`{\hat{\pi}}_{T}`$, the quasi-stationary distribution of $`T`$, 2) the uniform vector $`\mathbf{u}_{\text{ESCC}}`$, and 3) the normalized PageRank vector $`{\pi_{\text{ESCC}}\hspace{0pt}{(c)}}/{\|{\pi_{\text{ESCC}}\hspace{0pt}{(c)}}\|}`$. The first choice reflects the proximity of $`T`$ to a stochastic matrix. The second choice is inspired by definition of PageRank (restart from uniform distribution), and the third choice combines both these features.
If conditions of Proposition [3](#Thmproposition3 "Proposition 3 ‣ 5 PageRank mass of ESCC ‣ Distribution of PageRank Mass Among Principle Components of the Web") are satisfied, then ([17](#S5.E17 "In item (i) ‣ Proposition 3 ‣ 5 PageRank mass of ESCC ‣ Distribution of PageRank Mass Among Principle Components of the Web")) and ([18](#S5.E18 "In item (ii) ‣ Proposition 3 ‣ 5 PageRank mass of ESCC ‣ Distribution of PageRank Mass Among Principle Components of the Web")) hold, and thus the value of $`c^{\ast}`$ satisfying ([19](#S5.E19 "In 5 PageRank mass of ESCC ‣ Distribution of PageRank Mass Among Principle Components of the Web")) must be in the interval $`(c_{1},c_{2})`$, where
<table>
<tbody>
<tr>
<td></td>
<td>(1−c1)/(1−p1c1)=‖𝐯T‖,(1−c2)/(1−λ1c2)=‖𝐯T‖.formulae-sequence1subscript𝑐11subscript𝑝1subscript𝑐1norm𝐯𝑇1subscript𝑐21subscript𝜆1subscript𝑐2norm𝐯𝑇(1-c_{1})/(1-p_{1}c_{1})=||{\bf v}T||,\quad(1-c_{2})/(1-\lambda_{1}c_{2})=||{\bf v}T||.</td>
<td></td>
</tr>
</tbody>
</table>
Numerical results for all three choices of $`\mathbf{v}`$ are presented in Table [1](#S5.T1 "Table 1 ‣ 5 PageRank mass of ESCC ‣ Distribution of PageRank Mass Among Principle Components of the Web").
$`v`$ $`c`$ INRIA FrMathInfo
$`{\hat{\pi}}_{\text{ESCC}}`$ $`c_{1}`$ 0.0184 0.1956
$`c_{2}`$ 0.5001 0.5002
$`c^{\ast}`$ .02 .16
$`\mathbf{u}_{\text{ESCC}}`$ $`c_{1}`$ 0.5062 0.5009
$`c_{2}`$ 0.9820 0.8051
$`c^{\ast}`$ .604 .535
$`\pi_{\text{ESCC}}/{\|\pi_{\text{ESCC}}\|}`$ $`1/{({1 + \lambda_{1}})}`$ 0.5001 0.5002
$`1/{({1 + p_{1}})}`$ 0.5062 0.5009
Table 1: Values of $`c^{\ast}`$ with bounds.
If $`\mathbf{v} = {\hat{\pi}}_{\text{ESCC}}`$ then we have $`{\|{\mathbf{v}\hspace{0pt}T}\|} = \lambda_{1}`$, which implies $`c_{1} = {{({1 - \lambda_{1}})}/{({1 - {\lambda_{1}\hspace{0pt}p_{1}}})}}`$ and $`c_{2} = {1/{({\lambda_{1} + 1})}}`$. In this case, the upper bound $`c_{2}`$ is only slightly larger than $`1/2`$ and $`c^{\ast}`$ is close to zero in our data sets (see Tabel [1](#S5.T1 "Table 1 ‣ 5 PageRank mass of ESCC ‣ Distribution of PageRank Mass Among Principle Components of the Web")). Such small $`c`$ however leads to ranking that takes into account only local information about the Web graph (see e.g. \[[21](#bib.bib21)\]). The choice $`\mathbf{v} = {\hat{\pi}}_{\text{ESCC}}`$ does not seem to represent the dynamics of the system; probably because the “easily bored surfer” random walk that is used in PageRank computations never follows a quasi-stationary distribution since it often restarts itself from the uniform probability vector.
For the uniform vector $`\mathbf{v} = \mathbf{u}_{\text{ESCC}}`$, we have $`{\|{\mathbf{v}\hspace{0pt}T}\|} = p_{1}`$, which gives $`c_{1},c_{2},c^{\ast}`$ presented in Table [1](#S5.T1 "Table 1 ‣ 5 PageRank mass of ESCC ‣ Distribution of PageRank Mass Among Principle Components of the Web"). We have obtained a higher upper bound but the values of $`c^{\ast}`$ are still much smaller than $`0.85`$.
Finally, for the normalized PageRank vector $`\mathbf{v} = {\pi_{\text{ESCC}}/{\|\pi_{\text{ESCC}}\|}}`$, using ([16](#S5.E16 "In 5 PageRank mass of ESCC ‣ Distribution of PageRank Mass Among Principle Components of the Web")), we rewrite ([19](#S5.E19 "In 5 PageRank mass of ESCC ‣ Distribution of PageRank Mass Among Principle Components of the Web")) as
<table>
<tbody>
<tr>
<td></td>
<td>‖πESCC(c)‖normsubscript𝜋ESCC𝑐\displaystyle||\pi_{\mbox{\tiny ESCC}}(c)||</td>
<td>=γ‖πESCC(c)‖πESCC(c)T𝟏=γ2(1−c)‖πESCC(c)‖𝐮IN+SCC[I−cT]−1T𝟏,absent𝛾normsubscript𝜋ESCC𝑐subscript𝜋ESCC𝑐𝑇1superscript𝛾21𝑐normsubscript𝜋ESCC𝑐subscript𝐮IN+SCCsuperscriptdelimited-[]𝐼𝑐𝑇1𝑇1\displaystyle=\frac{\gamma}{||\pi_{\mbox{\tiny ESCC}}(c)||}\pi_{\mbox{\tiny ESCC}}(c)T{\bf 1}=\frac{\gamma^{2}(1-c)}{||\pi_{\mbox{\tiny ESCC}}(c)||}{{{\bf u}_{\mbox{\tiny IN+SCC}}}}[I-cT]^{-1}T{\bf 1},</td>
<td></td>
</tr>
</tbody>
</table>
Multiplying by $`\|{\pi_{\text{ESCC}}\hspace{0pt}{(c)}}\|`$, after some algebra we obtain
<table>
<tbody>
<tr>
<td></td>
<td>‖πESCC(c)‖2superscriptnormsubscript𝜋ESCC𝑐2\displaystyle||\pi_{\mbox{\tiny ESCC}}(c)||^{2}</td>
<td>=γc‖πESCC(c)‖−(1−c)γ2c.absent𝛾𝑐normsubscript𝜋ESCC𝑐1𝑐superscript𝛾2𝑐\displaystyle=\frac{\gamma}{c}\,||\pi_{\mbox{\tiny ESCC}}(c)||-\frac{(1-c)\gamma^{2}}{c}.</td>
<td></td>
<td></td>
</tr>
</tbody>
</table>
Solving the quadratic equation for $`\|{\pi_{\text{ESCC}}\hspace{0pt}{(c)}}\|`$, we get
<table>
<tbody>
<tr>
<td></td>
<td>‖πESCC(c)‖=r(c)={γif c≤1/2,γ(1−c)cif c>1/2.normsubscript𝜋ESCC𝑐𝑟𝑐cases𝛾if 𝑐12𝛾1𝑐𝑐if 𝑐12||\pi_{\mbox{\tiny ESCC}}(c)||=r(c)=\left\{\begin{array}[]{ll}\gamma&\mbox{if }c\leq 1/2,\\ \frac{\gamma(1-c)}{c}&\mbox{if }c>1/2.\end{array}\right.</td>
<td></td>
</tr>
</tbody>
</table>
Hence, the value $`c^{\ast}`$ solving ([19](#S5.E19 "In 5 PageRank mass of ESCC ‣ Distribution of PageRank Mass Among Principle Components of the Web")) corresponds to the point where the graphs of $`\|{\pi_{\text{ESCC}}\hspace{0pt}{(c)}}\|`$ and $`r\hspace{0pt}{(c)}`$ cross each other. There is only one such point on (0,1), and since $`\|{\pi_{\text{ESCC}}\hspace{0pt}{(c)}}\|`$ decreases very slowly unless $`c`$ is close to one, whereas $`r\hspace{0pt}{(c)}`$ decreases relatively fast for $`c > {1/2}`$, we expect that $`c^{\ast}`$ is only slightly larger than $`1/2`$. Under conditions of Proposition [3](#Thmproposition3 "Proposition 3 ‣ 5 PageRank mass of ESCC ‣ Distribution of PageRank Mass Among Principle Components of the Web"), $`r\hspace{0pt}{(c)}`$ first crosses the line $`{\gamma\hspace{0pt}{({1 - c})}}/{({1 - {\lambda_{1}\hspace{0pt}c}})}`$, then $`{\|{\pi_{T}\hspace{0pt}{(c)}}\|}_{1}`$, and then $`{\gamma\hspace{0pt}{({1 - c})}}/{({1 - {p_{1}\hspace{0pt}c}})}`$. Thus, we yield $`{({1 + \lambda_{1}})}^{- 1} < c^{\ast} < {({1 + p_{1}})}^{- 1}`$. Since both $`\lambda_{1}`$ and $`p_{1}`$ are large, this suggests that $`c`$ should be chosen around 1/2. This is also reflected in Tabel [1](#S5.T1 "Table 1 ‣ 5 PageRank mass of ESCC ‣ Distribution of PageRank Mass Among Principle Components of the Web").
Last but not least, to support our theoretical argument about the undeserved high ranking of pages from Pure OUT, we carry out the following experiment. In the $`I\hspace{0pt}N\hspace{0pt}R\hspace{0pt}I\hspace{0pt}A`$ dataset we have chosen an absorbing component in Pure OUT consisting just of two nodes. We have added an artificial link from one of these nodes to a node in the Giant SCC and recomputed the PageRank. In Table [2](#S5.T2 "Table 2 ‣ 5 PageRank mass of ESCC ‣ Distribution of PageRank Mass Among Principle Components of the Web") in the column “PR rank w/o link” we give a ranking of a page according to the PageRank value computed before the addition of the artificial link and in the column “PR rank with link” we give a ranking of a page according to the PageRank value computed after the addition of the artificial link. We have also analyzed the log file of the site INRIA Sophia Antipolis (www-sop.inria.fr) and ranked the pages according to the number of clicks for the period of one year up to May 2007. We note that since we have the access only to the log file of the INRIA Sophia Antipolis site, we use the PageRank ranking also only for the pages from the INRIA Sophia Antipolis site. For instance, for $`c = 0.85`$, the ranking of Page A without an artificial link is 731 (this means that 731 pages are ranked better than Page A among the pages of INRIA Sophia Antipolis). However, its ranking according to the number of clicks is much lower, 2588. This confirms our conjecture that the nodes in Pure OUT obtain unjustifiably high ranking. Next we note that the addition of an artificial link significantly diminishes the ranking. In fact, it brings it close to the ranking provided by the number of clicks. Finally, we draw the attention of the reader to the fact that choosing $`c = {1/2}`$ also significantly reduces the gap between the ranking by PageRank and the ranking by the number of clicks.
$`c`$ PR rank w/o link PR rank with link rank by no. of clicks
Node A
0.5 1648 2307 2588
0.85 731 2101 2588
0.95 226 2116 2588
Node B
0.5 1648 4009 3649
0.85 731 3279 3649
0.95 226 3563 3649
Table 2: Comparison between PR and click based rankings.
To summarize, our results indicate that with $`c = 0.85`$, the Pure OUT component receives an unfairly large share of the PageRank mass. Remarkably, in order to satisfy any of the three intuitive criteria of fairness presented above, the value of $`c`$ should be drastically reduced. The experiment with the log files confirms the same. Of course, a drastic reduction of $`c`$ also considerably accelerates the computation of PageRank by numerical methods \[[22](#bib.bib22), [5](#bib.bib5), [23](#bib.bib23)\].
## Acknowledgments
This work is supported by EGIDE ECO-NET grant no. 10191XC and by NWO Meervoud grant no. 632.002.401.
## References
- \[1\] Page, L., Brin, S., Motwani, R., Winograd, T.: The PageRank citation ranking: Bringing order to the web. Technical report, Stanford University (1998)
- \[2\] Kleinberg, J.M.: Authoritative sources in a hyperlinked environment.
Journal of the ACM 46(5) (1999) 604–632
- \[3\] Lempel, R., Moran, S.: The stochastic approach for link-structure analysis (SALSA) and the TKC effect.
Comput. Networks 33(1–6) (2000) 387–401
- \[4\] Langville, A.N., Meyer, C.D.: Deeper inside PageRank.
Internet Math. 1 (2003) 335–380
- \[5\] Langville, A.N., Meyer, C.D.: Google’s PageRank and beyond: the science of search engine rankings. Princeton University Press, Princeton, NJ (2006)
- \[6\] Boldi, P., Santini, M., Vigna, S.: PageRank as a function of the damping factor. In: Proc. of the Fourteenth International World Wide Web Conference, Chiba, Japan, ACM Press (2005)
- \[7\] Avrachenkov, K., Litvak, N.: The effect of new links on Google PageRank.
Stoch. Models 22(2) (2006) 319–331
- \[8\] Bianchini, M., Gori, M., Scarselli, F.: Inside PageRank.
ACM Trans. Inter. Tech. 5(1) (2005) 92–128
- \[9\] Broder, A., Kumar, R., Maghoul, F., Raghavan, P., Rajagopalan, S., Statac, R., Tomkins, A., J.Wiener: Graph structure in the Web.
Computer Networks 33 (2000) 309–320
- \[10\] Kumar, R., Raghavan, P., Rajagopalan, S., Sivakumar, D., Tomkins, A., Upfal, E.: The Web as a graph. In: Proc. 19th ACM SIGACT-SIGMOD-AIGART Symp. Principles of Database Systems, PODS, ACM Press (15–17 2000) 1–10
- \[11\] Avrachenkov, K.: Analytic Perturbation Theory and its Applications. PhD thesis, University of South Australia (1999)
- \[12\] Korolyuk, V.S., Turbin, A.F.: Mathematical foundations of the state lumping of large systems. Volume 264 of Mathematics and its Applications. Kluwer Academic Publishers Group, Dordrecht (1993)
- \[13\] Pervozvanskii, A.A., Gaitsgori, V.G.: Theory of suboptimal decisions. Volume 12 of Mathematics and its Applications (Soviet Series). Kluwer Academic Publishers Group, Dordrecht (1988)
- \[14\] Yin, G.G., Zhang, Q.: Discrete-time Markov chains. Volume 55 of Applications of Mathematics (New York). Springer-Verlag, New York (2005)
- \[15\] Chen, P., Xie, H., Maslov, S., Redner, S.: Finding scientific gems with google. Arxiv preprint physics/0604130 (2006)
- \[16\] Dill, S., Kumar, R., Mccurley, K.S., Rajagopalan, S., Sivakumar, D., Tomkins, A.: Self-similarity in the web.
ACM Trans. Inter. Tech. 2(3) (2002) 205–223
- \[17\] Haveliwala, T.: Topic-sensitive PageRank: A context-sensitive ranking algorithm for Web search.
IEEE Transactions on Knowledge and Data Engineering 15(4) (2003) 784–796
- \[18\] Moler, C., Moler, K.: Numerical Computing with MATLAB. SIAM (2003)
- \[19\] Eiron, N., McCurley, K., Tomlin, J.: Ranking the Web frontier. In: WWW ’04: Proceedings of the 13th international conference on World Wide Web, New York, NY, USA, ACM Press (2004) 309–318
- \[20\] Seneta, E.: Non-negative matrices and Markov chains. Springer Series in Statistics. Springer, New York (2006) Revised reprint of the second (1981) edition \[Springer-Verlag, New York; MR0719544\].
- \[21\] Fortunato, S., Flammini, A.: Random walks on directed networks: the case of PageRank. Technical Report 0604203, arXiv/physics (2006)
- \[22\] Avrachenkov, K., Litvak, N., Nemirovsky, D., Osipova, N.: Monte Carlo methods in PageRank computation: When one iteration is sufficient. SIAM J. Numer. Anal. (2007)
- \[23\] Berkhin, P.: A survey on PageRank computing.
Internet Math. 2 (2005) 73–120
## Appendix
### A.1 Results from Singular Perturbation Theory
######
Lemma A.1
Let $`{A\hspace{0pt}{(\varepsilon)}} = {A - {\varepsilon\hspace{0pt}C}}`$ be a perturbation of irreducible stochastic matrix $`A`$ such that $`A\hspace{0pt}{(\varepsilon)}`$ is substochastic. Then, for sufficiently small $`\varepsilon`$ the following Laurent series expansion holds
<table>
<tbody>
<tr>
<td></td>
<td>[I−A(ε)]−1=1εX−1+X0+εX1+…,superscriptdelimited-[]𝐼𝐴𝜀11𝜀subscript𝑋1subscript𝑋0𝜀subscript𝑋1…[I-A(\varepsilon)]^{-1}=\frac{1}{\varepsilon}X_{-1}+X_{0}+\varepsilon X_{1}+...\ ,</td>
<td></td>
</tr>
</tbody>
</table>
with
<table>
<tbody>
<tr>
<td></td>
<td>X−1=1μC𝟏𝟏μ,subscript𝑋11𝜇𝐶11𝜇X_{-1}=\frac{1}{\mu C{\bf 1}}{\bf 1}\mu,</td>
<td></td>
</tr>
</tbody>
</table>
where $`\mu`$ is the stationary distribution of $`A`$. It follows that
<table>
<tbody>
<tr>
<td></td>
<td>[I−A(ε)]−1=1μεC𝟏𝟏μ+O(1)as ε→0.formulae-sequencesuperscriptdelimited-[]𝐼𝐴𝜀11𝜇𝜀𝐶11𝜇O1→as 𝜀0[I-A(\varepsilon)]^{-1}=\frac{1}{\mu\varepsilon C{\bf 1}}{\bf 1}\mu+\mbox{O}(1)\quad\mbox{as }\varepsilon\to 0.</td>
<td></td>
<td>(A.1)</td>
</tr>
</tbody>
</table>
######
Lemma A.2
Let $`{A\hspace{0pt}{(\varepsilon)}} = {A + {\varepsilon\hspace{0pt}C}}`$ be a transition matrix of perturbed Markov chain.
The perturbed Markov chain is assumed to be ergodic for sufficiently small $`\varepsilon`$ different from zero. Let the unperturbed Markov chain $`({\varepsilon = 0})`$ have $`m`$ ergodic classes. Namely, the transition matrix $`A`$ can be written in the form
<table>
<tbody>
<tr>
<td></td>
<td>A=[A100⋱0Am0L1⋯LmE]∈ℝn×n.𝐴delimited-[]subscript𝐴1missing-subexpression00missing-subexpression⋱missing-subexpressionmissing-subexpression0missing-subexpressionsubscript𝐴𝑚0subscript𝐿1⋯subscript𝐿𝑚𝐸superscriptℝ𝑛𝑛A=\left[\begin{array}[]{cccc}A_{1}&&0&0\\ &\ddots&&\\ 0&&A_{m}&0\\ L_{1}&\cdots&L_{m}&E\end{array}\right]\in\mathbb{R}^{n\times n}.</td>
<td></td>
</tr>
</tbody>
</table>
Then, the stationary distribution of the perturbed Markov chain has a limit
<table>
<tbody>
<tr>
<td></td>
<td>limε→0π(ε)=[ν1μ1⋯νmμm 0],subscript→𝜀0𝜋𝜀delimited-[]subscript𝜈1subscript𝜇1⋯subscript𝜈𝑚subscript𝜇𝑚 0\lim_{\varepsilon\to 0}\pi(\varepsilon)=[\nu_{1}\mu_{1}\ \cdots\ \nu_{m}\mu_{m}\ 0],</td>
<td></td>
</tr>
</tbody>
</table>
where zeros correspond to the set of transient states in the unperturbed Markov chain, $`\mu_{i}`$ is a stationary distribution of the unperturbed Markov chain corresponding to the $`i`$-th ergodic set, and $`\nu_{i}`$ is the $`i`$-th element of the aggregated stationary distribution vector that can be found by solution
<table>
<tbody>
<tr>
<td></td>
<td>νD=ν,ν𝟏=1,formulae-sequence𝜈𝐷𝜈𝜈11\nu D=\nu,\quad\nu{\bf 1}=1,</td>
<td></td>
</tr>
</tbody>
</table>
where $`D = {M\hspace{0pt}C\hspace{0pt}B}`$ is the generator of the aggregated Markov chain and
<table>
<tbody>
<tr>
<td></td>
<td>M=[μ100⋱0μm0]∈ℝm×n,B=[𝟏0⋱0𝟏ϕ1⋯ϕm]∈ℝn×m.formulae-sequence𝑀delimited-[]subscript𝜇1missing-subexpression00missing-subexpression⋱missing-subexpressionmissing-subexpression0missing-subexpressionsubscript𝜇𝑚0superscriptℝ𝑚𝑛𝐵delimited-[]1missing-subexpression0missing-subexpressionmissing-subexpression⋱missing-subexpressionmissing-subexpression0missing-subexpression1missing-subexpressionsubscriptitalic-ϕ1⋯subscriptitalic-ϕ𝑚missing-subexpressionsuperscriptℝ𝑛𝑚M=\left[\begin{array}[]{cccc}\mu_{1}&&0&0\\ &\ddots&&\\ 0&&\mu_{m}&0\end{array}\right]\in\mathbb{R}^{m\times n},\quad B=\left[\begin{array}[]{cccc}{\bf 1}&&0\\ &\ddots&\\ 0&&{\bf 1}\\ \phi_{1}&\cdots&\phi_{m}\end{array}\right]\in\mathbb{R}^{n\times m}.</td>
<td></td>
</tr>
</tbody>
</table>
with $`\phi_{i} = {{\lbrack{I - E}\rbrack}^{- 1}\hspace{0pt}L_{i}\hspace{0pt}\mathbf{1}}`$.
The proof of this lemma can be found in \[[11](#bib.bib11), [12](#bib.bib12), [14](#bib.bib14)\].
### A.2 Proofs
Derivation of ([9](#S4.E9 "In 4 PageRank mass of IN+SCC ‣ Distribution of PageRank Mass Among Principle Components of the Web")). First, we observe that if $`\pi_{\text{IN+SCC}}`$ and $`\pi_{\text{DN}}\hspace{0pt}\mathbf{1}`$ are known then it is straightforward to calculate $`\pi_{\text{OUT}}`$. Namely, we have
<table>
<tbody>
<tr>
<td></td>
<td>πOUT=πIN+SCCcR[I−cQ]−1+(1−cn+πDN𝟏cn)𝟏T[I−cQ]−1.subscript𝜋OUTsubscript𝜋IN+SCC𝑐𝑅superscriptdelimited-[]𝐼𝑐𝑄11𝑐𝑛subscript𝜋DN1𝑐𝑛superscript1𝑇superscriptdelimited-[]𝐼𝑐𝑄1\pi_{\mbox{\tiny OUT}}=\pi_{\mbox{\tiny IN+SCC}}cR[I-cQ]^{-1}+\left(\frac{1-c}{n}+\pi_{\mbox{\tiny DN}}{\bf 1}\frac{c}{n}\right){\bf 1}^{T}[I-cQ]^{-1}.</td>
<td></td>
</tr>
</tbody>
</table>
Therefore, let us solve the equations ([7](#S4.E7 "In 4 PageRank mass of IN+SCC ‣ Distribution of PageRank Mass Among Principle Components of the Web")) and ([8](#S4.E8 "In 4 PageRank mass of IN+SCC ‣ Distribution of PageRank Mass Among Principle Components of the Web")). Towards this goal, we sum the elements of the vector equation ([8](#S4.E8 "In 4 PageRank mass of IN+SCC ‣ Distribution of PageRank Mass Among Principle Components of the Web")), which corresponds to the postmultiplication of equation ([8](#S4.E8 "In 4 PageRank mass of IN+SCC ‣ Distribution of PageRank Mass Among Principle Components of the Web")) by vector $`\mathbf{1}`$.
<table>
<tbody>
<tr>
<td></td>
<td>−πIN+SCCcS𝟏+πDN𝟏−cnπDN𝟏𝟏T𝟏=1−cn𝟏T𝟏subscript𝜋IN+SCC𝑐𝑆1subscript𝜋DN1𝑐𝑛subscript𝜋DNsuperscript11𝑇11𝑐𝑛superscript1𝑇1-\pi_{\mbox{\tiny IN+SCC}}cS{\bf 1}+\pi_{\mbox{\tiny DN}}{\bf 1}-\frac{c}{n}\pi_{\mbox{\tiny DN}}{\bf 1}{\bf 1}^{T}{\bf 1}=\frac{1-c}{n}{\bf 1}^{T}{\bf 1}</td>
<td></td>
</tr>
</tbody>
</table>
Now, denote by $`n_{O\hspace{0pt}U\hspace{0pt}T}`$, $`n_{S\hspace{0pt}C\hspace{0pt}C}`$ and $`n_{D\hspace{0pt}N}`$ the number of pages in OUT component, SCC component and the number of dangling nodes. Since $`{\mathbf{1}^{T}\hspace{0pt}\mathbf{1}} = n_{D\hspace{0pt}N}`$, we have
<table>
<tbody>
<tr>
<td></td>
<td>πDN𝟏=nn−cnDN(πIN+SCCcS𝟏+1−cnnDN).subscript𝜋DN1𝑛𝑛𝑐subscript𝑛𝐷𝑁subscript𝜋IN+SCC𝑐𝑆11𝑐𝑛subscript𝑛𝐷𝑁\pi_{\mbox{\tiny DN}}{\bf 1}=\frac{n}{n-cn_{DN}}(\pi_{\mbox{\tiny IN+SCC}}cS{\bf 1}+\frac{1-c}{n}n_{DN}).</td>
<td></td>
</tr>
</tbody>
</table>
Substituting the above expression for $`\pi_{\text{DN}}\hspace{0pt}\mathbf{1}`$ into ([7](#S4.E7 "In 4 PageRank mass of IN+SCC ‣ Distribution of PageRank Mass Among Principle Components of the Web")), we get
<table>
<tbody>
<tr>
<td></td>
<td>πIN+SCC[I−cP−c2n−cnDNS𝟏𝟏T]=cn−cnDN1−cnnDN𝟏T+1−cn𝟏T,subscript𝜋IN+SCCdelimited-[]𝐼𝑐𝑃superscript𝑐2𝑛𝑐subscript𝑛𝐷𝑁𝑆superscript11𝑇𝑐𝑛𝑐subscript𝑛𝐷𝑁1𝑐𝑛subscript𝑛𝐷𝑁superscript1𝑇1𝑐𝑛superscript1𝑇\pi_{\mbox{\tiny IN+SCC}}\left[I-cP-\frac{c^{2}}{n-cn_{DN}}S{\bf 1}{\bf 1}^{T}\right]=\frac{c}{n-cn_{DN}}\frac{1-c}{n}n_{DN}{\bf 1}^{T}+\frac{1-c}{n}{\bf 1}^{T},</td>
<td></td>
</tr>
</tbody>
</table>
which directly implies ([9](#S4.E9 "In 4 PageRank mass of IN+SCC ‣ Distribution of PageRank Mass Among Principle Components of the Web")).
Proof of Proposition [1](#Thmproposition1 "Proposition 1 ‣ 3 The structure of the hyper-link transition matrix ‣ Distribution of PageRank Mass Among Principle Components of the Web") First, we note that if we make a change of variables $`\varepsilon = {1 - c}`$ the Google matrix becomes a transition matrix of a singularly perturbed Markov chain as in Lemma [A.2](#Thmlemma2 "Lemma A.2 ‣ A.1 Results from Singular Perturbation Theory ‣ Appendix ‣ Distribution of PageRank Mass Among Principle Components of the Web") with $`C = {{\frac{1}{n}\hspace{0pt}\mathbf{1}\mathbf{1}^{T}} - P}`$. Let us calculate the aggregated generator matrix $`D`$:
<table>
<tbody>
<tr>
<td></td>
<td>D=MCQ=1n𝟏𝟏TQ−MPQ.𝐷𝑀𝐶𝑄1𝑛superscript11𝑇𝑄𝑀𝑃𝑄D=MCQ=\frac{1}{n}{\bf 1}{\bf 1}^{T}Q-MPQ.</td>
<td></td>
</tr>
</tbody>
</table>
Using $`{M\hspace{0pt}P} = M`$, $`{M\hspace{0pt}Q} = I`$, and $`{M\hspace{0pt}\mathbf{1}} = \mathbf{1}`$ where vectors $`\mathbf{1}`$ are of appropriate dimensions, we obtain
<table>
<tbody>
<tr>
<td></td>
<td>D𝐷\displaystyle D</td>
<td>=1n𝟏𝟏TQ−I=1n𝟏[n1+𝟏[I−T]−1R1𝟏,⋯,nm+𝟏[I−T~]−1Rm𝟏]−I,absent1𝑛superscript11𝑇𝑄𝐼1𝑛1subscript𝑛11superscriptdelimited-[]𝐼𝑇1subscript𝑅11⋯subscript𝑛𝑚1superscriptdelimited-[]𝐼~𝑇1subscript𝑅𝑚1𝐼\displaystyle=\frac{1}{n}{\bf 1}{\bf 1}^{T}Q-I=\frac{1}{n}{\bf 1}[n_{1}+{\bf 1}[I-{T}]^{-1}{R}_{1}{\bf 1},\cdots,n_{m}+{\bf 1}[I-\tilde{T}]^{-1}{R}_{m}{\bf 1}]-I,</td>
<td></td>
</tr>
</tbody>
</table>
where $`n_{i}`$ be the number of nodes in the block $`Q_{i}`$, $`i = {1,\ldots,m}`$. Since the aggregated transition matrix $`D + I`$ has identical rows, its stationary distribution $`\nu`$ is just equal to these rows. Thus, invoking Lemma [A.2](#Thmlemma2 "Lemma A.2 ‣ A.1 Results from Singular Perturbation Theory ‣ Appendix ‣ Distribution of PageRank Mass Among Principle Components of the Web") we obtain ([4](#S3.E4 "In Proposition 1 ‣ 3 The structure of the hyper-link transition matrix ‣ Distribution of PageRank Mass Among Principle Components of the Web")).
Proof of Proposition [2](#Thmproposition2 "Proposition 2 ‣ 4 PageRank mass of IN+SCC ‣ Distribution of PageRank Mass Among Principle Components of the Web") Multiplying both sides of ([15](#S4.E15 "In 4 PageRank mass of IN+SCC ‣ Distribution of PageRank Mass Among Principle Components of the Web")) by $`\mathbf{1}`$ and taking the derivatives, after some tedious algebra we obtain
<table>
<tbody>
<tr>
<td></td>
<td>‖π~IN+SCC′(c)‖=normsubscriptsuperscript~𝜋′IN+SCC𝑐absent\displaystyle||\tilde{\pi}^{\prime}_{\mbox{\tiny IN+SCC}}(c)||=</td>
<td>−a(c)+β1−cβ‖π~scc(c)‖,𝑎𝑐𝛽1𝑐𝛽normsubscript~𝜋𝑠𝑐𝑐𝑐\displaystyle-a(c)+\frac{\beta}{1-c\beta}\,||\tilde{\pi}_{scc}(c)||,</td>
<td></td>
<td>(A.2)</td>
</tr>
</tbody>
</table>
where the real-valued function $`a\hspace{0pt}{(c)}`$ is given by
<table>
<tbody>
<tr>
<td></td>
<td>a(c)=α1−cβ𝐮IN+SCC[I−cP]−1[I−P][I−cP]−1𝟏.𝑎𝑐𝛼1𝑐𝛽subscript𝐮IN+SCCsuperscriptdelimited-[]𝐼𝑐𝑃1delimited-[]𝐼𝑃superscriptdelimited-[]𝐼𝑐𝑃11a(c)=\frac{\alpha}{1-c\beta}{{\bf u}_{\mbox{\tiny IN+SCC}}}[I-cP]^{-1}[I-P][I-cP]^{-1}{\bf 1}.</td>
<td></td>
</tr>
</tbody>
</table>
Differentiating ([A.2](#Sx2.E2 "In A.2 Proofs ‣ Appendix ‣ Distribution of PageRank Mass Among Principle Components of the Web")) and substituting $`\frac{\beta}{1 - {c\hspace{0pt}\beta}}\hspace{0pt}{\|{{\overset{\sim}{\pi}}_{S\hspace{0pt}C\hspace{0pt}C}\hspace{0pt}{(c)}}\|}`$ from ([A.2](#Sx2.E2 "In A.2 Proofs ‣ Appendix ‣ Distribution of PageRank Mass Among Principle Components of the Web")) in the resulting expression, we get
<table>
<tbody>
<tr>
<td></td>
<td>‖π~IN+SCC′′(c)‖normsubscriptsuperscript~𝜋′′IN+SCC𝑐\displaystyle||\tilde{\pi}^{\prime\prime}_{\mbox{\tiny IN+SCC}}(c)||</td>
<td>={−a′(c)+β1−cβa(c)}+2β1−cβ‖π~SCC′(c)‖.absentsuperscript𝑎′𝑐𝛽1𝑐𝛽𝑎𝑐2𝛽1𝑐𝛽normsubscriptsuperscript~𝜋′𝑆𝐶𝐶𝑐\displaystyle=\left\{-a^{\prime}(c)+\frac{\beta}{1-c\beta}a(c)\right\}+\frac{2\beta}{1-c\beta}\,||\tilde{\pi}^{\prime}_{SCC}(c)||.</td>
<td></td>
</tr>
</tbody>
</table>
Note that the term in the curly braces is negative by definition of $`a\hspace{0pt}{(c)}`$. Hence, if $`{\|{{\overset{\sim}{\pi}}_{\text{IN+SCC}}'\hspace{0pt}{(c)}}\|} \leq 0`$ for some $`c \in {\lbrack 0,1\rbrack}`$ then $`{\|{{\overset{\sim}{\pi}}_{\text{IN+SCC}}^{\operatorname{\prime\prime}}\hspace{0pt}{(c)}}\|} < 0`$ for this value of $`c`$.
Proof of Proposition [3](#Thmproposition3 "Proposition 3 ‣ 5 PageRank mass of ESCC ‣ Distribution of PageRank Mass Among Principle Components of the Web") (i) The function $`{f\hspace{0pt}{(c)}} = {{\gamma\hspace{0pt}{({1 - c})}}/{({1 - {\lambda_{1}\hspace{0pt}c}})}}`$ is decreasing and concave, and so is $`\|{\pi_{\text{ESCC}}\hspace{0pt}{(c)}}\|`$. Also, $`{\|{\pi_{\text{ESCC}}\hspace{0pt}{(0)}}\|} = {f\hspace{0pt}{(0)}} = \gamma`$, and $`{\|{\pi_{\text{ESCC}}\hspace{0pt}{(1)}}\|} = {f\hspace{0pt}{(1)}} = 0`$. Thus, for $`c \in {(0,1)}`$, the plot of $`\|{\pi_{\text{ESCC}}\hspace{0pt}{(c)}}\|`$ is either entirely above or entirely below $`f\hspace{0pt}{(c)}`$. In particular, if the first derivatives satisfy $`{\|{\pi_{\text{ESCC}}'\hspace{0pt}{(0)}}\|} < {f'\hspace{0pt}{(0)}}`$, then $`{\|{\pi_{\text{ESCC}}\hspace{0pt}{(c)}}\|} < {f\hspace{0pt}{(c)}}`$ for any $`c \in {(0,1)}`$. Since $`{f'\hspace{0pt}{(0)}} = {\gamma\hspace{0pt}{({\lambda_{1} - 1})}}`$ and $`{\|{\pi_{\text{ESCC}}'\hspace{0pt}{(0)}}\|} = {\gamma\hspace{0pt}{({p_{1} - 1})}}`$, we see that $`p_{1} < \lambda_{1}`$ implies ([17](#S5.E17 "In item (i) ‣ Proposition 3 ‣ 5 PageRank mass of ESCC ‣ Distribution of PageRank Mass Among Principle Components of the Web")).
The proof of (ii) is similar. We consider a concave decreasing function $`{g\hspace{0pt}{(c)}} = {{\gamma\hspace{0pt}{({1 - c})}}/{({1 - {p_{1}\hspace{0pt}c}})}}`$ and note that $`{g\hspace{0pt}{(0)}} = \gamma`$, $`{g\hspace{0pt}{(1)}} = 0`$. Now, if the condition in (ii) holds then $`{g'\hspace{0pt}{(1)}} > {\|{\pi_{\text{ESCC}}'\hspace{0pt}{(1)}}\|}`$, which implies ([18](#S5.E18 "In item (ii) ‣ Proposition 3 ‣ 5 PageRank mass of ESCC ‣ Distribution of PageRank Mass Among Principle Components of the Web")). |
0710.3439 | cs.NI | cyber | # Utility-Based Wireless Resource Allocation for Variable Rate Transmission
Xiaolu Zhang, Meixia Tao, and Chun Sum Ng Manuscript received March 21, 2007; revised July 19, 2007; accepted September 17, 2007. The editor coordinating the review of this paper and approving it for publication is D. Wu. The paper was presented in part at the IEEE International Conference on Communications, Glasgow, June 2007. The authors are with the Department of Electrical and Computer Engineering, National University of Singapore, Singapore 117576, (e-mail: zhangxiaolu@nus.edu.sg; mxtao@nus.edu.sg; elengcs@nus.edu.sg).
###### Abstract
For most wireless services with variable rate transmission, both average rate and rate oscillation are important performance metrics. The traditional performance criterion, utility of average transmission rate, boosts the average rate but also results in high rate oscillations. We introduce a utility function of instantaneous transmission rates. It is capable of facilitating the resource allocation with flexible combinations of average rate and rate oscillation. Based on the new utility, we consider the time and power allocation in a time-shared wireless network. Two adaptation policies are developed, namely, time sharing (TS) and joint time sharing and power control (JTPC). An extension to quantized time sharing with limited channel feedback (QTSL) for practical systems is also discussed. Simulation results show that by controlling the concavity of the utility function, a tradeoff between the average rate and rate oscillation can be easily made.
###### Index Terms:
Utility function, time-sharing, power control, rate adaptive, fairness.
## I Introduction
An important aspect of wireless systems is dynamic channel characteristics. One promising approach for addressing this issue is to dynamically allocate limited resources based on channel information and system preferences. Traditional investigations on wireless resource allocation pay much attention to hard real-time services. Therein, the goal is to smooth out channel variation and build “bit pipes” that deliver data at a fixed rate. The rapid growth of the Internet has led to an increasing demand for supporting transmissions of best-effort service in wireless systems. These applications allow variable-rate transmission and are tolerant of high rate oscillations. Therefore, opportunistic communications \[[1](#bib.bib1)\] have been introduced to achieve higher system throughput. The concept of opportunistic communications is essentially to transmit more information in good channel states and less in poor ones. Hard real-time service and best-effort service may be viewed as two extremes of rate-oscillation sensitivity. However, services such as many audio and video applications generally expect a balance between average rate and rate oscillation. If constant-rate transmission algorithms are used, the transmission efficiency would be very low. On the other hand, opportunistic scheduling schemes, such as \[[2](#bib.bib2)\] and\[[3](#bib.bib3)\], whose objective is to maximize a utility of average rates, can improve efficiency in terms of average rate but result in high oscillation in instantaneous transmission rates. This thus motivated the need for a new criterion that can be used to facilitate the choice of the combinations of average rate and rate oscillation.
In this letter we propose a new network objective function, namely, Time-average Aggregate concave Utility of instantaneous transmission Rate (TAUR). To illustrate the underlying mechanism of the proposed objective function, let us consider transmitting a same data stream using two different schemes. For scheme one, the data stream is transmitted at a constant speed of $`1`$ Mbit/s during the interval of $`10`$ seconds. For scheme two, no data is transmitted in the first 9 seconds and $`10`$ Mbit/s is used for transmission in the last second. Obviously, the utilities of the two transmission schemes are identical if the utility is defined as a function of average transmission rate. However, the time-average concave utility as a function of instantaneous transmission rate for scheme one is higher than that for scheme two, which is expected if the degree of user satisfaction is concerned. Thus, the resource allocation based on TAUR should be able to balance the average rate and rate oscillation over time by adjusting the concavity of the utility function.
The TAUR-based resource allocation problems are studied in a multi-user wireless system in an adaptive time-division fashion. We first consider the optimal time sharing (TS)-based scheduling policy in a backlogged system with constant power allocation. For a strictly concave utility, our analysis shows that the TS policy allows users with relatively better channel conditions to share a same time frame. We then propose a joint optimal time sharing and power control (JTPC) strategy where both the time-sharing fraction and the transmit power can be varied over time. In addition, a quantized TS policy with limited channel feedback (QTSL) is proposed for the ease of practical implementation.
## II System Model
We consider a single cell consisting of $`N`$ mobile users communicating with a common base station. The communication link between each user and the base station is modelled as a slowly time-varying fading channel with additive white Gaussian noise (AWGN). The channel coefficients remain approximately unchanged during each time frame, but can vary from one frame to another. Let the instantaneous channel gain of user $`i`$ at any given time frame $`t`$ be denoted by $`g_{i}\hspace{0pt}{(t)}`$. The network channel gain is denoted by the $`N`$-tuple $`{\mathbf{g}\hspace{0pt}{(t)}} \triangleq {({g_{1}\hspace{0pt}{(t)}},{g_{2}\hspace{0pt}{(t)}},\ldots,{g_{N}\hspace{0pt}{(t)}})}`$, and has a joint probability density function (PDF) $`f\hspace{0pt}{(\mathbf{g})}`$. Let $`p_{i}\hspace{0pt}{(t)}`$ denote the transmit power allocated to or from user $`i`$. The achievable transmission rate of user $`i`$ in the absence of other users can be expressed as \[[4](#bib.bib4)\]
<table>
<tbody>
<tr>
<td></td>
<td>ci(t)=log2[1+pi(t)gi(t)βN0],subscript𝑐𝑖𝑡subscript21subscript𝑝𝑖𝑡subscript𝑔𝑖𝑡𝛽subscript𝑁0c_{i}(t)=\log_{2}\left[1+\frac{p_{i}(t)g_{i}(t)}{\beta N_{0}}\right],</td>
<td></td>
<td>(1)</td>
</tr>
</tbody>
</table>
where $`N_{0}`$ is the noise power, and $`\beta`$ is the signal-to-noise ratio (SNR) gap \[[4](#bib.bib4)\]. We assume that each time frame can be accessed by all the $`N`$ users in an adaptive time-sharing fashion. Let $`{{\mathbf{ρ}}\hspace{0pt}{(\mathbf{g})}} = {(\rho_{1},\rho_{2},\ldots,\rho_{N})}`$ denote the time-sharing adaptation policy with respect to the network channel gain $`\mathbf{g}`$, where $`\rho_{i}`$ represents the fraction of the frame duration allocated to user $`i`$. Without loss of generality, the interval of a time frame is normalized. The actual transmission rate of user $`i`$ at the $`t`$-th time frame, $`r_{i}\hspace{0pt}{(t)}`$, is given by $`{r_{i}\hspace{0pt}{(t)}} = {\rho_{i}\hspace{0pt}c_{i}\hspace{0pt}{(t)}}`$. The frame index $`t`$ in $`r_{i}\hspace{0pt}{(t)}`$ and $`g_{i}\hspace{0pt}{(t)}`$ may be omitted hereafter if no confusion occurs.
The utility considered here is a function of the instantaneous transmission rate. For user $`i`$, we denote its utility as $`U_{i}\hspace{0pt}{({r_{i}\hspace{0pt}{(t)}})}`$. The exact expression for the utility $`U_{i}\hspace{0pt}{( \cdot )}`$ is not crucial. The analysis throughout this paper is valid for any utility function that is increasing, differentiable and concave.
## III Time Sharing
In this section, we assume that the transmission powers between BS and mobiles are constant and identical for different users, i.e., $`{p_{i}\hspace{0pt}{(t)}} = {p,{\forall i},t}`$, and that the wireless network is fully loaded. We choose the aggregate utility, which is the sum of individual user utilities, as the performance measure. The goal is to find the optimal time-sharing adaptation policy $`{\mathbf{ρ}}^{\ast}\hspace{0pt}{(\mathbf{g})}`$ relative to the instantaneous network channel condition $`\mathbf{g}`$, so as to maximize the TAUR of the system. Since the channel processes are ergodic, the optimization problem can be expressed mathematically as
<table>
<tbody>
<tr>
<td></td>
<td>max𝝆(𝐠)ITSsubscript𝝆𝐠subscript𝐼TS\displaystyle\max_{\mbox{\boldmath$\rho$}({\mathbf{g}})}~{}~{}I_{\rm TS}</td>
<td>≜≜\displaystyle\triangleq</td>
<td>𝔼t[∑i=1NUi(ri(t))]subscript𝔼𝑡delimited-[]subscriptsuperscript𝑁𝑖1subscript𝑈𝑖subscript𝑟𝑖𝑡\displaystyle\mathbb{E}_{t}\left[\sum^{N}_{i=1}U_{i}(r_{i}(t))\right]</td>
<td></td>
<td></td>
</tr>
<tr>
<td></td>
<td></td>
<td>=\displaystyle=</td>
<td>∫𝐠∑i=1NUi(ρi(𝐠),gi)f(𝐠)d𝐠subscript𝐠subscriptsuperscript𝑁𝑖1subscript𝑈𝑖subscript𝜌𝑖𝐠subscript𝑔𝑖𝑓𝐠d𝐠\displaystyle\int_{\mathbf{g}}\sum^{N}_{i=1}U_{i}(\rho_{i}(\mathbf{g}),g_{i})f(\mathbf{g})\mbox{d}\mathbf{g}</td>
<td></td>
<td></td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td>s.t.</td>
<td></td>
<td>∑i=1Nρi(𝐠)=1.subscriptsuperscript𝑁𝑖1subscript𝜌𝑖𝐠1\displaystyle\sum^{N}_{i=1}\rho_{i}(\mathbf{g})=1.</td>
<td></td>
<td>(3)</td>
</tr>
</tbody>
</table>
where notation $`{\mathbb{E}}_{t}\hspace{0pt}{\lbrack \cdot \rbrack}`$ represents the time average.
Since the constraint ([3](#S3.E3 "In III Time Sharing ‣ Utility-Based Wireless Resource Allocation for Variable Rate Transmission")) is defined for all channel states, the average aggregate utility maximization in ([III](#S3.Ex1 "III Time Sharing ‣ Utility-Based Wireless Resource Allocation for Variable Rate Transmission")) is equivalent to maximizing the instantaneous aggregate utility for every channel state. Furthermore, since the utility $`U_{i}\hspace{0pt}{( \cdot )}`$ is a concave function of $`r_{i}`$ by assumption, $`U_{i}\hspace{0pt}{( \cdot )}`$ is also concave in $`\rho_{i}`$. Therefore, taking the derivative of the Lagrangian $`{\sum_{i = 1}^{N}{U_{i}\hspace{0pt}{({\rho_{i}\hspace{0pt}{(\mathbf{g})}},g_{i})}}} + {\lambda\hspace{0pt}{\lbrack{1 - {\sum_{i = 1}^{N}{\rho_{i}\hspace{0pt}{(\mathbf{g})}}}}\rbrack}}`$, and equating it to zero, we obtain $`\rho_{i}^{\ast}`$ as
<table>
<tbody>
<tr>
<td></td>
<td>ρi∗(𝐠)=[(∂Ui(ρi,gi)∂ρi)−1(λ)]+,i=1,2,…,N.formulae-sequencesuperscriptsubscript𝜌𝑖𝐠superscriptdelimited-[]superscriptsubscript𝑈𝑖subscript𝜌𝑖subscript𝑔𝑖subscript𝜌𝑖1𝜆𝑖12…𝑁\rho_{i}^{*}(\mathbf{g})=\left[\left(\frac{\partial U_{i}(\rho_{i},g_{i})}{\partial\rho_{i}}\right)^{-1}\left(\lambda\right)\right]^{+},~{}~{}i=1,2,\ldots,N.</td>
<td></td>
<td>(4)</td>
</tr>
</tbody>
</table>
In ([4](#S3.E4 "In III Time Sharing ‣ Utility-Based Wireless Resource Allocation for Variable Rate Transmission")), $`\left( {\partial{U_{i}/{\partial\rho_{i}}}} \right)^{- 1}\hspace{0pt}{( \cdot )}`$ is the inverse function of $`{({\partial{U_{i}/{\partial\rho_{i}}}})}\hspace{0pt}{( \cdot )}`$
<sup>1</sup>
<sup>1</sup>1When the utility function is strictly concave, $`{({\partial{U_{i}/{\partial\rho_{i}}}})}\hspace{0pt}{( \cdot )}`$ is a monotonically decreasing function of $`\rho`$ and, hence, its inverse exists.
, and $`{(x)}^{+} \triangleq {\max{(0,x)}}`$. The Lagrange multiplier $`\lambda`$ can be determined using the constraint ([3](#S3.E3 "In III Time Sharing ‣ Utility-Based Wireless Resource Allocation for Variable Rate Transmission")). As can be seen, the optimal time-sharing policy is only a function of the instantaneous channel conditions and is independent of the channel statistics. The explicit solution of the proposed optimal TS policy in a two-user network with log utility is discussed in \[[5](#bib.bib5)\].
We now compare the proposed TS policy for maximizing the time-average aggregate utility of instantaneous rate with the existing gradient scheduling (GS) policy \[[3](#bib.bib3)\] for maximizing the aggregate utility of average transmission rate. As shown in \[[3](#bib.bib3)\], the GS policy maximizes the weighted sum of instantaneous rates in the system. In a time-shared wireless network, this results in choosing the user satisfying the following condition to transmit during the whole time frame:
<table>
<tbody>
<tr>
<td></td>
<td>i∗(t)=argmax1≤i≤N∂Ui(Ri(t))∂Ri(t)ci(t).superscript𝑖𝑡subscript1𝑖𝑁subscript𝑈𝑖subscript𝑅𝑖𝑡subscript𝑅𝑖𝑡subscript𝑐𝑖𝑡i^{*}(t)=\arg\max_{1\leq i\leq N}{\frac{\partial U_{i}(R_{i}(t))}{\partial R_{i}(t)}c_{i}(t)}.</td>
<td></td>
<td>(5)</td>
</tr>
</tbody>
</table>
Here, $`R_{i}\hspace{0pt}{(t)}`$ is updated as $`{R_{i}\hspace{0pt}{(t)}} = {{{({1 - \alpha})}\hspace{0pt}R_{i}\hspace{0pt}{({t - 1})}} + {\alpha\hspace{0pt}c_{i}\hspace{0pt}{({t - 1})}}}`$ for $`i = {i^{\ast}\hspace{0pt}{({t - 1})}}`$ and $`{R_{i}\hspace{0pt}{(t)}} = {{({1 - \alpha})}\hspace{0pt}R_{i}\hspace{0pt}{({t - 1})}}`$ for $`{\forall i} \neq {i^{\ast}\hspace{0pt}{({t - 1})}}`$, with arbitrary initial value $`R_{i}\hspace{0pt}{(0)}`$, and $`\alpha > 0`$ is a fixed small parameter. On the other hand, in the proposed TS policy, it shall be clear in Section V that the decreasing marginal utility gives opportunity to the users in poor channel condition to share the time frame.
Fig. [1](#A0.F1 "Figure 1 ‣ Utility-Based Wireless Resource Allocation for Variable Rate Transmission") shows the simulated TAUR in the network by using the proposed TS policy and the existing GS policy. Here the utility of both policies is specified to have the form
<table>
<tbody>
<tr>
<td></td>
<td>U(r)=ln(1+rA),𝑈𝑟1𝑟𝐴U(r)=\ln\left(1+\frac{r}{A}\right),</td>
<td></td>
<td>(6)</td>
</tr>
</tbody>
</table>
where $`A > 0`$ is a concavity indicator. The channels are assumed to be Rayleigh fading and the SNR gap is set to $`8.2`$ dB for all users which corresponds to a bit-error-rate requirement of $`10^{- 5}`$ when adaptive quadrature amplitude modulation (QAM) is used. The number of users in the network varies from 8 to 32 and their channel conditions are symmetric. The concavity indicator $`A`$ is set at $`0.1`$. Fig. [2](#A0.F2 "Figure 2 ‣ Utility-Based Wireless Resource Allocation for Variable Rate Transmission") compares the mean and the standard deviation of the transmission rate achieved by TS and GS policies for $`A = {0.1,1}`$ and 10 with $`N = 32`$. The performances of the GS policy at different values of $`A`$ are identical due to the channel symmetry among the users. In fact, it is an extreme case of our TS policy when $`A\rightarrow\infty`$ . Although GS can obtain the maximum average rate, the standard deviation of the rate increases rapidly as the average SNR increases. The TS policy, on the other hand, has a flexible balance between the average rate and the rate oscillation through adjusting the concavity of the utility function.
## IV joint time-sharing and power control
In this section we allow both the transmission time and power to change with respect to channel conditions in each time frame. The optimization problem in Section [III](#S3 "III Time Sharing ‣ Utility-Based Wireless Resource Allocation for Variable Rate Transmission") is extended to finding the joint optimal time-sharing and power control policy. Uplink and downlink transmission are considered separately due to different power constraints. For the uplink, the power source is, generally, rechargeable batteries attached to the mobile devices. Thus, the optimization is subject to each user’s average power constraint. Mathematically, this can be represented as
<table>
<tbody>
<tr>
<td></td>
<td>max(𝝆,𝐩)subscript𝝆𝐩\displaystyle\max_{(\mbox{\boldmath$\rho$},\mathbf{p})}</td>
<td></td>
<td>IJTPC≜∫𝐠∑i=1NUi(ρi(𝐠),pi(𝐠),gi)f(𝐠)d𝐠≜subscript𝐼JTPCsubscript𝐠subscriptsuperscript𝑁𝑖1subscript𝑈𝑖subscript𝜌𝑖𝐠subscript𝑝𝑖𝐠subscript𝑔𝑖𝑓𝐠d𝐠\displaystyle I_{\rm JTPC}\triangleq\int_{\mathbf{g}}\sum^{N}_{i=1}U_{i}(\rho_{i}(\mathbf{g}),p_{i}(\mathbf{g}),g_{i})f(\mathbf{g})\mbox{d}\mathbf{g}</td>
<td></td>
<td>(7)</td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td>s.t.formulae-sequencest\displaystyle{\rm s.t.}</td>
<td></td>
<td>∑i=1Nρi(𝐠)=1subscriptsuperscript𝑁𝑖1subscript𝜌𝑖𝐠1\displaystyle\sum^{N}_{i=1}\rho_{i}(\mathbf{g})=1</td>
<td></td>
<td rowspan="2">(9)</td>
</tr>
<tr>
<td></td>
<td></td>
<td></td>
<td>∫𝐠ρi(𝐠)pi(𝐠)f(𝐠)d𝐠=p¯i,i=1,2,…,N,formulae-sequencesubscript𝐠subscript𝜌𝑖𝐠subscript𝑝𝑖𝐠𝑓𝐠d𝐠subscript¯𝑝𝑖𝑖12…𝑁\displaystyle\int_{\mathbf{g}}\rho_{i}(\mathbf{g})p_{i}(\mathbf{g})f(\mathbf{g})\mbox{d}\mathbf{g}=\bar{p}_{i},~{}i=1,2,\ldots,N,</td>
<td></td>
</tr>
</tbody>
</table>
where $`{\overline{p}}_{i}`$ is the average power constraint of user $`i`$.
Note that the utility function $`U\hspace{0pt}{(\rho,p,g)}`$ is concave in $`\rho`$ and $`p`$ separately based on our assumption, but not in both $`\rho`$ and $`p`$. Moreover, the equality constraints in ([9](#S4.E9 "In IV joint time-sharing and power control ‣ Utility-Based Wireless Resource Allocation for Variable Rate Transmission")) are nonlinear. To make the problem more tractable, we define $`s = {p\hspace{0pt}\rho}`$. It can be shown that $`U\hspace{0pt}{(\rho,s,g)}`$ is concave in both $`\rho`$ and $`s`$ (since its Hessian matrix is negative semidefinite). The problem thus falls into the classic calculus of variations \[[6](#bib.bib6)\]. Applying the Euler-Lagrange equation results in the following necessary and sufficient conditions for the optimal solution $`\rho_{i}^{\ast}`$ and $`s_{i}^{\ast}`$:
<table>
<tbody>
<tr>
<td></td>
<td>∂Ui∂si+λi=0,i=1,2,…,Nformulae-sequencesubscript𝑈𝑖subscript𝑠𝑖subscript𝜆𝑖0𝑖12…𝑁\displaystyle\frac{\partial U_{i}}{\partial s_{i}}+\lambda_{i}=0,~{}~{}~{}i=1,2,\ldots,N</td>
<td></td>
<td>(10)</td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td>∂Ui∂ρif(𝐠)+λ0(𝐠)=0,i=1,2,…,Nformulae-sequencesubscript𝑈𝑖subscript𝜌𝑖𝑓𝐠subscript𝜆0𝐠0𝑖12…𝑁\displaystyle\frac{\partial U_{i}}{\partial\rho_{i}}f(\mathbf{g})+\lambda_{0}(\mathbf{g})=0,~{}~{}~{}i=1,2,\ldots,N</td>
<td></td>
<td>(11)</td>
</tr>
</tbody>
</table>
where $`\lambda_{0}\hspace{0pt}{(\mathbf{g})}`$ and $`\lambda_{i}`$ are Lagrange multipliers and determined by constraints ([9](#S4.E9 "In IV joint time-sharing and power control ‣ Utility-Based Wireless Resource Allocation for Variable Rate Transmission")) and ([9](#S4.E9 "In IV joint time-sharing and power control ‣ Utility-Based Wireless Resource Allocation for Variable Rate Transmission")). The closed-form solutions to the above equations are generally difficult to obtain due to nonlinearity of the utility function $`U`$ in $`\rho`$ and $`s`$. The nonlinear Gauss-Seidel algorithm \[[7](#bib.bib7)\] can be used to search for the optimal time-sharing vector $`{\mathbf{ρ}}^{\ast}`$ and vector $`\mathbf{s}^{\ast} = {\lbrack s_{1}^{\ast},s_{2}^{\ast},\ldots,s_{N}^{\ast}\rbrack}`$ under the average power constraint $`\overline{\mathbf{p}} = {\lbrack{\overline{p}}_{1},{\overline{p}}_{2},\ldots,{\overline{p}}_{N}\rbrack}`$. It is outlined as follows.
1. 1\.
*Initialize $`\mathbf{s}`$*: set $`\mathbf{p}^{(0)} = {N\hspace{0pt}\overline{\mathbf{p}}}`$, $`\rho_{i} = \left. 1/N \right.`$, $`\forall i`$, let $`j = 0`$ and calculate the initial $`\mathbf{s}^{(\mathbf{0})}`$.
2. 2\.
*Update $`\rho^{(j)}`$* : given $`\mathbf{s}^{(j)}`$, find the optimal time-sharing vector $`{\mathbf{ρ}}^{(j)}`$ using (<a href="#S4.E11" style="font-size:90%;" title="In IV joint time-sharing and power control ‣ Utility-Based Wireless Resource Allocation for Variable Rate Transmission">11</a>)
<table>
<tbody>
<tr>
<td></td>
<td>ρi(j)(𝐠)=[(∂Ui∂ρi)−1(λ0′,si(j))]+,i=1,2,…,N,formulae-sequencesuperscriptsubscript𝜌𝑖𝑗𝐠superscriptdelimited-[]superscriptsubscript𝑈𝑖subscript𝜌𝑖1superscriptsubscript𝜆0′subscriptsuperscript𝑠𝑗𝑖𝑖12…𝑁\rho_{i}^{(j)}(\mathbf{g})=\left[\left(\frac{\partial U_{i}}{\partial\rho_{i}}\right)^{-1}\left(\lambda_{0}^{\prime},s^{(j)}_{i}\right)\right]^{+},~{}~{}i=1,2,\ldots,N,</td>
<td></td>
</tr>
</tbody>
</table>
where $`\lambda_{0}'`$ satisfies (<a href="#S4.E9" style="font-size:90%;" title="In IV joint time-sharing and power control ‣ Utility-Based Wireless Resource Allocation for Variable Rate Transmission">9</a>)
3. 3\.
*Compute the average aggregate utility $`I^{(j)}`$ using ([7](#S4.E7 "In IV joint time-sharing and power control ‣ Utility-Based Wireless Resource Allocation for Variable Rate Transmission"))*
4. 4\.
*Update $`\mathbf{s}^{({j + 1})}`$*: given $`{\mathbf{ρ}}^{(j)}`$, find the optimal power control vector $`\mathbf{s}^{({j + 1})}`$ using (<a href="#S4.E10" style="font-size:90%;" title="In IV joint time-sharing and power control ‣ Utility-Based Wireless Resource Allocation for Variable Rate Transmission">10</a>)
<table>
<tbody>
<tr>
<td></td>
<td>si(j+1)(𝐠)=[(∂Ui∂si)−1(λi,ρi(j))]+fori=1,2,…,N,formulae-sequencesuperscriptsubscript𝑠𝑖𝑗1𝐠superscriptdelimited-[]superscriptsubscript𝑈𝑖subscript𝑠𝑖1subscript𝜆𝑖subscriptsuperscript𝜌𝑗𝑖for𝑖12…𝑁s_{i}^{(j+1)}(\mathbf{g})=\left[\left(\frac{\partial U_{i}}{\partial s_{i}}\right)^{-1}\left(\lambda_{i},\rho^{(j)}_{i}\right)\right]^{+}~{}\mbox{for}~{}i=1,2,\ldots,N,</td>
<td></td>
</tr>
</tbody>
</table>
where $`\lambda_{i}`$ satisfies (<a href="#S4.E9" style="font-size:90%;" title="In IV joint time-sharing and power control ‣ Utility-Based Wireless Resource Allocation for Variable Rate Transmission">9</a>) for $`i = {1,2,\ldots,N}`$ and let $`j = {j + 1}`$.
5. 5\.
*Repeat Steps 2)-4) until $`{I^{({j + 1})} - I^{(j)}} < \Delta`$, where $`\Delta`$ is a small number.*
At each iteration, the optimization of $`\mathbf{ρ}`$ and $`\mathbf{s}`$ is carried out successively. Steps 2) and 4) involve only the calculation of a one-dimensional maximization problem whose solution is given in Section [III](#S3 "III Time Sharing ‣ Utility-Based Wireless Resource Allocation for Variable Rate Transmission"). The condition that $`U`$ is continuously differentiable and concave in ($`\rho`$, $`s`$) guarantees the convergence of the nonlinear Gauss-Seidel algorithm. The proof can be seen in \[[7](#bib.bib7), Prop 3.9 in Section 3.3\].
The problem formulation for the downlink differs from the uplink only in the power constraint, which is given by $`{\int_{\mathbf{g}}{\sum_{i = 1}^{N}{\rho_{i}\hspace{0pt}{(\mathbf{g})}\hspace{0pt}p_{i}\hspace{0pt}{(\mathbf{g})}\hspace{0pt}f\hspace{0pt}{(\mathbf{g})}\hspace{0pt}\text{d}\hspace{0pt}\mathbf{g}}}} = \overline{p}`$. A similar problem-solving approach to the one proposed for the uplink can also be obtained and hence is omitted.
Although JTPC utilizes two degrees of freedom in resource allocation and has much higher computational complexity, its performance is not expected much higher than that of TS in the high SNR region. This is attributed to the fact that the transmission rate is linear in time, but concave in transmission power. That is, at high SNR, the gain from power control is smaller than from time sharing adaptation. This is also verified by the simulation results shown in Fig. [3](#A0.F3 "Figure 3 ‣ Utility-Based Wireless Resource Allocation for Variable Rate Transmission"), where we compare the average aggregate utilities obtained by the proposed TS and JTPC schemes. It is observed that at high SNR the performance gain of JTPC over TS is not noticeable. Hence, in Section [V](#S5 "V Implementation Issues ‣ Utility-Based Wireless Resource Allocation for Variable Rate Transmission"), we assume the absence of power control.
## V Implementation Issues
We have provided the analytical results for TS and JTPC policies. In this section, we address some important implementation issues, including quantized time sharing fractions, limited channel feedback, and fairness.
###
V-A
Quantized Time Sharing With Limited Channel Feedback
In scheduling the downlink transmission, the BS needs to know each user’s channel state information (CSI). This could be gained by sending the CSI from each user to the base station through a feedback channel upon channel estimation at each user terminal. In practice, perfect channel feedback is not feasible due to limited capacity of the feedback links. We assume in this subsection that the channel estimate of each user is quantized into $`K = 2^{M}`$ regions using $`M`$ bits. Let $`{\mathbb{G}} = {\{ G_{1},G_{2},\ldots,G_{K + 1}\}}`$ be the set of channel gain thresholds in increasing order with $`G_{1} = 0`$ and $`G_{K + 1} = \infty`$. If the channel gain of user $`i`$ falls into range $`\lbrack G_{k},G_{k + 1})`$, we say user $`i`$ is in channel state $`k`$, and denote it as $`S_{i} = k`$. Suppose we apply the equal-probability method to do the channel partitioning and the channel gains follow exponential distribution, the threshold set $`\mathbb{G}`$ can be determined easily.
Furthermore, time sharing fractions in practice cannot be an arbitrary number, but are restricted to a finite set of values due to switching latency and difficulties in rigid synchronization. Therefore, we assume that a time frame is partitioned into $`L`$ slots with equal length. Correspondingly, the number of users which can transmit in the same frame is limited by $`L`$. At the beginning of each frame, the base station computes the optimal time-sharing vector $`{\mathbf{ρ}}^{\ast}`$ defined in ([12](#S5.E12 "In V-A Quantized Time Sharing With Limited Channel Feedback ‣ V Implementation Issues ‣ Utility-Based Wireless Resource Allocation for Variable Rate Transmission")) upon obtaining the network channel states $`\mathbf{S} = {\lbrack S_{1},S_{2},\ldots,S_{N}\rbrack}`$.
<table>
<tbody>
<tr>
<td></td>
<td>𝝆∗(𝐒)superscript𝝆𝐒\displaystyle\mbox{\boldmath$\rho$}^{*}(\mathbf{S})</td>
<td>=\displaystyle=</td>
<td>argmax𝝆(𝐒)∫GS1GS1+1∫GS2GS2+1…∫GSNGSN+1[∑i=1NUi(ρi(𝐠),gi)]d𝐠argsubscript𝝆𝐒superscriptsubscriptsubscript𝐺subscript𝑆1subscript𝐺subscript𝑆11superscriptsubscriptsubscript𝐺subscript𝑆2subscript𝐺subscript𝑆21…superscriptsubscriptsubscript𝐺subscript𝑆𝑁subscript𝐺subscript𝑆𝑁1delimited-[]superscriptsubscript𝑖1𝑁subscript𝑈𝑖subscript𝜌𝑖𝐠subscript𝑔𝑖d𝐠\displaystyle\mbox{arg}\max_{\mbox{\boldmath$\rho$}(\mathbf{S})}~{}\int_{G_{S_{1}}}^{G_{S_{1}+1}}\int_{G_{S_{2}}}^{G_{S_{2}+1}}\ldots\int_{G_{S_{N}}}^{G_{S_{N}+1}}\left[\sum_{i=1}^{N}U_{i}(\rho_{i}(\mathbf{g}),g_{i})\right]\mbox{d}\mathbf{g}</td>
<td></td>
<td>(12)</td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td>s.t.</td>
<td></td>
<td>ρi∈{0,1L,2L,…,1},∀i∈{1,…,N}and∑i=1Nρi=1.formulae-sequencesubscript𝜌𝑖01𝐿2𝐿…1for-all𝑖1…𝑁andsuperscriptsubscript𝑖1𝑁subscript𝜌𝑖1\displaystyle\rho_{i}\in\left\{0,\frac{1}{L},\frac{2}{L},\ldots,1\right\},~{}\forall i\in\{1,\ldots,N\}~{}\mbox{and}~{}\sum_{i=1}^{N}\rho_{i}=1.</td>
<td></td>
<td></td>
</tr>
</tbody>
</table>
The time sharing policy considered here maps the current channel states $`\mathbf{S} \in {\mathbb{R}}_{+}^{N}`$ to a time-sharing vector To avoid the exponential complexity in exhaustive search, an online greedy algorithm with complexity of $`\mathcal{O}\hspace{0pt}{({L\hspace{0pt}N})}`$ is proposed. Beginning with an initial solution $`{\mathbf{ρ}} = {\lbrack 0,0,\ldots,0\rbrack}`$, each time slot is assigned at one iteration to the most favorable user that maximizes the increment of the current objective till the total $`L`$ slots are traversed. The greedy algorithm is outlined below:
1. 1\.
*Initialization*
Let $`v = 0`$ (the index of the time slot), $`\rho_{i}^{(0)} = 0`$ and $`U_{i}^{(0)} = {0\hspace{0pt}\left( {{\forall i} \in \left\{ 1,\ldots,N \right\}} \right)}`$.
2. 2\.
*Allocate the $`{({v + 1})}^{t\hspace{0pt}h}`$ time slot to the user indexed by $`i^{\ast}`$*
<table>
<tbody>
<tr>
<td></td>
<td>i∗=argmaxi∈{1,…,N}∫GSiGSi+1[Ui(ρi(v)+1/L)−Ui(ρi(v))]dgi.superscript𝑖argsubscript𝑖1…𝑁superscriptsubscriptsubscript𝐺subscript𝑆𝑖subscript𝐺subscript𝑆𝑖1delimited-[]subscript𝑈𝑖superscriptsubscript𝜌𝑖𝑣1𝐿subscript𝑈𝑖superscriptsubscript𝜌𝑖𝑣dsubscript𝑔𝑖i^{*}=\mbox{arg}~{}\max_{i\in\{1,\ldots,N\}}~{}\int_{G_{S_{i}}}^{G_{S_{i}+1}}\bigg{[}U_{i}(\rho_{i}^{(v)}+1/L)-U_{i}(\rho_{i}^{(v)})\bigg{]}\mbox{d}{g_{i}}.</td>
<td></td>
<td>(13)</td>
</tr>
</tbody>
</table>
Let $`\rho_{i^{\ast}}^{({v + 1})} = {\rho_{i^{\ast}}^{(v)} + \left. 1/L \right.}`$ and $`\rho_{i}^{({v + 1})} = \rho_{i}^{(v)}`$ for $`i \neq i^{\ast}`$.
3. 3\.
*Let $`v = {v + 1}`$, and return to Step 2) until $`v = L`$*
It is shown in the appendix that this algorithm leads to the optimal solution to Problem ([12](#S5.E12 "In V-A Quantized Time Sharing With Limited Channel Feedback ‣ V Implementation Issues ‣ Utility-Based Wireless Resource Allocation for Variable Rate Transmission")).
Fig. [4](#A0.F4 "Figure 4 ‣ Utility-Based Wireless Resource Allocation for Variable Rate Transmission") shows the performance of the quantized time sharing with limited channel feedback (QTSL) policy. The CSI is quantized at 3-bit resolution, and the number of time slots is the same as the number of users in the network. It is seen that the performance of QTSL approaches that obtained by the optimal TS policy. In Fig. [5](#A0.F5 "Figure 5 ‣ Utility-Based Wireless Resource Allocation for Variable Rate Transmission"), we illustrate the performance of the QTSL policy when the number of time slots in a time frame is half of the number of users in the network. This time we also vary the number of channel feedback bits from $`1`$ to $`3`$. There is a performance gap between QTSL and TS, but the average aggregate utility obtained by QTSL with $`N = 32`$ and $`L = 16`$ is still higher than that of the optimal TS policy with $`N = 16`$. In addition, the performance gain is limited when the CSI is quantized using more than two bits.
###
V-B
Fairness Guarantee
The scheduling schemes based on the aggregate utility maximization developed in previous sections may not guarantee fairness for users with different channel statistics, such as path loss and shadowing. This is because the TS scheme tends to allocate more time to the user with higher average SNR. A commonly used fairness criterion in computer networks is the max-min fairness \[[8](#bib.bib8)\]. Unlike wireline networks, the wireless network suffers from time-varying channel impairments. Thus, time-average utility max-min (TUMM) fairness, defined below, is meaningful in wireless networks.
*Definition 1*: A time sharing policy $`\mathbf{ρ}`$ is time-average utility max-min fair if for each $`i \in {\{ 1,2,\ldots,N\}}`$ and any other time sharing policy $`\overline{\mathbf{ρ}}`$ for which $`{{\mathbb{E}}\hspace{0pt}{\lbrack{U_{i}\hspace{0pt}{(\rho_{i})}}\rbrack}} < {{\mathbb{E}}\hspace{0pt}{\lbrack{U_{i}\hspace{0pt}{(\overline{\rho_{i}})}}\rbrack}}`$, there exists some $`j`$ with $`{{\mathbb{E}}\hspace{0pt}{\lbrack{U_{i}\hspace{0pt}{(\rho_{i})}}\rbrack}} \geq {{\mathbb{E}}\hspace{0pt}{\lbrack{U_{j}\hspace{0pt}{(\rho_{j})}}\rbrack}} > {{\mathbb{E}}\hspace{0pt}{\lbrack{U_{j}\hspace{0pt}{({\overline{\rho}}_{j})}}\rbrack}}`$, i.e., *increasing some component $`{\mathbb{E}}\hspace{0pt}{({U_{i}\hspace{0pt}{(\rho_{i})}})}`$ must be at the expense of decreasing some already smaller component $`{\mathbb{E}}\hspace{0pt}{\lbrack{U_{j}\hspace{0pt}{(\overline{\rho_{j}})}}\rbrack}`$*
Since the resource allocation based on TUMM fairness does not concern the fairness at any instant, it allows the scheduler to exploit the instantaneous fluctuation of the channel conditions. We formulate the TUMM-based time sharing as a multiple-objective programming problem:
<table>
<tbody>
<tr>
<td></td>
<td>max𝝆(𝐠)subscript𝝆𝐠\displaystyle\max_{\mbox{\boldmath$\rho$}(\mathbf{g})}</td>
<td></td>
<td>a𝑎\displaystyle a</td>
<td></td>
<td>(14)</td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td>s.t.</td>
<td></td>
<td>𝔼[Ui(ρi(𝐠),gi)]=a,i=1,2,…,Nformulae-sequence𝔼delimited-[]subscript𝑈𝑖subscript𝜌𝑖𝐠subscript𝑔𝑖𝑎𝑖12…𝑁\displaystyle\mathbb{E}[U_{i}(\rho_{i}(\mathbf{g}),g_{i})]=a,~{}i=1,2,\ldots,N</td>
<td></td>
<td></td>
</tr>
<tr>
<td></td>
<td></td>
<td></td>
<td>∑i=1Nρi(𝐠)=1,subscriptsuperscript𝑁𝑖1subscript𝜌𝑖𝐠1\displaystyle\sum^{N}_{i=1}\rho_{i}(\mathbf{g})=1,</td>
<td></td>
<td></td>
</tr>
</tbody>
</table>
where $`a`$ is a variable to be maximized. When $`U_{i}\hspace{0pt}{( \cdot )}`$’s are all strictly concave, there exists a unique solution.
A frequently used method in multiple objective programming problem is the point estimate weighted-sum approach \[[9](#bib.bib9)\]. In this method, each objective is multiplied by a weight $`w_{i}`$. Then, the $`N`$ weighted objectives are summed to form a weighted-sum objective function, denoted as $`\sum_{i = 1}^{N}{w_{i}\hspace{0pt}{\mathbb{E}}\hspace{0pt}{\lbrack{U_{i}\hspace{0pt}{(\rho_{i})}}\rbrack}}`$. The weights $`w_{i}`$’s are chosen such that ([](#S5.Ex5 "V-B Fairness Guarantee ‣ V Implementation Issues ‣ Utility-Based Wireless Resource Allocation for Variable Rate Transmission")
V-B
) is satisfied. It can be proven that the solution to maximize the average aggregate weighted utility is also Pareto-optimal. Take the two-user case for example. The notion of fairness can be realized by dynamically adjusting the weights, i.e., dynamically adjusting the moving direction towards the intersection of the Pareto-optimal frontier and the straight line $`{{\mathbb{E}}\hspace{0pt}{\lbrack U_{1}\rbrack}} = {{\mathbb{E}}\hspace{0pt}{\lbrack U_{2}\rbrack}}`$ while keeping the two users’ average utility on the Pareto-optimal frontier. This weight adaptation method can be used to guarantee fairness when the users have different channel distributions.
## VI Conclusion
We develop a new framework for resource allocation in wireless networks for variable-rate transmission. The time-average aggregate utility of instantaneous transmission rate is proposed to jointly optimize the resulting average rate and rate oscillation. In particular, a time-sharing policy and a joint time-sharing and power control policy are designed to exploit the channel fluctuation. The effects of partial channel state information and discrete time sharing fractions are also studied. Furthermore, an adaptive method to guarantee strict fairness among users with different channel statistics is discussed.
\[Optimality Proof of The Greedy Algorithm\]
Consider one realization of $`\mathbf{S}`$ and $`\rho_{i} \in {\{ 0,{1/L},{2/L},\ldots,1\}}`$. For simplicity, we define $`{{\overset{\sim}{U}}_{i}\hspace{0pt}{(\rho_{i})}}:={\int_{G_{S_{i}}}^{G_{S_{i} + 1}}{U_{i}\hspace{0pt}{({\rho_{i}\hspace{0pt}{(\mathbf{g})}},g_{i})}\hspace{0pt}\text{d}\hspace{0pt}g_{i}}}`$ and $`{d_{i}\hspace{0pt}{(\rho)}} = {{{\overset{\sim}{U}}_{i}\hspace{0pt}{(\rho)}} - {{\overset{\sim}{U}}_{i}\hspace{0pt}\left( {\rho - {1/L}} \right)}}`$. Let $`D_{L}`$ denote the set of $`L`$ largest elements in $`{D = \left. \{{d_{i}\hspace{0pt}{(\rho)}} \middle| {{i = {1,2,\ldots,N}},{\rho = {0,{1/L},{2/L},{\ldots\hspace{0pt}1}}}}\} \right.}.`$ Then, the Lagrangian can be written as:
<table>
<tbody>
<tr>
<td></td>
<td>∑i=1N(U~i(ρi)−λρi)=∑i=1N[U~i(0)+(di(1/L)−λ)+…+(di(ρi)−λ)].superscriptsubscript𝑖1𝑁subscript~𝑈𝑖subscript𝜌𝑖𝜆subscript𝜌𝑖superscriptsubscript𝑖1𝑁delimited-[]subscript~𝑈𝑖0subscript𝑑𝑖1𝐿𝜆…subscript𝑑𝑖subscript𝜌𝑖𝜆\sum_{i=1}^{N}(\tilde{U}_{i}(\rho_{i})-\lambda\rho_{i})=\sum_{i=1}^{N}[\tilde{U}_{i}(0)+(d_{i}(1/L)-\lambda)+\ldots+(d_{i}(\rho_{i})-\lambda)].</td>
<td></td>
<td>(16)</td>
</tr>
</tbody>
</table>
Let $`\lambda`$ be the smallest $`d_{i}\hspace{0pt}{(\rho)}`$ in $`D_{L}`$, then
<table>
<tbody>
<tr>
<td></td>
<td>di(ρ)−λ{≥0ifdi(ρ)∈DL<0otherwise.subscript𝑑𝑖𝜌𝜆casesabsent0ifsubscript𝑑𝑖𝜌subscript𝐷𝐿absent0otherwise\displaystyle d_{i}(\rho)-\lambda\left\{\begin{array}[]{ll}\geq 0&\mbox{if}~{}d_{i}(\rho)\in D_{L}\\ <0&\textrm{otherwise}\end{array}\right..</td>
<td></td>
<td>(19)</td>
</tr>
</tbody>
</table>
Therefore the Lagrangian ([16](#A0.E16 "In VI Conclusion ‣ Utility-Based Wireless Resource Allocation for Variable Rate Transmission")) is maximized by
<table>
<tbody>
<tr>
<td></td>
<td>ρi∗={0ifdi(1/L)∉DL1ifdi(1)∈DLρifdi(ρ)∈DLanddi(ρ+1/L)∉DL..superscriptsubscript𝜌𝑖cases0ifsubscript𝑑𝑖1𝐿subscript𝐷𝐿1ifsubscript𝑑𝑖1subscript𝐷𝐿𝜌ifsubscript𝑑𝑖𝜌subscript𝐷𝐿andsubscript𝑑𝑖𝜌1𝐿subscript𝐷𝐿\displaystyle\rho_{i}^{*}=\left\{\begin{array}[]{ll}0&\mbox{if}~{}d_{i}(1/L)\notin D_{L}\\ 1&\mbox{if}~{}d_{i}(1)\in D_{L}\\ \rho&\mbox{if}~{}d_{i}(\rho)\in D_{L}~{}\mbox{and}~{}d_{i}(\rho+1/L)\notin D_{L}.\end{array}\right..</td>
<td></td>
<td>(23)</td>
</tr>
</tbody>
</table>
Due to the concavity of the utility function, we have $`{d_{i}\hspace{0pt}{({1/L})}} > {d_{i}\hspace{0pt}{({2/L})}} > \ldots > {d_{i}\hspace{0pt}{({N_{i}/L})}} > {d_{i}\hspace{0pt}{({{({N_{i} + 1})}/L})}} > \ldots > {d_{i}\hspace{0pt}{(1)}}`$. Here suppose that $`{d_{i}\hspace{0pt}{({N_{i}/L})}} \in D_{L}`$ and $`{d_{i}\hspace{0pt}{({{({N_{i} + 1})}/L})}} \notin D_{L}`$, then $`{\sum_{i}N_{i}} = {|D_{L}|}`$. Since $`\rho_{i}^{\ast} = {N_{i}/L}`$ from ([23](#A0.E23 "In VI Conclusion ‣ Utility-Based Wireless Resource Allocation for Variable Rate Transmission")), $`{\sum_{i}\rho_{i}^{\ast}} = {{|D_{L}|}/L} = 1`$ holds for the chosen $`\lambda`$. $`\rho_{i}^{\ast}`$ is an optimal solution of Problem ([12](#S5.E12 "In V-A Quantized Time Sharing With Limited Channel Feedback ‣ V Implementation Issues ‣ Utility-Based Wireless Resource Allocation for Variable Rate Transmission")).
Since Step 2) computes the $`L`$ largest $`d_{i}\hspace{0pt}{(\rho)}`$ in the decreasing order of their values, the greedy algorithm obtains the optimal solution.
## References
- \[1\] P. Viswanath, D. Tse, and R. Laroia, “Opportunistic beamforming using dumb antennas,” *IEEE Trans. on Info. Theory*, vol. 48, no. 6, pp. 1277–1294, June 2002.
- \[2\] A. Jalali, R. Padovani, and R. Pankaj, “Data throughput of cdma-hdr a high efficiency-high data rate personal communication wireless system,” in *Proc. Veh. Technol. (VTC)*, 2000.
- \[3\] A. L. Stolyar, “On the asymptotic optimality of the gradient scheduling algorithm for multiuser throughput allocation,” *Operation Research*, vol. 53, no. 1, pp. 12–25, Jan. 2005.
- \[4\] X. Qiu and K. Chawla, “On the performance of adaptive modulation in cellular systems,” *IEEE Trans. on Commun.*, vol. 47, no. 6, pp. 884–895, June 1999.
- \[5\] X. Zhang, M. Tao, and C. S. Ng, “Time sharing policy in wireless networks for variable rate transmission,” in *Proc. IEEE ICC 2007*, June 2007.
- \[6\] B. S. Gottfried and J. Weisman, *Introduction to optimization theory*. Prentice-Hall, 1973.
- \[7\] D. P. Bertsekas and J. N. Tsitsiklis, *Parallel and Distributed Computation: Numerical Methods*. Athena Scientific, 1997.
- \[8\] J. M. Jaffe, “Bottleneck flow control,” *IEEE Trans. Commun.*, vol. 29, no. 7, pp. 954–962, July 1981.
- \[9\] R. E. Steuer, *Multiple Criteria Optimization: Theory, Computation, and Application*. Wiley, New York, 1986.
[Refer to caption](/html/0710.3439/assets/x1.png)
Figure 1: Performance comparison of different resource allocation schemes
[Refer to caption](/html/0710.3439/assets/x2.png)
Figure 2: Average rate and standard deviation of rate versus average SNR using TS and GS
[Refer to caption](/html/0710.3439/assets/x3.png)
Figure 3: Average aggregate utility versus average SNR using TS and JTPA for two-user scenario
[Refer to caption](/html/0710.3439/assets/x4.png)
Figure 4: Performance comparison of the optimal TS policy and QTSL policy for $`N = 32`$ and 16 users, with $`L = N`$ slots and $`M = 3`$ feedback bits
[Refer to caption](/html/0710.3439/assets/x5.png)
Figure 5: Performance comparison of $`N = 32`$ users with $`L = 16`$ and $`N = 16`$ users with $`L = 8`$ time slots, and with $`M = {1,2}`$ and 3 feedback bits |
0708.2395 | cs.CR | cyber | Key Agreement and Authentication Schemes Using Non-Commutative Semigroups
M.M. Chowdhury
Abstract
We give a new two-pass authentication scheme, which is a generalisation of an authentication scheme of Sibert-Dehornoy-Girault based on the Diffie-Hellman conjugacy problem. Compared to the above scheme, for some parameters it is more efficient with respect to multiplications. We sketch a proof that our authentication scheme is secure.
1\. Introduction
In recent years various cryptographic protocols using infinite non-abelian groups have been proposed. For example the seminal algebraic key establishment protocol given in \[[1](#bib.bib1)\], and Artin’s braid groups have been popular choices for such protocols. Braid groups are a popular choice because they are not too complicated to work with and they are more complicated than abelian groups. In particular the conjugacy problem in braid groups is algorithmically difficult and hence gives a one-way function.
We give a new authentication scheme (by using equations (2) and (3) below which form a main part of this paper), which is a generalisation of the authentication scheme $`I`$ of SDG (Sibert-Dehornoy-Girault) given in \[[10](#bib.bib10)\]. We do not claim that equations (2) and (3) are totally original; for example a simpler version of equation (3) is used in the authentication scheme in \[[14](#bib.bib14)\]. In this paper we refer to the authentication scheme $`I`$ in \[[10](#bib.bib10)\] as the SDG scheme. Based on the proof of the SDG scheme given in \[[10](#bib.bib10)\] we sketch a proof that our new authentication scheme is a perfectly honest-verifier ZK interactive proof of knowledge of the prover’s secret. Two other provably secure schemes which are scheme $`I\hspace{0pt}I`$ and scheme $`I\hspace{0pt}I\hspace{0pt}I`$ are given \[[10](#bib.bib10)\] and they differ from our authentication scheme because they are zero-knowledge in an theoretical infinite framework, and they are iterated three pass schemes. Other differences are that scheme $`I\hspace{0pt}I`$ is based on different hard problems compared to our authentication scheme. We refer to \[[10](#bib.bib10)\] for further details of these schemes. Two related authentication schemes were proposed in \[[11](#bib.bib11)\] and it was shown in \[[12](#bib.bib12)\] one $``\hspace{0pt}``$authentication scheme$`"`$ in \[[11](#bib.bib11)\] is totally insecure and the other authentication scheme as shown in \[[12](#bib.bib12)\] can be broken by solving a specialisation of the decomposition problem defined in \[[4](#bib.bib4)\] (see below). An authentication scheme is given in \[[14](#bib.bib14)\] which has security based on a generalisation of the discrete logarithm problem in non-abelian finite groups. The main difference of our authentication scheme from the authentication scheme in \[[14](#bib.bib14)\] and the unbroken scheme in \[[11](#bib.bib11)\], \[[12](#bib.bib12)\] is that it is based on our version of the Diffie-Hellman decomposition problem defined below.
2\. Hard Problems in Non-Abelian Groups
We now define the following known hard problems. The notation $`{\lbrack I,J\rbrack} = 1`$ (resp. $`{\lbrack I,J\rbrack} \neq 1`$) means that the subsets $`I`$ and $`J`$ of a semigroup $`G`$ commute (resp. do not commute). We may consider $`G = B_{n}`$ (the braid group of index $`n`$) in the following known problems, because the problems are then hard and by hard we mean there is no known algorithm to solve the problem such that a cryptographic protocol based on the problem would be insecure for practical use. When $`G = B_{n}`$ then WLOG the usual choices for $`A`$ and $`B`$ are the braid subgroups $`L\hspace{0pt}B_{n}`$ and $`U\hspace{0pt}B_{n}`$ (defined in section 3) as these are the choices that are used in \[[2](#bib.bib2)\] and \[[6](#bib.bib6)\] but the choices of $`A`$ and $`B`$ may be different. For our protocol to be secure at the very least $`G`$ should be non-commutative.
The DP (Decomposition Problem) \[[4](#bib.bib4)\] is defined as follows.
Public Information: $`G`$ is a semigroup, $`A`$ is a subset of $`G`$. $`{x,y} \in G`$ with $`y = {a\hspace{0pt}x\hspace{0pt}b}`$.
Secret information: $`{a,b} \in A`$.
Objective: find elements $`{f,g} \in A`$ such that $`{f\hspace{0pt}x\hspace{0pt}g} = y`$.
The definition of the $`D\hspace{0pt}P`$ above generalises the definition of a less general version of the $`D\hspace{0pt}P`$ given in \[[6](#bib.bib6)\], \[[3](#bib.bib3)\] and \[[5](#bib.bib5)\]. The less general version only differs from the above definition of $`D\hspace{0pt}P`$ because $`G`$ is a group and $`A`$ is a subgroup. In our notation in all of this paper we omit the binary operation $`\ast`$ when writing products so for example $`f \ast x \ast g`$ is understood to mean $`f\hspace{0pt}x\hspace{0pt}g`$. We require that $`\ast`$ is efficiently computable.
The CSP (Conjugacy Search Problem) \[[1](#bib.bib1)\], \[[3](#bib.bib3)\] is defined as follows.
Public Information: $`G`$ is a group. $`{x,y} \in G`$ with $`y = {f^{- 1}\hspace{0pt}x\hspace{0pt}f}`$.
Secret Information: $`f \in G`$.
Objective: find an element $`g \in G`$ such that $`{g^{- 1}\hspace{0pt}x\hspace{0pt}g} = y`$.
The DH-DP (Diffie-Hellman Decomposition Problem) \[[6](#bib.bib6)\], \[[3](#bib.bib3)\] is defined as follows.
Public Information: $`G`$ is a group. $`A,B`$ are subgroups of $`G`$ with $`{\lbrack A,B\rbrack} = 1`$. $`{x,y_{a},y_{b}} \in G`$ with $`{y_{a} = {a\hspace{0pt}x\hspace{0pt}b}},`$ $`y_{b} = {c\hspace{0pt}x\hspace{0pt}d}`$.
Secret Information: $`{a,b} \in A`$, $`{c,d} \in B`$.
Objective: find the element $`cy_{a}d{( = ay_{b}b = acxbd)}.`$
The DH-CP (Diffie-Hellman Conjugacy Problem) is the specialisation of the DH-DP \[[6](#bib.bib6)\] with $`a = b^{- 1}`$ and $`c = d^{- 1}`$.
We now re-define the DP and DH-DP above as used in our authentication scheme. In the rest of this paper below the DP and DH-DP will mean their re-definitions.
The re-definition of the DP is as follows.
Public Information: $`G`$ is a semigroup. $`A`$, $`B`$ are subsets of $`G`$. $`{x,y} \in G`$ with $`y = {a\hspace{0pt}x\hspace{0pt}b}`$.
Secret Information: $`a \in A`$, $`b \in B`$.
Objective: find elements $`f \in A`$, $`g \in B`$ such that $`{f\hspace{0pt}x\hspace{0pt}g} = y`$.
The re-definition of the DH-DP is as follows.
Public information: $`G`$ is a semigroup. $`A,B,C,D`$ are subsets of $`G`$. $`{x,y_{a},y_{b}} \in G`$ with $`y_{a} = {a\hspace{0pt}x\hspace{0pt}b}`$, $`y_{b} = {c\hspace{0pt}x\hspace{0pt}d}`$.
Secret Information: $`a \in A`$, $`b \in B`$, $`c \in C`$, $`d \in D`$.
Objective: find the element $`c\hspace{0pt}y_{a}\hspace{0pt}d`$ $`({= {a\hspace{0pt}y_{b}\hspace{0pt}b} = {a\hspace{0pt}c\hspace{0pt}x\hspace{0pt}b\hspace{0pt}d}})`$. So if we can find $`f \in A`$, $`g \in B`$ such that $`{f\hspace{0pt}x\hspace{0pt}g} = y_{a}`$ or $`h \in C`$, $`i \in D`$ such that $`{h\hspace{0pt}x\hspace{0pt}i} = y_{b}`$ then this is sufficient to break our schemes which have security based upon the DH-DP.
A variant of the above re-definition of the DH-DP which we refer to as the DH-DP’ on which the security of our variant protocols is based upon is as follows.
Public information: $`G`$ is a semigroup. $`A,B,C,D`$ are subsets of $`G`$. $`{x,y_{a},y_{b}} \in G`$ with $`y_{a} = {a\hspace{0pt}x\hspace{0pt}b}`$, $`y_{b} = {c\hspace{0pt}a\hspace{0pt}x\hspace{0pt}b\hspace{0pt}d}`$.
Secret Information: $`a \in A`$, $`b \in B`$, $`c \in C`$, $`d \in D`$. ($`a,b`$ have an inverse).
Objective: find the element $`a^{- 1}y_{b}b^{- 1}{( = cxd,y_{a} = axb)}`$. So if we can find $`f \in A`$, $`g \in B`$ $`h \in C`$, $`i \in D`$ such that $`{h\hspace{0pt}x\hspace{0pt}i} = {c\hspace{0pt}x\hspace{0pt}d}`$ or $`{{f^{- 1}\hspace{0pt}y_{b}\hspace{0pt}g^{- 1}} = {c\hspace{0pt}x\hspace{0pt}d}},`$ $`{f\hspace{0pt}x\hspace{0pt}g} = y_{a}`$ then this is sufficient to break our schemes which have security based upon the DH-DP’.
In all of this paper for our authentication scheme the DP, DH-DP and DH-DP’ are considered with commutativity conditions such as (2) and (3) defined below. We assume the DP, DH-DP and DH-DP’ are hard. The connection between the DH-DP and the DP is similar to the one between the Diffie-Hellman problem and the discrete logarithm problem. The DH-DP, DH-DP’ is obviously reducible to the DP, but we assume that it is as hard for general $`G`$. Hence checking that the DP is hard for $`x`$ is supposed to ensure that the DH-DP also is.
The security of the modified key exchange protocol on page 2 of \[[4](#bib.bib4)\] is based on the DP with the additional condition that $`{\lbrack A,B\rbrack} = 1`$ and its security is also based on the DH-DP.
3\. The Sibert-Dehornoy-Girault Authentication Scheme
All the details of implementation in braid groups for the SDG authentication scheme $`I`$ are given in \[[10](#bib.bib10)\] so we do not reproduce them all here; we restrict to the details we require. For $`n \geq 2`$, $`B_{n}`$ is defined to be the group with the presentation with $`n - 1`$ generators (plus the identity $`e`$), denoted $`\sigma_{i}`$ for $`i = {1,2,\ldots,{n - 1}}`$ and the defining relationships
<table>
<tbody>
<tr>
<td></td>
<td>σiσj=σjσi|i−j|>1σiσjσi=σjσiσj|i−j|=1subscript𝜎𝑖subscript𝜎𝑗subscript𝜎𝑗subscript𝜎𝑖𝑖𝑗1subscript𝜎𝑖subscript𝜎𝑗subscript𝜎𝑖subscript𝜎𝑗subscript𝜎𝑖subscript𝜎𝑗𝑖𝑗1\begin{array}[t]{ll}\sigma_{i}\sigma_{j}=\sigma_{j}\sigma_{i}&|i-j|>1\\ \sigma_{i}\sigma_{j}\sigma_{i}=\sigma_{j}\sigma_{i}\sigma_{j}&|i-j|=1\end{array}</td>
<td></td>
<td>(1)</td>
</tr>
</tbody>
</table>
We refer the reader to any textbook about braids, for instance \[[13](#bib.bib13)\]; each element of $`B_{n}`$ has the geometrical interpretation by an $`n`$-strand braid in the usual sense. This geometrical interpretation is that any $`n`$-strand braid diagram can be first sliced into a concatenation of elementary diagrams with one crossing each and then each elementary diagram can be used to give an encoding of the braid diagram as a word in one of the letters $`\sigma_{i}`$ or $`\sigma_{i}^{- 1}`$. $`\sigma_{i}`$ is used for the diagram where the $`i`$th strand crosses under the $`({i + 1})`$ st one, and $`\sigma_{i}^{- 1}`$ is used for the diagram where the $`i`$th strand crosses over the $`({i + 1})`$st one.
$`L\hspace{0pt}B_{n}`$ and $`U\hspace{0pt}B_{n}`$ are the two commuting subgroups of $`B_{n}`$ generated by the Artin generators $`\sigma_{1},\ldots,\sigma_{{\lfloor{n/2}\rfloor} - 1}`$ and $`\sigma_{\lfloor{n/2}\rfloor},\ldots,\sigma_{n - 1}`$ \[[10](#bib.bib10)\]. The SDG authentication is as follows. Following \[[10](#bib.bib10)\] we require that the DH-CP and CSP are hard in $`G`$ as the security of the authentication is based on these problems, let $`G = B_{n}`$, $`A = {L\hspace{0pt}B_{n}}`$, $`B = {U\hspace{0pt}B_{n}}`$ in the DH-CP and CSP. Let $`G^{'}`$ be a non-abelian group, let $`w \in G^{'}`$ be a publicly known word which has a secret word as one of its factors (an attacker may attempt to recover this secret word from $`w`$), let $`S`$ be a function which maps $`w`$ to an equivalent word, $`S`$ is called a scrambling function, the security of the protocol is based on the difficulty of recovering the above secret word from the word $`S\hspace{0pt}{(w)}`$ \[[10](#bib.bib10)\]. A choice for the scrambling function can be a normal form for $`w`$ \[[10](#bib.bib10)\]. All braids are expressed in rewritten form using the scrambling function such that the DH-CP and CSP are hard.
- •
Phase 1. Key generation
i) Choose a public $`b \in B_{n}`$.
ii) $`A`$(lice) chooses a secret braid $`s \in {L\hspace{0pt}B_{n}}`$, her private key; she publishes $`b^{'} = {s\hspace{0pt}b\hspace{0pt}s^{- 1}}`$; the pair $`(b,b^{'})`$ is the public key.
- •
Phase 2. Authentication phase
i) $`B`$(ob) chooses $`r \in {U\hspace{0pt}B_{n}}`$ and sends the challenge $`t = {r\hspace{0pt}b\hspace{0pt}r^{- 1}}`$ to $`A`$.
ii) $`A`$ sends the response $`y = {h\hspace{0pt}{({s\hspace{0pt}t\hspace{0pt}s^{- 1}})}}`$ to $`B`$, and $`B`$ checks $`y = {h\hspace{0pt}{({r\hspace{0pt}b^{'}\hspace{0pt}r^{- 1}})}}`$.
$`h`$ is a fixed collision-free hash function from braids to sequences of 0’s and 1’s or, possibly, to braids, for which this choice for $`h`$, $`G`$ must have an efficient solution for the word problem for use in phase 2ii).
A proof that the above authentication scheme is a perfectly honest-verifier ZK interactive proof of knowledge of $`s`$ is given in \[[10](#bib.bib10)\]. There is a linear algebraic algorithm to solve the DH-CP but the attack is not efficient enough to break the public-key cryptosystem with the proposed parameters of \[[2](#bib.bib2)\] in real time \[[6](#bib.bib6)\]. The SDG scheme has a similar structure to the public-key cryptosystem in \[[2](#bib.bib2)\] because both algorithms are based on the DH-CP. Hence the parameters of the SDG scheme are based on considerations of the parameters of the above public-key cryptosystem \[[10](#bib.bib10)\]. Hence parameters can be chosen for the SDG scheme so that it cannot be broken in real time by using the attack on the DH-CP in \[[6](#bib.bib6)\].
4\. New Authentication Scheme
The new authentication scheme is as follows. Let $`G`$ be a (infinite or finite) non-commutative semigroup. We define the scrambling function as in the SDG scheme, for our scheme, but with the modification that it is defined over the semigroup $`G`$ instead of only a group. All elements in $`G`$ in this section are rewritten using the scrambling function and parameters are chosen such that the DH-DP and DP are hard, as the security of our authentication scheme is based on these problems.
- •
Phase 0. Initial setup
i) $`G`$ is chosen and is publicly known. The users publicly agree on which method first, second or third (described below) will be used to select the subsets and publicly agree on which of the commutativity conditions (2) or (3) will be used.
A first method to select the parameters is to select publicly known subsets $`L_{A}`$, $`L_{B}`$, $`R_{A}`$, $`R_{B}`$ and $`Z`$ of $`G`$ are chosen for which either property a) below is true or property b) below is true. Let $`z \in Z`$ with $`z`$ the publicly known element which is the value of $`x`$ in the definition of the DH-DP used in the example of the DH-DP in our new authentication scheme.
Following \[[5](#bib.bib5)\] let $`g \in G`$ for $`G`$ a group, $`C_{G}\hspace{0pt}{(g)}`$ is the centraliser of $`g`$, we describe the modifications to the authentication scheme (and these apply to the key agreement protocol described below) to give two further methods to select the subgroups as follows. Publicly known subsets or privately known $`L_{A}`$, $`L_{B}`$, $`R_{A}`$, $`R_{B}`$ and $`Z`$ of $`G`$ are chosen for which either property 2) below is true or property 3) below for the second and third methods below.
Using the above first method the security of the protocol is based on the DH-DP. We now give the two further methods to select the subsets which result in two modifications to the protocol when the first method is used. The second method to select the subgroups is $`A`$ chooses $`{(a_{1},a_{2})} \in {G \times G}`$ and publishes the subgroups as a set of generators of the centralisers $`{{L_{B},R_{B},L_{B}} \subseteq {C_{G}\hspace{0pt}{(a_{1})}}},{{R_{B} \subseteq {C_{G}\hspace{0pt}{(a_{2})}}},{L_{B} = {\{\alpha_{1},\ldots,\alpha_{k}\}}}}`$ etc. $`B`$ chooses $`{(b_{1},b_{2})} \in {L_{B} \times R_{B}}`$, and hence can compute $`x`$ below etc. Following \[[5](#bib.bib5)\] there is no explicit indication of where to select $`a_{1},b_{1}`$ and/or $`a_{2},b_{2}`$ from. Hence before attempting something like a length based attack in this case the attacker has to compute the centraliser of $`L_{B}`$ and/or $`R_{B}`$.
So a third method (this method is given for the key agreement protocol in \[[5](#bib.bib5)\] ) to select the subgroups is
$`A`$ chooses $`{L_{A} = G},{a_{1} \in G}`$, and publishes $`{L_{B} \subseteq {C_{G}\hspace{0pt}{(a_{1})}}},{L_{B} = {\{\alpha_{1},\ldots,\alpha_{k}\}}}`$,
$`B`$ chooses $`{L_{B} = G},{b_{2} \in G}`$, and publishes $`{R_{A} \subseteq {C_{G}\hspace{0pt}{(b_{2})}}},{R_{A} = {\{\beta_{1},\ldots,\beta_{k}\}}}`$,
Hence $`A`$ chooses $`{(a_{1},a_{2})} \in {{G \times C_{G}}\hspace{0pt}{(b_{2})}}`$ and publishes the subgroup(s) as a set of generators of the centralisers $`B`$ chooses $`{(b_{1},b_{2})} \in {{C_{G}\hspace{0pt}{(a_{1})}} \times G}`$, and hence can compute $`x`$ etc. Again there is no explicit indication of where to select $`a_{1}`$ and/or $`b_{2}`$ from. Hence before attempting a length based attack in this case the attacker has to compute the centraliser of $`L_{B}`$ and/or $`R_{B}`$.
The above three methods are examples of the following method, in general if a user $`A`$ or $`B`$ selects an element $`a_{i}`$ or $`b_{i}`$ respectively as their secret element (which multiplies the public element $`z`$ at the left WLOG) key then the other user selects an element of the $`C_{G}\hspace{0pt}{(a_{i})}`$ or $`C_{G}\hspace{0pt}{(b_{i})}`$ respectively as their secret key for the corresponding secret commuting element to left multiply WLOG, and in general an attacker has no explicit indication of where to select $`a_{i}`$ and/or $`b_{i}`$ from. A potential disadvantage of using the above third method to select the subgroups is one user chooses the other users subgroups and this may aid the user who has selected the subgroup to find the others users secret key, and this secret key may be of use in say another attack. Hence before attempting a length based attack in this case the attacker has to compute the centraliser of $`C_{G}\hspace{0pt}{(a_{i})}`$ and/or $`C_{G}\hspace{0pt}{(b_{i})}`$. Note if the subgroups are selected in this in the second or third way (using centraliser computations) then its security is based on a variant of the DP, DH-DP and the difficulty of computing centralisers, for example if the third method above is used to select the subsets then following the attack given in \[[5](#bib.bib5)\] the security of the protocol to find $`A`$ or $`B'\hspace{0pt}s`$ private key may be found as follows
Attack on $`A`$’s Key. Find an element $`a_{1}'`$ which commutes with every element of the subgroup $`L_{B}`$ and an element $`a_{2}' \in R_{B}`$ such that $`z' = {a_{1}'\hspace{0pt}z\hspace{0pt}a_{2}'}`$ where $`a_{1}'\hspace{0pt}z\hspace{0pt}a_{2}'`$ above may be rewritten using a normal form. The pair $`(a_{1}',a_{2}')`$ is equivalent to the pair $`{(a_{1},a_{2})},`$ because $`{a_{1}'\hspace{0pt}z\hspace{0pt}a_{2}'} = {a_{1}\hspace{0pt}z\hspace{0pt}a_{2}}`$ this means an attacker can authenticate as Alice. The attack applies to the key exchange protocol below with the modification $`K_{A} = {a_{1}'\hspace{0pt}z\hspace{0pt}a_{2}'}`$ instead of $`z' = {a_{1}'\hspace{0pt}z\hspace{0pt}a_{2}'}`$ this gives a equivalent secret key for $`A`$ used to get the common secret key.
Attack on $`B`$’s Key in the key exchange protocol in the section below. Find an element $`b_{2}'`$ which commutes with every element of the subgroup $`R_{A}`$ and an element $`b_{1}' \in L_{B}`$ such that $`K_{B} = {b_{1}'\hspace{0pt}z\hspace{0pt}b_{2}'}`$ where $`a_{1}'\hspace{0pt}z\hspace{0pt}a_{2}'`$ above may be rewritten using a normal form. The pair $`(b_{1}',b_{2}')`$ is equivalent to the pair $`{(b_{1},b_{2})},`$ because $`{b_{1}'\hspace{0pt}z\hspace{0pt}b_{2}'} = {b_{1}\hspace{0pt}z\hspace{0pt}b_{2}}`$ this means an attacker can find the common secret key when this set up is used as part of a key exchange algorithm as described below.
Then (following \[[5](#bib.bib5)\]) the most obvious way to recover Bob private key (the attack for $`A`$ is similar)
B1. Compute the centraliser of $`{R_{A},R_{A}} \subseteq {C_{G}\hspace{0pt}{(b_{2})}}`$.
B2. Solve the search version of the membership problem in the double coset $`{{<L_{B}>} \cdot z \cdot C_{G}}\hspace{0pt}{(R_{A})}`$.
So for the protocol to be secure we want both the above problems to be computationally hard, for the problem B2 to be hard it is required the centraliser $`C_{G}\hspace{0pt}{(L_{A})}`$ should be large enough to resist a brute force attack. The key exchange protocol in section 5 has security based upon the above problem. The above attack can be used to attack the authentication scheme or recover $`B`$’s secret elements with the modification $`x = {b_{1}'\hspace{0pt}z'\hspace{0pt}b_{2}'}`$ instead of $`K_{B} = {b_{1}'\hspace{0pt}z\hspace{0pt}b_{2}'}`$ ($`x`$ is used instead of $`K_{B}`$ and $`z'`$ is used instead of $`z`$ etc.).
The attacks above are considered with commutativity condition for the DH-DP such as (2) and (3) below. A similar attack is discussed when the second method is used to choose the subsets. Hence the platform group $`G`$ should satisfy the requirements given in the security analysis in section 6.
a\) If $`z \neq e`$ we require the following conditions
<table>
<tbody>
<tr>
<td></td>
<td>[LA,LB]=1,[RA,RB]=1,[LB,Z]≠1,[LA,Z]≠1,[RB,Z]≠1,[RA,Z]≠1,[LA,RA]≠1,[LB,RB]≠1.subscript𝐿𝐴subscript𝐿𝐵1subscript𝑅𝐴subscript𝑅𝐵1subscript𝐿𝐵𝑍1subscript𝐿𝐴𝑍1subscript𝑅𝐵𝑍1subscript𝑅𝐴𝑍1subscript𝐿𝐴subscript𝑅𝐴1subscript𝐿𝐵subscript𝑅𝐵1\begin{array}[t]{ll}[L_{A},L_{B}]=1,&[R_{A},R_{B}]=1,\\ [L_{B},Z]\neq 1,&[L_{A},Z]\neq 1,\\ [R_{B},Z]\neq 1,&[R_{A},Z]\neq 1,\\ [L_{A},R_{A}]\neq 1,&[L_{B},R_{B}]\neq 1.\end{array}</td>
<td></td>
<td>(2)</td>
</tr>
</tbody>
</table>
All the above conditions for $`z \neq e`$ can arise by generalising from properties of subgroups used in either the SDG scheme or CKLHC scheme for example the second and third conditions in (2) arise from the observations that in general $`{{\lbrack{L\hspace{0pt}B_{n}},B_{n}\rbrack} \neq 1},{{\lbrack{L\hspace{0pt}B_{n}},{U\hspace{0pt}B_{n}}\rbrack} = 1}`$.
b) If $`z = e`$ we require the following conditions
<table>
<tbody>
<tr>
<td></td>
<td>[LA,LB]=1,[RA,RB]=1,[LA,RA]≠1,[LB,RB]≠1,[LB,RA]≠1,[LA,RB]≠1.subscript𝐿𝐴subscript𝐿𝐵1subscript𝑅𝐴subscript𝑅𝐵1subscript𝐿𝐴subscript𝑅𝐴1subscript𝐿𝐵subscript𝑅𝐵1subscript𝐿𝐵subscript𝑅𝐴1subscript𝐿𝐴subscript𝑅𝐵1\begin{array}[t]{ll}[L_{A},L_{B}]=1,&[R_{A},R_{B}]=1,\\ [L_{A},R_{A}]\neq 1,&[L_{B},R_{B}]\neq 1,\\ [L_{B},R_{A}]\neq 1,&[L_{A},R_{B}]\neq 1.\end{array}</td>
<td></td>
<td>(3)</td>
</tr>
</tbody>
</table>
We were unable to show that the DH-DP using conditions (2) or (3) was easy. Hence we assume that the DH-DP is hard with condition with the above conditions. Note condition (3) is condition (2) but with conditions of the form of a subset not commuting with $`Z`$ omitted (because $`z = e`$) and the additional conditions $`{{\lbrack L_{B},R_{A}\rbrack} \neq 1},{{\lbrack L_{A},R_{B}\rbrack} \neq 1}`$. The above additional conditions are required so the DH-DP is not easy and hence our authentication scheme is secure.
- •
Phase 1. Key generation
i) Choose a public $`z \in Z`$.
ii) $`A`$ chooses secret elements $`a_{1} \in L_{A}`$, $`a_{2} \in R_{A}`$, her private key; she publishes $`z^{'} = {a_{1}\hspace{0pt}z\hspace{0pt}a_{2}}`$; the pair $`(z,z^{'})`$ is the public key.
- •
Phase 2. Authentication phase
i) $`B`$ chooses $`b_{1} \in L_{B}`$, $`b_{2} \in R_{B}`$ and sends the challenge $`x = {b_{1}\hspace{0pt}z\hspace{0pt}b_{2}}`$ to $`A`$.
ii) $`A`$ sends the response $`w = {H\hspace{0pt}{({a_{1}\hspace{0pt}x\hspace{0pt}a_{2}})}}`$ to $`B`$, and $`B`$ checks $`w = {H\hspace{0pt}{({b_{1}\hspace{0pt}z^{'}\hspace{0pt}b_{2}})}}`$.
$`H`$ is a fixed collision-free hash function from elements of $`G`$ to sequences of 0$`^{'}`$s and 1$`^{'}`$s or, possibly, to elements of $`G`$, for which this choice of $`H`$, $`G`$ must have an efficient solution for the word problem for use in phase 2ii)
Proposition 4.1
Our Authentication Scheme is a perfectly honest-verifier ZK interactive proof of knowledge of $`a_{1}`$ and $`a_{2}`$.
Proof. (Sketch) Completeness. Assume that, at step 2(ii) $`w^{'}`$ is sent by $`A`$. Then $`B`$ accepts $`A`$’s key if and only if $`w^{'} = {H\hspace{0pt}{({b_{1}\hspace{0pt}z^{'}\hspace{0pt}b_{2}})}}`$ which is equivalent to
<table>
<tbody>
<tr>
<td></td>
<td>w′=H(b1(a1za2)b2).superscript𝑤′𝐻subscript𝑏1subscript𝑎1𝑧subscript𝑎2subscript𝑏2w^{{}^{\prime}}=H(b_{1}(a_{1}za_{2})b_{2}).</td>
<td></td>
<td>(4)</td>
</tr>
</tbody>
</table>
By hypothesis $`a_{1} \in L_{A}`$, $`a_{2} \in R_{A}`$, $`b_{1} \in L_{B}`$, $`b_{2} \in R_{B}`$, so $`{{b_{1}\hspace{0pt}a_{1}} = {a_{1}\hspace{0pt}b_{1}}},{{b_{2}\hspace{0pt}a_{2}} = {a_{2}\hspace{0pt}b_{2}}}`$ holds and (4) is equivalent to $`w = {H\hspace{0pt}{({a_{1}\hspace{0pt}{({b_{1}\hspace{0pt}z\hspace{0pt}b_{2}})}\hspace{0pt}a_{2}})}}`$ i.e. $`w^{'} = w`$.
Soundness. Assume cheater $`A^{'}`$ is accepted with non-negligible probability. This means $`A^{'}`$ can compute $`H\hspace{0pt}{({b_{1}\hspace{0pt}z^{'}\hspace{0pt}b_{2}})}`$ with non-negligible probability. Since $`H`$ is supposed to be an ideal hash function, this means that $`A^{'}`$ can compute the element $`q`$ satisfying $`{H\hspace{0pt}{(q)}} = {H\hspace{0pt}{({b_{1}\hspace{0pt}z^{'}\hspace{0pt}b_{2}})}}`$ with non-negligible probability and this is because of two possibilities. The first possibility is that $`q = {b_{1}\hspace{0pt}z^{'}\hspace{0pt}b_{2}}`$ which contradicts the hypothesis that the DH-DP is hard. The second possibility is $`q \neq {b_{1}\hspace{0pt}z^{'}\hspace{0pt}b_{2}}`$ which means that $`A^{'}`$ and $`B`$ are able to find a collision for $`H`$ which contradicts that $`H`$ is a collision free hash function.
Honest-verifier zero knowledge. Consider the probabilistic Turing machine defined as follows: it chooses random elements $`b_{1}`$ and $`b_{2}`$ using the same drawing as the honest verifier, and outputs the instances $`(b_{1},b_{2},{H\hspace{0pt}{({b_{1}\hspace{0pt}z^{'}\hspace{0pt}b_{2}})}})`$. So the instances generated by this simulator follow the same probability distribution as the ones generated by the interactive pair $`(A,B)`$. $`\square`$
4.1 Comparison of Our Authentication Scheme with the
Sibert-Dehornoy-Girault Authentication Scheme
The generalisation (not the variants) specialises to the SDG scheme (hence our authentication scheme can be as secure as the SDG authentication scheme) with the parameters $`H = h`$, $`{G = B_{n}},{Z = B_{n}}`$, $`{L_{A} = R_{A} = {L\hspace{0pt}B_{n}}},`$ $`L_{B} = R_{B} = {U\hspace{0pt}B_{n}}`$, $`b_{1} = r`$, $`b_{2} = r^{- 1}`$, $`a_{1} = s`$, $`a_{2} = s^{- 1}`$, $`z = b`$, $`z^{'} = b^{'}`$, $`x = t`$ , $`w = y`$, the first method is used to select the subsets and conditions (2) (or property a) is true) are used.
If $`z = e`$ then this implies the following. Condition (3) do not allow the subgroups used in the SDG scheme to be used in our authentication scheme, because if these subgroups are used as the incorrect choices for $`L_{A}`$, $`L_{B}`$, $`R_{A}`$ and $`R_{B}`$ in our authentication scheme then it is easy to see the DH-DP is easy and hence our authentication scheme is insecure. There is more control of the public parameters (the choices of the subsets) compared to the SDG scheme and this may be useful for selecting secure public keys. Compared to the SDG scheme, potentially our authentication scheme requires less memory and fewer multiplications to compute the challenge of $`B`$ and the public key of $`A`$. This is because the identity element may take little memory (compared to $`b \neq e`$) to represent depending on the representation of $`G`$. Alternatively we can omit the implementation of using $`Z`$ completely because $`x`$ and $`z^{'}`$ can be computed using two multiplications.
4.2 A Variant of the Authentication Scheme
A variant of the above authentication scheme is as follows
- •
Phase 0. Initial Setup. The phase 0 is the same as phase 0 of the authentication scheme in section 4.
- •
Phase 1. Key generation
i) Choose a public $`z \in Z`$.
ii) $`A`$ chooses (invertible) secret elements $`a_{1} \in L_{A}`$, $`a_{2} \in R_{A}`$, her private key; she publishes $`z^{'} = {a_{1}\hspace{0pt}z\hspace{0pt}a_{2}}`$; the pair $`(z,z^{'})`$ is the public key.
- •
Phase 2. Authentication phase
i) $`B`$ chooses $`b_{1} \in L_{B}`$, $`b_{2} \in R_{B}`$ and sends the challenge $`x = {b_{1}\hspace{0pt}z'\hspace{0pt}b_{2}}`$ to $`A`$.
ii) $`A`$ sends the response $`w = {H\hspace{0pt}{({a_{1}^{- 1}\hspace{0pt}x\hspace{0pt}a_{2}^{- 1}})}}`$ to $`B`$, and $`B`$ checks $`w = {H\hspace{0pt}{({b_{1}\hspace{0pt}z\hspace{0pt}b_{2}})}}`$, if the check is true he accepts if the check is false he rejects.
$`H`$ is a fixed collision-free hash function from elements of $`G`$ to sequences of 0$`^{'}`$s and 1$`^{'}`$s or, possibly, to elements of $`G`$, for which this choice of $`H`$, $`G`$ must have an efficient solution for the word problem for use in phase 2ii). The above variant authentication scheme has security based on the DH-DP’.
The above variant authentication scheme specialises to the authentication scheme in \[[14](#bib.bib14)\]. The above variant protocol specialises to the authentication scheme in \[[14](#bib.bib14)\] with the parameters, conditions (3) are used, $`G`$ a finite non-abelian group, Bob is user $`B`$ in our protocol, Alice is user $`A`$ in our protocol $`z = {i\hspace{0pt}d\hspace{0pt}e\hspace{0pt}n\hspace{0pt}t\hspace{0pt}i\hspace{0pt}t\hspace{0pt}y}`$ $`e\hspace{0pt}l\hspace{0pt}e\hspace{0pt}m\hspace{0pt}e\hspace{0pt}n\hspace{0pt}t`$, $`L_{A},R_{B}`$ are publicly known and is generated by $`a`$, $`R_{B}`$ is generated by $`b`$, $`{L_{B} = L_{A}},{R_{B} = R_{A}}`$ $`B`$ selects the secret element which is depends on the secret exponents $`{0 < r < n_{1}},{0 < s < n_{2}}`$, $`{a^{r} \in L_{A}},{b^{s} \in R_{B}}`$, $`A`$ selects the secret element which is depends on the secret exponents $`{0 < v < n_{1}},{0 < w < n_{2}}`$, $`{a^{v} \in L_{A}},{b^{w} \in R_{B}}`$, so the common secret key is $`z^{'} = {a^{v}\hspace{0pt}b^{w}}`$, where the notation $`e,f`$,is used in \[[14](#bib.bib14)\].
5\. New Key Agreement Protocol
The setup for the authentication protocols in the above section can be used for key agreement as follows.
- •
Phase 0. Initial Setup. The phase 0 is the same as phase 0 of the authentication scheme in section 4.
- •
Choose $`z \in G`$. Phase 1.
ii) $`A`$(lice) chooses a secret elements $`a_{1} \in L_{A}`$, $`a_{2} \in R_{A}`$, her private key; she publishes $`K_{A} = {a_{1}\hspace{0pt}z\hspace{0pt}a_{2}}`$; the pair $`(z,K_{A})`$ is the public key.
i) $`B`$(ob) chooses a secret elements $`b_{1} \in L_{B}`$, $`b_{2} \in R_{B}`$, her private key; she publishes $`K_{B} = {b_{1}\hspace{0pt}z\hspace{0pt}b_{2}}`$; the pair $`(z,K_{B})`$ is the public key.
iii) $`A`$ and $`B`$ can compute the common shared secret key $`\kappa`$ as $`\kappa = {a_{1}\hspace{0pt}K_{B}\hspace{0pt}a_{2}}`$ and $`\kappa = {b_{1}\hspace{0pt}K_{A}\hspace{0pt}b_{2}}`$ respectively. Optionally the alternative computation $`\kappa = {h\hspace{0pt}{({a_{1}\hspace{0pt}K_{B}\hspace{0pt}a_{2}})}}`$ and $`\kappa = {h\hspace{0pt}{({b_{1}\hspace{0pt}K_{A}\hspace{0pt}b_{2}})}}`$ can be done.
$`h`$ is a fixed collision-free hash function from braids to sequences of 0’s and 1’s or, possibly, to braids, for which this choice for $`h`$. Again the above protocol is considered with the commutativity conditions 2 or 3. Note the elements $`K_{A}`$ and $`K_{B}`$ are rewritten for example a normal form to make the protocol secure.
Because phase 0 is the same as the authentication scheme again its security is based on a variant of the DP, DH-DP and the difficulty of computing centralisers.
Our protocol specialises to the CKLHC protocol in \[[8](#bib.bib8)\] is the above protocol with the parameters $`G`$ the braid group, $`A`$ and $`B`$ commuting defined by $`L\hspace{0pt}B_{n}`$ and $`U\hspace{0pt}B_{n}`$ are the two commuting subgroups of $`B_{n}`$ generated by the Artin generators $`\sigma_{1},\ldots,\sigma_{{\lfloor{n/2}\rfloor} - 1}`$ and $`{\sigma_{\lfloor{n/2}\rfloor},\ldots,\sigma_{n - 1}},`$ the first method in phase 0 is used and the DH-DP uses the condition (2). The publicly transmitted information is rewritten using the left canonical form. The KLCHKP protocol \[[2](#bib.bib2)\] is the specialisation of the CKLHC protocol with $`a_{2} = a_{1}^{- 1}`$ and $`b_{2} = b_{1}^{- 1}`$ and hence our protocol generalises the KLCHKP protocol.
Our protocol specialises to the key agreement protocol given on page two of \[[4](#bib.bib4)\] with the parameters $`L_{A} = R_{B}`$, $`L_{B} = R_{A}`$, $`{{\lbrack L_{A},L_{B}\rbrack} = 1},`$ and commutativity condition (2), $`G`$ is a semi-group or $`G`$ is the Thompson group, the first method in phase 0 is used and hence our protocol generalise the protocol given in \[[4](#bib.bib4)\]. Using the notation used in \[[4](#bib.bib4)\] this is $`{L_{A} = A},{{L_{B} = B},{z = w}}`$. So the left and right secrets are taken from different subgroups the protocol.
Our protocol specialises to the key agreement protocol given in \[[5](#bib.bib5)\] when $`G`$ is a group, $`A`$ and $`B`$ are subgroups in the parameters of the DH-DP, using the third method in phase 0 to select the subsets/ modify the key agreement protocol described above, commutativity condition 2, $`{L_{B} \subseteq {C_{G}\hspace{0pt}{(a_{1})}}},{{L_{B} = {\{\alpha_{1},\ldots,\alpha_{k}\}}},{{R_{A} \subseteq {C_{G}\hspace{0pt}{(b_{2})}}},{R_{A} = {{\{\beta_{1},\ldots,\beta_{k}\}},K_{A}}}}}`$ and $`K_{B}`$ are rewritten using a normal form, using the notation given in \[[5](#bib.bib5)\] this is $`{{K_{A} = P_{A}},{{K_{B} = P_{B}},{{B = L_{B}},{A = R_{A}}}}}.`$
Hence we get new variants (which use less multiplications) of the protocols \[[4](#bib.bib4)\], \[[5](#bib.bib5)\], \[[2](#bib.bib2)\],\[[8](#bib.bib8)\] if we consider the parameters that our key agreement protocol above specialises to the protocols above but using condition (3) instead of conditions (2), for example if we consider the second method in phase 0 to modify the protocol with the parameters above to specialise, then compared to using the third method in phase 0 the advantage of this protocol is that user $`A`$ does not have to wait for $`B`$ to select the subgroup and so for example can create the certificate to prove that the public key of $`A`$ belongs to $`A`$.
5.1 Variant of Key Exchange
A variant of the key exchange is as follows
The above setup can be used for key agreement as follows the details are as follows.
- •
Phase 0. Initial Setup. The phase 0 is the same as phase 0 of the authentication scheme in section 4.
- •
Choose $`z \in G`$.
ii) $`A`$(lice) chooses a secret invertible elements $`a_{1} \in L_{A}`$, $`a_{2} \in R_{A}`$, her private key; she publishes $`K_{A} = {a_{1}\hspace{0pt}z\hspace{0pt}a_{2}}`$; the pair $`(z,K_{A})`$ is the public key.
i) $`B`$(ob) chooses a secret braid $`b_{1} \in L_{B}`$, $`b_{2} \in R_{B}`$, his private key; he publishes $`K_{B} = {b_{1}\hspace{0pt}K_{A}\hspace{0pt}b_{2}}`$; the pair $`(K_{A},K_{B})`$ is the public key.
iii) $`A`$ and $`B`$ can compute the common shared secret key $`\kappa`$ as $`\kappa = {a_{1}^{- 1}\hspace{0pt}K_{B}\hspace{0pt}a_{2}^{- 1}}`$ and $`\kappa = {b_{1}\hspace{0pt}z\hspace{0pt}b_{2}}`$ respectively. Optionally the alternative computation $`\kappa = {h\hspace{0pt}{({a_{1}^{- 1}\hspace{0pt}K_{B}\hspace{0pt}a_{2}^{- 1}})}}`$ and $`\kappa = {h\hspace{0pt}{({b_{1}\hspace{0pt}K_{A}\hspace{0pt}b_{2}})}}`$ can be done.
$`h`$ is a fixed collision-free hash function from braids to sequences of 0’s and 1’s or, possibly, to braids, for which this choice for $`h`$. Again the above protocol is considered with the commutativity conditions 2 or 3. Note the elements $`K_{A}`$ and $`K_{B}`$ are rewritten for example a normal form to make the protocol secure.
Again because phase 0 is the same as the authentication scheme we have sketched its security is based on a variant of the DP, DH-DP’ and/or the difficulty of computing centralisers.
The above variant protocol specialises to the key exchange in \[[9](#bib.bib9)\] with the parameters, conditions (2) or (3) are used, $`G`$ a finite non-abelian group, Bob is user $`B`$ in our protocol, Alice is user $`A`$ in our protocol, $`z = e`$, $`L_{A}`$ is publicly known and is generated by $`a`$, $`R_{B}`$ is generated by $`b`$, $`{L_{B} = L_{A}},{R_{B} = R_{A}}`$ $`B`$ selects the secret element which is depends on the secret exponents $`{0 < r < n_{1}},{0 < s < n_{2}}`$, $`{a^{r} \in L_{A}},{b^{s} \in R_{B}}`$, $`A`$ selects the secret element which is depends on the secret exponents $`{0 < v < n_{1}},{0 < w < n_{2}}`$, $`{a^{v} \in L_{A}},{b^{w} \in R_{B}}`$, so the common secret key is $`\kappa = f = {a^{v}\hspace{0pt}b^{w}}`$, where the notation $`e,f`$ is used in \[[9](#bib.bib9)\].
5.1.1 Groups With No Efficient Normal Form
If there is no efficient algorithm for a canonical form in $`G`$ then the secret elements such as Alice’s secret key must be disguised in another way such as using a scrambling function. The algorithm in \[[1](#bib.bib1)\] for when $`G`$ is a group, and there is no efficient normal form in $`G`$ but there is an efficient algorithm for the word problem in $`G`$, can be used to find a common shared secret key can be found as follows using the algorithm in \[[1](#bib.bib1)\] which is. Fix $`b`$ as $`b = 0`$ or $`b = 1`$.
1\. $`B`$ sends a rewritten form of $`\kappa`$ which is $`r`$ for $`r = \kappa`$ or a random word for $`{r,r} \neq \kappa`$.
2\. User $`A`$ checks if $`k = r`$ then this determines the bit $`b`$, otherwise the bit is $`1 - b`$.
3\. The steps 1 and 2 are repeated $`m`$ times so an $`m`$ bit key is exchanged.
As stated in \[[1](#bib.bib1)\] the protocol is probabilistic and slower compared to using a canonical form.
6\. Security Analysis
The attacks below are considered with commutativity condition for the DH-DP such as (2) and (3) below.
6.1 Attacks When Second Method is used to Choose Subsets
in Authentication Scheme and Key Exchange Scheme
If the second method above is used to select the subsets then following the attack given in \[[5](#bib.bib5)\] the security of the protocol to find $`A`$ or $`B'\hspace{0pt}s`$ private key may be found as follows
Attack on $`A`$’s Key. Find an element $`a_{1}'`$ which commutes with every element of the subgroup $`L_{B}`$ and an element $`a_{2}' \in R_{B}`$ which commutes with every element of the subgroup $`R_{B}`$ such that $`z' = {a_{1}'\hspace{0pt}z\hspace{0pt}a_{2}'}`$ where $`a_{1}'\hspace{0pt}z\hspace{0pt}a_{2}'`$ above may be rewritten using a normal form. The pair $`(a_{1}',a_{2}')`$ is equivalent to the pair $`{(a_{1},a_{2})},`$ because $`{a_{1}'\hspace{0pt}z\hspace{0pt}a_{2}'} = {a_{1}\hspace{0pt}z\hspace{0pt}a_{2}}`$ this means an attacker can authenticate as Alice. The attack applies to the key exchange protocol (when the second method is used to choose the subsets) with the modification $`K_{A} = {a_{1}'\hspace{0pt}z\hspace{0pt}a_{2}'}`$ instead of $`z' = {a_{1}'\hspace{0pt}z\hspace{0pt}a_{2}'}`$ this gives a equivalent secret key for $`A`$ used to get the common secret key.
Then (following \[[5](#bib.bib5)\]) the most obvious way to do the above attack Alice’s private key
A1. Compute the centraliser of $`{R_{B},R_{B}} \subseteq {C_{G}\hspace{0pt}{(a_{2})}}`$ and compute the centraliser of $`{L_{B},L_{B}} \subseteq {C_{G}\hspace{0pt}{(a_{1})}}`$
A2. Solve the search version of the membership problem in the double coset $`{<{C_{G}\hspace{0pt}{(L_{B})}}>} \cdot z \cdot {<{C_{G}\hspace{0pt}{(R_{B})}}>}`$.
So for the protocol to be secure we want both the above problems to be computationally hard, for the problem A1 to be hard it is required both the centralisers $`C_{G}\hspace{0pt}{(L_{B})}`$ and $`C_{G}\hspace{0pt}{(R_{B})}`$be large enough to resist a brute force type attack. The above attack can be used to attack the authentication scheme or recover $`A`$’s secret elements with the modification $`z^{'} = {a_{1}'\hspace{0pt}z\hspace{0pt}a_{2}'}`$ instead of $`K_{A} = {a_{1}'\hspace{0pt}z\hspace{0pt}a_{2}'}`$ ($`z'`$ is used instead of $`K_{A}`$ etc.) so this can be used to impersonate $`A`$.
Attack on $`B`$’s Key in the key exchange protocol. Find an element $`b_{1}'`$ $`\in L_{B}`$ and an element $`b_{2}' \in R_{B}`$ such that $`K_{B} = {b_{1}'\hspace{0pt}z\hspace{0pt}b_{2}'}`$ where $`b_{1}'\hspace{0pt}z\hspace{0pt}b_{2}'`$ above may be rewritten using a normal form. The pair $`(b_{1}',b_{2}')`$ is equivalent to the pair $`{(b_{1},b_{2})},`$ because $`{b_{1}'\hspace{0pt}z\hspace{0pt}b_{2}'} = {b_{1}\hspace{0pt}z\hspace{0pt}b_{2}}`$ this means an attacker can find the common secret key when this set up is used as part of a key exchange algorithm as described below.
Then (following \[[5](#bib.bib5)\]) the most obvious way to do the above attack, Bob private key
B1. Solve the search version of the membership problem in the double coset $`{<L_{B}>} \cdot z \cdot {<R_{B}>}`$.
So for the protocol to be secure we want the above problem to be computationally hard, for the problem B1 to be hard it is required that the elements cannot of $`L_{B}`$, $`R_{B}`$ all be tested (for an equivalent key pair $`b_{1}',b_{2}'`$) so to resist a brute force attack the possible values for $`b_{1}',b_{2}'`$ should be large enough. The above attack can be used to attack the authentication scheme or recover $`B`$’s secret elements with the modification $`x' = {b_{1}'\hspace{0pt}z'\hspace{0pt}b_{2}'}`$ instead of $`K_{B} = {b_{1}'\hspace{0pt}z\hspace{0pt}b_{2}'}`$ ($`x'`$ is used instead of $`K_{B}`$ and $`z'`$ is used instead of $`z`$ etc.) so this can be used to impersonate $`A`$.
6.2 Attack When Second Method is used to Choose Subsets in the
Variant Authentication Scheme and Variant Key Exchange Scheme
If the second method above is used to select the subsets then following variant of the attack given in \[[5](#bib.bib5)\] the security of the protocol to find $`A`$ or $`B'\hspace{0pt}s`$ private key may be found as follows
Attack on $`A`$’s Key. Find an element $`a_{1}'`$ which commutes with every element of the subgroup $`L_{B}`$ and an element $`a_{2}' \in R_{B}`$ which commutes with every element of the subgroup $`R_{B}`$ such that $`z' = {a_{1}'\hspace{0pt}z\hspace{0pt}a_{2}'}`$ where $`a_{1}'\hspace{0pt}z\hspace{0pt}a_{2}'`$ above may be rewritten using a normal form and $`a_{1}',a_{2}'`$ are both invertible elements.
The pair $`(a_{1}',a_{2}')`$ is equivalent to the pair $`{(a_{1},a_{2})},`$ because $`{a_{1}'\hspace{0pt}z\hspace{0pt}a_{2}'} = {a_{1}\hspace{0pt}z\hspace{0pt}a_{2}}`$ this means an attacker can authenticate as Alice because the attacker can compute $`a_{1}^{- 1}\hspace{0pt}x\hspace{0pt}a_{2}^{- 1}`$. The above attack applies to the key exchange protocol with the modification $`K_{A} = {a_{1}'\hspace{0pt}z\hspace{0pt}a_{2}'}`$ instead of $`z' = {a_{1}'\hspace{0pt}z\hspace{0pt}a_{2}'}`$ this gives an equivalent secret key for $`A`$ used to get the common secret key which can be computed as $`\kappa = {a_{1}^{- 1}\hspace{0pt}K_{B}\hspace{0pt}a_{2}^{- 1}}`$. (Another attack call this attack B, is to find elements $`a_{1}',a_{2}'`$ such that $`{a_{1}^{'{- 1}}\hspace{0pt}z'\hspace{0pt}a_{2}^{'{- 1}}} = z`$ , $`a_{1}',a_{2}'`$ can be used instead of $`a_{1},a_{2}`$ to get the common secret key, the attack is similar for the key agreement protocol with $`K_{A}`$ used in place of $`z^{'}`$).
Then (following \[[5](#bib.bib5)\]) the most obvious way to do the above attack Alice’s private key
A1. Compute the centraliser of $`{R_{B},R_{B}} \subseteq {C_{G}\hspace{0pt}{(a_{2})}}`$ and compute the centraliser of $`{L_{B},L_{B}} \subseteq {C_{G}\hspace{0pt}{(a_{1})}}`$
A2. Solve the search version of the membership problem in the double coset $`{<{C_{G}\hspace{0pt}{(L_{B})}}>} \cdot z \cdot {<{C_{G}\hspace{0pt}{(R_{B})}}>}`$ and/or $`{<{C_{G}\hspace{0pt}{(L_{B})}}>} \cdot K_{B} \cdot {<{C_{G}\hspace{0pt}{(R_{B})}}>}`$ .
(For attack B we do the search $`{<{C_{G}\hspace{0pt}{(L_{B})}}>} \cdot z^{'} \cdot {<{C_{G}\hspace{0pt}{(R_{B})}}>}`$ there is a variant of attack B when the third method is used to choose the subsets doing a search of the form $`{<L_{B}>} \cdot z^{'} \cdot {<{C_{G}\hspace{0pt}{(R_{B})}}>}`$) So for the protocol to be secure we want both the above problems to be computationally hard, for the problem A1 to be hard it is required both the centralisers $`C_{G}\hspace{0pt}{(L_{B})}`$ and $`C_{G}\hspace{0pt}{(R_{B})}`$ to be large to resist a brute force type attack.
Attack on $`B`$’s key in the key exchange protocol. Find an element $`b_{1}'`$ $`\in L_{B}`$ and an element $`b_{2}' \in R_{B}`$ such that $`K_{B} = {b_{1}'\hspace{0pt}z\hspace{0pt}b_{2}'}`$ where $`b_{1}'\hspace{0pt}z\hspace{0pt}b_{2}'`$ above may be rewritten using a normal form. The pair $`(b_{1}',b_{2}')`$ is equivalent to the pair $`{(b_{1},b_{2})},`$ because $`{b_{1}'\hspace{0pt}z\hspace{0pt}b_{2}'} = {b_{1}\hspace{0pt}z\hspace{0pt}b_{2}}`$ this means an attacker can find the common secret key when this set up is used as part of a key exchange algorithm as described below.
Then (following \[[5](#bib.bib5)\]) the most obvious way to do the above attack, Bob private key
B1. Solve the search version of the membership problem in the double coset $`{<L_{B}>} \cdot z \cdot {<R_{B}>}`$ and/or $`{<L_{B}>} \cdot K_{A} \cdot {<R_{B}>}`$.
So for the protocol to be secure we want the above problem to be computationally hard, for the problem B1 to be hard it is required that the elements of $`L_{B}`$,$`R_{B}`$ cannot all be tested (for an equivalent key, pair $`b_{1}',b_{2}'`$) so to resist a brute force attack. The above attack can be used to attack the authentication scheme or recover $`B`$’s secret elements with the modification $`x = {b_{1}'\hspace{0pt}z'\hspace{0pt}b_{2}'}`$ instead of $`K_{B} = {b_{1}'\hspace{0pt}z\hspace{0pt}b_{2}'}`$ ($`x'`$ is used instead of $`K_{B}`$ and $`z'`$ is used instead of $`z`$ etc.) so this can be used to impersonate $`A`$.
6.3 Attack When Third Method is used to Choose Subsets in the
Variant Authentication Scheme and Variant Key Exchange Scheme
Attack on $`A`$’s Key. Find an element $`a_{1}'`$ which commutes with every element of the subgroup $`L_{B}`$ and an element $`a_{2}' \in R_{B}`$ such that $`z' = {a_{1}'\hspace{0pt}z\hspace{0pt}a_{2}'}`$ where $`a_{1}'\hspace{0pt}z\hspace{0pt}a_{2}'`$ above may be rewritten using a normal form and $`a_{1},a_{2}`$ are both invertible elements. The pair $`(a_{1}',a_{2}')`$ is equivalent to the pair $`{(a_{1},a_{2})},`$ because $`{a_{1}'\hspace{0pt}z\hspace{0pt}a_{2}'} = {a_{1}\hspace{0pt}z\hspace{0pt}a_{2}}`$ this means an attacker can authenticate as Alice the attack applies to the key exchange protocol with the modification $`K_{A} = {a_{1}'\hspace{0pt}z\hspace{0pt}a_{2}'}`$ instead of $`z' = {a_{1}'\hspace{0pt}z\hspace{0pt}a_{2}'}`$ is solved this gives a equivalent secret key for $`A`$ used to get the common secret key by computing $`\kappa = {a_{1}^{'{- 1}}\hspace{0pt}K_{B}\hspace{0pt}a_{2}^{'{- 1}}}`$.
Then (following \[[5](#bib.bib5)\]) the most obvious way to do the above attack Alice’s private key
A1. Compute the centraliser of $`{L_{A},L_{A}} \subseteq {C_{G}\hspace{0pt}{(a_{1})}}`$.
A2. Solve the search version of the membership problem in the double coset $`{<{C_{G}\hspace{0pt}{(L_{A})}}>} \cdot z \cdot {<R_{A}>}`$ and/or $`{<{C_{G}\hspace{0pt}{(L_{A})}}>} \cdot K_{B} \cdot {<R_{A}>}`$ .
Attack on $`B`$’s key in the key exchange protocol is below. Find an element $`b_{2}'`$ which commutes with every element of the subgroup $`R_{A}`$ and an element $`b_{1}' \in L_{B}`$ such that $`K_{B} = {b_{1}'\hspace{0pt}z\hspace{0pt}b_{2}'}`$ where $`b_{1}'\hspace{0pt}z\hspace{0pt}b_{2}'`$ above may be rewritten using a normal form. The pair $`(b_{1}',b_{2}')`$ is equivalent to the pair $`{(b_{1},b_{2})},`$ because $`{b_{1}'\hspace{0pt}z\hspace{0pt}b_{2}'} = {b_{1}\hspace{0pt}z\hspace{0pt}b_{2}}`$ this means an attacker can find the common secret key.
Then (following \[[5](#bib.bib5)\]) the most obvious way to recover Bob private key
B1. Compute the centraliser of $`{R_{A},R_{A}} \subseteq {C_{G}\hspace{0pt}{(b_{2})}}`$.
B2. Solve the search version of the membership problem in the double coset $`{{<L_{B}>} \cdot z \cdot C_{G}}\hspace{0pt}{(R_{A})}`$.
So for the protocol to be secure we want both the above problems to be computationally hard, for the problem B2 to be hard it is required the centraliser $`C_{G}\hspace{0pt}{(R_{A})}`$ to be large enough to resist a brute force attack. The key exchange protocol in section 5 has security based upon the above problem. The above attack can be used to attack the authentication scheme or recover $`B`$’s secret elements with the modification $`x = {b_{1}'\hspace{0pt}z'\hspace{0pt}b_{2}'}`$ instead of $`K_{B} = {b_{1}'\hspace{0pt}z\hspace{0pt}b_{2}'}`$ is solved ($`x`$ is used instead of $`K_{B}`$ and $`z'`$ is used instead of $`z`$ etc.) so this can be used to impersonate Alice.
6.4 Requirements of the of Platform Group
Hence the platform group $`G`$ should satisfy at least the following properties in order for our key establishment protocol to be efficient and secure, we have taken the requirements from \[[5](#bib.bib5)\] so the properties are the same as given in \[[5](#bib.bib5)\] with the relevant modifications. At least one of property P7, P8, or P9 is true depending on the choice of protocol used.
(P1) $`G`$ should be a non-commutative group of at least exponential growth. The latter means that the number of elements of length $`n`$ in $`G`$ is at least exponential in $`n`$; this is needed to prevent attacks brute force type attacks on the key space.
(P2) This property may be optional. There should be an efficiently computable normal form for elements of $`G`$.
(P3) It should be computationally easy to perform group operations (multiplication and inversion) on normal forms.
(P4) It should be computationally easy to generate pairs $`(a,{\{ a_{1},\ldots,a_{k}\}})`$ such that $`{a\hspace{0pt}a_{i}} = {a_{i}\hspace{0pt}a}`$ for each $`i = {1,\ldots,k}`$. (Clearly, in this case the subgroup generated by $`a_{1},\ldots,a_{k}`$ centralizes a).
(P5) For a generic set $`\{ g_{1},\ldots,g_{k}\}`$ of elements of $`G`$ it should be difficult to compute $`{C\hspace{0pt}{(g_{1},\ldots,g_{n})}} = {{C\hspace{0pt}{(g_{1})}} \cap \ldots \cap {C\hspace{0pt}{(g_{k})}}}`$.
(P6) This property may be optional. Even if $`H = {C\hspace{0pt}{(g_{1},\ldots,g_{n})}}`$ is computed, it should be hard to find $`x \in H`$ and $`y \in H_{1}`$ (where $`H_{1}`$ is some fixed subgroup given by a generating set) such that $`{x\hspace{0pt}w\hspace{0pt}y} = w^{'}`$, i.e., to solve the membership search problem for a double coset.
(P7) This property may be optional. Even if $`H = {C\hspace{0pt}{(g_{1},\ldots,g_{n})}}`$ is computed, and $`H_{1} = {C\hspace{0pt}{(g_{1}',\ldots,g_{m}')}}`$ ($`g'`$ is a generator as usual) is computed and it should be hard to find $`x \in H`$ and $`y \in H_{1}`$ such that $`{x\hspace{0pt}w\hspace{0pt}y} = w^{'}`$, i.e., to solve the membership search problem for a double coset.
(P8) This property may be optional. Given $`H`$ and $`H_{1}`$ is some fixed subgroups given by a generating sets,and it should be hard to find $`x \in H`$ and $`y \in H_{1}`$ such that $`{x\hspace{0pt}w\hspace{0pt}y} = w^{'}`$, i.e., to solve the membership search problem for a double coset.
(P9)This property may be optional. Which is there should be an efficiently algorithm for the word problem in $`G`$.
6.5 Braid groups
We now consider braid groups as a possible platform group. Here we consider the properties (P1)-(P6) from the previous section.
(P1) For $`n > 2`$, braid groups $`B_{n}`$ are non-commutative groups of exponential growth.
(P2) There are several known normal forms for elements of B<sub>n</sub>, including Garside normal form (see \[[8](#bib.bib8)\]). The Garside normal form is efficiently computable.
(P3) There are efficient algorithms to multiply or invert normal forms of elements of $`B_{n}`$ \[[8](#bib.bib8)\].
(P4) It is not so easy to compute the whole centralizer of an element $`g`$ of $`G`$ \[[5](#bib.bib5)\]. The number of steps required to compute $`C_{G}\hspace{0pt}{(g)}`$ is proportional the size of the SSS (super summit set) of $`g`$. Generally speaking the “super summit set” is not of polynomial size in $`n`$ and the braid length. Nevertheless, there are approaches to finding “large parts” of $`C_{G}\hspace{0pt}{(g)}`$, e.g. one can generate a sufficiently large part of $`S\hspace{0pt}S\hspace{0pt}S\hspace{0pt}{(g)}`$.
(P5) For a generic subgroup $`A`$ there is no efficient algorithm to compute $`C_{G}\hspace{0pt}{(A)}`$.
(P6),(P7) and (P8) There is no known solution to the membership search problem for double cosets $`H\hspace{0pt}w\hspace{0pt}H'`$ in braid groups.
(P9) There are efficient algorithms for the word problem in braid groups, such as the practical handle reduction algorithm for the word problem described in \[[10](#bib.bib10)\].
7\. Conclusion
We have presented new two-pass authentication schemes and key exchange protocols. This paper is a work in progress, because further work we plan to do for our authentication scheme is to investigate potential semigroups (apart from braid groups) and parameters for which it is secure, when its security is based on the DH-DP or variants of the DH-DP.
## References
- \[1\] I. Anshel, M. Anshel, D. Goldfeld. An Algebraic Method for Public-key Cryptography, Mathematical Research Letters, 6, 1999, pp. 287-292.
- \[2\] K.K. Ko, S.J. Lee, J.H. Cheon, J.W. Han, J.S. Kang, C. Park. New public-key cryptosystem using braid groups. CRYPTO 2000. LNCS, 1880, 2000, pp. 166-183.
- \[3\] K. H. Ko, Tutorial on Braid Cryptosystems 3, PKC 2001, Korea, February 13-15, 2001. Available at www.ipkc.org/pre_conf/pkc2001/PKCtp_ko.ps
- \[4\] V. Shpilrain and A. Ushakov, “Thompson’s group and public key cryptography”, LNCS, 3531, Springer-Verlag, 2005, pp. 151-164.
- \[5\] V. Shpilrain and A. Ushakov, A new key exchange protocol based on the decomposition problem, available at http://eprint.iacr.org/2005/447.pdf (Accessed 20/05/2006)
- \[6\]
J. Cheon and B. Jun, “A Polynomial Time Algorithm for the Braid Diffie Hellman Conjugacy Problem”, Proceedings of Crypto 2003, LNCS, 2729, Springer-Verlag , 2003, pp. 212-215.
- \[7\] T. Thomas and A. K. Lal, Group Signature Schemes Using Braid Groups, arXiv:cs.CR/0602063, 2006
- \[8\] J.C. Cha, K.H. Ko, S.J. Lee, J.W. Han, J.H. Cheon, An efficient implementation of braid groups, Advances in cryptology: Proceedings of ASIACRYPT 2001,Lecture Notes in Computer Science, Springer-Verlag, 2248, pp.144-156, 2001
- \[9\] E. Stickel, A New Method for Exchanging Secrets, Proceedings of the Third International Conference on Information Technology and Applications, 2005
- \[10\]
H. Sibert, P. Dehornoy, M. Girault, “Entity Authentication Schemes Using Braid Word Reduction”, Discrete Applied Mathematics, 154, 2006, pp. 420-436
- \[11\] Lal and Chatervedi. Authentication schemes using braid groups, arxiv preprint, July 2005
- \[12\] B. Tsaban, On an authentication scheme based on the root problem in the braid groups, ArXiv preprint, September 2005
- \[13\]
K. Murasugi and B. Kurpita, A Study of Braids, Kluwer Academic Publishers, The Netherlands, 1999.
- \[14\] Stickel, A new public-key cryptosystem non-abelian groups, In proceedings of the 13th International Conference on Information Systems Development. Vilnius Technika, Vilnius 2004, September 2004 |
0710.5238 | cs.NI | cyber | # On The Linear Behaviour of the Throughput of IEEE 802.11 DCF in Non-Saturated Conditions
F. Daneshgaran, M. Laddomada, F. Mesiti, and M. Mondin F. Daneshgaran is with ECE Dept., CSU, Los Angeles, USA.M. Laddomada, F. Mesiti, and M. Mondin are with DELEN, Politecnico di Torino, Italy.
###### Abstract
We propose a linear model of the throughput of the IEEE 802.11 Distributed Coordination Function (DCF) protocol at the data link layer in non-saturated traffic conditions. We show that the throughput is a linear function of the packet arrival rate (PAR) $`\lambda`$ with a slope depending on both the number of contending stations and the average payload length. We also derive the interval of validity of the proposed model by showing the presence of a critical $`\lambda`$, above which the station begins operating in saturated traffic conditions.
The analysis is based on the multi-dimensional Markovian state transition model proposed by Liaw et al. with the aim of describing the behaviour of the MAC layer in unsaturated traffic conditions. Simulation results closely match the theoretical derivations, confirming the effectiveness of the proposed linear model.
###### Index Terms:
DCF, Distributed Coordination Function, IEEE 802.11, MAC, saturation, throughput, unsaturated.
## I Introduction
Modelling of the DCF at the MAC layer of the series of IEEE 802.11 standards has recently garnered interest in the scientific community \[[1](#bib.bib1)\]-\[[6](#bib.bib6)\]. After the seminal work by Bianchi \[[1](#bib.bib1)\] who proposed a bi-dimensional Markov model of the back-off stage procedure adopted by the DCF in saturated conditions, many papers have focused on various facets of basic access mechanism providing extensions to most recent versions of the IEEE 802.11 series of standards \[[7](#bib.bib7)\]. Recently, in \[[3](#bib.bib3)\] the authors proposed a novel fixed-point analysis of the DCF providing an effective framework for analyzing single cell IEEE 802.11 WLANs without resorting to the bi-dimensional contention model \[[1](#bib.bib1)\].
Practical networks usually operate in non-saturated conditions and data traffic is mainly bursty. Under these operating conditions, Bianchi’s model does not describe accurately the behaviour of the throughput at the MAC layer. In this respect, in \[[4](#bib.bib4)\]-\[[5](#bib.bib5)\] the authors proposed two different bi-dimensional Markov models accounting for unsaturated traffic conditions, extending the basic bi-dimensional model proposed in \[[1](#bib.bib1)\].
In this paper we take a different approach with respect to works \[[4](#bib.bib4)\]-\[[5](#bib.bib5)\]. Upon starting from the bi-dimensional model proposed by Liaw et al. in \[[4](#bib.bib4)\], we show that the behaviour of the throughput of the IEEE 802.11 DCF in unsaturated conditions can be described by a linear relation that, with respect to the PAR $`\lambda`$, depends on two network parameters: the number $`N`$ of contending stations, and the average size $`E\hspace{0pt}{\lbrack{P\hspace{0pt}L}\rbrack}`$ of the transmitted packets. This is one of the key contribution of the paper: no simulations are needed for throughput evaluation since it can be theoretically predicted employing the model $`{S\hspace{0pt}{(\lambda)}} = {{N \cdot E}\hspace{0pt}{\lbrack{P\hspace{0pt}L}\rbrack}\hspace{0pt}\lambda}`$ developed in Section III. Of course, the limit of validity of such a model has to be clearly identified, and it represents another contribution of this paper. To this end, we derive the interval of validity of the proposed model with respect to the PARs at the MAC layer. We demonstrate the existence of a critical PAR, $`\lambda_{c}`$, which discriminates the unsaturated region, characterized by the range $`\lambda \in {\lbrack 0,\lambda_{c})}`$, from the saturation zone identified by any $`\lambda \in {\lbrack\lambda_{c},{+ \infty})}`$.
For conciseness, we invite the interested reader to refer to \[[4](#bib.bib4)\] for many details on the considered bi-dimensional Markov model, and references therein to get a picture of the topic addressed in this letter. Briefly, Liaw et al. extended the saturated Bianchi’s model by introducing a new idle state, not present in the original Bianchi’s model, accounting for the case in which the station buffer is empty, after a successful completion of a packet transmission. The main advantages of such a model rely on its simplicity and the effectiveness in describing the dynamics of the DCF in unsaturated traffic conditions, while basic hypotheses are the same as in Bianchi’s model.
Paper outline is as follows. In section II, we briefly recall the main probabilities needed for developing the proposed linear model, evaluate the throughput and present the adopted traffic model. Finally, Section III presents the linear model of the throughput along with simulation results.
## II Problem formulation and Traffic model
The bi-dimensional contention Markov model proposed in \[[4](#bib.bib4)\] governs the behaviour of each contending station through a series of states indexed by the pair $`{{{{(i,k)},{\forall i}} \in {\lbrack 0,m\rbrack}},{k \in {\lbrack 0,{W_{i} - 1}\rbrack}}},`$ whereby $`i`$ identifies the backoff stage, and $`k \in {\lbrack 0,{W_{i} - 1}\rbrack}`$ the backoff counter. The other parameters needed in the proposed framework can be summarized as follows: $`\tau`$ is the probability that a station starts a transmission in a randomly chosen slot time (ST), $`q`$ is the probability that there is at least a packet in the queue after a successful transmission, $`{{W_{i} = {2^{i}\hspace{0pt}W_{0}}},{{\forall i} \in {\lbrack 1,m\rbrack}}},`$ is the size of the $`i`$th contention window, $`W_{0}`$ is the minimum size of the contention window, $`P_{I,0}`$ is the probability of having at least one packet to be transmitted in the queue when the system is in idle state, and $`p`$ is the collision probability defined as in \[[1](#bib.bib1)\]
<table>
<tbody>
<tr>
<td></td>
<td>p=1−(1−τ)N−1𝑝1superscript1𝜏𝑁1\small p=1-(1-\tau)^{N-1}</td>
<td></td>
<td>(1)</td>
</tr>
</tbody>
</table>
Stationary probability $`b_{I}`$ of being in the idle state is:
<table>
<tbody>
<tr>
<td></td>
<td>bI=(1−q)b0,0/PI,0subscript𝑏𝐼1𝑞subscript𝑏00subscript𝑃𝐼0\small b_{I}=(1-q)b_{0,0}/P_{I,0}</td>
<td></td>
<td>(2)</td>
</tr>
</tbody>
</table>
whereby $`b_{0,0}`$ is defined as follows:
<table>
<tbody>
<tr>
<td></td>
<td>b0,0=1−bIα,α=12[1−(2p)m+11−2pW0+1−pm+11−p]formulae-sequencesubscript𝑏001subscript𝑏𝐼𝛼𝛼12delimited-[]1superscript2𝑝𝑚112𝑝subscript𝑊01superscript𝑝𝑚11𝑝\small b_{0,0}=\frac{1-b_{I}}{\alpha},~{}\alpha=\frac{1}{2}\left[\frac{1-(2p)^{m+1}}{1-2p}W_{0}+\frac{1-p^{m+1}}{1-p}\right]</td>
<td></td>
<td>(3)</td>
</tr>
</tbody>
</table>
By employing the normalization condition \[[1](#bib.bib1)\], it is possible to obtain:
<table>
<tbody>
<tr>
<td></td>
<td>τ=∑i=0mbi,0=∑i=0mpi⋅b0,0=ε⋅b0,0,ε=1−pm+11−pformulae-sequence𝜏superscriptsubscript𝑖0𝑚subscript𝑏𝑖0superscriptsubscript𝑖0𝑚⋅superscript𝑝𝑖subscript𝑏00⋅𝜀subscript𝑏00𝜀1superscript𝑝𝑚11𝑝\small\tau=\sum_{i=0}^{m}b_{i,0}=\sum_{i=0}^{m}p^{i}\cdot b_{0,0}=\varepsilon\cdot b_{0,0},~{}\varepsilon=\frac{1-p^{m+1}}{1-p}</td>
<td></td>
<td>(4)</td>
</tr>
</tbody>
</table>
Next line of pursuit is the computation of the system throughput. Putting together Eq.s ([1](#S2.E1 "In II Problem formulation and Traffic model ‣ On The Linear Behaviour of the Throughput of IEEE 802.11 DCF in Non-Saturated Conditions")), ([4](#S2.E4 "In II Problem formulation and Traffic model ‣ On The Linear Behaviour of the Throughput of IEEE 802.11 DCF in Non-Saturated Conditions")), a nonlinear system can be defined and solved numerically, obtaining the values of $`\tau`$ and $`p`$. The solution of the previous system is used for evaluating the throughput, defined as the ratio between the average payload information transmitted in a ST and the average length, $`T_{a\hspace{0pt}v}`$, of a ST:
<table>
<tbody>
<tr>
<td></td>
<td>S𝑆\displaystyle\small S</td>
<td>=\displaystyle=</td>
<td>Pt⋅Ps⋅E[PL]/Tav⋅subscript𝑃𝑡subscript𝑃𝑠𝐸delimited-[]𝑃𝐿subscript𝑇𝑎𝑣\displaystyle P_{t}\cdot P_{s}\cdot E[PL]/T_{av}</td>
<td></td>
<td>(5)</td>
</tr>
</tbody>
</table>
whereby $`E\hspace{0pt}{\lbrack{P\hspace{0pt}L}\rbrack}`$ is the average packet payload length (expressed in bits), $`P_{t}`$ is the probability that there is at least one transmission in the considered ST, with $`N`$ stations contending for the channel, each transmitting with probability $`\tau`$, i.e., $`P_{t} = {1 - {({1 - \tau})}^{N}}`$. Probability $`P_{s}`$ is the conditional probability that a packet transmission occurring on the channel is successful:
<table>
<tbody>
<tr>
<td></td>
<td>Ps=Nτ(1−τ)N−1/Ptsubscript𝑃𝑠𝑁𝜏superscript1𝜏𝑁1subscript𝑃𝑡\small P_{s}=N\tau(1-\tau)^{N-1}/P_{t}</td>
<td></td>
<td>(6)</td>
</tr>
</tbody>
</table>
Upon noting that, in a given ST, a station can reside in one of three possible kind of states, namely the idle state $`I`$ where the station spends $`T_{I}`$, the backoff states where the station spends $`T_{B\hspace{0pt}O}`$, and the transmitting states in which the station spends $`T_{T\hspace{0pt}X}`$, then the average duration $`T_{a\hspace{0pt}v}`$ of a ST easily follows
<table>
<tbody>
<tr>
<td></td>
<td>Tavsubscript𝑇𝑎𝑣\displaystyle\small T_{av}</td>
<td>=\displaystyle=</td>
<td>bI⋅TI+τ⋅TTX+∑i=0m∑k=1Wi−1bi,k⋅TBO⋅subscript𝑏𝐼subscript𝑇𝐼⋅𝜏subscript𝑇𝑇𝑋superscriptsubscript𝑖0𝑚superscriptsubscript𝑘1subscript𝑊𝑖1⋅subscript𝑏𝑖𝑘subscript𝑇𝐵𝑂\displaystyle b_{I}\cdot T_{I}+\tau\cdot T_{TX}+\sum_{i=0}^{m}\sum_{k=1}^{W_{i}-1}b_{i,k}\cdot T_{BO}</td>
<td></td>
<td rowspan="2">(7)</td>
</tr>
<tr>
<td></td>
<td></td>
<td>=\displaystyle=</td>
<td>bI⋅TI+[ε⋅TTX+θ⋅TBO]⋅b0,0⋅subscript𝑏𝐼subscript𝑇𝐼⋅delimited-[]⋅𝜀subscript𝑇𝑇𝑋⋅𝜃subscript𝑇𝐵𝑂subscript𝑏00\displaystyle b_{I}\cdot T_{I}+[\varepsilon\cdot T_{TX}+\theta\cdot T_{BO}]\cdot b_{0,0}</td>
<td></td>
</tr>
</tbody>
</table>
where $`\varepsilon`$ is defined in ([4](#S2.E4 "In II Problem formulation and Traffic model ‣ On The Linear Behaviour of the Throughput of IEEE 802.11 DCF in Non-Saturated Conditions")), and
<table>
<tbody>
<tr>
<td></td>
<td>θ=12[1−(2p)m+11−2pW0−1−pm+11−p]𝜃12delimited-[]1superscript2𝑝𝑚112𝑝subscript𝑊01superscript𝑝𝑚11𝑝\small\theta=\frac{1}{2}\left[\frac{1-(2p)^{m+1}}{1-2p}W_{0}-\frac{1-p^{m+1}}{1-p}\right]</td>
<td></td>
<td>(8)</td>
</tr>
</tbody>
</table>
Let us define the time durations $`T_{T\hspace{0pt}X},T_{B\hspace{0pt}O}`$ and $`T_{I}`$ in ([7](#S2.E7 "In II Problem formulation and Traffic model ‣ On The Linear Behaviour of the Throughput of IEEE 802.11 DCF in Non-Saturated Conditions")). Transmission time $`T_{T\hspace{0pt}X}`$ can be evaluated by noting that a station can experience two possible events: it successfully transmits over the channel or it encounters a collision. By doing so, $`T_{T\hspace{0pt}X}`$ can be defined as follows:
<table>
<tbody>
<tr>
<td></td>
<td>TTXsubscript𝑇𝑇𝑋\displaystyle\small T_{TX}</td>
<td>=\displaystyle=</td>
<td>(1−p)⋅Ts+p⋅Tc⋅1𝑝subscript𝑇𝑠⋅𝑝subscript𝑇𝑐\displaystyle(1-p)\cdot T_{s}+p\cdot T_{c}</td>
<td></td>
<td>(9)</td>
</tr>
</tbody>
</table>
whereby $`T_{c}`$ and $`T_{s}`$ are, respectively, the average time a channel is sensed busy due to a collision, and the successful data frame transmission time \[[1](#bib.bib1), [4](#bib.bib4)\].
Backoff time duration $`T_{B\hspace{0pt}O}`$ can be evaluated by considering the following two possibilities. A station can reside in a backoff slot of duration $`\sigma`$ if no other station is transmitting in the same ST, or for a time $`T_{T\hspace{0pt}X}`$ due to a collision with at least another station occupying the channel:
<table>
<tbody>
<tr>
<td></td>
<td>TBO=(1−Ptx[N−1])⋅σ+Ptx[N−1]⋅TTXsubscript𝑇𝐵𝑂⋅1subscript𝑃𝑡𝑥delimited-[]𝑁1𝜎⋅subscript𝑃𝑡𝑥delimited-[]𝑁1subscript𝑇𝑇𝑋\small T_{BO}=(1-P_{tx[N-1]})\cdot\sigma+P_{tx[N-1]}\cdot T_{TX}</td>
<td></td>
<td>(10)</td>
</tr>
</tbody>
</table>
whereby $`P_{t\hspace{0pt}x\hspace{0pt}{\lbrack{N - 1}\rbrack}} = {1 - {({1 - \tau})}^{N - 1}}`$ corresponds to the probability that at least a station, other than the tagged one, is transmitting in a ST.
For the sake of defining the traffic model employed for performance verification, we need to define both the access time $`T_{A}`$ (this is the average time a station spends through the various backoff stages before transmitting a packet) and the service time $`T_{S}`$. From \[[4](#bib.bib4)\], $`T_{A}`$ can be defined as follows:
<table>
<tbody>
<tr>
<td></td>
<td>TA=∑i=0mpi⋅Wi2⋅TBO∑i=0mpi=W02ε1−(2p)m+11−(2p)⋅TBOsubscript𝑇𝐴superscriptsubscript𝑖0𝑚⋅superscript𝑝𝑖subscript𝑊𝑖2subscript𝑇𝐵𝑂superscriptsubscript𝑖0𝑚superscript𝑝𝑖⋅subscript𝑊02𝜀1superscript2𝑝𝑚112𝑝subscript𝑇𝐵𝑂\small T_{A}=\frac{\sum_{i=0}^{m}p^{i}\cdot\frac{W_{i}}{2}\cdot T_{BO}}{\sum_{i=0}^{m}p^{i}}=\frac{W_{0}}{2\,\varepsilon}\frac{1-(2p)^{m+1}}{1-(2p)}\cdot T_{BO}</td>
<td></td>
<td>(11)</td>
</tr>
</tbody>
</table>
whereby $`\varepsilon`$ is as defined in ([4](#S2.E4 "In II Problem formulation and Traffic model ‣ On The Linear Behaviour of the Throughput of IEEE 802.11 DCF in Non-Saturated Conditions")). On the other hand, $`T_{S}`$ \[[6](#bib.bib6)\], i.e., the time elapsed from the moment a packet is taken from the queue to the instant in which it is successfully transmitted, can be defined as $`T_{S} = {T_{A} + T_{T\hspace{0pt}X}}`$.
As far as $`T_{I}`$ is concerned, we resort to the definition \[[4](#bib.bib4)\]:
<table>
<tbody>
<tr>
<td></td>
<td>TI=[ε⋅TTX+θ⋅TBO]⋅b0,01−bI=[ε⋅TTX+θ⋅TBO]αsubscript𝑇𝐼⋅delimited-[]⋅𝜀subscript𝑇𝑇𝑋⋅𝜃subscript𝑇𝐵𝑂subscript𝑏001subscript𝑏𝐼delimited-[]⋅𝜀subscript𝑇𝑇𝑋⋅𝜃subscript𝑇𝐵𝑂𝛼\small T_{I}=\frac{[\varepsilon\cdot T_{TX}+\theta\cdot T_{BO}]\cdot b_{0,0}}{1-b_{I}}=\frac{[\varepsilon\cdot T_{TX}+\theta\cdot T_{BO}]}{\alpha}</td>
<td></td>
<td>(12)</td>
</tr>
</tbody>
</table>
where last equality stems from ([3](#S2.E3 "In II Problem formulation and Traffic model ‣ On The Linear Behaviour of the Throughput of IEEE 802.11 DCF in Non-Saturated Conditions")).
The employed traffic model is $`M/G/1/K`$. Probabilities $`q`$ and $`P_{I,0}`$ in our model can be defined as follows:
<table>
<tbody>
<tr>
<td></td>
<td>q=1−π0=1−1−ρ1−ρK+1,PI,0=1−e−λ⋅TI𝑞absent1subscript𝜋011𝜌1superscript𝜌𝐾1subscript𝑃𝐼0absent1superscript𝑒⋅𝜆subscript𝑇𝐼\small\begin{array}[]{rlrl}q=&1-\pi_{0}=1-\frac{1-\rho}{1-\rho^{K+1}},&P_{I,0}=&1-e^{-\lambda\cdot T_{I}}\end{array}</td>
<td></td>
<td>(13)</td>
</tr>
</tbody>
</table>
where $`\pi_{0}`$ is the probability of an empty system \[[8](#bib.bib8)\], $`\rho = {\lambda \cdot T_{S}}`$, $`q`$ follows from M/G/1/K queuing theory \[[8](#bib.bib8)\], while $`P_{I\hspace{0pt}0}`$ stems from the fact that for exponentially distributed interarrival times with mean $`1/\lambda`$ the probability of having at least one packet arrival during time $`T`$ is equal to $`1 - e^{- {\lambda \cdot T}}`$.
Employing ([2](#S2.E2 "In II Problem formulation and Traffic model ‣ On The Linear Behaviour of the Throughput of IEEE 802.11 DCF in Non-Saturated Conditions")), for a $`K - 1`$-length queue, we have:
<table>
<tbody>
<tr>
<td></td>
<td>bI=1−ρ1−ρK+1⋅11−e−λ⋅TI⋅b0,0subscript𝑏𝐼⋅1𝜌1superscript𝜌𝐾111superscript𝑒⋅𝜆subscript𝑇𝐼subscript𝑏00\small b_{I}=\frac{1-\rho}{1-\rho^{K+1}}\cdot\frac{1}{1-e^{-\lambda\cdot T_{I}}}\cdot b_{0,0}</td>
<td></td>
<td>(14)</td>
</tr>
</tbody>
</table>
## III The Linear model and simulation results
TABLE I: Numerical Results.
<table>
<thead>
<tr>
<th>N𝑁N</th>
<th>10</th>
<th>20</th>
<th>30</th>
</tr>
</thead>
<tbody>
<tr>
<th>Smsubscript𝑆𝑚S_{m} [Mbps]</th>
<td>9.118</td>
<td>8.73</td>
<td>8.608</td>
</tr>
<tr>
<th>λcsubscript𝜆𝑐\lambda_{c} [pkt/s]</th>
<td>111.2</td>
<td>53.235</td>
<td>34.99</td>
</tr>
</tbody>
</table>
A model of the throughput in non-saturated traffic conditions, along with its dependence on some key network parameters, can be derived by analyzing ([5](#S2.E5 "In II Problem formulation and Traffic model ‣ On The Linear Behaviour of the Throughput of IEEE 802.11 DCF in Non-Saturated Conditions")) in the limit $`\lambda\rightarrow 0`$.
Let us write the throughput in ([5](#S2.E5 "In II Problem formulation and Traffic model ‣ On The Linear Behaviour of the Throughput of IEEE 802.11 DCF in Non-Saturated Conditions")) as a function of $`\tau`$. By employing ([6](#S2.E6 "In II Problem formulation and Traffic model ‣ On The Linear Behaviour of the Throughput of IEEE 802.11 DCF in Non-Saturated Conditions")), the numerator can be rewritten as $`{P_{t}\hspace{0pt}P_{s}\hspace{0pt}E\hspace{0pt}{\lbrack{P\hspace{0pt}L}\rbrack}} = {N\hspace{0pt}\tau\hspace{0pt}{({1 - \tau})}^{N - 1}\hspace{0pt}E\hspace{0pt}{\lbrack{P\hspace{0pt}L}\rbrack}}`$. As far as the denominator is concerned, upon substituting ([12](#S2.E12 "In II Problem formulation and Traffic model ‣ On The Linear Behaviour of the Throughput of IEEE 802.11 DCF in Non-Saturated Conditions")) in ([7](#S2.E7 "In II Problem formulation and Traffic model ‣ On The Linear Behaviour of the Throughput of IEEE 802.11 DCF in Non-Saturated Conditions")), and remembering that $`b_{0,0} = {{({1 - b_{I}})}/\alpha}`$, after some algebra, it is possible to obtain the following relation $`T_{a\hspace{0pt}v} = {{\lbrack{{\varepsilon \cdot T_{T\hspace{0pt}X}} + {\theta \cdot T_{B\hspace{0pt}O}}}\rbrack}/\alpha}`$. By collecting the previous relations, the throughput can be rewritten as:
<table>
<tbody>
<tr>
<td></td>
<td>S(τ)=Nτ(1−τ)N−1E[PL]α(τ)ε(τ)⋅TTX(τ)+θ(τ)⋅TBO(τ)𝑆𝜏𝑁𝜏superscript1𝜏𝑁1𝐸delimited-[]𝑃𝐿𝛼𝜏⋅𝜀𝜏subscript𝑇𝑇𝑋𝜏⋅𝜃𝜏subscript𝑇𝐵𝑂𝜏\small S(\tau)=\frac{N\tau(1-\tau)^{N-1}E[PL]\alpha(\tau)}{\varepsilon(\tau)\cdot T_{TX}(\tau)+\theta(\tau)\cdot T_{BO}(\tau)}</td>
<td></td>
<td>(15)</td>
</tr>
</tbody>
</table>
whereby we highlighted the dependence on $`\tau`$ of the terms $`\alpha,\varepsilon,T_{T\hspace{0pt}X},\theta`$, and $`T_{B\hspace{0pt}O}`$. Upon noting that ([3](#S2.E3 "In II Problem formulation and Traffic model ‣ On The Linear Behaviour of the Throughput of IEEE 802.11 DCF in Non-Saturated Conditions")) yields $`{\lim_{\lambda\rightarrow 0}b_{0,0}} = 0`$, from ([4](#S2.E4 "In II Problem formulation and Traffic model ‣ On The Linear Behaviour of the Throughput of IEEE 802.11 DCF in Non-Saturated Conditions")) it follows $`\tau\rightarrow 0`$ as well. In the limit $`\tau\rightarrow 0`$, it is straightforward to demonstrate the following relations: $`p\rightarrow 0`$ (from ([1](#S2.E1 "In II Problem formulation and Traffic model ‣ On The Linear Behaviour of the Throughput of IEEE 802.11 DCF in Non-Saturated Conditions"))), $`{\varepsilon\hspace{0pt}{(\tau)}}\rightarrow 1`$ (from ([4](#S2.E4 "In II Problem formulation and Traffic model ‣ On The Linear Behaviour of the Throughput of IEEE 802.11 DCF in Non-Saturated Conditions"))), $`{\theta\hspace{0pt}{(\tau)}}\rightarrow\frac{W_{0} - 1}{2}`$ (from ([8](#S2.E8 "In II Problem formulation and Traffic model ‣ On The Linear Behaviour of the Throughput of IEEE 802.11 DCF in Non-Saturated Conditions"))), $`{\alpha\hspace{0pt}{(\tau)}}\rightarrow\frac{W_{0} + 1}{2}`$ (from ([3](#S2.E3 "In II Problem formulation and Traffic model ‣ On The Linear Behaviour of the Throughput of IEEE 802.11 DCF in Non-Saturated Conditions"))), $`{T_{T\hspace{0pt}X}\hspace{0pt}{(\tau)}}\rightarrow T_{s}`$ (from ([9](#S2.E9 "In II Problem formulation and Traffic model ‣ On The Linear Behaviour of the Throughput of IEEE 802.11 DCF in Non-Saturated Conditions"))), and $`{T_{B\hspace{0pt}O}\hspace{0pt}{(\tau)}}\rightarrow\sigma`$ (from ([10](#S2.E10 "In II Problem formulation and Traffic model ‣ On The Linear Behaviour of the Throughput of IEEE 802.11 DCF in Non-Saturated Conditions"))).
Upon substituting the derivations above in ([15](#S3.E15 "In III The Linear model and simulation results ‣ On The Linear Behaviour of the Throughput of IEEE 802.11 DCF in Non-Saturated Conditions")), the limit $`{\lim_{\tau\rightarrow 0}{S\hspace{0pt}{(\tau)}}} = 0 = \left\lbrack {{S\hspace{0pt}{(\tau)}}\mid}_{\tau = 0} \right\rbrack`$ easily follows.
Upon employing the Taylor’s formula around $`\tau \approx 0`$, the throughput $`S\hspace{0pt}{(\tau)}`$ can be well approximated as follows:
<table>
<tbody>
<tr>
<td></td>
<td>S(τ)≈[S(τ)∣τ=0]+[∂S(τ)∂τ∣τ=0]τ=[∂S(τ)∂τ∣τ=0]τ𝑆𝜏delimited-[]evaluated-at𝑆𝜏𝜏0delimited-[]evaluated-at𝑆𝜏𝜏𝜏0𝜏delimited-[]evaluated-at𝑆𝜏𝜏𝜏0𝜏S(\tau)\approx\left[S(\tau)\mid_{\tau=0}\right]+\left[\frac{\partial S(\tau)}{\partial\tau}\mid_{\tau=0}\right]\tau=\left[\frac{\partial S(\tau)}{\partial\tau}\mid_{\tau=0}\right]\tau</td>
<td></td>
</tr>
</tbody>
</table>
As far as $`\tau \approx 0`$, $`{({1 - \tau})}^{N - 1} \approx {{1 - {{({N - 1})}\hspace{0pt}\tau}} + {o\hspace{0pt}{(\tau)}}}`$; therefore, the approximation $`{N\hspace{0pt}\tau\hspace{0pt}{({1 - \tau})}^{N - 1}} \approx {{N\hspace{0pt}\tau} + {o\hspace{0pt}{(\tau)}}}`$ holds as well. As a consequence, ([15](#S3.E15 "In III The Linear model and simulation results ‣ On The Linear Behaviour of the Throughput of IEEE 802.11 DCF in Non-Saturated Conditions")) can be rewritten as follows:
TABLE II: Typical network parameters
<table>
<tbody>
<tr>
<th>MAC header</th>
<th>28 bytes</th>
<td>Propag. delay τpsubscript𝜏𝑝\tau_{p}</td>
<td>1 μs𝜇𝑠\mu s</td>
</tr>
<tr>
<th>PLCP Preamble</th>
<th>144 bit</th>
<td>PLCP Header</td>
<td>48 bit</td>
</tr>
<tr>
<th>PHY header</th>
<th>24 bytes</th>
<td>Slot time</td>
<td>20 μs𝜇𝑠\mu s</td>
</tr>
<tr>
<th>PLCP rate</th>
<th>1Mbps</th>
<td>W0</td>
<td>32</td>
</tr>
<tr>
<th>No. back-off stages, m</th>
<th>5</th>
<td>Wmax</td>
<td>1024</td>
</tr>
<tr>
<th>Payload size</th>
<th>1025 bytes</th>
<td>SIFS</td>
<td>10 μs𝜇𝑠\mu s</td>
</tr>
<tr>
<th>ACK</th>
<th>14 bytes</th>
<td>DIFS</td>
<td>50 μs𝜇𝑠\mu s</td>
</tr>
<tr>
<th>ACK timeout</th>
<th>364μs𝜇𝑠\mu s</th>
<td>EIFS</td>
<td>364 μs𝜇𝑠\mu s</td>
</tr>
</tbody>
</table>
<table>
<tbody>
<tr>
<td></td>
<td>S(τ)≈N⋅E[PL]⋅W0+12Ts+W0−12⋅σ⋅τ𝑆𝜏⋅⋅⋅𝑁𝐸delimited-[]𝑃𝐿subscript𝑊012subscript𝑇𝑠⋅subscript𝑊012𝜎𝜏\small S(\tau)\approx\frac{N\cdot E[PL]\cdot\frac{W_{0}+1}{2}}{T_{s}+\frac{W_{0}-1}{2}\cdot\sigma}\cdot\tau</td>
<td></td>
<td>(16)</td>
</tr>
</tbody>
</table>
Next line of pursuit consists in expressing $`\tau`$ in terms of the PAR $`\lambda`$. By using the MacLaurin expansion of the exponential $`e^{- {\lambda\hspace{0pt}T_{I}}} \approx {{1 - {\lambda\hspace{0pt}T_{I}}} + {o\hspace{0pt}{(\lambda)}}}`$, from ([14](#S2.E14 "In II Problem formulation and Traffic model ‣ On The Linear Behaviour of the Throughput of IEEE 802.11 DCF in Non-Saturated Conditions")) it is $`b_{0,0} \approx {\lambda\hspace{0pt}T_{I}}`$ as $`\lambda\rightarrow 0`$. On the other hand, Equ. ([12](#S2.E12 "In II Problem formulation and Traffic model ‣ On The Linear Behaviour of the Throughput of IEEE 802.11 DCF in Non-Saturated Conditions")) yields $`{\lim_{\lambda\rightarrow 0}T_{I}} = {{\lbrack{T_{s} + {\frac{W_{0} - 1}{2}\hspace{0pt}\sigma}}\rbrack}/\frac{W_{0} + 1}{2}}`$. Finally, in the limit $`\lambda\rightarrow 0`$, Equ. ([4](#S2.E4 "In II Problem formulation and Traffic model ‣ On The Linear Behaviour of the Throughput of IEEE 802.11 DCF in Non-Saturated Conditions")) yields $`\tau\rightarrow b_{0,0} \approx {{\lambda \cdot {\lbrack{T_{s} + {\frac{W_{0} - 1}{2}\hspace{0pt}\sigma}}\rbrack}}/\frac{W_{0} + 1}{2}}`$. Upon substituting the previous mathematical derivations in ([16](#S3.E16 "In III The Linear model and simulation results ‣ On The Linear Behaviour of the Throughput of IEEE 802.11 DCF in Non-Saturated Conditions")), the throughput can be approximated as follows:
<table>
<tbody>
<tr>
<td></td>
<td>S(λ)≃N⋅E[PL]⋅λsimilar-to-or-equals𝑆𝜆⋅⋅𝑁𝐸delimited-[]𝑃𝐿𝜆\small S(\lambda)\simeq N\cdot E[PL]\cdot\lambda</td>
<td></td>
<td>(17)</td>
</tr>
</tbody>
</table>
It is interesting to estimate the interval of validity $`\lbrack 0,\lambda_{c}\rbrack`$ of the linear throughput model proposed in ([17](#S3.E17 "In III The Linear model and simulation results ‣ On The Linear Behaviour of the Throughput of IEEE 802.11 DCF in Non-Saturated Conditions")). An appropriate value of $`\lambda_{c}`$ can be obtained by finding the abscissa $`\lambda`$ corresponding to the intersection of the straight line ([17](#S3.E17 "In III The Linear model and simulation results ‣ On The Linear Behaviour of the Throughput of IEEE 802.11 DCF in Non-Saturated Conditions")) with the horizontal line passing through the maximum $`S_{m}`$ of the throughput. To this end, after expressing any term involved in ([5](#S2.E5 "In II Problem formulation and Traffic model ‣ On The Linear Behaviour of the Throughput of IEEE 802.11 DCF in Non-Saturated Conditions")) as a function of $`\tau`$, the maximum throughput $`S_{m}`$ can be obtained in two steps. First, one has to find the value $`\tau_{m}`$ for which the throughput gets maximized. This can be easily obtained by equating to zero the derivative of ([5](#S2.E5 "In II Problem formulation and Traffic model ‣ On The Linear Behaviour of the Throughput of IEEE 802.11 DCF in Non-Saturated Conditions")) with respect to $`\tau`$, and then solving for $`\tau`$. In the second step, the maximizing value $`\tau_{m}`$ can be substituted into $`S\hspace{0pt}{(\tau)}`$ for obtaining $`S_{m} = {S\hspace{0pt}(\tau)_{|_{\tau_{m}}}}`$. Finally, $`\lambda_{c}`$ is the value of $`\lambda`$ for which the following holds:
[Refer to caption](/html/0710.5238/assets/x1.png)
Figure 1: Throughput for the 2-way mechanism as a function of the PAR $`\lambda`$, for three different number $`N`$ of contending stations. Straight dashed lines refer to the linear model of the throughput in ([17](#S3.E17 "In III The Linear model and simulation results ‣ On The Linear Behaviour of the Throughput of IEEE 802.11 DCF in Non-Saturated Conditions")).
<table>
<tbody>
<tr>
<td></td>
<td>λc=λ:N⋅E[PL]⋅λ=Sm:subscript𝜆𝑐𝜆⋅⋅𝑁𝐸delimited-[]𝑃𝐿𝜆subscript𝑆𝑚\small\lambda_{c}=\lambda:N\cdot E[PL]\cdot\lambda=S_{m}</td>
<td></td>
<td>(18)</td>
</tr>
</tbody>
</table>
Table [I](#S3.T1 "TABLE I ‣ III The Linear model and simulation results ‣ On The Linear Behaviour of the Throughput of IEEE 802.11 DCF in Non-Saturated Conditions") shows the values of $`\lambda_{c}`$, found numerically, along with the respective values of $`S_{m}`$, for various values of $`N`$. We considered the bit rate 54Mbps for the protocol IEEE 802.11g. The linear model ([17](#S3.E17 "In III The Linear model and simulation results ‣ On The Linear Behaviour of the Throughput of IEEE 802.11 DCF in Non-Saturated Conditions")) shows a close agreement with both theoretical (continuous curves) and simulated throughput curves. Fig. [1](#S3.F1 "Figure 1 ‣ III The Linear model and simulation results ‣ On The Linear Behaviour of the Throughput of IEEE 802.11 DCF in Non-Saturated Conditions") shows the straight lines in ([17](#S3.E17 "In III The Linear model and simulation results ‣ On The Linear Behaviour of the Throughput of IEEE 802.11 DCF in Non-Saturated Conditions")) (dashed lines) for three different values of $`N`$. The figure also shows the values of $`S_{m}`$ (horizontal dash-dot lines) along with the three values of $`\lambda_{c}`$ deduced from ([18](#S3.E18 "In III The Linear model and simulation results ‣ On The Linear Behaviour of the Throughput of IEEE 802.11 DCF in Non-Saturated Conditions")) and noted in Table [I](#S3.T1 "TABLE I ‣ III The Linear model and simulation results ‣ On The Linear Behaviour of the Throughput of IEEE 802.11 DCF in Non-Saturated Conditions"). Simulated values (star-marked points) have been obtained with ns-2 by using settings noted in Table [II](#S3.T2 "TABLE II ‣ III The Linear model and simulation results ‣ On The Linear Behaviour of the Throughput of IEEE 802.11 DCF in Non-Saturated Conditions"), along with standard 54Mbps 802.11 parameterizations \[[7](#bib.bib7)\], which are also the standard parameters defined in ns-2. Within ns-2, $`N`$ stations have been randomly placed in a square area with edge size equal to $`50`$m using a uniform distribution. All simulation results in Fig. [1](#S3.F1 "Figure 1 ‣ III The Linear model and simulation results ‣ On The Linear Behaviour of the Throughput of IEEE 802.11 DCF in Non-Saturated Conditions") are obtained with a 95% confidence interval lower than $`15`$kbps.
Fig. [1](#S3.F1 "Figure 1 ‣ III The Linear model and simulation results ‣ On The Linear Behaviour of the Throughput of IEEE 802.11 DCF in Non-Saturated Conditions") shows that the maximum achievable throughput does not exceed 10Mbps (@N=10) despite the maximum bit rate employed (54Mbps): this throughput penalty is essentially due to the fact that the control packet and the PLCL header are transmitted at 1Mbps no matter the operating transmission mode. As a reference throughput performance, we show the maximum saturation throughput (labelled ”Sat. Through. \[1\]” in Fig. [1](#S3.F1 "Figure 1 ‣ III The Linear model and simulation results ‣ On The Linear Behaviour of the Throughput of IEEE 802.11 DCF in Non-Saturated Conditions")) found in \[[1](#bib.bib1)\] in case the $`N`$ contending stations transmit at the optimal transmission probability $`\tau`$. Notice that, throughput penalty with respect to 54Mbps is well predicted by the theoretical formulation presented in \[[1](#bib.bib1)\].
## References
- \[1\]
G. Bianchi, ”Performance analysis of the IEEE 802.11 distributed coordination function”, IEEE JSAC, Vol.18, No.3, March 2000.
- \[2\]
G. Bianchi and I. Tinnirello, ”Remarks on IEEE 802.11 DCF performance analysis”, IEEE Comm. Letters, Vol.9, No.8, Aug. 2005.
- \[3\]
A. Kumar, E. Altman, D. Miorandi, and M. Goyal, ”New insights from a fixed-point analysis of single cell IEEE 802.11 WLANs”, to appear on IEEE/ACM Trans. on Networking, 2007, now available on IEEExplorer.
- \[4\]
Y.S. Liaw, A. Dadej, and A.Jayasuriya, ”Performance analysis of IEEE 802.11 DCF under limited load”, In Proc. of Asia-Pacific Conference on Communications, Vol.1, pp.759 - 763, 03-05 Oct. 2005.
- \[5\]
D. Malone, K. Duffy, and D.J. Leith, ”Modeling the 802.11 distributed coordination function in non-saturated heterogeneous conditions”, IEEE-ACM Trans. on Networking, vol. 15, No. 1, pp. 159 172, Feb. 2007.
- \[6\]
W. Lee, C. Wang, and K. Sohraby, ”On use of traditional M/G/1 model for IEEE 802.11 DCF in unsaturated traffic conditions”, In Proc. of IEEE WCNC, 2006, pp. 1933 1937.
- \[7\]
IEEE Standard for Wireless LAN Medium Access Control (MAC) and Physical Layer (PHY) Specifications, November 1997, P802.11
- \[8\]
G. Bolch, S. Greiner, H. de Meer, and K.S. Trivedi, Queueing Networks and Markov Chains, Wiley-Interscience, 2nd edition, 2006. |
0710.5327 | cs.NI | cyber | # Escalating The War On SPAM Through Practical POW Exchange
Paul Gardner-Stephen (paul@infoeng.flinders.edu.au),
School of Informatics & Engineering,
Flinders University, Adelaide, Australia
###### Abstract
Proof-of-work (POW) schemes have been proposed in the past. One prominent system is HASHCASH (Back, [2002](#bib.bib2)) which uses cryptographic puzzles . However, work by Laurie and Clayton ([2004](#bib.bib9)) has shown that for a uniform proof-of-work scheme on email to have an impact on SPAM, it would also be onerous enough to impact on senders of “legitimate” email. I suggest that a non-uniform proof-of-work scheme on email may be a solution to this problem, and describe a framework that has the potential to limit SPAM, without unduly penalising legitimate senders, and is constructed using only current SPAM filter technology, and a small change to the SMTP (Simple Mail Transfer Protocol). Specifically, I argue that it is possible to make sending SPAM 1,000 times more expensive than sending “legitimate” email (so called HAM). Also, unlike the system proposed by Liu and Camp ([2006](#bib.bib10)), it does not require the complications of maintaining a reputation system.
## 1 Introduction
### 1.1 A Brief Introduction To Proof-Of-Work As A SPAM Counter Measure
Proof-of-Work (POW) systems were suggested as a counter measure to junk email (SPAM) as early as 1992 (Dwork and Naor, [1993](#bib.bib4)), and the concept has been rediscovered and developed since then, e.g, (Rivest and Shamir, [1996](#bib.bib12); Rivest et al., [1996](#bib.bib11); Jakobsson and Juels, [1999](#bib.bib6); Back, [2002](#bib.bib2); Dwork et al., [2002](#bib.bib3); Wang and Reiter, [2003](#bib.bib13); Abadi et al., [2005](#bib.bib1); Ioannidis et al., [2005](#bib.bib5)).
In simple terms a POW system uses a puzzle challenge that is hard to solve, but easy to verify. The solution to the puzzle is presented to the receiving email server as proof of having completed a certain amount of work. The intention is that this will limit the number of messages that a Spammer can send per unit time. When applied to all email messages, I call the scheme a *Uniform Cost Proof Of Work* Scheme. That is, the computational burden is equal for every message, whether it is wanted by the recipient (HAM), or unwanted by the recipient (SPAM).
In reality, due to differences in computational capacity (CPU speed) of mail servers, the burden is not uniform. These differences can be great, probably around two orders of magnitude between the fastest and slowest computers that send email. However, the difference between the fastest and slowest random access memories are much less than the difference between the fastest and slowest CPUs. Classes of puzzles have been suggested that use this property to keep the burden as uniform as possible, to reduce the performance difference to a factor of less than five Abadi et al. ([2005](#bib.bib1)). This would seem to be a reasonable achievement.
### 1.2 Problems With Uniform-Cost Proof-Of-Work As A SPAM Counter Measure
Assuming for the moment, then, that (approximately) Uniform-Cost Proof-of-Work systems exist, we turn to the problems they face. The following is a list of problems with Uniform-Cost Proof-of-Work systems that, with the exception of message latency, are drawn from the work of Laurie and Clayton ([2004](#bib.bib9)) and Krawetz ([2004](#bib.bib8)):
#### 1.2.1 Message Latency
By applying a computational delay to every message, the implied real-time delivery semantic of email is broken. This makes a Uniform-Cost Proof-of-Work scheme socially unattractive.
#### 1.2.2 Inequitable Burden
The problem of inequitable burden of proof-of-work schemes due to server speed has already been mentioned, and even though Abadi et al. ([2005](#bib.bib1)) showed how to reduce the margin to less than five-fold, that is still a significant inequality.
#### 1.2.3 Mailing Lists
Mailing lists send a great many messages. If each recipient invoked a uniform email proof-of-work burden, then large mailing lists would become expensive, if not impractical. That is, the indiscriminate nature of the proof-of-work scheme causes mailing lists to be impacted to the same degree as a Spammer who sends the same quantity of messages.
#### 1.2.4 Robot Armies
Moreover, because many Spammer control hundreds of thousands of compromised computers around the world, it would seem that Spammer have access to much larger CPU resources than most legitimate senders. Thus a uniformly applied proof-of-work scheme on email may actually hurt legitimate senders more than it hurts Spammer. Also, Spammer are able to maintain legitimate CPU resources, further compounding the problem.
#### 1.2.5 Summary
To summarise, we discover what Laurie and Clayton ([2004](#bib.bib9)) have established from a theoretical perspective, that for any uniform email proof-of-work scheme to be sufficient to reduce SPAM (assuming 2004 levels of roughly equal SPAM and HAM), then the several percent of legitimate email senders who send relatively many email messages will be prevented from doing so. In other words, a Uniform-Cost Proof-of-Work scheme, regardless of the quality of the algorithms it uses, cannot succeed, precisely because it burdens SPAM and HAM uniformly. Therefore, if Proof-of-Work is to work, it cannot be with a Uniform-Cost model.
### 1.3 Introducing Targeted-Cost Proof-Of-Work As A Feasible SPAM Counter Measure
Having argued that a Uniform-Cost Proof-of-Work scheme cannot succeed, I suggest that Proof-of-Work schemes, generally, are not fatally flawed. Note that for every objection in the previous text, that the problem is the indiscriminate application of a uniform burden on all email messages. The logical alternative is to apply the burden on a more intelligent basis. An ideal implementation would place no burden on legitimate messages (HAM), and an infinite and unavoidable burden on SPAM. Such a system would immediately, by definition, stop all SPAM — unfortunately this result cannot be produced with current technology.
While we have no 100% accurate method to discriminate HAM and SPAM, we do have effective heuristic email classification systems, such as SpamAssassin
<sup>1</sup>
<sup>1</sup>1HTTP://spamassassin.apache.org/
, CRM114
<sup>2</sup>
<sup>2</sup>2HTTP://crm114.sourceforge.net/
and many others. Thus, by using the judgements of an existing SPAM filter, as to the probability of a given message being SPAM, it is possible to dynamically determine the burden to apply to that message. That is, we can still place the vast majority of the burden on Spammers, with the precise proportion determined by the quality of the SPAM filters. As with previous Proof-of-Work schemes, the work would take the form of solving a specially created cryptographic puzzle with the desired degree of difficulty.
Such a scheme has a subtle but important difference to just using SPAM filters alone: The failure mode for falsely classifying HAM as SPAM is more robust, as the delivery of an incorrectly classified message is only resisted, instead of being refused. For example, it would be unrealistic to configure a SPAM filter to reject messages that are only 5% likely to be SPAM, even though that may be the threshold required to reject practically all SPAM. However, it is completely reasonable to resist the delivery of messages that are 5% likely to be SPAM. I call this method of the selective application of a Proof-of-Work scheme Targeted-Cost Proof-of-Work.
This assumes that all computers can perform the work required by a Proof-of-Work scheme at a uniform rate. But computers vary in speed. Fortunately, if both small hand held devices and super-computers are excluded from consideration (this seems reasonable, since few legitimate mail senders use hand held devices, and few spammers have sustained access to super computers), most computers connected to the Internet vary in clock speed by no more than about one order of magnitude. Also, research has been conducted into creating problems that are limited by a computers random access memory speed, which is a quantity that varies much less than CPU speed (Dwork et al., [2002](#bib.bib3); Abadi et al., [2005](#bib.bib1)). Moreover, there seems to be no reason to suggest that the average computer being used to send SPAM will be much faster than the average computer being used to send legitimate email. Thus, the system
Some SPAM filters have demonstrated \>99.9% accuracy
<sup>3</sup>
<sup>3</sup>3HTTP://crm114.sourceforge.net/
. Therefore, by using SPAM filters to inform a Targeted-Cost Proof-of-Work scheme, it should be possible to correctly resist 99.9% of SPAM, but only resist 0.1% of HAM. Thus Spammers are burdened almost 1,000 times more per message, on average, than legitimate senders.
By shifting the vast majority of the computation burden from legitimate senders to Spammers, Proof-of-Work should be feasible. Consider the calculations presented by Laurie and Clayton ([2004](#bib.bib9)) that showed how a Uniform-Cost Proof-of-Work scheme would hurt legitimate senders who send more than 250 emails per day. The graph presented by Laurie and Clayton ([2004](#bib.bib9)) suggests that 1.56% of legitimate senders would fall into this category. If the cost to send (on average) were reduced by a factor of 1,000, then legitimate senders could send of the order of 250,000 messages per computer per day, while still limiting Spammers to of the order of 250 messages per computer per day.
Moreover, if the advantage is 1,000 times, then some legitimate sending capacity can be sacrificed to further limit the sending of SPAM. I suggest that a Targeted-Cost Proof-of-Work scheme require proof of approximately 1 hour of work for messages suspected of being SPAM (assuming 99.9% accurate classification). In that case, legitimate senders could deliver about 24,000 messages per computer per day, while Spammers would be limited to only 24 messages per computer per day. Assuming that Spammers have about 10 million computers at their disposal, this would limit daily spam volumes to only 24 million messages per day — about one SPAM for every twenty Internet users.
If the system is detuned to accommodate a SPAM classifier that is less accurate, then the advantage would be reduced, either by limiting the daily capacity of legitimate senders, or by allowing Spammers to send more SPAM, or perhaps a combination of the two.
Even a classification accuracy of only 95% gives a 20:1 advantage to HAM over SPAM, which would allow legitimate senders to deliver several thousand messages per computer per day, while limiting Spammers to only a couple of hundred messages per computer per day. These limits should accommodate practically all legitimate senders (Laurie and Clayton ([2004](#bib.bib9)) suggest that fewer than 0.1% of users send more than a couple of thousand email messages per day), while limiting SPAM to similar volumes as HAM (since the vast majority of legitimate senders send many fewer messages per day (according to Laurie and Clayton ([2004](#bib.bib9)), the median is about 90 messages per day).
On balance, requiring about an hour of work for each suspected SPAM seems reasonable, as it severely limits SPAM delivery capacity, while not burdening legitimate senders too much: If SPAM filter accuracy of 99.9% is sustainable, then a cost of 1 hour applies to only 0.1% of legitimate email, giving an average delivery cost of only 3.6 seconds.
Moreover, as will be described below; a) legitimate bulk senders can act to reduce their burden; and b) it is possible for each receiving mail server to configure this behaviour independently, reflecting the accuracy of their SPAM filters, or to satisfy local policy. That is, a Targeted-Cost Proof-of-Work scheme can improve the HAM to SPAM advantage somewhat beyond what is immediately apparent. However, there are risks that must be managed.
## 2 Managing Risks
### 2.1 Variable SPAM Filter Performance
The effectiveness of a Targeted-Cost Proof-of-Work system is proportional to the accuracy of the SPAM filter that underlies it. But SPAM filters differ in accuracy from site to site. Some of this variability could be fixed if all sites used the current best SPAM filter software. However, some variability is due to the kind of email that sites receive (consider the difference between what staff at a marketing company, a SPAM researcher, and staff at a medical centre would consider SPAM and HAM, and also the difference between the kinds of messages they might receive).
Fortunately, a Targeted-Cost Proof-of-Work system need not consider these issues. This is because the receiving mail server specifies whether to resist delivery of a message, and what the level of resistance is. The local message administrator can tune the threshold at which the resistance applies, and the level of the resistance (perhaps introducing a sliding resistance scale based on the spamminess of a message, and white lists of senders who are never resisted). Therefore each mail server can optimise the system to minimise the amount of SPAM that is delivered, without unduly burdening legitimate senders.
### 2.2 When Delivery Of Legitimate Mail Is Too Expensive
This leads to an important issue: What happens when a receiving mail server imposes a burden that a legitimate sender is not willing to meet? This is the one failure mode that is undesirable (SPAM that is accepted is less of a problem, because the receiving mail server may still discard the message or mark it as SPAM, even after it has been accepted from the sender). Ideally, the sending mail server should alert the user that their message was not delivered because it was too spammy. The user could then re-draft and re-send the message. Alternatively, they could send a fresh message to the recipient asking to be added to their receiving mail servers white list so that future messages will not be resisted.
### 2.3 Senders Of “Pressed HAM” (PHAM)
A related problem is legitimate senders who send messages that are, perhaps unavoidably, spammy. I call such messages Pressed HAM (or PHAM for short) — they are not quite SPAM, but like Pressed HAM, they are not as palatable as real HAM. Many solicited advertisements and business newsletters would fall into this category. First, it is observed that business senders are the group most able to meet the delivery burden, assuming that they have a sound and profitable business model. Secondly, the resistance to delivery creates an economic incentive for such senders to avoid some of the evils of PHAM. Much PHAM can be rendered more palatable by including a link to the information, rather than hundreds of kilobytes of HTML and images.
### 2.4 Mailing Lists
Mailing lists are relatively straight forward: A correctly configured mailing list, that carries HAM, will not be greatly penalised. In contrast, if a Spammer uses a mailing list to attempt to deliver SPAM to a broad audience, then resistance will be applied to most deliveries. To alleviate this, and to reduce resource requirements, a mailing list could refuse to deliver messages that invoke too much resistance during delivery (and alert the sender of this). This may seem severe, however, some existing mailing list providers already implement similarly harsh policies, e.g., never re-trying delivery after a temporary failure.
### 2.5 Leaking SPAM Filter Information To Spammers
Because the resistance is selectively applied to SPAM, it is possible that Spammers may try submitting successive refinements of their messages to a given mail server in order to try to reduce the spamminess of the message, and so avoid encountering resistance during delivery. To mitigate this risk, it is recommended that in practise a mail server use a single degree of resistance: Either delivery is resisted, or it is not. If graduated resistance is absolutely required, the more coarsely grained the graduation, the better. The level of resistance could include a random factor to make it more difficult to reverse engineer HAM flavoured SPAM. Having had said this, a finely graduated resistance scheme makes a lot of sense, in that it encourages senders to reduce the overall spamminess of their messages, and provides a more graceful failure mode for misclassification of messages. Therefore we briefly consider the likely impact of using a graduated resistance scheme.
Regardless of the resistance scheme that is applied, the impact of any leaked information about the SPAM filter must be considered. There are two possible scenarios: 1) There are few distinct SPAM filters; or 2) There are many distinct SPAM filters.
If there are few distinct SPAM filters, then the information leaked by mail servers is already available to Spammers, because it is feasible for them to run their message through each and (much more easily) minimise the spamminess of their message. Alternatively, there are too many SPAM filters for a Spammer to do this. In that case, it seems reasonable to assume that minimising the spamminess of a message against any one filter will be of only limited value, as other filters will target different message features. Moreover, there are characteristics of a message that a Spammer cannot readily change, such as the IP address they are sending from. Thus the task of reducing the spamminess of a given message sufficient to avoid widespread classification as SPAM is particularly difficult.
However, let us be pessimistic, and assume that a Spammer does come up with a process that enables them, by repeated submission of a message, reduce its spamminess sufficient to avoid resistance during delivery. In that situation it is possible to use a “sin bin”, that blocks delivery when this anti-social behaviour is observed, e.g., repeatedly declining to meet the delivery burden imposed on a message. Hosts in the sin bin may not submit any more mail until a timeout elapses
<sup>4</sup>
<sup>4</sup>4In theory, a particularly well organised Spammer could use an army of robots to send each successive refinement from a different IP address. However, the complexity of this arrangement, combined with the further increase in bandwidth requirements in order for the robots to coordinate their actions, would limit the value of this approach.
.
However, even if it were possible to minimise the spamminess of a message by repeated submission, the repeated delivery attempts would multiply the network bandwidth required to deliver each message, thus reducing the amount of SPAM that can be send per unit time. Therefore, the effects of any information leaked via a Non-Uniform Proof-of-Work scheme are mitigated at multiple levels.
## 3 Implementing The SPAM Friction
### 3.1 Integrating SPAM Friction Into The SMTP
A Targeted-Cost Proof-of-Work system must see the the body of an email before it can decide whether to charge for delivery. For this reason, it is not possible to incorporate it into the existing SMTP (Klensin, [2001](#bib.bib7)). Specifically, while a receiving mail server could use a 211 message to notify the sender that they must provide Proof-of-Work, there is no mechanism in SMTP that would allow the sender to deliver the Proof-of-Work receipt. Thus a new keyword would be required in SMTP, or an old one must take on a new meaning. The former seems preferable, since if the SMTP must be changed, then it should not be done crudely. I suggest the new keyword POW as the new keyword that is used to deliver a Proof-of-Work receipt to the receiving mail server. The following shows an example conversation using the new keyword:
-
250 ESMTP Server Ready
EHLO sending-mail.com
250-receiving-mail.com Hello sending-mail.com \[10.1.2.3\]
250-SIZE 52428800
250-AUTH PLAIN LOGIN
250-STARTTLS
250-SPAMFRICTION ALG0, ALG1, ALG2
250 HELP
POW ISUPPORT ALG0, ALG1, ALG4
250 OK
MAIL FROM: sender@sending-mail.com
250 OK
RCPT TO: receiver@receiving-mail.com
250 Accepted
DATA
354 Enter message, ending with ‘‘.’’ on a line by itself
Spam, spam, eggs and spam
.
211 POW Required (SPAM) 0:21:892734982734987
POW RECEIPT 0:21:892734982734987:193287436879263
250 OK id=1HCVqn-000436-HX
The POW command is issued in response to a 211 message that requests proof of work. The string 0:21:892734982734987 is the puzzle that the sender must solve, where 0 is the algorithm number, 21 is the difficulty, and the long string of numbers the remainder of the problem specification. The response contains the puzzle concatenated with the solution strings. Otherwise, the conversation conforms to the SMTP.
### 3.2 Transitional Considerations
A critical issue for any changes to the SMTP, is that backward compatibility is retained. This is why the POW capability is issued using a 250 message, and the receiver is required to acknowledge if it also supports the capability. A legacy sender that does not support the capability makes no offer, alerting the receiver of the fact by its silence. Legacy senders would be handled by introducing a lengthly delay before offering to accept a message. To prevent abuse by spammers, a receiving mail server may take several courses of action when waiting for a Proof-of-Work receipt, or during the time out period when communicating with a legacy sender: a) refuse to accept additional connections from the same sending host; b) temporarily reject messages sent from the same sending host; c) accept additional messages, but increasing the delivery resistance in accordance with the number of messages in this state. In any case, it is possible to prevent Spammers from abusing the time out too much. More to the point, the time out still makes it harder to send SPAM than is currently the case.
Finally, for the scheme to be fully effective it would require participation by the majority of SMTP servers on the Internet. The interoperability with the existing SMTP just described helps to make it possible to progressively implement the scheme, with value increasing with each SMTP server that adopts it. Thus if the protocol was implemented in the major SMTP server software packages, then wide spread adoption would be possible.
## 4 Conclusions
In this paper I have argued that a Targeted-Cost Proof-of-Work scheme has the potential to dramatically reduce SPAM volumes. Moreover, its requirements are modest: a) SPAM filters that are $`\geq 99.9`$% accurate (which already exist); b) relatively small changes to the SMTP and mail server software and configurations. The risk-benefit ratio is compelling: The chance to make SPAM 1,000 times harder to send than HAM, and limiting overall SPAM volumes to around 24 messages per sending computer per day. The obvious next step is to implement this system, including selecting at least one initial problem class, and assess its effectiveness.
## 5 Acknowledgements
I would like to acknowledge the valuable feedback and encouragement of the my colleagues in the School of Informatics & Engineering and the LabF community.
## References
- Abadi et al. \[2005\] Martin Abadi, Mike Burrows, Mark Manasse, and Ted Wobber. Moderately hard, memory-bound functions. *ACM Trans. Inter. Tech.*, 5(2):299–327, 2005. ISSN 1533-5399. doi: http://doi.acm.org/10.1145/1064340.1064341.
- Back \[2002\] A. Back. Hash cash - a denial of service counter-measure, 2002.
URL citeseer.ist.psu.edu/back02hashcash.html.
- Dwork et al. \[2002\] C. Dwork, A. Goldberg, and M. Naor. On memory-bound functions for fighting spam, 2002.
URL citeseer.ist.psu.edu/dwork02memorybound.html.
- Dwork and Naor \[1993\] Cynthia Dwork and Moni Naor. Pricing via processing or combatting junk mail. In *CRYPTO ’92: Proceedings of the 12th Annual International Cryptology Conference on Advances in Cryptology*, pages 139–147, London, UK, 1993. Springer-Verlag. ISBN 3-540-57340-2.
- Ioannidis et al. \[2005\] John Ioannidis, Angelos D. Keromytis, and Moti Yung, editors. *Applied Cryptography and Network Security, Third International Conference, ACNS 2005, New York, NY, USA, June 7-10, 2005, Proceedings*, volume 3531 of *Lecture Notes in Computer Science*, 2005. ISBN 3-540-26223-7.
- Jakobsson and Juels \[1999\] M. Jakobsson and A. Juels. Proofs of work and bread pudding protocols, 1999.
URL citeseer.ist.psu.edu/238810.html.
- Klensin \[2001\] J. Klensin. Simple Mail Transfer Protocol. RFC 2821 (Proposed Standard), April 2001.
URL http://www.ietf.org/rfc/rfc2821.txt.
- Krawetz \[2004\] Neal Krawetz. Anti-spam solutions and security, part 2, March 2004.
URL http://www.securityfocus.com/infocus/1766.
- Laurie and Clayton \[2004\] B. Laurie and R. Clayton. Proof-of-work proves not to work, May 2004.
URL citeseer.ist.psu.edu/642739.html.
- Liu and Camp \[2006\] Debin Liu and L. Jean Camp. When proof of work works. Working Papers 06-18, NET Institute, October 2006. available at http://ideas.repec.org/p/net/wpaper/0618.html.
- Rivest et al. \[1996\] R. L. Rivest, A. Shamir, and D. A. Wagner. Time-lock puzzles and timed-release crypto. Technical report, Cambridge, MA, USA, 1996.
- Rivest and Shamir \[1996\] Ronald L. Rivest and Adi Shamir. Payword and micromint: Two simple micropayment schemes. In *Security Protocols Workshop*, pages 69–87, 1996.
URL citeseer.ist.psu.edu/rivest96payword.html.
- Wang and Reiter \[2003\] XiaoFeng Wang and Michael K. Reiter. Defending against denial-of-service attacks with puzzle auctions. In *SP ’03: Proceedings of the 2003 IEEE Symposium on Security and Privacy*, page 78, Washington, DC, USA, 2003. IEEE Computer Society. ISBN 0-7695-1940-7. |
0709.0961 | cs.NI | cyber | # A New Distributed Topology Control Algorithm for Wireless Environments with Non-Uniform Path Loss and Multipath Propagation
Harish Sethu and Thomas Gerety
Department of Electrical and Computer Engineering
Drexel University
Philadelphia, PA 19104-2875
Email: {sethu, thomas.gerety}@drexel.edu
###### Abstract
Each node in a wireless multi-hop network can adjust the power level at which it transmits and thus change the topology of the network to save energy by choosing the neighbors with which it directly communicates. Many previous algorithms for distributed topology control have assumed an ability at each node to deduce some location-based information such as the direction and the distance of its neighbor nodes with respect to itself. Such a deduction of location-based information, however, cannot be relied upon in real environments where the path loss exponents vary greatly leading to significant errors in distance estimates. Also, multipath effects may result in different signal paths with different loss characteristics, and none of these paths may be line-of-sight, making it difficult to estimate the direction of a neighboring node. In this paper, we present Step Topology Control (STC), a simple distributed topology control algorithm which reduces energy consumption while preserving the connectivity of a heterogeneous sensor network without use of any location-based information. The STC algorithm avoids the use of GPS devices and also makes no assumptions about the distance and direction between neighboring nodes. We show that the STC algorithm achieves the same or better order of communication and computational complexity when compared to other known algorithms that also preserve connectivity without the use of location-based information. We also present a detailed simulation-based comparative analysis of the energy savings and interference reduction achieved by the algorithms. The results show that, in spite of not incurring a higher communication or computational complexity, the STC algorithm performs better than other algorithms in uniform wireless environments and especially better when path loss characteristics are non-uniform.
## I Introduction
In a multi-hop wireless sensor network, a node communicates with another node across one or more consecutive wireless links with messages possibly passing through intermediate nodes. The topology of such a network can be viewed as a graph with an edge connecting any pair of nodes that can communicate with each other directly without going through any intermediate nodes. Each node in such a network can choose its own neighbors and thus control the topology by changing the power at which it makes its transmissions or, in the case of nodes capable of directional transmissions, by also changing the set of directions in which it will allow transmissions. The goal of such topology control is to employ algorithms that each node can execute in a distributed manner for the purposes of reducing energy consumption, maintaining connectivity, and increasing network lifetime and/or capacity.
In recent years, a large number of topology control algorithms have been proposed and studied for a diverse set of goals \[[1](#bib.bib1)\]. Early work on topology control assumed that accurate location information about its neighbors will be available to the nodes, such as through the use of GPS devices \[[2](#bib.bib2), [3](#bib.bib3), [4](#bib.bib4), [5](#bib.bib5), [6](#bib.bib6)\]. This assumption adds to the expense of the nodes and also results in high delays due to the acquiring and tracking of satellite signals. Also, one cannot rely on GPS in many real application environments such as inside buildings or thick forests. Some other topology control protocols that preserve connectivity rely on the more likely ability of a node to estimate the distance and direction to its neighbors. For example, in the cone-based distributed topology control (CBTC) algorithms, a node $`u`$ transmits with the minimum power $`p_{u,\alpha}`$ required to ensure that there is some node it can reach within every cone of degree $`\alpha`$ around $`u`$ \[[7](#bib.bib7)\]. Assuming a specific loss propagation model, the Euclidean distance to a neighbor can be deduced with knowledge of the power at which a transmission is made by a neighbor and the power at which the signal is received. The direction of a neighbor with respect to itself can be deduced from the angle of arrival of a signal.
Wireless communication, however, is often characterized by the phenomenon of multipath propagation wherein a signal reaches the receiving antenna via two or more paths \[[8](#bib.bib8)\]. In addition, there are several other kinds of radio irregularities that have an impact on the topology control algorithms \[[9](#bib.bib9)\]. The different paths, with differences in delay, attenuation, and phase shift, make it difficult for the receiving node to deduce its distance from the sender and the direction of the sender. In this paper, we focus on the design of topology control algorithms that work without the use of any location-based information so that they can be employed in the presence of multipath propagation or when path loss exponents are non-uniform in the region of interest. More specifically, we focus on connectivity-preserving algorithms that make no assumptions about the availability of GPS devices in nodes and no assumptions on the ability of a receiving node to deduce either the distance or the direction of the sender. Besides accommodating the environmental causes of radio irregularities, we seek to also meet the requirements of a heterogeneous sensor network where (a) different nodes may have different maximum transmission powers, and (b) variations in node/antenna configurations may lead to different reception thresholds in different directions.
###
I-A
Problem Statement
Assume that each node $`u \in V`$ is associated with a certain maximum power $`P_{u}`$ with which it is capable of making an omni-directional transmission (for ease of discussion, we use omni-directional transmissions but our problem statement and the proposed algorithm can be readily adapted for directional transmissions). Note that we allow $`P_{u}`$ to be different for different nodes, allowing a heterogeneous sensor network environment as in \[[10](#bib.bib10)\]. Consider the nodes in the network as vertices of a directed graph $`G_{m\hspace{0pt}a\hspace{0pt}x} = {(V,E_{m\hspace{0pt}a\hspace{0pt}x})}`$ in which nodes $`u`$ and $`v`$ are connected by a directed edge from $`u`$ to $`v`$ if and only if (a) an omni-directional transmission from $`u`$ at its maximum power $`P_{u}`$ can directly reach $`v`$, and (b) an omni-directional transmission from $`v`$ at its maximum power $`P_{v}`$ can directly reach $`u`$. Note that if $`{(u,v)} \in E_{m\hspace{0pt}a\hspace{0pt}x}`$, then $`{(v,u)} \in E_{m\hspace{0pt}a\hspace{0pt}x}`$. Many widely deployed MAC and address resolution protocols in wireless networks not only assume bidirectional links but also assume two-way handshakes and acknowledgments \[[11](#bib.bib11)\]. Therefore, with bidirectional communication assumed between directly communicating nodes, $`G_{m\hspace{0pt}a\hspace{0pt}x}`$ represents a realistic communication topology at maximum node powers.
Given $`G_{m\hspace{0pt}a\hspace{0pt}x} = {(V,E_{m\hspace{0pt}a\hspace{0pt}x})}`$, the goal of topology control in this paper can be thought of as a multi-objective optimization problem where we seek a new weighted directed graph $`G = {(V,E)}`$, $`E \subseteq E_{m\hspace{0pt}a\hspace{0pt}x}`$. We define $`P_{G}\hspace{0pt}{(u)}`$ as the minimum power with which node $`u`$ should make an omni-directional transmission so as to reach all of its neighbors in graph $`G`$. Let $`C_{G}\hspace{0pt}{(u)}`$ denote the energy cost of a transmission by $`u`$ at power $`P_{G}\hspace{0pt}{(u)}`$.
<sup>1</sup>
<sup>1</sup>1Note that the power at which a node makes its transmissions is the energy consumed in the transmission per unit time. When the context is a single node, reducing power and reducing energy become the same goal because when we reduce transmission power at a given node, we also reduce energy consumption at the node. The energy consumption by the overall network, however, is not always reduced when the transmission powers at the nodes are reduced. In this paper, we use the goals of reducing the energy cost and reducing the transmission power interchangeably when the context is limited to the transmissions by a single node. In the context of an entire network, we speak of the goal of reducing energy costs.
Assign weight $`C_{G}\hspace{0pt}{(u)}`$ to each directed edge $`{(u,v)} \in E`$ (all edges starting from a node have the same weight since nodes in a sensor network broadcast messages at a certain constant power level determined by the topology control algorithm). Let $`C_{G}\hspace{0pt}{({s\rightarrow{t,{E\hspace{0pt}n\hspace{0pt}e\hspace{0pt}r\hspace{0pt}g\hspace{0pt}y}}})}`$ denote the sum of $`C_{G}\hspace{0pt}{(i)}`$ for each transmitting node $`i`$ in the minimum-energy path from node $`s`$ to node $`t`$ in the graph $`G`$ (this represents the minimum energy cost for a message to traverse from a source $`s`$ to a destination $`t`$).
The multi-objective problem is to find $`G`$ with the following two objectives:
1. 1\.
Minimize the average power, computed over all nodes, required in an omni-directional transmission by a node $`u`$ to reach all of its neighbors in $`G`$. This objective can be expressed as:
<table>
<tbody>
<tr>
<td></td>
<td>min[∑u∈VCG(u)]subscript𝑢𝑉subscript𝐶𝐺𝑢\min\left[\sum_{u\in V}C_{G}(u)\right]</td>
<td></td>
</tr>
</tbody>
</table>
2. 2\.
Minimize the average energy cost, computed over all source-destination pairs, in a minimum-energy path from a source node to a destination node. This objective can be expressed as:
<table>
<tbody>
<tr>
<td></td>
<td>min[∑s,t∈VCG(s→t,Energy)]subscript𝑠𝑡𝑉subscript𝐶𝐺→𝑠𝑡E𝑛𝑒𝑟𝑔𝑦\min\left[\sum_{s,t\in V}C_{G}(s\rightarrow t,{\mathrm{E}nergy})\right]</td>
<td></td>
</tr>
</tbody>
</table>
under the following constraints:
1. 1\.
$`E \subseteq E_{m\hspace{0pt}a\hspace{0pt}x}`$.
2. 2\.
If $`{(u,v)} \in E`$, then $`{(v,u)} \in E`$, as is generally expected by MAC layer protocols.
3. 3\.
If there exists a path between $`u`$ and $`v`$ in $`G_{m\hspace{0pt}a\hspace{0pt}x}`$, then there also exists a path between $`u`$ and $`v`$ in $`G`$ (preserves connectivity).
The two objectives of minimizing the average transmission power of a node and minimizing the average energy cost along a minimum-energy path often work against each other. For example, minimizing average transmission power can lead to a very sparse graph with very long energy-expensive multi-hop paths between some pairs of nodes.
###
I-B
Related Work
01:
Function STC at node $`(u)`$:
02:
$`G\leftarrow(V,\phi)`$ /\* directed graph with no edges \*/
03:
Compile outTupleList$`(u)`$ and inTupleList$`(u)`$
04:
Broadcast outTupleList$`(u)`$ and inTupleList$`(u)`$ at maximum power $`P_{u}`$
05:
Receive outTupleList$`(n)`$ and inTupleList$`(n)`$ from each neighbor $`n`$ in $`G_{m\hspace{0pt}a\hspace{0pt}x}`$
06:
Compute fPairOfPaths$`(n)`$ for each $`n`$ two or fewer hops away from $`u`$
07:
Compute bPairOfPaths$`(n)`$ for each $`n`$ two or fewer hops away from $`u`$
08:
Sort outTupleList$`(u)`$
09:
$`k\leftarrow`$ degree of $`u`$ in $`G_{m\hspace{0pt}a\hspace{0pt}x}`$
10:
do $`k - 1`$ times
11:
$`{t\hspace{0pt}(u,v)}\leftarrow`$ the largest tuple in outTupleList$`(u)`$
12:
remove $`t\hspace{0pt}(u,v)`$ from outTupleList$`(u)`$
13:
vSet $`= \left\{ {i\hspace{0pt}\ } \middle| {{{\ \hspace{0pt}t\hspace{0pt}(i,v)} \in {i\hspace{0pt}n\hspace{0pt}T\hspace{0pt}u\hspace{0pt}p\hspace{0pt}l\hspace{0pt}e\hspace{0pt}L\hspace{0pt}i\hspace{0pt}s\hspace{0pt}t\hspace{0pt}(v)}},{{t\hspace{0pt}(i,v)} < {t\hspace{0pt}(u,v)}}} \right\}`$
14:
NoForwardPath $`=`$ True
15:
NoBackwardPath $`=`$ True
16:
for $`n \in`$ vSet
17:
$`p =`$ the first path in fPairOfPaths$`(n)`$ without $`v`$
18:
if $`{\max\left\{ {\left. r \middle| r \right. \in p} \right\}} < {t\hspace{0pt}(u,v)}`$
19:
NoForwardPath $`=`$ False
20:
break (out of for loop)
21:
end if
22:
end for
23:
vSet $`= \left\{ {i\hspace{0pt}\ } \middle| {{{\ \hspace{0pt}t\hspace{0pt}(v,i)} \in {o\hspace{0pt}u\hspace{0pt}t\hspace{0pt}T\hspace{0pt}u\hspace{0pt}p\hspace{0pt}l\hspace{0pt}e\hspace{0pt}L\hspace{0pt}i\hspace{0pt}s\hspace{0pt}t\hspace{0pt}(v)}},{{t\hspace{0pt}(v,i)} < {t\hspace{0pt}(v,u)}}} \right\}`$
24:
for $`n \in`$ vSet
25:
$`p =`$ the first path in bPairOfPaths$`(n)`$ without $`v`$
26:
if $`{\max\left\{ {\left. r \middle| r \right. \in p} \right\}} < {t\hspace{0pt}(v,u)}`$
27:
NoBackwardPath $`=`$ False
28:
break (out of for loop)
29:
end if
30:
end for
31:
if NoForwardPath or NoBackwardPath
32:
add a directed edge $`(u,v)`$ to $`G`$
33:
end if
34:
end do
Figure 1: Pseudo-code description of the Step Topology Control algorithm at node $`u`$.
Let $`C\hspace{0pt}{(u,v)}`$ denote the minimum energy cost of a successful transmission from $`u`$ to $`v`$. A topology control algorithm that minimizes energy consumption will remove an edge $`(u,v)`$ if and only if there exists a path between $`u`$ and $`v`$ through an intermediate set of nodes $`n_{1},{n_{2}\hspace{0pt}\ldots\hspace{0pt}n_{k}}`$ such that $`{{C\hspace{0pt}{(u,n_{1})}} + {C\hspace{0pt}{(n_{1},n_{2})}} + {\ldots\hspace{0pt}C\hspace{0pt}{(n_{k},v)}}} < {C\hspace{0pt}{(u,v)}}`$. Accomplishing this requires a significant exchange of information between nodes and in such a case, the topology control algorithm is indistinguishable from a routing algorithm. As a result, a number of distributed topology control algorithms have been proposed where nodes rely on lesser exchange of information between neighbors \[[1](#bib.bib1)\]. In this subsection, we focus on the subset of these protocols that can be adapted to work without the exchange of any location-based information between neighboring nodes.
The KNEIGH protocol is based on determining the number of neighbors that each node should have in order to achieve full connectivity with a high probability \[[12](#bib.bib12)\]. This protocol, however, does not guarantee connectivity even though it does achieve connectivity with a high probability. The Small Minimum-Energy Communication Network (SMECN) protocol \[[4](#bib.bib4)\] seeks to achieve a lower energy cost while guaranteeing connectivity by removing an edge $`(u,v)`$ if and only if there exists a path $`u\rightarrow n\rightarrow v`$ such that $`{{C\hspace{0pt}{(u,n)}} + {C\hspace{0pt}{(n,v)}}} < {C\hspace{0pt}{(u,v)}}`$. As proposed in \[[4](#bib.bib4)\], the protocol requires the use of GPS devices but the same results can be accomplished if each node exchanges information with each of its neighbors regarding the energy costs of reaching all of its neighbors. When used with some of the widely available routing algorithm implementations such as AODV or DSR that base their decisions on the number of hops in a path rather than the total energy cost of the path, the SMECN protocol does not necessarily result in a significant reduction in energy consumption. As will be shown in Section [IV](#S4 "IV Simulation Results ‣ A New Distributed Topology Control Algorithm for Wireless Environments with Non-Uniform Path Loss and Multipath Propagation"), the energy savings achieved by SMECN is not significantly greater even if the routing algorithm were to choose a minimum-energy path.
The Directed Relative Neighborhood Graph (DRNG) protocol \[[10](#bib.bib10)\] removes an edge $`(u,v)`$ if and only if there exists a path $`u\rightarrow n\rightarrow v`$ such that $`{\max{\{{C\hspace{0pt}{(u,n)}},{C\hspace{0pt}{(n,v)}}\}}} < {C\hspace{0pt}{(u,v)}}`$. In the Directed Local Spanning Subgraph (DLSS) protocol \[[10](#bib.bib10)\], each node creates a local spanning tree from the subgraph induced by itself and its neighbors. A node $`u`$ retains the edge $`(u,v)`$ in the topology-controlled graph if and only if the edge $`(u,v)`$ exists in the local spanning tree generated at node $`u`$. As proved in \[[10](#bib.bib10)\], any edge that is removed by DRNG is also removed by DLSS and therefore, DLSS achieves a sparser graph than DRNG. The fewer edges in DLSS often leads to a lower energy consumption, but it can sometimes lead to longer paths and therefore higher energy consumption than DRNG. A generalization of the DRNG protocol is the XTC protocol which was conceived independently and uses the notion of ‘link quality’ in making its topology control decision instead of energy costs \[[13](#bib.bib13)\]. When link quality between two nodes is measured by the energy cost of a transmission between them, the XTC protocol reduces to DRNG.
Among other attempts to accommodate the irregularities of a real wireless environment in topology control are some that allow for uncertainties in whether or not a nearby node is reachable if the distance to that node is above a certain threshold \[[14](#bib.bib14), [15](#bib.bib15)\]. However, these algorithms assume that each node can know the distances to other nearby nodes, something that cannot be relied upon in a real environments with multipath effects. Some other works have also considered realistic wireless models but they have only presented centralized algorithms \[[16](#bib.bib16)\].
###
I-C
Contributions
In this paper, we present the Step Topology Control (STC) algorithm in which a node $`u`$ removes an edge $`(u,v)`$ if and only if there exists:
- •
a path with three or fewer hops from $`u`$ to $`v`$ such that the energy cost across each hop is less than $`C\hspace{0pt}{(u,v)}`$.
- •
a path with three or fewer hops from $`v`$ to $`u`$ such that the energy cost across each hop is less than $`C\hspace{0pt}{(v,u)}`$.
The STC algorithm relies on each node exchanging information with each of its neighbors regarding the energy costs of communication to all of its neighbors.
The STC algorithm may be seen as an extension of the DRNG algorithm to allow a search for three-hop paths and to ensure bidirectional communication between directly communicating nodes. However, our implementation of the algorithm ensures that the order of communication and the computational complexity do not increase despite a search for three-hop paths. This search for three-hop paths makes only a small impact on performance when the wireless environment is uniform across the network. However, it makes a significant impact when a more realistic scenario is assumed with multipath propagation and other irregularities in the path loss characteristics in the wireless environment. In such environments, the STC algorithm achieves a much lower energy consumption and interference than other topology control algorithms in spite of maintaining the same or better order of communication and computational complexity. We also show that this improved performance of the STC algorithm exists largely independent of the size of the network.
Section [II](#S2 "II Step Topology Control ‣ A New Distributed Topology Control Algorithm for Wireless Environments with Non-Uniform Path Loss and Multipath Propagation") presents the STC algorithm along with a pseudo-code description. Section [III](#S3 "III Comparative Analysis ‣ A New Distributed Topology Control Algorithm for Wireless Environments with Non-Uniform Path Loss and Multipath Propagation") presents an analysis of the communication and computational complexity of the STC algorithm and presents a comparative analysis. Section [IV](#S4 "IV Simulation Results ‣ A New Distributed Topology Control Algorithm for Wireless Environments with Non-Uniform Path Loss and Multipath Propagation") presents several simulation-based results that provide a thorough comparison of the energy consumption properties of the STC algorithm and other existing topology control strategies that can be adapted to use no location-based information. In particular, we examine the algorithms in environments with both constant and varied path loss exponents and study their scalability in performance as the number of nodes increases. Section [V](#S5 "V Concluding Remarks ‣ A New Distributed Topology Control Algorithm for Wireless Environments with Non-Uniform Path Loss and Multipath Propagation") concludes the paper with a summary of its findings and future research directions. A proof of the relationship between the STC algorithm and a CBTC algorithm with all applicable optimizations is presented in the Appendix.
## II Step Topology Control
Denote by $`P_{m\hspace{0pt}i\hspace{0pt}n}\hspace{0pt}{(u,v)}`$, the minimum power necessary for a transmission from $`u`$ to reach $`v`$, otherwise known as transmission power threshold. We allow that $`P_{m\hspace{0pt}i\hspace{0pt}n}\hspace{0pt}{(u,v)}`$ is not necessarily the same as $`P_{m\hspace{0pt}i\hspace{0pt}n}\hspace{0pt}{(v,u)}`$. The basic idea behind the Step Topology Control (STC) algorithm is to find both forward and backward paths with three or fewer hops between $`u`$ and $`v`$ such that each hop requires a lower energy cost than that required for an equivalent direct transmission between $`u`$ and $`v`$. If such multi-hop paths exist between $`u`$ and $`v`$, node $`u`$ drops the edge $`(u,v)`$ from the directed graph it generates and node $`v`$ similarly drops the edge $`(v,u)`$.
The STC algorithm relies on being able to uniquely order the energy costs of transmissions across nodes. Since the power levels at which transmissions are made may take on only certain discrete values in some systems, we add additional identifiers to permit a unique ordering. We assume that each node $`u`$ is uniquely identified by an integer ID<sub>u</sub>. For each ordered pair of nodes $`u`$ and $`v`$, we associate an ordered tuple $`{t\hspace{0pt}{(u,v)}} = {(t_{1},t_{2},t_{3})}`$, where $`t_{1} = {P_{m\hspace{0pt}i\hspace{0pt}n}\hspace{0pt}{(u,v)}}`$, $`t_{2} = {ID}_{u}`$, and $`t_{3} = {ID}_{v}`$. We say that $`{(t_{1},t_{2},t_{3})} < {(t_{1}',t_{2}',t_{3}')}`$ if and only if (1) $`t_{1} < t_{1}'`$, or (2) $`t_{1} = t_{1}'`$ and $`t_{2} < t_{2}'`$, or (3) $`t_{1} = t_{1}'`$, $`t_{2} = t_{2}'`$, and $`t_{3} < t_{3}'`$. For the sake of completeness, we define $`{t\hspace{0pt}{(u,u)}} < {t\hspace{0pt}{(x,y)}}`$ for any $`u`$ and $`x \neq y`$ since a transmission to itself should cost less energy than a transmission to another node. Note that $`t\hspace{0pt}{(u,v)}`$ and $`t\hspace{0pt}{(v,u)}`$ are strictly ordered by the above lexicographic rule and not equal even if the minimum power required for transmission between $`u`$ and $`v`$ is the same in either direction. We call $`t\hspace{0pt}{(u,v)}`$ a transmission tuple. A path $`p`$ in the graph consists of an ordered sequence of transmission tuples. Denote by maxTuple$`(p)`$ the largest tuple in path $`p`$.
Consider any two nodes $`u`$ and $`v`$ such that $`{(u,v)} \in E_{m\hspace{0pt}a\hspace{0pt}x}`$. We assume that $`u`$ can determine the minimum power necessary for its transmission to reach $`v`$ as well as the minimum power required for a transmission from $`v`$ to reach itself. This is accomplished by transmitting beacon messages at increasing powers and noting the power at which each neighbor is first discovered. Each beacon message carries within it the power at which it is transmitted so that the discovered neighbor may also note the minimum power necessary for it to be reached by a neighboring node. Each node $`u`$ can thus compile two lists of transmission tuples: outTupleList$`(u)`$ containing $`t\hspace{0pt}{(u,v)}`$ for all $`{(u,v)} \in E_{m\hspace{0pt}a\hspace{0pt}x}`$, and inTupleList$`(u)`$ containing $`t\hspace{0pt}{(v,u)}`$ for all $`{(v,u)} \in E_{m\hspace{0pt}a\hspace{0pt}x}`$. This process of exchanging power level information is feasible in practice and is part of many proposed energy-aware MAC layer protocols as well as topology control algorithms such as DLSS.
Figure [1](#S1.F1 "Figure 1 ‣ I-B Related Work ‣ I Introduction ‣ A New Distributed Topology Control Algorithm for Wireless Environments with Non-Uniform Path Loss and Multipath Propagation") presents a pseudo-code description of the STC algorithm. Once a node $`u`$ compiles outTupleList$`(u)`$ and inTupleList$`(u)`$, it begins execution of the algorithm. Each node $`u`$ first broadcasts both its inTupleList and outTupleList at its maximum power $`P_{u}`$ to reach all of its neighbors (line 04). The node also collects the inTupleLists and outTupleLists from each of its neighbors (line 05).
Given all this information about the energy costs, node $`u`$ computes two forward paths (denoted fPairOfPaths$`(n)`$) and two backward paths (denoted bPairOfPaths$`(n)`$) for each node $`n`$ that is reachable by two or fewer hops (lines 06–07). We describe below in greater detail the construction of the forward path data structures for node $`n`$ (the construction of the backward pair of paths is similar). Let $`p_{1}`$ denote the first of the pair of paths in fPairOfPaths$`(n)`$ and let $`p_{2}`$ denote the second. We choose $`p_{1}`$ and $`p_{2}`$ such that maxTuple$`{(p_{1})} <`$ maxTuple$`{(p_{2})} <`$ maxTuple$`(p)`$ where $`p`$ is any path of two or fewer hops from $`u`$ to $`n`$ other than $`p_{1}`$ and $`p_{2}`$. In lines 16–22, this data structure allows the node to quickly determine if there exists a path $`s`$ of three or fewer hops between $`u`$ and another node $`v`$ such that maxTuple$`{(s)} < {t\hspace{0pt}{(u,v)}}`$. The reason we need a pair of paths instead of just one path is because one of these paths may be through $`v`$ and one would not choose to replace the edge $`(u,v)`$ with a path that goes through $`v`$.
Node $`u`$ then orders the tuples in its outTupleList$`(u)`$ (line 08) and considers each of the edges $`(u,v)`$ in reverse lexicographical order of the associated tuples $`t\hspace{0pt}{(u,v)}`$, i.e., the neighbor that requires the largest power to be reached is considered first (line 11). As each neighbor $`v`$ is processed, the corresponding tuple $`t\hspace{0pt}{(u,v)}`$ is removed from outTupleList$`(u)`$ (line 12).
To determine if an edge $`(u,v)`$ should be removed, a node $`u`$ looks for a forward path $`u\rightarrow n_{1}\rightarrow n_{2}\rightarrow v`$, where the nodes $`n_{1}`$ and $`n_{2}`$ may or may not be distinct (lines 13–22). The condition that the path should satisfy is $`{\max{\{{t\hspace{0pt}{(u,n_{1})}},{t\hspace{0pt}{(n_{1},n_{2})}},{t\hspace{0pt}{(n_{2},v)}}\}}} < {t\hspace{0pt}{(u,v)}}`$. The node similarly seeks to find a reverse path via nodes $`n_{3}`$ and $`n_{4}`$ (lines 23–33) such that $`{\max{\{{t\hspace{0pt}{(v,n_{3})}},{t\hspace{0pt}{(n_{3},n_{4})}},{t\hspace{0pt}{(n_{4},u)}}\}}} < {t\hspace{0pt}{(v,u)}}`$. Note that the number of distinct nodes among $`n_{1}`$, $`n_{2}`$, $`n_{3}`$ and $`n_{4}`$ may range anywhere between 1 and 4. If both forward and backward two- or three-hop paths are found satisfying the desired conditions on the tuples, the edge $`(u,v)`$ is removed from the graph (lines 31–33).
To determine if there exists a forward path from $`u`$ to $`v`$ satisfying the desired conditions, node $`u`$ first constructs the set vSet consisting of nodes $`i`$ that are neighbors of $`v`$ such that $`{t\hspace{0pt}{(i,v)}} < {t\hspace{0pt}{(u,v)}}`$ (line 13). Now, if there exists a path, $`p`$, of two or fewer hops from $`u`$ to any node in vSet such that the path is not through $`v`$ and maxTuple$`{(p)} < {t\hspace{0pt}{(u,v)}}`$, then the condition for the forward path is satisfied. This is determined in lines 16–22 and the existence of a backward path is similarly determined in lines 24–30.
## III Comparative Analysis
In this section, we discuss the communication and computational complexity of the STC algorithm in comparison to other topology control algorithms that also preserve connectivity while being capable of operation in real environments with multipath propagation or non-uniform path loss exponents. We also discuss any provable relationships between STC and other algorithms regarding the set of edges removed from $`G_{m\hspace{0pt}a\hspace{0pt}x}`$ by the algorithm.
###
III-A
Complexity
We consider the complexity of the following algorithms and present a comparison to the STC algorithm:
- •
Small Minimum Energy Communication Network (SMECN).
- •
Directed Relative Neighborhood Graph (DRNG).
- •
Directed Local Spanning Subgraph (DLSS).
Each of the above algorithms relies on each node first determining the energy cost to each of its neighbors. As explained in Section [II](#S2 "II Step Topology Control ‣ A New Distributed Topology Control Algorithm for Wireless Environments with Non-Uniform Path Loss and Multipath Propagation"), this can be accomplished by transmitting beacon messages at increasing powers and noting the power at which each neighbor is first discovered. When beacon messages carry the power at which they are being transmitted, each node can also learn the energy cost of the communication from each of its neighbors to itself. Depending on the granularity of power levels at which transmissions can be made, a variety of strategies based on linear or binary search methods may be employed in these steps to minimize interference and to maximize speed of convergence to the correct power values. We do not include this step in our complexity analysis because it is common to all topology control protocols above and cannot be considered a distinguishing feature of STC or any other specific protocol. Topology control algorithms begin with the assumption that each node can determine and will know the minimum power at which it should transmit to reach another particular node. For the STC algorithm, for example, we consider its complexity at node $`u`$ after the compilation of outTupleList$`(u)`$ and inTupleList$`(u)`$ is complete.
In the original graph, $`G_{m\hspace{0pt}a\hspace{0pt}x}`$, on which a topology control algorithm is executed, the in-degree of a node is the same as its out-degree (since bidirectional communication is assumed) and therefore, the maximum out-degree ($`\Delta^{+}`$) and the maximum in-degree ($`\Delta^{-}`$) are identical. We define the node-degree of a node in $`G_{m\hspace{0pt}a\hspace{0pt}x}`$ as the number of its neighbors that it communicates with, i.e., its out-degree or its in-degree. In the following, we denote the maximum node-degree of the original graph by $`\Delta = \Delta^{+} = \Delta^{-}`$. Further, in our analysis, we do not include the ID size in computing the communication complexity. This is because the IDs on sensor nodes are likely to be globally unique for each node and assigned by the manufacturer, as in the case of Ethernet cards. We expect the ID sizes, therefore, to be independent of the size of the sensor network deployment.
######
Theorem III.1
The communication complexity of the STC algorithm is $`O\hspace{0pt}{(\Delta^{2})}`$.
###### Proof:
For any given node $`n`$, the lists inTupleList$`(n)`$ and outTupleList$`(n)`$ are each of length $`O\hspace{0pt}{(\Delta)}`$. Broadcasting the lists (line 04 in Figure [1](#S1.F1 "Figure 1 ‣ I-B Related Work ‣ I Introduction ‣ A New Distributed Topology Control Algorithm for Wireless Environments with Non-Uniform Path Loss and Multipath Propagation")), therefore, is $`O\hspace{0pt}{(\Delta)}`$ in communication complexity and receiving the lists from each of up to $`\Delta`$ neighbors (line 05) is $`O\hspace{0pt}{(\Delta^{2})}`$ in communication complexity. The overall communication complexity, therefore, is $`O\hspace{0pt}{(\Delta^{2})}`$. ∎
######
Theorem III.2
The computational complexity of the STC algorithm is $`O\hspace{0pt}{(\Delta^{2})}`$.
###### Proof:
For each neighbor $`n'`$ of $`u`$, one can process the entries in the outTupleList$`(n')`$ to create the pair of paths in fPairOfPaths. The determination of whether or not a path to node $`n`$ should be included in fPairOfPaths and whether it should be the first or the second of the pair of paths can be made in $`O\hspace{0pt}{(1)}`$ time since it involves no more than two tuple comparisons. Let $`h`$ denote the number of one-hop or two-hop paths starting from $`u`$. Since $`h = {O\hspace{0pt}{(\Delta^{2})}}`$, creating the fPairOfPaths will require a total time of $`O\hspace{0pt}{(\Delta^{2})}`$. Creating the bPairOfPaths will similarly require a total time of $`O\hspace{0pt}{(\Delta^{2})}`$.
Sorting of the outTupleList$`(u)`$ in line 08 takes time $`O\hspace{0pt}{({\Delta\hspace{0pt}{\log\Delta}})}`$ since the size of the list is $`O\hspace{0pt}{(\Delta)}`$.
<table>
<thead>
<tr>
<th>Algorithm</th>
<th>Communication Complexity</th>
<th>Computational Complexity</th>
</tr>
</thead>
<tbody>
<tr>
<th>SMECN</th>
<td>O(Δ2)𝑂superscriptΔ2O(\Delta^{2})</td>
<td>O(Δ2)𝑂superscriptΔ2O(\Delta^{2})</td>
</tr>
<tr>
<th>DRNG</th>
<td>O(Δ2)𝑂superscriptΔ2O(\Delta^{2})</td>
<td>O(Δ2)𝑂superscriptΔ2O(\Delta^{2})</td>
</tr>
<tr>
<th>DLSS</th>
<td>O(Δ2)𝑂superscriptΔ2O(\Delta^{2})</td>
<td>O(Δ2logΔ)𝑂superscriptΔ2ΔO(\Delta^{2}\log\Delta)</td>
</tr>
<tr>
<th>STC</th>
<td>O(Δ2)𝑂superscriptΔ2O(\Delta^{2})</td>
<td>O(Δ2)𝑂superscriptΔ2O(\Delta^{2})</td>
</tr>
</tbody>
</table>
TABLE I: A Comparison between Topology Control Algorithms
We now show that the $`2\hspace{0pt}h`$ pairs of paths created above can allow the inner for loops of lines 16–22 and 24–30 to complete in time $`O\hspace{0pt}{(\Delta)}`$. It is readily verified that each iteration of the for loops completes in $`O\hspace{0pt}{(1)}`$ time (lines 17–21 and 25–29). Since the size of the vSet is $`O\hspace{0pt}{(\Delta)}`$, these inner for loops execute in $`O\hspace{0pt}{(\Delta)}`$ time. The only other non-loop component inside the outer do loop between lines 10–34 that takes more than $`O\hspace{0pt}{(1)}`$ time is the creation of the vSet itself, which takes $`O\hspace{0pt}{(\Delta)}`$ time. Since the do loop iterates $`k - 1`$ times where $`k \leq \Delta`$, the entire do loop completes in time $`O\hspace{0pt}{(\Delta^{2})}`$.
The overall computational complexity of the STC algorithm, therefore, is $`O\hspace{0pt}{(\Delta^{2})}`$.
∎
Table [I](#S3.T1 "TABLE I ‣ Proof: ‣ III-A Complexity ‣ III Comparative Analysis ‣ A New Distributed Topology Control Algorithm for Wireless Environments with Non-Uniform Path Loss and Multipath Propagation") presents a comparison of the communication and computational complexities of topology control algorithms that preserve connectivity and which do not employ the exchange of any location-based information between neighbors. All involve a communication complexity of $`O\hspace{0pt}{(\Delta^{2})}`$ since they all require that each node collect a list of energy costs from each of its neighbors. In the computational complexity analysis of DLSS, we consider the number of nodes in the local graph as $`O\hspace{0pt}{(\Delta)}`$ and the number of edges as $`O\hspace{0pt}{(\Delta^{2})}`$. The computation of the local minimum spanning tree, assuming Kruskal’s algorithm \[[17](#bib.bib17)\], is $`O\hspace{0pt}{({\Delta^{2}\hspace{0pt}{\log\Delta}})}`$.
###
III-B
Set of edges removed
The SMECN, DRNG and DLSS algorithms allow directed edges in the graph they generate. Since STC assumes bidirectional links motivated by the need of real MAC layer protocols, in order to make a fair comparison, we will assume that energy costs are the same in both forward and reverse directions (so that SMECN, DRNG and DLSS will also lead to only undirected edges in the graphs they generate). Under this assumption, the STC algorithm removes any edge from the original graph that is also removed by SMECN or DRNG. The SMECN algorithm removes an edge $`(u,v)`$ if there exists a two-hop path from $`u`$ to $`v`$ through $`n`$ such that $`{{C\hspace{0pt}{(u,n)}} + {C\hspace{0pt}{(n,v)}}} < {C\hspace{0pt}{(u,v)}}`$. Therefore, we have $`{C\hspace{0pt}{(u,n)}} < {C\hspace{0pt}{(u,v)}}`$ and $`{C\hspace{0pt}{(n,v)}} < {C\hspace{0pt}{(u,v)}}`$ which would be the condition that will lead the DRNG or STC algorithm to also remove edge $`(u,v)`$. Assuming that the energy costs are the same in both forward and reverse directions, the STC algorithm becomes a simple extension of the DRNG algorithm, and therefore, it is easily argued that it will remove all edges that are removed by DRNG. Thus, the STC algorithm yields a lower energy cost per transmission at any given node than either SMECN or DRNG under these conditions. However, the STC algorithm does not always remove an edge that would be removed by DLSS. For a comparative analysis with the DLSS algorithm, we rely on the simulation results presented in the next section.
It is of interest to note that the STC algorithm is related to the OPT-CBTC($`{5\hspace{0pt}\pi}/6`$) algorithm (which is the CBTC($`{5\hspace{0pt}\pi}/6`$) algorithm with all applicable optimizations). The STC algorithm removes all edges that would be removed by OPT-CBTC($`{5\hspace{0pt}\pi}/6`$). As a result of this relationship, proved in Appendix A, the STC algorithm exhibits some of the same angular properties as OPT-CBTC($`{5\hspace{0pt}\pi}/6`$).
Since DRNG, DLSS and STC do not necessarily reduce energy costs in the path from $`u`$ to $`v`$ each time they remove an edge $`(u,v)`$, rare pathological cases are possible where these algorithms significantly increase energy costs. Consider an edge $`(u,v)`$ corresponding to an energy cost of $`C\hspace{0pt}{(u,v)}`$. It is possible that the STC algorithm removes this edge because there exist three other edges, each corresponding to an energy cost of $`{C\hspace{0pt}{(u,v)}} - \epsilon`$, where $`\epsilon`$ is an infinitesimal quantity (thus almost tripling the energy cost between $`u`$ and $`v`$). These three edges, in turn may be removed because each can be replaced by three edges corresponding to energy costs of $`{C\hspace{0pt}{(u,v)}} - {2\hspace{0pt}\epsilon}`$, which now multiplies the total energy cost between $`u`$ and $`v`$ by nine. Such scenarios are rare in real situations and our simulation results, described in the next section, show that when nodes are scattered randomly in a region, the STC algorithm actually performs significantly better than the other algorithms considered in this paper.
## IV Simulation Results
In this section, we present a simulation-based comparative study of the following distributed topology control algorithms:
- •
Step Topology Control (STC), the algorithm presented in this paper.
- •
OPT-CBTC($`{5\hspace{0pt}\pi}/6`$), as a representative CBTC algorithm and because of its relationship to the STC algorithm.
- •
Small Minimum Energy Communication Network (SMECN).
- •
Directed Relative Neighborhood Graph (DRNG).
- •
Directed Local Spanning Subgraph (DLSS).
The XTC protocol uses the notion of ‘link quality’ in generating the output topology \[[13](#bib.bib13)\]. When the link quality is measured by the energy cost of a transmission across the link, the XTC protocol reduces to the DRNG protocol. Therefore, the XTC protocol is not separately included in these comparisons.
In addition to the above, we use the following two additional algorithmic metrics as references on the performance bounds of topology control algorithms:
- •
Minimum Spanning Tree (MST): The MST achieves the first stated objective of the topology control problem described in Section [](#S1.SS1 "I-A Problem Statement ‣ I Introduction ‣ A New Distributed Topology Control Algorithm for Wireless Environments with Non-Uniform Path Loss and Multipath Propagation")
I-A
. The MST preserves connectivity while minimizing the average power (computed over all nodes) with which a node makes its transmissions.
- •
MinReach: In MinReach, each transmission from a node to one of its neighbors uses exactly the minimum energy required to reach that particular neighbor (as opposed to each node making all its transmissions to any neighbor at the same power level). The total energy cost of transmissions from a source to a destination along the minimum energy path using MinReach is a lower bound (though, not a tight lower bound) on the total energy cost along a path between the source-destination pair
<sup>2</sup>
<sup>2</sup>2Some other topology control algorithms, such as a distributed algorithm for constructing a Gabriel Graph \[[18](#bib.bib18)\], cannot serve as optimal solutions or lower bounds in our case where different nodes may have different maximum transmission powers and where path loss exponents may vary in different directions for the same transmitting node.
.
###
IV-A
Metrics
Assume $`P_{u} = P`$ for all nodes $`u`$ (recall that $`P_{u}`$ denotes the maximum power with which a node $`u`$ can transmit). If the power $`P`$ is assumed to be an arbitrary quantity in our simulations, the reduction in energy achieved by topology control algorithms can be misleading (if $`P`$ is arbitrarily large, the reduction will be similarly large). Therefore, in our simulations, we assume the smallest possible value of $`P`$ so that the original graph used as the input to the topology algorithm is connected. This offers a standardized approach to estimating the effectiveness of topology control algorithms so that arbitrarily large reductions in energy consumption cannot be claimed by topology control algorithms by simply using a dense highly-connected $`G_{m\hspace{0pt}a\hspace{0pt}x}`$. For each network, the baseline for our comparisons is the initial graph, $`H`$, defined as follows. Consider graph $`G`$ generated by creating an edge $`(u,v)`$ from $`u`$ to $`v`$ if and only if a transmission from $`u`$ at power $`p`$ can reach $`v`$ and a transmission from $`v`$ to $`u`$ at power $`p`$ can reach $`u`$. Note that, when $`p = P`$, $`G = G_{m\hspace{0pt}a\hspace{0pt}x}`$. Let $`P_{H}`$ denote the minimum value of $`p`$ at which $`G`$ is connected. The graph $`G`$ generated by each node transmitting at power $`P_{H}`$ is the initial graph, $`H`$.
Given a graph, $`T'`$, generated by the execution of the topology control algorithm, we define $`P_{T'}\hspace{0pt}{(u)}`$ as the minimum power with which node $`u`$ should make an omni-directional transmission so as to reach all of its neighbors in graph $`T'`$. Let $`C_{T'}\hspace{0pt}{(u)}`$ denote the energy cost of a transmission by $`u`$ at power $`P_{T'}\hspace{0pt}{(u)}`$.
Given the graph $`T'`$ above, we define a new graph $`T`$ as follows: a node $`u`$ is connected by an edge to $`v`$ in $`T`$ if and only if $`{C_{T'}\hspace{0pt}{(u)}} > {C\hspace{0pt}{(u,v)}}`$ or $`{C_{T'}\hspace{0pt}{(v)}} > {C\hspace{0pt}{(v,u)}}`$. Thus, given omni-directional transmissions, $`T`$ represents the graph that is actually relevant for performance comparisons and, especially, interference comparisons. $`T`$, as opposed to $`T'`$, also enforces the ability for bidirectional transmissions between communicating nodes. We call $`T`$ the cover graph generated by the topology control algorithm. All of our simulation results use the cover graph.
Define $`P_{T}\hspace{0pt}{(u)}`$ as the minimum power with which node $`u`$ should make an omni-directional transmission so as to reach all of its neighbors in the cover graph $`T`$. Let $`C_{T}\hspace{0pt}{(u)}`$ denote the energy cost of a transmission by $`u`$ at power $`P_{T}\hspace{0pt}{(u)}`$. Note that $`{P_{T}\hspace{0pt}{(u)}} \geq {P_{T'}\hspace{0pt}{(u)}}`$ and $`{C_{T}\hspace{0pt}{(u)}} \geq {C_{T'}\hspace{0pt}{(u)}}`$. In our simulations, we use $`P_{T}\hspace{0pt}{(u)}`$ as the power at which node $`u`$ makes all its transmissions after the execution of a topology control algorithm generating cover graph $`T`$.
$`{P_{T}\hspace{0pt}{(u)}}/P_{H}`$, denoted by $`P_{T/H}\hspace{0pt}{(u)}`$, is the ratio of the power at which node $`u`$ transmits after the execution of the topology control algorithm that generates cover graph $`T`$ and the power at which it transmits in the initial graph, $`H`$. With this normalization to the energy costs in the initial graph, this ratio captures the energy savings per transmission due to the topology control algorithm.
Let $`C_{T}\hspace{0pt}{({u\rightarrow{v,{H\hspace{0pt}o\hspace{0pt}p\hspace{0pt}s}}})}`$ denote the sum of $`C_{T}\hspace{0pt}{(i)}`$ for each transmitting node $`i`$ in the minimum-hop path from $`u`$ to $`v`$ in cover graph $`T`$. $`{{C_{T}\hspace{0pt}{({u\rightarrow{v,{H\hspace{0pt}o\hspace{0pt}p\hspace{0pt}s}}})}}/C_{H}}\hspace{0pt}{({u\rightarrow{v,{H\hspace{0pt}o\hspace{0pt}p\hspace{0pt}s}}})}`$, denoted by $`C_{T/H}\hspace{0pt}{({u\rightarrow{v,{H\hspace{0pt}o\hspace{0pt}p\hspace{0pt}s}}})}`$, is the ratio of the energy cost along the minimum-hop path from $`u`$ to $`v`$ in cover graph $`T`$ generated by the topology control algorithm and the corresponding cost along the minimum-hop path from $`u`$ to $`v`$ in the initial graph, $`H`$. This ratio captures the energy savings along a path due to the topology control algorithm. In our simulation experiments, we examine both the minimum-energy paths and the minimum-hop paths. The ratio for the minimum-energy paths is computed similarly as above and is denoted by $`C_{T/H}\hspace{0pt}{({u\rightarrow{v,{E\hspace{0pt}n\hspace{0pt}e\hspace{0pt}r\hspace{0pt}g\hspace{0pt}y}}})}`$.
[Refer to caption](/html/0709.0961/assets/x1.png)
\(a\)
Average $`P_{T/H}\hspace{0pt}{(u)}`$ over all $`u`$.
[Refer to caption](/html/0709.0961/assets/x2.png)
\(b\)
$`C_{T/H}\hspace{0pt}\left( {u\rightarrow{v,{E\hspace{0pt}n\hspace{0pt}e\hspace{0pt}r\hspace{0pt}g\hspace{0pt}y}}} \right)`$ averaged over all pairs of nodes $`u`$ and $`v`$.
[Refer to caption](/html/0709.0961/assets/x3.png)
\(c\)
Average node-degree.
[Refer to caption](/html/0709.0961/assets/x4.png)
\(d\)
$`C_{T/H}\hspace{0pt}\left( {u\rightarrow{v,{H\hspace{0pt}o\hspace{0pt}p\hspace{0pt}s}}} \right)`$ averaged over all pairs of nodes $`u`$ and $`v`$.
[Refer to caption](/html/0709.0961/assets/x5.png)
\(e\)
$`I_{T/H}\hspace{0pt}\left( {u\rightarrow{v,{H\hspace{0pt}o\hspace{0pt}p\hspace{0pt}s}}} \right)`$ averaged over all pairs of nodes $`u`$ and $`v`$.
[Refer to caption](/html/0709.0961/assets/x6.png)
\(f\)
$`I_{T/H}\hspace{0pt}\left( {u\rightarrow{v,{E\hspace{0pt}n\hspace{0pt}e\hspace{0pt}r\hspace{0pt}g\hspace{0pt}y}}} \right)`$ averaged over all pairs of nodes $`u`$ and $`v`$.
Figure 2: Graphs showing the effectiveness of topology control algorithms when the path loss exponent is the same across the entire region of the network (in order to allow comparisons with the CBTC algorithms). The networks are generated with 200 randomly located nodes in a unit square area. Each data point represents an average of one hundred random networks. Note that a longer distance between two nodes does not necessarily mean higher energy cost across the pair since path loss exponents follow a random Gaussian distribution.
[Refer to caption](/html/0709.0961/assets/x7.png)
\(a\)
Average $`P_{T/H}\hspace{0pt}{(u)}`$ over all $`u`$.
[Refer to caption](/html/0709.0961/assets/x8.png)
\(b\)
$`C_{T/H}\hspace{0pt}\left( {u\rightarrow{v,{E\hspace{0pt}n\hspace{0pt}e\hspace{0pt}r\hspace{0pt}g\hspace{0pt}y}}} \right)`$ averaged over all pairs of nodes $`u`$ and $`v`$.
[Refer to caption](/html/0709.0961/assets/x9.png)
\(c\)
Average node-degree.
[Refer to caption](/html/0709.0961/assets/x10.png)
\(d\)
$`C_{T/H}\hspace{0pt}\left( {u\rightarrow{v,{H\hspace{0pt}o\hspace{0pt}p\hspace{0pt}s}}} \right)`$ averaged over all pairs of nodes $`u`$ and $`v`$.
[Refer to caption](/html/0709.0961/assets/x11.png)
\(e\)
$`I_{T/H}\hspace{0pt}\left( {u\rightarrow{v,{H\hspace{0pt}o\hspace{0pt}p\hspace{0pt}s}}} \right)`$ averaged over all pairs of nodes $`u`$ and $`v`$.
[Refer to caption](/html/0709.0961/assets/x12.png)
\(f\)
$`I_{T/H}\hspace{0pt}\left( {u\rightarrow{v,{E\hspace{0pt}n\hspace{0pt}e\hspace{0pt}r\hspace{0pt}g\hspace{0pt}y}}} \right)`$ averaged over all pairs of nodes $`u`$ and $`v`$.
Figure 3: Graphs showing the effectiveness of topology control algorithms in reducing energy costs when the path loss exponents exhibit a Gaussian distribution with a mean of 3.1 in the range $`\lbrack 2.7,3.5\rbrack`$ for various values of the standard deviation (indoor propagation \[[19](#bib.bib19)\]). The networks are generated with 200 randomly located nodes in a unit square area.
[Refer to caption](/html/0709.0961/assets/x13.png)
\(a\)
Average $`P_{T/H}\hspace{0pt}{(u)}`$ over all $`u`$.
[Refer to caption](/html/0709.0961/assets/x14.png)
\(b\)
$`C_{T/H}\hspace{0pt}\left( {u\rightarrow{v,{E\hspace{0pt}n\hspace{0pt}e\hspace{0pt}r\hspace{0pt}g\hspace{0pt}y}}} \right)`$ averaged over all pairs of nodes $`u`$ and $`v`$.
Figure 4: Graphs showing the effectiveness of topology control algorithms in reducing energy costs when the path loss exponent varies around a mean of 3.1 with a standard deviation of 0.16 (indoor propagation \[[19](#bib.bib19)\]). The graphs show that the effectiveness of the STC algorithm (for that matter, all localized topology control algorithms) do not change with increase in the number of nodes.
Our measure of interference derives from the definition in \[[20](#bib.bib20)\], which is a refinement of that used in \[[21](#bib.bib21)\]. Define the span, span$`(e)`$, of an edge $`e = {(g,h)}`$ as the number of nodes that are neighbors of at least one of $`g`$ and $`h`$. This represents the number of nodes that would have to remain silent to enable a successful transmission between $`g`$ and $`h`$. We measure interference, $`I_{T}\hspace{0pt}{({u\rightarrow{v,{H\hspace{0pt}o\hspace{0pt}p\hspace{0pt}s}}})}`$, along the minimum-hop path from $`u`$ to $`v`$ in cover graph $`T`$ as the sum of span$`(e)`$ for all $`e`$ in the path. $`I_{T/H}\hspace{0pt}{({u\rightarrow{v,{H\hspace{0pt}o\hspace{0pt}p\hspace{0pt}s}}})}`$ denotes the ratio of the interference along the minimum-hop path from $`u`$ to $`v`$ in cover graph $`T`$ generated by the topology control algorithm and the corresponding cost along the minimum-hop path from $`u`$ to $`v`$ in the initial graph, $`H`$. This ratio represents the reduction in interference achieved due to the topology control algorithm. The ratio for the minimum-energy paths is computed similarly as above and is denoted by $`I_{T/H}\hspace{0pt}{({u\rightarrow{v,{E\hspace{0pt}n\hspace{0pt}e\hspace{0pt}r\hspace{0pt}g\hspace{0pt}y}}})}`$.
###
IV-B
Wireless Model
The pertinent issue in the modeling of the wireless environment in the context of this paper is the path loss model. The log-distance path loss model based on the path loss as a logarithmic function of the distance $`d`$ has been confirmed both theoretically and by measurements in a large variety of environments \[[19](#bib.bib19)\]. In this model, the path loss at distance $`d`$, $`P\hspace{0pt}L\hspace{0pt}{(d)}`$ is expressed as:
<table>
<tbody>
<tr>
<td></td>
<td>PL(d)=PL(d0)+10γlog10(d/d0)𝑃𝐿𝑑𝑃𝐿subscript𝑑010𝛾subscript10𝑑subscript𝑑0PL(d)=PL(d_{0})+10\gamma\log_{10}(d/d_{0})</td>
<td></td>
</tr>
</tbody>
</table>
where the constant $`d_{0}`$ is an arbitrary reference distance and $`\gamma`$ is called the path loss exponent.
The path loss exponent is an important parameter in modeling a wireless environment and varies within a region depending upon a number of factors including antenna characteristics, transmission frequency, the nature of the obstructions, multipath, and shadowing effects. While most of these effects are location-specific and difficult to generalize across different environments \[[22](#bib.bib22)\], empirical observations in several real environments do reveal that the distribution of path loss exponents within a region of interest is Gaussian \[[23](#bib.bib23), [24](#bib.bib24), [25](#bib.bib25), [26](#bib.bib26)\]. Our simulation models, accordingly, use a Gaussian distribution when the path loss exponent is assumed to vary across the region of interest.
In our simulation studies, the specific values of the path loss exponents and the standard deviation of their distribution are based on results from empirical studies of indoor propagation described in \[[19](#bib.bib19)\]. We choose the indoor propagation model because the constraints on topology control specifically considered in this paper, such as having to avoid a GPS receiver and not being able to discern the direction of signal arrival, are most applicable to indoor wireless environments. Our study uses three sets of simulation experiments in comparing the effectiveness of different topology control algorithms.
In the first set of experiments, we assume that the path loss exponent is the same across the entire network and study the topology control algorithms for different values of the path loss exponent between 1.5 to 3.5. These experiments are intended to reveal the dependence of topology control performance on the path loss exponent and to allow a comparison to CBTC algorithms (since they assume a uniform path loss exponent across the entire region). This first set of experiments is not intended to simulate a real wireless environment.
In the second set of experiments, we vary the path loss exponent randomly using a truncated Gaussian distribution, as suggested in \[[23](#bib.bib23)\] for simulation experiments. Based on empirical measurements of indoor propagation at 634 locations in buildings \[[19](#bib.bib19)\], the path loss exponents in our experiment follow a truncated Gaussian distribution between 2.7 and 3.5 with a mean of 3.1. We vary the standard deviation of the distribution from 0 to 0.4 to study the impact of the spatial variation of path loss exponents on the performance of topology control algorithms.
In the third set of experiments, we study the effectiveness of the algorithms as the number of nodes increases. The networks used in our first two sets of experiments consist of 200 nodes located randomly in a unit square area. In the third set of experiments, we vary the number of nodes from 100 to 500 to study the scalability of the algorithms across a five-fold increase in the number of nodes. In these experiments also, we use a truncated Gaussian distribution for the path loss exponents ranging from 2.7 to 3.5 with a mean of 3.1. However, we use a single standard deviation of 0.16 (based on scatter plots from the empirical measurements of indoor propagation on a single floor of a building, described in \[[19](#bib.bib19)\]). Each data point in this paper represents an average of one hundred different randomly generated networks.
Finally, we note that DRNG and STC algorithms can be generalized by a parameter $`k`$ so that an edge $`(u,v)`$ is removed when there exists a path of $`k`$ hops or less from $`u`$ to $`v`$ and from $`v`$ to $`u`$ provided the energy cost across each of these hops is less than that corresponding to a direct transmission between $`u`$ and $`v`$. In the fourth set of experiments, we examine the dependence of our key performance metrics on $`k`$.
Our simulation model, because it is based on empirical measurements in real wireless environments, encapsulates multipath and other phenomena such as shadowing common in real environments. The following describes each of the simulation experiments and the results in greater detail.
###
IV-C
Experiment 1
We include the CBTC algorithms in our first set of simulation experiments. The CBTC algorithms, however, assume that the loss propagation characteristics of the medium are uniform across the entire region. In fairness to the CBTC algorithms, therefore, we conduct this simulation experiment with identical path loss exponents between any pair of directly communicating nodes in the network.
For path loss exponents ranging from 1.5 to 3.5, we first provide results on the two objectives of the topology control problem statement described in [](#S1.SS1 "I-A Problem Statement ‣ I Introduction ‣ A New Distributed Topology Control Algorithm for Wireless Environments with Non-Uniform Path Loss and Multipath Propagation")
I-A
. Figure [2(a)](#S4.F2.sf1 "In Figure 2 ‣ IV-A Metrics ‣ IV Simulation Results ‣ A New Distributed Topology Control Algorithm for Wireless Environments with Non-Uniform Path Loss and Multipath Propagation") presents the ratio $`P_{T/H}\hspace{0pt}{(u)}`$ averaged over all nodes. The cover graph of the minimum spanning tree (MST) represents the lower bound of the achievable average $`P_{T/H}\hspace{0pt}{(u)}`$ and is also plotted in the graph. Figure [2(b)](#S4.F2.sf2 "In Figure 2 ‣ IV-A Metrics ‣ IV Simulation Results ‣ A New Distributed Topology Control Algorithm for Wireless Environments with Non-Uniform Path Loss and Multipath Propagation") plots the ratio $`C_{T/H}\hspace{0pt}{({u\rightarrow{v,{E\hspace{0pt}n\hspace{0pt}e\hspace{0pt}r\hspace{0pt}g\hspace{0pt}y}}})}`$ averaged over all pairs of nodes $`u`$ and $`v`$. The result obtained for MinReach represents the lower bound on the average energy costs along a path. Note that the MST graph which achieves the lower bound for the first objective does very poorly on the second objective of minimizing the energy cost along paths. This is because the MST graph has far too few edges leading to unnecessarily long multi-hop paths. In fact, the MST graph is worse than no topology control at all for low values of path loss exponents.
Figure [2(c)](#S4.F2.sf3 "In Figure 2 ‣ IV-A Metrics ‣ IV Simulation Results ‣ A New Distributed Topology Control Algorithm for Wireless Environments with Non-Uniform Path Loss and Multipath Propagation") plots the average node-degree of a node in the cover graphs generated by the topology control algorithms. SMECN is the only algorithm for which the average node degree varies substantially with a change in the path loss exponent. This is because SMECN makes a comparison between the energy cost along an edge with the sum of the energy costs along other edges. The result of such a comparison changes with the assumed value of the path loss exponent and therefore, SMECN results in different topologies under different values of the path loss exponent. All other algorithms are based on one-to-one comparison of edges as regards their energy costs, and the result of such comparisons is the same independent of the path loss exponent as long as the same path loss exponent characterizes the entire network.
Figure [2(d)](#S4.F2.sf4 "In Figure 2 ‣ IV-A Metrics ‣ IV Simulation Results ‣ A New Distributed Topology Control Algorithm for Wireless Environments with Non-Uniform Path Loss and Multipath Propagation") plots the ratio $`C_{T/H}\hspace{0pt}{({u\rightarrow{v,{H\hspace{0pt}o\hspace{0pt}p\hspace{0pt}s}}})}`$ averaged over all pairs of nodes $`u`$ and $`v`$ along minimum-hop paths. This represents the energy costs along a path when using routing algorithms that minimize the hop count. Figures [2(e)](#S4.F2.sf5 "In Figure 2 ‣ IV-A Metrics ‣ IV Simulation Results ‣ A New Distributed Topology Control Algorithm for Wireless Environments with Non-Uniform Path Loss and Multipath Propagation") and [2(f)](#S4.F2.sf6 "In Figure 2 ‣ IV-A Metrics ‣ IV Simulation Results ‣ A New Distributed Topology Control Algorithm for Wireless Environments with Non-Uniform Path Loss and Multipath Propagation") similarly present results on the average interference encountered in a path.
These results show that STC performs better—though only slightly better—than the other topology control algorithms when the path loss exponents are constant across the network. These results also show that even while the STC algorithm generates a sparser graph than other existing algorithms, it manages to keep the paths from lengthening unnecessarily and thus achieves an overall reduction in energy consumption. The same cannot be said of the MST graph, which performs very poorly as regards interference. The MST graph has so few edges in comparison to the original graph that paths between node pairs become significantly longer, causing even more interference than one would have with no topology control at all.
Based on this set of experiments, we do not consider MST or a distributed version of it as a candidate for topology control. However, it is a useful metric as a lower bound on the average transmission power of each node after the execution of a topology control algorithm. Therefore, in the simulation results that follow, we plot the results for MST only when presenting the ratio $`P_{T/H}\hspace{0pt}{(u)}`$ averaged over all nodes (the first objective of topology control in this paper).
###
IV-D
Experiment 2
In our second set of experiments, we allow a non-uniform value of the path loss exponent in the region of interest. For each pair of directly communicating nodes, we choose a random value of the path loss exponent assuming a truncated Gaussian distribution with a mean of 3.1 and ranging between 2.7 and 3.5 (based on the empirical measurements cited in Section [](#S4.SS2 "IV-B Wireless Model ‣ IV Simulation Results ‣ A New Distributed Topology Control Algorithm for Wireless Environments with Non-Uniform Path Loss and Multipath Propagation")
IV-B
). We consider standard deviations of the Gaussian distribution ranging from 0 to 0.4. Since the CBTC algorithms work under the assumption that the loss propagation characteristics of the medium are uniform across the entire region, we do not include CBTC algorithms in this set of experiments.
If the energy cost of transmission from $`u`$ to $`v`$ is not the same as that from $`v`$ to $`u`$, under DRNG or DLSS it is sometimes possible for a node $`u`$ to keep a directed edge to node $`v`$, but for node $`v`$ to drop the directed edge to $`u`$. However, since many MAC layer protocols expect bidirectional communication between directly communicating nodes \[[11](#bib.bib11)\], a topology control algorithm should ideally generate a graph in which a directed edge from $`u`$ to $`v`$ exists if and only if a directed edge from $`v`$ to $`u`$ exists. Therefore, in order to ensure a fair comparison when we simulate DRNG or DLSS in our studies, as in Section [](#S3.SS1 "III-A Complexity ‣ III Comparative Analysis ‣ A New Distributed Topology Control Algorithm for Wireless Environments with Non-Uniform Path Loss and Multipath Propagation")
III-A
, we assume that the energy cost or the path loss exponent between two directly communicating nodes is the same in both directions because the number and type of obstructions along each ray path is likely the same even in multipath environments \[[27](#bib.bib27)\].
Figure [3(a)](#S4.F3.sf1 "In Figure 3 ‣ IV-A Metrics ‣ IV Simulation Results ‣ A New Distributed Topology Control Algorithm for Wireless Environments with Non-Uniform Path Loss and Multipath Propagation") plots the ratio $`P_{T/H}\hspace{0pt}{(u)}`$ averaged over all nodes for different degrees of variation in path loss exponents. It is of interest to note that even though the STC algorithm does not improve performance by much when the path loss exponents are uniform across the network, it does make a significant difference when the path loss exponents vary.
Figure [3(b)](#S4.F3.sf2 "In Figure 3 ‣ IV-A Metrics ‣ IV Simulation Results ‣ A New Distributed Topology Control Algorithm for Wireless Environments with Non-Uniform Path Loss and Multipath Propagation") plots the ratio $`C_{T/H}\hspace{0pt}{({u\rightarrow{v,{E\hspace{0pt}n\hspace{0pt}e\hspace{0pt}r\hspace{0pt}g\hspace{0pt}y}}})}`$ averaged over all pairs of nodes $`u`$ and $`v`$. Figure [3(c)](#S4.F3.sf3 "In Figure 3 ‣ IV-A Metrics ‣ IV Simulation Results ‣ A New Distributed Topology Control Algorithm for Wireless Environments with Non-Uniform Path Loss and Multipath Propagation") plots the average node degree in the cover graphs generated by the various topology control algorithms. Figure [3(d)](#S4.F3.sf4 "In Figure 3 ‣ IV-A Metrics ‣ IV Simulation Results ‣ A New Distributed Topology Control Algorithm for Wireless Environments with Non-Uniform Path Loss and Multipath Propagation") plots the ratio $`C_{T/H}\hspace{0pt}{({u\rightarrow{v,{H\hspace{0pt}o\hspace{0pt}p\hspace{0pt}s}}})}`$ averaged over all pairs of nodes $`u`$ and $`v`$. Figures [3(e)](#S4.F3.sf5 "In Figure 3 ‣ IV-A Metrics ‣ IV Simulation Results ‣ A New Distributed Topology Control Algorithm for Wireless Environments with Non-Uniform Path Loss and Multipath Propagation") and [3(f)](#S4.F3.sf6 "In Figure 3 ‣ IV-A Metrics ‣ IV Simulation Results ‣ A New Distributed Topology Control Algorithm for Wireless Environments with Non-Uniform Path Loss and Multipath Propagation") similarly plot the average interference along the minimum-hop and the minimum-energy paths, respectively. These plots again indicate that even though the STC algorithm generates a sparser graph, it does not result in paths that are so much longer that the energy consumed or the interference along a path actually increases. In fact, we find that the STC algorithm reduces the energy cost and the interference along a path significantly in comparison to other topology control algorithms that can be employed in real urban environments. In particular, the effectiveness of the STC algorithm in cutting down the interference as well as the energy costs along a path improves with increase in the variation in path loss exponents (especially for small values of the standard deviation in path loss exponent distribution).
Since the DLSS algorithm is the closest in performance to the STC algorithm, it is worthwhile discussing the reasons behind the significant difference between their performances in the presence of variation in path loss exponents. Even though both DLSS and STC algorithms incur the same order of initial information exchange overhead, only in the STC algorithm does a node $`u`$ use the information about a node that is not directly reachable by $`u`$ but is a common neighbor of two or more neighbors of $`u`$. In DLSS, the local subgraph at node $`u`$ used for generating a localized spanning tree in DLSS is one that is induced by the neighbors of $`u`$ and does not contain a node that is not directly reachable by $`u`$. STC performs better than DLSS because, in an irregular wireless environment, it is more likely that a node, say $`n`$, is unreachable by a node $`u`$ even if it is reachable by more than one neighbor of $`u`$. Thus, DLSS ignores node $`n`$ in the creation of the local spanning tree while the STC algorithm will consider it as long as $`n`$ is reachable by some neighbor of $`u`$.
###
IV-E
Experiment 3
[Refer to caption](/html/0709.0961/assets/x15.png)
\(a\)
Initial graph, $`H`$
[Refer to caption](/html/0709.0961/assets/x16.png)
\(b\)
OPT-CBTC($`{5\hspace{0pt}\pi}/6`$)
[Refer to caption](/html/0709.0961/assets/x17.png)
\(c\)
SMECN
[Refer to caption](/html/0709.0961/assets/x18.png)
\(d\)
DRNG
[Refer to caption](/html/0709.0961/assets/x19.png)
\(e\)
DLSS
[Refer to caption](/html/0709.0961/assets/x20.png)
\(f\)
STC
[Refer to caption](/html/0709.0961/assets/x21.png)
\(g\)
MST
[Refer to caption](/html/0709.0961/assets/x22.png)
\(h\)
DLSS (cover graph)
[Refer to caption](/html/0709.0961/assets/x23.png)
\(i\)
STC (cover graph)
[Refer to caption](/html/0709.0961/assets/x24.png)
\(j\)
MST (cover graph)
Figure 5: Graphs generated by the topology control algorithms when the path loss exponents vary around a mean of 3.1 and a standard deviation of 0.16 (indoor propagation \[[19](#bib.bib19)\]). For the OPT-CBTC($`{5\hspace{0pt}\pi}/6`$) algorithm which assumes a uniform path loss exponent in the entire region, the plot shows the graph generated when the path loss exponent is 3.1 for all node pairs. Figures (a)–(g) depict the graphs generated by the respective topology control algorithms while Figures (h)–(j) depict the cover graphs. Each network consists of 200 randomly located nodes in a unit square area.
In this set of experiments, we study the effectiveness of the algorithms across a five-fold increase in the number of nodes from 100 to 500. In these experiments, as described earlier, we use a truncated Gaussian distribution with a mean of 3.1 and a standard deviation of 0.16 based on empirical measurements of indoor propagation reported in \[[19](#bib.bib19)\]. Figure [4(a)](#S4.F4.sf1 "In Figure 4 ‣ IV-A Metrics ‣ IV Simulation Results ‣ A New Distributed Topology Control Algorithm for Wireless Environments with Non-Uniform Path Loss and Multipath Propagation") shows that the STC algorithm performs better than other algorithms independent of the number of nodes as regards the first objective of the topology control algorithm specified in Section [](#S1.SS1 "I-A Problem Statement ‣ I Introduction ‣ A New Distributed Topology Control Algorithm for Wireless Environments with Non-Uniform Path Loss and Multipath Propagation")
I-A
. Figure [4(b)](#S4.F4.sf2 "In Figure 4 ‣ IV-A Metrics ‣ IV Simulation Results ‣ A New Distributed Topology Control Algorithm for Wireless Environments with Non-Uniform Path Loss and Multipath Propagation") plots the ratio $`C_{T/H}\hspace{0pt}{({u\rightarrow{v,{E\hspace{0pt}n\hspace{0pt}e\hspace{0pt}r\hspace{0pt}g\hspace{0pt}y}}})}`$ averaged over all pairs of nodes $`u`$ and $`v`$ for minimum-energy paths.
In fact, all of the topology control algorithms in our study scale similarly since they are all localized algorithms where no control information propagates beyond more than two hops. The plots for interference and average node degrees are similarly flat against the number of nodes for each of the topology control algorithms. All of the plots indicate that the STC achieves lower energy costs and interference than other algorithms independent of the number of nodes in the network.
Finally, Figures [5(a)](#S4.F5.sf1 "In Figure 5 ‣ IV-E Experiment 3 ‣ IV Simulation Results ‣ A New Distributed Topology Control Algorithm for Wireless Environments with Non-Uniform Path Loss and Multipath Propagation")–[5(j)](#S4.F5.sf10 "In Figure 5 ‣ IV-E Experiment 3 ‣ IV Simulation Results ‣ A New Distributed Topology Control Algorithm for Wireless Environments with Non-Uniform Path Loss and Multipath Propagation") present a pictorial representation of the graphs generated by the topology control algorithms. For these figures, as described earlier, we use a truncated Gaussian distribution with a mean of 3.1 and a standard deviation of 0.16 based on empirical measurements of indoor propagation reported in \[[19](#bib.bib19)\], except for the case of OPT-CBTC($`{5\hspace{0pt}\pi}/6`$) for which we used the uniform value of 3.1 between all node pairs (because CBTC algorithms assume that the path loss exponents are the same across the entire region of the network). Figures [5(a)](#S4.F5.sf1 "In Figure 5 ‣ IV-E Experiment 3 ‣ IV Simulation Results ‣ A New Distributed Topology Control Algorithm for Wireless Environments with Non-Uniform Path Loss and Multipath Propagation")–[5(g)](#S4.F5.sf7 "In Figure 5 ‣ IV-E Experiment 3 ‣ IV Simulation Results ‣ A New Distributed Topology Control Algorithm for Wireless Environments with Non-Uniform Path Loss and Multipath Propagation") are not the cover graphs since it is harder to observe the distinct edges and the sparseness achieved by the algorithms in the more dense cover graphs. Figures [5(h)](#S4.F5.sf8 "In Figure 5 ‣ IV-E Experiment 3 ‣ IV Simulation Results ‣ A New Distributed Topology Control Algorithm for Wireless Environments with Non-Uniform Path Loss and Multipath Propagation")–[5(j)](#S4.F5.sf10 "In Figure 5 ‣ IV-E Experiment 3 ‣ IV Simulation Results ‣ A New Distributed Topology Control Algorithm for Wireless Environments with Non-Uniform Path Loss and Multipath Propagation") depict the cover graphs generated by the DLSS, STC and MST, the ones that generate the most sparse cover graphs.
###
IV-F
Experiment 4
[Refer to caption](/html/0709.0961/assets/x25.png)
\(a\)
Average $`P_{T/H}\hspace{0pt}{(u)}`$ over all $`u`$.
[Refer to caption](/html/0709.0961/assets/x26.png)
\(b\)
$`C_{T/H}\hspace{0pt}\left( {u\rightarrow{v,{E\hspace{0pt}n\hspace{0pt}e\hspace{0pt}r\hspace{0pt}g\hspace{0pt}y}}} \right)`$ averaged over all pairs of nodes $`u`$ and $`v`$.
Figure 6: STC and DRNG are specific instances of a class of protocols parametrized by a positive integer $`k`$, in which an edge $`(u,v)`$ is dropped iff there exists a path of $`k`$ hops or less from $`u`$ to $`v`$ and from $`v`$ to $`u`$, and iff the transmission across each of these hops involves an energy cost smaller than that for a direct transmission between $`u`$ and $`v`$. These graphs make comparisons between the energy performance achieved for different values of $`k`$.
STC and DRNG are specific instances of a class of protocols parametrized by a positive integer $`k`$, in which an edge $`(u,v)`$ is dropped if there exists a path of $`k`$ hops or less from $`u`$ to $`v`$ and from $`v`$ to $`u`$, and if the transmission across each of these hops involves an energy cost smaller than that for a direct transmission between $`u`$ and $`v`$. A legitimate question one may wish to address is if larger values of $`k`$ will yield better performance than $`k = 2`$ (DRNG or XTC) and $`k = 3`$ (STC).
Figure [6(a)](#S4.F6.sf1 "In Figure 6 ‣ IV-F Experiment 4 ‣ IV Simulation Results ‣ A New Distributed Topology Control Algorithm for Wireless Environments with Non-Uniform Path Loss and Multipath Propagation") presents the ratio $`P_{T/H}\hspace{0pt}{(u)}`$ averaged over all nodes for values of $`k`$ from 2 to 6. Figure [6(b)](#S4.F6.sf2 "In Figure 6 ‣ IV-F Experiment 4 ‣ IV Simulation Results ‣ A New Distributed Topology Control Algorithm for Wireless Environments with Non-Uniform Path Loss and Multipath Propagation") plots the ratio $`C_{T/H}\hspace{0pt}{({u\rightarrow{v,{E\hspace{0pt}n\hspace{0pt}e\hspace{0pt}r\hspace{0pt}g\hspace{0pt}y}}})}`$ averaged over all pairs of nodes $`u`$ and $`v`$ for the same values of $`k`$. As expected, the average power at which nodes have to make their transmissions reduces as $`k`$ increases (although the reductions after $`k = 3`$ are not as significant as that from $`k = 2`$ to $`k = 3`$). The average energy cost across an energy-optimal path, however, increases with $`k = 6`$ after reaching its best values at $`k = 5`$. While the two figures indicate that the best performance may be achieved at $`k = 5`$ in terms of energy, it would come at the expense of a significant increase in computational and communication costs (which themselves may negate the energy savings with the reduced number of edges). On the other hand, as shown in Section [](#S3.SS1 "III-A Complexity ‣ III Comparative Analysis ‣ A New Distributed Topology Control Algorithm for Wireless Environments with Non-Uniform Path Loss and Multipath Propagation")
III-A
, moving from $`k = 2`$ to $`k = 3`$ does not incur an increase in the communication or computational complexity while achieving a significant increase in energy savings.
## V Concluding Remarks
Real wireless environments are characterized by a variety of irregularities and by the phenomenon of multipath propagation. Without topology control algorithms in these environments with large variations in loss characteristics, the nodes in a network may be forced to use very high power levels in their transmissions to ensure communication and network connectivity. Most topology control algorithms do not accommodate for the unique requirements of real wireless environments and often assume the ability of a node to deduce spatial information about its neighbors. The Step Topology Control (STC) algorithm presented in this paper makes no location-based assumptions and achieves a reduction in the overall energy consumption without the use of GPS devices or estimations of distance and direction.
We have presented simulation results studying the energy consumption properties of the STC algorithm in comparison to other algorithms that similarly do not employ location-based information. While the STC algorithm certainly performs better than other algorithms when the loss characteristics are uniform in the region of the network, it performs significantly better when there exists a variation in these loss characteristics. This makes the STC algorithm especially desirable in the presence of multipath propagation or when the path loss exponents are non-uniform in the region of interest. A theoretical analysis of the energy savings achieved by the STC algorithm in the presence of varying path loss exponents remains an open problem.
Most interestingly, the performance advantages of the STC algorithm come without an increase in the order of communication or computational complexity. In fact, DLSS, the topology control algorithm that comes closest to the STC algorithm in performance, has a higher computational complexity than the STC algorithm.
We show that the STC algorithm is related to the OPT-CBTC($`{5\hspace{0pt}\pi}/6`$) algorithm and retains some of the angular properties of the CBTC algorithms. When the path loss characteristics are known to a node within its neighborhood, these properties may be employed in an optimization of the discovery process to determine when the local neighborhood is fully discovered for topology control purposes.
## Appendix
RELATIONSHIP TO CONE-BASED TOPOLOGY CONTROL
In the cone-based topology control (CBTC) algorithm, a node $`u`$ determines the minimum power $`p_{u,\alpha}`$ at which it can make an omni-directional broadcast and successfully reach at least one neighbor node in each cone/sector of angle $`\alpha`$. An edge $`(u,v)`$ is removed from the network topology in an execution of CBTC($`\alpha`$) if $`u`$ cannot reach $`v`$ with power $`p_{u,\alpha}`$ and $`v`$ cannot reach $`u`$ at power $`p_{v,\alpha}`$. It is proved in \[[7](#bib.bib7)\] that when $`\alpha \leq {{5\hspace{0pt}\pi}/6}`$, the network connectivity is preserved. Additional optimizations to further reduce the energy consumption at each node and remove more edges are possible and these include:
- •
Shrink-Back Operation: If after the execution of CBTC($`\alpha`$) at a node $`u`$ not all cones of angle $`\alpha`$ contain a neighbor node (i.e., there is an $`\alpha`$-gap in the cone coverage as in the case of nodes at the boundary of the network), node $`u`$ may unnecessarily transmit at its maximum power. With this optimization, a node transmits at the power at which further increasing its transmission power does not increase the cone coverage.
- •
Pairwise Edge Removal: If there is an edge from $`u`$ to $`v_{1}`$ and from $`u`$ to $`v_{2}`$, then this operation removes the longer edge if $`{\angle\hspace{0pt}v_{1}\hspace{0pt}u\hspace{0pt}v_{2}} < {\pi/3}`$, even if there is no edge between $`v_{1}`$ and $`v_{2}`$. It is proved in \[[7](#bib.bib7)\] that this preserves the connectivity.
OPT-CBTC($`{5\hspace{0pt}\pi}/6`$) is the cone-based topology control algorithm that uses CBTC($`{5\hspace{0pt}\pi}/6`$) along with the applicable optimizations of the shrink-back operation and pairwise edge removal. The CBTC algorithms assume uniform loss characteristics across the entire region of the network and also assume that the maximum power $`P_{u}`$ of each node $`u`$ is the same. We will make the same assumptions in the following to prove the relationship between STC and OPT-CBTC($`{5\hspace{0pt}\pi}/6`$).
Let $`G_{{CBTC}\hspace{0pt}{({{5\hspace{0pt}\pi}/6})}} = {(V,E_{{CBTC}\hspace{0pt}{({{5\hspace{0pt}\pi}/6})}})}`$ denote the graph obtained by CBTC($`{5\hspace{0pt}\pi}/6`$). Similarly, let $`G_{{OPT} - {{CBTC}\hspace{0pt}{({{5\hspace{0pt}\pi}/6})}}} = {(V,E_{{OPT} - {{CBTC}\hspace{0pt}{({{5\hspace{0pt}\pi}/6})}}})}`$ denote the graph obtained by OPT-CBTC($`{5\hspace{0pt}\pi}/6`$). Let $`G = {(V,E)}`$ denote the graph obtained by the STC algorithm. Let $`d\hspace{0pt}{(u,v)}`$ denote the distance between two nodes $`u`$ and $`v`$. We first restate the lemma from \[[7](#bib.bib7)\] that helps in drawing the relationship we seek.
######
Lemma V.1
Any edge $`{(u,v)} \in E_{m\hspace{0pt}a\hspace{0pt}x}`$ either belongs to $`E_{{CBTC}\hspace{0pt}{({{5\hspace{0pt}\pi}/6})}}`$ or there exist $`u'`$, $`v' \in V`$ such that (a) $`{d\hspace{0pt}{(u',v')}} < {d\hspace{0pt}{(u,v)}}`$, (b) either $`u' = u`$ or $`{(u,u')} \in E_{{CBTC}\hspace{0pt}{({{5\hspace{0pt}\pi}/6})}}`$, and (c) either $`v' = v`$ or $`{(v,v')} \in E_{{CBTC}\hspace{0pt}{({{5\hspace{0pt}\pi}/6})}}`$.
The above lemma is proved in \[[7](#bib.bib7)\]. We now proceed to prove the relationship between STC and OPT-CBTC($`{5\hspace{0pt}\pi}/6`$) by first proving the relationship between STC and CBTC($`{5\hspace{0pt}\pi}/6`$).
######
Lemma V.2
If an edge $`{(u,v)} \notin E_{{CBTC}\hspace{0pt}{({{5\hspace{0pt}\pi}/6})}}`$, then $`{(u,v)} \notin E`$.
###### Proof:
Assume that edge $`{(u,v)} \notin E_{{CBTC}\hspace{0pt}{({{5\hspace{0pt}\pi}/6})}}`$. When an edge $`{(u,n)} \in E_{{CBTC}\hspace{0pt}{({{5\hspace{0pt}\pi}/6})}}`$ and $`{(u,v)} \notin E_{{CBTC}\hspace{0pt}{({{5\hspace{0pt}\pi}/6})}}`$, we know that $`{d\hspace{0pt}{(u,v)}} > {d\hspace{0pt}{(u,n)}}`$ since $`n`$ is discovered by $`u`$ at a certain power level but $`v`$ is not discovered by $`u`$. Therefore, using Lemma [V.1](#Sx1.Thmtheorem1 "Lemma V.1 ‣ Appendix ‣ A New Distributed Topology Control Algorithm for Wireless Environments with Non-Uniform Path Loss and Multipath Propagation"), we know that there exist nodes $`u'`$ and $`v'`$ such that $`{d\hspace{0pt}{(u',v')}} < {d\hspace{0pt}{(u,v)}}`$ and in addition, one of the following three conditions is satisfied:
Case 1: $`u' = u`$, $`v' \neq v`$, $`{d\hspace{0pt}{(v,v')}} < {d\hspace{0pt}{(u,v)}}`$. In this case, we have a two-hop path between $`u`$ and $`v`$ through $`v'`$ such that both hops are of distance less than $`d\hspace{0pt}{(u,v)}`$.
Case 2: $`u' \neq u`$, $`v' = v`$, $`{d\hspace{0pt}{(u,u')}} < {d\hspace{0pt}{(u,v)}}`$. In this case also, we have a two-hop path between $`u`$ and $`v`$ through $`u'`$ such that both hops are of distance less than $`d\hspace{0pt}{(u,v)}`$.
Case 3: $`u' \neq u`$, $`v' \neq v`$, $`{d\hspace{0pt}{(u,u')}} < {d\hspace{0pt}{(u,v)}}`$, and $`{d\hspace{0pt}{(v,v')}} < {d\hspace{0pt}{(u,v)}}`$. We now have a three-hop path between $`u`$ and $`v`$ through $`u'`$ and $`v'`$ such that each of the three hops is of distance less than $`d\hspace{0pt}{(u,v)}`$.
Since there exist nodes such that either a two-hop or a three-hop path exists between $`u`$ and $`v`$ with each hop corresponding to a distance smaller than $`d\hspace{0pt}{(u,v)}`$, $`{(u,v)} \notin E`$.
∎
######
Theorem V.3
If $`{(u,v)} \notin E_{{OPT} - {{CBTC}\hspace{0pt}{({{5\hspace{0pt}\pi}/6})}}}`$, then $`{(u,v)} \notin E`$.
###### Proof:
Since we know from Lemma [V.2](#Sx1.Thmtheorem2 "Lemma V.2 ‣ Appendix ‣ A New Distributed Topology Control Algorithm for Wireless Environments with Non-Uniform Path Loss and Multipath Propagation") that STC removes all the edges that are removed by CBTC($`{5\hspace{0pt}\pi}/6`$), we only have to prove that STC also removes the edges removed by the optimizations of the shrink-back operation and pairwise edge removal.
When an edge $`(u,v)`$ is removed as part of the Shrink-Back operation, the cone coverage around $`u`$ does not change. Therefore, there exist nodes $`u'`$ and $`v'`$ such that $`{d\hspace{0pt}{(u',v')}} < {d\hspace{0pt}{(u,v)}}`$ and in addition, one of the three cases listed in the proof of Lemma [V.2](#Sx1.Thmtheorem2 "Lemma V.2 ‣ Appendix ‣ A New Distributed Topology Control Algorithm for Wireless Environments with Non-Uniform Path Loss and Multipath Propagation") applies. Exactly as in the proof of Lemma [V.2](#Sx1.Thmtheorem2 "Lemma V.2 ‣ Appendix ‣ A New Distributed Topology Control Algorithm for Wireless Environments with Non-Uniform Path Loss and Multipath Propagation"), this implies that $`{(u,v)} \notin E`$.
When the Pairwise Edge Removal operation removes an edge $`(u,v)`$, it implies there is another neighbor node $`(u,n)`$ such that $`{d\hspace{0pt}{(u,n)}} < {d\hspace{0pt}{(u,v)}}`$ and $`{\angle\hspace{0pt}n\hspace{0pt}u\hspace{0pt}v} < {\pi/3}`$. Since $`{\angle\hspace{0pt}n\hspace{0pt}u\hspace{0pt}v} < {\pi/3}`$, edge $`(n,v)`$ is not the longest edge in the triangle $`n\hspace{0pt}u\hspace{0pt}v`$. Since $`d\hspace{0pt}{(u,n)}`$ is also smaller than $`d\hspace{0pt}{(u,v)}`$, we have a two-hop path between $`u`$ and $`v`$ through $`n`$ where the distance across each hop is less than $`d\hspace{0pt}{(u,v)}`$. Therefore, edge $`(u,v)`$ is removed by STC as well. ∎
## Acknowledgment
This work was supported in part by NSF Awards CNS-0322797 and CNS-0626548.
## References
- \[1\] P. Santi, *Topology Control in Wireless Ad Hoc and Sensor Networks*. John Wiley and Sons, 2006.
- \[2\] N. Li, J. C. Hou, and L. Sha, “Design and analysis of an MST-based topology control algorithm,” *IEEE Transactions on Wireless Communications*, vol. 4, no. 3, pp. 1195–1206, May 2005.
- \[3\] V. Rodoplu and T. H. Meng, “Minimum energy mobile wireless networks,” *IEEE Journal on Selected Areas in Communications*, vol. 17, no. 8, pp. 1333–1344, August 1999.
- \[4\] L. Li and J. Y. Halpern, “Minimum energy mobile wireless networks revisited,” in *Proceedings of the IEEE International Conference on Communications (ICC)*. IEEE, 2001, pp. 278–283.
- \[5\] L. Jia, R. Rajaraman, and C. Scheideler, “On local algorithms for topology control and routing in ad hoc networks,” in *Proceedings of the ACM symposium on Parallel Algorithms and Architectures (SPAA)*. ACM Press, 2003, pp. 220–229.
- \[6\] X.-Y. Li, W.-Z. Song, and Y. Wang, “Efficient topology control for ad-hoc wireless networks with non-uniform transmission ranges,” *Wireless Networks*, vol. 11, no. 3, pp. 255–264, 2005.
- \[7\] L. E. Li, J. Y. Halpern, P. Bahl, Y.-M. Wang, and R. Wattenhofer, “A cone-based distributed topology control algorithm for wireless multi-hop networks,” *IEEE/ACM Transactions on Networking*, vol. 13, no. 1, pp. 147–159, February 2005.
- \[8\] D. Puccinelli and M. Haenggi, “Multipath fading in wireless sensor networks: measurements and interpretation,” in *Proceedings of the 2006 International Conference on Communications and Mobile Computing*, 2006, pp. 1039–1044.
- \[9\] G. Zhou, T. He, S. Krishnamurthy, and J. A. Stankovic, “Models and solutions for radio irregularity in wireless sensor networks,” *ACM Transactions on Sensor Networks*, vol. 2, no. 2, pp. 221–262, 2006.
- \[10\] N. Li and J. C. Hou, “Localized topology control algorithms for heterogeneous wireless networks,” *IEEE/ACM Transactions on Networking*, vol. 13, no. 6, pp. 1313–1324, December 2005.
- \[11\] LAN MAN Standards Committee of the IEEE Computer Society, “Wireless LAN medium access control (MAC) and physical layer (PHY) specifications,” 2003. \[Online\]. Available: http://standards.ieee.org
- \[12\] D. M. Blough, M. Leoncini, G. Resta, and P. Santi, “The $`k`$-Neigh protocol for symmetric topology control in ad hoc networks,” in *Proceedings of the International Symposium on Mobile Ad Hoc Networking and Computing (ACM MobiHoc)*. ACM Press, 2003, pp. 141–152.
- \[13\] R. Wattenhofer and A. Zollinger, “XTC: A practical topology control algorithm for ad hoc networks,” in *Proceedings of the International Parallel and Distributed Processing Symposium (IPDPS)*. IEEE, 2004, pp. 216–223.
- \[14\] M. Damian, S. Pandit, and S. Pemmaraju, “Local approximation schemes for topology control,” in *Proceedings of the Annual ACM Symposium on Principles of Distributed Computing (PODC)*. ACM Press, 2006, pp. 208–217.
- \[15\] F. Kuhn and A. Zollinger, “Ad-hoc networks beyond unit disk graphs,” in *Proceedings of the Joint Workshop on Foundations of Mobile Computing (DIALM-POMC)*. ACM Press, 2003, pp. 69–78.
- \[16\] D. M. Blough, M. Leoncini, G. Resta, and P. Santi, “Topology control with better radio models: Implications for energy and multi-hop interference,” in *Proceedings of the 8th ACM International Symposium on Modeling, Analysis and Simulation of Wireless and Mobile Systems (MSWiM)*. ACM Press, 2005, pp. 260–268.
- \[17\] J. B. Kruskal, “On the shortest spanning subtree of a graph and the traveling salesman problem,” *Proceedings of the American Mathematical Society*, vol. 7, no. 1, pp. 48–50, February 1956.
- \[18\] W.-Z. Song, Y. Wang, X.-Y. Li, and O. Frieder, “Localized algorithms for energy efficient topology in wireless ad hoc networks,” in *MobiHoc ’04: Proceedings of the 5th ACM international symposium on Mobile ad hoc networking and computing*. New York, NY, USA: ACM, 2004, pp. 98–108.
- \[19\] T. S. Rappaport, *Wireless Communications: Principles and Practice*, 2nd ed. Prentice Hall, 2002.
- \[20\] T. Johansson and L. Carr-Motyčková, “Reducing interference in ad hoc networks through topology control,” in *Proceedings of the Joint Workshop on Foundations of Mobile Computing (DIALM-POMC)*. ACM Press, 2005, pp. 17–23.
- \[21\] M. Burkhart, P. von Rickenbach, R. Wattenhofer, and A. Zollinger, “Does topology control reduce interference?” in *Proceedings of the ACM International Symposium on Mobile Ad Hoc Networking and Computing (MobiHoc)*. ACM Press, 2004, pp. 9–19.
- \[22\] S. Haykin and M. Moher, *Modern Wireless Communications*. Prentice Hall, 2005.
- \[23\] V. Erceg, L. J. Greenstein, S. Y. Tjandra, S. R. Parkoff, A. Gupta, B. Kulic, A. A. Julius, and R. Bianchi, “An empirically based path loss model for wireless channels in suburban environments,” *IEEE Journal on Selected Areas in Communications*, vol. 17, no. 7, pp. 1205–1211, 1999.
- \[24\] J. A. Dabin, A. M. Haimovich, and H. Grebel, “A statistical ultra-wideband indoor channel model and the effects of antenna directivity on path loss and multipath propagation,” *IEEE Journal on Selected Areas in Communications*, vol. 24, no. 4, pp. 752–758, 2006.
- \[25\] L. C. Liechty, E. Reifsnider, and G. Durgin, “Developing the best 2.4 GHz propagation model from active network measurements,” in *Proceedings of the Vehicular Technology Conference*. IEEE, 2007, pp. 894–896.
- \[26\] S. Geng and P. Vainikainen, “Experimental investigation of the properties of multiband UWB propagation channels,” in *Proceedings of the Annual IEEE Symposium on Personal, Indoor and Mobile Radio Communications*. IEEE, 2007, pp. 1–5.
- \[27\] Z. Irahhauten, G. Bellusci, G. J. M. Janssen, H. Nikookar, and C. Tiberius, “Investigation of UWB ranging in dense indoor multipath environments,” in *Procedings of the IEEE Singapore International Conference on Communication systems*. IEEE, 2006, pp. 1–5. |
0710.2716 | cs.NI | cyber | Cost and Effects of Pinning Control for Network Synchronization
<sup>\*</sup>
<sup>\*</sup>\* This work is supported by the National Science Foundation of China under grants 60674093, 60334030.
Rong Li $`^{\text{a}}`$
<sup>†</sup>
<sup>†</sup>† Corresponding author: lirong@pku.edu.cn
, Zhisheng Duan $`^{\text{a}}`$, Guanrong Chen $`^{\text{b}}`$
$`^{\text{a}}`$ Department of Mechanics and Aerospace Technology, Peking University, Beijing, P. R. China
$`^{\text{b}}`$ Department of Electronic Engineering, City University of Hong Kong, Hong Kong
Abstract. In this paper, the problem of pinning control for synchronization of complex dynamical networks is discussed. A cost function of the controlled network is defined by the feedback gain and the coupling strength of the network. An interesting result is that lower cost is achieved by the control scheme of pinning nodes with smaller degrees. Some rigorous mathematical analysis is presented for achieving lower cost in the synchronization of different star-shaped networks. Numerical simulations on some non-regular complex networks generated by the Barabási-Albert model and various star-shaped networks are shown for verification and illustration.
Keywords: complex dynamical network, pinning control, exponential stability.
## 1 Introduction and problem formulation
Complex networks are currently being studied across many fields of sciences, including physics, chemistry, biology, mathematics, sociology and engineering \[[1](#bib.bib1), [2](#bib.bib2), [3](#bib.bib3), [5](#bib.bib5), [9](#bib.bib9), [15](#bib.bib15), [17](#bib.bib17), [19](#bib.bib19)\]. A complex network is a large set of interconnected nodes, in which a node is a fundamental unit with specific contents. Examples of complex networks include the Internet, food webs, cellular neural networks, biological neural networks, electrical power grids, telephone cell graphs, etc. Recently, synchronization of complex networks of dynamical systems has received a great deal of attention from the nonlinear dynamics community \[[10](#bib.bib10), [12](#bib.bib12), [16](#bib.bib16), [18](#bib.bib18), [20](#bib.bib20)\]. A special control strategy called pinning control is used to achieve synchronization of complex networks; that is, only a fraction of the nodes or even a single node is controlled over the whole network \[[4](#bib.bib4), [6](#bib.bib6), [11](#bib.bib11), [21](#bib.bib21)\]. This control method has become a common technique for control, stabilization and synchronization of coupled dynamical systems. In general, different nodes have different degrees in a network, thus a natural question is how different the effect would be when nodes with different degrees are pinned.
Consider a dynamical network consisting of N identical and diffusively coupled nodes, with each node being an $`n`$-dimensional dynamical system. The state equations of the network are
<table>
<tbody>
<tr>
<td></td>
<td>x˙i(t)=f(xi(t),t)+c∑j=1NaijΓxj(t),i=1,2,⋯,N,formulae-sequencesubscript˙𝑥𝑖𝑡𝑓subscript𝑥𝑖𝑡𝑡𝑐subscriptsuperscript𝑁𝑗1subscript𝑎𝑖𝑗Γsubscript𝑥𝑗𝑡𝑖12⋯𝑁\dot{x}_{i}(t)=f(x_{i}(t),t)+c\sum^{N}_{j=1}a_{ij}\Gamma x_{j}(t),\qquad i=1,2,\cdots,N,</td>
<td></td>
<td>(1)</td>
</tr>
</tbody>
</table>
where $`f\hspace{0pt}{( \cdot )}`$ is the dynamical function of an isolated node, $`x_{i} = {(x_{i\hspace{0pt}1},x_{i\hspace{0pt}2},\cdots,x_{i\hspace{0pt}n})} \in R^{n}`$ are the state variables of node $`i`$, constant $`c > 0`$ represents the coupling strength, and $`\Gamma \in R^{N \times N}`$ is the inner linking matrix. Moreover, the coupling matrix $`A = {(a_{i\hspace{0pt}j})} \in R^{N \times N}`$ represents the coupling configuration of the network: If there is a connection between node $`i`$ and node $`j`$ $`({i \neq j})`$, then $`a_{i\hspace{0pt}j} = a_{j\hspace{0pt}i} = 1`$; otherwise, $`a_{i\hspace{0pt}j} = a_{j\hspace{0pt}i} = 0`$ $`({i \neq j})`$; the diagonal entries of $`A`$ are defined by
<table>
<tbody>
<tr>
<td></td>
<td>aii=−∑j=1j≠iNaij,i=1,2,⋯,N.formulae-sequencesubscript𝑎𝑖𝑖subscriptsuperscript𝑁FRACOP𝑗1𝑗𝑖subscript𝑎𝑖𝑗𝑖12⋯𝑁a_{ii}=-\sum^{N}_{j=1\atop j\neq i}a_{ij},\qquad i=1,2,\cdots,N.</td>
<td></td>
<td>(2)</td>
</tr>
</tbody>
</table>
Suppose that the network is connected in the sense of having no isolated clusters. Then, the coupling matrix $`A`$ is irreducible. From Lemma 2 of \[[20](#bib.bib20)\], it can be proved that zero is an eigenvalue of $`A`$ with multiplicity one and all the other eigenvalues of $`A`$ are strictly negative.
Network ([1](#S1.E1 "In 1 Introduction and problem formulation")) is said to achieve (asymptotical) synchronization if
<table>
<tbody>
<tr>
<td></td>
<td>x1(t)→x2(t)→⋯→xN(t)→s(t),ast→∞,formulae-sequence→subscript𝑥1𝑡subscript𝑥2𝑡→⋯→subscript𝑥𝑁𝑡→𝑠𝑡→as𝑡x_{1}(t)\rightarrow x_{2}(t)\rightarrow\cdots\rightarrow x_{N}(t)\rightarrow s(t),\qquad\textrm{as}\quad t\rightarrow\infty,</td>
<td></td>
<td>(3)</td>
</tr>
</tbody>
</table>
where, because of the diffusive coupling configuration, $`s\hspace{0pt}{(t)}`$ is a solution of an isolated node, which can be an equilibrium, a periodic or a chaotic orbit. As shown in \[[4](#bib.bib4), [6](#bib.bib6), [11](#bib.bib11), [21](#bib.bib21)\], this can be achieved by controlling several nodes (or even only one node) of the network. Without loss of generality, suppose that the controllers are added on the last $`N - k`$ nodes of the network, so that the equations of the controlled network can be written as
<table>
<tbody>
<tr>
<td></td>
<td>x˙i(t)=f(xi(t),t)+c∑j=1NaijΓxj(t),i=1,2,⋯,k,x˙i(t)=f(xi(t),t)+c∑j=1NaijΓxj(t)−cεiΓ(xi(t)−s(t)),i=k+1,k+2,⋯,N,formulae-sequencesubscript˙𝑥𝑖𝑡𝑓subscript𝑥𝑖𝑡𝑡𝑐superscriptsubscript𝑗1𝑁subscript𝑎𝑖𝑗Γsubscript𝑥𝑗𝑡𝑖12⋯𝑘missing-subexpressionformulae-sequencesubscript˙𝑥𝑖𝑡𝑓subscript𝑥𝑖𝑡𝑡𝑐superscriptsubscript𝑗1𝑁subscript𝑎𝑖𝑗Γsubscript𝑥𝑗𝑡𝑐subscript𝜀𝑖Γsubscript𝑥𝑖𝑡𝑠𝑡𝑖𝑘1𝑘2⋯𝑁\begin{array}[]{l}\dot{x}_{i}(t)=f(x_{i}(t),t)+c\sum_{j=1}^{N}a_{ij}\Gamma x_{j}(t),\qquad\qquad\qquad\qquad\qquad\;i=1,2,\cdots,k,\\ \\ \dot{x}_{i}(t)=f(x_{i}(t),t)+c\sum_{j=1}^{N}a_{ij}\Gamma x_{j}(t)-c\varepsilon_{i}\Gamma(x_{i}(t)-s(t)),\qquad i=k+1,k+2,\cdots,N,\end{array}</td>
<td></td>
<td>(4)</td>
</tr>
</tbody>
</table>
where the feedback gains $`\varepsilon_{i}`$ are positive constants. It can be seen that synchronizing all states $`x_{i}\hspace{0pt}{(t)}`$ to $`s\hspace{0pt}{(t)}`$ is determined by the dynamics of an isolated node, the coupling strength $`c > 0`$, the inner linking matrix $`\Gamma`$, the feedback gains $`\varepsilon_{i} \geq 0`$, and the coupling matrix $`A`$.
As discussed in \[[6](#bib.bib6), [11](#bib.bib11), [21](#bib.bib21)\], to achieve synchronization of complex dynamical networks, the controllers are generally preferred to be added to the nodes with larger degrees. However, it is also known that, to achieve a certain synchronizability of the network, the feedback gains $`\varepsilon_{i}`$ usually have to be quite large. In \[[4](#bib.bib4)\], when a single controller is used, the coupling strength $`c`$ has to be quite large in general. From the view point of realistic applications, these are not expected and sometimes cannot be realized. Practically, a designed control strategy is expected to be effective and also easily implementable. In this paper, for various star-shaped networks and non-regular complex networks, a new concept of cost function is introduced to evaluate the efficiency of the designed controllers. It is found that surprisingly the cost can be much lower by controlling nodes with smaller degrees than controlling nodes with larger degrees. As will be seen, moreover, both the feedback gains $`\varepsilon_{i}`$ and the coupling strength $`c`$ can be much smaller than those used in \[[4](#bib.bib4), [6](#bib.bib6), [11](#bib.bib11), [21](#bib.bib21)\].
The outline of this paper is as follows. In Section 2, a new definition of cost function and some mathematical preliminaries are given. Stability of different star-shaped networks controlled by pinning some nodes with small degrees are analyzed in Sections 3 and 4, respectively, where some simulated examples of dynamical networks are compared for illustration and verification. In Section 5, pinning control of non-regular complex dynamical networks of chaotic oscillators is studied through numerical simulations. Finally, Section 6 concludes the paper.
## 2 The cost of pinning control
Denote $`{e_{i}\hspace{0pt}{(t)}} = {{x_{i}\hspace{0pt}{(t)}} - {s\hspace{0pt}{(t)}}}`$, where $`s\hspace{0pt}{(t)}`$ satisfies $`{\overset{˙}{s}\hspace{0pt}{(t)}} = {f\hspace{0pt}{({s\hspace{0pt}{(t)}})}}`$. Then, the error equations of network ([1](#S1.E1 "In 1 Introduction and problem formulation")) can be written as
<table>
<tbody>
<tr>
<td></td>
<td>e˙i(t)=f(xi(t),t)−f(s(t))+c∑j=1NaijΓej(t),i=1,2,⋯,N,formulae-sequencesubscript˙𝑒𝑖𝑡𝑓subscript𝑥𝑖𝑡𝑡𝑓𝑠𝑡𝑐subscriptsuperscript𝑁𝑗1subscript𝑎𝑖𝑗Γsubscript𝑒𝑗𝑡𝑖12⋯𝑁\dot{e}_{i}(t)=f(x_{i}(t),t)-f(s(t))+c\sum^{N}_{j=1}a_{ij}\Gamma e_{j}(t),\qquad\qquad\qquad\;\;i=1,2,\cdots,N,</td>
<td></td>
</tr>
</tbody>
</table>
while the error equations of the controlled network ([4](#S1.E4 "In 1 Introduction and problem formulation")) can be written as
<table>
<tbody>
<tr>
<td></td>
<td>e˙i(t)=f(xi(t),t)−f(s(t))+c∑j=1Na~ijΓej(t),i=1,2,⋯,N,formulae-sequencesubscript˙𝑒𝑖𝑡𝑓subscript𝑥𝑖𝑡𝑡𝑓𝑠𝑡𝑐subscriptsuperscript𝑁𝑗1subscript~𝑎𝑖𝑗Γsubscript𝑒𝑗𝑡𝑖12⋯𝑁\dot{e}_{i}(t)=f(x_{i}(t),t)-f(s(t))+c\sum^{N}_{j=1}\tilde{a}_{ij}\Gamma e_{j}(t),\qquad\qquad\qquad\;\;i=1,2,\cdots,N,</td>
<td></td>
<td>(5)</td>
</tr>
</tbody>
</table>
where $`{{\overset{\sim}{a}}_{i\hspace{0pt}i} = {a_{i\hspace{0pt}i} - \varepsilon_{i}}},{{\varepsilon_{i} > 0},{i = {{k + 1},{k + 2},\cdots,N}}}`$, and $`{\overset{\sim}{a}}_{i\hspace{0pt}j} = a_{i\hspace{0pt}j}`$ otherwise. Let $`\overset{\sim}{A} = {({\overset{\sim}{a}}_{i\hspace{0pt}j})} \in R^{N \times N}`$, and denote $`{e\hspace{0pt}{(t)}} = {({e_{1}\hspace{0pt}{(t)}},{e_{2}\hspace{0pt}{(t)}},\cdots,{e_{N}\hspace{0pt}{(t)}})}^{T}`$.
Differentiating ([5](#S2.E5 "In 2 The cost of pinning control")) along $`s\hspace{0pt}{(t)}`$ gives
<table>
<tbody>
<tr>
<td></td>
<td>e˙(t)=D(f(s(t)))e(t)+cΓe(t)A~T.˙𝑒𝑡𝐷𝑓𝑠𝑡𝑒𝑡𝑐Γ𝑒𝑡superscript~𝐴𝑇\dot{e}(t)=D(f(s(t)))e(t)+c\Gamma e(t)\tilde{A}^{T}.</td>
<td></td>
<td>(6)</td>
</tr>
</tbody>
</table>
By analyzing the matrix $`{\overset{\sim}{A}}^{T}`$, it is easy to see that all the eigenvalues of $`{\overset{\sim}{A}}^{T}`$ are negative, which are denoted by
<table>
<tbody>
<tr>
<td></td>
<td>0>λ1≥⋯≥λN.0subscript𝜆1⋯subscript𝜆𝑁0>\lambda_{1}\geq\cdots\geq\lambda_{N}.</td>
<td></td>
</tr>
</tbody>
</table>
There exists an orthogonal matrix $`U`$ such that $`{\overset{\sim}{A}}^{T} = {U\hspace{0pt}J\hspace{0pt}U^{- 1}}`$, where $`J = {d\hspace{0pt}i\hspace{0pt}a\hspace{0pt}g\hspace{0pt}{\{\lambda_{1},\lambda_{2},\cdots,\lambda_{N}\}}}`$. Let $`{\overset{\sim}{e}\hspace{0pt}{(t)}} = {e\hspace{0pt}{(t)}\hspace{0pt}U}`$. Then, from ([6](#S2.E6 "In 2 The cost of pinning control")), one has
<table>
<tbody>
<tr>
<td></td>
<td>e~˙i(t)=[Df(s(t))+cλiΓ]e~i(t),i=1,⋯,N.formulae-sequencesubscript˙~𝑒𝑖𝑡delimited-[]𝐷𝑓𝑠𝑡𝑐subscript𝜆𝑖Γsubscript~𝑒𝑖𝑡𝑖1⋯𝑁\dot{\tilde{e}}_{i}(t)=[Df(s(t))+c\lambda_{i}\Gamma]\tilde{e}_{i}(t),\qquad\qquad\qquad\;\;i=1,\cdots,N.</td>
<td></td>
<td>(7)</td>
</tr>
</tbody>
</table>
Therefore, the local stability problem of network ([4](#S1.E4 "In 1 Introduction and problem formulation")) is converted into the stability problem of the $`N`$ independent linear systems ([7](#S2.E7 "In 2 The cost of pinning control")). When the system function of an isolated node and the inner linking matrix $`\Gamma`$ are fixed, the stability problem of systems ([7](#S2.E7 "In 2 The cost of pinning control")) are dependent on the coupling strength $`c`$ and the eigenvalues of $`\overset{\sim}{A}`$. Clearly, the smaller the eigenvalues of the matrix $`\overset{\sim}{A}`$ are ($`{\lambda_{i} < 0},{i = {1,\cdots,N}}`$), the smaller the coupling strength $`c > 0`$ is needed to guarantee the same synchronizability of the network ([4](#S1.E4 "In 1 Introduction and problem formulation")), if systems ([7](#S2.E7 "In 2 The cost of pinning control")) have unbounded synchronization regions \[[18](#bib.bib18)\]. In the following, a cost function is introduced to describe the efficiency of the controllers.
Definition 1 (Cost Function) Suppose that the feedback gain matrix is $`G = diag{\{\varepsilon_{1},\cdots,}`$ $`\varepsilon_{N}\}`$, where $`{\varepsilon_{i} \geq 0},{i = {1,\cdots,N}}`$, are given as in ([4](#S1.E4 "In 1 Introduction and problem formulation")). The Cost Function is defined as
<table>
<tbody>
<tr>
<td></td>
<td>CF=c∑i=1Nεi.𝐶𝐹𝑐superscriptsubscript𝑖1𝑁subscript𝜀𝑖CF=c\sum_{i=1}^{N}\varepsilon_{i}.</td>
<td></td>
</tr>
</tbody>
</table>
Remark 1 The smaller the $`C\hspace{0pt}F`$, the more efficient a control strategy to achieve the same goal of control, and the easier to be implemented.
In order to discuss the effects of pinning control, the following Lemmas are needed.
Lemma 2 \[[8](#bib.bib8)\] Let $`A = {\lbrack a_{i\hspace{0pt}j}\rbrack} \in C^{n \times n}`$ be Hermitian, and let $`a_{n\hspace{0pt}n} \leq \cdots \leq a_{22} \leq a_{11}`$ be a rearrangement of its diagonal entries in increasing order. Let the eigenvalues of $`A`$ be ordered as
<table>
<tbody>
<tr>
<td></td>
<td>λmin=λn≤λn−1≤⋯≤λ2≤λ1=λmax.subscript𝜆subscript𝜆𝑛subscript𝜆𝑛1⋯subscript𝜆2subscript𝜆1subscript𝜆\lambda_{\min}=\lambda_{n}\leq\lambda_{n-1}\leq\cdots\leq\lambda_{2}\leq\lambda_{1}=\lambda_{\max}.</td>
<td></td>
<td>(8)</td>
</tr>
</tbody>
</table>
Then
<table>
<tbody>
<tr>
<td></td>
<td>(i)λn≤aii≤λ1for alli=1,⋯,n,(ii)a11+a22≤λ2,ifλ1=0.formulae-sequence(i)subscript𝜆𝑛subscript𝑎𝑖𝑖subscript𝜆1for all𝑖1⋯𝑛formulae-sequence(ii)subscript𝑎11subscript𝑎22subscript𝜆2ifsubscript𝜆10\begin{array}[]{l}\textrm{(i)}\quad\,\lambda_{n}\leq a_{ii}\leq\lambda_{1}\quad\textrm{for all}\quad i=1,\cdots,n,\\ \textrm{(ii)}\quad a_{11}+a_{22}\leq\lambda_{2},\quad\textrm{if}\quad\lambda_{1}=0.\end{array}</td>
<td></td>
</tr>
</tbody>
</table>
Remark 2 For diffusive networks, the smaller the $`\lambda_{2}`$, the easier the synchronization, if the network has an unbound synchronized region \[[18](#bib.bib18)\]. However, Lemma 2 shows that $`\lambda_{2}`$ is related to $`a_{11} + a_{22}`$, where $`a_{11}`$ and $`a_{22}`$ are determined by two smallest nodes. Therefore, in this case, in order to improve the synchronizability, these small nodes should be pinned.
Lemma 3 Let $`A \in C^{n \times n}`$ be Hermitian, and
<table>
<tbody>
<tr>
<td></td>
<td>A~=[A1A12A21TA2+D],~𝐴delimited-[]subscript𝐴1subscript𝐴12superscriptsubscript𝐴21𝑇subscript𝐴2𝐷\tilde{A}=\left[\begin{array}[]{cc}A_{1}&A_{12}\\ A_{21}^{T}&A_{2}+D\end{array}\right],</td>
<td></td>
</tr>
</tbody>
</table>
where $`A_{1} \in R^{k \times k}`$, $`A_{12} \in R^{k \times {({N - k})}}`$, $`A_{2} \in R^{{({N - k})} \times {({N - k})}}`$ and $`D = {d\hspace{0pt}i\hspace{0pt}a\hspace{0pt}g\hspace{0pt}{\{\varepsilon_{k + 1},\cdots,\varepsilon_{N}\}}}`$. Then, $`{\overset{\sim}{A} < {- {\alpha\hspace{0pt}I}}},{\alpha > 0}`$, if and only if $`A_{1} < {- {\alpha\hspace{0pt}I_{1}}}`$ and $`{{A_{2} + D} - {A_{12}^{T}\hspace{0pt}{({A_{1} + {\alpha\hspace{0pt}I_{1}}})}^{- 1}\hspace{0pt}A_{12}}} < {- {\alpha\hspace{0pt}I_{2}}}`$, where $`I_{1}`$ and $`I_{2}`$ are identity matrixes with appropriate dimensions.
Remark 3 Matrix $`D`$, i.e. $`\varepsilon_{i}`$ can be determined by the LMI method \[[7](#bib.bib7)\]. Suppose the eigenvalues of $`A_{1}`$ are $`\beta_{1} \geq \beta_{2} \geq \cdots \geq \beta_{k}`$. Then $`{- \alpha} > \beta_{1}`$; and if $`\varepsilon_{i}\rightarrow{- \infty}`$, then $`{{- \alpha}\rightarrow\beta_{1}}.`$
## 3 Simple star-shaped networks
Now, consider a network consisting of $`N`$ identical nodes with a simple star-shaped coupling configuration: there exists a central node which connects to all the other non-central nodes, and there are no direct connections among the non-central nodes. Fig. 1 gives an example of such a network.
Fig. 1 $`N = 9`$
The coupling matrix $`A`$ of network ([1](#S1.E1 "In 1 Introduction and problem formulation")) with a simple star-shaped coupling configuration can been written as
<table>
<tbody>
<tr>
<td></td>
<td>A=[A1A12A21A2],𝐴delimited-[]subscript𝐴1subscript𝐴12subscript𝐴21subscript𝐴2A=\left[\begin{array}[]{cc}A_{1}&A_{12}\\ A_{21}&A_{2}\end{array}\right],</td>
<td></td>
<td>(9)</td>
</tr>
</tbody>
</table>
where $`A_{1} = {{- N} + 1}`$, $`A_{12} = {(1,1,\cdots,1)} \in R^{({N - 1})}`$, $`A_{21} = A_{12}^{T}`$, and $`{A_{2} = {- I_{{({N - 1})} \times {({N - 1})}}}}.`$
Let the feedback gain matrix be $`{G = {d\hspace{0pt}i\hspace{0pt}a\hspace{0pt}g\hspace{0pt}{\{ 0,\varepsilon,\cdots,\varepsilon\}}} \in R^{N \times N}};`$ that is, the controllers are added to all the non-central nodes for simplicity in this discussion. Then, the coupling matrix of the controlled network ([4](#S1.E4 "In 1 Introduction and problem formulation")) is
<table>
<tbody>
<tr>
<td></td>
<td>A~=[A1A12A21A2−εI].~𝐴delimited-[]subscript𝐴1subscript𝐴12subscript𝐴21subscript𝐴2𝜀𝐼\tilde{A}=\left[\begin{array}[]{cc}A_{1}&A_{12}\\ A_{21}&A_{2}-\varepsilon I\end{array}\right].</td>
<td></td>
</tr>
</tbody>
</table>
Theorem 1 If there are constants $`\varepsilon > 0`$ and $`\overset{\sim}{k} > 0`$, such that $`{N - \overset{\sim}{k} - 1} > 0`$, $`\varepsilon > {\overset{\sim}{k} - 1}`$ and $`\varepsilon > \frac{\overset{\sim}{k}\hspace{0pt}{({N - \overset{\sim}{k}})}}{N - \overset{\sim}{k} - 1}`$, then $`{{\lambda_{i}\hspace{0pt}{(\overset{\sim}{A})}} < {- \overset{\sim}{k}}},{i = {1,\cdots,N}}`$, where $`{{{{\lambda_{i}\hspace{0pt}{(\overset{\sim}{A})}},i} = 1},{\cdots,N}},`$ are the eigenvalues of the matrix $`\overset{\sim}{A}`$.
Proof: Consider the matrix
<table>
<tbody>
<tr>
<td></td>
<td>A~+k~I=[A1+k~A12A21(−1−ε+k~)I].~𝐴~𝑘𝐼delimited-[]subscript𝐴1~𝑘subscript𝐴12subscript𝐴211𝜀~𝑘𝐼\tilde{A}+\tilde{k}I=\left[\begin{array}[]{cc}A_{1}+\tilde{k}&A_{12}\\ A_{21}&(-1-\varepsilon+\tilde{k})I\end{array}\right].</td>
<td></td>
</tr>
</tbody>
</table>
By the assumptions of the theorem, $`{{({{{- 1} - \varepsilon} + \overset{\sim}{k}})}\hspace{0pt}I} < 0`$. Let $`{\overline{A} = {{({A_{1} + {\overset{\sim}{k}\hspace{0pt}I}})} - {A_{12}\hspace{0pt}{({{({{{- 1} - \varepsilon} + \overset{\sim}{k}})}\hspace{0pt}I})}^{- 1}\hspace{0pt}A_{21}}}}.`$ Then
<table>
<tbody>
<tr>
<td></td>
<td>A¯=−N+1+k~+N−11+ε−k~<−N+1+k~+N−k~−1N−1(N−1)=0.¯𝐴𝑁1~𝑘𝑁11𝜀~𝑘missing-subexpression𝑁1~𝑘𝑁~𝑘1𝑁1𝑁1missing-subexpression0\begin{array}[]{ccl}\bar{A}&=&-N+1+\tilde{k}+\frac{N-1}{1+\varepsilon-\tilde{k}}\\ &<&-N+1+\tilde{k}+\frac{N-\tilde{k}-1}{N-1}(N-1)\\ &=&0.\end{array}</td>
<td></td>
</tr>
</tbody>
</table>
From the Schur complement Lemma, $`{\overline{A} + {\overset{\sim}{k}\hspace{0pt}I}} < 0`$, which leads to the assertion of the theorem. $`\square`$
Remark 4 Theorem 1 shows that under the control strategy of adding controllers to all the non-central nodes, as the constant $`\varepsilon > 0`$ increases, the eigenvalues of the matrix $`\overset{\sim}{A}`$ will approach $`- {({N - 1})}`$. Further, suppose that $`N > 2`$ and take $`\overset{\sim}{k} = 1`$. Then, based on Theorem 1, to guarantee that $`{{\lambda_{i}\hspace{0pt}{(\overset{\sim}{A})}} < {- 1}},{i = {1,\cdots,N}}`$, the feedback gain $`\varepsilon`$ only needs to satisfy
<table>
<tbody>
<tr>
<td></td>
<td>ε>N−1N−2.𝜀𝑁1𝑁2\varepsilon>\frac{N-1}{N-2}.</td>
<td></td>
<td>(10)</td>
</tr>
</tbody>
</table>
In this case, if the coupling strength of network ([9](#S3.E9 "In 3 Simple star-shaped networks")) is $`c`$, then the cost function is $`{C\hspace{0pt}F} = {c\hspace{0pt}\varepsilon\hspace{0pt}{({N - 1})}}`$. From ([10](#S3.E10 "In 3 Simple star-shaped networks")) and through direct calculations, it is easy to see that under the control strategy of adding controllers to the non-central nodes, $`{C\hspace{0pt}F} > \frac{{({N - 1})}^{2}}{({N - 2})}`$ suffices to guarantee that $`{{\lambda_{i}\hspace{0pt}{(\overset{\sim}{A})}} < {- 1}},{i = {1,\cdots,N}}`$.
Remark 5 Consider the control strategy that a single controller is added to the central node. In this case, the feedback gain matrix is given by $`G = {d\hspace{0pt}i\hspace{0pt}a\hspace{0pt}g\hspace{0pt}{\{\varepsilon,0,\cdots,0\}}} \in R^{N \times N}`$ and $`{C\hspace{0pt}F} = {c\hspace{0pt}\varepsilon}`$. Let $`\lambda_{1}\hspace{0pt}{(\overset{\sim}{A})}`$ represent the largest eigenvalue of matrix $`\overset{\sim}{A}`$. Then, based on Lemma 2, $`{\lambda_{1}\hspace{0pt}{(\overset{\sim}{A})}} \geq {- 1}`$ no matter how large the $`\varepsilon`$, or equally the cost function $`C\hspace{0pt}F`$, is taken. This shows that in order to guarantee $`\lambda_{i}`$ be small, the small nodes should be pinned.
Some simulations on the network ([9](#S3.E9 "In 3 Simple star-shaped networks")), shown in Fig. 1, are presented in Figs. 2$`\sim`$3, where the dynamics of an isolated node is a chaotic oscillator \[[13](#bib.bib13)\], $`\Gamma = {d\hspace{0pt}i\hspace{0pt}a\hspace{0pt}g\hspace{0pt}{\{ 0,\, 1,\, 0\}}}`$, and the synchronized state is set to be $`{s\hspace{0pt}{(t)}} = {\lbrack 7.9373\;7.9373\;21\rbrack}`$, which is an unstable equilibrium point of an isolated node \[[14](#bib.bib14)\].
\(a\) $`{C\hspace{0pt}F} = {c\hspace{0pt}\varepsilon} = {300\hspace{0pt}c} = 3000`$. (b) $`{C\hspace{0pt}F} = {c\hspace{0pt}\varepsilon\hspace{0pt}{({N - 1})}} = {1.5\hspace{0pt}{({N - 1})}\hspace{0pt}c} = 120`$. Fig. 2 Synchronization of the simple star-shaped coupling network ([9](#S3.E9 "In 3 Simple star-shaped networks")) with coupling strength $`c = 10`$: (a) adding a single controller to the central node; (b) adding controllers to all the non-central nodes.
\(a\) $`{C\hspace{0pt}F} = {c\hspace{0pt}\varepsilon} = {500\hspace{0pt}c} = 3500`$. (b) $`{C\hspace{0pt}F} = {c\hspace{0pt}\varepsilon\hspace{0pt}{({N - 1})}} = {1.5\hspace{0pt}{({N - 1})}\hspace{0pt}c} = 84`$. Fig. 3 Synchronization of the simple star-shaped coupling network ([9](#S3.E9 "In 3 Simple star-shaped networks")) with coupling strength $`c = 7`$: (a) pinning the central node; (b) pinning all the non-central nodes.
From Figs. 2$`\sim`$3, it can be seen that even with both a small coupling strength $`c`$ and a small cost function $`C\hspace{0pt}F`$, the synchronization of network ([9](#S3.E9 "In 3 Simple star-shaped networks")) can be faster to achieve when the controllers are added to all the non-central nodes (nodes with smaller degrees) than the case that a single controller is added to the central node (the node with a larger degree). Although the more controllers are needed for pinning nodes with smaller degrees, the total cost is still lower and the control effect is better in comparison.
## 4 Clusters of star-shaped networks with global coupling
Consider a network consisting of $`N`$ identical nodes in clusters of star-shaped coupling configuration: there are $`k`$ central nodes which are connected to each other; each central node may have different numbers of non-central nodes attached to it; there are no direct connections among the non-central nodes; and any central node has no direct connection to a non-central node attached to another central node. Fig. 4 gives an example of such a network for the case of $`k = 3`$.
Fig. 4 A cluster of star-shaped network with $`k = 3`$
Without loss of generality, let the first $`k`$ nodes be central nodes, each connecting to $`n_{i}`$ non-central nodes, where $`n_{i}`$ satisfy that $`{n_{1} \leq n_{2} \leq \cdots \leq n_{k}}.`$ Clearly, $`N = {k + {\sum_{i = 1}^{k}n_{i}}}`$. The coupling matrix $`A`$ of network ([1](#S1.E1 "In 1 Introduction and problem formulation")) in such clusters of star-shaped coupling configuration can been written as
<table>
<tbody>
<tr>
<td></td>
<td>A=[A1A12A21A2],𝐴delimited-[]subscript𝐴1subscript𝐴12subscript𝐴21subscript𝐴2A=\left[\begin{array}[]{cc}A_{1}&A_{12}\\ A_{21}&A_{2}\end{array}\right],</td>
<td></td>
<td>(11)</td>
</tr>
</tbody>
</table>
where $`A_{1} \in R^{k \times k}`$, $`A_{12} \in R^{k \times {({N - k})}}`$, $`A_{21} = A_{12}^{T} \in R^{{({N - k})} \times k}`$, and $`{A_{2} \in R^{{({N - k})} \times {({N - k})}}}.`$ Let $`\alpha_{i} = {(1,1,\cdots,1)} \in R^{n_{i}}`$, $`i = {1,\cdots,k}`$. Then
$`A_{1} = \begin{bmatrix}
{{{- k} + 1} - n_{1}} & 1 & \cdots & 1 \\
1 & {{{- k} + 1} - n_{2}} & \cdots & 1 \\
\vdots & \vdots & \ddots & \vdots \\
1 & 1 & \cdots & {{{- k} + 1} - n_{k}}
\end{bmatrix}`$, $`{A_{12} = \begin{bmatrix}
\alpha_{1} & & & \\
& \alpha_{2} & & \\
& & \ddots & \\
& & & \alpha_{k}
\end{bmatrix}},`$
$`A_{21} = \begin{bmatrix}
\alpha_{1}^{T} & & & \\
& \alpha_{2}^{T} & & \\
& & \ddots & \\
& & & \alpha_{k}^{T}
\end{bmatrix}`$, $`{A_{2} = {- I_{({N - k})}}}.`$
Let the feedback gain matrix be $`{G = \begin{bmatrix}
0 & 0 \\
0 & {\varepsilon\hspace{0pt}I_{({N - k})}}
\end{bmatrix}}.`$ Then, the corresponding coupling matrix $`\overset{\sim}{A}`$ of the controlled network ([4](#S1.E4 "In 1 Introduction and problem formulation")) is
<table>
<tbody>
<tr>
<td></td>
<td>A~=A−G.~𝐴𝐴𝐺\tilde{A}=A-G.</td>
<td></td>
<td>(12)</td>
</tr>
</tbody>
</table>
Theorem 2 If there are constants $`\varepsilon > 0`$ and $`\overset{\sim}{k} > 0`$, such that $`n_{1} > \overset{\sim}{k}`$, $`\varepsilon > {\overset{\sim}{k} - 1}`$ and $`\varepsilon > \frac{\overset{\sim}{k}\hspace{0pt}{({{n_{1} + 1} - \overset{\sim}{k}})}}{n_{1} - \overset{\sim}{k}}`$, then $`{{\lambda_{i}\hspace{0pt}{(\overset{\sim}{A})}} < {- \overset{\sim}{k}}},{i = {1,\cdots,N}}`$.
Proof: Consider the matrix
<table>
<tbody>
<tr>
<td></td>
<td>A~+k~I=[A1+k~IA12A21(−1−ε+k~)I].~𝐴~𝑘𝐼delimited-[]subscript𝐴1~𝑘𝐼subscript𝐴12subscript𝐴211𝜀~𝑘𝐼\tilde{A}+\tilde{k}I=\left[\begin{array}[]{cc}A_{1}+\tilde{k}I&A_{12}\\ A_{21}&(-1-\varepsilon+\tilde{k})I\end{array}\right].</td>
<td></td>
</tr>
</tbody>
</table>
By the assumptions of the theorem, $`{{({{{- 1} - \varepsilon} + \overset{\sim}{k}})}\hspace{0pt}I} < 0`$. Let $`\overline{A} = {{({A_{1} + {\overset{\sim}{k}\hspace{0pt}I}})} - {A_{12}\hspace{0pt}{({{({{{- 1} - \varepsilon} + \overset{\sim}{k}})}\hspace{0pt}I})}^{- 1}\hspace{0pt}A_{21}}}`$. Then $`\overline{A}`$ can be expressed as
<table>
<tbody>
<tr>
<td></td>
<td>A¯=[−k+1−n1+k~+n11+ε−k~1⋯11−k+1−n2+k~+n21+ε−k~⋯1⋮⋮⋱⋮11⋯−k+1−nk+k~+nk1+ε−k~].¯𝐴delimited-[]𝑘1subscript𝑛1~𝑘subscript𝑛11𝜀~𝑘1⋯11𝑘1subscript𝑛2~𝑘subscript𝑛21𝜀~𝑘⋯1⋮⋮⋱⋮11⋯𝑘1subscript𝑛𝑘~𝑘subscript𝑛𝑘1𝜀~𝑘\bar{A}=\left[\begin{array}[]{cccc}-k+1-n_{1}+\tilde{k}+\frac{n_{1}}{1+\varepsilon-\tilde{k}}&1&\cdots&1\\ 1&-k+1-n_{2}+\tilde{k}+\frac{n_{2}}{1+\varepsilon-\tilde{k}}&\cdots&1\\ \vdots&\vdots&\ddots&\vdots\\ 1&1&\cdots&-k+1-n_{k}+\tilde{k}+\frac{n_{k}}{1+\varepsilon-\tilde{k}}\end{array}\right].</td>
<td></td>
</tr>
</tbody>
</table>
Let $`{{m_{i} = {{{{- k} + 1} - n_{i}} + \overset{\sim}{k} + \frac{n_{i}}{{1 + \varepsilon} - \overset{\sim}{k}} + {{({k - 1})} \cdot 1}}},{i = {1,2,\cdots,k}}}.`$ Then
<table>
<tbody>
<tr>
<td></td>
<td>mi=−ni+k~+ni1+ε−k~<−ni+k~+n1−k~n1ni=ni(−k~n1)+k~=k~(1−nin1)≤0.subscript𝑚𝑖subscript𝑛𝑖~𝑘subscript𝑛𝑖1𝜀~𝑘missing-subexpressionsubscript𝑛𝑖~𝑘subscript𝑛1~𝑘subscript𝑛1subscript𝑛𝑖missing-subexpressionsubscript𝑛𝑖~𝑘subscript𝑛1~𝑘missing-subexpression~𝑘1subscript𝑛𝑖subscript𝑛10\begin{array}[]{ccl}m_{i}&=&-n_{i}+\tilde{k}+\frac{n_{i}}{1+\varepsilon-\tilde{k}}\\ &<&-n_{i}+\tilde{k}+\frac{n_{1}-\tilde{k}}{n_{1}}n_{i}\\ &=&n_{i}(\frac{-\tilde{k}}{n_{1}})+\tilde{k}\\ &=&\tilde{k}(1-\frac{n_{i}}{n_{1}})\leq 0.\end{array}</td>
<td></td>
</tr>
</tbody>
</table>
According to the Gerschgorin disc theorem \[[8](#bib.bib8)\], $`\overline{A} < 0`$. By the Schur complement Lemma, $`{\overset{\sim}{A} + {\overset{\sim}{k}\hspace{0pt}I}} < 0`$, which leads to the assertion of the theorem. $`\square`$
Some simulations on the network ([11](#S4.E11 "In 4 Clusters of star-shaped networks with global coupling")) shown in Fig. 4 are presented in Fig. 5, where the dynamics of an isolated node, the inner liking matrix $`\Gamma`$, and the synchronized state are the same as that given in Section 2.
\(a\) $`{C\hspace{0pt}F} = {c\hspace{0pt}\varepsilon\hspace{0pt}k} = {300\hspace{0pt}k\hspace{0pt}c} = 9000`$. (b) $`{C\hspace{0pt}F} = {c\hspace{0pt}\varepsilon\hspace{0pt}{({\sum_{i = 1}^{k}n_{i}})}} = {2.5\hspace{0pt}{({\sum_{i = 1}^{k}n_{i}})}\hspace{0pt}c} = 225`$. Fig. 5 Synchronization of network ([11](#S4.E11 "In 4 Clusters of star-shaped networks with global coupling")) with coupling strength $`c = 10`$: (a) pinning $`k`$ central nodes; (b) pinning all the non-central nodes.
From Fig. 5, it can be seen that although all the non-central nodes are pinned, the cost is lower and the effect is better than the case where the central nodes are pinned.
By using the Kronecker product, the controlled network ([5](#S2.E5 "In 2 The cost of pinning control")) can be rewritten as
<table>
<tbody>
<tr>
<td></td>
<td>e˙(t)=F(x,s)+c(A−G)⊗Γe(t),˙𝑒𝑡𝐹𝑥𝑠tensor-product𝑐𝐴𝐺Γ𝑒𝑡\dot{e}(t)=F(x,s)+c(A-G)\otimes\Gamma e(t),</td>
<td></td>
<td>(13)</td>
</tr>
</tbody>
</table>
where $`{e\hspace{0pt}{(t)}} = {({e_{1}^{T}\hspace{0pt}{(t)}},\cdots,{e_{N}^{T}\hspace{0pt}{(t)}})}^{T}`$, $`{F\hspace{0pt}{(x,s)}} = {({({{f\hspace{0pt}{({x_{1}\hspace{0pt}{(t)}})}} - {f\hspace{0pt}{({s\hspace{0pt}{(t)}})}}})}^{T},\cdots,{({{f\hspace{0pt}{({x_{N}\hspace{0pt}{(t)}})}} - {f\hspace{0pt}{({s\hspace{0pt}{(t)}})}}})}^{T})}^{T}`$, $`G = {d\hspace{0pt}i\hspace{0pt}a\hspace{0pt}g\hspace{0pt}{\{\underset{k}{\underbrace{0,\cdots,0}},\underset{N - k}{\underbrace{\varepsilon,\cdots,\varepsilon}}\}}} \in R^{N \times N}`$ and $`\Gamma = {d\hspace{0pt}i\hspace{0pt}a\hspace{0pt}g\hspace{0pt}{\{ r_{1},r_{2},\cdots,r_{n}\}}}`$, $`r_{i} = 1`$ or $`{{0,i} = 1},{\cdots,n}`$.
Theorem 3 If there are a positive diagonal matrix $`P`$ and constants $`{\mu > 0},{{\varepsilon > 0},{{\overset{\sim}{k} > 0},{c > 0}}}`$, such that $`n_{1} > \overset{\sim}{k}`$, $`\varepsilon > {\overset{\sim}{k} - 1}`$, $`\varepsilon > \frac{\overset{\sim}{k}\hspace{0pt}{({{n_{1} + 1} - \overset{\sim}{k}})}}{n_{1} - \overset{\sim}{k}}`$ and
<table>
<tbody>
<tr>
<td></td>
<td>(x−y)TP(f(x,t)−f(y,t)−ck~Γ(x−y))≤−μ(x−y)T(x−y),superscript𝑥𝑦𝑇𝑃𝑓𝑥𝑡𝑓𝑦𝑡𝑐~𝑘Γ𝑥𝑦𝜇superscript𝑥𝑦𝑇𝑥𝑦(x-y)^{T}P(f(x,t)-f(y,t)-c\tilde{k}\Gamma(x-y))\leq-\mu(x-y)^{T}(x-y),</td>
<td></td>
</tr>
</tbody>
</table>
then the controlled network ([4](#S1.E4 "In 1 Introduction and problem formulation")) is globally exponentially synchronized to $`s\hspace{0pt}{(t)}`$.
Proof: Choose a Lyapunov function as $`{V\hspace{0pt}{(t)}} = {\frac{1}{2}\hspace{0pt}e^{T}\hspace{0pt}{(t)}\hspace{0pt}{({I_{N} \otimes P})}\hspace{0pt}e\hspace{0pt}{(t)}}`$. Its time derivative is
<table>
<tbody>
<tr>
<td></td>
<td>V˙(t)=eT(t)(IN⊗P)e˙(t)=eT(t)(IN⊗P)(F(x,s)+c(A−G)⊗Γe(t))=eT(t)(IN⊗P)F(x,s)+ceT(t)((A−G)⊗PΓ)e(t)=eT(t)(IN⊗P)(F(x,s)−c(k~In⊗Γ)eT(t))+ceT(t)((A−G)+k~In)⊗PΓe(t).˙𝑉𝑡superscript𝑒𝑇𝑡tensor-productsubscript𝐼𝑁𝑃˙𝑒𝑡missing-subexpressionsuperscript𝑒𝑇𝑡tensor-productsubscript𝐼𝑁𝑃𝐹𝑥𝑠tensor-product𝑐𝐴𝐺Γ𝑒𝑡missing-subexpressionsuperscript𝑒𝑇𝑡tensor-productsubscript𝐼𝑁𝑃𝐹𝑥𝑠𝑐superscript𝑒𝑇𝑡tensor-product𝐴𝐺𝑃Γ𝑒𝑡missing-subexpressionsuperscript𝑒𝑇𝑡tensor-productsubscript𝐼𝑁𝑃𝐹𝑥𝑠𝑐tensor-product~𝑘subscript𝐼𝑛Γsuperscript𝑒𝑇𝑡tensor-product𝑐superscript𝑒𝑇𝑡𝐴𝐺~𝑘subscript𝐼𝑛𝑃Γ𝑒𝑡\begin{array}[]{ccl}\dot{V}(t)&=&e^{T}(t)(I_{N}\otimes P)\dot{e}(t)\\ &=&e^{T}(t)(I_{N}\otimes P)(F(x,s)+c(A-G)\otimes\Gamma e(t))\\ &=&e^{T}(t)(I_{N}\otimes P)F(x,s)+ce^{T}(t)((A-G)\otimes P\Gamma)e(t)\\ &=&e^{T}(t)(I_{N}\otimes P)(F(x,s)-c(\tilde{k}I_{n}\otimes\Gamma)e^{T}(t))+ce^{T}(t)((A-G)+\tilde{k}I_{n})\otimes P\Gamma e(t).\end{array}</td>
<td></td>
</tr>
</tbody>
</table>
By Theorem 2, $`{{({A - G})} + {\overset{\sim}{k}\hspace{0pt}I_{n}}} < 0`$. So,
<table>
<tbody>
<tr>
<td></td>
<td>V˙(t)≤−μ∑i=1NeiT(t)ei(t)≤−μ∑i=1N1max{Pj}1≤j≤neiT(t)Pei(t)=−μmax{Pj}1≤j≤neiT(t)(In⊗P)ei(t)=−μmax{Pj}1≤j≤nV(t),\begin{array}[]{ccl}\dot{V}(t)&\leq&-\mu\sum_{i=1}^{N}e_{i}^{T}(t)e_{i}(t)\leq-\mu\sum_{i=1}^{N}\frac{1}{\max\{P_{j}\}_{1\leq j\leq n}}e_{i}^{T}(t)Pe_{i}(t)\\ \\ &=&-\frac{\mu}{\max\{P_{j}\}_{1\leq j\leq n}}e_{i}^{T}(t)(I_{n}\otimes P)e_{i}(t)=-\frac{\mu}{\max\{P_{j}\}_{1\leq j\leq n}}V(t),\end{array}</td>
<td></td>
</tr>
</tbody>
</table>
and the theorem is thus proved. $`\square`$
## 5 Some non-regular complex networks
In this section, first, a non-regular coupled network consisting of 20 nodes is generated following the procedure of the well-known BA model. Fig. 6 shows the synchronization of the network with controllers being added to the three “biggest” nodes of degrees 15, 13 and 10, respectively. The dynamics of an isolated node, the inner liking matrix $`\Gamma`$, and the synchronized state are the same as that given in Section 2. From Figs. 6(c) and 6(d), it can be seen that the network does not synchronize any faster, although a larger feedback gain is used. In Fig. 7, a different control scheme is applied: eleven nodes with smaller degrees in the network are pinned with coupling strength $`c = 6`$ and feedback gain $`\varepsilon = 5`$, yielding cost function $`{C\hspace{0pt}F} = 330`$. Compared with Fig. 6, a better control performance is obtained with a much smaller cost function. Figs. 8 and 9 also show that it is more efficient by pinning nodes with smaller degrees than pinning nodes with larger degrees. As shown in Fig. 8, to achieve a similar synchronization effect, a much smaller feedback gain is needed in pinning “smaller” nodes than pinning “larger” ones, thereby the cost function is also smaller in the former control scheme than that in the latter. Fig. 9 shows that although the coupling strength, the feedback gain and the cost function are similar, the synchronization effect is better in pinning “smaller” nodes than pinning “larger” ones.
\(a\) (b)
\(c\) (d) Fig. 6 Pinning the three “biggest” nodes with degrees 15, 13 and 10 in a 20-node non-regular coupled network: (a) $`{{c = 0},{{\varepsilon = 0},{{C\hspace{0pt}F} = 0}}}.`$ (b) $`{{c = 6},{{\varepsilon = 0},{{C\hspace{0pt}F} = 0}}}.`$ (c) $`{{c = 6},{{\varepsilon = 500},{{C\hspace{0pt}F} = 9000}}}.`$ (d) $`{{c = 6},{{\varepsilon = 1000},{{C\hspace{0pt}F} = 18000}}}.`$
Fig. 7 Pinning the eleven “smaller” nodes in a 20-node non-regular coupled network: $`c = 6`$, $`\varepsilon = 5`$, $`{{C\hspace{0pt}F} = 330}.`$
\(a\) $`{{c = 8},{{\varepsilon = 500},{{C\hspace{0pt}F} = 12000}}}.`$ (b) $`{{c = 6},{{\varepsilon = 8},{{C\hspace{0pt}F} = 528}}}.`$ Fig. 8 (a) Pinning the three “biggest” nodes with degrees 15, 13 and 10. (b) Pinning the eleven “smaller” nodes.
\(a\) $`{{c = 6},{{\varepsilon = 55},{{C\hspace{0pt}F} = 660}}}.`$ (b) $`{{c = 6},{{\varepsilon = 22},{{C\hspace{0pt}F} = 660}}}.`$ (c) $`{{c = 6},{{\varepsilon = 10},{{C\hspace{0pt}F} = 660}}}.`$ Fig. 9 (a) Pinning the two “biggest” nodes with degrees 15 and 13. (b) Pinning the three “biggest” nodes with degrees 15, 13 and 10, and the two “smallest” nodes with the same degrees 3. (c) Pinning the eleven “smaller” nodes.
It seems more efficient to use the control scheme of pinning nodes with smaller degrees than pinning nodes with larger degrees for achieving the desired synchronous states of some non-regular coupled dynamical networks. This is an interesting phenomenon that was not noticed before.
## 6 Conclusions
In this paper, pinning control to achieve synchronization of complex dynamical networks is further investigated. In contrary to the general perception, the nodes with smaller degrees play an important role in the synchronizability of some networks. It has been shown by computer simulations in this paper that to achieve a similar synchronization effect on the networks considered, a smaller coupling strength, a smaller feedback gain and thus a lower cost function are needed in the control scheme of pinning nodes with smaller degrees comparing with those needed in pinning nodes with larger degrees. In other words, it seems more efficient to use the control scheme of pinning nodes with smaller degrees than that of pinning nodes with larger degrees for some dynamical networks to achieve synchronization, an interesting phenomenon that deserves further investigation in the future.
## References
- \[1\]
Barabási, A. L., Albert, R. $`\&`$ Jeong, H. \[1999\] “Mean-field theory for scale-free random networks”, Phys. A, Vol. 272, 173-187.
- \[2\]
Boccaletti, S., Latora, V., Moreno, Y., Chavez, M., $`\&`$ Hwang D. -U. \[2006\] “Complex networks: structure and dynamics”, Physics Reports, 424, 175-308.
- \[3\]
Chen, G., Fan, Z. P. $`\&`$ Li, X. \[2004\] “Modelling the complex Internet topology”, in Complex Dynamics in Communication Network, ed. by L. Kocarev and G. Vattay, New York: Springer-Verlag, 213-234.
- \[4\]
Chen, T. P., Liu, X. W. $`\&`$ Lu, W. L. \[2006\] “Pinning complex networks by a single controller”, arXiv: math-ph/0601030
- \[5\]
Chua, L. O. \[1998\] CNN: A paradigm for complexity. Singapore: World Scientific.
- \[6\]
Fan, Z. P. $`\&`$ Chen, G. \[2005\] “Pinning control of scale-free complex networks”, Proceedings of IEEE International Symposium on Circuits and Systems, Vol. 1, 284-287.
- \[7\]
Gahinet, P., Nemirovski, A. $`\&`$ Laub, A. J. \[1995\] LMI control toolbox. Natick, Mass: The Math Works, Inc.
- \[8\]
Horn, R. $`\&`$ Johnson, C. \[1985\] Matrix analysis. Cambridge: Cambridge Univ. Press.
- \[9\]
Li, X., Jin, Y. Y. $`\&`$ Chen, G. \[2004\] “Complexity and synchronization of the world trade web”, Phys. A, Vol. 328, 367-378.
- \[10\]
Li, X. $`\&`$ Chen, G. \[2003\] “Synchronization and desynchronization of complex dynamical networks: an engineering viewpoint”, IEEE Trans. on Circuits and Systems I-Fund. Th. Appl., Vol. 50, No. 11, 1381-1390.
- \[11\]
Li, X., Wang, X. F. $`\&`$ Chen, G. \[2004\] “Pinning a complex dynamical network to its equilibrium”, IEEE Trans. on Circuits and Systems I-Fund. Th. Appl., Vol. 51, No. 10, 2074-2087.
- \[12\]
Lü, J. H. $`\&`$ Chen, G. \[2004\] “Chaos synchronization of general complex dynamical networks”, Phys. A, Vol. 334, 281-302.
- \[13\]
Lü, J. H., Chen, G. $`\&`$ Zhang, S. \[2002\] “The compound structure of a new chaotic attractor”, Chaos, Solotons and Fractals Vol. 14, No. 5, 669-672.
- \[14\]
Lü, J. H., Chen, G. $`\&`$ Zhang, S. \[2002\] “Dynamical analysis of a new chaotic attractor”, Int. J. of Bifurcation and Chaos, Vol. 12, No. 5, 1001-1015.
- \[15\]
Newman M E J. \[2003\] “The structure and function of complex networks”, SIAM Review, 45, 167-256.
- \[16\]
Posadas-Castillo, C., Cruz-Henández, C. $`\&`$ López-Mancilla, D. \[2005\], “Sincronización de múltiples osciladores de Rössler”, Yearly Congress of the Association of Automatic Control Mexico,Cuernavaca Morelos, Oct. 19-21.
- \[17\]
Strogatz, S. H. \[2001\] “Exploring complex networks”, Nature, Vol. 410, 268-176.
- \[18\]
Wang, X. F. $`\&`$ Chen, G. \[2002\] “Synchronization in samll-world dynamical networks”, Int. J. Bifurcation and Chaos, Vol. 12, No. 1, 187-192.
- \[19\]
Watts, D. J. $`\&`$ Strogatz, S. H. \[1998\] “Collective dynamics of ‘small world’ network”, Nature, Vol. 393, 440-442.
- \[20\]
Wu, C. W. $`\&`$ Chua, L. O. \[1995\] “Synchronization in an array of linearly coupled dynamical systems”, IEEE Trans. on Circuits and Systems I-Fund. Th. Appl., Vol. 42, No. 8, 430-447.
- \[21\] Xiang, L. Y., Liu, Z. X., Chen. Z. Q., Chen, G. $`\&`$ Yuan, Z. Z. \[2006\] “Pinning a complex dynamical network by a small number of feedback controllers”, to appear. |
0710.1190 | cs.NI | cyber | # Power Efficient Scheduling under Delay Constraints over Multi-user Wireless Channels
Nitin Salodkar, Abhay Karandikar, and V. S. Borkar, Nitin Salodkar is with Department of Computer Science and Engineering, IIT Bombay, Powai, Mumbai India - 400076. Email: nitins@cse.iitb.ac.in Abhay Karandikar is with Department of Electrical Engineering, IIT Bombay, Powai, Mumbai India - 400076. Email: karandi@ee.iitb.ac.in V. S. Borkar is with the School of Technology and Computer Science, Tata Institute of Fundamental Research, Homi Bhabha Road, Mumbai, India - 400005. Email: borkar@tifr.res.in
###### Abstract
In this paper, we consider the problem of power efficient uplink scheduling in a Time Division Multiple Access (TDMA) system over a fading wireless channel. The objective is to minimize the power expenditure of each user subject to satisfying individual user delay. We make the practical assumption that the system statistics are unknown, i.e., the probability distributions of the user arrivals and channel states are unknown. The problem has the structure of a Constrained Markov Decision Problem (CMDP). Determining an optimal policy under for the CMDP faces the problems of state space explosion and unknown system statistics. To tackle the problem of state space explosion, we suggest determining the transmission rate of a particular user in each slot based on its channel condition and buffer occupancy only. The rate allocation algorithm for a particular user is a learning algorithm that learns about the buffer occupancy and channel states of that user during system execution and thus addresses the issue of unknown system statistics. Once the rate of each user is determined, the proposed algorithm schedules the user with the best rate. Our simulations within an IEEE 802.16 system demonstrate that the algorithm is indeed able to satisfy the user specified delay constraints. We compare the performance of our algorithm with the well known M-LWDF algorithm. Moreover, we demonstrate that the power expended by the users under our algorithm is quite low.
###### Index Terms:
Multi-user Fading Channel, Markov Decision Process, Energy Efficient Scheduling
## I Introduction
Broadband wireless networks like IEEE 802.16 \[[1](#bib.bib1)\] and 3G cellular \[[2](#bib.bib2)\] are expected to provide Quality of Service (QoS) for emerging multimedia applications. One of the challenges in providing QoS is the time varying nature of the wireless channel due to multipath fading \[[3](#bib.bib3)\]. Moreover, for portable and hand-held devices, energy efficiency is also an important consideration.
In a multi-user wireless system, recent studies \[[4](#bib.bib4), [5](#bib.bib5)\] suggest that since the wireless channel fades independently across different users, this diversity can be exploited by *opportunistically* scheduling the user with the best channel gain. This leads to significant performance gain in terms of total system throughput. Such scheduling algorithms that exploit the characteristics of the physical channel to satisfy some network level QoS performance metrics are refererred to as cross layer scheduling algorithms \[[6](#bib.bib6)\]. Power required to transmit *reliably* at a certain rate under better channel conditions is much less than that required under poorer channel conditions at the same rate \[[3](#bib.bib3)\]. This suggests that in order to save power, one should transmit at higher rates under better channel conditions, this leads to queuing delays. Moreover, since transmission power is an increasing and strictly convex function of the transmission rate \[[3](#bib.bib3)\], power efficiency can also be achieved by transmitting the data at lower rates, albeit at the cost of increased queuing delay thus leading to a power-delay tradeoff.
In this paper, we consider a single cell multi-user wireless uplink system with Time Division Multiple Access (TDMA). For such a system, we consider the problem of determining the user to be scheduled in each time slot so that the average transmission power expended by *each user* is minimized subject to a constraint on the average queuing delay experienced by *each user*. Moreover, we assume a peak power constraint, i.e., in each slot the transmission power of a user is less than or equal to a certain maximum value. This scenario may correspond to a base station scheduling users on an uplink in an IEEE 802.16 system to satisfy delay constraint of each user.
There is a copious literature on cross layer scheduling algorithms. See \[[7](#bib.bib7)\] for a succinct review. The scheduling problem is typically formulated as an optimization problem with an objective of efficiently allocating resources such as time, frequency bands, power, codes etc. to the users under physical layer (wireless channel) and/or network layer QoS constraints. Various QoS constraints have been considered in the literature like system throughput, minimum rate, maximum delay, delay bound, queue stability and fairness. A *scheduling policy* is an allocation rule that allocates these resources based on parameters like channel conditions of the users, their queue lengths etc. In this paper, we concentrate on efficiently allocating power and rates to users based on their channel condition and queue length. The power allocation policy is considered feasible if it satisfies certain average or peak power constraints. On the other hand, the rate allocation policy is considered feasible if the physical layer can deliver the data reliably to the users at a given rate. The set of all feasible rate tuples is called the feasible capacity region \[[7](#bib.bib7)\].
A scheduling policy is considered *stable* if the expected queue lengths are bounded under the policy. Many scheduling policies proposed in the literature have considered stability as a QoS criterion. In \[[4](#bib.bib4)\], the authors determine the throughput capacity region of a multi-access system, i.e., the set of all feasible rates with average power constraints. In \[[8](#bib.bib8)\], the authors have shown that the throughput capacity region is same as the multi-access stability region (i.e., the set of all arrival vectors for which there exists some rate and power allocation policies that keep the system stable.). A scheduler is termed *thoughput-optimal* if it can maintain the stability of the system as long as the arrival rate is within the stability region. Thoughput optimal scheduling policies have been explored in \[[4](#bib.bib4)\], \[[9](#bib.bib9)\]. Longest Connected Queue (LCQ) \[[10](#bib.bib10)\] , Exponential (EXP) \[[11](#bib.bib11)\] , Longest Weighted Queue Highest Possible Rate (LWQHPR) \[[12](#bib.bib12)\] and Modified Longest Weighted Delay First (M-LWDF) \[[13](#bib.bib13)\] are other well known throughput optimal scheduling policies. In \[[14](#bib.bib14)\], the authors define the notion of *delay limited capacity* of a multi-access system, i.e., the maximum rate achievable such that the delay is independent of the fading characteristics.
While thoughput-optimal scheduling policies maintain the stability of the queueing system, they do not necessarily guarantee small queue lengths and consequently lower delays. Delay-optimal scheduling deals with optimal rate and power allocation such that the average queue length and hence average delay are minimized for arrival rates within the stability region under average and peak power constraints. Due to the nature of the constraints, there is no loss of optimality in choosing the rate and power allocation policies separately \[[7](#bib.bib7)\]. Hence to simplify the problem, one can choose any stationary power allocation policy that satisfies the peak and average power constraints. The delay optimal policy therefore deals with with optimal rate allocation for minimizing delays under a given power allocation policy. It has been shown that the Longest Queue Highest Possible Rate (LQHPR) policy \[[15](#bib.bib15)\] (besides being thoughput optimal) is also delay optimal for any symmetric power control under symmetric fading provided that the packet arrival process is Poisson and packet length is exponentially distributed.
Apart from throughput and delay optimal policies, opportunistic scheduling with various fairness constraints have been explored in \[[16](#bib.bib16), [17](#bib.bib17)\].
In this paper, our focus is on rate allocation with a constraint on peak power as well as average queueing delay which acts as the QoS metric. This problem for a single user wireless channel without the peak power constraints has been explored in the pioneering work of \[[18](#bib.bib18)\]. The problem with many generalizations on arrival and channel gain processes have been considered in subsequent papers \[[19](#bib.bib19), [20](#bib.bib20), [21](#bib.bib21), [22](#bib.bib22), [23](#bib.bib23), [24](#bib.bib24), [25](#bib.bib25)\]. In most of these papers, the scheduling policy has been formulated as a control policy within the Markov Decision Process (MDP) framework. However, only structural results of the optimal policy are available under various assumptions and that too for a single user scenario only. There is very little work for extending the vast body of literature on delay constrained power efficient scheduling to multi-user scenario. Recently, in \[[26](#bib.bib26)\], the author has extended the asymptotic analysis of Berry-Gallger \[[18](#bib.bib18)\] for exploiting the power-delay tradeoff in multi-user system. The objective is to minimize the total power on the *downlink* subject to user queue stability constraints. The author using the concept of Lyapunov Drift Steering has also given an algorithm that comes within a logarithmic factor of achieving the Berry-Gallager power-delay bound. However, on the downlink, the base station typically transmits with a constant maximum power sufficient to reach the farthest user and hence power minimization in not a major concern. Moreover, minimizing the sum power can lead to unfairness, i.e., users with better average channel conditions might get a far higher share of the bandwidth than the users who have relatively poor average channel conditions. On the other hand, for the *uplink*, the problem is to minimize the power of each user subject to individual delay constraint, which has not been addressed in the literature so far.
In \[[27](#bib.bib27)\], the author has extended the analysis for single user case to the multi-user case, albeit with only two users which can be applicable for the uplink also. Beyond two users, the problem becomes unwieldy to gain any useful insight, primarily due to large state space. For the two user case, the author has given an elegant near optimal policy where each user’s rate allocation is determined by the joint channel states across users and the user’s own queue state. Thus each user’s queue evolution process behaves as if it were controlled by a single user policy. However, computation of user’s transmission power still takes into account the joint channel and queue state processes.
Even for the single user case \[[19](#bib.bib19), [20](#bib.bib20), [21](#bib.bib21), [22](#bib.bib22), [23](#bib.bib23), [24](#bib.bib24), [25](#bib.bib25)\], practical implementation of optimal policy is far from simple. This is because a knowledge of the probability distributions of the arrival and channel gain process is required for computing the optimal policy. This knowledge is not available in practice. While, we have addressed this limitation by formulating an on-line algorithm within stochastic approximation framework in \[[28](#bib.bib28)\], this algorithm still deals with the single user scenario. This algorithm does not assume any explicit knowledge of the probability distributions of the channel gain and arrival processes. In this paper, we consider a multi-user wireless system. The state of the system is defined as the minimum information required by the scheduler for making scheduling decisions. For the multi-user scenario considered in this paper the state space is considerably large as compared to that for the single user scenario. We illustrate this with a simple example. Let us assume that the channel condition of a user can be represented using $`8`$ states. This is a practical assumption and has been justified in \[[29](#bib.bib29)\]. Let us assume that each user has a buffer in which at most $`50`$ packets can be stored. For a single user system, the channel state and buffer occupancy of the user forms the state of the system in any time slot. The number of states is $`{8 \times 50} = 400`$. Now consider a multi-user system with $`4`$ users. In this case, the state of the system consists of the channel state and buffer occupancy of each user. The state space consists of $`{50^{4} \times 8^{4}} = {2.56 \times 10^{10}}`$ states. Furthermore, the number of states increases exponentially with the users. Hence determining the optimal policy by estimating the dynamic programming value function would take prohibitively long time. Hence in this paper, we propose a alternate approach.
In our approach, each user’s queue evolution behaves as if it were controlled by a single user policy. Depending on each user’s channel state and queue size, the algorithm allocates a certain rate to each user in a slot using the single user algorithm outlined in this paper. The algorithm then schedules the user with the highest rate in a slot. From the structural properties of optimal policy for a single user scenario, it is well known that the optimal policy is increasing in queue length and channel gain \[[27](#bib.bib27)\]. Thus more number of packets are transmitted when the queue length is greater or the channel gain is higher. Hence a user transmitting at a high rate has either very good channel condition, or large queue length. Scheduling such a user, therefore, either leads to its power savings or aids in satisfying the delay constraint.
The scheduling algorithms proposed in the literature like EXP scheduler \[[11](#bib.bib11)\], LQHPR scheduler \[[15](#bib.bib15)\], M-LWDF \[[13](#bib.bib13)\] scheduler require the queue length information for determining the scheduling decision. In the downlink scenario, this information is readily available to the scheduler residing at the base station. However, in the uplink scenario, this information needs to be communicated by the users to the scheduler. Communicating the queue length information poses a significant overhead. In our approach, each user determines the rate at which it would transmit if it were scheduled in a slot. All the users inform these rates to the base station. The base station then schedules the user with the highest rate. Thus by communicating the rates directly, we avoid the queue length communication overhead.
The IEEE 802.16 system is an emerging system for broadband wireless access and is expected to provide QoS to the users. Through our simulations in an IEEE 802.16 system, we demonstrate that the algorithm is indeed able to satisfy the delay constraints of the users. Moreover, we demonstrate that the power expenditure of a user is commensurate with its delay requirement, the average arrival rate and average channel conditions. The higher the delay, lower the average arrival rate and better the average channel conditions, the lower is the power expenditure.
The contributions of this paper are summarized as follows:
1. 1\.
We formulate the problem of minimizing the average power expended by each user subject to a constraint on individual user delay as a constrained optimization problem. To the best of our knowledge, this problem has not been studied in a multi-user uplink scenario.
2. 2\.
We propose an online algorithm that does not require the knowledge of the probability distributions of the channel states and the arrivals of the users.
3. 3\.
The computational complexity of our approach increases only linearly with the number of users.
4. 4\.
The communication overhead of our approach is low and hence the algorithm is suitable for practical implementation. The algorithm satisfies the delay constraints of the users. We demonstrate the power efficiency our our algorithm through comparison with the M-LWDF algorithm within an IEEE 802.16 system simulation.
The rest of the paper is organized as follows. In Section [II](#S2 "II System Model ‣ Power Efficient Scheduling under Delay Constraints over Multi-user Wireless Channels"), we present the system model. We formulate the problem as an optimization problem in Section [III](#S3 "III Problem Formulation ‣ Power Efficient Scheduling under Delay Constraints over Multi-user Wireless Channels"), where we show that the problem has the structure of a Constrained Markov Decision Problem (CMDP). We discuss the issues like large state space and unknown system model in determining an optimal solution using the traditional CMDP solution techniques. In Section [IV](#S4 "IV Single User Scenario in Presence of Transmitter Errors ‣ Power Efficient Scheduling under Delay Constraints over Multi-user Wireless Channels"), we consider and extension to the traditional single user scenario based on transmitter induced errors. In Section [V](#S5 "V An Online Primal Dual Algorithm For the Multi-user Problem ‣ Power Efficient Scheduling under Delay Constraints over Multi-user Wireless Channels"), we propose an online algorithm that is based on the extension to the single user scenario detailed in Section [IV](#S4 "IV Single User Scenario in Presence of Transmitter Errors ‣ Power Efficient Scheduling under Delay Constraints over Multi-user Wireless Channels"). We also discuss the implementation issues. We present the simulation setup and discuss results in Section [VI](#S6 "VI Simulation Setup and Results ‣ Power Efficient Scheduling under Delay Constraints over Multi-user Wireless Channels"). Finally, we conclude in Section [VII](#S7 "VII Conclusion ‣ Power Efficient Scheduling under Delay Constraints over Multi-user Wireless Channels").
## II System Model
[Refer to caption](/html/0710.1190/assets/x1.png)
Figure 1: System Model
As illustrated in Figure [1](#S2.F1 "Figure 1 ‣ II System Model ‣ Power Efficient Scheduling under Delay Constraints over Multi-user Wireless Channels"), we consider uplink transmissions in a TDMA system with $`N`$ users, i.e., time is divided into slots of equal duration and only one user is allowed to transmit in a slot. We assume that the slot duration is normalized to unity. The base station is a centralized entity that schedules the users in every slot. We assume a fading wireless channel where the channel gain is assumed to remain constant for the duration of the slot and to change in an independent and identically distributed (i.i.d.) manner across slots. This model is called as the *block fading* model \[[18](#bib.bib18)\]. We assume that the fading across users is also i.i.d. Under these assumptions, if a user $`i`$ transmits a signal $`y_{n}^{i}`$ in slot $`n`$, then the received signal $`Y_{n}^{i}`$ can be expressed as,
<table>
<tbody>
<tr>
<td></td>
<td>Yni=Hniyni+Gn,subscriptsuperscript𝑌𝑖𝑛subscriptsuperscript𝐻𝑖𝑛subscriptsuperscript𝑦𝑖𝑛subscript𝐺𝑛Y^{i}_{n}=H^{i}_{n}y^{i}_{n}+G_{n},</td>
<td></td>
<td>(1)</td>
</tr>
</tbody>
</table>
where $`H_{n}^{i}`$ denotes the complex channel gain due to fading and $`G_{n}`$ denotes the complex additive white Gaussian noise with zero mean and variance $`N_{0}`$. Let $`X_{n}^{i} = {|H_{n}^{i}|}^{2}`$ be the channel state for user $`i`$ in slot $`n`$. Practically $`H_{n}^{i}`$ is a continuous random variable and hence so is $`X_{n}^{i}`$. However, in this paper we assume that $`X_{n}^{i}`$ takes only finite and discrete values from a set $`\mathcal{X}`$. This assumption has been justified in \[[18](#bib.bib18)\], \[[19](#bib.bib19)\]. In this paper, we assume that the distribution of $`H_{n}^{i}`$ and hence that of $`X_{n}^{i}`$ is unknown.
Each user possesses a finite buffer of $`B`$ bits. Bits arrive into the user buffer and are queued until they are transmitted. The arrival process for each user is assumed to be i.i.d. across slots. Let $`A_{n}^{i}`$ denote the number of bits arriving into the user $`i`$ buffer in slot $`n`$. We assume that the random variable $`A_{n}^{i}`$ takes values from a finite and discrete set $`\mathcal{A}\overset{\Delta}{=}{\{ 0,\ldots,A\}}`$. Like $`X_{n}^{i}`$, we assume that the distribution of $`A_{n}^{i}`$ is unknown. Let $`Q_{n}^{i}`$ denote the queue length or buffer occupancy of user $`i`$ in slot $`n`$. Let $`U_{n}^{i}`$ denote the number of bits transmitted by user $`i`$ in slot $`n`$. We assume that $`U_{n}^{i}`$ takes values from the set $`\mathcal{U}\overset{\Delta}{=}{\{ 0,\ldots,B\}}`$. Let $`I_{n}^{i}`$ be an indicator variable that is set to $`1`$ if user $`i`$ is scheduled in slot $`n`$ and is set to $`0`$ otherwise. Let $`\mathbf{I}_{n}`$ be the vector $`\lbrack I_{n}^{1},\ldots,I_{n}^{N}\rbrack`$. Note that since only one user can transmit in a slot, only one element of $`\mathbf{I}_{n}`$ is equal to $`1`$ and the rest are $`0`$. Let $`\mathcal{I}`$ be the set of all possible $`N`$ dimensional vectors with one element equal to $`1`$ and the rest being $`0`$. Let $`K_{n}^{i}`$ denote the number of bits that the user $`i`$ transmits in a slot if it is scheduled. Then $`U_{n}^{i}`$ can be represented as $`U_{n}^{i} = {I_{n}^{i}\hspace{0pt}K_{n}^{i}}`$. Moreover, since a user can at most transmit all the bits in its buffer in a slot, $`K_{n}^{i} \leq Q_{n}^{i}`$. Since we assume that the slot length is normalized to unity, $`U_{n}^{i}`$ is the rate at which user $`i`$ transmits in slot $`n`$. Let $`\mathbf{U}_{n}`$ be the vector $`\lbrack U_{n}^{1},\ldots,U_{n}^{N}\rbrack`$, $`\mathbf{U}_{n} \in \mathcal{U}^{N}`$
The buffer evolution equation for user $`i`$ can be expressed as,
<table>
<tbody>
<tr>
<td></td>
<td>Qn+1isubscriptsuperscript𝑄𝑖𝑛1\displaystyle Q^{i}_{n+1}</td>
<td>=\displaystyle=</td>
<td>max{Qni−Uni,0}+An+1i.subscriptsuperscript𝑄𝑖𝑛subscriptsuperscript𝑈𝑖𝑛0subscriptsuperscript𝐴𝑖𝑛1\displaystyle\max\left\{Q^{i}_{n}-U^{i}_{n},0\right\}+A^{i}_{n+1}.</td>
<td></td>
<td>(2)</td>
</tr>
</tbody>
</table>
The buffer size $`B`$ is large as compared to the arrival rate, thus we can neglect the buffer overflow in the buffer evolution equation.
From \[[18](#bib.bib18)\], the power required for error-free or reliable communication at a rate $`K_{n}^{i} = u`$ bits/sec when $`X_{n}^{i} = x`$ is given by,
<table>
<tbody>
<tr>
<td></td>
<td>P(x,u)=WN0x(2uW−1).𝑃𝑥𝑢𝑊subscript𝑁0𝑥superscript2𝑢𝑊1P(x,u)=\frac{WN_{0}}{x}\big{(}2^{\frac{u}{W}}-1\big{)}.</td>
<td></td>
<td>(3)</td>
</tr>
</tbody>
</table>
where $`W`$ is the bandwidth in Hz. Note that for a given $`x`$, the transmission power $`P\hspace{0pt}{(x,u)}`$ is an increasing and strictly convex function of $`u`$. Let $`\hat{P}`$ denote the peak power constraint. Let $`{\hat{K}}_{n}^{i}`$ be the maximum rate at which user $`i`$ can transmit in a slot $`n`$ when the channel condition is $`X_{n}^{i}`$ while satisfying the peak power constraint (i.e., $`{P\hspace{0pt}{({\hat{K}}_{n}^{i},X_{n}^{i})}} \leq \hat{P}`$). Then the set of feasible rates for user $`i`$ in slot $`n`$, $`\mathcal{F}_{n}^{i}\overset{\Delta}{=}{\{ 0,\ldots,{\min{({\hat{K}}_{n}^{i},Q_{n}^{i})}}\}}`$.
We assume that the users specify their QoS requirements in terms of the average packet delay requirements. These delay requirements of the users are known a priori to the scheduler. By Little’s law \[[30](#bib.bib30)\], the average delay $`\overline{D}`$ is related to the average queue length $`\overline{Q}`$ as,
<table>
<tbody>
<tr>
<td></td>
<td>Q¯=a¯D¯,¯𝑄¯𝑎¯𝐷\bar{Q}=\bar{a}\bar{D},</td>
<td></td>
<td>(4)</td>
</tr>
</tbody>
</table>
where $`\overline{a}`$ is the average arrival rate. In the rest of the paper, we treat average delay as synonymous with average queue length and ignore the proportionality constant $`\overline{a}`$.
## III Problem Formulation
In this section, we formulate the problem as a constrained optimization problem within the Constrained Markov Decision Process (CMDP) framework.
###
III-A
Formulation as a Constrained Optimization Problem
The user devices have limited battery power, hence it is essential to design transmission policies that conserve battery power. The power-delay tradeoff can be exploited to save power at the expense of extra delay. Moreover, multi-user diversity can be exploited to schedule a user with better channel state. Such a user requires lesser power while transmitting at a certain rate as compared to when it has a poorer channel state. However, this also incurs additional delay. The objective is to design a joint rate allocation and scheduling scheme that minimizes the power expenditure of each user subject to the satisfaction of the individual delay constraints. The average power consumed by a user $`i`$ over a long period of time can be expressed as,
<table>
<tbody>
<tr>
<td></td>
<td>P¯isuperscript¯𝑃𝑖\displaystyle\bar{P}^{i}</td>
<td>=\displaystyle=</td>
<td>lim supM→∞1M𝐄∑n=1MP(Xni,IniKni).subscriptlimit-supremum→𝑀1𝑀𝐄superscriptsubscript𝑛1𝑀𝑃subscriptsuperscript𝑋𝑖𝑛subscriptsuperscript𝐼𝑖𝑛subscriptsuperscript𝐾𝑖𝑛\displaystyle\limsup_{M\rightarrow\infty}\frac{1}{M}\mathbf{E}\sum_{n=1}^{M}P(X^{i}_{n},I^{i}_{n}K^{i}_{n}).</td>
<td></td>
<td>(5)</td>
</tr>
</tbody>
</table>
The average queue length of a user $`i`$ over a long period of time can be expressed as,
<table>
<tbody>
<tr>
<td></td>
<td>Q¯isuperscript¯𝑄𝑖\displaystyle\bar{Q}^{i}</td>
<td>=\displaystyle=</td>
<td>lim supM→∞1M𝐄∑n=1MQni.subscriptlimit-supremum→𝑀1𝑀𝐄superscriptsubscript𝑛1𝑀subscriptsuperscript𝑄𝑖𝑛\displaystyle\limsup_{M\rightarrow\infty}\frac{1}{M}\mathbf{E}\sum_{n=1}^{M}Q^{i}_{n}.</td>
<td></td>
<td>(6)</td>
</tr>
</tbody>
</table>
Each user $`i`$ wants its average queue length to remain below a certain value, say, $`{\overline{\delta}}^{i}`$. Hence the problem becomes,
<table>
<tbody>
<tr>
<td></td>
<td>MinimizeP¯isubjecttoQ¯i≤δ¯i,i=1,…,N.formulae-sequenceMinimizesuperscript¯𝑃𝑖subjecttosuperscript¯𝑄𝑖superscript¯𝛿𝑖𝑖1…𝑁\mathrm{Minimize}\;\;\bar{P}^{i}\;\;\mathrm{subject\;to}\;\;\bar{Q}^{i}\leq\bar{\delta}^{i},\;\;i=1,\ldots,N.</td>
<td></td>
<td>(7)</td>
</tr>
</tbody>
</table>
######
Remark 1
Dependence between problems: Note that the $`N`$ problems formulated in ([7](#S3.E7 "In III-A Formulation as a Constrained Optimization Problem ‣ III Problem Formulation ‣ Power Efficient Scheduling under Delay Constraints over Multi-user Wireless Channels")) are not independent. This is because in a TDMA system, only one user can be scheduled in a slot. Consequently, the scheduling decision in a slot impacts the buffer occupancy of all the users in the future slots.
###
III-B
Notion of Optimal Solution
The problem in ([7](#S3.E7 "In III-A Formulation as a Constrained Optimization Problem ‣ III Problem Formulation ‣ Power Efficient Scheduling under Delay Constraints over Multi-user Wireless Channels")) is a multi-objective optimization problem with $`N`$ objectives and $`N`$ constraints. There can be multiple average power vectors that can be considered as optimal. Hence it is necessary to precisely define the properties of an optimal solution sought by us. We seek *Pareto optimal* solutions \[[31](#bib.bib31)\]. Let vector $`\lbrack{\overline{P}}_{\psi}^{1},\ldots,{\overline{P}}_{\psi}^{N}\rbrack`$ be the power expenditure vector under the rate allocation policy $`\psi`$. We say that the rate allocation policy $`\psi`$ is Pareto optimal if and only if there exists no rate allocation policy $`\zeta`$ with the corresponding power expenditure vector $`\lbrack{\overline{P}}_{\zeta}^{1},\ldots,{\overline{P}}_{\zeta}^{N}\rbrack`$ having the following properties,
<table>
<tbody>
<tr>
<td></td>
<td>∀i∈{1,…,N}Pζi≤Pψi∧∃i∈{1,…,N}|Pζi<Pψi.for-all𝑖1…𝑁subscriptsuperscript𝑃𝑖𝜁subscriptsuperscript𝑃𝑖𝜓𝑖conditional1…𝑁subscriptsuperscript𝑃𝑖𝜁subscriptsuperscript𝑃𝑖𝜓\displaystyle\forall i\in\{1,\ldots,N\}P^{i}_{\zeta}\leq P^{i}_{\psi}\;\;\land\;\;\exists i\in\{1,\ldots,N\}|P^{i}_{\zeta}<P^{i}_{\psi}.</td>
<td></td>
<td>(8)</td>
</tr>
</tbody>
</table>
The Pareto optimal solution is generally not unique and the set of Pareto optimal solutions is called the set of non-dominated solutions. The weighted sum approach is a common approach for solving a multi-objective optimization problem \[[31](#bib.bib31)\]. In this approach, one aggregates the $`N`$ objective functions into a single objective function. The resultant problem has a single objective function and $`N`$ constraints and can be expressed as,
<table>
<tbody>
<tr>
<td></td>
<td>MinimizeP¯Minimize¯𝑃\displaystyle\mathrm{Minimize}\;\;\bar{P}</td>
<td>=\displaystyle=</td>
<td>γ1P¯1+…+γNP¯N,superscript𝛾1superscript¯𝑃1…superscript𝛾𝑁superscript¯𝑃𝑁\displaystyle\gamma^{1}\bar{P}^{1}+\ldots+\gamma^{N}\bar{P}^{N},</td>
<td></td>
<td></td>
</tr>
<tr>
<td></td>
<td></td>
<td></td>
<td>subjectto,subjectto\displaystyle\mathrm{subject\;to},</td>
<td></td>
<td></td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td>Q¯isuperscript¯𝑄𝑖\displaystyle\bar{Q}^{i}</td>
<td>≤\displaystyle\leq</td>
<td>δ¯i,i=1,…,N.formulae-sequencesuperscript¯𝛿𝑖𝑖1…𝑁\displaystyle\bar{\delta}^{i},\;\;i=1,\ldots,N.</td>
<td></td>
<td>(9)</td>
</tr>
</tbody>
</table>
where $`{\mathbf{γ}}\overset{\Delta}{=}{\lbrack\gamma^{1},\ldots,\gamma^{N}\rbrack}`$ is the weight vector. It is generally assumed that $`{\gamma^{i} \in {{\lbrack 0,1\rbrack},{\forall i}}},{{\sum_{i = 1}^{N}\gamma^{i}} = 1}`$ implying that $`\overline{P}`$ is a convex combination of the individual powers. In general, the non-dominated set (i.e., the set of all Pareto optimal policies) may be a non-convex set. By varying the weight vector in the weighted sum approach we can determine the Pareto optimal policies within a convex subset of the non-dominated set. However, choosing the weight vector in order to obtain a particular solution is not straight forward. In the next section, we formulate the problem in ([](#S3.Ex2 "III-B Notion of Optimal Solution ‣ III Problem Formulation ‣ Power Efficient Scheduling under Delay Constraints over Multi-user Wireless Channels")
III-B
) within the CMDP framework.
###
III-C
The CMDP Framework
Let $`\mathbf{X}_{n}\overset{\Delta}{=}{\lbrack X_{n}^{1},\ldots,X_{n}^{N}\rbrack}`$ and $`\mathbf{Q}_{n}\overset{\Delta}{=}{\lbrack Q_{n}^{1},\ldots,Q_{n}^{N}\rbrack}`$. The state of the above system $`\mathbf{S}_{n}`$ at time $`n`$ can be described by the tuple, $`\mathbf{S}_{n}\overset{\Delta}{=}{\lbrack\mathbf{Q}_{n},\mathbf{X}_{n}\rbrack}`$, comprising of the queue length and the channel state of each of the $`N`$ users. Note that the system state space $`\mathcal{S} = {\mathcal{Q}^{N} \times \mathcal{X}^{N}}`$ is discrete and finite. Let $`\{\mathbf{S}_{n}\}`$ denote the state process. In each slot, the scheduler sets the rate vector vector $`\mathbf{U}_{n}\overset{\Delta}{=}{\lbrack U_{n}^{1},\ldots,U_{n}^{N}\rbrack}`$, where $`U_{n}^{i} = {I_{n}^{i}\hspace{0pt}K_{n}^{i}}`$. $`\mathbf{U}_{n}`$ takes values from the finite action space $`\mathcal{U}^{N}`$. $`\{\mathbf{U}_{n}\}`$ denotes the control process. This problem has the structure of a CMDP with finite state and action spaces. Since we are considering average power expended and average delay suffered, it is an average cost CMDP. The scheduler objective is to determine an optimal *rate allocation policy*, i.e., a mapping from past history of states and actions to a rate allocation vector $`\mathbf{U}_{n}`$ in every slot $`n`$. For a CMDP with finite state and action spaces, it is well known that an optimal stationary randomized policy exists \[[32](#bib.bib32)\], i.e., the rate allocation policy is a mapping from the current system state to a probability distribution on the set of rate allocation vectors. However, the traditional computational approaches based on Linear Programming \[[32](#bib.bib32)\] cannot be used to determine the optimal policy because of the following reasons:
1. 1\.
*Large state space:* The system state space is large even for very few users. We have already illustrated this with an example in Section [I](#S1 "I Introduction ‣ Power Efficient Scheduling under Delay Constraints over Multi-user Wireless Channels"). Moreover, the state space grows exponentially with number of users, hence the computational complexity of the traditional approaches also grows exponentially with number of users.
2. 2\.
*Unknown user/system statistics:* The probability distributions of $`X_{n}^{i}`$ and $`A_{n}^{i}`$ are unknown. The traditional approaches rely on the knowledge of these distributions for determining the optimal policy.
We now extend the approach suggested in \[[28](#bib.bib28)\] to determine an optimal policy for the problem in ([](#S3.Ex2 "III-B Notion of Optimal Solution ‣ III Problem Formulation ‣ Power Efficient Scheduling under Delay Constraints over Multi-user Wireless Channels")
III-B
).
Let user $`i`$ be scheduled in slot $`n`$. Then the state of the system immediately after user $`i`$ transmits the data can be represented as $`{\overset{\sim}{\mathbf{S}}}_{n} = {\lbrack{\overset{\sim}{\mathbf{Q}}}_{n},{\overset{\sim}{\mathbf{X}}}_{n}\rbrack} = {\lbrack{(q_{n}^{1},\ldots,{(\max{(0,q_{n}^{i} - u_{n}^{i})},\ldots,q_{n}^{N},x_{n}^{1},\ldots,x_{n}^{N}\rbrack}}}`$. Let $`\mathbf{A}_{n} = {\lbrack a_{n}^{1},\ldots,a_{n}^{N}\rbrack}`$ denote the arrival vector in slot $`n`$. The state of the system at the beginning of slot $`n + 1`$ can be represented as, $`\mathbf{S}_{n + 1} = {\lbrack\mathbf{Q}_{n + 1},\mathbf{X}_{n + 1}\rbrack} = {\lbrack{(q_{n}^{1} + a_{n + 1}^{1},\ldots,{(\max{(0,q_{n}^{i} - u_{n}^{i} + a_{n + 1}^{i})},\ldots,q_{n}^{N} + a_{n + 1}^{N},x_{n + 1}^{1},\ldots,x_{n + 1}^{N})}\rbrack}}`$. The queue transition equation in the vector form can be written as,
<table>
<tbody>
<tr>
<td></td>
<td>𝐐n+1=max(0,𝐐n+𝐀n+1−𝐔n).subscript𝐐𝑛10subscript𝐐𝑛subscript𝐀𝑛1subscript𝐔𝑛\mathbf{Q}_{n+1}=\max(0,\mathbf{Q}_{n}+\mathbf{A}_{n+1}-\mathbf{U}_{n}).</td>
<td></td>
<td>(10)</td>
</tr>
</tbody>
</table>
Let $`\mathbf{\Delta} = {\lbrack{\overline{\delta}}^{1},\ldots,{\overline{\delta}}^{N}\rbrack}`$ denote the delay constraint vector. The problem in ([](#S3.Ex2 "III-B Notion of Optimal Solution ‣ III Problem Formulation ‣ Power Efficient Scheduling under Delay Constraints over Multi-user Wireless Channels")
III-B
) can be converted into an unconstrained problem using the Lagrangian approach. The unconstrained problem can be expressed as,
<table>
<tbody>
<tr>
<td></td>
<td>MinimizeP¯+∑i=1Nλi(Q¯i−δ¯i).Minimize¯𝑃superscriptsubscript𝑖1𝑁superscript𝜆𝑖superscript¯𝑄𝑖superscript¯𝛿𝑖\mathrm{Minimize}\;\;\bar{P}+\sum_{i=1}^{N}\lambda^{i}(\bar{Q}^{i}-\bar{\delta}^{i}).</td>
<td></td>
<td>(11)</td>
</tr>
</tbody>
</table>
The immediate cost $`b\hspace{0pt}{( \cdot , \cdot , \cdot )}`$ incurred in scheduling a user $`i`$ in state $`\mathbf{S}_{n}`$ when the LM vector is $`\mathbf{λ}`$ and rate vector $`\mathbf{U}_{n}`$ (user $`i`$ is scheduled in slot $`n`$) can be expressed as,
<table>
<tbody>
<tr>
<td></td>
<td>b(𝝀,𝐒n,𝐔n)=P(xni,uni)+∑i=1Nλi(Qni−δ¯i)𝑏𝝀subscript𝐒𝑛subscript𝐔𝑛𝑃subscriptsuperscript𝑥𝑖𝑛subscriptsuperscript𝑢𝑖𝑛superscriptsubscript𝑖1𝑁superscript𝜆𝑖subscriptsuperscript𝑄𝑖𝑛superscript¯𝛿𝑖b(\mbox{\boldmath{$\lambda$}},\mathbf{S}_{n},\mathbf{U}_{n})=P(x^{i}_{n},u^{i}_{n})+\sum_{i=1}^{N}\lambda^{i}(Q^{i}_{n}-\bar{\delta}^{i})</td>
<td></td>
<td>(12)</td>
</tr>
</tbody>
</table>
Let $`{\overset{\sim}{\mathbf{s}}}^{0} = {\lbrack\mathbf{q}^{0},\mathbf{x}^{0}\rbrack}`$ denote a fixed state. Let $`\overset{\sim}{V}\hspace{0pt}{( \cdot )}`$ denote the dynamic programming value function based on the state $`\overset{\sim}{\mathbf{S}}`$ reached immediately after taking the scheduling decision but before the arrivals. Let $`\mathcal{F}_{n} = {\lbrack\mathcal{F}_{n}^{1},\ldots,\mathcal{F}_{n}^{N}\rbrack}`$ be the set of feasible
<sup>1</sup>
<sup>1</sup>1A rate is feasible based on the peak power constraints and the buffer occupancy
rate vectors in slot $`n`$. Let $`\{ f_{n}\}`$ and $`\{ e_{n}\}`$ be two sequences that have the following properties,
<table>
<tbody>
<tr>
<td></td>
<td>fn→0,en→0,∑n(fn)2<∞,∑n(en)2<∞,formulae-sequence→subscript𝑓𝑛0formulae-sequence→subscript𝑒𝑛0formulae-sequencesubscript𝑛superscriptsubscript𝑓𝑛2subscript𝑛superscriptsubscript𝑒𝑛2\displaystyle f_{n}\rightarrow 0,\quad e_{n}\rightarrow 0,\quad\sum_{n}(f_{n})^{2}<\infty,\quad\sum_{n}(e_{n})^{2}<\infty,</td>
<td></td>
<td>(13)</td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td>∑nfn=∞,∑nen=∞,formulae-sequencesubscript𝑛subscript𝑓𝑛subscript𝑛subscript𝑒𝑛\displaystyle\sum_{n}f_{n}=\infty,\quad\sum_{n}e_{n}=\infty,</td>
<td></td>
<td>(14)</td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td>∑n(fn2+en2)<∞,limn→∞enfn→0.formulae-sequencesubscript𝑛superscriptsubscript𝑓𝑛2superscriptsubscript𝑒𝑛2→subscript→𝑛subscript𝑒𝑛subscript𝑓𝑛0\displaystyle\quad\sum_{n}({f_{n}}^{2}+{e_{n}}^{2})<\infty,\quad\lim_{n\rightarrow\infty}\frac{e_{n}}{f_{n}}\to 0.</td>
<td></td>
<td>(15)</td>
</tr>
</tbody>
</table>
The significance of these properties is explained later. We now present an optimal online primal-dual algorithm for solving the constrained problem in ([](#S3.Ex2 "III-B Notion of Optimal Solution ‣ III Problem Formulation ‣ Power Efficient Scheduling under Delay Constraints over Multi-user Wireless Channels")
III-B
):
<table>
<tbody>
<tr>
<td></td>
<td>𝐔n+1subscript𝐔𝑛1\displaystyle\mathbf{U}_{n+1}</td>
<td>=\displaystyle=</td>
<td>argmin𝐕∈ℱn{(1−fn)V~(𝐒~n)+fn×\displaystyle\arg\min_{\mathbf{V}\in\mathcal{F}_{n}}\Big{\{}(1-f_{n})\tilde{V}(\mathbf{\tilde{S}}_{n})+f_{n}\times</td>
<td></td>
<td rowspan="4">(16)</td>
</tr>
<tr>
<td></td>
<td></td>
<td></td>
<td>{b(𝝀,(𝐐~n+𝐀n+1,𝐗n+1),𝐕)\displaystyle\big{\{}b(\mbox{\boldmath{$\lambda$}},(\mathbf{\tilde{Q}}_{n}+\mathbf{A}_{n+1},\mathbf{X}_{n+1}),\mathbf{V})</td>
<td></td>
</tr>
<tr>
<td></td>
<td></td>
<td></td>
<td>+V~n((𝐐~n+𝐀n+1−𝐕,𝐗n+1))subscript~𝑉𝑛subscript~𝐐𝑛subscript𝐀𝑛1𝐕subscript𝐗𝑛1\displaystyle+\tilde{V}_{n}((\mathbf{\tilde{Q}}_{n}+\mathbf{A}_{n+1}-\mathbf{V},\mathbf{X}_{n+1}))</td>
<td></td>
</tr>
<tr>
<td></td>
<td></td>
<td></td>
<td>−V~n(𝐬~0)}},\displaystyle-\tilde{V}_{n}(\mathbf{\tilde{s}}^{0})\big{\}}\Big{\}},</td>
<td></td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td>V~n+1(𝐒~n)subscript~𝑉𝑛1subscript~𝐒𝑛\displaystyle\tilde{V}_{n+1}(\mathbf{\tilde{S}}_{n})</td>
<td>=\displaystyle=</td>
<td>(1−fn)V~n(𝐒~n)+fn×\displaystyle(1-f_{n})\tilde{V}_{n}(\mathbf{\tilde{S}}_{n})+f_{n}\times</td>
<td></td>
<td rowspan="4">(17)</td>
</tr>
<tr>
<td></td>
<td></td>
<td></td>
<td>{b(𝝀,(𝐐~n+𝐀n+1,𝐗n+1),𝐔n+1)\displaystyle\Big{\{}b(\mbox{\boldmath{$\lambda$}},(\mathbf{\tilde{Q}}_{n}+\mathbf{A}_{n+1},\mathbf{X}_{n+1}),\mathbf{U}_{n+1})</td>
<td></td>
</tr>
<tr>
<td></td>
<td></td>
<td></td>
<td>+V~n(𝐐~n+𝐀n+1−𝐔n+1,𝐗n+1)subscript~𝑉𝑛subscript~𝐐𝑛subscript𝐀𝑛1subscript𝐔𝑛1subscript𝐗𝑛1\displaystyle+\tilde{V}_{n}(\mathbf{\tilde{Q}}_{n}+\mathbf{A}_{n+1}-\mathbf{U}_{n+1},\mathbf{X}_{n+1})</td>
<td></td>
</tr>
<tr>
<td></td>
<td></td>
<td></td>
<td>−V~n(𝐬~0)}.\displaystyle-\tilde{V}_{n}(\mathbf{\tilde{s}}^{0})\Big{\}}.</td>
<td></td>
</tr>
</tbody>
</table>
The algorithm in ([16](#S3.E16 "In III-C The CMDP Framework ‣ III Problem Formulation ‣ Power Efficient Scheduling under Delay Constraints over Multi-user Wireless Channels")) and ([17](#S3.E17 "In III-C The CMDP Framework ‣ III Problem Formulation ‣ Power Efficient Scheduling under Delay Constraints over Multi-user Wireless Channels")) is an online version of the well known Relative Value Iteration Algorithm (RVIA) \[[33](#bib.bib33)\]. It iteratively determines the optimal value function and hence the optimal policy one state at a time for a fixed value of the LM vector $`\mathbf{λ}`$. To determine the optimal LM vector we augment the above algorithm with a dual LM iteration:
<table>
<tbody>
<tr>
<td></td>
<td>𝝀n+1=Λ[𝝀n+en(𝐐n−𝚫)],subscript𝝀𝑛1Λdelimited-[]subscript𝝀𝑛subscript𝑒𝑛subscript𝐐𝑛𝚫\mbox{\boldmath{$\lambda$}}_{n+1}=\Lambda[\mbox{\boldmath{$\lambda$}}_{n}+e_{n}\left(\mathbf{Q}_{n}-\mbox{\boldmath{$\Delta$}}\right)],</td>
<td></td>
<td>(18)</td>
</tr>
</tbody>
</table>
where $`\Lambda`$ is a projection operator for ensuring that the LMs are non-negative and finite. The properties of the update sequences in ([14](#S3.E14 "In III-C The CMDP Framework ‣ III Problem Formulation ‣ Power Efficient Scheduling under Delay Constraints over Multi-user Wireless Channels")) ensure that the sequences $`\{ f_{n}\}`$ and $`\{ e_{n}\}`$ converge to $`0`$ sufficiently fast to eliminate the noise effects when the iterates are close to their optimal values $`\overset{\sim}{V}\hspace{0pt}{( \cdot , \cdot )}`$ and $`{\mathbf{λ}}^{\ast}`$, while those in ([13](#S3.E13 "In III-C The CMDP Framework ‣ III Problem Formulation ‣ Power Efficient Scheduling under Delay Constraints over Multi-user Wireless Channels")) ensure that they do not approach $`0`$ too rapidly to avoid convergence of the algorithm to non-optimal values. Furthermore, ([15](#S3.E15 "In III-C The CMDP Framework ‣ III Problem Formulation ‣ Power Efficient Scheduling under Delay Constraints over Multi-user Wireless Channels")) ensures that the update rates of primal iterations, i.e., the value function iterations and the dual iterations, i.e., the LM iterations are different. Since $`e_{n}`$ approaches $`0`$ much faster than $`f_{n}`$, the update rate of the value function iterations is much higher than the update rate of the LM iterations. This ensures that even though both the primals and duals are updated simultaneously, both converge to their optimal values \[[34](#bib.bib34)\]. ([16](#S3.E16 "In III-C The CMDP Framework ‣ III Problem Formulation ‣ Power Efficient Scheduling under Delay Constraints over Multi-user Wireless Channels")), ([17](#S3.E17 "In III-C The CMDP Framework ‣ III Problem Formulation ‣ Power Efficient Scheduling under Delay Constraints over Multi-user Wireless Channels")) and ([18](#S3.E18 "In III-C The CMDP Framework ‣ III Problem Formulation ‣ Power Efficient Scheduling under Delay Constraints over Multi-user Wireless Channels")) constitute the optimal algorithm. The proof of optimality is exactly similar to that in \[[28](#bib.bib28)\].
However, compared to the single user case, the state space here is too large for the algorithm to converge in reasonable number of iterations. We therefore motivate an alternate approach. We incorporate the possibility of transmitter induced errors in the single user scenario. We then motivate the multi-user solution by making use of this extension to the single user scenario.
## IV Single User Scenario in Presence of Transmitter Errors
We describe the scenario in brief. Consider a point to point transmission system over a fading wireless channel. Time is divided into slots of unit duration. We consider the block fading model as described in Section [II](#S2 "II System Model ‣ Power Efficient Scheduling under Delay Constraints over Multi-user Wireless Channels"). The scheduler is unaware of the probability distribution of the arrivals and the channel state in each slot. The objective is to minimize the average transmission power subject to average packet delay constraints. In \[[28](#bib.bib28)\], we have determined an online algorithm that determines the optimal transmission rate in each slot so as to minimize the average power expenditure subject to average packet delay constraints. We consider the following extension to the above problem: suppose that after the online algorithm has determined the rate $`U_{n} \in \mathcal{F}_{n}`$, with a certain unknown random probability $`\theta_{n} \in {\lbrack 0,1\rbrack}`$, the transmitter is unable to proceed with the transmission. We assume that the probability distribution of $`\theta_{n}`$ is not known. Under this assumption, the queue evolution equation can be expressed as,
<table>
<tbody>
<tr>
<td></td>
<td>Qn+1=Qn+An+1−InUn,subscript𝑄𝑛1subscript𝑄𝑛subscript𝐴𝑛1subscript𝐼𝑛subscript𝑈𝑛Q_{n+1}=Q_{n}+A_{n+1}-I_{n}U_{n},</td>
<td></td>
<td>(19)</td>
</tr>
</tbody>
</table>
where $`I_{n}`$ is an indicator variable that is set to $`1`$ if the transmitter is successful in transmitting the packets and is set to $`0`$ otherwise. We now formulate the rate allocation problem for this scenario. The long term power expenditure can be expressed as,
<table>
<tbody>
<tr>
<td></td>
<td>P¯esubscript¯𝑃𝑒\displaystyle\bar{P}_{e}</td>
<td>=\displaystyle=</td>
<td>lim supM→∞1M𝐄∑n=1MP(Xn,InUn)subscriptlimit-supremum→𝑀1𝑀𝐄superscriptsubscript𝑛1𝑀𝑃subscript𝑋𝑛subscript𝐼𝑛subscript𝑈𝑛\displaystyle\limsup_{M\rightarrow\infty}\frac{1}{M}\mathbf{E}\sum_{n=1}^{M}P(X_{n},I_{n}U_{n})</td>
<td></td>
<td>(20)</td>
</tr>
</tbody>
</table>
The average queue length over a long period of time can be expressed as,
<table>
<tbody>
<tr>
<td></td>
<td>Q¯esubscript¯𝑄𝑒\displaystyle\bar{Q}_{e}</td>
<td>=\displaystyle=</td>
<td>lim supM→∞1M𝐄∑n=1MQn,subscriptlimit-supremum→𝑀1𝑀𝐄superscriptsubscript𝑛1𝑀subscript𝑄𝑛\displaystyle\limsup_{M\rightarrow\infty}\frac{1}{M}\mathbf{E}\sum_{n=1}^{M}Q_{n},</td>
<td></td>
<td>(21)</td>
</tr>
</tbody>
</table>
Hence the rate allocation problem can be stated as,
<table>
<tbody>
<tr>
<td></td>
<td>MinimizeP¯esubjecttoQ¯e≤δ¯.Minimizesubscript¯𝑃𝑒subjecttosubscript¯𝑄𝑒¯𝛿\mathrm{Minimize}\;\;\bar{P}_{e}\;\;\mathrm{subject\;to}\;\;\bar{Q}_{e}\leq\bar{\delta}.</td>
<td></td>
<td>(22)</td>
</tr>
</tbody>
</table>
Note that the problem in ([22](#S4.E22 "In IV Single User Scenario in Presence of Transmitter Errors ‣ Power Efficient Scheduling under Delay Constraints over Multi-user Wireless Channels")) has the structure of a CMDP with a state space for the single user case and average cost criterion. The objective is to determine an optimal policy $`\mu^{\ast}`$ such that the power expended under this policy is minimum possible while satisfying the delay constraint.
###
IV-A
The Primal Dual Approach
The constrained problem in ([22](#S4.E22 "In IV Single User Scenario in Presence of Transmitter Errors ‣ Power Efficient Scheduling under Delay Constraints over Multi-user Wireless Channels")) can be converted into an unconstrained problem using the Lagrangian approach \[[32](#bib.bib32)\]. Let $`\lambda \geq 0`$ be a real number called as the Lagrange Multiplier (LM). Let $`\mathcal{B}`$ be the set $`\{ 0,1\}`$. $`c:{{\mathcal{R}^{+} \times \mathcal{Q} \times \mathcal{X} \times \mathcal{B} \times \mathcal{U}}\rightarrow\mathcal{R}}`$ be defined as the following,
<table>
<tbody>
<tr>
<td></td>
<td>c(λ,Qn,Xn,In,Un)=ΔP(Xn,InUn)+λ(Qn−δ¯),superscriptΔ𝑐𝜆subscript𝑄𝑛subscript𝑋𝑛subscript𝐼𝑛subscript𝑈𝑛𝑃subscript𝑋𝑛subscript𝐼𝑛subscript𝑈𝑛𝜆subscript𝑄𝑛¯𝛿\displaystyle c(\lambda,Q_{n},X_{n},I_{n},U_{n})\stackrel{{\scriptstyle\Delta}}{{=}}P(X_{n},I_{n}U_{n})+\lambda(Q_{n}-\bar{\delta}),</td>
<td></td>
<td>(23)</td>
</tr>
</tbody>
</table>
where $`U_{n}`$ is determined using the rate allocation policy $`\mu:{{\mathcal{Q} \times \mathcal{X}}\rightarrow\mathcal{U}}`$. The unconstrained problem is to minimize,
<table>
<tbody>
<tr>
<td></td>
<td>L(μ,λ)=lim supM→∞1M∑n=1Mc(λ,Qn,Xn,In,μ(Qn,Xn)).𝐿𝜇𝜆subscriptlimit-supremum→𝑀1𝑀superscriptsubscript𝑛1𝑀𝑐𝜆subscript𝑄𝑛subscript𝑋𝑛subscript𝐼𝑛𝜇subscript𝑄𝑛subscript𝑋𝑛L(\mu,\lambda)=\limsup_{M\rightarrow\infty}\frac{1}{M}\sum_{n=1}^{M}c(\lambda,Q_{n},X_{n},I_{n},\mu(Q_{n},X_{n})).</td>
<td></td>
<td>(24)</td>
</tr>
</tbody>
</table>
$`L\hspace{0pt}{( \cdot , \cdot )}`$ is called the Lagrangian. Our objective is to determine the optimal rate allocation policy $`\mu^{\ast}`$ and optimal LM $`\lambda^{\ast}`$ such that the following saddle point optimality condition is satisfied,
<table>
<tbody>
<tr>
<td></td>
<td>L(μ∗,λ)≤L(μ∗,λ∗)≤L(μ,λ∗).𝐿superscript𝜇𝜆𝐿superscript𝜇superscript𝜆𝐿𝜇superscript𝜆L(\mu^{*},\lambda)\leq L(\mu^{*},\lambda^{*})\leq L(\mu,\lambda^{*}).</td>
<td></td>
<td>(25)</td>
</tr>
</tbody>
</table>
For a fixed LM $`\lambda`$, the problem is an unconstrained Markov Decision Problem (MDP) with finite state and action spaces with the average cost criterion. The following dynamic programming equation \[[35](#bib.bib35)\] gives a necessary condition for optimality of a solution,
<table>
<tbody>
<tr>
<td></td>
<td>V(q,x)=minr∈ℱ[c(λ,q,x,I,r)−β+\displaystyle V(q,x)=\min_{r\in\mathcal{F}}\Big{[}c(\lambda,q,x,I,r)-\beta+</td>
<td></td>
<td></td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td>∑a′,x′p((q,x),r,(q+a′−r,x′))×\displaystyle\sum_{{a}^{\prime},{x}^{\prime}}p((q,x),r,(q+{a}^{\prime}-r,{x}^{\prime}))\times</td>
<td></td>
<td></td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td>V(q+a′−r,x′)],a′∈𝒜,x′∈𝒳,\displaystyle V(q+{a}^{\prime}-r,{x}^{\prime})\Big{]},{a}^{\prime}\in\mathcal{A},\;{x}^{\prime}\in\mathcal{X},</td>
<td></td>
<td>(26)</td>
</tr>
</tbody>
</table>
where $`V\hspace{0pt}{( \cdot , \cdot )}`$ is the value function, $`\beta \in \mathcal{R}`$ is the unique optimal power expenditure. Let $`{(q^{0},x^{0})} \in {\mathcal{Q} \times \mathcal{X}}`$ be a fixed state. If we impose $`{V\hspace{0pt}{(q^{0},x^{0})}} = 0`$, then $`V\hspace{0pt}{( \cdot , \cdot )}`$ is unique \[[35](#bib.bib35)\]. $`p\hspace{0pt}{(s,r,s')}`$ is the probability of reaching a state $`s'`$ upon taking an action $`r`$ in state $`s`$. The traditional approaches for computing the optimal policy for an unconstrained average cost MDP such as the Relative Value Iteration Algorithm \[[35](#bib.bib35)\] require the knowledge of $`p\hspace{0pt}{( \cdot , \cdot , \cdot )}`$ which in this case is dependent on the probability distributions of the arrivals and channel states which is not known. Note that determining the optimal value function as defined in ([](#S4.Ex9 "IV-A The Primal Dual Approach ‣ IV Single User Scenario in Presence of Transmitter Errors ‣ Power Efficient Scheduling under Delay Constraints over Multi-user Wireless Channels")
IV-A
) is not sufficient because the unconstrained solution for a particular $`\lambda`$ does not ensure that the constraints would be satisfied. To ensure constraint satisfaction, the optimal LM needs to be determined.
###
IV-B
The Online Rate Allocation Algorithm
We now present the rate allocation algorithm. Let the user state at the beginning of slot $`n`$ be $`{(Q_{n},X_{n})} = {(q,x)}`$. Suppose that $`u`$ bits are transmitted in slot $`n`$. The following primal-dual algorithm can be used to compute the rate $`U_{n + 1} = r_{n + 1}`$ at which the transmitter should transmit in slot $`n + 1`$,
<table>
<tbody>
<tr>
<td></td>
<td>rn+1subscript𝑟𝑛1\displaystyle r_{n+1}</td>
<td>=\displaystyle=</td>
<td>argminv∈ℱn+1{(1−fn)V~n(q~,x~)+fn×\displaystyle\arg\min_{v\in\mathcal{F}_{n+1}}\Big{\{}(1-f_{n})\tilde{V}_{n}(\tilde{q},\tilde{x})+f_{n}\times</td>
<td></td>
<td rowspan="4">(27)</td>
</tr>
<tr>
<td></td>
<td></td>
<td></td>
<td>{c(λn,q~+An+1,Xn+1,1,v)\displaystyle\big{\{}c(\lambda_{n},\tilde{q}+A_{n+1},X_{n+1},1,v)</td>
<td></td>
</tr>
<tr>
<td></td>
<td></td>
<td></td>
<td>+V~n(q~+An+1−v,Xn+1)subscript~𝑉𝑛~𝑞subscript𝐴𝑛1𝑣subscript𝑋𝑛1\displaystyle+\tilde{V}_{n}(\tilde{q}+A_{n+1}-v,X_{n+1})</td>
<td></td>
</tr>
<tr>
<td></td>
<td></td>
<td></td>
<td>−V~n(q~0,x~0)}},\displaystyle-\tilde{V}_{n}(\tilde{q}^{0},\tilde{x}^{0})\big{\}}\Big{\}},</td>
<td></td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td>V~n+1(q~,x~)subscript~𝑉𝑛1~𝑞~𝑥\displaystyle\tilde{V}_{n+1}(\tilde{q},\tilde{x})</td>
<td>=\displaystyle=</td>
<td>(1−fn)V~n(q~,x~)+fn×\displaystyle(1-f_{n})\tilde{V}_{n}(\tilde{q},\tilde{x})+f_{n}\times</td>
<td></td>
<td rowspan="4">(28)</td>
</tr>
<tr>
<td></td>
<td></td>
<td></td>
<td>{c(λn,q~+An+1,Xn+1,In+1,rn+1)\displaystyle\Big{\{}c(\lambda_{n},\tilde{q}+A_{n+1},X_{n+1},I_{n+1},r_{n+1})</td>
<td></td>
</tr>
<tr>
<td></td>
<td></td>
<td></td>
<td>+V~n(q~+An+1−In+1rn+1,Xn+1)−limit-fromsubscript~𝑉𝑛~𝑞subscript𝐴𝑛1subscript𝐼𝑛1subscript𝑟𝑛1subscript𝑋𝑛1\displaystyle+\tilde{V}_{n}(\tilde{q}+A_{n+1}-I_{n+1}r_{n+1},X_{n+1})-</td>
<td></td>
</tr>
<tr>
<td></td>
<td></td>
<td></td>
<td>V~n(q~0,x~0)},\displaystyle\tilde{V}_{n}(\tilde{q}^{0},\tilde{x}^{0})\Big{\}},</td>
<td></td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td>λn+1subscript𝜆𝑛1\displaystyle\lambda_{n+1}</td>
<td>=\displaystyle=</td>
<td>Λ[λn+en(Qn−δ¯)].Λdelimited-[]subscript𝜆𝑛subscript𝑒𝑛subscript𝑄𝑛¯𝛿\displaystyle\Lambda[\lambda_{n}+e_{n}\left(Q_{n}-\bar{\delta}\right)].</td>
<td></td>
<td>(29)</td>
</tr>
</tbody>
</table>
These equations are explained below:
1. 1\.
([27](#S4.E27 "In IV-B The Online Rate Allocation Algorithm ‣ IV Single User Scenario in Presence of Transmitter Errors ‣ Power Efficient Scheduling under Delay Constraints over Multi-user Wireless Channels")), ([28](#S4.E28 "In IV-B The Online Rate Allocation Algorithm ‣ IV Single User Scenario in Presence of Transmitter Errors ‣ Power Efficient Scheduling under Delay Constraints over Multi-user Wireless Channels")) and ([29](#S4.E29 "In IV-B The Online Rate Allocation Algorithm ‣ IV Single User Scenario in Presence of Transmitter Errors ‣ Power Efficient Scheduling under Delay Constraints over Multi-user Wireless Channels")) constitute the rate allocation algorithm. It consists of two phases: *rate determination phase* and *update phase*. ([27](#S4.E27 "In IV-B The Online Rate Allocation Algorithm ‣ IV Single User Scenario in Presence of Transmitter Errors ‣ Power Efficient Scheduling under Delay Constraints over Multi-user Wireless Channels")) constitutes the rate determination phase of the algorithm, i.e., it is used to determine the rate at which a user transmits in a slot *if the transmission is successful*. ([28](#S4.E28 "In IV-B The Online Rate Allocation Algorithm ‣ IV Single User Scenario in Presence of Transmitter Errors ‣ Power Efficient Scheduling under Delay Constraints over Multi-user Wireless Channels")) is the primal iteration to determine the optimal value function and thereby the optimal policy, while ([29](#S4.E29 "In IV-B The Online Rate Allocation Algorithm ‣ IV Single User Scenario in Presence of Transmitter Errors ‣ Power Efficient Scheduling under Delay Constraints over Multi-user Wireless Channels")) is the coupled dual iteration for determining the optimal LM. They constitute the update phase of the algorithm.
2. 2\.
If in a state $`{(Q_{n},X_{n})} = {(q,x)}`$, the transmitter decides to transmit $`u \leq q`$ bits, then $`\overset{\sim}{q}\overset{\Delta}{=}{q - u}`$, and $`\overset{\sim}{x}\overset{\Delta}{=}x`$.
3. 3\.
([28](#S4.E28 "In IV-B The Online Rate Allocation Algorithm ‣ IV Single User Scenario in Presence of Transmitter Errors ‣ Power Efficient Scheduling under Delay Constraints over Multi-user Wireless Channels")) determines the optimal value function based on this new virtual state $`(\overset{\sim}{q},\overset{\sim}{x})`$. Note that the value function for this new state is related to the usual value function as $`{{\overset{\sim}{V}}_{n}\hspace{0pt}{(\overset{\sim}{q},\overset{\sim}{x})}} = {\mathbf{E}^{A,X}\hspace{0pt}{\lbrack{V_{n}\hspace{0pt}{(Q,X)}}\rbrack}}`$.
4. 4\.
The rate determination phase ([27](#S4.E27 "In IV-B The Online Rate Allocation Algorithm ‣ IV Single User Scenario in Presence of Transmitter Errors ‣ Power Efficient Scheduling under Delay Constraints over Multi-user Wireless Channels")) determines the rate assuming that the transmitter would be successful in transmitting in slot $`n + 1`$ ($`I_{n + 1}`$ is assumed to be equal to $`1`$). However in ([28](#S4.E28 "In IV-B The Online Rate Allocation Algorithm ‣ IV Single User Scenario in Presence of Transmitter Errors ‣ Power Efficient Scheduling under Delay Constraints over Multi-user Wireless Channels")), updating the value function requires the knowledge of whether the transmission is successful or not. This because the immediate cost function $`c\hspace{0pt}{( \cdot , \cdot , \cdot , \cdot , \cdot )}`$ depends on $`I_{n + 1}`$, i.e., whether the transmission is successful or not. Thus the update phase updates the value function and LM in each slot based on the success of the transmission.
5. 5\.
$`({\overset{\sim}{q}}^{0},{\overset{\sim}{x}}^{0})`$ is any pre-designated state. On the RHS in ([28](#S4.E28 "In IV-B The Online Rate Allocation Algorithm ‣ IV Single User Scenario in Presence of Transmitter Errors ‣ Power Efficient Scheduling under Delay Constraints over Multi-user Wireless Channels")), the value function corresponding to this state is subtracted in order to keep the iterates bounded.
6. 6\.
The LM iteration in ([29](#S4.E29 "In IV-B The Online Rate Allocation Algorithm ‣ IV Single User Scenario in Presence of Transmitter Errors ‣ Power Efficient Scheduling under Delay Constraints over Multi-user Wireless Channels")) ensures that the specified delay constraint is satisfied.
7. 7\.
The sequences $`f_{n}`$ and $`e_{n}`$ have properties specified in ([13](#S3.E13 "In III-C The CMDP Framework ‣ III Problem Formulation ‣ Power Efficient Scheduling under Delay Constraints over Multi-user Wireless Channels")), ([14](#S3.E14 "In III-C The CMDP Framework ‣ III Problem Formulation ‣ Power Efficient Scheduling under Delay Constraints over Multi-user Wireless Channels")) and ([15](#S3.E15 "In III-C The CMDP Framework ‣ III Problem Formulation ‣ Power Efficient Scheduling under Delay Constraints over Multi-user Wireless Channels")). The reasons for imposing these properties have been explained in Section [](#S3.SS3 "III-C The CMDP Framework ‣ III Problem Formulation ‣ Power Efficient Scheduling under Delay Constraints over Multi-user Wireless Channels")
III-C
.
###
IV-C
Proof of Convergence
######
Theorem 1
For the rate determination algorithm ([27](#S4.E27 "In IV-B The Online Rate Allocation Algorithm ‣ IV Single User Scenario in Presence of Transmitter Errors ‣ Power Efficient Scheduling under Delay Constraints over Multi-user Wireless Channels")), ([28](#S4.E28 "In IV-B The Online Rate Allocation Algorithm ‣ IV Single User Scenario in Presence of Transmitter Errors ‣ Power Efficient Scheduling under Delay Constraints over Multi-user Wireless Channels")) and ([29](#S4.E29 "In IV-B The Online Rate Allocation Algorithm ‣ IV Single User Scenario in Presence of Transmitter Errors ‣ Power Efficient Scheduling under Delay Constraints over Multi-user Wireless Channels")), the iterates $`{(V_{n},\lambda_{n})}\rightarrow{(V,\lambda^{\ast})}`$.
###### Proof:
The proof of convergence is exactly similar to that in \[[28](#bib.bib28)\]. The probability of transmission failure in each slot serves as an extra *noise* term. The algorithm being a stochastic approximation based online algorithm, averages out this extra noise term and determines the optimal policy and the optimal LM. ∎
## V An Online Primal Dual Algorithm For the Multi-user Problem
In this section, we propose a suboptimal approach to solve the problem in ([7](#S3.E7 "In III-A Formulation as a Constrained Optimization Problem ‣ III Problem Formulation ‣ Power Efficient Scheduling under Delay Constraints over Multi-user Wireless Channels")). To avoid the state space explosion, in the proposed algorithm, we determine the rate $`R_{n}^{i} \in \mathcal{F}_{n}^{i}`$ for a user $`i`$ in a slot $`n`$, *if it is scheduled*, based on its *state* $`S_{n}^{i}\overset{\Delta}{=}{\lbrack Q_{n}^{i},X_{n}^{i}\rbrack}`$ *alone* instead of the entire system state $`\mathbf{S}_{n} = {\lbrack\mathbf{Q}_{n},\mathbf{X}_{n}\rbrack}`$. Note that $`S_{n}^{i} \in {\mathcal{Q} \times \mathcal{X}}`$. The rate $`R_{n}^{i}`$ is determined using a rate allocation policy $`\rho^{i}`$, i.e., a mapping from the history of states and rate allocations for user $`i`$ to its transmission rate. Once the rate $`R_{n}^{i}`$ for each user $`i`$ is determined, the next task is to determine a user to be scheduled in that slot. The user selection policy $`\kappa`$ is a mapping, $`\kappa:{{\mathcal{F}_{n}^{1} \times \ldots \times \mathcal{F}_{n}^{N}}\rightarrow\mathcal{I}}`$.
###
V-A
Rate Allocation Algorithm for a User
The rate allocation algorithm for each user behaves as if it were controlled by a single user policy in the presence of transmitter errors as explained in Section [IV](#S4 "IV Single User Scenario in Presence of Transmitter Errors ‣ Power Efficient Scheduling under Delay Constraints over Multi-user Wireless Channels"). Each user $`i`$ determines the rate $`R_{n + 1}^{i}`$ at which it would transmit in slot $`n + 1`$ if it were to be scheduled in slot $`n + 1`$ and informs this rate to the base station. The base station uses the user selection algorithm to schedule a user. The users who are not scheduled in a slot update their value function assuming transmitter errors, while the user who is scheduled updates its value function assuming successful transmission.
######
Remark 2
In the case of the single user scenario with transmitter errors, the probability with which a transmission is unsuccessful is independent of the scheduler action, i.e., the transmission rate determined by the online algorithm. In the multi-user scenario, this independence does not hold. This makes the problem a multi-agent learning problem \[[36](#bib.bib36)\], \[[37](#bib.bib37)\] where each agent (user) attempts to learn the optimal strategy and the actions taken by an agent (a user) influences the actions taken by the other agents (users).
###
V-B
User Selection Algorithm
The user selection algorithm is simple: select the user with the largest $`R_{n}^{i}`$, i.e., select the user with the best rate. The intuition behind this is the following. The rate allocation algorithm of a user $`i`$ would direct it to transmit at a high rate $`R_{n}^{i}`$ under two circumstances: either the channel condition for that user is very good, in which case, transmission at high rate saves power, or the delay constraints of that user are not being satisfied. Thus selecting a user with a high rate results in either power savings or the user delay constraint being satisfied.
[Refer to caption](/html/0710.1190/assets/x2.png)
Figure 2: Solution schematic
###
V-C
Algorithm Details and Implementation
The rate allocation algorithm is implemented on the user devices while the user selection algorithm is implemented at the base station, as illustrated in Figure [2](#S5.F2 "Figure 2 ‣ V-B User Selection Algorithm ‣ V An Online Primal Dual Algorithm For the Multi-user Problem ‣ Power Efficient Scheduling under Delay Constraints over Multi-user Wireless Channels"). From ([27](#S4.E27 "In IV-B The Online Rate Allocation Algorithm ‣ IV Single User Scenario in Presence of Transmitter Errors ‣ Power Efficient Scheduling under Delay Constraints over Multi-user Wireless Channels")) note that the rate determination phase requires $`X_{n}^{i}`$, i.e., the knowledge of the channel state at the base station. The communication overhead incurred by the base station in informing a user the channel state perceived by it depends on the number of states used to represent the channel. We represent the channel using $`8`$ states. Hence the base station needs $`3`$ bits per slot in order to inform a user the channel state perceived by it. The users inform the base station the rate at which they would transmit if they were to be scheduled. We allocate $`3`$ bits for conveying this information, i.e., the system can employ $`8`$ rates. The user selection algorithm then determines the user to be scheduled and all the users are informed about this decision. The rate allocation algorithm at each user then enters the update phase where the value function and the LM for each user are appropriately updated using ([28](#S4.E28 "In IV-B The Online Rate Allocation Algorithm ‣ IV Single User Scenario in Presence of Transmitter Errors ‣ Power Efficient Scheduling under Delay Constraints over Multi-user Wireless Channels")) and ([29](#S4.E29 "In IV-B The Online Rate Allocation Algorithm ‣ IV Single User Scenario in Presence of Transmitter Errors ‣ Power Efficient Scheduling under Delay Constraints over Multi-user Wireless Channels")). The algorithm thus continues in each slot $`n`$. The rate allocation algorithm that is executed at each user device is illustrated in Algorithm [1](#alg1 "Algorithm 1 ‣ V-C Algorithm Details and Implementation ‣ V An Online Primal Dual Algorithm For the Multi-user Problem ‣ Power Efficient Scheduling under Delay Constraints over Multi-user Wireless Channels") where, steps 4-8 represent the rate determination phase, while steps 10-14 represent the update phase. The user selection algorithm executed at the base station is detailed in Algorithm [2](#alg2 "Algorithm 2 ‣ V-C Algorithm Details and Implementation ‣ V An Online Primal Dual Algorithm For the Multi-user Problem ‣ Power Efficient Scheduling under Delay Constraints over Multi-user Wireless Channels").
Algorithm 1
The Rate Allocation Algorithm at the User $`i`$ Device
1:
Initialize the value function matrix $`{{{\overset{\sim}{V}}^{i}\hspace{0pt}{(q,x)}}\leftarrow{0\hspace{0pt}{\forall q}} \in \mathcal{Q}},{x \in \mathcal{X}}`$
2:
Initialize LM $`\lambda^{i}\leftarrow 0`$
3:
Initialize slot counter $`n\leftarrow 1`$
4:
Initialize queue length $`q^{i}\leftarrow 0`$
5:
Initialize channel states $`{x^{i}\leftarrow 0},{x_{}^{i'}\leftarrow 0}`$
6:
Reference state $`{\overset{\sim}{s}}^{i,0} = {(0,x^{1})}`$, where $`x^{1} \in \mathcal{X}`$
7:
while TRUE do
8:
while Base station has not informed the channel state $`{}x_{}^{i'}`$ do
9:
wait
10:
end while
11:
Determine the number of arrivals $`A_{n + 1}^{i} = a^{i}`$ in the current slot
12:
Determine the queue length in the current slot $`Q_{n}^{i} = q^{i}`$
13:
Use the rate determination phase of the rate allocation algorithm, i.e., ([27](#S4.E27 "In IV-B The Online Rate Allocation Algorithm ‣ IV Single User Scenario in Presence of Transmitter Errors ‣ Power Efficient Scheduling under Delay Constraints over Multi-user Wireless Channels")) to determine the rate $`r^{i}`$, for transmission
14:
Determine the power $`P\hspace{0pt}{(x_{}^{i'},r^{i})}`$ required to transmit $`r^{i}`$ bits using ([3](#S2.E3 "In II System Model ‣ Power Efficient Scheduling under Delay Constraints over Multi-user Wireless Channels"))
15:
Inform the base station of the rate $`r^{i}`$
16:
while Base station has not scheduled a user do
17:
wait
18:
end while
19:
if user $`i`$ is scheduled then
20:
$`u^{i}\leftarrow r^{i}`$
21:
else
22:
$`u^{i}\leftarrow 0`$
23:
end if
24:
Update the component $`(q^{i},x^{i})`$ of the value function matrix $`{\overset{\sim}{V}}^{i}`$ using ([28](#S4.E28 "In IV-B The Online Rate Allocation Algorithm ‣ IV Single User Scenario in Presence of Transmitter Errors ‣ Power Efficient Scheduling under Delay Constraints over Multi-user Wireless Channels")). Rest of the components of the matrix remain unchanged
25:
Update the LM $`\lambda^{i}`$ using ([29](#S4.E29 "In IV-B The Online Rate Allocation Algorithm ‣ IV Single User Scenario in Presence of Transmitter Errors ‣ Power Efficient Scheduling under Delay Constraints over Multi-user Wireless Channels")) ($`Q_{n}^{i} = q^{i}`$)
26:
$`q^{i}\leftarrow{{q^{i} + a^{i}} - u^{i}}`$
27:
$`x^{i}\leftarrow x_{}^{i'}`$
28:
$`n\leftarrow{n + 1}`$
29:
end while
Algorithm 2
The User Selection Algorithm at the Base Station
1:
while TRUE do
2:
for $`i \in {1,\ldots,N}`$ do
3:
Estimate the channel state $`X_{n + 1}^{i} = x_{}^{i'}`$ in the current slot for user $`i`$
4:
Inform $`{}x_{}^{i'}`$ to user $`i`$
5:
end for
6:
while Rate of each user is not known do
7:
wait
8:
end while
9:
Determine the user $`k`$ who has the highest rate
10:
Schedule user $`k`$ in the current slot
11:
end while
###
V-D
Discussion
Here we discuss certain aspects of the online algorithm:
1. 1\.
*Computational complexity:* The computational complexity of the rate allocation algorithm executed at a user device is independent of the number of users in the system. This is because the rate allocation algorithm for any user $`i`$ is directly dependent on the user $`i`$ state $`S^{i}`$ only and is independent of the states of the other users. The user selection algorithm has to determine the maximum of $`N`$ numbers and hence is linear in $`N`$. Thus the computational complexity of the user selection algorithm grows only linearly with the number of users.
2. 2\.
*An auctioning interpretation:* The solution can be interpreted as an auction, where the user selection algorithm auctions each time slot. The users bid in the form of their transmission rates to the user selection algorithm, which allocates the time slot to the user bidding the highest rate. The rate bid by a user is dependent on its channel state and queue length constraint violation (i.e., the difference between the current queue length and the queue length constraint). If the channel state is quite good and queue constraint violation is large, the user bids a high rate. This is because transmitting at a high rate when the channel state is good saves power, while doing it when the queue length constraint violation is large aids in satisfying the delays. Note that the users do not bid unnecessarily high rates because that might result in higher power consumption. For a user, not winning an auction in a certain slot, implies that other users either have better channel conditions or higher queue length constraint violation or both. If a user does not win the auction for a certain number of slots successively, its queue length grows thus forcing it to bid a higher rate. Motivated by this interpretation, we refer to the scheduling scheme proposed in this paper as the Auctioning Algorithm (AA).
## VI Simulation Setup and Results
We demonstrate the performance of our algorithm under the IEEE 802.16 \[[1](#bib.bib1)\] framework through our simulations in a discrete event simulator. Specifically, we intend to demonstrate the following:
1. 1\.
The algorithm satisfies the delay constraints of all the users.
2. 2\.
The algorithm is efficient in terms of the power consumed for each of the users. Moreover, power consumed is commensurate with the delay requirements, average arrival rates and the channel states of the users.
3. 3\.
The sum power consumed under our algorithm is marginally more than that consumed under the optimal algorithm in Section [](#S3.SS3 "III-C The CMDP Framework ‣ III Problem Formulation ‣ Power Efficient Scheduling under Delay Constraints over Multi-user Wireless Channels")
III-C
.
4. 4\.
Average power expended by the users under our algorithm, is much less than that expended under the M-LWDF scheme \[[13](#bib.bib13)\].
###
VI-A
M-LWDF Algorithm Details
The M-LWDF scheduler \[[13](#bib.bib13)\] attempts to minimize the user delay. It also considers the probability with which a user’s queue length is allowed to exceed a certain target queue length. We assume that this probability is the same for all the users and hence adapt the M-LWDF scheme for our scenario by ignoring it in the present simulations. Specifically, the adapted M-LWDF scheme schedules a user $`i`$ in each slot such that,
<table>
<tbody>
<tr>
<td></td>
<td>i=argmaxjτnj×Unj,𝑖subscript𝑗subscriptsuperscript𝜏𝑗𝑛subscriptsuperscript𝑈𝑗𝑛i=\arg\max_{j}\tau^{j}_{n}\times U^{j}_{n},</td>
<td></td>
<td>(30)</td>
</tr>
</tbody>
</table>
where $`\tau_{n}^{j}`$ is the delay experienced by the head of the line packet for user $`j`$. M-LWDF scheme transmits at a constant maximum power in each time slot. We first determine the average delays experienced by the users under the M-LWDF scheme for various transmission powers. We consider the values of these delays to be the delay constraints for our algorithm. We determine the average delays experienced by the users under our algorithm and also the power expended by the users under our algorithm for the same maximum transmission power in each slot as in the M-LWDF scheme. We compare the average power expended by the users under under our algorithm, with that expended under the M-LWDF scheme. We perform the simulations within the framework of an IEEE 802.16 system. Next, we provide some details regarding the IEEE 802.16 system.
###
VI-B
The IEEE 802.16 System
The IEEE 802.16 standard specifies two modes for sharing the wireless medium: point-to-multipoint (PMP) and mesh. In this paper, we concentrate on the PMP mode where a centralized base station (BS) serves multiple subscriber stations (SSs). We consider the uplink (UL) transmissions. IEEE 802.16 medium access control (MAC) specfies four different scheduling services in order to meet the QoS requirements of various applications. These are: unsolicited grant service (UGS) (for real-time applications with strict delay requirements), real-time polling service (rtPS) (for real-time applications with less stringent delay requirements), non real-time polling service (nrtPS) and best effort (BE) (for applications that do not have any delay requirements). However, unlike BE connection, nrtPS connection is reserved a minimum amount of bandwidth. We consider the *residential* scenario as in \[[38](#bib.bib38)\]. It consists of a BS providing Internet access to the subscribers. Although the standard does not specify any QoS class for providing average delays, we argue that the nrtPS must be extended to cater to the average delay requirements of the users. The unicast polling service of nrtPS can be extended to inform a user the channel state perceived by the base station as well as to determine the rate at which a user would transmit if it were to be scheduled. The scheduling algorithm can thus be implemented as a part of nrtPS.
The system can be operated in either time division duplex (TDD) or frequency division duplex (FDD) mode. We assume the FDD mode of operation where all SSs have full-duplex capability. We consider a single carrier system (WirelessMAN-SC) with a frame duration of $`1`$ msec and bandwidth $`W`$ of $`10`$ MHz. We assume that the users transmit at a rate such that data is delivered reliably to the base station. Hence we do not consider retransmissions and Automatic Repeat Request (ARQ). The SSs employ the following modulations: 64 Quadrature Amplitude Modulation (QAM), 16 QAM, Quadrature Phase Shift Keying (QPSK), QPSK with 1/2 rate convolutional code which provide us with $`4`$ rates of transmission.
###
VI-C
Simulation Setup and Results
Internet traffic is modeled as a web traffic source \[[38](#bib.bib38), [39](#bib.bib39)\]. Packet sizes are drawn from a truncated Pareto distribution (shape factor $`1.2`$, mode = $`2000`$ bits, cutoff threshold = $`10000`$ bits) which provides us with an average packet size of $`3860`$ bits. In each time frame, we generate the arrivals for all the users using Poisson distribution
<sup>2</sup>
<sup>2</sup>2AA does not rely on the Poisson arrival process of the users, we simulate using the Poisson process only for the purpose of illustration.
. Arrivals are generated in an i.i.d. manner across frames. We fragment the packets into fragments of size $`2000`$ bits each. Fragments of size less than $`2000`$ bits are padded with extra bits to make them of size $`2000`$ bits. Since all fragments are of equal size, we determine the transmission rate for users in terms of number of fragments per frame instead of bits per frame. We simulate a Rayleigh fading channel
<sup>3</sup>
<sup>3</sup>3AA does not rely on the Rayleigh channel, we simulate using a Rayleigh channel only for the purpose of illustration.
for each user. For a Rayleigh model, channel state $`X^{i}`$ is an exponentially distributed random variable with probability density function given by $`{f_{X^{i}}\hspace{0pt}{(x)}} = {\frac{1}{\alpha^{i}}\hspace{0pt}e^{- \frac{x}{\alpha^{i}}}}`$, where $`\alpha^{i}`$ is the mean of $`X^{i}`$. We know from ([3](#S2.E3 "In II System Model ‣ Power Efficient Scheduling under Delay Constraints over Multi-user Wireless Channels")) that the power required for transmitting $`u`$ fragments of size $`\ell`$ bits when the channel state is $`x`$ is given by, $`{P\hspace{0pt}{(x,u)}} = {\frac{N_{0}\hspace{0pt}W}{x}\hspace{0pt}\left( {2^{{u\hspace{0pt}\ell}/W} - 1} \right)}`$, where $`N_{0}`$ is the power spectral density of the additive white Gaussian noise and $`W`$ is the received signal bandwidth. We assume that the product $`W\hspace{0pt}N_{0}`$ is normalized to $`1`$. We measure the sum of queuing and transmission delays of the packets and ignore the propagation delays. In all the scenarios described below, a single simulation run consists of running the algorithm for $`100000`$ frames and the results are obtained after averaging over $`20`$ simulation runs.
######
Scenario 1
Comparison with the Optimal Algorithm: This scenario demonstrates that the sum power expended by AA is very close to that expended by the optimal algorithm (OA) suggested in Section [](#S3.SS3 "III-C The CMDP Framework ‣ III Problem Formulation ‣ Power Efficient Scheduling under Delay Constraints over Multi-user Wireless Channels")
III-C
. For this scenario, we assume $`N = 2`$, i.e., two users. The channel state can be either bad ($`\alpha^{1}`$ = $`0.1422`$ ($`- 8.47`$ dB)) or good ($`\alpha^{2}`$ = $`2.0796`$ ($`3.18`$ dB)). We assume a buffer of size $`10`$ packets ($`B = 10`$) at each user. In each frame the arrivals are generated using the Poisson distribution with mean $`0.05`$ packets/msec ($`0.184`$ Mbits/sec/user). Packet lengths are Pareto distributed with parameters as discussed previously in this section. In each frame, we generate a Rayleigh random variable with mean $`0.9817`$ ($`- 0.08`$ dB). If the value taken by the random variable is greater than $`2.0796`$, the channel state is assumed to be good, else channel state is assumed to be bad. The peak transmission power in any slot is fixed at $`3`$ Watts. We compare the sum power for the two users for the two schemes in Table [I](#S6.T1 "TABLE I ‣ Scenario 1 ‣ VI-C Simulation Setup and Results ‣ VI Simulation Setup and Results ‣ Power Efficient Scheduling under Delay Constraints over Multi-user Wireless Channels"). It can be seen that both the schemes satisfy the delay constraints. The power required by AA is marginally more than that required for the OA.
TABLE I:
Comparison between Optimal Algorithm (OA) and the Auctioning Algorithm (AA)
<table>
<tbody>
<tr>
<td>Delay</td>
<td>Achieved</td>
<td>Achieved</td>
<td>Power for</td>
<td>Power for</td>
</tr>
<tr>
<td>Constraint</td>
<td>Delay for OA</td>
<td>Delay for AA</td>
<td>OA</td>
<td>AA</td>
</tr>
<tr>
<td>3 msec</td>
<td>3.09119</td>
<td>3.09700</td>
<td>0.26359</td>
<td>0.26522</td>
</tr>
<tr>
<td>5 msec</td>
<td>3.57960</td>
<td>3.40470</td>
<td>0.24756</td>
<td>0.26097</td>
</tr>
</tbody>
</table>
For the rest of the scenarios, we discretize the channel into eight equal probability bins, with the boundaries specified by { (-$`\infty`$, $`- 8.47`$ dB), \[$`- 8.47`$ dB, $`- 5.41`$ dB), \[$`- 5.41`$ dB, $`- 3.28`$ dB), \[$`- 3.28`$ dB, $`- 1.59`$ dB), \[$`- 1.5`$ dB, $`- 0.08`$ dB), \[$`- 0.08`$ dB, $`1.42`$ dB), \[$`1.42`$ dB, $`3.18`$ dB), \[$`3.18`$ dB, $`\infty`$ ) }. For each bin, we associate a channel state and the state space $`\mathcal{X}`$ = { $`- 13`$ dB, $`- 8.47`$ dB, $`- 5.41`$ dB, $`- 3.28`$ dB, $`- 1.59`$ dB, $`- 0.08`$ dB, $`1.42\hspace{0pt}d\hspace{0pt}B`$, $`3.18`$ dB}. This discretization of the state space of $`X^{i}`$ has been justified in \[[19](#bib.bib19)\]. We assume $`N = 20`$, i.e., a system with $`20`$ users and thereby $`20`$ UL connections. We assume that the number of users do not change during the course of the simulation. Users are divided into two groups (Group 1 and Group 2) of $`10`$ users each.
######
Scenario 2
Comparison with the M-LWDF Algorithm: We compare the power consumed by the M-LWDF algorithm with the power consumed by the AA. We first simulate the M-LWDF scheme \[[13](#bib.bib13)\]. In each frame, arrivals are generated with a Poisson distribution with mean $`0.1`$ packets/msec. Packet lengths are Pareto distributed as explained above. This results in an arrival rate of $`0.386`$ Mbits/sec/user. We choose $`\alpha^{i} = {0.9817\hspace{0pt}{({- {0.08\hspace{0pt}{dB}}})}\hspace{0pt}{\forall i}}`$. In each slot we generate $`X^{i}`$ using exponential distribution with mean $`\alpha^{i}`$. We determine the channel state based on the bin that contains $`X^{i}`$ as explained above. We perform multiple experiments. In successive experiments, we fix the maximum transmission power at $`1.5,2,2.5,3,4,4.5`$ Watts respectively. For each experiment, we determine the average delays experienced by the users. Next we fix the achieved delays by the M-LWDF algorithm as the delay constraints for our algorithm. The maximum transmission power is the same as that for the M-LWDF algorithm. We determine the average power expended for each user under the AA and compare it to that for the M-LWDF algorithm in Table [II](#S6.T2 "TABLE II ‣ Scenario 2 ‣ VI-C Simulation Setup and Results ‣ VI Simulation Setup and Results ‣ Power Efficient Scheduling under Delay Constraints over Multi-user Wireless Channels") (Power constraint and average power expenditure expressed in Watts, achieved delays expressed in milli-seconds (msec)). It can be seen that for all the experiments, the AA satisfies the delay constraints. Moreover, the power expended by each user under the AA is much less than that expended under the M-LWDF algorithm. This is because the AA is able to adapt the transmission power based on the channel state and delay constraint violation.
TABLE II:
Comparison between M-LWDF and the AA
<table>
<tbody>
<tr>
<td>Power</td>
<td>Achieved Delay</td>
<td>Achieved</td>
<td>Average Power</td>
<td>Average</td>
</tr>
<tr>
<td>Constr.</td>
<td>- M-LWDF</td>
<td>Delay - AA</td>
<td>- M-LWDF</td>
<td>Power - AA</td>
</tr>
<tr>
<td>1.5</td>
<td>28.71532</td>
<td>28.12900</td>
<td>0.07499</td>
<td>0.04206</td>
</tr>
<tr>
<td>2</td>
<td>28.18142</td>
<td>28.01677</td>
<td>0.09999</td>
<td>0.04737</td>
</tr>
<tr>
<td>2.5</td>
<td>22.57460</td>
<td>22.03332</td>
<td>0.12499</td>
<td>0.05530</td>
</tr>
<tr>
<td>3</td>
<td>22.12825</td>
<td>18.27730</td>
<td>0.14999</td>
<td>0.07007</td>
</tr>
<tr>
<td>3.5</td>
<td>21.99025</td>
<td>16.36487</td>
<td>0.17499</td>
<td>0.07026</td>
</tr>
<tr>
<td>4</td>
<td>20.09445</td>
<td>16.39282</td>
<td>0.19999</td>
<td>0.07073</td>
</tr>
<tr>
<td>4.5</td>
<td>20.09445</td>
<td>14.74980</td>
<td>0.22497</td>
<td>0.07074</td>
</tr>
</tbody>
</table>
######
Scenario 3
In this scenario, we demonstrate that the AA satisfies the various user specified delay constraints. We consider two cases: symmetric and asymmetric. In each frame arrivals are generated with a Poisson distribution with mean $`0.1`$ packets/msec. Packet lengths are Pareto distributed with parameters detailed above. This results in an arrival rate of $`0.386`$ Mbits/sec/user. We choose $`\alpha^{i} = {0.9817\hspace{0pt}{({- {0.08\hspace{0pt}{dB}}})}\hspace{0pt}{\forall i}}`$. In each slot we generate $`X^{i}`$ using exponential distribution with mean $`\alpha^{i}`$. We determine the channel state based on the bin that contains $`X^{i}`$ as explained above. We perform multiple experiments. In the symmetric scenario, in successive experiments, the delay constraints of all the users are fixed at $`25,50,75,100,125,150,175`$ msec respectively. We measure the average delay experienced and the average power expended by each user in each experiment. These quantities for a user selected at random are plotted in Figure [3](#S7.F3 "Figure 3 ‣ Power Efficient Scheduling under Delay Constraints over Multi-user Wireless Channels"). In the asymmetric case, the delay constraint of the users in Group $`1`$ are fixed at $`100`$ msec in each experiment, while the delay constraints of the users in Group $`2`$ are fixed at $`25,50,75,100,125,150,175`$ msec in successive experiments. Average delay suffered by a user selected at random from Group 1 and Group 2 and power consumed by them are plotted in Figure [4](#S7.F4 "Figure 4 ‣ Power Efficient Scheduling under Delay Constraints over Multi-user Wireless Channels"). It can be observed from Figures [3(a)](#S7.F3.sf1 "In Figure 3 ‣ Power Efficient Scheduling under Delay Constraints over Multi-user Wireless Channels") and [4(a)](#S7.F4.sf1 "In Figure 4 ‣ Power Efficient Scheduling under Delay Constraints over Multi-user Wireless Channels") that the delay constraints are satisfied in both the cases. Moreover, from Figures [3(b)](#S7.F3.sf2 "In Figure 3 ‣ Power Efficient Scheduling under Delay Constraints over Multi-user Wireless Channels") and [4(b)](#S7.F4.sf2 "In Figure 4 ‣ Power Efficient Scheduling under Delay Constraints over Multi-user Wireless Channels") it can be observed that power expended is a convex decreasing function of the delay constraint imposed by the user. Larger delay constraints imply that much lesser power is required to satisfy the constraint.
######
Scenario 4
In this scenario, we demonstrate that the AA satisfies the user specified delay constraints for various channel conditions. We consider two cases: symmetric and asymmetric. The delay constraints of all the users are kept constant at $`100`$ msec. For the symmetric case, we fix $`\alpha^{i}`$ as $`- {13\hspace{0pt}{dB}}`$, $`- {8.47\hspace{0pt}{dB}}`$, $`- {5.41\hspace{0pt}{dB}}`$, $`- {3.28\hspace{0pt}{dB}}`$, $`- {1.59\hspace{0pt}{dB}}`$, $`- {0.08\hspace{0pt}{dB}}`$, $`1.42\hspace{0pt}{dB}`$, $`\forall i`$ in successive experiments. Rest of the parameters are the same as in Scenario [3](#Thmscn3 "Scenario 3 ‣ VI-C Simulation Setup and Results ‣ VI Simulation Setup and Results ‣ Power Efficient Scheduling under Delay Constraints over Multi-user Wireless Channels"). We measure the average delay suffered by each of the users and the average power consumed by each of them. These quantities are plotted in Figure [5](#S7.F5 "Figure 5 ‣ Power Efficient Scheduling under Delay Constraints over Multi-user Wireless Channels"). In the asymmetric case, we maintain the average channel state for users in Group 1 constant for all the experiments, i.e., $`{\alpha^{i} = {- {0.08\hspace{0pt}{dB}}}},{i \in {1,\ldots,10}}`$. For the users in Group 2, i.e., $`\alpha^{i}`$ for $`i \in {11,\ldots,20}`$, the average channel state is fixed at $`\alpha^{i}`$ $`=`$ $`- {13\hspace{0pt}{dB}}`$, $`- {8.47\hspace{0pt}{dB}}`$, $`- {5.41\hspace{0pt}{dB}}`$, $`- {3.28\hspace{0pt}{dB}}`$, $`- {1.59\hspace{0pt}{dB}}`$, $`- {0.08\hspace{0pt}{dB}}`$, $`1.42\hspace{0pt}{dB}`$, in successive experiments. Average delay suffered by a user in Group 1 and in Group 2 and power consumed by them are plotted in Figure [6](#S7.F6 "Figure 6 ‣ Power Efficient Scheduling under Delay Constraints over Multi-user Wireless Channels"). It can be observed from Figures [5(a)](#S7.F5.sf1 "In Figure 5 ‣ Power Efficient Scheduling under Delay Constraints over Multi-user Wireless Channels") and [6(a)](#S7.F6.sf1 "In Figure 6 ‣ Power Efficient Scheduling under Delay Constraints over Multi-user Wireless Channels") that the delay constraints are satisfied even for extremely poor channel conditions. Moreover, from Figures [5(b)](#S7.F5.sf2 "In Figure 5 ‣ Power Efficient Scheduling under Delay Constraints over Multi-user Wireless Channels") and [6(b)](#S7.F6.sf2 "In Figure 6 ‣ Power Efficient Scheduling under Delay Constraints over Multi-user Wireless Channels") it can be observed that the scheme is able to satisfy the delay constraints above a certain average channel state
<sup>4</sup>
<sup>4</sup>
4
This average channel state is dependent on the peak transmission power.
. Better channel conditions imply that much lesser power is required to satisfy the delay constraints.
######
Scenario 5
In this scenario we demonstrate the range of arrival rates for which the AA satisfies the user specified delay constraint of $`100`$ msec. We consider two cases - symmetric and asymmetric. In the symmetric case, the arrival rates of all the users are fixed at $`0.2702\hspace{0pt}{to0}\hspace{0pt}.5018`$ Mbits/sec ($`0.05\hspace{0pt}{to0}\hspace{0pt}.12`$ packets/msec) in successive experiments. Rest of the parameters are same as in Scenario [3](#Thmscn3 "Scenario 3 ‣ VI-C Simulation Setup and Results ‣ VI Simulation Setup and Results ‣ Power Efficient Scheduling under Delay Constraints over Multi-user Wireless Channels"). We measure the average delay suffered and average power expended by each user. These quantities for a user chosen at random are plotted in Figure [7](#S7.F7 "Figure 7 ‣ Power Efficient Scheduling under Delay Constraints over Multi-user Wireless Channels"). In the asymmetric case, the arrival rate of the users in Group 1 is fixed at $`0.386`$ Mbits/sec ($`0.15`$ packets/msec) for all the experiments, while the arrival rates of the users in Group 2 are increased from $`0.1351 - 0.2509`$ Mbits/sec ($`0.07 - 0.13`$ packets/msec) in $`8`$ steps in successive experiments. Rest of the parameters are same as in Scenario [3](#Thmscn3 "Scenario 3 ‣ VI-C Simulation Setup and Results ‣ VI Simulation Setup and Results ‣ Power Efficient Scheduling under Delay Constraints over Multi-user Wireless Channels"). Average delay suffered by a user from Group1 and Group 2 (each selected at random) and power consumed by them are plotted in Figure [8](#S7.F8 "Figure 8 ‣ Power Efficient Scheduling under Delay Constraints over Multi-user Wireless Channels"). It can be observed from Figures [7(a)](#S7.F7.sf1 "In Figure 7 ‣ Power Efficient Scheduling under Delay Constraints over Multi-user Wireless Channels") and [8(a)](#S7.F8.sf1 "In Figure 8 ‣ Power Efficient Scheduling under Delay Constraints over Multi-user Wireless Channels") that the delay constraints are satisfied in both the cases. From Figures [7(b)](#S7.F7.sf2 "In Figure 7 ‣ Power Efficient Scheduling under Delay Constraints over Multi-user Wireless Channels") and [8(b)](#S7.F8.sf2 "In Figure 8 ‣ Power Efficient Scheduling under Delay Constraints over Multi-user Wireless Channels") it can be seen that power expended is an increasing function of the average arrival rates for the same delay constraint. Higher the arrival rate, higher is the power expended.
## VII Conclusion
In this paper, we have considered the problem of power efficient uplink scheduling in a TDMA system over a fading wireless channel with the objective of minimizing user power expenditure under individual delay constraints. We have assumed that the user statistics are unknown, i.e., the probability distributions of the user arrivals and channel states are unknown. We have formulated the problem under the CMDP framework. Determining the optimal policy under the CMDP framework faces two problems: state space explosion and unknown user statistics. To tackle state space explosion, we have suggested performing the rate allocation for a particular user based on its buffer occupancy and channel state only. The rate allocation algorithm is a learning algorithm that learns about the channel state and buffer occupancy of a user during system execution and determines its rate of transmission and hence takes care of the unknown user statistics. Once the rate allocation for all the users is is done, the algorithm schedules a user with the highest rate in a slot. We have performed simulations in the IEEE 802.16 system. Our simulation results have demonstrated that the system is indeed able to satisfy the user specified delay constraints and comparison with the M-LWDF scheme indicates that power expended by the users is low. Moreover, the power expended is commensurate with the QoS requirement, lower average arrival rates, better average channel conditions and higher average delay requirements lead to lower power expenditure.
## References
- \[1\] LAN/MAN Standards Committee, *IEEE standard for Local and Metropolitan Area Networks - Part 16: Air Interface for Fixed Broadband Wireless Access Systems (IEEE P802.16-REVd/D5)*. IEEE Computer Society, May 2004.
- \[2\]
“Third Generation Partnership Project.” \[Online\]. Available: http://www.3gpp.org/
- \[3\] D. Tse and P. Viswanath, *Fundamentals of Wireless Communication*. Cambridge University Press, 2005.
- \[4\] D. Tse and S. Hanly, “Multi-access Fadaing Channels - Part I: Polymatroid Structure, Optimal Resource Allocation and Throughput Capacities,” *IEEE Transactions on Information Theory*, vol. 44, no. 7, pp. 2796–2815, 1998.
- \[5\] R. Knopp and P. A. Humblet, “Information Capacity and Power Control in Single-Cell Multiuser Communications,” in *Proceedings of IEEE ICC*, vol. 1, Seattle, USA, June 1995, pp. 331–335.
- \[6\] S. Shakkottai, T. S. Rappaport, and P. C. Karlsson, “Cross-Layer Design for Wireless Networks,” *IEEE Communications Magazine*, vol. 41, no. 10, pp. 74–80, 2003.
- \[7\] R. Berry and E. M. Yeh, “Cross-Layer Wireless Resource Allocation,” *IEEE Signal Processing Magazine*, vol. 21, no. 5, pp. 59–68, 2004.
- \[8\] E. Yeh and A. Cohen, “Throughput and Delay Optimal Resource Allocation in Multiaccess Fading Channels,” in *Proceedings of International Symposium on Information Theory*, 2003, p. 245.
- \[9\] M. J. Neely, E. Modiano, and C. Rohrs, “Power and Server Allocation in Multi-beam Satellite with Time Varying Channels,” in *Proceedings of IEEE INFOCOM*, San Francisco, USA, 2002, pp. 138–152.
- \[10\] L. Tassiulas and A. Ephremides, “Dynamic Server Allocation to Parallel Queues with Randomly Varying Connectivity,” *IEEE Transactions on Information Theory*, no. 2, pp. 466–478, 1993.
- \[11\] S. Shakkottai and A. Stolyar, “Scheduling for Multiple Fows Sharing a Time-varying Channel: the Exponential Rule,” *AMS Translations Series 2*, vol. 207, 2002.
- \[12\] E. Yeh and A. Cohen, “Delay Optimal Rate Allocation in Multiaccess Fading Communications,” in *Proceedings of IEEE Workshop on Multimedia Signal Processing*, Oct. 2002, pp. 404–407.
- \[13\] M. Andrews, K. Kumaran, K. Ramanan, A. Stolyar, P. Whiting, and R. Vijayakumar, “Providing Quality of Service over a Shared Wireless Link,” *IEEE Communications Magazine*, vol. 39, Feb. 1996.
- \[14\] D. Tse and S. Hanly, “Multi-access Fadaing Channels - Part II: Delay-Limited Capacities,” *IEEE Transactions on Information Theory*, vol. 44, no. 7, pp. 2816–2831, 1998.
- \[15\] E. Yeh, “Multiaccess and Fading in Communication Networks,” PhD Thesis, Massachusetts Institute of Technology, September 2001.
- \[16\] E. F. Chaponniere, P. Black, J. M. Holtzman, and D. Tse, “Transmitter directed Multiple Receiver System using Path Diversity to Equitably Maximize Throughput ,” Sept. 2002, U.S. Patent No. 6449490.
- \[17\] X. Liu, E. Chong, and N. Shroff, “Opportunistic Transmission Scheduling with Resource-Sharing Constraints in Wireless Networks,” *IEEE Journal on Selected Areas in Communication*, vol. 19, no. 10, pp. 2053–2064, 2001.
- \[18\] R. A. Berry and R. G. Gallager, “Communication over Fading Channels with Delay Constraints,” *IEEE Transactions on Information Theory*, vol. 48, no. 5, pp. 1135–1149, 2002.
- \[19\] H. Wang and N. Mandayam, “A Simple Packet Transmission Scheme for Wireless Data over Fading Channels,” *IEEE Transcations on Communications*, vol. 52, no. 7, pp. 1055–1059, 2004.
- \[20\] M. Goyal, A. Kumar, and V. Sharma, “Power Constrained and Delay Optimal Policies for Scheduling Transmissions over a Fading Channel,” in *Proceedings of IEEE INFOCOM*, vol. 1, Mar. 2003, pp. 311–320.
- \[21\]
M. Agarwal, A. Karandikar, and V. S. Borkar, “Structural Properties of Optimal Transmission Policies over a Randomly Varying channel,” *Under review in IEEE Transactions on Automatic Control*, 2006. \[Online\]. Available: http://www.ee.iitb.ac.in/~karandi/pubs˙dir/mukul.ps.gz
- \[22\] D. V. Djonin and V. Krishnamurthy, “Structural Results on the Optimal Transmission Scheduling Policies and Costs for Correlated Sources and Channels,” in *IEEE CDC*, Dec. 2005, pp. 3231–3236.
- \[23\] B. Prabhakar, E. U. Biyikoglu, and A. E. Gamal, “Energy-Efficient Transmission over a Wireless Link via Lazy Packet Scheduling,” in *Proceedings of IEEE INFOCOM*, vol. 1, Apr. 2001, pp. 386–394.
- \[24\] B. E. Collins and R. L. Cruz, “Transmission Policies for Time Varying Channels with Average Delay Constraints,” in *Proceedings of 1999 Allerton Conference on Communication, Control, and Computation*, 1999.
- \[25\] G. Rajadhyaksha and V. S. Borkar, “Transmission Rate Control over Randomly Varying Channels,” *Probability in the Engineering and Informational Sciences*, vol. 19, no. 1, pp. 73–82, Jan. 2005.
- \[26\] M. J. Neely, “Optimal Energy and Delay Tradeoffs for Multi-user Wireless Downlinks,” in *Proceedings of IEEE INFOCOM*, 2006, pp. 1–13.
- \[27\] R. Berry, “Power and Delay Trade-offs in Fading Channels,” June 2000, PhD Thesis, Massachusetts Institute of Technology.
- \[28\]
N. Salodkar, A. Bhorkar, A. Karandikar, and V. S. Borkar, “On-Line Learning Algorithm for Energy Efficient Delay Constrained Scheduling over Fading Channel,” May 2007, Submitted. \[Online\]. Available: http://www.it.iitb.ac.in/~nitins/post˙decision-1.pdf
- \[29\] H. S. Wang and N. Moayeri, “Finite-State Markov Channel – A Useful Model for Radio Communication Channels,” *IEEE Transactions on Vehicular Technology*, vol. 44, no. 1, pp. 163–171, Feb. 1995.
- \[30\] D. P. Bertsekas and R. Gallager, *Data Networks*. Prentice Hall, 1987.
- \[31\] C. A. C. Coello, “A Comprehensive Survey of Evolutionary-Based Multiobjective Optimization Techniques,” *Knowledge and Information Systems*, vol. 3, no. 1, pp. 269–308, 1999.
- \[32\] E. Altman, *Constrained Markov Decision Processes*. Boca Raton, FL: Chapman and Hall/CRC Press, 1999.
- \[33\] D. P. Bertsekas, *Dynamic Programming and Optimal Control*. Belmont, MA: Athena Scientific, 1995, vol. 1.
- \[34\] V. S. Borkar, “Stochastic Approximation with Two Time Scales,” *Systems and Control Letters*, vol. 29, pp. 291–294, 1996.
- \[35\] M. Puterman, *Markov Decision Processes*. New York: Wiley, 1994.
- \[36\] H. P. Young, *Strategic Learning and its Limits*. New York: Oxford University Press, 2004.
- \[37\] ——, “The Possible and the Impossible in Multi-Agent Learning,” *Artificial Intelligence*, vol. 171, no. 7, pp. 429–433, 2007.
- \[38\] C. Cicconeti, L. Lenzini, E. Mingozzi, and C. Eklund, “Quality of Service Support in IEEE 802.16 Networks,” *IEEE Network*, pp. 50–55, March/April 2006.
- \[39\] T. G. Neame, M. Zukerman, and R. G. Addie, “Modeling Broadband Traffic Streams,” in *Proceedings of IEEE GLOBECOM*, 1999, pp. 1048–1052.
[Refer to caption](/html/0710.1190/assets/x3.png)
\(a\)
Achieved delay of a user
[Refer to caption](/html/0710.1190/assets/x4.png)
\(b\)
Power consumed by a user
Figure 3: Variation of achieved delay and power consumed for various delay constraints - symmetric case
[Refer to caption](/html/0710.1190/assets/x5.png)
\(a\)
Achieved delay of a user
[Refer to caption](/html/0710.1190/assets/x6.png)
\(b\)
Power consumed by a user
Figure 4: Variation of achieved delay and power consumed for various delay constraints - asymmetric case
[Refer to caption](/html/0710.1190/assets/x7.png)
\(a\)
Achieved delay of a user
[Refer to caption](/html/0710.1190/assets/x8.png)
\(b\)
Power consumed by a user
Figure 5: Variation of achieved delay and power consumed for various channel conditions - symmetric case
[Refer to caption](/html/0710.1190/assets/x9.png)
\(a\)
Achieved delay of a user
[Refer to caption](/html/0710.1190/assets/x10.png)
\(b\)
Power consumed by a user
Figure 6: Variation of achieved delay and power consumed for various channel conditions - asymmetric case
[Refer to caption](/html/0710.1190/assets/x11.png)
\(a\)
Achieved delay of a user
[Refer to caption](/html/0710.1190/assets/x12.png)
\(b\)
Power consumed by a user
Figure 7: Variation of achieved delay and power consumed for various arrival rates - symmetric case
[Refer to caption](/html/0710.1190/assets/x13.png)
\(a\)
Achieved delay of a user
[Refer to caption](/html/0710.1190/assets/x14.png)
\(b\)
Power consumed by a user
Figure 8: Variation of achieved delay and power consumed for various arrival rates - asymmetric case |
0710.0386 | cs.NI | cyber | # Comparing Maintenance Strategies for Overlays
<sup>†</sup>
<sup>†</sup>thanks: This work is funded by the 6th FP EVERGROW project.
Supriya Krishnamurthy<sup>1,3</sup>, Sameh El-Ansary<sup>1</sup>, Erik Aurell<sup>1,2</sup> and Seif Haridi<sup>1,3</sup>
<sup>1</sup> Swedish Institute of Computer Science (SICS), Sweden
<sup>2</sup> Department of Computational Biology, KTH-Royal Institute of Technology, Sweden
<sup>3</sup> IMIT, KTH-Royal Institute of Technology, Sweden
{supriya,sameh,eaurell,seif}@sics.se
###### Abstract
In this paper, we present an analytical tool for understanding the performance of structured overlay networks under churn based on the master-equation approach of physics. We motivate and derive an equation for the average number of hops taken by lookups during churn, for the Chord network. We analyse this equation in detail to understand the behaviour with and without churn. We then use this understanding to predict how lookups will scale for varying peer population as well as varying the sizes of the routing tables. We then consider a change in the maintenance algorithm of the overlay, from periodic stabilisation to a reactive one which corrects fingers only when a change is detected. We generalise our earlier analysis to understand how the reactive strategy compares with the periodic one.
## I Introduction
A crucial part of assessing the performance of a structured P2P system (aka DHT) is evaluating how it copes with churn. Extensive simulation is currently the prevalent tool for gaining such knowledge. Examples include the work of Li et al. \[[10](#bib.bib10)\], Rhea et al. \[[12](#bib.bib12)\], and Rowstron et al. \[[5](#bib.bib5)\]. There has also been some theoretical analyses done, albeit less frequently. For instance, Liben-Nowell et al. \[[11](#bib.bib11)\] prove a lower bound on the maintenance rate required for a network to remain connected in the face of a given churn rate. Aspnes et al. \[[4](#bib.bib4)\] give upper and lower bounds on the number of messages needed to locate a node/data item in a DHT in the presence of node or link failures. The value of theoretical studies of this nature is that they provide insights neutral to the details of any particular DHT.
We have chosen to adopt a slightly different approach to theoretical work on DHTs. We concentrate not on establishing bounds, but rather on a more precise prediction of the relevant quantities in such dynamically evolving systems. Our approach is based mainly on the Master-Equation approach used in the analysis of physical systems. We have previously introduced our approach in in \[[7](#bib.bib7), [8](#bib.bib8)\] where we presented a detailed analysis of the Chord system \[[13](#bib.bib13)\]. In this paper, we show that the approach is applicable to other systems as well. We do this by comparing the periodic stabilization maintenance technique of Chord with the correction-on-change maintenance technique of DKS \[[3](#bib.bib3)\].
Due to space limitations, we assume reader familiarity with Chord and DKS, including such terminology as successors, finger starts and finger nodes etc.
The rest of the paper is organised as follows. In Section [II](#S2 "II The Master-Equation Approach for Structured Overlays ‣ Comparing Maintenance Strategies for Overlays This work is funded by the 6th FP EVERGROW project."), we introduce the Master-Equation approach. In Section [III](#S3 "III Related Work ‣ Comparing Maintenance Strategies for Overlays This work is funded by the 6th FP EVERGROW project."), we mention some related work. In section [IV](#S4 "IV The Lookup Equation for Chord ‣ Comparing Maintenance Strategies for Overlays This work is funded by the 6th FP EVERGROW project.") we begin by briefly reviewing some of our previously published results on predicting the performance of the Chord network as a function of the failed pointers in the system in the case that the nodes use a periodic maintenance scheme. We then show some new results on how this complicated equation can be simplified to get quick predictions for varying number of peers and varying number of links per node. We relegate some of the details of this analysis to Appendix [VII](#S7 "VII Appendix ‣ Comparing Maintenance Strategies for Overlays This work is funded by the 6th FP EVERGROW project."). In section [V](#S5 "V ’Correction-on-Change’ Maintenance Strategy ‣ Comparing Maintenance Strategies for Overlays This work is funded by the 6th FP EVERGROW project."), we explain how to use the Master-Equation approach to analyse the reactive maintenance strategy of interest and present our results on how this strategy compares with the periodic case analysed earlier. We summarise our results in Section [VI](#S6 "VI Summary ‣ Comparing Maintenance Strategies for Overlays This work is funded by the 6th FP EVERGROW project.").
## II The Master-Equation Approach for Structured Overlays
In a complicated system like a P2P network, in which there are many participants, and in which there are many inter-leaved processes happening in time, predicting the state of the network (or of any quantity of interest) can at best be done by specifying the probability distribution function (PDF) of the quantity in the steady state (when the system, though changing continually in time, is stationary on average). For example, one quantity of interest for us when analysing such a network, is the fraction of failed links between nodes, in the steady state. This quantity does not take some deterministic value in the steady state. Instead it is specified by a PDF, which can then be used to determine the average value. The problem is thus to calculate the PDF (and then to understand how it affects the performance of the network, as explained below).
In general this is not an easy task, since the probability is affected by a number of inter-leaved processes in any time-varying system. In \[[7](#bib.bib7), [8](#bib.bib8)\], we demonstrated how we could analyse a P2P network like Chord \[[13](#bib.bib13)\], using a Master-Equation based approach. This approach is generally used in physics to understand a system evolving in time, by means of equations specifying the time-evolution of the probabilities of finding the system in a specific state. These equations require as an input, the rates of various processes affecting the state of the system. For example, in a peer-to-peer network, these processes could be the join and failure rates of the member nodes, the rate at which each node performs maintenance as well as the rate at which lookups are done in the network (the latter rate is relevant only if the lookups affect the state of the network in some way). Given these rates, the equation for the time-evolution of the probability of the quantity of interest can be written by keeping track of how these rates affect this quantity (such as the number of failed pointers in the system) in an infinitesimal interval of time, when only a limited number of processes (typically one) can be expected to occur simultaneously.
With this approach, we were able to quantify very accurately the probabilities of any connection in the network (either fingers or successors) having failed. We then demonstrated how we could use this information to predict the performance of the network—the number of hops including time outs which a lookup takes on average — as a function of the rates (of join, failure and stabilization) of all the processes happening in the network, as well as of all the parameters specifying the network (such as how many pointers a node has on average). The analysis was done for a specific maintenance strategy, called periodic maintenance (or eager maintenance)
In this paper, we generalise our approach so as to be able to compare networks using different maintenance strategies. In particular, we compare our earlier results for periodic maintenance with a reactive maintenance strategy proposed in \[[6](#bib.bib6)\]. Combining this with some of our previous results, we are also, as a by product, able to compare the performance of networks specified by different numbers of peers, different number of pointers per node and/or different maintenance strategies. As we show below, which system is better depends both on the value of the parameters as well as the level of churn. The approach we propose is thus a useful tool for the quantitative and fair comparison of networks specified by different parameters and using different algorithms.
## III Related Work
In \[[2](#bib.bib2)\], an analysis, very similar in spirit to the one done in this paper, is carried out in the context of P-Grid \[[1](#bib.bib1)\]. An equation is written for system performance in the state of dynamic equilibrium for various maintenance strategies. However for each maintenance strategy, the analysis has to be entirely redone. In contrast, a master equation description provides a foundation for the theoretical analysis of overlays, which does not have to be entirely rebuilt each time any given algorithm is changed. As we show in this paper, we can carry over a lot of our earlier analysis, when the maintenance scheme is changed from a periodic to a reactive one. In addition, the master equation description can be made arbitrarily precise to include non-linear effects as well. And as we show, non linear effects are important when churn is high.
## IV The Lookup Equation for Chord
We quantify the performance of the network, by the number of hops required on average from the originator of the query to the node with the answer. This is just the total number of nodes contacted per query (or equivalently, the total number of pointers used per query) including the total number of failed pointers used en route. This latter quantity (which arises because of the churn in the network) is the reason that the hop count per query increases with high dynamism and is hence an important quantity to understand. In the case of the periodic maintenance scheme, this quantity is a function of $`{({1 - \beta})}\hspace{0pt}r`$ where $`r`$ is the ratio of the stabilisation rate to the join (or failure) rate and $`1 - \beta`$ is the fraction of times a node stabilises its finger, when performing maintenance, as mentioned in Section [I](#S1 "I Introduction ‣ Comparing Maintenance Strategies for Overlays This work is funded by the 6th FP EVERGROW project."). We demonstrate how this quantity can be calculated in Section [V](#S5 "V ’Correction-on-Change’ Maintenance Strategy ‣ Comparing Maintenance Strategies for Overlays This work is funded by the 6th FP EVERGROW project."), in the context of the reactive maintenance policy, which is a simple generalisation of how it is calculated earlier in\[[7](#bib.bib7), [8](#bib.bib8)\], for the periodic maintenance scheme. In this section, we briefly review our earlier results on how the performance of the network (as exemplified by the average hopcount per query), can be determined once the fraction of failed pointers is known.
The key to predicting the performance of the network is to write a recursive equation for the expected cost $`C_{t}\hspace{0pt}{(r,\beta)}`$ (also denoted $`C_{t}`$) for a given node to reach some target, $`t`$ keys away from it. (For example, $`C_{1}`$ is the cost of looking up the adjacent key which is $`1`$ key away).
The Lookup Equation for the expected cost of reaching a general distance $`t`$ is then derived by following closely the Chord protocol which is a greedy strategy designed to reduce the distance to the query at every step without overshooting the target . A lookup for $`t`$ thus proceeds by first finding the closest preceding finger. The node that this finger points to is then asked to continue the query, if it is alive. If this node is dead, the originator of the query uses the next closest preceding finger and the query proceeds in this manner.
For the purposes of the analysis, it is easier to think in terms of the closest preceding start. Let us hence define $`\xi`$ to be the *s*tart of the finger (say the $`k^{t\hspace{0pt}h}`$) that most closely precedes $`t`$. Hence $`\xi = {2^{k - 1} + n}`$ and $`t = {\xi + m}`$, i.e. there are $`m`$ keys between the sought target $`t`$ and the start of the most closely preceding finger. With that, we can write a recursion relation for $`C_{\xi + m}`$ as follows:
<table>
<tbody>
<tr>
<td></td>
<td>Cξ+m=Cξ[1−a(m)]+(1−fk)a(m)[1+∑i=0m−1bc(i,m)Cm−i]+fka(m)[1+∑i=1k−1hk(i)∑l=0ξ/2i−1bc(l,ξ/2i)(1+(i−1)+Cξi−l+m)+O(hk(k))]subscript𝐶𝜉𝑚subscript𝐶𝜉delimited-[]1𝑎𝑚1subscript𝑓𝑘𝑎𝑚delimited-[]1superscriptsubscript𝑖0𝑚1𝑏𝑐𝑖𝑚subscript𝐶𝑚𝑖subscript𝑓𝑘𝑎𝑚delimited-[]1superscriptsubscript𝑖1𝑘1subscriptℎ𝑘𝑖superscriptsubscript𝑙0𝜉superscript2𝑖1𝑏𝑐𝑙𝜉superscript2𝑖1𝑖1subscript𝐶subscript𝜉𝑖𝑙𝑚𝑂subscriptℎ𝑘𝑘\begin{split}&C_{\xi+m}=C_{\xi}\left[1-a(m)\right]\\ &+(1-f_{k})a(m)\left[1+\sum_{i=0}^{m-1}bc(i,m)C_{m-i}\right]\\ &+f_{k}a(m)\biggl{[}1+\sum_{i=1}^{k-1}h_{k}(i)\\ &\sum_{l=0}^{\xi/2^{i}-1}bc(l,\xi/2^{i})(1+(i-1)+C_{\xi_{i}-l+m})+O(h_{k}(k))\biggr{]}\end{split}</td>
<td></td>
<td>(1)</td>
</tr>
</tbody>
</table>
where $`\xi_{i} \equiv {\sum_{m = {1,i}}{\xi/2^{m}}}`$ and $`h_{k}\hspace{0pt}{(i)}`$ is the probability that a node is forced to use its $`k - i^{t\hspace{0pt}h}`$ finger owing to the death of its $`k^{t\hspace{0pt}h}`$ finger.
The probabilities $`a,{b\hspace{0pt}c}`$ can be derived from the internode interval distribution \[[7](#bib.bib7), [8](#bib.bib8)\] which is just the distribution of distances between adjacent nodes. Given a ring of $`\mathcal{K}`$ keys and $`N`$ nodes (on average), where nodes can join and leave independently, the probability that two adjacent nodes are a distance $`x`$ apart on the ring is simply $`{P\hspace{0pt}{(x)}} = {\rho^{x - 1}\hspace{0pt}{({1 - \rho})}}`$ where $`\rho = \frac{\mathcal{K} - N}{\mathcal{K}}`$. Using this distribution, its easy to estimate the probability that there is definitely atleast one node in an interval of length $`x`$. This is: $`{a\hspace{0pt}{(x)}} \equiv {1 - \rho^{x}}`$. The probability that the first node encountered from any key is at a distance $`i`$ from that key is then $`b_{i} \equiv {\rho^{i}\hspace{0pt}{({1 - \rho})}}`$. Hence the conditional probability that the first node from a given key is at a distance $`i`$ given that there is atleast one node in the interval is $`{b\hspace{0pt}c\hspace{0pt}{(i,x)}} \equiv {{{b\hspace{0pt}{(i)}}/a}\hspace{0pt}{(x)}}`$.
The probability $`h_{k}\hspace{0pt}{(i)}`$ is easy to compute given the probability $`a`$ as well as the probabilities $`f_{k}`$’s of the $`k^{t\hspace{0pt}h}`$ finger being dead.
<table>
<tbody>
<tr>
<td></td>
<td>hk(i)=a(ξ/2i)(1−fk−i)×Πs=1,i−1(1−a(ξ/2s)+a(ξ/2s)fk−s),i<khk(k)=Πs=1,k−1(1−a(ξ/2s)+a(ξ/2s)fk−s)formulae-sequencesubscriptℎ𝑘𝑖𝑎𝜉superscript2𝑖1subscript𝑓𝑘𝑖subscriptΠ𝑠1𝑖11𝑎𝜉superscript2𝑠𝑎𝜉superscript2𝑠subscript𝑓𝑘𝑠𝑖𝑘subscriptℎ𝑘𝑘subscriptΠ𝑠1𝑘11𝑎𝜉superscript2𝑠𝑎𝜉superscript2𝑠subscript𝑓𝑘𝑠\begin{split}h_{k}(i)=&a(\xi/2^{i})(1-f_{k-i})\\ \times&\Pi_{s=1,i-1}(1-a(\xi/2^{s})+a(\xi/2^{s})f_{k-s}),i<k\\ h_{k}(k)=&\Pi_{s=1,k-1}(1-a(\xi/2^{s})+a(\xi/2^{s})f_{k-s})\end{split}</td>
<td></td>
<td>(2)</td>
</tr>
</tbody>
</table>
Eqn.[2](#S4.E2 "In IV The Lookup Equation for Chord ‣ Comparing Maintenance Strategies for Overlays This work is funded by the 6th FP EVERGROW project.") accounts for all the reasons that a node may have to use its $`k - i^{t\hspace{0pt}h}`$ finger instead of its $`k^{t\hspace{0pt}h}`$ finger. This could happen because the intervening fingers were either dead or not distinct (fingers $`k`$ and $`k - 1`$ are not distinct if they have the same entry in the finger table. Though the starts of the two fingers are different, if there is no node in the interval between the starts, the entry in the finger table will be the same). The probabilities $`h_{k}\hspace{0pt}{(i)}`$ satisfy the constraint $`{\sum_{i = 1}^{k}{h_{k}\hspace{0pt}{(i)}}} = 1`$. $`h_{k}\hspace{0pt}{(k)}`$, is the probability that a node cannot use any earlier entry in its finger table,in which case it has to fall back on its successor list instead. We indicate this case by the last term in Eq. [1](#S4.E1 "In IV The Lookup Equation for Chord ‣ Comparing Maintenance Strategies for Overlays This work is funded by the 6th FP EVERGROW project.") which is $`O\hspace{0pt}{({h_{k}\hspace{0pt}{(k)}})}`$. In practise, the probability for this is extremely small except for targets very close to $`n`$. Hence this does not significantly affect the value of general lookups and we ignore it for the moment.
The cost for general lookups is
<table>
<tbody>
<tr>
<td></td>
<td>L(r,β)=Σi=1𝒦−1Ci(r,β)𝒦𝐿𝑟𝛽superscriptsubscriptΣ𝑖1𝒦1subscript𝐶𝑖𝑟𝛽𝒦L(r,\beta)=\frac{\Sigma_{i=1}^{{\cal K}-1}C_{i}(r,\beta)}{\cal K}</td>
<td></td>
</tr>
</tbody>
</table>
The lookup equation is solved recursively numerically, using the expressions for $`a`$, $`b\hspace{0pt}c`$, $`h_{k}\hspace{0pt}{(i)}`$ and $`C_{1}`$. In Fig. [1](#S4.F1 "Figure 1 ‣ IV The Lookup Equation for Chord ‣ Comparing Maintenance Strategies for Overlays This work is funded by the 6th FP EVERGROW project."), we have plotted the theoretical prediction of Equation [1](#S4.E1 "In IV The Lookup Equation for Chord ‣ Comparing Maintenance Strategies for Overlays This work is funded by the 6th FP EVERGROW project.") versus what we get from simulating Chord. Here we have used $`N \sim 1000`$ and $`K = 2^{20}`$. As can be seen the the theoretical results match the simulation results very well.
[Refer to caption](/html/0710.0386/assets/x1.png)
Figure 1: Theory and Simulation for $`L\hspace{0pt}{(r,\beta)}`$
In Fig. [2](#S4.F2 "Figure 2 ‣ IV The Lookup Equation for Chord ‣ Comparing Maintenance Strategies for Overlays This work is funded by the 6th FP EVERGROW project.") we also show the theoretical predictions for some larger values of $`N`$.
[Refer to caption](/html/0710.0386/assets/x2.png)
Figure 2: Lookup cost, theoretical curve, for $`N = {1000,2000,4000,8000,160000}`$ peers. The rationale for the fits is explained later in the text.
On general grounds, it is easy to argue from the structure of Equation [1](#S4.E1 "In IV The Lookup Equation for Chord ‣ Comparing Maintenance Strategies for Overlays This work is funded by the 6th FP EVERGROW project."), that the dependence of the average lookup on churn comes entirely from the presence of the terms $`f_{k}`$. Since $`f_{k} \sim f`$ is independent of $`k`$ for large fingers, we can approximate the average lookup length by the functional form $`{L\hspace{0pt}{(r,\beta)}} = {A + {B\hspace{0pt}f} + {C\hspace{0pt}f^{2}} + \cdots}`$. The coefficients $`A,B,C`$ etc can be recursively computed by solving the lookup equation to the required order in $`f`$. They depend only on $`N`$ the number of nodes, $`1 - \rho`$ the density of peers and $`b`$ the base or equivalently the size of the finger table of each node. The advantage of writing the lookup length this way is that churn-specific details such as how new joinees construct a finger table or how exactly stabilizations are done in the system, can be isolated in the expression for $`f`$. If we were to change our stabilization strategy, as we will demonstrate below, we could immediately estimate the lookup length by plugging in the new expression for $`f`$ in the above relation.
Another advantage of having a simple expression such as the above, is that if we can estimate $`A,B,{C\hspace{0pt}\cdots}`$ accurately, we can make use of the expression for $`L`$ to estimate the churn (or the value of $`r`$) in the system, hence using a local measure to estimate a global quantity. The logic in doing so is the inverse of the reasoning we have used so far. So far, we have used the churn as the input for finding $`f_{k}`$ and hence $`L`$. But we can also reverse the logic and try and estimate churn, if we know the value of the average lookup length $`L`$. If $`L`$ has the above simple expression, then given $`A`$ and $`B`$ to $`O\hspace{0pt}{(f)}`$, we have $`f = \frac{L - A}{B}`$. From the expression for $`f`$ (see section [V](#S5 "V ’Correction-on-Change’ Maintenance Strategy ‣ Comparing Maintenance Strategies for Overlays This work is funded by the 6th FP EVERGROW project.") for how to evaluate $`f`$), we can now get the value of $`r`$. Hence any peer can make an estimate of the churn that the system is facing if it knows how long its lookups are taking on average, and if it has an estimate of $`N`$.
To get $`A`$, we need to consider Eqn [1](#S4.E1 "In IV The Lookup Equation for Chord ‣ Comparing Maintenance Strategies for Overlays This work is funded by the 6th FP EVERGROW project.") with no churn (all $`f_{k}`$’s set to zero). In Appendix [VII](#S7 "VII Appendix ‣ Comparing Maintenance Strategies for Overlays This work is funded by the 6th FP EVERGROW project."), we study the lookup equation [1](#S4.E1 "In IV The Lookup Equation for Chord ‣ Comparing Maintenance Strategies for Overlays This work is funded by the 6th FP EVERGROW project.") in some detail to understand the behaviour without churn and obtain the value of $`A`$ for any base $`b`$. This is useful on several counts. First, the value of $`A`$ is needed to predict the lookup costs as explained above. Secondly, if $`b`$ changes ( a system of base $`b`$ has a finger table of size $`\mathcal{M} = {{({b - 1})}\hspace{0pt}l\hspace{0pt}o\hspace{0pt}g_{b}\hspace{0pt}{(\mathcal{K})}}`$), all else remaining the same, the only major change in the lookup cost is due to the change in $`A`$. So estimating $`A`$ precisely has the benefit that we can predict the lookup cost for any base $`b`$. Thirdly, the analysis confirms that Equation [1](#S4.E1 "In IV The Lookup Equation for Chord ‣ Comparing Maintenance Strategies for Overlays This work is funded by the 6th FP EVERGROW project.") does indeed reproduce well known results for the lookup hop count in Chord, such as for example, that the average lookup cost is $`0.5 \ast {\log{(N)}}`$ without churn \[[13](#bib.bib13)\]. Infact as demonstrated in Appendix [VII](#S7 "VII Appendix ‣ Comparing Maintenance Strategies for Overlays This work is funded by the 6th FP EVERGROW project."), for any $`N`$, the average lookup cost as predicted by Eq. [1](#S4.E1 "In IV The Lookup Equation for Chord ‣ Comparing Maintenance Strategies for Overlays This work is funded by the 6th FP EVERGROW project.") is indeed $`0.5 \ast {\log{(N)}}`$ plus some $`\rho`$-dependent corrections which though small are accurately predicted.
A simple estimate for $`B`$ and $`C`$ can be made in the following manner. Let every finger be dead with some finite probability $`f`$. Each lookup encounters on average $`A`$ fingers, where $`A`$ is the average lookup length without churn. Each of these fingers could be alive (in which case it contributes a cost of $`1`$), dead with a probability $`f`$ in which case it contributes a cost of $`2`$ if the next finger chosen is alive (with probability $`1 - f`$) and so on. Its trivial to verify that this estimates the look-up cost to be $`A\hspace{0pt}{({1 + f + f^{2} + \cdots})}`$. Comparing with our expression for $`L`$, this gives an estimate of $`{B = A},{C = {A,\cdots}}`$.
In general if $`L = {A + {{B \ast g}\hspace{0pt}{(f)}}}`$, then if we scale $`L`$ by plotting $`{({L - A})}/B`$ for varying $`N`$, we should get an estimate of $`g\hspace{0pt}{(f)}`$. Note that $`f`$ depends on $`\rho`$ and $`\mathcal{M}`$ the number of fingers. In addition if $`{g\hspace{0pt}{(f)}} = {{a_{1}\hspace{0pt}f} + {a_{2}\hspace{0pt}f^{2}} + \cdots}`$, the coefficients $`a_{1}`$,$`a_{2}`$, etc can also depend on $`\rho`$. However for $`{1 - \rho}\operatorname{<<}1`$, these dependences on $`\rho`$ are small and the curves for different $`N`$ collapse onto the same curve on scaling. In Fig. [3](#S4.F3 "Figure 3 ‣ IV The Lookup Equation for Chord ‣ Comparing Maintenance Strategies for Overlays This work is funded by the 6th FP EVERGROW project.") we have scaled the curves ploted in Fig. [2](#S4.F2 "Figure 2 ‣ IV The Lookup Equation for Chord ‣ Comparing Maintenance Strategies for Overlays This work is funded by the 6th FP EVERGROW project.") in the above manner, using $`B = A`$. The values of $`A`$ used are derived from the analysis of the previous section. As can be seen the curves collapse onto one curve which is well approximated by the function $`{g\hspace{0pt}{(f)}} = {f + {3 \ast f^{2}}}`$, giving $`a_{1} = 1`$ and $`a_{2} = 3`$. The fits in Fig [2](#S4.F2 "Figure 2 ‣ IV The Lookup Equation for Chord ‣ Comparing Maintenance Strategies for Overlays This work is funded by the 6th FP EVERGROW project.") are also according to this functional form. It should be emphasized however that this approximation for $`g\hspace{0pt}{(f)}`$ is good only for $`{1 - \rho}\operatorname{<<}1`$. For higher values of peer density, the curves for different $`N`$ will not collapse onto one curve and any $`\rho`$-dependence of the coefficients $`a_{i}`$’s will show up as well.
[Refer to caption](/html/0710.0386/assets/x3.png)
Figure 3: Scaled Lookup cost, for $`N = {1000,2000,4000,8000,160000}`$ peers.
We can use the above functional form to predict how lookups would behave if we change the base $`b`$ (the size of the routing table) of the system. In Fig [4](#S4.F4 "Figure 4 ‣ IV The Lookup Equation for Chord ‣ Comparing Maintenance Strategies for Overlays This work is funded by the 6th FP EVERGROW project.") we plot the functional form $`A\hspace{0pt}{(b)}\hspace{0pt}{({1 + {f\hspace{0pt}{(b)}} + {3\hspace{0pt}f\hspace{0pt}{(b)}^{2}}})}`$ for $`b = {2,4,16}`$. The coefficient $`A\hspace{0pt}{(b)}`$ is accurately predicted by Eq. [11](#S7.E11 "In VII Appendix ‣ Comparing Maintenance Strategies for Overlays This work is funded by the 6th FP EVERGROW project.")(in Appendix [VII](#S7 "VII Appendix ‣ Comparing Maintenance Strategies for Overlays This work is funded by the 6th FP EVERGROW project.")), with the definition of $`\xi\hspace{0pt}{({i + 1})}`$ taken appropriately. $`f\hspace{0pt}{(b)}`$ is affected by the base $`b`$ because the number of fingers increases with $`b`$.
As can be seen, when churn is low, a large $`b`$ is an advantage and significantly improves the lookup length. However when churn is high, the flip side of having a larger routing table is that it needs more maintenance. Hence beyond some value of churn, the larger the value of $`b`$, the larger the lookup latency.
This is similar to the spirit of the numerical investigations done in \[[9](#bib.bib9)\]. However when comparing different bases for Chord, Li et al \[[9](#bib.bib9)\] find that while base $`2`$ is the best for high churn (as we find here), base $`8`$ is the best for low churn. Increasing the base beyond this does not seem to improve the cost. The discrepancy between this finding and ours is due to the details of the periodic maintenance scheme which we use. In our case, we have taken the simplest scenario in which each node needs to stabilise $`\mathcal{M}`$ fingers and the order in which this is done is random. In practice only $`\sim {\log N}`$ of the $`\mathcal{M}`$ fingers are distinct, so only $`\sim {\log N}`$ stabilisations need be done by each node. In addition, in \[[9](#bib.bib9)\], finger stabilisations are done only if the finger is pinged and found to be dead.
[Refer to caption](/html/0710.0386/assets/x4.png)
Figure 4: Lookup cost, for $`N = 1000`$ peers for base $`b = {2,4,16}`$.
## V ’Correction-on-Change’ Maintenance Strategy
In this section, we analyse a different maintenance strategy using the master-equation formalism. The strategy we have analysed so far is periodic stabilisation of successors as well as fingers. We now consider a strategy where a node periodically stabilises its successors but does not do so for its fingers. Instead, for maintaining its fingers, it relies on other nodes for updates \[[6](#bib.bib6)\]. Whenever a node $`n`$ detects that its first successor $`n.s_{1}`$ is wrong (failed or incorrect), it sends out messages to all the nodes that are pointing to its wrong first successor, so that they can update their affected finger. The node sending messages can either do so by broadcasting these messages to all affected nodes simultaneously, or by scheduling messages periodically at some rate. We analyse the latter option in this paper, since it provides a more intuitive and broader framework for the comparison of the two schemes
For a system with id-size $`\mathcal{K}`$, there are of the order of $`\mathcal{M} = {\log_{2}\mathcal{K}}`$ fingers pointing to any node (there can be more than this if node spacings are smaller than average. However, as we argue below, for our purpose this is not important). Of course, not all $`\mathcal{M}`$ of these fingers are distinct. Several of these fingers belong to node $`n`$ itself. However to keep the analysis simple (and in keeping with the spirit of our analysis of the periodic stabilisation scheme), we assume that every node that detects a wrong successor needs to send out exactly $`\mathcal{M}`$ messages (even if some of these ’messages’ are sent to itself).
To find out where the nodes that point to $`n.s_{1}`$ are located, $`n`$ needs to do a lookup. For example, to find the node with the $`k^{t\hspace{0pt}h}`$ finger pointing to $`n.s_{1}`$, $`n`$ can do a lookup for the id $`n - 2^{k - 1}`$. On obtaining the first successor (lets call it node $`p`$) of this id, it would immediately know if the $`k^{t\hspace{0pt}h}`$ finger of $`p`$ indeed needs to be updated. We think of each lookup as a ’correction message’. If there is more than one node that needs its $`k^{t\hspace{0pt}h}`$ successor updated (because for example, the successors of $`p`$ also happen to point to $`n.s_{1}`$), $`n`$ could leave the responsibility of informing these other nodes to $`p`$. We could take into account the probability that a correction action leads to more than $`\mathcal{M}`$ messages. But for the moment we ignore this point (We could argue that once it is $`p`$’s responsibilities to check that its successors know about $`n.s_{1}`$, it could piggy-back this information when it does a successor stabilisation, which does not affect the number of messages sent).
Whenever a node receives a message updating its information about a finger, it immediately corrects the appropriate entry in its routing table.
In the following, we demonstrate how we can analyse such a strategy. We would like to ultimately compare its performance to periodic stabilisation in the face of churn. To make such a comparisn meaningful, we need to quantify the concept of ’maintenance-effort’ per node, and compare the two schemes at a given level of churn and at the same value of the maintenance effort per node.We elaborate on this a little later in Section [](#S5.SS2 "V-B Comparison of Correction-on-change and Periodic Stabilisation ‣ V ’Correction-on-Change’ Maintenance Strategy ‣ Comparing Maintenance Strategies for Overlays This work is funded by the 6th FP EVERGROW project.")
V-B
.
Another point to note is how to quantify system performance. We have previously done it in terms of lookup hops. But a more correct way might be to ask for the latency for consistent lookups (since some of the lookups could be inconsistent). However we have checked that , within our analytical framework, this does not change the results qualiltatively.
###
V-A
Analysis of the Correction-on-change strategy
To generalise the analysis to meet the situation when some nodes are sending messages while others are not, we say that a node can be in state $`S_{1}`$ or $`S_{2}`$. In state $`S_{1}`$, a node can stabilise its first successor at rate $`\alpha\hspace{0pt}\lambda_{s}`$, fail at rate $`\lambda_{f}`$ and assist in joins at rate $`\lambda_{j}`$ as before. In state $`S_{2}`$, a node can stabilise its first successor at rate $`a\hspace{0pt}\lambda_{s}`$, fail at rate $`\lambda_{f}`$, assist in joins at rate $`\lambda_{j}`$ and in addition, send correction messages (which is essentially equivalent to doing one lookup ) at rate $`\lambda_{M} \equiv {c\hspace{0pt}\lambda_{s}}`$. As we show in Section [](#S5.SS2 "V-B Comparison of Correction-on-change and Periodic Stabilisation ‣ V ’Correction-on-Change’ Maintenance Strategy ‣ Comparing Maintenance Strategies for Overlays This work is funded by the 6th FP EVERGROW project.")
V-B
, if we want to compare the two maintenance strategies in a fair manner then the most general values that these parameters can take is $`\alpha = 1`$ and $`{a + c} = 1`$.
Let $`N_{S_{1}}`$ be the number of nodes in state $`S_{1}`$ and $`N_{S_{2}}`$ the number of nodes in state $`S_{2}`$. Clearly $`{N_{S_{1}} + N_{S_{2}}} = N`$, the total number of nodes in the system.
We can further partition $`S_{2}`$ into $`S_{2}^{1}`$,$`S_{2}^{2}`$, $`S_{2}^{3}`$, $`\cdots`$, $`S_{2}^{\mathcal{M}}`$. $`S_{2}^{1}`$ is the state of the node which has yet to send its first correction message, $`S_{2}^{2}`$ the state of the node which has sent its first correction message but is yet to send its second, etc.
Consider the gain and loss terms for $`N_{S_{1}}`$. These are summarised in table [I](#S5.T1 "TABLE I ‣ V-A Analysis of the Correction-on-change strategy ‣ V ’Correction-on-Change’ Maintenance Strategy ‣ Comparing Maintenance Strategies for Overlays This work is funded by the 6th FP EVERGROW project.").
<table>
<thead>
<tr>
<th>NS1(t+Δt)subscript𝑁subscript𝑆1𝑡Δ𝑡N_{S_{1}}(t+\Delta t)</th>
<th>Probability of Occurence</th>
</tr>
</thead>
<tbody>
<tr>
<th>=NS1(t)−1absentsubscript𝑁subscript𝑆1𝑡1=N_{S_{1}}(t)-1</th>
<td>c1.1=(λfNS1Δt)subscript𝑐1.1subscript𝜆𝑓subscript𝑁subscript𝑆1Δ𝑡c_{1.1}=(\lambda_{f}N_{S_{1}}\Delta t)</td>
</tr>
<tr>
<th>=NS1(t)+1absentsubscript𝑁subscript𝑆1𝑡1=N_{S_{1}}(t)+1</th>
<td>c1.2=(λjNΔt)subscript𝑐1.2subscript𝜆𝑗𝑁Δ𝑡c_{1.2}=(\lambda_{j}N\Delta t)</td>
</tr>
<tr>
<th>=NS1(t)+1absentsubscript𝑁subscript𝑆1𝑡1=N_{S_{1}}(t)+1</th>
<td>c1.3=(λMNS2ℳΔt)subscript𝑐1.3subscript𝜆𝑀subscript𝑁superscriptsubscript𝑆2ℳΔ𝑡c_{1.3}=(\lambda_{M}N_{S_{2}^{\cal M}}\Delta t)</td>
</tr>
<tr>
<th>=NS1−1absentsubscript𝑁subscript𝑆11=N_{S_{1}}-1</th>
<td>c1.4=(αλsNS1Δt)w1subscript𝑐1.4𝛼subscript𝜆𝑠subscript𝑁subscript𝑆1Δ𝑡subscript𝑤1c_{1.4}=(\alpha\lambda_{s}N_{S_{1}}\Delta t)w_{1}</td>
</tr>
<tr>
<th>=NS1(t)absentsubscript𝑁subscript𝑆1𝑡=N_{S_{1}}(t)</th>
<td>1−(c1.1+c1.2+c1.3+c1.4)1subscript𝑐1.1subscript𝑐1.2subscript𝑐1.3subscript𝑐1.41-(c_{1.1}+c_{1.2}+c_{1.3}+c_{1.4})</td>
</tr>
</tbody>
</table>
TABLE I: Gain and loss terms for $`N_{S_{1}}`$ the number of nodes in state $`S_{1}`$.
Term $`c_{1.1}`$ is the probability that an $`S_{1}`$ node is lost because it failed. Term $`c_{1.2}`$ is the probability that a join occurs thus adding to the number of $`S_{1}`$ nodes in the system (since a new joinee is always an $`S_{1}`$-type node). Term $`c_{1.3}`$ is the probability that an $`S_{2}^{\mathcal{M}}`$ node sent its last message at rate $`\lambda_{M}`$ and converted into an $`S_{1}`$ node. The last term $`c_{1.4}`$ is the probability that an $`S_{1}`$-type node did a stabilisation at rate $`\alpha\hspace{0pt}\lambda_{s}`$, found a wrong first successor with probability $`w_{1}`$ and hence converted into an $`S_{2}`$ node. $`w_{1}`$ is the fraction of wrong successor pointers of an $`S_{1}`$-type node.
Defining $`{\lambda_{s}/\lambda_{f}} = r`$ and $`{\lambda_{M}/\lambda_{f}} = {c\hspace{0pt}r}`$ the steady state equation predicted by table [I](#S5.T1 "TABLE I ‣ V-A Analysis of the Correction-on-change strategy ‣ V ’Correction-on-Change’ Maintenance Strategy ‣ Comparing Maintenance Strategies for Overlays This work is funded by the 6th FP EVERGROW project.") is:
<table>
<tbody>
<tr>
<td></td>
<td>PS1(1+αrw1)=1+crPS2ℳsubscript𝑃subscript𝑆11𝛼𝑟subscript𝑤11𝑐𝑟subscript𝑃superscriptsubscript𝑆2ℳP_{S_{1}}(1+\alpha rw_{1})=1+crP_{S_{2}^{\cal M}}</td>
<td></td>
<td>(3)</td>
</tr>
</tbody>
</table>
where $`P_{S_{1}} = {N_{S_{1}}/N}`$.
We can write a similar equation $`N_{S_{2}}`$ which however does not give us any new information since $`{N_{S_{1}} + N_{S_{2}}} = N`$.
Writing a gain-loss equation for each of the $`N_{S_{2}^{i}}`$’s in turn, we obtain,
<table>
<tbody>
<tr>
<td></td>
<td>PS21=PS1(αrw1−arw1′)1+cr+arw1′+arw1′1+cr+arw1′subscript𝑃superscriptsubscript𝑆21subscript𝑃subscript𝑆1𝛼𝑟subscript𝑤1𝑎𝑟superscriptsubscript𝑤1′1𝑐𝑟𝑎𝑟superscriptsubscript𝑤1′𝑎𝑟superscriptsubscript𝑤1′1𝑐𝑟𝑎𝑟superscriptsubscript𝑤1′P_{S_{2}^{1}}=\frac{P_{S_{1}}(\alpha rw_{1}-arw_{1}^{\prime})}{1+cr+arw_{1}^{\prime}}+\frac{arw_{1}^{\prime}}{1+cr+arw_{1}^{\prime}}</td>
<td></td>
<td>(4)</td>
</tr>
</tbody>
</table>
and
<table>
<tbody>
<tr>
<td></td>
<td>PS2i=PS21(cr1+cr+arw1′)i−1subscript𝑃superscriptsubscript𝑆2𝑖subscript𝑃superscriptsubscript𝑆21superscript𝑐𝑟1𝑐𝑟𝑎𝑟superscriptsubscript𝑤1′𝑖1P_{S_{2}^{i}}=P_{S_{2}^{1}}\left(\frac{cr}{1+cr+arw_{1}^{\prime}}\right)^{i-1}</td>
<td></td>
<td>(5)</td>
</tr>
</tbody>
</table>
, for $`2 \leq i \leq \mathcal{M}`$.
Here $`w_{1}`$ is the fraction of $`S_{1}`$ nodes with wrong pointers and $`w_{1}'`$ is the fraction of $`S_{2}`$ nodes with wrong pointers. We have made a simplification here in assuming that the fraction of wrong pointers of $`S_{2}`$ nodes is the same, irrespective of the state of the $`S_{2}`$ node. In practice (especially if $`a = 0`$), this will not be the case. However for the parameter ranges we are interested in ($`r\operatorname{>>}1`$), this is not crucial.
Clearly $`{\sum_{1}^{\mathcal{M}}P_{S_{2}^{i}}} = P_{S_{2}}`$. A quantity of interest in our analysis is
<table>
<tbody>
<tr>
<td></td>
<td>PS2ℳ/PS2=1−(1−g1ℳ−1)1−g1ℳsubscript𝑃superscriptsubscript𝑆2ℳsubscript𝑃subscript𝑆211superscriptsubscript𝑔1ℳ11superscriptsubscript𝑔1ℳP_{S_{2}^{\cal M}}/P_{S_{2}}=1-\frac{(1-g_{1}^{{\cal M}-1})}{1-g_{1}^{\cal M}}</td>
<td></td>
<td>(6)</td>
</tr>
</tbody>
</table>
where $`g_{1} = \frac{c\hspace{0pt}r}{({1 + {c\hspace{0pt}r} + {a\hspace{0pt}r\hspace{0pt}w_{1}'}})}`$.
To solve for $`P_{S_{1}}`$ etc, we need to solve for $`w_{1}`$ and $`w_{1}'`$.
However, consider first the equation for $`W_{T}`$ – the total number of wrong successor pointers in the system (irrespective of whether the pointer belongs to an $`S_{1}`$ or an $`S_{2}`$ type node. The gain and loss terms for $`W_{T}`$ are shown in table [II](#S5.T2 "TABLE II ‣ V-A Analysis of the Correction-on-change strategy ‣ V ’Correction-on-Change’ Maintenance Strategy ‣ Comparing Maintenance Strategies for Overlays This work is funded by the 6th FP EVERGROW project."). $`w = {W_{T}/N}`$ is the fraction of wrong succesor pointers in the system.
TABLE II: Gain and loss terms for $`W_{T}`$: the total number of wrong first successor pointers in the system.
<table>
<thead>
<tr>
<th>Change in WTsubscript𝑊𝑇W_{T}</th>
<th>Probability of Occurrence</th>
</tr>
</thead>
<tbody>
<tr>
<td>WT(t+Δt)=WT(t)+1subscript𝑊𝑇𝑡Δ𝑡subscript𝑊𝑇𝑡1W_{T}(t+\Delta t)=W_{T}(t)+1</td>
<td>c2.1=(λjNΔt)(1−w)subscript𝑐2.1subscript𝜆𝑗𝑁Δ𝑡1𝑤c_{2.1}=(\lambda_{j}N\Delta t)(1-w)</td>
</tr>
<tr>
<td>WT(t+Δt)=W1(t)+1subscript𝑊𝑇𝑡Δ𝑡subscript𝑊1𝑡1W_{T}(t+\Delta t)=W_{1}(t)+1</td>
<td>c2.2=(λfNΔt)(1−w)2subscript𝑐2.2subscript𝜆𝑓𝑁Δ𝑡superscript1𝑤2c_{2.2}=(\lambda_{f}N\Delta t)(1-w)^{2}</td>
</tr>
<tr>
<td>WT(t+Δt)=W1(t)−1subscript𝑊𝑇𝑡Δ𝑡subscript𝑊1𝑡1W_{T}(t+\Delta t)=W_{1}(t)-1</td>
<td>c2.3=(λfNΔt)subscript𝑐2.3subscript𝜆𝑓𝑁Δ𝑡c_{2.3}=(\lambda_{f}N\Delta t)</td>
</tr>
<tr>
<td>W1(t+Δt)=W1(t)−1subscript𝑊1𝑡Δ𝑡subscript𝑊1𝑡1W_{1}(t+\Delta t)=W_{1}(t)-1</td>
<td>c2.4=(αλsΔt)NS1w1+(aλsΔt)NS2w1′subscript𝑐2.4𝛼subscript𝜆𝑠Δ𝑡subscript𝑁subscript𝑆1subscript𝑤1𝑎subscript𝜆𝑠Δ𝑡subscript𝑁subscript𝑆2superscriptsubscript𝑤1′c_{2.4}=(\alpha\lambda_{s}\Delta t){N_{S_{1}}}w_{1}+(a\lambda_{s}\Delta t){N_{S_{2}}}w_{1}^{\prime}</td>
</tr>
<tr>
<td>W1(t+Δt)=W1(t)subscript𝑊1𝑡Δ𝑡subscript𝑊1𝑡W_{1}(t+\Delta t)=W_{1}(t)</td>
<td>1−(c2.1+c2.2+c2.3+c2.4)1subscript𝑐2.1subscript𝑐2.2subscript𝑐2.3subscript𝑐2.41-(c_{2.1}+c_{2.2}+c_{2.3}+c_{2.4})</td>
</tr>
</tbody>
</table>
This gives the following equation
<table>
<tbody>
<tr>
<td></td>
<td>(3+αr)w1PS1+(3+ar)w1′PS2=23𝛼𝑟subscript𝑤1subscript𝑃subscript𝑆13𝑎𝑟superscriptsubscript𝑤1′subscript𝑃subscript𝑆22(3+\alpha r)w_{1}P_{S_{1}}+(3+ar)w_{1}^{\prime}P_{S_{2}}=2</td>
<td></td>
<td>(7)</td>
</tr>
</tbody>
</table>
The gain and loss terms $`W_{1}'`$. – the number of $`S_{2}`$ nodes with wrong successor pointers – are written in much the same way except for a few small changes. Table [III](#S5.T3 "TABLE III ‣ V-A Analysis of the Correction-on-change strategy ‣ V ’Correction-on-Change’ Maintenance Strategy ‣ Comparing Maintenance Strategies for Overlays This work is funded by the 6th FP EVERGROW project.") details the changes that occur in $`W_{1}'`$. in time $`\Delta\hspace{0pt}t`$.
TABLE III: Gain and loss terms for $`W_{1}'`$: the number of wrong first successor pointers of $`S_{2}`$-type nodes.
<table>
<thead>
<tr>
<th>Change in W1subscript𝑊1W_{1}</th>
<th>Probability of Occurrence</th>
</tr>
</thead>
<tbody>
<tr>
<td>W1′(t+Δt)=W1′(t)+1superscriptsubscript𝑊1′𝑡Δ𝑡superscriptsubscript𝑊1′𝑡1W_{1}^{\prime}(t+\Delta t)=W_{1}^{\prime}(t)+1</td>
<td>c2.1=(λjNS2Δt)(1−w1′)subscript𝑐2.1subscript𝜆𝑗subscript𝑁subscript𝑆2Δ𝑡1superscriptsubscript𝑤1′c_{2.1}=(\lambda_{j}{N_{S_{2}}}\Delta t)(1-w_{1}^{\prime}).</td>
</tr>
<tr>
<td>W1′(t+Δt)=W1′(t)+1superscriptsubscript𝑊1′𝑡Δ𝑡superscriptsubscript𝑊1′𝑡1W_{1}^{\prime}(t+\Delta t)=W_{1}^{\prime}(t)+1</td>
<td>c2.2=λfNS2(1−w1′)2PS2subscript𝑐2.2subscript𝜆𝑓subscript𝑁subscript𝑆2superscript1superscriptsubscript𝑤1′2subscript𝑃subscript𝑆2c_{2.2}=\lambda_{f}{N_{S_{2}}}(1-w_{1}^{\prime})^{2}P_{S_{2}}</td>
</tr>
<tr>
<td></td>
<td>+(1−w1)(1−w1′)PS1)Δt+(1-w_{1})(1-w_{1}^{\prime})P_{S_{1}})\Delta t</td>
</tr>
<tr>
<td>W1′(t+Δt)=W1′(t)−1superscriptsubscript𝑊1′𝑡Δ𝑡superscriptsubscript𝑊1′𝑡1W_{1}^{\prime}(t+\Delta t)=W_{1}^{\prime}(t)-1</td>
<td>c2.3=λfNS2(w1′2PS2+w1w1′PS1)Δtsubscript𝑐2.3subscript𝜆𝑓subscript𝑁subscript𝑆2superscriptsuperscriptsubscript𝑤1′2subscript𝑃subscript𝑆2subscript𝑤1superscriptsubscript𝑤1′subscript𝑃subscript𝑆1Δ𝑡c_{2.3}=\lambda_{f}{N_{S_{2}}}{(w_{1}^{\prime}}^{2}P_{S_{2}}+w_{1}w_{1}^{\prime}P_{S_{1}})\Delta t</td>
</tr>
<tr>
<td>W1′(t+Δt)=W1′(t)−1superscriptsubscript𝑊1′𝑡Δ𝑡superscriptsubscript𝑊1′𝑡1W_{1}^{\prime}(t+\Delta t)=W_{1}^{\prime}(t)-1</td>
<td>c2.4=aλsNS2w1′Δtsubscript𝑐2.4𝑎subscript𝜆𝑠subscript𝑁subscript𝑆2superscriptsubscript𝑤1′Δ𝑡c_{2.4}=a\lambda_{s}{N_{S_{2}}}w_{1}^{\prime}\Delta t</td>
</tr>
<tr>
<td>W1′(t+Δt)=W1′(t)−1superscriptsubscript𝑊1′𝑡Δ𝑡superscriptsubscript𝑊1′𝑡1W_{1}^{\prime}(t+\Delta t)=W_{1}^{\prime}(t)-1</td>
<td>c2.5=λMNS2ℳw1′Δtsubscript𝑐2.5subscript𝜆𝑀superscriptsubscript𝑁subscript𝑆2ℳsuperscriptsubscript𝑤1′Δ𝑡c_{2.5}=\lambda_{M}{N_{S_{2}}^{\cal M}}w_{1}^{\prime}\Delta t</td>
</tr>
<tr>
<td>W1(t+Δt)=W1(t)subscript𝑊1𝑡Δ𝑡subscript𝑊1𝑡W_{1}(t+\Delta t)=W_{1}(t)</td>
<td>1−(c2.1+c2.2+c2.3+c2.4+c2.5)1subscript𝑐2.1subscript𝑐2.2subscript𝑐2.3subscript𝑐2.4subscript𝑐2.51-(c_{2.1}+c_{2.2}+c_{2.3}+c_{2.4}+c_{2.5})</td>
</tr>
</tbody>
</table>
The terms here are much the same as derived earlier except that we now have to keep track of whether the node that is failing (in terms $`c_{2.2}`$ and $`c_{2.3}`$) is a $`S_{1}`$ or an $`S_{2}`$-type node. In addition term $`c_{2.5}`$ is the probability that an $`S_{2}^{\mathcal{M}}`$-type node has a wrong successor pointer, but sends a message and hence turns into an $`S_{1}`$ node with a wrong pointer.
Table [III](#S5.T3 "TABLE III ‣ V-A Analysis of the Correction-on-change strategy ‣ V ’Correction-on-Change’ Maintenance Strategy ‣ Comparing Maintenance Strategies for Overlays This work is funded by the 6th FP EVERGROW project.") gives us the following equation for $`w_{1}'`$ in the steady state
<table>
<tbody>
<tr>
<td></td>
<td>2=w1′(3+ar+crPS2ℳPS2)+(w1−w1′)PS12superscriptsubscript𝑤1′3𝑎𝑟𝑐𝑟subscript𝑃superscriptsubscript𝑆2ℳsubscript𝑃subscript𝑆2subscript𝑤1superscriptsubscript𝑤1′subscript𝑃subscript𝑆12=w_{1}^{\prime}\left(3+ar+cr\frac{P_{S_{2}^{\cal M}}}{P_{S_{2}}}\right)+(w_{1}-w_{1}^{\prime})P_{S_{1}}</td>
<td></td>
<td>(8)</td>
</tr>
</tbody>
</table>
We can write a similar equation for $`w_{1}`$ which however does not contain any new information since $`w_{1}`$ and $`w_{1}'`$ satisfy equation [7](#S5.E7 "In V-A Analysis of the Correction-on-change strategy ‣ V ’Correction-on-Change’ Maintenance Strategy ‣ Comparing Maintenance Strategies for Overlays This work is funded by the 6th FP EVERGROW project.").
So in effect we have three equations, Eqn. [3](#S5.E3 "In V-A Analysis of the Correction-on-change strategy ‣ V ’Correction-on-Change’ Maintenance Strategy ‣ Comparing Maintenance Strategies for Overlays This work is funded by the 6th FP EVERGROW project."), Eq. [7](#S5.E7 "In V-A Analysis of the Correction-on-change strategy ‣ V ’Correction-on-Change’ Maintenance Strategy ‣ Comparing Maintenance Strategies for Overlays This work is funded by the 6th FP EVERGROW project.") and [8](#S5.E8 "In V-A Analysis of the Correction-on-change strategy ‣ V ’Correction-on-Change’ Maintenance Strategy ‣ Comparing Maintenance Strategies for Overlays This work is funded by the 6th FP EVERGROW project.") for three unknowns $`P_{S_{1}}`$, $`w_{1}`$ and $`w_{1}'`$. In practice this set of equations is very hard to solve exactly because of the appearance of terms such as $`g_{1}^{\mathcal{M}}`$ in Eq. [6](#S5.E6 "In V-A Analysis of the Correction-on-change strategy ‣ V ’Correction-on-Change’ Maintenance Strategy ‣ Comparing Maintenance Strategies for Overlays This work is funded by the 6th FP EVERGROW project.").
In the following we will solve the set of equation to $`O\hspace{0pt}{({1/r})}`$ by expanding Eq. [6](#S5.E6 "In V-A Analysis of the Correction-on-change strategy ‣ V ’Correction-on-Change’ Maintenance Strategy ‣ Comparing Maintenance Strategies for Overlays This work is funded by the 6th FP EVERGROW project.") to first order in $`w_{1}'`$. In this case,
<table>
<tbody>
<tr>
<td></td>
<td>PS2ℳ/PS2=1ℳ−(ℳ−12ℳ)1+arw1′crsubscript𝑃superscriptsubscript𝑆2ℳsubscript𝑃subscript𝑆21ℳℳ12ℳ1𝑎𝑟superscriptsubscript𝑤1′𝑐𝑟P_{S_{2}^{\cal M}}/P_{S_{2}}=\frac{1}{\cal M}-\left(\frac{{\cal M}-1}{2{\cal M}}\right)\frac{1+arw_{1}^{\prime}}{cr}</td>
<td></td>
<td>(9)</td>
</tr>
</tbody>
</table>
We can now solve the set of three coupled equations to get a quartic equation for $`w_{1}'`$ as a function of $`a,\alpha,\mathcal{M}`$ and $`r`$. Only one of the roots of the quartic equation is a true solution satisfying all the conditions above. The details of the calculations though straight forward are tedious and not shown here.
To calculate the cost of lookups, we still need to calculate the probability that a finger is dead. The loss and gain terms for this calculation are almost exactly the same as carried out earlier, in \[[7](#bib.bib7), [8](#bib.bib8)\] (except for term $`c_{3.2}`$) and are shown in table [IV](#S5.T4 "TABLE IV ‣ V-A Analysis of the Correction-on-change strategy ‣ V ’Correction-on-Change’ Maintenance Strategy ‣ Comparing Maintenance Strategies for Overlays This work is funded by the 6th FP EVERGROW project.").
TABLE IV: The relevant gain and loss terms for $`F_{k}`$, the number of nodes whose $`k\hspace{0pt}t\hspace{0pt}h`$ fingers are pointing to a failed node for $`k > 1`$.
<table>
<thead>
<tr>
<th>Fk(t+Δt)subscript𝐹𝑘𝑡Δ𝑡F_{k}(t+\Delta t)</th>
<th>Probability of Occurence</th>
</tr>
</thead>
<tbody>
<tr>
<th>=Fk(t)+1absentsubscript𝐹𝑘𝑡1=F_{k}(t)+1</th>
<td>c3.1=(λjNΔt)∑i=1kp𝑗𝑜𝑖𝑛(i,k)fisubscript𝑐3.1subscript𝜆𝑗𝑁Δ𝑡superscriptsubscript𝑖1𝑘subscript𝑝𝑗𝑜𝑖𝑛𝑖𝑘subscript𝑓𝑖c_{3.1}=(\lambda_{j}{N}\Delta t)\sum_{i=1}^{k}p_{\it join}(i,k)f_{i}</td>
</tr>
<tr>
<th>=Fk(t)−1absentsubscript𝐹𝑘𝑡1=F_{k}(t)-1</th>
<td>c3.2=fk∑kfk(λMNS2(1−w1′)A(w1,w1′)Δt)subscript𝑐3.2subscript𝑓𝑘subscript𝑘subscript𝑓𝑘subscript𝜆𝑀subscript𝑁subscript𝑆21superscriptsubscript𝑤1′𝐴subscript𝑤1superscriptsubscript𝑤1′Δ𝑡c_{3.2}=\frac{f_{k}}{\sum_{k}f_{k}}(\lambda_{M}{N_{S_{2}}}(1-w_{1}^{\prime})A(w_{1},w_{1}^{\prime})\Delta t)</td>
</tr>
<tr>
<th>=Fk(t)+1absentsubscript𝐹𝑘𝑡1=F_{k}(t)+1</th>
<td>c3.3=(1−fk)2[1−p1(k)](λfNΔt)subscript𝑐3.3superscript1subscript𝑓𝑘2delimited-[]1subscript𝑝1𝑘subscript𝜆𝑓𝑁Δ𝑡c_{3.3}=(1-f_{k})^{2}[1-p_{1}(k)](\lambda_{f}{N}\Delta t)</td>
</tr>
<tr>
<th>=Fk(t)+2absentsubscript𝐹𝑘𝑡2=F_{k}(t)+2</th>
<td>c3.4=(1−fk)2(p1(k)−p2(k))(λfNΔt)subscript𝑐3.4superscript1subscript𝑓𝑘2subscript𝑝1𝑘subscript𝑝2𝑘subscript𝜆𝑓𝑁Δ𝑡c_{3.4}=(1-f_{k})^{2}(p_{1}(k)-p_{2}(k))(\lambda_{f}{N}\Delta t)</td>
</tr>
<tr>
<th>=Fk(t)+3absentsubscript𝐹𝑘𝑡3=F_{k}(t)+3</th>
<td>c3.5=(1−fk)2(p2(k)−p3(k))(λfNΔt)subscript𝑐3.5superscript1subscript𝑓𝑘2subscript𝑝2𝑘subscript𝑝3𝑘subscript𝜆𝑓𝑁Δ𝑡c_{3.5}=(1-f_{k})^{2}(p_{2}(k)-p_{3}(k))(\lambda_{f}{N}\Delta t)</td>
</tr>
<tr>
<th>=Fk(t)absentsubscript𝐹𝑘𝑡=F_{k}(t)</th>
<td>1−(c3.1+c3.2+c3.3+c3.4+c3.5)1subscript𝑐3.1subscript𝑐3.2subscript𝑐3.3subscript𝑐3.4subscript𝑐3.51-(c_{3.1}+c_{3.2}+c_{3.3}+c_{3.4}+c_{3.5})</td>
</tr>
</tbody>
</table>
The term $`c_{3.2}`$ is the probability that a message is sent ($`\lambda_{M}\hspace{0pt}N_{S_{2}}`$) times the probability that a $`k^{t\hspace{0pt}h}`$ pointer gets this message (with probability $`f_{k}/{\sum f_{k}}`$ since only nodes with wrong pointers get the messages), times the probability that the message is not outdated ($`1 - w_{1}'`$), times the probability that the predecessor of the node which has to receive the message has a correct successor pointer. This last quantity is denoted by $`{A\hspace{0pt}{(w_{1},w_{1}')}} = {1 - {({{w_{1}\hspace{0pt}P_{S_{1}}} + {w_{1}'\hspace{0pt}P_{S_{2}}}})}}`$, since the predecessor could have been an $`S_{1}`$ or an $`S_{2}`$ type node.
An estimate for $`\sum f_{k}`$ is simply $`\sim {{\mathcal{M}\hspace{0pt}N_{S_{2}}}/N}`$. Substituting this in term $`c_{3.2}`$, this term becomes $`= {\lambda_{M}\hspace{0pt}N\hspace{0pt}\Delta\hspace{0pt}t\hspace{0pt}{({f_{k}/\mathcal{M}})}\hspace{0pt}{({1 - w_{1}'})}\hspace{0pt}A\hspace{0pt}{(w_{1},w_{1}')}}`$
Solving for $`f_{k}`$ in the steady state, and substituting for $`w_{1}'`$, we get $`f_{k}`$ as a function of the parameters. As mentioned earlier a quick and precise estimate of the lookup length is then obtained by taking $`L = {A\hspace{0pt}{({1 + f + {3\hspace{0pt}f^{2}}})}}`$.
###
V-B
Comparison of Correction-on-change and Periodic Stabilisation
In order to compare how the two strategies perform under churn, we need to make sure that we are comparing lookup latencies for the same number of total maintenance messages sent.
Let us assume that the maximum rate for sending messages per node is $`C`$. In the case of periodic stabilisation, this implies that the rate of doing successor stabilisations $`\lambda_{s_{1}}`$ and finger stabilisations $`\lambda_{s_{2}}`$ must in total not exceeed $`C`$. This implies that $`{{\lambda_{s_{1}}/C} + {\lambda_{s_{2}}/C}} \leq 1`$. If we assume that all nodes always send messages up to their maximum capacity, then clearly $`{{\lambda_{s_{1}}/C} + {\lambda_{s_{2}}/C}} = 1`$. Suppose we define $`r \equiv {C/\lambda_{j}}`$ and $`{r_{1} \equiv {\lambda_{s_{1}}/\lambda_{j}}},{r_{2} \equiv {\lambda_{s_{2}}/\lambda_{j}}}`$. Then for a given value of $`r`$, $`{r_{1} + r_{2}} = r`$. Hence if finger stabilisations are done at rate $`{({1 - \beta})}\hspace{0pt}r`$, the successor stabilisations need to be done at rate $`\beta\hspace{0pt}r`$, where the parameter $`\beta`$ can be varied from $`0`$ to $`1`$.
In the case of correction-on-change, we need to impose the same maximum rate $`C`$ no matter which state the nodes are in. In this case, let $`\lambda_{S_{1}}`$ be the rate of successor stabilisation in state $`S_{1}`$, $`\lambda_{S_{2}}`$ the rate of successor stabilisation in state $`S_{2}`$ and $`\lambda_{S_{3}}`$ be the rate of sending messages in state $`S_{2}`$. Clearly $`\lambda_{S_{1}} = C`$ and $`{\lambda_{S_{2}} + \lambda_{S_{3}}} = C`$. Defining $`r`$ as before, we get $`{\lambda_{s_{1}}/\lambda_{j}} = r`$ and $`{{\lambda_{s_{2}}/\lambda_{j}} + {\lambda_{s_{3}}/\lambda_{j}}} = r`$. Hence comparing with our parameters $`\alpha = 1`$ and $`{a + c} = 1`$.
In Fig. [5](#S5.F5 "Figure 5 ‣ V-B Comparison of Correction-on-change and Periodic Stabilisation ‣ V ’Correction-on-Change’ Maintenance Strategy ‣ Comparing Maintenance Strategies for Overlays This work is funded by the 6th FP EVERGROW project."), we have plotted the function $`L = {A\hspace{0pt}{({1 + f + {3\hspace{0pt}f^{2}}})}}`$ with the value of the lookup length without churn $`A = 5.846`$ for $`N = 1000`$ nodes, for $`a = 0`$ (and $`c = 1`$) and for $`\beta = 0.4`$. $`f`$ is calculated separately for the two maintenance techniques.
[Refer to caption](/html/0710.0386/assets/x5.png)
Figure 5: Comparison of the Lookup cost for the two maintenance strategies, for $`N = 1000`$.
As can be seen, correction-on-change is better than periodic stabilisation when churn is low but not when churn is high. On comparing lookup lengths for several different $`a`$, it becomes evident (see yFig. [6](#S5.F6 "Figure 6 ‣ V-B Comparison of Correction-on-change and Periodic Stabilisation ‣ V ’Correction-on-Change’ Maintenance Strategy ‣ Comparing Maintenance Strategies for Overlays This work is funded by the 6th FP EVERGROW project.")) that $`a \sim 0.2`$ is the optimum value for the correction-on-change strategy.
[Refer to caption](/html/0710.0386/assets/x6.png)
Figure 6: Comparison of the Lookup cost for different values of the parameter $`a`$, as explained in the text.
So interestingly, for nodes in state $`S_{2}`$, it is not the best strategy to increase $`c`$ as much as possible. Its a better strategy to spend some of the bandwidth on maintaining a correct successor.
## VI Summary
In summary, we have demonstrated the usefulness of the master-equation approach for understanding churn in overlay networks. Our analysis can take into account most details of the algorithms used by these networks, to provide predictions for how the performance depends on the parameters. There are several directions in which we can extend the present analysis. One of the more important ones is to model congestion on the links. This could affect the performance of the two compared maintenance strategies differently. The periodic case may not be as affected as much as the reactive case, which could suffer from congestion collapse.
Acknowledgments We would like to thank Ali Ghodsi for several very useful discussions.
## References
- \[1\]
Karl Aberer, *P-Grid: A self-organizing access structure for p2p information systems*, InProceedings of the Sixth International Conference on Cooperative Information Systems (CoopIS 2001) (Trento, Italy), 2001.
- \[2\]
Karl Aberer, Anwitaman Datta, and Manfred Hauswirth, *Efficient, self-contained handling of identity in peer-to-peer systems*, IEEE Transactions on Knowledge and Data Engineering 16 (2004), no. 7, 858–869.
- \[3\]
Luc Onana Alima, Sameh El-Ansary, Per Brand, and Seif Haridi, *DKS(N; k; f): A Family of Low Communication, Scalable and Fault-Tolerant Infrastructures for P2P Applications*, The 3rd International Workshop On Global and Peer-To-Peer Computing on Large Scale Distributed Systems (CCGRID 2003) (Tokyo, Japan), May 2003.
- \[4\]
James Aspnes, Zoë Diamadi, and Gauri Shah, *Fault-tolerant routing in peer-to-peer systems*, Proceedings of the twenty-first annual symposium on Principles of distributed computing, ACM Press, 2002, pp. 223–232.
- \[5\]
Miguel Castro, Manuel Costa, and Antony Rowstron, *Performance and dependability of structured peer-to-peer overlays*, Proceedings of the 2004 International Conference on Dependable Systems and Networks (DSN’04), IEEE Computer Society, 2004.
- \[6\]
Ali Ghodsi, Luc Onana Alima, and Seif Haridi, *Low- bandwdith topology maintenance for robustness in structured overlay networks*, 38th International HICSS Conference, Springer-Verlag, 2005.
- \[7\]
Supriya Krishnamurthy, Sameh El-Ansary, Erik Aurell, and Seif Haridi, *A statistical theory of chord under churn*, The 4th International Workshop on Peer-to-Peer Systems (IPTPS’05) (Ithaca, New York), February 2005.
- \[8\]
, *An analytical study of a strutured overlay in the presence of dynamic embership*, IEEE Joint Transaction on Networking (2007).
- \[9\]
Jinyang Li, Jeremy Stribling, Thomer M. Gil, Robert Morris, and Frans Kaashoek, *Comparing the performance of distributed hash tables under churn*, The 3rd International Workshop on Peer-to-Peer Systems (IPTPS’02) (San Diego, CA), Feb 2004.
- \[10\]
Jinyang Li, Jeremy Stribling, Robert Morris, M. Frans Kaashoek, and Thomer M. Gil, *A performance vs. cost framework for evaluating dht design tradeoffs under churn*, Proceedings of the 24th Infocom (Miami, FL), March 2005.
- \[11\]
David Liben-Nowell, Hari Balakrishnan, and David Karger, *Analysis of the evolution of peer-to-peer systems*, ACM Conf. on Principles of Distributed Computing (PODC) (Monterey, CA), July 2002.
- \[12\]
Sean Rhea, Dennis Geels, Timothy Roscoe, and John Kubiatowicz, *Handling churn in a DHT*, Proceedings of the 2004 USENIX Annual Technical Conference(USENIX ’04) (Boston, Massachusetts, USA), June 2004.
- \[13\]
Ion Stoica, Robert Morris, David Liben-Nowell, David Karger, M. Frans Kaashoek, Frank Dabek, and Hari Balakrishnan, *Chord: A scalable peer-to-peer lookup service for internet applications*, IEEE Transactions on Networking 11 (2003).
## VII Appendix
Equation [1](#S4.E1 "In IV The Lookup Equation for Chord ‣ Comparing Maintenance Strategies for Overlays This work is funded by the 6th FP EVERGROW project.") with the churn-dependent terms set to zero becomes:
<table>
<tbody>
<tr>
<td></td>
<td>Cξ+m=Cξ[1−a(m)]+a(m)+∑i=0m−1b(i)Cm−isubscript𝐶𝜉𝑚subscript𝐶𝜉delimited-[]1𝑎𝑚𝑎𝑚superscriptsubscript𝑖0𝑚1𝑏𝑖subscript𝐶𝑚𝑖C_{\xi+m}=C_{\xi}\left[1-a(m)\right]+a(m)+\sum_{i=0}^{m-1}b(i)C_{m-i}</td>
<td></td>
<td>(10)</td>
</tr>
</tbody>
</table>
After some rewriting of this, it is easily seen that the cost for any key $`i + 1`$ can be written as the following recursion relation:
<table>
<tbody>
<tr>
<td></td>
<td>Ci+1=ρCi+(1−ρ)+(1−ρ)Ci+1−ξ(i+1)subscript𝐶𝑖1𝜌subscript𝐶𝑖1𝜌1𝜌subscript𝐶𝑖1𝜉𝑖1C_{i+1}=\rho C_{i}+(1-\rho)+(1-\rho)C_{i+1-\xi(i+1)}</td>
<td></td>
<td>(11)</td>
</tr>
</tbody>
</table>
Here we have used the definition of $`a`$ and $`b`$ from the internode-interval distribution and the notation $`\xi\hspace{0pt}{({i + 1})}`$ refers to the start of the finger most closely preceding $`i + 1`$. For instance, for $`{i + 1} = 4`$, $`{\xi\hspace{0pt}{({i + 1})}} = 2`$ and for $`{i + 1} = 11`$, $`{\xi\hspace{0pt}{({i + 1})}} = 8`$ etc.
[Refer to caption](/html/0710.0386/assets/x7.png)
Figure 7: Theory and Simulation for the lookup cost without churn for a key space of size $`\mathcal{K} = 2^{14}`$ for varying $`N`$. Plotted as reference is the curve $`0.5\hspace{0pt}{\log_{2}{(N)}}`$. Note that on the y axis we have actually plotted $`L - 1`$ for convenience.
[Refer to caption](/html/0710.0386/assets/x8.png)
Figure 8: The average cost $`C_{i}`$ (the number hops for looking up an item $`i`$ keys away) in a network of $`\mathcal{N} = 1000`$ nodes and $`\mathcal{K} = 2^{20}`$ keys without churn obtained from the recurrence relation ([11](#S7.E11 "In VII Appendix ‣ Comparing Maintenance Strategies for Overlays This work is funded by the 6th FP EVERGROW project.")). The average lookup length $`L`$ is also plotted as a reference.
We are interested in solving the recursion relation and computing $`L = {\frac{1}{\mathcal{K}}\hspace{0pt}{\sum_{i = 1}^{\mathcal{K} - 1}C_{i}}}`$. To do this, we decompose this sum into the following partial sums:
<table>
<tbody>
<tr>
<td></td>
<td>s0=C1=1s1=C2s2=C3+C4s3=C5+C6+C7+C8…sℳ=C2ℳ−1+1+…+C𝒦−1subscript𝑠0subscript𝐶11subscript𝑠1subscript𝐶2subscript𝑠2subscript𝐶3subscript𝐶4subscript𝑠3subscript𝐶5subscript𝐶6subscript𝐶7subscript𝐶8…subscript𝑠ℳsubscript𝐶superscript2ℳ11…subscript𝐶𝒦1\begin{split}s_{0}&=C_{1}=1\\ s_{1}&=C_{2}\\ s_{2}&=C_{3}+C_{4}\\ s_{3}&=C_{5}+C_{6}+C_{7}+C_{8}\\ \ldots\\ s_{\cal M}&=C_{2^{{\cal M}-1}+1}+\ldots+C_{{\cal K}-1}\end{split}</td>
<td></td>
<td>(12)</td>
</tr>
</tbody>
</table>
Substituting the expressions for the $`C`$’s in the above, we find:
<table>
<tbody>
<tr>
<td></td>
<td>s0=1s1=ρ1−ρ[C1−C2]+1+s0s2=ρ1−ρ[C2−C4]+2+[s0+s1]…si=ρ1−ρ[C2i−1−C2i]+2i−1+∑j=0j−1sjsubscript𝑠01subscript𝑠1𝜌1𝜌delimited-[]subscript𝐶1subscript𝐶21subscript𝑠0subscript𝑠2𝜌1𝜌delimited-[]subscript𝐶2subscript𝐶42delimited-[]subscript𝑠0subscript𝑠1…subscript𝑠𝑖𝜌1𝜌delimited-[]subscript𝐶superscript2𝑖1subscript𝐶superscript2𝑖superscript2𝑖1superscriptsubscript𝑗0𝑗1subscript𝑠𝑗\begin{split}s_{0}&=1\\ s_{1}&=\frac{\rho}{1-\rho}[C_{1}-C_{2}]+1+s_{0}\\ s_{2}&=\frac{\rho}{1-\rho}[C_{2}-C_{4}]+2+[s_{0}+s_{1}]\\ \ldots\\ s_{i}&=\frac{\rho}{1-\rho}[C_{2^{i-1}}-C_{2^{i}}]+2^{i-1}+\sum_{j=0}^{j-1}s_{j}\end{split}</td>
<td></td>
<td>(13)</td>
</tr>
</tbody>
</table>
By substituting serially the expressions for $`s_{j}`$ (where $`0 \leq j \leq {i - 1}`$), the expression for $`s_{i}`$ (for $`i \geq 2`$) becomes:
<table>
<tbody>
<tr>
<td></td>
<td>si=ρ1−ρ[2i−2C1−C2i−∑j=1i−2si−2−jC2j]+2i+(i−1)2i−2subscript𝑠𝑖𝜌1𝜌delimited-[]superscript2𝑖2subscript𝐶1subscript𝐶superscript2𝑖superscriptsubscript𝑗1𝑖2superscript𝑠𝑖2𝑗subscript𝐶superscript2𝑗superscript2𝑖𝑖1superscript2𝑖2\begin{split}s_{i}&=\frac{\rho}{1-\rho}[2^{i-2}C_{1}-C_{2^{i}}-\sum_{j=1}^{i-2}s^{i-2-j}C_{2^{j}}]\\ &+2^{i}+(i-1)2^{i-2}\end{split}</td>
<td></td>
<td>(14)</td>
</tr>
</tbody>
</table>
Hence
<table>
<tbody>
<tr>
<td></td>
<td>∑i=0ℳsi=−ρ+[2ℳ+1−1]+ℳ2ℳ−1−[2ℳ−1]+ρ1−ρ[(2ℳ−1−1)C1−∑i=2M−1C2i−C𝒦−1−(2ℳ−2−1)C2−(2ℳ−3−1)C4−…]superscriptsubscript𝑖0ℳsubscript𝑠𝑖𝜌delimited-[]superscript2ℳ11ℳsuperscript2ℳ1delimited-[]superscript2ℳ1𝜌1𝜌delimited-[]superscript2ℳ11subscript𝐶1superscriptsubscript𝑖2𝑀1subscript𝐶superscript2𝑖subscript𝐶𝒦1superscript2ℳ21subscript𝐶2superscript2ℳ31subscript𝐶4…\begin{split}\sum_{i=0}^{\cal M}s_{i}&=-\rho+[2^{{\cal M}+1}-1]+{\cal M}2^{{\cal M}-1}-[2^{{\cal M}}-1]\\ &+\frac{\rho}{1-\rho}\biggl{[}(2^{{\cal M}-1}-1)C_{1}-\sum_{i=2}^{{\cal}M-1}C_{2^{i}}-C_{{\cal K}-1}\\ &-(2^{{\cal M}-2}-1)C_{2}-(2^{{\cal M}-3}-1)C_{4}-\dots\biggr{]}\end{split}</td>
<td></td>
<td>(15)</td>
</tr>
</tbody>
</table>
Therefore
<table>
<tbody>
<tr>
<td></td>
<td>∑i=0ℳsi=−ρ+2ℳ+ℳ2ℳ−1+ρ1−ρ[(2ℳ−1−1)C1−∑i=2M−1C2i−C𝒦−1−∑j=2ℳ−2(2ℳ−j−1)C2j−1]superscriptsubscript𝑖0ℳsubscript𝑠𝑖𝜌superscript2ℳℳsuperscript2ℳ1𝜌1𝜌delimited-[]superscript2ℳ11subscript𝐶1superscriptsubscript𝑖2𝑀1subscript𝐶superscript2𝑖subscript𝐶𝒦1superscriptsubscript𝑗2ℳ2superscript2ℳ𝑗1subscript𝐶superscript2𝑗1\begin{split}\sum_{i=0}^{\cal M}s_{i}&=-\rho+2^{\cal M}+{\cal M}2^{{\cal M}-1}\\ &+\frac{\rho}{1-\rho}\biggl{[}(2^{{\cal M}-1}-1)C_{1}-\sum_{i=2}^{{\cal}M-1}C_{2^{i}}-C_{{\cal K}-1}\\ -&\sum_{j=2}^{{\cal M}-2}(2^{{\cal M}-j}-1)C_{2^{j-1}}\biggr{]}\end{split}</td>
<td></td>
<td>(16)</td>
</tr>
</tbody>
</table>
The equation for the average lookup length without churn is thus,
<table>
<tbody>
<tr>
<td></td>
<td>L=∑s𝒦=−ρ𝒦+1+12ℳ+ρ1−ρ[2ℳ−1−1𝒦C1−1𝒦∑i=2ℳ−1C2i−1𝒦C𝒦−1−∑j=2ℳ−22ℳ−j−1𝒦C2j−1]𝐿𝑠𝒦𝜌𝒦112ℳ𝜌1𝜌delimited-[]superscript2ℳ11𝒦subscript𝐶11𝒦superscriptsubscript𝑖2ℳ1subscript𝐶superscript2𝑖1𝒦subscript𝐶𝒦1superscriptsubscript𝑗2ℳ2superscript2ℳ𝑗1𝒦subscript𝐶superscript2𝑗1\begin{split}L&=\frac{\sum s}{\cal K}\\ &=-\frac{\rho}{\cal K}+1+\frac{1}{2}{\cal M}\\ &+\frac{\rho}{1-\rho}\biggl{[}\frac{2^{{\cal M}-1}-1}{\cal K}C_{1}-\frac{1}{\cal K}\sum_{i=2}^{{\cal M}-1}C_{2^{i}}-\frac{1}{\cal K}C_{{\cal K}-1}\\ -&\sum_{j=2}^{{\cal M}-2}\frac{2^{{\cal M}-j}-1}{\cal K}C_{2^{j-1}}\biggr{]}\end{split}</td>
<td></td>
<td>(17)</td>
</tr>
</tbody>
</table>
If we can take the limit $`\mathcal{K}\rightarrow\infty`$, we can throw away some of the terms.
<table>
<tbody>
<tr>
<td></td>
<td>lim𝒦→∞L=1+12ℳ+ρ1−ρ[C12−1𝒦∑i=1ℳ−1C2i+C2𝒦−1𝒦C𝒦−1−∑j=2ℳ−22ℳ−j𝒦C2j−1+∑j=2ℳ−2C2j−1𝒦]≈1+12ℳ+ρ1−ρ[C12−C24−C48…−C2ℳ−32ℳ−2]subscript→𝒦𝐿112ℳ𝜌1𝜌delimited-[]subscript𝐶121𝒦superscriptsubscript𝑖1ℳ1subscript𝐶superscript2𝑖subscript𝐶2𝒦1𝒦subscript𝐶𝒦1superscriptsubscript𝑗2ℳ2superscript2ℳ𝑗𝒦subscript𝐶superscript2𝑗1superscriptsubscript𝑗2ℳ2subscript𝐶superscript2𝑗1𝒦112ℳ𝜌1𝜌delimited-[]subscript𝐶12subscript𝐶24subscript𝐶48…subscript𝐶superscript2ℳ3superscript2ℳ2\begin{split}\lim_{{\cal K}\to\infty}L&=1+\frac{1}{2}{\cal M}\\ &+\frac{\rho}{1-\rho}\biggl{[}\frac{C_{1}}{2}-\frac{1}{\cal K}\sum_{i=1}^{{\cal M}-1}C_{2^{i}}+\frac{C_{2}}{\cal K}-\frac{1}{\cal K}C_{{\cal K}-1}\\ -&\sum_{j=2}^{{\cal M}-2}\frac{2^{{\cal M}-j}}{\cal K}C_{2^{j-1}}+\sum_{j=2}^{{\cal M}-2}\frac{C_{2^{j-1}}}{\cal K}\biggr{]}\\ \approx&1+\frac{1}{2}{\cal M}+\frac{\rho}{1-\rho}\left[\frac{C_{1}}{2}-\frac{C_{2}}{4}-\frac{C_{4}}{8}\ldots-\frac{C_{2^{{\cal M}-3}}}{2^{{\cal M}-2}}\right]\end{split}</td>
<td></td>
<td>(18)</td>
</tr>
</tbody>
</table>
Since $`C_{1} = 1`$, we can write
<table>
<tbody>
<tr>
<td></td>
<td>L=1+12ℳ−ρ2(1−ρ)[C2−12+C4−14+…+C2ℳ−3−12ℳ−3]𝐿112ℳ𝜌21𝜌delimited-[]subscript𝐶212subscript𝐶414…subscript𝐶superscript2ℳ31superscript2ℳ3\begin{split}L&=1+\frac{1}{2}{\cal M}-\frac{\rho}{2(1-\rho)}\biggl{[}\frac{C_{2}-1}{2}+\frac{C_{4}-1}{4}+\ldots\\ +&\frac{C_{2^{{\cal M}-3}}-1}{2^{{\cal M}-3}}\biggr{]}\end{split}</td>
<td></td>
<td>(19)</td>
</tr>
</tbody>
</table>
From the recursion relation for the $`C_{i}`$’s, it is easy to see that
<table>
<tbody>
<tr>
<td></td>
<td>(Ci−1)=(1−ρ)gi(1)(ρ)+(1−ρ)2gi(2)(ρ)+…subscript𝐶𝑖11𝜌superscriptsubscript𝑔𝑖1𝜌superscript1𝜌2superscriptsubscript𝑔𝑖2𝜌…(C_{i}-1)=(1-\rho)g_{i}^{(1)}(\rho)+(1-\rho)^{2}g_{i}^{(2)}(\rho)+\ldots</td>
<td></td>
<td>(20)</td>
</tr>
</tbody>
</table>
where the $`g_{i}`$’s are functions only of $`\rho`$.
Hence if ($`1 - \rho`$) is small ($`\frac{N}{\mathcal{K}}\rightarrow 0`$), we need only compute the $`C_{i}`$’s to first order in ($`1 - \rho`$) to get the leading order effect and second order in ($`1 - \rho`$) to get the correction etc.
Hence in general the, the expression for $`L`$ is:
<table>
<tbody>
<tr>
<td></td>
<td>L=1+12ℳ−ρ2[e1(ρ)+(1−ρ)e2(ρ)+(1−ρ)2e3(ρ)…]𝐿112ℳ𝜌2delimited-[]subscript𝑒1𝜌1𝜌subscript𝑒2𝜌superscript1𝜌2subscript𝑒3𝜌…\begin{split}L&=1+\frac{1}{2}{\cal M}-\frac{\rho}{2}\biggl{[}e_{1}(\rho)+(1-\rho)e_{2}(\rho)+(1-\rho)^{2}e_{3}(\rho)\ldots\biggr{]}\end{split}</td>
<td></td>
<td>(21)</td>
</tr>
</tbody>
</table>
Where $`{e_{1}\hspace{0pt}{(\rho)}} = {\sum_{i = 1}^{\mathcal{M} - 3}{g_{2^{i}}^{(1)}\hspace{0pt}{(\rho)}}}`$ etc.
We evaluate this expression numerically by solving recursion relation ([11](#S7.E11 "In VII Appendix ‣ Comparing Maintenance Strategies for Overlays This work is funded by the 6th FP EVERGROW project.")) and compare it with simulations done at zero churn. As can be seen the prediction of the equation is very accurate (Figure [7](#S7.F7 "Figure 7 ‣ VII Appendix ‣ Comparing Maintenance Strategies for Overlays This work is funded by the 6th FP EVERGROW project.")).
Let us now compute $`e_{1}\hspace{0pt}{(\rho)}`$ to see what the leading order effect is. We now need to solve recursion relation ([11](#S7.E11 "In VII Appendix ‣ Comparing Maintenance Strategies for Overlays This work is funded by the 6th FP EVERGROW project.")) only to order $`1 - \rho`$, which gives:
<table>
<tbody>
<tr>
<td></td>
<td>C2−1=(1−ρ)C4−1=(1−ρ)[1+ρ+ρ2]C8−1=(1−ρ)[1+ρ+ρ2+⋯+ρ6]…Ci−1=(1−ρ)[1+ρ+ρ2+⋯+ρi−2]subscript𝐶211𝜌subscript𝐶411𝜌delimited-[]1𝜌superscript𝜌2subscript𝐶811𝜌delimited-[]1𝜌superscript𝜌2⋯superscript𝜌6…subscript𝐶𝑖11𝜌delimited-[]1𝜌superscript𝜌2⋯superscript𝜌𝑖2\begin{split}C_{2}-1&=(1-\rho)\\ C_{4}-1&=(1-\rho)\left[1+\rho+\rho^{2}\right]\\ C_{8}-1&=(1-\rho)\left[1+\rho+\rho^{2}+\dots+\rho^{6}\right]\\ \ldots\\ C_{i}-1&=(1-\rho)\left[1+\rho+\rho^{2}+\dots+\rho^{i-2}\right]\\ \end{split}</td>
<td></td>
<td>(22)</td>
</tr>
</tbody>
</table>
Therefore,
<table>
<tbody>
<tr>
<td></td>
<td>L=1+12ℳ+ρ2[12+1+ρ+ρ24+…]𝐿112ℳ𝜌2delimited-[]121𝜌superscript𝜌24…\begin{split}L&=1+\frac{1}{2}{\cal M}+\frac{\rho}{2}\biggl{[}\frac{1}{2}+\frac{1+\rho+\rho^{2}}{4}+\ldots\biggr{]}\end{split}</td>
<td></td>
<td>(23)</td>
</tr>
</tbody>
</table>
Consider the expression inside the brackets. We are computing this in the approximation $`\frac{N}{\mathcal{K}} = \epsilon\rightarrow 0`$, i.e. $`\rho = {1 - \epsilon}`$, therefore $`\rho^{x} = {({1 - \epsilon})}^{x} \approx e^{- {\epsilon\hspace{0pt}x}}`$. If $`x > \frac{1}{\epsilon}`$, then $`\rho^{x}\rightarrow 0`$, therefore if $`x > \frac{\mathcal{K}}{N}`$, then $`\rho^{x}\rightarrow 0`$. Hence, the terms inside the brackets become:
<table>
<tbody>
<tr>
<td></td>
<td>∑j=1T2j−12j+(2T−1)∑j=T+1ℳ−312jsuperscriptsubscript𝑗1𝑇superscript2𝑗1superscript2𝑗superscript2𝑇1superscriptsubscript𝑗𝑇1ℳ312𝑗\begin{split}\sum_{j=1}^{T}\frac{2^{j}-1}{2^{j}}+(2^{T}-1)\sum_{j=T+1}^{{\cal M}-3}\frac{1}{2}j\end{split}</td>
<td></td>
<td>(24)</td>
</tr>
</tbody>
</table>
Where $`T \equiv {{\ln_{2}\mathcal{K}} - {\ln_{2}N}}`$ and we have put $`\rho^{x} \approx 1`$ for $`x < \frac{\mathcal{K}}{N}`$ and $`\rho\rightarrow 0`$ for $`x > \frac{\mathcal{K}}{N}`$. This is clearly an overestimation and so we expect the result to over estimate the exact expression [21](#S7.E21 "In VII Appendix ‣ Comparing Maintenance Strategies for Overlays This work is funded by the 6th FP EVERGROW project.").
Expression [24](#S7.E24 "In VII Appendix ‣ Comparing Maintenance Strategies for Overlays This work is funded by the 6th FP EVERGROW project.") becomes:
<table>
<tbody>
<tr>
<td></td>
<td>T−[1−(12)ℳ−3]+[1−(12)ℳ−3−T]≈T𝑇delimited-[]1superscript12ℳ3delimited-[]1superscript12ℳ3𝑇𝑇T-\left[1-(\frac{1}{2})^{{\cal M}-3}\right]+\left[1-(\frac{1}{2})^{{\cal M}-3-T}\right]\approx T</td>
<td></td>
</tr>
</tbody>
</table>
Therefore:
<table>
<tbody>
<tr>
<td></td>
<td>L=1+12ln2𝒦−12[ln2𝒦−ln2N]≈1+12ln2N𝐿112subscript2𝒦12delimited-[]subscript2𝒦subscript2𝑁112subscript2𝑁\begin{split}L&=1+\frac{1}{2}{\ln_{2}{\cal K}}-\frac{1}{2}\left[\ln_{2}{\cal K}-\ln_{2}N\right]\\ &\approx 1+\frac{1}{2}\ln_{2}N\end{split}</td>
<td></td>
<td>(25)</td>
</tr>
</tbody>
</table>
Which is the known result for the average lookup length of Chord.
Another important parameter in the performance of DHTs in general is the base. By increasing the base, the number of fingers per node increases which leads to a shorter lookup path length. The effect of varying the base has been studied in \[[3](#bib.bib3), [10](#bib.bib10)\]. So far, we have considered in this analysis base-$`2`$ Chord. We can likewise carry out this analysis for any base.
In general, we have base-$`b`$ with $`{({b - 1})}\hspace{0pt}l\hspace{0pt}o\hspace{0pt}g_{b}\hspace{0pt}{(\mathcal{K})}`$ fingers per node. Consider as an example $`b = 4`$. Here we can define the the partial sums again in the following manner:
<table>
<tbody>
<tr>
<td></td>
<td>Δ0=s0=C1=1Δ1=s1+s2+s3Δ2=s4+s5+s6…subscriptΔ0subscript𝑠0subscript𝐶11subscriptΔ1subscript𝑠1subscript𝑠2subscript𝑠3subscriptΔ2subscript𝑠4subscript𝑠5subscript𝑠6…\begin{split}\Delta_{0}&=s_{0}=C_{1}=1\\ \Delta_{1}&=s_{1}+s_{2}+s_{3}\\ \Delta_{2}&=s_{4}+s_{5}+s_{6}\\ \ldots&\end{split}</td>
<td></td>
<td>(26)</td>
</tr>
</tbody>
</table>
where
<table>
<tbody>
<tr>
<td></td>
<td>s1=C2=ρC1+(1−ρ)+(1−ρ)C1s2=C3=ρC2+(1−ρ)+(1−ρ)C1s3=C4=ρC3+(1−ρ)+(1−ρ)C1s4=C5+C6+C7+C8s5=C9+C10+C11+C12s6=C13+C14+C15+C16…subscript𝑠1subscript𝐶2𝜌subscript𝐶11𝜌1𝜌subscript𝐶1subscript𝑠2subscript𝐶3𝜌subscript𝐶21𝜌1𝜌subscript𝐶1subscript𝑠3subscript𝐶4𝜌subscript𝐶31𝜌1𝜌subscript𝐶1subscript𝑠4subscript𝐶5subscript𝐶6subscript𝐶7subscript𝐶8subscript𝑠5subscript𝐶9subscript𝐶10subscript𝐶11subscript𝐶12subscript𝑠6subscript𝐶13subscript𝐶14subscript𝐶15subscript𝐶16…\begin{split}s_{1}&=C_{2}=\rho C_{1}+(1-\rho)+(1-\rho)C_{1}\\ s_{2}&=C_{3}=\rho C_{2}+(1-\rho)+(1-\rho)C_{1}\\ s_{3}&=C_{4}=\rho C_{3}+(1-\rho)+(1-\rho)C_{1}\\ s_{4}&=C_{5}+C_{6}+C_{7}+C_{8}\\ s_{5}&=C_{9}+C_{10}+C_{11}+C_{12}\\ s_{6}&=C_{13}+C_{14}+C_{15}+C_{16}\\ \ldots&\end{split}</td>
<td></td>
<td>(27)</td>
</tr>
</tbody>
</table>
Therefore
<table>
<tbody>
<tr>
<td></td>
<td>Δ0=C1Δ1=ρ[Δ1+C1−C4]+3(1−ρ)+3(1−ρ)[Δ0]Δ2=ρ[Δ2+C4−C16]+12(1−ρ)+3(1−ρ)[Δ0+Δ1]…subscriptΔ0subscript𝐶1subscriptΔ1𝜌delimited-[]subscriptΔ1subscript𝐶1subscript𝐶431𝜌31𝜌delimited-[]subscriptΔ0subscriptΔ2𝜌delimited-[]subscriptΔ2subscript𝐶4subscript𝐶16121𝜌31𝜌delimited-[]subscriptΔ0subscriptΔ1…\begin{split}\Delta_{0}&=C_{1}\\ \Delta_{1}&=\rho\left[\Delta_{1}+C_{1}-C_{4}\right]+3(1-\rho)+3(1-\rho)\left[\Delta_{0}\right]\\ \Delta_{2}&=\rho\left[\Delta_{2}+C_{4}-C_{16}\right]+12(1-\rho)+3(1-\rho)\left[\Delta_{0}+\Delta_{1}\right]\\ \ldots&\end{split}</td>
<td></td>
<td>(28)</td>
</tr>
</tbody>
</table>
In general for a base $`b`$, define $`B \equiv {b - 1}`$ and $`b^{\mathcal{M}} = \mathcal{K}`$. Then we have:
<table>
<tbody>
<tr>
<td></td>
<td>Δj=ρ1−ρ[Cbj−1−Cbj]+B(B+1)j−1+B[Δ0+Δ1+⋯+Δj−1]subscriptΔ𝑗𝜌1𝜌delimited-[]subscript𝐶superscript𝑏𝑗1subscript𝐶superscript𝑏𝑗𝐵superscript𝐵1𝑗1𝐵delimited-[]subscriptΔ0subscriptΔ1⋯subscriptΔ𝑗1\begin{split}\Delta_{j}&=\frac{\rho}{1-\rho}\left[C_{b^{j-1}}-C_{b^{j}}\right]\\ +&B(B+1)^{j-1}+B\left[\Delta_{0}+\Delta_{1}+\dots+\Delta_{j-1}\right]\end{split}</td>
<td></td>
<td>(29)</td>
</tr>
</tbody>
</table>
Following much the same procedure as before, we find
<table>
<tbody>
<tr>
<td></td>
<td>L=1𝒦∑j=0ℳΔj≈1+BB+1ℳ−BB+1ρ1−ρ[Cb−1B+1+Cb2−1(B+1)2+…]𝐿1𝒦superscriptsubscript𝑗0ℳsubscriptΔ𝑗1𝐵𝐵1ℳ𝐵𝐵1𝜌1𝜌delimited-[]subscript𝐶𝑏1𝐵1subscript𝐶superscript𝑏21superscript𝐵12…\begin{split}L=&\frac{1}{\cal K}\sum_{j=0}^{\cal M}\Delta_{j}\\ \approx&1+\frac{B}{B+1}{\cal M}-\frac{B}{B+1}\frac{\rho}{1-\rho}\left[\frac{C_{b}-1}{B+1}+\frac{C_{b^{2}}-1}{(B+1)^{2}}+\ldots\right]\end{split}</td>
<td></td>
<td>(30)</td>
</tr>
</tbody>
</table>
for $`\mathcal{K}\rightarrow\infty`$ as the analogue of ([19](#S7.E19 "In VII Appendix ‣ Comparing Maintenance Strategies for Overlays This work is funded by the 6th FP EVERGROW project.")). Again we can simplify and slightly overestimate the sum by assuming that $`\rho^{x} \approx 0`$ for $`x > \frac{\mathcal{K}}{N}`$ and $`\rho^{x} \approx 1`$ for $`x < \frac{\mathcal{K}}{N}`$. Then we get:
<table>
<tbody>
<tr>
<td></td>
<td>L≈1+b−1bln2Nln2b𝐿1𝑏1𝑏subscript2𝑁subscript2𝑏\begin{split}L\approx 1+\frac{b-1}{b}\frac{\ln_{2}N}{\ln_{2}b}\end{split}</td>
<td></td>
<td>(31)</td>
</tr>
</tbody>
</table>
This is the analogue of Eq. [25](#S7.E25 "In VII Appendix ‣ Comparing Maintenance Strategies for Overlays This work is funded by the 6th FP EVERGROW project.") for any base $`b`$. |
0711.0692 | cs.CR | cyber | <sup>1</sup>
<sup>1</sup>institutetext: Loria-INPL
Équipe Carte
<sup>1</sup>
<sup>1</sup>email: anne.bonfante@libertysurf.fr,Jean-Yves.Marion@loria.fr
<sup>2</sup>
<sup>2</sup>institutetext: École Nationale Supérieure
Mines de Nancy
# On the defence notion
<sup>†</sup>
<sup>†</sup>thanks: This work has been supported by project ARA Virus.
Anne Bonfante 11
Jean-Yves Marion 1122
## 1 Computer virology and art of war
“Trojan horses”, “logic bombs”, “armoured viruses” and “cryptovirology” are terms recalling war gears. In fact, concepts of attack and defence drive the world of computer virology, which looks like a war universe in an information society. This war has several shapes, from invasions of a network by worms, to military and industrial espionage…
For convenience, and because the term “virus” has a bigger impact from an epistemological point of view,
(the word “virus” was adopted after theoretical works of Cohen \[[1](#bib.bib1)\] and Adleman \[[2](#bib.bib2)\]) we will refer to “virus” or all the different kinds of computer infections. The reader may consult Filiol’s books \[[3](#bib.bib3), [16](#bib.bib16)\] for further information about the definitions and classifications of malware.
We could think at first sight that questions about computer security, and particularly about fights against computer viruses, are just a matter of highly trained security computer officers, and that a victory just depends on scientific and technical knowledge. But sometimes algorithmic and programming are not enough to catch the whole picture. Indeed, the study of malware strategies shows that they are close to the battle of the Horacii against the Curiacii…
A different point of view could then bring to computer virology a new prospect. This study is a part of a general investigation, which tries to understand viral strategies in order to anticipate attacks and improve defences, and to define the philosophical and political issues. Reading famous strategists and philosophers could give precious information about conscious- or unconscious viral practices, explicit- and implicit behaviour of the malware’s writers, whatever kind of attacks they launch. For this, a case study is a good way to make the connection between the technical or scientific profiles of a virus and the kind of strategy used (the meaning of the word ”strategy” is here conveyed by political and war philosophies). The case of the Whale virus can be extended to the Bradley concept, its epistemological model, which is meaningful.
The question of War, as well as the best way to do it and the political- and human issues related to it, is very old in the history of philosophy, as we can already find it in the Peloponnesian war \[[4](#bib.bib4)\], written by Thucydide, a Greek general and historian in 430 BC, or Plato’s Republic \[[5](#bib.bib5)\] (420-340 BC). Philosophy of war has always been an active theme of philosophy and was the main issue for some authors like Machiavelli (1469-1527), who wrote the Art of war \[[6](#bib.bib6)\] in 1519.
Are the writers of malware Machiavelli’s readers? This question sounds like a joke, but is legitimate, as we shall see. Indeed, the more dangerous viruses are the ones which seem to apply principles and to use strategies of the art of war, like Machiavelli’s, or from other war theorists like Clausewitz \[[7](#bib.bib7)\].
## 2 Bradley defences
The Whale virus challenged Scotland Yard for about two weeks during the 1990’s \[[8](#bib.bib8)\]. This virus is a typical case study, because it has its own defensive weapons. From the Whale virus, a virus shape named Bradley was proposed by Filiol \[[9](#bib.bib9)\]. This shape, which could also be considered as an epistemological model, will guide us in our study on virus defences. Indeed, it uses very complex notions coming from mathematics and computer science in order to defend and attack.
Bradley protection systems are first used to prevent the detection of the viruses by an antivirus software, and then to bypass the defences of the host system, which is under attack. We will study the two most interesting types of defence: the armoured code aspect and the virus’s one because they are typical and relevant. It is worth to notice that the furtive aspect was a Whale feature.
An armoured code consists in protecting a code from an analysis, which can be static by disassembling, or dynamic, by monitoring executions.
Protecting a code against static analysis is a scientific challenge. The paper of Barak & al \[[10](#bib.bib10)\] on obfuscation is a good illustration of these difficulties. With the benefit of hindsight, all the protection methods used lean on the same approach. The code protection is performed by an obfuscator. In fact, the latter is a kind of compiler, which transforms a source programme $`s`$ into a scrambled programme $`s'`$ in such a way that:
1. 1\.
The scrambled programme $`s'`$ computes the same thing as the source programme $`s`$. (in other words, there are semantically equivalent).
2. 2\.
The runtime of the scrambled programme $`s'`$ is close to the one of $`s`$. (that is up to a polynomial).
3. 3\.
The programme $`s'`$ is unreadable as well for a human analyst as for a de-obfuscation programme.
The third point corresponds to the obfuscation clause, which turns out to be difficult to formalize. Obfuscation consists in rewriting the code in order to make it less understandable. Cryptography plays a crucial role in this task. The interested reader may consult the recent paper of Beaucamps and Filiol \[[12](#bib.bib12)\] about the practical obfuscation methods. Moreover, a virus may mutate when it duplicates with polymorphism or metamorphism techniques.
An obfuscator protects a programme in a way that we could call passive, unlike furtivity techniques. It aims to camouflage a code in order to slow the analysis of its adversaries down. Advantage for the attacker with this method: the guarantee of immunity for a while, because the time spent to camouflage is ever shorter than the one spent by the adversary to detect it. Thus, the attacker protects itself against antivirus software, which, on the one hand search for a signature of a known malware. On the other hand, it protects itself against the work of analysts, which are trying to understand the meaning of its code. This method used by virus looks like both a camouflage and a shield, which are designed to resist an attack as long as possible. It appears then to be a method of passive defence.
A Bradley virus also combines an active defence method: ”furtivity”. It consists in deceiving a host system, by modifying, for example, system interruptions. In the case of Whale, techniques of furtivity allow to detect a debugger. The defence becomes active, because a virus can trigger an action off, if it is aware that some agent is trying to analyse it. Furtivity then allows the prevention of behavioural analysis techniques used by antivirus. Moreover, the resistance of a virus to such an analysis allows the collection of information about the defence methods of the host system. Consequently, we may suppose that this fact is a piece of its attack plan and the reasons why the virus was launched… We must then look at the reasons for which a virus like Bradley integrates so many defence techniques.
## 3 Why does a virus defend itself?
The part of the virus code which purpose is to infiltrate a system and to deceive the antivirus can be considered as both an offensive weapon (it enters a fortress) and a defensive weapon (camouflage). A particularity of Whale, as well as the Bradley concept, is that its code contains advanced defensive functionalities, which go beyond the traditional ones used to enter a system. If a virus was a mere offensive weapon, why allowing such defence mechanisms then?
If we send troops to attack a place, and whatever the issues are, it seems normal to arm them and to give them some means to protect themselves. Such defensive means are necessary in order to protect troops and avoid casualties. However, in the context of a conflict in the information society, what does the attacker try to protect?
If the only goal is to get some information back, the loss of a virus is a part of the operation. Why are viruses protected then? We think that the host systems should address those questions in order to prepare its defence.
This brings us to a first hypothesis. The attacker is a Machiavelli’s reader without knowing it. Indeed, one of the fundamental principles of the art of war is that we must have both offensive and defensive weaponries. A good defence is meaningless if it is not armed: the only guarantee of autonomy and independence is the ability to be in a defensive position as well as in an offensive one if necessary. Computer security questions should then necessarily consider the possibility to prepare and use offensive methods. However, state legislations, like the French one, do not allow attacks and force security officers to set only defensive methods with no possibility of counter-attack. The efficiency of the defence is also related to the questioning about this ban on possessing and training with its own (offensive) weapons. Machiavelli’s readers know how important it is to keep control over one’s own weapons because it is a sine qua non condition for security and independence.
The second hypothesis is that a virus provides a defence because it has some hostile intentions. If there are as sophisticated high-level defence mechanisms as the ones mentioned above, it means that the virus was made for a very important purpose. We could use this criterion in order to set a typology of viral attacks: the more elaborated the defence is, the more important the objective is.
Lastly, the third hypothesis is that we should call security policy views into question. If a virus based on the Bradley concept integrates such elaborated sequences of defence, it is because it probably expects a serious counter-attack. The defences are thus stronger than what they claim or than they are known to be (but it is maybe a trick: fooling the adversary while looking more vulnerable than you are in reality.)
Other hypotheses are conceivable, and some show that ’viral war’ takes sometimes paradoxical or new shapes we should study in order to anticipate the future.
## 4 New time scales?
Information technology war implies to have to reconsider the scales of the conflicts, particularly the time scale, but also the one of the space. The time necessary to decipher and analyse a virus (remember the two weeks which were necessary to analyse Whale!) is a “long” time. It means that it corresponds to a human time-scale. On the other hand, if we consider that a viral attack symbolically represents a hostile operation, the length of the attack is very short (between a second and a minute). There are at least two levels of analysis to bring out in order to understand the defence question:
- •
The human scale. The conception of dangerous viruses is a long process because they integrate a lot of mathematics and a big knowledge in computer science. It is hard to capture a virus and difficult to analyse it. This is the case of Whale.
- •
The computer scale. A virus is inside a system, and must fight against an anti-virus. The fight lasts a few seconds, or minutes, programme versus programme.
We could think at first sight that a viral attack is a completely new kind of war, since it takes place in an entirely different way compared to traditional war. Can we use the term of war again, since there is nothing in common between a few-second strike and a real battle in the field?
A virus writer finds time to design a virus efficiently and cannot ignore that it will require a lot of time to decipher and analyse the virus, if it is captured. Therefore, this time was anticipated and should be considered as a genuine part of the attack. It can be used to create a diversion or to have more time (indeed, during the phase of analysis, the attacker can do something else). Then this type of virus corresponds more to a traditional war, and consequently sets conflicts back in a long time-scale and not in an instantaneous one. Isn’t it the aim of the attacker? So don’t we have to consider that a viral attack is just a weapon as another one, or just a part, of a global conflict? Such a viral attack is a way to set things back in a better-known field, which is more traditional, in a way more human: the field of men, logistics and “normal” time.
This shift of scale between viral attacks and traditional fights allows us to ask further questions. For example, look at the ratio between the 14 days of analysis, which is the time necessary to understand Whale, and 1 minute, which will be considered like a time-reference at the computer level. The ratio between these scales is around a million. The change of scale directly implies that when a viral attack is launched, it cannot be controlled at the human level anymore, because it is too fast. Admittedly, the use of a gun implies that its process is autonomous and does not depend on a human (when a gun is fired, the bullet is unavoidable…except in ”Matrix”!). But in the case of a computer war, we should not consider a single virus, but a set of viruses against defences. Each of these agents communicates with the others, reacts to its environment and takes decisions, independently from any human interventions. Moreover, if the computer scale is very short, the conflict area could be the whole net \[[13](#bib.bib13)\]! This change of the scales maybe implies that we should change our classical concepts of war analysis, as we traditionally find them in books of great authors like Clausewitz \[[7](#bib.bib7)\].
## 5 A central question: the losses
Cyberwar implies to reconsider a crucial notion of every armed conflict: the one of the casualties. Indeed, why does someone program an armoured virus? What is the attacker’s objective? What may he lose, and which losses does he want to avoid?
In the setting of traditional war with different armed forces, the question of the casualties is one of the most important to make strategic and tactical choices. This question is coupled with an obvious moral issue. This issue is represented by the assessment of the inherent ratio in any operation: the (human-, technical- or practical) cost of an action must be related to the interest or to the benefit of this action. So, this assumes the responsibility of the military leaders and of the governments, particularly a moral responsibility, since casualties should be avoided. Taking a risk should then be justified with respect to the current issue, especially when human lives are concerned. The moral responsibility is more difficult to take, but it just exists in the case of real wars, because we can always know the number of deaths, without always really understanding the benefit of an operation. Cyberwar changes traditional war categories, since the moral and the casualty question as well as the question of the just cause, do not come up anymore.
Inversely, the heroic figure or the heroic action supposes that the loss of a life is there for a great collective benefit. The greater the sacrifice of the hero is, the more he takes risks; the more conscious he is that he will not survive to his action, the more heroic he is. We can ask if hackers identify to the issues of the hero, as we already mentioned, by overcoming all difficulties and responding to all challenges. Indeed, the point for a hacker is here to accomplish a technical exploit giving him the feeling of being a hero, even if his actions are reprehensible. Recent movies show that clearly: it is the heroic figure of Neo in Matrix.
It seems to us that we should understand the specificity of modern wars with respect to how they consider casualties. In the case of a virus, the loss of the latter is somehow immaterial, maybe anticipated, or even programmed by its designer. It means, first, that we do not consider losses as real ones (like events to avoid), or at least that it is a part of the strategy, especially as we can always replace a virus: losses are duplicable because a virus can be duplicated.
Another characteristic of computer-war is that it does not lean on economical and industrial resources, which are decisive factors in the case of a “real” war, even if they are not the only ones. In the case of a viral war, the attacker’s strength is not directly related to economical- or industrial resources, but to a scientific and technical knowledge, which then becomes a possible cause of war, or at least reasons for a competition.
When an attacker ”loses” a virus, the latter can be analysed. From then on, we may reproduce quite easily its design. However, there are some interests in avoiding the loss of a viral weapon in order to protect knowledge, in terms of cryptography for example. The attacker, as a war leader or a strategist, measures the interest of an action and determines the risk he runs. One of the clearest specificities of a modern or post-modern conflict is that the technical questions aim more and more at replacing the question of force. Violent confrontation is shifted to technical considerations.
## 6 Counter-attack and issues
The viral defence system often allows the attacker to stay anonymous, which prevents possible counter-attacks. But this fact changes the concept of war because the enemy is not identified. What is peculiar to war is to have an enemy, a known opponent. Once the enemy remains anonymous, confrontation does not correspond to the representation of war, but more to guerrilla warfare: the enemy is not conventional and cannot be spotted.
Furthermore, if the attacker tries to hide his strategy as long as possible, it means that the host system itself can be protected. Indeed, a virus is able to delude a defence in such a way that it can be in position to observe tools used for defence: a virus is able to be behind the line, and so is able to list all defensive tools. Therefore, this means that the virus code has a mechanism the function of which is to detect the tools used by the defence in order to analyse intruders, viruses and so on.
We could conclude for the time being that the strategy used by the virus in this case is a prospective one. Detecting the defence tools has a meaning only if another attack is planned in the near future. Consequently, we should analyse the aims of the virus from a temporal point of view. A viral attack can be an initial assault before another one, so we should not simply consider what represents a given attack at a given time.
Lastly, the issues of computer virology are directly related to other domains where computer networks are used as tools to exchange information. It is a way to say that the field of viral war is also economic and financial, and that it implies political positions, in particular questions about individual rights. Economic issues are not always perceptible behind some behaviour, which may have some moral justifications: for example, it is the case of Sony rootkit \[[15](#bib.bib15)\]. The reason of the integration of computer virology was to protect author rights, and to protect the data contained in the disk. The reality was that this technology allowed to spy the user’s data (it is of great interest from a commercial point of view…). The ambiguity is in the fact that Sony was perfectly able to justify this functionality, which is usually used by malware, in order to protect author rights. This worry is legitimate, but nothing proves that the real purpose was this one… So, computer virology is the place of different kinds of wars, but the hostility logic often remains the same.
## 7 Evolutions and perspectives
Viral defence, as a large part of computer security, depends on the results of theoretical computer science. For the record, there are Cohen’s results on viral detection undecidability: it is not possible to construct perfect anti-virus software. Other results depend on conjectures of the algorithmic complexity theory (like if P=PSPACE, then practically all cryptographic codes will be broken). It is interesting to see once more that we face a scale problem. Indeed, conjectures related to the algorithmic complexity theory imply that a given problem is computable but the runtime is beyond belief and is greater than the age of the universe! Science acts as a guarantor for defence soundness. If a defence depends on the algorithmic complexity theory, so it must take into account the time scales that we have mentioned previously. A viral defence may be conceived to resist an analysis at a human scale, and from then on can be considered as immune with respect to the computer time scale (where the time scale is about a minute). On the other hand, during a computer attack, and at a computer time scale, this heavy defence slows the virus execution down. So, the issue is to choose the right weaponry according to the targets that are aimed. This subject was widely discussed in many strategy books, as Machiavelli did in his Art of war.
As we saw it above, viral war integrates the necessity of a prospective strategy: on the one hand, because it could use the information the virus found, pursuing the elaboration of its attack’s tactics according to the environment it is located in. In this sense, a virus is a modern weapon since it is adaptable: there are interactions between a virus and its environment. There is an epistemological and scientific dimension as soon as we think about a viral attack as a deployment of a large number of autonomous and cooperative agents. This direction recalls analogies with biological immune systems \[[14](#bib.bib14)\]. Another possibility is to consider the agents of a network with divergent interests like in the algorithmic game theory. The game theory, which tries to predict behaviours of a population in a given environment, could then be used to analyse political and sociological behaviours of agents running viral attacks, and also of the users who are concerned by those attacks.
To come back to the question of the interaction of a virus in its environment, we can say it is inseparable from a forecast. That is that a virus adapts to its environment in order to pursue an attack, or to send information back with the view to make another attack. This temporal forecast, even if it is implicit, is the sign that a viral arm cannot be understood, or analysed, without thinking of the one who possesses it. The attacker’s intention (or the intention of the ones who order this action) is revealed by the question, which seems harmless at first glance: what is the purpose of a weapon? If, to be specific, it can be used to make another attack, or just to pursue the current one: this attack is a step. It also means that it is necessary for the ones who are fighting against viral attacks, not only to view an attack here and now, but always to think about the intentions that could be implied. To give another conclusion to this discussion, we should analyse a viral attack by what it implicitly discloses: the virus’s method for the attack indicates what its writer knows, or what he thinks he knows about the target he aims. Anybody thinking he is a potential target should integrate into its security policy what he wants to show to the others, in such a way that he directs the kind of attack he can be the target of. The issue at stake here is to apply the good old strategy: a false lure, a false target, or a noticeable weakness, which transforms stratagems into an art of playing with representations and to trick enemies. We must always come back to this Machiavelli’s principle: “Never believe that the enemy doesn’t know what he does”, always assume that you have to deal with a smart-, surely crafty enemy, whose intentions are not always visible. So, the security of a target should not only stem from the strength of a defence: the attackers will always succeed to penetrate by a way or another. We should also know how to use trickery, how to play with psychological effects and make our opponent believe that defences are different from what they really are…
## Acknowledgement
We would like to thank first lieutenant de Marqueissac for her valuable help in improving the correctness of our English.
## References
- \[1\] F. Cohen. Computer Viruses. PhD thesis, University of Southern California, January 1986.
- \[2\] L. Adleman. An abstract theory of computer viruses. In Crypto’88 - Advances in Cryptology, Lecture Notes in Computer Science 403, 1988.
- \[3\] E. Filiol. Computer viruses: from theory to applications, IRIS International Serie, Springer, 2004.
- \[4\] Thucydide. The Peloponnesian war.
- \[5\] Platon. The Republic.
- \[6\] N. Machiavel. The art of war.
- \[7\] C. Clausewitz. On war.
- \[8\]
E. Filiol. Whale : le virus se rebiffe. Journal de la sécurité informatique MISC 19, Mai 2005.
- \[9\] E. Filiol. Strong Cryptography Armoured Computer Viruses Forbidding Code Analysis: the Bradley Virus, In Proceedings of the 14th EICAR Conference, pp 216-227, Malta, 2005.
- \[10\] B. Barak, O. Goldreich, R. Impagliazzo, S. Rudich, A. Sahay, S. Vadhan and K. Yang, On the (Im)possibility of Obfuscating Programs. Lecture Notes in Computer Science, 2139, Crypto, 2001.
- \[11\] Cryptographie malicieuse. Journal de la sécurité informatique MISC 20, juillet 2005.
- \[12\] P. Beaucamps and E. Filiol. On the possibility of practically obfuscating programs: towards a unified perspective of code protection. TCV 2006 Special Issue, Journal of computer virology, 3 (1), pp. 3–21, 2007.
- \[13\]
I. Balepin. Superworms and Cryptovirology: a Deadly Combination. http://www.csif.cs.ucdavis.edu/~balepin/files/worms-cryptovirology.pdf, 2003.
- \[14\] Editorial, Making connections. Nature Immunology 3-10, Octobre 2002.
- \[15\]
Sony Rootkit Case: http://www.sysinternals.com/blog/2005/10/sony-rootkits-and-digital-rights.html
- \[16\] E. Filiol, Techniques virales avancées, Collection IRIS, Springer, 2007 (an English translation is due on december 2007). |
0710.5242 | cs.NI | cyber | # A Model of the IEEE 802.11 DCF in Presence of Non Ideal Transmission Channel and Capture Effects
F. Daneshgaran, M. Laddomada, F. Mesiti, and M. Mondin
Accepted to IEEE Globecom 2007This work has been supported by PRIN-2005 ICONA.F. Daneshgaran is with ECE Dept., California State University, Los Angeles, USA.M. Laddomada (laddomada@polito.it), F. Mesiti and M. Mondin are with DELEN, Politecnico di Torino, Italy.
###### Abstract
In this paper, we provide a throughput analysis of the IEEE 802.11 protocol at the data link layer in non-saturated traffic conditions taking into account the impact of both transmission channel and capture effects in Rayleigh fading environment. Impacts of both non-ideal channel and capture become important in terms of the actual observed throughput in typical network conditions whereby traffic is mainly unsaturated, specially in an environment of high interference.
We extend the multi-dimensional Markovian state transition model characterizing the behavior at the MAC layer by including transmission states that account for packet transmission failures due to errors caused by propagation through the channel, along with a state characterizing the system when there are no packets to be transmitted in the buffer of a station.
###### Index Terms:
Capture, DCF, Distributed Coordination Function, fading, IEEE 802.11, MAC, Rayleigh, rate adaptation, saturation, throughput, unsaturated, non-saturated.
## I Introduction
Wireless Local Area Networks (WLANs) using the IEEE802.11 series of standards have experienced an exponential growth in the recent past \[[1](#bib.bib1)\]-\[[13](#bib.bib13)\]. The Medium Access Control (MAC) layer of many wireless protocols resemble that of IEEE802.11. Hence, while we focus on this protocol, it is evident that the results easily extend to other protocols with similar MAC layer operation.
The most relevant works to what is presented here are \[[3](#bib.bib3), [4](#bib.bib4)\]. In \[[3](#bib.bib3)\] the author provided an analysis of the saturation throughput of the basic 802.11 protocol assuming a two dimensional Markov model at the MAC layer, while in \[[4](#bib.bib4)\] the authors extended the underlying model in order to consider unsaturated traffic conditions by introducing a new idle state, not present in the original Bianchi’s model, accounting for the case in which the station buffer is empty, after a successful completion of a packet transmission. In the modified model, however, a packet is discarded after $`m`$ backoff stages, while in the Bianchi’s model, the station keeps iterating in the $`m`$-th backoff stage until the packet gets successfully transmitted.
In \[[5](#bib.bib5)\], the authors look at the impact of channel induced errors and the received SNR on the achievable throughput in a system with rate adaptation whereby the transmission rate of the terminal is adapted based on either direct or indirect measurements of the link quality. In \[[6](#bib.bib6)\], the authors deal with the extension of Bianchi’s Markov model in order to account for channel errors. Paper \[[7](#bib.bib7)\] proposes an extension of the Bianchi’s model considering a new state for each backoff stage accounting for the absence of new packets to be transmitted, i.e., in unloaded traffic conditions.
In real networks, traffic is mostly unsaturated, so it is important to derive a model accounting for practical network operations. In this paper \[[2](#bib.bib2)\], we extend the previous works on the subject by looking at all the three issues outlined before together, namely real channel conditions, unsaturated traffic, and capture effects. Our assumptions are essentially similar to those of Bianchi \[[3](#bib.bib3)\] with the exception that we do assume the presence of both channel induced errors and capture effects due to the transmission over Rayleigh fading channel.
The paper is organized as follows. Section II extends the Markov model initially proposed by Bianchi, presenting modifications that account for transmission errors and capture effects over Rayleigh fading channels employing the 2-way handshaking technique in unsaturated traffic conditions. Section III provides an expression for the unsaturated throughput of the link. In section IV we present simulation results where typical MAC layer parameters for IEEE802.11b are used to obtain throughput values as a function of various system level parameters, capture probability, and SNR under typical traffic conditions. Finally, Section V is devoted to conclusions.
## II Development of the Markov Model
In \[[3](#bib.bib3)\], an analytical model is presented for the computation of the throughput of a WLAN using the IEEE 802.11 DCF under ideal channel conditions. By virtue of the strategy employed for reducing the collision probability of the packets transmitted from the stations attempting to access the channel simultaneously, a random process $`b\hspace{0pt}{(t)}`$ is used to represent the backoff counter of a given station. Backoff counter is decremented at the start of every idle backoff slot and when it reaches zero, the station transmits and a new value for $`b\hspace{0pt}{(t)}`$ is set.
The value of $`b\hspace{0pt}{(t)}`$ after each transmission depends on the size of the contention window from which it is drawn. Therefore it depends on the station s transmission history, rendering it a non-Markovian process. To overcome this problem and get to the definition of a Markovian process, a second process $`s\hspace{0pt}{(t)}`$ is defined representing the size of the contention window from which $`b\hspace{0pt}{(t)}`$ is drawn, $`({{W_{i} = {2^{i}\hspace{0pt}W}},{i = {s\hspace{0pt}{(t)}}}})`$.
A two-dimensional Markov process $`({s\hspace{0pt}{(t)}},{b\hspace{0pt}{(t)}})`$ can now be defined, based on two assertions:
1. 1\.
the probability $`\tau`$ that a station will attempt transmission in a generic time slot is constant across all time slots;
2. 2\.
the probability $`P_{c\hspace{0pt}o\hspace{0pt}l}`$ that any transmission experiences a collision is constant and independent of the number of collisions already suffered.
The main aim of this section is to propose an effective modification of the bi-dimensional Markov process proposed in \[[3](#bib.bib3)\] in order to account for unsaturated traffic conditions, channel error propagation and capture effects over a Rayleigh fading channel under the hypothesis of employing a 2-way handshaking access mechanism.
On the basis of this assumption, collisions can occur with probability $`P_{c\hspace{0pt}o\hspace{0pt}l}`$ on the transmitted packets, while transmission errors due to the channel, can occur with probability $`P_{e}`$. We assume that collisions and transmission error events are statistically independent. In this scenario, a packet is successfully transmitted if there is no collision (this event has probability $`1 - P_{c\hspace{0pt}o\hspace{0pt}l}`$) and the packet encounters no channel errors during transmission (this event has probability $`1 - P_{e}`$). The probability of successful transmission is therefore equal to $`{({1 - P_{e}})}\hspace{0pt}{({1 - P_{c\hspace{0pt}o\hspace{0pt}l}})}`$, from which we can set an equivalent probability of failed transmission as $`P_{e\hspace{0pt}q} = {{P_{e} + P_{c\hspace{0pt}o\hspace{0pt}l}} - {P_{e}\hspace{0pt}P_{c\hspace{0pt}o\hspace{0pt}l}}}`$.
Furthermore, in mobile radio environment, it may happen that the channel is captured by a station whose power level is stronger than other stations transmitting at the same time. This may be due to relative distances and/or channel conditions for each user and may happen whether or not the terminals exercise power control. Capture effect often reduces the collision probability on the channel since the stations whose power level at the receiver are very low due to path attenuation, shadowing and fading, are considered as interferers at the AP raising the noise floor.
[Refer to caption](/html/0710.5242/assets/x1.png)
Figure 1: Markov chain for the contention model in unsaturated traffic conditions, based on the 2-way handshaking technique, considering the effects of capture and channel induced errors.
To simplify the analysis, we make the assumption that the impact of the channel induced errors on the RTS, CTS and Acknowledgment (ACK) packets are negligible because of their short length. This is justified on the basis of the assumption that the bit errors inflicting the transmitted data are independent of each other. We note that with sufficient interleaving we can always ensure that the errors inflicting individual bits in a data packet are independent of each other \[[14](#bib.bib14), [17](#bib.bib17)\].
Practical networks operate in unsaturated traffic conditions. In this case, Bianchi’s model \[[3](#bib.bib3)\] assuming the presence of a packet to be transmitted in each and every station’s buffer, is not valid anymore. However, the simplicity of such a model can be retained also in unsaturated conditions by introducing a new state, labelled $`I`$, accounting for the following two situations:
- •
immediately after a successful transmission, the buffer of the transmitting station is empty;
- •
the station is in an idle state with an empty buffer until a new packet arrives at the buffer for transmission.
With these considerations in mind, let us discuss the Markov model shown in Fig. [1](#S2.F1 "Figure 1 ‣ II Development of the Markov Model ‣ A Model of the IEEE 802.11 DCF in Presence of Non Ideal Transmission Channel and Capture Effects"), modelling unsaturated traffic condition. The Markov Process of Fig. [1](#S2.F1 "Figure 1 ‣ II Development of the Markov Model ‣ A Model of the IEEE 802.11 DCF in Presence of Non Ideal Transmission Channel and Capture Effects") is governed by the following transition probabilities
<sup>1</sup>
<sup>1</sup>1$`P_{i,{k|{j,n}}}`$ is short for $`P\hspace{0pt}\left. \{{{{s\hspace{0pt}{({t + 1})}} = i},{{b\hspace{0pt}{({t + 1})}} = k}} \middle| {{{s\hspace{0pt}{(t)}} = j},{{b\hspace{0pt}{(t)}} = n}}\} \right.`$.
:
<table>
<tbody>
<tr>
<td></td>
<td>Pi,k|i,k+1=1,k∈[0,Wi−2],i∈[0,m]P0,k|i,0=q(1−Peq)/W0,k∈[0,W0−1],i∈[0,m]Pi,k|i−1,0=Peq/Wi,k∈[0,Wi−1],i∈[1,m]Pm,k|m,0=Peq/Wm,k∈[0,Wm−1]PI|i,0=(1−q)(1−Peq),i∈[0,m]P0,k|I=q/W0,k∈[0,W0−1]PI|I=1−qsubscript𝑃𝑖conditional𝑘𝑖𝑘1absent1formulae-sequence𝑘0subscript𝑊𝑖2𝑖0𝑚subscript𝑃0conditional𝑘𝑖0absent𝑞1subscript𝑃𝑒𝑞subscript𝑊0formulae-sequence𝑘0subscript𝑊01𝑖0𝑚subscript𝑃𝑖conditional𝑘𝑖10absentsubscript𝑃𝑒𝑞subscript𝑊𝑖formulae-sequence𝑘0subscript𝑊𝑖1𝑖1𝑚subscript𝑃𝑚conditional𝑘𝑚0absentsubscript𝑃𝑒𝑞subscript𝑊𝑚𝑘0subscript𝑊𝑚1subscript𝑃conditional𝐼𝑖0absent1𝑞1subscript𝑃𝑒𝑞𝑖0𝑚subscript𝑃0conditional𝑘𝐼absent𝑞subscript𝑊0𝑘0subscript𝑊01subscript𝑃conditional𝐼𝐼absent1𝑞missing-subexpression\small\begin{array}[]{lll}P_{i,k|i,k+1}&=1,&~{}k\in[0,W_{i}-2],~{}i\in[0,m]\\ P_{0,k|i,0}&=q(1-P_{eq})/W_{0},&~{}k\in[0,W_{0}-1],~{}i\in[0,m]\\ P_{i,k|i-1,0}&=P_{eq}/W_{i},&~{}k\in[0,W_{i}-1],~{}i\in[1,m]\\ P_{m,k|m,0}&=P_{eq}/W_{m},&~{}k\in[0,W_{m}-1]\\ P_{I|i,0}&=(1-q)(1-P_{eq}),&~{}i\in[0,m]\\ P_{0,k|I}&=q/W_{0},&~{}k\in[0,W_{0}-1]\\ P_{I|I}&=1-q&\end{array}</td>
<td></td>
<td>(1)</td>
</tr>
</tbody>
</table>
The first equation in ([1](#S2.E1 "In II Development of the Markov Model ‣ A Model of the IEEE 802.11 DCF in Presence of Non Ideal Transmission Channel and Capture Effects")) states that, at the beginning of each slot time, the backoff time is decremented. The second equation accounts for the fact that after a successful transmission, a new packet transmission starts with backoff stage 0 with probability $`q`$, in case there is a new packet in the buffer to be transmitted . Third and fourth equations deal with unsuccessful transmissions and the need to reschedule a new contention stage. The fifth equation deals with the practical situation in which after a successful transmission, the buffer of the station is empty, and as a consequence, the station transits in the idle state $`I`$ waiting for a new packet arrival. The sixth equation models the situation in which a new packet arrives in the station buffer, and a new backoff procedure is scheduled. Finally, the seventh equation models the situation in which there are no packets to be transmitted and the station is in the idle state.
## III Markovian Process Analysis and Throughput Computation
Next line of pursuit consists in finding a solution of the stationary distribution:
<table>
<tbody>
<tr>
<td></td>
<td>bi,k=limt→∞P[s(t)=i,b(t)=k],∀k∈[0,Wi−1],∀i∈[0,m]formulae-sequencesubscript𝑏𝑖𝑘subscript→𝑡𝑃delimited-[]formulae-sequence𝑠𝑡𝑖𝑏𝑡𝑘formulae-sequencefor-all𝑘0subscript𝑊𝑖1for-all𝑖0𝑚b_{i,k}=\lim_{t\rightarrow\infty}P[s(t)=i,b(t)=k],~{}\forall k\in[0,W_{i}-1],~{}\forall i\in[0,m]</td>
<td></td>
</tr>
</tbody>
</table>
that is, the probability of a station occupying a given state at any discrete time slot. First, we note the following relations:
<table>
<tbody>
<tr>
<td></td>
<td>bi,0=Peq⋅bi−1,0=Peqi⋅b0,0,∀i∈[1,m−1]bm,0=Peqm1−Peq⋅b0,0,i=msubscript𝑏𝑖0⋅subscript𝑃𝑒𝑞subscript𝑏𝑖10⋅superscriptsubscript𝑃𝑒𝑞𝑖subscript𝑏00for-all𝑖1𝑚1subscript𝑏𝑚0⋅superscriptsubscript𝑃𝑒𝑞𝑚1subscript𝑃𝑒𝑞subscript𝑏00𝑖𝑚\begin{array}[]{rcll}b_{i,0}&=&P_{eq}\cdot b_{i-1,0}=P_{eq}^{i}\cdot b_{0,0},&\forall i\in[1,m-1]\\ b_{m,0}&=&\frac{P_{eq}^{m}}{1-P_{eq}}\cdot b_{0,0},&i=m\\ \end{array}</td>
<td></td>
<td>(2)</td>
</tr>
</tbody>
</table>
whereby $`P_{e\hspace{0pt}q}`$ is the equivalent probability of failed transmission, that takes into account the need for a new contention due to either packet collision ($`P_{c\hspace{0pt}o\hspace{0pt}l}`$) or channel errors ($`P_{e}`$), i.e., $`P_{e\hspace{0pt}q} = {{P_{c\hspace{0pt}o\hspace{0pt}l} + P_{e}} - {P_{e} \cdot P_{c\hspace{0pt}o\hspace{0pt}l}}}`$.
State $`b_{I}`$ in Fig. [1](#S2.F1 "Figure 1 ‣ II Development of the Markov Model ‣ A Model of the IEEE 802.11 DCF in Presence of Non Ideal Transmission Channel and Capture Effects") considers both the situation in which after a successful transmission there are no packets to be transmitted, and the situation in which the packet queue is empty and the station is waiting for new packet arrival. The stationary probability to be in state $`b_{I}`$ can be evaluated as follows:
<table>
<tbody>
<tr>
<td></td>
<td>bI=(1−q)(1−Peq)⋅∑i=0mbi,0+(1−q)⋅bI=(1−q)(1−Peq)q⋅∑i=0mbi,0subscript𝑏𝐼⋅1𝑞1subscript𝑃𝑒𝑞superscriptsubscript𝑖0𝑚subscript𝑏𝑖0⋅1𝑞subscript𝑏𝐼missing-subexpression⋅1𝑞1subscript𝑃𝑒𝑞𝑞superscriptsubscript𝑖0𝑚subscript𝑏𝑖0\begin{array}[]{rcl}b_{I}&=&(1-q)(1-P_{eq})\cdot\sum_{i=0}^{m}b_{i,0}+(1-q)\cdot b_{I}\\ &=&\frac{(1-q)(1-P_{eq})}{q}\cdot\sum_{i=0}^{m}b_{i,0}\\ \end{array}</td>
<td></td>
<td>(3)</td>
</tr>
</tbody>
</table>
The expression above reflects the fact that state $`b_{I}`$ can be reached after a successful packet transmission from any state $`{b_{i,0},{\forall i}} \in {\lbrack 0,m\rbrack}`$ with probability $`{({1 - q})}\hspace{0pt}{({1 - P_{e\hspace{0pt}q}})}`$, or because the station is waiting in idle state with probability $`({1 - q})`$, whereby $`q`$ is the probability of having at least one packet to be transmitted in the buffer. The statistical model of $`q`$ will be discussed in the next section.
The other stationary probabilities for any $`k \in {\lbrack 1,{W_{i} - 1}\rbrack}`$ follow by resorting to the state transition diagram shown in Fig. [1](#S2.F1 "Figure 1 ‣ II Development of the Markov Model ‣ A Model of the IEEE 802.11 DCF in Presence of Non Ideal Transmission Channel and Capture Effects"):
<table>
<tbody>
<tr>
<td></td>
<td>bi,k=Wi−kWi{q(1−Peq)⋅∑i=0mbi,0+q⋅bI,i=0Peq⋅bi−1,0,i∈[1,m−1]Peq(bm−1,0+bm,0),i=msubscript𝑏𝑖𝑘subscript𝑊𝑖𝑘subscript𝑊𝑖cases⋅𝑞1subscript𝑃𝑒𝑞superscriptsubscript𝑖0𝑚subscript𝑏𝑖0⋅𝑞subscript𝑏𝐼𝑖0⋅subscript𝑃𝑒𝑞subscript𝑏𝑖10𝑖1𝑚1subscript𝑃𝑒𝑞subscript𝑏𝑚10subscript𝑏𝑚0𝑖𝑚b_{i,k}=\frac{W_{i}-k}{W_{i}}\left\{\begin{array}[]{ll}q(1-P_{eq})\cdot\sum_{i=0}^{m}b_{i,0}+q\cdot b_{I},&i=0\\ P_{eq}\cdot b_{i-1,0},&i\in[1,m-1]\\ P_{eq}(b_{m-1,0}+b_{m,0}),&i=m\\ \end{array}\right.</td>
<td></td>
<td>(4)</td>
</tr>
</tbody>
</table>
Employing the normalization condition, after lengthy algebra, and remembering the relation $`{\sum_{i = 0}^{m}b_{i,0}} = \frac{b_{0,0}}{1 - P_{e\hspace{0pt}q}}`$, it is possible to obtain:
<table>
<tbody>
<tr>
<td></td>
<td>1=b0,02[W(∑i=0m−1(2Peq)i+(2Peq)m1−Peq)+11−Peq+2(1−q)q]1subscript𝑏002delimited-[]𝑊superscriptsubscript𝑖0𝑚1superscript2subscript𝑃𝑒𝑞𝑖superscript2subscript𝑃𝑒𝑞𝑚1subscript𝑃𝑒𝑞11subscript𝑃𝑒𝑞21𝑞𝑞\small 1=\frac{b_{0,0}}{2}\left[W\left(\sum_{i=0}^{m-1}(2P_{eq})^{i}+\frac{(2P_{eq})^{m}}{1-P_{eq}}\right)+\frac{1}{1-P_{eq}}+\frac{2(1-q)}{q}\right]</td>
<td></td>
<td>(5)</td>
</tr>
</tbody>
</table>
Normalization condition yields the following equation for computation of $`b_{0,0}`$:
<table>
<tbody>
<tr>
<td></td>
<td>2(1−Peq)(1−2Peq)qq[(W+1)(1−2Peq)+WPeq(1−(2Peq)m)]+2(1−q)(1−Peq)(1−2Peq)21subscript𝑃𝑒𝑞12subscript𝑃𝑒𝑞𝑞𝑞delimited-[]𝑊112subscript𝑃𝑒𝑞𝑊subscript𝑃𝑒𝑞1superscript2subscript𝑃𝑒𝑞𝑚21𝑞1subscript𝑃𝑒𝑞12subscript𝑃𝑒𝑞missing-subexpression\begin{array}[]{ll}\frac{2(1-P_{eq})(1-2P_{eq})q}{q[(W+1)(1-2P_{eq})+WP_{eq}(1-(2P_{eq})^{m})]+2(1-q)(1-P_{eq})(1-2P_{eq})}\end{array}</td>
<td></td>
<td>(6)</td>
</tr>
</tbody>
</table>
Equ. ([6](#S3.E6 "In III Markovian Process Analysis and Throughput Computation ‣ A Model of the IEEE 802.11 DCF in Presence of Non Ideal Transmission Channel and Capture Effects")) is then used to compute $`\tau`$, the probability that a station starts a transmission in a randomly chosen time slot:
<table>
<tbody>
<tr>
<td></td>
<td>τ=2(1−2Peq)qq[(W+1)(1−2Peq)+WPeq(1−(2Peq)m)]+2(1−q)(1−Peq)(1−2Peq)𝜏212subscript𝑃𝑒𝑞𝑞𝑞delimited-[]𝑊112subscript𝑃𝑒𝑞𝑊subscript𝑃𝑒𝑞1superscript2subscript𝑃𝑒𝑞𝑚21𝑞1subscript𝑃𝑒𝑞12subscript𝑃𝑒𝑞missing-subexpression\begin{array}[]{ll}\tau=\frac{2(1-2P_{eq})q}{q[(W+1)(1-2P_{eq})+WP_{eq}(1-(2P_{eq})^{m})]+2(1-q)(1-P_{eq})(1-2P_{eq})}\end{array}</td>
<td></td>
<td>(7)</td>
</tr>
</tbody>
</table>
The collision probability needed to compute $`\tau`$ can be found considering that using a 2-way hand-shaking mechanism, a packet from a transmitting station encounters a collision if in a given time slot, at least one of the remaining $`({N - 1})`$ stations transmits simultaneously another packet, and there is no capture. In our model, we assume that capture is a subset of the collision events. This is indeed justified by the fact that there is no capture without collision, and that capture occurs only because of collisions between a certain number of transmitting stations attempting to transmit simultaneously on the channel.
<table>
<tbody>
<tr>
<td></td>
<td>Pcol=1−(1−τ)N−1−Pcapsubscript𝑃𝑐𝑜𝑙1superscript1𝜏𝑁1subscript𝑃𝑐𝑎𝑝P_{col}=1-(1-\tau)^{N-1}-P_{cap}</td>
<td></td>
<td>(8)</td>
</tr>
</tbody>
</table>
As far as the capture effects are concerned, we resort to the mathematical formulation proposed in \[[9](#bib.bib9), [10](#bib.bib10)\]. In particular, under the hypothesis of power-controlled stations in infrastructure mode, the capture probability conditioned on $`i`$ interfering frames can be defined as follows:
<table>
<tbody>
<tr>
<td></td>
<td>Pcp(γ>zog(Sf)|i)=1[1+z0g(Sf)]isubscript𝑃𝑐𝑝𝛾conditionalsubscript𝑧𝑜𝑔subscript𝑆𝑓𝑖1superscriptdelimited-[]1subscript𝑧0𝑔subscript𝑆𝑓𝑖P_{cp}\left(\gamma>z_{o}g(S_{f})|i\right)=\frac{1}{{\left[1+z_{0}g(S_{f})\right]}^{i}}</td>
<td></td>
<td>(9)</td>
</tr>
</tbody>
</table>
whereby, $`\gamma`$, defined as $`P_{u}/{\sum_{k = 1}^{i}P_{k}}`$, is the ratio of the power $`P_{u}`$ of the useful signal and the sum of the powers of the $`i`$ interfering channel contenders transmitting simultaneously $`i`$ frames, $`g\hspace{0pt}{(S_{f})}`$ is the inverse of the processing gain of the correlation receiver, and $`z_{0}`$ is the capture ratio, i.e., the value of the signal-to-interference power ratio identifying the capture threshold at the receiver. Notice that ([9](#S3.E9 "In III Markovian Process Analysis and Throughput Computation ‣ A Model of the IEEE 802.11 DCF in Presence of Non Ideal Transmission Channel and Capture Effects")) signifies the fact that capture probability corresponds to the probability that the power ratio $`\gamma`$ is above the capture threshold $`z_{0}\hspace{0pt}g\hspace{0pt}{(S_{f})}`$ which considers the inverse of the processing gain $`g\hspace{0pt}{(S_{f})}`$. For Direct Sequence Spread Spectrum (DSSS) using a 11-chip spreading factor ($`S_{f} = 11`$), we have $`{g\hspace{0pt}{(S_{f})}} = \frac{2}{3\hspace{0pt}S_{f}}`$.
Upon defining the probability of generating exactly $`i + 1`$ interfering frames over $`N`$ contending stations in a generic slot time:
<table>
<tbody>
<tr>
<td></td>
<td>(Ni+1)τi+1(1−τ)N−i−1binomial𝑁𝑖1superscript𝜏𝑖1superscript1𝜏𝑁𝑖1{N\choose i+1}\tau^{i+1}(1-\tau)^{N-i-1}</td>
<td></td>
</tr>
</tbody>
</table>
the frame capture probability $`P_{c\hspace{0pt}a\hspace{0pt}p}`$ can be obtained as follows:
<table>
<tbody>
<tr>
<td></td>
<td>Pcap=∑i=1N−1(Ni+1)τi+1(1−τ)N−i−1Pcp(γ>zog(Sf)|i)subscript𝑃𝑐𝑎𝑝superscriptsubscript𝑖1𝑁1binomial𝑁𝑖1superscript𝜏𝑖1superscript1𝜏𝑁𝑖1subscript𝑃𝑐𝑝𝛾conditionalsubscript𝑧𝑜𝑔subscript𝑆𝑓𝑖P_{cap}=\sum_{i=1}^{N-1}{N\choose i+1}\tau^{i+1}(1-\tau)^{N-i-1}P_{cp}\left(\gamma>z_{o}g(S_{f})|i\right)</td>
<td></td>
<td>(10)</td>
</tr>
</tbody>
</table>
Putting together Equ.s ([7](#S3.E7 "In III Markovian Process Analysis and Throughput Computation ‣ A Model of the IEEE 802.11 DCF in Presence of Non Ideal Transmission Channel and Capture Effects")), ([8](#S3.E8 "In III Markovian Process Analysis and Throughput Computation ‣ A Model of the IEEE 802.11 DCF in Presence of Non Ideal Transmission Channel and Capture Effects")), and ([10](#S3.E10 "In III Markovian Process Analysis and Throughput Computation ‣ A Model of the IEEE 802.11 DCF in Presence of Non Ideal Transmission Channel and Capture Effects")), along with $`P_{e\hspace{0pt}q}`$, the following nonlinear system can be defined and solved numerically, obtaining the values of $`\tau`$, $`P_{c\hspace{0pt}o\hspace{0pt}l}`$, $`P_{c\hspace{0pt}a\hspace{0pt}p}`$, and $`P_{e\hspace{0pt}q}`$:
<table>
<tbody>
<tr>
<td></td>
<td>{τ=2(1−2Peq)qq[(W+1)(1−2Peq)+WPeq(1−(2Peq)m)]+2(1−q)(1−Peq)(1−2Peq)Pcol=1−(1−τ)N−1−PcapPeq=Pcol+Pe−Pe⋅PcolPcap=∑i=1N−1(Ni+1)τi+1(1−τ)N−i−11(1+z0g(Sf))icases𝜏212subscript𝑃𝑒𝑞𝑞𝑞delimited-[]𝑊112subscript𝑃𝑒𝑞𝑊subscript𝑃𝑒𝑞1superscript2subscript𝑃𝑒𝑞𝑚21𝑞1subscript𝑃𝑒𝑞12subscript𝑃𝑒𝑞missing-subexpressionsubscript𝑃𝑐𝑜𝑙1superscript1𝜏𝑁1subscript𝑃𝑐𝑎𝑝missing-subexpressionsubscript𝑃𝑒𝑞subscript𝑃𝑐𝑜𝑙subscript𝑃𝑒⋅subscript𝑃𝑒subscript𝑃𝑐𝑜𝑙missing-subexpressionsubscript𝑃𝑐𝑎𝑝superscriptsubscript𝑖1𝑁1binomial𝑁𝑖1superscript𝜏𝑖1superscript1𝜏𝑁𝑖11superscript1subscript𝑧0𝑔subscript𝑆𝑓𝑖missing-subexpression\left\{\begin{array}[]{ll}\tau=\frac{2(1-2P_{eq})q}{q[(W+1)(1-2P_{eq})+WP_{eq}(1-(2P_{eq})^{m})]+2(1-q)(1-P_{eq})(1-2P_{eq})}\\ P_{col}=1-(1-\tau)^{N-1}-P_{cap}\\ P_{eq}=P_{col}+P_{e}-P_{e}\cdot P_{col}\\ P_{cap}=\sum_{i=1}^{N-1}{N\choose i+1}\tau^{i+1}(1-\tau)^{N-i-1}\frac{1}{{(1+z_{0}g(S_{f}))}^{i}}\end{array}\right.</td>
<td></td>
<td>(11)</td>
</tr>
</tbody>
</table>
The final step in the analysis is the computation of the normalized system throughput, defined as the fraction of time the channel is used to successfully transmit payload bits:
<table>
<tbody>
<tr>
<td></td>
<td>S=Pt⋅Ps⋅(1−Pe)E[PL](1−Pt)σ+Pt(1−Ps)Tc+PtPs(1−Pe)Ts+PtPsPeTe𝑆⋅subscript𝑃𝑡subscript𝑃𝑠1subscript𝑃𝑒𝐸delimited-[]𝑃𝐿1subscript𝑃𝑡𝜎subscript𝑃𝑡1subscript𝑃𝑠subscript𝑇𝑐subscript𝑃𝑡subscript𝑃𝑠1subscript𝑃𝑒subscript𝑇𝑠subscript𝑃𝑡subscript𝑃𝑠subscript𝑃𝑒subscript𝑇𝑒\small S=\frac{P_{t}\cdot P_{s}\cdot(1-P_{e})E[PL]}{(1-P_{t})\sigma+P_{t}(1-P_{s})T_{c}+P_{t}P_{s}(1-P_{e})T_{s}+P_{t}P_{s}P_{e}T_{e}}</td>
<td></td>
<td>(12)</td>
</tr>
</tbody>
</table>
where the meaning of the underlined symbols is as follows.
$`P_{t}`$ is the probability that there is at least one transmission in the considered time slot, with $`N`$ stations contending for the channel, each transmitting with probability $`\tau`$:
<table>
<tbody>
<tr>
<td></td>
<td>Pt=1−(1−τ)Nsubscript𝑃𝑡1superscript1𝜏𝑁P_{t}=1-(1-\tau)^{N}</td>
<td></td>
<td>(13)</td>
</tr>
</tbody>
</table>
$`P_{s}`$ is the conditional probability that a packet transmission occurring on the channel is successful. This event corresponds to the case in which exactly one station transmits in a given time slot, or two or more stations transmit simultaneously and capture by the desired station occurs. These conditions yields the following probability:
<table>
<tbody>
<tr>
<td></td>
<td>Ps=Nτ(1−τ)N−1+PcapPtsubscript𝑃𝑠𝑁𝜏superscript1𝜏𝑁1subscript𝑃𝑐𝑎𝑝subscript𝑃𝑡P_{s}=\frac{N\tau(1-\tau)^{N-1}+P_{cap}}{P_{t}}</td>
<td></td>
<td>(14)</td>
</tr>
</tbody>
</table>
$`T_{c}`$, $`T_{e}`$ and $`T_{s}`$ are the average times a channel is sensed busy due to a collision, error affected data frame transmission time and successful data frame transmission times, respectively. Knowing the time durations for ACK frames, ACK timeout, DIFS, SIFS, $`\sigma`$, data packet length ($`P\hspace{0pt}L`$) and PHY and MAC headers duration ($`H`$), and propagation delay $`\tau_{p}`$, $`T_{c}`$, $`T_{s}`$, and $`T_{e}`$ can be computed as suggested in \[[11](#bib.bib11)\]. $`E\hspace{0pt}{\lbrack{P\hspace{0pt}L}\rbrack}`$ is the average packet payload length. $`\sigma`$ is the duration of an empty time slot.
###
III-A
Modelling offered load and estimation of probability $`q`$
In our analysis, the offered load related to each station is characterized by parameter $`\lambda`$ representing the rate at which packets arrive at the station buffer from the upper layers, and measured in $`{p\hspace{0pt}k\hspace{0pt}t}/s`$. The time between two packet arrivals is defined as interarrival time, and its mean value is evaluated as $`\frac{1}{\lambda}`$. One of the most commonly used traffic models assumes packet arrival process is Poisson. The resulting interarrival times are exponentially distributed.
In the proposed model, we need a probability, identified by $`q`$, that indicates if at the end of a given transmission there is at least one packet in the queue to be transmitted. Probability $`q`$ can be well approximated in a situation with small buffer size \[[13](#bib.bib13)\] through the following relation:
<table>
<tbody>
<tr>
<td></td>
<td>q=1−e−λE[Sts]𝑞1superscript𝑒𝜆𝐸delimited-[]subscript𝑆𝑡𝑠q=1-e^{-\lambda E[S_{ts}]}</td>
<td></td>
<td>(15)</td>
</tr>
</tbody>
</table>
where, $`E\hspace{0pt}{\lbrack S_{t\hspace{0pt}s}\rbrack}`$ is the expected time per slot, which is useful to relate the state of the Markov chain with the actual time spent in each state.
A more accurate model can be derived upon considering different values of $`q`$ for each backoff state. However, a reasonable solution consists in using a mean probability valid for the whole Markov model \[[13](#bib.bib13)\], derived from $`E\hspace{0pt}{\lbrack S_{t\hspace{0pt}s}\rbrack}`$. The value of $`E\hspace{0pt}{\lbrack S_{t\hspace{0pt}s}\rbrack}`$ can be obtained by resorting to the durations and the respective probabilities of the idle slot ($`\sigma`$), the successful transmission slot ($`T_{s}`$), the error slot due to collision ($`T_{c}`$), and the error slot due to channel ($`T_{e}`$), as follows:
<table>
<tbody>
<tr>
<td></td>
<td>E[Sts]=(1−Pt)⋅σ+Pt(1−Ps)⋅Tc++PtPsPe⋅Te+PtPs(1−Pe)⋅Ts𝐸delimited-[]subscript𝑆𝑡𝑠⋅1subscript𝑃𝑡𝜎limit-from⋅subscript𝑃𝑡1subscript𝑃𝑠subscript𝑇𝑐missing-subexpressionmissing-subexpression⋅subscript𝑃𝑡subscript𝑃𝑠subscript𝑃𝑒subscript𝑇𝑒⋅subscript𝑃𝑡subscript𝑃𝑠1subscript𝑃𝑒subscript𝑇𝑠\begin{array}[]{rcl}E[S_{ts}]&=&(1-P_{t})\cdot\sigma+P_{t}(1-P_{s})\cdot T_{c}+\\ &&+P_{t}P_{s}P_{e}\cdot T_{e}+P_{t}P_{s}(1-P_{e})\cdot T_{s}\end{array}</td>
<td></td>
<td>(16)</td>
</tr>
</tbody>
</table>
Upon recalling that packet inter-arrival times are exponentially distributed, we can use the average slot time to calculate the probability $`q`$ that in such a time interval a given station receives a packet from upper layers in its transmission queue. The probability that in a generic time $`T`$, $`k`$ events occur, is:
<table>
<tbody>
<tr>
<td></td>
<td>P{a(T)=k}=e−λT(λT)kk!𝑃𝑎𝑇𝑘superscript𝑒𝜆𝑇superscript𝜆𝑇𝑘𝑘P\{a(T)=k\}=e^{-\lambda T}\frac{(\lambda T)^{k}}{k!}</td>
<td></td>
<td>(17)</td>
</tr>
</tbody>
</table>
from which we obtain the relation ([15](#S3.E15 "In III-A Modelling offered load and estimation of probability q ‣ III Markovian Process Analysis and Throughput Computation ‣ A Model of the IEEE 802.11 DCF in Presence of Non Ideal Transmission Channel and Capture Effects")) referred to earlier:
<table>
<tbody>
<tr>
<td></td>
<td>q=1−P{a(E[Sts])=0}=1−e−λE[Sts]𝑞1𝑃𝑎𝐸delimited-[]subscript𝑆𝑡𝑠01superscript𝑒𝜆𝐸delimited-[]subscript𝑆𝑡𝑠q=1-P\{a(E[S_{ts}])=0\}=1-e^{-\lambda E[S_{ts}]}</td>
<td></td>
<td>(18)</td>
</tr>
</tbody>
</table>
## IV Simulation Results and Model Validations
TABLE I: Typical network parameters
<table>
<tbody>
<tr>
<th>MAC header</th>
<th>24 bytes</th>
<td>τpsubscript𝜏𝑝\tau_{p}</td>
<td>1 μ𝜇\mus</td>
</tr>
<tr>
<th>PHY header</th>
<th>16 bytes</th>
<td>Slot time</td>
<td>20 μ𝜇\mus</td>
</tr>
<tr>
<th>Payload size, E[PL]</th>
<th>1024 bytes</th>
<td>SIFS</td>
<td>10 μ𝜇\mus</td>
</tr>
<tr>
<th>ACK</th>
<th>14 bytes</th>
<td>DIFS</td>
<td>50 μ𝜇\mus</td>
</tr>
<tr>
<th>RTS</th>
<th>20 bytes</th>
<td>EIFS</td>
<td>300 μ𝜇\mus</td>
</tr>
<tr>
<th>CTS</th>
<th>14 bytes</th>
<td>ACK,CTS timeout</td>
<td>300 μ𝜇\mus</td>
</tr>
<tr>
<th>m𝑚m</th>
<th>5</th>
<td>W𝑊W</td>
<td>32</td>
</tr>
<tr>
<th>Tssubscript𝑇𝑠T_{s}</th>
<th>8.812 ms</th>
<td>Tcsubscript𝑇𝑐T_{c}</td>
<td>8.812 ms</td>
</tr>
</tbody>
</table>
This section focuses on simulation results for validating the theoretical models and derivations presented in the previous sections. We have developed a C++ simulator modelling both the DCF protocol details in 802.11b and the backoff procedures of a specific number of independent transmitting stations. The simulator also takes into account all real operations of each transmitting station, including physical propagation delays, etc.
Typical MAC layer parameters for the lowest rate IEEE802.11b are given in Table [I](#S4.T1 "TABLE I ‣ IV Simulation Results and Model Validations ‣ A Model of the IEEE 802.11 DCF in Presence of Non Ideal Transmission Channel and Capture Effects") \[[1](#bib.bib1)\]. In so far as the computation of the FER is concerned, it should be noted that data transmission rate of various packet types differ. For simplicity, we assume that data packets transmitted by different stations are affected by the same FER.
The FER as a function of the SNR can be computed as follows:
<table>
<tbody>
<tr>
<td></td>
<td>Pe(SNR)=1−[1−Pe(PLCP,SNR)]⋅[1−Pe(DATA,SNR)]subscript𝑃𝑒𝑆𝑁𝑅1⋅delimited-[]1subscript𝑃𝑒𝑃𝐿𝐶𝑃𝑆𝑁𝑅delimited-[]1subscript𝑃𝑒𝐷𝐴𝑇𝐴𝑆𝑁𝑅\small P_{e}(SNR)=1-\left[1-P_{e}(PLCP,SNR)\right]\cdot\left[1-P_{e}(DATA,SNR)\right]</td>
<td></td>
<td>(19)</td>
</tr>
</tbody>
</table>
where,
<table>
<tbody>
<tr>
<td></td>
<td>Pe(PLCP,SNR)=1−[1−Pb(BPSK,SNR)]8×PLCP,subscript𝑃𝑒𝑃𝐿𝐶𝑃𝑆𝑁𝑅1superscriptdelimited-[]1subscript𝑃𝑏𝐵𝑃𝑆𝐾𝑆𝑁𝑅8𝑃𝐿𝐶𝑃\small P_{e}(PLCP,SNR)=1-\left[1-P_{b}(BPSK,SNR)\right]^{8\times PLCP},</td>
<td></td>
<td>(20)</td>
</tr>
</tbody>
</table>
and
<table>
<tbody>
<tr>
<td></td>
<td>Pe(DATA,SNR)=1−[1−Pb(TYPE,SNR)]8×(DATA+MAC).subscript𝑃𝑒𝐷𝐴𝑇𝐴𝑆𝑁𝑅1superscriptdelimited-[]1subscript𝑃𝑏𝑇𝑌𝑃𝐸𝑆𝑁𝑅8𝐷𝐴𝑇𝐴𝑀𝐴𝐶\small P_{e}(DATA,SNR)=1-\left[1-P_{b}(TYPE,SNR)\right]^{8\times(DATA+MAC)}.</td>
<td></td>
<td>(21)</td>
</tr>
</tbody>
</table>
$`P_{b}\hspace{0pt}{({B\hspace{0pt}P\hspace{0pt}S\hspace{0pt}K},{S\hspace{0pt}N\hspace{0pt}R})}`$ is the BER as a function of SNR for the lowest data transmit rate employing DBPSK modulation, DATA denotes the data block size in bytes, and any other constant byte size in above expression represents overhead. Note that the FER, $`P_{e}\hspace{0pt}{({S\hspace{0pt}N\hspace{0pt}R})}`$, implicitly depends on the modulation format used. Hence, for each supported rate, one curve for $`P_{e}\hspace{0pt}{({S\hspace{0pt}N\hspace{0pt}R})}`$ as a function of SNR can be generated. $`P_{b}\hspace{0pt}{({T\hspace{0pt}Y\hspace{0pt}P\hspace{0pt}E},{S\hspace{0pt}N\hspace{0pt}R})}`$ is modulation dependent whereby the parameter $`T\hspace{0pt}Y\hspace{0pt}P\hspace{0pt}E`$ can be any of the following $`{T\hspace{0pt}Y\hspace{0pt}P\hspace{0pt}E} \in {\{{D\hspace{0pt}B\hspace{0pt}P\hspace{0pt}S\hspace{0pt}K},{D\hspace{0pt}Q\hspace{0pt}P\hspace{0pt}S\hspace{0pt}K},{C\hspace{0pt}C\hspace{0pt}K\hspace{0pt}5.5},{C\hspace{0pt}C\hspace{0pt}K\hspace{0pt}11}\}}`$
<sup>2</sup>
<sup>2</sup>2The acronyms are short for Differential Binary Phase Shift Keying, Differential Quadrature Phase Shift Keying and Complementary Code Keying, respectively.
.
For DBPSK and DQPSK modulation formats, $`P_{b}\hspace{0pt}{({T\hspace{0pt}Y\hspace{0pt}P\hspace{0pt}E},{S\hspace{0pt}N\hspace{0pt}R})}`$ can be well approximated by \[[18](#bib.bib18)\]:
<table>
<tbody>
<tr>
<td></td>
<td>2max(log2M,2)∑i=1max(M4,1)1π∫0π211+γ1sin2θlog2Msin2((2i−1)πM)𝑑θ2subscript2𝑀2superscriptsubscript𝑖1𝑀411𝜋superscriptsubscript0𝜋211𝛾1superscript2𝜃subscript2𝑀superscript22𝑖1𝜋𝑀differential-d𝜃\small\frac{2}{\max(\log_{2}M,2)}\sum_{i=1}^{\max(\frac{M}{4},1)}\frac{1}{\pi}\int_{0}^{\frac{\pi}{2}}\frac{1}{1+\gamma\frac{1}{\sin^{2}\theta}\log_{2}M\sin^{2}\left(\frac{(2i-1)\pi}{M}\right)}d\theta</td>
<td></td>
<td>(22)</td>
</tr>
</tbody>
</table>
whereby $`M`$ is the number of bits per modulated symbols, $`\gamma`$ is the signal-to-noise ratio, and $`\theta`$ is the signal direction over the Rayleigh fading channel.
[Refer to caption](/html/0710.5242/assets/x2.png)
Figure 2: Theoretical and simulated throughput for the 2-way mechanism as a function of the packet rate $`\lambda`$, for three different number of contending stations and two different values of SNR as noted in the legend. Capture thresholds is $`z_{0} = 6`$dB. Simulated points are identified by star-markers over the respective theoretical curves. Payload size is 1024 bytes.
Fig. [2](#S4.F2 "Figure 2 ‣ IV Simulation Results and Model Validations ‣ A Model of the IEEE 802.11 DCF in Presence of Non Ideal Transmission Channel and Capture Effects") shows the behavior of the throughput as a function of $`\lambda`$, i.e., the packet rate, for three different values of the number of contending stations and for two values of SNR. The capture threshold is $`z_{0} = 6`$dB. Beside noting the throughput improvement achievable for high SNR, notice that for a specified number of contending stations, the throughput manifests a linear behaviour for low values of packet rates with a slope depending mainly on the number of stations $`N`$. However, for increasing values of $`\lambda`$, the saturation behavior occurs quite fast. Notice that, as exemplified in ([18](#S3.E18 "In III-A Modelling offered load and estimation of probability q ‣ III Markovian Process Analysis and Throughput Computation ‣ A Model of the IEEE 802.11 DCF in Presence of Non Ideal Transmission Channel and Capture Effects")), $`q\rightarrow 1`$ as $`\lambda\rightarrow\infty`$. Actually, saturated traffic conditions are achieved quite fast for values of $`\lambda`$ on the order of ten packets per second with a number of contending stations greater than or equal to 10.
Fig. [3](#S4.F3 "Figure 3 ‣ IV Simulation Results and Model Validations ‣ A Model of the IEEE 802.11 DCF in Presence of Non Ideal Transmission Channel and Capture Effects") shows the behavior of the throughput as a function of $`\lambda`$, for three different values of the number of contending stations and for two different SNRs. The capture threshold is $`z_{0} = 24`$dB. We can draw conclusions similar to the ones derived for Fig. [2](#S4.F2 "Figure 2 ‣ IV Simulation Results and Model Validations ‣ A Model of the IEEE 802.11 DCF in Presence of Non Ideal Transmission Channel and Capture Effects"). Upon comparing the curves shown in Fig.s [2](#S4.F2 "Figure 2 ‣ IV Simulation Results and Model Validations ‣ A Model of the IEEE 802.11 DCF in Presence of Non Ideal Transmission Channel and Capture Effects") and [3](#S4.F3 "Figure 3 ‣ IV Simulation Results and Model Validations ‣ A Model of the IEEE 802.11 DCF in Presence of Non Ideal Transmission Channel and Capture Effects"), it is easily seen that capture effects allow the system throughput to be almost the same independently from the number of stations in saturated conditions, i.e., for high values of $`\lambda`$.
Fig. [3](#S4.F3 "Figure 3 ‣ IV Simulation Results and Model Validations ‣ A Model of the IEEE 802.11 DCF in Presence of Non Ideal Transmission Channel and Capture Effects") also shows the presence of a peak in the throughput as a function of $`\lambda`$, which characterizes the transition between the linear and saturated throughput. Such a peak tends to manifest itself for increasing values of $`\lambda`$ as the number of stations $`N`$ decreases. A comparative analysis of the curves shown in Fig.s [2](#S4.F2 "Figure 2 ‣ IV Simulation Results and Model Validations ‣ A Model of the IEEE 802.11 DCF in Presence of Non Ideal Transmission Channel and Capture Effects") and [3](#S4.F3 "Figure 3 ‣ IV Simulation Results and Model Validations ‣ A Model of the IEEE 802.11 DCF in Presence of Non Ideal Transmission Channel and Capture Effects") reveals that the peak of the throughput tends to disappear because of the presence of capture effects during transmission.
[Refer to caption](/html/0710.5242/assets/x3.png)
Figure 3: Theoretical and simulated throughput for the 2-way mechanism as a function of the packet rate $`\lambda`$, for three different number of contending stations and two different values of SNR as noted in the legend. Capture threshold is $`z_{0} = 24`$dB. Simulated points are identified by star-markers over the respective theoretical curves. Payload size is 1024 bytes.
## V Conclusions
In this paper, we have provided an extension of the Markov model characterizing the DCF behavior at the MAC layer of the IEEE802.11 series of standards by accounting for channel induced errors and capture effects typical of fading environments under unsaturated traffic conditions. The modelling allows taking into consideration the impact of channel contention in throughput analysis which is often not considered or it is considered in a static mode by using a mean contention period. Simulation results confirm the validity of the proposed theoretical models.
## References
- \[1\]
IEEE Standard for Wireless LAN Medium Access Control (MAC) and Physical Layer (PHY) Specifications, November 1997, P802.11
- \[2\]
F. Daneshgaran, M. Laddomada, F. Mesiti, and M. Mondin, ”Unsaturated Throughput Analysis of IEEE 802.11 in Presence of Non Ideal Transmission Channel and Capture Effects”, IEEE Trans. on Wireless Communications, 2008.
- \[3\]
G. Bianchi, ”Performance analysis of the IEEE 802.11 distributed coordination function”, IEEE JSAC, Vol.18, No.3, March 2000.
- \[4\]
L. Yong Shyang, A. Dadej, and A.Jayasuriya, ”Performance analysis of IEEE 802.11 DCF under limited load”, In Proc. of Asia-Pacific Conference on Communications, Vol.1, pp.759 - 763, 03-05 Oct. 2005.
- \[5\]
D. Qiao, S. Choi, and K.G. Shin ”Goodput analysis and link adaptation for IEEE 802.11a wireless LANs”, IEEE Trans. On Mobile Computing, Vol.1, No.4, Oct.-Dec. 2002.
- \[6\]
P. Chatzimisios, A.C. Boucouvalas, and V. Vitsas, ”Influence of channel BER on IEEE 802.11 DCF”, IEE Electronics Letters, Vol.39, No.23, pp.1687-1689, Nov. 2003.
- \[7\]
K. Duffy, D. Malone, and D.J. Leith, ”Modeling the 802.11 distributed coordination function in non-saturated conditions”, IEEE Communications Letters, Vol.9, No.8, pp.715-717, Aug. 2005.
- \[8\]
Jae Hyun Kim and Jong Kyu Lee ”Capture effects of wireless CSMA/CA protocols in Rayleigh and shadow fading channels”, IEEE Trans. on Veh. Tech., Vol.48, No.4, pp.1277-1286, July 1999.
- \[9\]
M. Zorzi and R.R. Rao, ”Capture and retransmission control in mobile radio”, IEEE JSAC, Vol.12, No.8, pp.1289 - 1298, Oct. 1994.
- \[10\]
Z. Hadzi-Velkov and B. Spasenovski, ”Capture effect in IEEE 802.11 basic service area under influence of Rayleigh fading and near/far effect”, In Proc. of 13th IEEE PMRC, Vol.1, pp.172 - 176, Sept. 2002.
- \[11\]
Z. Kong, D.H.K. Tsang, B.Bensaou, and D.Gao, ”Performance analysis of IEEE 802.11e contention-based channel access”, IEEE JSAC, Vol.22, No.10, Dec. 2004.
- \[12\]
T. S. Ho and K. C. Chen, ”Performance evaluation and enhancement of the CSMA/CA MAC protocol for 802.11 wireless LAN s”, In Proc. of IEEE PIMRC, Taipei, Taiwan, Oct. 1996, pp. 392 396.
- \[13\]
D. Malone, K. Duffy, and D.J. Leith, ”Modeling the 802.11 distributed coordination function in non-saturated heterogeneous conditions”, To appear on IEEE-ACM Trans. on Networking.
- \[14\]
F. Daneshgaran and M. Laddomada, “Reduced complexity interleaver growth algorithm for turbo codes,” IEEE Trans. on Wireless Communications, vol.4, no.3, pp.954-964, May 2005.
- \[15\]
F. Daneshgaran, M. Laddomada, and M. Mondin, “Interleaver design for serially concatenated convolutional codes: Theory and application,” IEEE Trans. on Inf. Theory, Vol.50, No. 6, pp. 1177-1188, June 2004.
- \[16\]
F. Daneshgaran, M. Laddomada, and M. Mondin, “An extensive search for good punctured rate k/k+1 recursive convolutional codes for serially concatenated convolutional codes,” IEEE Trans. on Inform. Theory, vol.50, no.1, pp.208-217, January 2004.
- \[17\]
F. Daneshgaran, M. Laddomada, and M. Mondin, “High rate recursive convolutional codes for concatenated channel codes,” IEEE Trans. on Communications, vol.52, no.11, pp.1846-1850, Nov. 2004.
- \[18\]
M.K. Simon and M. Alouini, Digital Communication over Fading Channels: A Unified Approach to Performance Analysis, Wiley-Interscience, 1st edition, 2000. |
0710.5235 | cs.NI | cyber | # Unsaturated Throughput Analysis of IEEE 802.11 in Presence of Non Ideal Transmission Channel and Capture Effects
F. Daneshgaran, M. Laddomada, F. Mesiti, and M. Mondin
This work has been partially supported by Euroconcepts, S.r.l. (http://www.euroconcepts.it) and ISMB through the European IST project, OBAN (http://oban.prz.tu-berlin.de/index.html), IST 6FP Contract No. 001889.To appear on IEEE Transactions on Wireless Communications, 2008F. Daneshgaran is with Euroconcepts, S.r.l., Italy.M. Laddomada (laddomada@polito.it), F. Mesiti and M. Mondin are with DELEN, Politecnico di Torino, Italy.
###### Abstract
In this paper, we provide a throughput analysis of the IEEE 802.11 protocol at the data link layer in non-saturated traffic conditions taking into account the impact of both transmission channel and capture effects in Rayleigh fading environment. The impact of both non-ideal channel and capture become important in terms of the actual observed throughput in typical network conditions whereby traffic is mainly unsaturated, especially in an environment of high interference.
We extend the multi-dimensional Markovian state transition model characterizing the behavior at the MAC layer by including transmission states that account for packet transmission failures due to errors caused by propagation through the channel, along with a state characterizing the system when there are no packets to be transmitted in the buffer of a station. Finally, we derive a linear model of the throughput along with its interval of validity.
Simulation results closely match the theoretical derivations confirming the effectiveness of the proposed model.
###### Index Terms:
Capture, DCF, Distributed Coordination Function, fading, IEEE 802.11, MAC, Rayleigh, rate adaptation, saturation, throughput, unsaturated, non-saturated.
## I Introduction
Wireless Local Area Networks (WLANs) using the IEEE802.11 series of standards have experienced an exponential growth in the recent past \[1-23\]. The Medium Access Control (MAC) layer of many wireless protocols resemble that of IEEE802.11. Hence, while we focus on this protocol, it is evident that the results easily extend to other protocols with similar MAC layer operation.
The IEEE802.11 MAC presents two options \[[1](#bib.bib1)\], namely the Distributed Coordination Function (DCF) and the Point Coordination Function (PCF). PCF is generally a complex access method that can be implemented in an infrastructure network. DCF is similar to Carrier Sense Multiple Access with Collision Avoidance (CSMA/CA) and is the focus of this paper.
With this background, let us provide a quick survey of the recent literature related to the problem addressed here. This survey is by no means exhaustive and is meant to simply provide a sampling of the literature in this fertile area. We invite the interested readers to refer to the references we provide in the following and references therein.
The most relevant works to what is presented here are \[[13](#bib.bib13), [14](#bib.bib14), [15](#bib.bib15)\]. In \[[13](#bib.bib13)\] the author provided an analysis of the saturation throughput of the basic 802.11 protocol assuming a two dimensional Markov model at the MAC layer, while in \[[14](#bib.bib14)\] the authors extended the underlying model in order to consider unsaturated traffic conditions by introducing a new idle state, not present in the original Bianchi’s model, accounting for the case in which the station buffer is empty, after a successful completion of a packet transmission. In the modified model, however, a packet is discarded after $`m`$ backoff stages, while in Bianchi’s model, the station keeps iterating in the $`m`$-th backoff stage until the packet gets successfully transmitted. In \[[15](#bib.bib15)\], authors propose a novel Markov model for the DCF of IEEE 802.11, using a IEEE 802.11a PHY, in a scenario with various stations contending for the channel and transmitting with different transmission rates. An admission control mechanism is also proposed for maximizing the throughput while guaranteeing fairness to the involved transmitting stations.
In \[[16](#bib.bib16)\], the authors extend the work of Bianchi to multiple queues with different contention characteristics in the 802.11e variant of the standard with provisions of QoS. In \[[17](#bib.bib17)\], the authors present an analytical model, in which most new features of the Enhanced Distributed Channel Access (EDCA) in 802.11e such as virtual collision, different arbitration interframe spaces (AIFS), and different contention windows are taken into account. Based on the model, the throughput performance of differentiated service traffic is analyzed and a recursive method capable of calculating the mean access delay is presented. Both articles referenced assume the transmission channel to be ideal.
In \[[18](#bib.bib18)\], the authors look at the impact of channel induced errors and the received SNR on the achievable throughput in a system with rate adaptation whereby the transmission rate of the terminal is adapted based on either direct or indirect measurements of the link quality. In \[[19](#bib.bib19)\], the authors deal with the extension of Bianchi’s Markov model in order to account for channel errors. In \[[20](#bib.bib20)\], the authors investigate the saturation throughput in both congested and error-prone Gaussian channel, by proposing a simple and accurate analytical model of the behaviour of the DCF at the MAC layer. The PHY-layer is based on the parameters of the IEEE 802.11a protocol.
Paper \[[21](#bib.bib21)\] proposes an extension of the Bianchi’s model considering a new state for each backoff stage accounting for the absence of new packets to be transmitted, i.e., in unloaded traffic conditions.
In real networks, traffic is mostly unsaturated, so it is important to derive a model accounting for practical network operations. In this paper, we extend the previous works on the subject by looking at all the three issues outlined before together, namely real channel conditions, unsaturated traffic, and capture effects. Our assumptions are essentially similar to those of Bianchi \[[13](#bib.bib13)\] with the exception that we do assume the presence of both channel induced errors and capture effects due to the transmission over a Rayleigh fading channel. As a reference standard, we use network parameters belonging to the IEEE802.11b protocol, even though the proposed mathematical models hold for any flavor of the IEEE802.11 family or other wireless protocols with similar MAC layer functionality. We also demonstrate that the behavior of the throughput versus the packet rate, $`\lambda`$, is a linear function of $`\lambda`$ up to a critical packet rate, $`\lambda_{c}`$, above which throughput enters saturated conditions. Furthermore, in the linear region, the slope of the throughput depends only on the average packet length and on the number of contending stations.
Paper outline is as follows. After a brief review of the functionalities of the contention window procedure at MAC layer, section II extends the Markov model initially proposed by Bianchi, presenting modifications that account for transmission errors and capture effects over Rayleigh fading channels employing the 2-way handshaking technique in unsaturated traffic conditions. Section III provides an expression for the unsaturated throughput of the link. In section IV we present simulation results where typical MAC layer parameters for IEEE802.11b are used to obtain throughput values as a function of various system level parameters, capture probability, and SNR under typical traffic conditions. Section V derives a linear model of the throughput along with its interval of validity. Finally, Section VI is devoted to conclusions.
## II Development of the Markov Model
In this section, we present the basic rationales of the proposed bi-dimensional Markov model useful for evaluating the throughput of the DCF under unsaturated traffic conditions. We make the following assumptions: the number of terminals is finite, channel is prone to errors, capture in Rayleigh fading transmission scenario can occur, and packet transmission is based on the 2-way handshaking access mechanism. For conciseness, we will limit our presentation to the ideas needed for developing the proposed model. The interested readers can refer to \[[1](#bib.bib1), [13](#bib.bib13)\] for many details on the operating functionalities of the DCF.
###
II-A
Markovian Model Characterizing the MAC Layer under unsaturated traffic conditions, Real Transmission Channel and Capture Effects
In \[[13](#bib.bib13)\], an analytical model is presented for the computation of the throughput of a WLAN using the IEEE 802.11 DCF under ideal channel conditions. By virtue of the strategy employed for reducing the collision probability of the packets transmitted from the stations attempting to access the channel simultaneously, a random process $`b\hspace{0pt}{(t)}`$ is used to represent the backoff counter of a given station. Backoff counter is decremented at the start of every idle backoff slot and when it reaches zero, the station transmits and a new value for $`b\hspace{0pt}{(t)}`$ is set.
The value of $`b\hspace{0pt}{(t)}`$ after each transmission depends on the size of the contention window from which it is drawn. Therefore it depends on the stations’ transmission history, rendering it a non-Markovian process. To overcome this problem and get to the definition of a Markovian process, a second process $`s\hspace{0pt}{(t)}`$ is defined representing the size of the contention window from which $`b\hspace{0pt}{(t)}`$ is drawn, $`({{W_{i} = {2^{i}\hspace{0pt}W}},{i = {s\hspace{0pt}{(t)}}}})`$.
We recall that a backoff time counter is initialized depending on the number of failed transmissions for the transmitted packet. It is chosen in the range $`\lbrack 0,{W_{i} - 1}\rbrack`$ following a uniform distribution, where $`W_{i}`$ is the contention window at the backoff stage $`i`$. At the first transmission attempt (i.e., for $`i = 0`$), the contention window size is set equal to a minimum value $`W_{0} = W`$, and the process $`s\hspace{0pt}{(t)}`$ takes on the value $`{s\hspace{0pt}{(t)}} = i = 0`$.
The backoff stage $`i`$ is incremented in unitary steps after each unsuccessful transmission up to the maximum value $`m`$, while the contention window is doubled at each stage up to the maximum value $`{C\hspace{0pt}W_{m\hspace{0pt}a\hspace{0pt}x}} = {2^{m}\hspace{0pt}W}`$.
The backoff time counter is decremented as long as the channel is sensed idle and stopped when a transmission is detected. The station transmits when the backoff time counter reaches zero.
A two-dimensional Markov process $`({s\hspace{0pt}{(t)}},{b\hspace{0pt}{(t)}})`$ can now be defined, based on two assertions:
1. 1\.
the probability $`\tau`$ that a station will attempt transmission in a generic time slot is constant across all time slots;
2. 2\.
the probability $`P_{c\hspace{0pt}o\hspace{0pt}l}`$ that any transmission experiences a collision is constant and independent of the number of collisions already suffered.
Bianchi’s model relies on the following fundamental assumptions: 1) the mobile stations always have something to transmit (i.e., the saturation condition), 2) there are no hidden terminals and there is no capture effect (i.e., a terminal which perceives a higher signal-to-noise ratio (SNR) relative to other terminals capturing the channel \[[22](#bib.bib22), [23](#bib.bib23), [24](#bib.bib24)\] and limiting access to other terminals, similar to the near-far problem in cellular networks), 3) at each transmission attempt and regardless of the number of retransmissions suffered, each packet collides with constant and independent probability, and 4) the transmission channel is ideal and packet errors are only due to collisions. Clearly, the first, the second and fourth assumptions are not valid in any real setting, specially when there is mobility and when the transmission channel suffers from fading effects.
The main aim of this section is to propose an effective modification of the bi-dimensional Markov process proposed in \[[13](#bib.bib13)\] in order to account for unsaturated traffic conditions, channel error propagation and capture effects over a Rayleigh fading channel under the hypothesis of employing a 2-way handshaking access mechanism.
It is useful to briefly recall the 2-way handshaking mechanism. A station that wants to transmit a packet, waits until the channel is sensed idle for a Distributed InterFrame Space (DIFS), follows the backoff rules and then transmits a packet. After the successful reception of a data frame, the receiver sends an ACK frame to the transmitter. Only upon a correct ACK frame reception, the transmitter assumes successful delivery of the corresponding data frame. If an ACK frame is received in error or if no ACK frame is received, due possibly to an erroneous reception of the preceding data frame, the transmitter will contend again for the medium.
On the basis of this assumption, collisions can occur with probability $`P_{c\hspace{0pt}o\hspace{0pt}l}`$ on the transmitted packets, while transmission errors due to the channel, can occur with probability $`P_{e}`$. We assume that collisions and transmission error events are statistically independent. In this scenario, a packet is successfully transmitted if there is no collision (this event has probability $`1 - P_{c\hspace{0pt}o\hspace{0pt}l}`$) and the packet encounters no channel errors during transmission (this event has probability $`1 - P_{e}`$). The probability of successful transmission is therefore equal to $`{({1 - P_{e}})}\hspace{0pt}{({1 - P_{c\hspace{0pt}o\hspace{0pt}l}})}`$, from which we can set an equivalent probability of failed transmission as $`P_{e\hspace{0pt}q} = {{P_{e} + P_{c\hspace{0pt}o\hspace{0pt}l}} - {P_{e}\hspace{0pt}P_{c\hspace{0pt}o\hspace{0pt}l}}}`$.
Furthermore, in mobile radio environment, it may happen that the channel is captured by a station whose power level is stronger than other stations transmitting at the same time. This may be due to relative distances and/or channel conditions for each user and may happen whether or not the terminals exercise power control. Capture effect often reduces the collision probability on the channel since the stations whose power level at the receiver are very low due to path attenuation, shadowing and fading, are considered as interferers at the access point (AP) raising the noise floor.
[Refer to caption](/html/0710.5235/assets/x1.png)
Figure 1: Markov chain for the contention model in unsaturated traffic conditions, based on the 2-way handshaking technique, considering the effects of capture and channel induced errors.
To simplify the analysis, we make the assumption that the impact of the channel induced errors on the packet headers are negligible because of their short length with respect to the data payload size. This is justified on the basis of the assumption that the bit errors affecting the transmitted data are independent of each other. Hence, the packet or frame error rate, identified respectively with the acronyms PER or FER, is a function of the packet length, with shorter packets having exponentially smaller probability of error compared to longer packets.
Practical networks operate in unsaturated traffic conditions. In this case, Bianchi’s model \[[13](#bib.bib13)\] assuming the presence of a packet to be transmitted in each and every station’s buffer, is not valid anymore. However, the simplicity of such a model can be retained also in unsaturated conditions by introducing a new state, labelled $`I`$, accounting for the following two situations:
- •
immediately after a successful transmission, the buffer of the transmitting station is empty;
- •
the station is in an idle state with an empty buffer until a new packet arrives at the buffer for transmission.
With these considerations in mind, let us discuss the Markov model shown in Fig. [1](#S2.F1 "Figure 1 ‣ II-A Markovian Model Characterizing the MAC Layer under unsaturated traffic conditions, Real Transmission Channel and Capture Effects ‣ II Development of the Markov Model ‣ Unsaturated Throughput Analysis of IEEE 802.11 in Presence of Non Ideal Transmission Channel and Capture Effects"), modelling unsaturated traffic condition. Notice that in our Markov chain we do not model the 802.11 post-backoff feature. However, post-backoff has negligible effects on the theoretical results developed in the paper as exemplified by the simulation results obtained in the following. Similar to the model in \[[13](#bib.bib13)\], $`({m + 1})`$ different backoff stages are considered (this includes the zero-th stage). The maximum contention windows (CW) size is $`{C\hspace{0pt}W\hspace{0pt}m\hspace{0pt}a\hspace{0pt}x} = {2^{m}\hspace{0pt}W}`$, and the notation $`W_{i} = {2^{i}\hspace{0pt}W}`$ is used to define the $`i^{t\hspace{0pt}h}`$ contention window size. A packet transmission is attempted only in the $`(i,0)`$ states, $`{\forall i} = {0,\ldots,m}`$. If collision occurs, or transmission is unsuccessful due to channel errors, the backoff stage is incremented, so that the new state can be $`({i + 1},k)`$ with probability $`P_{e\hspace{0pt}q}/W_{i + 1}`$, since a uniform distribution between the states in the same backoff stage is assumed. We consider capture as a subset of the event of a collision (i.e., capture implicitly implies a collision). In other word, the capture event can happen in the presence of collision by allowing the transmitting station with the highest received power level at the access point to capture the channel. In the event of capture, no collision is detected and the Markov model transits into one of the transmitting states $`(i,0)`$ depending on the current contention stage. If no collision occurs (or is detected), a data frame can be transmitted and the transmitting station enters state $`(i,0)`$ based on its backoff stage. From state $`(i,0)`$ the transmitting station re-enters the initial backoff stage $`i = 0`$ if the transmission is successful and at least one packet is present in the buffer, otherwise the station transit in the state labelled $`I`$ waiting for a new packet arrival. Otherwise, if errors occur during transmission, the ACK packet is not sent, an ACK-timeout occurs, and the backoff stage is changed to $`({i + 1},k)`$ with probability $`P_{e\hspace{0pt}q}/W_{i + 1}`$.
The Markov Process of Fig. [1](#S2.F1 "Figure 1 ‣ II-A Markovian Model Characterizing the MAC Layer under unsaturated traffic conditions, Real Transmission Channel and Capture Effects ‣ II Development of the Markov Model ‣ Unsaturated Throughput Analysis of IEEE 802.11 in Presence of Non Ideal Transmission Channel and Capture Effects") is governed by the following transition probabilities
<sup>1</sup>
<sup>1</sup>1$`P_{i,{k|{j,n}}}`$ is short for $`P\hspace{0pt}\left. \{{{{s\hspace{0pt}{({t + 1})}} = i},{{b\hspace{0pt}{({t + 1})}} = k}} \middle| {{{s\hspace{0pt}{(t)}} = j},{{b\hspace{0pt}{(t)}} = n}}\} \right.`$.
:
<table>
<tbody>
<tr>
<td></td>
<td>Pi,k|i,k+1=1,k∈[0,Wi−2],i∈[0,m]P0,k|i,0=q(1−Peq)/W0,k∈[0,W0−1],i∈[0,m]Pi,k|i−1,0=Peq/Wi,k∈[0,Wi−1],i∈[1,m]Pm,k|m,0=Peq/Wm,k∈[0,Wm−1]PI|i,0=(1−q)(1−Peq),i∈[0,m]P0,k|I=q/W0,k∈[0,W0−1]PI|I=1−qsubscript𝑃𝑖conditional𝑘𝑖𝑘1absent1formulae-sequence𝑘0subscript𝑊𝑖2𝑖0𝑚subscript𝑃0conditional𝑘𝑖0absent𝑞1subscript𝑃𝑒𝑞subscript𝑊0formulae-sequence𝑘0subscript𝑊01𝑖0𝑚subscript𝑃𝑖conditional𝑘𝑖10absentsubscript𝑃𝑒𝑞subscript𝑊𝑖formulae-sequence𝑘0subscript𝑊𝑖1𝑖1𝑚subscript𝑃𝑚conditional𝑘𝑚0absentsubscript𝑃𝑒𝑞subscript𝑊𝑚𝑘0subscript𝑊𝑚1subscript𝑃conditional𝐼𝑖0absent1𝑞1subscript𝑃𝑒𝑞𝑖0𝑚subscript𝑃0conditional𝑘𝐼absent𝑞subscript𝑊0𝑘0subscript𝑊01subscript𝑃conditional𝐼𝐼absent1𝑞missing-subexpression\small\begin{array}[]{lll}P_{i,k|i,k+1}&=1,&~{}k\in[0,W_{i}-2],~{}i\in[0,m]\\ P_{0,k|i,0}&=q(1-P_{eq})/W_{0},&~{}k\in[0,W_{0}-1],~{}i\in[0,m]\\ P_{i,k|i-1,0}&=P_{eq}/W_{i},&~{}k\in[0,W_{i}-1],~{}i\in[1,m]\\ P_{m,k|m,0}&=P_{eq}/W_{m},&~{}k\in[0,W_{m}-1]\\ P_{I|i,0}&=(1-q)(1-P_{eq}),&~{}i\in[0,m]\\ P_{0,k|I}&=q/W_{0},&~{}k\in[0,W_{0}-1]\\ P_{I|I}&=1-q&\end{array}</td>
<td></td>
<td>(1)</td>
</tr>
</tbody>
</table>
The first equation in ([1](#S2.E1 "In II-A Markovian Model Characterizing the MAC Layer under unsaturated traffic conditions, Real Transmission Channel and Capture Effects ‣ II Development of the Markov Model ‣ Unsaturated Throughput Analysis of IEEE 802.11 in Presence of Non Ideal Transmission Channel and Capture Effects")) states that, at the beginning of each slot time, the backoff time is decremented. The second equation accounts for the fact that after a successful transmission, a new packet transmission starts with backoff stage 0 with probability $`q`$, in case there is a new packet in the buffer to be transmitted. Third and fourth equations deal with unsuccessful transmissions and the need to reschedule a new contention stage. The fifth equation deals with the practical situation in which after a successful transmission, the buffer of the station is empty, and as a consequence, the station transits in the idle state $`I`$ waiting for a new packet arrival. The sixth equation models the situation in which a new packet arrives in the station buffer, and a new backoff procedure is scheduled. Finally, the seventh equation models the situation in which there are no packets to be transmitted and the station is in the idle state.
## III Markovian Process Analysis and Throughput Computation
Next line of pursuit consists in finding a solution of the stationary distribution:
<table>
<tbody>
<tr>
<td></td>
<td>bi,k=limt→∞P[s(t)=i,b(t)=k],∀k∈[0,Wi−1],∀i∈[0,m]formulae-sequencesubscript𝑏𝑖𝑘subscript→𝑡𝑃delimited-[]formulae-sequence𝑠𝑡𝑖𝑏𝑡𝑘formulae-sequencefor-all𝑘0subscript𝑊𝑖1for-all𝑖0𝑚b_{i,k}=\lim_{t\rightarrow\infty}P[s(t)=i,b(t)=k],~{}\forall k\in[0,W_{i}-1],~{}\forall i\in[0,m]</td>
<td></td>
</tr>
</tbody>
</table>
that is, the probability of a station occupying a given state at any discrete time slot.
First, we note the following relations:
<table>
<tbody>
<tr>
<td></td>
<td>bi,0=Peq⋅bi−1,0=Peqi⋅b0,0,∀i∈[1,m−1]bm,0=Peqm1−Peq⋅b0,0,i=msubscript𝑏𝑖0⋅subscript𝑃𝑒𝑞subscript𝑏𝑖10⋅superscriptsubscript𝑃𝑒𝑞𝑖subscript𝑏00for-all𝑖1𝑚1subscript𝑏𝑚0⋅superscriptsubscript𝑃𝑒𝑞𝑚1subscript𝑃𝑒𝑞subscript𝑏00𝑖𝑚\begin{array}[]{rcll}b_{i,0}&=&P_{eq}\cdot b_{i-1,0}=P_{eq}^{i}\cdot b_{0,0},&\forall i\in[1,m-1]\\ b_{m,0}&=&\frac{P_{eq}^{m}}{1-P_{eq}}\cdot b_{0,0},&i=m\\ \end{array}</td>
<td></td>
<td>(2)</td>
</tr>
</tbody>
</table>
whereby, $`P_{e\hspace{0pt}q}`$
<sup>2</sup>
<sup>2</sup>2For simplicity, we assume that at each transmission attempt any station will encounter a constant and independent probability of failed transmission, $`P_{e\hspace{0pt}q}`$, independently from the number of retransmissions already suffered from each station.
is the equivalent probability of failed transmission, that takes into account the need for a new contention due to either packet collision ($`P_{c\hspace{0pt}o\hspace{0pt}l}`$) or channel errors ($`P_{e}`$), i.e.,
<table>
<tbody>
<tr>
<td></td>
<td>Peq=Pcol+Pe−Pe⋅Pcolsubscript𝑃𝑒𝑞subscript𝑃𝑐𝑜𝑙subscript𝑃𝑒⋅subscript𝑃𝑒subscript𝑃𝑐𝑜𝑙P_{eq}=P_{col}+P_{e}-P_{e}\cdot P_{col}</td>
<td></td>
<td>(3)</td>
</tr>
</tbody>
</table>
State $`b_{I}`$ in Fig. [1](#S2.F1 "Figure 1 ‣ II-A Markovian Model Characterizing the MAC Layer under unsaturated traffic conditions, Real Transmission Channel and Capture Effects ‣ II Development of the Markov Model ‣ Unsaturated Throughput Analysis of IEEE 802.11 in Presence of Non Ideal Transmission Channel and Capture Effects") considers both the situation in which after a successful transmission there are no packets to be transmitted, and the situation in which the packet queue is empty and the station is waiting for new packet arrival. The stationary probability to be in state $`b_{I}`$ can be evaluated as follows:
<table>
<tbody>
<tr>
<td></td>
<td>bI=(1−q)(1−Peq)⋅∑i=0mbi,0+(1−q)⋅bI=(1−q)(1−Peq)q⋅∑i=0mbi,0subscript𝑏𝐼⋅1𝑞1subscript𝑃𝑒𝑞superscriptsubscript𝑖0𝑚subscript𝑏𝑖0⋅1𝑞subscript𝑏𝐼missing-subexpression⋅1𝑞1subscript𝑃𝑒𝑞𝑞superscriptsubscript𝑖0𝑚subscript𝑏𝑖0\begin{array}[]{rcl}b_{I}&=&(1-q)(1-P_{eq})\cdot\sum_{i=0}^{m}b_{i,0}+(1-q)\cdot b_{I}\\ &=&\frac{(1-q)(1-P_{eq})}{q}\cdot\sum_{i=0}^{m}b_{i,0}\\ \end{array}</td>
<td></td>
<td>(4)</td>
</tr>
</tbody>
</table>
The expression above reflects the fact that state $`b_{I}`$ can be reached after a successful packet transmission from any state $`{b_{i,0},{\forall i}} \in {\lbrack 0,m\rbrack}`$ with probability $`{({1 - q})}\hspace{0pt}{({1 - P_{e\hspace{0pt}q}})}`$, or because the station is waiting in idle state with probability $`({1 - q})`$, whereby $`q`$ is the probability of having at least one packet to be transmitted in the buffer. The statistical model of $`q`$ will be discussed in the next section.
The other stationary probabilities for any $`k \in {\lbrack 1,{W_{i} - 1}\rbrack}`$ follow by resorting to the state transition diagram shown in Fig. [1](#S2.F1 "Figure 1 ‣ II-A Markovian Model Characterizing the MAC Layer under unsaturated traffic conditions, Real Transmission Channel and Capture Effects ‣ II Development of the Markov Model ‣ Unsaturated Throughput Analysis of IEEE 802.11 in Presence of Non Ideal Transmission Channel and Capture Effects"):
<table>
<tbody>
<tr>
<td></td>
<td>bi,k=Wi−kWi{q(1−Peq)⋅∑i=0mbi,0+q⋅bI,i=0Peq⋅bi−1,0,i∈[1,m−1]Peq(bm−1,0+bm,0),i=msubscript𝑏𝑖𝑘subscript𝑊𝑖𝑘subscript𝑊𝑖cases⋅𝑞1subscript𝑃𝑒𝑞superscriptsubscript𝑖0𝑚subscript𝑏𝑖0⋅𝑞subscript𝑏𝐼𝑖0⋅subscript𝑃𝑒𝑞subscript𝑏𝑖10𝑖1𝑚1subscript𝑃𝑒𝑞subscript𝑏𝑚10subscript𝑏𝑚0𝑖𝑚b_{i,k}=\frac{W_{i}-k}{W_{i}}\left\{\begin{array}[]{ll}q(1-P_{eq})\cdot\sum_{i=0}^{m}b_{i,0}+q\cdot b_{I},&i=0\\ P_{eq}\cdot b_{i-1,0},&i\in[1,m-1]\\ P_{eq}(b_{m-1,0}+b_{m,0}),&i=m\\ \end{array}\right.</td>
<td></td>
<td>(5)</td>
</tr>
</tbody>
</table>
Upon substituting ([4](#S3.E4 "In III Markovian Process Analysis and Throughput Computation ‣ Unsaturated Throughput Analysis of IEEE 802.11 in Presence of Non Ideal Transmission Channel and Capture Effects")) in ([5](#S3.E5 "In III Markovian Process Analysis and Throughput Computation ‣ Unsaturated Throughput Analysis of IEEE 802.11 in Presence of Non Ideal Transmission Channel and Capture Effects")), $`b_{0,k}`$ can be rewritten as follows:
<table>
<tbody>
<tr>
<td></td>
<td>q(1−Peq)⋅∑i=0mbi,0+q⋅bI==q(1−Peq)⋅∑i=0mbi,0+q⋅(1−q)(1−Peq)q⋅∑i=0mbi,0=(1−Peq)[q+(1−q)]⋅∑i=0mbi,0=(1−Peq)⋅∑i=0mbi,0⋅𝑞1subscript𝑃𝑒𝑞superscriptsubscript𝑖0𝑚subscript𝑏𝑖0⋅𝑞subscript𝑏𝐼absentmissing-subexpressionabsent⋅𝑞1subscript𝑃𝑒𝑞superscriptsubscript𝑖0𝑚subscript𝑏𝑖0⋅𝑞1𝑞1subscript𝑃𝑒𝑞𝑞superscriptsubscript𝑖0𝑚subscript𝑏𝑖0missing-subexpressionabsent⋅1subscript𝑃𝑒𝑞delimited-[]𝑞1𝑞superscriptsubscript𝑖0𝑚subscript𝑏𝑖0missing-subexpressionabsent⋅1subscript𝑃𝑒𝑞superscriptsubscript𝑖0𝑚subscript𝑏𝑖0missing-subexpression\begin{array}[]{ll}q(1-P_{eq})\cdot\sum_{i=0}^{m}b_{i,0}+q\cdot b_{I}=&\\ =q(1-P_{eq})\cdot\sum_{i=0}^{m}b_{i,0}+q\cdot\frac{(1-q)(1-P_{eq})}{q}\cdot\sum_{i=0}^{m}b_{i,0}&\\ =(1-P_{eq})[q+(1-q)]\cdot\sum_{i=0}^{m}b_{i,0}&\\ =(1-P_{eq})\cdot\sum_{i=0}^{m}b_{i,0}&\end{array}</td>
<td></td>
<td>(6)</td>
</tr>
</tbody>
</table>
Employing the normalization condition, after some mathematical manipulations, and remembering the relation $`{\sum_{i = 0}^{m}b_{i,0}} = \frac{b_{0,0}}{1 - P_{e\hspace{0pt}q}}`$, it is possible to obtain:
<table>
<tbody>
<tr>
<td></td>
<td>1=∑i=0m∑k=0Wi−1bi,k+bI=b0,02[W(∑i=0m−1(2Peq)i+(2Peq)m1−Peq)+11−Peq]++(1−q)(1−Peq)q⋅∑i=0mbi,0=b0,02[W(∑i=0m−1(2Peq)i+(2Peq)m1−Peq)+11−Peq]++(1−q)(1−Peq)q⋅b0,01−Peq=b0,02[W(∑i=0m−1(2Peq)i+(2Peq)m1−Peq)+11−Peq+2(1−q)q]1superscriptsubscript𝑖0𝑚superscriptsubscript𝑘0subscript𝑊𝑖1subscript𝑏𝑖𝑘subscript𝑏𝐼missing-subexpressionlimit-fromsubscript𝑏002delimited-[]𝑊superscriptsubscript𝑖0𝑚1superscript2subscript𝑃𝑒𝑞𝑖superscript2subscript𝑃𝑒𝑞𝑚1subscript𝑃𝑒𝑞11subscript𝑃𝑒𝑞missing-subexpressionmissing-subexpression⋅1𝑞1subscript𝑃𝑒𝑞𝑞superscriptsubscript𝑖0𝑚subscript𝑏𝑖0missing-subexpressionlimit-fromsubscript𝑏002delimited-[]𝑊superscriptsubscript𝑖0𝑚1superscript2subscript𝑃𝑒𝑞𝑖superscript2subscript𝑃𝑒𝑞𝑚1subscript𝑃𝑒𝑞11subscript𝑃𝑒𝑞missing-subexpressionmissing-subexpression⋅1𝑞1subscript𝑃𝑒𝑞𝑞subscript𝑏001subscript𝑃𝑒𝑞missing-subexpressionsubscript𝑏002delimited-[]𝑊superscriptsubscript𝑖0𝑚1superscript2subscript𝑃𝑒𝑞𝑖superscript2subscript𝑃𝑒𝑞𝑚1subscript𝑃𝑒𝑞11subscript𝑃𝑒𝑞21𝑞𝑞\small\begin{array}[]{lcl}1&=&\sum_{i=0}^{m}\sum_{k=0}^{W_{i}-1}b_{i,k}+b_{I}\\ &=&\frac{b_{0,0}}{2}\left[W\left(\sum_{i=0}^{m-1}(2P_{eq})^{i}+\frac{(2P_{eq})^{m}}{1-P_{eq}}\right)+\frac{1}{1-P_{eq}}\right]+\\ &&+\frac{(1-q)(1-P_{eq})}{q}\cdot\sum_{i=0}^{m}b_{i,0}\\ &=&\frac{b_{0,0}}{2}\left[W\left(\sum_{i=0}^{m-1}(2P_{eq})^{i}+\frac{(2P_{eq})^{m}}{1-P_{eq}}\right)+\frac{1}{1-P_{eq}}\right]+\\ &&+\frac{(1-q)(1-P_{eq})}{q}\cdot\frac{b_{0,0}}{1-P_{eq}}\\ &=&\frac{b_{0,0}}{2}\left[W\left(\sum_{i=0}^{m-1}(2P_{eq})^{i}+\frac{(2P_{eq})^{m}}{1-P_{eq}}\right)+\frac{1}{1-P_{eq}}+\frac{2(1-q)}{q}\right]\\ \end{array}</td>
<td></td>
<td>(7)</td>
</tr>
</tbody>
</table>
Normalization condition yields the following equation for computation of $`b_{0,0}`$:
<table>
<tbody>
<tr>
<td></td>
<td>b0,0=2W[(1−Peq)∑i=0m−1(2Peq)i+(2Peq)m]+1+21−qq(1−Peq)1−Peq=2(1−Peq)(W−PeqW)1−(2Peq)m(1−2Peq)+W(2Peq)m+1+2(1−q)(1−Peq)q=2(1−Peq)(1−2Peq)qq[(W+1)(1−2Peq)+WPeq(1−(2Peq)m)]+2(1−q)(1−Peq)(1−2Peq)subscript𝑏002𝑊delimited-[]1subscript𝑃𝑒𝑞superscriptsubscript𝑖0𝑚1superscript2subscript𝑃𝑒𝑞𝑖superscript2subscript𝑃𝑒𝑞𝑚121𝑞𝑞1subscript𝑃𝑒𝑞1subscript𝑃𝑒𝑞missing-subexpressionabsent21subscript𝑃𝑒𝑞𝑊subscript𝑃𝑒𝑞𝑊1superscript2subscript𝑃𝑒𝑞𝑚12subscript𝑃𝑒𝑞𝑊superscript2subscript𝑃𝑒𝑞𝑚121𝑞1subscript𝑃𝑒𝑞𝑞missing-subexpressionabsent21subscript𝑃𝑒𝑞12subscript𝑃𝑒𝑞𝑞𝑞delimited-[]𝑊112subscript𝑃𝑒𝑞𝑊subscript𝑃𝑒𝑞1superscript2subscript𝑃𝑒𝑞𝑚21𝑞1subscript𝑃𝑒𝑞12subscript𝑃𝑒𝑞missing-subexpression\begin{array}[]{ll}b_{0,0}=\frac{2}{\frac{W[(1-P_{eq})\sum_{i=0}^{m-1}(2P_{eq})^{i}+(2P_{eq})^{m}]+1+2\frac{1-q}{q}(1-P_{eq})}{1-P_{eq}}}&\\ =\frac{2(1-P_{eq})}{(W-P_{eq}W)\frac{1-(2P_{eq})^{m}}{(1-2P_{eq})}+W(2P_{eq})^{m}+1+2\frac{(1-q)(1-P_{eq})}{q}}&\\ =\frac{2(1-P_{eq})(1-2P_{eq})q}{q[(W+1)(1-2P_{eq})+WP_{eq}(1-(2P_{eq})^{m})]+2(1-q)(1-P_{eq})(1-2P_{eq})}&\end{array}</td>
<td></td>
<td>(8)</td>
</tr>
</tbody>
</table>
As a side note, if $`q\rightarrow 1`$, i.e., the station is approaching saturated traffic conditions and assuming packet transmission errors are only due to collisions, i.e., $`P_{e} = 0`$ and $`P_{e\hspace{0pt}q} = P_{c\hspace{0pt}o\hspace{0pt}l}`$, ([8](#S3.E8 "In III Markovian Process Analysis and Throughput Computation ‣ Unsaturated Throughput Analysis of IEEE 802.11 in Presence of Non Ideal Transmission Channel and Capture Effects")) we get:
<table>
<tbody>
<tr>
<td></td>
<td>limq→1b0,0→2(1−Pcol)(1−2Pcol)(W+1)(1−2Pcol)+WPcol(1−(2Pcol)m)→subscript→𝑞1subscript𝑏0021subscript𝑃𝑐𝑜𝑙12subscript𝑃𝑐𝑜𝑙𝑊112subscript𝑃𝑐𝑜𝑙𝑊subscript𝑃𝑐𝑜𝑙1superscript2subscript𝑃𝑐𝑜𝑙𝑚missing-subexpression\begin{array}[]{ll}\lim_{q\rightarrow 1}b_{0,0}\rightarrow\frac{2(1-P_{col})(1-2P_{col})}{(W+1)(1-2P_{col})+WP_{col}(1-(2P_{col})^{m})}\end{array}</td>
<td></td>
<td>(9)</td>
</tr>
</tbody>
</table>
which corresponds to the stationary state probability $`b_{0,0}`$ found by Bianchi \[[13](#bib.bib13)\] under saturated conditions.
Equ. ([8](#S3.E8 "In III Markovian Process Analysis and Throughput Computation ‣ Unsaturated Throughput Analysis of IEEE 802.11 in Presence of Non Ideal Transmission Channel and Capture Effects")) is then used to compute $`\tau`$, the probability that a station starts a transmission in a randomly chosen time slot. In fact, taking into account that a packet transmission occurs when the backoff counter reaches zero, we have:
<table>
<tbody>
<tr>
<td></td>
<td>τ=∑i=0mbi,0=b0,01−Peq==2(1−2Peq)qq[(W+1)(1−2Peq)+WPeq(1−(2Peq)m)]+2(1−q)(1−Peq)(1−2Peq)𝜏absentsuperscriptsubscript𝑖0𝑚subscript𝑏𝑖0subscript𝑏001subscript𝑃𝑒𝑞absent212subscript𝑃𝑒𝑞𝑞𝑞delimited-[]𝑊112subscript𝑃𝑒𝑞𝑊subscript𝑃𝑒𝑞1superscript2subscript𝑃𝑒𝑞𝑚21𝑞1subscript𝑃𝑒𝑞12subscript𝑃𝑒𝑞\begin{array}[]{ll}\tau&=\sum_{i=0}^{m}b_{i,0}=\frac{b_{0,0}}{1-P_{eq}}=\\ =&\frac{2(1-2P_{eq})q}{q[(W+1)(1-2P_{eq})+WP_{eq}(1-(2P_{eq})^{m})]+2(1-q)(1-P_{eq})(1-2P_{eq})}\\ \end{array}</td>
<td></td>
<td>(10)</td>
</tr>
</tbody>
</table>
Once again as a consistency check, note that if $`m = 0`$, i.e., no exponential backoff is employed, $`q\rightarrow 1`$, i.e., the station is approaching saturated traffic conditions, and under the condition that packet transmission errors are only due to collisions, i.e., $`P_{e} = 0`$ and $`P_{e\hspace{0pt}q} = P_{c\hspace{0pt}o\hspace{0pt}l}`$, ([10](#S3.E10 "In III Markovian Process Analysis and Throughput Computation ‣ Unsaturated Throughput Analysis of IEEE 802.11 in Presence of Non Ideal Transmission Channel and Capture Effects")) we get:
<table>
<tbody>
<tr>
<td></td>
<td>limq→1τ→2W+1→subscript→𝑞1𝜏2𝑊1\lim_{q\rightarrow 1}\tau\rightarrow\frac{2}{W+1}</td>
<td></td>
<td>(11)</td>
</tr>
</tbody>
</table>
which is the result found in \[[26](#bib.bib26)\] for the constant backoff setting, showing that the transmission probability $`\tau`$ is independent of the collision probability $`P_{c\hspace{0pt}o\hspace{0pt}l}`$.
The collision probability needed to compute $`\tau`$ can be found considering that using a 2-way hand-shaking mechanism, a packet from a transmitting station encounters a collision if in a given time slot, at least one of the remaining $`({N - 1})`$ stations transmits simultaneously another packet, and there is no capture. In our model, we assume that capture is a subset of the collision events. This is indeed justified by the fact that there is no capture without collision, and that capture occurs only because of collisions between a certain number of transmitting stations attempting to transmit simultaneously on the channel.
<table>
<tbody>
<tr>
<td></td>
<td>Pcol=1−(1−τ)N−1−Pcapsubscript𝑃𝑐𝑜𝑙1superscript1𝜏𝑁1subscript𝑃𝑐𝑎𝑝P_{col}=1-(1-\tau)^{N-1}-P_{cap}</td>
<td></td>
<td>(12)</td>
</tr>
</tbody>
</table>
As far as the capture effects are concerned, we resort to the mathematical formulation proposed in \[[23](#bib.bib23), [24](#bib.bib24)\].
We consider a scenario in which $`N`$ stations are uniformly distributed in a circular area of radius $`R`$, and transmit toward a common access point placed in the center of such an area. When signal transmission is affected by Rayleigh fading, the instantaneous power of the signal received by the receiver placed at a mutual distance $`r_{i}`$ from the transmitter is exponentially distributed as:
<table>
<tbody>
<tr>
<td></td>
<td>f(x)=1poe−xpo,x>0formulae-sequence𝑓𝑥1subscript𝑝𝑜superscript𝑒𝑥subscript𝑝𝑜𝑥0f(x)=\frac{1}{p_{o}}e^{-\frac{x}{p_{o}}},x>0</td>
<td></td>
</tr>
</tbody>
</table>
whereby $`p_{o}`$ is the local-mean power determined by the equation:
<table>
<tbody>
<tr>
<td></td>
<td>po=A⋅ri−npPtxsubscript𝑝𝑜⋅𝐴superscriptsubscript𝑟𝑖subscript𝑛𝑝subscript𝑃𝑡𝑥p_{o}=A\cdot r_{i}^{-n_{p}}P_{tx}</td>
<td></td>
<td>(13)</td>
</tr>
</tbody>
</table>
where, $`n_{p}`$ is the path-loss exponent (which is typically greater or equal to $`3.5`$ in indoor propagation conditions in the absence of the direct signal path), $`P_{t\hspace{0pt}x}`$ is the transmitted power, and $`A \cdot r_{i}^{- n_{p}}`$ is the deterministic path-loss \[[29](#bib.bib29)\]. Both $`A`$ and $`P_{t\hspace{0pt}x}`$ are identical for all transmitted frames.
Under the hypothesis of power-controlled stations in infrastructure mode and Rayleigh fading, the capture probability conditioned on $`i`$ interfering frames can be defined as follows:
<table>
<tbody>
<tr>
<td></td>
<td>Pcp(γ>zog(Sf)|i)=1[1+z0g(Sf)]isubscript𝑃𝑐𝑝𝛾conditionalsubscript𝑧𝑜𝑔subscript𝑆𝑓𝑖1superscriptdelimited-[]1subscript𝑧0𝑔subscript𝑆𝑓𝑖P_{cp}\left(\gamma>z_{o}g(S_{f})|i\right)=\frac{1}{{\left[1+z_{0}g(S_{f})\right]}^{i}}</td>
<td></td>
<td>(14)</td>
</tr>
</tbody>
</table>
whereby, $`\gamma`$, defined as
<table>
<tbody>
<tr>
<td></td>
<td>γ=Pu/∑k=1iPk𝛾subscript𝑃𝑢superscriptsubscript𝑘1𝑖subscript𝑃𝑘\gamma=P_{u}/\sum_{k=1}^{i}P_{k}</td>
<td></td>
<td>(15)</td>
</tr>
</tbody>
</table>
is the ratio of the power $`P_{u}`$ of the useful signal and the sum of the powers of the $`i`$ interfering channel contenders transmitting simultaneously $`i`$ frames, $`g\hspace{0pt}{(S_{f})}`$ is the inverse of the processing gain of the correlation receiver, and $`z_{0}`$ is the capture ratio, i.e., the value of the signal-to-interference power ratio identifying the capture threshold at the receiver. Notice that ([14](#S3.E14 "In III Markovian Process Analysis and Throughput Computation ‣ Unsaturated Throughput Analysis of IEEE 802.11 in Presence of Non Ideal Transmission Channel and Capture Effects")) signifies the fact that capture probability corresponds to the probability that the power ratio $`\gamma`$ is above the capture threshold $`z_{0}\hspace{0pt}g\hspace{0pt}{(S_{f})}`$ which considers the inverse of the processing gain $`g\hspace{0pt}{(S_{f})}`$. For Direct Sequence Spread Spectrum (DSSS) using a 11-chip spreading factor ($`S_{f} = 11`$), we have $`{g\hspace{0pt}{(S_{f})}} = \frac{2}{3\hspace{0pt}S_{f}}`$ \[[25](#bib.bib25)\].
Upon defining the probability of generating exactly $`i + 1`$ interfering frames over $`N`$ contending stations in a generic slot time:
<table>
<tbody>
<tr>
<td></td>
<td>(Ni+1)τi+1(1−τ)N−i−1binomial𝑁𝑖1superscript𝜏𝑖1superscript1𝜏𝑁𝑖1{N\choose i+1}\tau^{i+1}(1-\tau)^{N-i-1}</td>
<td></td>
</tr>
</tbody>
</table>
the frame capture probability $`P_{c\hspace{0pt}a\hspace{0pt}p}`$ can be obtained as follows:
<table>
<tbody>
<tr>
<td></td>
<td>Pcap=∑i=1N−1(Ni+1)τi+1(1−τ)N−i−1Pcp(γ>zog(Sf)|i)subscript𝑃𝑐𝑎𝑝superscriptsubscript𝑖1𝑁1binomial𝑁𝑖1superscript𝜏𝑖1superscript1𝜏𝑁𝑖1subscript𝑃𝑐𝑝𝛾conditionalsubscript𝑧𝑜𝑔subscript𝑆𝑓𝑖P_{cap}=\sum_{i=1}^{N-1}{N\choose i+1}\tau^{i+1}(1-\tau)^{N-i-1}P_{cp}\left(\gamma>z_{o}g(S_{f})|i\right)</td>
<td></td>
<td>(16)</td>
</tr>
</tbody>
</table>
Putting together Equ.s ([3](#S3.E3 "In III Markovian Process Analysis and Throughput Computation ‣ Unsaturated Throughput Analysis of IEEE 802.11 in Presence of Non Ideal Transmission Channel and Capture Effects")), ([10](#S3.E10 "In III Markovian Process Analysis and Throughput Computation ‣ Unsaturated Throughput Analysis of IEEE 802.11 in Presence of Non Ideal Transmission Channel and Capture Effects")), ([12](#S3.E12 "In III Markovian Process Analysis and Throughput Computation ‣ Unsaturated Throughput Analysis of IEEE 802.11 in Presence of Non Ideal Transmission Channel and Capture Effects")), and ([16](#S3.E16 "In III Markovian Process Analysis and Throughput Computation ‣ Unsaturated Throughput Analysis of IEEE 802.11 in Presence of Non Ideal Transmission Channel and Capture Effects")) the following nonlinear system can be defined and solved numerically, obtaining the values of $`\tau`$, $`P_{c\hspace{0pt}o\hspace{0pt}l}`$, $`P_{c\hspace{0pt}a\hspace{0pt}p}`$, and $`P_{e\hspace{0pt}q}`$:
<table>
<tbody>
<tr>
<td></td>
<td>{τ=2(1−2Peq)qq[(W+1)(1−2Peq)+WPeq(1−(2Peq)m)]+2(1−q)(1−Peq)(1−2Peq)Pcol=1−(1−τ)N−1−PcapPeq=Pcol+Pe−Pe⋅PcolPcap=∑i=1N−1(Ni+1)τi+1(1−τ)N−i−11(1+z0g(Sf))icases𝜏212subscript𝑃𝑒𝑞𝑞𝑞delimited-[]𝑊112subscript𝑃𝑒𝑞𝑊subscript𝑃𝑒𝑞1superscript2subscript𝑃𝑒𝑞𝑚21𝑞1subscript𝑃𝑒𝑞12subscript𝑃𝑒𝑞missing-subexpressionsubscript𝑃𝑐𝑜𝑙1superscript1𝜏𝑁1subscript𝑃𝑐𝑎𝑝missing-subexpressionsubscript𝑃𝑒𝑞subscript𝑃𝑐𝑜𝑙subscript𝑃𝑒⋅subscript𝑃𝑒subscript𝑃𝑐𝑜𝑙missing-subexpressionsubscript𝑃𝑐𝑎𝑝superscriptsubscript𝑖1𝑁1binomial𝑁𝑖1superscript𝜏𝑖1superscript1𝜏𝑁𝑖11superscript1subscript𝑧0𝑔subscript𝑆𝑓𝑖missing-subexpression\left\{\begin{array}[]{ll}\tau=\frac{2(1-2P_{eq})q}{q[(W+1)(1-2P_{eq})+WP_{eq}(1-(2P_{eq})^{m})]+2(1-q)(1-P_{eq})(1-2P_{eq})}\\ P_{col}=1-(1-\tau)^{N-1}-P_{cap}\\ P_{eq}=P_{col}+P_{e}-P_{e}\cdot P_{col}\\ P_{cap}=\sum_{i=1}^{N-1}{N\choose i+1}\tau^{i+1}(1-\tau)^{N-i-1}\frac{1}{{(1+z_{0}g(S_{f}))}^{i}}\end{array}\right.</td>
<td></td>
<td>(17)</td>
</tr>
</tbody>
</table>
The final step in the analysis is the computation of the normalized system throughput, defined as the fraction of time the channel is used to successfully transmit payload bits:
<table>
<tbody>
<tr>
<td></td>
<td>S=Pt⋅Ps⋅(1−Pe)E[PL](1−Pt)σ+Pt(1−Ps)Tc+PtPs(1−Pe)Ts+PtPsPeTe𝑆⋅subscript𝑃𝑡subscript𝑃𝑠1subscript𝑃𝑒𝐸delimited-[]𝑃𝐿1subscript𝑃𝑡𝜎subscript𝑃𝑡1subscript𝑃𝑠subscript𝑇𝑐subscript𝑃𝑡subscript𝑃𝑠1subscript𝑃𝑒subscript𝑇𝑠subscript𝑃𝑡subscript𝑃𝑠subscript𝑃𝑒subscript𝑇𝑒\small S=\frac{P_{t}\cdot P_{s}\cdot(1-P_{e})E[PL]}{(1-P_{t})\sigma+P_{t}(1-P_{s})T_{c}+P_{t}P_{s}(1-P_{e})T_{s}+P_{t}P_{s}P_{e}T_{e}}</td>
<td></td>
<td>(18)</td>
</tr>
</tbody>
</table>
where,
- •
$`P_{t}`$ is the probability that there is at least one transmission in the considered time slot, with $`N`$ stations contending for the channel, each transmitting with probability $`\tau`$:
<table>
<tbody>
<tr>
<td></td>
<td>Pt=1−(1−τ)Nsubscript𝑃𝑡1superscript1𝜏𝑁P_{t}=1-(1-\tau)^{N}</td>
<td></td>
<td>(19)</td>
</tr>
</tbody>
</table>
- •
$`P_{s}`$ is the conditional probability that a packet transmission occurring on the channel is successful. This event corresponds to the case in which exactly one station transmits in a given time slot, or two or more stations transmit simultaneously and capture by the desired station occurs. These conditions yields the following probability:
<table>
<tbody>
<tr>
<td></td>
<td>Ps=Nτ(1−τ)N−1+PcapPtsubscript𝑃𝑠𝑁𝜏superscript1𝜏𝑁1subscript𝑃𝑐𝑎𝑝subscript𝑃𝑡P_{s}=\frac{N\tau(1-\tau)^{N-1}+P_{cap}}{P_{t}}</td>
<td></td>
<td>(20)</td>
</tr>
</tbody>
</table>
- •
$`T_{c}`$, $`T_{e}`$ and $`T_{s}`$ are the average times a channel is sensed busy due to a collision, error affected data frame transmission time and successful data frame transmission times, respectively. Knowing the time durations for ACK frames, ACK timeout, DIFS, SIFS, $`\sigma`$, data packet length ($`P\hspace{0pt}L`$) and PHY and MAC headers duration ($`H`$), and propagation delay $`\tau_{p}`$, $`T_{c}`$, $`T_{s}`$, and $`T_{e}`$ can be computed as follows \[[17](#bib.bib17)\]:
<table>
<tbody>
<tr>
<td></td>
<td>Tc=H+PL+ACKtimeoutTe=H+PL+ACKtimeoutTs=H+PL+SIFS+τp+ACK+DIFS+τpsubscript𝑇𝑐absent𝐻𝑃𝐿𝐴𝐶subscript𝐾𝑡𝑖𝑚𝑒𝑜𝑢𝑡subscript𝑇𝑒absent𝐻𝑃𝐿𝐴𝐶subscript𝐾𝑡𝑖𝑚𝑒𝑜𝑢𝑡subscript𝑇𝑠absent𝐻𝑃𝐿𝑆𝐼𝐹𝑆subscript𝜏𝑝𝐴𝐶𝐾𝐷𝐼𝐹𝑆subscript𝜏𝑝\begin{array}[]{ll}T_{c}&=H+PL+ACK_{timeout}\\ T_{e}&=H+PL+ACK_{timeout}\\ T_{s}&=H+PL+SIFS+\tau_{p}+ACK+DIFS+\tau_{p}\\ \end{array}</td>
<td></td>
<td>(21)</td>
</tr>
</tbody>
</table>
- •
$`E\hspace{0pt}{\lbrack{P\hspace{0pt}L}\rbrack}`$ is the average packet payload length.
- •
$`\sigma`$ is the duration of an empty time slot.
The setup described above is used in Section [IV](#S4 "IV Simulation Results and Model Validations ‣ Unsaturated Throughput Analysis of IEEE 802.11 in Presence of Non Ideal Transmission Channel and Capture Effects") for DCF simulation at the MAC layer.
###
III-A
Modelling offered load and estimation of probability $`q`$
In our analysis, the offered load related to each station is characterized by parameter $`\lambda`$ representing the rate at which packets arrive at the station buffer from the upper layers, and measured in $`{p\hspace{0pt}k\hspace{0pt}t}/s`$. The time between two packet arrivals is defined as interarrival time, and its mean value is evaluated as $`\frac{1}{\lambda}`$. One of the most commonly used traffic models assumes packet arrival process is Poisson. The resulting interarrival times are exponentially distributed.
In the proposed model, we need a probability $`q`$ that indicates if there is at least one packet to be transmitted in the queue. Probability $`q`$ can be well approximated in a situation with small buffer size \[[27](#bib.bib27)\] through the following relation:
<table>
<tbody>
<tr>
<td></td>
<td>q=1−e−λE[Sts]𝑞1superscript𝑒𝜆𝐸delimited-[]subscript𝑆𝑡𝑠q=1-e^{-\lambda E[S_{ts}]}</td>
<td></td>
<td>(22)</td>
</tr>
</tbody>
</table>
where, $`E\hspace{0pt}{\lbrack S_{t\hspace{0pt}s}\rbrack}`$ is the expected time per slot, which is useful to relate the state of the Markov chain with the actual time spent in each state.
A more accurate model can be derived upon considering different values of $`q`$ for each backoff state. However, a reasonable solution consists in using a mean probability valid for the whole Markov model \[[27](#bib.bib27)\], derived from $`E\hspace{0pt}{\lbrack S_{t\hspace{0pt}s}\rbrack}`$. The value of $`E\hspace{0pt}{\lbrack S_{t\hspace{0pt}s}\rbrack}`$ can be obtained by resorting to the durations and the respective probabilities of the idle slot ($`\sigma`$), the successful transmission slot ($`T_{s}`$), the error slot due to collision ($`T_{c}`$), and the error slot due to channel ($`T_{e}`$), as follows:
<table>
<tbody>
<tr>
<td></td>
<td>E[Sts]=(1−Pt)⋅σ+Pt(1−Ps)⋅Tc++PtPsPe⋅Te+PtPs(1−Pe)⋅Ts𝐸delimited-[]subscript𝑆𝑡𝑠⋅1subscript𝑃𝑡𝜎limit-from⋅subscript𝑃𝑡1subscript𝑃𝑠subscript𝑇𝑐missing-subexpressionmissing-subexpression⋅subscript𝑃𝑡subscript𝑃𝑠subscript𝑃𝑒subscript𝑇𝑒⋅subscript𝑃𝑡subscript𝑃𝑠1subscript𝑃𝑒subscript𝑇𝑠\begin{array}[]{rcl}E[S_{ts}]&=&(1-P_{t})\cdot\sigma+P_{t}(1-P_{s})\cdot T_{c}+\\ &&+P_{t}P_{s}P_{e}\cdot T_{e}+P_{t}P_{s}(1-P_{e})\cdot T_{s}\end{array}</td>
<td></td>
<td>(23)</td>
</tr>
</tbody>
</table>
Upon recalling that packet inter-arrival times are exponentially distributed, we can use the average slot time to calculate the probability $`q`$ that in such a time interval a given station receives a packet from upper layers in its transmission queue. The probability that in a generic time $`T`$, $`k`$ events occur, is:
<table>
<tbody>
<tr>
<td></td>
<td>P{a(T)=k}=e−λT(λT)kk!𝑃𝑎𝑇𝑘superscript𝑒𝜆𝑇superscript𝜆𝑇𝑘𝑘P\{a(T)=k\}=e^{-\lambda T}\frac{(\lambda T)^{k}}{k!}</td>
<td></td>
<td>(24)</td>
</tr>
</tbody>
</table>
from which we obtain the relation ([22](#S3.E22 "In III-A Modelling offered load and estimation of probability q ‣ III Markovian Process Analysis and Throughput Computation ‣ Unsaturated Throughput Analysis of IEEE 802.11 in Presence of Non Ideal Transmission Channel and Capture Effects")) referred to earlier:
<table>
<tbody>
<tr>
<td></td>
<td>q=1−P{a(E[Sts])=0}=1−e−λE[Sts]𝑞1𝑃𝑎𝐸delimited-[]subscript𝑆𝑡𝑠01superscript𝑒𝜆𝐸delimited-[]subscript𝑆𝑡𝑠q=1-P\{a(E[S_{ts}])=0\}=1-e^{-\lambda E[S_{ts}]}</td>
<td></td>
<td>(25)</td>
</tr>
</tbody>
</table>
## IV Simulation Results and Model Validations
TABLE I: Typical network parameters
<table>
<tbody>
<tr>
<th>MAC header</th>
<td>24 bytes</td>
</tr>
<tr>
<th>PHY header</th>
<td>16 bytes</td>
</tr>
<tr>
<th>Payload size</th>
<td>1024 bytes</td>
</tr>
<tr>
<th>ACK</th>
<td>14 bytes</td>
</tr>
<tr>
<th>RTS</th>
<td>20 bytes</td>
</tr>
<tr>
<th>CTS</th>
<td>14 bytes</td>
</tr>
<tr>
<th>τpsubscript𝜏𝑝\tau_{p}</th>
<td>1 μs𝜇𝑠\mu s</td>
</tr>
<tr>
<th>Slot time</th>
<td>20 μs𝜇𝑠\mu s</td>
</tr>
<tr>
<th>SIFS</th>
<td>10 μs𝜇𝑠\mu s</td>
</tr>
<tr>
<th>DIFS</th>
<td>50 μs𝜇𝑠\mu s</td>
</tr>
<tr>
<th>EIFS</th>
<td>300 μs𝜇𝑠\mu s</td>
</tr>
<tr>
<th>ACK timeout</th>
<td>300 μs𝜇𝑠\mu s</td>
</tr>
<tr>
<th>CTS timeout</th>
<td>300 μs𝜇𝑠\mu s</td>
</tr>
</tbody>
</table>
This section focuses on simulation results for validating the theoretical models and derivations presented in the previous sections. We have developed a C++ simulator modelling both the DCF protocol details in 802.11b and the backoff procedures of a specific number of independent transmitting stations. The simulator is designed to implement the main tasks accomplished at both MAC and PHY layers of a wireless network in a more versatile and customizable manner than ns-2, where the lack of a complete physical layer makes difficult a precise configuration at this level.
The simulator considers an Infrastructure BSS (Basic Service Set) with an AP and a certain number of mobile stations which communicates only with the AP. For the sake of simplicity, inside each station there are only three fundamental working levels: traffic model generator, MAC and PHY layers. Traffic is generated following the exponential distribution for the packet interarrival times. Moreover, the MAC layer is managed by a state machine which follows the main directives specified in the standard \[[1](#bib.bib1)\], namely waiting times (DIFS, SIFS, EIFS), post-backoff, backoff, basic and RTS/CTS access mode.
Typical MAC layer parameters for IEEE802.11b are given in Table [I](#S4.T1 "TABLE I ‣ IV Simulation Results and Model Validations ‣ Unsaturated Throughput Analysis of IEEE 802.11 in Presence of Non Ideal Transmission Channel and Capture Effects") \[[1](#bib.bib1)\]. In so far as the computation of the FER is concerned, it should be noted that data transmission rate of various packet types differ. For simplicity, we assume that data packets transmitted by different stations are affected by the same FER. This way, channel errors on the transmitted packets can be accounted for as it is done within ns-2 \[[30](#bib.bib30)\]. In other words, a uniformly distributed binary random variable $`X_{e}`$ is generated in order to decide if a transmitted packet is received erroneously. The statistic of such a random variable is $`{P\hspace{0pt}{({X_{e} = 1})}} = {P_{e}\hspace{0pt}{({S\hspace{0pt}N\hspace{0pt}R})}}`$ (as specified in ([26](#S4.E26 "In IV Simulation Results and Model Validations ‣ Unsaturated Throughput Analysis of IEEE 802.11 in Presence of Non Ideal Transmission Channel and Capture Effects"))), and $`{P\hspace{0pt}{({X_{e} = 0})}} = {1 - {P_{e}\hspace{0pt}{({S\hspace{0pt}N\hspace{0pt}R})}}}`$.
As far as capture effects are concerned, the investigated scenario is as follows. In our simulator, $`N`$ stations are randomly placed in a circular area of radius $`R`$ (in the simulation results presented below we assume $`R = 10`$m), while the AP is placed at the center of the circle. When two or more station transmissions collide, the value of $`\gamma`$ as defined in ([15](#S3.E15 "In III Markovian Process Analysis and Throughput Computation ‣ Unsaturated Throughput Analysis of IEEE 802.11 in Presence of Non Ideal Transmission Channel and Capture Effects")) is evaluated for any transmitting station given their relative distance $`r_{i}`$ from the AP. Let $`\gamma_{j}`$ be the value of $`\gamma`$ for the $`j`$-th transmitting station among the $`i + 1`$ colliding stations. The power between a transmitter and a receiver is evaluated in accordance to ([13](#S3.E13 "In III Markovian Process Analysis and Throughput Computation ‣ Unsaturated Throughput Analysis of IEEE 802.11 in Presence of Non Ideal Transmission Channel and Capture Effects")) with a path-loss exponent equal to $`3.5`$. The values of $`\gamma_{j}`$ for each colliding station are compared with the threshold $`{z_{o} \cdot g}\hspace{0pt}{(S_{f})}`$. Then, the transmitting station for which $`\gamma_{j}`$ is above the threshold captures the channel.
The physical layer (PHY) of the basic 802.11b standard is based on the spread spectrum technology. Two options are specified, the Frequency Hopped Spread Spectrum (FHSS) and the Direct Sequence Spread Spectrum (DSSS). The FHSS uses Frequency Shift Keying (FSK) while the DSSS uses Differential Phase Shift Keying (DPSK) or Complementary Code Keying (CCK). The 802.11b employs DSSS at various rates including one employing CCK encoding 4 and 8 bits on one CCK symbol. The four supported data rates in 802.11b are 1, 2, 5.5 and 11 Mbps.
In particular, the following data are transmitted at lowest rate (1 Mbps) in IEEE802.11b: PLCP=16 bytes (PLCP plus header), ACK Headers=16 bytes, ACK=14 bytes, RTS=20 bytes, CTS=14 bytes (RTS and CTS are only for four way handshake).
The FER as a function of the SNR can be computed as follows:
<table>
<tbody>
<tr>
<td></td>
<td>Pe(SNR)=1−[1−Pe(PLCP,SNR)]⋅[1−Pe(DATA,SNR)]subscript𝑃𝑒𝑆𝑁𝑅1⋅delimited-[]1subscript𝑃𝑒𝑃𝐿𝐶𝑃𝑆𝑁𝑅delimited-[]1subscript𝑃𝑒𝐷𝐴𝑇𝐴𝑆𝑁𝑅\small P_{e}(SNR)=1-\left[1-P_{e}(PLCP,SNR)\right]\cdot\left[1-P_{e}(DATA,SNR)\right]</td>
<td></td>
<td>(26)</td>
</tr>
</tbody>
</table>
where,
<table>
<tbody>
<tr>
<td></td>
<td>Pe(PLCP,SNR)=1−[1−Pb(BPSK,SNR)]8×PLCP,subscript𝑃𝑒𝑃𝐿𝐶𝑃𝑆𝑁𝑅1superscriptdelimited-[]1subscript𝑃𝑏𝐵𝑃𝑆𝐾𝑆𝑁𝑅8𝑃𝐿𝐶𝑃\small P_{e}(PLCP,SNR)=1-\left[1-P_{b}(BPSK,SNR)\right]^{8\times PLCP},</td>
<td></td>
<td>(27)</td>
</tr>
</tbody>
</table>
and
<table>
<tbody>
<tr>
<td></td>
<td>Pe(DATA,SNR)=1−[1−Pb(TYPE,SNR)]8×(DATA+MAC).subscript𝑃𝑒𝐷𝐴𝑇𝐴𝑆𝑁𝑅1superscriptdelimited-[]1subscript𝑃𝑏𝑇𝑌𝑃𝐸𝑆𝑁𝑅8𝐷𝐴𝑇𝐴𝑀𝐴𝐶\small P_{e}(DATA,SNR)=1-\left[1-P_{b}(TYPE,SNR)\right]^{8\times(DATA+MAC)}.</td>
<td></td>
<td>(28)</td>
</tr>
</tbody>
</table>
$`P_{b}\hspace{0pt}{({B\hspace{0pt}P\hspace{0pt}S\hspace{0pt}K},{S\hspace{0pt}N\hspace{0pt}R})}`$ is the BER as a function of SNR for the lowest data transmit rate employing DBPSK modulation, DATA denotes the data block size in bytes, and any other constant byte size in above expression represents overhead. Note that the FER, $`P_{e}\hspace{0pt}{({S\hspace{0pt}N\hspace{0pt}R})}`$, implicitly depends on the modulation format used. Hence, for each supported rate, one curve for $`P_{e}\hspace{0pt}{({S\hspace{0pt}N\hspace{0pt}R})}`$ as a function of SNR can be generated. $`P_{b}\hspace{0pt}{({T\hspace{0pt}Y\hspace{0pt}P\hspace{0pt}E},{S\hspace{0pt}N\hspace{0pt}R})}`$ is modulation dependent whereby the parameter $`T\hspace{0pt}Y\hspace{0pt}P\hspace{0pt}E`$ can be any of the following $`{T\hspace{0pt}Y\hspace{0pt}P\hspace{0pt}E} \in {\{{D\hspace{0pt}B\hspace{0pt}P\hspace{0pt}S\hspace{0pt}K},{D\hspace{0pt}Q\hspace{0pt}P\hspace{0pt}S\hspace{0pt}K},{C\hspace{0pt}C\hspace{0pt}K\hspace{0pt}5.5},{C\hspace{0pt}C\hspace{0pt}K\hspace{0pt}11}\}}`$
<sup>3</sup>
<sup>3</sup>3The acronyms are short for Differential Binary Phase Shift Keying, Differential Quadrature Phase Shift Keying and Complementary Code Keying, respectively.
.
For DBPSK and DQPSK modulation formats, $`P_{b}\hspace{0pt}{({T\hspace{0pt}Y\hspace{0pt}P\hspace{0pt}E},{S\hspace{0pt}N\hspace{0pt}R})}`$ can be well approximated by \[[28](#bib.bib28)\]:
[Refer to caption](/html/0710.5235/assets/x2.png)
Figure 2: Theoretical and simulated throughput for the 2-way mechanism as a function of the SNR, for two different packet rates and capture thresholds, shown in the legend. Simulated points are identified by star-markers over the respective theoretical curves. Payload size is 1024 bytes, while the number of contending stations is $`N = 10`$.
<table>
<tbody>
<tr>
<td></td>
<td>2max(log2M,2)∑i=1max(M4,1)1π∫0π211+γlog2Msin2θsin2((2i−1)πM)𝑑θ2subscript2𝑀2superscriptsubscript𝑖1𝑀411𝜋superscriptsubscript0𝜋211𝛾subscript2𝑀superscript2𝜃superscript22𝑖1𝜋𝑀differential-d𝜃\small\frac{2}{\max(\log_{2}M,2)}\sum_{i=1}^{\max(\frac{M}{4},1)}\frac{1}{\pi}\int_{0}^{\frac{\pi}{2}}\frac{1}{1+\frac{\gamma\log_{2}M}{\sin^{2}\theta}\sin^{2}\left(\frac{(2i-1)\pi}{M}\right)}d\theta</td>
<td></td>
<td>(29)</td>
</tr>
</tbody>
</table>
whereby, $`M`$ is the number of bits per modulated symbols, $`\gamma`$ is the signal-to-noise ratio, and $`\theta`$ is the signal direction over the Rayleigh fading channel.
We note that the proposed DCF model is valid with any other PHY setup. Actually, all we need is simply the packet error rate probability $`P_{e}\hspace{0pt}{({D\hspace{0pt}A\hspace{0pt}T\hspace{0pt}A},{S\hspace{0pt}N\hspace{0pt}R})}`$ in ([26](#S4.E26 "In IV Simulation Results and Model Validations ‣ Unsaturated Throughput Analysis of IEEE 802.11 in Presence of Non Ideal Transmission Channel and Capture Effects")) for the specific PHY and channel transmission conditions model. All the mathematical derivations proposed in this paper are specified with respect to $`P_{e}\hspace{0pt}{({D\hspace{0pt}A\hspace{0pt}T\hspace{0pt}A},{S\hspace{0pt}N\hspace{0pt}R})}`$, and so they are valid for any kind of transmission model at the PHY layer.
In what follows, we shall present theoretical and simulation results for the lowest supported data rate. We note that by repeating the process, similar curves can be generated for all other types of modulation formats. All we need is really the BER as a function of SNR for each modulation format and the corresponding raw data rate over the channel. If the terminals use rate adaptation, then under optimal operating condition, the achievable throughput for a given SNR is the maximum over the set of modulation formats supported. We have verified a close match between theoretical and simulated performances for other transmitting data rates as well. In the results presented below we assume the following values for the contention window: $`{C\hspace{0pt}W_{m\hspace{0pt}i\hspace{0pt}n}} = 32`$, $`m = 5`$, and $`{C\hspace{0pt}W_{m\hspace{0pt}a\hspace{0pt}x}} = {{2^{m} \cdot C}\hspace{0pt}W_{m\hspace{0pt}i\hspace{0pt}n}} = 1024`$.
Fig. [2](#S4.F2 "Figure 2 ‣ IV Simulation Results and Model Validations ‣ Unsaturated Throughput Analysis of IEEE 802.11 in Presence of Non Ideal Transmission Channel and Capture Effects") shows the behavior of the throughput for the 2-way mechanism as a function of the SNR, for two different capture thresholds, $`z_{0}`$, and packet rates, $`\lambda`$, as noted in the legend, and for $`N = 10`$ transmitting stations. Simulated points are marked by star on the respective theoretical curves. Throughput increases as a function of the three parameters SNR, $`\lambda`$, and $`z_{0}`$ until saturation point.
[Refer to caption](/html/0710.5235/assets/x3.png)
Figure 3: Theoretical and simulated throughput for the 2-way mechanism as a function of the packet rate $`\lambda`$, for three different number of contending stations and two different values of SNR as noted in the legend. Capture thresholds is $`z_{0} = 6`$dB. Simulated points are identified by star-markers over the respective theoretical curves. Payload size is 1024 bytes.
Fig. [3](#S4.F3 "Figure 3 ‣ IV Simulation Results and Model Validations ‣ Unsaturated Throughput Analysis of IEEE 802.11 in Presence of Non Ideal Transmission Channel and Capture Effects") shows the behavior of the throughput as a function of $`\lambda`$, i.e., the packet rate, for three different values of the number of contending stations and for two values of SNR. The capture threshold is $`z_{0} = 6`$dB. Beside noting the throughput improvement achievable for high SNR, notice that for a specified number of contending stations, the throughput manifests a linear behavior for low values of packet rates with a slope depending mainly on the number of stations $`N`$. However, for increasing values of $`\lambda`$, the saturation behavior occurs quite fast. Notice that, as exemplified in ([25](#S3.E25 "In III-A Modelling offered load and estimation of probability q ‣ III Markovian Process Analysis and Throughput Computation ‣ Unsaturated Throughput Analysis of IEEE 802.11 in Presence of Non Ideal Transmission Channel and Capture Effects")), $`q\rightarrow 1`$ as $`\lambda\rightarrow\infty`$. Actually, saturated traffic conditions are achieved quite fast for values of $`\lambda`$ on the order of ten packets per second with a number of contending stations greater than or equal to 10.
Fig. [4](#S4.F4 "Figure 4 ‣ IV Simulation Results and Model Validations ‣ Unsaturated Throughput Analysis of IEEE 802.11 in Presence of Non Ideal Transmission Channel and Capture Effects") shows the behavior of the throughput as a function of $`\lambda`$, for three different values of the number of contending stations and for two different SNRs. The capture threshold is $`z_{0} = 24`$dB. We can draw conclusions similar to the ones derived for Fig. [3](#S4.F3 "Figure 3 ‣ IV Simulation Results and Model Validations ‣ Unsaturated Throughput Analysis of IEEE 802.11 in Presence of Non Ideal Transmission Channel and Capture Effects"). Upon comparing the curves shown in Figs [3](#S4.F3 "Figure 3 ‣ IV Simulation Results and Model Validations ‣ Unsaturated Throughput Analysis of IEEE 802.11 in Presence of Non Ideal Transmission Channel and Capture Effects") and [4](#S4.F4 "Figure 4 ‣ IV Simulation Results and Model Validations ‣ Unsaturated Throughput Analysis of IEEE 802.11 in Presence of Non Ideal Transmission Channel and Capture Effects"), it is easily seen that capture effects allow the system throughput to be almost the same independently from the number of stations in saturated conditions, i.e., for high values of $`\lambda`$.
Fig. [4](#S4.F4 "Figure 4 ‣ IV Simulation Results and Model Validations ‣ Unsaturated Throughput Analysis of IEEE 802.11 in Presence of Non Ideal Transmission Channel and Capture Effects") also shows the presence of a peak in the throughput as a function of $`\lambda`$, which characterizes the transition between the linear and saturated throughput. Such a peak tends to manifest itself for increasing values of $`\lambda`$ as the number of stations $`N`$ increases. A comparative analysis of the curves shown in Figs [3](#S4.F3 "Figure 3 ‣ IV Simulation Results and Model Validations ‣ Unsaturated Throughput Analysis of IEEE 802.11 in Presence of Non Ideal Transmission Channel and Capture Effects") and [4](#S4.F4 "Figure 4 ‣ IV Simulation Results and Model Validations ‣ Unsaturated Throughput Analysis of IEEE 802.11 in Presence of Non Ideal Transmission Channel and Capture Effects") reveals that the peak of the throughput tends to disappear because of the presence of capture effects during transmission.
Notice also that the throughput behavior exhibits saturation quite fast for low values of SNR independently from capture effects; this is essentially due to the fact that channel propagation errors tend to dominate over both collisions and capture when the SNR is low.
[Refer to caption](/html/0710.5235/assets/x4.png)
Figure 4: Theoretical and simulated throughput for the 2-way mechanism as a function of the packet rate $`\lambda`$, for three different number of contending stations and two different values of SNR as noted in the legend. Capture threshold is $`z_{0} = 24`$dB. Simulated points are identified by star-markers over the respective theoretical curves. Payload size is 1024 bytes.
In order to assess throughput performances as a function of the payload size, Fig. [5](#S4.F5 "Figure 5 ‣ IV Simulation Results and Model Validations ‣ Unsaturated Throughput Analysis of IEEE 802.11 in Presence of Non Ideal Transmission Channel and Capture Effects") shows the behavior of the throughput as a function of $`\lambda`$, for a payload size equal to 128 bytes. The other simulation parameters are noted in the legend. Beside noting similar conclusions on the effects of both capture and number of contending stations over the throughput behavior as for the previous mentioned figures, notice also that a smaller payload size reduces proportionally the slope of the throughput before saturated conditions, and the maximum achievable throughput in saturated conditions. This issue will be dealt with in the next section.
[Refer to caption](/html/0710.5235/assets/x5.png)
Figure 5: Theoretical and simulated throughput for the 2-way mechanism as a function of the packet rate $`\lambda`$, for three different number of contending stations and two different capture thresholds noted in the legend. Simulated points are identified by star-markers over the respective theoretical curves. Payload size is 128 bytes, while SNR=$`45`$dB.
###
IV-A
Simulation results within ns-2
For the sake of validating the proposed model and our developed simulator against a widely adopted network simulator, while providing simulation results for a new set of parameters, we have performed new simulations using ns-2 (version 2.29) for throughput evaluations. Due to the lack of a complete PHY layer implementation, we made some modifications in the original ns-2 source code in order to account for a specific FER on the transmitted packets. In this respect, we adopted the suggestions proposed in \[[30](#bib.bib30)\] for simulating a 802.11b channel within ns-2. Briefly, channel errors on the transmitted packets are accounted for by using a uniformly distributed binary random variable $`X_{e}`$ with the following probabilities: $`{P\hspace{0pt}{({X_{e} = 1})}} = {P_{e}\hspace{0pt}{({S\hspace{0pt}N\hspace{0pt}R})}}`$ (as specified in ([26](#S4.E26 "In IV Simulation Results and Model Validations ‣ Unsaturated Throughput Analysis of IEEE 802.11 in Presence of Non Ideal Transmission Channel and Capture Effects"))), and $`{P\hspace{0pt}{({X_{e} = 0})}} = {1 - {P_{e}\hspace{0pt}{({S\hspace{0pt}N\hspace{0pt}R})}}}`$. For simplicity, we assume that data packets transmitted by different stations are affected by the same FER.
The employed propagation model is identified by the label `Propagation/Shadowing` within ns-2. We considered a path-loss exponent equal to 3.5 with a zero standard deviation in compliance with the theoretical capture scenario described above. Notice that the standard ns-2 simulator provides only Ad-Hoc support and related routing protocols; for this reason we applied a patch for implementing an Infrastructure BSS without any routing protocol between the various involved stations. This patch is identified by the acronym NOAH, standing for NO Ad-Hoc, and can be downloaded from http://icapeople.epfl.ch/widmer/uwb/ns-2/noah/.
In connection with the employed traffic model, we used an exponential distribution for simulating packet interarrival times.
Throughput evaluation is accomplished by averaging over $`100`$ sample scenarios, whereby any transmitting scenario considers a set of $`N`$ randomly distributed (with a uniform pdf) stations over a circular area of radius 10 m. Simulation results along with theoretical curves are shown in Figs [6](#S4.F6 "Figure 6 ‣ IV-A Simulation results within ns-2 ‣ IV Simulation Results and Model Validations ‣ Unsaturated Throughput Analysis of IEEE 802.11 in Presence of Non Ideal Transmission Channel and Capture Effects")-[8](#S4.F8 "Figure 8 ‣ IV-A Simulation results within ns-2 ‣ IV Simulation Results and Model Validations ‣ Unsaturated Throughput Analysis of IEEE 802.11 in Presence of Non Ideal Transmission Channel and Capture Effects"). The employed parameters are noted in the respective legends of each figure. While the main conclusions related to the throughput behavior as a function of the number of contending stations, capture threshold, traffic load, and FER, are similar to the ones already derived above with our C++ simulator, here we notice a good agreement between theoretical and ns-2 simulation results for a wide range of packet arrival rates, $`\lambda`$.
Figs [6](#S4.F6 "Figure 6 ‣ IV-A Simulation results within ns-2 ‣ IV Simulation Results and Model Validations ‣ Unsaturated Throughput Analysis of IEEE 802.11 in Presence of Non Ideal Transmission Channel and Capture Effects")-[7](#S4.F7 "Figure 7 ‣ IV-A Simulation results within ns-2 ‣ IV Simulation Results and Model Validations ‣ Unsaturated Throughput Analysis of IEEE 802.11 in Presence of Non Ideal Transmission Channel and Capture Effects") also depict the theoretical saturated throughput derived by Bianchi in \[[13](#bib.bib13)\] in order to underline a good matching between our theoretical model, valid for any traffic load in the investigated range of $`\lambda`$ up to the saturation condition for $`\lambda\rightarrow\infty`$.
[Refer to caption](/html/0710.5235/assets/x6.png)
Figure 6: Theoretical and simulated (ns-2) throughput for the 2-way mechanism as a function of the packet rate $`\lambda`$, for two different payload sizes and packet error probabilities, shown in the legend, and without capture. Simulated points are identified by cross-markers over the respective theoretical curves. The number of contending stations is $`N = 10`$.
[Refer to caption](/html/0710.5235/assets/x7.png)
Figure 7: Theoretical and simulated (ns-2) throughput for the 2-way mechanism as a function of the packet rate $`\lambda`$, for two different payload sizes and packet error probabilities, shown in the legend, and without capture. Simulated points are identified by cross-markers over the respective theoretical curves. The number of contending stations is $`N = 5`$.
[Refer to caption](/html/0710.5235/assets/x8.png)
Figure 8: Both subplots show theoretical and simulated (ns-2) throughput for the 2-way mechanism as a function of the packet rate $`\lambda`$, for two different payload sizes and fixed packet error probability, as noted in the legend. Capture threshold $`z_{0} = 1`$dB in both subplots. Simulated points are identified by cross-markers over the respective theoretical curves. Upper subplot refers to a number of contending stations equal to $`N = 10`$, whereas the lower subplot corresponds to $`N = 5`$.
## V Modelling the linear behavior of the throughput in unsaturated conditions
The linear behavior of the throughput in unsaturated traffic conditions, along with its dependence on some key network parameters, can be better understood by analyzing the throughput in ([18](#S3.E18 "In III Markovian Process Analysis and Throughput Computation ‣ Unsaturated Throughput Analysis of IEEE 802.11 in Presence of Non Ideal Transmission Channel and Capture Effects")) as $`\lambda\rightarrow 0`$, i.e., in unloaded traffic conditions. From ([10](#S3.E10 "In III Markovian Process Analysis and Throughput Computation ‣ Unsaturated Throughput Analysis of IEEE 802.11 in Presence of Non Ideal Transmission Channel and Capture Effects")), it is easily seen that the probability $`\tau\rightarrow 0`$ when $`\lambda\rightarrow 0`$, since $`q\rightarrow 0`$. In particular, for very small values of $`\lambda`$, the following approximation from ([10](#S3.E10 "In III Markovian Process Analysis and Throughput Computation ‣ Unsaturated Throughput Analysis of IEEE 802.11 in Presence of Non Ideal Transmission Channel and Capture Effects")) holds:
<table>
<tbody>
<tr>
<td></td>
<td>τ≃q1−Pesimilar-to-or-equals𝜏𝑞1subscript𝑃𝑒\tau\simeq\frac{q}{1-P_{e}}</td>
<td></td>
</tr>
</tbody>
</table>
In addition, it is straightforward to verify that $`P_{c\hspace{0pt}a\hspace{0pt}p}\rightarrow 0`$ (from ([16](#S3.E16 "In III Markovian Process Analysis and Throughput Computation ‣ Unsaturated Throughput Analysis of IEEE 802.11 in Presence of Non Ideal Transmission Channel and Capture Effects"))). As $`\lambda`$ approaches zero, the throughput can be approximated by:
<table>
<tbody>
<tr>
<td></td>
<td>S≃N(1−Pe)E[PL]στ=N⋅E[PL]σq=N⋅E[PL]⋅λsimilar-to-or-equals𝑆𝑁1subscript𝑃𝑒𝐸delimited-[]𝑃𝐿𝜎𝜏⋅𝑁𝐸delimited-[]𝑃𝐿𝜎𝑞⋅⋅𝑁𝐸delimited-[]𝑃𝐿𝜆S\simeq\frac{N(1-P_{e})E[PL]}{\sigma}\tau=\frac{N\cdot E[PL]}{\sigma}q=N\cdot E[PL]\cdot\lambda</td>
<td></td>
<td>(30)</td>
</tr>
</tbody>
</table>
since $`q = {1 - e^{- {\lambda\hspace{0pt}E\hspace{0pt}{\lbrack S_{t\hspace{0pt}s}\rbrack}}}}`$ can be approximated by $`q \simeq {E\hspace{0pt}{\lbrack S_{t\hspace{0pt}s}\rbrack}\hspace{0pt}\lambda}`$ upon employing MacLaurin approximation of the exponential, and $`{E\hspace{0pt}{\lbrack S_{t\hspace{0pt}s}\rbrack}}\rightarrow\sigma`$ for $`\lambda\rightarrow 0`$ as it can be deduced from ([23](#S3.E23 "In III-A Modelling offered load and estimation of probability q ‣ III Markovian Process Analysis and Throughput Computation ‣ Unsaturated Throughput Analysis of IEEE 802.11 in Presence of Non Ideal Transmission Channel and Capture Effects")). The main conclusion that can be drawn from this equation is that as $`\lambda\rightarrow 0`$, the throughput tends to manifest a linear behavior relative to the packet rate $`\lambda`$ with a slope depending on both the number of contending stations and the average payload size $`E\hspace{0pt}{\lbrack{P\hspace{0pt}L}\rbrack}`$.
It is interesting to estimate the interval of validity $`\lbrack 0,\lambda_{c}\rbrack`$ of the linear throughput model proposed in ([30](#S5.E30 "In V Modelling the linear behavior of the throughput in unsaturated conditions ‣ Unsaturated Throughput Analysis of IEEE 802.11 in Presence of Non Ideal Transmission Channel and Capture Effects")). To this end, let us rewrite ([18](#S3.E18 "In III Markovian Process Analysis and Throughput Computation ‣ Unsaturated Throughput Analysis of IEEE 802.11 in Presence of Non Ideal Transmission Channel and Capture Effects")) as follows:
<table>
<tbody>
<tr>
<td></td>
<td>S=E[PL][Ts−Tc1−Pe+TePe1−Pe]+σ1−PtPt+TcPs(1−Pe)𝑆𝐸delimited-[]𝑃𝐿delimited-[]subscript𝑇𝑠subscript𝑇𝑐1subscript𝑃𝑒subscript𝑇𝑒subscript𝑃𝑒1subscript𝑃𝑒𝜎1subscript𝑃𝑡subscript𝑃𝑡subscript𝑇𝑐subscript𝑃𝑠1subscript𝑃𝑒\begin{array}[]{rcl}S&=&\frac{E[PL]}{\left[T_{s}-\frac{T_{c}}{1-P_{e}}+\frac{T_{e}P_{e}}{1-P_{e}}\right]+\frac{\sigma\frac{1-P_{t}}{P_{t}}+T_{c}}{P_{s}(1-P_{e})}}\\ \end{array}</td>
<td></td>
<td>(31)</td>
</tr>
</tbody>
</table>
Since $`T_{s},T_{c},T_{e}`$ and $`P_{e}`$ are independent of $`\tau`$, $`S`$ can be maximized by minimizing the function:
<table>
<tbody>
<tr>
<td></td>
<td>F(τ)=σ1−PtPt+TcPs(1−Pe)𝐹𝜏𝜎1subscript𝑃𝑡subscript𝑃𝑡subscript𝑇𝑐subscript𝑃𝑠1subscript𝑃𝑒F(\tau)=\frac{\sigma\frac{1-P_{t}}{P_{t}}+T_{c}}{P_{s}(1-P_{e})}</td>
<td></td>
<td>(32)</td>
</tr>
</tbody>
</table>
or equivalently, by maximizing the function:
TABLE II: Values of $`\lambda_{c}`$ from ([38](#S5.E38 "In V Modelling the linear behavior of the throughput in unsaturated conditions ‣ Unsaturated Throughput Analysis of IEEE 802.11 in Presence of Non Ideal Transmission Channel and Capture Effects")) for the setup considered in Figs [4](#S4.F4 "Figure 4 ‣ IV Simulation Results and Model Validations ‣ Unsaturated Throughput Analysis of IEEE 802.11 in Presence of Non Ideal Transmission Channel and Capture Effects")-[5](#S4.F5 "Figure 5 ‣ IV Simulation Results and Model Validations ‣ Unsaturated Throughput Analysis of IEEE 802.11 in Presence of Non Ideal Transmission Channel and Capture Effects").
<table>
<tbody>
<tr>
<td>E[PL]=1024 byte</td>
<td>N=4</td>
<td>N=10</td>
<td>N=20</td>
</tr>
<tr>
<td>SNR=20dB</td>
<td>13.9187</td>
<td>5.5372</td>
<td>2.7638</td>
</tr>
<tr>
<td>SNR=45dB</td>
<td>26.7546</td>
<td>10.6444</td>
<td>5.3132</td>
</tr>
<tr>
<td>E[PL]=128 byte</td>
<td>N=4</td>
<td>N=10</td>
<td>N=20</td>
</tr>
<tr>
<td>SNR=45dB</td>
<td>135.0307</td>
<td>53.3990</td>
<td>26.6039</td>
</tr>
</tbody>
</table>
<table>
<tbody>
<tr>
<td></td>
<td>F1(τ)=1F(τ)=Ps(1−Pe)σ1−PtPt+Tcsubscript𝐹1𝜏1𝐹𝜏subscript𝑃𝑠1subscript𝑃𝑒𝜎1subscript𝑃𝑡subscript𝑃𝑡subscript𝑇𝑐F_{1}(\tau)=\frac{1}{F(\tau)}=\frac{P_{s}(1-P_{e})}{\sigma\frac{1-P_{t}}{P_{t}}+T_{c}}</td>
<td></td>
<td>(33)</td>
</tr>
</tbody>
</table>
with respect to $`\tau`$. Upon substituting ([19](#S3.E19 "In 1st item ‣ III Markovian Process Analysis and Throughput Computation ‣ Unsaturated Throughput Analysis of IEEE 802.11 in Presence of Non Ideal Transmission Channel and Capture Effects")) and ([20](#S3.E20 "In 2nd item ‣ III Markovian Process Analysis and Throughput Computation ‣ Unsaturated Throughput Analysis of IEEE 802.11 in Presence of Non Ideal Transmission Channel and Capture Effects")) in $`F_{1}\hspace{0pt}{(\tau)}`$, it is possible to write:
<table>
<tbody>
<tr>
<td></td>
<td>F1(τ)=Nτ(1−Pe)σ(1−τ)+Tc(1−τ)1−N−Tc(1−τ)subscript𝐹1𝜏𝑁𝜏1subscript𝑃𝑒𝜎1𝜏subscript𝑇𝑐superscript1𝜏1𝑁subscript𝑇𝑐1𝜏F_{1}(\tau)=\frac{N\tau(1-P_{e})}{\sigma(1-\tau)+T_{c}(1-\tau)^{1-N}-T_{c}(1-\tau)}</td>
<td></td>
<td>(34)</td>
</tr>
</tbody>
</table>
Differentiating $`F_{1}\hspace{0pt}{(\tau)}`$ with respect to $`\tau`$ and equating it to zero, it is possible to write:
<table>
<tbody>
<tr>
<td></td>
<td>(1−τ)N(σ−Tc)−Tc(Nτ−1)=0superscript1𝜏𝑁𝜎subscript𝑇𝑐subscript𝑇𝑐𝑁𝜏10\begin{array}[]{rcl}(1-\tau)^{N}(\sigma-T_{c})-T_{c}(N\tau-1)&=&0\\ \end{array}</td>
<td></td>
<td>(35)</td>
</tr>
</tbody>
</table>
Under the hypothesis $`\tau \ll 1`$, the following approximation holds:
<table>
<tbody>
<tr>
<td></td>
<td>(1−τ)N≈1−Nτ+N(N−1)2τ2+o(τ)superscript1𝜏𝑁1𝑁𝜏𝑁𝑁12superscript𝜏2𝑜𝜏(1-\tau)^{N}\approx 1-N\tau+\frac{N(N-1)}{2}\tau^{2}+o(\tau)</td>
<td></td>
</tr>
</tbody>
</table>
By substituting the previous approximation in ([35](#S5.E35 "In V Modelling the linear behavior of the throughput in unsaturated conditions ‣ Unsaturated Throughput Analysis of IEEE 802.11 in Presence of Non Ideal Transmission Channel and Capture Effects")), it is possible to obtain the value of $`\tau`$ maximizing $`F_{1}\hspace{0pt}{(\tau)}`$:
<table>
<tbody>
<tr>
<td></td>
<td>τm=σ−σ[Nσ−2(N−1)(σ−Tc)]/N(N−1)(σ−Tc)subscript𝜏𝑚𝜎𝜎delimited-[]𝑁𝜎2𝑁1𝜎subscript𝑇𝑐𝑁𝑁1𝜎subscript𝑇𝑐\tau_{m}=\frac{\sigma-\sqrt{\sigma\left[N\sigma-2(N-1)(\sigma-T_{c})\right]/N}}{(N-1)(\sigma-T_{c})}</td>
<td></td>
<td>(36)</td>
</tr>
</tbody>
</table>
Substituting $`\tau_{m}`$ in ([31](#S5.E31 "In V Modelling the linear behavior of the throughput in unsaturated conditions ‣ Unsaturated Throughput Analysis of IEEE 802.11 in Presence of Non Ideal Transmission Channel and Capture Effects")), it is possible to obtain the maximum throughput $`S_{m}`$:
<table>
<tbody>
<tr>
<td></td>
<td>Sm=E[PL][Ts−Tc1−Pe+TePe1−Pe]+(σ−Tc)(1−τm)N+TcNτm(1−τm)N−1(1−Pe)subscript𝑆𝑚𝐸delimited-[]𝑃𝐿delimited-[]subscript𝑇𝑠subscript𝑇𝑐1subscript𝑃𝑒subscript𝑇𝑒subscript𝑃𝑒1subscript𝑃𝑒𝜎subscript𝑇𝑐superscript1subscript𝜏𝑚𝑁subscript𝑇𝑐𝑁subscript𝜏𝑚superscript1subscript𝜏𝑚𝑁11subscript𝑃𝑒S_{m}=\frac{E[PL]}{\left[T_{s}-\frac{T_{c}}{1-P_{e}}+\frac{T_{e}P_{e}}{1-P_{e}}\right]+\frac{(\sigma-T_{c})(1-\tau_{m})^{N}+T_{c}}{N\tau_{m}(1-\tau_{m})^{N-1}(1-P_{e})}}</td>
<td></td>
<td>(37)</td>
</tr>
</tbody>
</table>
Equating $`S_{m}`$ to the linear model in ([30](#S5.E30 "In V Modelling the linear behavior of the throughput in unsaturated conditions ‣ Unsaturated Throughput Analysis of IEEE 802.11 in Presence of Non Ideal Transmission Channel and Capture Effects")), it is possible to obtain the value $`\lambda_{c}`$ for which the linear model reaches the beginning of the saturation zone, i.e., the maximum $`\lambda`$ above which the linear model loses validity:
<table>
<tbody>
<tr>
<td></td>
<td>λc=1N[Ts−Tc1−Pe+TePe1−Pe]+(σ−Tc)(1−τm)N+Tcτm(1−τm)N−1(1−Pe)subscript𝜆𝑐absent1𝑁delimited-[]subscript𝑇𝑠subscript𝑇𝑐1subscript𝑃𝑒subscript𝑇𝑒subscript𝑃𝑒1subscript𝑃𝑒𝜎subscript𝑇𝑐superscript1subscript𝜏𝑚𝑁subscript𝑇𝑐subscript𝜏𝑚superscript1subscript𝜏𝑚𝑁11subscript𝑃𝑒\begin{array}[]{rl}\lambda_{c}=&\frac{1}{N\left[T_{s}-\frac{T_{c}}{1-P_{e}}+\frac{T_{e}P_{e}}{1-P_{e}}\right]+\frac{(\sigma-T_{c})(1-\tau_{m})^{N}+T_{c}}{\tau_{m}(1-\tau_{m})^{N-1}(1-P_{e})}}\\ \end{array}</td>
<td></td>
<td>(38)</td>
</tr>
</tbody>
</table>
The linear model ([30](#S5.E30 "In V Modelling the linear behavior of the throughput in unsaturated conditions ‣ Unsaturated Throughput Analysis of IEEE 802.11 in Presence of Non Ideal Transmission Channel and Capture Effects")) shows a close agreement with both theoretical and simulated throughput curves. Fig. [9](#S5.F9 "Figure 9 ‣ V Modelling the linear behavior of the throughput in unsaturated conditions ‣ Unsaturated Throughput Analysis of IEEE 802.11 in Presence of Non Ideal Transmission Channel and Capture Effects") shows the straight lines ([30](#S5.E30 "In V Modelling the linear behavior of the throughput in unsaturated conditions ‣ Unsaturated Throughput Analysis of IEEE 802.11 in Presence of Non Ideal Transmission Channel and Capture Effects")) for three different values of the number of contending stations $`N`$, compared with the theoretical curves already depicted in Fig. [4](#S4.F4 "Figure 4 ‣ IV Simulation Results and Model Validations ‣ Unsaturated Throughput Analysis of IEEE 802.11 in Presence of Non Ideal Transmission Channel and Capture Effects"). The figure also shows the values of $`S_{m}`$ as derived from ([37](#S5.E37 "In V Modelling the linear behavior of the throughput in unsaturated conditions ‣ Unsaturated Throughput Analysis of IEEE 802.11 in Presence of Non Ideal Transmission Channel and Capture Effects")) along with the three values of $`\lambda_{c}`$ deduced from ([38](#S5.E38 "In V Modelling the linear behavior of the throughput in unsaturated conditions ‣ Unsaturated Throughput Analysis of IEEE 802.11 in Presence of Non Ideal Transmission Channel and Capture Effects")). Notice that $`\lambda_{c}`$, besides being the limit of validity of the linear model, can also be interpreted as the value of packet rate above which the throughput enters saturation conditions.
In order to further verify the values of $`\lambda_{c}`$ deduced from ([38](#S5.E38 "In V Modelling the linear behavior of the throughput in unsaturated conditions ‣ Unsaturated Throughput Analysis of IEEE 802.11 in Presence of Non Ideal Transmission Channel and Capture Effects")), Table [II](#S5.T2 "TABLE II ‣ V Modelling the linear behavior of the throughput in unsaturated conditions ‣ Unsaturated Throughput Analysis of IEEE 802.11 in Presence of Non Ideal Transmission Channel and Capture Effects") shows the values of $`\lambda_{c}`$ related to both simulation and theoretical results depicted in Figs [4](#S4.F4 "Figure 4 ‣ IV Simulation Results and Model Validations ‣ Unsaturated Throughput Analysis of IEEE 802.11 in Presence of Non Ideal Transmission Channel and Capture Effects")-[5](#S4.F5 "Figure 5 ‣ IV Simulation Results and Model Validations ‣ Unsaturated Throughput Analysis of IEEE 802.11 in Presence of Non Ideal Transmission Channel and Capture Effects"). In particular, values shown in the upper part of the table, i.e., the ones labelled E\[PL\]=1024, are related to the curves in Fig. [4](#S4.F4 "Figure 4 ‣ IV Simulation Results and Model Validations ‣ Unsaturated Throughput Analysis of IEEE 802.11 in Presence of Non Ideal Transmission Channel and Capture Effects"), whereas the other values are related to the results in Fig. [5](#S4.F5 "Figure 5 ‣ IV Simulation Results and Model Validations ‣ Unsaturated Throughput Analysis of IEEE 802.11 in Presence of Non Ideal Transmission Channel and Capture Effects"). Notice the good agreement between theoretical and simulated results even in the presence of capture. The key observation here can be deduced by a comparative analysis of the results shown in Figs. [3](#S4.F3 "Figure 3 ‣ IV Simulation Results and Model Validations ‣ Unsaturated Throughput Analysis of IEEE 802.11 in Presence of Non Ideal Transmission Channel and Capture Effects")-[4](#S4.F4 "Figure 4 ‣ IV Simulation Results and Model Validations ‣ Unsaturated Throughput Analysis of IEEE 802.11 in Presence of Non Ideal Transmission Channel and Capture Effects"): capture tends to simply shape the peak on which the throughput reaches its maximum, leaving the abscissa of the maximum, that is $`\lambda_{c}`$, in the same position.
[Refer to caption](/html/0710.5235/assets/x9.png)
Figure 9: Throughput for the 2-way mechanism as a function of the packet rate $`\lambda`$, for three different number of contending stations. Capture threshold and SNR are set to very high values since in pre-saturation regime. Payload size is 1024 bytes. Straight lines refer to the linear model of the throughput derived in ([30](#S5.E30 "In V Modelling the linear behavior of the throughput in unsaturated conditions ‣ Unsaturated Throughput Analysis of IEEE 802.11 in Presence of Non Ideal Transmission Channel and Capture Effects")).
## VI Conclusions
In this paper, we have provided an extension of the Markov model characterizing the DCF behavior at the MAC layer of the IEEE802.11 series of standards by accounting for channel induced errors and capture effects typical of fading environments under unsaturated traffic conditions. The modelling allows taking into consideration the impact of channel contention in throughput analysis which is often not considered or it is considered in a static mode by using a mean contention period. Subsequently, based on justifiable assumptions, the stationary probability of the Markov chain is calculated to obtain the behavior of the throughput in both unsaturated and saturated conditions. The closed form expressions allow derivation of the throughput as a function of a multitude of system level parameters including packet and header sizes for a variety of applications. Simulation results confirm the validity of the proposed theoretical models.
## References
- \[1\]
IEEE Standard for Wireless LAN Medium Access Control (MAC) and Physical Layer (PHY) Specifications, November 1997, P802.11
- \[2\]
Wei Zhuang, Yung-Sze Gan, Kok-Jeng Loh, and Kee-Chaing Chua, ”Policy-based QoS-management architecture in an integrated UMTS and WLAN environment”, IEEE Communications Magazine, Vol.41, No.11, pp.118-125, Nov. 2003.
- \[3\]
W. Liu, W. Lou, X. Chen, and Y. Fang, ”A QoS-enabled MAC architecture for prioritized service in IEEE 802.11 WLANs”, Proc. IEEE GLOBECOM’03, Vol.7, pp.3802-3807, Dec. 2003.
- \[4\]
G. Fodor, A. Eriksson, and A. Tuoriniemi, ”Providing quality of service in always best connected networks”, IEEE Communications Magazine, Vol.41, No.7, pp.154-163, July 2003.
- \[5\]
J. Zhao, Z. Guo, Q. Zhang, and W. Zhu, ”Performance study of MAC for service differentiation in IEEE 802.11”, In Proc. of IEEE GLOBECOM’02, Vol.1, pp.778-782, Nov. 2002.
- \[6\]
W. Pattara-Atikom, P. Krishnamurthy, and S. Banerjee, ”Distributed mechanisms for quality of service in wireless LANs”, IEEE Personal Communications, Vol.10, No.3, pp.26-34, June 2003.
- \[7\]
Ye Ge and J. Hou, ”An analytical model for service differentiation in IEEE 802.11”, Proc. ICC ’03, Vol.2, pp.1157-1162, May 2003.
- \[8\]
H. L. Truong and G. Vannuccini, ”Performance evaluation of the QoS enhanced IEEE 802.11e MAC layer”, In Proc. of IEEE VTC 2003-Spring, Vol.2, pp.940-944, April 2003.
- \[9\]
Q. Li and M. VanderSchaar, ”Providing adaptive QoS to layered video over wireless local area networks through real-time retry limit adaptation”, IEEE Trans. on Multimedia, Vol.6, No.2, pp.278-290, April 2004.
- \[10\]
Q. Zhang, C. Guo, Z. Guo, and W. Zhu, ”Efficient mobility management for vertical handoff between WWAN and WLAN”, IEEE Communications Magazine, Vol.41, No.11, pp.102-108, Nov. 2003.
- \[11\]
H. Honkasalo, K. Pehkonen, M. T. Niemi, and A.T. Leino, ”WCDMA and WLAN for 3G and beyond”, IEEE Personal Communications, Vol.9, No.2, pp.14-18, April 2002.
- \[12\]
S. Park, K. Kim, D.C. Kim, S. Choi, and S. Hong, ”Collaborative QoS architecture between DiffServ and 802.11e wireless LAN”, In Proc. of IEEE VTC 2003-Spring, Vol.2, pp.945-949, April 2003.
- \[13\]
G. Bianchi, ”Performance analysis of the IEEE 802.11 distributed coordination function”, IEEE JSAC, Vol.18, No.3, March 2000.
- \[14\]
L. Yong Shyang, A. Dadej, and A.Jayasuriya, ”Performance analysis of IEEE 802.11 DCF under limited load”, In Proc. of Asia-Pacific Conference on Communications, Vol.1, pp.759 - 763, 03-05 Oct. 2005.
- \[15\]
M. Ergen and P. Varaiya, ”Throughput analysis and admission control in IEEE 802.11a”, Springer Mobile Networks and Applications, vol. 10, no. 5., pp. 705-706, October 2005.
- \[16\]
J.W. Robinson and T.S. Randhawa ”Saturation throughput analysis of IEEE 802.11e enhanced distributed coordination function”, IEEE JSAC, Vol.22, No.5, June 2004.
- \[17\]
Z. Kong, D.H.K. Tsang, B.Bensaou, and D.Gao, ”Performance analysis of IEEE 802.11e contention-based channel access”, IEEE JSAC, Vol.22, No.10, Dec. 2004.
- \[18\]
D. Qiao, S. Choi, and K.G. Shin ”Goodput analysis and link adaptation for IEEE 802.11a wireless LANs”, IEEE Trans. On Mobile Computing, Vol.1, No.4, Oct.-Dec. 2002.
- \[19\]
P. Chatzimisios, A.C. Boucouvalas, and V. Vitsas, ”Influence of channel BER on IEEE 802.11 DCF”, IEE Electronics Letters, Vol.39, No.23, pp.1687-1689, Nov. 2003.
- \[20\]
Q. Ni, T. Li, T. Turletti, and Y. Xiao, ”Saturation throughput analysis of error-prone 802.11 wireless networks”, Wiley Journal of Wireless Communications and Mobile Computing, Vol. 5, No. 8, pp. 945-956, Dec. 2005.
- \[21\]
K. Duffy, D. Malone, and D.J. Leith, ”Modeling the 802.11 distributed coordination function in non-saturated conditions”, IEEE Communications Letters, Vol.9, No.8, pp.715-717, Aug. 2005.
- \[22\]
Jae Hyun Kim and Jong Kyu Lee ”Capture effects of wireless CSMA/CA protocols in Rayleigh and shadow fading channels”, IEEE Trans. on Veh. Tech., Vol.48, No.4, pp.1277-1286, July 1999.
- \[23\]
M. Zorzi and R.R. Rao, ”Capture and retransmission control in mobile radio”, IEEE JSAC, Vol.12, No.8, pp.1289 - 1298, Oct. 1994.
- \[24\]
Z. Hadzi-Velkov and B. Spasenovski, ”Capture effect in IEEE 802.11 basic service area under influence of Rayleigh fading and near/far effect”, In Proc. of 13th IEEE International Symposium on Personal, Indoor and Mobile Radio Communications, Vol.1, pp.172 - 176, Sept. 2002.
- \[25\]
M. B. Pursley, ”Performance evaluation for phase-coded spread spectrum multiple-access communication Part I: System analysis”, IEEE Trans. Commun., vol. COM-25, pp. 795 799, Aug. 1977.
- \[26\]
T. S. Ho and K. C. Chen, ”Performance evaluation and enhancement of the CSMA/CA MAC protocol for 802.11 wireless LAN s”, In Proc. of IEEE PIMRC, Taipei, Taiwan, Oct. 1996, pp. 392 396.
- \[27\]
D. Malone, K. Duffy, and D.J. Leith, ”Modeling the 802.11 distributed coordination function in non-saturated heterogeneous conditions”, IEEE-ACM Trans. on Networking, vol. 15, No. 1, pp. 159 172, Feb. 2007.
- \[28\]
M.K. Simon and M. Alouini, Digital Communication over Fading Channels: A Unified Approach to Performance Analysis, Wiley-Interscience, 1st edition, 2000.
- \[29\]
T. Rappaport, Wireless Communications, Prentice Hall, 1999.
- \[30\] Wu Xiuchao, ”Simulate 802.11b Channel within NS-2”, available online at `http://www.comp.nus.edu.sg/~wuxiucha/research/` `reactive/report/80211ChannelinNS2_new.pdf` |
0711.1533 | cs.NI | cyber | # N3Logic: A Logical Framework For the World Wide Web
Tim Berners-Lee
Dan Connolly
Lalana Kagal
Yosi Scharf
Computer Science and Artificial Intelligence lab
Massachusetts Institute of Technology
{timbl
connolly
lkagal
syosi}@csail.mit.edu
Jim Hendler
Rensselaer Polytechnic Institute
hendler@cs.rpi.edu
(11 Sep 2007; 18 October 2007)
###### Abstract
The Semantic Web drives towards the use of the Web for interacting with logically interconnected data. Through knowledge models such as Resource Description Framework (RDF), the Semantic Web provides a unifying representation of richly structured data. Adding logic to the Web implies the use of rules to make inferences, choose courses of action, and answer questions. This logic must be powerful enough to describe complex properties of objects but not so powerful that agents can be tricked by being asked to consider a paradox. The Web has several characteristics that can lead to problems when existing logics are used, in particular, the inconsistencies that inevitably arise due to the openness of the Web, where anyone can assert anything. N3Logic is a logic that allows rules to be expressed in a Web environment. It extends RDF with syntax for nested graphs and quantified variables and with predicates for implication and accessing resources on the Web, and functions including cryptographic, string, math. The main goal of N3Logic is to be a minimal extension to the RDF data model such that the same language can be used for logic and data. In this paper, we describe N3Logic and illustrate through examples why it is an appropriate logic for the Web.
###### keywords:
logic, Web, Semantic Web, scoped negation, quoting, RDF, N3
## 1 Introduction
The Semantic Web is an enhancement of the World Wide Web to broaden its sharing capacity to application data sharing as well as human-focused data sharing. Through RDF the Semantic Web provides a unifying representation of richly structured data. And through the Web, application developers can share these data representations so that their applications can make “decisions” on the basis of combining the many different kinds of data published on the Web.
RDF builds on the fundamental pointer mechanism of the Web; the Uniform Resource Identifier (URI). In the initial incarnation of the Web, URIs were generally thought to refer to documents and parts of documents via hypertext anchors. The Semantic Web makes it explicit that URIs can be used to name anything – from abstract concepts (“color”) to physical objects (“the building in which MIT’s CSAIL personnel work”) to electronic objects (“the machine code that implements the Linux operating system”). RDF uses URIs to give names to relationships between objects as well as to the objects themselves.
The abstract representation of RDF \[[Carroll and Klyne (2004)](#bib.bib14)\] is a directed labeled graph – i.e. nodes and arcs. A subset of the nodes are URIs; these are “named” nodes. Other nodes may not be identifying URIs but the graph can still describe properties of such nodes including relationships with other named and un-named nodes. The properties and relationships are the edges in the graph. Every edge type has its own label and these same edge type labels themselves can be used to name other nodes in the graph that represent the edge type. This permits properties of edge types themselves to be represented in an RDF graph.
N3 is a compact and readable alternative to RDF’s XML syntax \[[Berners-Lee (1998)](#bib.bib8)\]. N3 allows RDF to be expressed but emphasizes readability and symmetry. It also allows quoting or statements to be made about statements. This quoting feature allows users to distinguish between what they believe to be true and what someone else including a website states or believes.
N3Logic uses N3 syntax and extends RDF with a vocabulary of predicates. N3 aims to do for logical information what RDF does for data: to provide a common data model and a common syntax, so that extensions of the language are made simply by defining new terms in an ontology. Declarative programming languages such as Scheme \[[(27)](#bib.bib27)\] already do this. However, they differ in their choice of pairs rather than the RDF binary relational model for data, and lack the use of universal identifiers as symbols. N3Logic allows rules to be integrated smoothly with RDF and provides certain essential built-in functions that allow information from the Web to be accessed and reasoned over. The main goal of N3Logic is to make a minimal extension to the RDF data model so the same language could be used for logic and data, and to do so in a way that is compatible with the architectural principles of the Web
In this paper, we discuss the features of N3Logic with the help of examples and describe its informal semantics.
## 2 N3Logic Overview
One of the main motivations of N3Logic is to be useful as a tool in the open Web environment. The Web contains many sources of information, with different characteristics and relationships to any given reader. Whereas a closed system may be built based on a single knowledge base of believed facts, an open Web-based system exists in an unbounded sea of interconnected information resources. This requires that an entity be aware of the provenance of information and responsible for its disposition. A language for use in this environment typically requires the ability to express which document or message asserts what. We found that quoting provides a pragmatic solution to this. However, quotation and reference, with its inevitable possibility of direct or indirect self-reference, if added directly to first order logic presents problems such as paradox traps. To avoid this, N3Logic has deliberately been kept to limited expressive power: it currently contains no general first order negation.
Another goal of N3Logic is that information, such as but not limited to rules, which requires greater expressive power than the RDF graph, should be sharable in the same way as RDF can be shared. This means that one person should be able to express knowledge in N3Logic for a certain purpose, and later independently someone else can reuse that knowledge for a different unforeseen purpose. As the context of the latter use is unknown, this prevents us from making implicit closed assumptions about the total set of knowledge in the system as a whole.
Further, users of the Web have the ability to express new knowledge without affecting systems that are already built. We’ve chosen to adopt a monotonicity requirement for N3Logic because we find it scales well. This implies that the addition of new information from elsewhere cannot silently change the meaning of the original knowledge, though it might cause an inconsistency by contradicting the old information. The non-monotonicity of many existing systems follows from a form of negation as failure (NAF) in which a sentence is deemed false if it not held within (or, derivable from) the current knowledge base. It is this concept of current knowledge base, which is a variable quantity, and the ability to indirectly make reference to it which causes the non-monotonicity. In N3Logic, while a current knowledge base is a fine concept, there is no ability to make references to it implicitly in the negative. The negation provided is called Scoped Negation As Failure (SNAF) and is the ability for a specific given document (or, essentially, some abstract formula) to objectively determine whether or not it holds, or allows one to derive, a given fact. However, negated forms of many of the built-in functions are available. (Please refer to Sections [4.3](#S4.SS3 "4.3 N3Logic Vocabulary ‣ 4 Notions and Terminology ‣ N3Logic: A Logical Framework For the World Wide Web") and [5.5](#S5.SS5 "5.5 Graph Functions ‣ 5 Informal Semantics of N3Logic ‣ N3Logic: A Logical Framework For the World Wide Web") for more information about scoped negation and built-in functions)
## 3 Motivating Example
We describe a Web-based scenario that will be used to illustrate the different features of N3 logic. Consider a conference management system that handles different aspects of registration for conferences. It allows people to register by specifying their names, addresses, affiliations, and their Friend-of-a-Friend (foaf) page. A foaf page includes information such as the organization the registrant works for, her/his current and past projects, and her/his interests \[[Brickley and Miller (1999)](#bib.bib13), [Dumbill (2002)](#bib.bib15)\]. Using this information, the conference management system goes out onto the Web and retrieves relevant information. By reasoning over this information it is able to make inferences about the registrant such as whether the registrant is a vegetarian or not, which workshops she/he would be most interested in, whether she/he is a member of a certain professional organization, and whether the registrant is a student. This allows the conference management system to provide greater support in the registration process by figuring out what registration fees are applicable, whether to order vegetarian meals, and by suggesting appropriate workshops. In the following sections, we define N3Logic in detail, using the example of this conference system to illustrate key features.
## 4 Notions and Terminology
N3 is based on the abstract syntax of RDF. The concrete syntax of N3 includes a number of other abbreviations. Please refer to the Appendix [11](#S11 "11 APPENDIX : N3 Concrete Syntax ‣ N3Logic: A Logical Framework For the World Wide Web") or to the N3 Primer \[[Berners-Lee (2005)](#bib.bib10)\] for a tutorial introduction.
### 4.1 Basic Concepts from RDF
- •
The atomic formulas in the RDF abstract syntax are called “triples”; they are analogous to one 3-place holds(s, p, o) predicate. For example:
```
<http://dig.csail.mit.edu/2006/Papers/TPLP/example/exconf#ExConf>
<http://www.w3.org/1999/02/22-rdf-syntax-ns#type> <http://example.org/conf#Conference>
```
- •
RDF also has conjunctions of formulas. URI terms can be abbreviated using namespaces, and the keyword ’a’ is short for $`<`$http://www.w3.org/1999/02/22-rdf-syntax-ns#type$`>`$.
```
@prefix conf: <http://example.org/conf#> .
@prefix : <http://dig.csail.mit.edu/2006/Papers/TPLP/example/exconf#> .
ExConf a conf:Conference .
ExConf conf:homepage <http://www.l3s.de/~olmedilla/events/MTW06_Workshop.html>.
ExConf conf:registrant Judy.
Judy a foaf:Person.
<http://dig.csail.mit.edu/2005/09/rein/examples/judy-foaf.rdf> foaf:maker Judy.
```
- •
RDF also has literal terms; for example:
```
ExConf conf:eventName "WWW2006 Workshop on Models of Trust for the Web".
ExConf conf:numOfRegistrants 65.
```
- •
Finally, RDF abstract syntax allows existential quantification. Certain syntaxes (e.g. “\[“…”\]”) allow an existential variable to be introduced without having to name it: this is known as a blank node. The following are equivalent:
```
@forSome X. j:Joe foaf:knows X. X foaf:name "Fred" .
j:Joe foaf:knows [ foaf:name "Fred" ] .
```
### 4.2 N3 Extension to RDF
N3 extends the abstract syntax of RDF in two ways:
- •
It has all of the terms of RDF plus quoted formulas. For example:
```
b:mary says { j:Joe foaf:schoolHomepage <http://example.edu> } .
```
- •
It has all of the formulas of RDF plus universally quantified formulas. In simple cases, the @forAll quantifier can be left implicit. The following are equivalent:
```
@forAll X. { X a Man } log:implies { X a Mortal }.
{ ?X a Man } log:implies { ?X a Mortal }.
```
The N3 example below declares namespace prefixes and defines ExConf an instance of the Conference class as defined in conf namespace.
```
@keywords a.
@prefix conf: <http://example.org/conf#> .
@prefix : <http://dig.csail.mit.edu/2006/Papers/TPLP/example/exconf#> .
ExConf a conf:Conference;
conf:eventName "WWW2006 Workshop on Models of Trust for the Web";
conf:acronym "MTW06";
conf:address "mtw@www.org";
conf:homepage <http://www.l3s.de/~olmedilla/events/MTW06_Workshop.html> .
```
### 4.3 N3Logic Vocabulary
N3Logic uses the N3 syntax and also includes a set of predicates. Its vocabulary is union of the N3 syntax and the set of URI references defined in the log: (http://www.w3.org/2000/10/swap/log#), crypto: (http://www.w3.org/2000/10/swap/crypto#), list: (http://www.w3.org/2000/10/swap/list#), math: (http://www.w3.org/2000/10/swap/math#), os: (http://www.w3.org/2000/10/swap/os#), string: (http://www.w3.org/2000/10/swap/string#), and time: (http://www.w3.org/2000/10/swap/time#) namespaces as shown in Table [1](#T1 "Table 1 ‣ 4.3 N3Logic Vocabulary ‣ 4 Notions and Terminology ‣ N3Logic: A Logical Framework For the World Wide Web").
<table>
<tbody>
<tr>
<td> log:conclusion, log:content, log:includes, log:semantics, log:notIncludes, log:supports … crypto:md5, crypto:sign, crypto:verify … list:in, list:last … math:lessThan, math:greaterThan … os:argv, os:environ … string:contains, string:endsWith, string:scrape … time:day, time:hour, time:minute …</td>
</tr>
<tr>
<td> </td>
</tr>
</tbody>
</table>
Table 1: Some N3Logic predicates
While N3Logic properties can be used simply as ground facts in a database, is very useful to take advantage of the fact that in fact they can be calculated. N3Logic includes axiom schemas for each of these terms; reasoners can use these axioms to evaluate formulas and bind variables. These are called built-in functions and they can be used to provide a variety of functionality, for example the crypto:sha1 built-in allows the object to be computed as the SHA-1 hash of the subject.
## 5 Informal Semantics of N3Logic
Various vocabularies, notably RDFS and OWL, have defined RDF predicates with logical semantics, such as rdfs:range, and owl:sameAs, etc. Of these, N3 logic uses rdf:type and owl:sameAs, and defines further predicates to allow rules, web access and built-in calculated functions. N3Logic extends RDF in two ways (i) a syntax namely N3, and (2) a vocabulary of new predicates, which can be used to talk about the provenance of information, contents of documents on the Web, and provide a variety of useful functionality such as string, cryptographic, and mathematic functions.
N3Logic allows statements to be made about, and to query, other statements such as the content of data in information resources on the Web. Formulae provide the ability to represent such sets of statements. To allow statements about them, some of the relationships defined are given URIs so that these statements and queries can be written in N3.
The fact that the rule language and the data language are the same gives a certain simplicity (there is only one syntax) and completeness (rules can operate on themselves, anything written in N3 can be queried in N3). This would be broken if a special syntax were added for built-in functions and operators. Instead, these are simply represented as RDF properties. Rules may have full N3, even with nested graphs, on both sides of the implication. This gives a form of completeness as rules can generate rules.
### 5.1 Relationship with RDF, RDFS, and OWL
N3 syntax allows RDF to be expressed, however, it does not make use of the full RDF vocabulary.
- •
In N3 the keyword “a” is a shorthand for rdf:type and can be replaced with a direct use of the full URI symbol for rdf:type.
- •
The shorthand notation of “=” refers to owl:sameAs. It is used to state that the subject and object are equal.
- •
N3 uses rdf:List, rdf:first, rdf:rest, and rdf:nil for describing lists. Implementations may treat list as a data type rather than just a ladder of rdf:first and rdf:rest properties. The use of rdf:first and rdf:rest can be seen as a reification of the list datatype. This use of lists requires more axioms than are actually defined in the RDF specification. These axioms, described in N3, are given below
- –
Existence of lists
```
(?X).
{ ?L rdf:rest [ ] } => { [ ] rdf:first ?X; rdf:rest ?L }.
```
- –
Uniqueness of lists
```
{ ?L1 rdf:first ?X; rdf:rest ?R.
?L1 rdf:first ?X; rdf:rest ?R. } => { ?L1 = ?L2 }.
```
- –
Uniqueness of the rdf:first and rdf:rest of a list
```
{ ?S rdf:first ?O1, ?O2 } => { ?O1 = ?O2 }.
{ ?S rdf:rest ?O1, ?O2 } => { ?O1 = ?O2 }.
```
- •
The semantics of RDFS can be easily expressed in N3Logic. A set of N3Logic rules for defining rdf:domain and rdf:range are as follows
```
{ ?S [ s:domain ?C ] ?O } => { ?S a ?C } .
{ ?S [ s:range ?C ] ?O } => { ?O a ?C } .
{ ?S a [ s:subClassOf ?C ] } => { ?S a ?C } .
```
### 5.2 Quoted N3 Formulae
Quoting is an important feature provided by N3. Various forms of literal value are allowed in RDF graphs, however the RDF standard does not itself provide for another RDF graph to be a data value. Remedying this allows one to express relationships between graphs, for example that a given graph is the RDF content of a particular document. Every RDF graph is a subclass of N3 formula. A quoted N3 formula is an N3 formula enclosed in braces “{“ “}”. Some examples of N3 formulae and quoted N3 formulae are shown in Table [2](#T2 "Table 2 ‣ 5.2 Quoted N3 Formulae ‣ 5 Informal Semantics of N3Logic ‣ N3Logic: A Logical Framework For the World Wide Web").
Note that, however, substitution of equal (owl:sameAs) terms occurring in a formula does not take place within nested quoted formulae: formulae are not referentially transparent.
In our conference management example, we assume that there are several administrators. Each administrator specifies a list of people who may not register due to some reason. The example below states that Joe says that Peter is not permitted to register. Even though Peter is equal to John, it does not imply that Joe says that John is not permitted to register. This is an example of a quoted N3 formula.
```
j:Joe says { mit:Peter policy:notpermitted conf:Register }.
mit:Peter = cmu:John.
```
<table>
<tbody>
<tr>
<td> N3 formula : c1 is a Car as defined in the “ex” namespace and its color is green. @keywords a. @prefix ex: <http://example.org/car.n3#> . c1 a ex:Car; ex:color "green".</td>
</tr>
<tr>
<td> N3 formula : The semantics of <<http://www.example.org/myfoaf.rdf>> as expressed in N3 is stored into a variable, F. @forAll F. <http://www.example.org/myfoaf.rdf> log:semantics F .</td>
</tr>
<tr>
<td> Quoted N3 formula : Joe believes that Peter is a graduate student. j:Joe believes { mit:Peter a school:GraduateStudent } .</td>
</tr>
<tr>
<td> Quoted N3 Formula : Mary believes that Joe believes that Peter is a graduate student. Mary believes { j:Joe believes { mit:Peter a school:GraduateStudent } } .</td>
</tr>
<tr>
<td> </td>
</tr>
</tbody>
</table>
Table 2: N3 Formula Examples
### 5.3 Dereferencing URIs
The Web is exposed as a mapping between URIs and the information returned when such a URI is dereferenced, using appropriate protocols. In N3Logic, the information resource is identified by a symbol, which is in fact is its URI. The information is represented in formulae, the information retrieved is also represented as a formula. Not all information on the Web is, of course in N3. If we assume that N3 is the interlingua, then from the point of view of this system, the semantics of a document is exactly what can be expressed in N3.
log:semantics is a logical property that represents the relation between a document and the logical expression which represents its meaning expressed as N3. The log:semantics of an N3 document is the formula achieved by parsing the representation of the document.
In order to access a foaf file and store its N3 representation to a variable, log:semantics is used in the following manner
```
<http://www.example.org/myfoaf.rdf> log:semantics ?F.
```
The Architecture of the World Wide Web defines algorithms by which a machine can determine representations of document given its symbol - URI. For a representation in N3, this is the formula which corresponds to the document production of the grammar. For a representation in RDF/XML, it is the formula which is the entire graph parsed. For any other languages, it may be calculated as long as a specification exists that defines the equivalent N3 semantics for files in that language. The N3 semantics of other languages for Web documents such as GRDDL \[[Hazaël-Massieux and Connolly (2005)](#bib.bib19)\] and RDF/A \[[Adida and Birbeck (2006)](#bib.bib2)\] may be defined so that they are also calculable.
The N3 formula of a document is not the semantics of the document in any absolute sense. It is the semantics expressed in N3. In turn, the full semantics of an N3 formula are grounded in the definitions of the properties and classes used by the formula. In the HTTP space, in which URIs are minted by an authority, definitive information about those definitions may be found by dereferencing the URIs. This information may be in natural language, in some machine-processable logic, or a mixture.
[Refer to caption](/html/0711.1533/assets/x1.png)
Figure 1: Relationship between resource, URI, string, N3 representation, and logic properties.
Other logical properties related to dereferencing include log:content, log:parsedAsN3, log:N3String, and log:uri.
- •
log:content of a document, is the string which was returned as the document contents when the URI of the resource was looked up.
- •
log:parsedAsN3 is the logical property that returns the formula got by parsing the string as a Notation3 document.
- •
log:N3String returns the string which expresses a certain formula in N3.
- •
log:uri is the URI of a resource. Normal logic processing should not look at URIs but in some cases one needs to. Though a URI itself has semantics and can be used to obtain additional information, log:uri must be used carefully because it is a level-breaker: it lets an N3 system look at its infrastructure. It is a function which one can evaluate either way: resource to URI or URI to resource.
```
<http://example.org/> log:uri "http://example.org/"
```
Figure [1](#F1 "Figure 1 ‣ 5.3 Dereferencing URIs ‣ 5 Informal Semantics of N3Logic ‣ N3Logic: A Logical Framework For the World Wide Web") illustrates the relationship between URI, resource, string, N3 representation, and the different logical properties.
### 5.4 Rules
N3 extends RDF with variables and nested graphs to enable the descriptions of rules. In its blank nodes (items in the graph not directly identified by a URI) an RDF graph has a form of existential variable. Extending the language to allow variables existentially or universally quantified over a graph allows N3 to be used for a form of logic. The drive for this initially was so that, given variables, a rule is just a relation between two graphs. Variables are defined such that when substitution occurs in a graph, it also occurs in any nested graph.
The log:implies property expresses a rule, its subject being the antecedent graph, and the object being the consequent graph. It relates two formulae, expressing implication. A statement using log:implies cannot be calculated. It is not a built-in function, but a predicate which allows the expression of a rule. N3 allows blank nodes in the conclusion of a rule, hence allowing the creation of new objects.
Continuing with our earlier example, though anyone can specify a list of people who cannot register, the conference management system only trusts administrators of the system with this information.
```
@forAll A, X.
{ A a conf:Administrator.
A conf:says { X policy:notpermitted conf:Register }.
} => { X policy:notpermitted conf:Register }.
```
To do some inference within another set of rules, a useful relationship is log:conclusion. It is a property between a formula, and the result of running any rules in the formula on all the data recursively - the deductive closure. To make up the initial formula, log:conjunction can be used to merge a list of formulae.
For example, if the statements made by administrators is in file $`<`$statements.n3$`>`$, the list of administrators in file $`<`$admin.rdf$`>`$, and the rule described about in $`<`$rule.n3$`>`$, then we can use log:conjunction and log:conclusion to decide whether or not someone should be permitted to register. log:supports is a relationship that combines log:conclusion and log:includes to check whether the conclusion of the formula includes a certain subgraph. (Please refer to section [5.5](#S5.SS5 "5.5 Graph Functions ‣ 5 Informal Semantics of N3Logic ‣ N3Logic: A Logical Framework For the World Wide Web") for more information about log:includes).
```
@forAll S, A, R, C, X.
{ <statements.n3> log:semantics S.
<admin.rdf> log:semantics A.
<rule.n3> log:semantics R.
(S A R) log:conjunction C.
C log:supports { X policy:notpermitted conf:Register }.
} => { X policy:notpermitted conf:Register }.
```
As it is possible to have blank nodes in the conclusion of a rule, log:conclusion and log:supports are undecidable and may run forever. However, it is possible to restrict N3Logic to a decidable subset language in which blank nodes are not allowed in the conclusion.
### 5.5 Graph Functions
The logic function, log:includes, checks whether one formula can be N3-derived from another formula. (Please refer to Section [5.7](#S5.SS7 "5.7 N3 Derivation ‣ 5 Informal Semantics of N3Logic ‣ N3Logic: A Logical Framework For the World Wide Web") for more information about N3-derivation) Together, log:semantics and log:includes allow rules to access the Web, and to objectively check the contents of documents, without having to load them and believe everything they say.
The following rule states that if the home page of a registrant says that she is a vegetarian, then she/he is a vegetarian. We find the URI of her homepage on her foaf page.
```
@forAll X, FOAF, F, H, HS.
{ ExConf conf:registrant X.
FOAF foaf:maker X.
FOAF log:semantics F.
F log:includes { X foaf:homepage H }.
H log:semantics HS.
HS log:includes { X a ex:Vegetarian }
} => { X a ex:Vegetarian}.
```
Whereas some datasets (such as a list of members of a club) are definitively complete, others (such as a set of temperature measurements) are not. This aspect of the Semantic Web makes negation as failure meaningless unless it is associated to a specific dataset. As a formula is of finite size, it can be tested for what it does not say. As a form of negation, log:notIncludes is completely monotonic. It can be evaluated by a mathematical calculation on the value of the two terms: no other knowledge gained can influence the result. This is the scoped negation as failure mentioned in the introduction. This is not a non-monotonic negation as failure.
Figure [2](#F2 "Figure 2 ‣ 5.5 Graph Functions ‣ 5 Informal Semantics of N3Logic ‣ N3Logic: A Logical Framework For the World Wide Web") shows how log:includes and log:notIncludes relate to quoting.
[Refer to caption](/html/0711.1533/assets/x2.png)
Figure 2: Quoting and its relation to log:includes and log:notIncludes
We continue with our example. We assume that every school website has a property linking to its directory. The directory provides information about people including their foaf pages, their designation (such as student, faculty, associate), and their email address. Below we have a rule that states if the directory of the school does not specify that the person under consideration is a student, the system gives the person the academic rate but not the student rate.
```
@forAll X, FOAF, F, H, HS, D, DS.
{ ExConf conf:registrant X.
FOAF foaf:maker X.
FOAF log:semantics F.
F log:includes { X foaf:schoolHomepage H }.
H log:semantics HS.
HS log:includes { H school:directory D }.
D log:semantics DS.
DS log:notIncludes { X a school:Student }
} => { X conf:registrationRate conf:AcademicRate }.
```
The effect of a default with an explicit domain is achieved with log:notIncludes. Defaults can be handled by first running rules to work out everything that is specified, and then doing a log:notIncludes on the result as shown in the example above.
### 5.6 Built-ins
N3Logic also provides other built-ins for additional functionality. Some examples include
- •
crypto functions - md5, sign, and verify
- •
xmath functions - cos, greaterThan, notGreaterThan, and difference
- •
os functions for retrieving environment information - argv, baseAbsolute, and baseRelative
- •
string functions - concatenation, matches, and startsWith
- •
time functions - dayOfWeek, gmTime, and localTime
The following example describes a rule that states that papers submitted to the conference that are not more than 6 pages are valid if authorized by a program chair of the conference.
```
@prefix math: <http://www.w3.org/2000/10/swap/math#> .
@forAll PAPER, LEN, CHAIR.
{ ExConf conf:submittedPaper PAPER.
PAPER conf:pageLength LEN.
LEN math:notGreaterThan 6.
PAPER conf:authorized CHAIR.
CHAIR conf:chair ExConf.
} => { PAPER a conf:ValidPaper }.
```
### 5.7 N3 Derivation
N3-derivation is not aimed at moving all the way to the powerful inference capabilities of an expressive logic, but rather extends textual inclusion to include some simple inferences that are standard in most logics. These include:
- •
Conjunction Elimination (CE): For any formulae A and B, given conjunction A and B, then A follows, and B follows.
- •
Conjunction Introduction (CI): For any formulae A and B, given A, and given B, then the conjunction A and B follows.
- •
Universal Elimination (UE): Given any formula A, a universal variable x that is used in A, and a ground term t, A$`{{}_{t}^{}/}x`$ follows. i.e. @forAll x, A and ground term t, then A$`{{}_{t}^{}/}x`$ follows.
- •
Existential Introduction (EI): Given any formula A containing a ground term t, and an existential variable v that does not occur in A, then A$`{{}_{x}^{}/}t`$ follows.
- •
Variable Renaming (VR): For any formula A, and variables x and x’, A = A’ where A’ is a formula derived by Subst(x/x’, A).
N3 derivation is any finite number of applications of of CE, CI, UE, EI or VR. An N3 derivation operator T is defined as any operator which is the successive application of any sequence (possibly empty) of such operators. A formula F n3-derives a formula T F, which implies that by a combination of CE, EI, UE and VR, T F logically follows from F.
As RDF Graph is a subclass of N3 formula, if F and G are RDF graphs, only CE and EI apply and n3-derivation reduces to simple entailment from RDF Semantics.
The implementation of built-in functions is not in general required for any implementation of N3Logic, as they can always soundly be treated as ground facts. However, their usefulness is derived from their implementation. For example, { 1 math:negation -1 } is derived by calculation. Like other RDF properties, the set is designed to be extensible, enabling others to use URIs for new functions. When a triple can be evaluated, or a variable bound because its predicate is a built-in function, then the derivation of the statement is said to be by calculated derivation. N3Logic-derivation is N3-derivation with modus ponens and calculated derivation.
### 5.8 Symmetry of Triples
When designing an ontology in RDF the direction chosen for a given property is arbitrary - one can either define “parent” or “child”, “employee” or “employer”. Our philosophy (from the Enquire design of 1980 \[[Berners-Lee (1980)](#bib.bib7)\]) is that one should not favor one way over another. On the other hand, one should not encourage people to declare both a property and its inverse. Therefore, a design choice in N3 is to allow forward and backward links to be expressed with equal ease. It does this by providing keywords “is” and “of” that allow one to reverse the direction of the description of a triple. This also enables the serialization of any acyclic graphs with blank nodes without requiring them to have generated node identifiers.
### 5.9 Extensibility
The extensibility of RDF is deliberate so that a document may draw on predicates from many sources. The statement {s p o} expresses that the relationship denoted by p holds between the things denoted by s and o. The meaning of the statement {s p o} in general is defined by any specification for p. The Architecture of the Web specifies informally how information about the relation can be discovered \[[Jacobs and Walsh (2004)](#bib.bib20)\]. In a similar fashion, N3Logic allows external predicates - predicates not defined within N3Logic - to be used. The definitions of these external predicates can be discovered by looking up their URI on the Web and used as long as their semantics are defined in N3Logic. Clearly a system which includes further logical predicates, beyond those defined in N3Logic, whose meaning introduces greater logical expressiveness would change the properties of the logic.
By having rules and data in the same languages, N3Logic provides simplicity in syntax and completeness as rules can operate on themselves and anything written in N3 can be queried in N3. A rule engine, when analyzing a rule prior to running it, can treat specially those properties it knows as calculable functions which occur in the antecedent. This allows N3 to be used to develop specific languages such as query languages. For example, we can create a language for expressing graph differences and updates by simply defining two new properties, diff:insertion and diff:deletion \[[Berners-Lee and Connolly (2004)](#bib.bib12)\]. These properties provide a way to uniquely identify what is changing and a way to distinguish between the pieces added and those subtracted.
In the following example, everyone who is paying a student rate, will also be given accommodation in the dormitory but will not be given a ticket to attend the social event of the conference.
```
@forAll X.
{ X conf:registrationRate conf:StudentRate }
diff:insertion { X conf:accommodation conf:Dormitory };
diff:deletion { X conf:ticket conf:SocialEvent } .
```
In many languages similar to N3, there is a risk of ambiguity as to whether a naked alphanumeric string is a keyword or an identifier. Serious version management problems occur when new keywords are added to a language, changing things which were identifiers into keywords. N3 is designed to be extended in the future. For this reason, an N3 document can declare which keywords it uses without the “@” sign. This allows N3 to be extended without the danger that existing documents be incorrectly interpreted by future systems, or future documents by existing systems.
## 6 N3Logic Example: Access control policy
We extend our conference management system example with a policy for controlling access to pictures that were taken during the conference. This policy states that only people who registered for the conference can view pictures taken at the conference. During the registration process, the system records registrant’s foaf pages. In order to prove that they have registered for the conference, users must be able to prove ownership of a registered foaf page. This can be done either using a decentralized authentication mechanism such as Open ID \[[(16)](#bib.bib16)\] or using a mechanism by which users must present a secret key whose hash is on their foaf page. In this policy, we use the latter. A user request consists of her secret key and the URI of the picture being requested. If the picture is one of the pictures taken at the conference and the secret key is the digest of the session:hexdigest value on a registered foaf page, then the request is considered valid and the picture is returned to the user. For the entire working example, please refer to http://dig.csail.mit.edu/2006/Papers/TPLP/example/.
```
@forAll REQ, PHOTO, WHO, FOAF, X, TXT, CONF, C.
{ REQ a rein:Request.
REQ rein:resource PHOTO.
<http://dig.csail.mit.edu/2006/Papers/TPLP/example/exconf.n3> log:semantics C.
C log:includes
{ CONF a conf:Conference. PHOTO a conf:GroupPicture; conf:taken CONF }.
REQ rein:requester WHO.
WHO session:secret ?S.
?S crypto:md5 TXT.
C log:includes
{ CONF conf:registrant X. FOAF foaf:maker X }.
FOAF log:semantics [ log:includes
{ FOAF foaf:maker [ session:hexdigest TXT ] }
].
} => { REQ rein:requester [ policy:permitted-to-view PHOTO ]. REQ a ValidRequest }.
```
## 7 Implementations
We have developed cwm \[[Berners-Lee (2000)](#bib.bib9)\] a forward-chained reasoner in python \[[(1)](#bib.bib1)\] for N3 and N3Logic. It is a general-purpose reasoner for the Semantic Web that can be used for querying, checking, transforming and filtering information. Currently, cwm parses RDF/XML, and N3 and its subsets. A number of tools, for example SWOOP \[[Kalyanpur et al. (2005)](#bib.bib22)\], support Turtle \[[Beckett (2006)](#bib.bib6)\], a fragment of N3 that is equivalent to RDF/XML.
Being based on a more expressive logic language adds a host of features to cwm not available to other RDF processing tools: accessing Web resources, and filtering RDF graphs after merging them, for example. Since N3Logic is expressive enough so that positive datalog-like rules can be expressed in it, cwm is able to reason using a first order logic but without classical negation. Combining this reasoning functionality with its ability to retrieve documents form the Web as needed, the system can be considered a reasoner for the Web. It has grown from a proof of concept application to a popular rule engine, used in major research projects such as Policy Aware Web \[[Kolovski et al. (2005)](#bib.bib23), [Kagal et al. (2006)](#bib.bib21)\], and the Technical Report Automation project at W3C (http://www.w3.org/TR/).
Euler is an inference engine supporting logic based proofs. Unlike cwm, it is a backward-chaining reasoner enhanced with Euler path detection \[[Roo (2005)](#bib.bib25)\].
Pychinko is a Python implementation of the classic Rete pattern matching algorithm \[[Parsia et al. (2005)](#bib.bib24)\]. Rete has shown to be, in many cases, the most efficient way to apply rules to a set of facts–the basic functionality of an expert system. Pychinko employs an optimized implementation of the algorithm to handle facts, expressed as triples, and process them using a set of N3 rules. Pychinko tries to closely mimic the features available in Cwm, as it is one of the most widely used rule engines in the RDF community. Pychinko has proven to be faster than Cwm, however it’s limitation lies in its expressivity: Pychinko cannot handle most of the cwm built-ins. It is worth mentioning here that the RETE engine used in Pychinko has been ported to Cwm - thus Cwm can now boast the same performance improvements.
## 8 Related Work
Several logics related to N3Logic exist including OWL, Simple Common Logic (SCL) \[[Altheim et al. (2005)](#bib.bib3)\], and Knowledge Interface Format (KIF) \[[Genesereth (1998)](#bib.bib18)\]. OWL is built on top of RDFS and provides a vocabulary for describing the characteristics of properties and classes, the relationships between classes, and relationships between properties. OWL is based on Description Logic \[[Baader et al. (2003)](#bib.bib5)\], which is a subset of First Order Logic (FOL) \[[Shapiro (2005)](#bib.bib26)\]. OWL provides limited expressivity for a Web-like environment as it does not support quantified variables, rules, or a mechanism to distinguish which document or person asserts what. KIF is a framework for exchanging of declarative knowledge among heterogeneous computer systems. It is a version of first order predicate calculus with extensions to support non-monotonic reasoning and quoting. The key differences between KIF and N3Logic are that KIF does not include operators for Web access and it supports non-monotonic reasoning. SCL is aimed at providing a standard logical interchange language based on XML. It has a higher-order syntax that provides integration between different representation languages but SCL gives this syntax a completely first-order semantics. This syntax can be used to provide quoting functionality. Proof Carrying Authorization (PCA) proposes that the underlying framework of a distributed authorization system be a higher-order logic and that different domains in this system use different application-specific logics that are subsets of the higher-order logic \[[Appel and Felten (1999)](#bib.bib4)\]. They also propose that clients develop proofs of access using these application specific logics and send them to servers to validate. N3Logic draws inspiration from PCA but modifies it to leverage the distributed nature and linkability of the Web.
A formal categorization of N3Logic is complicated as it differs from most traditional logics in expressivity. It is clearly more expressive than Datalog \[[Gallaire and Minker (1978)](#bib.bib17)\] but less expressive than traditional FOL. Much Semantic Web work uses DL expressivity, and like DL, N3Logic is a subset of FOL, although the quoting mechanisms provide higher order features (which we believe are actually limited to FOL in the same manner as SCL). However, unlike DL, N3Logic is not decidable, limiting expressivity in other ways motivated by the Web considerations discuss in this paper. As such, developing a formal model theory for N3Logic is quite challenging, and is the focus of current work.
## 9 Conclusion
The main goal of N3Logic is to extend the RDF data model, so that the same language can be used for logic and data. N3Logic uses the N3 syntax, which provides quoting, variables, and the implication operator. N3Logic also includes built-in functions that allow rules to access Web resources, define which inference can be drawn from specific Web documents, and other useful functionality such as mathematic, cryptographic, and string. In this paper, we described the N3 syntax and give the informal semantics of N3Logic.
The use of log:notIncludes to allow default reasoning without non-monotonic behavior achieves a design goal for distributed rule systems. The N3Logic language has been found to have some useful practical properties. The separation between the N3 extensions to RDF and the logic properties has allowed N3Logic to be extended with other properties to provide functionality such as the expression of graph differences and updates \[[Berners-Lee and Connolly (2004)](#bib.bib12)\].
## 10 Acknowledgements
We are grateful to Vladimir Kolovski, Sean Palmer, Dave Beckett, and Jos de Roo for feedback on the N3 language resulting from their implementations; to the Data Access Working Group for feedback resulting from their adoption of N3 syntax for part of SPARQL grammar; to the RDF working group for their co-operation in keeping NTriples a subset of N3; and to many in the W3C Semantic Web Interest Group for helpful advice and suggestions. This work is supported in part by funding from US Defense Advanced Research Projects Agency (DARPA) and Air Force Research Laboratory, Air Force Materiel Command, USAF, under agreement number F30602-00-2-0593, Semantic Web Development. This work was also funded under NSF ITR 04-012.
## References
- \(1\) The Official Python Programming Language Website.
http://www.python.org/.
- Adida and Birbeck (2006)
Adida, B. and Birbeck, M. 2006.
RDF/A Primer 1.0. Embedding RDF in XHTML. W3C Working Draft.
http://www.w3.org/TR/xhtml-rdfa-primer/.
- Altheim et al. (2005)
Altheim, M., Anderson, B., Hayes, P., Menzel, C., Sowa, J. F., and Tammet, T. 2005.
SCL: Simple Common Logic.
http://www.ihmc.us/users/phayes/CL/SCL2004.html.
- Appel and Felten (1999)
Appel, A. W. and Felten, E. W. 1999.
Proof-Carrying Authentication.
In 6th ACM Conference on Computer and Communications Security.
- Baader et al. (2003)
Baader, F., Calvanese, D., McGuinness, D., Nardi, D., and Patel-Schneider, P., Eds. 2003.
The Description Logic Handbook.
Cambridge University Press.
- Beckett (2006)
Beckett, D. 2006.
Turtle - Terse RDF Triple Language.
http://www.dajobe.org/2004/01/turtle/.
- Berners-Lee (1980)
Berners-Lee, T. 1980.
ENQUIRE V 1.1 Manual.
http://www.w3.org/History/1980/Enquire/.
- Berners-Lee (1998)
Berners-Lee, T. 1998.
Notation 3 (N3) A readable RDF Syntax.
http://www.w3.org/DesignIssues/Notation3.html.
- Berners-Lee (2000)
Berners-Lee, T. 2000.
Cwm : General-purpose Data Processor for the Semantic Web.
http://www.w3.org/2000/10/swap/doc/cwm.
- Berners-Lee (2005)
Berners-Lee, T. 2005.
Primer: Getting into RDF and Semantic Web using N3.
http://www.w3.org/2000/10/swap/Primer.
- Berners-Lee (2006)
Berners-Lee, T. 2006.
N3 BNF.
http://www.w3.org/2000/10/swap/grammar/n3-report.html.
- Berners-Lee and Connolly (2004)
Berners-Lee, T. and Connolly, D. 2004.
Delta: an ontology for the distribution of differences between RDF graphs.
http://www.w3.org/DesignIssues/Diff.
- Brickley and Miller (1999)
Brickley, D. and Miller, L. 1999.
Friend of a friend (foaf) project.
http://www.foaf-project.org/.
- Carroll and Klyne (2004)
Carroll, J. J. and Klyne, G. 2004.
Resource Description Framework (RDF): Concepts and Abstract Syntax . Tech. Rep. http://www.w3.org/TR/2004/REC-rdf-concepts-20040210/, W3C. February.
- Dumbill (2002)
Dumbill, E. 2002.
Finding friends with XML and RDF.
http://www-128.ibm.com/developerworks/xml/library/x-foaf.html.
- \(16\)
Foundation, O.
OpenID.
http://openid.net/.
- Gallaire and Minker (1978)
Gallaire, H. and Minker, J., Eds. 1978.
Logic and Data Bases.
Perseus Publishing.
- Genesereth (1998)
Genesereth, M. R. 1998.
Knowledge Interchange Format. Draft proposed American National Standard (dpANS). NCITS.T2/98-004.
http://logic.stanford.edu/kif/dpans.html.
- Hazaël-Massieux and Connolly (2005)
Hazaël-Massieux, D. and Connolly, D. 2005.
Gleaning Resource Descriptions from Dialects of Languages (GRDDL), W3C Team Submission.
http://www.w3.org/TeamSubmission/grddl/.
- Jacobs and Walsh (2004)
Jacobs, I. and Walsh, N. 2004.
The Architecture of the World Wide Web. W3C Recommendation.
http://www.w3.org/TR/webarch/.
- Kagal et al. (2006)
Kagal, L., Berners-Lee, T., Connolly, D., and Weitzner, D. 2006.
Using Semantic Web Technologies for Open Policy Management on the Web.
In 21st National Conference on Artificial Intelligence.
- Kalyanpur et al. (2005)
Kalyanpur, A., Parsia, B., Sirin, E., Cuenca-Grau, B., and Hendler, J. 2005.
Swoop: A ’Web’ Ontology Editing Browser.
Journal of Web Semantics Vol 4(2).
- Kolovski et al. (2005)
Kolovski, V., Katz, Y., Hendler, J., Weitzner, D., and Berners-Lee, T. 2005.
Towards a Policy-Aware Web.
In Semantic Web and Policy Workshop at the 4th International Semantic Web Conference.
- Parsia et al. (2005)
Parsia, B., Katz, Y., and Clark, K. 2005.
Pychinko: Rete-based RDF friendly rule engine.
http://www.mindswap.org/~katz/pychinko/.
- Roo (2005)
Roo, J. D. 2005.
Euler proof mechanism.
http://www.agfa.com/w3c/euler/.
- Shapiro (2005)
Shapiro, S. Fall 2005.
Classical Logic.
In The Stanford Encyclopedia of Philosophy, E. N. Zalta, Ed.
- \(27\)
Steele, G. L. and Sussman, G. J.
Scheme.
http://www.swiss.ai.mit.edu/projects/scheme/.
## 11 APPENDIX : N3 Concrete Syntax
N3 provides a human-readable syntax for RDF and is a language that uses conventional unix-style punctuation, which is both more easily writable and readable than the RDF/XML syntax \[[Berners-Lee (2006)](#bib.bib11)\]. It provides quantified variables and allows quoting so that statements about statements can be made.
- •
It provides URI abbreviation using prefixes which are bound to a namespace.
```
@prefix j: <http://example.org/joe-foaf#> .
@prefix rdf: <http://www.w3.org/1999/02/22-rdf-syntax-ns#> .
@prefix rdfs: <http://www.w3.org/2000/01/rdf-schema#> .
@prefix conf: <http://example.org/conf#> .
@prefix : <http://dig.csail.mit.edu/2006/Papers/TPLP/example/exconf#> .
```
The prefixes used in the above example are assumed throughout the paper.
- •
Qualified names using the default namespace have an empty prefix and start with a colon.
```
@prefix : <http://dig.csail.mit.edu/2006/Papers/TPLP/example/exconf#> .
ExConf a conf:Conference .
```
- •
An N3 formula is comprised of a conjunctive set of triples.
- •
Triples are formed of terms, which are symbols (URIs), strings, blank nodes, numeric literals, lists or quoted formulae.
- •
A triple may be terminated by a period.
```
ExConf conf:eventName "WWW2006 Workshop on Models of Trust for the Web" .
```
- •
The repetition of another object for the same subject and predicate is possible using a comma “,”
```
ExConf conf:cochair <http://csail.mit.edu/~lkagal/foaf#me>,
<http://umbc.edu/~finin/foaf#tim>.
```
- •
The repetition of another predicate for the same subject is done using a semicolon “;”
```
ExConf conf:eventName "WWW2006 Workshop on Models of Trust for the Web",
conf:acronym "MTW06" .
```
- •
Blank nodes with certain properties can be defined by just putting the properties between “\[“ and “\]”. They can be used in two ways: \[ \] together followed by the properties or \[ \] around the properties. The following example describes the same subject in these two ways.
```
[ ] conf:homepage <http://www.l3s.de/~olmedilla/events/MTW06_Workshop.html> .
[ conf:homepage <http://www.l3s.de/~olmedilla/events/MTW06_Workshop.html> ] .
```
- •
N3 allows has a special \_: namespace prefix. An identifier of such a form (e.g. \_:xyz) represents an blank node.
- •
An empty URI reference, which refers to the document it is written in, can be written using “$`<`$$`>`$”. The example document declares itself to be a Conference as defined in the conf namespace.
```
<> a conf:Conference .
```
- •
Quoted formulae allow N3 formulae to be quoted within N3 formulae using “{“ and “}”. In a quoted N3 formula, s p o, both s and o can be RDF graphs or N3 formulae.
```
j:Joe says
{ ex:ExConf conf:homepage <http://www.l3s.de/~olmedilla/events/MTW06_Workshop.html;
conf:eventName "WWW2006 Workshop on Models of Trust for the Web";
conf:acronym "MTW06" } .
```
- •
N3 formula can have both existential and universal quantifiers. Existential variables can be indicated by an @forSome declaration and universal variables can be indicated by @forAll declarations. ?X is a shorthand notation and implies universal quantification in the enclosing parent of the current formula. It can be used without an explicit @forAll declaration. The scope of the @forAll, which is used to define universal variables, is outside the scope of any @forSome, which is used to define existential variables. If both universal and existential quantification are specified for the same context, then the scope of the universal quantification is outside the scope of the existentials.
- •
Keywords are a very limited set of alphanumeric strings which are in the grammar prefixed by an “@” sign. If no @keywords directive is given, all qualified names need a “:” before them and all keywords except “a”, “is”, and “of” require an “@”. If the @keywords directive is given, then the given set of bare strings (without either “:” or “@” before them) are keywords and the others are qualified names in the default namespace.
- •
The keywords “a”, “is”, and “of” can be used without an “@” even if the @keywords directive is not given. The keyword “a” maps to rdf:type whereas “of” and “is” provide syntactic sugar to describe triples in the reverse direction such as { object is predicate of subject}.
- •
The keywords @true and @false are boolean literals.
- •
Strings are defined within double quotes “ “ such as “Joe Smith” and within triple double quotes “ “ “ for multi-line values or values containing quote marks.
- •
Numerical literals such as integers, floats, and decimals are also supported.
- •
Comments are identified with “#”. Everything that follows the “#” is ignored till the end of the line.
- •
The shorthand =$`>`$ may be used for the implies property defined in the log: namespace (http://www.w3.org/2000/10/swap/log#). (Please refer to Section [5.4](#S5.SS4 "5.4 Rules ‣ 5 Informal Semantics of N3Logic ‣ N3Logic: A Logical Framework For the World Wide Web") for more information).
- •
“=” is a shorthand notation for the sameAs property defined in owl: namespace (http://www.w3.org/2002/07/owl#). (Please refer to Section [5.1](#S5.SS1 "5.1 Relationship with RDF, RDFS, and OWL ‣ 5 Informal Semantics of N3Logic ‣ N3Logic: A Logical Framework For the World Wide Web") for more information).
- •
N3 supports RDF collections and uses them frequently as ordered containers, as argument lists to N-ary functions such as crypto:sign in which the subject is a list of two things, a hash string and a key (containing private and public parts) and the object is calculated as a signature string by signing the hash with the key’s private part.
For example, to describe Joe’s interests we would use a list as follows
```
j:Joe interests ( "AI" "Semantic Web" "Logic" ) .
``` |
0710.3955 | cs.NI | cyber | # On the Behavior of the Distributed Coordination Function of IEEE 802.11 with Multirate Capability under General Transmission Conditions
F. Daneshgaran, M. Laddomada, F. Mesiti, and M. Mondin
This work has been supported by Euroconcepts, S.r.l. (http://www.euroconcepts.it)F. Daneshgaran is with ECE Dept., California State University, Los Angeles, USA.M. Laddomada, F. Mesiti, and M. Mondin are with DELEN, Politecnico di Torino, Italy.
###### Abstract
The aim of this paper is threefold. First, it presents a multi-dimensional Markovian state transition model characterizing the behavior of the IEEE 802.11 protocol at the Medium Access Control layer which accounts for packet transmission failures due to channel errors modeling both saturated and non-saturated traffic conditions. Second, it provides a throughput analysis of the IEEE 802.11 protocol at the data link layer in both saturated and non-saturated traffic conditions taking into account the impact of both the physical propagation channel and multirate transmission in Rayleigh fading environment. The general traffic model assumed is M/M/1/K. Finally, it shows that the behavior of the throughput in non-saturated traffic conditions is a linear combination of two system parameters; the payload size and the packet rates, $`\lambda^{(s)}`$, of each contending station. The validity interval of the proposed model is also derived.
Simulation results closely match the theoretical derivations, confirming the effectiveness of the proposed models.
###### Index Terms:
DCF, Distributed Coordination Function, fading, IEEE 802.11, MAC, Rayleigh fading, rate adaptation, saturation, throughput, unsaturated, non-saturated.
## I Introduction
The IEEE802.11 Medium Access Control (MAC) layer \[[1](#bib.bib1)\] presents a mandatory option, namely the Distributed Coordination Function (DCF), based on the Carrier Sense Multiple Access Collision Avoidance CSMA/CA access method, that has received considerable attention in the past years \[[3](#bib.bib3)\]-\[[18](#bib.bib18)\].
Many papers, following the seminal work by Bianchi \[[3](#bib.bib3)\], have addressed the problem of modeling the DCF in a variety of traffic load and channel transmission conditions. Most of them focus on a scenario presenting $`N`$ saturated stations that transmit towards a common Access Point (AP) under the hypotheses that the packet rates, along with the probability of transmission in a randomly chosen time slot, is common to all the involved stations, while the error events on the transmitted packets are mainly due to collisions between packets belonging to different stations.
Real networks are different in many respects. First, traffic is mostly non-saturated, so it is important to derive a model accounting for practical network operations. Channel conditions are far from being ideal and often packet transmission has to be rescheduled until data is correctly received. Due to Rayleigh and shadow fading conditions, a real scenario presents stations transmitting at different bit rates, because of multirate adaptation foreseen at the physical layer of WLAN protocols such as IEEE 802.11b. Furthermore, other scenarios usually occur in a real network. As an example, different stations usually operate with different loads, i.e., they have different packet rates, while the transmitting bit rate can also differ between the contending stations. In all these situations the common hypothesis, widely employed in the literature, that all the contending stations have the same probability of transmitting in a randomly chosen time slot, does not hold anymore. The aim of this paper is to provide a model and theoretical analysis under much more realistic scenarios. With this background, let us provide a quick survey of the recent literature related to the problem addressed in this paper.
In \[[3](#bib.bib3)\], the author provides an analysis of the saturation throughput of the basic 802.11 protocol assuming a two dimensional Markov model at the MAC layer. Papers \[[4](#bib.bib4)\]-\[[6](#bib.bib6)\] model the influence of real channel conditions on the throughput of the DCF operating in saturated traffic conditions, while \[[7](#bib.bib7)\]-\[[9](#bib.bib9)\] thoroughly analyze the influence of capture on the throughput of wireless transmission systems.
The behavior of the DCF of IEEE 802.11 WLANs in unsaturated traffic conditions has been analyzed in a number of papers \[[10](#bib.bib10)\]-\[[12](#bib.bib12)\]. In \[[10](#bib.bib10)\] the authors extend the Bianchi’s model by introducing a new state in order to consider non-saturated traffic conditions, accounting for the case in which the station queue is empty after successful completion of a packet transmission. In the modified model however, a packet is discarded after $`m`$ backoff stages, while in Bianchi’s model the station keeps iterating in the $`m`$-th backoff stage until the packet gets successfully transmitted. These models rely on the basic hypotheses of ideal channel transmission and on the absence of capture effects. Paper \[[11](#bib.bib11)\] proposes to model non-saturated traffic conditions by adding a new state for each backoff stage accounting for the absence of new packets to be transmitted. In \[[12](#bib.bib12)\]-\[[13](#bib.bib13)\], the authors extend the multi-dimensional Markovian state transition model characterizing the MAC layer behavior by including transmission states that account for packet transmission failures due to errors caused by propagation through the channel, along with a state characterizing the situation when a station has no packets to transmit.
In \[[14](#bib.bib14)\], the authors look at the impact of channel induced errors and of the received SNR on the achievable throughput in a system with rate adaptation, whereby the transmission rate of the terminal is modified depending on either direct or indirect measurements of the link quality. In \[[15](#bib.bib15)\], the authors were the first to observe that in multirate networks the aggregate throughput is strongly influenced by that of the slowest contending station; such a phenomenon is denoted as “performance anomaly of the DCF of IEEE 802.11 protocol”. In \[[16](#bib.bib16)\], authors provide an analytical framework for analyzing the link delay of multirate networks. In \[[18](#bib.bib18)\], authors provide DCF models for finite load sources with multirate capabilities, while in \[[17](#bib.bib17)\] authors propose a DCF model for networks with multirate stations and derive the saturation throughput. Remedies to performance anomalies are also discussed. In both previous works, packet errors are only due to collisions between different contending stations.
In this paper, we substantially extend the previous work proposed in the companion papers \[[12](#bib.bib12)\]-\[[13](#bib.bib13)\] by looking at all the three issues outlined before together, namely, real channel conditions, saturated and non-saturated traffic, and multirate capabilities. Our basic assumptions are essentially similar to those of Bianchi \[[3](#bib.bib3)\], with the exception that we do assume the presence of both channel induced errors due to the transmission over Rayleigh fading channel, and we consider a general traffic model. As a reference standard, we use network parameters belonging to the IEEE802.11b protocol, even though the proposed mathematical model holds for any flavor of the IEEE802.11 family or other wireless protocols with similar MAC layer functionality. We also derive a simple model of the aggregate throughput in unloaded traffic conditions and show that essentially, the throughput depends on the linear combination of the packet rates $`\lambda^{(s)}`$ of each contending station, each properly weighted by its payload size, while packet errors due to imperfect channel conditions do not affect the aggregate throughput. The validity interval of the proposed model is also derived.
The paper is organized as follows. After a brief review of the functionalities of the contention window procedure at MAC layer, Section [II](#S2 "II Markovian Model Characterizing the MAC Layer under general traffic conditions and Real Transmission Channel ‣ On the Behavior of the Distributed Coordination Function of IEEE 802.11 with Multirate Capability under General Transmission Conditions") substantially extends the Markov model initially proposed by Bianchi, presenting modifications that account for transmission errors over Rayleigh fading channels employing the 2-way handshaking technique in a variety of traffic conditions. Section [III](#S3 "III Markovian Process Analysis ‣ On the Behavior of the Distributed Coordination Function of IEEE 802.11 with Multirate Capability under General Transmission Conditions") is devoted to the solution of the proposed bi-dimensional Markov chain related to each contending station, and provides an expression for the aggregate throughput of the link. After defining the basic time parameters needed for system performance evaluation, Section [IV](#S4 "IV Estimating the Average Time Slot Duration ‣ On the Behavior of the Distributed Coordination Function of IEEE 802.11 with Multirate Capability under General Transmission Conditions") estimates the expected time slot duration needed for the evaluation of the aggregate throughput of the link. The adopted traffic model is discussed in Section [V](#S5 "V Traffic Model ‣ On the Behavior of the Distributed Coordination Function of IEEE 802.11 with Multirate Capability under General Transmission Conditions").
Section [VI](#S6 "VI Physical Layer Modeling ‣ On the Behavior of the Distributed Coordination Function of IEEE 802.11 with Multirate Capability under General Transmission Conditions") briefly addresses the modeling of the physical layer of IEEE 802.11b in a variety of channel conditions. In Section [VII](#S7 "VII Simulation Results and Model Validation ‣ On the Behavior of the Distributed Coordination Function of IEEE 802.11 with Multirate Capability under General Transmission Conditions") we present simulation results where typical MAC layer parameters for IEEE802.11b are used to obtain throughput values as a function of various system level parameters and SNR, under typical traffic conditions. Section [VIII](#S8 "VIII A simple model of the aggregate throughput in non-saturated conditions ‣ On the Behavior of the Distributed Coordination Function of IEEE 802.11 with Multirate Capability under General Transmission Conditions") derives a simple model of the aggregate throughput in unsaturated traffic conditions, along with its interval of validity. Insights on the DCF behavior with multiple contending stations are also given, considering a multirate setting. Finally, Section [IX](#S9 "IX Conclusions ‣ On the Behavior of the Distributed Coordination Function of IEEE 802.11 with Multirate Capability under General Transmission Conditions") is devoted to conclusions.
## II Markovian Model Characterizing the MAC Layer under general traffic conditions and Real Transmission Channel
In a previous paper \[[12](#bib.bib12)\], we proposed a bi-dimensional Markov model for characterizing the behavior of the DCF under a variety of real traffic conditions, both non-saturated and saturated, with packet queues of small sizes, and considered the IEEE 802.11b protocol with the basic 2-way handshaking mechanism. Many of the basic hypotheses were the same as those adopted by Bianchi in the seminal paper \[[3](#bib.bib3)\].
As a starting point for the derivations which follow, we adopt the bi-dimensional model proposed in the companion paper \[[12](#bib.bib12)\], appropriately modified in order to account for a scenario of $`N`$ contending stations, each one employing a specific bit rate and a different transmission packet rate. As in \[[12](#bib.bib12)\], we consider the IEEE 802.11b protocol employing the basic 2-way handshaking mechanism. For conciseness, we refer the interested to \[[12](#bib.bib12)\] for many details on the considered bi-dimensional Markov model.
Consider the following scenario: $`N`$ stations transmit towards a common AP in infrastructure mode, whereby each station, characterized by its own traffic load, can access the channel using a data rate in the set $`\{ 1,2,5.5,11\}`$ Mbps, depending on channel conditions. Each bit rate is associated with a different modulation format, whereas the basic rate is 1 Mbps with DBPSK modulation (2Mbps with DQPSK if short preamble is used) \[[1](#bib.bib1)\]. We identify a generic station with the index $`s \in S = {\{ 1,2,\cdots,N\}}`$, where $`N`$ is the number of stations in the network, and $`S`$ is the set of station indexes. As far as the transmission data rate is concerned, we define four rate-classes identified by a rate-class identifier $`r`$ taking values in the set $`R = {\{ 1,2,3,4\}}`$ ordered by increasing values of data rates $`R_{D} = {\{ 1,2,5.5,11\}}`$ Mbps (as an example, rate-class $`r = 3`$ is related to the bit rate 5.5 Mbps). Concerning control packets and PLCP header transmissions, the basic rate is identified by $`R_{C}`$. Two basic rates are available for control packets, namely 1 and 2 Mbps, whereby the latter is adopted for short preambles \[[1](#bib.bib1)\].
The traffic load of the s-th station is identified by a Packet Arrival Rate (PAR) $`\lambda^{(s)}`$. Upon defining both rate-classes and traffic, we can associate a generic station $`s`$ to a rate-class $`r \in R`$ and traffic load $`\lambda^{(s)}`$. As a consequence, with respect to the model in \[[12](#bib.bib12)\], we need to specify the probabilities along with different Markov chains for each station in the network.
The two sources of errors on the transmitted packets are collisions between packets and channel induced errors. Considering the $`s`$-th station, collisions can occur with probability $`P_{c\hspace{0pt}o\hspace{0pt}l}^{(s)}`$, while transmission errors due to imperfect channel transmissions can occur with probability $`P_{e}^{(s)}`$. Notice that $`P_{e}^{(s)}`$ depends upon the station rate-class $`r`$, which in turn depends on the Signal-to-Noise Ratio (SNR) received by the AP (appropriate expressions will be provided in Section [VI](#S6 "VI Physical Layer Modeling ‣ On the Behavior of the Distributed Coordination Function of IEEE 802.11 with Multirate Capability under General Transmission Conditions") for each rate-class). We assume that collisions and transmission error events are statistically independent. In this scenario, a packet from the $`s`$-th station is successfully transmitted if there is no collision (this event has probability $`1 - P_{c\hspace{0pt}o\hspace{0pt}l}^{(s)}`$) and the packet encounters no channel errors during transmission (this event has probability $`1 - P_{e}^{(s)}`$). The probability of successful transmission is therefore equal to $`{({1 - P_{e}^{(s)}})}\hspace{0pt}{({1 - P_{c\hspace{0pt}o\hspace{0pt}l}^{(s)}})}`$, while the probability of failed transmission is,
<table>
<tbody>
<tr>
<td></td>
<td>Peq(s)=Pcol(s)+Pe(s)−Pe(s)⋅Pcol(s)subscriptsuperscript𝑃𝑠𝑒𝑞subscriptsuperscript𝑃𝑠𝑐𝑜𝑙subscriptsuperscript𝑃𝑠𝑒⋅subscriptsuperscript𝑃𝑠𝑒subscriptsuperscript𝑃𝑠𝑐𝑜𝑙P^{(s)}_{eq}=P^{(s)}_{col}+P^{(s)}_{e}-P^{(s)}_{e}\cdot P^{(s)}_{col}</td>
<td></td>
<td>(1)</td>
</tr>
</tbody>
</table>
To simplify the analysis, we make the assumption that the impact of the channel induced errors on the packet headers are negligible because of their short length compared to the data payload size. This is justified on the basis of the assumption that the bit errors inflicting the transmitted data are independent of each other. Hence, the packet or frame error rate, identified respectively with the acronyms PER or FER, is a function of the packet length, with shorter packets having exponentially smaller probability of error compared to longer packets. We note that with sufficient interleaving we can always ensure that the errors affecting individual bits in a data packet are independent of each other.
[Refer to caption](/html/0710.3955/assets/x1.png)
Figure 1: Markov chain for the contention model of the generic $`s`$-th station in general traffic conditions, based on the 2-way handshaking technique, considering the effects of channel induced errors.
The modified Markov model related to the $`s`$-th contending station is depicted in Fig. [1](#S2.F1 "Figure 1 ‣ II Markovian Model Characterizing the MAC Layer under general traffic conditions and Real Transmission Channel ‣ On the Behavior of the Distributed Coordination Function of IEEE 802.11 with Multirate Capability under General Transmission Conditions"). We consider $`({m + 1})`$ different backoff stages, including the zero-th stage. The maximum Contention Window (CW) size is $`W_{m\hspace{0pt}a\hspace{0pt}x} = {2^{m}\hspace{0pt}W_{0}}`$, and the notation $`W_{i} = {2^{i}\hspace{0pt}W_{0}}`$ is used to define the $`i^{t\hspace{0pt}h}`$ contention window size ($`W_{0}`$ is the minimum contention window size). A packet transmission is attempted only in the $`(i,0)`$ states, $`{\forall i} = {0,\ldots,m}`$. In case of collision, or due to the fact that transmission is unsuccessful because of channel errors, the backoff stage is incremented, so that the new state can be $`({i + 1},k)`$ with uniform probability $`P_{e\hspace{0pt}q}^{(s)}/W_{i + 1}`$. The contention window is assumed to be common to all the stations in the network; for this reason the station index $`s`$ is dropped from the contention model depicted in Fig. [1](#S2.F1 "Figure 1 ‣ II Markovian Model Characterizing the MAC Layer under general traffic conditions and Real Transmission Channel ‣ On the Behavior of the Distributed Coordination Function of IEEE 802.11 with Multirate Capability under General Transmission Conditions").
In order to account for non-saturated traffic conditions, we have introduced a new state labelled $`I`$, accounting for the following two situations:
- •
Immediately after a successful transmission, the queue of the transmitting station is empty. This event occurs with probability $`{({1 - q^{(s)}})}\hspace{0pt}{({1 - P_{e\hspace{0pt}q}^{(s)}})}`$, whereby $`q^{(s)}`$ is the probability that there is at least one packet in the queue after a successful transmission (see Section [V](#S5 "V Traffic Model ‣ On the Behavior of the Distributed Coordination Function of IEEE 802.11 with Multirate Capability under General Transmission Conditions")).
- •
The station is in an idle state with an empty queue until a new packet arrives in the queue. Probability $`P_{I,0}^{(s)}`$ represents the probability that while the station resides in the idle state $`I`$ there is at least one packet arrival, and a new backoff procedure is scheduled.
We note that the probability $`P_{I,0}`$ of staying in the idle state is strictly related to the adopted traffic model. This issue will be further investigated.
The transition probabilities
<sup>1</sup>
<sup>1</sup>1$`P_{i,{k|{j,n}}}^{(s)}`$ is short for $`P^{(s)}\hspace{0pt}\left. \{{{{s\hspace{0pt}{({t + 1})}^{(s)}} = i},{{b\hspace{0pt}{({t + 1})}^{(s)}} = k}} \middle| {{{s\hspace{0pt}{(t)}^{(s)}} = j},{{b\hspace{0pt}{(t)}^{(s)}} = n}}\} \right.`$.
for the generic $`s`$-th station’s Markov process in Fig. [1](#S2.F1 "Figure 1 ‣ II Markovian Model Characterizing the MAC Layer under general traffic conditions and Real Transmission Channel ‣ On the Behavior of the Distributed Coordination Function of IEEE 802.11 with Multirate Capability under General Transmission Conditions") are as follow:
<table>
<tbody>
<tr>
<td></td>
<td>Pi,k|i,k+1(s)=1,k∈[0,Wi−2],i∈[0,m]P0,k|i,0(s)=q(s)(1−Peq(s))/W0,k∈[0,W0−1],i∈[0,m]Pi,k|i−1,0(s)=Peq(s)/Wi,k∈[0,Wi−1],i∈[1,m]Pm,k|m,0(s)=Peq(s)/Wm,k∈[0,Wm−1]PI|i,0(s)=(1−q(s))(1−Peq(s)),i∈[0,m]P0,k|I(s)=PI,0(s)/W0,k∈[0,W0−1]PI|I(s)=1−PI,0(s)superscriptsubscript𝑃𝑖conditional𝑘𝑖𝑘1𝑠absent1formulae-sequence𝑘0subscript𝑊𝑖2𝑖0𝑚superscriptsubscript𝑃0conditional𝑘𝑖0𝑠absentsuperscript𝑞𝑠1superscriptsubscript𝑃𝑒𝑞𝑠subscript𝑊0formulae-sequence𝑘0subscript𝑊01𝑖0𝑚superscriptsubscript𝑃𝑖conditional𝑘𝑖10𝑠absentsuperscriptsubscript𝑃𝑒𝑞𝑠subscript𝑊𝑖formulae-sequence𝑘0subscript𝑊𝑖1𝑖1𝑚superscriptsubscript𝑃𝑚conditional𝑘𝑚0𝑠absentsuperscriptsubscript𝑃𝑒𝑞𝑠subscript𝑊𝑚𝑘0subscript𝑊𝑚1superscriptsubscript𝑃conditional𝐼𝑖0𝑠absent1superscript𝑞𝑠1superscriptsubscript𝑃𝑒𝑞𝑠𝑖0𝑚superscriptsubscript𝑃0conditional𝑘𝐼𝑠absentsuperscriptsubscript𝑃𝐼0𝑠subscript𝑊0𝑘0subscript𝑊01superscriptsubscript𝑃conditional𝐼𝐼𝑠absent1superscriptsubscript𝑃𝐼0𝑠missing-subexpression\small\begin{array}[]{lll}P_{i,k|i,k+1}^{(s)}&=1,&~{}k\in[0,W_{i}-2],~{}i\in[0,m]\\ P_{0,k|i,0}^{(s)}&=q^{(s)}(1-P_{eq}^{(s)})/W_{0},&~{}k\in[0,W_{0}-1],~{}i\in[0,m]\\ P_{i,k|i-1,0}^{(s)}&=P_{eq}^{(s)}/W_{i},&~{}k\in[0,W_{i}-1],~{}i\in[1,m]\\ P_{m,k|m,0}^{(s)}&=P_{eq}^{(s)}/W_{m},&~{}k\in[0,W_{m}-1]\\ P_{I|i,0}^{(s)}&=(1-q^{(s)})(1-P_{eq}^{(s)}),&~{}i\in[0,m]\\ P_{0,k|I}^{(s)}&=P_{I,0}^{(s)}/W_{0},&~{}k\in[0,W_{0}-1]\\ P_{I|I}^{(s)}&=1-P_{I,0}^{(s)}&\end{array}</td>
<td></td>
<td>(2)</td>
</tr>
</tbody>
</table>
The first equation in ([2](#S2.E2 "In II Markovian Model Characterizing the MAC Layer under general traffic conditions and Real Transmission Channel ‣ On the Behavior of the Distributed Coordination Function of IEEE 802.11 with Multirate Capability under General Transmission Conditions")) states that at the beginning of each backoff slot time, the backoff time is decremented. The second equation accounts for the fact that after a successful transmission, a new packet transmission starts with backoff stage 0 with probability $`q^{(s)}`$, in case there is a new packet in the queue to be transmitted. Third and fourth equations deal with unsuccessful transmissions and the need to reschedule a new contention stage. The fifth equation deals with the practical situation in which after a successful transmission, the queue of the station is empty, and as a consequence, the station transits in the idle state $`I`$ waiting for a new packet arrival. The sixth equation models the situation in which a new packet arrives in the queue while the station is in the idle state. Finally, the seventh equation models the situation in which there are no packets to be transmitted and the station is in the idle state.
## III Markovian Process Analysis
This section focuses on the evaluation of the stationary state distribution of the Markov model proposed in the previous section. The objective is to find the probability that a station occupies a given state at any discrete time slot:
<table>
<tbody>
<tr>
<td></td>
<td>bi,k(s)=limt→∞P[s(s)(t)=i,b(s)(t)=k],∀k∈[0,Wi−1],∀i∈[0,m]formulae-sequencesuperscriptsubscript𝑏𝑖𝑘𝑠subscript→𝑡𝑃delimited-[]formulae-sequencesuperscript𝑠𝑠𝑡𝑖superscript𝑏𝑠𝑡𝑘formulae-sequencefor-all𝑘0subscript𝑊𝑖1for-all𝑖0𝑚b_{i,k}^{(s)}=\lim_{t\rightarrow\infty}P[s^{(s)}(t)=i,b^{(s)}(t)=k],~{}\forall k\in[0,W_{i}-1],~{}\forall i\in[0,m]</td>
<td></td>
</tr>
</tbody>
</table>
along with the stationary probability $`b_{I}^{(s)}`$ of being in the idle state. This mathematical derivation is at the core of finding the probability $`\tau^{(s)}`$ that a station will attempt transmission in a randomly chosen time slot. In order to simplify the notation, in what follows we will omit the superscript $`(s)`$ since the mathematical derivations are valid for any contending station $`s = {1,\ldots,N}`$.
For future developments, from the model depicted in Fig. [1](#S2.F1 "Figure 1 ‣ II Markovian Model Characterizing the MAC Layer under general traffic conditions and Real Transmission Channel ‣ On the Behavior of the Distributed Coordination Function of IEEE 802.11 with Multirate Capability under General Transmission Conditions") we note the following relations:
<table>
<tbody>
<tr>
<td></td>
<td>bi,0=Peq⋅bi−1,0=Peqi⋅b0,0,∀i∈[1,m−1]bm,0=Peqm1−Peq⋅b0,0,i=msubscript𝑏𝑖0⋅subscript𝑃𝑒𝑞subscript𝑏𝑖10⋅superscriptsubscript𝑃𝑒𝑞𝑖subscript𝑏00for-all𝑖1𝑚1subscript𝑏𝑚0⋅superscriptsubscript𝑃𝑒𝑞𝑚1subscript𝑃𝑒𝑞subscript𝑏00𝑖𝑚\begin{array}[]{rcll}b_{i,0}&=&P_{eq}\cdot b_{i-1,0}=P_{eq}^{i}\cdot b_{0,0},&\forall i\in[1,m-1]\\ b_{m,0}&=&\frac{P_{eq}^{m}}{1-P_{eq}}\cdot b_{0,0},&i=m\\ \end{array}</td>
<td></td>
<td>(3)</td>
</tr>
</tbody>
</table>
whereby, $`P_{e\hspace{0pt}q}`$ is the equivalent probability of failed transmission
<sup>2</sup>
<sup>2</sup>2For simplicity, we assume that at each transmission attempt, any station will encounter a constant and independent probability of failed transmission, $`P_{e\hspace{0pt}q}`$, independently from the number of retransmissions already suffered by each station.
that takes into account the need for a new contention due to either packet collision ($`P_{c\hspace{0pt}o\hspace{0pt}l}`$) or channel errors ($`P_{e}`$).
Let us focus on the meaning of the idle state $`I`$ noted in Fig. [1](#S2.F1 "Figure 1 ‣ II Markovian Model Characterizing the MAC Layer under general traffic conditions and Real Transmission Channel ‣ On the Behavior of the Distributed Coordination Function of IEEE 802.11 with Multirate Capability under General Transmission Conditions") to which the stationary probability $`b_{I}`$ is associated. This state considers both the situation in which after a successful transmission there are no packets in the station queue to be transmitted, and the situation in which the packet queue is empty and the station is waiting for new packet arrivals. The stationary probability of being in state $`b_{I}`$ can be evaluated as follows:
<table>
<tbody>
<tr>
<td></td>
<td>bI=(1−q)(1−Peq)∑i=0mbi,0+(1−PI,0)bI=(1−q)(1−Peq)PI,0⋅∑i=0mbi,0subscript𝑏𝐼1𝑞1subscript𝑃𝑒𝑞superscriptsubscript𝑖0𝑚subscript𝑏𝑖01subscript𝑃𝐼0subscript𝑏𝐼missing-subexpression⋅1𝑞1subscript𝑃𝑒𝑞subscript𝑃𝐼0superscriptsubscript𝑖0𝑚subscript𝑏𝑖0\begin{array}[]{lll}b_{I}&=&(1-q)(1-P_{eq})\sum_{i=0}^{m}b_{i,0}+(1-P_{I,0})b_{I}\\ &=&\frac{(1-q)(1-P_{eq})}{P_{I,0}}\cdot\sum_{i=0}^{m}b_{i,0}\end{array}</td>
<td></td>
<td>(4)</td>
</tr>
</tbody>
</table>
Upon employing the probabilities $`b_{i,0}`$ noted in ([3](#S3.E3 "In III Markovian Process Analysis ‣ On the Behavior of the Distributed Coordination Function of IEEE 802.11 with Multirate Capability under General Transmission Conditions")), it is straightforward to obtain:
<table>
<tbody>
<tr>
<td></td>
<td>∑i=0mbi,0=b0,0[∑i=0m−1Peqi+Peqm1−Peq]=b0,01−Peqsuperscriptsubscript𝑖0𝑚subscript𝑏𝑖0subscript𝑏00delimited-[]superscriptsubscript𝑖0𝑚1subscriptsuperscript𝑃𝑖𝑒𝑞superscriptsubscript𝑃𝑒𝑞𝑚1subscript𝑃𝑒𝑞subscript𝑏001subscript𝑃𝑒𝑞\sum_{i=0}^{m}b_{i,0}=b_{0,0}\left[\sum_{i=0}^{m-1}P^{i}_{eq}+\frac{P_{eq}^{m}}{1-P_{eq}}\right]=\frac{b_{0,0}}{1-P_{eq}}</td>
<td></td>
<td>(5)</td>
</tr>
</tbody>
</table>
By using the previous result, ([4](#S3.E4 "In III Markovian Process Analysis ‣ On the Behavior of the Distributed Coordination Function of IEEE 802.11 with Multirate Capability under General Transmission Conditions")) simplifies to
<table>
<tbody>
<tr>
<td></td>
<td>bI=1−qPI,0⋅b0,0subscript𝑏𝐼⋅1𝑞subscript𝑃𝐼0subscript𝑏00b_{I}=\frac{1-q}{P_{I,0}}\cdot b_{0,0}</td>
<td></td>
<td>(6)</td>
</tr>
</tbody>
</table>
Equ. ([4](#S3.E4 "In III Markovian Process Analysis ‣ On the Behavior of the Distributed Coordination Function of IEEE 802.11 with Multirate Capability under General Transmission Conditions")) reflects the fact that state $`b_{I}`$ can be reached after a successful packet transmission from any state $`{b_{i,0},{\forall i}} \in {\lbrack 0,m\rbrack}`$ with probability $`{({1 - q})}\hspace{0pt}{({1 - P_{e\hspace{0pt}q}})}`$, or because the station remains in idle state with probability $`({1 - P_{I,0}})`$, whereby $`P_{I,0}`$ is the probability of having at least one packet to be transmitted in the queue. The statistical model of $`P_{I,0}`$ will be discussed in Section [V](#S5 "V Traffic Model ‣ On the Behavior of the Distributed Coordination Function of IEEE 802.11 with Multirate Capability under General Transmission Conditions"). It is anticipated that $`P_{I,0}`$ is related to the specific traffic model chosen for modeling packet arrivals in the station queue.
The other stationary probabilities for any $`k \in {\lbrack 1,{W_{i} - 1}\rbrack}`$ follow by resorting to the state transition diagram shown in Fig. [1](#S2.F1 "Figure 1 ‣ II Markovian Model Characterizing the MAC Layer under general traffic conditions and Real Transmission Channel ‣ On the Behavior of the Distributed Coordination Function of IEEE 802.11 with Multirate Capability under General Transmission Conditions"):
<table>
<tbody>
<tr>
<td></td>
<td>bi,k=Wi−kWi{q(1−Peq)⋅∑i=0mbi,0++PI,0⋅bI,i=0Peq⋅bi−1,0,i∈[1,m−1]Peq(bm−1,0+bm,0),i=msubscript𝑏𝑖𝑘subscript𝑊𝑖𝑘subscript𝑊𝑖caseslimit-from⋅𝑞1subscript𝑃𝑒𝑞superscriptsubscript𝑖0𝑚subscript𝑏𝑖0missing-subexpression⋅subscript𝑃𝐼0subscript𝑏𝐼𝑖0⋅subscript𝑃𝑒𝑞subscript𝑏𝑖10𝑖1𝑚1subscript𝑃𝑒𝑞subscript𝑏𝑚10subscript𝑏𝑚0𝑖𝑚b_{i,k}=\frac{W_{i}-k}{W_{i}}\left\{\begin{array}[]{ll}q(1-P_{eq})\cdot\sum_{i=0}^{m}b_{i,0}+&\\ +P_{I,0}\cdot b_{I},&i=0\\ P_{eq}\cdot b_{i-1,0},&i\in[1,m-1]\\ P_{eq}(b_{m-1,0}+b_{m,0}),&i=m\\ \end{array}\right.</td>
<td></td>
<td>(7)</td>
</tr>
</tbody>
</table>
Upon substituting ([4](#S3.E4 "In III Markovian Process Analysis ‣ On the Behavior of the Distributed Coordination Function of IEEE 802.11 with Multirate Capability under General Transmission Conditions")) in ([7](#S3.E7 "In III Markovian Process Analysis ‣ On the Behavior of the Distributed Coordination Function of IEEE 802.11 with Multirate Capability under General Transmission Conditions")), $`b_{0,k}`$ can be obtained as follows:
<table>
<tbody>
<tr>
<td></td>
<td>q(1−Peq)⋅∑i=0mbi,0+PI,0⋅bI=q(1−Peq)⋅∑i=0mbi,0+PI,0⋅(1−q)(1−Peq)PI,0⋅∑i=0mbi,0==(1−Peq)⋅∑i=0mbi,0⋅𝑞1subscript𝑃𝑒𝑞superscriptsubscript𝑖0𝑚subscript𝑏𝑖0⋅subscript𝑃𝐼0subscript𝑏𝐼absentmissing-subexpression⋅𝑞1subscript𝑃𝑒𝑞superscriptsubscript𝑖0𝑚subscript𝑏𝑖0⋅subscript𝑃𝐼01𝑞1subscript𝑃𝑒𝑞subscript𝑃𝐼0superscriptsubscript𝑖0𝑚subscript𝑏𝑖0absentmissing-subexpressionabsent⋅1subscript𝑃𝑒𝑞superscriptsubscript𝑖0𝑚subscript𝑏𝑖0missing-subexpression\begin{array}[]{ll}q(1-P_{eq})\cdot\sum_{i=0}^{m}b_{i,0}+P_{I,0}\cdot b_{I}=&\\ q(1-P_{eq})\cdot\sum_{i=0}^{m}b_{i,0}+P_{I,0}\cdot\frac{(1-q)(1-P_{eq})}{P_{I,0}}\cdot\sum_{i=0}^{m}b_{i,0}=&\\ =(1-P_{eq})\cdot\sum_{i=0}^{m}b_{i,0}&\end{array}</td>
<td></td>
<td>(8)</td>
</tr>
</tbody>
</table>
Employing the normalization condition, after some mathematical manipulations, and remembering ([5](#S3.E5 "In III Markovian Process Analysis ‣ On the Behavior of the Distributed Coordination Function of IEEE 802.11 with Multirate Capability under General Transmission Conditions")), it is possible to obtain:
<table>
<tbody>
<tr>
<td></td>
<td>11\displaystyle 1</td>
<td>=\displaystyle=</td>
<td>∑i=0m∑k=0Wi−1bi,k+bIsuperscriptsubscript𝑖0𝑚superscriptsubscript𝑘0subscript𝑊𝑖1subscript𝑏𝑖𝑘subscript𝑏𝐼\displaystyle\sum_{i=0}^{m}\sum_{k=0}^{W_{i}-1}b_{i,k}+b_{I}</td>
<td></td>
<td rowspan="3">(9)</td>
</tr>
<tr>
<td></td>
<td></td>
<td>=\displaystyle=</td>
<td>b0,02{W0[∑i=0m−1(2Peq)i+(2Peq)m1−Peq]+11−Peq}+bIsubscript𝑏002subscript𝑊0delimited-[]superscriptsubscript𝑖0𝑚1superscript2subscript𝑃𝑒𝑞𝑖superscript2subscript𝑃𝑒𝑞𝑚1subscript𝑃𝑒𝑞11subscript𝑃𝑒𝑞subscript𝑏𝐼\displaystyle\frac{b_{0,0}}{2}\left\{W_{0}\left[\sum_{i=0}^{m-1}(2P_{eq})^{i}+\frac{(2P_{eq})^{m}}{1-P_{eq}}\right]+\frac{1}{1-P_{eq}}\right\}+b_{I}</td>
<td></td>
</tr>
<tr>
<td></td>
<td></td>
<td>=\displaystyle=</td>
<td>α⋅b0,0+bI⋅𝛼subscript𝑏00subscript𝑏𝐼\displaystyle\alpha\cdot b_{0,0}+b_{I}</td>
<td></td>
</tr>
</tbody>
</table>
whereby,
<table>
<tbody>
<tr>
<td></td>
<td>α=12{W0[1−(2Peq)m1−2Peq+(2Peq)m1−Peq]+11−Peq}𝛼12subscript𝑊0delimited-[]1superscript2subscript𝑃𝑒𝑞𝑚12subscript𝑃𝑒𝑞superscript2subscript𝑃𝑒𝑞𝑚1subscript𝑃𝑒𝑞11subscript𝑃𝑒𝑞\alpha=\frac{1}{2}\left\{W_{0}\left[\frac{1-(2P_{eq})^{m}}{1-2P_{eq}}+\frac{(2P_{eq})^{m}}{1-P_{eq}}\right]+\frac{1}{1-P_{eq}}\right\}</td>
<td></td>
<td>(10)</td>
</tr>
</tbody>
</table>
From ([9](#S3.E9 "In III Markovian Process Analysis ‣ On the Behavior of the Distributed Coordination Function of IEEE 802.11 with Multirate Capability under General Transmission Conditions")), the following equation for computation of $`b_{0,0}`$ easily follows:
<table>
<tbody>
<tr>
<td></td>
<td>b0,0=1−bIαsubscript𝑏001subscript𝑏𝐼𝛼b_{0,0}=\frac{1-b_{I}}{\alpha}</td>
<td></td>
<td>(11)</td>
</tr>
</tbody>
</table>
Equ. ([11](#S3.E11 "In III Markovian Process Analysis ‣ On the Behavior of the Distributed Coordination Function of IEEE 802.11 with Multirate Capability under General Transmission Conditions")) is used to compute $`\tau^{(s)}`$, the probability that the $`s`$-th station starts a transmission in a randomly chosen time slot. In fact, taking into account that a packet transmission occurs when the backoff counter reaches zero, we have:
<table>
<tbody>
<tr>
<td></td>
<td>τ(s)superscript𝜏𝑠\displaystyle\tau^{(s)}</td>
<td>=\displaystyle=</td>
<td>∑i=0mbi,0(s)=b0,0(s)1−Peq(s)=1−bI(s)α(s)(1−Peq(s))=superscriptsubscript𝑖0𝑚superscriptsubscript𝑏𝑖0𝑠superscriptsubscript𝑏00𝑠1superscriptsubscript𝑃𝑒𝑞𝑠1superscriptsubscript𝑏𝐼𝑠superscript𝛼𝑠1superscriptsubscript𝑃𝑒𝑞𝑠absent\displaystyle\sum_{i=0}^{m}b_{i,0}^{(s)}=\frac{b_{0,0}^{(s)}}{1-P_{eq}^{(s)}}=\frac{1-b_{I}^{(s)}}{\alpha^{(s)}(1-P_{eq}^{(s)})}=</td>
<td></td>
</tr>
<tr>
<td></td>
<td></td>
<td>=\displaystyle=</td>
<td>2(1−bI(s))(1−2Peq(s))(W0+1)(1−2Peq(s))+W0Peq(s)(1−(2Peq(s))m)21superscriptsubscript𝑏𝐼𝑠12superscriptsubscript𝑃𝑒𝑞𝑠subscript𝑊0112superscriptsubscript𝑃𝑒𝑞𝑠subscript𝑊0superscriptsubscript𝑃𝑒𝑞𝑠1superscript2superscriptsubscript𝑃𝑒𝑞𝑠𝑚\displaystyle\frac{2(1-b_{I}^{(s)})(1-2P_{eq}^{(s)})}{(W_{0}+1)(1-2P_{eq}^{(s)})+W_{0}P_{eq}^{(s)}(1-(2P_{eq}^{(s)})^{m})}</td>
<td></td>
</tr>
</tbody>
</table>
whereby, we re-introduced the superscript $`(s)`$ since this expression will be used in what follows.
The collision probability $`P_{c\hspace{0pt}o\hspace{0pt}l}^{(s)}`$ needed to compute $`\tau^{(s)}`$ can be found considering that using a 2-way hand-shaking mechanism, a packet from a transmitting station encounters a collision if in a given time slot, at least one of the remaining $`({N - 1})`$ stations transmits one packet. Since each station has its own $`\tau^{(s)}`$, the collision probability for the $`s`$-th contending station depends on the transmission probabilities of the remaining stations as follows:
<table>
<tbody>
<tr>
<td></td>
<td>Pcol(s)=1−∏j=1j≠sN(1−τ(j))superscriptsubscript𝑃𝑐𝑜𝑙𝑠1superscriptsubscriptproduct𝑗1𝑗𝑠𝑁1superscript𝜏𝑗P_{col}^{(s)}=1-\prod_{\begin{subarray}{c}j=1\\ j\neq s\end{subarray}}^{N}(1-\tau^{(j)})</td>
<td></td>
<td>(13)</td>
</tr>
</tbody>
</table>
Given the set of $`N`$ equations ([1](#S2.E1 "In II Markovian Model Characterizing the MAC Layer under general traffic conditions and Real Transmission Channel ‣ On the Behavior of the Distributed Coordination Function of IEEE 802.11 with Multirate Capability under General Transmission Conditions")) and ([III](#S3.Ex4 "III Markovian Process Analysis ‣ On the Behavior of the Distributed Coordination Function of IEEE 802.11 with Multirate Capability under General Transmission Conditions")), a non-linear system of $`2\hspace{0pt}N`$ equations can be solved in order to determine the values of $`\tau^{(s)}`$ and $`P_{c\hspace{0pt}o\hspace{0pt}l}^{(s)}`$ for any $`s = {1,\ldots,N}`$; this is the operating point corresponding to the $`N`$ stations in the network needed in order to determine the aggregate throughput of the network, defined as the fraction of time the channel is used to successfully transmit payload bits:
<table>
<tbody>
<tr>
<td></td>
<td>S=∑s=1N1TavPs(s)⋅(1−Pe(s))⋅PL𝑆superscriptsubscript𝑠1𝑁⋅1subscript𝑇𝑎𝑣superscriptsubscript𝑃𝑠𝑠1superscriptsubscript𝑃𝑒𝑠𝑃𝐿S=\sum_{s=1}^{N}\frac{1}{T_{av}}P_{s}^{(s)}\cdot(1-P_{e}^{(s)})\cdot PL</td>
<td></td>
<td>(14)</td>
</tr>
</tbody>
</table>
whereby, the summation is over the throughput related to the $`N`$ contending stations, $`P\hspace{0pt}L`$ is the average payload size, and $`T_{a\hspace{0pt}v}`$ is the expected time per slot defined in the following.
Probabilities involved in ([14](#S3.E14 "In III Markovian Process Analysis ‣ On the Behavior of the Distributed Coordination Function of IEEE 802.11 with Multirate Capability under General Transmission Conditions")) are as follows: $`P_{e}^{(s)}`$ is the PER of the $`s`$-th station due to imperfect channel transmissions and $`P_{s}^{(s)}`$ is the probability of successful packet transmission from the $`s`$-th station.
In the next section, we derive the mathematical relations defining both $`T_{a\hspace{0pt}v}`$ and the probabilities involved in ([14](#S3.E14 "In III Markovian Process Analysis ‣ On the Behavior of the Distributed Coordination Function of IEEE 802.11 with Multirate Capability under General Transmission Conditions")).
## IV Estimating the Average Time Slot Duration
In order to proceed further, we need to evaluate the average time that a station spends in any possible state, i.e., the average time slot duration $`T_{a\hspace{0pt}v}`$, in terms of the key probabilities involved in the proposed model. This is the focus of the current section.
The average duration $`T_{a\hspace{0pt}v}`$ of a time slot (or expected time per slot) can be evaluated by weighting the times spent by a station in a particular state with the probability of being in that state. It is possible to note four kinds of time slots.
- •
The average idle slot duration, identified by $`T_{I}`$, in which no station is transmitting over the channel.
- •
The average collision slot duration, identified by $`T_{C}`$, in which more than one station is attempting to gain access to the channel.
- •
The average duration of the slot due to erroneous transmissions because of imperfect channel conditions, identified by $`T_{E}`$.
- •
The average slot duration of a successful transmission, identified by $`T_{S}`$.
###
IV-A
The average idle slot duration
The average idle slot duration can be evaluated as the probability $`({1 - P_{t}})`$ that no station is attempting to gain access to the channel times the duration $`\sigma`$ of an empty slot time.
Let $`P_{t}`$ be the probability that the channel is busy in a slot because at least one station is transmitting. Then, we have:
<table>
<tbody>
<tr>
<td></td>
<td>Pt=1−∏s=1N(1−τ(s))subscript𝑃𝑡1superscriptsubscriptproduct𝑠1𝑁1superscript𝜏𝑠P_{t}=1-\prod_{s=1}^{N}(1-\tau^{(s)})</td>
<td></td>
<td>(15)</td>
</tr>
</tbody>
</table>
The average idle slot duration can be defined as:
<table>
<tbody>
<tr>
<td></td>
<td>TI=PI⋅σ=(1−Pt)⋅σsubscript𝑇𝐼⋅subscript𝑃𝐼𝜎⋅1subscript𝑃𝑡𝜎T_{I}=P_{I}\cdot\sigma=(1-P_{t})\cdot\sigma</td>
<td></td>
<td>(16)</td>
</tr>
</tbody>
</table>
where each idle slot is assumed to have duration $`\sigma`$.
###
IV-B
The average slot duration of a successful transmission
Consider a tagged station in the set of $`N`$ stations in the analyzed network, and let $`s`$ be its index in the set $`\{ 1,\ldots,N\}`$. Then, the average slot duration of a successful transmission can be found upon evaluating the probability that only the $`s`$-th tagged station is successfully transmitting over the channel, i.e.,
<table>
<tbody>
<tr>
<td></td>
<td>Ps(s)=τ(s)∏j=1j≠sN(1−τ(j))superscriptsubscript𝑃𝑠𝑠superscript𝜏𝑠superscriptsubscriptproduct𝑗1𝑗𝑠𝑁1superscript𝜏𝑗P_{s}^{(s)}=\tau^{(s)}\prod_{\begin{subarray}{c}j=1\\ j\neq s\end{subarray}}^{N}(1-\tau^{(j)})</td>
<td></td>
<td>(17)</td>
</tr>
</tbody>
</table>
times the duration of a successful transmission from the $`s`$-th station, $`T_{s}^{(s)}`$. The latter depends on the rate-class $`(r)`$ which the tagged station belong to, and can be evaluated as follows:
<table>
<tbody>
<tr>
<td></td>
<td>Ts(s)superscriptsubscript𝑇𝑠𝑠\displaystyle T_{s}^{(s)}</td>
<td>=\displaystyle=</td>
<td>HPHYRC+HMAC+PLRD(s)+δ+subscript𝐻𝑃𝐻𝑌subscript𝑅𝐶subscript𝐻𝑀𝐴𝐶𝑃𝐿superscriptsubscript𝑅𝐷𝑠limit-from𝛿\displaystyle\frac{H_{PHY}}{R_{C}}+\frac{H_{MAC}+PL}{R_{D}^{(s)}}+\delta+</td>
<td></td>
</tr>
<tr>
<td></td>
<td></td>
<td></td>
<td>+SIFS+HPHY+ACKRC+δ+DIFS𝑆𝐼𝐹𝑆subscript𝐻𝑃𝐻𝑌𝐴𝐶𝐾subscript𝑅𝐶𝛿𝐷𝐼𝐹𝑆\displaystyle+SIFS+\frac{H_{PHY}+ACK}{R_{C}}+\delta+DIFS</td>
<td></td>
</tr>
</tbody>
</table>
whereby, $`P\hspace{0pt}L`$ is the average payload length, $`H_{P\hspace{0pt}H\hspace{0pt}Y}`$ and $`H_{M\hspace{0pt}A\hspace{0pt}C}`$ are, respectively, the physical and MAC header sizes, $`\tau_{p}`$ is the propagation delay, DIFS is the duration of the Distributed InterFrame Space, $`R_{C}`$ is the basic data rate used for transmitting protocol data, and $`R_{D}^{(s)}`$ is the data rate of the $`s`$-th station.
With this setup, the average slot duration of a successful transmission can be evaluated as:
<table>
<tbody>
<tr>
<td></td>
<td>TS=∑i=1NPs(i)(1−Pe(i))⋅Ts(i)subscript𝑇𝑆superscriptsubscript𝑖1𝑁⋅superscriptsubscript𝑃𝑠𝑖1superscriptsubscript𝑃𝑒𝑖superscriptsubscript𝑇𝑠𝑖T_{S}=\sum_{i=1}^{N}P_{s}^{(i)}\left(1-P_{e}^{(i)}\right)\cdot T_{s}^{(i)}</td>
<td></td>
<td>(19)</td>
</tr>
</tbody>
</table>
whereby, $`\left( {1 - P_{e}^{(i)}} \right)`$ accounts for the probability of packet transmission without channel induced errors.
###
IV-C
The average collision slot duration
In a network of stations transmitting equal length packets with different data rates, the average duration of $`T_{C}`$ is largely dominated by the slowest transmitting stations. This phenomenon is called performance anomaly of 802.11b, and it has been first observed in \[[15](#bib.bib15)\].
As an example, suppose that a frame transmitted by a station using the rate 1 Mbps (class 1) collides with the one of a station transmitting at rate 11 Mbps one (class 4). Of course, both frames get corrupted while the channel appears busy to the remaining sensing stations for the whole duration of the frame transmitted by the low rate station. This implies that fast stations (high classes) are penalized by slow stations (low classes), causing a drop of the aggregate throughput. The duration of the collision depends on the frame duration of the lowest rate stations. In order to evaluate the collision probability, we define the class $`(r)`$ collision duration as follows:
<table>
<tbody>
<tr>
<td></td>
<td>Tc(r)=HPHYRC+HMAC+PLRD(r)+ACKtimeoutsuperscriptsubscript𝑇𝑐𝑟subscript𝐻𝑃𝐻𝑌subscript𝑅𝐶subscript𝐻𝑀𝐴𝐶𝑃𝐿superscriptsubscript𝑅𝐷𝑟𝐴𝐶subscript𝐾𝑡𝑖𝑚𝑒𝑜𝑢𝑡T_{c}^{(r)}=\frac{H_{PHY}}{R_{C}}+\frac{H_{MAC}+PL}{R_{D}^{(r)}}+ACK_{timeout}</td>
<td></td>
<td>(20)</td>
</tr>
</tbody>
</table>
which takes into account basic rate $`R_{C}`$ and data rate $`R_{D}^{(r)}`$ of class $`(r)`$.
For the derivations which follow, we consider a set of indexes which identify the stations transmitting with the $`r`$-th data rate:
<table>
<tbody>
<tr>
<td></td>
<td>n(r)={identifiers of stations belonging to rate-class (r)}𝑛𝑟identifiers of stations belonging to rate-class (r)n(r)=\{\mbox{identifiers of stations belonging to rate-class (r)}\}</td>
<td></td>
</tr>
</tbody>
</table>
$`{\forall r} \in R = {\{ 1,\ldots,4\}}`$ such that $`{\sum_{r = 1}^{N_{R}}{|n^{(r)}|}} = N`$ ($`| \cdot |`$ is the cardinality of the embraced set).
With this setup, there are two different kinds of collisions:
- •
intra-class collisions between at least two frames belonging to same rate class $`(r)`$;
- •
inter-class collisions between at least one frame of class $`(r)`$ and at least one frame of class $`{(j)} > {(r)}`$
As far as intra-class $`(r)`$ collisions are concerned, the collision probability $`P_{c\hspace{0pt}1}^{(r)}`$ can be evaluated as follows:
<table>
<tbody>
<tr>
<td></td>
<td>{1−[∏s∈n(r)(1−τ(s))+∑s∈n(r)τ(s)∏j∈n(r)j≠s(1−τ(j))]}⋅\displaystyle\small\left\{1-\left[\prod_{s\in n(r)}(1-\tau^{(s)})+\sum_{s\in n(r)}\tau^{(s)}\prod_{\begin{subarray}{c}j\in n(r)\\ j\neq s\end{subarray}}(1-\tau^{(j)})\right]\right\}\cdot</td>
<td></td>
<td></td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td>⋅∏s∈{S−n(r)}(1−τ(s))\displaystyle\cdot\prod_{s\in\{S-n(r)\}}(1-\tau^{(s)})</td>
<td></td>
<td>(21)</td>
</tr>
</tbody>
</table>
Notice that the latter is the probability that stations not belonging to the same data rate set $`n\hspace{0pt}{(r)}`$ do not transmit, times the probability that there are at least two stations in the same rate class $`n\hspace{0pt}{(r)}`$ transmitting over the channel \[[17](#bib.bib17)\]. Notice that the first product within brace brackets accounts for the scenario in which no stations with rate in $`n\hspace{0pt}{(r)}`$ transmit, or there is only one station transmitting with rate in $`n\hspace{0pt}{(r)}`$. As a side note, notice that $`P_{c\hspace{0pt}1}^{(r)} = 0`$ in case there are no collisions between stations belonging to the same rate class.
Following a similar reasoning, the inter-class $`(r)`$ collision probability $`P_{c\hspace{0pt}2}^{(r)}`$ can be evaluated as:
<table>
<tbody>
<tr>
<td></td>
<td>[1−∏s∈n(r)(1−τ(s))]⋅[1−∏j=r+1NR∏s∈n(j)(1−τ(s))]⋅\displaystyle\left[1-\prod_{s\in n(r)}(1-\tau^{(s)})\right]\cdot\left[1-\prod_{j=r+1}^{N_{R}}\prod_{s\in n(j)}(1-\tau^{(s)})\right]\cdot</td>
<td></td>
<td></td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td>⋅[∏j=1r−1∏s∈n(j)(1−τ(s))]⋅absentdelimited-[]superscriptsubscriptproduct𝑗1𝑟1subscriptproduct𝑠𝑛𝑗1superscript𝜏𝑠\displaystyle\cdot\left[\prod_{j=1}^{r-1}\prod_{s\in n(j)}(1-\tau^{(s)})\right]</td>
<td></td>
<td>(22)</td>
</tr>
</tbody>
</table>
which considers the scenario in which at least one station of class $`(r)`$ and at least one station belonging to a higher rate class $`(j)`$ (i.e., $`{(j)} > {(r)}`$) transmit in the same time slot, while all the other stations belonging to lower indexed classes (i.e., with $`{(i)} < {(r)}`$) are silent. As a side note, notice that $`P_{c\hspace{0pt}2}^{(r)} = 0`$ in case there are no collisions between stations belonging to different rate classes.
The total class $`(r)`$ collision probability is the sum of the previous two probabilities:
<table>
<tbody>
<tr>
<td></td>
<td>Pc(r)=Pc1(r)+Pc2(r)superscriptsubscript𝑃𝑐𝑟superscriptsubscript𝑃𝑐1𝑟superscriptsubscript𝑃𝑐2𝑟P_{c}^{(r)}=P_{c1}^{(r)}+P_{c2}^{(r)}</td>
<td></td>
<td>(23)</td>
</tr>
</tbody>
</table>
while the average collision slot duration can be computed considering the whole set of classes $`r \in R`$ and their collision probabilities weighted by their durations:
<table>
<tbody>
<tr>
<td></td>
<td>TC=∑r=1NRPc(r)⋅Tc(r)subscript𝑇𝐶superscriptsubscript𝑟1subscript𝑁𝑅⋅superscriptsubscript𝑃𝑐𝑟superscriptsubscript𝑇𝑐𝑟T_{C}=\sum_{r=1}^{N_{R}}P_{c}^{(r)}\cdot T_{c}^{(r)}</td>
<td></td>
<td>(24)</td>
</tr>
</tbody>
</table>
###
IV-D
The average duration of the slot due to erroneous transmissions
The average duration of the slot due to erroneous transmissions can be evaluated in a way similar to the one used for evaluating $`T_{S}`$ and $`T_{C}`$:
<table>
<tbody>
<tr>
<td></td>
<td>TE=∑i=1NPs(i)⋅Pe(i)⋅Te(i)subscript𝑇𝐸superscriptsubscript𝑖1𝑁⋅superscriptsubscript𝑃𝑠𝑖superscriptsubscript𝑃𝑒𝑖superscriptsubscript𝑇𝑒𝑖T_{E}=\sum_{i=1}^{N}P_{s}^{(i)}\cdot P_{e}^{(i)}\cdot T_{e}^{(i)}</td>
<td></td>
<td>(25)</td>
</tr>
</tbody>
</table>
whereby, $`P_{s}^{(i)}`$ is defined in ([17](#S4.E17 "In IV-B The average slot duration of a successful transmission ‣ IV Estimating the Average Time Slot Duration ‣ On the Behavior of the Distributed Coordination Function of IEEE 802.11 with Multirate Capability under General Transmission Conditions")), and $`T_{e}^{(s)}`$ is assumed to be equal to $`T_{c}^{(s)}`$ since when a channel error occurs, the transmitting station does not receive the acknowledgment before the end of the ACK timeout.
###
IV-E
Average time slot duration
Given the expected slots derived in the previous sections, the average duration of a slot time is simply:
<table>
<tbody>
<tr>
<td></td>
<td>Tav=TI+TC+TS+TEsubscript𝑇𝑎𝑣subscript𝑇𝐼subscript𝑇𝐶subscript𝑇𝑆subscript𝑇𝐸T_{av}=T_{I}+T_{C}+T_{S}+T_{E}</td>
<td></td>
<td>(26)</td>
</tr>
</tbody>
</table>
## V Traffic Model
This section presents the traffic model employed in our setup along with the derivation of the key probabilities $`q^{(s)}`$ and $`P_{I,0}^{(s)}`$ shown in Fig. [1](#S2.F1 "Figure 1 ‣ II Markovian Model Characterizing the MAC Layer under general traffic conditions and Real Transmission Channel ‣ On the Behavior of the Distributed Coordination Function of IEEE 802.11 with Multirate Capability under General Transmission Conditions"). The offered load related to each station is characterized by the parameter $`\lambda^{(s)}`$ representing the rate at which packets arrive at the $`s`$-th station buffer from the upper layers, and measured in packets per second. The time between two packet arrivals is defined as interarrival time, and its mean value is evaluated as $`\frac{1}{\lambda^{(s)}}`$. One of the most commonly used traffic models assumes that the packet arrival process is Poisson. The resulting interarrival times are exponentially distributed.
In the proposed model shown in Fig. [1](#S2.F1 "Figure 1 ‣ II Markovian Model Characterizing the MAC Layer under general traffic conditions and Real Transmission Channel ‣ On the Behavior of the Distributed Coordination Function of IEEE 802.11 with Multirate Capability under General Transmission Conditions"), we need to know the probability $`q^{(s)}`$ that there is at least one packet to be transmitted in the queue. Probability $`q^{(s)}`$ can be well approximated in a situation with small buffer size \[[11](#bib.bib11)\] through the following relation:
<table>
<tbody>
<tr>
<td></td>
<td>q(s)=1−e−λ(s)Tavsuperscript𝑞𝑠1superscript𝑒superscript𝜆𝑠subscript𝑇𝑎𝑣q^{(s)}=1-e^{-\lambda^{(s)}T_{av}}</td>
<td></td>
<td>(27)</td>
</tr>
</tbody>
</table>
where, $`T_{a\hspace{0pt}v}`$ is the expected time per slot, which is useful to relate the state of the Markov chain with the actual time spent in each state. Such a time has been derived in ([26](#S4.E26 "In IV-E Average time slot duration ‣ IV Estimating the Average Time Slot Duration ‣ On the Behavior of the Distributed Coordination Function of IEEE 802.11 with Multirate Capability under General Transmission Conditions")). Under the hypothesis of small queue systems, probabilities $`q^{(s)}`$ and $`P_{I,0}^{(s)}`$ can be evaluated considering that the probability of having at least one packet arrival in the queue at the end of a successful packet transmission can be approximated with the probability of having at least one packet arrival in an average time slot duration. As a result of this simple approximation, we have $`q^{(s)} = P_{I,0}^{(s)}`$. Upon remembering ([6](#S3.E6 "In III Markovian Process Analysis ‣ On the Behavior of the Distributed Coordination Function of IEEE 802.11 with Multirate Capability under General Transmission Conditions")) and ([11](#S3.E11 "In III Markovian Process Analysis ‣ On the Behavior of the Distributed Coordination Function of IEEE 802.11 with Multirate Capability under General Transmission Conditions")), $`\tau^{(s)}`$ in ([III](#S3.Ex4 "III Markovian Process Analysis ‣ On the Behavior of the Distributed Coordination Function of IEEE 802.11 with Multirate Capability under General Transmission Conditions")) can be evaluated as:
<table>
<tbody>
<tr>
<td></td>
<td>τ(s)=q(s)(1−Peq(s))−1q(s)(α(s)−1)+1=2(1−2Peq(s))q(s)D(q(s),W0,m,Peq(s))superscript𝜏𝑠superscript𝑞𝑠superscript1superscriptsubscript𝑃𝑒𝑞𝑠1superscript𝑞𝑠superscript𝛼𝑠11212superscriptsubscript𝑃𝑒𝑞𝑠superscript𝑞𝑠𝐷superscript𝑞𝑠subscript𝑊0𝑚superscriptsubscript𝑃𝑒𝑞𝑠\displaystyle\tau^{(s)}=\frac{q^{(s)}\left(1-P_{eq}^{(s)}\right)^{-1}}{q^{(s)}(\alpha^{(s)}-1)+1}=\frac{2(1-2P_{eq}^{(s)})q^{(s)}}{D(q^{(s)},W_{0},m,P_{eq}^{(s)})}</td>
<td></td>
<td>(28)</td>
</tr>
</tbody>
</table>
whereby,
<table>
<tbody>
<tr>
<td></td>
<td>D(q(s),W0,m,Peq(s))=q(s)[(W0+1)(1−2Peq(s))+W0Peq(s)⋅(1−(2Peq(s))m)]+2(1−q(s))(1−Peq(s))(1−2Peq(s))\small\begin{array}[]{ll}D(q^{(s)},W_{0},m,P_{eq}^{(s)})=q^{(s)}\left[(W_{0}+1)(1-2P_{eq}^{(s)})+W_{0}P_{eq}^{(s)}\cdot\right.&\\ \left.(1-(2P_{eq}^{(s)})^{m})\right]+2(1-q^{(s)})(1-P_{eq}^{(s)})(1-2P_{eq}^{(s)})&\\ \end{array}</td>
<td></td>
<td>(29)</td>
</tr>
</tbody>
</table>
Though simple, this approximation has been verified by simulation, proving to be quite effective for predicting the aggregate throughput.
## VI Physical Layer Modeling
In a scenario with $`N`$ contending stations randomly distributed around a common access point, throughput performance depends on the channel conditions experienced by each station. In what follows we briefly recall the main signal propagation issues in order to evaluate the PER experienced by a generic station in the network. This will serve as the basis for the simulated scenarios discussed in Section [VII](#S7 "VII Simulation Results and Model Validation ‣ On the Behavior of the Distributed Coordination Function of IEEE 802.11 with Multirate Capability under General Transmission Conditions").
###
VI-A
Signal Propagation
Consider a contending station at distance $`d`$ from an access point. Given the one-sided noise power spectral density
<sup>3</sup>
<sup>3</sup>3In what follows we will set the effective antenna temperature $`T = {273\hspace{0pt}K}`$ and $`N_{o} = {- 174}`$dBm.
, $`N_{o}`$, the received SNR can be evaluated as \[[23](#bib.bib23)\]:
<table>
<tbody>
<tr>
<td></td>
<td>SNRdB=P(d)|−dBmNo−Bw|−dBNFSNR_{\textrm{dB}}=P(d)\left|{}_{\textrm{dBm}}\right.-N_{o}-B_{w}\left|{}_{\textrm{dB}}\right.-N_{F}</td>
<td></td>
<td>(30)</td>
</tr>
</tbody>
</table>
whereby, $`N_{F}`$ is the receiver noise figure (10dB), while the power received at a distance $`d`$ denoted $`\left. P{(d)} \middle|_{\text{dBm}} \right.`$ is given by,
<table>
<tbody>
<tr>
<td></td>
<td>P(d)|=dBmPtx|−dBmL(d)|dBP(d)\left|{}_{dBm}\right.=P_{tx}\left|{}_{dBm}\right.-L(d)\left|{}_{dB}\right.</td>
<td></td>
<td>(31)</td>
</tr>
</tbody>
</table>
In the previous equation, $`\left. L{(d)} \middle|_{d\hspace{0pt}B} \right.`$ is the path-loss at distance $`d`$ from the transmitter
<table>
<tbody>
<tr>
<td></td>
<td>L(d)|=dBLo|+dB10⋅nplog10(dd0)L(d)\left|{}_{dB}\right.=L_{o}\left|{}_{dB}\right.+10\cdot n_{p}\log_{10}\left(\frac{d}{d_{0}}\right)</td>
<td></td>
</tr>
</tbody>
</table>
and $`L_{o}`$ is defined as,
<table>
<tbody>
<tr>
<td></td>
<td>Lo|=dB−10log10(GtGrλ2(4π)2d0np)L_{o}\left|{}_{dB}\right.=-10\log_{10}\left(\frac{G_{t}G_{r}\lambda^{2}}{(4\pi)^{2}d_{0}^{n_{p}}}\right)</td>
<td></td>
</tr>
</tbody>
</table>
where, $`d_{0}`$ is a reference distance (usually selected equal to 1m) with path-loss $`L_{o}`$.
The other symbols are defined as follows; $`G_{t}`$ is the transmitter antenna gain, $`G_{r}`$ is the receiver antenna gain, $`n_{p}`$ is the path-loss exponent, and $`\lambda = {c/f}`$ is the wavelength given by the ratio between the light velocity and the carrier frequency. The path-loss exponent $`n_{p}`$ depends on the specific propagation environment and it ranges from 2 (free space propagation) to 3.5-4 for non-line-of-sight propagation, or multi-path fast fading conditions in indoor environments \[[23](#bib.bib23)\]. Furthermore, based on FCC regulations, in the 2.4GHz ISM band the transmitted power $`\left. P_{t\hspace{0pt}x} \middle|_{d\hspace{0pt}B\hspace{0pt}m} \right.`$ is $`20`$dBm or equivalently, $`100`$mW.
The SNR per transmitted bit $`\gamma`$ accounts for the spreading gain $`C_{s}/B_{s}`$, and is defined as:
<table>
<tbody>
<tr>
<td></td>
<td>γ|=dBSNRdB+10log10(CsBs)\gamma\left|{}_{dB}\right.=SNR_{dB}+10\log_{10}\left(\frac{C_{s}}{B_{s}}\right)</td>
<td></td>
<td>(32)</td>
</tr>
</tbody>
</table>
whereby, $`C_{s}`$ is the number of chips per symbol while $`B_{s}`$ is the number of bits per transmitted symbol (see Table [I](#S6.T1 "TABLE I ‣ VI-A Signal Propagation ‣ VI Physical Layer Modeling ‣ On the Behavior of the Distributed Coordination Function of IEEE 802.11 with Multirate Capability under General Transmission Conditions")).
TABLE I: PHY setup of the IEEE 802.11b standard.
<table>
<tbody>
<tr>
<td>Frequency [GHz]</td>
<td>2.4</td>
<td>2.4</td>
<td>2.4</td>
<td>2.4</td>
</tr>
<tr>
<td>Bit rate [Mbps]</td>
<td>1</td>
<td>2</td>
<td>5.5</td>
<td>11</td>
</tr>
<tr>
<td>Modulation</td>
<td>DBSPK</td>
<td>DQPSK</td>
<td>CCK</td>
<td>CCK</td>
</tr>
<tr>
<td>Chips per symbol, Cssubscript𝐶𝑠C_{s}</td>
<td>11</td>
<td>11</td>
<td>8</td>
<td>8</td>
</tr>
<tr>
<td>Bits per symbol, Bssubscript𝐵𝑠B_{s}</td>
<td>1</td>
<td>2</td>
<td>4</td>
<td>8</td>
</tr>
<tr>
<td>Channel band [MHz]</td>
<td>22</td>
<td>22</td>
<td>22</td>
<td>22</td>
</tr>
<tr>
<td>Minimum power of received</td>
<td>-85</td>
<td>-82</td>
<td>-80</td>
<td>-76</td>
</tr>
<tr>
<td>signal (Sensitivity) [dBm]</td>
<td></td>
<td></td>
<td></td>
<td></td>
</tr>
</tbody>
</table>
###
VI-B
IEEE 802.11b Physical Layer
The physical layer (PHY) of the basic 802.11b standard is based on the spread spectrum technology. Two options are specified, the Frequency Hopped Spread Spectrum (FHSS) and the Direct Sequence Spread Spectrum (DSSS). The FHSS uses Frequency Shift Keying (FSK) while the DSSS uses Differential Phase Shift Keying (DPSK) or Complementary Code Keying (CCK).
DSSS transmits signals in the 2.4-GHz ISM band (i.e., 2.4000 -2.4835 GHz). The basic 802.11 DSSS gives data rates of 1 and 2 Mbps. The 802.11b extension \[[1](#bib.bib1)\] employs DSSS at various rates including one employing CCK encoding on 4 and 8 bits for each CCK symbol, or optionally employing packet binary convolutional coding. The four supported data rates in 802.11b are 1, 2, 5.5 and 11 Mbps.
BER performance of the various transmitting modes of IEEE802.11b are shown in Table [II](#S6.T2 "TABLE II ‣ VI-B IEEE 802.11b Physical Layer ‣ VI Physical Layer Modeling ‣ On the Behavior of the Distributed Coordination Function of IEEE 802.11 with Multirate Capability under General Transmission Conditions") for various channel models \[[2](#bib.bib2), [21](#bib.bib21), [22](#bib.bib22)\]. In Table [II](#S6.T2 "TABLE II ‣ VI-B IEEE 802.11b Physical Layer ‣ VI Physical Layer Modeling ‣ On the Behavior of the Distributed Coordination Function of IEEE 802.11 with Multirate Capability under General Transmission Conditions"), the SNR in ([32](#S6.E32 "In VI-A Signal Propagation ‣ VI Physical Layer Modeling ‣ On the Behavior of the Distributed Coordination Function of IEEE 802.11 with Multirate Capability under General Transmission Conditions")) is denoted as $`\gamma`$ while $`I_{k}\hspace{0pt}{({a\hspace{0pt}b})}`$ is the modified Bessel function of $`k`$th order.
TABLE II: BER $`P_{b}\hspace{0pt}{(\gamma)}`$ for various channel models
<table>
<thead>
<tr>
<th>Channel Model</th>
<th>DBPSK</th>
<th>DQPSK</th>
<th>CCK-5.5/11 Mbps</th>
</tr>
</thead>
<tbody>
<tr>
<th>AWGN</th>
<td>12erfc(γ)12erfc𝛾\frac{1}{2}\textrm{erfc}\left(\sqrt{\gamma}\right)</td>
<td>Q1(a,b)−12I0(ab)e−12a2−12b2subscript𝑄1𝑎𝑏12subscript𝐼0𝑎𝑏superscript𝑒12superscript𝑎212superscript𝑏2Q_{1}(a,b)-\frac{1}{2}I_{0}(ab)e^{-\frac{1}{2}a^{2}-\frac{1}{2}b^{2}}</td>
<td>1−∫−γ+∞[∫−(z+γ)+(z+γ)e−η22𝑑η]α2−1e−z22𝑑z1superscriptsubscript𝛾superscriptdelimited-[]superscriptsubscript𝑧𝛾𝑧𝛾superscript𝑒superscript𝜂22differential-d𝜂𝛼21superscript𝑒superscript𝑧22differential-d𝑧1-\int_{-\sqrt{\gamma}}^{+\infty}\left[\int_{-(z+\sqrt{\gamma})}^{+(z+\sqrt{\gamma})}e^{-\frac{\eta^{2}}{2}}d\eta\right]^{\frac{\alpha}{2}-1}e^{-\frac{z^{2}}{2}}dz</td>
</tr>
<tr>
<th></th>
<td></td>
<td>a=2γ(1−12),b=2γ(1+12)formulae-sequence𝑎2𝛾112𝑏2𝛾112a=\sqrt{2\gamma\left(1-\sqrt{\frac{1}{2}}\right)},~{}b=\sqrt{2\gamma\left(1+\sqrt{\frac{1}{2}}\right)}</td>
<td>α={4,5.5 Mbps8,11 Mbps𝛼cases45.5 Mbps811 Mbps\alpha=\left\{\begin{array}[]{ll}4,&\textrm{5.5 Mbps}\\ 8,&\textrm{11 Mbps}\end{array}\right.</td>
</tr>
<tr>
<th></th>
<td></td>
<td>Q1(a,b)=e−a2+b22∑k=0+∞(ab)kIk(ab)subscript𝑄1𝑎𝑏superscript𝑒superscript𝑎2superscript𝑏22superscriptsubscript𝑘0superscript𝑎𝑏𝑘subscript𝐼𝑘𝑎𝑏Q_{1}(a,b)=e^{-\frac{a^{2}+b^{2}}{2}}\sum_{k=0}^{+\infty}\left(\frac{a}{b}\right)^{k}I_{k}(ab)</td>
<td></td>
</tr>
<tr>
<th></th>
<td></td>
<td>I0(ab)=1+∑k=1+∞[(ab/2)kk!]2subscript𝐼0𝑎𝑏1superscriptsubscript𝑘1superscriptdelimited-[]superscript𝑎𝑏2𝑘𝑘2I_{0}\left(ab\right)=1+\sum_{k=1}^{+\infty}\left[\frac{\left(ab/2\right)^{k}}{k!}\right]^{2}</td>
<td></td>
</tr>
<tr>
<th>Rayleigh fading</th>
<td>12(1+γ)121𝛾\frac{1}{2\left(1+\gamma\right)}</td>
<td>12[1−γ221+γ22]12delimited-[]1𝛾221𝛾22\frac{1}{2}\left[1-\sqrt{\frac{\gamma\frac{\sqrt{2}}{2}}{1+\gamma\frac{\sqrt{2}}{2}}}\right]</td>
<td>2(α−1)2α−1∑i=1α−1(−1)i+1Ciα−11+i+i⋅γsuperscript2𝛼1superscript2𝛼1superscriptsubscript𝑖1𝛼1superscript1𝑖1subscriptsuperscript𝐶𝛼1𝑖1𝑖⋅𝑖𝛾\frac{2^{(\alpha-1)}}{2^{\alpha}-1}\sum_{i=1}^{\alpha-1}\frac{(-1)^{i+1}C^{\alpha-1}_{i}}{1+i+i\cdot\gamma}</td>
</tr>
<tr>
<th></th>
<td></td>
<td></td>
<td>α={4,5.5 Mbps8,11 Mbps𝛼cases45.5 Mbps811 Mbps\alpha=\left\{\begin{array}[]{ll}4,&\textrm{5.5 Mbps}\\ 8,&\textrm{11 Mbps}\end{array}\right.</td>
</tr>
<tr>
<th></th>
<td></td>
<td></td>
<td>Ciα−1=(α−1)!i!⋅(α−1−i)!subscriptsuperscript𝐶𝛼1𝑖𝛼1⋅𝑖𝛼1𝑖C^{\alpha-1}_{i}=\frac{(\alpha-1)!}{i!\cdot(\alpha-1-i)!}</td>
</tr>
</tbody>
</table>
The FER as a function of the SNR can be computed as follows:
<table>
<tbody>
<tr>
<td></td>
<td>Pe(DATA,SNR)=1−[1−Pe(PLCP,SNR)]⋅[1−Pe(PSDU,SNR)]subscript𝑃𝑒DATASNR1⋅delimited-[]1subscript𝑃𝑒PLCPSNRdelimited-[]1subscript𝑃𝑒PSDUSNR\small P_{e}(\textrm{DATA},\textrm{SNR})=1-\left[1-P_{e}(\textrm{PLCP},\textrm{SNR})\right]\cdot\left[1-P_{e}(\textrm{PSDU},\textrm{SNR})\right]</td>
<td></td>
<td>(33)</td>
</tr>
</tbody>
</table>
where,
<table>
<tbody>
<tr>
<td></td>
<td>Pe(PLCP,SNR)=1−[1−Pb(DBPSK,SNR)]8×PLCP,subscript𝑃𝑒PLCPSNR1superscriptdelimited-[]1subscript𝑃𝑏DBPSKSNR8PLCP\small P_{e}(\textrm{PLCP},\textrm{SNR})=1-\left[1-P_{b}(\textrm{DBPSK},\textrm{SNR})\right]^{8\times\textrm{PLCP}},</td>
<td></td>
<td>(34)</td>
</tr>
</tbody>
</table>
and
<table>
<tbody>
<tr>
<td></td>
<td>Pe(PSDU,SNR)=1−[1−Pb(TYPE,SNR)]8×PSDU.subscript𝑃𝑒PSDUSNR1superscriptdelimited-[]1subscript𝑃𝑏TYPESNR8PSDU\small P_{e}(\textrm{PSDU},\textrm{SNR})=1-\left[1-P_{b}(\textrm{TYPE},\textrm{SNR})\right]^{8\times\textrm{PSDU}}.</td>
<td></td>
<td>(35)</td>
</tr>
</tbody>
</table>
In the previous equations, PLCP is the sum of the lengths of the PLCP Preamble and the PLCP header, PSDU is the sum of the lengths of the payload size PL and the MAC header and $`\text{DATA} = {\text{PSDU} + \text{PLCP}}`$.
$`P_{b}\hspace{0pt}{(\text{DBPSK},\text{SNR})}`$ is the BER as a function of SNR for the lowest data transmit rate employing DBPSK modulation (or DQPSK if the basic rate 2Mbps is adopted). Note that the FER, $`P_{e}\hspace{0pt}{(\text{DATA},\text{SNR})}`$, implicitly depends on the modulation format used. Hence, for each supported rate, one curve for $`P_{e}\hspace{0pt}{(\text{DATA},\text{SNR})}`$ as a function of SNR can be generated. $`P_{b}\hspace{0pt}{(\text{TYPE},\text{SNR})}`$ is modulation dependent whereby the parameter TYPE can be any of the following $`\text{TYPE} \in {\{\text{DBPSK},\text{DQPSK},\text{CCK5.5},\text{CCK11}\}}`$
<sup>4</sup>
<sup>4</sup>4The acronyms are short for Differential Binary Phase Shift Keying, Differential Quadrature Phase Shift Keying and Complementary Code Keying with rate $`5.5`$ or $`11`$ Mbps, respectively.
.
## VII Simulation Results and Model Validation
TABLE III: Typical network parameters
<table>
<tbody>
<tr>
<th>MAC header</th>
<th>28 bytes</th>
<td>Propag. delay τpsubscript𝜏𝑝\tau_{p}</td>
<td>1 μs𝜇𝑠\mu s</td>
</tr>
<tr>
<th>PLCP Preamble</th>
<th>144 bit</th>
<td>PLCP Header</td>
<td>48 bit</td>
</tr>
<tr>
<th>PHY header</th>
<th>24 bytes</th>
<td>Slot time</td>
<td>20 μs𝜇𝑠\mu s</td>
</tr>
<tr>
<th>PLCP rate</th>
<th>1Mbps</th>
<td>W0</td>
<td>32</td>
</tr>
<tr>
<th>No. back-off stages, m</th>
<th>5</th>
<td>Wmax</td>
<td>1024</td>
</tr>
<tr>
<th>Payload size</th>
<th>1028 bytes</th>
<td>SIFS</td>
<td>10 μs𝜇𝑠\mu s</td>
</tr>
<tr>
<th>ACK</th>
<th>14 bytes</th>
<td>DIFS</td>
<td>50 μs𝜇𝑠\mu s</td>
</tr>
<tr>
<th>ACK timeout</th>
<th>364μs𝜇𝑠\mu s</th>
<td>EIFS</td>
<td>364 μs𝜇𝑠\mu s</td>
</tr>
</tbody>
</table>
This section focuses on simulation results for validating the theoretical models and derivations presented in the previous sections. We have developed a C++ simulator modeling both the DCF protocol details in 802.11b and the backoff procedures of a specific number of independent transmitting stations. The simulator considers an Infrastructure BSS (Basic Service Set) with an AP and a certain number of fixed stations which communicate only with the AP. For the sake of simplicity, inside each station there are only three fundamental working levels; traffic model generator, MAC and PHY layers. Traffic is generated following the exponential distribution for the packet interarrival times. Moreover, the MAC layer is managed by a state machine which follows the main directives specified in the standard \[[1](#bib.bib1)\], namely waiting times (DIFS, SIFS, EIFS), post-backoff, backoff, basic and RTS/CTS access modes. The typical MAC layer parameters for IEEE802.11b noted in Table [III](#S7.T3 "TABLE III ‣ VII Simulation Results and Model Validation ‣ On the Behavior of the Distributed Coordination Function of IEEE 802.11 with Multirate Capability under General Transmission Conditions") \[[1](#bib.bib1)\] have been used for performance validation.
For conciseness, in this paper we present a set of results related to the three scenarios described in the following.
The first investigated scenario considers $`N = 10`$ contending stations. Nine stations are randomly placed on the perimeter of a circle of radius $`R`$, while the AP is placed at the center of the transmission area. Upon employing equations ([30](#S6.E30 "In VI-A Signal Propagation ‣ VI Physical Layer Modeling ‣ On the Behavior of the Distributed Coordination Function of IEEE 802.11 with Multirate Capability under General Transmission Conditions"))-([32](#S6.E32 "In VI-A Signal Propagation ‣ VI Physical Layer Modeling ‣ On the Behavior of the Distributed Coordination Function of IEEE 802.11 with Multirate Capability under General Transmission Conditions")) with $`n_{p} = 4`$ (typical of heavily faded Rayleigh channel conditions), we have chosen a distance $`R = 20`$ m in such a way that the SNR per transmitted bit is above the minimum sensitivity specified in Table [I](#S6.T1 "TABLE I ‣ VI-A Signal Propagation ‣ VI Physical Layer Modeling ‣ On the Behavior of the Distributed Coordination Function of IEEE 802.11 with Multirate Capability under General Transmission Conditions"), relative to $`11`$ Mbps bit rate. Such stations are in saturated conditions and have a packet rate $`\lambda = 8`$ kpkt/s. The payload size, assumed to be common to all the transmitting stations, is $`{P\hspace{0pt}L} = 1028`$ bytes. The tenth station in the following is identified as the slow station and is placed at 4 different distances from the AP so that its transmission occurs with four different bit rates envisaged in the IEEE802.11b protocol.
The theoretical aggregate throughput achieved by the system in this scenario is depicted in Fig. [2](#S7.F2 "Figure 2 ‣ VII Simulation Results and Model Validation ‣ On the Behavior of the Distributed Coordination Function of IEEE 802.11 with Multirate Capability under General Transmission Conditions") as a function of the packet rate of the slow station. Curves in both subplots have been parameterized with respect to the bit rate of the slow station. Simulated points are denoted with cross-points over the respective theoretical curves. The upper curves refer to ideal channel conditions, i.e., PER=0, while the lower subplot represents a scenario in which the packets transmitted by all the stations are affected by a PER equal to $`8 \cdot 10^{- 2}`$, which is the worst-case situation with regards to the minimum sensitivity \[[1](#bib.bib1)\].
Some considerations are in order. Both subplots show that the aggregate throughput is significantly lower than 11 Mbps even though all the stations transmit at the highest bit rate (continuous curve). This is essentially due to overhead and control data transmitted at the basic rate $`R_{C}`$. Moreover, throughput reduces as the PAR $`\lambda_{s\hspace{0pt}l\hspace{0pt}o\hspace{0pt}w}`$ increases, reaching saturation values strongly influenced by the rate of the slowest station. This phenomenon is called performance anomaly of 802.11b, and it has been first observed in \[[15](#bib.bib15)\]. A comparative analysis of the set of curves depicted in both subplots reveals that a small throughput reduction is due to the presence of channel induced errors.
[Refer to caption](/html/0710.3955/assets/x2.png)
Figure 2: Theoretical and simulated throughput for the 2-way mechanism as a function of the packet rate $`\lambda_{s\hspace{0pt}l\hspace{0pt}o\hspace{0pt}w}`$ of the slow station, for four different bit rates shown in the legends, and different PER values. Simulated points are identified by cross-markers over the respective theoretical curves. Payload size is 1028 bytes for all the $`N = 10`$ contending stations, whereby 9 stations are saturated and transmit with a packet rate of 8 kpkt/s at the maximum bit rate of $`11`$ Mbps. In the upper subplot $`\text{PER} = 0`$, while in the lower one $`\text{PER} = {8 \times 10^{- 2}}`$.
The second investigated scenario is as follows. The four bit rates are equally distributed between $`8`$ contending stations ($`2`$ stations with each bit rate). The packet rate $`\lambda`$ is the same for all the contending stations and represents the independent variable against which the aggregate throughput in Fig. [3](#S7.F3 "Figure 3 ‣ VII Simulation Results and Model Validation ‣ On the Behavior of the Distributed Coordination Function of IEEE 802.11 with Multirate Capability under General Transmission Conditions") is drawn. Curves have been parameterized with respect to the PER identified in the legends, assumed to be equal for all the contending stations. PER equal to zero represents ideal channel transmission conditions, while PER$`= {8 \times 10^{- 2}}`$ is the PER specified in the standard \[[1](#bib.bib1)\], corresponding to the minimum receiver sensitivity.
In addition to noting the expected throughput reduction due to the packet error rate, notice that the throughput manifests a linear behavior for low values of the packet rates with a slope depending mainly on the packet size of the contending stations. This observation will be further analyzed in the next section where a linear model for the throughput will be developed in unsaturated conditions. Moreover, note that in unsaturated traffic conditions, the aggregate throughput is quite independent of the packet errors due to non ideal channel conditions.
[Refer to caption](/html/0710.3955/assets/x3.png)
Figure 3: Theoretical and simulated throughput for the 2-way mechanism as a function of the packet rate $`\lambda`$ of the 8 contending stations, organized so that the four bit rates are equally distributed between the involved stations. Simulated points are identified by cross-markers over the respective theoretical curves. Payload size is $`\text{PL} = 1028`$ bytes for all the contending stations.
The last investigated scenario considers three saturated stations transmitting over a Rayleigh fading channel. Two stations transmit at the maximum bit rate of 11 Mbps, since they are located at 5 m distance from the AP. The third station is assumed to move radially from the AP, and to switch between the four envisaged bit rates as its distance from the AP increases. The aggregate throughput is shown in Fig. [4](#S7.F4 "Figure 4 ‣ VII Simulation Results and Model Validation ‣ On the Behavior of the Distributed Coordination Function of IEEE 802.11 with Multirate Capability under General Transmission Conditions") as a function of the distance of the third station from the AP. While the theoretical throughput estimated with the proposed model is shown as a continuous curve, simulated results are noted with star-marked points. The radial distances from the AP at which rate switching occurs are noted with vertical lines in the same plot. Rate switching has been accomplished upon evaluating the minimum distance at which the per-station PER (according to the appropriate BER relation summarized in Table [II](#S6.T2 "TABLE II ‣ VI-B IEEE 802.11b Physical Layer ‣ VI Physical Layer Modeling ‣ On the Behavior of the Distributed Coordination Function of IEEE 802.11 with Multirate Capability under General Transmission Conditions") along with expressions ([33](#S6.E33 "In VI-B IEEE 802.11b Physical Layer ‣ VI Physical Layer Modeling ‣ On the Behavior of the Distributed Coordination Function of IEEE 802.11 with Multirate Capability under General Transmission Conditions"))-([35](#S6.E35 "In VI-B IEEE 802.11b Physical Layer ‣ VI Physical Layer Modeling ‣ On the Behavior of the Distributed Coordination Function of IEEE 802.11 with Multirate Capability under General Transmission Conditions"))) was above $`8 \times 10^{- 2}`$ for each employed bit rate, as specified in the standard \[[1](#bib.bib1), [2](#bib.bib2)\].
As already noted in the results presented in the previous plots, the aggregate throughput is strongly influenced by the bit rate of the slowest station. Up to about 26 m, the 3 stations transmit at the maximum bit rate and the aggregate throughput slightly decreases because of the worst BER conditions due to the SNR reduction of the moving station. As the distance increases and the bit rate of the moving stations decreases, the behavior changes. In fact, at higher bit rates, the larger the distance, the more the moving station tends to go through backoff stages because of the increasingly worse channel conditions. The channel inactivity of the moving station is in favor of the other two fastest stations and the aggregate throughput tends to slightly increase.
[Refer to caption](/html/0710.3955/assets/x4.png)
Figure 4: Theoretical (continuous curve) and simulated (star-marked points) throughput for the 2-way mechanism for three saturated stations as a function of the distance from the AP of the third station. Payload size is $`\text{PL} = 1028`$ bytes for all the contending stations.
## VIII A simple model of the aggregate throughput in non-saturated conditions
This section derives a simple model of the aggregate throughput in unsaturated conditions. To this end, consider the aggregate throughput in ([14](#S3.E14 "In III Markovian Process Analysis ‣ On the Behavior of the Distributed Coordination Function of IEEE 802.11 with Multirate Capability under General Transmission Conditions"))
<table>
<tbody>
<tr>
<td></td>
<td>S(τ(1),…,τ(N))=∑s=1N1TavPs(s)⋅(1−Pe(s))⋅PL𝑆superscript𝜏1…superscript𝜏𝑁superscriptsubscript𝑠1𝑁⋅1subscript𝑇𝑎𝑣superscriptsubscript𝑃𝑠𝑠1superscriptsubscript𝑃𝑒𝑠𝑃𝐿S\left(\tau^{(1)},\ldots,\tau^{(N)}\right)=\sum_{s=1}^{N}\frac{1}{T_{av}}P_{s}^{(s)}\cdot(1-P_{e}^{(s)})\cdot PL</td>
<td></td>
<td>(36)</td>
</tr>
</tbody>
</table>
whereby, we have emphasized the dependence of $`S\hspace{0pt}\left( \tau^{(1)},\ldots,\tau^{(N)} \right)`$ on the set of $`N`$ probabilities $`\tau^{(1)},\ldots,\tau^{(N)}`$, with $`\tau^{(s)}`$ defined in ([III](#S3.Ex4 "III Markovian Process Analysis ‣ On the Behavior of the Distributed Coordination Function of IEEE 802.11 with Multirate Capability under General Transmission Conditions")). Of course, probabilities $`\tau^{(1)},\ldots,\tau^{(N)}`$ depend on the traffic rates $`\lambda^{(1)},\ldots,\lambda^{(N)}`$ of the $`N`$ contending stations as exemplified by ([27](#S5.E27 "In V Traffic Model ‣ On the Behavior of the Distributed Coordination Function of IEEE 802.11 with Multirate Capability under General Transmission Conditions")) and ([28](#S5.E28 "In V Traffic Model ‣ On the Behavior of the Distributed Coordination Function of IEEE 802.11 with Multirate Capability under General Transmission Conditions")).
Let us find an expression for $`S\hspace{0pt}\left( \tau^{(1)},\ldots,\tau^{(N)} \right)`$ when the overall system approaches unsaturated traffic conditions. In this respect, let us find the expressions for $`P_{s}^{(s)}`$, $`P_{e}^{(s)}`$ and $`T_{a\hspace{0pt}v}`$ in the limit $`\overline{\tau}\rightarrow\overline{0}`$, whereby, the previous compact relation is used to signify the fact that all the probabilities $`\tau^{(1)},\ldots,\tau^{(N)}`$ approaches very small values.
First of all, notice that from ([27](#S5.E27 "In V Traffic Model ‣ On the Behavior of the Distributed Coordination Function of IEEE 802.11 with Multirate Capability under General Transmission Conditions")) and ([28](#S5.E28 "In V Traffic Model ‣ On the Behavior of the Distributed Coordination Function of IEEE 802.11 with Multirate Capability under General Transmission Conditions")), the following approximations easily follow:
<table>
<tbody>
<tr>
<td></td>
<td>q(s)≈λ(s)Tav,λ(s)→0,∀s=1,…,Nτ(s)≈λ(s)Tav1−Peq(s),λ(s)→0,∀s=1,…,Nsuperscript𝑞𝑠superscript𝜆𝑠subscript𝑇𝑎𝑣formulae-sequence→superscript𝜆𝑠0for-all𝑠1…𝑁superscript𝜏𝑠superscript𝜆𝑠subscript𝑇𝑎𝑣1superscriptsubscript𝑃𝑒𝑞𝑠formulae-sequence→superscript𝜆𝑠0for-all𝑠1…𝑁\begin{array}[]{ll}q^{(s)}\approx\lambda^{(s)}T_{av},&\lambda^{(s)}\rightarrow 0,~{}\forall s=1,\ldots,N\\ \tau^{(s)}\approx\frac{\lambda^{(s)}T_{av}}{1-P_{eq}^{(s)}},&\lambda^{(s)}\rightarrow 0,~{}\forall s=1,\ldots,N\end{array}</td>
<td></td>
<td>(37)</td>
</tr>
</tbody>
</table>
From ([13](#S3.E13 "In III Markovian Process Analysis ‣ On the Behavior of the Distributed Coordination Function of IEEE 802.11 with Multirate Capability under General Transmission Conditions")), it easily follows
<table>
<tbody>
<tr>
<td></td>
<td>limτ¯→0¯Pcol(s)=0,∀s=1,…,Nformulae-sequencesubscript→¯𝜏¯0superscriptsubscript𝑃𝑐𝑜𝑙𝑠0for-all𝑠1…𝑁\lim_{\overline{\tau}\rightarrow\overline{0}}P_{col}^{(s)}=0,~{}\forall s=1,\ldots,N</td>
<td></td>
</tr>
</tbody>
</table>
Given the relation $`P_{e\hspace{0pt}q}^{(s)} = {{P_{e}^{(s)} + P_{c\hspace{0pt}o\hspace{0pt}l}^{(s)}} - {P_{e}^{(s)}\hspace{0pt}P_{c\hspace{0pt}o\hspace{0pt}l}^{(s)}}}`$ in ([1](#S2.E1 "In II Markovian Model Characterizing the MAC Layer under general traffic conditions and Real Transmission Channel ‣ On the Behavior of the Distributed Coordination Function of IEEE 802.11 with Multirate Capability under General Transmission Conditions")), as $`\overline{\lambda}\rightarrow\overline{0}`$ (which is tantamount to considering $`\overline{\tau}\rightarrow\overline{0}`$ in light of the relation ([37](#S8.E37 "In VIII A simple model of the aggregate throughput in non-saturated conditions ‣ On the Behavior of the Distributed Coordination Function of IEEE 802.11 with Multirate Capability under General Transmission Conditions"))) we have:
<table>
<tbody>
<tr>
<td></td>
<td>Peq(s)→Pe(s),∀s=1,…,Nformulae-sequence→superscriptsubscript𝑃𝑒𝑞𝑠superscriptsubscript𝑃𝑒𝑠for-all𝑠1…𝑁P_{eq}^{(s)}\rightarrow P_{e}^{(s)},~{}\forall s=1,\ldots,N</td>
<td></td>
</tr>
</tbody>
</table>
From ([17](#S4.E17 "In IV-B The average slot duration of a successful transmission ‣ IV Estimating the Average Time Slot Duration ‣ On the Behavior of the Distributed Coordination Function of IEEE 802.11 with Multirate Capability under General Transmission Conditions")), as $`\overline{\tau}\rightarrow\overline{0}`$ the following relation holds,
[Refer to caption](/html/0710.3955/assets/x5.png)
Figure 5: Aggregate throughput for $`N = 2`$, 11Mbps contending stations over the bi-dimensional set $`{\lambda^{(1)} \in {{\lbrack 0,1000\rbrack}\hspace{0pt}\text{pkt/s}}},{\lambda^{(2)} \in {{\lbrack 0,1000\rbrack}\hspace{0pt}\text{pkt/s}}}`$, for $`{P\hspace{0pt}L} = 1028`$ bytes. The other transmission parameters are noted in Table [III](#S7.T3 "TABLE III ‣ VII Simulation Results and Model Validation ‣ On the Behavior of the Distributed Coordination Function of IEEE 802.11 with Multirate Capability under General Transmission Conditions").
<table>
<tbody>
<tr>
<td></td>
<td>Ps(s)≈τ(s)≈λ(s)Tav1−Pe(s),∀s=1,…,Nformulae-sequencesuperscriptsubscript𝑃𝑠𝑠superscript𝜏𝑠superscript𝜆𝑠subscript𝑇𝑎𝑣1superscriptsubscript𝑃𝑒𝑠for-all𝑠1…𝑁P_{s}^{(s)}\approx\tau^{(s)}\approx\frac{\lambda^{(s)}T_{av}}{1-P_{e}^{(s)}},~{}\forall s=1,\ldots,N</td>
<td></td>
</tr>
</tbody>
</table>
Upon employing the previous relations, ([36](#S8.E36 "In VIII A simple model of the aggregate throughput in non-saturated conditions ‣ On the Behavior of the Distributed Coordination Function of IEEE 802.11 with Multirate Capability under General Transmission Conditions")) can be rewritten as:
<table>
<tbody>
<tr>
<td></td>
<td>S(λ(1),…,λ(N))≈PL⋅∑s=1Nλ(s)𝑆superscript𝜆1…superscript𝜆𝑁⋅𝑃𝐿superscriptsubscript𝑠1𝑁superscript𝜆𝑠S\left(\lambda^{(1)},\ldots,\lambda^{(N)}\right)\approx PL\cdot\sum_{s=1}^{N}\lambda^{(s)}</td>
<td></td>
<td>(38)</td>
</tr>
</tbody>
</table>
which is valid under the hypothesis that the packet length $`P\hspace{0pt}L`$ is the same for all the $`N`$ contending stations. In case of different packet lengths, the aggregate throughput is the weighted sum of each station packet rate times the respective packet length.
Notice that in unsaturated conditions, the aggregate throughput is independent from the packet error rate affecting data transmission from each contending station in the network.
Equ. ([38](#S8.E38 "In VIII A simple model of the aggregate throughput in non-saturated conditions ‣ On the Behavior of the Distributed Coordination Function of IEEE 802.11 with Multirate Capability under General Transmission Conditions")) states that for very low values of $`{{{\lambda^{(s)},{\forall s}} = 1},{\ldots,N}},`$ the aggregate throughput behaves as a linear combination of the packet rates of the $`N`$ contending stations times the average payload length $`P\hspace{0pt}L`$. As a reference example, consider the aggregate throughput
<sup>5</sup>
<sup>5</sup>5The throughput has been plotted using ([14](#S3.E14 "In III Markovian Process Analysis ‣ On the Behavior of the Distributed Coordination Function of IEEE 802.11 with Multirate Capability under General Transmission Conditions")) evaluated by employing the solution of the non-linear system of $`2\hspace{0pt}N`$ equations discussed at the end of Section [III](#S3 "III Markovian Process Analysis ‣ On the Behavior of the Distributed Coordination Function of IEEE 802.11 with Multirate Capability under General Transmission Conditions").
shown in Fig. [5](#S8.F5 "Figure 5 ‣ VIII A simple model of the aggregate throughput in non-saturated conditions ‣ On the Behavior of the Distributed Coordination Function of IEEE 802.11 with Multirate Capability under General Transmission Conditions") for a scenario with $`N = 2`$ (11Mbps) contending stations. In this reference example, we have considered ideal channel conditions (i.e., $`{P_{e}^{(s)} = 0},{{\forall s} = {1,2}}`$) so that only collisions can yield packet losses. The other transmission parameters are noted in Table [III](#S7.T3 "TABLE III ‣ VII Simulation Results and Model Validation ‣ On the Behavior of the Distributed Coordination Function of IEEE 802.11 with Multirate Capability under General Transmission Conditions"). There are various points that are worth noting from this figure. When both stations approach saturated conditions, i.e., for $`\lambda^{(1)}`$ and $`\lambda^{(2)}`$ tending to infinity, the maximum aggregate throughput is less than half the maximum bit rate of each contending station due to overhead and control data transmitted at the basic rate $`R_{C}`$. On the other hand, when both $`\lambda^{(1)}`$ and $`\lambda^{(2)}`$ take on very small values, the aggregate throughput points define a plane predicted by ([38](#S8.E38 "In VIII A simple model of the aggregate throughput in non-saturated conditions ‣ On the Behavior of the Distributed Coordination Function of IEEE 802.11 with Multirate Capability under General Transmission Conditions")) in the region $`{(\lambda^{(1)},\lambda^{(2)})} \in D_{2}`$ defined as follows
<table>
<tbody>
<tr>
<td></td>
<td>D2={λ(1)∈[0,λc(1)2)0≤λ(2)≤−λc(2)λc(1)λ(1)+λc(2)2subscript𝐷2casessuperscript𝜆10superscriptsubscript𝜆𝑐12missing-subexpression0superscript𝜆2superscriptsubscript𝜆𝑐2superscriptsubscript𝜆𝑐1superscript𝜆1superscriptsubscript𝜆𝑐22missing-subexpressionD_{2}=\left\{\begin{array}[]{ll}\lambda^{(1)}\in\left[0,\frac{\lambda_{c}^{(1)}}{2}\right)&\\ 0\leq\lambda^{(2)}\leq-\frac{\lambda_{c}^{(2)}}{\lambda_{c}^{(1)}}\lambda^{(1)}+\frac{\lambda_{c}^{(2)}}{2}&\end{array}\right.</td>
<td></td>
<td>(39)</td>
</tr>
</tbody>
</table>
whereby, $`\lambda_{c}^{(1)}`$ and $`\lambda_{c}^{(2)}`$ marked on Fig. [5](#S8.F5 "Figure 5 ‣ VIII A simple model of the aggregate throughput in non-saturated conditions ‣ On the Behavior of the Distributed Coordination Function of IEEE 802.11 with Multirate Capability under General Transmission Conditions"), represents the limiting values of the packet rate above which each station approaches saturated traffic conditions
<sup>6</sup>
<sup>6</sup>6For the considerations that follow, let us suppose the values of $`\lambda_{c}^{(1)}`$ and $`\lambda_{c}^{(2)}`$ are known. Their values will actually be derived later.
. Region $`D_{2}`$ along with the line
<sup>7</sup>
<sup>7</sup>7This is the equation of the line passing through the two points $`(0,{\lambda_{c}^{(2)}/2})`$ and $`({\lambda_{c}^{(1)}/2},0)`$.
$`\lambda^{(2)} = {{- {\frac{\lambda_{c}^{(2)}}{\lambda_{c}^{(1)}}\hspace{0pt}\lambda^{(1)}}} + \frac{\lambda_{c}^{(2)}}{2}}`$ is identified as unsaturated operating area in Fig. [5](#S8.F5 "Figure 5 ‣ VIII A simple model of the aggregate throughput in non-saturated conditions ‣ On the Behavior of the Distributed Coordination Function of IEEE 802.11 with Multirate Capability under General Transmission Conditions").
Consider now the previous scenario with two 11Mbps contending stations transmitting at a fixed packet rate $`\lambda^{(1)} = \lambda^{(2)} = 50`$ pkt/s (unsaturated conditions for both stations), and a third station starts transmitting at a bit rate equal to 1 Mbps. As above, we have considered ideal channel conditions for the two 11 Mbps stations (i.e., $`{P_{e}^{(s)} = 0},{{\forall s} = {1,2}}`$) while the third station is assumed to experience three different PER values $`P_{e}^{(3)} = {0,10^{- 1},10^{- 2}}`$.
Fig. [6](#S8.F6 "Figure 6 ‣ VIII A simple model of the aggregate throughput in non-saturated conditions ‣ On the Behavior of the Distributed Coordination Function of IEEE 802.11 with Multirate Capability under General Transmission Conditions") shows the aggregate throughput as a function of the packet rate $`\lambda^{(3)}`$ of the third station. It can be observed that when $`\lambda^{(3)}`$ takes on low values, the overall system can be considered in unsaturated conditions, and therefore model ([38](#S8.E38 "In VIII A simple model of the aggregate throughput in non-saturated conditions ‣ On the Behavior of the Distributed Coordination Function of IEEE 802.11 with Multirate Capability under General Transmission Conditions")) holds. In this case, $`S\hspace{0pt}{(\lambda^{(1)},\lambda^{(2)},\lambda^{(3)})}`$ can be considered as a line with slope equal to the average packet length $`P\hspace{0pt}L`$, while $`{P\hspace{0pt}L\hspace{0pt}\left( {\lambda^{(1)} + \lambda^{(2)}} \right)} = {{P\hspace{0pt}L} \cdot 100}`$ pkt/s is the value of $`S\hspace{0pt}{(\lambda^{(1)},\lambda^{(2)},\lambda^{(3)})}`$ for $`\lambda^{(3)} = 0`$. So long as $`\lambda^{(3)} \in {\lbrack 0,{\lambda_{c}^{(3)}/2})}`$, model ([38](#S8.E38 "In VIII A simple model of the aggregate throughput in non-saturated conditions ‣ On the Behavior of the Distributed Coordination Function of IEEE 802.11 with Multirate Capability under General Transmission Conditions")) predicts the aggregate throughput quite well without the necessity of solving the non-linear system described at the end of Section [III](#S3 "III Markovian Process Analysis ‣ On the Behavior of the Distributed Coordination Function of IEEE 802.11 with Multirate Capability under General Transmission Conditions"). On the other hand, when $`\lambda^{(3)}`$ grows above $`\lambda_{c}^{(3)}/2`$, the system approaches saturated conditions quite fast and the overall throughput drops asymptotically to $`1.3`$Mbps, as shown in Fig. [6](#S8.F6 "Figure 6 ‣ VIII A simple model of the aggregate throughput in non-saturated conditions ‣ On the Behavior of the Distributed Coordination Function of IEEE 802.11 with Multirate Capability under General Transmission Conditions") for $`\lambda^{(3)}\rightarrow\infty`$. Notice that the aggregate throughput is strongly influenced by that of the slowest contending station (1 Mbps in this specific case), confirming the rate anomaly problem already noted in literature by resorting to simulations \[[15](#bib.bib15)\]. Also note that the PER experienced by the third station as already demonstrated theoretically above, does not affect the linear zone of the aggregate throughput, while it reduces the aggregate throughput obtained in saturated conditions. Also shown in the same figure is the throughput of the third station (when the first two stations are silent, i.e., $`\lambda^{(1)} = \lambda^{(2)} = 0`$ pkt/s) along with the linear model of the throughput valid for values of $`\lambda^{(3)}`$ lower than $`\lambda_{c}^{(3)}`$.
[Refer to caption](/html/0710.3955/assets/x6.png)
Figure 6: Theoretical and Simulated Throughput for a scenario with three contending stations with the following transmission parameters: the first two stations transmit with a bit rate of 11Mbps, while the third one is a 1Mbps station. Payload size is 1028 bytes for all stations. Packet rates are noted in the figure for the two scenarios investigated. Dash-dotted line refers to the linear model of the throughput derived in ([38](#S8.E38 "In VIII A simple model of the aggregate throughput in non-saturated conditions ‣ On the Behavior of the Distributed Coordination Function of IEEE 802.11 with Multirate Capability under General Transmission Conditions")).
For $`N`$ contending stations, the region $`D_{N}`$ is the volume contained in the region $`{(\lambda^{(1)},\ldots,\lambda^{(N)})} \in D_{N}`$ defined as
<table>
<tbody>
<tr>
<td></td>
<td>DN={λ(1)∈[0,λc(1)2)…λ(N−1)∈[0,λc(N−1)2)0≤λ(N)≤λc(N)2−λc(N)∑s=1N−1λ(s)λc(s)subscript𝐷𝑁casessuperscript𝜆10superscriptsubscript𝜆𝑐12missing-subexpression…missing-subexpressionsuperscript𝜆𝑁10superscriptsubscript𝜆𝑐𝑁12missing-subexpression0superscript𝜆𝑁superscriptsubscript𝜆𝑐𝑁2superscriptsubscript𝜆𝑐𝑁superscriptsubscript𝑠1𝑁1superscript𝜆𝑠superscriptsubscript𝜆𝑐𝑠missing-subexpressionD_{N}=\left\{\begin{array}[]{ll}\lambda^{(1)}\in\left[0,\frac{\lambda_{c}^{(1)}}{2}\right)&\\ \ldots&\\ \lambda^{(N-1)}\in\left[0,\frac{\lambda_{c}^{(N-1)}}{2}\right)&\\ 0\leq\lambda^{(N)}\leq\frac{\lambda_{c}^{(N)}}{2}-\lambda_{c}^{(N)}\sum_{s=1}^{N-1}\frac{\lambda^{(s)}}{\lambda_{c}^{(s)}}&\end{array}\right.</td>
<td></td>
<td>(40)</td>
</tr>
</tbody>
</table>
whereby, the last equation is the hyperplane passing through the $`N`$-dimensional points
<table>
<tbody>
<tr>
<td></td>
<td>(λc(1)/2,0,…,0),(0,λc(2)/2,…,0),…,(0,…,0,λc(N)/2)superscriptsubscript𝜆𝑐120…00superscriptsubscript𝜆𝑐22…0…0…0superscriptsubscript𝜆𝑐𝑁2(\lambda_{c}^{(1)}/2,0,\ldots,0),(0,\lambda_{c}^{(2)}/2,\ldots,0),\ldots,(0,\ldots,0,\lambda_{c}^{(N)}/2)</td>
<td></td>
</tr>
</tbody>
</table>
The natural question at this point is the determination of the critical packet rates $`{{{\lambda_{c}^{(s)},{\forall s}} = 1},{\ldots,N}},`$ needed in order to identify the interval of validity of the aggregate throughput in ([38](#S8.E38 "In VIII A simple model of the aggregate throughput in non-saturated conditions ‣ On the Behavior of the Distributed Coordination Function of IEEE 802.11 with Multirate Capability under General Transmission Conditions")). Consider the $`s`$-th station in the network. Such a station is in saturation when there is always a packet to be transmitted in its buffer. A simple threshold on $`\lambda^{(s)}`$ discerning unsaturated from saturated operation is $`\lambda_{c}^{(s)} = \frac{1}{T_{s\hspace{0pt}t}}`$, whereby $`T_{s\hspace{0pt}t}`$ is the service time of the $`s`$-th station. The reason is simple; the service time is the time interval from when a packet is taken from the queue head, to when it is successfully transmitted. If the packet inter-arrival time in the buffer is greater than $`T_{s\hspace{0pt}t}`$, then on the average, the buffer will contain at most one packet waiting for transmission. On the other hand, when the packet inter-arrival time is less than $`T_{s\hspace{0pt}t}`$, then packets will be buffered for successive transmissions. In this respect, $`T_{s\hspace{0pt}t}`$ is the equilibrium interval based on which on average a packet is transmitted soon after its arrival in the buffer.
The service time $`T_{s\hspace{0pt}t}`$ can be defined as follows \[[10](#bib.bib10), [19](#bib.bib19)\]:
<table>
<tbody>
<tr>
<td></td>
<td>Tst=TA+TTXsubscript𝑇𝑠𝑡subscript𝑇𝐴subscript𝑇𝑇𝑋T_{st}=T_{A}+T_{TX}</td>
<td></td>
</tr>
</tbody>
</table>
whereby, $`T_{A}`$ is the average time that a station spends through the various backoff stages before transmitting a packet, i.e., the so called MAC access time, while $`T_{T\hspace{0pt}X}`$ is the average duration of a transmission.
As far as $`T_{A}`$ is concerned, we note that despite the existence of $`m + 1`$ backoff stages (see Fig. [1](#S2.F1 "Figure 1 ‣ II Markovian Model Characterizing the MAC Layer under general traffic conditions and Real Transmission Channel ‣ On the Behavior of the Distributed Coordination Function of IEEE 802.11 with Multirate Capability under General Transmission Conditions")), in unsaturated conditions where the collisions between stations rarely occur, a packet is transmitted in the $`0`$-th backoff stage after an average number of backoff slots equal to $`W_{0}/2`$. Based on these considerations, the average MAC access time in unloaded traffic conditions can be simply defined as
<table>
<tbody>
<tr>
<td></td>
<td>TA=W02σsubscript𝑇𝐴subscript𝑊02𝜎T_{A}=\frac{W_{0}}{2}\sigma</td>
<td></td>
<td>(41)</td>
</tr>
</tbody>
</table>
whereby, $`\sigma`$ is the duration of an empty slot.
The average duration of a transmission, $`T_{T\hspace{0pt}X}`$, can be evaluated by observing that a station can experience two possible situations; successful transmission or collision. Upon neglecting the effects of collisions based on the considerations stated above in connection to $`T_{A}`$, $`T_{T\hspace{0pt}X}`$ can be defined as in ([](#S4.Ex5 "IV-B The average slot duration of a successful transmission ‣ IV Estimating the Average Time Slot Duration ‣ On the Behavior of the Distributed Coordination Function of IEEE 802.11 with Multirate Capability under General Transmission Conditions")
IV-B
); $`T_{T\hspace{0pt}X} = T_{s}^{(s)}`$.
The values of $`\lambda_{c}^{(r)}`$ derived under these hypotheses, are shown in Table [IV](#S8.T4 "TABLE IV ‣ VIII A simple model of the aggregate throughput in non-saturated conditions ‣ On the Behavior of the Distributed Coordination Function of IEEE 802.11 with Multirate Capability under General Transmission Conditions") as a function of the station transmission bit rate.
TABLE IV: Critical Packet Rates.
<table>
<tbody>
<tr>
<th>Station bit rate</th>
<td>1 Mbps</td>
<td>2 Mbps</td>
<td>5.5 Mbps</td>
<td>11 Mbps</td>
</tr>
<tr>
<th>λc(r)superscriptsubscript𝜆𝑐𝑟\lambda_{c}^{(r)}-[pkt/s]</th>
<td>107.4</td>
<td>196.5</td>
<td>416.3</td>
<td>612.0</td>
</tr>
</tbody>
</table>
## IX Conclusions
In this paper, we have presented a multi-dimensional Markovian state transition model characterizing the DCF behavior at the MAC layer of the IEEE802.11 series of standards by accounting for channel induced errors and multirate transmission typical of fading environments, under both non-saturated and saturated traffic conditions. The model provided in the paper allows taking into consideration the impact of channel contention in throughput analysis which is often not considered or is considered with a static model by using a mean contention period. Subsequently, based on justifiable assumptions, the stationary probability of the Markov chain is calculated obtaining the throughput in both non-saturated and saturated conditions.
Finally, we have shown that the behavior of the aggregate throughput in non-saturated traffic conditions is a linear combination of the payload size along with the packet rates of each contending station in the network. This result is significant for system level optimization of the network. Theoretical derivations were supported by simulations.
## References
- \[1\]
IEEE Standard for Wireless LAN Medium Access Control (MAC) and Physical Layer (PHY) Specifications, November 1997, P802.11
- \[2\]
IEEE Std 802.15.2-2003; Part 15.2: Coexistence of Wireless Personal Area Networks with Other Wireless Devices Operating in Unlicensed Frequency Bands, August 2003.
- \[3\]
G. Bianchi, “Performance analysis of the IEEE 802.11 distributed coordination function”, IEEE JSAC, Vol.18, No.3, March 2000.
- \[4\]
Ha Cheol Lee, “Impact of bit errors on the DCF throughput in wireless LAN over ricean fading channels”, In Proc. of IEEE ICDT ’06, 2006.
- \[5\]
Q. Ni, T. Li, T. Turletti, and Y. Xiao, “Saturation throughput analysis of error-prone 802.11 wireless networks”, Wiley Journal of Wireless Communications and Mobile Computing, Vol. 5, No. 8, pp. 945-956, Dec. 2005.
- \[6\]
P. Chatzimisios, A.C. Boucouvalas, and V. Vitsas, “Influence of channel BER on IEEE 802.11 DCF”, IEE Electronics Letters, Vol.39, No.23, pp.1687-1689, Nov. 2003.
- \[7\]
M. Zorzi and R.R. Rao, “Capture and retransmission control in mobile radio”, IEEE JSAC, Vol.12, No.8, pp.1289 - 1298, Oct. 1994.
- \[8\]
Jae Hyun Kim and Jong Kyu Lee “Capture effects of wireless CSMA/CA protocols in Rayleigh and shadow fading channels”, IEEE Trans. on Veh. Tech., Vol.48, No.4, pp.1277-1286, July 1999.
- \[9\]
Z. Hadzi-Velkov and B. Spasenovski, “Capture effect in IEEE 802.11 basic service area under influence of Rayleigh fading and near/far effect”, In Proc. of 13th IEEE International Symposium on Personal, Indoor and Mobile Radio Communications, Vol.1, pp.172 - 176, Sept. 2002.
- \[10\]
L. Yong Shyang, A. Dadej, and A.Jayasuriya, “Performance analysis of IEEE 802.11 DCF under limited load”, In Proc. of Asia-Pacific Conference on Communications, Vol.1, pp.759 - 763, 03-05 Oct. 2005.
- \[11\]
D. Malone, K. Duffy, and D.J. Leith, “Modeling the 802.11 distributed coordination function in non-saturated heterogeneous conditions”, IEEE-ACM Trans. on Networking, vol. 15, No. 1, pp. 159 172, Feb. 2007.
- \[12\]
F. Daneshgaran, M. Laddomada, F. Mesiti, and M. Mondin, “Unsaturated Throughput Analysis of IEEE 802.11 in Presence of Non Ideal Transmission Channel and Capture Effects,” to appear on IEEE Trans. on Wireless Communications, 2008.
- \[13\]
F. Daneshgaran, M. Laddomada, F. Mesiti, and M. Mondin, “A Model of the IEEE 802.11 DCF in presence of non ideal transmission channel and capture effects,” In Proc. of IEEE Globecom 07, Washington DC, November 2007.
- \[14\]
D. Qiao, S. Choi, and K.G. Shin, “Goodput analysis and link adaptation for IEEE 802.11a wireless LANs”, IEEE Trans. On Mobile Computing, Vol.1, No.4, Oct.-Dec. 2002.
- \[15\]
Heusse M.,Rousseau F., Berger-Sabbatel G. and Duda A., “Performance anomaly of 802.11b” In Proc. of IEEE INFOCOM 2003, pp. 836-843.
- \[16\]
T. Joshi, A. Mukherjee, and D.P. Agrawal, “Analytical modeling of the link delay characteristics for IEEE 802.11 DCF multi-rate WLANs,” In Proc. of IEEE CCECE-CCGEI, pp. 2164-2167, May 2006.
- \[17\]
D.-Y. Yang, T.-J. Lee, K. Jang, J.-B. Chang, and S. Choi, “Performance enhancement of multirate IEEE 802.11 WLANs with geographically scattered stations,” IEEE Trans. on Mobile Computing, vol.5, no.7, pp.906-919, July 2006.
- \[18\]
G.R. Cantieni, Q. Ni, C. Barakat, and T. Turletti, “Performance analysis under finite load and improvements for multirate 802.11,” Computer Communications, Elsevier, vol.28, pp.1095-1109, 2005.
- \[19\]
W. Lee, C. Wang, and K. Sohraby, “On use of traditional M/G/1 model for IEEE 802.11 DCF in unsaturated traffic conditions”, In Proc. of IEEE WCNC, 2006, pp. 1933 1937.
- \[20\]
G. Bolch, S. Greiner, H. de Meer, and K.S. Trivedi, Queueing Networks and Markov Chains, Wiley-Interscience, 2nd edition, 2006.
- \[21\]
M.K. Simon and M. Alouini, Digital Communication over Fading Channels: A Unified Approach to Performance Analysis, Wiley-Interscience, 1st edition, 2000.
- \[22\]
M. Fainberg, A Performance analysis of the IEEE 802.11b local area network in the presence of bluetooth personal area network, Available at `http://eeweb.poly.edu/dgoodman/fainberg.pdf`.
- \[23\]
T. S. Rappaport, Wireless Communications, Principles and Practice, Prentice-Hall, 2nd edition, USA, 2002.
- \[24\] Wu Xiuchao, “Simulate 802.11b Channel within NS-2”, available online at `http://www.comp.nus.edu.sg/~wuxiucha/research/` `reactive/report/80211ChannelinNS2_new.pdf` |
0710.2736 | cs.NI | cyber | $`L_{2}`$ norm performance index of synchronization and optimal control synthesis of complex networks
<sup>\*</sup>
<sup>\*</sup>\*This work is supported by both the National Science Foundation of China under grants 60674093, 60334030 and the City University of Hong Kong under the Research Enhancement Scheme and SRG grant 7002134.
Chao Liu $`^{\text{a}}`$
<sup>†</sup>
<sup>†</sup>† Corresponding author: chaoliu@pku.edu.cn
, Zhisheng Duan $`^{\text{a}}`$, Guanrong Chen $`^{\text{b}}`$, Lin Huang $`^{\text{a}}`$
$`^{\text{a}}`$ State Key Laboratory for Turbulence and Complex Systems, Department of Mechanics and
Aerospace Engineering, College of Engineering, Peking University, Beijing 100871, P. R. China
$`^{\text{b}}`$ Department of Electronic Engineering, City University of Hong Kong, Hong Kong, SAR, P. R. China
Abstract.
In this paper, the synchronizability problem of dynamical networks is addressed, where better synchronizability means that the network synchronizes faster with lower-overshoot. The $`L_{2}`$ norm of the error vector $`e`$ is taken as a performance index to measure this kind of synchronizability. For the equilibrium synchronization case, it is shown that there is a close relationship between the $`L_{2}`$ norm of the error vector $`e`$ and the $`H_{2}`$ norm of the transfer function $`G`$ of the linearized network about the equilibrium point. Consequently, the effect of the network coupling topology on the $`H_{2}`$ norm of the transfer function $`G`$ is analyzed. Finally, an optimal controller is designed, according to the so-called LQR problem in modern control theory, which can drive the whole network to its equilibrium point and meanwhile minimize the $`L_{2}`$ norm of the output of the linearized network.
Keywords. network, synchronizability, $`L_{2}`$ norm, $`H_{2}`$ norm, LQR problem, optimal control.
## 1 Introduction
The topic of synchronizability of dynamical networks has attracted increasing interest recently (see \[[1](#bib.bib1), [3](#bib.bib3), [6](#bib.bib6), [10](#bib.bib10), [11](#bib.bib11), [12](#bib.bib12), [14](#bib.bib14)\] and references therein), which is mostly referred to as how easy it is for the network to synchronize. Technically, this is a problem whether there is a wide range of coupling strength in which the synchronization is stable, and the wider the range, the better the synchronizability of the network. Accordingly, for some cases when the synchronized region is unbounded, a better synchronizability means that the synchronization can be achieved with a smaller coupling strength \[[10](#bib.bib10)\]. Generally, the synchronizability of a network depends on the underlying coupling topology of the network. Studies show that the ability for a network to synchronize is related to the spectral properties of the outer coupling matrix of the network \[[10](#bib.bib10), [11](#bib.bib11)\] and thus influenced by the structural properties of the network, such as average distance \[[14](#bib.bib14)\], degree homogeneity \[[6](#bib.bib6)\], clustering coefficient \[[12](#bib.bib12)\], degree correlation \[[1](#bib.bib1)\], betweenness centrality \[[3](#bib.bib3)\], etc.
In this paper, the synchronizability problem of dynamical networks is considered from a different point of view: suppose that the coupling strength of a network belongs to the range in which the synchronization is stable; then, how fast the synchronization will be achieved? This problem is important for the reason that in practical engineering implementation (such as communications via chaotic synchronization), synchronization is expected to be achieved not only easily but also swiftly. In \[[7](#bib.bib7)\], it is qualitatively pointed out that networks with diagonalizable outer coupling matrices may synchronize faster than the ones with non-diagonalizable outer coupling matrices. Clearly, a measure is needed for a quantitative description of the swiftness of network synchronization. To meet this objective, the $`L_{2}`$ norm of the error vector $`e`$, denoted as $`{\| e\|}_{2}`$, is taken in this paper as a performance index of this kind of network synchronizability. As will be seen later, the quantity $`{\| e\|}_{2}`$ presents a suitable measure of both swiftness and overshoot (referring to the largest difference among various node dynamics before the synchronization is achieved): the smaller the quantity $`{\| e\|}_{2}`$, the faster with smaller overshoot the network synchronization.
Furthermore, as shown by the numerical examples given below, the quantity $`{\| e\|}_{2}`$ is influenced by the coupling topology of the network. Thus, the investigation on the relationship between $`{\| e\|}_{2}`$ and the network structure is of significance. In this paper, for the case that the synchronous state is an equilibrium point, it is pointed out that $`{\| e\|}_{2}`$ is upper-bounded by the product of the vector 2-norm of the initial error vector $`e_{0}`$ and the $`H_{2}`$ norm of the transfer function $`G\hspace{0pt}{(s)}`$, denoted as $`{\|{G\hspace{0pt}{(s)}}\|}_{2}`$ or simply $`{\| G\|}_{2}`$, of the linearized network about the equilibrium point. Thus, the smaller the $`{\| G\|}_{2}`$, the smaller the $`{\| e\|}_{2}`$ as well. As pointed out in \[[2](#bib.bib2)\], the relationship between $`{\|{G\hspace{0pt}{(s)}}\|}_{2}`$ and the network structure is quite complicated. Under some assumptions, it is proved in this paper (see Theorem 1 and Example 4) that $`{\| G\|}_{2}`$ will not increase as the real eigenvalues of the symmetrical outer coupling matrix increase.
For a linear time-invariant system, the linear quadratic regulator problem, or simply the LQR problem, is a classical problem in modern control theory. The objective of the LQR problem is to find an optimal control law $`u\hspace{0pt}{(t)}`$ such that the state $`x\hspace{0pt}{(t)}`$ is driven into a (small) neighborhood of the origin while minimizing a quadratic performance ($`L_{2}`$ performance) index on $`u`$ and $`x`$. In fact, the LQR problem is posed traditionally as the minimization problem of the $`L_{2}`$ norm of the regulator output of the system. In this paper, based on the techniques of the LQR problem, an optimal controller design is developed so as to drive the network dynamics onto some homogenous stationary states while minimizing the $`L_{2}`$ norm of the output of the linearized network.
The rest of the paper is organized as follows. In Section 2, some preliminary definitions and lemmas necessary for successive development are presented. In Section 3, some numerical examples are provided to illustrate that the quantity $`{\| e\|}_{2}`$ presents a suitable measure of both swiftness and overshoot of the network synchronization. For the equilibrium synchronization case, the relationship between $`{\| e\|}_{2}`$ and the network structure is investigated in Section 4. Based on the results of the LQR problem, an LQR optimal controller is proposed in Section 5. The paper is concluded by the last Section.
## 2 Preliminaries
$`L_{2}\hspace{0pt}{\lbrack a,b\rbrack}`$ is an infinite-dimensional Hilbert space, which consists of all square-integrable and Lebesgue measurable functions defined on an interval $`\lbrack a,b\rbrack`$ with the scalar inner product
<table>
<tbody>
<tr>
<td></td>
<td>⟨f,g⟩=∫abf(t)∗g(t)𝑑t,𝑓𝑔superscriptsubscript𝑎𝑏𝑓superscript𝑡𝑔𝑡differential-d𝑡\langle f,g\rangle=\int_{a}^{b}f(t)^{*}g(t)dt,</td>
<td></td>
</tr>
</tbody>
</table>
while if the functions are vector or matrix-valued, the inner product is defined as
<table>
<tbody>
<tr>
<td></td>
<td>⟨f,g⟩=∫abtrace[f(t)∗g(t)]𝑑t,𝑓𝑔superscriptsubscript𝑎𝑏tracedelimited-[]𝑓superscript𝑡𝑔𝑡differential-d𝑡\langle f,g\rangle=\int_{a}^{b}\textrm{trace}[f(t)^{*}g(t)]dt,</td>
<td></td>
</tr>
</tbody>
</table>
where $`\ast`$ denotes complex conjugate transpose, and the induced norm is defined as
<table>
<tbody>
<tr>
<td></td>
<td>‖f‖2=⟨f,f⟩,subscriptnorm𝑓2𝑓𝑓\|f\|_{2}=\sqrt{\langle f,f\rangle},</td>
<td></td>
<td>(1)</td>
</tr>
</tbody>
</table>
for $`{f,g} \in {L_{2}\hspace{0pt}{\lbrack a,b\rbrack}}`$.
Consider a continuous-time linear system,
<table>
<tbody>
<tr>
<td></td>
<td>x˙=Ax+Bu,y=Cx,formulae-sequence˙𝑥𝐴𝑥𝐵𝑢𝑦𝐶𝑥\dot{x}=Ax+Bu,\quad y=Cx,</td>
<td></td>
<td>(2)</td>
</tr>
</tbody>
</table>
where $`x`$, $`u`$ and $`y`$ are the state, input and output of the system, respectively, and $`A \in {\mathbb{R}}^{n \times n}`$, $`B \in {\mathbb{R}}^{n \times m}`$ and $`C \in {\mathbb{R}}^{l \times n}`$ are given constant matrices. The transfer function from $`u`$ to $`y`$ is $`{G\hspace{0pt}{(s)}} = {C\hspace{0pt}{({{s\hspace{0pt}I} - A})}^{- 1}\hspace{0pt}B}`$. If $`A`$ is stable, then the $`H_{2}`$ norm of system ([2](#S2.E2 "In 2 Preliminaries")) is represented by the $`H_{2}`$ norm of the transfer function $`G\hspace{0pt}{(s)}`$, which is defined by
<table>
<tbody>
<tr>
<td></td>
<td>‖G(s)‖2=12π∫−∞+∞trace{G∗(jω)G(jω)}𝑑ω.subscriptnorm𝐺𝑠212𝜋superscriptsubscripttracesuperscript𝐺𝑗𝜔𝐺𝑗𝜔differential-d𝜔\|G(s)\|_{2}=\sqrt{\frac{1}{2\pi}\int_{-\infty}^{+\infty}\textrm{trace}\{G^{*}(j\omega)G(j\omega)\}d\omega}.</td>
<td></td>
</tr>
</tbody>
</table>
It can be proved that $`{\|{G\hspace{0pt}{(s)}}\|}_{2}^{2}`$ equals the overall output energy of the system response to the impulse input. For computing $`\|{G\hspace{0pt}{(s)}}\|`$, the following formula is convenient.
Lemma 1 \[[13](#bib.bib13)\]: If $`A`$ is stable, then $`{\|{G\hspace{0pt}{(s)}}\|}_{2}^{2} = {\text{trace}\hspace{0pt}{({B^{T}\hspace{0pt}Y\hspace{0pt}B})}}`$, where matrix $`Y`$ is the solution to the following Lyapunov equation:
<table>
<tbody>
<tr>
<td></td>
<td>YA+ATY+CTC=0.𝑌𝐴superscript𝐴𝑇𝑌superscript𝐶𝑇𝐶0YA+A^{T}Y+C^{T}C=0.</td>
<td></td>
<td>(3)</td>
</tr>
</tbody>
</table>
Equivalently,
<table>
<tbody>
<tr>
<td></td>
<td>‖G(s)‖22=infP>0{trace(BTPB):PA+ATP+CTC≤0}.superscriptsubscriptnorm𝐺𝑠22subscriptinfimum𝑃0conditional-settracesuperscript𝐵𝑇𝑃𝐵𝑃𝐴superscript𝐴𝑇𝑃superscript𝐶𝑇𝐶0\|G(s)\|_{2}^{2}=\inf\limits_{P>0}\left\{\textrm{trace}(B^{T}PB):\;PA+A^{T}P+C^{T}C\leq 0\right\}.</td>
<td></td>
<td>(4)</td>
</tr>
</tbody>
</table>
The so-called linear quadratic regulator (LQR) problem is an optimal control problem with a quadratic performance ($`H_{2}`$ norm) criterion. For the linear time-invariant system
<table>
<tbody>
<tr>
<td></td>
<td>x˙=Ax+Bu,x(t0)=x0,formulae-sequence˙𝑥𝐴𝑥𝐵𝑢𝑥subscript𝑡0subscript𝑥0\dot{x}=Ax+Bu,\;x(t_{0})=x_{0},</td>
<td></td>
</tr>
</tbody>
</table>
where $`x_{0}`$ is arbitrarily given, the regulator problem refers to finding a control function $`u\hspace{0pt}{(t)}`$ defined on $`\lbrack t_{0},T\rbrack`$, which can be a function of the state $`x\hspace{0pt}{(t)}`$, such that the state $`x\hspace{0pt}{(t)}`$ is driven into a (small) neighborhood of origin at time $`T`$, $`T < \infty`$. Since every physical system has energy limitation, and large control action (even if it is realizable) can easily drive the system out of its valid region, certain limitations have to be imposed on the control in practical engineering implementation. For these reasons, the regulator problem is usually posed as an optimal control problem with a certain combined performance index on $`u`$ and $`x`$. Focusing on the infinite time regulator problem (i.e., $`T\rightarrow\infty`$) and without loss of generality assuming $`t_{0} = 0`$, the LQR problem is formulated as follows: Find a control $`u\hspace{0pt}{(t)}`$ defined on $`\lbrack 0,\infty)`$ such that the state $`x\hspace{0pt}{(t)}`$ is driven to the origin at $`t\rightarrow\infty`$ and the following performance index is minimized:
<table>
<tbody>
<tr>
<td></td>
<td>minu∫0∞[x(t)u(t)]∗[QSS∗R][x(t)u(t)]𝑑tsubscript𝑢superscriptsubscript0superscriptdelimited-[]𝑥𝑡𝑢𝑡delimited-[]𝑄𝑆superscript𝑆𝑅delimited-[]𝑥𝑡𝑢𝑡differential-d𝑡\min\limits_{u}\int_{0}^{\infty}\left[\begin{array}[]{c}x(t)\\ u(t)\end{array}\right]^{*}\left[\begin{array}[]{cc}Q&S\\ S^{*}&R\end{array}\right]\left[\begin{array}[]{c}x(t)\\ u(t)\end{array}\right]dt</td>
<td></td>
<td>(5)</td>
</tr>
</tbody>
</table>
for some $`Q = {Q^{\ast},S}`$ and $`R = R^{\ast} > 0`$. Here, $`R > 0`$ emphasizes that the control energy has to be finite, i.e., $`{u\hspace{0pt}{(t)}} \in {L_{2}\hspace{0pt}{\lbrack 0,\infty)}}`$. Moreover, it is assumed that
<table>
<tbody>
<tr>
<td></td>
<td>[QSS∗R]≥0,R>0.formulae-sequencedelimited-[]𝑄𝑆superscript𝑆𝑅0𝑅0\left[\begin{array}[]{cc}Q&S\\ S^{*}&R\end{array}\right]\geq 0,\;R>0.</td>
<td></td>
<td>(6)</td>
</tr>
</tbody>
</table>
Then, ([6](#S2.E6 "In 2 Preliminaries")) can be factored as
<table>
<tbody>
<tr>
<td></td>
<td>[QSS∗R]=[C∗D∗][CD]delimited-[]𝑄𝑆superscript𝑆𝑅delimited-[]superscript𝐶superscript𝐷delimited-[]𝐶𝐷\left[\begin{array}[]{cc}Q&S\\ S^{*}&R\end{array}\right]=\left[\begin{array}[]{c}C^{*}\\ D^{*}\end{array}\right]\left[\begin{array}[]{cc}C&D\end{array}\right]</td>
<td></td>
</tr>
</tbody>
</table>
and ([5](#S2.E5 "In 2 Preliminaries")) can be written as
<table>
<tbody>
<tr>
<td></td>
<td>minu(t)∈L2[0,∞)‖Cx+Du‖22.subscript𝑢𝑡subscript𝐿20superscriptsubscriptnorm𝐶𝑥𝐷𝑢22\min\limits_{u(t)\in L_{2}[0,\;\infty)}\|Cx+Du\|_{2}^{2}.</td>
<td></td>
</tr>
</tbody>
</table>
Traditionally, the LQR problem is posed as the following minimization problem:
<table>
<tbody>
<tr>
<td></td>
<td>minu(t)∈L2[0,∞)‖y‖22subject to:{x˙=Ax+Bu,x(t0)=x0y=Cx+Du.subscript𝑢𝑡subscript𝐿20superscriptsubscriptnorm𝑦22subject to:casesformulae-sequence˙𝑥𝐴𝑥𝐵𝑢𝑥subscript𝑡0subscript𝑥0𝑦𝐶𝑥𝐷𝑢\begin{array}[]{l}\min\limits_{u(t)\in L_{2}[0,\;\infty)}\|y\|_{2}^{2}\\ \textrm{subject to:}\left\{\begin{array}[]{l}\dot{x}=Ax+Bu,\;x(t_{0})=x_{0}\\ y=Cx+Du.\end{array}\right.\end{array}</td>
<td></td>
<td>(7)</td>
</tr>
</tbody>
</table>
For the above LQR problem, the following lemma is useful.
Lemma 2 \[[13](#bib.bib13)\]: Suppose that in ([7](#S2.E7 "In 2 Preliminaries")):
(A1) $`(A,B)`$ is stabilizable;
(A2) $`D`$ has full column rank with $`\lbrack{D\hspace{0pt}D_{\bot}}\rbrack`$ being unitary;
(A3) $`(C,A)`$ is detectable;
(A4) $`\begin{bmatrix}
{A - {j\hspace{0pt}\omega\hspace{0pt}I}} & B \\
C & D
\end{bmatrix}`$ has full column rank for all $`\omega`$.
Then, there exists a unique optimal control $`u = {F\hspace{0pt}x}`$ for the LQR problem ([7](#S2.E7 "In 2 Preliminaries")), where
<table>
<tbody>
<tr>
<td></td>
<td>F=−(B∗X+DC)𝐹superscript𝐵𝑋𝐷𝐶F=-(B^{*}X+DC)</td>
<td></td>
<td>(8)</td>
</tr>
</tbody>
</table>
and $`X`$ is the stabilizing solution to the following Riccati equation:
<table>
<tbody>
<tr>
<td></td>
<td>(A−BD∗C)∗X+X(A−BD∗C)−XBB∗X+C∗D⊥D⊥∗C=0.superscript𝐴𝐵superscript𝐷𝐶𝑋𝑋𝐴𝐵superscript𝐷𝐶𝑋𝐵superscript𝐵𝑋superscript𝐶subscript𝐷bottomsuperscriptsubscript𝐷bottom𝐶0(A-BD^{*}C)^{*}X+X(A-BD^{*}C)-XBB^{*}X+C^{*}D_{\bot}D_{\bot}^{*}C=0.</td>
<td></td>
<td>(9)</td>
</tr>
</tbody>
</table>
Moreover, the minimized $`L_{2}`$ norm of the output $`y\hspace{0pt}{(t)}`$ is given by
<table>
<tbody>
<tr>
<td></td>
<td>minu(t)∈L2[0,∞)‖y‖22=‖Gcx0‖22,subscript𝑢𝑡subscript𝐿20superscriptsubscriptnorm𝑦22superscriptsubscriptnormsubscript𝐺𝑐subscript𝑥022\min\limits_{u(t)\in L_{2}[0,\;\infty)}\|y\|_{2}^{2}=\|G_{c}x_{0}\|_{2}^{2},</td>
<td></td>
<td>(10)</td>
</tr>
</tbody>
</table>
where $`G_{c}`$ is the transfer function of the system
<table>
<tbody>
<tr>
<td></td>
<td>{x˙=(A+BF)x+Ix0δ(t),x(0−)=0y=(C+DF)x,casesformulae-sequence˙𝑥𝐴𝐵𝐹𝑥𝐼subscript𝑥0𝛿𝑡𝑥subscript00𝑦𝐶𝐷𝐹𝑥\left\{\begin{array}[]{l}\dot{x}=(A+BF)x+Ix_{0}\delta(t),\;x(0_{-})=0\\ y=(C+DF)x,\end{array}\right.</td>
<td></td>
</tr>
</tbody>
</table>
with $`I`$ being the identity matrix and $`\delta\hspace{0pt}{(t)}`$ the impulse function.
## 3 $`L_{2}`$ norm performance index of network synchronization
Consider a network of $`N`$ identical dynamical nodes, described by
<table>
<tbody>
<tr>
<td></td>
<td>x˙i=f(xi)−σ∑j=1NmijΓxj,i=1,2,⋯,N,formulae-sequencesubscript˙𝑥𝑖𝑓subscript𝑥𝑖𝜎superscriptsubscript𝑗1𝑁subscript𝑚𝑖𝑗Γsubscript𝑥𝑗𝑖12⋯𝑁\dot{x}_{i}=f(x_{i})-\sigma\sum_{j=1}^{N}m_{ij}\Gamma x_{j},\;i=1,2,\cdots,N,</td>
<td></td>
<td>(11)</td>
</tr>
</tbody>
</table>
where $`f\hspace{0pt}{(x)}`$ governs the dynamics of each individual node, $`\sigma`$ is the coupling strength, $`\Gamma`$ is the inner linking matrix, and $`M = \left\lbrack m_{i\hspace{0pt}j} \right\rbrack`$ is the outer coupling matrix.
Let
<table>
<tbody>
<tr>
<td></td>
<td>ei(t)=xi(t)−x1(t),i=1,⋯,N,formulae-sequencesubscript𝑒𝑖𝑡subscript𝑥𝑖𝑡subscript𝑥1𝑡𝑖1⋯𝑁e_{i}(t)=x_{i}(t)-x_{1}(t),\;i=1,\cdots,N,</td>
<td></td>
<td>(12)</td>
</tr>
</tbody>
</table>
and
<table>
<tbody>
<tr>
<td></td>
<td>e(t)=[e1Te2T⋯eNT]T𝑒𝑡superscriptdelimited-[]superscriptsubscript𝑒1𝑇superscriptsubscript𝑒2𝑇⋯superscriptsubscript𝑒𝑁𝑇𝑇e(t)=[e_{1}^{T}\;e_{2}^{T}\;\cdots\;e_{N}^{T}]^{T}</td>
<td></td>
<td>(13)</td>
</tr>
</tbody>
</table>
denote the error vector of network ([11](#S3.E11 "In 3 𝐿₂ norm performance index of network synchronization")). Then, network ([11](#S3.E11 "In 3 𝐿₂ norm performance index of network synchronization")) is said to achieve (asymptotical) synchronization if
<table>
<tbody>
<tr>
<td></td>
<td>e(t)→0,ast→∞.formulae-sequence→𝑒𝑡0→as𝑡e(t)\rightarrow 0,\;\textrm{as}\;t\rightarrow\infty.</td>
<td></td>
</tr>
</tbody>
</table>
Let $`{\|{e\hspace{0pt}{(t)}}\|}_{2}\hspace{0pt}{({\lbrack 0,T\rbrack})}`$, or simply $`{\|{e\hspace{0pt}{(t)}}\|}_{2}`$ when no confusion may be caused, denote the $`L_{2}`$ norm of $`e\hspace{0pt}{(t)}`$ on a given interval $`\lbrack 0,T\rbrack`$. Then, according to ([1](#S2.E1 "In 2 Preliminaries")),
<table>
<tbody>
<tr>
<td></td>
<td>‖e(t)‖2=⟨e,e⟩=∫0Ttrace[e(t)∗e(t)]𝑑t.subscriptnorm𝑒𝑡2𝑒𝑒superscriptsubscript0𝑇tracedelimited-[]𝑒superscript𝑡𝑒𝑡differential-d𝑡\|e(t)\|_{2}=\sqrt{\langle e,e\rangle}=\sqrt{\int_{0}^{T}\textrm{trace}[e(t)^{*}e(t)]dt}.</td>
<td></td>
<td>(14)</td>
</tr>
</tbody>
</table>
In fact, $`{\|{e\hspace{0pt}{(t)}}\|}_{2}`$ represents the energy of the error signal $`e`$ on the interval $`\lbrack 0,T\rbrack`$, which is in proportion to the area between the error function $`e`$ and the time axis. Hence, $`{\|{e\hspace{0pt}{(t)}}\|}_{2}`$ can be used as a quantitative measure of the swiftness and overshoot of the network synchronization. In what follows, two examples are first given for illustration.
Example 1: Suppose that each single node in network ([11](#S3.E11 "In 3 𝐿₂ norm performance index of network synchronization")) is a Chua’s oscillator. In the dimensionless form, Chua’s oscillator is described by
<table>
<tbody>
<tr>
<td></td>
<td>{x˙1=α(−x1+x2−f(x1)),x˙2=x1−x2+x3,x˙3=−βx2−γx3,casessubscript˙𝑥1𝛼subscript𝑥1subscript𝑥2𝑓subscript𝑥1subscript˙𝑥2subscript𝑥1subscript𝑥2subscript𝑥3subscript˙𝑥3𝛽subscript𝑥2𝛾subscript𝑥3\left\{\begin{array}[]{l}\dot{x}_{1}=\alpha(-x_{1}+x_{2}-f(x_{1})),\\ \dot{x}_{2}=x_{1}-x_{2}+x_{3},\\ \dot{x}_{3}=-\beta x_{2}-\gamma x_{3},\\ \end{array}\right.</td>
<td></td>
<td>(15)</td>
</tr>
</tbody>
</table>
where $`f\hspace{0pt}{( \cdot )}`$ is a piecewise linear function:
<table>
<tbody>
<tr>
<td></td>
<td>f(x1)=m1x1+12(m2−m1)(|x1+1|−|x1−1|).𝑓subscript𝑥1subscript𝑚1subscript𝑥112subscript𝑚2subscript𝑚1subscript𝑥11subscript𝑥11f(x_{1})=m_{1}x_{1}+\frac{1}{2}(m_{2}-m_{1})(|x_{1}+1|-|x_{1}-1|).</td>
<td></td>
<td>(16)</td>
</tr>
</tbody>
</table>
Take parameters $`{\alpha = 9},{{\beta = 14},{{\gamma = 0.01},{m_{1} = {- 0.714}}}}`$, and $`m_{2} = {- 1.14}`$, so that Chua’s oscillator ([15](#S3.E15 "In 3 𝐿₂ norm performance index of network synchronization")) generates a double-scroll chaotic attractor. Set $`\Gamma = {\text{diag}\hspace{0pt}{(1,1,1)}}`$ in ([11](#S3.E11 "In 3 𝐿₂ norm performance index of network synchronization")). Fig. 1 shows the different synchronization performances of network ([11](#S3.E11 "In 3 𝐿₂ norm performance index of network synchronization")) with the same coupling strength $`\sigma`$ ($`\sigma = 6`$) but different coupling configurations. The corresponding values of $`{\|{e\hspace{0pt}{(t)}}\|}_{2}`$ are computed numerically as given in Table 1.
Nearest-neighbor coupling Star-shaped coupling Global coupling
Nearest-neighbor coupling Star-shaped coupling Global coupling
Fig. 1 Synchronization of the state variables of network ([11](#S3.E11 "In 3 𝐿₂ norm performance index of network synchronization")) with 20 nodes and the same coupling strength in different coupling configurations.
<table>
<thead>
<tr>
<th>Coupling</th>
<th>Nearest-neighbor</th>
<th>Star-shaped</th>
<th>Global</th>
</tr>
</thead>
<tbody>
<tr>
<th>‖e‖2subscriptnorm𝑒2\|e\|_{2}</th>
<td>9.2744</td>
<td>0.8030</td>
<td>0.4142</td>
</tr>
</tbody>
</table>
Table 1 Values of $`{\| e\|}_{2}`$.
As to the case of equilibrium synchronization of network ([11](#S3.E11 "In 3 𝐿₂ norm performance index of network synchronization")), since the synchronous state $`s`$ is known, the error vector can be defined in the following way:
<table>
<tbody>
<tr>
<td></td>
<td>ei=xi(t)−s,i=1,⋯,N,e(t)=[e1Te2T⋯eNT]T.formulae-sequencesubscript𝑒𝑖subscript𝑥𝑖𝑡𝑠𝑖1⋯𝑁𝑒𝑡superscriptdelimited-[]superscriptsubscript𝑒1𝑇superscriptsubscript𝑒2𝑇⋯superscriptsubscript𝑒𝑁𝑇𝑇\begin{array}[]{l}e_{i}=x_{i}(t)-s,\;i=1,\cdots,N,\\ e(t)=[e_{1}^{T}\;e_{2}^{T}\;\cdots\;e_{N}^{T}]^{T}.\end{array}</td>
<td></td>
<td>(17)</td>
</tr>
</tbody>
</table>
Example 2: Consider a Lur’e system,
<table>
<tbody>
<tr>
<td></td>
<td>{x˙1=(A1−2B1C1)x1+B1f1(y1),y1=C1x1,casessubscript˙𝑥1subscript𝐴12subscript𝐵1subscript𝐶1subscript𝑥1subscript𝐵1subscript𝑓1subscript𝑦1subscript𝑦1subscript𝐶1subscript𝑥1\left\{\begin{array}[]{l}\dot{x}_{1}=(A_{1}-2B_{1}C_{1})x_{1}+B_{1}f_{1}(y_{1}),\\ y_{1}=C_{1}x_{1},\end{array}\right.</td>
<td></td>
<td>(18)</td>
</tr>
</tbody>
</table>
where $`x_{1}`$ is the state, $`y_{1}`$ is the measured output,
<table>
<tbody>
<tr>
<td></td>
<td>A1=[01−a−b],B1=[b1b2],C1=[c1c2],Γ=[1001],formulae-sequencesubscript𝐴1delimited-[]01𝑎𝑏formulae-sequencesubscript𝐵1delimited-[]subscript𝑏1subscript𝑏2formulae-sequencesubscript𝐶1delimited-[]subscript𝑐1subscript𝑐2Γdelimited-[]1001A_{1}=\left[\begin{array}[]{cc}0&1\\ -a&-b\end{array}\right],\;B_{1}=\left[\begin{array}[]{c}b_{1}\\ b_{2}\end{array}\right],\;C_{1}=\left[\begin{array}[]{cc}c_{1}&c_{2}\end{array}\right],\;\Gamma=\left[\begin{array}[]{cc}1&0\\ 0&1\end{array}\right],</td>
<td></td>
</tr>
</tbody>
</table>
and the nonlinear function $`{f_{1}\hspace{0pt}{(y_{1})}} = {{|{y_{1} + 1}|} - {|{y_{1} - 1}|}}`$.
A network with system ([18](#S3.E18 "In 3 𝐿₂ norm performance index of network synchronization")) as individual nodes is given as follows:
<table>
<tbody>
<tr>
<td></td>
<td>{x˙=(IN⊗(A1−2B1C1)−σM⊗Γ)x+(IN⊗B1)f(y),y=(IN⊗C1)x,cases˙𝑥tensor-productsubscript𝐼𝑁subscript𝐴12subscript𝐵1subscript𝐶1tensor-product𝜎𝑀Γ𝑥tensor-productsubscript𝐼𝑁subscript𝐵1𝑓𝑦𝑦tensor-productsubscript𝐼𝑁subscript𝐶1𝑥\left\{\begin{array}[]{l}\dot{x}=(I_{N}\otimes(A_{1}-2B_{1}C_{1})-\sigma M\otimes\Gamma)x+(I_{N}\otimes B_{1})f(y),\\ y=(I_{N}\otimes C_{1})x,\end{array}\right.</td>
<td></td>
<td>(19)</td>
</tr>
</tbody>
</table>
where $`x = {(x_{1}^{T},\cdots,x_{N}^{T})}^{T}`$, $`y = {(y_{1},\cdots,y_{N})}^{T}`$ and $`{f\hspace{0pt}{(y)}} = {({f_{1}\hspace{0pt}{(y_{1})}},\cdots,{f_{N}\hspace{0pt}{(y_{N})}})}^{T}`$.
Network ([19](#S3.E19 "In 3 𝐿₂ norm performance index of network synchronization")) can be viewed as a large-scale system with measured output and feedback. Assume that system ([19](#S3.E19 "In 3 𝐿₂ norm performance index of network synchronization")) is observable. Then, the states of network nodes achieve synchronization if and only if all the outputs of network nodes achieve synchronization. Thus, one only needs to consider the outputs of network ([19](#S3.E19 "In 3 𝐿₂ norm performance index of network synchronization")).
Replace $`x_{i}`$ with $`y_{i}`$ and set $`s = 0`$ in ([17](#S3.E17 "In 3 𝐿₂ norm performance index of network synchronization")). Then, $`e_{i} = y_{i}`$ and $`e = y`$. Let $`{{a = 10},{{b = 3},{{{b\hspace{0pt}1} = 0},{{{b\hspace{0pt}2} = 1},{c_{1} = c_{2} = 1}}}}},`$ and $`\sigma = 2`$ in ([18](#S3.E18 "In 3 𝐿₂ norm performance index of network synchronization")). Fig. 2 shows the different performances of outputs of network ([19](#S3.E19 "In 3 𝐿₂ norm performance index of network synchronization")) with the same coupling strength $`\sigma`$ ($`\sigma`$=2) but different coupling configurations. The corresponding values of $`{\| y\|}_{2}`$ are listed in Table 2.
Nearest-neighbor coupling Star-shaped coupling Global coupling
Fig. 2 Synchronization of the outputs of network ([19](#S3.E19 "In 3 𝐿₂ norm performance index of network synchronization")) with 20 nodes having the same coupling strength but different coupling configurations.
<table>
<thead>
<tr>
<th>Coupling</th>
<th>Nearest-neighbor</th>
<th>Star-shaped</th>
<th>Global</th>
</tr>
</thead>
<tbody>
<tr>
<th>‖y‖2subscriptnorm𝑦2\|y\|_{2}</th>
<td>0.7296</td>
<td>0.5768</td>
<td>0.3990</td>
</tr>
</tbody>
</table>
Table 2 Values of $`{\| y\|}_{2}`$.
$`\square`$
Remark 1: As indicated by the above examples, in the same situation of stable synchronization, the performances of synchronization of networks can be quite different.
Remark 2: It is also clear that swiftness and overshoot are important indexes for describing the synchronous behaviors. The $`L_{2}`$ norm of the error vector $`e`$, i.e., $`{\| e\|}_{2}`$ as defined in ([13](#S3.E13 "In 3 𝐿₂ norm performance index of network synchronization")) or ([17](#S3.E17 "In 3 𝐿₂ norm performance index of network synchronization")) for the equilibrium synchronization case, can properly measure the swiftness and overshoot of the network synchronization: the smaller the $`{\| e\|}_{2}`$, the faster with lower overshoot the network synchronization. Thus, the quantity $`{\| e\|}_{2}`$ can be taken as a performance index of network synchronizability.
Remark 3: The quantity $`{\| e\|}_{2}`$ is influenced by many factors of the network, such as
\(1\) network structure, particularly the coupling strength $`\sigma`$ and the outer coupling matrix $`M`$;
\(2\) dynamical components, particularly the individual dynamics determined by $`f`$, the synchronous state $`s\hspace{0pt}{(t)}`$, and the inner linking matrix $`\Gamma`$.
## 4 Local synchronization to network equilibrium point
In the literature where the eigenvalues of the outer coupling matrix are used to measure the network synchronizability, it is a topic of great interest to investigate the relationship between the eigenvalues and the network structure thereby finding suitable ways to enhance the synchronizability. In this section, for the case of equilibrium synchronization, the relationship between the quantity $`{\| e\|}_{2}`$ and some network parameters is explored.
### 4.1 $`{\| G\|}_{2}`$ as a constraint of $`{\| e\|}_{2}`$
Suppose that each single node in network ([11](#S3.E11 "In 3 𝐿₂ norm performance index of network synchronization")) has a measured output $`y = {C\hspace{0pt}x}`$. Then, the equations of network ([11](#S3.E11 "In 3 𝐿₂ norm performance index of network synchronization")) can be written as follows:
<table>
<tbody>
<tr>
<td></td>
<td>{x˙i=f(xi)−σ∑j=1NmijΓxj,xi(0)=xi0,yi=Cxi,i=1,⋯,N.formulae-sequencecasesformulae-sequencesubscript˙𝑥𝑖𝑓subscript𝑥𝑖𝜎superscriptsubscript𝑗1𝑁subscript𝑚𝑖𝑗Γsubscript𝑥𝑗subscript𝑥𝑖0subscript𝑥𝑖0subscript𝑦𝑖𝐶subscript𝑥𝑖𝑖1⋯𝑁\left\{\begin{array}[]{l}\dot{x}_{i}=f(x_{i})-\sigma\sum\limits_{j=1}^{N}m_{ij}\Gamma x_{j},\;x_{i}(0)=x_{i0},\\ y_{i}=Cx_{i},\end{array}\qquad i=1,\cdots,N.\right.</td>
<td></td>
<td>(20)</td>
</tr>
</tbody>
</table>
If $`C = I_{n}`$, then $`y_{i} = x_{i}`$ is just the state of the $`i`$th node.
Let $`s`$ denote an equilibrium point of the individual node, satisfying $`{f\hspace{0pt}{(s)}} = 0`$. Then, the linearized equations of ([20](#S4.E20 "In 4.1 ‖𝐺‖₂ as a constraint of ‖𝑒‖₂ ‣ 4 Local synchronization to network equilibrium point")) about the synchronous solutions $`x_{i} = s`$ are as follows:
<table>
<tbody>
<tr>
<td></td>
<td>{η˙i=Df(s)ηi−σ∑j=1NmijΓηj,ηi(0)=xi0−s=ηi0,ei=Cηi,i=1,⋯,N,formulae-sequencecasesformulae-sequencesubscript˙𝜂𝑖𝐷𝑓𝑠subscript𝜂𝑖𝜎superscriptsubscript𝑗1𝑁subscript𝑚𝑖𝑗Γsubscript𝜂𝑗subscript𝜂𝑖0subscript𝑥𝑖0𝑠subscript𝜂𝑖0subscript𝑒𝑖𝐶subscript𝜂𝑖𝑖1⋯𝑁\left\{\begin{array}[]{l}\dot{\eta}_{i}=Df(s)\eta_{i}-\sigma\sum\limits_{j=1}^{N}m_{ij}\Gamma\eta_{j},\;\eta_{i}(0)=x_{i0}-s=\eta_{i0},\\ e_{i}=C\eta_{i},\end{array}\qquad i=1,\cdots,N,\right.</td>
<td></td>
<td>(21)</td>
</tr>
</tbody>
</table>
where $`\eta_{i}`$ ($`\eta_{i} = {x_{i} - s}`$) and $`e_{i}`$ are the state error vector and the output error vector to the $`i`$th node, respectively. Viewing the impulse function $`\eta_{i\hspace{0pt}0}\hspace{0pt}\delta\hspace{0pt}{(t)}`$ as an input to system ([21](#S4.E21 "In 4.1 ‖𝐺‖₂ as a constraint of ‖𝑒‖₂ ‣ 4 Local synchronization to network equilibrium point")), system ([21](#S4.E21 "In 4.1 ‖𝐺‖₂ as a constraint of ‖𝑒‖₂ ‣ 4 Local synchronization to network equilibrium point")) can be equivalently written as
<table>
<tbody>
<tr>
<td></td>
<td>{η˙i=Df(s)ηi−σ∑j=1NmijΓηj+ηi0δ(t),ηi(0−)=0,ei=Cηi,i=1,⋯,N.formulae-sequencecasesformulae-sequencesubscript˙𝜂𝑖𝐷𝑓𝑠subscript𝜂𝑖𝜎superscriptsubscript𝑗1𝑁subscript𝑚𝑖𝑗Γsubscript𝜂𝑗subscript𝜂𝑖0𝛿𝑡subscript𝜂𝑖subscript00subscript𝑒𝑖𝐶subscript𝜂𝑖𝑖1⋯𝑁\left\{\begin{array}[]{l}\dot{\eta}_{i}=Df(s)\eta_{i}-\sigma\sum\limits_{j=1}^{N}m_{ij}\Gamma\eta_{j}+\eta_{i0}\delta(t),\;\eta_{i}(0_{-})=0,\\ e_{i}=C\eta_{i},\end{array}\qquad i=1,\cdots,N.\right.</td>
<td></td>
<td>(22)</td>
</tr>
</tbody>
</table>
Using the Kronecker product, the error system ([22](#S4.E22 "In 4.1 ‖𝐺‖₂ as a constraint of ‖𝑒‖₂ ‣ 4 Local synchronization to network equilibrium point")) can be rewritten as
<table>
<tbody>
<tr>
<td></td>
<td>{η˙=(IN⊗Df(s)−σM⊗Γ)η+(IN⊗In)u(t),e=(IN⊗C)η,cases˙𝜂tensor-productsubscript𝐼𝑁𝐷𝑓𝑠tensor-product𝜎𝑀Γ𝜂tensor-productsubscript𝐼𝑁subscript𝐼𝑛𝑢𝑡𝑒tensor-productsubscript𝐼𝑁𝐶𝜂\left\{\begin{array}[]{l}\dot{\eta}=(I_{N}\otimes Df(s)-\sigma M\otimes\Gamma)\eta+(I_{N}\otimes I_{n})u(t),\\ e=(I_{N}\otimes C)\eta,\end{array}\right.</td>
<td></td>
<td>(23)</td>
</tr>
</tbody>
</table>
where $`\eta = {\lbrack{\eta_{1}^{T}\hspace{0pt}\eta_{2}^{T}\hspace{0pt}\cdots\hspace{0pt}\eta_{N}^{T}}\rbrack}^{T}`$, $`e = {\lbrack{e_{1}^{T}\hspace{0pt}e_{2}^{T}\hspace{0pt}\cdots\hspace{0pt}e_{N}^{T}}\rbrack}^{T}`$ and $`{u\hspace{0pt}{(t)}} = {\eta_{0}\hspace{0pt}\delta\hspace{0pt}{(t)}}`$.
As in Example 2, suppose that system ([23](#S4.E23 "In 4.1 ‖𝐺‖₂ as a constraint of ‖𝑒‖₂ ‣ 4 Local synchronization to network equilibrium point")) is observable. Then, $`{\| e\|}_{2}`$ can be taken as a measure of the swiftness and overshoot of the synchronization of network ([20](#S4.E20 "In 4.1 ‖𝐺‖₂ as a constraint of ‖𝑒‖₂ ‣ 4 Local synchronization to network equilibrium point")). Let $`G\hspace{0pt}{(s)}`$ denote the transfer function of the error equation ([23](#S4.E23 "In 4.1 ‖𝐺‖₂ as a constraint of ‖𝑒‖₂ ‣ 4 Local synchronization to network equilibrium point")) from $`u`$ to $`e`$. Then
<table>
<tbody>
<tr>
<td></td>
<td>G(s)=(IN⊗C)[s(IN⊗In)−(IN⊗Df(s)−σM⊗Γ)]−1(IN⊗In).𝐺𝑠tensor-productsubscript𝐼𝑁𝐶superscriptdelimited-[]𝑠tensor-productsubscript𝐼𝑁subscript𝐼𝑛tensor-productsubscript𝐼𝑁𝐷𝑓𝑠tensor-product𝜎𝑀Γ1tensor-productsubscript𝐼𝑁subscript𝐼𝑛G(s)=(I_{N}\otimes C)[s(I_{N}\otimes I_{n})-(I_{N}\otimes Df(s)-\sigma M\otimes\Gamma)]^{-1}(I_{N}\otimes I_{n}).</td>
<td></td>
<td>(24)</td>
</tr>
</tbody>
</table>
Let the transfer function $`G\hspace{0pt}{(s)}`$ be given in ([24](#S4.E24 "In 4.1 ‖𝐺‖₂ as a constraint of ‖𝑒‖₂ ‣ 4 Local synchronization to network equilibrium point")), then Lemma 3: The inequality
<table>
<tbody>
<tr>
<td></td>
<td>‖e‖2≤‖G‖2‖η0‖2subscriptnorm𝑒2subscriptnorm𝐺2subscriptnormsubscript𝜂02\|e\|_{2}\leq\|G\|_{2}\|\eta_{0}\|_{2}</td>
<td></td>
<td>(25)</td>
</tr>
</tbody>
</table>
holds, where $`{\| e\|}_{2}`$ is the $`L_{2}`$ norm of the error vector $`e\hspace{0pt}{(t)}`$ on the interval $`\lbrack 0,\infty)`$, $`{\| G\|}_{2}`$ is the $`H_{2}`$ norm of the transfer function $`G\hspace{0pt}{(s)}`$, and $`{\|\eta_{0}\|}_{2}`$ is the vector 2-norm of the initial error vector $`\eta_{0}`$.
Proof: Since $`{u\hspace{0pt}{(t)}} = {\eta_{0}\hspace{0pt}\delta\hspace{0pt}{(t)}}`$, $`{e\hspace{0pt}{(t)}} = {g\hspace{0pt}{(t)}\hspace{0pt}\eta_{0}}`$, where $`g\hspace{0pt}{(t)}`$ denotes the corresponding bilateral Laplace transform of $`G\hspace{0pt}{(s)}`$. Then, by Parseval’s identity,
<table>
<tbody>
<tr>
<td></td>
<td>‖e(t)‖2=‖g(t)η0‖2=‖Gη0‖2≤‖G‖2‖η0‖2.subscriptnorm𝑒𝑡2subscriptnorm𝑔𝑡subscript𝜂02subscriptnorm𝐺subscript𝜂02subscriptnorm𝐺2subscriptnormsubscript𝜂02\|e(t)\|_{2}=\|g(t)\eta_{0}\|_{2}=\|G\eta_{0}\|_{2}\leq\|G\|_{2}\|\eta_{0}\|_{2}.</td>
<td></td>
</tr>
</tbody>
</table>
$`\square`$
Thus, the quantity $`{\| e\|}_{2}`$ is upper-bounded by the product $`{\| G\|}_{2}\hspace{0pt}{\|\eta_{0}\|}_{2}`$. Since $`\eta_{0}`$ is the given initial error vector, $`{\| G\|}_{2}`$ can be taken as a constraint of $`{\| e\|}_{2}`$. In fact, as introduced in Sec.1, $`{\| G\|}_{2}^{2}`$ equals the overall output energy of the system response to the impulse input.
Remark 4: The advantages of using the quantity $`{\| G\|}_{2}`$ include:
\(1\) $`{\| G\|}_{2}`$ can be numerically computed;
\(2\) the synchronizability is affected by many factors of a network, while $`{\| G\|}_{2}`$ can be seen as an overall reflection of these network factors consisting of both structural and dynamical ones.
Example 3: Consider Example 2 again. The linearized equation of network ([19](#S3.E19 "In 3 𝐿₂ norm performance index of network synchronization")) about the equilibrium point $`s = {\lbrack 0\;0\;0\rbrack}`$ is given as follows:
<table>
<tbody>
<tr>
<td></td>
<td>{η˙=(IN⊗A1−σM⊗Γ)η+(IN⊗In)η0δ(t),η(0−)=0,e=(IN⊗C)η.casesformulae-sequence˙𝜂tensor-productsubscript𝐼𝑁subscript𝐴1tensor-product𝜎𝑀Γ𝜂tensor-productsubscript𝐼𝑁subscript𝐼𝑛subscript𝜂0𝛿𝑡𝜂subscript00𝑒tensor-productsubscript𝐼𝑁𝐶𝜂\left\{\begin{array}[]{l}\dot{\eta}=(I_{N}\otimes A_{1}-\sigma M\otimes\Gamma)\eta+(I_{N}\otimes I_{n})\eta_{0}\delta(t),\;\;\eta(0_{-})=0,\\ e=(I_{N}\otimes C)\eta.\end{array}\right.</td>
<td></td>
<td>(26)</td>
</tr>
</tbody>
</table>
The corresponding values of $`{\| G\|}_{2}`$ with the three different network configurations are listed in Table 3.
<table>
<thead>
<tr>
<th>Coupling</th>
<th>Nearest-neighbor</th>
<th>Star-shaped</th>
<th>Global</th>
</tr>
</thead>
<tbody>
<tr>
<th>‖G‖2subscriptnorm𝐺2\|G\|_{2}</th>
<td>2.5091</td>
<td>2.4383</td>
<td>1.2565</td>
</tr>
<tr>
<th>‖e‖2subscriptnorm𝑒2\|e\|_{2}</th>
<td>0.7296</td>
<td>0.5768</td>
<td>0.3990</td>
</tr>
</tbody>
</table>
Table 3 Values of $`{\| G\|}_{2}`$ of network ([19](#S3.E19 "In 3 𝐿₂ norm performance index of network synchronization")) with different network configurations.
### 4.2 Relationship between $`{\| G\|}_{2}`$ and network structure
Example 3 shows that $`{\| G\|}_{2}`$ is influenced by the network configurations. Thus, the relationship between $`{\| G\|}_{2}`$ and the network structure is a problem deserving further investigation.
In this section, it is always assumed that the outer coupling matrix $`M`$ is symmetrical. Then, there exists a unitary matrix $`U \in {\mathbb{R}}^{N \times N}`$ such that $`M = {U\hspace{0pt}\Delta\hspace{0pt}U^{- 1}}`$, where $`\Delta = {\text{diag}\hspace{0pt}{(\lambda_{1},\lambda_{2},\cdots,\lambda_{N})}}`$ is a diagonal matrix with the diagonal entries $`{{\lambda_{i},i} = 1},{\cdots,N}`$, being the eigenvalues of matrix $`M`$. Let $`\eta = {{({U \otimes I_{n}})}\hspace{0pt}\xi}`$, $`u = {{({U \otimes I_{n}})}\hspace{0pt}\omega}`$, and $`z = {{({U^{- 1} \otimes I_{l}})}\hspace{0pt}e}`$. Then, the error equation ([23](#S4.E23 "In 4.1 ‖𝐺‖₂ as a constraint of ‖𝑒‖₂ ‣ 4 Local synchronization to network equilibrium point")) can be rewritten as follows:
<table>
<tbody>
<tr>
<td></td>
<td>{ξ˙=(IN⊗Df(s)−σΔ⊗Γ)ξ+(IN⊗In)ω,z=(IN⊗C)ξ.cases˙𝜉tensor-productsubscript𝐼𝑁𝐷𝑓𝑠tensor-product𝜎ΔΓ𝜉tensor-productsubscript𝐼𝑁subscript𝐼𝑛𝜔𝑧tensor-productsubscript𝐼𝑁𝐶𝜉\left\{\begin{array}[]{l}\dot{\xi}=(I_{N}\otimes Df(s)-\sigma\Delta\otimes\Gamma)\xi+(I_{N}\otimes I_{n})\omega,\\ z=(I_{N}\otimes C)\xi.\end{array}\right.</td>
<td></td>
<td>(27)</td>
</tr>
</tbody>
</table>
Note that system ([27](#S4.E27 "In 4.2 Relationship between ‖𝐺‖₂ and network structure ‣ 4 Local synchronization to network equilibrium point")) is composed of $`N`$ uncoupled subsystems:
<table>
<tbody>
<tr>
<td></td>
<td>{ξ˙i=(Df(s)−σλiΓ)ξi+Inωi,zi=Cξi,i=1,⋯,N,formulae-sequencecasessubscript˙𝜉𝑖𝐷𝑓𝑠𝜎subscript𝜆𝑖Γsubscript𝜉𝑖subscript𝐼𝑛subscript𝜔𝑖subscript𝑧𝑖𝐶subscript𝜉𝑖𝑖1⋯𝑁\left\{\begin{array}[]{l}\dot{\xi}_{i}=(Df(s)-\sigma\lambda_{i}\Gamma)\xi_{i}+I_{n}\omega_{i},\\ z_{i}=C\xi_{i},\end{array}\qquad i=1,\cdots,N,\right.</td>
<td></td>
<td>(28)</td>
</tr>
</tbody>
</table>
where $`\omega_{i}`$ is the $`i`$th component of the input vector $`\omega = {{({U^{- 1} \otimes I_{n}})}\hspace{0pt}u} = {{({U^{- 1} \otimes I_{n}})}\hspace{0pt}\eta_{0}\hspace{0pt}\delta\hspace{0pt}{(t)}}`$. Let
<table>
<tbody>
<tr>
<td></td>
<td>Ei=diag(0,⋯,0,1,0,⋯,0),subscript𝐸𝑖diag0⋯010⋯0E_{i}=\textrm{diag}(0,\cdots,0,1,0,\cdots,0),</td>
<td></td>
</tr>
</tbody>
</table>
be a matrix with the $`i`$th diagonal entry being 1 and all the other entries being zero. Then, $`\omega_{i} = {{({E_{i} \otimes I_{n}})}\hspace{0pt}{({U^{- 1} \otimes I_{n}})}\hspace{0pt}\omega} = {{({{E_{i}\hspace{0pt}U^{- 1}} \otimes I_{n}})}\hspace{0pt}\eta_{0}\hspace{0pt}\delta\hspace{0pt}{(t)}}`$ and $`\omega_{i\hspace{0pt}0} = {{({{E_{i}\hspace{0pt}U^{- 1}} \otimes I_{n}})}\hspace{0pt}\eta_{0}}`$.
From ([28](#S4.E28 "In 4.2 Relationship between ‖𝐺‖₂ and network structure ‣ 4 Local synchronization to network equilibrium point")), the condition for ensuring the stable synchronous state
<table>
<tbody>
<tr>
<td></td>
<td>x1=x2=⋯=xN=ssubscript𝑥1subscript𝑥2⋯subscript𝑥𝑁𝑠x_{1}=x_{2}=\cdots=x_{N}=s</td>
<td></td>
</tr>
</tbody>
</table>
is that the $`N`$ matrices
<table>
<tbody>
<tr>
<td></td>
<td>Df(s)−σλiΓ,i=1,⋯,N,formulae-sequence𝐷𝑓𝑠𝜎subscript𝜆𝑖Γ𝑖1⋯𝑁Df(s)-\sigma\lambda_{i}\Gamma,\qquad i=1,\cdots,N,</td>
<td></td>
<td>(29)</td>
</tr>
</tbody>
</table>
are all stable.
Let $`T\hspace{0pt}{(s)}`$ denote the transfer function of ([27](#S4.E27 "In 4.2 Relationship between ‖𝐺‖₂ and network structure ‣ 4 Local synchronization to network equilibrium point")) from $`\omega`$ to $`z`$. Then
<table>
<tbody>
<tr>
<td></td>
<td>T(s)=(IN⊗C)[s(IN⊗In)−(IN⊗Df(s)−σΔ⊗Γ)]−1(IN⊗In)=(U−1⊗Il)G(s)(U⊗Im).𝑇𝑠absenttensor-productsubscript𝐼𝑁𝐶superscriptdelimited-[]𝑠tensor-productsubscript𝐼𝑁subscript𝐼𝑛tensor-productsubscript𝐼𝑁𝐷𝑓𝑠tensor-product𝜎ΔΓ1tensor-productsubscript𝐼𝑁subscript𝐼𝑛missing-subexpressionabsenttensor-productsuperscript𝑈1subscript𝐼𝑙𝐺𝑠tensor-product𝑈subscript𝐼𝑚\begin{array}[]{rl}T(s)&=(I_{N}\otimes C)[s(I_{N}\otimes I_{n})-(I_{N}\otimes Df(s)-\sigma\Delta\otimes\Gamma)]^{-1}(I_{N}\otimes I_{n})\\ &=(U^{-1}\otimes I_{l})G(s)(U\otimes I_{m}).\end{array}</td>
<td></td>
</tr>
</tbody>
</table>
Since both $`({U^{- 1} \otimes I_{l}})`$ and $`({U \otimes I_{m}})`$ are unitary transformations, one has
<table>
<tbody>
<tr>
<td></td>
<td>‖T(s)‖2=‖G(s)‖2.subscriptnorm𝑇𝑠2subscriptnorm𝐺𝑠2\|T(s)\|_{2}=\|G(s)\|_{2}.</td>
<td></td>
<td>(30)</td>
</tr>
</tbody>
</table>
Let $`{T_{i}\hspace{0pt}{(s)}} = {C\hspace{0pt}{\lbrack{{s\hspace{0pt}I_{n}} - {({{D\hspace{0pt}f\hspace{0pt}{(s)}} - {\sigma\hspace{0pt}\lambda_{i}\hspace{0pt}\Gamma}})}}\rbrack}^{- 1}\hspace{0pt}I_{n}}`$ denote the transfer function of the $`i`$th subsystem in ([28](#S4.E28 "In 4.2 Relationship between ‖𝐺‖₂ and network structure ‣ 4 Local synchronization to network equilibrium point")). Then
<table>
<tbody>
<tr>
<td></td>
<td>T(s)=[T1(s)⋱TN(s)]𝑇𝑠delimited-[]subscript𝑇1𝑠missing-subexpressionmissing-subexpressionmissing-subexpression⋱missing-subexpressionmissing-subexpressionmissing-subexpressionsubscript𝑇𝑁𝑠T(s)=\left[\begin{array}[]{ccc}T_{1}(s)&&\\ &\ddots&\\ &&T_{N}(s)\end{array}\right]</td>
<td></td>
</tr>
</tbody>
</table>
and
<table>
<tbody>
<tr>
<td></td>
<td>‖T(s)‖22=∑i=1N‖Ti(s)‖22.superscriptsubscriptnorm𝑇𝑠22superscriptsubscript𝑖1𝑁superscriptsubscriptnormsubscript𝑇𝑖𝑠22\|T(s)\|_{2}^{2}=\sum\limits_{i=1}^{N}\|T_{i}(s)\|_{2}^{2}.</td>
<td></td>
<td>(31)</td>
</tr>
</tbody>
</table>
Assumption 1: Suppose that the outer coupling matrix $`M`$ is symmetrical, diffusive and irreducible, with the off-diagonal entries $`m_{i\hspace{0pt}j} \leq 0`$ and the diagonal entries $`m_{i\hspace{0pt}i} = {- {\sum_{{j = 1},{j \neq i}}^{N}m_{i\hspace{0pt}j}}}`$, for $`{{i,j} = 1},{2,\cdots,N}`$.
The outer coupling matrix $`M`$ satisfying Assumption 1 has a zero eigenvalue of multiplicity $`1`$, and its other eigenvalues are all positive.
Theorem 1: Suppose that the inner linking matrix $`\Gamma = {k\hspace{0pt}I}`$, where $`k > 0`$ is a constant and $`I`$ the identity matrix. Suppose that the outer coupling matrix $`M`$ satisfies Assumption 1 with the eigenvalues given as follows:
<table>
<tbody>
<tr>
<td></td>
<td>0=λ1<λ2≤⋯≤λN.0subscript𝜆1subscript𝜆2⋯subscript𝜆𝑁0=\lambda_{1}<\lambda_{2}\leq\cdots\leq\lambda_{N}.</td>
<td></td>
<td>(32)</td>
</tr>
</tbody>
</table>
Then
<table>
<tbody>
<tr>
<td></td>
<td>‖Ti(s)‖22≤‖Tj(s)‖22,superscriptsubscriptnormsubscript𝑇𝑖𝑠22superscriptsubscriptnormsubscript𝑇𝑗𝑠22\|T_{i}(s)\|_{2}^{2}\leq\|T_{j}(s)\|_{2}^{2},</td>
<td></td>
<td>(33)</td>
</tr>
</tbody>
</table>
where $`1 \leq j < i \leq N`$, and
<table>
<tbody>
<tr>
<td></td>
<td>‖T(s)‖22≥‖T1(s)‖22+(N−1)‖TN(s)‖22≥N‖TN(s)‖22,‖T(s)‖22≤‖T1(s)‖22+(N−1)‖T2(s)‖22≤N‖T1(s)‖22.superscriptsubscriptnorm𝑇𝑠22superscriptsubscriptnormsubscript𝑇1𝑠22𝑁1superscriptsubscriptnormsubscript𝑇𝑁𝑠22𝑁superscriptsubscriptnormsubscript𝑇𝑁𝑠22superscriptsubscriptnorm𝑇𝑠22superscriptsubscriptnormsubscript𝑇1𝑠22𝑁1superscriptsubscriptnormsubscript𝑇2𝑠22𝑁superscriptsubscriptnormsubscript𝑇1𝑠22\begin{array}[]{l}\|T(s)\|_{2}^{2}\geq\|T_{1}(s)\|_{2}^{2}+(N-1)\|T_{N}(s)\|_{2}^{2}\geq N\|T_{N}(s)\|_{2}^{2},\\ \|T(s)\|_{2}^{2}\leq\|T_{1}(s)\|_{2}^{2}+(N-1)\|T_{2}(s)\|_{2}^{2}\leq N\|T_{1}(s)\|_{2}^{2}.\end{array}</td>
<td></td>
<td>(34)</td>
</tr>
</tbody>
</table>
Proof: Let $`P`$ be an arbitrary positive definite matrix such that
<table>
<tbody>
<tr>
<td></td>
<td>P(Df(s)−σλjΓ)+(Df(s)−σλjΓ)TP+CTC<0.𝑃𝐷𝑓𝑠𝜎subscript𝜆𝑗Γsuperscript𝐷𝑓𝑠𝜎subscript𝜆𝑗Γ𝑇𝑃superscript𝐶𝑇𝐶0P(Df(s)-\sigma\lambda_{j}\Gamma)+(Df(s)-\sigma\lambda_{j}\Gamma)^{T}P+C^{T}C<0.</td>
<td></td>
</tr>
</tbody>
</table>
By the assumption that $`\Gamma = {k\hspace{0pt}I}`$, one has
<table>
<tbody>
<tr>
<td></td>
<td>PΓ+ΓTP>0.𝑃ΓsuperscriptΓ𝑇𝑃0P\Gamma+\Gamma^{T}P>0.</td>
<td></td>
</tr>
</tbody>
</table>
Then
<table>
<tbody>
<tr>
<td></td>
<td>P(Df(s)−σλiΓ)+(Df(s)−σλiΓ)TP+CTC=P(Df(s)−σλjΓ)+(Df(s)−σλjΓ)TP+CTC−σ(λi−λj)(PΓ+ΓTP)<−σ(λi−λj)(PΓ+ΓTP)≤0.missing-subexpression𝑃𝐷𝑓𝑠𝜎subscript𝜆𝑖Γsuperscript𝐷𝑓𝑠𝜎subscript𝜆𝑖Γ𝑇𝑃superscript𝐶𝑇𝐶missing-subexpressionabsent𝑃𝐷𝑓𝑠𝜎subscript𝜆𝑗Γsuperscript𝐷𝑓𝑠𝜎subscript𝜆𝑗Γ𝑇𝑃superscript𝐶𝑇𝐶𝜎subscript𝜆𝑖subscript𝜆𝑗𝑃ΓsuperscriptΓ𝑇𝑃missing-subexpressionabsent𝜎subscript𝜆𝑖subscript𝜆𝑗𝑃ΓsuperscriptΓ𝑇𝑃0\begin{array}[]{rl}&P(Df(s)-\sigma\lambda_{i}\Gamma)+(Df(s)-\sigma\lambda_{i}\Gamma)^{T}P+C^{T}C\\ &=P(Df(s)-\sigma\lambda_{j}\Gamma)+(Df(s)-\sigma\lambda_{j}\Gamma)^{T}P+C^{T}C-\sigma(\lambda_{i}-\lambda_{j})(P\Gamma+\Gamma^{T}P)\\ &<-\sigma(\lambda_{i}-\lambda_{j})(P\Gamma+\Gamma^{T}P)\leq 0.\end{array}</td>
<td></td>
</tr>
</tbody>
</table>
The above inequality and Lemma 1 together leads to the assertions of the theorem. $`\square`$
Remark 5: Since $`\lambda_{1} = 0`$, $`{\|{T_{1}\hspace{0pt}{(s)}}\|}_{2}`$ in ([34](#S4.E34 "In 4.2 Relationship between ‖𝐺‖₂ and network structure ‣ 4 Local synchronization to network equilibrium point")) represents the $`H_{2}`$ norm of the transfer function of each individual node in the network.
Remark 6: Theorem 1 provides a simple way for comparing $`{\| G\|}_{2}`$ of a network with its different possible structures. It can be deduced from Theorem 1 that $`{\| G\|}_{2}`$ will not increase as the eigenvalues of the outer coupling matrix increase.
Example 4: Consider Example 2 again. The eigenvalues of the global coupling matrix $`M_{g\hspace{0pt}l\hspace{0pt}o}`$ are $`\lambda_{1} = 0`$ and $`\lambda_{2} = \cdots = \lambda_{N} = N`$, while the eigenvalues of the star-shaped coupling matrix $`M_{s\hspace{0pt}t\hspace{0pt}a}`$ are $`\lambda_{1} = 0`$, $`\lambda_{2} = \cdots = \lambda_{N - 1} = 1`$ and $`\lambda_{N} = N`$. Let $`G_{g\hspace{0pt}l\hspace{0pt}o}`$ and $`G_{s\hspace{0pt}t\hspace{0pt}a}`$ denote the transfer functions of the global coupling and the star-shaped coupling networks, respectively. Then, by Theorem 1,
<table>
<tbody>
<tr>
<td></td>
<td>‖Gglo‖22=‖Tglo‖22=‖T1(s)‖22+(N−1)‖Tλ=N‖22≤‖T1(s)‖22+‖Tλ=N‖22+(N−2)‖Tλ=1‖22=‖Tsta‖22=‖Gsta‖22.superscriptsubscriptnormsubscript𝐺𝑔𝑙𝑜22superscriptsubscriptnormsubscript𝑇𝑔𝑙𝑜22absentsuperscriptsubscriptnormsubscript𝑇1𝑠22𝑁1superscriptsubscriptnormsubscript𝑇𝜆𝑁22missing-subexpressionabsentsuperscriptsubscriptnormsubscript𝑇1𝑠22superscriptsubscriptnormsubscript𝑇𝜆𝑁22𝑁2superscriptsubscriptnormsubscript𝑇𝜆122missing-subexpressionabsentsuperscriptsubscriptnormsubscript𝑇𝑠𝑡𝑎22superscriptsubscriptnormsubscript𝐺𝑠𝑡𝑎22\begin{array}[]{rl}\|G_{glo}\|_{2}^{2}=\|T_{glo}\|_{2}^{2}&=\|T_{1}(s)\|_{2}^{2}+(N-1)\|T_{\lambda=N}\|_{2}^{2}\\ &\leq\|T_{1}(s)\|_{2}^{2}+\|T_{\lambda=N}\|_{2}^{2}+(N-2)\|T_{\lambda=1}\|_{2}^{2}\\ &=\|T_{sta}\|_{2}^{2}=\|G_{sta}\|_{2}^{2}.\end{array}</td>
<td></td>
</tr>
</tbody>
</table>
It is consistent with the numerical results given in Example 3.
For Example 2, through integral computations, an analytical expression of $`{\| T_{i}\|}_{2}^{2}`$ can be obtained as follows:
<table>
<tbody>
<tr>
<td></td>
<td>‖Ti‖22=2(c12+c22)σ2λi2+[(3c12+c22)b+2c1c2(1−a)]σλi+(c1b−c2a)2+(c12+c22)a+c124σ3λi3+6bσ2λi2+(4a+2b2)σλi+2ab.superscriptsubscriptnormsubscript𝑇𝑖222superscriptsubscript𝑐12superscriptsubscript𝑐22superscript𝜎2superscriptsubscript𝜆𝑖2delimited-[]3superscriptsubscript𝑐12superscriptsubscript𝑐22𝑏2subscript𝑐1subscript𝑐21𝑎𝜎subscript𝜆𝑖superscriptsubscript𝑐1𝑏subscript𝑐2𝑎2superscriptsubscript𝑐12superscriptsubscript𝑐22𝑎superscriptsubscript𝑐124superscript𝜎3superscriptsubscript𝜆𝑖36𝑏superscript𝜎2superscriptsubscript𝜆𝑖24𝑎2superscript𝑏2𝜎subscript𝜆𝑖2𝑎𝑏\|T_{i}\|_{2}^{2}=\frac{2(c_{1}^{2}+c_{2}^{2})\sigma^{2}\lambda_{i}^{2}+[(3c_{1}^{2}+c_{2}^{2})b+2c_{1}c_{2}(1-a)]\sigma\lambda_{i}+(c_{1}b-c_{2}a)^{2}+(c_{1}^{2}+c_{2}^{2})a+c_{1}^{2}}{4\sigma^{3}\lambda_{i}^{3}+6b\sigma^{2}\lambda_{i}^{2}+(4a+2b^{2})\sigma\lambda_{i}+2ab}.</td>
<td></td>
<td>(35)</td>
</tr>
</tbody>
</table>
By ([35](#S4.E35 "In 4.2 Relationship between ‖𝐺‖₂ and network structure ‣ 4 Local synchronization to network equilibrium point")) and ([31](#S4.E31 "In 4.2 Relationship between ‖𝐺‖₂ and network structure ‣ 4 Local synchronization to network equilibrium point")), Table 4 gives the different values of $`{\| G_{g\hspace{0pt}l\hspace{0pt}o}\|}_{2}`$ and $`{\| G_{s\hspace{0pt}t\hspace{0pt}a}\|}_{2}`$ of network ([19](#S3.E19 "In 3 𝐿₂ norm performance index of network synchronization")) as the node number $`N`$ increases.
<table>
<thead>
<tr>
<th>N</th>
<th>1</th>
<th>10</th>
<th>100</th>
<th>1000</th>
<th>∞\infty</th>
</tr>
</thead>
<tbody>
<tr>
<td>‖Gglo‖2subscriptnormsubscript𝐺𝑔𝑙𝑜2\|G_{glo}\|_{2}</td>
<td>1.0801</td>
<td>1.3190</td>
<td>1.4492</td>
<td>1.4696</td>
<td>1.2910</td>
</tr>
<tr>
<td>‖Gsta‖2subscriptnormsubscript𝐺𝑠𝑡𝑎2\|G_{sta}\|_{2}</td>
<td>1.0801</td>
<td>2.2619</td>
<td>6.9840</td>
<td>22.0434</td>
<td>∞\infty</td>
</tr>
</tbody>
</table>
Table 4 Values of $`{\| G\|}_{2}`$ of network ([19](#S3.E19 "In 3 𝐿₂ norm performance index of network synchronization")) with different network configurations.
$`\square`$
## 5 Optimal controller design
So far, the pinning control strategy is extensively used for achieving synchronization of dynamical networks \[[5](#bib.bib5), [8](#bib.bib8)\]. The main advantage of the pinning control strategy is that only a few network nodes are needed to be controlled. AS far as the control effects (referring to the swiftness and overshoot of the synchronization) and the control cost (represented by the $`L_{2}`$ norm of the control input) are concerned, however, pinning control may not be the best choice. It is revealed by the LQR problem, as illuminated in ([7](#S2.E7 "In 2 Preliminaries")), when matrices $`C`$ and $`D`$ are properly selected, both the control effects and the control cost can be simultaneously evaluated by the $`L_{2}`$ norm of the measured output $`y`$ of the controlled network. The smaller the $`\| y\|`$, the better the control effects and the lower the control cost. In this section, based on the optimal solution to the LQR problem, an LQR optimal controller is designed for network ([11](#S3.E11 "In 3 𝐿₂ norm performance index of network synchronization")), which can drive the network onto some homogenous stationary states while minimizing the quantity $`\| y\|`$.
Suppose that the controlled network is given as follows:
<table>
<tbody>
<tr>
<td></td>
<td>{x˙i=f(xi)−σ∑j=1NmijΓxj+Buiyi=Cxi+Dui,i=1,⋯,N,formulae-sequencecasessubscript˙𝑥𝑖𝑓subscript𝑥𝑖𝜎superscriptsubscript𝑗1𝑁subscript𝑚𝑖𝑗Γsubscript𝑥𝑗𝐵subscript𝑢𝑖subscript𝑦𝑖𝐶subscript𝑥𝑖𝐷subscript𝑢𝑖𝑖1⋯𝑁\left\{\begin{array}[]{l}\dot{x}_{i}=f(x_{i})-\sigma\sum\limits_{j=1}^{N}m_{ij}\Gamma x_{j}+Bu_{i}\\ y_{i}=Cx_{i}+Du_{i},\end{array}\qquad i=1,\cdots,N,\right.</td>
<td></td>
<td>(36)</td>
</tr>
</tbody>
</table>
where $`B \in {\mathbb{R}}^{n \times m}`$, $`C \in {\mathbb{R}}^{l \times n}`$, $`D \in {\mathbb{R}}^{l \times m}`$ are given constant matrices, $`u = {(u_{1}^{T},\cdots,u_{N}^{T})}^{T}`$ is the controller to be designed, and $`y = {(y_{1}^{T},\cdots,y_{N}^{T})}^{T}`$ is the measured output of the LQR problem. As in Sec. 4, let $`s`$ denote an equilibrium point of the individual node and let $`{\eta = {(\eta_{1}^{T},\cdots,\eta_{N}^{T})}^{T}},`$ where $`{{\eta_{i} = {x_{i} - s}},{i = {1,\cdots,N}}},`$ be the state error vector.
Suppose that the controller is a feedback law of the state error, i.e., $`u = {F\hspace{0pt}\eta}`$, where $`F`$ is the feedback gain matrix to be determined. Then, by the Kroneck product, the equation of the linearized system of ([36](#S5.E36 "In 5 Optimal controller design")) about the synchronous solution $`x_{1} = \cdots = x_{N} = s`$ is given as follows:
<table>
<tbody>
<tr>
<td></td>
<td>{η˙=Acη+Bcu,y=Ccη+Dcu,cases˙𝜂subscript𝐴𝑐𝜂subscript𝐵𝑐𝑢𝑦subscript𝐶𝑐𝜂subscript𝐷𝑐𝑢\left\{\begin{array}[]{l}\dot{\eta}=A_{c}\eta+B_{c}u,\\ y=C_{c}\eta+D_{c}u,\end{array}\right.</td>
<td></td>
<td>(37)</td>
</tr>
</tbody>
</table>
where $`A_{c} = {{{I_{N} \otimes D}\hspace{0pt}f\hspace{0pt}{(s)}} - {{\sigma\hspace{0pt}M} \otimes \Gamma}}`$, $`B_{c} = {I_{N} \otimes B}`$, $`C_{c} = {I_{N} \otimes C}`$ and $`D_{c} = {I_{N} \otimes D}`$. Furthermore, suppose that the matrices $`A_{c},B_{c},C_{c}`$ and $`D_{c}`$ satisfy the assumptions (A1-A4) as given in Lemma 2. Then, by Lemma 2, the feedback gain $`F`$ for the LQR problem of the linearized system ([37](#S5.E37 "In 5 Optimal controller design")) is obtained as
<table>
<tbody>
<tr>
<td></td>
<td>F=−(Bc∗X+DcCc),𝐹superscriptsubscript𝐵𝑐𝑋subscript𝐷𝑐subscript𝐶𝑐F=-(B_{c}^{*}X+D_{c}C_{c}),</td>
<td></td>
<td>(38)</td>
</tr>
</tbody>
</table>
where $`X`$ is the stabilizing solution to the following Riccati equation:
<table>
<tbody>
<tr>
<td></td>
<td>(Ac−BcDc∗Cc)∗X+X(Ac−BcDc∗Cc)−XBcBc∗X+Cc∗(Dc)⊥(Dc)⊥∗Cc=0.superscriptsubscript𝐴𝑐subscript𝐵𝑐superscriptsubscript𝐷𝑐subscript𝐶𝑐𝑋𝑋subscript𝐴𝑐subscript𝐵𝑐superscriptsubscript𝐷𝑐subscript𝐶𝑐𝑋subscript𝐵𝑐superscriptsubscript𝐵𝑐𝑋superscriptsubscript𝐶𝑐subscriptsubscript𝐷𝑐bottomsuperscriptsubscriptsubscript𝐷𝑐bottomsubscript𝐶𝑐0(A_{c}-B_{c}D_{c}^{*}C_{c})^{*}X+X(A_{c}-B_{c}D_{c}^{*}C_{c})-XB_{c}B_{c}^{*}X+C_{c}^{*}(D_{c})_{\bot}(D_{c})_{\bot}^{*}C_{c}=0.</td>
<td></td>
<td>(39)</td>
</tr>
</tbody>
</table>
Remark 7: The optimal controller $`u = {F\hspace{0pt}\eta}`$ with $`F`$ given in ([38](#S5.E38 "In 5 Optimal controller design")) has the property that it controls network ([36](#S5.E36 "In 5 Optimal controller design")) to the synchronous state $`x_{1} = \cdots = x_{N} = s`$ and meanwhile minimizes the $`L_{2}`$ norm of the output $`y`$ of the linearized system ([37](#S5.E37 "In 5 Optimal controller design")).
As the node number $`N`$ of network ([36](#S5.E36 "In 5 Optimal controller design")) increases, the direct computation of the feedback gain $`F`$ as given in ([38](#S5.E38 "In 5 Optimal controller design")) will become harder. For the case when the outer matrix $`M`$ is symmetrical, the feedback gain $`F`$ can be alternatively given in terms of the $`N`$ uncoupled subsystems. The main advantage of this approach is that the decentralized feedback gain $`F_{i}`$ will be given only based on the information of the $`i`$th subsystem, $`i = {1,2,\cdots,N}`$.
By using the unitary transformation, as used in Sec. 4, system ([37](#S5.E37 "In 5 Optimal controller design")) becomes
<table>
<tbody>
<tr>
<td></td>
<td>{ξ˙=(IN⊗Df(s)−σΔ⊗Γ)ξ+(IN⊗B)ω,z=(IN⊗C)ξ+(IN⊗D)ω,cases˙𝜉tensor-productsubscript𝐼𝑁𝐷𝑓𝑠tensor-product𝜎ΔΓ𝜉tensor-productsubscript𝐼𝑁𝐵𝜔𝑧tensor-productsubscript𝐼𝑁𝐶𝜉tensor-productsubscript𝐼𝑁𝐷𝜔\left\{\begin{array}[]{l}\dot{\xi}=(I_{N}\otimes Df(s)-\sigma\Delta\otimes\Gamma)\xi+(I_{N}\otimes B)\omega,\\ z=(I_{N}\otimes C)\xi+(I_{N}\otimes D)\omega,\end{array}\right.</td>
<td></td>
<td>(40)</td>
</tr>
</tbody>
</table>
where $`\eta = {{({U \otimes I_{n}})}\hspace{0pt}\xi}`$, $`u = {{({U \otimes I_{n}})}\hspace{0pt}\omega}`$, and $`z = {{({U^{- 1} \otimes I_{l}})}\hspace{0pt}y}`$. The $`N`$ uncoupled controlled subsystems are as follows:
<table>
<tbody>
<tr>
<td></td>
<td>{ξ˙i=Aiξi+Bωi,zi=Cξi+Dωi,i=1,⋯,N,formulae-sequencecasessubscript˙𝜉𝑖subscript𝐴𝑖subscript𝜉𝑖𝐵subscript𝜔𝑖subscript𝑧𝑖𝐶subscript𝜉𝑖𝐷subscript𝜔𝑖𝑖1⋯𝑁\left\{\begin{array}[]{l}\dot{\xi}_{i}=A_{i}\xi_{i}+B\omega_{i},\\ z_{i}=C\xi_{i}+D\omega_{i},\end{array}\qquad i=1,\cdots,N,\right.</td>
<td></td>
<td>(41)</td>
</tr>
</tbody>
</table>
where $`A_{i} = {{D\hspace{0pt}f\hspace{0pt}{(s)}} - {\sigma\hspace{0pt}\lambda_{i}\hspace{0pt}\Gamma}}`$. By the analysis in Sec. 4, one has
<table>
<tbody>
<tr>
<td></td>
<td>‖y‖22=‖z‖22=∑i=1N‖zi‖22.superscriptsubscriptnorm𝑦22superscriptsubscriptnorm𝑧22superscriptsubscript𝑖1𝑁superscriptsubscriptnormsubscript𝑧𝑖22\|y\|_{2}^{2}=\|z\|_{2}^{2}=\sum_{i=1}^{N}\|z_{i}\|_{2}^{2}.</td>
<td></td>
</tr>
</tbody>
</table>
Suppose that the constant matrices $`A_{i},B,C`$ and $`D`$ in ([41](#S5.E41 "In 5 Optimal controller design")) satisfy the assumptions (A1-A4) as given in Lemma 2, for $`{i = {1,\cdots,N}}.`$ Then, the controller $`F_{i}`$ to the $`i`$th subsystem can be given as follows:
<table>
<tbody>
<tr>
<td></td>
<td>Fi:=−(B∗Xi+DC),assignsubscript𝐹𝑖superscript𝐵subscript𝑋𝑖𝐷𝐶F_{i}:=-(B^{*}X_{i}+DC),</td>
<td></td>
<td>(42)</td>
</tr>
</tbody>
</table>
where $`X_{i}`$ is the stabilizing solution to the following Riccati equation:
<table>
<tbody>
<tr>
<td></td>
<td>(Df(s)−λiΓ−BD∗C)∗Xi+Xi(A−λiΓ−BD∗C)−XiBB∗Xi+C∗D⊥D⊥∗C=0,superscript𝐷𝑓𝑠subscript𝜆𝑖Γ𝐵superscript𝐷𝐶subscript𝑋𝑖subscript𝑋𝑖𝐴subscript𝜆𝑖Γ𝐵superscript𝐷𝐶subscript𝑋𝑖𝐵superscript𝐵subscript𝑋𝑖superscript𝐶subscript𝐷bottomsuperscriptsubscript𝐷bottom𝐶0(Df(s)-\lambda_{i}\Gamma-BD^{*}C)^{*}X_{i}+X_{i}(A-\lambda_{i}\Gamma-BD^{*}C)-X_{i}BB^{*}X_{i}+C^{*}D_{\bot}D_{\bot}^{*}C=0,</td>
<td></td>
<td>(43)</td>
</tr>
</tbody>
</table>
for $`i = {1,2,\cdots,N}`$. Moreover, the feedback gain matrix $`F`$ is given by
<table>
<tbody>
<tr>
<td></td>
<td>F=(U⊗Im)[F1⋱FN](U∗⊗In).𝐹tensor-product𝑈subscript𝐼𝑚delimited-[]subscript𝐹1missing-subexpressionmissing-subexpressionmissing-subexpression⋱missing-subexpressionmissing-subexpressionmissing-subexpressionsubscript𝐹𝑁tensor-productsuperscript𝑈subscript𝐼𝑛F=(U\otimes I_{m})\left[\begin{array}[]{ccc}F_{1}&&\\ &\ddots&\\ &&F_{N}\end{array}\right](U^{*}\otimes I_{n}).</td>
<td></td>
<td>(44)</td>
</tr>
</tbody>
</table>
Example 5: Consider Example 1 again. Suppose that the network is with the global coupling configuration. The objective is to design a controller such that network ([11](#S3.E11 "In 3 𝐿₂ norm performance index of network synchronization")) with Chua’s oscillators can be driven to the synchronous solution $`x_{1} = \cdots = x_{N} = s = {\lbrack 0\;0\;0\rbrack}^{T}`$.
It is easy to deduce that
<table>
<tbody>
<tr>
<td></td>
<td>Df(s)=[−α(m2+1)α01−110−β−γ].𝐷𝑓𝑠delimited-[]𝛼subscript𝑚21𝛼01110𝛽𝛾Df(s)=\left[\begin{array}[]{ccc}-\alpha(m_{2}+1)&\alpha&0\\ 1&-1&1\\ 0&-\beta&-\gamma\end{array}\right].</td>
<td></td>
</tr>
</tbody>
</table>
Consequently, the linearized equation of the controlled network is given by
<table>
<tbody>
<tr>
<td></td>
<td>{η˙=(IN⊗Df(s)−σM⊗Γ)η+(IN⊗B)u,y=(IN⊗C)η+(IN⊗D)u,cases˙𝜂tensor-productsubscript𝐼𝑁𝐷𝑓𝑠tensor-product𝜎𝑀Γ𝜂tensor-productsubscript𝐼𝑁𝐵𝑢𝑦tensor-productsubscript𝐼𝑁𝐶𝜂tensor-productsubscript𝐼𝑁𝐷𝑢\left\{\begin{array}[]{l}\dot{\eta}=(I_{N}\otimes Df(s)-\sigma M\otimes\Gamma)\eta+(I_{N}\otimes B)u,\\ y=(I_{N}\otimes C)\eta+(I_{N}\otimes D)u,\end{array}\right.</td>
<td></td>
<td>(45)</td>
</tr>
</tbody>
</table>
where the constant matrices $`B = C = D = I_{3}`$ in this example.
By the above analysis, the LQR optimal controller can be designed by solving the Riccati equation ([39](#S5.E39 "In 5 Optimal controller design")). For comparison, a pinning control strategy is also considered, where $`6`$ nodes are randomly selected to be pinned by controllers $`u_{i} = {- {{\sigma \cdot d \cdot I_{3}}\hspace{0pt}\eta_{i}}}`$ with $`\sigma = 1`$ and $`d = 20`$.
Fig. 3 shows the different performances of the synchronous behavior of the two controlled networks, by the LQR optimal controller and the pinning controller, respectively. The corresponding values of $`{\| y\|}_{2}`$ are listed in Table 5.
LQR optimal control
pinning control
Fig. 3 Different effects of the LQR and pinning controllers.
<table>
<thead>
<tr>
<th>Controller</th>
<th>LQR optimal control</th>
<th>Pinning control</th>
</tr>
</thead>
<tbody>
<tr>
<th>‖y‖2subscriptnorm𝑦2\|y\|_{2}</th>
<td>0.1309</td>
<td>4.9799</td>
</tr>
</tbody>
</table>
Table 5 Values of $`{\| y\|}_{2}`$.
$`\square`$
## 6 Conclusion
In this paper, synchronizability of dynamical networks is considered based on some new measures: the swiftness and overshoot of the network synchronization. The quantity $`{\| e\|}_{2}`$, which represents the $`L_{2}`$ norm of the synchronization error vector $`e\hspace{0pt}{(t)}`$, is taken as the performance index of this kind of synchronizability. It has been shown by several numerical examples, $`{\| e\|}_{2}`$ presents a suitable measure of both swiftness and overshoot of network synchronization: the smaller the values of $`{\| e\|}_{2}`$, the faster with smaller overshoot the network synchronization. For the case when the synchronous state is an equilibrium point, $`{\| e\|}_{2}`$ is upper-bounded by the product of the vector 2-norm of the initial error vector $`e_{0}`$ and the $`H_{2}`$ norm of the transfer function $`G\hspace{0pt}{(s)}`$, denoted as $`{\|{G\hspace{0pt}{(s)}}\|}_{2}`$, of the linearized network about the equilibrium point. The relationship between $`{\|{G\hspace{0pt}{(s)}}\|}_{2}`$ and the network structure has also been discussed. Under some assumptions, it has been proved that $`{\|{G\hspace{0pt}{(s)}}\|}_{2}`$ will not increase as the real eigenvalues of the outer coupling matrix increase. Finally, based on the techniques of the LQR control theory, an optimal controller has been suggested to drive the network onto some homogenous stationary states, which , in the mean time, can minimize the $`L_{2}`$ norm of the output of the linearized network. Further research along this direction seems to be quite promising as long as the network energy and performance are concerned, therefore deserves further efforts.
## References
- \[1\]
M. di Bernardo, F. Garofalo and F. Sorrentino, Synchronizability of degree correlated networks, arXiv: cond-mat/0504335.
- \[2\]
Z. S. Duan, J. Z. Whang, G. R. Chen and L. Huang, Complexity in linearly coupled dynamical networks: some unusual phenomena in energy accumulation, 2007, submitted.
- \[3\]
H. Hong, B. J. Kim, M. Y. Choi and H. Park, Factors that predict better synchronizability on complex networks, Phys. Rev. E, vol. 69, 2004, 067105.
- \[4\]
G. Hu, J. Z. Yang and W. J. Liu, Instability and controllability of linearly coupled oscillators: eigenvalues analysis, Phys. Rev. E, vol. 58, no. 4, 1998, pp. 4440-4453.
- \[5\]
X. Li, X. F. Wang and G. R. Chen, Pinning a complex dynamical network to its equilibrium, IEEE Trans. Circuits Syst.-I , vol. 51, no. 48, 2004, pp. 2074-2087.
- \[6\]
T. Nishikawa, A. E. Motter, Y. C. Duan and F. C. Hoppensteadt, Heterogeneity in oscillator networks: are smaller worlds easier to synchronize? Phys. Rev. Lett., vol. 91, no. 1, 2003, 014101.
- \[7\]
T. Nishikawa and A. E. Motter, Synchronization is optimal in nondiagonalizable networks, Phys. Rev. Lett., vol. 91, no. 1, 2003, 014101.
- \[8\]
X. F. Wang and G. R. Chen, Pinning control of scale-free dynamical networks, Physica A, vol. 310, 2002, pp. 521-531.
- \[9\]
C. W. Wu and L. O. Chua, A unified framework for synchronization and control of dynamical systems, Int. J. Bifurc. Chaos, vol. 4, no. 4, 1994, pp. 979-989.
- \[10\]
C. W. Wu, Perturbation of coupling matrices and its effect on the synchronizability in arrays of coupled chaotic systems, Phys. Lett. A, vol. 319, 2003, pp. 495-503.
- \[11\]
C. W. Wu, Synchronizability of networks of chaotic systems coupled via a graph with a prescribed degree sequence, Phys. Lett. A, vol. 346, 2005, pp. 281-287.
- \[12\]
X. Wu, B. Wang, T. Zhou, W. Wang, M. Zhao and H. Yang, Synchronizability of highly clusterd scale-free networks, Chinese Phys. Lett., vol. 23, no. 4, 2006, pp. 1046-1049.
- \[13\]
K. M. Zhou, J. C. Doyle and K. Glover, Robust and optimal control, Englewood Cliffs: Prentice-Hall, 1996.
- \[14\]
T. Zhou, M. Zhao, and B. H. Wang, Better synchronizability predicted by crossed double cycle, Phys. Rev. E, vol. 73, 2006, 037101. |
0710.5241 | cs.NI | cyber | # Connection Between System Parameters and Localization Probability in Network of Randomly Distributed Nodes
F. Daneshgaran, M. Laddomada, and M. Mondin,
To appear on IEEE Transactions on Wireless Communications, November 2007This work has been supported by NSF grant entitled ”Hourglass: an Infrastructure for Sensor Networks”.Fred Daneshgaran is with the ECE Dept., Calif. State Univ., Los Angeles, USA.Massimiliano Laddomada (laddomada@polito.it) and Marina Mondin are with the Dipartimento di Elettronica, Politecnico di Torino, Italy.
###### Abstract
This article deals with localization probability in a network of randomly distributed communication nodes contained in a bounded domain. A fraction of the nodes denoted as L-nodes are assumed to have localization information while the rest of the nodes denoted as NL nodes do not. The basic model assumes each node has a certain radio coverage within which it can make relative distance measurements. We model both the case radio coverage is fixed and the case radio coverage is determined by signal strength measurements in a Log-Normal Shadowing environment. We apply the probabilistic method to determine the probability of NL-node localization as a function of the coverage area to domain area ratio and the density of L-nodes. We establish analytical expressions for this probability and the transition thresholds with respect to key parameters whereby marked change in the probability behavior is observed. The theoretical results presented in the article are supported by simulations.
###### Index Terms:
Ad-hoc network, connectivity, GPS, iterative localization, LBS, localization, location based services, positioning, probabilistic method, random arrays, range-free localization, sensor networks.
## I Introduction
This article deals with a network of randomly distributed nodes some of which have a-priori knowledge about their position (these are the so called L-nodes) while others are supposed to localize themselves (these are the so called non-localized or NL-nodes).
This general framework can be recognized in many practical scenarios and in different types of communication networks, such as Distributed Sensor Networks (DSNs) and wireless networks. Generally speaking, we may classify the localization algorithms in at least two ways, centralized or distributed \[[1](#bib.bib1)\] and range-free or based on ranging techniques \[[2](#bib.bib2)\]. The most common techniques are based on measured range whereby the location of nodes are estimated through some standard methods such as triangulation. The ranging technique itself can be based on: 1) Received Signal Strength (RSS) measurements, 2) Time Of Arrival (TOA) measurements requiring clock synchronization among nodes, 3) Time Difference Of Arrival (TDOA) techniques requiring relative synchronicity of L-nodes, 4) Angle Of Arrival (AOA) measurements, and 4) hybrid combination of the above.
In range-free localization, connectivity between nodes is a binary event, either two nodes are within communication range of each other or they are not \[[3](#bib.bib3)\]. For simplicity, we may view this as a hard quantization of for instance RSS. If RSS is above a certain detection threshold, the nodes can communicate, otherwise they cannot. Of course, the nature of path loss and the terrain characteristics influences both the coverage radius and the deviation of the coverage zone from the idealistic circular geometry. Various range free algorithms have been proposed in the literature including the centroid algorithm \[[4](#bib.bib4)\], the DV-HOP algorithm \[[5](#bib.bib5)\], the Amorphous positioning algorithm \[[6](#bib.bib6)\], APIT \[[7](#bib.bib7)\], and ROCRSSI \[[2](#bib.bib2)\]. The fundamental characteristic of these algorithms is their ability to provide coarse (compared to ranging techniques) localization information with minimal per node communication and computation requirements.
Regardless of whether ranging based or range-free techniques are used for localization, several global issues of concern remain to be addressed in their application. In particular, this article deals with broader questions of whether a given NL-node can localize itself in a random array of nodes, a percentage of which have localization information. What are the requirements on the density of nodes over a terrain and coverage area of individual nodes that would allow node localization with very high probability?
These questions and others like them are at the heart of the localization problem and by their very nature, require a probabilistic setting \[[8](#bib.bib8)\]. To have a frame of reference, we present a communication scenario which is needed before any localization of NL-nodes can take place. In particular, we assume: 1) ranging techniques are employed and the coverage radius of each node is dependent on the path loss characteristic. While the coverage radius may vary from node to node, we assume that the coverage region is circular. The fact that the coverage geometry may not be perfectly circular has no significant bearing on the calculated probabilities. This is because of our assumption of uniform distribution of nodes on the terrain whereby what matters in so far as probability of finding other nodes in a given area around a reference node is concerned, is the area itself and not its geometric shape; 2) the nodes have calibrated power levels possibly achieved via an external beacon signal, 3) a CSMA/CA link layer protocol is employed for communication among nodes, hence, while collisions might occur, we can assume that the nodes within communication range can communicate without difficulty, and 4) periodic broadcasts by L-nodes are used to inform NL-nodes that can hear them of the coordinate location of the L-nodes relative to some absolute reference frame.
A variety of localization techniques for both indoor and outdoor environments are currently available, offering various trade-offs between accuracy, cost and complexity. In this article we assume that each NL-node needs to communicate with at least three other L-nodes in order to be able to localize itself. A review of various localization techniques proposed in the literature may be found in \[[9](#bib.bib9)\]. In \[[10](#bib.bib10)\], the authors propose an approach based on connectivity information, to derive the locations of nodes in a network. In \[[11](#bib.bib11)\], the authors present some work in the field of source localization in sensor networks.
The rest of the article is organized as follows. In section-II we present the theoretical analysis for the localization probability. Section-III is devoted to presenting the simulation results in support of the theoretical analysis and section-IV to conclusions.
## II Analysis of Localization Probability Using the Probabilistic Method
The basic parameter set defining the problem is as follows.
- •
Total number of nodes distributed uniformly over a circular disk of radius $`R`$ (denoted as the domain) is $`n`$.
- •
$`k`$ of the $`n`$ nodes are assumed to be L-nodes (implying they have localization information relative to some coordinate frame. How this localization is established is irrelevant to problem formulation), the rest are denoted as non-localized nodes (NL-nodes) and need to localize themselves.
- •
The radio coverage radius is denoted as $`d`$ and is defined by $`d = \sqrt{A_{c\hspace{0pt}o\hspace{0pt}v}/\pi}`$ (we assume $`d\operatorname{<<}R`$). In particular, the shape of the coverage area $`A_{c\hspace{0pt}o\hspace{0pt}v}`$ is irrelevant in so far as the calculation of the probabilities is concerned due to the assumption of the uniform distribution of the nodes in the terrain.
- •
The localization problem is two dimensional and three distance measurements relative to nodes with known positions is sufficient to solve for the $`(X,Y)`$ coordinates of the NL-node unambiguously.
- •
We shall neglect the boundary problem in the sense that the nodes near the boundary of the domain can only see other network elements within the domain. Hence in effect, their radio coverage area within which they may identify other nodes is reduced. This assumption is validated to be reasonably good, even when the total number of nodes is not very large.
Various localization techniques have been reported in the literature. The basic strategy assumed in this article is as follows \[[12](#bib.bib12)\].
1. 1\.
The graph corresponding to the set of nodes is created with bidirectional arcs based on range. In particular, any pair of nodes whose true distances are within a certain limit are connected by a bi-directional edge and can communicate with each other.
2. 2\.
A given NL-node that has not localized itself, stores two sets of information generated locally: i) absolute coordinates of the L-nodes within its radio coverage (if any), ii) distance estimates to all other nodes within its radio coverage based on mean power measurements. We shall assume the measured distance information to be accurate.
The basic question that we wish to answer in this article is as follows; given that we are at a NL-node that is interior to our domain, what is the probability that the node can localize itself? To localize itself a NL-node needs at least three L-nodes within its radio coverage. Given the existence of degenerate cases, the formulas presented in the Theorems below really represent lower bounds to the node localization failure probability. However, since the degenerate cases are very rare, the calculated probabilities are very close to the true values (i.e., the bound is tight), and we have verified this experimentally.
Theorem 2.1: Under the assumption of a uniform distribution of $`n`$ nodes, $`k < n`$ of which are L-nodes while the rest are NL-nodes, over a circular domain of radius $`R`$, and per node radio coverage radius of $`d < R`$, the NL-node localization failure probability is tightly lower bounded by:
<table>
<tbody>
<tr>
<td></td>
<td>PF≥{(1−a)22∂2∂a2+(1−a)∂∂a+1}∘{[b2a+(1−b2)]n−1}=[1−(1−a)⋅b2]n−3⋅[1+b2(1−a)(n−3)++b4(1−a)2(n−1)(n−2)2]\small\begin{array}[]{lll}P_{F}&\geq&\left\{\frac{(1-a)^{2}}{2}\frac{\partial^{2}}{\partial a^{2}}+(1-a)\frac{\partial}{\partial a}+1\right\}\circ\left\{[b^{2}a+(1-b^{2})]^{n-1}\right\}\\ &=&[1-(1-a)\cdot b^{2}]^{n-3}\cdot\left[1+b^{2}(1-a)(n-3)+\right.\\ &&\left.+b^{4}(1-a)^{2}\frac{(n-1)(n-2)}{2}\right]\\ \end{array}</td>
<td></td>
<td>(1)</td>
</tr>
</tbody>
</table>
where $`a = {({1 - \frac{k}{n}})}`$ (i.e., the fraction of the NL-nodes), $`b = \left( \frac{d}{R} \right)`$ and the operator notation of vector calculus is used to simplify the expression.
Proof. The assumption of uniform distribution of nodes over our domain implies that the probability of having a node within a circular disk (i.e., the radio coverage zone of a given node) of radius $`d`$ is given by $`\left( \frac{A_{c\hspace{0pt}o\hspace{0pt}v}}{A_{d\hspace{0pt}o\hspace{0pt}m\hspace{0pt}a\hspace{0pt}i\hspace{0pt}n}} \right) = \left( \frac{d}{R} \right)^{2}`$. In $`p`$ independent trials, the probability of having $`p`$ nodes within the coverage zone is simply $`\left( \frac{d}{R} \right)^{2\hspace{0pt}p}`$. Probability of a NL-node localization failure is bounded by:
<table>
<tbody>
<tr>
<td></td>
<td>PF≥P{p≤2or(p≥3and at most two nodesare L-nodes while the rest are NL-nodes)}\begin{array}[]{llll}P_{F}&\geq&P\{p\leq 2\;or\;(p\geq 3\;\mbox{and at most two nodes}\\ &&\mbox{are L-nodes while the rest are NL-nodes})\}\end{array}</td>
<td></td>
<td>(2)</td>
</tr>
</tbody>
</table>
The probability that $`p = 0`$, is the probability that all other $`({n - 1})`$ nodes are outside the coverage zone of the NL-node and is given by $`{\lbrack{1 - b^{2}}\rbrack}^{n - 1}`$. Similarly, the probability that $`p \geq 1`$ nodes are within the coverage radius is:
<table>
<tbody>
<tr>
<td></td>
<td>(n−1p)b2p[1−b2]n−1−p𝑛1𝑝superscript𝑏2𝑝superscriptdelimited-[]1superscript𝑏2𝑛1𝑝\left(\begin{array}[]{c}n-1\\ p\end{array}\right)b^{2p}[1-b^{2}]^{n-1-p}</td>
<td></td>
</tr>
</tbody>
</table>
If in addition, we require that of the $`p`$ nodes at most two be L-nodes while the other $`({p - 2})`$ nodes are NL-nodes, then the probability of the event of interest is:
<table>
<tbody>
<tr>
<td></td>
<td>(n−1p)b2p[1−b2]n−1−p[ap+pap−1(1−a)++p(p−1)2ap−2(1−a)2]\begin{array}[]{ll}\left(\begin{array}[]{c}n-1\\ p\end{array}\right)b^{2p}[1-b^{2}]^{n-1-p}&\left[a^{p}+pa^{p-1}(1-a)+\right.\\ &\left.+\frac{p(p-1)}{2}a^{p-2}(1-a)^{2}\right]\end{array}</td>
<td></td>
</tr>
</tbody>
</table>
Putting the pieces together, after some algebra the NL-node localization failure probability can be written as (assuming $`n \geq 4`$):
<table>
<tbody>
<tr>
<td></td>
<td>PF≥∑p=0(n−1)(n−1p)b2p[1−b2]n−1−p[ap++pap−1(1−a)+p(p−1)2ap−2(1−a)2]\begin{array}[]{lll}P_{F}&\geq&\sum_{p=0}^{(n-1)}\left(\begin{array}[]{c}n-1\\ p\end{array}\right)b^{2p}[1-b^{2}]^{n-1-p}\left[a^{p}+\right.\\ &&\left.+pa^{p-1}(1-a)+\frac{p(p-1)}{2}a^{p-2}(1-a)^{2}\right]\end{array}</td>
<td></td>
<td>(3)</td>
</tr>
</tbody>
</table>
This expression can further be written in the form:
<table>
<tbody>
<tr>
<td></td>
<td>PFsubscript𝑃𝐹\displaystyle P_{F}</td>
<td>≥\displaystyle\geq</td>
<td>(1−a)22∂2∂a2{[b2⋅a+(1−b2)]n−1}+limit-fromsuperscript1𝑎22superscript2superscript𝑎2superscriptdelimited-[]⋅superscript𝑏2𝑎1superscript𝑏2𝑛1\displaystyle\frac{(1-a)^{2}}{2}\frac{\partial^{2}}{\partial a^{2}}\left\{[b^{2}\cdot a+(1-b^{2})]^{n-1}\right\}+</td>
<td></td>
<td rowspan="3">(4)</td>
</tr>
<tr>
<td></td>
<td></td>
<td></td>
<td>+(1−a)∂∂a{[b2⋅a+(1−b2)]n−1}+limit-from1𝑎𝑎superscriptdelimited-[]⋅superscript𝑏2𝑎1superscript𝑏2𝑛1\displaystyle+(1-a)\frac{\partial}{\partial a}\left\{[b^{2}\cdot a+(1-b^{2})]^{n-1}\right\}+</td>
<td></td>
</tr>
<tr>
<td></td>
<td></td>
<td></td>
<td>[b2⋅a+(1−b2)]n−1superscriptdelimited-[]⋅superscript𝑏2𝑎1superscript𝑏2𝑛1\displaystyle\left[b^{2}\cdot a+(1-b^{2})\right]^{n-1}</td>
<td></td>
</tr>
</tbody>
</table>
$`\square`$
When $`{{({1 - a})} \cdot b^{2}}\operatorname{<<}{2/n}`$ we have the approximation:
<table>
<tbody>
<tr>
<td></td>
<td>PF≃1−[(n−3)⋅(1−a)⋅b2]2similar-to-or-equalssubscript𝑃𝐹1superscriptdelimited-[]⋅𝑛31𝑎superscript𝑏22P_{F}\simeq 1-\left[(n-3)\cdot(1-a)\cdot b^{2}\right]^{2}</td>
<td></td>
<td>(5)</td>
</tr>
</tbody>
</table>
Fig. [1](#S2.F1 "Figure 1 ‣ II Analysis of Localization Probability Using the Probabilistic Method ‣ Connection Between System Parameters and Localization Probability in Network of Randomly Distributed Nodes") depicts the localization probability $`({1 - P_{F}})`$ as a function of the fraction of the NL-nodes, $`a`$, for exponentially increasing values of $`b`$ defined via the expression $`{{b\hspace{0pt}{(j)}} = {2^{j/22} - 1}},{j = {3,4,\ldots,20}}`$ in a network containing $`n = 300`$ nodes. It is evident from the figure that the localization probability increases as the fraction of the NL-nodes decreases for a fixed $`b`$, and for a given fraction of NL-nodes, the localization probability increases rapidly as a function of increasing $`b`$.
[Refer to caption](/html/0710.5241/assets/x1.png)
Figure 1: Localization probability as a function of $`a`$ for various values of $`b`$ for a network containing $`n = 300`$ nodes. The leftmost curve is associated with the smallest $`b`$ ($`b = 0.099`$), the rightmost curve is associated with the largest $`b`$ ($`b = 0.878`$), with $`b`$ varying as $`{{b\hspace{0pt}{(j)}} = {2^{j/22} - 1}},{j = {3,4,\ldots,20}}`$. Increasing $`b`$ monotonically increases the localization probability for a given $`a`$.
###
II-A
Extension to Log-Normal Shadowing
Our analysis above on the node localization failure probability may really be taken to represent a probability conditioned on two things: 1) the total number of nodes $`n`$ is known, and 2) the radio coverage area of the NL-node is known. In this subsection, we wish to demonstrate that it is relatively straightforward to remove these conditions. In particular, we shall take a particular propagation model and remove the conditioning on the knowledge of the coverage area. To this end, consider the problem of localizing a specific node by signal power measurements. Let us consider the received power $`P\hspace{0pt}{(d)}`$ at a distance $`d`$ from a specific point $`{P\hspace{0pt}{(d)}} = {{P_{o} - {10\hspace{0pt}n_{p}\hspace{0pt}{\log_{10}{(d)}}}} + X_{s}}`$ whereby $`P_{o}`$ is the signal power at a reference distance $`d_{o}`$ normalized to one for simplicity, $`n_{p}`$ is the path loss exponent, and $`X_{s}`$ is a Gaussian-distributed random variable taking into account the shadowing effect, i.e, $`X_{s} \sim {N\hspace{0pt}{(\mu_{s},\sigma_{s}^{2})}}`$ with $`\mu_{s} = 0`$. An estimator of $`d`$ is given by $`\hat{d} = 10^{\frac{P_{o} - P}{10\hspace{0pt}n_{p}}}`$, where $`P`$ is the measured power. With simple mathematical calculation, it is possible to obtain: $`\hat{d} = 10^{\frac{P_{o} - P}{10\hspace{0pt}n_{p}}} = 10^{{\log_{10}{(d)}} - \frac{X_{s}}{10\hspace{0pt}n_{p}}} = {d \cdot 10^{- \frac{X_{s}}{10\hspace{0pt}n_{p}}}}`$ and then $`{\hat{d}/d} = 10^{- \frac{X_{s}}{10\hspace{0pt}n_{p}}}`$. The random variable $`X_{1} = {X_{s}/n_{p}}`$ is a gaussian random variable $`N\hspace{0pt}{(0,\sigma_{1}^{2})}`$ with zero mean and variance $`\sigma_{1}^{2} = {\sigma_{s}^{2}/n_{p}^{2}}`$. We take $`d`$ to be the true distance that would be measured if $`X_{s} = 0`$. Hence, $`X_{s}`$ behaves as a perturbation affecting the measured distance $`\hat{d}`$. Let the receiver have a detection threshold of $`\gamma`$. Hence, if the received power is below $`\gamma`$, the receiver does not detect the presence of a signal. This puts an upper limit on the estimated distance as $`{\hat{d}}_{m\hspace{0pt}a\hspace{0pt}x} = 10^{\frac{P_{o} - \gamma}{10\hspace{0pt}n_{p}}}`$, leading to an upper limit of $`\frac{\hat{d}}{d} \leq \frac{{\hat{d}}_{m\hspace{0pt}a\hspace{0pt}x}}{d}`$. We note that if the received power is below $`\gamma`$, the transmit and receive nodes do not see each other. Hence the probability of finding a node in the coverage zone of the transmit node is zero. Since our probability values are proportional to the size of the coverage area, we interpret this as the event that $`\hat{d} = 0`$.
With this setup, the ratio between the estimated distance (when measurable) and actual one $`d`$, i.e., $`Y = {\hat{d}/d}`$ is a random variable with log-normal probability density function $`f_{Y}\hspace{0pt}{(y)}`$ (initially we assume $`\gamma = {- \infty}`$ dB). In fact, considering the following transformation applied to the random variable $`X_{1} \sim {N\hspace{0pt}{(0,\sigma_{1}^{2})}}`$; $`Y = {g\hspace{0pt}{(X_{1})}} = 10^{- \frac{X_{1}}{10}}`$ it is simple to see that $`\frac{\hat{d}}{d}`$ is distributed as follows:
<table>
<tbody>
<tr>
<td></td>
<td>fY(y)=fX1(x1)|g′(x1)||x1=g−1(y)=α2πσ12ye−(10⋅log10(y))22σ12f_{Y}(y)=\frac{f_{X_{1}}(x_{1})}{|g^{{}^{\prime}}(x_{1})|}_{|_{x_{1}=g^{-1}(y)}}=\frac{\alpha}{\sqrt{2\pi\sigma_{1}^{2}}y}e^{\frac{-(10\cdot\log_{10}(y))^{2}}{2\sigma_{1}^{2}}}</td>
<td></td>
<td>(6)</td>
</tr>
</tbody>
</table>
whereby $`\alpha = \frac{10}{\ln{(10)}}`$, $`{f_{X_{1}}\hspace{0pt}{(x_{1})}} \sim {N\hspace{0pt}{(0,\sigma_{1}^{2})}}`$, and $`{|{g^{'}\hspace{0pt}{(x_{1})}}|} = {\frac{\ln{(10)}}{10}\hspace{0pt}10^{- \frac{x_{1}}{10}}}`$.
An important parameter in the analysis that follows is the coverage to domain radius ratio. Consider the random variable: $`\frac{\hat{d}}{R} = \hat{b} = {\left( \frac{\hat{d}}{d} \right) \cdot \left( \frac{d}{R} \right)}`$ whereby $`b_{o} = \frac{d}{R}`$ is a fixed constant representing the true coverage to domain radius ratio. With this setup, we have $`{\hat{d}/R} = \hat{b} = {b_{o} \cdot Y}`$. Define:
<table>
<tbody>
<tr>
<td></td>
<td>f1(x)=α2πσ12xe−(10⋅log10(x)−10⋅log10(bo))22σ12subscript𝑓1𝑥𝛼2𝜋superscriptsubscript𝜎12𝑥superscript𝑒superscript⋅10subscript10𝑥⋅10subscript10subscript𝑏𝑜22superscriptsubscript𝜎12f_{1}(x)=\frac{\alpha}{\sqrt{2\pi\sigma_{1}^{2}}x}e^{\frac{-\left(10\cdot\log_{10}(x)-10\cdot\log_{10}(b_{o})\right)^{2}}{2\sigma_{1}^{2}}}</td>
<td></td>
</tr>
</tbody>
</table>
The probability density function (pdf) $`f_{\hat{b}}\hspace{0pt}{(\hat{b})}`$ of $`\hat{b}`$ is:
<table>
<tbody>
<tr>
<td></td>
<td>{α2πσ12b^e−(10⋅log10(b^)−10⋅log10(bo))22σ12++(∫d^maxR∞f1(x)𝑑x)δ(b^),0≤b^≤b^max0b^>b^maxcaseslimit-from𝛼2𝜋superscriptsubscript𝜎12^𝑏superscript𝑒superscript⋅10subscript10^𝑏⋅10subscript10subscript𝑏𝑜22superscriptsubscript𝜎12missing-subexpressionsuperscriptsubscriptsubscript^𝑑𝑚𝑎𝑥𝑅subscript𝑓1𝑥differential-d𝑥𝛿^𝑏0^𝑏subscript^𝑏𝑚𝑎𝑥0^𝑏subscript^𝑏𝑚𝑎𝑥\left\{\begin{array}[]{ll}\frac{\alpha}{\sqrt{2\pi\sigma_{1}^{2}}\widehat{b}}e^{\frac{-\left(10\cdot\log_{10}(\widehat{b})-10\cdot\log_{10}(b_{o})\right)^{2}}{2\sigma_{1}^{2}}}+&\\ +\left(\int_{\frac{\widehat{d}_{max}}{R}}^{\infty}f_{1}(x)dx\right)\delta(\widehat{b}),&0\leq\widehat{b}\leq\widehat{b}_{max}\\ 0&\widehat{b}>\widehat{b}_{max}\end{array}\right.</td>
<td></td>
<td>(7)</td>
</tr>
</tbody>
</table>
Note that $`\hat{b}`$ is a mixed random variable. The probability $`{P\hspace{0pt}{({\hat{b} = 0})}} \neq 0`$. This follows from the argument above that the event that the received signal is below the detection threshold is equivalent to the event that $`\hat{d} = 0`$ implying $`\hat{b} = 0`$. Again, considering the detection threshold, we have the upper limit $`\hat{b} \leq {{\hat{d}}_{m\hspace{0pt}a\hspace{0pt}x}/R}`$. To proceed further, we make the basic assumption that $`\frac{{\hat{d}}_{m\hspace{0pt}a\hspace{0pt}x}}{R} < 1`$ (this is almost always true) and that the coverage region is entirely contained in our domain. Note that the detection threshold $`\gamma`$ and path loss exponent have a major impact on $`{\hat{d}}_{m\hspace{0pt}a\hspace{0pt}x}`$, but, we always assume $`R`$ is large enough that the above condition holds.
Theorem 2.2: Under the assumption of a uniform distribution of $`n`$ nodes, $`k < n`$ of which are L-nodes while the rest are NL-nodes, over a circular domain of radius $`R`$, and per node radio coverage governed by log-normal shadowing with $`\frac{{\hat{d}}_{m\hspace{0pt}a\hspace{0pt}x}}{R} < 1`$, the NL-node localization failure probability $`P_{F}`$ is tightly lower bounded by:
<table>
<tbody>
<tr>
<td></td>
<td>∑l=0n−3(n−3l)(−k1)l{E[b^2l]+k2E[b^2l+2]+k3E[b^2l+4]}superscriptsubscript𝑙0𝑛3𝑛3𝑙superscriptsubscript𝑘1𝑙𝐸delimited-[]superscript^𝑏2𝑙subscript𝑘2𝐸delimited-[]superscript^𝑏2𝑙2subscript𝑘3𝐸delimited-[]superscript^𝑏2𝑙4\small\sum_{l=0}^{n-3}\left(\begin{array}[]{c}n-3\\ l\\ \end{array}\right)\left(-k_{1}\right)^{l}\left\{E\left[\widehat{b}^{2l}\right]+k_{2}E\left[\widehat{b}^{2l+2}\right]+k_{3}E\left[\widehat{b}^{2l+4}\right]\right\}</td>
<td></td>
<td>(8)</td>
</tr>
</tbody>
</table>
where $`a = {({1 - \frac{k}{n}})}`$ (i.e., the fraction of the NL-nodes), $`k_{1} = {1 - a}`$, $`k_{2} = {{({1 - a})} \cdot {({n - 3})}}`$, and $`k_{3} = {{({1 - a})}^{2} \cdot \frac{({{n^{2} - {3\hspace{0pt}n}} + 2})}{2}}`$.
Proof. From Theorem 2.1 the NL-node localization failure probability conditioned on $`\hat{b}`$ can be written as follows:
<table>
<tbody>
<tr>
<td></td>
<td>PF|b^≥[1−k1⋅b^2]n−3⋅[1+k2b^2+k3b^4]subscript𝑃conditional𝐹^𝑏⋅superscriptdelimited-[]1⋅subscript𝑘1superscript^𝑏2𝑛3delimited-[]1subscript𝑘2superscript^𝑏2subscript𝑘3superscript^𝑏4P_{F|\widehat{b}}\geq\left[1-k_{1}\cdot\widehat{b}^{2}\right]^{n-3}\cdot\left[1+k_{2}\widehat{b}^{2}+k_{3}\widehat{b}^{4}\right]</td>
<td></td>
</tr>
</tbody>
</table>
Using the fact that:
<table>
<tbody>
<tr>
<td></td>
<td>[1−k1⋅b^2]n−3=∑l=0n−3(n−3l)(−k1b^2)lsuperscriptdelimited-[]1⋅subscript𝑘1superscript^𝑏2𝑛3superscriptsubscript𝑙0𝑛3𝑛3𝑙superscriptsubscript𝑘1superscript^𝑏2𝑙\left[1-k_{1}\cdot\widehat{b}^{2}\right]^{n-3}=\sum_{l=0}^{n-3}\left(\begin{array}[]{c}n-3\\ l\\ \end{array}\right)\left(-k_{1}\widehat{b}^{2}\right)^{l}</td>
<td></td>
</tr>
</tbody>
</table>
it is possible to obtain the following relation:
<table>
<tbody>
<tr>
<td></td>
<td>PF|b^≥∑l=0n−3(n−3l)(−k1)l⋅[b^2l+k2b^2l+2+k3b^2l+4]subscript𝑃conditional𝐹^𝑏superscriptsubscript𝑙0𝑛3⋅𝑛3𝑙superscriptsubscript𝑘1𝑙delimited-[]superscript^𝑏2𝑙subscript𝑘2superscript^𝑏2𝑙2subscript𝑘3superscript^𝑏2𝑙4P_{F|\widehat{b}}\geq\sum_{l=0}^{n-3}\left(\begin{array}[]{c}n-3\\ l\\ \end{array}\right)\left(-k_{1}\right)^{l}\cdot\left[\widehat{b}^{2l}+k_{2}\widehat{b}^{2l+2}+k_{3}\widehat{b}^{2l+4}\right]</td>
<td></td>
<td>(9)</td>
</tr>
</tbody>
</table>
With this setup, the NL-node localization probability failure $`P_{F}`$ can be lower bounded as follows:
<table>
<tbody>
<tr>
<td></td>
<td>PF=∫0+∞PF|b^⋅fb^(b^)𝑑b^.subscript𝑃𝐹superscriptsubscript0⋅subscript𝑃conditional𝐹^𝑏subscript𝑓^𝑏^𝑏differential-d^𝑏P_{F}=\int_{0}^{+\infty}P_{F|\widehat{b}}\cdot f_{\widehat{b}}(\widehat{b})d\widehat{b}.</td>
<td></td>
</tr>
</tbody>
</table>
By substituting ([9](#S2.E9 "In II-A Extension to Log-Normal Shadowing ‣ II Analysis of Localization Probability Using the Probabilistic Method ‣ Connection Between System Parameters and Localization Probability in Network of Randomly Distributed Nodes")) in the previous equation, it is possible to obtain:
<table>
<tbody>
<tr>
<td></td>
<td>PF≥∑l=0n−3(n−3l)(−k1)l[∫0+∞b^2lfb^(b^)db^++k2∫0+∞b^2l+2fb^(b^)db^+k3∫0+∞b^2l+4fb^(b^)db^]\begin{array}[]{lll}P_{F}&\geq&\sum_{l=0}^{n-3}\left(\begin{array}[]{c}n-3\\ l\\ \end{array}\right)\left(-k_{1}\right)^{l}\left[\int_{0}^{+\infty}\widehat{b}^{2l}f_{\widehat{b}}(\widehat{b})d\widehat{b}+\right.\\ &&\left.+k_{2}\int_{0}^{+\infty}\widehat{b}^{2l+2}f_{\widehat{b}}(\widehat{b})d\widehat{b}+k_{3}\int_{0}^{+\infty}\widehat{b}^{2l+4}f_{\widehat{b}}(\widehat{b})d\widehat{b}\right]\end{array}</td>
<td></td>
<td>(10)</td>
</tr>
</tbody>
</table>
Note that the contribution of the delta function of $`f_{\hat{b}}\hspace{0pt}{(\hat{b})}`$ in above integrals is zero since we are looking at non-central moments. This way a closed form bound for $`P_{F}`$ can be written as follows:
<table>
<tbody>
<tr>
<td></td>
<td>PF≥∑l=0n−3(n−3l)(−k1)l{E[b^2l]++k2E[b^2l+2]+k3E[b^2l+4]}\begin{array}[]{lll}P_{F}&\geq&\sum_{l=0}^{n-3}\left(\begin{array}[]{c}n-3\\ l\\ \end{array}\right)\left(-k_{1}\right)^{l}\left\{E\left[\widehat{b}^{2l}\right]+\right.\\ &&\left.+k_{2}E\left[\widehat{b}^{2l+2}\right]+k_{3}E\left[\widehat{b}^{2l+4}\right]\right\}\end{array}</td>
<td></td>
<td>(11)</td>
</tr>
</tbody>
</table>
$`\square`$
Under certain conditions the expectations above are approximated by the moments of a log-normal random variable $`\hat{b}`$. The k-th order moment of a log-normal random variable is given by,
<table>
<tbody>
<tr>
<td></td>
<td>E[yk]=ekαμ+12(kα)2σ12𝐸delimited-[]superscript𝑦𝑘superscript𝑒𝑘𝛼𝜇12superscript𝑘𝛼2superscriptsubscript𝜎12E\left[y^{k}\right]=e^{\frac{k}{\alpha}\mu+\frac{1}{2}\left(\frac{k}{\alpha}\right)^{2}\sigma_{1}^{2}}</td>
<td></td>
<td>(12)</td>
</tr>
</tbody>
</table>
whereby $`\mu`$ is the mean of the random variable $`10\hspace{0pt}{\log_{10}y}`$. In particular, provided that the number of nodes $`n`$ is small (less than 10) and the variance of shadow variable $`X_{s}`$ is sufficiently small and recalling that $`\alpha = \frac{10}{\ln 10}`$ and $`\mu = {10 \cdot {\log_{10}{(b_{o})}}}`$, we have the approximation:
<table>
<tbody>
<tr>
<td></td>
<td>PF≃∑l=0n−3(n−3l)(−k1)l[e2lαμ+12(2lα)2σ12++k2e2l+2αμ+12(2l+2α)2σ12+k3e2l+4αμ+12(2l+4α)2σ12]\begin{array}[]{ll}P_{F}\simeq&\sum_{l=0}^{n-3}\left(\begin{array}[]{c}n-3\\ l\\ \end{array}\right)\left(-k_{1}\right)^{l}\left[e^{\frac{2l}{\alpha}\mu+\frac{1}{2}\left(\frac{2l}{\alpha}\right)^{2}\sigma_{1}^{2}}+\right.\\ &\left.+k_{2}e^{\frac{2l+2}{\alpha}\mu+\frac{1}{2}\left(\frac{2l+2}{\alpha}\right)^{2}\sigma_{1}^{2}}+k_{3}e^{\frac{2l+4}{\alpha}\mu+\frac{1}{2}\left(\frac{2l+4}{\alpha}\right)^{2}\sigma_{1}^{2}}\right]\end{array}</td>
<td></td>
<td>(13)</td>
</tr>
</tbody>
</table>
The above expression can be written in the alternate form:
<table>
<tbody>
<tr>
<td></td>
<td>PF≃∑l=0n−3(n−3l)(−k1)l[bo2leσ12(2l)237.72++k2bo(2l+2)eσ12(2l+2)237.72+k3bo(2l+4)eσ12(2l+4)237.72]\begin{array}[]{lll}P_{F}&\simeq&\sum_{l=0}^{n-3}\left(\begin{array}[]{c}n-3\\ l\\ \end{array}\right)\left(-k_{1}\right)^{l}\left[b_{o}^{2l}e^{\frac{\sigma_{1}^{2}(2l)^{2}}{37.72}}+\right.\\ &&\left.+k_{2}b_{o}^{(2l+2)}e^{\frac{\sigma_{1}^{2}(2l+2)^{2}}{37.72}}+k_{3}b_{o}^{(2l+4)}e^{\frac{\sigma_{1}^{2}(2l+4)^{2}}{37.72}}\right]\end{array}</td>
<td></td>
<td>(14)</td>
</tr>
</tbody>
</table>
[Refer to caption](/html/0710.5241/assets/x2.png)
Figure 2: Localization probability as a function of $`b_{o}`$ for $`a = 0.2`$ under both Log-Normal shadowing (solid line) and no shadowing (dashed line).
Note that the limiting behavior is consistent in that if $`X_{s}`$ has variance of zero, $`\sigma_{1} = 0`$ and the above expression reduces to the already known result as expressed in ([9](#S2.E9 "In II-A Extension to Log-Normal Shadowing ‣ II Analysis of Localization Probability Using the Probabilistic Method ‣ Connection Between System Parameters and Localization Probability in Network of Randomly Distributed Nodes")).
For a numerical example, consider a scenario with the following parameters: $`\gamma = {- 80}`$ dBm, $`P_{0} = 0`$ dBm, $`d_{0} = 10`$ cm, $`n_{p} = 3.5`$, $`\sigma_{s} = 12`$ dB, $`\sigma_{1} = 3.43`$ dB, $`R = 40`$ m, $`n = 50`$ and $`k = 10`$. The computed $`{\hat{d}}_{m\hspace{0pt}a\hspace{0pt}x} = 19.3`$ m and $`{\hat{b}}_{m\hspace{0pt}a\hspace{0pt}x} = 0.48`$. Fig. [2](#S2.F2 "Figure 2 ‣ II-A Extension to Log-Normal Shadowing ‣ II Analysis of Localization Probability Using the Probabilistic Method ‣ Connection Between System Parameters and Localization Probability in Network of Randomly Distributed Nodes") depicts the localization probability $`({1 - P_{F}})`$ obtained using numerical integration for computation of the Expectations above, as a function of $`b_{o}`$. For comparison purposes, we also show the results obtained from Theorem 2.1 where there is no Log-Normal shadowing. Two immediate observations from the figure are: 1) shadowing actually improves localization probability when $`b_{o}`$ is small, and 2) shadowing is detrimental to localization probability when $`b_{o}`$ is relatively large.
###
II-B
Threshold Conditions
Returning to our analysis where we assume the knowledge of the radio coverage area of a given NL-node, a common characteristic of many problems tackled using the probabilistic method is the existence of a transition threshold where the characteristic of interest exhibits a large variation. The transition threshold observable specially for large values of $`b`$ can be obtained by taking the second partial derivative of $`P_{F}`$ with respect to $`a`$ and setting the result to zero. Given that $`P_{F}`$ is defined through a differential equation, it is better to work directly with this form. This leads to the condition:
<table>
<tbody>
<tr>
<td></td>
<td>(1−a)2∂4∂a4{[b2⋅a+(1−b2)]n−1}=∂3∂a3{[b2⋅a+(1−b2)]n−1}1𝑎2superscript4superscript𝑎4superscriptdelimited-[]⋅superscript𝑏2𝑎1superscript𝑏2𝑛1absentmissing-subexpressionsuperscript3superscript𝑎3superscriptdelimited-[]⋅superscript𝑏2𝑎1superscript𝑏2𝑛1missing-subexpression\begin{array}[]{ll}\frac{(1-a)}{2}\frac{\partial^{4}}{\partial a^{4}}\left\{[b^{2}\cdot a+(1-b^{2})]^{n-1}\right\}=&\\ \frac{\partial^{3}}{\partial a^{3}}\left\{[b^{2}\cdot a+(1-b^{2})]^{n-1}\right\}&\end{array}</td>
<td></td>
<td>(15)</td>
</tr>
</tbody>
</table>
The resulting closed form expression for the transition threshold on $`a`$ is given by:
<table>
<tbody>
<tr>
<td></td>
<td>a∗=1−1b2(0.5n−1)superscript𝑎11superscript𝑏20.5𝑛1a^{*}=1-\frac{1}{b^{2}(0.5n-1)}</td>
<td></td>
<td>(16)</td>
</tr>
</tbody>
</table>
At the value $`a^{\ast}`$, there is a marked change in the localization probability. Values of $`a`$ below threshold lead to a high probability of localization of the NL-nodes. Values of $`a`$ greater than $`a^{\ast}`$ lead to very low probability of one shot localization. The transition threshold is plotted in Fig. [3](#S3.F3 "Figure 3 ‣ III Simulation Results ‣ Connection Between System Parameters and Localization Probability in Network of Randomly Distributed Nodes") as a function of the coverage to domain radius ratio. The noticeable feature is the rapid increase in this threshold at a value of $`b \simeq 0.15`$.
Fig. [4](#S3.F4 "Figure 4 ‣ III Simulation Results ‣ Connection Between System Parameters and Localization Probability in Network of Randomly Distributed Nodes") depicts the localization probability $`({1 - P_{F}})`$ as a function of the coverage to domain radius ratio, $`b`$, for exponentially increasing values of $`a`$ defined via the expression $`{{a\hspace{0pt}{(j)}} = {2^{j/22} - 1}},{j = {1,2,\ldots,20}}`$ in a network containing $`n = 300`$ nodes.
The transition threshold in this scenario can be obtained by taking the second partial derivative of $`P_{F}`$ with respect to $`b`$ and setting the result to zero. This leads to the closed form expression:
<table>
<tbody>
<tr>
<td></td>
<td>b∗={4n2−n−152(1−a)(2n3−8n2+10.5n−4.5)[1++1+6(n−9)(2n3−8n2+10.5n−4.5)(4n2−n−15)2]}1/2\begin{array}[]{ll}b^{*}=\left\{\frac{4n^{2}-n-15}{2(1-a)(2n^{3}-8n^{2}+10.5n-4.5)}\left[1+\right.\right.&\\ \left.\left.+\sqrt{1+\frac{6(n-9)(2n^{3}-8n^{2}+10.5n-4.5)}{(4n^{2}-n-15)^{2}}}\right]\right\}^{1/2}&\end{array}</td>
<td></td>
<td>(17)</td>
</tr>
</tbody>
</table>
For large values of $`n`$, the above expression simplifies to:
<table>
<tbody>
<tr>
<td></td>
<td>b∗≃{1(1−a)n[1+1.75]}1/2similar-to-or-equalssuperscript𝑏superscript11𝑎𝑛delimited-[]11.7512b^{*}\simeq\left\{\frac{1}{(1-a)n}\left[1+\sqrt{1.75}\right]\right\}^{1/2}</td>
<td></td>
<td>(18)</td>
</tr>
</tbody>
</table>
For a fixed $`a`$, there is a transition threshold with respect to $`b`$. At values of $`b`$ below $`b^{\ast}`$, the localization probability is small while for values of $`b`$ greater than $`b^{\ast}`$ the localization probability increases rapidly. The transition threshold is plotted in Fig. [5](#S3.F5 "Figure 5 ‣ III Simulation Results ‣ Connection Between System Parameters and Localization Probability in Network of Randomly Distributed Nodes") as a function of the percentage of the NL-nodes. The noticeable feature is the relatively gradual increase in this threshold as a function of $`a`$.
So far we have looked at what we may call a one-shot localization of the NL-nodes. What we mean is that our calculated localization probability looks at the event that a given NL-node finds at least three L-nodes within its radio coverage. Once a NL-node localizes itself, technically, it could become a reference localization node and change from the category of NL-node, to the category of L-node. The process of NL-node localization can then repeat itself with the potentially newly added L-nodes. The study of this iterative localization process is beyond the scope of the present work. The asymptotic behavior however in the limiting case of infinite iterations is relatively easy to deduce. In particular, there is always a non-zero probability that at least one node may not get localized even if all the other nodes are L nodes. This provides the lower bound:
<table>
<tbody>
<tr>
<td></td>
<td>PF(∞)≥∑k=02(n−1k)b2k[1−b2]n−1−k.subscript𝑃𝐹superscriptsubscript𝑘02𝑛1𝑘superscript𝑏2𝑘superscriptdelimited-[]1superscript𝑏2𝑛1𝑘P_{F}(\infty)\geq\sum_{k=0}^{2}\left(\begin{array}[]{c}n-1\\ k\end{array}\right)b^{2k}[1-b^{2}]^{n-1-k}.</td>
<td></td>
</tr>
</tbody>
</table>
## III Simulation Results
We have simulated the node localization problem as outlined in the previous section. The details of our simulation setup is as follows.
[Refer to caption](/html/0710.5241/assets/x3.png)
Figure 3: Probability transition threshold $`a^{\ast}`$ as a function of $`b`$ for a network containing $`n = 300`$ nodes.
- •
We consider uniformly distributing $`n`$ nodes over a domain of radius $`R`$ normalized to $`R = 1`$. To achieve this, we generate $`n`$ pairs of uniformly distributed independent random variables $`(r_{i},\theta_{i})`$. Random variable $`r_{i}`$ is uninform in the interval $`\lbrack 0,1\rbrack`$ while $`\theta_{i}`$ is uniform in the interval $`\lbrack{- \pi},\pi\rbrack`$. To get a uniform distribution of nodes over the domain, the variable $`r_{i}`$ is transformed to $`\sqrt{r_{i}}`$ such that the pair $`(\sqrt{r_{i}},\theta_{i})`$ denotes the polar coordinates of the $`i`$-th node within our domain.
- •
$`k`$ of the $`n`$ nodes are identified as the L-nodes by selecting at random a set of $`k`$ unique indices in the range $`\{ 1,2,\ldots,n\}`$.
- •
For a given NL-node, the set of distances to the $`k`$ L-nodes are calculated and the number of L-nodes (if any) that fall within the coverage radius $`d`$ is obtained. This is repeated for all the NL-nodes and the count of the total number of NL-nodes that can localize themselves is obtained.
- •
To get good statistical averages, we have generated 1000 realizations of the localization problem for a given set of parameters, $`n,k,a,b`$. We have then obtained the empirical estimate of node localization failure probability by averaging the results over 1000 realizations.
[Refer to caption](/html/0710.5241/assets/x4.png)
Figure 4: Localization probability as a function of $`b`$ for various values of $`a`$ in the set $`{{a\hspace{0pt}{(j)}} = {2^{j/22} - 1}},{j = {1,2,\ldots,20}}`$ for a network containing $`n = 300`$ nodes. The leftmost curve is associated with the smallest $`a`$. Increasing $`a`$ monotonically decreases the localization probability for a given $`b`$.
As an example of sample simulation results while assessing the impact of the density of nodes over the domain, Fig. [6](#S4.F6 "Figure 6 ‣ IV Conclusions ‣ Connection Between System Parameters and Localization Probability in Network of Randomly Distributed Nodes") depicts the NL-node localization probability versus $`a`$ for three different values of $`n`$ at a value of $`b = 0.05`$. Simulation points are marked by ”\*” while the analytical curves obtained using the formulas derived in the previous section are shown as solid lines. Note that at $`n = 500`$ the curve is convex, at $`n = 1000`$ it is almost linear and at $`n = 3000`$ it is almost concave. The other noticeable effect is the reduction in the localization probability for a fixed $`a`$ as the density of the nodes within the domain is reduced. The gap between theory and simulation can be attributed to the boundary points which as stated previously, have an effectively smaller coverage area than the interior points.
[Refer to caption](/html/0710.5241/assets/x5.png)
Figure 5: Probability transition threshold $`b^{\ast}`$ as a function of $`a`$ for a network containing $`n = 300`$ nodes.
## IV Conclusions
In this article we have presented a probabilistic setup for NL-node localization in a randomly distributed network of mixed localized and non-localized nodes. The probabilistic method is then used to answer some fundamental questions regarding the feasibility of NL-node localization in such a network based on some basic parameters such as the density of the nodes and coverage to domain radius ratios. We have derived expressions for transition thresholds with respect to several key parameters whereby marked change in the localization probability is observed.
[Refer to caption](/html/0710.5241/assets/x6.png)
Figure 6: Simulated and theoretical $`({1 - P_{F}})`$ as a function of percentage of NL-nodes $`a`$ (simulated points are marked by ”\*”) for three different values of $`n`$ at $`b = 0.05`$.
## References
- \[1\]
N. Patwari, J. N. Ash, S. Kyperountas, A. O. Hero III, R. L. Moses, and N. S. Correal, “Locating the nodes: Cooperative localization in wireless sensor networks,” IEEE Signal Proc. Magazine, pp. 54–69, July 2005.
- \[2\]
C. Liu and K. Wu, “Performance evaluation of range-free localization schemes for wireless sensor networks,” Proc. of IEEE Intern. Perf. Computing and Commun. Conf., pp. 59–66, 2005.
- \[3\]
N. Patwari and A. O. Hero III, “Using proximity and quantized RSS for sensor localization in wireless networks,” Proc. IEEE/ACM 2nd Workshop on Wireless Sensor Networks and Applications, pp. 20–29, Sept. 2003.
- \[4\]
N. Bulusu, J. Heidemann, and D. Estrin, “GPS-less low cost outdoor localization for very small devices ,” IEEE Personal Commun. Magazine, Vol.7, No.5, pp. 28–34, Oct. 2000.
- \[5\]
D. Niculescu and B. Nath, “DV based positioning in ad hoc networks,” J. of Telecomm. Systems, Vol.1, 2003.
- \[6\]
R. Nagpal, “Organizing a global positioning system from local information on an amorphous computer,” A.I. Memo1666, MIT A.I. Laboratory, August 1999.
- \[7\]
T. He, C. Huang, B.M. Blum, J.A. Stankovic, and T. Abdelzaher, “Range-free localization schemes for large scale sensor networks,” Proc. MobiCom 2003, pp. 81-95, Sept. 2003.
- \[8\]
N. Alon and J. H. Spencer, The Probabilistic Method, Second Edition, John Wiley & Sons, ISBN 0-471-37046-0, 2000.
- \[9\]
N. Bulusu, J. Heidemann, and D. Estrin, “GPS-less low-cost outdoor localization for very small devices,” IEEE Personal Communications, Vol. 7, No. 5, pp. 28 - 34, Oct. 2000.
- \[10\]
Yi Shang, W. Ruml, Ying Zhang, and M. Fromherz, “Localization from connectivity in sensor networks,” IEEE Transactions on Parallel and Distributed Systems, Vol. 15, No. 11, pp. 961 - 974, Nov. 2004.
- \[11\]
J.C. Chen, Kung Yao, and R.E. Hudson, “Source localization and beamforming,” IEEE Signal Processing Magazine, Vol. 19, No. 2, pp. 30 - 39, March 2002.
- \[12\]
V. Crespi and G. Cybenko, “Decentralized algorithms for sensor registration,” Proc. Intern. Joint Conf. on Neural Networks, IJCNN2003, Vol.1, pp.266-271, Portlane, Oregon, July 2003. |
0710.0270 | cs.NI | cyber | # An Analytical Study of a Structured Overlay in the Presence of Dynamic Membership
<sup>†</sup>
<sup>†</sup>thanks: This work is funded by the European 6th FP EVERGROW project. ©IEEE. Personal use of this material is permitted. However, permission to reprint/republish this material for advertising or promotional purposes or for creating new collective works for resale or redistribution to servers or lists, or to reuse any copyrighted component of this work in other works, must be obtained from the IEEE.
Supriya Krishnamurthy<sup>1,3</sup>, Sameh El-Ansary<sup>1</sup>, Erik Aurell<sup>1,2</sup> and Seif Haridi<sup>1,3</sup>
<sup>1</sup> Swedish Institute of Computer Science (SICS), Sweden
<sup>2</sup> Department of Physics, KTH-Royal Institute of Technology, Sweden
<sup>3</sup> IMIT, KTH-Royal Institute of Technology, Sweden
{supriya,sameh,eaurell,seif}@sics.se
###### Abstract
In this paper we present an analytical study of dynamic membership (aka churn) in structured peer-to-peer networks. We use a fluid model approach to describe steady-state or transient phenomena, and apply it to the Chord system. For any rate of churn and stabilization rates, and any system size, we accurately account for the functional form of the probability of network disconnection as well as the fraction of failed or incorrect successor and finger pointers. We show how we can use these quantities to predict both the performance and consistency of lookups under churn. All theoretical predictions match simulation results. The analysis includes both features that are generic to structured overlays deploying a ring as well as Chord-specific details, and opens the door to a systematic comparative analysis of, at least, ring-based structured overlay systems under churn.
## I Introduction
An intrinsic property of Peer-to-Peer systems is the process of never-ceasing dynamic membership. Structured Peer-to-Peer Networks (aka Distributed Hash Tables (DHTs)) have the underlying principle of arranging nodes in an overlay graph of known topology and diameter. This knowledge results in the provision of performance guarantees. However, dynamic membership continuously “corrupts/churns” the overlay graph and every DHT strives to provide a technique to “correct/maintain” the graph in the face of this perturbation.
Both theoretical and empirical studies have been conducted to analyze the performance of DHTs undergoing “churn” and simultaneously performing “maintenance”. Liben-Nowell et al. \[[11](#bib.bib11)\] prove a lower bound on the maintenance rate required for a network to remain connected in the face of a given dynamic membership rate. Aspnes et al. \[[3](#bib.bib3)\] give upper and lower bounds on the number of messages needed to locate a node/data item in a DHT in the presence of node or link failures. The value of such theoretical studies is that they provide insights neutral to the details of any particular DHT. Empirical studies have also been conducted to complement these theoretical studies by showing how within the asymptotic bounds, the performance of a DHT may vary substantially depending on different DHT designs and implementation decisions. Examples include the work of: Li et al. \[[10](#bib.bib10)\], Rhea et al. \[[14](#bib.bib14)\], and Rowstron et al. \[[5](#bib.bib5)\].
In this paper, we present a fluid model of Chord \[[15](#bib.bib15)\], a specific DHT, under churn. Fluid models have been used to model data communication systems at least since the early ’80ies \[[2](#bib.bib2)\], and in some sense since the work of Erlang \[[4](#bib.bib4)\]. More recently, in the context of P2P systems, it has been used to model the performance of BitTorrent \[[13](#bib.bib13)\] and the Squirrel caching system \[[6](#bib.bib6)\]. This technique has much in common with macroscopic and mesoscopic descriptions of physical and chemical phenomena (from where the term fluid has obviously been borrowed), and carries the same advantages of conciseness and computability relative to an underlying more exact description. Our analysis is directly based on the master equation approach of physical kinetics, see e.g. the text book \[[12](#bib.bib12)\], which provides a scheme for taking the various dynamical processes involved systematically into account.
The fluid model requires the notion of a state of the system. This is just a listing of the quantities one would need to know for a description of the system at a given level of detail. For Chord, we use grosso modo a level of description which requires keeping track of how many nodes there are in the system and what the state (whether correct, incorrect or failed) of each of the pointers of those nodes is. This information is not enough to draw a unique graph of network-connections because, for example, if we know that a given node has an ’incorrect’ successor pointer, this still does not tell us which node it is pointing to. However, as we will see, beginning at this level of description is sufficient to keep track of most of the details of the Chord protocols. Having defined a state, the fluid model is simply a set of equations for the evolution of the probability of finding the system in this state, given the details of the dynamics. The master equation approach is useful for keeping track of the contribution of all the events which can bring about changes in the probability in a micro-instant of time i.e., evaluating all the terms in the dynamics leading to a gain or loss of this probability.
Using this formalism we investigate a probabilistic model in which peers arrive independently, distributed as a Poisson process, and life-times are exponentially distributed. While this setup is not necessary fully realistic (more realistic models can also be analyzed using master equation techniques), it is standard in modeling, as it typically brings out the salient features of the system with as few obscuring details from the probabilistic model as possible. We then derive the functional forms of the following: $`(i)`$ Chord-specific inter-node distribution properties and $`({i\hspace{0pt}i})`$ for every outgoing pointer of a Chord node, the probability that it is in any one of its possible states. This probability is different for each of the successor and finger pointers. We then use this information to predict other quantities such as $`({i\hspace{0pt}i\hspace{0pt}i})`$ the probability that the network gets disconnected, $`({i\hspace{0pt}v})`$ lookup consistency (number of failed lookups), and $`(v)`$ lookup performance (latency). All quantities are computed as a function of the parameters involved and all results are verified by simulations.
## II Related Work
Closest in spirit to our work is the informal derivation in the original Chord paper \[[15](#bib.bib15)\] of the average number of timeouts encountered by a lookup. This quantity was approximated there by the product of the average number of fingers used in a lookup times the probability that a given finger points to a departed node. Our methodology not only allows us to derive the latter quantity systematically but also demonstrates how this probability depends on which finger (or successor) is involved. Further we are able to derive a precise relation relating this probability to lookup performance and consistency accurately at any value of the system parameters.
In the works of Aberer et al. \[[1](#bib.bib1)\] and Wang *e*t al. \[[16](#bib.bib16)\], DHTs are analyzed under churn and the results are compared with simulations. These analyses can also be classified as fluid models. However the main parameter is the probability that a random selected entry of a routing table is stale. In our analysis, we determine this quantity from system details and churn rates.
A brief announcement of the results presented in this paper, has appeared earlier in \[[8](#bib.bib8)\].
## III Our Implementation of Chord
The Chord Ring. The general philosophy of DHTs is to map a set of data items onto a set of nodes where the insertion and lookup of items is done using the unique keys that the items are given. Chord’s realization of that philosophy is as follows. Peers and data items are given unique keys (usually obtained by a cryptographic hash of unique attribute like the IP address or public key for nodes, and filename or checksum for items) drawn from a circular key space of size $`\mathcal{K}`$. The Chord system dictates that the right place for storing an item is at the first alive node whose key succeeds the key of the item. Since we refer to nodes and items by their keys, the insertion and lookup of items becomes a matter of locating the right “successor” of a key. All nodes have successor and predecessor pointers. For $`N`$ nodes, using only the successor pointers to lookup items requires $`\frac{1}{2}\hspace{0pt}N`$ hops on average.
Fingers. To reduce the average lookup path length, nodes keep $`\mathcal{M} = {\log_{2}\mathcal{K}}`$ pointers known as the “fingers”. Using these fingers, a node can retrieve any key in $`O\hspace{0pt}{({\log N})}`$ hops. The fingers of a node $`n`$ (where $`n \in {{0\hspace{0pt}\cdots\hspace{0pt}\mathcal{K}} - 1}`$) point to exponentially increasing distances of keys away from $`n`$. That is, $`\forall i \in 1..\mathcal{M}`$, $`n`$ points to a node whose key is equal to $`n + 2^{i - 1}`$. We denote that key by $`n.{f\hspace{0pt}i\hspace{0pt}n_{i}}.{s\hspace{0pt}t\hspace{0pt}a\hspace{0pt}r\hspace{0pt}t}`$. However, for a certain $`i`$, there might not be a node in the network whose key is equal to $`n + 2^{i - 1}`$. Therefore, $`n`$ points to the first successor of $`n + 2^{i - 1}`$ which we denote by $`n.{f\hspace{0pt}i\hspace{0pt}n_{i}}.{n\hspace{0pt}o\hspace{0pt}d\hspace{0pt}e}`$.
The Successor List Moreover, each node keeps a list of the $`\mathcal{S} = {O\hspace{0pt}{({\log{(N)}})}}`$ immediate successors as backups for its first successor. We use the notation $`n.s`$ to refer to this list and $`n.s_{i}`$ to refer to the $`i^{t\hspace{0pt}h}`$ element in the list. Finally we use the notation $`n.p`$ to refer to the predecessor.
Stabilization, Churn & Steady State. To keep the pointers up-to-date in the presence of churn, each node performs periodic stabilization of its successors and fingers. In our analysis, we define $`\lambda_{j}`$ as the rate of joins per node, $`\lambda_{f}`$ the rate of failures per node and $`\lambda_{s}`$ the rate of stabilizations per node. The fraction of stabilizations which act on the successors is $`\alpha`$, such that the rate of successor stabilizations is $`\alpha\hspace{0pt}\lambda_{s}`$, and the rate of finger stabilizations is $`{({1 - \alpha})}\hspace{0pt}\lambda_{s}`$. In all that follows, we impose the steady state condition $`\lambda_{j} = \lambda_{f}`$ unless otherwise stated. Further it is useful to define $`r \equiv \frac{\lambda_{s}}{\lambda_{f}}`$ which is the relevant ratio on which all the quantities we are interested in will depend, e.g, $`r = 50`$ means that a join/fail event takes place every half an hour for a stabilization which takes place once every $`36`$ seconds. Throughout the paper we will use the terms $`\lambda_{j}\hspace{0pt}N\hspace{0pt}\Delta\hspace{0pt}t`$, $`\lambda_{f}\hspace{0pt}N\hspace{0pt}\Delta\hspace{0pt}t`$, $`\alpha\hspace{0pt}\lambda_{s}\hspace{0pt}N\hspace{0pt}\Delta\hspace{0pt}t`$ and $`{({1 - \alpha})}\hspace{0pt}\lambda_{s}\hspace{0pt}N\hspace{0pt}\Delta\hspace{0pt}t`$ to denote the respective probabilities that a join, failure, a successor stabilization, or a finger stabilization take place anywhere on the ring during a micro period of time of length $`\Delta\hspace{0pt}t`$.
Parameters. The parameters of the problem are hence: $`\mathcal{K}`$, $`N`$, $`\alpha`$ and $`r`$. All relevant measurable quantities should be entirely expressible in terms of these parameters.
Simulation Since we are collecting statistics like the probability of a particular finger pointer to be wrong, we need to repeat each experiment $`100`$ times before obtaining well-averaged results. The total simulation sequential real time for obtaining the results of this paper was about $`1800`$ hours that was parallelized on a cluster of $`14`$ nodes where we had $`N = 1000`$, $`\mathcal{K} = 2^{20}`$, $`\mathcal{S} = 6`$, $`200 \leq r \leq 2000`$ and $`0.25 \leq \alpha \leq 0.75`$.
While the main outlines of the chord protocol are provided by its authors in \[[15](#bib.bib15)\], an exact analysis necessitates the provision of a deeper level of detail and adopted assumptions which we provide in the following subsections.
###
III-A
Joins, Failures & Ring Stabilization
Initialization. Initially, a node knows its key and at least one node with key $`c`$ that already exists in the network and is alive. The knowledge of such a node is assumed to be acquired through some out-of-band method. The predecessor $`p`$, successors ($`s_{1..\mathcal{S}}`$) and fingers ($`{f\hspace{0pt}i\hspace{0pt}n_{1..\mathcal{M}}}.{n\hspace{0pt}o\hspace{0pt}d\hspace{0pt}e}`$) are all assigned to $`n\hspace{0pt}i\hspace{0pt}l`$.
Joins (Fig. [1](#S3.F1 "Figure 1 ‣ III-A Joins, Failures & Ring Stabilization ‣ III Our Implementation of Chord ‣ An Analytical Study of a Structured Overlay in the Presence of Dynamic Membership This work is funded by the European 6th FP EVERGROW project. ©IEEE. Personal use of this material is permitted. However, permission to reprint/republish this material for advertising or promotional purposes or for creating new collective works for resale or redistribution to servers or lists, or to reuse any copyrighted component of this work in other works, must be obtained from the IEEE.")). A new node $`n`$ joins by looking up its successor using the initial random contact node $`c`$. It also starts its first stabilization of the successors and initializes its fingers.
Stabilization of Successors (Fig. [1](#S3.F1 "Figure 1 ‣ III-A Joins, Failures & Ring Stabilization ‣ III Our Implementation of Chord ‣ An Analytical Study of a Structured Overlay in the Presence of Dynamic Membership This work is funded by the European 6th FP EVERGROW project. ©IEEE. Personal use of this material is permitted. However, permission to reprint/republish this material for advertising or promotional purposes or for creating new collective works for resale or redistribution to servers or lists, or to reuse any copyrighted component of this work in other works, must be obtained from the IEEE.")). The function fixSuccessors is triggered periodically with rate $`\alpha\hspace{0pt}\lambda_{s}`$. A node $`n`$ tells its first alive successor $`y`$ that it believes itself to be $`y`$’s predecessor and expects as an answer $`y`$’s predecessor $`y.p`$ and successors $`y.s`$. The response of $`y`$ can lead to three actions:
Case A. Some node exists between $`n`$ and $`y`$ (i.e., $`n`$’s belief is wrong), so $`n`$ prepends $`y.p`$ to its successor list as a first successor and retries *fixSuccessors*.
Case B. $`y`$ confirms $`n`$’s belief and informs $`n`$ of $`y`$’s old predecessor $`y.p`$. Therefore $`n`$ considers $`y.p`$ as an alternative/initial predecessor for $`n`$. Finally, $`n`$ reconciles its successor list with $`y.s`$.
Case C. $`y`$ agrees that $`n`$ is its predecessor and the only task of $`n`$ is to update its successor list by reconciling it with $`y.s`$.
By calling *iThinkIamYourPred* (Fig. [1](#S3.F1 "Figure 1 ‣ III-A Joins, Failures & Ring Stabilization ‣ III Our Implementation of Chord ‣ An Analytical Study of a Structured Overlay in the Presence of Dynamic Membership This work is funded by the European 6th FP EVERGROW project. ©IEEE. Personal use of this material is permitted. However, permission to reprint/republish this material for advertising or promotional purposes or for creating new collective works for resale or redistribution to servers or lists, or to reuse any copyrighted component of this work in other works, must be obtained from the IEEE.")), some node $`x`$ informs $`n`$ that it believes itself to be $`n`$’s predecessor. If $`n`$’s predecessor $`p`$ is not alive or $`n\hspace{0pt}i\hspace{0pt}l`$, then $`n`$ accepts $`x`$ as a predecessor and informs $`x`$ about this agreement by returning $`x`$. Alternatively, if $`n`$’s predecessor $`p`$ is alive (discovering that will be explained shortly in section [](#S3.SS3 "III-C Failures ‣ III Our Implementation of Chord ‣ An Analytical Study of a Structured Overlay in the Presence of Dynamic Membership This work is funded by the European 6th FP EVERGROW project. ©IEEE. Personal use of this material is permitted. However, permission to reprint/republish this material for advertising or promotional purposes or for creating new collective works for resale or redistribution to servers or lists, or to reuse any copyrighted component of this work in other works, must be obtained from the IEEE.")
III-C
), then there are two possibilities: The first is that $`x`$ is in the region between $`n`$ and its current predecessor $`p`$, therefore $`n`$ should accept $`x`$ as a new predecessor and inform $`x`$ about its old predecessor. The second is that $`p`$ is already pointing to $`x`$ so the state is correct at both parties and $`n`$ confirms that to $`x`$ by informing it that $`x`$ is the predecessor of $`n`$. In all cases the function returns a predecessor and a successor list.
The function *firstAliveSuccessor* (Fig. [1](#S3.F1 "Figure 1 ‣ III-A Joins, Failures & Ring Stabilization ‣ III Our Implementation of Chord ‣ An Analytical Study of a Structured Overlay in the Presence of Dynamic Membership This work is funded by the European 6th FP EVERGROW project. ©IEEE. Personal use of this material is permitted. However, permission to reprint/republish this material for advertising or promotional purposes or for creating new collective works for resale or redistribution to servers or lists, or to reuse any copyrighted component of this work in other works, must be obtained from the IEEE.")) iterates through the successor list. In each iteration, if the first successor $`s_{1}`$ is alive, it is returned. Otherwise, the dead successor is dropped from the list and nil is appended to the end of the list. If the first successor is *nil* this means that all immediate successors are dead and that the ring is disconnected.
$`n`$
.join($`c`$)
$`s_{1}`$ = $`c`$.findSuccessor($`n`$)
fixSuccessors()
initFingers($`s_{1}`$)
$`n`$
.fixSuccessors()
$`y =`$ firstAliveSuccessor()
$`\left\{ {y.{p,y}.s} \right\}`$ = $`y`$.iThinkIamYourPred($`n`$)
if ($`y.{p \in \left( {m\hspace{0pt}e},y \right)}`$) //Case A
prepend($`y.p`$)
fixSuccessors()
elsif ($`y.{p \in \left( y,{m\hspace{0pt}e} \right)}`$) //Case B
considerANewPred($`y.p`$)
reconcilce($`y.s`$)
else //Case C: $`y.p = = me`$
reconcile($`y.s`$)
$`n`$
.firstAliveSuccessor()
while (true)
if ($`s_{1} = = nil`$)
//Broken Ring!!
if (isAlive($`s_{1}`$))
return ($`s_{1}`$)
$`\forall i \in 1..(\mathcal{S} - 1)`$
$`s_{i} = s_{i + 1}`$
$`s_{S} = {n\hspace{0pt}i\hspace{0pt}l}`$
$`n`$
.iThinkIAmYourPred($`x`$)
if ((isNotAlive($`p`$) or ($`p = = nil`$))
$`p = x`$
return($`\left\{ s,x \right\}`$)
if ($`x \in \left( p,{m\hspace{0pt}e} \right)`$)
$`{o\hspace{0pt}l\hspace{0pt}d\hspace{0pt}p} = p`$
$`p = x`$
return($`\left\{ s,{o\hspace{0pt}l\hspace{0pt}d\hspace{0pt}p} \right\}`$)
else
return($`\left\{ s,p \right\}`$)
$`n`$
.considerANewPred($`x`$)
if (isNotAlive($`p`$)
or ($`p = = nil`$)
or ($`x \in (p,n)`$))
$`p = x`$
$`n`$
.reconcile($`s'`$)
for $`i = 1..(\mathcal{S} - 1)`$
$`s_{i + 1} = s_{i}'`$
$`n`$
.prepend($`y`$)
for $`i = {\mathcal{S}\hspace{0pt}..2}`$
$`s_{i} = s_{i - 1}`$
$`s_{1} = y`$
Figure 1: Joins and Ring Stabilization Algorithms.
###
III-B
Lookups and Stabilization of Fingers
Stabilization of Fingers (Fig. LABEL:fig:fixfingers). Stabilization of fingers occurs at a rate $`{({1 - \alpha})}\hspace{0pt}\lambda_{s}`$. Each time the *fixFingers* function is triggered, a random finger $`f\hspace{0pt}i\hspace{0pt}n_{i}`$ is chosen and a lookup for $`{f\hspace{0pt}i\hspace{0pt}n_{i}}.{s\hspace{0pt}t\hspace{0pt}a\hspace{0pt}r\hspace{0pt}t}`$ is performed and the result is used to update $`{f\hspace{0pt}i\hspace{0pt}n_{i}}.{n\hspace{0pt}o\hspace{0pt}d\hspace{0pt}e}`$.
$`n`$
.findSuccessor($`k`$)
//Case A: $`k`$ is exactly equal to $`n`$
if ($`k`$ == $`n`$)
return($`n`$)
//Case B: $`k`$ is between $`n`$ and $`s_{1}`$
if ($`k \in \left( n,s_{1} \right\rbrack`$)
return(firstAliveSuccessorNoChange());
//Case C: Forward to the lookup to
//the closest preceding alive finger
$`c\hspace{0pt}p\hspace{0pt}f`$ = closestAlivePrecedingFinger($`k`$);
if ($`cpf = = nil`$)
$`y`$ = firstAliveSuccessorNoChange();
if ($`k \in (n,y\rbrack`$)
return($`y`$);
$`c\hspace{0pt}p\hspace{0pt}f`$ = closestAlivePrecedingSucc(k);
return($`c\hspace{0pt}p\hspace{0pt}f`$.findSuccessor(k))
else
return ($`c\hspace{0pt}p\hspace{0pt}f`$.findSuccessor(k));
$`n`$
.firstAliveSuccessorNoChange()
$`i = 1`$
while (true)
if ($`s_{i} = = nil`$)
//Broken Ring!!
if (isAlive($`s_{i}`$))
return ($`s_{i}`$)
$`i + +`$
$`n`$
.closestAlivePrecedingFinger($`k`$)
for $`i = {\mathcal{M}\hspace{0pt}..1}`$
if (($`{f\hspace{0pt}i\hspace{0pt}n_{i}} \in (n,k)`$)
and ($`{f\hspace{0pt}i\hspace{0pt}n_{i}} \neq {n\hspace{0pt}i\hspace{0pt}l}`$)
and isAlive($`f\hspace{0pt}i\hspace{0pt}n_{i}`$))
return($`f\hspace{0pt}i\hspace{0pt}n_{i}`$)
return(nil)
$`n`$
.closestAlivePrecedingSucc($`k`$)
for $`i = {\mathcal{S}\hspace{0pt}..1}`$
if (($`s_{i} \in (n,k)`$)
and ($`s_{i} \neq {n\hspace{0pt}i\hspace{0pt}l}`$)
and isAlive($`s_{i}`$))
return($`s_{i}`$)
return(cpf)
Figure 3: The Lookup Algorithm.
Initialization of Fingers (Fig. LABEL:fig:fixfingers). After having initialized its first successor $`s_{1}`$, a node $`n`$ sets all fingers with starts between $`n`$ and $`s_{1}`$ to $`s_{1}`$. The rest of the fingers are initialized by taking a copy of the finger table of $`s_{1}`$ and finding an approximate successor to every finger from that finger table.
Lookups (Fig. [3](#S3.F3 "Figure 3 ‣ III-B Lookups and Stabilization of Fingers ‣ III Our Implementation of Chord ‣ An Analytical Study of a Structured Overlay in the Presence of Dynamic Membership This work is funded by the European 6th FP EVERGROW project. ©IEEE. Personal use of this material is permitted. However, permission to reprint/republish this material for advertising or promotional purposes or for creating new collective works for resale or redistribution to servers or lists, or to reuse any copyrighted component of this work in other works, must be obtained from the IEEE.")). A lookup operation is a fundamental operation that is used to find the successor of a key. It is used by many other routines and its performance and consistency are the main quantities of interest in the evaluation of any DHT. A node $`n`$ looking up the successor of $`k`$ runs the *findSuccessor* algorithm which can lead to the following cases:
Case A. If $`k`$ is equal to $`n`$ then $`n`$ is trivially the successor of $`k`$.
Case B. If $`k \in {(n,s_{1}\rbrack}`$ then $`n`$ has found the successor of $`k`$, but it could be that $`s_{1}`$ has failed and $`n`$ has not yet discovered this. However, entries in the successor list can act as backups for the first successor. Therefore, the first alive successor of $`n`$ is the successor of $`k`$. Note that, in this case, while we try to find the first alive successor, we do not change the entries in the successor list. This is mainly because, to simplify the analysis, we want the successor list to be changed at a fixed rate rate $`\alpha\hspace{0pt}\lambda_{s}`$ only by the *fixSuccessors* function.
Case C. The lookup should be forwarded to a node closer to $`k`$, namely the closest alive finger preceding $`k`$ in $`n`$’s finger table. The call to the function *closestAlivePrecedingFinger* returns such a node if possible and the lookup is forwarded to it. However, it could be the case that all alive preceding fingers to $`k`$ are dead. In that case, we need to use the successor list as a last resort for the lookup. Therefore, we locate the first alive successor $`y`$ and if $`k \in {(n,y\rbrack}`$ then $`y`$ is the successor of $`k`$. Otherwise, we locate the closest alive preceding successor to $`k`$ and forward the lookup to it.
###
III-C
Failures
Throughout the code we use the call $`i\hspace{0pt}s\hspace{0pt}A\hspace{0pt}l\hspace{0pt}i\hspace{0pt}v\hspace{0pt}e`$ and $`i\hspace{0pt}s\hspace{0pt}N\hspace{0pt}o\hspace{0pt}t\hspace{0pt}A\hspace{0pt}l\hspace{0pt}i\hspace{0pt}v\hspace{0pt}e`$. A simple interpretation of those routines would be to equate them to a performance of a ping. However, a correct implementation for them is that they are discovered by performing the operation required. For instance, a call to $`f\hspace{0pt}i\hspace{0pt}r\hspace{0pt}s\hspace{0pt}t\hspace{0pt}A\hspace{0pt}l\hspace{0pt}i\hspace{0pt}v\hspace{0pt}e\hspace{0pt}S\hspace{0pt}u\hspace{0pt}c\hspace{0pt}c\hspace{0pt}e\hspace{0pt}s\hspace{0pt}o\hspace{0pt}r`$ in Fig. [1](#S3.F1 "Figure 1 ‣ III-A Joins, Failures & Ring Stabilization ‣ III Our Implementation of Chord ‣ An Analytical Study of a Structured Overlay in the Presence of Dynamic Membership This work is funded by the European 6th FP EVERGROW project. ©IEEE. Personal use of this material is permitted. However, permission to reprint/republish this material for advertising or promotional purposes or for creating new collective works for resale or redistribution to servers or lists, or to reuse any copyrighted component of this work in other works, must be obtained from the IEEE.") is performed to retrieve a node $`y`$ and then call $`y.{i\hspace{0pt}T\hspace{0pt}h\hspace{0pt}i\hspace{0pt}n\hspace{0pt}k\hspace{0pt}I\hspace{0pt}a\hspace{0pt}m\hspace{0pt}Y\hspace{0pt}o\hspace{0pt}u\hspace{0pt}r\hspace{0pt}P\hspace{0pt}r\hspace{0pt}e\hspace{0pt}d}`$, so alternatively the first alive successor could be discovered by iterating on the successor list and calling $`i\hspace{0pt}T\hspace{0pt}h\hspace{0pt}i\hspace{0pt}n\hspace{0pt}k\hspace{0pt}I\hspace{0pt}a\hspace{0pt}m\hspace{0pt}Y\hspace{0pt}o\hspace{0pt}u\hspace{0pt}r\hspace{0pt}P\hspace{0pt}r\hspace{0pt}e\hspace{0pt}d`$.
## IV The Analysis
###
IV-A
Distributional Properties of Inter-Node Distances
In this section we will assume that all keys are populated by peers with independent and equal probability, and, furthermore, that this probability does not change with time. The first condition is a natural consequence of peers joining and leaving/failing independently. The last condition, on the other hand, does not hold strictly since the number of peers present under churn is a fluctuating quantity, Nevertheless, it can be expected to hold to good accuracy in sufficiently large systems. A detailed analysis along these lines will be given elsewhere.
######
Definition IV.1
Given two keys $`{u,v} \in {\{{{0\hspace{0pt}\ldots\hspace{0pt}\mathcal{K}} - 1}\}}`$, the “distance” between them is $`u - v`$ (with modulo-$`\mathcal{K}`$ arithmetic). We interchangeably say that $`u`$ and $`v`$ form an “interval” of length $`u - v`$. Hence the number of keys inside an interval of length $`\ell`$ is $`\ell - 1`$ keys.
######
Property IV.1
The probability $`P\hspace{0pt}{(x)}`$ of finding an interval of length $`x`$ is: $`{P\hspace{0pt}{(x)}} = {\rho^{x - 1}\hspace{0pt}{({1 - \rho})}}`$ where $`\rho = \frac{\mathcal{K} - N}{\mathcal{K}}`$.
Under the stated conditions, each key will be populated with the same probability $`\frac{N}{\mathcal{K}} = {1 - \rho}`$, for $`N\operatorname{<<}K`$. An interval of length $`x`$ then involves $`x - 1`$ consecutive unpopulated keys, and then one populated key, which explains the formula.
[Refer to caption](/html/0710.0270/assets/x1.png)
Figure 4: (a) Case when $`n`$ and $`p`$ have the same value of $`{f\hspace{0pt}i\hspace{0pt}n_{k}}.{n\hspace{0pt}o\hspace{0pt}d\hspace{0pt}e}`$. (b) Case where a newly joined node $`p`$ copies the $`k^{t\hspace{0pt}h}`$ entry of its successor node $`n`$ as the best approximation for its own $`k^{t\hspace{0pt}h}`$ entry (by the join protocol). In this case, there could be a node $`o`$ which is the ’correct’ entry for $`p.{f\hspace{0pt}i\hspace{0pt}n_{k}}.{n\hspace{0pt}o\hspace{0pt}d\hspace{0pt}e}`$. However, since $`p`$ is newly joined, the only information it has access to is the finger table of $`n`$.
We now derive some properties of this distribution which will be used in the ensuing analysis.
######
Property IV.2
For any two keys $`u`$ and $`v`$, where $`v = {u + x}`$, let $`b_{i}`$ be the probability that the first node encountered in between these two keys is at $`u + i`$ (where $`0 \leq i < x`$). Then $`b_{i} \equiv {\rho^{i}\hspace{0pt}{({1 - \rho})}}`$. The probability that there is definitely at least one node between $`u`$ and $`v`$ is: $`{a\hspace{0pt}{(x)}} \equiv {1 - \rho^{x}}`$. Hence the conditional probability that the first node is at a distance $`i`$ given that there is at least one node in the interval is $`{b\hspace{0pt}c\hspace{0pt}{(i,x)}} \equiv {{{b\hspace{0pt}{(i)}}/a}\hspace{0pt}{(x)}}`$.
######
Property IV.3
The probability that a node and at least one of its immediate predecessors share the same $`k^{t\hspace{0pt}h}`$ finger is $`{p_{1}\hspace{0pt}{(k)}} \equiv {\frac{\rho}{1 + \rho}\hspace{0pt}{({1 - \rho^{2^{k} - 2}})}}`$. The explanation for this property goes as follows. If the distance between node $`n`$ and its predecessor $`p`$ is $`x`$, the distance between $`n.{f\hspace{0pt}i\hspace{0pt}n_{k}}`$.*start* and $`p.{f\hspace{0pt}i\hspace{0pt}n_{k}}`$.*start* is also $`x`$ (see Fig. [4](#S4.F4 "Figure 4 ‣ IV-A Distributional Properties of Inter-Node Distances ‣ IV The Analysis ‣ An Analytical Study of a Structured Overlay in the Presence of Dynamic Membership This work is funded by the European 6th FP EVERGROW project. ©IEEE. Personal use of this material is permitted. However, permission to reprint/republish this material for advertising or promotional purposes or for creating new collective works for resale or redistribution to servers or lists, or to reuse any copyrighted component of this work in other works, must be obtained from the IEEE.")(a)). If there is no node in between $`n.{f\hspace{0pt}i\hspace{0pt}n_{k}}`$.*start* and $`p.{f\hspace{0pt}i\hspace{0pt}n_{k}}`$.*start* then $`n.{f\hspace{0pt}i\hspace{0pt}n_{k}}`$.*node* and $`p.{f\hspace{0pt}i\hspace{0pt}n_{k}}`$.*node* will share the same value. From Property [IV.1](#S4.Thmproperty1 "Property IV.1 ‣ IV-A Distributional Properties of Inter-Node Distances ‣ IV The Analysis ‣ An Analytical Study of a Structured Overlay in the Presence of Dynamic Membership This work is funded by the European 6th FP EVERGROW project. ©IEEE. Personal use of this material is permitted. However, permission to reprint/republish this material for advertising or promotional purposes or for creating new collective works for resale or redistribution to servers or lists, or to reuse any copyrighted component of this work in other works, must be obtained from the IEEE."), the probability that the distance between $`n`$ and $`p`$ is $`x`$ is $`\rho^{x - 1}\hspace{0pt}{({1 - \rho})}`$. However, $`x`$ has to be less than $`2^{k - 1}`$, otherwise $`p.{f\hspace{0pt}i\hspace{0pt}n_{k}}`$.*node* will be equal to $`n`$. The probability that no node exists between $`n.{f\hspace{0pt}i\hspace{0pt}n_{k}}`$.*start* and $`p.{f\hspace{0pt}i\hspace{0pt}n_{k}}`$.*start* is $`\rho^{x}`$ (by Property [IV.2](#S4.Thmproperty2 "Property IV.2 ‣ IV-A Distributional Properties of Inter-Node Distances ‣ IV The Analysis ‣ An Analytical Study of a Structured Overlay in the Presence of Dynamic Membership This work is funded by the European 6th FP EVERGROW project. ©IEEE. Personal use of this material is permitted. However, permission to reprint/republish this material for advertising or promotional purposes or for creating new collective works for resale or redistribution to servers or lists, or to reuse any copyrighted component of this work in other works, must be obtained from the IEEE.")). Therefore the probability that the $`n.{f\hspace{0pt}i\hspace{0pt}n_{k}}`$.*node* and $`p.{f\hspace{0pt}i\hspace{0pt}n_{k}}`$.*node* share the same value is: $`{\sum_{x = 1}^{2^{k - 1} - 1}{\rho^{x - 1}\hspace{0pt}{({1 - \rho})}\hspace{0pt}\rho^{x}}} = {\frac{\rho}{1 + \rho}\hspace{0pt}{({1 - \rho^{2^{k} - 2}})}}`$. It is straightforward (though tedious) to derive similar expressions for $`p_{2}\hspace{0pt}{(k)}`$ the probability that a node and at least two of its immediate predecessors share the same $`k^{t\hspace{0pt}h}`$ finger, $`p_{3}\hspace{0pt}{(k)}`$ and so on.
######
Property IV.4
We can similarly assess the probability that the join protocol (see Section [](#S3.SS2 "III-B Lookups and Stabilization of Fingers ‣ III Our Implementation of Chord ‣ An Analytical Study of a Structured Overlay in the Presence of Dynamic Membership This work is funded by the European 6th FP EVERGROW project. ©IEEE. Personal use of this material is permitted. However, permission to reprint/republish this material for advertising or promotional purposes or for creating new collective works for resale or redistribution to servers or lists, or to reuse any copyrighted component of this work in other works, must be obtained from the IEEE.")
III-B
) results in further replication of the $`k^{t\hspace{0pt}h}`$ pointer. Let us define the probability $`p_{join}\hspace{0pt}{(i,k)}`$ as the probability that a newly joined node, chooses the $`i^{t\hspace{0pt}h}`$ entry of its successor’s finger table for its own $`k^{t\hspace{0pt}h}`$ entry. Note that this is unambiguous even in the case that the successor’s $`i^{t\hspace{0pt}h}`$ entry is repeated. All we are asking is, when is the $`k^{t\hspace{0pt}h}`$ entry of the new joinee the same as the $`i^{t\hspace{0pt}h}`$ entry of the successor? Clearly $`i \leq k`$. In fact for the larger fingers, we only need to consider $`p_{join}\hspace{0pt}{(k,k)}`$, since $`{p_{join}\hspace{0pt}{(i,k)}} \sim 0`$ for $`i < k`$. Using the interval distribution we find, for large $`k`$, $`{p_{j\hspace{0pt}o\hspace{0pt}i\hspace{0pt}n}\hspace{0pt}{(k,k)}} \sim {{{\rho\hspace{0pt}{({1 - \rho^{2^{k - 2} - 2}})}} + {{({1 - \rho})}\hspace{0pt}{({1 - \rho^{2^{k - 2} - 2}})}}} - {{({1 - \rho})}\hspace{0pt}\rho\hspace{0pt}{({2^{k - 2} - 2})}\hspace{0pt}\rho^{2^{k - 2} - 3}}}`$. This function goes to $`1`$ for large $`k`$.
We can also analogously compute $`p_{join}\hspace{0pt}{(i,k)}`$ for any $`i`$. The only trick here is to estimate the probability that starting from $`i`$, the last distinct entry of $`n`$’s finger table does not give $`p`$ a better choice for its $`k_{t\hspace{0pt}h}`$ entry. This can again readily be computed using property [IV.2](#S4.Thmproperty2 "Property IV.2 ‣ IV-A Distributional Properties of Inter-Node Distances ‣ IV The Analysis ‣ An Analytical Study of a Structured Overlay in the Presence of Dynamic Membership This work is funded by the European 6th FP EVERGROW project. ©IEEE. Personal use of this material is permitted. However, permission to reprint/republish this material for advertising or promotional purposes or for creating new collective works for resale or redistribution to servers or lists, or to reuse any copyrighted component of this work in other works, must be obtained from the IEEE."), but we do not do the computation here since for our purposes $`p_{join}\hspace{0pt}{(k,k)}`$ suffices.
###
IV-B
Successor Pointers
[Refer to caption](/html/0710.0270/assets/x2.png)
Figure 5: Changes in $`W_{1}`$, the number of wrong (failed or outdated) $`s_{1}`$ pointers, due to joins, failures and stabilizations.
We now turn to estimating various quantities of interest for Chord. In all that follows we will evaluate various average quantities, as a function of the parameters. To do this we need to understand how the dynamical evolution of the system affects these quantities.
In the case of Chord, we only need to consider one of three kinds of events happening at any micro-instant: a join, a failure or a stabilization. One assumption made in the following is that such a micro-instant of time exists, or in other words, that we can divide time till we have an interval small enough that in this interval, only one of these three processes occurs anywhere in the system. Implicit in this is the assumption that a stabilization (either of successors or fingers) is done faster than the time-scales over which joins and fails occur.
Another aspect of this system which simplifies analysis is that successor pointers of adjacent nodes are independent of each other. That is, the state of the first successor pointer of a given node does not affect the state of the first successor pointer of either its predecessor or its successor. The same logic also works for the state of the second successor pointers of adjacent nodes and so on. On the other hand, the state of the second successor pointer of a node is clearly related to the state of its first successor pointer as well the state of the first successor pointer of the successor. This is taken into account in the analysis of second and higher successor pointers. In characterizing the states of higher successors, we look for the leading order behavior in terms of the parameter $`r`$.
Consider first the successor pointers. Let $`w_{k}\hspace{0pt}{(r,\alpha)}`$ denote the fraction of nodes having a *wrong* $`k^{t\hspace{0pt}h}`$ successor pointer and $`d_{k}\hspace{0pt}{(r,\alpha)}`$ the fraction of nodes having a *failed* successor pointer. Also, let $`W_{k}\hspace{0pt}{(r,\alpha)}`$ be the number of nodes having a *wrong* $`k^{t\hspace{0pt}h}`$ successor pointer and $`D_{k}\hspace{0pt}{(r,\alpha)}`$ the number of nodes having a *failed* successor pointer. A *failed* pointer is one which points to a departed node while a *wrong* pointer points either to an incorrect node (alive but not correct) or a dead one. As we will see, both these quantities play a role in predicting lookup consistency and lookup length.
By the protocol for stabilizing successors in Chord, a node periodically contacts its first successor, possibly correcting it and reconciling with its successor list. Therefore, the number of wrong $`k^{t\hspace{0pt}h}`$ successor pointers are not independent quantities but depend on the number of wrong first successor pointers.
TABLE I: Gain and loss terms for $`W_{1}\hspace{0pt}{(r,\alpha)}`$: the number of wrong first successors as a function of $`r`$ and $`\alpha`$.
<table>
<thead>
<tr>
<th>Change in W1(r,α)subscript𝑊1𝑟𝛼W_{1}(r,\alpha)</th>
<th>Probability of Occurrence</th>
</tr>
</thead>
<tbody>
<tr>
<td>W1(t+Δt)=W1(t)+1subscript𝑊1𝑡Δ𝑡subscript𝑊1𝑡1W_{1}(t+\Delta t)=W_{1}(t)+1</td>
<td>c1.1=(λjNΔt)(1−w1)subscript𝑐1.1subscript𝜆𝑗𝑁Δ𝑡1subscript𝑤1c_{1.1}=(\lambda_{j}N\Delta t)(1-w_{1})</td>
</tr>
<tr>
<td>W1(t+Δt)=W1(t)+1subscript𝑊1𝑡Δ𝑡subscript𝑊1𝑡1W_{1}(t+\Delta t)=W_{1}(t)+1</td>
<td>c1.2=λfN(1−w1)2Δtsubscript𝑐1.2subscript𝜆𝑓𝑁superscript1subscript𝑤12Δ𝑡c_{1.2}=\lambda_{f}N(1-w_{1})^{2}\Delta t</td>
</tr>
<tr>
<td>W1(t+Δt)=W1(t)−1subscript𝑊1𝑡Δ𝑡subscript𝑊1𝑡1W_{1}(t+\Delta t)=W_{1}(t)-1</td>
<td>c1.3=λfNw12Δtsubscript𝑐1.3subscript𝜆𝑓𝑁superscriptsubscript𝑤12Δ𝑡c_{1.3}=\lambda_{f}Nw_{1}^{2}\Delta t</td>
</tr>
<tr>
<td>W1(t+Δt)=W1(t)−1subscript𝑊1𝑡Δ𝑡subscript𝑊1𝑡1W_{1}(t+\Delta t)=W_{1}(t)-1</td>
<td>c1.4=αλsNw1Δtsubscript𝑐1.4𝛼subscript𝜆𝑠𝑁subscript𝑤1Δ𝑡c_{1.4}=\alpha\lambda_{s}Nw_{1}\Delta t</td>
</tr>
<tr>
<td>W1(t+Δt)=W1(t)subscript𝑊1𝑡Δ𝑡subscript𝑊1𝑡W_{1}(t+\Delta t)=W_{1}(t)</td>
<td>1−(c1.1+c1.2+c1.3+c1.4)1subscript𝑐1.1subscript𝑐1.2subscript𝑐1.3subscript𝑐1.41-(c_{1.1}+c_{1.2}+c_{1.3}+c_{1.4})</td>
</tr>
</tbody>
</table>
We write an equation for $`W_{1}\hspace{0pt}{(r,\alpha)}`$ by accounting for all the events that can change it in a micro event of time $`\Delta\hspace{0pt}t`$. An illustration of the different cases in which changes in $`W_{1}`$ take place due to joins, failures and stabilizations is provided in Fig. [5](#S4.F5 "Figure 5 ‣ IV-B Successor Pointers ‣ IV The Analysis ‣ An Analytical Study of a Structured Overlay in the Presence of Dynamic Membership This work is funded by the European 6th FP EVERGROW project. ©IEEE. Personal use of this material is permitted. However, permission to reprint/republish this material for advertising or promotional purposes or for creating new collective works for resale or redistribution to servers or lists, or to reuse any copyrighted component of this work in other works, must be obtained from the IEEE."). In some cases $`W_{1}`$ increases/decreases while in others it stays unchanged. For each increase/decrease, Table [I](#S4.T1 "TABLE I ‣ IV-B Successor Pointers ‣ IV The Analysis ‣ An Analytical Study of a Structured Overlay in the Presence of Dynamic Membership This work is funded by the European 6th FP EVERGROW project. ©IEEE. Personal use of this material is permitted. However, permission to reprint/republish this material for advertising or promotional purposes or for creating new collective works for resale or redistribution to servers or lists, or to reuse any copyrighted component of this work in other works, must be obtained from the IEEE.") provides the corresponding probabilities.
By our implementation of the join protocol, a new node $`n_{y}`$, joining between two nodes $`n_{x}`$ and $`n_{z}`$, always has a correct $`s_{1}`$ pointer after the join. However the state of $`n_{x}.s_{1}`$ before the join makes a difference. If $`n_{x}.s_{1}`$ was correct (pointing to $`n_{z}`$) before the join, then after the join it will be wrong and therefore $`W_{1}`$ increases by $`1`$. If $`n_{x}.s_{1}`$ was wrong before the join, then it will remain wrong after the join and $`W_{1}`$ is unaffected. Thus, we need to account for the former case only. The probability that $`n_{x}.s_{1}`$ is correct is $`1 - w_{1}`$ and term $`c_{1.1}`$ follows from this.
For failures, we have $`4`$ cases. To illustrate them we use nodes $`n_{x}`$, $`n_{y}`$, $`n_{z}`$ and assume that $`n_{y}`$ is going to fail. First, if both $`n_{x}.s_{1}`$ and $`n_{y}.s_{1}`$ were correct, then the failure of $`n_{y}`$ will make $`n_{x}.s_{1}`$ wrong and hence $`W_{1}`$ increases by $`1`$. Second, if $`n_{x}.s_{1}`$ and $`n_{y}.s_{1}`$ were both wrong, then the failure of $`n_{y}`$ will decrease $`W_{1}`$ by one, since one wrong pointer disappears. Third, if $`n_{x}.s_{1}`$ was wrong and $`n_{y}.s_{1}`$ was correct, then $`W_{1}`$ is unaffected. Fourth, if $`n_{x}.s_{1}`$ was correct and $`n_{y}.s_{1}`$ was wrong, then the wrong pointer of $`n_{y}`$ disappears and $`n_{x}.s_{1}`$ becomes wrong, therefore $`W_{1}`$ is unaffected. For the first case to happen, we need to pick two nodes with correct pointers, the probability of this is $`{({1 - w_{1}})}^{2}`$. For the second case to happen, we need to pick two nodes with wrong pointers, the probability of this is $`w_{1}^{2}`$. From these probabilities follow the terms $`c_{1.2}`$ and $`c_{1.3}`$.
Finally, a successor stabilization does not affect $`W_{1}`$, unless the stabilizing node had a wrong pointer. The probability of picking such a node is $`w_{1}`$. From this follows the term $`c_{1.4}`$.
Hence the equation for $`W_{1}\hspace{0pt}{(r,\alpha)}`$ is:
<table>
<tbody>
<tr>
<td></td>
<td>dW1Ndt=λj(1−w1)+λf(1−w1)2−λfw12−αλsw1𝑑subscript𝑊1𝑁𝑑𝑡subscript𝜆𝑗1subscript𝑤1subscript𝜆𝑓superscript1subscript𝑤12subscript𝜆𝑓superscriptsubscript𝑤12𝛼subscript𝜆𝑠subscript𝑤1\frac{dW_{1}}{Ndt}=\lambda_{j}(1-w_{1})+\lambda_{f}(1-w_{1})^{2}-\lambda_{f}w_{1}^{2}-\alpha\lambda_{s}w_{1}</td>
<td></td>
</tr>
</tbody>
</table>
Solving for $`w_{1}`$ in the steady state and putting $`\lambda_{j} = \lambda_{f}`$, we get:
<table>
<tbody>
<tr>
<td></td>
<td>w1(r,α)=23+rα≈2rαsubscript𝑤1𝑟𝛼23𝑟𝛼2𝑟𝛼w_{1}(r,\alpha)=\frac{2}{3+r\alpha}\approx\frac{2}{r\alpha}</td>
<td></td>
<td>(1)</td>
</tr>
</tbody>
</table>
This expression matches well with the simulation results as shown in Fig. [6](#S4.F6 "Figure 6 ‣ IV-B Successor Pointers ‣ IV The Analysis ‣ An Analytical Study of a Structured Overlay in the Presence of Dynamic Membership This work is funded by the European 6th FP EVERGROW project. ©IEEE. Personal use of this material is permitted. However, permission to reprint/republish this material for advertising or promotional purposes or for creating new collective works for resale or redistribution to servers or lists, or to reuse any copyrighted component of this work in other works, must be obtained from the IEEE."). $`d_{1}\hspace{0pt}{(r,\alpha)}`$ is then $`\approx {\frac{1}{2}\hspace{0pt}w_{1}\hspace{0pt}{(r,\alpha)}}`$ since when $`\lambda_{j} = \lambda_{f}`$, about half the number of wrong pointers are incorrect and about half point to dead nodes. Thus $`{d_{1}\hspace{0pt}{(r,\alpha)}} \approx \frac{1}{r\hspace{0pt}\alpha}`$ which also matches well the simulations as shown in Fig. [6](#S4.F6 "Figure 6 ‣ IV-B Successor Pointers ‣ IV The Analysis ‣ An Analytical Study of a Structured Overlay in the Presence of Dynamic Membership This work is funded by the European 6th FP EVERGROW project. ©IEEE. Personal use of this material is permitted. However, permission to reprint/republish this material for advertising or promotional purposes or for creating new collective works for resale or redistribution to servers or lists, or to reuse any copyrighted component of this work in other works, must be obtained from the IEEE.").
[Refer to caption](/html/0710.0270/assets/x3.png)
Figure 6: Theory and simulation for the probability of wrong $`1^{s\hspace{0pt}t}`$ successor $`w_{1}\hspace{0pt}{(r,\alpha)}`$ and failed $`1^{s\hspace{0pt}t}`$ successor $`d_{1}\hspace{0pt}{(r,\alpha)}`$.
The fraction of wrong second successors can be estimated in an analogous manner. Consider, for a node $`n`$, the possible states of the successor, $`n.s_{1}`$, the successor of the successor, $`\ast {(n.s_{1})}.s_{1}`$, and the second successor, $`n.s_{2}`$. In a fully correct state, $`\ast {(n.s_{1})}.s_{1}`$ and $`n.s_{2}`$ of course point to the same node. If in such a state either $`n.s_{1}`$ or $`\ast {(n.s_{1})}.s_{1}`$ becomes incorrect through the action of a join or a failure, then $`n.s_{2}`$ is also incorrect. On the other hand, $`n.s_{2}`$ cannot be corrected by the stabilization protocol unless both $`n.s_{1}`$ and $`\ast {(n.s_{1})}.s_{1}`$ are both already corrected. Hence, $`n.s_{2}`$ is wrong if either $`n.s_{1}`$ or $`\ast {(n.s_{1})}.s_{1}`$ are wrong, and also if both $`n.s_{1}`$ and $`\ast {(n.s_{1})}.s_{1}`$ are correct, but $`n.s_{2}`$ has not yet been corrected. If the number of such non-stabilized configurations is $`N_{2}`$ and the fraction is $`n_{2}`$, we have
<table>
<tbody>
<tr>
<td></td>
<td>w2=2w1−w12+n2subscript𝑤22subscript𝑤1superscriptsubscript𝑤12subscript𝑛2w_{2}=2w_{1}-w_{1}^{2}+n_{2}</td>
<td></td>
<td>(2)</td>
</tr>
</tbody>
</table>
To estimate $`n_{2}`$ we consider how these configurations might be gained or lost. The gain term arises from stabilizations of configurations where $`n.s_{1}`$ is correct but $`\ast {(n.s_{1})}.s_{1}`$ is wrong. A stabilization performed by node $`n.s_{1}`$ then results in the gain of a $`N_{2}`$ configuration. On the other hand, non-stabilized configurations are lost either by a stabilization performed by node $`n`$ (when it gets the correct successor list from its successor and hence corrects $`n.s_{2}`$), or by corrupting either $`n.s_{1}`$ or $`\ast {(n.s_{1})}.s_{1}`$ (by a join or failure). The latter possibility gives terms of order $`\frac{1}{r^{2}}`$ and we can ignore it in the limit that stabilizations happens on a much faster time scale than joins and failures (i.e., $`r`$ much larger than unity). The equation for $`N_{2}`$ is hence
<table>
<tbody>
<tr>
<td></td>
<td>dN2dt≈αλsw1(1−w1)−αλsn2𝑑subscript𝑁2𝑑𝑡𝛼subscript𝜆𝑠subscript𝑤11subscript𝑤1𝛼subscript𝜆𝑠subscript𝑛2\frac{dN_{2}}{dt}\approx\alpha\lambda_{s}w_{1}(1-w_{1})-\alpha\lambda_{s}n_{2}</td>
<td></td>
<td>(3)</td>
</tr>
</tbody>
</table>
which implies $`n_{2} \approx w_{1}`$ to order $`\frac{1}{r}`$. Thus, we have $`w_{2} \approx \frac{6}{r}`$.
For higher successors we reason similarly by considering the state of the $`k - 1^{s\hspace{0pt}t}`$ successor pointer of node $`n`$, the successor pointer of the $`k - 1^{s\hspace{0pt}t}`$ successor, and the $`k^{t\hspace{0pt}h}`$ successor pointer of node $`n`$. We can write a recursion equation for $`w_{k}`$ the fraction of nodes with wrong $`k^{t\hspace{0pt}h}`$ successor pointer
<table>
<tbody>
<tr>
<td></td>
<td>wk=w1+wk−1−wk−1w1+nksubscript𝑤𝑘subscript𝑤1subscript𝑤𝑘1subscript𝑤𝑘1subscript𝑤1subscript𝑛𝑘w_{k}=w_{1}+w_{k-1}-w_{k-1}w_{1}+n_{k}</td>
<td></td>
<td>(4)</td>
</tr>
</tbody>
</table>
where $`n_{k}`$ is the density of configurations where the $`k - 1^{s\hspace{0pt}t}`$ successor pointer of node $`n`$ and the first successor pointer of the $`k - 1^{s\hspace{0pt}t}`$ successor are both correct, but this information has not yet been used to correct the $`k^{t\hspace{0pt}h}`$ successor pointer of node $`n`$. If node $`n`$ does not as yet have the correct information about its $`k^{t\hspace{0pt}h}`$ successor, that means that either all the nodes in between $`n`$ and its $`k - 1^{s\hspace{0pt}t}`$ successor have the correct information but node $`n`$ has not as yet stabilized, or that the stabilization has propagated back from the $`k - 1^{s\hspace{0pt}t}`$ successor to some node in between but not as yet to $`n.s_{1}`$. To elaborate on this further, there is the case where the second successor pointer of the $`k - 2^{n\hspace{0pt}d}`$ successor has not been corrected, then the case where this has been done, but the third successor pointer of the $`k - 3^{r\hspace{0pt}d}`$ successor has not been corrected, and so on. Each of these is analogous to $`n_{2}`$ and each occurs with density $`{({1 - w_{k - 1}})}\hspace{0pt}w_{1}`$, if joins and failures are neglected compared to stabilizations. Hence, if to leading order in $`\frac{1}{r}`$ we have $`w_{k} \sim \frac{c_{k}}{\alpha\hspace{0pt}r}`$, then
<table>
<tbody>
<tr>
<td></td>
<td>ck=ck−1+kc1subscript𝑐𝑘subscript𝑐𝑘1𝑘subscript𝑐1c_{k}=c_{k-1}+kc_{1}</td>
<td></td>
<td>(5)</td>
</tr>
</tbody>
</table>
which leads to
<table>
<tbody>
<tr>
<td></td>
<td>wk≈k(k+1)αrsubscript𝑤𝑘𝑘𝑘1𝛼𝑟w_{k}\approx\frac{k(k+1)}{\alpha r}</td>
<td></td>
<td>(6)</td>
</tr>
</tbody>
</table>
. We note that this expression obviously depends on the details of the stabilization scheme, and is in principle only valid up to $`k \sim \sqrt{r}`$. As shown in Fig. [7](#S4.F7 "Figure 7 ‣ IV-B Successor Pointers ‣ IV The Analysis ‣ An Analytical Study of a Structured Overlay in the Presence of Dynamic Membership This work is funded by the European 6th FP EVERGROW project. ©IEEE. Personal use of this material is permitted. However, permission to reprint/republish this material for advertising or promotional purposes or for creating new collective works for resale or redistribution to servers or lists, or to reuse any copyrighted component of this work in other works, must be obtained from the IEEE."), the agreement between theory and simulation is still however quite reasonable at $`k = 5`$ and $`r = 100`$.
[Refer to caption](/html/0710.0270/assets/x4.png)
Figure 7: Theory and simulation for the probability of a wrong $`k^{t\hspace{0pt}h}`$ successor $`w_{k}\hspace{0pt}{(r,\alpha)}`$.
###
IV-C
Break-up (Network Disconnection) Probability
TABLE II: Gain and loss terms for $`N_{b\hspace{0pt}u}\hspace{0pt}{(2,r,\alpha)}`$: the number of nodes with dead first and second successors.
<table>
<thead>
<tr>
<th>Change in Nbu(r,α)subscript𝑁𝑏𝑢𝑟𝛼N_{bu}(r,\alpha)</th>
<th>Probability of Occurrence</th>
</tr>
</thead>
<tbody>
<tr>
<td>Nbu(t+Δt)=Nbu(t)+1subscript𝑁𝑏𝑢𝑡Δ𝑡subscript𝑁𝑏𝑢𝑡1N_{bu}(t+\Delta t)=N_{bu}(t)+1</td>
<td>c2.1=(λfNΔt)d1(r,α)subscript𝑐2.1subscript𝜆𝑓𝑁Δ𝑡subscript𝑑1𝑟𝛼c_{2.1}=(\lambda_{f}N\Delta t)d_{1}(r,\alpha)</td>
</tr>
<tr>
<td>Nbu(t+Δt)=Nbu(t)+1subscript𝑁𝑏𝑢𝑡Δ𝑡subscript𝑁𝑏𝑢𝑡1N_{bu}(t+\Delta t)=N_{bu}(t)+1</td>
<td>c2.2=λfNΔt(1−d1)d2subscript𝑐2.2subscript𝜆𝑓𝑁Δ𝑡1subscript𝑑1subscript𝑑2c_{2.2}=\lambda_{f}N\Delta t(1-d_{1})d_{2}</td>
</tr>
<tr>
<td>Nbu(t+Δt)=Nbu(t)−1subscript𝑁𝑏𝑢𝑡Δ𝑡subscript𝑁𝑏𝑢𝑡1N_{bu}(t+\Delta t)=N_{bu}(t)-1</td>
<td>c2.3=αλsNΔtPbu(2,r,α)subscript𝑐2.3𝛼subscript𝜆𝑠𝑁Δ𝑡subscript𝑃𝑏𝑢2𝑟𝛼c_{2.3}=\alpha\lambda_{s}N\Delta tP_{bu}(2,r,\alpha)</td>
</tr>
<tr>
<td>Nbu(t+Δt)=Nbu(t)subscript𝑁𝑏𝑢𝑡Δ𝑡subscript𝑁𝑏𝑢𝑡N_{bu}(t+\Delta t)=N_{bu}(t)</td>
<td>1−(c2.1+c2.2+c2.3)1subscript𝑐2.1subscript𝑐2.2subscript𝑐2.31-(c_{2.1}+c_{2.2}+c_{2.3})</td>
</tr>
</tbody>
</table>
We demonstrate below, how calculating $`d_{k}\hspace{0pt}{(r,\alpha)}`$: the fraction of nodes with dead $`k^{t\hspace{0pt}h}`$ pointers, helps in estimating the probability that the network gets disconnected for any value of $`r`$ and $`\alpha`$. Let $`P_{b\hspace{0pt}u}\hspace{0pt}{(n,r,\alpha)}`$ be the probability that $`n`$ consecutive nodes fail. If $`n = \mathcal{S}`$, the length of the successor list, then clearly the node whose successor list this is, gets disconnected from the network and the network breaks up. For the range of $`r`$ considered in Fig. [6](#S4.F6 "Figure 6 ‣ IV-B Successor Pointers ‣ IV The Analysis ‣ An Analytical Study of a Structured Overlay in the Presence of Dynamic Membership This work is funded by the European 6th FP EVERGROW project. ©IEEE. Personal use of this material is permitted. However, permission to reprint/republish this material for advertising or promotional purposes or for creating new collective works for resale or redistribution to servers or lists, or to reuse any copyrighted component of this work in other works, must be obtained from the IEEE."), $`{P_{b\hspace{0pt}u}\hspace{0pt}{(\mathcal{S},r,\alpha)}} \sim 0`$. However should we go lower, this starts becoming finite. The master equation analysis introduced here can be used to estimate $`P_{b\hspace{0pt}u}\hspace{0pt}{(n,r,\alpha)}`$ for any $`1 \leq n \leq \mathcal{S}`$. We indicate how this might be done by first considering the case $`n = 2`$. Let $`N_{b\hspace{0pt}u}\hspace{0pt}{(2,r,\alpha)}`$ be the number of configurations in which a node has both $`s_{1}`$ and $`s_{2}`$ dead and $`P_{b\hspace{0pt}u}\hspace{0pt}{(2,r,\alpha)}`$ be the fraction of such configurations. Table [II](#S4.T2 "TABLE II ‣ IV-C Break-up (Network Disconnection) Probability ‣ IV The Analysis ‣ An Analytical Study of a Structured Overlay in the Presence of Dynamic Membership This work is funded by the European 6th FP EVERGROW project. ©IEEE. Personal use of this material is permitted. However, permission to reprint/republish this material for advertising or promotional purposes or for creating new collective works for resale or redistribution to servers or lists, or to reuse any copyrighted component of this work in other works, must be obtained from the IEEE.") indicates how this is estimated within the present framework.
[Refer to caption](/html/0710.0270/assets/x5.png)
Figure 8: Theory and simulation for the probability of failure of the $`2^{n\hspace{0pt}d}`$ successor, $`d_{2}\hspace{0pt}{(r,\alpha)}`$.
A join event does not affect this probability in any way. So we only need to consider the effect of failures or stabilization events. The term $`c_{2.1}`$ accounts for the situation when the first successor of a node is dead (which happens with probability $`d_{1}\hspace{0pt}{(r,\alpha)}`$ as explained above). A failure event can then kill its second successor as well and this happens with probability $`c_{2.1}`$. The second term is the situation that the first successor is alive (with probability $`1 - d_{1}`$) but the second successor is dead (with probability $`d_{2}`$). The logic used to estimate $`d_{2}`$ (or $`d_{k}`$ in general) is very similar to the reasoning we used to estimate the $`w_{k}`$’s. So we have
<table>
<tbody>
<tr>
<td></td>
<td>dk=d1+(k−1)d1=kd1subscript𝑑𝑘subscript𝑑1𝑘1subscript𝑑1𝑘subscript𝑑1d_{k}=d_{1}+(k-1)d_{1}=kd_{1}</td>
<td></td>
<td>(7)</td>
</tr>
</tbody>
</table>
Thus the $`k^{t\hspace{0pt}h}`$ successor of a node is dead if the $`k - 1^{s\hspace{0pt}t}`$ successor’s successor is dead, or the $`k - 1^{s\hspace{0pt}t}`$ successor’s successor is not dead but the intermediate nodes think it is because they haven’t stabilized. Hence $`d_{2} \sim {{2/\alpha}\hspace{0pt}r}`$. This estimate for $`d_{2}`$ matches the simulation results very well, as shown in Fig. [8](#S4.F8 "Figure 8 ‣ IV-C Break-up (Network Disconnection) Probability ‣ IV The Analysis ‣ An Analytical Study of a Structured Overlay in the Presence of Dynamic Membership This work is funded by the European 6th FP EVERGROW project. ©IEEE. Personal use of this material is permitted. However, permission to reprint/republish this material for advertising or promotional purposes or for creating new collective works for resale or redistribution to servers or lists, or to reuse any copyrighted component of this work in other works, must be obtained from the IEEE.").
Coming back to counting the gain and loss terms for $`N_{b\hspace{0pt}u}\hspace{0pt}{(2,r,\alpha)}`$, a stabilization event reduces the number of such configurations by one, if the node doing the stabilization had such a configuration to begin with.
Solving the equation for $`N_{b\hspace{0pt}u}\hspace{0pt}{(2,r,\alpha)}`$, one hence obtains that $`{P_{b\hspace{0pt}u}\hspace{0pt}{(2,r,\alpha)}} \sim {3/{({\alpha\hspace{0pt}r})}^{2}}`$. As Fig. [9](#S4.F9 "Figure 9 ‣ IV-C Break-up (Network Disconnection) Probability ‣ IV The Analysis ‣ An Analytical Study of a Structured Overlay in the Presence of Dynamic Membership This work is funded by the European 6th FP EVERGROW project. ©IEEE. Personal use of this material is permitted. However, permission to reprint/republish this material for advertising or promotional purposes or for creating new collective works for resale or redistribution to servers or lists, or to reuse any copyrighted component of this work in other works, must be obtained from the IEEE.") shows, this is a precise estimate.
We can similarly estimate the probabilities for three consecutive nodes failing, etc, and hence also the general disconnection probability $`P_{b\hspace{0pt}u}\hspace{0pt}{(\mathcal{S},r,\alpha)}`$. In fact $`P_{b\hspace{0pt}u}\hspace{0pt}{(\mathcal{S},r,\alpha)}`$ may be written in terms of the $`d_{k}\hspace{0pt}{(r,\alpha)}`$ as:
<table>
<tbody>
<tr>
<td></td>
<td>Pbu(𝒮)=(𝒮−1)!∑1𝒮di(r,α)(αr)𝒮−1subscript𝑃𝑏𝑢𝒮𝒮1superscriptsubscript1𝒮subscript𝑑𝑖𝑟𝛼superscript𝛼𝑟𝒮1P_{bu}({\cal S})=({{\cal S}-1})!\frac{\sum_{1}^{\cal S}d_{i}(r,\alpha)}{(\alpha r)^{{\cal S}-1}}</td>
<td></td>
<td>(8)</td>
</tr>
</tbody>
</table>
The logic behind this equation is similar to that used for solving for $`P_{b\hspace{0pt}u}\hspace{0pt}{(2)}`$, namely that for $`\mathcal{S}`$ consecutive nodes to fail, any $`\mathcal{S} - 1`$ of the $`\mathcal{S}`$ nodes should have failed first, and then a failure event kills the remaining node. ([8](#S4.E8 "In IV-C Break-up (Network Disconnection) Probability ‣ IV The Analysis ‣ An Analytical Study of a Structured Overlay in the Presence of Dynamic Membership This work is funded by the European 6th FP EVERGROW project. ©IEEE. Personal use of this material is permitted. However, permission to reprint/republish this material for advertising or promotional purposes or for creating new collective works for resale or redistribution to servers or lists, or to reuse any copyrighted component of this work in other works, must be obtained from the IEEE.")) is readily solved by substituting the values of the $`d_{k}`$’s to get
<table>
<tbody>
<tr>
<td></td>
<td>Pbu(𝒮)=(𝒮+1)!2(αr)𝒮subscript𝑃𝑏𝑢𝒮𝒮12superscript𝛼𝑟𝒮P_{bu}({\cal S})=\frac{({{\cal S}+1})!}{2(\alpha r)^{{\cal S}}}</td>
<td></td>
<td>(9)</td>
</tr>
</tbody>
</table>
As mentioned above this is again correct only to leading order. Namely there will be correction terms of the order $`r^{\mathcal{S} + 1}`$ which we haven’t computed at this level of approximation. The Master Equation formalism thus affords the possibility of making a precise prediction for when the system runs the danger of getting disconnected, as a function of the parameters.
[Refer to caption](/html/0710.0270/assets/x6.png)
Figure 9: Theory and simulation for the break-up probability $`P_{b\hspace{0pt}u}\hspace{0pt}{(2,r,\alpha)}`$.
Lookup Consistency By the lookup protocol, a lookup is inconsistent if the immediate predecessor of the sought key has a wrong $`s_{1}`$ pointer. However, we need only consider the case when the $`s_{1}`$ pointer is pointing to an alive (but incorrect) node since our implementation of the protocol always requires the lookup to return an alive node as an answer to the query. The probability that a lookup is inconsistent $`I\hspace{0pt}{(r,\alpha)}`$ is hence $`{w_{1}\hspace{0pt}{(r,\alpha)}} - {d_{1}\hspace{0pt}{(r,\alpha)}}`$. This prediction matches the simulation results very well, as shown in Fig. [10](#S4.F10 "Figure 10 ‣ IV-C Break-up (Network Disconnection) Probability ‣ IV The Analysis ‣ An Analytical Study of a Structured Overlay in the Presence of Dynamic Membership This work is funded by the European 6th FP EVERGROW project. ©IEEE. Personal use of this material is permitted. However, permission to reprint/republish this material for advertising or promotional purposes or for creating new collective works for resale or redistribution to servers or lists, or to reuse any copyrighted component of this work in other works, must be obtained from the IEEE.").
[Refer to caption](/html/0710.0270/assets/x7.png)
Figure 10: Theory and simulation for inconsistent lookups $`I\hspace{0pt}{(r,\alpha)}`$.
###
IV-D
Failure of Fingers
[Refer to caption](/html/0710.0270/assets/x8.png)
Figure 11: Changes in $`F_{k}`$, the number of failed $`f\hspace{0pt}i\hspace{0pt}n_{k}`$ pointers, due to joins, failures and stabilizations.
We now turn to estimating the fraction of finger pointers which point to failed nodes. As we will see this is an important quantity for predicting lookups, since failed fingers cause timeouts and increase the lookup length. However, we only need to consider fingers pointing to dead nodes. Unlike members of the successor list, alive fingers even if outdated, always bring a query closer to the destination and do not affect consistency or substantially even the lookup length. Therefore we consider fingers in only two states, alive or dead (failed). By our implementation of the stabilization protocol (see Sections [](#S3.SS1 "III-A Joins, Failures & Ring Stabilization ‣ III Our Implementation of Chord ‣ An Analytical Study of a Structured Overlay in the Presence of Dynamic Membership This work is funded by the European 6th FP EVERGROW project. ©IEEE. Personal use of this material is permitted. However, permission to reprint/republish this material for advertising or promotional purposes or for creating new collective works for resale or redistribution to servers or lists, or to reuse any copyrighted component of this work in other works, must be obtained from the IEEE.")
III-A
and [](#S3.SS2 "III-B Lookups and Stabilization of Fingers ‣ III Our Implementation of Chord ‣ An Analytical Study of a Structured Overlay in the Presence of Dynamic Membership This work is funded by the European 6th FP EVERGROW project. ©IEEE. Personal use of this material is permitted. However, permission to reprint/republish this material for advertising or promotional purposes or for creating new collective works for resale or redistribution to servers or lists, or to reuse any copyrighted component of this work in other works, must be obtained from the IEEE.")
III-B
), fingers and successors are stabilized entirely independently of each other to simplify the analysis. Thus even though the first finger is also always the first successor, this information is not used by the node in updating the finger. Fingers of nodes far apart are independent of each other. Fingers of adjacent nodes can be correlated and we take this into account. The only assumption in this section is in connection with the join protocol as explained below.
Let $`f_{k}\hspace{0pt}{(r,\alpha)}`$ denote the fraction of nodes whose $`k^{t\hspace{0pt}h}`$ finger points to a failed node and $`F_{k}\hspace{0pt}{(r,\alpha)}`$ denote the respective number. For notational simplicity, we write these as simply $`F_{k}`$ and $`f_{k}`$. We can predict this function for any $`k`$ by again estimating the gain and loss terms for this quantity, caused by a join, failure or stabilization event, and keeping only the most relevant terms. These are listed in Table [III](#S4.T3 "TABLE III ‣ IV-D Failure of Fingers ‣ IV The Analysis ‣ An Analytical Study of a Structured Overlay in the Presence of Dynamic Membership This work is funded by the European 6th FP EVERGROW project. ©IEEE. Personal use of this material is permitted. However, permission to reprint/republish this material for advertising or promotional purposes or for creating new collective works for resale or redistribution to servers or lists, or to reuse any copyrighted component of this work in other works, must be obtained from the IEEE.") and illustrated in Fig. [11](#S4.F11 "Figure 11 ‣ IV-D Failure of Fingers ‣ IV The Analysis ‣ An Analytical Study of a Structured Overlay in the Presence of Dynamic Membership This work is funded by the European 6th FP EVERGROW project. ©IEEE. Personal use of this material is permitted. However, permission to reprint/republish this material for advertising or promotional purposes or for creating new collective works for resale or redistribution to servers or lists, or to reuse any copyrighted component of this work in other works, must be obtained from the IEEE.")
TABLE III: The relevant gain and loss terms for $`F_{k}`$, the number of nodes whose $`k\hspace{0pt}t\hspace{0pt}h`$ fingers are pointing to a failed node for $`k > 1`$.
<table>
<thead>
<tr>
<th>Fk(t+Δt)subscript𝐹𝑘𝑡Δ𝑡F_{k}(t+\Delta t)</th>
<th>Probability of Occurence</th>
</tr>
</thead>
<tbody>
<tr>
<th>=Fk(t)+1absentsubscript𝐹𝑘𝑡1=F_{k}(t)+1</th>
<td>c3.1=(λjNΔt)∑i=1kp𝑗𝑜𝑖𝑛(i,k)fisubscript𝑐3.1subscript𝜆𝑗𝑁Δ𝑡superscriptsubscript𝑖1𝑘subscript𝑝𝑗𝑜𝑖𝑛𝑖𝑘subscript𝑓𝑖c_{3.1}=(\lambda_{j}N\Delta t)\sum_{i=1}^{k}p_{\it join}(i,k)f_{i}</td>
</tr>
<tr>
<th>=Fk(t)−1absentsubscript𝐹𝑘𝑡1=F_{k}(t)-1</th>
<td>c3.2=(1−α)1ℳfk(λsNΔt)subscript𝑐3.21𝛼1ℳsubscript𝑓𝑘subscript𝜆𝑠𝑁Δ𝑡c_{3.2}=(1-\alpha)\frac{1}{{\cal M}}f_{k}(\lambda_{s}N\Delta t)</td>
</tr>
<tr>
<th>=Fk(t)+1absentsubscript𝐹𝑘𝑡1=F_{k}(t)+1</th>
<td>c3.3=(1−fk)2[1−p1(k)](λfNΔt)subscript𝑐3.3superscript1subscript𝑓𝑘2delimited-[]1subscript𝑝1𝑘subscript𝜆𝑓𝑁Δ𝑡c_{3.3}=(1-f_{k})^{2}[1-p_{1}(k)](\lambda_{f}N\Delta t)</td>
</tr>
<tr>
<th>=Fk(t)+2absentsubscript𝐹𝑘𝑡2=F_{k}(t)+2</th>
<td>c3.4=(1−fk)2(p1(k)−p2(k))(λfNΔt)subscript𝑐3.4superscript1subscript𝑓𝑘2subscript𝑝1𝑘subscript𝑝2𝑘subscript𝜆𝑓𝑁Δ𝑡c_{3.4}=(1-f_{k})^{2}(p_{1}(k)-p_{2}(k))(\lambda_{f}N\Delta t)</td>
</tr>
<tr>
<th>=Fk(t)+3absentsubscript𝐹𝑘𝑡3=F_{k}(t)+3</th>
<td>c3.5=(1−fk)2(p2(k)−p3(k))(λfNΔt)subscript𝑐3.5superscript1subscript𝑓𝑘2subscript𝑝2𝑘subscript𝑝3𝑘subscript𝜆𝑓𝑁Δ𝑡c_{3.5}=(1-f_{k})^{2}(p_{2}(k)-p_{3}(k))(\lambda_{f}N\Delta t)</td>
</tr>
<tr>
<th>=Fk(t)absentsubscript𝐹𝑘𝑡=F_{k}(t)</th>
<td>1−(c3.1+c3.2+c3.3+c3.4+c3.5)1subscript𝑐3.1subscript𝑐3.2subscript𝑐3.3subscript𝑐3.4subscript𝑐3.51-(c_{3.1}+c_{3.2}+c_{3.3}+c_{3.4}+c_{3.5})</td>
</tr>
</tbody>
</table>
A join event can play a role here by increasing the number of $`F_{k}`$ pointers if the successor of the joinee had a failed $`i^{t\hspace{0pt}h}`$ pointer (occurs with probability $`f_{i}`$) and the joinee replicated this from the successor as the joinee’s $`k^{t\hspace{0pt}h}`$ pointer. (occurs with probability $`p_{join}\hspace{0pt}{(i,k)}`$ from property [IV.4](#S4.Thmproperty4 "Property IV.4 ‣ IV-A Distributional Properties of Inter-Node Distances ‣ IV The Analysis ‣ An Analytical Study of a Structured Overlay in the Presence of Dynamic Membership This work is funded by the European 6th FP EVERGROW project. ©IEEE. Personal use of this material is permitted. However, permission to reprint/republish this material for advertising or promotional purposes or for creating new collective works for resale or redistribution to servers or lists, or to reuse any copyrighted component of this work in other works, must be obtained from the IEEE.")). For large enough $`k`$, this probability is one only for $`p_{join}\hspace{0pt}{(k,k)}`$, that is, the new joinee mostly only replicates the successor’s $`k`$th pointer as its own $`k^{t\hspace{0pt}h}`$ pointer. This is what we consider here.
A stabilization evicts a failed pointer if there was one to begin with. The stabilization rate is divided by $`\mathcal{M}`$, since a node stabilizes any one finger randomly, every time it decides to stabilize a finger at rate $`{({1 - \alpha})}\hspace{0pt}\lambda_{s}`$.
Given a node $`n`$ with an alive $`k^{t\hspace{0pt}h}`$ finger (occurs with probability $`1 - f_{k}`$), when the node pointed to by that finger fails, the number of failed $`k^{t\hspace{0pt}h}`$ fingers ($`F_{k}`$) increases. The amount of this increase depends on the number of immediate predecessors of $`n`$ that were pointing to the failed node with their $`k^{t\hspace{0pt}h}`$ finger. That number of predecessors could be $`0`$, $`1`$, $`2`$,.. etc. Using property [IV.3](#S4.Thmproperty3 "Property IV.3 ‣ IV-A Distributional Properties of Inter-Node Distances ‣ IV The Analysis ‣ An Analytical Study of a Structured Overlay in the Presence of Dynamic Membership This work is funded by the European 6th FP EVERGROW project. ©IEEE. Personal use of this material is permitted. However, permission to reprint/republish this material for advertising or promotional purposes or for creating new collective works for resale or redistribution to servers or lists, or to reuse any copyrighted component of this work in other works, must be obtained from the IEEE.") the respective probabilities of those cases are: $`1 - {p_{1}\hspace{0pt}{(k)}}`$, $`{p_{1}\hspace{0pt}{(k)}} - {p_{2}\hspace{0pt}{(k)}}`$, $`{p_{2}\hspace{0pt}{(k)}} - {p_{3}\hspace{0pt}{(k)}}`$,… etc.
Solving for $`f_{k}`$ in the steady state, we get:
<table>
<tbody>
<tr>
<td></td>
<td>fk=[2P~rep(k)+2−p𝑗𝑜𝑖𝑛(k)+r(1−α)ℳ]2(1+P~rep(k))−[2P~rep(k)+2−p𝑗𝑜𝑖𝑛(k)+r(1−α)ℳ]2−4(1+P~rep(k))22(1+P~rep(k))subscript𝑓𝑘delimited-[]2subscript~𝑃𝑟𝑒𝑝𝑘2subscript𝑝𝑗𝑜𝑖𝑛𝑘𝑟1𝛼ℳ21subscript~𝑃𝑟𝑒𝑝𝑘superscriptdelimited-[]2subscript~𝑃𝑟𝑒𝑝𝑘2subscript𝑝𝑗𝑜𝑖𝑛𝑘𝑟1𝛼ℳ24superscript1subscript~𝑃𝑟𝑒𝑝𝑘221subscript~𝑃𝑟𝑒𝑝𝑘\begin{split}&f_{k}=\frac{\left[2\tilde{P}_{rep}(k)+2-p_{\it join}(k)+\frac{r(1-\alpha)}{{\cal M}}\right]}{2(1+\tilde{P}_{rep}(k))}\\ &-\frac{\sqrt{\left[2\tilde{P}_{rep}(k)+2-p_{\it join}(k)+\frac{r(1-\alpha)}{{\cal M}}\right]^{2}-4(1+\tilde{P}_{rep}(k))^{2}}}{2(1+\tilde{P}_{rep}(k))}\end{split}</td>
<td></td>
<td>(10)</td>
</tr>
</tbody>
</table>
where $`{{\overset{\sim}{P}}_{r\hspace{0pt}e\hspace{0pt}p}\hspace{0pt}{(k)}} = {\Sigma\hspace{0pt}p_{i}\hspace{0pt}{(k)}}`$. In practice, it is enough to keep the first three terms in this sum. To first order in $`\frac{1}{r}`$ we have, in analogy to ([6](#S4.E6 "In IV-B Successor Pointers ‣ IV The Analysis ‣ An Analytical Study of a Structured Overlay in the Presence of Dynamic Membership This work is funded by the European 6th FP EVERGROW project. ©IEEE. Personal use of this material is permitted. However, permission to reprint/republish this material for advertising or promotional purposes or for creating new collective works for resale or redistribution to servers or lists, or to reuse any copyrighted component of this work in other works, must be obtained from the IEEE.")),
<table>
<tbody>
<tr>
<td></td>
<td>fk≈(1+P~rep(k))ℳ(1−α)rsubscript𝑓𝑘1subscript~𝑃𝑟𝑒𝑝𝑘ℳ1𝛼𝑟f_{k}\approx\frac{(1+\tilde{P}_{rep}(k)){\cal M}}{(1-\alpha)r}</td>
<td></td>
<td>(11)</td>
</tr>
</tbody>
</table>
This expression simply says that the fraction of dead fingers is inversely proportional to the rate of finger stabilizations, $`{({1 - \alpha})}\hspace{0pt}r`$, and proportional to how many fingers there are to stabilize, $`\mathcal{M}`$, with the proportionality factor $`({1 + {{\overset{\sim}{P}}_{r\hspace{0pt}e\hspace{0pt}p}\hspace{0pt}{(k)}}})`$ depending only on $`\rho`$.
To sum up, the computation of the fraction of dead $`k^{t\hspace{0pt}h}`$ finger pointers is analogous to the calculation of the fraction of wrong first successor pointer, albeit a bit more involved. No recursion is involved, in contrast to the calculation of the fraction of wrong higher successor pointers. The above expressions, ([10](#S4.E10 "In IV-D Failure of Fingers ‣ IV The Analysis ‣ An Analytical Study of a Structured Overlay in the Presence of Dynamic Membership This work is funded by the European 6th FP EVERGROW project. ©IEEE. Personal use of this material is permitted. However, permission to reprint/republish this material for advertising or promotional purposes or for creating new collective works for resale or redistribution to servers or lists, or to reuse any copyrighted component of this work in other works, must be obtained from the IEEE.")) match very well with the simulation results (Fig. [13](#S4.F13 "Figure 13 ‣ IV-E Cost of Finger Stabilizations and Lookups ‣ IV The Analysis ‣ An Analytical Study of a Structured Overlay in the Presence of Dynamic Membership This work is funded by the European 6th FP EVERGROW project. ©IEEE. Personal use of this material is permitted. However, permission to reprint/republish this material for advertising or promotional purposes or for creating new collective works for resale or redistribution to servers or lists, or to reuse any copyrighted component of this work in other works, must be obtained from the IEEE.")).
###
IV-E
Cost of Finger Stabilizations and Lookups
[Refer to caption](/html/0710.0270/assets/x9.png)
Figure 12: Cases that a lookup can encounter with the respective probabilities and costs.
[Refer to caption](/html/0710.0270/assets/x10.png)
[Refer to caption](/html/0710.0270/assets/x11.png)
Figure 13: Theory and simulation for probability of failure of the $`k^{t\hspace{0pt}h}`$ finger $`f_{k}\hspace{0pt}{(r,\alpha)}`$, and the lookup length $`L\hspace{0pt}{(r,\alpha)}`$.
In this section, we demonstrate how the information about the failed fingers and successors can be used to predict the cost of stabilizations, lookups or in general the cost for reaching any key in the id space. By cost we mean the number of hops needed to reach the destination including the number of timeouts encountered en-route. Timeouts occur every time a query is passed to a dead node. The node does not answer and the originator of the query has to use another finger instead. For this analysis, we consider timeouts and hops to add equally to the cost. We can easily generalize this analysis to investigate the case when a timeout costs some factor $`\gamma`$ times the cost of a hop.
Define $`C_{t}\hspace{0pt}{(r,\alpha)}`$ (also denoted by $`C_{t}`$) to be the expected cost for a given node to reach some target key which is $`t`$ keys away from it (which means reaching the first successor of this key). For example, $`C_{1}`$ would then be the cost of looking up the adjacent key ($`1`$ key away). Since the adjacent key is always stored at the first alive successor, therefore if the first successor is alive (which occurs with probability $`1 - d_{1}`$), the cost will be $`1`$ hop. If the first successor is dead but the second is alive (occurs with probability $`d_{1}\hspace{0pt}{({1 - d_{2}})}`$), the cost will be 1 hop + 1 timeout = $`2`$ and the *expected* cost is $`{2 \times d_{1}}\hspace{0pt}{({1 - d_{2}})}`$ and so forth. Therefore, we have $`C_{1} = {{1 - d_{1}} + {{2 \times d_{1}}\hspace{0pt}{({1 - d_{2}})}} + {{3 \times d_{1}}\hspace{0pt}d_{2}\hspace{0pt}{({1 - d_{3}})}} + \cdots} \approx {1 + d_{1}} = {1 + {1/{({\alpha\hspace{0pt}r})}}}`$.
To find the expected cost for reaching a general distance $`t`$ we need to closely follow the Chord protocol, which would lookup $`t`$ by first finding the closest preceding finger. For the purposes of the analysis, we will find it easier to think in terms of the closest preceding start. Let us hence define $`\xi`$ to be the *s*tart of the finger (say the $`k^{t\hspace{0pt}h}`$) that most closely precedes $`t`$. Hence $`\xi = {2^{k - 1} + n}`$ and $`t = {\xi + m}`$ i.e., there are $`m`$ keys between the sought target $`t`$ and the start of the closest preceding finger. With that, we can write a recursion relation for $`C_{\xi + m}`$ as follows:
<table>
<tbody>
<tr>
<td></td>
<td>Cξ+m=Cξ[1−a(m)]+(1−fk)a(m)[1+∑i=0m−1bc(i,m)Cm−i]+fka(m)[1+∑i=1k−1hk(i)∑l=0ξ/2i−1bc(l,ξ/2i)(1+(i−1)+Cξi−l+m)+O(hk(k))]subscript𝐶𝜉𝑚subscript𝐶𝜉delimited-[]1𝑎𝑚1subscript𝑓𝑘𝑎𝑚delimited-[]1superscriptsubscript𝑖0𝑚1𝑏𝑐𝑖𝑚subscript𝐶𝑚𝑖subscript𝑓𝑘𝑎𝑚delimited-[]1superscriptsubscript𝑖1𝑘1subscriptℎ𝑘𝑖superscriptsubscript𝑙0𝜉superscript2𝑖1𝑏𝑐𝑙𝜉superscript2𝑖1𝑖1subscript𝐶subscript𝜉𝑖𝑙𝑚𝑂subscriptℎ𝑘𝑘\begin{split}&C_{\xi+m}=C_{\xi}\left[1-a(m)\right]\\ &+(1-f_{k})a(m)\left[1+\sum_{i=0}^{m-1}bc(i,m)C_{m-i}\right]\\ &+f_{k}a(m)\biggl{[}1+\sum_{i=1}^{k-1}h_{k}(i)\\ &\sum_{l=0}^{\xi/2^{i}-1}bc(l,\xi/2^{i})(1+(i-1)+C_{\xi_{i}-l+m})+O(h_{k}(k))\biggr{]}\end{split}</td>
<td></td>
<td>(12)</td>
</tr>
</tbody>
</table>
where $`\xi_{i} \equiv {\sum_{m = {1,i}}{\xi/2^{m}}}`$ and $`h_{k}\hspace{0pt}{(i)}`$ is the probability that a node is forced to use its $`k - i^{t\hspace{0pt}h}`$ finger owing to the death of its $`k^{t\hspace{0pt}h}`$ finger. The probabilities $`a,b,{b\hspace{0pt}c}`$ have already been introduced in Section [IV](#S4 "IV The Analysis ‣ An Analytical Study of a Structured Overlay in the Presence of Dynamic Membership This work is funded by the European 6th FP EVERGROW project. ©IEEE. Personal use of this material is permitted. However, permission to reprint/republish this material for advertising or promotional purposes or for creating new collective works for resale or redistribution to servers or lists, or to reuse any copyrighted component of this work in other works, must be obtained from the IEEE."), and we define the probability $`h_{k}\hspace{0pt}{(i)}`$ below.
The lookup equation though rather complicated at first sight merely accounts for all the possibilities that a Chord lookup will encounter, and deals with them exactly as the protocol dictates.
The first term (Fig. [12](#S4.F12 "Figure 12 ‣ IV-E Cost of Finger Stabilizations and Lookups ‣ IV The Analysis ‣ An Analytical Study of a Structured Overlay in the Presence of Dynamic Membership This work is funded by the European 6th FP EVERGROW project. ©IEEE. Personal use of this material is permitted. However, permission to reprint/republish this material for advertising or promotional purposes or for creating new collective works for resale or redistribution to servers or lists, or to reuse any copyrighted component of this work in other works, must be obtained from the IEEE.") (a)) accounts for the eventuality that there is no node intervening between $`\xi`$ and $`\xi + m`$ (occurs with probability $`1 - {a\hspace{0pt}{(m)}}`$). In this case, the cost of looking for $`\xi + m`$ is the same as the cost for looking for $`\xi`$.
The second term (Fig. [12](#S4.F12 "Figure 12 ‣ IV-E Cost of Finger Stabilizations and Lookups ‣ IV The Analysis ‣ An Analytical Study of a Structured Overlay in the Presence of Dynamic Membership This work is funded by the European 6th FP EVERGROW project. ©IEEE. Personal use of this material is permitted. However, permission to reprint/republish this material for advertising or promotional purposes or for creating new collective works for resale or redistribution to servers or lists, or to reuse any copyrighted component of this work in other works, must be obtained from the IEEE.") (b)) accounts for the situation when a node does intervene in between (with probability $`a\hspace{0pt}{(m)}`$), and this node is alive (with probability $`1 - f_{k}`$). Then the query is passed on to this node (with $`1`$ added to register the increase in the number of hops) and then the cost depends on the length of the distance between this node and $`t`$.
The third term (Fig. [12](#S4.F12 "Figure 12 ‣ IV-E Cost of Finger Stabilizations and Lookups ‣ IV The Analysis ‣ An Analytical Study of a Structured Overlay in the Presence of Dynamic Membership This work is funded by the European 6th FP EVERGROW project. ©IEEE. Personal use of this material is permitted. However, permission to reprint/republish this material for advertising or promotional purposes or for creating new collective works for resale or redistribution to servers or lists, or to reuse any copyrighted component of this work in other works, must be obtained from the IEEE.") (c)) accounts for the case when the intervening node is dead (with probability $`f_{k}`$). Then the cost increases by $`1`$ (for a timeout) and the query needs to find an alternative lower finger that most closely precedes the target. Let the $`k - i^{t\hspace{0pt}h}`$ finger (for some $`i`$, $`1 \leq i \leq {k - 1}`$) be such a finger. This happens with probability $`h_{k}\hspace{0pt}{(i)}`$ i.e., the probability that the lookup is passed back to the $`k - i^{t\hspace{0pt}h}`$ finger either because the intervening fingers are dead or share the same finger table entry as the $`k^{t\hspace{0pt}h}`$ finger is denoted by $`h_{k}\hspace{0pt}{(i)}`$. The start of the $`k - i^{t\hspace{0pt}h}`$ finger is at $`\xi/2^{i}`$ and the distance between $`\xi/2^{i}`$ and $`\xi`$ is equal to $`\sum_{m = {1,i}}{\xi/2^{m}}`$ which we denote by $`\xi_{i}`$. Therefore, the distance from the start of the $`k - i^{t\hspace{0pt}h}`$ to the target is equal to $`\xi_{i} + m`$. However, note that $`{f\hspace{0pt}i\hspace{0pt}n_{k - i}}.{n\hspace{0pt}o\hspace{0pt}d\hspace{0pt}e}`$ could be $`l`$ keys away (with probability $`b\hspace{0pt}c\hspace{0pt}{(l,{\xi/2^{i}})}`$) from $`{f\hspace{0pt}i\hspace{0pt}n_{k - i}}.{s\hspace{0pt}t\hspace{0pt}a\hspace{0pt}r\hspace{0pt}t}`$ (for some $`l`$, $`0 \leq l < {\xi/2^{i}}`$). Therefore, after making one hop to $`{f\hspace{0pt}i\hspace{0pt}n_{k - i}}.{n\hspace{0pt}o\hspace{0pt}d\hspace{0pt}e}`$, the remaining distance to the target is $`{\xi_{i} + m} - l`$. The increase in cost for this operation is $`1 + {({i - 1})}`$; the $`1`$ indicates the cost of taking up the query again by $`{f\hspace{0pt}i\hspace{0pt}n_{k - i}}.{n\hspace{0pt}o\hspace{0pt}d\hspace{0pt}e}`$, and the $`i - 1`$ indicates the cost for trying and discarding each of the $`i - 1`$ intervening fingers. The probability $`h_{k}\hspace{0pt}{(i)}`$ is easy to compute given property [IV.2](#S4.Thmproperty2 "Property IV.2 ‣ IV-A Distributional Properties of Inter-Node Distances ‣ IV The Analysis ‣ An Analytical Study of a Structured Overlay in the Presence of Dynamic Membership This work is funded by the European 6th FP EVERGROW project. ©IEEE. Personal use of this material is permitted. However, permission to reprint/republish this material for advertising or promotional purposes or for creating new collective works for resale or redistribution to servers or lists, or to reuse any copyrighted component of this work in other works, must be obtained from the IEEE.") and the expression for the $`f_{k}`$’s computed in the previous section.
<table>
<tbody>
<tr>
<td></td>
<td>hk(i)=a(ξ/2i)(1−fk−i)×Πs=1,i−1(1−a(ξ/2s)+a(ξ/2s)fk−s),i<khk(k)=Πs=1,k−1(1−a(ξ/2s)+a(ξ/2s)fk−s)formulae-sequencesubscriptℎ𝑘𝑖𝑎𝜉superscript2𝑖1subscript𝑓𝑘𝑖subscriptΠ𝑠1𝑖11𝑎𝜉superscript2𝑠𝑎𝜉superscript2𝑠subscript𝑓𝑘𝑠𝑖𝑘subscriptℎ𝑘𝑘subscriptΠ𝑠1𝑘11𝑎𝜉superscript2𝑠𝑎𝜉superscript2𝑠subscript𝑓𝑘𝑠\begin{split}h_{k}(i)=&a(\xi/2^{i})(1-f_{k-i})\\ \times&\Pi_{s=1,i-1}(1-a(\xi/2^{s})+a(\xi/2^{s})f_{k-s}),i<k\\ h_{k}(k)=&\Pi_{s=1,k-1}(1-a(\xi/2^{s})+a(\xi/2^{s})f_{k-s})\end{split}</td>
<td></td>
<td>(13)</td>
</tr>
</tbody>
</table>
In ([13](#S4.E13 "In IV-E Cost of Finger Stabilizations and Lookups ‣ IV The Analysis ‣ An Analytical Study of a Structured Overlay in the Presence of Dynamic Membership This work is funded by the European 6th FP EVERGROW project. ©IEEE. Personal use of this material is permitted. However, permission to reprint/republish this material for advertising or promotional purposes or for creating new collective works for resale or redistribution to servers or lists, or to reuse any copyrighted component of this work in other works, must be obtained from the IEEE.")) we account for all the reasons that a node may have to use its $`k - i^{t\hspace{0pt}h}`$ finger instead of its $`k^{t\hspace{0pt}h}`$ finger. This could happen because the intervening fingers were either dead or not distinct. The probabilities $`h_{k}\hspace{0pt}{(i)}`$ satisfy the constraint $`{\sum_{i = 1}^{k}{h_{k}\hspace{0pt}{(i)}}} = 1`$ since clearly, either a node uses any one of its fingers or it doesn’t. This latter probability is $`h_{k}\hspace{0pt}{(k)}`$, that is the probability that a node cannot use any earlier entry in its finger table. In this case, $`n`$ proceeds to its successor list. The query is now passed on to the first alive successor and the new cost is a function of the distance of this node from the target $`t`$. We indicate this case by the last term in [12](#S4.E12 "In IV-E Cost of Finger Stabilizations and Lookups ‣ IV The Analysis ‣ An Analytical Study of a Structured Overlay in the Presence of Dynamic Membership This work is funded by the European 6th FP EVERGROW project. ©IEEE. Personal use of this material is permitted. However, permission to reprint/republish this material for advertising or promotional purposes or for creating new collective works for resale or redistribution to servers or lists, or to reuse any copyrighted component of this work in other works, must be obtained from the IEEE.") which is $`O\hspace{0pt}{({h_{k}\hspace{0pt}{(k)}})}`$. This can again be computed from the inter-node distribution and from the functions $`d_{k}\hspace{0pt}{(r,\alpha)}`$ computed earlier. However in practice, the probability for this is extremely small except for targets very close to $`n`$. Hence this does not significantly affect the value of general lookups and we ignore it in our analysis.
The cost for general lookups is hence
<table>
<tbody>
<tr>
<td></td>
<td>L(r,α)=Σi=1𝒦−1Ci(r,α)𝒦𝐿𝑟𝛼superscriptsubscriptΣ𝑖1𝒦1subscript𝐶𝑖𝑟𝛼𝒦L(r,\alpha)=\frac{\Sigma_{i=1}^{{\cal K}-1}C_{i}(r,\alpha)}{\cal K}</td>
<td></td>
</tr>
</tbody>
</table>
The lookup equation is solved recursively numerically, given the coefficients and $`C_{1}`$. In Fig. [13](#S4.F13 "Figure 13 ‣ IV-E Cost of Finger Stabilizations and Lookups ‣ IV The Analysis ‣ An Analytical Study of a Structured Overlay in the Presence of Dynamic Membership This work is funded by the European 6th FP EVERGROW project. ©IEEE. Personal use of this material is permitted. However, permission to reprint/republish this material for advertising or promotional purposes or for creating new collective works for resale or redistribution to servers or lists, or to reuse any copyrighted component of this work in other works, must be obtained from the IEEE."), we compare theoretical results with simulation for $`N = 1000`$. It is seen that the theory matches the simulation results very well.
In Fig. [14](#S4.F14 "Figure 14 ‣ IV-E Cost of Finger Stabilizations and Lookups ‣ IV The Analysis ‣ An Analytical Study of a Structured Overlay in the Presence of Dynamic Membership This work is funded by the European 6th FP EVERGROW project. ©IEEE. Personal use of this material is permitted. However, permission to reprint/republish this material for advertising or promotional purposes or for creating new collective works for resale or redistribution to servers or lists, or to reuse any copyrighted component of this work in other works, must be obtained from the IEEE.") we also show the theoretical predictions for some larger values of $`N`$. From the structure of Equation [12](#S4.E12 "In IV-E Cost of Finger Stabilizations and Lookups ‣ IV The Analysis ‣ An Analytical Study of a Structured Overlay in the Presence of Dynamic Membership This work is funded by the European 6th FP EVERGROW project. ©IEEE. Personal use of this material is permitted. However, permission to reprint/republish this material for advertising or promotional purposes or for creating new collective works for resale or redistribution to servers or lists, or to reuse any copyrighted component of this work in other works, must be obtained from the IEEE."), it is clear that the dependence of the average lookup on churn comes entirely from the presence of the terms $`f_{k}`$. Since $`f_{k} \sim f`$ is independent of $`k`$ for large fingers, we can approximate the average lookup length by the functional form $`{L\hspace{0pt}{(r,\alpha)}} = {A + {B\hspace{0pt}f} + {C\hspace{0pt}f^{2}} + \cdots}`$. The coefficients $`A,B,C`$ etc can be recursively computed by solving the lookup equation to the required order in $`f`$ and depend only on $`N`$ the number of nodes, $`1 - \rho`$ the density of peers and $`b`$ the base or equivalently the size of the finger table of each node. The advantage of writing the lookup length this way is that churn-specific details such as how new joinees construct a finger table or how exactly stabilizations are done in the system, can be isolated in the expression for $`f`$. If we were to change our stabilization strategy for example \[[9](#bib.bib9)\], we could immediately estimate the lookup length by plugging in the new expression for $`f`$ in the above relation.
The coefficient $`A`$, which is the lookup cost without churn can be obtained very precisely for any base $`b`$, from analyzing ([12](#S4.E12 "In IV-E Cost of Finger Stabilizations and Lookups ‣ IV The Analysis ‣ An Analytical Study of a Structured Overlay in the Presence of Dynamic Membership This work is funded by the European 6th FP EVERGROW project. ©IEEE. Personal use of this material is permitted. However, permission to reprint/republish this material for advertising or promotional purposes or for creating new collective works for resale or redistribution to servers or lists, or to reuse any copyrighted component of this work in other works, must be obtained from the IEEE.")) in the zero-churn case. This analysis is rather laborious and will be presented elsewhere \[[9](#bib.bib9)\]. It confirms the well-known result $`A = {\frac{1}{2}\hspace{0pt}{\log_{2}N}}`$ and in addition reproduces small deviations from this behavior previously observed by us in numerical simulations \[[7](#bib.bib7)\]. The values of $`A`$ in Fig. [14](#S4.F14 "Figure 14 ‣ IV-E Cost of Finger Stabilizations and Lookups ‣ IV The Analysis ‣ An Analytical Study of a Structured Overlay in the Presence of Dynamic Membership This work is funded by the European 6th FP EVERGROW project. ©IEEE. Personal use of this material is permitted. However, permission to reprint/republish this material for advertising or promotional purposes or for creating new collective works for resale or redistribution to servers or lists, or to reuse any copyrighted component of this work in other works, must be obtained from the IEEE.") are taken from this analysis.
$`B`$ can be qualitatively estimated as follows : every sufficiently long finger is dead with some finite probability $`f`$ given by ([10](#S4.E10 "In IV-D Failure of Fingers ‣ IV The Analysis ‣ An Analytical Study of a Structured Overlay in the Presence of Dynamic Membership This work is funded by the European 6th FP EVERGROW project. ©IEEE. Personal use of this material is permitted. However, permission to reprint/republish this material for advertising or promotional purposes or for creating new collective works for resale or redistribution to servers or lists, or to reuse any copyrighted component of this work in other works, must be obtained from the IEEE.")). If $`A`$ is the average value of the lookup length without churn, then each look-up encounters $`f\hspace{0pt}A`$ dead fingers on average. This estimate predicts a look-up cost of approximately $`A\hspace{0pt}{({1 + f})}`$, giving $`B = A`$ and $`C`$ and all other coefficients equal to $`0`$..
[Refer to caption](/html/0710.0270/assets/x12.png)
Figure 14: Lookup cost, theoretical curve, for $`1000`$,$`2000`$,$`4000`$,$`8000`$ and $`16000`$ peers. The rationale for the fits is explained in the text.
In Fig. [14](#S4.F14 "Figure 14 ‣ IV-E Cost of Finger Stabilizations and Lookups ‣ IV The Analysis ‣ An Analytical Study of a Structured Overlay in the Presence of Dynamic Membership This work is funded by the European 6th FP EVERGROW project. ©IEEE. Personal use of this material is permitted. However, permission to reprint/republish this material for advertising or promotional purposes or for creating new collective works for resale or redistribution to servers or lists, or to reuse any copyrighted component of this work in other works, must be obtained from the IEEE.") we show that the best fit to the data is obtained in fact by taking $`B = A`$ and $`C = {3\hspace{0pt}A}`$. The expression for $`f`$ is taken from [10](#S4.E10 "In IV-D Failure of Fingers ‣ IV The Analysis ‣ An Analytical Study of a Structured Overlay in the Presence of Dynamic Membership This work is funded by the European 6th FP EVERGROW project. ©IEEE. Personal use of this material is permitted. However, permission to reprint/republish this material for advertising or promotional purposes or for creating new collective works for resale or redistribution to servers or lists, or to reuse any copyrighted component of this work in other works, must be obtained from the IEEE.") for large $`k`$ (for a system with $`20`$ fingers, the expression for $`f_{k}`$ becomes independent of $`k`$ for $`k \geq 13`$). In general, as mentioned earlier, $`B`$ and $`C`$ can be obtained accurately for any value of the system parameters by the numerical solution of Eq. [12](#S4.E12 "In IV-E Cost of Finger Stabilizations and Lookups ‣ IV The Analysis ‣ An Analytical Study of a Structured Overlay in the Presence of Dynamic Membership This work is funded by the European 6th FP EVERGROW project. ©IEEE. Personal use of this material is permitted. However, permission to reprint/republish this material for advertising or promotional purposes or for creating new collective works for resale or redistribution to servers or lists, or to reuse any copyrighted component of this work in other works, must be obtained from the IEEE.") to the required order.
## V Discussion and Conclusion
In this paper we have presented a detailed theoretical analysis of a DHT-based P2P system, Chord, using a fluid model. The technique for deriving the fluid model has been borrowed from the master equation approach of physics, which helps in systematically taking different dynamical effects into account. This analysis differs from previous theoretical work done on DHTs in that it aims not at establishing bounds, but on precise determination of the relevant quantities in this dynamically evolving system. From the match of our theory and the simulations, it can be seen that we can predict with an accuracy of greater than $`1\%`$ in most cases. Though this analysis is not exact, since it takes only some (and not all) correlations into account, yet it provides a methodology for keeping track of most of the relevant details of the system. We expect that a similar analysis can be done for most other DHT’s, thus helping to establish quantitative guidelines for their comparison.
The main conclusions for the analysis of Chord in a statistically steady state are the following.
######
Property V.1
As a function of $`r`$, the ratio of the rate of stabilizations to the rate of failures, the fraction of wrong pointers of any kind (successors or fingers) is to leading order and good approximation $`\text{Const.}/r`$, where the constant depends on the pointer.
######
Property V.2
The probability of break up of a ring can be estimated from the knowledge of the fraction of wrong first successors, wrong second successors, etc. This probability is generally very low when every node has a sufficient number of successors, indicating that Chord is robust against ring break-up.
######
Property V.3
At a given value of $`r`$, the fraction of wrong successors, $`w_{k}`$, and the fraction of dead fingers, $`f_{k}`$, increases with $`k`$. The fraction of wrong successors increases indefinitely, and becomes of order one at $`k`$ about $`\sqrt{r}`$ for the particular stabilization strategy that we have used. The fraction of dead fingers on the other hand tends to a constant for sufficiently large $`k`$.
######
Property V.4
The look-up cost, which is the expected number of hops including time-outs, can be computed by numerical recursion. The fraction of incorrect finger pointers $`f_{k}`$ ($`\sim f`$ for large $`k`$) is a required input for this recursion. The lookup cost tends to the well-known average number of hops without churn when $`f`$ is small (or churn is low) and increases when $`f`$ is large. We show that it can be well described by the formula $`A\hspace{0pt}{({1 + {g\hspace{0pt}{(f)}}})}`$, where $`A`$ is the value of the lookup cost without churn and $`g\hspace{0pt}{(f)}`$ is well approximated by $`f + {3\hspace{0pt}f^{2}}`$ for $`N\operatorname{<<}K`$. In general $`g\hspace{0pt}{(f)}`$ can be obtained accurately to any desired order by solving Eq. [12](#S4.E12 "In IV-E Cost of Finger Stabilizations and Lookups ‣ IV The Analysis ‣ An Analytical Study of a Structured Overlay in the Presence of Dynamic Membership This work is funded by the European 6th FP EVERGROW project. ©IEEE. Personal use of this material is permitted. However, permission to reprint/republish this material for advertising or promotional purposes or for creating new collective works for resale or redistribution to servers or lists, or to reuse any copyrighted component of this work in other works, must be obtained from the IEEE.") recursively to the required order in $`f`$.
######
Property V.5
The preceding note brings out the following simple feature of Chord: under any state of churn, sufficiently long fingers are all dead with essentially the same probability. Hence, in a sufficiently large system, a look-up will almost surely encounter one or more dead fingers, leading to time-outs. For applications where time-outs should be the exception and not the norm, this paper helps in estimating how much stabilization is necessary under a given level of churn, to achieve such a level of performance.
######
Property V.6
The preceding note also brings out the additional feature that by writing the lookup cost in the above simplified form, we can isolate the effects of churn-specific details in the expression for $`f`$. Changing details in the join protocol or changing the maintenance strategy \[[9](#bib.bib9)\] merely cause a change in the expression for $`f`$. The lookup cost with this new strategy can then be immediately assessed for any $`r`$, by plugging in the new expression for $`f`$ in the expression for the lookup cost (as opposed to solving Eq. [12](#S4.E12 "In IV-E Cost of Finger Stabilizations and Lookups ‣ IV The Analysis ‣ An Analytical Study of a Structured Overlay in the Presence of Dynamic Membership This work is funded by the European 6th FP EVERGROW project. ©IEEE. Personal use of this material is permitted. However, permission to reprint/republish this material for advertising or promotional purposes or for creating new collective works for resale or redistribution to servers or lists, or to reuse any copyrighted component of this work in other works, must be obtained from the IEEE.") each time for each value of $`r`$).
The impact of this work can be summarized as follows: given that periodic stabilization is a fundamental technique for topology maintenance in DHTs, the question: ”How often should a DHT node perform periodic stabilization?” is of great practical relevance. The answer to this question depends on several factors. First we need to know where the DHT is deployed, in a LAN, in a cooperative milieu, or among public non-trusting partners, i.e., what is the expected join/failure rate (churn)? Secondly, since DHTs involve different types of stabilizations, we need to know which of these rates is of interest to optimize. For example, in the DHT studied in this paper, there is both ring stabilization as well as finger stabilization. Thirdly, we also need to know whether we have performance goals which require us to know how much stabilization is needed, or constraints on bandwidth which necessitate a knowledge of the expected performance. Previous analytical attempts (see Section [II](#S2 "II Related Work ‣ An Analytical Study of a Structured Overlay in the Presence of Dynamic Membership This work is funded by the European 6th FP EVERGROW project. ©IEEE. Personal use of this material is permitted. However, permission to reprint/republish this material for advertising or promotional purposes or for creating new collective works for resale or redistribution to servers or lists, or to reuse any copyrighted component of this work in other works, must be obtained from the IEEE.")) have addressed these question through the identification of general (algorithm/system-neutral) bounds on stabilization rates.
In this paper, we have taken another point of view. We have traded-off generality for accuracy. That is, we have produced results that can describe to a very high degree of accuracy quantities like the probability of inconsistent look-ups and the expected look-up length as functions of the stabilization and churn rates. Many of the insights we get from this analysis such as most of the points listed above, would be very hard to come by from simulations alone. So for instance, the formulae produced in this paper could directly be used by a system administrator or the person in charge of deploying a DHT as a guide for configuring stabilization rates. While the results are based on Chord, all analyses concerning the ring (break-up and inconsistency) are applicable to many other systems, since consistent hashing on a ring is a recurring component in many other DHTs.
## VI Limitations and Future Work
The main limitation of this work stems from the fact that the results are inherently dependent on the intricate details of the analyzed algorithms. While some changes in the algorithms can be easily accommodated without redoing the analysis (as explained in [V.6](#S5.Thmproperty6 "Property V.6 ‣ V Discussion and Conclusion ‣ An Analytical Study of a Structured Overlay in the Presence of Dynamic Membership This work is funded by the European 6th FP EVERGROW project. ©IEEE. Personal use of this material is permitted. However, permission to reprint/republish this material for advertising or promotional purposes or for creating new collective works for resale or redistribution to servers or lists, or to reuse any copyrighted component of this work in other works, must be obtained from the IEEE.")), others such as a different lookup strategy or a different placement of fingers would necessitate recalculating all the quantities again. However, results concerning the ring-related aspects like successor lists, break-up probability and inter-node distributions are likely to be reusable in other variations of the Chord protocols as well other systems using a ring geometry.
For the future, the authors’ research agenda include the introduction of extensions to the current model to be able to account for locality-awareness and different topology maintenance techniques. Some work towards the latter goal has already been done in \[[9](#bib.bib9)\]. Relatedly, a useful application for this work is to enable systems to dynamically self-tune their stabilization rates and choose the best maintenance technique to achieve a desired hop count.
## References
- \[1\]
Karl Aberer, Anwitaman Datta, and Manfred Hauswirth, *Efficient, self-contained handling of identity in peer-to-peer systems*, IEEE Transactions on Knowledge and Data Engineering 16 (2004), no. 7, 858–869.
- \[2\]
D. Anick, D. Mitra, and M.M. Sondhi, *Stochastic theory of data-handling systems with multiple sources*, Bell Systems Technical Journal 61 (1982), 1871–1894.
- \[3\]
James Aspnes, Zoë Diamadi, and Gauri Shah, *Fault-tolerant routing in peer-to-peer systems*, Proceedings of the twenty-first annual symposium on Principles of distributed computing, ACM Press, 2002, pp. 223–232.
- \[4\]
E. Brockmeyer, H.L. Halstrom, and Arns Jensen, *The life and works of A.K. Erlang*, The Copenhagen Telephone Company, 1948.
- \[5\]
Miguel Castro, Manuel Costa, and Antony Rowstron, *Performance and dependability of structured peer-to-peer overlays*, Proceedings of the 2004 International Conference on Dependable Systems and Networks (DSN’04), IEEE Computer Society, 2004.
- \[6\]
Florence Clévenot and Philippe Nain, *A simple fluid model for the analysis of the squirrel peer-to-peer caching system*, IEEE INFOCOM 2004, 2004.
- \[7\]
Sameh El-Ansary, Erik Aurell, and Seif Haridi, *A physics-inspired performace evaluation of a structured peer-to-peer overlay network*, The International Conference on Parallel and Distributed Computing and Networks (PDCN 2005), 2005.
- \[8\]
Supriya Krishnamurthy, Sameh El-Ansary, Erik Aurell, and Seif Haridi, *A statistical theory of chord under churn*, The 4th International Workshop on Peer-to-Peer Systems (IPTPS’05) (Ithaca, New York), February 2005.
- \[9\]
, *Comparing maintenance strategies for overlays*, Tech. report, Swedish Institute of Computer Science, in preparation 2007.
- \[10\]
Jinyang Li, Jeremy Stribling, Robert Morris, M. Frans Kaashoek, and Thomer M. Gil, *A performance vs. cost framework for evaluating dht design tradeoffs under churn*, Proceedings of the 24th Infocom (Miami, FL), March 2005.
- \[11\]
David Liben-Nowell, Hari Balakrishnan, and David Karger, *Analysis of the evolution of peer-to-peer systems*, ACM Conf. on Principles of Distributed Computing (PODC) (Monterey, CA), July 2002.
- \[12\]
N.G. van Kampen, *Stochastic Processes in Physics and Chemistry*, North-Holland Publishing Company, 1981, ISBN-0-444-86200-5.
- \[13\]
Dongyu Qui and R. Srikant, *Modeling and performance analysis of bittorrent-like peer-to-peer networks*, SIGCOMM’04 (Portland, Oregon), August 2004.
- \[14\]
Sean Rhea, Dennis Geels, Timothy Roscoe, and John Kubiatowicz, *Handling churn in a DHT*, Proceedings of the 2004 USENIX Annual Technical Conference(USENIX ’04) (Boston, Massachusetts, USA), June 2004.
- \[15\]
Ion Stoica, Robert Morris, David Liben-Nowell, David Karger, M. Frans Kaashoek, Frank Dabek, and Hari Balakrishnan, *Chord: A scalable peer-to-peer lookup service for internet applications*, IEEE Transactions on Networking 11 (2003).
- \[16\]
Shengquan Wang, Dong Xuan, and Wei Zhao, *On resilience of structured peer-to-peer systems*, GLOBECOM 2003 - IEEE Global Telecommunications Conference, Dec 2003, pp. 3851–3856. |
0710.5236 | cs.NI | cyber | # Saturation Throughput Analysis of IEEE 802.11 in Presence of Non Ideal Transmission Channel and Capture Effects
F. Daneshgaran, M. Laddomada, F. Mesiti, M. Mondin, and M. Zanolo
This work has been partially supported by Euroconcepts, S.r.l. (http://www.euroconcepts.it) and ISMB through the European IST project, OBAN (http://oban.prz.tu-berlin.de/index.html), IST 6FP Contract No. 001889.To appear on IEEE Transactions on Communications, 2008F. Daneshgaran is with ECE Dept., CSU, Los Angeles, USA.M. Laddomada (laddomada@polito.it), F. Mesiti, M. Mondin and M. Zanolo are with DELEN, Politecnico di Torino, Italy.
###### Abstract
In this paper, we provide a saturation throughput analysis of the IEEE 802.11 protocol at the data link layer by including the impact of both transmission channel and capture effects in Rayleigh fading environment. Impacts of both non-ideal channel and capture effects, specially in an environment of high interference, become important in terms of the actual observed throughput. As far as the 4-way handshaking mechanism is concerned, we extend the multi-dimensional Markovian state transition model characterizing the behavior at the MAC layer by including transmission states that account for packet transmission failures due to errors caused by propagation through the channel. This way, any channel model characterizing the physical transmission medium can be accommodated, including AWGN and fading channels. We also extend the Markov model in order to consider the behavior of the contention window when employing the basic 2-way handshaking mechanism.
Under the usual assumptions regarding the traffic generated per node and independence of packet collisions, we solve for the stationary probabilities of the Markov chain and develop expressions for the saturation throughput as a function of the number of terminals, packet sizes, raw channel error rates, capture probability, and other key system parameters. The theoretical derivations are then compared to simulation results confirming the effectiveness of the proposed models.
###### Index Terms:
Capture, DCF, Distributed Coordination Function, fading, IEEE 802.11, MAC, Rayleigh, rate adaptation, saturation, throughput.
## I Introduction
Wireless Local Area Networks (WLANs) combine data connectivity with user mobility. The Institute of Electrical and Electronic Engineering (IEEE) 802.11 series of standards define both an infrastructure mode, with at least one central Access Point (AP) connected to a wired network, and an ad-hoc or peer-to-peer mode, in which a set of wireless stations communicate directly with another one without needing a central access point or wired network connection.
WLANs have experienced an exponential growth in the recent past \[1-20\]. The fundamental impetus has been to replicate the enormous success of wired LANs with obvious advantages a wireless paradigm would bring. The first series of standards by IEEE was ratified in 1997 but had relatively low data rates (1 or 2 Mbps). Two high rate versions, namely, the IEEE802.11a and IEEE802.11b were later ratified in 1999 and they have found widespread acceptance and use. The enormous success of these standards has prompted multiple working groups to extend some aspects of the basic protocol.
While in the ensuing presentation we shall focus on the basic IEEE802.11b protocol, the analysis is easily extensible to other versions of the basic protocol employing the same Medium Access Control (MAC) mechanisms.
The MAC is central to embedding Quality of Service (QoS) features and has two sub-layers \[[1](#bib.bib1)\]. The Distributed Coordination Function (DCF) and Point Coordination Function (PCF). PCF is generally a complex access method that can be implemented in an infrastructure network. DCF is similar to Carrier Sense Multiple Access with Collision Avoidance (CSMA/CA) method and is further explained in the following.
With this background, let us provide a quick survey of the recent literature related to the problem addressed in this paper. This survey is by no means exhaustive and is meant to simply provide a sampling of the literature in this fertile area.
The most relevant work to what is presented here is \[[2](#bib.bib2)\], where the author provided an analysis of the saturation throughput of the basic 802.11 protocol assuming a two dimensional Markov model at the MAC layer, making several fundamental assumptions: 1) the mobile stations always have something to transmit (i.e., the saturation condition), 2) there are no hidden terminals and there is no capture effect (i.e., a terminal which perceives a higher signal-to-noise ratio (SNR) relative to other terminals captures the channel \[[3](#bib.bib3), [4](#bib.bib4), [5](#bib.bib5)\] and limits access to other terminals), 3) at each transmission attempt and regardless of the number of retransmissions suffered, each packet collides with constant and independent probability, and 4) the transmission channel is ideal and the only cause of packet errors are collisions. Clearly, the second and fourth assumption is not valid in any real setting, specially when there is mobility and when the transmission channel suffers from fading effects.
In \[[6](#bib.bib6)\], the authors extend the work of Bianchi to multiple queues with different contention characteristics in the 802.11e variant of the standard with provisions of QoS. In \[[7](#bib.bib7)\], the authors present an analytical model, in which most new features of the Enhanced Distributed Channel Access (EDCA) in 802.11e such as virtual collision, different arbitration interframe space (AIFS), and different contention windows are taken into account. Based on the model, the throughput performance of differentiated service traffic is analyzed and a recursive method capable of calculating the mean access delay is presented. Both articles referenced assume the transmission channel to be ideal.
In \[[8](#bib.bib8)\], the authors look at the impact of channel induced errors and the received SNR on the achievable throughput in a system with rate adaptation whereby the transmission rate of the terminal is adapted based on either direct or indirect measurements of the link quality. In \[[9](#bib.bib9)\], the authors deal with the extension of Bianchi’s Markov model in order to account for channel errors.
In this article we extend the previous work on the subject as exemplified by the articles referenced above by looking at all the three issues outlined before together. Our assumptions are essentially similar to those of Bianchi \[[2](#bib.bib2)\] with the exception that we do assume the presence of both channel induced errors and capture effects due to the transmission over Rayleigh fading channels. Simulation results confirm the effectiveness of the proposed models. As a reference standard, we use network parameters belonging to the IEEE802.11b protocol, even though the proposed mathematical models hold for any flavor of the IEEE802.11 family.
The rest of the article is organized as follows. Section II reviews the functionalities of the contention window procedure at MAC layer. Section III extends the Markov model initially proposed by Bianchi, presenting modifications that account for transmission errors and capture effects over Rayleigh fading channels employing the 4-way handshaking technique. This section also provides an expression for the saturation throughput of the link. In section IV we briefly extend the analysis of the two dimensional Markov chain for the basic 2-way handshaking mechanism. In section V we present simulation results where typical MAC layer parameters for IEEE802.11b are used to obtain throughput values as a function of various system level parameters, capture probability, and SNR. Section VI is devoted to conclusions.
## II Development of the Markov Model
In this section, we present the rationales at the basis of the proposed bi-dimensional Markov model useful for evaluating the throughput of the DCF under the assumptions of finite number of operating terminals, channel error conditions, capture effects in Rayleigh fading transmission scenario, and packet transmission based on the four-way handshaking access mechanism. For conciseness, we will limit our presentation to the ideas needed for developing the proposed model. We invite the interested readers to refer to \[[1](#bib.bib1), [2](#bib.bib2)\] for many details on the operating functionalities of the DCF.
###
II-A
Markovian Model Characterizing the MAC Layer: Perfect Transmission Channel
In \[[2](#bib.bib2)\], a discrete-time bi-dimensional Markov process $`({s\hspace{0pt}{(t)}},{b\hspace{0pt}{(t)}})`$ is presented for the computation of the throughput of a WLAN using the IEEE 802.11 DCF under ideal channel conditions.
It is based on two key assertions:
1. 1\.
the probability $`\tau`$ that a station will attempt transmission in a generic time slot is constant across all time slots;
2. 2\.
the probability $`P_{c\hspace{0pt}o\hspace{0pt}l}`$ that any transmission experiences a collision is constant and independent of the number of collisions already suffered.
The random process $`b\hspace{0pt}{(t)}`$ is used to represent the backoff counter of each station. Backoff counter is decremented at the start of every idle backoff slot and when it reaches zero, the station transmits and a new value for $`b\hspace{0pt}{(t)}`$ is set. The value of $`b\hspace{0pt}{(t)}`$ after each transmission depends on the size of the contention window from which it is drawn. A second process $`s\hspace{0pt}{(t)}`$ is defined, representing the size of the contention window $`({{W_{i} = {2^{i}\hspace{0pt}W}},{i = {s\hspace{0pt}{(t)}}}})`$ from which $`b\hspace{0pt}{(t)}`$ is drawn.
We recall that a backoff time counter is initialized depending on the number of failed transmissions for the transmitted packet. It is chosen from the range $`\lbrack 0,{W_{i} - 1}\rbrack`$ following a uniform Probability Mass Function (PMF), where $`W_{i}`$ is the contention window size at the backoff stage $`i`$. At the first transmission attempt (i.e., for $`i = 0`$), the contention window size is set equal to a minimum value $`W_{0} = W`$, and the process $`s\hspace{0pt}{(t)}`$ takes on the value $`{s\hspace{0pt}{(t)}} = i = 0`$.
The backoff stage $`i`$ is incremented in unitary steps after each unsuccessful transmission up to the maximum value $`m`$, while the contention window is doubled at each stage up to the maximum value $`{C\hspace{0pt}W_{m\hspace{0pt}a\hspace{0pt}x}} = {2^{m}\hspace{0pt}W}`$.
The backoff time counter is decremented as long as the channel is sensed idle and stopped when a transmission is detected. The station transmits when the backoff time counter reaches zero.
## III Markovian Model Characterizing the MAC Layer under Real Transmission Channel and Capture Effects Using the 4-Way Handshaking Mechanism
The main aim of this section is to propose an effective modification of the bi-dimensional Markov process proposed in \[[2](#bib.bib2)\] in order to account for channel error conditions and capture effects over Rayleigh fading channel under the hypothesis of employing the four-way handshaking access mechanism. It is useful to briefly recall the four-way handshaking mechanism for highlighting the hypotheses at the basis of the proposed Markov model.
A station that wants to transmit a packet, waits until the channel is sensed idle for a Distributed InterFrame Space (DIFS), follows the backoff rules and then preliminarily transmits a short Request To Send (RTS) frame. The receiving station, after having received an RTS frame, responds with a Clear To Send (CTS) frame, after a Short InterFrame Space (SIFS). The transmitting station is allowed to transmit its packet only after having correctly received the CTS frame. Both RTS and CTS frames carry the information about the length of the packet to be transmitted. This information is available to any listening station, so that they can update their Network Allocation Vector (NAV) with the information about the period of time in which the channel will be busy. Therefore, every station, hidden from either the transmitting or the receiving station, by detecting just one frame among the RTS and CTS frames, can delay transmissions by an appropriate amount, thus avoiding collisions.
On the basis of this assumption, collisions can only occur with probability $`P_{c\hspace{0pt}o\hspace{0pt}l}`$ on RTS packets, while transmission errors, due to the channel, can occur with probability $`P_{e}`$ only on the data frames. Therefore, if a station is free to send a data frame (that means that the transmitting station received the CTS frame), it can start the data frame transmission. This frame is protected from collision, but not from transmission errors. In this work, we assume that collision and transmission error events are statistically independents.
Furthermore, in mobile radio environment, it may happen that the channel is captured by a station whose power level is stronger than the other stations transmitting at the same time. This may be due to relative distances and/or channel conditions for each user and may happen whether or not the terminals exercise power control. As a matter of fact, capture effect reduces the collision probability on the channel, since the stations whose power level at the receiver is very low due to path attenuation, shadowing and fading, are considered as interferers from the AP, thus raising the noise floor.
To account for these phenomena, our Markov chain model depicted in Fig. [1](#S6.F1 "Figure 1 ‣ Saturation Throughput Analysis of IEEE 802.11 in Presence of Non Ideal Transmission Channel and Capture Effects") includes a Transmit (TX) state, identified by $`{{{(i,{- 1})},{\forall i}} = 0},{\ldots,m}`$, which is not considered in the original work of Bianchi in \[[2](#bib.bib2)\]. To reach the transmit state, there should be no collision or a collision in which the desired station captures the channel (hence, in effect no collision is detected even though there is one).
To simplify the analysis, we make the assumption that the impact of the channel induced errors on the RTS, CTS and Acknowledgment (ACK) packets are negligible because of their short length. This is justified on the basis of the assumption that the bit errors inflicting the transmitted data are independent of each other. Hence, the packet or frame error rate, identified respectively with the acronyms PER or FER, is a function of the packet length, with shorter packets having exponentially smaller probability of error compared to longer packets.
Notice that FER in general also depends on the forward error correcting code employed at the physical layer \[[21](#bib.bib21), [22](#bib.bib22)\]. However, for convolutional coding, as optionally foreseen in the IEEE 802.11b standard \[[1](#bib.bib1)\], bit error rate (BER) is independent from the code block size, and, as a consequence, FER assumes increasing values for larger and larger data packet sizes. It is known that the BER for convolutional codes depends on the distance spectrum of the code \[[23](#bib.bib23)\]-\[[25](#bib.bib25)\] which is independent of the block size.
We note that with sufficient interleaving we can always ensure that the errors inflicting individual bits in a data packet are independent of each other. Since the length of the ACK, CTS and RTS frames are typically less than 20 bytes, while the information baring packets are typically several hundreds of bytes long, the assumption is clearly justified. Including the possibility that the ACK, RTS and CTS frames can themselves be erroneous unduly complicates an already complex Markov chain with no clear benefits.
Let us discuss the Markov model shown in Fig. [1](#S6.F1 "Figure 1 ‣ Saturation Throughput Analysis of IEEE 802.11 in Presence of Non Ideal Transmission Channel and Capture Effects"). Similar to the model in \[[2](#bib.bib2)\], $`({m + 1})`$ different backoff stages are considered (this includes the zero-th stage). The maximum contention windows (CW) sizes is then $`{C\hspace{0pt}W\hspace{0pt}m\hspace{0pt}a\hspace{0pt}x} = {2^{m}\hspace{0pt}W}`$, and the relation $`W_{i} = {2^{i}\hspace{0pt}W}`$ is used to define the $`i^{t\hspace{0pt}h}`$ contention window size. An RTS can be transmitted only in the $`(i,0)`$ states, $`{\forall i} = {0,\ldots,m}`$, while data can only be transmitted in the $`(i,{- 1})`$ states, $`{\forall i} = {0,\ldots,m}`$ upon a successful reception of the CTS. If collision occurs, the backoff stage is incremented, so that the new state can be $`({i + 1},k)`$ with probability $`P_{c\hspace{0pt}o\hspace{0pt}l}/W_{i + 1}`$, since a uniform distribution between the states in the same backoff stage is considered. We consider capture as a subset of the event of a collision. In other words, the capture event can happen in the presence of collision, by allowing the transmitting station with the higher received power level at the access point to capture the channel. In this case, there is no collision and the Markov model transits into one of the transmitting states $`(i,{- 1})`$ depending on the current contention stage. If no collision occurs, a data frame can be transmitted, and the transmitting station enters state $`(i,{- 1})`$, based on the backoff stage it belongs to. From state $`(i,{- 1})`$, if the transmission is successful, the transmitting station re-enters the initial backoff stage, identified by $`i = 0`$. Otherwise, if errors occur during transmission, the ACK packet is not sent, an ACK-timeout occurs, and the backoff stage is changed to $`({i + 1},k)`$ with probability $`P_{e}/W_{i + 1}`$, where $`P_{e}`$ is the packet error probability for the employed physical layer.
The Markov Process of Fig. [1](#S6.F1 "Figure 1 ‣ Saturation Throughput Analysis of IEEE 802.11 in Presence of Non Ideal Transmission Channel and Capture Effects") is governed by the following transition probabilities
<sup>1</sup>
<sup>1</sup>1$`P_{i,{k|{j,n}}}`$ is short for $`P\hspace{0pt}\left. \{{{{s\hspace{0pt}{({t + 1})}} = i},{{b\hspace{0pt}{({t + 1})}} = k}} \middle| {{{s\hspace{0pt}{(t)}} = j},{{b\hspace{0pt}{(t)}} = n}}\} \right.`$.
:
<table>
<tbody>
<tr>
<td></td>
<td>Pi,k|i,k+1=1,k∈[0,Wi−2],i∈[0,m]Pi,−1|i,0=1−Pcol,k=0,i∈[0,m]P0,k|i,−1=(1−Pe)/W0,k∈[0,W0−1],i∈[0,m]Pi,k|i−1,0=Pcol/Wi,k∈[0,Wi−1],i∈[1,m]Pi,k|i−1,−1=Pe/Wi,k∈[0,Wi−1],i∈[1,m]Pm,k|m,0=Pcol/Wm,k∈[0,Wm−1]Pm,k|m,−1=Pe/Wm,k∈[0,Wm−1]subscript𝑃𝑖conditional𝑘𝑖𝑘1absent1formulae-sequence𝑘0subscript𝑊𝑖2𝑖0𝑚subscript𝑃𝑖conditional1𝑖0absent1subscript𝑃𝑐𝑜𝑙formulae-sequence𝑘0𝑖0𝑚subscript𝑃0conditional𝑘𝑖1absent1subscript𝑃𝑒subscript𝑊0formulae-sequence𝑘0subscript𝑊01𝑖0𝑚subscript𝑃𝑖conditional𝑘𝑖10absentsubscript𝑃𝑐𝑜𝑙subscript𝑊𝑖formulae-sequence𝑘0subscript𝑊𝑖1𝑖1𝑚subscript𝑃𝑖conditional𝑘𝑖11absentsubscript𝑃𝑒subscript𝑊𝑖formulae-sequence𝑘0subscript𝑊𝑖1𝑖1𝑚subscript𝑃𝑚conditional𝑘𝑚0absentsubscript𝑃𝑐𝑜𝑙subscript𝑊𝑚𝑘0subscript𝑊𝑚1subscript𝑃𝑚conditional𝑘𝑚1absentsubscript𝑃𝑒subscript𝑊𝑚𝑘0subscript𝑊𝑚1\small\begin{array}[]{lll}P_{i,k|i,k+1}&=1,&~{}k\in[0,W_{i}-2],~{}i\in[0,m]\\ P_{i,-1|i,0}&=1-P_{col},&~{}k=0,~{}i\in[0,m]\\ P_{0,k|i,-1}&=(1-P_{e})/W_{0},&~{}k\in[0,W_{0}-1],~{}i\in[0,m]\\ P_{i,k|i-1,0}&=P_{col}/W_{i},&~{}k\in[0,W_{i}-1],~{}i\in[1,m]\\ P_{i,k|i-1,-1}&=P_{e}/W_{i},&~{}k\in[0,W_{i}-1],~{}i\in[1,m]\\ P_{m,k|m,0}&=P_{col}/W_{m},&~{}k\in[0,W_{m}-1]\\ P_{m,k|m,-1}&=P_{e}/W_{m},&~{}k\in[0,W_{m}-1]\end{array}</td>
<td></td>
<td>(1)</td>
</tr>
</tbody>
</table>
The first equation in ([1](#S3.E1 "In III Markovian Model Characterizing the MAC Layer under Real Transmission Channel and Capture Effects Using the 4-Way Handshaking Mechanism ‣ Saturation Throughput Analysis of IEEE 802.11 in Presence of Non Ideal Transmission Channel and Capture Effects")) states that, at the beginning of each slot time, the backoff time is decremented. The second equation accounts for the fact that after a successful RTS transmission and CTS reception, the data transmission state is $`({{{s\hspace{0pt}{(t)}} = i},{{b\hspace{0pt}{(t)}} = {- 1}}})`$ with probability $`1 - P_{c\hspace{0pt}o\hspace{0pt}l}`$. The third equation accounts for the fact that after a successful transmission, a new packet transmission starts with backoff stage 0. The other equations deal with unsuccessful transmissions. Capture effects are accounted for in the definition of the collision probability. The fourth and sixth equations deal with RTS frame collisions. In this case, a new backoff stage is scheduled. Finally, the fifth and seventh equations model failed transmissions due to channel induced errors. In this situation, a new backoff stage is scheduled as well.
###
III-A
Markovian Process Analysis and Throughput Computation
Next line of pursuit consists in finding a solution of the stationary distribution:
<table>
<tbody>
<tr>
<td></td>
<td>bi,k=limt→∞P[s(t)=i,b(t)=k],∀k∈[−1,Wi−1],∀i∈[0,m]formulae-sequencesubscript𝑏𝑖𝑘subscript→𝑡𝑃delimited-[]formulae-sequence𝑠𝑡𝑖𝑏𝑡𝑘formulae-sequencefor-all𝑘1subscript𝑊𝑖1for-all𝑖0𝑚b_{i,k}=\lim_{t\rightarrow\infty}P[s(t)=i,b(t)=k],~{}\forall k\in[-1,W_{i}-1],~{}\forall i\in[0,m]</td>
<td></td>
</tr>
</tbody>
</table>
that is the probability of a station occupying a given state at any discrete time.
For the sake of simplifying the evaluation of the normalization condition of the bidimensional Markov chain, let us express all the probabilities $`b_{i,k}`$ as a function of $`b_{0,0}`$. To this end, note the following relations:
<table>
<tbody>
<tr>
<td></td>
<td>bi,−1=bi,0(1−Pcol),∀i∈[0,m]bi,0=bi−1,−1Pe+bi−1,0Pcol,∀i∈[1,m)formulae-sequencesubscript𝑏𝑖1subscript𝑏𝑖01subscript𝑃𝑐𝑜𝑙for-all𝑖0𝑚missing-subexpressionformulae-sequencesubscript𝑏𝑖0subscript𝑏𝑖11subscript𝑃𝑒subscript𝑏𝑖10subscript𝑃𝑐𝑜𝑙for-all𝑖1𝑚missing-subexpression\begin{array}[]{ll}b_{i,-1}=b_{i,0}\left(1-P_{col}\right),~{}\forall i\in[0,m]&\\ b_{i,0}=b_{i-1,-1}P_{e}+b_{i-1,0}P_{col},~{}\forall i\in[1,m)&\\ \end{array}</td>
<td></td>
<td>(2)</td>
</tr>
</tbody>
</table>
while for $`i = m`$ we have:
<table>
<tbody>
<tr>
<td></td>
<td>bm,0=bm−1,−1Pe+bm−1,0Pcol++bm,−1Pe+bm,0Pcolsubscript𝑏𝑚0subscript𝑏𝑚11subscript𝑃𝑒limit-fromsubscript𝑏𝑚10subscript𝑃𝑐𝑜𝑙missing-subexpressionsubscript𝑏𝑚1subscript𝑃𝑒subscript𝑏𝑚0subscript𝑃𝑐𝑜𝑙missing-subexpression\begin{array}[]{ll}b_{m,0}=b_{m-1,-1}P_{e}+b_{m-1,0}P_{col}+&\\ +b_{m,-1}P_{e}+b_{m,0}P_{col}&\end{array}</td>
<td></td>
</tr>
</tbody>
</table>
from which it is possible to obtain:
<table>
<tbody>
<tr>
<td></td>
<td>bm,0(1−Pcol)=bm−1,−1Pe+bm,−1Pe++bm−1,0Pcolsubscript𝑏𝑚01subscript𝑃𝑐𝑜𝑙subscript𝑏𝑚11subscript𝑃𝑒limit-fromsubscript𝑏𝑚1subscript𝑃𝑒missing-subexpressionsubscript𝑏𝑚10subscript𝑃𝑐𝑜𝑙missing-subexpression\begin{array}[]{ll}b_{m,0}\left(1-P_{col}\right)=b_{m-1,-1}P_{e}+b_{m,-1}P_{e}+&\\ +b_{m-1,0}P_{col}&\end{array}</td>
<td></td>
</tr>
</tbody>
</table>
These relations will be used later for solving the normalization condition applied to the bi-dimensional Markov model shown in Fig. [1](#S6.F1 "Figure 1 ‣ Saturation Throughput Analysis of IEEE 802.11 in Presence of Non Ideal Transmission Channel and Capture Effects").
Upon defining an equivalent probability of failed transmission, $`P_{e\hspace{0pt}q}`$
<sup>2</sup>
<sup>2</sup>2For simplicity, we assume that at each transmission attempt any station will encounter a constant and independent probability of failed transmission, $`P_{e\hspace{0pt}q}`$, independently from the number of retransmissions already suffered from each station.
, that takes into account the need for a new contention due to either RTS collision ($`P_{c\hspace{0pt}o\hspace{0pt}l}`$) or channel errors affecting the data frame ($`P_{e}`$), i.e.,
<table>
<tbody>
<tr>
<td></td>
<td>Peq=Pcol+Pe−Pe⋅Pcolsubscript𝑃𝑒𝑞subscript𝑃𝑐𝑜𝑙subscript𝑃𝑒⋅subscript𝑃𝑒subscript𝑃𝑐𝑜𝑙P_{eq}=P_{col}+P_{e}-P_{e}\cdot P_{col}</td>
<td></td>
<td>(3)</td>
</tr>
</tbody>
</table>
after some simple algebra involving the previous relations, it is possible to obtain the following relations:
<table>
<tbody>
<tr>
<td></td>
<td>bi,0=Peqbi−1,0,∀i∈[1,m)bm,0=Peqbm−1,0+Peqbm,0,i=msubscript𝑏𝑖0subscript𝑃𝑒𝑞subscript𝑏𝑖10for-all𝑖1𝑚subscript𝑏𝑚0subscript𝑃𝑒𝑞subscript𝑏𝑚10subscript𝑃𝑒𝑞subscript𝑏𝑚0𝑖𝑚\begin{array}[]{ll}b_{i,0}=P_{eq}b_{i-1,0},&~{}\forall i\in[1,m)\\ b_{m,0}=P_{eq}b_{m-1,0}+P_{eq}b_{m,0},&~{}i=m\end{array}</td>
<td></td>
<td>(4)</td>
</tr>
</tbody>
</table>
Proceeding backward, the following results follow by inspection:
<table>
<tbody>
<tr>
<td></td>
<td>bi,0=Peqib0,0,∀i∈[1,m)bm,0=Peqm1−Peqb0,0,i=msubscript𝑏𝑖0superscriptsubscript𝑃𝑒𝑞𝑖subscript𝑏00for-all𝑖1𝑚subscript𝑏𝑚0superscriptsubscript𝑃𝑒𝑞𝑚1subscript𝑃𝑒𝑞subscript𝑏00𝑖𝑚\begin{array}[]{ll}b_{i,0}=P_{eq}^{i}b_{0,0},&~{}\forall i\in[1,m)\\ b_{m,0}=\frac{P_{eq}^{m}}{1-P_{eq}}b_{0,0},&~{}i=m\end{array}</td>
<td></td>
<td>(5)</td>
</tr>
</tbody>
</table>
The other stationary probabilities for any $`k \in {\lbrack 1,{W_{i} - 1}\rbrack}`$ follow by resorting to the state transition diagram:
<table>
<tbody>
<tr>
<td></td>
<td>bi,k=Wi−kWi{(1−Pe)∑j=0mbj,−1,i=0Pcolbi−1,0+Pebi−1,−1,i∈[1,m)Pcol(bm−1,0+bm,0)++Pe(bm−1,−1+bm,−1),i=msubscript𝑏𝑖𝑘subscript𝑊𝑖𝑘subscript𝑊𝑖cases1subscript𝑃𝑒superscriptsubscript𝑗0𝑚subscript𝑏𝑗1𝑖0subscript𝑃𝑐𝑜𝑙subscript𝑏𝑖10subscript𝑃𝑒subscript𝑏𝑖11𝑖1𝑚limit-fromsubscript𝑃𝑐𝑜𝑙subscript𝑏𝑚10subscript𝑏𝑚0missing-subexpressionsubscript𝑃𝑒subscript𝑏𝑚11subscript𝑏𝑚1𝑖𝑚b_{i,k}=\frac{W_{i}-k}{W_{i}}\left\{\begin{array}[]{ll}(1-P_{e})\sum_{j=0}^{m}b_{j,-1},&~{}i=0\\ P_{col}b_{i-1,0}+P_{e}b_{i-1,-1},&~{}i\in[1,m)\\ P_{col}(b_{m-1,0}+b_{m,0})+\\ +P_{e}(b_{m-1,-1}+b_{m,-1}),&~{}i=m\end{array}\right.</td>
<td></td>
<td>(6)</td>
</tr>
</tbody>
</table>
Employing the normalization condition, after some mathematical manipulations, it is possible to obtain:
<table>
<tbody>
<tr>
<td></td>
<td>1=∑i=0m∑k=−1wi−1bi,k=b0,02(1−Peq)W(1−(2Peq)m−1)Peq+Px(1−Pcol)((1−Pe)(W+1)+2)+Px(1−Peqm−1)Peq+(2mW+1)PxPeqm2Px(1−Peq)1superscriptsubscript𝑖0𝑚superscriptsubscript𝑘1subscript𝑤𝑖1subscript𝑏𝑖𝑘subscript𝑏0021subscript𝑃𝑒𝑞𝑊1superscript2subscript𝑃𝑒𝑞𝑚1subscript𝑃𝑒𝑞subscript𝑃𝑥1subscript𝑃𝑐𝑜𝑙1subscript𝑃𝑒𝑊12subscript𝑃𝑥1superscriptsubscript𝑃𝑒𝑞𝑚1subscript𝑃𝑒𝑞superscript2𝑚𝑊1subscript𝑃𝑥superscriptsubscript𝑃𝑒𝑞𝑚2subscript𝑃𝑥1subscript𝑃𝑒𝑞\small\begin{array}[]{l}1=\sum_{i=0}^{m}\sum_{k=-1}^{w_{i}-1}b_{i,k}=b_{0,0}\frac{2(1-P_{eq})W(1-(2P_{eq})^{m-1})P_{eq}+P_{x}(1-P_{col})\left((1-P_{e})(W+1)+2\right)+P_{x}(1-P_{eq}^{m-1})P_{eq}+(2^{m}W+1)P_{x}P_{eq}^{m}}{2P_{x}(1-P_{eq})}\end{array}</td>
<td></td>
<td>(7)</td>
</tr>
</tbody>
</table>
whereby, $`P_{x} = {1 - {2\hspace{0pt}P_{e\hspace{0pt}q}}}`$. Normalization condition yields the following equation for computation of $`b_{0,0}`$:
<table>
<tbody>
<tr>
<td></td>
<td>b0,0=2Px(1−Peq)2(1−Peq)W(1−(2Peq)m−1)Peq+Px(1−Pcol)((1−Pe)(W+1)+2)+Px(1−Peqm−1)Peq+(2mW+1)PxPeqmsubscript𝑏002subscript𝑃𝑥1subscript𝑃𝑒𝑞21subscript𝑃𝑒𝑞𝑊1superscript2subscript𝑃𝑒𝑞𝑚1subscript𝑃𝑒𝑞subscript𝑃𝑥1subscript𝑃𝑐𝑜𝑙1subscript𝑃𝑒𝑊12subscript𝑃𝑥1superscriptsubscript𝑃𝑒𝑞𝑚1subscript𝑃𝑒𝑞superscript2𝑚𝑊1subscript𝑃𝑥superscriptsubscript𝑃𝑒𝑞𝑚missing-subexpression\small\begin{array}[]{ll}b_{0,0}=\frac{2P_{x}(1-P_{eq})}{2(1-P_{eq})W(1-(2P_{eq})^{m-1})P_{eq}+P_{x}(1-P_{col})\left((1-P_{e})(W+1)+2\right)+P_{x}(1-P_{eq}^{m-1})P_{eq}+(2^{m}W+1)P_{x}P_{eq}^{m}}\end{array}</td>
<td></td>
<td>(8)</td>
</tr>
</tbody>
</table>
This result is then used to compute $`\tau`$, the probability that a station starts a transmission in a randomly chosen slot time. In fact, taking into account that an RTS transmission occurs when the backoff counter reaches zero, we have:
<table>
<tbody>
<tr>
<td></td>
<td>τ=∑i=0mbi,0=b0,01−Peq=2Px2(1−Peq)W(1−(2Peq)m−1)Peq+Px(1−Pcol)((1−Pe)(W+1)+2)+Px(1−Peqm−1)Peq+(2mW+1)PxPeqm𝜏superscriptsubscript𝑖0𝑚subscript𝑏𝑖0subscript𝑏001subscript𝑃𝑒𝑞2subscript𝑃𝑥21subscript𝑃𝑒𝑞𝑊1superscript2subscript𝑃𝑒𝑞𝑚1subscript𝑃𝑒𝑞subscript𝑃𝑥1subscript𝑃𝑐𝑜𝑙1subscript𝑃𝑒𝑊12subscript𝑃𝑥1superscriptsubscript𝑃𝑒𝑞𝑚1subscript𝑃𝑒𝑞superscript2𝑚𝑊1subscript𝑃𝑥superscriptsubscript𝑃𝑒𝑞𝑚missing-subexpression\begin{array}[]{ll}\tau=\sum_{i=0}^{m}b_{i,0}=\frac{b_{0,0}}{1-P_{eq}}=\frac{2P_{x}}{2(1-P_{eq})W(1-(2P_{eq})^{m-1})P_{eq}+P_{x}(1-P_{col})\left((1-P_{e})(W+1)+2\right)+P_{x}(1-P_{eq}^{m-1})P_{eq}+(2^{m}W+1)P_{x}P_{eq}^{m}}\end{array}</td>
<td></td>
<td>(9)</td>
</tr>
</tbody>
</table>
The collision probability, needed to compute $`\tau`$, can be found considering that using a 4-way hand-shaking an RTS frame from a transmitting station encounters a collision if in a time slot, at least one of the remaining $`({N - 1})`$ stations transmit simultaneously an RTS frame, and there is no capture. In our model, we assume that capture is a subset of the collision events. This is indeed justified by the fact that there is no capture without collision, and that capture occurs only because of collisions between a certain number of transmitting stations attempting to transmit simultaneously on the channel.
<table>
<tbody>
<tr>
<td></td>
<td>Pcol=1−(1−τ)N−1−Pcapsubscript𝑃𝑐𝑜𝑙1superscript1𝜏𝑁1subscript𝑃𝑐𝑎𝑝P_{col}=1-(1-\tau)^{N-1}-P_{cap}</td>
<td></td>
<td>(10)</td>
</tr>
</tbody>
</table>
As far as capture effects are concerned, we resort to the mathematical formulation proposed in \[[4](#bib.bib4), [5](#bib.bib5)\]. In particular, under the hypothesis of power-controlled stations in infrastructure mode, the capture probability conditioned on $`i`$ interfering frames can be defined as follows:
<table>
<tbody>
<tr>
<td></td>
<td>Pcp(γ>zog(Sf)|i)=1[1+z0g(Sf)]isubscript𝑃𝑐𝑝𝛾conditionalsubscript𝑧𝑜𝑔subscript𝑆𝑓𝑖1superscriptdelimited-[]1subscript𝑧0𝑔subscript𝑆𝑓𝑖P_{cp}\left(\gamma>z_{o}g(S_{f})|i\right)=\frac{1}{{\left[1+z_{0}g(S_{f})\right]}^{i}}</td>
<td></td>
<td>(11)</td>
</tr>
</tbody>
</table>
whereby $`\gamma`$, defined as $`P_{u}/{\sum_{k = 1}^{i}P_{k}}`$, is the ratio of the power $`P_{u}`$ of the useful signal and the sum of the powers of the $`i`$ interfering channel contenders transmitting simultaneously $`i`$ frames, $`g\hspace{0pt}{(S_{f})}`$ is the inverse of the processing gain of the correlation receiver, and $`z_{0}`$ is the capture ratio, i.e., the value of the signal-to-interference power ratio identifying the capture threshold at the receiver. Notice that ([11](#S3.E11 "In III-A Markovian Process Analysis and Throughput Computation ‣ III Markovian Model Characterizing the MAC Layer under Real Transmission Channel and Capture Effects Using the 4-Way Handshaking Mechanism ‣ Saturation Throughput Analysis of IEEE 802.11 in Presence of Non Ideal Transmission Channel and Capture Effects")) signifies the fact that capture probability corresponds to the probability that the power ratio $`\gamma`$ is above the capture threshold $`z_{0}\hspace{0pt}g\hspace{0pt}{(S_{f})}`$ which considers the inverse of the processing gain $`g\hspace{0pt}{(S_{f})}`$. For Direct Sequence Spread Spectrum (DSSS) using a 11-chip spreading factor ($`S_{f} = 11`$), we have $`{g\hspace{0pt}{(S_{f})}} = \frac{2}{3\hspace{0pt}S_{f}}`$.
Upon defining the probability of generating exactly $`i + 1`$ interfering frames over $`N`$ contending stations in a generic slot time:
<table>
<tbody>
<tr>
<td></td>
<td>(Ni+1)τi+1(1−τ)N−i−1binomial𝑁𝑖1superscript𝜏𝑖1superscript1𝜏𝑁𝑖1{N\choose i+1}\tau^{i+1}(1-\tau)^{N-i-1}</td>
<td></td>
</tr>
</tbody>
</table>
the frame capture probability $`P_{c\hspace{0pt}a\hspace{0pt}p}`$ can be obtained as follows:
<table>
<tbody>
<tr>
<td></td>
<td>Pcap=∑i=1N−1(Ni+1)τi+1(1−τ)N−i−1Pcp(γ>zog(Sf)|i)subscript𝑃𝑐𝑎𝑝superscriptsubscript𝑖1𝑁1binomial𝑁𝑖1superscript𝜏𝑖1superscript1𝜏𝑁𝑖1subscript𝑃𝑐𝑝𝛾conditionalsubscript𝑧𝑜𝑔subscript𝑆𝑓𝑖P_{cap}=\sum_{i=1}^{N-1}{N\choose i+1}\tau^{i+1}(1-\tau)^{N-i-1}P_{cp}\left(\gamma>z_{o}g(S_{f})|i\right)</td>
<td></td>
<td>(12)</td>
</tr>
</tbody>
</table>
Putting together equations ([3](#S3.E3 "In III-A Markovian Process Analysis and Throughput Computation ‣ III Markovian Model Characterizing the MAC Layer under Real Transmission Channel and Capture Effects Using the 4-Way Handshaking Mechanism ‣ Saturation Throughput Analysis of IEEE 802.11 in Presence of Non Ideal Transmission Channel and Capture Effects")), ([9](#S3.E9 "In III-A Markovian Process Analysis and Throughput Computation ‣ III Markovian Model Characterizing the MAC Layer under Real Transmission Channel and Capture Effects Using the 4-Way Handshaking Mechanism ‣ Saturation Throughput Analysis of IEEE 802.11 in Presence of Non Ideal Transmission Channel and Capture Effects")), ([10](#S3.E10 "In III-A Markovian Process Analysis and Throughput Computation ‣ III Markovian Model Characterizing the MAC Layer under Real Transmission Channel and Capture Effects Using the 4-Way Handshaking Mechanism ‣ Saturation Throughput Analysis of IEEE 802.11 in Presence of Non Ideal Transmission Channel and Capture Effects")) and ([12](#S3.E12 "In III-A Markovian Process Analysis and Throughput Computation ‣ III Markovian Model Characterizing the MAC Layer under Real Transmission Channel and Capture Effects Using the 4-Way Handshaking Mechanism ‣ Saturation Throughput Analysis of IEEE 802.11 in Presence of Non Ideal Transmission Channel and Capture Effects")), the following nonlinear system can be defined and solved numerically, obtaining the values of $`\tau`$ (defined in ([9](#S3.E9 "In III-A Markovian Process Analysis and Throughput Computation ‣ III Markovian Model Characterizing the MAC Layer under Real Transmission Channel and Capture Effects Using the 4-Way Handshaking Mechanism ‣ Saturation Throughput Analysis of IEEE 802.11 in Presence of Non Ideal Transmission Channel and Capture Effects"))), $`P_{c\hspace{0pt}o\hspace{0pt}l}`$, $`P_{e\hspace{0pt}q}`$, and $`P_{c\hspace{0pt}a\hspace{0pt}p}`$:
<table>
<tbody>
<tr>
<td></td>
<td>{τPcol=1−(1−τ)N−1−PcapPeq=Pcol+Pe−Pe⋅PcolPcap=∑i=1N−1(Ni+1)τi+1(1−τ)N−i−11(1+z0g(Sf))icases𝜏missing-subexpressionsubscript𝑃𝑐𝑜𝑙absent1superscript1𝜏𝑁1subscript𝑃𝑐𝑎𝑝subscript𝑃𝑒𝑞absentsubscript𝑃𝑐𝑜𝑙subscript𝑃𝑒⋅subscript𝑃𝑒subscript𝑃𝑐𝑜𝑙subscript𝑃𝑐𝑎𝑝absentsuperscriptsubscript𝑖1𝑁1binomial𝑁𝑖1superscript𝜏𝑖1superscript1𝜏𝑁𝑖11superscript1subscript𝑧0𝑔subscript𝑆𝑓𝑖\left\{\begin{array}[]{ll}\tau&\\ P_{col}&=1-(1-\tau)^{N-1}-P_{cap}\\ P_{eq}&=P_{col}+P_{e}-P_{e}\cdot P_{col}\\ P_{cap}&=\sum_{i=1}^{N-1}{N\choose i+1}\tau^{i+1}(1-\tau)^{N-i-1}\frac{1}{{(1+z_{0}g(S_{f}))}^{i}}\end{array}\right.</td>
<td></td>
<td>(13)</td>
</tr>
</tbody>
</table>
The final step in the analysis is the computation of the normalized system throughput, defined as the fraction of time the channel is used to successfully transmit payload bits:
<table>
<tbody>
<tr>
<td></td>
<td>S=Pt⋅Ps⋅(1−Pe)E[PL](1−Pt)σ+Pt(1−Ps)Tc+PtPs(1−Pe)Ts+PtPsPeTe𝑆⋅subscript𝑃𝑡subscript𝑃𝑠1subscript𝑃𝑒𝐸delimited-[]𝑃𝐿1subscript𝑃𝑡𝜎subscript𝑃𝑡1subscript𝑃𝑠subscript𝑇𝑐subscript𝑃𝑡subscript𝑃𝑠1subscript𝑃𝑒subscript𝑇𝑠subscript𝑃𝑡subscript𝑃𝑠subscript𝑃𝑒subscript𝑇𝑒\small S=\frac{P_{t}\cdot P_{s}\cdot(1-P_{e})E[PL]}{(1-P_{t})\sigma+P_{t}(1-P_{s})T_{c}+P_{t}P_{s}(1-P_{e})T_{s}+P_{t}P_{s}P_{e}T_{e}}</td>
<td></td>
<td>(14)</td>
</tr>
</tbody>
</table>
where
- •
$`P_{t}`$ is the probability that there is at least one transmission in the considered slot time, with $`N`$ stations contending for the channel, each transmitting with probability $`\tau`$:
<table>
<tbody>
<tr>
<td></td>
<td>Pt=1−(1−τ)Nsubscript𝑃𝑡1superscript1𝜏𝑁P_{t}=1-(1-\tau)^{N}</td>
<td></td>
<td>(15)</td>
</tr>
</tbody>
</table>
- •
$`P_{s}`$ is the conditional probability that an RTS transmission occurring on the channel is successful. This event corresponds to the case in which exactly one station transmits in a given slot time, or two or more stations transmit simultaneously and capture by the desired station occurs. These conditions yields the following probability:
<table>
<tbody>
<tr>
<td></td>
<td>Ps=Nτ(1−τ)N−1+PcapPtsubscript𝑃𝑠𝑁𝜏superscript1𝜏𝑁1subscript𝑃𝑐𝑎𝑝subscript𝑃𝑡P_{s}=\frac{N\tau(1-\tau)^{N-1}+P_{cap}}{P_{t}}</td>
<td></td>
<td>(16)</td>
</tr>
</tbody>
</table>
- •
$`T_{c}`$, $`T_{e}`$ and $`T_{s}`$ are the average time a channel is sensed busy due to an RTS collision, error affected data frame transmission time and successful data frame transmission times, respectively. Knowing the time durations for RTS, CTS, ACK frames, CTS and ACK timeout, DIFS, SIFS, $`\sigma`$, data packet length ($`P\hspace{0pt}L`$) and PHY and MAC headers duration ($`H`$), and propagation delay $`\tau_{p}`$, $`T_{c}`$, $`T_{s}`$, and $`T_{e}`$ can be computed as follows \[[7](#bib.bib7)\]:
<table>
<tbody>
<tr>
<td></td>
<td>Tc=RTS+CTStimeoutTe=RTS+SIFS+τp+CTS+SIFS+τp++H+PL+ACKtimeoutTs=RTS+SIFS+τp+CTS+SIFS+τp++H+PL+SIFS+τp+ACK+DIFS+τpsubscript𝑇𝑐absent𝑅𝑇𝑆𝐶𝑇subscript𝑆𝑡𝑖𝑚𝑒𝑜𝑢𝑡subscript𝑇𝑒absent𝑅𝑇𝑆𝑆𝐼𝐹𝑆subscript𝜏𝑝𝐶𝑇𝑆𝑆𝐼𝐹𝑆limit-fromsubscript𝜏𝑝missing-subexpression𝐻𝑃𝐿𝐴𝐶subscript𝐾𝑡𝑖𝑚𝑒𝑜𝑢𝑡subscript𝑇𝑠absent𝑅𝑇𝑆𝑆𝐼𝐹𝑆subscript𝜏𝑝𝐶𝑇𝑆𝑆𝐼𝐹𝑆limit-fromsubscript𝜏𝑝missing-subexpression𝐻𝑃𝐿𝑆𝐼𝐹𝑆subscript𝜏𝑝𝐴𝐶𝐾𝐷𝐼𝐹𝑆subscript𝜏𝑝\begin{array}[]{ll}T_{c}&=RTS+CTS_{timeout}\\ T_{e}&=RTS+SIFS+\tau_{p}+CTS+SIFS+\tau_{p}+\\ &+H+PL+ACK_{timeout}\\ T_{s}&=RTS+SIFS+\tau_{p}+CTS+SIFS+\tau_{p}+\\ &+H+PL+SIFS+\tau_{p}+ACK+DIFS+\tau_{p}\\ \end{array}</td>
<td></td>
<td>(17)</td>
</tr>
</tbody>
</table>
- •
$`E\hspace{0pt}{\lbrack{P\hspace{0pt}L}\rbrack}`$ is the average packet payload length.
- •
$`\sigma`$ is the duration of an empty slot time.
The setup described above is used in Section [V](#S5 "V Simulation Results and Model Validations ‣ Saturation Throughput Analysis of IEEE 802.11 in Presence of Non Ideal Transmission Channel and Capture Effects") for DCF simulation at the MAC layer.
## IV Analysis of the Markov chain for the basic 2-way handshaking mechanism in presence of channel errors and capture effects
This section briefly provides the extension of the Markov model for the contention model to the case of using the basic two-way handshaking mechanism.
Let us review the basic functionalities of the two-way handshaking technique. No RTS/CTS exchange of packets is used. After the successful reception of a data frame, the receiver sends an ACK frame to the transmitter. Only upon a correct ACK frame reception, the transmitter assumes successful delivery of the corresponding data frame. If an ACK frame is received in error or if no ACK frame is received, due possibly to an erroneous reception of the preceding data frame, the transmitter will contend again for the medium.
Based on these assumptions, using the two-way handshaking mechanism there is no evident distinction between the transmission of information and protocol data. In this respect, the Markov chain model for the backoff window size can be represented as in Bianchi’s model \[[2](#bib.bib2)\] upon exchanging $`p`$, the collision probability, with the probability $`P_{e\hspace{0pt}q}`$ defined in ([3](#S3.E3 "In III-A Markovian Process Analysis and Throughput Computation ‣ III Markovian Model Characterizing the MAC Layer under Real Transmission Channel and Capture Effects Using the 4-Way Handshaking Mechanism ‣ Saturation Throughput Analysis of IEEE 802.11 in Presence of Non Ideal Transmission Channel and Capture Effects")). Notice that transmission fails when collision occurs between packets, or when data transmission is affected by channel errors with probability $`P_{e}`$. As far as collision is concerned, capture effects reduce the probability of collision by an amount equal to $`P_{c\hspace{0pt}a\hspace{0pt}p}`$, i.e., the probability of capture. So, in this respect, rationales are the same as for the 4-way handshaking mechanism discussed above.
Upon defining $`P_{e\hspace{0pt}q}`$ as in ([3](#S3.E3 "In III-A Markovian Process Analysis and Throughput Computation ‣ III Markovian Model Characterizing the MAC Layer under Real Transmission Channel and Capture Effects Using the 4-Way Handshaking Mechanism ‣ Saturation Throughput Analysis of IEEE 802.11 in Presence of Non Ideal Transmission Channel and Capture Effects")), the stationary distribution of the Bianchi’s Markov chain can be evaluated by following the same mathematical derivations dealt with in Bianchi’s model \[[2](#bib.bib2)\] whereby $`P_{e\hspace{0pt}q}`$ is used in place of $`p`$. Finally, the throughput is evaluated through ([14](#S3.E14 "In III-A Markovian Process Analysis and Throughput Computation ‣ III Markovian Model Characterizing the MAC Layer under Real Transmission Channel and Capture Effects Using the 4-Way Handshaking Mechanism ‣ Saturation Throughput Analysis of IEEE 802.11 in Presence of Non Ideal Transmission Channel and Capture Effects")) upon solving the non-linear system of equations in ([13](#S3.E13 "In III-A Markovian Process Analysis and Throughput Computation ‣ III Markovian Model Characterizing the MAC Layer under Real Transmission Channel and Capture Effects Using the 4-Way Handshaking Mechanism ‣ Saturation Throughput Analysis of IEEE 802.11 in Presence of Non Ideal Transmission Channel and Capture Effects")).
###
IV-A
Markovian Process Analysis and Throughput Computation
The objective is finding a solution of the stationary distribution:
<table>
<tbody>
<tr>
<td></td>
<td>bi,k=limt→∞P[s(t)=i,b(t)=k],∀i∈[0,m],∀k∈[0,Wi−1],formulae-sequencesubscript𝑏𝑖𝑘subscript→𝑡𝑃delimited-[]formulae-sequence𝑠𝑡𝑖𝑏𝑡𝑘formulae-sequencefor-all𝑖0𝑚for-all𝑘0subscript𝑊𝑖1b_{i,k}=\lim_{t\rightarrow\infty}P[s(t)=i,b(t)=k],~{}\forall i\in[0,m],~{}\forall k\in[0,W_{i}-1],</td>
<td></td>
</tr>
</tbody>
</table>
that is the probability of a station occupying a given state at any discrete slot time.
First of all, note that for any $`i \in {\lbrack 1,m)}`$ the following relation holds:
<table>
<tbody>
<tr>
<td></td>
<td>bi,0=bi−1,0Peq,∀i∈[1,m)bm,0=bm−1,0Peq+bm,0Peq,i=msubscript𝑏𝑖0subscript𝑏𝑖10subscript𝑃𝑒𝑞for-all𝑖1𝑚subscript𝑏𝑚0subscript𝑏𝑚10subscript𝑃𝑒𝑞subscript𝑏𝑚0subscript𝑃𝑒𝑞𝑖𝑚\begin{array}[]{ll}b_{i,0}=b_{i-1,0}P_{eq},&~{}\forall i\in[1,m)\\ b_{m,0}=b_{m-1,0}P_{eq}+b_{m,0}P_{eq},&~{}i=m\end{array}</td>
<td></td>
<td>(18)</td>
</tr>
</tbody>
</table>
Proceeding backward, the following results follow by inspection:
<table>
<tbody>
<tr>
<td></td>
<td>bi,0=Peqi⋅b0,0,∀i∈[1,m)bm,0=Peqm1−Peq⋅b0,0,i=msubscript𝑏𝑖0⋅superscriptsubscript𝑃𝑒𝑞𝑖subscript𝑏00for-all𝑖1𝑚subscript𝑏𝑚0⋅superscriptsubscript𝑃𝑒𝑞𝑚1subscript𝑃𝑒𝑞subscript𝑏00𝑖𝑚\begin{array}[]{ll}b_{i,0}=P_{eq}^{i}\cdot b_{0,0},&~{}\forall i\in[1,m)\\ b_{m,0}=\frac{P_{eq}^{m}}{1-P_{eq}}\cdot b_{0,0},&~{}i=m\end{array}</td>
<td></td>
<td>(19)</td>
</tr>
</tbody>
</table>
The stationary probabilities for any $`k \in {\lbrack 1,{W_{i} - 1}\rbrack}`$ follow by resorting to the state transition diagram:
<table>
<tbody>
<tr>
<td></td>
<td>bi,k=Wi−kWi{(1−Peq)∑j=0mbj,0,i=0Peqbi−1,0,i∈[1,m)Peq(bm−1,0+bm,0),i=msubscript𝑏𝑖𝑘subscript𝑊𝑖𝑘subscript𝑊𝑖cases1subscript𝑃𝑒𝑞superscriptsubscript𝑗0𝑚subscript𝑏𝑗0𝑖0subscript𝑃𝑒𝑞subscript𝑏𝑖10𝑖1𝑚subscript𝑃𝑒𝑞subscript𝑏𝑚10subscript𝑏𝑚0𝑖𝑚b_{i,k}=\frac{W_{i}-k}{W_{i}}\left\{\begin{array}[]{ll}(1-P_{eq})\sum_{j=0}^{m}b_{j,0},&~{}i=0\\ P_{eq}b_{i-1,0},&~{}i\in[1,m)\\ P_{eq}(b_{m-1,0}+b_{m,0}),&~{}i=m\end{array}\right.</td>
<td></td>
<td>(20)</td>
</tr>
</tbody>
</table>
Upon utilizing the normalization condition:
<table>
<tbody>
<tr>
<td></td>
<td>1=∑i=0m∑k=0Wi−1bi,k=b0,02[W(∑i=0m−1(2Peq)i+(2Peq)m1−Peq)+11−Peq]1superscriptsubscript𝑖0𝑚superscriptsubscript𝑘0subscript𝑊𝑖1subscript𝑏𝑖𝑘subscript𝑏002delimited-[]𝑊superscriptsubscript𝑖0𝑚1superscript2subscript𝑃𝑒𝑞𝑖superscript2subscript𝑃𝑒𝑞𝑚1subscript𝑃𝑒𝑞11subscript𝑃𝑒𝑞missing-subexpression\begin{array}[]{ll}1=\sum_{i=0}^{m}\sum_{k=0}^{W_{i}-1}b_{i,k}=\frac{b_{0,0}}{2}\left[W\left(\sum_{i=0}^{m-1}(2P_{eq})^{i}+\frac{(2P_{eq})^{m}}{1-P_{eq}}\right)+\frac{1}{1-P_{eq}}\right]&\end{array}</td>
<td></td>
<td>(21)</td>
</tr>
</tbody>
</table>
the following equation for computation of $`b_{0,0}`$ results:
<table>
<tbody>
<tr>
<td></td>
<td>b0,0=2(1−Peq)(1−2Peq)(1−2Peq)(W+1)+PeqW(1−(2Peq)m)subscript𝑏0021subscript𝑃𝑒𝑞12subscript𝑃𝑒𝑞12subscript𝑃𝑒𝑞𝑊1subscript𝑃𝑒𝑞𝑊1superscript2subscript𝑃𝑒𝑞𝑚b_{0,0}=\frac{2(1-P_{eq})(1-2P_{eq})}{(1-2P_{eq})(W+1)+P_{eq}W\left(1-(2P_{eq})^{m}\right)}</td>
<td></td>
<td>(22)</td>
</tr>
</tbody>
</table>
This result is then used to compute $`\tau`$, the probability that a station starts a transmission in a randomly chosen time slot. In fact, taking into account that a data packet transmission occurs when the backoff counter reaches zero, we have:
<table>
<tbody>
<tr>
<td></td>
<td>τ=∑i=0mbi,0=b0,01−Peq==2(1−2Peq)(1−2Peq)(W+1)+PeqW(1−(2Peq)m)𝜏superscriptsubscript𝑖0𝑚subscript𝑏𝑖0subscript𝑏001subscript𝑃𝑒𝑞absentmissing-subexpressionabsent212subscript𝑃𝑒𝑞12subscript𝑃𝑒𝑞𝑊1subscript𝑃𝑒𝑞𝑊1superscript2subscript𝑃𝑒𝑞𝑚missing-subexpression\begin{array}[]{ll}\tau=\sum_{i=0}^{m}b_{i,0}=\frac{b_{0,0}}{1-P_{eq}}=&\\ =\frac{2(1-2P_{eq})}{(1-2P_{eq})(W+1)+P_{eq}W(1-(2P_{eq})^{m})}&\end{array}</td>
<td></td>
<td>(23)</td>
</tr>
</tbody>
</table>
The nonlinear system of equations needed to compute $`\tau`$, $`P_{c\hspace{0pt}o\hspace{0pt}l}`$, $`P_{c\hspace{0pt}a\hspace{0pt}p}`$, and $`P_{e\hspace{0pt}q}`$ is defined as follows:
<table>
<tbody>
<tr>
<td></td>
<td>{τ=2(1−2Peq)(1−2Peq)(W+1)+PeqW(1−(2Peq)m)Pcol=1−(1−τ)N−1−PcapPeq=Pcol+Pe−Pe⋅PcolPcap=∑i=1N−1(Ni+1)τi+1(1−τ)N−i−11(1+z0g(Sf))icases𝜏absent212subscript𝑃𝑒𝑞12subscript𝑃𝑒𝑞𝑊1subscript𝑃𝑒𝑞𝑊1superscript2subscript𝑃𝑒𝑞𝑚subscript𝑃𝑐𝑜𝑙absent1superscript1𝜏𝑁1subscript𝑃𝑐𝑎𝑝subscript𝑃𝑒𝑞absentsubscript𝑃𝑐𝑜𝑙subscript𝑃𝑒⋅subscript𝑃𝑒subscript𝑃𝑐𝑜𝑙subscript𝑃𝑐𝑎𝑝absentsuperscriptsubscript𝑖1𝑁1binomial𝑁𝑖1superscript𝜏𝑖1superscript1𝜏𝑁𝑖11superscript1subscript𝑧0𝑔subscript𝑆𝑓𝑖\left\{\begin{array}[]{ll}\tau&=\frac{2(1-2P_{eq})}{(1-2P_{eq})(W+1)+P_{eq}W(1-(2P_{eq})^{m})}\\ P_{col}&=1-(1-\tau)^{N-1}-P_{cap}\\ P_{eq}&=P_{col}+P_{e}-P_{e}\cdot P_{col}\\ P_{cap}&=\sum_{i=1}^{N-1}{N\choose i+1}\tau^{i+1}(1-\tau)^{N-i-1}\frac{1}{{(1+z_{0}g(S_{f}))}^{i}}\end{array}\right.</td>
<td></td>
<td>(24)</td>
</tr>
</tbody>
</table>
whereby $`P_{c\hspace{0pt}o\hspace{0pt}l}`$, $`P_{c\hspace{0pt}a\hspace{0pt}p}`$, and $`P_{e\hspace{0pt}q}`$ are defined as for the 4-way handshaking mechanism.
The final step in the analysis is the computation of the normalized system throughput defined in ([14](#S3.E14 "In III-A Markovian Process Analysis and Throughput Computation ‣ III Markovian Model Characterizing the MAC Layer under Real Transmission Channel and Capture Effects Using the 4-Way Handshaking Mechanism ‣ Saturation Throughput Analysis of IEEE 802.11 in Presence of Non Ideal Transmission Channel and Capture Effects")), where all the terms are defined as for the 4-way handshaking, excepts for the definitions of the following time intervals for the 2-way mechanism \[[7](#bib.bib7)\]:
<table>
<tbody>
<tr>
<td></td>
<td>Tc=H+PL+ACKtimeoutTe=H+PL+ACKtimeoutTs=H+PL+SIFS+τp+ACK+DIFS+τpsubscript𝑇𝑐absent𝐻𝑃𝐿𝐴𝐶subscript𝐾𝑡𝑖𝑚𝑒𝑜𝑢𝑡subscript𝑇𝑒absent𝐻𝑃𝐿𝐴𝐶subscript𝐾𝑡𝑖𝑚𝑒𝑜𝑢𝑡subscript𝑇𝑠absent𝐻𝑃𝐿𝑆𝐼𝐹𝑆subscript𝜏𝑝𝐴𝐶𝐾𝐷𝐼𝐹𝑆subscript𝜏𝑝\begin{array}[]{ll}T_{c}&=H+PL+ACK_{timeout}\\ T_{e}&=H+PL+ACK_{timeout}\\ T_{s}&=H+PL+SIFS+\tau_{p}+ACK+DIFS+\tau_{p}\end{array}</td>
<td></td>
<td>(25)</td>
</tr>
</tbody>
</table>
## V Simulation Results and Model Validations
This section presents some simulation results for validating the theoretical models and derivations presented in the previous sections. We have developed a C++ simulator modelling both the DCF protocol details in 802.11b and the backoff procedures of a specific number of independent transmitting stations. The simulator also takes into account all real operations of each transmitting station, including physical propagation delays, etc.
###
V-A
Simulation Setup
Let us discuss the main functionalities of the developed simulator. It considers an Infrastructure BSS (Basic Service Set) with an Access Point (AP) and a pre-specified number of mobile stations which communicates only with the AP under the hypothesis that each station has always a packet to be transmitted, i.e., saturated conditions. The MAC layer is managed by a state machine which follows the main directives specified in the standard \[[1](#bib.bib1)\], namely waiting times (DIFS, SIFS, EIFS), post-backoff, backoff, basic and RTS/CTS access mode. As far as simulation results are concerned, we have employed MAC layer parameters for IEEE802.11b as noted in Table [I](#S6.T1 "TABLE I ‣ Saturation Throughput Analysis of IEEE 802.11 in Presence of Non Ideal Transmission Channel and Capture Effects") \[[1](#bib.bib1)\].
For the sake of simulating capture effects, the $`N`$ contending stations are randomly placed in a circular area of radius $`R`$ (in the simulation results presented below we assume $`R = 50`$m), while the AP is placed at the center of the transmission area. When two or more station transmissions collide, the value of $`\gamma`$, defined as
<table>
<tbody>
<tr>
<td></td>
<td>γ=Pu/∑k=1iPk𝛾subscript𝑃𝑢superscriptsubscript𝑘1𝑖subscript𝑃𝑘\gamma=P_{u}/\sum_{k=1}^{i}P_{k}</td>
<td></td>
<td>(26)</td>
</tr>
</tbody>
</table>
is evaluated for any transmitting station given their relative distance $`r_{i}`$ from the AP. Notice that $`\gamma`$ in ([26](#S5.E26 "In V-A Simulation Setup ‣ V Simulation Results and Model Validations ‣ Saturation Throughput Analysis of IEEE 802.11 in Presence of Non Ideal Transmission Channel and Capture Effects")) is the ratio of the power $`P_{u}`$ of the useful signal and the sum of the powers of the $`i`$ interfering channel contenders transmitting simultaneously $`i`$ frames. Let $`\gamma_{j}`$ be the value of $`\gamma`$ for the $`j`$-th transmitting station among the $`i + 1`$ colliding stations. The power between a transmitter and a receiver is the local mean power denoted $`p_{o}`$, and defined as:
<table>
<tbody>
<tr>
<td></td>
<td>po=A⋅ri−npPtxsubscript𝑝𝑜⋅𝐴superscriptsubscript𝑟𝑖subscript𝑛𝑝subscript𝑃𝑡𝑥p_{o}=A\cdot r_{i}^{-n_{p}}P_{tx}</td>
<td></td>
<td>(27)</td>
</tr>
</tbody>
</table>
In ([27](#S5.E27 "In V-A Simulation Setup ‣ V Simulation Results and Model Validations ‣ Saturation Throughput Analysis of IEEE 802.11 in Presence of Non Ideal Transmission Channel and Capture Effects")), $`n_{p}`$ is the path-loss exponent
<sup>3</sup>
<sup>3</sup>3In the simulated scenarios, we used the value $`n_{p} = 3.5`$.
(which is typically greater than or equal to $`3.5`$ in indoor propagation conditions in the absence of the direct signal path \[[27](#bib.bib27)\]), $`P_{t\hspace{0pt}x}`$ is the transmitted power, and $`A \cdot r_{i}^{- n_{p}}`$ is the deterministic path-loss \[[27](#bib.bib27)\]. Both $`A`$ and $`P_{t\hspace{0pt}x}`$ are identical for all transmitted frames. When signal transmission is affected by Rayleigh fading, the instantaneous power of the signal received by the receiver placed at a mutual distance $`r_{i}`$ from the transmitter is exponentially distributed as:
<table>
<tbody>
<tr>
<td></td>
<td>f(x)=1poe−xpo,x>0formulae-sequence𝑓𝑥1subscript𝑝𝑜superscript𝑒𝑥subscript𝑝𝑜𝑥0f(x)=\frac{1}{p_{o}}e^{-\frac{x}{p_{o}}},x>0</td>
<td></td>
</tr>
</tbody>
</table>
The values of $`\gamma_{j}`$ for each colliding station are compared with the threshold $`{z_{o} \cdot g}\hspace{0pt}{(S_{f})}`$. Then, the transmitting station for which $`\gamma_{j}`$ is above the threshold captures the channel.
Throughput simulations are accomplished by averaging over $`100`$ sample scenarios, whereby any transmitting scenario considers a set of $`N`$ randomly distributed (with a uniform probability mass function) stations over a circular area of radius 50 m as specified above.
The physical layer (PHY) of the basic 802.11b standard is based on the spread spectrum technology. Two options are specified, the Frequency Hopped Spread Spectrum (FHSS) and the DSSS (Direct Sequence Spread Spectrum). The FHSS uses Frequency Shift Keying (FSK) while the DSSS uses Differential Phase Shift Keying (DPSK) or Complementary Code Keying (CCK). The 802.11b employs DSSS at various rates including one employing CCK encoding 4 and 8 bits on one CCK symbol. The four supported data rates in 802.11b are 1, 2, 5.5 and 11 Mbps.
The FER as a function of the SNR can be computed as follows:
<table>
<tbody>
<tr>
<td></td>
<td>Pe(SNR)=1−[1−Pe(PLCP,SNR)]⋅[1−Pe(DATA,SNR)]subscript𝑃𝑒𝑆𝑁𝑅1⋅delimited-[]1subscript𝑃𝑒𝑃𝐿𝐶𝑃𝑆𝑁𝑅delimited-[]1subscript𝑃𝑒𝐷𝐴𝑇𝐴𝑆𝑁𝑅\small P_{e}(SNR)=1-\left[1-P_{e}(PLCP,SNR)\right]\cdot\left[1-P_{e}(DATA,SNR)\right]</td>
<td></td>
<td>(28)</td>
</tr>
</tbody>
</table>
where,
<table>
<tbody>
<tr>
<td></td>
<td>Pe(PLCP,SNR)=1−[1−Pb(BPSK,SNR)]8×PLCP,subscript𝑃𝑒𝑃𝐿𝐶𝑃𝑆𝑁𝑅1superscriptdelimited-[]1subscript𝑃𝑏𝐵𝑃𝑆𝐾𝑆𝑁𝑅8𝑃𝐿𝐶𝑃\small P_{e}(PLCP,SNR)=1-\left[1-P_{b}(BPSK,SNR)\right]^{8\times PLCP},</td>
<td></td>
<td>(29)</td>
</tr>
</tbody>
</table>
and
<table>
<tbody>
<tr>
<td></td>
<td>Pe(DATA,SNR)=1−[1−Pb(TYPE,SNR)]8×(DATA+MAC).subscript𝑃𝑒𝐷𝐴𝑇𝐴𝑆𝑁𝑅1superscriptdelimited-[]1subscript𝑃𝑏𝑇𝑌𝑃𝐸𝑆𝑁𝑅8𝐷𝐴𝑇𝐴𝑀𝐴𝐶\small P_{e}(DATA,SNR)=1-\left[1-P_{b}(TYPE,SNR)\right]^{8\times(DATA+MAC)}.</td>
<td></td>
<td>(30)</td>
</tr>
</tbody>
</table>
$`P_{b}\hspace{0pt}{({B\hspace{0pt}P\hspace{0pt}S\hspace{0pt}K},{S\hspace{0pt}N\hspace{0pt}R})}`$ is the BER as a function of SNR for the lowest data transmit rate employing DBPSK modulation, DATA denotes the data block size in bytes, and any other constant byte size in above expression represents overhead. Note that the FER, $`P_{e}\hspace{0pt}{({S\hspace{0pt}N\hspace{0pt}R})}`$, implicitly depends on the modulation format used. Hence, for each supported rate, one curve for $`P_{e}\hspace{0pt}{({S\hspace{0pt}N\hspace{0pt}R})}`$ as a function of SNR can be generated. $`P_{b}\hspace{0pt}{({T\hspace{0pt}Y\hspace{0pt}P\hspace{0pt}E},{S\hspace{0pt}N\hspace{0pt}R})}`$ is modulation dependent whereby the parameter $`T\hspace{0pt}Y\hspace{0pt}P\hspace{0pt}E`$ can be any of the following $`{T\hspace{0pt}Y\hspace{0pt}P\hspace{0pt}E} \in {\{{D\hspace{0pt}B\hspace{0pt}P\hspace{0pt}S\hspace{0pt}K},{D\hspace{0pt}Q\hspace{0pt}P\hspace{0pt}S\hspace{0pt}K},{C\hspace{0pt}C\hspace{0pt}K\hspace{0pt}5.5},{C\hspace{0pt}C\hspace{0pt}K\hspace{0pt}11}\}}`$
<sup>4</sup>
<sup>4</sup>4The acronyms are short for Differential Binary Phase Shift Keying, Differential Quadrature Phase Shift Keying and Complementary Code Keying, respectively.
.
For DBPSK and DQPSK modulation formats, $`P_{b}\hspace{0pt}{({T\hspace{0pt}Y\hspace{0pt}P\hspace{0pt}E},{S\hspace{0pt}N\hspace{0pt}R})}`$ can be well approximated by \[[26](#bib.bib26)\]:
<table>
<tbody>
<tr>
<td></td>
<td>2max(log2M,2)∑i=1max(M4,1)1π∫0π211+γ1sin2θlog2Msin2((2i−1)πM)𝑑θ2subscript2𝑀2superscriptsubscript𝑖1𝑀411𝜋superscriptsubscript0𝜋211𝛾1superscript2𝜃subscript2𝑀superscript22𝑖1𝜋𝑀differential-d𝜃\small\frac{2}{\max(\log_{2}M,2)}\sum_{i=1}^{\max(\frac{M}{4},1)}\frac{1}{\pi}\int_{0}^{\frac{\pi}{2}}\frac{1}{1+\gamma\frac{1}{\sin^{2}\theta}\log_{2}M\sin^{2}\left(\frac{(2i-1)\pi}{M}\right)}d\theta</td>
<td></td>
<td>(31)</td>
</tr>
</tbody>
</table>
whereby, $`M`$ is the number of bits for modulated symbols, $`\gamma`$ is the signal-to-noise ratio, and $`\theta`$ is the signal direction over the Rayleigh fading channel.
In so far as the computation of the FER is concerned, it should be noted that data packet error rates of different contending stations differ. For simplicity, we assume that data packets transmitted by different stations are affected by the same FER. Of course, this is a simplifying assumption which is indeed justified by the following pragmatic considerations. We have verified by simulation that the aggregate throughput is quite insensitive on the fact that different stations experience different FER values on the transmitted packets, provided that the maximum FER value affecting the packets of whatever station in the network, is lower than
<sup>5</sup>
<sup>5</sup>5Notice that $`8 \times 10^{- 2}`$ corresponds, for $`{E\hspace{0pt}{\lbrack{P\hspace{0pt}L}\rbrack}} = 1024`$ bytes, to the worst allowed FER associated to the receiver sensitivity specified in the IEEE 802.11b standard \[[1](#bib.bib1)\].
$`8 \times 10^{- 2}`$. As a reference scenario, we considered $`N = 9`$ contending stations divided uniformly in three groups, with the assumption that the stations in each group transmit packets affected by the same FERs noted in the first column of Table [II](#S6.T2 "TABLE II ‣ Saturation Throughput Analysis of IEEE 802.11 in Presence of Non Ideal Transmission Channel and Capture Effects"). Then, we evaluated through simulation the aggregate throughput, shown in the second column, in six configurations, three of which are characterized by a maximum FER equal to $`10^{- 2}`$, while the others are associated to a maximum FER equal to $`10^{- 3}`$. The other network parameters are noted in the next section. Upon comparing the scenario in which all the stations transmit packets affected by the same FER, with a scenario whereby some packets are affected by lower FER values, it is possible to note that the obtained throughput does not change significantly, confirming the simplifying assumption considered in this paper. Furthermore, this is a common hypothesis widely used in the literature \[[9](#bib.bib9)\].
Channel errors on the transmitted packets have been accounted for as it is done within ns-2 \[[28](#bib.bib28)\] simulator. In other words, a uniformly distributed binary random variable $`X_{e}`$ is generated in order to decide if a transmitted packet is received erroneously. The statistic of such a random variable is $`{P\hspace{0pt}{({X_{e} = 1})}} = {P_{e}\hspace{0pt}{({S\hspace{0pt}N\hspace{0pt}R})}}`$ (as specified in ([28](#S5.E28 "In V-A Simulation Setup ‣ V Simulation Results and Model Validations ‣ Saturation Throughput Analysis of IEEE 802.11 in Presence of Non Ideal Transmission Channel and Capture Effects"))), and $`{P\hspace{0pt}{({X_{e} = 0})}} = {1 - {P_{e}\hspace{0pt}{({S\hspace{0pt}N\hspace{0pt}R})}}}`$.
###
V-B
Simulation Results
In what follows, we shall present theoretical and simulation results for the lowest supported data rate. We note that by repeating the process, similar curves can be generated for all other types of modulation formats. All we need is really the BER as a function of SNR for each modulation format and the corresponding raw data rate over the channel. If the terminals use rate adaptation, then under optimal operating condition, the achievable throughput for a given SNR is the maximum over the set of modulation formats supported.
In the following, we present simulation results for the raw data rate of 1 Mbps. We have verified a close match between theoretical and simulated performance for other transmitting data rates as well. In the simulation results presented below we assume the following values for the contention window: CWmin=32, m=5, and CWmax = $`2^{m} \cdot`$CWmin = 1024.
The theoretical behavior of the probabilities $`\tau`$, $`P_{e\hspace{0pt}q}`$, and $`P_{c\hspace{0pt}o\hspace{0pt}l}`$ are depicted in Figs. [2](#S6.F2 "Figure 2 ‣ Saturation Throughput Analysis of IEEE 802.11 in Presence of Non Ideal Transmission Channel and Capture Effects") and [3](#S6.F3 "Figure 3 ‣ Saturation Throughput Analysis of IEEE 802.11 in Presence of Non Ideal Transmission Channel and Capture Effects") for both 2 way and 4 way hand-shaking mechanisms as a function of the channel SNR. All the curves have been drawn for $`N = 20`$ contending stations and for a payload size equal to 1024 bytes. The curves have been parameterized with respect to the capture threshold $`z_{0}`$. A comparative analysis of the two figures reveals that for increasing values of the quality of the channel, as exemplified by the SNR, the transmission probability $`\tau`$ increases up to reaching a saturated level around $`40`$dB, above which the channel quality can be well assumed as ideal. Furthermore, this probability increases for lower values of the capture threshold $`z_{0}`$, i.e., for higher values of the capture probability.
A comparative analysis of the curves shown in Fig. [3](#S6.F3 "Figure 3 ‣ Saturation Throughput Analysis of IEEE 802.11 in Presence of Non Ideal Transmission Channel and Capture Effects") reveals that the equivalent transmission failure probability decreases for increasing values of SNR, achieving asymptotic values essentially equal to collision probability at high SNR values. Notice that the relation defining probability $`\tau`$ for the 2 way mechanism, as defined in ([9](#S3.E9 "In III-A Markovian Process Analysis and Throughput Computation ‣ III Markovian Model Characterizing the MAC Layer under Real Transmission Channel and Capture Effects Using the 4-Way Handshaking Mechanism ‣ Saturation Throughput Analysis of IEEE 802.11 in Presence of Non Ideal Transmission Channel and Capture Effects")), tends to the relation found by Bianchi \[[2](#bib.bib2)\]:
<table>
<tbody>
<tr>
<td></td>
<td>τ=2(1−2p)(1−2p)(W+1)+pW(1−(2p)m)𝜏212𝑝12𝑝𝑊1𝑝𝑊1superscript2𝑝𝑚\tau=\frac{2(1-2p)}{(1-2p)(W+1)+pW(1-(2p)^{m})}</td>
<td></td>
</tr>
</tbody>
</table>
when $`{S\hspace{0pt}N\hspace{0pt}R}\rightarrow\infty`$, i.e., for $`{P_{e}\hspace{0pt}{({S\hspace{0pt}N\hspace{0pt}R})}}\rightarrow 0`$, and $`P_{c\hspace{0pt}a\hspace{0pt}p}\rightarrow 0`$. In this case, $`P_{e\hspace{0pt}q}\rightarrow p`$, where $`p`$ is the collision probability as defined by Bianchi, and here identified by $`P_{c\hspace{0pt}o\hspace{0pt}l}`$. In this respect, our model is more general and also embraces Bianchi’s model as an asymptotic case. This is clearly highlighted in Figs. [2](#S6.F2 "Figure 2 ‣ Saturation Throughput Analysis of IEEE 802.11 in Presence of Non Ideal Transmission Channel and Capture Effects") and [3](#S6.F3 "Figure 3 ‣ Saturation Throughput Analysis of IEEE 802.11 in Presence of Non Ideal Transmission Channel and Capture Effects") whereby the Bianchi’s probabilities as specified in \[[2](#bib.bib2)\] are depicted as horizontal lines due to the independence of the Bianchi’s model on both capture effects and channel errors. On the other hand, for very low values of SNR, the equivalent failure probability $`P_{e\hspace{0pt}q}`$ is practically 1 no matter what values are assumed by the other parameters.
Fig. [4](#S6.F4 "Figure 4 ‣ Saturation Throughput Analysis of IEEE 802.11 in Presence of Non Ideal Transmission Channel and Capture Effects") shows the behavior of the saturation throughput for the 2 way mechanism as a function of the SNR, for three different capture thresholds, $`z_{0}`$, and for 5 transmitting stations. Simulated points are marked by star on the respective theoretical curves. As expected, throughput improves as the SNR increases up to 35 dB. Furthermore, throughput improves because of the capture effects.
Fig. [5](#S6.F5 "Figure 5 ‣ Saturation Throughput Analysis of IEEE 802.11 in Presence of Non Ideal Transmission Channel and Capture Effects") shows the behavior of the saturation throughput for the 2 way mechanism as a function of the SNR, for three different capture thresholds, $`z_{0}`$, and for 20 transmitting stations. Upon comparing the curves shown in Fig.s [4](#S6.F4 "Figure 4 ‣ Saturation Throughput Analysis of IEEE 802.11 in Presence of Non Ideal Transmission Channel and Capture Effects") and [5](#S6.F5 "Figure 5 ‣ Saturation Throughput Analysis of IEEE 802.11 in Presence of Non Ideal Transmission Channel and Capture Effects"), it is easily seen that capture effects allow the system throughput to be almost the same independently from the number of stations, since throughput curves related to $`z_{0} = 1`$ dB are very close. On the other hand, when capture probability is low, i.e., for $`z_{0} = 24`$dB, collision probability reduces the throughput as the number of contending stations increase.
With the aim of comparing the throughput predicted by the proposed model in a variety of transmission conditions to the Bianchi’s throughput, Fig. [5](#S6.F5 "Figure 5 ‣ Saturation Throughput Analysis of IEEE 802.11 in Presence of Non Ideal Transmission Channel and Capture Effects") also shows the Bianchi’s throughput along with the maximum throughput $`S_{m}`$ achievable upon optimizing the minimum contention window, $`W`$, as discussed in \[[2](#bib.bib2)\]. The maximum throughput evaluated by Bianchi is defined as:
<table>
<tbody>
<tr>
<td></td>
<td>Sm=E[PL]Ts+σK+Tc(K(e1/K−1)−1)subscript𝑆𝑚𝐸delimited-[]𝑃𝐿subscript𝑇𝑠𝜎𝐾subscript𝑇𝑐𝐾superscript𝑒1𝐾11S_{m}=\frac{E[PL]}{T_{s}+\sigma K+T_{c}\left(K(e^{1/K}-1)-1\right)}</td>
<td></td>
<td>(32)</td>
</tr>
</tbody>
</table>
whereby $`K = \sqrt{\frac{T_{c}}{2\hspace{0pt}\sigma}}`$, while $`T_{c}`$ and $`T_{s}`$ are defined in ([25](#S4.E25 "In IV-A Markovian Process Analysis and Throughput Computation ‣ IV Analysis of the Markov chain for the basic 2-way handshaking mechanism in presence of channel errors and capture effects ‣ Saturation Throughput Analysis of IEEE 802.11 in Presence of Non Ideal Transmission Channel and Capture Effects")) for the 2-way mechanism, and ([17](#S3.E17 "In 3rd item ‣ III-A Markovian Process Analysis and Throughput Computation ‣ III Markovian Model Characterizing the MAC Layer under Real Transmission Channel and Capture Effects Using the 4-Way Handshaking Mechanism ‣ Saturation Throughput Analysis of IEEE 802.11 in Presence of Non Ideal Transmission Channel and Capture Effects")) for the 4-way mechanism. Notice that, as already suggested by Bianchi, $`S_{m}`$ is independent of the number $`N`$ of contending stations.
First of all, notice that as $`{S\hspace{0pt}N\hspace{0pt}R}\rightarrow\infty`$ and in absence of capture, the throughput predicted by our model tends to the Bianchi’s throughput, which also corresponds to about $`z_{0} = 24`$dB, i.e., a transmission scenario in which capture threshold is so high that capture probability is very low. On the other hand, for decreasing values of $`z_{0}`$, throughput tends to increase, as expected, in the presence of capture. Secondly, for values of $`z_{0} = 6`$dB, capture effects are such that the throughput reaches the value $`S_{m}`$ predicted by Bianchi under the hypothesis of employing an optimal minimum contention window size, even though in our simulation setup the minimum contention window is not chosen to be the optimal value yielding ([32](#S5.E32 "In V-B Simulation Results ‣ V Simulation Results and Model Validations ‣ Saturation Throughput Analysis of IEEE 802.11 in Presence of Non Ideal Transmission Channel and Capture Effects")). For increasing values of capture probabilities, as exemplified by $`z_{0} = 1`$dB, throughput goes above $`S_{m}`$. This is essentially due to the fact that capture tends to reduce the collision probability experienced by the stations which attempt to access the channel simultaneously, validating the key hypothesis suggested in ([10](#S3.E10 "In III-A Markovian Process Analysis and Throughput Computation ‣ III Markovian Model Characterizing the MAC Layer under Real Transmission Channel and Capture Effects Using the 4-Way Handshaking Mechanism ‣ Saturation Throughput Analysis of IEEE 802.11 in Presence of Non Ideal Transmission Channel and Capture Effects")). Indeed, simulation results confirm such a hypothesis.
In order to assess throughput performance as a function of the payload size, Fig. [6](#S6.F6 "Figure 6 ‣ Saturation Throughput Analysis of IEEE 802.11 in Presence of Non Ideal Transmission Channel and Capture Effects") shows throughput performance as a function of the SNR, for two different values of the payload sizes. Depending on the channel quality as exemplified by the SNR on the abscissa in Fig. [6](#S6.F6 "Figure 6 ‣ Saturation Throughput Analysis of IEEE 802.11 in Presence of Non Ideal Transmission Channel and Capture Effects"), it could be preferable to transmit shorter packets for low SNRs. On the other hand, when channel quality improves, longer packets allow one to increase throughput.
Finally, Fig.s [7](#S6.F7 "Figure 7 ‣ Saturation Throughput Analysis of IEEE 802.11 in Presence of Non Ideal Transmission Channel and Capture Effects") and [8](#S6.F8 "Figure 8 ‣ Saturation Throughput Analysis of IEEE 802.11 in Presence of Non Ideal Transmission Channel and Capture Effects") show the behavior of the saturation throughput as a function of the number of the contending stations for 2 and 4 way mechanism, respectively, for various key values of the parameters SNR and $`z_{0}`$. In both figures, SNR$`= \infty`$ is used to signify ideal transmission (i.e., $`P_{e} = 0`$), while $`z_{0} = \infty`$ means absence of capture effects. Notice that, as expected, in the 4 way mechanism, saturation throughput performance is less sensitive to capture effects. For comparison purposes, Fig.s [7](#S6.F7 "Figure 7 ‣ Saturation Throughput Analysis of IEEE 802.11 in Presence of Non Ideal Transmission Channel and Capture Effects") shows Bianchi’s throughput along with the maximum achievable throughput in ([32](#S5.E32 "In V-B Simulation Results ‣ V Simulation Results and Model Validations ‣ Saturation Throughput Analysis of IEEE 802.11 in Presence of Non Ideal Transmission Channel and Capture Effects")). Notice that Bianchi’s throughput is superimposed on the curve labelled $`{S\hspace{0pt}N\hspace{0pt}R} = \infty`$ and $`z_{0} = \infty`$, as already noted from the simulation results shown in the previous figures. Achievable throughput in a variety of transmission conditions is usually less than $`S_{m}`$ in ([32](#S5.E32 "In V-B Simulation Results ‣ V Simulation Results and Model Validations ‣ Saturation Throughput Analysis of IEEE 802.11 in Presence of Non Ideal Transmission Channel and Capture Effects")), specially when $`S\hspace{0pt}N\hspace{0pt}R`$ is lower than $`45`$dB: in this case, channel errors tend to dominate reducing the throughput. On the other hand, throughput can be higher than $`S_{m}`$ for a low number of contending stations (in the considered scenario, $`N \leq 20`$ contending stations) and for low values of capture thresholds $`z_{0}`$, say $`z_{0} \leq 6`$dB. On the other hand, employing the 4 way mechanism, throughput is very close to the theoretical Bianchi’s value $`S_{m} = 0.86`$ Mbps no matter what the capture threshold $`z_{0}`$. This is due to the impact of the 4 way mechanism mitigating the effects of collisions between multiple contending stations.
## VI Conclusions
In this paper, we provided an extension of the Markov model characterizing the DCF behavior at the MAC layer of the IEEE802.11 series of standards accounting for channel induced errors and capture effects typical of fading environments. The modelling allows taking into consideration the impact of channel contention in throughput analysis which is often not considered or it is considered in a static mode by using a mean contention period. Subsequently, based on justifiable assumptions, the stationary probability of the Markov chain is calculated to obtain the saturation throughput. The closed form expressions allow derivation of the throughput as a function of a multitude of system level parameters including packet and header sizes for a variety of applications. Simulation results confirm the validity of the proposed theoretical models.
## References
- \[1\]
IEEE Standard for Wireless LAN Medium Access Control (MAC) and Physical Layer (PHY) Specifications, November 1997, P802.11
- \[2\]
G. Bianchi, ”Performance analysis of the IEEE 802.11 distributed coordination function”, IEEE JSAC, Vol.18, No.3, March 2000.
- \[3\]
Jae Hyun Kim and Jong Kyu Lee ”Capture effects of wireless CSMA/CA protocols in Rayleigh and shadow fading channels”, IEEE Trans. on Veh. Tech., Vol.48, No.4, pp.1277-1286, July 1999.
- \[4\]
M. Zorzi and R.R. Rao, ”Capture and retransmission control in mobile radio”, IEEE JSAC, Vol.12, No.8, pp.1289 - 1298, Oct. 1994.
- \[5\]
Z. Hadzi-Velkov and B. Spasenovski, ”Capture effect in IEEE 802.11 basic service area under influence of Rayleigh fading and near/far effect”, In Proc. of 13th IEEE International Symposium on Personal, Indoor and Mobile Radio Communications, Vol.1, pp.172 - 176, Sept. 2002.
- \[6\]
J.W. Robinson and T.S. Randhawa ”Saturation throughput analysis of IEEE 802.11e enhanced distributed coordination function”, IEEE JSAC, Vol.22, No.5, June 2004.
- \[7\]
Z. Kong, D.H.K. Tsang, B.Bensaou, and D.Gao, ”Performance analysis of IEEE 802.11e contention-based channel access”, IEEE JSAC, Vol.22, No.10, Dec. 2004.
- \[8\]
D. Qiao, S. Choi, and K.G. Shin ”Goodput analysis and link adaptation for IEEE 802.11a wireless LANs”, IEEE Trans. On Mobile Computing, Vol.1, No.4, Oct.-Dec. 2002.
- \[9\]
P. Chatzimisios, A.C. Boucouvalas, and V. Vitsas, ”Influence of channel BER on IEEE 802.11 DCF”, IEE Electronics Letters, Vol.39, No.23, pp.1687-1689, Nov. 2003.
- \[10\]
Wei Zhuang, Yung-Sze Gan, Kok-Jeng Loh, and Kee-Chaing Chua, ”Policy-based QoS-management architecture in an integrated UMTS and WLAN environment”, IEEE Communications Magazine, Vol.41, No.11, pp.118-125, Nov. 2003.
- \[11\]
W. Liu, W. Lou, X. Chen, and Y. Fang, ”A QoS-enabled MAC architecture for prioritized service in IEEE 802.11 WLANs”, Proc. IEEE GLOBECOM’03, Vol.7, pp.3802-3807, Dec. 2003.
- \[12\]
G. Fodor, A. Eriksson, and A. Tuoriniemi, ”Providing quality of service in always best connected networks”, IEEE Communications Magazine, Vol.41, No.7, pp.154-163, July 2003.
- \[13\]
J. Zhao, Z. Guo, Q. Zhang, and W. Zhu, ”Performance study of MAC for service differentiation in IEEE 802.11”, In Proc. of IEEE GLOBECOM’02, Vol.1, pp.778-782, Nov. 2002.
- \[14\]
W. Pattara-Atikom, P. Krishnamurthy, and S. Banerjee, ”Distributed mechanisms for quality of service in wireless LANs”, IEEE Personal Communications, Vol.10, No.3, pp.26-34, June 2003.
- \[15\]
Ye Ge and J. Hou, ”An analytical model for service differentiation in IEEE 802.11”, Proc. ICC ’03, Vol.2, pp.1157-1162, May 2003.
- \[16\]
H. L. Truong and G. Vannuccini, ”Performance evaluation of the QoS enhanced IEEE 802.11e MAC layer”, In Proc. of IEEE VTC 2003-Spring, Vol.2, pp.940-944, April 2003.
- \[17\]
Q. Li and M. VanderSchaar, ”Providing adaptive QoS to layered video over wireless local area networks through real-time retry limit adaptation”, IEEE Trans. on Multimedia, Vol.6, No.2, pp.278-290, April 2004.
- \[18\]
Q. Zhang, C. Guo, Z. Guo, and W. Zhu, ”Efficient mobility management for vertical handoff between WWAN and WLAN”, IEEE Communications Magazine, Vol.41, No.11, pp.102-108, Nov. 2003.
- \[19\]
H. Honkasalo, K. Pehkonen, M. T. Niemi, and A.T. Leino, ”WCDMA and WLAN for 3G and beyond”, IEEE Personal Communications, Vol.9, No.2, pp.14-18, April 2002.
- \[20\]
S. Park, K. Kim, D.C. Kim, S. Choi, and S. Hong, ”Collaborative QoS architecture between DiffServ and 802.11e wireless LAN”, In Proc. of IEEE VTC 2003-Spring, Vol.2, pp.945-949, April 2003.
- \[21\]
F. Daneshgaran and M. Laddomada, “Reduced complexity interleaver growth algorithm for turbo codes,” IEEE Trans. on Wireless Communications, vol.4, no.3, pp.954-964, May 2005.
- \[22\]
F. Daneshgaran, M. Laddomada, and M. Mondin, “Interleaver design for serially concatenated convolutional codes: Theory and application,” IEEE Trans. on Information Theory, Vol.50, No. 6, pp. 1177-1188, June 2004.
- \[23\]
Proakis, J. G., “Digital Communications,” McGraw Hill, Fourth Edition, New York, 2001.
- \[24\]
F. Daneshgaran, M. Laddomada, and M. Mondin, “An extensive search for good punctured rate k/k+1 recursive convolutional codes for serially concatenated convolutional codes,” IEEE Trans. on Inform. Theory, vol.50, no.1, pp.208-217, January 2004.
- \[25\]
F. Daneshgaran, M. Laddomada, and M. Mondin, “High rate recursive convolutional codes for concatenated channel codes,” IEEE Trans. on Communications, vol.52, no.11, pp.1846-1850, Nov. 2004.
- \[26\]
M.K. Simon and M. Alouini, Digital Communication over Fading Channels: A Unified Approach to Performance Analysis, Wiley-Interscience, 1st edition, 2000.
- \[27\]
T. Rappaport, Wireless Communications, Prentice Hall, 1999.
- \[28\] Wu Xiuchao, ”Simulate 802.11b Channel within NS-2”, available online at `http://www.comp.nus.edu.sg/~wuxiucha/research/` `reactive/report/80211ChannelinNS2_new.pdf`
[Refer to caption](/html/0710.5236/assets/x1.png)
Figure 1: Markov chain for the contention model based on the 4-way handshaking technique, considering the effects of capture and channel induced errors.
[Refer to caption](/html/0710.5236/assets/x2.png)
Figure 2: Theoretical behavior of the probability $`\tau`$ as a function of SNR for both 2 way (curves labelled 2W) and 4 way (curves labelled 4W) handshaking mechanisms. Curves have been obtained for the capture threshold $`z_{0}`$, payload size equal to 1024 bytes and for $`N = 20`$ contending stations.
[Refer to caption](/html/0710.5236/assets/x3.png)
Figure 3: Theoretical behaviour of the equivalent failure transmission probability $`P_{e\hspace{0pt}q}`$ as a function of SNR for both 2 way (curves labelled 2W) and 4 way (curves labelled 4W) handshaking mechanisms. Curves have been obtained for the capture threshold $`z_{0}`$, payload size equal to 1024 bytes and for $`N = 20`$ contending stations.
[Refer to caption](/html/0710.5236/assets/x4.png)
Figure 4: Saturation throughput for the 2 way mechanism as a function of the SNR, for three different capture thresholds $`z_{0}`$. Simulated points are identified by star-markers over the respective theoretical curves. Payload size is 1024 bytes, while the number of contending stations is $`N = 5`$.
[Refer to caption](/html/0710.5236/assets/x5.png)
Figure 5: Saturation throughput for the 2 way mechanism as a function of the SNR, for three different capture thresholds $`z_{0}`$. Simulated points are identified by star-markers over the respective theoretical curves. Payload size is 1024 bytes, while the number of contending stations is $`N = 20`$.
[Refer to caption](/html/0710.5236/assets/x6.png)
Figure 6: Saturation throughput for the 2 way mechanism as a function of the SNR, for payload sizes 128 and 1024 bytes, and $`z_{0} = {6,24}`$ dB. Simulated points are identified by star-markers over the respective theoretical curves.
[Refer to caption](/html/0710.5236/assets/x7.png)
Figure 7: Saturation throughput for the 2 way mechanism as a function of the number of contending stations, for various values of the SNR and $`z_{0}`$ parameters. Payload size is 1024 bytes. Simulated points are identified by star-markers over the respective theoretical curves.
[Refer to caption](/html/0710.5236/assets/x8.png)
Figure 8: Saturation throughput for the 4 way mechanism as a function of the number of contending stations, for various values of the SNR and $`z_{0}`$ parameters. Payload size is 1024 bytes. Simulated points are identified by star-markers over the respective theoretical curves.
TABLE I: Typical network parameters
<table>
<thead>
<tr>
<th>MAC header</th>
<th>24 bytes</th>
</tr>
</thead>
<tbody>
<tr>
<th>PHY header</th>
<td>16 bytes</td>
</tr>
<tr>
<th>Payload size</th>
<td>1024 bytes</td>
</tr>
<tr>
<th>ACK</th>
<td>14 bytes</td>
</tr>
<tr>
<th>RTS</th>
<td>20 bytes</td>
</tr>
<tr>
<th>CTS</th>
<td>14 bytes</td>
</tr>
<tr>
<th>propagation delay τpsubscript𝜏𝑝\tau_{p}</th>
<td>1 μs𝜇𝑠\mu s</td>
</tr>
<tr>
<th>Slot time</th>
<td>20 μs𝜇𝑠\mu s</td>
</tr>
<tr>
<th>SIFS</th>
<td>10 μs𝜇𝑠\mu s</td>
</tr>
<tr>
<th>DIFS</th>
<td>50 μs𝜇𝑠\mu s</td>
</tr>
<tr>
<th>EIFS</th>
<td>300 μs𝜇𝑠\mu s</td>
</tr>
<tr>
<th>ACK timeout</th>
<td>300 μs𝜇𝑠\mu s</td>
</tr>
<tr>
<th>CTS timeout</th>
<td>300 μs𝜇𝑠\mu s</td>
</tr>
</tbody>
</table>
TABLE II: Throughput dependence on the per-station FER in a scenario with $`N = 9`$ contending stations using the 2-way mechanism without capture. Packet size is $`{E\hspace{0pt}{\lbrack{P\hspace{0pt}L}\rbrack}} = 1024`$ bytes.
<table>
<thead>
<tr>
<th>FER</th>
<th>Sisubscript𝑆𝑖S_{i} [Mbps]</th>
</tr>
</thead>
<tbody>
<tr>
<th>10−2,10−2,10−2superscript102superscript102superscript10210^{-2},10^{-2},10^{-2}</th>
<td>0.777</td>
</tr>
<tr>
<th>10−2,10−3,10−4superscript102superscript103superscript10410^{-2},10^{-3},10^{-4}</th>
<td>0.781</td>
</tr>
<tr>
<th>10−2,10−4,10−5superscript102superscript104superscript10510^{-2},10^{-4},10^{-5}</th>
<td>0.781</td>
</tr>
<tr>
<th>10−3,10−3,10−3superscript103superscript103superscript10310^{-3},10^{-3},10^{-3}</th>
<td>0.784</td>
</tr>
<tr>
<th>10−3,10−4,10−5superscript103superscript104superscript10510^{-3},10^{-4},10^{-5}</th>
<td>0.786</td>
</tr>
<tr>
<th>10−3,10−5,10−6superscript103superscript105superscript10610^{-3},10^{-5},10^{-6}</th>
<td>0.785</td>
</tr>
</tbody>
</table> |
0710.3824 | cs.CR | cyber | INSTITUT NATIONAL DE RECHERCHE EN INFORMATIQUE ET EN AUTOMATIQUE
Deterministic Secure Positioning in Wireless Sensor Networks
Sylvie Delaët — Partha Sarathi Mandal — Mariusz Rokicki — Sébastien TixeuilN° 9999
Octobre 2007
Deterministic Secure Positioning in Wireless Sensor Networks
Sylvie Delaët
<sup>†</sup>
<sup>†</sup>thanks: Univ. Paris-Sud XI, France
, Partha Sarathi Mandal
<sup>†</sup>
<sup>†</sup>thanks: INRIA Futurs & Univ. Paris-Sud XI, France
, Mariusz Rokicki
<sup>†</sup>
<sup>†</sup>thanks: CNRS & Univ. Paris-Sud XI, France
, Sébastien Tixeuil
<sup>†</sup>
<sup>†</sup>thanks: Univ. Pierre & Marie Curie, INRIA Futurs, France
Thème NUM — Systèmes numériques
Projet Grand Large
Rapport de recherche n° 9999 — Octobre 2007 — ?? pages
Abstract: Properly locating sensor nodes is an important building block for a large subset of wireless sensor networks (WSN) applications. As a result, the performance of the WSN degrades significantly when misbehaving nodes report false location and distance information in order to fake their actual location. In this paper we propose a general distributed deterministic protocol for accurate identification of faking sensors in a WSN. Our scheme does *not* rely on a subset of *trusted* nodes that are not allowed to misbehave and are known to every node in the network. Thus, any subset of nodes is allowed to try faking its position. As in previous approaches, our protocol is based on distance evaluation techniques developed for WSN.
On the positive side, we show that when the received signal strength (RSS) technique is used, our protocol handles at most $`{\lfloor\frac{n}{2}\rfloor} - 2`$ faking sensors. Also, when the time of flight (ToF) technique is used, our protocol manages at most $`{\lfloor\frac{n}{2}\rfloor} - 3`$ misbehaving sensors. On the negative side, we prove that no deterministic protocol can identify faking sensors if their number is $`{\lceil\frac{n}{2}\rceil} - 1`$. Thus our scheme is almost optimal with respect to the number of faking sensors.
We discuss application of our technique in the trusted sensor model. More precisely our results can be used to minimize the number of trusted sensors that are needed to defeat faking ones.
Key-words: Wireless Sensor Network, Secure Positioning, Distributed Protocol, Faking Sensor.
Localisation déterministe et sécurisée dans les réseaux de capteurs
Résumé : Localiser correctement des capteurs autonomes est une brique de base importante pour un grand nombre d’applications dans les réseaux de capteurs (WSN). En effet, l’efficacité du WSN est significativement dégradée quand des nœuds malicieux rapportent de fausses positions et de fausses informations de distance de manière à simuler une localisation fictive. Dans cet article, nous proposons une solution algorithmique distribuée pour l’identification exacte des capteurs malicieux dans un WSN. Notre approche n’est pas basée sur l’utilisation d’un sous-ensemble de nœuds “de confiance” qui serait connu de chaque autre nœud du WSN. Ainsi, tout sous-ensemble des participants peut essayer de tricher sur sa position. Comme dans les approches précédentes, notre protocole est basé sur des techniques d’évaluation des distances développées pour les WSN.
Nous montrons que quand la technique de la force du signal reçu (RSS) est utilisée, notre protocole peut tolérer au plus $`{\lfloor\frac{n}{2}\rfloor} - 2`$ nœuds malicieux. De plus, quand la technique du temps de vol (ToF) est utilisée, notre protocole peut gérer au plus $`{\lfloor\frac{n}{2}\rfloor} - 3`$ tricheurs. Nous montrons également qu’il est impossible pour un protocole déterministe d’identifier les nœuds malicieux si leur nombre est au moins égal à $`{\lceil\frac{n}{2}\rceil} - 1`$, ce qui rend notre résultat presque optimal en ce qui concerne le nombre de nœuds malicieux tolérés.
Nous discutons l’application de notre technique au modèle où il existe des nœuds de confiance. Plus précisément, nos résultats peuvent être utilisés pour minimiser le nombre de nœuds de confiance nécessaires à la détection sans faille des nœuds malicieux.
Mots-clés : Réseaux de capteurs sans fil, localisation sécurisée, algorithme distribué, capteurs malicieux.
## Chapter 1 Introduction
Properly locating sensor nodes is an important building block for a large subset of wireless sensor networks (WSN) applications. For example, environment and habitat monitoring \[[20](#bib.bib20)\], surveillance and tracking for military \[[10](#bib.bib10)\] or civilian purpose, both require the knowledge of the location where a particular event takes place. Location of nodes in a WSN can also be used for location based routing algorithms (such as geographic routing \[[14](#bib.bib14)\]), or location based services.
Most of existing position verification protocols rely on distance evaluation techniques (*e.g.* \[[1](#bib.bib1), [9](#bib.bib9), [11](#bib.bib11), [19](#bib.bib19), [21](#bib.bib21), [22](#bib.bib22)\]). Received signal strength (RSS) \[[1](#bib.bib1)\] and time of flight (ToF) \[[9](#bib.bib9)\] techniques are relatively easy to implement yet very precise (one or two meters). In the RSS technique, receiving sensor estimates the distance of the sender on the basis of sending and receiving signal strengths. In the ToF technique, sensor estimates distance based on message delay and radio signal propagation time. Position verification using the aforementioned distance estimation techniques is relatively straighforward provided that *all* sensors cooperate. However, this task becomes challenging in the presence of misbehaving nodes that are allowed to report false position and distance information in order to fake their actual position. In the following such nodes are denoted as *faking* or *cheating* nodes.
Such misbehaviors could occur due to several factors: a sensor may malfunction due to improper sensor deployment, partial communication problem due objects in the vicinity, or inaccurate position (coordinates) estimation. We consider that misbehaving sensors are unaware that they are malfunctioning, so locally they properly execute the protocol that is given to all nodes. Nevertheless, they can report incorrect position, change signal strength (when the RSS technique is used), or report incorrect transmission time (when the ToF technique is used).
### 1.1 Related Work
Most methods \[[3](#bib.bib3), [4](#bib.bib4), [16](#bib.bib16), [15](#bib.bib15)\] existing in the literature that use distance estimation techniques to detect and filter out faking nodes are based on the availability of a few fixed trusted entities (or *verifiers*), that are equipped with GPS. We refer to this model as the *trusted sensor* (or *TS*) model. In this model, the faking nodes may use attacks not available to regular nodes, such as radio signal jamming or using directional antenas, that permit to implement *e.g.* wormhole attack \[[12](#bib.bib12)\] and Sybil attack \[[8](#bib.bib8)\]. Lazos and Poovendran \[[15](#bib.bib15)\] present a secure range-independent localization scheme, where each sensor computes its position based on received beacons messages from locators. Sensors compute the center of gravity of beacons’s intersection region, and the computed location becomes the estimated location of the sensor. Probabilistic analysis of the protocol demonstrate that it is resilient to wormhole and Sybil attacks, with high probability. Lazos *et al.* \[[16](#bib.bib16)\] further refine this scheme with multilateration to reduce the number of required locator, while maintaining probabilistic guarantees. The protocol of Capkun and Hubaux \[[4](#bib.bib4)\] relies on a distance bounding technique proposed by Brands and Chaum \[[2](#bib.bib2)\]. Each sensor $`v`$ measures its distance to a (potential) faking sensor $`u`$ based on its message round-trip delay and radio signal propagation time, thus enabling the faking node $`u`$ only to *enlarge* the distance to $`v`$. Then, if the faking node is located inside the triangle formed by verifiers and its faked position is also located within the triangle, then at least one of the three verifiers detects an inconsistency. Capkun, Cagalj, Srivastava \[[3](#bib.bib3)\] is supported by powerful verifiers, that know their positions and communicate with some wired channels that prevent faking nodes to locate them or to listen their transmissions. Then, each verifier $`v`$ measures the arrival time $`t_{v}`$ of the (potential) faking node transmission. Verifiers exchange all such arrival times and check consistency of the declared position. However, the TS model presents several drawback in WSNs: first the network can not self-organize in an entirely distributed manner, and second the trusted nodes have to be checked regularly and manually to actually remain trusted.
Relaxing the assumption of trusted nodes makes the problem more challenging, and to our knowledge, has only been investigated very recently \[[13](#bib.bib13)\]. We call this model where no trusted node preexists the *no trusted sensor* (or *NTS*) model. The approach of \[[13](#bib.bib13)\] is randomized and consists of two phases: distance measurement and filtering. In the distance measurement phase, sensors measure their distances to their neighbors, faking sensors being allowed to corrupt the distance measure technique. In the filtering phase each correct sensor randomly picks up $`2`$ so-called *pivot* sensors. Next each sensor $`v`$ uses trilateration with respect to the chosen pivot sensors to compute the location of its neighbor $`u`$. If there is a match between the announced location and the computed location, the $`(u,v)`$ link is added to the network, otherwise it is discarded. Of course, the chosen pivot sensors could be faking and lying, so the protocol may only give probabilistic guarantee.
In this paper we present a deterministic protocol that performs in the NTS model and where every correct (*i.e.* non faking) node: *(i)* identifies the positions (coordinates) of all correct nodes, and *(ii)* identifies the faking nodes (if any). The goal of the faking nodes is to convince the correct nodes that they are located in a fake position.
### 1.2 Our results
The main contribution of this paper is a secure deterministic positioning protocol, FindMap, in the NTS model. To the best of our knowledge, it is the first deterministic protocol for this problem in the NTS model. The basic version of the protocol assumes that faking sensors are not able to mislead distance evaluation techniques. Then, our protocol correctly filters out faking sensors provided they are at most $`{\lceil\frac{n}{2}\rceil} - 2`$. Conversely, we show evidence that it in the same setting, it is impossible to deterministically solve the problem when the number of faking sensors is at least $`{\lceil\frac{n}{2}\rceil} - 1`$. We then extend the protocol do deal with faking sensors that are also allowed to corrupt the distance measure technique (RSS or ToF). In the case of RSS, our protocol tolerates at most $`{\lfloor\frac{n}{2}\rfloor} - 2`$ faking sensors (provided that no four sensors are located on the same circle and no three sensors are co-linear). In the case of ToF, our protocol may handle up to $`{\lfloor\frac{n}{2}\rfloor} - 3`$ faking sensors (provided that no six sensors are located on the same hyperbola and no three sensors are co-linear).
Our results have significant impact on secure positioning in the TS model as well. The TS protocol presented by Capkun *et al.* \[[3](#bib.bib3)\] relies on set of hidden stations, that detect inconsistencies between measured distance and distance computed from claimed coordinates, using ToF-like technique to estimate the distance. Our detailed analysis shows that six hidden stations (verifiers) are sufficient to detect inconsistency in the same setting. In \[[3](#bib.bib3)\], the authors conjecture that the ToF-like technique could be replaced with RSS technique. Our results anwser positively to the open question of \[[3](#bib.bib3)\], improving the number of needed stations to four. So, in the TS model, our results can be used to efficiently deploy a minimal number trusted stations.
## Chapter 2 Technical preliminaries
We assume that every node is able to communicate to every other node in the WSN. The size of the WSN is $`n`$ and is known to every node. Each node is also aware of its own geographic coordinates, and those coordinates are used to identify nodes. The WSN is partially synchronous: every node operates in rounds. In one round, every node is able to send exactly one message to every other node wihout collision occuring. For each transmission, a correct nodes uses the same transmission power $`S_{s}`$.
Faking nodes are allowed to transmit incorrect coordinates (and thus incorrect identifier) to the other nodes. In the basic protocol, faking nodes can not corrupt distance measure techniques, while in Section [4](#Ch4 "Chapter 4 Protocol based on RSS ranging technique") we relax this assumption and allow faking sensors to change its radio transmitter power and send a related fake position to the correct nodes. In Section [5](#Ch5 "Chapter 5 Protocol based on ToF-like ranging techniques") a faking sensor also can report incorrect transmission time. Also, we assume that faking nodes may cooperate between themselves in an omniscient manner (*i.e.* without exchanging messages) in order to fool the correct nodes in the WSN.
We assume that all distance estimation techniques are perfect with respect to precision. The distance computed by node $`v`$ to node $`u`$ based on a distance estimation technique is denoted by $`\hat{d}\hspace{0pt}{(v,u)}`$. The distance computed by $`v`$ to the node $`u`$ using coordinates provided by $`u`$ is denoted by $`d\hspace{0pt}{(v,u)}`$. A particular sensor $`v`$ detects inconsistency on distance (*i.e.* position) of sensor $`u`$ if $`{d\hspace{0pt}{(v,u)}} \neq {\hat{d}\hspace{0pt}{(v,u)}}`$. Our protocols rely on detecting and reporting such inconsistencies.
In the remaining of the paper, we use three distance estimation techniques:
1. 1\.
In the *received signal strength* (*RSS*) technique we assume that each node can precisely measure the distance to the transmitting node from RSS by Frii’s transmission equation [2.1](#Ch2.E1 "In item 1 ‣ Chapter 2 Technical preliminaries") \[[17](#bib.bib17)\]:
<table>
<tbody>
<tr>
<td></td>
<td>Sr=Ss(λ4πd)2subscript𝑆𝑟subscript𝑆𝑠superscript𝜆4𝜋𝑑2\displaystyle S_{r}=S_{s}\left(\frac{\lambda}{4\pi d}\right)^{2}</td>
<td></td>
<td>(2.1)</td>
</tr>
</tbody>
</table>
Where $`S_{s}`$ is the transmission power of the sender, $`S_{r}`$ is the remaining power or receive signal strength (RSS) of the wave at receiver, $`\lambda`$ is wave length and $`d`$ is distance between sender and receiver.
2. 2\.
The *synchronous time of flight* (*SToF*) technique relies on propagation time of the radio signal. For this technique we assume that sensors are synchronized by global time. Sender $`u`$ attaches the time of transmission, $`t_{s}`$ to the message. The receiver $`v`$ records the message arrival time $`t_{r}`$ of the message. Next $`v`$ computes the distance $`d = {t \ast s}`$ of $`u`$ based on time delay $`t = {t_{r} - t_{s}}`$ of the message and radio signal speed $`s`$.
3. 3\.
The *different arrival time* (*DAT*) technique provides similar guarantees as SToF. The advantage of DAT over SToF is that DAT does not require synchronization. In the DAT technique each sensor transmits its message with two types of signals that differ on propagation speed *e.g.* radio signal (RF) and ultra sound signal (US). Sender sensor $`u`$ transmits its message with RF and US signal simultaneously. Receiver sensor $`v`$, which estimates its distance to sender $`u`$, records arrival time $`t_{r}`$ of RF signal and arrival time $`t_{u}`$ of US signal from $`u`$. Then, based on the propagation speed $`s_{r}`$ of RF, propagation speed $`s_{u}`$ of US and difference of arrival times $`t = {t_{u} - t_{r}}`$ sensor $`v`$ can compute distance to sensor $`u`$. Equation [2.2](#Ch2.E2 "In item 3 ‣ Chapter 2 Technical preliminaries") show the relation.
<table>
<tbody>
<tr>
<td></td>
<td>t=d^sr−d^su𝑡^𝑑subscript𝑠𝑟^𝑑subscript𝑠𝑢\displaystyle t=\frac{\hat{d}}{s_{r}}-\frac{\hat{d}}{s_{u}}</td>
<td></td>
<td>(2.2)</td>
</tr>
</tbody>
</table>
## Chapter 3 Basic Protocol
In this section we present the protocol FindMap, that essentially performs by majority voting. The protocol detects all faking sensors provided that $`{n - 2 - f} > f`$. Thus the total number of faking sensors is at most $`{\lceil\frac{n}{2}\rceil} - 2`$. In this section we consider the relatively simpler case where faking sensors are not able to cheat the distance estimation techniques (see above) that are used by the correct nodes. Our second key assumption is that no three correct sensors are co-linear. This assumption allows to formulate the following fact.
######
Fact 1
If a faking sensor transmits a message with a fake position then at least one of three correct sensors can detect an inconsistency (see Figure [3.1](#Ch3.F1 "Figure 3.1 ‣ Chapter 3 Basic Protocol")).
[Refer to caption](/html/0710.3824/assets/x1.png)
Figure 3.1: Example in which sensor $`F`$ consistently fakes its location to $`F'`$ against sensors $`P_{1}`$ and $`P_{2}`$. However the third sensor $`P_{3}`$ always detects an inconsistency since no three correct sensors are co-linear.
Based on Fact [1](#Thmfact1 "Fact 1 ‣ Chapter 3 Basic Protocol"), we can develop
FindMap(threshold)
. The protocol operates in two rounds. The protocol is paremeterized by a threshold parameter. In Round 1 all sensors exchange their coordinates by transmitting an initial message. Next each node $`v`$ computes the distances $`\hat{d}\hspace{0pt}{(v,u)}`$ (from the distance estimation technique) and $`d\hspace{0pt}{(v,u)}`$ (from the obtained node coordinates) of $`u`$ and compare them. If $`{\hat{d}\hspace{0pt}{(v,u)}} \neq {d\hspace{0pt}{(v,u)}}`$ then $`v`$ accuses $`u`$ to fake its position. Otherwise $`v`$ does not accuse $`u`$. To keep record of its accusations, each node $`v`$ maintain an array $`a\hspace{0pt}c\hspace{0pt}c\hspace{0pt}u\hspace{0pt}s_{v}`$ of size $`n`$. In Round 2 each node $`v`$ exchanges its array of accusations. Next each node $`v`$ counts accusations toward every other node $`u`$ including its own accusations. A sensor $`v`$ detects a sensor $`u`$ as faking if the number of accusations is at least equal to the threshold parameter. For our basic FindMap protocol we use $`\text{threshold} = {\lfloor\frac{n}{2}\rfloor}`$.
Protocol FindMap($`\text{threshold} = {\lfloor\frac{n}{2}\rfloor}`$)
Round 1:
1\. $`v`$ exchange coordinates by transmiting $`i\hspace{0pt}n\hspace{0pt}i\hspace{0pt}t_{v}`$ and receiving $`n - 1`$ messages.
2\. for each received message $`i\hspace{0pt}n\hspace{0pt}i\hspace{0pt}t_{u}`$:
3\. compute $`\hat{d}\hspace{0pt}{(v,u)}`$ and $`d\hspace{0pt}{(v,u)}`$ using the coordinates of $`u`$.
4\.
if ($`{\hat{d}\hspace{0pt}{(v,u)}} \neq {d\hspace{0pt}{(v,u)}}`$) then $`{a\hspace{0pt}c\hspace{0pt}c\hspace{0pt}u\hspace{0pt}s_{v}\hspace{0pt}{\lbrack u\rbrack}}\leftarrow{t\hspace{0pt}r\hspace{0pt}u\hspace{0pt}e}`$
5\.
else $`{a\hspace{0pt}c\hspace{0pt}c\hspace{0pt}u\hspace{0pt}s_{v}\hspace{0pt}{\lbrack u\rbrack}}\leftarrow{f\hspace{0pt}a\hspace{0pt}l\hspace{0pt}s\hspace{0pt}e}`$
Round 2:
6\. $`v`$ exchange accusations by transmiting $`a\hspace{0pt}c\hspace{0pt}c\hspace{0pt}u\hspace{0pt}s_{v}`$ and receiving $`n - 1`$ accusations.
7\. for each received $`a\hspace{0pt}c\hspace{0pt}c\hspace{0pt}u\hspace{0pt}s_{u}`$:
8\. for $`r = {1\hspace{0pt}t\hspace{0pt}o\hspace{0pt}n}`$
9\.
if $`{a\hspace{0pt}c\hspace{0pt}c\hspace{0pt}u\hspace{0pt}s_{u}\hspace{0pt}{\lbrack r\rbrack}} = {t\hspace{0pt}r\hspace{0pt}u\hspace{0pt}e}`$ then $`{{N\hspace{0pt}u\hspace{0pt}m\hspace{0pt}A\hspace{0pt}c\hspace{0pt}c\hspace{0pt}u\hspace{0pt}s_{r}} +} = 1`$
10\. for each sensor $`u`$:
11\.
if ($`\text{threshold} \leq {N\hspace{0pt}u\hspace{0pt}m\hspace{0pt}A\hspace{0pt}c\hspace{0pt}c\hspace{0pt}u\hspace{0pt}s_{u}}`$) then $`v`$ considers $`u`$ is faking.
######
Theorem 1
Protocol FindMap($`\lfloor\frac{n}{2}\rfloor`$) identifies all the faking sensors and finds the position of correct sensors provided $`{n - f - 2} > f`$.
Proof: First we will show that each faking sensors will be accused by proper number of correct sensors. In each subset of three correct sensors there exists at least one which detects inconsistency on distance to a faking sensors. This is guaranteed by fact [1](#Thmfact1 "Fact 1 ‣ Chapter 3 Basic Protocol"). Thus each faking sensors will be accused by at least $`n - f - 2`$ correct sensors. Inequality $`{n - f - 2} > f`$ guarantees that number of correct sensors is at least $`\lfloor\frac{n}{2}\rfloor`$. We can also observe that each correct sensors can be accused by at most $`{\lceil\frac{n}{2}\rceil} - 2`$ faking sensors. However this is not enough to find a correct sensors faking. $`\square`$
Next we show that it is impossible to detect the real location of correct sensors and filter out the faking one when $`{n - 2 - f} \leq f`$. The assumption that faking sensors cannot corrupt the distance ranging technique makes this result even stronger. Our protocol is synchronous but this impossibility result holds for asynchronous settings too.
######
Theorem 2
If $`{n - f - 2} \leq f`$ then the real location of the correct sensors cannot be detected by a deterministic protocol.
Proof:
[Refer to caption](/html/0710.3824/assets/x2.png)
Figure 3.2: First execution.
Let us assume that correct sensors run a protocol $`\mathcal{P}`$, which allows to detect location of correct sensors and identify the faking sensors even when $`{n - f - 2} = f`$. In case $`{n - f - 2} < f`$ we make some faking sensors correct to achieve equality and in case $`n`$ is odd one of the faking sensors will remain silent. Let us consider the first execution (see figure [3.2](#Ch3.F2 "Figure 3.2 ‣ Chapter 3 Basic Protocol")). There are two correct sensors $`v`$ and $`u`$ located on the straight line $`l`$. There are two sets of sensors $`C`$-correct sensors and $`\Gamma`$-faking sensors located on the lower half of the plane. The sizes of the sets are equal $`{|C|} = {|\Gamma|} = f`$. The sensors in $`\Gamma`$ are trying to convince sensors $`v`$ and $`u`$ that they are located in $`\Gamma'`$ on the other side of the straight line $`l`$ symmetrically. Each sensor in $`\Gamma`$ behave as if it was a correct sensor reflected symmetrically against straight line $`l`$. The sensors in $`\Gamma'`$ are called virtual faking sensors. Virtual sensors in $`\Gamma'`$ execute the protocol as if sensors in $`C`$ were faking and their correct location was in $`C'`$, which is symmetric reflection of $`C`$ against straight line $`l`$. Construction of the second execution will clarify why we need such behavior of sensors in $`\Gamma'`$. We can see that sensors $`v`$ and $`u`$ are not able to detect inconsistency directly on the distance of virtual faking sensors since symmetry preserves their distances from $`v`$ and $`u`$. By our assumption about correctness of the protocol $`\mathcal{P}`$ sensors $`v`$ and $`u`$ are able to verify that sensors in $`\Gamma'`$ are faking.
[Refer to caption](/html/0710.3824/assets/x3.png)
Figure 3.3: Second execution.
Now let us consider the second execution (see figure [3.3](#Ch3.F3 "Figure 3.3 ‣ Chapter 3 Basic Protocol")). In the second execution sensors in $`C`$ and $`\Gamma'`$ are swapped. Thus sensors in $`\Gamma`$ has to be located on the other side of straight line $`l`$ symmetrically. Now virtual faking sensors in $`\Gamma'`$ can imitate the first execution of the correct sensors in $`C`$. Correct sensors in $`C`$ behave like virtual sensors in $`\Gamma'`$ in first execution. This is because the virtual sensors in $`\Gamma'`$ in the first execution behaved like correct sensors and additionally they claimed that sensors from $`C`$ were located in $`C'`$ (see figure [3.3](#Ch3.F3 "Figure 3.3 ‣ Chapter 3 Basic Protocol")). Now $`\Gamma`$ is really located in the previous location of $`C'`$ and the sensors in $`C`$ are correct. Thus sensors $`v`$ and $`u`$ are not able to distinguish between the first and the second execution. Sensors $`v`$ and $`u`$ will have to decide that $`C`$ is set of faking sensors. This is because $`v`$ and $`u`$ have made such decision in first execution and $`v`$ and $`u`$ is not able to distinguish between these two executions. $`\square`$
## Chapter 4 Protocol based on RSS ranging technique
In this section, we consider that sensors use RSS technique to measure distance. We are assuming that each correct sensor has a fixed common transmission signal strength of $`S_{s}`$. The faking sensors can change their transmission signal strength and send suitable fake position to other sensors. Let $`F`$ be a faking sensor that changes its signal strength $`S_{s}^{'}`$ and sends a suitable fake position $`F'`$ to other correct sensors. Sensor $`v`$ can estimate the distance, $`\hat{d}`$ from the receive signal strength (RSS) by Frii’s transmission equation assuming the common signal strength $`S_{s}`$ has been used, according to the assumption in section [2](#Ch2 "Chapter 2 Technical preliminaries").
<table>
<tbody>
<tr>
<td></td>
<td>d^2=cSsSr⟹d^2=SsSs′d2…(2)superscript^𝑑2𝑐subscript𝑆𝑠subscript𝑆𝑟superscript^𝑑2subscript𝑆𝑠subscriptsuperscript𝑆′𝑠superscript𝑑2…2{\hat{d}}^{2}=c\frac{S_{s}}{S_{r}}\implies{\hat{d}}^{2}=\frac{S_{s}}{S^{{}^{\prime}}_{s}}d^{2}\ldots(2)</td>
<td></td>
</tr>
</tbody>
</table>
where $`c = \left( \frac{\lambda}{4\hspace{0pt}\pi} \right)^{2}`$, $`S_{r} = {c\hspace{0pt}\frac{S_{s}^{'}}{d^{2}}}`$, and $`d`$ is the distance from $`v`$ to the actual position of $`F`$.
We show that Protocol FindMap($`{\lceil\frac{n}{2}\rceil} - 1`$) can be adapted to this model provided that $`{n - 3 - f} > f`$, *i.e.* the total number of faking sensors is at most $`{\lfloor\frac{n}{2}\rfloor} - 2`$ and no four correct sensors are located on a particular circle. In this variant of the protocol, a sensor $`v`$ considers sensor $`u`$ faking if the number of accusations messages for $`u`$ is at least $`{\lceil\frac{n}{2}\rceil} - 1`$.
######
Lemma 1
Let $`F`$ be a faking sensor, and $`P_{1}`$ and $`P_{2}`$ be two correct sensors. There exists a position $`(x_{f},y_{f})`$ for $`F`$ such that $`F`$ is always able to fake a position $`F' = {(x_{f}',y_{f}')}`$ to both $`P_{1}`$ and $`P_{2}`$, with $`x_{f} \neq x_{f}'`$, and $`y_{f} \neq y_{f}'`$ by changing its signal strength from S to $`S'`$.
Proof:
[Refer to caption](/html/0710.3824/assets/x4.png)
Figure 4.1: An example showing a faking sensor $`F`$ can supply its suitable false position $`F'`$ to correct sensors $`P_{1}`$ and $`P_{2}`$ by changing its signal strength.
The faking sensor, $`F`$ changes its signal strength from $`S_{s}`$ to $`S_{s}^{'}`$ and sends a corresponding fake position $`(x_{f}',y_{f}')`$ to $`P_{1}`$ and $`P_{2}`$ such that
<table>
<tbody>
<tr>
<td></td>
<td>d1^2=SsSs′d12andd2^2=SsSs′d22formulae-sequencesuperscript^subscript𝑑12subscript𝑆𝑠subscriptsuperscript𝑆′𝑠superscriptsubscript𝑑12andsuperscript^subscript𝑑22subscript𝑆𝑠subscriptsuperscript𝑆′𝑠superscriptsubscript𝑑22{\hat{d_{1}}}^{2}=\frac{S_{s}}{S^{{}^{\prime}}_{s}}{d_{1}}^{2}\leavevmode\nobreak\ \leavevmode\nobreak\ {\rm and}\leavevmode\nobreak\ \leavevmode\nobreak\ {\hat{d_{2}}}^{2}=\frac{S_{s}}{S^{{}^{\prime}}_{s}}{d_{2}}^{2}</td>
<td></td>
</tr>
</tbody>
</table>
Where $`\hat{d_{1}}`$ and $`\hat{d_{2}}`$ are the estimated distances measured by $`P_{1}`$ and $`P_{2}`$ respectively from the RSS of $`F`$ and $`(x_{f}',y_{f}')`$ is the point of intersection of the two circles centering at $`P_{1}`$ and $`P_{2}`$ with radius $`\hat{d_{1}}`$ and $`\hat{d_{2}}`$ respectively according to the figure [4.1](#Ch4.F1 "Figure 4.1 ‣ Chapter 4 Protocol based on RSS ranging technique"), $`d_{1}`$ and $`d_{2}`$ are the distances from the actual position $`(x_{f},y_{f})`$ of $`F`$ to $`P_{1}`$ and $`P_{2}`$ respectively
Then $`P_{1}`$ and $`P_{2}`$ can not able to detect the inconsistency of the fake position $`(x_{f}',y_{f}')`$ of $`F`$ such that $`x_{f} \neq x_{f}'`$, and $`y_{f} \neq y_{f}'`$. $`\square`$
######
Lemma 2
Let $`F`$ be a faking sensor, and $`P_{1}`$ and $`P_{2}`$ be two correct sensors. There exists a position $`(x_{f},y_{f})`$ for $`F`$ such that $`F`$ can always choose a fake position $`F' = {(x_{f}',y_{f}')}`$ for both $`P_{1}`$ and $`P_{2}`$, with $`x_{f} \neq x_{f}'`$, and $`y_{f} \neq y_{f}'`$ by changing its signal strength. Then the possible fake locations for $`F'`$ are placed on a circular arc.
Proof: From lemma [1](#Thmlemma1 "Lemma 1 ‣ Chapter 4 Protocol based on RSS ranging technique") we know that $`{\frac{\hat{d_{1}}}{d_{1}} = {\sqrt{\left( \frac{S_{s}}{S_{s}^{'}} \right)}\quad{and}}}\quad{\frac{\hat{d_{2}}}{d_{2}} = \sqrt{\left( \frac{S_{s}}{S_{s}^{'}} \right)}}`$ that is $`\frac{\hat{d_{1}}}{d_{1}} = \frac{\hat{d_{2}}}{d_{2}}`$ or $`\frac{\hat{d_{1}}}{\hat{d_{2}}} = \frac{d_{1}}{d_{2}}`$ implies $`\frac{\hat{d_{1}}}{\hat{d_{2}}} = \delta`$ where $`\delta = \frac{d_{1}}{d_{2}}`$ = constant, for a pair of sensors $`P_{1}`$ and $`P_{2}`$.
[Refer to caption](/html/0710.3824/assets/x5.png)
Figure 4.2: An example showing possible locations ($`F_{1}\hspace{0pt}F\hspace{0pt}F_{2}`$) of the fake position $`(x_{f}',y_{f}')`$ than can be supplied by faking sensor $`F`$ for a pair correct sensors $`P_{1}`$ and $`P_{2}`$ by changing its signal strength.
If $`(x_{1},y_{1})`$ and $`(x_{2},y_{2})`$ are the coordinates of $`P_{1}`$ and $`P_{2}`$ then the possible location of the $`(x_{f}',y_{f}')`$ is
<table>
<tbody>
<tr>
<td></td>
<td>(x−x1)2+(y−y1)2(x−x2)2+(y−y2)2=δ2superscript𝑥subscript𝑥12superscript𝑦subscript𝑦12superscript𝑥subscript𝑥22superscript𝑦subscript𝑦22superscript𝛿2\frac{(x-x_{1})^{2}+(y-y_{1})^{2}}{(x-x_{2})^{2}+(y-y_{2})^{2}}=\delta^{2}</td>
<td></td>
</tr>
</tbody>
</table>
$`\Longrightarrow{{{x^{2} + y^{2}} - {2\hspace{0pt}\left( \frac{x_{1} - {\delta^{2}\hspace{0pt}x_{2}}}{1 - \delta^{2}} \right)\hspace{0pt}x} - {2\hspace{0pt}\left( \frac{y_{1} - {\delta^{2}\hspace{0pt}y_{2}}}{1 - \delta^{2}} \right)\hspace{0pt}y}} + \frac{{x_{1}^{2} + y_{1}^{2}} - {\delta^{2}\hspace{0pt}{({x_{2}^{2} + y_{2}^{2}})}}}{1 - \delta^{2}}} = 0`$
Which is an equation of circle, where $`\delta = \sqrt{\frac{{({x_{f} - x_{1}})}^{2} + {({y_{f} - y_{1}})}^{2}}{{({x_{f} - x_{2}})}^{2} + {({y_{f} - y_{2}})}^{2}}}`$.
Now we have to prove that $`(x_{f}',y_{f}')`$ can lay only on $`F_{1}\hspace{0pt}F\hspace{0pt}F_{2}`$ part of circular arc as shown in figure [4.2](#Ch4.F2 "Figure 4.2 ‣ Chapter 4 Protocol based on RSS ranging technique"). Where $`F_{1}`$ and $`F_{2}`$ are the point of intersection of two circle of transmission range centering at $`P_{1}`$ and $`P_{2}`$ such that at least one of the circles is its maximum transmission range.
We can prove this by contradiction. Suppose, $`(x_{f}',y_{f}')`$ laying on the counterpart of the circular arc $`F_{1}\hspace{0pt}F\hspace{0pt}F_{2}`$. Then it is not possible by $`F`$ to pretend its fake position to $`P_{1}`$ and $`P_{2}`$ simultaneously. Since counterpart of the circular arc $`F_{1}\hspace{0pt}F\hspace{0pt}F_{2}`$ does not belong to the common transmission of $`P_{1}`$ and $`P_{2}`$, hence proved. $`\square`$
######
Lemma 3
Let $`F`$ be a faking sensor, and $`P_{1}`$, $`P_{2}`$, $`P_{3}`$ be three correct sensors on a circle. There exists a position $`(x_{f},y_{f})`$ for $`F`$ and positions $`(x_{1},y_{1})`$, $`(x_{2},y_{2})`$ and $`(x_{3},y_{3})`$ such that $`F`$ is always able to fake a position $`F' = {(x_{f}',y_{f}')}`$ to $`P_{1}`$, $`P_{2}`$ and $`P_{3}`$ such that $`x_{f} \neq x_{f}'`$, and $`y_{f} \neq y_{f}'`$.
Proof:
[Refer to caption](/html/0710.3824/assets/x6.png)
[Refer to caption](/html/0710.3824/assets/x7.png)
Figure 4.3: An example showing a faking sensor $`F`$ can lie about its position by changing signal strength to three correct sensors.
From Lemma [1](#Thmlemma1 "Lemma 1 ‣ Chapter 4 Protocol based on RSS ranging technique") and [2](#Thmlemma2 "Lemma 2 ‣ Chapter 4 Protocol based on RSS ranging technique"), faking sensor $`F = {(x_{f},y_{f})}`$ can fake its position $`F' = {(x_{f}',y_{f}')}`$ to two correct sensors $`P_{1}`$, $`P_{2}`$ by changing its signal strength from $`S_{s}`$ to $`S_{s}'`$ such that $`{P_{1}\hspace{0pt}F'}:{{P_{1}\hspace{0pt}F} = \lambda}`$ and $`{P_{2}\hspace{0pt}F'}:{{P_{2}\hspace{0pt}F} = \lambda}`$ where $`\lambda = \sqrt{\frac{S_{s}}{S_{s}^{'}}}`$ and $`{P_{1}\hspace{0pt}F'} = \hat{d_{1}}`$, $`{P_{1}\hspace{0pt}F} = d_{1}`$, $`{P_{2}\hspace{0pt}F'} = \hat{d_{2}}`$, $`{P_{2}\hspace{0pt}F} = d_{2}`$.
We have to prove that there exist a sensor $`P_{3}`$ with coordinates ($`x_{3},y_{3}`$) such that $`P_{3}`$ can not able to detect the inconsistency of fake position $`(x_{f}',y_{f}')`$, i.e., $`P_{3}`$ has to locate at a position like $`P_{1}`$ and $`P_{2}`$ such that $`{P_{3}\hspace{0pt}F'}:{{P_{3}\hspace{0pt}F} = \lambda}`$ as shown in figure [4.3](#Ch4.F3 "Figure 4.3 ‣ Chapter 4 Protocol based on RSS ranging technique"). Therefore $`{F\hspace{0pt}F'}:{{F'\hspace{0pt}P_{3}} = {({1 - \lambda})}}:\lambda`$ Therefore $`{(x_{3},y_{3})} = \left( \frac{x_{f}' - {\lambda\hspace{0pt}x_{f}}}{1 - \lambda},\frac{y_{f}' - {\lambda\hspace{0pt}y_{f}}}{1 - \lambda} \right)`$. From geometry we know that only one circle pass through three fix points, hence proved. $`\square`$
######
Lemma 4
Let $`F`$ be a faking sensor, and $`P_{1}`$, $`P_{2}`$ be correct sensors. There exists a position $`(x_{f},y_{f})`$ for $`F`$ and positions $`(x_{1},y_{1})`$, $`(x_{2},y_{2})`$ such that $`F`$ is always able to fake a position $`F' = {(x_{f}',y_{f}')}`$ to $`P_{1}`$ and $`P_{2}`$ such that $`x_{f} \neq x_{f}'`$, and $`y_{f} \neq y_{f}'`$. Then $`F`$ also can fake the position $`(x_{f}',y_{f}')`$ to more $`P_{i}`$’s if and only if they lay on a particular circle.
Proof: Lemma [2](#Thmlemma2 "Lemma 2 ‣ Chapter 4 Protocol based on RSS ranging technique") implies that faking sensor $`F`$ can fix a fake position $`F'`$ on the circular arc $`F_{1}\hspace{0pt}F\hspace{0pt}F_{2}`$ with a suitable changed signal strength ($`S'`$) such that $`P_{1}`$ and $`P_{2}`$ can not able to detect the inconsistency as shown in figure [4.4](#Ch4.F4 "Figure 4.4 ‣ Chapter 4 Protocol based on RSS ranging technique").
[Refer to caption](/html/0710.3824/assets/x8.png)
Figure 4.4: An example showing a faking sensor $`F`$ can lie about its position by changing signal strength to multiple number of correct sensors which are laying on a particular circle.
Let $`P`$ is a variable point such that it keeps the same ratio $`\left. \sqrt{}\frac{S_{s}}{S_{s}^{'}}{( = \lambda} \right.`$) like $`P_{1}`$ and $`P_{2}`$ with $`F`$ and $`F'`$. Then $`P`$ also can not able to detect the inconsistency of the fake position $`F'`$. If $`\hat{d_{p}}`$ is the distance between $`P`$ and $`F'`$ and $`d_{p}`$ is the distance between $`P`$ and $`F`$ then $`\frac{\hat{d_{p}}}{d_{p}} = \lambda`$
Therefore the possible location of the point $`P`$ is $`\frac{{({x - x_{f}'})}^{2} + {({y - y_{f}'})}^{2}}{{({x - x_{f}})}^{2} + {({y - y_{f}})}^{2}} = \lambda^{2}`$
$`\Longrightarrow{{{x^{2} + y^{2}} - {2\hspace{0pt}\left( \frac{x_{f}' - {\lambda^{2}\hspace{0pt}x_{f}}}{1 - \lambda^{2}} \right)\hspace{0pt}x} - {2\hspace{0pt}\left( \frac{y_{f}' - {\lambda^{2}\hspace{0pt}y_{f}}}{1 - \lambda^{2}} \right)\hspace{0pt}y}} + \frac{{x_{f}^{'2} + y_{f}^{'2}} - {\lambda^{2}\hspace{0pt}{({x_{f}^{2} + y_{f}^{2}})}}}{1 - \lambda^{2}}} = 0`$
This is an equation of circle with respect to the given fake position $`F'`$ of $`F`$ and $`P_{1}`$ and $`P_{2}`$ as shown in figure [4.4](#Ch4.F4 "Figure 4.4 ‣ Chapter 4 Protocol based on RSS ranging technique"). Therefore, $`F`$ pretends the fake position $`F'`$ to the sensors which are laying only on the particular circle. $`\square`$
######
Theorem 3
Let $`F`$ be a faking sensor, and $`P_{1}`$, $`P_{2}`$, $`P_{3}`$ be three correct sensors on a circle. If there exist a sensor $`P_{4}`$ which does not lay on the same circle, $`P_{4}`$ is able to detect the inconsistency of $`F`$.
Proof: From lemma [3](#Thmlemma3 "Lemma 3 ‣ Chapter 4 Protocol based on RSS ranging technique") faking sensor $`F`$ can convey the fake position $`F'`$ to $`P_{1}`$, $`P_{2}`$, $`P_{3}`$, provided circles with radius $`\hat{d_{1}} = {\lambda\hspace{0pt}d_{1}}`$, $`\hat{d_{2}} = {\lambda\hspace{0pt}d_{2}}`$, and $`\hat{d_{3}} = {\lambda\hspace{0pt}d_{3}}`$ centering at $`P_{1}`$, $`P_{1}`$, and $`P_{1}`$ respectively intersect at $`F'`$, where $`\lambda = \sqrt{\left( \frac{S_{s}}{S_{s}^{'}} \right)}`$.
[Refer to caption](/html/0710.3824/assets/x9.png)
Figure 4.5: An example showing a that if four sensors $`P_{1}`$, $`P_{2}`$, $`P_{3}`$, $`P_{4}`$ do not lay in a particular circle then faking sensor $`F`$ can be detected by sensor $`P_{4}`$ which is not laying on the circle.
As $`P_{4}`$ not on the circle then $`\hat{d_{4}} \neq {\lambda\hspace{0pt}d_{4}}`$ as in figure [4.5](#Ch4.F5 "Figure 4.5 ‣ Chapter 4 Protocol based on RSS ranging technique") implies $`\hat{d_{4}} \neq {d\hspace{0pt}{(P_{4},F')}}`$, where $`d\hspace{0pt}{(P_{4},F')}`$ is the distance from $`P_{4}`$ to $`F'`$ calculated from coordinates of $`F'`$. Hence $`P_{4}`$ can able to detect the inconsistency of faking node $`F`$. $`\square`$
######
Corollary 1
The protocol FindMap($`{\lceil\frac{n}{2}\rceil} - 1`$) identifies all faking sensors in the model where faking sensors can corrupt RSS ranging technique, provided that $`{n - f - 3} > f`$ and no four sensors are located on the same circle and no three sensors are co-linear.
Proof: Let us consider a faking sensor $`F`$, which fakes its transmission power. Theorem [3](#Thmtheorem3 "Theorem 3 ‣ Chapter 4 Protocol based on RSS ranging technique") guarantees that in each set of four correct sensors there exists a sensor, which detects inconsistency on distance to $`F`$. Thus each faking sensor will be accused by at least $`n - f - 3`$ correct sensors. By inequality $`{n - f - 3} > f`$ the number of correct sensors that accuse $`F`$ is at least $`{\lceil\frac{n}{2}\rceil} - 1`$ and the number of faking sensors is at most $`{\lfloor\frac{n}{2}\rfloor} - 2`$. Thus each faking sensor will be found faking and no correct sensor will be found faking. If faking node $`F`$ does not change its transmission power but only lies about its position then at least one on three no-linear correct sensors will detect inconsistency. $`\square`$
Theorem [3](#Thmtheorem3 "Theorem 3 ‣ Chapter 4 Protocol based on RSS ranging technique") can be also applied in the protocol for the model of trusted sensors. In the protocol presented in \[[3](#bib.bib3)\], we can use theorem [3](#Thmtheorem3 "Theorem 3 ‣ Chapter 4 Protocol based on RSS ranging technique") to find deployment of the minimum number of hidden stations required to detect faking nodes.
######
Corollary 2
If the four hidden stations are not located on the same circle and no three stations are co-linear then one of the stations will always detect a faking node.
Corollary [2](#Thmcorollary2 "Corollary 2 ‣ Chapter 4 Protocol based on RSS ranging technique") remains true provided the faking node’s transmission reaches all hidden stations and it is not allowed to use directional antennas. Since the verifiers are hidden to the faking node in the model of \[[3](#bib.bib3)\], the latter has very low chances to consistently fake its position even with directional antennas.
## Chapter 5 Protocol based on ToF-like ranging techniques
In this chapter, we first discuss how faking sensors can corrupt the two SToF and DAT ranging techniques:
1. 1\.
In case the *SToF* ranging technique is used by Sensor $`u`$, $`u`$ first transmits a message attaching the time of transmission $`t_{s}`$ into the message. Sensor $`v`$, which receives the message from sensor $`u`$ at time $`t_{r}`$, estimates the distance based on delay $`t = {t_{r} - t_{s}}`$ and radio signal propagation speed $`s_{r}`$, $`{\hat{d}\hspace{0pt}{(v,u)}} = {s_{r}\hspace{0pt}t}`$. So, it is possible that a faking sensor can prevent sensor $`v`$ from computing the real distance by faking the transmission time $`t_{s}`$.
2. 2\.
In case the *DAT* ranging technique is used, Sensor $`u`$ transmits each message simultaneously with two signals (*e.g.* RF and US signals). Sensor $`v`$ then records the difference of arrival time $`t`$ between RF signal and US signal. This can be done using only a local clock at $`v`$. Thus no global time is required. Then, Sensor $`v`$ computes distance $`\hat{d}\hspace{0pt}{(v,u)}`$ based on $`t`$, propagation speed $`s_{r}`$ of RF signal and propagation speed $`s_{u}`$ of US signal. In this case, a faking sensor may prevent a correct sensor $`v`$ from computing real distance by delaying one of the two simultaneous transmissions.
Now we show that corrupting SToF and DAT ranging technique has the same affect on correct sensors.
######
Lemma 5
If the ranging is evaluated with SToF technique and faking sensor $`F`$ shifts real transmission time then all correct sensors compute the real distance to sensor $`F`$ increased or decreased by the same length $`b`$.
Proof: Let us assume that faked sensor $`F`$ shifts its real transmission time by $`t'`$. Then all the correct sensors will compute the distance modified by $`b = {s_{r}\hspace{0pt}t'}`$, where $`s_{r}`$ is the radio signal propagation speed. $`\square`$
######
Lemma 6
If the ranging is evaluated with DAT technique and faking sensor $`F`$ introduces shift $`t' \neq 0`$ between the RF and US transmissions, then all correct sensors compute the real distance to the sensor $`F`$ increased or decreased by the same length $`b`$.
Proof: Since the faking sensor shifts the two transmissions by time $`t'`$ then the difference in arrivals time of the signals will be $`t + t'`$ where $`t`$ is original difference for $`t' = 0`$. Each correct sensor will compute $`\hat{d}`$ based on the following equation.
<table>
<tbody>
<tr>
<td></td>
<td>t+t′=d^sr−d^su𝑡superscript𝑡′^𝑑subscript𝑠𝑟^𝑑subscript𝑠𝑢t+t^{\prime}=\frac{\hat{d}}{s_{r}}-\frac{\hat{d}}{s_{u}}</td>
<td></td>
</tr>
</tbody>
</table>
Thus the real distance will be modified by
<table>
<tbody>
<tr>
<td></td>
<td>b=t′1/sr−1/su𝑏superscript𝑡′1subscript𝑠𝑟1subscript𝑠𝑢b=\frac{t^{\prime}}{1/s_{r}-1/s_{u}}</td>
<td></td>
</tr>
</tbody>
</table>
in all correct sensors. $`\square`$
Since the corruption on SToF and DAT has the same result we can formulate the following theorem for both ranging techniques.
######
Theorem 4
If the distance evaluation is done with SToF or DAT techniques and no six sensors are located on the same hyperbola and no three sensors are co-linear, then at least one of six correct sensors detects inconsistency in faked transmission.
Proof:
Let us assume that faking sensor $`F`$ enlarges its distance against the correct sensors by $`b`$. The case when sensor reduces its distance is symmetric. By lemma [5](#Thmlemma5 "Lemma 5 ‣ Chapter 5 Protocol based on ToF-like ranging techniques") and [6](#Thmlemma6 "Lemma 6 ‣ Chapter 5 Protocol based on ToF-like ranging techniques") there are at most two faked locations $`F'`$ and $`F^{\operatorname{\prime\prime}}`$ for faking sensor $`F`$, which guarantee consistency against sensors $`P_{1}`$ and $`P_{2}`$ (see figure [5.1](#Ch5.F1 "Figure 5.1 ‣ Chapter 5 Protocol based on ToF-like ranging techniques")). Let us assume that sensor $`F`$ decides for faked location $`F'`$.
[Refer to caption](/html/0710.3824/assets/x10.png)
Figure 5.1: Figure shows that sensor $`F`$ can change its position to $`F'`$ and consistently lie against sensor $`P_{3}`$ which is located in the middle of segment $`F\hspace{0pt}B`$. Length of segment $`F'\hspace{0pt}B`$ is $`b`$.
Now we will find the set of correct sensors, which will not detect the inconsistency. We consider two cases:
1. 1\.
The first case is when distance $`c`$ between $`F'`$ and $`F`$ is strictly larger than $`b`$ (see figure [5.2](#Ch5.F2 "Figure 5.2 ‣ item 1 ‣ Chapter 5 Protocol based on ToF-like ranging techniques")). Each correct sensors $`P`$, which cannot detect inconsistency on distance to $`F`$, has to meet $`{d\hspace{0pt}{(P,F)}} = {\hat{d}\hspace{0pt}{(P,F)}}`$. The condition $`{d\hspace{0pt}{(P,F)}} = {\hat{d}\hspace{0pt}{(P,F)}}`$ can be transformed into the distances on the plane $`{|{F'\hspace{0pt}P}|} = {{|{F\hspace{0pt}P}|} + b}`$. Based on this condition we can came up with system of equations for sensors in $`S = {\{ P:{{d\hspace{0pt}{(P,F)}} = {\hat{d}\hspace{0pt}{(P,F)}}}\}}`$.
[Refer to caption](/html/0710.3824/assets/x11.png)
Figure 5.2: We assume that $`{|{F\hspace{0pt}F'}|} > b`$. Figure shows set $`S`$ of correct sensors located on the hyperbola, which cannot detect inconsistency. That is for each correct sensor $`P`$ located on the hyperbola the distance $`|{F'\hspace{0pt}P}|`$ is equal to $`{|{F\hspace{0pt}P}|} + b`$
<table>
<tbody>
<tr>
<td></td>
<td>x2+y2superscript𝑥2superscript𝑦2\displaystyle x^{2}+y^{2}</td>
<td>=\displaystyle=</td>
<td>z2superscript𝑧2\displaystyle z^{2}</td>
<td></td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td>x2+(y−c)2superscript𝑥2superscript𝑦𝑐2\displaystyle x^{2}+(y-c)^{2}</td>
<td>=\displaystyle=</td>
<td>(z+b)2superscript𝑧𝑏2\displaystyle(z+b)^{2}</td>
<td></td>
</tr>
</tbody>
</table>
Where $`{|{F\hspace{0pt}P}|} = z`$, $`x,y`$ are the coordinates of correct sensor $`P \in S`$. We assume that $`F = {(0,0)}`$ and $`F' = {(0,c)}`$. Next we can find the equation of the hyperbola.
<table>
<tbody>
<tr>
<td></td>
<td>x2+(y−c)2superscript𝑥2superscript𝑦𝑐2\displaystyle x^{2}+(y-c)^{2}</td>
<td>=\displaystyle=</td>
<td>(x2+y2+b)2superscriptsuperscript𝑥2superscript𝑦2𝑏2\displaystyle(\sqrt{x^{2}+y^{2}}+b)^{2}</td>
<td></td>
<td></td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td>x2+y2−2yc+c2superscript𝑥2superscript𝑦22𝑦𝑐superscript𝑐2\displaystyle x^{2}+y^{2}-2yc+c^{2}</td>
<td>=\displaystyle=</td>
<td>x2+y2+2bx2+y2+b2superscript𝑥2superscript𝑦22𝑏superscript𝑥2superscript𝑦2superscript𝑏2\displaystyle x^{2}+y^{2}+2b\sqrt{x^{2}+y^{2}}+b^{2}</td>
<td></td>
<td></td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td>(−2yc+c2−b2)2superscript2𝑦𝑐superscript𝑐2superscript𝑏22\displaystyle(-2yc+c^{2}-b^{2})^{2}</td>
<td>=\displaystyle=</td>
<td>4b2(x2+y2)4superscript𝑏2superscript𝑥2superscript𝑦2\displaystyle 4b^{2}(x^{2}+y^{2})</td>
<td></td>
<td></td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td>4y2c2−4yc(c2−b2)+(c2−b2)24superscript𝑦2superscript𝑐24𝑦𝑐superscript𝑐2superscript𝑏2superscriptsuperscript𝑐2superscript𝑏22\displaystyle 4y^{2}c^{2}-4yc(c^{2}-b^{2})+(c^{2}-b^{2})^{2}</td>
<td>=\displaystyle=</td>
<td>4b2(x2+y2)4superscript𝑏2superscript𝑥2superscript𝑦2\displaystyle 4b^{2}(x^{2}+y^{2})</td>
<td></td>
<td></td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td>4y2c2−4yc(c2−b2)+(c2−b2)2−4b2y24superscript𝑦2superscript𝑐24𝑦𝑐superscript𝑐2superscript𝑏2superscriptsuperscript𝑐2superscript𝑏224superscript𝑏2superscript𝑦2\displaystyle 4y^{2}c^{2}-4yc(c^{2}-b^{2})+(c^{2}-b^{2})^{2}-4b^{2}y^{2}</td>
<td>=\displaystyle=</td>
<td>4b2x24superscript𝑏2superscript𝑥2\displaystyle 4b^{2}x^{2}</td>
<td></td>
<td></td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td>4(c2−b2)y2−4c(c2−b2)y+(c2−b2)24superscript𝑐2superscript𝑏2superscript𝑦24𝑐superscript𝑐2superscript𝑏2𝑦superscriptsuperscript𝑐2superscript𝑏22\displaystyle 4(c^{2}-b^{2})y^{2}-4c(c^{2}-b^{2})y+(c^{2}-b^{2})^{2}</td>
<td>=\displaystyle=</td>
<td>4b2x24superscript𝑏2superscript𝑥2\displaystyle 4b^{2}x^{2}</td>
<td></td>
<td></td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td>(c2−b2)(4y2−4cy+c2−b2)superscript𝑐2superscript𝑏24superscript𝑦24𝑐𝑦superscript𝑐2superscript𝑏2\displaystyle(c^{2}-b^{2})(4y^{2}-4cy+c^{2}-b^{2})</td>
<td>=\displaystyle=</td>
<td>4b2x24superscript𝑏2superscript𝑥2\displaystyle 4b^{2}x^{2}</td>
<td></td>
<td></td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td>(c2−b2)((2y−c)2−b2)superscript𝑐2superscript𝑏2superscript2𝑦𝑐2superscript𝑏2\displaystyle(c^{2}-b^{2})((2y-c)^{2}-b^{2})</td>
<td>=\displaystyle=</td>
<td>4b2x24superscript𝑏2superscript𝑥2\displaystyle 4b^{2}x^{2}</td>
<td></td>
<td></td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td>(c2−b2)(2y−c)2−b2(c2−b2)superscript𝑐2superscript𝑏2superscript2𝑦𝑐2superscript𝑏2superscript𝑐2superscript𝑏2\displaystyle(c^{2}-b^{2})(2y-c)^{2}-b^{2}(c^{2}-b^{2})</td>
<td>=\displaystyle=</td>
<td>4b2x24superscript𝑏2superscript𝑥2\displaystyle 4b^{2}x^{2}</td>
<td></td>
<td></td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td>(c2−b2)(2y−c)2−4b2x2superscript𝑐2superscript𝑏2superscript2𝑦𝑐24superscript𝑏2superscript𝑥2\displaystyle(c^{2}-b^{2})(2y-c)^{2}-4b^{2}x^{2}</td>
<td>=\displaystyle=</td>
<td>b2(c2−b2)superscript𝑏2superscript𝑐2superscript𝑏2\displaystyle b^{2}(c^{2}-b^{2})</td>
<td></td>
<td></td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td>(c2−b2)(2y−c)2−4b2x2superscript𝑐2superscript𝑏2superscript2𝑦𝑐24superscript𝑏2superscript𝑥2\displaystyle(c^{2}-b^{2})(2y-c)^{2}-4b^{2}x^{2}</td>
<td>=\displaystyle=</td>
<td>b2(c2−b2)superscript𝑏2superscript𝑐2superscript𝑏2\displaystyle b^{2}(c^{2}-b^{2})</td>
<td></td>
<td></td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td>(2y−c)2b2−4x2c2−b2superscript2𝑦𝑐2superscript𝑏24superscript𝑥2superscript𝑐2superscript𝑏2\displaystyle\frac{(2y-c)^{2}}{b^{2}}-\frac{4x^{2}}{c^{2}-b^{2}}</td>
<td>=\displaystyle=</td>
<td>11\displaystyle 1</td>
<td></td>
<td>(5.2)</td>
</tr>
</tbody>
</table>
The five sensors uniquely determine the hyperbola. Thus the sixth sensor, which is not located on the hyperbola by our assumption, will detect inconsistency.
2. 2\.
The second case is when distance $`c`$ between $`F'`$ and $`F`$ is at most $`b`$ (see figure [5.3](#Ch5.F3 "Figure 5.3 ‣ item 2 ‣ Chapter 5 Protocol based on ToF-like ranging techniques")). We will show that $`P_{1}`$ or $`P_{2}`$ will have to detect inconsistency. The distance measured using coordinates by $`P_{i}`$ for $`i = {1,2}`$ has to be exactly $`{|{F\hspace{0pt}P_{i}}|} + b`$ to prevent sensor $`P_{i}`$ from detecting inconsistency. By triangle inequality we have $`{{|{F'\hspace{0pt}F}|} + {|{F\hspace{0pt}P_{i}}|}} \geq {|{F'\hspace{0pt}P_{i}}|}`$ for $`i = {1,2}`$. Thus the distance $`|{F'\hspace{0pt}P_{i}}|`$ measured by $`P_{i}`$ with a ranging technique is at most $`{|{F\hspace{0pt}P_{i}}|} + b`$. Sensor $`P_{i}`$ for $`i = {1,2}`$ will measure required distance when sensors $`F'`$, $`F`$ and $`P_{i}`$ are co-linear. This will happen for at most one sensor. This is because we assume that no three sensors are co-linear.
[Refer to caption](/html/0710.3824/assets/x12.png)
Figure 5.3: We assume $`{F\hspace{0pt}F'} \leq b`$ Figure shows that faking sensor $`F`$ cannot change its position to $`F'`$ consistently against sensors $`P_{1}`$ and $`P_{2}`$. That is $`{F'\hspace{0pt}P_{1}} < {{|{F\hspace{0pt}P_{1}}|} + b}`$ or $`{F'\hspace{0pt}P_{2}} < {{|{F\hspace{0pt}P_{2}}|} + b}`$ allowing sensor $`P_{1}`$ or $`P_{2}`$ to detect inconsistency.
$`\square`$
Theorem [4](#Thmtheorem4 "Theorem 4 ‣ Chapter 5 Protocol based on ToF-like ranging techniques") allows us to modify the protocol FindMap so that it works in the model in which faking sensors can corrupt the SToF or DAT ranging technique.
######
Corollary 3
The protocol FindMap($`{\lceil\frac{n}{2}\rceil} - 2`$) identifies all faking sensors, in the model where faking sensors can corrupt SToF or DAT ranging techniques, provided $`{n - f - 5} > f`$ and no six sensors are located on the same hyperbola and no three sensors are co-linear.
Proof: Let us consider a faking sensor $`F`$. Theorem [4](#Thmtheorem4 "Theorem 4 ‣ Chapter 5 Protocol based on ToF-like ranging techniques") guarantees that in each set of six correct sensors there exists a sensor which detects inconsistency on distance to $`F`$. Thus each faking sensor will be accused by at least correct $`n - f - 5`$ sensors. By inequality $`{n - f - 5} > f`$ the number of correct sensors that accuse $`F`$ is at least $`{\lceil\frac{n}{2}\rceil} - 2`$ and the number of faking sensors is at most $`{\lfloor\frac{n}{2}\rfloor} - 3`$. Thus each faking sensor will be find faking and no correct sensor will be found faking. $`\square`$
Theorem [4](#Thmtheorem4 "Theorem 4 ‣ Chapter 5 Protocol based on ToF-like ranging techniques") can be also applied in the protocol for the model of trusted sensors \[[3](#bib.bib3)\]. We can use theorem [4](#Thmtheorem4 "Theorem 4 ‣ Chapter 5 Protocol based on ToF-like ranging techniques") to compute the deployment of the minimum number of hidden stations required to detect faking nodes.
######
Corollary 4
If the six hidden stations are not located on the same hyperbola and no three stations are co-linear then one of the stations always detect a faking node.
Corollary [4](#Thmcorollary4 "Corollary 4 ‣ Chapter 5 Protocol based on ToF-like ranging techniques") is true provided the attacker’s transmission reaches all the hidden stations and attacker is not allowed to use directional antennas. Since the verifiers are hidden to the faking node, the latter has very low chance to consistently fake its position even with directional antennas.
## Chapter 6 Concluding Remarks
We proposed a secure positioning deterministic protocol for WSN that performs in the most general NTS model. Although the previous protocol of Hwang *et al.* \[[13](#bib.bib13)\] is probabilistic (and thus, unlike ours, can not give *certain* results), it is interesting to see if the certainty of the result comes with a price (with respect to the number of exchanged messages to solve the problem). In \[[13](#bib.bib13)\], each sensor announces one distance at a time in a round robin fashion (otherwise the faking node could hold its own announcement, collect all correct nodes informations, and send a consistent range claim), inducing $`n\hspace{0pt}{({n - 1})}`$ sent messages, an overall $`O\hspace{0pt}{(n^{2})}`$ message complexity. In our case, $`n`$ coordinate messages are sent in round one, and $`n`$ accusation messages are sent in round two, overall a $`O\hspace{0pt}{(n)}`$ message complexity. However, from a information complexity point of view, the two approaches are equivalent, since the exchanged messages in our protocol can be $`n`$-sized (inducing $`n^{2}`$ information in both cases).
To conclude, we would like to mention two interesting open questions:
1. 1\.
Our protocol makes some synchrony hypotheses to separate between rounds and filter faking nodes. It is worth investigating to determine the exact model assumptions that are necessary and sufficient to solve the same problem in the NTS model with respect to synchrony.
2. 2\.
Our network model assumes that correct nodes are within range of every other node. Extending our result to WSN with fixed ranges for every node is a challenging task, especially since previous results on networks facing intermittent failures and attacks \[[6](#bib.bib6), [7](#bib.bib7), [18](#bib.bib18)\] are written for rather stronger models (*i.e.* wired secure communications) than that of this paper.
## References
- \[1\] P. Bahl and V. N. Padmanabhan. Radar: An in-building rf-based user location and tracking system.
In INFOCOM, volume 2, pages 775–784. IEEE, 2000.
- \[2\] S. Brands and D. Chaum. Distance-bounding protocols.
In EUROCRYPT ’93: Workshop on the theory and application of cryptographic techniques on Advances in cryptology, pages 344–359, Secaucus, NJ, USA, 1994. Springer-Verlag New York, Inc.
- \[3\] S. Capkun, M. Cagalj, and M. B. Srivastava. Secure localization with hidden and mobile base stations.
In INFOCOM. IEEE, 2006.
- \[4\] S. Capkun and J. Hubaux. Secure positioning in wireless networks.
IEEE Journal on Selected Areas in Communications: Special Issue on Security in Wireless Ad Hoc Networks, 24(2):221–232, 2006.
- \[5\] Sylvie Delaët, Partha Sarathi Mandal, Mariusz Rokicki, and Sébastien Tixeuil. Deterministic secure positionning in wireless sensor networks. Technical report, INRIA, October 2007.
- \[6\] Sylvie Delaët and Sébastien Tixeuil. Tolerating transient and intermittent failures.
J. Parallel Distrib. Comput., 62(5):961–981, 2002.
- \[7\] Sylvie Delaët, Bertrand Ducourthial, and Sébastien Tixeuil. Self-stabilization with r-operators revisited.
Journal of Aerospace Computing, Information, and Communication, 2006.
- \[8\] J. R. Douceur. The sybil attack.
In Peter Druschel, M. Frans Kaashoek, and Antony I. T. Rowstron, editors, IPTPS ’01: International Workshop on Peer-to-Peer Systems, volume 2429 of Lecture Notes in Computer Science, pages 251–260, London, UK, 2002. Springer-Verlag.
- \[9\] R. J. Fontana, E. Richley, and J. Barney. Commercialization of an ultra wideband precision asset location system. pages 369–373, 2003.
- \[10\] T. He, S. Krishnamurthy, J. A. Stankovic, T. Abdelzaher, L. Luo, R. Stoleru, T. Yan, L. Gu, J. Hui, and B. Krogh. An energy-efficient surveillance system using wireless sensor networks.
In MobiSys ’04: Proceedings of the 2nd Int. Conf. on Mobile systems, applications, and services, pages 270–283, New York, NY, USA, 2004. ACM Press.
- \[11\] J. Hightower, R. Want, and G. Borriello. SpotON: An indoor 3d location sensing technology based on RF signal strength. UW CSE 00-02-02, University of Washington, Department of Computer Science and Engineering, Seattle, WA, February 2000.
- \[12\] Y. Hu, A. Perrig, and D. B. Johnson. Packet leashes: A defense against wormhole attacks in wireless networks.
In INFOCOM. IEEE, 2003.
- \[13\] J. Hwang, T. He, and Y. Kim. Detecting phantom nodes in wireless sensor networks.
In INFOCOM, pages 2391–2395. IEEE, 2007.
- \[14\] B. Karp and H. T. Kung. Gpsr: greedy perimeter stateless routing for wireless networks.
In MobiCom ’00: Proceedings of the 6th Annual Int. Conf. on Mobile Computing and Networking, pages 243–254, New York, NY, USA, 2000. ACM Press.
- \[15\] L. Lazos and R. Poovendran. Serloc: Robust localization for wireless sensor networks.
ACM Trans. Sen. Netw., 1(1):73–100, 2005.
- \[16\] L. Lazos, R. Poovendran, and S. Capkun. Rope: robust position estimation in wireless sensor networks.
In IPSN, pages 324–331. IEEE, 2005.
- \[17\] C. H. Liu and D. J. Fang. Propagation. in antenna handbook: Theory, applications, and design.
Van Nostrand Reinhold, Chapter 29:1–56, 1988.
- \[18\] M. Nesterenko and S. Tixeuil. Discovering network topology in the presence of byzantine faults.
In Paola Flocchini and Leszek Gasieniec, editors, SIROCCO, volume 4056 of Lecture Notes in Computer Science, pages 212–226. Springer, 2006.
- \[19\] N. B. Priyantha, A. Chakraborty, and H. Balakrishnan. The cricket location-support system.
In 6th ACM MOBICOM, Boston, MA, August 2000. ACM.
- \[20\] R. Szewczyk, A. Mainwaring, J. Polastre, J. Anderson, , and D. Culler. An analysis of a large scale habitat monitoring application.
In SenSys ’04: Proceedings of the 2nd Int. Conf. on Embedded Networked Sensor Systems, pages 214–226, New York, NY, USA, 2004. ACM Press.
- \[21\] R. Want, A. Hopper, ao V. Falc and J. Gibbons. The active badge location system.
ACM Trans. Inf. Syst., 10(1):91–102, 1992.
- \[22\] A. Ward, A. Jones, and A. Hopper. A new location technique for the active office.
Personal Communications, IEEE \[see also IEEE Wireless Communications\], 4(5):42–47, 1997. |
0710.3279 | cs.NI | cyber | # Resource Allocation for Delay Differentiated Traffic in Multiuser OFDM Systems
Meixia Tao, Ying-Chang Liang and Fan Zhang This work was presented in part at the IEEE International Conference on Communications, Istanbul, Turkey, June 2006. The corresponding author is Meixia Tao, who is with the department of Electrical and Computer Engineering, National University of Singapore, Singapore 117576. Email: mxtao@nus.edu.sg.
###### Abstract
Most existing work on adaptive allocation of subcarriers and power in multiuser orthogonal frequency division multiplexing (OFDM) systems has focused on homogeneous traffic consisting solely of either delay-constrained data (guaranteed service) or non-delay-constrained data (best-effort service). In this paper, we investigate the resource allocation problem in a heterogeneous multiuser OFDM system with both delay-constrained (DC) and non-delay-constrained (NDC) traffic. The objective is to maximize the sum-rate of all the users with NDC traffic while maintaining guaranteed rates for the users with DC traffic under a total transmit power constraint. Through our analysis we show that the optimal power allocation over subcarriers follows a multi-level water-filling principle; moreover, the valid candidates competing for each subcarrier include only one NDC user but all DC users. By converting this combinatorial problem with exponential complexity into a convex problem or showing that it can be solved in the dual domain, efficient iterative algorithms are proposed to find the optimal solutions. To further reduce the computational cost, a low-complexity suboptimal algorithm is also developed. Numerical studies are conducted to evaluate the performance the proposed algorithms in terms of service outage probability, achievable transmission rate pairs for DC and NDC traffic, and multiuser diversity.
###### Index Terms:
Orthogonal frequency division multiplexing (OFDM), constant-rate transmission, variable-rate transmission, power control, convex optimization, water filling.
## I Introduction
Future broadband wireless networks are expected to support a wide variety of communication services with diverse quality-of-service (QoS) requirements. Applications such as voice transmission and real-time video streaming are very delay-sensitive and need guaranteed throughput. On the other hand, applications like file transfer and email services are relatively delay tolerant so variable-rate transmission is acceptable. From the physical layer point of view, transmission of delay-tolerant or non-delay-constrained (NDC) traffic can be viewed as an ergodic capacity problem \[[1](#bib.bib1)\], where maximizing the long-term average transmission rate is the goal. Thus, wireless resources, such as transmission power and frequency bandwidth, can be dynamically allocated so as to exploit the time or frequency selectivities of broadband wireless fading channels. Likewise, transmission of delay-sensitive or delay-constrained (DC) traffic can be regarded as a delay-limited capacity problem \[[2](#bib.bib2)\] in which a constant transmission rate should be maintained with probability one regardless of channel variations. In this case, it is desirable to allocate more transmission power and frequency bandwidth when the channel experiences deep fade and to allocate less resources when the channel is under favorable conditions. We investigate in this work resource allocation in a broadband wireless network that supports simultaneous transmission of users with delay differentiated traffic. Our focus is on the formulation of an analytical framework from the physical layer perspective as well as the design of efficient and practical algorithms.
Multicarrier transmission in the shape of orthogonal frequency division multiplexing (OFDM) is a leading technique to provide spectrally efficient modulation as well as user multiplexing in future wireless systems. With OFDM technique, the broadband wireless channel is divided into a set of orthogonal narrowband subcarriers. In a single user system, since the channel frequency responses are different at different subcarriers, the system performance can be significantly enhanced by adapting the transmission parameters such as modulation, coding, and power over each subcarrier. For instance, the transmitter can send at higher transmission rates over subcarriers with better channel condition while lower rates or no data over subcarriers in deep fade. This follows the well-known water-filling principle. In a multiuser system, different subcarriers can be allocated to different users to provide a multiple access method, also known as OFDMA. As the channels on each subcarrier are likely independent for different users, the subcarriers experiencing deep fade for one user may not be in deep fade for other users. As a result, each subcarrier could be in good condition for some users in a multiuser OFDM system. By adaptively allocating the subcarriers among multiple users based on instantaneous channel information, multiuser diversity can be utilized to boost the overall system efficiency.
Adaptive resource allocation in multiuser OFDM systems has focused on homogeneous traffic only. In such systems, the traffic consists solely of either DC data requiring constant-rate transmission \[[3](#bib.bib3)\], or variable-rate NDC data which can be served in a best-effort manner \[[4](#bib.bib4), [5](#bib.bib5), [6](#bib.bib6)\]. For systems with pure DC traffic, the problem is to minimize the total transmit power while satisfying a basic transmission rate for each user, which is often referred to as *margin adaption* \[[7](#bib.bib7)\]. In \[[3](#bib.bib3)\], an iterative algorithm was proposed to allocate each user a set of subcarriers and then determine the power and rate for each user on its allocated subcarriers. For systems with pure NDC traffic, the problem is often formulated as maximizing the sum-rate of the system subject to a total transmit power constraint. This formulation is also known as *rate adaptation* \[[7](#bib.bib7)\]. In \[[4](#bib.bib4)\], it was shown that the total sum-rate of a multiuser OFDM system is maximized when each subcarrier is allocated to the user with the best channel gain for that subcarrier. The total transmit power is then distributed over the subcarriers using the water-filling algorithm. This result holds, however, only for single-antenna systems. It is no longer optimal when multiple antennas are deployed at the base station due to the spatial multiplexing gain \[[5](#bib.bib5), [6](#bib.bib6), [8](#bib.bib8)\]. Other problem formulations for systems with pure NDC traffic take user fairness into account. For example, \[[9](#bib.bib9)\] studied the max-min criterion which aims to maximize the transmission rate of the bottleneck user. In \[[10](#bib.bib10)\], it was proposed to maintain proportional rates among users for each channel realization. A utility-function based optimization framework to balance system efficiency and user fairness was also discussed in\[[11](#bib.bib11)\].
In this paper, we consider the subcarrier and power allocation problem in a heterogeneous multiuser OFDM system where DC and NDC traffic is supported simultaneously. Users in the system are classified into DC users and NDC users based on their traffic delay requirements. We assume that the total transmit power from the base station is fixed. Our objective is to maximize the sum-rate of all the NDC users while maintaining the basic transmission rates of all the DC users over every transmission frame. A similar problem was studied in \[[12](#bib.bib12)\]. However, it assumed static subcarrier allocation so only the transmit power adaptation was discussed. Our work, instead, considers joint subcarrier and power adaptation in multiuser OFDM systems, which is one step forward of the previous work. This multiuser subcarrier and power allocation problem is a mixed integer programming problem, the complexity of which increases exponentially with the number of subcarriers. To make the problem more tractable, we transform it into a convex programming problem by introducing time-sharing variables. We show that, for a given subcarrier assignment, the optimal power distribution is achieved by multi-level water-filling. In particular, the water level of each DC user depends explicitly on the channel gains of its assigned subcarriers and its basic rate requirement, and can differ from one another. On the other hand, the water levels of all NDC users are the same. We also show that, for the optimal subcarrier assignment, the set of valid user candidates competing for each subcarrier consists of all the users from the DC group and one from the NDC group with the best channel gain. Using these properties, we propose an efficient iterative algorithm to compute the optimal solution numerically. Alternatively, the original problem is solved in the dual domain by using dual decomposition. It is shown that the dual updates can be done efficiently using an ellipsoid algorithm. In addition, we present a suboptimal algorithm with linear complexity in the number of subcarriers and the number of users.
The rest of this paper is organized as follows. In Section II we introduce the system model and describe the problem formulation. In Section III we formulate the resource allocation problem as a convex optimization problem by using time-sharing technique and present analytical frameworks of the optimal solution. An iterative algorithm to search the optimal solution is also presented. In Section IV, we attempt to solve the problem using dual approach. A low-complexity suboptimal algorithm is given in Section V. In Section VI, we present numerical results of our proposed algorithms in a multiuser OFDM system. Finally conclusion and discussions are given in Section VII.
## II System Model and Problem Formulation
We consider the downlink of a multiuser OFDM system with block diagram shown in Fig. [1](#A0.F1 "Figure 1 ‣ Resource Allocation for Delay Differentiated Traffic in Multiuser OFDM Systems"). The system consists of $`K`$ mobile users. The first $`K_{1}`$ users have DC traffic, which requires a constant transmission rate of $`R_{k}`$ ($`k = {1,\ldots,K_{1}}`$) bits per OFDM symbol, respectively. The traffic of the remaining $`K - K_{1}`$ users has no delay constraint and can be delivered in a best-effort manner. Note that $`K`$ is the number of users that are scheduled for transmission during a certain transmission interval. The total number of users in a practical system may be much larger than $`K`$ and, hence, other multiple access techniques such as time-division multiple access (TDMA) are needed in conjunction with OFDMA. The data streams from the $`K`$ users are serially fed into the encoder block at the base station transmitter. The total channel bandwidth is $`B`$ Hz and is divided into $`N`$ orthogonal subcarriers, which are shared among the $`K`$ users. The transmission is on a time-frame basis, where each frame consists of multiple OFDM symbols. The fading coefficients of all users are assumed to remain unchanged within each transmission frame but can vary from one frame to another. All channel information is assumed perfectly known at the central controller, which can be embedded with the base station. Typically, the channel information can be collected by estimating it at each user terminal and sending it to the base station via a feedback channel, or through channel estimation of the uplink in a time-division duplex system. Based on the instantaneous channel inputs, the central controller allocates different subcarriers to different users and determines the amount of power/bits to be transmitted on each subcarrier through the subcarrier and power/bit allocation algorithm. The resulting allocation information is used to configure the encoder block at the base station transmitter and to facilitate the subcarrier selector and decoder at each user receiver. Note that this allocation information may be sent to each user via a separate channel. The output data symbols from the encoder are then modulated by the inverse fast Fourier transform (IFFT). Guard interval is inserted to ensure orthogonality between OFDM symbols. The total transmit power from the base station is fixed and is given by $`P_{T}`$.
The broadband wireless channel between the base station and each user terminal is assumed to be frequency-selective Rayleigh fading. However, the channel in each subcarrier is narrow enough to experience flat fading. Let $`r_{k,n}`$ denote the transmission rate of user $`k`$ on subcarrier $`n`$ in bits per OFDM symbol. It depends on the channel gain $`h_{k,n}`$ and the allocated power $`P_{k,n}`$ of user $`k`$ on subcarrier $`n`$. In general, $`r_{k,n}`$ can be expressed as
<table>
<tbody>
<tr>
<td></td>
<td>rk,n=log2(1+Pk,n|hk,n|2ΓN0B/N),subscript𝑟𝑘𝑛subscript21subscript𝑃𝑘𝑛superscriptsubscriptℎ𝑘𝑛2Γsubscript𝑁0𝐵𝑁r_{k,n}=\log_{2}{\left(1+\frac{P_{k,n}|h_{k,n}|^{2}}{\Gamma N_{0}B/N}\right)},</td>
<td></td>
<td>(1)</td>
</tr>
</tbody>
</table>
where $`N_{0}`$ is the power spectral density of additive white Gaussian noise and $`\Gamma`$ is a constant, usually called the signal-to-noise ratio (SNR) gap \[[7](#bib.bib7)\]. When instantaneous mutual information is used to characterize the achievable transmission rate, we have $`\Gamma = 1`$ (0 dB). If practical signal constellations are used, $`\Gamma`$ is a constant related to a given bit-error-rate (BER) requirement. For example, when uncoded QAM modulation is used we have $`\Gamma = {- {{\ln{({5 \cdot {BER}})}}/1.5}}`$. In general, the gap serves as a convenient mechanism for analyzing the difference between the SNR needed to achieve a certain data rate for a practical system and the theoretical limit. Throughout this paper we use ([1](#S2.E1 "In II System Model and Problem Formulation ‣ Resource Allocation for Delay Differentiated Traffic in Multiuser OFDM Systems")) as a unified form to characterize both the theoretical mutual information and practical transmission rate.
The problem we consider here is to optimize the allocation of subcarriers and power under the total transmit power constraint so as to maximize the sum-rate of all the $`K - K_{1}`$ NDC users while satisfying the individual rate requirement for each of the $`K_{1}`$ DC users. Mathematically, the given problem can be formulated as
<table>
<tbody>
<tr>
<td></td>
<td>max{Ωk,Pk,n}subscriptsubscriptΩ𝑘subscript𝑃𝑘𝑛\displaystyle\max_{\{\Omega_{k},P_{k,n}\}}</td>
<td></td>
<td>∑k=K1+1K∑n∈Ωkrk,nsuperscriptsubscript𝑘subscript𝐾11𝐾subscript𝑛subscriptΩ𝑘subscript𝑟𝑘𝑛\displaystyle\sum_{k=K_{1}+1}^{K}\sum_{n\in\Omega_{k}}{r_{k,n}}</td>
<td></td>
<td>(2)</td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td>subjecttosubjectto\displaystyle{\rm subject~{}to}</td>
<td></td>
<td>∑n∈Ωkrk,n≥Rk,k=1,…,K1formulae-sequencesubscript𝑛subscriptΩ𝑘subscript𝑟𝑘𝑛subscript𝑅𝑘𝑘1…subscript𝐾1\displaystyle\sum_{n\in\Omega_{k}}r_{k,n}\geq R_{k},~{}k=1,\ldots,K_{1}</td>
<td></td>
<td></td>
</tr>
<tr>
<td></td>
<td></td>
<td></td>
<td>∑k=1K∑n∈ΩkPk,n=PTsuperscriptsubscript𝑘1𝐾subscript𝑛subscriptΩ𝑘subscript𝑃𝑘𝑛subscript𝑃𝑇\displaystyle\sum_{k=1}^{K}\sum_{n\in\Omega_{k}}{P_{k,n}}=P_{T}</td>
<td></td>
<td></td>
</tr>
<tr>
<td></td>
<td></td>
<td></td>
<td>Pk,n≥0,∀k,nsubscript𝑃𝑘𝑛0for-all𝑘𝑛\displaystyle P_{k,n}\geq 0,\forall k,n</td>
<td></td>
<td></td>
</tr>
<tr>
<td></td>
<td></td>
<td></td>
<td>Ω1∪Ω2∪⋯∪ΩK⊆{1,2,…,N}subscriptΩ1subscriptΩ2⋯subscriptΩ𝐾12…𝑁\displaystyle\Omega_{1}\cup\Omega_{2}\cup\cdots\cup\Omega_{K}\subseteq\{1,2,\ldots,N\}</td>
<td></td>
<td></td>
</tr>
<tr>
<td></td>
<td></td>
<td></td>
<td>Ω1,Ω2,…,ΩKaredisjointsubscriptΩ1subscriptΩ2…subscriptΩ𝐾aredisjoint\displaystyle\Omega_{1},\Omega_{2},\ldots,\Omega_{K}{~{}}{\rm are~{}disjoint}</td>
<td></td>
<td></td>
</tr>
</tbody>
</table>
where $`\Omega_{k}`$ is the set of subcarriers assigned to user $`k`$. $`\Omega_{k}`$’s must be mutually exclusive since each subcarrier is allowed to be used by one user only. In general, it is necessary to share the same subcarrier among multiple users in order to achieve the multiuser capacity region \[[13](#bib.bib13)\]. This suggests that superposition coding together with high complexity decoding should be used. However, there is only a small range of frequency with overlapping sharing according to \[[13](#bib.bib13)\] when optimal power control is used. We therefore focus on mutually exclusive subcarrier assignment schemes, which can also simplify transmitter and receiver implementation for practical systems.
## III Time-Sharing Based Optimal Subcarrier and Power Allocation
Finding the optimization variables $`\Omega_{k}`$ and $`P_{k,n}`$ for all $`k`$ and $`n`$ in ([2](#S2.E2 "In II System Model and Problem Formulation ‣ Resource Allocation for Delay Differentiated Traffic in Multiuser OFDM Systems")) is a mixed integer programming problem. In the system with $`K`$ users and $`N`$ subcarriers, there are $`K^{N}`$ possible subcarrier assignments since each subcarrier can be used by one user only. For each subcarrier assignment, the total power will be allocated to meet the individual rate requirement for each DC user and at the same time to maximize the sum-rate of the NDC users. The subcarrier assignment together with its associated power allocation that results in the largest sum-rate while satisfying all the constraints is the optimal solution.
An approach to make the problem more tractable is to relax the constraint that each subcarrier is used by one user only. We introduce a sharing factor $`\rho_{k,n} \in {\lbrack 0,1\rbrack}`$ indicating the portion of time that subcarrier $`n`$ is assigned to user $`k`$ during each transmission frame. This time-sharing technique was first proposed in \[[3](#bib.bib3)\] and has been frequently used in the context of subcarrier assignment in multiuser OFDM systems to convert a mixed integer programming problem into a convex optimization problem \[[9](#bib.bib9), [6](#bib.bib6), [5](#bib.bib5), [10](#bib.bib10)\]. In addition, we introduce a variable $`s_{k,n}`$ and define it as $`s_{k,n} = {\rho_{k,n}\hspace{0pt}P_{k,n}}`$ for all $`k`$ and $`n`$. Clearly, $`s_{k,n}`$ becomes the actual amount of power allocated to user $`k`$ on subcarrier $`n`$, whereas $`P_{k,n}`$ is the power as if subcarrier $`n`$ is occupied by user $`k`$ only. If $`\rho_{k,n} = 0`$, we always have $`s_{k,n} = 0`$ but $`P_{k,n}`$ is not necessarily equal to zero. For notation brevity, we let $`\alpha_{k,n} = {{|h_{k,n}|}^{2}/{({{\Gamma_{k}\hspace{0pt}N_{0}\hspace{0pt}B}/N})}}`$ for all $`k`$ and $`n`$, which is called the effective *channel-to-noise ratio* (CNR) of user $`k`$ on subcarrier $`n`$. Here, for the purpose of generality, the subindex $`k`$ is added to the SNR gap $`\Gamma`$ to include the case when each user has different BER requirements if adaptive modulation and coding is used. With the aid of time-sharing factors $`\rho_{k,n}`$ we now readily transform the original problem ([2](#S2.E2 "In II System Model and Problem Formulation ‣ Resource Allocation for Delay Differentiated Traffic in Multiuser OFDM Systems")) into:
<table>
<tbody>
<tr>
<td></td>
<td>max{ρk,n,sk,n}subscriptsubscript𝜌𝑘𝑛subscript𝑠𝑘𝑛\displaystyle\max_{\{\rho_{k,n},s_{k,n}\}}</td>
<td></td>
<td>∑k=K1+1K∑n=1Nρk,nlog2(1+sk,nαk,nρk,n)superscriptsubscript𝑘subscript𝐾11𝐾superscriptsubscript𝑛1𝑁subscript𝜌𝑘𝑛subscript21subscript𝑠𝑘𝑛subscript𝛼𝑘𝑛subscript𝜌𝑘𝑛\displaystyle\sum_{k=K_{1}+1}^{K}\sum_{n=1}^{N}{\rho_{k,n}\log_{2}{\left(1+\frac{s_{k,n}\alpha_{k,n}}{\rho_{k,n}}\right)}}</td>
<td></td>
<td>(3)</td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td>subjecttosubjectto\displaystyle{\rm subject~{}to}</td>
<td></td>
<td>∑n=1Nρk,nlog2(1+sk,nαk,nρk,n)≥Rk,1≤k≤K1formulae-sequencesuperscriptsubscript𝑛1𝑁subscript𝜌𝑘𝑛subscript21subscript𝑠𝑘𝑛subscript𝛼𝑘𝑛subscript𝜌𝑘𝑛subscript𝑅𝑘1𝑘subscript𝐾1\displaystyle\sum_{n=1}^{N}\rho_{k,n}\log_{2}{\Big{(}1+\frac{s_{k,n}\alpha_{k,n}}{\rho_{k,n}}\Big{)}}\geq R_{k},1\leq k\leq K_{1}</td>
<td></td>
<td rowspan="4">(7)</td>
</tr>
<tr>
<td></td>
<td></td>
<td></td>
<td>∑k=1K∑n=1Nsk,n=PTsuperscriptsubscript𝑘1𝐾superscriptsubscript𝑛1𝑁subscript𝑠𝑘𝑛subscript𝑃𝑇\displaystyle\sum_{k=1}^{K}\sum_{n=1}^{N}{s_{k,n}}=P_{T}</td>
<td></td>
</tr>
<tr>
<td></td>
<td></td>
<td></td>
<td>∑k=1Kρk,n=1,∀nsuperscriptsubscript𝑘1𝐾subscript𝜌𝑘𝑛1for-all𝑛\displaystyle\sum_{k=1}^{K}{\rho_{k,n}}=1,~{}\forall n</td>
<td></td>
</tr>
<tr>
<td></td>
<td></td>
<td></td>
<td>sk,n≥0,0≤ρk,n≤1,∀k,n.formulae-sequenceformulae-sequencesubscript𝑠𝑘𝑛00subscript𝜌𝑘𝑛1for-all𝑘𝑛\displaystyle s_{k,n}\geq 0,~{}0\leq\rho_{k,n}\leq 1,~{}\forall k,n~{}.</td>
<td></td>
</tr>
</tbody>
</table>
The objective function ([3](#S3.E3 "In III Time-Sharing Based Optimal Subcarrier and Power Allocation ‣ Resource Allocation for Delay Differentiated Traffic in Multiuser OFDM Systems")) is a sum of functions of the form $`{f\hspace{0pt}{(\rho_{k,n},s_{k,n})}} = {\rho_{k,n}\hspace{0pt}{\log_{2}\left( {1 + {{C\hspace{0pt}s_{k,n}}/\rho_{k,n}}} \right)}}`$, where $`C`$ is some positive constant. By evaluating the Hessian matrix of $`f\hspace{0pt}{(\rho_{k,n},s_{k,n})}`$ at $`\rho_{k,n}`$ and $`s_{k,n}`$, we can prove that $`f\hspace{0pt}{(\rho_{k,n},s_{k,n})}`$ is concave \[[14](#bib.bib14)\]. Thus, the objective function is also concave since any positive linear combination of concave functions is concave. Moreover, since the inequality constraint functions in ([7](#S3.E7 "In III Time-Sharing Based Optimal Subcarrier and Power Allocation ‣ Resource Allocation for Delay Differentiated Traffic in Multiuser OFDM Systems")) are convex and the constraints in ([7](#S3.E7 "In III Time-Sharing Based Optimal Subcarrier and Power Allocation ‣ Resource Allocation for Delay Differentiated Traffic in Multiuser OFDM Systems"))-([7](#S3.E7 "In III Time-Sharing Based Optimal Subcarrier and Power Allocation ‣ Resource Allocation for Delay Differentiated Traffic in Multiuser OFDM Systems")) are all affine, the feasible set of this optimization problem is convex. Therefore, the problem in ([3](#S3.E3 "In III Time-Sharing Based Optimal Subcarrier and Power Allocation ‣ Resource Allocation for Delay Differentiated Traffic in Multiuser OFDM Systems"))-([7](#S3.E7 "In III Time-Sharing Based Optimal Subcarrier and Power Allocation ‣ Resource Allocation for Delay Differentiated Traffic in Multiuser OFDM Systems")) is a convex optimization problem and there exists a unique optimal solution, which can be obtained in polynomial time. In the following we derive some desirable properties of the optimal solution.
The Lagrangian of the above problem is given by
<table>
<tbody>
<tr>
<td></td>
<td>J1({ρk,n},{sk,n},𝜷,μ,𝒗)subscript𝐽1subscript𝜌𝑘𝑛subscript𝑠𝑘𝑛𝜷𝜇𝒗\displaystyle J_{1}\left(\{\rho_{k,n}\},\{s_{k,n}\},\boldsymbol{\beta},\mu,\boldsymbol{v}\right)</td>
<td>=\displaystyle=</td>
<td>∑k=K1+1K∑n=1Nρk,nlog2(1+sk,nαk,nρk,n)+limit-fromsuperscriptsubscript𝑘subscript𝐾11𝐾superscriptsubscript𝑛1𝑁subscript𝜌𝑘𝑛subscript21subscript𝑠𝑘𝑛subscript𝛼𝑘𝑛subscript𝜌𝑘𝑛\displaystyle\sum_{k=K_{1}+1}^{K}\sum_{n=1}^{N}{\rho_{k,n}\log_{2}{\left(1+\frac{s_{k,n}\alpha_{k,n}}{\rho_{k,n}}\right)}}+</td>
<td></td>
<td rowspan="3">(8)</td>
</tr>
<tr>
<td></td>
<td></td>
<td></td>
<td>∑k=1K1βk[∑n=1Nρk,nlog2(1+sk,nαk,nρk,n)−Rk]+limit-fromsuperscriptsubscript𝑘1subscript𝐾1subscript𝛽𝑘delimited-[]superscriptsubscript𝑛1𝑁subscript𝜌𝑘𝑛subscript21subscript𝑠𝑘𝑛subscript𝛼𝑘𝑛subscript𝜌𝑘𝑛subscript𝑅𝑘\displaystyle\sum_{k=1}^{K_{1}}{\beta_{k}\left[\sum_{n=1}^{N}{\rho_{k,n}\log_{2}{\left(1+\frac{s_{k,n}\alpha_{k,n}}{\rho_{k,n}}\right)-R_{k}}}\right]}+</td>
<td></td>
</tr>
<tr>
<td></td>
<td></td>
<td></td>
<td>μ(PT−∑k=1K∑n=1Nsk,n)+∑n=1Nvn(1−∑k=1Kρk,n),𝜇subscript𝑃𝑇superscriptsubscript𝑘1𝐾superscriptsubscript𝑛1𝑁subscript𝑠𝑘𝑛superscriptsubscript𝑛1𝑁subscript𝑣𝑛1superscriptsubscript𝑘1𝐾subscript𝜌𝑘𝑛\displaystyle\mu\left(P_{T}-\sum_{k=1}^{K}\sum_{n=1}^{N}{s_{k,n}}\right)+\sum_{n=1}^{N}{v_{n}\left(1-\sum_{k=1}^{K}{\rho_{k,n}}\right)},</td>
<td></td>
</tr>
</tbody>
</table>
where $`{\mathbf{β}} = {(\beta_{1},\ldots,\beta_{K_{1}})} \succeq 0`$, $`\mu \geq 0`$ and $`{\mathbf{v}} = {(v_{1},\ldots,v_{N})}`$ are the Lagrange multipliers for the constraints ([7](#S3.E7 "In III Time-Sharing Based Optimal Subcarrier and Power Allocation ‣ Resource Allocation for Delay Differentiated Traffic in Multiuser OFDM Systems")), ([7](#S3.E7 "In III Time-Sharing Based Optimal Subcarrier and Power Allocation ‣ Resource Allocation for Delay Differentiated Traffic in Multiuser OFDM Systems")) and ([7](#S3.E7 "In III Time-Sharing Based Optimal Subcarrier and Power Allocation ‣ Resource Allocation for Delay Differentiated Traffic in Multiuser OFDM Systems")), respectively. The boundary constraints ([7](#S3.E7 "In III Time-Sharing Based Optimal Subcarrier and Power Allocation ‣ Resource Allocation for Delay Differentiated Traffic in Multiuser OFDM Systems")) will be absorbed in the Karush-Kuhn-Tucker (KKT) conditions \[[14](#bib.bib14)\] as shown later. Let $`\rho_{k,n}^{\ast}`$ and $`s_{k,n}^{\ast}`$ denote the optimal solution, if it exists, for $`1 \leq k \leq K`$, $`1 \leq n \leq N`$. Applying the KKT conditions, we can obtain the necessary and sufficient conditions for $`\rho_{k,n}^{\ast}`$ and $`s_{k,n}^{\ast}`$. Specifically, $`\rho_{k,n}^{\ast}`$ and $`s_{k,n}^{\ast}`$ should satisfy the following equations:
<table>
<tbody>
<tr>
<td></td>
<td>∂J1(…)∂sk,n∗{=0,sk,n∗>0<0,sk,n∗=0,∀k,nsubscript𝐽1…superscriptsubscript𝑠𝑘𝑛casesformulae-sequenceabsent0superscriptsubscript𝑠𝑘𝑛0formulae-sequenceabsent0superscriptsubscript𝑠𝑘𝑛0for-all𝑘𝑛\displaystyle\frac{\partial J_{1}(\ldots)}{\partial s_{k,n}^{*}}\left\{\begin{array}[]{l}=0,~{}s_{k,n}^{*}>0\\ <0,~{}s_{k,n}^{*}=0\\ \end{array}\right.,~{}~{}\forall k,n</td>
<td></td>
<td>(11)</td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td>∂J1(…)∂ρk,n∗{<0,ρk,n∗=0=0,0<ρk,n∗<1>0,ρk,n∗=1,∀k,nsubscript𝐽1…superscriptsubscript𝜌𝑘𝑛casesformulae-sequenceabsent0superscriptsubscript𝜌𝑘𝑛0formulae-sequenceabsent00superscriptsubscript𝜌𝑘𝑛1formulae-sequenceabsent0superscriptsubscript𝜌𝑘𝑛1for-all𝑘𝑛\displaystyle\frac{\partial J_{1}(\ldots)}{\partial\rho_{k,n}^{*}}\left\{\begin{array}[]{l}<0,~{}\rho_{k,n}^{*}=0\\ =0,~{}0<\rho_{k,n}^{*}<1\\ >0,~{}\rho_{k,n}^{*}=1\end{array}\right.,~{}~{}\forall k,n</td>
<td></td>
<td>(15)</td>
</tr>
</tbody>
</table>
<table>
<tbody>
<tr>
<td></td>
<td>βk[∑n=1Nρk,n∗log2(1+sk,n∗αk,nρk,n∗)−Rk]=0,1≤k≤K1.formulae-sequencesubscript𝛽𝑘delimited-[]superscriptsubscript𝑛1𝑁superscriptsubscript𝜌𝑘𝑛subscript21superscriptsubscript𝑠𝑘𝑛subscript𝛼𝑘𝑛superscriptsubscript𝜌𝑘𝑛subscript𝑅𝑘01𝑘subscript𝐾1\displaystyle\beta_{k}\left[\sum_{n=1}^{N}\rho_{k,n}^{*}\log_{2}{\Big{(}1+\frac{s_{k,n}^{*}\alpha_{k,n}}{\rho_{k,n}^{*}}\Big{)}}-R_{k}\right]=0,~{}~{}1\leq k\leq K_{1}.</td>
<td></td>
<td>(16)</td>
</tr>
</tbody>
</table>
###
III-A
Optimal Power Allocation for Given Subcarrier Assignment
In this subsection, we present the optimal power distribution when subcarrier assignment is given.
Let $`\{\rho_{k,n}\}`$ be any given subcarrier assignment scheme. Differentiating the Lagrangian in ([8](#S3.E8 "In III Time-Sharing Based Optimal Subcarrier and Power Allocation ‣ Resource Allocation for Delay Differentiated Traffic in Multiuser OFDM Systems")) with respect to $`s_{k,n}`$ and substituting the result into the KKT condition ([11](#S3.E11 "In III Time-Sharing Based Optimal Subcarrier and Power Allocation ‣ Resource Allocation for Delay Differentiated Traffic in Multiuser OFDM Systems")), we obtain:
<table>
<tbody>
<tr>
<td></td>
<td>Pk,n∗superscriptsubscript𝑃𝑘𝑛\displaystyle P_{k,n}^{*}</td>
<td>=\displaystyle=</td>
<td>sk,n∗ρk,n=(βkμln2−1αk,n)+superscriptsubscript𝑠𝑘𝑛subscript𝜌𝑘𝑛superscriptsubscript𝛽𝑘𝜇21subscript𝛼𝑘𝑛\displaystyle\frac{s_{k,n}^{*}}{\rho_{k,n}}=\left(\frac{\beta_{k}}{\mu\ln 2}-\frac{1}{\alpha_{k,n}}\right)^{+}</td>
<td></td>
<td>(17)</td>
</tr>
</tbody>
</table>
for $`k = {1,\ldots,K_{1}}`$ and $`n = {1,\ldots,N}`$, and
<table>
<tbody>
<tr>
<td></td>
<td>Pk,n∗superscriptsubscript𝑃𝑘𝑛\displaystyle P_{k,n}^{*}</td>
<td>=\displaystyle=</td>
<td>sk,n∗ρk,n=(1μln2−1αk,n)+superscriptsubscript𝑠𝑘𝑛subscript𝜌𝑘𝑛superscript1𝜇21subscript𝛼𝑘𝑛\displaystyle\frac{s_{k,n}^{*}}{\rho_{k,n}}=\left(\frac{1}{\mu\ln 2}-\frac{1}{\alpha_{k,n}}\right)^{+}</td>
<td></td>
<td>(18)</td>
</tr>
</tbody>
</table>
for $`k = {{K_{1} + 1},\ldots,K}`$ and $`n = {1,\ldots,N}`$. Here, $`{(x)}^{+} \triangleq {\max{(0,x)}}`$.
Equations ([17](#S3.E17 "In III-A Optimal Power Allocation for Given Subcarrier Assignment ‣ III Time-Sharing Based Optimal Subcarrier and Power Allocation ‣ Resource Allocation for Delay Differentiated Traffic in Multiuser OFDM Systems")) and ([18](#S3.E18 "In III-A Optimal Power Allocation for Given Subcarrier Assignment ‣ III Time-Sharing Based Optimal Subcarrier and Power Allocation ‣ Resource Allocation for Delay Differentiated Traffic in Multiuser OFDM Systems")) clearly show that the optimal power allocation follows the standard water-filling approach, except that the allocated power is only on for $`\rho_{k,n}`$ portion of time. For each user, more power will be allocated to the subcarriers with higher CNRs and vice versa. But the water levels of different users can be different. Specifically, the water level of each DC user is given by $`L_{k} = {\beta_{k}/{({\mu\hspace{0pt}{\ln 2}})}}`$, for $`k = {1,\ldots,K_{1}}`$, and it should ensure the basic rate requirement $`R_{k}`$ in ([7](#S3.E7 "In III Time-Sharing Based Optimal Subcarrier and Power Allocation ‣ Resource Allocation for Delay Differentiated Traffic in Multiuser OFDM Systems")). Substituting ([17](#S3.E17 "In III-A Optimal Power Allocation for Given Subcarrier Assignment ‣ III Time-Sharing Based Optimal Subcarrier and Power Allocation ‣ Resource Allocation for Delay Differentiated Traffic in Multiuser OFDM Systems")) into the KKT condition ([16](#S3.E16 "In III Time-Sharing Based Optimal Subcarrier and Power Allocation ‣ Resource Allocation for Delay Differentiated Traffic in Multiuser OFDM Systems")) and in view of $`\beta_{k} \neq 0`$, we obtain the closed-form expression for $`L_{k}`$ given by:
<table>
<tbody>
<tr>
<td></td>
<td>Lksubscript𝐿𝑘\displaystyle L_{k}</td>
<td>=\displaystyle=</td>
<td>[2Rk∏n∈Ωk(αk,n)ρk,n]1/∑n∈Ωkρk,n,superscriptdelimited-[]superscript2subscript𝑅𝑘subscriptproduct𝑛subscriptΩ𝑘superscriptsubscript𝛼𝑘𝑛subscript𝜌𝑘𝑛1subscript𝑛subscriptΩ𝑘subscript𝜌𝑘𝑛\displaystyle\left[\frac{2^{R_{k}}}{\prod_{n\in\Omega_{k}}{(\alpha_{k,n})^{\rho_{k,n}}}}\right]^{1/\sum_{n\in\Omega_{k}}{\rho_{k,n}}},</td>
<td></td>
<td>(19)</td>
</tr>
</tbody>
</table>
where $`\Omega_{k}`$ is the set of subcarriers that is assigned to user $`k`$ with $`\rho_{k,n} > 0`$ and satisfies $`\alpha_{k,n} > {1/L_{k}}`$. In the case where the given subcarrier assignment is mutually exclusive, i.e., all the $`\rho_{k,n}`$’s only take one or zero, the water levels $`L_{k}`$ can be re-expressed as
<table>
<tbody>
<tr>
<td></td>
<td>Lksubscript𝐿𝑘\displaystyle L_{k}</td>
<td>=\displaystyle=</td>
<td>(2Rk∏n′=1gkαk,n′)1/gk,superscriptsuperscript2subscript𝑅𝑘superscriptsubscriptproductsuperscript𝑛′1subscript𝑔𝑘subscript𝛼𝑘superscript𝑛′1subscript𝑔𝑘\displaystyle\left(\frac{2^{R_{k}}}{\prod_{n^{\prime}=1}^{g_{k}}{\alpha_{k,n^{\prime}}}}\right)^{1/g_{k}},</td>
<td></td>
<td>(20)</td>
</tr>
</tbody>
</table>
where $`\alpha_{k,1'} \geq \alpha_{k,2'} \geq \ldots, \geq \alpha_{k,{|\Omega_{k}|}'}`$ are the ordered CNRs of the $`k`$-th DC user on its allocated subcarrier set $`\Omega_{k}`$, and $`g_{k}\mspace{7mu}{({\leq {|\Omega_{k}|}})}`$ is the largest integer satisfying $`\alpha_{k,g_{k}'} > {1/L_{k}}`$. The water level of all the NDC users, on the other hand, is observed from ([18](#S3.E18 "In III-A Optimal Power Allocation for Given Subcarrier Assignment ‣ III Time-Sharing Based Optimal Subcarrier and Power Allocation ‣ Resource Allocation for Delay Differentiated Traffic in Multiuser OFDM Systems")) to be the same and is given by $`L_{0} = {1/{({\mu\hspace{0pt}{\ln 2}})}}`$. Let $`P_{{DC},T} = {\sum_{k = 1}^{K_{1}}{\sum_{n = 1}^{N}{\rho_{k,n}\hspace{0pt}P_{k,n}^{\ast}}}}`$ represent the actual total power consumed by the $`K_{1}`$ DC users. Then the water level of NDC users, $`L_{0}`$, can be obtained numerically by using the total power constraint $`{P_{{DC},T} + {\sum_{k = {K_{1} + 1}}^{K}{\sum_{n = 1}^{N}{\rho_{k,n}\hspace{0pt}P_{k,n}^{\ast}}}}} = P_{T}`$.
Fig. [2](#A0.F2 "Figure 2 ‣ Resource Allocation for Delay Differentiated Traffic in Multiuser OFDM Systems") illustrates the optimal power distribution based on multi-level water-filling in a multiuser OFDM system with $`K = 4`$ users and $`N = 8`$ subcarriers. In the figure, the height of each blank region represents the inverse of the channel-to-noise ratio and the height of shadowed ones is the allocated power. The water levels of the two DC users, $`k = 1`$ and $`k = 2`$, are given by $`L_{1}`$ and $`L_{2}`$, respectively, and they are determined explicitly by the basic rate targets. The water level of the two NDC users, $`k = 3`$ and $`k = 4`$, is given by $`L_{0}`$, and it depends on the total available power after the subtraction of the power consumed by DC users. This interpretation on the determination of water levels implicitly imposes higher priorities on DC users, for which the basic rate targets must be guaranteed all the time.
###
III-B
Optimal Subcarrier Assignment
The subcarriers and power should be allocated jointly to achieve the optimal solution of the problem formulated in ([3](#S3.E3 "In III Time-Sharing Based Optimal Subcarrier and Power Allocation ‣ Resource Allocation for Delay Differentiated Traffic in Multiuser OFDM Systems"))-([7](#S3.E7 "In III Time-Sharing Based Optimal Subcarrier and Power Allocation ‣ Resource Allocation for Delay Differentiated Traffic in Multiuser OFDM Systems")). The previous subsection discussed the analytical expressions of the optimal power allocation for a given subcarrier assignment. In this subsection, we derive the optimal strategy for subcarrier assignment assuming the power allocation is optimized.
Taking the partial derivative of the Lagrangian in ([8](#S3.E8 "In III Time-Sharing Based Optimal Subcarrier and Power Allocation ‣ Resource Allocation for Delay Differentiated Traffic in Multiuser OFDM Systems")) with respect to $`\rho_{k,n}`$, we have:
<table>
<tbody>
<tr>
<td></td>
<td>∂J1(…)∂ρk,nsubscript𝐽1…subscript𝜌𝑘𝑛\displaystyle\frac{\partial J_{1}(\ldots)}{\partial\rho_{k,n}}</td>
<td>=\displaystyle=</td>
<td>β~k[log2(1+sk,nαk,nρk,n)−sk,nαk,nln2(ρk,n+sk,nαk,n)]−vn,subscript~𝛽𝑘delimited-[]subscript21subscript𝑠𝑘𝑛subscript𝛼𝑘𝑛subscript𝜌𝑘𝑛subscript𝑠𝑘𝑛subscript𝛼𝑘𝑛2subscript𝜌𝑘𝑛subscript𝑠𝑘𝑛subscript𝛼𝑘𝑛subscript𝑣𝑛\displaystyle{\tilde{\beta}}_{k}\left[\log_{2}{\Big{(}1+\frac{s_{k,n}\alpha_{k,n}}{\rho_{k,n}}\Big{)}}-\frac{s_{k,n}\alpha_{k,n}}{\ln 2(\rho_{k,n}+s_{k,n}\alpha_{k,n})}\right]-v_{n},</td>
<td></td>
<td>(21)</td>
</tr>
</tbody>
</table>
where $`{\overset{\sim}{\beta}}_{k} = \beta_{k}`$ for $`1 \leq k \leq K_{1}`$ and $`{\overset{\sim}{\beta}}_{k} = 1`$ otherwise. Now we substitute the optimal power allocation ([17](#S3.E17 "In III-A Optimal Power Allocation for Given Subcarrier Assignment ‣ III Time-Sharing Based Optimal Subcarrier and Power Allocation ‣ Resource Allocation for Delay Differentiated Traffic in Multiuser OFDM Systems")) and ([18](#S3.E18 "In III-A Optimal Power Allocation for Given Subcarrier Assignment ‣ III Time-Sharing Based Optimal Subcarrier and Power Allocation ‣ Resource Allocation for Delay Differentiated Traffic in Multiuser OFDM Systems")) into ([21](#S3.E21 "In III-B Optimal Subcarrier Assignment ‣ III Time-Sharing Based Optimal Subcarrier and Power Allocation ‣ Resource Allocation for Delay Differentiated Traffic in Multiuser OFDM Systems")) and apply the KKT condition ([15](#S3.E15 "In III Time-Sharing Based Optimal Subcarrier and Power Allocation ‣ Resource Allocation for Delay Differentiated Traffic in Multiuser OFDM Systems")), then we get:
<table>
<tbody>
<tr>
<td></td>
<td>ρk,n∗superscriptsubscript𝜌𝑘𝑛\displaystyle\rho_{k,n}^{*}</td>
<td>=\displaystyle=</td>
<td>{1,vn<Hk,n(L0,L~k)0,vn>Hk,n(L0,L~k).cases1subscript𝑣𝑛subscript𝐻𝑘𝑛subscript𝐿0subscript~𝐿𝑘0subscript𝑣𝑛subscript𝐻𝑘𝑛subscript𝐿0subscript~𝐿𝑘\displaystyle\left\{\begin{array}[]{ll}1,&v_{n}<H_{k,n}(L_{0},{\tilde{L}}_{k})\\ 0,&v_{n}>H_{k,n}(L_{0},{\tilde{L}}_{k})\end{array}\right..</td>
<td></td>
</tr>
</tbody>
</table>
Here, the variable $`{\overset{\sim}{L}}_{k}`$, for $`k = {1,\ldots,K}`$ is defined as $`{\overset{\sim}{L}}_{k} = L_{k}`$ for $`1 \leq k \leq K_{1}`$ and $`{\overset{\sim}{L}}_{k} = L_{0}`$ otherwise, and the function $`H_{k,n}\hspace{0pt}{(L_{0},{\overset{\sim}{L}}_{k})}`$ is given by:
<table>
<tbody>
<tr>
<td></td>
<td>Hk,n(L0,L~k)subscript𝐻𝑘𝑛subscript𝐿0subscript~𝐿𝑘\displaystyle H_{k,n}(L_{0},{\tilde{L}}_{k})</td>
<td>=\displaystyle=</td>
<td>L~kL0{[log2(αk,nL~k)]+−1ln2(1−1αk,nL~k)+}.subscript~𝐿𝑘subscript𝐿0superscriptdelimited-[]subscript2subscript𝛼𝑘𝑛subscript~𝐿𝑘12superscript11subscript𝛼𝑘𝑛subscript~𝐿𝑘\displaystyle\frac{{\tilde{L}}_{k}}{L_{0}}\left\{\left[\log_{2}{\big{(}\alpha_{k,n}{\tilde{L}}_{k}\big{)}}\right]^{+}-\frac{1}{\ln 2}\left(1-\frac{1}{\alpha_{k,n}{\tilde{L}}_{k}}\right)^{+}\right\}.</td>
<td></td>
<td>(23)</td>
</tr>
</tbody>
</table>
Due to the constraint in ([7](#S3.E7 "In III Time-Sharing Based Optimal Subcarrier and Power Allocation ‣ Resource Allocation for Delay Differentiated Traffic in Multiuser OFDM Systems")), we conclude that the optimal subcarrier assignment is decomposed into $`N`$ independent problems. That is, for each subcarrier $`n`$, if $`H_{k,n}\hspace{0pt}{(L_{0},{\overset{\sim}{L}}_{k})}`$’s, for $`k = {1,\ldots,K}`$, are all distinct, then only the user with the largest $`H_{k,n}\hspace{0pt}{(L_{0},{\overset{\sim}{L}}_{k})}`$ can use that subcarrier. In other words, we have:
<table>
<tbody>
<tr>
<td></td>
<td>ρk′(n),n∗=1,ρk,n∗=0,formulae-sequencesuperscriptsubscript𝜌superscript𝑘′𝑛𝑛1superscriptsubscript𝜌𝑘𝑛0\displaystyle\rho_{k^{\prime}(n),n}^{*}=1,{~{}}\rho_{k,n}^{*}=0,</td>
<td></td>
<td>∀k≠k′(n)for-all𝑘superscript𝑘′𝑛\displaystyle\forall k\neq k^{\prime}(n)</td>
<td></td>
<td>(24)</td>
</tr>
</tbody>
</table>
where $`{k'\hspace{0pt}{(n)}} = {{\arg{\max_{1 \leq k \leq K}H_{k,n}}}\hspace{0pt}{(L_{0},{\overset{\sim}{L}}_{k})}}`$. Hence, it follows that for a given set of water levels $`\{ L_{0},L_{1},\ldots,L_{K_{1}}\}`$, or equivalently, a given set of Lagrange multipliers $`\{{\mathbf{β}},\mu\}`$, we can determine the optimal subcarrier allocation using ([24](#S3.E24 "In III-B Optimal Subcarrier Assignment ‣ III Time-Sharing Based Optimal Subcarrier and Power Allocation ‣ Resource Allocation for Delay Differentiated Traffic in Multiuser OFDM Systems")). However, the optimal solution obtained may not satisfy the individual rate constraint ([7](#S3.E7 "In III Time-Sharing Based Optimal Subcarrier and Power Allocation ‣ Resource Allocation for Delay Differentiated Traffic in Multiuser OFDM Systems")) and the total power constraint ([7](#S3.E7 "In III Time-Sharing Based Optimal Subcarrier and Power Allocation ‣ Resource Allocation for Delay Differentiated Traffic in Multiuser OFDM Systems")).
Function $`H_{k,n}\hspace{0pt}{(L_{0},{\overset{\sim}{L}}_{k})}`$ in ([23](#S3.E23 "In III-B Optimal Subcarrier Assignment ‣ III Time-Sharing Based Optimal Subcarrier and Power Allocation ‣ Resource Allocation for Delay Differentiated Traffic in Multiuser OFDM Systems")) plays a key role in finding the optimal solution of subcarrier and power allocation. We now take a closer look at its features. Firstly, by differentiating $`H_{k,n}\hspace{0pt}{(L_{0},{\overset{\sim}{L}}_{k})}`$ with respect to $`\alpha_{k,n}`$ at $`\alpha_{k,n} > {1/{\overset{\sim}{L}}_{k}}`$, it is seen that $`H_{k,n}\hspace{0pt}{(L_{0},{\overset{\sim}{L}}_{k})}`$ is monotonically increasing in $`\alpha_{k,n}`$. As a result, for each subcarrier, the user with a larger CNR is more likely to be assigned this subcarrier. In the extreme case where DC users are absent (i.e. $`K_{1} = 0`$), each subcarrier will be assigned to the user with the largest CNR. This agrees with the previous result in \[[4](#bib.bib4)\]. In the general case, the candidate pool for accessing each subcarrier consists of all the $`K_{1}`$ DC users and the only one NDC user with the largest CNR. This consequently suggests that the exhaustive search for optimal subcarrier assignment in the original problem ([2](#S2.E2 "In II System Model and Problem Formulation ‣ Resource Allocation for Delay Differentiated Traffic in Multiuser OFDM Systems")) has a complexity of $`{({K_{1} + 1})}^{N}`$ rather than $`K^{N}`$. Secondly, we also observe that $`H_{k,n}\hspace{0pt}{(L_{0},{\overset{\sim}{L}}_{k})}`$ is a non-decreasing function of $`L_{k}`$ and $`L_{0}`$, respectively, when $`1 \leq k \leq K_{1}`$ and $`K_{1} < k \leq K`$. This is obtained by differentiating $`H_{k,n}\hspace{0pt}{(L_{0},{\overset{\sim}{L}}_{k})}`$ with respect to $`L_{k}`$ and $`L_{0}`$ for $`1 \leq k \leq K_{1}`$ and $`K_{1} < k \leq K`$, respectively. Thus, increasing the water level $`L_{k}`$ of DC user $`k`$ increases its chance to occupy more subcarriers. Similarly, increasing the water level $`L_{0}`$ of NDC users allow them to access more subcarriers.
Using the above observations, we propose in the next subsection an iterative algorithm to find the optimal water levels and determine the corresponding subcarrier assignment so that all the rate and power constraints are satisfied.
###
III-C
Iterative Algorithm
The thrust of the algorithm is to obtain the set of optimal water levels $`\{ L_{0},L_{1},\ldots,L_{K_{1}}\}`$ using two nested loops. The outer loop varies $`L_{0}`$ to meet the total power constraint. The inner loop searches $`\{ L_{1},\ldots,L_{K_{1}}\}`$ and determines the optimal $`\rho_{k,n}`$ for all $`k`$ and $`n`$ at a given value of $`L_{0}`$ to satisfy the basic rate requirement for every DC user. The algorithm is outlined as follows.
Optimal Subcarrier Assignment Algorithm
*Main function*
- a\)
Set $`L_{LB} = L_{UB} = {{\min_{\begin{matrix}
{K_{1} < k \leq K} \\
{1 \leq n \leq N}
\end{matrix}}{\{{1/\alpha_{k,n}}\}}} - \Delta}`$
<sup>1</sup>
<sup>1</sup>1$`\Delta`$ is a very small number.
- b\)
Find the optimal $`\{ L_{k},\rho_{k,n}\}`$ at $`L_{0} = L_{UB}`$
Compute the actual power consumption $`P_{T}'`$ using ([17](#S3.E17 "In III-A Optimal Power Allocation for Given Subcarrier Assignment ‣ III Time-Sharing Based Optimal Subcarrier and Power Allocation ‣ Resource Allocation for Delay Differentiated Traffic in Multiuser OFDM Systems")) and ([18](#S3.E18 "In III-A Optimal Power Allocation for Given Subcarrier Assignment ‣ III Time-Sharing Based Optimal Subcarrier and Power Allocation ‣ Resource Allocation for Delay Differentiated Traffic in Multiuser OFDM Systems"))
if $`P_{T}' < P_{T}`$, update $`L_{UB} = {2\hspace{0pt}L_{UB}}`$, and repeat Step b)
else, go to Step c)
- c\)
Find the optimal $`\{ L_{k},\rho_{k,n}\}`$ at $`L_{0} = {{({L_{LB} + L_{UB}})}/2}`$
Compute $`P_{T}'`$ using ([17](#S3.E17 "In III-A Optimal Power Allocation for Given Subcarrier Assignment ‣ III Time-Sharing Based Optimal Subcarrier and Power Allocation ‣ Resource Allocation for Delay Differentiated Traffic in Multiuser OFDM Systems")) and ([18](#S3.E18 "In III-A Optimal Power Allocation for Given Subcarrier Assignment ‣ III Time-Sharing Based Optimal Subcarrier and Power Allocation ‣ Resource Allocation for Delay Differentiated Traffic in Multiuser OFDM Systems"))
if $`P_{T}' > P_{T}`$, let $`L_{UB} = L_{0}`$
elseif $`P_{T}' < P_{T}`$, let $`L_{LB} = L_{0}`$
Repeat Step c) until $`P_{T}' = P_{T}`$
*Function*: find the optimal $`\{ L_{k},\rho_{k,n}\}`$ at a given $`L_{0}`$
- 1\)
Set $`L_{k}`$ using ([20](#S3.E20 "In III-A Optimal Power Allocation for Given Subcarrier Assignment ‣ III Time-Sharing Based Optimal Subcarrier and Power Allocation ‣ Resource Allocation for Delay Differentiated Traffic in Multiuser OFDM Systems")) with $`\Omega_{k} = {\{ 1,\ldots,N\}}`$, for $`k = {1,\ldots,K_{1}}`$
Compute $`H_{k,n}`$ using ([23](#S3.E23 "In III-B Optimal Subcarrier Assignment ‣ III Time-Sharing Based Optimal Subcarrier and Power Allocation ‣ Resource Allocation for Delay Differentiated Traffic in Multiuser OFDM Systems")), $`{\forall k},n`$
Obtain $`\rho_{k,n}`$ using ([24](#S3.E24 "In III-B Optimal Subcarrier Assignment ‣ III Time-Sharing Based Optimal Subcarrier and Power Allocation ‣ Resource Allocation for Delay Differentiated Traffic in Multiuser OFDM Systems")), $`{\forall k},n`$
- 2\)
Compute $`R_{k}' = {\sum_{n = 1}^{N}{\rho_{k,n}\hspace{0pt}{\lbrack{\log_{2}{({L_{k}\hspace{0pt}\alpha_{k,n}})}}\rbrack}^{+}}}`$ for $`k = {1,\ldots,K_{1}}`$
- 3a)
Find $`k^{\ast}`$ with $`R_{k^{\ast}}' < R_{k^{\ast}}`$ and $`{R_{k^{\ast}}' - R_{k^{\ast}}} \leq {R_{k}' - R_{k}}`$ for all $`1 \leq k \leq K_{1}`$
- 3b)
Find the subcarrier set $`\mathcal{A}_{k^{\ast}} = \left. \{ n \middle| {\rho_{k^{\ast},n} < 1}\} \right.`$ for the found $`k^{\ast}`$
for each $`n \in \mathcal{A}_{k^{\ast}}`$
Let $`{k\hspace{0pt}{(n)}} = {\arg\hspace{0pt}{\max_{1 \leq k \leq K}H_{k,n}}}`$ and obtain $`L_{k^{\ast}}\hspace{0pt}{(n)}`$ such that $`H_{k^{\ast},n} = {{({1 + \Delta})}\hspace{0pt}H_{{k\hspace{0pt}{(n)}},n}}`$
- 3c)
while $`R_{k^{\ast}}' < R_{k^{\ast}}`$ for the found $`k^{\ast}`$
Denote $`m = {\arg\hspace{0pt}{\min_{n \in \mathcal{A}_{k^{\ast}}}L_{k^{\ast}}}\hspace{0pt}{(n)}}`$
Update $`L_{k^{\ast}} = {L_{k^{\ast}}\hspace{0pt}{(m)}}`$
Update $`\rho_{k^{\ast},m} = 1`$, $`\rho_{k,m} = 0`$ for $`{\forall k} \neq k^{\ast}`$, and $`\mathcal{A}_{k^{\ast}} = {\mathcal{A}_{k^{\ast}} - {\{ m\}}}`$
Compute $`R_{k^{\ast}}' = {\sum_{n = 1}^{N}{\rho_{k^{\ast},n}\hspace{0pt}{\lbrack{\log_{2}{({L_{k^{\ast}}\hspace{0pt}\alpha_{k^{\ast},n}})}}\rbrack}^{+}}}`$
- 3d)
if $`R_{k^{\ast}}' > R_{k^{\ast}}`$, let $`R_{k^{\ast}}' = {R_{k^{\ast}}' - {\log_{2}{({L_{k^{\ast}}\hspace{0pt}\alpha_{k^{\ast},m}})}}}`$
if $`R_{k^{\ast}}' < R_{k^{\ast}}`$
Update $`\rho_{k^{\ast},m} = {{({R_{k^{\ast}} - R_{k^{\ast}}'})}/{\log_{2}{({L_{k^{\ast}}\hspace{0pt}\alpha_{k^{\ast},m}})}}}`$ and $`\rho_{{k\hspace{0pt}{(m)}},m} = {1 - \rho_{k^{\ast},m}}`$
elseif $`R_{k^{\ast}}' > R_{k^{\ast}}`$
Update $`\rho_{k^{\ast},m} = 0`$, $`\rho_{{k\hspace{0pt}{(m)}},m} = 1`$ and compute $`L_{k^{\ast}}`$ using ([19](#S3.E19 "In III-A Optimal Power Allocation for Given Subcarrier Assignment ‣ III Time-Sharing Based Optimal Subcarrier and Power Allocation ‣ Resource Allocation for Delay Differentiated Traffic in Multiuser OFDM Systems"))
- 4\)
Repeat Steps 2)-3) until $`R_{k}' = R_{k}`$ for all $`k = {1,\ldots,K_{1}}`$
In the outer loop (main function), we initialize $`L_{0}`$ to a value slightly below the minimum reciprocal of the CNRs of all NDC users over all subcarriers so that none of the NDC users is assigned any power resource based on ([18](#S3.E18 "In III-A Optimal Power Allocation for Given Subcarrier Assignment ‣ III Time-Sharing Based Optimal Subcarrier and Power Allocation ‣ Resource Allocation for Delay Differentiated Traffic in Multiuser OFDM Systems")). By doing this, the subcarrier and power will be initially allocated to all DC users as if NDC users were absent. We then keep increasing $`L_{0}`$ until the actual total power consumption $`P_{T}'`$ exceeds the total available power $`P_{T}`$ and an upper bound of $`L_{0}`$ is obtained. Note that if the number of subcarriers in a practical system is large enough, we can always find at least one subcarrier fully occupied by one NDC user and, therefore, an analytical upper bound of $`L_{0}`$ can be derived from ([18](#S3.E18 "In III-A Optimal Power Allocation for Given Subcarrier Assignment ‣ III Time-Sharing Based Optimal Subcarrier and Power Allocation ‣ Resource Allocation for Delay Differentiated Traffic in Multiuser OFDM Systems")) as $`L_{UB} = {P_{T} + {\max_{\begin{matrix}
{K_{1} < k \leq K} \\
{1 \leq n \leq N}
\end{matrix}}{\{{1/\alpha_{k,n}}\}}}}`$. The algorithm then proceeds to use the bisection method to update $`L_{0}`$ so that $`P_{T}'`$ converges to $`P_{T}`$. The outer loop converges because the actual total power consumption $`P_{T}'`$ increases monotonically with $`L_{0}`$ given the target data rates for all DC users are satisfied.
The operation in the inner loop (function: find the optimal $`\{ L_{k},\rho_{k,n}\}`$ at a given $`L_{0}`$) is similar to the algorithm introduced in \[[3](#bib.bib3)\]. Each $`L_{k}`$ is first initialized to the minimum water level needed by DC user $`k`$ to achieve its target data rate, which happens when all the $`N`$ subcarriers in the system are assigned to this user. We then gradually increase $`L_{k}`$ for one of the DC users until the target data rate for this user is satisfied. Increasing $`L_{k}`$ is carried out step by step and the increment at each step is the minimum required value allowing only one more subcarrier to be added to this DC user. During this process, there are chances that the newly added subcarrier is not fully needed and should be time-shared with other users. It is also possible that this subcarrier is not needed at all and it only needs to adjust $`L_{k}`$ using ([19](#S3.E19 "In III-A Optimal Power Allocation for Given Subcarrier Assignment ‣ III Time-Sharing Based Optimal Subcarrier and Power Allocation ‣ Resource Allocation for Delay Differentiated Traffic in Multiuser OFDM Systems")). The algorithm then switches to another DC user and repeats the process until the target data rates for all DC users are satisfied. The inner loop converges because for a given DC user $`k`$, as $`L_{k}`$ increases, $`H_{k,n}`$ for all $`n`$ increases and more $`\rho_{k,n}`$ become one. Hence the data rate $`R_{k}' = {\sum_{n = 1}^{N}{\rho_{k,n}\hspace{0pt}{\lbrack{\log_{2}{({L_{k}\hspace{0pt}\alpha_{k,n}})}}\rbrack}^{+}}}`$ increases. On the other hand, the rates of some other DC users may drop due to their $`\rho_{k,n}`$ changing from one to zero. Nevertheless, as all the $`L_{k}`$’s increase, the rate of each DC user increases and hence the optimal $`L_{k}`$’s can be approached iteratively.
The complexity of the above algorithm lies in the number of iterations needed to update $`L_{0}`$ in the outer loop and the number of iterations to update $`\{ L_{k},\rho_{k,n}\}`$ for each $`L_{0}`$ in the inner loop. Since $`L_{0}`$ is searched by the bisection method, it requires $`\log_{2}{({1/\epsilon})}`$ iterations to converge, where $`\epsilon`$ is the accuracy. The empirical study in the Appendix shows that the exact number of iterations to find the optimal $`\{ L_{k},\rho_{k,n}\}`$ for each $`L_{0}`$ can vary for different values of $`L_{0}`$ and different CNR realizations, but the averaged total number of iterations required to update the set of water levels $`\{ L_{0},L_{1},\ldots,L_{K_{1}}\}`$ in the whole algorithm can be well approximated by $`\mathcal{O}\hspace{0pt}{({{K_{1}^{2}/\sqrt{N}}\hspace{0pt}{\log_{2}{({1/\epsilon})}}})}`$. Since the computational load in each iteration is linear in $`K\hspace{0pt}N`$, the overall complexity of the proposed algorithm is $`\mathcal{O}\hspace{0pt}{({K_{1}^{2}\hspace{0pt}K\hspace{0pt}N^{1/2}\hspace{0pt}{\log_{2}{({1/\epsilon})}}})}`$.
###
III-D
Feasibility and Service Outage
Similar to the delay-limited capacity problem in information theory \[[2](#bib.bib2)\], the constant-rate transmission of DC users considered in this paper can only be guaranteed in a probabilistic manner since the total transmit power is fixed and finite. The service is said to be in an outage if any of the basic rate requirements cannot be satisfied. Thus, the feasibility of the optimization problem in ([3](#S3.E3 "In III Time-Sharing Based Optimal Subcarrier and Power Allocation ‣ Resource Allocation for Delay Differentiated Traffic in Multiuser OFDM Systems"))-([7](#S3.E7 "In III Time-Sharing Based Optimal Subcarrier and Power Allocation ‣ Resource Allocation for Delay Differentiated Traffic in Multiuser OFDM Systems")) is directly related to the condition that $`P_{T} \geq P_{{DC},\min}`$, where $`P_{{DC},\min}`$ is the minimum total power needed to support all $`R_{k}`$’s in the absence of NDC users. Finding $`P_{{DC},\min}`$ reduces to the margin adaption problem \[[3](#bib.bib3)\]. The algorithm proposed in the previous subsection is able to detect the service outage and obtain the outage probability numerically in an efficient way. Specifically, if $`P_{T}'`$ computed in Step b) of the main function is greater than $`P_{T}`$ when $`L_{UB}`$ is still given by the initial value set in Step a), the algorithm will terminate and declare an outage.
When outage occurs, one may ignore all the $`K_{1}`$ DC users and allocate all subcarrier and power resources to the $`K - K_{1}`$ NDC users only. Alternatively, one may ignore one or more DC users from the user list so that at the current channel condition the resources are sufficient to provide the basic rates of the remaining DC users. Those ignored DC users may be re-scheduled for transmission at the next transmission frame by a higher layer scheduler. Further analysis on higher layer scheduling is out of the scope of this work.
## IV Subcarrier and power Allocation using Dual Decomposition
The convex relaxation technique in Section [III](#S3 "III Time-Sharing Based Optimal Subcarrier and Power Allocation ‣ Resource Allocation for Delay Differentiated Traffic in Multiuser OFDM Systems") permits time sharing of each subcarrier. The system model it employs thus differs from the original OFDMA system where only mutually exclusive subcarrier assignment is allowed. As a result, the solution gives an upper bound on the achievable maximum sum-rate of all NDC users under the individual rate requirement for each DC user and the total transmit power constraint. Recently, it is shown in \[[15](#bib.bib15)\] that the duality gap of non-convex resource optimization problems in multicarrier systems is nearly zero if the number of subcarriers is sufficiently large. Thus, the original problem can be solved in the dual domain using decomposition method. Applying this result, the authors in \[[16](#bib.bib16)\] developed efficient algorithms to solve the weighted sum-rate maximization and weighted sum-power minimization problems in the downlink of multiuser OFDM systems. In this section we shall apply the result from \[[15](#bib.bib15)\] and solve our original problem ([2](#S2.E2 "In II System Model and Problem Formulation ‣ Resource Allocation for Delay Differentiated Traffic in Multiuser OFDM Systems")) using the dual decomposition method. Note that the subcarrier and power allocation solution in this section provides a lower bound on the maximum sum-rate of all NDC users when all the target rates for DC users and the total power constraint are satisfied.
Define $`\mathcal{D}`$ as a set of all non-negative $`P_{k,n}`$’s for $`1 \leq k \leq K`$ and $`1 \leq n \leq N`$ such that for each $`n`$ only one $`P_{k,n}`$ is positive. The Lagrange dual function of the problem ([2](#S2.E2 "In II System Model and Problem Formulation ‣ Resource Allocation for Delay Differentiated Traffic in Multiuser OFDM Systems")) is given by:
<table>
<tbody>
<tr>
<td></td>
<td>g(𝜷,μ)𝑔𝜷𝜇\displaystyle g\left(\boldsymbol{\beta},\mu\right)</td>
<td>=\displaystyle=</td>
<td>max{Pk,n}∈𝒟J2({Pk,n},𝜷,μ)subscriptsubscript𝑃𝑘𝑛𝒟subscript𝐽2subscript𝑃𝑘𝑛𝜷𝜇\displaystyle\max_{\{P_{k,n}\}\in{\cal D}}J_{2}\left(\{P_{k,n}\},\boldsymbol{\beta},\mu\right)</td>
<td></td>
</tr>
<tr>
<td></td>
<td></td>
<td>=\displaystyle=</td>
<td>max{Pk,n}∈𝒟[∑k=K1+1K∑n=1Nrk,n+∑k=1K1βk(∑n=1Nrk,n−Rk)\displaystyle\max_{\{P_{k,n}\}\in{\cal D}}\Bigg{[}\sum_{k=K_{1}+1}^{K}\sum_{n=1}^{N}{r_{k,n}}+\sum_{k=1}^{K_{1}}{\beta_{k}}\Big{(}\sum_{n=1}^{N}r_{k,n}-R_{k}\Big{)}</td>
<td></td>
</tr>
<tr>
<td></td>
<td></td>
<td></td>
<td>+μ(PT−∑k=1K∑n=1NPk,n)],\displaystyle+\mu\left(P_{T}-\sum_{k=1}^{K}\sum_{n=1}^{N}P_{k,n}\right)\Bigg{]},</td>
<td></td>
</tr>
</tbody>
</table>
where $`J_{2}\hspace{0pt}{({\{ P_{k,n}\}},{\mathbf{β}},\mu)}`$ is the Lagrangian and the dual variables $`\{{\mathbf{β}},\mu\}`$ are defined in the same way as in ([8](#S3.E8 "In III Time-Sharing Based Optimal Subcarrier and Power Allocation ‣ Resource Allocation for Delay Differentiated Traffic in Multiuser OFDM Systems")). The dual optimization problem is then formulated as:
<table>
<tbody>
<tr>
<td></td>
<td>minimizeminimize\displaystyle\mathrm{minimize}</td>
<td></td>
<td>g(𝜷,μ)𝑔𝜷𝜇\displaystyle{g\left(\boldsymbol{\beta},\mu\right)}</td>
<td></td>
<td>(26)</td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td>subjecttosubjectto\displaystyle\mathrm{subject~{}to}</td>
<td></td>
<td>𝜷⪰0,μ≥0.formulae-sequencesucceeds-or-equals𝜷0𝜇0\displaystyle\boldsymbol{\beta}\succeq 0,\mu\geq 0.</td>
<td></td>
<td></td>
</tr>
</tbody>
</table>
Note that the Lagrangian $`J_{2}\hspace{0pt}{(\ldots)}`$ is linear in $`\beta_{k}`$ and $`\mu`$ for fixed $`P_{k,n}`$, and $`g\hspace{0pt}({\mathbf{β}},\mu)`$ is the maximum of these linear functions, so the dual problem ([26](#S4.E26 "In IV Subcarrier and power Allocation using Dual Decomposition ‣ Resource Allocation for Delay Differentiated Traffic in Multiuser OFDM Systems")) is convex.
To solve the dual problem, we first decompose the dual function into $`N`$ independent optimization problems:
<table>
<tbody>
<tr>
<td></td>
<td>g(𝜷,μ)𝑔𝜷𝜇\displaystyle g\left(\boldsymbol{\beta},\mu\right)</td>
<td>=\displaystyle=</td>
<td>∑n=1Ngn(𝜷,μ)−∑k=1K1βkRk+μPT,superscriptsubscript𝑛1𝑁subscript𝑔𝑛𝜷𝜇superscriptsubscript𝑘1subscript𝐾1subscript𝛽𝑘subscript𝑅𝑘𝜇subscript𝑃𝑇\displaystyle\sum_{n=1}^{N}{g_{n}\left(\boldsymbol{\beta},\mu\right)}-\sum_{k=1}^{K_{1}}{\beta_{k}R_{k}}+\mu P_{T},</td>
<td></td>
<td>(27)</td>
</tr>
</tbody>
</table>
where
<table>
<tbody>
<tr>
<td></td>
<td>gn(𝜷,μ)=max{Pk,n}∈𝒟[∑k=K1+1Krk,n+∑k=1K1βkrk,n−μ∑k=1KPk,n].subscript𝑔𝑛𝜷𝜇subscriptsubscript𝑃𝑘𝑛𝒟superscriptsubscript𝑘subscript𝐾11𝐾subscript𝑟𝑘𝑛superscriptsubscript𝑘1subscript𝐾1subscript𝛽𝑘subscript𝑟𝑘𝑛𝜇superscriptsubscript𝑘1𝐾subscript𝑃𝑘𝑛g_{n}\left(\boldsymbol{\beta},\mu\right)=\max_{\{P_{k,n}\}\in{\cal D}}\left[\sum_{k=K_{1}+1}^{K}r_{k,n}+\sum_{k=1}^{K_{1}}\beta_{k}r_{k,n}-\mu\sum_{k=1}^{K}P_{k,n}\right].</td>
<td></td>
<td>(28)</td>
</tr>
</tbody>
</table>
Suppose subcarrier $`n`$ is assigned to user $`k`$. Using the KKT condition similar to ([11](#S3.E11 "In III Time-Sharing Based Optimal Subcarrier and Power Allocation ‣ Resource Allocation for Delay Differentiated Traffic in Multiuser OFDM Systems")), the optimal $`P_{k,n}^{\ast}`$ that maximizes the object of the max operation in ([28](#S4.E28 "In IV Subcarrier and power Allocation using Dual Decomposition ‣ Resource Allocation for Delay Differentiated Traffic in Multiuser OFDM Systems")) for fixed $`\mathbf{β}`$ and $`\mu`$ can be readily obtained and is given by ([17](#S3.E17 "In III-A Optimal Power Allocation for Given Subcarrier Assignment ‣ III Time-Sharing Based Optimal Subcarrier and Power Allocation ‣ Resource Allocation for Delay Differentiated Traffic in Multiuser OFDM Systems")) when $`1 \leq k \leq K_{1}`$, or ([18](#S3.E18 "In III-A Optimal Power Allocation for Given Subcarrier Assignment ‣ III Time-Sharing Based Optimal Subcarrier and Power Allocation ‣ Resource Allocation for Delay Differentiated Traffic in Multiuser OFDM Systems")) when $`K_{1} < k \leq K`$. Substituting ([17](#S3.E17 "In III-A Optimal Power Allocation for Given Subcarrier Assignment ‣ III Time-Sharing Based Optimal Subcarrier and Power Allocation ‣ Resource Allocation for Delay Differentiated Traffic in Multiuser OFDM Systems")) and ([18](#S3.E18 "In III-A Optimal Power Allocation for Given Subcarrier Assignment ‣ III Time-Sharing Based Optimal Subcarrier and Power Allocation ‣ Resource Allocation for Delay Differentiated Traffic in Multiuser OFDM Systems")) into ([28](#S4.E28 "In IV Subcarrier and power Allocation using Dual Decomposition ‣ Resource Allocation for Delay Differentiated Traffic in Multiuser OFDM Systems")) and comparing all the $`K`$ possible user assignments of this subcarrier, we obtain
<table>
<tbody>
<tr>
<td></td>
<td>gn(𝜷,μ)=max1≤k≤K{β~k[log2(αk,nβ~kμln2)]+−μ(β~kμln2−1αk,n)+},subscript𝑔𝑛𝜷𝜇subscript1𝑘𝐾subscript~𝛽𝑘superscriptdelimited-[]subscript2subscript𝛼𝑘𝑛subscript~𝛽𝑘𝜇2𝜇superscriptsubscript~𝛽𝑘𝜇21subscript𝛼𝑘𝑛g_{n}\left(\boldsymbol{\beta},\mu\right)=\max_{1\leq k\leq K}\left\{\tilde{\beta}_{k}\left[\log_{2}{\left(\frac{\alpha_{k,n}\tilde{\beta}_{k}}{\mu\ln 2}\right)}\right]^{+}-\mu\left(\frac{\tilde{\beta}_{k}}{\mu\ln 2}-\frac{1}{\alpha_{k,n}}\right)^{+}\right\},</td>
<td></td>
<td>(29)</td>
</tr>
</tbody>
</table>
where $`{\overset{\sim}{\beta}}_{k} = \beta_{k}`$ for $`k = {1,\ldots,K_{1}}`$ and $`{\overset{\sim}{\beta}}_{k} = 1`$ for $`k = {{K_{1} + 1},\ldots,K}`$.
Once ([29](#S4.E29 "In IV Subcarrier and power Allocation using Dual Decomposition ‣ Resource Allocation for Delay Differentiated Traffic in Multiuser OFDM Systems")) is solved for all $`n`$’s, the dual function $`g\hspace{0pt}({\mathbf{β}},\mu)`$ can be obtained using ([27](#S4.E27 "In IV Subcarrier and power Allocation using Dual Decomposition ‣ Resource Allocation for Delay Differentiated Traffic in Multiuser OFDM Systems")). Since it is convex, a gradient-type algorithm can minimize $`g\hspace{0pt}({\mathbf{β}},\mu)`$ by updating $`\{{\mathbf{β}},\mu\}`$ simultaneously along some appropriate search directions, which is guaranteed to converge to the optimal solution. In general, $`g\hspace{0pt}({\mathbf{β}},\mu)`$ is not differentiable, and thus its gradient does not exist. Nevertheless, we can resort to the subgradient derived in the following proposition.
*Proposition 1*: For the dual problem ([26](#S4.E26 "In IV Subcarrier and power Allocation using Dual Decomposition ‣ Resource Allocation for Delay Differentiated Traffic in Multiuser OFDM Systems")) with primal defined in ([2](#S2.E2 "In II System Model and Problem Formulation ‣ Resource Allocation for Delay Differentiated Traffic in Multiuser OFDM Systems")), the following is a subgradient of $`g\hspace{0pt}({\mathbf{β}},\mu)`$
<table>
<tbody>
<tr>
<td></td>
<td>ΔβkΔsubscript𝛽𝑘\displaystyle\Delta\beta_{k}</td>
<td>=\displaystyle=</td>
<td>∑n=1Nrk,n∗−Rk,k=1…,K1,formulae-sequencesuperscriptsubscript𝑛1𝑁superscriptsubscript𝑟𝑘𝑛subscript𝑅𝑘𝑘1…subscript𝐾1\displaystyle\sum_{n=1}^{N}{r_{k,n}^{*}}-R_{k},~{}k=1\ldots,K_{1},</td>
<td></td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td>ΔμΔ𝜇\displaystyle\Delta\mu</td>
<td>=\displaystyle=</td>
<td>PT−∑k=1N∑n=1NPk,n∗,subscript𝑃𝑇superscriptsubscript𝑘1𝑁superscriptsubscript𝑛1𝑁superscriptsubscript𝑃𝑘𝑛\displaystyle P_{T}-\sum_{k=1}^{N}\sum_{n=1}^{N}P_{k,n}^{*},</td>
<td></td>
</tr>
</tbody>
</table>
where $`P_{k,n}^{\ast}`$ maximizes the Lagrangian $`J_{2}\hspace{0pt}{(\ldots)}`$ over $`\mathcal{D}`$ at $`\mathbf{β}`$ and $`\mu`$, and $`r_{k,n}^{\ast} = {\log_{2}{({1 + {P_{k,n}^{\ast}\hspace{0pt}\alpha_{k,n}}})}}`$.
###### Proof:
By definition of $`g\hspace{0pt}({\mathbf{β}},\mu)`$ in ([IV](#S4.Ex11 "IV Subcarrier and power Allocation using Dual Decomposition ‣ Resource Allocation for Delay Differentiated Traffic in Multiuser OFDM Systems")):
<table>
<tbody>
<tr>
<td></td>
<td>g(𝜷′,μ′)𝑔superscript𝜷′superscript𝜇′\displaystyle g\left(\boldsymbol{\beta}^{\prime},\mu^{\prime}\right)</td>
<td>≥\displaystyle\geq</td>
<td>∑k=K1+1K∑n=1Nrk,n∗+∑k=1K1βk′(∑n=1Nrk,n∗−Rk)+μ′(PT−∑k=1K∑n=1NPk,n∗)superscriptsubscript𝑘subscript𝐾11𝐾superscriptsubscript𝑛1𝑁superscriptsubscript𝑟𝑘𝑛superscriptsubscript𝑘1subscript𝐾1superscriptsubscript𝛽𝑘′superscriptsubscript𝑛1𝑁superscriptsubscript𝑟𝑘𝑛subscript𝑅𝑘superscript𝜇′subscript𝑃𝑇superscriptsubscript𝑘1𝐾superscriptsubscript𝑛1𝑁superscriptsubscript𝑃𝑘𝑛\displaystyle\sum_{k=K_{1}+1}^{K}\sum_{n=1}^{N}{r_{k,n}^{*}}+\sum_{k=1}^{K_{1}}{\beta_{k}^{\prime}\left(\sum_{n=1}^{N}{r_{k,n}^{*}}-R_{k}\right)}+\mu^{\prime}\left(P_{T}-\sum_{k=1}^{K}\sum_{n=1}^{N}{P_{k,n}^{*}}\right)</td>
<td></td>
</tr>
<tr>
<td></td>
<td></td>
<td>=\displaystyle=</td>
<td>g(𝜷,μ)+∑k=1K1(βk′−βk)(∑n=1Nrk,n∗−Rk)+(μ′−μ)(PT−∑k=1K∑n=1NPk,n∗).𝑔𝜷𝜇superscriptsubscript𝑘1subscript𝐾1superscriptsubscript𝛽𝑘′subscript𝛽𝑘superscriptsubscript𝑛1𝑁superscriptsubscript𝑟𝑘𝑛subscript𝑅𝑘superscript𝜇′𝜇subscript𝑃𝑇superscriptsubscript𝑘1𝐾superscriptsubscript𝑛1𝑁superscriptsubscript𝑃𝑘𝑛\displaystyle g\left(\boldsymbol{\beta},\mu\right)+\sum_{k=1}^{K_{1}}{(\beta_{k}^{\prime}-\beta_{k})\left(\sum_{n=1}^{N}{r_{k,n}^{*}}-R_{k}\right)}+(\mu^{\prime}-\mu)\left(P_{T}-\sum_{k=1}^{K}\sum_{n=1}^{N}{P_{k,n}^{*}}\right).</td>
<td></td>
</tr>
</tbody>
</table>
Proposition 1 is hence proven by using the definition of subgradient. ∎
With the above subgradient, both the subgradient and ellipsoid methods \[[17](#bib.bib17)\] can be used to update $`\{{\mathbf{β}},\mu\}`$. Here we choose the ellipsoid method which converges in $`\mathcal{O}\hspace{0pt}{({({K_{1} + 1})}^{2})}`$ iterations. The algorithm details can be found in \[[17](#bib.bib17)\]. The following lemma leads to a suitable choice of the initial ellipsoid.
*Lemma 1*: The optimal dual variables $`\{{\mathbf{β}}^{\ast},\mu^{\ast}\}`$ must satisfy
<table>
<tbody>
<tr>
<td></td>
<td>00\displaystyle 0</td>
<td>≤\displaystyle\leq</td>
<td>μ∗≤μmax=1ln2maxK1<k≤K1≤n≤N{αk,n},superscript𝜇superscript𝜇12subscriptsubscript𝐾1𝑘𝐾1𝑛𝑁subscript𝛼𝑘𝑛\displaystyle\mu^{*}\leq\mu^{\max}=\frac{1}{\ln{2}}\max_{\begin{subarray}{c}K_{1}<k\leq K\\ 1\leq n\leq N\end{subarray}}{\{\alpha_{k,n}\}},</td>
<td></td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td>00\displaystyle 0</td>
<td>≤\displaystyle\leq</td>
<td>βk∗≤βkmax=maxK1<k≤K1≤n≤N{αk,n}[PT+1min1≤n≤N{αk,n}].superscriptsubscript𝛽𝑘superscriptsubscript𝛽𝑘subscriptsubscript𝐾1𝑘𝐾1𝑛𝑁subscript𝛼𝑘𝑛delimited-[]subscript𝑃𝑇1subscript1𝑛𝑁subscript𝛼𝑘𝑛\displaystyle\beta_{k}^{*}\leq\beta_{k}^{\max}=\max_{\begin{subarray}{c}K_{1}<k\leq K\\ 1\leq n\leq N\end{subarray}}{\{\alpha_{k,n}\}}\left[P_{T}+\frac{1}{\min_{1\leq n\leq N}{\{\alpha_{k,n}\}}}\right].</td>
<td></td>
</tr>
</tbody>
</table>
###### Proof:
The dual variables $`\{{\mathbf{β}}^{\ast},\mu^{\ast}\}`$ must satisfy the KKT conditions in order to be optimal. Taking the partial derivative of $`J_{2}\hspace{0pt}{(\ldots)}`$ in ([IV](#S4.Ex11 "IV Subcarrier and power Allocation using Dual Decomposition ‣ Resource Allocation for Delay Differentiated Traffic in Multiuser OFDM Systems")) with respect to $`P_{k,n}`$ results in
<table>
<tbody>
<tr>
<td></td>
<td>αk,nln2(1+αk,nPk,n)subscript𝛼𝑘𝑛21subscript𝛼𝑘𝑛subscript𝑃𝑘𝑛\displaystyle\frac{\alpha_{k,n}}{\ln{2}\left(1+\alpha_{k,n}P_{k,n}\right)}</td>
<td>=\displaystyle=</td>
<td>μ,𝜇\displaystyle\mu,</td>
<td></td>
<td>(30)</td>
</tr>
</tbody>
</table>
if user $`k`$, for $`K_{1} < k \leq K`$, is active in subcarrier $`n`$, or
<table>
<tbody>
<tr>
<td></td>
<td>βkαk,nln2(1+αk,nPk,n)subscript𝛽𝑘subscript𝛼𝑘𝑛21subscript𝛼𝑘𝑛subscript𝑃𝑘𝑛\displaystyle\frac{\beta_{k}\alpha_{k,n}}{\ln{2}\left(1+\alpha_{k,n}P_{k,n}\right)}</td>
<td>=\displaystyle=</td>
<td>μ,𝜇\displaystyle\mu,</td>
<td></td>
<td>(31)</td>
</tr>
</tbody>
</table>
if user $`k`$, for $`1 \leq k \leq K_{1}`$, is active in subcarrier $`n`$. Since $`P_{k,n}`$ must always satisfy $`0 \leq P_{k,n} \leq P_{T}`$ due to the power constraint, we obtain the upper bound $`\mu^{\max}`$ by letting $`P_{k,n} = 0`$ in ([30](#S4.E30 "In Proof: ‣ IV Subcarrier and power Allocation using Dual Decomposition ‣ Resource Allocation for Delay Differentiated Traffic in Multiuser OFDM Systems")) and the upper bound $`\beta_{k}^{\max}`$ by substituting $`\mu^{\max}`$ into ([31](#S4.E31 "In Proof: ‣ IV Subcarrier and power Allocation using Dual Decomposition ‣ Resource Allocation for Delay Differentiated Traffic in Multiuser OFDM Systems")) and letting $`P_{k,n} = P_{T}`$. ∎
Using Lemma 1, one may choose an initial ellipsoid that encloses the hyper-cuboid where $`\{{\mathbf{β}}^{\ast},\mu^{\ast}\}`$ resides, namely, $`{E\hspace{0pt}{(\mathbf{A}_{0},\mathbf{z}_{0})}} = \left. \{\mathbf{x} \middle| {{{({\mathbf{x} - \mathbf{z}_{0}})}^{T}\hspace{0pt}\mathbf{A}_{0}^{- 1}\hspace{0pt}{({\mathbf{x} - \mathbf{z}_{0}})}} \leq 1}\} \right.`$, where
<table>
<tbody>
<tr>
<td></td>
<td>𝐀0subscript𝐀0\displaystyle\mathbf{A}_{0}</td>
<td>=\displaystyle=</td>
<td>diag[(1+K1)(12β1max)2,⋯,(1+K1)(12βK1max)2,(1+K1)(12μmax)2]diagdelimited-[]1subscript𝐾1superscript12superscriptsubscript𝛽12⋯1subscript𝐾1superscript12superscriptsubscript𝛽subscript𝐾121subscript𝐾1superscript12superscript𝜇2\displaystyle\mathrm{diag}\left[\begin{array}[]{cccc}(1+K_{1})\Big{(}\frac{1}{2}\beta_{1}^{\max}\Big{)}^{2},&\cdots,&(1+K_{1})\Big{(}\frac{1}{2}\beta_{K_{1}}^{\max}\Big{)}^{2},&(1+K_{1})\Big{(}\frac{1}{2}\mu^{\max}\Big{)}^{2}\end{array}\right]</td>
<td></td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td>𝐳0subscript𝐳0\displaystyle\mathbf{z}_{0}</td>
<td>=\displaystyle=</td>
<td>[12β1max,⋯,12βK1max,12μmax]T.superscriptdelimited-[]12superscriptsubscript𝛽1⋯12superscriptsubscript𝛽subscript𝐾112superscript𝜇𝑇\displaystyle\left[\begin{array}[]{cccc}\frac{1}{2}\beta_{1}^{\max},&\cdots,&\frac{1}{2}\beta_{K_{1}}^{\max},&\frac{1}{2}\mu^{\max}\end{array}\right]^{T}.</td>
<td></td>
</tr>
</tbody>
</table>
Due to duality gap, after obtaining the optimal dual variables $`\{{\mathbf{β}}^{\ast},\mu^{\ast}\}`$ that minimize the dual function, it remains to find the optimal primal solutions $`\{ P_{k,n}^{\ast}\}`$ that maximize the Lagrangian $`J_{2}\hspace{0pt}{(\ldots)}`$ and satisfy all the rate and power constraints in the original problem ([2](#S2.E2 "In II System Model and Problem Formulation ‣ Resource Allocation for Delay Differentiated Traffic in Multiuser OFDM Systems")). We can solve this by first identifying the subcarrier assignment $`\{\Omega_{k}^{\ast}\}`$ using ([29](#S4.E29 "In IV Subcarrier and power Allocation using Dual Decomposition ‣ Resource Allocation for Delay Differentiated Traffic in Multiuser OFDM Systems")) with $`\{{\mathbf{β}}^{\ast},\mu^{\ast}\}`$ substituted and then determining the power allocation $`\{ P_{k,n}^{\ast}\}`$ using the results derived in Section III-A.
## V Suboptimal Subcarrier and Power Allocation
In this section we present a suboptimal allocation algorithm that has a much lower computational cost compared with both the optimal iterative algorithm in Section III-C and the dual update algorithm in Section IV. The idea is to first obtain the subcarrier assignment for the DC users by assuming that the power is equally distributed over all subcarriers and that all the NDC users are absent. After that, the power distribution for each DC user over its assigned subcarrier set is individually refined. The purpose of the refinement is to minimize the power while maintaining the basic transmission rates. At last, the residual subcarriers and power are distributed among the NDC users using the optimal resource allocation algorithm in \[[4](#bib.bib4)\] to maximize the sum-rate. This algorithm is suboptimal because the subcarrier assignment for DC users in the first step is obtained by assuming equal power allocation. The decoupling of subcarrier assignment and power allocation for DC users carried out in the first two steps, though being suboptimal, can greatly simplify the complexity and is often used for resource allocation in multiuser OFDM systems such as \[[9](#bib.bib9), [10](#bib.bib10), [18](#bib.bib18)\]
The outline of the proposed suboptimal subcarrier assignment scheme for DC users is presented below.
Suboptimal Subcarrier Allocation Algorithm for DC users
1. 1\.
set $`R_{k}' = 0`$, $`\Omega_{k} = \varnothing`$ for all $`k = {1,\ldots,K_{1}}`$ and $`\mathcal{A} = {\{ 1,2,\ldots,N\}}`$
2. 2\.
while $`\mathcal{A} \neq \varnothing`$ and $`R_{k}' < R_{k}`$ for any $`1 \leq k \leq K_{1}`$
- a\)
find $`k^{\ast}`$ with $`R_{k^{\ast}}' < R_{k^{\ast}}`$ and $`{R_{k^{\ast}}' - R_{k^{\ast}}} \leq {R_{k}' - R_{k}}`$ for all $`1 \leq k \leq K_{1}`$
- b\)
for the found $`k^{\ast}`$, find $`n^{\ast}`$ satisfying $`\alpha_{k^{\ast},n^{\ast}} \geq \alpha_{k^{\ast},n}`$ for $`n \in \mathcal{A}`$
- c\)
for the found $`k^{\ast}`$ and $`n^{\ast}`$, update $`\Omega_{k^{\ast}} = {\Omega_{k^{\ast}} \cup {\{ n^{\ast}\}}}`$, $`\mathcal{A} = {\mathcal{A} - {\{ n^{\ast}\}}}`$ and $`R_{k^{\ast}}' = {R_{k^{\ast}}' + {\log_{2}{({1 + \frac{\alpha_{k^{\ast},n^{\ast}}\hspace{0pt}P_{T}}{N}})}}}`$
At each iteration of Step $`2)`$ in the above algorithm, the DC user whose current data rate $`R_{k}'`$ is the farthest away from its target rate $`R_{k}`$ will be allowed to pick a new subcarrier from the available subcarrier set. Preferably, the subcarrier with the highest CNR will be chosen.
Upon acquiring $`\Omega_{k}`$ for $`1 \leq k \leq K_{1}`$, the power distribution for each DC user is then adjusted using the analytical solution ([17](#S3.E17 "In III-A Optimal Power Allocation for Given Subcarrier Assignment ‣ III Time-Sharing Based Optimal Subcarrier and Power Allocation ‣ Resource Allocation for Delay Differentiated Traffic in Multiuser OFDM Systems")) and ([20](#S3.E20 "In III-A Optimal Power Allocation for Given Subcarrier Assignment ‣ III Time-Sharing Based Optimal Subcarrier and Power Allocation ‣ Resource Allocation for Delay Differentiated Traffic in Multiuser OFDM Systems")). In the case where $`g_{k} < {|\Omega_{k}|}`$ for some $`k`$, the above suboptimal subcarrier assignment algorithm over-allocates subcarriers to DC user $`k`$. To efficiently utilize all the wireless resources, the remaining $`{|\Omega_{k}|} - g_{k}`$ subcarriers should be returned to the residual subcarrier set $`\mathcal{A}`$, which will be distributed among the $`K - K_{1}`$ NDC users. Let $`P_{{DC},T}`$ denote the actual power consumption of all the $`K_{1}`$ DC users. If $`P_{{DC},T}`$ is larger than the total power limit $`P_{T}`$, a service outage occurs. Otherwise, the residual transmit power $`P_{T} - P_{{DC},T}`$ together with the residual subcarrier set $`\mathcal{A}`$ are allocated over the $`K - K_{1}`$ NDC users. Specifically, each subcarrier in $`\mathcal{A}`$ is assigned to the NDC user with the highest CNR, and the power is distributed over these subcarriers in the form of water-filling ([18](#S3.E18 "In III-A Optimal Power Allocation for Given Subcarrier Assignment ‣ III Time-Sharing Based Optimal Subcarrier and Power Allocation ‣ Resource Allocation for Delay Differentiated Traffic in Multiuser OFDM Systems")), where the water level can be determined by $`P_{T} - P_{{DC},T}`$.
The number of iterations involved in finding the suboptimal $`\Omega_{k}`$’s for $`k = {1,\ldots,K}`$ is limited by $`N`$ since $`N`$ is the total number of subcarriers available. That is, the proposed suboptimal algorithm only performs a fixed number of iterations rather than iterating till convergence. The power allocation for given $`\{\Omega_{k}\}`$ has explicit analytical solution as shown in Section III-A and its complexity is linear in $`K\hspace{0pt}N`$. Therefore, the overall complexity of this suboptimal algorithm is only linear in $`K`$ and $`N`$.
## VI Numerical Results
In this section we present numerical performance results of the proposed optimal and suboptimal resource allocation algorithms. We consider a multiuser OFDM system with $`N = 64`$ subcarriers and $`K = 8`$ users. Therein, $`K_{1} = 4`$ users have DC traffic and the others have NDC traffic. For simplicity, we let the rate requirements of all DC users be identical and equal to $`R_{DC}/K_{1}`$ bits/OFDM symbol, where $`R_{DC}`$ denotes the sum of the basic rates. In all simulations, the channel from the base station to each user terminal is modeled by the HiperLan/2 channel model A \[[19](#bib.bib19)\], which is an 8-tap channel with exponential power delay profile, 20MHz sampling frequency and 50ns rms (root-mean-square) delay spread. The channels for different users are assumed to be independent. We also assume that the path losses from the base station to all user terminals are the same. The average channel gain on each subcarrier is normalized. The system total transmit SNR is defined as $`P_{T}/{({N_{0}\hspace{0pt}B})}`$. The SNR gap in the rate function ([1](#S2.E1 "In II System Model and Problem Formulation ‣ Resource Allocation for Delay Differentiated Traffic in Multiuser OFDM Systems")) is set to 6.6 (8.2 dB) for both DC users and NDC users. In practice, when uncoded QAM constellation is used the SNR gap of 8.2 dB corresponds to a BER requirement of $`10^{- 5}`$.
To evaluate the performance of the three proposed adaptive resource allocation algorithms, we also present the results for two non-adaptive schemes in comparison. In both schemes, the subcarrier assignment is predetermined but the power allocation for each user over its predetermined set of subcarriers follows the optimal approach derived in Section [](#S3.SS1 "III-A Optimal Power Allocation for Given Subcarrier Assignment ‣ III Time-Sharing Based Optimal Subcarrier and Power Allocation ‣ Resource Allocation for Delay Differentiated Traffic in Multiuser OFDM Systems")
III-A
. In the first scheme, all the 8 users are treated equally and each is assigned 8 subcarriers. We refer to this scheme as Fixed Subcarrier Assignment with Optimal Power Allocation (FSA-OPA). In the other scheme, DC users are given higher priority than NDC users and each is assigned 12 subcarriers, whereas each NDC user is allocated 4 subcarriers only. This scheme is called Fixed Subcarrier Assignment with *Priority* and with Optimal Power Allocation (FSAP-OPA). In addition, for both schemes, we let the predetermined subcarriers for each user spread over the entire bandwidth in a comb pattern \[[20](#bib.bib20)\]. This prevents the situation where all subcarriers of a user are in deep fade.
We first compare the performance in terms of service outage behavior. Fig. [3](#A0.F3 "Figure 3 ‣ Resource Allocation for Delay Differentiated Traffic in Multiuser OFDM Systems") illustrates the service outage probability versus total transmit SNR when the total target transmission rate of DC users is $`R_{DC} = 80`$ bits/OFDM symbol. It is first observed that the time-sharing based optimal algorithm and the dual method perform almost identically. This suggests that two algorithms result in almost the same subcarrier assignment solution for DC users. This observation is expected because the duality gap vanishes when $`N`$ is sufficiently large and, as a result, both the upper bound given by the optimal algorithm with time sharing and the lower bound given by the dual method approach the truly optimal solution. One can also see that the performance loss due to the suboptimal subcarrier assignment in the suboptimal algorithm is marginal. In particular, at a service outage probability of $`1\%`$, the SNR loss is within $`0.5`$ dB. In addition, it is seen from Fig. [3](#A0.F3 "Figure 3 ‣ Resource Allocation for Delay Differentiated Traffic in Multiuser OFDM Systems") that the proposed adaptive algorithms significantly outperform the two fixed subcarrier assignment (FSA) schemes. At moderate and high SNR regions, the service outage probability is more than an order of magnitude lower. Besides, the FSA scheme with priority outperforms the one without priority as more subcarriers are assigned to DC users in the former.
In Fig. [4](#A0.F4 "Figure 4 ‣ Resource Allocation for Delay Differentiated Traffic in Multiuser OFDM Systems"), we plot the minimum required total transmit SNR for different $`R_{DC}`$ at a given service outage probability of $`1\%`$. It is again observed that the optimal algorithm with time sharing and the dual method have almost identical performance. Therefore, only the results of the former will be illustrated hereafter. From Fig. [4](#A0.F4 "Figure 4 ‣ Resource Allocation for Delay Differentiated Traffic in Multiuser OFDM Systems") we observe that, for a wide range of $`R_{DC}`$ that the multiuser OFDM system can support with $`1\%`$ outage probability, the difference on the minimum required SNR between the optimal and suboptimal algorithms is consistently less than $`0.5`$ dB. In particular, as $`R_{DC}`$ decreases, the performance of the suboptimal algorithm becomes closer to that of the optimal algorithm. This is expected as the suboptimality of the proposed suboptimal algorithm lies only in the subcarrier allocation for DC users. If the rate requirement for DC users is small, the suboptimal algorithm will become nearly optimal. Fig. [4](#A0.F4 "Figure 4 ‣ Resource Allocation for Delay Differentiated Traffic in Multiuser OFDM Systems") also shows that, as $`R_{DC}`$ increases, the minimum required total SNR of the proposed adaptive algorithms increases at a much lower speed than that of the two FSA schemes.
We next study the achievable transmission rates of the heterogenous multiuser OFDM system with the proposed adaptive subcarrier and power allocation algorithms. Fig. [5](#A0.F5 "Figure 5 ‣ Resource Allocation for Delay Differentiated Traffic in Multiuser OFDM Systems") shows the achievable pairs of the basic sum-rate for DC traffic $`R_{DC}`$ and the average sum-rate for NDC traffic $`{\overline{R}}_{NDC}`$ at a total transmit SNR of $`20`$ dB. The average sum-rate for NDC traffic $`{\overline{R}}_{NDC}`$ is obtained by averaging the instantaneous sum-rates of NDC users over $`500`$ independent channel realizations. To ensure a service outage probability of $`1\%`$ or below, the maximum value of $`R_{DC}`$ in our simulation is set to $`176`$ bits/OFDM symbol for the proposed algorithms, and to $`112`$ and $`80`$ bits/OFDM symbols, respectively, for FSAP-OPA and FSA-OPA. The maximum achievable $`R_{DC}`$ with an acceptable service outage probability, for example $`1\%`$, at a given total SNR can be obtained from Fig. [4](#A0.F4 "Figure 4 ‣ Resource Allocation for Delay Differentiated Traffic in Multiuser OFDM Systems"). From Fig. [5](#A0.F5 "Figure 5 ‣ Resource Allocation for Delay Differentiated Traffic in Multiuser OFDM Systems") one can observe that, compared with the optimal subcarrier and power allocation algorithm, the loss of the average achievable NDC traffic rate at a given $`R_{DC}`$ by using the suboptimal algorithm is within $`{2\%} \sim {9\%}`$. On the other hand, compared with the two FSA schemes, both the proposed optimal and suboptimal adaptive algorithms demonstrate substantially larger achievable rate regions. We also observe that, at the same $`R_{DC}`$, the $`{\overline{R}}_{NDC}`$ of FSA-OPA is larger than that of FSAP-OPA. This is because NDC users have fewer subcarriers in FSAP-OPA. However, the maximum $`R_{DC}`$ FSAP-OPA can support is larger than that of FSA-OPA.
Finally, we demonstrate the multiuser diversity exploited by our algorithms. We let the number of DC users in the system be fixed at $`K_{1} = 4`$ and vary the number of NDC users between $`4`$ and $`16`$. Fig. [6](#A0.F6 "Figure 6 ‣ Resource Allocation for Delay Differentiated Traffic in Multiuser OFDM Systems") presents the average sum-rate $`{\overline{R}}_{NDC}`$ as a function of the number of NDC users at $`R_{DC} = 32`$ bits/OFDM symbols. Same to Fig. [5](#A0.F5 "Figure 5 ‣ Resource Allocation for Delay Differentiated Traffic in Multiuser OFDM Systems"), the total transmit SNR is $`20`$ dB and $`{\overline{R}}_{NDC}`$ is obtained by averaging $`500`$ independent channel realizations. In the two FSA schemes, the subcarrier allocation for DC users is the same as before, but the rest of the subcarriers are all allocated to one NDC user, which is selected in a round-robin fashion at each transmission frame. The values of $`{\overline{R}}_{NDC}`$ for the two FSA schemes remain constant since no multiuser diversity is achieved. On the contrary, $`{\overline{R}}_{NDC}`$ obtained by the proposed adaptive algorithms increases as the number of NDC users increases, which clearly shows the multiuser diversity. In particular, the achievable $`{\overline{R}}_{NDC}`$ of the optimal algorithm is about $`110\%`$ and $`140\%`$ higher than that of the FSAP-OPA scheme when the system has $`4`$ and $`16`$ NDC users, respectively.
## VII Conclusion and Discussions
Supporting communication services with diverse QoS requirements in future broadband wireless networks is crucial. This paper considered the resource allocation problem in an OFDM based downlink system that supports simultaneous transmission of users with DC traffic at constant rates and users with NDC traffic at variable rates. We investigated this problem from the physical layer perspective and aimed to maximize the sum-rate of NDC traffic while maintaining individual basic rates of DC traffic for each channel realization under a total power constraint. It was shown that the optimal power allocation over the subcarriers in such a heterogeneous system has the form of multi-level water-filling; moreover, the set of valid user candidates competing for each subcarrier consists of only one NDC user but all DC users. We converted this combinatorial problem with exponential complexity into a convex problem using the time-sharing technique and developed an efficient iterative algorithm with polynomial complexity. We also solved the original problem using dual decomposition method which leads to polynomial complexity as well. To further speed up the resource allocation and make it more suitable for practical systems, we then proposed a suboptimal algorithm whose computation load is only linear in the number of users and subcarriers in the system.
The performance of our algorithms was evaluated in terms of service outage probability, achievable DC and NDC traffic rate pairs, and multiuser diversity. The numerical results showed that the convex relaxation technique with time sharing and the dual decomposition approach obtained almost the same solution and that the suboptimal algorithm has the near optimal performance. Results also demonstrated that the proposed adaptive subcarrier and power allocation algorithms significantly outperform the schemes with adaptive power allocation but fixed subcarrier assignment.
This paper adopted the continuous rate function ([1](#S2.E1 "In II System Model and Problem Formulation ‣ Resource Allocation for Delay Differentiated Traffic in Multiuser OFDM Systems")), which greatly helped to derive the insights of optimal resource allocation. If discrete rates are used in practical systems, our algorithms can be modified accordingly. In particular, since the proposed suboptimal algorithm has near-optimal performance at significantly lower complexity, it is more desirable to modify the suboptimal one. For instance, one can obtain the subcarrier assignment using the proposed suboptimal algorithm and then apply the greedy bit loading algorithm for each single user as in \[[3](#bib.bib3)\]. Nevertheless, our continuous rate formulation provides the performance upper bound for systems with discrete rates.
We have also assumed that the channels from the base station to all the NDC users have the same path loss. By symmetry, our formulation also equalizes the long-term average throughput among all the NDC user. To achieve fairness when their channel path losses are different, we can simply modify our cost function ([2](#S2.E2 "In II System Model and Problem Formulation ‣ Resource Allocation for Delay Differentiated Traffic in Multiuser OFDM Systems")) by dividing the channel-to-noise ratio with the path loss. By doing so, the effective channel gains for all NDC users are normalized. Therefore, only the user whose current channel condition is at its peak level will be selected to compete with DC users for each subcarrier. This is similar to the concept of “riding on the channel peak” in opportunistic scheduling.
\[Empirical Study on the convergence speed of the iterative algorithm in Section III-C\]
Simulation settings: $`P_{T} = 100`$, $`2`$ NDC users, $`R_{k} = 16`$ bits/ OFDM symbol for $`1 \leq k \leq K_{1}`$, $`K_{1} \in {\{ 1,\ldots,12\}}`$, $`N \in {\{ 16,32,64,128\}}`$. $`\alpha_{k,n} = {N\hspace{0pt}{|h_{k,n}|}^{2}}`$ are randomly generated with $`h_{k,n}`$ modeled as complex Gaussian variables of zero mean and unit variance and independent for all $`k`$ and $`n`$. The accuracy of bisection-searching $`L_{0}`$ is set to $`\epsilon = 10^{- 7}`$, and it leads to $`26`$ iterations in the main function throughout this simulation study. The parameter $`\Delta`$ in Step 3b) of the inner function is set to $`0.005`$. Note that it typically takes very few iterations in the while loop of Step 3c) in the inner function to find the $`L_{k^{\ast}}`$ for DC user $`k^{\ast}`$ that meets its rate requirement. Thus we choose to count the number of times it repeats for Steps 2)-3) as the number of iterations to update $`\{ L_{k},\rho_{k,n}\}`$ at a given $`L_{0}`$. Fig. [7](#A0.F7 "Figure 7 ‣ Resource Allocation for Delay Differentiated Traffic in Multiuser OFDM Systems") shows the snapshot of iterations to update $`\{ L_{k},\rho_{k,n}\}`$ at each updating step of $`L_{0}`$ for three random channel realizations. The number of iterations varies for different $`L_{0}`$ and different channel realizations and, in general, more iterations are needed when $`K_{1}`$ increases. To extract the rules on how the number of iterations change with $`N`$ and $`K_{1}`$, we plot in Fig. [8](#A0.F8 "Figure 8 ‣ Resource Allocation for Delay Differentiated Traffic in Multiuser OFDM Systems") the averaged total iterations needed to find the set of optimal solutions $`\{ L_{0},L_{k},\rho_{k,n}\}`$, where each value is obtained by averaging over $`20`$ independent channel realizations. For comparison we also plot the curves generated using the analytical expression $`{c\hspace{0pt}K_{1}^{2}}/\sqrt{N}`$ with the constant $`c`$ being $`c = {26 \times 5.1} = 132.6`$. It is observed that the analytical expression provides a very good approximation on the shape of the simulated curves. Therefore, we conclude that the proposed time-sharing based optimal subcarrier assignment algorithm converges in $`\mathcal{O}\hspace{0pt}{({{K_{1}^{2}/\sqrt{N}}\hspace{0pt}{\log_{2}{({1/\epsilon})}}})}`$ iterations.
## References
- \[1\] A. J. Goldsmith and P. Varaiya, “Capacity of fading channels with channel side information,” *IEEE Trans. on Infor. Theory*, vol. 43, no. 6, pp. 1986–1992, Nov. 1997.
- \[2\] G. Caire, G. Taricco, and E. Biglieri, “Optimal power control over fading channels,” *IEEE Trans. on Infor. Theory*, vol. 45, no. 5, pp. 1468–1489, July 1999.
- \[3\] C. Y. Wong, R. S. Cheng, K. B. Letaief, and R. D. Murch, “Multiuser OFDM with adaptive subcarrier, bit and power allocation,” *IEEE Journal on Selected Areas in Comm.*, vol. 17, no. 10, pp. 1747–1758, Oct. 1999.
- \[4\] J. Jang and K. B. Lee, “Transmit power adaptation for multiuser OFDM systems,” *IEEE Journal on Selected Areas in Comm.*, vol. 21, no. 2, pp. 171–178, Feb. 2003.
- \[5\] P. W. C. Chan and R. S. Cheng, “Optimal power allocation in zero-forcing MIMO-OFDM downlink with multiuser diversity,” in *Proc. of IST Mobile & Wireless Communications Summit*, Dresden, June 2005.
- \[6\] Y. M. Tsang and R. S. Cheng, “Optimal resouce allocation in SDMA/MIMO/OFDM systems under QoS and power constraints,” in *Proc. of IEEE WCNC*, 2004.
- \[7\] J. M. Cioffi, “Digital communications,” *EE379 Course Reader, Stanford University*, 2003.
- \[8\] Y. J. Zhang and K. B. Letaief, “An efficient resource allocation scheme for spatial multiuser access in MIMO/OFDM systems,” *IEEE Trans. Comm.*, vol. 53, no. 1, pp. 107–116, Jan. 2005.
- \[9\] W. Rhee and J. M. Cioffi, “Increase in capacity of multiuser OFDM system using dynamic subchannel allocation,” in *Proc. of IEEE VTC*, 2000.
- \[10\] Z. Shen, J. G. Andrews, and B. L. Evans, “Adaptive resource allocation in multiuser OFDM systems with proportional fairness,” *IEEE Trans. on Wireless Comm.*, vol. 4, no. 6, pp. 2726–2737, Nov. 2005.
- \[11\] G. Song and Y. Li, “Cross-layer optimization for OFDM wireless networks: Part II: Algorithm development,” *IEEE Trans. on Wireless Comm.*, vol. 4, no. 2, pp. 625–634, March 2005.
- \[12\] M. Anas, K. Kim, S. Shin, and K. Kim, “QoS aware power allocation for combined guaranteed performance and best effort users in OFDMA systems,” in *Proc. International Symposium on Intelligent Signal Processing and Comm. Systems*, Nov. 2004.
- \[13\] A. J. Goldsmith and M. Effros, “The capacity region of broadcast channels with intersymbol interference and colored Gaussian noise,” *IEEE Trans. on Infor. Theory*, vol. 47, pp. 219–240, 2001.
- \[14\] S. Boyd and L. Vandenberghe, *Convex Optimization*. Cambridge University Press, 2004.
- \[15\] W. Yu and R. Lui, “Dual methods for nonconvex spectrum optimization of multicarrier systems,” *IEEE Trans. on Comm.*, vol. 54, no. 7, pp. 1310–1322, July 2006.
- \[16\] K. Seong, M. Mhoseni, and J. M. Cioffi, “Optimal resource allocation for OFDMA downlink systems,” in *Proc. of ISIT’06*, Seattle, USA, July 2006.
- \[17\] S. Boyd, “Convex optimization ii,” *EE364B Course Note, Stanford University. Available at http://www.stanford.edu/class/ee364b/*, 2006.
- \[18\] K. Kim, H. Kang, and K. Kim, “Providing quality of service in adaptive resource allocation for OFDMA systems,” in *Proc. of IEEE VTC’04*, vol. 3, May 2004, pp. 1612–1615.
- \[19\] E. N. Committee, “Channel models for HIPERLAN/2 in different indoor scenarios,” *European Telecom. Standards Institute, Sophia-Antipolis, Valbonne, France, Norme ETSI, doc 3ERI085B*, 1998.
- \[20\] E. Lawrey, “Multiuser OFDM,” in *Proc. of International Symposium on Signal Processing and its Applications (ISSPA’99)*, Brisbane, Australia, Aug. 1999.
[Refer to caption](/html/0710.3279/assets/x1.png)
Figure 1: Block diagram of a downlink multiuser OFDM system
[Refer to caption](/html/0710.3279/assets/x2.png)
Figure 2: Illustration of multi-level water-filling for given subcarrier assignment in a multiuser OFDM system with 2 DC users, 2 NDC users and 8 subcarriers.
[Refer to caption](/html/0710.3279/assets/x3.png)
Figure 3: Service outage probability versus total transmit SNR at $`R_{DC} = 80`$ bits/OFDM symbol.
[Refer to caption](/html/0710.3279/assets/x4.png)
Figure 4: Minimum required total transmit SNR versus $`R_{DC}`$ at a service outage probability of $`1\%`$.
[Refer to caption](/html/0710.3279/assets/x5.png)
Figure 5: Achievable ($`R_{DC}`$, $`{\overline{R}}_{NDC}`$) rate pairs at a total transmit SNR of 20 dB.
[Refer to caption](/html/0710.3279/assets/x6.png)
Figure 6: Achievable $`{\overline{R}}_{NDC}`$ versus the number of NDC users at a total transmit SNR of 20 dB with $`4`$ DC users.
[Refer to caption](/html/0710.3279/assets/x7.png)
Figure 7: Iterations required to update $`\{ L_{k},\rho_{k,n}\}`$ at each $`L_{0}`$ update.
[Refer to caption](/html/0710.3279/assets/x8.png)
Figure 8: Total iterations required to update $`\{ L_{0},L_{k},\rho_{k,n}\}`$. |
0711.0892 | cs.NI | cyber | # Routing in Outer Space: Improved Security and Energy-Efficiency in Multi-Hop Wireless Networks
Alessandro Mei
Department of Computer Science
Sapienza University of Rome, Italy
Email: mei@di.uniroma1.it
Julinda Stefa
Department of Computer Science
Sapienza University of Rome, Italy
Email: stefa@di.uniroma1.it
###### Abstract
In this paper we consider security-related and energy-efficiency issues in multi-hop wireless networks. We start our work from the observation, known in the literature, that shortest path routing creates congested areas in multi-hop wireless networks. These areas are critical—they generate both security and energy efficiency issues. We attack these problems and set out routing in outer space, a new routing mechanism that transforms any shortest path routing protocol (or approximated versions of it) into a new protocol that, in case of uniform traffic, guarantees that every node of the network is responsible for relaying the same number of messages, on expectation. We can show that a network that uses routing in outer space does not have congested areas, does not have the associated security-related issues, does not encourage selfish positioning, and, in spite of using more energy globally, lives longer of the same network using the original routing protocol.
###### Index Terms:
Multi-hop wireless networks, energy-efficiency, routing, load-balancing, analysis, simulations.
## I Introduction
During the past years the interest in multi-hop wireless networks has been growing significantly. These types of networks have an important functionality that is the possibility to use other nodes as relays in order to deliver messages and data from sources to destinations. This functionality makes multi-hop wireless networks not only scalable but also usable in various areas and contexts. One of the most representative and important examples of multi-hop wireless networks are wireless sensor networks where small devices equipped with a radio transmitter and a battery are deployed in an geographic area for monitoring or measuring of some desired property like temperature, pressure etc. \[[1](#bib.bib1), [2](#bib.bib2)\].
The routing on a wireless sensor network is one of the most interesting and difficult issues to solve due to the limited resources and capacities of the nodes. Hence, protocols that use less information possible and need minimal energy consumption of nodes have become more than valuable in this context. Much research work has been devoted to finding energy-efficient routing protocols for this kind of networks. Often, these protocols tend to find an approximation of the shortest path between the source and the destination of the message. In \[[3](#bib.bib3)\], the authors analyze the impact of shortest path routing in a large multi-hop wireless network. They show that relay traffic induces congested areas in the network. If the traffic pattern is uniform, i. e. every message has a random source and a random destination uniformly and independently chosen, and the network area is a disk, then the center of the disk is a congested area, where the nodes has to relay much more messages that the other nodes of the network. We have the same problem if the network area is a square, or a rectangle, or any other two-dimensional convex surface. Our experiments show that, when using geographic routing \[[4](#bib.bib4)\] on a network deployed in a square, 25% of the messages are relayed by the nodes in a small central congested region whose area is $`3\%`$ of the total area of the square.
Congested areas are bad for a number of important reasons. They raise security-related issues: If a large number of messages are relayed by the nodes deployed in a relatively small congested region, then jamming can be a vicious attack. It is usually expensive to jam a large geographical area, it is much cheaper and effective to jam a small congested region. In the square, for example, it is enough to jam 3% of the network area to stop 25% of the messages. Moreover, if an attacker has the goal of getting control over as many communications as possible, then it is enough to control 3% of the network nodes to handle 25% of the messages.
There are also energy-efficiency issues: Aside from re-transmissions, that are costly and, in congested areas, often more frequent, the nodes have to relay a much larger number of messages. Therefore these nodes will die earlier than the other nodes in the network, exacerbating the problem for the nodes in the congested region that are still operational. In the long run, this results in holes in the network and in a faster, and less graceful, death of the system. Note that these problems are not solved by trying to balance the load just locally, as done by a few protocols in the literature (like GEAR \[[5](#bib.bib5)\], for example)—these protocols are useful, they can be used in any case (in our protocols as well), and are efficient in smoothing the energy requirements among neighbors, while they can’t do much against congested areas and they don’t help to alleviate the above discussed security-related issues.
Lastly, there are other concerns in the contexts where the nodes are carried by individual independent entities. In this paper we do not consider mobility. However, if the position of the node can be chosen by the node itself in such a way to maximize its own advantage, and if energy is an issue, then no node would stay in the center of the square, the highly congested region. If the nodes are selfish, an uneven distribution of load in the network area leads to an irregular distribution of the nodes—there is no point in positioning in the place where the battery is going to last the shortest. Selfish behavior is a recent concern in the network community and it is rapidly gaining importance \[[4](#bib.bib4), [6](#bib.bib6), [7](#bib.bib7)\]. Most of these contributions show how to devise mechanisms such that selfish nodes can’t help but truthfully execute the protocol. For the best of our knowledge, here we are raising a new concern, that can be important in mobile networks or whenever the position of the node can be an independent and selfish choice.
Solving these issues—security, energy-efficiency, and tolerance to (a particular case of) selfish behavior—is an important and non-trivial problem, and, at least partially, our goal. In this paper we attack this problem and set out *routing in outer space*, a new routing mechanism that transforms any shortest path routing protocol (or approximated versions of it) into a new protocol that, in case of uniform traffic, guarantees that every node of the network is responsible for relaying the same number of messages, on expectation. We can show that a network that uses routing in outer space does not have congested areas, does not have the associated security issues, and, in spite of using more energy globally, lives longer of the same network using the original routing protocol—that is, it is more energy-efficient. We support our claims by showing routing in outer space based on geographic routing, and performing a large set of experiments.
The rest of the paper is organized as follows: In Section [II](#S2 "II Related Work ‣ Routing in Outer Space: Improved Security and Energy-Efficiency in Multi-Hop Wireless Networks") we report on the relevant literature in this area; in Section [III](#S3 "III Routing in Outer Space ‣ Routing in Outer Space: Improved Security and Energy-Efficiency in Multi-Hop Wireless Networks") we present the theoretical idea behind our work, we come up with routing in outer space and prove its mathematical properties; In Section [IV](#S4 "IV Routing in Outer Space in Practice ‣ Routing in Outer Space: Improved Security and Energy-Efficiency in Multi-Hop Wireless Networks"), after describing our node and network assumptions and our simulation environment, we discuss on on the practical issues related in implementing routing in outer space starting from geographic routing; lastly, we present an extensive set of experiments, fully supporting our claims.
## II Related Work
Routing in multi-hop wireless networks is one of most important, interesting, and challenging tasks due to network devices limitations and network dynamics. As a matter of fact this is one of the most studied topics in this area, and the literature on routing protocols for multi-hop wireless networks is vast. There have been proposed protocols that maintain routes continuously (based on distance vector) \[[8](#bib.bib8), [9](#bib.bib9), [10](#bib.bib10)\], that create routes on-demand \[[11](#bib.bib11), [12](#bib.bib12), [13](#bib.bib13)\] or a hybrid \[[14](#bib.bib14)\]. For a good survey and comparison see \[[15](#bib.bib15), [16](#bib.bib16)\]. Other examples of routing protocols for multi-hop wireless networks are those based on link-state like OLSR \[[17](#bib.bib17)\], etc.
Geographic routing or position-based routing, where nodes locally decide the next relay on the basis based on information obtained through some GPS (Global Positioning System) or other location determination techniques \[[18](#bib.bib18)\], seems to be one of the most feasible and studied approach. Examples of reasearch work on this approach are protocols like GEAR (Geographical Energy Aware Routing) \[[5](#bib.bib5)\], GAF (Geographical Adaptive Fidelity) \[[19](#bib.bib19)\]. For a good starting survey see \[[20](#bib.bib20)\].
All these protocols try to approximate shortest path between sources and destinations over the network. In \[[21](#bib.bib21)\], the authors analytically study the impact of shortest single path routing on a node by approximating single paths to line segments and characterizing the deviation of routes from line segments. However, one of the parameters in their model is not analytically quantified and would need to be estimated via simulations. In \[[22](#bib.bib22)\] the authors introduce an analytical model to evaluate the imposed load at a node by approximating a shortest path route to a narrow rectangle, where the load of a node is defined as the number of paths going through the node. The length of the rectangle is defined as the distance between source and destination and the width of the rectangle parameterizes the deviation of paths from the line segments between sources and destinations. However, since the number of nodes is not parameterized in the analytic model, the model does not provide an explicit relationship between the load and the number of nodes. In \[[3](#bib.bib3)\], the authors analyze the load for a homogeneous multi-hop wireless network for the case of straight line routing as in \[[23](#bib.bib23), [24](#bib.bib24), [21](#bib.bib21), [22](#bib.bib22), [25](#bib.bib25), [3](#bib.bib3)\] (where Shortest path routing is approximated to straight line routing in large multi-hop wireless networks). Assuming uniform traffic, in \[[3](#bib.bib3)\] the authors prove that relays induce so called hotspots or congested areas in the network. The creation of these types of congested areas not only affect the overall network throughput by generating energy-efficiency problems, but also create the serious security problems that we mentioned before. Of course, geographic routing (which, in dense networks, approximates the shortest path between source and destination) also suffers of the same problems. A lot of work has been done regarding to the energy-efficiency issues, and several approaches that try to solve the problem locally have been proposed like \[[5](#bib.bib5), [26](#bib.bib26), [27](#bib.bib27)\].
## III Routing in Outer Space
We model the multi-hop wireless network as a undirected graph $`G = {(V,E)}`$, where $`V`$ is the set of nodes and $`E`$ is the set of edges. The nodes are ad-hoc deployed on the network area $`S`$. Formally, it is enough to assume that $`S`$ is a metric space with distance $`d_{S}`$ and that every node is a point on $`S`$. Given two nodes $`{u,v} \in V`$ deployed on $`S`$, we will denote the distance between their positions on the space with $`d_{S}\hspace{0pt}{(u,v)}`$. The nodes have a transmission range $`r`$—two nodes $`{u,v} \in V`$ are connected by a wireless link $`{u\hspace{0pt}v} \in E`$ if $`{d_{S}\hspace{0pt}{(u,v)}} \leq r`$, that is, their distance is at most $`r`$. The common practice in the literature is to take a convex surface as $`S`$, usually a square, a rectangle, or a disk, with the usual Euclidean distance. In this paper we assume that the nodes know their position, either by being equipped with a GPS unit, or by using one of the many localization protocols \[[28](#bib.bib28), [29](#bib.bib29)\]. And that they know the boundaries of the network area $`S`$; this is possible either by pre-loading this information on the nodes before deployment, or by using one of the protocols in \[[30](#bib.bib30), [31](#bib.bib31), [32](#bib.bib32)\].
We started from the observation that shortest path routing on the square, or even an approximate version of it, generates congested areas on the center of the network. We have already discussed that this phenomenon is not desirable. The same problem is present on the rectangle, on the disk, and on any two dimensional convex deployment of the network, which is the common case in practice. Here, the idea is to relinquish shortest paths so as to get rid of congested areas, with the goal of improving security, energy efficiency, and tolerance to selfish behavior of the multi-hop wireless network. As the first step, we have to realize that there do exist metric spaces that do not present the problem. First, we need a formal definition of the key property of the metric space we are looking for.
######
Definition 1
Consider a multi-hop wireless network deployed on a space $`S`$. Fix a node $`u`$ and choose its position on $`S`$ arbitrarily. Then, deploy the other nodes of the network uniformly and independently at random. We will say that $`S`$ is *symmetric* if, chosen two nodes $`v_{1}`$ and $`v_{2}`$ uniformly at random in the network, the probability that $`u`$ is on the shortest path from $`v_{1}`$ to $`v_{2}`$ does not depend on its position.
Clearly, the disk is not a symmetric space as in the above definition. It has been clearly shown in \[[3](#bib.bib3)\]—if node $`u`$ is on the center of the circle or nearby, the probability that $`u`$ is traversed by a message routed along the shortest path from a random source node $`v_{1}`$ to a random destination $`v_{2}`$ is larger than that of a node away from the center of the network area. Clearly, the square has exactly the same problem. This claim is confirmed by our experiments: $`25\%`$ of the shortest paths traverse a relatively small central disk whose area is $`3\%`$ of the entire square.
To solve these problems, our idea is to map the network nodes onto a symmetric space (the *outer space*) through a mapping that preserves the initial network properties (such as distribution, number of nodes, and, with some limitations, distances between them). The second step is to route messages through the shortest paths as they are defined on the outer space. When the outer space and the corresponding mapping are clear from the context, we will call these paths the *outer space shortest paths*. Since the outer space is symmetric, we can actually prove that every node in the network has the same probability of being traversed by an outer space shortest path. In the following section we will see that, based on this idea, we can design practical routing protocols that do not have highly congested areas, weaker security, and all the problems we have been discussing here. Now, let’s make a step back and proceed formally.
Let $`S`$ be the original space where the network is deployed, and let $`T`$ be the outer space, an abstract space we use to describe routes, both metric spaces with respective distances $`d_{S}`$ and $`d_{T}`$. We are looking for a mapping function $`\phi:{S\mapsto T}`$ with the following properties:
1. 1\.
$`\phi`$ is an injection;
2. 2\.
if $`x`$ is a point taken uniformly at random on $`S`$, then $`\phi\hspace{0pt}{(x)}`$ is also taken uniformly at random on $`T`$;
3. 3\.
for every $`r > 0`$, and every $`{u,v} \in T`$, $`u \neq v`$, if $`{d_{T}\hspace{0pt}{({\phi\hspace{0pt}{(u)}},{\phi\hspace{0pt}{(v)}})}} \leq r`$ then $`{d_{S}\hspace{0pt}{(u,v)}} \leq r`$.
Property [1](#S3.I1.i1 "item 1 ‣ III Routing in Outer Space ‣ Routing in Outer Space: Improved Security and Energy-Efficiency in Multi-Hop Wireless Networks") guarantees that $`\phi^{- 1}`$ is well-defined, Property [2](#S3.I1.i2 "item 2 ‣ III Routing in Outer Space ‣ Routing in Outer Space: Improved Security and Energy-Efficiency in Multi-Hop Wireless Networks") guarantees that a uniform traffic on $`S`$ is still a uniform traffic when mapped onto $`T`$ through $`\phi`$, and Property [3](#S3.I1.i3 "item 3 ‣ III Routing in Outer Space ‣ Routing in Outer Space: Improved Security and Energy-Efficiency in Multi-Hop Wireless Networks") says that paths on $`T`$ are paths on $`S`$, when mapped through $`\phi^{- 1}`$. We’ll see later why these properties are important.
######
Definition 2
A mapping $`\phi:{S\longrightarrow T}`$ is *fair* if it enjoys Properties [1](#S3.I1.i1 "item 1 ‣ III Routing in Outer Space ‣ Routing in Outer Space: Improved Security and Energy-Efficiency in Multi-Hop Wireless Networks"), [2](#S3.I1.i2 "item 2 ‣ III Routing in Outer Space ‣ Routing in Outer Space: Improved Security and Energy-Efficiency in Multi-Hop Wireless Networks"), and [3](#S3.I1.i3 "item 3 ‣ III Routing in Outer Space ‣ Routing in Outer Space: Improved Security and Energy-Efficiency in Multi-Hop Wireless Networks").
Once such a fair mapping has been fixed, any message from node $`u`$ to node $`v`$ can be routed following the shortest path $`{\phi\hspace{0pt}{(u)}},\ldots,{\phi\hspace{0pt}{(v)}}`$ between the images of $`u`$ and $`v`$ and through the images of the nodes of the network under $`\phi`$ on space $`T`$. Being $`\phi`$ a fair mapping, the path $`{\phi^{- 1}\hspace{0pt}{(u)}},\ldots,{\phi^{- 1}\hspace{0pt}{(v)}}`$ is a well defined path on $`S`$. Indeed, $`\phi^{- 1}\hspace{0pt}{(u)}`$ is unique for all $`u`$, since $`\phi`$ is injective; and any two consecutive nodes in the shortest path on $`T`$ are neighbors in $`S`$ as well, thanks to Property [3](#S3.I1.i3 "item 3 ‣ III Routing in Outer Space ‣ Routing in Outer Space: Improved Security and Energy-Efficiency in Multi-Hop Wireless Networks"). If $`T`$ is symmetric as in Definition [1](#Thmdefinition1 "Definition 1 ‣ III Routing in Outer Space ‣ Routing in Outer Space: Improved Security and Energy-Efficiency in Multi-Hop Wireless Networks"), the routing through $`\phi`$ would be well distributed over $`T`$. Hence, this path can be used to route messages on $`S`$, giving as a result a homogeneous distribution of the message flow over all the original network area.
######
Theorem 1
Let $`\phi:{S\longrightarrow T}`$ be a mapping from source metric space $`S`$ to target metric space $`T`$. Assume that $`\phi`$ is fair and $`T`$ is symmetric. Fixed a node $`u \in S`$, deployed the other nodes of the network uniformly at random, and taken a source $`v_{1} \in S`$ and a destination $`v_{2} \in S`$ uniformly at random, the probability that the outer space shortest path $`{\phi^{- 1}\hspace{0pt}{(v_{1})}},\ldots,{\phi^{- 1}\hspace{0pt}{(v_{2})}}`$ traverses $`u`$ is independent of the position of $`u`$ on $`S`$.
The above theorem gives an important hint on how to build a routing protocol on a not symmetric network area, in such a way that the message flow is distributed homogeneously over all the network. What is needed is to determine a symmetric space (the outer space) and a fair mapping for it, and then to “transform” the shortest paths on the original network area into the corresponding outer space shortest paths.
We assume that the original network area is a square of side 1. An excellent candidate as a symmetric outer space is the torus. A *torus* is a 3-dimensional surface that we can model as $`T = {{\lbrack 0,t\rbrack} \times {\lbrack 0,t\rbrack}}`$. Let $`u_{x}`$ and $`u_{y}`$ be the coordinates of the position of node $`u`$ on the torus. We can endow $`T`$ with the following distance $`d_{T}`$:
<table>
<tbody>
<tr>
<td></td>
<td>dT(u,v)=dx2+dy2,subscript𝑑𝑇𝑢𝑣superscriptsubscript𝑑𝑥2superscriptsubscript𝑑𝑦2d_{T}(u,v)=\sqrt{d_{x}^{2}+d_{y}^{2}},</td>
<td></td>
<td>(1)</td>
</tr>
</tbody>
</table>
where
<table>
<tbody>
<tr>
<td></td>
<td>dxsubscript𝑑𝑥\displaystyle d_{x}</td>
<td>=min{|ux−vx|,t−|ux−vx|},andabsentsubscript𝑢𝑥subscript𝑣𝑥𝑡subscript𝑢𝑥subscript𝑣𝑥and\displaystyle=\min\{\left|u_{x}-v_{x}\right|,t-\left|u_{x}-v_{x}\right|\},\;\textrm{and}</td>
<td></td>
<td>(2)</td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td>dysubscript𝑑𝑦\displaystyle d_{y}</td>
<td>=min{|uy,vy|,t−|uy,vy|}.\displaystyle=\min\{\left|u_{y},v_{y}\right|,t-\left|u_{y},v_{y}\right|\}.</td>
<td></td>
<td>(3)</td>
</tr>
</tbody>
</table>
The common way to visualize a torus is to consider a square, and then to fold it in such a way that the left side is glued together with the right side, and that the top side is glued together with the bottom side. In the following, we will picture the torus unfolded, just like a square, as it is commonly done to easily see this 3-dimensional surface as a 2-dimensional one.
######
Fact 1
A torus surface is *symmetric* as in Definition [1](#Thmdefinition1 "Definition 1 ‣ III Routing in Outer Space ‣ Routing in Outer Space: Improved Security and Energy-Efficiency in Multi-Hop Wireless Networks").
Clearly, virtually no wireless network in real life is deployed on a torus. Here, we are using the torus just as an abstract space. We are *not* making any unreasonable assumption on the nodes of the network being phisically placed on a torus like area with continuous boundaries, nor are we assuming that the network area becomes suddenly a torus. Indeed, we assume that the real network is deployed on the square, where the nodes close to one side *cannot* communicate with the nodes close to the opposite side. Crucially, the paths used to deliver the messages are computed as they are defined through a fair mapping onto the torus, the outer space. Coming back to our idea, now that the target symmetric outer space has been chosen, what is left to do is to find a fair mapping $`\phi_{S\hspace{0pt}T}`$ from the square to the torus.
Let $`S = {{\lbrack 0,1\rbrack} \times {\lbrack 0,1\rbrack}}`$ be a square, and let $`T = {{\lbrack 0,2\rbrack} \times {\lbrack 0,2\rbrack}}`$ be a torus. As the mapping $`\phi_{S\hspace{0pt}T}`$ from $`S`$ to $`T`$ we propose the following:
$`{\phi_{S\hspace{0pt}T}\hspace{0pt}{({(x,y)})}} = {(x',y')}`$ where:
<table>
<tbody>
<tr>
<td></td>
<td>x′={xwith probability 1/22−xwith probability 1/2,superscript𝑥′cases𝑥with probability 122𝑥with probability 12x^{\prime}=\begin{cases}x&\textrm{with probability }1/2\\ 2-x&\textrm{with probability }1/2,\end{cases}</td>
<td></td>
</tr>
</tbody>
</table>
and
<table>
<tbody>
<tr>
<td></td>
<td>y′={ywith probability 1/22−ywith probability 1/2.superscript𝑦′cases𝑦with probability 122𝑦with probability 12y^{\prime}=\begin{cases}y&\textrm{with probability }1/2\\ 2-y&\textrm{with probability }1/2.\end{cases}</td>
<td></td>
</tr>
</tbody>
</table>
An example of such a mapping can be seen in Figure [1](#S3.F1 "Figure 1 ‣ III Routing in Outer Space ‣ Routing in Outer Space: Improved Security and Energy-Efficiency in Multi-Hop Wireless Networks"), where a node on the square is mapped to one of the four equally probably images on the torus.
[Refer to caption](/html/0711.0892/assets/x1.png)
Figure 1: Example of transformation of a point from the square to the torus through the mapping $`\phi_{S\hspace{0pt}T}`$. Point $`(x,y)`$ on the square $`S = {{\lbrack 0,1\rbrack} \times {\lbrack 0,1\rbrack}}`$ has four possible and equally probable images on the torus $`T = {{\lbrack 0,2\rbrack} \times {\lbrack 0,2\rbrack}}`$. According on $`\phi_{S\hspace{0pt}T}`$, only one of the images will actually appear on $`T`$.
######
Theorem 2
$`\phi_{S\hspace{0pt}T}`$ is a *fair* mapping with probability one.
###### Proof:
The proof of this claim follows directly from the definition of the mapping $`\phi_{S\hspace{0pt}T}`$. The full proof is technical, without adding much to the understanding of this work. Therefore, for the sake of brevity, we omit the details. ∎
It is interesting to note that it is *not* true that two points that are neighbors on the square are also neighbors on the torus when mapped through $`\phi_{S\hspace{0pt}T}`$. Generally speaking, it is impossible to build a mapping with both this property and Property [3](#S3.I1.i3 "item 3 ‣ III Routing in Outer Space ‣ Routing in Outer Space: Improved Security and Energy-Efficiency in Multi-Hop Wireless Networks"), since the square and the torus are topologically different.
The outer space shortest path between two nodes may be different from the corresponding shortest path. Clearly, it can’t be shorter by definition of shortest path on $`S`$. A natural question to ask is whether we can bound the stretch, that is, how much longer may the outer space shortest path be compared with the corresponding shortest path? Unfortunately, the answer is that the stretch cannot be bounded by a constant. However, quite surprisingly, we can prove a very good constant bound in the case when many messages are sent through the network, that is the common case in practice. Indeed, while in the worst case the stretch can be high, it is not on average if we assume a uniform traffic. This claim is formalized in the following theorem, where we show that, on expectation, the distance of the images under $`\phi_{S\hspace{0pt}T}`$ of two nodes taken uniformly and independently at random is at most the double of the original distance.
######
Theorem 3
If $`u,v`$ are taken uniformly at random on the square $`S = {{\lbrack 0,1\rbrack} \times {\lbrack 0,1\rbrack}}`$, and $`{\phi_{S\hspace{0pt}T}\hspace{0pt}{(u)}},{\phi_{S\hspace{0pt}T}\hspace{0pt}{(v)}}`$ are their respective images under $`\phi_{S\hspace{0pt}T}`$ on the torus $`T = {{\lbrack 0,2\rbrack} \times {\lbrack 0,2\rbrack}}`$, then
<table>
<tbody>
<tr>
<td></td>
<td>E[dT(ϕST(u),ϕST(v)]≤2E[dS(u,v)].E[d_{T}(\phi_{ST}(u),\phi_{ST}(v)]\leq 2E[d_{S}(u,v)].</td>
<td></td>
</tr>
</tbody>
</table>
###### Proof:
Let $`{u,v} \in S`$ be two nodes whose position is taken uniformly at random, and let $`{E\hspace{0pt}{\lbrack{d_{S}\hspace{0pt}{(u,v)}}\rbrack}} = \mu`$ be the expectation of their distance on $`S`$. Since $`\phi_{S\hspace{0pt}T}`$ is fair, also $`\phi_{S\hspace{0pt}T}\hspace{0pt}{(u)}`$ and $`\phi_{S\hspace{0pt}T}\hspace{0pt}{(v)}`$ are taken uniformly at random in the torus. Clearly, the distance between $`\phi_{S\hspace{0pt}T}\hspace{0pt}{(u)}`$ and $`\phi_{S\hspace{0pt}T}\hspace{0pt}{(v)}`$ on the torus cannot be larger of the distance of $`\phi_{S\hspace{0pt}T}\hspace{0pt}{(u)}`$ and $`\phi_{S\hspace{0pt}T}\hspace{0pt}{(v)}`$ on a square $`S' = {{\lbrack 0,2\rbrack} \times {\lbrack 0,2\rbrack}}`$. Indeed, every path on the torus is also a path on the square (the opposite is not true); and the average distance of two random points in a square of edge two is the double of the average distance of two random points in a square of edge one. Therefore,
<table>
<tbody>
<tr>
<td></td>
<td>E[dT(ϕST(u),ϕST(v)]\displaystyle E[d_{T}(\phi_{ST}(u),\phi_{ST}(v)]</td>
<td>≤E[dS′(ϕST(u),ϕST(v)]\displaystyle\leq E[d_{S^{\prime}}(\phi_{ST}(u),\phi_{ST}(v)]</td>
<td></td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td></td>
<td>=2E[dS(u,v)]absent2𝐸delimited-[]subscript𝑑𝑆𝑢𝑣\displaystyle=2E[d_{S}(u,v)]</td>
<td></td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td></td>
<td>=2μ.absent2𝜇\displaystyle=2\mu.</td>
<td></td>
</tr>
</tbody>
</table>
∎
In the following, we will see with experiments that the actual average stretch is even smaller.
Of course, it is always possible to use the outer space shortest path only when the stretch of that particular path is small, and to use the classical shortest path when the stretch is high and the outer space shortest path is going to cost a lot more. However, we do not perform these kind of optimizations—even though they may reduce the global energy required by the network to deliver the messages, they also unbalance the load among the nodes. In the following, we will implement our idea in a practical routing protocol derived from geographical routing, and show its performance by means of experiments.
## IV Routing in Outer Space in Practice
[Refer to caption](/html/0711.0892/assets/x2.png)
\(a\)
Geographic routing between node $`u`$ and $`v`$.
[Refer to caption](/html/0711.0892/assets/x3.png)
\(b\)
Four equally probable outer space geographic routings between node $`u`$ and node $`v`$.
Figure 2: Transformation of a geographic route on the square into four possible outer space geographic routes between nodes $`u`$ and $`v`$. As can be seen, depending on which possible $`\phi_{S\hspace{0pt}T}\hspace{0pt}{(u)}`$ and $`\phi_{S\hspace{0pt}T}\hspace{0pt}{(v)}`$ images are chosen for $`u`$ and $`v`$ on the torus surface, there are four possible outer space geographic routes on the square between $`u`$ and $`v`$. The network is made of $`6,441`$ nodes.
In order to design a practical protocol that makes use of our ideas we start from geographic routing, a simple protocol that, when the network is dense enough, can be shown to approximate shortest path routing quite well \[[4](#bib.bib4)\]. Here, we define *outer space geographic routing*, its outer space counterpart.
In geographic routing, the destination of a message is a geographical position in the network area—in the square in our case. Every relay node performs a very simple protocol: send the message to the node that is closer to destination. If such a node does not exist, then the message is delivered. If the network is dense, every message is delivered to the node closest to destination. It is known that this simple version of geographic routing sometimes is not able to deliver the message to the node closest to destination, and there are plenty of ways to overcome this problem in the literature. However, we do not consider these extensions (outer space geographic routing could as well be based on these more complex and complete versions), since the increased complexity do not add much to this work.
Outer space geographic routing works quite as simply. Every relay node looks at the destination $`x`$ of the message, and forwards it to the node $`u`$ that minimizes $`d_{T}\hspace{0pt}{({\phi_{S\hspace{0pt}T}\hspace{0pt}{(x)}},{\phi_{S\hspace{0pt}T}\hspace{0pt}{(u)}})}`$. Just like geographic routing, implemented on the outer space.
Take, as an example, a message from node $`u`$ destined to a geographic position close to node $`v`$. According to the definition of $`\phi_{S\hspace{0pt}T}`$, each node on the square $`S`$ has four possible and equally probable images on the torus $`T`$. This implies that for each pair $`u`$, $`v`$ of nodes on $`S`$ there are four possible and equally probable pairs of images $`\phi_{S\hspace{0pt}T}\hspace{0pt}{(u)}`$, $`\phi_{S\hspace{0pt}T}\hspace{0pt}{(v)}`$ on $`T`$
<sup>1</sup>
<sup>1</sup>1Actually, there are 16 possible and equiprobable such couples up to isomorphism, which fall into 4 different classes of symmetry.
. This yields four possible and different outer space geographic routes between the images $`u`$ and $`v`$ under $`\phi_{S\hspace{0pt}T}`$. Hence, between any two nodes on the square there is one out of four different and equally probable outer space routes. To see an example of the four routes, see Figure [2](#S4.F2 "Figure 2 ‣ IV Routing in Outer Space in Practice ‣ Routing in Outer Space: Improved Security and Energy-Efficiency in Multi-Hop Wireless Networks").
To implement such a routing, it is enough that the nodes know their position in the square. Then, computing $`\phi_{S\hspace{0pt}T}`$ for itself and the neighbors is trivial and fast. Note that it is not really important that the nodes agree on which of the four possible images is actually chosen for any particular node (except for the destination, but the problem can easily be fixed). However, to get this agreement it is enough that every node uses the same pseudo-random number generator, seeded with the id of the node being mapped.
###
IV-A
Node and Network Properties, Assumptions, and Simulation Environment
We model our network node as a sensor. A typical example can be the Mica2DOT node (outdoor range $`150\hspace{0pt}\text{m}`$, 3V coin cell battery). These nodes have been widely used in sensor network academic research and real testbeds. For our experiments, we have considered networks with up to 10,000 nodes, distributed on a square of side $`1,{500\hspace{0pt}\text{m}}`$. In the following, we will assume for the sake of simplicity that the side of the square is 1, and that the node transmission range is 0.1. The nodes are placed according to a Poisson distribution with density $`\rho`$, chosen in such a way that every node has 30–40 neighbors on average. This is a reasonable and commonly used density for this kind of networks.
We inject a *uniform traffic* in the network—every message has a random source and a random destination uniformly and independently chosen. This type of traffic distribution is highly used in the simulations, for example when the goal is to study network capacity limits, optimal routing, and security properties \[[23](#bib.bib23), [33](#bib.bib33), [34](#bib.bib34)\]. We assume that the nodes know their position on the network area. Therefore, they need to know both their absolute position, and their position within the square. The nodes can get the absolute position either in hardware, by using a GPS (Global Positioning System), or in software. There exist several techniques for location sensing like those based on proximity or triangulation using different types of signals like radio, infrared acoustic, etc. Based on these techniques, several location systems have been proposed in the literature like infrastructure-based localization systems \[[35](#bib.bib35), [36](#bib.bib36)\] and ad-hoc localization systems \[[28](#bib.bib28), [29](#bib.bib29)\]. In \[[18](#bib.bib18)\] you can find a survey on these systems. Once the absolute position is known, we can get the nodes to know their relative position within the square by pre-loading the information on the deployment area, or by using one of the several techniques for boundary detection based on geometry methods, statistical methods, and topological methods (see \[[30](#bib.bib30), [31](#bib.bib31), [32](#bib.bib32)\]).
In the next two sections we present the results of the experiments we have performed, comparing our routing scheme with geographic routing over the same networks and with the same set of messages to route. For the experiments we have used our own event-based simulator. The assumptions and network properties listed above have been exactly reflected in the behavior of the simulator.
###
IV-B
Security-Related Experiments
[Refer to caption](/html/0711.0892/assets/images/securitySquare.jpg)
\(a\)
Geographic routing.
[Refer to caption](/html/0711.0892/assets/images/securityTorus.jpg)
\(b\)
Outer space geographic routing.
Figure 3: The average fraction of the messages whose routing path traverses the selected sub-areas of a network of $`1,336`$ nodes, in the case of geographic routing and in the case of outer space geographic routing.
In these experiments, we measure the number of messages whose routing path traverses five sub-areas of the same size in the network area. Every sub-area is a circle of radius $`0.1`$ (incidentally, the same of the transmission radius of a network node), that corresponds to an area of $`3.14\%`$ of the whole network surface. The sub-areas are centered in some “crucial” points of the network area: The center and the middle-half-diagonals points. The center of the network is known to be the most congested area. We want to test whether the middle-half-diagonal centered areas handle a significantly smaller number of messages. More specifically we consider the sub-areas centered in the points of coordinates $`(0.5,0.5)`$, $`(0.25,0.25)`$, $`(0.25,0.75)`$, $`(0.75,0.25)`$, $`(0.75,0.75)`$, assuming a square of side one. Our experiments are done on networks with different number of nodes (from 1,000 to 10,000). For each network we have launched both geographic routing and outer space geographic routing on message sets of different cardinality (from $`50,000`$ to $`1,000,000`$ of messages, generated as an instance of uniform traffic). In Figure [3](#S4.F3 "Figure 3 ‣ IV-B Security-Related Experiments ‣ IV Routing in Outer Space in Practice ‣ Routing in Outer Space: Improved Security and Energy-Efficiency in Multi-Hop Wireless Networks") we present the average of the results obtained with a network of $`1,336`$ nodes generated by a Poisson process, but we stress out that exactly the same results are obtained for networks with up to $`10,000`$ nodes. As it can be seen, the experiments fully support the findings in \[[3](#bib.bib3)\]. Geographic routing (see Figure [3(a)](#S4.F3.sf1 "In Figure 3 ‣ IV-B Security-Related Experiments ‣ IV Routing in Outer Space in Practice ‣ Routing in Outer Space: Improved Security and Energy-Efficiency in Multi-Hop Wireless Networks")) concentrates a relevant fraction of the messages on a small central area of the network, while the other sub-areas handle on average little more than the half. We have already discussed why this is dangerous, and important to avoid.
Figure [3(b)](#S4.F3.sf2 "In Figure 3 ‣ IV-B Security-Related Experiments ‣ IV Routing in Outer Space in Practice ‣ Routing in Outer Space: Improved Security and Energy-Efficiency in Multi-Hop Wireless Networks") shows the result with the same set of messages and the same network deployment, this time using outer space geographic routing. The message load in the central sub-area is $`30\%`$ lower compared with the load of the same sub-area in the case of the geographic routing. Outer space geographic routing seems to transform the network area in a symmetric surface, making sure that the number of message handled by all the sub-areas remains reasonably low, $`17\%`$, and equally distributed. As a result, the load among network nodes is equally balanced and there are no “over-loaded” areas. This network is intuitively stronger than the same network using geographic routing, there are no areas that are clearly more rewarding as objective of a malicious attack, and no network areas have more “responsabilities” than others.
Furthermore, Figure [3(a)](#S4.F3.sf1 "In Figure 3 ‣ IV-B Security-Related Experiments ‣ IV Routing in Outer Space in Practice ‣ Routing in Outer Space: Improved Security and Energy-Efficiency in Multi-Hop Wireless Networks") clearly shows that, with geographic routing, it is not a good strategy to stay in the center of the network if you want to save your battery. If the nodes are selfish, it is a much better strategy to position in one of the sub-central areas, for example, where the battery is going to last $`66\%`$ longer. Even better if you move towards the side of the square. Conversely, when using outer space geographic routing, there is no advantage in choosing one position or the other, which is exactly our goal to guarantee an even distribution of the nodes, although part of them are selfish.
###
IV-C
How to Live Longer by Consuming More Energy
Our main motivation is related to security. However, it is always important to understand what is the energy overhead of getting rid of congested areas. Indeed, what Theorem [3](#Thmtheorem3 "Theorem 3 ‣ III Routing in Outer Space ‣ Routing in Outer Space: Improved Security and Energy-Efficiency in Multi-Hop Wireless Networks") says in a word is that the paths using outer space geographic routing are on average (at most) twice as long as the paths using geographic routing. This should have an immediate consequence on energy consumption: Messages routed with outer space geographic routing should make network nodes consume more energy, up to twice as much. And it actually is so. What it turns out with our experiments is that the overall energy consumption is about $`1.4`$ times larger with outer space geographic routing, see Figure [4](#S4.F4 "Figure 4 ‣ IV-C How to Live Longer by Consuming More Energy ‣ IV Routing in Outer Space in Practice ‣ Routing in Outer Space: Improved Security and Energy-Efficiency in Multi-Hop Wireless Networks"). Like before, the figure shows the result with a network of $`1,625`$ nodes, but we have done more experiments with different sizes, up to $`10,000`$ nodes, and the result does not change.
Usually, when a wireless network consumes more energy, its life is shorter. However, it is not always the case. Sometimes it is better to consume more energy, if this is done more equally in the network. This is exactly what happens with outer space geographic routing. We consider two measure of network longevity: time to first node death, time to loss of efficiency in delivery messages. These measures are well-known and used in the literature \[[37](#bib.bib37), [38](#bib.bib38), [39](#bib.bib39)\]. We have made two sets of experiments, each using one of the above way to measure the longevity of the network. In each of the experiments we count the number of messages that are successfully delivered before network “death”, where network death is defined according to the above two measures.
[Refer to caption](/html/0711.0892/assets/x4.png)
Figure 4: Global energy consumption of the network nodes after running geographic and outer space geographic routing, respectively, on sets of $`50,000`$ messages each.The network considered is made of $`1,625`$ nodes.
The first set of experiments is done according to the first measure. We have generated the network, the uniform traffic, and injected the traffic into two copies of the same network, one using geographic routing and one using outer space geographic routing. This have been iterated several times with networks of different sizes. The result is shown in Figure [5](#S4.F5 "Figure 5 ‣ IV-C How to Live Longer by Consuming More Energy ‣ IV Routing in Outer Space in Practice ‣ Routing in Outer Space: Improved Security and Energy-Efficiency in Multi-Hop Wireless Networks"), where we show the number of messages delivered on average by a network of $`1,625`$ nodes (the result does not change by considering network of different size), using the two routing protocols under evaluation.
[Refer to caption](/html/0711.0892/assets/x5.png)
Figure 5: Time to first node death. The time is measured as the number of messages delivered to destination before the death of the first node. The network consists of $`1,625`$ nodes.
As you can see, the network lifetime of outer space geographic routing is $`22.57\%`$ longer, on average, than simple geographic routing. As a matter of fact, the number of messages successfully delivered by the network until the very first node death is much greater with outer space routing.
[Refer to caption](/html/0711.0892/assets/x6.png)
Figure 6: Time to delivered percentage. The time is measured as the number of messages delivered to destination before the delivery success ratio falls under $`95\%`$. The network considered is made of $`1,625`$ nodes.
Since security usually comes at a price, this is somewhat surprising. Routing in outer space seems to deliver a more secure routing and a more energy-efficient network, simultaneously.
Figure [6](#S4.F6 "Figure 6 ‣ IV-C How to Live Longer by Consuming More Energy ‣ IV Routing in Outer Space in Practice ‣ Routing in Outer Space: Improved Security and Energy-Efficiency in Multi-Hop Wireless Networks") shows the result we get when considering the second definition of network lifetime. In this case, we consider the network dead when it is not efficient any more in delivering messages. Note that geographic routing (and similarly its outer space version) has the problem of “dead ends”, places where the message cannot proceed because there is no node closer to destination, while the destination is still far. There are a number of solutions to this problem, and there do exists more sophisticated versions of geographic routing that know how to detect this situation and deliver the message whenever there is a path between source and destination. However, this mechanisms are usually costly. When the network is not able any longer to deliver messages with simple geographic routing, that means that there has been enough deterioration to create many dead ends in the network itself. We use this as a measure of the quality of its structure. In this set of experiments we count the number of messages that reach destination until the percentage of delivery falls under some threshold (in our case $`95\%`$). As can be seen in the figure, even in this case outer space geographic routing prolongs the life of the network to a value that is on average $`11.14\%`$ larger than the one achieved with geographic routing.
## V Conclusions
Uniform traffic injected into multi-hop wireless networks generates congested areas. These areas carry a number of non-trivial issues about security, energy-efficiency, and tolerance to (a particular case of) selfish behavior. In this paper we describe routing in outer space, a mechanism to transform shortest path routing protocols into new protocols that do not have the above mentioned problems.
Routing in outer space guarantees that every node of the network is responsible for relaying the same number of messages, on expectation. We can show that a network that uses routing in outer space does not have congested areas, does not have the associated security-related issues, does not encourage selfish positioning, and, in spite of using more energy globally, lives longer of the same network using the original routing protocol.
## References
- \[1\]
I. Akyildiz, W. Su, Y. Sankarasubramaniam, and E. Cayirci, “A survey on sensor networks,” IEEE Communcations Magazine, vol. 40, pp. 102–114, August 2002.
- \[2\]
A. Mainwaring, D. Culler, J. Polastre, R. Szewczyk, and J. Anderson, “Wireless sensor networks for habitat monitoring,” in WSNA ’02: Proceedings of the 1st ACM international workshop on Wireless sensor networks and applications, (New York, NY, USA), pp. 88–97, ACM Press, 2002.
- \[3\]
S. Kwon and N. B. Shroff, “Paradox of shortest path routing for large multi-hop wireless networks,” in IEEE INFOCOM’07 Anchorage, May 2007.
- \[4\]
B. Karp and H. T. Kung, “Gpsr: Greedy perimeter stateless routing for wireless networks,” in Proceedings of the 6th Annual ACM/IEEE International Conference on Mobile Computing and Networking (MobiCom ’00), (Boston, MA, USA), August 2000.
- \[5\] Y. Yu, R. Govindan, and D. Estrin, “Geographical and energy aware routing: A recursive data dissemination protocol for wireless sensor networks,” Tech. Rep. UCLA/CSD-TR-01-0023, UCLA Computer Science Department, May 2001.
- \[6\]
V. Srinivasan, P. Neggehalli, C. F. Chiasserini, and R. R. Rao, “Cooperation in wireless ad hoc wireless networks,” in Proceedings of the Twenty-Second Annual Joint Conference of the IEEE Computer and Communications Societies (INFOCOM 2003), 2003.
- \[7\]
W. Wang, X. Li, and Y. Wang, “Truthful multicast routing in selfish wireless networks,” in Proceedings of the 10th Annual ACM/IEEE International Conference on Mobile Computing and Networking (MobiCom ’04), (Philadelphia, PA, USA), September 2004.
- \[8\]
C. E. Perkins and P. Bhagwat, “Highly dynamic destination-sequenced distance-vector routing (dsdv) for mobile computers,” in SIGCOMM ’94: Proceedings of the conference on Communications architectures, protocols and applications, (New York, NY, USA), pp. 234–244, ACM Press, 1994.
- \[9\]
M. G. Zapata and N. Asokan, “Securing ad hoc routing protocols,” in Proceedings of the ACM Workshop on Wireles Security (WiSe), 2002.
- \[10\]
K. Sanzgiri, B. Dahill, B. Levine, C. Shields, and E. Belding-Royer, “A secure routing protocol for ad hoc networks,” in Proceedings of the International Conference on Network Protocols (ICNP), 2002.
- \[11\]
D. B. Johnson and D. A. Maltz, “Dynamic source routing in ad hoc wireless networks,” in Mobile Computing (Imielinski and Korth, eds.), vol. 353, Kluwer Academic Publishers, 1996.
- \[12\]
V. D. Park and M. S. Corson, “A highly adaptive distributed routing algorithm for mobile wireless networks,” in INFOCOM ’97: Proceedings of the INFOCOM ’97. Sixteenth Annual Joint Conference of the IEEE Computer and Communications Societies. Driving the Information Revolution, (Washington, DC, USA), p. 1405, IEEE Computer Society, 1997.
- \[13\]
C. Perkins and E. Royer, “Ad hoc on-demand distance vector routing,” in Proceedings of the IEEE Workshop on Mobile Computing Systems and Applications, pp. 90–100, February 1999.
- \[14\]
Z. Haas, “A new routing protocol for the reconfigurable wireless networks,” in Proc. of the IEEE Int. Conf. on Universal Personal Communications, October 1997.
- \[15\]
J. Broch, D. A. Maltz, D. B. Johnson, Y.-C. Hu, and J. Jetcheva, “A performance comparison of multi-hop wireless ad hoc network routing protocols,” in MobiCom ’98: Proceedings of the 4th annual ACM/IEEE international conference on Mobile computing and networking, (New York, NY, USA), pp. 85–97, ACM Press, 1998.
- \[16\] E. Royer and C. Toh, “A review of current routing protocols for ad-hoc mobile wireless networks,” April 1999.
- \[17\]
P. Jacquet, P. Mühlethaler, T. Clausen, A. Laouiti, A. Qayyum, and L. Viennot, “Optimized link state routing protocol for ad hoc networks,” in Proceedings of the 5th IEEE Multi Topic Conference (INMIC 2001), 2001.
- \[18\]
J. Hightower and G. Borriello, “Location systems for ubiquitous computing,” IEEE Computer, vol. 34, pp. 57–66, August 2001.
- \[19\]
Y. Xu, J. Heidemann, and D. Estrin, “Geography-informed energy conservation for ad hoc routing,” in Proceedings of the 7th Annual ACM/IEEE International Conference on Mobile Computing and Networking (MobiCom ’01), (Rome, Italy), July 2001.
- \[20\] K. Seada and A. Helmy, “Geographic protocols in sensor networks,” tech. rep., USC, July 2004.
- \[21\]
P. P. Pham and S. Perreau, “Increasing the network performance using multi-path routing mechanism with load balance,” Ad Hoc Networks, vol. 2, pp. 433–459, October 2004.
- \[22\]
Y. Ganjali and A. Keshavarzian, “Load balancing in ad hoc networks: single-path routing vs. multi-path routing,” in Proceedings of the Twenty-third Annual Joint Conference of the IEEE Computer and Communications Societies (INFOCOM 2004), vol. 2, pp. 1120–1125 vol.2, 2004.
- \[23\]
P. Gupta and P. R. Kumar, “The capacity of wireless networks,” IEEE Transactions On Information Theory, vol. 46, March 2000.
- \[24\]
A. Rajeswaran and R. Negi, “Capacity of power constrained ad-hoc networks,” in Proceedings INFOCOM, 2004., vol. 1, pp. 7–11, March 2004.
- \[25\]
M. Franceschetti, O. Dousse, D. Tse, and P. Thiran, “On the throughput capacity of random wireless networks,” IEEE Transactions on Information Theory, vol. 52, june 2006.
- \[26\]
M. Zorzi and R. R. Rao, “Geographic random forwarding (geraf) for ad hoc and sensor networks: Energy and latency performance,” IEEE Transactions on Mobile Computing, vol. 2, no. 4, pp. 349–365, 2003.
- \[27\]
P. Casari, A. Marcucci, M. Nati, C. Petrioli, and M. Zorzi, “A detailed simulation study of geographic random forwarding (geraf) in wireless sensor networks,” in Military Communications Conference, 2005. MILCOM 2005.IEEE, vol. 1, pp. 59–68, October 2005.
- \[28\]
N. Bulusu, J. Heidemann, D. Estrin, and T. Tran, “Self–configuring localization systems: Design and experimental evaluation,” Trans. on Embedded Computing Sys., vol. 3, no. 1, pp. 24–60, 2004.
- \[29\]
A. Savvides, C.-C. Han, and M. B. Srivastava, “Dynamic fine-grain localization in ad-hoc networks of sensors,” in Proceedings of the Seventh Annual ACM/IEEE International Conference on Mobile Computing and Networking (Mobicom), 2001.
- \[30\]
Q. Fang, J. Gao, and L. J. Guibas, “Locating and bypassing holes in sensor networks,” Mob. Netw. Appl., vol. 11, no. 2, pp. 187–200, 2006.
- \[31\]
A. Kröller, S. P. Fekete, D. Pfisterer, and S. Fischer, “Deterministic boundary recognition and topology extraction for large sensor networks,” in SODA ’06: Proceedings of the seventeenth annual ACM-SIAM symposium on Discrete algorithm, (New York, NY, USA), pp. 1000–1009, ACM Press, 2006.
- \[32\]
Y. Wang, J. Gao, and J. S. Mitchell, “Boundary recognition in sensor networks by topological methods,” in MobiCom ’06: Proceedings of the 12th annual international conference on Mobile computing and networking, (New York, NY, USA), pp. 122–133, ACM Press, 2006.
- \[33\]
A. Zemlianov and G. de Veciana, “Capacity of ad hoc wireless networks with infrastructure support,” IEEE Journal on selected areas in Communications, vol. 23, March 2005.
- \[34\]
X. Hong, P. Wang, J. Kong, Q. Zheng, and J. Liu, “Effective probabilistic approach protecting sensor traffic,” Military Communications Conference, 2005. MILCOM 2005. IEEE, vol. 1, pp. 169–175, October 2005.
- \[35\] A. Ward, A. Jones, and A. Hopper, “A new location technique for the active office,” 1997.
- \[36\]
N. Priyantha, A. Chakraborty, and H. Balakrishnan, “The cricket location-support system,” in Proceedings of the 6th Annual ACM International Conference on Mobile Computing and Networking (MobiCom ’00), August 2000.
- \[37\]
M. Bhardwaj and A. Chandrakasan, “Bounding the lifetime of sensor networks via optimal role assignments,” in Proceedings of the Twenty-First Annual Joint Conference of the IEEE Computer and Communications Societies (INFOCOM 2002), vol. 3, pp. 1587–1596, 2002.
- \[38\]
D. Blough and P. Santi, “Investigating upper bounds on network lifetime extension for cell-based energy conservation techniques in adhoc networks,” in Proceedings of the Eighth Annual International Conference on Mobile Computing and Networking (ACM MobiCom 2002), 2002.
- \[39\]
H. Zhang and J. Hou, “On deriving the upper bound of $`\alpha`$—lifetime for large sensor networks,” in MobiHoc ’04: Proceedings of the 5th ACM international symposium on Mobile ad hoc networking and computing, (New York, NY, USA), pp. 121–132, ACM Press, 2004. |
0711.1273 | cs.NI | cyber | # Practical Resource Allocation Algorithms for QoS in OFDMA-based Wireless Systems
Tolga Girici2, Chenxi Zhu1,Jonathan R. Agre1, Anthony Ephremides2
2 Institute for Systems Research
A.V.William Building
University of Maryland
Email:{tgirici,etony}@eng.umd.edu 1 Fujitsu Labs of America
8400 Baltimore Ave., Suite 302
College Park, Maryland 20740
Email:{czhu,jagre}@fla.fujitsu.com
###### Abstract
In this work we propose an efficient resource allocation algorithm for OFDMA based wireless systems supporting heterogeneous traffic. The proposed algorithm provides proportionally fairness to data users and short term rate guarantees to real-time users. Based on the QoS requirements, buffer occupancy and channel conditions, we propose a scheme for rate requirement determination for delay constrained sessions. Then we formulate and solve the proportional fair rate allocation problem subject to those rate requirements and power/bandwidth constraints. Simulations results show that the proposed algorithm provides significant improvement with respect to the benchmark algorithm.
## I Introduction
Broadband wireless networks are designed to be able to provide high rate and heterogenous services to mobile users that have various quality of service (QoS) requirements. Two notable examples of broadband wireless technologies are 3GPP and Mobile WiMax(802.16e). Transmissions in Long Term Evolution (3GPP) and 802.16-based wireless technologies are based on OFDM, where several modulation, coding and power allocation schemes are allowed to give more degrees of freedom to resource allocation \[[1](#bib.bib1)\]. Fully taking advantage of this degree of freedom is an important problem and has been studied previously in \[[2](#bib.bib2), [3](#bib.bib3), [4](#bib.bib4), [5](#bib.bib5), [6](#bib.bib6)\]. Papers \[[2](#bib.bib2)\] and \[[4](#bib.bib4)\] address maximizing total throughput subject to power and subcarrier constraints. Above works consider maximizing total capacity for data traffic but do not address fairness for data traffic or QoS for real time traffic. The authors in \[[3](#bib.bib3)\], \[[5](#bib.bib5)\], \[[6](#bib.bib6)\] studied proportional fair scheduling. However these schemes also do not guarantee any short or long term transmission rates. The scheduling rules do not apply sufficiently to different QoS requirements and heterogeneous traffic.
In OFDMA, a wideband channel is divided into a number of narrow-band carriers and these carriers are allocated to users. Typically the carriers that are close in the frequency spectrum have correlated channel conditions. In order to make the allocation easier carriers are grouped into subchannels. There are various ways of subchannelization, e.g. contiguous grouping (i.e. Band AMC), where adjacent carriers are grouped into a single subchannel. By this method it is safe to assume that each subchannel is subject to independent and identically distributed fading. This method fully takes the advantage of OFDMA by frequency selectivity. Another method is the distributed grouping (i.e. PUSC/FUSC) where a subchannel is formed by sampling carriers across the whole range of subcarriers according to a permutation, or randomly, so that each subchannel has the same average fading with respect to a user. Most of the previous works has considered the first method in their models, however it has two main disadvantages for mobile networks. First, the proposed algorithms become too complex when each subchannel has different fading. We choose permutational method for subchannelization. Therefore our question becomes how many subchannels to allocate instead of which subchannels , which makes our resource allocation algorithms more practical. Second, for a mobile channel with fast fading, channel estimation and feedback becomes more practical using distributed grouping.
Motivated by the above issues we propose a resource allocation algorithm, that satisfies delay requirements for real time traffic, while providing proportional fair rate allocation for elastic traffic. Our algorithm is based on user selection and rate requirement determination for voice users and solution of a proportional fair rate allocation problem subject to those rate requirements and power/bandwidth constraints.
## II System Model
We consider a cellular system consisting of a single base station transmitting to $`N`$ mobile users. Time is divided into frames of length $`T_{f}`$ and at each time frame base station allocates the total bandwidth $`W`$ and total power $`P`$ among the users. In the simulations we keep the users fixed, however we simulate mobility by fast and slow fading. Fast fading is Rayleigh distributed and slow fading is log-normal distributed. Total channel gain is the product of distance attenuation, fast and slow fading. Let $`h_{i}\hspace{0pt}{(t)}`$ be the channel gain of user i at time t. For an AWGN channel with noise p.s.d. $`N_{0}`$, signal to interference plus noise ratio (SINR) is,
<table>
<tbody>
<tr>
<td></td>
<td>SINRi=pi(t)hi(t)N0wi(t),𝑆𝐼𝑁subscript𝑅𝑖subscript𝑝𝑖𝑡subscriptℎ𝑖𝑡subscript𝑁0subscript𝑤𝑖𝑡SINR_{i}=\frac{p_{i}(t)h_{i}(t)}{N_{0}w_{i}(t)},</td>
<td></td>
<td>(1)</td>
</tr>
</tbody>
</table>
where $`p_{i}\hspace{0pt}{(t)}`$ and $`w_{i}\hspace{0pt}{(t)}`$ are the power and bandwidth allocated to user i at time t.
We use the following (modulation,coding,repetition) pairs \[QPSK,1/2,6$`\times`$ - QPSK,1/2,4$`\times`$ - QPSK,1/2,2$`\times`$ - QPSK,1/2,1$`\times`$1 - QPSK,3/4,1$`\times`$ - 16QAM,1/2,1$`\times`$ - 16QAM,3/4,1$`\times`$ - 64QAM,2/3, 1$`\times`$ - 64QAM,3/4,1$`\times`$\] corresponding to the following SINR levels: \[-2.78, -1.0, 2.0, 5, 6, 10.5, 14, 18, 20\] dB \[[7](#bib.bib7)\]. For instance QPSK,1/2,6$`\times`$ corresponds to a bandwidth efficiency of 1/6 bps/Hz. In the problem formulation, we will use the following rate function.
<table>
<tbody>
<tr>
<td></td>
<td>ri(pi(t),wi(t))=wi(t)log(1+βpi(t)hi(t)N0wi(t)),subscript𝑟𝑖subscript𝑝𝑖𝑡subscript𝑤𝑖𝑡subscript𝑤𝑖𝑡1𝛽subscript𝑝𝑖𝑡subscriptℎ𝑖𝑡subscript𝑁0subscript𝑤𝑖𝑡r_{i}(p_{i}(t),w_{i}(t))=w_{i}(t)\log\left(1+\beta\frac{p_{i}(t)h_{i}(t)}{N_{0}w_{i}(t)}\right),</td>
<td></td>
<td>(2)</td>
</tr>
</tbody>
</table>
which is the Shannon capacity expression with an SINR factor $`\beta < 1`$. If we choose $`\beta = 0.25`$, this rate function approximates the above values quite well. After allocating the power and bandwidth we quantize the SINR to the values above. Bandwidth also is quantized to multiples of subchannel bandwidth, $`W_{s\hspace{0pt}u\hspace{0pt}b}`$.
The network can support different traffic types such as real time (VoIP), video streaming, data applications with some rate requirements (FTP) and best effort traffic. We assume that each user demands a single type of traffic. We will consider the following traffic types: 1) Best Effort (BE): Non real time traffic with no minimum rate requirements. 2) Video Streaming: Bursty real time traffic with delay constraint. 3) VoIP: Constant bit rate (CBR) traffic with delay constraint.
We classify the traffic into two groups as elastic and non-elastic traffic. BE traffic is elastic, that is, a BE user can use any available traffic. Fairness and throughput are the performance objectives for BE traffic. Proportional fairness provides a good balances between the two. Voice traffic is non-elastic; it is a CBR traffic with strict delay requirements. If a voice user can receive its short term required rate level, it doesn’t need excessive resources. On the other hand Video streaming traffic is in between the two types. It has a basic rate requirement with certain delay constraints, however it is possible to achieve higher quality video transmission if the user experiences good channel conditions. In this work we aim to satisfy the basic rate requirement for voice and video users, while treating excessive rate allocation for video users similarly as BE users. Typical rates for these traffic types are listed in Table [II](#S5.T2 "TABLE II ‣ V Numerical Evaluation ‣ Practical Resource Allocation Algorithms for QoS in OFDMA-based Wireless Systems").
## III User Selection
Our proposed scheduling algorithm consists of user selection and rate allocation. After selecting the users, the subchannels and power is allocated.
###
III-A
Modified Largest Delay First - Proportional Fairness
In single channel systems Largest Weighted Delay First (LWDF) is shown to be throughput optimal \[[8](#bib.bib8)\]. In this scheme at each frame the user maximizing the following quantity transmits
<table>
<tbody>
<tr>
<td></td>
<td>aiDiHOL(t)ri(P,W),subscript𝑎𝑖superscriptsubscript𝐷𝑖𝐻𝑂𝐿𝑡subscript𝑟𝑖𝑃𝑊a_{i}D_{i}^{HOL}(t)r_{i}(P,W),</td>
<td></td>
<td>(3)</td>
</tr>
</tbody>
</table>
where $`D_{i}^{H\hspace{0pt}O\hspace{0pt}L}\hspace{0pt}{(t)}`$ is the head of line packet delay and $`r_{i}\hspace{0pt}{(P,W)}`$ is the channel capacity of user i at frame t (calculated from ([2](#S2.E2 "In II System Model ‣ Practical Resource Allocation Algorithms for QoS in OFDMA-based Wireless Systems")), where P and W is the fixed transmission power and channel bandwidth). The parameter $`a_{i}`$ is a positive constant. If QoS is defined as
<table>
<tbody>
<tr>
<td></td>
<td>P(Di>Dimax)<δi,𝑃subscript𝐷𝑖superscriptsubscript𝐷𝑖𝑚𝑎𝑥subscript𝛿𝑖P(D_{i}>D_{i}^{max})<\delta_{i},</td>
<td></td>
<td>(4)</td>
</tr>
</tbody>
</table>
where $`D_{i}^{m\hspace{0pt}a\hspace{0pt}x}`$ is the delay constraint and $`\delta_{i}`$ is the probability of exceeding this constraint (typically 0.05), then the constant $`a_{i}`$ can be defined as $`a_{i} = {- \frac{\log{(\delta_{i})}}{D_{i}^{m\hspace{0pt}a\hspace{0pt}x}\hspace{0pt}R_{i}\hspace{0pt}{(t)}}}`$ , which is referred to as M-LWDF-PF \[[8](#bib.bib8)\]. Here, $`R_{i}\hspace{0pt}{(t)}`$ is the average received rate. Averaged (filtered) values of long term received rates of users, which is computed as follows:
<table>
<tbody>
<tr>
<td></td>
<td>Ri(t+1)=αiRi(t)+(1−αi)ri(pi(t),wi(t))subscript𝑅𝑖𝑡1subscript𝛼𝑖subscript𝑅𝑖𝑡1subscript𝛼𝑖subscript𝑟𝑖subscript𝑝𝑖𝑡subscript𝑤𝑖𝑡R_{i}(t+1)=\alpha_{i}R_{i}(t)+(1-\alpha_{i})r_{i}(p_{i}(t),w_{i}(t))</td>
<td></td>
<td>(5)</td>
</tr>
</tbody>
</table>
The equation above can be considered as a filter with time constant $`1/{({1 - \alpha_{i}})}`$ for user $`i`$. The constant $`\alpha_{i}`$ should be chosen such that the average received rate is detected earlier than the delay constraint in terms of frame durations. We choose 100msec, 400 msec and 1000 msec as the delay constraints of voice, streaming and BE users. Converting these values into number of frames of 1msec we get the $`\alpha`$ values in Table [II](#S5.T2 "TABLE II ‣ V Numerical Evaluation ‣ Practical Resource Allocation Algorithms for QoS in OFDMA-based Wireless Systems"). M-LWDF-PF can be adapted to OFDMA systems as follows. Power is distributed equally to all subchannels. Starting from the first subchannel , the subchannel is allocated to the user maximizing ([3](#S3.E3 "In III-A Modified Largest Delay First - Proportional Fairness ‣ III User Selection ‣ Practical Resource Allocation Algorithms for QoS in OFDMA-based Wireless Systems")). Then the received rate $`R\hspace{0pt}{(t)}`$ is updated according to ([5](#S3.E5 "In III-A Modified Largest Delay First - Proportional Fairness ‣ III User Selection ‣ Practical Resource Allocation Algorithms for QoS in OFDMA-based Wireless Systems")). All the subchannels are allocated one-by-one according to this rule. We will use this algorithm as benchmark in our simulations.
###
III-B
Proposed Algorithm - Delay and Rate Based Resource Allocation
There are two main disadvantages of M-LWDF-PF algorithm. First, the power is divided equally to over subcarriers. Performance can be increased by power control. Secondly, data users are much different than video and voice in terms of QoS requirements. Therefore it is hard to use the same metric for data and real time users. We propose a Delay and Rate based Resource Allocation algorithm (DRA). We first choose the users to be served in the current frame according to the following user satisfaction value.
<table>
<tbody>
<tr>
<td></td>
<td>USVi(t)=LiDiHOLlog(1+βpi(t)hi(t)N0wi(t))ri0Ri(t)𝑈𝑆subscript𝑉𝑖𝑡subscript𝐿𝑖superscriptsubscript𝐷𝑖𝐻𝑂𝐿1𝛽subscript𝑝𝑖𝑡subscriptℎ𝑖𝑡subscript𝑁0subscript𝑤𝑖𝑡superscriptsubscript𝑟𝑖0subscript𝑅𝑖𝑡USV_{i}(t)=L_{i}D_{i}^{HOL}\log\left(1+\frac{\beta p_{i}(t)h_{i}(t)}{N_{0}w_{i}(t)}\right)\frac{r_{i}^{0}}{R_{i}(t)}</td>
<td></td>
<td>(6)</td>
</tr>
</tbody>
</table>
Here $`L_{i} = {- \frac{\log{(\delta_{i})}}{D_{i}^{m\hspace{0pt}a\hspace{0pt}x}}}`$ and $`r_{i}^{0}`$ is the basic rate requirement for user i. Let $`U_{D}`$, $`U_{S}`$ and $`U_{V}`$ be the BE, Video and Voice users. Let $`U_{R} = {U_{S} \cup U_{V}}`$ be the set of real time users. Let $`U_{E}`$ and $`\overline{U_{E}}`$ be the set of users demanding elastic traffic and the rest, respectively.
We use a simple formula to determine the fraction $`F_{R}\hspace{0pt}{(t)}`$ of real time users scheduled in each time slot,
<table>
<tbody>
<tr>
<td></td>
<td>FR(t)=1|UR|∑i∈URI(qi(t)>0.5Dimaxri0)subscript𝐹𝑅𝑡1subscript𝑈𝑅subscript𝑖subscript𝑈𝑅𝐼subscript𝑞𝑖𝑡0.5superscriptsubscript𝐷𝑖𝑚𝑎𝑥superscriptsubscript𝑟𝑖0F_{R}(t)=\frac{1}{|U_{R}|}\sum_{i\in U_{R}}I(q_{i}(t)>0.5D_{i}^{max}r_{i}^{0})</td>
<td></td>
<td>(7)</td>
</tr>
</tbody>
</table>
Here $`q_{i}\hspace{0pt}{(t)}`$ is the queue size in bits and $`0.5\hspace{0pt}D_{i}^{m\hspace{0pt}a\hspace{0pt}x}\hspace{0pt}r_{i}^{0}`$ denotes a queue size threshold in bits and $`I{(.)}`$ is the indicator function taking value one if the argument inside is true. As more users exceed this threshold, more fraction of real time users are scheduled. For data users, the BS simply chooses a fraction of 0.2 of users.
## IV Joint Power and Bandwidth Allocation
After the users are chosen, joint power and bandwidth allocation is performed. Let $`U_{D}'`$, $`U_{S}'`$ and $`U_{V}'`$ be the chosen users that belong to all three traffic classes. The algorithm is as follows:
###
IV-A
Basic Rate Allocation for Real Time Users
For the selected real time users ($`i \in U_{R}'`$) the rate requirements are determined first. Rate requirement for real time user $`i`$ is,
<table>
<tbody>
<tr>
<td></td>
<td>ric(qi(t),ωi(t))=(qi(t)Ts,ri0ωi(t),),i∈UR′r_{i}^{c}(q_{i}(t),\omega_{i}(t))=\left(\frac{q_{i}(t)}{T_{s}},\frac{r_{i}^{0}}{\omega_{i}(t)},\right),~{}i\in U_{R}^{\prime}</td>
<td></td>
<td>(8)</td>
</tr>
</tbody>
</table>
Here $`q_{i}\hspace{0pt}{(t)}`$ is the queue size and $`\omega_{i}\hspace{0pt}{(t)}`$ is the transmission frequency of user i, which is updated as follows:
<table>
<tbody>
<tr>
<td></td>
<td>ωi(t)=αiωi(t−1)+(1−αi)I(ri(t)>0),subscript𝜔𝑖𝑡subscript𝛼𝑖subscript𝜔𝑖𝑡11subscript𝛼𝑖𝐼subscript𝑟𝑖𝑡0\omega_{i}(t)=\alpha_{i}\omega_{i}(t-1)+(1-\alpha_{i})I(r_{i}(t)>0),</td>
<td></td>
<td>(9)</td>
</tr>
</tbody>
</table>
where $`I\hspace{0pt}{({{r_{i}\hspace{0pt}{(t)}} > 0})}`$ is the function that takes value one if the node receives packets in time slot $`t`$, zero otherwise. Therefore this frequency decreases if the node transmits less and less frequently. Using this frequency expression in the basic rate function, we compensate for the lack of transmission in the previous time slots possibly due to bad channel conditions.
For the chosen real time users with non-elastic traffic ($`i \in {\overline{U_{E}} \cap U_{R}'}`$) basic resource allocation is enough to support the session. For these users we allocate the basic resource as follows, and don’t include them in the rate allocation which will be defined later. First, the nominal SNR $`\gamma_{i}^{0}`$ is determined according to the uniform power per bandwidth allocation as $`\gamma_{i}^{0} = \frac{P\hspace{0pt}h_{i}\hspace{0pt}{(t)}}{N_{0}\hspace{0pt}W}`$. Then $`\gamma_{i}^{0}`$ is quantized by decreasing $`\frac{P\hspace{0pt}h_{i}\hspace{0pt}{(t)}}{N_{0}\hspace{0pt}W}`$ to the closest SNR level in Section [II](#S2 "II System Model ‣ Practical Resource Allocation Algorithms for QoS in OFDMA-based Wireless Systems"). If $`\gamma_{i}^{0}`$ is smaller than the smallest SNR level, then the ceiling is taken. Based on this nominal SINR, nominal bandwidth efficiency $`S_{i}^{0}\hspace{0pt}{(t)}`$ (in bps/Hz) is determined again using the values above. Using this basic rate and the nominal bandwidth efficiency, basic bandwidth for non-elastic traffic is determined as $`{w_{i}^{m\hspace{0pt}i\hspace{0pt}n} = \frac{r_{i}^{m\hspace{0pt}i\hspace{0pt}n}\hspace{0pt}{(t)}}{S_{i}^{0}\hspace{0pt}{(t)}}},{i \in {\overline{U_{E}} \cap U_{R}'}}`$. Then this bandwidth is quantized to a multiple of subchannel bandwidth by $`w_{i}^{m\hspace{0pt}i\hspace{0pt}n} = {{\max{(1,{\lfloor w_{i}^{m\hspace{0pt}i\hspace{0pt}n}\rfloor})}}\hspace{0pt}W_{s\hspace{0pt}u\hspace{0pt}b}}`$. Minimal power for this user is then $`{p_{i}^{m\hspace{0pt}i\hspace{0pt}n} = {{{\gamma_{i}^{0}\hspace{0pt}w_{i}^{m\hspace{0pt}i\hspace{0pt}n}\hspace{0pt}N_{0}}/h_{i}}\hspace{0pt}{(t)}}},{{\forall i} \in {\overline{U_{E}} \cap U_{R}'}}`$. Hence $`p_{i} = p_{i}^{m\hspace{0pt}i\hspace{0pt}n}`$ and $`w_{i} = w_{i}^{m\hspace{0pt}i\hspace{0pt}n}`$ for these users.
<sup>1</sup>
<sup>1</sup>1After the basic allocation, if the total bandwidth or power is greater then the available resource, the user with the largest power is chosen, bandwidth is decreased by one subchannel and the power is also decreased in order to keep the SINR fixed. This process is continued until the total bandwidth and power for voice and video users becomes smaller than the available resources.
Let the residual power and bandwidth after non-elastic real time traffic allocations be $`P' = {\sum_{i \in {\overline{U_{E}} \cap U_{R}'}}p_{i}^{m\hspace{0pt}i\hspace{0pt}n}}`$ and $`W' = {\sum_{i \in {\overline{U_{E}} \cap U_{R}'}}w_{i}^{m\hspace{0pt}i\hspace{0pt}n}}`$. For real time users with elastic traffic ($`i \in {U_{R}' \cap U_{E}}`$) we include the basic rate as a constraint in joint residual bandwidth-power allocation, which will be explained next.
###
IV-B
Proportional Fair Resource Allocation for Data and Video Streaming
At this stage the residual power ($`P'`$) and bandwidth ($`W'`$) is allocated among the chosen users demanding elastic traffic in a proportional fair manner. The PF resource allocation problem in ([10](#S4.E10 "In IV-B Proportional Fair Resource Allocation for Data and Video Streaming ‣ IV Joint Power and Bandwidth Allocation ‣ Practical Resource Allocation Algorithms for QoS in OFDMA-based Wireless Systems")) is solved among the chosen streaming and data users.
Find ($`\mathbf{p}^{\ast},\mathbf{w}^{\ast}`$) such that:
<table>
<tbody>
<tr>
<td></td>
<td>max𝐩,𝐰∏i∈UE∩(UR′∪UD′)(wilog(1+piniwi))ϕisubscript𝐩𝐰subscriptproduct𝑖subscript𝑈𝐸superscriptsubscript𝑈𝑅′superscriptsubscript𝑈𝐷′superscriptsubscript𝑤𝑖1subscript𝑝𝑖subscript𝑛𝑖subscript𝑤𝑖subscriptitalic-ϕ𝑖\max_{\mathbf{p},\mathbf{w}}\prod_{i\in U_{E}\cap(U_{R}^{\prime}\cup U_{D}^{\prime})}{\left(w_{i}\log\left({1+\frac{p_{i}}{n_{i}w_{i}}}\right)\right)^{\phi_{i}}}</td>
<td></td>
<td>(10)</td>
</tr>
</tbody>
</table>
subject to,
<table>
<tbody>
<tr>
<td></td>
<td>wilog(1+piniwi)subscript𝑤𝑖1subscript𝑝𝑖subscript𝑛𝑖subscript𝑤𝑖\displaystyle w_{i}\log\left({1+\frac{p_{i}}{n_{i}w_{i}}}\right)</td>
<td>≥\displaystyle\geq</td>
<td>rimin,∀i∈UE∩UR′superscriptsubscript𝑟𝑖𝑚𝑖𝑛for-all𝑖subscript𝑈𝐸superscriptsubscript𝑈𝑅′\displaystyle r_{i}^{min},~{}\forall i\in U_{E}\cap U_{R}^{\prime}</td>
<td></td>
<td>(11)</td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td>∑i∈UE∩(UR′∪UD′)pisubscript𝑖subscript𝑈𝐸superscriptsubscript𝑈𝑅′superscriptsubscript𝑈𝐷′subscript𝑝𝑖\displaystyle\sum_{i\in U_{E}\cap(U_{R}^{\prime}\cup U_{D}^{\prime})}{p_{i}}</td>
<td>≤\displaystyle\leq</td>
<td>P′superscript𝑃′\displaystyle P^{\prime}</td>
<td></td>
<td>(12)</td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td>∑i∈UE∩(UR′∪UD′)wisubscript𝑖subscript𝑈𝐸superscriptsubscript𝑈𝑅′superscriptsubscript𝑈𝐷′subscript𝑤𝑖\displaystyle\sum_{i\in U_{E}\cap(U_{R}^{\prime}\cup U_{D}^{\prime})}{w_{i}}</td>
<td>≤\displaystyle\leq</td>
<td>W′superscript𝑊′\displaystyle W^{\prime}</td>
<td></td>
<td>(13)</td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td>pi,wisubscript𝑝𝑖subscript𝑤𝑖\displaystyle p_{i},w_{i}</td>
<td>≥\displaystyle\geq</td>
<td>0,∀i∈UE∩(UR′∪UD′)0for-all𝑖subscript𝑈𝐸superscriptsubscript𝑈𝑅′superscriptsubscript𝑈𝐷′\displaystyle 0,\forall i\in U_{E}\cap(U_{R}^{\prime}\cup U_{D}^{\prime})</td>
<td></td>
<td>(14)</td>
</tr>
</tbody>
</table>
Here log-sum is written as a product. The above problem is a convex optimization problem with a concave objective function and convex set \[[9](#bib.bib9)\]. In this optimization we also included the parameter $`\phi_{i}`$, which depends on the traffic type. Since data users typically can tolerate more rate and video users are already allocated basic bandwidth, we can give higher $`\phi_{i}`$ for data users. We can solve this problem using the Lagrange multipliers.
###
IV-C
Bandwidth and SINR quantization and Reshuffling
After the resources are allocated, first the bandwidth for data and video streaming users is quantized as $`w_{i} = {{\max{(1,{\lfloor w_{i}\rfloor})}}\hspace{0pt}W_{s\hspace{0pt}u\hspace{0pt}b}}`$. Then the SINR is quantized and transmit power is determined. Unlike best effort transmission, queue size plays an important role in real time transmissions. As a result of the above optimization some streaming time users may get more rates than that is enough to transmit all bits in the queue. Some of the bandwidth is taken from video users in order to obey this queue constraint. After these modifications, if the total bandwidth is greater than the available, then the user with the highest power is found and its bandwidth decreased. Power is recalculated in order to keep the SINR fixed. This process is continued until bandwidth constraint is satisfied. If total power is still greater than the available then again choosing the user with highest power and decreasing bandwidth, power constraint is satisfied. If after these processes there is a leftover bandwidth, then choosing the user that has the highest channel a subchannel is added and power is increased accordingly (if there is enough power to do so). If there is some leftover power, then starting from the user with lower channel gains, SINR is boosted to the next power level (if there is enough power to do so). For the real time users we don’t increase bandwidth or power if there isn’t enough buffer content.
## V Numerical Evaluation
For the numerical evaluations we divide the users to 5 classes according to the distances, 0.3,0.6,0.9,1.2,1.5 km. There are equal number of users at each class. We use the parameters in Table [I](#S5.T1 "TABLE I ‣ V Numerical Evaluation ‣ Practical Resource Allocation Algorithms for QoS in OFDMA-based Wireless Systems").
<table>
<thead>
<tr>
<th>Parameter</th>
<th>Value</th>
</tr>
</thead>
<tbody>
<tr>
<th>Cell radius</th>
<td>1.5km</td>
</tr>
<tr>
<th>User Distances</th>
<td>0.3,0.6,0.9,1.2,1.5 km</td>
</tr>
<tr>
<th>Total power (P)</th>
<td>20 W</td>
</tr>
<tr>
<th>Total bandwidth (W)</th>
<td>10 MHz</td>
</tr>
<tr>
<th>Frame Length</th>
<td>1 msec</td>
</tr>
<tr>
<th>Voice Traffic</th>
<td>CBR 32kbps</td>
</tr>
<tr>
<th>Video Traffic</th>
<td>802.16 - 128kbps</td>
</tr>
<tr>
<th>Best effort File</th>
<td>5 MB</td>
</tr>
<tr>
<th>AWGN p.s.d.(N0subscript𝑁0N_{0})</th>
<td>-169dBm/Hz</td>
</tr>
<tr>
<th>Pathloss exponent (γ𝛾\gamma)</th>
<td>3.5</td>
</tr>
<tr>
<th>ψDB∼N(μψdB,σψdB)similar-tosubscript𝜓𝐷𝐵𝑁subscript𝜇subscript𝜓𝑑𝐵subscript𝜎subscript𝜓𝑑𝐵\psi_{DB}\sim N(\mu_{\psi_{dB}},\sigma_{\psi_{dB}})</th>
<td>N(0dB,8dB)</td>
</tr>
<tr>
<th>Coherent Time (Fast/Slow)</th>
<td>(5msec/300msec.)</td>
</tr>
<tr>
<th>Pathloss(dB, d in meters)</th>
<td>−31.5−35log10d+ψdB31.535subscript10𝑑subscript𝜓𝑑𝐵-31.5-35\log_{10}d+\psi_{dB}</td>
</tr>
</tbody>
</table>
TABLE I: Simulation Parameters
We performed the simulations using MATLAB. We compared our algorithm with the benchmark M-LWDF algorithm with proportional fairness. Delay exceeding probability is taken as $`\delta_{i} = 0.05`$ for all users. The traffic and resource allocation parameters are listed in Table [II](#S5.T2 "TABLE II ‣ V Numerical Evaluation ‣ Practical Resource Allocation Algorithms for QoS in OFDMA-based Wireless Systems"). Since we choose data users separately from others, the parameters $`L_{i}`$ and head of line delay $`D_{i}^{H\hspace{0pt}O\hspace{0pt}L}`$ are not used for data users.
<table>
<thead>
<tr>
<th>Traffic</th>
<th>r0(kbps)superscript𝑟0𝑘𝑏𝑝𝑠r^{0}(kbps)</th>
<th>rmax(kbps)superscript𝑟𝑚𝑎𝑥𝑘𝑏𝑝𝑠r^{max}(kbps)</th>
<th>Dmax(s)superscript𝐷𝑚𝑎𝑥𝑠D^{max}(s)</th>
<th>Lisubscript𝐿𝑖L_{i}</th>
<th>ϕisubscriptitalic-ϕ𝑖\phi_{i}</th>
<th>αisubscript𝛼𝑖\alpha_{i}</th>
</tr>
</thead>
<tbody>
<tr>
<td>VoIP</td>
<td>32</td>
<td>32</td>
<td>0.1</td>
<td>13</td>
<td>-</td>
<td>0.98</td>
</tr>
<tr>
<td>Streaming</td>
<td>128</td>
<td>1024</td>
<td>0.4</td>
<td>3.25</td>
<td>1</td>
<td>0.995</td>
</tr>
<tr>
<td>BE</td>
<td>0</td>
<td>∞\infty</td>
<td>2</td>
<td>0.65</td>
<td>-</td>
<td>0.998</td>
</tr>
</tbody>
</table>
TABLE II: Minimum required and maximum sustained rates for different types of traffic.
The measured performance metrics are $`95^{t\hspace{0pt}h}`$ percentile delay for real time users and total throughput for data users. We will observe these parameters with respect to number of video users. For the delay, we observe the users in the range 0.3-1.2 separately as good users and the ones at 1.5km as bad users.
###
V-A
Fixed Rate Video Traffic
In the first part of the simulations we considered the video traffic rate fixed at 128kbps and treated it as non-elastic. We consider CBR voice traffic, where a fixed length packet arrives periodically. For the Video traffic we used the model in IEEE 802.16e system evaluation methodology. Packet lengths, and interarrival times truncated Pareto distributed such that average rate is 128kbps. For the BE traffic we assume that there are unlimited number of packets in the queue.
In Figure [1](#S5.F1 "Figure 1 ‣ V-A Fixed Rate Video Traffic ‣ V Numerical Evaluation ‣ Practical Resource Allocation Algorithms for QoS in OFDMA-based Wireless Systems"), we plotted the 95 percentile delays of real time users vs increasing number of video users. For this simulation we kept the number of data and Voice users fixed at 20. Again we observe that $`95^{t\hspace{0pt}h}`$ percentile delay for video users increases exponentially with number video users, while delays for the users at the edge is within the acceptable range for DRA unlike M-LWDF.
[Refer to caption](/html/0711.1273/assets/x1.png)
Figure 1: 95 percentile delay(msec) vs. number of video users
In figure [2](#S5.F2 "Figure 2 ‣ V-A Fixed Rate Video Traffic ‣ V Numerical Evaluation ‣ Practical Resource Allocation Algorithms for QoS in OFDMA-based Wireless Systems") we see that total data rate decreases linearly with increasing video users. Data performance of DRA is again better than M-LWDF.
[Refer to caption](/html/0711.1273/assets/x2.png)
Figure 2: Total throughput(Mbos) vs. number of video users
In Figure [3](#S5.F3 "Figure 3 ‣ V-A Fixed Rate Video Traffic ‣ V Numerical Evaluation ‣ Practical Resource Allocation Algorithms for QoS in OFDMA-based Wireless Systems"), $`95^{t\hspace{0pt}h}`$ percentile delay for video and voice users are plotted for increasing number of data users. The number of Streaming and Voice users are kept fixed at 20. We observe a linear increase in the delay w.r.t. number of data users with M-LWDF. The delay increase is negligible for DRA.
[Refer to caption](/html/0711.1273/assets/x3.png)
Figure 3: 95 percentile delay(msec) vs. number of data users
###
V-B
Elastic Video Traffic
In the second part of the simulations we considered video traffic rate that varies with packet delays. We implemented a simple rate control scheme that looks at the average head of line packet delay and increases or decreases input rate according to a threshold policy. We defined rate levels $`r_{i}^{0}\hspace{0pt}\lambda_{i}`$, ($`\lambda_{i} \in {\{ 1,2,\ldots,8\}}`$) that are integer multiples of 128kbps. Interarrival times are the same for level $`1`$ and $`k`$, however for level $`k`$ packet size is $`k`$ times larger for each packet. For each user $`i \in {U_{E} \cap U_{R}}`$ and at each update instant.
- •
if $`{\overline{D_{i}^{H\hspace{0pt}O\hspace{0pt}L}}\hspace{0pt}{(t)}} < {0.125\hspace{0pt}D_{i}^{m\hspace{0pt}a\hspace{0pt}x}}`$ then $`\lambda_{i} = {\min{\{{\lambda_{i} + 1},\lambda^{m\hspace{0pt}a\hspace{0pt}x}\}}}`$
- •
if $`{\overline{D_{i}^{H\hspace{0pt}O\hspace{0pt}L}}\hspace{0pt}{(t)}} > {0.25\hspace{0pt}D_{i}^{m\hspace{0pt}a\hspace{0pt}x}}`$ then $`\lambda_{i} = {\max{\{{\lambda_{i} - 1},1\}}}`$
- •
else, $`\lambda_{i} = \lambda_{i}`$
Here $`\overline{D_{i}^{H\hspace{0pt}O\hspace{0pt}L}}\hspace{0pt}{(t)}`$ denotes mean HOL packet delay in the last $`400`$ frames. The updates are made at each $`200`$ frames.
Figure [4](#S5.F4 "Figure 4 ‣ V-B Elastic Video Traffic ‣ V Numerical Evaluation ‣ Practical Resource Allocation Algorithms for QoS in OFDMA-based Wireless Systems") shows the evolution of rate levels along with queue sizes for video users at distances 300, 900 and 1500 meters. We observe that users closer to the BS can achieve higher rates.
[Refer to caption](/html/0711.1273/assets/x4.png)
Figure 4: Evolution of Video rate along with queue sizes for users at 300, 600 and 900meters
In Figure [5](#S5.F5 "Figure 5 ‣ V-B Elastic Video Traffic ‣ V Numerical Evaluation ‣ Practical Resource Allocation Algorithms for QoS in OFDMA-based Wireless Systems") we observe the comparison of delay and throughput for the DRA and LWDF schemes.We see that DRA system satisfies delay constraints for voice users unlike LWDF. As for throughput, we see that DRA can provide significantly better throughput for video users at all distances. Total data/video throughput and log-sum throughput (proportional fairness) is also better for DRA scheme.
[Refer to caption](/html/0711.1273/assets/x5.png)
Figure 5: $`95^{t\hspace{0pt}h}`$ percentile delay and average throughput for users at different distances.
## References
- \[1\] A. Ghosh, D. Wolter, J. G. Andres, and R Chen. Broadband Wireless Access with WiMax/802.16: Current Performance Benchmarks and Future Potential.
IEEE Communications Magazine, Feb. 2005.
- \[2\] W. Rhee and J. M. Cioffi. Increase in capacity of multiuser OFDM system using dynamic subchannel allocation.
Vehicular Technology Conference Proceedings, 2000. VTC 2000-Spring Tokyo. 2000 IEEE 51st, pages 1085–1089, 15-18 May 2000.
- \[3\] Z. Shen, J. G. Andrews, and B. L. Evans. Adaptive resource allocation in multiuser OFDM systems with proportional rate constraints.
Wireless Communications, IEEE Transactions on, pages 2726–2737, Nov. 2005.
- \[4\] H. Kim, Y. Han, and S. Kim. Joint subcarrier and power allocation in uplink OFDMA systems.
IEEE Communication Letters, pages 526–528, June 2005.
- \[5\] G. Song. Cross-Layer Resource Allocation and Scheduling in Wireless Multicarrier Networks.
Ph.D. Dissertation, Georgia Institute of Technology, Apr. 2005.
- \[6\] C. Zhu and J. Agre. Proportional-Fair Scheduling Algorithms for OFDMA-based Wireless Systems.
Preprint, Fujitsu Labs, 2006.
- \[7\] IEEE 802.16e. IEEE Standard for Local and Metropolitan Area Networks, Part 16: Air Interface for Fixed and Mobile Broadband Wireless Access Systesm, Amendment 2: Physical and Medium Access Control Layers for Combined Fixed and Mobiel Operation in Licensed Bands and Corrigendum 1.
IEEE, Feb. 2006.
- \[8\] M. Andrews, K. Kumaran, K. Ramanan, A. Stolyar, R. Vijayakumar, and P. Whiting. Providing Quality of Service over a Shared Wireless Link.
IEEE Communcations Magazine, pages 150–154, Feb. 2001.
- \[9\] L. Vanderberghe S. Boyd.
Convex Optimization.
March 8, 2004. |
0711.2780 | cs.NI | cyber | <sup>1</sup>
<sup>1</sup>institutetext: Scuola Superiore di Catania
Via S. Nullo 5/i, 95123, Catania, Italy
<sup>1</sup>
<sup>1</sup>email: sascellato@ssc.unict.it
<sup>2</sup>
<sup>2</sup>institutetext: Department of Computer Science, University College London
Gower Street, London, WC1E 6BT, United Kingdom
<sup>2</sup>
<sup>2</sup>email: c.mascolo@cs.ucl.ac.uk
<sup>3</sup>
<sup>3</sup>institutetext: Department of Computer Science, Dartmouth College
6211 Sudikoff Laboratory, Hanover NH 03755 USA
<sup>3</sup>
<sup>3</sup>email: musolesi@cs.dartmouth.edu
<sup>4</sup>
<sup>4</sup>institutetext: Dipartimento di Fisica e Astronomia, Università di Catania
and INFN Sezione di Catania
Via S. Sofia 64, 95125, Catania, Italy
<sup>4</sup>
<sup>4</sup>email: latora@ct.infn.it
# Epcast: Controlled Dissemination in Human-based Wireless Networks by means of Epidemic Spreading Models
Salvatore Scellato<sup>1</sup>
Cecilia Mascolo<sup>2</sup>
Mirco Musolesi<sup>3</sup>
Vito Latora<sup>4</sup>
###### Abstract
Epidemics-inspired techniques have received huge attention in recent years from the distributed systems and networking communities. These algorithms and protocols rely on probabilistic message replication and redundancy to ensure reliable communication. Moreover, they have been successfully exploited to support group communication in distributed systems, broadcasting, multicasting and information dissemination in fixed and mobile networks. However, in most of the existing work, the probability of infection is determined heuristically, without relying on any analytical model. This often leads to unnecessarily high transmission overheads.
In this paper we show that models of epidemic spreading in complex networks can be applied to the problem of tuning and controlling the dissemination of information in wireless ad hoc networks composed of devices carried by individuals, i.e., human-based networks. The novelty of our idea resides in the evaluation and exploitation of the structure of the underlying human network for the automatic tuning of the dissemination process in order to improve the protocol performance. We evaluate the results using synthetic mobility models and real human contacts traces.
Keywords: Wireless Networks, Mobile Ad Hoc Networks, Epidemic Spreading Models, Complex Networks, Data Dissemination
## 1 Introduction
Mobile human networks (i.e., ad hoc networks composed by devices carried by individuals) can be frequently and temporarily disconnected. Traditional routing protocol, including the basic flooding, fail to offer any sort of reliability when this happens. Epidemic-style protocols instead, being store and forward approaches and inherently delay tolerant \[[11](#bib.bib11)\], allow for communication in dynamic and mobile networks, also in presence of temporary disconnections or network partitions. A desired feature of the protocols is the ability to control the information spreading. For example, in emergency scenarios, when the network infrastructure has failed, it may be sufficient to send the messages only to a percentage of the rescue team members (e.g., 50% of the doctors). In other situations, there might be a need to reach all the deployed emergency personnel with the minimum overhead to avoid to collapse the network. Up to our knowledge, no solutions exploiting the minimal necessary and sufficient number of replicated messages, given the emergent network structure to guarantee a desired level of reliability exist.
The analogy between information dissemination in mobile systems and epidemics transmission in social systems is apparent. Information spreading can be modelled with a simple model for disease spreading, the so-called SIR (Susceptible-Infected-Recovered) model \[[2](#bib.bib2)\]: a host is initially Susceptible to new information, then it becomes Infected when he actually receives it, and finally it can stop the store-and-forward dissemination process becoming Recovered and, therefore, immune to further infections. Epidemics-inspired techniques have received huge attention in recent years from the distributed systems community \[[9](#bib.bib9)\]. These algorithms and protocols rely on probabilistic message replication and redundancy to ensure reliable communication. Epidemic techniques were firstly exploited to guarantee consistency in distributed databases \[[8](#bib.bib8)\]. More recently, these algorithms have been applied to support group communication in distributed systems. In particular, several protocols have been proposed for broadcasting, multicasting and information dissemination in fixed networks \[[10](#bib.bib10)\].
A few attempts have been made to apply epidemic based techniques for information dissemination in mobile ad hoc networks \[[17](#bib.bib17), [7](#bib.bib7), [3](#bib.bib3)\]. However, existing epidemic algorithms do not permit to control the spreading of the information depending on the desired reliability and the network structure. This is partly due to the fact that these approaches are fundamentally based on empirical experiments and not on analytical models: the input parameters that control the dissemination process are selected by using experimental results and are not based on any mathematical model. This implies that the message replication process cannot be tuned with accuracy in a dynamic way: for instance, it is not possible to set the parameters of the dissemination process in order to reach only a certain desired percentage of the hosts in a prefixed amount of time. Moreover, these approaches do not exploit the information on the underlying network topology \[[1](#bib.bib1), [4](#bib.bib4), [5](#bib.bib5)\]. The use of epidemic spreading models based on the structure of the underlying network allows us to devise accurate mechanisms for controlling the message replication process. In other words, the number of the replicas in the network and their persistence can be tuned to achieve a desired delivery ratio.
In \[[15](#bib.bib15)\] we have presented initial results based on the so-called SIS (Susceptible-Infected-Susceptible), a model of disease spreading not considering the recovered state. In this paper, we propose a refined version of the algorithm based on a SIR model. The use of SIR, in coordination with the ability to decide to constrain the epidemics to a percentage of hosts, allows us to lower the message overhead considerably with respect to both our previous work and other approaches, as shown in our results section. We present an extended evaluation based on synthetic models and real traces of connectivity of the Dartmouth College \[[14](#bib.bib14)\] and National University of Singapore \[[16](#bib.bib16)\] campuses.
This paper is structured as follows. In Section [2](#S2 "2 Primitives for Controlled Epidemic Dissemination ‣ Epcast: Controlled Dissemination in Human-based Wireless Networks by means of Epidemic Spreading Models") we describe the implementation of the middleware interface supporting the epidemic dissemination process. Section [3](#S3 "3 Dissemination Techniques based on Epidemic Models ‣ Epcast: Controlled Dissemination in Human-based Wireless Networks by means of Epidemic Spreading Models") presents briefly the models of epidemic spreading in complex networks that are at the basis of our dissemination algorithm. The implementation issues are discussed in Section [4](#S4 "4 Implementation ‣ Epcast: Controlled Dissemination in Human-based Wireless Networks by means of Epidemic Spreading Models"). The proposed dissemination algorithm is evaluated analytically and by means of simulations in Section [5](#S5 "5 Evaluation ‣ Epcast: Controlled Dissemination in Human-based Wireless Networks by means of Epidemic Spreading Models"). Section [6](#S6 "6 Concluding Remarks ‣ Epcast: Controlled Dissemination in Human-based Wireless Networks by means of Epidemic Spreading Models") concludes the paper.
## 2 Primitives for Controlled Epidemic Dissemination
Our goal is to provide a primitive that allows developers to tune information dissemination in human networks according to their specific application requirements. Our aim is to ensure the spreading of information from a source $`A`$ to a certain percentage $`\Psi`$ of the mobile hosts of the system in a given interval time defined by a timeout $`t^{\ast}`$.
We introduce a primitive for probabilistic anycast communication as follows:
```
epcast(message,percentageOfHosts,time)
```
where `message` is the message that has to be sent to a certain percentage of hosts equal to the value defined in `percentageOfHosts` in a bounded time interval equal to `time`.
By using these basic primitives, more complex programming interfaces and communication infrastructures can be designed, such as publish/subscribe systems or service discovery protocols.
The infectivity of the epidemics (i.e., the probability of being infected by a host that is in the same radio range, like in human diseases spreading) can be used to control the anycast probabilistic communication mechanism. Given a percentage of hosts that has to be infected equal to $`\Psi`$, we are able to accurately calculate the value of the infectivity in order to obtain an infection rate equal to a proportion of the total number of the hosts in the network.
As we will discuss in the next section, these primitives rely on a probabilistic algorithm based on the transmission of a minimal, and, at the same time, sufficient, number of messages. Existing epidemic-style protocols usually achieve 100% delivery, but they waste resources by sending a large number of messages on the network, whereas our approach succeeds to send only the amount of messages necessary to inform the desired percentage of hosts in the given time.
## 3 Dissemination Techniques based on Epidemic Models
In this section we introduce the mathematical models at the basis of the design of the communication API presented in Section [2](#S2 "2 Primitives for Controlled Epidemic Dissemination ‣ Epcast: Controlled Dissemination in Human-based Wireless Networks by means of Epidemic Spreading Models"). In order to model the message replication mechanisms, we exploit mathematical models that have been devised to describe the dynamics of infections in human populations \[[2](#bib.bib2)\]. The study of mathematical models of biological phenomena has been pioneered by Kermack and McKendrick in the first half of the last century. In the last years, researchers have focused their interest in the modeling of disease spreading in networks characterised by well-defined structural properties \[[4](#bib.bib4), [5](#bib.bib5)\].
According to the classic Kermack and McKendrick model, an individual can be in three states: infected, (i.e., an individual is infected with the disease) susceptible (i.e., an individual is prone to be infected) and removed (i.e., an individual is immune, as it recovered from the disease). This kind of model is usually referred to as the Susceptible-Infected-Removed (SIR) model \[[2](#bib.bib2)\]. Removing the possibility of permanently recovering from the disease a different version of the model is obtained, according to which individuals can exist in only two possible states, infected and susceptible. In the literature, this model is usually referred to as Susceptible-Infected-Susceptible (SIS) model \[[2](#bib.bib2)\]. The SIR model can guarantee the same delivery of the SIS model with a substantially lower number of messages as shown by the generic epidemic process depicted in Figures [2](#S3.F2 "Figure 2 ‣ 3 Dissemination Techniques based on Epidemic Models ‣ Epcast: Controlled Dissemination in Human-based Wireless Networks by means of Epidemic Spreading Models") and [2](#S3.F2 "Figure 2 ‣ 3 Dissemination Techniques based on Epidemic Models ‣ Epcast: Controlled Dissemination in Human-based Wireless Networks by means of Epidemic Spreading Models"). This is due to the fact that the model introduces the possibility of having hosts that are recovered, i.e., hosts that will not participate in spreading the infection after having receiving a message $`M`$ and deleted it from the buffer. In other words, in the SIR model the number of broadcasting nodes decreases after a given peak of infected nodes; instead in the SIS model, the number of broadcasting nodes at the end of the infection is (approximately) equal to the number of nodes to be infected (i.e., desired percentage of nodes in the `epcast` primitive).
[Refer to caption](/html/0711.2780/assets/x1.png)
Figure 1: Infection spreading in the SIS (top) and SIR (bottom) model, under equivalent conditions. The number of infected (infected, removed and infected + removed) individuals is reported as a function of time for the SIS (SIR) model (with $`\gamma = 0.05`$ and tuning the value of $`\beta`$ in order to have an infection of 100%).
[Refer to caption](/html/0711.2780/assets/x2.png)
Figure 2: Number of replicas per host per message for the SIS and the SIR model. Same conditions as in Fig.1
In the remainder of this paper we will substitute the term individual, used by epidemiologists, with the term host. A host is considered infected if it holds the message and susceptible if it does not. If the message is deleted from buffer of the host, the host becomes recovered and cannot be infected by the same message anymore. The information is spreaded among all infectives and recovered, while susceptibles are still unaware of that: it is now clear that the dissemination results depend on both infectives and recovered hosts, since these are the actual recipients of the messages that have been sent. It is useful to define a host as reached if it is either an infective or a recovered, since in both cases it has already received the message. Moreover, it is worth noting that only infectives contribute to message replication and spreading, while recovered hosts do not.
The main assumptions of our model are the following:
- •
all susceptibles in the population are equally at risk of infection from any infected host (this hypothesis is usually defined by epidemiologists as homogeneous mixing);
- •
all infectives in the population have equal chances to recover;
- •
the infectivity of a single host, per message, is constant
<sup>1</sup>
<sup>1</sup>1Note that the infectivity per single message (i.e., a disease) is constant, but not per single host. In other words, a host usually stores messages characterised by different infectivities in its buffer.
;
- •
the initial number of the nodes in the network is known a priori by each host
<sup>2</sup>
<sup>2</sup>2The initial number of hosts can be usually estimated in occasion of sport events, rallies, etc. for example by evaluating the seating capacity of the venues or the size of the area when the event takes place. Statistical data are also usually available for many application scenarios, such as number of passengers that uses a station or an airport in a certain time of the day, etc. Alternatively, this number can be estimated using distributed algorithms for the calculation of the approximated network size such as \[[13](#bib.bib13)\].
;
- •
every host collaborates to the delivery process and no malicious nodes are present;
- •
each node has a buffer of the same size;
- •
the number of hosts is considered constant during the spreading of the infection
<sup>3</sup>
<sup>3</sup>3This is a realistic assumption, since users usually require that the information will be disseminated in a limited time.
;
Under the assumptions above, the system dynamics, in the case of a scenario composed of $`N`$ active hosts, can be approximately
<sup>4</sup>
<sup>4</sup>4This is rigorously justifiable in a network only for complete graphs in large population limit. However, the model provides a good approximation also in scenarios composed of a limited number of hosts.
described by the following system of non-linear differential equations \[[2](#bib.bib2)\]:
<table>
<tbody>
<tr>
<td>(1)</td>
<td></td>
<td>{dS(t)dt=−βS(t)I(t)dI(t)dt=βS(t)I(t)−γI(t)dR(t)dt=γI(t)S(t)+I(t)+R(t)=Ncases𝑑𝑆𝑡𝑑𝑡𝛽𝑆𝑡𝐼𝑡𝑑𝐼𝑡𝑑𝑡𝛽𝑆𝑡𝐼𝑡𝛾𝐼𝑡𝑑𝑅𝑡𝑑𝑡𝛾𝐼𝑡𝑆𝑡𝐼𝑡𝑅𝑡𝑁\left\{\begin{array}[]{l}\frac{dS(t)}{dt}=-\beta S(t)I(t)\\ \frac{dI(t)}{dt}=\beta S(t)I(t)-\gamma I(t)\\ \frac{dR(t)}{dt}=\gamma I(t)\\ S(t)+I(t)+R(t)=N\\ \end{array}\right.</td>
<td></td>
</tr>
</tbody>
</table>
where $`{S\hspace{0pt}{(t)}},{I\hspace{0pt}{(t)}},{R\hspace{0pt}{(t)}}`$ are respectively the number of susceptible, infectives and removed hosts at time $`t`$, $`\beta`$ is the average number of contacts with susceptible hosts that leads to a new infected host per unit of time per infective, and $`\gamma`$ is the average rate of removal of infectives per unit of time per infectives in the population. The equations of the system state that the decaying rate of susceptibles and the growth rate of infectives are affected only by the infectivity $`\beta`$, the number of susceptibles $`S\hspace{0pt}{(t)}`$ and the number of infectives $`I\hspace{0pt}{(t)}`$; the decaying rate of infectives and the relative growth of recovered is proportional to the removal rate $`\gamma`$ and the number of infectives $`I\hspace{0pt}{(t)}`$. The last equation states that actually only two equations are needed to completely define the problem, since the sum of the number of hosts in the three classes is constant. We furthermore set the initial conditions: $`{S\hspace{0pt}{(0)}} = {N - 1}`$, $`{I\hspace{0pt}{(0)}} = 1`$, and $`{R\hspace{0pt}{(0)}} = 0`$, with the condition $`{I\hspace{0pt}{(0)}} = 1`$ representing the first copy of the message that is inserted in its buffer by the sender.
The numerical solution of system ([1](#S3.E1 "In 3 Dissemination Techniques based on Epidemic Models ‣ Epcast: Controlled Dissemination in Human-based Wireless Networks by means of Epidemic Spreading Models")) can be easily obtained by standard ODE solver routines. This allows to compute the number of infectives and recovered at instant $`t`$ as a function of the infectivity $`\beta`$ and of the removal rate $`\gamma`$. The value of $`\gamma`$ is usually fixed by the local properties of the hosts
<sup>5</sup>
<sup>5</sup>5If overflow phenomena do not occur (i.e., in the case of sufficiently large buffers), the model can be simplified with $`\gamma = 0`$ and, therefore, no host will never become recovered.
. Instead, the value of $`\beta`$, that is the fundamental parameter of the message replication algorithm, can be tuned in order to have, after a specific length of time $`t^{\ast}`$, a number of reached hosts (i.e., hosts that have received the message) equal to $`{I\hspace{0pt}{(t^{\ast})}} + {R\hspace{0pt}{(t^{\ast})}}`$ or, in other words, a fraction of reached hosts equal to $`{({{I\hspace{0pt}{(t^{\ast})}} + {R\hspace{0pt}{(t^{\ast})}}})}/N`$.
In order to effectively exploit the model just described, the actual connectivity of each host has to be taken into account. We will assume a mobile system with a homogeneous network structure, described by a connectivity distribution $`P\hspace{0pt}{(k)}`$, strongly peaked at an average value $`\langle k\rangle`$. This is a realistic assumption in cases characterized by a high density of hosts, and where the movement is well described as an uncorrelated random process, such as in large outdoor spaces (i.e., squares, stations, airports or around sport venues) \[[12](#bib.bib12), [15](#bib.bib15)\]. In this case, the degree $`k`$ of each node can be approximated quite precisely with the average degree $`\langle k\rangle`$. In order to include the effect of the connectivity on the spreading, the system ([1](#S3.E1 "In 3 Dissemination Techniques based on Epidemic Models ‣ Epcast: Controlled Dissemination in Human-based Wireless Networks by means of Epidemic Spreading Models")) can be rewritten by substituting $`\beta`$ with $`\lambda\hspace{0pt}\frac{\langle k\rangle}{N}`$ \[[4](#bib.bib4)\]:
<table>
<tbody>
<tr>
<td>(2)</td>
<td></td>
<td>{dS(t)dt=−λ⟨k⟩NS(t)I(t)dI(t)dt=λ⟨k⟩NS(t)I(t)−γI(t)dR(t)dt=γI(t)S(t)+I(t)+R(t)=Ncases𝑑𝑆𝑡𝑑𝑡𝜆delimited-⟨⟩𝑘𝑁𝑆𝑡𝐼𝑡𝑑𝐼𝑡𝑑𝑡𝜆delimited-⟨⟩𝑘𝑁𝑆𝑡𝐼𝑡𝛾𝐼𝑡𝑑𝑅𝑡𝑑𝑡𝛾𝐼𝑡𝑆𝑡𝐼𝑡𝑅𝑡𝑁\left\{\begin{array}[]{l}\frac{dS(t)}{dt}=-\lambda\dfrac{\langle k\rangle}{N}S(t)I(t)\\ \frac{dI(t)}{dt}=\lambda\dfrac{\langle k\rangle}{N}S(t)I(t)-\gamma I(t)\\ \frac{dR(t)}{dt}=\gamma I(t)\\ S(t)+I(t)+R(t)=N\\ \end{array}\right.</td>
<td></td>
</tr>
</tbody>
</table>
where $`\lambda`$ represents the probability of infecting a neighbouring host during a unit of time, and $`\frac{\langle k\rangle}{N}`$ gives the probability of being in contact with a certain host. In other words, in this model, by substituting $`\beta`$ with $`\lambda\hspace{0pt}\frac{\langle k\rangle}{N}`$, we have separated, in a sense, the event of being connected to a certain host and the infective process \[[4](#bib.bib4)\].
In conclusion, the main idea is to calculate the value of $`\lambda`$ as a function of $`{I\hspace{0pt}{(t^{\ast})}} + {R\hspace{0pt}{(t^{\ast})}}`$ and $`\langle k\rangle`$. It is also interesting to note that in homogeneous networks, every host knows its value of $`k`$ and, consequently, it has a good estimate of $`\langle k\rangle`$. We will exploit this property to tune the spreading of message replicas in the system.
## 4 Implementation
Every time the middleware primitive defined in Section [2](#S2 "2 Primitives for Controlled Epidemic Dissemination ‣ Epcast: Controlled Dissemination in Human-based Wireless Networks by means of Epidemic Spreading Models") is invoked, the middleware calculates the value of the infectivity $`\lambda`$ that is necessary and sufficient to spread the information to the desired fraction of hosts in the specified time interval (specified in the field `percentageOfHosts` of the `epcast` primitive), by evaluating the current average degree of connectivity and the current removal rate of messages from the buffer. The message identifiers, the value of the calculated infectivity, the timestamp containing the value specified in `time` expressing its temporal validity are inserted in the corresponding headers of the message in the infectivity field. Then, the message is inserted in the local buffer.
The epidemic spreading protocol is executed periodically with a period equal to $`\tau`$. With respect to the calculation of the message infectivity, we assume $`\tau`$ as time unit in the formulae presented in Section [3](#S3 "3 Dissemination Techniques based on Epidemic Models ‣ Epcast: Controlled Dissemination in Human-based Wireless Networks by means of Epidemic Spreading Models"). In other words, assuming, for example, $`\tau = 10`$, a timestamp equal to one minute corresponds to six time units. The value of $`\tau`$ can be set by the application developer during the deployment of the platform. Clearly, the choice of the values of $`\tau`$ influences the accuracy of the model, since it relies on a probabilistic process. For this reason, given a minimum value of timestamp equal to $`t_{M\hspace{0pt}I\hspace{0pt}N}`$, developers should ensure $`\tau\operatorname{<<}t_{M\hspace{0pt}I\hspace{0pt}N}`$. The number of rounds will be equal to $`t^{\ast}/\tau`$. For the Law of the Large Numbers, we obtain a better accuracy of the estimation of the evolution of the epidemics as the number of rounds (i.e., from a probabilistic point of view, the number of trials) increases.
Every $`\tau`$ seconds each infected host broadcasts the message and its neighbours receive the message. If the message is not already present in their buffer, they store it with a probability $`\lambda`$: moreover, they will not store it if the message has been already present in buffer in the past, although it is not present at current time. This behaviour maps quite well the SIR epidemic spreading model, since a node receives a new message, actively spreads it for some time and then it deletes the message from the buffer (i.e. to make room for new messages), never accepting it again. Therefore, a node has to store the identifiers of all messages received in a defined time window, which is a reasonable given the limited occupation of the vector of the message identifiers.
## 5 Evaluation
### 5.1 Analytical Evaluation
An interesting quantitative parameter is the total number of messages needed to disseminate messages to a certain percentage of hosts. A message is broadcasted by an infective host in every round: as soon as the host deletes the message it does not accept the same message again.
Considering an infection process repeated for a number of times equal to $`r`$ number of rounds, indicating with $`t_{r}`$ the time length of the $`r^{t\hspace{0pt}h}`$ round, the total number of replicas per single type of message can be estimated as follows:
<table>
<tbody>
<tr>
<td>(3)</td>
<td></td>
<td> Number Of Replicas=∫t=0t=trI(t)𝑑t Number Of Replicassuperscriptsubscript𝑡0𝑡subscript𝑡𝑟𝐼𝑡differential-d𝑡\text{ Number Of Replicas}=\int_{t=0}^{t=t_{r}}I(t)dt</td>
<td></td>
</tr>
</tbody>
</table>
From a graphical point of view, the number of copies is equal to the area under the curves in Figure [2](#S3.F2 "Figure 2 ‣ 3 Dissemination Techniques based on Epidemic Models ‣ Epcast: Controlled Dissemination in Human-based Wireless Networks by means of Epidemic Spreading Models") and [2](#S3.F2 "Figure 2 ‣ 3 Dissemination Techniques based on Epidemic Models ‣ Epcast: Controlled Dissemination in Human-based Wireless Networks by means of Epidemic Spreading Models"). A comparison between SIR- and SIS-based protocols shows that while for both cases the formula [3](#S5.E3 "In 5.1 Analytical Evaluation ‣ 5 Evaluation ‣ Epcast: Controlled Dissemination in Human-based Wireless Networks by means of Epidemic Spreading Models") helds, in the former case the total number of replicas sent is much lower. This is the result of the recovering process, which enables hosts to stop message spreading when the epidemics is already growing but, at the same time, still guarantees that the final result will be guaranteed.
### 5.2 Experimental Evaluation
#### 5.2.1 Description of the Simulation
In order to test the performance of these techniques, we defined a square simulation area with a side of 1 km and a transmission range equal to 200 m. The simulation was set to run several runs for each mobile scenario in order to obtain a statistically meaningful set of results (with a maximum 5% error). All simulations are written in Python using NetworkX
<sup>6</sup>
<sup>6</sup>6 http://networkx.lanl.gov
, a package for the creation, manipulation, and study of the structure, dynamics, and functions of complex networks. We analysed scenarios characterised by different number of hosts (more precisely 64, 128, 256, 512). These input parameters model typical deployment settings of mobile ad hoc networked systems. We do not model explicitly the failures in the system, since we assume that during the infection process, the number of hosts remains constant.
The movements of the hosts are generated using a Random Way-Point mobility model \[[6](#bib.bib6)\]; every host moves at a speed that is randomly generated by using a uniform distribution. The range of the possible speeds is $`{{\lbrack 1,6\rbrack}\hspace{0pt}m}/s`$. We selected this mobility model, since as discussed in \[[12](#bib.bib12)\], its emergent topology has a Poisson degree distribution. Therefore, in this scenario, the properties of the network can be studied with a good approximation by assuming a homogeneous network model. The accuracy of the approximation increases as the density of population increases, since, considering the finite and limited simulated time, we obtain a scenario characterised by a time series of degree of connectivity values with lower variance. Moreover, the so-called border effects, due to the host that moves at the boundaries of the simulated scenarios, have less influence as the density of population increases.
Each node uses a buffer of 5 messages, managed as a FIFO (first in first out) queue, and 20 different messages are sent in the initial round by random chosen nodes.
[Refer to caption](/html/0711.2780/assets/x3.png)
Figure 3: Delivery ratio vs population density with desired reliability equal to 100.
[Refer to caption](/html/0711.2780/assets/x4.png)
Figure 4: Delivery ratio vs population density with desired reliability equal to 50.
[Refer to caption](/html/0711.2780/assets/x5.png)
Figure 5: Number of replicas per host per message vs population density with desired reliability equal to 100.
[Refer to caption](/html/0711.2780/assets/x6.png)
Figure 6: Number of replicas per host per message vs population density with desired reliability equal to 50.
#### 5.2.2 Analysis of Simulation Results
In this subsection we will analyse the results of our simulations, discussing the performance of the proposed techniques. We will study the variations of some performance indicators, such as the delivery ratio and the number of messages sent as functions of the density of hosts (i.e., the number of the hosts in the simulation area).
Figure [4](#S5.F4 "Figure 4 ‣ 5.2.1 Description of the Simulation ‣ 5.2 Experimental Evaluation ‣ 5 Evaluation ‣ Epcast: Controlled Dissemination in Human-based Wireless Networks by means of Epidemic Spreading Models") and [4](#S5.F4 "Figure 4 ‣ 5.2.1 Description of the Simulation ‣ 5.2 Experimental Evaluation ‣ 5 Evaluation ‣ Epcast: Controlled Dissemination in Human-based Wireless Networks by means of Epidemic Spreading Models") show the delivery ratio (i.e., the desired percentage of hosts in the `epcast` primitive) in terms of population density, for the case of a desired percentage of hosts equal to 100 and 50, respectively, with $`t^{\ast} = {10\hspace{0pt}m\hspace{0pt}i\hspace{0pt}n}`$. The performance in terms of delivery ratio are close to the desired ones. Also in this case, the better approximation of the assumption of homogeneous network, obtained when the density of population increases, leads to better results (i.e., a more accurate estimation) for the case of 512 nodes.
The number of replicas per host per message are plotted in Figure [6](#S5.F6 "Figure 6 ‣ 5.2.1 Description of the Simulation ‣ 5.2 Experimental Evaluation ‣ 5 Evaluation ‣ Epcast: Controlled Dissemination in Human-based Wireless Networks by means of Epidemic Spreading Models") and [6](#S5.F6 "Figure 6 ‣ 5.2.1 Description of the Simulation ‣ 5.2 Experimental Evaluation ‣ 5 Evaluation ‣ Epcast: Controlled Dissemination in Human-based Wireless Networks by means of Epidemic Spreading Models"). These diagrams illustrate the scalability of our approach, since the number of replicas is slightly decreasing when more nodes are added.
#### 5.2.3 Evaluation with Dartmouth traces
In order to evalute our approach on real data we run simulations using a source of data describing how real users move between different locations, i.e. wireless access points. A large amount of traces from the 802.11b campus network of Dartmouth College is available through the CRAWDAD project \[[14](#bib.bib14)\]. We selected all the contacts between 9 am and 6 pm in one day during the term (Wednesday 21 April 2004), discarding contacts with duration less than 60 seconds. Two users are connected only if they are associated with the same access point during a time slot: epidemics spreading is therefore performed among users co-located with access points. Our resulting data set had 2201 unique MACs and 11572 contacts with all access points. We assume that each MAC address corresponds to a unique user. The other simulation parameters are the same of the previous analysis.
<table>
<thead>
<tr>
<th>Type</th>
<th>Desired fraction</th>
<th>Delivered fraction</th>
<th>Messages sent</th>
</tr>
</thead>
<tbody>
<tr>
<td>epcast</td>
<td>0.50</td>
<td>0.43</td>
<td>17132</td>
</tr>
<tr>
<td>epcast</td>
<td>0.75</td>
<td>0.68</td>
<td>24738</td>
</tr>
<tr>
<td>epcast</td>
<td>1.00</td>
<td>0.90</td>
<td>32475</td>
</tr>
<tr>
<td>epcast(heterogeneous)</td>
<td>1.00</td>
<td>0.90</td>
<td>57342</td>
</tr>
<tr>
<td>Epidemic (β=0.25𝛽0.25\beta=0.25)</td>
<td>1.00</td>
<td>0.64</td>
<td>95969</td>
</tr>
<tr>
<td>Epidemic (β=0.50𝛽0.50\beta=0.50)</td>
<td>1.00</td>
<td>0.87</td>
<td>121873</td>
</tr>
<tr>
<td>Epidemic (β=1.00𝛽1.00\beta=1.00)</td>
<td>1.00</td>
<td>0.92</td>
<td>155446</td>
</tr>
</tbody>
</table>
Table 1: Comparation of performances on the real dataset of Dartmouth College traces
In Table [1](#S5.T1 "Table 1 ‣ 5.2.3 Evaluation with Dartmouth traces ‣ 5.2 Experimental Evaluation ‣ 5 Evaluation ‣ Epcast: Controlled Dissemination in Human-based Wireless Networks by means of Epidemic Spreading Models") we show the performances of our approach: the percentage of host actually reached is slight less than the desired fraction of population and this can be explained by observing that these hosts are not always connected during all the simulation time and may be easily absent from the underlying network. In other words, the underpinning hypothesis of the epidemic spreading model that we are using are only approximately satisfied. We run a simulation with a standard epidemic approach where infectivity is not tuned using the SIR model but it is set to 0.25, 0.50 and 1.00 respectively. It is interesting to note that the number of messages is in all three cases higher; only the case with infectivity equal to 1.00, the standard epidemic protocol is able to reach all the hosts. This is also demonstrate how it is difficult to choose the right value of the infectivity in a purely heuristic way to reach all the hosts of the system.
We run also some simulations using a dataset from the National University of Singapore\[[16](#bib.bib16)\], which contains contact pattern of 22341 students inferred from the information on class schedules and class rosters for the Spring semester of 2006. Two students are connected if they attend the same class during a time slot. However, in this dataset a large fraction of students is not included in the instantaneous underlying network, since they are not attending any class. The result is that in this case the epidemics fails to start using our model based on the assumption of homogeneous mixing. Additional virtual point of aggregation can be included in the simulations, grouping a percentage of the students that are not attending lectures during a particular timeslot: this modification ensures homogeneous mixing, providing good results for our algorithm. However, this is only a conjecture given the nature of the traces.
#### 5.2.4 Heterogeneous Networks
The results and the solutions discussed in this paper rely on the assumption of homogeneous networks, that are emerging from the random movements of the nodes. We now show that the proposed approach can be extended to the general case of heterogeneous networks \[[4](#bib.bib4), [5](#bib.bib5)\]. These structures are emerging in presence of small clusters of people or communities.
For heterogeneous networks the approximation $`k \approx {\langle k\rangle}`$ is not valid. However, the same probabilistic communication primitives introduced in Section [2](#S2 "2 Primitives for Controlled Epidemic Dissemination ‣ Epcast: Controlled Dissemination in Human-based Wireless Networks by means of Epidemic Spreading Models") could be used, with a different semantics. This relies on the following observations: given $`k`$ fluctuating in the range $`\lbrack k_{M\hspace{0pt}I\hspace{0pt}N},k_{M\hspace{0pt}A\hspace{0pt}X}\rbrack`$, we observe that for a value of the infectivity corresponding to $`k = k_{M\hspace{0pt}I\hspace{0pt}N}`$, the obtained spreading of the infection $`I\hspace{0pt}{(t^{\ast},k_{M\hspace{0pt}I\hspace{0pt}N})}`$ will always be greater than the one obtained with another $`k`$. In other words, if $`k_{M\hspace{0pt}I\hspace{0pt}N}`$ is selected in the calculation of the value of the infectivity, the value of `Reliability` can be considered approximately as a guaranteed lower bound of the reliability level.
The value of $`k_{M\hspace{0pt}I\hspace{0pt}N}`$ can be dynamically retrieved and set by the middleware by monitoring the connectivity of the hosts composing the mobile system. We plan to investigate these adaptive mechanisms further in the future.
## 6 Concluding Remarks
In this paper we have shown how models of epidemic spreading in complex networks can be applied effectively to the problem of disseminating information to subset of hosts (or to all the hosts) in a wireless network, controlling at the same time the number of the copies in the system. We have presented an analytical and experimental evaluation of our approach using a synthetic random model and real traces, showing the effectiveness of our approach.
Acknowledgements Cecilia Mascolo and Mirco Musolesi acknowledge the support of EPSRC through the CREAM Project. Salvo Scellato thanks UCL for the financial support as Visiting Student.
## References
- \[1\]
R. Albert and A.-L. Barabasi.
Statistical Mechanics of Complex Networks.
Review of Modern Physics, 74:47–97, 2002.
- \[2\]
R. M. Anderson and R. M. May.
Infectious Diseases of Humans: Dynamics and Control.
Oxford University Press, 1992.
- \[3\]
S. Baehni, C. Chabra, and R. Guerraoui.
Frugal Event Dissemination in a Mobile Environment.
In Proceedings of ACM Middleware’05, 2005.
- \[4\]
M. Barthélemy, A. Barrat, R. Pastor-Satorras, and A. Vespignani.
Dynamic Patterns of Epidemic Outbreaks in Complex Heterogeneous Networks.
Journal of Theoretical Biology, 2005.
- \[5\]
S. Boccaletti, V. Latora, Y. Moreno, M. Chavez, and D.-U. Hwang.
Complex networks : Structure and dynamics.
Phys. Rep., 424(4-5):175–308, Elsevier 2006.
- \[6\]
T. Camp, J. Boleng, and V. Davies.
A Survey of Mobility Models for Ad Hoc Network Research.
Wireless Communication and Mobile Computing, 2(5):483–502, 2002.
- \[7\]
P. Costa and G. P. Picco.
Semi-probabilistic Content-Based Publish-Subscribe.
In Proceedings of ICDCS’05, pages 575–585, 2005.
- \[8\]
A. Demers, D. Greene, C. Hauser, W. Irish, J. Larson, S. Shenker, H. Sturgis, D. Swinehart, and D. Terry.
Epidemic Algorithms for Replicated Database Maintenance.
ACM SIGOPS Operating Systems Review, 22(1), January 1988.
- \[9\]
P. T. Eugster, R. Guerraoui, A.-M. Kermarrec, and L. Massoulié.
Epidemic Information Dissemination in Distributed Systems.
IEEE Computer, May 2004.
- \[10\]
P. T. Eugster, S. Handurukande, R. Guerraoui, A.-M. Kermarrec, and P. Kouznetsov.
Lightweight Probabilistic Broadcast.
ACM Transactions on Computer Systems, 21(4):341–374, 2003.
- \[11\]
K. Fall.
A delay-tolerant network architecture for challenged internets.
In Proceedings of the SIGCOMM’03, pages 27–34, New York, NY, USA, 2003. ACM Press.
- \[12\]
I. Glauche, W. Krause, R. Sollacher, and M. Greiner.
Continuum Percolation of Wireless Ad Hoc Communication Networks.
Physica A, 325:577–600, 2003.
- \[13\]
M. Jelasity and A. Montresor.
Epidemic-style proactive aggregation in large overlay networks.
In Proceedings of ICDCS’04, pages 102–109, Tokyo, Japan, Mar. 2004. IEEE Computer Society.
- \[14\]
D. Kotz, T. Henderson, and I. Abyzov.
CRAWDAD data set dartmouth/campus (v. 2007-02-08).
Downloaded from http://crawdad.cs.dartmouth.edu/dartmouth/campus, Feb. 2007.
- \[15\]
M. Musolesi and C. Mascolo.
Controlled Epidemic-style Dissemination Middleware for Mobile Ad Hoc Networks.
In Proceedings of MOBIQUITOUS 2006. ACM Press, July 2006.
- \[16\]
V. Srinivasan, M. Motani, and W. T. Ooi.
CRAWDAD data set nus/contact (v. 2006-08-01).
Downloaded from http://crawdad.cs.dartmouth.edu/nus/contact, Aug. 2006.
- \[17\]
A. Vahdat and D. Becker.
Epidemic Routing for Partially Connected Ad Hoc Networks.
Technical Report CS-2000-06, Department of Computer Science, Duke University, 2000. |
0711.1269 | cs.NI | cyber | # Fair Scheduling in OFDMA-based Wireless Systems with QoS Constraints
Tolga Girici2, Chenxi Zhu1, Jonathan R. Agre1 and Anthony Ephremides2
tgirici@glue.umd.edu
2 Institute for Systems Research
University of Maryland 1 Fujitsu Labs of America
College Park, Maryland 20740
###### Abstract
In this work we consider the problem of downlink resource allocation for proportional fairness of long term received rates of data users and quality of service for real time sessions in an OFDMA-based wireless system. The base station allocates available power and bandwidth to individual users based on long term average received rates, QoS based rate constraints and channel conditions. We solve the underlying constrained optimization problem and propose an algorithm that achieves the optimal allocation. Numerical evaluation results show that the proposed algorithm provides better QoS to voice and video sessions while providing more and fair rates to data users in comparison with existing schemes.
## I Introduction
Recent IEEE 802.16 standard defines the air interface and medium access control (MAC) specifications for wireless metropolitan area networks. Such networks intend to provide high speed voice, data and on demand video streaming services for end users. IEEE 802.16 standard is often referred to as WiMax and it provides substantially higher rates than cellular networks. Besides it eliminates the costly infrastructure to deploy cables, therefore it is becoming an alternative to cabled networks, such as fiber optic and DSL systems \[[1](#bib.bib1)\],\[[2](#bib.bib2)\].
Transmissions in IEEE 802.16 networks is based on Orthogonal frequency division multiplexing (OFDM), which is a multicarrier transmission technique that is proposed for high speed wireless transmission. It is based on a large number of orthogonal subchannels, each working at a different frequency. Within OFDMA framework, the resource allocated to the users come in three dimensions: time, frequency and power. This requires the scheduler to operate with higher degree of freedom but also makes the notion of resource fairness obsolete and makes the problem more involved.
There are three main issues that need to be considered in multiple access resource allocation, which are spectral efficiency, fairness and Quality of Service (QoS). These are often contradicting aims and achieving these in an OFDMA based system is a challenging task. OFDMA-based scheduling for systems with heterogeneous traffic has not been studied sufficiently. Recently in \[[3](#bib.bib3)\], \[[4](#bib.bib4)\], \[[5](#bib.bib5)\], proportional fair scheduling was studied for multicarrier systems. However, the scheduling rules in these works do not apply sufficiently to different QoS requirements and heterogeneous traffic. Besides they only provide fairness for short term received rate, instead of long term. In this work our goal is to find algorithms that provide proportional fairness to data users and QoS for real time sessions such as voice and video streaming.
## II System Model
We consider a downlink system, where a base station transmits to N users. Each user is assumed to demand a single type of traffic. Let $`W`$ and $`P`$ denote the total bandwidth and power, respectively. Total bandwidth $`W`$ is divided into $`N_{s\hspace{0pt}u\hspace{0pt}b}`$ subchannels of length $`W_{s\hspace{0pt}u\hspace{0pt}b}`$, each consisting of a group of carriers. We consider PUSC as the mode of subchannelization, which is more suitable for mobile users than AMC. It provides frequency diversity and inter-cell interference averaging. This minimizes the performance degradation due to fast fading characteristics of mobile environments. We assume flat fading, which is a reasonable assumption with PUSC.
The noise and interference power density is $`N_{0}`$, and the channel gain averaged over the entire band from the BS to user $`i`$ at time $`t`$ is $`h_{i}\hspace{0pt}{(t)}`$, where $`h_{i}\hspace{0pt}{(t)}`$ includes path loss, shadowing (lognormal fading) and fast fading. There are three classes of users. Users in the classes $`U_{D}`$, $`U_{S}`$ and $`U_{V}`$ demand data, voice and video traffic, respectively. The system that we consider is time slotted with frame length $`T_{s}`$. The scheduler makes a resource allocation decision at each frame. Active period in a voice conversation, streaming duration and file size are both very long with respect to the frame size. Therefore it is realistic that during the course of optimization the number of active sessions are fixed.
In problem formulation we will adopt the Shannon channel capacity for AWGN channel as a function for bandwidth and transmission power assigned to user $`i`$:
<table>
<tbody>
<tr>
<td></td>
<td>ri(wi(t),pi(t))=wi(t)log(1+βpi(t)hi(t)N0wi(t))subscript𝑟𝑖subscript𝑤𝑖𝑡subscript𝑝𝑖𝑡subscript𝑤𝑖𝑡1𝛽subscript𝑝𝑖𝑡subscriptℎ𝑖𝑡subscript𝑁0subscript𝑤𝑖𝑡r_{i}(w_{i}(t),p_{i}(t))=w_{i}(t)\log\left(1+\beta\frac{p_{i}(t)h_{i}(t)}{N_{0}w_{i}(t)}\right)</td>
<td></td>
<td>(1)</td>
</tr>
</tbody>
</table>
The reason for using Shannon capacity is its simplicity, and it also approximates the above set of rate-SINR relation with $`\beta = 0.25`$. The parameter $`0 < \beta < 1`$ compensates the rate gap between Shannon capacity and rate achieved by practical modulation and coding pairs listed in IEEE 802.16a/e standards. After finding power and bandwidth pairs by solving the convex programming problem, we quantize the SINRs to the thresholds determined in the standard. In this paper, we propose a Fair and QoS-based Power and Subchannel Allocation (FQPSA) Algorithm. Objective of the algorithm is to maximize proportional fairness for data users subject to rate constraints of real time users selected each frame according to a metric.
## III Problem Formulation
###
III-A
Data Traffic
The objective for the data users is to optimize log sum of the exponential averaged rates of the users Average received user rates are updated at each frame according to the exponential averaging formula:
<table>
<tbody>
<tr>
<td></td>
<td>Ri(t)=αiRi(t−1)+(1−αi)ri(wi(t),pi(t)),∀i,tsubscript𝑅𝑖𝑡subscript𝛼𝑖subscript𝑅𝑖𝑡11subscript𝛼𝑖subscript𝑟𝑖subscript𝑤𝑖𝑡subscript𝑝𝑖𝑡for-all𝑖𝑡R_{i}(t)=\alpha_{i}R_{i}(t-1)+(1-\alpha_{i})r_{i}(w_{i}(t),p_{i}(t)),\forall i,t</td>
<td></td>
<td>(2)</td>
</tr>
</tbody>
</table>
This way we consider both current rate as well as rates given to the user in the past. Observed at time $`t`$, the highest consideration is given to the current rate $`r\hspace{0pt}{(t)}`$, and the rates received at the past $`t - 1`$, $`t - 2`$,… carry diminishing importance.
<table>
<tbody>
<tr>
<td></td>
<td>C(𝐑(t))=∑i=1NlogRi(t)=C(𝐑(t−1))+∑i=1Nlog(α+α¯ri(wi(t),pi(t))Ri(t−1))𝐶𝐑𝑡superscriptsubscript𝑖1𝑁subscript𝑅𝑖𝑡𝐶𝐑𝑡1superscriptsubscript𝑖1𝑁𝛼¯𝛼subscript𝑟𝑖subscript𝑤𝑖𝑡subscript𝑝𝑖𝑡subscript𝑅𝑖𝑡1C(\mathbf{R}(t))=\sum_{i=1}^{N}\log R_{i}(t)=\\ C(\mathbf{R}(t-1))+\sum_{i=1}^{N}\log\left(\alpha+\frac{\overline{\alpha}r_{i}(w_{i}(t),p_{i}(t))}{R_{i}(t-1)}\right)</td>
<td></td>
<td>(3)</td>
</tr>
</tbody>
</table>
As a matter of fact, we limit ourself to greedy schemes in the sense that at frame $`t`$, we try to maximize the proportional fair capacity $`C\hspace{0pt}{({\mathbf{R}\hspace{0pt}{(t)}})}`$ without considering the future frames $`t + 1`$,$`t + 2`$, etc. Only the second term in ([3](#S3.E3 "In III-A Data Traffic ‣ III Problem Formulation ‣ Fair Scheduling in OFDMA-based Wireless Systems with QoS Constraints")) needs consideration. Here the log-sum can be replaced with product.
###
III-B
Real Time Traffic
The Base Station first chooses a number of voice and video users according to a metric that reflects current channel condition, head of line packet delay and long term received rate. Then for the selected users the number of bits to be transmitted in the current frame is determined. These procedures are explained in \[[6](#bib.bib6)\], in detail. Let $`U_{R}'`$ be the set of real time sessions selected and let $`r_{i}^{c}`$ be the rate requirement of real time session $`i`$.
The resulting optimization problem considering all types of sessions is as follows: Find,
<sup>1</sup>
<sup>1</sup>1Here $`p_{i}`$,$`w_{i}`$, $`n_{i}`$ are the values at time $`t`$. The time index is not shown for convenience.
:
<table>
<tbody>
<tr>
<td></td>
<td>(𝐩∗,𝐰∗)=argmax𝐩,𝐰∏i∈UD(αi+αi¯wilog(1+piniwi)Ri)superscript𝐩superscript𝐰subscript𝐩𝐰subscriptproduct𝑖subscript𝑈𝐷subscript𝛼𝑖¯subscript𝛼𝑖subscript𝑤𝑖1subscript𝑝𝑖subscript𝑛𝑖subscript𝑤𝑖subscript𝑅𝑖(\mathbf{p^{*},w^{*}})=\arg\max_{\mathbf{p},\mathbf{w}}\prod_{i\in U_{D}}{\left(\alpha_{i}+\frac{\overline{\alpha_{i}}w_{i}\log\left({1+\frac{p_{i}}{n_{i}w_{i}}}\right)}{R_{i}}\right)}</td>
<td></td>
</tr>
</tbody>
</table>
subject to,
<table>
<tbody>
<tr>
<td></td>
<td>∑i∈UD∪UR′pi∗subscript𝑖subscript𝑈𝐷superscriptsubscript𝑈𝑅′subscriptsuperscript𝑝𝑖\displaystyle\sum_{i\in U_{D}\cup U_{R}^{\prime}}{p^{*}_{i}}</td>
<td>≤\displaystyle\leq</td>
<td>P𝑃\displaystyle P</td>
<td></td>
<td>(4)</td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td>∑i∈UD∪UR′wi∗subscript𝑖subscript𝑈𝐷superscriptsubscript𝑈𝑅′subscriptsuperscript𝑤𝑖\displaystyle\sum_{i\in U_{D}\cup U_{R}^{\prime}}{w^{*}_{i}}</td>
<td>≤\displaystyle\leq</td>
<td>W𝑊\displaystyle W</td>
<td></td>
<td>(5)</td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td>wi∗log(1+pi∗niwi∗)subscriptsuperscript𝑤𝑖1subscriptsuperscript𝑝𝑖subscript𝑛𝑖subscriptsuperscript𝑤𝑖\displaystyle w^{*}_{i}\log\left({1+\frac{p^{*}_{i}}{n_{i}w^{*}_{i}}}\right)</td>
<td>≥\displaystyle\geq</td>
<td>ric,i∈UR′superscriptsubscript𝑟𝑖𝑐𝑖superscriptsubscript𝑈𝑅′\displaystyle r_{i}^{c},i\in U_{R}^{\prime}</td>
<td></td>
<td>(6)</td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td>pi∗,wi∗subscriptsuperscript𝑝𝑖subscriptsuperscript𝑤𝑖\displaystyle p^{*}_{i},w^{*}_{i}</td>
<td>≥\displaystyle\geq</td>
<td>0,∀i∈UD∪UR′0for-all𝑖subscript𝑈𝐷superscriptsubscript𝑈𝑅′\displaystyle 0,\forall i\in U_{D}\cup U_{R}^{\prime}</td>
<td></td>
<td>(7)</td>
</tr>
</tbody>
</table>
where $`n_{i} = \frac{N_{0}}{\beta\hspace{0pt}h_{i}}`$.
## IV Solution of the Optimization Problem
The objective function ([](#S3.Ex1 "III-B Real Time Traffic ‣ III Problem Formulation ‣ Fair Scheduling in OFDMA-based Wireless Systems with QoS Constraints")
III-B
) is an increasing function of $`(w,p)`$, therefore the maximum is achieved only when the constraints ([4](#S3.E4 "In III-B Real Time Traffic ‣ III Problem Formulation ‣ Fair Scheduling in OFDMA-based Wireless Systems with QoS Constraints"), [5](#S3.E5 "In III-B Real Time Traffic ‣ III Problem Formulation ‣ Fair Scheduling in OFDMA-based Wireless Systems with QoS Constraints"), [6](#S3.E6 "In III-B Real Time Traffic ‣ III Problem Formulation ‣ Fair Scheduling in OFDMA-based Wireless Systems with QoS Constraints")) are all met with equality. For this reason we can replace these inequalities with equalities in the discussion below.
Actually there is no guarantee that a solution can be found that satisfies both rate and power constraints, given the rate requirements and channel conditions. The rate requirement can be too high that it may be impossible to satisfy with the given channel conditions. To start with, we assume that the problem is feasible. We will discuss about how to detect infeasibility of the problem and what to do in that case in the next section. We can write the Lagrangian of the problem as:
<table>
<tbody>
<tr>
<td></td>
<td>L(𝐰,𝐩,λp,λw,λr)=∏i∈UD(αi+αi¯wilog(1+piniwi)Ri)+λp(P−∑i∈UD∪UR′pi)+λw(W−∑i∈UD∪UR′wi)+∑i∈UR′λir(wilog(1+piniwi)−ri0).𝐿𝐰𝐩subscript𝜆𝑝subscript𝜆𝑤superscript𝜆𝑟subscriptproduct𝑖subscript𝑈𝐷subscript𝛼𝑖¯subscript𝛼𝑖subscript𝑤𝑖1subscript𝑝𝑖subscript𝑛𝑖subscript𝑤𝑖subscript𝑅𝑖subscript𝜆𝑝𝑃subscript𝑖subscript𝑈𝐷superscriptsubscript𝑈𝑅′subscript𝑝𝑖subscript𝜆𝑤𝑊subscript𝑖subscript𝑈𝐷superscriptsubscript𝑈𝑅′subscript𝑤𝑖subscript𝑖superscriptsubscript𝑈𝑅′superscriptsubscript𝜆𝑖𝑟subscript𝑤𝑖1subscript𝑝𝑖subscript𝑛𝑖subscript𝑤𝑖superscriptsubscript𝑟𝑖0L(\mathbf{w},\mathbf{p},\lambda_{p},\lambda_{w},\mathbf{\lambda}^{r})=\prod_{i\in U_{D}}{\left(\alpha_{i}+\frac{\overline{\alpha_{i}}w_{i}\log\left({1+\frac{p_{i}}{n_{i}w_{i}}}\right)}{R_{i}}\right)}\\ +\lambda_{p}\left(P-\sum_{i\in U_{D}\cup U_{R}^{\prime}}p_{i}\right)+\lambda_{w}\left(W-\sum_{i\in U_{D}\cup U_{R}^{\prime}}w_{i}\right)\\ +\sum_{i\in U_{R}^{\prime}}\lambda_{i}^{r}\left(w_{i}\log\left(1+\frac{p_{i}}{n_{i}w_{i}}\right)-r_{i}^{0}\right).</td>
<td></td>
<td>(8)</td>
</tr>
</tbody>
</table>
From now on we assume that all rates are in nats/sec and logarithms are natural. Taking the derivatives of $`L\hspace{0pt}{(\mathbf{p},\mathbf{w},\lambda_{p},\lambda_{w},\lambda^{r})}`$, we get the following:
###
IV-A
For users $`i \in U_{D}`$
<table>
<tbody>
<tr>
<td></td>
<td>∂L(𝐩,𝐰,λp,λw,λr)∂pi|(𝐩∗,𝐰∗)=0⇒L∗λp=(niwi∗+pi∗)(Riα~iwi∗+log(1+pi∗niwi∗))evaluated-at𝐿𝐩𝐰subscript𝜆𝑝subscript𝜆𝑤superscript𝜆𝑟subscript𝑝𝑖superscript𝐩superscript𝐰0⇒superscript𝐿subscript𝜆𝑝subscript𝑛𝑖superscriptsubscript𝑤𝑖superscriptsubscript𝑝𝑖subscript𝑅𝑖subscript~𝛼𝑖superscriptsubscript𝑤𝑖1superscriptsubscript𝑝𝑖subscript𝑛𝑖superscriptsubscript𝑤𝑖\left.\frac{\partial L(\mathbf{p},\mathbf{w},\lambda_{p},\lambda_{w},\mathbf{\lambda}^{r})}{\partial p_{i}}\right|_{(\mathbf{p}^{*},\mathbf{w}^{*})}=0~{}~{}\Rightarrow\\ \frac{L^{*}}{\lambda_{p}}=(n_{i}w_{i}^{*}+p_{i}^{*})\left(\frac{R_{i}\widetilde{\alpha}_{i}}{w_{i}^{*}}+\log\left(1+\frac{p_{i}^{*}}{n_{i}w_{i}^{*}}\right)\right)</td>
<td></td>
<td>(9)</td>
</tr>
</tbody>
</table>
<table>
<tbody>
<tr>
<td></td>
<td>∂L(𝐩,𝐰,λp,λw,λr)∂wi|(𝐩∗,𝐰∗)=0⇒L∗λw=(niwi∗+pi∗)(Riα~i+wi∗log(1+pi∗niwi∗))(niwi∗+pi∗)log(1+pi∗niwi∗)−pievaluated-at𝐿𝐩𝐰subscript𝜆𝑝subscript𝜆𝑤superscript𝜆𝑟subscript𝑤𝑖superscript𝐩superscript𝐰0⇒superscript𝐿subscript𝜆𝑤subscript𝑛𝑖superscriptsubscript𝑤𝑖superscriptsubscript𝑝𝑖subscript𝑅𝑖subscript~𝛼𝑖superscriptsubscript𝑤𝑖1superscriptsubscript𝑝𝑖subscript𝑛𝑖superscriptsubscript𝑤𝑖subscript𝑛𝑖superscriptsubscript𝑤𝑖superscriptsubscript𝑝𝑖1superscriptsubscript𝑝𝑖subscript𝑛𝑖superscriptsubscript𝑤𝑖subscript𝑝𝑖\left.\frac{\partial L(\mathbf{p},\mathbf{w},\lambda_{p},\lambda_{w},\mathbf{\lambda}^{r})}{\partial w_{i}}\right|_{(\mathbf{p}^{*},\mathbf{w}^{*})}=0~{}~{}\Rightarrow\\ \frac{L^{*}}{\lambda_{w}}=\frac{(n_{i}w_{i}^{*}+p_{i}^{*})\left(R_{i}\widetilde{\alpha}_{i}+w_{i}^{*}\log\left(1+\frac{p_{i}^{*}}{n_{i}w_{i}^{*}}\right)\right)}{(n_{i}w_{i}^{*}+p_{i}^{*})\log\left(1+\frac{p_{i}^{*}}{n_{i}w_{i}^{*}}\right)-p_{i}}</td>
<td></td>
<td>(10)</td>
</tr>
</tbody>
</table>
where $`{\overset{\sim}{\alpha}}_{i} = \frac{\alpha_{i}}{1 - \alpha_{i}}`$. Let $`\Lambda_{p} = \frac{L^{\ast}}{\lambda_{p}}`$ and $`\Lambda_{w} = \frac{L^{\ast}}{\lambda_{w}}`$. By dividing ([9](#S4.E9 "In IV-A For users i∈U_D ‣ IV Solution of the Optimization Problem ‣ Fair Scheduling in OFDMA-based Wireless Systems with QoS Constraints")) with ([10](#S4.E10 "In IV-A For users i∈U_D ‣ IV Solution of the Optimization Problem ‣ Fair Scheduling in OFDMA-based Wireless Systems with QoS Constraints")) we can write for all $`i \in U_{D}`$:
<table>
<tbody>
<tr>
<td></td>
<td>ΛpΛw=Λa=ni((1+xi∗)log(1+xi∗)−xi∗),subscriptΛ𝑝subscriptΛ𝑤subscriptΛ𝑎subscript𝑛𝑖1superscriptsubscript𝑥𝑖1superscriptsubscript𝑥𝑖superscriptsubscript𝑥𝑖\frac{\Lambda_{p}}{\Lambda_{w}}=\Lambda_{a}=n_{i}\left(\left(1+x_{i}^{*}\right)\log\left(1+x_{i}^{*}\right)-x_{i}^{*}\right),</td>
<td></td>
<td>(11)</td>
</tr>
</tbody>
</table>
where $`x_{i}^{\ast} = \frac{p_{i}^{\ast}}{n_{i}\hspace{0pt}w_{i}^{\ast}}`$ denotes the optimal effective SINR, which is the SINR multiplied by the SINR gap parameter $`\beta`$.
###
IV-B
For users $`i \in U_{R}'`$
<table>
<tbody>
<tr>
<td></td>
<td>∂L(𝐩,𝐰,λp,λw,λr)∂pi|(𝐩∗,𝐰∗)⇒λpλiv=1ni11+pi∗niwi∗⇒evaluated-at𝐿𝐩𝐰subscript𝜆𝑝subscript𝜆𝑤superscript𝜆𝑟subscript𝑝𝑖superscript𝐩superscript𝐰subscript𝜆𝑝superscriptsubscript𝜆𝑖𝑣1subscript𝑛𝑖11superscriptsubscript𝑝𝑖subscript𝑛𝑖superscriptsubscript𝑤𝑖\left.\frac{\partial L(\mathbf{p},\mathbf{w},\lambda_{p},\lambda_{w},\mathbf{\lambda}^{r})}{\partial p_{i}}\right|_{(\mathbf{p}^{*},\mathbf{w}^{*})}\Rightarrow\frac{\lambda_{p}}{\lambda_{i}^{v}}=\frac{1}{n_{i}}\frac{1}{1+\frac{p_{i}^{*}}{n_{i}w_{i}^{*}}}</td>
<td></td>
<td>(12)</td>
</tr>
</tbody>
</table>
<table>
<tbody>
<tr>
<td></td>
<td>∂L(𝐩,𝐰,λp,λw,λr)∂wi|(𝐩∗,𝐰∗)⇒λwλir=log(1+pi∗niwi∗)−pi∗niwi∗1+pi∗niwi∗⇒evaluated-at𝐿𝐩𝐰subscript𝜆𝑝subscript𝜆𝑤superscript𝜆𝑟subscript𝑤𝑖superscript𝐩superscript𝐰subscript𝜆𝑤superscriptsubscript𝜆𝑖𝑟1superscriptsubscript𝑝𝑖subscript𝑛𝑖superscriptsubscript𝑤𝑖superscriptsubscript𝑝𝑖subscript𝑛𝑖superscriptsubscript𝑤𝑖1superscriptsubscript𝑝𝑖subscript𝑛𝑖superscriptsubscript𝑤𝑖\left.\frac{\partial L(\mathbf{p},\mathbf{w},\lambda_{p},\lambda_{w},\mathbf{\lambda}^{r})}{\partial w_{i}}\right|_{(\mathbf{p}^{*},\mathbf{w}^{*})}\Rightarrow\\ \frac{\lambda_{w}}{\lambda_{i}^{r}}=\log\left(1+\frac{p_{i}^{*}}{n_{i}w_{i}^{*}}\right)-\frac{\frac{p_{i}^{*}}{n_{i}w_{i}^{*}}}{1+\frac{p_{i}^{*}}{n_{i}w_{i}^{*}}}</td>
<td></td>
<td>(13)</td>
</tr>
</tbody>
</table>
With $`\frac{\Lambda_{p}}{\Lambda_{w}} = \frac{\lambda_{w}}{\lambda_{p}}`$, dividing equation ([13](#S4.E13 "In IV-B For users i∈U_R' ‣ IV Solution of the Optimization Problem ‣ Fair Scheduling in OFDMA-based Wireless Systems with QoS Constraints")) to ([12](#S4.E12 "In IV-B For users i∈U_R' ‣ IV Solution of the Optimization Problem ‣ Fair Scheduling in OFDMA-based Wireless Systems with QoS Constraints")) for all $`i \in U_{R}'`$ again gives ([11](#S4.E11 "In IV-A For users i∈U_D ‣ IV Solution of the Optimization Problem ‣ Fair Scheduling in OFDMA-based Wireless Systems with QoS Constraints")).
By dividing ([13](#S4.E13 "In IV-B For users i∈U_R' ‣ IV Solution of the Optimization Problem ‣ Fair Scheduling in OFDMA-based Wireless Systems with QoS Constraints")) to ([12](#S4.E12 "In IV-B For users i∈U_R' ‣ IV Solution of the Optimization Problem ‣ Fair Scheduling in OFDMA-based Wireless Systems with QoS Constraints")) we can eliminate $`\lambda_{i}^{v}`$’s from the problem. It is worth noting that we get the same relation between $`\Lambda_{a}/n_{i}`$ and $`x_{i}`$ for both data and real time sessions. At this point it is convenient to define the function $`f_{a}\hspace{0pt}{(x)}`$ as:
<table>
<tbody>
<tr>
<td></td>
<td>fa(x)=(1+x)log(1+x)−x.subscript𝑓𝑎𝑥1𝑥1𝑥𝑥f_{a}(x)=(1+x)\log(1+x)-x.</td>
<td></td>
<td>(14)</td>
</tr>
</tbody>
</table>
Then we have,
<table>
<tbody>
<tr>
<td></td>
<td>xi=fa−1(Λa/ni),∀i∈UD∪UR′formulae-sequencesubscript𝑥𝑖superscriptsubscript𝑓𝑎1subscriptΛ𝑎subscript𝑛𝑖for-all𝑖subscript𝑈𝐷superscriptsubscript𝑈𝑅′x_{i}=f_{a}^{-1}(\Lambda_{a}/n_{i}),\forall i\in U_{D}\cup U_{R}^{\prime}</td>
<td></td>
<td>(15)</td>
</tr>
</tbody>
</table>
Signal to noise ratio ($`x_{i}`$) is a monotonic increasing function of $`\Lambda_{a}`$ for all users $`i \in {U_{D} \cup U_{R}'}`$.
For real time users we also have:
<table>
<tbody>
<tr>
<td></td>
<td>∂L(𝐩,𝐰,λp,λw,λr)∂λiv|(𝐩∗,𝐰∗)=0⇒ri0=wi∗log(1+pi∗niwi∗),∀i∈UR′formulae-sequenceevaluated-at𝐿𝐩𝐰subscript𝜆𝑝subscript𝜆𝑤superscript𝜆𝑟superscriptsubscript𝜆𝑖𝑣superscript𝐩superscript𝐰0⇒superscriptsubscript𝑟𝑖0superscriptsubscript𝑤𝑖1superscriptsubscript𝑝𝑖subscript𝑛𝑖superscriptsubscript𝑤𝑖for-all𝑖superscriptsubscript𝑈𝑅′\left.\frac{\partial L(\mathbf{p},\mathbf{w},\lambda_{p},\lambda_{w},\mathbf{\lambda}^{r})}{\partial\lambda_{i}^{v}}\right|_{(\mathbf{p}^{*},\mathbf{w}^{*})}=0~{}~{}\Rightarrow\\ r_{i}^{0}=w_{i}^{*}\log\left(1+\frac{p_{i}^{*}}{n_{i}w_{i}^{*}}\right),\forall i\in U_{R}^{\prime}</td>
<td></td>
<td>(16)</td>
</tr>
</tbody>
</table>
###
IV-C
For all nodes $`i \in {U_{D} \cup U_{R}'}`$
<table>
<tbody>
<tr>
<td></td>
<td>∂L(𝐩,𝐰,λp,λw,λr)∂λp|(𝐩∗,𝐰∗)=0⇒P=∑i∈Upi∗evaluated-at𝐿𝐩𝐰subscript𝜆𝑝subscript𝜆𝑤superscript𝜆𝑟subscript𝜆𝑝superscript𝐩superscript𝐰0⇒𝑃subscript𝑖𝑈superscriptsubscript𝑝𝑖\left.\frac{\partial L(\mathbf{p},\mathbf{w},\lambda_{p},\lambda_{w},\mathbf{\lambda}^{r})}{\partial\lambda_{p}}\right|_{(\mathbf{p}^{*},\mathbf{w}^{*})}=0~{}~{}\Rightarrow P=\sum_{i\in U}p_{i}^{*}</td>
<td></td>
<td>(17)</td>
</tr>
</tbody>
</table>
<table>
<tbody>
<tr>
<td></td>
<td>∂L(𝐩,𝐰,λp,λw,λr)∂λw|(𝐩∗,𝐰∗)=0⇒W=∑i∈Uwi∗evaluated-at𝐿𝐩𝐰subscript𝜆𝑝subscript𝜆𝑤superscript𝜆𝑟subscript𝜆𝑤superscript𝐩superscript𝐰0⇒𝑊subscript𝑖𝑈superscriptsubscript𝑤𝑖\left.\frac{\partial L(\mathbf{p},\mathbf{w},\lambda_{p},\lambda_{w},\mathbf{\lambda}^{r})}{\partial\lambda_{w}}\right|_{(\mathbf{p}^{*},\mathbf{w}^{*})}=0~{}~{}\Rightarrow W=\sum_{i\in U}w_{i}^{*}</td>
<td></td>
<td>(18)</td>
</tr>
</tbody>
</table>
From Equation ([9](#S4.E9 "In IV-A For users i∈U_D ‣ IV Solution of the Optimization Problem ‣ Fair Scheduling in OFDMA-based Wireless Systems with QoS Constraints")) for data users we can write:
<table>
<tbody>
<tr>
<td></td>
<td>[Λp−ni(1+xi∗)Riα~i]+log(1+xi∗)(1+xi∗)nisuperscriptdelimited-[]subscriptΛ𝑝subscript𝑛𝑖1superscriptsubscript𝑥𝑖subscript𝑅𝑖subscript~𝛼𝑖1superscriptsubscript𝑥𝑖1superscriptsubscript𝑥𝑖subscript𝑛𝑖\displaystyle\frac{[\Lambda_{p}-n_{i}(1+x_{i}^{*})R_{i}\widetilde{\alpha}_{i}]^{+}}{\log(1+x_{i}^{*})(1+x_{i}^{*})n_{i}}</td>
<td>=\displaystyle=</td>
<td>wi∗,i∈UDsuperscriptsubscript𝑤𝑖𝑖subscript𝑈𝐷\displaystyle w_{i}^{*},i\in U_{D}</td>
<td></td>
<td>(19)</td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td>[Λp−ni(1+xi∗)Riα~i]+xi∗log(1+xi∗)(1+xi∗)superscriptdelimited-[]subscriptΛ𝑝subscript𝑛𝑖1superscriptsubscript𝑥𝑖subscript𝑅𝑖subscript~𝛼𝑖superscriptsubscript𝑥𝑖1superscriptsubscript𝑥𝑖1superscriptsubscript𝑥𝑖\displaystyle\frac{[\Lambda_{p}-n_{i}(1+x_{i}^{*})R_{i}\widetilde{\alpha}_{i}]^{+}x_{i}^{*}}{\log(1+x_{i}^{*})(1+x_{i}^{*})}</td>
<td>=\displaystyle=</td>
<td>pi∗,i∈UDsuperscriptsubscript𝑝𝑖𝑖subscript𝑈𝐷\displaystyle p_{i}^{*},i\in U_{D}</td>
<td></td>
<td>(20)</td>
</tr>
</tbody>
</table>
The $`{\lbrack.\rbrack}^{+}`$ operator in Equations ([19](#S4.E19 "In IV-C For all nodes i∈U_D∪U_R' ‣ IV Solution of the Optimization Problem ‣ Fair Scheduling in OFDMA-based Wireless Systems with QoS Constraints")),([20](#S4.E20 "In IV-C For all nodes i∈U_D∪U_R' ‣ IV Solution of the Optimization Problem ‣ Fair Scheduling in OFDMA-based Wireless Systems with QoS Constraints")) guarantees that $`{w_{i},p_{i}} \geq 0`$ for all users. Given $`\Lambda_{p}`$ and $`\Lambda_{a}`$ we can compute the power and bandwidth for users $`i \in U_{D}`$ using ([19](#S4.E19 "In IV-C For all nodes i∈U_D∪U_R' ‣ IV Solution of the Optimization Problem ‣ Fair Scheduling in OFDMA-based Wireless Systems with QoS Constraints")) and ([20](#S4.E20 "In IV-C For all nodes i∈U_D∪U_R' ‣ IV Solution of the Optimization Problem ‣ Fair Scheduling in OFDMA-based Wireless Systems with QoS Constraints")). Given $`\Lambda_{a}`$, we can calculate the power and bandwidth for users $`i \in U_{R}'`$ using ([16](#S4.E16 "In IV-B For users i∈U_R' ‣ IV Solution of the Optimization Problem ‣ Fair Scheduling in OFDMA-based Wireless Systems with QoS Constraints")). Please note that just from ([16](#S4.E16 "In IV-B For users i∈U_R' ‣ IV Solution of the Optimization Problem ‣ Fair Scheduling in OFDMA-based Wireless Systems with QoS Constraints")), ([19](#S4.E19 "In IV-C For all nodes i∈U_D∪U_R' ‣ IV Solution of the Optimization Problem ‣ Fair Scheduling in OFDMA-based Wireless Systems with QoS Constraints")) and ([20](#S4.E20 "In IV-C For all nodes i∈U_D∪U_R' ‣ IV Solution of the Optimization Problem ‣ Fair Scheduling in OFDMA-based Wireless Systems with QoS Constraints")), the bandwidth and power constraints ([5](#S3.E5 "In III-B Real Time Traffic ‣ III Problem Formulation ‣ Fair Scheduling in OFDMA-based Wireless Systems with QoS Constraints")) ([4](#S3.E4 "In III-B Real Time Traffic ‣ III Problem Formulation ‣ Fair Scheduling in OFDMA-based Wireless Systems with QoS Constraints")) are not necessarily satisfied. We need to find the right $`\Lambda_{a}`$ and $`\Lambda_{p}`$ so that the power and bandwidth constraints are satisfied. Let $`S_{p}\hspace{0pt}{(\Lambda_{a},\Lambda_{p})}`$ and $`S_{w}\hspace{0pt}{(\Lambda_{a},\Lambda_{p})}`$ be the total bandwidth and total power corresponding to $`\Lambda_{a}`$ and $`\Lambda_{p}`$:
<table>
<tbody>
<tr>
<td></td>
<td>Sw(Λa,Λp)=∑i∈UD∪UR′wi(Λa,Λp)=∑i∈UD[Λp−ni(1+xi)Riα~i]+log(1+xi)(1+xi)ni+∑i∈UR′riclog(1+xi)subscript𝑆𝑤subscriptΛ𝑎subscriptΛ𝑝subscript𝑖subscript𝑈𝐷superscriptsubscript𝑈𝑅′subscript𝑤𝑖subscriptΛ𝑎subscriptΛ𝑝subscript𝑖subscript𝑈𝐷superscriptdelimited-[]subscriptΛ𝑝subscript𝑛𝑖1subscript𝑥𝑖subscript𝑅𝑖subscript~𝛼𝑖1subscript𝑥𝑖1subscript𝑥𝑖subscript𝑛𝑖subscript𝑖superscriptsubscript𝑈𝑅′superscriptsubscript𝑟𝑖𝑐1subscript𝑥𝑖S_{w}(\Lambda_{a},\Lambda_{p})=\sum_{i\in U_{D}\cup U_{R}^{\prime}}{w_{i}(\Lambda_{a},\Lambda_{p})}=\\ \sum_{i\in U_{D}}\frac{[\Lambda_{p}-n_{i}(1+x_{i})R_{i}\widetilde{\alpha}_{i}]^{+}}{\log(1+x_{i})(1+x_{i})n_{i}}+\sum_{i\in U_{R}^{\prime}}\frac{r_{i}^{c}}{\log{(1+x_{i})}}</td>
<td></td>
<td>(21)</td>
</tr>
</tbody>
</table>
<table>
<tbody>
<tr>
<td></td>
<td>Sp(Λa,Λp)=∑i∈UD∪UR′pi(Λa,Λp)=∑i∈UD[Λp−ni(1+xi)Riα~i]+xilog(1+xi)(1+xi)+∑i∈UR′ricxinilog(1+xi)subscript𝑆𝑝subscriptΛ𝑎subscriptΛ𝑝subscript𝑖subscript𝑈𝐷superscriptsubscript𝑈𝑅′subscript𝑝𝑖subscriptΛ𝑎subscriptΛ𝑝subscript𝑖subscript𝑈𝐷superscriptdelimited-[]subscriptΛ𝑝subscript𝑛𝑖1subscript𝑥𝑖subscript𝑅𝑖subscript~𝛼𝑖subscript𝑥𝑖1subscript𝑥𝑖1subscript𝑥𝑖subscript𝑖superscriptsubscript𝑈𝑅′superscriptsubscript𝑟𝑖𝑐subscript𝑥𝑖subscript𝑛𝑖1subscript𝑥𝑖S_{p}(\Lambda_{a},\Lambda_{p})=\sum_{i\in U_{D}\cup U_{R}^{\prime}}{p_{i}(\Lambda_{a},\Lambda_{p})}=\\ \sum_{i\in U_{D}}\frac{[\Lambda_{p}-n_{i}(1+x_{i})R_{i}\widetilde{\alpha}_{i}]^{+}x_{i}}{\log(1+x_{i})(1+x_{i})}+\sum_{i\in U_{R}^{\prime}}\frac{r_{i}^{c}x_{i}n_{i}}{\log{(1+x_{i})}}</td>
<td></td>
<td>(22)</td>
</tr>
</tbody>
</table>
where $`x_{i} = {f_{a}^{- 1}\hspace{0pt}{({\Lambda_{a}/n_{i}})}}`$ is the SNR of user $`i`$. As a result, the problem is finding $`\Lambda_{a}^{\ast}`$ and $`\Lambda_{p}^{\ast}`$ such that
<table>
<tbody>
<tr>
<td></td>
<td>Sw(Λa∗,Λp∗)subscript𝑆𝑤subscriptsuperscriptΛ𝑎subscriptsuperscriptΛ𝑝\displaystyle S_{w}(\Lambda^{*}_{a},\Lambda^{*}_{p})</td>
<td>=\displaystyle=</td>
<td>W𝑊\displaystyle W</td>
<td></td>
<td>(23)</td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td>Sp(Λa∗,Λp∗)subscript𝑆𝑝subscriptsuperscriptΛ𝑎subscriptsuperscriptΛ𝑝\displaystyle S_{p}(\Lambda^{*}_{a},\Lambda^{*}_{p})</td>
<td>=\displaystyle=</td>
<td>P𝑃\displaystyle P</td>
<td></td>
<td>(24)</td>
</tr>
</tbody>
</table>
using Equations ([15](#S4.E15 "In IV-B For users i∈U_R' ‣ IV Solution of the Optimization Problem ‣ Fair Scheduling in OFDMA-based Wireless Systems with QoS Constraints")),([23](#S4.E23 "In IV-C For all nodes i∈U_D∪U_R' ‣ IV Solution of the Optimization Problem ‣ Fair Scheduling in OFDMA-based Wireless Systems with QoS Constraints")) and ([24](#S4.E24 "In IV-C For all nodes i∈U_D∪U_R' ‣ IV Solution of the Optimization Problem ‣ Fair Scheduling in OFDMA-based Wireless Systems with QoS Constraints")). Note that although $`\Lambda_{a}`$ and $`\Lambda_{p}`$ are independent variables that determine power and bandwidth for each node, they become dependent when the power and bandwidth constraints ([23](#S4.E23 "In IV-C For all nodes i∈U_D∪U_R' ‣ IV Solution of the Optimization Problem ‣ Fair Scheduling in OFDMA-based Wireless Systems with QoS Constraints")) ([24](#S4.E24 "In IV-C For all nodes i∈U_D∪U_R' ‣ IV Solution of the Optimization Problem ‣ Fair Scheduling in OFDMA-based Wireless Systems with QoS Constraints")) need to be satisfied.
###
IV-D
Feasibility of the Solution
From ([19](#S4.E19 "In IV-C For all nodes i∈U_D∪U_R' ‣ IV Solution of the Optimization Problem ‣ Fair Scheduling in OFDMA-based Wireless Systems with QoS Constraints")) and ([20](#S4.E20 "In IV-C For all nodes i∈U_D∪U_R' ‣ IV Solution of the Optimization Problem ‣ Fair Scheduling in OFDMA-based Wireless Systems with QoS Constraints")), $`\Lambda_{p} = 0`$ corresponds to the case that no bandwidth and power is allocated to data sessions. If the problem is feasible, then there exists a $`\Lambda_{a}`$ that satisfies the following:
<table>
<tbody>
<tr>
<td></td>
<td>Sw(Λa,0)subscript𝑆𝑤subscriptΛ𝑎0\displaystyle S_{w}(\Lambda_{a},0)</td>
<td>=\displaystyle=</td>
<td>W,𝑊\displaystyle W,</td>
<td></td>
<td>(25)</td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td>Sp(Λa,0)subscript𝑆𝑝subscriptΛ𝑎0\displaystyle S_{p}(\Lambda_{a},0)</td>
<td>≤\displaystyle\leq</td>
<td>P.𝑃\displaystyle P.</td>
<td></td>
<td>(26)</td>
</tr>
</tbody>
</table>
Let $`\Lambda_{a}^{0}`$ be the smallest $`\Lambda_{a}`$ that satisfies ([25](#S4.E25 "In IV-D Feasibility of the Solution ‣ IV Solution of the Optimization Problem ‣ Fair Scheduling in OFDMA-based Wireless Systems with QoS Constraints")) and ([26](#S4.E26 "In IV-D Feasibility of the Solution ‣ IV Solution of the Optimization Problem ‣ Fair Scheduling in OFDMA-based Wireless Systems with QoS Constraints")). In the report \[[6](#bib.bib6)\] we proved the existence and uniqueness of $`\Lambda_{a}^{0}`$ for a feasible problem using monotonicity properties of the functions $`S_{w}\hspace{0pt}{(\Lambda_{a},\Lambda_{p})}`$ and $`S_{p}\hspace{0pt}{(\Lambda_{a},\Lambda_{p})}`$.
## V Proposed Algorithm
Using ([15](#S4.E15 "In IV-B For users i∈U_R' ‣ IV Solution of the Optimization Problem ‣ Fair Scheduling in OFDMA-based Wireless Systems with QoS Constraints")),([21](#S4.E21 "In IV-C For all nodes i∈U_D∪U_R' ‣ IV Solution of the Optimization Problem ‣ Fair Scheduling in OFDMA-based Wireless Systems with QoS Constraints")) and ([22](#S4.E22 "In IV-C For all nodes i∈U_D∪U_R' ‣ IV Solution of the Optimization Problem ‣ Fair Scheduling in OFDMA-based Wireless Systems with QoS Constraints")) we can develop an algorithm to determine the power and bandwidth allocation. The algorithm is also able to detect infeasibility if there no solution exists.
Algorithm:
1. 1\.
Find $`\Lambda_{a}^{0}`$. If $`{S_{p}\hspace{0pt}{(\Lambda_{a}^{0},0)}} < P`$ then the problem is feasible. If not, find the real time session that consumes the largest power and decrease its rate. Repeat Step 1 and 2 until feasibility is achieved.
2. 2\.
Using two nested binary searches find the unique $`(\Lambda_{a}^{\ast},{\Lambda_{p}^{w}\hspace{0pt}{(\Lambda_{a}^{\ast})}})`$ that satisfies power and bandwidth constraints.
3. 3\.
Compute optimal power and bandwidth using the optimal $`(\Lambda_{a}^{\ast},{\Lambda_{p}^{w}\hspace{0pt}{(\Lambda_{a}^{\ast})}})`$.
Let us summarize the nested binary search. We start from a $`\Lambda_{a}`$, then we find the $`\Lambda_{p}^{w}\hspace{0pt}{(\Lambda_{a})}`$ that satisfies the bandwidth constraint. If the power constraint is exceeded then we decrease $`\Lambda_{a}'`$, else increase it, and search for $`\Lambda_{p}^{w}\hspace{0pt}{(\Lambda_{a}')}`$. We continue until both constraints are satisfied.
Since the problem is convex (proved in \[[6](#bib.bib6)\]), the problem has a unique solution. We need the monotonicity property of function $`S_{p}\hspace{0pt}{(\Lambda_{a},{\Lambda_{p}^{w}\hspace{0pt}{(\Lambda_{a})}})}`$ (sum of powers) as a function of $`\Lambda_{a}`$ in order to be able to find it using binary search. Those properties are also verified in \[[6](#bib.bib6)\].
## VI Performance Evaluation
For the numerical evaluations we equally divide the users to 5 classes according to the distances, 0.3,0.6,0.9,1.2,1.5 km. We use the parameters in Table [I](#S6.T1 "TABLE I ‣ VI Performance Evaluation ‣ Fair Scheduling in OFDMA-based Wireless Systems with QoS Constraints").
<table>
<thead>
<tr>
<th>Parameter</th>
<th>Value</th>
</tr>
</thead>
<tbody>
<tr>
<th>Cell radius</th>
<td>1.5km</td>
</tr>
<tr>
<th>User Distances</th>
<td>0.3,0.6,0.9,1.2,1.5 km</td>
</tr>
<tr>
<th>Total power (P)</th>
<td>20 W</td>
</tr>
<tr>
<th>Total bandwidth (W)</th>
<td>8.3 MHz</td>
</tr>
<tr>
<th>Frame Length</th>
<td>1 msec</td>
</tr>
<tr>
<th>Voice Traffic</th>
<td>CBR 32kbps</td>
</tr>
<tr>
<th>Video Traffic</th>
<td>802.16 - 128kbps</td>
</tr>
<tr>
<th>FTP File</th>
<td>5 MB</td>
</tr>
<tr>
<th>AWGN p.s.d.(N0subscript𝑁0N_{0})</th>
<td>-174dBm/Hz</td>
</tr>
<tr>
<th>Pathloss exponent (γ𝛾\gamma)</th>
<td>3.5</td>
</tr>
<tr>
<th>ψDB∼N(μψdB,σψdB)similar-tosubscript𝜓𝐷𝐵𝑁subscript𝜇subscript𝜓𝑑𝐵subscript𝜎subscript𝜓𝑑𝐵\psi_{DB}\sim N(\mu_{\psi_{dB}},\sigma_{\psi_{dB}})</th>
<td>N(0dB,8dB)</td>
</tr>
<tr>
<th>Coherent Time (Fast/Slow Fading)</th>
<td>(5msec/300msec.)</td>
</tr>
<tr>
<th>Pathloss(dB, d in meters)</th>
<td>−31.5−35log10d+ψdB31.535subscript10𝑑subscript𝜓𝑑𝐵-31.5-35\log_{10}d+\psi_{dB}</td>
</tr>
</tbody>
</table>
TABLE I: Simulation Parameters
We will use M-LWDF-PF scheme as benchmark in our simulations. In this scheme at each frame the user maximizing the following quantity transmits.
<table>
<tbody>
<tr>
<td></td>
<td>aiDiHOL(t)log(1+βP/N0W),subscript𝑎𝑖superscriptsubscript𝐷𝑖𝐻𝑂𝐿𝑡1𝛽𝑃subscript𝑁0𝑊a_{i}D_{i}^{HOL}(t)\log(1+\beta P/N_{0}W),</td>
<td></td>
<td>(27)</td>
</tr>
</tbody>
</table>
where $`D_{i}^{H\hspace{0pt}O\hspace{0pt}L}\hspace{0pt}{(t)}`$ is the head of line packet delay and $`r_{i}\hspace{0pt}{(t)}`$ is the channel capacity of user i at frame t. The parameter $`a_{i}`$ is a positive constant. The constant $`a_{i}`$ is defined as $`a_{i} = {- \frac{\log{(\delta_{i})}}{D_{i}^{m\hspace{0pt}a\hspace{0pt}x}\hspace{0pt}R_{i}\hspace{0pt}{(t)}}}`$ , which is referred to as M-LWDF-PF \[[7](#bib.bib7)\] \[[4](#bib.bib4)\]. M-LWDF-PF can be adapted to OFDMA systems as follows. Power is distributed equally to all subchannels. Starting from the first subchannel , the subchannel is allocated to the user maximizing ([27](#S6.E27 "In VI Performance Evaluation ‣ Fair Scheduling in OFDMA-based Wireless Systems with QoS Constraints")). Then the received rate $`R\hspace{0pt}{(t)}`$ is updated according to ([2](#S3.E2 "In III-A Data Traffic ‣ III Problem Formulation ‣ Fair Scheduling in OFDMA-based Wireless Systems with QoS Constraints")). All the subchannels are allocated one-by-one according to this rule. Delay exceeding probability is taken as $`\delta_{i} = 0.05`$ for all users. Delay constraint for voice, video and data users are 0.1, 0.4 and 1 seconds, respectively.
Tables [II](#S6.T2 "TABLE II ‣ VI Performance Evaluation ‣ Fair Scheduling in OFDMA-based Wireless Systems with QoS Constraints"), [III](#S6.T3 "TABLE III ‣ VI Performance Evaluation ‣ Fair Scheduling in OFDMA-based Wireless Systems with QoS Constraints") and [IV](#S6.T4 "TABLE IV ‣ VI Performance Evaluation ‣ Fair Scheduling in OFDMA-based Wireless Systems with QoS Constraints") show the effects of increasing the number of voice, video streaming and data users, respectively. In these tables we observe that FPQS Algorithm is better than M-LWDF algorithm in terms of both delay performance and data performance. With FQPS delay for voice and video sessions stay within acceptable bounds, while with M-LWDF, it exceeds the bounds for the user at the cell edge when $`V \geq 30`$ or $`S > 40`$. Besides, FQPSA provides at least 10 percent increase in total throughput. Total throughput decreases linearly with increasing number of voice and video users. Although 10 voice users adds up to 0.32 Mbps, adding 10 users decreases the total throughput approximately by 1.2-1.4 Mbps. This is because voice has a very strict delay requirement and a voice session may have to be transmitted despite bad channel conditions. Throughput for LWDF decreases with a little bit slower rate but that reflects to the voice and video performance negatively. Log-sum performance of FQPSA is also better than that of M-LWDF, which shows that our algorithm provides fairness.
<table>
<tbody>
<tr>
<td>V =</td>
<td>10</td>
<td>20</td>
<td>30</td>
<td>40</td>
<td>50</td>
<td>60</td>
</tr>
<tr>
<td>FQPS Voice (G)</td>
<td>36</td>
<td>33</td>
<td>30</td>
<td>32</td>
<td>31</td>
<td>31</td>
</tr>
<tr>
<td>FQPS Voice (B)</td>
<td>88</td>
<td>66</td>
<td>62</td>
<td>81</td>
<td>82</td>
<td>80</td>
</tr>
<tr>
<td>LWDF Voice (G)</td>
<td>41</td>
<td>41</td>
<td>41</td>
<td>41</td>
<td>41</td>
<td>41</td>
</tr>
<tr>
<td>LWDF Voice (B)</td>
<td>89</td>
<td>91</td>
<td>93</td>
<td>110</td>
<td>133</td>
<td>151</td>
</tr>
<tr>
<td>FQPS Video (G)</td>
<td>148</td>
<td>132</td>
<td>130</td>
<td>132</td>
<td>132</td>
<td>131</td>
</tr>
<tr>
<td>FQPS Video (B)</td>
<td>295</td>
<td>271</td>
<td>250</td>
<td>263</td>
<td>266</td>
<td>252</td>
</tr>
<tr>
<td>LWDF Video (G)</td>
<td>180</td>
<td>187</td>
<td>192</td>
<td>200</td>
<td>202</td>
<td>208</td>
</tr>
<tr>
<td>LWDF Video (B)</td>
<td>384</td>
<td>399</td>
<td>408</td>
<td>410</td>
<td>423</td>
<td>435</td>
</tr>
<tr>
<td>FQPS (Mbps)</td>
<td>26.9</td>
<td>25.2</td>
<td>23.8</td>
<td>22.7</td>
<td>21.4</td>
<td>20.1</td>
</tr>
<tr>
<td>FQPS log-sum</td>
<td>280</td>
<td>279</td>
<td>278</td>
<td>277</td>
<td>276</td>
<td>275</td>
</tr>
<tr>
<td>LWDF (Mbps)</td>
<td>23.9</td>
<td>22.7</td>
<td>21.6</td>
<td>20.5</td>
<td>19.4</td>
<td>18.5</td>
</tr>
<tr>
<td>LWDF Log-sum</td>
<td>275</td>
<td>278</td>
<td>276</td>
<td>274</td>
<td>272</td>
<td>269</td>
</tr>
</tbody>
</table>
TABLE II: Effects of increasing $`\#`$ of Voice users for $`{S = 20},{D = 20}`$. Units: Delay (msec), t.put (Mbps)
<table>
<tbody>
<tr>
<td>S =</td>
<td>10</td>
<td>20</td>
<td>30</td>
<td>40</td>
<td>50</td>
<td>60</td>
</tr>
<tr>
<td>FQPS Voice (G)</td>
<td>31</td>
<td>33</td>
<td>39</td>
<td>43</td>
<td>49</td>
<td>51</td>
</tr>
<tr>
<td>FQPS Voice (B)</td>
<td>57</td>
<td>66</td>
<td>71</td>
<td>87</td>
<td>93</td>
<td>108</td>
</tr>
<tr>
<td>LWDF Voice (G)</td>
<td>41</td>
<td>41</td>
<td>41</td>
<td>41</td>
<td>41</td>
<td>41</td>
</tr>
<tr>
<td>LWDF Voice (B)</td>
<td>86</td>
<td>91</td>
<td>94</td>
<td>101</td>
<td>142</td>
<td>117</td>
</tr>
<tr>
<td>FQPS Video (G)</td>
<td>122</td>
<td>132</td>
<td>150</td>
<td>167</td>
<td>174</td>
<td>190</td>
</tr>
<tr>
<td>FQPS Video (B)</td>
<td>217</td>
<td>271</td>
<td>275</td>
<td>301</td>
<td>338</td>
<td>392</td>
</tr>
<tr>
<td>LWDF Video (G)</td>
<td>171</td>
<td>187</td>
<td>190</td>
<td>194</td>
<td>202</td>
<td>223</td>
</tr>
<tr>
<td>LWDF Video (B)</td>
<td>395</td>
<td>399</td>
<td>399</td>
<td>414</td>
<td>463</td>
<td>519</td>
</tr>
<tr>
<td>FQPS D(Mbps)</td>
<td>27.2</td>
<td>25.2</td>
<td>23.5</td>
<td>21.5</td>
<td>19.5</td>
<td>17.5</td>
</tr>
<tr>
<td>FQPS Log-sum</td>
<td>281</td>
<td>279</td>
<td>278</td>
<td>276</td>
<td>274</td>
<td>272</td>
</tr>
<tr>
<td>LWDF (Mbps)</td>
<td>25.1</td>
<td>22.7</td>
<td>20.7</td>
<td>18.6</td>
<td>16.7</td>
<td>14.6</td>
</tr>
<tr>
<td>LWDF log-sum</td>
<td>279</td>
<td>277</td>
<td>277</td>
<td>276</td>
<td>275</td>
<td>273</td>
</tr>
</tbody>
</table>
TABLE III: Effects of increasing $`\#`$ of Video Streaming users for $`{D = 20},{V = 20}`$. Units: Delay (msec), t.put (Mbps)
In Table [IV](#S6.T4 "TABLE IV ‣ VI Performance Evaluation ‣ Fair Scheduling in OFDMA-based Wireless Systems with QoS Constraints") we can observe the effects of increasing the number of data users.We observe that delay for both voice and video streaming sessions stay approximately constant. Delay performance is much better than that of M-LWDF algorithm. Data performance is 10 percent better than M-LWDF. Total throughput increases with number of data users, but the increase diminishes as D increases.
<table>
<tbody>
<tr>
<td>D =</td>
<td>20</td>
<td>30</td>
<td>40</td>
<td>50</td>
<td>60</td>
<td>70</td>
</tr>
<tr>
<td>FQPS Voice (G)</td>
<td>38</td>
<td>33</td>
<td>34</td>
<td>34</td>
<td>36</td>
<td>39</td>
</tr>
<tr>
<td>FQPS Voice (B)</td>
<td>72</td>
<td>66</td>
<td>69</td>
<td>71</td>
<td>71</td>
<td>91</td>
</tr>
<tr>
<td>LWDF Voice (G)</td>
<td>41</td>
<td>47</td>
<td>50</td>
<td>57</td>
<td>61</td>
<td>67</td>
</tr>
<tr>
<td>LWDF Voice (B)</td>
<td>91</td>
<td>106</td>
<td>126</td>
<td>146</td>
<td>170</td>
<td>182</td>
</tr>
<tr>
<td>FQPS Video (G)</td>
<td>155</td>
<td>132</td>
<td>143</td>
<td>145</td>
<td>142</td>
<td>168</td>
</tr>
<tr>
<td>FQPS Video (B)</td>
<td>278</td>
<td>271</td>
<td>246</td>
<td>270</td>
<td>270</td>
<td>323</td>
</tr>
<tr>
<td>LWDF Video (G)</td>
<td>187</td>
<td>232</td>
<td>265</td>
<td>297</td>
<td>335</td>
<td>357</td>
</tr>
<tr>
<td>LWDF Video (B)</td>
<td>399</td>
<td>460</td>
<td>537</td>
<td>557</td>
<td>612</td>
<td>656</td>
</tr>
<tr>
<td>FQPS (Mbps)</td>
<td>24.5</td>
<td>25.2</td>
<td>25.8</td>
<td>26.1</td>
<td>26.2</td>
<td>26.6</td>
</tr>
<tr>
<td>FQPS Log-sum</td>
<td>146</td>
<td>279</td>
<td>407</td>
<td>532</td>
<td>654</td>
<td>775</td>
</tr>
<tr>
<td>LWDF (Mbps)</td>
<td>22.7</td>
<td>23.3</td>
<td>23.7</td>
<td>24.0</td>
<td>24.1</td>
<td>24.3</td>
</tr>
<tr>
<td>LWDF Log-sum</td>
<td>146</td>
<td>278</td>
<td>405</td>
<td>530</td>
<td>653</td>
<td>772</td>
</tr>
</tbody>
</table>
TABLE IV: Effects of increasing $`\#`$ of Data users for $`{S = 20},{V = 20}`$. Units: Delay (msec), t.put (Mbps)
## References
- \[1\] C. Eklund, R. B. Marks, K.L. Stanwood, and S. Wang.
IEEE Standard 802.16: A Technical Overview of the WirelessMAN Air Interface for Broadband Wireless Access.
IEEE Communications Magazine, June 2002.
- \[2\] A. Ghosh, D. Wolter, J. G. Andres, and R Chen. Broadband Wireless Access with WiMax/802.16: Current Performance Benchmarks and Future Potential.
IEEE Communications Magazine, Feb. 2005.
- \[3\] H. Kim and Y. Han. A Proportional Fair Scheduling for Multicarrier Transmission Systems.
IEEE Communication Letters, pages 210–212, Mar. 2005.
- \[4\] G. Song and G. Li.
Utility-Based Resource Allocation and Schedulng in OFDM-Based Wireless Broadband Networks.
IEEE Communications Magazine, Dec. 2005.
- \[5\] C. Zhu and J. Agre. Proportional-Fair Scheduling Algorithms for OFDMA-based Wireless Systems.
Preprint, Fujitsu Labs, 2006.
- \[6\] T. Girici, C. Zhu, J. Agre, and A. Ephremides. Proportional Fair Scheduling Algorithm in OFDMA-based Wireless Systems with QoS Constraints.
Preprint, Fujitsu Labs, 2006.
- \[7\] M. Andrews, K. Kumaran, K. Ramanan, A. Stolyar, and P. Whiting. Providing Quality of Service over a Shared Wireless Link.
IEEE Communcations Magazine, pages 150–154, Feb. 2001. |
0711.0301 | cs.NI | cyber | # Throughput Optimal On-Line Algorithms for Advanced Resource Reservation in Ultra High-Speed Networks
Reuven Cohen<sup>1,2</sup>, Niloofar Fazlollahi<sup>1</sup> and David Starobinski<sup>1</sup>
<sup>1</sup>Dept. of Electrical and Computer Engineering
Boston University, Boston, MA 02215
<sup>2</sup>Dept. of Physics, MIT, Cambridge, MA 02139
Email: reuven@mit.edu,{nfazl,staro}@bu.edu
###### Abstract
Advanced channel reservation is emerging as an important feature of ultra high-speed networks requiring the transfer of large files. Applications include scientific data transfers and database backup. In this paper, we present two new, on-line algorithms for advanced reservation, called $`\mathtt{B}\mathtt{a}\mathtt{t}\mathtt{c}\mathtt{h}\mathtt{A}\mathtt{l}\mathtt{l}`$ and $`\mathtt{B}\mathtt{a}\mathtt{t}\mathtt{c}\mathtt{h}\mathtt{L}\mathtt{i}\mathtt{m}`$, that are guaranteed to achieve optimal throughput performance, based on multi-commodity flow arguments. Both algorithms are shown to have polynomial-time complexity and provable bounds on the maximum delay for $`1 + \varepsilon`$ bandwidth augmented networks. The $`\mathtt{B}\mathtt{a}\mathtt{t}\mathtt{c}\mathtt{h}\mathtt{L}\mathtt{i}\mathtt{m}`$ algorithm returns the completion time of a connection immediately as a request is placed, but at the expense of a slightly looser competitive ratio than that of $`\mathtt{B}\mathtt{a}\mathtt{t}\mathtt{c}\mathtt{h}\mathtt{A}\mathtt{l}\mathtt{l}`$. We also present a simple approach that limits the number of parallel paths used by the algorithms while provably bounding the maximum reduction factor in the transmission throughput. We show that, although the number of different paths can be exponentially large, the actual number of paths needed to approximate the flow is quite small and proportional to the number of edges in the network. Simulations for a number of topologies show that, in practice, 3 to 5 parallel paths are sufficient to achieve close to optimal performance. The performance of the competitive algorithms are also compared to a greedy benchmark, both through analysis and simulation.
## I Introduction
TCP/IP has been so far considered and used as the core infrastructure for all sorts of data transfer including bulk FTP applications. Yet, it has recently been observed that in ultra high speed networks, there exists a large gap between the capacity of network links and the maximum end-to-end throughput achieved by TCP \[[1](#bib.bib1), [2](#bib.bib2)\]. This gap, which is mainly attributed to the shared nature of Internet traffic, is becoming increasingly problematic for modern Grid and data backup applications requiring transfer of extremely large datasets on the orders of terabytes and more.
These limitations have raised efforts for development of an alternative protocol stack to complement TCP/IP. This protocol stack is based on the concept of *advanced reservation* \[[2](#bib.bib2), [3](#bib.bib3)\], and is specifically tailored for large file transfers and other high throughput applications. The most important property of advanced reservation is to offer hosts and users the ability to reserve in advance *dedicated* paths to connect their resources. The importance of advanced reservation in supporting dedicated connections has been made evident by the growing number of testbeds related to this technology, such as UltraScience Net \[[2](#bib.bib2)\], On-demand Secure Circuits and Advanced Reservation Systems (OSCARS) \[[4](#bib.bib4)\], and Dynamic Resource Allocation via GMPLS Optical Networks (DRAGON) \[[5](#bib.bib5)\].
Several protocols and algorithms have been proposed in the literature to support advance reservation, see, e.g., \[[2](#bib.bib2), [6](#bib.bib6), [3](#bib.bib3)\] and references therein. To the authors’ knowledge, none of them guarantees the same throughput performance as an optimal off-line algorithm. Instead, most are based on greedy approaches, whereas each request is allocated a path guaranteeing the earliest completion time at the time the request is placed.
Our first contribution in this paper is to uncover fundamental limitations of greedy algorithms. Specifically, we show that there exist network topologies and request patterns for which the maximum throughput of these algorithms can be arbitrarily smaller than the optimal throughput, as the network size grows.
Next, we present competitive advanced reservation algorithms that provably achieve the maximum throughput for any network and request profile. The first algorithm, called $`\mathtt{B}\mathtt{a}\mathtt{t}\mathtt{c}\mathtt{h}\mathtt{A}\mathtt{l}\mathtt{l}`$, provides a competitive ratio on the maximum delay experienced by each request in an augmented resource network with respect to the optimal off-line algorithm in the original network. $`\mathtt{B}\mathtt{a}\mathtt{t}\mathtt{c}\mathtt{h}\mathtt{A}\mathtt{l}\mathtt{l}`$ groups requests into batches. In each batch, paths are efficiently assigned based on a maximum concurrent flow optimization. All the requests arriving during a current batch must wait until its completion before being assigned to the next batch.
The $`\mathtt{B}\mathtt{a}\mathtt{t}\mathtt{c}\mathtt{h}\mathtt{A}\mathtt{l}\mathtt{l}`$ algorithm does not return the connection completion time to the user at the time a request is placed, but only when the connection actually starts. Our second algorithm, called $`\mathtt{B}\mathtt{a}\mathtt{t}\mathtt{c}\mathtt{h}\mathtt{L}\mathtt{i}\mathtt{m}`$, returns the completion time immediately as a request is placed, but at the expense of a slightly looser competitive ratio. This algorithm operates by limiting the length of each batch.
The presented competitive algorithms are based upon multicommodity flow algorithms, and therefore can be performed efficiently in polynomial time in the size of the network for each request.
Our model assumes that the network infrastructure supports path dispersion, i.e., multiple paths in parallel can be used to route data. Obviously, a too large path dispersion may be undesirable, as it may entail fragmenting a file into a large number of segments and reassembling them at the destination. To address this issue, we present a simple approach, based on the max-flow min-cut theorem, that limits the number of parallel paths while bounding the maximum reduction factor in the transmission throughput. We prove that this bound is tight. We, then, propose two algorithms $`\mathtt{B}\mathtt{a}\mathtt{t}\mathtt{c}\mathtt{h}\mathtt{A}\mathtt{l}\mathtt{l}\mathtt{D}\mathtt{i}\mathtt{s}\mathtt{p}`$ and $`\mathtt{B}\mathtt{a}\mathtt{t}\mathtt{c}\mathtt{h}\mathtt{L}\mathtt{i}\mathtt{m}\mathtt{D}\mathtt{i}\mathtt{s}\mathtt{p}`$, based upon $`\mathtt{B}\mathtt{a}\mathtt{t}\mathtt{c}\mathtt{h}\mathtt{A}\mathtt{l}\mathtt{l}`$ and $`\mathtt{B}\mathtt{a}\mathtt{t}\mathtt{c}\mathtt{h}\mathtt{L}\mathtt{i}\mathtt{m}`$ respectively. These algorithms perform similarly to the original algorithms, in terms of the batching process. However, after filling each batch, the algorithms will limit the dispersion of each flow. Although these algorithms are not throughput optimal anymore, they are still throughput competitive.
While the main emphasis of this paper is on the derivation of theoretical performance guarantees on the maximum throughput of advanced reservation protocols, we also provide simulation results illustrating the performance of these protocols in terms of the average delay. The simulations show that the proposed competitive algorithms compare favorably to a greedy benchmark, although the benchmark does sometime exhibit superior performance, especially in sparse topologies or at low load. With respect to path dispersion, we show that excellent performance can be achieved with as few as five or so parallel paths per connection.
Note that preliminary findings leading to this work were reported in a two-page abstract \[[7](#bib.bib7)\]. The only overlap is Lemma [IV.1](#S4.Thmtheorem1 "Lemma IV.1 ‣ IV-A Capacity Bounds ‣ IV Competitive Algorithms ‣ Throughput Optimal On-Line Algorithms for Advanced Resource Reservation in Ultra High-Speed Networks") and Theorem [IV.3](#S4.Thmtheorem3 "Theorem IV.3 ‣ IV-B The 𝙱𝚊𝚝𝚌𝚑𝙰𝚕𝚕 Competitive Algorithm ‣ IV Competitive Algorithms ‣ Throughput Optimal On-Line Algorithms for Advanced Resource Reservation in Ultra High-Speed Networks"), which were presented without proof.
This paper is organized as follows. In Section [II](#S2 "II Related Work ‣ Throughput Optimal On-Line Algorithms for Advanced Resource Reservation in Ultra High-Speed Networks"), we briefly scan the related work on advanced reservation, competitive approaches, and path dispersion. In Section [III](#S3 "III Greedy Algorithms and their Limitations ‣ Throughput Optimal On-Line Algorithms for Advanced Resource Reservation in Ultra High-Speed Networks"), we introduce our model and notation used throughout the paper. We also present natural greedy approaches and demonstrate their inefficiency. In Section [IV](#S4 "IV Competitive Algorithms ‣ Throughput Optimal On-Line Algorithms for Advanced Resource Reservation in Ultra High-Speed Networks"), we describe the $`\mathtt{B}\mathtt{a}\mathtt{t}\mathtt{c}\mathtt{h}\mathtt{A}\mathtt{l}\mathtt{l}`$ and $`\mathtt{B}\mathtt{a}\mathtt{t}\mathtt{c}\mathtt{h}\mathtt{L}\mathtt{i}\mathtt{m}`$ algorithms and derive results on their competitive ratios. Our method for bounding path dispersion is described and analyzed in Section [V](#S5 "V Bounding Path Dispersion ‣ Throughput Optimal On-Line Algorithms for Advanced Resource Reservation in Ultra High-Speed Networks"). In Section [VI](#S6 "VI Simulations ‣ Throughput Optimal On-Line Algorithms for Advanced Resource Reservation in Ultra High-Speed Networks"), simulation results evaluating the performance of some of our algorithms under different network topologies and traffic parameters are presented. We conclude the paper in Section [VII](#S7 "VII Conclusion ‣ Throughput Optimal On-Line Algorithms for Advanced Resource Reservation in Ultra High-Speed Networks").
## II Related Work
Competitive algorithms for advanced reservation networks is the focus of \[[8](#bib.bib8)\]. This work discusses the lazy ftp problem, where reservations are made for channels for the transfer of different files. The algorithm presented there provides a 4-competitive algorithm for the makespan (the total completion time). However, \[[8](#bib.bib8)\] focuses on the case of fixed routes. When routing is also to be considered, the time complexity of the algorithm presented there may be exponential in the network size.
Another recent work \[[9](#bib.bib9)\], focusing on routing in packet switched network in an adversarial setting, discusses choosing routes for fixed size packets injected by an adversary. It enforces regularity limitations on the adversary which are stronger than the ones required here, and achieves the network capacity with a guarantee on the maximum queue size. It does not discuss the case of advance reservation with different job sizes or bandwidth requirements. It is based upon approximating an integer programming, which may not be extensible to a case where path reservation, rather than packet-based routing is involved.
Most of the works regarding competitive approaches to routing focused mainly on call admission, without the ability of advance reservation. For some results in this field see, e.g., \[[10](#bib.bib10), [11](#bib.bib11)\]. Some results involving advanced reservation are presented in \[[12](#bib.bib12)\]. However, the path selection there is based on several alternatives supplied by a user in the request rather than a path selection using an automated mechanism attempting to optimize performance, as discussed here. In \[[13](#bib.bib13)\] a combination of call admission and circuit switching is used to obtain a routing scheme with a logarithmic competitive ratio on the total revenue received. A competitive routing scheme in terms of the number of failed routes in the setting of ad-hoc networks is presented in \[[14](#bib.bib14)\]. A survey of on-line routing results is presented in \[[15](#bib.bib15)\]. A competitive algorithm for admission and routing in a multicasting setting is presented in \[[16](#bib.bib16)\]. Most of the other existing work in this area consists of heuristic approaches which main emphasis are on the algorithm correctness and computational complexity, without throughput guarantees.
In \[[17](#bib.bib17)\] a rate achieving scheme for packet switching at the switch level is presented. Their scheme is based on convergence to the optimal multicommodity flow using delayed decision for queued packets. Their results somewhat resemble our $`\mathtt{B}\mathtt{a}\mathtt{t}\mathtt{c}\mathtt{h}\mathtt{A}\mathtt{l}\mathtt{l}`$ algorithm. However, their scheme depends on the existence of an average packet size, whereas our scheme addresses the full routing-scheduling question for any size distribution and any (adversarial) arrival schedule. In \[[18](#bib.bib18)\] queuing analysis of several optical transport network architectures is conducted, and it is shown that under some conditions on the arrival process, some of the schemes can achieve the maximum network rate. As the previous one, this paper does not discuss the full routing-scheduling question discussed here and does not handle unbounded job sizes. Another difference is that our paper provides an algorithm, $`\mathtt{B}\mathtt{a}\mathtt{t}\mathtt{c}\mathtt{h}\mathtt{L}\mathtt{i}\mathtt{m}`$ discussed below, guaranteeing the ending time of a job at the time of arrival, which, as far as we know is the first such algorithm.
Many papers have discussed the issue of path dispersion and attempted to achieve good throughput with limited dispersion, a survey of some results in this field is given in \[[19](#bib.bib19)\]. In \[[20](#bib.bib20), [21](#bib.bib21)\] heuristic methods of controlling multipath routing and some quantitative measures are presented. As far as we know, our work proposes the first formal treatment allowing the approximation of a flow using a limited number of paths with any desired accuracy.
## III Greedy Algorithms and their Limitations
###
III-A
Network Model
We first present the network model that will be used throughout the paper (both for the greedy and competitive algorithms). The model consists of a general network topology, represented by a graph $`G\hspace{0pt}{(V,E)}`$, where $`V`$ is the set of nodes and $`E`$ is the set of links connecting the nodes. The graph $`G`$ can be directed or undirected. The capacity of each link $`e \in E`$ is $`C\hspace{0pt}{(e)}`$. A connection request, also referred to as job, contains the tuple $`(s,d,f)`$, where $`s \in V`$ is the source node, $`d \in {V - {\{ s\}}}`$ is the destination node, and $`f`$ is the file size.
Accordingly, an advance reservation algorithm computes a starting time at which the connection can be initiated, a set of paths used for the connection, and an amount of bandwidth allocated to each path. Our model supports path dispersion, i.e., multiple paths in parallel can be used to route data (in section [](#S5.SS2 "V-B Competitive Algorithms ‣ V Bounding Path Dispersion ‣ Throughput Optimal On-Line Algorithms for Advanced Resource Reservation in Ultra High-Speed Networks")
V-B
, we discuss practical methods to bound path dispersion while still providing performance guarantees). In addition, the greedy algorithms introduced below employ path switching, whereby a connection is allowed to switch between different paths throughout its duration.
In subsequent sections, we will make frequent use of multicommodity functions. The multicommodity flow problem is a linear planning problem returning $`\mathtt{t}\mathtt{r}\mathtt{u}\mathtt{e}`$ or $`\mathtt{f}\mathtt{a}\mathtt{l}\mathtt{s}\mathtt{e}`$ based upon the feasibility of transmitting concurrent flows from all pairs of sources and destination during a given time duration, such that the total flow through each link does not exceed its capacity. It is solved by a function $`{\mathtt{m}\mathtt{u}\mathtt{l}\mathtt{t}\mathtt{i}\mathtt{c}\mathtt{o}\mathtt{m}\mathtt{m}}\hspace{0pt}{(G,L,T)}`$, where $`L`$ is a list of jobs, each containing a source, a destination, and a file size, and $`T`$ is the time duration.
In the sequel of this paper, we will be interested in characterizing the saturation throughput of the various algorithms, which is defined as the maximum arrival rate of requests such that the average delay experienced by requests is still finite (delay is defined as the time elapsing between the arrival of a connection request until the completion of the corresponding connection).
The maximum concurrent flow is calculated by the function $`{\mathtt{m}\mathtt{a}\mathtt{x}\mathtt{f}\mathtt{l}\mathtt{o}\mathtt{w}}\hspace{0pt}{(G,L)}`$. It returns $`T_{\min}`$, the minimum value of $`T`$ such that $`{\mathtt{m}\mathtt{u}\mathtt{l}\mathtt{t}\mathtt{i}\mathtt{c}\mathtt{o}\mathtt{m}\mathtt{m}}\hspace{0pt}{(G,L,T)}`$ returns $`\mathtt{t}\mathtt{r}\mathtt{u}\mathtt{e}`$. Both the multicommodity and maximum concurrent flow problems are known to be computable in polynomial time \[[22](#bib.bib22), [23](#bib.bib23)\].
###
III-B
Algorithms
#### $`\mathtt{G}\mathtt{r}\mathtt{e}\mathtt{e}\mathtt{d}\mathtt{y}`$
A seemingly natural way to implement advanced channel reservation is to follow a greedy procedure, where, at the time a request is placed, the request is allocated a path (or set of paths) guaranteeing the earliest possible completion time. We refer to this approach as $`\mathtt{G}\mathtt{r}\mathtt{e}\mathtt{e}\mathtt{d}\mathtt{y}`$ and explain it next.
The $`\mathtt{G}\mathtt{r}\mathtt{e}\mathtt{e}\mathtt{d}\mathtt{y}`$ algorithm divides the time axis into slots delineated by event. Each event corresponds to a set up or tear down instance of a connection. Therefore, during each time slot the state of all links in the network remains unchanged. In general, the time axis will consist of $`n`$ time slots, where $`n \geq 1`$ is a variable and slot $`i`$ corresponds to time interval $`\lbrack t_{i},t_{i + 1}\rbrack`$. Note that $`t_{1} = t`$ (the time at which the current request is placed) and $`t_{n + 1} = \infty`$.
Let $`W_{i} = {\{{b_{i}\hspace{0pt}{(1)}},{b_{i}\hspace{0pt}{(2)}},\ldots,{b_{i}\hspace{0pt}{({|E|})}}\}}`$ be the vector of reserved bandwidth on all links at time slot $`i`$ where $`i = {1,\ldots,n}`$, and $`b_{i}\hspace{0pt}{(e)}`$ denote the reserved bandwidth on link $`e`$ during slot $`i`$, with $`e = {1,\ldots,{|E|}}`$.
For each slot $`i \in L`$, construct a graph $`G_{i}`$, where the capacity of link $`e \in E`$ is $`{C_{i}\hspace{0pt}{(e)}} = {{C\hspace{0pt}{(e)}} - {b_{i}\hspace{0pt}{(e)}}}`$, i.e., $`C_{i}\hspace{0pt}{(e)}`$ represents the available bandwidth on link $`e`$ during slot $`i`$.
In order to guarantee the earliest completion time, $`\mathtt{G}\mathtt{r}\mathtt{e}\mathtt{e}\mathtt{d}\mathtt{y}`$ repeatedly performs a max flow allocation between nodes $`s`$ and $`d`$, for as many time slots as needed until the entire file is transferred. This approach ensures that, in each time slot, the maximum possible number of bits is transmitted and, hence, the earliest possible completion time (at the time the request is placed) is achieved. The $`\mathtt{G}\mathtt{r}\mathtt{e}\mathtt{e}\mathtt{d}\mathtt{y}`$ algorithm can thus be concisely described with the following pseudo-code:
1. 1\.
Initialization
- •
Set initial time slot: $`i`$=1.
- •
Set initial size of remaining file: $`r = f`$.
2. 2\.
If $`{{\mathtt{m}\mathtt{a}\mathtt{x}\mathtt{f}\mathtt{l}\mathtt{o}\mathtt{w}}\hspace{0pt}{(G_{i},s,d,r)}} \geq {t_{i + 1} - t_{i}}`$ (i.e., all the remaining file can be transferred during the current time slot) then
3. 3\.
Exit step:
- •
Update $`W_{i}`$ by subtracting the used bandwidth from every link in the new flow and, if needed, create a new event $`t_{i + 1}`$ corresponding to end of connection.
- •
Exit procedure.
else
4. 4\.
Non-exit step:
- •
Update $`W_{i}`$ by subtracting the used bandwidth from every link in the new flow.
- •
$`r = {r - {{{r \times {({t_{i + 1} - t_{i}})}}/{\mathtt{m}\mathtt{a}\mathtt{x}\mathtt{f}\mathtt{l}\mathtt{o}\mathtt{w}}}\hspace{0pt}{(G_{i},s,d,r)}}}`$ (update size of remaining file).
- •
$`i = {i + 1}`$ (advance to the next time slot).
- •
Go back to step 2.
#### $`{\mathtt{G}\mathtt{r}\mathtt{e}\mathtt{e}\mathtt{d}\mathtt{y}}\hspace{0pt}\_\hspace{0pt}{\mathtt{s}\mathtt{h}\mathtt{o}\mathtt{r}\mathtt{t}\mathtt{e}\mathtt{s}\mathtt{t}}`$
The $`{\mathtt{G}\mathtt{r}\mathtt{e}\mathtt{e}\mathtt{d}\mathtt{y}}\hspace{0pt}\_\hspace{0pt}{\mathtt{s}\mathtt{h}\mathtt{o}\mathtt{r}\mathtt{t}\mathtt{e}\mathtt{s}\mathtt{t}}`$ algorithm is a variation of $`\mathtt{G}\mathtt{r}\mathtt{e}\mathtt{e}\mathtt{d}\mathtt{y}`$, where only shortest paths (in terms of number of hops) between the source $`s`$ and destination $`d`$ are utilized to route data. To implement $`{\mathtt{G}\mathtt{r}\mathtt{e}\mathtt{e}\mathtt{d}\mathtt{y}}\hspace{0pt}\_\hspace{0pt}{\mathtt{s}\mathtt{h}\mathtt{o}\mathtt{r}\mathtt{t}\mathtt{e}\mathtt{s}\mathtt{t}}`$, we employ exactly the same procedure as in $`\mathtt{G}\mathtt{r}\mathtt{e}\mathtt{e}\mathtt{d}\mathtt{y}`$, except that we prune all the links not belonging to one of the shortest paths using breadth first search. Note that for a given source $`s`$ and destination $`d`$, the pruned links are the same for all the graphs $`G_{i}`$.
###
III-C
Inefficiency Results
We next constructively show that there exist certain arrival patterns, for which both $`\mathtt{G}\mathtt{r}\mathtt{e}\mathtt{e}\mathtt{d}\mathtt{y}`$ and $`{\mathtt{G}\mathtt{r}\mathtt{e}\mathtt{e}\mathtt{d}\mathtt{y}}\hspace{0pt}\_\hspace{0pt}{\mathtt{s}\mathtt{h}\mathtt{o}\mathtt{r}\mathtt{t}\mathtt{e}\mathtt{s}\mathtt{t}}`$ achieve a saturation throughput significantly lower than optimal. Specifically, we present cases where the saturation throughput of $`\mathtt{G}\mathtt{r}\mathtt{e}\mathtt{e}\mathtt{d}\mathtt{y}`$ and $`{\mathtt{G}\mathtt{r}\mathtt{e}\mathtt{e}\mathtt{d}\mathtt{y}}\hspace{0pt}\_\hspace{0pt}{\mathtt{s}\mathtt{h}\mathtt{o}\mathtt{r}\mathtt{t}\mathtt{e}\mathtt{s}\mathtt{t}}`$ is $`\Omega\hspace{0pt}{({|V|})}`$ times smaller than the optimal throughput. Thus, the ratio of the maximum throughput of the greedy algorithms to that of the optimal algorithm can be arbitrarily small.
######
Theorem III.1
For any given vertex set with cardinality $`|V|`$, there exists a graph $`G\hspace{0pt}{(V,E)}`$ such that the saturation throughput of $`\mathtt{G}\mathtt{r}\mathtt{e}\mathtt{e}\mathtt{d}\mathtt{y}`$ is $`{|V|}/2`$ times smaller than the optimal saturation throughput.
###### Proof:
Consider the ring network shown in Fig. [1](#S3.F1 "Figure 1 ‣ III-C Inefficiency Results ‣ III Greedy Algorithms and their Limitations ‣ Throughput Optimal On-Line Algorithms for Advanced Resource Reservation in Ultra High-Speed Networks")(a). Suppose every link is an undirected 1 Gb/s link and that during each second requests arrive in this order: 1Gb request from node 1 to 2, 1Gb request from node 2 to 3, 1Gb request from node 3 to 4, $`\cdots`$, 1Gb request from node $`|V|`$ to 1.
The $`\mathtt{G}\mathtt{r}\mathtt{e}\mathtt{e}\mathtt{d}\mathtt{y}`$ algorithm will allocate the maximum flow to each request, meaning it will be split to two paths, half flowing directly, and half flowing through the alternate path along the entire ring. Therefore, each new request will have to wait to the previous one to end, resulting in a completion time of $`{|V|}/2`$ seconds.
On the other hand, the optimal time is just one second using the direct link between each pair of nodes. Assuming the above pattern of requests repeats periodically, then an optimal algorithm can support at most one request per second between each pair of neighboring nodes before reaching saturation, while $`\mathtt{G}\mathtt{r}\mathtt{e}\mathtt{e}\mathtt{d}\mathtt{y}`$ can support at most $`2/{|V|}`$ request per second, hence proving the theorem. A similar result can be proved for directed graphs. $`\square`$
We next show that restricting routing to shortest paths does not solve the inefficiency problem.
######
Theorem III.2
For any given vertex set cardinality $`|V|`$, there exists a graph $`G\hspace{0pt}{(V,E)}`$ such that the saturation throughput of $`{\mathtt{G}\mathtt{r}\mathtt{e}\mathtt{e}\mathtt{d}\mathtt{y}}\hspace{0pt}\_\hspace{0pt}{\mathtt{s}\mathtt{h}\mathtt{o}\mathtt{r}\mathtt{t}\mathtt{e}\mathtt{s}\mathtt{t}}`$ (or any other algorithm using only shortest path routing) is $`{|V|} - 2`$ times smaller than the optimal saturation throughput.
###### Proof:
Consider the network depicted in Fig. [1](#S3.F1 "Figure 1 ‣ III-C Inefficiency Results ‣ III Greedy Algorithms and their Limitations ‣ Throughput Optimal On-Line Algorithms for Advanced Resource Reservation in Ultra High-Speed Networks")(b), where all requests are from node 1 to 2, and only the direct path is used by the algorithm. In this scenario, an optimal algorithm would all $`{|V|} - 2`$ paths between nodes 1 and 2. Hence, the optimal algorithm can achieve a saturation throughput $`{|V|} - 2`$ higher than $`{\mathtt{G}\mathtt{r}\mathtt{e}\mathtt{e}\mathtt{d}\mathtt{y}}\hspace{0pt}\_\hspace{0pt}{\mathtt{s}\mathtt{h}\mathtt{o}\mathtt{r}\mathtt{t}\mathtt{e}\mathtt{s}\mathtt{t}}`$. $`\square`$
[Refer to caption](/html/0711.0301/assets/x1.png)
[Refer to caption](/html/0711.0301/assets/x2.png)
Figure 1: Examples for the disadvantages of the greedy approach.
## IV Competitive Algorithms
In the previous section, we uncovered some of the limitations of greedy algorithms that immediately allocate resources to each arriving request. In this section, we present a new family of on-line, polynomial-time algorithms that rely on the idea of “batching” requests. Thus, instead of immediately reserving a path for each incoming request as in the greedy algorithms, we accumulate several arrivals in a batch and assign a more efficient set of flow paths to the whole batch. The proposed algorithms guarantee that the maximum delay experienced by each request is within a finite multiplicative factor of the value achieved with an optimal off-line algorithm. Thus, these algorithms reach the maximum throughput achievable by any algorithm.
###
IV-A
Capacity Bounds
As a preparatory step, we present a bound on the maximum throughput achievable by any algorithm. This bound, based on a multicommodity flow formulation, will be used to compare the performance of the proposed on-line competitive algorithms to the optimal off-line algorithm.
######
Lemma IV.1
If during any time interval $`T`$, each node $`i \in V`$ sends on average $`\gamma_{i\hspace{0pt}j}`$ bits of information per unit time to node $`j \in {V - {\{ i\}}}`$, then there exists a multi-commodity flow allocation on graph $`G`$ with flow values $`f_{i\hspace{0pt}j} = \gamma_{i\hspace{0pt}j}`$.
###### Proof:
First, we prove that for any given arrangement of sources and destinations in a network and at any time $`t`$, the transmission rates in bits per unit time comply with the properties of multi-commodity flows. Take the transmission rate between node $`i`$ and $`j`$ going through the (directed) edge $`k`$ at time $`t`$ to be $`R_{i,j}^{k}\hspace{0pt}{(t)}`$. The total throughput passing each link can not exceed the link capacity, $`{\sum_{i,j}{R_{i,j}^{k}\hspace{0pt}{(t)}}} \leq {C\hspace{0pt}{(k)}}`$. According to the information conservation property, the total transmission rate into any node is equal to the out-going rate except for the source and destination of any of the transmissions. Now, the average transmissions during any time interval $`T`$ satisfy the same properties as above because of linearity. Thus, we obtain a valid multi-commodity flow problem, where the time average of each flow represents an admissible commodity. $`\square`$
######
Corollary IV.2
The average transmission rate between all pairs in the network for any time interval is a feasible multicommodity flow.
###
IV-B
The $`\mathtt{B}\mathtt{a}\mathtt{t}\mathtt{c}\mathtt{h}\mathtt{A}\mathtt{l}\mathtt{l}`$ Competitive Algorithm
In case no deterministic knowledge on the future requests is given, one would like to give some bounds on the performance of the algorithm compared to the performance of a “perfect” off-line algorithm (i.e., with full knowledge of the future and unlimited computational resources). We present here an algorithm, called $`\mathtt{B}\mathtt{a}\mathtt{t}\mathtt{c}\mathtt{h}\mathtt{A}\mathtt{l}\mathtt{l}`$ (since it batches together all pending requests), giving bounds on this measure.
The algorithm can be described as follows (we assume that there is initially no pending request):
1. 1\.
For a request $`l = {\{ s,d,f\}}`$ arriving at time $`t`$, give an immediate starting time and an ending time of $`t_{c} = {t + {{\mathtt{m}\mathtt{a}\mathtt{x}\mathtt{f}\mathtt{l}\mathtt{o}\mathtt{w}}\hspace{0pt}{(G,l)}}}`$.
2. 2\.
Set $`L\leftarrow\text{null}`$.
3. 3\.
While $`t < t_{c}`$ and another request $`l' = {\{ s',d',f'\}}`$ arrives at time $`t`$:
- •
Set $`{L\leftarrow{L \cup l'}}.`$
- •
Mark $`t_{c}`$ as its starting time and add it to the waiting batch.
4. 4\.
When $`t = t_{c}`$, calculate $`t' = {{\mathtt{m}\mathtt{a}\mathtt{x}\mathtt{f}\mathtt{l}\mathtt{o}\mathtt{w}}\hspace{0pt}{(G,L)}}`$.
- •
If $`t' = 0`$ (i.e., there is no pending request) go back to step [1](#S4.I1.i1 "item 1 ‣ IV-B The 𝙱𝚊𝚝𝚌𝚑𝙰𝚕𝚕 Competitive Algorithm ‣ IV Competitive Algorithms ‣ Throughput Optimal On-Line Algorithms for Advanced Resource Reservation in Ultra High-Speed Networks").
- •
Else $`t_{c} = {t_{c} + t'}`$ and go back to step [2](#S4.I1.i2 "item 2 ‣ IV-B The 𝙱𝚊𝚝𝚌𝚑𝙰𝚕𝚕 Competitive Algorithm ‣ IV Competitive Algorithms ‣ Throughput Optimal On-Line Algorithms for Advanced Resource Reservation in Ultra High-Speed Networks").
We note that upon the arrival of a request, the $`\mathtt{B}\mathtt{a}\mathtt{t}\mathtt{c}\mathtt{h}\mathtt{A}\mathtt{l}\mathtt{l}`$ algorithm returns only the starting time of the connection. The allocated paths and completion time are computed only when the connection starts (in the next section, we present another competitive algorithm with a slightly looser competitive ratio but which returns the completion time at the arrival time of a request).
We next compare the performance of an optimal off-line algorithm in the network with that of the $`\mathtt{B}\mathtt{a}\mathtt{t}\mathtt{c}\mathtt{h}\mathtt{A}\mathtt{l}\mathtt{l}`$ algorithm in an augmented network. The augmented network is similar to the original network, other than that it has a capacity of $`{({1 + \varepsilon})}\hspace{0pt}C\hspace{0pt}{(e)}`$ at any link, $`e`$, that has capacity $`C\hspace{0pt}{(e)}`$ in the original network. This implies that the performance of the competitive algorithm is comparable if one allows a factor $`\varepsilon`$ extra capacity in the links, or, alternatively, one may say that the performance of the algorithm is comparable to the optimal off-line algorithm in some lower capacity network, allowing the maximum rate of only $`{({1 + \varepsilon})}^{- 1}\hspace{0pt}C\hspace{0pt}{(e)}`$ for each link. The $`\mathtt{B}\mathtt{a}\mathtt{t}\mathtt{c}\mathtt{h}\mathtt{A}\mathtt{l}\mathtt{l}`$ algorithm satisfies the following theoretical property on the maximum waiting time:
######
Theorem IV.3
In a network with augmented resources where every edge has bandwidth $`({1 + \varepsilon})`$ times the original bandwidth (for all $`\varepsilon > 0`$), the maximum waiting time from request arrival to the end of the transmission using $`\mathtt{B}\mathtt{a}\mathtt{t}\mathtt{c}\mathtt{h}\mathtt{A}\mathtt{l}\mathtt{l}`$, for requests arriving up to any time $`t^{\ast}`$, is no more than $`2/\varepsilon`$ times the maximum waiting time for the optimal algorithm in the original network.
###### Proof:
Consider the augmented resource network. Take the maximum length batch that accommodates requests arriving before time $`t^{\ast}`$, say $`i`$, and mark its length by $`T`$. Since the batch before this one was of length at most $`T`$ (if a request arrives when no batch is executing, the batch is considered to be of size 0), and all requests in batch $`i`$ were received during the execution of previous batch $`i - 1`$, then the total waiting time of each of these requests was at most $`2\hspace{0pt}T`$. By Corollary [IV.2](#S4.Thmtheorem2 "Corollary IV.2 ‣ IV-A Capacity Bounds ‣ IV Competitive Algorithms ‣ Throughput Optimal On-Line Algorithms for Advanced Resource Reservation in Ultra High-Speed Networks"), the total time for handling all requests received during the execution of batch $`i - 1`$ must have been at least $`T`$ in the augmented network, or $`{({1 + \varepsilon})}\hspace{0pt}T`$ in the original network. Since all of these requests arrived during a time of at most $`T`$, one of them must have been continuing at least a time of $`\varepsilon\hspace{0pt}T`$ after the last request under any algorithm. Therefore the ratio between the maximum waiting time is at most $`2/\varepsilon`$. $`\square`$
######
Corollary IV.4
The saturation throughput of $`\mathtt{B}\mathtt{a}\mathtt{t}\mathtt{c}\mathtt{h}\mathtt{A}\mathtt{l}\mathtt{l}`$ is optimal because for any arbitrarily small $`\epsilon > 0`$, the delay of a request is guaranteed to be at most a finite multiplicative factor larger than the maximum delay of the optimal algorithm in the reduced resources network.
We next show that the computational complexity of $`\mathtt{B}\mathtt{a}\mathtt{t}\mathtt{c}\mathtt{h}\mathtt{A}\mathtt{l}\mathtt{l}`$ is practical.
######
Theorem IV.5
The computational complexity of algorithm $`\mathtt{B}\mathtt{a}\mathtt{t}\mathtt{c}\mathtt{h}\mathtt{A}\mathtt{l}\mathtt{l}`$ per request is at most polynomial in the size of the network.
###### Proof:
For each batch, the $`\mathtt{B}\mathtt{a}\mathtt{t}\mathtt{c}\mathtt{h}\mathtt{A}\mathtt{l}\mathtt{l}`$ algorithm solves a maximum concurrent flow problem that can be computed in polynomial time \[[22](#bib.bib22)\]. Noting that the number of batches cannot exceed the number of requests (since each batch contains at least one request), the theorem is proven. $`\square`$
###
IV-C
The $`\mathtt{B}\mathtt{a}\mathtt{t}\mathtt{c}\mathtt{h}\mathtt{L}\mathtt{i}\mathtt{m}`$ Competitive Algorithm
The following algorithm, called $`\mathtt{B}\mathtt{a}\mathtt{t}\mathtt{c}\mathtt{h}\mathtt{L}\mathtt{i}\mathtt{m}`$, operates in a similar setup to $`\mathtt{B}\mathtt{a}\mathtt{t}\mathtt{c}\mathtt{h}\mathtt{A}\mathtt{l}\mathtt{l}`$. However, it gives a guarantee on the finishing time of a job when it is submitted.
The algorithm is based on creating increasing size batches in case of high loads. When the rate of request arrival is high, each batch created will be approximately twice as long as the previous one, and thus larger batches, giving a good approximation of the total flow will be created. When the load is low only a single batch of pending requests exists at most times, and its size will remain approximately constant or even decrease for decreasing load.
The algorithm maintains a list of times, $`t_{i}`$, $`i = {1,\ldots,n}`$, where for every interval, $`\lbrack t_{i},t_{i + 1}\rbrack`$, a batch of jobs is assigned. When a new request between a source $`s`$ and a destination $`d`$ arrives at time $`t`$, then an attempt is first made to add it to one of the existing intervals by using a multicommodity flow calculation. If the attempt fails, a new time, then a new batch is created and $`t_{n + 1} = {\max{({t_{n} + {({t_{n} - t})}},{t_{n} + M})}}`$, where $`M = {{f/{\mathtt{m}\mathtt{a}\mathtt{x}\mathtt{f}\mathtt{l}\mathtt{o}\mathtt{w}}}\hspace{0pt}{(G,s,d)}}`$ is the minimum time for the job completion, is added to the list, and the job is assigned to the interval $`\lbrack t_{n},t_{n + 1}\rbrack`$.
We next provide a detailed description of how the algorithm handles a new request. We use the tuple $`l = {\{ s,f,d\}}`$ to denote this job and the list $`L_{i}`$ to denote the set of jobs already assigned to slot $`i`$.
1. 1\.
Initialization: set $`i\leftarrow 1`$.
2. 2\.
While $`i \leq n`$
- •
If $`{{\mathtt{m}\mathtt{u}\mathtt{l}\mathtt{t}\mathtt{i}\mathtt{c}\mathtt{o}\mathtt{m}\mathtt{m}}\hspace{0pt}{(G,{L_{i} \cup l},{t_{i + 1} - t_{i}})}} = {\mathtt{t}\mathtt{r}\mathtt{u}\mathtt{e}}`$ then return $`i`$ and exit (check if request can be accommodated during slot $`i`$).
- •
$`i\leftarrow{i + 1}`$.
3. 3\.
Set $`n\leftarrow{n + 1}`$ (create a new slot).
4. 4\.
Set $`M = {{\mathtt{m}\mathtt{a}\mathtt{x}\mathtt{f}\mathtt{l}\mathtt{o}\mathtt{w}}\hspace{0pt}{(G,l)}}`$.
5. 5\.
If $`M > {t_{n - 1} - t}`$
then $`t_{n + 1}\leftarrow{t_{n} + M}`$;
else $`t_{n + 1}\leftarrow{t_{n} + {({t_{n} - t})}}`$.
6. 6\.
Return $`n`$.
Fig. [2](#S4.F2 "Figure 2 ‣ IV-C The 𝙱𝚊𝚝𝚌𝚑𝙻𝚒𝚖 Competitive Algorithm ‣ IV Competitive Algorithms ‣ Throughput Optimal On-Line Algorithms for Advanced Resource Reservation in Ultra High-Speed Networks") illustrates runs of the $`\mathtt{B}\mathtt{a}\mathtt{t}\mathtt{c}\mathtt{h}\mathtt{A}\mathtt{l}\mathtt{l}`$ and $`\mathtt{B}\mathtt{a}\mathtt{t}\mathtt{c}\mathtt{h}\mathtt{L}\mathtt{i}\mathtt{m}`$ algorithms for the same set of requests.
The advantage of the $`\mathtt{B}\mathtt{a}\mathtt{t}\mathtt{c}\mathtt{h}\mathtt{L}\mathtt{i}\mathtt{m}`$ algorithm over $`\mathtt{B}\mathtt{a}\mathtt{t}\mathtt{c}\mathtt{h}\mathtt{A}\mathtt{l}\mathtt{l}`$ is the assignment of a completion time to a request upon its arrival. Furthermore, the proof of the algorithm requires only an attempt to assign a new job to the last interval, rather than to all of them. Therefore, it is also possible to give the details of the selected paths to the source as soon as a new batch is created.
We introduce the following technical Lemma which will be used to prove the main theorem.
######
Lemma IV.6
\(a\) For any $`i`$, $`{t_{i} - t_{i - 1}} \geq {{({t_{i} - t})}/2}`$. (b) For any $`i`$, $`{t_{i + 1} - t_{i}} \leq {\max{({2\hspace{0pt}{({t_{i} - t_{i - 1}})}},M)}}`$, where $`M`$ is the minimum completion time for the largest job in the interval $`\lbrack t_{i + 1},t_{i}\rbrack`$.
###### Proof:
\(a\) When the time slot $`\lbrack t_{i - 1},t_{i}\rbrack`$ is added, its length is the maximum of $`M`$ and $`\lbrack t,t_{i - 1}\rbrack`$. As $`t`$ increases $`{t_{i} - t_{i - 1}} \geq {t_{i - 1} - t}`$, and therefore $`{2\hspace{0pt}{({t_{i} - t_{i - 1}})}} \geq {{({t_{i} - t_{i - 1}})} + {({t_{i - 1} - t})}} = {t_{i} - t}`$.
\(b\) From (a) follows that at the time $`t`$ when $`\lbrack t_{i + 1},t_{i}\rbrack`$ is added $`{t_{i} - t_{i - 1}} \geq {{({t_{i} - t})}/2}`$. Since $`{t_{i + 1} - t_{i}} = {\max{({t_{i} - t},M)}}`$ the Lemma follows. $`\square`$
######
Theorem IV.7
For every $`\varepsilon > 0`$ and every time $`t^{\ast}`$, the maximum waiting time from the request time to the end of a job for any request arriving up to time $`t^{\ast}`$ and for a network with augmented resources, is no more than $`4/\varepsilon`$ times the maximum waiting time for the optimal algorithm in the original network.
###### Proof:
When creating a new batch, at the interval $`\lbrack t_{n},t_{n + 1}\rbrack`$ there are two possibilities:
1. 1\.
$`t_{n + 1} = {t_{n} + M}`$, where $`M`$ is the minimum time for the new job completion. In this case, the minimum time for the completion of the job in the original network would have been $`{({1 + \varepsilon})}\hspace{0pt}M`$. By Lemma [IV.6](#S4.Thmtheorem6 "Lemma IV.6 ‣ IV-C The 𝙱𝚊𝚝𝚌𝚑𝙻𝚒𝚖 Competitive Algorithm ‣ IV Competitive Algorithms ‣ Throughput Optimal On-Line Algorithms for Advanced Resource Reservation in Ultra High-Speed Networks")(a) the total waiting time for the job here is at most $`2\hspace{0pt}M`$, and therefore the ratio of waiting times is $`{2/{({1 + \varepsilon})}} < {4/\varepsilon}`$.
2. 2\.
$`t_{n + 1} = {t_{n} + {({t_{n} - t})}}`$, in this case, there exists a request among the set of tasks in batch $`n`$ that according to theorem [IV.3](#S4.Thmtheorem3 "Theorem IV.3 ‣ IV-B The 𝙱𝚊𝚝𝚌𝚑𝙰𝚕𝚕 Competitive Algorithm ‣ IV Competitive Algorithms ‣ Throughput Optimal On-Line Algorithms for Advanced Resource Reservation in Ultra High-Speed Networks") waits for a time of at least $`({t_{n} - t_{n - 1}})`$. Therefore, the maximum waiting time in the original network is at least $`\varepsilon\hspace{0pt}{({t_{n} - t_{n - 1}})}`$. Since $`{t_{n} - t_{n - 1}} \geq {{({t_{n} - t})}/2}`$, and $`t_{n + 1} = {{t_{n} + t_{n}} - t}`$, it follows that $`{t_{n + 1} - t} \leq {4\hspace{0pt}{({t_{n} - t_{n - 1}})}}`$. Therefore, the ratio is at most $`4/\varepsilon`$.
$`\square`$
[Refer to caption](/html/0711.0301/assets/x3.png)
Figure 2: Illustration of the batching process by the algorithms $`\mathtt{B}\mathtt{a}\mathtt{t}\mathtt{c}\mathtt{h}\mathtt{A}\mathtt{l}\mathtt{l}`$ and $`\mathtt{B}\mathtt{a}\mathtt{t}\mathtt{c}\mathtt{h}\mathtt{L}\mathtt{i}\mathtt{m}`$. The odd numbered jobs request one hour of maximum bandwidth traffic between nodes $`A`$ and $`B`$ and the even numbered jobs request one hour maximum bandwidth traffic between $`B`$ and $`C`$. In $`\mathtt{B}\mathtt{a}\mathtt{t}\mathtt{c}\mathtt{h}\mathtt{A}\mathtt{l}\mathtt{l}`$, a new batch is created for all requests arriving during the running of a previous batch. In $`\mathtt{B}\mathtt{a}\mathtt{t}\mathtt{c}\mathtt{h}\mathtt{L}\mathtt{i}\mathtt{m}`$, for each new request an attempt is made to add it to one of the existing windows, and if it fails, a new window is appended at the end.
## V Bounding Path Dispersion
###
V-A
Approximating flows by single circuits
The algorithms presented in the previous sections (both greedy and competitive), do not limit the path dispersion, i.e., the number of number of paths simultaneously used by a connection. In practice, it is desirable to minimize the number of such paths due to the cost of setting up many paths and the need to split a into many low capacity channels. The following suggests a method of achieving this goal.
######
Lemma V.1
In every flow of size $`F`$ between two nodes on a directed graph $`G\hspace{0pt}{(V,E)}`$ there exists a path of capacity at least $`F/{|E|}`$ between these nodes.
###### Proof:
Remove all edges of capacity (strictly) less than $`F/{|E|}`$ from the graph. The total capacity of these edges is smaller than $`{{F/{|E|}} \times {|E|}} = F`$. By the Max-Flow–Min-Cut theorem, the maximum flow equals the minimum cut in the network. Therefore, since the total flow is $`F`$ there must remain at least one path between the nodes after the removal. All edges in this path have capacity of at least $`F/{|E|}`$. Therefore, the path capacity is at least $`F/{|E|}`$. $`\square`$
Lemma [V.1](#S5.Thmtheorem1 "Lemma V.1 ‣ V-A Approximating flows by single circuits ‣ V Bounding Path Dispersion ‣ Throughput Optimal On-Line Algorithms for Advanced Resource Reservation in Ultra High-Speed Networks") is tight in order in both $`|E|`$ and $`|V|`$ as can be seen for the network in Fig. [3](#S5.F3 "Figure 3 ‣ V-A Approximating flows by single circuits ‣ V Bounding Path Dispersion ‣ Throughput Optimal On-Line Algorithms for Advanced Resource Reservation in Ultra High-Speed Networks"). In this network each path amounts for only $`{O\hspace{0pt}{({|E|}^{- 1})}} = {O\hspace{0pt}{({|V|}^{- 2})}}`$ of the total flow between nodes $`1`$ and $`2`$. Therefore, to obtain a flow that consists of a given fraction of the maximum flow at least $`{O\hspace{0pt}{({|E|})}} = {O\hspace{0pt}{({|V|}^{2})}}`$ of paths must be used.
[Refer to caption](/html/0711.0301/assets/x4.png)
Figure 3: A network demonstrating the optimality of Lemma [V.1](#S5.Thmtheorem1 "Lemma V.1 ‣ V-A Approximating flows by single circuits ‣ V Bounding Path Dispersion ‣ Throughput Optimal On-Line Algorithms for Advanced Resource Reservation in Ultra High-Speed Networks").
The following theorem establishes the maximum number of paths needed to achieve a throughput with a constant factor of that achieved by the original flow.
######
Theorem V.2
For every flow of size $`F`$ between two nodes on a directed graph $`G\hspace{0pt}{(V,E)}`$ and for every $`\alpha > 0`$ there exists a set of at most $`\lceil{\alpha\hspace{0pt}{|E|}}\rceil`$ paths achieving a flow of at least $`{({1 - e^{- \alpha}})}\hspace{0pt}F`$.
###### Proof:
Apply Lemma [V.1](#S5.Thmtheorem1 "Lemma V.1 ‣ V-A Approximating flows by single circuits ‣ V Bounding Path Dispersion ‣ Throughput Optimal On-Line Algorithms for Advanced Resource Reservation in Ultra High-Speed Networks"), $`\lceil{\alpha\hspace{0pt}{|E|}}\rceil`$ times. Each time an appropriate path is found, its flow is removed from all of its links. For each such path the remaining flow is multiplied by at most $`{1 - \frac{1}{|E|}} \leq {\exp\left( \frac{1}{|E|} \right)}`$. $`\square`$
###
V-B
Competitive Algorithms
Theorem [V.2](#S5.Thmtheorem2 "Theorem V.2 ‣ V-A Approximating flows by single circuits ‣ V Bounding Path Dispersion ‣ Throughput Optimal On-Line Algorithms for Advanced Resource Reservation in Ultra High-Speed Networks") provides both an algorithm for reducing the number of paths and a bound on the throughput loss. To approximate the flow by a limited number of paths remove all edges with capacity less than $`F/{|E|}`$ and find a remaining path. This process can be repeated to obtain improving approximations of the flow. The maximum number of paths needed to achieve an approximation of the flow to within a constant factor is linear in $`|E|`$, while the number of possible paths may be exponential in $`|E|`$.
Using the above approximation in conjunction with the competitive algorithms one can devise two new algorithms $`\mathtt{B}\mathtt{a}\mathtt{t}\mathtt{c}\mathtt{h}\mathtt{A}\mathtt{l}\mathtt{l}\mathtt{D}\mathtt{i}\mathtt{s}\mathtt{p}`$ and $`\mathtt{B}\mathtt{a}\mathtt{t}\mathtt{c}\mathtt{h}\mathtt{L}\mathtt{i}\mathtt{m}\mathtt{D}\mathtt{i}\mathtt{s}\mathtt{p}`$, based upon $`\mathtt{B}\mathtt{a}\mathtt{t}\mathtt{c}\mathtt{h}\mathtt{A}\mathtt{l}\mathtt{l}`$ and $`\mathtt{B}\mathtt{a}\mathtt{t}\mathtt{c}\mathtt{h}\mathtt{L}\mathtt{i}\mathtt{m}`$ respectively. These algorithms perform similarly to $`\mathtt{B}\mathtt{a}\mathtt{t}\mathtt{c}\mathtt{h}\mathtt{A}\mathtt{l}\mathtt{l}`$ and $`\mathtt{B}\mathtt{a}\mathtt{t}\mathtt{c}\mathtt{h}\mathtt{L}\mathtt{i}\mathtt{m}`$, in terms of the batching process. However, after filling each batch, the algorithms will limit the dispersion of each flow, and approximate the flow. To achieve a partial flow, each time we select the widest path from remaining edges (where the weights are determined by the link utilization in the solution of the multicommodity flow) and reserve the total path bandwidth. We repeat this until either the desired number of paths is reached or the entire flow is routed. Setting the dispersion bound to $`\alpha\hspace{0pt}{|E|}`$, Theorem [V.2](#S5.Thmtheorem2 "Theorem V.2 ‣ V-A Approximating flows by single circuits ‣ V Bounding Path Dispersion ‣ Throughput Optimal On-Line Algorithms for Advanced Resource Reservation in Ultra High-Speed Networks") guarantees that the time allocated for the slot will be no more than $`{{\mathtt{m}\mathtt{a}\mathtt{x}\mathtt{f}\mathtt{l}\mathtt{o}\mathtt{w}}\hspace{0pt}{(G,L)}}/{({1 - e^{- \alpha}})}`$, where $`{\mathtt{m}\mathtt{a}\mathtt{x}\mathtt{f}\mathtt{l}\mathtt{o}\mathtt{w}}\hspace{0pt}{(G,L)}`$ is the original slot duration. Algorithms $`\mathtt{B}\mathtt{a}\mathtt{t}\mathtt{c}\mathtt{h}\mathtt{A}\mathtt{l}\mathtt{l}\mathtt{D}\mathtt{i}\mathtt{s}\mathtt{p}`$ and $`\mathtt{B}\mathtt{a}\mathtt{t}\mathtt{c}\mathtt{h}\mathtt{L}\mathtt{i}\mathtt{m}\mathtt{D}\mathtt{i}\mathtt{s}\mathtt{p}`$ will achieve a throughput within a factor of $`1 - e^{- \alpha}`$ of the maximum throughput (and will no longer be throughput optimal as $`\mathtt{B}\mathtt{a}\mathtt{t}\mathtt{c}\mathtt{h}\mathtt{A}\mathtt{l}\mathtt{l}`$ and $`\mathtt{B}\mathtt{a}\mathtt{t}\mathtt{c}\mathtt{h}\mathtt{L}\mathtt{i}\mathtt{m}`$, but only throughput competitive).
## VI Simulations
In this section, we present simulation results illustrating the performance of the algorithms described in this paper. While the emphasis of the previous sections was on the throughput optimality (or competitiveness) of our proposed algorithms, here we are also interested in evaluating their performance for other metrics, especially average delay. Average delay is defined as the average time elapsing from the point a request is placed until the corresponding connection is completed.
The main points of interest are as follows: (1) how do the competitive algorithms fare compared to the greedy approach of Section [III](#S3 "III Greedy Algorithms and their Limitations ‣ Throughput Optimal On-Line Algorithms for Advanced Resource Reservation in Ultra High-Speed Networks")? (2) what value of path dispersion is needed to ensure good performance? It is important to emphasize that we do not expect the throughput optimal competitive algorithms to always outperform the greedy approach in terms of average delay.
###
VI-A
Simulation Set-Up
We have written our own in simulator in C++. The simulator uses the COIN-OR Linear Program Solver (CLP) library \[[24](#bib.bib24)\] to solve multi-commodity optimization problems and allows evaluating our algorithms under various topological settings and traffic conditions. The main simulation parameters are as follows:
[Refer to caption](/html/0711.0301/assets/x5.png)
\(a\)
8 node
[Refer to caption](/html/0711.0301/assets/x6.png)
\(b\)
11 node LambdaRail
Figure 4: Simulation topologies.
- •
*Topology:* our simulator supports arbitrary topologies. In this paper, we consider the two topologies depicted in Figure [4](#S6.F4 "Figure 4 ‣ VI-A Simulation Set-Up ‣ VI Simulations ‣ Throughput Optimal On-Line Algorithms for Advanced Resource Reservation in Ultra High-Speed Networks"). One is a fully connected graph (clique) of eight nodes and the other is an 11-node topology, similar to the National LambdaRail testbed \[[25](#bib.bib25)\]. Each link on these graphs is full-duplex and assumed to have a capacity of 20 Gb/s.
- •
*Arrival process:* we assume that the aggregated arrival of requests to the network forms a Poisson process (this can easily be changed, if desired). The mean rate of arrivals is adjustable. Our delay measurements are carried out at different arrival rates, referred to as *network load*, in units of requests per hour.
- •
*File size distribution:*
We consider two models for the file size distribution:
1. 1\.
Pareto:
<table>
<tbody>
<tr>
<td></td>
<td>F(x)=1−(xmx−γ)β, where x≥xm+γ.𝐹𝑥1superscriptsubscript𝑥𝑚𝑥𝛾𝛽, where 𝑥subscript𝑥𝑚𝛾F(x)=1-\left(\frac{x_{m}}{x-\gamma}\right)^{\beta}\mbox{, where }x\geq x_{m}+\gamma.</td>
<td></td>
</tr>
</tbody>
</table>
In the simulations, we set $`\beta = 2.5`$, $`x_{m} = 1.48`$ TB (terabyte) and $`\gamma = {6.25 \ast 10^{- 3}}`$ TB, implying that the mean file size is $`2.475`$ TB.
2. 2\.
Exponential:
<table>
<tbody>
<tr>
<td></td>
<td>F(x)=exp(−λx), where x≥0.𝐹𝑥𝜆𝑥, where 𝑥0F(x)=\exp(-\lambda x)\mbox{, where }x\geq 0.</td>
<td></td>
</tr>
</tbody>
</table>
In the simulations, $`{1/\lambda} = 2.475`$ TB.
- •
*Source and Destination*: for each request, the source and destination are selected uniformly at random, except that they must be different nodes.
All the simulations are run for a total of at least $`10^{6}`$ requests.
###
VI-B
Results
We first present simulation results for the clique topology. Figure [5](#S6.F5 "Figure 5 ‣ VI-B Results ‣ VI Simulations ‣ Throughput Optimal On-Line Algorithms for Advanced Resource Reservation in Ultra High-Speed Networks") depicts the average delay as a function of the network load for the algorithms $`\mathtt{B}\mathtt{a}\mathtt{t}\mathtt{c}\mathtt{h}\mathtt{A}\mathtt{l}\mathtt{l}`$ and $`\mathtt{G}\mathtt{r}\mathtt{e}\mathtt{e}\mathtt{d}\mathtt{y}`$, assuming a Pareto file size distribution. The figure indicates that $`\mathtt{G}\mathtt{r}\mathtt{e}\mathtt{e}\mathtt{d}\mathtt{y}`$’s average delay increases sharply around 140 requests/hour, while $`\mathtt{B}\mathtt{a}\mathtt{t}\mathtt{c}\mathtt{h}\mathtt{A}\mathtt{l}\mathtt{l}`$ can sustain higher network load since it allocates flows more efficiently. On the other hand, at lower load, $`\mathtt{G}\mathtt{r}\mathtt{e}\mathtt{e}\mathtt{d}\mathtt{y}`$ achieves a lower average delay than $`\mathtt{B}\mathtt{a}\mathtt{t}\mathtt{c}\mathtt{h}\mathtt{A}\mathtt{l}\mathtt{l}`$. This result can be explained by the fact that $`\mathtt{G}\mathtt{r}\mathtt{e}\mathtt{e}\mathtt{d}\mathtt{y}`$ does not wait to establish a connection. This results illustrates the existence of a delay-throughput trade-off between the two algorithms.
[Refer to caption](/html/0711.0301/assets/x7.png)
Figure 5: Performance comparison of algorithms $`\mathtt{B}\mathtt{a}\mathtt{t}\mathtt{c}\mathtt{h}\mathtt{A}\mathtt{l}\mathtt{l}`$ and $`\mathtt{G}\mathtt{r}\mathtt{e}\mathtt{e}\mathtt{d}\mathtt{y}`$ for the 8-node clique and Pareto file size distribution.
Figure [6](#S6.F6 "Figure 6 ‣ VI-B Results ‣ VI Simulations ‣ Throughput Optimal On-Line Algorithms for Advanced Resource Reservation in Ultra High-Speed Networks") compares the performance of the $`\mathtt{B}\mathtt{a}\mathtt{t}\mathtt{c}\mathtt{h}\mathtt{A}\mathtt{l}\mathtt{l}`$ and $`\mathtt{B}\mathtt{a}\mathtt{t}\mathtt{c}\mathtt{h}\mathtt{L}\mathtt{i}\mathtt{m}`$ algorithms. The file size distribution is exponential. The figure shows that $`\mathtt{B}\mathtt{a}\mathtt{t}\mathtt{c}\mathtt{h}\mathtt{L}\mathtt{i}\mathtt{m}`$ is less efficient than $`\mathtt{B}\mathtt{a}\mathtt{t}\mathtt{c}\mathtt{h}\mathtt{A}\mathtt{l}\mathtt{l}`$ in terms of average delay, especially at low load. This result is somewhat expected given that $`\mathtt{B}\mathtt{a}\mathtt{t}\mathtt{c}\mathtt{h}\mathtt{L}\mathtt{i}\mathtt{m}`$ uses a less efficient batching process and its delay ratio is looser. However, since $`\mathtt{B}\mathtt{a}\mathtt{t}\mathtt{c}\mathtt{h}\mathtt{L}\mathtt{i}\mathtt{m}`$ is throughput optimal, its performance approaches that of $`\mathtt{B}\mathtt{a}\mathtt{t}\mathtt{c}\mathtt{h}\mathtt{A}\mathtt{l}\mathtt{l}`$ at high load.
[Refer to caption](/html/0711.0301/assets/x8.png)
Figure 6: Performance comparison of algorithms $`\mathtt{B}\mathtt{a}\mathtt{t}\mathtt{c}\mathtt{h}\mathtt{A}\mathtt{l}\mathtt{l}`$ and $`\mathtt{B}\mathtt{a}\mathtt{t}\mathtt{c}\mathtt{h}\mathtt{L}\mathtt{i}\mathtt{m}`$ for the 8-node clique and exponential file size distribution.
Figure [7](#S6.F7 "Figure 7 ‣ VI-B Results ‣ VI Simulations ‣ Throughput Optimal On-Line Algorithms for Advanced Resource Reservation in Ultra High-Speed Networks") evaluates and compares the performance of the $`\mathtt{B}\mathtt{a}\mathtt{t}\mathtt{c}\mathtt{h}\mathtt{A}\mathtt{l}\mathtt{l}`$ and $`\mathtt{B}\mathtt{a}\mathtt{t}\mathtt{c}\mathtt{h}\mathtt{A}\mathtt{l}\mathtt{l}\mathtt{D}\mathtt{i}\mathtt{s}\mathtt{p}`$ algorithms. For $`\mathtt{B}\mathtt{a}\mathtt{t}\mathtt{c}\mathtt{h}\mathtt{A}\mathtt{l}\mathtt{l}\mathtt{D}\mathtt{i}\mathtt{s}\mathtt{p}`$, results are presented for the cases where the path dispersion bound is set to either 1, 3, or 5 paths per connection. The file size distribution is exponential. The figure shows also a fluid bound on capacity derived in \[[26](#bib.bib26)\]. Its value represents an upper bound on the maximum network load for which the average delay of requests is still bounded.
[Refer to caption](/html/0711.0301/assets/x9.png)
Figure 7: Performance evaluation of algorithms $`\mathtt{B}\mathtt{a}\mathtt{t}\mathtt{c}\mathtt{h}\mathtt{A}\mathtt{l}\mathtt{l}`$ and $`\mathtt{B}\mathtt{a}\mathtt{t}\mathtt{c}\mathtt{h}\mathtt{A}\mathtt{l}\mathtt{l}\mathtt{D}\mathtt{i}\mathtt{s}\mathtt{p}`$ with bounds 1, 3 or 5 paths per connection for the 8-node clique and exponential file size distribution. A fluid bound on capacity is also depicted.
The figure shows that $`\mathtt{B}\mathtt{a}\mathtt{t}\mathtt{c}\mathtt{h}\mathtt{A}\mathtt{l}\mathtt{l}`$ approaches the capacity bound at a reasonable delay value and that a path dispersion of $`5`$ per connection (corresponding to $`\alpha = 0.089`$) is sufficient for $`\mathtt{B}\mathtt{a}\mathtt{t}\mathtt{c}\mathtt{h}\mathtt{A}\mathtt{l}\mathtt{l}\mathtt{D}\mathtt{i}\mathtt{s}\mathtt{p}`$ to achieve performance close to $`\mathtt{B}\mathtt{a}\mathtt{t}\mathtt{c}\mathtt{h}\mathtt{A}\mathtt{l}\mathtt{l}`$. It is worth mentioning that, in this topology, there exist $`1957`$ possible paths between any two nodes. Thus, with 5 paths, $`\mathtt{B}\mathtt{a}\mathtt{t}\mathtt{c}\mathtt{h}\mathtt{A}\mathtt{l}\mathtt{l}\mathtt{D}\mathtt{i}\mathtt{s}\mathtt{p}`$ uses only $`0.25\%`$ of the total paths possible. The figure also demonstrates the importance of multi-path routing: the performance achieved using a single path per connections is far worse.
[Refer to caption](/html/0711.0301/assets/x10.png)
Figure 8: Performance evaluation of algorithms $`\mathtt{B}\mathtt{a}\mathtt{t}\mathtt{c}\mathtt{h}\mathtt{A}\mathtt{l}\mathtt{l}`$, $`\mathtt{B}\mathtt{a}\mathtt{t}\mathtt{c}\mathtt{h}\mathtt{A}\mathtt{l}\mathtt{l}\mathtt{D}\mathtt{i}\mathtt{s}\mathtt{p}`$ and $`{\mathtt{G}\mathtt{r}\mathtt{e}\mathtt{e}\mathtt{d}\mathtt{y}}\hspace{0pt}\_\hspace{0pt}{\mathtt{s}\mathtt{h}\mathtt{o}\mathtt{r}\mathtt{t}\mathtt{e}\mathtt{s}\mathtt{t}}`$ for the 11-node topology [4(b)](#S6.F4.sf2 "In Figure 4 ‣ VI-A Simulation Set-Up ‣ VI Simulations ‣ Throughput Optimal On-Line Algorithms for Advanced Resource Reservation in Ultra High-Speed Networks") and Pareto file size. Algorithm $`\mathtt{B}\mathtt{a}\mathtt{t}\mathtt{c}\mathtt{h}\mathtt{A}\mathtt{l}\mathtt{l}\mathtt{D}\mathtt{i}\mathtt{s}\mathtt{p}`$ is plotted with bounds of 3, 2 and 1 on path dispersion.
Figure [8](#S6.F8 "Figure 8 ‣ VI-B Results ‣ VI Simulations ‣ Throughput Optimal On-Line Algorithms for Advanced Resource Reservation in Ultra High-Speed Networks") depicts the performance of the various algorithms and the fluid bound for the 11-node topology of figure [4(b)](#S6.F4.sf2 "In Figure 4 ‣ VI-A Simulation Set-Up ‣ VI Simulations ‣ Throughput Optimal On-Line Algorithms for Advanced Resource Reservation in Ultra High-Speed Networks"). The file size follows a Pareto distribution. In this case, we observe that $`\mathtt{B}\mathtt{a}\mathtt{t}\mathtt{c}\mathtt{h}\mathtt{A}\mathtt{l}\mathtt{l}\mathtt{D}\mathtt{i}\mathtt{s}\mathtt{p}`$ with as few as 3 paths per connection (or $`\alpha = 0.107`$) approximates $`\mathtt{B}\mathtt{a}\mathtt{t}\mathtt{c}\mathtt{h}\mathtt{A}\mathtt{l}\mathtt{l}`$ very closely. Since this network is sparser than the previous one, it is reasonable to obtain good performance with a smaller path dispersion. Another observation is that for this topology and the range of network loads under consideration, $`\mathtt{G}\mathtt{r}\mathtt{e}\mathtt{e}\mathtt{d}\mathtt{y}`$ outperforms $`\mathtt{B}\mathtt{a}\mathtt{t}\mathtt{c}\mathtt{h}\mathtt{A}\mathtt{l}\mathtt{l}`$ in terms of average delay. Although $`\mathtt{B}\mathtt{a}\mathtt{t}\mathtt{c}\mathtt{h}\mathtt{A}\mathtt{l}\mathtt{l}`$ achieves the optimal saturation throughput, its superiority over $`\mathtt{G}\mathtt{r}\mathtt{e}\mathtt{e}\mathtt{d}\mathtt{y}`$ may only happen at very large delays. The advantage of $`\mathtt{G}\mathtt{r}\mathtt{e}\mathtt{e}\mathtt{d}\mathtt{y}`$ in this scenario can be explained by the sparsity of the underlying topology. Seemingly, the added efficiency in throughput achieved with batching is not sufficient to compensate the cost incurred by delaying requests.
## VII Conclusion
Advance reservation of dedicated network resources is emerging as one of the most important features of the next-generation of network architectures. In this paper, we have made several advances in this area. Specifically, we have proposed two new on-line algorithms for advanced reservation, called $`\mathtt{B}\mathtt{a}\mathtt{t}\mathtt{c}\mathtt{h}\mathtt{A}\mathtt{l}\mathtt{l}`$ and $`\mathtt{B}\mathtt{a}\mathtt{t}\mathtt{c}\mathtt{h}\mathtt{L}\mathtt{i}\mathtt{m}`$, that are guaranteed to achieve optimal throughput performance. We have proven that both algorithms achieve a competitive ratio on the maximum delay until job completion in $`1 + \varepsilon`$ bandwidth augmented networks. While $`\mathtt{B}\mathtt{a}\mathtt{t}\mathtt{c}\mathtt{h}\mathtt{L}\mathtt{i}\mathtt{m}`$ has a slightly looser competitive ratio than that of $`\mathtt{B}\mathtt{a}\mathtt{t}\mathtt{c}\mathtt{h}\mathtt{A}\mathtt{l}\mathtt{l}`$ (i.e., $`4/\epsilon`$ instead of $`2/\epsilon`$), it has the distinct advantage of computing the completion time of a connection immediately as a request is placed.
A key observation of this paper is that path dispersion is essential to achieve full network utilization. However, splitting a transmission into too many different paths may render a flow-based approach inapplicable in many real-world environments. Thus, we have presented a rigorous, theoretical approach to address the path dispersion problem and presented a method for approximating the maximum multicommodity flow using a limited number of paths. Specifically, while the number of paths between two nodes in a network scales exponentially with the number of edges, we have shown that throughput competitiveness up to any desired ratio factor can be achieved with a number of paths scaling linearly with the total number of edges. In practice, our simulations indicate that 3 paths (in sparse graphs) to 5 paths (in dense graphs) may be sufficient.
As shown in our simulations, for some topologies, a greedy approach may perform better than the competitive algorithms, especially at low load, in sparse topologies, or with uncorrelated traffic between different pairs. However, we have shown that there exist scenarios where greedy strategies can be highly inefficient. This can never be the case for the competitive algorithms.
We conclude by noting that the algorithms proposed in this paper can be either run in a centralized fashion (a reasonable solution in small networks) or using link-state routing and distributed signaling mechanism, such as enhanced versions of GMPLS \[[5](#bib.bib5)\] or RSVP \[[27](#bib.bib27)\]. Distributed approximations of the multicommodity flow have also been discussed extensively in the literature (see, e.g., \[[28](#bib.bib28), [29](#bib.bib29), [30](#bib.bib30)\]). Part of our future work will be to investigate these implementation issues into more detail.
## References
- \[1\] “Network Provisioning and Protocols for DOE Large-Science Applications,” in *Provisioning for Large-Scale Science Applications*, N. S. Rao and W. R. Wing, Eds. Springer, New-York, April 2003, argonne, IL.
- \[2\] N.S.V. Rao and W.R. Wing and S.M. Carter and Q. Wu, “UltraScience Net: Network Testbed for Large-Scale Science Applications,” *IEEE Communications Magazine*, vol. , 2005.
- \[3\] R. Cohen and N. Fazlollahi and D. Starobinski, “Graded Channel Reservation with Path Switching in Ultra High Capacity Networks,” in *Gridnets’06*, October 2006, san-Jose, CA.
- \[4\] “On-demand secure circuits and advanced reservation systems,” http://www.es.net/oscars/index.html.
- \[5\] “Dynamic resource allocation via gmpls optical networks,” http://dragon.maxgigapop.net.
- \[6\] Guerin, R. A. and Orda, A., “Networks With Advance Reservations: The Routing Perspective,” in *Proceedings of INFOCOM’00*, March 2000, tel-Aviv, Israel.
- \[7\] N. Fazlollahi and R. Cohen and D. Starobinski, “On the Capacity Limits of Advanced Channel Reservation Architectures,” May 2007, proc. of the High-Speed Networks Workshop. http://people.bu.edu/staro/Starobinski-abstract.pdf.
- \[8\] A. Goel and M. Henzinger and S. Plotkin and E. Tardos, “Scheduling Data Transfers in a Network and the Set Scheduling Problem,” in *Proc. of the 31st Annual ACM Symposium on the Theory of Computing*, 1999, pp. 189–199.
- \[9\]
M. Andrews and A. Fernandez and A. Goel and L. Zhang, “Source routing and scheduling in packet networks,” *Journal of the ACM (JACM)*, vol. 52, no. 4, pp. 582–601, July 2005.
- \[10\] Y. Azar and J. Naor and R. Rom, “Routing strategies for fast networks,” in *Proc. of IEEE Infocom*, 1992.
- \[11\]
B. Awerbuch, Y. Azar, A. Fiat, S. Leonardi, A. Rosen, “On-line competitive algorithms for call admission in optical networks,” *Algorithmica*, vol. 31, no. 1, pp. 29–43, 2001.
- \[12\] Erlebach, T., “Call admission control for advance reservation requests with alternatives,” ETH, Zurich, Tech. Rep. 142, 2002.
- \[13\]
S. A. Plotkin, “Competitive Routing of Virtual Circuits in ATM Networks,” *IEEE J. on Selected Areas in Communication (JSAC)*, vol. 13, no. 6, pp. 1128–1136, 1995.
- \[14\] B. Awerbuch, D. Holmer, H. Rubens, and R. D. Kleinberg, “Provably competitive adaptive routing.” in *INFOCOM*, 2005, pp. 631–641.
- \[15\] S. Leonardi, “On-line network routing.” in *Online Algorithms*, 1996, pp. 242–267.
- \[16\] A. Goel, M. R. Henzinger, and S. A. Plotkin, “An online throughput-competitive algorithm for multicast routing and admission control.” *J. Algorithms*, vol. 55, no. 1, pp. 1–20, 2005.
- \[17\] Ganjali, Y. and Keshavarzian, A. and Shah, D., “Input Queued Switches: Cell Switching vs. Packet Switching,” in *Proceedings of INFOCOM’03*, March 2000, san-Fancisco, CA, USA.
- \[18\] Weichenberg, G. and Chan, V. W. S. and Médard, M., “On the Capacity of Optical Networks: A Framework for Comparing Different Transport Architectures,” in *Proceedings of INFOCOM’06*, March 2000, barcelona, Spain.
- \[19\]
E. Gustafsson and G. Karlsson, “A Literature Survey on Traffic Dispersion,” *IEEE Network*, vol. 11, no. 2, pp. 28–36, March-April 1997.
- \[20\]
J. Shen and J. Shi and J. Crowcroft, “Proactive Multi-path Routing,” *Lect. Notes in Comp. Sci.*, vol. 2511, pp. 145–156, 2002.
- \[21\] R. Chow and C. W. Lee and and J. C. L. Liu, “Traffic Dispersion Strategies for Multimedia Streaming,” in *Proceedings of the 8th IEEE Workshop on Future Trends of Distributed Computing Systems*, 2001, p. 18.
- \[22\]
F. Shahrokhi and D. W. Matula, “The Maximum Concurrent Flow Problem,” *Journal of the ACM (JACM)*, vol. 37, no. 2, April 1990.
- \[23\]
T. Leighton and S. Rao, “Multicommodity max-flow min-cut theorems and their use in designing approximation algorithms,” *Journal of the ACM (JACM)*, vol. 46, no. 4, pp. 787 – 832, November 1999.
- \[24\] “Coin-Or project,” http://www.coin-or.org/.
- \[25\] “National LambdaRail Inc.” http://www.nlr.net/.
- \[26\] N. Fazlollahi and R. Cohen and D. Starobinski, “On the Capacity Limits of Advanced Channel Reservation Architectures,” in *Proc. of the High-Speed Networks Workshop*, May 2007.
- \[27\] A. Schill and S. Kuhn and F. Breiter, “Resource Reservation in Advance in Heterogeneous Networks with Partial ATM Infrastructures,” in *Proceedings of INFOCOM’97*, April 1997, kobe, Japan.
- \[28\] B. Awerbuch and T. Leighton, “Improved approximation algorithms for the multi-commodity flow problem and local competitive routing in dynamic networks,” in *Proceedings of the twenty-sixth annual ACM symposium on Theory of computing (STOC)*, 1994, pp. 487 – 496.
- \[29\] A. Kamath and O. Palmon and S. Plotkin, “Fast approximation algorithm for minimum cost multicommodity flow,” in *Proc. of the sixth ACM-SIAM Symposium on the Discrete Algorithms (SODA)*, 1995, pp. 493–501.
- \[30\] B. Awerbuch and R. Khandekar and S. Rao, “Distributed Algorithms for Multicommodity Flow Problems via Approximate Steepest Descent Framework,” in *Proceedings of the ACM-SIAM symposium on Discrete Algorithms (SODA)*, 2007. |
0712.1224 | cs.CR | cyber | # Evaluating the Utility of Anonymized Network Traces for Intrusion Detection
Kiran Lakkaraju Department of Computer Science University of Illinois, Urbana-Champaign
Adam Slagell National Center for Supercomputing Applications University of Illinois, Urbana-Champaign
###### Abstract
Recently, it has become increasingly important for computer security researchers and incident investigators to have access to larger and more diverse data sets. At the same time, trends towards protecting customer privacy have grown as a result of many embarrassing releases—of supposedly anonymous information—which has been traced back to individual computer users \[[7](#bib.bib7), [15](#bib.bib15)\]. This has further increased reluctance of data owners to release large data sets to the research community or to share logs relevant to attacks from a common threat. The burgeoning field of data sanitization has helped alleviate some of these problems as it has recently provided many new tools for anonymizing sensitive data, but there is still a difficult trade-off to be negotiated between the data owner’s need for privacy and security and the analyst’s need for high utility data. Data sanitization policies must be created that are secure enough for the first party, but do not result in too much information loss to be usable to the second.
Necessary to solving this problem of negotiating policies for data sanitization is the ability to analyze the effects of anonymization on both the security of the sanitized data and the utility left after anonymization. In this paper, we focus on analyzing the utility of network traces post-anonymization. Of course, any such measure of utility will naturally be subjective to the type of analysis being performed. So this work scopes the problem to utility for the task of attack detection. We employ a methodology we developed that analyzes the effect of anonymization on Intrusion Detection Systems (IDS), and we provide the first rigorous analysis of single field anonymization on IDS effectiveness. This work will begin to answer the questions of whether the field affects anonymization more than the algorithm; which fields have a larger impact on utility; and which anonymization algorithms have a larger impact on utility.
## 1 Introduction
The ability to safely share log files and network traces has become increasing important to several communities: networking research, computer security research, incident response, and education \[[23](#bib.bib23)\]. Synthetically generated data is abundant, but has been highly criticized for many uses \[[14](#bib.bib14)\]. Honeynets can be useful in generating exercises for students and help meet the needs of educators \[[1](#bib.bib1)\], but very few honeynets have been setup on a scale to generate some of the large, cross-sectional data sets needed by computer security researchers \[[25](#bib.bib25)\]. Furthermore, using a honeynet does not necessarily release the owner from all legal responsibility when sharing data \[[19](#bib.bib19)\]. Lastly, nothing but real data will suit the needs of the incident responder who must share data about specific attacks under investigation of real machines. Therefore, there is a high demand for methods to share real log files and network traces within several communities.
At the same time as there is an increased need to share these data sets, there is increased reluctance. First, many data owners recognize the inherent security risks of releasing detailed information that could be used to map out their own networks, services or sensor locations \[[3](#bib.bib3), [4](#bib.bib4), [8](#bib.bib8), [10](#bib.bib10), [16](#bib.bib16), [17](#bib.bib17), [26](#bib.bib26), [28](#bib.bib28)\]. Secondly, there are serious privacy concerns that companies have about releasing customer data, especially in light of recent incidents where some publicly released data—believed to be anonymous—leaked information about specific users \[[7](#bib.bib7), [15](#bib.bib15)\]. Lastly, recent research has questioned the legality of releasing much of the data as has been done up until now \[[19](#bib.bib19)\].
The burgeoning field of data sanitization has helped address this tension by providing organizations, who wish to share their data, with new tools to anonymize computer and network logs \[[4](#bib.bib4), [5](#bib.bib5), [12](#bib.bib12), [16](#bib.bib16), [18](#bib.bib18), [20](#bib.bib20), [21](#bib.bib21), [22](#bib.bib22), [22](#bib.bib22), [26](#bib.bib26), [27](#bib.bib27)\]. However, little has been done to help users negotiate the difficult trade-off between the data owner’s need for security and privacy, and the data analyst’s need for high quality data—what is called the utility vs. security trade-off \[[20](#bib.bib20)\]. As anonymization is an inherently lossy process, and the data analyst wants information as close to the original as possible, there is always this tension and a need to negotiate policies that meet the needs of both parties.
Necessary to solving this problem of negotiating policies for data sanitization is the ability to analyze the effects of anonymization on both the security of the sanitized data and the utility left after anonymization. In this paper, we focus upon the latter problem of evaluating the effects of anonymization on the utility of the data sets to be shared. Of course, utility is subjective since it depends upon who is using the data, or more specifically, for what purpose it is being used. Hence, what is important to a researcher in the network measurements community may be completely irrelevant to the incident responder. Therefore, we have scoped this work to evaluating utility for attack detection.
The task of attack detection is an important part of the incident responder’s daily job. When investigating broad attacks, of which their organization is only a part, they may have to settle for anonymized logs from the other sites involved. It would be a similar case when using a distributed or collaborative intrusion detection system that crosses organizational boundaries. Output from the sensors may need to be anonymized. Not only is attack detection important in these “real world” applications, but it is important to the intrusion detection research community, as well. This community has often complained that their only good data sets to test new technologies against are synthetic. However, if they can still do their analysis with anonymized data, then they are more likely to obtain large, real data sets.
The utility of a data set is not only constrained by the type of analysis being done with it but also the type of data being shared. We have chosen to look at the effects of anonymization on one of the most commonly shared and general types of data, the pcap formatted network trace. From this type of data, many others can be derived (e.g., NetFlows) \[[2](#bib.bib2)\].
To quantitatively measure the ability to identify attacks in anonymized data, we developed what we call the IDS Utility Metric. This measurement evaluates and compares the false positive and negative rates of a baseline unanonymized data set with that of an anonymized data set. By doing this, we can automate an objective process to help us answer several questions: (1) How does anonymization of a particular field affect the ability to detect an attack, (2) are there unique effects when certain pairs or triplets of fields are anonymized together, and (3) how does the use of different types of anonymization algorithms affect attack identification? In this paper, we present the results of anonymizing a portion of the 1999 MIT/Lincoln Labs DARPA data set—by using the FLAIM \[[20](#bib.bib20)\] anonymization framework—with over 150 separate anonymization policies to help us answer these questions.
The rest of the paper is organized as follows. Section [2](#S2 "2 FLAIM: Framework for Log Anonymization and Information Management ‣ Evaluating the Utility of Anonymized Network Traces for Intrusion Detection") describes the anonymization algorithms used by FLAIM in our experiments, while section [3](#S3 "3 Methodology ‣ Evaluating the Utility of Anonymized Network Traces for Intrusion Detection") describes the methodology and setup of our experiments. In section [4](#S4 "4 Results and Analysis ‣ Evaluating the Utility of Anonymized Network Traces for Intrusion Detection"), we present our results and analysis. We survey the related work in section [6](#S6 "6 Related Work ‣ Evaluating the Utility of Anonymized Network Traces for Intrusion Detection"), and state our conclusions, as well as scope out future work to be done, in sections [7](#S7 "7 Conclusions ‣ Evaluating the Utility of Anonymized Network Traces for Intrusion Detection") and [8](#S8 "8 Future Work ‣ Evaluating the Utility of Anonymized Network Traces for Intrusion Detection").
## 2 FLAIM: Framework for Log Anonymization and Information Management
We chose to use FLAIM \[[20](#bib.bib20)\] as our anonymization engine for several reasons. First, we could easily script its execution for a multitude of tests. Second, it has a very flexible XML policy language that made it simple to generate hundreds of unique anonymization policies. Third, it can anonymize as many or more fields in PCAP traces as any other anonymization tool. Lastly, FLAIM has a very rich set of anonymization algorithms that can be applied to all these fields. With these properties, it was the ideal tool for our experiments.
### 2.1 Anonymization Algorithms
FLAIM implements a plethora of anonymization algorithms for several data types. The three basic data types available to most anonymization algorithms are *binary, string,* and *numeric*. Additionally, there are a few special data types like timestamps. Binary data is just treated as a string of bits with no special structure. Algorithms anonymizing binary data output binary data of the same length. String data are variable length, terminated by a null character. Anonymization algorithms that take in strings will also output stings. However, the length may change. Numeric data is interpreted as a number of a given base, specified in the anonymization policy. This is useful, for example, when working with decimal numbers like a port number. There, one may want to act on individual digits, rather than bits (e.g., replacing the last 3 digits with 0’s).
Table [10](#S9.T10 "Table 10 ‣ 9 Appendix ‣ Evaluating the Utility of Anonymized Network Traces for Intrusion Detection") lists the different anonymization algorithms in FLAIM along with the data types they operate upon. Further information on these algorithms can be found in \[[6](#bib.bib6)\].
### 2.2 Anonymization Policies in FLAIM
FLAIM provides an expressive and powerful method for specifying anonymization policies that can be modified at run time, thus enabling efficient automation. An anonymization policy is an XML file that specifies the anonymization algorithms that should be applied to the various fields in the log, along with any special parameters to be passed to those algorithms.
## 3 Methodology
[Refer to caption](/html/0712.1224/assets/x1.png)
Figure 1: This diagram illustrates the process by which we compare anonymized logs versus non-anonymized logs.
For all of our experiments, we used a subset of the 1999 DARPA evaluation data set. The Defense Advanced Research Projects Office (DARPA) created an Intrusion Detection Evaluation testbed in 1998 and 1999. Data was captured from a simulated network that was subjected to various attacks. This data set has been frequently used in evaluating intrusion detection systems since its creation \[[11](#bib.bib11)\]. Thus, we found it appropriate to use in evaluating the effects anonymization of data can have on intrusion detection.
As we mention, we used but a portion of the 1999 data set. Specifically, we used the inside tcpdump data from Wednesday of the second week of the evaluation. Since FLAIM currently just anonymizes TCP, UDP and ICMP, we filtered out all other network protocols before running our experiments, whose methodology is depicted in Figure [1](#S3.F1 "Figure 1 ‣ 3 Methodology ‣ Evaluating the Utility of Anonymized Network Traces for Intrusion Detection").
The IDS Utility Metric is constructed as follows. First, the unanonymized data set is processed by Snort to produce the “baseline” set of alerts, using the default Snort rule sets. We consider this the ideal set of alerts for the data set. Next, we take the same data set, and anonymize it—in our case, with FLAIM. We then take the anonymized data set, and we run Snort against it with the same rule set as before. The alerts generated from the unanonymized file are used as a baseline against which the alerts generated by the anonymized file are compared. The difference between the alerts in the anonymized file, versus the unanonymized file, is used as a measure of the loss of utility in the log. The larger the difference, the more information was not available in the logs in order for the IDS to correctly identify an attack. This process is depicted in Figure [1](#S3.F1 "Figure 1 ‣ 3 Methodology ‣ Evaluating the Utility of Anonymized Network Traces for Intrusion Detection").
FLAIM schemas specify a set of anonymization algorithms that are appropriate for each field in a pcap log. These are summarized in Figure [4](#S9.F4 "Figure 4 ‣ 9 Appendix ‣ Evaluating the Utility of Anonymized Network Traces for Intrusion Detection"). In this evaluation, we only consider anonymization policies that transform single fields. Before we can evaluate the affect of multi-field policies, we must come to an understanding of single field policies. There are 152 single field policies that can be generated for pcap data (Figure [4](#S9.F4 "Figure 4 ‣ 9 Appendix ‣ Evaluating the Utility of Anonymized Network Traces for Intrusion Detection") in the Appendix lists all the fields and anonymization algorithms used for testing). Each anonymization algorithm also has parameters that affect how it anonymizes the field. We will not go over the parameters in detail, but instead refer readers to the FLAIM manual \[[6](#bib.bib6)\]. Table [9](#S9.T9 "Table 9 ‣ 9 Appendix ‣ Evaluating the Utility of Anonymized Network Traces for Intrusion Detection") in the Appendix summarizes the parameter settings for the anonymization algorithms.
We iterate the process described above over each of the 152 anonymization policies, comparing the results pre- and post-anonymization. We describe how we compare these data sets in the next section, while the section after discuses the actual metric in more detail.
### 3.1 Comparing Snort Alerts
Alerts generated by Snort are defined by several properties (a full list is included in the appendix as table [8](#S9.T8 "Table 8 ‣ 9 Appendix ‣ Evaluating the Utility of Anonymized Network Traces for Intrusion Detection")). Each alert is associated with a specific packet. The relevant alert fields—for our purposes—are listed below:
timestamp
: the timestamp from the offending packet.
sig_generator
: the part of Snort generating alert.
sig_id
: the Id. number of the signature that was fired.
msg
: description of the alert.
proto
: the protocol of the offending packet.
src
: the source IP address of the offending packet.
srcport
: The source port of the offending packet.
dst
: the destination IP address of the offending packet.
dstport
: the destination port of the offending packet.
id
: Packet Id.
To determine whether two alert sets are equal we need a way of determining if two alerts are equal. Normally this can be done by comparing each field of the alerts. However, in this case the alerts generated from the anonymized log will cause alerts that are actually equal to appear unequal. To overcome this, we compare alerts on fields which will not change due to anonymization. This leads to two distinct field sets that must be used when comparing alerts. They are shown in Table [1](#S3.T1 "Table 1 ‣ 3.1 Comparing Snort Alerts ‣ 3 Methodology ‣ Evaluating the Utility of Anonymized Network Traces for Intrusion Detection"). Field Set 1 is used when the timestamp field has not been anonymized. Field Set 2 is used when the timestamp field has been anonymized.
<table>
<thead>
<tr>
<th>Field Set 1</th>
<th>Field Set 2</th>
</tr>
</thead>
<tbody>
<tr>
<td>timestamp</td>
<td>sig_id</td>
</tr>
<tr>
<td>sig_id</td>
<td>src</td>
</tr>
<tr>
<td>id</td>
<td>srcport</td>
</tr>
<tr>
<td></td>
<td>dst</td>
</tr>
<tr>
<td></td>
<td>dstport</td>
</tr>
<tr>
<td></td>
<td>id</td>
</tr>
<tr>
<td></td>
<td>tcpseq</td>
</tr>
</tbody>
</table>
Table 1: The two field sets that are used to compare alerts.
### 3.2 Metrics for evaluating utility
The purpose of anonymization is to share logs while hiding sensitive information. Anonymization, while inherently an information reducing procedure, must be measured in terms of the amount of information that is lost in the file. However, by anonymizing we can introduce new false patterns into the data. The “best” anonymization policy should minimize information loss, while not adding any new false patterns to the log.
We can consider the IDS process as a pattern classification process. The data set is input, and the IDS classifies each packet as malicious or not. The alerts generated from the unanonymized data (which we call the baseline data) are considered to be the correct analysis of the data set. We then compare the alerts generated by the anonymized data to the baseline data alert set.
Let the set of alerts generated from the baseline file be called $`\mathcal{A}_{b\hspace{0pt}a\hspace{0pt}s\hspace{0pt}e\hspace{0pt}l\hspace{0pt}i\hspace{0pt}n\hspace{0pt}e}`$ and the set of alerts generated from the anonymized file be called $`\mathcal{A}_{a\hspace{0pt}n\hspace{0pt}o\hspace{0pt}n\hspace{0pt}y}`$. In terms of alerts we should compare $`\mathcal{A}_{a\hspace{0pt}n\hspace{0pt}o\hspace{0pt}n\hspace{0pt}y}`$vs. $`\mathcal{A}_{b\hspace{0pt}a\hspace{0pt}s\hspace{0pt}e\hspace{0pt}l\hspace{0pt}i\hspace{0pt}n\hspace{0pt}e}`$. The best result would be for the anonymized alerts to match, exactly, the alerts generated in the baseline set. Let us consider the baseline alerts to be the target set. The alerts from the anonymized file will be the generated set. Then we can define several metrics:
True Positive
$`TP = \mathcal{A}_{a\hspace{0pt}n\hspace{0pt}o\hspace{0pt}n\hspace{0pt}y} \cap \mathcal{A}_{b\hspace{0pt}a\hspace{0pt}s\hspace{0pt}e\hspace{0pt}l\hspace{0pt}i\hspace{0pt}n\hspace{0pt}e}|`$ The number of alerts in $`\mathcal{A}_{a\hspace{0pt}n\hspace{0pt}o\hspace{0pt}n\hspace{0pt}y}`$that are also in $`\mathcal{A}_{b\hspace{0pt}a\hspace{0pt}s\hspace{0pt}e\hspace{0pt}l\hspace{0pt}i\hspace{0pt}n\hspace{0pt}e}`$.
False Positive
$`{F\hspace{0pt}P} = {|{\mathcal{A}_{a\hspace{0pt}n\hspace{0pt}o\hspace{0pt}n\hspace{0pt}y} - {({\mathcal{A}_{a\hspace{0pt}n\hspace{0pt}o\hspace{0pt}n\hspace{0pt}y} \cap \mathcal{A}_{b\hspace{0pt}a\hspace{0pt}s\hspace{0pt}e\hspace{0pt}l\hspace{0pt}i\hspace{0pt}n\hspace{0pt}e}})}}|}`$ The number of alerts that were generated by the anonymized file, but were not in the baseline file.
False Negative
$`{F\hspace{0pt}N} = {|{\mathcal{A}_{b\hspace{0pt}a\hspace{0pt}s\hspace{0pt}e\hspace{0pt}l\hspace{0pt}i\hspace{0pt}n\hspace{0pt}e} - {({\mathcal{A}_{a\hspace{0pt}n\hspace{0pt}o\hspace{0pt}n\hspace{0pt}y} \cap \mathcal{A}_{b\hspace{0pt}a\hspace{0pt}s\hspace{0pt}e\hspace{0pt}l\hspace{0pt}i\hspace{0pt}n\hspace{0pt}e}})}}|}`$ The number of alerts that were not caught by the anonymized file.
The True Positive rate indicates how much of the information was preserved in the anonymized log. The False Positive rate indicates how many additional patterns were added to the log through anonymization. The False Negative rate indicate the amount of information that was removed from the log. A good anonymization policy should make sure both False Positive and False Negative are low while maximizing the True Positive rate.
While the false positive rate is an important factor of primary importance is the false negative rate, as it indicates the loss of information through anonymization. For the remainder of this paper, we use both False Positive and False Negative rates as a measure of the utility of a log post-anonymization, but focus more on the false negative rate.
## 4 Results and Analysis
In this section, we describe the results of our experiments with single field anonymization policies. These experiments provide a substantive start to answering these questions:
- •
What affects utility more, the fields that are anonymized or the anonymization algorithm?
- •
Which fields have a larger impact on the utility of a log?
- •
Which anonymization algorithms have a larger impact on the utility of a log?
To answer these questions, we evaluated all 152 pairs of fields and anonymization algorithms. For each pair, the number of alerts generated by Snort was calculated. The alerts generated for each pair were compared with the baseline alerts (See Table [2](#S4.T2 "Table 2 ‣ 4 Results and Analysis ‣ Evaluating the Utility of Anonymized Network Traces for Intrusion Detection") and Table [3](#S4.T3 "Table 3 ‣ 4 Results and Analysis ‣ Evaluating the Utility of Anonymized Network Traces for Intrusion Detection")). False Positives/False Negatives were calculated based on the definitions above, whose detailed results we discuss later.
The unanonymized file produced 81 alerts. The number and types of alerts produced are summarized in Table [4](#S4.T4 "Table 4 ‣ 4 Results and Analysis ‣ Evaluating the Utility of Anonymized Network Traces for Intrusion Detection"). There are a total of 81 alerts generated in the baseline file, but only 19 unique types of alerts.
Table 2: Alerts generated for Anonymization-Field pairs. AnonyAlg is the anonymization algorithm used; Field is the field which it was applied on; NumAlerts is the number of alerts that were generated; NumTypesOfAlerts is the number of different *types* of alerts generated. Table 1 of 2.
<table>
<thead>
<tr>
<th>AnonyAlg</th>
<th>Field</th>
<th>NumAlerts</th>
<th>NumTypes OfAlerts</th>
</tr>
</thead>
<tbody>
<tr>
<td>BinaryBlackMarker</td>
<td>SRC_MAC</td>
<td>81</td>
<td>19</td>
</tr>
<tr>
<td>BytesTruncation</td>
<td>SRC_MAC</td>
<td>81</td>
<td>19</td>
</tr>
<tr>
<td>Annihilation</td>
<td>SRC_MAC</td>
<td>81</td>
<td>19</td>
</tr>
<tr>
<td>MacRandomPermutation</td>
<td>SRC_MAC</td>
<td>81</td>
<td>19</td>
</tr>
<tr>
<td>BinaryBlackMarker</td>
<td>DST_MAC</td>
<td>81</td>
<td>19</td>
</tr>
<tr>
<td>BytesTruncation</td>
<td>DST_MAC</td>
<td>81</td>
<td>19</td>
</tr>
<tr>
<td>Annihilation</td>
<td>DST_MAC</td>
<td>81</td>
<td>19</td>
</tr>
<tr>
<td>MacRandomPermutation</td>
<td>DST_MAC</td>
<td>81</td>
<td>19</td>
</tr>
<tr>
<td>IPv4PrefixPreserving</td>
<td>IPV4_SRC_IP</td>
<td>1014</td>
<td>5</td>
</tr>
<tr>
<td>BinaryBlackMarker</td>
<td>IPV4_SRC_IP</td>
<td>9</td>
<td>4</td>
</tr>
<tr>
<td>Annihilation</td>
<td>IPV4_SRC_IP</td>
<td>9</td>
<td>4</td>
</tr>
<tr>
<td>RandomPermutation</td>
<td>IPV4_SRC_IP</td>
<td>9</td>
<td>4</td>
</tr>
<tr>
<td>NumericTruncation</td>
<td>IPV4_SRC_IP</td>
<td>9</td>
<td>4</td>
</tr>
<tr>
<td>IPv4PrefixPreserving</td>
<td>IPV4_DST_IP</td>
<td>759</td>
<td>5</td>
</tr>
<tr>
<td>BinaryBlackMarker</td>
<td>IPV4_DST_IP</td>
<td>9</td>
<td>4</td>
</tr>
<tr>
<td>Annihilation</td>
<td>IPV4_DST_IP</td>
<td>9</td>
<td>4</td>
</tr>
<tr>
<td>RandomPermutation</td>
<td>IPV4_DST_IP</td>
<td>14</td>
<td>5</td>
</tr>
<tr>
<td>NumericTruncation</td>
<td>IPV4_DST_IP</td>
<td>9</td>
<td>4</td>
</tr>
<tr>
<td>BinaryBlackMarker</td>
<td>IPV4_ID</td>
<td>81</td>
<td>19</td>
</tr>
<tr>
<td>Annihilation</td>
<td>IPV4_ID</td>
<td>81</td>
<td>19</td>
</tr>
<tr>
<td>NumericTruncation</td>
<td>IPV4_ID</td>
<td>81</td>
<td>19</td>
</tr>
<tr>
<td>RandomPermutation</td>
<td>IPV4_ID</td>
<td>81</td>
<td>19</td>
</tr>
<tr>
<td>Classify</td>
<td>IPV4_ID</td>
<td>81</td>
<td>19</td>
</tr>
<tr>
<td>Annihilation</td>
<td>IPV4_OFFSET</td>
<td>81</td>
<td>19</td>
</tr>
<tr>
<td>BinaryBlackMarker</td>
<td>IPV4_TTL</td>
<td>81</td>
<td>19</td>
</tr>
<tr>
<td>Annihilation</td>
<td>IPV4_TTL</td>
<td>81</td>
<td>19</td>
</tr>
<tr>
<td>NumericTruncation</td>
<td>IPV4_TTL</td>
<td>81</td>
<td>19</td>
</tr>
<tr>
<td>RandomPermutation</td>
<td>IPV4_TTL</td>
<td>81</td>
<td>19</td>
</tr>
<tr>
<td>Classify</td>
<td>IPV4_TTL</td>
<td>81</td>
<td>19</td>
</tr>
<tr>
<td>Annihilation</td>
<td>IPV4_CHECKSUM</td>
<td>81</td>
<td>19</td>
</tr>
<tr>
<td>BinaryBlackMarker</td>
<td>TCP_DST_PORT</td>
<td>520290</td>
<td>2</td>
</tr>
<tr>
<td>NumericTruncation</td>
<td>TCP_DST_PORT</td>
<td>457410</td>
<td>6</td>
</tr>
<tr>
<td>Substitution</td>
<td>TCP_DST_PORT</td>
<td>923033</td>
<td>2</td>
</tr>
<tr>
<td>Annihilation</td>
<td>TCP_DST_PORT</td>
<td>923033</td>
<td>2</td>
</tr>
<tr>
<td>RandomPermutation</td>
<td>TCP_DST_PORT</td>
<td>18</td>
<td>2</td>
</tr>
<tr>
<td>Classify</td>
<td>TCP_DST_PORT</td>
<td>521670</td>
<td>2</td>
</tr>
<tr>
<td>BinaryBlackMarker</td>
<td>TCP_SRC_PORT</td>
<td>380527</td>
<td>4</td>
</tr>
<tr>
<td>NumericTruncation</td>
<td>TCP_SRC_PORT</td>
<td>294519</td>
<td>6</td>
</tr>
</tbody>
</table>
Table 3: Alerts generated for Anonymization-Field pairs. AnonyAlg is the anonymization algorithm used; Field is the field which it was applied on; NumAlerts is the number of alerts that were generated; NumTypesOfAlerts is the number of different *types* of alerts generated. Table 2 of 2
<table>
<thead>
<tr>
<th>AnonyAlg</th>
<th>Field</th>
<th>NumAlerts</th>
<th>NumTypes OfAlerts</th>
</tr>
</thead>
<tbody>
<tr>
<td>Substitution</td>
<td>TCP_SRC_PORT</td>
<td>922171</td>
<td>6</td>
</tr>
<tr>
<td>Annihilation</td>
<td>TCP_SRC_PORT</td>
<td>922171</td>
<td>6</td>
</tr>
<tr>
<td>RandomPermutation</td>
<td>TCP_SRC_PORT</td>
<td>20</td>
<td>4</td>
</tr>
<tr>
<td>Classify</td>
<td>TCP_SRC_PORT</td>
<td>381954</td>
<td>4</td>
</tr>
<tr>
<td>BinaryBlackMarker</td>
<td>TCP_SEQUENCE</td>
<td>81</td>
<td>19</td>
</tr>
<tr>
<td>NumericTruncation</td>
<td>TCP_SEQUENCE</td>
<td>65</td>
<td>15</td>
</tr>
<tr>
<td>Annihilation</td>
<td>TCP_SEQUENCE</td>
<td>65</td>
<td>15</td>
</tr>
<tr>
<td>Classify</td>
<td>TCP_SEQUENCE</td>
<td>81</td>
<td>19</td>
</tr>
<tr>
<td>BinaryBlackMarker</td>
<td>TCP_ACK_NO</td>
<td>62</td>
<td>16</td>
</tr>
<tr>
<td>NumericTruncation</td>
<td>TCP_ACK_NO</td>
<td>56</td>
<td>14</td>
</tr>
<tr>
<td>Annihilation</td>
<td>TCP_ACK_NO</td>
<td>56</td>
<td>14</td>
</tr>
<tr>
<td>Classify</td>
<td>TCP_ACK_NO</td>
<td>81</td>
<td>19</td>
</tr>
<tr>
<td>BinaryBlackMarker</td>
<td>TCP_FLAGS</td>
<td>5</td>
<td>3</td>
</tr>
<tr>
<td>NumericTruncation</td>
<td>TCP_FLAGS</td>
<td>5</td>
<td>3</td>
</tr>
<tr>
<td>Annihilation</td>
<td>TCP_FLAGS</td>
<td>5</td>
<td>3</td>
</tr>
<tr>
<td>BinaryBlackMarker</td>
<td>TCP_WINDOW</td>
<td>81</td>
<td>19</td>
</tr>
<tr>
<td>NumericTruncation</td>
<td>TCP_WINDOW</td>
<td>81</td>
<td>19</td>
</tr>
<tr>
<td>Annihilation</td>
<td>TCP_WINDOW</td>
<td>81</td>
<td>19</td>
</tr>
<tr>
<td>Classify</td>
<td>TCP_WINDOW</td>
<td>81</td>
<td>19</td>
</tr>
<tr>
<td>Annihilation</td>
<td>TCP_CHECKSUM</td>
<td>81</td>
<td>19</td>
</tr>
<tr>
<td>BinaryBlackMarker</td>
<td>TCP_URGENT</td>
<td>81</td>
<td>19</td>
</tr>
<tr>
<td>NumericTruncation</td>
<td>TCP_URGENT</td>
<td>81</td>
<td>19</td>
</tr>
<tr>
<td>Annihilation</td>
<td>TCP_URGENT</td>
<td>81</td>
<td>19</td>
</tr>
<tr>
<td>Classify</td>
<td>TCP_URGENT</td>
<td>81</td>
<td>19</td>
</tr>
<tr>
<td>Annihilation</td>
<td>TCP_OPTIONS</td>
<td>81</td>
<td>19</td>
</tr>
<tr>
<td>BinaryBlackMarker</td>
<td>UDP_DST_PORT</td>
<td>81</td>
<td>19</td>
</tr>
<tr>
<td>NumericTruncation</td>
<td>UDP_DST_PORT</td>
<td>81</td>
<td>19</td>
</tr>
<tr>
<td>Substitution</td>
<td>UDP_DST_PORT</td>
<td>81</td>
<td>19</td>
</tr>
<tr>
<td>Annihilation</td>
<td>UDP_DST_PORT</td>
<td>81</td>
<td>19</td>
</tr>
<tr>
<td>RandomPermutation</td>
<td>UDP_DST_PORT</td>
<td>81</td>
<td>19</td>
</tr>
<tr>
<td>Classify</td>
<td>UDP_DST_PORT</td>
<td>81</td>
<td>19</td>
</tr>
<tr>
<td>BinaryBlackMarker</td>
<td>UDP_SRC_PORT</td>
<td>81</td>
<td>19</td>
</tr>
<tr>
<td>NumericTruncation</td>
<td>UDP_SRC_PORT</td>
<td>81</td>
<td>19</td>
</tr>
<tr>
<td>Substitution</td>
<td>UDP_SRC_PORT</td>
<td>81</td>
<td>19</td>
</tr>
<tr>
<td>Annihilation</td>
<td>UDP_SRC_PORT</td>
<td>81</td>
<td>19</td>
</tr>
<tr>
<td>RandomPermutation</td>
<td>UDP_SRC_PORT</td>
<td>81</td>
<td>19</td>
</tr>
<tr>
<td>Classify</td>
<td>UDP_SRC_PORT</td>
<td>81</td>
<td>19</td>
</tr>
<tr>
<td>Annihilation</td>
<td>UDP_CHECKSUM</td>
<td>81</td>
<td>19</td>
</tr>
<tr>
<td>RandomTimeShift</td>
<td>TS_SEC</td>
<td>81</td>
<td>19</td>
</tr>
<tr>
<td>TimeUnitAnnihilation</td>
<td>TS_SEC</td>
<td>81</td>
<td>19</td>
</tr>
<tr>
<td>Annihilation</td>
<td>TS_SEC</td>
<td>81</td>
<td>19</td>
</tr>
<tr>
<td>BinaryBlackMarker</td>
<td>TS_SEC</td>
<td>83</td>
<td>20</td>
</tr>
<tr>
<td>TimeEnumeration</td>
<td>TS_SEC</td>
<td>399</td>
<td>19</td>
</tr>
<tr>
<td>Annihilation</td>
<td>TS_USEC</td>
<td>81</td>
<td>19</td>
</tr>
</tbody>
</table>
Table 4: Alerts generated in the baseline unanonymized file. AlertID is the id of the alert; Num is the number of that type of alert generated; Desc is a description of the alert
<table>
<thead>
<tr>
<th>AlertID</th>
<th>Num</th>
<th>Desc</th>
</tr>
</thead>
<tbody>
<tr>
<td>1</td>
<td>1</td>
<td>(portscan) TCP Portscan 111This alert was generated from generator 122</td>
</tr>
<tr>
<td>323</td>
<td>1</td>
<td>FINGER root query</td>
</tr>
<tr>
<td>330</td>
<td>1</td>
<td>FINGER redirection attempt</td>
</tr>
<tr>
<td>332</td>
<td>1</td>
<td>FINGER 0 query</td>
</tr>
<tr>
<td>356</td>
<td>1</td>
<td>FTP passwd retrieval attempt</td>
</tr>
<tr>
<td>359</td>
<td>1</td>
<td>FTP satan scan</td>
</tr>
<tr>
<td>503</td>
<td>4</td>
<td>MISC Source Port 20 to ¡1024</td>
</tr>
<tr>
<td>1200</td>
<td>9</td>
<td>ATTACK-RESPONSES Invalid URL</td>
</tr>
<tr>
<td>1201</td>
<td>12</td>
<td>ATTACK-RESPONSES 403 Forbidden</td>
</tr>
<tr>
<td>1288</td>
<td>4</td>
<td>WEB-FRONTPAGE /_vti_bin/ access</td>
</tr>
<tr>
<td>1292</td>
<td>30</td>
<td>ATTACK-RESPONSES directory listing</td>
</tr>
<tr>
<td>1418</td>
<td>2</td>
<td>SNMP request tcp</td>
</tr>
<tr>
<td>1420</td>
<td>2</td>
<td>SNMP trap tcp</td>
</tr>
<tr>
<td>1421</td>
<td>1</td>
<td>SNMP AgentX/tcp request</td>
</tr>
<tr>
<td>2467</td>
<td>1</td>
<td>NETBIOS SMB D$ unicode share access</td>
</tr>
<tr>
<td>2470</td>
<td>1</td>
<td>NETBIOS SMB C$ unicode share access</td>
</tr>
<tr>
<td>2473</td>
<td>1</td>
<td>NETBIOS SMB ADMIN$ unicode share access</td>
</tr>
<tr>
<td>3151</td>
<td>6</td>
<td>FINGER / execution attempt</td>
</tr>
<tr>
<td>3441</td>
<td>2</td>
<td>FTP PORT bounce attempt</td>
</tr>
</tbody>
</table>
### 4.1 Fields or Anonymization Algorithms?
It is important to understand which causes a greater impact on utility; the field that is being anonymized or the anonymization algorithm that is being applied.
To evaluate the effects of anonymizing a field we calculate the *marginal* of a field. The marginal of a field is the average number of false positive/false negatives over all anonymization algorithms. Similarly, the marginal of an anonymization algorithm is the false positive/false negative rate averaged over all fields. The marginal provides a concise summary of the effect of anonymizing a particular field or using a particular anonymization algorithm.
Figure [2](#S4.F2 "Figure 2 ‣ 4.1 Fields or Anonymization Algorithms? ‣ 4 Results and Analysis ‣ Evaluating the Utility of Anonymized Network Traces for Intrusion Detection") shows the false positive and false negative marginals of the fields. The full data for these graphs is in Table [6](#S4.T6 "Table 6 ‣ 4.3 What anonymization algorithms have higher impact on utility? ‣ 4 Results and Analysis ‣ Evaluating the Utility of Anonymized Network Traces for Intrusion Detection"). Figure [3](#S4.F3 "Figure 3 ‣ 4.1 Fields or Anonymization Algorithms? ‣ 4 Results and Analysis ‣ Evaluating the Utility of Anonymized Network Traces for Intrusion Detection") shows the false positive and false negative marginals of the anonymization algorithms. The full data for these graphs is in Table [5](#S4.T5 "Table 5 ‣ 4.3 What anonymization algorithms have higher impact on utility? ‣ 4 Results and Analysis ‣ Evaluating the Utility of Anonymized Network Traces for Intrusion Detection").
<table>
<tbody>
<tr>
<td>Refer to caption</td>
<td>Refer to caption</td>
</tr>
</tbody>
</table>
Figure 2: The left hand chart shows the marginal of each field with respect to false positives (i.e, the average number of false positives for a field, averaged over all anonymization algorithms). The right hand chart shows the marginal of each field with respect to false negatives.
<table>
<tbody>
<tr>
<td>Refer to caption</td>
<td>Refer to caption</td>
</tr>
</tbody>
</table>
Figure 3: The left hand chart shows the marginal of all the anonymization algorithms, with respect to false positives. The right hand chart shows the marginal of all anonymization algorithms with respect to false negatives.
The fact that the majority of fields have no impact on utility, even when anonymized in numerous ways, indicates that fields are more important than the anonymization algorithm in determining the utility of an anonymized log. If it was the other way around, we would expect false positives and false negatives to be more evenly distributed over all the fields.
Figure [2](#S4.F2 "Figure 2 ‣ 4.1 Fields or Anonymization Algorithms? ‣ 4 Results and Analysis ‣ Evaluating the Utility of Anonymized Network Traces for Intrusion Detection") indicates that the majority of fields generate no false positives or false negatives. The fields for which this is true (such as DST_MAC) did not affect the utility of the log under any anonymization algorithm. We can conclude that anonymizing these fields has no impact on the utility of a log, with respect to the IDS metric.
Consider the false positive rate first. We can see that very few fields generated any false positives. This indicates that anonymization, usually, does not add new patterns.
### 4.2 Which fields have higher impact on utility?
As Figure [2](#S4.F2 "Figure 2 ‣ 4.1 Fields or Anonymization Algorithms? ‣ 4 Results and Analysis ‣ Evaluating the Utility of Anonymized Network Traces for Intrusion Detection") and Table [6](#S4.T6 "Table 6 ‣ 4.3 What anonymization algorithms have higher impact on utility? ‣ 4 Results and Analysis ‣ Evaluating the Utility of Anonymized Network Traces for Intrusion Detection") clearly show, the majority of fields do not generate false positives. The fields that do are TCP_DST_PORT, TCP_SRC_PORT, IPV4_SRC_IP, IPV4_DST_IP, TS_SEC, and TCP_ACK_NO.
We can see that several of the fields resulted in an average of 81 alerts. These fields (shown in italics in the table) had 0 error. Judging from these results, we can see that this set of fields did not affect the utility of the log as measured by the IDS metric.
TCP_DST_PORT and TCP_SRC_PORT generated the most false alerts on average. Upon inspection of the generated alert files, we can see that most of the pairs generated only 2 types of alerts. In the case of BinaryBlackMarker there were 520286 alerts of type 524. Alert 524 is “BAD-TRAFFIC tcp port 0 traffic”. The other anonymization algorithms produced the same pattern—the majority of alerts were of type 524.
The reason for this is the value we substitute for the port field. The BinaryBlackMarker, Substitution, NumericTruncation, and Annihilation algorithms all replaced the port field with $`0`$. This resulted in the alert being triggered for nearly all the packets in the log (See Table [9](#S9.T9 "Table 9 ‣ 9 Appendix ‣ Evaluating the Utility of Anonymized Network Traces for Intrusion Detection") for the parameter settings of the anonymization algorithms). The same problem occurs for the TCP_SRC_PORT field.
The RandomPermutation algorithm replaced the port number with another, random port number, thus infrequently causing the generation of the BAD-TRAFFIC alert. It is clear from this that the false positive rate was greatly affected by the choice of the substitution port. However, the effect would have been less pronounced had we counted the types of new alerts rather than the raw alerts, themselves.
Arguably, it is the false negative count that is most important in determining the utility of a log. A low false negative count indicates that little information was lost in the process of anonymization. In terms of false negatives, we find that there are $`8`$ fields that have an impact on the average false negative rate: TCP_DST_PORT,TCP_FLAGS,TCP_SRC_PORT, IPV4_DST_IP, IPV4_SRC_IP,TCP_ACK_NO, TCP_SEQUENCE, and TS_SEC.
### 4.3 What anonymization algorithms have higher impact on utility?
Figure [3](#S4.F3 "Figure 3 ‣ 4.1 Fields or Anonymization Algorithms? ‣ 4 Results and Analysis ‣ Evaluating the Utility of Anonymized Network Traces for Intrusion Detection") summarizes the false positive and false negative rates with respect to anonymization algorithm. For each anonymization algorithm, the average false positive/false negative rate is calculated over all the fields. Table [5](#S4.T5 "Table 5 ‣ 4.3 What anonymization algorithms have higher impact on utility? ‣ 4 Results and Analysis ‣ Evaluating the Utility of Anonymized Network Traces for Intrusion Detection") contains the data for the graphs.
We can see from these that most anonymization algorithms have an impact on the utility of a log. In contrast, the field data that we saw before showed strong structure in what fields affected utility. When viewed from the perspective of anonymization algorithms, there is no single anonymization algorithm that stands out.
We can see from these that most anonymization algorithms have an impact on the utility of a log. In contrast, the field data that we saw before showed strong structure in what fields affected utility. When viewed from the perspective of anonymization algorithms, there is no single anonymization algorithm that stands out.
It might seem like the substitution algorithm has the largest effect with a huge number of alerts. However, this is because of the parameter setting used. By looking at the false negative rate we can see that while substitution still has a large effect on utility, most of the other algorithms have an effect as well.
Table 5: False negative/positive counts for an anonymization algorithm, aggregated over all log fields.
<table>
<tbody>
<tr>
<th>Anonymization Alg.</th>
<td>False Negatives</td>
<td>False Positive</td>
</tr>
<tr>
<th>Annihilation</th>
<td>10.62</td>
<td>47312.69</td>
</tr>
<tr>
<th>BinaryBlackMarker</th>
<td>13.43</td>
<td>30027.37</td>
</tr>
<tr>
<th>BytesTruncation</th>
<td>0</td>
<td>0</td>
</tr>
<tr>
<th>Classify</th>
<td>8.5</td>
<td>50200.83</td>
</tr>
<tr>
<th>IPv4PrefixPreserving</th>
<td>28.8</td>
<td>351</td>
</tr>
<tr>
<th>MacRandomPermutation</th>
<td>0</td>
<td>0</td>
</tr>
<tr>
<th>NumericTruncation</th>
<td>17.96</td>
<td>32692.13</td>
</tr>
<tr>
<th>RandomPermutation</th>
<td>16.72</td>
<td>2.11</td>
</tr>
<tr>
<th>RandomTimeShift</th>
<td>0</td>
<td>0</td>
</tr>
<tr>
<th>Substitution</th>
<td>38</td>
<td>461298.5</td>
</tr>
<tr>
<th>TimeEnumeration</th>
<td>1.25</td>
<td>80.75</td>
</tr>
<tr>
<th>TimeUnitAnnihilation</th>
<td>0</td>
<td>0</td>
</tr>
</tbody>
</table>
Table 6: False negative/positive counts for a field, aggregated over all anonymization algorithms
<table>
<tbody>
<tr>
<th>Field</th>
<td>False Negatives</td>
<td>False Positive</td>
</tr>
<tr>
<th>DST_MAC</th>
<td>0</td>
<td>0</td>
</tr>
<tr>
<th>IPV4_CHECKSUM</th>
<td>0</td>
<td>0</td>
</tr>
<tr>
<th>IPV4_DST_IP</th>
<td>72</td>
<td>151</td>
</tr>
<tr>
<th>IPV4_ID</th>
<td>0</td>
<td>0</td>
</tr>
<tr>
<th>IPV4_OFFSET</th>
<td>0</td>
<td>0</td>
</tr>
<tr>
<th>IPV4_SRC_IP</th>
<td>72</td>
<td>201</td>
</tr>
<tr>
<th>IPV4_TTL</th>
<td>0</td>
<td>0</td>
</tr>
<tr>
<th>SRC_MAC</th>
<td>0</td>
<td>0</td>
</tr>
<tr>
<th>TCP_ACK_NO</th>
<td>20</td>
<td>2.75</td>
</tr>
<tr>
<th>TCP_CHECKSUM</th>
<td>0</td>
<td>0</td>
</tr>
<tr>
<th>TCP_DST_PORT</th>
<td>77.67</td>
<td>557572.33</td>
</tr>
<tr>
<th>TCP_FLAGS</th>
<td>76</td>
<td>0</td>
</tr>
<tr>
<th>TCP_OPTIONS</th>
<td>0</td>
<td>0</td>
</tr>
<tr>
<th>TCP_SEQUENCE</th>
<td>8</td>
<td>0</td>
</tr>
<tr>
<th>TCP_SRC_PORT</th>
<td>75.5</td>
<td>483554.83</td>
</tr>
<tr>
<th>TCP_URGENT</th>
<td>0</td>
<td>0</td>
</tr>
<tr>
<th>TCP_WINDOW</th>
<td>0</td>
<td>0</td>
</tr>
<tr>
<th>TS_SEC</th>
<td>1.2</td>
<td>65.2</td>
</tr>
<tr>
<th>TS_USEC</th>
<td>0</td>
<td>0</td>
</tr>
<tr>
<th>UDP_CHECKSUM</th>
<td>0</td>
<td>0</td>
</tr>
<tr>
<th>UDP_DST_PORT</th>
<td>0</td>
<td>0</td>
</tr>
<tr>
<th>UDP_SRC_PORT</th>
<td>0</td>
<td>0</td>
</tr>
</tbody>
</table>
## 5 Multi-Field Policy Analysis
By multi-field anonymization policies, we are referring to anonymization schemes that transform two or more fields in a log. The bulk of this work has focused on single field policy analysis. However, single field policy analysis will lay the groundwork for understanding more complex multi-field policies.
Table 7: This table shows that a multi-field policy has uncertain affects on the utility of a log
<table>
<thead>
<tr>
<th>Policy</th>
<th>Alerts</th>
</tr>
</thead>
<tbody>
<tr>
<td>IPV4_SRC_IP-IPv4PrefixPreserving</td>
<td>1014</td>
</tr>
<tr>
<td>TCP_SRC_PORT-RandomPermutation</td>
<td>7</td>
</tr>
<tr>
<td>TS_SEC-TimeEnumeration</td>
<td>399</td>
</tr>
<tr>
<td>IPV4_SRC_IP,TCP_SRC_PORT</td>
<td>1016</td>
</tr>
<tr>
<td>IPV4_SRC_IP,TCP_SRC_PORT and TS_SEC</td>
<td>1010</td>
</tr>
</tbody>
</table>
Multi-field policies are difficult to analyze because the fields are often related in subtle ways, not just very direct ways, such as between the ACK and SEQ numbers. The anonymization of additional fields certainly does not affect the total number of false positive alerts in a linear way. For example, examine Table [7](#S5.T7 "Table 7 ‣ 5 Multi-Field Policy Analysis ‣ Evaluating the Utility of Anonymized Network Traces for Intrusion Detection").
Here, we see that anonymizing 3 fields separately produces 1014, 7 and 399 false alerts, respectively. If we anonymize the first 2 fields, we get $`1016 \neq {1014 + 7}`$ alerts. This may not be too surprising. However, if we do all 3 fields together, we get only $`1010`$ alerts—rather than something around $`1420 = {1014 + 7 + 399}`$. This is actually fewer false alerts than any one field being anonymized in isolation.
So while there is always more information loss when anonymizing more fields, there may be a critical point at which one starts getting fewer false positives as they do additional anonymization. We suspect that this is always the case since complete anonymization will leave nothing to alert upon. But clearly, more analysis must be done on multi-field anonymization, and our future work will be focused on evaluating multi-field anonymization policies.
## 6 Related Work
Work in data sanitization for computer and network logs to date has focused almost entirely upon development of tools \[[20](#bib.bib20), [9](#bib.bib9), [27](#bib.bib27), [13](#bib.bib13), [5](#bib.bib5), [17](#bib.bib17)\] and anonymization algorithms \[[26](#bib.bib26)\], with little to no work doing any formal analysis of the effects of anonymization. This is an important aspect missing from the research body since without it, we just have a lot of tools to haphazardly anonymize data without knowing how to do it wisely or effectively. This is also in stark contrast to k-anonymity \[[24](#bib.bib24)\] which is usually applied to medical and census type data and has received much more attention from the research community.
While writing this paper, a new piece of work closely related to ours appeared \[[27](#bib.bib27)\]. Though this is mostly a paper presenting yet another anonymization tool for pcap logs, at the end they introduce a similar method of analyzing the effects of anonymization of pcap traces by use of an IDS. Theirs is a cursory analysis, only considering anonymization of one field at a time and for only a few different fields. Also, their analysis only considered the number of alerts, where the more alerts are generated the more “security analysis” is provided. Clearly, more than just the number of alerts needs to be considered when evaluating utility, such as the false positive and negative rates. We were unable to reproduce their results (which may be because they used an unspecified subset of the LBNL data set
<sup>2</sup>
<sup>2</sup>2http://www.icir.org/enterprise-tracing/
), but more troubling is the use of this data set in the first place. To evaluate the effects of anonymization, one must start with unanonymized data. However, this LBNL data set is already anonymized. So one has no clean baseline with which to compare the anonymized data in this case.
## 7 Conclusions
Anonymization can be a powerful tool to allow greater cooperation between organizations. The need for cooperation is strikingly clear. However, a clear understanding of the needs of the data provider and the client is necessary before flexible, effective sharing between organizations can occur.
The objective of this paper has been to begin to formally evaluate the utility vs. security trade off. The *IDS metric* is simple yet effective in evaluating the difference in utility when anonymizing different fields in different ways.
In this paper, we have focused on answering three questions: whether the field affects anonymization more than the algorithm; which fields have a larger impact on utility; and which anonymization algorithms have a larger impact on utility.
We have provided a thorough evaluation of single field anonymization polices upon pcap formatted network traces. We found that the primary impact on the utility of a log is not the particular anonymization algorithm, but rather the field that was anonymized.
In addition, we were able to empirically show a range of utilities for a log based on the field that was anonymized. The loss of utility was largest for ports and IP addresses. There was some loss of utility for the fields of ID, sequence number, flags, timestamp, and ACK number. However, for many of the fields there was no change in utility when anonymized.
This empirical evaluation provides the basis for further work on studying the impact of more complex anonymization schemes on the utility of a log.
## 8 Future Work
There are numerous ways in which this work can be extended. First of all, it is clear that evaluating utility via Snort generated IDS alerts will cause the utility to depend upon the rule set. In fact, it is as of yet unclear exactly how the rule set impacts the utility measure. Though, we suspect there is a significant effect since we found the specific fields anonymized has more effect than how it was anonymized, and Snort rules usually focus on a one or two fields.
While we used Snort for all of our experiments, it is just one IDS and only one type. It would be very interesting to investigate whether anomaly-based IDSs are affected in similar ways.
The strength of an anonymization algorithm is a measure of how difficult it is to “break” or “de-anonymize” a log that has been anonymized via the algorithm. It is clear that we want strong anonymization algorithms so that attackers will have a difficult time to break the algorithm. As \[[20](#bib.bib20)\] points out, there is a trade off between the security of an anonymization algorithm and the utility of the log. We have not discussed the strength of an anonymization algorithm in this work.
Our work is currently limited to anonymization policies for just one field. Our next step would be to extend this work to multiple field anonymization policies and to work with more realistic data. The DARPA evaluation data set is useful because it is supervised, however it is still synthetic. In the future, we will be gaining access to other large unanonymized data sets that we can use instead of the DARPA data. However, it is important for this research to have unanonymized baseline sets.
Finally, we have still focused on utility for just one task, attack detection. An incident responder does more than just detect attacks, and in the future, we could look at how anonymizing logs affects other important security related tasks—such as alert correlation.
## References
- \[1\] Honeynet project challenge of the month. http://www.honeynet.org/misc/chall.html, March 2003.
- \[2\] Argus: Qosient software. http://www.qosient.com, March 2008.
- \[3\] J. Bethencourt, J. Franklin, and M. Vernon. Mapping internet sensors with probe response attacks.
In Proceedings of the 14th USENIX Security Symposium, August 2005.
- \[4\] S. Coull, C. V. Wright, F. Monrose, M. P. Collins, and M. Reiter. Playing devil’s advocate: Inferring sensitive information from anonymized network traces.
In Proceedings of the 14th Annual Network and Distributed System Security Symposium, February 2007.
- \[5\] M. Foukarakis, D. Antoniades, S. Antonatos, and E. P. Markatos. Flexible and high-performance anonymization of netflow records using anontool.
In Proceedings of the 3rd SECOVAL Workshop, September 2007.
- \[6\] L. W. Group. Flaim core user’s guide, March 2008.
- \[7\] K. Hafner. Researchers yearn to use aol logs, but they hesitate. New York Times, August 23 2006.
- \[8\] T. Kohno, A. Broido, and K. C. Claffy. Remote physical device fingerprinting.
In Proceedings of the IEEE Symposium on Security and Privacy, May 2005.
- \[9\] Y. Li, A. Slagell, K. Luo, and W. Yurcik. Canine: A combined conversion and anonymization tool for processing netflows for security.
In Proceedings of the 10th International Conference on Telecommunication Systems, Modeling and Analysis, November 2005.
- \[10\] P. Lincoln, P. Porras, and V. Shmatikov. Privacy-preserving sharing and correlation of security alerts.
In Proceedings of the 13th USENIX Security Symposium, August 2004.
- \[11\] R. Lippmann, J. W. Haines, D. J. Fried, J. Korba, and K. Das. The 1999 darpa off-line intrusion detection evaluation.
Computer Networks, 2000.
- \[12\] E. Lundin and E. Jonsson. Privacy vs. intrusion detection analysis.
In Proceedings of the 2nd International Workshop on Recent Advances in Intrusion Detection, September 1999.
- \[13\] K. Luo, Y. Li, C. Ermopoulos, W. Yurcik, and A. Slagell. Scrub-pa: A multi-level, multi-dimensional anonymization tool for process accounting.
In ACM Computing Research Repository (CoRR), January 2006.
- \[14\] J. McHugh. Testing intrusion detection systems: A critique of the 1998 and 1999 darpa intrusion detection system evaluations as performed by lincoln laboratories.
ACM Transactions on Information and System Security, 3(4):262–294, 2000.
- \[15\] A. Narayanan and V. Shmatikov. How to break anonymity of the netflix prize dataset.
In ACM Computing Research Repository, October 2006.
- \[16\] R. Pang, M. Allman, V. Paxson, and J. Lee. The devil and packet trace anonymization.
Computer Communication Review, January 2006.
- \[17\] R. Pang and V. Paxson. A high-level programming environment for packet trace anonymization and transformation.
In Proceedings of the AMC SIGCOMM Conference, August 2003.
- \[18\] V. E. Seeberg and S. Petrovic. A new classification scheme for anonymization of real data used in ids benchmarking.
In Proceedings of the 2nd International Conference on Availability, Reliability and Security, April 2007.
- \[19\] D. Sicker, P. Ohm, and D. Grunwald. Legal issues surrounding monitoring during network research.
In Proceedings of the 7th ACM SIGCOMM conference on Internet measurement, August 2007.
- \[20\] A. Slagell, K. Lakkaraju, and X. Luo. Flaim: A multi-level anonymization framework for computer and network logs.
In Proceedings of the 20th USENIX Large Installation System Administration Conference, December 2006.
- \[21\] A. Slagell, Y. Li, and K. Luo. Sharing network logs for computer forensics: A new tool for the anonymization of netflow records.
In Proceedings of the Computer Network Forensics Research Workshop, August 2005.
- \[22\] A. Slagell, J. Wang, and W. Yurcik. Network log anonymization: Application of crypto-pan to cisco netflows.
In Proceedings of the NSF/AFRL Workshop on Secure Knowledge Management, September 2004.
- \[23\] A. Slagell and W. Yurcik. Sharing computer network logs for security and privacy: A motivation for new methodologies of anonymization.
In Proceedings of SECOVAL: The Workshop on the Value of Security through Collaboration, August 2005.
- \[24\] L. Sweeney. K-anonymity: A model for protecting privacy.
International Journal on Uncertainty, Fuzziness and Knowledge-based Systems, 10, 2002.
- \[25\] M. Vrable, J. Ma, J. Chen, D. Moore, E. Vandekieft, A. C. Snoeren, G. M. Voelker, and S. Savage. Scalability, fidelity, and containment in the potemkin virtual honeyfarm.
In Proceedings of the 20th Symposium on Operating System Principles, October 2005.
- \[26\] J. Xu, J. Fan, M. H. Ammar, and S. B. Moon. Prefix-preserving ip address anonymization: Measurement-based security evaluation and a new cryptography-based scheme.
In Proceedings of the 10th IEEE International Conference on Network Protocols, November 2002.
- \[27\] W. Yurcik and etc. Scrub-tcpdump: A multi-level packet anonymizer demonstrating privacy/analysis tradeoffs.
In Proceedings of the 3rd SECOVAL Workshop, September 2007.
- \[28\] M. Zalewski. p0f: Passive os fingerprinting tool. http://lcamtuf.coredump.cx/p0f.shtml, March 2008.
## 9 Appendix
Table 8: Fields in a SNORT alert.
<table>
<tbody>
<tr>
<td>• timestamp • sig_generator • sig_id • sig_rev • msg • proto • src • srcport • dst • dstport</td>
<td>• ethsrc • ethdst • ethlen • tcpflags • tcpseq • tcpack • tcplen • tcpwindow • ttl</td>
<td>• tos • id • dgmlen • iplen • icmptype • icmpcoe • icmpid • icmpseq</td>
</tr>
</tbody>
</table>
[Refer to caption](/html/0712.1224/assets/x6.png)
Figure 4: PCAP Fields and Anonymization Algorithms. Each section contains the fields (on the left) on which any of the anonymization algorithms on the right can be applied. For instance, Only the anonymization algorithms BinaryBlackMarker and Annihilation can be applied to the ICMP_ORIG_DATA field (the bottom left section)
Table 9: Parameters to the Anonymization Algorithms
<table>
<thead>
<tr>
<th>Anonymization Algorithm</th>
<th>Parameters</th>
<th>Description</th>
</tr>
</thead>
<tbody>
<tr>
<td>IPV4PrefixPreserving</td>
<td>Passphrase: foobar</td>
<td>Sets the anonymization passphrase.</td>
</tr>
<tr>
<td>MacRandomPermutation</td>
<td>None</td>
<td>No Parameters.</td>
</tr>
<tr>
<td>RandomTimeShift</td>
<td>lowerTimeShiftLimit: 16250000 upperTimeShiftLimit: 31500000</td>
<td>Shift time by a random amount between the lower and upper limits.</td>
</tr>
<tr>
<td>TimeUnitAnnihilation</td>
<td>timeField: years</td>
<td>Annihilate the years portion of the timestamp.</td>
</tr>
<tr>
<td>NumericTruncation</td>
<td>numShift: 5 radix: 2</td>
<td>Shorten the field by 5 bits.</td>
</tr>
<tr>
<td>TimeEnumeration</td>
<td>baseTime: 0 intervalSize: 1</td>
<td>Set the time of the oldest record to 0. One timestp is equal to adding 1 to the timestamp field.</td>
</tr>
<tr>
<td>RandomPermutation</td>
<td>None</td>
<td>No Parameters.</td>
</tr>
<tr>
<td>Annihilation</td>
<td>None</td>
<td>No Parameters.</td>
</tr>
<tr>
<td>Classify</td>
<td>configString: 1024:0,65536:65535</td>
<td>Set elements less than 1024 to 0, and all others to 65535.</td>
</tr>
<tr>
<td>BinaryBlackMarker</td>
<td>numMarks: 8 replacement: 0</td>
<td>Mark 8 bits as 0</td>
</tr>
<tr>
<td>BytesTruncation</td>
<td>numbits: 20 direction: left</td>
<td>Remove 20 bits, starting from the left.</td>
</tr>
<tr>
<td>Substitution</td>
<td>substitute: 0</td>
<td>Substitute field with 0.</td>
</tr>
</tbody>
</table>
Table 10: Anonymization algorithms with applicable data types.
<table>
<tbody>
<tr>
<td>Anonymization Alg.</td>
<td>Data Type(s)</td>
<td>Description</td>
</tr>
<tr>
<td>Prefix-preserving</td>
<td>binary</td>
<td>Implements prefix-preserving permuta-</td>
</tr>
<tr>
<td></td>
<td></td>
<td>tion described in [26]</td>
</tr>
<tr>
<td>Truncation</td>
<td>binary</td>
<td>Removes suffix or prefix of data by</td>
</tr>
<tr>
<td></td>
<td>string</td>
<td>specified number of units.</td>
</tr>
<tr>
<td></td>
<td>numeric</td>
<td></td>
</tr>
<tr>
<td>Hash</td>
<td>binary</td>
<td>Outputs cryptographic has of data.</td>
</tr>
<tr>
<td></td>
<td>string</td>
<td></td>
</tr>
<tr>
<td></td>
<td>hostname</td>
<td></td>
</tr>
<tr>
<td>Black Marker</td>
<td>binary</td>
<td>Overwrites specified number of units</td>
</tr>
<tr>
<td></td>
<td>string</td>
<td>with specified constant.</td>
</tr>
<tr>
<td></td>
<td>hostname</td>
<td></td>
</tr>
<tr>
<td>Time Unit Annihilation</td>
<td>timestamp</td>
<td>Annihilates particular time units</td>
</tr>
<tr>
<td></td>
<td></td>
<td>(e.g., hour and minute units).</td>
</tr>
<tr>
<td>Random Time Shift</td>
<td>timestamp</td>
<td>Randomly shifts timestamps within</td>
</tr>
<tr>
<td></td>
<td></td>
<td>given window by same amount.</td>
</tr>
<tr>
<td>Enumeration</td>
<td>timestamp</td>
<td>Preserves order, but not distance</td>
</tr>
<tr>
<td></td>
<td></td>
<td>between elements.</td>
</tr>
<tr>
<td>Random Permutation</td>
<td>binary</td>
<td>Creates random 1-to-1 mapping.</td>
</tr>
<tr>
<td>Annihilation</td>
<td>binary</td>
<td>Replaces field with NULL value.</td>
</tr>
<tr>
<td></td>
<td>string</td>
<td></td>
</tr>
<tr>
<td>Classify</td>
<td>numeric</td>
<td>Partitions data into multiple non-</td>
</tr>
<tr>
<td></td>
<td></td>
<td>overlapping subsets.</td>
</tr>
<tr>
<td>Substitution</td>
<td>binary</td>
<td>Replaces all instances with a particu-</td>
</tr>
<tr>
<td></td>
<td>numeric</td>
<td>lar constant value.</td>
</tr>
</tbody>
</table> |
0710.4828 | cs.CR | cyber | <sup>1</sup>
<sup>1</sup>footnotetext: Correspondence should be addressed to hhaji@sbu.ac.ir.
<sup>2</sup>
<sup>2</sup>footnotetext: <sup>∗</sup> This paper is partially supported by Shahid Beheshti University.
Bounds for Visual Cryptography Schemes
Hossein Hajiabolhassan <sup>∗</sup> and Abbas Cheraghi
Department of Mathematical Sciences
Shahid Beheshti University, G.C.
P.O. Box 1983963113, Tehran, Iran
hhaji@sbu.ac.ir
Department of Mathematics
Faculty of Khansar
University of Isfahan
Isfahan, Iran
cheraghi@sci.ui.ac.ir
###### Abstract
In this paper, we investigate the best pixel expansion of the various models of visual cryptography schemes. In this regard, we consider visual cryptography schemes introduced by Tzeng and Hu \[[13](#bib.bib13)\]. In such a model, only minimal qualified sets can recover the secret image and that the recovered secret image can be darker or lighter than the background. Blundo et al. \[[4](#bib.bib4)\] introduced a lower bound for the best pixel expansion of this scheme in terms of minimal qualified sets. We present another lower bound for the best pixel expansion of the scheme. As a corollary, we introduce a lower bound, based on an induced matching of hypergraph of qualified sets, for the best pixel expansion of the aforementioned model and the traditional model of visual cryptography realized by basis matrices. Finally, we study access structures based on graphs and we present an upper bound for the smallest pixel expansion in terms of strong chromatic index.
-
Key words: visual cryptography, secret sharing scheme, hypergraph, basis matrices, pixel expansion, contrast.
-
Subject classification: 94A62.
## 1 Introduction
Visual cryptography schemes (VCS) are a special kind of secret sharing schemes in which secret is an image. For a set $`\mathcal{P}`$ of $`n`$ participants, a VCS encrypts a secret image into $`n`$ transparencies which constitute the shares given to the $`n`$ participants. The power set of participants is usually divided into qualified sets, which can visually recover the secret image by stacking their transparencies without any cryptography knowledge, and forbidden sets which have no information on the secret image.
The fascinating idea of visual cryptography was first introduced by Naor and Shamir \[[12](#bib.bib12)\]. Naor and Shamir \[[12](#bib.bib12)\] have proved that the pixel expansion of any visual $`k`$ out of $`k`$ scheme must be at least $`2^{k - 1}`$. Also, they have presented a visual $`k`$ out of $`k`$ scheme with pixel expansion $`2^{k - 1}`$.
Most papers on visual cryptography investigate two parameters, the pixel expansion and the contrast. The pixel expansion is the number of subpixels used to encode each pixel of the secret image in a share, that should be as small as possible. The contrast measures the “difference” between a black and a white pixel in the reconstructed image. Several results on the contrast and the pixel expansion of VCSs can be found in \[[1](#bib.bib1), [2](#bib.bib2), [3](#bib.bib3), [4](#bib.bib4), [6](#bib.bib6), [8](#bib.bib8), [12](#bib.bib12), [14](#bib.bib14)\]. Finding the best pixel expansion in the different models of VCS is the main challenge in visual cryptography. The problem of determining the best visual contrast (regardless of pixel expansion) is completely resolved \[[9](#bib.bib9), [10](#bib.bib10)\], so that, for the sake of completeness, it is interesting to find bound for the pixel expansion.
In this paper, we investigate the best pixel expansion of the different models of visual cryptography schemes. In the second section, we present several models of visual cryptography schemes. In the third section, we introduce some lower bounds for the best pixel expansion of different models of visual cryptography schemes. In this regard, we consider visual cryptography schemes introduced by Tzeng and Hu \[[13](#bib.bib13)\]. In such a model, only minimal qualified sets can recover the secret image and that the recovered secret image can be darker or lighter than the background. Blundo et al. \[[4](#bib.bib4)\] introduced a lower bound for the best pixel expansion of this scheme in terms of minimal qualified sets. We present another lower bound for the best pixel expansion of the scheme. As a corollary, we introduce a lower bound, based on an induced matching of hypergraph of qualified sets, for the best pixel expansion of the aforementioned model and the traditional model of visual cryptography realized by basis matrices. Finally, we study access structures based on graphs and we present an upper bound for the smallest pixel expansion in terms of strong chromatic index.
## 2 Models and Notations
First, we mention some of definitions and notations which are referred to throughout the paper. Hereafter, the symbol $`\mathcal{P}`$ stands for the set of participants. Furthermore, we assume that $`\mathcal{P} = {\{ 1,2,\ldots,n\}}`$ and let $`2^{\mathcal{P}}`$ denote the power set of $`\mathcal{P}`$, i.e., the set of all subsets of $`\mathcal{P}`$. A family $`\mathcal{Q} \subseteq 2^{\mathcal{P}}`$ is said to be monotone if for any $`A \in \mathcal{Q}`$ and any $`B \subseteq \mathcal{P}`$ such that $`A \subseteq B`$, it holds that $`B \in \mathcal{Q}`$. Throughout the paper we assume that $`{\mathcal{F},\mathcal{Q}} \subseteq 2^{\mathcal{P}}`$, where $`{\mathcal{Q} \cap \mathcal{F}} = \varnothing`$, $`{\mathcal{Q} \cup \mathcal{F}} = 2^{\mathcal{P}}`$ and $`\mathcal{Q}`$ is monotone. The members of $`\mathcal{Q}`$ and $`\mathcal{F}`$ are termed qualified sets and forbidden sets, respectively. Denote the minimal qualified sets of $`\mathcal{Q}`$ by $`\mathcal{Q}_{0}`$. Also, we call $`\Gamma = {(\mathcal{P},\mathcal{Q},\mathcal{F})}`$ the access structure of the scheme.
Let $`M = M_{n \times m}`$ be an $`n \times m`$ Boolean matrix. The $`i`$th row vector of M is denoted by $`M_{i}`$. Set $`M_{i} \circ M_{j}`$ to be the bit-wise “OR” of vectors $`M_{i}`$ and $`M_{j}`$. Suppose $`X = {\{ i_{1},i_{2},\ldots,i_{q}\}} \subseteq {\{ 1,2,\ldots,n\}}`$, and define $`M_{_{X}}\overset{def}{=}{M_{i_{1}} \circ M_{i_{2}} \circ \cdots \circ M_{i_{q}}}`$, whereas $`{M\hspace{0pt}{\lbrack X\rbrack}}\overset{def}{=}{M\hspace{0pt}{\lbrack X\rbrack}\hspace{0pt}{\lbrack{\{ 1,\ldots,m\}}\rbrack}}`$ denotes the $`{|X|} \times m`$ matrix obtained from $`M`$ by considering only the rows corresponding to members of $`X`$. Let $`A||B`$ denote the concatenation of two matrices $`A`$ and $`B`$ of the same number of rows. Denote the Hamming weight of row vector $`v`$ by $`w\hspace{0pt}{(v)}`$. For two vectors $`u`$ and $`v`$, denote their inner product by $`u \cdot v`$. Let $`T`$ be a set of vectors. The vector space generated by $`T`$ is denoted by $`s\hspace{0pt}p\hspace{0pt}a\hspace{0pt}n\hspace{0pt}{(T)}`$.
In visual cryptography schemes we assume that the secret image consists of a collection of black and white pixels. Each pixel of secret image appears in $`n`$ modified versions called shares, one for each transparency, and each share is divided into $`m`$ black and white subpixels. The $`m`$ subpixels of shares can be represented by an $`n \times m`$ Boolean matrix $`M = {\lbrack m_{i\hspace{0pt}j}\rbrack}`$ where $`m_{i\hspace{0pt}j} = 1`$ if and only if $`j`$th subpixel in the $`i`$th transparency is black. The resultant shares should meet the properties of visual cryptography. The conventional definition for VCS is as follows.
######
Definition 1
. Let $`\Gamma = {(\mathcal{P},\mathcal{Q},\mathcal{F})}`$ be an access structure where $`{|\mathcal{P}|} = n`$. Two collections (multisets) $`\mathcal{C}^{0}`$ and $`\mathcal{C}^{1}`$ of $`n \times m`$ Boolean matrices constitute a $`(\Gamma,m)`$-$`{VCS}_{1}`$ if there exist a value $`{\alpha\hspace{0pt}{(m)}} > 0`$ and a set $`{\{{(X,t_{X})}\}}_{X \in \mathcal{Q}}`$ satisfying
1. 1\.
Any qualified set $`X = {\{ i_{1},i_{2},\ldots,i_{q}\}} \in \mathcal{Q}`$ can recover the shared image by stacking their transparencies. Formally, for any $`{M \in \mathcal{C}^{0}},{{\omega\hspace{0pt}{(M_{X})}} \leq {t_{X} - {{\alpha\hspace{0pt}{(m)}} \cdot m}}}`$, whereas for any $`M' \in \mathcal{C}^{1}`$, $`{\omega\hspace{0pt}{(M_{X}')}} \geq t_{X}`$.
2. 2\.
Any forbidden set $`X = {\{ i_{1},i_{2},\ldots,i_{q}\}} \in \mathcal{F}`$ has no information on the shared image. Formally, the two collections $`{\mathcal{D}^{t},t} \in {\{ 0,1\}}`$, of $`q \times m`$ matrices obtained by restricting each $`n \times m`$ matrix in $`M \in \mathcal{C}^{t}`$ to rows $`i_{1},i_{2},\ldots,i_{q}`$, are indistinguishable in the sense that they contain the same matrices with the same frequencies.
The value $`m`$ is called pixel expansion, and the value $`\alpha\hspace{0pt}{(m)}`$ is termed contrast. The first and second conditions are called contrast and security, respectively. The notation $`m_{1}\hspace{0pt}{(\Gamma)}`$ stands for the minimum value of $`m`$ for which such a $`\Gamma`$-$`{VCS}_{1}`$ exists and called the best pixel expansion of $`\Gamma`$-$`{VCS}_{1}`$. $`\spadesuit`$
The most of constructions in this paper are based on two $`n \times m`$ matrices, $`S^{0}`$ and $`S^{1}`$ called basis matrices. In this case, the collections $`\mathcal{C}^{0}`$ and $`\mathcal{C}^{1}`$ are generated by permuting the columns of the corresponding basis matrices $`S^{0}`$ and $`S^{1}`$, respectively, in all possible ways.
######
Definition 2
. Let $`\Gamma = {(\mathcal{P},\mathcal{Q},\mathcal{F})}`$ be an access structure where $`{|\mathcal{P}|} = n`$. A $`(\Gamma,m)`$-$`{VCS}_{2}`$ is realized using two basis matrices $`S^{0}`$ and $`S^{1}`$ of $`n \times m`$ Boolean matrices if there exist a value $`{\alpha\hspace{0pt}{(m)}} > 0`$ and a set $`{\{{(X,t_{X})}\}}_{X \in \mathcal{Q}}`$ satisfying
1. 1\.
Any qualified set $`X = {\{ i_{1},i_{2},\ldots,i_{q}\}} \in \mathcal{Q}`$ can recover the shared image by stacking their transparencies. Formally, $`{\omega\hspace{0pt}{(S_{X}^{0})}} \leq {t_{X} - {{\alpha\hspace{0pt}{(m)}} \cdot m}}`$, whereas $`{\omega\hspace{0pt}{(S_{X}^{1})}} \geq t_{X}`$.
2. 2\.
Any forbidden set $`X = {\{ i_{1},i_{2},\ldots,i_{q}\}} \in \mathcal{F}`$ has no information on the shared image. Formally, the two $`q \times m`$ matrices obtained by restricting $`S^{0}`$ and $`S^{1}`$ to rows $`i_{1},i_{2},\ldots,i_{q}`$ are equal up to a column permutation.
The notation $`m_{2}\hspace{0pt}{(\Gamma)}`$ stands for the minimum value of $`m`$ for which such a $`\Gamma`$-$`{VCS}_{2}`$ with basis matrices exists and called the best pixel expansion of $`\Gamma`$-$`{VCS}_{2}`$. $`\spadesuit`$
Now, we recall the definition of visual cryptography scheme defined in \[[13](#bib.bib13)\]. In this scheme, only the sets in $`\mathcal{Q}_{0}`$ can recover the secret image by stacking their transparencies and that the image, which is revealed by stacking the transparencies of a minimal qualified set, can be darker or lighter than the background. Note that any non-minimal qualified set, by stacking their transparencies, has no information on the shared image, i.e., cannot distinguish a white pixel from a black one. Here is the formal definition \[[13](#bib.bib13)\].
######
Definition 3
. Let $`\Gamma = {(\mathcal{P},\mathcal{Q},\mathcal{F})}`$ be an access structure where $`{|\mathcal{P}|} = n`$. A $`(\Gamma,m)`$-$`{VCS}_{3}`$ is realized using two basis matrices $`S^{0}`$ and $`S^{1}`$ of $`n \times m`$ Boolean matrices if there exist a value $`{\alpha\hspace{0pt}{(m)}} > 0`$ and a set $`{\{{(X,t_{X})}\}}_{X \in \mathcal{Q}_{0}}`$ satisfying
1. 1\.
Any minimal qualified set $`X = {\{ i_{1},i_{2},\ldots,i_{q}\}} \in \mathcal{Q}_{0}`$ can recover the shared image by stacking their transparencies. Formally, $`{\omega\hspace{0pt}{(S_{X}^{0})}} = t_{X}`$, whereas either, $`{\omega\hspace{0pt}{(S_{X}^{1})}} \geq {t_{X} + {{\alpha\hspace{0pt}{(m)}} \cdot m}}`$ or, $`{\omega\hspace{0pt}{(S_{X}^{1})}} \leq {t_{X} - {{\alpha\hspace{0pt}{(m)}} \cdot m}}`$.
2. 2\.
Any forbidden set $`X = {\{ i_{1},i_{2},\ldots,i_{q}\}} \in \mathcal{F}`$ has no information on the shared image. Formally, the two $`q \times m`$ matrices obtained by restricting $`S^{0}`$ and $`S^{1}`$ to rows $`i_{1},i_{2},\ldots,i_{q}`$ are equal up to a column permutation.
3. 3\.
Any non-minimal qualified set $`X = {\{ i_{1},i_{2},\ldots,i_{q}\}} \in {\mathcal{Q} \smallsetminus \mathcal{Q}_{0}}`$, by stacking their transparencies, has no information on shared image. Formally, the two $`1 \times m`$ vectors $`V_{0}`$ and $`V_{1}`$, obtained by OR-ing the rows $`i_{1},i_{2},\ldots,i_{q}`$ of the matrix $`S^{0}`$ and $`S^{1}`$, respectively, are indistinguishable in the sense that they have the same Hamming weight.
Also, we will use the notation $`m_{3}\hspace{0pt}{(\Gamma)}`$ to denote the minimum pixel expansion of basis matrices of $`\Gamma`$-$`{VCS}_{3}`$ and called the best pixel expansion of $`{VCS}_{3}`$. $`\spadesuit`$
To achieve the smallest pixel expansion in the different models of visual cryptography schemes, we consider the following definition in which the minimal qualified subsets can recover the shared image by stacking their transparencies. Also, the revealed image can be darker or lighter than the background as well. Moreover, we don’t mind whether non-minimal qualified subsets can obtain the secret.
######
Definition 4
. Let $`\Gamma = {(\mathcal{P},\mathcal{Q},\mathcal{F})}`$ be an access structure where $`{|\mathcal{P}|} = n`$. Two collections (multisets) $`\mathcal{C}^{0}`$ and $`\mathcal{C}^{1}`$ of $`n \times m`$ Boolean matrices constitute a $`(\Gamma,m)`$-$`{VCS}_{4}`$ if there exist a value $`{\alpha\hspace{0pt}{(m)}} > 0`$ and a set $`{\{{(X,t_{X})}\}}_{X \in \mathcal{Q}_{0}}`$ satisfying
1. 1\.
Any minimal qualified set $`X = {\{ i_{1},i_{2},\ldots,i_{q}\}} \in \mathcal{Q}_{0}`$ can recover the shared image by stacking their transparencies. Formally, for any $`M \in \mathcal{C}^{0}`$, $`{\omega\hspace{0pt}{(M_{X})}} = t_{X}`$, whereas for any $`M' \in \mathcal{C}^{1}`$, either, $`{\omega\hspace{0pt}{(M_{X}')}} \geq {t_{X} + {{\alpha\hspace{0pt}{(m)}} \cdot m}}`$ or, $`{\omega\hspace{0pt}{(M_{X}')}} \leq {t_{X} - {{\alpha\hspace{0pt}{(m)}} \cdot m}}`$.
2. 2\.
Any forbidden set $`X = {\{ i_{1},i_{2},\ldots,i_{q}\}} \in \mathcal{F}`$ has no information on the shared image. Formally, the two collections $`{\mathcal{D}^{t},t} \in {\{ 0,1\}}`$, of $`q \times m`$ matrices obtained by restricting each $`n \times m`$ matrix in $`M \in \mathcal{C}^{t}`$ to rows $`i_{1},i_{2},\ldots,i_{q}`$, are indistinguishable in the sense that they contain the same matrices with the same frequencies.
The notation $`m_{4}\hspace{0pt}{(\Gamma)}`$ stands for the minimum value of $`m`$ for which such a $`\Gamma`$-$`{VCS}_{4}`$ exists and called the best pixel expansion of $`\Gamma`$-$`{VCS}_{4}`$. $`\spadesuit`$
Most constructions in this paper are realized using basis matrices; hence, we consider the following definition as well.
######
Definition 5
. Let $`\Gamma = {(\mathcal{P},\mathcal{Q},\mathcal{F})}`$ be an access structure where $`{|\mathcal{P}|} = n`$. A $`(\Gamma,m)`$-$`{VCS}_{5}`$ is realized using two basis matrices $`S^{0}`$ and $`S^{1}`$ of $`n \times m`$ Boolean matrices if there exist a value $`{\alpha\hspace{0pt}{(m)}} > 0`$ and a set $`{\{{(X,t_{X})}\}}_{X \in \mathcal{Q}_{0}}`$ satisfying
1. 1\.
Any minimal qualified set $`X = {\{ i_{1},i_{2},\ldots,i_{q}\}} \in \mathcal{Q}_{0}`$ can recover the shared image by stacking their transparencies. Formally, $`{\omega\hspace{0pt}{(S_{X}^{0})}} = t_{X}`$, whereas either, $`{\omega\hspace{0pt}{(S_{X}^{1})}} \geq {t_{X} + {{\alpha\hspace{0pt}{(m)}} \cdot m}}`$ or, $`{\omega\hspace{0pt}{(S_{X}^{1})}} \leq {t_{X} - {{\alpha\hspace{0pt}{(m)}} \cdot m}}`$.
2. 2\.
Any forbidden set $`X = {\{ i_{1},i_{2},\ldots,i_{q}\}} \in \mathcal{F}`$ has no information on the shared image. Formally, the two $`q \times m`$ matrices obtained by restricting $`S^{0}`$ and $`S^{1}`$ to rows $`i_{1},i_{2},\ldots,i_{q}`$ are equal up to a column permutation.
Also, we will use the notation $`m_{5}\hspace{0pt}{(\Gamma)}`$ to denote the minimum pixel expansion of basis matrices of $`\Gamma`$-$`{VCS}_{5}`$ and called the best pixel expansion of $`{VCS}_{5}`$. $`\spadesuit`$
It is instructive to add some notes on different models we have introduced so far. First, $`{VCS}_{1}`$ has been considered. In fact, $`{VCS}_{1}`$ is the traditional model of VCS which was introduced by M. Naor and A. Shamir \[[12](#bib.bib12)\]. Constructing the families $`\mathcal{C}^{0}`$ and $`\mathcal{C}^{1}`$, mentioned in $`{VCS}_{1}`$, may seem a daunting task. However, it can be more convenient to handle with basis matrices. Hence, basis matrices are used in the most constructions of VCS found in the literature. That is why we consider the models $`{VCS}_{2}`$, $`{VCS}_{3}`$ and $`{VCS}_{5}`$. Finally, we consider $`{VCS}_{4}`$ to achieve the smallest pixel expansion among the different models of visual cryptography schemes.
## 3 Lower Bounds for Pixel Expansion
In this section, we introduce some lower bounds for the best pixel expansion of the different models of visual cryptography schemes. First, for a given access structure $`\Gamma = {(\mathcal{P},\mathcal{Q},\mathcal{F})}`$, we present a lower bound for the best pixel expansion of $`\Gamma`$-$`{VCS}_{2}`$ as follows.
######
Theorem 1
. Let $`\Gamma = {(\mathcal{P},\mathcal{Q},\mathcal{F})}`$ be an access structure and let $`\Omega = {\{ F_{1},F_{2},\ldots,F_{t}\}}`$ be a collection of forbidden sets such that $`{\bigcup\limits_{i = 1}^{t}F_{i}} \in \mathcal{F}`$. Also, assume that for any two disjoint non-empty subsets $`{A,B} \subset \Omega`$, there exists a forbidden set $`F' \in \mathcal{F}`$ such that at least one of the following conditions holds
- •
For any $`F_{i} \in A`$, $`{F_{i} \cup F'} \in \mathcal{F}`$ and there exists an $`F_{j} \in B`$ such that $`{F_{j} \cup F'} \in \mathcal{Q}`$.
- •
For any $`F_{i} \in B`$, $`{F_{i} \cup F'} \in \mathcal{F}`$ and there exists an $`F_{j} \in A`$ such that $`{F_{j} \cup F'} \in \mathcal{Q}`$.
Then we have $`{m_{2}\hspace{0pt}{(\Gamma)}} \geq {t + 1}`$.
######
Proof.
Let $`S^{0}`$ and $`S^{1}`$ be $`{n \times m_{2}}\hspace{0pt}{(\Gamma)}`$ the basis matrices of access structure $`\Gamma = {(\mathcal{P},\mathcal{Q},\mathcal{F})}`$. Consider the following sets
<table>
<tbody>
<tr>
<td></td>
<td>T0=def{SFi0| 1≤i≤t},superscriptdefsuperscript𝑇0conditional-setsubscriptsuperscript𝑆0subscript𝐹𝑖1𝑖𝑡T^{0}\mbox{$\ \stackrel{{\scriptstyle\rm def}}{{=}}\ $}\{S^{0}_{{}_{F_{i}}}\ |\ 1\leq i\leq t\},</td>
<td></td>
</tr>
</tbody>
</table>
<table>
<tbody>
<tr>
<td></td>
<td>T1=def{SFi1| 1≤i≤t}.superscriptdefsuperscript𝑇1conditional-setsubscriptsuperscript𝑆1subscript𝐹𝑖1𝑖𝑡T^{1}\mbox{$\ \stackrel{{\scriptstyle\rm def}}{{=}}\ $}\{S^{1}_{{}_{F_{i}}}\ |\ 1\leq i\leq t\}.</td>
<td></td>
</tr>
</tbody>
</table>
We claim that the vectors of $`T^{0}`$ (resp. $`T^{1}`$) are linearly independent over the real numbers. On the contrary, suppose that there are some real coefficients $`a_{_{F_{i}}}`$ such that $`{\sum\limits_{i = 1}^{t}{a_{_{F_{i}}}\hspace{0pt}S_{_{F_{i}}}^{0}}} = 0`$. Define $`Y = {\bigcup\limits_{i = 1}^{t}F_{i}}`$. Since $`Y \in \mathcal{F}`$, $`S^{0}\hspace{0pt}{\lbrack Y\rbrack}`$ and $`S^{1}\hspace{0pt}{\lbrack Y\rbrack}`$ are equal up to a permutation of columns; hence, $`{\sum\limits_{i = 1}^{t}{a_{_{F_{i}}}\hspace{0pt}S_{_{F_{i}}}^{1}}} = 0`$. Therefore,
<table>
<tbody>
<tr>
<td></td>
<td>∑aFi>0aFiSFi0=−∑aFj<0aFjSFj0&∑aFi>0aFiSFi1=−∑aFj<0aFjSFj1.formulae-sequencesubscriptsubscript𝑎subscript𝐹𝑖0subscript𝑎subscript𝐹𝑖subscriptsuperscript𝑆0subscript𝐹𝑖subscriptsubscript𝑎subscript𝐹𝑗0subscript𝑎subscript𝐹𝑗subscriptsuperscript𝑆0subscript𝐹𝑗subscriptsubscript𝑎subscript𝐹𝑖0subscript𝑎subscript𝐹𝑖subscriptsuperscript𝑆1subscript𝐹𝑖subscriptsubscript𝑎subscript𝐹𝑗0subscript𝑎subscript𝐹𝑗subscriptsuperscript𝑆1subscript𝐹𝑗\displaystyle{\sum_{a_{{}_{F_{i}}}>0}}a_{{}_{F_{i}}}S^{0}_{{}_{F_{i}}}=-\displaystyle{\sum_{a_{{}_{F_{j}}}<0}}a_{{}_{F_{j}}}S^{0}_{{}_{F_{j}}}\quad\&\quad\displaystyle{\sum_{a_{{}_{F_{i}}}>0}}a_{{}_{F_{i}}}S^{1}_{{}_{F_{i}}}=-\displaystyle{\sum_{a_{{}_{F_{j}}}<0}}a_{{}_{F_{j}}}S^{1}_{{}_{F_{j}}}.</td>
<td></td>
</tr>
</tbody>
</table>
Set $`A\overset{def}{=}\left. \{ F_{j} \middle| {a_{_{F_{j}}} < 0}\} \right.`$ and $`B\overset{def}{=}\left. \{ F_{i} \middle| {a_{_{F_{i}}} > 0}\} \right.`$. Without loss of generality, assume that there exists a forbidden set $`F'`$ such that for any $`F_{j} \in A`$, $`{F_{j} \cup F'} \in \mathcal{F}`$ and there exists a member of $`B`$ such as $`F^{\operatorname{\prime\prime}}`$ such that $`{F^{\operatorname{\prime\prime}} \cup F'} \in \mathcal{Q}`$. We have
<table>
<tbody>
<tr>
<td></td>
<td>SF′0⋅∑aFi>0aFiSFi0=SF′0⋅∑aFj<0−aFjSFj0&SF′1⋅∑aFi>0aFiSFi1=SF′1⋅∑aFj<0−aFjSFj1.formulae-sequence⋅subscriptsuperscript𝑆0superscript𝐹′subscriptsubscript𝑎subscript𝐹𝑖0subscript𝑎subscript𝐹𝑖subscriptsuperscript𝑆0subscript𝐹𝑖⋅subscriptsuperscript𝑆0superscript𝐹′subscriptsubscript𝑎subscript𝐹𝑗0subscript𝑎subscript𝐹𝑗subscriptsuperscript𝑆0subscript𝐹𝑗⋅subscriptsuperscript𝑆1superscript𝐹′subscriptsubscript𝑎subscript𝐹𝑖0subscript𝑎subscript𝐹𝑖subscriptsuperscript𝑆1subscript𝐹𝑖⋅subscriptsuperscript𝑆1superscript𝐹′subscriptsubscript𝑎subscript𝐹𝑗0subscript𝑎subscript𝐹𝑗subscriptsuperscript𝑆1subscript𝐹𝑗S^{0}_{{}_{F^{\prime}}}\cdot\displaystyle{\sum_{a_{{}_{F_{i}}}>0}}a_{{}_{F_{i}}}S^{0}_{{}_{F_{i}}}=S^{0}_{{}_{F^{\prime}}}\cdot\displaystyle{\sum_{a_{{}_{F_{j}}}<0}}-a_{{}_{F_{j}}}S^{0}_{{}_{F_{j}}}\ \ \&\ \ S^{1}_{{}_{F^{\prime}}}\cdot\displaystyle{\sum_{a_{{}_{F_{i}}}>0}}a_{{}_{F_{i}}}S^{1}_{{}_{F_{i}}}=S^{1}_{{}_{F^{\prime}}}\cdot\displaystyle{\sum_{a_{{}_{F_{j}}}<0}}-a_{{}_{F_{j}}}S^{1}_{{}_{F_{j}}}.</td>
<td></td>
<td>(1)</td>
</tr>
</tbody>
</table>
Note that $`{w\hspace{0pt}{(S_{_{F'}}^{0})}} = {w\hspace{0pt}{(S_{_{F'}}^{1})}}`$. Also, for any $`F_{j} \in A`$, $`{F' \cup F_{j}} \in \mathcal{F}`$; consequently, $`S^{0}\hspace{0pt}{\lbrack{F_{j} \cup F'}\rbrack}`$ and $`S^{1}\hspace{0pt}{\lbrack{F_{j} \cup F'}\rbrack}`$ are equal up to a permutation of columns. Hence, $`{S_{_{F'}}^{0} \cdot S_{_{F_{j}}}^{0}} = {S_{_{F'}}^{1} \cdot S_{_{F_{j}}}^{1}}`$. Moreover, there exists an $`F^{\operatorname{\prime\prime}} \in B`$ such that $`{F^{\operatorname{\prime\prime}} \cup F'} \in \mathcal{Q}`$. Thus,
<table>
<tbody>
<tr>
<td></td>
<td>SF′0⋅SF′′0>SF′1⋅SF′′1.⋅subscriptsuperscript𝑆0superscript𝐹′subscriptsuperscript𝑆0superscript𝐹′′⋅subscriptsuperscript𝑆1superscript𝐹′subscriptsuperscript𝑆1superscript𝐹′′S^{0}_{{}_{F^{\prime}}}\cdot S^{0}_{{}_{F^{\prime\prime}}}>S^{1}_{{}_{F^{\prime}}}\cdot S^{1}_{{}_{F^{\prime\prime}}}.</td>
<td></td>
</tr>
</tbody>
</table>
In view of Equations [1](#S3.E1 "In Proof. ‣ 3 Lower Bounds for Pixel Expansion"), one can obtain that for any $`F_{i} \in \Omega`$, $`a_{_{F_{i}}} = 0`$; accordingly, $`{\dim{({{span}\hspace{0pt}{(T^{0})}})}} = {\dim{({{span}\hspace{0pt}{(T^{1})}})}} = t`$. In addition, $`{F' \cup Y} \in \mathcal{Q}`$; therefore, $`{w\hspace{0pt}{(S_{_{F' \cup Y}}^{0})}} \neq {w\hspace{0pt}{(S_{_{F' \cup Y}}^{1})}}`$. Also, $`{w\hspace{0pt}{(S_{_{F'}}^{0})}} = {w\hspace{0pt}{(S_{_{F'}}^{1})}}`$ and the matrices $`S^{0}\hspace{0pt}{\lbrack Y\rbrack}`$ and $`S^{1}\hspace{0pt}{\lbrack Y\rbrack}`$ are equal up to a permutation of columns. Now, it is easy to check that $`{\dim{({{span}\hspace{0pt}{({T^{1} \cup {\{ S_{_{F'}}^{1}\}}})}})}} = {t + 1}`$. On the other hand, $`{m_{2}\hspace{0pt}{(\Gamma)}} \geq {\dim{({{span}\hspace{0pt}{({T^{1} \cup {\{ S_{_{F'}}^{1}\}}})}})}}`$. Thus, $`{m_{2}\hspace{0pt}{(\Gamma)}} \geq {t + 1}`$. $`\blacksquare`$
For a given access structure $`\Gamma = {(\mathcal{P},\mathcal{Q},\mathcal{F})}`$, a lower bound for the best pixel expansion of $`\Gamma`$-$`{VCS}_{3}`$ has been introduced in \[[4](#bib.bib4)\] as follows.
######
Theorem A
. \[[4](#bib.bib4)\] Let $`\Gamma = {(\mathcal{P},\mathcal{Q},\mathcal{F})}`$ be an access structure. The best pixel expansion of $`\Gamma`$-$`{VCS}_{3}`$ satisfies
<table>
<tbody>
<tr>
<td></td>
<td>m3(Γ)≥⌈|𝒬0|2⌉.subscript𝑚3Γsubscript𝒬02m_{3}(\Gamma)\geq\lceil{|{\cal Q}_{0}|\over 2}\rceil.</td>
<td></td>
</tr>
</tbody>
</table>
Note that the aforementioned theorem is not effective when $`|\mathcal{Q}_{0}|`$ is small. Now, we present a theorem which can be considered as a counterpart of Theorem [A](#Thmprealphthm1 "Theorem A ‣ 3 Lower Bounds for Pixel Expansion").
######
Theorem 2
. Let $`\Gamma = {(\mathcal{P},\mathcal{Q},\mathcal{F})}`$ be an access structure and let $`\Omega = {\{ F_{1},F_{2},\ldots,F_{t}\}}`$ be a collection of forbidden sets such that $`{\bigcup\limits_{i = 1}^{t}F_{i}} \in \mathcal{F}`$. Also, assume that for any non-empty subset $`A \subset \Omega`$, there exist two forbidden sets $`F' \in A`$ and $`F^{\operatorname{\prime\prime}} \in \mathcal{F}`$ such that $`{F' \cup F^{\operatorname{\prime\prime}}} \in \mathcal{Q}_{0}`$ and for any $`F_{i} \in {A \smallsetminus {\{ F'\}}}`$, $`{F_{i} \cup F^{\operatorname{\prime\prime}}} \notin \mathcal{Q}_{0}`$. Then $`{m_{3}\hspace{0pt}{(\Gamma)}} \geq {t + 1}`$.
######
Proof.
It is simple to prove that $`{m_{3}\hspace{0pt}{(\Gamma)}} \geq 2`$ whenever $`t = 1`$; hence, assume that $`t \geq 2`$. Let $`S^{0}`$ and $`S^{1}`$ be $`{n \times m_{3}}\hspace{0pt}{(\Gamma)}`$ the basis matrices of access structure $`\Gamma = {(\mathcal{P},\mathcal{Q},\mathcal{F})}`$. Set $`F_{t + 1}\overset{def}{=}{F_{1} \cup \cdots \cup F_{t}}`$. Consider the following sets
<table>
<tbody>
<tr>
<td></td>
<td>T0=def{SFi0| 1≤i≤t+1},superscriptdefsuperscript𝑇0conditional-setsubscriptsuperscript𝑆0subscript𝐹𝑖1𝑖𝑡1T^{0}\mbox{$\ \stackrel{{\scriptstyle\rm def}}{{=}}\ $}\{S^{0}_{{}_{F_{i}}}\ |\ 1\leq i\leq t+1\},</td>
<td></td>
</tr>
</tbody>
</table>
<table>
<tbody>
<tr>
<td></td>
<td>T1=def{SFi1| 1≤i≤t+1}.superscriptdefsuperscript𝑇1conditional-setsubscriptsuperscript𝑆1subscript𝐹𝑖1𝑖𝑡1T^{1}\mbox{$\ \stackrel{{\scriptstyle\rm def}}{{=}}\ $}\{S^{1}_{{}_{F_{i}}}\ |\ 1\leq i\leq t+1\}.</td>
<td></td>
</tr>
</tbody>
</table>
We claim that the vectors of $`T^{0}`$ (resp. $`T^{1}`$) are linearly independent over the real numbers. Suppose that there are some real coefficients $`a_{_{F_{i}}}`$ such that $`{\sum\limits_{i = 1}^{t + 1}{a_{_{F_{i}}}\hspace{0pt}S_{_{F_{i}}}^{0}}} = 0`$. Since $`F_{t + 1} \in \mathcal{F}`$, $`S^{0}\hspace{0pt}{\lbrack F_{t + 1}\rbrack}`$ and $`S^{1}\hspace{0pt}{\lbrack F_{t + 1}\rbrack}`$ are equal up to a permutation of columns; consequently, $`{\sum\limits_{i = 1}^{t + 1}{a_{_{F_{i}}}\hspace{0pt}S_{_{F_{i}}}^{1}}} = 0`$. Set $`A\overset{def}{=}\left. \{ F_{i} \middle| {{i \leq t},{a_{_{F_{i}}} \neq 0}}\} \right.`$. If $`A = \varnothing`$, then the assertion follows easily. Hence, assume that $`A \neq \varnothing`$ and there exist two forbidden sets $`F' \in A`$ and $`F^{\operatorname{\prime\prime}} \in \mathcal{F}`$ such that $`{F' \cup F^{\operatorname{\prime\prime}}} \in \mathcal{Q}_{0}`$ and for any $`F_{i} \in {A \smallsetminus {\{ F'\}}}`$, $`{F_{i} \cup F^{\operatorname{\prime\prime}}} \notin \mathcal{Q}_{0}`$. We have
<table>
<tbody>
<tr>
<td></td>
<td>SF′′0⋅∑aFi≠0aFiSFi0=0&SF′′1⋅∑aFi≠0aFiSFi1=0.formulae-sequence⋅subscriptsuperscript𝑆0superscript𝐹′′subscriptsubscript𝑎subscript𝐹𝑖0subscript𝑎subscript𝐹𝑖subscriptsuperscript𝑆0subscript𝐹𝑖0⋅subscriptsuperscript𝑆1superscript𝐹′′subscriptsubscript𝑎subscript𝐹𝑖0subscript𝑎subscript𝐹𝑖subscriptsuperscript𝑆1subscript𝐹𝑖0S^{0}_{{}_{F^{\prime\prime}}}\cdot\displaystyle{\sum_{a_{{}_{F_{i}}}\not=0}}a_{{}_{F_{i}}}S^{0}_{{}_{F_{i}}}=0\ \ \quad\&\ \ \quad S^{1}_{{}_{F^{\prime\prime}}}\cdot\displaystyle{\sum_{a_{{}_{F_{i}}}\not=0}}a_{{}_{F_{i}}}S^{1}_{{}_{F_{i}}}=0.</td>
<td></td>
<td>(2)</td>
</tr>
</tbody>
</table>
Note that $`{w\hspace{0pt}{(S_{_{F^{\operatorname{\prime\prime}}}}^{0})}} = {w\hspace{0pt}{(S_{_{F^{\operatorname{\prime\prime}}}}^{1})}}`$. Also, for any $`F_{i} \in {{A \cup {\{ F_{t + 1}\}}} \smallsetminus {\{ F'\}}}`$, $`{F^{\operatorname{\prime\prime}} \cup F_{i}} \notin \mathcal{Q}_{0}`$; consequently, $`{S_{_{F^{\operatorname{\prime\prime}}}}^{0} \cdot S_{_{F_{i}}}^{0}} = {S_{_{F^{\operatorname{\prime\prime}}}}^{1} \cdot S_{_{F_{i}}}^{1}}`$. Moreover, $`{F^{\operatorname{\prime\prime}} \cup F'} \in \mathcal{Q}_{0}`$. Thus,
<table>
<tbody>
<tr>
<td></td>
<td>SF′′0⋅SF′0≠SF′′1⋅SF′1.⋅subscriptsuperscript𝑆0superscript𝐹′′subscriptsuperscript𝑆0superscript𝐹′⋅subscriptsuperscript𝑆1superscript𝐹′′subscriptsuperscript𝑆1superscript𝐹′S^{0}_{{}_{F^{\prime\prime}}}\cdot S^{0}_{{}_{F^{\prime}}}\not=S^{1}_{{}_{F^{\prime\prime}}}\cdot S^{1}_{{}_{F^{\prime}}}.</td>
<td></td>
</tr>
</tbody>
</table>
In view of Equations [2](#S3.E2 "In Proof. ‣ 3 Lower Bounds for Pixel Expansion"), one can obtain that for any $`F_{i} \in A`$, $`a_{_{F_{i}}} = 0`$; accordingly, $`{\dim{({{span}\hspace{0pt}{(T^{0})}})}} = {\dim{({{span}\hspace{0pt}{(T^{1})}})}} = {t + 1}`$ which implies that $`{m_{3}\hspace{0pt}{(\Gamma)}} \geq {t + 1}`$. $`\blacksquare`$
In the language of hypergraph theory, for a given access structure $`\Gamma = {(\mathcal{P},\mathcal{Q},\mathcal{F})}`$, one can introduce a lower bound for the best pixel expansion of $`\Gamma`$-$`{VCS}_{2}`$ and $`\Gamma`$-$`{VCS}_{3}`$ in terms of an induced matching of the hypergraph $`(\mathcal{P},\mathcal{Q})`$.
######
Theorem 3
. Let $`\Gamma = {(\mathcal{P},\mathcal{Q},\mathcal{F})}`$ be an access structure. Also, assume that there exist disjoint qualified sets $`A_{1},\ldots,A_{t}`$ such that for any qualified set $`B \subseteq {A_{1} \cup \cdots \cup A_{t}}`$ one should have $`A_{i} \subseteq B`$ for some $`1 \leq i \leq t`$. Then
<table>
<tbody>
<tr>
<td></td>
<td>min{m2(Γ),m3(Γ)}≥∑i=1t2|Ai|−1−(t−1).subscript𝑚2Γsubscript𝑚3Γsuperscriptsubscript𝑖1𝑡superscript2subscript𝐴𝑖1𝑡1\min\{m_{2}(\Gamma),m_{3}(\Gamma)\}\geq\displaystyle\sum_{i=1}^{t}2^{|A_{i}|-1}-(t-1).</td>
<td></td>
</tr>
</tbody>
</table>
######
Proof.
Suppose that $`S^{0}`$ and $`S^{1}`$ are $`{n \times m_{3}}\hspace{0pt}{(\Gamma)}`$ basis matrices for access structure $`\Gamma = {(\mathcal{P},\mathcal{Q},\mathcal{F})}`$. Let $`{|A_{i}|} = r_{i}`$ and $`A_{i} = {\{ p_{i\hspace{0pt}1},\ldots,p_{i\hspace{0pt}r_{i}}\}}`$. Define $`Y_{i}\overset{def}{=}{\{ p_{i\hspace{0pt}1},\ldots,p_{i\hspace{0pt}{({r_{i} - 1})}}\}}`$. Consider the non-empty members of power set of $`Y_{i}`$’s. Set $`{\{ F_{i\hspace{0pt}1},\ldots,F_{i\hspace{0pt}{({2^{r_{i} - 1} - 1})}}\}}\overset{def}{=}{2^{Y_{i}} \smallsetminus \varnothing}`$ such that $`{|F_{i\hspace{0pt}1}|} \leq {|F_{i\hspace{0pt}2}|} \leq \cdots \leq {|F_{i\hspace{0pt}{({2^{r_{i} - 1} - 1})}}|}`$. Define
<table>
<tbody>
<tr>
<td></td>
<td>Ω=def{Fij| 1≤i≤t, 1≤j≤2ri−1−1}.superscriptdefΩconditional-setsubscript𝐹𝑖𝑗formulae-sequence1𝑖𝑡1𝑗superscript2subscript𝑟𝑖11\Omega\mbox{$\ \stackrel{{\scriptstyle\rm def}}{{=}}\ $}\{F_{ij}|\ 1\leq i\leq t,\ 1\leq j\leq 2^{r_{{}_{i}}-1}-1\}.</td>
<td></td>
</tr>
</tbody>
</table>
Consider the following ordering for $`\Omega`$,
<table>
<tbody>
<tr>
<td></td>
<td>F11≤⋯≤F1(2r1−1−1)≤F21≤⋯≤F2(2r2−1−1)≤⋯≤Ft(2rt−1−1).subscript𝐹11⋯subscript𝐹1superscript2subscript𝑟111subscript𝐹21⋯subscript𝐹2superscript2subscript𝑟211⋯subscript𝐹𝑡superscript2subscript𝑟𝑡11F_{11}\leq\cdots\leq F_{1(2^{r_{{}_{1}}-1}-1)}\leq F_{21}\leq\cdots\leq F_{2(2^{r_{{}_{2}}-1}-1)}\leq\cdots\leq F_{t(2^{r_{{}_{t}}-1}-1)}.</td>
<td></td>
</tr>
</tbody>
</table>
Assume that $`A`$ is a non-empty subsets of $`\Omega`$. Without loss of generality, suppose that $`F_{r\hspace{0pt}s}`$ is the largest member of $`A`$. Set $`F'\overset{def}{=}{A_{r} \smallsetminus F_{r\hspace{0pt}s}}`$. It is straightforward to check that for any $`F_{i\hspace{0pt}j} \in {A \smallsetminus {\{ F_{r\hspace{0pt}s}\}}}`$ we have $`{F_{i\hspace{0pt}j} \cup F'} \in \mathcal{F}`$, whereas $`{F_{r\hspace{0pt}s} \cup F'} \in \mathcal{Q}_{0}`$. In view of Theorem [2](#Thmprethm2 "Theorem 2 ‣ 3 Lower Bounds for Pixel Expansion"), one can conclude that $`{m_{3}\hspace{0pt}{(\Gamma)}} \geq {{\sum_{i = 1}^{t}2^{{|A_{i}|} - 1}} - {({t - 1})}}`$. Similarly, one can show that $`{m_{2}\hspace{0pt}{(\Gamma)}} \geq {{\sum_{i = 1}^{t}2^{{|A_{i}|} - 1}} - {({t - 1})}}`$, as desired. $`\blacksquare`$
Access structure $`\Gamma = {(\mathcal{P},\mathcal{Q},\mathcal{F})}`$ with $`{|\mathcal{P}|} = k`$ and $`\mathcal{Q} = {\{\mathcal{P}\}}`$ is well-known as $`k`$ out of $`k`$ scheme. Theorem [3](#Thmprethm3 "Theorem 3 ‣ 3 Lower Bounds for Pixel Expansion") presents a simple proof that the pixel expansion of basis matrices of $`k`$ out of $`k`$ scheme is at least $`2^{k - 1}`$.
######
Corollary 1
\[[12](#bib.bib12)\] Let $`\Gamma = {(\mathcal{P},\mathcal{Q},\mathcal{F})}`$ be a $`k`$ out of $`k`$ scheme. Then
<table>
<tbody>
<tr>
<td></td>
<td>min{m2(Γ),m3(Γ)}≥2k−1.subscript𝑚2Γsubscript𝑚3Γsuperscript2𝑘1\min\{m_{2}(\Gamma),m_{3}(\Gamma)\}\geq 2^{k-1}.</td>
<td></td>
</tr>
</tbody>
</table>
Now, we introduce a lower bound for the best pixel expansion of $`\Gamma`$-$`{VCS}_{5}`$. One can deduce the following theorem whose proof is almost identical to that of Theorem [2](#Thmprethm2 "Theorem 2 ‣ 3 Lower Bounds for Pixel Expansion") and the proof is omitted for the sake of brevity.
######
Theorem 4
. Let $`\Gamma = {(\mathcal{P},\mathcal{Q},\mathcal{F})}`$ be an access structure and let $`\Omega = {\{ F_{1},F_{2},\ldots,F_{t}\}}`$ be a collection of forbidden sets such that $`{\bigcup\limits_{i = 1}^{t}F_{i}} \in \mathcal{F}`$. Also, assume that for any non-empty subset $`A \subset \Omega`$, there exist two forbidden sets $`F' \in A`$ and $`F^{\operatorname{\prime\prime}} \in \mathcal{F}`$ such that $`{F' \cup F^{\operatorname{\prime\prime}}} \in \mathcal{Q}_{0}`$ and for any $`F_{i} \in {A \smallsetminus {\{ F'\}}}`$, $`{F_{i} \cup F^{\operatorname{\prime\prime}}} \in \mathcal{F}`$. Then we have $`{m_{5}\hspace{0pt}{(\Gamma)}} \geq t`$.
## 4 Graph Access Structure
In this section, we study access structures based on graphs. To begin, some definitions are given which are used throughout this section. A graph access structure is an access structure for which the set of participants is the vertex set $`V\hspace{0pt}{(G)}`$ of a graph $`G = {({V\hspace{0pt}{(G)}},{E\hspace{0pt}{(G)}})}`$, and the edge set of $`G`$ constitutes the minimal qualified subsets of access structure, i.e., the qualified subsets are precisely those containing an edge of G. From $`G`$, one can define an access structure $`G = {({V\hspace{0pt}{(G)}},\mathcal{Q},\mathcal{F})}`$ where $`\mathcal{Q}_{0} = {E\hspace{0pt}{(G)}}`$.
Throughout the paper the word graph is used for a finite simple graph. A subgraph $`H`$ of a graph $`G`$ is said to be induced if for any pair of vertices $`u`$ and $`v`$ of $`H`$, $`\{ u,v\}`$ is an edge of $`H`$ if and only if $`\{ u,v\}`$ is an edge of $`G`$. Two graphs $`G`$ and $`H`$ are called disjoint if they have no vertex in common. A matching $`M_{n}`$ is a set of $`n`$ disjoint edges, that is, no two edges share a common vertex. A subgraph of $`G`$ whose edge set is non-empty and forms a complete bipartite graph is called a biclique of $`G`$. A biclique cover $`\mathcal{B}`$ of $`G`$ is a collection of bicliques covering $`E\hspace{0pt}{(G)}`$ (every edge of G belongs to at least one biclique of the collection). The biclique covering number of $`G`$, $`b\hspace{0pt}c\hspace{0pt}{(G)}`$, is the fewest number of bicliques among all biclique covers of $`G`$.
A homomorphism $`f:{G\longrightarrow H}`$ from a graph $`G`$ to a graph $`H`$ is a map $`f:{{V\hspace{0pt}{(G)}}\longrightarrow{V\hspace{0pt}{(H)}}}`$ such that $`{\{ u,v\}} \in {E\hspace{0pt}{(G)}}`$ implies $`{\{{f\hspace{0pt}{(u)}},{f\hspace{0pt}{(v)}}\}} \in {E\hspace{0pt}{(H)}}`$. Notation $`{Hom}^{e}\hspace{0pt}{(G,H)}`$ denotes the sets of onto–edges homomorphisms from $`G`$ to $`H`$, for more on graph homomorphism see \[[5](#bib.bib5), [7](#bib.bib7)\].
######
Lemma 1
. Let $`G`$ and $`H`$ be two graphs such that $`{{Hom}^{e}\hspace{0pt}{(G,H)}} \neq \varnothing`$. Then $`{m_{1}\hspace{0pt}{(G)}} \geq {m_{1}\hspace{0pt}{(H)}}`$ and $`{m_{2}\hspace{0pt}{(G)}} \geq {m_{2}\hspace{0pt}{(H)}}`$.
######
Proof.
First, we show that $`{m_{1}\hspace{0pt}{(G)}} \geq {m_{1}\hspace{0pt}{(H)}}`$. Without loss of generality, suppose that $`H`$ does not have any isolated vertex. Assume that $`{V\hspace{0pt}{(G)}} = {\{ 1,\ldots,n\}}`$, $`{V\hspace{0pt}{(H)}} = {\{ 1,\ldots,n'\}}`$, and $`\sigma \in {{Hom}^{e}\hspace{0pt}{(G,H)}}`$. Also, let $`\mathcal{C}^{0}`$ and $`\mathcal{C}^{1}`$ be two collections (multisets) of $`{n \times m_{1}}\hspace{0pt}{(G)}`$ Boolean matrices constitute a $`(G,{m_{1}\hspace{0pt}{(G)}})`$-$`{VCS}_{1}`$. For any $`{n \times m_{1}}\hspace{0pt}{(G)}`$ matrix $`M`$, define $`{n' \times m_{1}}\hspace{0pt}{(G)}`$ matrix $`M_{\sigma}`$ as follows. For any $`1 \leq i \leq n'`$, the $`i`$th row of $`M_{\sigma}`$ is the vector $`M_{\sigma^{- 1}\hspace{0pt}{(i)}}`$; i.e., the vector obtained by considering the bit-wise “OR” of the vectors corresponding to participants in $`\sigma^{- 1}\hspace{0pt}{(i)}`$. Now, construct two collections (multisets) of $`{n' \times m_{1}}\hspace{0pt}{(G)}`$ Boolean matrices $`\mathcal{D}^{0}`$ and $`\mathcal{D}^{1}`$ as follows.
<table>
<tbody>
<tr>
<td></td>
<td>𝒟0=def{Mσ|M∈𝒞0}&𝒟1=def{Nσ|N∈𝒞1}.formulae-sequencesuperscriptdefsuperscript𝒟0conditional-setsubscript𝑀𝜎𝑀superscript𝒞0superscriptdefsuperscript𝒟1conditional-setsubscript𝑁𝜎𝑁superscript𝒞1{\cal D}^{0}\mbox{$\ \stackrel{{\scriptstyle\rm def}}{{=}}\ $}\{M_{\sigma}|\ M\in{\cal C}^{0}\}\ \ \quad\&\ \ \quad{\cal D}^{1}\mbox{$\ \stackrel{{\scriptstyle\rm def}}{{=}}\ $}\{N_{\sigma}|\ N\in{\cal C}^{1}\}.</td>
<td></td>
</tr>
</tbody>
</table>
It is easy to check that $`\mathcal{D}^{0}`$ and $`\mathcal{D}^{1}`$ constitute an $`(H,{m_{1}\hspace{0pt}{(G)}})`$-$`{VCS}_{1}`$. Similarly, one can show that $`{m_{2}\hspace{0pt}{(G)}} \geq {m_{2}\hspace{0pt}{(H)}}`$, as claimed. $`\blacksquare`$
Now, we provide an upper bound for $`m_{4}\hspace{0pt}{(G)}`$. First, we specify the exact value of $`m_{4}\hspace{0pt}{(M_{n})}`$ as follows.
######
Lemma 2
. Let $`G`$ be a graph such that each connected component of $`G`$ is a biclique or an isolated vertex. Then $`{m_{4}\hspace{0pt}{(G)}} = 2`$.
######
Proof.
First, we prove the assertion when $`G`$ is a matching. Let $`M_{n}`$ be a matching with $`n`$ edges where $`{V\hspace{0pt}{(M_{n})}} = {\{ 1,2,\ldots,{2\hspace{0pt}n}\}}`$ and $`{E\hspace{0pt}{(M_{n})}} = \left. \{{\{{{2\hspace{0pt}i} - 1},{2\hspace{0pt}i}\}} \middle| {1 \leq i \leq n}\} \right.`$. Set
<table>
<tbody>
<tr>
<td></td>
<td>𝒟0={(1100),(0011)}superscript𝒟011000011{\cal D}^{0}=\{\left({\begin{array}[]{cc}1&1\\ 0&0\end{array}}\right),\ \left({\begin{array}[]{cc}0&0\\ 1&1\end{array}}\right)\}</td>
<td></td>
</tr>
</tbody>
</table>
and
<table>
<tbody>
<tr>
<td></td>
<td>𝒟1={(1001),(0110)}.superscript𝒟110010110{\cal D}^{1}=\{\left({\begin{array}[]{cc}1&0\\ 0&1\end{array}}\right),\ \left({\begin{array}[]{cc}0&1\\ 1&0\end{array}}\right)\}.</td>
<td></td>
</tr>
</tbody>
</table>
For $`i = {0,1}`$, define
<table>
<tbody>
<tr>
<td></td>
<td>𝒞i=def{(A1i‖⋯‖Ani)T|Aji∈𝒟i, 1≤j≤n},superscriptdefsuperscript𝒞𝑖formulae-sequenceconditionalsuperscriptsubscriptsuperscript𝐴𝑖1norm⋯subscriptsuperscript𝐴𝑖𝑛𝑇subscriptsuperscript𝐴𝑖𝑗superscript𝒟𝑖1𝑗𝑛{\cal C}^{i}\mbox{$\ \stackrel{{\scriptstyle\rm def}}{{=}}\ $}\{(A^{i}_{1}||\cdots||A^{i}_{n})^{T}|\ A^{i}_{j}\in{\cal D}^{i},\ 1\leq j\leq n\},</td>
<td></td>
</tr>
</tbody>
</table>
where $`{({A_{1}^{i}\hspace{0pt}{\|\cdots\|}\hspace{0pt}A_{n}^{i}})}^{T}`$ means the transpose of the matrix $`A_{1}^{i}\hspace{0pt}{\|\cdots\|}\hspace{0pt}A_{n}^{i}`$. Now, one can check that $`\mathcal{C}^{0}`$ and $`\mathcal{C}^{1}`$ constitute an $`(M_{n},2)`$-$`{VCS}_{4}`$; that is, $`{m_{4}\hspace{0pt}{(M_{n})}} = 2`$. Note that adding isolated vertices does not alter the pixel expansion of $`G`$. Since if $`G`$ has isolated vertices, then one can add a zero row to any matrix of $`\mathcal{C}^{0}`$ and $`\mathcal{C}^{1}`$ corresponding to any isolated vertex. It is readily seen that the new collections of matrices constitute a $`(G,2)`$-$`{VCS}_{4}`$. Similarly, if each connected components of $`G`$ is a biclique or an isolated vertex, then one can show that $`{m_{4}\hspace{0pt}{(G)}} = 2`$. $`\blacksquare`$
A strong edge coloring of a graph $`G`$ is an edge coloring in which every color class is an induced matching; that is, any two vertices belonging to distinct edges with the same color are not adjacent. The strong chromatic index $`s'\hspace{0pt}{(G)}`$ is the minimum number of colors in a strong edge coloring of $`G`$. It is well-known that $`{s'\hspace{0pt}{(G)}} \leq {2\hspace{0pt}\Delta\hspace{0pt}{(G)}^{2}}`$, see \[[11](#bib.bib11)\]. Now, we present an upper bound for $`m_{4}\hspace{0pt}{(G)}`$ in terms of strong chromatic index and biclique covering number.
######
Theorem 5
. Let $`G`$ be a non-empty graph. Then $`{m_{4}\hspace{0pt}{(G)}} \leq {\min{\{{2\hspace{0pt}b\hspace{0pt}c\hspace{0pt}{(G)}},{2\hspace{0pt}s'\hspace{0pt}{(G)}}\}}}`$.
######
Proof.
It is well-known that $`{m_{4}\hspace{0pt}{(G)}} \leq {m_{1}\hspace{0pt}{(G)}} \leq {2\hspace{0pt}b\hspace{0pt}c\hspace{0pt}{(G)}}`$, see \[[1](#bib.bib1)\]. Consider a strong edge coloring with $`s'\hspace{0pt}{(G)}`$ colors. Let $`I_{1},\ldots,I_{s'\hspace{0pt}{(G)}}`$ be the color classes of the strong edge coloring. For any $`1 \leq i \leq {s'\hspace{0pt}{(G)}}`$, one can extend any $`I_{i}`$ to a spanning subgraph of $`G`$, say $`J_{i}`$, such that $`I_{i}`$ is an induced subgraph of $`J_{i}`$ and $`{{E\hspace{0pt}{(J_{i})}} \smallsetminus {E\hspace{0pt}{(I_{i})}}} = \varnothing`$. In view of Lemma [2](#Thmprelem2 "Lemma 2 ‣ 4 Graph Access Structure"), for any $`1 \leq i \leq {s'\hspace{0pt}{(G)}}`$, there exist two collections $`\mathcal{C}_{i}^{0}`$ and $`\mathcal{C}_{i}^{1}`$ of matrices which constitute a $`(J_{i},2)`$-$`{VCS}_{4}`$. For $`i = {0,1}`$, set
<table>
<tbody>
<tr>
<td></td>
<td>𝒟i=def{A1i||⋯||As′(G)i|Aji∈𝒞ji, 1≤j≤s′(G)}.superscriptdefsuperscript𝒟𝑖conditional-setsubscriptsuperscript𝐴𝑖1formulae-sequence⋯subscriptsuperscript𝐴𝑖superscript𝑠′𝐺subscriptsuperscript𝐴𝑖𝑗subscriptsuperscript𝒞𝑖𝑗1𝑗superscript𝑠′𝐺{\cal D}^{i}\mbox{$\ \stackrel{{\scriptstyle\rm def}}{{=}}\ $}\{A^{i}_{1}||\cdots||A^{i}_{s^{\prime}(G)}|\ A^{i}_{j}\in{\cal C}^{i}_{j},\ 1\leq j\leq s^{\prime}(G)\}.</td>
<td></td>
</tr>
</tbody>
</table>
It is easy to see that $`\mathcal{D}^{0}`$ and $`\mathcal{D}^{1}`$ constitute a $`(G,{2\hspace{0pt}s'\hspace{0pt}{(G)}})`$-$`{VCS}_{4}`$. $`\blacksquare`$
A $`t`$-strong biclique covering of a graph $`G`$ is an edge covering, $`{E\hspace{0pt}{(G)}} = {\cup_{i = 1}^{t}{E\hspace{0pt}{(H_{i})}}}`$, where each $`H_{i}`$ is a set of disjoint bicliques, say $`H_{i\hspace{0pt}1},\ldots,H_{i\hspace{0pt}r_{i}}`$, such that the graph $`G`$ has no edges between $`H_{i\hspace{0pt}k}`$ and $`H_{i\hspace{0pt}j}`$ for any $`1 \leq j < k \leq r_{i}`$. The strong biclique covering number $`s\hspace{0pt}{(G)}`$ is the minimum number $`t`$ for which there exists a $`t`$-strong biclique covering of $`G`$. It is easy to verify that $`{s\hspace{0pt}{(G)}} \leq {\min{\{{b\hspace{0pt}c\hspace{0pt}{(G)}},{s'\hspace{0pt}{(G)}}\}}}`$. The proof of the next theorem is identical to that of Theorem [5](#Thmprethm5 "Theorem 5 ‣ 4 Graph Access Structure") and the proof is omitted for the sake of brevity. Here is a generalization of Theorem [5](#Thmprethm5 "Theorem 5 ‣ 4 Graph Access Structure").
######
Theorem 6
. Let $`G`$ be a non-empty graph. Then we have $`{m_{4}\hspace{0pt}{(G)}} \leq {2\hspace{0pt}s\hspace{0pt}{(G)}}`$.
Suppose that $`P_{_{4}}`$ is a path with the vertex set $`\{ v_{_{1}},v_{_{2}},v_{_{3}},v_{_{4}}\}`$ and the edge set $`\{{\{ v_{_{1}},v_{_{2}}\}},{\{ v_{_{2}},v_{_{3}}\}},{\{ v_{_{3}},v_{_{4}}\}}\}`$. Set $`F{{}_{1}^{}\overset{def}{=}}{\{ v_{_{1}}\}}`$ and $`F{{}_{2}^{}\overset{def}{=}}{\{ v_{_{3}}\}}`$. It is easy to see that $`F_{1}`$ and $`F_{2}`$ satisfy Theorem [1](#Thmprethm1 "Theorem 1 ‣ 3 Lower Bounds for Pixel Expansion"); consequently, $`{m_{2}\hspace{0pt}{(P_{_{4}})}} \geq 3`$. Furthermore, it is easy to check that
<table>
<tbody>
<tr>
<td></td>
<td>S0=(010011011001)andS1=(010101011100)formulae-sequencesuperscript𝑆0010011011001andsuperscript𝑆1010101011100S^{0}=\left({\begin{array}[]{ccc}0&1&0\\ 0&1&1\\ 0&1&1\\ 0&0&1\end{array}}\right)\quad\quad{\rm and}\quad\quad S^{1}=\left({\begin{array}[]{ccc}0&1&0\\ 1&0&1\\ 0&1&1\\ 1&0&0\end{array}}\right)</td>
<td></td>
</tr>
</tbody>
</table>
are the basis matrices of $`(P_{_{4}},3)`$-$`{VCS}_{2}`$. Thus, $`{m_{2}\hspace{0pt}{(P_{_{4}})}} = 3`$ which implies that the lower bound mentioned in Theorem [1](#Thmprethm1 "Theorem 1 ‣ 3 Lower Bounds for Pixel Expansion") is sharp.
The following corollary is a special case of Theorem [3](#Thmprethm3 "Theorem 3 ‣ 3 Lower Bounds for Pixel Expansion").
######
Corollary 2
Let $`G`$ be a graph access structure and $`e_{1},\ldots,e_{t}`$ be an induced matching of $`G`$. Then we have
<table>
<tbody>
<tr>
<td></td>
<td>min{m2(G),m3(G)}≥t+1.subscript𝑚2𝐺subscript𝑚3𝐺𝑡1\min\{m_{2}(G),m_{3}(G)\}\geq t+1.</td>
<td></td>
</tr>
</tbody>
</table>
Now, we show that $`{m_{3}\hspace{0pt}{(M_{2})}} = 3`$. Consider the following matrices
<table>
<tbody>
<tr>
<td></td>
<td>S0=(101011110110)andS1=(101101110011).formulae-sequencesuperscript𝑆0101011110110andsuperscript𝑆1101101110011S^{0}=\left({\begin{array}[]{ccc}1&0&1\\ 0&1&1\\ 1&1&0\\ 1&1&0\end{array}}\right)\quad\quad{\rm and}\quad\quad S^{1}=\left({\begin{array}[]{ccc}1&0&1\\ 1&0&1\\ 1&1&0\\ 0&1&1\end{array}}\right).</td>
<td></td>
</tr>
</tbody>
</table>
One can check that $`S^{0}`$ and $`S^{1}`$ are the basis matrices of $`(M_{2},3)`$-$`{VCS}_{3}`$ which implies that the lower bound mentioned in Corollary [2](#Thmprecor2 "Corollary 2 ‣ 4 Graph Access Structure") is sharp.
Acknowledgement: The authors wish to thank anonymous referees for their invaluable comments.
## References
- \[1\]
G. Ateniese, C. Blundo, A. De Santis, and D.R. Stinson, Visual cryptography for general access structures, Information and Computation, 129 (1996), 86–106.
- \[2\]
C. Blundo, A. De Santis, and D.R. Stinson, On the contrast in visual cryptography schemes, Journal of Cryptology, 12 (1999), 261–289.
- \[3\]
C. Blundo, P. D’Arco, A. De Santis, and D.R. Stinson, Contrast optimal threshold visual cryptography schemes, SIAM Journal Discrete Mathematics, 16 (2003), 224–261.
- \[4\]
C. Blundo, S. Cimato, and A. De Santis, Visual Cryptography Schemes with Optimal Pixel Expansion, Theoretical Computer Science, 369 (2006), 169–182.
- \[5\]
A. Daneshgar and H. Hajiabolhassan, Graph homomorphims through random walks, J. Graph Theory, 44 (2003), 15–38.
- \[6\]
S. Droste, New results on visual cryptography, In Proceedings of Advances in Cryptology–CRYPTO 96, LNCS, Springer–Verlag, 1109 (1996), 401 -415.
- \[7\]
P. Hell and J. Nešetřil, Graphs and Homomorphisms, Oxford Lecture Series in Mathematics and its Applications, 28, Oxford University press, Oxford (2004).
- \[8\]
T. Hofmeister, M. Krause, and H.U. Simon, Contrast-optimal $`k`$ out of $`n`$ secret sharing schemes in visual cryptography, Computing and combinatorics (Shanghai, 1997), Theoretical Computer Science, 240 (2000), 471–485.
- \[9\]
M. Krause and H. Ulrich Simon, Determining the optimal contrast for secret sharing schemes in visual cryptography, Combinatorics, Probability and Computing, 12 (2003), 285–299.
- \[10\]
C. Kuhlmann and H. Ulrich Simon, Construction of visual secret sharing schemes with almost optimal contrast, Proceedings of the 11th Annual ACM-SIAM Symposium on Discrete Algorithms, (2000), 262–272.
- \[11\]
M. Molloy and B. Reed, A bound on the strong chromatic index of a graph, J. Combin. Theory Ser. B, 69 (1997), 103–109.
- \[12\]
M. Naor and A. Shamir, Visual cryptography, in: Advances in Cryptology–EUROCRYPT 94, Lecture Notes in Computer Science, 950, Springer, Berlin, (1995), 197–202.
- \[13\]
W.- G. Tzeng and C.- M. Hu, A New Approach for Visual Cryptography, Designs, Codes and Cryptography, 27 (2002), 207 -227.
- \[14\]
E.R. Verheul and H.C.A. Van Tilborg, Constructions and properties of k out of n visual secret sharing schemes, Designs, Codes and Cryptography, 11 (1997), 179 -196. |
0710.0658 | cs.NI | cyber | # Detailed Network Measurements Using Sparse Graph Counters: The Theory
Yi Lu, Andrea Montanari and Balaji Prabhakar
Yi Lu is with the Department of Electrical Engineering, Stanford University, yi.lu@stanford.edu. Andrea Montanari is with Departments of Electrical Engineering and Statistics, Stanford University, montanari@stanford.edu. Balaji Prabhakar is with the Department of Electrical Engineering, Stanford University, balaji@stanford.edu.
###### Abstract
Measuring network flow sizes is important for tasks like accounting/billing, network forensics and security. Per-flow accounting is considered hard because it requires that many counters be updated at a very high speed; however, the large fast memories needed for storing the counters are prohibitively expensive. Therefore, current approaches aim to obtain approximate flow counts; that is, to detect large elephant flows and then measure their sizes.
Recently the authors and their collaborators have developed \[[1](#bib.bib1)\] a novel method for per-flow traffic measurement that is fast, highly memory efficient and accurate. At the core of this method is a novel counter architecture called “counter braids.” In this paper, we analyze the performance of the counter braid architecture under a Maximum Likelihood (ML) flow size estimation algorithm and show that it is optimal; that is, the number of bits needed to store the size of a flow matches the entropy lower bound. While the ML algorithm is optimal, it is too complex to implement. In \[[1](#bib.bib1)\] we have developed an easy-to-implement and efficient message passing algorithm for estimating flow sizes.
## I Introduction
This paper addresses a theoretical problem arising in a novel approach to network traffic measurement the authors and their collaborators have recently developed. We refer the reader to \[[1](#bib.bib1)\] for technological background, motivation, related literature and other details. In order to keep this paper self-contained, we summarize the background and restrict the literature survey to what is relevant for the results of this paper.
Background. Measuring the sizes of network flows on high speed links is known to be a technologically challenging problem \[[2](#bib.bib2)\]. The nature of the data to be measured is as follows: At any given time several 10s or 100s of thousands of flows can be active on core Internet links. Packets arrive at the rate of one in every 40-50 nanoseconds on these links which currently run at 10 Gbps. Finally, flow size distributions are heavy-tailed, giving rise to the well-known decomposition of flows into a large number of short “mice” and a few large “elephants.” As a rule of thumb, network traffic follows an “80-20 rule”: 80% of the flows are small, and the remaining 20% of the large flows bring about 80% of the packets or bytes.
This implies that measuring flow sizes accurately requires a large array of counters which can be updated at very high speeds, and a good counter management algorithm for updating counts, installing new counters when flows initiate and uninstalling them when flows terminate.
Since high-speed large memories are either too expensive or simply infeasible with the current technology, the bulk of research on traffic measurement has focused on approximate counting methods. These approaches aim at detecting elephant flows and measuring their sizes.
Counter braids. In \[[1](#bib.bib1)\] we develop a novel counter architecture, called “counter braids”, which is fast, very efficient with memory usage and gives an accurate measurement of all flow sizes, not just the elephants. We will briefly review this architecture using the following simple example.
Suppose we are given 5 numbers and are told that four of them are no more than 2 bits long while the fifth can be 8 bits long. We are not told which is which!
Figures [1](#S1.F1 "Figure 1 ‣ I Introduction ‣ Detailed Network Measurements Using Sparse Graph Counters: The Theory") and [2](#S1.F2 "Figure 2 ‣ I Introduction ‣ Detailed Network Measurements Using Sparse Graph Counters: The Theory") present two approaches for storing the values of the 5 numbers. The first one corresponds to a traditional array of counters, whereby the same number of memory registers is allocated to each measured variable (flow). The structure in Fig. [2](#S1.F2 "Figure 2 ‣ I Introduction ‣ Detailed Network Measurements Using Sparse Graph Counters: The Theory") is more efficient in memory, but retrieving the count values is less straightforward, requiring a flow size estimation algorithm.
[Refer to caption](/html/0710.0658/assets/x1.png)
[img.svg](data:image...)
Figure 1: A simple counter structure: to each flow size we associate its binary representation (filled circle $`= 1`$, empty circle $`= 0`$).
[Refer to caption](/html/0710.0658/assets/x2.png)
[img.svg](data:image...)
Figure 2: Counters braid.
Viewed from an information-theoretic perspective, the design of an efficient counting scheme and a good flow size estimation is equivalent to the design of an efficient source code \[[3](#bib.bib3)\]. However, the applications we consider impose a stringent constraint on such a code: each time the size of a flow changes (because a new packet arrives) a small number of operations must be sufficient to update the stored information. This is not the case with standard source codes, where changing a single letter in the source stream may alter completely the compressed version.
In this paper we prove that, under a probabilistic model for the flow sizes (namely that they form a vector of iid random variables), counter braids achieve a compression rate equal to the entropy of the flow sizes distribution, in the large system limit. Namely, for any rate larger than the flow entropy, the flow sizes can be recovered from the counter values, with error probability vanishing in the large system limit. Further, we prove optimal compression can be achieved by using braids that are *sparse*. The result is non-obvious, since counter braids form a pretty restrictive family of architectures.
Our treatment makes use of techniques from the theory of low-density parity check codes, and the whole construction is inspired by that of LDPC \[[4](#bib.bib4), [5](#bib.bib5)\]. The construction of LDPC codes has an analogy in the source coding problem thanks to standard equivalence between coding over discrete memoryless symmetric channels, and compressing iid discrete random variables \[[6](#bib.bib6)\]. However, the key ideas in the present paper have been developed to deal with the problem that the flow sizes are *a priori* unbounded. In the channel coding language, this would be equivalent to use a countable but infinite input alphabet.
Finally, we insist on using sparse braids for two reasons. First, this allows the stored values to be updated with a *small* (typically bounded) number of operations. Second, it is easy to realize that ML decoding of counter braids is NP-hard, since it has ML decoding of linear codes as a special case \[[7](#bib.bib7)\]. However, thanks to the sparseness of the underlying graph, one may use iterative message passing techniques \[[8](#bib.bib8)\]. Indeed, a simple message passing algorithm for estimating flow sizes is described and analyzed using real and synthetic network traces in \[[1](#bib.bib1)\].
## II Counter Braids: Basic definitions
######
Definition 1
.
A *counter braid* is a couple $`(G,q)`$ where $`q \geq 2`$ is an integer (register capacity) and $`G`$ is a directed acyclic graph on vertex sets $`I`$ (input nodes) and $`R`$ (registers), with the input nodes having in-degree zero. We write $`G = {(I,R,E)}`$, with $`E`$ the set of directed edges in $`G`$.
For any node $`i \in {I \cup R}`$, we will denote by $`{\partial_{+}i} \equiv {\{ j:{{(i,j)} \in E}\}}`$ the set of descendants of $`i`$, and by $`{\partial_{-}i} \equiv {\{ j:{{(j,i)} \in E}\}}`$ the set of parents of $`i`$. Finally $`{\partial i} \equiv {\partial_{+} \cup {\partial_{-}i}}`$.
In the following we shall often omit the explicit reference to the register capacity and write $`G`$ for $`(G,q)`$. The input size of the braid is $`{|I|} \equiv n`$, and its storage size $`{|R|} \equiv m`$. An important parameter is its rate, which we measure in bits
<table>
<tbody>
<tr>
<td></td>
<td>r=|R|log2q|I|.𝑟𝑅subscript2𝑞𝐼\displaystyle r=\frac{|R|\log_{2}q}{|I|}\,.</td>
<td></td>
<td>(1)</td>
</tr>
</tbody>
</table>
We will say that a sequence of counters braids $`\{{G_{n} = {(I_{n},R_{n},E_{n})}}\}`$ is sparse if the number of edges per input node $`{|E_{n}|}/{|I_{n}|}`$ is bounded.
######
Definition 2
.
A *state* (or *configuration* of the counter braid $`G_{q}`$, with is an assignment $`(x,y)`$ of non-negative integers to the nodes in $`G`$, with $`x = {\{ x_{i}:{i \in I}\}} \in {\mathbb{N}}^{I}`$, and $`y = {\{ y_{j}:{j \in R}\}} \in {\mathbb{N}}^{R}`$. The state $`(x,y)`$ is *valid* if $`y_{j} \in {\{ 0,\ldots,{q - 1}\}}`$ for any register $`j \in R`$.
Notice that a valid register configuration can be regarded as an element of $`{({\mathbb{Z}}_{q})}^{R}`$ (where $`{\mathbb{Z}}_{q}`$ is the group of integers modulo $`q`$.) We denote by $`0`$ the zero vector in $`{\mathbb{N}}^{K}`$.
We want now to describe the braid behavior when one of the input nodes is incremented by one unity (i.e. when a packet arrives at input node $`i \in I`$.) Assume the braid $`(G,q)`$ to be in a valid state $`(x,y)`$. Given $`i \in I`$, we define the new state $`{(x',y')} = {\mathsf{T}_{i}\hspace{0pt}{(x,y)}}`$ by letting $`x_{i}' = {x_{i} + 1}`$, $`x_{j}' = x_{j}`$ for any $`j \neq i`$, and $`y'`$ be defined by the following procedure.
<table>
<thead>
<tr>
<th colspan="2">REGISTERS UPDATE (input: flow index i𝑖i)</th>
</tr>
</thead>
<tbody>
<tr>
<th>1:</th>
<td>yj(0)=yjsubscript𝑦𝑗0subscript𝑦𝑗y_{j}(0)=y_{j} for j∉∂+i𝑗subscript𝑖j\not\in{\partial_{+}}i,</td>
</tr>
<tr>
<th></th>
<td>and yj(0)=yj+1subscript𝑦𝑗0subscript𝑦𝑗1y_{j}(0)=y_{j}+1 otherwise.</td>
</tr>
<tr>
<th>2:</th>
<td>Set t=0𝑡0t=0.</td>
</tr>
<tr>
<th>3:</th>
<td>while y(t)𝑦𝑡y(t) is not valid</td>
</tr>
<tr>
<th>4:</th>
<td> Let j∈R𝑗𝑅j\in R be such that yj(t)≥qsubscript𝑦𝑗𝑡𝑞y_{j}(t)\geq q;</td>
</tr>
<tr>
<th>5:</th>
<td> Set yj(t+1)=yj(t)−qsubscript𝑦𝑗𝑡1subscript𝑦𝑗𝑡𝑞y_{j}(t+1)=y_{j}(t)-q;</td>
</tr>
<tr>
<th>6:</th>
<td> For any l∈∂+j𝑙subscript𝑗l\in{\partial_{+}}j, set yl(t+1)=yl(t)+1subscript𝑦𝑙𝑡1subscript𝑦𝑙𝑡1y_{l}(t+1)=y_{l}(t)+1;</td>
</tr>
<tr>
<th>7:</th>
<td> For any l∈R∖{j,∂+j}𝑙𝑅𝑗subscript𝑗l\in R\setminus\{j,{\partial_{+}}j\}, set yl(t+1)=yl(t)subscript𝑦𝑙𝑡1subscript𝑦𝑙𝑡y_{l}(t+1)=y_{l}(t);</td>
</tr>
<tr>
<th>8:</th>
<td> Increment t:=t+1assign𝑡𝑡1t:=t+1;</td>
</tr>
<tr>
<th>9:</th>
<td>end</td>
</tr>
<tr>
<th>10:</th>
<td>return y(t)𝑦𝑡y(t).</td>
</tr>
</tbody>
</table>
Notice that this definition is ambiguous in that we did not specify which register to pick among the ones with $`{y_{j}\hspace{0pt}{(t)}} \geq q`$ at step 4 in the registers update routine. However this is not necessary, as stated in the following lemma (the proof is omitted from this extended abstract).
######
Lemma 1
.
The update procedure above halts after a finite number of steps. Further its output $`\mathsf{T}_{i}\hspace{0pt}{(x,y)}`$ does not depend on the order of update of the registers.
With an abuse of notation we shall write $`x' = {\mathsf{T}_{i}\hspace{0pt}{(x)}}`$, $`y' = {\mathsf{T}_{i}\hspace{0pt}{(y)}}`$, when $`{(x',y')} = {\mathsf{T}_{i}\hspace{0pt}{(x,y)}}`$.
When input values $`x`$ are incremented sequentially, the stored information $`y`$ is updated according to the above procedure. From now on we shall take a static view and assume a certain input $`x`$. The corresponding stored information $`y`$ is obtained through the mapping defined below.
######
Definition 3
.
Given a counter braid $`(G,q)`$, the associated *storage function* $`\mathsf{F}_{G}:{{\mathbb{N}}^{I}\rightarrow{\mathbb{Z}}_{q}^{R}}`$ returns, for any input configuration $`x \in {\mathbb{N}}^{I}`$ a register configuration $`y = {\mathsf{F}_{G}\hspace{0pt}{(x)}} \in {\mathbb{Z}}_{q}^{R}`$ defined as follows. Let $`x^{(0)} = 0`$, $`x^{(1)},\ldots`$, $`x^{(N)} = x`$ be a sequence of input configurations such that $`x^{({s + 1})}`$ is obtained from $`x^{(s)}`$ by incrementing its entry $`i\hspace{0pt}{(s)}`$. Then
<table>
<tbody>
<tr>
<td></td>
<td>𝖥G(x)≡𝖳i(N)∘𝖳i(N−1)∘⋯∘𝖳i(1)(0).subscript𝖥𝐺𝑥subscript𝖳𝑖𝑁subscript𝖳𝑖𝑁1⋯subscript𝖳𝑖10\displaystyle{\sf F}_{G}(x)\equiv{\sf T}_{i(N)}\circ{\sf T}_{i(N-1)}\circ\cdots\circ{\sf T}_{i(1)}(0)\,.</td>
<td></td>
<td>(2)</td>
</tr>
</tbody>
</table>
We shall drop the subscript $`G`$ from $`\mathsf{F}_{G}`$ whenever clear from the context. A priori it is not obvious that the mapping $`\mathsf{F}_{G}`$ is well defined. In particular, it is not obvious that it does not depend on the order in which input values are incremented, i.e. on the sequence $`\{{i\hspace{0pt}{(1)}},\ldots,{i\hspace{0pt}{(N)}}\}`$. This is nevertheless the case (the proof is omitted.)
######
Definition 4
.
Given a counter braid $`(G,q)`$, a *reconstruction* (or *decoding*) *function* is a function $`\hat{\mathsf{F}}:{{\mathbb{Z}}_{q}^{R}\rightarrow{\mathbb{N}}}`$.
###
II-A
Main results
Throughout this paper, we shall model the input values as iid integer random variables $`{(X_{1},\ldots,X_{n})} \equiv X`$ (identifying $`V = {\lbrack n\rbrack}`$) with common distribution $`p`$. The (binary) entropy of this distribution will be denoted by $`{H_{2}\hspace{0pt}{(p)}} \equiv {- {\sum_{x}{p\hspace{0pt}{(x)}\hspace{0pt}{\log_{2}p}\hspace{0pt}{(x)}}}}`$.
######
Definition 5
.
A sequence of counters braids $`\{{G_{n} = {(I_{n},R_{n},E_{n})}}\}`$, with $`{|I_{n}|} = n`$ has *design rate* $`r`$ if
<table>
<tbody>
<tr>
<td></td>
<td>r=limn→∞|Rn||In|log2q.𝑟subscript→𝑛subscript𝑅𝑛subscript𝐼𝑛subscript2𝑞\displaystyle r=\lim_{n\to\infty}\frac{|R_{n}|}{|I_{n}|}\log_{2}q\,.</td>
<td></td>
<td>(3)</td>
</tr>
</tbody>
</table>
It is *reliable* for the distribution $`p`$ if there exists a sequence of reconstruction functions $`{\hat{\mathsf{F}}}_{n} \equiv {\hat{\mathsf{F}}}_{G_{n}}`$ such that, for $`X`$ a random input and $`Y \equiv {\mathsf{F}_{G_{n}}\hspace{0pt}{(X)}}`$
<table>
<tbody>
<tr>
<td></td>
<td>Perr(Gn,𝖥^n)≡ℙ{𝖥^n(Y)≠X}→n0.subscriptPerrsubscript𝐺𝑛subscript^𝖥𝑛ℙsubscript^𝖥𝑛𝑌𝑋superscript→𝑛0\displaystyle{\rm P}_{{\rm\footnotesize err}}(G_{n},\widehat{\sf F}_{n})\equiv\mathbb{P}\{\widehat{\sf F}_{n}(Y)\neq X\}\stackrel{{\scriptstyle n}}{{\to}}0\,.</td>
<td></td>
<td>(4)</td>
</tr>
</tbody>
</table>
Shannon’s source coding theorem implies that there cannot exist reliable counter braids with asymptotic rate $`r < {h_{2}\hspace{0pt}{(p)}}`$. However, the achievability of such rates is far from obvious, since counter braids are a fairly specific compression scheme. The main theorem of this paper establishes achievability, even under the restriction that the braid is sparse.
In order to avoid technical complication, we make two assumptions on the input distribution $`p`$:
1. 1\.
It has *at most power-law tails*. By this we mean that $`{{\mathbb{P}}\hspace{0pt}{\{{X_{i} \geq x}\}}} \leq {A\hspace{0pt}x^{- \epsilon}}`$ for some $`\epsilon > 0`$.
2. 2\.
It has *decreasing digit entropy*. Let $`X_{i} = {\sum_{a \geq 0}{X_{i}\hspace{0pt}{(a)}\hspace{0pt}q^{a}}}`$ be the $`q`$-ary expansion of $`X_{i}`$, and $`h_{l}`$ be the $`q`$-ary entropy of $`X_{i}\hspace{0pt}{(l)}`$. Then $`h_{l}`$ is monotonically decreasing in $`l`$ for any $`q`$ large enough.
We call a distribution $`p`$ with this two properties *admissible*. While this class does not cover all possible distributions, it is likely to include any case of practical interest.
######
Theorem 1
.
For any admissible input distribution $`p`$, and any rate $`r > {H_{2}\hspace{0pt}{(p)}}`$ there exist a sequence of reliable sparse counter braids with asymptotic rate $`r`$.
As stressed above, we insist on the braid being sparse for two reasons: $`(i)`$ It allows to update the registers content $`y`$ with a small number of operations, whenever one entry of $`x`$ is incremented (i.e. the storage function can be efficiently recomputed); $`({i\hspace{0pt}i})`$ It opens the way to using low-complexity message passing algorithms for estimating the input vector $`x`$, given the stored information (i.e. for evaluating the recovery function $`{\hat{\mathsf{F}}}_{G}`$).
## III The architecture
###
III-A
Layering
We will consider *layered* architectures. By this we mean that the set of register is the disjoint union of $`L`$ layers $`R = {R^{1} \cup R^{2} \cup \cdots \cup R^{L}}`$ and that directed edges are either from $`I`$ to $`R^{1}`$ or from $`R^{l}`$ to $`R^{l + 1}`$ for some $`l \in {\{ 1,\ldots,{L - 1}\}}`$ (we shall sometimes adopt the convention $`R^{0} \equiv I`$). We denote by $`y^{(l)} = {\{ y_{i}:{i \in R^{l}}\}}`$ the vector of register values in layer $`l`$. We further let $`m_{l} \equiv {|R^{l}|}`$ denote the size of the $`l`$-th layer (with $`m_{0} \equiv n`$).
The graph structure is conveniently encoded in $`L`$ matrices $`{\mathbb{H}}_{1}`$, …$`{\mathbb{H}}_{L}`$, whereby $`{\mathbb{H}}_{l}`$ is the $`m_{l} \times m_{l - 1}`$ adjacency matrix of the subgraph induced by $`R^{l} \cup R^{l - 1}`$. We further let $`{\mathbb{H}}^{l} = {{{\mathbb{H}}_{l} \cdot {\mathbb{H}}_{l - 1}}\hspace{0pt}\cdots\hspace{0pt}{\mathbb{H}}_{1}}`$. The storage function $`\mathsf{F}`$ can be characterized as follows.
######
Lemma 2
.
Consider an $`L`$-layers counters braid, let $`x`$ be its input, and define the sequence of vectors $`z^{(l)} \in {\mathbb{N}}^{R^{l}}`$, by $`z^{(0)} = x`$ and
<table>
<tbody>
<tr>
<td></td>
<td>z(l)=⌊(ℍlz(l−1))/q⌋.superscript𝑧𝑙subscriptℍ𝑙superscript𝑧𝑙1𝑞\displaystyle z^{(l)}=\lfloor({\mathbb{H}}_{l}z^{(l-1)})/q\rfloor\,.</td>
<td></td>
<td>(5)</td>
</tr>
</tbody>
</table>
(the division and floor operation being component-wise on the vector $`{\mathbb{H}}_{l}\hspace{0pt}z^{({l - 1})}`$.) Then, the register values are $`y^{(l)} = {{{\mathbb{H}}_{l}\hspace{0pt}z^{({l - 1})}}\operatorname{mod}q}`$.
###
III-B
Recovery function
We now describe the recovery function $`\hat{\mathsf{F}}`$. Since in this paper we are only interested in achievability, we will neglect complexity considerations.
####
III-B1
One layer
Let us start from a one-layer braid and assume the inputs to be iid with common distribution $`p_{\ast}`$ supported on $`{\{ 0,{{\ldots\hspace{0pt}q} - 1}\}} \ni x_{i}`$. Then the register values are $`y = {{\mathbb{H}}\hspace{0pt}x}`$ $`\operatorname{mod}q`$, where $`\mathbb{H}`$ is the adjacency matrix of the braid. Fix $`\gamma \in {(0,1)}`$. We say that the input $`x \in {\{ 0,\ldots,{q - 1}\}}^{n}`$ is *typical*, and write $`x \in {\mathsf{T}_{n}\hspace{0pt}{(p_{\ast})}}`$ if its type $`\theta_{x}`$ satisfies $`D{(\theta_{x}||p_{\ast})} \leq n^{- \gamma}`$ (here the Kullback-Leibler divergence is computed in natural base). Denote by $`\mathsf{T}_{n}\hspace{0pt}{(p_{\ast};y)}`$ the set of input vectors that are typical and such that $`{{\mathbb{H}}\hspace{0pt}x} = y`$ $`\operatorname{mod}q`$. The ‘typical set decoder’ returns a vector $`\hat{x}`$ if this is the the unique element in $`\mathsf{T}_{n}\hspace{0pt}{(p_{\ast};y)}`$ and a standard error message otherwise. In formulae
<table>
<tbody>
<tr>
<td></td>
<td>𝖥^(y)={x^if 𝖳n(p∗;y)={x^},∗if |𝖳n(p∗;y)|≠1.^𝖥𝑦cases^𝑥if 𝖳n(p∗;y)={x^},∗if |𝖳n(p∗;y)|≠1.\displaystyle\widehat{\sf F}(y)=\left\{\begin{array}[]{ll}\widehat{x}&\mbox{if ${\sf T}_{n}(p_{*};y)=\{\widehat{x}\}$,}\\ \ast&\mbox{if $|{\sf T}_{n}(p_{*};y)|\neq 1$.}\end{array}\right.</td>
<td></td>
<td>(8)</td>
</tr>
</tbody>
</table>
####
III-B2
Multi-layer
Consider now a multi-layer braid and $`x \in {\mathbb{N}}^{I}`$ (inputs not restricted to be smaller than $`q`$) with $`x_{i}`$’s distributed independently according to $`p`$. It is convenient to write the input vector in base $`q`$
<table>
<tbody>
<tr>
<td></td>
<td>x=∑a≥0x(a)qa.𝑥subscript𝑎0𝑥𝑎superscript𝑞𝑎\displaystyle x=\sum_{a\geq 0}x(a)\,q^{a}\,.</td>
<td></td>
<td>(9)</td>
</tr>
</tbody>
</table>
where $`{x\hspace{0pt}{(a)}} = {\{{x_{i}\hspace{0pt}{(a)}}:{i \in V}\}}`$ with $`{x_{i}\hspace{0pt}{(a)}} \in {\{ 0,\ldots,{q - 1}\}}`$. Notice that, for each $`a \geq 0`$, the vector $`x\hspace{0pt}{(a)}`$ has iid entries. Let $`p_{a}`$ be the distribution on $`x_{i}\hspace{0pt}{(a)}`$ when $`x_{i}`$ has distribution $`p`$.
We’ll apply typical set decoding recursively, determining the $`q`$-ary vectors $`x\hspace{0pt}{(0)}`$, $`x\hspace{0pt}{(1)}`$, $`x\hspace{0pt}{(2)}`$, $`\ldots`$ in this order. Consider first $`x\hspace{0pt}{(0)}`$. It is clear from Lemma [2](#Thmlemma2 "Lemma 2. ‣ III-A Layering ‣ III The architecture ‣ Detailed Network Measurements Using Sparse Graph Counters: The Theory") that $`y^{(1)} = {{\mathbb{H}}_{1}\hspace{0pt}x\hspace{0pt}{(0)}} = {{\mathbb{H}}^{1}\hspace{0pt}x\hspace{0pt}{(0)}}`$ $`\operatorname{mod}q`$. We then apply typical set decoding to the determination of $`x\hspace{0pt}{(0)}`$. More precisely, we look for a solution of $`{{\mathbb{H}}^{1}\hspace{0pt}x} = y^{(1)}`$ $`\operatorname{mod}q`$ that is typical under distribution $`p_{0}`$. If there is a unique such solution, we declare it our estimate of $`x\hspace{0pt}{(0)}`$ and denote by $`\hat{x}\hspace{0pt}{(0)}`$. Otherwise we declare an error.
Consider now the determination of $`x\hspace{0pt}{(l)}`$ and assume the lower order terms in the expansion ([9](#S3.E9 "In III-B2 Multi-layer ‣ III-B Recovery function ‣ III The architecture ‣ Detailed Network Measurements Using Sparse Graph Counters: The Theory")) have already been estimated to be $`\hat{x}\hspace{0pt}{(0)}`$, $`\hat{x}\hspace{0pt}{(1)}`$, $`\ldots`$, $`\hat{x}\hspace{0pt}{({l - 1})}`$. Let $`{\hat{z}}^{(0)} \equiv {\sum_{a = 0}^{l - 1}{\hat{x}\hspace{0pt}{(a)}\hspace{0pt}q^{a}}}`$, and $`{\hat{z}}^{(a)}`$, $`a \geq 1`$ be determined through the same recursion as in Eq. ([5](#S3.E5 "In Lemma 2. ‣ III-A Layering ‣ III The architecture ‣ Detailed Network Measurements Using Sparse Graph Counters: The Theory")). Further let $`{\hat{y}}^{(a)} = {{{\mathbb{H}}_{a}\hspace{0pt}{\hat{z}}^{({a - 1})}}\operatorname{mod}q}`$ (this are nothing but the register values on input $`{\hat{z}}^{(0)}`$).
Assume the estimates $`\hat{x}\hspace{0pt}{(0)}`$, $`\hat{x}\hspace{0pt}{(1)}`$, $`\ldots`$, $`\hat{x}\hspace{0pt}{({l - 1})}`$ to be correct. It is then easy to realize that $`{\hat{y}}^{(a)} = y^{(a)}`$ for $`a = {1,{\ldots\hspace{0pt}l}}`$. Further $`z^{(l)} = {{\hat{z}}^{(l)} + {{\mathbb{H}}^{l}\hspace{0pt}x\hspace{0pt}{(l)}}}`$ $`\operatorname{mod}q`$, hence
<table>
<tbody>
<tr>
<td></td>
<td>y(l+1)=y^(l+1)+ℍl+1x(l)modq.superscript𝑦𝑙1modulosuperscript^𝑦𝑙1superscriptℍ𝑙1𝑥𝑙𝑞\displaystyle y^{(l+1)}=\widehat{y}^{(l+1)}+{\mathbb{H}}^{l+1}x(l)\,\mod q\,.</td>
<td></td>
<td>(10)</td>
</tr>
</tbody>
</table>
We therefore proceed to compute $`y^{({l + 1})} - {\hat{y}}^{({l + 1})}`$ $`\operatorname{mod}q`$. If the linear system $`{{\mathbb{H}}^{l + 1}\hspace{0pt}x\hspace{0pt}{(l)}} = {y^{({l + 1})} - {\hat{y}}^{({l + 1})}}`$ $`\operatorname{mod}q`$ admits more than one or no solution that is typical with respect to the distribution $`p_{l}`$, an error is returned. Otherwise, the next term in the expansion ([9](#S3.E9 "In III-B2 Multi-layer ‣ III-B Recovery function ‣ III The architecture ‣ Detailed Network Measurements Using Sparse Graph Counters: The Theory")) is estimated through the unique typical solution of such linear system.
The recovery algorithm is summarized below, with one improvement with respect to the description above. Instead of recomputing $`{\hat{z}}^{(0)}`$, $`\ldots,{\hat{z}}^{(l)}`$, at stage $`l`$ we only compute the vector $`{\hat{z}}^{(l)}`$ that is needed at the present stage.
<table>
<thead>
<tr>
<th colspan="2">RECOVERY (input: register values y𝑦y)</th>
</tr>
</thead>
<tbody>
<tr>
<th>1:</th>
<td>Initialize z^(a)=0superscript^𝑧𝑎0\widehat{z}^{(a)}=0 for a≥0𝑎0a\geq 0;</td>
</tr>
<tr>
<th>2:</th>
<td>for l∈{0,…L}𝑙0…𝐿l\in\{0,\dots L\}</td>
</tr>
<tr>
<th>3:</th>
<td> Set y^(l+1)=ℍl+1z^(l)^𝑦𝑙1subscriptℍ𝑙1superscript^𝑧𝑙\widehat{y}(l+1)={\mathbb{H}}_{l+1}\widehat{z}^{(l)} modqmoduloabsent𝑞\mod q;</td>
</tr>
<tr>
<th>4:</th>
<td> Let 𝖳lsubscript𝖳𝑙{\sf T}_{l} be the set of plsubscript𝑝𝑙p_{l}-typical</td>
</tr>
<tr>
<th></th>
<td> solutions of ℍl+1x^=y(l+1)−y^(l+1)superscriptℍ𝑙1^𝑥superscript𝑦𝑙1superscript^𝑦𝑙1{\mathbb{H}}^{l+1}\widehat{x}=y^{(l+1)}-\widehat{y}^{(l+1)}, modqmoduloabsent𝑞\mod q;</td>
</tr>
<tr>
<th>5:</th>
<td> If 𝖳l={x^}subscript𝖳𝑙^𝑥{\sf T}_{l}=\{\widehat{x}\} let x^(l)=x^^𝑥𝑙^𝑥\widehat{x}(l)=\widehat{x}</td>
</tr>
<tr>
<th></th>
<td> otherwise if |𝖳l|≠1subscript𝖳𝑙1|{\sf T}_{l}|\neq 1 return error;</td>
</tr>
<tr>
<th>6:</th>
<td> Set z^(l+1)=⌊{ℍl+1z^(l)+ℍl+1x^(l)}/q⌋superscript^𝑧𝑙1superscriptℍ𝑙1superscript^𝑧𝑙subscriptℍ𝑙1^𝑥𝑙𝑞\widehat{z}^{(l+1)}=\lfloor\{{\mathbb{H}}^{l+1}\widehat{z}^{(l)}+{\mathbb{H}}_{l+1}\widehat{x}(l)\}/q\rfloor;</td>
</tr>
<tr>
<th>7:</th>
<td>end</td>
</tr>
<tr>
<th>8:</th>
<td>return x^=∑ix^(i)qi^𝑥subscript𝑖^𝑥𝑖superscript𝑞𝑖\widehat{x}=\sum_{i}\widehat{x}(i)\,q^{i}.</td>
</tr>
</tbody>
</table>
###
III-C
Sparse graph ensemble and choice of the parameters
The optimal compression rate in Theorem [1](#Thmthm1 "Theorem 1. ‣ II-A Main results ‣ II Counter Braids: Basic definitions ‣ Detailed Network Measurements Using Sparse Graph Counters: The Theory") is achieved with the following random sparse graph construction. Fix the registers capacity $`q`$ and an integer $`k \geq 2`$. Then for $`l = {1,\ldots,L_{0}}`$ the graph induced by vertices $`R_{l - 1} \cup R_{l}`$ has a random edge set that is sampled by connecting each $`i \in R_{l - 1}`$ to $`k`$ iid uniformly random vertices in $`R_{l}`$ (all edges being directed from $`R_{l - 1}`$ to $`R_{l}`$). In other words, the $`m_{l} \times m_{l - 1}`$, $`0 - 1`$ matrix $`{\mathbb{H}}_{l}`$ has independent columns, each sampled by incrementing $`k`$ iid positions.
The choice of this ensemble is motivated by implementation concerns. In the flow counting problem, we do not know a priori the exact number of flows that needs to be stored. The above structure, this number can be changed without modifying existing links. Further, for each new flow, the subset of $`k`$ registers it is connected to can be chosen through a simple hash function.
To these $`L_{0}`$ stages, we add further $`L_{1}`$ stages, all of the same size $`m_{L_{0} + 1} = \cdots = m_{L_{0} + L_{1}} = m_{\ast}`$, with edges connecting each node in $`R_{l - 1}`$ to a different node in $`R_{l}`$. Equivalently, we take $`{\mathbb{H}}_{l}`$ to be the identity matrix in these stages.
It remains to specify the number of stages $`L_{0}`$, $`L_{1}`$ and their sizes $`m_{1}`$,…, $`m_{L_{0}}`$. Let $`p_{l}`$ be the distribution of the $`l`$-th least significant digit in the $`q`$-ary expansion of $`X_{i}`$. Recall that we defined $`h_{l}`$ to be the $`q`$-ary entropy of the distribution $`p_{l}`$, i.e.
<table>
<tbody>
<tr>
<td></td>
<td>hl≡−∑x=0q−1pl(x)logqpl(x).subscriptℎ𝑙superscriptsubscript𝑥0𝑞1subscript𝑝𝑙𝑥subscript𝑞subscript𝑝𝑙𝑥\displaystyle h_{l}\equiv-\sum_{x=0}^{q-1}p_{l}(x)\log_{q}p_{l}(x)\,.</td>
<td></td>
<td>(11)</td>
</tr>
</tbody>
</table>
Finally, in the achievability proof, we shall assume that $`q`$ is a prime number, large enough for $`h_{l}`$ to be monotonically decreasing.
######
Lemma 3
.
Assume $`{{\mathbb{P}}\hspace{0pt}{\{{X_{1} \geq x}\}}} \leq {A\hspace{0pt}x^{- \epsilon}}`$. Then there exists constants $`B,C`$ that only depend on $`A,\epsilon`$, such that for all $`l \geq 1`$, and all $`q`$ large enough
<table>
<tbody>
<tr>
<td></td>
<td>hl≤Blq−lϵ,subscriptℎ𝑙𝐵𝑙superscript𝑞𝑙italic-ϵ\displaystyle h_{l}\leq B\,l\,q^{-l\epsilon}\;,</td>
<td></td>
<td>(12)</td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td>|h2(p)−∑l≥0hllog2q|≤Cq−ϵ(log2q)2.subscriptℎ2𝑝subscript𝑙0subscriptℎ𝑙subscript2𝑞𝐶superscript𝑞italic-ϵsuperscriptsubscript2𝑞2\displaystyle\Big{|}h_{2}(p)-\sum_{l\geq 0}h_{l}\log_{2}q\Big{|}\leq C\,q^{-\epsilon}(\log_{2}q)^{2}\,.</td>
<td></td>
<td>(13)</td>
</tr>
</tbody>
</table>
The proof of this simple Lemma is deferred to Section [VI](#S6 "VI Some auxiliary results ‣ Detailed Network Measurements Using Sparse Graph Counters: The Theory").
######
Lemma 4
.
Let $`p_{\ast}`$ be a distribution over $`\{ 0,\ldots,q\}`$, with $`q`$-ary entropy $`H\hspace{0pt}{(p_{\ast})}`$, and $`\mathsf{T}_{n}\hspace{0pt}{(p_{\ast})}`$ be the set of $`p_{\ast}`$-typical vectors defined as in Sec [](#S3.SS2 "III-B Recovery function ‣ III The architecture ‣ Detailed Network Measurements Using Sparse Graph Counters: The Theory")
III-B
. Let $`|{\mathsf{T}_{n}\hspace{0pt}{(p_{\ast})}}|`$ be the size of this set. Recall that $`x \in {\mathsf{T}_{n}\hspace{0pt}{(p_{\ast})}}`$ if its type $`\theta_{x}`$ satisfies $`D{(\theta_{x}||p_{\ast})} \leq n^{- \gamma}`$. Then, for any $`\beta \in {({1 - {\gamma/2}},1)}`$, there exists $`A = {A\hspace{0pt}{(\beta,\gamma,q)}}`$ such that
<table>
<tbody>
<tr>
<td></td>
<td>|𝖳n(p∗)|≤qnH(p∗)+Anβ,.subscript𝖳𝑛subscript𝑝superscript𝑞𝑛𝐻subscript𝑝𝐴superscript𝑛𝛽\displaystyle|{\sf T}_{n}(p_{*})|\leq q^{nH(p_{*})+A\,n^{\beta}}\,,\,.</td>
<td></td>
</tr>
</tbody>
</table>
Further, if $`X = {(X_{1},\ldots,X_{n})}`$ is a vector with iid entries with common distribution $`p_{\ast}`$
<table>
<tbody>
<tr>
<td></td>
<td>ℙ{X∉𝖳n(p∗)}≤(n+1)qe−n1−γ.ℙ𝑋subscript𝖳𝑛subscript𝑝superscript𝑛1𝑞superscript𝑒superscript𝑛1𝛾\displaystyle\mathbb{P}\left\{X\not\in{\sf T}_{n}(p_{*})\right\}\leq(n+1)^{q}\,e^{-n^{1-\gamma}}\,.</td>
<td></td>
<td>(14)</td>
</tr>
</tbody>
</table>
In the following we will consider $`\gamma`$ and $`\beta`$ fixed once and for all, for instance by $`\gamma = {1/2}`$ and $`\beta = {7/8}`$.
Fix some $`\delta > 0`$, and let $`A\hspace{0pt}{(q)}`$ be a suitably large constant, we let, for $`l = {1,\ldots,L_{0}}`$,
<table>
<tbody>
<tr>
<td></td>
<td>mlsubscript𝑚𝑙\displaystyle m_{l}</td>
<td>≡\displaystyle\equiv</td>
<td>max{m¯l,⌈δml−1⌉},subscript¯𝑚𝑙𝛿subscript𝑚𝑙1\displaystyle\max\{\underline{m}_{l},\lceil\delta\,m_{l-1}\rceil\}\,,</td>
<td></td>
<td>(15)</td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td>m¯lsubscript¯𝑚𝑙\displaystyle\underline{m}_{l}</td>
<td>≡\displaystyle\equiv</td>
<td>(1+δ)⌈nhl−1(1+δ)+A(q)nβ⌉.1𝛿𝑛subscriptℎ𝑙11𝛿𝐴𝑞superscript𝑛𝛽\displaystyle(1+\delta)\left\lceil nh_{l-1}(1+\delta)+A(q)n^{\beta}\right\rceil\,.</td>
<td></td>
<td>(16)</td>
</tr>
</tbody>
</table>
The number of stages is such that
<table>
<tbody>
<tr>
<td></td>
<td>m¯l≤n(logn)−2 for any l≥L0.subscript¯𝑚𝑙𝑛superscript𝑛2 for any l≥L0\displaystyle\underline{m}_{l}\leq n(\log n)^{-2}\mbox{ for any $l\geq L_{0}$}\,.</td>
<td></td>
<td>(17)</td>
</tr>
</tbody>
</table>
This implies, by Lemma [3](#Thmlemma3 "Lemma 3. ‣ III-C Sparse graph ensemble and choice of the parameters ‣ III The architecture ‣ Detailed Network Measurements Using Sparse Graph Counters: The Theory"), $`L_{0} = {O\hspace{0pt}{({\log{\log n}})}}`$. To this we add $`L_{1} = {({\log n})}^{3/2}`$ stages within the second group, of size $`m_{\ast} = m_{L_{0}} \leq {n\hspace{0pt}{({\log n})}^{- 2}}`$. The total number of registers is therefore upper bounded as $`{|R|} \leq {{n\hspace{0pt}{({1 + \delta})}\hspace{0pt}{\sum_{l \geq 0}{{({h_{l} + {A\hspace{0pt}n^{\beta - 1}}})}\hspace{0pt}{({\sum_{i \geq 0}\delta^{i}})}}}} + {n\hspace{0pt}{({\log n})}^{- {1/2}}}}`$, and therefore the asymptotic rate of this architecture
<table>
<tbody>
<tr>
<td></td>
<td>r≤1+δ1−δ∑l≥0hllog2q.𝑟1𝛿1𝛿subscript𝑙0subscriptℎ𝑙subscript2𝑞\displaystyle r\leq\frac{1+\delta}{1-\delta}\sum_{l\geq 0}h_{l}\log_{2}q\,.</td>
<td></td>
<td>(18)</td>
</tr>
</tbody>
</table>
Since the right hand side can be made arbitrarily close to $`h\hspace{0pt}{(p)}`$ by Lemma [3](#Thmlemma3 "Lemma 3. ‣ III-C Sparse graph ensemble and choice of the parameters ‣ III The architecture ‣ Detailed Network Measurements Using Sparse Graph Counters: The Theory"), Theorem [1](#Thmthm1 "Theorem 1. ‣ II-A Main results ‣ II Counter Braids: Basic definitions ‣ Detailed Network Measurements Using Sparse Graph Counters: The Theory") follows from the following.
######
Theorem 2
.
For any input distribution $`p`$ with at most power-law tails and any choice of $`q \geq 2`$ and $`\delta > 0`$, there exists $`k \geq 2`$ such that the multi-layer braid described above is reliable.
## IV Analysis of one-layer architectures
In order to prove our main Theorem or, equivalently, Theorem [2](#Thmthm2 "Theorem 2. ‣ III-C Sparse graph ensemble and choice of the parameters ‣ III The architecture ‣ Detailed Network Measurements Using Sparse Graph Counters: The Theory"), we need first to prove a few preliminary results concerning a one-layer architecture. The proof here follows the technique of \[[9](#bib.bib9)\], the main tool being an estimate of the distance enumerator as in \[[4](#bib.bib4), [10](#bib.bib10), [11](#bib.bib11)\]. Distance enumerators for non-binary LDPC codes have been estimated in \[[12](#bib.bib12)\]. Unhappily we cannot here limit ourselves to citing these works, because the graph ensemble is different from the regular ones treated there.
Throughout this Section the source is a vector $`X = {(X_{1},\ldots,X_{n})}`$ with iid entries taking values in $`\{ 0,\ldots,{q - 1}\}`$ and distribution $`p_{\ast}`$ (in the application to multi-layer schemes $`p_{\ast}`$ will coincide with $`p_{l}`$ for some $`l \geq 0`$). We let $`\mathbb{H}`$ be an $`m \times n`$ matrix whose columns are independent vectors with integer entries. Each column is obtained by choosing $`k`$ positions independently and uniformly at random (eventually with repetition) and incrementing the corresponding entry by one. In other words, $`\mathbb{H}`$ is distributed as the adjacency matrix of a given layer in the multi-layer architecture.
Our first result is a simple combinatorial calculation. Let $`\overset{\rightarrow}{\lambda} = {\{\lambda_{z}:{z = {1,\ldots,{q - 1}}}\}}`$ be a vector in $`{\mathbb{R}}_{+}^{q - 1}`$. It is convenient to introduce the random variable $`\overset{\rightarrow}{W} = {\{ W_{z}:{z = {1,\ldots,{q - 1}}}\}}`$ taking values in $`{\mathbb{N}}^{q - 1}`$. The joint distribution of $`(W_{1},\ldots,W_{q - 1})`$ (to be denoted by $`{\mathbb{P}}_{\overset{\rightarrow}{\lambda}}`$) is the one of $`q - 1`$ Poisson random variables with means (respectively) $`\lambda_{1},\ldots,\lambda_{q - 1}`$, conditioned on $`{\sum_{z = 1}^{q - 1}{z\hspace{0pt}W_{z}}} = 0`$, $`\operatorname{mod}q`$.
######
Lemma 5
.
Let $`x \in {\{ 0,\ldots,{q - 1}\}}^{n}`$ be an input vector with $`n_{z}`$ entries equal to $`z`$, for $`z = {0,\ldots,{q - 1}}`$, and $`\mathbb{H}`$ be a random matrix as above. Define $`\overset{\rightarrow}{n} = {\{ n_{z}:{z = {1,\ldots,{q - 1}}}\}}`$. For any $`\overset{\rightarrow}{\lambda} \in {\mathbb{R}}_{+}^{q - 1}`$, let $`{\overset{\rightarrow}{W}}_{1}`$,…$`{\overset{\rightarrow}{W}}_{m}`$ be $`m`$ iid vectors with distribution $`{\mathbb{P}}_{\overset{\rightarrow}{\lambda}}`$. Then the probability that $`{{\mathbb{H}}\hspace{0pt}x} = 0`$ $`\operatorname{mod}q`$ is
<table>
<tbody>
<tr>
<td></td>
<td>ℙ{ℍx=0}=∏z=1q−1(knz)!emλz(mλz)knzQ(λ→)mℙλ→{∑i=1mW→i=kn→},ℙℍ𝑥0superscriptsubscriptproduct𝑧1𝑞1𝑘subscript𝑛𝑧superscript𝑒𝑚subscript𝜆𝑧superscript𝑚subscript𝜆𝑧𝑘subscript𝑛𝑧𝑄superscript→𝜆𝑚subscriptℙ→𝜆superscriptsubscript𝑖1𝑚subscript→𝑊𝑖𝑘→𝑛\displaystyle\mathbb{P}\{{\mathbb{H}}x=0\}=\prod_{z=1}^{q-1}\frac{(kn_{z})!\,e^{m\lambda_{z}}}{(m\lambda_{z})^{kn_{z}}}\,Q(\vec{\lambda})^{m}\,\mathbb{P}_{\vec{\lambda}}\left\{\sum_{i=1}^{m}\vec{W}_{i}=k\vec{n}\right\}\,,</td>
<td></td>
<td>(19)</td>
</tr>
</tbody>
</table>
where $`Q\hspace{0pt}{(\overset{\rightarrow}{\lambda})}`$ is the probability that $`{\sum_{z = 1}^{q - 1}{z\hspace{0pt}U_{z}}} = 0`$, $`\operatorname{mod}q`$ for independent Poisson random variables with means $`\lambda_{z}`$.
Further, for some universal constant $`C`$, and $`D_{q} = {1 - {\cos{{2\hspace{0pt}\pi}/q}}}`$, and $`n_{\ast} = {\sum_{z}n_{z}}`$
<table>
<tbody>
<tr>
<td></td>
<td>ℙ{ℍx=0}ℙℍ𝑥0\displaystyle\mathbb{P}\{{\mathbb{H}}x=0\}</td>
<td>≤\displaystyle\leq</td>
<td>(Ckn∗)q−12Q(kn→/m)mR(m,kq∑z=1q−1znz)superscript𝐶𝑘subscript𝑛𝑞12𝑄superscript𝑘→𝑛𝑚𝑚𝑅𝑚𝑘𝑞superscriptsubscript𝑧1𝑞1𝑧subscript𝑛𝑧\displaystyle(Ckn_{*})^{\frac{q-1}{2}}Q(k\vec{n}/m)^{m}R(m,\frac{k}{q}\sum_{z=1}^{q-1}zn_{z})\,</td>
<td></td>
<td></td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td>Q(kn→/m)𝑄𝑘→𝑛𝑚\displaystyle Q(k\vec{n}/m)</td>
<td>≤\displaystyle\leq</td>
<td>1q[1+(q−1)e−Dqkn∗m],1𝑞delimited-[]1𝑞1superscript𝑒subscript𝐷𝑞𝑘subscript𝑛𝑚\displaystyle\frac{1}{q}\left[1+(q-1)\,e^{-D_{q}\frac{kn_{*}}{m}}\right]\,,</td>
<td></td>
<td>(21)</td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td>R(m,N)𝑅𝑚𝑁\displaystyle R(m,N)</td>
<td>=\displaystyle=</td>
<td>min{1,(Cq2N/m)N}1superscript𝐶superscript𝑞2𝑁𝑚𝑁\displaystyle\min\left\{1,(Cq^{2}N/m)^{N}\right\}</td>
<td></td>
<td>(22)</td>
</tr>
</tbody>
</table>
###### Proof.
Due to the symmetry of the distribution of $`\mathbb{H}`$ with respect to permutation in its columns, $`{\mathbb{P}}\hspace{0pt}{\{{{{\mathbb{H}}\hspace{0pt}x} = 0}\}}`$ does depend on $`x`$ only through the number of ones, twos, etc. Without loss of generality we can assume the first $`n_{1}`$ coordinates to be ones, the next $`n_{2}`$ to be twos, and so on, and neglect the last $`n - {\sum_{z}n_{z}}`$ columns, corresponding to zeros. Think now of filling the matrix, by choosing its non-zero entries (edges in the associated graph). If we associate to each such entry the value of the corresponding coordinate in $`x`$, we want the probability for the sum of labels on each row to be $`0`$ $`\operatorname{mod}q`$. Since entries are independent and uniformly random, this is equal to the probability that each of $`m`$ urns is filled with balls whose labels add to $`0`$, when we throw $`k\hspace{0pt}n_{1}`$ balls labeled with $`1`$, $`k\hspace{0pt}n_{2}`$ labeled with $`2`$, and so on. It is an exercise in combinatorics to show that this is
<table>
<tbody>
<tr>
<td></td>
<td>∏z=1q−1(knz)!mknz𝖼𝗈𝖾𝖿𝖿{P(ξ1,…,ξq−1)m,ξ1kn1⋯ξq−1knq−1},superscriptsubscriptproduct𝑧1𝑞1𝑘subscript𝑛𝑧superscript𝑚𝑘subscript𝑛𝑧𝖼𝗈𝖾𝖿𝖿𝑃superscriptsubscript𝜉1…subscript𝜉𝑞1𝑚superscriptsubscript𝜉1𝑘subscript𝑛1⋯superscriptsubscript𝜉𝑞1𝑘subscript𝑛𝑞1\displaystyle\prod_{z=1}^{q-1}\frac{(kn_{z})!}{m^{kn_{z}}}\,{\sf coeff}\left\{P(\xi_{1},\dots,\xi_{q-1})^{m},\xi_{1}^{kn_{1}}\cdots\xi_{q-1}^{kn_{q-1}}\right\}\,,</td>
<td></td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td>P(⋯)≡∑l1…lq−1ξ1l1l1!…ξq−1lq−1lq−1!𝕀{∑z=1q−1zlz=0}.𝑃⋯subscriptsubscript𝑙1…subscript𝑙𝑞1superscriptsubscript𝜉1subscript𝑙1subscript𝑙1…superscriptsubscript𝜉𝑞1subscript𝑙𝑞1subscript𝑙𝑞1𝕀superscriptsubscript𝑧1𝑞1𝑧subscript𝑙𝑧0\displaystyle P(\cdots)\equiv\sum_{l_{1}\dots l_{q-1}}\frac{\xi_{1}^{l_{1}}}{l_{1}!}\dots\frac{\xi_{q-1}^{l_{q-1}}}{l_{q-1}!}\,{\mathbb{I}}\left\{\sum_{z=1}^{q-1}z\,l_{z}=0\right\}\,.</td>
<td></td>
</tr>
</tbody>
</table>
Equation ([19](#S4.E19 "In Lemma 5. ‣ IV Analysis of one-layer architectures ‣ Detailed Network Measurements Using Sparse Graph Counters: The Theory")) is then obtained by evaluating $`{\mathbb{P}}_{\overset{\rightarrow}{\lambda}}`$ and showing that it yields the above combinatorial expression.
In order to get Eq. (LABEL:eq:BoundProb), we denote $`{\mathbb{P}}_{\overset{\rightarrow}{\lambda}}\hspace{0pt}{\{\cdots\}}`$ by $`R`$, and use $`\lambda_{z} = {{k\hspace{0pt}n_{z}}/m}`$, thus leading to
<table>
<tbody>
<tr>
<td></td>
<td>ℙ{ℍx=0}=∏z=1q−1(knz)!eknz(knz)knzQ(kn→/m)mR.ℙℍ𝑥0superscriptsubscriptproduct𝑧1𝑞1𝑘subscript𝑛𝑧superscript𝑒𝑘subscript𝑛𝑧superscript𝑘subscript𝑛𝑧𝑘subscript𝑛𝑧𝑄superscript𝑘→𝑛𝑚𝑚𝑅\displaystyle\mathbb{P}\{{\mathbb{H}}x=0\}=\prod_{z=1}^{q-1}\frac{(kn_{z})!\,e^{kn_{z}}}{(kn_{z})^{kn_{z}}}\,Q(k\vec{n}/m)^{m}R\,.</td>
<td></td>
</tr>
</tbody>
</table>
Equation (LABEL:eq:BoundProb) follows from the observation that $`{N!} \leq {\sqrt{C\hspace{0pt}N}\hspace{0pt}{({N/e})}^{N}}`$ for some universal constant $`C`$.
In order to prove Eq. ([21](#S4.E21 "In Lemma 5. ‣ IV Analysis of one-layer architectures ‣ Detailed Network Measurements Using Sparse Graph Counters: The Theory")), notice that, by discrete Fourier transform
<table>
<tbody>
<tr>
<td></td>
<td>Q(λ→)𝑄→𝜆\displaystyle Q(\vec{\lambda})</td>
<td>=\displaystyle=</td>
<td>1q∑ℓ=0q−1𝔼{e2πiℓq∑z=1q−1zUz}1𝑞superscriptsubscriptℓ0𝑞1𝔼superscript𝑒2𝜋𝑖ℓ𝑞superscriptsubscript𝑧1𝑞1𝑧subscript𝑈𝑧\displaystyle\frac{1}{q}\sum_{\ell=0}^{q-1}\mathbb{E}\left\{e^{\frac{2\pi i\ell}{q}\sum_{z=1}^{q-1}zU_{z}}\right\}</td>
<td></td>
</tr>
<tr>
<td></td>
<td></td>
<td>=\displaystyle=</td>
<td>1q∑ℓ=0q−1exp{−∑z=1q−1λz(1−e2πiℓzq)}.1𝑞superscriptsubscriptℓ0𝑞1superscriptsubscript𝑧1𝑞1subscript𝜆𝑧1superscript𝑒2𝜋𝑖ℓ𝑧𝑞\displaystyle\frac{1}{q}\sum_{\ell=0}^{q-1}\exp\left\{-\sum_{z=1}^{q-1}\lambda_{z}(1-e^{\frac{2\pi i\ell z}{q}})\right\}\,.</td>
<td></td>
</tr>
</tbody>
</table>
The claim is proved by singling out the $`\ell = 0`$ term and bounding the others using $`{{Re}\hspace{0pt}{({1 - e^{\frac{2\hspace{0pt}\pi\hspace{0pt}i\hspace{0pt}\ell\hspace{0pt}z}{q}}})}} \geq D_{q}`$.
Let us finally prove Eq. ([22](#S4.E22 "In Lemma 5. ‣ IV Analysis of one-layer architectures ‣ Detailed Network Measurements Using Sparse Graph Counters: The Theory")). Obviously $`R \leq 1`$ since it is an upper bound on the probability $`{\mathbb{P}}_{\overset{\rightarrow}{\lambda}}\hspace{0pt}{\{\cdots\}}`$. If we let $`N \equiv {\frac{k}{q}\hspace{0pt}{\sum_{z = 1}^{q - 1}{z\hspace{0pt}n_{z}}}}`$, we can therefore assume, without loss of generality, that $`N`$ is an integer with $`{N/m} \leq {1/q}`$. Let $`V_{i}`$ be distributed as $`\sum_{z = 1}^{q - 1}{{W_{i,z}\hspace{0pt}z}/q}`$ conditioned on $`V_{i}`$ being an integer. Then the probability $`{\mathbb{P}}_{\overset{\rightarrow}{\lambda}}\hspace{0pt}{\{\cdots\}}`$ is upper bounded by
<table>
<tbody>
<tr>
<td></td>
<td>ℙ{∑i=1mVi≥N}ℙsuperscriptsubscript𝑖1𝑚subscript𝑉𝑖𝑁\displaystyle\mathbb{P}\left\{\sum_{i=1}^{m}V_{i}\geq N\right\}</td>
<td>≤\displaystyle\leq</td>
<td>(mN)ℙ{Vi≥1}Nbinomial𝑚𝑁ℙsuperscriptsubscript𝑉𝑖1𝑁\displaystyle\binom{m}{N}\mathbb{P}\{V_{i}\geq 1\}^{N}</td>
<td></td>
</tr>
<tr>
<td></td>
<td></td>
<td>≤\displaystyle\leq</td>
<td>(CmNℙ{Vi≥1})N.superscript𝐶𝑚𝑁ℙsubscript𝑉𝑖1𝑁\displaystyle\left(\frac{Cm}{N}\,\mathbb{P}\{V_{i}\geq 1\}\right)^{N}\,.</td>
<td></td>
</tr>
</tbody>
</table>
Recalling the definition of $`V_{i}`$, we have
<table>
<tbody>
<tr>
<td></td>
<td>ℙ{Vi≥1}ℙsubscript𝑉𝑖1\displaystyle\mathbb{P}\{V_{i}\geq 1\}</td>
<td>=\displaystyle=</td>
<td>ℙ{∑z=1q−1zUz≥q|∑z=1q−1zUz=0modq}ℙconditional-setsuperscriptsubscript𝑧1𝑞1𝑧subscript𝑈𝑧𝑞superscriptsubscript𝑧1𝑞1𝑧subscript𝑈𝑧modulo0𝑞\displaystyle\mathbb{P}\Big{\{}\sum_{z=1}^{q-1}zU_{z}\geq q\Big{|}\sum_{z=1}^{q-1}zU_{z}=0\mod q\Big{\}}</td>
<td></td>
</tr>
<tr>
<td></td>
<td></td>
<td>≤\displaystyle\leq</td>
<td>e∑z=1q−1λzℙ{∑z=1q−1zUz≥q}.superscript𝑒superscriptsubscript𝑧1𝑞1subscript𝜆𝑧ℙsuperscriptsubscript𝑧1𝑞1𝑧subscript𝑈𝑧𝑞\displaystyle e^{\sum_{z=1}^{q-1}\lambda_{z}}\mathbb{P}\Big{\{}\sum_{z=1}^{q-1}zU_{z}\geq q\Big{\}}\,.</td>
<td></td>
</tr>
</tbody>
</table>
But $`{\sum_{z = 1}^{q - 1}\lambda_{z}} = {{k\hspace{0pt}n_{\ast}}/m} \leq {{N\hspace{0pt}q}/m} \leq 1`$. Further, $`{\sum_{z = 1}^{q - 1}{z\hspace{0pt}U_{z}}} \geq q`$ only if $`{\sum_{z = 1}^{q - 1}U_{z}} \geq 2`$. Therefore we get
<table>
<tbody>
<tr>
<td></td>
<td>ℙ{Vi≥1}ℙsubscript𝑉𝑖1\displaystyle\mathbb{P}\{V_{i}\geq 1\}</td>
<td>≤\displaystyle\leq</td>
<td>eℙ{∑z=1q−1Uz≥2}𝑒ℙsuperscriptsubscript𝑧1𝑞1subscript𝑈𝑧2\displaystyle e\mathbb{P}\Big{\{}\sum_{z=1}^{q-1}U_{z}\geq 2\Big{\}}</td>
<td></td>
</tr>
<tr>
<td></td>
<td></td>
<td>≤\displaystyle\leq</td>
<td>C(∑z=1q−1λz)2≤C(kn∗/m)2.𝐶superscriptsuperscriptsubscript𝑧1𝑞1subscript𝜆𝑧2𝐶superscript𝑘subscript𝑛𝑚2\displaystyle C\Big{(}\sum_{z=1}^{q-1}\lambda_{z}\Big{)}^{2}\leq C(kn_{*}/m)^{2}\,.</td>
<td></td>
</tr>
</tbody>
</table>
The proof is completed by noticing that $`{({{k\hspace{0pt}n_{\ast}}/m})} \leq {({{N\hspace{0pt}q}/m})}`$. ∎
In the following, given a vector $`x = {(x_{1},\ldots,x_{n})}`$, we shall denote by $`{\| x\|}_{0}`$ is number of non-zero entries.
######
Lemma 6
.
Let $`\mathbb{H}`$ be a random $`m \times n`$ matrix distributed as above, with column weight $`k`$. Assume $`k`$ not to be a multiple of $`q`$, $`m \leq n`$, $`{({{m/n}\hspace{0pt}k})}^{1/k} \geq \Delta > 0`$ and $`3 \leq k \leq {m/{\log m}}`$. Then, there exists a constant $`B = {B\hspace{0pt}{(q,\Delta)}}`$, $`C = {C\hspace{0pt}{(q,\Delta)}}`$, such that, if
<table>
<tbody>
<tr>
<td></td>
<td>E≤Cmlog(nk/m),𝐸𝐶𝑚𝑛𝑘𝑚\displaystyle E\leq\frac{C\,m}{\log(nk/m)}\,,</td>
<td></td>
<td>(23)</td>
</tr>
</tbody>
</table>
then
<table>
<tbody>
<tr>
<td></td>
<td>ℙ{∃||z||0≤E:ℍz=0modq}≤n2(Bkm)kqℙconditional-set:evaluated-at𝑧0𝐸ℍ𝑧modulo0𝑞superscript𝑛2superscript𝐵𝑘𝑚𝑘𝑞\displaystyle\mathbb{P}\Big{\{}\exists||z||_{0}\leq E\,:\,{\mathbb{H}}z=0\mod q\Big{\}}\leq n^{2}\left(\frac{Bk}{m}\right)^{\frac{k}{q}}</td>
<td></td>
<td>(24)</td>
</tr>
</tbody>
</table>
(where it is understood that $`z \in {\{ 0,\ldots,{q - 1}\}}^{n}`$.)
###### Proof.
Throughout the proof, $`A`$ will denote a generic constant depending only on $`q`$ that can be chosen large enough to make the inequalities below hold.
Let $`z \in {\{ 0,\ldots,{q - 1}\}}^{n}`$ be such that $`{\| z\|}_{0} = \ell`$. We will upper bound the probability that $`{{\mathbb{H}}\hspace{0pt}z} = 0`$ $`\operatorname{mod}q`$ in different ways depending whether $`\ell \leq E_{0}`$ or $`\ell > E_{0}`$, where
<table>
<tbody>
<tr>
<td></td>
<td>E0=ρ(q)mk(mnk)2/(k−2),subscript𝐸0𝜌𝑞𝑚𝑘superscript𝑚𝑛𝑘2𝑘2\displaystyle E_{0}=\rho(q)\,\frac{m}{k}\,\left(\frac{m}{nk}\right)^{2/(k-2)}\,,</td>
<td></td>
<td>(25)</td>
</tr>
</tbody>
</table>
with $`\rho\hspace{0pt}{(q)}`$ a function to be determined. Notice that, under our hypotheses,
<table>
<tbody>
<tr>
<td></td>
<td>kE0m≥ρ(q)Δ2k/(k−2)𝑘subscript𝐸0𝑚𝜌𝑞superscriptΔ2𝑘𝑘2\displaystyle\frac{kE_{0}}{m}\geq\rho(q)\,\Delta^{2k/(k-2)}\,</td>
<td></td>
<td>(26)</td>
</tr>
</tbody>
</table>
is bounded away from $`0`$ (as $`2 < {{2\hspace{0pt}k}/{({k - 2})}} \leq 6`$ for $`k \geq 3`$.)
For $`{\| z\|}_{0} = \ell \leq E_{0}`$ (and $`z \neq 0`$) we use Lemma [5](#Thmlemma5 "Lemma 5. ‣ IV Analysis of one-layer architectures ‣ Detailed Network Measurements Using Sparse Graph Counters: The Theory"), Eq. (LABEL:eq:BoundProb), where we set $`{Q\hspace{0pt}{(\cdots)}} \leq 1`$, $`n_{\ast} = \ell`$ and $`\frac{k\hspace{0pt}\ell}{q} \leq {\frac{k}{q}\hspace{0pt}{\sum_{z = 1}^{q - 1}{z\hspace{0pt}n_{z}}}} \leq {k\hspace{0pt}\ell}`$. Further we assumed $`{{A\hspace{0pt}k\hspace{0pt}\ell}/m} \leq 1`$, which holds without loss of generality if we take $`{\rho\hspace{0pt}{(q)}} \leq {{1/A}\hspace{0pt}\Delta^{6}} \leq {{1/A}\hspace{0pt}\Delta^{{2\hspace{0pt}k}/{({k - 2})}}}`$ in Eq. ([25](#S4.E25 "In Proof. ‣ IV Analysis of one-layer architectures ‣ Detailed Network Measurements Using Sparse Graph Counters: The Theory")), thus getting
<table>
<tbody>
<tr>
<td></td>
<td>ℙ{ℍz=0}ℙℍ𝑧0\displaystyle\mathbb{P}\{{\mathbb{H}}z=0\}</td>
<td>≤\displaystyle\leq</td>
<td>(Akℓ)q−12(Akℓ/m)kℓ/q.superscript𝐴𝑘ℓ𝑞12superscript𝐴𝑘ℓ𝑚𝑘ℓ𝑞\displaystyle(Ak\ell)^{\frac{q-1}{2}}(Ak\ell/m)^{k\ell/q}\,.</td>
<td></td>
<td>(27)</td>
</tr>
</tbody>
</table>
Since $`{({k\hspace{0pt}\ell})}^{{({q - 1})}/2} \leq A^{{k\hspace{0pt}\ell}/q}`$, we have (by properly adjusting $`A`$)
<table>
<tbody>
<tr>
<td></td>
<td>ℙ{ℍz=0}ℙℍ𝑧0\displaystyle\mathbb{P}\{{\mathbb{H}}z=0\}</td>
<td>≤\displaystyle\leq</td>
<td>(Akℓ/m)kℓ/q.superscript𝐴𝑘ℓ𝑚𝑘ℓ𝑞\displaystyle(Ak\ell/m)^{k\ell/q}\,.</td>
<td></td>
<td>(28)</td>
</tr>
</tbody>
</table>
For $`{\| z\|} > E_{0}`$, we use Eq. (LABEL:eq:BoundProb) with $`{R\hspace{0pt}{(\cdots)}} \leq 1`$. Since $`{{k\hspace{0pt}\ell}/m} > {{k\hspace{0pt}E_{0}}/m}`$ is bounded away from $`0`$ by Eq. ([26](#S4.E26 "In Proof. ‣ IV Analysis of one-layer architectures ‣ Detailed Network Measurements Using Sparse Graph Counters: The Theory")), we have $`{Q\hspace{0pt}{(\cdots)}} \leq e^{- C}`$ for some $`C = {C\hspace{0pt}{(\Delta,q)}} > 0`$ and therefore
<table>
<tbody>
<tr>
<td></td>
<td>ℙ{ℍz=0}ℙℍ𝑧0\displaystyle\mathbb{P}\{{\mathbb{H}}z=0\}</td>
<td>≤\displaystyle\leq</td>
<td>(Akℓ)(q−1)/2e−Cm.superscript𝐴𝑘ℓ𝑞12superscript𝑒𝐶𝑚\displaystyle(Ak\ell)^{(q-1)/2}\,e^{-Cm}\,.</td>
<td></td>
<td>(29)</td>
</tr>
</tbody>
</table>
There are at most $`{\binom{n}{\ell}\hspace{0pt}{({q - 1})}^{\ell}} \leq \left( \frac{A\hspace{0pt}n}{\ell} \right)^{\ell}`$ vectors $`z`$ with $`{\| z\|}_{0} = \ell`$. If we denote by $`{\mathbb{P}}_{E_{1},E_{2}}`$ the probability of the event $`\left\{ {\exists z}:{{E_{1} \leq {\| z\|}_{0} \leq E_{2}},{{{\mathbb{H}}\hspace{0pt}z} = {0\operatorname{mod}q}}} \right\}`$, the probability in Eq. ([24](#S4.E24 "In Lemma 6. ‣ IV Analysis of one-layer architectures ‣ Detailed Network Measurements Using Sparse Graph Counters: The Theory")) is upper bounded by $`{\mathbb{P}}_{2,E_{0}} + {\mathbb{P}}_{E_{0},E}`$ (notice that if $`k`$ is not a multiple of $`q`$, $`{{\mathbb{H}}\hspace{0pt}z} = 0`$ is impossible for $`{\| z\|}_{0} = 1`$). By union bound we have
<table>
<tbody>
<tr>
<td></td>
<td>ℙ2,E0subscriptℙ2subscript𝐸0\displaystyle\mathbb{P}_{2,E_{0}}</td>
<td>≤\displaystyle\leq</td>
<td>∑ℓ=2E0(Anℓ)ℓ(Akℓm)kℓ/qsuperscriptsubscriptℓ2subscript𝐸0superscript𝐴𝑛ℓℓsuperscript𝐴𝑘ℓ𝑚𝑘ℓ𝑞\displaystyle\sum_{\ell=2}^{E_{0}}\left(\frac{An}{\ell}\right)^{\ell}\left(\frac{Ak\ell}{m}\right)^{k\ell/q}</td>
<td></td>
</tr>
<tr>
<td></td>
<td></td>
<td>≤\displaystyle\leq</td>
<td>(An2)2(2Akm)2k/q∑ℓ=2E0ξ(ℓ)ℓ−2,superscript𝐴𝑛22superscript2𝐴𝑘𝑚2𝑘𝑞superscriptsubscriptℓ2subscript𝐸0𝜉superscriptℓℓ2\displaystyle\left(\frac{An}{2}\right)^{2}\left(\frac{2Ak}{m}\right)^{2k/q}\sum_{\ell=2}^{E_{0}}\xi(\ell)^{\ell-2}\,,</td>
<td></td>
</tr>
</tbody>
</table>
where (using $`{({\ell/2})}^{{{2\hspace{0pt}k}/q} - 2} \leq A^{{k\hspace{0pt}{({\ell - 2})}}/q}`$ and eventually adjusting the constant $`A`$)
<table>
<tbody>
<tr>
<td></td>
<td>ξ(ℓ)≡nℓ(Akℓm)k/q.𝜉ℓ𝑛ℓsuperscript𝐴𝑘ℓ𝑚𝑘𝑞\displaystyle\xi(\ell)\equiv\frac{n}{\ell}\left(\frac{Ak\ell}{m}\right)^{k/q}\,.</td>
<td></td>
</tr>
</tbody>
</table>
For $`\ell \leq E_{0}`$, and choosing $`\rho\hspace{0pt}{(q)}`$ small enough in Eq. ([25](#S4.E25 "In Proof. ‣ IV Analysis of one-layer architectures ‣ Detailed Network Measurements Using Sparse Graph Counters: The Theory")), we obtain $`{\xi\hspace{0pt}{(\ell)}} \leq {1/2}`$ thus leading to $`{\mathbb{P}}_{2,E_{0}} \leq {n^{2}\hspace{0pt}{({{A\hspace{0pt}k}/m})}^{k/q}}`$.
Finally consider the contribution of vectors $`{\| z\|}_{0} \geq E_{0}`$. Proceeding as above, we have
<table>
<tbody>
<tr>
<td></td>
<td>ℙE0,Esubscriptℙsubscript𝐸0𝐸\displaystyle\mathbb{P}_{E_{0},E}</td>
<td>≤\displaystyle\leq</td>
<td>∑ℓ=E0E(Anℓ)ℓ(Akℓ)(q−1)/2e−Cmsuperscriptsubscriptℓsubscript𝐸0𝐸superscript𝐴𝑛ℓℓsuperscript𝐴𝑘ℓ𝑞12superscript𝑒𝐶𝑚\displaystyle\sum_{\ell=E_{0}}^{E}\left(\frac{An}{\ell}\right)^{\ell}(Ak\ell)^{(q-1)/2}\,e^{-Cm}</td>
<td></td>
</tr>
<tr>
<td></td>
<td></td>
<td>≤\displaystyle\leq</td>
<td>E(AnE)E(AkE)(q−1)/2e−Cm.𝐸superscript𝐴𝑛𝐸𝐸superscript𝐴𝑘𝐸𝑞12superscript𝑒𝐶𝑚\displaystyle E\left(\frac{An}{E}\right)^{E}(AkE)^{(q-1)/2}\,e^{-Cm}\,.</td>
<td></td>
</tr>
</tbody>
</table>
Here we bounded $`{({{A\hspace{0pt}n}/\ell})}^{\ell} = {\lbrack{({{A\hspace{0pt}n}/\ell})}^{{\ell/A}\hspace{0pt}n}\rbrack}^{A\hspace{0pt}n} \leq {({{A\hspace{0pt}n}/E})}^{E}`$, using the fact that $`x^{- x}`$ is an increasing function of $`x`$ for $`x \leq e^{- 1}`$, and that $`{{E/A}\hspace{0pt}n} = {{{C\hspace{0pt}m}/A}\hspace{0pt}n\hspace{0pt}{\log{({{n\hspace{0pt}k}/m})}}} \leq {{{C\hspace{0pt}m}/A}\hspace{0pt}n}`$ is smaller than $`e^{- 1}`$ for $`C`$ small enough.
Finally we bound $`E^{{({q + 1})}/2} \leq A^{E}`$ and $`k^{{({q - 1})}/2} \leq k^{E}`$ (which holds for $`m`$ large enough), thus getting
<table>
<tbody>
<tr>
<td></td>
<td>ℙE0,Esubscriptℙsubscript𝐸0𝐸\displaystyle\mathbb{P}_{E_{0},E}</td>
<td>≤\displaystyle\leq</td>
<td>(nkAm)Ee−Cm.superscript𝑛𝑘𝐴𝑚𝐸superscript𝑒𝐶𝑚\displaystyle\left(\frac{nkA}{m}\right)^{E}\,e^{-Cm}\,.</td>
<td></td>
</tr>
</tbody>
</table>
If we take $`E = {{{C\hspace{0pt}m}/2}\hspace{0pt}{\log{({{n\hspace{0pt}k\hspace{0pt}A}/m})}}}`$, we get $`{\mathbb{P}}_{E_{0},E_{1}} \leq e^{- {{C\hspace{0pt}m}/2}}`$, which is smaller than $`{({{B\hspace{0pt}k}/m})}^{\frac{k}{q}}`$ for a properly chosen constant $`B`$ and $`k \leq {m/{\log m}}`$ (indeed $`k \leq {m\hspace{0pt}\varepsilon_{m}}`$ would be enough for any $`\varepsilon_{m} \downarrow 0`$.) ∎
## V Analysis of multi-layer architectures and proof of Theorem [1](#Thmthm1 "Theorem 1. ‣ II-A Main results ‣ II Counter Braids: Basic definitions ‣ Detailed Network Measurements Using Sparse Graph Counters: The Theory")
###### Proof.
Let $`P_{err}^{(l)}`$ denote the probability that $`l`$-th term in the $`q`$-ary expansion of $`x`$ is decoded incorrectly by the decoder in Section [](#S3.SS2 "III-B Recovery function ‣ III The architecture ‣ Detailed Network Measurements Using Sparse Graph Counters: The Theory")
III-B
(i.e. that $`{\hat{x}\hspace{0pt}{(l)}} \neq {x\hspace{0pt}{(l)}}`$) given that $`x\hspace{0pt}{(0)}`$, …, $`x\hspace{0pt}{({l - 1})}`$ have been correctly recovered. We will prove that $`P_{err}^{(l)} = {O\hspace{0pt}{(n^{- A})}}`$ for some $`A > 0`$. Since the multi-layer architecture involves at most $`C\hspace{0pt}{({\log n})}^{\frac{3}{2}}`$ layers, this implies the thesis. Further, we shall consider only the first $`L_{0}`$ layers, since it will be clear from the derivation below that the error probability is decreasing for the last $`L_{1}`$ layers.
Let $`x`$ be the input. Since we are focusing on the $`l`$-th term in the $`q`$-ary expansion of the input, we will drop the index $`l`$, and take $`x \in {\{ 0,\ldots,{q - 1}\}}^{n}`$. This is just a vector whose entries are iid with distribution $`p_{l}`$.
The error probability $`P_{err}^{(l)}`$ is upper bounded by the probability that $`x \notin {\mathsf{T}_{n}\hspace{0pt}{(p_{l})}}`$ plus the probability that there exists $`x' \neq x`$ with $`{{\mathbb{H}}^{l}\hspace{0pt}x'} = {{\mathbb{H}}^{l}\hspace{0pt}x}`$ $`\operatorname{mod}q`$. The first contribution is bounded by Lemma [4](#Thmlemma4 "Lemma 4. ‣ III-C Sparse graph ensemble and choice of the parameters ‣ III The architecture ‣ Detailed Network Measurements Using Sparse Graph Counters: The Theory"), and we can therefore neglect it. Denoting the second contribution as $`P_{err}^{(l, \ast )}`$, and writing $`{\mathbb{E}}_{x}`$, $`\mathbb{P}`$ for (respectively) expectation with respect to $`x`$ and probability with respect to the matrices $`{\mathbb{H}}_{1}`$, …$`{\mathbb{H}}_{l}`$, we have (matrix multiplications below are understood to be modulo $`q`$)
<table>
<tbody>
<tr>
<td></td>
<td>Perr(l,∗)superscriptsubscriptPerr𝑙\displaystyle{\rm P}_{{\rm\footnotesize err}}^{(l,*)}</td>
<td>=\displaystyle=</td>
<td>𝔼xℙ{∃x′∈𝖳n(pl)∖{x} s.t. ℍlx′=ℍlx}subscript𝔼𝑥ℙsuperscript𝑥′subscript𝖳𝑛subscript𝑝𝑙𝑥 s.t. superscriptℍ𝑙superscript𝑥′superscriptℍ𝑙𝑥\displaystyle\mathbb{E}_{x}\mathbb{P}\left\{\exists x^{\prime}\in{\sf T}_{n}(p_{l})\setminus\{x\}\mbox{ s.t. }{\mathbb{H}}^{l}x^{\prime}={\mathbb{H}}^{l}x\right\}</td>
<td></td>
</tr>
<tr>
<td></td>
<td></td>
<td>=\displaystyle=</td>
<td>∑t=1lQt(l),superscriptsubscript𝑡1𝑙subscriptsuperscript𝑄𝑙𝑡\displaystyle\sum_{t=1}^{l}Q^{(l)}_{t}\,,</td>
<td></td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td>Qt(l)subscriptsuperscript𝑄𝑙𝑡\displaystyle Q^{(l)}_{t}</td>
<td>≡\displaystyle\equiv</td>
<td>𝔼xℙ{∃x′∈𝖳n(pl)∖{x} s.t.\displaystyle\mathbb{E}_{x}\mathbb{P}\big{\{}\exists x^{\prime}\in{\sf T}_{n}(p_{l})\setminus\{x\}\mbox{ s.t. }</td>
<td></td>
</tr>
<tr>
<td></td>
<td></td>
<td></td>
<td>ℍtx′=ℍtx,ℍt−1x′≠ℍt−1x}.\displaystyle\phantom{\mathbb{E}_{x}\mathbb{P}\big{\{}\exists x^{\prime}\in}{\mathbb{H}}^{t}x^{\prime}={\mathbb{H}}^{t}x,\,{\mathbb{H}}^{t-1}x^{\prime}\neq{\mathbb{H}}^{t-1}x\big{\}}\,.</td>
<td></td>
</tr>
</tbody>
</table>
Since, $`l \leq L = {O\hspace{0pt}{({\log n})}}`$, it is sufficient to show $`Q_{t}^{(l)} = {O\hspace{0pt}{(n^{- A})}}`$. In $`Q_{t}^{(l)}`$ we can separate error events due to input $`x'`$ such that $`d_{t} \equiv {d\hspace{0pt}{({{\mathbb{H}}^{t - 1}\hspace{0pt}x'},{{\mathbb{H}}^{t - 1}\hspace{0pt}x})}} \leq E`$ and the other ones. As a consequence $`Q_{t}^{(l)}`$ is upper bounded by
<table>
<tbody>
<tr>
<td></td>
<td>𝔼xsubscript𝔼𝑥\displaystyle\mathbb{E}_{x}</td>
<td>ℙ{∃x′∈𝖳n(pl), s.t. 1≤dt≤E,ℍtx′=ℍtx}\displaystyle\mathbb{P}\left\{\exists x^{\prime}\in{\sf T}_{n}(p_{l}),\mbox{ s.t. }1\leq d_{t}\leq E,\;{\mathbb{H}}^{t}x^{\prime}={\mathbb{H}}^{t}x\right\}</td>
<td></td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td>+\displaystyle+</td>
<td>𝔼xℙ{∃x′∈𝖳n(pl) s.t. E<dt,ℍlx′=ℍlx}\displaystyle\mathbb{E}_{x}\mathbb{P}\left\{\exists x^{\prime}\in{\sf T}_{n}(p_{l})\mbox{ s.t. }E<d_{t},\,{\mathbb{H}}^{l}x^{\prime}={\mathbb{H}}^{l}x\right\}</td>
<td></td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td></td>
<td>≤ℙ{∃z s.t. ||z||0≤E,ℍtz=0}absentℙconditional-set𝑧 s.t. formulae-sequenceevaluated-at𝑧0𝐸subscriptℍ𝑡𝑧0\displaystyle\leq\mathbb{P}\{\exists z\mbox{ s.t. }||z||_{0}\leq E\,,{\mathbb{H}}_{t}z=0\}</td>
<td></td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td></td>
<td>+|𝖳n(pl)|sup{ℙ{ℍtz=0}:‖z‖0>E}.subscript𝖳𝑛subscript𝑝𝑙supremumconditional-setℙsubscriptℍ𝑡𝑧0subscriptnorm𝑧0𝐸\displaystyle\phantom{aaaa}+|{\sf T}_{n}(p_{l})|\sup\big{\{}\mathbb{P}\{{\mathbb{H}}_{t}z=0\}\,:\,||z||_{0}>E\big{\}}.</td>
<td></td>
</tr>
</tbody>
</table>
Here $`z`$ is understood to be a $`m_{t - 1}`$ dimensional vector with entries in $`\{ 0,\ldots,{q - 1}\}`$.
Notice that $`{({{m_{t}/k}\hspace{0pt}m_{t - 1}})}^{1/k} \geq {({\delta/k})}^{1/k} \geq \delta`$. Next we choose $`E = {{C\hspace{0pt}{({{q,\Delta} = \delta})}\hspace{0pt}m_{t}}/{\log{({{m_{t - 1}\hspace{0pt}k}/m_{t}})}}}`$ with $`C\hspace{0pt}{(q,\Delta)}`$ as in the statement of Lemma [6](#Thmlemma6 "Lemma 6. ‣ IV Analysis of one-layer architectures ‣ Detailed Network Measurements Using Sparse Graph Counters: The Theory"). As a consequence the first term above is upper bounded by
<table>
<tbody>
<tr>
<td></td>
<td>mt−12(Bkmt)kq≤(Bk)kqδ−2mt−kq+2≤C(logn)kq−2n−kq+2,superscriptsubscript𝑚𝑡12superscript𝐵𝑘subscript𝑚𝑡𝑘𝑞superscript𝐵𝑘𝑘𝑞superscript𝛿2superscriptsubscript𝑚𝑡𝑘𝑞2𝐶superscript𝑛𝑘𝑞2superscript𝑛𝑘𝑞2\displaystyle m_{t-1}^{2}\left(\frac{Bk}{m_{t}}\right)^{\frac{k}{q}}\leq(Bk)^{\frac{k}{q}}\delta^{-2}m_{t}^{-{\frac{k}{q}}+2}\leq C(\log n)^{{\frac{k}{q}}-2}n^{-{\frac{k}{q}}+2}\,,</td>
<td></td>
</tr>
</tbody>
</table>
where we used $`m_{t - 1} \leq {m_{t}/\delta}`$ and $`m_{t} \geq {n/{({\log n})}^{2}}`$. The constant $`C`$ that depends uniquely on $`q`$, $`k`$, $`\delta`$, but not on $`n`$.
It remains to bound the second contribution, due to inputs $`x'`$ with $`{d\hspace{0pt}{(x',x)}} > E`$. Using Lemma [4](#Thmlemma4 "Lemma 4. ‣ III-C Sparse graph ensemble and choice of the parameters ‣ III The architecture ‣ Detailed Network Measurements Using Sparse Graph Counters: The Theory") (to bound $`\mathsf{T}_{n}\hspace{0pt}{(p_{l})}`$) and [5](#Thmlemma5 "Lemma 5. ‣ IV Analysis of one-layer architectures ‣ Detailed Network Measurements Using Sparse Graph Counters: The Theory") (to bound $`{\mathbb{P}}\hspace{0pt}{\{{{{\mathbb{H}}_{t}\hspace{0pt}z} = 0}\}}`$ for $`{\| z\|}_{0} > E`$)
<table>
<tbody>
<tr>
<td></td>
<td>𝔼xℙ{∃x′∈𝖳n(pl) s.t. E<dt,ℍlx′=ℍlx}\displaystyle\mathbb{E}_{x}\mathbb{P}\left\{\exists x^{\prime}\in{\sf T}_{n}(p_{l})\mbox{ s.t. }E<d_{t},\,{\mathbb{H}}^{l}x^{\prime}={\mathbb{H}}^{l}x\right\}</td>
<td></td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td>≤qnhl+Anβ(Ckn)q−12{1q[1+(q−1)e−DkE/mt]}mt,absentsuperscript𝑞𝑛subscriptℎ𝑙𝐴superscript𝑛𝛽superscript𝐶𝑘𝑛𝑞12superscript1𝑞delimited-[]1𝑞1superscript𝑒𝐷𝑘𝐸subscript𝑚𝑡subscript𝑚𝑡\displaystyle\leq q^{nh_{l}+An^{\beta}}(Ckn)^{\frac{q-1}{2}}\left\{\frac{1}{q}[1+(q-1)e^{-DkE/m_{t}}]\right\}^{m_{t}}\,,</td>
<td></td>
</tr>
</tbody>
</table>
By eventually enlarging the constant $`A`$ (in a way that depends on $`q`$), we can get rid of the term $`{({C\hspace{0pt}n})}^{\frac{q - 1}{2}}`$. By further using $`{({1 + x})} \leq q^{x/{\log q}}`$ we can upper bound the above by $`k^{q - {1/2}}\hspace{0pt}q^{\Phi}`$ where
<table>
<tbody>
<tr>
<td></td>
<td>Φ=nhl+A(q)nβ−mt+A′(q)mte−D(q)kE/mtΦ𝑛subscriptℎ𝑙𝐴𝑞superscript𝑛𝛽subscript𝑚𝑡superscript𝐴′𝑞subscript𝑚𝑡superscript𝑒𝐷𝑞𝑘𝐸subscript𝑚𝑡\displaystyle\Phi=nh_{l}+A(q)n^{\beta}-m_{t}+A^{\prime}(q)m_{t}e^{-D(q)kE/m_{t}}</td>
<td></td>
</tr>
</tbody>
</table>
with $`{A'\hspace{0pt}{(q)}} = {{({q - 1})}/{\log q}}`$. Notice that $`{{k\hspace{0pt}E}/m_{t}} = {{C\hspace{0pt}{(q,\delta)}\hspace{0pt}k}/{\log{({{k\hspace{0pt}m_{t - 1}}/m_{t}})}}}`$ can be made arbitrarily large by taking $`k`$ large enough. In particular, we can choose $`k_{\ast}\hspace{0pt}{(q,\delta)}`$ such that $`{A'\hspace{0pt}{(q)}\hspace{0pt}e^{- {{D\hspace{0pt}{(q)}\hspace{0pt}k\hspace{0pt}E}/m_{t}}}} \leq {\delta/3}`$ for any $`k \geq k_{\ast}`$. For such $`k`$, and using the fact that $`m_{t} \geq m_{l} = {{\lbrack{{n\hspace{0pt}h_{l}} + {A\hspace{0pt}{(q)}\hspace{0pt}n^{\beta}}}\rbrack}\hspace{0pt}{({1 + \delta})}}`$
<table>
<tbody>
<tr>
<td></td>
<td>Φ≤nhl+A(q)nβ−ml(1−δ/3)≤−13δ[nhl+A(q)nβ].Φ𝑛subscriptℎ𝑙𝐴𝑞superscript𝑛𝛽subscript𝑚𝑙1𝛿313𝛿delimited-[]𝑛subscriptℎ𝑙𝐴𝑞superscript𝑛𝛽\displaystyle\Phi\leq nh_{l}+A(q)n^{\beta}-m_{l}(1-\delta/3)\leq-\frac{1}{3}\delta[nh_{l}+A(q)n^{\beta}]\,.</td>
<td></td>
</tr>
</tbody>
</table>
Summing the various contributions, we obtain, for any $`k \geq {k_{\ast}\hspace{0pt}{(q,\delta)}}`$
<table>
<tbody>
<tr>
<td></td>
<td>Qt(l)≤C(q,k,δ)(logn)kq−2n−kq+2+kq−12q−δ(A(q)nβ+nhl)/3,subscriptsuperscript𝑄𝑙𝑡𝐶𝑞𝑘𝛿superscript𝑛𝑘𝑞2superscript𝑛𝑘𝑞2superscript𝑘𝑞12superscript𝑞𝛿𝐴𝑞superscript𝑛𝛽𝑛subscriptℎ𝑙3\displaystyle Q^{(l)}_{t}\leq C(q,k,\delta)(\log n)^{{\frac{k}{q}}-2}n^{-{\frac{k}{q}}+2}+k^{\frac{q-1}{2}}q^{-\delta(A(q)n^{\beta}+nh_{l})/3}\,,</td>
<td></td>
<td>(30)</td>
</tr>
</tbody>
</table>
which proves the thesis. ∎
## VI Some auxiliary results
###### Proof: Lemma [3](#Thmlemma3 "Lemma 3. ‣ III-C Sparse graph ensemble and choice of the parameters ‣ III The architecture ‣ Detailed Network Measurements Using Sparse Graph Counters: The Theory").
First consider Eq. ([12](#S3.E12 "In Lemma 3. ‣ III-C Sparse graph ensemble and choice of the parameters ‣ III The architecture ‣ Detailed Network Measurements Using Sparse Graph Counters: The Theory")). Let $`X_{1}`$ be an integer random variable with distribution $`p`$, $`X_{1}\hspace{0pt}{(l)}`$ its $`l`$-th least significant $`q`$-ary digit and $`Z`$ the indicator function on $`{X_{1}\hspace{0pt}{(l)}} \geq 0`$. From $`{H\hspace{0pt}{({X_{1}\hspace{0pt}{(l)}})}} = {{H\hspace{0pt}{(Z)}} + {H\hspace{0pt}{(\left. {X_{1}\hspace{0pt}{(l)}} \middle| Z \right.)}}}`$ it follows that, for $`{\overline{p}}_{l} = {{\mathbb{P}}\hspace{0pt}{\{{X_{1} > q^{l}}\}}}`$:
<table>
<tbody>
<tr>
<td></td>
<td>hl≤p¯llogq(q−1)−p¯llogqp¯l−(1−p¯l)logq(1−p¯l).subscriptℎ𝑙subscript¯𝑝𝑙subscript𝑞𝑞1subscript¯𝑝𝑙subscript𝑞subscript¯𝑝𝑙1subscript¯𝑝𝑙subscript𝑞1subscript¯𝑝𝑙\displaystyle h_{l}\leq\overline{p}_{l}\log_{q}(q-1)-\overline{p}_{l}\log_{q}\overline{p}_{l}-(1-\overline{p}_{l})\log_{q}(1-\overline{p}_{l})\,.</td>
<td></td>
</tr>
</tbody>
</table>
Choosing $`q`$ large enough so that $`{\overline{p}}_{l} \leq {A\hspace{0pt}q^{- \epsilon}} \leq {1/2}`$ for all $`l \geq 1`$, we can upper bound $`- {{({1 - {\overline{p}}_{l}})}\hspace{0pt}{\log_{q}{({1 - {\overline{p}}_{l}})}}}`$ by $`2\hspace{0pt}{\overline{p}}_{l}`$, thus getting
<table>
<tbody>
<tr>
<td></td>
<td>hl≤3p¯l−p¯llogqp¯l,subscriptℎ𝑙3subscript¯𝑝𝑙subscript¯𝑝𝑙subscript𝑞subscript¯𝑝𝑙\displaystyle h_{l}\leq 3\overline{p}_{l}-\overline{p}_{l}\log_{q}\overline{p}_{l}\,,</td>
<td></td>
</tr>
</tbody>
</table>
which implies Eq. ([12](#S3.E12 "In Lemma 3. ‣ III-C Sparse graph ensemble and choice of the parameters ‣ III The architecture ‣ Detailed Network Measurements Using Sparse Graph Counters: The Theory")) for $`{\overline{p}}_{l} \leq {A\hspace{0pt}q^{- {l\hspace{0pt}\epsilon}}}`$.)
In order to prove Eq. ([13](#S3.E13 "In Lemma 3. ‣ III-C Sparse graph ensemble and choice of the parameters ‣ III The architecture ‣ Detailed Network Measurements Using Sparse Graph Counters: The Theory")), first notice that $`{H\hspace{0pt}{(X_{1})}} = {H\hspace{0pt}{({\{{X_{1}\hspace{0pt}{(l)}}\}})}} \leq {\sum_{l \geq 0}{H\hspace{0pt}{({X_{1}\hspace{0pt}{(l)}})}}}`$ whence $`{h_{2}\hspace{0pt}{(p)}} \leq {\sum_{l \geq 0}{h_{l}\hspace{0pt}{\log_{2}q}}}`$. By the same argument $`{h_{2}\hspace{0pt}{(p)}} \geq {h_{0}\hspace{0pt}{\log_{2}q}}`$. The thesis follows by bounding $`\sum_{l \geq 0}h_{l}`$ using Eq. ([12](#S3.E12 "In Lemma 3. ‣ III-C Sparse graph ensemble and choice of the parameters ‣ III The architecture ‣ Detailed Network Measurements Using Sparse Graph Counters: The Theory")). ∎
###### Proof: Lemma [4](#Thmlemma4 "Lemma 4. ‣ III-C Sparse graph ensemble and choice of the parameters ‣ III The architecture ‣ Detailed Network Measurements Using Sparse Graph Counters: The Theory")..
The number of vectors with type $`\theta`$ is upper bounded by $`q^{n\hspace{0pt}H\hspace{0pt}{(\theta)}}`$. Since there are at most $`{({n + 1})}^{q}`$ distinct types, $`{|{\mathsf{T}_{n}\hspace{0pt}{(p_{\ast})}}|} \leq q^{{n\hspace{0pt}H\hspace{0pt}{(p_{\ast})}} + {n\hspace{0pt}K_{n}}}`$ where
<table>
<tbody>
<tr>
<td></td>
<td>Kn≡supθ{H(θ)−H(p∗):D(θ||p∗)≤n−γ}+logq(n+1)qn.\displaystyle K_{n}\equiv\sup_{\theta}\{H(\theta)-H(p_{*}):D(\theta||p_{*})\leq n^{-\gamma}\}+\frac{\log_{q}(n+1)^{q}}{n}\,.</td>
<td></td>
</tr>
</tbody>
</table>
The bound $`{{H\hspace{0pt}{(\theta)}} - {H\hspace{0pt}{(p_{\ast})}}} \leq {{\|{\theta - p_{\ast}}\|}_{1}\hspace{0pt}{\log{({q/{\|{\theta - p_{\ast}}\|}})}}}`$ and $`{\|{\theta - p_{\ast}}\|} \leq \sqrt{2D{(\theta||p_{\ast})}}`$ \[[3](#bib.bib3)\].
Equation ([14](#S3.E14 "In Lemma 4. ‣ III-C Sparse graph ensemble and choice of the parameters ‣ III The architecture ‣ Detailed Network Measurements Using Sparse Graph Counters: The Theory")) is just Sanov Theorem. ∎
## VII Acknoledgments
Yi Lu is supported by the Cisco Stanford Graduate Fellowship.
## References
- \[1\] S. Dharmapurikar, A. Kabbani, Y. Lu, A. Montanari and B. Prabhakar. “Passing Messages Through Counter Braids: A Novel Approach to Traffic Measurement.” Technical Report, TR06-ISL012201, March, 2007.
- \[2\]
C. Estan and G. Varghese. “New Directions in Traffic Measurement and Accounting: Focusing on the Elephants, Ignoring the Mice.” ACM Trans. on Comp. Syst., 21:270–313, 2003.
- \[3\] T. Cover and J. A. Thomas, *Elements of Information Theory*, Wiley Interscience, New York, 1991
- \[4\]
Robert G. Gallager. Low-Density Parity-Check Codes. MIT Press, Cambridge, Massachussetts, 1963.
- \[5\]
T. Richardson and R. Urbanke, Modern Coding Theory, draft available at http://lthcwww.epfl.ch/mct/index.php
- \[6\]
G. Caire, S. Shamai, and S. Verdú. “Noiseless data compression with low density parity check codes.” In P. Gupta and G. Kramer, editors, Dimacs Series in Mathematics and Theoretical Computer Science, pages 224–235. AMS, 2004.
- \[7\]
E. Berlekamp, R. J. McEliecee, and H. C.A. van Tilborg. “On the inherent intractability of certain coding problems.” IEEE Trans. Inform. Theory, IT-29:384–386, 1978.
- \[8\]
F. R. Kschischang, B. J. Frey and H-A. Loeliger, “Factor graphs and the sum-product algorithm” (2001), IEEE Trans. Inform. Theory 47, 498-519.
- \[9\]
S. Aji, H. Jin, A. Khandekar, D. J.C. MacKay, and R. J. McEliece. “BSC Thresholds For Code Ensembles Based on ‘Typical Pairs’ Decoding.” In Brian Marcus and Joachim Rosenthal, editors, Codes, Systems and Graphical Models, pages 195–210. Springer, 2001.
- \[10\] G. Miller and D. Burshtein, “Asymptotic enumeration method for analyzing LDPC codes,” *IEEE Trans. Inform. Theory*, vol. 50, no. 6, pp. 1115–1131, June 2004.
- \[11\] S. L. Litsyn and V. S. Shevelev, “On ensembles of low-density parity-check codes: asymptotic distance distributions,” *IEEE Trans. Inform. Theory*, vol. IT–48, pp. 887 –908, Apr. 2002.
- \[12\]
A. Bennatan and D. Burshtein. “On the application of LDPC Codes to Arbitrary Discrete-Memoryless Channels.” IEEE Trans. Inform. Theory, 50:417–438, 2004. |
0710.5002 | cs.CR | cyber | # The entropy of keys derived from laser speckle
B. Škorić
Philips Research Europe
Abstract
Laser speckle has been proposed in a number of papers as a high-entropy source of unpredictable bits for use in security applications. Bit strings derived from speckle can be used for a variety of security purposes such as identification, authentication, anti-counterfeiting, secure key storage, random number generation and tamper protection. The choice of laser speckle as a source of random keys is quite natural, given the chaotic properties of speckle. However, this same chaotic behaviour also causes reproducibility problems. Cryptographic protocols require either zero noise or very low noise in their inputs; hence the issue of error rates is critical to applications of laser speckle in cryptography. Most of the literature uses an error reduction method based on Gabor filtering. Though the method is successful, it has not been thoroughly analysed.
In this paper we present a statistical analysis of Gabor-filtered speckle patterns. We introduce a model in which perturbations are described as random phase changes in the source plane. Using this model we compute the second and fourth order statistics of Gabor coefficients. We determine the mutual information between perturbed and unperturbed Gabor coefficients and the bit error rate in the derived bit string. The mutual information provides an absolute upper bound on the number of secure bits that can be reproducibly extracted from noisy measurements.
Keywords: Physical Unclonable Function, PUF, speckle, Gabor transform, entropy, key extraction, fuzzy extractor
## 1 Introduction
### 1.1 Key generation from speckle patterns
In \[[11](#bib.bib11), [12](#bib.bib12)\] Pappu et al. proposed to use speckle patterns obtained from coherent multiple scattering in a token to authenticate persons and devices. In a typical scenario, a person carries an authentication token consisting of a transparent material with scattering particles inside, e.g. glass with air bubbles. When he wants to get access to some service, he presents his token to a reader device. The device shines laser light onto the token under some predetermined conditions (wave length, angle, focal distance, beam shape etc). This is called a ‘challenge’. The resulting speckle pattern (in transmission or reflection) under some predetermined angle is recorded by the device. The recorded image is called the ‘response’. The response is processed, e.g. by Gabor filtering, to yield a bit string that is reasonably insensitive to noise in the image. This bit string is compared to a previously enrolled bit string. If the strings are sufficiently similar, the token is authenticated. (A variant of this procedure, involving a scan over a length of paper, was developed in \[[1](#bib.bib1)\] for the authentication of paper documents.)
Often one cannot trust the reader device and/or the link between the reader and the verifier. In that case a Challenge-Response Pair (CRP) cannot be used safely more than once in the above scenario. In order to have a secure and practical token, it must be possible to obtain many different CRPs from one token. It must also be very hard to predict a CRP given previously observed CRPs. Further security requirements follow if one demands that it must be hard for an attacker to (i) extract all CRPs from a token in a short amount of time, and (ii) to extract enough information from the token either to physically clone it or to successfully compute its responses. Pappu introduced the name ‘PUF’ for a token (not necessarily optical) that satisfies all these security requirements. PUF stands for Unclonable Physical Function. Alternative names in the literature are Physical One-Way Function (POWF) and Physical Random Function. The word function stems from the fact that a response can be regarded as the evaluation of a complicated function of the argument; the function is parametrised by the physical structure of the token. It turns out that the physics of multiple scattering is compatible with all PUF requirements, especially if the token is created by a random mixing procedure of sufficiently small particles.
Going one step further than the simple matching procedure of \[[11](#bib.bib11), [12](#bib.bib12)\], it is possible to use a token’s response as a secret key in a cryptographic protocol. This is nontrivial, since any amount of noise is fatal to ordinary cryptographic primitives. Secure forms of error correction, in which the redundancy data does not leak (much) information on the secret key, were developed in \[[9](#bib.bib9), [10](#bib.bib10), [4](#bib.bib4)\]. These techniques are called fuzzy extractors or helper data schemes. Their application to optical PUFs was studied in \[[17](#bib.bib17), [13](#bib.bib13), [14](#bib.bib14)\]. Key generation from CRP measurements is an enabler for a wide variety of security applications such as authentication, brand protection, tamper protection, anti-counterfeiting, secure key storage and special forms of authenticated computation \[[6](#bib.bib6)\]. For an overview of the subject of security with noisy data we refer to \[[15](#bib.bib15)\].
In all of these examples, it is important to have a good understanding of the number of random bits that can be extracted from the measurements. Overestimation can lead to serious cryptographic weaknesses. Underestimation leads to waste of resources. A general framework for the computation of measurement entropy was set up in \[[16](#bib.bib16)\] and applied to transmissive optical PUFs.
A different approach was taken in \[[8](#bib.bib8)\], where Gabor-filtered speckle patterns were compressed using the Context Tree Weighting (CTW) method. The size of the compressed data gives an upper bound on the entropy.
A second important point is a good understanding of the noise that occurs in the response when the same challenge is applied multiple times. This noise determines how much of the total entropy of a response can be extracted in a reproducible way. Measurement noise is caused by many factors: temperature, moisture, stray light, mechanical misalignment, differences between reader devices, ageing etc. In \[[17](#bib.bib17)\] several methods were proposed to deal with noise in optical PUFs, e.g. alignment methods and efficient protocols. In \[[8](#bib.bib8)\] CTW compression was employed to estimate the mutual information between two (Gabor-filtered) noisy measurements of the same response. This information-theoretic quantity captures the shared entropy between two data sets and gives an upper limit on the length of the shared key that can be reproducibly extracted from these sets.
That work has resulted in a lot of practical know-how, sufficient to set up a secure key extraction system. What is lacking, however, is a theoretical understanding of the effects of measurement noise on the Gabor coefficients. In this paper we will address the issue of random perturbations and the statistical properties of Gabor-transformed speckle patterns.
### 1.2 Binarized Gabor coefficients
Bit strings can be extracted effectively by using a Gabor transform. This method was proposed in \[[11](#bib.bib11)\] and further studied in \[[17](#bib.bib17), [8](#bib.bib8)\]. Gabor Transforms are well suited since they are insensitive to small changes in an image and they reveal the locations as well as the orientations of structures at different spatial frequencies. They are used in a wide range of applications, such as iris recognition \[[3](#bib.bib3)\], texture analysis and image enhancement, coding and compression.
Here we briefly review the method used in the literature. A laser beam illuminates an object and, either by transmission or reflection, produces a speckle pattern. An image of the pattern is recorded in the ‘detection plane’. A point in the detection plane is denoted as two-dimensional vector $`\overset{\rightarrow}{x}`$. The light intensity in the detection plane is denoted as $`I\hspace{0pt}{(\overset{\rightarrow}{x})}`$.
A two-dimensional Gabor basis function $`\Gamma\hspace{0pt}{(w,\overset{\rightarrow}{k},{\overset{\rightarrow}{x}}_{0},\overset{\rightarrow}{x})}`$ is the product of a plane wave with wave vector $`\overset{\rightarrow}{k}`$ and a Gaussian with width $`w`$ centered on $`{\overset{\rightarrow}{x}}_{0}`$. We write the Gabor basis functions $`\Gamma_{IM}`$ and the Gabor coefficients $`G`$ as follows:
<table>
<tbody>
<tr>
<td></td>
<td>G(w,k→,x→0)𝐺𝑤→𝑘subscript→𝑥0\displaystyle G(w,\vec{k},\vec{x}_{0})</td>
<td>=\displaystyle=</td>
<td>∫d2xΓIM(w,k→,x→0,x→)I(x→)superscriptd2𝑥subscriptΓIM𝑤→𝑘subscript→𝑥0→𝑥𝐼→𝑥\displaystyle\int\!{{\rm d}}^{2}x\;\Gamma_{\rm IM}(w,\vec{k},\vec{x}_{0},\vec{x})I(\vec{x})</td>
<td></td>
<td>(1)</td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td>ΓIM(w,k→,x→0,x→)subscriptΓIM𝑤→𝑘subscript→𝑥0→𝑥\displaystyle\Gamma_{\rm IM}(w,\vec{k},\vec{x}_{0},\vec{x})</td>
<td>=\displaystyle=</td>
<td>12πw2sink→⋅(x→−x→0)exp[−(x→−x→0)22w2].⋅12𝜋superscript𝑤2→𝑘→𝑥subscript→𝑥0superscript→𝑥subscript→𝑥022superscript𝑤2\displaystyle\frac{1}{2\pi w^{2}}\sin\vec{k}\cdot(\vec{x}-\vec{x}_{0})\exp[-\frac{(\vec{x}-\vec{x}_{0})^{2}}{2w^{2}}].</td>
<td></td>
<td>(2)</td>
</tr>
</tbody>
</table>
Only the imaginary (sine) part of the transform is considered. This is motivated by the property that the imaginary part is invariant under spatially constant perturbations of the intensity.
Gabor coefficients $`G`$ are evaluated for a subset of parameters $`w`$, $`\overset{\rightarrow}{k}`$, $`{\overset{\rightarrow}{x}}_{0}`$, e.g. on a sub-lattice of positions $`{\overset{\rightarrow}{x}}_{0}`$, for two perpendicular choices of $`\overset{\rightarrow}{k}`$ (with equal modulus $`|\overset{\rightarrow}{k}|`$), and for one fixed $`w`$. Since the basis functions form an overcomplete set, such a restricted choice of parameters can capture almost all information available in a speckle image.
Coefficients are discarded if they do not exceed a certain threshold $`T`$, i.e. one only keeps $`{|G|} > T`$. The chosen coefficients are called ‘robust’, because they are unlikely to be affected by noise. Finally, the robust coefficients are binarized; positive values are mapped to ‘1’ and negative to ‘0’.
This procedure is applied to the speckle pattern photographed during enrollment, and again to the image obtained in the authentication measurement. The Hamming distance (number of bit flips) between the enrolled bitstring and the second bitstring depends on the threshold $`T`$ and the amount of measurement noise. Based on knowledge about the expected number of bit flips, one applies an error correction scheme that can cope with the noise. It is important to keep in mind that the robust bitstring can be deceptively long. The actual amount of information contained in it can be much less than the length, due to correlations between the Gabor coefficients \[[17](#bib.bib17)\].
### 1.3 Contributions and outline of this paper
In Section [2](#S2 "2 The random phase model ‣ The entropy of keys derived from laser speckle") we first briefly describe the random phase model and the intensity statistics that are obtained from it. We motivate our use of this model. In Sections [3](#S3 "3 Statistics of the Gabor coefficients ‣ The entropy of keys derived from laser speckle") and [4](#S4 "4 Perturbing speckle patterns ‣ The entropy of keys derived from laser speckle") we study the statistics of Gabor coefficients and the effects of random perturbations. This paper contains the following novel contributions:
- •
In Section [3.2](#S3.SS2 "3.2 Computation of higher moments ‣ 3 Statistics of the Gabor coefficients ‣ The entropy of keys derived from laser speckle") we analyse the statistical properties of the Gabor-transform ([1](#S1.E1 "In 1.2 Binarized Gabor coefficients ‣ 1 Introduction ‣ The entropy of keys derived from laser speckle")) of a speckle pattern. We present a procedure for computing arbitrary moments of the distribution. Computation of the first four moments shows that there are small deviations from the normal form.
- •
In Section [3.3](#S3.SS3 "3.3 Estimated entropy of a set of Gabor coefficients for given detector noise ‣ 3 Statistics of the Gabor coefficients ‣ The entropy of keys derived from laser speckle") we compute the information content of a set of Gabor coefficients, for a given noise level of the detector. The entropy per typical speckle area turns out to be proportional to the square of the logarithm of the signal to noise ratio.
- •
In Section [4.1](#S4.SS1 "4.1 Phase perturbations ‣ 4 Perturbing speckle patterns ‣ The entropy of keys derived from laser speckle") we introduce a method of perturbing a speckle pattern in the random phase model. Each $`\lambda^{2}`$ sized source region has its phase shifted by a small random amount $`\varepsilon`$, where $`\varepsilon`$ is drawn from a uniform distribution of width $`2\hspace{0pt}q`$. By tuning $`q \in {\lbrack{\frac{1}{2}\hspace{0pt}\bigtriangleup\hspace{0pt}\varphi},\pi\rbrack}`$ (where $`\bigtriangleup\hspace{0pt}\varphi`$ denotes the phase uncertainty due to the number-phase uncertainty relation), the magnitude of the perturbation is selected. We use this kind of perturbation to represent a misalignment, such as a shift or rotation of the token or the laser, or a change in the structure of a token.
- •
In Section [4.2](#S4.SS2 "4.2 Mutual information in the source plane ‣ 4 Perturbing speckle patterns ‣ The entropy of keys derived from laser speckle") we compute the mutual information between the original speckle ‘source’ and the perturbed one as a function of the noise strength $`q`$. The result is proportional to $`\log{({\pi/q})}`$.
- •
In Section [4.3](#S4.SS3 "4.3 Effect of perturbations on the intensity ‣ 4 Perturbing speckle patterns ‣ The entropy of keys derived from laser speckle") we show that the correlation between the intensity before and after a perturbation is given by $`\frac{\sin^{2}q}{q^{2}}`$. In Section [4.4](#S4.SS4 "4.4 Effect of perturbations on the Gabor coefficients ‣ 4 Perturbing speckle patterns ‣ The entropy of keys derived from laser speckle") the same correlation is obtained between Gabor coefficients before and after a perturbation.
- •
In Section [4.4.1](#S4.SS4.SSS1 "4.4.1 Mutual information between perturbed and unperturbed Gabor coefficients ‣ 4.4 Effect of perturbations on the Gabor coefficients ‣ 4 Perturbing speckle patterns ‣ The entropy of keys derived from laser speckle") we compute the mutual information between a set of Gabor coefficients before and after a perturbation.
- •
In Section [4.4.2](#S4.SS4.SSS2 "4.4.2 Bit error probability ‣ 4.4 Effect of perturbations on the Gabor coefficients ‣ 4 Perturbing speckle patterns ‣ The entropy of keys derived from laser speckle") we compute the bit flip probability for a binarized Gabor coefficient due to a random perturbation.
- •
In Section [5](#S5 "5 Comparison to experimental results ‣ The entropy of keys derived from laser speckle") we give experimental results. (i) It turns out that the empirical distribution function of the Gabor coefficients is consistent with theory. There is a noticeable deviation from the Gaussian form. The theoretical prediction of the variance matches very well with the data. (ii) We studied random perturbations by doing measurements on a sample whose surface structure slowly changes in time. Correlations were determined between the state before and after a perturbation. As expected from the theoretical results, there is a linear relation between the correlation function of the intensity and the correlation function of the Gabor coefficients.
## 2 The random phase model
### 2.1 Motivation and definitions
Throughout this paper we use the random phase model as described by Goodman \[[7](#bib.bib7)\], in the Fresnel approximation, in a free space geometry, for completely polarised light. This model has the advantage of being relatively simple while yielding intensity statistics that agree with experimental observations. We depart from the traditional approach only in one respect. In \[[7](#bib.bib7)\] the components of the electric field amplitude ($`A_{x}`$, $`A_{y}`$) in the detection plane are sometimes treated as the ‘fundamental’ degrees of freedom. For instance, second order and higher order intensity correlations are derived using the Gaussian distribution of $`\overset{\rightarrow}{A}\hspace{0pt}{(\overset{\rightarrow}{x})}`$. The traditional approach has the drawback that it is very difficult to keep track of the number of degrees of freedom: It looks as if there is one degree of freedom per (continuum!) location $`\overset{\rightarrow}{x}`$ in the source plane. In fact the physical degrees of freedom lie in the source plane (defined as the exit plane of a transmissive PUF, or as the surface that reflects the laser light) and they are very easy to identify and to count. Another drawback is that there is no natural way to introduce misalignment perturbations in terms of the $`\overset{\rightarrow}{A}\hspace{0pt}{(\overset{\rightarrow}{x})}`$ variables. On a more esthetic level, there is the drawback of having to rely on the Central Limit Theorem to get the Gaussian distribution of $`\overset{\rightarrow}{A}`$, while in fact the number of random amplitudes added together is not infinite but merely very large.
For these reasons we base our calculations on the random phases in the source plane as the fundamental degrees of freedom. All the well known speckle properties are of course reproduced in this approach. The model looks as follows. Diffused light leaves the PUF at the exit plane, through a disc-shaped region with radius $`R`$ which we call the ‘source’. We assume that the intensity is the same everywhere in the source. (We normalize the intensity to 1). Hence the source is modelled as a collection of random phases $`\varphi`$. The disc is divided into small regions of area $`\lambda^{2}`$. (The total number of regions is denoted as $`N_{reg} = {{\pi\hspace{0pt}R^{2}}/\lambda^{2}}`$). Together these generate the speckle pattern according to Huygens’ principle. The complex amplitudes $`\alpha`$ in each region are the basic degrees of freedom. We write
<table>
<tbody>
<tr>
<td></td>
<td>αa→=expiφa→,subscript𝛼→𝑎𝑖subscript𝜑→𝑎\alpha_{\vec{a}}=\exp i\varphi_{\vec{a}},</td>
<td></td>
<td>(3)</td>
</tr>
</tbody>
</table>
where the subscript $`\overset{\rightarrow}{a}`$ denotes a discrete two-dimensional coordinate in the source, with $`{|\overset{\rightarrow}{a}|} < R`$. The phases $`\varphi_{\overset{\rightarrow}{a}}`$ at all the locations $`\overset{\rightarrow}{a}`$ are independent stochastic variables, with a uniform distribution in the interval $`({- \pi},\pi\rbrack`$. We introduce the notation $`\left\langle \cdot \right\rangle_{\varphi}`$ for taking the expectation value with respect to the random phases. We have
<table>
<tbody>
<tr>
<td></td>
<td>⟨(αa→)n⟩φ=0forn∈ℕ+formulae-sequencesubscriptdelimited-⟨⟩superscriptsubscript𝛼→𝑎𝑛𝜑0for𝑛superscriptℕ\displaystyle\left\langle(\alpha_{\vec{a}})^{n}\right\rangle_{\varphi}=0\quad{\rm for}\;n\in{\mathbb{N}}^{+}</td>
<td>;</td>
<td>⟨αa→αb→∗⟩φ=δa→,b→.subscriptdelimited-⟨⟩subscript𝛼→𝑎superscriptsubscript𝛼→𝑏𝜑subscript𝛿→𝑎→𝑏\displaystyle\left\langle\alpha_{\vec{a}}\alpha_{\vec{b}}^{*}\right\rangle_{\varphi}=\delta_{\vec{a},\vec{b}}.</td>
<td></td>
<td>(4)</td>
</tr>
</tbody>
</table>
From these basic rules it is straightforward to derive many-point correlations.
Note that in adopting independent random phases we ignore the correlations that are known to exist between the phases as a consequence of either (a) multiple coherent scattering in a diffusive medium (see e.g. \[[5](#bib.bib5)\]), or (b) height and/or orientation correlations between microscopic pieces of a rough surface. These correlations lead to a reduction of the number of degrees of freedom. However, in this paper we are primarily interested in the influence that the parameter choices in ([2](#S1.E2 "In 1.2 Binarized Gabor coefficients ‣ 1 Introduction ‣ The entropy of keys derived from laser speckle")) have on the information that can be extracted from a Gabor-filtered speckle pattern. In this context the correlations between phases in the source plane are only of minor importance. Hence we will ignore them and work with ([4](#S2.E4 "In 2.1 Motivation and definitions ‣ 2 The random phase model ‣ The entropy of keys derived from laser speckle")).
The distance between the source and the detection plane is denoted as $`z`$. We assume $`z \gg \lambda`$ and use the Fresnel approximation. For the complex amplitude $`A = {A_{x} + {i\hspace{0pt}A_{y}}}`$ in a point $`\overset{\rightarrow}{x}`$ in the detection plane we then have
<table>
<tbody>
<tr>
<td></td>
<td>A(x→)=λz∑a→αa→exp−iπλz(x→−a→)2.𝐴→𝑥𝜆𝑧subscript→𝑎subscript𝛼→𝑎𝑖𝜋𝜆𝑧superscript→𝑥→𝑎2A(\vec{x})=\frac{\lambda}{z}\sum_{\vec{a}}\alpha_{\vec{a}}\exp-i\frac{\pi}{\lambda z}(\vec{x}-\vec{a})^{2}.</td>
<td></td>
<td>(5)</td>
</tr>
</tbody>
</table>
### 2.2 Entropy of the source
The entropy of the source is an upper bound on the entropy of the speckle pattern in a half- sphere. The entropy of the source is easily computed in the random phase model. The phase distribution is completely uniform. Therefore the entropy reduces to the logarithm of the number of possible states. It is well known that a coherent state with photon number $`N_{0}`$ has an uncertainty $`{\bigtriangleup\hspace{0pt}N} = \sqrt{N_{0}}`$ in the photon number. Using the number-phase uncertainty relation $`{\bigtriangleup\hspace{0pt}N\hspace{0pt}\bigtriangleup\hspace{0pt}\varphi} = \frac{1}{2}`$, we find a phase discretisation $`{\bigtriangleup\hspace{0pt}\varphi} = {1/{({2\hspace{0pt}\sqrt{N_{0}}})}}`$. Hence the entropy (expressed in bits) is given by
<table>
<tbody>
<tr>
<td></td>
<td>H[α]=log2(2π△φ)Nreg=Nreglog2(4πN0).H[\alpha]=\log_{2}(\frac{2\pi}{\triangle\varphi})^{N_{\rm reg}}=N_{\rm reg}\log_{2}(4\pi\sqrt{N_{0}}).</td>
<td></td>
<td>(6)</td>
</tr>
</tbody>
</table>
Note that this result is equivalent to the estimate in \[[17](#bib.bib17)\], where the light exiting the PUF was described in terms of transversal momentum modes. In \[[17](#bib.bib17)\] the correlations between modes were studied as well, and the resulting entropy reduction was estimated. In this paper we will not take such correlations into account.
In order to get some feeling for the orders of magnitude we substitute numbers into ([6](#S2.E6 "In 2.2 Entropy of the source ‣ 2 The random phase model ‣ The entropy of keys derived from laser speckle")). A laser with $`\lambda = 780`$nm produces an output power $`P =`$1mW, and a measurement with a CCD camera takes about $`{\bigtriangleup\hspace{0pt}t} =`$1ms. The total number of photons involved in one measurement is $`{P\hspace{0pt}\bigtriangleup\hspace{0pt}t}/{({{h\hspace{0pt}c}/\lambda})}`$, where $`h`$ is Planck’s constant and $`c`$ is the velocity of light. Thus we arrive at $`N_{0} = {{\lambda\hspace{0pt}P\hspace{0pt}\bigtriangleup\hspace{0pt}t}/{({h\hspace{0pt}c\hspace{0pt}N_{reg}})}}`$. Assuming a source diameter of 1mm, we get $`N_{reg} = {1.3 \cdot 10^{6}}`$ and $`N_{0} = {3 \cdot 10^{6}}`$ photons per region, yielding an entropy of approximately 14 bits per region of size $`\lambda^{2}`$.
### 2.3 Statistics of the intensity
All the well known statistical properties of the intensity can be derived from the random phase model. For completeness and for use in later sections, we briefly discuss how these properties are derived. The intensity at position $`\overset{\rightarrow}{x}`$ in the plane of detection is given by the squared modulus of the amplitude ([5](#S2.E5 "In 2.1 Motivation and definitions ‣ 2 The random phase model ‣ The entropy of keys derived from laser speckle")),
<table>
<tbody>
<tr>
<td></td>
<td>I(x→)=|A(x→)|2=λ2z2∑a→,b→αa→αb→∗expiπλz[b→2−a→2+2x→⋅(a→−b→)].𝐼→𝑥superscript𝐴→𝑥2superscript𝜆2superscript𝑧2subscript→𝑎→𝑏subscript𝛼→𝑎subscriptsuperscript𝛼→𝑏𝑖𝜋𝜆𝑧delimited-[]superscript→𝑏2superscript→𝑎2⋅2→𝑥→𝑎→𝑏I(\vec{x})=\left|A(\vec{x})\right|^{2}=\frac{\lambda^{2}}{z^{2}}\sum_{\vec{a},\vec{b}}\alpha_{\vec{a}}\alpha^{*}_{\vec{b}}\;\exp\frac{i\pi}{\lambda z}\left[\vec{b}^{2}-\vec{a}^{2}+2\vec{x}\cdot(\vec{a}-\vec{b})\right].</td>
<td></td>
<td>(7)</td>
</tr>
</tbody>
</table>
The average intensity $`I_{av}`$ is obtained by taking the expectation value $`\left\langle \cdot \right\rangle_{\varphi}`$ and directly applying ([4](#S2.E4 "In 2.1 Motivation and definitions ‣ 2 The random phase model ‣ The entropy of keys derived from laser speckle")). Summation $`\sum_{\overset{\rightarrow}{a}}`$ over a constant yields a factor $`N_{reg}`$. We obtain
<table>
<tbody>
<tr>
<td></td>
<td>Iav=⟨I(x→)⟩φ=Nregλ2z2=πR2z2.subscript𝐼avsubscriptdelimited-⟨⟩𝐼→𝑥𝜑subscript𝑁regsuperscript𝜆2superscript𝑧2𝜋superscript𝑅2superscript𝑧2I_{\rm av}=\left\langle I(\vec{x})\right\rangle_{\varphi}=\frac{N_{\rm reg}\lambda^{2}}{z^{2}}=\frac{\pi R^{2}}{z^{2}}.</td>
<td></td>
<td>(8)</td>
</tr>
</tbody>
</table>
For second order statistics of the intensity one needs 4th order correlations of the random amplitudes $`\alpha`$. In particular, from ([4](#S2.E4 "In 2.1 Motivation and definitions ‣ 2 The random phase model ‣ The entropy of keys derived from laser speckle")) it follows that
<table>
<tbody>
<tr>
<td></td>
<td>⟨αa→1αa→2αb→1∗αb→2∗⟩φ=δa→1,b→1δa→2,b→2+δa→1,b→2δa→2,b→1.subscriptdelimited-⟨⟩subscript𝛼subscript→𝑎1subscript𝛼subscript→𝑎2subscriptsuperscript𝛼subscript→𝑏1subscriptsuperscript𝛼subscript→𝑏2𝜑subscript𝛿subscript→𝑎1subscript→𝑏1subscript𝛿subscript→𝑎2subscript→𝑏2subscript𝛿subscript→𝑎1subscript→𝑏2subscript𝛿subscript→𝑎2subscript→𝑏1\left\langle\alpha_{\vec{a}_{1}}\alpha_{\vec{a}_{2}}\alpha^{*}_{\vec{b}_{1}}\alpha^{*}_{\vec{b}_{2}}\right\rangle_{\varphi}=\delta_{\vec{a}_{1},\vec{b}_{1}}\delta_{\vec{a}_{2},\vec{b}_{2}}+\delta_{\vec{a}_{1},\vec{b}_{2}}\delta_{\vec{a}_{2},\vec{b}_{1}}.</td>
<td></td>
<td>(9)</td>
</tr>
</tbody>
</table>
Using ([9](#S2.E9 "In 2.3 Statistics of the intensity ‣ 2 The random phase model ‣ The entropy of keys derived from laser speckle")), the well known results follow for the variance ($`\sigma_{I} = I_{av}`$) and for the intensity correlation function $`C_{I}`$,
<table>
<tbody>
<tr>
<td></td>
<td>CI(x→,x→′):=⟨I(x→)I(x→′)⟩φ−Iav2σI2=4[J1(|x→′−x→|/M)|x→′−x→|/M]2,assignsubscript𝐶𝐼→𝑥superscript→𝑥′subscriptdelimited-⟨⟩𝐼→𝑥𝐼superscript→𝑥′𝜑superscriptsubscript𝐼av2superscriptsubscript𝜎𝐼24superscriptdelimited-[]subscript𝐽1superscript→𝑥′→𝑥𝑀superscript→𝑥′→𝑥𝑀2C_{I}(\vec{x},\vec{x}^{\prime}):=\frac{\left\langle I(\vec{x})I(\vec{x}^{\prime})\right\rangle_{\varphi}-I_{\rm av}^{2}}{\sigma_{I}^{2}}=4\left[\frac{J_{1}(|\vec{x}^{\prime}-\vec{x}|/M)}{|\vec{x}^{\prime}-\vec{x}|/M}\right]^{2},</td>
<td></td>
<td>(10)</td>
</tr>
</tbody>
</table>
where $`M`$ is a constant proportional to the average speckle size,
<table>
<tbody>
<tr>
<td></td>
<td>M=λz2πR𝑀𝜆𝑧2𝜋𝑅M=\frac{\lambda z}{2\pi R}</td>
<td></td>
<td>(11)</td>
</tr>
</tbody>
</table>
and $`J_{1}`$ is a Bessel function. Higher order expectation values can also be computed. In particular, in order to derive the well known exponential probability density
<table>
<tbody>
<tr>
<td></td>
<td>p(I)=Iav−1exp(−I/Iav),𝑝𝐼superscriptsubscript𝐼av1𝐼subscript𝐼avp(I)=I_{\rm av}^{-1}\exp(-I/I_{\rm av}),</td>
<td></td>
<td>(12)</td>
</tr>
</tbody>
</table>
it has to be shown that $`\left\langle {\lbrack{I\hspace{0pt}{(\overset{\rightarrow}{x})}}\rbrack}^{n} \right\rangle_{\varphi} = {I_{av}^{n}\hspace{0pt}{n!}}`$. This is done using the following correlation function, which also follows from ([4](#S2.E4 "In 2.1 Motivation and definitions ‣ 2 The random phase model ‣ The entropy of keys derived from laser speckle")),
<table>
<tbody>
<tr>
<td></td>
<td>⟨αa→1…αa→n⋅αb→1∗…αb→n∗⟩φ=∑ν∈Sn∏k=1nδa→k,b→ν(k).subscriptdelimited-⟨⟩⋅subscript𝛼subscript→𝑎1…subscript𝛼subscript→𝑎𝑛subscriptsuperscript𝛼subscript→𝑏1…subscriptsuperscript𝛼subscript→𝑏𝑛𝜑subscript𝜈subscript𝑆𝑛superscriptsubscriptproduct𝑘1𝑛subscript𝛿subscript→𝑎𝑘subscript→𝑏𝜈𝑘\left\langle\alpha_{\vec{a}_{1}}\ldots\alpha_{\vec{a}_{n}}\cdot\alpha^{*}_{\vec{b}_{1}}\ldots\alpha^{*}_{\vec{b}_{n}}\right\rangle_{\varphi}=\sum_{\nu\in S_{n}}\prod_{k=1}^{n}\delta_{\vec{a}_{k},\vec{b}_{\nu(k)}}.</td>
<td></td>
<td>(13)</td>
</tr>
</tbody>
</table>
Here $`S_{n}`$ stands for the ‘symmetric group’ of all the $`n!`$ possible permutations of the numbers $`1\hspace{0pt}\ldots\hspace{0pt}n`$. Eq. ([13](#S2.E13 "In 2.3 Statistics of the intensity ‣ 2 The random phase model ‣ The entropy of keys derived from laser speckle")) is also used in the derivation of the joint probability distribution $`p\hspace{0pt}{({I\hspace{0pt}{(\overset{\rightarrow}{x})}},{I\hspace{0pt}{({\overset{\rightarrow}{x}}')}})}`$. This distribution follows from the expectation value
<table>
<tbody>
<tr>
<td></td>
<td>⟨[I(x→)]n[I(x→′)]m⟩φ=Iavn+mn!m!F12(−n,−m;1;CI(x→,x→′))subscriptdelimited-⟨⟩superscriptdelimited-[]𝐼→𝑥𝑛superscriptdelimited-[]𝐼superscript→𝑥′𝑚𝜑superscriptsubscript𝐼av𝑛𝑚𝑛𝑚subscriptsubscript𝐹12𝑛𝑚1subscript𝐶𝐼→𝑥superscript→𝑥′\left\langle[I(\vec{x})]^{n}[I(\vec{x}^{\prime})]^{m}\right\rangle_{\varphi}=I_{\rm av}^{n+m}n!m!\;\;{}_{2}F_{1}(-n,-m;1;C_{I}(\vec{x},\vec{x}^{\prime}))</td>
<td></td>
<td>(14)</td>
</tr>
</tbody>
</table>
and is given by
<table>
<tbody>
<tr>
<td></td>
<td>p(I(x→),I(x→′))=1Iav2(1−CI)exp[−I(x→)+I(x→′)Iav(1−CI)]I0(2I(x→)I(x→′)IavCI1−CI).𝑝𝐼→𝑥𝐼superscript→𝑥′1superscriptsubscript𝐼av21subscript𝐶𝐼𝐼→𝑥𝐼superscript→𝑥′subscript𝐼av1subscript𝐶𝐼subscript𝐼02𝐼→𝑥𝐼superscript→𝑥′subscript𝐼avsubscript𝐶𝐼1subscript𝐶𝐼p(I(\vec{x}),I(\vec{x}^{\prime}))=\frac{1}{I_{\rm av}^{2}(1-C_{I})}\exp\left[-\frac{I(\vec{x})+I(\vec{x}^{\prime})}{I_{\rm av}(1-C_{I})}\right]I_{0}\left(\frac{2\sqrt{I(\vec{x})I(\vec{x}^{\prime})}}{I_{\rm av}}\frac{\sqrt{C_{I}}}{1-C_{I}}\right).</td>
<td></td>
<td>(15)</td>
</tr>
</tbody>
</table>
Here $`C_{I}`$ is shorthand notation for $`C_{I}\hspace{0pt}{(\overset{\rightarrow}{x},{\overset{\rightarrow}{x}}')}`$ and $`I_{0}`$ is a Bessel function.
## 3 Statistics of the Gabor coefficients
### 3.1 Second order statistics
Second order statistics of the Gabor coefficients of a speckle pattern were calculated in \[[17](#bib.bib17)\]. In this section we study the higher order statistics. In particular we show that all the odd moments are zero and that the fourth moment is dominated by the Gaussian contribution; i.e. the probability distribution of a Gabor coefficient is ‘almost’ Gaussian.
We will often use shorthand notation $`G`$ for $`G\hspace{0pt}{(w,\overset{\rightarrow}{k},\overset{\rightarrow}{x})}`$. From the fact that $`\Gamma_{IM}`$ ([2](#S1.E2 "In 1.2 Binarized Gabor coefficients ‣ 1 Introduction ‣ The entropy of keys derived from laser speckle")) is an odd function in $`({\overset{\rightarrow}{x} - {\overset{\rightarrow}{x}}_{0}})`$ it is easily seen that $`\left\langle G \right\rangle_{\varphi} = 0`$. In \[[17](#bib.bib17)\] it was shown that the variance of $`G`$ is given by
<table>
<tbody>
<tr>
<td></td>
<td>σG2(w,k→,x→0)≈Iav2(1−2γ)e−(1−γ)w2k2sinhγw2k2,subscriptsuperscript𝜎2𝐺𝑤→𝑘subscript→𝑥0superscriptsubscript𝐼av212𝛾superscript𝑒1𝛾superscript𝑤2superscript𝑘2𝛾superscript𝑤2superscript𝑘2\sigma^{2}_{G}(w,\vec{k},\vec{x}_{0})\approx I_{\rm av}^{2}(1-2\gamma)e^{-(1-\gamma)w^{2}k^{2}}\sinh\gamma w^{2}k^{2},</td>
<td></td>
<td>(16)</td>
</tr>
</tbody>
</table>
where $`\gamma = {\frac{1}{2}\hspace{0pt}{\lbrack{1 + {{M^{2}\hspace{0pt}\Sigma^{2}}/{({2\hspace{0pt}w^{2}})}}}\rbrack}^{- 1}}`$ and $`\Sigma \approx 1.29`$ is a numerical constant. The constant $`\Sigma`$ originates from an approximation of the correlation ([10](#S2.E10 "In 2.3 Statistics of the intensity ‣ 2 The random phase model ‣ The entropy of keys derived from laser speckle")) by a Gaussian curve. The correlation between two different Gabor coefficients was also computed. The correlation is defined as
<table>
<tbody>
<tr>
<td></td>
<td>CG(w,w′,k→,k→′,x→,x→′):=⟨G(w,k→,x→)G(w′,k→′,x→′)⟩φσG(w,k→,x→)σG(w′,k→′,x→′),assignsubscript𝐶𝐺𝑤superscript𝑤′→𝑘superscript→𝑘′→𝑥superscript→𝑥′subscriptdelimited-⟨⟩𝐺𝑤→𝑘→𝑥𝐺superscript𝑤′superscript→𝑘′superscript→𝑥′𝜑subscript𝜎𝐺𝑤→𝑘→𝑥subscript𝜎𝐺superscript𝑤′superscript→𝑘′superscript→𝑥′C_{G}(w,w^{\prime},\vec{k},\vec{k}^{\prime},\vec{x},\vec{x}^{\prime}):=\frac{\left\langle G(w,\vec{k},\vec{x})G(w^{\prime},\vec{k}^{\prime},\vec{x}^{\prime})\right\rangle_{\varphi}}{\sigma_{G}(w,\vec{k},\vec{x})\sigma_{G}(w^{\prime},\vec{k}^{\prime},\vec{x}^{\prime})},</td>
<td></td>
<td>(17)</td>
</tr>
</tbody>
</table>
and the following result was obtained for $`w' = w`$,
<table>
<tbody>
<tr>
<td></td>
<td>CGsubscript𝐶𝐺\displaystyle C_{G}</td>
<td>≈\displaystyle\approx</td>
<td>exp[−γ2⋅(x→′−x→)2w2]×\displaystyle\exp\left[-\frac{\gamma}{2}\cdot\frac{(\vec{x}^{\prime}-\vec{x})^{2}}{w^{2}}\right]\times</td>
<td></td>
<td rowspan="2">(18)</td>
</tr>
<tr>
<td></td>
<td></td>
<td></td>
<td>eγw2k→⋅k→′cos[γ(x→′−x→)⋅(k→′+k→)]−e−γw2k→⋅k→′cos[γ(x→′−x→)⋅(k→′−k→)]2sinhγw2k2sinhγw2k′2.superscript𝑒⋅𝛾superscript𝑤2→𝑘superscript→𝑘′⋅𝛾superscript→𝑥′→𝑥superscript→𝑘′→𝑘superscript𝑒⋅𝛾superscript𝑤2→𝑘superscript→𝑘′⋅𝛾superscript→𝑥′→𝑥superscript→𝑘′→𝑘2𝛾superscript𝑤2superscript𝑘2𝛾superscript𝑤2superscript𝑘′2\displaystyle\frac{e^{\gamma w^{2}\vec{k}\cdot\vec{k}^{\prime}}\cos[\gamma(\vec{x}^{\prime}-\vec{x})\cdot(\vec{k}^{\prime}+\vec{k})]-e^{-\gamma w^{2}\vec{k}\cdot\vec{k}^{\prime}}\cos[\gamma(\vec{x}^{\prime}-\vec{x})\cdot(\vec{k}^{\prime}-\vec{k})]}{2\sqrt{\sinh\gamma w^{2}k^{2}}\sqrt{\sinh\gamma w^{2}k^{\prime 2}}}.</td>
<td></td>
</tr>
</tbody>
</table>
The result ([18](#S3.E18 "In 3.1 Second order statistics ‣ 3 Statistics of the Gabor coefficients ‣ The entropy of keys derived from laser speckle")) is accurate for small distances $`|{{\overset{\rightarrow}{x}}' - \overset{\rightarrow}{x}}|`$. For larger distances, the Gaussian tail underestimates the actual correlation.
### 3.2 Computation of higher moments
We present a procedure that allows for the computation of arbitrary moments of $`G`$. Substitution of ([7](#S2.E7 "In 2.3 Statistics of the intensity ‣ 2 The random phase model ‣ The entropy of keys derived from laser speckle")) into ([2](#S1.E2 "In 1.2 Binarized Gabor coefficients ‣ 1 Introduction ‣ The entropy of keys derived from laser speckle")) gives an expression for $`G`$ in terms of the random phases,
<table>
<tbody>
<tr>
<td></td>
<td>G(w,k→,x→0)𝐺𝑤→𝑘subscript→𝑥0\displaystyle G(w,\vec{k},\vec{x}_{0})</td>
<td>=\displaystyle=</td>
<td>iλ2z2e−12w2k2∑a→,b→αa→αb→∗expiπλz[b→2−a→2+2x→0⋅(a→−b→)]𝑖superscript𝜆2superscript𝑧2superscript𝑒12superscript𝑤2superscript𝑘2subscript→𝑎→𝑏subscript𝛼→𝑎subscriptsuperscript𝛼→𝑏𝑖𝜋𝜆𝑧delimited-[]superscript→𝑏2superscript→𝑎2⋅2subscript→𝑥0→𝑎→𝑏\displaystyle\frac{i\lambda^{2}}{z^{2}}e^{-\mbox{$\textstyle\frac{1}{2}$}w^{2}k^{2}}\sum_{\vec{a},\vec{b}}\alpha_{\vec{a}}\alpha^{*}_{\vec{b}}\exp\frac{i\pi}{\lambda z}[\vec{b}^{2}-\vec{a}^{2}+2\vec{x}_{0}\cdot(\vec{a}-\vec{b})]</td>
<td></td>
<td rowspan="2">(19)</td>
</tr>
<tr>
<td></td>
<td></td>
<td></td>
<td>exp[−(2π)2(λz)2(a→−b→)2]sinh[2πλzw2k→⋅(a→−b→)].superscript2𝜋2superscript𝜆𝑧2superscript→𝑎→𝑏2⋅2𝜋𝜆𝑧superscript𝑤2→𝑘→𝑎→𝑏\displaystyle\exp\left[-\frac{(2\pi)^{2}}{(\lambda z)^{2}}(\vec{a}-\vec{b})^{2}\right]\;\sinh\left[\frac{2\pi}{\lambda z}w^{2}\vec{k}\cdot(\vec{a}-\vec{b})\right].</td>
<td></td>
</tr>
</tbody>
</table>
Taking the $`n`$th power of ([19](#S3.E19 "In 3.2 Computation of higher moments ‣ 3 Statistics of the Gabor coefficients ‣ The entropy of keys derived from laser speckle")) and averaging over the random phases leads to a correlation function of the form ([13](#S2.E13 "In 2.3 Statistics of the intensity ‣ 2 The random phase model ‣ The entropy of keys derived from laser speckle")). Each $`{\overset{\rightarrow}{b}}_{j}`$ evaluates to $`{\overset{\rightarrow}{a}}_{\nu\hspace{0pt}{(j)}}`$. For each permutation, the sums $`\sum_{j}{({{\overset{\rightarrow}{a}}_{j} - {\overset{\rightarrow}{b}}_{j}})}`$ and $`\sum_{j}{({{\overset{\rightarrow}{b}}_{j}^{2} - {\overset{\rightarrow}{a}}_{j}^{2}})}`$ reduce to zero. In this way we obtain the following expression
<table>
<tbody>
<tr>
<td></td>
<td>⟨(G)n⟩φsubscriptdelimited-⟨⟩superscript𝐺𝑛𝜑\displaystyle\left\langle(G)^{n}\right\rangle_{\varphi}</td>
<td>=\displaystyle=</td>
<td>[iλ2z2]ne−nw2k2/2∑a→1⋯a→nexp[−w2M2R2∑j=1na→j2]superscriptdelimited-[]𝑖superscript𝜆2superscript𝑧2𝑛superscript𝑒𝑛superscript𝑤2superscript𝑘22subscriptsubscript→𝑎1⋯subscript→𝑎𝑛superscript𝑤2superscript𝑀2superscript𝑅2superscriptsubscript𝑗1𝑛superscriptsubscript→𝑎𝑗2\displaystyle\left[\frac{i\lambda^{2}}{z^{2}}\right]^{n}e^{-nw^{2}k^{2}/2}\sum_{\vec{a}_{1}\cdots\vec{a}_{n}}\exp\left[-\frac{w^{2}}{M^{2}R^{2}}\sum_{j=1}^{n}\vec{a}_{j}^{2}\right]</td>
<td></td>
</tr>
<tr>
<td></td>
<td></td>
<td></td>
<td>∑ν∈Snexp[w2M2R2∑j=1na→j⋅a→ν(j)]∏t=1nsinhw2MRk→⋅(a→t−a→ν(t)).subscript𝜈subscript𝑆𝑛superscript𝑤2superscript𝑀2superscript𝑅2superscriptsubscript𝑗1𝑛⋅subscript→𝑎𝑗subscript→𝑎𝜈𝑗superscriptsubscriptproduct𝑡1𝑛⋅superscript𝑤2𝑀𝑅→𝑘subscript→𝑎𝑡subscript→𝑎𝜈𝑡\displaystyle\sum_{\nu\in S_{n}}\exp\left[\frac{w^{2}}{M^{2}R^{2}}\sum_{j=1}^{n}\vec{a}_{j}\cdot\vec{a}_{\nu(j)}\right]\prod_{t=1}^{n}\sinh\frac{w^{2}}{MR}\vec{k}\cdot(\vec{a}_{t}-\vec{a}_{\nu(t)}).</td>
<td></td>
</tr>
</tbody>
</table>
It is immediately clear that odd moments vanish, since ([3.2](#S3.Ex3 "3.2 Computation of higher moments ‣ 3 Statistics of the Gabor coefficients ‣ The entropy of keys derived from laser speckle")) is real-valued only for even $`n`$.
We can simplify ([3.2](#S3.Ex3 "3.2 Computation of higher moments ‣ 3 Statistics of the Gabor coefficients ‣ The entropy of keys derived from laser speckle")) by noting that the product of sinh’s is zero for permutations that have a fixed point. In other words, all cycles of a permutation have to be longer than 1. Also, we note that the $`\overset{\rightarrow}{a}`$-summation factors into a product of independent sums in accordance with the cycle structure of the permutation. For instance, the permutation $`{(231)}\hspace{0pt}{(564)}`$ leads to the factorisation $`{({\sum_{{\overset{\rightarrow}{a}}_{1}\hspace{0pt}{\overset{\rightarrow}{a}}_{2}\hspace{0pt}{\overset{\rightarrow}{a}}_{3}}\cdots})}\hspace{0pt}{({\sum_{{\overset{\rightarrow}{a}}_{4}\hspace{0pt}{\overset{\rightarrow}{a}}_{5}\hspace{0pt}{\overset{\rightarrow}{a}}_{6}}\cdots})}`$, where the dots indicate an expression depending only on the three denoted variables. Furthermore, the outcome of each such factor depends only on the length of the cycle, and not on the identity of the summation variables.
The 4th moment is obtained as follows. Among the $`4!`$ permutations of $`\{ 1,2,3,4\}`$ there are 3 containing two cycles of length two and 6 containing one cycle of length four. The 3 permutations with two cycles give rise to a contribution $`3\hspace{0pt}{(\sigma_{G}^{2})}^{2}`$, which precisely corresponds to the 4th moment of a Gaussian distribution. The non-Gaussian contribution from the 6 remaining permutations (i.e. the 4th cumulant) is computed in appendices [A](#A1 "Appendix A Fourth moment of the Gabor coefficients; case 𝑤>𝑀 ‣ The entropy of keys derived from laser speckle") and [B](#A2 "Appendix B Fourth moment of the Gabor coefficients; case 𝑤≪𝑀 ‣ The entropy of keys derived from laser speckle"). In the regime $`w \ll M`$ it turns out that the ratio of the non-Gaussian part to the Gaussian is 1/64. In the regime $`w \gtrapprox {3\hspace{0pt}M}`$ this ratio is of order $`\mathcal{O}\hspace{0pt}{({M^{2}/w^{2}})}`$.
### 3.3 Estimated entropy of a set of Gabor coefficients for given detector noise
We use a Gaussian approximation for the distribution function of the Gabor coefficients in order to derive an upper bound on the entropy of a set of Gabor coefficients, at a given noise level of the detector. This is a useful exercise for two reasons. First, it is not a priori clear how much of the information present in the source ends up in the Gabor coefficients. Second, detector noise affects the Gabor coefficients in a nontrivial way.
As mentioned in Section [1.2](#S1.SS2 "1.2 Binarized Gabor coefficients ‣ 1 Introduction ‣ The entropy of keys derived from laser speckle"), a relatively small set of coefficients can capture almost all the information available in a speckle image. We will consider the example given in Section [1.2](#S1.SS2 "1.2 Binarized Gabor coefficients ‣ 1 Introduction ‣ The entropy of keys derived from laser speckle"), which is also the choice made in \[[11](#bib.bib11), [17](#bib.bib17), [8](#bib.bib8)\]. Taking more than one width $`w`$, or more than one wave number $`|\overset{\rightarrow}{k}|`$ does not make much sense, since all the features in a speckle pattern have more or less the same length scale, namely the average speckle size. It is also not very useful to take more than two angles of the wave vector: Eq. ([18](#S3.E18 "In 3.1 Second order statistics ‣ 3 Statistics of the Gabor coefficients ‣ The entropy of keys derived from laser speckle")) shows us that there is a strong correlation $`{\sinh{\lbrack{\gamma\hspace{0pt}w^{2}\hspace{0pt}k^{2}\hspace{0pt}{\cos\zeta}}\rbrack}}/{\sinh{\gamma\hspace{0pt}w^{2}\hspace{0pt}k^{2}}}`$ between Gabor coefficients at the same position $`\overset{\rightarrow}{x}`$, when their $`\overset{\rightarrow}{k}`$-vectors have a mutual angle $`\zeta`$. Hence, for small $`\zeta`$ there is a lot of redundancy.
We take a single width $`w`$, a single wave vector length $`k`$, two perpendicular directions $`\psi_{1}`$, $`\psi_{2}`$, and a lattice of positions $`\overset{\rightarrow}{x}`$. We introduce the following definitions:
<table>
<tbody>
<tr>
<td></td>
<td>ΣG[jj](x→,x→′)superscriptsubscriptΣ𝐺delimited-[]𝑗𝑗→𝑥superscript→𝑥′\displaystyle\Sigma_{G}^{[jj]}(\vec{x},\vec{x}^{\prime})</td>
<td>:=assign\displaystyle:=</td>
<td>⟨G(w,k→j,x→)G(w,k→j,x→′)⟩φsubscriptdelimited-⟨⟩𝐺𝑤subscript→𝑘𝑗→𝑥𝐺𝑤subscript→𝑘𝑗superscript→𝑥′𝜑\displaystyle\left\langle G(w,\vec{k}_{j},\vec{x})G(w,\vec{k}_{j},\vec{x}^{\prime})\right\rangle_{\varphi}</td>
<td></td>
<td></td>
</tr>
<tr>
<td></td>
<td></td>
<td>=\displaystyle=</td>
<td>Iav2(12−γ)exp[−γ2(x→′−x→)2w2]{e(2γ−1)w2k2cos2γk→j⋅(x→′−x→)−e−w2k2}superscriptsubscript𝐼av212𝛾𝛾2superscriptsuperscript→𝑥′→𝑥2superscript𝑤2⋅superscript𝑒2𝛾1superscript𝑤2superscript𝑘22𝛾subscript→𝑘𝑗superscript→𝑥′→𝑥superscript𝑒superscript𝑤2superscript𝑘2\displaystyle I_{\rm av}^{2}(\mbox{$\textstyle\frac{1}{2}$}-\gamma)\exp\left[-\frac{\gamma}{2}\frac{(\vec{x}^{\prime}-\vec{x})^{2}}{w^{2}}\right]\left\{e^{(2\gamma-1)w^{2}k^{2}}\cos 2\gamma\vec{k}_{j}\cdot(\vec{x}^{\prime}-\vec{x})-e^{-w^{2}k^{2}}\right\}</td>
<td></td>
<td></td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td>ΣG[12](x→,x→′)superscriptsubscriptΣ𝐺delimited-[]12→𝑥superscript→𝑥′\displaystyle\Sigma_{G}^{[12]}(\vec{x},\vec{x}^{\prime})</td>
<td>:=assign\displaystyle:=</td>
<td>⟨G(w,k→1,x→)G(w,k→2,x→′)⟩φsubscriptdelimited-⟨⟩𝐺𝑤subscript→𝑘1→𝑥𝐺𝑤subscript→𝑘2superscript→𝑥′𝜑\displaystyle\left\langle G(w,\vec{k}_{1},\vec{x})G(w,\vec{k}_{2},\vec{x}^{\prime})\right\rangle_{\varphi}</td>
<td></td>
<td rowspan="3">(21)</td>
</tr>
<tr>
<td></td>
<td></td>
<td>=\displaystyle=</td>
<td>−Iav2(1−2γ)e(γ−1)w2k2exp[−γ2(x→′−x→)2w2]superscriptsubscript𝐼av212𝛾superscript𝑒𝛾1superscript𝑤2superscript𝑘2𝛾2superscriptsuperscript→𝑥′→𝑥2superscript𝑤2\displaystyle-I_{\rm av}^{2}(1-2\gamma)e^{(\gamma-1)w^{2}k^{2}}\exp\left[-\frac{\gamma}{2}\frac{(\vec{x}^{\prime}-\vec{x})^{2}}{w^{2}}\right]</td>
<td></td>
</tr>
<tr>
<td></td>
<td></td>
<td></td>
<td>×sin[γk→1⋅(x→′−x→)]sin[γk→2⋅(x→′−x→)].absent⋅𝛾subscript→𝑘1superscript→𝑥′→𝑥⋅𝛾subscript→𝑘2superscript→𝑥′→𝑥\displaystyle\times\sin[\gamma\vec{k}_{1}\cdot(\vec{x}^{\prime}-\vec{x})]\sin[\gamma\vec{k}_{2}\cdot(\vec{x}^{\prime}-\vec{x})].</td>
<td></td>
</tr>
</tbody>
</table>
We define the combined covariance matrix $`\Sigma_{G}`$ as
<table>
<tbody>
<tr>
<td></td>
<td>ΣG=(ΣG[11]ΣG[12]ΣG[12]ΣG[22]).subscriptΣ𝐺matrixsuperscriptsubscriptΣ𝐺delimited-[]11superscriptsubscriptΣ𝐺delimited-[]12superscriptsubscriptΣ𝐺delimited-[]12superscriptsubscriptΣ𝐺delimited-[]22\Sigma_{G}=\left(\matrix{\Sigma_{G}^{[11]}&\Sigma_{G}^{[12]}\cr\Sigma_{G}^{[12]}&\Sigma_{G}^{[22]}}\right).</td>
<td></td>
<td>(22)</td>
</tr>
</tbody>
</table>
We consider the joint probability distribution for all the Gabor coefficients to be Gaussian. This, of course, is not true, as we see from the nonzero even moments in Section [3.2](#S3.SS2 "3.2 Computation of higher moments ‣ 3 Statistics of the Gabor coefficients ‣ The entropy of keys derived from laser speckle"). However, for a given mean and covariance matrix, the Gaussian distribution has a higher entropy than any other distribution. Hence our procedure yields an upper bound on the entropy.
We furthermore assume that the detector noise is Gaussian, independent of the intensity and independent for each pixel. We can then apply the well known channel capacity formula (see e.g. \[[2](#bib.bib2)\]), which expresses the mutual information between a source and a noisy detection as the logarithm of a signal to noise ratio. Let $`\overset{\rightarrow}{G}`$ be a vector of Gabor coefficients, and $`\delta\hspace{0pt}\overset{\rightarrow}{G}`$ the detector noise in these coefficients, then
<table>
<tbody>
<tr>
<td></td>
<td>𝐈(G→;G→+δG→)=12log2|det(𝟏+ΣN−1ΣG)|.𝐈→𝐺→𝐺𝛿→𝐺12subscript21superscriptsubscriptΣ𝑁1subscriptΣ𝐺{\bf I}(\vec{G};\vec{G}+\delta\vec{G})=\mbox{$\textstyle\frac{1}{2}$}\log_{2}|\det({\bf 1}+\Sigma_{N}^{-1}\Sigma_{G})|.</td>
<td></td>
<td>(23)</td>
</tr>
</tbody>
</table>
Here the matrix $`\Sigma_{N}`$ is the covariance matrix of the noise. We determine $`\Sigma_{N}`$ as follows. The independence of the noise in each pixel yields
<table>
<tbody>
<tr>
<td></td>
<td>⟨δI(x→)δI(x→′)⟩n=NI2tδ(x→−x→′),subscriptdelimited-⟨⟩𝛿𝐼→𝑥𝛿𝐼superscript→𝑥′nsuperscriptsubscript𝑁𝐼2𝑡𝛿→𝑥superscript→𝑥′\langle\delta I(\vec{x})\;\delta I(\vec{x}^{\prime})\rangle_{\rm n}=N_{I}^{2}t\;\delta(\vec{x}-\vec{x}^{\prime}),</td>
<td></td>
<td>(24)</td>
</tr>
</tbody>
</table>
where $`N_{I}`$ denotes the noise amplitude, $`t`$ is the area of a detector pixel and the notation $`{\langle \cdot \rangle}_{n}`$ denotes a noise average. Note that $`N_{I}`$ is lower bounded by the shot noise.
Using ([24](#S3.E24 "In 3.3 Estimated entropy of a set of Gabor coefficients for given detector noise ‣ 3 Statistics of the Gabor coefficients ‣ The entropy of keys derived from laser speckle")) and the definition of the Gabor coefficients ([1](#S1.E1 "In 1.2 Binarized Gabor coefficients ‣ 1 Introduction ‣ The entropy of keys derived from laser speckle")) we obtain the following expression for the covariance of the noise in the Gabor coefficients
<table>
<tbody>
<tr>
<td></td>
<td>⟨δG(w,k→1,x→1)δG(w,k→2,x→2)⟩n=NI2t∫d2xΓIM(w,k→1,x→1,x→)ΓIM(w,k→2,x→2,x→).subscriptdelimited-⟨⟩𝛿𝐺𝑤subscript→𝑘1subscript→𝑥1𝛿𝐺𝑤subscript→𝑘2subscript→𝑥2nsuperscriptsubscript𝑁𝐼2𝑡superscriptd2𝑥subscriptΓIM𝑤subscript→𝑘1subscript→𝑥1→𝑥subscriptΓIM𝑤subscript→𝑘2subscript→𝑥2→𝑥\left\langle\delta G(w,\vec{k}_{1},\vec{x}_{1})\;\delta G(w,\vec{k}_{2},\vec{x}_{2})\right\rangle_{\rm n}=N_{I}^{2}t\int\!{\rm d}^{2}x\;\Gamma_{\rm IM}(w,\vec{k}_{1},\vec{x}_{1},\vec{x})\Gamma_{\rm IM}(w,\vec{k}_{2},\vec{x}_{2},\vec{x}).</td>
<td></td>
<td>(25)</td>
</tr>
</tbody>
</table>
Evaluation of the integral is straightforward (it is equivalent to the derivation of $`\Sigma_{G}`$ in \[[17](#bib.bib17)\] in the limit $`\Sigma\rightarrow 0`$) and yields, in the special case $`{|{\overset{\rightarrow}{k}}_{1}|} = {|{\overset{\rightarrow}{k}}_{2}|} = k`$,
<table>
<tbody>
<tr>
<td></td>
<td>ΣN[jj](x→,x→′)superscriptsubscriptΣ𝑁delimited-[]𝑗𝑗→𝑥superscript→𝑥′\displaystyle\Sigma_{N}^{[jj]}(\vec{x},\vec{x}^{\prime})</td>
<td>=\displaystyle=</td>
<td>NI2t8πw2e−(x→′−x→)24w2[cosk→j⋅(x→′−x→)−e−w2k2]superscriptsubscript𝑁𝐼2𝑡8𝜋superscript𝑤2superscript𝑒superscriptsuperscript→𝑥′→𝑥24superscript𝑤2delimited-[]⋅subscript→𝑘𝑗superscript→𝑥′→𝑥superscript𝑒superscript𝑤2superscript𝑘2\displaystyle\hphantom{-}\frac{N_{I}^{2}t}{8\pi w^{2}}e^{-\frac{(\vec{x}^{\prime}-\vec{x})^{2}}{4w^{2}}}\left[\cos\vec{k}_{j}\cdot(\vec{x}^{\prime}-\vec{x})-e^{-w^{2}k^{2}}\right]</td>
<td></td>
<td>(26)</td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td>ΣN[12](x→,x→′)superscriptsubscriptΣ𝑁delimited-[]12→𝑥superscript→𝑥′\displaystyle\Sigma_{N}^{[12]}(\vec{x},\vec{x}^{\prime})</td>
<td>=\displaystyle=</td>
<td>−NI2t8πw2e−(x→′−x→)24w2e−12w2k2sin[12k→1⋅(x→′−x→)]sin[12k→2⋅(x→′−x→)].superscriptsubscript𝑁𝐼2𝑡8𝜋superscript𝑤2superscript𝑒superscriptsuperscript→𝑥′→𝑥24superscript𝑤2superscript𝑒12superscript𝑤2superscript𝑘2⋅12subscript→𝑘1superscript→𝑥′→𝑥⋅12subscript→𝑘2superscript→𝑥′→𝑥\displaystyle-\frac{N_{I}^{2}t}{8\pi w^{2}}e^{-\frac{(\vec{x}^{\prime}-\vec{x})^{2}}{4w^{2}}}e^{-\mbox{$\textstyle\frac{1}{2}$}w^{2}k^{2}}\sin[\mbox{$\textstyle\frac{1}{2}$}\vec{k}_{1}\cdot(\vec{x}^{\prime}-\vec{x})]\sin[\mbox{$\textstyle\frac{1}{2}$}\vec{k}_{2}\cdot(\vec{x}^{\prime}-\vec{x})].</td>
<td></td>
<td>(27)</td>
</tr>
</tbody>
</table>
$`\Sigma_{N}`$ has the same block structure as $`\Sigma_{G}`$. Notice that both $`\Sigma_{G}`$ and $`\Sigma_{N}`$ depend on $`\overset{\rightarrow}{x}`$ and $`{\overset{\rightarrow}{x}}'`$ only through the difference $`({{\overset{\rightarrow}{x}}' - \overset{\rightarrow}{x}})`$. This allows for efficient computation of the determinant in ([23](#S3.E23 "In 3.3 Estimated entropy of a set of Gabor coefficients for given detector noise ‣ 3 Statistics of the Gabor coefficients ‣ The entropy of keys derived from laser speckle")) by diagonalisation in the Fourier domain. Notice further that the off-diagonal blocks ($`{\overset{\rightarrow}{k}}_{1} \perp {\overset{\rightarrow}{k}}_{2}`$) have very small values compared to the diagonal blocks due to the presence of the sine factors.
We compute the Fourier transforms as follows. Strictly speaking, the transform is a summation over the finite $`\overset{\rightarrow}{x}`$-grid. We denote the size of the grid as $`L`$. However, we will sum to infinity, since the error introduced in this way is only an edge effect. The finiteness of $`L`$ is still reflected by the discretisation of the momentum $`\overset{\rightarrow}{p}`$ conjugate to $`{\overset{\rightarrow}{x}}' - \overset{\rightarrow}{x}`$. All momenta are multiples of $`\pi/L`$. The highest momentum is determined by the lattice constant $`\ell`$ of the $`\overset{\rightarrow}{x}`$- grid. As a second approximation, we will replace the summations over $`\overset{\rightarrow}{x}`$ and $`{\overset{\rightarrow}{x}}'`$ by integrations, i.e. $`\sum_{\overset{\rightarrow}{x}}\rightarrow{\ell^{- 2}\hspace{0pt}{\int{d^{2}\hspace{0pt}x}}}`$. This is a good approximation provided that the lattice constant is significantly smaller than the average speckle size. Using this procedure we obtain
<table>
<tbody>
<tr>
<td></td>
<td>Σ~G[jj](p→)superscriptsubscript~Σ𝐺delimited-[]𝑗𝑗→𝑝\displaystyle\tilde{\Sigma}_{G}^{[jj]}(\vec{p})</td>
<td>:=assign\displaystyle:=</td>
<td>ℓ−2∫d2(x→′−x→)ΣG[jj]e−ip→⋅(x→′−x→)superscriptℓ2superscriptd2superscript→𝑥′→𝑥superscriptsubscriptΣ𝐺delimited-[]𝑗𝑗superscript𝑒⋅𝑖→𝑝superscript→𝑥′→𝑥\displaystyle\ell^{-2}\int\!{\rm d}^{2}(\vec{x}^{\prime}-\vec{x})\;\Sigma_{G}^{[jj]}e^{-i\vec{p}\cdot(\vec{x}^{\prime}-\vec{x})}</td>
<td></td>
<td rowspan="2">(28)</td>
</tr>
<tr>
<td></td>
<td></td>
<td>=\displaystyle=</td>
<td>4πIav2w2ℓ2(12γ−1)e−w2k2e−w22γp2sinh2w2k→j⋅p→4𝜋superscriptsubscript𝐼av2superscript𝑤2superscriptℓ212𝛾1superscript𝑒superscript𝑤2superscript𝑘2superscript𝑒superscript𝑤22𝛾superscript𝑝2superscript2⋅superscript𝑤2subscript→𝑘𝑗→𝑝\displaystyle 4\pi I_{\rm av}^{2}\frac{w^{2}}{\ell^{2}}(\frac{1}{2\gamma}-1)e^{-w^{2}k^{2}}e^{-\frac{w^{2}}{2\gamma}p^{2}}\sinh^{2}w^{2}\vec{k}_{j}\cdot\vec{p}</td>
<td></td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td>Σ~N[jj](p→)superscriptsubscript~Σ𝑁delimited-[]𝑗𝑗→𝑝\displaystyle\tilde{\Sigma}_{N}^{[jj]}(\vec{p})</td>
<td>:=assign\displaystyle:=</td>
<td>ℓ−2∫d2(x→′−x→)ΣN[jj]e−ip→⋅(x→′−x→)superscriptℓ2superscriptd2superscript→𝑥′→𝑥superscriptsubscriptΣ𝑁delimited-[]𝑗𝑗superscript𝑒⋅𝑖→𝑝superscript→𝑥′→𝑥\displaystyle\ell^{-2}\int\!{\rm d}^{2}(\vec{x}^{\prime}-\vec{x})\;\Sigma_{N}^{[jj]}e^{-i\vec{p}\cdot(\vec{x}^{\prime}-\vec{x})}</td>
<td></td>
<td rowspan="2">(29)</td>
</tr>
<tr>
<td></td>
<td></td>
<td>=\displaystyle=</td>
<td>NI2tℓ2e−w2k2e−w2p2sinh2w2k→j⋅p→.superscriptsubscript𝑁𝐼2𝑡superscriptℓ2superscript𝑒superscript𝑤2superscript𝑘2superscript𝑒superscript𝑤2superscript𝑝2superscript2⋅superscript𝑤2subscript→𝑘𝑗→𝑝\displaystyle N_{I}^{2}\frac{t}{\ell^{2}}\;e^{-w^{2}k^{2}}e^{-w^{2}p^{2}}\sinh^{2}w^{2}\vec{k}_{j}\cdot\vec{p}.</td>
<td></td>
</tr>
</tbody>
</table>
Substitution of ([28](#S3.E28 "In 3.3 Estimated entropy of a set of Gabor coefficients for given detector noise ‣ 3 Statistics of the Gabor coefficients ‣ The entropy of keys derived from laser speckle")) and ([29](#S3.E29 "In 3.3 Estimated entropy of a set of Gabor coefficients for given detector noise ‣ 3 Statistics of the Gabor coefficients ‣ The entropy of keys derived from laser speckle")) into ([23](#S3.E23 "In 3.3 Estimated entropy of a set of Gabor coefficients for given detector noise ‣ 3 Statistics of the Gabor coefficients ‣ The entropy of keys derived from laser speckle")) gives us the mutual information for one Gabor direction,
<table>
<tbody>
<tr>
<td></td>
<td>𝐈(G[j];G[j]+δG[j])𝐈superscript𝐺delimited-[]𝑗superscript𝐺delimited-[]𝑗𝛿superscript𝐺delimited-[]𝑗\displaystyle{\bf I}(G^{[j]};G^{[j]}+\delta G^{[j]})</td>
<td>=\displaystyle=</td>
<td>12log2|det(1+(ΣN[jj])−1ΣG[jj])|=12trlog2|1+(ΣN[jj])−1ΣG[jj]|12subscript21superscriptsuperscriptsubscriptΣ𝑁delimited-[]𝑗𝑗1superscriptsubscriptΣ𝐺delimited-[]𝑗𝑗12trsubscript21superscriptsuperscriptsubscriptΣ𝑁delimited-[]𝑗𝑗1superscriptsubscriptΣ𝐺delimited-[]𝑗𝑗\displaystyle\mbox{$\textstyle\frac{1}{2}$}\log_{2}\left|\det\left(1+(\Sigma_{N}^{[jj]})^{-1}\Sigma_{G}^{[jj]}\right)\right|=\mbox{$\textstyle\frac{1}{2}$}{\rm tr}\;\log_{2}\left|1+(\Sigma_{N}^{[jj]})^{-1}\Sigma_{G}^{[jj]}\right|</td>
<td></td>
<td rowspan="2">(30)</td>
</tr>
<tr>
<td></td>
<td></td>
<td>=\displaystyle=</td>
<td>12∑p→log2(1+c1e−c2p2),12subscript→𝑝subscript21subscript𝑐1superscript𝑒subscript𝑐2superscript𝑝2\displaystyle\mbox{$\textstyle\frac{1}{2}$}\sum_{\vec{p}}\log_{2}\left(1+c_{1}e^{-c_{2}p^{2}}\right),</td>
<td></td>
</tr>
</tbody>
</table>
where we have defined the constants $`c_{1}`$, $`c_{2}`$ as
<table>
<tbody>
<tr>
<td></td>
<td>c1=(12γ−1)4πIav2w2NI2t=2πΣ2Iav2NI2M2tsubscript𝑐112𝛾14𝜋superscriptsubscript𝐼av2superscript𝑤2superscriptsubscript𝑁𝐼2𝑡2𝜋superscriptΣ2superscriptsubscript𝐼av2superscriptsubscript𝑁𝐼2superscript𝑀2𝑡\displaystyle c_{1}=(\frac{1}{2\gamma}-1)\frac{4\pi I_{\rm av}^{2}w^{2}}{N_{I}^{2}t}=2\pi\Sigma^{2}\frac{I_{\rm av}^{2}}{N_{I}^{2}}\frac{M^{2}}{t}</td>
<td>;absent\displaystyle\quad;\quad</td>
<td>c2=(12γ−1)w2=12M2Σ2.subscript𝑐212𝛾1superscript𝑤212superscript𝑀2superscriptΣ2\displaystyle c_{2}=(\frac{1}{2\gamma}-1)w^{2}=\mbox{$\textstyle\frac{1}{2}$}M^{2}\Sigma^{2}.</td>
<td></td>
<td>(31)</td>
</tr>
</tbody>
</table>
Note that $`c_{2}`$ is proportional to the average speckle area, while $`c_{1}`$ plays the role of a signal to noise ratio (SNR). Note too that all dependence on $`w`$ and $`\overset{\rightarrow}{k}`$ has disappeared from ([30](#S3.E30 "In 3.3 Estimated entropy of a set of Gabor coefficients for given detector noise ‣ 3 Statistics of the Gabor coefficients ‣ The entropy of keys derived from laser speckle")). The reason is that we are computing a generic upper bound. Computation of the actual amount of extracted information will in general depend on $`w`$ and $`\overset{\rightarrow}{k}`$.
The summation domain of $`\sum_{\overset{\rightarrow}{p}}`$ is given by $`{(p_{x},p_{y})} = {{(i,j)}\hspace{0pt}\frac{\pi}{L}}`$, with $`{i,j} \in {\mathbb{Z}}`$ and $`{|p_{x}|} < {\pi/\ell}`$, $`{|p_{y}|} < {\pi/\ell}`$. Eq. ([30](#S3.E30 "In 3.3 Estimated entropy of a set of Gabor coefficients for given detector noise ‣ 3 Statistics of the Gabor coefficients ‣ The entropy of keys derived from laser speckle")) can be further evaluated by approximating the momentum sum by an integration. Notice that the summand in ([30](#S3.E30 "In 3.3 Estimated entropy of a set of Gabor coefficients for given detector noise ‣ 3 Statistics of the Gabor coefficients ‣ The entropy of keys derived from laser speckle")) only depends on the length of $`\overset{\rightarrow}{p}`$. Hence the computation is simplified in polar coordinates,
<table>
<tbody>
<tr>
<td></td>
<td>∑p→f(p2)≈(Lπ)2∫π/Lπ/ℓdpx∫π/Lπ/ℓdpyf(p2)=L2π∫(π/L)2(π/ℓ)2dp2f(p2).subscript→𝑝𝑓superscript𝑝2superscript𝐿𝜋2superscriptsubscript𝜋𝐿𝜋ℓdifferential-dsubscript𝑝𝑥superscriptsubscript𝜋𝐿𝜋ℓdifferential-dsubscript𝑝𝑦𝑓superscript𝑝2superscript𝐿2𝜋superscriptsubscriptsuperscript𝜋𝐿2superscript𝜋ℓ2differential-dsuperscript𝑝2𝑓superscript𝑝2\sum_{\vec{p}}f(p^{2})\approx(\frac{L}{\pi})^{2}\int_{\pi/L}^{\pi/\ell}\!{\rm d}p_{x}\int_{\pi/L}^{\pi/\ell}\!{\rm d}p_{y}\;f(p^{2})=\frac{L^{2}}{\pi}\int_{(\pi/L)^{2}}^{(\pi/\ell)^{2}}{\rm d}p^{2}\;f(p^{2}).</td>
<td></td>
<td>(32)</td>
</tr>
</tbody>
</table>
Applying the approximation ([32](#S3.E32 "In 3.3 Estimated entropy of a set of Gabor coefficients for given detector noise ‣ 3 Statistics of the Gabor coefficients ‣ The entropy of keys derived from laser speckle")) to ([30](#S3.E30 "In 3.3 Estimated entropy of a set of Gabor coefficients for given detector noise ‣ 3 Statistics of the Gabor coefficients ‣ The entropy of keys derived from laser speckle")) we obtain
<table>
<tbody>
<tr>
<td></td>
<td>𝐈(G[j];G[j]+δG[j])≈L22π[−1c2Dilog(−expc2p2c1)−12c2p4+p2lnc1]p=π/Lπ/ℓ.𝐈superscript𝐺delimited-[]𝑗superscript𝐺delimited-[]𝑗𝛿superscript𝐺delimited-[]𝑗superscript𝐿22𝜋superscriptsubscriptdelimited-[]1subscript𝑐2Dilogsubscript𝑐2superscript𝑝2subscript𝑐112subscript𝑐2superscript𝑝4superscript𝑝2subscript𝑐1𝑝𝜋𝐿𝜋ℓ{\bf I}(G^{[j]};G^{[j]}+\delta G^{[j]})\approx\frac{L^{2}}{2\pi}\left[-\frac{1}{c_{2}}{\rm Dilog}(-\frac{\exp c_{2}p^{2}}{c_{1}})-\mbox{$\textstyle\frac{1}{2}$}c_{2}p^{4}+p^{2}\ln c_{1}\right]_{p=\pi/L}^{\pi/\ell}.</td>
<td></td>
<td>(33)</td>
</tr>
</tbody>
</table>
Here Dilog is the dilogarithm function. Note that ([33](#S3.E33 "In 3.3 Estimated entropy of a set of Gabor coefficients for given detector noise ‣ 3 Statistics of the Gabor coefficients ‣ The entropy of keys derived from laser speckle")) is not expressed in bits but in natural units (‘nats’), i.e. using the natural logarithm $`\ln`$ instead of the base-2 $`\log_{2}`$.
Eq. ([33](#S3.E33 "In 3.3 Estimated entropy of a set of Gabor coefficients for given detector noise ‣ 3 Statistics of the Gabor coefficients ‣ The entropy of keys derived from laser speckle")) can be further evaluated if $`c_{1}`$ is very large (we call this the ‘large SNR’ case), or when $`c_{1}`$ is very small (‘small SNR’). We define a quantity $`y`$ as
<table>
<tbody>
<tr>
<td></td>
<td>y=c1e−c2π2/ℓ2.𝑦subscript𝑐1superscript𝑒subscript𝑐2superscript𝜋2superscriptℓ2y=c_{1}e^{-c_{2}\pi^{2}/\ell^{2}}.</td>
<td></td>
<td>(34)</td>
</tr>
</tbody>
</table>
The crossover between the two regimes lies around $`y = 1`$.
#### Large SNR
In this case we have $`y \gg 1`$. Applying the asymptotic relation
<table>
<tbody>
<tr>
<td></td>
<td>Dilog(x)=x+𝒪(x2)for|x|≪1formulae-sequenceDilog𝑥𝑥𝒪superscript𝑥2formuch-less-than𝑥1{\rm Dilog}(x)=x+{\cal O}(x^{2})\quad\mbox{for}\quad|x|\ll 1</td>
<td></td>
<td>(35)</td>
</tr>
</tbody>
</table>
to ([33](#S3.E33 "In 3.3 Estimated entropy of a set of Gabor coefficients for given detector noise ‣ 3 Statistics of the Gabor coefficients ‣ The entropy of keys derived from laser speckle")) we obtain
<table>
<tbody>
<tr>
<td></td>
<td>𝐈(G[j];G[j]+δG[j])≈πL22ℓ2[lnc1−c2π22ℓ2+𝒪(y−1)+𝒪(ℓ2M2L4)+𝒪(ℓ2c1L2)].𝐈superscript𝐺delimited-[]𝑗superscript𝐺delimited-[]𝑗𝛿superscript𝐺delimited-[]𝑗𝜋superscript𝐿22superscriptℓ2delimited-[]subscript𝑐1subscript𝑐2superscript𝜋22superscriptℓ2𝒪superscript𝑦1𝒪superscriptℓ2superscript𝑀2superscript𝐿4𝒪superscriptℓ2subscript𝑐1superscript𝐿2{\bf I}(G^{[j]};G^{[j]}+\delta G^{[j]})\approx\frac{\pi L^{2}}{2\ell^{2}}\left[\ln c_{1}-\frac{c_{2}\pi^{2}}{2\ell^{2}}+{\cal O}(y^{-1})+{\cal O}(\frac{\ell^{2}M^{2}}{L^{4}})+{\cal O}(\frac{\ell^{2}}{c_{1}L^{2}})\right].</td>
<td></td>
<td>(36)</td>
</tr>
</tbody>
</table>
#### Small SNR
In this case we have $`y \ll 1`$. Using ([35](#S3.E35 "In Large SNR ‣ 3.3 Estimated entropy of a set of Gabor coefficients for given detector noise ‣ 3 Statistics of the Gabor coefficients ‣ The entropy of keys derived from laser speckle")) and the asymptotic behaviour
<table>
<tbody>
<tr>
<td></td>
<td>Dilog(−x)=−π2/6−12(lnx)2+𝒪(x−1)forx→∞formulae-sequenceDilog𝑥superscript𝜋2612superscript𝑥2𝒪superscript𝑥1for→𝑥{\rm Dilog}(-x)=-\pi^{2}/6-\mbox{$\textstyle\frac{1}{2}$}(\ln x)^{2}+{\cal O}(x^{-1})\quad\mbox{for}\quad x\rightarrow\infty</td>
<td></td>
<td>(37)</td>
</tr>
</tbody>
</table>
we get
<table>
<tbody>
<tr>
<td></td>
<td>𝐈(G[jj];G[jj]+δG[jj])≈L24πc2[(lnc1)2+π23+𝒪(y)+𝒪(1c1)+𝒪(M2L2)].𝐈superscript𝐺delimited-[]𝑗𝑗superscript𝐺delimited-[]𝑗𝑗𝛿superscript𝐺delimited-[]𝑗𝑗superscript𝐿24𝜋subscript𝑐2delimited-[]superscriptsubscript𝑐12superscript𝜋23𝒪𝑦𝒪1subscript𝑐1𝒪superscript𝑀2superscript𝐿2{\bf I}(G^{[jj]};G^{[jj]}+\delta G^{[jj]})\approx\frac{L^{2}}{4\pi c_{2}}\left[(\ln c_{1})^{2}+\frac{\pi^{2}}{3}+{\cal O}(y)+{\cal O}(\frac{1}{c_{1}})+{\cal O}(\frac{M^{2}}{L^{2}})\right].</td>
<td></td>
<td>(38)</td>
</tr>
</tbody>
</table>
[Refer to caption](/html/0710.5002/assets/x1.png)
Figure 1: Mutual information, per average speckle area, between a noiseless and a noisy vector $`G^{\lbrack j\rbrack}`$ of Gabor coefficients, plotted as a function of $`I_{av}/N_{I}`$. The other parameters are: $`L = 800`$ pixels, $`M = 3`$ pixels, $`\ell = 5`$ pixels, $`t = 1`$ pixel <sup>2</sup>. The radius of an average speckle has been set, somewhat arbitrarily, to $`M`$.
Eq. ([33](#S3.E33 "In 3.3 Estimated entropy of a set of Gabor coefficients for given detector noise ‣ 3 Statistics of the Gabor coefficients ‣ The entropy of keys derived from laser speckle")) is plotted in Fig. [1](#S3.F1 "Figure 1 ‣ Small SNR ‣ 3.3 Estimated entropy of a set of Gabor coefficients for given detector noise ‣ 3 Statistics of the Gabor coefficients ‣ The entropy of keys derived from laser speckle") on a logarithmic scale; The factor $`L^{2}`$ was replaced by $`\pi\hspace{0pt}M^{2}`$ to obtain a result per average speckle area. We see the transition from parabolic behaviour to linear behaviour around $`{I_{av}/N_{I}} = 1`$. The linear part of the curve is the usual ‘log SNR’ dependence, but here it occurs with a nonzero offset.
## 4 Perturbing speckle patterns
### 4.1 Phase perturbations
We introduce a method for perturbing a speckle pattern in the model of Section [2.1](#S2.SS1 "2.1 Motivation and definitions ‣ 2 The random phase model ‣ The entropy of keys derived from laser speckle"). We shift all the phases by random amounts,
<table>
<tbody>
<tr>
<td></td>
<td>φa→→φ^a→=φa→+εa→.→subscript𝜑→𝑎subscript^𝜑→𝑎subscript𝜑→𝑎subscript𝜀→𝑎\varphi_{\vec{a}}\rightarrow\hat{\varphi}_{\vec{a}}=\varphi_{\vec{a}}+\varepsilon_{\vec{a}}.</td>
<td></td>
<td>(39)</td>
</tr>
</tbody>
</table>
This is done independently for all locations $`\overset{\rightarrow}{a}`$. The perturbations $`\varepsilon`$ are chosen from a distribution that is uniform on the interval $`({- q},q\rbrack`$ and zero elsewhere. The parameter $`q`$ is the ‘strength’ of the perturbation. The maximum value is $`q = \pi`$, resulting in a completely independent speckle pattern. The minimum value is $`q = {{\bigtriangleup\hspace{0pt}\varphi}/2}`$ (see Section [2.2](#S2.SS2 "2.2 Entropy of the source ‣ 2 The random phase model ‣ The entropy of keys derived from laser speckle")), in accordance with the uncertainty relation, and represents no visible change at all. The exact relation between $`q`$ and the actual physical perturbation is hard to define. We will come back to this in Section [5](#S5 "5 Comparison to experimental results ‣ The entropy of keys derived from laser speckle").
We use ‘hat’ notation ($`\hat{}`$) for perturbed quantities, e.g. $`\hat{I}\hspace{0pt}{(\overset{\rightarrow}{x})}`$ is the intensity after perturbation. Note that the statistical properties of $`{\hat{\alpha}}_{\overset{\rightarrow}{a}}`$ are exactly the same as the properties of $`\alpha_{\overset{\rightarrow}{a}}`$, i.e. uniformly distributed over the unit circle in the complex plane. Hence the statistics of $`\hat{I}`$ and $`\hat{G}`$ are the same as for the unperturbed quantities. This is precisely what we want from our model, for there must be no ‘preference’ in the formalism for either the perturbed or unperturbed speckle pattern.
Definition ([39](#S4.E39 "In 4.1 Phase perturbations ‣ 4 Perturbing speckle patterns ‣ The entropy of keys derived from laser speckle")) allows us to study perturbations quantitatively, on a continuous scale as a function of $`q`$. Of special interest are the correlations between states before and after a perturbation.
We introduce the notation $`\left\langle \cdot \right\rangle_{\varepsilon}`$ for averaging with respect to the random variables $`\{\varepsilon_{\overset{\rightarrow}{a}}\}`$. We list several properties that will be useful in later sections. The average effect on $`\alpha_{\overset{\rightarrow}{a}}`$ is multiplication by the following factor:
<table>
<tbody>
<tr>
<td></td>
<td>⟨expiεa→⟩ε=∫−qqdεexpiε2q=sinqq.subscriptdelimited-⟨⟩𝑖subscript𝜀→𝑎𝜀superscriptsubscript𝑞𝑞differential-d𝜀𝑖𝜀2𝑞𝑞𝑞\left\langle\exp i\varepsilon_{\vec{a}}\right\rangle_{\varepsilon}=\int_{-q}^{q}\!\!{\rm d}\varepsilon\;\frac{\exp i\varepsilon}{2q}=\frac{\sin q}{q}.</td>
<td></td>
<td>(40)</td>
</tr>
</tbody>
</table>
From this the correlation between perturbed and unperturbed $`\alpha`$ follows,
<table>
<tbody>
<tr>
<td></td>
<td>C(α^a→,αb→):=⟨α^a→αb→∗⟩ε,φ=δa→,b→sinqq.assign𝐶subscript^𝛼→𝑎subscript𝛼→𝑏subscriptdelimited-⟨⟩subscript^𝛼→𝑎subscriptsuperscript𝛼→𝑏𝜀𝜑subscript𝛿→𝑎→𝑏𝑞𝑞C(\hat{\alpha}_{\vec{a}},\alpha_{\vec{b}}):=\left\langle\hat{\alpha}_{\vec{a}}\alpha^{*}_{\vec{b}}\right\rangle_{\varepsilon,\varphi}=\delta_{\vec{a},\vec{b}}\frac{\sin q}{q}.</td>
<td></td>
<td>(41)</td>
</tr>
</tbody>
</table>
Here $`\left\langle \cdot \right\rangle_{\varepsilon,\varphi}`$ indicates averaging first over $`\varepsilon`$ and then over $`\varphi`$. Other useful identities are
<table>
<tbody>
<tr>
<td></td>
<td>⟨expiνεa→⟩ε=sinνqνqsubscriptdelimited-⟨⟩𝑖𝜈subscript𝜀→𝑎𝜀𝜈𝑞𝜈𝑞\left\langle\exp i\nu\varepsilon_{\vec{a}}\right\rangle_{\varepsilon}=\frac{\sin\nu q}{\nu q}</td>
<td></td>
<td>(42)</td>
</tr>
</tbody>
</table>
and
<table>
<tbody>
<tr>
<td></td>
<td>⟨expi(εa→−εb→)⟩ε=δa→,b→+(1−δa→,b→)(sinqq)2.subscriptdelimited-⟨⟩𝑖subscript𝜀→𝑎subscript𝜀→𝑏𝜀subscript𝛿→𝑎→𝑏1subscript𝛿→𝑎→𝑏superscript𝑞𝑞2\left\langle\exp i(\varepsilon_{\vec{a}}-\varepsilon_{\vec{b}})\right\rangle_{\varepsilon}=\delta_{\vec{a},\vec{b}}+(1-\delta_{\vec{a},\vec{b}})\left(\frac{\sin q}{q}\right)^{2}.</td>
<td></td>
<td>(43)</td>
</tr>
</tbody>
</table>
For convenience later on we introduce the following shorthand notation,
<table>
<tbody>
<tr>
<td></td>
<td>Q:=sin2qq2.assign𝑄superscript2𝑞superscript𝑞2Q:=\frac{\sin^{2}q}{q^{2}}.</td>
<td></td>
<td>(44)</td>
</tr>
</tbody>
</table>
Note that $`Q = 1`$ in the case of zero perturbation, and $`Q = 0`$ when the perturbation has maximum strength ($`q = \pi`$).
### 4.2 Mutual information in the source plane
The mutual information \[[2](#bib.bib2)\] between two stochastic variables $`X`$ and $`Y`$ is denoted as $`\mathbf{I}\hspace{0pt}{(X;Y)}`$. It represents the amount of overlap in the information they carry, and it can be computed as
<sup>1</sup>
<sup>1</sup>1 $`H\hspace{0pt}{(X,Y)}`$ stands for the combined entropy of $`X`$ and $`Y`$. The notation $`H\hspace{0pt}{(\left. Y \middle| X \right.)}`$ denotes the uncertainty in $`Y`$ given knowledge of $`X`$.
<table>
<tbody>
<tr>
<td></td>
<td>𝐈(X;Y)=H(X)+H(Y)−H(X,Y)=H(Y)−H(Y|X).𝐈𝑋𝑌𝐻𝑋𝐻𝑌𝐻𝑋𝑌𝐻𝑌𝐻conditional𝑌𝑋{\bf I}(X;Y)=H(X)+H(Y)-H(X,Y)=H(Y)-H(Y|X).</td>
<td></td>
<td>(45)</td>
</tr>
</tbody>
</table>
This quantity is of great importance in cryptography. Let $`X`$ and $`Y`$ be two noisy versions of the same secret, possessed by two parties respectively. These parties can derive a common secret key from the variable that they hold. The theoretical maximum length of their common key is precisely given by the mutual information $`\mathbf{I}\hspace{0pt}{(X;Y)}`$.
In our case, $`X`$ is the unperturbed speckle source $`\alpha`$ and $`Y`$ is the perturbed source $`\hat{\alpha}`$. Their mutual information gives an absolute physical upper bound on the length of the key that can be derived from a speckle pattern in a reproducible way, for a given noise level $`q`$. The entropy $`H\hspace{0pt}{\lbrack\alpha\rbrack}`$ is given by ([6](#S2.E6 "In 2.2 Entropy of the source ‣ 2 The random phase model ‣ The entropy of keys derived from laser speckle")). The conditional entropy $`H\hspace{0pt}{\lbrack\left. \hat{\alpha} \middle| \alpha \right.\rbrack}`$ equals the uncertainty in the set $`\{{\hat{\alpha}}_{\overset{\rightarrow}{a}}\}`$ if $`\{\alpha_{\overset{\rightarrow}{a}}\}`$ is known. This is precisely the amount of information contained in the perturbation $`\{\varepsilon_{\overset{\rightarrow}{a}}\}`$. Thus we have
<table>
<tbody>
<tr>
<td></td>
<td>H[α^|α]=H[ε]=Nreglog2q△φ.𝐻delimited-[]conditional^𝛼𝛼𝐻delimited-[]𝜀subscript𝑁reg2𝑞△𝜑H[\hat{\alpha}|\alpha]=H[\varepsilon]=N_{\rm reg}\log\frac{2q}{\triangle\varphi}.</td>
<td></td>
<td>(46)</td>
</tr>
</tbody>
</table>
Substitution of ([6](#S2.E6 "In 2.2 Entropy of the source ‣ 2 The random phase model ‣ The entropy of keys derived from laser speckle")) and (conditional) into ([45](#S4.E45 "In 4.2 Mutual information in the source plane ‣ 4 Perturbing speckle patterns ‣ The entropy of keys derived from laser speckle")) yields the result,
<table>
<tbody>
<tr>
<td></td>
<td>𝐈(α^;α)=Nreglogπq.𝐈^𝛼𝛼subscript𝑁reg𝜋𝑞{\bf I}(\hat{\alpha};\alpha)=N_{\rm reg}\log\frac{\pi}{q}.</td>
<td></td>
<td>(47)</td>
</tr>
</tbody>
</table>
Clearly, if $`q`$ has its maximum strength ($`\pi`$) then the mutual information is zero. When $`q`$ has its minimum value, $`{\bigtriangleup\hspace{0pt}\varphi}/2`$, then the mutual information is equal to the full entropy ([6](#S2.E6 "In 2.2 Entropy of the source ‣ 2 The random phase model ‣ The entropy of keys derived from laser speckle")) of the source.
[Refer to caption](/html/0710.5002/assets/x2.png)
Figure 2: Mutual information between unperturbed and perturbed source, as a function of the perturbation strength $`q`$.
### 4.3 Effect of perturbations on the intensity
The effect of a perturbation on the intensity is computed as follows. First we average the perturbed intensity $`\hat{I}\hspace{0pt}{(\overset{\rightarrow}{x})}`$ over the perturbations $`\varepsilon_{\overset{\rightarrow}{a}}`$. Using the representation ([7](#S2.E7 "In 2.3 Statistics of the intensity ‣ 2 The random phase model ‣ The entropy of keys derived from laser speckle")) and the identity ([43](#S4.E43 "In 4.1 Phase perturbations ‣ 4 Perturbing speckle patterns ‣ The entropy of keys derived from laser speckle")) we get
<table>
<tbody>
<tr>
<td></td>
<td>⟨I^(x→)⟩ε=QI(x→)+(1−Q)Iav.subscriptdelimited-⟨⟩^𝐼→𝑥𝜀𝑄𝐼→𝑥1𝑄subscript𝐼av\left\langle\hat{I}(\vec{x})\right\rangle_{\varepsilon}=QI(\vec{x})+(1-Q)I_{\rm av}.</td>
<td></td>
<td>(48)</td>
</tr>
</tbody>
</table>
Eq. ([48](#S4.E48 "In 4.3 Effect of perturbations on the intensity ‣ 4 Perturbing speckle patterns ‣ The entropy of keys derived from laser speckle")) shows how, on average, the perturbed value $`\hat{I}\hspace{0pt}{(\overset{\rightarrow}{x})}`$ gradually changes from $`I\hspace{0pt}{(\overset{\rightarrow}{x})}`$ to the average intensity $`I_{av}`$ as a function of $`q`$.
Next we compute the two-point correlation function between the perturbed and unperturbed speckle pattern. Taking the expectation value $`\left\langle \cdot \right\rangle_{\varphi}`$ of ([48](#S4.E48 "In 4.3 Effect of perturbations on the intensity ‣ 4 Perturbing speckle patterns ‣ The entropy of keys derived from laser speckle")) multiplied with $`I\hspace{0pt}{({\overset{\rightarrow}{x}}')}`$ we obtain
<table>
<tbody>
<tr>
<td></td>
<td>C(I^(x→),I(x→′)):=⟨I^(x→)I(x→′)⟩ε,φ−Iav2Iav2=QCI(x→,x→′),assign𝐶^𝐼→𝑥𝐼superscript→𝑥′subscriptdelimited-⟨⟩^𝐼→𝑥𝐼superscript→𝑥′𝜀𝜑superscriptsubscript𝐼av2superscriptsubscript𝐼av2𝑄subscript𝐶𝐼→𝑥superscript→𝑥′C(\hat{I}(\vec{x}),I(\vec{x}^{\prime})):=\frac{\left\langle\hat{I}(\vec{x})I(\vec{x}^{\prime})\right\rangle_{\varepsilon,\varphi}-I_{\rm av}^{2}}{I_{\rm av}^{2}}=QC_{I}(\vec{x},\vec{x}^{\prime}),</td>
<td></td>
<td>(49)</td>
</tr>
</tbody>
</table>
where we have used the fact that the correlation with the constant number $`I_{av}`$ in the last term of ([48](#S4.E48 "In 4.3 Effect of perturbations on the intensity ‣ 4 Perturbing speckle patterns ‣ The entropy of keys derived from laser speckle")) vanishes. The result ([49](#S4.E49 "In 4.3 Effect of perturbations on the intensity ‣ 4 Perturbing speckle patterns ‣ The entropy of keys derived from laser speckle")) is intuitive as it states that the total correlation is the product of the ordinary two-point correlation $`C_{I}`$ ([10](#S2.E10 "In 2.3 Statistics of the intensity ‣ 2 The random phase model ‣ The entropy of keys derived from laser speckle")) and a perturbation effect.
The joint probability distribution $`p\hspace{0pt}{(\hat{I},I')}`$ is obtained by computing the moments
$`\left\langle {{\hat{I}}^{n}\hspace{0pt}{(I')}^{m}} \right\rangle_{\varepsilon,\varphi}`$ for all $`n`$ and $`m`$. This exercise is completely analogous to the derivation of ([14](#S2.E14 "In 2.3 Statistics of the intensity ‣ 2 The random phase model ‣ The entropy of keys derived from laser speckle")), but now with extra phase factors $`\exp{i\hspace{0pt}\varepsilon_{\overset{\rightarrow}{a}}}`$. Without showing the derivation, we mention that the averageing procedures $`\left\langle \cdot \right\rangle_{\varepsilon}`$ and $`\left\langle \cdot \right\rangle_{\varphi}`$ in this computation commute and give the result of the $`\varphi`$-average,
<table>
<tbody>
<tr>
<td></td>
<td>⟨[I^(x→)]n[I(x→′)]m⟩φ=Iavn+mn!m!F12(−n,−m;1;|1Nreg∑a→eiεa→eia→⋅(x→−x→′)RM|2).subscriptdelimited-⟨⟩superscriptdelimited-[]^𝐼→𝑥𝑛superscriptdelimited-[]𝐼superscript→𝑥′𝑚𝜑superscriptsubscript𝐼av𝑛𝑚𝑛𝑚subscriptsubscript𝐹12𝑛𝑚1superscript1subscript𝑁regsubscript→𝑎superscript𝑒𝑖subscript𝜀→𝑎superscript𝑒𝑖⋅→𝑎→𝑥superscript→𝑥′𝑅𝑀2\left\langle[\hat{I}(\vec{x})]^{n}[I(\vec{x}^{\prime})]^{m}\right\rangle_{\varphi}=I_{\rm av}^{n+m}n!m!\;{}_{2}F_{1}\left(-n,-m;1;\left|\frac{1}{N_{\rm reg}}\sum_{\vec{a}}e^{i\varepsilon_{\vec{a}}}e^{i\frac{\vec{a}\cdot(\vec{x}-\vec{x}^{\prime})}{RM}}\right|^{2}\right).</td>
<td></td>
<td>(50)</td>
</tr>
</tbody>
</table>
Taking the $`\varepsilon`$-average of ([50](#S4.E50 "In 4.3 Effect of perturbations on the intensity ‣ 4 Perturbing speckle patterns ‣ The entropy of keys derived from laser speckle")) is higly nontrivial in general. However, an estimate with errors of order $`{1/N_{reg}} \ll 1`$ is easily obtained by replacing averages of powers by powers of averages,
<table>
<tbody>
<tr>
<td></td>
<td>⟨[I^(x→)]n[I(x→′)]m⟩φ=Iavn+mn!m!F12(−n,−m;1;QCI(x→,x→′))⋅{1+𝒪(1Nreg)}.subscriptdelimited-⟨⟩superscriptdelimited-[]^𝐼→𝑥𝑛superscriptdelimited-[]𝐼superscript→𝑥′𝑚𝜑⋅superscriptsubscript𝐼av𝑛𝑚𝑛𝑚subscriptsubscript𝐹12𝑛𝑚1𝑄subscript𝐶𝐼→𝑥superscript→𝑥′1𝒪1subscript𝑁reg\left\langle[\hat{I}(\vec{x})]^{n}[I(\vec{x}^{\prime})]^{m}\right\rangle_{\varphi}=I_{\rm av}^{n+m}n!m!\;{}_{2}F_{1}\left(-n,-m;1;\;QC_{I}(\vec{x},\vec{x}^{\prime})\vphantom{\int}\right)\cdot\{1+{\cal O}(\frac{1}{N_{\rm reg}})\}.</td>
<td></td>
<td>(51)</td>
</tr>
</tbody>
</table>
This gives, up to errors of order $`\mathcal{O}\hspace{0pt}{(N_{reg}^{- 1})}`$, a joint probability distribution $`p\hspace{0pt}{({\hat{I}\hspace{0pt}{(\overset{\rightarrow}{x})}},{I\hspace{0pt}{({\overset{\rightarrow}{x}}')}})}`$ of the form ([15](#S2.E15 "In 2.3 Statistics of the intensity ‣ 2 The random phase model ‣ The entropy of keys derived from laser speckle")) with $`C_{I}`$ replaced by $`Q\hspace{0pt}C_{I}`$.
### 4.4 Effect of perturbations on the Gabor coefficients
The properties of the perturbed Gabor coefficients are readily computed. We define the correlation between a perturbed coefficient $`\hat{G}\hspace{0pt}{(s',{\overset{\rightarrow}{k}}',{\overset{\rightarrow}{x}}')}`$ and the unperturbed $`G\hspace{0pt}{(s,\overset{\rightarrow}{k},\overset{\rightarrow}{x})}`$ as
<table>
<tbody>
<tr>
<td></td>
<td>C(G,G^′):=⟨GG^′⟩ε,φ−⟨G⟩ε,φ⟨G^′⟩ε,φσG2,assign𝐶𝐺superscript^𝐺′subscriptdelimited-⟨⟩𝐺superscript^𝐺′𝜀𝜑subscriptdelimited-⟨⟩𝐺𝜀𝜑subscriptdelimited-⟨⟩superscript^𝐺′𝜀𝜑superscriptsubscript𝜎𝐺2C(G,\hat{G}^{\prime}):=\frac{\left\langle G\hat{G}^{\prime}\right\rangle_{\varepsilon,\varphi}-\left\langle G\right\rangle_{\varepsilon,\varphi}\left\langle\hat{G}^{\prime}\right\rangle_{\varepsilon,\varphi}}{\sigma_{G}^{2}},</td>
<td></td>
<td>(52)</td>
</tr>
</tbody>
</table>
with $`\sigma_{G}`$ as defined in ([16](#S3.E16 "In 3.1 Second order statistics ‣ 3 Statistics of the Gabor coefficients ‣ The entropy of keys derived from laser speckle")). We have again used shorthand notation $`G`$ for $`G\hspace{0pt}{(s,\overset{\rightarrow}{k},\overset{\rightarrow}{x})}`$. The averages in ([52](#S4.E52 "In 4.4 Effect of perturbations on the Gabor coefficients ‣ 4 Perturbing speckle patterns ‣ The entropy of keys derived from laser speckle")) are easily evaluated. First we note that $`\left\langle G \right\rangle_{\varphi} = \left\langle \hat{G} \right\rangle_{\varphi} = 0`$. Second, we factorize $`\left\langle {G\hspace{0pt}{\hat{G}}'} \right\rangle_{\varepsilon,\varphi} = \left\langle {G\hspace{0pt}\left\langle {\hat{G}}' \right\rangle_{\varepsilon}} \right\rangle_{\varphi}`$. It directly follws from the definition of $`G`$ ([1](#S1.E1 "In 1.2 Binarized Gabor coefficients ‣ 1 Introduction ‣ The entropy of keys derived from laser speckle")) and the perturbation- averaged intensity ([48](#S4.E48 "In 4.3 Effect of perturbations on the intensity ‣ 4 Perturbing speckle patterns ‣ The entropy of keys derived from laser speckle")) that
<table>
<tbody>
<tr>
<td></td>
<td>⟨G^′⟩ε=QG′.subscriptdelimited-⟨⟩superscript^𝐺′𝜀𝑄superscript𝐺′\left\langle\hat{G}^{\prime}\right\rangle_{\varepsilon}=QG^{\prime}.</td>
<td></td>
<td>(53)</td>
</tr>
</tbody>
</table>
Thus we obtain
<table>
<tbody>
<tr>
<td></td>
<td>C(G,G^′)=Q⋅CG(w,w′,k→,k→′,x→,x→′),𝐶𝐺superscript^𝐺′⋅𝑄subscript𝐶𝐺𝑤superscript𝑤′→𝑘superscript→𝑘′→𝑥superscript→𝑥′C(G,\hat{G}^{\prime})=Q\cdot C_{G}(w,w^{\prime},\vec{k},\vec{k}^{\prime},\vec{x},\vec{x}^{\prime}),</td>
<td></td>
<td>(54)</td>
</tr>
</tbody>
</table>
with $`C_{G}`$ as defined in ([17](#S3.E17 "In 3.1 Second order statistics ‣ 3 Statistics of the Gabor coefficients ‣ The entropy of keys derived from laser speckle")). Hence, as in the case of intensity correlations, the correlation between the unperturbed $`G`$ at location $`\overset{\rightarrow}{x}`$ and the perturbed $`{\hat{G}}'`$ at $`{\overset{\rightarrow}{x}}'`$ factorizes into a contribution from the perturbation and the ordinary correlation function $`C_{G}`$.
#### 4.4.1 Mutual information between perturbed and unperturbed Gabor coefficients
We estimate the mutual information between an $`n`$-dimensional vector $`G`$ of unperturbed Gabor coefficients and the vector $`\hat{G}`$ of corresponding perturbed coefficients. The vector consists of the same set as in Section [3.3](#S3.SS3 "3.3 Estimated entropy of a set of Gabor coefficients for given detector noise ‣ 3 Statistics of the Gabor coefficients ‣ The entropy of keys derived from laser speckle"), i.e. a single width $`w`$, two perpendicular wave vectors $`{\overset{\rightarrow}{k}}_{1}`$, $`{\overset{\rightarrow}{k}}_{2}`$ of equal length, and a grid of points $`\overset{\rightarrow}{x}`$. We approximate the joint probability distribution of $`G`$ and $`\hat{G}`$ by a Gaussian distribution. This is motivated by the results of Section [3.2](#S3.SS2 "3.2 Computation of higher moments ‣ 3 Statistics of the Gabor coefficients ‣ The entropy of keys derived from laser speckle").
We make use of a well known result from information theory. If $`X`$ is an $`n`$-component Gaussian-distributed vector with covariance matrix $`\Sigma_{X}`$, then the differential entropy \[[2](#bib.bib2)\] of $`X`$ is given by
<table>
<tbody>
<tr>
<td></td>
<td>h(X)=12log2[(2πe)n|detΣX|].ℎ𝑋12subscript2superscript2𝜋𝑒𝑛subscriptΣ𝑋h(X)=\mbox{$\textstyle\frac{1}{2}$}\log_{2}[(2\pi e)^{n}|\det\Sigma_{X}|].</td>
<td></td>
<td>(55)</td>
</tr>
</tbody>
</table>
Let $`Y`$ be a second vector of the same length, with covariance matrix $`\Sigma_{Y}`$, and let $`\Sigma_{X\hspace{0pt}Y}`$ be the covariance of $`X`$ and $`Y`$. Let $`\Sigma_{X}`$ and $`\Sigma_{X\hspace{0pt}Y}`$ commute. Then the mutual information between $`X`$ and $`Y`$ follows from the definition ([45](#S4.E45 "In 4.2 Mutual information in the source plane ‣ 4 Perturbing speckle patterns ‣ The entropy of keys derived from laser speckle")) and ([55](#S4.E55 "In 4.4.1 Mutual information between perturbed and unperturbed Gabor coefficients ‣ 4.4 Effect of perturbations on the Gabor coefficients ‣ 4 Perturbing speckle patterns ‣ The entropy of keys derived from laser speckle")),
<table>
<tbody>
<tr>
<td></td>
<td>𝐈(X;Y)=−12log2|det(1−ΣY−1ΣX−1ΣXY2)|.𝐈𝑋𝑌12subscript21superscriptsubscriptΣ𝑌1superscriptsubscriptΣ𝑋1superscriptsubscriptΣ𝑋𝑌2{\bf I}(X;Y)=-\mbox{$\textstyle\frac{1}{2}$}\log_{2}\left|\det(1-\Sigma_{Y}^{-1}\Sigma_{X}^{-1}\Sigma_{XY}^{2})\vphantom{M^{M}}\right|.</td>
<td></td>
<td>(56)</td>
</tr>
</tbody>
</table>
Note that in the special case $`Y = {X + N}`$, where $`N`$ is Gaussian noise uncorrelated to $`X`$, ([56](#S4.E56 "In 4.4.1 Mutual information between perturbed and unperturbed Gabor coefficients ‣ 4.4 Effect of perturbations on the Gabor coefficients ‣ 4 Perturbing speckle patterns ‣ The entropy of keys derived from laser speckle")) reduces to the form ([23](#S3.E23 "In 3.3 Estimated entropy of a set of Gabor coefficients for given detector noise ‣ 3 Statistics of the Gabor coefficients ‣ The entropy of keys derived from laser speckle")).
Now we consider the case where $`X = G`$ and $`Y = {\hat{G} + {\delta\hspace{0pt}G}}`$, where $`\delta\hspace{0pt}G`$ again is the Gaussian detector noise discussed in Section [3.3](#S3.SS3 "3.3 Estimated entropy of a set of Gabor coefficients for given detector noise ‣ 3 Statistics of the Gabor coefficients ‣ The entropy of keys derived from laser speckle"). Note that $`\hat{G}`$ has the same $`\varphi`$-ensemble statistics as $`G`$. Thus we have
<table>
<tbody>
<tr>
<td></td>
<td>ΣX→ΣG,→subscriptΣ𝑋subscriptΣ𝐺\displaystyle\Sigma_{X}\rightarrow\Sigma_{G},</td>
<td>ΣY→ΣG+ΣN,→subscriptΣ𝑌subscriptΣ𝐺subscriptΣ𝑁\displaystyle\Sigma_{Y}\rightarrow\Sigma_{G}+\Sigma_{N},</td>
<td>ΣXY→QΣG.→subscriptΣ𝑋𝑌𝑄subscriptΣ𝐺\displaystyle\Sigma_{XY}\rightarrow Q\Sigma_{G}.</td>
<td></td>
<td>(57)</td>
</tr>
</tbody>
</table>
Substitution into ([56](#S4.E56 "In 4.4.1 Mutual information between perturbed and unperturbed Gabor coefficients ‣ 4.4 Effect of perturbations on the Gabor coefficients ‣ 4 Perturbing speckle patterns ‣ The entropy of keys derived from laser speckle")) yields the mutual information
<table>
<tbody>
<tr>
<td></td>
<td>𝐈(G;G^+δG)=12log2|det(1+ΣN−1ΣG)|−12log2|det(1+[1−Q2]ΣN−1ΣG)|.𝐈𝐺^𝐺𝛿𝐺12subscript21superscriptsubscriptΣ𝑁1subscriptΣ𝐺12subscript21delimited-[]1superscript𝑄2superscriptsubscriptΣ𝑁1subscriptΣ𝐺{\bf I}(G;\hat{G}+\delta G)=\mbox{$\textstyle\frac{1}{2}$}\log_{2}\left|\det\left(1+\Sigma_{N}^{-1}\Sigma_{G}\right)\right|-\mbox{$\textstyle\frac{1}{2}$}\log_{2}\left|\det\left(1+[1-Q^{2}]\Sigma_{N}^{-1}\Sigma_{G}\right)\right|.</td>
<td></td>
<td>(58)</td>
</tr>
</tbody>
</table>
This represents an absolute upper bound on the information that can be reproducibly extracted from a speckle pattern, given that there is noise $`\delta\hspace{0pt}I`$ in the detector and perturbation noise $`\{\varepsilon_{\overset{\rightarrow}{a}}\}`$ in the source. Clearly, ([58](#S4.E58 "In 4.4.1 Mutual information between perturbed and unperturbed Gabor coefficients ‣ 4.4 Effect of perturbations on the Gabor coefficients ‣ 4 Perturbing speckle patterns ‣ The entropy of keys derived from laser speckle")) reduces to ([23](#S3.E23 "In 3.3 Estimated entropy of a set of Gabor coefficients for given detector noise ‣ 3 Statistics of the Gabor coefficients ‣ The entropy of keys derived from laser speckle")) in the case $`q \downarrow 0`$, and the mutual information goes to zero in the case $`q\rightarrow\pi`$.
We estimate ([58](#S4.E58 "In 4.4.1 Mutual information between perturbed and unperturbed Gabor coefficients ‣ 4.4 Effect of perturbations on the Gabor coefficients ‣ 4 Perturbing speckle patterns ‣ The entropy of keys derived from laser speckle")) using the approximation method given in ([32](#S3.E32 "In 3.3 Estimated entropy of a set of Gabor coefficients for given detector noise ‣ 3 Statistics of the Gabor coefficients ‣ The entropy of keys derived from laser speckle")). This yields, for one direction of the $`\overset{\rightarrow}{k}`$-vector (and expressed in natural units instead of bits),
<table>
<tbody>
<tr>
<td></td>
<td>𝐈(G[j];G^[j]+δG[j])𝐈superscript𝐺delimited-[]𝑗superscript^𝐺delimited-[]𝑗𝛿superscript𝐺delimited-[]𝑗\displaystyle{\bf I}(G^{[j]};\hat{G}^{[j]}+\delta G^{[j]})</td>
<td>≈\displaystyle\approx</td>
<td>πL22ℓ2(1−ℓ2L2)ln11−Q2𝜋superscript𝐿22superscriptℓ21superscriptℓ2superscript𝐿211superscript𝑄2\displaystyle\frac{\pi L^{2}}{2\ell^{2}}(1-\frac{\ell^{2}}{L^{2}})\ln\frac{1}{1-Q^{2}}</td>
<td></td>
<td rowspan="2">(59)</td>
</tr>
<tr>
<td></td>
<td></td>
<td>+\displaystyle+</td>
<td>L22πc2[−Dilog−ec2p2c1+Dilog−ec2p2(1−Q2)c1]p=π/Lπ/ℓ,superscript𝐿22𝜋subscript𝑐2superscriptsubscriptdelimited-[]Dilogsuperscript𝑒subscript𝑐2superscript𝑝2subscript𝑐1Dilogsuperscript𝑒subscript𝑐2superscript𝑝21superscript𝑄2subscript𝑐1𝑝𝜋𝐿𝜋ℓ\displaystyle\frac{L^{2}}{2\pi c_{2}}\left[-{\rm Dilog}\frac{-e^{c_{2}p^{2}}}{c_{1}}+{\rm Dilog}\frac{-e^{c_{2}p^{2}}}{(1-Q^{2})c_{1}}\right]_{p=\pi/L}^{\pi/\ell},</td>
<td></td>
</tr>
</tbody>
</table>
where the constants $`c_{1}`$ and $`c_{2}`$ are defined in ([31](#S3.E31 "In 3.3 Estimated entropy of a set of Gabor coefficients for given detector noise ‣ 3 Statistics of the Gabor coefficients ‣ The entropy of keys derived from laser speckle")). The result ([59](#S4.E59 "In 4.4.1 Mutual information between perturbed and unperturbed Gabor coefficients ‣ 4.4 Effect of perturbations on the Gabor coefficients ‣ 4 Perturbing speckle patterns ‣ The entropy of keys derived from laser speckle")) is plotted in Fig. [3](#S4.F3 "Figure 3 ‣ 4.4.1 Mutual information between perturbed and unperturbed Gabor coefficients ‣ 4.4 Effect of perturbations on the Gabor coefficients ‣ 4 Perturbing speckle patterns ‣ The entropy of keys derived from laser speckle"); As in Fig. [1](#S3.F1 "Figure 1 ‣ Small SNR ‣ 3.3 Estimated entropy of a set of Gabor coefficients for given detector noise ‣ 3 Statistics of the Gabor coefficients ‣ The entropy of keys derived from laser speckle"), the factor $`L^{2}`$ was replaced by $`\pi\hspace{0pt}M^{2}`$ to obtain a result per average speckle area. The mutual information decreases sharply as a function of the perturbation strength $`q`$.
[Refer to caption](/html/0710.5002/assets/x3.png)
Figure 3: Mutual information, per average speckle area, between a noiseless vector $`G^{\lbrack j\rbrack}`$ of Gabor coefficients and a noisy & perturbed version $`{\hat{G}}^{\lbrack j\rbrack} + {\delta\hspace{0pt}G^{\lbrack j\rbrack}}`$, plotted as a function of the perturbation strength $`q`$. The parameters are: $`L = 800`$ pixels, $`M = 3`$ pixels, $`\ell = 5`$ pixels, $`t = 1`$ pixel <sup>2</sup>, $`{I_{av}/N_{I}} = 10`$. The radius of an average speckle has been set, somewhat arbitrarily, to $`M`$.
#### 4.4.2 Bit error probability
We estimate the probability of a bit error in a binarized Gabor coefficient due to the random perturbation. Here we will neglect the detector noise. We use the shorthand notation $`G = {G\hspace{0pt}{(w,\overset{\rightarrow}{k},\overset{\rightarrow}{x})}}`$ and $`\hat{G} = {\hat{G}\hspace{0pt}{(w,\overset{\rightarrow}{k},\overset{\rightarrow}{x})}}`$. As in the previous section, we approximate the joint probability distribution $`\rho`$ of $`G`$ and $`\hat{G}`$ by a Gaussian, making use of the correlation ([54](#S4.E54 "In 4.4 Effect of perturbations on the Gabor coefficients ‣ 4 Perturbing speckle patterns ‣ The entropy of keys derived from laser speckle")),
<table>
<tbody>
<tr>
<td></td>
<td>ρ(G,G^)𝜌𝐺^𝐺\displaystyle\rho(G,\hat{G})</td>
<td>=\displaystyle=</td>
<td>12πdetAexp−12(G,G^)A−1(GG^)12𝜋𝐴12𝐺^𝐺superscript𝐴1matrix𝐺^𝐺\displaystyle\frac{1}{2\pi\sqrt{\det A}}\exp-\mbox{$\textstyle\frac{1}{2}$}(G,\hat{G})A^{-1}\left(\matrix{G\cr\hat{G}}\right)</td>
<td></td>
<td>(60)</td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td>A𝐴\displaystyle A</td>
<td>=\displaystyle=</td>
<td>σG2(1QQ1).superscriptsubscript𝜎𝐺2matrix1𝑄𝑄1\displaystyle\sigma_{G}^{2}\left(\matrix{1&Q\cr Q&1}\right).</td>
<td></td>
<td></td>
</tr>
</tbody>
</table>
A bit error occurs when the signs of $`G`$ and $`\hat{G}`$ are not equal, while $`{|G|} > T`$ (see Section [1.2](#S1.SS2 "1.2 Binarized Gabor coefficients ‣ 1 Introduction ‣ The entropy of keys derived from laser speckle")). The probability of this event is given by the following integral expression
<table>
<tbody>
<tr>
<td></td>
<td>Perr=Prob[G^<0|G>T]=Prob[G^<0,G>T]Prob[G>T]=∫−∞0dG^∫T∞dGρ∫−∞∞dG^∫T∞dGρ.subscript𝑃errProbdelimited-[]^𝐺inner-product0𝐺𝑇Probdelimited-[]formulae-sequence^𝐺0𝐺𝑇Probdelimited-[]𝐺𝑇superscriptsubscript0differential-d^𝐺superscriptsubscript𝑇differential-d𝐺𝜌superscriptsubscriptdifferential-d^𝐺superscriptsubscript𝑇differential-d𝐺𝜌P_{\rm err}={\rm Prob}[\hat{G}<0|G>T]=\frac{{\rm Prob}[\hat{G}<0,G>T]}{{\rm Prob}[G>T]}=\frac{\int_{-\infty}^{0}\!\!{\rm d}\hat{G}\;\int_{T}^{\infty}\!\!{\rm d}G\;\rho}{\int_{-\infty}^{\infty}\!\!{\rm d}\hat{G}\;\int_{T}^{\infty}\!\!{\rm d}G\;\rho}.</td>
<td></td>
<td>(61)</td>
</tr>
</tbody>
</table>
Evaluation of several of the integrals gives
<table>
<tbody>
<tr>
<td></td>
<td>Perrsubscript𝑃err\displaystyle P_{\rm err}</td>
<td>=\displaystyle=</td>
<td>∫T∞dG1σG2πe−12G2/σG2⋅12ErfcQG1−Q2σG212ErfcTσG2,superscriptsubscript𝑇⋅differential-d𝐺1subscript𝜎𝐺2𝜋superscript𝑒12superscript𝐺2superscriptsubscript𝜎𝐺212Erfc𝑄𝐺1superscript𝑄2subscript𝜎𝐺212Erfc𝑇subscript𝜎𝐺2\displaystyle\frac{\int_{T}^{\infty}\!\!{\rm d}G\;\frac{1}{\sigma_{G}\sqrt{2\pi}}e^{-\mbox{$\textstyle\frac{1}{2}$}G^{2}/\sigma_{G}^{2}}\cdot\mbox{$\textstyle\frac{1}{2}$}{\rm Erfc}\frac{QG}{\sqrt{1-Q^{2}}\sigma_{G}\sqrt{2}}}{\mbox{$\textstyle\frac{1}{2}$}{\rm Erfc}\frac{T}{\sigma_{G}\sqrt{2}}},</td>
<td></td>
<td>(62)</td>
</tr>
</tbody>
</table>
where $`Erfc`$ stands for the complementary error function. Fig. [4](#S4.F4 "Figure 4 ‣ 4.4.2 Bit error probability ‣ 4.4 Effect of perturbations on the Gabor coefficients ‣ 4 Perturbing speckle patterns ‣ The entropy of keys derived from laser speckle") shows the behaviour of $`P_{err}`$ as a function of $`q`$ and $`T`$. Exact evaluation of the leftover integral in ([62](#S4.E62 "In 4.4.2 Bit error probability ‣ 4.4 Effect of perturbations on the Gabor coefficients ‣ 4 Perturbing speckle patterns ‣ The entropy of keys derived from laser speckle")) is difficult in general. However, in some limiting cases analytic results can be obtained. For instance, for $`T = 0`$ the result is $`P_{err} = {\pi^{- 1}\hspace{0pt}{\arccos Q}}`$. Furthermore, in the two limiting cases $`\frac{Q}{\sqrt{1 - Q^{2}}} \gg \frac{\sigma_{G}\hspace{0pt}\sqrt{2}}{T}`$ and $`\frac{Q}{\sqrt{1 - Q^{2}}} \ll \frac{\sigma_{G}\hspace{0pt}\sqrt{2}}{T}`$ approximations can be obtained. In the former case we apply a large-argument asymptotic expansion of the $`Erfc`$ function; in the latter case a Taylor expansion. For the first case, let us define the small parameter $`\varepsilon \ll 1`$,
<table>
<tbody>
<tr>
<td></td>
<td>ε:=σG2T1−Q2Q.assign𝜀subscript𝜎𝐺2𝑇1superscript𝑄2𝑄\varepsilon:=\frac{\sigma_{G}\sqrt{2}}{T}\frac{\sqrt{1-Q^{2}}}{Q}.</td>
<td></td>
<td>(63)</td>
</tr>
</tbody>
</table>
After some straightforward but tedious algebra we obtain
<table>
<tbody>
<tr>
<td></td>
<td>Perr=e−1/ε2e−12T2/σG2Erfc[T/(σG2)]T2πσG2{ε3−ε5(3/2+12T2/σG2)+𝒪(ε7)}.subscript𝑃errsuperscript𝑒1superscript𝜀2superscript𝑒12superscript𝑇2superscriptsubscript𝜎𝐺2Erfcdelimited-[]𝑇subscript𝜎𝐺2𝑇2𝜋subscript𝜎𝐺2superscript𝜀3superscript𝜀53212superscript𝑇2superscriptsubscript𝜎𝐺2𝒪superscript𝜀7P_{\rm err}=\frac{e^{-1/\varepsilon^{2}}e^{-\mbox{$\textstyle\frac{1}{2}$}T^{2}/\sigma_{G}^{2}}}{{\rm Erfc}[T/(\sigma_{G}\sqrt{2})]}\frac{T}{2\pi\sigma_{G}\sqrt{2}}\left\{\varepsilon^{3}-\varepsilon^{5}(3/2+\mbox{$\textstyle\frac{1}{2}$}T^{2}/\sigma_{G}^{2})+{\cal O}(\varepsilon^{7})\right\}.</td>
<td></td>
<td>(64)</td>
</tr>
</tbody>
</table>
Expression ([64](#S4.E64 "In 4.4.2 Bit error probability ‣ 4.4 Effect of perturbations on the Gabor coefficients ‣ 4 Perturbing speckle patterns ‣ The entropy of keys derived from laser speckle")) is useful in the weak perturbation limit $`Q\rightarrow 1`$ and in the limit of large thresholds, $`{T/\sigma_{G}} \gg 1`$.
For the second case we write $`\eta:={1/\varepsilon} \ll 1`$. A Taylor expansion of the $`Erfc`$ function in ([62](#S4.E62 "In 4.4.2 Bit error probability ‣ 4.4 Effect of perturbations on the Gabor coefficients ‣ 4 Perturbing speckle patterns ‣ The entropy of keys derived from laser speckle")) yields, after some algebra,
<table>
<tbody>
<tr>
<td></td>
<td>Perr=12[1−(η−η33π[1+2σGT])σG2Te−12T2/σG212Erfc[T/(σG2)]+𝒪(η5)].subscript𝑃err12delimited-[]1𝜂superscript𝜂33𝜋delimited-[]12subscript𝜎𝐺𝑇subscript𝜎𝐺2𝑇superscript𝑒12superscript𝑇2superscriptsubscript𝜎𝐺212Erfcdelimited-[]𝑇subscript𝜎𝐺2𝒪superscript𝜂5P_{\rm err}=\frac{1}{2}\left[1-(\eta-\frac{\eta^{3}}{3\pi}[1+\frac{2\sigma_{G}}{T}])\frac{\sigma_{G}\sqrt{2}}{T}\frac{e^{-\mbox{$\textstyle\frac{1}{2}$}T^{2}/\sigma_{G}^{2}}}{\mbox{$\textstyle\frac{1}{2}$}{\rm Erfc}[T/(\sigma_{G}\sqrt{2})]}+{\cal O}(\eta^{5})\right].</td>
<td></td>
<td>(65)</td>
</tr>
</tbody>
</table>
Expression ([65](#S4.E65 "In 4.4.2 Bit error probability ‣ 4.4 Effect of perturbations on the Gabor coefficients ‣ 4 Perturbing speckle patterns ‣ The entropy of keys derived from laser speckle")) is useful in the strong perturbation limit $`Q\rightarrow 0`$ and in the limit of small thresholds, $`{T/\sigma_{G}} \ll 1`$.
[Refer to caption](/html/0710.5002/assets/x4.png)
Figure 4: Bit error probability $`P_{err}`$ according to ([62](#S4.E62 "In 4.4.2 Bit error probability ‣ 4.4 Effect of perturbations on the Gabor coefficients ‣ 4 Perturbing speckle patterns ‣ The entropy of keys derived from laser speckle")) as a function of the perturbation strength $`q`$ for various values of the threshold $`T`$.
## 5 Comparison to experimental results
In this section we briefly compare a number of theoretical results, obtained in the previous chapters, to actual experiments. The experimental data were obtained with a very simple setup, consisting of a laser, a sample holder and a detector. The laser has a wavelength of $`\lambda =`$780nm. A parallel beam shines on the sample at an angle of 45<sup>∘</sup>. The spot is circular, with a diameter $`D =`$1mm. The sample is a piece of paper. The detector is a CCD camera, mounted at a normal angle to the sample. The distance between the sample and the camera is $`z =`$10cm. The camera has a pixel pitch of 6.25$`\mu`$m and takes 1024$`\times`$768 pixel images with 256 gray scales. The typical speckle diameter at the location of the camera is of order $`{{\lambda\hspace{0pt}z}/D} =`$78$`\mu`$m, corresponding to 12 pixels in the image. The setup is not particularly well protected against background light.
#### Intensity distribution
In order to illustrate the quality of our data, we show in Fig. [5](#S5.F5 "Figure 5 ‣ Intensity distribution ‣ 5 Comparison to experimental results ‣ The entropy of keys derived from laser speckle") a histogram of the gray levels present in a single typical CCD image. The lowest gray scale present in the image was normalised to zero. The deviations from the theoretical curve ([12](#S2.E12 "In 2.3 Statistics of the intensity ‣ 2 The random phase model ‣ The entropy of keys derived from laser speckle")) at low intensity show that there is a noticeable effect of the background light.
[Refer to caption](/html/0710.5002/assets/x5.png)
Figure 5: Intensity histogram of a single speckle pattern. The histogram was made with a bin width of 5 gray values. The theoretical curve is the exponential distribution $`{({1/I_{av}})}\hspace{0pt}{\exp{({- {I/I_{av}}})}}`$.
#### Statistical distribution of the Gabor coefficients
Here we discuss the experimental verification of the theoretical results of Sections [3.1](#S3.SS1 "3.1 Second order statistics ‣ 3 Statistics of the Gabor coefficients ‣ The entropy of keys derived from laser speckle") and [3.2](#S3.SS2 "3.2 Computation of higher moments ‣ 3 Statistics of the Gabor coefficients ‣ The entropy of keys derived from laser speckle"). First we show that the theoretical prediction ([16](#S3.E16 "In 3.1 Second order statistics ‣ 3 Statistics of the Gabor coefficients ‣ The entropy of keys derived from laser speckle")) for $`\sigma_{G}`$ is accurate. Then we show that the distribution function of the Gabor coefficients has a noticeable deviation from the Gaussian form, with fatter tails than a Gaussian.
The empirical $`\varphi`$-ensemble probability distribution of $`G\hspace{0pt}{(w,\overset{\rightarrow}{k},{\overset{\rightarrow}{x}}_{0})}`$ should ideally be obtained as follows: Insert many samples; for each sample, measure $`G\hspace{0pt}{(w,\overset{\rightarrow}{k},{\overset{\rightarrow}{x}}_{0})}`$; finally make a histogram of all the Gabor coefficients; this yields the empirical distribution function for $`G\hspace{0pt}{(w,\overset{\rightarrow}{k},{\overset{\rightarrow}{x}}_{0})}`$.
We used a less labour-intensive approach. We took a single sample; from the single CCD image we derived Gabor coefficients $`G\hspace{0pt}{(w,\overset{\rightarrow}{k},{\overset{\rightarrow}{x}}_{0})}`$ for all $`{\overset{\rightarrow}{x}}_{0}`$; we made a histogram of the resulting set; we used this as the empirical distribution of $`G\hspace{0pt}{(w,\overset{\rightarrow}{k},{\overset{\rightarrow}{x}}_{0})}`$. This approach is motivated by (a) the fact that the $`\varphi`$-ensemble probability distribution does not depend on $`{\overset{\rightarrow}{x}}_{0}`$ and (b) the ergodicity property of laser speckle, i.e. the property that the spatial intensity distribution asymptotically tends to the $`\varphi`$-ensemble intensity distribution.
The result is shown in Fig. [6](#S5.F6 "Figure 6 ‣ Statistical distribution of the Gabor coefficients ‣ 5 Comparison to experimental results ‣ The entropy of keys derived from laser speckle"), where $`\sigma_{G}`$ is given as a function of $`k = {|\overset{\rightarrow}{k}|}`$ for a number of choices for $`w`$. The theoretical result ([16](#S3.E16 "In 3.1 Second order statistics ‣ 3 Statistics of the Gabor coefficients ‣ The entropy of keys derived from laser speckle")) is also plotted. The correspondence of theory vs. experiment is very good, except at large $`k`$. There we start to see the difference between the spatial continuum appraoch of the theory and the discrete pixellated nature of the CCD images. The theory uses spatial integration, while the data processing involves summation over pixels. For fast oscillations, the integration in ([1](#S1.E1 "In 1.2 Binarized Gabor coefficients ‣ 1 Introduction ‣ The entropy of keys derived from laser speckle")) averages out to zero more quickly than the summation.
Fig. [7](#S5.F7 "Figure 7 ‣ Statistical distribution of the Gabor coefficients ‣ 5 Comparison to experimental results ‣ The entropy of keys derived from laser speckle") shows the shape of the empirical distribution function of the Gabor coefficients. The curves were derived from a single image. It can be seen that the distribution has fatter tails than a Gaussian, as was derived in Section [3.2](#S3.SS2 "3.2 Computation of higher moments ‣ 3 Statistics of the Gabor coefficients ‣ The entropy of keys derived from laser speckle").
[Refer to caption](/html/0710.5002/assets/x6.png)
Figure 6: The variance $`\sigma_{G}`$ of the Gabor coefficients (normalized w.r.t. $`I_{av}`$) as a function of the spatial frequency parameter $`k`$, for several values of the Gaussian width $`w`$. (Here $`w`$ is measured in pixels, and $`k`$ in radians/pixel.) The solid curves represent the experimental data. The dotted curves are the theoretical result ([16](#S3.E16 "In 3.1 Second order statistics ‣ 3 Statistics of the Gabor coefficients ‣ The entropy of keys derived from laser speckle")), with the parameter choice $`M = 5`$ pixels.
[Refer to caption](/html/0710.5002/assets/x7.png)
[Refer to caption](/html/0710.5002/assets/x8.png)
Figure 7: Distribution of Gabor coefficients obtained from a single speckle pattern, by applying two perpendicular angle parameters. Left: $`\varphi = 0`$; Right: $`\varphi = 90^{\circ}`$. The solid curve depicts a histogram of the Gabor coefficients. The dotted curve is the least-squares Gaussian fit to that histogram.
#### Effect of perturbations
We investigated perturbations as follows. Our sample was a piece of paper whose surface structure changed over time. We took 40 pictures at half hour intervals. The changing surface structure can be regarded as a random perturbation as modelled in Section [4.1](#S4.SS1 "4.1 Phase perturbations ‣ 4 Perturbing speckle patterns ‣ The entropy of keys derived from laser speckle"). Unfortunately, it is not possible to experimentally regulate the perturbation strength $`q`$. We therefore used the following approach to compare theory and experiment. We looked at all pairs of CCD images; there are $`\binom{40}{2} = 780`$ pairs. For each pair $`(A,B)`$ we computed the empirical intensity correlation $`\Xi_{I}`$ and Gabor coefficient correlation $`\Xi_{G}`$,
<table>
<tbody>
<tr>
<td></td>
<td>ΞI[A,B]subscriptΞ𝐼𝐴𝐵\displaystyle\Xi_{I}[A,B]</td>
<td>=\displaystyle=</td>
<td>Npix−1∑iIiAIiB−(Npix−1∑iIiA)(Npix−1∑iIiB)Npix−1∑i[IiA]2−[Npix−1∑iIiA]2Npix−1∑j[IjB]2−[Npix−1∑jIjB]2superscriptsubscript𝑁pix1subscript𝑖subscriptsuperscript𝐼𝐴𝑖subscriptsuperscript𝐼𝐵𝑖superscriptsubscript𝑁pix1subscript𝑖subscriptsuperscript𝐼𝐴𝑖superscriptsubscript𝑁pix1subscript𝑖subscriptsuperscript𝐼𝐵𝑖superscriptsubscript𝑁pix1subscript𝑖superscriptdelimited-[]subscriptsuperscript𝐼𝐴𝑖2superscriptdelimited-[]superscriptsubscript𝑁pix1subscript𝑖subscriptsuperscript𝐼𝐴𝑖2superscriptsubscript𝑁pix1subscript𝑗superscriptdelimited-[]subscriptsuperscript𝐼𝐵𝑗2superscriptdelimited-[]superscriptsubscript𝑁pix1subscript𝑗subscriptsuperscript𝐼𝐵𝑗2\displaystyle\frac{N_{\rm pix}^{-1}\sum_{i}I^{A}_{i}I^{B}_{i}-(N_{\rm pix}^{-1}\sum_{i}I^{A}_{i})(N_{\rm pix}^{-1}\sum_{i}I^{B}_{i})}{\sqrt{N_{\rm pix}^{-1}\sum_{i}[I^{A}_{i}]^{2}-[N_{\rm pix}^{-1}\sum_{i}I^{A}_{i}]^{2}}\sqrt{N_{\rm pix}^{-1}\sum_{j}[I^{B}_{j}]^{2}-[N_{\rm pix}^{-1}\sum_{j}I^{B}_{j}]^{2}}}</td>
<td></td>
<td>(66)</td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td>ΞG[A,B]subscriptΞ𝐺𝐴𝐵\displaystyle\Xi_{G}[A,B]</td>
<td>=\displaystyle=</td>
<td>Npix−1∑iGiAGiB−(Npix−1∑iGiA)(Npix−1∑iGiB)Npix−1∑i[GiA]2−[Npix−1∑iGiA]2Npix−1∑j[GjB]2−[Npix−1∑jGjB]2.superscriptsubscript𝑁pix1subscript𝑖subscriptsuperscript𝐺𝐴𝑖subscriptsuperscript𝐺𝐵𝑖superscriptsubscript𝑁pix1subscript𝑖subscriptsuperscript𝐺𝐴𝑖superscriptsubscript𝑁pix1subscript𝑖subscriptsuperscript𝐺𝐵𝑖superscriptsubscript𝑁pix1subscript𝑖superscriptdelimited-[]subscriptsuperscript𝐺𝐴𝑖2superscriptdelimited-[]superscriptsubscript𝑁pix1subscript𝑖subscriptsuperscript𝐺𝐴𝑖2superscriptsubscript𝑁pix1subscript𝑗superscriptdelimited-[]subscriptsuperscript𝐺𝐵𝑗2superscriptdelimited-[]superscriptsubscript𝑁pix1subscript𝑗subscriptsuperscript𝐺𝐵𝑗2\displaystyle\frac{N_{\rm pix}^{-1}\sum_{i}G^{A}_{i}G^{B}_{i}-(N_{\rm pix}^{-1}\sum_{i}G^{A}_{i})(N_{\rm pix}^{-1}\sum_{i}G^{B}_{i})}{\sqrt{N_{\rm pix}^{-1}\sum_{i}[G^{A}_{i}]^{2}-[N_{\rm pix}^{-1}\sum_{i}G^{A}_{i}]^{2}}\sqrt{N_{\rm pix}^{-1}\sum_{j}[G^{B}_{j}]^{2}-[N_{\rm pix}^{-1}\sum_{j}G^{B}_{j}]^{2}}}.</td>
<td></td>
<td>(67)</td>
</tr>
</tbody>
</table>
Here $`N_{pix}`$ is the number of pixels in the image, $`\sum_{i}`$ stands for summation over all pixels, $`I_{i}^{A}`$ denotes the intensity in image $`A`$ at location $`{\overset{\rightarrow}{x}}_{i}`$, and $`G_{i}^{A}`$ is shorthand notation for $`G\hspace{0pt}{(w,\overset{\rightarrow}{k},{\overset{\rightarrow}{x}}_{i})}`$ for some fixed value of $`w`$ and $`\overset{\rightarrow}{k}`$. The empirical correlation $`\Xi_{I}`$ should be equivalent to the theoretical correlation ([49](#S4.E49 "In 4.3 Effect of perturbations on the intensity ‣ 4 Perturbing speckle patterns ‣ The entropy of keys derived from laser speckle")) with $`{\overset{\rightarrow}{x}}' = \overset{\rightarrow}{x}`$; Similarly, $`\Xi_{G}`$ should be equivalent to ([54](#S4.E54 "In 4.4 Effect of perturbations on the Gabor coefficients ‣ 4 Perturbing speckle patterns ‣ The entropy of keys derived from laser speckle")) with the substitution $`w' = w`$, $`{\overset{\rightarrow}{k}}' = \overset{\rightarrow}{k}`$ and $`{\overset{\rightarrow}{x}}' = \overset{\rightarrow}{x}`$. Hence, we expect $`{\Xi_{I}\hspace{0pt}{\lbrack A,B\rbrack}} = Q`$ and $`{\Xi_{G}\hspace{0pt}{\lbrack A,B\rbrack}} = Q`$. In Fig. [8](#S5.F8 "Figure 8 ‣ Effect of perturbations ‣ 5 Comparison to experimental results ‣ The entropy of keys derived from laser speckle") we have plotted $`\Xi_{G}`$ vs. $`\Xi_{I}`$ for all image pairs. The data points are clearly bunched together on a narrow band slightly above the theoretically expected line $`\Xi_{G} = \Xi_{I}`$. We hypothesize that this small difference is due to detector noise, which we did not take into account here. The Gabor coefficients, resulting from a spatial sum, are less sensitive to detector noise than the intensity itself. Hence the correlation $`\Xi_{G}`$ of the Gabor coefficients is larger than the intensity correlation $`\Xi_{I}`$.
[Refer to caption](/html/0710.5002/assets/x9.png)
Figure 8: Effect of a perturbation on the intensities and the Gabor coefficients. Horizontal axis: the intensity correlation function ([66](#S5.E66 "In Effect of perturbations ‣ 5 Comparison to experimental results ‣ The entropy of keys derived from laser speckle")). Vertical axis: The correlation function ([67](#S5.E67 "In Effect of perturbations ‣ 5 Comparison to experimental results ‣ The entropy of keys derived from laser speckle")) of the Gabor coefficients. Data from two perpendicular $`\overset{\rightarrow}{k}`$ vectors are plotted together. The solid line is the theoretical predicition for zero detector noise.
## 6 Summary
Laser speckle has been proposed in the security literature as a source of high-entropy bit strings for various (cryptographic) purposes. It is important to know how the physical properties of speckle affect the entropy of the extracted bit strings. More in particular, we need to know the mutual information between two repeated key extractions when noise is taken into account. Another important parameter is the bit error rate.
In this paper, we have developed a simple approach to address these issues. We have studied the case of key extraction using Gabor coefficients. We used a simple model for speckle, generated by a large number of independent random phases in a source plane. We have modeled perturbations of the speckle pattern as small, uniformly distributed perturbations of the random phases. Detector noise was modeled as being Gaussian, independent of the intensity and without correlations between the detector’s pixels.
Our main results are
- •
The Gabor coefficients have a distribution function that is close to Gaussian.
- •
We have derived an expression for the mutual information between an unperturbed and perturbed speckle source.
- •
We have obtained analytical expressions that give an upper bound on the mutual entropy of a set of Gabor coefficients (i) when there is detector noise but no perturbation and (ii) when there is detector noise as well as a perturbation.
- •
We have computed the bit error rate caused by perturbations of a speckle pattern.
Experimental data on the statistics of Gabor coefficients and on the correlation functions of Gabor coefficients and intensities are in accordance with theory.
The results of this paper, particularly the mutual information and error rate expressions, provide useful parameters for key extraction systems.
#### Acknowledgements
We thank Sjoerd Stallinga, Geert-Jan Schrijen, Wil Ophey, Pim Tuyls, Frans Willems and Tanya Ignatenko for useful comments.
## References
- \[1\] J. D. R Buchanan, R. P. Cowburn, A. Jausovec, D. Petit, P. Seem, G. Xiong, D. Atkinson, K. Fenton, D.A. Allwood, and M. T. Bryan. Forgery: ‘fingerprinting’ documents and packaging.
Nature, Brief Communications, 436:475, July 2005.
- \[2\] T. M. Cover and J. A. Thomas.
Elements of information theory.
Wiley Series in Telecommunications. Wiley & Sons, 1991.
- \[3\] J. Daugman. How iris recognition works.
IEEE Trans. Circuits and Systems for Video Technology, 14(1):21–30, 2004.
- \[4\] Y. Dodis, M. Reyzin, and A. Smith. Fuzzy extractors: How to generate strong keys from biometrics and other noisy data.
In C. Cachin and J. Camenisch, editors, Advances in Cryptology — Eurocrypt 2004, volume 3027 of LNCS, pages 523–540. Springer-Verlag, 2004.
- \[5\] S. Feng, C. Kane, P. A. Lee, and A. D. Stone. Correlations and fluctuations of coherent wave transmission through disordered media.
Phys.Rev.Lett., 61(7):834–837, 1988.
- \[6\] B. Gassend, D. E. Clarke, M. van Dijk, and S. Devadas. Controlled physical random functions.
In Annual Computer Security Applications Conference (ACSAC), pages 149–160, 2002.
- \[7\] J. W. Goodman.
Laser Speckle and Related Phenomena, chapter Statistical properties of laser speckle patterns.
Springer-Verlag, New York, 2nd edition, 1984.
- \[8\] T. Ignatenko, G.J. Schrijen, B. Škorić, P. Tuyls, and F. M. J. Willems. Estimating the secrecy rate of physical uncloneable functions with the context-tree weighting method.
In Proc. IEEE International Symposium on Information Theory 2006, pages 499–503, Seattle, USA, July 2006.
- \[9\] A. Juels and M. Wattenberg. A fuzzy commitment scheme.
In G. Tsudik, editor, Sixth ACM Conference on Computer and Communications Security, pages 28–36. ACM Press, 1999.
- \[10\] J.-P. M. G. Linnartz and P. Tuyls. New shielding functions to enhance privacy and prevent misuse of biometric templates.
In J. Kittler and M. Nixon, editors, Conference on Audio and Video Based Person Authentication, volume 2688 of LNCS, pages 238–250. Springer-Verlag, 2003.
- \[11\] R. Pappu.
Physical One-Way Functions.
PhD thesis, MIT, 2001.
- \[12\] R. Pappu, B. Recht, J. Taylor, and N. Gershenfeld. Physical one-way functions.
Science, 297:2026–2030, Sept. 2002.
- \[13\] P. Tuyls and B. Škorić.
AmIware: Hardware Technology Drivers of Ambient Intelligence, chapter Secret Key Generation from Classical Physics, pages 421–447.
Springer, 2006.
- \[14\] P. Tuyls and B. Škorić.
Security, Privacy and Trust in Modern Data Management, chapter Strong Authentication with Physical Unclonable Functions, pages 133–148.
Springer, 2007.
- \[15\] P. Tuyls, B. Škorić, and T. Kevenaar.
Security with Noisy Data: Private Biometrics, Secure Key Storage and Anti-Counterfeiting.
Springer, 2007.
- \[16\] P. Tuyls, B. Škorić, S. Stallinga, A.H.M. Akkermans, and W. Ophey. Information Theoretical Security Analysis of Physical Unclonable Functions.
In A.S. Patrick and M. Yung, editors, Financial Cryptography and Data Security — FC 2005, volume 3570 of LNCS, pages 141–155. Springer-Verlag, 2005.
- \[17\] B. Škorić, P. Tuyls, and W. Ophey. Robust key extraction from physical uncloneable functions.
In Proc. Applied Cryptography and Network Security (ACNS), pages 407–422, 2005.
## Appendix A Fourth moment of the Gabor coefficients; case $`w > M`$
In this appendix we calculate the non-Gaussian part of $`\left\langle G^{4} \right\rangle_{\varphi}`$. As discussed in section [3](#S3 "3 Statistics of the Gabor coefficients ‣ The entropy of keys derived from laser speckle"), ([13](#S2.E13 "In 2.3 Statistics of the intensity ‣ 2 The random phase model ‣ The entropy of keys derived from laser speckle")) has 6 permutations with one cycle. These are equivalent because of the possibility of relabeling the dummy variables $`{\overset{\rightarrow}{a}}_{1},\cdots,{\overset{\rightarrow}{a}}_{4}`$. Introducing the notation $`S_{i\hspace{0pt}j} = {{\sinh{\frac{w^{2}}{M\hspace{0pt}R}\hspace{0pt}\overset{\rightarrow}{k}}} \cdot {({{\overset{\rightarrow}{a}}_{i} - {\overset{\rightarrow}{a}}_{j}})}}`$, we can write the non-Gaussian part of $`\left\langle G^{4} \right\rangle_{\varphi}`$ as
<table>
<tbody>
<tr>
<td></td>
<td>6[iλ2z2]4e−2w2k2∑a→1,⋯,a→4exp[−w2M2R2∑j=14a→j2]6superscriptdelimited-[]𝑖superscript𝜆2superscript𝑧24superscript𝑒2superscript𝑤2superscript𝑘2subscriptsubscript→𝑎1⋯subscript→𝑎4superscript𝑤2superscript𝑀2superscript𝑅2superscriptsubscript𝑗14superscriptsubscript→𝑎𝑗2\displaystyle 6\left[\frac{i\lambda^{2}}{z^{2}}\right]^{4}e^{-2w^{2}k^{2}}\sum_{\vec{a}_{1},\cdots,\vec{a}_{4}}\exp\left[-\frac{w^{2}}{M^{2}R^{2}}\sum_{j=1}^{4}\vec{a}_{j}^{2}\right]</td>
<td></td>
<td></td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td>exp[w2M2R2(a→1⋅a→2+a→2⋅a→3+a→3⋅a→4+a→4⋅a→1)]S12S23S34S41.superscript𝑤2superscript𝑀2superscript𝑅2⋅subscript→𝑎1subscript→𝑎2⋅subscript→𝑎2subscript→𝑎3⋅subscript→𝑎3subscript→𝑎4⋅subscript→𝑎4subscript→𝑎1subscript𝑆12subscript𝑆23subscript𝑆34subscript𝑆41\displaystyle\exp\left[\frac{w^{2}}{M^{2}R^{2}}(\vec{a}_{1}\cdot\vec{a}_{2}+\vec{a}_{2}\cdot\vec{a}_{3}+\vec{a}_{3}\cdot\vec{a}_{4}+\vec{a}_{4}\cdot\vec{a}_{1})\right]S_{12}S_{23}S_{34}S_{41}.</td>
<td></td>
<td>(68)</td>
</tr>
</tbody>
</table>
Note that the expressions in the exponents are invariant under relabeling of the summation variables. This allows us to expand the product $`S_{12}\hspace{0pt}S_{23}\hspace{0pt}S_{34}\hspace{0pt}S_{41}`$ into sixteen terms, which can then be grouped together into contributions of the same ‘type’, i.e. equivalent under the summation. This gives
<table>
<tbody>
<tr>
<td></td>
<td>S12S23S34S41→18[1−2cosh2w2MRk→⋅(a→1−a→2)+cosh2w2MRk→⋅(a→1−a→2+a→3−a→4)].→subscript𝑆12subscript𝑆23subscript𝑆34subscript𝑆4118delimited-[]1⋅22superscript𝑤2𝑀𝑅→𝑘subscript→𝑎1subscript→𝑎2⋅2superscript𝑤2𝑀𝑅→𝑘subscript→𝑎1subscript→𝑎2subscript→𝑎3subscript→𝑎4S_{12}S_{23}S_{34}S_{41}\rightarrow\frac{1}{8}\left[1-2\cosh\frac{2w^{2}}{MR}\vec{k}\cdot(\vec{a}_{1}-\vec{a}_{2})+\cosh\frac{2w^{2}}{MR}\vec{k}\cdot(\vec{a}_{1}-\vec{a}_{2}+\vec{a}_{3}-\vec{a}_{4})\right].</td>
<td></td>
<td>(69)</td>
</tr>
</tbody>
</table>
Next we diagonalise the quadratic terms in the exponent of ([68](#A1.E68 "In Appendix A Fourth moment of the Gabor coefficients; case 𝑤>𝑀 ‣ The entropy of keys derived from laser speckle")). The exponent is of the form $`\exp{({- {\beta^{2}\hspace{0pt}a\hspace{0pt}\mu\hspace{0pt}a^{T}}})}`$, where $`a`$ denotes the four-component row vector $`({\overset{\rightarrow}{a}}_{1},\cdots,{\overset{\rightarrow}{a}}_{4})`$, $`\beta = {w/{({M\hspace{0pt}R})}}`$, and the matrix $`\mu`$ is given by
<table>
<tbody>
<tr>
<td></td>
<td>μ=(1−120−12−121−1200−121−12−120−121).𝜇matrix112012121120012112120121\mu=\left(\matrix{\hphantom{-}1&-\mbox{$\textstyle\frac{1}{2}$}&\hphantom{-}0&-\mbox{$\textstyle\frac{1}{2}$}\cr-\mbox{$\textstyle\frac{1}{2}$}&\hphantom{-}1&-\mbox{$\textstyle\frac{1}{2}$}&\hphantom{-}0\cr\hphantom{-}0&-\mbox{$\textstyle\frac{1}{2}$}&\hphantom{-}1&-\mbox{$\textstyle\frac{1}{2}$}\cr-\mbox{$\textstyle\frac{1}{2}$}&\hphantom{-}0&-\mbox{$\textstyle\frac{1}{2}$}&\hphantom{-}1}\right).</td>
<td></td>
<td>(70)</td>
</tr>
</tbody>
</table>
$`\mu`$ has one eigenvalue 0 with eigenvector $`(1,1,1,1)`$. The other three eigenvectors are
<table>
<tbody>
<tr>
<td></td>
<td>12(1,0,−1,0)121010\displaystyle{\textstyle\frac{1}{\sqrt{2}}}(1,0,-1,0)</td>
<td>12(0,1,0,−1)120101\displaystyle\frac{1}{\sqrt{2}}(0,1,0,-1)</td>
<td>12(1,−1,1,−1)121111\displaystyle\mbox{$\textstyle\frac{1}{2}$}(1,-1,1,-1)</td>
<td></td>
<td>(71)</td>
</tr>
</tbody>
</table>
with eigenvalues 1, 1 and 2 respectively. We defining the new summation variables $`{\overset{\rightarrow}{v}}_{0} = {\beta\hspace{0pt}{({{\overset{\rightarrow}{a}}_{1} + {\overset{\rightarrow}{a}}_{2} + {\overset{\rightarrow}{a}}_{3} + {\overset{\rightarrow}{a}}_{4}})}}`$, $`{\overset{\rightarrow}{v}}_{1} = {\beta\hspace{0pt}{({{\overset{\rightarrow}{a}}_{1} - {\overset{\rightarrow}{a}}_{3}})}}`$, $`{\overset{\rightarrow}{v}}_{2} = {\beta\hspace{0pt}{({{\overset{\rightarrow}{a}}_{2} - {\overset{\rightarrow}{a}}_{4}})}}`$ and $`{\overset{\rightarrow}{v}}_{3} = {\beta\hspace{0pt}{({{{{\overset{\rightarrow}{a}}_{1} - {\overset{\rightarrow}{a}}_{2}} + {\overset{\rightarrow}{a}}_{3}} - {\overset{\rightarrow}{a}}_{4}})}}`$. Next we approximate the summations by integrals: $`\sum_{\overset{\rightarrow}{a}}\rightarrow{\lambda^{- 2}\hspace{0pt}{\int{d^{2}\hspace{0pt}a}}}`$. Taking the Jacobian into account, we write $`{\int{d^{2}\hspace{0pt}a_{1}\hspace{0pt}\cdots\hspace{0pt}d^{2}\hspace{0pt}a_{4}}} = {\beta^{- 8}\hspace{0pt}{\int{d^{2}\hspace{0pt}v_{0}\hspace{0pt}\cdots\hspace{0pt}d^{2}\hspace{0pt}v_{3}}}}`$. Thus ([68](#A1.E68 "In Appendix A Fourth moment of the Gabor coefficients; case 𝑤>𝑀 ‣ The entropy of keys derived from laser speckle")) can be approximated as
<table>
<tbody>
<tr>
<td></td>
<td>34β−8λ−8[iλ2z2]4e−2w2k2∫d2v0⋯d2v3exp[−12(v→12+v→22+v→32)]×\displaystyle\frac{3}{4}\beta^{-8}\lambda^{-8}\left[\frac{i\lambda^{2}}{z^{2}}\right]^{4}e^{-2w^{2}k^{2}}\int\!{\rm d}^{2}v_{0}\cdots{\rm d}^{2}v_{3}\;\exp\left[-\mbox{$\textstyle\frac{1}{2}$}(\vec{v}_{1}^{2}+\vec{v}_{2}^{2}+\vec{v}_{3}^{2})\right]\times</td>
<td></td>
<td></td>
</tr>
</tbody>
<tbody>
<tr>
<td></td>
<td>[1+cosh2wk→⋅v→3−2coshwk→⋅(v→3+v→1−v→2)].delimited-[]1⋅2𝑤→𝑘subscript→𝑣3⋅2𝑤→𝑘subscript→𝑣3subscript→𝑣1subscript→𝑣2\displaystyle[1+\cosh 2w\vec{k}\cdot\vec{v}_{3}-2\cosh w\vec{k}\cdot(\vec{v}_{3}+\vec{v}_{1}-\vec{v}_{2})].</td>
<td></td>
<td>(72)</td>
</tr>
</tbody>
</table>
Given the finite summation intervals of $`{\overset{\rightarrow}{a}}_{1}\hspace{0pt}\cdots\hspace{0pt}{\overset{\rightarrow}{a}}_{4}`$, evaluation of the integrals in ([72](#A1.E72 "In Appendix A Fourth moment of the Gabor coefficients; case 𝑤>𝑀 ‣ The entropy of keys derived from laser speckle")) does not yield esthetic results. However, if $`w \gtrapprox {3\hspace{0pt}M}`$ then a substantial part of the Gaussian distribution is covered by the integration, and considering the interval to be infinite is not a bad approximation. The case $`w < M`$ is discussed separately in Appendix [B](#A2 "Appendix B Fourth moment of the Gabor coefficients; case 𝑤≪𝑀 ‣ The entropy of keys derived from laser speckle"). We only have to take into account the finiteness of the $`v_{0}`$-integral. The integrand in ([72](#A1.E72 "In Appendix A Fourth moment of the Gabor coefficients; case 𝑤>𝑀 ‣ The entropy of keys derived from laser speckle")) does not depend on $`{\overset{\rightarrow}{v}}_{0}`$, and this leads to a factor $`\pi\hspace{0pt}{({4\hspace{0pt}\beta\hspace{0pt}R})}^{2}`$, since $`{|{\overset{\rightarrow}{v}}_{0}|} < {4\hspace{0pt}\beta\hspace{0pt}R}`$. The remaining integrals are readily evaluated. The final result is
<table>
<tbody>
<tr>
<td></td>
<td>⟨G4⟩φ≈3σG4+32Iav4(Mw)6(1−2e−w2k2/2+e−2w2k2).subscriptdelimited-⟨⟩superscript𝐺4𝜑3superscriptsubscript𝜎𝐺432superscriptsubscript𝐼av4superscript𝑀𝑤612superscript𝑒superscript𝑤2superscript𝑘22superscript𝑒2superscript𝑤2superscript𝑘2\left\langle G^{4}\right\rangle_{\varphi}\approx 3\sigma_{G}^{4}+\frac{3}{2}I_{\rm av}^{4}(\frac{M}{w})^{6}(1-2e^{-w^{2}k^{2}/2}+e^{-2w^{2}k^{2}}).</td>
<td></td>
<td>(73)</td>
</tr>
</tbody>
</table>
From ([16](#S3.E16 "In 3.1 Second order statistics ‣ 3 Statistics of the Gabor coefficients ‣ The entropy of keys derived from laser speckle")) we see that $`\sigma_{G}^{2}`$ is asymptotically proportional to $`I_{av}^{2}\hspace{0pt}{({M/w})}^{2}`$ for small $`M/w`$. Hence the result ([73](#A1.E73 "In Appendix A Fourth moment of the Gabor coefficients; case 𝑤>𝑀 ‣ The entropy of keys derived from laser speckle")) is of the form $`3\hspace{0pt}\sigma_{G}^{4}\hspace{0pt}{({1 + {\mathcal{O}\hspace{0pt}{\lbrack{M^{2}/w^{2}}\rbrack}}})}`$.
## Appendix B Fourth moment of the Gabor coefficients; case $`w \ll M`$
In this appendix we compute $`\left\langle G^{4} \right\rangle_{\varphi}`$ in the limit where the length scale $`w`$ of the Gabor transform is very small compared to the average speckle size. In this limit, the intensity changes only slowly as a function of $`\overset{\rightarrow}{x}`$ within the Gaussian envelope. Around the point of interest $`{\overset{\rightarrow}{x}}_{0}`$ we can make a linear approximation
<table>
<tbody>
<tr>
<td></td>
<td>I(x→)≈I(x→0)+D→⋅(x→−x→0),𝐼→𝑥𝐼subscript→𝑥0⋅→𝐷→𝑥subscript→𝑥0I(\vec{x})\approx I(\vec{x}_{0})+\vec{D}\cdot(\vec{x}-\vec{x}_{0}),</td>
<td></td>
<td>(74)</td>
</tr>
</tbody>
</table>
with $`\overset{\rightarrow}{D} = {{\nabla I}\hspace{0pt}{({\overset{\rightarrow}{x}}_{0})}}`$. The Gabor transform ([1](#S1.E1 "In 1.2 Binarized Gabor coefficients ‣ 1 Introduction ‣ The entropy of keys derived from laser speckle")) then reduces to
<table>
<tbody>
<tr>
<td></td>
<td>G(w,k→,x→0)≈w2D→⋅k→e−w2k2/2.𝐺𝑤→𝑘subscript→𝑥0⋅superscript𝑤2→𝐷→𝑘superscript𝑒superscript𝑤2superscript𝑘22G(w,\vec{k},\vec{x}_{0})\approx w^{2}\;\vec{D}\cdot\vec{k}\;e^{-w^{2}k^{2}/2}.</td>
<td></td>
<td>(75)</td>
</tr>
</tbody>
</table>
Differentiating ([7](#S2.E7 "In 2.3 Statistics of the intensity ‣ 2 The random phase model ‣ The entropy of keys derived from laser speckle")) and taking the inner product with $`\overset{\rightarrow}{k}`$, we obtain
<table>
<tbody>
<tr>
<td></td>
<td>D→⋅k→=2πiλz3∑a→,b→αa→αb→∗k→⋅(a→−b→)expiπλz[b→2−a→2+2x→0⋅(a→−b→)].⋅→𝐷→𝑘2𝜋𝑖𝜆superscript𝑧3subscript→𝑎→𝑏⋅subscript𝛼→𝑎subscriptsuperscript𝛼→𝑏→𝑘→𝑎→𝑏𝑖𝜋𝜆𝑧delimited-[]superscript→𝑏2superscript→𝑎2⋅2subscript→𝑥0→𝑎→𝑏\vec{D}\cdot\vec{k}=\frac{2\pi i\lambda}{z^{3}}\sum_{\vec{a},\vec{b}}\alpha_{\vec{a}}\alpha^{*}_{\vec{b}}\vec{k}\cdot(\vec{a}-\vec{b})\exp\frac{i\pi}{\lambda z}\left[\vec{b}^{2}-\vec{a}^{2}+2\vec{x}_{0}\cdot(\vec{a}-\vec{b})\right].</td>
<td></td>
<td>(76)</td>
</tr>
</tbody>
</table>
We square ([75](#A2.E75 "In Appendix B Fourth moment of the Gabor coefficients; case 𝑤≪𝑀 ‣ The entropy of keys derived from laser speckle")), apply ([9](#S2.E9 "In 2.3 Statistics of the intensity ‣ 2 The random phase model ‣ The entropy of keys derived from laser speckle")) and replace the sums by integrals. In this way we obtain
<table>
<tbody>
<tr>
<td></td>
<td>σG2=⟨G2(w,k→,x→0)⟩≈4w4k2Iav2M−2e−w2k2.superscriptsubscript𝜎𝐺2delimited-⟨⟩superscript𝐺2𝑤→𝑘subscript→𝑥04superscript𝑤4superscript𝑘2superscriptsubscript𝐼av2superscript𝑀2superscript𝑒superscript𝑤2superscript𝑘2\sigma_{G}^{2}=\left\langle G^{2}(w,\vec{k},\vec{x}_{0})\right\rangle\approx 4w^{4}k^{2}I_{\rm av}^{2}M^{-2}e^{-w^{2}k^{2}}.</td>
<td></td>
<td>(77)</td>
</tr>
</tbody>
</table>
The fourth moment of ([75](#A2.E75 "In Appendix B Fourth moment of the Gabor coefficients; case 𝑤≪𝑀 ‣ The entropy of keys derived from laser speckle")) is obtained using ([13](#S2.E13 "In 2.3 Statistics of the intensity ‣ 2 The random phase model ‣ The entropy of keys derived from laser speckle")) and again replacing summations by integrations,
<table>
<tbody>
<tr>
<td></td>
<td>⟨G4(w,k→,x→0)⟩≈3σG4(1+164).delimited-⟨⟩superscript𝐺4𝑤→𝑘subscript→𝑥03superscriptsubscript𝜎𝐺41164\left\langle G^{4}(w,\vec{k},\vec{x}_{0})\right\rangle\approx 3\sigma_{G}^{4}(1+\frac{1}{64}).</td>
<td></td>
<td>(78)</td>
</tr>
</tbody>
</table> |
0711.2442 | cs.NI | cyber | Are networks with more edges easier to synchronize?
<sup>\*</sup>
<sup>\*</sup>
\*
This work is jointly supported by the National Natural Science Foundation of China under grant 60674093, the Key Projects of Educational Ministry under grant 107110 and the City University of Hong Kong under the Research Enhancement Scheme and SRG grant 7002134.
Zhisheng Duan<sup>1</sup>, Wenxu Wang<sup>2</sup>, Chao Liu<sup>1</sup>, and Guanrong Chen<sup>1,2</sup>
<sup>1</sup>State Key Laboratory for Turbulence and Complex Systems, Department of Mechanics and Aerospace Engineering, College of Engineering, Peking University, Beijing 100871, P. R. China
<sup>2</sup> Department of electronic Engineering, City University of Hong Kong, Hong Kong
Email: duanzs@pku.edu.cn, Fax: (8610)62765037
Abstract. In this paper, the relationship between the network synchronizability and the edge distribution of its associated graph is investigated. First, it is shown that adding one edge to a cycle definitely decreases the network sychronizability. Then, since sometimes the synchronizability can be enhanced by changing the network structure, the question of whether the networks with more edges are easier to synchronize is addressed. It is shown by examples that the answer is negative. This reveals that generally there are redundant edges in a network, which not only make no contributions to synchronization but actually may reduce the synchronizability. Moreover, an example shows that the node betweenness centrality is not always a good indicator for the network synchronizability. Finally, some more examples are presented to illustrate how the network synchronizability varies following the addition of edges, where all the examples show that the network synchronizability globally increases but locally fluctuates as the number of added edges increases.
Keywords. Complex network, Complementary graph, Synchronizability, Edge addition.
## 1 Introduction and problem formulation
Systems composing of dynamical units are ubiquitous in nature, ranging from physical to technological, and to biological fields. These systems can be naturally described by networks with nodes representing the dynamical units and links representing the interactions among them. The topologies of such networks have been extensively studied and some common architectures have been discovered \[[1](#bib.bib1), [2](#bib.bib2)\]. The small-world property, for example, characterized by short average distance and high clustering among nodes, is one of the most common properties shared by many real networks \[[3](#bib.bib3)\]. More significantly, many networks show high heterogeneity of node connectivity, which typically possesses a power-law distribution, named scale-free networks \[[4](#bib.bib4)\]. It is known that these topological characteristics have strong influence on the dynamics of the structured systems, such as epidemic spreading, traffic congestion, collective synchronization, and so on \[[5](#bib.bib5), [6](#bib.bib6)\]. From this viewpoint, systematically understanding the network structural effects on their dynamical processes is of both theoretical and practical importance.
Synchronization behavior, in particular, as a widely observed phenomenon in networked systems has received a great deal of attention in the past decades \[[7](#bib.bib7), [8](#bib.bib8), [9](#bib.bib9), [10](#bib.bib10), [11](#bib.bib11), [12](#bib.bib12), [13](#bib.bib13), [14](#bib.bib14), [15](#bib.bib15), [16](#bib.bib16), [17](#bib.bib17), [18](#bib.bib18)\]. Oscillator network models have been commonly used to characterize synchronization behaviors. In this setting, a synchonizability theorem provided by Pecora and Carroll \[[19](#bib.bib19)\] indicates that the collective synchronous behavior of a network is completely determined by the network structure, independent of individual oscillator dynamics, provided that the network coupling strength satisfies some strong conditions. In this framework, it has been found that, compared to regular lattices, small-world networks have remarkably better synchronizability \[[20](#bib.bib20)\]. In contrast to small-world networks, scale-free networks tend to inhibit synchronization, although they have much shorter average distances than small-world networks \[[21](#bib.bib21)\] which are generally deemed to be advantageous for synchronization. Therefore the node betweenness centrality was provided as a good indicator to the network synchronizability \[[22](#bib.bib22)\]. Since the synchronizability is correlated with many topological properties, a natural question is which property is the most significant to the synchronizability? Donetti et al. \[[23](#bib.bib23)\] gave a good answer to this question by an optimization argument. They pointed out that a network with optimized synchronizability should have an extremely homogeneous structure, i.e., the distributions of some fundamental topological properties should be very narrow. Their work provides a constructive approach to the issue of networked synchronization, making a big progress in this area. However, some issues still remain unclear, e.g., what is the most important topological property for the synchronizability? And what is the effect of the connectivity density on the synchronizability? In particular, as admitted by the authors, this approach cannot theoretically guarantee to find the optimal solution.
Motivated by the above works, this paper focuses on the relationships between the network synchronizability and the edge distribution of the associated graph. The effects of the connection patterns of graphs on the synchronizability are analyzed both theoretically and numerically. It is found that adding an edge to a cycle of size $`N \geq 5`$ definitely decreases the network synchronizability, but the synchronizability may be improved by changing the cyclic structure. However, a further example shows that, by arbitrarily optimizing the network structures, networks with more edges are not necessarily easier to synchronize. This implies that there are redundant edges in the network with respect to synchronization.
Consider a dynamical network consisting of $`N`$ coupled identical nodes, with each node being an $`n`$-dimensional dynamical system, described by
<table>
<tbody>
<tr>
<td></td>
<td>x˙i=f(xi)−c∑j=1NaijH(xj),i=1,2,⋯,N,formulae-sequencesubscript˙𝑥𝑖𝑓subscript𝑥𝑖𝑐superscriptsubscript𝑗1𝑁subscript𝑎𝑖𝑗𝐻subscript𝑥𝑗𝑖12⋯𝑁\dot{x}_{i}=f(x_{i})-c\sum_{j=1}^{N}a_{ij}H(x_{j}),\;i=1,2,\cdots,N,</td>
<td></td>
<td>(1)</td>
</tr>
</tbody>
</table>
where $`x_{i} = {(x_{i\hspace{0pt}1},x_{i\hspace{0pt}2},\cdots,x_{i\hspace{0pt}n})} \in {\mathbb{R}}^{n}`$ is the state vector of node $`i`$, $`{f\hspace{0pt}{( \cdot )}}:{{\mathbb{R}}^{n}\rightarrow{\mathbb{R}}^{n}}`$ is a smooth vector-valued function, constant $`c > 0`$ represents the coupling strength, $`{H\hspace{0pt}{( \cdot )}}:{{\mathbb{R}}^{n}\rightarrow{\mathbb{R}}^{n}}`$ is called the inner linking function, and $`A = {(a_{i\hspace{0pt}j})}_{N \times N}`$ is called the outer coupling matrix or topological matrix, which represents the coupling configuration of the entire network. This paper only considers the case that the network is diffusively connected, i.e., $`A`$ is irreducible and its entries satisfy
<table>
<tbody>
<tr>
<td></td>
<td>aii=−∑j=1,j≠iNaij,i=1,2,⋯,N.formulae-sequencesubscript𝑎𝑖𝑖superscriptsubscriptformulae-sequence𝑗1𝑗𝑖𝑁subscript𝑎𝑖𝑗𝑖12⋯𝑁a_{ii}=-\sum_{j=1,j\neq i}^{N}a_{ij},\;i=1,2,\cdots,N.</td>
<td></td>
</tr>
</tbody>
</table>
Further, suppose that, if there is an edge between node $`i`$ and node $`j`$, then $`a_{i\hspace{0pt}j} = a_{j\hspace{0pt}i} = {- 1}`$, i.e., $`A`$ is a Laplacian matrix. In this setting, $`0`$ is an eigenvalue of $`A`$ with multiplicity 1, and all the other eigenvalues of $`A`$ are strictly positive, which are denoted by
<table>
<tbody>
<tr>
<td></td>
<td>0=λ1<λ2≤λ3≤⋯≤λN.0subscript𝜆1subscript𝜆2subscript𝜆3⋯subscript𝜆𝑁0=\lambda_{1}<\lambda_{2}\leq\lambda_{3}\leq\cdots\leq\lambda_{N}.</td>
<td></td>
<td>(2)</td>
</tr>
</tbody>
</table>
The dynamical network ([1](#S1.E1 "In 1 Introduction and problem formulation")) is said to achieve (asymptotical) synchronization if $`{{{x_{1}\hspace{0pt}{(t)}}\rightarrow{x_{2}\hspace{0pt}{(t)}}\rightarrow\cdots\rightarrow{x_{N}\hspace{0pt}{(t)}}\rightarrow{s\hspace{0pt}{(t)}}},{{\text{as}\hspace{0pt}t}\rightarrow\infty}},`$ where, because of the diffusive coupling configuration, the synchronous state $`{s\hspace{0pt}{(t)}} \in {\mathbb{R}}^{n}`$ is a solution of an individual node, i.e., $`{\overset{˙}{s}\hspace{0pt}{(t)}} = {f\hspace{0pt}{({s\hspace{0pt}{(t)}})}}`$.
It is well known that the eigenratio $`{r\hspace{0pt}{(A)}} = \frac{\lambda_{2}}{\lambda_{N}}`$ of the network structural matrix $`A`$ characterizes the synchronizability. The larger the $`r\hspace{0pt}{(A)}`$ is, the better the synchronizability will be. The enhancement of the network synchronizability and the relationships between $`r\hspace{0pt}{(A)}`$ and the network structural characteristics such as average distance, node betweenness, degree distribution, clustering coefficient, etc., have been well studied \[[24](#bib.bib24), [25](#bib.bib25), [26](#bib.bib26), [27](#bib.bib27), [28](#bib.bib28), [29](#bib.bib29)\]. In particular, the graph-theoretical method was used to discuss the network synchronizability in \[[24](#bib.bib24), [25](#bib.bib25)\]. This paper further investigates the relationship between the network edges and its synchronizability by graph-theoretical tools.
Throughout this paper, for any given undirected graph $`G`$, eigenvalues of $`G`$ mean eigenvalues of its corresponding Laplacian matrix. Notations for graphs and their corresponding Laplacian matrices are not differentiated, and networks and their corresponding graphs are not distinguished, unless otherwise indicated.
## 2 Adding one edge to a cycle
It has been shown by examples that adding edges can either increase or decrease the network synchronizability \[[25](#bib.bib25)\], and it was found \[[30](#bib.bib30)\] that in scale-free networks where the nodes are coupled symmetrically, if some overloaded edges are removed, the network will become more synchronizable.
In this section, consider adding one edge to a given cycle with $`N`$ $`({N \geq 4})`$ nodes. In this case, the added edge definitely decreases the sychronizability. To show this, the following lemmas are needed.
Lemma 1 \[[31](#bib.bib31), [32](#bib.bib32)\] For any given connected graph $`G`$ of size $`N`$, its nonzero eigenvalues indexed as listed in ([2](#S1.E2 "In 1 Introduction and problem formulation")) grow monotonically with the number of added edges; that is, for any added edge $`e`$, $`{\lambda_{i}\hspace{0pt}{({G + e})}} \geq {\lambda_{i}\hspace{0pt}{(G)}}`$, $`i = {1,\cdots,N}`$.
Lemma 1 shows the eigenvalue changes of graphs due to the addition of edges, but it does not show any information about the eigenratio $`r\hspace{0pt}{(A)}`$. Therefore, this eigenratio needs to be studied in more detail.
Lemma 2 \[[27](#bib.bib27), [31](#bib.bib31)\] For any given connected graph $`G`$ of size $`N`$, its largest eigenvalue $`\lambda_{N}`$ satisfies $`\lambda_{N} \geq {d_{m\hspace{0pt}a\hspace{0pt}x} + 1}`$, with equality if and only if $`d_{m\hspace{0pt}a\hspace{0pt}x} = {N - 1}`$. Further, if $`G`$ is not a complete graph, then the smallest nonzero eigenvalue of $`G`$ satisfies $`\lambda_{2} \leq d_{m\hspace{0pt}i\hspace{0pt}n}`$. Here $`d_{m\hspace{0pt}a\hspace{0pt}x}`$ and $`d_{m\hspace{0pt}i\hspace{0pt}n}`$ denote the maximum and minimum degrees of $`G`$.
Lemma 3 \[[27](#bib.bib27)\] For any cycle $`C_{N}`$ with $`N`$ ($`\geq 4`$) nodes, its eigenvalues are given by $`\mu_{1},\cdots,\mu_{N}`$ (not necessarily ordered as in ([2](#S1.E2 "In 1 Introduction and problem formulation"))) with $`\mu_{1} = 0`$ and
<table>
<tbody>
<tr>
<td></td>
<td>μk+1=3−sin(3kπN)sin(kπN),k=1,⋯,N−1.formulae-sequencesubscript𝜇𝑘133𝑘𝜋𝑁𝑘𝜋𝑁𝑘1⋯𝑁1\mu_{k+1}=3-\frac{\sin(\frac{3k\pi}{N})}{\sin(\frac{k\pi}{N})},\,\,k=1,\cdots,N-1.</td>
<td></td>
</tr>
</tbody>
</table>
Lemma 4 Given a connected graph $`G`$, if the multiplicity of its smallest nonzero eigenvalue $`\lambda_{2}`$ is larger than or equal to 2, then adding one edge to $`G`$ can not change this eigenvalue, i.e., $`{{\lambda_{2}\hspace{0pt}{({G + e})}} = {\lambda_{2}\hspace{0pt}{(G)}}}.`$
Proof This lemma follows from the fact that $`{\text{rank~}\hspace{0pt}{({{\lambda_{2}\hspace{0pt}I} - {({G + e})}})}} \leq {{\text{rank~}\hspace{0pt}{({{\lambda_{2}\hspace{0pt}I} - G})}} + 1}`$. $`\square`$
By the above lemmas, one can get the following result for cycles.
Theorem 1 For any cycle $`C_{N}`$ with $`N \geq 4`$ nodes, adding one edge can not enhance its synchronizability $`r\hspace{0pt}{(C_{N})}`$; specifically, one has $`{r\hspace{0pt}{({C_{4} + e})}} = {r\hspace{0pt}{(C_{4})}}`$ and $`{r\hspace{0pt}{({C_{N} + e})}} < {r\hspace{0pt}{(C_{N})}}`$ ($`N \geq 5`$).
Proof $`{r\hspace{0pt}{({C_{4} + e})}} = {r\hspace{0pt}{(C_{4})}}`$ holds obviously. For the case of $`N \geq 5`$, by Lemma 2, one has $`{\lambda_{N}\hspace{0pt}{({C_{N} + e})}} > 4`$. But by Lemma 3, $`{\lambda_{N}\hspace{0pt}{(C_{N})}} \leq 4`$. And Lemma 3 shows that the multiplicity of the smallest nonzero eigenvalue $`\lambda_{2}`$ of $`C_{N}`$ is 2. By Lemma 4, $`{\lambda_{2}\hspace{0pt}{({C_{N} + e})}} = {\lambda_{2}\hspace{0pt}{(C_{N})}}`$. Therefore, $`{r\hspace{0pt}{({C_{N} + e})}} < {r\hspace{0pt}{(C_{N})}}`$ for all $`N \geq 5`$. $`\square`$
Theorem 1 shows that adding one edge a cycle with $`N \geq 5`$ nodes definitely decreases the network sychronizability, as shown by the two examples in Figs. 1-5. By simple computation, one obtains that $`{r\hspace{0pt}{(C_{5})}} = \frac{1.3820}{3.6180} = 0.3820`$ and $`{r\hspace{0pt}{({C_{5} + {e\hspace{0pt}{\{ 1,3\}}}})}} = \frac{1.3820}{4.6180} = 0.2993 < {r\hspace{0pt}{(C_{5})}}`$; $`{r\hspace{0pt}{(C_{6})}} = \frac{1}{4} = 0.25`$, $`{r\hspace{0pt}{({C_{6} + {e\hspace{0pt}{\{ 1,3\}}}})}} = \frac{1}{4.4142} = 0.2265 < {r\hspace{0pt}{(C_{6})}}`$ and $`{r\hspace{0pt}{({C_{6} + {e\hspace{0pt}{\{ 1,4\}}}})}} = \frac{1}{5} = 0.2 < {r\hspace{0pt}{(C_{6})}}`$.
Fig. 1 Graph $`C_{5}`$ Fig. 2 Graph $`C_{5} + {e\hspace{0pt}{\{ 1,3\}}}`$
Fig. 3 Graph $`C_{6}`$ Fig. 4 Graph $`C_{6} + {e\hspace{0pt}{\{ 1,3\}}}`$ Fig. 5 Graph $`C_{6} + {e\hspace{0pt}{\{ 1,4\}}}`$
From the above two examples, one can find that the synchronizability of cycles strictly decreases if only one edge is added, and the results vary depending upon where the edge is placed, e.g., $`{{r\hspace{0pt}{({C_{6} + {e\hspace{0pt}{\{ 1,3\}}}})}} > {r\hspace{0pt}{({C_{6} + {e\hspace{0pt}{\{ 1,4\}}}})}}}.`$ Considering the optimization of network structures, $`r\hspace{0pt}{({C_{6} + {e\hspace{0pt}{\{ 1,3\}}}})}`$ is still not the best one among all graphs with 7 edges connecting 6 nodes in a cycle, as demonstrated in the next section.
## 3 Changing the network structure to enhance its synchronizability
It is shown in the above section that adding one edge to a cycle decreases its synchronizability. A further question is whether the synchronizability can be enhanced by changing the network structure after edge addition. The answer is ‘yes’ in some cases. For example, one can change $`C_{5} + {e\hspace{0pt}{\{ 1,3\}}}`$ to $`C_{5\hspace{0pt}o}`$ as in Fig. 6, and $`C_{6} + {e\hspace{0pt}{\{ 1,3\}}}`$ to $`C_{6\hspace{0pt}o}`$ as in Fig. 7. Then, $`{r\hspace{0pt}{(C_{5\hspace{0pt}o})}} = \frac{2}{5} = 0.4`$ and $`{{r\hspace{0pt}{(C_{6\hspace{0pt}o})}} = \frac{1.2679}{4.7231} = 0.2684}.`$
Fig. 6 Graph $`C_{5\hspace{0pt}o}`$ Fig. 7 Graph $`C_{6\hspace{0pt}o}`$
Comparing with the graphs in Figs. 1-5, one can see that both the synchronizabilities of $`C_{5\hspace{0pt}o}`$ and $`C_{6\hspace{0pt}o}`$ have been improved. In fact, two cycles share a common edge in Figs. 2, 4 and 5. In this case, generally the betweenness centrality is large, or the node-to-node distances are not homogeneous. In comparison, the network structural characteristics are more homogeneous in Figs. 6 and 7. This is consistent with the result of \[[23](#bib.bib23)\]. For simple graphs with a few nodes and edges, as those shown above, one can compute their eigenvalues to find a good structure for the synchronizability. However, for a general graph, how to optimize the network structure toward the best possible synchronizability? Some optimal rules are provided based on an optimizing algorithm in \[[23](#bib.bib23)\]: to have homogeneous degree, node distance, betweenness, and loop distributions. But these rules are observed from simulations, theoretical proofs are not available by now. And, sometimes, these rules are contrary to each other. For example, comparing $`C_{6} + {e\hspace{0pt}{\{ 1,3\}}}`$ with $`C_{6} + {e\hspace{0pt}{\{ 1,4\}}}`$, one can find that the cycle of $`C_{6} + {e\hspace{0pt}{\{ 1,4\}}}`$ is more homogeneous, but the average node distance of $`C_{6} + {e\hspace{0pt}{\{ 1,3\}}}`$ is smaller. It seems that the importance of these rules should be ordered. Although some rules are provided in \[[23](#bib.bib23)\], optimizing the network structure for better synchronizability is still a hard problem, since it is possible that the optimizing algorithm converges to a suboptimal solution.
Other than the rules for optimization, complementary graphs can be used to characterize the network synchronizability \[[25](#bib.bib25)\]. For a given graph $`G`$, the complementary graph of $`G`$, denoted by $`G^{c}`$, is the graph containing all the nodes of $`G`$ and all the edges that are not in $`G`$. For eigenvalues of graphs and complementary graphs, the following lemma is useful (see \[[31](#bib.bib31), [32](#bib.bib32)\] and references therein).
Lemma 5 For any given graph $`G`$, the following statements hold:
\(i\) $`\lambda_{N}\hspace{0pt}{(G)}`$, the largest eigenvalue of $`G`$, satisfies $`{{\lambda_{N}\hspace{0pt}{(G)}} \leq N}.`$
\(ii\) $`{\lambda_{N}\hspace{0pt}{(G)}} = N`$ if and only if $`G^{c}`$ is disconnected.
(iii) If $`G^{c}`$ is disconnected and has (exactly) $`q`$ connected components, then the multiplicity of $`{\lambda_{N}\hspace{0pt}{(G)}} = N`$ is $`q - 1`$.
\(iv\) $`{{{\lambda_{i}\hspace{0pt}{(G^{c})}} + {\lambda_{{N - i} + 2}\hspace{0pt}{(G)}}} = N},{2 \leq i \leq N}`$.
For example, the complementary graph of $`C_{5\hspace{0pt}o}`$ is disconnected (see Fig. 8) and the largest eigenvalue of $`C_{5\hspace{0pt}o}`$ is 5, the number of nodes. Its smallest nonzero eigenvalue $`\lambda_{2} = 2`$ can be easily obtained by computing the largest eigenvalue of its complementary graph. Further, according to the complementary graph, adding one more edge to graph $`C_{5\hspace{0pt}o}`$ can not enhance its synchronizability. However, if adding two more edges to $`C_{5\hspace{0pt}o}`$, e.g., $`e\hspace{0pt}{\{ 1,5\}}`$ and $`e\hspace{0pt}{\{ 3,5\}}`$, then the synchronizability increases to $`r = \frac{3}{5}`$. The corresponding complementary graph becomes the complementary graph of cycle $`C_{4}`$ (see Fig. 9). Cycle $`C_{4}`$ and its complementary graph are very important in graph theory \[[32](#bib.bib32)\] (see the section below for their further applications).
For a given graph $`G`$, if its complementary graph is disconnected and includes two separated graphs $`G_{1}`$ and $`G_{2}`$, then by Lemma 5 the synchronizability of $`G`$ is $`{r\hspace{0pt}{(G)}} = \frac{N - {\max{\{{\lambda_{m\hspace{0pt}a\hspace{0pt}x}\hspace{0pt}{(G_{1})}},{\lambda_{m\hspace{0pt}a\hspace{0pt}x}\hspace{0pt}{(G_{2})}}\}}}}{N}`$, where $`N`$ is the number of nodes of $`G`$ and $`\lambda_{m\hspace{0pt}a\hspace{0pt}x}`$ denotes the maximum eigenvalue of the corresponding Laplacian matrix. It is well known that the complementary graphs of bipartite graphs are disconnected \[[25](#bib.bib25), [33](#bib.bib33)\], so the synchronizability of bipartite graphs can be simply analyzed by the above method. Actually, $`C_{5\hspace{0pt}o}`$ in Fig. 6 is a bipartite graph. Obviously, better understanding and careful manipulation of complementary graphs are useful for enhancing the network synchronizability (see the section below for further applications of complementary graphs).
Fig. 8 Graph $`C_{5\hspace{0pt}o}^{c}`$ Fig. 9 Graph $`C_{4}^{c}`$
## 4 Are networks with more edges easier to synchronize?
For a given graph $`G`$, let $`\mathcal{V}`$ and $`\mathcal{E}`$ denote the sets of nodes and edges of $`G`$, respectively. A graph $`G_{1}`$ is called an induced subgraph of $`G`$, if the node set $`\mathcal{V}_{1}`$ of $`G_{1}`$ is a subset of $`\mathcal{V}`$ and the edges of $`G_{1}`$ are all edges in $`\mathcal{E}`$. In this section, subgraphs and complementary graphs are used to discuss network synchronizability.
In the concern of optimizing network structures, an interesting question is whether networks with more edges are easier to synchronize. In order to answer this question, the following lemma is needed.
Lemma 6 For any given graph $`G`$, suppose $`G_{1}`$ is its induced subgraph including all nodes of $`G`$ with the maximum degree $`d_{m\hspace{0pt}a\hspace{0pt}x}`$. If $`G_{1}`$ includes a cycle $`C_{2\hspace{0pt}k}`$ with even nodes $`2\hspace{0pt}k\hspace{0pt}{({k \geq 2})}`$ as an induced subgraph, then the largest eigenvalue of $`G`$ satisfies $`{{\lambda_{N}\hspace{0pt}{(G)}} \geq {d_{m\hspace{0pt}a\hspace{0pt}x} + 2}}.`$
Proof By Lemma 3, for any cycle $`C_{2\hspace{0pt}k}`$ with even nodes, its largest eigenvalue is 4. And since the degree of every its node is 2, $`- 2`$ must be an eigenvalue of the adjacency matrix of $`C_{2\hspace{0pt}k}`$. Let $`L_{1}`$ be the sub-matrix of the Laplacian matrix of $`G`$ associated with all the nodes in $`G_{1}`$. By the assumption, one has
<table>
<tbody>
<tr>
<td></td>
<td>(dmax+2)I−L1=2I+A(G1)≯0,subscript𝑑𝑚𝑎𝑥2𝐼subscript𝐿12𝐼𝐴subscript𝐺1not-greater-than0(d_{max}+2)I-L_{1}=2I+A(G_{1})\not>0,</td>
<td></td>
</tr>
</tbody>
</table>
where $`A\hspace{0pt}{(G_{1})}`$ is the adjacency matrix of $`G_{1}`$. This implies that the largest eigenvalue of $`G_{1}`$ is larger than or equal to $`d_{m\hspace{0pt}a\hspace{0pt}x} + 2`$. Thus, Lemma 1 leads to the result directly. $`\square`$
Remark 1 Besides Lemma 2, there are few results on lower bounds of the largest eigenvalue of Laplacian matrices in graph theory \[[31](#bib.bib31), [33](#bib.bib33)\]. Since networks with good synchronizability always have homogeneous degree distributions, Lemma 6 is very useful for the study of network synchronization.
Theorem 2 For any graph $`G`$ with 16 edges on 10 nodes, its eigenratio is bounded by $`{r\hspace{0pt}{(G)}} < \frac{2}{5}`$.
Proof If the largest node degree of $`G`$ is $`d_{m\hspace{0pt}a\hspace{0pt}x} \geq 6`$, then the smallest node degree must stisfy $`d_{m\hspace{0pt}i\hspace{0pt}n} \leq 2`$. The conclusion follows directly from Lemma 2. In order to have good synchronizability, the degree distribution of $`G`$ should be homogeneous. Then, first suppose that $`G`$ has 8 nodes with degree 3 and two nodes with degree 4. In this case, by Lemma 2, the largest eigenvalue of $`G`$ is $`{\lambda_{10}\hspace{0pt}{(G)}} > 5`$.
In what follows, consider the largest eigenvalue of the complementary graph $`G^{c}`$. By the above discussion, $`G^{c}`$ must have 8 nodes with degree 6 and two nodes with degree 5. Suppose $`G_{1}`$ is the subgraph of $`G^{c}`$ composing of 8 degree-6 nodes. By direct computing, $`G_{1}`$ must have 19 or 20 edges, and the degree of every its node is at least 4. Hence, $`G_{1}^{c}`$ has 9 or 8 edges and the degree of every its node is at most 3. If the largest eigenvalue of $`G_{1}`$ is 8, i.e., $`G_{1}^{c}`$ is disconnected (Lemma 5), then the largest eigenvalue of $`G^{c}`$ is larger than or equal to 8. Therefore, the smallest nonzero eigenvalue of $`G`$ is $`{\lambda_{2}\hspace{0pt}{(G)}} \leq {10 - 8} = 2`$. By the above discussion, the theorem obviously holds. Hence, suppose both $`G_{1}`$ and $`G_{1}^{c}`$ are connected. Then, $`G_{1}`$ must have a cycle $`C_{4}`$ as an induced subgraph. This holds if and only if $`G_{1}^{c}`$ has $`C_{4}^{c}`$ (see Fig. 9) as an induced subgraph. With only 9 or 8 edges having node degree at most 3, drawing $`G_{1}^{c}`$ directly one can easily reach the conclusion. By Lemma 6, the largest eigenvalue of $`G^{c}`$ must be larger than or equal to 8. Repeating the above discussion concludes the proof.
When $`G`$ has 9 nodes with degree 3 and one node with degree 5, the proof can be similarly completed. $`\square`$
Remark 2 Theorem 2 shows that there is not a graph $`G`$ with 16 edges on 10 nodes whose synchronizability is $`{{r\hspace{0pt}{(G)}} \geq \frac{2}{5}}.`$ However, there does exist a graph $`\Gamma_{1}`$ with 15 edges on 10 nodes whose synchronizability is $`{r\hspace{0pt}{(\Gamma_{1})}} = \frac{2}{5}`$ (see Fig. 10), consistent with the result of \[[23](#bib.bib23)\]. This clearly shows that networks with more edges are not necessarily easier to synchronize. In fact, by the optimal result of \[[23](#bib.bib23)\], $`r = \frac{2}{5}`$ is the optimal synchronizability for graphs with 15 edges on 10 nodes. For any graph $`G`$ with 16 edges on 10 nodes, if both $`G`$ and $`G^{c}`$ have cycles with even nodes, then by Lemma 6 and Theorem 2, $`{r\hspace{0pt}{(G)}} \leq \frac{2}{6} = \frac{1}{3}`$. Therefore, adding one more edge definitely decreases the synchronizability. The existence of cycles with even nodes can be easily tested by drawing graphs, so Lemma 6 is very useful for analyzing the synchronizability of homogeneous networks. Actually, the graph shown in Fig. 10 is quite homogeneous in structure \[[23](#bib.bib23)\]. With one more edge being added, such a structure is destroyed. It is therefore easy to understand why adding more edges do not necessarily result in better synchronizability.
Fig. 10 Graph $`\Gamma_{1}`$, $`{r\hspace{0pt}{(\Gamma_{1})}} = \frac{2}{5}`$
Remark 3 Fig. 11 shows a new graph $`\Gamma_{2}`$ with 20 edges on 10 nodes. It also has quite homogeneous structural characteristics as discussed in \[[23](#bib.bib23)\]. In fact, the betweenness centrality of each node of $`\Gamma_{1}`$ is 6, larger than that of $`\Gamma_{2}`$, 5. But, the synchronizability of graph $`\Gamma_{2}`$ is worse than that of graph $`\Gamma_{1}`$, contrary to the result of \[[22](#bib.bib22)\]. So far, the existing theories \[[21](#bib.bib21), [22](#bib.bib22), [23](#bib.bib23), [28](#bib.bib28)\] can not explain why the synchronizability of $`\Gamma_{1}`$ is better than that of $`\Gamma_{2}`$. This shows the complexity of the relationships between the synchronizability and network structural characteristics. Although $`\Gamma_{2}`$ has the property of homogeneity, another question is whether there exists another graph with 20 edges on 10 nodes having better synchronizability than that of $`\Gamma_{1}`$ or $`\Gamma_{2}`$? If the answer is negative, it implies that generally there are many redundant edges in a network with respect to its synchronizability. This kind of questions are still open today.
Fig. 11 Graph $`\Gamma_{2}`$, $`{r\hspace{0pt}{(\Gamma_{2})}} = \frac{2.7639}{7.2361} \approx 0.382 < {r\hspace{0pt}{(\Gamma_{1})}} = \frac{2}{5}`$
## 5 Some examples
In this section, some examples are given to show the changes of the synchronizability versus the addition of edges.
Example 1 The synchronizability changes by adding edges to graphs with cycles are shown in Figs. 12 and 13, where their initial graphs are $`C_{10}`$ and $`C_{50}`$, respectively, and $`m_{a\hspace{0pt}d\hspace{0pt}d}`$ denotes the number of added edges. The figures in (a)s show the synchronizability changes during the process of adding edges with degree homogeneity (i.e., guaranteeing the node degrees be as homogeneous as possible during edge-adding). The figures in (b)s show the cases corresponding to random edge-adding. Naturally, the corresponding synchronizabilities in (a)s are better than those in (b)s, since degree homogeneity is an important property for networks to achieve good synchronizabilities. In all graphs, it is shown that the synchronizability globally increases but locally fluctuates. According to Theorem 2 and Remark 2, this is the expected phenomenon.
\(a\) Adding edges with degree homogeneity (b) Randomly adding edges
Fig. 12 The synchronizability changes of graphs obtained from $`C_{10}`$ by adding edges
\(a\) Adding edges with degree homogeneity (b) Randomly adding edges
Fig. 13 The synchronizability changes of graphs obtained from $`C_{50}`$ by adding edges
Example 2 The synchronizability changes of graphs obtained from a scale-free graph by randomly adding edges are shown in Fig. 14, for which, the same conclusion can be drawn as in Example 1.
\(a\) Adding edges with degree homogeneity (b) Randomly adding edges
Fig. 14 The synchronizability changes of graphs obtained from
a scale-free graph by adding edges
## 6 Conclusion
In this paper, the relationship between the network synchronizability and the edge distribution of the associated graphs has been studied. It has been proved that the synchronizability definitely decreases if one edge is being added to a cycle with $`N`$ ($`N \geq 5`$) nodes. However, it has also been shown that the synchronizability can be improved by changing the network structure. Further, some examples have shown that some networks with more edges, unexpectedly, have worse synchronizabilities even if the network structures are in some sense optimized. This implies that, for network synchronization, generally there are redundant edges, which do not make any contribution to synchronization but may actually destroy the synchronizability. In addition, an example of a graph with 20 edges on 10 nodes has been provided to show that the existing theories can not explain why it has worse synchronizability than that of a graph with 15 edges on 10 nodes. Some other examples have also been given to show that the network synchronizability globally increases but locally fluctuates due to edge-adding. According to these results, in practical synchronization problems, the synchronizability and the number of communication edges should have a coordinative relation. And one may utilize the redundant edges to improve robustness or other network properties. These kinds of important questions remain open for further research in the future.
## References
- \[1\]
R. Albert and A.-L. Barabási, Rev. Mod. Phys. 74, 47 (2002).
- \[2\]
S.N. Dorogovtsev and J.F.F. Mendes, Adv. Phys. 51, 1079 (2002).
- \[3\]
D. J. Watts and S. H. Strogatz, Collective dynamics of ‘small-world’ networks, Nature 393, 440 (1998).
- \[4\]
A. L. Barabási, R. Albert, and H. Jeong, Science 286, 509 (1999).
- \[5\]
M.E.J. Newman, SIAM Rev. 45, 167 (2003).
- \[6\]
S. Boccaletti, V. Latora, Y. Moreno, M. Chavez, and D. U. Hwang, Physics Reports 424, 175 (2006).
- \[7\]
M. Barahona and L. M. Pecora, Phys. Rev. Lett. 89, 054101 (2002).
- \[8\]
W. Lu and T. Chen, Physica D 198, 148 (2004).
- \[9\]
I. V. Belykh, E. Lange, and M. Hasler, Phys. Rev. Lett. 94, 188101 (2005).
- \[10\]
A. E. Motter, C. Zhou, and J. Kurths, Phys. Rev. E 71, 016116 (2005).
- \[11\]
M. Chavez, D.-U. Hwang, A. Amann, H. G. E. Hentschel, and S. Boccaletti, Phys. Rev. Lett. 94, 218701 (2005).
- \[12\]
C. Zhou, A. E. Motter, and J. Kurths, Phys. Rev. Lett. 96, 034101 (2006).
- \[13\]
F. M. Atay, T. Bıyıkoğlu, and J. Jost, Physica D 224, 35 (2006).
- \[14\] X. Wang, Y. C. Lai, and C. H. Lai, arXiv: nlin. CD/0612057v1.
- \[15\]
F. Sorrentino, M. di Bernardo, F. Garofalo, and G. Chen, Phys. Rev. E 75, 046103 (2007).
- \[16\]
A. Stefański, P. Perlikowski, and T. Kapitaniak, Phys. Rev. E 75, 016210 (2007).
- \[17\]
J. H. Lü, X. H. Yu, G. Chen and D. Z. Cheng, IEEE. Trans. Circuits Syst.-I 51, 787-796 (2004).
- \[18\]
X. F. Wang and G. Chen, IEEE Trans. Circuits Syst-I 49, 54-62 (2002).
- \[19\]
L. M. Pecora and T. L. Carroll, Phys. Rev. Lett. 80, 2109 (1998).
- \[20\]
X. Guardiola, A. Diaz-Guilera, M. Llas, and C. J. Perez, Phys. Rev. E 62, 5565 (2000).
- \[21\]
T. Nishikawa, A. E. Motter, Y. C. Lai, and F. C. Hoppensteadt, Phys. Rev. Lett. 91, 014101 (2003).
- \[22\]
H. Hong, B. J. Kim, M. Y. Choi, and H. Park, Phys. Rev. E 69, 067105 (2004).
- \[23\]
L. Donetti, P. I. Hurtado, and M. A. Muñoz, Phys. Rev. Lett. 95, 188701 (2005).
- \[24\]
F. M. Atay and T. Bıyıkoğlu, Phys. Rev. E 72, 016217 (2005).
- \[25\] Z. S. Duan, G. Chen, and L. Huang, arXiv:0706.2901v1.
- \[26\]
A. E. Motter, C. S. Zhou, and J. Kurths, Europhysics Letters 69, 334 (2005).
- \[27\]
C. W. Wu, Phys. Lett. A 346, 281 (2005).
- \[28\]
M. Zhao, T. Zhou, B. H. Wang, G. Yan, H. J. Yang, and W. J. Bai, Physica A 371, 773 (2006).
- \[29\]
C. S. Zhou and J. Kurths, Phys. Rev. Lett. 96, 164102 (2006).
- \[30\]
C. Y. Yin, W. X. Wang, G. Chen, and B. H. Wang, Phys. Rev. E 74, 047102 (2006).
- \[31\]
R. Merris, Linear Algebra and its Applications 197, 143 (1994).
- \[32\]
R. Merris, Linear Algebra and its Applications 278, 221 (1998).
- \[33\]
Y. L. Pan, Linear Algebra and its Applications 355, 287 (2002). |
0710.5830 | cs.NI | cyber | # Stability analysis of a max-min fair Rate Control Protocol (RCP) in a small buffer regime
Thomas Voice (T.Voice@statslab.cam.ac.uk)
Gaurav Raina (G.Raina@statslab.cam.ac.uk)
###### Abstract
In this note we analyse various stability properties of the max-min fair Rate Control Protocol (RCP) operating with small buffers. We first tackle the issue of stability for networks with arbitrary topologies. We prove that the max-min fair RCP fluid model is globally stable in the absence of propagation delays, and also derive a set of conditions for local stability when arbitrary heterogeneous propagation delays are present. The network delay stability result assumes that, at equilibrium, there is only one bottleneck link along each route. Lastly, in the simpler setting of a single link, single delay model, we investigate the impact of the loss of local stability via a Hopf bifurcation.
keywords rcp, max-min fairness, small buffers, stability, bifurcation.
## 1 Introduction
The Rate Control Protocol (RCP) \[[5](#bib.bib5), [6](#bib.bib6)\] takes a radically different approach towards managing flow and congestion control as compared to the existing Transmission Control Protocol (TCP) congestion avoidance framework. The TCP framework has, imbedded in it, an implicit mechanism for detecting congestion within the network. Loss of a packet, caused by the overflow of a buffer, is intended to provide the necessary feedback information. In sharp contrast, RCP aims to achieve fast flow completion times by communicating *explicit* rate feedback between routers and end-stations \[[4](#bib.bib4)\]. RCP is closely related to the eXplicit Control Protocol, XCP \[[11](#bib.bib11)\]; another algorithm which proposes the use of explicit rate feedback. Both the RCP and XCP algorithms intend to converge to a max-min fair resource allocation \[[2](#bib.bib2), [5](#bib.bib5), [11](#bib.bib11)\]; see \[[16](#bib.bib16), [17](#bib.bib17), [18](#bib.bib18), [22](#bib.bib22)\] and references therein for a sample of the literature exploring various issues related to different notions of fairness in a networking context.
The performance of congestion control algorithms is often coupled with the choice of certain parameters in routers in the network. For example, such parameters may correspond to different choices of Active Queue Management (AQM) schemes \[[3](#bib.bib3), [10](#bib.bib10), [14](#bib.bib14), [15](#bib.bib15)\], or the size of buffers in routers \[[8](#bib.bib8), [20](#bib.bib20)\]. Researchers have begun questioning the design rules for sizing buffers in core routers in the Internet \[[1](#bib.bib1)\] and some recent literature \[[7](#bib.bib7), [20](#bib.bib20), [25](#bib.bib25)\] suggests that buffers should indeed be much smaller as compared to the current design principles. In light of the separate developments made towards the understanding of RCP \[[2](#bib.bib2), [4](#bib.bib4), [5](#bib.bib5), [6](#bib.bib6)\] and buffer sizing \[[1](#bib.bib1), [7](#bib.bib7), [20](#bib.bib20), [25](#bib.bib25)\] it is natural to investigate the two, i.e., RCP and small buffers, together. Such an investigation is also motivated by the observation that previous buffer sizing studies have focussed primarily on the impact of smaller buffers with TCP. In this paper, the focus of our analysis will be on the limiting regime of a max-min fair RCP fluid model operating over a communication network with small buffers.
Previous control theoretic analysis of RCP has focussed on a single bottleneck, where the queue is modelled as a saturated integrator \[[2](#bib.bib2)\]. In our work, we assume that the size of the buffers is small enough so that it is no longer possible to explicitly model the queue. Rather, with such small buffers, at the time scale of operation of the congestion controllers, it is the distribution of queue size that plays the important role \[[25](#bib.bib25)\]. It is worth stating that we do not, in any way, contribute to the question of exactly how small the size of buffers ought to be. The model we analyse is simpler than the one analysed by \[[2](#bib.bib2)\] in that we do not explicitly model the queue, but more involved in the sense that our framework represents a heterogeneous network of arbitrary topology. In our fluid model for RCP, apart from the model for the queue, all other parameters are exactly the same as specified in the original RCP algorithm \[[5](#bib.bib5)\].
The styles of analysis we employ have all been popular in the study of congestion control: for example, global stability without propagation delays \[[21](#bib.bib21), [24](#bib.bib24)\], local stability with delays \[[8](#bib.bib8), [10](#bib.bib10), [12](#bib.bib12), [14](#bib.bib14), [15](#bib.bib15), [21](#bib.bib21), [23](#bib.bib23)\], and also an analysis of the dynamic system when stability may not be guaranteed \[[9](#bib.bib9), [13](#bib.bib13), [19](#bib.bib19)\].
We now outline the essence of our contribution. First, we prove that the fluid model of RCP is globally stable in the absence of propagation delays. Then, we derive a set of conditions for local stability when arbitrary heterogeneous propagation delays are present. The network delay stability result relies upon the assumption that, at equilibrium, there is only one bottleneck link along each route. Furthermore, in the simpler setting of a single link, single delay, model we are able to provide a more in depth study of the RCP algorithm. Using bifurcation analysis, we investigate the impact of the loss of local stability in a special case where we show that the RCP algorithm would always give rise to a super-critical Hopf bifurcation.
An *overview* of this paper follows. In Section $`2`$, we analyse the stability properties of RCP over networks of arbitrary topology. In Section $`3`$, we investigate local instability of RCP in a single link, single delay, model via a Hopf bifurcation analysis. Finally, in Section $`4`$ we summarise our contribution, and discuss some avenues for further research.
## 2 RCP over arbitrary topology networks
In this section, our objective is to analyse the stability of RCP in a small buffer regime, over networks with an arbitrary topology. We first show global stability for the RCP fluid model in the absence of propagation delays and then derive a set of conditions for local stability when arbitrary heterogeneous propagation delays are present. The network delay stability result relies upon the assumption that, at equilibrium, there is only one bottleneck link along each route. That is, at equilibrium, for each route there is a unique link with minimal per-flow bandwidth available. We begin with the model description.
### 2.1 Model description
At the level of theoretical abstraction, our set up for an Internet like communication network follows the commonly adopted framework \[[21](#bib.bib21)\].
We suppose that the communication network comprises of an interconnection of a set of *routes*, $`S`$, with a set of *links*, $`J`$. Each route $`r \in S`$ represents a user of the network. Associated with each route is a set of links which represents the path along which that user transmits information through the network. Further, a route $`r`$ has associated with it a flow rate $`{x_{r}\hspace{0pt}{(t)}} \geq 0`$, which represents a dynamic fluid approximation to the rate at which the user is sending packets along route $`r`$, at time $`t`$.
The flow rate, for each $`r \in S`$ is determined by the links $`l \in r`$, via the use of *explicit* rate feedback. Each link $`l \in J`$ has associated with it a flow rate $`R_{l}\hspace{0pt}{(t)}`$, which represents the maximum flow rate allowed for routes which pass through $`l`$. Each packet which is sent through the network carries, in its header, an explicit rate feedback variable. This variable is initially set to the maximum desired flow rate for $`r`$. As the packet passes through each $`l \in r`$, if the feedback variable is greater than $`R_{l}\hspace{0pt}{(t)}`$, then link $`l`$ sets it equal to $`R_{l}\hspace{0pt}{(t)}`$. When the packet reaches its destination, an acknowledgement packet (ack), containing the final value of the explicit rate feedback variable, is returned to the origin of $`r`$, and the flow rate $`x_{r}\hspace{0pt}{(t)}`$ is updated accordingly.
For each route $`r`$ and link $`l \in r`$, we let $`\tau_{r\hspace{0pt}l}`$ denote the propagation delay from the origin of $`r`$ to $`l`$, i.e. the length of time it takes for a packet to travel from the origin to link $`l`$ along route $`r`$. Let $`\tau_{l\hspace{0pt}r}`$ denote the propagation delay from $`l`$ to the origin of $`r`$, i.e. the length of time it takes for the explicit rate feedback information from link $`l`$ to reach the user who is transmitting along route $`r`$. In RCP, a packet must reach its destination before an acknowledgement packet (ack) is returned to its source. Furthermore, as we are interested in a network with small buffers, we may safely assume that queuing delays are a negligible component of the end to end delay. Thus for all $`l \in r`$, $`{\tau_{r\hspace{0pt}l} + \tau_{l\hspace{0pt}r}} = \tau_{r}`$, the round trip time for route $`r`$.
We now have the following model for the end-system behaviour of RCP. For each route $`r \in S`$,
<table>
<tbody>
<tr>
<td></td>
<td>xr(t)=minl∈rRl(t−τlr).subscript𝑥𝑟𝑡subscriptmin𝑙𝑟subscript𝑅𝑙𝑡subscript𝜏𝑙𝑟x_{r}(t)=\text{min}_{l\in r}R_{l}(t-\tau_{lr}).</td>
<td></td>
<td>(1)</td>
</tr>
</tbody>
</table>
Now, for each link $`l \in J`$, $`R_{l}\hspace{0pt}{(t)}`$ is updated depending upon the total aggregate flow through link $`l`$ at time $`t`$. Although the update rule is discrete, we can model it via a fluid approximation with the following differential equation \[[2](#bib.bib2)\]. For each link $`l \in J`$,
<table>
<tbody>
<tr>
<td></td>
<td>R˙l(t)=Rl(t)(αldlCl(Cl−yl(t))−βlql(t)dl2Cl)Rl(t)+,subscript˙𝑅𝑙𝑡subscript𝑅𝑙𝑡subscriptsuperscriptsubscript𝛼𝑙subscript𝑑𝑙subscript𝐶𝑙subscript𝐶𝑙subscript𝑦𝑙𝑡subscript𝛽𝑙subscript𝑞𝑙𝑡superscriptsubscript𝑑𝑙2subscript𝐶𝑙subscript𝑅𝑙𝑡\dot{R}_{l}(t)=R_{l}(t)\left(\frac{\alpha_{l}}{d_{l}C_{l}}(C_{l}-y_{l}(t))-\frac{\beta_{l}q_{l}(t)}{d_{l}^{2}C_{l}}\right)^{+}_{R_{l}(t)},</td>
<td></td>
<td>(2)</td>
</tr>
</tbody>
</table>
where $`\alpha_{l},\beta_{l}`$ are positive constants, $`C_{l}`$ is the capacity at link $`l`$, $`d_{l}`$ is the average round trip time of the *flows* passing through link $`l`$,
<table>
<tbody>
<tr>
<td></td>
<td>yl(t)=∑r:l∈rxr(t−τrl),subscript𝑦𝑙𝑡subscript:𝑟𝑙𝑟subscript𝑥𝑟𝑡subscript𝜏𝑟𝑙\displaystyle y_{l}(t)=\sum_{r:l\in r}x_{r}(t-\tau_{rl}),</td>
<td></td>
<td>(3)</td>
</tr>
</tbody>
</table>
and $`q_{l}\hspace{0pt}{(t)}`$ is the queue size at time $`t`$. Here we use the notation $`a = {(b)}_{c}^{+}`$ to mean that $`a = 0`$ if $`b < 0`$ and $`c \leq 0`$, otherwise $`a = b`$. In the original RCP model \[[5](#bib.bib5)\], $`\alpha_{l} = \alpha`$ and $`\beta_{l} = \beta`$ for each $`l \in J`$ for some $`\alpha`$ and $`\beta`$. We have allowed these constants to vary between links in order to study their effect on stability.
Model for the queue. The basis of our investigation is to study a regime where the buffers are so small that it is no longer possible to explicitly model the queue as a saturated integrator. This assumption simply expresses the idea that with small enough buffers, we may use the approximation
<table>
<tbody>
<tr>
<td></td>
<td>ql(t)=pl(yl(t)),subscript𝑞𝑙𝑡subscript𝑝𝑙subscript𝑦𝑙𝑡\displaystyle q_{l}(t)=p_{l}(y_{l}(t)),</td>
<td></td>
<td>(4)</td>
</tr>
</tbody>
</table>
where $`p_{l}\hspace{0pt}{( \cdot )}`$ is a continuously differentiable function representing the *mean queue length* of link $`l`$. This is consistent with the observation in \[[25](#bib.bib25)\] that in a small buffer regime, it is the distribution of the queue size that plays the prominent role in the dynamics of the congestion control framework. At the level of interest in this paper, we do not motivate any explicit functional form for the mean queue length. Different functional forms may be suitable candidates, but our primary focus is to investigate some stability properties of the dynamical system defined by ([1](#S2.E1 "In 2.1 Model description ‣ 2 RCP over arbitrary topology networks ‣ Stability analysis of a max-min fair Rate Control Protocol (RCP) in a small buffer regime")-[4](#S2.E4 "In 2.1 Model description ‣ 2 RCP over arbitrary topology networks ‣ Stability analysis of a max-min fair Rate Control Protocol (RCP) in a small buffer regime")).
### 2.2 Global stability without propagation delays
In this subsection, we seek to investigate the stability properties of RCP without taking into account the effects of propagation delays. This gives us some insight into the stability of the algorithm in general, and is also a plausible model for the special case when propagation delays are small in comparison to the update step size of the algorithm. In terms of the fluid model, this would mean
<table>
<tbody>
<tr>
<td></td>
<td>maxl∈JαldlCl+βlql(t)dl2Cl≪minr∈S1τr.much-less-thansubscript𝑙𝐽subscript𝛼𝑙subscript𝑑𝑙subscript𝐶𝑙subscript𝛽𝑙subscript𝑞𝑙𝑡superscriptsubscript𝑑𝑙2subscript𝐶𝑙subscript𝑟𝑆1subscript𝜏𝑟\max_{l\in J}\frac{\alpha_{l}}{d_{l}C_{l}}+\frac{\beta_{l}q_{l}(t)}{d_{l}^{2}C_{l}}\ll\min_{r\in S}\frac{1}{\tau_{r}}.</td>
<td></td>
</tr>
</tbody>
</table>
To model RCP without the effects of delays, we use ([1](#S2.E1 "In 2.1 Model description ‣ 2 RCP over arbitrary topology networks ‣ Stability analysis of a max-min fair Rate Control Protocol (RCP) in a small buffer regime")-[4](#S2.E4 "In 2.1 Model description ‣ 2 RCP over arbitrary topology networks ‣ Stability analysis of a max-min fair Rate Control Protocol (RCP) in a small buffer regime")), setting $`\tau_{r} = 0`$ for all $`r \in S`$, but leaving all other parameters fixed.
Ideally we would like to show that our delay-free model of RCP, at equilibrium, is globally stable. Unfortunately, as the following example demonstrates, we cannot expect the equilibrium points of ([1](#S2.E1 "In 2.1 Model description ‣ 2 RCP over arbitrary topology networks ‣ Stability analysis of a max-min fair Rate Control Protocol (RCP) in a small buffer regime")-[4](#S2.E4 "In 2.1 Model description ‣ 2 RCP over arbitrary topology networks ‣ Stability analysis of a max-min fair Rate Control Protocol (RCP) in a small buffer regime")) to be unique, or even isolated. Consider a network consisting of two identical links, $`l,j`$ and one route $`r = {\{ l,j\}}`$. Then, since $`p_{l}\hspace{0pt}{( \cdot )}`$ is strictly increasing, there exists a unique $`y`$ such that
<table>
<tbody>
<tr>
<td></td>
<td>αldlCl(Cl−y)−βlpl(y)dl2Cl=0.subscript𝛼𝑙subscript𝑑𝑙subscript𝐶𝑙subscript𝐶𝑙𝑦subscript𝛽𝑙subscript𝑝𝑙𝑦superscriptsubscript𝑑𝑙2subscript𝐶𝑙0\frac{\alpha_{l}}{d_{l}C_{l}}(C_{l}-y)-\frac{\beta_{l}p_{l}(y)}{d_{l}^{2}C_{l}}=0.</td>
<td></td>
</tr>
</tbody>
</table>
By inspection, there are equilibrium points at $`R = {(y,y')}`$ and $`R = {(y',y)}`$ for all $`y' \geq y`$.
Alternatively, consider the situation where $`l`$ and $`j`$ are not completely identical, and $`C_{j} > C_{l}`$. Then, there exists a unique $`y'`$ such that
<table>
<tbody>
<tr>
<td></td>
<td>αjdjCj(Cj−y′)−βjpj(y′)dj2Cj=0.subscript𝛼𝑗subscript𝑑𝑗subscript𝐶𝑗subscript𝐶𝑗superscript𝑦′subscript𝛽𝑗subscript𝑝𝑗superscript𝑦′superscriptsubscript𝑑𝑗2subscript𝐶𝑗0\frac{\alpha_{j}}{d_{j}C_{j}}(C_{j}-y^{\prime})-\frac{\beta_{j}p_{j}(y^{\prime})}{d_{j}^{2}C_{j}}=0.</td>
<td></td>
</tr>
</tbody>
</table>
Now, $`y' > y`$, so, if $`x_{r} \leq y`$ then $`{\overset{˙}{R}}_{j}`$ will be strictly positive. However, if $`x_{r} > y`$ then $`{\overset{˙}{R}}_{l}`$ will be strictly negative. Thus, for this example, no equilibrium point exists.
So, there may be an entire manifold of equilibrium vectors for $`R`$, or there may be no equilibrium point at all. However, the same is not true for the equilibrium vector of flow rates $`x`$.
######
Theorem 2.1
.
Suppose that $`x\hspace{0pt}{(t)}`$ evolves according to ([1](#S2.E1 "In 2.1 Model description ‣ 2 RCP over arbitrary topology networks ‣ Stability analysis of a max-min fair Rate Control Protocol (RCP) in a small buffer regime")-[4](#S2.E4 "In 2.1 Model description ‣ 2 RCP over arbitrary topology networks ‣ Stability analysis of a max-min fair Rate Control Protocol (RCP) in a small buffer regime")), with $`\tau_{r} = 0`$ for all $`r \in S`$. Then, there exists a unique vector $`\overline{x}`$ such that $`{x\hspace{0pt}{(t)}}\rightarrow\overline{x}`$ as $`t\rightarrow\infty`$.
###### Proof.
We prove this result recursively for the more general system where, for all $`l \in J`$, we replace ([2](#S2.E2 "In 2.1 Model description ‣ 2 RCP over arbitrary topology networks ‣ Stability analysis of a max-min fair Rate Control Protocol (RCP) in a small buffer regime")) with
<table>
<tbody>
<tr>
<td></td>
<td>R˙l(t)=Rl(t)(ul(t)+αldlCl(Cl−yl(t))−βlpl(yl(t))dl2Cl)Rl(t)+,subscript˙𝑅𝑙𝑡subscript𝑅𝑙𝑡subscriptsuperscriptsubscript𝑢𝑙𝑡subscript𝛼𝑙subscript𝑑𝑙subscript𝐶𝑙subscript𝐶𝑙subscript𝑦𝑙𝑡subscript𝛽𝑙subscript𝑝𝑙subscript𝑦𝑙𝑡superscriptsubscript𝑑𝑙2subscript𝐶𝑙subscript𝑅𝑙𝑡\displaystyle\dot{R}_{l}(t)=R_{l}(t)\left(u_{l}(t)+\frac{\alpha_{l}}{d_{l}C_{l}}(C_{l}-y_{l}(t))-\frac{\beta_{l}p_{l}(y_{l}(t))}{d_{l}^{2}C_{l}}\right)^{+}_{R_{l}(t)},</td>
<td></td>
<td>(5)</td>
</tr>
</tbody>
</table>
where $`{u_{l}\hspace{0pt}{(t)}}\rightarrow 0`$ as $`t\rightarrow\infty`$.
Let $`N_{l}`$ be the number of $`r \in S`$ such that $`l \in r`$. We set $`{\overline{y}}_{l}`$ equal to the unique value such that $`{{\overset{˙}{R}}_{l}\hspace{0pt}{(t)}} = 0`$ when $`{y_{l}\hspace{0pt}{(t)}} = {\overline{y}}_{l}`$, and set $`\overline{R} = {\min_{l \in J}{{\overline{y}}_{l}/N_{l}}}`$. We shall show that, for all $`r \in S`$ with $`\overline{R} = {{\overline{y}}_{l}/N_{l}}`$ for some $`l \in r`$, $`x_{r}\rightarrow\overline{R}`$ as $`t\rightarrow\infty`$. This allows us to remove such an $`x_{r}\hspace{0pt}{(t)}`$ from the system, by replacing it with $`\overline{R}`$ plus a vanishing term which we incorporate into the $`u_{l}\hspace{0pt}{(t)}`$, for each $`l \in r`$. Since this will always remove at least one $`r`$ from $`S`$, this is sufficient to prove our result.
It remains to show that $`x_{r}\rightarrow\overline{R}`$ for $`r \in S`$ with $`\overline{R} = {{\overline{y}}_{l}/N_{l}}`$, for some $`l \in r`$. Now, for any $`\epsilon > 0`$ there exists a $`T`$ such that
<table>
<tbody>
<tr>
<td></td>
<td>|ul(t)|<ϵαldlCl,subscript𝑢𝑙𝑡italic-ϵsubscript𝛼𝑙subscript𝑑𝑙subscript𝐶𝑙|u_{l}(t)|<\epsilon\frac{\alpha_{l}}{d_{l}C_{l}},</td>
<td></td>
</tr>
</tbody>
</table>
for all $`l \in J`$, $`t > T`$. From ([5](#S2.E5 "In Proof. ‣ 2.2 Global stability without propagation delays ‣ 2 RCP over arbitrary topology networks ‣ Stability analysis of a max-min fair Rate Control Protocol (RCP) in a small buffer regime")) we see that, if, for $`t > T`$, for any $`l \in J`$, $`{R_{l}\hspace{0pt}{(t)}} < {{({{\overline{y}}_{l}/N_{l}})} - {2\hspace{0pt}\epsilon}}`$, then $`{{\overset{˙}{R}}_{l}\hspace{0pt}{(t)}} > \delta`$ for some $`\delta > 0`$. Thus, for some $`T'`$, for all $`t > T'`$, $`l \in J`$, $`{R_{l}\hspace{0pt}{(t)}} \geq {{({{\overline{y}}_{l}/N_{l}})} - {2\hspace{0pt}\epsilon}}`$.
Let us assume that $`\epsilon`$ is small enough that, for any $`j \in J`$ with $`{({{\overline{y}}_{j}/N_{j}})} > \overline{R}`$,
<table>
<tbody>
<tr>
<td></td>
<td>y¯jNj−R¯>(2+2maxl∈JNl)ϵ.subscript¯𝑦𝑗subscript𝑁𝑗¯𝑅22subscript𝑙𝐽subscript𝑁𝑙italic-ϵ\frac{\overline{y}_{j}}{N_{j}}-\overline{R}>\left(2+2\max_{l\in J}N_{l}\right)\epsilon.</td>
<td></td>
</tr>
</tbody>
</table>
Now, let $`r \in S`$ be such that $`\overline{R} = {{\overline{y}}_{l}/N_{l}}`$ for some $`l \in r`$. Let $`r'`$ be the set of $`l \in r`$ with $`\overline{R} = {{\overline{y}}_{l}/N_{l}}`$. Suppose that, for some $`t > T'`$,
<table>
<tbody>
<tr>
<td></td>
<td>minl∈r′Rl(t)>R¯+2ϵmaxj∈JNj.subscript𝑙superscript𝑟′subscript𝑅𝑙𝑡¯𝑅2italic-ϵsubscript𝑗𝐽subscript𝑁𝑗\min_{l\in r^{\prime}}R_{l}(t)>\overline{R}+2\epsilon\max_{j\in J}N_{j}.</td>
<td></td>
</tr>
</tbody>
</table>
If $`{x_{r}\hspace{0pt}{(t)}} = {R_{j}\hspace{0pt}{(t)}}`$ and $`j \in r'`$ then $`{x_{r}\hspace{0pt}{(t)}} > {\overline{R} + {2\hspace{0pt}\epsilon\hspace{0pt}{\max_{j \in J}N_{j}}}}`$. Otherwise $`{x_{r}\hspace{0pt}{(t)}} = {R_{j}\hspace{0pt}{(t)}}`$ with $`j \notin r'`$, but since $`t > T'`$, $`x_{r} > {\overline{R} + {2\hspace{0pt}\epsilon\hspace{0pt}{\max_{j \in J}N_{j}}}}`$ still holds. For any $`l \in r'`$, $`x_{r}\hspace{0pt}{(t)}`$ is sufficiently high that $`y_{l}\hspace{0pt}{(t)}`$ must be greater than $`{N_{l}\hspace{0pt}\overline{R}} + {2\hspace{0pt}\epsilon}`$. Thus, $`{{\overset{˙}{R}}_{l}\hspace{0pt}{(t)}} < {- \delta}`$ for some $`\delta > 0`$. Therefore, for some $`T^{\operatorname{\prime\prime}}`$, for all $`t > T^{\operatorname{\prime\prime}}`$,
<table>
<tbody>
<tr>
<td></td>
<td>minl∈r′Rl(t)≤R¯+2ϵmaxj∈JNj.subscript𝑙superscript𝑟′subscript𝑅𝑙𝑡¯𝑅2italic-ϵsubscript𝑗𝐽subscript𝑁𝑗\min_{l\in r^{\prime}}R_{l}(t)\leq\overline{R}+2\epsilon\max_{j\in J}N_{j}.</td>
<td></td>
</tr>
</tbody>
</table>
Hence, for all $`t > T^{\operatorname{\prime\prime}}`$,
<table>
<tbody>
<tr>
<td></td>
<td>|xr(t)−R¯|≤2ϵmaxj∈JNj.subscript𝑥𝑟𝑡¯𝑅2italic-ϵsubscript𝑗𝐽subscript𝑁𝑗\left|x_{r}(t)-\overline{R}\right|\leq 2\epsilon\max_{j\in J}N_{j}.</td>
<td></td>
</tr>
</tbody>
</table>
Since $`\epsilon`$ was arbitrary, the result follows. ∎
Note, the vector $`\overline{x}`$ will be close to that of the max-min fair allocation of flow rates, because of its construction. However, the presence of the function $`p_{l}\hspace{0pt}{( \cdot )}`$ in ([2](#S2.E2 "In 2.1 Model description ‣ 2 RCP over arbitrary topology networks ‣ Stability analysis of a max-min fair Rate Control Protocol (RCP) in a small buffer regime")) means that link capacity will not be fully utilised at equilibrium.
### 2.3 Local stability with propagation delays
In this subsection, we derive conditions for the local stability of ([1](#S2.E1 "In 2.1 Model description ‣ 2 RCP over arbitrary topology networks ‣ Stability analysis of a max-min fair Rate Control Protocol (RCP) in a small buffer regime")-[4](#S2.E4 "In 2.1 Model description ‣ 2 RCP over arbitrary topology networks ‣ Stability analysis of a max-min fair Rate Control Protocol (RCP) in a small buffer regime")) when propagation delays are present. Our result relies upon the assumption that there is only one bottleneck link along each route, that is for each $`r \in S`$ there is only one $`l \in r`$ such that $`{x_{r}\hspace{0pt}{(t)}} = {R_{l}\hspace{0pt}{(t)}}`$ at equilibrium.
For each $`r \in S`$, $`l \in J`$, we let $`{\overline{x}}_{r}`$ be the equilibrium value of $`x_{r}\hspace{0pt}{(t)}`$ and $`{\overline{R}}_{l}`$ be the maximum of $`{\overline{x}}_{s}`$ for all $`s \in S`$ such that $`l \in s`$. For each $`l \in J`$, we let $`{\overline{y}}_{l}`$ be the unique value such that $`{{\overset{˙}{R}}_{l}\hspace{0pt}{(t)}} = 0`$ whenever $`{y_{l}\hspace{0pt}{(t)}} = {\overline{y}}_{l}`$. We can assume, without loss of generality, that for all $`l \in J`$,
<table>
<tbody>
<tr>
<td></td>
<td>y¯l=∑r:l∈rx¯r.subscript¯𝑦𝑙subscript:𝑟𝑙𝑟subscript¯𝑥𝑟\overline{y}_{l}=\sum_{r:l\in r}\overline{x}_{r}.</td>
<td></td>
</tr>
</tbody>
</table>
Otherwise, assuming the system is always local to equilibrium, $`R_{l}\hspace{0pt}{(t)}`$ will simply continually increase. If $`R_{l}\hspace{0pt}{(t)}`$ is initially large enough then $`{{\max_{r \in S}x_{r}}\hspace{0pt}{(t)}} < {R_{l}\hspace{0pt}{(t)}}`$ for all time and thus, $`R_{l}\hspace{0pt}{(t)}`$ has no effect on the rest of the system, and can be ignored.
Now, we have assumed that, for all $`r \in S`$, $`l \in r`$, if $`{\overline{x}}_{r} = {\overline{R}}_{l}`$, then $`{\overline{R}}_{j} > {\overline{R}}_{l}`$ for all $`j \in r`$, $`j \neq l`$. Furthermore, by definition, if $`{\overline{x}}_{r} \neq {\overline{R}}_{l}`$, we must have $`{\overline{x}}_{r} < {\overline{R}}_{l}`$. Thus, for all $`r \in S`$, $`l \in r`$, whenever the system is close to equilibrium, either $`{x_{r}\hspace{0pt}{(t)}} = {R_{l}\hspace{0pt}{({t - \tau_{l\hspace{0pt}r}})}}`$, or else $`{\overline{x}}_{r} < {\overline{R}}_{l}`$ and $`R_{l}\hspace{0pt}{({t - \tau_{l\hspace{0pt}r}})}`$ has no effect on $`x_{r}\hspace{0pt}{(t)}`$. So, we can isolate each $`l \in J`$ and model the effect of the rest of the system on $`R_{l}\hspace{0pt}{(t)}`$ as a vanishing perturbation. This allows us to find conditions for stability for the overall system using a recursive argument.
Accordingly, we initially restrict our attention to the single link case, $`J = {\{ l\}}`$, and we consider the following generalisation of ([2](#S2.E2 "In 2.1 Model description ‣ 2 RCP over arbitrary topology networks ‣ Stability analysis of a max-min fair Rate Control Protocol (RCP) in a small buffer regime")),
<table>
<tbody>
<tr>
<td></td>
<td>R˙l(t)=−Rl(t)(fl(yl(t)−NlR¯l)+ul(t))Rl(t)+,subscript˙𝑅𝑙𝑡subscript𝑅𝑙𝑡subscriptsuperscriptsubscript𝑓𝑙subscript𝑦𝑙𝑡subscript𝑁𝑙subscript¯𝑅𝑙subscript𝑢𝑙𝑡subscript𝑅𝑙𝑡\dot{R}_{l}(t)=-R_{l}(t)\bigg{(}f_{l}\Big{(}y_{l}(t)-N_{l}\overline{R}_{l}\Big{)}+u_{l}(t)\bigg{)}^{+}_{R_{l}(t)},</td>
<td></td>
<td>(6)</td>
</tr>
</tbody>
</table>
where $`y_{l}\hspace{0pt}{(t)}`$ is defined as in ([3](#S2.E3 "In 2.1 Model description ‣ 2 RCP over arbitrary topology networks ‣ Stability analysis of a max-min fair Rate Control Protocol (RCP) in a small buffer regime")), $`f_{l}\hspace{0pt}{( \cdot )}`$ is an increasing differentiable function with $`{f_{l}\hspace{0pt}{(0)}} = 0`$ and $`{u_{l}\hspace{0pt}{(t)}}\rightarrow 0`$ as $`t\rightarrow\infty`$. When the recursive argument is complete, the vector $`u\hspace{0pt}{(t)}`$ represents the behaviour of $`R_{j}\hspace{0pt}{(t)}`$ for $`j \in J`$ such that $`{\overline{R}}_{j} < {\overline{R}}_{l}`$.
######
Theorem 2.2
.
Consider the case where the network consists of only a single link, $`J = l`$, with flow rates following ([6](#S2.E6 "In 2.3 Local stability with propagation delays ‣ 2 RCP over arbitrary topology networks ‣ Stability analysis of a max-min fair Rate Control Protocol (RCP) in a small buffer regime")). If
<table>
<tbody>
<tr>
<td></td>
<td>fl′(0)R¯l∑r∈Sτr<1,superscriptsubscript𝑓𝑙′0subscript¯𝑅𝑙subscript𝑟𝑆subscript𝜏𝑟1f_{l}^{\prime}(0)\overline{R}_{l}\sum_{r\in S}\tau_{r}<1,</td>
<td></td>
</tr>
</tbody>
</table>
then for all $`\epsilon`$, there exists $`a`$ and $`u`$ such that, if $`\left| {{R_{l}\hspace{0pt}{(t)}} - {\overline{R}}_{l}} \right| \leq a`$ for all $`t \leq 0`$ and $`\left| {u_{l}\hspace{0pt}{(t)}} \right| < u`$ for all $`t`$, then $`\left| {{R_{l}\hspace{0pt}{(t)}} - {\overline{R}}_{l}} \right| \leq \epsilon`$ for all $`t`$, and $`{R_{l}\hspace{0pt}{(t)}}\rightarrow{\overline{R}}_{l}`$ as $`t\rightarrow\infty`$.
###### Proof.
Let $`\tau_{l}`$ be $`\max_{r \in S}\tau_{r}`$, the maximum round trip time.
Suppose for some $`t`$, $`\gamma \leq 1`$, $`a'`$ and $`u'`$; $`\left| {{R_{l}\hspace{0pt}{(t)}} - {\overline{R}}_{l}} \right| = {\gamma\hspace{0pt}a'}`$, $`\left| {{R_{l}\hspace{0pt}{(t')}} - {\overline{R}}_{l}} \right| \leq a'`$ and $`\left| {u_{l}\hspace{0pt}{(t')}} \right| < {u'\hspace{0pt}a'}`$ for all $`t' \in \left\lbrack {t - {2\hspace{0pt}\tau_{l}}},t \right\rbrack`$. Then, for all $`t' \in \left\lbrack {t - \tau_{l}},t \right\rbrack`$,
<table>
<tbody>
<tr>
<td></td>
<td>|R˙l(t′)|≤(R¯l+a′)(u′a′+fl′(0)|S|a′+o(a′)).subscript˙𝑅𝑙superscript𝑡′subscript¯𝑅𝑙superscript𝑎′superscript𝑢′superscript𝑎′superscriptsubscript𝑓𝑙′0𝑆superscript𝑎′𝑜superscript𝑎′\displaystyle\left|\dot{R}_{l}(t^{\prime})\right|\leq\left(\overline{R}_{l}+a^{\prime}\right)\Big{(}u^{\prime}a^{\prime}+f_{l}^{\prime}(0)|S|a^{\prime}+o(a^{\prime})\Big{)}.</td>
<td></td>
</tr>
</tbody>
</table>
Thus, for all $`r \in S`$,
<table>
<tbody>
<tr>
<td></td>
<td>|Rl(t−τr)−Rl(t)|≤τrR¯lfl′(0)|S|a′+τr(R¯l+a′)a′u′+o(a′).subscript𝑅𝑙𝑡subscript𝜏𝑟subscript𝑅𝑙𝑡subscript𝜏𝑟subscript¯𝑅𝑙superscriptsubscript𝑓𝑙′0𝑆superscript𝑎′subscript𝜏𝑟subscript¯𝑅𝑙superscript𝑎′superscript𝑎′superscript𝑢′𝑜superscript𝑎′\left|R_{l}(t-\tau_{r})-R_{l}(t)\right|\leq\tau_{r}\overline{R}_{l}f_{l}^{\prime}(0)|S|a^{\prime}+\tau_{r}\left(\overline{R}_{l}+a^{\prime}\right)a^{\prime}u^{\prime}+o(a^{\prime}).</td>
<td></td>
<td>(7)</td>
</tr>
</tbody>
</table>
So,
<table>
<tbody>
<tr>
<td></td>
<td>R˙l(t)Rl(t)subscript˙𝑅𝑙𝑡subscript𝑅𝑙𝑡\displaystyle\displaystyle\frac{\dot{R}_{l}(t)}{R_{l}(t)}</td>
<td>=\displaystyle=</td>
<td>fl′(0)(|S|R¯l−yl(t))−ul(t)+o(a′)superscriptsubscript𝑓𝑙′0𝑆subscript¯𝑅𝑙subscript𝑦𝑙𝑡subscript𝑢𝑙𝑡𝑜superscript𝑎′\displaystyle f_{l}^{\prime}(0)\left(|S|\overline{R}_{l}-y_{l}(t)\right)-u_{l}(t)+o(a^{\prime})</td>
<td></td>
</tr>
<tr>
<td></td>
<td></td>
<td>=\displaystyle=</td>
<td>fl′(0)|S|(R¯l−Rl(t))−ul(t)+fl′(0)∑r∈S(Rl(t)−Rl(t−τr))+o(a′)superscriptsubscript𝑓𝑙′0𝑆subscript¯𝑅𝑙subscript𝑅𝑙𝑡subscript𝑢𝑙𝑡superscriptsubscript𝑓𝑙′0subscript𝑟𝑆subscript𝑅𝑙𝑡subscript𝑅𝑙𝑡subscript𝜏𝑟𝑜superscript𝑎′\displaystyle f_{l}^{\prime}(0)|S|\left(\overline{R}_{l}-R_{l}(t)\right)-u_{l}(t)+f_{l}^{\prime}(0)\sum_{r\in S}\Big{(}R_{l}(t)-R_{l}(t-\tau_{r})\Big{)}+o(a^{\prime})</td>
<td></td>
</tr>
<tr>
<td></td>
<td></td>
<td>=\displaystyle=</td>
<td>fl′(0)|S|γa′+δ,superscriptsubscript𝑓𝑙′0𝑆𝛾superscript𝑎′𝛿\displaystyle f_{l}^{\prime}(0)|S|\gamma a^{\prime}+\delta,</td>
<td></td>
</tr>
</tbody>
</table>
where
<table>
<tbody>
<tr>
<td></td>
<td>|δ|𝛿\displaystyle\left|\delta\right|</td>
<td>≤\displaystyle\leq</td>
<td>fl′(0)R¯l∑r∈Sτrfl′(0)|S|a′+|S|τl(R¯l+a′)a′u′+o(a′)superscriptsubscript𝑓𝑙′0subscript¯𝑅𝑙subscript𝑟𝑆subscript𝜏𝑟superscriptsubscript𝑓𝑙′0𝑆superscript𝑎′𝑆subscript𝜏𝑙subscript¯𝑅𝑙superscript𝑎′superscript𝑎′superscript𝑢′𝑜superscript𝑎′\displaystyle f_{l}^{\prime}(0)\overline{R}_{l}\sum_{r\in S}\tau_{r}f_{l}^{\prime}(0)|S|a^{\prime}+|S|\tau_{l}\left(\overline{R}_{l}+a^{\prime}\right)a^{\prime}u^{\prime}+o(a^{\prime})</td>
<td></td>
</tr>
<tr>
<td></td>
<td></td>
<td>=\displaystyle=</td>
<td>fl′(0)2|S|2R¯ldla′+|S|τl(R¯l+a′)a′u′+o(a′).superscriptsubscript𝑓𝑙′superscript02superscript𝑆2subscript¯𝑅𝑙subscript𝑑𝑙superscript𝑎′𝑆subscript𝜏𝑙subscript¯𝑅𝑙superscript𝑎′superscript𝑎′superscript𝑢′𝑜superscript𝑎′\displaystyle f_{l}^{\prime}(0)^{2}|S|^{2}\overline{R}_{l}d_{l}a^{\prime}+|S|\tau_{l}\left(\overline{R}_{l}+a^{\prime}\right)a^{\prime}u^{\prime}+o(a^{\prime}).</td>
<td></td>
</tr>
</tbody>
</table>
Since $`{f_{l}'\hspace{0pt}{(0)}\hspace{0pt}{|S|}\hspace{0pt}{\overline{R}}_{l}\hspace{0pt}d_{l}} < 1`$, there exists a value of $`a < \epsilon`$, $`u'`$ and $`\gamma < 1`$ such that, if $`a' \leq a`$, then $`|\delta|`$ is guaranteed to be less than $`\gamma\hspace{0pt}f_{l}'\hspace{0pt}{(0)}\hspace{0pt}{|S|}\hspace{0pt}a'`$. In which case, $`{\overset{˙}{R}}_{l}\hspace{0pt}{(t)}`$ must have the same sign as $`{\overline{R}}_{l} - {R_{l}\hspace{0pt}{(t)}}`$.
So, we can take $`u = {u'\hspace{0pt}a}`$, with $`a`$ as given above, and if $`\left| {{R_{l}\hspace{0pt}{(t)}} - {\overline{R}}_{l}} \right| \leq a`$ for all $`t \leq 0`$ and $`\left| {u_{l}\hspace{0pt}{(t)}} \right| < u`$ for all $`t`$, then $`\left| {{R_{l}\hspace{0pt}{(t)}} - {\overline{R}}_{l}} \right| \leq \epsilon`$ for all $`t`$.
Furthermore, for all $`a' < a`$, there exists a time $`T`$ such that, for all $`t > T`$, $`{|{u_{l}\hspace{0pt}{(t)}}|} < {u'\hspace{0pt}a'}`$. From the above analysis, we know that after $`T`$, if $`R_{l}\hspace{0pt}{(t)}`$ is beyond $`a'`$ of $`{\overline{R}}_{l}`$, then it will converge to $`{\overline{R}}_{l}`$ at a rate of $`\gamma`$. Thus, there is some $`T' > T`$ such that, $`\left| {{R_{l}\hspace{0pt}{(t)}} - {\overline{R}}_{l}} \right| \leq a'`$ for all $`t > T'`$. Since $`a'`$ was arbitrary, $`{R_{l}\hspace{0pt}{(t)}}\rightarrow{\overline{R}}_{l}`$ as required. ∎
We now consider the general network case.
######
Theorem 2.3
.
Under ([1](#S2.E1 "In 2.1 Model description ‣ 2 RCP over arbitrary topology networks ‣ Stability analysis of a max-min fair Rate Control Protocol (RCP) in a small buffer regime")-[4](#S2.E4 "In 2.1 Model description ‣ 2 RCP over arbitrary topology networks ‣ Stability analysis of a max-min fair Rate Control Protocol (RCP) in a small buffer regime")), the equilibrium point $`\overline{R}`$ is locally asymptotically stable provided that, for each $`l \in J`$,
<table>
<tbody>
<tr>
<td></td>
<td>(αldlCl+βlpl′(y¯l)dl2Cl)R¯l∑r:l∈r,R¯l=x¯rτr<1.subscript𝛼𝑙subscript𝑑𝑙subscript𝐶𝑙subscript𝛽𝑙subscriptsuperscript𝑝′𝑙subscript¯𝑦𝑙superscriptsubscript𝑑𝑙2subscript𝐶𝑙subscript¯𝑅𝑙subscript:𝑟formulae-sequence𝑙𝑟subscript¯𝑅𝑙subscript¯𝑥𝑟subscript𝜏𝑟1\Bigl{(}\frac{\alpha_{l}}{d_{l}C_{l}}+\frac{\beta_{l}p^{\prime}_{l}(\overline{y}_{l})}{d_{l}^{2}C_{l}}\Bigr{)}\ \overline{R}_{l}\sum_{r:l\in r,\overline{R}_{l}=\overline{x}_{r}}\tau_{r}<1.</td>
<td></td>
<td>(8)</td>
</tr>
</tbody>
</table>
###### Proof.
Let $`\delta`$ be the minimum of $`\left| {{\overline{R}}_{l} - {\overline{R}}_{j}} \right|/2`$ for $`l \neq j \in J`$. Now, if, for all $`t`$, all $`l \in J`$, $`\left| {{R_{l}\hspace{0pt}{(t)}} - {\overline{R}}_{l}} \right| < \delta`$, then each $`l \in J`$ will evolve according to
<table>
<tbody>
<tr>
<td></td>
<td>R˙l(t)=−Rl(t)(fl(yleq(t)−y¯leq)+ul(t)),subscript˙𝑅𝑙𝑡subscript𝑅𝑙𝑡subscript𝑓𝑙subscriptsuperscript𝑦𝑒𝑞𝑙𝑡subscriptsuperscript¯𝑦𝑒𝑞𝑙subscript𝑢𝑙𝑡\dot{R}_{l}(t)=-R_{l}(t)\Big{(}f_{l}\big{(}y^{eq}_{l}(t)-\overline{y}^{eq}_{l}\big{)}+u_{l}(t)\Big{)},</td>
<td></td>
<td>(9)</td>
</tr>
</tbody>
</table>
where $`u_{l}\hspace{0pt}{(t)}`$ represents the effect of $`R_{j}\hspace{0pt}{(t)}`$ for all $`{\overline{R}}_{j} < {\overline{R}}_{l}`$, and
<table>
<tbody>
<tr>
<td></td>
<td>yleq(t)=∑r:l∈r,x¯r=R¯lxr(t−τrl),subscriptsuperscript𝑦𝑒𝑞𝑙𝑡subscript:𝑟formulae-sequence𝑙𝑟subscript¯𝑥𝑟subscript¯𝑅𝑙subscript𝑥𝑟𝑡subscript𝜏𝑟𝑙y^{eq}_{l}(t)=\sum_{r:l\in r,\overline{x}_{r}=\overline{R}_{l}}x_{r}(t-\tau_{rl}),</td>
<td></td>
</tr>
</tbody>
</table>
with $`{\overline{y}}_{l}^{e\hspace{0pt}q}`$ equal to the equilibrium value of $`y_{l}^{e\hspace{0pt}q}\hspace{0pt}{(t)}`$. For each $`l \in J`$, we have
<table>
<tbody>
<tr>
<td></td>
<td>fl(w)=αldlCl(w+y¯l−Cl)+βlpl(y¯l+w)dl2Cl.subscript𝑓𝑙𝑤subscript𝛼𝑙subscript𝑑𝑙subscript𝐶𝑙𝑤subscript¯𝑦𝑙subscript𝐶𝑙subscript𝛽𝑙subscript𝑝𝑙subscript¯𝑦𝑙𝑤superscriptsubscript𝑑𝑙2subscript𝐶𝑙f_{l}(w)=\frac{\alpha_{l}}{d_{l}C_{l}}(w+\overline{y}_{l}-C_{l})+\frac{\beta_{l}p_{l}(\overline{y}_{l}+w)}{d_{l}^{2}C_{l}}.</td>
<td></td>
</tr>
</tbody>
</table>
By definition, ([9](#S2.E9 "In Proof. ‣ 2.3 Local stability with propagation delays ‣ 2 RCP over arbitrary topology networks ‣ Stability analysis of a max-min fair Rate Control Protocol (RCP) in a small buffer regime")) is an example of ([6](#S2.E6 "In 2.3 Local stability with propagation delays ‣ 2 RCP over arbitrary topology networks ‣ Stability analysis of a max-min fair Rate Control Protocol (RCP) in a small buffer regime")), for the reduced network formed by $`l`$ and all routes $`r \in S`$ with $`l \in r`$ and $`{\overline{x}}_{r} = {\overline{R}}_{l}`$. Furthermore, ([8](#S2.E8 "In Theorem 2.3. ‣ 2.3 Local stability with propagation delays ‣ 2 RCP over arbitrary topology networks ‣ Stability analysis of a max-min fair Rate Control Protocol (RCP) in a small buffer regime")) is precisely the condition of Theorem [2.2](#S2.Thmproposition2 "Theorem 2.2. ‣ 2.3 Local stability with propagation delays ‣ 2 RCP over arbitrary topology networks ‣ Stability analysis of a max-min fair Rate Control Protocol (RCP) in a small buffer regime") for this reduced system.
Suppose $`\left| {{R_{l}\hspace{0pt}{(t)}} - {\overline{R}}_{l}} \right| < \delta`$ for all $`l \in J`$, for all $`t`$. By differentiability of $`p_{l}\hspace{0pt}{( \cdot )}`$, for any $`u`$, we can find $`\epsilon`$ such that if $`\left| {{R_{j}\hspace{0pt}{(t)}} - {\overline{R}}_{j}} \right| < \epsilon`$ for all $`t`$, for all $`j`$ such that $`{\overline{R}}_{j} < {\overline{R}}_{l}`$, then $`\left| {u_{l}\hspace{0pt}{(t)}} \right| < u`$ for all $`t`$. This allows us to prove local stability recursively.
We begin with the $`l \in J`$ such that $`{\overline{R}}_{l}`$ is maximal. We apply Theorem [2.2](#S2.Thmproposition2 "Theorem 2.2. ‣ 2.3 Local stability with propagation delays ‣ 2 RCP over arbitrary topology networks ‣ Stability analysis of a max-min fair Rate Control Protocol (RCP) in a small buffer regime") to find conditions under which $`R_{l}\hspace{0pt}{(t)}`$ converges to $`{\overline{R}}_{l}`$ and $`\left| {{R_{l}\hspace{0pt}{(t)}} - {\overline{R}}_{l}} \right| < \delta`$ for all $`t`$. This gives us $`a`$ and $`u`$, where the initial conditions of $`R_{l}\hspace{0pt}{(t)}`$ should be within $`a`$ of $`{\overline{R}}_{l}`$, and $`u_{l}\hspace{0pt}{(t)}`$ should tend to zero and always be bounded by $`u`$. We can find an $`\epsilon < \delta`$ so that if $`\left| {{R_{j}\hspace{0pt}{(t)}} - {\overline{R}}_{j}} \right| < \epsilon`$ for all $`t`$ and $`R_{j}\hspace{0pt}{(t)}`$ converges to $`{\overline{R}}_{j}`$ for all $`j \neq l`$, then $`\left| {u_{l}\hspace{0pt}{(t)}} \right| < u`$ for all $`t`$ and $`u_{l}\hspace{0pt}{(t)}`$ converges to $`0`$. So, we set $`{R_{l}\hspace{0pt}{(t)}} \in \left\lbrack {{\overline{R}}_{l} - a},{{\overline{R}}_{l} + a} \right\rbrack`$ for $`t \leq 0`$ as our initial condition for $`l`$, remove $`l`$ from the network and repeat this process for $`\delta' = \epsilon`$. Since we remove one link each time, eventually we will find suitable initial conditions for the entire network. ∎
Note, each link does not necessarily need to keep track of which flows are under its control in order to meet ([8](#S2.E8 "In Theorem 2.3. ‣ 2.3 Local stability with propagation delays ‣ 2 RCP over arbitrary topology networks ‣ Stability analysis of a max-min fair Rate Control Protocol (RCP) in a small buffer regime")). The condition holds if, for all $`l \in J`$,
<table>
<tbody>
<tr>
<td></td>
<td>(αldl+βlpl′(y¯l)dl2)y¯lCldlp<1,subscript𝛼𝑙subscript𝑑𝑙subscript𝛽𝑙subscriptsuperscript𝑝′𝑙subscript¯𝑦𝑙superscriptsubscript𝑑𝑙2subscript¯𝑦𝑙subscript𝐶𝑙superscriptsubscript𝑑𝑙𝑝1\left(\frac{\alpha_{l}}{d_{l}}+\frac{\beta_{l}p^{\prime}_{l}(\overline{y}_{l})}{d_{l}^{2}}\right)\frac{\overline{y}_{l}}{C_{l}}d_{l}^{p}<1,</td>
<td></td>
</tr>
</tbody>
</table>
where $`d_{l}^{p}`$ is the average per *packet* round trip time,
<table>
<tbody>
<tr>
<td></td>
<td>dlp=1y¯l∑r:l∈rx¯rτr.superscriptsubscript𝑑𝑙𝑝1subscript¯𝑦𝑙subscript:𝑟𝑙𝑟subscript¯𝑥𝑟subscript𝜏𝑟d_{l}^{p}=\frac{1}{\overline{y}_{l}}\sum_{r:l\in r}\overline{x}_{r}\tau_{r}.</td>
<td></td>
</tr>
</tbody>
</table>
This suggests that it should be $`d_{l}^{p}`$ not $`d_{l}`$ that appears in the RCP controller ([2](#S2.E2 "In 2.1 Model description ‣ 2 RCP over arbitrary topology networks ‣ Stability analysis of a max-min fair Rate Control Protocol (RCP) in a small buffer regime")). If, for each $`l \in J`$ we let $`\gamma_{l}`$ be such that $`{p_{l}'\hspace{0pt}{({\overline{y}}_{l})}} = {{\gamma_{l}\hspace{0pt}p_{l}\hspace{0pt}{({\overline{y}}_{l})}}/{\overline{y}}_{l}}`$ then,
<table>
<tbody>
<tr>
<td></td>
<td>pl′(y¯l)=γlαldlβlCl−y¯ly¯l<γlαldlβl,subscriptsuperscript𝑝′𝑙subscript¯𝑦𝑙subscript𝛾𝑙subscript𝛼𝑙subscript𝑑𝑙subscript𝛽𝑙subscript𝐶𝑙subscript¯𝑦𝑙subscript¯𝑦𝑙subscript𝛾𝑙subscript𝛼𝑙subscript𝑑𝑙subscript𝛽𝑙p^{\prime}_{l}(\overline{y}_{l})=\gamma_{l}\frac{\alpha_{l}d_{l}}{\beta_{l}}\frac{C_{l}-\overline{y}_{l}}{\overline{y}_{l}}<\frac{\gamma_{l}\alpha_{l}d_{l}}{\beta_{l}},</td>
<td></td>
</tr>
</tbody>
</table>
thus, ([8](#S2.E8 "In Theorem 2.3. ‣ 2.3 Local stability with propagation delays ‣ 2 RCP over arbitrary topology networks ‣ Stability analysis of a max-min fair Rate Control Protocol (RCP) in a small buffer regime")) is ensured if, for all $`l \in J`$,
<table>
<tbody>
<tr>
<td></td>
<td>αl=dlpdl(1+γl).subscript𝛼𝑙superscriptsubscript𝑑𝑙𝑝subscript𝑑𝑙1subscript𝛾𝑙\alpha_{l}=\frac{d_{l}^{p}}{d_{l}(1+\gamma_{l})}.</td>
<td></td>
<td>(10)</td>
</tr>
</tbody>
</table>
This parameter choice scheme is attractive, because it is fairly decentralised and only requires local information. Each link needs only to measure, estimate or be informed of the round trip times for packets passing through that link. However, these results rely on the weak assumption that each route has only *one* bottleneck link. However, any network with multiple bottleneck routes can easily be transformed into one without, for example, by making a small change in the parameters $`\beta_{l}`$, $`l \in J`$. Preliminary numerical results suggest that Theorem [2.3](#S2.Thmproposition3 "Theorem 2.3. ‣ 2.3 Local stability with propagation delays ‣ 2 RCP over arbitrary topology networks ‣ Stability analysis of a max-min fair Rate Control Protocol (RCP) in a small buffer regime") may hold in general; but finding an analytic result to confirm this, or a counter-example to disprove it, remains an open problem.
In the next section we investigate the impact of loss of local stability in RCP.
## 3 Local bifurcation analysis
A key focus in the choice of parameters for any congestion control proposal is to ensure that they lead to a stable equilibrium. Most proposals for congestion control, for example see \[[21](#bib.bib21)\], lead to the analysis of non-linear time delayed dynamical systems. For such non-linear systems, typically sufficient conditions for local stability guide parameter choices. Following local stability, a local bifurcation theoretic analysis can make us comfortable in running the system close to the *edge* of stability.
We first recapitulate a result about the loss of local stability in a non-linear retarded functional differential equation and then use it to analyse the RCP fluid model. Following the analysis in \[[19](#bib.bib19)\] we outline a local Hopf bifurcation result for the following non-linear delay equation
<table>
<tbody>
<tr>
<td></td>
<td>u˙(t)=η(−ξxu(t−τ)±ξxyu(t)u(t−τ)),˙𝑢𝑡𝜂plus-or-minussubscript𝜉𝑥𝑢𝑡𝜏subscript𝜉𝑥𝑦𝑢𝑡𝑢𝑡𝜏\displaystyle\dot{u}(t)=\eta\big{(}-\xi_{x}u(t-\tau)\pm\xi_{xy}u(t)u(t-\tau)\big{)},</td>
<td></td>
<td>(11)</td>
</tr>
</tbody>
</table>
where $`{{\eta,\tau,\xi_{x},\xi_{x\hspace{0pt}y}} > 0}.`$ The parameter $`\eta`$ has been intentionally introduced to just tip the above equation over the edge of (local) stability. This exogenous, non-dimensional, parameter will *act* as our bifurcation parameter.
We may state the following about equation ([11](#S3.E11 "In 3 Local bifurcation analysis ‣ Stability analysis of a max-min fair Rate Control Protocol (RCP) in a small buffer regime")).
######
Theorem 3.1
.
A necessary and sufficient condition for local stability is
<table>
<tbody>
<tr>
<td></td>
<td>ηξxτ<π/2,𝜂subscript𝜉𝑥𝜏𝜋2\eta\xi_{x}\tau<\pi/2,</td>
<td></td>
</tr>
</tbody>
</table>
and treating $`\eta`$ as the bifurcation parameter, the first local Hopf bifurcation occurs with period $`4\hspace{0pt}\tau`$ at $`\eta = \eta_{c}`$, where $`{{\eta_{c}\hspace{0pt}\xi_{x}\hspace{0pt}\tau} = {\pi/2}}.`$ Further, as the Hopf condition is just violated, the equation will always undergo a super-critical Hopf bifurcation where the amplitude of the stable bifurcating solutions will be proportional to
<table>
<tbody>
<tr>
<td></td>
<td>ξxξxy20π(η−ηc)3π−2.subscript𝜉𝑥subscript𝜉𝑥𝑦20𝜋𝜂subscript𝜂𝑐3𝜋2\displaystyle\frac{\xi_{x}}{\xi_{xy}}\sqrt{\frac{20\pi(\eta-\eta_{c})}{3\pi-2}}.</td>
<td></td>
</tr>
</tbody>
</table>
### 3.1 RCP: single link, single delay model
In our bifurcation analysis, for the sake of simplicity, we shall leave the queuing term out of the model by taking $`\beta = 0`$. Consider the following single link, single delay, RCP fluid model
<table>
<tbody>
<tr>
<td></td>
<td>R˙(t)=ηR(t)(αCτ(C−y(t))),˙𝑅𝑡𝜂𝑅𝑡𝛼𝐶𝜏𝐶𝑦𝑡\displaystyle\dot{R}(t)=\eta R(t)\left(\frac{\alpha}{C\tau}\left(C-y(t)\right)\right),</td>
<td></td>
<td>(12)</td>
</tr>
</tbody>
</table>
where $`{y\hspace{0pt}{(t)}} = {\sum_{s}{R\hspace{0pt}{({t - \tau})}}}`$ is the aggregate load at the link and $`\eta`$ is the non-dimensional bifurcation parameter. Define $`{u\hspace{0pt}{(t)}} = {{R\hspace{0pt}{(t)}} - \overline{R}}`$, and take a Taylor expansion of ([12](#S3.E12 "In 3.1 RCP: single link, single delay model ‣ 3 Local bifurcation analysis ‣ Stability analysis of a max-min fair Rate Control Protocol (RCP) in a small buffer regime")) to obtain
<table>
<tbody>
<tr>
<td></td>
<td>u˙(t)˙𝑢𝑡\displaystyle\dot{u}(t){}</td>
<td></td>
<td>=−ηατu(t−τ)−ηατR¯u(t)u(t−τ).absent𝜂𝛼𝜏𝑢𝑡𝜏𝜂𝛼𝜏¯𝑅𝑢𝑡𝑢𝑡𝜏\displaystyle=-\ \eta\frac{\alpha}{\tau}u(t-\tau)-\eta\frac{\alpha}{\tau\overline{R}}u(t)u(t-\tau).</td>
<td></td>
</tr>
</tbody>
</table>
Now using Theorem [3.1](#S3.Ex23 "Theorem 3.1. ‣ 3 Local bifurcation analysis ‣ Stability analysis of a max-min fair Rate Control Protocol (RCP) in a small buffer regime"), we may state the following about equation ([12](#S3.E12 "In 3.1 RCP: single link, single delay model ‣ 3 Local bifurcation analysis ‣ Stability analysis of a max-min fair Rate Control Protocol (RCP) in a small buffer regime")). A necessary and sufficient condition for local stability is
<table>
<tbody>
<tr>
<td></td>
<td>ηα<π/2,𝜂𝛼𝜋2\eta\alpha<\pi/2,</td>
<td></td>
</tr>
</tbody>
</table>
and treating $`\eta`$ as the bifurcation parameter, the first local Hopf bifurcation occurs with period $`4\hspace{0pt}\tau`$ at $`\eta = \eta_{c}`$, where $`{{\eta_{c}\hspace{0pt}\alpha} = {\pi/2}}.`$ If the Hopf condition is just violated, the equation will *always* undergo a super-critical Hopf bifurcation where the amplitude of the stable bifurcating solutions will be proportional to
<table>
<tbody>
<tr>
<td></td>
<td>R¯20π(η−ηc)3π−2.¯𝑅20𝜋𝜂subscript𝜂𝑐3𝜋2\overline{R}\sqrt{\frac{20\pi(\eta-\eta_{c})}{3\pi-2}}.</td>
<td></td>
</tr>
</tbody>
</table>
We highlight that equation ([12](#S3.E12 "In 3.1 RCP: single link, single delay model ‣ 3 Local bifurcation analysis ‣ Stability analysis of a max-min fair Rate Control Protocol (RCP) in a small buffer regime")) cannot produce a sub-critical Hopf bifurcation. However, in the bifurcation analysis we have omitted any additional non-linear effects that would arise if $`\beta > 0`$ in the RCP model.
We now summarise our contribution in this paper, and outline some avenues for further research.
## 4 Conclusion
It is noteworthy to observe that the small buffer regime has allowed us to tackle the question of stability for a large network with an arbitrary topology. First, we proved that the RCP fluid model is globally stable in the absence of propagation delays. Then, we derive a set of conditions for local stability when arbitrary heterogeneous propagation delays are present. The network delay stability result relies upon the weak assumption that, at equilibrium, there is only one bottleneck link along each route. An interesting avenue for research would be to show that either this result may hold in general, or provide a counter-example to disprove it. Finally in a single link, single delay, model we investigate the impact of the loss of local stability in a special case (setting $`\beta = 0`$) where we show that the RCP algorithm would *always* give rise to an innocuous looking super-critical Hopf bifurcation.
As RCP aims for max-min fairness it is appropriate to first consider a model that embodies the original formulation, as we did in this paper. We now outline some natural avenues for further research.
*Fairness and Stability.* A conveniently parameterized family of $`\alpha`$-fair rate allocations was introduced in \[[17](#bib.bib17)\]. The parameter $`\alpha`$ lies in the range $`(0,\infty)`$, and the cases $`\alpha\rightarrow 0`$, $`\alpha = 1`$ and $`\alpha\rightarrow\infty`$ correspond respectively to an allocation which achieves maximum throughput, is proportionally fair or is max-min fair \[[17](#bib.bib17)\]. In this paper we have only considered a max-min fair allocation mechanism, as has been originally specified \[[5](#bib.bib5)\]. An immediate direction of further research would be to incorporate different notions of fairness into the RCP framework, and analyse the stability of such networks.
*Impact of the RCP parameters: $`\alpha_{l}`$ and $`\beta_{l}`$.* We have only considered the limiting case of very small buffers. The choice of $`\beta`$ impacts the rate at which the queue is drained. Small values of $`\beta`$ drain the queue slowly; so with small $`\beta`$ and for large enough buffers it is appropriate to model the queue as a saturated integrator. However, for large values of $`\beta`$, the queue may drain fast enough so that at the timescale of operation of the congestion control protocols, finer queueing theoretic models may have to be developed. Such queuing models would certainly contribute to the non-linearity in the RCP dynamical system, and hence impact the stability of the time delayed network. Further, the results from local bifurcation theory could also be subtle. In our bifurcation theoretic analysis we set $`\beta = 0`$, thus removing any non-linearity that may arise from the queue.
Acknowledgements
The authors are grateful to Frank Kelly for comments on earlier drafts and acknowledge funding provided by the EPSRC grant GR/S$`86266/01`$. The usual caveat applies.
## References
- \[1\] G. Appenzeller, I. Keslassy, and N. McKeown, “Sizing router buffers,” *Computer Communication Review*, vol. 34, no. 4, pp. 281-292, 2004.
- \[2\] H. Balakrishnan, N. Dukkipati, N. McKeown, and C. Tomlin, “Stability Analysis of Explicit Congestion Control Protocols,” Stanford University Department of Aeronautics and Astronautics Report: SUDAAR 776, 2005. \[online\] http://yuba.stanford.edu/rcp/
- \[3\] S. Deb and R. Srikant, “Rate-based versus Queue-based models of congestion control,” *Proceedings of ACM Sigmetrics*, 2004.
- \[4\] N. Dukkipati and N. McKeown, “Why Flow-Completion Time is the Right Metric for Congestion Control,” *Computer Communication Review*, vol. 36, no. 1, pp. 59-62, 2006.
- \[5\] N. Dukkipati, M. Kobayashi, R. Zhang-Shen, and N. McKeown, “Processor
Sharing Flows in the Internet,” *Thirteenth International Workshop on
Quality of Service (IWQoS)*, 2005. \[online\] http://yuba.stanford.edu/rcp/
- \[6\] N. Dukkipati, N. McKeown, and A. G. Fraser, “RCP-AC: Congestion Control to make flows complete quickly in any environment,”*High-Speed Networking Workshop: The Terabits Challenge*, (in Conjunction with IEEE Infocom), 2006.
- \[7\] M. Enachescu, Y. Ganjali, A. Goel, N. McKeown, and T. Roughgarden, “Part III: routers with very small buffers,” *Computer Communication Review*, vol. 35, no. 3, pp. 83-90, 2005.
- \[8\] H. Han, C. V. Hollot, Y. Chait, and V. Misra, “TCP Networks Stabilized by Buffer-Based AQMs,” *Proceedings of IEEE Infocom*, 2004.
- \[9\] H. Han, C. V. Hollot, D. Towsley, and Y. Chait, “Synchronization of TCP Flows in Networks with Small DropTail Buffers,” *Proceedings of IEEE Conference on Decision and Control*, 2005.
- \[10\] C.V. Hollot, V. Misra, D. Towsley, and W. Gong, “Analysis and design of controllers for AQM routers supporting TCP flows,” *IEEE Transactions on Automatic Control*, vol. 47, no. 6, pp. 945-959, 2002.
- \[11\] D. Katabi, M. Handley, and C. Rohrs, “Internet Congestion Control for Future High Bandwidth-Delay Product Environments,” *Proceedings of ACM Sigcomm,* 2002.
- \[12\] S.S. Kunniyur, and R. Srikant, “Stable, scalable, fair congestion control and AQM schemes that achieve high utilization in the Internet,” *IEEE Transactions on Automatic Control*, vol. 48, no. 11, pp. 2024-2028, 2003.
- \[13\] R.J. La, “Instability of a Tandem network and its propagation under RED,” *IEEE Transactions on Automatic Control*, vol. 49, no. 6, pp. 1006-1011, 2004.
- \[14\] A. Lakshmikantha, C. Beck, and R. Srikant, “Robustness of Real and Virtual Queue based Active Queue Management Schemes,” *IEEE Transactions on Networking*, vol. 13, no. 1, pp. 81-93, 2005.
- \[15\] S. Liu, T. Basar, and R. Srikant, “Exponential-RED: a stabilizing AQM scheme for low- and high-speed TCP protocols,” *IEEE Transactions on Networking*, vol. 13, no. 5, pp. 1068 - 1081, 2005.
- \[16\] P. Marbach, “Priority Service and Max-Min Fairness,” *IEEE Transactions on Networking*, vol. 11, pp. 733-746, 2003.
- \[17\] J. Mo and J. Walrand, “Fair end-to-end window-based congestion control,” *IEEE Transactions on Networking*, vol. 8, no. 5, pp. 556-567, 2000.
- \[18\] B. Radunovic and J.-Y. Le Boudec, “Rate performance objectives of multihop wireless networks,” *IEEE Transactions on Mobile Computing*, vol. 3, no. 4, pp. 334-349, 2004.
- \[19\] G. Raina, “Local Bifurcation Analysis of Some Dual Congestion Control Algorithms,” *IEEE Transactions on Automatic Control*, vol. 50, no. 8, pp. 1135 - 1146, 2005.
- \[20\] G. Raina and D. Wischik, “Buffer sizes for large multiplexers: TCP queueing theory and instability analysis,” *Proceedings of EuroNGI*, 2005.
- \[21\] R. Srikant, *The Mathematics of Internet Congestion Control*. Birkhauser, 2004.
- \[22\] L. Tan, X. Zhang, L. Andrew, and M. Zukerman, “Price-based Max-Min Fair Rate Allocation in Wireless Multi-hop Networks,” *IEEE Communications Letters*, vol. 10, no.1, pp. 31-33, 2006.
- \[23\] G. Vinnicombe, “On the stability of networks operating TCP-like congestion control,” *Proceedings of International Federation of Automatic Control*, 2002.
- \[24\] T. Voice, “A Global Stability Result For Primal-Dual Congestion Control Algorithms With Routing,” *Computer Communication Review*, vol. 34, no. 3, pp. 35-41, 2004.
- \[25\] D. Wischik and N. McKeown, “Part I: buffer sizes for core routers,” *Computer Communication Review*, vol. 35, no. 3, pp. 75-78, 2005. |
0712.0693 | cs.CR | cyber | # Cryptanalysis of an image encryption scheme based on the Hill cipher
Chengqing Li
Dan Zhang
Guanrong Chen
Corresponding author: Chengqing Li (swiftsheep@hotmail.com). Department of Electronic Engineering, City University of Hong Kong, Kowloon Tong, Hong Kong SAR, China College of Computer Science and Technology, Zhejiang University, Hangzhou 310027, Zhejiang, China
###### Abstract
This paper studies the security of an image encryption scheme based on the Hill cipher and reports its following problems: 1) there is a simple necessary and sufficient condition that makes a number of secret keys invalid; 2) it is insensitive to the change of the secret key; 3) it is insensitive to the change of the plain-image; 4) it can be broken with only one known/chosen-plaintext; 5) it has some other minor defects.
###### keywords:
cryptanalysis , encryption , Hill cipher , known-plaintext attack. CLC: TN918, TP393.08.
<sup>†</sup>
<sup>†</sup>journal: J. of Zhejiang University SCIENCE
, , and
## 1 Introduction
The history of cryptography can be traced back to the secret communication among people thousands of years ago. With the development of human society and industrial technology, theories and methods of cryptography have been changed and improved gradually, and meanwhile cryptanalysis has also been developed. In 1949, Shannon published his seminar paper “Communication theory of secrecy systems” \[[1](#bib.bib1)\], which marked the beginning of the modern cryptology.
In the past two decades, the security of multimedia data has become more and more important. However, it has been recognized that the traditional text-encryption schemes cannot efficiently protect multimedia data due to some special properties of the multimedia data, such as strong redundancy and bulk size of the uncompressed data. To meet this challenge, a number of special image encryption schemes based on some nonlinear theories were proposed \[[2](#bib.bib2), [3](#bib.bib3), [4](#bib.bib4)\]. Yet, many of them are found to be insecure from the view point of cryptography \[[5](#bib.bib5), [6](#bib.bib6), [7](#bib.bib7), [8](#bib.bib8), [9](#bib.bib9), [10](#bib.bib10), [11](#bib.bib11), [12](#bib.bib12), [13](#bib.bib13), [14](#bib.bib14), [15](#bib.bib15), [16](#bib.bib16), [17](#bib.bib17)\].
In \[[18](#bib.bib18)\], Ismail et al. tried to encrypt images efficiently by modifying the classical Hill cipher \[[19](#bib.bib19)\]. This paper studies the security of the scheme proposed in \[[18](#bib.bib18)\] and reports the following findings: 1) there exist a number of invalid secret keys; 2) the scheme is insensitive to the change of the secret key; 3) the scheme is insensitive to the change of the plain-image; 4) the scheme can be broken with only one known/chosen plain-image; 5) the scheme has some other minor performance defects.
The rest of this paper is organized as follows. The next section briefly introduces the encryption scheme to be studied. Section [3](#S3 "3 Cryptanalysis ‣ Cryptanalysis of an image encryption scheme based on the Hill cipher") presents detailed cryptanalysis of the scheme. The last section concludes the paper.
## 2 The image encryption scheme to be studied
The scheme proposed in \[[18](#bib.bib18)\] scans the gray scales of a plain-image $`\mathbf{P}`$ (or one channel of a color image) of size $`M \times N`$ in a raster order and divides it into $`\lceil{{M\hspace{0pt}N}/m}\rceil`$ vectors of size $`m`$: $`{\{{\mathbf{P}}_{l}\}}_{l = 1}^{\lceil{{M\hspace{0pt}N}/m}\rceil}`$, where $`{\mathbf{P}}_{l} = {\{{{\mathbf{P}}\hspace{0pt}{({{{({l - 1})} \cdot m} + 1})}},\cdots,{{\mathbf{P}}\hspace{0pt}{({{{({l - 1})} \cdot m} + m})}}\}}`$ (the last vector is padded with some zero bytes if $`M\hspace{0pt}N`$ can not be divided by $`m`$). Then, the vectors $`{\{{\mathbf{P}}_{l}\}}_{l = 1}^{\lceil{{M\hspace{0pt}N}/m}\rceil}`$ are encrypted in increasing order with the following function:
<table>
<tbody>
<tr>
<td></td>
<td>𝑪l=(𝑷l⋅𝑲l)mod256,subscript𝑪𝑙modulo⋅subscript𝑷𝑙subscript𝑲𝑙256\bm{C}_{l}=(\bm{P}_{l}\cdot\bm{K}_{l})\bmod 256,</td>
<td></td>
<td>(1)</td>
</tr>
</tbody>
</table>
where $`{\mathbf{K}}_{1} = {({{\mathbf{K}}_{1}\hspace{0pt}{\lbrack i,j\rbrack}})}_{m \times m}`$, $`{{\mathbf{K}}_{1}\hspace{0pt}{\lbrack i,j\rbrack}} \in {\mathbb{Z}}_{256}`$, the initial state of $`{\mathbf{K}}_{l \geq 2}`$ is set to be $`{\mathbf{K}}_{l - 1}`$, and then every row of $`{\mathbf{K}}_{l}`$ is generated iteratively with the following function, for $`i = 1 \sim m`$:
<table>
<tbody>
<tr>
<td></td>
<td>𝑲l[i,:]=(𝑰𝑽⋅𝑲l)mod256,subscript𝑲𝑙𝑖:modulo⋅𝑰𝑽subscript𝑲𝑙256\bm{K}_{l}[i,:]=(\bm{IV}\cdot\bm{K}_{l})\bmod 256,</td>
<td></td>
<td>(2)</td>
</tr>
</tbody>
</table>
where $`{\mathbf{I}}\hspace{0pt}{\mathbf{V}}`$ is a vector of size $`1 \times m`$ and $`{{\mathbf{I}}\hspace{0pt}{\mathbf{V}}\hspace{0pt}{\lbrack i\rbrack}} \in {\mathbb{Z}}_{256}`$. Finally, the cipher-image is obtained as $`{\mathbf{C}} = {\{{\mathbf{C}}_{l}\}}_{l = 1}^{\lceil{{M\hspace{0pt}N}/m}\rceil}`$.
The secret key of the encryption scheme includes three parts: $`m`$, $`{\mathbf{K}}_{0}`$, and $`{\mathbf{I}}\hspace{0pt}{\mathbf{V}}`$.
The decryption procedure is the same as the above encryption procedure except that Eq. ([1](#S2.E1 "In 2 The image encryption scheme to be studied ‣ Cryptanalysis of an image encryption scheme based on the Hill cipher")) is replaced by the following function:
<table>
<tbody>
<tr>
<td></td>
<td>𝑷l=(𝑪l⋅𝑲l−1)mod256,subscript𝑷𝑙modulo⋅subscript𝑪𝑙superscriptsubscript𝑲𝑙1256\bm{P}_{l}=(\bm{C}_{l}\cdot\bm{K}_{l}^{-1})\bmod 256,</td>
<td></td>
<td>(3)</td>
</tr>
</tbody>
</table>
where $`{{({{\mathbf{K}}_{l} \cdot {\mathbf{K}}_{l}^{- 1}})}\operatorname{mod}256} = \mathbf{I}`$, the identity matrix.
## 3 Cryptanalysis
### 3.1 Some Defects of the Scheme
#### 3.1.1 Invalid keys
An invalid key is a key that fails to ensure the success of the encryption scheme.
From the following Fact [1](#Thmthm1 "Fact 1 ‣ 3.1.1 Invalid keys ‣ 3.1 Some Defects of the Scheme ‣ 3 Cryptanalysis ‣ Cryptanalysis of an image encryption scheme based on the Hill cipher") and Corollary [1](#Thmcorollary1 "Corollary 1 ‣ 3.1.1 Invalid keys ‣ 3.1 Some Defects of the Scheme ‣ 3 Cryptanalysis ‣ Cryptanalysis of an image encryption scheme based on the Hill cipher"), one can see that one secret key in the above-described scheme is invalid if and only if $`{\gcd{({\mathbf{K}}_{1},256)}} \neq 1`$ or $`{{{\mathbf{I}}\hspace{0pt}{\mathbf{V}}\hspace{0pt}{\lbrack i\rbrack}}\operatorname{mod}2} = 0`$.
######
Fact 1
A matrix $`\mathbf{K}`$ is invertible in $`{\mathbb{Z}}_{n}`$ if and only if $`{\gcd{({\det{(\mathbf{K})}},n)}} = 1`$.
######
Proposition 1
$`{\det{({\mathbf{K}}_{l})}} = {\left( {\prod\limits_{i = 1}^{m}{{\mathbf{I}}\hspace{0pt}{\mathbf{V}}\hspace{0pt}{\lbrack i\rbrack}}} \right)\hspace{0pt}{\det{({\mathbf{K}}_{l - 1})}}}`$.
Proof: According to Eq. ([2](#S2.E2 "In 2 The image encryption scheme to be studied ‣ Cryptanalysis of an image encryption scheme based on the Hill cipher")), there is a relation between $`{\mathbf{K}}_{l}`$ and $`{\mathbf{K}}_{l - 1}`$, as follows:
<table>
<tbody>
<tr>
<td></td>
<td>𝑲l=(∑i=1m𝑰𝑽[i]𝑲l−1[i,:]𝑰𝑽[1]𝑲l[1,:]+∑i=2m𝑰𝑽[i]𝑲l−1[i,:]⋮∑i=1m−1𝑰𝑽[i]𝑲l[i,:]+𝑰𝑽[m]𝑲l−1[m,:])mod256.subscript𝑲𝑙modulomissing-subexpressionsuperscriptsubscript𝑖1𝑚𝑰𝑽delimited-[]𝑖subscript𝑲𝑙1𝑖:missing-subexpression𝑰𝑽delimited-[]1subscript𝑲𝑙1:superscriptsubscript𝑖2𝑚𝑰𝑽delimited-[]𝑖subscript𝑲𝑙1𝑖:missing-subexpression⋮missing-subexpressionsuperscriptsubscript𝑖1𝑚1𝑰𝑽delimited-[]𝑖subscript𝑲𝑙𝑖:𝑰𝑽delimited-[]𝑚subscript𝑲𝑙1𝑚:256\bm{K}_{l}=\left(\begin{array}[]{cc}&\sum\limits_{i=1}^{m}\bm{IV}[i]\bm{K}_{l-1}[i,:]\\ &\bm{IV}[1]\bm{K}_{l}[1,:]+\sum\limits_{i=2}^{m}\bm{IV}[i]\bm{K}_{l-1}[i,:]\\ &\vdots\\ &\sum\limits_{i=1}^{m-1}\bm{IV}[i]\bm{K}_{l}[i,:]+\bm{IV}[m]\bm{K}_{l-1}[m,:]\end{array}\right)\bmod 256.</td>
<td></td>
<td>(4)</td>
</tr>
</tbody>
</table>
Subtracting $`\sum\limits_{i = 1}^{i_{0} - 1}{{\mathbf{I}}\hspace{0pt}{\mathbf{V}}\hspace{0pt}{\lbrack i\rbrack}\hspace{0pt}{\mathbf{K}}_{l}\hspace{0pt}{\lbrack i,:\rbrack}}`$ from $`{\mathbf{K}}_{l}\hspace{0pt}{\lbrack i_{0},:\rbrack}`$ for $`i_{0} = m \sim 2`$, one gets
<table>
<tbody>
<tr>
<td></td>
<td>𝑲l′=(∑i=1m𝑰𝑽[i]𝑲l−1[i,:]∑i=2m𝑰𝑽[i]𝑲l−1[i,:]⋮𝑰𝑽[m]𝑲l−1[m,:])mod256.subscriptsuperscript𝑲′𝑙modulomissing-subexpressionsuperscriptsubscript𝑖1𝑚𝑰𝑽delimited-[]𝑖subscript𝑲𝑙1𝑖:missing-subexpressionsuperscriptsubscript𝑖2𝑚𝑰𝑽delimited-[]𝑖subscript𝑲𝑙1𝑖:missing-subexpression⋮missing-subexpression𝑰𝑽delimited-[]𝑚subscript𝑲𝑙1𝑚:256\bm{K}^{\prime}_{l}=\left(\begin{array}[]{cc}&\sum\limits_{i=1}^{m}\bm{IV}[i]\bm{K}_{l-1}[i,:]\\ &\sum\limits_{i=2}^{m}\bm{IV}[i]\bm{K}_{l-1}[i,:]\\ &\vdots\\ &\bm{IV}[m]\bm{K}_{l-1}[m,:]\end{array}\right)\bmod 256.</td>
<td></td>
<td>(5)</td>
</tr>
</tbody>
</table>
Subtracting $`{\mathbf{K}}_{l}'\hspace{0pt}{\lbrack i_{0},:\rbrack}`$ from $`{\mathbf{K}}_{l}'\hspace{0pt}{\lbrack{i_{0} - 1},:\rbrack}`$ for $`i_{0} = 2 \sim m`$, one has
<table>
<tbody>
<tr>
<td></td>
<td>𝑲l′′=(𝑰𝑽[1]𝑲l−1[1,:]𝑰𝑽[2]𝑲l−1[2,:]⋮𝑰𝑽[m]𝑲l−1[m,:])mod256.subscriptsuperscript𝑲′′𝑙modulomissing-subexpression𝑰𝑽delimited-[]1subscript𝑲𝑙11:missing-subexpression𝑰𝑽delimited-[]2subscript𝑲𝑙12:missing-subexpression⋮missing-subexpression𝑰𝑽delimited-[]𝑚subscript𝑲𝑙1𝑚:256\bm{K}^{\prime\prime}_{l}=\left(\begin{array}[]{cc}&\bm{IV}[1]\bm{K}_{l-1}[1,:]\\ &\bm{IV}[2]\bm{K}_{l-1}[2,:]\\ &\vdots\\ &\bm{IV}[m]\bm{K}_{l-1}[m,:]\end{array}\right)\bmod 256.</td>
<td></td>
<td>(6)</td>
</tr>
</tbody>
</table>
Obviously, $`{\det{({\mathbf{K}}_{l})}} = {\det{({\mathbf{K}}_{l}')}} = {\det{({\mathbf{K}}_{l}^{\operatorname{\prime\prime}})}} = {\left( {\prod\limits_{i = 1}^{m}{{\mathbf{I}}\hspace{0pt}{\mathbf{V}}\hspace{0pt}{\lbrack i\rbrack}}} \right)\hspace{0pt}{\det{({\mathbf{K}}_{l - 1})}}}`$, which completes the proof of the proposition. $`\blacksquare`$
######
Corollary 1
$`{\det{({\mathbf{K}}_{l})}} = {\left( {\prod\limits_{i = 1}^{m}{{\mathbf{I}}\hspace{0pt}{\mathbf{V}}\hspace{0pt}{\lbrack i\rbrack}}} \right)^{l - 1}\hspace{0pt}{\det{({\mathbf{K}}_{1})}}}`$.
Proof: The result directly follows from Proposition [1](#Thmproposition1 "Proposition 1 ‣ 3.1.1 Invalid keys ‣ 3.1 Some Defects of the Scheme ‣ 3 Cryptanalysis ‣ Cryptanalysis of an image encryption scheme based on the Hill cipher"). $`\blacksquare`$
#### 3.1.2 Insensitivity to the change of the secret key
Although it is claimed in \[[18](#bib.bib18), Sec. 5\] that the encryption scheme is very sensitive to the change of the sub-keys $`{\mathbf{K}}_{1}`$, $`{\mathbf{I}}\hspace{0pt}{\mathbf{V}}`$, this is not true.
Let’s first study the influence on $`{\mathbf{K}}_{l \geq 2}`$ if only one bit of $`{\mathbf{K}}_{1}`$ is changed. Without loss of generality, assume that the $`n`$-th significant bit of $`{\mathbf{K}}_{1}\hspace{0pt}{(1,j_{0})}`$ is changed from zero to one, where $`0 \leq n \leq 7`$. Let $`{\overset{\sim}{\mathbf{K}}}_{l}`$ denote the modified version of $`{\mathbf{K}}_{l}`$. The change $`{\mathbf{D}}_{l} = {{\overset{\sim}{\mathbf{K}}}_{l} - {\mathbf{K}}_{l}}`$ can be presented by the following two equations:
<table>
<tbody>
<tr>
<td></td>
<td>𝑫l[:,j]≡0,for j≠j0,formulae-sequencesubscript𝑫𝑙:𝑗0for 𝑗subscript𝑗0\bm{D}_{l}[:,j]\equiv 0,\mbox{for }j\neq j_{0},</td>
<td></td>
<td>(7)</td>
</tr>
</tbody>
</table>
<table>
<tbody>
<tr>
<td></td>
<td>𝑫l[:,j0]=(∑i=1m𝑰𝑽[i]𝑫l−1[i,j0]𝑰𝑽[1]𝑫l[1,j0]+∑i=2m𝑰𝑽[i]𝑫l−1[i,j0]⋮∑i=1m−1𝑰𝑽[i]𝑫l[i,j0]+𝑰𝑽[m]𝑫l−1[m,j0])mod256,subscript𝑫𝑙:subscript𝑗0modulomissing-subexpressionsuperscriptsubscript𝑖1𝑚𝑰𝑽delimited-[]𝑖subscript𝑫𝑙1𝑖subscript𝑗0missing-subexpression𝑰𝑽delimited-[]1subscript𝑫𝑙1subscript𝑗0superscriptsubscript𝑖2𝑚𝑰𝑽delimited-[]𝑖subscript𝑫𝑙1𝑖subscript𝑗0missing-subexpression⋮missing-subexpressionsuperscriptsubscript𝑖1𝑚1𝑰𝑽delimited-[]𝑖subscript𝑫𝑙𝑖subscript𝑗0𝑰𝑽delimited-[]𝑚subscript𝑫𝑙1𝑚subscript𝑗0256\bm{D}_{l}[:,j_{0}]=\left(\begin{array}[]{cc}&\sum\limits_{i=1}^{m}\bm{IV}[i]\bm{D}_{l-1}[i,j_{0}]\\ &\bm{IV}[1]\bm{D}_{l}[1,j_{0}]+\sum\limits_{i=2}^{m}\bm{IV}[i]\bm{D}_{l-1}[i,j_{0}]\\ &\vdots\\ &\sum\limits_{i=1}^{m-1}\bm{IV}[i]\bm{D}_{l}[i,j_{0}]+\bm{IV}[m]\bm{D}_{l-1}[m,j_{0}]\end{array}\right)\bmod 256,</td>
<td></td>
<td>(8)</td>
</tr>
</tbody>
</table>
where $`{{\mathbf{D}}_{1}\hspace{0pt}{\lbrack 1,j_{0}\rbrack}} = 2^{n}`$, $`{{\mathbf{D}}_{1}\hspace{0pt}{\lbrack i,j_{0}\rbrack}} = 0`$, $`i = 2 \sim m`$.
Since $`{{{\mathbf{I}}\hspace{0pt}{\mathbf{V}}\hspace{0pt}{\lbrack i\rbrack}}\operatorname{mod}2} \neq 0`$, $`{{\mathbf{D}}_{l}\hspace{0pt}{\lbrack i,j_{0}\rbrack}} \neq 0`$ always exist. From Eq. ([8](#S3.E8 "In 3.1.2 Insensitivity to the change of the secret key ‣ 3.1 Some Defects of the Scheme ‣ 3 Cryptanalysis ‣ Cryptanalysis of an image encryption scheme based on the Hill cipher")), one can see that $`{{\mathbf{D}}_{l}\hspace{0pt}{\lbrack i,j_{0}\rbrack}} \geq 2^{n}`$ exists, which means that only the $`n_{0}`$-th bit of $`{\mathbf{C}}_{l}\hspace{0pt}{\lbrack j_{0}\rbrack}`$ may possibly be changed, where $`n_{0} \geq n`$. Note also that there is no influence on $`{\mathbf{C}}_{l}`$ if
<table>
<tbody>
<tr>
<td></td>
<td>(𝑷l𝑫l[:,j0])mod256=0.modulosubscript𝑷𝑙subscript𝑫𝑙:subscript𝑗02560(\bm{P}_{l}\bm{D}_{l}[:,j_{0}])\bmod 256=0.</td>
<td></td>
</tr>
</tbody>
</table>
To verify the above analysis, an experiment has been carried out using a plain-image “Lenna” with the secret key
<table>
<tbody>
<tr>
<td></td>
<td>m=4,𝑰𝑽=(391733),𝑲1=(112378519103201203119150792135).formulae-sequence𝑚4formulae-sequence𝑰𝑽matrix391733subscript𝑲1matrix112378519103201203119150792135m=4,\bm{IV}=(\begin{matrix}3&9&17&33\end{matrix}),\bm{K}_{1}=\left(\begin{matrix}11&2&3&7\\ 8&5&19&103\\ 201&203&119&150\\ 7&9&21&35\\ \end{matrix}\right).</td>
<td></td>
<td>(9)</td>
</tr>
</tbody>
</table>
Only the $`5`$-th significant bit of $`{\mathbf{K}}_{1}\hspace{0pt}{\lbrack 1,2\rbrack}`$ is changed, namely $`{{\overset{\sim}{\mathbf{K}}}_{1}\hspace{0pt}{\lbrack 1,2\rbrack}} = {{({{{\mathbf{K}}_{1}\hspace{0pt}{\lbrack 1,2\rbrack}} + 2^{5}})}\operatorname{mod}256}`$. Let $`\overset{\sim}{\mathbf{C}}`$ denote the cipher-image corresponding to $`{\overset{\sim}{\mathbf{K}}}_{1}`$. The bit-planes of difference $`|{\overset{\sim}{\mathbf{C}} - {\mathbf{C}}}|`$ are shown in Fig. [1](#S3.F1 "Figure 1 ‣ 3.1.2 Insensitivity to the change of the secret key ‣ 3.1 Some Defects of the Scheme ‣ 3 Cryptanalysis ‣ Cryptanalysis of an image encryption scheme based on the Hill cipher"), which demonstrates the very weak sensitivity of the encryption scheme with respect to $`{\mathbf{K}}_{1}`$.
[Refer to caption](/html/0712.0693/assets/x1.png)
a\) $`0 \sim 4`$-th
[Refer to caption](/html/0712.0693/assets/x2.png)
b\) $`5`$-th
[Refer to caption](/html/0712.0693/assets/x3.png)
c\) $`6`$-th
[Refer to caption](/html/0712.0693/assets/x4.png)
d\) $`7`$-th
Figure 1: The bit-planes of $`|{\overset{\sim}{\mathbf{C}} - {\mathbf{C}}}|`$ when one bit of $`{\mathbf{K}}_{1}`$ is changed.
Now, consider the influence on $`{\mathbf{K}}_{l \geq 2}`$ if only one bit of $`{\mathbf{I}}\hspace{0pt}{\mathbf{V}}`$ is changed. Without loss of generality, assume the $`n`$-th significant bit of $`{\mathbf{I}}\hspace{0pt}{\mathbf{V}}\hspace{0pt}{\lbrack 1\rbrack}`$ is changed from zero to one. Similarly, let $`{\mathbf{D}}_{l}`$ denote the change of $`{\mathbf{K}}_{l}`$. Due to the extremely complex formulation of $`{\mathbf{D}}_{l \geq 3}`$, only $`{\mathbf{D}}_{2}`$ is shown here.
<table>
<tbody>
<tr>
<td></td>
<td>𝑫2[:,j]=(𝑲1[1,j]2n𝑫2[1,j](𝑰𝑽[1]+2n)+𝑲2[1,j]2n𝑫2[2,j]+𝑰𝑽[2]𝑫2[2,j]⋮𝑫2[2,j]+∑i=2m−1𝑰𝑽[i]𝑫2[i,j])mod256,subscript𝑫2:𝑗modulomissing-subexpressionsubscript𝑲11𝑗superscript2𝑛missing-subexpressionsubscript𝑫21𝑗𝑰𝑽delimited-[]1superscript2𝑛subscript𝑲21𝑗superscript2𝑛missing-subexpressionsubscript𝑫22𝑗𝑰𝑽delimited-[]2subscript𝑫22𝑗missing-subexpression⋮missing-subexpressionsubscript𝑫22𝑗superscriptsubscript𝑖2𝑚1𝑰𝑽delimited-[]𝑖subscript𝑫2𝑖𝑗256\bm{D}_{2}[:,j]=\left(\begin{array}[]{cc}&\bm{K}_{1}[1,j]2^{n}\\ &\bm{D}_{2}[1,j](\bm{IV}[1]+2^{n})+\bm{K}_{2}[1,j]2^{n}\\ &\bm{D}_{2}[2,j]+\bm{IV}[2]\bm{D}_{2}[2,j]\\ &\vdots\\ &\bm{D}_{2}[2,j]+\sum\limits_{i=2}^{m-1}\bm{IV}[i]\bm{D}_{2}[i,j]\end{array}\right)\bmod 256,</td>
<td></td>
<td>(10)</td>
</tr>
</tbody>
</table>
where $`j = 1 \sim m`$.
To see the influence of the change of $`{\mathbf{I}}\hspace{0pt}{\mathbf{V}}`$, an experiment has been carried out using plain-image “Lenna”, with the same secret key shown in Eq. ([9](#S3.E9 "In 3.1.2 Insensitivity to the change of the secret key ‣ 3.1 Some Defects of the Scheme ‣ 3 Cryptanalysis ‣ Cryptanalysis of an image encryption scheme based on the Hill cipher")) above. Only the $`5`$-th significant bit of $`{\mathbf{I}}\hspace{0pt}{\mathbf{V}}\hspace{0pt}{\lbrack 1\rbrack}`$ is changed, namely $`{\overset{\sim}{{\mathbf{I}}\hspace{0pt}{\mathbf{V}}}\hspace{0pt}{\lbrack 1\rbrack}} = {{({{{\mathbf{I}}\hspace{0pt}{\mathbf{V}}\hspace{0pt}{\lbrack 1\rbrack}} + 2^{5}})}\operatorname{mod}256}`$. The bit-planes of difference between cipher-images corresponding to $`{\mathbf{I}}\hspace{0pt}{\mathbf{V}}`$ and $`\overset{\sim}{{\mathbf{I}}\hspace{0pt}{\mathbf{V}}}`$, respectively, are shown in Fig. [2](#S3.F2 "Figure 2 ‣ 3.1.2 Insensitivity to the change of the secret key ‣ 3.1 Some Defects of the Scheme ‣ 3 Cryptanalysis ‣ Cryptanalysis of an image encryption scheme based on the Hill cipher").
Comparing Fig. [1](#S3.F1 "Figure 1 ‣ 3.1.2 Insensitivity to the change of the secret key ‣ 3.1 Some Defects of the Scheme ‣ 3 Cryptanalysis ‣ Cryptanalysis of an image encryption scheme based on the Hill cipher") and Fig. [2](#S3.F2 "Figure 2 ‣ 3.1.2 Insensitivity to the change of the secret key ‣ 3.1 Some Defects of the Scheme ‣ 3 Cryptanalysis ‣ Cryptanalysis of an image encryption scheme based on the Hill cipher"), one can see that the sensitivty with respect to $`{\mathbf{I}}\hspace{0pt}{\mathbf{V}}`$ is much stronger than the one with respect to $`{\mathbf{K}}_{1}`$, which agrees with the above theoretical analysis. But one bit change of a sub-key of a secure cipher should cause every bit of the ciphertext changed with a probability of $`\frac{1}{2}`$. Obviously, the sensitivity of the encryption scheme with respect to sub-keys $`{\mathbf{K}}_{1}`$, $`{\mathbf{I}}\hspace{0pt}{\mathbf{V}}`$ is very far from this requirement.
[Refer to caption](/html/0712.0693/assets/x5.png)
a\) $`0 \sim 4`$-th
[Refer to caption](/html/0712.0693/assets/x6.png)
b\) $`5`$-th
[Refer to caption](/html/0712.0693/assets/x7.png)
c\) $`6`$-th
[Refer to caption](/html/0712.0693/assets/x8.png)
d\) $`7`$-th
Figure 2: The bit-planes of $`|{\overset{\sim}{\mathbf{C}} - {\mathbf{C}}}|`$ when one bit of $`{\mathbf{I}}\hspace{0pt}{\mathbf{V}}`$ is changed.
#### 3.1.3 Insensitivity to the change of the plain-image
This property is especially important for image encryption since an image and its watermarked version may be encrypted simultaneously.
Since the role of $`{\mathbf{P}}_{l}`$ in Eq. ([1](#S2.E1 "In 2 The image encryption scheme to be studied ‣ Cryptanalysis of an image encryption scheme based on the Hill cipher")) is exactly the same as that of $`{\mathbf{I}}\hspace{0pt}{\mathbf{V}}`$ in Eq. ([2](#S2.E2 "In 2 The image encryption scheme to be studied ‣ Cryptanalysis of an image encryption scheme based on the Hill cipher")), the analysis about its insensitivity to the change of the plain-image can be carried out just like the case about the sub-key $`{\mathbf{I}}\hspace{0pt}{\mathbf{V}}`$ discussed above.
#### 3.1.4 Some other problems
The encryption scheme has the following additional problems:
1. 1\.
cannot encrypt plain-image of a fixed value zero;
2. 2\.
efficiency of implementation is low: From\[[20](#bib.bib20), Thorem 2.3.3\], one can see that the number of invertible matrices of size $`m \times m`$ in $`{\mathbb{Z}}_{256}`$ is
<table>
<tbody>
<tr>
<td></td>
<td>|GL(m,ℤ256)|=27m2∏k=0m−1(2m−2k).𝐺𝐿𝑚subscriptℤ256superscript27superscript𝑚2superscriptsubscriptproduct𝑘0𝑚1superscript2𝑚superscript2𝑘|GL(m,\mathbb{Z}_{256})|=2^{7m^{2}}\prod_{k=0}^{m-1}(2^{m}-2^{k}).</td>
<td></td>
<td>(11)</td>
</tr>
</tbody>
</table>
Thus, the probability that a matrix of size $`m \times m`$ in $`{\mathbb{Z}}_{256}`$ is invertible is
<table>
<tbody>
<tr>
<td></td>
<td>pm=27m2∏k=0m−1(2m−2k)28m2=∏k=1m(1−2−k)≈13.subscript𝑝𝑚superscript27superscript𝑚2superscriptsubscriptproduct𝑘0𝑚1superscript2𝑚superscript2𝑘superscript28superscript𝑚2superscriptsubscriptproduct𝑘1𝑚1superscript2𝑘13p_{m}=\frac{2^{7m^{2}}\prod_{k=0}^{m-1}(2^{m}-2^{k})}{2^{8m^{2}}}=\prod_{k=1}^{m}(1-2^{-k})\approx\frac{1}{3}.</td>
<td></td>
<td>(12)</td>
</tr>
</tbody>
</table>
So, it needs $`O\hspace{0pt}{({3\hspace{0pt}m^{2}})}`$ and $`O\hspace{0pt}{({{m^{2} \cdot M}\hspace{0pt}N})}`$ times of computations, respectively, for checking the reversibility of $`{\mathbf{K}}_{1}`$ and for calculating $`{\{{\mathbf{K}}_{l}^{- 1}\}}_{l = 1}^{\lceil{{M\hspace{0pt}N}/m}\rceil}`$. Note that these computations have no direct contributions to protecting the plain-image.
3. 3\.
the scope of sub-key $`m`$ is limited: As discussed above, the larger the value $`m`$ the higher the computational cost.
4. 4\.
the confusion capability is weak: This problem is caused by the linearity of the main encryption function. To demonstrate this defect, the encryption result of one special plain-image is shown in Fig. [3](#S3.F3 "Figure 3 ‣ 3.1.4 Some other problems ‣ 3.1 Some Defects of the Scheme ‣ 3 Cryptanalysis ‣ Cryptanalysis of an image encryption scheme based on the Hill cipher"), where Figure [3](#S3.F3 "Figure 3 ‣ 3.1.4 Some other problems ‣ 3.1 Some Defects of the Scheme ‣ 3 Cryptanalysis ‣ Cryptanalysis of an image encryption scheme based on the Hill cipher")b) also effectively disproves the conclusion about the quality of encryption results given in \[[18](#bib.bib18), Sec. 4\].
[Refer to caption](/html/0712.0693/assets/x9.png)
a\) plain-image
[Refer to caption](/html/0712.0693/assets/x10.png)
b\) cipher-image
Figure 3: A special test image, “Test_pattern”.
### 3.2 Known/Chosen-Plaintext Attack
The known/chosen-plaintext attack works by reconstructing the secret key or its equivalent based on some known/chosen plaintexts and their corresponding ciphertexts.
For this encryption scheme, the equivalent key $`{\{{\mathbf{K}}_{l}\}}_{l = 1}^{\lceil{{M\hspace{0pt}N}/m}\rceil}`$ can be reconstructed from $`m`$ plain-images $`{\mathbf{P}}^{(1)} \sim {\mathbf{P}}^{(m)}`$ and their corresponding cipher-images $`{\mathbf{C}}^{(1)} \sim {\mathbf{C}}^{(m)}`$ by using
<table>
<tbody>
<tr>
<td></td>
<td>𝑲l=(𝑷l(B)⋅(𝑪l(1)𝑪l(2)⋮𝑪l(m)))mod256,subscript𝑲𝑙modulo⋅subscriptsuperscript𝑷𝐵𝑙missing-subexpressionsubscriptsuperscript𝑪1𝑙missing-subexpressionsubscriptsuperscript𝑪2𝑙missing-subexpression⋮missing-subexpressionsubscriptsuperscript𝑪𝑚𝑙256\bm{K}_{l}=\left(\bm{P}^{(B)}_{l}\cdot\left(\begin{array}[]{cc}&\bm{C}^{(1)}_{l}\\ &\bm{C}^{(2)}_{l}\\ &\vdots\\ &\bm{C}^{(m)}_{l}\end{array}\right)\right)\bmod 256,</td>
<td></td>
<td>(13)</td>
</tr>
</tbody>
</table>
where
<table>
<tbody>
<tr>
<td></td>
<td>𝑷l(B)=(𝑷l(1)𝑷l(2)⋮𝑷l(m))−1.subscriptsuperscript𝑷𝐵𝑙superscriptmissing-subexpressionsubscriptsuperscript𝑷1𝑙missing-subexpressionsubscriptsuperscript𝑷2𝑙missing-subexpression⋮missing-subexpressionsubscriptsuperscript𝑷𝑚𝑙1\bm{P}^{(B)}_{l}=\left(\begin{array}[]{cc}&\bm{P}^{(1)}_{l}\\ &\bm{P}^{(2)}_{l}\\ &\vdots\\ &\bm{P}^{(m)}_{l}\end{array}\right)^{-1}.</td>
<td></td>
<td>(14)</td>
</tr>
</tbody>
</table>
The reversibility of $`{\mathbf{P}}_{l}^{(B)}`$ can be ensured by utilizing more than $`m`$ plain-images or by choosing $`m`$ special plain-images. Note that the above known/chosen-plaintext attack can be carried out with only one know/chosen plain-image due to the very short period of sequence $`{\{{{\mathbf{K}}_{l}\hspace{0pt}{\lbrack:,j\rbrack}}\}}_{l = 1}^{\lceil{{M\hspace{0pt}N}/m}\rceil}`$ for $`j = 1 \sim m`$. To study the period of this sequence, 10,000 tests have been done for a given value of $`{\mathbf{I}}\hspace{0pt}{\mathbf{V}}`$ of size $`1 \times 3`$, where $`{\mathbf{K}}_{1}`$ is selected randomly. The numbers of tests where the corresponding sequence $`{\{{{\mathbf{K}}_{l}\hspace{0pt}{(:,1)}}\}}_{l = 1}^{\lceil{{M\hspace{0pt}N}/m}\rceil}`$ has period $`p`$, $`N_{p}`$, with some values of $`{\mathbf{I}}\hspace{0pt}{\mathbf{V}}`$, is shown in Table [1](#S3.T1 "Table 1 ‣ 3.2 Known/Chosen-Plaintext Attack ‣ 3 Cryptanalysis ‣ Cryptanalysis of an image encryption scheme based on the Hill cipher"), which shows that the period of $`{\{{{\mathbf{K}}_{l}\hspace{0pt}{\lbrack:,j\rbrack}}\}}_{l = 1}^{\lceil{{M\hspace{0pt}N}/m}\rceil}`$ is indeed very short.
Table 1: Values of $`N_{p}`$ with some values of $`{\mathbf{I}}\hspace{0pt}{\mathbf{V}}`$, $`p = 2^{s}`$, $`s = 3 \sim 9`$.
<table>
<tbody>
<tr>
<th>𝑰𝑽𝑰𝑽\bm{IV}</th>
<td>N8subscript𝑁8N_{8}</td>
<td>N16subscript𝑁16N_{16}</td>
<td>N32subscript𝑁32N_{32}</td>
<td>N64subscript𝑁64N_{64}</td>
<td>N128subscript𝑁128N_{128}</td>
<td>N256subscript𝑁256N_{256}</td>
<td>N512subscript𝑁512N_{512}</td>
</tr>
<tr>
<th>(91, 63, 45)</th>
<td>0</td>
<td>0</td>
<td>0</td>
<td>0</td>
<td>0</td>
<td>1463</td>
<td>8537</td>
</tr>
<tr>
<th>(113, 25, 219)</th>
<td>14</td>
<td>34</td>
<td>127</td>
<td>561</td>
<td>3561</td>
<td>5703</td>
<td>0</td>
</tr>
<tr>
<th>(253, 115, 17)</th>
<td>6</td>
<td>20</td>
<td>72</td>
<td>284</td>
<td>1081</td>
<td>8537</td>
<td>0</td>
</tr>
<tr>
<th>(1, 3, 5)</th>
<td>0</td>
<td>0</td>
<td>98</td>
<td>284</td>
<td>1081</td>
<td>8537</td>
<td>0</td>
</tr>
<tr>
<th>(5, 121, 247)</th>
<td>7</td>
<td>36</td>
<td>132</td>
<td>561</td>
<td>3561</td>
<td>5703</td>
<td>0</td>
</tr>
</tbody>
</table>
## 4 Conclusion
In this paper, the security and performance of an image encryption scheme based on the Hill cipher have been analyzed in detail. It has been found that the scheme can be broken with only one known/chosen plain-image. There is a simple necessary and sufficient condition that makes a number of secret keys invalid. In addition, the scheme is insensitive to the change of the secret key/plain-image. Some other performance defects have also been found. In conclusion, the encryption scheme under study actually has much weaker security than the original Hill cipher, therefore is not recommended for applications.
## 5 Acknowledgement
This research was supported by the City University of Hong Kong under the SRG project 7002134.
## References
- \[1\] C. E. Shannon, Communication theory of secrecy systems, Bell System Technical Journal 28 (4) (1949) 656–715.
- \[2\]
S. Li, G. Chen, X. Zheng, Chaos-based encryption for digital images and videos, in: B. Furht, D. Kirovski (Eds.), Multimedia Security Handbook, CRC Press, LLC, 2004, Ch. 4, pp. 133–167, preprint available online at http://www.hooklee.com/pub.html.
- \[3\]
S. Li, Analyses and new designs of digital chaotic ciphers, Ph.D. thesis, School of Electronic and Information Engineering, Xi’an Jiaotong University, Xi’an, China, available online at http://www.hooklee.com/pub.html (2003).
- \[4\]
C. Li, Cryptanalyses of some multimedia encryption schemes, Master’s thesis, Department of Mathematics, Zhejiang University, Hangzhou, China, available online at http://eprint.iacr.org/2006/340 (May 2005).
- \[5\]
C. Li, S. Li, D. Zhang, G. Chen, Cryptanalysis of a chaotic neural network based multimedia encryption scheme, Lecture Notes in Computer Science 3333 (2004) 418–425.
- \[6\] C. Li, S. Li, G. Chen, G. Chen, L. Hu, Cryptanalysis of a new signal security system for multimedia data transmission, EURASIP Journal on Applied Signal Processing 2005 (8) (2005) 1277–1288.
- \[7\]
C. Li, S. Li, D. Zhang, G. Chen, Chosen-plaintext cryptanalysis of a clipped-neural-network-based chaotic cipher, Lecture Notes in Computer Science 3497 (2005) 630–636.
- \[8\] C. Li, S. Li, D.-C. Lou, On the security of the Yen-Guo’s domino signal encryption algorithm (DSEA), Elsevier Journal of Systems and Software 79 (2) (2006) 253–258.
- \[9\] S. Li, C. Li, K.-T. Lo, G. Chen, Cryptanalysis of an image encryption scheme, Journal of Electronic Imaging 15 (4) (2006) article number 043012.
- \[10\] C. Li, S. Li, G. Álvarez, G. Chen, K.-T. Lo, Cryptanalysis of two chaotic encryption schemes based on circular bit shift and xor operations, Physics Letters A 369 (1-2) (2007) 23–30.
- \[11\] G. Alvarez, S. Li, Some basic cryptographic requirements for chaos-based cryptosystems, International Journal of Bifurcation and Chaos 16 (8) (2006) 2129–2151.
- \[12\]
D. Arroyo, C. Li, S. Li, G. Alvarez, Cryptanalysis of a computer cryptography scheme based on a filter bank, available online at http://arxiv.org/abs/0710.5471 (2007).
- \[13\]
C. Li, S. Li, M. Asim, J. Nunez, G. Álvarez, G. Chen, On the security defects of an image encryption scheme, Cryptology ePrint Archive: Report 2007/397, available online at http://eprint.iacr.org/2007/397 (2007).
- \[14\]
S. Li, C. Li, K.-T. Lo, G. Chen, Cryptanalysis of an image scrambling scheme without bandwidth expansion, accepted by IEEE Transactions on Circuits and Systems for Video Technology, available online at http://eprint.iacr.org/2006/215 (2007).
- \[15\] J. Zhou, Z. Liang, Y. Chen, A. O. C., Security analysis of multimedia encryption schemes based on multiple huffman table, IEEE Signal Processing Letters 14 (3) (2007) 201–204.
- \[16\] S. Li, G. Chen, A. Cheung, B. Bhargava, K.-T. Lo, On the design of perceptual MPEG-video encryption algorithms, IEEE Transactions on Circuits and Systems for Video Technology 17 (2) (2007) 214–223.
- \[17\] G. Alvarez, S. Li, L. Hernandez, Analysis of security problems in a medical image encryption system, Computers in Biology and Medicine 37 (3) (2007) 424–427.
- \[18\] I. A. Ismail, M. Amin, H. Diab, How to repair the hill cipher, Journal of Zhejiang University SCIENCE A 7 (12) (2006) 2022–2030.
- \[19\] L. S. Hill, Cryptography in an algebraic alphabet, The American Mathematical Monthly 36 (1929) 306–312.
- \[20\] J. Overbey, W. Traves, J. Wojdylo, On the keyspace of the hill cipher, Cryptologia 29 (1) (2005) 59–72. |
0711.4562 | cs.NI | cyber | # Near-Deterministic Inference of AS Relationships
Yuval Shavitt
Eran Shir
Udi Weinsberg
School of Electrical Engineering Tel-Aviv University, Israel <shavitt@eng.tau.ac.il> School of Electrical Engineering Tel-Aviv University, Israel <shire@eng.tau.ac.il> School of Electrical Engineering Tel-Aviv University, Israel <udiw@eng.tau.ac.il>
###### Abstract
The discovery of Autonomous Systems (ASes) interconnections and the inference of their commercial Type-of-Relationships (ToR) has been extensively studied during the last few years. The main motivation is to accurately calculate AS-level paths and to provide better topological view of the Internet. An inherent problem in current algorithms is their extensive use of heuristics. Such heuristics incur unbounded errors which are spread over all inferred relationships. We propose a near-deterministic algorithm for solving the ToR inference problem. Our algorithm uses as input the Internet core, which is a dense sub-graph of top-level ASes. We test several methods for creating such a core and demonstrate the robustness of the algorithm to the core’s size and density, the inference period, and errors in the core.
We evaluate our algorithm using AS-level paths collected from RouteViews BGP paths and DIMES traceroute measurements. Our proposed algorithm deterministically infers over 95% of the approximately 58,000 AS topology links. The inference becomes stable when using a week worth of data and as little as 20 ASes in the core. The algorithm infers 2$`\sim`$3 times more peer-to-peer relationships in edges discovered only by DIMES than in RouteViews edges, validating the DIMES promise to discover periphery AS edges.
## 1 Introduction
Today’s Internet consists of thousands of networks administrated by various Autonomous Systems (ASes). ASes are assigned with one or more blocks of IP prefixes and communicate routing information to each other using Border Gateway Protocol (BGP). Each AS uses a set of local policies for selecting the best route for each reachable prefix. Typically, these policies are based on the Type-of-Relationship (ToR) that exists between ASes and on a shortest path criteria. In order to calculate the paths between ASes, one needs to obtain the ToR between all neighboring ASes. Since ToRs are regarded as proprietary information, deducing them is an important yet difficult problem.
Typically \[[17](#bib.bib17)\], there are three major commercial relationships between neighboring ASes: customer-to-provider (c2p), peer-to-peer (p2p), and sibling-to-sibling (s2s). In the c2p category, a customer AS pays a provider AS (usually larger than the customer) for traffic that is sent between the two. In the p2p category, two ASes freely exchange traffic between themselves and their customers, but do not exchange traffic from or to their providers or other peers. In s2s, two ASes administratively belong to the same organization and freely exchange traffic between their providers, customers, peers, or other siblings.
Gao \[[12](#bib.bib12)\] was the first to study the AS relationships inference problem and deduced that every BGP path must comply with the following hierarchical pattern: an uphill segment of zero or more c2p or s2s links, followed by zero or one p2p links, followed by a downhill segment of zero or more p2c or s2s links. Paths with this hierarchical structure are called *valley-free* or valid. Paths that do not follow this hierarchical structure are called invalid and may result from BGP misconfigurations or from BGP policies that are more complex and do not distinctly fall into the above classification. Most work in this field (section [2](#S2 "2 Related Work ‣ Near-Deterministic Inference of AS Relationships")) follows the valley free routing principle.
Current relationships inference algorithms attempt to solve the ToR problem either by using heuristic assumptions or by optimizing some aspects of the ToR assignments. Optimization is usually achieved by minimizing the number of paths that violate the valley free routing property \[[21](#bib.bib21)\] while not allowing cycles to be created \[[8](#bib.bib8), [18](#bib.bib18)\] in the resulting directed relationships graph.
Using heuristic assumptions throughout the relationships inference process causes the erroneous ToRs to be spread over all interconnecting ASes links. The optimization models fail to capture the true Internet hierarchy \[[10](#bib.bib10)\] and have a relatively low p2p inference accuracy \[[23](#bib.bib23)\]. The result is that both solutions fail to provide an insight, or a bound on the inference errors.
Typically, AS relationships are not published by AS operators, hence the validation of such results is done either by sending queries to the operators of a small subset of ASes \[[9](#bib.bib9)\] or by comparing the results to partial information that is available on the Internet \[[23](#bib.bib23)\]. Although these methods can give a good approximation on the correctness of the results, one cannot assume a bounded mistake.
This paper aims to improve on existing methods by providing a near-deterministic inference scheme for solving the ToR problem. The input for our algorithm is the Internet *Core*, a sub-graph that consists of the globally top-level providers of the Internet graph and their interconnecting edges with their already inferred relationship types. Theoretically, given an accurate core with no relationships errors, the algorithm *deterministically* infers most of the remaining AS relationships using the AS-level paths relative to this core, without incurring additional inference errors. In real-world scenarios, where the core and AS-level paths can contain errors (due to misconfigurations or measurements mistakes), the algorithm introduces minimal inference mistakes. The core can be approximated in several ways, as described in section [3.4](#S3.SS4 "3.4 Core Graph Construction ‣ 3 AS Relationships Inference ‣ Near-Deterministic Inference of AS Relationships"), or extracted from public databases. We show that our algorithm has relaxed requirements from the core, and proves to be robust under changes in its definition, size and density. Since the top-level ASes are a small and stable group, accurately revealing the core members and their mutual types of relationships is fairly easy. For the remaining set of relationships that cannot be inferred deterministically, a heuristic inference method is deployed. Since this group is relatively small, it is possible to provide a strict bound on the inference error. In order to increase the number of vantage points from which we see the Internet, we use both RouteViews (RV) \[[2](#bib.bib2)\] BGP data and DIMES \[[20](#bib.bib20)\] AS-level traceroutes. We expect that over time, the group of non-deterministic inferred relationships will even further decrease.
The remaining of this paper is organized as follows. Section [2](#S2 "2 Related Work ‣ Near-Deterministic Inference of AS Relationships") provides several related works concerning AS relationships inference. Section [3](#S3 "3 AS Relationships Inference ‣ Near-Deterministic Inference of AS Relationships") provides a detailed description of our deterministic inference algorithm and discusses the methods used to infer on the remaining unclassified edges. Section [4](#S4 "4 Experimental Results ‣ Near-Deterministic Inference of AS Relationships") provides a detailed evaluation of the proposed algorithms and Section [5](#S5 "5 Conclusion and Future Work ‣ Near-Deterministic Inference of AS Relationships") concludes the paper and discusses future work.
## 2 Related Work
As mentioned above, Gao’s pioneering work \[[12](#bib.bib12)\] was the first to study the AS relationships inference problem. Gao proposed an inference heuristic that identified top providers and peering links based on AS size, which is proportional to its degree (the number of immediate neighbors of a vertex), and the valley-free nature of routing paths. Gao used this heuristic to infer relationship between ASes in the Internet by traversing advertised BGP routes, locally identifying the top provider for each path, and classifying edges (i.e., inferring the relationships represented by the edges) as going uphill to the top provider and downhill afterwards. Xia and Gao \[[23](#bib.bib23)\] later proposed to use partially available information regarding AS relationships in order infer the unknown relations. It is not clear that this information can be obtained and validated periodically, unlike our suggestion to use the almost constant relationships in the Internet core. Our sole reliance on the core produces simpler inference rules that are less prone to inference errors.
Following Gao’s work, Subramanian et al. \[[21](#bib.bib21)\] formally defined the Type-of-Relation (*ToR*) maximization problem that attempts to maximize the number of valid (valley-free) routing paths for a given AS graph. Their approach takes as input the BGP tables collected at different vantage points and computes a rank for every AS. This rank is a measure of how close to the graph core an AS lies (equivalent to vertex coreness \[[3](#bib.bib3)\]), and is heuristically used to infer AS relationships by comparing ranks of adjacent ASes. If the ranks are similar, the algorithm classifies the link as p2p, otherwise it is classified as c2p or p2c.
Battista et al. \[[6](#bib.bib6)\] showed that the decision version of the ToR problem (*ToR-D*) is an NP-complete problem in the general case. Motivated by the hardness of the general problem, they proposed approximation algorithms and reduced the ToR-D problem to a 2SAT formula by mapping any two adjacent edges in all input AS-level routing paths into a clause with two literals, while adding heuristics based inference.
Dimitropoulos et al. \[[10](#bib.bib10)\] addressed a problem in current ToR algorithms. They showed that although ToR algorithms produce a directed Internet graph with a very small number of invalid paths, the resulting AS relationships are far from reality. This led them to the conclusion that simply trying to maximize the number of valid paths (namely improving the result of the ToR algorithms) does not produce realistic results. Later in \[[9](#bib.bib9)\] they showed that ToR has no means to deterministically select the most realistic solution when facing multiple possible solutions. In order to solve this problem, the authors suggested a new objective function by adding a notion of "AS importance", which is the AS degree "gradient" in the original undirected Internet graph. The modified ToR algorithm directs the edges from low importance AS to a higher one. The authors showed that although they have high success rate in p2c inference (96.5%) and in s2s inference (90.3%), the p2p inference success rate (82.8%) is relatively low. Moreover, the authors surveyed some ASes operators and mention that for some of them, the BGP tables, which are the source for AS-level routing paths for most works in this research field, miss up to 86.2% of the true relationships between adjacent ASes, most of which are of p2p type.
These observations match the evaluation work done in \[[23](#bib.bib23)\], and highly motivate our work, driving us not only to seek an algorithm that better captures the true AS relationships in the Internet while reducing the usage of heuristics for inference, but also to add a different, complementary data source for routing paths, that has the ability to capture much of the missing links.
## 3 AS Relationships Inference
In this section we describe our ToR inference algorithm in details. We start with explaining the deterministic algorithm, and proceed with the heuristics we employ for edges that the deterministic algorithm fails to classify.
### 3.1 Deterministic Classification
Our deterministic algorithm receives as input two undirected AS-level graphs and a set of AS-level routing paths, denoted by $`S`$. The first graph, denoted by $`G\hspace{0pt}{(V_{G},E_{G})}`$, contains the set of vertices that represent all ASes, and the interconnecting edges that need to be classified. The second graph, denoted by $`C\hspace{0pt}o\hspace{0pt}r\hspace{0pt}e\hspace{0pt}{(V_{C},E_{C})}`$, holds the vertices and interconnecting edges that represent the core of $`G`$, and is assumed to contain all the top-level ASes.
Prior to starting the relationships inference algorithm, we infer s2s relationships, since ignoring these relationships might cause proliferation of erroneous inference \[[9](#bib.bib9)\]. We use s2s data collected from \[[1](#bib.bib1)\]. These s2s classifications are obtained from IRR databases, namely RIPE, ARIN and APNIC. Although these databases are not always up-to-date, they are reasonably steady and accurate for the s2s inference. Once classified, the s2s edges are removed from the edges set $`E_{G}`$, and the two adjacent vertices are united to form a single vertex that inherits the connectivity of both.
Following the assumption that the input core consists of all the global top-level ASes and using the valley-free model of Internet routing, the algorithm classifies most of the edges in $`G\hspace{0pt}{(V_{G},E_{G})}`$ without using heuristic assumptions:
Phase 1. All paths that *pass through the core* are split into a segment of zero or more uphill c2p edges towards the core, at most one p2p edge in the core and a downhill segment of zero or more p2c edges from the core. The algorithm, shown in Alg. [1](#alg1 "Algorithm 1 ‣ 3.1 Deterministic Classification ‣ 3 AS Relationships Inference ‣ Near-Deterministic Inference of AS Relationships"), traverses only paths that pass through the core. It starts with the uphill segment of the path, classifying each edge as c2p, until reaching the core. Once reached the core, the uphill segment finishes and the core segment starts. Inside the core the algorithm classifies edges that are not already classified as p2p (the default type for core edges). The downhill segment starts with an AS that does not belong to the core, and is traversed until the end of the path. Invalid paths are detected when an edge is directed towards the core (uphill) during the downhill segment. Each path that is classified in this phase is removed from the set of paths $`S`$. Note that the algorithm does not use direct inference but a voting technique (see section [3.2](#S3.SS2 "3.2 Assigning Type-of-Relationship to Edges ‣ 3 AS Relationships Inference ‣ Near-Deterministic Inference of AS Relationships")) in order to resolve ambiguities resulted from incorrect paths.
[Refer to caption](/html/0711.4562/assets/x1.png)
\(a\)
Phase 1
[Refer to caption](/html/0711.4562/assets/x2.png)
\(b\)
Phase 2
Figure 1: Deterministic ToR inference algorithm
An example for edges that are classified during this phase is illustrated in Fig. [1(a)](#S3.F1.sf1 "In Figure 1 ‣ 3.1 Deterministic Classification ‣ 3 AS Relationships Inference ‣ Near-Deterministic Inference of AS Relationships"). Path *P* traverses the core and consists of seven AS hops, numbered 1 to 7. The segment between hop 1 and hop 4 is identified as an uphill segment, resulting in the classification of the edges (1,2), (2,3), and (3,4) as c2p (illustrated as an arrow pointed from the customer to the provider). ASes 4 and 5 are inside the core, therefore are already classified (by default to p2p). The segment starting in AS 5 and ending in AS 7 is considered as a downhill segment, resulting in the classification of the edges (5,6) and (6,7) as p2c.
1:
Graphs $`G\hspace{0pt}\left( V_{G},E_{G} \right)`$, $`{C\hspace{0pt}o\hspace{0pt}r\hspace{0pt}e\hspace{0pt}\left( V_{C},E_{C} \right)} \subset G`$, Paths set $`S`$
2:
Edges $`E_{G}`$ with votes for relationship types
3:
foreach $`{p\hspace{0pt}a\hspace{0pt}t\hspace{0pt}h} \in S`$ do
4:
if $`{\exists e} \in {{p\hspace{0pt}a\hspace{0pt}t\hspace{0pt}h} \mid e} \in {E_{C}\hspace{0pt}\text{or}\hspace{0pt}{\exists v}} \in {{p\hspace{0pt}a\hspace{0pt}t\hspace{0pt}h} \mid v} \in V_{C}`$ then
5:
$`{u\hspace{0pt}p\hspace{0pt}H\hspace{0pt}i\hspace{0pt}l\hspace{0pt}l}\leftarrow{T\hspace{0pt}R\hspace{0pt}U\hspace{0pt}E}`$
6:
$`{d\hspace{0pt}o\hspace{0pt}w\hspace{0pt}n\hspace{0pt}H\hspace{0pt}i\hspace{0pt}l\hspace{0pt}l}\leftarrow{F\hspace{0pt}A\hspace{0pt}L\hspace{0pt}S\hspace{0pt}E}`$
7:
$`{i\hspace{0pt}n\hspace{0pt}C\hspace{0pt}o\hspace{0pt}r\hspace{0pt}e}\leftarrow{F\hspace{0pt}A\hspace{0pt}L\hspace{0pt}S\hspace{0pt}E}`$
8:
foreach $`{e\hspace{0pt}d\hspace{0pt}g\hspace{0pt}e} \in {P\hspace{0pt}a\hspace{0pt}t\hspace{0pt}h}`$ do
9:
$`{{A\hspace{0pt}S\hspace{0pt}1}\leftarrow{e\hspace{0pt}d\hspace{0pt}g\hspace{0pt}e}}.{f\hspace{0pt}i\hspace{0pt}r\hspace{0pt}s\hspace{0pt}t\hspace{0pt}A\hspace{0pt}S}`$
10:
$`{{A\hspace{0pt}S\hspace{0pt}2}\leftarrow{e\hspace{0pt}d\hspace{0pt}g\hspace{0pt}e}}.{s\hspace{0pt}e\hspace{0pt}c\hspace{0pt}o\hspace{0pt}n\hspace{0pt}d\hspace{0pt}A\hspace{0pt}S}`$
11:
if $`{e\hspace{0pt}d\hspace{0pt}g\hspace{0pt}e} \in E_{C}`$ then
12:
$`{u\hspace{0pt}p\hspace{0pt}H\hspace{0pt}i\hspace{0pt}l\hspace{0pt}l}\leftarrow{F\hspace{0pt}A\hspace{0pt}L\hspace{0pt}S\hspace{0pt}E}`$
13:
$`{i\hspace{0pt}n\hspace{0pt}C\hspace{0pt}o\hspace{0pt}r\hspace{0pt}e}\leftarrow{T\hspace{0pt}R\hspace{0pt}U\hspace{0pt}E}`$
14:
else if $`{A\hspace{0pt}S\hspace{0pt}1} \in {V_{C}\hspace{0pt}{\mathbf{a}\mathbf{n}\mathbf{d}}\hspace{0pt}A\hspace{0pt}S\hspace{0pt}2} \notin V_{C}`$ then
15:
$`{u\hspace{0pt}p\hspace{0pt}H\hspace{0pt}i\hspace{0pt}l\hspace{0pt}l}\leftarrow{F\hspace{0pt}A\hspace{0pt}L\hspace{0pt}S\hspace{0pt}E}`$
16:
$`{i\hspace{0pt}n\hspace{0pt}C\hspace{0pt}o\hspace{0pt}r\hspace{0pt}e}\leftarrow{F\hspace{0pt}A\hspace{0pt}L\hspace{0pt}S\hspace{0pt}E}`$
17:
$`{d\hspace{0pt}o\hspace{0pt}w\hspace{0pt}n\hspace{0pt}H\hspace{0pt}i\hspace{0pt}l\hspace{0pt}l}\leftarrow{T\hspace{0pt}R\hspace{0pt}U\hspace{0pt}E}`$
18:
else if $`{d\hspace{0pt}o\hspace{0pt}w\hspace{0pt}n\hspace{0pt}H\hspace{0pt}i\hspace{0pt}l\hspace{0pt}l\hspace{0pt}{\mathbf{a}\mathbf{n}\mathbf{d}}\hspace{0pt}A\hspace{0pt}S\hspace{0pt}2} \in V_{C}`$ then
19:
$`v\hspace{0pt}o\hspace{0pt}t\hspace{0pt}e\hspace{0pt}F\hspace{0pt}o\hspace{0pt}r\hspace{0pt}I\hspace{0pt}n\hspace{0pt}v\hspace{0pt}a\hspace{0pt}l\hspace{0pt}i\hspace{0pt}d\hspace{0pt}\left( {e\hspace{0pt}d\hspace{0pt}g\hspace{0pt}e} \right)`$
20:
end if
21:
if $`u\hspace{0pt}p\hspace{0pt}H\hspace{0pt}i\hspace{0pt}l\hspace{0pt}l`$ then
22:
$`v\hspace{0pt}o\hspace{0pt}t\hspace{0pt}e\hspace{0pt}F\hspace{0pt}o\hspace{0pt}r\hspace{0pt}C\hspace{0pt}u\hspace{0pt}s\hspace{0pt}t\hspace{0pt}o\hspace{0pt}m\hspace{0pt}e\hspace{0pt}r\hspace{0pt}T\hspace{0pt}o\hspace{0pt}P\hspace{0pt}r\hspace{0pt}o\hspace{0pt}v\hspace{0pt}i\hspace{0pt}d\hspace{0pt}e\hspace{0pt}r\hspace{0pt}\left( {e\hspace{0pt}d\hspace{0pt}g\hspace{0pt}e} \right)`$
23:
else if $`i\hspace{0pt}n\hspace{0pt}C\hspace{0pt}o\hspace{0pt}r\hspace{0pt}e\hspace{0pt}\text{and}\hspace{0pt}n\hspace{0pt}o\hspace{0pt}t\hspace{0pt}C\hspace{0pt}l\hspace{0pt}a\hspace{0pt}s\hspace{0pt}s\hspace{0pt}i\hspace{0pt}f\hspace{0pt}i\hspace{0pt}e\hspace{0pt}d\hspace{0pt}\left( {e\hspace{0pt}d\hspace{0pt}g\hspace{0pt}e} \right)`$ then
24:
$`v\hspace{0pt}o\hspace{0pt}t\hspace{0pt}e\hspace{0pt}F\hspace{0pt}o\hspace{0pt}r\hspace{0pt}P\hspace{0pt}e\hspace{0pt}e\hspace{0pt}r\hspace{0pt}T\hspace{0pt}o\hspace{0pt}P\hspace{0pt}e\hspace{0pt}e\hspace{0pt}r\hspace{0pt}\left( {e\hspace{0pt}d\hspace{0pt}g\hspace{0pt}e} \right)`$
25:
else
26:
$`v\hspace{0pt}o\hspace{0pt}t\hspace{0pt}e\hspace{0pt}F\hspace{0pt}o\hspace{0pt}r\hspace{0pt}P\hspace{0pt}r\hspace{0pt}o\hspace{0pt}v\hspace{0pt}i\hspace{0pt}d\hspace{0pt}e\hspace{0pt}r\hspace{0pt}T\hspace{0pt}o\hspace{0pt}C\hspace{0pt}u\hspace{0pt}s\hspace{0pt}t\hspace{0pt}o\hspace{0pt}m\hspace{0pt}e\hspace{0pt}r\hspace{0pt}\left( {e\hspace{0pt}d\hspace{0pt}g\hspace{0pt}e} \right)`$
27:
end if
28:
end for
29:
$`S\leftarrow{S \smallsetminus {p\hspace{0pt}a\hspace{0pt}t\hspace{0pt}h}}`$
30:
end if
31:
end for
Algorithm 1
Phase 1 of ToR Inference Algorithm
Since the remaining paths in $`S`$ do not traverse the core, they do not provide us with a direct method for classification. However, amongst these, there are paths that partly overlap other paths that traverse the core. Meaning that some of the remaining paths already contain edges that were classified as either c2p or p2c in the first phase of the algorithm. We use these edges for the second phase of the algorithm:
Phase 2. For a given path, edges that precede a c2p edge must reside in an uphill segment, and be of type c2p. Edges that follow a p2c edge must be in a downhill segment, and be of type p2c. The algorithm, listed in Alg. [2](#alg2 "Algorithm 2 ‣ 3.1 Deterministic Classification ‣ 3 AS Relationships Inference ‣ Near-Deterministic Inference of AS Relationships"), traverses one path at a time, and looks for an already inferred c2p or p2c edges. If a c2p edge is detected, all unclassified edges in the path before this edge, temporarily stored in the *suspectC2P* list, are classified as c2p. If a p2c edge is detected, all unclassified edges in the path after this edge, temporarily stored in the *suspectP2C* list, are classified as p2c.
Since this phase uses classified edges in order to classify unclassified edges, it is repeated for all paths in $`S`$ that still have unclassified edges, until there are no more edges that can be classified using this method.
1:
Graph $`G\hspace{0pt}\left( V_{G},E_{G} \right)`$, Remaining set of paths $`S`$
2:
Edges $`E_{G}`$ with votes for relationship types
3:
foreach $`{p\hspace{0pt}a\hspace{0pt}t\hspace{0pt}h} \in S`$ do
4:
$`{s\hspace{0pt}u\hspace{0pt}s\hspace{0pt}p\hspace{0pt}e\hspace{0pt}c\hspace{0pt}t\hspace{0pt}C\hspace{0pt}2\hspace{0pt}P}\leftarrow\varnothing`$
5:
$`{s\hspace{0pt}u\hspace{0pt}s\hspace{0pt}p\hspace{0pt}e\hspace{0pt}c\hspace{0pt}t\hspace{0pt}P\hspace{0pt}2\hspace{0pt}C}\leftarrow\varnothing`$
6:
$`{p\hspace{0pt}a\hspace{0pt}s\hspace{0pt}s\hspace{0pt}e\hspace{0pt}d\hspace{0pt}P\hspace{0pt}2\hspace{0pt}C}\leftarrow{F\hspace{0pt}A\hspace{0pt}L\hspace{0pt}S\hspace{0pt}E}`$
7:
foreach $`{e\hspace{0pt}d\hspace{0pt}g\hspace{0pt}e} \in {P\hspace{0pt}a\hspace{0pt}t\hspace{0pt}h}`$ do
8:
if $`{m\hspace{0pt}a\hspace{0pt}x\hspace{0pt}V\hspace{0pt}o\hspace{0pt}t\hspace{0pt}e\hspace{0pt}s\hspace{0pt}C\hspace{0pt}2\hspace{0pt}P\hspace{0pt}\left( {e\hspace{0pt}d\hspace{0pt}g\hspace{0pt}e} \right)\hspace{0pt}{\mathbf{a}\mathbf{n}\mathbf{d}}\hspace{0pt}s\hspace{0pt}u\hspace{0pt}s\hspace{0pt}p\hspace{0pt}e\hspace{0pt}c\hspace{0pt}t\hspace{0pt}C\hspace{0pt}2\hspace{0pt}P} \neq \varnothing`$ then
9:
foreach $`e \in {s\hspace{0pt}u\hspace{0pt}s\hspace{0pt}p\hspace{0pt}e\hspace{0pt}c\hspace{0pt}t\hspace{0pt}C\hspace{0pt}2\hspace{0pt}P}`$ do
10:
$`v\hspace{0pt}o\hspace{0pt}t\hspace{0pt}e\hspace{0pt}F\hspace{0pt}o\hspace{0pt}r\hspace{0pt}C\hspace{0pt}u\hspace{0pt}s\hspace{0pt}t\hspace{0pt}o\hspace{0pt}m\hspace{0pt}e\hspace{0pt}r\hspace{0pt}T\hspace{0pt}o\hspace{0pt}P\hspace{0pt}r\hspace{0pt}o\hspace{0pt}v\hspace{0pt}i\hspace{0pt}d\hspace{0pt}e\hspace{0pt}r\hspace{0pt}\left( {e\hspace{0pt}d\hspace{0pt}g\hspace{0pt}e} \right)`$
11:
end for
12:
$`{s\hspace{0pt}u\hspace{0pt}s\hspace{0pt}p\hspace{0pt}e\hspace{0pt}c\hspace{0pt}t\hspace{0pt}C\hspace{0pt}2\hspace{0pt}P}\leftarrow\varnothing`$
13:
else if $`m\hspace{0pt}a\hspace{0pt}x\hspace{0pt}V\hspace{0pt}o\hspace{0pt}t\hspace{0pt}e\hspace{0pt}s\hspace{0pt}P\hspace{0pt}2\hspace{0pt}C\hspace{0pt}\left( {e\hspace{0pt}d\hspace{0pt}g\hspace{0pt}e} \right)`$ then
14:
$`{s\hspace{0pt}u\hspace{0pt}s\hspace{0pt}p\hspace{0pt}e\hspace{0pt}c\hspace{0pt}t\hspace{0pt}C\hspace{0pt}2\hspace{0pt}P}\leftarrow\varnothing`$
15:
$`{p\hspace{0pt}a\hspace{0pt}s\hspace{0pt}s\hspace{0pt}e\hspace{0pt}d\hspace{0pt}P\hspace{0pt}2\hspace{0pt}C}\leftarrow{T\hspace{0pt}R\hspace{0pt}U\hspace{0pt}E}`$
16:
end if
17:
if $`n\hspace{0pt}o\hspace{0pt}C\hspace{0pt}l\hspace{0pt}a\hspace{0pt}s\hspace{0pt}s\hspace{0pt}i\hspace{0pt}f\hspace{0pt}i\hspace{0pt}c\hspace{0pt}a\hspace{0pt}t\hspace{0pt}i\hspace{0pt}o\hspace{0pt}n\hspace{0pt}V\hspace{0pt}o\hspace{0pt}t\hspace{0pt}e\hspace{0pt}s\hspace{0pt}\left( {e\hspace{0pt}d\hspace{0pt}g\hspace{0pt}e} \right)`$ then
18:
if $`{p\hspace{0pt}a\hspace{0pt}s\hspace{0pt}s\hspace{0pt}e\hspace{0pt}d\hspace{0pt}P\hspace{0pt}2\hspace{0pt}C} = {F\hspace{0pt}A\hspace{0pt}L\hspace{0pt}S\hspace{0pt}E}`$ then
19:
$`{s\hspace{0pt}u\hspace{0pt}s\hspace{0pt}p\hspace{0pt}e\hspace{0pt}c\hspace{0pt}t\hspace{0pt}C\hspace{0pt}2\hspace{0pt}P}\leftarrow{{s\hspace{0pt}u\hspace{0pt}s\hspace{0pt}p\hspace{0pt}e\hspace{0pt}c\hspace{0pt}t\hspace{0pt}C\hspace{0pt}2\hspace{0pt}P} \cup {e\hspace{0pt}d\hspace{0pt}g\hspace{0pt}e}}`$
20:
else
21:
$`{s\hspace{0pt}u\hspace{0pt}s\hspace{0pt}p\hspace{0pt}e\hspace{0pt}c\hspace{0pt}t\hspace{0pt}P\hspace{0pt}2\hspace{0pt}C}\leftarrow{{s\hspace{0pt}u\hspace{0pt}s\hspace{0pt}p\hspace{0pt}e\hspace{0pt}c\hspace{0pt}t\hspace{0pt}P\hspace{0pt}2\hspace{0pt}C} \cup {e\hspace{0pt}d\hspace{0pt}g\hspace{0pt}e}}`$
22:
end if
23:
end if
24:
end for
25:
if $`{s\hspace{0pt}u\hspace{0pt}s\hspace{0pt}p\hspace{0pt}e\hspace{0pt}c\hspace{0pt}t\hspace{0pt}P\hspace{0pt}2\hspace{0pt}C} ≢ \varnothing`$ then
26:
foreach $`e \in {s\hspace{0pt}u\hspace{0pt}s\hspace{0pt}p\hspace{0pt}e\hspace{0pt}c\hspace{0pt}t\hspace{0pt}P\hspace{0pt}2\hspace{0pt}C}`$ do
27:
$`v\hspace{0pt}o\hspace{0pt}t\hspace{0pt}e\hspace{0pt}F\hspace{0pt}o\hspace{0pt}r\hspace{0pt}P\hspace{0pt}r\hspace{0pt}o\hspace{0pt}v\hspace{0pt}i\hspace{0pt}d\hspace{0pt}e\hspace{0pt}r\hspace{0pt}T\hspace{0pt}o\hspace{0pt}C\hspace{0pt}u\hspace{0pt}s\hspace{0pt}t\hspace{0pt}o\hspace{0pt}m\hspace{0pt}e\hspace{0pt}r\hspace{0pt}\left( {e\hspace{0pt}d\hspace{0pt}g\hspace{0pt}e} \right)`$
28:
end for
29:
end if
30:
end for
Algorithm 2
Phase 2 of Classification Algorithm
Fig. [1(b)](#S3.F1.sf2 "In Figure 1 ‣ 3.1 Deterministic Classification ‣ 3 AS Relationships Inference ‣ Near-Deterministic Inference of AS Relationships") illustrates an example for a path that contains edges that will be classified during Phase 2 of the algorithm. Path *P* does not traverse the core but contain edges that are already classified. The edge (1,2) precedes the edge (2,3) which is classified as c2p, therefore is classified as c2p. The edge (6,7) follows a p2c edge, therefore is classified as p2c. The edges (3,4) and (4,5) cannot be classified, since we are unable to determine which AS is the top-level provider and whether the edges (3,4) and (4,5) represent a p2p relationship.
### 3.2 Assigning Type-of-Relationship to Edges
The data we use might be "noisy" and reflect transient routing effects or changes in the commercial relationships between ASes, especially when performing relationships inference over a long time frame. To avoid incorrect inferences resulting from these effects, we use voting technique \[[12](#bib.bib12)\] instead of direct relationship inference. Meaning, that the above methods vote for the ToR of each traversed edge. Once the algorithm is finished, we count the votes and assign each edge with the type that received a relative votes count that passes a given threshold.
Note that the voting technique is not used as a heuristic inference method, but rather as a method to avoid inference bias caused by incorrect or transient paths. The deterministic algorithm does not attempt to classify edges that have a vote count that does not pass the given threshold, but rather it employs the non-deterministic methods discussed in section [3.3](#S3.SS3 "3.3 Non-Deterministic Inference of Remaining Relationships ‣ 3 AS Relationships Inference ‣ Near-Deterministic Inference of AS Relationships"). Setting the exact value of the threshold, in order to achieve these goals is further discussed in section [4.2](#S4.SS2 "4.2 Voting Threshold ‣ 4 Experimental Results ‣ Near-Deterministic Inference of AS Relationships").
### 3.3 Non-Deterministic Inference of Remaining Relationships
The deterministic algorithm fails to classify several types of edges. The first type are edges that appear in paths that do not traverse the core, and reside between a c2p edge and a p2c edge (see Fig. [2(a)](#S3.F2.sf1 "In Figure 2 ‣ 3.3 Non-Deterministic Inference of Remaining Relationships ‣ 3 AS Relationships Inference ‣ Near-Deterministic Inference of AS Relationships")). This can be a result of a path that does not traverse the core and has no overlapping edges with other paths, or overlaps other paths in edges that are close to the beginning or end of the path. Alternatively, the path may have a p2p relationship between its two top-level vertices.
[Refer to caption](/html/0711.4562/assets/x3.png)
\(a\)
Peer-to-Peer
[Refer to caption](/html/0711.4562/assets/x4.png)
\(b\)
Non-Valley-Free
Figure 2: Non-Deterministic ToR inferences
In order to infer relationships related to these edges we use the following assumption (which is backed from observations, as we show in section [4.5](#S4.SS5 "4.5 Analysis of Non-Deterministically Inferred Relationships ‣ 4 Experimental Results ‣ Near-Deterministic Inference of AS Relationships")): a c2p or p2c edges should participate in, at least, one path that pass through the core. For this not to happen the following should occur: 1) a client will rarely route through a certain edge to its provider and thus may not expose this link in a DIMES measurement path that passes through the core, and 2) at the same time the paths through the core that contain this edge will be filtered by BGP in the direction of the speakers sampled by RouteViews. Thus, we assume that most c2p and p2c edges are already classified by our deterministic algorithm. Following this assumption we can infer that in paths that do not pass through the core, and have a *single* remaining unclassified edge (which must reside between a c2p and a p2c edges), this edge should be classified as a p2p edge (this is illustrated as the edge between ASes 3 and 4 in Fig. [2(a)](#S3.F2.sf1 "In Figure 2 ‣ 3.3 Non-Deterministic Inference of Remaining Relationships ‣ 3 AS Relationships Inference ‣ Near-Deterministic Inference of AS Relationships")). In case there is more than a single edge between the c2p and the p2c (as illustrated in Fig. [1(b)](#S3.F1.sf2 "In Figure 1 ‣ 3.1 Deterministic Classification ‣ 3 AS Relationships Inference ‣ Near-Deterministic Inference of AS Relationships")), we leave the edges unclassified, since we cannot determine which of the vertices is the provider.
The second type of unclassified edges are the ones that have a similar number of votes for two or more types of relationships. This can be the result of changes in the commercial relationship between adjacent ASes over the measurements period, or due to more complex peering agreements that can cause the same edge to behave differently as seen from different view points in the Internet \[[9](#bib.bib9)\].
To resolve these ambiguities, we use heuristic-based methods suggested by other works. Although we chose to use the AS degree in the graph \[[12](#bib.bib12)\], and the k-shell index \[[22](#bib.bib22), [7](#bib.bib7)\], any other method can be employed. Analysis of this inference technique is further discussed in the experimental results in section [4](#S4 "4 Experimental Results ‣ Near-Deterministic Inference of AS Relationships").
The third type of unclassified edges are edges that appear in non-valley-free paths (Fig. [2(b)](#S3.F2.sf2 "In Figure 2 ‣ 3.3 Non-Deterministic Inference of Remaining Relationships ‣ 3 AS Relationships Inference ‣ Near-Deterministic Inference of AS Relationships")), possibly the result of valid paths that pass a malformed core, or invalid paths that pass an accurate core. Since these invalid paths occur in only a small fraction of paths (less than 1% on average from the investigated paths per week), we leave the classification of these "valley edges" to future work.
### 3.4 Core Graph Construction
Motivated by the need to capture the true global hierarchal structure of the Internet we looked for an accurate global decomposition of the Internet AS-level graph. There have been several attempts to characterize the core of the Internet AS graph \[[21](#bib.bib21), [22](#bib.bib22), [7](#bib.bib7), [16](#bib.bib16), [15](#bib.bib15)\]. We use three core construction methods, that result in cores that vary in size and density. We analyze the effect that each core has on the classification algorithm.
Tauro et al. \[[22](#bib.bib22)\] proposed the *Jellyfish* conceptual model in which they identified a topological center and classified vertices into layers with respect to the center. The authors defined core as a clique of high-degree vertices, and constructed it by sorting the vertices in non-increasing degree order. The first vertex in the core is the one with the highest degree. Then, they examine each vertex in that order; a vertex is added to the core only if it forms a clique with the vertices already in the core. The resulting core is a clique but not necessarily the maximal clique of the graph. We refer to this core as *GreedyMaxClique*.
Carmi et al. \[[7](#bib.bib7)\] indicated that using the popular vertex’s degree (which was encouraged by the finding of the Internet’s power-law distribution \[[11](#bib.bib11)\]) as an indicator of the vertex’s importance can be misleading. The authors presented the new *Medusa* model, that uses a $`k`$-pruning algorithm to decompose the Internet AS graph and extract a nucleus (the $`K_{m\hspace{0pt}a\hspace{0pt}x}`$-Core) which is a very well connected globally distributed subgraph. Note that this algorithm extracts a core by looking at the entire graph, unlike GreedyMaxClique that takes a local approach. The properties listed for this model are useful for AS relationship inference, mainly due to the finding that the nucleus plays a critical role in BGP routing, since its vertices lie in a large fraction of the paths that connect different ASes. We refer to this core as *k*-Core.
The last core we use is constructed from most of the ASes and interconnecting edges that exhibit p2p relationship under the inference method in \[[9](#bib.bib9)\]. We use the Automated AS ranking provided by CAIDA \[[1](#bib.bib1)\] and constructed a graph that contains all the edges classified as p2p and their adjacent AS vertices. We then selected the largest connected component that contains some of the largest tier-1 ASes, namely AS701 ($`U\hspace{0pt}U\hspace{0pt}N\hspace{0pt}E\hspace{0pt}T`$) and AS7018 ($`{A\hspace{0pt}T}\& T`$). We refer to this core as *CP* (CAIDA Peers).
The three core types vary in size and density as an attempt to capture different inference behaviors. Using a small, dense core reduces the probability that a non-top-level AS is wrongfully considered as a top-level AS for all paths that pass through it, thus causing incorrect inferences. However, a small core might miss top-level ASes, thus cause non-valley-free paths. On the other hand, when using a large core, a trace might have several hops in the core. In this case we follow \[[21](#bib.bib21)\] and assume that two ASes may have an "indirect peering" relation, meaning they have p2p relationship through an intermediate AS, such as an exchange point. Traces with more than three hops in the core are considered invalid.
## 4 Experimental Results
In this section we evaluate the deterministic algorithm and the additional heuristics inferences using data from the first five weeks of 2007. We evaluate its accuracy by comparing the results to the classification algorithm proposed in \[[9](#bib.bib9)\], referred to as $`C\hspace{0pt}A\hspace{0pt}I\hspace{0pt}D\hspace{0pt}A`$. We start by discussing the data sources, and the three different type of cores we use as inputs to the algorithm. We then analyze the effect of the core size (number of vertices) and density (the number of edges divided by the full clique size) on the algorithm, and check the transient effects caused by aggregating data for changing time frames. Finally we check the sensitivity of the algorithm to increasing mistake in the core.
### 4.1 Data Sources
For this work we combined data from the RouteViews (RV) \[[2](#bib.bib2)\] and DIMES \[[20](#bib.bib20)\] projects to maximize the size of our AS topology. We used BGP paths collected by the RV project, similar to most other previous work \[[10](#bib.bib10), [13](#bib.bib13), [14](#bib.bib14), [22](#bib.bib22), [11](#bib.bib11), [21](#bib.bib21), [19](#bib.bib19)\]. We created weekly batches of AS-level paths by downloading one RV file that was generated daily at 20:00, and merged all 7 files, making sure that each path appears exactly once. We parse RV’s files and use only AS paths marked as "valid".
Since BGP paths miss many of the actual links (primarily of type p2p) caused by non-advertised links in BGP \[[9](#bib.bib9), [8](#bib.bib8)\] we use additional data from DIMES. DIMES is a large-scale distributed measurements effort that measures and tracks the evolution of the Internet from hundreds of different view-points, in an attempt to overcome the "law of diminishing returns" \[[5](#bib.bib5)\]. DIMES daily collects over 2 million traceroute and ping measurements targeted at a set of over 5 million IP addresses, which are spread over all the allocated IP prefixes.
In order to create AS-level paths from the IP-level traceroutes provided by DIMES agents, we preform AS resolution for each hop in all paths. Although IP-to-AS mapping can be a difficult task \[[24](#bib.bib24)\] we take a strait-forward resolution approach, and then filter out paths that exhibit various animalities. AS resolution is done by first performing longest-prefix-matching against BGP tables obtained from RV archive. This resolves approximately 98% of the IP addresses. For the remaining 2%, we query against two WhoIs databases, namely RIPE and RADB, that resolves additional 1.5% of the IP addresses. The remaining 0.5% unresolved IP addresses are discarded and do not participate in the inference algorithm.
The raw DIMES data was filtered in order to reduce inference mistakes and inclusion of false links. We filtered for some measurements artifacts by only including edges which were seen from at least two agents. In addition we trimmed all traces that exhibit known traceroute problems \[[4](#bib.bib4)\], namely routing loops and destination impersonation, keeping only the section of the path preceding the identified problem.
Table [1](#S4.T1 "Table 1 ‣ 4.1 Data Sources ‣ 4 Experimental Results ‣ Near-Deterministic Inference of AS Relationships") shows the number of ASes and interconnecting links gathered during the first five weeks of 2007, obtained by using both RouteViews and DIMES. On average, the data set consists of over 24,000 AS vertices and approximately 58,000 links (undirected edges). Approximately 44% of the edges exist only in RV paths, about 12% exist only in the filtered DIMES paths and the remaining 44% of the edges exist in both RV and DIMES. In section [4.5](#S4.SS5 "4.5 Analysis of Non-Deterministically Inferred Relationships ‣ 4 Experimental Results ‣ Near-Deterministic Inference of AS Relationships") we analyze the edges seen only by DIMES in order to understand the type of these additional links.
Table 1: ASes and links collected by Dimes and RouteViews during the first five weeks of 2007
<table>
<thead>
<tr>
<th>Week</th>
<th>ASes</th>
<th>Links</th>
<th>RV links</th>
<th>DIMES links</th>
<th>RV&DIMES links</th>
</tr>
</thead>
<tbody>
<tr>
<th>1</th>
<td>24391</td>
<td>57875</td>
<td>24282</td>
<td>6964</td>
<td>26629</td>
</tr>
<tr>
<th>2</th>
<td>24451</td>
<td>57920</td>
<td>24313</td>
<td>6986</td>
<td>26621</td>
</tr>
<tr>
<th>3</th>
<td>24492</td>
<td>57921</td>
<td>24609</td>
<td>6630</td>
<td>26682</td>
</tr>
<tr>
<th>4</th>
<td>24581</td>
<td>59058</td>
<td>24913</td>
<td>7288</td>
<td>26857</td>
</tr>
<tr>
<th>5</th>
<td>24716</td>
<td>59779</td>
<td>25528</td>
<td>7331</td>
<td>26920</td>
</tr>
</tbody>
</table>
On a weekly average, we filtered approximately 5,100 DIMES edges that were measured only once, which is over 15% of the edges measured by DIMES. Around half of these edges appear in RouteViews, providing a testimony to our conservatism.
### 4.2 Voting Threshold
In order to validate the usage of the voting technique described in section [3.2](#S3.SS2 "3.2 Assigning Type-of-Relationship to Edges ‣ 3 AS Relationships Inference ‣ Near-Deterministic Inference of AS Relationships") and set a proper threshold value, we tested the distribution of votes to inference types. For each edge we calculated the ratio of p2c votes. Fig. [3](#S4.F3 "Figure 3 ‣ 4.2 Voting Threshold ‣ 4 Experimental Results ‣ Near-Deterministic Inference of AS Relationships") shows the number of edges for each p2c ratio. Clearly, the vast majority of the edges are uniquely classified as either p2c or c2p. This remains true when running the algorithm on longer time frames.
[Refer to caption](/html/0711.4562/assets/x5.png)
Figure 3: Types-of-Relations voting distribution
Looking at the data backing up this graph, we see that on average over 94% of the edges have votes for exactly one relationship type, and almost 99% of the edges have over 80% of the votes casted for a single relationship type, which provides a very high level of confidence for this selected type. Thus, a threshold value of 0.8 covers almost 99% of the edges, and leaves approximately 1% of the edges to be classified using heuristic methods, or remain unclassified.
### 4.3 Sensitivity Analysis
Since the construction of the core graph is a major building block for the algorithm, we evaluate the effect that the core has on the inference process. We start by examining the overall algorithm performance and stability over consecutive weeks. We then evaluate the optimal core size, i.e., a core that results in a minimal inference mistake while achieving a high classification percentage. Finally, we test the sensitivity of the algorithm to errors in the core by randomly replacing core vertices.
We start by looking at the result of executing the algorithm using the first five weeks of 2007, each time with a different core type. As expected, most of the AS relationships are inferred in phase 1 of the deterministic algorithm using paths that traverse the core. These paths comprise a large percentage of all available paths, ranging from over 98% for k-Core and CP to 81% for the smaller GreedyMaxClique core.
Table [2](#S4.T2 "Table 2 ‣ 4.3 Sensitivity Analysis ‣ 4 Experimental Results ‣ Near-Deterministic Inference of AS Relationships") shows the structure of the different core types used and the effect it has on the deterministic inference algorithm. The percent of classified edges and edges matching CAIDA’s inference is calculated out of the total number of edges (including edges that are unclassified by CAIDA). It shows that the smallest GreedyMaxClique core results in the lowest deterministic inference percentage while the largest CP core have the highest percentage. This is the result of the larger cores having more paths that traverse them, therefore can be deterministically inferred. k-Core provides an excellent overall inference percentage (over 95% deterministically inferred and around 75% matching CAIDA). Additionally, the results are stable over the measured fix weeks period. The drop in the percentage of edges matching CAIDA in week 5 is caused due to a decrease in the number of edges classified by CAIDA.
Table 2: Structure of input cores and its effect on the deterministic inference algorithm
<table>
<tbody>
<tr>
<th>Core</th>
<th>Week→→\rightarrow</th>
<td>1</td>
<td>2</td>
<td>3</td>
<td>4</td>
<td>5</td>
</tr>
<tr>
<th>k-Core</th>
<th>Core Vertices</th>
<td>57</td>
<td>56</td>
<td>54</td>
<td>58</td>
<td>54</td>
</tr>
<tr>
<th></th>
<th>Core Edges</th>
<td>2260</td>
<td>2198</td>
<td>2076</td>
<td>2344</td>
<td>2134</td>
</tr>
<tr>
<th></th>
<th>Classified</th>
<td>95.59%</td>
<td>95.76%</td>
<td>95.34%</td>
<td>95.24%</td>
<td>94.22%</td>
</tr>
<tr>
<th></th>
<th>Match CAIDA</th>
<td>75.23%</td>
<td>75.32%</td>
<td>75.08%</td>
<td>73.76%</td>
<td>62.76%</td>
</tr>
<tr>
<th>Greedy</th>
<th>Core Vertices</th>
<td>17</td>
<td>17</td>
<td>17</td>
<td>18</td>
<td>17</td>
</tr>
<tr>
<th>Max</th>
<th>Core Edges</th>
<td>272</td>
<td>272</td>
<td>272</td>
<td>306</td>
<td>272</td>
</tr>
<tr>
<th>Clique</th>
<th>Classified</th>
<td>89.64%</td>
<td>89.87%</td>
<td>89.77%</td>
<td>89.62%</td>
<td>88.87%</td>
</tr>
<tr>
<th></th>
<th>Match CAIDA</th>
<td>73.73%</td>
<td>73.82%</td>
<td>73.60%</td>
<td>72.56%</td>
<td>61.68%</td>
</tr>
<tr>
<th>CP</th>
<th>Core Vertices</th>
<td>1067</td>
<td>1053</td>
<td>1068</td>
<td>1056</td>
<td>1087</td>
</tr>
<tr>
<th></th>
<th>Core Edges</th>
<td>6158</td>
<td>6110</td>
<td>6012</td>
<td>5844</td>
<td>6138</td>
</tr>
<tr>
<th></th>
<th>Classified</th>
<td>98.29%</td>
<td>98.55%</td>
<td>98.45%</td>
<td>98.0%</td>
<td>97.39%</td>
</tr>
<tr>
<th></th>
<th>Match CAIDA</th>
<td>79.77%</td>
<td>79.78%</td>
<td>79.43%</td>
<td>77.93%</td>
<td>67.19%</td>
</tr>
</tbody>
</table>
Although CP core seems to result in the best overall performance, constructing the CP core revealed that only a few p2p edges out of the approximately 6,000 edges were not a part of the largest connected component. This suggests that CAIDA incorrectly infers AS relationships as p2p, since it is highly unlikely that almost all p2p edges are connected. This causes a bias, resulting in more inference errors.
Table [3](#S4.T3 "Table 3 ‣ 4.3 Sensitivity Analysis ‣ 4 Experimental Results ‣ Near-Deterministic Inference of AS Relationships") shows that less than 6% of the edges were differently classified using two cores in each week, and the difference between k-Core and GreedyMaxClique is much smaller. This shows that the algorithm results are relatively consistent regardless of the input core.
Table 3: Percentage of edges that change classification comparing different core types
<table>
<thead>
<tr>
<th>Cores ||| Week→→\rightarrow</th>
<th>1</th>
<th>2</th>
<th>3</th>
<th>4</th>
<th>5</th>
</tr>
</thead>
<tbody>
<tr>
<td>k-Core - GreedyMaxClique</td>
<td>1.77%</td>
<td>1.66%</td>
<td>1.58%</td>
<td>1.8%</td>
<td>1.64%</td>
</tr>
<tr>
<td>k-Core - CP</td>
<td>5.94%</td>
<td>5.89%</td>
<td>5.84%</td>
<td>5.7%</td>
<td>5.81%</td>
</tr>
<tr>
<td>GreedyMaxClique - CP</td>
<td>3.53%</td>
<td>3.45%</td>
<td>3.37%</td>
<td>3.34%</td>
<td>3.51%</td>
</tr>
</tbody>
</table>
In order to find the best core size, we run the algorithm with a growing core size starting at four vertices. We do this for two of our core types - k-Core and GreedyMaxClique, using the first week of 2007. We start with the highest degree vertices and add vertices in a non-increasing degree order. Using k-Core, we first add vertices from the $`K_{m\hspace{0pt}a\hspace{0pt}x} - {C\hspace{0pt}o\hspace{0pt}r\hspace{0pt}e}`$ and then proceed to shells with lower indexes.
[Refer to caption](/html/0711.4562/assets/x6.png)
\(a\)
k-Core
[Refer to caption](/html/0711.4562/assets/x7.png)
\(b\)
GreedyMaxClique
Figure 4: Robustness of the algorithm to changes in the size of the core
Fig. [4](#S4.F4 "Figure 4 ‣ 4.3 Sensitivity Analysis ‣ 4 Experimental Results ‣ Near-Deterministic Inference of AS Relationships") shows the robustness of the algorithm relative to the size of the input core. The vertical dashed line marks the true core size. For both cores, it shows that for more than 20 vertices in the core the algorithm classification success and similarity to CAIDA do not significantly change, while the number of deterministically classified edges increases. However, this increase comes with an increase in the percentage of non-valley-free paths as shown in Fig. [5](#S4.F5 "Figure 5 ‣ 4.3 Sensitivity Analysis ‣ 4 Experimental Results ‣ Near-Deterministic Inference of AS Relationships"). This implies that the core must be kept small enough to decrease the number of invalid paths.
[Refer to caption](/html/0711.4562/assets/x8.png)
Figure 5: Non-Valley-Free paths
Overall we showed that the algorithm is consistent over time and various cores. Additionally, a core containing approximately 20 top-level ASes is sufficient to obtain excellent inference results.
### 4.4 Time Aggregation Analysis
We wish to find a time frame for which the algorithm captures best the relationships between ASes. A short time frame results in a fast running algorithm but might miss AS links and AS paths, especially in the DIMES data. This results in a low vote count, possibly decreasing the success of the algorithm. On the other hand, a long time frame captures two effects that can also cause a decrease in the algorithm’s success: 1) commercial relationships can change and complex routing behaviors may occur over long durations, and 2) possible measurements mistakes can pile up and skew the results.
We executed the algorithm on an increasing time frame. We started with the first day of 2007 and aggregated single days until the end of the first week (for this daily analysis we used three RouteViews files a day). Then, we aggregated a week at a time, until reaching 10 consecutive weeks. DIMES provides approximately 1.5M non-unique tracroutes each day, reaching over 100M traceroutes for the 10 weeks period. RouteViews provides approximately 1.2M *unique* paths regardless of the time frame used.
[Refer to caption](/html/0711.4562/assets/x9.png)
Figure 6: Deterministically classified edges over an increasing time frame
The percentage of deterministically inferred edges over the aggregated time frame is shown in Fig. [6](#S4.F6 "Figure 6 ‣ 4.4 Time Aggregation Analysis ‣ 4 Experimental Results ‣ Near-Deterministic Inference of AS Relationships"). Using data from a single week (marked as the vertical dashed line) results in over 90% of the edges being classified for all core types, having CP obtaining the best percentage and GreedyMaxClique the worst. This is directly related to the size of the core, since a larger core results in more paths that traverse through it, yielding more deterministically inferred relationships.
We evaluate the deterministic algorithm over this time frame by looking at the edges that are identically classified by the deterministic algorithm and CAIDA, out of the edges that are classified by both. Fig. [7](#S4.F7 "Figure 7 ‣ 4.4 Time Aggregation Analysis ‣ 4 Experimental Results ‣ Near-Deterministic Inference of AS Relationships") shows that for all cores and any time frame, the algorithms agree on over 92% of the edges. Obviously, using CP gives the best match to CAIDA’s inference. It is interesting to see that although k-Core has better overall deterministic inference success than GreedyMaxClique (shown in Fig. [6](#S4.F6 "Figure 6 ‣ 4.4 Time Aggregation Analysis ‣ 4 Experimental Results ‣ Near-Deterministic Inference of AS Relationships")), it results in a lower match rate. This is probably due to the relatively small, local and degree-based GreedyMaxClique core, which is more related to the heuristic used by CAIDA’s inference methods.
[Refer to caption](/html/0711.4562/assets/x10.png)
Figure 7: Percentage of deterministically inferred edges matching CAIDA out of edges that are inferred by both algorithms
Finally, we looked at the consistency of the inference results over the time frame by comparing edges that are classified in both time frames. We found that over 98% of the inferences remain constant between consecutive time frames. This suggests that there are only a few commercial relationships that change over time. Short-term routing changes have very little effect, since they statistically "disappear" as the more common routes become dominant over time.
### 4.5 Analysis of Non-Deterministically Inferred Relationships
Edges that the deterministic algorithm fails to classify are classified using the two heuristic-based inference methods described is section [3.3](#S3.SS3 "3.3 Non-Deterministic Inference of Remaining Relationships ‣ 3 AS Relationships Inference ‣ Near-Deterministic Inference of AS Relationships"). The first method breaks voting ties and the second infers p2p relationships.
To break voting ties we compared adjacent AS degrees (similar to \[[12](#bib.bib12)\]) and inferred the relationship between them as p2p if the degrees ratio is between 0.8 and 1.2, or p2c otherwise (marking the provider as the AS with the higher degree). For k-Core we also compared the $`k`$-Shell index, and inferred the relationship to be p2p if the two ASes have the same $`k`$-Shell index, or p2c otherwise (marking the provider as the AS with the higher $`k`$-Shell index) and note very little difference between the two heuristics.
We estimate the accuracy and robustness of the algorithm by intentionally increasing the mistake in the core. We do this by randomly replacing ASes in the core and see how the number of relationships inferred and their correlation with CAIDA’s inference change. We start by replacing one core AS with one random AS (that is connected to at least one of the remaining core ASes) and gradually replace more ASes until we have a core that consists of completely random but still connected ASes. Fig. [8](#S4.F8 "Figure 8 ‣ 4.5 Analysis of Non-Deterministically Inferred Relationships ‣ 4 Experimental Results ‣ Near-Deterministic Inference of AS Relationships") shows the percentage of classified edges, deterministically classified edges and the percentage of edges that match CAIDA’s inference using k-Core and GreedyMaxClique.
[Refer to caption](/html/0711.4562/assets/x11.png)
\(a\)
k-Core
[Refer to caption](/html/0711.4562/assets/x12.png)
\(b\)
GreedyMaxClique
Figure 8: Robustness of the algorithm to errors in the core
Interestingly, while the algorithm’s performance decreases as we increase the randomness of the core, the overall degradation is not as high as one would expect. However, Fig. [9](#S4.F9 "Figure 9 ‣ 4.5 Analysis of Non-Deterministically Inferred Relationships ‣ 4 Experimental Results ‣ Near-Deterministic Inference of AS Relationships") shows a rising trend of the percentage of unclassified, p2p and tie-breaking heuristic edges as we inject errors to the core. As more errors are injected, the algorithm needs to use more heuristics. Particulary, when there are approximately 50% random vertices in the core, the effect of the increasing mistake becomes more noticeable. However, even with a completely random core, the overall heuristically inferred edges account for less than 20% of all edges.
[Refer to caption](/html/0711.4562/assets/x13.png)
\(a\)
k-Core
[Refer to caption](/html/0711.4562/assets/x14.png)
\(b\)
GreedyMaxClique
Figure 9: Percentage of p2p, heuristically classified and unclassified edges
These results indicate that although the algorithm seems quite robust to the mistake in the core, it is still significantly affected once there are more than 50% incorrect core vertices. Looking at p2p edges showed us that as we inject more errors, the percentage of p2p edges that are classified differently by CAIDA increases from around 16% to over 40%.
Finally, we analyzed the distribution of relationship types for edges that appear only in DIMES, and are not seen in the BGP routing tables, to understand what are the types of relationships that DIMES discovers. We found that while on average the p2p relationships comprise 4-5% of the total number of edges, it goes up to around 12% of the edges that appear only in DIMES. Moreover, approximately 40% of the p2p edges inferred by our algorithm, do not appear in the RV tables. This means that utilizing DIMES significantly improves the ability to detect p2p links between ASes, mainly since DIMES agents are spread over the Internet and contribute AS links that are either not collected by the RouteViews routers or even not published in the BGP protocol.
## 5 Conclusion and Future Work
The common weakness of previously proposed AS relationships inference algorithms is their lack of guarantee on inference errors introduced during the process. This work improves on existing methods by providing a near-deterministic algorithm that, given a classified error-free input core, does not introduce additional inference errors. We investigate various input cores and show that the proposed algorithm provides accurate inferences that are robust under changes in the core’s size and creation technique. We show that a core containing as little as 20 almost fully-connected ASes is sufficient for good inference results. Additionally, we show that heuristic methods can still play an important role in inferring the remaining relationships. Using data collected from a single week (containing approximately 1.2M BGP paths and over 10M DIMES AS-level traceroutes), the algorithm runs for only about 2 hours and yields over 95% deterministically inferred relationships.
As the Internet grows larger, many providers interconnect at multiple locations for traffic engineering and embrace the usage of exchange points. The relationships and policies used in these interconnection points might not conform to either provider-to-customer or peer-to-peer relationships. Moreover, it might not even conform to the valley-free property. The data provided from the DIMES project can reveal these complex relationships and seed other large scale Internet analysis work.
## References
- \[1\] CAIDA. Automated Autonomous System (AS) ranking. Research Project. http://as-rank.caida.org.
- \[2\] University of Oregon RouteViews Project (http://www.routeviews.org/).
- \[3\] J. I. Alvarez-Hamelin, L. Dall’Asta, A. Barrat, and A. Vespignani. k-core decomposition: a tool for the analysis of large scale internet graphs, 2005.
- \[4\] B. Augustin, X. Cuvellier, B. Orgogozo, F. Viger, T. Friedman, M. Latapy, C. Magnien, and R. Teixeira. Avoiding traceroute anomalies with paris traceroute.
In IMC’06, pages 153–158, 2006.
- \[5\] P. Barford, A. Bestavros, J. Byers, and M. Crovella. On the marginal utility of network topology measurements.
In IMW’01: the 1st ACM SIGCOMM Workshop on Internet Measurement, pages 5–17, 2001.
- \[6\] G. D. Battista, M. Patrignani, and M. Pizzonia. Computing the types of the relationships between autonomous systems. Technical Report RT-DIA-73-2002, Dipartimento di Informatica e Automazione, Universita di Roma Tre, 2002., 2002.
- \[7\] S. Carmi, S. Havlin, S. Kirkpatrick, Y. Shavitt, and E. Shir. Medusa - new model of internet topology using k-shell decomposition, 2006.
- \[8\] R. Cohen and D. Raz. Acyclic type of relationships between autonomous systems.
In IEEE INFOCOM, 2007.
- \[9\] X. Dimitropoulos, D. Krioukov, M. Fomenkov, B. Huffaker, Y. Hyun, kc claffy, and G. Riley. As relationships: Inference and validation.
ACM SIGCOMM Computer Communications Review, 37:2007, 2006.
- \[10\] X. Dimitropoulos, D. Krioukov, B. Huffaker, kc claffy, and G. Riley. Inferring as relationships: Dead end or lively beginning?
LNCS, 3503:113, 2005.
- \[11\] M. Faloutsos, P. Faloutsos, and C. Faloutsos. On power-law relationships of the internet topology.
SIGCOMM Comput. Commun. Rev., 29(4):251–262, 1999.
- \[12\] L. Gao. On inferring autonomous system relationships in the internet.
IEEE/ACM Transactions on Networking, 9(6):733–745, 2001.
- \[13\] L. Gao, T. Griffin, and J. Rexford. Inherently safe backup routing with BGP.
In INFOCOM, pages 547–556, 2001.
- \[14\] L. Gao and J. Rexford. Stable internet routing without global coordination.
In Measurement and Modeling of Computer Systems, pages 307–317, 2000.
- \[15\] Z. Ge, D. Figueiredo, S. Jaiwal, and L. Gao. On the hierarchical structure of the logical internet graph.
In SPIE ITCOM, Aug. 2001.
- \[16\] R. Govindan and A. Reddy. An analysis of internet inter-domain topology and route stability.
In INFOCOM, pages 850–857, 1997.
- \[17\] G. Huston. Interconnection, peering, and settlements.
In INET, San Jose, CA, USA, June 1999.
- \[18\] S. Kosub, M. G. Maaß, and H. Täubig. Acyclic type-of-relationship problems on the internet.
In The 3rd Workshop on Combinatorial and Algorithmic Aspects of Networking (CAAN’06), volume 4235 of LNCS, pages 98–111. Springer-Verlag, July 2006.
- \[19\] Z. M. Mao, L. Qiu, J. Wang, and Y. Zhang. On as-level path inference.
In SIGMETRICS’05, pages 339–349, 2005.
- \[20\] Y. Shavitt and E. Shir. Dimes: let the internet measure itself.
ACM SIGCOMM Computer Communications Review, 35(5):71–74, 2005.
- \[21\] L. Subramanian, S. Agarwal, J. Rexford, and R. H. Katz. Characterizing the internet hierarchy from multiple vantage points.
In IEEE INFOCOM 2002, New York, NY, USA, Apr. 2002.
- \[22\] L. Tauro, C. Palmer, G. Siganos, and M. Faloutsos. A simple conceptual model for the internet topology.
- \[23\] J. Xia and L. Gao. On the evaluation of as relationship inferences.
In IEEE Globecom, Dallas, TX, USA, Nov. 2004.
- \[24\] J. R. J. W. Zhuoqing Mao, David Johnson and R. H. Katz. Scalable and accurate identification of as-level forwarding paths.
In INFOCOM. |
0712.2231 | cs.CR | cyber | # Trust for Location-based Authorisation
Andreas U. Schmidt1, Nicolai Kuntze1, Jörg Abendroth2,
1Fraunhofer Institute for Secure Information Technology SIT
Rheinstraße 75, 64295 Darmstadt, Germany
Email: {andreas.schmidt,nicolai.kuntze}@sit.fraunhofer.de 2Nokia Siemens Networks GmbH & Co. KG
Otto-Hahn-Ring 6, 81737 Munich, Germany
Email: joerg.abendroth@nsn.com
###### Abstract
We propose a concept for authorisation using the location of a mobile device and the enforcement of location-based policies. Mobile devices enhanced by Trusted Computing capabilities operate an autonomous and secure location trigger and policy enforcement entity. Location determination is two-tiered, integrating cell-based triggering at handover with precision location measurement by the device.
## I Introduction
Ambient technologies propose to enrich the mobile environment with different computer devices and networks. One use case is to utilise a user device for remote control of other systems — to automatically turn on the light when entering a room, or to enable access to the emergency controls of a power plant while repairing the pipes. Often, a natural requirement is to restrict possibilities to authorised personnel in a certain location. Vice versa sometimes one wants to restrict the functionality of devices based on its location, e.g., disable the camera while entering a sensitive area. All these scenarios require location based authorisation and sometimes access control, a subject which has attracted some attention in the literature, cf. \[[1](#bib.bib1)\] and references therein.
The implementation of location based authorisation can be based on various methods for location determination, e.g., triangulation (e.g. strength of the device signal at six base stations \[[2](#bib.bib2), [3](#bib.bib3)\]), calculating the round trip time \[[4](#bib.bib4)\], or deploying a GPS module locally to the client. The first method lacks the precision, and the second requires new sensors to be installed in the location-based authorisation enabled areas. The last has not been deployed because relying on the client side for location information is deemed insecure. Event though hybrid approaches like combination of triangulation and GPS exist, the latter argument applies to them. In conclusion current approaches are generally unfit for authorisation and access control based on location with scalable strength of enforcement. We propose a solution, that combines three technologies. First a network-side location trigger, which only has cell granularity, and second a device-side enforcing agent using de-central (e.g. GPS) or co-operative (e.g. A-GPS) location methods and concurrently enforces the pertinent policies. Third security is established by applying Trusted Computing concepts of Attestation. Section [II](#S2 "II TLTA Concept ‣ Trust for Location-based Authorisation") presents the proposed concept on the level of functional building blocks, architecture, and operational scenario. Section [III](#S3 "III Protocols and Procedures ‣ Trust for Location-based Authorisation") defines the required communication protocols and their integration into pcs networks. Security and efficiency of the concept and application scenarios are discussed in Section [IV](#S4 "IV Discussion ‣ Trust for Location-based Authorisation"). Section [V](#S5 "V Conclusion ‣ Trust for Location-based Authorisation") concludes the paper.
## II TLTA Concept
In order to fell authorisation decisions based on device location, an entity called location trigger must have access to location information and be able to provide this information to other entities enforcing according policies, either on the device or on the part of some service. The concept of an Trusted Location Trigger Authorisation (TLTA) we propose in the following, rests on these key ideas. First, the location trigger in TLTA operates on two separate levels: i) Handover-based localisation within a network cell and ii) GPS-based (here always meaning either full, device operated GPS or A-GPS) localisation within an area circumscribed by a perimeter of cells. Second, a trusted entity on the device, called the Location Trigger Enforcer (LTE), embodying the location trigger functionality and enforcing, possibly in co-operation with a network-side counterpart, authorisation policies. The trust in the LTE rests on security properties of the device, which we assume to be a trusted computing (TC) platform for the purpose of location-based authorisation. The latter notion is, for our subject, specified by the standards of the Trusted Computing Group’s (TCG) Mobile Phone Working Group (MPWG, see \[[5](#bib.bib5)\]). The necessary architectural principals of the MPWG for TLTA are explained in Section [](#S2.SS2 "II-B Trusted MTs ‣ II TLTA Concept ‣ Trust for Location-based Authorisation")
II-B
. It should be noted however, that the TLTA concept is applicable resting on any other trusted platform architecture which fulfils the core requirements in the namely section.
In the following, we will take an agnostic stance with respect to the actual restrictions that are to be enforced and just speak of ‘the policies’ to be enforced. The next subsection sketches the architecture of a TLTA system and security-relevant assumptions. The state sequence of a device operating under the TLTA scheme is developed in Section [](#S2.SS3 "II-C Mobility process ‣ II TLTA Concept ‣ Trust for Location-based Authorisation")
II-C
.
###
II-A
Entities and their functions
We here choose the Long Term Evolution of 3G Networks specified by the 3GPP standardisation (3GPPLTE, see \[[6](#bib.bib6)\]) as the technical framework for the presentation of TLTA. Within it, various improvements over common 2G and 3G network architectures are specified. One of the aims is to simplify the overall architecture. This flat, all-IP based infrastructure is referred to as System Architecture Evolution (SAE). Meeting technical and performance requirements arising from an analysis of the current state of mobile technology resulted in a reduction of the number of network nodes involved in data processing and transport. A flatter network architecture leads to improved data latency (the transmission delay between the transmitter sending data and the time of reception) and better support of delay-sensitive, interactive and real-time communications.
[Refer to caption](/html/0712.2231/assets/x1.png)
Figure 1: SAE model and its basic components
The SAE model is shown in Figure [1](#S2.F1 "Figure 1 ‣ II-A Entities and their functions ‣ II TLTA Concept ‣ Trust for Location-based Authorisation") and consists of two types of network elements. Access of user devices, called Mobile Terminals (MT) is facilitated by enhanced base stations, referred to as enhanced Node-Bs (eNB). They provide the air interface and performs the resource management thereof. The second class of elements are the Access Gate Ways (AGW) which implement management functionalities and act as data gateways for user packages. The AGW establishes connections to Service Providers. Within this core network model the MNO can offer different services, for instance A-GPS.
Figure [2](#S2.F2 "Figure 2 ‣ II-A Entities and their functions ‣ II TLTA Concept ‣ Trust for Location-based Authorisation") depicts the interaction between the basic SAE architecture (shown as an MNO cloud) and the components added by TLTA. The customer who requests the enforcement of certain policies on devices in a geographic area, called the protected zone, is termed TLTAC service requester (TLTSR). The TLTSR handles for instance the provisioning of content to devices. TLTA introduces on network operator side the Trusted Location Trigger Authorisation Centre (TLTAC) which establishes a pre-defined geographic area within the area covered by the MNO network. The TLTAC receives (i) the physical geometry of the protected zone and (ii) the desired policies to be enforced within this zone (A in the figure). Based on these data the TLTAC determines configures certain eNBs of the MNO in Step B. After the setup TLTAC’s main task is to authenticate the end users devices and to maintain a list of them for authorisation purposes.
A Location Trigger Enforcer (LTE) is integrated in the end users MT which is the device-side counterpart of the TLTAC (The abbreviation LTE used in this paper should not be confounded with Long Term Evolution in the context of the 3GPP). LTE (i) receives and holds the policies from TLTAC, (ii) receives and holds the protection zone geometry, (iii) determines entry and exit events to the protection zone by some high-precision location mechanism, e.g., A-GPS, and (iv) enforces the policies on the device.
[Refer to caption](/html/0712.2231/assets/x2.png)
Figure 2: Interaction of an existing MNO network and the TLTA components
After this setup, MTs entering the protected zone may acquire certain services. The access to these services is controlled by the LTE. The basic scheme is that the device which enters the protected zone performs an attestation (see below) of its internal state and configuration to the MNO (Step 1 in Figure [2](#S2.F2 "Figure 2 ‣ II-A Entities and their functions ‣ II TLTA Concept ‣ Trust for Location-based Authorisation")). If this attestation succeeds the device is registered at the TLTAC (Step 2) and also registered with the A-GPS service, or any other location providing service which is used to determine the accurate location of the device (Step 3). These registration steps are mainly required to protect the location service against fraud and to enable certain payment methods and schemes.
###
II-B
Trusted MTs
The TLTAC and LTE together form a Policy Decision Point (PDP) / Policy Enforcement Point (PEP) related model based on a local enforcement of policies. From the perspective of the content provider the trustworthiness of the local LTE must be testified and sufficiently strong. Therefore, TLTA requires that the MT constitutes a trusted platform resting on certain capabilities constituting a Trusted Computing Base (TCB), which for the purpose of this report comprise:
####
II-B1
Secure boot
The boot process is the time span from the power on to the stage when the user is able to interact with the device. Within this time span, secure boot ensures that fundamental capabilities are in place which are critical for the reliability of the system, comprising in particular means for resource isolation and management. During secure boot, as defined in \[[5](#bib.bib5)\], each component is measured in order of execution. These measurement values are then verified by a local verification agent using so called Reference Integrity Measurement (RIM) digital certificates. If the values match then the processing continues and the started component is registered in a log as well as in a corresponding system state register in a secure storage space. Otherwise the boot will fail and the system switches into a failed state. In case of a failed state the system may offer a Pristine Boot which as a fallback. Pristine Boot defines a device state after initial factory installation where the life cycle of a device begins. Another use case for the Pristine Boot is an update process to rebuild the required credentials (, e.g., RIM certificates) used during secure boot. If the secure boot of an MT succeeds the system on top can rely on the integrity of the system underneath. This includes all subsystems integrated in the platform including the LTE.
[Refer to caption](/html/0712.2231/assets/x3.png)
Figure 3: Mobile station travelling through the zones and perimeters of a TLTA system.
####
II-B2
Remote attestation
A system which has performed a secure boot process on an appropriate platform constitutes a trustworthy base for other operations. This is not useful without means to convey this state in a reliable way to other entities. This process is called attestation in the context of TC. TC uses two concepts to implement this functionality: Platform Configuration Registers (PCR) and Attestation Identities (AI). A PCR reflects the system configuration reported by the measurement entity and is protected within the Mobile Trusted Module (MTM) against tampering. PCR values are created by combining digest values representing the measured component. Each value added to the PCR is also recorded in a log. Log and the corresponding PCR value together are testifying to the system state. To simplify the verification process, the MPWG has introduced a device side verifier. This verifier works as an agent close with the measurement system and holds RIM certificates. An Attestation Identity is embodied in a credential, and usually certified by a trusted third party, e.g., a Privacy CA, to attest that the owning MT contains a live, unaltered MTM. Producing the AI credential together with PCR values and measurement logs and signed by the AI private key to an external verifier, constitutes the attestation process proper. The verifier, in the particular case at hand, then has assurance that the MT operates an unaltered LTE on top of a set of trustworthy resources.
###
II-C
Mobility process
The protected zone, denoted $`p\hspace{0pt}z`$, is for simplicity taken as the area circumscribed by a closed polygon in the plane which in turn is covered by cells of the mobile network. Now, assign the number $`1`$ to all cells on the boundary of the closed, connected set of cells covering the $`p\hspace{0pt}z`$ and call this set of cells $`c_{1}`$ The surveillance perimeter ($`s\hspace{0pt}p`$) is defined as outer boundary of $`c_{1}`$. Call $`c_{0}`$ the set of cells adjacent to $`s\hspace{0pt}p`$, $`c_{- 1}`$ the set of cells adjacent to the outer boundary of $`c_{0}`$, and so on. Finally let the outbound perimeter $`o\hspace{0pt}p`$ be a closed polygon such that $`c_{1} \subset \overline{o\hspace{0pt}p}`$ and $`{{dist}{({o\hspace{0pt}p},c_{1})}} > 0`$. This geometry is shown in Figure [3](#S2.F3 "Figure 3 ‣ II-B1 Secure boot ‣ II-B Trusted MTs ‣ II TLTA Concept ‣ Trust for Location-based Authorisation"), wherein, $`o\hspace{0pt}p`$ results from simply scaling $`p\hspace{0pt}z`$.
Before TLTA can operate according to a TLTSR request, steps of device preparation and node configuration must be executed. For the latter, TLTAC transmits a request to the AGW which configures the eNBs of $`c_{0}`$ and $`c_{1}`$ to execute the trust-enhanced handover described in Section [](#S3.SS1 "III-A Trust-enhanced handover at the perimeter ‣ III Protocols and Procedures ‣ Trust for Location-based Authorisation")
III-A
below, when an MT crosses the surveillance perimeter inbound. Device preparation is not required for trusted MTs already equipped with an LTE, thus it reduces to LTE enrolment. This can be done by secure software download, which is a major use case for TC-enabled devices, at any stage in the mobility process, e.g., when a device enters a $`c_{0}`$ cell.
[Refer to caption](/html/0712.2231/assets/x4.png)
Figure 4: Network and device states during the travel.
Figure [3](#S2.F3 "Figure 3 ‣ II-B1 Secure boot ‣ II-B Trusted MTs ‣ II TLTA Concept ‣ Trust for Location-based Authorisation") shows the travel of a mobile device MT through the zones. We describe this journey on a functional level, deferring details of communication protocols and procedures to Section [III](#S3 "III Protocols and Procedures ‣ Trust for Location-based Authorisation"). The states of MT and the mobile network relevant to TLTA are shown in Figure [4](#S2.F4 "Figure 4 ‣ II-C Mobility process ‣ II TLTA Concept ‣ Trust for Location-based Authorisation"). At the point marked $`(1)`$, MT enters the mobile network in a normal operational mode, with deactivated LTE and A-GPS functions. Upon crossing $`s\hspace{0pt}p`$ at $`(2)`$, the handover between a cell of $`c_{0}`$ and the corresponding cell of $`c_{1}`$ is augmented by the activation of the LTE and the A-GPS, a remote attestation process, registration of MT with the TLTAC, and download of the authorisation policy, and the geometry of $`p\hspace{0pt}z`$ to MT. From there on the device autonomously operates the location trigger to detect $`p`$ using A-GPS and with the aid of an A-GPS support server (possibly a function of TLTAC). When MT enters $`p\hspace{0pt}z`$ at $`(3)`$, LTE enforces the policy, for instance unlocks a preassigned credential for service access. During sojourn in $`p\hspace{0pt}z`$, MT follows the policy and can for instance use a LBS whenever the user desires. At some points $`(4)`$, MT may leave and re-enter $`p\hspace{0pt}z`$ and LTE switches the enforcement of the policy off and on. LTE finally releases policy enforcement at point $`5`$, but is deactivated only at crossing of $`o\hspace{0pt}p`$. At $`(6)`$, LTE also actively de-registers MT with TLTAC. The triggering of location is cell-based only between points $`(1)`$ and $`(2)`$ and after $`(6)`$, otherwise it relies on (assisted) GPS and MT’ autonomous decision via LTE.
## III Protocols and Procedures
Here we present the technical concepts at the heart of TLTA, namely integration with the handover at $`s\hspace{0pt}p`$, operation in $`p\hspace{0pt}z`$, and release at $`o\hspace{0pt}p`$.
###
III-A
Trust-enhanced handover at the perimeter
The trusted activation of the LTE, i.e., the activation of the LTE on the MT and a trustworthy report of that fact toward the network, when the device crosses the $`s\hspace{0pt}p`$, is at the heart of the TLTA concept. The main idea is to integrate this process into the normal handover of MT between a cell of $`c_{0}`$ and of $`c_{1}`$. Various handover protocols have been specified, and here we rely on 3GPPLTE, which offers advanced flexibility. 3GPPLTE handover (HO) is divided in two basic stages, HO preparation and HO execution. In the present case it involves an eNB of $`c_{0}`$ (eNB0) and of $`c_{1}`$ (eNB1), as shown in Figure [5](#S3.F5 "Figure 5 ‣ III-A Trust-enhanced handover at the perimeter ‣ III Protocols and Procedures ‣ Trust for Location-based Authorisation").
[Refer to caption](/html/0712.2231/assets/x5.png)
Figure 5: Handover and attestation at the surveillance perimeter.
This process is augmented by two stages of remote attestation of the MT, regarding the presence and activity of the LTE, in our scheme. This design maximally separates the tasks between the eNBs of $`c_{0}`$ and $`c_{1}`$. The first one, attestation preparation, sets up eNB1 and MT to perform LTE activation and remote attestation. The second one, attestation execution, is performed after HO and does not interfere with this time-critical process.
When eNB1 is notified of an HO request of some MT, it issues an HO response to eNB0, and additionally notifies eNB0 of the necessity of a remote attestation of MT. The sole task of eNB0 is then to generate and distribute a particular nonce, as specified by the TCG’s remote attestation protocol \[[7](#bib.bib7)\], for this purpose to MT and eNB1. This is a modification of the original TCG attestation protocol which combines eNB0 and eNB1 into one verifying party. eNB0 issues the \[TCG conforming\] attestation request to MT, along with the Att_nonce. The activation of LTE and attestation thereof is concluded only after HO. This means that the device loads the LTE (if it is not already present in its memory), generates an attestation package\[[7](#bib.bib7)\], and sends it to eNB1 only after the initial signalling with this B-Node is complete and acknowledged by the MT.
By attestation, eNB1 can be convinced that MT is in a trustworthy state and has an untampered LTE present. It then sends an acknowledgement to MT, including the signal to activate the LTE for operation within the $`s\hspace{0pt}p`$.
###
III-B
Protected zone operation and release
Registration of the device at the TLTAC and download of the policy to the LTE on the MT are shown in Figure [6](#S3.F6 "Figure 6 ‣ III-B Protected zone operation and release ‣ III Protocols and Procedures ‣ Trust for Location-based Authorisation") and can be initiated by the AGW, eNB1 or the device itself. We have chosen the second variant in the figure. After registration with TLTAC, and optionally the A-GPS service, the MT/TLTAC pair can operate as a combination of policy enforcement point/policy decision point of an authorisation system in any desired separation of duties.
[Refer to caption](/html/0712.2231/assets/x6.png)
Figure 6: Process after the activation of LTE and its interaction with TLTAC
Policy enforcement can be divided into two levels for the area between $`s\hspace{0pt}p`$ and $`p\hspace{0pt}z`$ and the interior of $`p\hspace{0pt}z`$ with two separate policies called $`\text{P}_{\text{sp}}`$ and $`\text{P}_{\text{pz}}`$, respectively. Authorisation decisions can be felled either by LTE alone or by LTE in collaboration with LTACC in both cases. For the latter, LTE would transmit with an authorisation request a binary attribute to LTACC, stating whether the device is in pz or not.
The location trigger is operated autonomously by the LTE, which periodically performs GPS measurements and requests geopositioning from the A-GPS service. When the MT detects entry to $`p\hspace{0pt}z`$, the enforced policy is switched from $`\text{P}_{\text{sp}}`$ to $`\text{P}_{\text{pz}}`$. MT can then, for instance, consume services from TLTSR. Upon MT crossing $`o\hspace{0pt}p`$, it notifies TLTAC, which de-registers it and signals the exit event to the A-GPS service (optionally, also TLTSR could be signalled).
## IV Discussion
In this section we discuss the security of TLTA, questions of performance and practical issues, and finally sketch some concrete application scenarios.
###
IV-A
Security
The two-tiered architecture of TLTA naturally divides security in two levels. Security on the sp rests on the integrity of the involved B-nodes eNB0 and eNB1, which is a precondition. Security in pz on the other hand depends on the separation of tasks between TLTAC and the LTE. The security of the LTE is enabled by the unique features of trust-enhanced MTs, namely secure boot and attestation, features which are fundamental for TLTA. When a registered device enters or leaves pz from or to sp, policy enforcement by LTE switches between the two policies $`\text{P}_{\text{sp}}`$ and $`\text{P}_{\text{pz}}`$ above. We discuss the security properties of the enforcement of $`\text{P}_{\text{pz}}`$.
The design entails that a security assessment of TLTA must distinguish between policies for functional enforcement and access control, respectively. The former means that certain functions on all devices in pz shall be enabled or disabled. The latter allows certain content or services to be delivered or made available to the MTs registered with TLTAC. The essential difference between the two is that, in order to extend functional enforcement to all MTs in pz, LTE activation must be ensured on all of them or the policy must be enforced otherwise, e.g., by out-of-band measures and processes. Access control is less security-critical. In contrast to functional enforcement, an access control policy is effectively enforced by the TLTAC, which will yield access only to MTs with active LTE accepting $`\text{P}_{\text{pz}}`$. In this case, LTE provides the added security that proliferation of access to content or services beyond pz is effectively prevented.
Functional enforcement in TLTA is susceptible to certain attacks which do not apply to access control and whose general strategy is to subvert sp. The simplest and most effective is to cross sp and enter pz with a disabled MT, or an MT shielded against network communication. Though one could attempt to detect such rogue MTs via the network when they are activated inside pz, this might not be very efficient nor effective. A more sophisticated attacker could try to prevent the handover at the sp, e.g., using a beam antenna to tamper the MT’s measurement of the base stations. In principle, this kind of attacks can only be mitigated by additional, physical access control procedures on pz’s boundary.
A general caveat of architectures relying on a trusted platform built upon a hardware security anchor like an MTM is that it is more susceptible to side channel attacks than, e.g., a smart card. Applied to the TLTA concept this argument could mean to imply that, e.g., location data is tampered with while in the MT’s memory or in transit between processor and memory, respectively, the wireless modem. A hardening of the MT preventing this could be achieved using secure channels \[[8](#bib.bib8), [9](#bib.bib9), [10](#bib.bib10)\] (Texas Instruments’ proprietary M-Shield technology is another example which could sensibly be combined with Trusted Computing for TLTA and other purposes).
###
IV-B
Efficiency
The activities during the central stage of TLTA, the entrance to the surveillance perimeter, offer a great variety of implementation options. In particular it should be noted that though we used (A)-GPS as a prime example, any other autonomous or MT/eNB co-operative method for detecting entry to pz such as triangulation, is applicable.
Let us briefly discuss the design choices taken in Sections [II](#S2 "II TLTA Concept ‣ Trust for Location-based Authorisation") and [III](#S3 "III Protocols and Procedures ‣ Trust for Location-based Authorisation"). The de-centralised architecture and operation of TLTA is designed with the central aim of reducing network load. This is only possibly with the desired level of security and enforcement if trust-enhanced MTs are used as central functional building blocks. TLTA is loosely coupled to the top-level of the cellular infrastructure. The TLTAC operates on the level of the network’s Home/Visitor Location Register (HLR/VLR) and Mobile Switching Centre (MSC) And could be co-located or integrated with either one.
Signalling cost is a major concern for location management in cellular overlay networks \[[11](#bib.bib11), [12](#bib.bib12)\], which poses similar demands as location based authorisation. In our case, the two-tiered method of TLTA yields the largest advantage in this respect by leveraging the capacities of the edge of the network. In particular communication with MTs at the sp is completely borne by the eNBs and thus maximally de-centralised, and moreover MTs operate largely autonomous inside pz. Furthermore, the protocol is designed such that the eNB of $`c_{0}`$ is passive, its sole essential task being the generation of the `Att_nonce`. This saves communication effort at the stage of node configuration, since only the eNBs of $`c_{1}`$ need to be activated to handle entry events. Another feature is the separation of inbound (sp) and outbound perimeter. It avoids judder, i.e., frequent switching on and off of the LTE and according (de)registration with the TLTAC.
A bottleneck could ensue if the physical access to the pz is restricted, in the extreme case to a single cell and a large number of MTs enters sp. This can happen in particular in application scenarios for functional enforcement (see below). This concentrates the load for the TLTA entry protocol to a single eNB0, eNB1 pair. The problem can be alleviated by a) organisational measures to distribute entry points over the perimeter of pz, or b) shifitng the sp further away from pz, i.e., to the boundary between cell layers $`c_{- n}`$ $`c_{{- n} + 1}`$, $`n \geq 1`$. A third option is to introduce extra cells ad hoc, e.g., mobile nano/pico-cells\[[13](#bib.bib13), [14](#bib.bib14), [15](#bib.bib15)\]. This is also an interesting option if the network’s deployed B-nodes in the area are not capable to execute the TLTA protocol.
Regarding the mobile users respectively device owners, the forceful activation of the LTE is a potential infringement of the proprietor’s power of disposal. A similar consideration holds for Trusted Computing technology proper, on any platform. It has led the TCG to mandate activation and take ownership procedures for the hardware security anchor TPM/MTM. These concepts can sensibly be lifted to the level of applications intimately connected with and depending on TC. Furthermore, procedures for user notification and acknowledgement should be combined with TLTA.
###
IV-C
Application Scenarios
The TLTA technology is of basic character, hence it is easy to envisage application scenarios in many sectors ranging from general m-commerce to public safety. We note but a few.
####
IV-C1
High-tech trade fair
Functional restriction policies can be enforced with exemptions. Consider a high-tech trade fair. pz is determined by the fair’s host so as to coincide with the physical entry barriers of the fair area. sp is a larger area so that visitors with TLTAC-enabled MTs will already be registered with TLTAC when passing the entry control. $`\text{P}_{\text{pz}}`$ stipulates that cameras and sound recording facilities of devices are to be switched off by LTE. They then can freely pass through, while unprotected (or non-networked devices, e.g., cameras) are detected at the entry barrier and are confiscated. Countering industrial espionage is another use case. A company may use TLTA to disable MT cameras within the borders of their campus, while keeping the MTs of visiting executives operative for communication purposes.
####
IV-C2
Sports stadium and concert hall
The voice commentary on a sports event is broadcast encrypted to mobile devices in a stadium, which is the pz. Only TLTA-enabled devices receive the decryption keys with the $`\text{P}_{\text{pz}}`$ data managed securely by LTE. The TLTSR and service provider can simply broadcast the content and has no need to operate own access control facilities. In a concert hall, the live audio broadcast should add to the experience of the audience, and can be personalised in terms of language and otherwise. This shall not be able to be received outside of the concert hall, to avoid bootlegging.
####
IV-C3
Mission critical communication
MTs in a disaster area can be configured to receive emergency and alarming messages. This can be used to inform the public or emergency action forces but also to locate them within the danger zone. Standard mobile devices can co-operate with the special ones of action forces and inter-work over network boundaries, e.g., with TETRA networks \[[16](#bib.bib16), [17](#bib.bib17)\].
####
IV-C4
Mobile gaming
A game spread out over the area of a city requires to adjust the device side policies based on narrow location, e.g., when the user has reached a certain waypoint.
These scenarios have in common that they could in principle also be realised by authorisation not based on location, e.g., organisational processes, yet with varying methods to reach the desired enforcement level. The benefit of TLTA is a unified authorisation method scaling with security requirements.
## V Conclusion
We have presented a novel system for the enforcement of authorisation policies according to location attributes. The salient features of i) largely de-central operation, and ii) strong enforcement level rest on a few key ideas. First, trust-enhanced devices combine functions of secure location triggering with with policy enforcement. Second, establishment of the protected zone policy enforcement is efficiently integrated in the handover between cells. Third, location trigger operation is two-tiered combining cell-based location with precision location methods like (A)-GPS. These concepts optimally leverage network and device resources and concurrently provide a high security level. Potential applications abound, some of which are only possible with TLTA or related concepts employing trusted devices. Implementation of TLTA seems feasible and economically viable, once TC-enabled devices become widely available.
## References
- \[1\] M. Anisetti, C.A. Ardagna, V. Bellandi, E. Damiani: *OpenAmbient: a Pervasive Access Control Architecture*. In: Long-Term and Dynamical Aspects of Information Security: Emerging Trends in Information and Communication Security, A. U. Schmidt, M. Kreutzer, R. Accorsi (eds.), Nova Science Publishers, Hauppauge, NY, 2007.
- \[2\] G. Wölfle, R. Hoppeand, D. Zimmermann, F. M. Landstorfer, *Enhanced Localization Technique within Urban and Indoor Environments based on Accurate and Fast Propagation Models*. European Wireless 2002, Firence, Italy, February 2002.
- \[3\] P. Wertz, R. Hoppe, D. Zimmermann, G. Wölfle, F. M. Landstorfer, *Enhanced Localization Technique within Urban and Indoor Environments*. 3rd COST 237 MCMMeeting, Guildford, UK, January 2002.
- \[4\] J. Vidal, J., M. Najar, R. Jativa, *High resolution time-of-arrival detection for wireless positioning systems*. In: Proc. of the 56th IEEE Vehicular Technology Conference, 2002.
- \[5\] Trusted Computing Group, *TCG Mobile Reference Architecture*. Specification version 1.0, Revision 1. 12 June 2007. <a href="https://www.trustedcomputinggroup.org/specs/mobilephone/" target="_blank">https://www.trustedcomputinggroup.org/specs/mobilephone/</a>
- \[6\]
H. Ekstrom, A. Furuskar, J. Karlsson, M. Meyer, S. Parkvall, J. Torsner, M. Wahlqvist, *Technical solutions for the 3G long-term evolution*, IEEE Communications Magazine 44, no.3, 38–45, March 2006.
- \[7\] Trusted Computing Group, *TPM Main Part 2 TPM Structures*, Specification Version 1.2 Revision 94, March 2006. *TPM Main Part 3 Commands*, Specification Version 1.2 Revision 94, March 2006.
- \[8\] A. P. Wise, *Memory encryption for digital video*. U.S. Patent US2007140477 (A1). 2007-06-20.
- \[9\] Data sheets from the ATMEL Corporation. <a href="http://www.atmel.com/dyn/products/datasheets.asp?family_id=646" target="_blank">http://www.atmel.com/dyn/products/datasheets.asp?family_id=646</a>
- \[10\]
J. Yang, L. Gao, Y. Zhang, *Improving Memory Encryption Performance in Secure Processors*. IEEE Transactions on Computers 54 (2005) 630 - 640.
- \[11\]
I. F. Akyildiz, W. Wang, *A dynamic location management scheme for next-generation multitier PCS systems*. IEEE Transactions on Wireless Communications 1, no.1, 178–189, January 2002.
- \[12\]
D. Morris, A. H. Aghvami, *Location Management Strategies for Cellular Overlay Networks — A Signaling Cost Analysis*. IEEE Transactions on Broadcasting 53, no.2, 480–493, June 2007.
- \[13\] C. Hartmann, O. Schlegelmilch, *Hierarchical cell structures with adaptive radio resource management*. 52nd IEEE Vehicular Technology Conference, 2000. VTS-Fall VTC 2000. Vol.4, pp. 1764–1771, 2000.
- \[14\] D. Molkdar, *Simulation results of a typical GSM picocellular system*. 52nd IEEE Vehicular Technology Conference, 2000. VTS-Fall VTC 2000. Vol.4, pp. 1590–1596, 2000.
- \[15\] T. Yamada, *Another double tier to avoid local congestion in a unified microcellular network*. IEEE 16th International Symposium on Personal, Indoor and Mobile Radio Communications, 2005. PIMRC 2005. Vol.4, pp. 2232–2236, 11-14 Sept. 2005
- \[16\]
A. K. Salkintzis, *Evolving public safety communication systems by integrating WLAN and TETRA networks*, IEEE Communications Magazine 44, no.1, 38–46, Jan. 2006.
- \[17\]
C. Edwards, *Wireless - Building on Tetra*, Engineering & Technology 1, no.2, 32–36, May 2006. |
0712.3037 | cs.CR | cyber | # Comments on “Improved Efficient Remote User Authentication Schemes”
Manik Lal Das
*Dhirubhai Ambani Institute of Information and Communication Technology*
*Gandhinagar - 382007, Gujarat, India.*
*Email: maniklal_das@daiict.ac.in*
###### Abstract
Recently, Tian et al. presented an article, in which they discussed some security weaknesses of Yoon et al.’s scheme and subsequently proposed two “improved” schemes. In this paper, we show that the Tian et al.’s schemes are insecure and vulnerable than the Yoon et al.’s scheme.
Keywords: Authentication, Smart card, Timestamp.
## 1 Introduction
Remote system authentication is a process by which a remote system gains confidence about the identity (or login request) of the communicating partner. Since the introduction of Lamport’s scheme \[[1](#bib.bib1)\], several new proposals and improvements on remote systems authentication \[[2](#bib.bib2)\], \[[3](#bib.bib3)\], \[[4](#bib.bib4)\], \[[5](#bib.bib5)\], \[[6](#bib.bib6)\] have been proposed. Recently, Tian et al. \[[7](#bib.bib7)\] presented an article by observing some flaws of the Yoon et al.’s scheme \[[8](#bib.bib8)\], and subsequently suggested two improved schemes. The basis of the Tian et al.’s observation on Yoon et al.’s scheme was on this assumption: If an attacker steals a user’s smart card and extracts the values stored in it through some means \[[9](#bib.bib9)\], \[[10](#bib.bib10)\] without being noticed, then the attacker can either masquerade as the user to forge a valid login request, or masquerade as the server to forge a valid reply message.
In this paper, we show that the Tian et al.’s schemes are insecure with the above mentioned arguments what they had considered, in fact, more vulnerable than \[[8](#bib.bib8)\]. The remainder of the paper is organized as follows. In the next section, we review the Tian et al.’s schemes. In section 3, we show the security weaknesses of the schemes. We conclude the paper with the section 4.
## 2 The Tian et al.’s Schemes
The schemes consists of four phases: Registration, Login, Authentication and Password change. The registration and password change phases are same for both the schemes.
Registration phase: A new user can register to the remote server by the following steps.
- R1.
A user $`U_{i}`$ submits his identity $`I\hspace{0pt}D_{i}`$ and password $`P\hspace{0pt}W_{i}`$ to the server ($`S`$) through a secure channel.
- R2.
Then $`S`$ chooses four distinct cryptographic one-way hash functions $`h\hspace{0pt}{( \cdot )}`$, $`h_{1}\hspace{0pt}{( \cdot )}`$, $`h_{2}\hspace{0pt}{( \cdot )}`$, and $`h_{3}\hspace{0pt}{( \cdot )}`$.
- R3.
$`S`$ computes $`R_{i} = {h\hspace{0pt}{({I\hspace{0pt}D_{i}},x_{s})}}`$, $`H_{i} = {h\hspace{0pt}{(R_{i})}}`$ and $`X_{i} = {R_{i} \oplus {h\hspace{0pt}{({I\hspace{0pt}D_{i}},{P\hspace{0pt}W_{i}})}}}`$, where $`\oplus`$ denotes the bitwise exclusive-OR operation.
- R4.
Then $`S`$ personalizes a smart card with $`< ID_{i},H_{i},X_{i},h{( \cdot )},h_{1}{( \cdot )},h_{2}{( \cdot )},h_{3}{( \cdot )} >`$ and sends it to $`U_{i}`$ in a secure manner.
Password change phase: This phase is invoked when a user $`U_{i}`$ wants to change his password from $`P\hspace{0pt}W_{i}`$ to $`P\hspace{0pt}W_{i}'`$. The user attaches his smart card to the card reader and enters $`P\hspace{0pt}W_{i}`$, then the smart card performs the following operations:
- P1.
Compute $`R_{i}' = {X_{i} \oplus {h\hspace{0pt}{({I\hspace{0pt}D_{i}},{P\hspace{0pt}W_{i}})}}}`$ and $`H_{i}' = {h\hspace{0pt}{(R_{i}')}}`$.
- P2.
Compare $`H_{i}'`$ with $`H_{i}`$. If they are equal, then the user enters a new password $`P\hspace{0pt}W_{i}'`$, otherwise it rejects the password change request.
- P3.
Compute $`X_{i}' = {R_{i} \oplus {h\hspace{0pt}{({I\hspace{0pt}D_{i}},{P\hspace{0pt}W_{i}'})}}}`$. Then, store $`X_{i}'`$ in smart card in place of $`X_{i}`$.
### 2.1 The First Scheme
This scheme uses the timestamp mechanism to avoid the replay attack (assuming the user and server time synchronization is proper).
Login phase: $`U_{i}`$ attaches his smart card to the card reader and enters password $`P\hspace{0pt}W_{i}^{\ast}`$. Then the smart card performs the following operations:
- LF1.
Compute $`R_{i}' = {X_{i} \oplus {h\hspace{0pt}{({I\hspace{0pt}D_{i}},{P\hspace{0pt}W_{i}^{\ast}})}}}`$ and $`H_{i}' = {h\hspace{0pt}{(R_{i}')}}`$.
- LF2.
Compare $`H_{i}'`$ with $`H_{i}`$. If they are equal, then the smart card proceeds to the next step, otherwise it terminates the operation.
- LF3.
Compute $`C_{1} = {h_{1}\hspace{0pt}{(S,{I\hspace{0pt}D_{i}},R_{i},T)}}`$, where $`T`$ is the timestamp.
- LF4.
$`U_{i}`$ sends the login request $`< ID_{i},T,C_{1} >`$ to $`S`$ over a public channel.
Authentication phase: Upon receiving the login request $`< ID_{i},T,C_{1} >`$, the server $`S`$ and the user $`U_{i}`$ perform the following steps for mutual authentication:
- AF1.
$`S`$ checks the validity of $`I\hspace{0pt}D_{i}`$ and $`T`$. If both are correct then proceeds to the next step, otherwise rejects the login request.
- AF2.
$`S`$ computes $`R_{i} = {h\hspace{0pt}{({I\hspace{0pt}D_{i}},x_{s})}}`$ and checks whether $`C_{1} = {h_{1}\hspace{0pt}{(S,{I\hspace{0pt}D_{i}},R_{i},T)}}`$. If this check holds, $`S`$ assures that $`U_{i}`$ is authentic and proceeds to the next step, otherwise it rejects the request.
- AF3.
$`S`$ computes $`C_{2} = {h_{2}\hspace{0pt}{({I\hspace{0pt}D_{i}},S,R_{i},T')}}`$, where $`T'`$ is a timestamp. Then, $`S`$ sends $`< T',C_{2} >`$ back to $`U_{i}`$ through the public channel.
- AF4.
Upon receiving $`S`$’s response message $`< T',C_{2} >`$, $`U_{i}`$’s smart card first checks the validity of $`T'`$ and then whether $`C_{2} = {h_{2}\hspace{0pt}{({I\hspace{0pt}D_{i}},S,R_{i},T')}}`$. If these checks hold, $`U_{i}`$ assures the authenticity of $`S`$ and the mutual authentication is done, otherwise it rejects the connection.
- AF5.
Once the mutual authentication is completed, $`U_{i}`$ and $`S`$ use $`h_{3}\hspace{0pt}{({I\hspace{0pt}D_{i}},S,R_{i},T,T')}`$ as the session key.
### 2.2 The Second Scheme
This scheme uses a nonce based challenge-response mechanism, so it avoids the time synchronization problem.
Login phase: $`U_{i}`$ attaches his smart card to the card reader and enters password $`P\hspace{0pt}W_{i}`$. Then the smart card performs the following operations:
- LS1.
Compute $`R_{i}' = {X_{i} \oplus {h\hspace{0pt}{({I\hspace{0pt}D_{i}},{P\hspace{0pt}W_{i}})}}}`$ and $`H_{i}' = {h\hspace{0pt}{(R_{i}')}}`$.
- LS2.
Compare $`H_{i}'`$ with $`H_{i}`$. If they are equal, proceeds to the next step, otherwise it terminates the operation.
- LS3.
Send the login request $`< ID_{i},N_{i} >`$ to $`S`$ over a public channel, where $`N_{i}`$ is a nonce selected by $`U_{i}`$.
Authentication phase: Upon receiving the login request $`< ID_{i},N_{i} >`$, the server $`S`$ and the user $`U_{i}`$ perform the following steps for mutual authentication:
- AS1.
$`S`$ checks the validity of $`I\hspace{0pt}D_{i}`$.
- AS2.
$`S`$ chooses a nonce $`N_{s}`$, computes $`R_{i} = {h\hspace{0pt}{({I\hspace{0pt}D_{i}},x_{s})}}`$, $`C_{1} = {h_{1}\hspace{0pt}{(S,{I\hspace{0pt}D_{i}},R_{i},N_{i},N_{s})}}`$ and sends $`< C_{1},N_{s} >`$ to $`U_{i}`$ over a public channel.
- AS3.
Upon receiving $`< C_{1},N_{s} >`$, $`U_{i}`$ checks whether $`C_{1} = {h_{1}\hspace{0pt}{(S,{I\hspace{0pt}D_{i}},R_{i},N_{i},N_{s})}}`$. If this check holds correct, $`U_{i}`$ assures the authenticity of $`S`$, otherwise terminates the operation.
- AS4.
$`U_{i}`$ computes $`C_{2} = {h_{2}\hspace{0pt}{({I\hspace{0pt}D_{i}},S,R_{i},N_{s},N_{i})}}`$ and sends it to $`S`$.
- AS5.
Upon receiving $`C_{2}`$, $`S`$ checks whether $`C_{2} = {h_{2}\hspace{0pt}{({I\hspace{0pt}D_{i}},S,R_{i},N_{s},N_{i})}}`$. $`U_{i}`$ authentic if the check passes and the mutual authentication is done, otherwise $`S`$ terminates the operation.
- AS6.
After the mutual authentication, the user and the server use $`h_{3}\hspace{0pt}{({I\hspace{0pt}D_{i}},S,R_{i},N_{i},N_{s})}`$ as the session key.
## 3 Security Weaknesses
The basis of the following attacks is based on this risk of smart card stored information:
A legitimate user could extract the values stored in smart card by some means \[[9](#bib.bib9)\], \[[10](#bib.bib10)\], then he/she could act as the role of server to register any number of users. We note that the Tian et al.’s scheme also assumed a similar risk.
1. Attacks by a legitimate user
In the registration phase, $`X_{i} = {R_{i} \oplus {h\hspace{0pt}{({I\hspace{0pt}D_{i}},{P\hspace{0pt}W_{i}})}}}`$ is stored in $`U_{i}`$’s smart card. Once $`U_{i}`$ extracts $`X_{i}`$ from his smart card by some means \[[9](#bib.bib9)\], \[[10](#bib.bib10)\], then he/she can easily get $`R_{i}`$ by computing $`R_{i} = {X_{i} \oplus {h\hspace{0pt}{({I\hspace{0pt}D_{i}},{P\hspace{0pt}W_{i}})}}}`$. After that, no remote server is required to register a new user. Now, $`U_{i}`$ who has $`R_{i}`$, could register any number of users by distributing $`R_{i}`$ and $`I\hspace{0pt}D_{i}`$. In fact, smart card and password are not required at all to login $`S`$ those who got $`R_{i}`$ and $`I\hspace{0pt}D_{i}`$ from $`U_{i}`$. Because, a valid login message is $`< ID_{i},T,C_{1} >`$, where $`T`$ is a timestamp (for the first scheme) and $`C_{1} = {h_{1}\hspace{0pt}{(S,{I\hspace{0pt}D_{i}},R_{i},T)}}`$. For the second scheme, the challenge-response comprises with the secret $`R_{i}`$ only, other parameters are public. Therefore, the server secret is virtually compromised by a legitimate user’s smart card.
2. Attacks by an adversary
Suppose an attacker steals $`U_{i}`$’s smart card and intercepts $`C_{1} = {(S,{I\hspace{0pt}D_{i}},R_{i},T)}`$ from a valid login request. Now the attacker extracts the information stored in the smart card and launches an offline guessing attacks of $`P\hspace{0pt}W_{i}`$ in order to obtain the value of $`R_{i}`$. The attacker guesses a password and obtains an $`R_{i}^{\ast}`$, and then checks whether $`C_{1} = {h_{1}\hspace{0pt}{({S\hspace{0pt}I\hspace{0pt}D_{i}},R_{i}^{\ast},T)}}`$. Once the guess succeeds, then the attacker has a valid $`R_{i}`$ and can create any number of valid login request.
3. No two-factor authentication
Two-factor authentication is a technique that requires two independent factors (e.g. password, smart card) to establish identity and privileges. Common implementations of two-factor authentication use ‘something you know: password’ as one of the two factors, and use either ‘something you have: smart card’ or ‘something you are: biometric’ as the other factor. A common example of two-factor authentication is a bank card (credit card, debit card); the card itself is the physical item, and the personal identification number (PIN) is the data that goes with it.
In Tian et al.’s scheme, we observe that once a party has information of $`I\hspace{0pt}D_{i}`$ and $`R_{i}`$, then he does not require password and a valid smart card at all. Without password and smart card, one can easily pass the mutual authentication and establish the session key. Therefore, the schemes lack two-factor authentication.
## 4 Conclusion
The threat of smart card security \[[9](#bib.bib9)\], \[[10](#bib.bib10)\], \[[11](#bib.bib11)\] is a crucial concern, where some secret information is stored in the memory of smart cards. However, to the best of my knowledge, one can still use smart card to store some secret data by considering the applications requirement and scope/value of the secret information stored in the smart card. It is also important to judge the financial cost and time to extract the secret data from the smart card. If the cost as well as time is tolerable or higher than the cost of the secret inside the smart card, then one can take that risk while using smart card to store some secret data. If extracting a secret from the card leads to collapse the whole system (e.g. Tian et al.’s schemes) then definitely some additional counter measure should be taken while designing the scheme. Of course, smart card vendors are quite aware of these threats and they are also taking counter measure continuously to safe guard the cards security.
We have shown that the Tian et al.’s scheme is insecure by several weaknesses. Just by extracting a secret data from a smart card can collapse the whole system’s security.
## References
- \[1\]
L. Lamport, “Password authentication with insecure communication,” Communication of ACM, vol.24, pp.770–772, 1981.
- \[2\]
D. P. Jablon, “Strong password-only authenticated key exchange,” ACM Computer Communications Review, vol.26, no.5, pp.5–26, 1996.
- \[3\]
M. S. Hwang, and L. H. Li, “A New Remote User Authentication Scheme Using Smart Cards,” IEEE Trans. on Consumer Electron., vol.46, no.1, pp.28–30, 2000.
- \[4\]
C. C. Lee, L. H. Li, and M. S. Hwang, “A Remote User Authentication Scheme Using Hash Functions,” ACM Operating Systems Review, vol.36, no.4, pp. 23–29, 2002.
- \[5\]
A. K. Awasthi, and S. Lal, “An enhanced remote user authentication scheme using smart cards,” IEEE Trans. Consumer Electron., vol.50, no.2, pp.583–586, 2004.
- \[6\]
M. L. Das, A. Saxena, and V. P. Gulati, “A Dynamic ID-based Remote User Authentication Scheme,” IEEE Trans. on Consumer Electron., vol.50, no.2, pp. 629–631, 2004.
- \[7\]
X. Tian, R. W. Zhu, and D. S. Wong, “Improved Effcient Remote User Authentication Schemes,” International Journal of Network Security, vol.4, no.2, PP.149 -154, 2007.
- \[8\]
E. J. Yoon, E. K. Ryu, and K. Y. Yoo, “Efficient remote user authentication scheme based on generalized ElGamal signature scheme,” IEEE Trans. on Consumer Electron., vol.50, no.2, pp.568–570, 2004.
- \[9\]
P. Kocher, J. Jaffe, and B. Jun, “Differential power analysis,” in The Proceedings of Advances in Cryptology (CRYPTO’99), LNCS 1666, pp.388–397, 1999.
- \[10\]
T. S. Messerges, E. A. Dabbish, and R. H. Sloan, “Examining smart-card security under the threat of power analysis attacks,” IEEE Trans. on Computers, vol.51, no.5, pp.541–552, 2002.
- \[11\]
M. Joye, and F. Olivier, “Side-channel analysis,” Encyclopedia of Cryptography and Security: Kluwer Academic Publishers, pp.571–576, 2005 |
0712.2054 | cs.NI | cyber | # Distributed Fair Scheduling Using Variable Transmission Lengths in Carrier-Sensing-based Wireless Networks
Libin Jiang and Jean Walrand
Dept. Electrical Engineering and Computer Sciences, University of California, Berkeley
{ljiang, wlr}@eecs.berkeley.edu
\* This work was supported in part by NSF grants NeTS-WN 0435478 and NeTS-FIND 0627161.
###### Abstract
The fairness of IEEE 802.11 wireless networks (including Wireless LAN and Ad-hoc networks) is hard to predict and control because of the randomness and complexity of the MAC contentions and dynamics. Moreover, asymmetric channel conditions such as those caused by capture and channel errors often lead to severe unfairness among stations. In this paper we propose a novel distributed scheduling algorithm that we call VLS, for “variable-length scheduling”, that provides weighted fairness to all stations despite the imperfections of the MAC layer and physical channels. Distinct features of VLS include the use of variable transmission lengths based on distributed observations, compatibility with 802.11’s contention window algorithm, opportunistic scheduling to achieve high throughput in time-varying wireless environments, and flexibility and ease of implementation. Also, VLS makes the throughput of each station more smooth, which is appealing to real-time applications such as video and voice. Although the paper mostly assumes 802.11 protocol, the idea generally applies to wireless networks based on CSMA (Carrier Sensing Multiple Access).
###### Index Terms:
Distributed Fair Scheduling, Variable Transmission Lengths, Carrier Sensing Multiple Access, IEEE 802.11, Wireless Channel
## I INTRODUCTION
In this paper, we propose a simple distributed scheduling algorithm that provides weighted fairness in IEEE 802.11 \[[1](#bib.bib1)\] wireless networks, despite the unpredictability of the 802.11 MAC layer and physical channels.
In 802.11 wireless networks, MAC-layer contention, dynamics and bandwidth allocation are hard to predict. For such networks, the fixed-point model in \[[4](#bib.bib4)\] gives a method for computing the long-term throughput of the Binary Exponential Backoff (BEB) algorithm \[[1](#bib.bib1)\]. However, the short-term dynamics and unfairness are quite unpredictable. In a certain period, a station may randomly backoff more than others, and therefore have a smaller chance of winning the channel, which in turn makes that station backoff even more.
Meanwhile, the BEB amplifies the unfairness caused by the impairments of the wireless channels. This aggravation is an unintentional side effect of BEB that was designed to reduce collisions, not to guarantee fairness. The following two effects cause the unfairness:
- \(1\)
Capture: Capture occurs when the signals from different transmitters have very different strengths at a receiver \[[8](#bib.bib8)\]. For instance, a ratio of 2 in distances from the stations to the AP can lead to approximately a ratio of 16 in received signal strengths. When more than one station transmit packets to the AP at the same time, the AP may be able to capture and correctly decode the packet from the closer station, while ignoring the other packets. This effect increases the aggregate throughput since the AP receives one packet even when multiple transmissions overlap in time. However, capture may result in unfairness since the stations that are further away backoff more with the BEB algorithm, and consequently obtain much less throughput than closer stations \[[7](#bib.bib7)\].
- \(2\)
Channel errors: In addition to packet collisions, channel errors are another important cause of packet loss. A more lossy channel to the AP drops more packets because of channel errors. The transmitting station interprets all packet losses as collisions and doubles its contention window. Accordingly, the BEB algorithm magnifies the asymmetry of the lossy channels. To alleviate this problem, reference \[[11](#bib.bib11)\] describes a way to differentiate the two kinds of packet losses (due to collisions or channel errors). The algorithm proposed in this paper provides a simpler solution.
With a more complicated MAC, IEEE 802.11e \[[2](#bib.bib2)\] provides Differentiated Service (DiffServ), by adopting different minimum Contention Windows ($`C\hspace{0pt}W_{m\hspace{0pt}i\hspace{0pt}n}`$) and inter-frame Spaces (IFS) for different service classes such as voice, video and data. This protocol provides relative performance differentiation among different classes: the classes with smaller $`C\hspace{0pt}W_{m\hspace{0pt}i\hspace{0pt}n}`$ and IFS have a relative priority over others. To evaluate the performance of 802.11e, reference \[[14](#bib.bib14)\] provides a simulation study; while reference \[[13](#bib.bib13)\] uses an analytical model (a Markov chain) to find the saturated throughput of 802.11e. However, the model there is quite complicated, indicating that the “amount” of relative priority is hard to quantify and control. For instance, it is not clear how much more bandwidth the protocol gives to video with a particular setting of $`C\hspace{0pt}W_{m\hspace{0pt}i\hspace{0pt}n}`$ and IFS, nor how to adjust the amount of priority by varying these parameters.
In this paper, we describe a simple, easy to implement, distributed fair scheduling algorithm that we call VLS, for“variable-length scheduling,” to cope with the above problems. VLS provides exact weighted fairness despite the unpredictability of the 802.11 MAC layer and physical channels.
## II Variable-Length Scheduling (VLS): Bringing Order To Random Access
In this section, we assume that there is only one collision domain. That is, each station can sense the transmissions of other stations. (We consider the case of multiple collision domains in section [V](#S5 "V VLS in Multiple Collision domain ‣ Distributed Fair Scheduling Using Variable Transmission Lengths in Carrier-Sensing-based Wireless Networks").) There are two versions of the scheduling algorithm: without and with an access point (AP). The latter is an adaptation of the former that utilizes the AP to simplify the algorithm.
###
II-A
Distributed algorithm
The algorithm is based on the concept of “virtual slot.” By definition, a station sees a virtual slot when it senses a collision, a burst of transmissions (i.e., one DATA-ACK exchange, or a series of DATA-ACK exchanges separated by SIFS), or when it transmits a burst of packets itself. Mini-slots are not counted as virtual slots. In other words, a station counts a virtual slot whenever if senses the channel as “idle” for an interval equal to DIFS (DIFS$`>`$SIFS \[[1](#bib.bib1)\]) and is involved in a contention process (i.e., when the station’s backoff counter goes down until it transmits a packet or senses other transmissions). In Fig. [1](#S2.F1 "Figure 1 ‣ II-A Distributed algorithm ‣ II Variable-Length Scheduling (VLS): Bringing Order To Random Access ‣ Distributed Fair Scheduling Using Variable Transmission Lengths in Carrier-Sensing-based Wireless Networks"), for example, there are 3 virtual slots. “Virtual slots” are similar to “busy slots” except that a burst of transmissions is counted as one “virtual slot.”
[Refer to caption](/html/0712.2054/assets/x1.png)
Figure 1: Virtual slots. (There are 3 virtual slots in this figure.)
The notion of “virtual slot” is particularly useful because every station in a single collision domain sees the same number of virtual slots, assuming that the stations are always backlogged. (If not, the station starts the algorithm only when it has a backlog and stops it when its backlog is cleared.) Therefore, virtual slots can serve as a “clock” for scheduling. We design the distributed algorithm as follows.
- •
Each station $`j`$ is assigned a “weight” $`W_{j}`$ \[[15](#bib.bib15)\]. (If there are multiple flows outgoing from station $`j`$, then let $`W_{j}`$ be the sum of the weights of all individual flows.) And each station $`j`$ keeps track of a value $`m_{j}`$ that is initially equal to 1.
- •
If station $`j`$ gets an ACK after it transmits a packet, it keeps transmitting a burst of $`m_{j}\hspace{0pt}W_{j}`$ packets separated by SIFS and then resets $`m_{j}`$ to the value 1.
- •
If it does not get an ACK after it transmits a packet, or if it does not get to transmit (i.e., it does not win a contention), station $`j`$ increments $`m_{j}`$ by one whenever it sees a virtual slot.
For example, Fig. [2](#S2.F2 "Figure 2 ‣ II-A Distributed algorithm ‣ II Variable-Length Scheduling (VLS): Bringing Order To Random Access ‣ Distributed Fair Scheduling Using Variable Transmission Lengths in Carrier-Sensing-based Wireless Networks") shows the process with 2 stations with weights $`w_{1}`$ and $`w_{2}`$ respectively. In the figure, each block represents a virtual slot (aligned across different stations). The number in each block indicates the transmission length in that virtual slot. (Note that the size of a block here does not reflect the actual length of the virtual slot.) Assume station 1 starts transmission with the 1st virtual slot, while station 2 starts transmission with the 2nd.
[Refer to caption](/html/0712.2054/assets/x2.png)
Figure 2: Variable-length Scheduling. ($`w_{1}`$, $`w_{2}`$ are the weights. Each block represents a variable-length virtual slot. The number in a block indicates the transmission length.)
In a time period when station $`j`$ is backlogged, we can see that the total number of packets it has transmitted is equal to the number of virtual slots it has seen so far times the weight $`W_{j}`$. (So station $`j`$ virtually transmits $`W_{j}`$ packets per slot. This is why we use the name “virtual slot”.) Since all stations see the same virtual slots, the bandwidth allocation is weighted fair. That is $`T_{j}/W_{j}`$ is the same for all the stations where $`T_{j}`$ is the average rate of packets that station $`j`$ transmits.
Note that VLS guarantees the weighted fairness, no matter what happens in the MAC and physical layers. Thus, VLS automatically adjusts for the randomness of the MAC protocol and the asymmetry of the physical channels.
###
II-B
Algorithm with an AP
In a wireless LAN with an AP, the above algorithm can be adapted so that the client stations need not count the virtual slots. In this variation, the AP counts the virtual slots for the stations and piggybacks that count in the MAC-layer ACKs to the stations. The algorithm works as follows.
- •
The AP keeps counting the virtual slots. It increments the count ($`v + +`$) for each virtual slot.
- •
A station can start and stop the algorithm at any time. When station $`j`$ starts, it contends for the channel and sends the first $`W_{j}`$ packets to the AP. In one ACK, the AP piggybacks the current value of $`v`$, denoted as $`v_{i}^{j}`$. The next time station $`j`$ wins the channel, it sends $`W_{j}`$ packets first. The AP, again, piggybacks the current value of $`v`$, denoted as $`v_{i + 1}^{j}`$. Then, the station sends $`W_{j}\hspace{0pt}{({v_{i + 1}^{j} - v_{i}^{j} - 1})}`$ more packets in the burst. Since $`{v_{i + 1}^{j} - v_{i}^{j} - 1} \geq 0`$, station $`j`$ sends at least $`W_{j}`$ packets per burst.
###
II-C
Considerations on Burst Length
Suppose that, at each virtual slot, every station $`i`$ has a probability $`p_{i}`$ of winning the channel. Then on average, station $`i`$ accumulates $`W_{i}/p_{i}`$ units of “credit” before it wins the channel (since the average number of virtual slots it waits for is $`1/p_{i}`$). It can then spent the credits by transmitting (on average) $`W_{i}/p_{i}`$ packets. When the number of active stations $`N`$ increases in a wireless networks, $`p_{i}`$ decreases (approximately $`\propto {1/N}`$). As a result, the average burst length increases, thus causing more delays for all the stations.
To avoid this effect, we define a system-wide parameter $`c > 0`$, called “speed of the clock”, and we modify the protocol as follows:
- •
Instead of transmitting $`m_{j}\hspace{0pt}W_{j}`$ packets in a burst as in subsection [](#S2.SS1 "II-A Distributed algorithm ‣ II Variable-Length Scheduling (VLS): Bringing Order To Random Access ‣ Distributed Fair Scheduling Using Variable Transmission Lengths in Carrier-Sensing-based Wireless Networks")
II-A
, station $`j`$ transmits $`{c \cdot m_{j}}\hspace{0pt}W_{j}`$ packets. If $`{c \cdot m_{j}}\hspace{0pt}W_{j}`$ is not an integer, then it transmits $`\left\lfloor {{c \cdot m_{j}}\hspace{0pt}W_{j}} \right\rfloor`$ packets, and saves the extra credits $`{{c \cdot m_{j}}\hspace{0pt}W_{j}} - \left\lfloor {{c \cdot m_{j}}\hspace{0pt}W_{j}} \right\rfloor`$ for the next time. Essentially, $`c`$ controls the speed of clock in the whole network. (Therefore, the delay is proportional to $`c`$.)
- •
The stations can adjust the value of parameter $`c`$ in several different ways:
- –
If station $`j`$ knows the number of backlogged stations $`N`$, then it can choose $`c_{j} = {1/N}`$ and broadcast $`c_{j}`$ to the network. The other stations will then follow the parameter.
- –
Common TCP flows are usually not sensitive to the burst length and delay. But if a station (say, station $`j`$) has delay-sensitive flows and some other stations’ burst lengths are causing too much delay to it, it computes a new value of $`c_{j}`$ and broadcasts it to the network. (Assume the current average delay for station $`j`$ is $`d_{j}`$, and its targeted delay is $`D_{j}`$, then set $`c_{j} = {{c_{0} \cdot D_{j}}/d_{j}}`$, where $`c_{0}`$ is the current parameter of the system.)
- –
If the network has an AP, the AP can act as a controller to adjust $`c`$.
- –
If there are more than one stations broadcasting $`c_{j}`$, each station follow the lowest $`c_{j}`$ it has received (i.e., $`c = {\min_{j}c_{j}}`$).
- •
Further details about the implementation of broadcasting:
- –
Station $`j`$ embeds $`c_{j}`$ in a packet (or piggybacked in a usual data packet), along with its ID/address.
- –
To increase reliability, this packet can be repeated multiple times. Also, although stations in a single collision domain may “carrier-sense” each other, they may not be able to “decode” the packets of each other. Therefore, the packets containing $`c_{j}`$ are transmitted with a higher power, or a lower data rate, than usual packets.
- –
If a station has broadcast a $`c_{j}`$ before and wants to update it, it simply broadcasts the new $`c_{j}`$. Since other stations know the ID of the sender, they update the old parameter of the same sender, and follow the lowest $`c_{j}`$ in their records.
In addition, we can impose a limit on the burst length, $`B_{j}`$, of each station $`j`$. In this case, station $`j`$ can transmit up to $`m\hspace{0pt}i\hspace{0pt}n\hspace{0pt}{(B_{j},{{c \cdot m_{j}}\hspace{0pt}W_{j}})}`$ packets in a burst (and the remaining credit is left for future transmissions). This mechanism smooths out the randomness of the burst lengths, which may otherwise be (randomly) long or short, even if the network has a proper value of $`c`$. But in this situation, $`c`$ needs to be small enough to avoid the instability of credits (i.e., the remaining credits should not go to infinity). In particular, a simple inequality needs to be satisfied. We discuss this issue in Section [IV](#S4 "IV Stability of the Credits ‣ Distributed Fair Scheduling Using Variable Transmission Lengths in Carrier-Sensing-based Wireless Networks").
###
II-D
Generalized Fair Scheduling
As mentioned before, virtual slots act as a clock for scheduling. Using this synchronization mechanism, VLS has the flexibility to achieve many forms of fairness. The scheduler above uses the number of packets as a fairness metric. VLS can also provide weighted fairness in terms of the number of bits or the “air-time” occupied by different stations. If different stations use very different data rates (e.g., 1Mbps vs. 11Mbps) in a shared-medium wireless network, providing fairness in terms of bits leads to very low efficiency (throughput) of the whole network \[[9](#bib.bib9)\]. In this case, \[[9](#bib.bib9)\] shows that allocating equal air-time to different stations strikes a good balance between fairness and efficiency, and is actually equivalent to achieving proportional fairness \[[6](#bib.bib6)\].
## III Performance Evaluation
###
III-A
Short-term fairness
In a wireless LAN, or an ad-hoc network with a single collision domain, the long-term saturated throughput should the same for all stations, by symmetry \[[4](#bib.bib4)\]. However, the collisions and the dynamics of contention windows (with BEB) are quite unpredictable, leading to fluctuations of short-term throughput (Fig. [3](#S3.F3 "Figure 3 ‣ III-B Weighted Fairness ‣ III Performance Evaluation ‣ Distributed Fair Scheduling Using Variable Transmission Lengths in Carrier-Sensing-based Wireless Networks")(a)). Also, the volumes of data that different stations send drift away from each other (Fig. [3](#S3.F3 "Figure 3 ‣ III-B Weighted Fairness ‣ III Performance Evaluation ‣ Distributed Fair Scheduling Using Variable Transmission Lengths in Carrier-Sensing-based Wireless Networks")(a)). This means that, although the average throughput of the different stations are equal in the long term, the average throughputs may differ over a considerable time window (20 sec in the figure). With VLS, short-term and long-term fairness have been clearly improved (Fig. [3](#S3.F3 "Figure 3 ‣ III-B Weighted Fairness ‣ III Performance Evaluation ‣ Distributed Fair Scheduling Using Variable Transmission Lengths in Carrier-Sensing-based Wireless Networks")(b)).
###
III-B
Weighted Fairness
Without VLS, weighted fairness may be implemented by using different $`C\hspace{0pt}W_{m\hspace{0pt}i\hspace{0pt}n}`$’s. Approximately, the throughput of an individual station is inversely proportional to its minimum CW, assuming that each station has the same average packet size and use the same IFS (Fig. [4](#S3.F4 "Figure 4 ‣ III-B Weighted Fairness ‣ III Performance Evaluation ‣ Distributed Fair Scheduling Using Variable Transmission Lengths in Carrier-Sensing-based Wireless Networks")(a)). However, the approximation is not accurate (especially when $`C\hspace{0pt}W_{m\hspace{0pt}i\hspace{0pt}n}`$’s are small), and one can expect it to be vulnerable to physical layer factors such as capture effect and channel errors. With VLS, the weighted fairness is exact, and easy to adjust and control (Fig. [4](#S3.F4 "Figure 4 ‣ III-B Weighted Fairness ‣ III Performance Evaluation ‣ Distributed Fair Scheduling Using Variable Transmission Lengths in Carrier-Sensing-based Wireless Networks")(b)).
[Refer to caption](/html/0712.2054/assets/x3.png)
\(a\)
w/0 VLS
[Refer to caption](/html/0712.2054/assets/x4.png)
\(b\)
VLS
Figure 3: Improvement of short-term and long-term fairness (10 stations, $`W_{j} = {1,{\forall j}}`$) (Throughout the paper, “S1” means “Station 1”, etc.)
[Refer to caption](/html/0712.2054/assets/x5.png)
\(a\)
w/o VLS (The vector of $`C\hspace{0pt}W_{m\hspace{0pt}i\hspace{0pt}n}`$’s of the 10 stations is (128 64 128 64 42 32 128 64 26 42))
[Refer to caption](/html/0712.2054/assets/x6.png)
\(b\)
VLS (The vector of weights of the 10 stations is (1 2 1 2 3 4 1 2 5 3))
Figure 4: Weighted Fairness
###
III-C
Solving Unfairness Problems
Besides improving short-term fairness and providing weighted fairness, our algorithm can readily solve many other unfairness problems in wireless networks.
####
III-C1
Unfairness due to Capture Effect
In a WLAN, different stations may have very different distances from the AP. Also, the channel qualities between stations and the AP may differ greatly even if the distances are similar, for instance because of multipath or obstructions. The above effects result in different signal strengths from different stations as received by the AP. When more than one stations transmit packets to the AP at the same time , so that these transmissions collide, the AP may still be able to capture and correctly decode the packet with the strongest signal, and send back an ACK. This feature is helpful in terms of the aggregate throughput since one packet is received even in the event of a collision, but may exacerbate the unfairness. The weaker stations tend to backoff more with the BEB algorithm, and therefore obtain much less throughput than stronger ones (see Fig. [5](#S3.F5 "Figure 5 ‣ III-C1 Unfairness due to Capture Effect ‣ III-C Solving Unfairness Problems ‣ III Performance Evaluation ‣ Distributed Fair Scheduling Using Variable Transmission Lengths in Carrier-Sensing-based Wireless Networks")(a)).
If our VLS algorithm is enabled, it can overcome the unfairness problem, as well as *retaining the throughput benefit* provided by capture. Since both strong and weak stations share the same view of the virtual slots, they share the bandwidth in a fair way (see Fig. [5](#S3.F5 "Figure 5 ‣ III-C1 Unfairness due to Capture Effect ‣ III-C Solving Unfairness Problems ‣ III Performance Evaluation ‣ Distributed Fair Scheduling Using Variable Transmission Lengths in Carrier-Sensing-based Wireless Networks")(b)).
[Refer to caption](/html/0712.2054/assets/x7.png)
\(a\)
w/o VLS
[Refer to caption](/html/0712.2054/assets/x8.png)
\(b\)
VLS
Figure 5: Overcoming unfairness due to capture effect (S1 is the strong station. When S1’s packets collide with other stations’ packets, S1’s packet is captured)
####
III-C2
Unfairness due to Channel Errors
In wireless networks, in addition to packet collisions, channel errors are another important cause of packet losses. If a station has a lossy channel to the AP, its packets are dropped with higher probability because of channel errors. Similarly to the capture effect, these losses also result in an asymmetry among different stations, and therefore in unfairness, aggravated by the BEB algorithm. This effect is shown in Fig. [7](#S3.F7 "Figure 7 ‣ III-C2 Unfairness due to Channel Errors ‣ III-C Solving Unfairness Problems ‣ III Performance Evaluation ‣ Distributed Fair Scheduling Using Variable Transmission Lengths in Carrier-Sensing-based Wireless Networks")(a) where only station 1 suffers from a loss probability $`p = 0.15`$ caused by channel errors.
A simple model of a time-varying wireless channel is a Markov chain with two states: “good” and “bad”. For simplicity, we assume that in the “good” state all packets can be received and that in the “bad” state all packets are dropped. Each state has an exponentially-distributed duration before it transits to the other state. Fig. [6](#S3.F6 "Figure 6 ‣ III-C2 Unfairness due to Channel Errors ‣ III-C Solving Unfairness Problems ‣ III Performance Evaluation ‣ Distributed Fair Scheduling Using Variable Transmission Lengths in Carrier-Sensing-based Wireless Networks") shows the state transition diagram. In the scenario simulated, only station 1 has such a noisy channel, with $`\lambda_{g} = {{20/s}\hspace{0pt}e\hspace{0pt}c}`$, $`\lambda_{b} = {{113/s}\hspace{0pt}e\hspace{0pt}c}`$ (therefore the average loss probability is $`p = 0.15`$); and other stations have perfect channels ($`p = 0`$). It turns out that our algorithm not only maintains fairness, but also utilize the channel “*opportunistically*” to get a high throughput: When the weak station wins a contention, but meets the “bad” state, it transmits a packet, without receiving the ACK, and relinquishes the channel immediately. When it wins a contention and meets the “good” state, it can transmit a burst of packets. Since the channel is likely to stay in the “good” state for some time, the station has the opportunity to compensate for its past losses. If at some point the state goes back to “bad”, the station stops immediately and waits for the next opportunity (see Fig. [7](#S3.F7 "Figure 7 ‣ III-C2 Unfairness due to Channel Errors ‣ III-C Solving Unfairness Problems ‣ III Performance Evaluation ‣ Distributed Fair Scheduling Using Variable Transmission Lengths in Carrier-Sensing-based Wireless Networks")(b)). In other words, the weak station transmits more when the channel is good, and less when the channel is bad, thus utilizing the noisy channel more efficiently. As a result, the total throughput in Fig. [7](#S3.F7 "Figure 7 ‣ III-C2 Unfairness due to Channel Errors ‣ III-C Solving Unfairness Problems ‣ III Performance Evaluation ‣ Distributed Fair Scheduling Using Variable Transmission Lengths in Carrier-Sensing-based Wireless Networks")(b) is only slightly less than that in Fig. [7](#S3.F7 "Figure 7 ‣ III-C2 Unfairness due to Channel Errors ‣ III-C Solving Unfairness Problems ‣ III Performance Evaluation ‣ Distributed Fair Scheduling Using Variable Transmission Lengths in Carrier-Sensing-based Wireless Networks")(c), where the channels are perfect for all stations.
[Refer to caption](/html/0712.2054/assets/x9.png)
Figure 6: Markov Channel Model
It is evident that if the average duration of the “good” state is too short, the weak station may still receive unfair throughput. We can derive the condition under which fairness is guaranteed (see Section [IV](#S4 "IV Stability of the Credits ‣ Distributed Fair Scheduling Using Variable Transmission Lengths in Carrier-Sensing-based Wireless Networks")). Nevertheless, VLS always improves the fairness.
When the channel of one station is significantly worse than that of the others, providing throughput-fairness to different stations may drag down the total throughput of the network. (This is a tradeoff between fairness and efficiency.) In that case, providing time-fairness would be more suitable. VLS has the flexibility to provide time-fairness, as mentioned in subsection [](#S2.SS4 "II-D Generalized Fair Scheduling ‣ II Variable-Length Scheduling (VLS): Bringing Order To Random Access ‣ Distributed Fair Scheduling Using Variable Transmission Lengths in Carrier-Sensing-based Wireless Networks")
II-D
.
In the above, we have assumed that the stations do not know the state of the channel before transmitting. If the channel state information is known in advance, the station can avoid transmitting in the “bad” state. In that case, the throughput performance is further improved.
[Refer to caption](/html/0712.2054/assets/x10.png)
\(a\)
w/o VLS, only S1 suffers a loss rate $`p`$=0.15 due to channel errors
[Refer to caption](/html/0712.2054/assets/x11.png)
\(b\)
With VLS, $`p`$=0.15 for S1
[Refer to caption](/html/0712.2054/assets/x12.png)
\(c\)
With VLS, $`p`$=0 for all stations
Figure 7: Overcoming unfairness due to channel errors (Note that in (b) fairness is achieved, also, the throughput is close to (c) w/o channel errors.)
## IV Stability of the Credits
In VLS, each active station keeps accumulating and spending credit, where the accumulated credit is proportional to a station’s weight, the number of virtual slots it has observed, and the speed of clock $`c`$. In this section, we consider whether the credits are stable, that is, whether the credits of some stations keep increasing and go to infinity. If that happens, the bandwidth allocation may not be fair, since the extra credits of some stations are not spent.
First, if there is no limit on the length of transmission at each burst, the credits must be stable, given that each station has non-zero probability of winning the channel. Suppose at each virtual slot, a station $`i`$ has a probability of $`p_{i}`$ of winning the channel, then on average, it has accumulated $`{{W_{i} \cdot c}/p_{i}} < \infty`$ units of credit. It then spent all of them. In fact, the probability that the credits reach $`G`$ (before it is spent) is approximately $`{({1 - p_{i}})}^{G/{({W_{i} \cdot c})}}`$. Therefore $`{P\hspace{0pt}r\hspace{0pt}{\{{G = \infty}\}}} = 0`$.
But if there is a pre-defined limit of the burst length $`B_{i}`$ for station $`i`$, then instability is possible. In particular, if $`B_{i} \leq {{W_{i} \cdot c}/p_{i}}`$, the credits of station $`i`$ go to infinity. This can be readily proved by law of large numbers. Also, this inequality tells us how to avoid instability: ensuring
<table>
<tbody>
<tr>
<td></td>
<td>Bi>Wi⋅c/pi,∀isubscript𝐵𝑖⋅subscript𝑊𝑖𝑐subscript𝑝𝑖for-all𝑖B_{i}>W_{i}\cdot c/p_{i},\forall i</td>
<td></td>
<td>(1)</td>
</tr>
</tbody>
</table>
For implementation, each station monitors its $`p_{i}`$ (average over a period of time). If the inequality is violated, it computes a proper value of $`c`$ and broadcasts it to the network. Then every station uses the new value of $`c`$. Another simple implementation is to monitor the credits. If the credits of station $`i`$ keep increasing, it knows a smaller $`c`$ is needed to stabilize its credit.
The above analysis can be extended to the case of capture effect or channel errors.
1. 1\.
Capture effect. Capture effect will affect the values of the $`p_{i}`$’s. The weaker stations have smaller $`p_{i}`$’s, which, in turn, may entail adjustment of $`c`$.
2. 2\.
Channel errors. Here, we define $`p_{i}`$ as station $`i`$’s probability of winning the contention AND meet the “good” state of channel $`i`$, in a given virtual slot. Accordingly, channel errors clearly affect the $`p_{i}`$’s. The stations with noisy channels have smaller $`p_{i}`$’s (due to both channel errors and BEB). Also, the time-variation of channel quality imposes another limit on the burst length. Denote the length of the “good” period of channel $`i`$ as $`T_{i}`$, which is a random variable. Then the following inequality is required:
<table>
<tbody>
<tr>
<td></td>
<td>𝐄[min(Bi,Ti)]>Wi⋅c/pi,∀i𝐄delimited-[]𝑚𝑖𝑛subscript𝐵𝑖subscript𝑇𝑖⋅subscript𝑊𝑖𝑐subscript𝑝𝑖for-all𝑖{\bf E}[min(B_{i},T_{i})]>W_{i}\cdot c/p_{i},\forall i</td>
<td></td>
<td>(2)</td>
</tr>
</tbody>
</table>
To analyze the values of $`p_{i}`$’s, one can adapt Bianchi’s fixed point model \[[4](#bib.bib4)\]. In practice, the stations do not need to compute $`p_{i}`$’s. They only need to adjust $`c`$ according to their extra credits.
## V VLS in Multiple Collision domain
If a wireless network has multiple collision domains, a station may not be able to hear all the other stations’ transmissions. Therefore, different stations may have different views of the virtual slots, which makes virtual-slot-based scheduling more difficult. (Two stations have the same view of virtual slots only if they can hear the same set of stations.) In IEEE 802.11, this may cause severe unfairness problem (as will be shown later). So, in this case, we devise a variable-length scheduling algorithm based on the realized throughputs, instead of the number of virtual slots. The basic idea is similar to \[[3](#bib.bib3)\]\[[5](#bib.bib5)\]. (There, the minimum Contention Windows, $`C\hspace{0pt}W_{m\hspace{0pt}i\hspace{0pt}n}`$’s, are dynamically adjusted.)
Say we have a set of $`J`$ stations, with respective weights $`{{W_{j},j} = 1},{2,\ldots,J}`$. The weights can be pre-determined by optimizing some global objective of the network. For example, they can be the solution of the utility optimization problem \[[6](#bib.bib6)\]\[[10](#bib.bib10)\]
<table>
<tbody>
<tr>
<td></td>
<td></td>
<td>max𝑚𝑎𝑥\displaystyle max</td>
<td>∑jUj(Wj)subscript𝑗subscript𝑈𝑗subscript𝑊𝑗\displaystyle\sum_{j}U_{j}(W_{j})</td>
<td></td>
</tr>
<tr>
<td></td>
<td></td>
<td>st𝑠𝑡\displaystyle st</td>
<td>∑k∈𝒞mWk≤1,∀,mm=1,2,…,M\displaystyle\sum_{k\in{\cal C}_{m}}W_{k}\leq 1,\forall{}_{m},m=1,2,\dots,M</td>
<td></td>
</tr>
</tbody>
</table>
where $`U_{j}`$ is the concave “utility function” of station $`j`$, and each set $`\mathcal{C}_{m}`$ is a “clique” (in a clique, only one station can transmit at a time). Note that with the constraints, we have assumed that the contention graph is a “perfect graph” \[[12](#bib.bib12)\] (otherwise, another set of constraints based on “independent set” should be used \[[12](#bib.bib12)\]). We have also omitted some details such as packet collisions. This optimization problem can be solved in a distributed manner, similar to wired network \[[6](#bib.bib6)\]. A “clique” here is analogous to a “link” in wired network, therefore to solve the problem, some communication among stations in the same clique is needed.
In this VLS algorithm, each station $`j`$ monitors the aggregate throughput of its neighbors $`\mathcal{N}_{j}`$ (i.e, $`\sum_{k \in \mathcal{N}_{j}}S_{k}`$), as well as its own throughput $`S_{j}`$. (This can be done in several ways: (a) station $`j`$ can overhear the packets sent by its neighbors, if possible; (b) otherwise, stations can explicitly exchange information about their throughput with their neighbors periodically.) Then, it adjusts the burst length as follows.
<table>
<tbody>
<tr>
<td></td>
<td>bj(t+1)=bj(t)−α(t)bj(t)(Sj∑k∈𝒩jSk−Wj∑k∈𝒩jWk)subscript𝑏𝑗𝑡1subscript𝑏𝑗𝑡𝛼𝑡subscript𝑏𝑗𝑡subscript𝑆𝑗subscript𝑘subscript𝒩𝑗subscript𝑆𝑘subscript𝑊𝑗subscript𝑘subscript𝒩𝑗subscript𝑊𝑘b_{j}(t+1)=b_{j}(t)-\alpha(t)b_{j}(t)(\frac{S_{j}}{\sum_{k\in{\cal N}_{j}}S_{k}}-\frac{W_{j}}{\sum_{k\in{\cal N}_{j}}W_{k}})</td>
<td></td>
</tr>
</tbody>
</table>
where $`b_{j}\hspace{0pt}{(t)}`$ denotes the burst length of station $`j`$ at time $`t`$, and $`\alpha\hspace{0pt}{(t)}`$ is the step size. Clearly, $`b\hspace{0pt}{(t)}`$ converges when the actual ratio of throughputs is equal to the target ratio of weights.
In the following simulation, we compare the throughput allocation with fixed-length scheduling and VLS. In VLS, we use a discrete version of the above algorithm: each node adjusts its transmission length every 4ms, and the throughput $`S_{j}`$ is an average over the last 40ms.
In the network simulated, stations 1, 2, 3 belong to Collision Domain 1, while stations 3, 4, 5 belong to Collision Domain 2, as shown in Fig. [8](#S5.F8 "Figure 8 ‣ V VLS in Multiple Collision domain ‣ Distributed Fair Scheduling Using Variable Transmission Lengths in Carrier-Sensing-based Wireless Networks")(a). Note that station 3 faces contentions from both domains. The destinations of the flows from node 1, 2, 3, 4, 5 are assumed to be node 2, 1, 2, 5, 4, respectively. The data rate is 11Mbps, and the initial transmission length is 1ms for all stations.
Without VLS, station 3’s throughput is very low compared to others (Fig. [8](#S5.F8 "Figure 8 ‣ V VLS in Multiple Collision domain ‣ Distributed Fair Scheduling Using Variable Transmission Lengths in Carrier-Sensing-based Wireless Networks")(b)). The reason is that at most of the time, station 3 senses the medium as busy, due to the transmissions in both collision domains. Therefore it does not have much chance to transmit its packets. Then, we use VLS, and require the weights of all stations to be 1/3. As shown in Fig. [8](#S5.F8 "Figure 8 ‣ V VLS in Multiple Collision domain ‣ Distributed Fair Scheduling Using Variable Transmission Lengths in Carrier-Sensing-based Wireless Networks")(c), after a short period of time in the beginning (for convergence), the target weights are achieved. On average, station 3’ transmission length is 1.54ms, while others’ are about 0.5ms after convergence.
[Refer to caption](/html/0712.2054/assets/x13.png)
\(a\)
A network with 2 collision domains
[Refer to caption](/html/0712.2054/assets/x14.png)
\(b\)
Without VLS
[Refer to caption](/html/0712.2054/assets/x15.png)
\(c\)
VLS, with equal weights for all stations
Figure 8: VLS in Multiple Collision Domains
## VI Conclusions
In this paper, we have proposed a distributed algorithm, VLS, for fair scheduling in 802.11 wireless networks. We have shown that, by varying the transmission lengths of different stations, it is feasible to provide fairness without careful modeling of MAC-layer contention, physical channel variations, and the effect of multiple collision domains. Many existing fairness problems in 802.11 networks are therefore overcome, including short-term unfairness, unfairness introduced by physical channel’s asymmetry, such as capture effect and channel errors. It makes the throughput of each station more smooth, which is appealing to real-time service such as video and voice. It can easily provide weighted fairness for different services. For networks with multiple collision domains, VLS gives a way to avoid the starvation of those stations that are in several collision domains.
Since VLS can avoid excessively long transmissions by tuning the parameter $`c`$, we should be able to achieve certain objectives on packet delay. However, this problem has not been studied thoroughly in this paper, and is a subject for further study.
Since VLS can be used in 802.11e networks, the advantages of both protocols can be achieved: VLS can provide weighted fairness (a prioritization) in terms of throughput, while 802.11e can provide prioritization in terms of delay, by using different $`C\hspace{0pt}W_{m\hspace{0pt}i\hspace{0pt}n}`$’s and IFS’s (Inter-Frame Spaces). Therefore, throughputs and delays can be controlled separately by two protocols, instead of being coupled in a complicated way as in 802.11e \[[13](#bib.bib13)\]. More analytical/experimental study of this issue is interesting for future research.
## References
- \[1\] Information technology- telecommunications and information exchange between systems- local and metropolitan area networks- specific requirements- part 11: Wireless lan medium access control (mac) and physical layer (phy) specifications.
ANSI/IEEE Std 802.11, 1999 Edition (R2003), pages i–513, 2003.
- \[2\] Unapproved draft amendment standard for information technology– telecommunications and information exchange between systems–lan/man specific requirements– part 11 wireless medium access control (mac) and physical layer (phy) specifications: Medium access control (mac) quality of service (qos) enhancements (replaced by approved draft 802.11e/d13.0).
IEEE Std P802.11e/D13.0, 2005.
- \[3\] B. Bensaou, Wang Yu, and Ko Chi Chung. Fair medium access in 802.11 based wireless ad-hoc networks.
In Mobile and Ad Hoc Networking and Computing, 2000. MobiHOC. 2000 First Annual Workshop on, pages 99–106, 2000.
- \[4\] G. Bianchi. Performance analysis of the ieee 802.11 distributed coordination function.
Selected Areas in Communications, IEEE Journal on, 18(3):535–547, 2000.
0733-8716.
- \[5\] Rajarshi Gupta and Jean Walrand.
Achieving Fairness in a Distributed Ad-Hoc MAC.
”Advances in Control, Communication Networks, and Transportation Systems,” E.H. Abed (Ed.), Systems and Control: Foundations and Applications Series, Springer-Birkhauser, Boston. July 2005.
- \[6\] F. Kelly. Charging and rate control for elastic traffic.
European Transactions on Telecommunications, 8:33–37, January 1997.
- \[7\] A. Kochut, A. Vasan, A. U. Shankar, and A. Agrawala. Sniffing out the correct physical layer capture model in 802.11b.
In Network Protocols, 2004. ICNP 2004. Proceedings of the 12th IEEE International Conference on, pages 252–261, 2004.
- \[8\] C. T. Lau and C. Leung. Capture models for mobile packet radio networks.
Communications, IEEE Transactions on, 40(5):917–925, 1992.
0090-6778.
- \[9\] Jiang Li Bin and Liew Soung Chang. Proportional fairness in wireless lans and ad hoc networks.
In Wireless Communications and Networking Conference, 2005 IEEE, volume 3, pages 1551–1556 Vol. 3, 2005.
- \[10\] X. Lin, N.B. Shroff, and R. Srikant. A tutorial on cross-layer optimization in wireless networks.
Selected Areas in Communications, IEEE Journal on, 24(8):1452–1463, 2006.
- \[11\] Liew S.C Qixiang Pang, Leung V.C.M. Improvement of WLAN contention resolution by loss differentiation.
IEEE Transactions on Wireless Communications, 5:3605 – 3615, December 2006.
- \[12\] John Musacchio Rajarshi Gupta and Jean Walrand. Sufficient rate constraints for QoS flows in ad-hoc networks.
AD HOC NETWORKS, 5(4):429–443, May 2007.
- \[13\] J. W. Robinson and T. S. Randhawa. Saturation throughput analysis of ieee 802.11e enhanced distributed coordination function.
Selected Areas in Communications, IEEE Journal on, 22(5):917–928, 2004.
0733-8716.
- \[14\] Choi Sunghyun, J. del Prado, N. Sai Shankar, and S. Mangold. ”IEEE 802.11 e contention-based channel access (EDCF) performance evaluation”.
In Communications, 2003. ICC ’03. IEEE International Conference on, volume 2, pages 1151–1156 vol.2, 2003.
- \[15\] N. Vaidya, A. Dugar, S. Gupta, and P. Bahl. Distributed fair scheduling in a wireless LAN.
Mobile Computing, IEEE Transactions on, 4(6):616–629, 2005. |
0712.2113 | cs.CR | cyber | # On the deployment of Mobile Trusted Modules
Andreas U. Schmidt, Nicolai Kuntze, Michael Kasper
Fraunhofer Institute for Secure Information Technology SIT
Rheinstraße 75, 64295 Darmstadt, Germany
Email: {andreas.schmidt,nicolai.kuntze,michael.kasper}@sit.fraunhofer.de
###### Abstract
In its recently published TCG Mobile Reference Architecture, the TCG Mobile Phone Work Group specifies a new concept to enable trust into future mobile devices. For this purpose, the TCG devises a trusted mobile platform as a set of trusted engines on behalf of different stakeholders supported by a physical trust-anchor. In this paper, we present our perception on this emerging specification. We propose an approach for the practical design and implementation of this concept and how to deploy it to a trustworthy operating platform. In particular we propose a method for the take-ownership of a device by the user and the migration (i.e., portability) of user credentials between devices.
## I Introduction
As a first deliverable within their scope and work programme, the Mobile Phone Work Group of the Trusted Computing Group (TCG MPWG) has published a specification \[[1](#bib.bib1)\], which offers new potentials for implementing trust in mobile computing platforms by introducing a new, hardware-based trust anchor for mobile phones and devices. This trust anchor, called a Mobile Trusted Module (MTM), has properties and features comparable to a Trusted Platform Module (TPM, see \[[2](#bib.bib2), [3](#bib.bib3)\]). Concurrently the MPWG issued a much more universal security architecture for mobile phones and devices on a higher abstraction level. The pertinent specification is called TCG Mobile Reference Architecture (RA) \[[4](#bib.bib4)\] and abstracts a trusted mobile platform as a set of tamper resistant trusted engines operating on behalf of different stakeholders. This architecture offers a high degree on flexibility and modularity in design and implementation of the trusted components to all participants in hard- and software development.
An important aspect of the TCG Mobile Reference Architecture is the potential to virtualise significant parts of a trusted mobile platform as trusted software applications and services. The trusted execution chain for this rests on the MTM. The implementation of this chip depends on the security requirements of its specific use-case. For high levels of protection and isolation, an MTM could be implemented as a slightly modified Trusted Platform Module (TPM). This enables cost-effective implementation of new security-critical applications and various innovative business models, in both the mobile and generic computing domain \[[5](#bib.bib5), [6](#bib.bib6), [7](#bib.bib7)\].
The present paper discusses the main structural features of the RA, highlighting the capabilities of the MTM as the main functional building block. After this technology review, we propose two basic methods for usage of the RA, namely the set-up of a trusted subsystem on a device by a remote owner, and its migration from one device to another.
This paper is organised as follows. In Section [II](#S2 "II TCG MPWG Reference Architecture ‣ On the deployment of Mobile Trusted Modules"), we explore the significant parts of the MPWG Reference Architecture. It is divided into four parts. Subsection [](#S2.SS1 "II-A Architectural Overview ‣ II TCG MPWG Reference Architecture ‣ On the deployment of Mobile Trusted Modules")
II-A
gives an overview of the security architecture, and subsection [](#S2.SS2 "II-B Mobile Trusted Module ‣ II TCG MPWG Reference Architecture ‣ On the deployment of Mobile Trusted Modules")
II-B
details the concepts of the proposed architectural approach for an MTM and the requirements to virtualise its functionality, whereby a high security and isolation level is maintained. Furthermore, we propose a model for remote stakeholder take-ownership in [](#S2.SS3 "II-C Setup and Take-Ownership of a Trusted Subsystem ‣ II TCG MPWG Reference Architecture ‣ On the deployment of Mobile Trusted Modules")
II-C
and migration of trusted subsystems in [](#S2.SS4 "II-D Migration of a Trusted-Subsystem ‣ II TCG MPWG Reference Architecture ‣ On the deployment of Mobile Trusted Modules")
II-D
. In Section [III](#S3 "III Design of Mobile Trusted Modules on Trustworthy Operating Platforms ‣ On the deployment of Mobile Trusted Modules"), we show how such an architecture can be implemented on trustworthy operating platforms.
## II TCG MPWG Reference Architecture
The TCG MPWG has developed an architecture on a high level of abstraction for a trusted mobile platform, which offers numerous variations for design and implementation. In this section, we reflect essential parts of this architecture and an overview of significant platform components in terms of our objective.
###
II-A
Architectural Overview
A trusted mobile platform is characterised as a set of multiple tamper-resistant engines, each acting on behalf of a different stakeholder. Broadly, such an platform has several major components: trusted engines $`\mathcal{T}\hspace{0pt}\mathcal{E}`$, trusted services $`\mathcal{T}\hspace{0pt}\mathcal{S}`$ customised by trusted resources $`\mathcal{T}\hspace{0pt}\mathcal{R}`$. A general trusted mobile platform is illustrated in Figure [1](#S2.F1 "Figure 1 ‣ II-A Architectural Overview ‣ II TCG MPWG Reference Architecture ‣ On the deployment of Mobile Trusted Modules").
[Refer to caption](/html/0712.2113/assets/x1.png)
Figure 1: Trusted Mobile Platform Architecture
We define a trusted subsystem $`\mathcal{T}\hspace{0pt}\mathcal{S}\hspace{0pt}\mathcal{S}_{\sigma}`$ as a logical unit of a trusted engine together with its interrelated hardware compartment. A $`\mathcal{T}\hspace{0pt}\mathcal{S}\hspace{0pt}\mathcal{S}`$ of a stakeholder $`\sigma`$ can formally described by a tuple
<table>
<tbody>
<tr>
<td></td>
<td>𝒯𝒮𝒮σ𝒯𝒮subscript𝒮𝜎\displaystyle\mathcal{TSS_{\sigma}}</td>
<td>=\displaystyle=</td>
<td>{𝒯ℰσ,𝒯𝒮σ,𝒯𝒮ϵ,𝒯ℛσ,𝒮𝒫σ,𝒮𝒞σ}𝒯subscriptℰ𝜎𝒯subscript𝒮𝜎𝒯subscript𝒮italic-ϵ𝒯subscriptℛ𝜎𝒮subscript𝒫𝜎𝒮subscript𝒞𝜎\displaystyle\left\{\mathcal{TE_{\sigma}},\mathcal{TS_{\sigma}},\mathcal{TS_{\epsilon}},\mathcal{TR_{\sigma}},\mathcal{SP_{\sigma}},\mathcal{SC_{\sigma}}\right\}</td>
<td></td>
</tr>
</tbody>
</table>
In each trusted subsystem $`\mathcal{T}\hspace{0pt}\mathcal{S}\hspace{0pt}\mathcal{S}`$ either a remote or local entity acts as a stakeholder, who is able to configure its own subsystem and define his security policy $`\mathcal{S}\hspace{0pt}\mathcal{P}_{\sigma}`$ and system configuration $`\mathcal{S}\hspace{0pt}\mathcal{C}_{\sigma}`$ within an isolated and protected environment. The MPWG Reference Architecture specifies the following principal entities: the local stakeholders Device Owner $`\mathcal{D}\hspace{0pt}\mathcal{O}`$ and User $`\mathcal{U}`$; and the remote stakeholders Device Manufacturer $`\mathcal{D}\hspace{0pt}\mathcal{M}`$, and more general Remote Owners $`\mathcal{R}\hspace{0pt}\mathcal{O}`$ (e.g. a communication carrier, or service provider). The functionality of a $`\mathcal{T}\hspace{0pt}\mathcal{S}\hspace{0pt}\mathcal{S}`$ is either based on dedicated resources of an embedded engine $`\mathcal{T}\hspace{0pt}\mathcal{E}_{\sigma}`$, or may be provided by trusted services $`\mathcal{T}\hspace{0pt}\mathcal{S}_{\epsilon}`$ of external engines.
Each subsystem is able to enforce its security policy $`\mathcal{S}\hspace{0pt}\mathcal{P}_{\sigma}`$ and subsystem configuration $`\mathcal{S}\hspace{0pt}\mathcal{C}_{\sigma}`$. As a consequence, the functionality of a trusted subsystem $`\mathcal{T}\hspace{0pt}\mathcal{S}\hspace{0pt}\mathcal{S}_{\sigma}`$ is constrained by the available resources $`\mathcal{T}\hspace{0pt}\mathcal{R}_{\sigma}`$ with their derived trusted services $`\mathcal{T}\hspace{0pt}\mathcal{S}_{\sigma}`$, by the offered functionality of external trusted services $`\mathcal{T}\hspace{0pt}\mathcal{S}_{\epsilon}`$, by the security policy $`\mathcal{S}\hspace{0pt}\mathcal{P}_{\sigma}`$, and finally the system configuration $`\mathcal{S}\hspace{0pt}\mathcal{C}_{\sigma}`$ of an engine’s stakeholder.
All internal functions executed inside $`\mathcal{T}\hspace{0pt}\mathcal{S}\hspace{0pt}\mathcal{S}_{\sigma}`$ are isolated from other subsystems by the underlying security layer and is only accessible if a proper service interface is defined and exported. A $`\mathcal{T}\hspace{0pt}\mathcal{S}\hspace{0pt}\mathcal{S}_{\sigma}`$ relies on the reputation of the stakeholder $`\sigma`$ as basis for that trust. Therefore, each stakeholder issues a security policy $`\mathcal{S}\hspace{0pt}\mathcal{P}_{\sigma}`$ and a set of credentials belonging to embedded trusted components of its subsystem $`\mathcal{T}\hspace{0pt}\mathcal{S}\hspace{0pt}\mathcal{S}_{\sigma}`$. This policy contains reference measurements (RIM), quality assertions and security-critical requirements.
####
II-A1
Trusted Engines
The most important concept within the MPWG Reference Architecture is that of trusted engines.
[Refer to caption](/html/0712.2113/assets/x2.png)
Figure 2: Generic Trusted Engine
The purpose of a trusted engine is to provide confidence in all its embedded services, which are internally or externally provided by the engine. It is a protected entity on behalf of a specific stakeholder that has special abilities to manipulate and store data, and to provide evidence of its trustworthiness and the current state of the engine. Figure [2](#S2.F2 "Figure 2 ‣ II-A1 Trusted Engines ‣ II-A Architectural Overview ‣ II TCG MPWG Reference Architecture ‣ On the deployment of Mobile Trusted Modules") shows a generic trusted engine. In general, each engine has at least following abilities:
- •
implement arbitrary software functionalities as trusted and/or normal services,
- •
provide the evidence for its trustworthiness,
- •
report the evidence of its current state,
- •
obtain and use Endorsement Keys (EK) and/or Attestation Identity Keys (AIK),
- •
access a set of trusted resources, and
- •
import and/or export services, shielded capabilities and protected functionality.
In order to establish a definite categorisation, the MPWG differentiates engines according to their functional dispensability. Therefore, an engine is either dedicated to a mandatory (of $`\mathcal{D}\hspace{0pt}\mathcal{O}`$ or $`\mathcal{D}\hspace{0pt}\mathcal{M}`$) or a discretionary domain (of $`\mathcal{D}\hspace{0pt}\mathcal{O}`$). Engines inside a mandatory domain are permanently located on a trusted platform and hold security-critical and essential functionality. All essential services of a trusted mobile platform should be located inside the mandatory domain, which does *not permit a local stakeholder to remove a remote owner from the engine*. Mandatory engines have access to a Mobile Remote owner Trusted Module (MRTM) to guarantee that a valid and trustworthy engine state is always present.
Non-essential engines and services are replaceable by the device owner $`\mathcal{D}\hspace{0pt}\mathcal{O}`$ and should be located inside the discretionary domain. Engines inside the discretionary domain are controlled by the device owner $`\mathcal{D}\hspace{0pt}\mathcal{O}`$. Discretionary engines are required to be supported by a Mobile Local owner Trusted Module (MLTM).
####
II-A2
Trusted Resources
As illustrated in Figure [2](#S2.F2 "Figure 2 ‣ II-A1 Trusted Engines ‣ II-A Architectural Overview ‣ II TCG MPWG Reference Architecture ‣ On the deployment of Mobile Trusted Modules"), an internal trusted service has access to several trusted resources. The TCG calls these resources Root-Of-Trusts (RoT) representing the trusted components acting on base of the trusted execution chain and providing functionality for measurement, storing, reporting, verification and enforcement that affect the trustworthiness of the platform. The following RoTs are defined for the mobile domain:
- •
Root of Trust for Storage (RTS),
- •
Root of Trust for Reporting (RTR),
- •
Root of Trust for Measurement (RTM),
- •
Root of Trust for Verification (RTV), and
- •
Root of Trust for Enforcement (RTE)
Each RoT vouches its trustworthiness either directly by supplied secrets (EK, AIK) and associated credentials, which are only accessible by authenticated subjects of the stakeholder, or indirectly by measurements of other trusted resources. These resources are only mutable by authorised entities of a stakeholder.
In this paper, we group several logically self-contained RoTs to simplify the presentation of interfaces and the communication layer. In a typical arrangement, the RTS and RTR represent one unit, while the RTM and RTV build another unit within an $`\mathcal{T}\hspace{0pt}\mathcal{S}\hspace{0pt}\mathcal{S}_{\sigma}`$. However, note that the RTV and the RTM depend on protected storage mechanisms, which are provided by the RTS. Thus, it is also plausible to implement all RoTs together as a common unit within an engine.
RTS/RTR are the trusted resources that are responsible for secure storage and reliable reporting of information about the state of trusted mobile platform. An RTS provide PCRs and protected storage for an engine and stores the measurements made by the RTM, cryptographic keys, and security sensitive data. An RTR signs the measurements with cryptographic signature keys of $`T\hspace{0pt}S\hspace{0pt}S_{\sigma}`$.
RTM/RTV In general, an RTM is a reliable instance to measure software components and provide evidence of the current state of a trusted engine and its embedded services. In the mobile domain, to avoid communication costs, this functionality is extended by a local verifier, which checks the measurements against a given Reference Integrity Metrics (RIM). This process can be done instantly as the measurements are performed employing a combination of RTM and RTV. Figure [3](#S2.F3 "Figure 3 ‣ II-A2 Trusted Resources ‣ II-A Architectural Overview ‣ II TCG MPWG Reference Architecture ‣ On the deployment of Mobile Trusted Modules") depicts such a Measure$`\rightarrow`$Verify$`\rightarrow`$Extend process.
[Refer to caption](/html/0712.2113/assets/x3.png)
Figure 3: Measurement and Verification Process
An RTE is required if an engine uses allocated resources and services. In this case, such RoT acts as a trusted boot loader and ensures the availability of all allocated trusted resources and services within that trusted subsystem.
####
II-A3
Services of a Trusted Engine
A trusted engine integrates all functionality by customising available platform resources as software services. Such a service offers computation, storage, or communication channels to other internal or external services and applications based on dedicated or allocated resources. The MPWG categorises them into: trusted, normal, and measured services.
A trusted service customises trusted resources. Thus, a trusted service is implicitly supplied with an $`E\hspace{0pt}K`$ or $`A\hspace{0pt}I\hspace{0pt}K`$ in order to attest its trustworthiness. Trusted services are intended to provide reliable measurements of their current state and to provide evidence of the state of other normal services or resources.
Normal services are customising normal resources and implement functionality, but are not able to provide evidence of their trustworthiness by own capabilities. However, normal services can access internal trusted services to use their provided functionality. Therefore, an internal normal services is able to vouch its trustworthiness by associated integrity metrics that have been measured by a trusted service.
###
II-B
Mobile Trusted Module
The generic term Mobile Trusted Module (MTM) refers to a dedicated hardware-based trust-anchor. It is typically composed of an RTS and RTR and has characteristics comparable to a TPM. According to their design objective the MPWG distinguishes between MRTM and MLTM. Both must support a subset of TPM commands as specified in \[[4](#bib.bib4)\]. Additionally, an MRTM has to support a set of commands to enable local verification and specific mobile device functionality.
The TCG MPWG Reference Architecture does not exclude to utilise a TPM v1.2 (or even a TPM v1.1) as an MTM, if an appropriate interface consisting of a set of commands conforming to the MPWG specification and associated data structures are provided. Although it is possible to implement this architecture upon a standard TPM, we here focus on a more sophisticated solution based on a Trustworthy Computing Platform such as EMSCB/Turaya \[[8](#bib.bib8)\]. In this context, we expect three different solutions for isolation, key management and protection of $`\mathcal{T}\hspace{0pt}\mathcal{S}\hspace{0pt}\mathcal{S}_{\sigma}`$ .
A Standard TPM-based Model uses a non-modified standard TPM to build the trusted computing base. The secret keys are stored into a single key-hierarchy on behalf of $`\mathcal{D}\hspace{0pt}\mathcal{O}`$ as specified in \[[1](#bib.bib1)\]. In this case, an adversary or malicious local owner may be able to access the secret keys of a remote stakeholder and take control of a remote owner compartment. A $`\mathcal{D}\hspace{0pt}\mathcal{O}`$ can also disable the whole MTM or corrupt mandatory engines of remote stakeholders.
A Software-based MTM-Emulation Model uses a software-based allocated $`M\hspace{0pt}T\hspace{0pt}M`$-emulation with an isolated key-hierarchy. All sensitive and security-critical, such as $`E\hspace{0pt}K`$ or $`S\hspace{0pt}R\hspace{0pt}K`$, are only protected by software mechanisms outside of the tamper-resistant environment of a dedicated MTM \[[9](#bib.bib9), [10](#bib.bib10)\].
Generic MTM-based Model supporting multiple stakeholders and virtual MTMs. In order to circumvent resulting drawbacks and mitigate attacks, we favour a solution with a higher level of security. For this reason, we adopt the proposed secure co-processor variant of \[[9](#bib.bib9)\] and describe a generic MTM with support for multiple stakeholder environments.
[Refer to caption](/html/0712.2113/assets/x4.png)
Figure 4: MTM Architecture supporting multiple Stakeholders
In a cost-efficient scenario, the trusted mobile platform is implementable based on a single generic MTM and several virtualised MTMs for each trusted engine. Hence, at least one dedicated MTM has to be available and additionally a unique vMTM has to be instantiated in each trusted subsystem $`\mathcal{T}\hspace{0pt}\mathcal{S}_{\sigma}`$. In such case, a physically bounded MTM in the platform acts as a master trust anchor and offers MRTM and MLTM functionality with respect to its specific use case.
A Trusted Software Layer offers a vMTM Proxy Service to all embedded trusted engines $`\mathcal{T}\hspace{0pt}\mathcal{E}_{\sigma}`$. The main task of this service is to route MTM commands from a $`\mathcal{T}\hspace{0pt}\mathcal{E}_{\sigma}`$ to its dedicated instance $`v\hspace{0pt}M\hspace{0pt}T\hspace{0pt}M_{\sigma}`$. The advantage is that all security-critical MTM commands are tunnelled to $`v\hspace{0pt}M\hspace{0pt}T\hspace{0pt}M_{\sigma}`$ and are executed within the protected environment of the dedicated MTM.
Figure [4](#S2.F4 "Figure 4 ‣ II-B Mobile Trusted Module ‣ II TCG MPWG Reference Architecture ‣ On the deployment of Mobile Trusted Modules") illustrates the architecture of a generic MTM with isolated vMTM compartments. This architecture requires a slightly modified TPM. Mainly, we add a trusted component, the vMTM Instance Manager, which is responsible to separate vMTM instances from each other. This includes administration, isolated execution, memory management and access control for each stakeholder compartment. Thus, a vMTM instance is able to hold an autonomous and hardware-protected key hierarchy to store its secrets and protect the execution of security-critical data (e.g. signature and encryption algorithms).
###
II-C
Setup and Take-Ownership of a Trusted Subsystem
The take-ownership operation establishes the trust relationship between a stakeholder and trusted mobile platform. Currently, the MPWG Reference Architecture does not define how a remote owner is to perform this initial setup and take-ownership of its $`\mathcal{T}\hspace{0pt}\mathcal{S}\hspace{0pt}\mathcal{S}_{\sigma}`$. Hence we propose a method in this section. The main idea behind our procedure is to install and instantiate a ’blank’ trusted subsystem $`\mathcal{T}\hspace{0pt}\mathcal{S}\hspace{0pt}\mathcal{S}_{\sigma}^{\ast}`$ containing a pristine engine $`\mathcal{T}\hspace{0pt}\mathcal{E}_{\mathcal{R}\hspace{0pt}\mathcal{O}}^{\ast}`$ with a set of generic trusted services $`\mathcal{T}\hspace{0pt}\mathcal{S}_{\sigma}^{\ast}`$. This subsystem is then certified by a remote owner, if the platform is able to provide evidence of its pristine configuration and policy conformance with respect to $`\mathcal{R}\hspace{0pt}\mathcal{O}`$. Figure [5](#S2.F5 "Figure 5 ‣ II-C Setup and Take-Ownership of a Trusted Subsystem ‣ II TCG MPWG Reference Architecture ‣ On the deployment of Mobile Trusted Modules") illustrates this process, which we now descirbe.
[Refer to caption](/html/0712.2113/assets/x5.png)
Figure 5: Remote Stakeholder Take-Ownership Protocol
Platform and Protocol Precondition. In a preliminary stage, the trusted mobile platform has carried out the boot process and has loaded the trusted computing base and the engine $`\mathcal{T}\hspace{0pt}\mathcal{E}_{\mathcal{D}\hspace{0pt}\mathcal{M}}`$ with its trusted services. The trusted platform has checked that the installed hardware and running software are in a trustworthy state and configuration. It is able to report and attest this state, if challenged by an authorised entity.
Remote Stakeholder Take-Ownership Protocol. In the first phase, the trusted engine $`\mathcal{T}\hspace{0pt}\mathcal{E}_{\mathcal{D}\hspace{0pt}\mathcal{M}}`$ carries out a take-ownership preparation for the remote stakeholder. A ’blank’ engine $`\mathcal{T}\hspace{0pt}\mathcal{E}_{\mathcal{R}\hspace{0pt}\mathcal{O}}^{\ast}`$ is installed and booted by the $`R\hspace{0pt}T\hspace{0pt}E_{\mathcal{D}\hspace{0pt}\mathcal{M}}`$, and a clean $`v\hspace{0pt}M\hspace{0pt}T\hspace{0pt}M_{\mathcal{R}\hspace{0pt}\mathcal{O}}`$ instance is activated inside the dedicated $`M\hspace{0pt}T\hspace{0pt}M`$. An initial setup prepares the pristine engine $`\mathcal{T}\hspace{0pt}\mathcal{E}_{\mathcal{R}\hspace{0pt}\mathcal{O}}^{\ast}`$. A endorsement key-pair $`E\hspace{0pt}K_{\mathcal{R}\hspace{0pt}\mathcal{O}}^{\ast}`$ is generated within $`v\hspace{0pt}M\hspace{0pt}T\hspace{0pt}M_{\mathcal{R}\hspace{0pt}\mathcal{O}}`$ with a corresponding certificate $`C\hspace{0pt}e\hspace{0pt}r\hspace{0pt}t_{\mathcal{T}\hspace{0pt}\mathcal{S}\hspace{0pt}\mathcal{S}_{\mathcal{R}\hspace{0pt}\mathcal{O}}}`$
<sup>1</sup>
<sup>1</sup>1Typically, the key generation needs a so-called Owner Authentication. Because this is problematic in a remote-owner scenario, authentication of command execution may be enforced by challenge-response mechanisms between $`\mathcal{R}\hspace{0pt}\mathcal{O}`$ and $`\mathcal{T}\hspace{0pt}\mathcal{S}\hspace{0pt}\mathcal{S}_{\mathcal{R}\hspace{0pt}\mathcal{O}}`$.
.
Next, $`\mathcal{T}\hspace{0pt}\mathcal{E}_{\mathcal{R}\hspace{0pt}\mathcal{O}}^{\ast}`$ performs an attestation of its current state. The attestation can be done by the local verifier $`R\hspace{0pt}T\hspace{0pt}V_{\mathcal{D}\hspace{0pt}\mathcal{M}}`$ inside the $`\mathcal{T}\hspace{0pt}\mathcal{S}\hspace{0pt}\mathcal{S}_{D\hspace{0pt}M}`$ using $`R\hspace{0pt}I\hspace{0pt}M`$ certificates of the remote stakeholder $`\mathcal{R}\hspace{0pt}\mathcal{O}`$. If no suitable $`R\hspace{0pt}I\hspace{0pt}M`$ and corresponding $`R\hspace{0pt}I\hspace{0pt}M`$-certificate are available for an pristine engine, alternatively a remote attestation with an associated Privacy CA is also possible.
$`\mathcal{T}\hspace{0pt}\mathcal{E}_{\mathcal{R}\hspace{0pt}\mathcal{O}}^{\ast}`$ creates a symmetric key $`K_{{\mathcal{R}\hspace{0pt}\mathcal{O}},{t\hspace{0pt}e\hspace{0pt}m\hspace{0pt}p}}`$ and encrypts the public portion of the endorsement key $`E\hspace{0pt}K_{\mathcal{R}\hspace{0pt}\mathcal{O}}^{\ast}`$, the corresponding certificate $`C\hspace{0pt}e\hspace{0pt}r\hspace{0pt}t_{\mathcal{T}\hspace{0pt}\mathcal{S}\hspace{0pt}\mathcal{S}_{\mathcal{R}\hspace{0pt}\mathcal{O}}}`$, attestation and purpose information. Next, $`\mathcal{T}\hspace{0pt}\mathcal{E}_{\mathcal{R}\hspace{0pt}\mathcal{O}}^{\ast}`$ encrypts $`K_{{\mathcal{R}\hspace{0pt}\mathcal{O}},{t\hspace{0pt}e\hspace{0pt}m\hspace{0pt}p}}`$ with a public key $`K_{{\mathcal{R}\hspace{0pt}\mathcal{O}},{P\hspace{0pt}K}}`$ and sends both messages to the remote owner. After reception by the remote stakeholder, the messages are decrypted using the private portion of key $`K_{{\mathcal{R}\hspace{0pt}\mathcal{O}},{P\hspace{0pt}K}}`$. We assume that this key is either available through a protected communication channel or pre-installed by the device manufacturer.
In a next step, $`\mathcal{R}\hspace{0pt}\mathcal{O}`$ verifies the attestation data and checks the intended purpose of $`\mathcal{T}\hspace{0pt}\mathcal{E}_{\mathcal{R}\hspace{0pt}\mathcal{O}}^{\ast}`$. If the engine and device attestation data is valid and the intended purpose is acceptable, the $`\mathcal{R}\hspace{0pt}\mathcal{O}`$ generates an individual security policy $`\mathcal{S}\hspace{0pt}\mathcal{P}_{R\hspace{0pt}O}`$. The $`\mathcal{R}\hspace{0pt}\mathcal{O}`$ signs the $`C\hspace{0pt}e\hspace{0pt}r\hspace{0pt}t_{\mathcal{T}\hspace{0pt}\mathcal{S}\hspace{0pt}\mathcal{S}_{\mathcal{R}\hspace{0pt}\mathcal{O}}}`$ and creates $`R\hspace{0pt}I\hspace{0pt}M`$ certificates for local verification of a ’complete’ $`\mathcal{T}\hspace{0pt}\mathcal{S}\hspace{0pt}\mathcal{S}_{\sigma}`$. Furthermore, $`\mathcal{R}\hspace{0pt}\mathcal{O}`$ creates a Setup Configuration $`S\hspace{0pt}C_{\mathcal{T}\hspace{0pt}\mathcal{S}\hspace{0pt}\mathcal{S}_{\mathcal{R}\hspace{0pt}\mathcal{O}}}`$, which enforces the engine to individualise its services and complete its configuration with respect to the intended purpose and given security policy. In this step, $`\mathcal{R}\hspace{0pt}\mathcal{O}`$ encrypts the messages with the public portion of the $`K_{{\mathcal{R}\hspace{0pt}\mathcal{O}},{E\hspace{0pt}K}}`$ and transfers this package to the engine $`\mathcal{T}\hspace{0pt}\mathcal{E}_{R\hspace{0pt}O}`$.
Finally, the trusted engine $`\mathcal{T}\hspace{0pt}\mathcal{E}_{\mathcal{R}\hspace{0pt}\mathcal{O}}^{\ast}`$ decrypts the received package and installs it inside the $`T\hspace{0pt}S\hspace{0pt}S_{\mathcal{R}\hspace{0pt}\mathcal{O}}`$ and thus completes its instantiation.
###
II-D
Migration of a Trusted-Subsystem
If a stakeholder wants to move a source $`\mathcal{T}\hspace{0pt}\mathcal{S}\hspace{0pt}\mathcal{S}_{\sigma,S}`$ to another MTM-enabled platform, for instance to port user credentials from device to device, all security-critical information including the Storage Root Key (SRK) has to be migrated to the target $`\mathcal{T}\hspace{0pt}\mathcal{S}\hspace{0pt}\mathcal{S}_{\sigma,D}`$. In our scenario, we assume the same remote owner (e.g. mobile network operator) on both subsystems $`\mathcal{T}\hspace{0pt}\mathcal{S}\hspace{0pt}\mathcal{S}_{{R\hspace{0pt}O},S}`$ and $`\mathcal{T}\hspace{0pt}\mathcal{S}\hspace{0pt}\mathcal{S}_{{R\hspace{0pt}O},D}`$.
[Refer to caption](/html/0712.2113/assets/x6.png)
Figure 6: Trusted Subsystem Migration Protocol
To be able to securely migrate the SRK, we suggest a modification of the current MPWG specification to allow inter-stakeholder-migration of a complete isolated key hierarchy. Thus, an isolated key hierarchy is (1) migratable between environments of identical stakeholders, (2) if and only if an entitling security policy on both platforms exists. The advantage of migration between identical stakeholder subsystems is that the migration process doesn’t require a trusted third party. We only involve the owner in combination with local verification mechanisms of the $`\mathcal{T}\hspace{0pt}\mathcal{S}\hspace{0pt}\mathcal{S}_{R\hspace{0pt}O}`$ to migrate the trusted subsystem (including the SRK) to another platform. This enables for instance direct, device-to-device porting of credentials, e.g. using short-range communication. We here propose a complete, multilateral and secure migration protocol, which is illustrated in Figure [6](#S2.F6 "Figure 6 ‣ II-D Migration of a Trusted-Subsystem ‣ II TCG MPWG Reference Architecture ‣ On the deployment of Mobile Trusted Modules").
Platform and Protocol Precondition. Similar to section [](#S2.SS3 "II-C Setup and Take-Ownership of a Trusted Subsystem ‣ II TCG MPWG Reference Architecture ‣ On the deployment of Mobile Trusted Modules")
II-C
, the trusted mobile platform has carried out the same initial steps as mentioned above. Furthermore, the remote owner has performed an remote take-ownership procedure as described in [](#S2.SS3 "II-C Setup and Take-Ownership of a Trusted Subsystem ‣ II TCG MPWG Reference Architecture ‣ On the deployment of Mobile Trusted Modules")
II-C
.
Trusted Subsystem Migration Protocol. At the beginning of the migration protocol, the device owner $`\mathcal{D}\hspace{0pt}\mathcal{O}_{S}`$ of the source platform $`\mathcal{T}\hspace{0pt}\mathcal{P}_{S}`$ initialises the migration procedure and requests an appropriate migration service of $`\mathcal{T}\hspace{0pt}\mathcal{S}\hspace{0pt}\mathcal{S}_{{R\hspace{0pt}O},S}`$. Next, the trusted platform $`\mathcal{T}\hspace{0pt}\mathcal{P}_{S}`$ is instructed by $`\mathcal{T}\hspace{0pt}\mathcal{S}\hspace{0pt}\mathcal{S}_{R\hspace{0pt}O}`$ to establish a secure channel to the target platform $`\mathcal{T}\hspace{0pt}\mathcal{P}_{D}`$. After the connection is available, $`\mathcal{T}\hspace{0pt}\mathcal{S}\hspace{0pt}\mathcal{S}_{{R\hspace{0pt}O},S}`$ activates the corresponding migration service of $`\mathcal{T}\hspace{0pt}\mathcal{S}\hspace{0pt}\mathcal{S}_{{R\hspace{0pt}O},D}`$ to perform the import procedure. Thereupon, the target subsystem $`\mathcal{T}\hspace{0pt}\mathcal{S}\hspace{0pt}\mathcal{S}_{\sigma,D}`$ performs a local verification of $`\mathcal{T}\hspace{0pt}\mathcal{S}\hspace{0pt}\mathcal{S}_{{R\hspace{0pt}O},S}`$. If revoked, it replies with an error message and halts the protocol. Otherwise $`\mathcal{T}\hspace{0pt}\mathcal{E}_{{R\hspace{0pt}O},D}`$ requests an confirmation from the local owner $`\mathcal{D}\hspace{0pt}\mathcal{O}_{D}`$.
Next, the target subsystem $`\mathcal{T}\hspace{0pt}\mathcal{S}\hspace{0pt}\mathcal{S}_{{R\hspace{0pt}O},D}`$ generates a nonce $`N_{{R\hspace{0pt}O},D}`$. In order to provide evidence of its trustworthiness, $`\mathcal{T}\hspace{0pt}\mathcal{S}\hspace{0pt}\mathcal{S}_{{R\hspace{0pt}O},D}`$ sends all necessary information to the source subsystem $`\mathcal{T}\hspace{0pt}\mathcal{S}\hspace{0pt}\mathcal{S}_{{R\hspace{0pt}O},S}`$. This includes the current state $`S_{{R\hspace{0pt}O},D}`$, a certificate of $`\mathcal{T}\hspace{0pt}\mathcal{S}\hspace{0pt}\mathcal{S}_{{R\hspace{0pt}O},D}`$, security policy $`\mathcal{S}\hspace{0pt}\mathcal{P}_{{R\hspace{0pt}O},D}`$ and the nonce $`N_{{R\hspace{0pt}O},D}`$. Having received the target subsystem’s message, $`\mathcal{T}\hspace{0pt}\mathcal{S}\hspace{0pt}\mathcal{S}_{{R\hspace{0pt}O},S}`$ verifies the state of $`\mathcal{T}\hspace{0pt}\mathcal{S}\hspace{0pt}\mathcal{S}_{{R\hspace{0pt}O},D}`$. If the target system is in a trustworthy state and holds an acceptable security policy and system configuration, the current state of $`\mathcal{T}\hspace{0pt}\mathcal{S}\hspace{0pt}\mathcal{S}_{{R\hspace{0pt}O},S}`$ is locked to nonce $`N_{{R\hspace{0pt}O},D}`$.
The $`\mathcal{T}\hspace{0pt}\mathcal{S}\hspace{0pt}\mathcal{S}_{{R\hspace{0pt}O},S}`$ generates a symmetric migration key $`K_{M}`$, serialises its instance and encrypts it with the migration key, which is bound to an acceptable configuration of $`\mathcal{T}\hspace{0pt}\mathcal{S}\hspace{0pt}\mathcal{S}_{{R\hspace{0pt}O},D}`$. Next, the key-blob and the encrypted instance are sent to the destination $`\mathcal{T}\hspace{0pt}\mathcal{S}\hspace{0pt}\mathcal{S}_{{R\hspace{0pt}O},D}`$. In particular, this includes the whole isolated key-hierarchy $`\mathcal{K}_{{\mathcal{R}\hspace{0pt}\mathcal{O}},S}`$ with $`S\hspace{0pt}R\hspace{0pt}K_{{\mathcal{R}\hspace{0pt}\mathcal{O}},S}`$, the security policy $`\mathcal{S}\hspace{0pt}\mathcal{P}_{{R\hspace{0pt}O},S}`$, and the required subsystem configuration $`\mathcal{S}\hspace{0pt}\mathcal{C}_{{R\hspace{0pt}O},S}`$.
Finally, the target subsystem $`\mathcal{T}\hspace{0pt}\mathcal{S}\hspace{0pt}\mathcal{S}_{{R\hspace{0pt}O},D}`$ decrypts the received blob and uses $`S\hspace{0pt}R\hspace{0pt}K_{{R\hspace{0pt}O},S}`$ as its own $`S\hspace{0pt}R\hspace{0pt}K`$. The subsystem verifies the obtained security policy $`\mathcal{S}\hspace{0pt}\mathcal{P}_{{R\hspace{0pt}O},S}`$ and the subsystem configuration $`\mathcal{S}\hspace{0pt}\mathcal{C}_{{R\hspace{0pt}O},S}`$. With this information, $`\mathcal{T}\hspace{0pt}\mathcal{S}\hspace{0pt}\mathcal{S}_{{R\hspace{0pt}O},D}`$ rebuilds the internal structure of the source.
The source system should then be notified of the success of migration and ultimately delete the migrated key hierarchy (or even do it before sending the migration package as indicated for simplicity in Figure [6](#S2.F6 "Figure 6 ‣ II-D Migration of a Trusted-Subsystem ‣ II TCG MPWG Reference Architecture ‣ On the deployment of Mobile Trusted Modules")). Otherwise one obtains replicated trusted subsystems, by themselves indistinguishable to the remote owner. But this may depend on the policies to be enforced in the particular use case.
[Refer to caption](/html/0712.2113/assets/x7.png)
Figure 7: Trustworthy Operating Platform with multiple Trusted Engines
## III Design of Mobile Trusted Modules on Trustworthy Operating Platforms
A prototypical implementation of the trusted engines and the specified trusted services was realised as an extension to the existing EMSCB / Turaya Computing Platform. Turaya is an implementation of the EMSCB security architecture. It provides fundamental security mechanisms and a protected and isolated execution environment, which meet the requirements of the MPWG Reference Architecture \[[8](#bib.bib8), [11](#bib.bib11), [12](#bib.bib12)\].
Figure [7](#S2.F7 "Figure 7 ‣ II-D Migration of a Trusted-Subsystem ‣ II TCG MPWG Reference Architecture ‣ On the deployment of Mobile Trusted Modules") illustrates our model, in which a hypervisor/microkernel executes a legacy operating system in coexistence with a running instance of the EMSCB-based security architecture. The latter controls a virtual machine with several trusted engines and services compliant to the MPWG requirements \[[4](#bib.bib4), [1](#bib.bib1)\]. In the following paragraphs, we outline the significant platform layers concerning our approach.
The Hardware Layer of our model includes a generic MTM as described in Section [](#S2.SS2 "II-B Mobile Trusted Module ‣ II TCG MPWG Reference Architecture ‣ On the deployment of Mobile Trusted Modules")
II-B
, in addition to conventional hardware components. This MTM acts as a dedicated master trust anchor for the complete trusted mobile platform.
The Virtualisation Layer provides generic hardware abstraction, between the physical hardware of a trusted mobile platform and the Trusted Software Layer below. The EMSCB project supports microkernels of the L4-family \[[14](#bib.bib14)\] such as hypervisors \[[15](#bib.bib15)\]. In general, all solutions provide mechanisms for resource management, inter-process-communication (IPC), virtual machines, memory management and scheduling. In our specific case, the virtualisation layer includes also a fully functional device driver for a dedicated generic MTM. Furthermore, it is responsible for instantiation of both the trusted software layer and the legacy operating system.
The Trusted Software Layer provides security functionality and is responsible for isolation of embedded applications and software compartments. It also implements the vMTM Proxy Service as described in Section [](#S2.SS2 "II-B Mobile Trusted Module ‣ II TCG MPWG Reference Architecture ‣ On the deployment of Mobile Trusted Modules")
II-B
. Currently, EMSCB/Turaya provide an excellent foundation by its security services (trust manager, compartment manager, storage manager), which are required by the RTR and RTV, Protected Storage and Trusted Engines Management Agent of $`\mathcal{T}\hspace{0pt}\mathcal{E}_{D\hspace{0pt}M}`$. Therefore, it is reasonable to build the significant parts of the device manufacturer engine $`\mathcal{T}\hspace{0pt}\mathcal{E}_{D\hspace{0pt}M}`$ within this layer.
Trusted engines $`\mathcal{T}\hspace{0pt}\mathcal{E}_{\sigma}`$ within the Application Layer are implemented as parallel and isolated L4Linux compartments \[[13](#bib.bib13)\] on behalf of different stakeholders. Each compartment has access to its vMTM instance through an embedded client-side device driver. This driver constrains the functionality with respect to its specific use case (MRTLM or MLTM). Furthermore, $`\mathcal{T}\hspace{0pt}\mathcal{E}_{\sigma}`$ has an $`R\hspace{0pt}T\hspace{0pt}E_{\sigma}`$, which is responsible for building all required allocated trusted resources and services depending of its specific system configuration $`\mathcal{S}\hspace{0pt}\mathcal{C}_{\sigma}`$ and security policy $`\mathcal{S}\hspace{0pt}\mathcal{P}_{\sigma}`$.
## IV Conclusion and Further Work
We have introduced the Trusted Engines and MTMs in terms of our objective. In this context, we have exposed significant parts of the MPWG Reference Architecture and how it can be implemented on a (very slightly modified) TPM trust-anchor. We have shown how to deploy trusted virtualised compartments on devices and exhibited basic operations required in the mobile domain, such as migration.
Using a vMTM in lieu of a Subscriber Identity Module (SIM) as a trusted and protected software allows expansion to a much wider field of authentication and identification management systems even on standard PC platforms \[[16](#bib.bib16)\]. Supporting online transactions by authentication via credentials held in a vMTM may be one attractive use case. However, there are some privacy and security challenges associated with this implementation on a desktop computer, which require further research. Finally, replacing SIMs/USIMs by multi-purpose vSIMs may be attractive even for genuine mobile devices.
## References
- \[1\] TCG. TCG MPWG Mobile Trusted Module specification, version 1.0, Revision 1, 12 June 2007. <a href="https://www.trustedcomputinggroup.org/specs/mobilephone/tcg-mobile-trusted-module-1.0.pdf" target="_blank">https://www.trustedcomputinggroup.org/specs/mobilephone/tcg-mobile-trusted-module-1.0.pdf</a>
- \[2\] TCG. TCG specification architecture overview, specification revision 1.2, 2004.
- \[3\] TCG. Trusted Platform Module specification (TPM), version 1.2, revision 94, 2005. <a href="https://www.trustedcomputinggroup.org/specs/TPM" target="_blank">https://www.trustedcomputinggroup.org/specs/TPM</a>
- \[4\] Trusted Computing Group, *TCG Mobile Reference Architecture*. Specification version 1.0, Revision 1. 12 June 2007. <a href="https://www.trustedcomputinggroup.org/specs/mobilephone/tcg-mobile-reference-architecture-1.0.pdf" target="_blank">https://www.trustedcomputinggroup.org/specs/mobilephone/tcg-mobile-reference-architecture-1.0.pdf</a>
- \[5\] Nicolai Kuntze and Andreas U. Schmidt. Transitive trust in mobile scenarios.
In Günter Müller, editor, Proceedings of the International Conference on Emerging Trends in Information and Communication Security (ETRICS 2006), volume 3995 of Lecture Notes in Computer Science (LNCS), pages 73–85. Springer-Verlag, 2006.
- \[6\] Kuntze, N., Schmidt, A.U.: Trusted computing in mobile action. In Venter, H.S., Eloff, J.H.P., Labuschagne, L., Eloff, M.M., eds.: Proceedings of the Information Security South Africa (ISSA) Conference (2006)
- \[7\] Nicolai Kuntze and Andreas U. Schmidt. Trusted ticket systems and applications.
In To appear in: New Approaches for Security, Privacy, and Trust in Complex Systems. Proceedings of the IFIP sec2007. Sandton, South Africa 14-16 May 2007. Springer-Verlag, 2007.
- \[8\] EMSCB. European multilaterally secure computing base - towards trustworthy systems with open standards and trusted computing. 2006. <a href="http://www.emscb.com/" target="_blank">http://www.emscb.com/</a>
- \[9\] Stefan Berger, Ramon Caceres, Kenneth A. Goldman, Ronald Perez, Reiner Sailer, and Leendert van Doorn. vTPM: Virtualizing the trusted platform module.
- \[10\] Mario Strasser. Linux TPM emulator, 2005.
- \[11\] Ahmad-Reza Sadeghi, Christian Stüble, and Norbert Pohlmann. European multilateral secure computing base - open trusted computing for you and me.
Datenschutz und Datensicherheit (DuD), pages 548–554, 2004.
- \[12\] PERSEUS. The perseus security framework. http://www.perseus-os.org. Applied Data Security Group, Ruhr-University Bochum.
- \[13\] Michael Hohmuth.
Linux-Emulation auf einem Mikrokern.
PhD thesis, TU Dresden, Fakultät Informatik, Lehrstuhl Betriebssysteme, 1996.
- \[14\] R. Baumgartl, M. Borriss, Cl.-J. Hamann, M. Hohmuth, L. Reuther, S. Schönberg, and J. Wolter. Dresden realtime operating system (drops).
In Workshop of System-Designed Automation, (SDA’98), 1998.
- \[15\] P. Barham. Xen and the art of virtualization, 2003.
- \[16\] Jane Dashevsky, Edward C. Epp, Jose Puthenkulam, and Mrudula Yelamanchi. SIM trust parameters.
Intel Developer Update Magazine, 2003. |
0801.3924 | cs.CR | cyber | # Increased security through open source
<sup>†</sup>
<sup>†</sup>thanks: Id: oss-acm.tex,v 1.16 2005/04/26 17:44:29 jhh Exp
Jaap-Henk Hoepman, Bart Jacobs
Security of Systems (SoS) group
Institute for Computing and Information Sciences
Radboud University Nijmegen
P.O. Box 9010, 6500 GL Nijmegen, the Netherlands
{jhh,bart}@cs.ru.nl
## 1 Introduction
The last few years have shown a worldwide rise in the attention for, and actual use of, open source software (OSS), most notably of the operating system Linux and various applications running on top of it. Various major companies and governments are adopting OSS. As a result, there are many publications concerning its advantages and disadvantages. The ongoing discussions cover a wide range of topics, such as Windows versus Linux, cost issues, intellectual property rights, development methods, etc. Here we wish to focus on security issues surrounding OSS. It has become a reasonably well-established conviction within the computer security community that publishing designs and protocols contributes to the security of systems built on them. But should one go all the way and publish source code as well? That is the fundamental question that we wish to address in this paper.
The following analogies may help to introduce the issues and controversies surrounding the open source debate.
- •
Would you, while travelling far from home, take medicines of an unknown brand given to you by a self-proclaimed “doctor”, without documentation, and hence without (independent) assurance about the nature and proper working of the ingredients?
- •
Who would you trust most? A locksmith who keeps the working of his locks secret, so that thieves cannot exploit this knowledge? Or a locksmith who publishes the workings of his locks, so that everyone (including thieves) can judge how good/bad they are (so you exclusively rely on the complexity of the keys for protection)?
In the remainder of this paper we will discuss the impact of open source on both the security and transparency of a software system. We focus on the more technical aspects of this issue (and refer to Glass \[[4](#bib.bib4)\] for a discussion of the economical perspective of open source), combining and extending arguments developed over the years \[[12](#bib.bib12)\]. We stress that our discussion of the problem only applies to software for general purpose computing systems. For embedded systems, where the software usually cannot easily be patched or upgraded, different considerations may apply.
### 1.1 Security through obscurity: design vs. implementation
Through the centuries, secrecy was the predominant methodology surrounding the design of any secure system. Security of military communication systems, for example, was mostly based on the fact that only few people knew how it worked, and not on any inherently secure method of communication. Ciphers in those days were very insecure.
In 1883, Auguste Kerckhoffs \[[5](#bib.bib5)\] extensively argued that any secure military system ”… must not require secrecy and can be stolen by the enemy without causing trouble”. In the academic security community Kerckhoffs’ Principle is widely supported: in the design of a system, *security through obscurity* is considered bad practice, for many reasons similar to the ones we will discuss later on. This point is starting to get across to industry as well, witnessed by the fact that, for instance, the security of the third generation of cellular telephone networks (UMTS) is based on open and published standards.
In Kerckhoffs’ days there was hardly any difference between the *design* of a system and its actual *implementation*. These days, however, the difference is huge: system designs are already very complex, and their implementation is hard to get completely right. The question then arises whether Kerckhoffs’ Principle applies only to the design of a system, or also to its implementation. In other words, should secure systems also be open source, or not?
There is no agreement on the answer to this question even in the academic community \[[1](#bib.bib1)\]. According to us, the answer is: ”absolutely!”. In the remainder of this paper we will argue why.
### 1.2 Security, risk and exposure
When discussing whether open source makes systems more secure, we have to be precise about what we mean by that. In fact, for the purpose of this discussion we need to distinguish between the *security* of a system, the *exposure* of that system, and the *risk* associated with using that system. We define these terms next.
The ultimate decisive factor that determines whether a system is “secure enough” is the *risk* associated with using that system. This risk is defined as a combination of the likelihood of a successful attack on a system together with the damage to the assets resulting from it.
The *exposure* of a system completely ignores the damage that is incurred by a successful attack, and is defined as just the likelihood of a successful attack. This depends on several factors, like the number and severity of vulnerabilities in the system, but also whether these vulnerabilities are known to attackers, how hard it is to exploit such a vulnerability, and whether the system is a high-profile target or not.
Finally, we consider the *security* of a system to be an objective measure of the number of its vulnerabilities and their severity (i.e., the privileges obtained by exploiting the vulnerability).
To summarise, exposure combines security with the likelihood of attack, and risk combines exposure with the damage sustained by the attack. We note that in other papers on this and similar topics, security has been used to mean either security proper, exposure or risk as defined above. With these definitions in place, we see that opening the source clearly does not change the security of a system (simply because it doesn’t introduce new bugs), while the exposure is likely to increase in the short term (because it makes the existing bugs more visible). The question is what happens to the security and the exposure of an open source system in the long run.
### 1.3 Open vs. closed source
The increased attention paid to open source in the media and by society at large has made open source an almost catch-all phrase. Here, we use it in its original, rather specific, meaning. Open source software is software for which the corresponding source (and all relevant documentation) is available for inspection, use, modification and redistribution by the user
<sup>1</sup>
<sup>1</sup>1See <a href="http://www.opensource.org/" target="_blank">http://www.opensource.org/</a>.
. We do not distinguish between any kind of development methodology (e.g., the Cathedral or the Bazaar\[[10](#bib.bib10)\]). Neither do we care about the pricing model (freeware, shareware, etc.). We do assume however that users (in principle) are allowed and able to rebuild the system from the (modified) sources, and that they have access to the proper tools to do so.
In some cases, allowing the user to redistribute the modified sources (in full, or through patches) is also necessary (e.g., Free Software and the GNU Public License
<sup>2</sup>
<sup>2</sup>2<a href="http://www.gnu.org/copyleft/gpl.html" target="_blank">http://www.gnu.org/copyleft/gpl.html</a>
). Most of our arguments also hold for *source available* software, where the license does not allow redistribution of the (modified) source.
## 2 Open source necessary for security
We believe that using open source software is a necessary requirement to build systems that are more secure. Our main argument is that opening the source allows independent assessment of the exposure of a system and the risk associated with using the system, makes patching bugs easier and more likely, and forces software developers to spend more effort on the quality of their code. The remainder of this paper is devoted to arguing our case in detail.
We will first review arguments in favour of keeping the source closed, and then discuss reasons why open source does (in the long run) increase security. As noted in the introduction, there is a distinction between making the design of a system public and also making its implementation public. We focus on the latter case, but note that most (but not all) of these arguments also apply to the question whether the design should be open or not.
### 2.1 Keep the source closed: arguments against open source
First of all, keeping the source closed prevents the attacker from having easy access to information that may be helpful to successfully launch an attack \[[2](#bib.bib2)\]. Opening the source gives the attacker a wealth of information to search for vulnerabilities and/or bugs, like potential buffer overflows, and thus increases the exposure of the system.
Also, there is a huge difference between openness of the design and openness of the source. Openness of the design may reveal logical errors in the security in the worst case. With proper review, these errors can and usually are found. For source code, this is not, or at least not completely, the case. In the foreseeable future, source code will continue to contain bugs, no matter how hard we look, test or verify.
Moreover, opening the source gives unfair advantage to the attacker. The attacker needs to find but *one* vulnerability to successfully attack the system. The defender needs to patch *all* vulnerabilities to protect himself completely. This is considered an uneven battle.
Fourth, there is no direct guarantee that the binary object code running in the computer corresponds to the source code that has been evaluated \[[11](#bib.bib11)\]. People unable or unwilling to compile from source must rely on a trusted third party to vouch for this
<sup>3</sup>
<sup>3</sup>3Or could use tools like systrace to confine untrusted object code, and to enforce a security policy nevertheless \[[9](#bib.bib9)\].
.
Also, making the source public does not guarantee that any qualified person will actually look at the source and evaluate (let alone improve) it. There are many open source projects that, after a brief flurry of activity, are only marginally maintained and quickly sink into oblivion. The attackers, on the other hand, most surely *will* scrutinise the source.
In bazaar style open source projects, back-doors may be sneaked into the source by hackers posing as trustful contributors. That this is not an idle threat became clear in November 2003, when Linux kernel developers discovered a back-door in a harmlessly looking error-checking feature added to a system call
<sup>4</sup>
<sup>4</sup>4<a href="http://www.securityfocus.com/news/7388" target="_blank">http://www.securityfocus.com/news/7388</a>.
.
Finally, and more generally, the quality of a piece of software (and patches to it) depend on the skills of the programmers working on it \[[8](#bib.bib8)\]. For many open source projects there is no a priori selection of programmers based on their skill. Usually any help is appreciated, and there is only rudimentary quality control.
### 2.2 Closed is not so closed: arguments against closed source
Let us first review the arguments put forward against open source in the previous section. The last two arguments against open source are actually aimed at the development methodology instead. The systems developed in that manner would also be more insecure if they were closed source. We assume a minimal standard of proper coding practices, project management, change control and quality control. In fact, one of our main points is that by opening up the source, software projects cannot get away with poor project management and poor quality control so easily.
Now turning to the first argument against closed source, we note that keeping the source closed for a long time appears to be hard \[[7](#bib.bib7)\]. Last year, source code for certain types of voting machines manufactured by Diebold were distributed on the Internet, and subsequent research on that source code revealed horrible programming errors and security vulnerabilities \[[6](#bib.bib6)\]. Recently even parts of the source to Microsoft Windows NT became public. Within days the first exploit based on this source code was published. The Diebold case also revealed how bad coding standards of current closed source systems can be, and how they lead to awfully insecure systems.
Even if the source remains closed, vulnerabilities of such closed source systems will eventually be found and become known to a larger public after a while. Vulnerabilities in existing closed source software are announced on a daily basis. In fact, tools like debuggers and disassemblers allow attackers to find vulnerabilities in applications without access to the source relatively quickly. Moreover, not all vulnerabilities that are discovered will be published. Their discoverers may keep them secret to avoid patches for them, allowing use of the vulnerability to exploit systems for a prolonged period of time. We see that while the perceived exposure of a closed source system may appear to be low, the actual exposure eventually becomes much higher (approaching the exposure that would exist initially if the system were completely open source).
Even worse, only the producer of closed source software can release patches for any vulnerabilities that are found. Many of those patches are released weeks or months after the vulnerability is discovered, if at all. The latter case occurs for instance with legacy software for which the company producing it no longer exists (or refuses to give support for it after a while, e.g., Microsoft Windows NT Server 4.0 or Netscape Calendar). The consequence is that systems stay exposed longer, increasing the risk of using that system.
We see that keeping the source closed actually hurts the defender much more than the attacker: while a determined attacker can still discover weaknesses easily, the defender is prevented from patching them.
Finally, closed source software severely limits the user of such software to evaluate its security for or by himself. The situation improves if at least the design of the system is open. If the system is evaluated by an independent party according to some generally accepted methodology (like the Common Criteria), this gives the user another basis for trusting the security of the software. However such evaluations are rare (because they are expensive), and usually limited to certain restricted usage scenarios or parameter settings that may not correspond to the actual operating environment of a particular user. Moreover, such evaluations apply only to a specific version of the software: new versions need to be reevaluated.
### 2.3 The way forward: arguments supporting open source
We see that the arguments against ”security through obscurity” generally apply to the implementation of a system as well. It is a widely held design principle that the security of a system should only depend on the secrecy of the (user-specific) keys, on the ground that all other information of the system is shared by many other people and therefore will become public as a matter of course.
Moreover, open source enables users to evaluate the security by themselves, or to hire a party of *their* choice to evaluate the security for them. Open source even enables several different and independent teams of people to evaluate the security of the system, removing the dependence on a single party to decide in favour or against a certain system. All this does not decrease the security or exposure of the system. However, it does help to asses the real exposure of the system, closing the gap between perceived and actual exposure.
Open source enables users to find bugs, and to patch these bugs themselves. There is also a potential network effect: if users submit their patches to a central repository, all other users can update their system to include this patch, increasing their security too. Given that different users are likely to find different bugs, many bugs are potentially removed. This leads to more and faster patches, and hence more secure code (this corresponds to ”Linus’s Law”: ”Given enough eyeballs, bugs are shallow” \[[10](#bib.bib10)\]). Evidence suggests that patches for open source software are released almost twice as fast as for closed source software, thus halving the vulnerability period \[[12](#bib.bib12)\].
If a user is unable to patch a bug himself, open source at least enables him to communicate about bugs with developers more efficiently (because both can use the same frame of reference — i.e., the source code — for communication \[[10](#bib.bib10)\]).
Also, open source software enables users to add extra security measures. Several tools exist to enhance the security of existing systems, provided the source is available \[[3](#bib.bib3)\]. These tools do not rely on static checking of the code. Instead, they add generic runtime checks to the code to detect e.g., buffer overflows or stack frame corruptions. Moreover, open source software allows the user to limit the complexity of the system (and thereby increasing its security) by removing unneeded parts.
Finally, and importantly, open source forces developer communities to be more careful, and to use the best possible tools to secure their systems. It also forces them to use clean coding styles (”sloppy” code is untrustworthy), and to put more effort into quality control. Otherwise, companies and individual programmers alike will loose respect and credibility. As a side effect, this will stimulate research and development in new, improved tools for software development, testing and evaluation and perhaps even verification.
## 3 Conclusions
We conclude that opening the source of existing systems will at first increase their exposure, due to the fact that more information about vulnerabilities becomes available to attackers. However, this exposure (and the associated risk of using the system) can now be determined publicly. With closed source systems the perceived exposure may appear to be low, while the actual exposure (due to increased knowledge of the attackers) may be much higher.
Moreover, because the source is open, all interested parties can assess the exposure of a system, hunt for bugs and issue patches for them, or otherwise increase the security of the system. Security fixes will quickly be available, so that the period of increased exposure is short.
In the long run, openness of the source will increase its security. Sloppy code is visible to everyone, and questions even the overall quality of it. Any available tools to validate the source will be used more often by the producers. If not, the users will do it themselves, afterwards. New, much more advanced, tools will be developed to improve the security of software even further. Open source allows users to make a much more informed choice about the security of a system, based on their own or on independent judgement.
It is our conviction that all these benefits outweigh the disadvantages of a short period of increased exposure.
## References
- \[1\] Ross Anderson. Security in open versus closed systems — the dance of Boltzmann, Coase and Moore.
In Conference on Open Source Software Economics, Toulouse (France), June 20–21 2002.
- \[2\] Kenneth Brown. Opening the open source debate. Technical report, Alexis de Tocqueville Institution, June 2002.
- \[3\] Crispin Cowan. Software security for open-source systems.
IEEE J. Security & Privacy, 1(1):38–45, 2003.
- \[4\] Robert L. Glass. A look at the economics of open source.
Comm. ACM, 47(2):25–27, February 2004.
- \[5\] Auguste Kerckhoffs. La cryptographie militaire.
Journal des sciences militaires, IX, 1983.
pp. 5–38, Jan. 1883, and pp. 161–191, Feb. 1883.
- \[6\] T. Kohno, A. Stubblefield, A. D. Rubin, and D. S. Wallach. Analysis of an electronic voting system.
In IEEE Security & Privacy, Berkeley/Oakland, CA, USA, May 9–12 2004. IEEE.
- \[7\] Rebecca T. Mercuri and Peter G. Neumann. Inside Risks: Security by obscurity.
Comm. ACM, 46(11):160, December 2003.
- \[8\] Peter G. Neumann. Inside Risks: Information system security redux.
Comm. ACM, 46(10):136, October 2003.
- \[9\] Niels Provos. Improving host security with system call policies.
In 12th USENIX Sec. Symp., Washington D.C., USA, August 2003. USENIX.
- \[10\] Eric S. Raymond. The cathedral and the bazaar, 2000.
- \[11\] Ken Thompson. Reflections on trusting trust.
Comm. ACM, 27(8):761–763, August 1984.
- \[12\] Brian Witten, Carl Landwehr, and Michael Caloyannides. Does open source improve system security?
IEEE Software, pages 57–61, September – October 2001. |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.