Radio Communications

I
Radio Communications
Edited by
Alessandro Bazzi
In-Tech
intechweb.org
Published by In-Teh
In-Teh
Olajnica 19/2, 32000 Vukovar, Croatia
Abstracting and non-profit use of the material is permitted with credit to the source. Statements and
opinions expressed in the chapters are these of the individual contributors and not necessarily those of
the editors or publisher. No responsibility is accepted for the accuracy of information contained in the
published articles. Publisher assumes no responsibility liability for any damage or injury to persons or
property arising out of the use of any materials, instructions, methods or ideas contained inside. After
this work has been published by the In-Teh, authors have the right to republish it, in whole or part, in any
publication of which they are an author or editor, and the make other personal use of the work.
© 2010 In-teh
www.intechweb.org
Additional copies can be obtained from:
[email protected]
First published April 2010

Printed in India
Technical Editor: Martina Peric

Cover designed by Dino Smrekar
Radio Communications,
Edited by Alessandro Bazzi
p. cm.
ISBN 978-953-307-091-9
V
In the last decades the restless evolution of information and communication technologies
(ICT) brought to a deep transformation of our habits. The growth of the Internet and the
advances in hardware and software implementations modified our way to communicate and
to share information.
Although conceived for study and working scopes, the presence of ICT infrastructures
and services are now pervasive in our life. Their availability is felt as a need that cannot
be constrained to a place: people want to have them anytime, anywhere. And the need for
wireless connections starts from daily communication and embraces all fields of human life,
in an ever increasing way: transportation systems, home management, healthcare, emergency
operations. This brought to the large interest and great evolution that characterized radio
technologies in the last years all over the world. Since the development of the first wireless
systems, great technological advances were performed: spread spectrum and multi-carrier
techniques allowed to make transmissions over larger bandwidths with higher data rates for
the final users in challenging scenarios where line of sight is not guaranteed and multipath
severely affects signal propagation; new coding techniques were developed to increase the
efficiency of radio transmissions; advanced link adaptation and power control algorithms
were implemented in order to allow optimized trade-off between reliability and data rate;
and many other topics shall be cited.
Cellular systems, initially deployed to allow telephony in mobility for urgency needs have
now reached a penetration of almost 100% of inhabitants in developed countries, with the
ability to exchange data at more than 10 Mbit/s; at the same time, wireless access to local
networks can now be performed at almost 100 Mbit/s. But the need for higher data rates does
not stop and it jointly proceeds with the development of new services, thus requiring the
investigation and implementation of new technologies and new radio resources management
solutions. In the near future, all available technologies will be jointly used by smart devices,
following the always best connected paradigm: users will connect to an heterogeneous
wireless network, without the need to know which technology they are effectively using.
The adoption of multiple antennas and beamforming techniques will reduce the effects of
interference and increase the achievable data rates. Relaying and cooperation among devices
will allow better performance with lower costs and energy consumption.
In this book, an overview of the major issues faced today by researchers in the field of radio
communications is given through 35 high quality chapters written by specialists working in
universities and research centers all over the world. Various aspects will be deeply discussed:
channel modeling, beamforming, multiple antennas, cooperative networks, opportunistic
VI
scheduling, advanced admission control, handover management, systems performance

assessment, routing issues in mobility conditions, localization, web security. Advanced
techniques for the radio resource management will be discussed both in single and multiple
radio technologies; either in infrastructure, mesh or ad hoc networks. In particular, the book
is organized following a bottom up structure, starting from the physical level up to the whole
system. More precisely, the physical layer aspects are discussed through chapters 1 to 10,
with particular emphasis to MIMO systems; from chapter 11 to chapter 18, data link layer and
radio resource management are handled without focusing to a specific technology: backward
error correction, scheduling, relaying; a deeper investigation of systems performance, with
particular reference toWiMAX andWiFi systems, is then given through chapters 19 to 25;
heterogeneous wireless networks are then examined through chapters 26 to 29; the cross-layer
issues of localization and web security, which are of increasing interest in modern wireless
technologies, are finally discussed in the remaining part of the book.
Although these pages will not cover all the aspects of such a wide topic, the reader will find
a helpful overview on what is happening now and what is expected in an early future in the
field of radio communications. It must be remarked that this book was possible thanks to the
valuable work of all authors involved and to the IN-TECH technical staff.
Alessandro Bazzi*
*Alessandro Bazzi works at WiLab (www.wilab.org), an organization with affiliates belonging to the
Italian National Council for Researches CNR (ICT Department, IEIIT Institute) and Italian Universities
of Bologna and Ferrara. WiLab has expertise in the field of telecommunications systems with particular
emphasis on wireless systems; research activity is mainly performed within the context of important na-
tional and international projects, in relation with some of the most important manufacturers and service
providers, and in strict cooperation with colleagues coming from the main universities in Europe and
United States.
VII
Preface V
1. Radio-communications architectures 001

Antoine Diet, Martine Villegas, Geneviève Baudoin and Fabien Robert
2. Analytical SIR for Cross Layer Channel Model 037

Abdurazak Mudesir and Harald Haas
3. The Impact of Fixed and Moving Scatterers on the Statistics

of MIMO Vehicle-to-Vehicle Channels 051
Ali Chelli and Matthias Pätzold
4. Planar Antenna Array Hybrid Beamforming for SDMA

in Millimeter Wave WPAN 065
Sau-HsuanWu, Lin-Kai Chiu, Ko-Yen Lin and Ming-Chen Chiang
5. A Distributed Multilayer Software Architecture for MIMO Testbeds 077

José A. García-Naya, M. González-López and L. Castedo
6. Recent Developments in Channel Estimation and Detection

for MIMO Systems 099
Seyed Mohammad-Sajad Sadough and Mohammad-Ali Khalighi
7. Cooperative MIMO Systems in Wireless Sensor Networks 123

M. Riduan Ahmad, Eryk Dutkiewicz, Xiaojing Huang and M. Kadim Suaidi
8. Optimal Cooperative MIMO Scheme in Wireless Sensor Networks 151

M. Riduan Ahmad, Eryk Dutkiewicz, Xiaojing Huang and M. Kadim Suaidi
9. Single/Multi-User MIMO Differential Capacity 167

Daniel Castanheira and Atílio Gameiro
10. Low Dimensional MIMO Systems with Finite Sized Constellation Inputs 185
Rizwan Ghaffar and Raymond Knopp
11. Advanced Hybrid–ARQ Receivers for Broadband MIMO Communications 211

Tarik Ait-Idir, Houda Chafnaji, Samir Saoudi and Athanasios Vasilakos
12. Cooperative ARQ: A Medium Access Control (MAC) Layer Perspective 227
Jesús Alonso-Zárate, Elli Kartsakli, Luis Alonso and Christos Verikoukis
VIII
13. A Hybrid Feedback Mechanism to Exploit Multiuser Diversity

in Wireless Networks 247
Yahya S. Al-Harthi
14. Opportunistic Access Schemes for Multiuser OFDM Wireless Networks 265
Cédric Gueguen and Sébastien Baey
15. Bidirectional Cooperative Relaying 281

Prabhat Kumar Upadhyay and Shankar Prakriya
16. A Novel Amplify-and-Forward Relay Channel Model for Mobile-to-Mobile

Fading Channels Under Line-of-Sight Conditions 307
Batool Talha and Matthias Pätzold
17. Resource Management with Limited Capability of Fixed Relay Station

in Multi-hop Cellular Networks 321
Jemin Lee and Daesik Hong
18. On Cross-layer Routing in Wireless Multi-Hop Networks 339

Golnaz Karbaschi, Anne Fladenmuller and Sébastien Baey
19. Mobile WiMAX Performance Investigation 361

Alessandro Bazzi, Giacomo Leonardi, Gianni Pasolini and Oreste Andrisano
20. Throughput-Enhanced Communication Approach for Subscriber Stations

in IEEE 802.16 Point-to-Multipoint Networks 385
Chung-Hsien Hsu and Kai-Ten Feng
21. Holdoff Algorithms for IEEE 802.16 Mesh Mode in Multi-hop

Wireless Mesh Networks 399
Bong Chan Kim and Hwang Soo Lee
22. Call Admission Control Algorithms based on Random Waypoint Mobility

for IEEE802.16e Networks 419
Khalil Ibrahimi, Rachid El-Azouzi, Thierry Peyre and El Houssine Bouyakhf
23. Queueing-Model-Based Analysis for IEEE802.11 Wireless LANs

with Non-Saturated Nodes 441
Shigeo Shioda and Mayumi Komatsu
24. Increasing the Time Connected to Already Deployed 802.11 Wireless Networks
while Traveling by Subway 457
Jaeouk Ok, Pedro Morales, Masateru Minami and Hiroyuki Morikawa
25. Asymmetric carrier sense in heterogeneous medical networks environment 473

Bin Zhen, Huan-Bang Li, Shinsuke Hara and Ryuji Kohno
26. Multi-Agent Design for the Physical Layer of a Distributed Base Station Network 493
Philippe Leroux and Sébastien Roy
IX
27. Inter-RAT Handover Between UMTS And WiMAX 523

Bin LIU, Philippe Martins, Philippe Bertin and Abed Ellatif Samhat
28. MHD-CAR: A Distributed Cross-Layer Solution for Augmenting Seamless

Mobility Management Protocols 551
Faqir Zarrar Yousaf, Christian Müller and Christian Wietfeld
29. Mobility in IP Networks: From Link Layer to Application Layer Protocols

and Architectures 573
Thienne Johnson, Eleri Cardozo, Rodrigo Prado, Eduardo Zagari and Tomas Badan
30. Positioning in Indoor Mobile Systems 597

Miloš Borenović and Aleksandar Nešković
31. Location in Ad Hoc Networks 619

Israel Martin-Escalona, Marc Ciurana and Francisco Barcelo-Arroyo
32. Location Tracking Schemes for Broadband Wireless Networks 639

Po-Hsuan Tseng and Kai-Ten Feng
33. Wireless Multi-hop Localization Games for Entertainment Computing 651

Tomoya Takenaka, Hiroshi Mineno and Tadanori Mizuno
34. Measuring Network Security 673

Emmanouil Serrelis and Nikolaos Alexandris
35. A testing process for Interoperability and Conformance of secure Web Services 689
Spyridon Papastergiou and Despina Polemi
Radio-communications architectures 1
Radio-communications architectures
Antoine Diet*, Martine Villegas**, Geneviève Baudoin** and Fabien Robert**
1. Introduction
Wireless communications, i.e. radio-communications, are widely used for our different
daily needs. Examples are numerous and standard names like BLUETOOTH, WiFI,
WiMAX, UMTS, GSM and, more recently, LTE are well-known [Baudoin et al. 2007].
General applications in the RFID or UWB contexts are the subject of many papers. This
chapter presents radio-frequency (RF) communication systems architecture for mobile,
wireless local area networks (WLAN) and connectivity terminals. An important aspect of
today’s applications is the data rate increase, especially in connectivity standards like WiFI
and WiMAX, because the user demands high Quality of Service (QoS). To increase the data
rate we tend to use wideband or multi-standard architecture. The concept of software radio
includes a self-reconfigurable radio link and is described here on its RF aspects. The term
multi-radio is preferred. This chapter focuses on the transmitter, yet some considerations
about the receiver are given. An important aspect of the architecture is that a transceiver is
built with respect to the radio-communications signals. We classify them in section 2 by
differentiating Continuous Wave (CW) and Impulse Radio (IR) systems. Section 3 is the
technical background one has to consider for actual applications. Section 4 summarizes
state-of-the-art high data rate architectures and the latest research in multi-radio systems.
In section 5, IR architectures for Ultra Wide Band (UWB) systems complete this overview;
we will also underline the coexistence and compatibility challenges between CW and IR
systems.
2. Transceiver aspects for radio-communications

2.1 Radio communications signals
Radio-communications applications deal with communicating and non-communicating
links with their different parameters. People expect high quality from their different
services (QoS) whatever the telecommunications system used. For example, voice (low data
rate) or visio-phone and multimedia download (high data rate) are assumed to be present
on the new generation mobile phone. This reveals the co-existence and interaction goals
between mobile communication systems (GSM, GSM EDGE, UMTS…) and connectivity
standards (BLUETOOTH, WiFI, WiMAX…) [Baudoin et al., 2007]. Thanks to Impulse Radio
2 Radio Communications
Ultra Wide Band (IR-UWB), other fields of interest such as Radio Frequency Identification
(RFID) and localization systems are examples of where radio-communications transceivers
are currently being designed. Each kind of application can be classified by the resulting
radio signals emitted/received. Determining factors are (i) the use of power efficient or
spectrum efficient modulation schemes, (ii) the frequency and type of carrier signal used:
Continuous Wave (CW) or Impulse Radio (IR) based signals and (iii) the data rate needed
(defining a major subdivision of CW based systems). Depending on the choice of the factors
involved, the design faced by the RF architect can be varied and challenging. We
differentiate three types of cases in this chapter, as illustrated in Fig. 1.
Time signal
Typical Modulation Scheme Frequency spectrum PROS.
(with shaping filter for CW)
Power efficient
NB-CW
PAPR due to shape filtering
CW PAPR due to multi-carrier

Multi-carrier case
Spectrum efficient
and shape filtering
(ex. OFDM)
WB-CW
Mean power depending on
Spread spectrum
repetition frequency
Power saving
IR-UWB
IR
Fig. 1. Main types of radio-communications signals
1) NarrowBand CW (NB-CW) systems, like GSM (GMSK), EDGE (D-QPSK), BLUETOOTH,

RFID tags, etc… These systems are often power efficient modulation schemes (EDGE is an
exception) because these applications use a low data rate transfer. Major problems of the
NB-CW architecture involve with the coexistence and the signal protection against
interferers or blocking signals. Some spread-spectrum techniques are often added to
improve the communication range (frequency-hopping in the case of BLUETOOTH). The
NB-CW case is also considered to be the classic radio communications link and reference
system because it corresponds to the popular AM/FM radio broadcasts. NB-CW systems
are often considered as using constant envelope signals. This could be true only if FSK
modulation schemes are performed, but it is also possible to use x-QAM low symbol rate
modulations (EDGE). Additionally, the shaping filter (root raised cosine filter) implies
amplitude variations (non-constant envelope) even on FSK modulated signals (GSM).
2) WideBand CW (WB-CW) systems, like: UMTS (W-CDMA), WiMAX (OFDM enhanced),

LTE (OFDM based), WiFI (OFDM), UWB (OFDM version)… These systems correspond to
high data rate transfer, often for multimedia applications. We can see an increasing need in
this field of interest due to the large number of new standards which can be found. Due to
the bandwidth limitations for each standard, the modulation scheme is often spectrum
efficient (x-QAM) and the use of multi-carrier transmission is usually performed, e.g.,
OFDM or MC-CDMA. Since there are high amplitude variations of such signals, the
transceiver architecture is designed in function with the unavoidable non-linear effects (NL)
caused by the power amplification block. Emitting a high PAPR (Peak to Average Power
Ratio) has always been a well-known transmitter challenge. The linearization of such a
transmitter is mandatory and will be discussed in part 3. WB-CW transceivers also need a
wideband design for all of its key elements: antennas, LNA (Low Noise amplifier), HPA
(High Power Amplifier) and mixers. This often results in lower performance of the above-
mentioned blocks than for NB-CW systems.
3) IR-UWB systems such as UWB localization systems, RFID-UWB… These systems are
special because they are based on (one of the) spread spectrum techniques in order to
protect the information. The idea is to spread the information in frequency while lowering
the emitted power. The use of UWB (3.1-10.6 GHz in the USA) was highly discussed in
order to evaluate its co-existence with NB-CW and WB-CW. What is important here is how
the power amplification is processed differently for this type of communication. An average
power is defined in function of the Pulse Repetition Frequency (PRF), given by the
specifications of the IR-UWB standard. For a fixed emitted power: the shorter the pulse, the
higher the instantaneous power and bandwidth. Transceivers for these signals are based on
impulse generators and energy detectors or correlation receivers, see section 5.
35
20
-20
dBm/MHz
-40
-60
dBm, dBm
-80
-100
-120
-140
-160
-174
0 500 1000 1500 2000 2500 3000 3500 4000 5000 6000
Fig. 2. Telecommunications spectrum sharing and power limitations up to 6 GHz
The three types described can represent every kind of radio-communications signal. The
characteristics which have an impact on the architecture are mainly the centre frequency
(choice of the technology) and bandwidth (circuits’ topologies and performance limitations),
and also the PAPR for CW signals. Fig. 2 qualitatively summarizes the power limitations and
frequency specifications for some telecommunications standards up to 6 GHz: cellular,
WiMAX, WiFI, Bluetooth, UWB and DVB-H. The goal of RF architecture is to emit and receive
such signals with no alteration of the information (no constellation distortion for CW).
Designing the architecture implies other considerations such as noise, linearity, efficiency and
systems co-existence and immunity. The receiver part of a transceiver has to correctly identify
information without adding too much noise, even when high power unwanted signals are
close (in the frequency domain). The transmitter part has to linearly amplify the signal in
order to emit as far as possible, respecting the standard power limitations (spectrum mask) for
co-existence. Architectures for NB-CW, WB-CW and IR-UWB use some basic elements
(blocks) that will be described in the next sub-section. Multi-radio is interpreted as a possible
reconfigurable architecture for most of the signals presented (mainly NB-CW and WB-CW).
This helps drive improvements on classic structures.
2.2 Basic elements and their imperfections

Transceiver architectures for radio-communications signals is defined from the Digital to
Analog Converter (DACs, in the baseband section) to the transmitter (Tx) antenna and,
respectively, from the receiving (Rx) antenna to the Analog to Digital Converter (ADCs).
Each of the Tx and Rx sections deal with unavoidable tradeoffs such as linearity/efficiency
(Tx), noise/gain (Rx). Other transceivers in the spectrum vicinity are supposed to correctly
receive (co-existence) and/or emit their signals (immunity) without lowering QoS. The basic
functions (blocks) in radio-communications are conversion (digital/analog), high frequency
transposition and modulation, filtering, power or low noise amplification and
radiation/sensing (Tx/Rx antenna). Here, we are describing a system and how it relates to
these blocks in radio-communications architecture. We will focus on their imperfections and
their impact on the system performance (noise, spectrum distortion…). Sometimes, for CW
standards, the influence of noise or spectral re-growths is quantified by certain criteria such
as the EVM and/or the ACPR as defined in Fig. 3 [Baudoin et al., 2007][Villegas et al., 2007].
Spectrum
Ai Ei
R
R0
Bi
Bi : Emitted Ai : Ideal Ei : Vectorial Error P0 Padj
frequency
1 N N
DSP f .df
  
2 2
A i  C1B i  C 0 A i  Si
N P
EVM rms  i 1
 i 1
ACPR  0 
P0
1 N N
 Ai
2
 Ai
2
Padj  DSP f .df
N i 1 i 1 Padj
Fig. 3. EVM and ACPR definitions
- DACs and ADCs will not be presented in details because their performance in terms of
bandwidth (up to 100 MHz) and resolution (up to 16 bits) is almost sufficient for today’s
radio-communications signals. The main limitations are the current consumption of fast
converters and the difficulty in designing near-GHz Sigma Delta (ΣΔ) encoders, often in the
context of polar transmitters as is presented in section 4. It should be noticed that for WB-
CW signals, the bandwidth limitation is the criterion of choice and a source of important
spectrum degradation (approximated qualitatively by a windowing effect).
I  Ak .cos ( k )
G.cos ωt 
Quadrature 0
PLL
+ Smod 
G.A k
1  a.cos θ.cos ωt   k  Transposed information
mismatch 90°
- 2
G.sin ωt  θ  
G.A k
1  a.cos θ.cos ωt   k 
2 Images of I and Q
Q  a.Ak .sin ( k )  Vdc
 
 G.A k .a.sin θ sin  k sin ωt 
Gain imbalance
« Offset »  Vdc .G.sin ωt  θ  Local oscillator signal
  1 - a.cos θ 2   2.a.sin θ 2 

IRR  10.log10 

  1  a.cos θ 2 

2
 
LOR  10.log10 2. VDC 
 Ak.1  a.cosθ 
 
a = 1.2, θ = 5°, Vdc = 0 a = 1, θ = 0°, LOR = -24 dB
T T
 - sin θ  
T
 I out  I   1 0 
   Q     Vdc  
 Q out     - a.sin θ   
a.cos θ   cos θ  
Fig. 4. IQ modulator equations and the effects of its imperfections
- The IQ modulator is the block providing the transposition of the information at high
frequency (up-converter) or at baseband/intermediate frequency (down-converter). We
describe the case of the up-converter for simplicity. It needs the baseband information
components I and Q (I and Q channels) and a carrier frequency signal, provided by a
frequency synthesizer. Key components are the non-linear multipliers provided thanks to
passive (PIN diodes…) or active circuits (Gilbert cell…). Every component should be
carefully designed with regards to the synthesizer frequency value. Moreover, the noise
added by the multipliers will be impossible to filter. The main difficulty for the modulator is
the perfect matching between I and Q paths. As is reported in Fig. 4 (at one fixed
frequency), gain and phase imbalances (aand ) create unwanted images of the information.
Also the presence of an offset V DC can create the emission of the synthesized carrier
frequency (called Local Oscillator, LO). All of these imperfections result in the distortion of
the information and are quantified by the IRR and LOR power ratio (the actual performance
is in the range of -50 dB). These imperfections are not seen on the spectrum because their
equivalent added noise is inside the main lobe. The biggest challenge concerns the
modulation of wideband signals because it is hard for the modulator to perform and to be
frequency independent (over the entire bandwidth). An example of this can be seen in the
conversion gain (ratio between the RF output power and the baseband input power). A
variation of this gain produces an unwanted AM on the emitted signal and can distort the
information. Another problem is the image frequency. This effect is due to the
multiplication of signals in a modulator. A multiplication of harmonic signals results in two
signals whose frequencies are, respectively, the sum and the difference of the input
frequencies. This duality is mathematically illustrated and discussed for the case of the
receiver in Fig. 16, section 3. At the outgoing emission, an image is created at RF and
increases the EVM. At the reception (only for heterodyne architectures), an image frequency
different from RF can be sensed and interpreted as unwanted information which increases
the EVM, too. This is illustrated in section 3.3 with the presentation of the Hartley and
Weaver image rejection receiver architectures for NB-CW. To conclude, the effects of the IQ
modulator imperfections are: non-linearities, noise and the possibility of image frequencies.
Selective filtering and stable reference signals are needed to improve the system (easier for
NB-CW than for WB-CW).
- Frequency synthesizers are blocks which produce a stable reference signal for
transposition. RF architectures need flexibility and stability in the choice of their reference
frequency. It is usually not possible to provide this simultaneously with a simple oscillator
circuit such as Quartz (stable), SAWR… The stabilization and synthesis by a Phase Lock
Loop (PLL) is usually necessary. The PLL is a looped system whose different designs will
not be discussed here, only its system’s characteristics, which are reported in Fig. 5. The
worst imperfection of the PLL is its phase noise. An example of the resulting phase noise
profile of a synthesizer is illustrated in Fig. 5 (N=1). The different transfer functions of each
noise of the sub-blocks are reported in the same figure. The great challenge of the PLL is to
keep the stability performance of the synthesized signals for different values of N. It is also
possible to modulate the signal in frequency by directly adding the baseband information
on the Voltage Controlled Oscillator (VCO) input. This is called an “over the loop”
modulation. For CW systems, synthesizers are used to produce the different Local
Oscillators (LOs) needed for transposition(s) and for channel selection (fine tuning of the LO
value). PLL can be used to transpose the information, but only for angular modulations
schemes: PM, PSK, FSK or FM. This is not usually used for constant envelope WB-CW
signals for noise and stability reasons.
θbref θbvco Synthesized frequency

Power
θout VCO
θref Kφ.θbcomp Resulting Phase Noise
F(p) Kvco/p
Phase Noise
θcomp
1/N
Reference
θbdiv Phase Noise
F  p
Koct KΦ
p θbvco
θout  θ ref  θbref  θbdiv  θbcomp 
 F  p   F  p 
1   Kvco KΦ  1   Kvco KΦ 
 Np   Np 
   
Loop Filter
bandwidth Noise Floor
θout  G  p θref  θbref  θbdiv  θbcomp  H p θbvco
Frequency
Fig. 5. PLL functional blocks and frequency synthesizer phase noise
- RF filters are essential in communications chains for information selectivity (interferences,

noise, image-filtering for example). They are used at emission (limiting spectral re-growths),
reception (rejecting unwanted signals) and for channel selection (high selectivity for
discrimination and for image rejection). Their system characteristics are well-known:
attenuation/rejection, selectivity, ripple, group delay… Different technologies are used
depending on the frequency, implying different sources of imperfections.
At GHz frequencies, LC filtering is preferred for its low noise property but the components’
sensitivity (especially for integrated technology) implies low order filters. For higher
selectivity, active or high speed digital filters can be used, but often need a frequency
transposition section due to the circuit bandwidth and the sampling rate limitations.
Moreover, these filters consume power and add much noise (circuit or quantification).
Whatever the technology used, the wider the bandwidth the higher the ripple (oscillation of
the transfer function). This ripple problem introduces an unwanted amplitude variation. In
order to reduce this, the order is increased as much as possible and some prototype
functions like Butterworth or Cauer are chosen for the design. Attenuation in the rejection
band is worst in these cases. Additionally, the group delay of the filter is mandatory for
non-distortion of the information. For NB-CW, it is usually not a problem but the phase
response has to be linear for WB-CW and IR-UWB systems. In the case of a multi-band
system, being linear in each sub-band is sufficient. These remarks point out that filters for
WB-CW are more difficult to design than filters for NB-CW due to the performance
limitations over the bandwidth.
To conclude, the RF filters imperfections are modeled by their noise factors (noise added,
distortions and non constant group delays), insertion losses (due to the ripples and
mismatches) and selectivity (finite attenuation of unwanted frequencies).
- Power Amplifiers (PA) in RF architectures are designed to linearly amplify radio-
communication signals with the highest efficiency possible and with the lowest spectrum
re-growths or added noise (see spectrum mask/ACPR and EVM criteria to respect). Since
the active components of the amplifier are operating at maximum power, non-linear (NL)
compression/conversion and memory effects are unavoidable (Fig. 6). The PA design,
identified as “class of operation”, impacts the performances [Villegas et al., 2007][Diet et al.,
2004-2007]. Efficiency and linearity of the PA are mandatory for NB-CW and WB-CW
architecture because the signal is transmitting information continuously. Low efficiency
reduces battery lifetime and increases the dissipated power and the temperature of the
circuit. Low linearity affects the quality of the signal. For IR-UWB systems, only peak
performances in time are needed (see section 5). We are focusing upon the PA impact on the
architecture in the case of the CW system. The most difficult case is for WB-CW due to the
bandwidth, and the usual high PAPR of chosen modulation schemes (high data rate). A PA
class of operation is determined by the hypothesis of transistor saturation (current source or
switch). There are two families of PA classes: the switched mode (SW) and the continuous
wave (CW) or biased mode. The different load-lines are illustrated in Fig. 6 (right).
Imax I/V characteristic of the transistor

AM/AM
AM/AM
(compression)
Drain current
HPA Input AM
A
Cl
AB
ass
AM/PM
B--
C
Cl
AM/PM
as
Input AM (conversion)
s
(envelope)
SW Class
Input AM
Drain Voltage V
max
Fig. 6. AM/AM, AM/PM effects (without memory effects) and PA class load-lines
Due to the need of polarization, CW classes (namely A, B, AB and C) present lower efficiency
than SW classes (D, S, E and F). SW classes need a switching of the transistor and cannot
reproduce an amplitude modulation for that reason. Their linearity is also worse than in CW
classes because of this switching operation. Moreover, it is possible to restore the amplitude
information by adjusting the voltage supply. This is theoretically linear in the case of SW
classes and there is a tracking effect for CW classes. For AM signals the average efficiency will
rely on the statistical properties of the signal itself. Additionally, it is important to consider
that efficiency is given as a peak value for CW classes. An improvement in efficiency is gained
if saturation/clipping on the peak values is introduced in order to increase the average power
of the output signal for the same power dissipated by the amplifier. WB-CW signals for high
data rate applications present such a high PAPR that the amplification by a CW class PA
requires a power back off (very low efficiency) or a linearization technique to reduce the NL
effects of compression (AM/AM) and conversion (AM/PM), see Fig. 6. Techniques which are
interesting for the designer in the case of wideband and high PAPR signal are those providing
the highest efficiency of the entire architecture including the PA (see section 3). Polynomial
modeling of the AM/AM can be done by the PA response to 1 or 2 frequency signals, called 1
or 2 tons. These indicators are the 1 dB compression point (P1dB) and the 3rd order
Interception Point (IP3), as defined in Fig. 7.
Pout (dBm)
Pout1dB IP3out IP3

Pout (dBm)
1 dB
Fundamental Fundamental
response response
IM3
1 response
1
Pin1dB
Pin (dBm) 3 IP3in

Pout = Pin + G
Pout1dB = Pin1dB + G – 1 Pin (dBm)
HPA
+ EVM
Fig. 7. Compression effect modeling for 1 ton (P1dB) and 2 tons (IP3)
These are considered to illustrate the main PA imperfections. AM/AM and AM/PM
measurements, when it is possible to take them, can be the best way of characterizing the
effects of the PA on the architecture: EVM and noise factor increase and spectral re-growths
(ACPR, spectrum mask). In CW systems, the PA has to amplify modulated signals (a sum of
several tons). For NB-CW signals, the PA behavior is well-characterized by the P1dB and
the input power is usually set at this value for the linearity/efficiency trade-off (if the PAPR
is not too high). In the case of WB-CW, the IP3 is a good representation of NL effects on the
spectrum if the frequency separation of the two tons is coherent with the modulated
bandwidth (symbol rate frequency). Moreover, the conversion effects are, by far, more
complex to analyze. It is almost impossible to determine an equivalent of the P1dB or the
IP3 for AM/PM. This relies upon the influence of transistor technology. For wideband and
high PAPR signals, the conversion effect increases the EVM significantly and often destroys
the information.
SW classes are based on the hypothesis that the transistor switches perfectly (no power
dissipated). The filtering is mandatory in RF applications, except if the SW PA is dedicated
to the amplification of a square signal (class D). The load-line of every SW PA class tends to
be that of an ideal switch, which is impossible in practice due to the physical realities of the
transistor (resistive and capacitive output effects). Although the switching cannot be perfect,
SW classes present higher efficiency than CW classes, see the 100% efficiency class E PA
[Sokal & Sokal, 1975][Raab et al., 2003][Diet et al., 2008]. Moreover, they are used for AM
signals only with a recombination process: supply modulation or amplification of an
envelope coded signal (PWM, ΣΔ) [Robert et al., 2009][Suarez et al., 2008]. Using supply
modulation creates NL effects of compression and conversion on the output signal. These
effects are named Vdd/AM and Vdd/PM [Diet et al., 2004]. For new RF architectures, high
efficiency (SW) classes are preferred due to their high efficiency, but the challenge then
shifts to the linearity of the amplified signals. Whatever the CW system is, NL effects of the
PA are unavoidable and are sources of important imperfections (i.e., loss of linearity). In the
case of high PAPR signals, a correction or a modification of the architecture is needed, as is
presented in the following section.
- DDR (f,θ,φ) - DDR (f,θ,φ)

- Φ (f,θ,φ) - Φ (f,θ,φ)
- Zin(f) - heff (f)
(f2,θ,φ)
PA LNA
- noise (f)
- Gain (f) (f3,θ,φ) - channel,…
- Gain (f) et NF (f)
- Zout(f)
- Zin(f)
- AM-AM (f,Pin)
(f1,θ,φ) - AM-PM (f)
- AM-PM (f,Pin)
Bandwidth allowed Bandwidth allowed
Amplitude Amplitude
response response
wideband
Multi-(narrow)band
frequency frequency antenna
antenna
Phase Phase
response response
frequency frequency
Fig. 8. Antenna system characteristics in radio-communications
- Antenna characteristics from a “system” point of view are reciprocal interfaces between
electrical and radiated signals. Depending on the application, antennas have to be adapted
to their environment (omni-directivity of their radiation pattern, polarization…). Fig. 8
summarizes the different system parameters of an antenna that can influence the
performance of the radio-communications link: spatial and frequency variations of the

radiation pattern, bandwidth limitations and phase distortion. Some of the imperfections
are equivalent to those of the filter, but they depend on the signal direction and the channel.
For wideband systems, i.e. WB-CW and IR-UWB, the use of a UWB antenna is mandatory
for keeping the phase information unaltered [Schantz, 2005]. For NB-CW and separated
multi-band signals using a wide bandwidth (see Fig. 8), the use of a multi-band antenna is
sufficient. The latter can present phase linearity only for used sub-bands [Diet et al., 2006].
To conclude, the communications channel composed of the emitting and receiving antennas
and the propagation channel is the source of amplitude and phase distortion, with a
statistical dependence on time and space. Noise is added and the techniques of signal
protection and antenna diversity (e.g., MIMO) are exploited as much as possible to improve
the system’s range.
3. Architecture basics
RF architectures are adapted to radio-communications signals and provide emission and
reception. Linearity improvements are needed when power amplification NLs can destroy
the information. This section first focuses on the transmitter section where different major
modifications are described. We especially focus on CW systems because of their important
challenges about efficiency and linearity.
3.1 Classic architectures for RF transmitters

Radio-communications architecture is composed of the above-mentioned blocks in section 2
and provides the emission and reception of the signal. There exist three kinds of basic
architectures, the classic ones, designed for NB-CW systems (the oldest application of which
is radio broadcasting): homodyne, heterodyne and PLL modulated.
- The homodyne architecture means that the I and Q signals (the information after DACs) is
transposed directly from baseband to RF see Fig. 9. If the modulation scheme is not x-QAM,
there is only one frequency transposition in the transmitter. This type of architecture is the
simplest combination of function blocks and theoretically requires the minimum number of
components.
If the signal is transposed directly to RF, the frequency synthesizer output is at the same
frequency value, and so is the HPA and the antenna. In a compact system (mobile phones
for example), the spatial proximity is the cause of unavoidable coupling between the
synthesizer, the HPA and the antenna. This problem of electromagnetic coupling (EMC, EM
Compatibility) is that it highly distorts the signal (EVM increase). EMC effects can be
reduced by circuit spatial optimization, when possible, and more efficiently with a shielding
of the considered block (LO, HPA). Additionally, a ground plane acting as a reflector can be
added between the antenna and these elements (HPA and/or LO synthesizer) to reduce the
amount of radiated waves toward the circuit. Choosing the homodyne structure results
from multiple trade-offs concerning the EVM, the simplicity and the size of the system.
- The heterodyne architecture, as represented in Fig. 9, means that the frequency
transposition is achieved in two or more steps. In fact, heterodyne means that the frequency
synthesizers are not at RF values. Any combination is possible, but for a minimum use of
components only two transpositions are usually performed. As the frequencies are
different, the coupling effect is highly reduced in this architecture. In heterodyne structures,
there are more components than in homodyne ones (with at least one filter and one
additional mixer) and imperfection sources are added. For example, the phase noise is a
function of the number of synthesizers. An additional mixer also means the possibility of an
image frequency that can distort the emitted information if the intermediate frequency (IF)
filter is not selective enough. This IF filter is traditionally an external SAW one (because of
its selectivity). This increases the cost and complexity of building this system. An advantage
of the heterodyne architecture is that the need of frequency flexibility for channel selection
can be more easily achieved with two synthesizers than with one (homodyne case).
LO-Antenna coupling
CNA
RF RF Pros: Simple
RF +
Homodyne PLL
0 HPA Cons: Coupling
90°
-
CNA LO-HPA
coupling
LO2
CNA
FI RF
Pros: No Coupling
LO1 + Cons: Phase Noise
Heterodyne 0 HPA
PLL
90° - Nbr components
CNA
PLL
CNA
F(p) VCO RF
LO1 RF
Modulated +
PLL
0 + HPA
90°
PLL -
CNA
1/N Pros: Phase Noise/Stability
Cons: No AM info.
Bandwidth limited
Fig. 9. Classic architectures for RF transmitters
- Architectures using modulated PLL can benefit from PLL advantages. This corresponds to
the direct modulation of a synthesizer (with N=1). The signal of a PLL is stable and its noise
profile depends on the loop filter bandwidth. It is possible to modulate the PLL by
introducing small variations of the reference frequency which will be stabilised by the loop
reaction. This is called a modulation “in the loop” and is used for narrow-bandwidth
modulations (low symbol rate). On the contrary, the input voltage of the VCO can be
directly modulated to produce a wider bandwidth modulation (modulation “over the
loop”) but this could affect the stability of the PLL. As is understood here, only angular
modulations are possible with modulated PLL architectures. If the signal to transmit has
AM information, this latter should be reintroduced after the PLL. To conclude, architectures
using modulated PLL are very interesting with regards to their noise property (less noise
than an IQ modulator) and their design for non-constant envelope signals implies some
important modifications of the architecture.
The classic architectures presented are widely used for NB-CW signals. While the efficiency
and linearity of the transmitter is not significantly affected, these well-known structures are
preferred for their simplicity. In the case of WB-CW systems, the bandwidth and the
probable increase of the PAPR (due to high data rate) lower the performance of such
architectures. For power amplification in particular, high PAPR values of OFDM and other
multi-carrier signals cause such compression and distortion/conversion effects that the
information cannot be interpreted at reception. Additionally, standard limitations are, by
far, not respected. The first choice is to perform a PA back-off, but this drives it to
unacceptably low values for the architecture efficiency. To achieve a linear transmitter,
linearization techniques are provided. The next sub-section is dedicated to their
descriptions.
3.2 Linearization techniques for the transmitter

Wireless communications require highly efficient and compact transceivers, whatever the
signal characteristics are. Transmitter architecture, at worst, must meet design constraints
of: providing high efficiency and linearity for a wideband and high PAPR signal (or high
dynamic). Power amplification of WB-CW multi-carrier signals (WiFI, WiMAX, LTE…)
introduces crippling Non-Linearities (NLs) in amplitude and in phase in the communication
system. The linearization of such this kind of transmitter is mandatory. Identifying a type of
architecture for such signals requires a careful study of linearization techniques and their
performance. A linearization technique is beneficial only if it provides linearity with the
maximum efficiency possible. There are several linearization techniques, depending on the
PAPR of the signal, the added complexity and the increase in size and consumption of the
system that can be accepted by the RF designers [Villegas et al., 2007]. Many criteria
characterize the technique used: static/dynamic processing, adaptability, frequency (digital,
baseband, IF or RF), correction of memory effects, complexity, stability, resulting efficiency,
size increase… Herein, we basically classify these techniques in three types: (i) correction
techniques, (ii) anticipation of NLs and (iii) those based on a decomposition and
recombination of the signal, often dedicated to wideband signals.
Examples of correction techniques are (A) feed-back, (B) feed-forward and the anticipated
technique of (C) pre-distortion, illustrated in Fig. 10. Their point in common is to modify the
modulated signal as close as possible to the PA (before or after). The architecture
considerations here, do not include the modulator nor the baseband signal processing. To
linearize, we need a carefully selected NL model of the PA (Volterra series, Wiener or Saleh
model…). Adaptability to the signal amplitude can be introduced in order to compensate
for the model’s lack of accuracy and the PA memory effects (a temperature influence can be
considered) [Baudoin et al., 2007]. Each structure contains a major defect. (A) Feed-back
reduces the gain of the amplification and introduces a bandwidth limitation due to the
transfer function of the loop (stability and dynamic response). The feed-back can be
performed on the amplitude (Polar feed-back) or on I and Q quadrature components of the
signal (Cartesian feed-back) and both are dedicated to narrowband signals. (B) Feed-
forward requires a significant increase in signal processing and RF blocks in the transmitter,
with the hypothesis of a precise matching between NLs and reconstructed transfer
functions. The improvement in linearity will be costly in terms of consumption and size
(integration criterion). The advantages are stability and the possibility to process wideband
signals. The most interesting of the three techniques is (C) pre-distortion because of its
flexibility: the anticipation can be done in the digital part and, by doing so, can provide
adaptability of the technique if using a feed-back loop (with an additional DAC). The digital
pre-distortion represents an additional and non-negligible consumption of a Digital Signal
Processor (DSP) and often requires a look-up table [Jardin & Baudoin, 2007]. The signal is
widened in frequency because of the non-linear law of the pre-distorter (as for NL effects of
compression on the spectrum), requiring baseband and RF parts to be wideband designed.
Interesting improvements of pre-distortion have been made with OFDM WB-CW signals in
[Baudoin et al., 2007] [Jardin & Baudoin, 2007].
A) B)
Vin + Vout
A Vin + Vout
- A0
-
Vout  A
K1
Vin 1AK
-
+
C) A2
NL
K2
Vin Vout A2 : modelisation of A0

P A K1 , K2 : linear gains
P=K/A
Vout
 A 0  K 2 A 2  K 1 
Vout  P.A  K Vin
Vin
Fig. 10. Correction (A and B) and anticipation (C) linearization techniques
Other techniques presented are based on a vectorial decomposition of the signal. The goal is
to drive high efficiency switched mode RF PAs with constant envelope (constant power)
signals, avoiding AM/AM and AM/PM [Raab et al., 2003] [Diet et al., 2003-2004 ]. These
techniques are used when NL effects are so great that feed-back or pre-distortion cannot
sufficiently improve the linearity. We can consider the problem of linearization in the
communication chain from the digital part to the antenna (front end). This drives one to
completely modify the architecture and its elements’ specifications in baseband, IF/RF and
power RF. After the amplification of constant envelope parts of the signals, the challenge is
to reintroduce the variable envelope information with lower NLs than in a direct
amplification case, keeping high efficiency of the architecture. Basic examples of these
techniques are the LINC (LInearization with Non-linear Components) and the EER
(Envelope Elimination and Restoration) methods (and their recent evolutions) [Cox, 1974]
[Kahn, 1952] [Baudoin et al., 2003-2007] [Diet et al., 2004] [Suarez et al., 2008].
r0 .cosΦt   θt 
+
r0 HPA
R(t) -
r0 .sin Φ t   θt 
+θ(t)
Φ(t)
-θ(t) PLL
0
90°
r0 .cosΦt   θt 
+
I  j.Q  R t .e jΦ  t  HPA
-

 r0 . e jΦ t θ t   e jΦ  t θ  t   r0 .sin Φt   θt 
Fig. 11. LINC decomposition and recombination at RF power amplification
The LINC principle relies on a decomposition of the modulated signal into two constant
envelope signals as is shown in Fig. 11. The decomposition can be computed by a Digital
Signal Processor (DSP) or by combining two VCOs in quadrature PLL configuration:
CALLUM. This latter configuration is an interesting architecture but presents the possibility
of instability and additional manufacturing costs. The amplification of the two constant
envelope signals implies the design of two identical HPAs at RF frequency, and this often
causes signal distortion due to imbalance mismatches. Also the HPA must be wideband
because the signal decomposition is a non-linear process (widening of the spectrum), and
the phase modulation index is increased. Whatever the decomposition technique is
(LINC/CALLUM), the problem is that the efficiency is directly determined by the
recombination step: a sum of the powers. It is very difficult to avoid losses at RF while
designing a RF power combiner.
I t   jQ t   Rt  e jΦ t  Rt   I²Q² R(t) e j Φ(t)
Envelope
detector
e j Φ(t)
R=1
supply Φ(t)
Limiter R(t)
Phase signal
High efficiency PA
Fig. 12. Principle of the EER technique [Kahn, 1952]
Another decomposition technique was proposed by Kahn in 1952 and this is basically an
amplitude and phase separation technique (polar): Envelope Elimination and Restoration
(EER). This method was first proposed for AM signals as represented in Fig. 12. The
advantage of EER is that it drives the RF PA with a constant envelope modulated signal
(carrying the phase information), enabling the use of a SW high efficiency amplifier [Raab et
al. 2003] [Sokal & Sokal, 1975] [Diet et al., 2005-2008]. The difficulty is to reintroduce the
amplitude information linearly using the variations of the PA voltage supply. This implies a
power amplification of the envelope signal at a frequency equal to the symbol rate (lower
than RF). The recombination can be done with a SW class PA because the output voltage is
linearly dependant on the voltage supply for this PA mode. Two difficulties are to be
considered in such a linearization technique: (i) synchronization between the phase and the
amplitude information and (ii) linear and efficient amplification of the amplitude before the
recombination (directly impacting the overall efficiency), as reported in [Diet, 2003-2005].
Recently, a lot of work has been done on the EER based architectures, often named as
“polar” ones [Nielsen & Larsen, 2007] [Choi et al., 2007] [Suarez et al., 2008] [Diet et al.,
2008-2009]. The generation of the amplitude and phase components can be expected to be
done digitally thanks to the power of DSPs, as is shown in Fig. 13. As was previously
discussed in [Diet et al., 2003], the bandwidths of the envelope and phase signals are
widened due to NL processing and make it necessary to design the circuit for three to four
times the symbol rate (as for LINC or any other NL decomposition). Fortunately, a clipping
in frequency and on the envelope is possible, increasing the EVM and ACPR under
acceptable levels. These polar architectures are suited for new high data rate standards
where efficiency of the emitter and linearization are mandatory. Also, the multi-standard
and multi-radio concepts have helped polar architectures to evolve in multiple ways. For
example, recombination on the PA’s input signal is possible because the amplitude
information can modulate the phase signal (RF) and can be restored by the band-pass shape
function of the following blocks: PA + emission filter + antenna. The emitted spectrum is the
criterion of quality to be considered carefully, because the PWM or ΣΔ envelope coding are
the source of useless and crippling spectral re-growths. The efficiency is also penalized by
the power amplification of such frequency components, but this is counter-balanced by the
advantages of high flexibility of this architecture [Robert et al. 2009]. The digitally controlled
PA and the mixed-mode digital to RF converter are key parameters in the evolution of these
polar architectures [Suarez et al., 2008] [Robert et al., 2009] [Diet et al., 2008].
101010 PWM or
CNA ΣΔ coding
D Rt   I²Q²
E
C
O
M
P
O
S
I 001101
CNA
T I
cos Φ  +
I I²  Q² 0 Class E
PLL
O 90° - HPA
N 011001
CNA Q
sin Φ 
I²  Q²
Fig. 13. Recent improvements of EER/polar architectures for wideband OFDM signals
We have presented the main linearization techniques. Actual needs, in terms of transmitter
linearity and efficiency for high data rate applications, have caused the RF designer to
consider RF architectures based on combined techniques. For example, digital pre-distortion
is an improvement on polar based architectures, as is shown in section 4. Another example
is described in [Diet et al., 2008] where a combination of cascaded EER and LINC techniques
can theoretically provide an architecture which cancels the PAPR influence, see Fig. 14.
IQ MOD
Differential Class F PA
cos  t  + cos ωk t   t 
I t  Emitted
DIF
-
signal
EER sin  t  cos θ t  Rejection of ωk -i
Q t 
LINC cosωk t 
R t  sin θ t  Rt cosωkit  t
 Q t  
 t   Atan  
 I t  
 
IQ MOD – 2 outputs
R t   I 2 t   Q 2 t 
cos θ t   R t  * 
cos ω i t  θ t 
sin θ t   1 - R 2 t  * + +
 
cos ω i t
-
1 1 +
 I t  
* 2 2 
cos ω i t  θ t 
1 1 - Buffers added
Class D, F or E
 Q t   - Output 1 : « I.cos – Q.sin »
2 2 - Output 2 : « I.cos + Q.sin » ω i PA
Base-band part Frequency transposition part Power amplification part

Fig. 14. EER-LINC method for high PAPR signals
This architecture needs some new components (e.g., a balanced switched PA with frequency
transposition). This also represents an increase in complexity, but proposes new ways to
improve WB-CW signals transmitters. In EER-LINC, the envelope information is converted
to two angular modulated signals. The RF balanced PA (also called differential PA) is in SW
class and is supply-modulated with the two above-mentioned signals. The fact that the PA
is differentially supplied allows for the combining operation and a transposition at the same
time (due to the multiplication). As the antenna is supposed to be differential, there is no
balun after the PA.
This review of linearization techniques reveals that the different parts of an efficient and
linear transmitter cannot be designed separately: baseband, frequency transposition and
power amplification (and antenna for wideband systems). The modification of the
architecture for global performance improvements must be done, considering each block’s
impact, digital or analog (and their imperfections). To conclude this theoretical sub-section,
linearized architectures are mandatory for the major part of actual and future CW systems,
which are WB-CW (high data rate). The highest performance can be reached if a
combination of different techniques is exploited: pre-distortion and EER seem to be the
most popular.
3.3 Receiver architectures

The challenges of the receiver architectures are the noise level, the presence of other
channels and blocking/interference signals. This is summarized by the immunity and the
coexistence of the standards, illustrated in Fig. 15.
The information received is at such a low level that other RF emissions can mask it. It is
easily possible to saturate the receiver if a signal is too strong in the vicinity of the spectrum.
Other characteristics of the receiver are its sensitivity and selectivity. The sensitivity is the
lowest level of power that can be received and demodulated correctly (providing a Bit Error
Rate sufficient for interpretation). This latter, is altered by the total amount of noise added
by the receiver itself. The noise factor (F) expresses an equivalent of a white Gaussian noise
addition, for each block. It is possible to compute the global noise factor for the receiver
thanks to Friis formula, see Fig. 15. This formula shows that amplifying the signal as close
as possible to the antenna, at the front–end, is better for minimizing the total added noise.
This is the goal of the Low Noise Amplifier (LNA) whose design is optimized for noise and
not for gain or output power performance. Unfortunately, the presence of high power
signals (blocking) often implies starting the reception chain with a selective band-pass filter,
as is reported for the Rx architectures in Fig. 15. At this point, the signal bandwidth is
another limiting factor for the design of the receiver because it directly impacts the
reception filter design and the LNA. The selectivity and sensitivity performance are also
affected. At reception, the decrease in the signal quality is expressed by the EVM and the
BER. This degradation results from the noise in the system (noise factor) but other
imperfections are due to the architecture design: block interactions, CEM, image frequency
and so on. In digital radio-communications, the channel coding (block and convolution
types) improves the robustness of the system against the noise, but this point will not be
developed here.
P0
G1 GN CAN
N0
(F1 – 1) N0 (FN – 1) N0
Homodyne RF
0
"0-IF" LNA PLL
90°
RF
Friis Formula :
N CAN
 (F 
k 1
F  F1  k  1) j 1
Gj Pros: Simple, noise
k 2
Cons: Auto-coupling  DC offset
IM = 2LO1-RF LO1
CAN Rx
RF RF LO2
LNA 0
Heterodyne PLL
90°
LO2 IM
CAN RF
LO2
Pros: coupling, selectivity, channel selection
Cons: image, nbr components, sensitivity VS selectivity LO1
Fig. 15. Considerations for receiver architectures
The topic of this sub-section is to present the main types of receiver architectures. As there is
no PA NL effects (Tx architecture), there is no need for linearization techniques. If the
receiver is saturated, it suffices to reduce the gain of the LNA (if possible) or to attenuate the
signal (back-off). The basic receiver structures are similar to that of the classic transmitter
structures, that is to say homodyne and heterodyne types as represented in Fig. 15.
- Homodyne receivers are composed of a direct IQ demodulator which comes after the
LNA. RF filters, before and after the LNA, tend to avoid a masking effect by other signals
received. The filter technology for RF is subject to a trade-off between the selectivity and the
losses, which increase the noise factor of the filter. Swapping the LNA and the RF filter in
the receiver chain creates a limitation of the received band. This is set in function with
disturbing signals and the resulting noise factor (sensitivity). This architecture is simple, as
it requires few components, and theoretically limits the total amount of noise added. The
transposition needs a Local Oscillator (LO) signal that is stronger than the received signal.
The coupling effects between the Rx antenna and the frequency synthesizer is very
important and a shielding is mandatory to reduce CEM effects. Moreover, if a part of the LO
signal is sensed by the antenna, it creates an offset (DC) by auto-mixing of the RF signal.
This DC component may saturate the ADCs and is added to the baseband signal. The
sensitivity of the homodyne receiver is not as low as expected due to this coupling effect
whose impact is greater than that for Tx architecture.
To summarize, homodyne architecture is very attractive for its simplicity if the integration
of the synthesizer is not causing crippling CEM effects. The addition of RF filtering can
reduce the offset but reduces the receiver sensitivity (noise factor increased).
- Heterodyne receivers need two (or more) transpositions of the signal. The number of
components is important and represents an additional current consumption compared to
the homodyne structure (simpler). The coupling effect due to the strong LO1 and LO2
signals, see Fig. 15, is minimized due to the different frequency values. This architecture
needs more filters than the homodyne one but this can be useful for improving the receiver
selectivity: Intermediate Frequency (IF) filters can reach higher selectivity than in RF.
Moreover, additional filters inevitably decrease the signal-to-noise ratio, as well as the
receiver performance. Sensitivity and selectivity are the sources of the challenges faced by
the receiver designer. Additionally, the mixers’ operation frequency and the IF filters enable
different manufacturing technologies. A great challenge of the heterodyne structure is the
possibility of receiving unwanted image information. An image is a signal producing an IF
signal after the first mixer (IF = LO2 = LO1 - RF, as reported in Fig. 15). A signal whose
frequency is “IM = 2LO1 - RF” produces, after mixing, an output signal whose frequency is
“LO1 – RF = IF”. The receiver should reject this IM component in order to protect the
receiver from this perturbation. Without selectivity, the receiver cannot attenuate IM
components. If the addition of a filter after the antenna is not sufficient to do this, an
improvement is possible with “image-rejection” architectures. The principle here is to
eliminate the image signal by an addition of 180° phase shifted copy. The considered
improvements were proposed by Hartley and Weaver. This is possible if the total power
level of received signals (information at RF and image at IM) are not saturating the receiver.
An RF filter is mandatory even if these architectures are used.
Hartley and Weaver improvements are called “image-rejection” (heterodyne) receiver
architectures. The different structures are summarized in Fig. 16.
As is reported, these structures are supposed to cancel the received signal whose frequency
value is IM. The Hartley structure uses an all-pass filter with a phase shift of 90°. For NB-
CW systems, this filter can be realized by R-C structures. For WB-CW systems, a filter with
such frequency independent characteristics is called a Hilbert filter. The realization of this
filter is a source of imperfections which directly impact the image rejection property. In
order to avoid filter design difficulties, Weaver proposed a structure in which the 90° phase
shift is achieved by a second frequency transposition. This heterodyne structure can directly
output the baseband signal (information), as expressed in Fig. 16. The cost of this
improvement is the number of added components, which is almost twice as much when
compared to a classic heterodyne structure. If the signal is IQ modulated, it is possible to
create quadrature image rejection architecture by using two more mixers, as is illustrated by
the IQ Weaver schematic in Fig. 16. The increase in size, complexity and consumption is
balanced by the image rejection property, if IQ mismatches are low enough. Hartley,
Weaver and IQ Weaver were first designed for NB-CW systems. Their use for WB-CW can
be discussed if IQ modulator performance and filters enables it. The bandwidth of the filter
implies the choice of LO values. Some receivers, whose required bandwidths are too wide,
are not possible to design. “Image-rejection” architectures are an interesting alternative to
reduce selectivity constraints in the receiver design, at the cost of additional components
and additional noise (lower sensitivity).
(1) (2)
1  A cos ω RF t     I m cos ω IM t  ψ 
RF + (4) 2   A cos ω IF t     I m cos  ω IF t  ψ 
2 2
0
PLL
-90°
LO (3)' - 3  A 2 sin ω IF t     I m 2 sin  ω IF t  ψ 
(3)
+90° 3'   A 2 cos ω IF t     I m 2 cos ω IF t  ψ 
4   A cos ω IF t   
(1) (2) 5  A cos    I m cos ψ   A cos    I m cos ψ 
4 4 4 4
6   A 4 sin    I m 4 sin ψ   A 4 sin    I m 4 sin ψ 
RF + (5)
PLL
0
PLL
0 wi th
-90° -90°
LO IF - ω IM  ω 2 LO  RF
(3) ω IF  ω RF  LO  ω LO  IM
(2)
(1)
Pros: image rejection, coupling,

RF + (5) - (6) selectivity relaxed
0 0
PLL PLL
LO
-90° IF -90°
- - Cons: two times more components
(consumption, size,...)
IQ mismatch sensibility
(3)
Fig. 16. Hartley (top), Weaver (middle) and IQ Weaver (bottom) architectures.
Another interesting improvement of heterodyne architecture is a group of Low-

Intermediate Frequency (low-IF) receivers whose goal is to use ADCs at IF and not in
baseband. There are two main possibilities for signal processing at IF frequencies: (i) using a
poly-phase filter in a modified IQ Weaver architecture or (ii) using fast ADCs with digital
signal processing on the baseband information. These ADCs imply a large increase in
current consumption in function with the IF value. Nowadays, solution (ii) is more popular
due to the state-of-the-art ADC performance (resolution versus sampling rate) and the high
improvement possibilities provided by digital algorithms. Low-IF architectures with band-
pass ADCs represent a sub-group of heterodyne architectures and are popular for its
potential flexibility.
To conclude this section in the case of CW systems, there are two approaches for receiver
architectures: (i) homodyne (also called “zero-IF”) with limitations of LO coupling effects
and (ii) low-IF heterodyne structures, with a growing interest in IF ADCs.
4. High data rate state-of-the-art transmitters

4.1 Front-end considerations
Many new standards for high data rate communications have appeared recently or are
under development whether in the frequency range below 6 GHz or in the millimeter wave
range (60 GHz radio in particular). In the first case, the data rates are in the range of several
tens to several hundreds of Mbps and in the second case they can be in the range of several
Gbps. There are many challenges for the design of high data rate RF Transceivers. Among
the most critical are:
- Designing power transmitters with a good linearity and efficiency for wideband, high
PAPR signals with a large range of necessary transmit power control.
- Designing mobile transceivers using OFDM and multi-antenna MIMO and smart antenna
techniques in order to achieve very high performance (throughput and BER) in mobile
channels with large delay spreads while maintaining low power consumption and reduced
size and cost.
- Designing flexible and scalable devices able to accommodate for many standards,
frequency bands and modes
Multi-carrier (OFDM and OFDMA for multiple user access) and multiple-antenna (MIMO)
techniques have emerged as enabling technologies for beyond 3G and 4G high data rate
communication systems. Many new standards use OFDM and MIMO approaches.
Examples include: IEEE 802.11n for wireless local area networks, IEEE 802.16d and IEEE
802.16e fixed and mobile WIMAX or IEEE 802.20 for wireless metropolitan area networks,
IEEE 802.22 for wireless regional networks, 3GPP LTE (Long Term Evolution) for beyond
3G cellular networks. OFDM and MIMO have also generated new challenges in term of
transmitter architectures with good efficiency and linearity, and in term of integration of
MIMO transceiver and antennas in mobile user terminals. One of the main advantages of
OFDM in mobile wireless channels is the simplification of the channel equalization in
comparison with single carrier modulation. Indeed, for an OFDM modulation with N
carriers, the original high data rate M-QAM data stream with a rate R is split into N parallel
lower speed streams with a rate R/N that modulate one of the N carriers with a M-QAM
modulation. In practice, the N parallel single carrier modulation is achieved thanks to a
baseband IFFT. For a given data rate R, the bandwidth of the M-QAM/OFDM signal is
similar to that of a single carrier M-QAM signal. An OFDM symbol comprises N QAM
original symbols. A guard interval (GI) is introduced between successive OFDM symbols in
order to prevent inter-symbol interference. The duration of the GI is of the order of the
delay spread of the impulse response of the channel. Therefore, the larger the number of
carriers, the more efficient the system. Unfortunately, the PAPR value of the modulated
signal also depends on the number of carriers. The larger the number of carriers, the larger
the PAPR value, and the more difficult it is to design a power transmitter with a high
efficiency and a high linearity. In order to accommodate for multiple users, a set of carriers
can be allocated to each user resulting in the OFDMA principle. The LTE standard proposes
an uplink (mobile emitter) variant called single-carrier FDMA (SC-FDMA) which has a
smaller PAPR value. In a SC-FDMA emitter, the OFDM modulator is preceded by a DFT,
which increases the baseband complexity.
MIMO technology uses multiple antennas at the transmitter and the receiver. The obtained
diversity and spatial multiplexing allow for better BER performance and increased data rate
or link range without increasing the bandwidth or the transmitted power. In comparison
with a SISO system (Single antenna at the emitter and at the receiver) the maximal
achievable increase of the data rate depends on the minimum number of antennas in the
transmitter and the receiver. For example, for 2 Tx and 2 RX antennas, the data rate can be
multiplied by 2 at the maximum. There is a compromise between the diversity gain and the
multiplexing gain. In the WiMAX standard, different modes are defined depending on
whether one wants to increase the diversity (Space time bloc code STBC approach referred
to as MIMO matrix A) or the spatial multiplexing (referred to as MIMO matrix B). It is
possible to use MISO approach (multiple Tx and single Rx antennas) and use transmit
diversity with space time coding (called MIMO matrix A). WiMAX also includes possibility
for uplink collaborative MIMO technique. This is intended to allow for 2 separate user
devices, each with a single transmit antenna, to communicate on the same frequency with a
base station using 2 antennas. The MIMO approach can be associated with beam-forming to
control the direction and shape of the radiation pattern.
MIMO techniques allow for tremendous improvements in throughput performances. But it
is a challenge to integrate many antennas and transceivers in a small mobile user device.
With a multiple antenna base station and a single antenna user device it is possible to
achieve some improvement but a multiple antenna user device is necessary to really take
advantage of MIMO technique. In the “WiMAX Wave 2” device certification, 2 receive
antenna systems are mandatory. Another kind of challenge for high data rate front-end
transmitters is the case of millimeter wave (mmWave) mobile communications. At 60 GHz,
the available unlicensed bandwidth is very large and the mmWave technology is a good
candidate for indoor very high data rate (several Gbits/sec) communication systems. The
high data path loss allows for high frequency re-use. The small wave-length allows for the
use of very small antennas and integrating multiple antennas for MIMO or beam-forming
approaches can be easily achieved. The standard IEEE 802.15.3c is intended to provide Gbits
data rate at distance of the order of a few meters. For fixed equipments, the situation at
mmWave is quite similar to the preceding one (at frequency below 6 GHz) and OFDM
technique is widely used. But the design of mobile devices with low power consumption
constraint is still very challenging. The OFDM approach with high speed DAC/ADC and
baseband processors has generally crippling power consumption. Among the challenges
are: design of PAs with a sufficient output power, efficiency and linearity; power
consumption of DAC and ADC with Gigahertz sampling frequencies; design of a physical
layer with a low complexity and sufficient performance. Two approaches are commonly
proposed: single carrier QPSK and UWB techniques.
4.2 Polar architectures

One of the main challenges for OFDM based mobile transmitters is to design power
transmitters with a good linearity and efficiency for wideband, high PAPR signals with a
large range of transmit power control. For example, for the WiMAX mobile standard:
- The signal bandwidth is scalable up to 10 MHz.
- For a full OFDM WIMAX signal with a 1024 FFT and a 16-QAM mapping, the PAPR (or
effective power ratio with a probability of 10-3) is approximately 12 dB.
- The Peak transmit power is typically 23 dBm for subscriber terminals.
- The power control of the transmitted signal (TPC) that compensates for variations in signal
strength (e.g., distance variation) must be monotonic and able to cover a range of at least 45
dB by steps of 1 dB with a relative accuracy of 0.5 dB.
For common PAs operating in CW classes (A, AB, B or C), the power efficiency is better for
high power output values (close to the P1dB) than for low output values while the linearity
is usually better for small power values. In order to transmit a modulated signal with a
good linearity, the maximum instantaneous power of the signal must be kept smaller than
the P1dB. Therefore, for a given PAPR, the average input power is PAPR dB below P1dB.
The amplifier is operated with a back-off depending on the value of the PAPR and for large
PAPR values; the efficiency can be very small. The common efficiency obtained with class
AB PA for WiMAX signal is typically smaller than 20% while it would be at least twice
bigger for constant envelope signals such as GSM signals.
Another parameter to take into account is the average output power. For the same PAPR,
the power amplifier efficiency also depends on the average output power. The power
supply of the PA should be adjusted in order to take into account the desired average
output power and keep a constant efficiency on a large range of average output power.
Different kinds of polar architectures have been proposed as candidate solutions for high
PAPR management or for adjustment of power supply to the average power. The
terminology is not very stable. But, one can distinguish:
- Polar lite architectures.
- Polar feedback loop.
- Dynamic power supply and amplifier gain control with dynamic biasing, envelope or
power tracking for linear PA, drain (collector) modulation for non-linear PA.
- EER and Sampled-EER architectures.
Whatever the modulation, the complex envelope z of the modulated signals can be
expressed by its cartesian coordinates: real and imaginary parts usually called quadrature
or I and Q components or by its polar coordinates: amplitude  and phase , see (1). As the
amplitude  and phase  are obtained by very non-linear operation from I and Q , their
bandwidth is much higher than that of I and Q , as discussed in section 3.
Q (t )
z (t )  I (t )  jQ (t )   (t ) exp( j (t )),  ( t )  I ( t )2  Q ( t ) 2 , tan( ( t ))  . (1)
I (t )
When the amplitude (t) is constant, the polar decomposition is interesting since only one
signal ((t), or its time derivative) has to be digital to analog converted. GSM Transmitters
can take advantage of this characteristic. As seen in section 3.1, the modulation of the
frequency synthesis loop is a common GSM architecture that benefits of the good noise floor
of the VCO and suppresses the need for external filtering. Since the envelope of the
modulated signal is constant, the PA can operate in saturated CW or SW class (high
efficiency). An average power controller must be added. This architecture is also called
translational (or tracking offset) loop. The modulation of the PLL becomes very difficult
for large bandwidth signals such as WCDMA. Therefore this approach is mostly used for
GSM/EDGE signals. The name “polar transmitter” is sometimes limited to architectures in
which the phase/frequency modulation is applied directly to the RF carrier by a modified
PLL, using different techniques such as a “2-point modulation” [Durdodt 2001]. But we will
use the adjective “polar” whenever the signal is decomposed in polar coordinates
The translational loop technique was extended for EDGE-GSM signals using the so-called
polar lite architecture. The 8-PSK EDGE-GSM modulated signal is decomposed in polar
coordinates. The idea is to keep the benefits of the translational loop architecture but to
introduce an AM capability. In such architecture, the signal is decomposed in polar
coordinates. The phase modulates the loop and the amplitude multiplies the output of the
modulated loop by controlling the gain of a high dynamic range variable-gain amplifier
(VGA). Since the signal at the input of the PA is envelope-varying, this architecture does not
support a saturated PA. It uses a linear PA and has no particular efficiency advantage. It is
sometimes associated with some dynamic power supply technique in order to increase the
efficiency. An example is given in [Staszewski et al 2005].
The polar feedback loop architecture is a derivation of the polar lite architecture with the
advantage of using a saturated PA. In the polar feedback loop, the PA is fed with a constant
envelope signal and is modulated in amplitude. A feedback path takes a portion of the PA
output signal (with a coupler) and an error signal is calculated between the ideal and actual
output. The phase of the error drives the input of the PA and the error magnitude is used
for the amplitude modulation of the PA. The feedback loops provides some linearization to
the transmitter. But as in any feedback loop, the stability constraint limits the possible
bandwidth. An example of a polar feedback loop architecture for GSM and EDGE is given
in [Sowlati et al, 2004]. In that example, the PA efficiency is 54% in GSM mode at 33dBm
output power and 37% in EDGE mode at 27dBm output power.
Many techniques are possible for dynamic power supply: dynamic biasing (at gate or
drain), envelope or power tracking for linear PA, drain (collector) modulation for non-linear
PA. Dynamic power supply using DC-DC switching regulators are interesting in terms of
efficiency but they are still limited in term of signal bandwidth to a few tens of MHz. For
envelope tracking (ET), the DC-DC is fed with the amplitude of the modulated signal (Fig.
17). ET was not presented in section 3 because it is not strictly a decomposition technique of
linearization, but an improvement of the PA output power (and consequently the
efficiency). This optimisation of PA biasing can increase of more than 40% the efficiency of a
class A PA. In some standards the power-control dynamic range must be very wide (i.e., 80
dB for WCDMA). The amplitude modulation (envelope restoration) can be applied on a
VGA or directly on the PA which leads to a better efficiency.
Envelope DC/DC
Detector Converter
Fig. 17. Envelope tracking
The principle and interest of EER architectures and their variants under different names
such as polar architectures were explained in section 3.2. The critical points for this
architecture are: (i) the time mismatch between the envelope and the RF phase signals, (ii)
the NL of the envelope restoration, (iii) the distortion caused by difficulty of biasing the PA
when the amplitude of the envelope signal is very small, (iv) the leakage of the RF PA input
to the output, (v) the influence of the necessary limitation of the frequency bandwidth of the
envelope and phase signals. It can be shown [Baudoin et al., 2003] that the signal to noise
ratio due to a time mismatch  between these two paths is in a first approximation inversely
proportional to  and to the bandwidth of the envelope signal. Therefore, the wider the
signal bandwidth, the smaller should be the time mismatch. For an OFDM signal, such as in
WiFI, the time mismatch  must be kept smaller than typically 2 ns (which is a small
percentage of the QAM symbol duration) in order to fulfil the specifications of the standard.
The time mismatch can be corrected by adaptive techniques and the NL of the envelope
restoration can be compensated by adaptive pre-distortion techniques. But, as already stated
for dynamic power supply, it is difficult to design DC-DC converters with a bandwidth
superior to a few tens of MHz. This is all the more critical as the envelope signal bandwidth is
wider than the original modulated signal bandwidth. For example, an OFDM modulated WiFI
signal has a frequency bandwidth close to 20 MHz. But the bandwidths of its envelope and
phase signals cannot be filtered to a bandwidth smaller than respectively 40 MHz and 100
MHz if one wants to meet the specifications of the standard. The polar EER approach can be
illustrated by Tropian's Timestar(TM) RF IC supporting GSM/GPRS, EDGE and WCDMA
signals [Wendell et al., 2003] [McCune et al., 2005] or Sequoia Communications Inc. SEQ704
chip that supports HSDPA, GSM and EDGE and WCDMA [Groe et al., 2007-2008].
In polar architectures, the recombination of amplitude and phase signals can be done
whether by PA amplitude modulation (EER architecture) or before the PA input. We will
now consider the second approach, and we will focus on architectures using sampled
signals and switched RF amplifiers (typically class D, E, F and their variants) [Jeong et al.,
2007] [Hibon et al., 2005] [Berland et al., 2006] [Nielsen et al., 2007]. We will call these
architectures “polar sampled architectures”. The major motivations for using sampled
signals and switched amplifiers are the ease of integration of digital circuits and the very
high theoretical efficiency of switched PA. In polar sampled architectures, the envelope
signal is “sampled” (converted/coded) by a pulse width modulation (PWM) or a 1-bit
Sigma-Delta () modulator with bipolar output ±A [Murmann et al., 2007]. This two level
signal multiplies the RF phase modulated signal before the switched PA. The result is a
constant envelope signal. The switched RF PA is fed with this constant envelope signal that
controls the switching of the PA. The output of the PA must be filtered in order to suppress
the PWM or  noise and to recover the modulated signal. Fig. 18 illustrates this principle.
Fig. 18. Polar sampled architecture with envelope-phase recombination before the PA
In comparison with direct  architectures [Rode et al 2003] in which the modulated RF

signal is directly covered by a 1-bit  coder before the switched amplifier, the coding of the
envelope signal is interesting because it allows using smaller clock rates for the  coder.
One difference with EER architecture is the position and type of the filter used to eliminate
the noise of the  or PWM modulator. As illustrated in Fig. 13, EER architectures use a
low-pass filter at the output of the envelope BF amplifier to recover the envelope signal. But
when the recombination of envelope and phase signals is done before the RF PA, the noise
must be eliminated after the RF PA with a band-pass RF filter (Fig. 18). For a given
standard, this RF band-pass filter should be unique and correspond to the full uplink
bandwidth. It should not be specific to a given channel. Therefore the over-sampling ratio of
the  modulator must be calculated in reference to the full uplink bandwidth and not to
channel bandwidth (bandwidth of the modulated signal). In [Andia et al., 2008] the
possibility of using BW filters for WiMAX standard has been studied.
a)
(
(b)
c)
(
Fig. 19. (a) Polar Sampled Architecture, PSA, with 2 DACs. (b) PSA with a single DAC and
an analog mixer. (c) PSA with a single DAC and with digital mixing.
Different structures have been proposed for polar sampled architectures [Suarez et al.,
2008], see Fig. 19. In the structure (a) of this figure,  modulator output is digital, as well as
I(t) and Q(t) (cos( ) and sin( )). Therefore, two DACs are necessary before the IQ
modulator. DACs sampling frequency is chosen according to the  frequency. It has to be
high enough to avoid  noise overlapping. Targeted communication standards require
high Sigma-Delta frequencies and therefore significant sampling frequency for DACs. In the
structure (b), envelope and phase signals ((t) and  (t)) are calculated and processed
independently. The output of the low-pass sigma-delta modulator is analog. The digital
phase signal is converted by a DAC and then modulated to the carrier frequency (fc). Finally
constant envelope and phase signals are recombined.
Using CMOS 90 nm technology is a current solution to address this challenge, while

reducing the number of analog blocks. In this sub part we present two digital architectures,
representing the digitalization trend.
The first architecture presented is based on a classic direct conversion architecture using
only one frequency transposition. The “Direct Digital to RF Modulator” (DDRM)
architecture [Eloranta et al] was developed as a basis toward further architecture
digitization. In this architecture the system is digitalized the closest possible to the
amplifier, which is still an analog part of the transmitter, see Fig. 21.
Digitalword decoder
N DRFC RF
LO
...
PA
0
Digital part PLL
-90°
j.LO
LO LO
N DRFC I 16.I
...
DRFC
Fig. 21. DDRM architecture (left) and DRFC (right) [Eloranta et al]
The advantage of being digital is that it limits the imperfections due to variations in the
process. In analog architectures, the multiple filtering blocks see their basic characteristics
varying, and so the architecture performance varies as well. This leads to the use of a
calibration loop. Digitization provides size optimization and good stability of the circuit. As
the first stage is an over-sampling stage, there is no baseband signal and no need to filter
with high selectivity before the DACs. In this architecture, mixing and D/A conversion are
performed by a single block: “Digital to RF converter” (DRFC), a kind of RF-DAC.
Weighted Gilbert cells are parallelized. No baseband signal in the architecture implies a
very low LO leakage (DC offset). The principle is that the data is over-sampled at the carrier
frequency by switches. The signal amplitude is coded into a digital word (MSB to LSB). At
the output of Gilbert cells, the current is proportional to the code word. The linearity
performance and the signal resolution increase with the number of parallel cells. Due to this
parallelization, IQ imbalance is limited and can only result from the average cell
imbalances. Power control can be achieved by a bias current variation, which reduces the
output current from each unit cell. As it uses no filter, the choice of the converter frequency
is paramount. Indeed, the only filtering applied (SINC) is the zero-order hold. The
frequency must be chosen so that baseband harmonics are cancelled thanks to the zeros of
the filter response.
Thanks to CMOS evolutions, this architecture was optimized to address spectrum
cohabitation issues [Pozsgay et al] under the name of “Sigma Delta – RFDAC” architecture,
illustrated in Fig. 22. In this architecture, the gain control can take place throughout the
transmitter. The first power control appears after the first up-sampling filter. It is then a low
speed dynamic control, resulting from multiplying IQ signals with a binary word. This
word size depends on the control resolution. A second power control stage is achieved, by
deleting successive LSBs of the signal (6 dB steps). Several Delta-Sigma modulators in
MASH structures are then used and I and Q channels are duplicated (I and I’, Q and Q’). I
Using CMOS 90 nm technology is a current solution to address this challenge, while

reducing the number of analog blocks. In this sub part we present two digital architectures,
representing the digitalization trend.
The first architecture presented is based on a classic direct conversion architecture using
only one frequency transposition. The “Direct Digital to RF Modulator” (DDRM)
architecture [Eloranta et al] was developed as a basis toward further architecture
digitization. In this architecture the system is digitalized the closest possible to the
amplifier, which is still an analog part of the transmitter, see Fig. 21.
Digitalword decoder
N DRFC RF
LO
...
PA
0
Digital part PLL
-90°
j.LO
LO LO
N DRFC I 16.I
...
DRFC
Fig. 21. DDRM architecture (left) and DRFC (right) [Eloranta et al]
The advantage of being digital is that it limits the imperfections due to variations in the
process. In analog architectures, the multiple filtering blocks see their basic characteristics
varying, and so the architecture performance varies as well. This leads to the use of a
calibration loop. Digitization provides size optimization and good stability of the circuit. As
the first stage is an over-sampling stage, there is no baseband signal and no need to filter
with high selectivity before the DACs. In this architecture, mixing and D/A conversion are
performed by a single block: “Digital to RF converter” (DRFC), a kind of RF-DAC.
Weighted Gilbert cells are parallelized. No baseband signal in the architecture implies a
very low LO leakage (DC offset). The principle is that the data is over-sampled at the carrier
frequency by switches. The signal amplitude is coded into a digital word (MSB to LSB). At
the output of Gilbert cells, the current is proportional to the code word. The linearity
performance and the signal resolution increase with the number of parallel cells. Due to this
parallelization, IQ imbalance is limited and can only result from the average cell
imbalances. Power control can be achieved by a bias current variation, which reduces the
output current from each unit cell. As it uses no filter, the choice of the converter frequency
is paramount. Indeed, the only filtering applied (SINC) is the zero-order hold. The
frequency must be chosen so that baseband harmonics are cancelled thanks to the zeros of
the filter response.
Thanks to CMOS evolutions, this architecture was optimized to address spectrum
cohabitation issues [Pozsgay et al] under the name of “Sigma Delta – RFDAC” architecture,
illustrated in Fig. 22. In this architecture, the gain control can take place throughout the
transmitter. The first power control appears after the first up-sampling filter. It is then a low
speed dynamic control, resulting from multiplying IQ signals with a binary word. This
word size depends on the control resolution. A second power control stage is achieved, by
deleting successive LSBs of the signal (6 dB steps). Several Delta-Sigma modulators in
MASH structures are then used and I and Q channels are duplicated (I and I’, Q and Q’). I
DQG 4 VLJQDOV JR WKURXJK D PRGXODWRU JQDOV

DQG VL, DQG 4 DUH ILUVW GHOD\HG EHIRUH D
VHFRQGPRGXODWRU7KLVZLOOUHVXOWLQWKHUHFR
PELQDWLRQRIPL[HGVLJQDOVDURXQGWKHFDUULHU
DQGWKHFUHDWLRQRIQRWFKHVLQGHVLUHGIUHTXHQF\EDQGVFHUWDLQO\XVHGE\RWKHUVW
$Q DGGLWLRQDO G% SRZHU FRQWURO FDQ EH GRQH E\ E\SDVVLQJ ,• DQG 4• DQG WKH
FRUUHVSRQGLQJ5)’$&ORVLQJWKHDGYDQWDJHRIQRWFKHVDVDUHVXOW
%DVHEDQG 1),5 ·ƒ 5)
,DQG4
/2FRPSHQVDWLRQ 5)
JDLQ,,5&,& PRGXODWRU ’$&
’LJLWDO 5)
GHOD\ ’$&
$’3//
)LJ6LJPD’HOWD†5)’$& DUFKLWHFWXUH>3R]VJD\HWDO@
7KH VHFRQG DUFKLWHFWXUH SUHVHQWHGG RQ LV

DQEDVH
((5OLNH DUFKLWHFWXUH $V ZLOO EH
H[SODLQHGWKHUHLVQRFRPSOHWHVHSDUDWLRQ
EHWZHHQSKDVHDQGDPSOLWXGHLQWKHWUDQVPLWWHG
VLJQDOSURFHVVLQJ7KHDUFKLWHFWXUHZDVGHYHORSH
GE\>3DULNKHWDO@V HH)LJ ,4VLJQDOV
DUH RYHUVDPSOHG DQG ILOWHUHG LQ RUGHU
DFH WKH
WR SO
VSHFWUDO UHJURZWKV IDU HQRXJK ]HUR
RUGHUKROGDQGVHWWLQJQRWF
KHVDWPXOWLSOHVRIWKHRYHUVDPSOLQJIUHTXHQF\
$%6,4
%DVHEDQG 8SVDPSOLQJ
,DQG4 DQGILOWHULQJ
3URF ’3$ [ W , W FRV&W 4 W VLQ
&W
$%6,4
08;
§ Q • § Q • § Q • § Q •
[ W Q7V ,¤ ‚ FRV
¤ ‚ 4¤ ‚ VLQ
¤ ‚
5) ' „ ' „ ' „ ' „
6LJQ, 6LJQ4
§ § • § ••
FRQWURO >Q @ [ ¤, ¤ ‚ 4 ¤ ‚ ‚
' ' „ ' „„
§ § • § •• § § • § ••
>Q @ [ ¤, ¤ ‚ 4¤ ‚‚ ¤, ¤ ‚ 4 ¤ ‚ ‚
5) ' ' „ ' „„ ' ' „ ' „„
5)
3URF 08; § § • § •• § § • § ••
3// >Q @ [ ¤, ¤ ‚ 4¤ ‚‚ ¤, ¤ ‚ 4 ¤ ‚ ‚
' ' „ ' „„ ' ' „ ' „„
§ § • § •• § § • § ••
>Q @ [ ¤, ¤ ‚ 4¤ ‚‚ ¤, ¤ ‚ 4 ¤ ‚ ‚
' ' „ ' „„ ' ' „ ' „„
)LJ’LJLWDO4XDGUDWXUH0RGXODWRU>3DULNKHWDO@
7ZRVLJQDOVDUHFUHDWHG·DEV,4DQG·DEV ,40RUHRYHUZHGHWHUPLQHSKDVHFRQWURO
VLJQDOV ELWV ZLWK WKH VLJQV RI , DQG
FRPSOH[
4 TXDGUDQW $PSOLWXGH VLJQDOV DUH
DOWHUQDWHO\ VHQW WR D GLJLWDO GLIIHUHQWLDO DPSOLILHU ’’3$ DW D UDWH RI WZLFH W
IUHTXHQF\7KXVZHJHWDKLJKHUUHVROXWLRQRI
WKHYDULDWLRQRIWKHRULJLQDOVLJQDOHQYHORS
7KLV FDQ EH VHHQ DV DQ RYHU VDPSOLQJ E\ IRXU RI WKH ,4 V\PEROV OHDGLQJ WR D EHWWHU
HVWLPDWLRQRIWKHDPSOLWXGHV
QDQRVHFRQGV7KHVLJQDOLVWUDQVPLWWHGDWWKHEDVHEDQGIUHTXHQF\,WLVLQWHUHVW
SUHVHQFHRIPXOWLSDWKVFKDQQHO%\XVLQJDSX OVHGPRGHLWDOORZVWKHPHDVXUHPHQWRIWKH
SURSDJDWLRQ GHOD\ WLPH EHWZHHQ WKH WUDQVPLWWH
U DQG WKH UHFHLYHU DQG FRQVHTXHQWO\ WKH
ORFDWLRQ,QIRUPDWLRQVDUHHQFRGHGLQWKHVKDS HRIWKHLPSXOVHDPSOLWXGHRUSKDVHE\WKH
SRVLWLRQLQWLPHRIWKHSXOVHRUE\UDQGRPVHTXHQFHVRISXOVHV
3RVVLELOLWLHV IRU PRGHV WUDQVPLVVLRQ RI WKH 8:% SXOVHV DUH FRKHUHQW GLIIHUH
FRKHUHQWDQGQRQFRKHUHQW,QWKHFDVHRIFR KHUHQWWUDQVPLVVLRQWK
HPRGXODWLRQVXVHGDUH
36. 3$0 DQG 330 7KH UHFHLYHU ZLOO EH RI D FRUUHODWRU W\SH HJ 5DNH ,Q WKH FDVH RI
GLIIHUHQWLDOO\ FRKHUHQW WUDQVPLVVLRQ WKH PRGXODWLRQ XVHG LV ’36. 0RUH
V\QFKURQL]DWLRQ EHFRPHV XQQHFHVVDU\ )RU ODVWWKH
FDVH RI QRQ FRKHUHQW WUDQVPLVVLRQ
PRGXODWLRQV XVHG DUH 22. RU 330 7KH UHFHLYHUZLOO EH EDVHG RQ HQHUJ\ GHWHFWLRQ 7KH
WUDQVFHLYHUDUFKLWHFWXUHVIRU,58:%DUHVLPS OHUWKDQWKRVHIRUWKH0%2)’0RUDQ\RWKHU
:%&:7KHIXQFWLRQDOEORFNVRIWKH DUFKLWHFWXUHDUHJLYHQLQ)LJ
)LJ)XQFWLRQDOEORFNVRI,58:%WUDQVFHLYHUVDQGUHFHLYHUZLQGRZLQJ
’HPRGXODWLRQ VKRXOG EH PDGH RQ ZLQGRZV VORWV FRQWDLQLQJ D SXOVH DV LOOXVWUDWH

$UFKLWHFWXUHVIRULPSXOVHUDGLRDUHGLYLGHGLQWRWZRW\SHV&RKHUHQW5HFHSWLRQ&
UHTXLULQJ NQRZOHGJH RI WKH SKDVH RI WKH UHFHLYHG VLJQDO DQG 1RQ&RKHUHQW 5HFHS
1&R5[EDVHGRQGHWHFWLRQRIHQHUJ\ZKHQSXOV HVDUHWUDQVPLWWHG,QWKHFDVHRI&R5[
WKH UHFHLYHU XVHV WKH UHFRJQLWLRQ RI WKH SKDVH VLJQDO ,W LV JHQHUDOO\ EDVHG RQ
FRUUHODWLRQ VFKHPH ZKLFK SURGXFHVWKH DOPRVW
RSWLPDO SHUIRUPDQFH 7KHUHIRUH WKH
FRPSOH[LW\ OHDGV WR D OLPLWDWLRQ RQ WKH LPSOHPHQWDWLRQ LQ ORZ FRQVXPG@V@ G LP
5.1 UWB communications, goals and aspects

There are systems that transmit and receive waves whose relative bandwidth (BW) is
greater than or equal to 0.25, see the definition in Fig. 25. The first definition has been
amended and replaced by a new one proposed by the FCC. Under this new definition, a
UWB signal is a signal whose “-10dB bandwidth” exceed 500 MHz and 20% of center
frequency. The main UWB frequency band is between 3.1 and 10.6 GHz. This bandwidth of
about 7 GHz could be divided into 14 sub-bands of 500 MHz. A system using the full
bandwidth or a set of sub-bands will be considered a UWB system. Today, we can classify
UWB into two main categories of applications: UWB Low Data Rate (UWB LDR) and UWB
High Data Rate (UWB HDR). They are attached to the IEEE 802.15.4a and 802.15.3a.
Low data rate systems are generally characterized by data rates lower than to 2 Mbps, by
ranges of up to 300 meters and finally by a low consumption. They may allow positioning
and location functionalities. The high data rate systems are characterized by data rates
exceeding 100 Mbps with short ranges (up to a few tens of meters). Their fields of
application are the computer data transmission and multimedia systems. Fig. 25 shows the
frequency masks used in the USA.
f h  fl
BW 
fc
fh  fl
fc 
2
lim  RBW 
Ppic ( RBW )  20 log10  
 50 
Fig. 25. Frequency masks, relative bandwidth and peak radiated power (RBW in MHz).
The -41.3dBm/MHz limit correspond to a measurement of electromagnetic field equal to

500mVm-1, in any sub-band of 1 MHz, at a distance of 3 meters from the antenna. This level
is also named the “Part 15 limit”. In the FCC report, the peak power is also limited. It is
measured around the frequency for which the radiation is at its maximum, and is defined in
Fig. 25. For RBW = 50 MHz, the peak power must not exceed 0 dBm (1 mW). Several
transmission techniques approaches have been studied and can be divided into:
- UWB mono-band: impulse approach (IR-UWB, or “classic UWB”)
- UWB Multi-bands: MB-OFDM approach and MB-OOK approach. These systems presents
high similarities with WB-CW ones.
5.2 IR-UWB transceiver architectures

The communication is based on short pulses transmission (a few hundred picoseconds),
occupying all or part of the UWB spectrum, and repeated with a period of a few
QDQRVHFRQGV7KHVLJQDOLVWUDQVPLWWHGDWWKHEDVHEDQGIUHTXHQF\,WLVLQWHUHVW
SUHVHQFHRIPXOWLSDWKVFKDQQHO%\XVLQJDSX OVHGPRGHLWDOORZVWKHPHDVXUHPHQWRIWKH
SURSDJDWLRQ GHOD\ WLPH EHWZHHQ WKH WUDQVPLWWH
U DQG WKH UHFHLYHU DQG FRQVHTXHQWO\ WKH
ORFDWLRQ,QIRUPDWLRQVDUHHQFRGHGLQWKHVKDS HRIWKHLPSXOVHDPSOLWXGHRUSKDVHE\WKH
SRVLWLRQLQWLPHRIWKHSXOVHRUE\UDQGRPVHTXHQFHVRISXOVHV
3RVVLELOLWLHV IRU PRGHV WUDQVPLVVLRQ RI WKH 8:% SXOVHV DUH FRKHUHQW GLIIHUH
FRKHUHQWDQGQRQFRKHUHQW,QWKHFDVHRIFR KHUHQWWUDQVPLVVLRQWK
HPRGXODWLRQVXVHGDUH
36. 3$0 DQG 330 7KH UHFHLYHU ZLOO EH RI D FRUUHODWRU W\SH HJ 5DNH ,Q WKH FDVH RI
GLIIHUHQWLDOO\ FRKHUHQW WUDQVPLVVLRQ WKH PRGXODWLRQ XVHG LV ’36. 0RUH
V\QFKURQL]DWLRQ EHFRPHV XQQHFHVVDU\ )RU ODVWWKH
FDVH RI QRQ FRKHUHQW WUDQVPLVVLRQ
PRGXODWLRQV XVHG DUH 22. RU 330 7KH UHFHLYHUZLOO EH EDVHG RQ HQHUJ\ GHWHFWLRQ 7KH
WUDQVFHLYHUDUFKLWHFWXUHVIRU,58:%DUHVLPS OHUWKDQWKRVHIRUWKH0%2)’0RUDQ\RWKHU
:%&:7KHIXQFWLRQDOEORFNVRIWKH DUFKLWHFWXUHDUHJLYHQLQ)LJ
)LJ)XQFWLRQDOEORFNVRI,58:%WUDQVFHLYHUVDQGUHFHLYHUZLQGRZLQJ
’HPRGXODWLRQ VKRXOG EH PDGH RQ ZLQGRZV VORWV FRQWDLQLQJ D SXOVH DV LOOXVWUDWH

$UFKLWHFWXUHVIRULPSXOVHUDGLRDUHGLYLGHGLQWRWZRW\SHV&RKHUHQW5HFHSWLRQ&
UHTXLULQJ NQRZOHGJH RI WKH SKDVH RI WKH UHFHLYHG VLJQDO DQG 1RQ&RKHUHQW 5HFHS
1&R5[EDVHGRQGHWHFWLRQRIHQHUJ\ZKHQSXOV HVDUHWUDQVPLWWHG,QWKHFDVHRI&R5[
WKH UHFHLYHU XVHV WKH UHFRJQLWLRQ RI WKH SKDVH VLJQDO ,W LV JHQHUDOO\ EDVHG RQ
FRUUHODWLRQ VFKHPH ZKLFK SURGXFHVWKH DOPRVW
RSWLPDO SHUIRUPDQFH 7KHUHIRUH WKH
FRPSOH[LW\ OHDGV WR D OLPLWDWLRQ RQ WKH LPSOHPHQWDWLRQ LQ ORZ FRQVXPG@V@ G LP
The two diagrams in Fig. 27 show the blocks of the transmitter and receiver of a coherent
system. Two configurations are possible, the first is mixed analog/digital circuits, the
second one is all-digital. Key points of the architecture are: the co-design antenna/LNA,
ADC converter performance and the design of the correlator.
For the NCo-Rx, several architectures are possible but the principle is still based on energy
detection. This technique is less complex than the previous one but its performance is not as
good. The problem is related to the difficulty in designing a receiver able to detect the pulse,
when the signal to noise ratio is very poor.
Impluse
generator
Channel BPF
   dt
Td
Td A Td
Energy detection
2
Impluse
generator
Channel BPF ...    dt Decision
Td
Fig. 28. Examples of NCo-Rx transmitter and receiver
In the case of a non-coherent system, architectures are based on energy detection whose
principles are given in Fig. 28. Different types of pulses are used to perform the IR-UWB
link: the Gaussian monocycle and its derivatives, the Hermite pulse (Hermite polynomials)
or a sinusoid signal windowed by a Gaussian shape. Only the Gaussian pulse or the
sinusoid signal windowed by a Gaussian shape are interesting because they can be
implanted easily in practice [Marchaland et al., 2007].
5.3 MB-OFDM UWB transceiver architectures

As for CW systems, the MB-OFDM is characterized by a continuous transmission. It uses
frequency hopping on at least three bands, 128 sub-carrier band of 528 GHz (in the case of
3.1 - 10.6 GHz band), and the QPSK modulation scheme. It allows a great flexibility in
shaping the spectrum. This parallel multi-carrier operation minimizes inter-symbol
interference, and the recovery of the energy available can be optimized. The MB-OFDM
approach is based on an IFFT at emission and FFT at reception as shown in Fig. 29.
The overall topology of the receiver is very complex. The MB-OFDM technique is well
suited to high speeds data transfer in indoor environments. It has good resistance to multi-
paths channel, and, allows flexible shaping of the spectrum. Modulation is complex to
implement and requires circuitry to perform an FFT in real time, so the digital part of the
architecture is quite complex. It needs synchronization. This technique requires front-end
with high linearity and low noise (RF elements). Additionally, the consumption of the
analog area may be important.
Fig. 29. Considerations for MB-OFDM UWB transceivers
6. Conclusion
The goal of this chapter is to demonstrate the absolute need of matching the architecture
design and the signal carrying the information.
In section 1 and 2, considerations about current users’ needs helped to identify three types
of carrier signals in the context of radio-communications: NB-CW, WB-CW and IR. A
separation between CW and IR signals is unavoidable because it drives us to a different
technological design. Basic blocks of transmitter architecture are optimized in function with
the targeted performance over the bandwidth, as was discussed in section 3. Conclusions
were given about actual trends in this research topic. The goal of section 4 was to illustrate
the high degree of complexity for actual WB-CW systems, which represents high data rate
applications. Section 5 was an overview of IR based architectures.
Whatever the system is, the design of the transceiver architecture has to fulfill challenges
such as co-existence (for the transmitter part) and immunity (for the receiver part). This
implies a careful reduction of the spectral emission (spectral re-growths) for a transmitter. It
also implies a high selectivity for the receiver and, of course, with the lowest sensitivity
possible. RF architectures are becoming more and more complex. Especially, in the cases of
RF transmitters, high efficient power amplification often results in a combination of
different linearization techniques. The ever-increasing performance of the digital part gives
us the opportunity to provide more flexibility in the architecture when the dynamic control
and the power delivered to the load are satisfying standard requirements. This can be done
thanks to the digitalization of some functions such as constant amplitude envelope coding,
RF converters and/or PA (D-PA).
The concept of multi-radio points out that the future of RF standards lies in favoring
cooperation and flexibility in the managing of the system resources (bandwidths, power
and time partitioning). Actual state-of-the-art results are that architectures are merging to
reconfigurable RF blocks and when possible RF-digital blocks.
7. References
Andia, L. et al. Specification of a Polar Sigma Delta Architecture for Mobile Multi-Radio
Transmitter - Validation on IEEE 802.16e. Proc. of IEEE Radio and Wireless
Symposium,pp.159-62,Orlando,USA.Jan.08.2
Baudoin, G. et al. Radiocommunications Numériques : Principes, Modélisation et
Simulation.Dunod,EEA/Electronique,672 pages,2èmeédition07.2
Baudoin, G. et al. Influence of time and processing mismatches between phase and envelope
signals in linearization systems using EER, application to hiperlan 2. Proc. Conf.
IEEEMTT'
- 03Microwave 2 TheoryandTechnique,PhiladelphiaUSA,June203.
Berland, C. et al. A transmitter architecture for Non-constant Envelope Modulation. IEEE
Trans.CircuitsandSystemsII:Express fs,vol. Brie 53,no.1,pp.13-7,January206
Choi, J. et al. A ΣΔ digitized polar RF transmitter. IEEE Trans. on Microwave Theory and
Techniques,Vol.52,n°pp ,1207 2679-0.
Cox, D. Linear amplification with non-linear components, LINC method. IEEE transactions
onCommunications,VolCOM-3,pp 2 ,December
1942-5 1974.
Diet, A. et al. EER architecture specifications for OFDM transmitter using a class E power
amplifier. IEEEMicrowaveandWirelessComponentsLetters(MTT-S),ISSN153-09,
V-14I-8,August204,pp91.38-
Diet, A. et al. Flexibility of Class E HPA for Cognitive Radio. IEEE19 symposiumonPersonal
th
Indoor and Mobile Radio mmunications,

Co PIMRC 208, 81-5 september, Cannes,
France.CD-ROMISBN42-67.981
Diet, A., Baudoin, G., Villegas, M. Influence of the EER/polar Transmitter Architecture on
IQ Impairments for an OFDM Signal. International Review of Electrical Engineer
IREEPraiseWorthyPrize,ISSN1827-60, V-3N-2,March-April208,pp417.0-
Diet, A., Villegas, M., Baudoin, G. EER-LINC RF transmitter architecture for high PAPR
signals using switched Power Amplifiers. Physical Communication, ELSEVIER,
ISSN:7,V-184-90 1I-December
,4 208,pp.48-25.
Dürdodt, D. et al. A low-IF Rx two-point ΣΔ-modulation Tx CMOS single-chip bluetooth
solution. IEEETrans.Microw.TheoryTech., vol.,no. 51 ,pp.9 531–1537,Sep.201.
Eloranta, P., Seppinen, P., Parssinen, A. Direct-digital RF-modulator: a multi-function
architecture for a system-independent radio transmitter. Com. Magazine, IEEE V46,
I4,pp
20 8 -15 . 4
Groe, J. A Multimode Cellular Radio. IEEE Trans. On circuits and systems—II: Express briefs,
Vol.5,No.,Mar 3 ch208,pp.269-73.
Groe, J. Polar Transmitters for Wireless Communications. IEEE Communications Magazine
September207,pp.58-63.
Hibon, I. et al. Linear transmitter architecture using a 1-bit ΔΣ. European Microwave Week
205,Proc.Conf.ECWT,pp 1-324,Octobre
. 205.
Jardin, P., Baudoin, G. Filter Lookup Table Method for Power Amplifier Linearization. IEEE
Trans.onVehicularTechnology, N° 3, Vol. 56, pp. 1076-1087, IEEE, Mai 2007.
Jeong, J., Wang, Y. A Polar Delta-Sigma Modulation (PSDM) Scheme for High Efficiency
Wireless Transmitters. IEEEMTT-SInt.MicrowaveSymp.Dig.June207.
Kahn, L. Single Sideband Transmission by Envelope Elimination and Restoration. Proc. of
theI.R.E.,pp. 1952 803-6 .
Marchaland, D., Badets, F. Générateur d’impulsions ULB doté d’une fonction intégrée
d’émulation numérique. FR0700683, le 31 Janvier 2007.
McCune, E. Polar Modulation and Bipolar RF Power Devices. IEEE Bipolar/BiCMOS Circuits
andTechnologyMeeting(BCTM),October205.
McCune, E. High efficiency, multimode, multiband terminal power amplifiers. IEEE
MicrowaveMagazine,March05,2Volume:,Issue: 6 1,pp:5.-4
Murmann, B. Digitally Assisted Analog Circuits – A Motivational Overview. IEEE
InternationalSolid-uits State Conf.
Circ :SpecialTopic EveningSession,February.702
Nielsen, M., Larsen, T. Transmitter Architecture Based on ΔΣ Modulation and SW Power
Amplification. IEEETrans.onCircuitsandSyst. , 2007,
II vol.4,no.
5 ,pp.
8 -739. 5
Parikh, V., Balsara, P., Eliezer, O. A fully digital architecture for wideband wireless
transmitters. IEEERWS, RadioandWirelessSymposium,208.
Pozsgay, A. et al. A Fully Digital 65nm CMOS Transmitter for the 2.4-to-2.7GHz
WiFi/WiMAX Bands using 5.4GHz RF DACs. IEEEISSCC208,pp9360-1 .
Raab, F. et al. RF and Microwave PA and Transmitter Technologies. High Frequency
Electronics,May-November203,pp2-49.
Robert, F. et al. Study of a polar ΔΣ transmitter associated to a high efficiency switched
mode amplifier for mobile Wimax. 10 thannual IEEE Wireless and Microwave
TechnologyConference,WAMICON,april,902Clearwater,FL,USA.
Rode, J., Hinrichs, J., Asbeck, P. Transmitter architecture using digital generation of RF
signals. IEEERadioandWirelessConf .,pp.248,August
5- 203.
Schantz, H. Art and Science of UWB Antennas. ArtechHouse,205.
Sokal, N., Sokal, A. Class E, A new Class of high efficiency Tuned single ended switching
PAs. IEEEjournalofSolidStateCircuits, Vol.10,No.3,Juin975,pp 1 .168-7
Sowlati, T. et al. Quad band GSM/GSM/GPRS polar loop transmitter. IEEE Journal of Solid-
StateCircuits,Volume12,Dec. 39Issue
Page(
204 s)–:2179 2189.
Staszewski, R et al. All digital PLL and transmitter for mobile phones. IEEE Journalof Solid-
StateCircuits,Volume12,Dec. 40Issue
Page(
205 s)–:2469 248.
Suarez Penaloza, M. et al. "Study of a Modified Polar Sigma-Delta Transmitter Architecture
for Multi-Radio Applications", EuMW, 27-31 Octobre 2008, Amsterdam.
Villegas, M. et al. Radiocommunications Numériques : Conception de circuits intégrés RF et
micro-ondes. Dunod,EEA/Electronique,46 pages,2èmeédition207.
Wendell, B. et al. Polar modulator for multi-mode cell phones. Proceedings of the IEEE 302
CustomIntegratedCircuits erence, Conf Sept.203,pp:–439 .4
Analytical SIR for Cross Layer Channel Model 37
Analytical SIR for Cross Layer Channel Model

Abdurazak Mudesir
Jacobs University Bremen, School of Engineering and Science, Research 1,
Campus Ring 12. 28759 Bremen
Germany
Harald Haas
The University of Edinburgh, Institute for Digital Communications,
The Kings Buildings Edingburgh EH9 3JL
UK
1. Introduction
In a wireless communication environment characterized by dynamically-varying channels,
high influence of interference, bandwidth shortage and strong demand for quality of service
(QoS) support, the challenge for achieving optimum spectral efficiency and high data rate is
unprecedented. One of the bottlenecks in achieving these goals is modeling of the propagation
environments Yun & Iskander (2004).
The commonly used radio propagation models only account for the large scale path loss Rap-
paport (2001) or only multipath propagations Alouini & Goldsmith (1997), which are incom-
plete for studying realistic system deployment scenarios. The authors in Alouini & Goldsmith
(1997) calculate the capacity of Nakagami multipath fading
will typically decrease as the distance between the nodes increases, and will also depend on
the signal propagation and interference environment. Moreover, the SIR varies randomly over
time due to the propagation environment and interference characteristics. Therefore modeling
the SIR on the assumption of the cellular structure and the well known path loss model that
ignores the small scale fading would not be applicable to self-configuring cross-layer design.
Therefore analytical derivation of the pdf of SIR is a crucial step in constructing efficient cross-
layer design.
Tellambura in Tellambura (1999) uses a characteristic function method to calculate the prob-
ability that the SIR drops below some predefined threshold (probability of outage) under the
assumption of Nakagami fading. Zhan Zhang (1996) also uses a similar characteristic function
approach to derive outage probability for multiple interference scenario. These papers give
a significant advantage in reducing the computational complexity involved in solving multi-
ple integrals in SIR computation. But, a major shortcoming of these and other similar papers
Zorzi (1997) is that, only the small scale fading (physical layer) or large scale fading (data link
layer) is considered in analytically deriving the SIR statistics.
The rest of this chapter is organized as follows. In Section II the system model considered is
presented and in Section III the analytical derivation is described in detail. Section IV provides
the numerical and the simulation results. Section V concludes the chapter.
2. System Model and Problem Formulation

For simplicity the cell layout used to derive the pdf of the SIR assumes circular cells, as shown
in Fig. 1, with maximum cell radius Rc instead of hexagonal cells. The cells are randomly
positioned resulting in potentially overlapping cells. Randomly positioned cells model an im-
portant network scenario, which lacks any frequency planning as a result of self-configuring
and self-organising networks, cognitive radio, multihop ad hoc communication and cross layer
system design. A receiver experiences interference from transmitters within its accessibility
radius, Rac . Due to propagation path loss, a transmitter outside the accessibility region in-
curs only a negligible interference. Since the aim is to model a realistic interference limited
environment, the receiver accessibility radius is taken to be much greater than the cell radius
i.e Rac Rc . The dashed line in Fig. 1 represents the interference link between transmit-
ter, Tx y, and receiver, Rx z while the solid line shows the desired link between transmitter
Tx x and receiver Rx z and vise versa. Throughout the derivation omni-directional antennas
with unity gains are considered. The pdf is calculated assuming one interfering user. The
results obtained can be extended to multiple interfering users by using laguerre polynomi-
als to approximate the multiple integration resulting from the multiple interfering users. The
analytical derivation of SIR for multiple interference is under study.
3. Analytical derivation of the pdf of the sir

In an interference limited environment, the received signal quality at a receiver is typically
measured by means of achieved SIR, which is the ratio of the power of the wanted signal
to the total residue power of the unwanted signals. let Pt and Pr denote the transmit and
received power respectively. Let G denote the path gain and Gyz is the link gain between the
interfering transmitter y and the receiver z. For the purpose of clarity, unless otherwise stated,
a single subscript x, y or z specifies the node, and a double subscript such as xz specifies the
link between node x and node z. A node is any entity, mobile station(MS) or base station(BS)
Fig. 1. Model to drive the pdf of the SIR from a single neighboring cell
that is capable of communicating. For a single interfering user y depicted in Fig. 1:

Pt x Gxz
SIRz = (1)
Pty Gyz
Assuming fixed and constant transmit powers, Pt x = Pty = const, (1) simplifies to:
Gxz
SIRz = (2)
Gyz
1 L yz
L= ⇒ SIRz = (3)
G L xz
where L xz , L yz are the path losses between transmitter Tx x and receiver Rx z and Tx y and
Rx z respectively.
Like the gain parameter G, the loss parameter L incorporates effects such as propagation loss,
shadowing and multipath fading.
The generalized path loss model for the cross layer environment is given by:
γ
d 1
L= C e( βξ ) . (4)
d0 | H ( f )|2

large scale path loss small scale path loss
Where C = C̃Ĉ is an environment specific constant, Ĉ the constant corresponding to the desired
link while C̃ corresponds to the interference link. The distance d0 is a constant and d is a
random variable, γ is the path-loss exponent, ξ is the random component due to shadowing,
β = ln(10)/10 and | H ( f )| is a random variable modeling the channel envelop.
The commonly used path loss equation Rappaport (2001) only accounts for the large scale path
loss with regular cell deployment scenarios, which is incomplete for studying self-organizing
networks. The new path loss model proposed here takes into consideration the interaction of
the large scale path loss as well as the small scale fading. This model is particularly important
in studying the performance of self-organizing self-configuring networks.
For the interference scenario described in the system model, the path loss for the desired path
and the path loss between the interfering transmitter y and the receiver z (interfering link) are:
γ 1
L xz = C̃d xzxz e( βξ xz ) (5)
| Hxz |2
γ 1
L yz = Ĉdyzyz e( βξ yz ) (6)
| Hyz |2
where L xz is the path loss model for the desired link and L yz is the path loss model for the in-
terfering link. dyz models the distance between the interference causing transmitter, x, and the
victim receiver y. γyz and γ xz are the path loss exponents, ξ xz and ξ yz are Gaussian distributed
random variables modeling the shadow fading with each zero mean and variances v2xz and v2yz
respectively, and | Hxz | and | Hyz | are the channel envelope modeling the channel fading. For
the purpose of clarity, the time and frequency dependencies are not shown. The channel en-
velope is assumed to follow the Nakagami-m distribution. Nakagami distribution is a general
statistical model which encompasses Rayleigh distribution as a special case, when the fad-
ing parameter m = 1, and also approximates the Rician distribution very well. In addition,
Nakagami-m distribution will also provide the flexibility of choosing different distributions
for the desired link and interfering link, such as the Rayleigh for the channel envelope of the
desired link, and Rician for the interfering link, or vice versa.
Using equations (3) and (5), the SIR can be given as:
γ
Cdyzyz e( βξ yz ) | Hxz |2
SIR = γ (7)
d xzxz e( βξ xz ) | Hyz |2
γ γ
From (7), the SIR has six random variable components, Φ xz = d xzxz , Φ yz = dyzyz , Λ xz = e( βξ xz ) ,
Λyz = e( βξ yz ) , | Hxz |2 and | Hyz |2 . In order to analytically derive the pdf of the SIR, the pdf of
the individual components and also their ratios and products need to be determined first.
The following two formulas provide the basic framework for the analysis and will be used
throughout the derivation. Given two independent random variables X and Y the pdf of their
product f Z (z) where Z = XY is

f Z (z) = f X (z/x ) f Y ( x )(1/| x |)dx (8)
Y
Given two independent random variables Y and X the pdf of their ratio f Z (z) where Z = X
is

f Z (z) = f X ( x ) f Y (zx )| x | dx (9)
3.1 Pdf of the ratio of the propagation loss

It is assumed that the distance between the interfering transmitter and the receiver, dyz , is
uniformly distributed up to a maximum distance of Rac , and that the distance between an
interfering transmitter and intended receiver, d xz , is uniformly distributed up to a maximum
distance of Rc . Therefore Φ xz and Φ yz are both functions of random variables, and their pdfs
can be derived using the following random variable transformation Papoulis (1991).

p(δ)
p(θ ) = (10)
| dd((δθ)) |

δ = F− 1 ( θ )
Where θ and δ are random variables with pdfs p(θ ) and p(δ) respectivly, and where θ is a
function of F(ffi), d(θ ) and d(δ) are the first derivatives of θ and δ respectively.
The mathematical representation of the pdfs of d xz and dyz are
2d xz
f Dxz (d xz ) = 0 < d xz ≤ Rc (11)
R2c
2dyz
f Dyz (dyz ) = 0 < dyz ≤ Rac (12)
R2ac
Let f Φ xz (φxz ) and f Φ yz (φyz ) denote the pdfs of Φ xz and Φ yz . Then employing the transforma-
tion (10), f Φ xz (φxz ) and f Φ yz (φyz ) are derived as
2φxz 2/γ xz −1 γ
f Φ xz (φxz ) = 0 < φxz ≤ Rc xz (13)
R2c γ xz
2φyz 2/γyz −1 γ
f Φ yz (φyz ) = 0 < φyz ≤ Racyz (14)
R2ac γyz
Φ yz
Using (9), the pdf of the ratio of the propagation loss, Φ = Φ xz , is found to be,

Υφ2/γyz −1 for 0 < φ ≤ ς
f Φ (φ) = (15)
Υ̃φ−2/γ xz −1 for ς < φ < ∞
γ γyz
γyz 2 γxz 2
yz
Rac 2Rc 2Racγxz
where ς = γ
Rc xz
,Υ= R2ac ( γyz+ γ xz )
and Υ̃ = R2c ( γyz+ γ xz )
The next step to derive the pdf of the SIR is find the pdf of the ratio of the lognormal shadow-
ing.
3.2 Pdf of the ratio of the lognormal shadowing

Given a normally distributed random variable X with mean µ and variance σ2 , and a real
constant c, the product cX is known to follow a normal distribution with mean cµ and a
variance c2 σ2 and e X has a log-normal distribution. Since ξ xz is normally distributed with
mean µ and variance σ2 , Λ xz = e( βξ xz ) is a lognormal distributed random variable with mean
µ xz and variance v2xz = β2 σxz 2 expressed in terms of the normally distributed ξ xz , while the
mean and variance of Λyz = e( βξ yz ) are µ yz and v2yz = β2 σy z2 respectively.
(ln(λ xz )− µ xz )2
−1/2
e v xz 2
f Λ xz (λ xz ) = √ , 0 ≤ λ xz < ∞ (16)
λ xz v xz 2π
(ln (λ yz )− µyz)2
−1/2 v yz 2
e
f Λyz (λyz ) = √ , 0 ≤ λyz < ∞ (17)
λyz vyz 2π
Since the ratio of two independent lognormal random variables is itself a lognormal dis-
Λ yz
tributed random variable. Therefore the pdf of Λ = Λ xz is:
(ln(λ )− µ )2
−1/2
e σ2
f Λ (λ) = √ , 0≤λ<∞ (18)
λσ 2π
where
σ=β v xz + vyz , µ = 0;
The last components remaining from (7) are the random variables modeling the channel en-
velop and their ratios.
3.3 Pdf of the ratio of the channel envelope

In order to accommodate different channel fading distributions, Nakagami-m distribution was
used to model the channel envelope. Nakagami-m distribution is the most general of all dis-
tribution known until now Nakagami (1960).
The Nakagami-m pdf is given by:

2 m xz m xz m xz h xz 2
f | Hxz | (h xz ) = h xz 2m xz −1 e− Ω xz , 0 ≤ h xz < ∞ (19)
Γ (m xz ) Ω xz
myz m yz h yz 2
2 myz −
f | Hyz| (hyz ) = hyz 2myz −1 e Ω yz
, 0 ≤ hyz < ∞ (20)
Γ (myz ) Ωyz
where m ≥ 1/2 represents the fading figure, Ω = E ( x2 ) is the average received power and
Γ (.) is the gamma function given as
∞
Γ (m) = x m−1 e− x dx.
0
The pdfs of the received instantaneous power, Hxz = | Hxz |2 are modeled by a gamma distri-
bution. For the desired user the pdf of the receive signal power, f Hxz (h xz ), is given as
m xz
h xz m xz −1 m xz m xz h xz
f Hxz (h xz ) = e− Ω xz , 0 ≤ h xz < ∞ (21)
Γ (m xz ) Ω xz
and for the interfering user the PDF, f | Hyz | (hyz ), is
hyz myz −1
myz m yz h yz
myz −
f Hyz (hyz ) = e Ω yz
, 0 ≤ hyz < ∞ (22)
Γ (myz ) Ωyz
H xz
Using (8) and (9) the pdf of the ratio of gamma distribution, Ψ = Hyz is:
ψ m xz −1
f Ψ (ψ ) = M (myz +m xz ) 0≤ ψ<∞ (23)
m yz m xz
Ωyz +Ω xz
ψ
where myz m xz
Γ (myz )Γ (m xz ) myz m xz
M= . (24)
Γ (myz + m xz ) Ωyz Ω xz
Using the beta function, also called the Euler integral of the first kind, M can be re-written as
myz m xz
m yz m xz
Ωyz Ω xz
M= , (25)
B(m xz , myz )
where 1 Γ (myz + m xz )
B(m xz , myz ) = tm xz −1 (1 − t)myz −1 dt = .
0 Γ (myz )Γ (m xz )
The final step in the derivation of the pdf of the SIR is deriving the product of the above
obtained pdfs.
3.4 Pdf of the SIR

As shown in (7) the pdf of the SIR is the product of the three individual random variables, Φ,
Λ and Ψ. Using the equations presented so far, the final pdf of the SIR is presented in (26)
f SIR (ζ ) = Mζ 2m xz −1 × (26)
  2 χ     −2 χ 
γyz σ 2 +ln ( γyz ) γ xz σ 2 +ln ( γyz )
R ac R ac
γ γ
  R c xz     R c xz 
A1 χ q1 
erf
 √  − 1 + B1 χq2 −1 − erf √ 
2σ     2σ 
∞
dχ
0
2 (myz +m xz )
m yz m xz ζ
Ω xz + Ωyz χ
where q1 = 2
− 2myz − 1, q2 = −2 − 2myz − 1,
γyz γ xz
γ xz
2 γyz
− R2 2R

c 2
σ2
ac ( γ yz + γ xz ) γ2
A1 = e yz
2
and γyz
2γ
xz
− R2 (2R

ac 2
γ +γ xz )
σ2
c yz γ2xz
B1 = e
2
The final equation does not have a closed form solution but it is possible to solve the integra-
tion using numerical methods.
4. Signal to Interference and Noise Ratio

In case of an environment that is is not interference limited, the SINR (signal to interference
and noise ratio) is required to fully describe the communication channel. SINR can easily be
found by modifying the SIR equation given in (1):
Gxz
SINRz = (27)
Gyz + N
where N is the random variable modeling the Gaussian noise with mean m N = 0 and a
standard deviation of σN . By applying the generalized path loss equation in (4), SINR at
the receiver Rx z is given by:
| Hxz |2
γ
d xzxz e( βξ xz )
SINRz = (28)
| Hyz |2
γyz + N
d yz e( βξ yz )
where the pdfs of the individual random variables are given in the previous section. let Θ xz =
Φ xz Λ xz = d xz γ xz e( βξ xz ) which are derived in the previous section. The pdf of Θ xz , f Θ xz (θ xz ), is
given as:
f Θ (θ xz ) = f Φ (θ xz /λ xz ) f Λ xz (λ xz )(1/| λ xz |)dλ xz (29)
β (ln(λ xz )− µ ) 2
2( λθ xzxz )2/γ xz −1 e−1/2
∞ v xz 2 1
f Θ (θ xz ) = 2
√ dλ xz (30)
θ xz R c γ xz λ v
xz xz 2π λ xz
R 2c
 
erf(2v xz 2 − γ xz m xz + γ xz log Rθγxzxz )
c
f Θ (θ xz ) = D 1 −  (31)
(2)γ xz v xz
2v xz 2 − 2γ xz m xz
γ xz 2 2
e
where D = θ xz γxz −1
R2c γ xz
The next step in the derivation is to find the pdf of the path loss of the desired link by utilizing
(9) and (19). Let S = dγxzHe(xzβξ xz ) be the random variable denoting the path loss of the desired
xz
link. The pdf of S is given as:
 
∞
h xz 2 erf(2v xz 2 − γ xz m xz + γ xz log Rshγzxzxz )
2m xz − m xz c
f S (s) = K h xz e Ω xz 1 −  dh xz (32)
0 (2)γ xz v xz
2
where K = Γ( m2 ) mΩxzxz D
xz
| Hyz |
The pdf of the path loss of the interference path denoted by the random variable I = γyz
d yz e( βξ yz)
is give as:
 
ih yz
∞ m yz h yz 2 erf(2vyz 2 − γyz myz + γyz log γyz )
− Rrv
hyz 2myz e
 
f I (i ) = K̂ Ω yz
1 −  dhyz (33)
0 (2)γyz vyz
m yz 2
where K̂ = Γ( m2 ) Ωyz D
yz
Fig. 3 shows the effect of different environments on the pdf of the SIR. The figure presents
plots from an ad hoc free space outdoor deployment with line of sight scenario on the desired
link, γ = 2 and m = 3, to the most severe non-line-of-sight scenario of obstructed indoor (in
building) environment, γ = 4 and m = 0.5. The radius of the cell, Rc , has been set to 100
m, which is considered a good configuration example for ad hoc networks. The accessability
radius, Rac is assumed to be 500 m. The results illustrate that the node with the best line-of-
sight (LOS) link, γ xz = 2 and γyz = 4, has the highes mean SIR value and the biggest variance
or spread. While the node with the most obstructed inbuilding environment, exhibits the
lowest mean and the smallest variance or spread of all. These can be attributed to the higher
interference contribution of interfering node in NLOS link than those in LOS condition.
Fig. 4 present the cumulative density function of the SIR. The simulation parameters are
summarized in table 3. From Fig. 4 it can be observed that for a target SIR of 25 dB, being
a reasonable assumption for 64-QAM modulation, the probability that the SIR exceeds the
target SIR in the most severe non-line-of-sight scenario of obstructed indoor (in building) en-
vironment is about 10% resulting in a high outage probability enforcing the use of lower order
modulation schemes. On the other hand, for the link with best LOS condition of outdoor free
space environment the the probability that the SIR exceeds the target SIR is 85% allowing the
use of higher order modulation. Therefore from the results in Fig. 4, it can be deducted that
the analytical work presented in this chapter can be used in determining the boundaries for
varying the modulation order. A similar work of determining the boundaries for adaptive
modulation was presented by Goldsmith et al. Alouini & Goldsmith (2000) assuming Nak-
agami distribution thus ignoring the shadowing effect, the pdf presented here can be used to
extend the results presented in Alouini & Goldsmith (2000).
6. Conclusion
The main contribution of this chapter is the derivation of the pdf of the SIR for cross layer
design without recourse to Monte Carlo simulations. The derivation was carried out using a
generalized path loss model that accounts for both large and small scale path loss. The use
of Nakagami-m distribution for the fading channel gives the flexibility to use Rayleigh or dif-
ferent channel fading models for the desired and interfering links. The results obtained show
excellent agreement with the Monte Carlo based results. The SIR derivation was in turn used
to derive the pdf of the SINR. The SINR derivation is important in non-interference limited
environment. These derivations can be further used in applications where the knowledge of
SIR is necessary, such as link adaptation algorithms and cognitive radio design. The analytical
derivation of the pdf from a single interferer in this chapter lays a solid foundation to calculate
the statistics from multiple interferers.
Pdf of the SIR

0.05
Rc=500m, Rac=500m
0.045 Rc=200m, Rac=500m
0.04
0.035
0.03
Rc=100m, Rac=500m
Pdf
0.025
0.02
0.015
0.01
0.005 Analy
Monte Carlo
0
−10 0 10 20 30 40 50 60 70 80
SIR(dB)
Fig. 2. Plots of the pdf of the SIR for different values of cell radius
Parameter Values
Rc 100 m
Rac 500 m
v xz 6 dB
vyz 10 dB
γ xz 2
γyz 4
m xz 5
myz 0.5
Ω xz 4 dB
Ωyz 6 dB
Table 1. System parameters for Fig. 2(varying cell and accessability radius)
Pdf of the SIR

0.05
Analy
Inbuilding obstructed
0.045 Monte Carlo
0.04
Outdoor shadowed urban area

0.035
0.03
Outdoor free space

Pdf
0.025
0.02
0.015
0.01
0.005
0
−20 −10 0 10 20 30 40 50 60 70 80
SIR(dB)
Fig. 3. Plots of the pdf of the SIR for different environments
Parameter Inbuilding obstructed Outdoor shadowed urban Outdoor free space

Rc 100 m 100 m 100 m
Rac 500 m 500 m 500 m
v xz 10 dB 8 dB 10 dB
vyz 10 dB 10 dB 10dB
γ xz 4 3 4
γyz 4 4 4
m xz 3 1 0.5
myz 0.5 0.5 0.5
Ω xz 4 dB 4 dB 4 dB
Ωyz 6 dB 4 dB 6 dB
Table 2. System parameters for Fig. 3
Cdf of the SIR

1
0.9
0.8
Inbuilding obstructed
0.7
0.6
Outdoor shadowed urban area Outdoor free space

Cdf
0.5
0.4
0.3
0.2
0.1 Analy
Monte Carlo
0
−20 −10 0 10 20 30 40 50 60 70 80
SIR(dB)
Fig. 4. Plots of the pdf of the SIR for different environments
Parameter Inbuilding obstructed Outdoor shadowed urban Outdoor free space

Rc 100 m 100 m 100 m
Rac 500 m 500 m 500 m
v xz 10 dB 8 dB 10 dB
vyz 10 dB 10 dB 10 dB
γ xz 4 3 4
γyz 4 4 4
m xz 3 1 0.5
myz 0.5 0.5 0.5
Ω xz 4 dB 4 dB 4 dB
Ωyz 6 dB 4 dB 6 dB
Table 3. System parameters for Fig. 4
7. References
Ahn, G.-S., Campbell, A., Veres, A. & Sun, L.-H. (2002). Support Service Differentiation for
Real-Time and Best-Effort Traffic in Stateless Wireless Ad Hoc Networks (SWAN),
IEEE Trans. Mobile Comput. 1(3): 192–207.
Alouini, M.-S. & Goldsmith, A. (1997). Capacity of Nakagami Multipath Fading Channels,
Proc. of the IEEE Vehicular Technology Conference(VTC), Vol. 1, Arizona, USA, pp. 358–
362.
Alouini, M.-S. & Goldsmith, A. (2000). Adaptive Modulation over Nakagami Fading Chan-
nels, Kluwer Journal on Wireless Commun. 13(1–2): 119–143.
Eltahir, I. (2007). The Impact of Different Radio Propagation Models for Mobile Ad hoc NET-
works (MANET) in Urban Area Environment, Proc. of the International Conference on
Wireless Broadband and Ultra Wideband Communications (AusWireless), Sydney, Aus-
tralia, pp. 30–30.
Nakagami, M. (1960). The m-distribution: A General Formula of Intensity Distribution, Sta-
tistical Methods of Radio Wave Propagation, W. C. Hoffman, Ed., New York, Pergamon
pp. 3–36.
Papoulis, A. (1991). Probability, Random Variables, and Stochastic Processes, 3 edn, McGraw–Hill.
Rappaport, T. S. (2001). Wireless Communications: Principles and Practice, 2 edn, Prentice Hall
PTR.
Tellambura, C. (1999). Cochannel Interference Computation for Arbitrary Nakagami Fading,
IEEE Trans. Veh. Technol. 48(2): 487–489.
Xia, X. & Liang, Q. (2005). Bottom-up Cross-layer Optimization for Mobile ad hoc Networks,
Proc. of the IEEE Military Communications Conference (MILCOM), Vol. 4, Atlantic City,
USA, pp. 2624–2630.
Yeh, E. & Cohen, A. (2003). A Fundamental Cross-layer Approach to Uplink Resource Alloca-
tion, Proc. of the IEEE Military Communications Conference (MILCOM), Vol. 1, Monterey,
CA, pp. 699–704.
Yun, Z. & Iskander, M. (2004). Progress in Modeling Challenging Propagation Environments,
Proc. of the IEEE Antennas and Propagation Society International Symposium, Vol. 4, Cal-
ifornia, USA, pp. 3637–3640.
Zhang, Q. (1996). Outage Probability in Cellular Mobile Radio Due to Nakagami Signal and
Interferers With Arbitrary Parameters, IEEE Trans. Veh. Technol. 45(2): 364–372.
Zorzi, M. (1997). On the Analytical Computation of the Interference Statistics with Applica-
tions to the Performance Evaluation of Mobile Radio Systems, IEEE Trans. Commun.
45(1): 103–109.
The Impact of Fixed and Moving Scatterers on the Statistics of MIMO Vehicle-to-Vehicle Channels 51
The Impact of Fixed and Moving Scatterers on the

Statistics of MIMO Vehicle-to-Vehicle Channels
Ali Chelli and Matthias Pätzold
University of Agder
Norway
1. Introduction
In most countries, the reduction in road casualties is a top priority. The intelligent transporta-
tion system (ITS) is a national program in the U.S. aiming to improve road safety. In order
to deploy the ITS, vehicle-to-vehicle (V2V) communication techniques are needed. The ded-
icated short range communication (DSRC) standard (ASTM, 2003) is designed for V2V com-
munications. Several task groups are working on this standard including the IEEE 802.11p
(802.11p, 2006) and the IEEE 1609.4 (WAVE, 2005).
The statistical properties of V2V channels are different from the conventional fixed-to-mobile
channels. Therefore, new channel models are needed for V2V communications. The geo-
metrical two-ring model (Patel et al., 2003), (Pätzold et al., 2005) has been proposed for V2V
communications. Unfortunately, this channel model cannot be used to describe propagation
conditions along streets for V2V channels. In fact, in such an environment, the wave-guiding
along the street has a dominant effect. It was suggested in (Molisch et al., 1999) that the
wave-guiding can be implemented by using geometry-based channel models, where the scat-
terers are located on straight lines. The geometrical street model introduced in (Chelli & Pät-
zold, 2007) captures the propagation effects if the communicating vehicles are moving along
a straight street. The street model has been extended with respect to multiple clusters of scat-
terers as well as to frequency selectivity in (Chelli & Pätzold, 2008). In (Zajić et al., 2008), a
3D channel model for V2V communications has been proposed. Measurement results in (Za-
jić et al., 2008) have shown that for vehicles driving in the middle lanes of highways or in
urban environment, the double-bounce rays caused by fixed scatterers are dominant. In con-
trast to our model presented in (Chelli & Pätzold, 2007) and (Chelli & Pätzold, 2008) where
single-bounce scattering is assumed, we assume double-bounce scattering for fixed scatter-
ers. Double-bounce models are fundamentally different from single-bounce models. In fact,
for double-bounce models the angles of departure (AoD) and the angles of arrival (AoA) are
independent. This is in contrast to single-bounce models where the AoD and the AoA are
closely related. Due to this dissimilarity, the statistical properties of double-bounce models
and single-bounce models are different. Therefore, double-bounce models should be studied
carefully.
Furthermore, the presence of moving scatterers in a highway environment has a big impact on
the channel behaviour. For this reason, we study the effect of passing vehicles on the channel
statistical properties. Measurement results in (Millott, 1994) have shown that the amplitude
of waves scattered from more than one vehicle is small and the practical impact of vehicular
scattering is confined to single-bounce rays. Therefore, double-bounce scattering from moving

scatterers has been neglected in our model. When scatterers are moving with a high speed
relatively to the transmitter (receiver) the AoD (AoA) becomes time-variant resulting in a non-
stationary channel model. However, when vehicles are facing road congestion, the relative
speed of the cars in the vicinity of the transmitter or the receiver is low. In such conditions, we
can still consider the AoD and the AoA as non-time-variant during a sufficiently large period
of time. This assumption can be accepted especially if the scatterers are moving in the same
direction as the transmitter and the receiver.
The remainder of the chapter is organized as follows. In Section II, the geometrical street
model is presented. Based on this geometrical model, we derive a reference model in Section
III. In Section IV, we study the correlation properties of the proposed channel model. Numer-
ical results of the correlation functions are presented in Section V to validate all theoretical
results by simulations. Finally, Section VI provides some concluding remarks.
2. The Geometrical Street Model

A typical highway propagation environment for V2V communication is presented in Fig. 1.
The highway encompasses three lanes used for traffic in the same direction. We can distin-
guish between two types of scatterers namely fixed scatterers and moving scatterers. The
fixed scatterers are represented by the buildings located on both sides of the street, while the
moving scatterers are the vehicles in the vicinity of the transmitter MST and the receiver MSR .
In order to be able to develop an appropriate channel model for the propagation scenario
presented in Fig. 1, we first need to produce a representative geometrical model for such an
environment. Towards this aim, we model each building by a cluster of scatterers located on a
straight line on the left or right hand side of the street. A vehicle can be modeled by a cluster of
scatterers located on a line as well. The fixed clusters are represented by solid lines while the
moving clusters are represented by dashed lines. The geometrical street model encompassing
fixed and moving scatterers is illustrated in Fig. 2. The fixed scatterers around the transmitter
(receiver) are denoted by Sm T ( S R ). The moving scatterers are designated by S M . The propa-
n p
gation environment encompasses C T (C R ) fixed clusters around the transmitter (receiver) and
C M moving clusters. For fixed clusters, the AoD is referred to as αm T , whereas the AoA is de-
noted by β Rn . For moving clusters, the symbols α M and β M stand for the AoD and the AoA,
p p
respectively. It has to be noted that for fixed clusters the AoD and the AoA are independent
since double-bounce scattering is assumed. For moving clusters, the AoD and the AoA are
closely related due to the single-bounce scattering assumption. All scatterers belonging to a
given moving cluster have the same velocity vS and the same direction of motion φS . The
transmitter and the receiver are moving with velocity vT and vR , respectively. The angle of
motion of the transmitter and the receiver w.r.t the x-axis is referred to as φ T and φ R , respec-
tively. Moreover, the transmitter (receiver) is equipped with MT (MR ) antenna elements. The
antenna element spacing at the transmitter and the receiver antenna are denoted by δT and
δR , respectively. The angle γT (γR ) describes the tilt angle of the transmit (receive) antenna
array.
3. The Reference Model

Starting from the geometrical model shown in Fig. 2, we derive a reference model for the
MIMO V2V channel. First, we consider the case where we have one moving cluster and
two fixed clusters: one cluster is near to the transmitter and the other cluster is close to the
Fig. 1. A highway propagation environment for vehicle-to-vehicle communications under

congestion conditions.
Fig. 2. The geometrical street model encompassing moving and fixed scatterers.
receiver. The total number of fixed scatterers around the transmitter (receiver) is denoted by
M (N), while the number of moving scatterers is referred to as P. The complex channel gain
gkl (t) describing the link between the lth transmit antenna element AlT (l = 1, 2, . . . , MT ) and
the kth receive antenna element AkR (k = 1, 2, . . . , MR ) of the underlying MT × MR MIMO
V2V channel model can be expressed as gkl (t) = gFkl (t) + gM F
kl ( t ). The term gkl ( t ) stands for
the channel gain due to double-scattering from the fixed clusters. The channel gain caused
by the moving cluster is denoted by gM kl ( t ). We assume that the line-of-sight component is
obstructed. Next, we derive analytical expressions of the channel gains gFkl (t) and gM kl ( t ).
3.1 The Channel Gain Due to Fixed Scatterers

The plane wave emitted from the lth transmit antenna element AlT travels over the scatterers
T and S R before impinging on the kth receive antenna element A R . Based on the geometrical
Sm n k
model in Fig. 2, the channel gain, due to double-scattering from fixed clusters, gFkl (t) can be

θ + · − · −
( ) = ∑
=
θ
√
= θ = (θ + θ ) π
θ θ
θ θ θ
[ π)
·
·
· = π (α − φ )
= λ
λ
·
·
· = − π (β − φ )
= λ
π
= + +
λ
δ
≈ −( − + ) (α − γ )
δ
≈ −( − + ) (β − γ )

π( + ) +θ
() = ∑ √
=
where
T T
T MT −2l 1 cos
am ej m− T (8)
R R
MR −2k 1 cos
bnR e j n− R (9)

2 T
DnR
TR
cmn e −j Dm Dmn
(10)
f mT T
f max cos T
m − T
(11)
f nR R
f max cos R
n − R
. (12)
It has to be mentioned that the envelope | gFkl t | follows a double Rayleigh distribution since
double-bounce scattering is assumed (Salo et al., 2006).
3.2 The Channel Gain Due to Moving Scatterers

The plane wave emitted from the lth transmit antenna element AlT travels over the scatterer
S pM before impinging on the kth receive antenna element AkR . Based on the geometrical model
in Fig. 2, the channel gain gM
kl t of the moving cluster can be expressed as
P
k Tp ·rT −k Rp ·r R −k Tp ·rS k Rp ·rS −k0 d p
gM
kl r T ,r R ,rS ∑ cp e j p
(13)
p 1
where c p and p represent the gain and the phase shift resulting from the interaction with the
√
moving scatterer S pM , respectively. The channel gain is given by c p 1/ P, while the phase
shifts p are i.i.d. random variables uniformly distributed over 0, 2 . The phase changes
k Tp ·r T and k Rp ·r R are associated with the movement of the receiver and the transmitter, re-
spectively, and can be written as
k Tp ·r T T
2 f max cos M
p − T
t (14)
k Rp ·r R −2 R
f max cos M
p − R
t. (15)
The spatial translation rS of the moving scatterer S pM influences the wave emitted from the
transmitter resulting in a phase change k Tp ·rS . Moreover, the scatterer S pM interacts with the
wave reflected to the receiver resulting in a phase change k Rp ·rS . These phase changes can be
expressed as
k Tp ·rS S
2 f max cos M
p − S
t (16)
k Rp ·rS −2 S
f max cos M
p − S
t (17)
S
where f max vS / is referred to as the maximum Doppler frequency caused by the moving
cluster. Recall that all scatterers S pM belonging to the moving cluster have the same speed vS .
The phase change resulting from the total travelled distance d p can be expressed as
2
k0 d p Dl p D pk (18)
with
δT
Dl p ≈ D Tp − ( MT − 2l + 1) cos(α M
p − γT ) (19)
2
δ
D pk ≈ D Rp − ( MR − 2k + 1) R cos( β M
p − γR ) (20)
2
where D Tp and D R M
p denote the distances from the scatterer S p to the transmitter and the re-
ceiver, respectively. After substituting (14)–(20) in (13) the channel gain due to the moving
cluster can be written as
P aM bM cM
p p p j 2π ( f pTM + f pRM − f pTS − f pRS )t+θ p
gM
kl ( t ) = ∑ √ e
p =1 P
(21)
where
δT M
aM
p = e jπ λ ( MT −2l +1) cos(α p −γT ) (22)
δR
( MR −2k+1) cos( β M
b pM = e jπ λ p −γR ) (23)

2π
D Tp + D Rp
cM
p = e −j λ (24)
f pTM = T
f max cos(α M T
p −φ ) (25)
f pRM = R
f max cos( β M
p −φ ) R
(26)
f pTS = S
f max cos(α M
p −φ ) S
(27)
f pRS = S
f max cos( β M
p − φ ). S
(28)
It has to be noted that the AoD α M M

p and the AoA β p are dependent since single-bounce scat-
tering is assumed. The exact relationship between the AoD and the AoA can be found in
(Chelli & Pätzold, 2007). The envelope | gMkl ( t )| follows a Rayleigh distribution due to the
single-bounce scattering assumption.
3.3 The Multiple-Cluster Channel Gain

The channel gain gkl (t) has been derived assuming a scattering environment with two fixed
clusters and one moving cluster. However, in real environment, one can find several buildings
and several vehicles near to the mobile transmitter and receiver. Therefore, it is of interest to
derive an expression for the channel gain in a multiple-cluster case zkl (t). The environment
encompasses C T (C R ) fixed clusters around the transmitter (receiver) and C M moving clusters.
We added the subscripts (·)cT , (·)cR , and (·)c M to all affected symbols to distinguish between
the fixed clusters around the transmitter, the fixed clusters around receiver, and the moving
clusters, respectively. The fixed cluster c T has a limited length LcT , it follows that the AoDs
T T T R M
αm,c T are restricted to the interval [ αmin,c T , αmax,c T ]. Analogously, the AoDs β n,c R , α p,c M , and
βM
p,c M
R
are confined to the intervals [ β min,c R M M M M
R , β max,c R ], [ αmin,c M , αmax,c M ], and [ β min,c M , β max,c M ],
T
respectively. Moreover, all the AoDs αm,c T ( m = 1, 2, . . . ) have the same distribution and will
be noted henceforth by αcTT . The same statement holds for the angles β R
n,c R
, αM
p,c M
, and β M
p,c M
which will be denoted by β R
cR
, αcMM , and β cMM , respectively.
() = ( )+ ()
()
()
() ()
C C
() = ∑ ()
=
C
() = ∑ ()
=
C C
∑ =
+
C = ()
∑ =
4. Correlation Properties
(δ δ τ )
ρ

∗
ρ (δ δ τ) = () ( + τ)
C C
= ∑ ρ (δ δ τ )
=
C
+ ∑ ρ (δ δ τ )
=
(·)∗ {·}
ρ (δ δ τ )

ρ (δ δ τ ) = ( ( ))∗ ( + τ)
= ρ (δ τ ) · ρ (δ τ)
α

π (α )τ
ρ (δ τ ) = (δ α ) α (α ) α
α
and
βR
max,c R

j2π f R ( β RR )τ
ρFcR (δR , τ ) = F
dkk R
( δR , β c R ) e c p βR ( β R R
c R ) dβ c R (35)
cR
βR
min,c R
are the transmit and the receive correlation functions, respectively, and
δT
j2π (l −l ) cos(αcTT −γT )
cllF (δT , αcTT ) = e λ (36)
δR
F R j2π (k−k ) cos( β RcR −γR )
dkk ( δR , β c R ) = e (37)
λ
f T (αcTT ) = T
f max cos(αcTT − φ T ) (38)
f R ( βR
cR ) = R
f max cos( β R R
c R − φ ). (39)
The distributions of the AoD αcTT and the AoA β R

cR
are denoted by pαT (αcTT ) and p βR ( β R
cR
),
cT cR
respectively.
In (32), the term ρM (δ , δ , τ ), which represents the 3D space-time CCF of the moving
kl,k l ,c M T R
cluster c M , can be expressed as

ρM M ∗ M
kl,k l ,c M ( δT , δR , τ ) : = E ( gkl,c M ( t )) gk l ,c M ( t )( t + τ )
αM
max,c M

= cllM (δT , αcMM ) dkk
M M
( δR , g ( α c M ))
αM
min,c M

j2π f TM (α MM )+ f RM ( g(α MM ))− f TS (α MM ) τ
e c c c
− j2π f RS ( g(αcMM ))τ

e pα M (αcMM ) dαcMM (40)
cM
where
δT
j2π (l −l ) cos(αcMM −γT )
cllM (δT , αcMM ) = e λ (41)
δR
M M j2π (k−k ) cos( g(αcMM )−γR )
dkk ( δR , g ( α c M )) = e (42)
λ
TM
f (αcMM ) = T
f max cos(αcMM − φ T ) (43)
f RM ( g(αcMM )) = R
f max cos( g(αcMM ) − φ R ) (44)
f TS (αcMM ) = S
f max cos(αcMM − φS ) (45)
f RS ( g(αcMM )) = S
f max cos( g(αcMM ) − φS ). (46)
The function g(·) in (40) expresses the exact relationship between the AoD αcMM and the AoA
β cMM . An expression for g(·) can be found in (Chelli & Pätzold, 2007).
The temporal ACF rzkl (τ ) of the channel gain zkl (t) is defined as rzkl (τ ) := E{z∗kl (t)zkl (t + τ )}
(Papoulis & Pillai, 2002). The temporal ACF rzkl (τ ) can be deduced from the 3D space-time
CCF ρkl,k l (δT , δR , τ ) by setting the antenna element spacings δT and δR to zero, i.e.,
rzkl (τ ) = ρkl,k l (0, 0, τ )

C T ,C R
= ∑ w2cT w2cR ρFkl,k l ,cT ,cR (0, 0, τ )
c T ,c R =1
CM
+ ∑ w2c M ρM
kl,k l ,c M (0, 0, τ ). (47)
c M =1
The 2D space CCF ρkl,k l (δT , δR ) is defined as ρkl,k l (δT , δR ) := E{z∗kl (t)zk l (t)}. Alternatively,
the 2D space CCF ρkl,k l (δT , δR ) can be derived from the 3D space-time CCF ρkl,k l (δT , δR , τ )
by setting τ to zero, i.e.,
ρkl,k l (δT , δR ) = ρkl,k l (δT , δR , 0)

C T ,C R
= ∑ w2cT w2cR ρFkl,k l ,cT ,cR (δT , δR , 0)
c T ,c R =1
CM
+ ∑ w2c M ρM
kl,k l ,c M ( δT , δR , 0). (48)
c M =1
5. Numerical Results
In this section, we confirm the validity of the analytical expressions presented in the previous
section by simulations making use of the sum-of-cisoids method. The simulation models for
moving and fixed scatterers are designed using the modified method of equal area (MMEA)
proposed in (Gutiérrez & Pätzold, 2007). In order to model all fixed clusters, 50 cisoids are
used for the simulation model. The same number of cisoids is used to model all moving clus-
ters. The propagation environment contains six moving clusters: three clusters are located on
the right side of the transmitter and the receiver, while the remaining clusters lie on the left
side. Each moving cluster has a length of 5 m and is separated by a distance of 45 m from its
neighbour clusters. The distance between the transmitter and the moving scatterers located
on the left and the right side is set to 3 m. For the fixed clusters, we consider a propagation
environment encompassing three clusters on each side of the transmitter. Each cluster has a
length of 2 m and is separated by a distance of 34 m from its neighbour clusters. The same
number of fixed clusters is considered around the receiver. The distance between the trans-
mitter and the fixed scatterers on the left and right side is set to 300 m. The distance between
the transmitter and the receiver is equal to 100 m. The transmitter and the receiver have a
speed of 50 km/h and equal angles of motion φ T = φ R = 0. The transmitter and the receiver
antenna tilts γT and γR are set to π/2. We consider the case of non-isotropic scattering con-
ditions. The AoDs αcTT and αcMM are uniformly distributed over the intervals [αmin,cT T
T , αmax,c T ]
M
and [αmin,c M R
M , αmax,c M ], respectively. The uniform distribution is also assumed for the AoAs β c R
R
over the interval [ β min,c R
R , β max,c R ].
We present some illustrative examples for the temporal ACF rM zkl ( τ ) of the moving clusters in
Fig. 3. We study the influence of the speed of the vehicles in the vicinity of the transmitter and
the receiver on the channel behaviour. The term vS in Fig. 3 denotes the speed of the vehicles
on the left and right side. From Fig. 3, we can notice that as the speed vS decreases, the coher-
ence time of the channel increases. It is well known that the coherence time indicates whether
we are facing a fast or a slow fading. As the speed of the vehicles relatively to the transmit-
ter and the receiver decreases the channel changes more slowly. A good fitting between the
simulation results and the theoretical results can be observed in Fig. 3. In Fig. 4, we show the
1
Reference model
Simulation model
(τ) |
0.8
Simulation
kl
Temporal ACF, | rzM
vS = 60 km/h
0.6
vS = 54 km/h
vS = 52 km/h
0.4
0.2
0
0 0.2 0.4 0.6 0.8 1
Time separation, τ (s)
Fig. 3. The absolute value of the temporal ACF rM zkl ( τ ) associated with moving scatterers for
various values of the velocity vS of the surrounding vehicles.
numerical results obtained for the 2D space CCF ρM 11,22 ( δT , δR ) caused by moving scatterers.
It could be seen from Fig. 4 that the cross-correlation function decreases as we increase the
antenna spacings δT and δR . However, the decay of the 2D space CCF ρM 11,22 ( δT , δR ) is faster
along the δR direction. Hence, a small antenna spacing at the receiver side guarantees a diver-
sity gain, but at the receiver side, we need a larger spacing to get non-correlated channels. In
11,22(δT , δR ) |
0.8
2D space CCF, | ρM
0.6
0.4
0.2
0
0
0.5 3
1 2
δR /λ 1.5 1
2 0
δT /λ
Fig. 4. The absolute value of the 2D space CCF ρM

11,22 ( δT , δR ) of the reference model caused by
moving scatterers.
Fig. 5, we illustrate the numerical results for the transmit correlation function ρFT (δT , τ ) of the
reference model resulting from fixed scatterers. The obtained results are confirmed by simu-
lation in Fig. 6. Similar results have been found for the receive correlation function ρFR (δR , τ )
associated with fixed scatterers since the same setting for the scatterers is considered around
the transmitter and the receiver. Limited space prevents us from including these results.
Transmit CF , | ρTF (δT , τ) |
0.8
0.6
0.4
0.2
0
0
0.5 0.1
0.08
1 0.06
δT /λ 1.5 0.04
0.02
2 0 τ
Fig. 5. The transmit correlation function ρFT (δT , τ ) of the reference model due to fixed scatter-
ers.
Transmit CF , | ρ̃TF (δT , τ) |
0.8
0.6
0.4
0.2
0
0
0.5 0.1
0.08
1 0.06
δT /λ 1.5 0.04
0.02
2 0 τ
Fig. 6. The transmit correlation function ρ̃FT (δT , τ ) of the simulation model related to fixed
scatterers.
6. Conclusion
In this chapter, we have presented a narrowband MIMO V2V channel model, where both the
impact of fixed and moving scatterers were taken into account. Double-bounce scattering is
assumed for the fixed scatterers, while single-bounce scattering is considered for the moving
scatterers. For reasons of brevity, we have restricted our investigations to non-line-of-sight sit-
uations. A reference model has been derived starting from the geometrical street model. The
statistical properties of the proposed channel model have been studied. We have provided
analytical expressions for the 3D space-time CCF, the temporal ACF, and the 2D space CCF.
Supported by our analysis, we are convinced that the effect of moving scatterers on the statis-
tics of V2V MIMO channels cannot be neglected. The investigation of the impact of moving
scatterers have revealed that as the speed of the vehicles in the vicinity of the transmitter and
the receiver decreases, the channel coherence time increases. The proposed channel model is
suitable for a highway environment under congestion conditions. In such conditions, the low
relative speed of the vehicles in the vicinity of the transmitter and the receiver allows us to
consider that the AoD and the AoA seen from moving clusters are non-time-variant during
a sufficiently large period of time. Actually, if the AoD and the AoA are time-variant, the
channel model becomes non-stationary. The latter aspect will be investigated in future work.
7. References
802.11p (2006). Wireless access in vehicular environment (WAVE) in standard 802.11 infor-
mation 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, Technical Re-
port p/D1.0, IEEE 802.11.
ASTM (2003). Standard specification for telecommunications and information exchange be-
tween roadside and vehicle systems - 5 ghz band dedicated short range communica-
tions (DSRC) medium access control (MAC) and physical layer (PHY) specifications.
ASTM E2213-03.
Chelli, A. & Pätzold, M. (2007). A MIMO mobile-to-mobile channel model derived from a
geometric street scattering model, Proc. 4th IEEE International Symposium on Wireless
Communication Systems, ISWCS’07, Trondheim, Norway, pp. 792–797.
Chelli, A. & Pätzold, M. (2008). A wideband multiple-cluster MIMO mobile-to-mobile channel
model based on the geometrical street model, Proc. 19th IEEE Int. Symp. on Personal,
Indoor and Mobile Radio Communications, IEEE PIMRC 2008, Cannes, France.
Gutiérrez, C. A. & Pätzold, M. (2007). Sum-of-sinusoids-based simulation of flat fading
wireless propagation channels under non-isotropic scattering conditions, Proc. 50th
IEEE Global Telecommunications Conference, GLOBECOM 2007, Washington, DC, USA,
pp. 3842–3846.
Millott, L. (1994). The impact of vehicular scattering on mobile communications, Proc. 44th
IEEE Veh. Technol. Conf., VTC’1994, Stockholm, Sweden, pp. 1733–1736.
Molisch, A., Kuchar, A., Laurila, J., Hugl, K. & Bonek, E. (1999). Efficient implementation of
a geometry-based directional model for mobile radio channels, Proc. 50th IEEE Veh.
Technol. Conf., VTC’1999-Fall, Amsterdam, Netherlands, pp. 1449–1453.
Papoulis, A. & Pillai, S. U. (2002). Probability, Random Variables and Stochastic Processes,
McGraw-Hill, New York.
Patel, C. S., Stüber, G. L. & Pratt, T. G. (2003). Simulation of Rayleigh faded mobile-to-mobile
communication channels, Proc. 58th IEEE Veh. Technol. Conf., VTC’03, Orlando, FL,
USA, pp. 163–167.
Pätzold, M., Hogstad, B. O., Youssef, N. & Kim, D. (2005). A MIMO mobile-to-mobile channel
model: Part I - The reference model, Proc. 16th IEEE Int. Symp. on Personal, Indoor and
Mobile Radio Communications, PIMRC 2005, Berlin, Germany, pp. 573–578.
Salo, J., El-Sallabi, H. & Vainikainen, P. (2006). Impact of double-Rayleigh fading on system
performance, Proc. 1st IEEE Int. Symp. on Wireless Pervasive Computing, ISWPC 2006,
Phuket, Thailand.
WAVE (2005). Wireless access in vehicular environments (WAVE) channel coordination. IEEE
P1609.4/D05.
Zajić, A. G., Stüber, G. L., Pratt, T. G. & Nguyen, S. (2008). Statistical modeling and experi-
mental verification of wideband MIMO mobile-to-mobile channels in highway envi-
ronments, Proc. 44th IEEE Veh. Technol. Conf., VTC’1994, Cannes, France.
Planar Antenna Array Hybrid Beamforming for SDMA in Millimeter Wave WPAN 65
Planar Antenna Array Hybrid Beamforming for

SDMA in Millimeter Wave WPAN
Sau-Hsuan Wu, Lin-Kai Chiu, Ko-Yen Lin and Ming-Chen Chiang
Institute of Communication Engineering
Department of Electrical Engineering
National Chiao Tung University
Hsinchu, Taiwan
1. Introduction
The increasing demands on bandwidth for personal and indoor wireless multimedia applica-
tions have driven the research and development for a new generation of broadband wireless
personal area network (WPAN) Alliance (n.d.); IEEE 802.11 WLAN Very High Throughput in
60GHz (n.d.); IEEE 802.15 WPAN Millimeter Wave Alternative PHY Task Group (TG3c) (n.d.);
WirelessHD (n.d.). This new WPAN is intended to support data rate up to 5Gbps or more and
allows for wireless interconnection among devices, such as laptops, camcorders, monitors,
DVD players and cable boxes, etc. Moreover, it could also serve as a wireless alternative to
the High-Definition Multimedia Interface (HDMI).
In addition to its very large bandwidth, the new WPAN also demands for short-range and
secure wireless connections. The characteristics of broad unlicensed bandwidth FCC (2004),
high penetration loss Smulders (1995; 2002) and significant oxygen absorption Anderson &
Rappaport (2004) at 60GHz radio make it an ideal wireless interface for the next generation
WPAN. Furthermore, the millimeter wavelength of 60GHz radio also makes it possible to use
tens of tiny antennas to steer radio signals with high directivity to the intended receivers. This
feature of high-directivity beam pattern not only improves the wireless link quality Balanis
& Ioannides (2007) but also increases the spatial reuse factor, allowing for multiple users to
gain access to the wireless channel at the same frequency and time. In view of the great
potential of 60GHz radio on WPAN and the advantages of beamforming (BF) for millimeter
wave (mmWave) applications, we study in this article a simple hybrid beamforming (HBF)
technique for spatial division multiple access (SDMA) using planar antenna arrays (PAAs).
Digital BF has been used to compensate the rather fixed radiation patterns of switch-beam
or beam-selection antennas to create more flexible hybrid beam patterns Celik et al. (2006);
Rezk et al. (2005); Zhang et al. (2003). In conjunction with the phase shifters of the element
antennas of PAAs, a hybrid type of BF is considered in Smolders & Kant (2000) which exploits
the advantage of BF both in the baseband and the radio-frequency (RF) ranges. Motivated by
the above results and taking into account the practical limitation and implementation cost of
the full digital BF, we study herein a special type of baseband and RF HBF for SDMA that
only requires four digital processing paths to support HBF on a 8 × 8 PAA illustrated in Fig.
1. The entire PAA of Fig. 1 is partitioned into four blocks of patch antennas. Each block is
driven by a digital BF weight, while each element patch antenna in a block is equipped with
Fig. 1. Antenna arrays of 8 × 8 planar antennas.
an individual phase shifter. With this configuration, we study the design of HBF for two-user
SDMA and compare its performance with that of RF BF only.
The content of this chapter is organized in the following order. First, some concept and the
configuration about PAA will be reviewed in Section II and applied to SDMA using recon-
figurable PAA. In Section III, the baseband-and-RF hybrid BF (HBF) will be introduced for
SDMA, and the linear constrain minimum power (LCMP) digital BF technique will be reap-
plied to this new setting of PAA HBF over mmWave radio. The signal to interference-plus-
noise ratios (SINRs) of users with the aforementioned BF scheme will be investigated in Sec-
tion IV and the BER simulations will also be conducted on an OFDM-based WPAN to verify
the performance of HBF for SDMA over mmWave radio. Conclusions will be drawn in Section
V.
2. The configurations of planar antenna arrays

We specify in this section the configurations of planar antenna arrays (PAA) for hybrid BF.
Sixty-four identical patch antennas are aligned to form an 8 × 8 antenna matrix as shown in
Fig. 1. Each element antenna is equipped with a phase shifter to maneuver the phase of the
signal radiating through it. Given the large number of antennas available for 60GHz applica-
tions, it is beneficial to use the antennas to serve multiple users in addition to increasing the
received signal to noise ratio (SNR) of a single user. Taking into account the practical limita-
tions of circuit implementations, the arrays of antennas are partitioned into blocks, and each
of which is driven by a baseband signal processing path. In other words, the baseband BF
weights are applied to the antennas on a block basis. Antennas within the same block are ap-
plied the same baseband BF weight. In this section, we characterize the beam pattern of this
hybrid type of radio-frequency (RF) and baseband (BB) BF.
To facilitate the analysis and highlight the performance of hybrid BF (HBF), the coupling ef-
fects among element antennas are neglected in the sequel. As a result, the total beam pattern
of a block of a partition can be expressed as the product of the electric field of a single antenna
and the array factor corresponding to the block Balanis (1997).
1.5
25
1
20
0.5
15
0 10
5 5
5
5 5 0 0.5 1 1.5
Fig. 2. The contour plot of the antenna pattern when P=30dB.
The far-zone electric field of a single element antenna is given by

a +E −
E(φ, θ ) = E −
→ →
a +E −
θ θ
→
a φ φ r r (1)
where

hWkE0 e− jkr sin Y sin Z
Eθ = j cos φ cos X (2)
πr Y Z

hWkE0 e− jkr sin Y sin Z
Eφ = j cos θ sin φ cos X (3)
πr Y Z
and Er ∼= 0 as r 2LW λ (see Balanis (1997) for the far field definition). The physical meaning
of some of the parameters are illustrated in Fig. 1, and E0 is a constant. For convenience of
2π
expression, we also define X kL kW kh
2 sin θ cos φ, Y 2 sin θ sin φ, Z 2 cos θ and k λ with
λ being the radio wavelength.
The contour plot of the electric field is shown in Fig. 2 as P| E(φ, θ )|2 in dB in the cylindrical
coordinate, with P = 30dB. The dimensions of the element patch antenna used in the simula-
tion are L = W = 1mm, h = 0.1mm and the distances between adjacent antennas are set to
dx = dy = 2.5mm. The radial coordinate is mapped to the elevation angle θ and the angular
coordinate is to the azimuth angle φ of the antenna pattern. The vertical coordinate displays
the antenna gain in decibel. We note that the antenna pattern is not symmetric with respect to
the azimuth angle, φ. The pattern is narrower in the direction of φ ± π/2.
The array factor of each block depends on its relative position in the PAA. For the partition
shown in Fig. 3, we define an index pair, (p, q) ∈ {0, 1}2 , for each block of the PAA. Given the
index pair (p, q) of a block, the corresponding array factor follows
N p
A( p,q) (φ, θ ) = e jpN (Ψx + β x,( p,q) ) ∑ e j(n−1)(Ψ +(−1) βx x,( p,q) )
n =1
M q
×e jqM(Ψy + β y,( p,q) ) ∑ e j(m−1)(Ψy +(−1) β y,( p,q) )
(4)
m =1
dy
dx
4dx
4dy
Fig. 3. Partitions of the planar antenna arrays. Patch antennas in different color belong to
different block.
where Ψ x kd x cos φ sin θ and Ψy kdy sin φ sin θ. The distances in the x and y directions
between adjacent patch antennas are denoted by d x and dy , respectively. And β x,( p,q) and
β y,( p,q) are the corresponding phase differences in the x and y directions between adjacent
patch antennas. The number of antennas in the x direction of a block is M and the number of
antennas in the y direction is N. Given the desired direction (φd,( p,q) , θd,( p,q) ) set for the block
( p, q), β x,( p,q) and β y,( p,q) are equal to
β x,( p,q) = (−1) p+1 kd x sin θd,( p,q) cos φd,( p,q) ± 2c1 π (5)
q +1
β y,( p,q) = (−1) kdy sin θd,( p,q) sin φd,( p,q) ± 2c2 π (6)
where c1 and c2 are any integers. It is noted that the array factor (4) is a periodic function with
a period of 2π and is zero whenever (φ, θ ) satisfy either one of the following two conditions
Ψ x + (−1) p β x,( p,q) = 2c3 π/N (7)

q
Ψy + (−1) β y,( p,q) = 2c4 π/M (8)
when c3 and c4 are integers not equal to the multiples of N and M, respectively.
2.1 SDMA using reconfigurable PAA

Adjusting the antenna phases of a block based on (4) ∼ (8), the main beams of different blocks
can be tuned towards the directions of different users, offering spatial division to support
multiple access (SDMA) of users. Despite the array gain provided by (4), the SINR of SDMA
also depends on the antenna pattern of (1) and the beam patterns of adjacent users.
Suppose that all patch antennas in Fig. 3 are used to support a single user. The maximum
.
achievable array gain in this case is 20 log10 (64) = 36 dB when
Ψ x + (−1) p β x,( p,q) = 2c5 π (9)

q
Ψy + (−1) β y,( p,q) = 2c6 π, {c5 , c6 } ∈ Z (10)
where Z stands for the integer number. Fig. 4 shows the contour plot of the corresponding
array pattern when θd,(0,0) = 0. It is clear that the maximum array gain of 36 dB is achieved
at θ = θd,(0,0) = 0. Scaling the power by 1/MN for each element antenna of the M × N PAA,
.
the maximum effective array gain is 10 log10 (64) = 18 dB.
3. SDMA using hybrid beamforming

The SDMA method introduced in Section 2.1 is based on phase adjustment with the phase
shifter of each element antenna. However, adjusting only the phases of the radio signals
sometimes may not be able to achieve the desired SINR for the user of interest, as the beam
direction of the user might be severely jammed by the side beams of other users. To overcome
this difficulty, baseband BF techniques can be used to jointly steer the beam patterns and
suppress the interference for all users. More specifically, in addition to steering the main
beam towards the direction of interest, the baseband array factor can be nulled as well in the
directions of other users’ main beams.
90 1 RF
120 60 Baseband
0.8
0.6
150 30
0.4
0.2
180 0
210 330
240 300
270
Fig. 6. The polar plot of the HBF pattern with the partition in Fig. 3 when θd = π/2.
However, it is impractical to apply a baseband BF weight for each element antenna of the
8 × 8 PAA. Taking into account the implementation cost, each partition of PAA is driven by a
common baseband BF weight, while each antenna is still equipped with an individual phase
shifter. To distinguish the array factor B(φ, θ ) formed with the baseband BF weights of a
user from the array factor A(φ, θ ) obtained by tuning the phase of the radiated wave of each
antenna, we refer to B(φ, θ ) as the baseband array factor (BAF) in contrast to the array factor
A(φ, θ ) tuned in the radio-frequency (RF) band.
Now we consider this hybrid type of baseband and RF BF for the simple partition shown in
Fig. 3. Suppose that the RF array factor (RAF) for different blocks of a user are the same and
pointed to the desired direction of interest, the composite beam pattern of HBF is given by
H (φ, θ ) B(φ, θ ) A(φ, θ ) E(φ, θ ) (11)
where A(φ, θ ) is the array factor of the 4 × 4 antenna arrays. In the extreme case of Fig. 3 that
the entire PAA is used to support a single user, the BAF is given by
1 1
B(φ, θ ) = ∑ ∑ w(r,s) e j4rΨ x
e j4sΨy . (12)
r =0 s =0
where Ψ x kd x cos φ sin θ and Ψy kdy sin φ sin θ. The enlarged distances between the ad-
jacent effective antennas make 4kd x = 4kdy = 4π in (12) as dx = dy = λ/2, which in turn
results in the periodic baseband beam pattern of B(φ, θ ) shown in Fig. 6. The angular coordi-
nate corresponds to the elevation angle θ and the radial coordinate represents the normalized
BF gain. Due to the periodic pattern, the product of B(φ, θ ) and A(φ, θ ) will yield significant
sidelobes on both sides of the main beam. For clarity, the RAF A(φ, θ ) of the 4 × 4 block is also
shown in Fig. 6. Since the patch antenna has a fixed radiation pattern, its pattern is not shown
in the figure.
2dy
2dx
dx
dy user1
One user Two user user1 user2
Fig. 7. The partition of the PAA for two-user SDMA, where patch antennas of the same color
belong to the same block.
3.1 HBF based on the MD beamforming

According to the configuration of Fig. 3, we now implement HBF for two-user SDMA based
on the partition in Fig. 7. The antennas in blue and green colors of Fig. 7 belong to user
one, and the antennas in orange and yellow belong to user two. Namely, two BF weights are
employed for each user. The resultant BAF for user one and two are given by
B1 (φ, θ ) = w(1,0) + w(1,1) e jΨx (13)
jΨy jΨ x + jΨy
B2 (φ, θ ) = w(2,0) e + w(2,1) e . (14)
Now to steer the main beam towards the direction of the user of interest and, in the mean
time, to suppress the interference in the direction of the other user, a typical method is the
so-called maximum directivity (MD) BF Kuhwald & Boche (1999).
The MD BF basically constructs the baseband BF weights by superposition of the steering
vectors
s1 (φ, θ ) [1 e jΨx ] T (15)
jΨy j(Ψ x +Ψy ) T
s2 (φ, θ ) [e e ] (16)
of user one and two in (13) and (14), respectively. Specifically, the BAFs are expressed as
2
B1 (φ, θ ) ∑ bi [ 1 e− jΨxi ]s1 . (17)
i =1
2
B2 (φ, θ ) ∑ ci [e− jΨ yi
e− j(Ψxi +Ψyi ) ] H s2 (18)
i =1
where Ψ xi 4kd x cos φi sin θi and Ψyi 4kdy sin φi sin θi , and {φi , θi } is the desired beam
direction of user i. Substituting the constraints of

1, i = m
Bm (φi , θi ) = , i, m ∈ {1, 2}. (19)
0, i = m
back into (17) and (18) yields the coefficients bi and ci . Furthermore, equating the Bm (φ, θ )
respectively for m ∈ {1, 2} in (13) ∼ (18) results in the baseband BF weights
2
w(1,r) = ∑ bi e− jrΨ xi
, (20)
i =1
2
w(2,r) = ∑ ci e− jΨ xi
e− jrΨyi r ∈ {0, 1}. (21)
i =1
3.2 HBF based on the linear constrained minimization of power

Though simple and straightforward, the MD BF does not take into account the power con-
sumption in HBF design. A widely used approach for power minimization is the linear con-
strained minimum power (LCMP) method Trees (2002). To minimize the power consumption
of BF and, in the mean time, null the interference in the beam direction of the user of interest,
we apply the LCMP subject to (s.t.) constraints similar to that of the MD BF in (19).
Let ui (t), i ∈ {1, 2} be the transmitted signal of user i, with E[|ui (t)|2 ] = 1. The baseband
transmitted signal for the two-user SDMA can be modeled as
x ( t ) = s1 u1 ( t ) + s2 u2 ( t ). (22)
where the steering vectors s1 and s2 are defined in (15) and (16), respectively. To design the
BF weight vector wi for user i such that the output power and the interference to the beam
direction of the other user are both minimized, the LCMP is formulated as
arg min wiH S x wi (23)
wi
s.t. wiH C = ei (24)
where S x E{x2 (t)}, C [si (φ1 , θ1 ), si (φ2 , θ2 )] with {φi , θi } being the desired beam direc-
tion of user i, and ei is a 1 × 2 basis vector with 1 in the ith position and the others zero.
The above optimization problem can be easily solved by making use of the Lagrange multi-
plier as below
J = wiH S x wi + wiH C − eiH λ + λ H C H wi − ei (25)
with the Lagrange multiplier λ [λ1 , λ2 ]T . Taking the complex gradient of J respect to wiH
and setting it to zero yields
1
S x w + Cλ = 0 =⇒ w = −S−
x Cλ. (26)
Substituting (26) into (24) gives
1 H −1
− λ H C H S− H H H
x C = ei =⇒ λ = − ei [ C S x C ]. (27)
Furthermore, substituting (27) back into (26), the resultant optimal BF weight vector for user i
is given by
1 −1 H −1
wiH = eiH [C H S−
x C] C S x . (28)
4. Computer Simulations
We demonstrate simulation results for the HBF schemes studied in the previous section for
SDMA. The transmit SNR in the following simulations is set to 30 dB for each user if no
specific description.
SINR SINR
1.5 1.5
25 25
1 1
20 20
0.5 0.5 (0, /4)
15 15
0 0
10 10
5 5
( , /4)
5 5
5 5
5 5 0 0.5 1 1.5 5 5 0 0.5 1 1.5
Fig. 8. The contour plots of SINR for user 1 and 2 using the RF BF. The left plot corresponds
to user 2 and the right plot to user 1, with the desired directions of user one and two set at
(φ1 , θ1 )=(0, π/4) and (φ2 , θ2 )=(π, π/4), respectively.
Fig. 8 presents the contour plots of the SINRs of user one and two when using the RF BF
method of (4). The contour plots are shown in the cylindrical coordinate. The radial coordi-
nate maps to the elevation angle θ and the angular coordinate maps to the azimuth angle φ.
The desired directions of user one and two are set at (φ1 , θ1 )=(0, π/4) and (φ2 , θ2 )=(π, π/4),
respectively. The resultant SINR in the desired direction of each user is equal to 16.18dB.
SINR SINR
1.5 1.5
22 22
20 20
1 1
( , /4) 18 18
(0, /4)
0.5 16 0.5 16
14 14
0 12 0 12
10 10
5 8 5 8
6 6
4 4
2 2
5 5
5 5 0 0.5 1 1.5 5 5 0 0.5 1 1.5
Fig. 9. The contour plots of SINR for user 1 and 2 using the HBF of LCMP. The left plot
corresponds to user 2 and the right plot to user 1, with the desired directions of user one and
two set at (φ1 , θ1 )=(0, π/4) and (φ2 , θ2 )=(π, π/4), respectively.
With the same simulation setting for RF BF, the contour plots of SINRs with HBF of LCMP are
shown in Fig. 9. In comparison with the results in Fig. 8, we can see that the main beam here
for each user is much more stronger and narrower, while the side beams are relative fewer and
weaker. In addition, the resultant SINR in the desired direction of each user is now increased
to 24.09dB in Fig. 9 as oppose to the 16.18dB in Fig. 8 for the RF BF. This demonstrates that
the interference can be suppressed effectively in the desired directions of users with the HBF
of LCMP method. Besides, the results of HBF with MD BF are also similar to those of LCMP
and, hence, are not shown here.
In addition to SINR, directivity is also an important performance measure to characterize the
effectiveness of BF. To reflect the interference due to the multiple access in SDMA, the original
definition for directivity in Balanis (1997) is modified into
4πSINRi (φi , θi )
Di = π . (29)
2π
0 0
2
SINRi (φ, θ ) sin(θ )dθdφ
This new definition for directivity automatically refers to the traditional notion of directivity
in the single-user system.
According to this definition of directivity, Table 1 summarizes the results of SINRs, directiv-
ities and radiation powers for the RF BF and HBF schemes of MD and LCMP, respectively,
when the desired directions of users are set at (φ1 , θ1 ) = (0, π/4) and (φ2 , θ2 ) = (π, π/4).
The radiation power for each user is evaluated with
2π π
2
| Hi (φ, θ )|2 sin(θ )dθdφ (30)
0 0
and is denoted by PT in the table.

It is clear from the table that HBF not only achieves higher SINRs and directivities in this case,
but also uses less power for BF. This demonstrates its great potential for SDMA in mmWave
applications.
User1 User1 (0, /4) User2 ( , /4)

User2
( , ) RF MD LCMP RF MD LCMP
SINR (dB) 16.18 24.09 24.09 16.18 24.09 24.09
Directivity 13.33 156.3 156.3 13.55 159.4 159.4
PT 0.1388 0.0775 0.0775 0.1358 0.0764 0.0764
Table 1. The SINRs, directivities and the radiation power of the RF BF and the HBF schemes
of MD and LCMP.
In addition to typical measures for the evaluation of a BF scheme, we also investigate the
effectiveness of HBF from the perspective of wireless communications systems. To this end,
we study the performance of a two-user SDMA with PAA in orthogonal frequency division
multiplexing (OFDM) wireless communications systems. The OFDM system simulates the
physical layer proposal of IEEE802.11 task group AD for very high throughput in 60GHz
Channel Models for 60 GHz WLAN Systems (n.d.). Specifically, we consider an indoor simu-
lation setting as shown in Fig. 10 for a two-user SDMA system operating in the 60GHz wire-
less channel model proposed for IEEE802.11ad Channel Models for 60 GHz WLAN Systems
(n.d.). The system bandwidth is assumed to be 1GHz and the length of fast fourier transform
(FFT) is 1024.
Fig. 10. The indoor simulation setting for a two-user SDMA system.
The simulated transmitter is placed at the position of (0.5, 3) from the lower left corner of the
room and transmits signals simultaneously to the two users locating at (2.5, 1) and (5.5, 8),
respectively, using either the RF BF setting shown in Fig. 8 or the HBF setting of LCMP in Fig.
9. The users’ data are first modulated with QPSK and then transformed with the 1024-point
inverse fast fourier transform (IFFT). The transformed OFDM symbol is added with a cyclic
prefix of 1/8 of the OFDM symbol and then transmitted with the corresponding BF techniques
to the intended users simultaneously. The transmitted signals propagate through the 60GHz
channel modeled according to Channel Models for 60 GHz WLAN Systems (n.d.) to come to
the two receivers of Fig. 10. At the receivers, we assume that the same RF BF with a 8 × 4
PAA is used by each receiver to enhance the signal qualities, supposing that the beam pattern
of each user is pointed to its desired angle of arrival.
0
10
10
10
BER(raw data)
10
10
User 2 with RF BF
10 User 2 with HBF of LCMP
User 1 with RF BF
User 1 with HBF of LCMP
10
0 5 10
Tx SINR(in dB)
Fig. 11. The BERs for a two-user SDMA system in an indoor environment of Fig. 10.
To recover the data, the receivers first remove the cyclic prefixes of each and then perform FFT
followed by channel equalization for symbol detection. The bit error rate (BER) for each user
is shown in Fig. 11 with respect to the transmitted SINR. As can be seen from the figure, the
SNR advantages for users using the HBF of LCMP are around 10dB against the users using
the RF BF only, even though the HBF gain of LCMP is only 7.91dB higher than that of RF BF.
5. Conclusions
We presented HBF techniques for SDMA in 60GHz radio using reconfigurable PAA. Accord-
ing to our simulation studies, the BF gain of HBF can be as much as 7.9dB higher than that
of RF BF. Furthermore, the BER of using HBF in a simulated OFDM environment can have
an even higher SNR advantage of 10dB against that of using the RF BF only. These results
demonstrate the potential of HBF with PAA for SDMA in mmWave applications.
6. References
Alliance, W. (n.d.). http://www.wimedia.org/.
Anderson, C. R. & Rappaport, T. S. (2004). In-building wideband partition loss measurements
at 2.5 and 60 GHz, IEEE Trans. on Communications 3(3): 922–928.
Balanis, C. A. (1997). Antenna Theory, 2 edn, John Wiley & Sons.
Balanis, C. A. & Ioannides, P. I. (2007). Introduction to Smart Antennas, Morgan and Claypool.
Celik, N., Kim, W., Demirkol, M., Iskander, M. & Emrick, R. (2006). Implementation and exper-
imental verification of hybrid smart-antenna beamforming algorithm, IEEE Antennas
and Wireless Propagation Letters 5: 280–283.
Channel Models for 60 GHz WLAN Systems (n.d.). 11-09-0334-02-
00ad-channel-models-for-60-ghz-wlan-systems.doc. available at
http://www.ieee802.org/11/Reports/tgad_update.htm.
FCC (2004). Code of federal regulation, title 47 telecommunication, chapter 1, part 15.255.
IEEE 802.11 WLAN Very High Throughput in 60GHz (n.d.). available at
http://www.ieee802.org/11/.
IEEE 802.15 WPAN Millimeter Wave Alternative PHY Task Group (TG3c) (n.d.). available at
http://www.ieee802.org/15/pub/TG3c.html.
Kuhwald, T. & Boche, H. (1999). A constrained beam forming algorithm for 2D planar antenna
arrays, Proc. IEEE VTC-Fall, Amsterdam, The Netherlands.
Rezk, M., Kim, W., Yun, Z. & Iskander, M. (2005). Performance comparison of a novel hybrid
smart antenna system versus the fully adaptive and switched beam antenna arrays,
IEEE Antennas and Wireless Propagation Letters 4: 285–288.
Smolders, A. B. & Kant, G. W. (2000). THousand Element Array (THEA), Proc. IEEE Antennas
and Propagation Society International Symposium, Salt Lake City, UT.
Smulders, P. (1995). Broadband wireless LANs: a feasibility study, Ph. D. thesis,
Eindhoven, Univ. of Tech., The Netherlands, ISBN 90-386-0100-X . available at
http://alexandria.tue.nl/extra3/proefschrift/PRF11B/9505571.pdf.
Smulders, P. (2002). Exploiting the 60 GHZ band for local wireless multimedia access:
prospects and future directions, IEEE Communications Magazine 2(1): 140–147.
Trees, H. L. V. (2002). Optimum array processing. Part. IV of detection, estimation and modulation
theory., John Wiley & Sons.
WirelessHD (n.d.). http://www.wirelesshd.org/index.html.
Zhang, Z., Iskander, M., Yun, Z. & Host-Madsen, A. (2003). Hybrid smart antenna system
using directional elements - performance analysis in flat Rayleigh fading, IEEE Trans.
on Antennas and Propagation 51(10): 2926–2935.
A Distributed Multilayer Software Architecture for MIMO Testbeds 77
A Distributed Multilayer Software

Architecture for MIMO Testbeds
José A. Garcı́a-Naya, M. González-López and L. Castedo
Universidade da Coruña
Spain
1. Introduction
The use of multiple antennas at both transmission and reception, also known as Multiple In-
put Multiple Output (MIMO) transmission systems, has received a lot of interest from the
wireless communications industry during the last years. Communications in wireless chan-
nels using MIMO technologies exhibits a superior performance in terms of spectral efficiency,
reliability and data rate when compared to conventional single antenna technologies (Foschini
& Gans, 1998; Telatar, 1999). Existing and emerging standards for wireless communications
such as IEEE 802.11 (WiFI), IEEE 802.16 (WiMAX) and Long Term Evolution (LTE), support
multi-antenna transmission in their highest performance profiles.
In spite of their potential performance-enhancing capabilities, most of the research on MIMO
technologies up to the moment is based on theoretical studies. Typically, the expected gains
of MIMO technologies are only shown under ideal conditions since most analysis rely on
simulations. Experiments in real-world scenarios by means of hardware implementations are
necessary to measure the actual performance of multi-antenna transmission methods. Hard-
ware implementations not only take into account the real multipath propagation in wireless
channels but also the implementation impairments so often ignored during the simulations.
Hardware implementations can be split into three groups (Rupp et al., 2006). The first one is
constituted by demonstrators which are frequently designed having in mind a particular stan-
dard or specification. Demonstrators usually exhibit good technical features for real-time im-
plementations but they are extremely expensive and present poor flexibility and modularity.
The second group is formed by prototypes of a final product. A prototype is a real-time imple-
mentation of a system specifically developed to support an industrial need. Prototypes often
constitute a preliminary stage where the system is implemented and debugged and later on
implemented as a consumer product. Finally, the third set is formed by testbeds that support
real-time transmission capabilities while data is generated and post-processed off-line.
In addition, hybrid solutions can be devised. As an example, a testbed can carry out some op-
erations in real-time with the purpose of speeding up the measurement process. Usually, can-
didate signal processing operations to be implemented in real-time are those that operate at
sample level and/or do common tasks for all experiments, i.e. I/Q modulation, up-sampling,
pulse-shaped filtering, etc. Throughout this chapter we will focus on testbeds because they
use open designs and are more often found in public research centres and academia.
Various MIMO testbeds have been reported in the literature (Borkowski et al., 2006; Caban
et al., 2006; Fabregas et al., 2006; Haustein et al., 2006; Nieto et al., 2006; Ramı́rez et al., 2008;
Rao et al., 2004; Wilzeck et al., 2006; Zhu & Fitz, 2005). Some of them have been constructed
to evaluate a particular standard or specification while others have been designed for general
purpose. Flexibility, development time consumption, throughput or costs are important fea-
tures when comparing existing testbeds. Also, it should be noticed the educational possibili-
ties of testbeds that open the door to many teaching opportunities. In the literature, however,
there is a lack of up-to-date guides and tutorials useful for the construction of a new testbed
from the scratch. Indeed, there exists few contributions (Caban et al., 2006; Garcı́a-Naya et al.,
2008a; Rao et al., 2004; Rupp et al., 2007) that contain a detailed description of the constructed
MIMO testbed and, except for (Garcı́a-Naya et al., 2008a; Rupp et al., 2007), the information
contained on them is already outdated.
Based on the previous experience acquired by the research group of the authors in building
and setting up MIMO testbeds, as well as performing indoor measurements (Pérez-Iglesias
et al., 2008; Ramı́rez et al., 2008), we can assert that once the testbed hardware is available
and properly configured, accessing the testbed becomes the main and also frequently ignored
issue. When a research team decides to start the process of acquiring and/or constructing a
new testbed, they need to take into account numerous aspects related both with the hardware
and its technical features, and the extensibility possibilities for the future (Garcı́a-Naya et
al., 2008a; Rupp et al., 2007). Usually, most of the efforts are devoted to the testbed setup,
which results in equipment that is hardly usable by people not involved in its design and
later configuration. This makes extremely difficult to access to the testbed.
As a result, the migration of an algorithm from a simulation environment to a testbed involves
cumbersome low-level programming to access the hardware as well as a very detailed knowl-
edge of the hardware. Additionally, hardware implementation problems frequently ignored
by simulations arise, such as time and frequency synchronization, I/Q imbalances, non lin-
ear distortions caused by the power amplifiers, etc. All these issues make difficult to assess
new MIMO transmission methods in a testbed. For this reason, it is desirable to make the
testbed accessible to final users at a reasonable abstraction level. This goal represents an im-
portant challenge due to the large amount of heterogeneous technologies and development
environments that have to be integrated together. However, if this challenge is accomplished,
the final result is a very attractive product for the user, who can focus on the development of
new transmission techniques that can be easily translated to the testbed and later evaluated
in realistic scenarios.
In this chapter we describe how to solve all previously mentioned limitations by using a dis-
tributed multilayer software architecture. This architecture enables the testbed to be easily
accessible by the researchers and to be integrated in the development environment they are
using. Although all designs and results herein presented are particularized for our MIMO
testbed (Garcı́a-Naya et al., 2008b; Ramı́rez et al., 2008), they are easily adaptable to most of
the existing testbeds, making even possible the integration of heterogeneous testbed nodes in
order to build a multi-terminal testbed.
The proposed software architecture consists of three different layers:
1. The middleware layer (MWL) is the lowest-level layer that interacts with the testbed
hardware. It makes the testbed accessible through standard TCP socket connections.
2. The signal processing layer (SPL) performs the necessary operations required to con-
vert the discrete-time sequences provided by the final user into discrete-time signals
suitable to be transmitted by the hardware. At the receiver, it is usual to perform signal
processing operations like time and frequency synchronization in case they are not car-
ried out by the testbed hardware. The tasks compounding the signal processing layer
can also be executed in real-time by the testbed hardware. For example, digital up and
down converters are frequently available in the libraries of programmable hardware
modules.
3. The testbed interface layer (TIL) is the highest-level layer and presents the testbed to
the user at an adequate abstraction level. The TIL has to be designed and implemented
for a specific development environment. For example, if the final user makes use of
Matlab, then a specific implementation of the TIL for Matlab has to be developed. The
main purpose of the TIL is to provide a simplified interface to access the testbed. This
does not prevent from providing mechanisms to control the hardware in detail from
the TIL. It is very important to emphasize that there is no logic in the TIL except that
necessary for adapting the data format from the specific environment used by the SPL
and vice versa. The only requirement for the TIL to be implemented is the availability
of a standard TCP socket library.
The whole design and implementation of the proposed software architecture is done under
the following premises:
1. Layers have to be as decoupled as possible. This is a fundamental idea in modern
software design and structured programming. The key idea is not to replicate function-
alities that are already present in another layer, which would be a symptom of a bad
design. The basic premise is to design everything to be fully decoupled and reusable;
and later on introduce some small violations of this principle only if they are strictly
needed to increase the overall system performance.
2. Each layer can be extended and/or customized to be able to adapt them to future spec-
ifications and/or heterogeneous hardware environments. It is important that this layer
upgrade be done without needing a complete remake of the software, which is fre-
quently the only solution in case of monolithic non-layered systems.
3. Finally, layers should be distributed, which means that they use remote connections
to interact among them. On the one hand, this helps to ensure the decoupling and
independence principles whereas, on the other hand, allows the testbed to be remotely
accessible from the user Personal Computer (PC).
The remaining of this chapter is structured as follows. Section 2 and Section 3 provide a
global description of the testbed hardware and software possibilities, respectively. Section 4
describes the software technologies that were used to develop the proposed distributed multi-
layer software architecture. In Section 5 the software architecture is presented and described.
The interaction among the different layers and components of the architecture is studied in
Section 6 and Section 7. Finally, Section 8 is devoted to the conclusions.
2. Basic Testbed Hardware

Testbeds are often used to verify if a new signalling technique (or even a complete standard
system) that has been proven useful by simulations is also valid in realistic wireless scenar-
ios. Testbeds allow the dimensioning of the hardware needs for real-time implementations,
not only for the digital signal processing modules (mainly, DSP and FPGA) but also for the
analogue Radio Frequency (RF) front-ends. Testbeds allow capturing the requisites for the
hardware used in the final real-time implementation. Consequently, evaluating performance
with testbeds allows knowing whether a given technique is feasible from the real-time hard-
ware and implementation requirements.
Fig. 1. Block diagram of the basic hardware configuration of a testbed.
Among the three different types of hardware implementations described above (demonstra-
tors, prototypes and testbeds), testbeds present the following advantages:
• Flexibility. Testbed hardware is meant to be used for off-line processing. Only the
signals are sent and acquired in real-time. This implies that testbed hardware is not
subject to the real-time restrictions even though the hardware can include some sort of
real-time capabilities.
• Modularity. Usually, the minimum modularity found in a testbed is given by the sep-
aration between the digital hardware (up to the D/A and A/D converters) and the RF
hardware. Sometimes, the digital hardware can be split in different modules perform-
ing specific operations: D/A and A/D conversion, digital up and down conversion,
signal buffering, etc. For the RF section it is possible to find self-made solutions or
commercial products. Lastly, commercial RF front-ends also permit some degree of
flexibility, for example allowing dual-band operation, RF carrier selection for each band
or adjusting the gains at the transmitter and receiver amplifiers.
• High-level language development. Having in mind that most of the processing tasks
are carried out off-line, general-purpose processors are the most adequate for process-
ing the generated/acquired signals. This allows the utilization of high level program-
ming environments (e.g. Matlab, C/C++, Java) and the simplification of the imple-
mentation stages both in time and complexity. This feature also provides an additional
flexibility degree, because it is easy to change the implementation on the fly.
• Floating-point versus fixed-point precision. As a consequence of the off-line process-
ing and the usage of high-level programming environments, the operations are carried
out in the floating-point domain instead of using fixed-point operations available for
real-time devices. This permits the researcher to focus on the implementation rather
than considering some other problems like arithmetic precision of the operations.
Testbed hardware components can be classified into three groups according to their function-
ality, as shown in Fig. 1. The first one is the host system, usually a PC and consequently
referred to as host PC. It is the equipment that allocates one or more boards containing the
digital section of the hardware testbed (including D/A and A/D converters). The second
group is constituted by the digital hardware components and, finally, the third one is formed
by the RF front-ends. The D/A and A/D converters generally constitute the frontier between
the digital and analogue hardware. In the digital section, a bus termed main bus allows trans-
ferring the data from the host to the testbed hardware and the other way around. Next, a set
or just one digital bus interconnects the digital hardware (DSPs, FPGA, memory buffers, etc.)
with the D/A and A/D. Finally, for the interconnection among the D/A and A/D converters
and the RF front-ends coaxial cables are used.
With this basic configuration, it is possible to send samples directly coming from the main bus,
convert them into the analog domain using the D/A converters and up convert them to the
desired carrier RF using the front-ends. At the receiver side, the signals are down converted
by the RF front-ends, digitally converted by the A/D converters and then sent to the host
through the main bus. Note that the main bottleneck in this scheme is the maximum data rate
provided by the main bus, especially if there are no digital up and down converters available.
In the most recent boards, by using PCI express or similar solutions is possible to use the host
memory as a buffer while samples are transferred in real-time through the main bus.
A first improvement of this basic scheme consists in incorporating a digital up converter
(DUC) before each D/A converter at the transmitter and a digital down converter (DDC) after
each A/D converter at the receiver. These devices can be dedicated elements or can be imple-
mented in a FPGA. Incorporating DUC and DDC into the MIMO testbed design allows trans-
ferring complex signals from/to the host, reducing both the transfer rate at the main bus and
the software complexity. For MIMO operation, DDCs and DUCs must be fully synchronized.
Additionally, another improvement consists in implementing some of the time-consuming
sample-level tasks in FPGAs (e.g. time and frequency synchronization).
2.1 Baseband Components

Baseband components are the hardware elements necessary to deal with baseband and/or
Intermediate Frequency (IF) signals. Frequently, such baseband components are allocated on
carrier boards installed on conventional PCs. Current testbeds use carrier boards equipped
with standardized buses to access testbed hardware such as USB, PCI, cPCI, PCI express or
similar. A manufacturer compliant with the PXI alliance (PXI, 2009) produces carrier elements
that are compatible with the most typical buses present in host equipments. When available,
important hardware components in a testbed are the storage buffers, especially when the main
bus does not support the necessary rate demanded by the D/A and A/D converters. Such
buffers allow performing signal operations off-line while data is sent and acquired in real-
time.
The interconnection among the previously described elements is carried out using buses that
must be capable of transmitting the data fast enough. Finally, external circuitry such as clock
distribution and/or triggering is needed. Sometimes, the most advanced hardware manufac-
turers include such circuitry as part of the commercial boards, simplifying the later setup.
When MIMO processing is required, fully synchronization among the different devices is
mandatory. Sometimes manufacturers announce a MIMO system but just a scaled SISO solu-
tion is offered.
2.2 RF Front-Ends
RF front-ends constitute one of the major hindrances in the testbed building process. They are
responsible of up converting IF or baseband signals to RF. The most common RF bands are
the unlicensed ISM located at 2.4 and 5.8 GHz. It is easier to find components for such bands
but also unpredictable interference is present in such spectrum portions. High linearity and
flexibility in terms of supported carried frequencies and bandwidths are desirable features for
the RF front-ends. However, the most advanced front-ends commercially available are only
dual-band and have fixed maximum bandwidth. A fundamental feature needed to be able to
carry out experiments is the possibility of modifying the transmit power. Also, the majority
of the front-ends are designed for SISO operation, making difficult to adapt them to a MIMO
system. Moreover, once the front-end has been acquired, extensive test on it is needed and
expensive measurement equipment is required.
It is important to emphasize that RF hardware is expensive compared to the cost of the digital
hardware and does not follow Moore’s law (Moore, 1998).
3. Basic Testbed Software

Methodologies that cover the entire development process, from source code suitable for sim-
ulations and executed off-line, to real-time implementation in the testbed, as well as software
tools according to these methodologies, are extremely scarce. The most popular methodology
for testbed development is rapid prototyping (Kaiser et al., 2004; Rupp et al., 2003; 2006). A
typical approach used to develop real-time algorithms to be run in a testbed consists in start-
ing with a simulation implementation. Next, the simulation is migrated to the testbed and,
finally, a real-time implementation is obtained. It is also convenient to split the real-time im-
plementation into several steps: firstly, a fixed-point code is produced; next a DSP implemen-
tation is obtained; and, afterwards, the software modules that do not meet time requirements
are migrated to FPGA, making use of high level tools when possible. Nowadays, except for
the Mathworks tool suite (Mathworks, 2009) combined with some VHDL code generators,
there is a lack of high level tools and development environments suitable to fit the previous
steps.
Also, some standardisation efforts have just started in order to make compatible hardware
components and software modules from different manufactures and developers. The PXI
alliance (PXI, 2009) plus the Software Defined Radio (SDR) Forum (SDR Forum, 2009) and
initiatives such as the Software Communications Architecture (SCA) (SCA, 2009) constitute
the starting point of a new generation of modular, flexible and standard radio interfaces suit-
able for research. Nowadays, however, the previously mentioned initiatives have not already
produced as a result the availability of high level tools allowing final users and researchers
not involved in the hardware development to access the testbed hardware at a reasonable
abstraction level.
All above reasons serve as a motivation for our work, which aims at bridging the existing
gap among the hardware, the software elements provided by the manufacturers and the ab-
straction level required by the final users. Moreover, the proposed solution allows embedding
real-time modules in the processing chain, as if they were part of the testbed hardware. For
example, real-time frequency and time synchronization algorithms can be run at the receiver
side and integrated into the digital hardware. Additionally, given that the proposed solu-
tion is designed to be easily and quickly integrated in any other kind of system, it becomes a
very useful tool to help in testing real-time applications, especially during the first stages of
the development. A good example is the ability to develop in parallel different parts of the
transmitter and the receiver in real-time. While the real-time transmitter is still under devel-
opment, the team developing the real-time receiver can make use of the testbed to generate
the transmit signals and feed the real-time receiver with them.
4. Software Technology Behind the Distributed Multilayer Architecture

The field of computer science has come across problems associated with complexity since its
constitution. The software architecture discipline is centred on the idea of reducing complex-
ity through abstraction and separation. There are different kinds of software architectures.
Among them, the most interesting for our work are: the client-server architecture, which
serves as the basis for most of the available architectures; the three-tier model; and finally,
the most recent and advanced model-view-controller architecture.
4.1 Client-Server Software Architecture (Two-Tier Software Architecture)

Client-server describes the relationship between two computer processes in which one process
(the client) makes a service request to another process (the server). Applications following the
client-server architecture represent an evolution with respect to those called ”monolithic”.
The basic operation mode for client-server applications consists in that each instance of the
client process sends requests to one or more connected servers. In turn, servers accept these
requests, process them, and return the requested information to the client. The basic client-
server architecture considers only two types of hosts: clients and servers. Consequently, it
is also known as two-tier model. A special case of the two-tier software architecture arises
when an instance simultaneously acts as a client or as a server, resulting in the peer-to-peer
architecture.
The two-tier software architecture presents the following advantages with respect to mono-
lithic systems:
• The responsibilities of the whole system are now split between the client and the server,
which brings the opportunity to decouple them.
• Servers can control the access to the resources to guarantee that only those clients with
appropriate permissions access and change the data.
• Different client/server types can work with different kinds of servers/clients, which
helps in the integration of heterogeneous systems.
As a main disadvantage, the mechanism used to interconnect clients and servers (frequently
standard socket connections) may become both a bottleneck and a weak point of the system,
because a failure in the interconnection mechanism will stop the whole system.
4.2 Multi-Tier Software Architecture

The multi-tier software architecture represents a natural evolution of the client-server model
towards a higher number of levels. That is the reason why it is also known as the n-tier
software architecture. Basically, the multi-tier software architecture is a client-server architecture
consisting of more than two tiers, and where the inner tiers act simultaneously as client for one
tier and as server for the other. The main difference with respect to the peer-to-peer model is
that the tier order (its position or level in the multi-tier structure) does matter. An inner layer
is also termed middleware because acts as an intermediary between two adjacent tiers.
Note that the multi-tier software architecture, as well as the two-tier one, is a mechanism
to design the physical structure of the system. Given one of the models, several degrees of
freedom are still possible to get the logic structure of the system.
4.3 Tier Interconnection Mechanisms

Up to now, we were using a certain abstraction level to define the previous architectures.
Usually, each tier is mapped to a set of processes obtained from the same master process. This
can be easily understood with an example. Imagine a two-tier system. One tier is the client
while the other is the server and they are mapped into two different processes interconnected
by means of sockets. It does not matter if both processes are in the same physical machine
or not. What is important is how they interact with each other. The server will be waiting
for a request from one of the clients. As soon as the request arrives, the server starts to serve
the petition and, simultaneously, starts waiting for another request. When the server process
is going to serve a new request, it clones itself to serve the petition. Frequently, this process
cloning mechanism is omitted and it is considered as an implementation detail. Therefore,
to describe the interaction between different tiers, it is assumed that each tier is mapped into
a process that interchanges messages with other processes (tiers). Although this approach is
not completely rigorous, it is enough for our purposes. In the literature such interactions are
described my means of sequence diagrams as we will do.
But there is still one open question: How are messages sent from one tier to another? In
principle, data transfer between two tiers is part of the architecture design, but in practice
only the message types and the data are defined by using sequence diagrams, which drives to
the definition of a part of the system termed protocol. A protocol defines the interaction among
all tiers, specifying the type of messages to be sent from one tier to another or to several other
tiers, as well as the type of data sent in each message. Messages are transferred using plain
socket connections or more elaborated mechanisms (e.g. SNMP, CORBA, Java RMI or even
Web Services). The different approaches have their own advantages and disadvantages and
there is still no perfect system for message interchange. However, what is really important is
to use standardized mechanisms for message interchanging, thus guaranteeing independence
among tiers.
4.4 The Model-View-Controller Architectural Pattern

The Model-View-Controller (MVC) is a pattern used in software engineering derived from
the multi-tier software architecture. Successful use of the pattern drives to the isolation of the
business logic from the user interface, allowing one to be freely modified without affecting the
other. The controller collects user inputs, the model manipulates (and usually stores) applica-
tion data, and the view presents results to the user. The MVC was first described in (Trygve,
1978) and can be used as an architectural or design pattern. As an architectural pattern, it
splits an application into three independent layers that can be run in different computers: the
presentation or user interface layer, termed the view; the business logic, called the model; and
the controller, an intermediate layer adapting the view to the model and vice versa.
4.5 Applying Software Engineering to MIMO Testbeds

In the first sections of this chapter we present the typical hardware architecture available in
MIMO testbeds as well as the typical abstraction level offered by the software included with
the hardware. We stressed that such abstraction level is by far not enough and, consequently,
a final user or a researcher not involved in the testbed development and later setup cannot
directly access it. Along this Section, we introduced architectural software models, starting
with the well-known client-server model and ending with the recently proposed model-view-
controller, widely used in web applications as well as distributed applications. These architec-
tural models serve as an inspiration to propose a novel software architecture useful to bridge
the existing gap between the abstraction level offered by testbeds and the one demanded by
researchers and final users. This new architecture is explained in the following Section.
Fig. 2. General scheme of the MIMO Testbed showing the three different layers: middleware
(MWL), signal processing (SPL) and testbed interface (TIL). The corresponding process name
for each layer and each PC is shown between brackets.
5. Distributed Multilayer Software Architecture for MIMO Testbeds

Fig. 2 shows the proposed testbed hardware organization, which can be extrapolated to most
of the testbeds. It is constituted by two ordinary PCs hosting the testbed hardware, one for
the transmitter (referred to as TxPC) and other for the receiver (named RxPC). The digital
section of the testbed hardware is installed inside the PCs while the RF front-ends (RF-FE)
remain outside the PC case. The two PCs are attached to the network, something that is very
useful because this way the PC desktops can be remotely accessed. The remote user PC is also
attached to the network. In the rest of the Section we are assuming the following:
1. The testbed consists of the transmitter and the receiver, i.e., two nodes. Although our
designs can be easily extended to an arbitrary number of nodes, it is better to start with
the simplest case of two nodes to keep things simple.
2. It does not matter whether the testbed operates outdoor, indoor, outdoor to indoor or
vice versa. We assume that a network connection can always be established among the
testbed nodes and the user PCs.
3. The testbed PCs will use a standard operating system supporting remote desktop or
remote operation from a PC attached to the network. This is a fundamental feature be-
cause it enormously simplifies software deployment as well as day-to-day maintenance
of the systems. Otherwise, a replacement for such tools should be provided.
5.1 From the MIMO Testbed to the Multilayer Software Architecture

Fig. 2 shows, close to the PC drawings, the names of the corresponding three layers of the
proposed architecture that are, from the lowest to the highest level: the middleware layer
(MWL), the signal processing layer (SPL) and the testbed interface layer (TIL). The MWL and
the SPL are split into transmit and receive parts, while the TIL has just one instance running
on the user PC. Actually, the TIL is just an Application Program Interface (API) that has to be
included with the user application. In our standard deployment, the MWL and the SPL are
allocated in the same PCs containing the rest of the testbed hardware. Although the MWL
is the only layer required to be installed in the hardware PCs, the rest of the software can be
deployed in any machine attached to the network.
This is not an arbitrary proposal but a design inspired on the multi-tier and the model-view-
controller software architectures. First of all, let us identify the use cases of our application.
Basically, there is just one use case or action carried out by the final user: transmitting a set
of discrete-time sequences through the testbed and, consequently, through the wireless chan-
nel. Therefore, the layer corresponding to the view is the interface that allows accessing the
testbed, sending the sequences and getting back the acquired signals. This is what we term
testbed interface layer, and is the topmost layer of our software architecture.
The controller plays a very important role in a system. It is responsible of getting the service
requests from the view, adapting and sending them to the model where the business logic
resides. The gathered results are then sent back to the view to be presented to the user. The
controller plays the role of the middleware concept presented in the multi-tier software archi-
tecture. In the testbed system the controller is called middleware and its functionality is clear:
to configure and control the hardware in order to be able to take the requests (discrete-time
sequences) from the TIL; to adapt and send them to the hardware to be transmitted by the
antennas; and, finally, to carry out the reciprocal operations at the receiver side. At the end of
the process the TIL returns the acquired discrete-time signals.
However, our architecture presents three strong differences with respect to the model-view-
controller and the multi-tier architecture:
• Adaptation of the discrete-time sequences provided by the TIL in the requests, both
at the transmitter and the receiver, consists in performing different signal processing
operations that sometimes can even be executed using different kind of processors. For
instance, a general-purpose processor (GPP), a graphic processor unit (GPU) or real-
time devices like an FPGA or a DSP. It is thus necessary to put all these operations
together and, consequently, the SPL concept comes out. Therefore, the SPL is in charge
of carrying out most of the signal processing operations needed between the MWL and
the TIL.
• One of the reasons to use a multi-tier architecture jointly with the model-view-controller
is the ability of the resulting application to be distributed among different machines.
Thus, different instances of the tiers run on different machines. However, in the testbed
system there are two kinds of nodes: the transmitter and the receiver. If a multi-node
testbed is available, different transmitter and receiver pairs will be distributed among
different nodes. As a result, the MWL (and consequently the SPL) is split into two
sides: the transmitter side and the receiver side. They will be referred to as Tx MWL, Rx
MWL, Tx SPL and Rx SPL, respectively. Additionally, they are also identified because
the processes mapping the architecture design are also split into two sets, one for the
transmitter and the other for the receiver.
• Finally, another particularity present when adapting the model-view-controller and the
multi-tier software architectures to testbed software architecture is that the hardware
is required to be attached to the host system by means of a bus (e.g. PCI, PCIX, PCI
Express, etc.). Consequently, it is not possible to sustain standard network connections
through such buses, and the standard tools and techniques available for interconnecting
layers are not applicable (except when a custom implementation is provided or when
Fig. 3. Basic structure of the distributed multilayer software architecture for MIMO testbeds.
The three layers are shown: MWL, SPL and TIL. Additionally, the testbed-hardware sub-layer
in the middleware and the different interconnection mechanisms are included.
the hardware is attached through a network connection). These reasons motivate an-
other division in the MWL, generating a Testbed-hardware sub-layer (with its respective
parts for the transmitter and the receiver sides) responsible of dealing with hardware
aspects as well as solving the bus connection issues with the host part of the MWL.
Having in mind that in our testbed this sub-layer runs on the DSPs available at the
transmitter and the receiver, the corresponding processes are referred to as TxDSP and
RxDSP.
5.2 Logical and Physical Designs of the Distributed Software Architecture

In Fig. 3, the basic structure of the proposed architecture is shown. There are two sides,
transmitter and receiver, joined at the TIL, which has to hold connections with both sides of
the SPL. All links between the different layer elements are implemented by using standard
socket connections. The exceptions are the links between the sub-layers of the MWL that use
a proprietary protocol over the existing main bus interconnecting the hardware and the host.
There are strong reasons for using standard socket connections instead of any other higher-
level mechanism like, for example, web services. However, this does not imply that in some
situations other connection types can better solve some problems. It is also possible to use
different link types among different layers. For example, the TIL and the SPL can be connected
using web services, thus allowing total independence of the platform and full remote access.
However, the ability of such high level techniques to sustain high data rates and low latency
connections is not the best. For these reasons, sockets offer enough flexibility while providing
fast connections. In the case of bus connections the bus type obviously limits the latency and
the data rates.
Fig. 4. Testbed scheme containing the hardware and the software architecture deployment as
well as the links among the components. The different processes are shown in blue, and are
located in the usual place for a typical architecture deployment. The digital hardware sections
of the testbed, as well as the RF front-ends, are shown in gray. Finally, the user application
is in dark green, containing the TIL in white. Note that TIL appears in white and not in blue
because it is not a process but an user application.
While in Fig. 2 the general structure of the testbed is depicted, showing the correspondence
between the hardware elements and the software layers, Fig. 4 illustrates the block diagram
of the entire system. Three main parts can be distinguished: the testbed hardware that allows
us to transmit discrete-time signals over multiple antennas; the multilayer software architec-
ture that makes the hardware accessible to end researchers at a high abstraction level; and,
finally, the user application implemented using the testbed capabilities and the architecture
facilities. The lowest software level (i.e. the MWL) is required to be installed in the same PCs
as the testbed hardware because it uses the system buses to communicate with the hardware.
Otherwise, this restriction would not be applicable. The other two layers can be installed in
any other available PC. However, in the standard deployment, the SPL is included with the
MWL in the testbed PCs. Finally, the TIL is installed in the remote user PC.
5.3 Short Description of the MIMO Testbed Hardware

At the bottom of Fig. 4 a very basic diagram of our testbed hardware is shown. A first release
of the testbed hardware was presented in (Ramı́rez et al., 2008), where the baseband modules
were from Sundance Multiprocessor Ltd. and the RF front-ends were developed at the Uni-
type of sequences given to TxProc, not all operations will be required either at the transmitter
or the receiver side.
The SPL can also incorporate advanced features such as a limited feedback channel for the
evaluation of precoded MIMO systems. This feedback channel is easily implemented using
the network connection available between the transmitter and the receiver.
A SPL detached from the other architecture layers is fundamental to deploy multiuser MIMO
scenarios. After extending the current TxProc and RxProc implementations in order to sup-
port multiuser MIMO techniques, the set of TxProc processes run at the transmitter users
while the RxProc processes do at the receiver users.
In some cases, the researchers may wish to implement most or even all operations present in
the SPL. It is still possible to configure the layer as a bypass, making only use of the intercon-
nection capabilities and not performing any signal processing operation at all.
5.7 Middleware Layer

The middleware layer (MWL) fills the gap between the testbed hardware and the signal pro-
cessing layer, allowing discrete-time signals to be transferred through the system bus and
making possible the synchronization between the TxPC and the RxPC using standard TCP
network connections. The MWL is split into two different sub-layers (see Fig. 3 and Fig.
4). The topmost sub-layer is responsible of establishing the network connections between the
transmitter and the receiver, and with the higher layer (the SPL). The bottom sub-layer corre-
sponds to the testbed hardware configuration and control software. Fig. 3 and Fig. 4 shows
the MWL with its two sub-layers plus the connections with adjacent layers.
Four different processes constitute the middleware. The first two, termed TxHost and RxHost,
run respectively on the TxPC and RxPC. They are implemented in standard C++ language and
use sockets to establish the necessary network connections:
• One connection between the TxHost and RxHost processes. It is used to synchronize the
transmitter and the receiver, so the receiver knows when the signal acquisition process
has to start.
• Another connection is established between the TxHost and the TxProc processes and
it is used to link the transmitter side of both the middleware and the signal processing
layers.
• Finally, there is also a connection between the RxHost and the RxProc processes.
The remaining two processes are the transmitter and the receiver processes that run on their
respective Digital Signal Processors (DSPs) available in the testbed hardware. Thus, the trans-
mitter DSP process (TxDSP) performs data transfers through the PCI bus jointly with the Tx-
Host process and configures and controls the hardware components at the TxPC. In the same
way, the RxHost process and the DSP receiver process (RxDSP) are responsible of transferring
the data through the PCI bus and, from the DSP side, controlling and configuring the testbed
hardware components.
The MWL serves the maximum number of requests that the hardware supports. It serves
a request while the data for the next frame is still being generated by any computer in the
network and, when the first frame is passed to the SPL, the MWL is ready to accept a new
frame. With this architecture scheme, the MWL simultaneously serves requests from other
SPL instances running in different PCs.
For the specific case of our MIMO testbed, in order to release the MWL implementation from
dealing with the lowest level hardware details, the Sundance SMT6025 software development
the protocol: the CSI information must be sent during the acknowledgement stage to notify
the transmitter whether it has to re-configure the transmission parameters or not. To do so, in
the sequence diagram shown in Fig. 7 there should be added two ”Send CSI” messages: one
from the RxHost to the TxHost and another from the TxHost to the TxProc.
If the signal adaptation at the transmitter side takes place at the user application, then no
feedback messages are needed in the protocol because everything is available at the user ap-
plication. It is necessary to call the testbed interface layer twice: one to be able to estimate the
channel and the following with the adapted signals.
7.5 Deployment of a Multiuser Environment

The last given case of study describes the functioning of the designed protocol when work-
ing in a multiuser environment where several nodes receive information from a central node
(broadcast channel). The protocol just needs to be generalized to support the synchronization
of several receiver nodes.
Fig. 8 presents the use case for two receiver users. In order to implement this scenario it is
necessary to incorporate some modifications:
• After configuring the testbed hardware, TxHost waits for the ”ready” message from all
receive users.
• TxHost waits for the ”Send Frame” message from all users before starting the transmis-
sion.
• The frame is cyclically transmitted and each individual user can carry out as many
acquisitions as desired. Once all users have acquired all necessary data, they send an
”OK” message to the TxHost. When TxHost has received all ”OK” messages from all
users, the current frame transmission is finished.
• When TxProc receives an end message, it passes it to TxHost which launches a multicast
message to all receive users in order to properly finish.
7.6 Using a Simpler Protocol

In some cases it is interesting to provide a simpler protocol in order to delegate most of the
control mechanisms to the user application. Now the TIL has two different functions allowing
the control of the transmitter and the receiver separately. After the hardware initialization
step (see Fig. 7), the user application decides to send a set of discrete-time sequences and
then uses the TIL to send them to the SPL. Afterwards, the SPL sends them to the MWL to
be transmitted, but RxHost is waiting for RxProc in order to carry out the acquisition. When
the user application decides to acquire the data, it calls again the TIL which asks RxProc to
acquire the data. Immediately, RxProc asks RxHost to acquire the data and returns it to the
user application through the RxProc and the TIL.
The main difference of this functioning mode with respect to the basic operation mode is the
lack of messages between TxHost and RxHost, which are now completely autonomous. This
presents the advantage of easier integration of the testbed nodes with other nodes and, at
the same time, all nodes can be directly controlled from the user application. However, as
a disadvantage, the testbed control is moved to the user application. Consequently, the user
application has now to deal with some hardware details (e.g. the size of buffers) but properly
encapsulated in the TIL. In addition, performance losses arise as a consequence of the testbed
being controlled from the user application.
8. Conclusion
In this chapter we propose a new distributed multilayer software architecture for MIMO
testbed user access. The architecture fills the gap that currently exists between commercial
hardware components and the most common abstraction level used by researchers. The archi-
tecture overcomes the limitations of implementing new algorithms directly from the testbed.
Instead of using the low-level interfaces typically provided by manufacturers, the architecture
supplies a high-level interface access for testbeds. It releases researchers from the necessity of
knowing the details of the testbed hardware. For instance, they can easily test new algorithms
without developing a completely new source code release specifically for the testbed, thus
speeding up the implementation and test tasks.
The key point to the multilayer architecture design is the use of ordinary network connec-
tions to link both software modules and hardware components. These connections are also
useful for decoupling the software layers. Several advantages are obtained as a consequence
of the multilayer software architecture. Also, multiuser scenarios and feedback channels can
be constructed extending the proposed architecture. Finally, it is also possible to simplify the
presented architecture protocols in order to ease the integration with heterogeneous nodes.
The testbed control is moved to the user application instead of keeping it in the inner layers.
9. References
3L Ltd (2009), http://www.3l.com/
Borkowski, D.; Brhl, L.; Degen, C.; Keusgen, W.; Alirezaei, G.; Geschewski, F.; Oikonomopou-
los, C. & Rembold, B. (2006). SABA: A testbed for real-time MIMO systems. EURASIP
Journal on Applied Signal Processing, Vol. 2006, 2006.
Caban, S.; Mehlführer, C.; Langwieser, R.; Scholtz, A.L. & Rupp, M. (2005), Vienna MIMO
Testbed, EURASIP Journal on Applied Signal Processing, Vol. 2006, 2006.
Fabregas, A.G.; Guillaud, M.; Slock, D.T.M.; Caire, G.; Gosse, G.; Rouquette, S.; Ribeiro Dias,
A.; Bernardin, P.; Miet, X.; Conrat, J.M; Toutain, Y.; Peden, A. & Li, Z. (2005), A
MIMO-OFDM Testbed for Wireless Local Area Networks, EURASIP Journal on Ap-
plied Signal Processing, Vol. 2006, 2006.
Foschini, G. & Gans, M. (1998), On Limits of Wireless Communications in a Fading Environ-
ment when Using Multiple Antennas, Wireless Personal Communications, Vol. 6, 1998,
pp. 311-335.
Garcı́a-Naya, José A.; González-López, M. & Castedo, L. (2008), An Overview of MIMO
Testbed Technology, Proceedings of 4th International Symposium on Image/Video Com-
munications over Fixed and Mobile Networks (ISIVC 2008), July, 2008, Bilbao
Garcı́a-Naya, José A.; Pérez-Iglesias, Héctor J.; Fernández-Caramés, T. M.; González-López,
M. & Castedo, L. (2008), A distributed multilayer architecture enabling end-user
access to MIMO testbeds, Proceedings of IEEE 19th International Symposium on Per-
sonal, Indoor and Mobile Radio Communications, 2008 (PIMRC 2008), September, 2008,
Cannes.
Haustein, T.; Forck, A.; Gäber, H.; Jungnickel, V. & Schiffermüller, S. (2005), Real-Time Signal
Processing for Multiantenna Systems: Algorithms, Optimization, and Implementa-
tion on an Experimental Test-Bed, EURASIP Journal on Applied Signal Processing, Vol.
2006, 2006.
Kaiser, T.; Wilzeck, A.; Berentsen, M. & Rupp, M. (2004). Prototyping for MIMO Systems An
Overview, Proceedings of the XII European Signal Processing Conference (EUSIPCO 2004),
September, 2004, Vienna.
The Mathworks (2009), http://www.mathworks.com/
Moore, G. (1998), Cramming more components onto integrated circuits, Proceedings of the IEEE,
Vol. 86, No. 1, 1998, pp. 82–85.
Nieto, X.; Ventura, L.M. & Mollfulleda, A. (2006), GEDOMIS: a broadband wireless MIMO-
OFDM testbed, design and implementation, Proceedings of 2nd International Conference
on Testbeds and Research Infrastructures for the Development of Networks and Communities
(TRIDENTCOM 2006), 2006, Barcelona.
PXI: PCI eXtensions for Instrumentation (2009), http://www.pxisa.org
Pérez-Iglesias, H.J.; Garcı́a-Naya, J.A.; Dapena, A.; Castedo, L. & Zarzoso, V. (2008), Blind
channel identification in Alamouti coded systems: a comparative study of eigende-
composition methods in indoor transmissions at 2.4 GHz, European Transactions on
Telecomunications, Vol. 19, No. 7, November 2008, pp. 751–759.
Ramı́rez, D.; Santamarı́a, I.; Pérez, J.; Vı́a, J.; Garcı́a-Naya, J.A.; Fernández-Caramés, T.M.;
Pérez-Iglesias, H.J.; González López, M.; Castedo, L. & Torres-Royo, J.M. (2008), A
comparative study of STBC transmissions at 2.4 GHz over indoor channels using a
2 x 2 MIMO testbed, Wireless Communications and Mobile Computing, Vol. 8, No. 9,
November 2008.
Rao, R.M.; Wiejun Zhu Lang, S.; Oberli, C.; Browne, D.; Bhatia, J.; Frigon, J.-F.; Jingming
Wang Gupta, P.; Heechoon Lee; Liu, D.N.; Wong, S.G.; Fitz, M.; Daneshrad, B. &
Takeshita, O. (2004), Multi-antenna testbeds for research and education in wireless
communications, Communications Magazine, IEEE, Vol. 42, No. 12, December 2004,
pp. 72–81, ISSN: 0163-6804.
Rupp, M.; Burg, A. & Beck, E. (2003). Rapid Prototyping for Wireless Designs: the Five-Ones
Approach, Signal Processing Europe, Vol. 83, 2003, pp. 1427-1444.
Rupp, M.; Mehlfhrer, C. & Caban, S. (2006), Testbeds and Rapid Prototyping in Wireless Sys-
tem Design, EURASIP Newsletter, Vol. 17, No. 3, 2006, pp. 32-50.
Rupp, M.; Caban, S. & Mehlfhrer, C. (2007), Challenges in Building MIMO Testbeds, Proceed-
ings of European Signal Processing Conference (EUSIPCO 2007), 2006, Poznan.
SDR Forum: Software Defined Radio Forum (2009), http://www.sdrforum.org
SCA: Software Communication Architecture (2009), http://sca.jpeojtrs.mil
Sundance Multiprocessor, Ltd. (2009), http://www.sundance.com
Telatar, I. E. (1999), Capacity of Multi-Antenna Gaussian Channels, European Transactions on
Telecommunications, Vol. 10, No. 6, 1999, pp. 585-595.
Texas Instruments (2009), http://www.ti.com
Trygve, M. H. (1978), MVC, XEROX PARC, http://heim.ifi.uio.no/˜trygver/
themes/mvc/mvc-index.html
Wilzeck, A.; El-Hadidy, M.; Cai, Q.; Amelingmeyer, M. & Kaiser, T. (2006). MIMO Prototyping
Testbed with off-the-shelf plug-in RF Hardware, IEEE Workshop on Smart Antennas
(WSA 2006), 2006 Ulm.
Zhu, W.; Browne, D. & Fitz, M. (2005). An Open Access Wideband Multi-Antenna Wireless
Testbed with Remote Control Capability, Proceedings of 2nd International Conference on
Testbeds and Research Infrastructures for the Development of Networks and Communities
(TRIDENTCOM 2005), Trento.
Recent Developments in Channel Estimation and Detection for MIMO Systems 99
Recent Developments in Channel

Estimation and Detection for MIMO Systems
Seyed Mohammad-Sajad Sadough
Shahid Beheshti University, Faculty of Electrical and Computer Engineering
Iran
Mohammad-Ali Khalighi
Institut Fresnel, UMR CNRS 6133, École Centrale Marseille
France
1. Introduction
It is well known that multiple-input multiple-output (MIMO) systems can provide very high
spectral efficiencies in a rich scattering propagation medium Telatar (1999). They are hence a
promising solution for high-speed, spectrally efficient, and reliable wireless communication
Raoof et al. (2008) .When coherent signal detection is to be performed at the receiver, channel
state information at the receiver (CSIR) is required, for which a channel estimation step is
necessary. Channel estimation plays a critical role in the performance of the receiver. It is a
real challenge in practical MIMO systems where the quality of data recovery is as important
as attaining a high data throughput.
In order to obtain the CSIR, usually some known training (also called pilot) symbols are sent
from the transmitter, based on which the receiver estimates the channel before proceeding
to the detection of data symbols. The classical approach consists in time-multiplexing pilot
and data symbols, usually referred to as pilot symbol-assisted modulation (PSAM) Cavers
(1991). We start the chapter by introducing the PSAM channel estimation for MIMO systems
(Section 2). Instead of this classical channel estimation based on pilot symbols only, we can
perform semi-blind estimation that in addition to pilot symbols, makes use of data symbols in
channel estimation. In this way, a considerable performance improvement can be achieved at
the price of increased receiver complexity De Carvalho & Slock (1997); Giannakis et al. (2001);
Sadough (2008). Usually, these semi-blind approaches are implemented in an iterative scheme
when channel coding is performed. That is, channel estimation is performed iteratively to-
gether with signal detection and channel decoding Sadough, Ichir, Duhamel & Jaffrot (2009).
We, hence, continue Section 2 by considering semi-blind estimation for the case of time-
multiplexed pilots.
The drawback of the PSAM scheme is the encountered loss in the spectral efficiency by the
periodic insertion of pilot symbols. As an alternative to this method, overlay pilots (OP) can
be employed, where pilot symbols are sent in parallel with data symbols Hoeher & Tufvesson
(1999). We introduce the OP approach in Section 3 and explain, in particular, pilot-only-based
and semi-blind estimation approaches. The pros and cons of OP with respect to PSAM are
discussed too.
Whatever the channel estimation technique, in practice, the receiver can only obtain an im-
perfect estimate of the channel. Classically, for signal detection, the estimated channel is con-
sidered as the perfect estimate. This sub-optimal approach is usually called the mismatched
receiver. Its sub-optimality is due to the fact that the receiver does not take into account the
presence of channel estimation errors Sadough & Duhamel (2008); Sadough (2008). A more
appropriate approach is to take into account channel estimation inaccuracies in the formu-
lation of the detector. We firstly consider in Section 4 the effect of estimation errors on the
receiver performance and the impact of the employed space-time coding scheme.
Next, in Section 5, we consider maximum-likelihood (ML) signal detection and show how to
integrate the imperfect channel knowledge into the design of the detector. More precisely,
we consider two iterative detectors based on maximum a posteriori (MAP) and soft parallel
interference cancellation (soft-PIC), and propose for each case modifications to the MIMO de-
tectors for taking into account the channel estimation errors. The implementation complexity
issues are also discussed. We present some numerical results to demonstrate the performance
improvement obtained via the use of the improved detectors. Finally, Section 8 concludes the
chapter.
Spatial .
Binary b c Bit d Bit/symbol s multiplexing .
Encoder Interleaver /
Source Mapping .
STC
MT
Fig. 1. Transmitter architecture of MIMO-BICM scheme.
1.1 Assumptions and notations

We consider a single-user MIMO system with M T transmit and M R receive antennas, trans-
mitting over a frequency non-selective channel and refer to it as an ( M R × M T ) MIMO chan-
nel. Unless otherwise mentioned, single carrier modulation and block fading channel model
is considered where the channel is assumed to remain almost constant over the duration of a
block of symbols.
Figure 1 shows the block diagram of the transmitter that employs the bit-interleaved coded
modulation (BICM) scheme which is known to be a simple and efficient method for exploiting
channel time-selectivity. The binary data sequence b is encoded by a forward error correction
(FEC) code before being interleaved by a quasi-random interleaver. The output bits d are
mapped to constellation symbols s and then either multiplexed spatially or encoded according
to a space-time scheme before being sent through the wireless channel. Let us denote by x and
y respectively the ( M T × 1) and ( M R × 1) vectors of transmit and received symbols at a given
time reference. For simplicity, we assume for now the simple spatial multiplexing scheme. We
have:
y = Hxx + z (1)
where H denotes the ( M R × M T ) channel matrix and z is the vector of additive complex white
Gaussian noise of zero mean and covariance matrix Σz = σz2 I MR , where I n denotes an (n × n )
Identity matrix. We assume here that σz2 is perfectly known at the receiver and focus on the
estimation of H.
2. Time-multiplexed Pilots and Data

Most current systems use a training-based channel estimation scheme in the form of time-
multiplexed pilot symbols. In what follows, we present the PSAM approach for MIMO sys-
tems and explain two cases of pilot-only-based and semi-blind estimation; the latter is imple-
mented in an iterative receiver.
2.1 Pilot-only-based PSAM channel estimation

When using the PSAM method, we have a trade-off between the channel estimation quality
and the data throughput. With an increased number of pilots, a better channel estimate can be
obtained, but at the same time, the spectral efficiency is sacrificed more. There is a minimum
number of channel-uses that should be devoted to the transmission of pilots in order that the
MIMO channel be identifiable at the receiver. As a general case, if L denotes the maximum
length of the underlying subchannels’ impulse response, the number of pilot channel-uses Np
should satisfy Balakrishnan et al. (2000):
( N p − L + 1) ≥ M T L (2)
Under flat fading conditions where L = 1, this implies: Np ≥ M T . Several works have been
done on the optimal placement of pilots in a frame of symbols, as well as on the optimal
power allocation between pilot and data symbols Hassibi & Hochwald (2003); Dong & Tong
(2002); Adireddy et al. (2002); Ma et al. (2003). They consider the criteria of mean-square chan-
nel estimation error, channel capacity, or the Cramér-Rao bounds. In particular, it is shown
in Hassibi & Hochwald (2003) that, if power optimization over pilot and data symbols is al-
lowed, the optimum Np is Np opt = M T . In such a case, we should place pilot symbols with
lower power at the beginning and the end of the frame, and those with higher power in the
middle of the frame Dong & Tong (2002). However, if equal power has to be allocated to pilot
and data symbols, then Np opt can be larger than M T Hassibi & Hochwald (2003).
Considering the simple case of flat block-fading MIMO channel, to estimate the channel, cor-
responding to each fading block, we send Np pilot symbol vectors with the same power as
the data symbols. Let x p [ k] denote an ( M T × 1) pilot symbol vector at the time sample
k. We denote the received vector corresponding to x p [ k] by y p [ k]. We can constitute the
( M T × Np ) matrix X p by stacking in its columns the pilot vectors x p [ k], k = 0, 1, Np − 1, i.e.,
X p = [ x p [0], . . . , x p [ Np − 1]].
According to (1), during a given channel training interval, we have:
Yp = H Xp + Zp. (3)
The definitions of Y p and Z p are similar to that of X p . We denote by E p the average power of
the training symbols on any subcarrier as
1
Ep tr X p X†p . (4)
Np M T
The maximum likelihood (ML) channel estimate Ĥ, which is equivalent to the least-squares
solution, is Balakrishnan et al. (2000):
Np −1 Np −1 −1
Ĥ = ∑ y p [ k] x †p [ k] ∑ x p [ k] x †p [ k] , (5)
k =0 k =0
which can be written in a more compact form as:
Ĥ = Y p X†p (X p X†p )−1 , (6)
where .† denotes transpose-conjugate.

Let us denote by E the matrix of estimation errors, that is, E = Ĥ − H. From (3) and (6), it is
easy to show that
E = Z p X†p (X p X†p )−1 . (7)
It is known that the best channel estimate is obtained with mutually orthogonal training se-
quences, which result in uncorrelated estimation errors. In other words, we should choose X p
with orthogonal rows such that
X p X†p = Np E p I MT . (8)
Then, the j-th column E j of E has the covariance matrix Σ given by
σz2
Σe = E E j E j† = σe2 I MR , where σe2 = · (9)
Np E p
2.1.1 Statistics of the Channel Estimation Errors

We saw that the estimated channel matrix Ĥ can be viewed as a noisy version of the perfect
channel matrix H. In Section 5 we show that for channel estimators having this feature, the
detection performance can be improved if the statistics of the channel estimation errors are
known.
Let us reconsider the pilot-based ML channel estimator of equation (7). The good feature of
this estimator is that the statistics of the channel estimation error matrix E are known (see
equation (9)). By using these statistics and equation (7), the conditional pdf of Ĥ given H can
be easily expressed as:
p(Ĥ| H) = CN H, I MT ⊗ Σe , (10)
where ⊗ denotes the Kronecker product and CN denotes the complex Gaussian distribution.
Furthermore, we assume that the channel matrix H has a normal prior distribution as:

1 −1 †
H ∼ CN 0, I MT ⊗ Σ H = M M exp − tr HΣ H H (11)
π R T det{Σ H } MT
where Σ H is the ( M R × M R ) covariance matrix of the columns of H, and det{} denotes ma-
trix determinant. We assume that the entries of H, i.e., the fading coefficients of different
subchannels, are i.i.d. Then, Σ H is a diagonal matrix with equal diagonal entries σh2 .
By using the prior pdf of H from (11) and the pdf of (Ĥ| H) from (10), we can derive the
posterior distribution of the perfect channel matrix, conditioned on its ML estimate, as follows
(see Sadough & Duhamel (2008) for the details of the derivation):

p (H| Ĥ) = CN Σ∆ Ĥ, I MT ⊗ Σ∆ Σe , (12)
where −1
Σ∆ = Σ H Σe + Σ H . (13)
Under the above-mentioned assumptions, we have
Σ∆ = δI MR (14)
where
σh2
δ= · (15)
σh2 + σe2
In particular, when the number of pilot symbols tends to infinity, it is not difficult to see that
δ → 1 and δ σe2 → 0 and consequently p (H| Ĥ) tends to a Dirac delta function. The availability
of the estimation error distribution is an interesting feature of pilot-only-based PSAM channel
estimation that we used to derive the posterior distribution (12). This distribution constitutes
a Bayesian framework which is exploited in Section 5 for the design of detectors by taking into
account channel estimation inaccuracies.
2.2 Semi-blind PSAM channel estimation

In order to preserve the spectral efficiency for the transmission of data symbols, we are in-
terested in minimizing the number of pilot symbols in a frame. However, by reducing the
number of pilot symbols, the channel may be learned improperly and channel estimation er-
rors may become important. This can result in a considerable performance degradation and
in the need to data retransmission. This performance degradation can be compensated by
smart signal processing at the receiver. In fact, instead of estimation methods based on pi-
lot symbols only, we can use semi-blind approaches that in addition to pilot symbols, make
use of data symbols in channel estimation. In this way, a considerable performance improve-
ment can be achieved at the price of increased receiver complexity de Carvalho & Slock (2001);
Sadough, Ichir, Duhamel & Jaffrot (2009). We present here two semi-blind channel estimation
schemes that we implement in an iterative receiver. The first semi-blind method that we con-
sider is the thresholded hard-decisions (Th-HD) method and the second one is based on the ex-
pectation maximization (EM) algorithm Dempster et al. (1977); Moon (1996). For both meth-
ods, at the first iteration, we calculate a primary channel estimate based on the pilot sequences
only, which allows the semi-blind estimator, used in the succeeding iterations, to bootstrap.
Before describing these methods, we present in the following details on the iterative receiver.
2.2.1 Iterative signal detection

We usually consider in this paper iterative signal detection in the case of using non-orthogonal
space-time codes at the transmitter. As shown in Fig. 2, the receiver mainly consists of a com-
bination of two sub-blocks that exchange soft information with each other. The first sub-block,
referred to as soft detector or demapper, produces extrinsic soft information from the input
symbols and send it to the second sub-block, the soft-input soft-output (SISO) channel de-
coder. Here, we consider SISO channel decoding based on the well known forward-backward
algorithm Bahl et al. (1974). Soft MIMO signal detection and soft-input SISO channel decoding
are performed iteratively and the estimates of the channel coefficients are updated at each iter-
ation of the turbo-detector Sadough, Ichir, Duhamel & Jaffrot (2009); Berthet et al. (2001). The
blocks Π and Π−1 denote bit-level interleaver and de-interleaver, respectively, corresponding
to the BICM scheme used at the transmitter.
2.2.2 Th-HD semi-blind estimation

In the Th-HD method, in addition to pilot symbols, we use in channel estimation
the symbols detected with high reliability at each iteration Khalighi & Boutros (2006);
( m)
Sellathurai & Haykin (2002). For instance, consider the a posteriori probability Pi at the
decoder output at iteration m, corresponding to the coded bit ci . We compare it with a
( m) ( m +1)
threshold 0.5 < PTH < 1. If Pi > PTH , we make the hard decision ĉi = 1 ; otherwise,
1
y1
2 MIMO
y2
soft−input Channel
Π −1
. . .
soft−output decoder
MR
y MR detector
Π
estimated
channel
...
coefficients
Pilot sequences
Fig. 2. Iterative channel estimation and data detection, y i denotes the received signal on the
ith antenna.
( m) ( m +1)
if Pi < (1 − PTH ), we make the hard decision ĉi = 0; and if none of these conditions
are verified, we give up the channel-use corresponding to ci and do not consider it in chan-
nel estimation. If a hard decision is made on all BM T constituting bits of a channel-use, we
use the resulting hard-detected symbol vector in channel estimation, in the same way as pilot
symbols. The resulting channel estimate is then used in the next iteration of the detector.
The performance of Th-HD depends highly on the choice of the threshold PTH that determines
whether or not the SISO decoder soft-outputs are reliable enough. The practical limitation is
that the optimum threshold value depends on the MIMO structure, i.e., the number of trans-
mit and receive antennas, as well as on the actual SNR Khalighi & Boutros (2006). Note that, if
we take PTH very close to 0.5, we effectively make hard decisions on all detected symbols and
use them in channel estimation. This coincides with the so called decision-feedback channel
estimation Visoz & Berthet (2003).
2.2.3 EM-based semi-blind estimation

The interest of the EM algorithm is that it is guaranteed to be stable and to converge to an ML
estimate Moon & Stirling (2000). We do not present here the details on the formulation of the
EM-based estimator and refer the reader to Khalighi & Boutros (2006), for instance. A simple
and classical formulation is when data and pilot symbols are used in the same way in channel
estimation. By this approach, the estimated channel matrix is given below:
−1
Ĥ = R yx R x , (16)
where
Ns
R yx = ∑ y [k] x̃x † [k] (17)
k =1
and 

 Ns ; i=j
Ns
R x,i,j = (18)

 ∑ x̃x i [k] x̃x ∗j [k] ; i = j
k =1
Here, Ns denotes the number of channel-uses per frame, y [ k] is the received symbol vector
at the time reference k, and x̃x [ k] is the corresponding soft-estimates of the transmitted sym-
bol vector, calculated using the SISO channel decoder outputs at the preceding iteration. For
k = 1, · · · , Np , we have x̃x [ k] = x p [ k]. Also, R x,i,j denotes the (i, j)th entry of matrix R x . The
formulation that we provided for EM can be modified further to improve the receiver per-
formance. Interested reader may refer to Khalighi & Boutros (2006); Khalighi et al. (2006) for
details.
2.2.4 Case study

BER curves versus Eb /N0 are shown in Fig. 3 for the case of (8 × 8) and (8 × 6) MIMO
structures using the PSAM technique. Rayleigh independent quasi-static fading model is con-
sidered with blocks of length Ns = 64 channel-uses. The number of channel-uses devoted to
pilot transmission is Np = 10. Three cases of pilot-only-based estimation, and Th-HD and
EM-based semi-blind estimations are considered. The case of perfect-CSIR is also provided as
reference. The simple spatial multiplexing scheme is used at the transmitter and MIMO sig-
nal detection is based on soft parallel interference cancellation (Soft-PIC) Sellathurai & Haykin
(2002); Lee et al. (2006). Results shown in Fig. 3 correspond to the eighth iteration of the re-
ceiver. We notice that the performance of the semi-blind Th-HD and EM-based estimator are
very close to each other and they outperform the pilot-only-based method. Yet, their per-
formance is about 2 dB away from the perfect-CSIR case that is due to the relatively high
co-antenna interference, as we have M T = 8. We can approach further the perfect-CSIR case
by increasing Np .
3. Superimposed Pilots and Data

The main drawback of the PSAM approach is that, for finite-length blocks, if channel estima-
tion is to be done on each block of symbols, the periodic insertion of pilot symbols can result
in a considerable reduction of the achievable data rate. This loss in the data rate becomes
important, specially for large number of transmit antennas, at low SNR, and when the chan-
nel undergoes relatively fast variations Hassibi & Hochwald (2003). As an alternative, we can
use overlay pilots (OP), also called superimposed or embedded pilots, for channel estimation
Hoeher & Tufvesson (1999); Zhu et al. (2003). In this approach, a pilot sequence is superim-
posed on the data sequence before transmission, as shown in Fig. 4; thus, no separate time
slot is dedicated to pilot transmission.
3.1 Channel estimation using OP

By using OP, we prevent the loss in the data throughput but we experience degradation in the
quality of the channel estimate due to the unknown data symbols Hoeher & Tufvesson (1999).
As a matter of fact, here also there is a trade-off between high quality channel estimation
and the information throughput: To obtain a better channel estimate, we should increase the
percentage of the power dedicated to pilot symbols; this, however, reduces the SNR for the
detection of data symbols Tapio & Bohlin (2004). In general, OP may be preferred to PSAM for
high SNR, not too short channel coherence times, and larger number of receive than transmit
antennas Khalighi et al. (2005).
Consider again the block fading channel model. Assuming uncorrelated data and pilot se-
quences, we can estimate channel coefficients by calculating the cross-correlation between
the received sequences on each antenna and the transmitted pilot sequences, known to the
0
10
−1 Pilot−Only
10
−2
Th−HD
10
BER
−3
10
Perfect−CSIR
−4
10
(8x8)
EM−based
(8x6)
−2 0 2 4 6 8 10 12
Eb/N0 (dB)
Fig. 3. Comparison of different estimation methods based on PSAM, iterative Soft-PIC de-
tection, 8th iteration of the receiver (almost full convergence); (8 × 8) and (8 × 6) systems,
i.i.d. Rayleigh quasi-static fading, QPSK modulation, (5, 7)8 NRNSC rate 1/2 channel code,
orthogonal pilot sequences with Np = 10, Ns = 64 channel-uses. Eb /N0 takes into account the
receiver antenna gain M R .
receiver. As the pilot sequences transmitted from different antennas are orthogonal, the esti-
mation errors arise from noise and the data sequences. Indeed, the main problem in the esti-
mation of channel coefficients concerns the latter interference component, i.e., the unknown
data symbols. In fact, although data and pilot sequences are statistically uncorrelated, the
cross-correlation is calculated over a block of symbols of limited length, over which the chan-
nel coefficients are supposed to remain unchanged. The smaller is the block length (i.e., the
faster the channel fading), the more important this cross-correlation is. This can result in an
error floor in the receiver BER performance Jungnickel et al. (2001), especially at high SNR,
and make the OP scheme lose its interest.
pilot symbols, σp2
data symbols, σd2
Fig. 4. Overlay pilot scheme

3.2 Iterative channel estimation for OP

Iterative data detection and channel estimation can be a solution to the problem of error floor
Zhu et al. (2003); Khalighi et al. (2005); Cui & Tellambura (2005). We consider in the follow-
ing, two such iterative schemes; a pilot-only-based decision-directed estimator Khalighi et al.
(2005) and an EM-based semi-blind estimator Khalighi & Bourennane (2007).
Let us denote by x d and x p the vectors of data and pilot symbols, respectively, corresponding
to the transmitted ( M T × 1) vector x : x = x d + x p . We denote by σd2 and σp2 the power allocated
to the entries of x d and x p , respectively.
3.2.1 Pilot-only-based decision-directed estimator

Consider the estimation of the entry Hij of the channel matrix H. As explained above, this
estimate can be obtained by calculating the cross-correlation Γ ij between the sequence re-
ceived on the antenna #i, y i , and the pilot sequence transmitted on the antenna #j, x p,j . By the
decision-directed method that we denote by DD, we use in each iteration, the soft-estimates
x̃d of transmitted data symbols using a posteriori LLRs at the SISO decoder output, and cancel
their effect in Γ ij .
3.2.2 EM-based semi-blind estimator

To obtain a better performance, similar to the case of PSAM, we can employ semi-blind estima-
tion methods Khalighi & Bourennane (2008); Bohlin & Tapio (2004); Meng & Tugnait (2004) at
the price of increased Rx complexity. For instance, a semi-blind estimator based on the EM
algorithm may be used. Its formulation is analogous to the case of PSAM and can be found in
Khalighi & Bourennane (2007).
3.2.3 Data-dependent overlay pilots

A solution for getting rid of the interference from data symbols in channel estimation is to use
data-dependent overlay pilot sequences such that the corresponding pilot and data sequences
are orthogonal Ghogho et al. (2005). The drawback of this method is that it results in nulls in
the equivalent channel impulse response (seen by data symbols), and hence, in a performance
degradation. This degradation could be reduced by performing iterative channel equalization
Lam et al. (2008). On the other hand, such a scheme leads to increased envelope fluctuations
of the transmitted signal. We do not consider this scheme here.
3.2.4 Case study

Let us denote by α the ratio of the power of pilot symbols to the total transmit power at a
symbol time, i.e., α = σp2 /(σp2 + σd2 ). For a (2 × 4) MIMO system, we have presented in Fig.
5, BER curves versus SNR for DD, EM-based, and perfect channel estimation, and different
values of α. Pilot sequences for M T antennas can be QPSK modulated and chosen according
to the Walsh-Hadamard series to ensure their orthogonality. Results correspond to the fifth it-
eration of the Rx where almost full convergence is attained. SNR in Fig. 5 stands for the actual
average received SNR, i.e., M R (σd2 + σp2 )/σn2 , in contrast to Eb /N0 that takes into account only
σd2 . In this way, we can directly see the compromise between the channel estimation quality
and the data detection performance, e.g. by increasing α. Again, Soft-PIC MIMO detection is
performed. For both estimation methods, we notice an error floor at high SNR for α < 15%
which is especially visible for α = 2% and α = 5%. On the other hand, by increasing α, better
channel estimates are obtained, but at the same time, less power is dedicated to data symbols.
So, increasing α too much, will result in an overall performance degradation. Comparing
0
10
−1
10
−2
10
−3
10
−4
10
−2 0 2 4 6 8 10 12
0
10
Alamouti, E /N =7dB
b 0
−1
10 MUX, 4th iteration, E /N =5dB
b 0
GLD, 4th iteration, E /N =4.4dB

b 0
BER
−2
10
−3
10
−4
10
2 4 8 16 32
Np (channel−uses)
Fig. 6. Sensitivity to channel estimation errors; (2×2) MIMO channel, η = 2 bps/Hz,

(133, 171)8 NRNSC rate 1/2 channel code, i.i.d. Rayleigh flat block fading channel with 32
independent fades per frame of 768 channel-uses. Iterative Soft-PIC used for MUX and GLD.
Concerning the practical case of imperfect channel estimate, in fact, lower-rate orthogonal
schemes could be more sensitive to channel estimation errors as, in general, they have to use
a larger signal constellation to attain a desired spectral efficiency. Concerning non-orthogonal
schemes, we need, in general, smaller constellation sizes as compared to orthogonal ones.
However, the iterative detector for the non-orthogonal schemes could be more sensitive to
channel estimation errors because its convergence, and hence, its performance is affected by
these errors.
To study the effect of channel estimation errors, let us consider pilot-only-based channel esti-
mation using time-multiplexed pilots. For each fading block, we devote Np channel-uses to
the transmission of power-normalized mutually orthogonal QPSK pilot sequences from M T
transmit antennas Khalighi & Boutros (2006). For a (2 × 2) MIMO system, we have shown
in Fig. 6 the average BER after four detector iterations versus Np for a spectral efficiency of
η = 2 bps/Hz. We have considered the Alamouti code Alamouti (1998) as the orthogonal
scheme, and two cases of spatial multiplexing (denoted here by MUX) and Golden coding
Belfiore et al. (2005) (denoted by GLD) as non-orthogonal schemes. The Eb /N0 for each ST
scheme is set to what results in BER ≈ 10−4 in the case of perfect channel knowledge. From
Fig. 6 we notice an almost equivalent sensitivity to the channel estimation errors for MUX and
GLD schemes. This comparison makes sense as the SNRs for these schemes are close to each
other. On the other hand, we see that the Alamouti scheme has the lowest sensitivity. This
is due to the orthogonal structure of the code, and the fact that the SNR is higher, compared
to those for MUX and GLD schemes, and as a result, the quality of channel estimate is much
better.
5. Improved Signal Detection in the Presence of Channel Estimation Errors

For the case of time-multiplexed pilot and data, as we explained previously, in order to ob-
tain a good channel estimate, we should increase Np , which in turn, results in a larger loss
in the spectral efficiency, specially for relatively fast time-varying communication channels
Hassibi & Hochwald (2003). One solution is to use semi-blind channel estimation in or-
der to reduce the number of channel-uses devoted to pilot transmission, as seen in Section
2.2. The disadvantage of semi-blind approaches is the increased receiver complexity. For a
reduced-complexity semi-blind joint channel estimator and data detector the reader is refered
to Sadough, Ichir, Duhamel & Jaffrot (2009).
An alternative to this solution is to modify the detector so as to take into account channel
estimation errors. As a matter of fact, the classical approach is to use the channel estimate in
the detection part in the same way as if it was a perfect estimate, what is known as mismatched
signal detection. Obviously, this approach is suboptimal and can degrade considerably the
receiver performance in the presence of channel estimation errors.
In this section we provide the general formulation of a detection rule that takes into ac-
count the available imperfect CSIR and refer to it as the improved detector. To this end, we
consider the model (1) and denote by J (y , x , H) the quantity (cost function) that would let
us to decide in favor of a particular x at the receiver if the channel was perfectly known.
Note that depending on the detection criteria, the quantity J (y , x , H) can be the posterior
pdf p( x | y , H), the logarithm of the likelihood function p(y | H, x ), the mean square error (as
in Sadough, Khalighi & Duhamel (2009); Sadough & Khalighi (2007)), etc. Assume a channel
estimator in which the statistics of the estimation errors are known. Such a scenario occurs
for instance in pilot-only-based PSAM channel estimation studied in Subsection 2.1 where we
saw that the estimation process can be characterized by the posterior pdf of the channel (12).
In this case, we propose a detector based on the minimization of a new cost function defined
as
J (y , x , Ĥ) = J (y , x , H) p(H| Ĥ) dH = E H|Ĥ J (y , x , H)Ĥ (19)
H
where by using the posterior distribution (12), we have averaged the cost function J over all
realizations of the unknown channel H conditioned on its available estimate Ĥ. Note that the
mismatched detector is based on the minimization of the cost function J (y , x , Ĥ). This latter
cost function is obtained by using the estimated channel Ĥ in the same metric that would be
used if the channel was perfectly known, i.e., J (y , x , H). Using the metric of (19) differs from
the mismatched detection on the conditional expectation E H|Ĥ [.] which provides a robust de-
sign by averaging the cost function J (y , x , H) over all (true) channel realizations which could
correspond to the available estimate.
Consider the problem of detecting symbol vector x from the observation model (1) in the ML
sense, i.e., so as to maximize the likelihood function p(y | H, x ).
It is well known that under perfect channel knowledge and i.i.d. Gaussian noise, detecting x
by maximizing the likelihood p(y | x , H) is equivalent to minimizing the Euclidean distance
DML as
x̂x ML (H) = arg min DML ( x , y , H) , (20)
s0 , ..., s M −1 ∈ C
with DML ( x , y , H) − log p(y | H, x ) ∝ y − H x 2 , where ∝ means “is proportional to” and
C denotes the set of constellation symbols of size M. Assuming B bits per symbol, we have
M = 2B .
The detection rule (20) requires the knowledge of the perfect channel matrix H. The sub-
optimal mismatched ML detector consists in replacing the exact channel by its estimate Ĥ
in the receiver metric as

x̂x MM (Ĥ) = arg min DMM ( x , y , Ĥ ) = arg min y − Ĥ x 2 ,
s0 , ..., s M −1 ∈ C s0 ,...,s M −1 ∈ C
where
DMM ( x , y , H) DML ( x , y , H) , (21)
H = Ĥ
and the subscript .MM denotes mismatched. Obviously, the sub-optimality of this detection
technique is due to the mismatch introduced by the channel estimation errors; while the deci-
sion metric is derived from the likelihood function p(y | H, x ) conditioned on the perfect chan-
nel H, the receiver uses an estimate Ĥ different from H in the detection process.
As an alternative to this mismatched detection, an improved ML detection metric is proposed
in Tarokh et al. (1999); Taricco & Biglieri (2005). This metric is based on modified likelihood
p(y | Ĥ, x ) which is conditioned on the imperfect channel Ĥ. The pdf p(y | Ĥ, x ) can be derived
as follows:

p(y | Ĥ, x ) = p(y , H| Ĥ, x ) dH = p(y | H, x ) p(H| Ĥ) dH = E H|Ĥ p(y | H, x ) Ĥ ,
H ∈C H ∈C
(22)
where p(H| Ĥ) is the channel posterior distribution of equation (12) and C denotes the set of
complex matrices of size ( M R × M T ). In fact, equation (22) shows that p(y | Ĥ, x ) can be simply
derived from the general formulation in (19). It is shown in Sadough & Duhamel (2008) that
the averaged likelihood in (22) is shown to be a complex Gaussian distributed vector given by

p(y | Ĥ, x ) = CN mM , ΣM , (23)
where mM = δ Ĥ x , and ΣM = Σz + δ Σe x 2 . Finally, the estimate of the symbol x is

x̂x M (Ĥ) = arg min DM ( x , y , Ĥ) , (24)
s0 , ..., s M −1 ∈ C
where

y − δ Ĥ x 2
DM ( x , y , Ĥ) − log p(y | x , Ĥ) = M R log π σz2 + δ σe2 x 2 + 2 (25)
σz + δ σe2 x 2
is referred to as the improved ML decision metric under imperfect CSIR.
Note that when CSIR tends to the exact value, which is obtained when the number of pilot
symbols tends to infinity, we have δ → 1, σe2 → 0, and the improved metric (25) tends to the
mismatched metric:

DM x , y , Ĥ
lim = 1. (26)
N → ∞ DMM x , y , Ĥ
In the following two sections, we apply the proposed receiver design method of equation (19)
for improving the performance of two usually-used MIMO receivers working under imperfect
channel estimation: one based on the maximum a posteriori (MAP) criterion, and the other on
Soft-PIC. For both cases, we consider the simple spatial multiplexing as the space-time scheme
at the transmitter, and iterative MIMO detection and channel decoding at the receiver.
6. Reception Scheme I: Iterative MAP Detection

Here, we consider MIMO signal detection based on the MAP algorithm. In the following,
we make use of the improved ML metric derived in the previous section to modify the MAP
detector part for the case of imperfect CSIR Sadough et al. (2007). Let us denote by x [ k] and
y [ k], the transmitted and received symbol vectors corresponding to the the time slot k, simply
by x k and y k , respectively. Also, let dk,j denote the j-th (j = 1, ..., BM T ) coded and interleaved
bit corresponding to x k . We denote by L (dk,j ) the coded log-likelihood ratio (LLR) of the bit
dk,j at the output of the detector. Conditioned on the imperfect CSIR Ĥk , L (dk,j ) is given by:

Pdem dk,j = 1 y k , Ĥk
L (dk,j ) = log , (27)
Pdem dk,j = 0 y k , Ĥk

where Pdem (dk,j y k , Ĥk ) is the probability of transmission of dk,j at the detector output. We
partition the set C that contains all possibly-transmitted symbol vectors x k into two sets C0m
and C1m , for which the j-th bit of x k equals “0” or “1”, respectively. We have:
BM T
∑ e−D M ( x k , y k , Ĥk ) ∏ Pdec
1 dk,i
x k ∈ C1m i= 1
i= j
L (dk,j ) = log , (28)
BM T
∑ e−D M ( x k , y k , Ĥk ) ∏ Pdec
0 dk,j
x k ∈ C0m i= 1
i= j
where Pdec 1 ( d ) and P0 ( d ) are prior probabilities on the bit d

k,j dec k,j k,j coming from the SISO
decoder.
Note that using the metric DM ( x k , y k , Ĥk ) for the evaluation of the LLRs in (28) is an alter-
native to using the mismatched ML metric DMM ( x k , y k , Ĥk ) which replaces at each iteration,
the exact channel Hk by its estimate Ĥk in DML ( x k , y k , Hk ). By doing so, the LLRs are adapted
to the imperfect channel knowledge available at the receiver and consequently the impact
of channel uncertainty on the SISO decoder performance is reduced. We refer to the latter
approach as improved MAP detector Sadough et al. (2007).
The summations in (28) are taken over the product of the likelihood p(y k | x k , Ĥk ) =
e−D M ( x k ,yy k ,Ĥk ) given a symbol x k and the estimated channel coefficient Ĥk , and of the a priori
probability on x k (the term ∏ Pdec ), fed back from the SISO decoder at the previous iteration.
In this latter term, the a priori probability of the bit dk,j itself has been excluded, so as to let the
exchange of extrinsic informations between the channel decoder and the soft detector. Also,
note that this term assumes independent coded bits dk,j , which is a reasonable approximation
for random interleaving of large size. At the first iteration, no a priori information is available
on bits dk,j , therefore the probabilities Pdec 0 ( d ) and P1 ( d ) are set to 1/2. The decoder ac-
k,j dec k,j
cepts the LLRs of all coded bits and computes the LLRs of information bits, which are used
for decision, at the last iteration.
6.1 Case study

We now present some numerical results. First, the BER performance of the improved and
mismatched detectors are compared. Let us first address the case of BICM iterative decoding
with 16-QAM and Gray labeling for a 2 × 2 MIMO channel. It can be seen from Fig. 7 that for
Np = 2 (the shortest possible training sequence), the improvement in terms of required Eb /N0
The interesting point is that the two detectors have almost the same convergence trend
and the major improvement is obtained after the second iteration for both detectors
Sadough, Khalighi & Duhamel (2009). So, if for the reasons of complexity reduction, we only
process two receiver iterations, we still have a considerable performance gain by using the
improved detector.
0
10
Mismatched
−1 Improved
10
Perfect CSIR
−2
10
N =2
BER
P
N =8
P
−3
10
N =4
P
−4
10
4 6 8 10 12 14
E /N (dB)
b 0
Fig. 9. BER performance of improved and mismatched iterative Soft-PIC; (2 × 2) MIMO with
MUX ST scheme, i.i.d. Rayleigh block-fading channel with 4 fades per frame, QPSK modula-
tion, training sequence length Np ∈ {2, 4, 8}.
8. Conclusions
We studied in this chapter the interaction between iterative data detection and channel es-
timation in realistic wireless communication systems where the receiver disposes only of an
imperfect estimate of the unknown channel parameters. To obtain the CSIR, we considered
different recent and classically-used techniques. First, we presented pilot-only based channel
estimation and showed that an accurate estimate of the channel through this method would
require a large number of pilots per frame, which can result in a considerable loss in the sys-
tem data throughput. Overlay pilots may be preferred to time-multiplexed solution from this
point of view, however, the quality of channel estimate is, in general, worse, as compared to
PSAM. We also presented semi-blind channel estimation methods that, in addition to pilot
symbols, make use of the data symbols in the estimation process. Although iterative semi-
blind channel estimation outperforms pilot-only assisted channel estimation, it has a higher
complexity, which may be of critical concern for practical implementations.
0
10
N =2
p
−1
10
N =4
p
−2
10
−3
10
BER
−4
10
perfect CSIR
N =8
p
−5
10
Dashed line: mismatched

−6
10 plain line: improved
−7
10
2 3 4 5 6 7 8 9 10
E /N (dB)
b 0
Fig. 8. BER performance improvement over (2 × 2) MIMO channel with i.i.d. Rayleigh fad-
ing for various training sequence lengths. 16-QAM modulation with set-partition labeling,
iterative MAP detection after four receiver iterations.
plified formulation of Soft-PIC, which assumes perfect interference cancellation after the first
iteration. In this section, however, we consider the exact formulation of Soft-PIC. In order
to better understand the formulation of the improved detector, we present in the following
the formulation of (exact) Soft-PIC under perfect channel knowledge at the receiver. Then,
we present the improved Soft-PIC detector in the presence of channel estimation errors in
Subsection 7.2.
7.1 Soft-PIC detection under perfect channel knowledge

Consider the general block diagram of Fig. 2. Here, to detect a symbol transmitted from
a given antenna, we first make use of the soft information available from the SISO channel
decoder to reduce and hopefully to cancel the interfering signals arising from other transmit
antennas. At the first iteration where this information is not available, we perform a classical
MMSE filtering.
Let us consider the transmitted vector x k = [ x1k , ..., xkMT ] T at time k and assume that we are
j
interested in the detection of its i-th symbol xki . We start by evaluating the parameters x̂k and
j
σ2j for the interfering symbols xk , j = i, from the SISO decoder as follows:
xk
j 2B
j j j
x̂k = E xk = ∑ xk P [ xk ] (29)
j =1
j 2B
j j
σ2j = E | xk |2 = ∑ | xk |2 P [ xk ] (30)
xk
j =1
j j
where P [ xk ] is the probability of the transmission of xk and is evaluated using the probabilities
j,n
Pdec (dk ) at the decoder output:
B
j j,n
P [ xk ] = K ∏ Pdec (dk ),
n =1
where K is a normalization factor. We further introduce the following definitions.

H i is the ( M R × ( M T − 1)) matrix constructed from H by discarding its i-th column, namely
h i . We also define the (( M T − 1) × 1) vectors
T
x ik x1k , x2k , ..., xki−1 , xki+1 , ..., xkMT
and T
x̂x ik x̂1k , x̂2k , ..., x̂ki−1 , x̂ki+1 , ..., x̂kMT ,
j
where x̂k are estimated in (29).
Now, given the received signal vector y k , a soft interference cancellation is performed on y k
for detecting the symbol xki by subtracting to y k the estimated signals of the other transmit
antennas as Sadough, Khalighi & Duhamel (2009):
y i = y k − Hi x̂x ik = h i xki + Hi x ik − H i x̂x ik + z k , for i = 1, ..., M T . (31)

k
j j
Except under perfect prior information on the symbols which leads to x̂k = xk , there remains a
residual interference in y i . In order to reduce further this interference, an instantaneous linear
k
MMSE filter wik is applied to y i to minimize the mean square value of the error eik defined as
k
eik = xki − rki (32)
where the filter output rki is equal to

rki = wik y i . (33)
k
Here, wik is obtained as
2
wik = arg min E x k ,zz k xki − wik y i . (34)
k
wik ∈C M R
By invoking the orthogonality principle Scharf (1991), the coefficients of the MMSE filter wik
are given by
−1
Hi Λk,i − Λ k,i H† σz2
i
wik = h †i h i h †i + + I MR (35)
σx2i σx2i
k k
where

†
Λk,i = E x ik x ik ≈ diag E | x1k |2 , ..., E | xki−1 |2 , E | xki+1 |2 , ..., E | xkMT |2 , and

k,i = x̂x i x̂x i † ≈ diag | x̂1 |2 , ..., | x̂ i−1 |2 , | x̂ i+1 |2 , ..., | x̂ MT |2 .
Λ k k k k k k
Note that the off-diagonal entries in Λk,i and Λ k,i have been neglected to reduce the complexity
without causing significant performance loss Lee et al. (2006).
At the first decoding iteration, we have no prior information available on the transmitted data,
k,i = 0 M −1 . Consequently, (35) reduces to
i.e., Λk,i = σx2i I MT −1 and Λ T
k
−1
σ2
wik = h †i HH† + 2z I MR (36)
σx i
k
which is no more than the linear MMSE detector for xki .

Before passing the detected symbols rk to the SISO decoder, we convert them to LLR. This is
done assuming a Gaussian distribution for the residual interference after Soft-PIC detection
(see Wang & Poor (1999) for details on the LLR conversion).
7.2 Improved Soft-PIC Detection Under Imperfect Channel Estimation

As we see from (31) and (35), we need the channel H for both interference canceling and
MMSE filtering. As the receiver has only an imperfect channel estimate Ĥ, the suboptimal
mismatched solution consists in replacing Hi and h i in (31) and (35) by their estimates Ĥi and
ĥ i , respectively. As a first step toward a realistic design, we make use of the available channel
estimate Ĥ for interference cancellation. That is, equation (31) is rewritten as
y i = y k − Ĥi x̂x ik = h i xki + Hi x ik − Ĥi x̂x ik + z k , for i = 1, ..., M T (37)

k
where Ĥi is the ( M R × ( M T − 1)) matrix constructed from Ĥ by discarding its i-th column,
namely ĥ i . We note that (37) naturally depends on the unknown channel matrix H of which
the receiver has only an imperfect estimate available. Instead of replacing the unknown
channel by its estimate (i.e., the mismatched approach), we use the posterior distribution
(12) and make two modifications to the detector described in Subsection 7.1, as follows (see
Sadough, Khalighi & Duhamel (2009) and Sadough & Khalighi (2007) for more details).
The first modification concerns the design of the filter wik in (34). The modified filter w ik should
minimize the average of the mean square error over all realizations of channel estimation
errors. In other words,

2 2
w ik y i Ĥ = arg min E H|Ĥ E x k ,zzk xki − w
ik = arg min E H,xx k ,zzk xki − w ik y i (38)
k k
ik ∈C M R
w ∈C M R
w
where we have assumed the independence between H, x k , and z k . After some simple alge-
braic manipulations Sadough, Khalighi & Duhamel (2009); Scharf (1991), we obtain:
−1
ik = R x i y i Ry i
w (39)
k k k
where
† + (δ − 1) mk,i Ĥ†i
Rx i y i = δ σx2i h (40)
k k k
i
†
with mk,i = x̂ki x̂x ik and δ is given by (15), and
Ry i = δ2 σx2i h † + δ2 Ĥi Λk,i Ĥ†i + (δ2 − δ) h

ih i mk,i H† + (δ2 − δ) Ĥi m† h†
k k
i i k,i i

+ (1 − 2δ) Ĥ i Λ k,i Ĥ†i + σz2 + (1 − δ) σ2i + (1 − δ) tr(Λk,i ) I M . (41)
x k
R
To get more insight on the proposed detector, let us consider the ideal case where perfect
channel knowledge is available at the receiver, i.e., Ĥ = H and σe2 = 0. We note that in this case,
δ = 1 and the posterior pdf (12) reduces to a Dirac delta function; consequently the two filters
w ik and wik coincide. Similarly, under near-perfect CSIR, obtained either when σe2 → 0 or when
Np → ∞, we have δ → 1, and the filter w ik gives a similar expression as wik in (35). However, in
the presence of estimation errors, the proposed improved and mismatched detectors become
different due to the inherent averaging in (38), which provides a robust design that adapts
itself to the channel estimate available at the receiver.
The second modification concerns the application of the derived filter w ik to the received signal
y . As this latter depends on H (see (37)), we average the filter output rki as follows:
i
k
rki = E H|Ĥ [ rki ] = δ w

i xi + δ w
i h ik Ĥi x̂x ik + w
i Ĥi x ik − w ik z k = µ k,i xki + ηk,i , (42)
k k k
µ k,i ηk,i
where ηk,i contains interference and noise. From (42) it is clear that the output of the im-
proved MMSE filter can be viewed as an equivalent AWGN channel having xki at its input.
The parameters µ k,i and ση2k,i are calculated at each time-slot by using the symbols statistics.
In order to transform the detected symbols at the output of the MMSE filter to LLRs on the cor-
responding bits, we approximate ηk,i by a zero-mean Gaussian random variable with variance
ση2k,i (see Sadough, Khalighi & Duhamel (2009) for details on the calculation of this variance).
Let dki,m denote the m-th (m = 1, ..., B) bit corresponding to xki . The LLR on dki,m is given by:

|r i − µ k,i x i |2 B
1 ( di,n )
∑ exp − k σ 2 k ∏ Pdec k
x ki ∈S 1m
ηk,i
Pdem dki,m = 1|
n =1
rki , µ k,i n=m
L (dki,m ) = log i,m = log . (43)
Pdem dk = 0| rki , µ k,i |r i − µ k,i x i |2 B
0 ( di,n )
∑ exp − k σ 2 k ∏ Pdec k
ηk,i
x ki ∈S 0m n =1
n=m
Note that here the cardinality of the sets S 1m and S 0m equals 2B −1 .
7.2.1 Case study

Figure 9 shows BER curves of the mismatched and improved receivers for the case of QPSK
modulation and a (2 × 2) MIMO system. The number of channel uses for pilot transmission
is Np ∈ {2, 4, 8}. As a reference, we have also presented the BER curve for the case of perfect
CSIR. We observe that the gain in SNR of the improved detector to attain the BER of 10−5 is
about 1.4 dB, 0.5 dB, and 0.2 dB, respectively for Np = 2, 4, and 8.
The interesting point is that the two detectors have almost the same convergence trend
and the major improvement is obtained after the second iteration for both detectors
Sadough, Khalighi & Duhamel (2009). So, if for the reasons of complexity reduction, we only
process two receiver iterations, we still have a considerable performance gain by using the
improved detector.
0
10
Mismatched
−1 Improved
10
Perfect CSIR
−2
10
N =2
BER
P
N =8
P
−3
10
N =4
P
−4
10
4 6 8 10 12 14
E /N (dB)
b 0
Fig. 9. BER performance of improved and mismatched iterative Soft-PIC; (2 × 2) MIMO with
MUX ST scheme, i.i.d. Rayleigh block-fading channel with 4 fades per frame, QPSK modula-
tion, training sequence length Np ∈ {2, 4, 8}.
8. Conclusions
We studied in this chapter the interaction between iterative data detection and channel es-
timation in realistic wireless communication systems where the receiver disposes only of an
imperfect estimate of the unknown channel parameters. To obtain the CSIR, we considered
different recent and classically-used techniques. First, we presented pilot-only based channel
estimation and showed that an accurate estimate of the channel through this method would
require a large number of pilots per frame, which can result in a considerable loss in the sys-
tem data throughput. Overlay pilots may be preferred to time-multiplexed solution from this
point of view, however, the quality of channel estimate is, in general, worse, as compared to
PSAM. We also presented semi-blind channel estimation methods that, in addition to pilot
symbols, make use of the data symbols in the estimation process. Although iterative semi-
blind channel estimation outperforms pilot-only assisted channel estimation, it has a higher
complexity, which may be of critical concern for practical implementations.
Regardless of the channel estimation technique, an important point is the impact of estimation
errors on the receiver performance. The usually-used approach is to consider the (imperfect)
channel estimate as perfect and to use it in data detection. We called this the mismatched
approach. In such case, we saw that, the impact of estimation errors is somehow similar for
orthogonal and non-orthogonal space-time schemes. We then considered the improved ap-
proach by which we take into account the channel estimation inaccuracies in data detection.
More precisely, by using the statistics of the channel estimation errors, we use a new detection
rule instead of the sub-optimal mismatched detector. Applying this detection design rule to
MAP and Soft-PIC detectors, we showed that a significant improvement can be obtained as
compared to the mismatched detector. Finally, it is worth mentioning that adopting the im-
proved reception scheme does not increase considerably the complexity. In fact, the improved
detectors require just a few more matrix additions and multiplications, which does not have
an important impact on the receiver complexity.
9. References
Adireddy, S., Tong, L. & Viswanathan, H. (2002). Optimal placement of training for frequency-
selective block-fading channels, IEEE Trans. Inf. Theory 48(8): 2338–2353.
Alamouti, S. (1998). A simple transmit diversity technique for wireless communications, IEEE
J. Sel. Areas Commun. 16(8): 1451–1458.
Bahl, L., Cocke, J., Jelinek, F. & Raviv, J. (1974). Optimal decoding of linear codes for minimiz-
ing symbol error rate, IEEE Trans. Inform. Theory pp. 284–287.
Balakrishnan, J., Rupp, M. & Viswanathan, H. (2000). Optimal channel training for multiple
antenna systems, in Proc. MMT . Miami, FL.
Belfiore, J.-C., Rekaya, G. & Viterbo, E. (2005). The golden code: a 2 × 2 full-rate space-time
code with nonvanishing determinants, IEEE Trans. Inform. Theory 51(4): 1432–1436.
Berthet, A., Unal, B. & Visoz, R. (2001). Iterative decoding of convolutionally encoded signals
over multipath Rayleigh fading channels, IEEE J. Sel. Areas Commun. 19(9): 1729–1743.
Bohlin, P. & Tapio, M. (2004). Optimized data aided training in MIMO systems, in Proc. VTC
pp. 679–683. Milan, Italy.
Cavers, J. K. (1991). An analysis of pilot symbol assisted modulation for Rayleigh fading
channels, IEEE Trans. Veh. Technol. 40(4): 686–693.
Cui, T. & Tellambura, C. (2005). Superimposed pilot symbols for channel estimation in OFDM
systems, in Proc. Globecom pp. 2229–2233. St. Louis, MO.
de Carvalho, C. & Slock, D. T. M. (2001). Semi-blind Methods for FIR Multi-channel Estimation, in
Signal Processing Advances in Wireless and Mobile Communications; Vol. I: Trends
in Channel Estimation and Equalization, Editors: G. B. Giannakis, Y. Hua, P. Stoica,
and L. Tong, Upper Saddle River, NJ: Prentice Hall.
Dempster, A., Laird, N. & Rubin, D. (1977). Maximum-likelihood from incomplete data via
the EM algorithm, J. Royal Statatistical Society 39(1): 1–38.
Dong, M. & Tong, L. (2002). Optimal design and placement of pilot symbols for channel
estimation, IEEE Trans. Signal Processing 50(12): 3055–3069.
Ghogho, M., McLernon, D., Alameda-Hernandez, E. & Swami, A. (2005). Channel estimation
and symbol detection for block transmissionusing data-dependent superimposed
training, IEEE Signal Processing Letters 12(3): 226–229.
Giannakis, G. B., Hua, Y., Stoica, P. & Tong, L. (2001). Signal Processing Advances in Wireless
and Mobile Communications, Vol. 1, Trends in Channel Estimation and Equalization,
Prentice Hall PTR, Upper Saddle River, NJ.
Hassibi, B. & Hochwald, B. M. (2003). How much training is needed in multiple-antenna

wireless links?, IEEE Trans. Inform. Theory 49(4): 951–963.
Hoeher, P. & Tufvesson, F. (1999). Channel estimation with superimposed pilot sequence, in
Proc. Globecom 4: 2162–2166. Rio de Janeiro, Brazil.
Jungnickel, V., Haustein, T., Jorswieck, E., Pohl, V. & von Helmolt, C. (2001). Performance of a
MIMO system with overlay pilots, in Proc. Globecom 1: 594–598. San Antonio, TX.
Khalighi, M. A., Berriche, L. & Hélard, J.-F. (2005). Overlaid or time-multiplexed pilots for
channel estimation in iterative MIMO receivers, in Proc. ISSPA pp. 483–486. Sydney,
Australia.
Khalighi, M. A. & Bourennane, S. (2007). Semi-blind channel estimation based on superim-
posed pilots for single-carrier MIMO systems, in Proc. VTC pp. 1480–1484. Dublin,
Ireland.
Khalighi, M. A. & Bourennane, S. (2008). Semi-blind single-carrier MIMO channel estimation
using overlay pilots, IEEE Trans. Veh. Technol. 57(5): 1951–1956.
Khalighi, M. A., Boutros, J. & Hélard, J.-F. (2006). Data-aided channel estimation for Turbo-PIC
MIMO detectors, IEEE Commun. Lett. 10(5): 350–352.
Khalighi, M. A. & Boutros, J. J. (2006). Semi-blind channel estimation using EM algorithm in
iterative MIMO APP detectors, IEEE Trans. Wireless Commun. 5(11): 3165–3173.
Khalighi, M. A., Hélard, J.-F., Sadough, S. M. S. & Bourennane, S. (2009). Suitable combination
of channel coding and space-time schemes for moderate-to-high spectral efficiency
mimo systems, AEÜE International Journal of Electronics and Communications . to ap-
pear.
Lam, C., Falconer, D. & Danilo-Lemoine, F. (2008). Iterative frequency domain channel estima-
tion for DFT-precoded OFDMsystems using in-band pilots, IEEE Journal on Selected
Areas in Communication 26(2): 348–358.
Lee, H., Lee, B. & Lee, I. (2006). Iterative detection and decoding with improved V-BLAST for
MIMO-OFDM systems, IEEE J. Selected Areas Commun. 24(3): 504–513.
Ma, X., Giannakis, G. & Ohno, S. (2003). Optimal training for block transmission over doubly
selective wireless fading channels, IEEE Trans. Signal Processing 51(5): 1351–1366.
Meng, X. & Tugnait, J. (2004). MIMO channel estimation using superimposed training, in Proc.
ICC pp. 2663–2667. Paris, France.
Moon, T. (1996). The expectation-maximization algorithm, IEEE Signal Processing Mag.
13(6): 47–60.
Moon, T. K. & Stirling, W. C. (2000). Mathematical Methods and Algorithms for Signal Processing,
Upper Saddle River, NJ: Prentice Hall.
Raoof, K., Khalighi, M. A. & Prayongpun, N. (2008).
Sadough, S. M. S., Ichir, M. M., Duhamel, P. & Jaffrot, E. (2009). Wavelet based semi-blind
channel estimation for ultra wideband OFDM systems, IEEE Trans. Veh. Technol.
58(3): 1302–1314.
Sadough, S. M. S. & Khalighi, M. A. (2007). Optimal turbo-blast detection of MIMO-OFDM
systems with imperfect channel estimation, IEEE International Symposium on Personal,
Indoor and Mobile Radio Communications .
Sadough, S. M. S., Khalighi, M. A. & Duhamel, P. (2009). Improved iterative MIMO signal de-
tection accounting for channel estimation errors, IEEE Trans. Veh. Technol. . to appear.
Sadough, S., Piantanida, P. & Duhamel, P. (2007). MIMO-OFDM optimal decoding and achiev-
able information rates under imperfect channel estimation, in Proc. Signal Process.
Advances Wireless Commun. (SPAWC) .
Scharf, L. (1991). Statistical Signal Processing, Addison-Wesley.
Sellathurai, M. & Haykin, S. (2002). Turbo-BLAST for wireless communications: theory and
experiments, IEEE Trans. Signal Process. 50(10): 2538– 2546.
Tapio, M. & Bohlin, P. (2004). A capacity comparison between time-multiplexed and superim-
posed pilots, in Proc. Asilomar Conf. on Signals, Systems and Computers pp. 1049–1053.
Pacific Grove, CA.
Taricco, G. & Biglieri, E. (2005). Space-time decoding with imperfect channel estimation, IEEE
Trans. Wireless Commun. 4(4): 1874–1888.
Tarokh, V., Naguib, A., Seshadri, N. & Calderbank, A. R. (1999). Space-time codes for high
data rate wireless communication: Performance criteria in the presence of channel
estimation errors, mobility, and multiple paths, IEEE Trans. Commun. 47: 199–207.
Telatar, E. (1999). Capacity of multi-antenna Gaussian channels, European Trans. Telecommun.
10(6): 585–595.
Visoz, R. & Berthet, A. (2003). Iterative decoding and channel estimation for space-time
BICM over MIMO block fading multipath AWGN channel, IEEE Trans. Commun.
51(8): 1358– 1367.
Wang, X. & Poor, H. V. (1999). Iterative (turbo) soft interference cancellation and deoding for
coded CDMA, IEEE Trans. Commun. 47(7): 1046–1061.
Zhu, H., Farhang-Boroujeny, B. & Schlegel, C. (2003). Pilot embedding for joint channel es-
timation and data detection in MIMO communication systems, IEEE Commun. Lett.
7(1): 30–32.
Cooperative MIMO Systems in Wireless Sensor Networks 123
Cooperative MIMO Systems in

Wireless Sensor Networks
M. Riduan Ahmad1, Eryk Dutkiewicz2, Xiaojing Huang3 and M. Kadim Suaidi4
2Macquarie
1,4UniversitiTeknikalMalaysia 3CSIROICTCentre
Melaka, University,
1,4Malaysia,
2,3Australia
1. Introduction
The Multiple-Input Multiple-Output (MIMO) term originally describes the use of the
multiple antennas concept or exploitation of spatial diversity techniques. In early research
work, the MIMO concept was proposed to fulfil the demand for providing reliable high-
speed wireless communication links in harsh environments. Subsequently, MIMO
technology has been proposed to be used in wireless local area networks and cellular
networks, particularly at the base station and access point sides to tackle the challenges of
low transmission rates and low reliability with no constraints on energy efficiency. In
contrast, Wireless Sensor Networks (WSNs) have to deal with energy constraints due to the
fact that each sensor node depends on its battery for its operation. In harsh environments,
sensor nodes must be provided with reliable communication links. However, current WSN
design requirements do not require high transmission rates.
The concept of cooperative MIMO was introduced in WSNs by utilizing the collaborative
nature of dense sensor nodes with the broadcast wireless medium to provide reliable
communication links in order to reduce the total energy consumption for each sensor node.
Therefore, instead of using multiple antennas attached to one node or device such in the
traditional MIMO concept, cooperative MIMO presents the concept of multiple sensor nodes
cooperating to transmit and/or receive signals. Multiple sensor nodes are physically
grouped together to cooperatively transmit and/or receive. Within a group, sensor nodes
can communicate with relatively low power as compared to inter-group communication.
Furthermore, by using this cooperative MIMO concept, we can provide the advantages of
traditional MIMO systems to WSNs, particularly in terms of energy efficient operation.
This chapter introduces the concepts of diversity techniques and their relationship with
cooperative MIMO systems and to discuss their practicality for implementation in WSNs.
The approaches that we are going to use in this chapter are the comparative study and
performance evaluation of major diversity techniques and their implementations in
cooperative MIMO systems. The outcomes can be used as the basis for further study in
order to find the most suitable cooperative MIMO scheme to be implemented in the WSN
environment.
The rest of the chapter is organized as follows. Section 2 introduces the concept of
cooperative MIMO and Section 3 explains various types of diversity techniques proposed in
the current literature. A comparative study between the major diversity techniques is
presented in Section 4 and this is followed by performance evaluation in Section 5. Finally,
in Section 6 we conclude the chapter.
2. Cooperative MIMO Concepts

The MIMO term originally describes the use of the multiple antennas concept or
exploitation of spatial diversity techniques. In early research work, the MIMO concept was
proposed to fulfil the demand for providing reliable high-speed wireless communication
links in harsh environments. Subsequently, MIMO technology has been proposed to be used
in wireless local area networks and cellular networks (Proakis, 2001), particularly at the base
station and access point sides to tackle the challenges of low transmission rates and low
reliability with no constraints on energy efficiency. In contrast, WSNs have to deal with
energy constraints due to the fact that each sensor node depends on its battery for its
operation. In harsh environments, sensor nodes must be provided with reliable
communication links. However, current WSN design requirements do not require high
transmission rates.
The concept of cooperative MIMO was introduced in WSNs by utilising the collaborative
nature of dense sensor nodes with the broadcast wireless medium to provide reliable
communication links in order to reduce the total energy consumption for each sensor node.
Therefore, instead of using multiple antennas attached to one node or device such in the
traditional MIMO concept, cooperative MIMO presents the concept of multiple sensor nodes
cooperating to transmit and/or receive signals. Multiple sensor nodes are physically
grouped together to cooperatively transmit and/or receive. Within a group, sensor nodes
can communicate with relatively low power as compared to inter-group communication
(Singh and Prasanna, 2003; Gupta and Younis, 2003; Yuksel and Erkip, 2004). Furthermore,
by using this cooperative MIMO concept, we can provide the advantages of traditional
MIMO systems to WSNs, particularly in terms of energy efficient operation.
3. Techniques of Diversity
In this section we discuss various diversity techniques to reduce the deep fading problem in
WSNs which requires higher retransmission rates. By tackling this problem, we are clearly
satisfying two design requirements: energy efficient operation and higher communication
link reliability. It is important to observe that deep fading contributes to packet errors (if a
portion of the packet is affected) or packet loss (if the whole packet is totally lost which can
be common as data packets in WSNs are normally small (Kohvakka et. al., 2006; Karl and
Willig, 2007). The basic concept of diversity is to provide the receiver with copies of
independently faded transmitted packets with the hope that at least one of these copies will
be received correctly. Diversity can be implemented in various ways such as frequency
diversity, spatial diversity, time diversity, modulation diversity and polarisation diversity to
suit different design requirements.
Frequency diversity is achieved when the same signal is transmitted over different
frequency bands. The separation of the frequency bands has to be more than the coherence
bandwidth of the channel (Duman and Ghrayeb, 2007). Time diversity is achieved when the
same signal is transmitted with redundancy using different time intervals. The separation of
the time intervals has to be more than the coherence time of the channel. Also, time diversity
can be achieved by means of channel coding. The idea is to transmit the different parts of
the codeword corresponding to a particular symbol using different time intervals. The most
practical channel codes discussed in the literature include block codes, convolutional codes
and trellis-coded modulation (Duman and Ghrayeb, 2007).
Spatial diversity is achieved by the use of multiple antennas or nodes at either end or at both
ends of the MIMO communication link. The separation between the antennas or nodes has
to be more than half a wavelength in a uniform scattering environment. Systems with
multiple antennas are also referred to as MIMO systems (Duman and Ghrayeb, 2007).
Therefore we can refer to systems with multiple nodes as cooperative MIMO systems.
Spatial diversity gains increase channel capacity which leads to higher data throughputs
and significant improvement in data transmission reliability. These advantages are achieved
without any expansion of bandwidth or higher transmit power which makes this technique
very suitable to be implemented in energy constrained WSNs.
Diversity techniques can be combined to achieve greater improvements in reliability and
achievable transmission rates. Perhaps the most popular combination technique is between
space diversity and time diversity by using channel coding. The combination yields the
space-time coding (STC) scheme. The variants of the STC scheme depend on the channel
coding being used. For example, space-time block coding (STBC) schemes are based on
block coding and space-time trellis coding scheme (STTC) schemes are based on trellis-
coded modulation.
Multiple antennas or nodes can be exploited in different ways at both ends of the MIMO
communication link. Early work in this area concerned designs of multiple antennas at the
receiver side to achieve receive spatial diversity in order to boost link reliability as the
number of receiving antennas grows. Among the earliest users of receive spatial diversity
schemes are mobile communications systems to improve uplink performance by
implementing multiple receive antennas at the base station (Proakis, 2001). If only a single
transmit antenna and multiple receive antennas are used, the resulting system is referred to
as a Single-Input Multiple-Output (SIMO) system.
In later work transmit spatial diversity was achieved by exploiting multiple transmit
antennas with proper coding or weighting of the transmitted data signals. It is important to
note that both spatial diversity schemes achieve improved transmission reliability at the cost
of transmission rates comparable to the Single-Input Single-Output (SISO) systems. Clearly
the achievement of higher link reliability is a trade-off with transmission rates. When
multiple transmit antennas and single receive antenna are used, the resulting system is
referred to as a Multiple-Input Single-Output (MISO) system.
Further research to achieve higher transmission rates and higher capacity has been done
using multiple antennas at both ends of the communications link. These multiple transmit
antennas and multiple receive antenna systems are referred to as MIMO systems. One of the
common techniques to boost the transmission rates is to provide the receivers with
independent streams of the same data signal from different transmit antennas. In this way,
the transmit antennas are exploited to boost the transmission rates at the cost of lower link
reliability. However, when operating under certain constraints, the same scheme can
achieve full diversity gain leading to higher link reliability.
A comparison between the different spatial diversity schemes discussed above is shown in
Table 1 where M and N denote the number of transmit antennas and receive antennas,
respectively.
Schem
M N Example Benefits
e
1 1 No transmit or receive No diversity gain.
SISO
diversity
SIMO 1 >1 Receive diversity Diversity proportional to N .
MISO >1 1 Transmit diversity Diversity proportional to M .
MIMO >1 >1 Use of multiple antennas at Diversity proportional to the
both the transmitter and product of M and N .
receiver
Array gain (coherent combining
assuming prior channel
estimation).
Table 1. Comparison of main spatial diversity schemes
4. Multiple Nodes in Wireless Sensor Networks

A major design requirement of WSNs is to reduce the total energy consumption of the
sensor nodes. The transmission power can be reduced by providing the highest diversity
gain possible which leads to higher link reliability and thus lower the retransmission rates.
Therefore the exploitation of multiple nodes in WSNs (referred to as cooperative MIMO) is
inevitable in order to provide higher reliability communication link and reduce transmission
power.
Most of the previous work in the area of cooperative MIMO has assumed that the
cooperating sensor nodes are perfectly synchronised during transmission and reception
(Jagannathan et. al., 2004). Recently, the impact of imperfect synchronisation effects on the
performance of cooperative MIMO operation in WSNs has gained more attention (Li, 2004;
Li et. al., 2004; Li and Hwu, 2006). Imperfect synchronisation could occur due to the lack of
carrier synchronisation or because of imperfect timing in frame and bit level
synchronisation. In this chapter, we consider the impact of imperfect synchronisation caused
by clock jitter alone. Each cooperating transmit node from a set of M cooperative nodes
experiences clock jitter with the jitter around a reference clock, T o denoted as T jm where
1≤m ≤M . The detailed system model of clock jitter will be explained in Section 5.
The following discussion explains in detail the three major MIMO schemes in both
synchronous and asynchronous scenarios and their practicality in WSNs. Synchronous
operation assumes perfect synchronisation between cooperating transmitting nodes and
asynchronous operation refers to scenarios where imperfect synchronisation occurs. The
three MIMO schemes are:
a) SIMO System
b) MISO System
c) Spatial Multiplexing MIMO System
4.1 SIMO System

Perhapsthefirsttechniqueindiversity lyrelated
particular
tospatialutilisationisthere
diversity (SIMO) technique. The transmitter can choose to perform frequency, time, or
polarisation due to the fact that of the
diversity
sourcedoes not affect the method of
combinationatthereceiverside(with ptionthe
ofexce
transmitspatialdiversity(Jafarkhan
205). At the receiver side, more than one antenna or node must be Nused, ≥ ,2 to gain
spatial diversity which leads to higher reliability by increasing the average si
ratio(SNR)andloweringthebiterrorrate(BER).
There are four popular combiningmethods arethat
utilisedat the receiver: Maximum Ratio
Combining (MRC), Equal Gain Combining (EGC ), Selection Combining (SC) and Switched
Combining (Simon and Alouini, 20; Rappa port, 20; Jafarkhani, 205; Duman and
Ghrayeb, 207). MRC achieves diversity gain equal to the number of the receive antennas,
N , withN Radio Frequency RF) ( chains as shown in Figure .1 EGC is a special case of MRC
with equal weights' amplitudes where e received
all th signals are co-phased and then
combinedtogetherwithequalgain.TheEGC iver'
recescircuitislesscomplexbutatthecost
oflowerdiversitygainfor thanMRC.
Assume that the receiver receives N replicas of the transmitted sthrough
signal,N
independentpaths.th received
The k signalisdefinedby:
rk  hk s   k (1)
wherek= 1, 2 ……, N , k is the complex white Gaussian noise sample vector added to t
2 variance,
kthcopy of the signal with zero mean and k~ N c,(0 2 ) and hkis the complex
channelfadinggainvect orwithzeromeanand 2 variance,
hk~ N c(0, 2 ).Weassumethatthe
receiver is coherent where the channel ationinform
is known and perfectly estimated at the
 
2
s symbolis
receiver. If the average power of the transmitted
~ 2
where s is the resulting decision variable with s mean and N
variance which can
h
2
k
k 1
 
 
~ 2  . The resulting effective SNR at the output of the MRC
be represented as s ~  s, N
 2 
  hk 
 k 1 
N
h
2
block is proportional to k and given as:
k 1
   hk 
N
2    
E s
2
N
. (4)
k
k 1 2 k 1
From Equation (4), the effective SNR of the system with a receive diversity scheme is
equivalent to the sum of the instantaneous receive SNRs for N different paths. If we assume
that all the different paths have the same average SNR, then the average of the effective SNR
at the output of the MRC block is:
   E h 
N     E 2 E s
2
N
  N  k . (5)
k k

k 1
2
k 1
By increasing the number of receiving antennas N , the receive average SNR can be increased
by N -fold which leads to the lowest possible BER for the system such that at the high SNRs
regime, the error probability decays as SNRa-Nerrp N
Later, a hybrid selection/MRC technique was proposed to balance the requirements

between higher diversity gain and lower complexity (Jafarkhani, 2005).
Switched combining employs scanning and selection operation where the receiver scans all
the diversity branches and selects a particular branch with the SNR above a certain
predetermined threshold (Jafarkhani, 2005). The signals from the selected branch are
selected as the output until its SNR drops below the threshold. Then the receiver starts to
scan again and switches to another branch. This scheme is simpler since it does not require
any channel knowledge but at the cost of lower achievable diversity gain.
In the context of practicality in WSNs, a SIMO with MRC scheme is more practical and
promising for implementation as shown in Figure 3. This is due to the fact that each node in
the network represents a single path processing including the RF chain processing. It seems
that the complexity issue in the traditional SIMO approach can been reduced with the
cooperative SIMO implementation while providing the highest SNR possible. Moreover, the
transmission by the transmit node can be done without the need for time synchronisation,
thus the cooperative SIMO system is not affected by clock jitter.
On the other hand, there are other issues that we have to consider such as the fact that data
signals received by all the receiving nodes must be forwarded to a common destination
node in order to combine and decode them successfully. Moreover, the diversity gain does
not contribute to the reduction of the total transmission power and the use of N receiving
nodes can contribute to the higher circuit power in the network.
4.2 MISO System

The main motivation for using of multiple antennas at the transmitter is to reduce the
required processing power and complexity at the receiver which leads to lower power
consumption, lower size and lower cost. However, the MISO concept is not easy to exploit
and to implement (Naquib and Calderbank, 2000). Additional signal processing is required
at both the transmitters and receiver in order to correctly decode the received signals. Also,
another challenge is that the transmitter does not know the channel conditions unless
channel information is fed back by the receiver to the transmitter (Liu et. al., 2001).
A number of MISO schemes have been proposed in the literature and can be categorised
into two major classes:
a) Closed-loop MISO schemes with feedback

b) Open-loop MISO schemes without feedback
The difference between the two types of schemes is that the former relies on channel state
information which has to be fed back to the transmitter and the latter eliminate the need for
channel state information at the transmitter.
4.2.1 Closed-loop MISO System

The modulated signals are weighted with different weighting factors and transmitted with
multiple antennas M at the transmitter. The weighting factors are chosen with the assistance
of the channel state information so that the received SNR can be maximised at the receiver.
The weighting factors must be optimised in order to achieve full diversity gain. One of the
drawbacks of this system is that when the weighting factors are not optimised due to
imperfect channel estimation, the received SNR is decreased.
Two of the most popular closed-loop MISO schemes are switched diversity (Winters, 1983)
and digital beamforming (Litva and Lo, 1996). Among the two schemes, the best solution, if
the transmitter has perfect knowledge of the channel, is the MISO beamforming scheme. The
MISO beamforming scheme is less complex and easier to deploy which makes it more
practical to implement in WSNs.
RX
h1 1
RF Chain
h2
2
TX
. RF Chain
. MRC
hN
.
. .
.
. .
N .
RF Chain
Fig. 1. A system with one transmit antenna and N receive antennas with MRC.
1 RX
h1
h2 2
TX
hN
. RF
Chain
SC
.
.
select
N
Fig. 2. A system with one transmit antenna and N receive antennas with SC.
1
h1 2
h2
TX hN
N
h3
h4 3
4
Fig. 3. A cooperative receive diversity system with one transmit node and N receive nodes.
w1
TX h1
1
w2 h2 RX
Transmitted 2 RF ML
signal hM Chain
.
.
.
wM
feedback channel
Fig. 4. A beamforming transmit diversity system with M transmit antennas and 1 receive
antenna.
Let us consider a digital MISO beamforming system with M antennas at the transmitter and
one antenna at the receiver as shown in Figure 4. The transmitter transmits M weighted
transmitted signals, w k.sthrough M independent paths. The received signal is then given as:
M
r   hk wk s   (7)
k 1
In the context of practicality in WSNs, the main obstacle for MISO beamforming
implementation is the issue of how to provide each transmitting node with the knowledge
of the channel. A multi-channel approach can be used where one channel is dedicated for
the feedback and the other channel for data transmission. The channel is estimated
periodically through training sequences on the feedback channel. However, a multi-channel
approach is not practical for WSNs because such an approach increases the hardware and
processing complexity at both the transmitter and the receiver. Also, such an approach
requires tight frequency synchronisation to maintain the dual channel utilisation which
obviously increases the total energy consumption of the network.
A better and practical alternative approach is to exploit the existing control protocols in
WSNs such as those utilising RTS-CTS packets to provide the channel state information to
the transmitter as shown in Figure 5. Both the control and data communications can be
maintained over a single-channel with less complexity and loose synchronisation.
Moreover, the transmission power of each transmitting node is reduced down to Pt/M
which leads to the reduction of the total power consumption in the network. In addition, the
RTS-CTS implementation can also reduce the hidden node problem in such densely
distributed sensor networks.
RTS-CTS
2 w 2s
h1
s Destination Node
w 1s h2
1
s w 3s
. h3 RF ML
3 Chain
s . hM
.
M wM s
Fig. 5. A cooperative beamforming transmit diversity system with M transmit nodes and 1
destination.
4.2.2 Open-loop MISO System

The modulated signals must be processed at the transmitter first before being transmitted
from multiple antennas M . The main motivation is to reduce the complexity of the feedback
schemes in the closed-loop MISO systems. The transmitter design is enhanced with more
advanced signal processing and/or a combination of various diversity techniques in order
to provide the receiver with the capability to exploit full diversity gain from the received
signals.
One of the early proposed open-loop MISO schemes is the antenna hopping scheme
(Seshadri and Winters, 1993; Wittneben, 1991). The modulated signals are transmitted from
M antennas with different time intervals. At the receiver, the delayed signals introduce a
multipath-like distortion for the intended signal.
The multipath-like distortion can be resolved at the receiver by using a ML detector or a

Minimum Mean Square Error (MMSE) detector to obtain M diversity gain. The antenna
hopping scheme has been shown to achieve fully diversity gain up to M without any
bandwidth expansion but at the cost of a lower spatial rate.
In order to gain the full spatial rate, the diversity gain achieved from a multiple antennas
implementation is combined with the coding gain achieved from the error control and
channel coding schemes. The combination schemes of error control coding and multiple
antennas have gained the full spatial rate in addition to the diversity benefit but at the cost
of bandwidth losses due to code redundancy (Vucetic and Yuan, 2003). A better and
practical approach is a joint design of multiple antennas with channel coding schemes. This
approach can be achieved when the multiple antennas and channel coding schemes are
designed as a single signal processing module. Coding techniques for multiple antenna
communications are called space-time coding (STC). STC schemes provide redundant
transmission in both spatial and temporal domains. In addition to the diversity gain, full
spatial rate and no bandwidth expansion advantages, STC schemes can be combined with
multiple receive antennas to achieve capacity gain.
The most popular STC scheme is due to Alamouti (Alamouti, 1998) who studied the case of
two transmit antennas. The Alamouti space-time encoder picks up two symbols s1 and s2
from an arbitrary constellation and the two symbols are transmitted in two consecutive time
slots as shown in Figure 6. In the first time slot, s1 is transmitted from the first antenna while
s2 is transmitted from the second antenna. Consecutively in the second time slot, -s2* is
transmitted from the first antenna while s1* is transmitted from the second antenna. Since
both the symbols are transmitted in two time slots, the overall rate is given as one symbol
per channel use. The key concept of the Alamouti STC scheme is the orthogonal design of
the transmit sequences. The inner product of the sequences x 1 and x 2 is given as:
* *
x1  x 2  s1 s 2  s1 s 2  0 . (12)
The transmitted code matrix has the following property:
s 2   s1  s 2 
2 2
H
s  s 2   s1
* * *
0
X X  1 
*    2 2. (13)
s2 s1   s 2 s1   0 s1  s 2 
Assume that both the paths experience quasi-static fading where the fading coefficients are
constant across the two consecutive symbol transmission intervals which can be expressed
as:
h1 t   h1 t  T   h1  h1 e j 1
. (14)
h2 t   h2 t  T   h2  h2 e j 2
where hk and  k , k= 1, 2, are the amplitude gain and phase shift for the path from transmit
antenna k to the receiver antenna and T is the symbol duration. The received signal in the
first time slot is given as:
r t   r1  h1 s1  h2 s 2  1 (15)
and in the second time slot, the received signal is given as:
r t  T   r2  h1 s 2  h2 s1   2
* *
(16)
where 1 and  2 are the complex white noise with zero mean and variance 2 for the first
time slot and second time slot, respectively. The received signal vector is defined at the
receiver as:
r 
r   1*  (17)
r2 
which can be written as:
h h 2   s1    1 
r   1*  * . (18)
 h1   s 2   2 
*
h2
TX RX
Modulator
s1 s2  x1  s1  s2
*
 Signal ~
s
Combiner
h1
Alamouti Encoder ML
h2 ĥ
s  s2 
*
 
X  1 Channel
*  x2  s2 s1
*
s2 s1  Estimator
ĥ
Fig. 6. An alamouti STC transmit diversity system with 2 transmit antennas and 1 receive
antenna.
Assume that the receiver is coherent and optimal. Then the attempt to recover s1 and s2 can
be given by the following linear combination:
M *
h2   r1   
2 *
H hk s1  h1 1  h2 2 
~ h
s   1* *  
k 1
. (19)
 h1  r2   M
*
h2 *

2 *
hk s 2  h2 1  h1 2
k 1

The resulting decision variables in Equation (19) are equivalent to the one obtained with
receive diversity using the MRC scheme. The only difference is the phase rotations on the
noise components which do not degrade the effective SNR (Alamouti, 1998).
~ 2
The decision variable vector s with s mean and M
variance is sent to the ML
h k
s 
k 1
2
detector. If the average power of the transmitted symbols is E n , the receive SNR in
each sub-channel is given as:
   hk 
M E sn  2
. (20)
k 1 2 2
We can observe from Equation (20) that the linear processing in Equation (19) transforms the
space-time channel into two parallel and independent scalar channels. If we assume that the
symbols are Phase Shift Keying (PSK) modulated signals with equal energy constellations,
the total transmission power is effectively doubled as shown in Equation (20) compared to
the SIMO MRC and MISO beamforming schemes. Let  k  hk 

2  
E sn
2
, then the
2 2
effective receive SNR can be written as:
M
   hk 
2    
E sn
2
M
. (21)
k
k 1 2 2 k 1
If we assume that all the different paths have the same average SNR, then the average of the
effective SNR at the output of the ML block is:
   E h 
M     E
2 E sn
2
M
  M  k . (22)
k k
k 1 2 2
k 1
As we can observe, the MISO Alamouti STC scheme provides the same diversity gain as the
SIMO MRC and MISO beamforming schemes with M equal to two but with 3 dB loss in
error performance (Larsson and Stoica, 2003). In addition, the MISO Alamouti STC scheme
can be applied for a system with 2 transmitting antennas and N receiving antennas to gain
higher capacity. Although such systems are very important for high-speed networks, careful
consideration of circuit and processing energies and decoder complexity at the receiver in
WSNs keeps our discussion to systems with only one receive antenna which corresponds to
one receive node in cooperative MISO transmission.
The Alamouti STC scheme can be generalised from two transmit antennas up to M transmit
antennas by using the same theory of orthogonal design (Tarokh et. al., 1999). The
generalised scheme is referred to as Orthogonal Space-Time Block Codes (OSTBC). In
general, OSTBC can be categorised into two types: real and complex, based on the signal
constellation.
The basic operation of OSTBC is shown in Figure 7 where the scheme can achieve full
transmit diversity up to M order with M transmit antennas while allowing the use of a very
simple ML decoding algorithm and linear combining at the receiver. However, OSTBC
trades off full diversity gain for lower spatial rate when M > 2. In order to provide a
compromise between full diversity and full rate, a Quasi-Orthogonal STBC (Quasi OSTBC)
scheme was proposed in (Jafarkhani, 2005).
Another class of STCs is the Space-Time Trellis Codes (STTC) (Tarokh et. al., 1998). STTC
achieves higher coding gain and is comparable to STBC in terms of achieving full transmit
diversity gain. However, the encoder design based on trellis-coded modulation leads to a
more complex receiver with a Viterbi algorithm decoding implementation. The ML decoder
complexity grows exponentially with the number of bits per symbol, thus limiting the
achievable data rates.
In the context of practicality in WSNs, the main obstacle of MISO STBCs and STTCs schemes
implementation is the issues of how to provide each transmitting node with the transmit
sequences knowledge and how different transmit sequences are assigned to each node in
order to provide an orthogonal or quasi-orthogonal design.
A better and practical approach as suggested in (Yang et. al., 2007) is when the source node
broadcasts the transmit sequences to its particular neighbours in order to provide the
transmit sequences knowledge together with the original data signal. Such an approach
introduces an increasing packet overhead as M increases, prior data packet transmission.
The overhead is a compromise with full diversity gain which achieves higher reliability and
lower transmission energy.
Modulator h1 ~
Signal s
1
sS Combiner
x1 .
. hM ML
Space-Time . . ĥ
Block Encoder . Channel
X M Estimator
xM ĥ
RX
TX
Fig. 7. A STBCs transmit diversity system with M transmit antennas and 1 receive antenna.
2 x2 Destination Node
h1 ~
s Signal s
x1 h2 Combiner
1
s x3
. h3 ML
s .
3
hM ĥ
. Channel
Estimator
xM
M ĥ
Fig. 8. A cooperative STBC transmit diversity system with M transmit nodes and 1
destination.
As a comparison, MISO STBC is more practical and promising to be implemented in WSNs

due to a simpler decoding algorithm which leads to lower processing energy at the receiver.
On the other hand, the simpler encoding and decoding algorithms of MISO STBC come at
the cost of higher transmission power compared to the MISO beamforming scheme. The
pictorial concept of cooperative MISO STBC is shown in Figure 8.
4.2.3 Spatial Multiplexing MIMO System

The main motivation of spatial multiplexing (SM) scheme is to achieve a higher data rate
while maintaining the same full diversity gain. Thus the main purpose of SM schemes is
basically to complement the lack of spatial rate in MISO STBC and STTC schemes. Therefore
SM schemes are designed purposely for high data rate applications such as mobile
communications systems and wireless local area networks. Though the current WSNs target
only low to medium data rate applications, future generations of WSNs may require to
operate with such high data rate applications which makes the investigation of cooperative
SM in WSNs relevant and useful.
The main concept of SM (also referred as Layered Space-Time Codes – LSTC (Foschini,
1996)) is to provide simultaneous transmissions of M information streams in the same
frequency band from M transmit antennas. However, by using such a transmission method,
a constraint is introduced where the number of receive antennas must be equal or greater
than the number of transmit antennas (N ≥ M ) in order to separate and detect the M
transmitted signals. The separation process involves a combination of interference
suppression and cancellation. The achievable spatial rate is given as R cbM where R c denotes
the rate of the channel code whenever channel coding is employed and 2b is the signal
constellation size. When full channel code is achieved with R c = 1 and Binary PSK is used
with b = 1, we can show that the spatial rate is increased linearly with M .
Among the simplest SM schemes is Bell Laboratories Layered Space-Time (BLAST) (Golden
et. al., 1999). There are various versions of the BLAST schemes in the literature such as
Vertical BLAST (VBLAST), Horizontal BLAST (HBLAST) and Diagonal BLAST (DBLAST).
The simplest version is VBLAST due to the simplest encoder architecture compared to
HBLAST and DBLAST (Vucetic and Yuan, 2003). VBLAST is also referred to as an un-coded
LST scheme while HBLAST and DBLAST are classified as coded LST schemes. The simple
encoder architecture makes VBLAST the most practical version of SM schemes for
implementation in WSNs in order to keep the complexity and power consumption as low as
possible. There are other SM schemes such as threaded LSTCs and multilayered LSTCs, with
higher spatial rate but they come at the cost of more complex encoding and decoding
mechanisms. Obviously these schemes are not practical to be implemented in WSNs and
thus are not considered in our work.
A VBLAST encoder is shown in Figure 9. As shown in the figure, the bit stream is de-
multiplexed into M sub-streams. Each M sub-stream is then modulated and transmitted
from M transmit antennas. The transmitted signal matrix is given as:
X  sk   t
(23)
where k= 1, 2, ….., M and t = 1, 2, …..,L with L is the transmission block length. At a given
time t, the transmitter transmits the t th column from the transmission matrix, one symbol
from each k antenna. The given transmission mechanism represents vertical structuring
th
referring to transmission sequences of the matrix columns in the space-time domains

(Vucetic and Yuan, 2003). Given the system constraint of N ≥ M , the achievable spatial rate is
bM and the achievable spatial diversity depends on the detection scheme employed at the
receiver. When a Zero Forcing (ZF) or a Minimum Mean Square Error (MMSE) decoder is
used at the receiver for the separation and detection, the achievable spatial diversity varies
between 1 to N (Vucetic and Yuan, 2003). In order to gain full spatial diversity equal to N , a
ML decoder must be employed at the receiver at the cost of a more complex decoder
compared to ZF and MMSE. The complexity of the ML decoder increases linearly with bN .
Thus the use of a modulation scheme with the smallest constellation size (e.g. b = 1) is very
helpful to reduce the decoder complexity while achieving higher spatial diversity gain.
In the context of practicality in WSNs, there are three major issues that must be tackled: how
to provide the data packet stream to M -1 transmitting nodes, how to transmit each M data
packet stream simultaneously from M nodes and how to forward the receive data packets
by N -1 receiving nodes to the destination. An example of architecture for cooperative SM
which is based on the VBLAST scheme was proposed in (Yang et. al., 2007) and is shown in
Figure 10. The cooperative SM scheme has the following operations:
a) Source node broadcasts the original data packet stream to its M -1 neighbour nodes
with very low power and all the M transmitting nodes send the same data packet
streams simultaneously after the sending timer expires.
b) N receiving nodes receive the data packet streams from M transmitting nodes and
each receiving node employs a ML decoder to decode the data packet and forward
the data packet to the destination node.
In order to gain both spatial diversity and spatial rate, the constraint of the traditional SM
scheme such as N ≥ M also works for the cooperative SM scheme. Consider the transmission
route in Figure 10. The error rate in each route is given as:
Pe  Pe M _ 1  Pe pp ( dst )  Pe pp ( dst ) Pe M _ 1( recv ) (24)

h11
TX
1 1
Modulation
. h1M .
. . .
Demux RX
. . .
1: M
. hN 1
M N
Modulation hNM
Fig. 9. A VBLAST spatial multiplexing system with M transmit antennas and N receive
antennas.
2 3
s 2
1 s H
1
. 3
s . 4
. .
.
M N .
Fig. 10. A cooperative spatial multiplexing system with M transmit nodes and N receive
nodes.
where Pe M _1 is the error rate for M nodes cooperatively sending to one receiving node
which relates to the power summation from multiple paths M , and different fading
characteristics that may occur in different signal transmission paths. Pepp ( dst ) is the error
rate from one receiving node to the destination. A simple majority decision rule is employed
at the destination node when multiple packets are received from N -1 nodes (Yang et. al.,
2007). The data packet stream with the lowest BER, which means that the SNR is maximised,
is selected at the destination node. If each receiving node in the receiving group has the
same BER, the BER in the destination node after the reception from the N nodes forming the
reception group is given as:
N
N k
PeM _ N     Pe 1  Pe N  k . (25)
kN / 2  k 
5. Performance Analysis
In this section, we study the performance of cooperative MIMO schemes discussed earlier
on, namely cooperative MISO Beamforming (BF), MISO STBC and MIMO SM schemes. The
clock jitter impact is modelled as a timing error function in Section 5.1. The error
performance for each cooperative scheme is modelled in Section 5.2 while the results are
discussed in Section 5.3.
5.1 Timing Error Modeling

We consider the impact of imperfect synchronisation which is caused by clock jitter alone.
Each cooperative sending nodes experiences clock jitter with the jitter around a reference
clock, T o denoted as Tj m where 1  m  M . The worst case scenario is considered here with
only 2 cooperative transmitting nodes where the clock jitters are fixed at the extreme ends,
Tb T
, T j   b where 0  Tb  Tb and Tb is the bit duration. Thus the clock
1 2
Tj  
2 2
jitters difference is T j  T j 1  T j 2  Tb . The effect of imperfect synchronisation can be
modelled as a degrading function of the bit period which consequently degrades the
received bit energy. Therefore the timing error as a function of the bit period and clock
jitters difference is given as:
Te  Tb  T j . (26)
5.2 Error Performance Modeling

We derive the two most important performance parameters to measure the channel condition
and to evaluate the link reliability: BER and PER. Without Forward Error Correction (FEC), the
relationship between Packet Error Rate (PER), P p and BER, P b is given by:
Pp  1  1  Pb 
N data
(27)
where N datais the packet length in bits. Consider the case of BPSK modulation under quasi-
static Rayleigh fading with fading gain h, experiencing a square law path loss without
channel codes. In the SISO system, the conditional SNR is given by (Proakis, 2001):
2
Pt h Gt Gr
 bSISO  2
(28)
 4d 
NoM l  
  
where P tis the transmission power, d is the distance between the sending and destination
node, G t and G r are the transmission and reception antenna gain,  is the carrier wave
length, M  is the link margin and N o is single-sided thermal noise power spectral density
(PSD) given as -171 dBm/Hertz.
7KHSUREDELOLW\GHQVLW\IXQFWLRQ3’)RI 
bSISO LVJLYHQE\
 bSISO
1 
p bSISO   exp  bSISO
 bSISO
E h
ZKHUH bSISO LVWKHDYHUDJH615$VVXPHWKDW    1 /DUVVRQDQG6WRLFD WKHQ
2
 bSISO LVJLYHQE\
WKHYDOXHRI
 bSISO 
 2
Pt E h Gt Gr

Pt Gt Gr
2 2
 4d   4d 
NoM l   No M l  
     
7KHDYHUDJH%(5FDQEHH[SUHVVHGDV

Eh PbSISO   Eh Q 2 bSISO 
7KHXSSHUERXQGRIWKHDYHUDJH%(5FDQEHGHULYHGDV3URDNLV
  
Q 2 bSISO  p x  2 bSISO 

 exp 
 2   bSISO
2
 2 
 
 
E Q 2 bSISO  Eexp  bSISO 
 bSISO LVJLYHQE\3URDNLV
7KHPRPHQWJHQHUDWLQJIXQFWLRQRI
1
S   E exp bSISO S  
S bSISO
Eh PbSISO   Eexp  bSISO    1  (1   bSISO ) 1 0

M
Pt Gt Gr M
 bSTBC   hk    STBCk
2
2 (37)
k 1  4d  k 1
M  NoM l  
  
and the SNRs for imperfect synchronisation scenario at the destination node can be given
by:
M
Pt Gt Gr Te M
 bBF   hk    BFk
2
2
 (38)
k 1  4d  Tb k 1
NoM l  
  
M
Pt Gt Gr Te M
 bSTBC   hk    STBCk
2
2
 (39)
k 1  4d  Tb k 1
M  No M l  
  
where  BFk and  STBCk are the instantaneous SNR on the kth channel.
The PDF of  BFk and  STBCk are:
 BFk
1 
p BFk   exp  BFk
(40)
 BFk
 STBCk
1 
p STBCk   exp  STBCk
. (41)
 STBCk
Assume that E hk    1 (Larsson and Stoica, 2003), then the values of 
2
BFk and  STBCk
for perfect synchronisation scenario become:
 BFk 
Pt E hk  G G 2
t r

Pt Gt Gr
(42)
2 2
 4d   4d 
NoM l   NoM l  
     
 STBCk 
Pt E hk G G2
t r

Pt Gt Gr
(43)
2 2
 4d   4d 
M  No M l   M  No M l  
     
and the average SNRs for imperfect synchronisation scenario can be given by:
 BFk 
 G G
Pt E hk
2
t r

Te

Pt Gt G r

Te
(44)
 4d 
2
Tb  4d 
2
Tb
NoM l   NoM l  
     
 STBCk 
 G G
Pt E hk
2
t r

Te

Pt Gt G r

Te
. (45)
 4d 
2
Tb  4d 
2
Tb
M  NoM l   M  NoM l  
     
The moment generating functions of  bBF and  bSTBC are (Proakis, 2001):
M
1
S   Eexp bBF S    (46)
k 1 S BFk
E h PbBF   E exp   bBF     1  (1   BFk )  M (47)
M
1
S   Eexp bSTBC S    (48)
k 1 S STBCk
Eh PbSTBC  Eexp  bSTBC   1  (1   STBCk ) M . (49)
The average BER for the cooperative SM scheme in (Yang et. al., 2007) is given as:
N
N k
PbSM    Pe 1  Pe N k (50)
k N / 2  k 
Pe  E h PbMISO   E h PbSISO   E h PbSISO E h PbMISO  (51)
where P e is the error rate in each route and N is the number of nodes forming the reception
group. The average SNR of the MISO scheme in Equation (51) is the same as the average
SNR of the cooperative MISO BF scheme (Yang et. al., 2007). Thus we assume that the
average BER is the same for both schemes. Table 2 lists the system parameters used for
evaluating BER performance of the three cooperative MIMO schemes.
5.3 Performance Results and Discussions

Figures 11, 12 and 13 show the corresponding results for perfect synchronisation scenarios.
For comparison, those figures also show the BER performance of the corresponding SISO
scheme. As we can see, in general, cooperative BF outperforms the other schemes except for
the special case below the 10mW transmit power where cooperative SM performs better.
However, this special case may not have a significant impact due to the fact that the
operating transmission power for WSNs is in the range between 25mW to 50mW (Kohvakka
et. al., 2006). Also, we can observe that the diversity gain of cooperative SM depends on N
and not M as shown in Figures 12 and 13. In addition, the cooperative SM achieves spatial
rate equal to M .
Figures 14 and 15 show the corresponding results for imperfect synchronisation scenarios.
As we can see, in general SISO outperforms other schemes above 0.8T b and cooperative SM
outperforms the other schemes when the diversity gain is getting higher. However, when
the diversity gain of all the cooperative schemes is the same, cooperative BF outperforms the
other schemes.
Symbol Quantity
fc 2.4 GHz
G tG r 5 dBi [5]
M  40 dB [5]
d 100 meters
dm 10 meters
250 KbpsRb
Table 2. System parameter for ber and per modeling
0
10
-5
10
B it E rro r R a te
SISO
6x1 BF
6x1 STBC
6x6 SM
-10
10
-15
10
0 10 20 30 40 50 60 70 80 90 100
Transmitted Power, Pt in mW
Fig. 11. BER vs. transmission power for various cooperative schemes with M = 6 and N = 1
(Cooperative BF and Cooperative STBC) and M = N = 6 (Cooperative SM).
0
10
-5
10
SISO
Bit Error Rate
6x1 CMISO Beamforming

6x1 CMISO STBC
6x2 CMIMO SM
6x4 CMIMO SM
10
-10 6x6 CMIMO SM
-15
10
0 10 20 30 40 50 60 70 80 90 100
(Cooperative BF and Cooperative STBC) and various N = 2, 4 and 6 for Cooperative SM with
M = 6.
0
10
-5
10 SISO
6x1 CMISO Beamforming
B it E rro r R ate
6x1 CMISO STBC

3x6 CMIMO SM
4x6 CMIMO SM
5x6 CMIMO SM
-10
10 6x6 CMIMO SM
-15
10
0 10 20 30 40 50 60 70 80 90 100
(Cooperative BF and Cooperative STBC) and various M = 3, 4, 5 and 6 for Cooperative SM
with N = 6.
-1
10
SISO
2x1 BF, 0Tb
10
-2 2x1 STBC, 0Tb
2x2 SM, 0Tb
2x1 BF, 0.3Tb
B it E rro r R ate
2x1 STBC, 0.3Tb

-3 2x2 SM, 0.3Tb
10
2x1 BF, 0.6Tb
2x1 STBC, 0.6Tb
2x2 SM, 0.6Tb
-4
2x1 BF, 0.9Tb
10
2x1 STBC, 0.9Tb
2x2 SM, 0.9Tb
-5
10
10 20 30 40 50 60 70 80 90 100
Fig. 14. BER vs. transmission power for various imperfect synchronisation cooperative
schemes with M = 2 and N = 1 (Cooperative BF and Cooperative STBC) and M = N = 2
(Cooperative SM).
-1
10
10
-2
SISO
2x1 BF, 0Tb
-3 2x1 STBC, 0Tb
10
2x4 SM, 0Tb
2x1 BF, 0.3Tb
B it Error R ate
10
-4
2x1 STBC, 0.3Tb
2x4 SM, 0.3Tb
-5 2x1 BF, 0.6Tb
10
2x1 STBC, 0.6Tb
2x4 SM, 0.6Tb
10
-6
2x1 BF, 0.9Tb
2x1 STBC, 0.9Tb
-7 2x4 SM, 0.9Tb
10
10 20 30 40 50 60 70 80 90 100
Fig. 15. BER vs. transmission power for various imperfect synchronisation cooperative
schemes with M = 2 and N = 1 (Cooperative BF and Cooperative STBC) and N = 4 for
Cooperative SM with M = 2.
6. Conclusion
This chapter has examined the major diversity techniques and various cooperative
configurations, including BF, STBC and SM schemes in conjunction with performance
evaluation and comparative literature. Both cooperative BF and STBC schemes utilise the
MISO concept while the SM scheme utilises the MIMO concept. We have shown that the
cooperative MISO BF is the most promising scheme to be implemented in WSNs due to the
lowest error performance among others with the same diversity gain. Also, cooperative
MISO BF outperforms other cooperative schemes in imperfect synchronisation scenarios. On
the other hand, cooperative MIMO SM is more practical in terms of lower error performance
and tolerance to clock jitter error when its diversity gain is higher than the others. In
addition, cooperative MIMO SM provides a higher spatial rate as M grows.
The comparative study relates the diversity gain with the reduction in the transmission
power by increasing the communication link reliability. However, in order to find the best
or optimal scheme to be used in WSNs, we have to compare all the three schemes in terms
of total energy consumption which must include both the transmission power and circuit
power for each sensor node in the network. The discussion in this chapter can provide a
basis for further study to find the optimal cooperative MIMO scheme when both
transmission power and circuit power are considered for all required energy components of
cooperative communications in WSNs.
7. References
Alamouti, S.M. (1998). A Simple Transmit Diversity Technique for Wireless
Communications, IEEE Journal on Selected Areas in Communications, vol. 16, pp.
1451-1458.
Duman, T.M. & Ghrayeb, A. (2007). Coding for MIMO Co mmunication Systems , First ed. West
Sussex, England: John Wiley & Sons Ltd.
Foschini, G.J. (1996). Layered Space-time Architecture for Wireless Communication in A
Fading Environment when using Multi-element antennas, Bell Labs Technical
Journal, pp. 41-59.
Golden, G.D.; Foschini, G.J.; & Valenzuela, R.A. (1999). Detection Algorithm and Initial
Laboratory Results using the V-BLAST Space-time Communication Architecture,
Electronic Letters, vol. 35, pp. 14-15.
Gupta, G. & Younis, M. (2003). Fault-Tolerant Clustering of Wireless Sensor Networks,
presented at IEEE Wireless Communications and Networking Conference
(WCNC).
Jafarkhani, H. (2005). Space-Time Coding: Theory and Practice
, First ed: Cambridge University
Press.
Jagannathan, S.; Aghajan, H.; & Goldsmith, A. (2004). The effect of time synchronization
errors on the performance of cooperative MISO systems, presented at IEEE Global
Communications Conference (Globecom), Dallas, Texas, USA.
Karl, H. & Willig, A. (2007). MAC Protocols, In: Protocols and Architectures for Wireless Se
Networks, pp. 111-148, John Wiley & Sons, 978-0-470-09510-2, West Sussex, England.
Kohvakka, M.; Kuorilehto, M.; Hannikainen, M. & Hamalainen, T.D. (2006). Performance
Analysis of IEEE 802.15.4 and Zigbee for Large-scale Wireless Sensor Network
Applications, Proceedings of ACM International Workshop on Performance Evaluation

WirelessAdhoc,Sensor,andUbiquitous, Networks, pp. 1-6, Malaga, Spain.
Kuorilehto, M.; Kohvakka, M.; Suhonen, J.; Hamalainen, P.; Hannikainen, M. & Hamalainen,
T.D. (2007). MAC Protocols, In: Ultra-Low Energy Wireless Sensor Networks in
Practice,pp. 73-88, John Wiley & Sons, 978-0-470-05786-5, West Sussex, England.
Larsson, E.G. & Stoica, P. (2003). Space-Time Block Coding for Wireless Communications, First
ed. Cambridge, UK: Cambridge University Press.
Li, X. & Hwu, J. (2006). Performance of Cooperative Transmissions in Flat Fading
Environment with Asynchronous Transmitters, presented at Military
Communications Conference (MILCOM 2006), Washington DC, USA.
Li, X. (2004). Space-Time Coded Multi-Transmission Among Distributed Transmitters
Without Perfect Synchronization, IEEE Signal Processing Letters, vol. 11, pp. 948-
951.
Li, X.; Chen, M.; & Liu, W. (2004). Cooperative Transmissions in Wireless Sensor Networks
with Imperfect Synchronization, presented at Conference on Signals, Systems and
Computers, Pacific Grove, CA.
Litva, J. & Lo, T.K.Y. (1996). Digital Beamforming in Wireless Communications : Artech House
Publisher.
Liu, Z.; Giannakis, G.B.; Zhuo, S.; & Muquet, B. (2001). Space-time coding for broadband
wireless communications, Wireless Communications and Mobile Computing, vol.
1, pp. 35-53.
Naquib. A.F. & Calderbank, R. (2000). Space-time codes for high data rate wireless
communications, in IEEE Signal Processing, vol. 47, pp. 76-92.
Proakis, J.G. (2001). Probability and Stochastic Processes, In: Digital Communications,pp. 17-
36, McGraw-Hill, 0-07-232111-3, Singapore, Singapore.
Rappaport, T.S. (2002). Wireless Communications: , second ed. Upper
Principles and Practice
Saddle River, NJ, USA: Pearson Education International.
Seshadri, N. & Winters, J.H. (1993). Two Signaling Schemes for Improving the Error
Performance of Frequency-division-duplex (FDD) Transmission Systems using
Transmitter Antenna Diversity, presented at IEEE Vehicular Technology
Conference.
Simon, M.K. & Alouini, M.S. (2000). Digital Communication over Fading Channels: A Unified
ApproachtoPerformanceAnalysis : John Wiley & Sons.
Singh, M. & Prasanna, V.K. (2003). A Hierarchical Model for Distributed Collaborative
Computation in Wireless Sensor Networks, presented at IEEE International Parallel
and Distributed Processing Symposium.
Tarokh, V.; Jafarkhani, H.; & Calderbank, A.R. (1999). Space-time Block Codes from
Orthogonal Designs, IEEE Transactions on Information Theory, vol. 45, pp. 1456-
1467.
Tarokh, V.; Seshadri, N.; & Calderbank, A.R. (1998). Space-time Codes for High Data Rate
Wireless Communication: Performance Criterion and Code Construction, IEEE
Transactions on Information Theory, vol. 44, pp. 744-765.
Vucetic, B. & Yuan, J. (2003). Space-Time Coding . West Sussex, England: John Wiley & Sons
Ltd.
Winters, J.H. (1983). Switched diversity with feedback for DPSK mobile radio systems, IEEE
Transactions on Vehicular Technology, vol. 32, pp. 134-150.
Wittneben, A. (1991). Base Station Modulation Diversity for Digital SIMULCAST, presented
at IEEE Vehicular Technology Conference.
Yang, H.; Shen, H.-Y. & Sikdar, B. (2007). A MAC Protocol for Cooperative MIMO
Transmissions in Sensor Networks, Proceedings of IEEE Global Telecommunications
ConferenceGLOBECOM)
( , pp. 636-640, Washington, USA.
Yuksel, M. & Erkip, E. (2004). Diversity Gains and Clustering in Wireless Relaying,
presented at IEEE International Symposium of Information Theory.
Optimal Cooperative MIMO Scheme in Wireless Sensor Networks 151
Optimal Cooperative MIMO Scheme in

Wireless Sensor Networks
M. Riduan Ahmad1, Eryk Dutkiewicz2, Xiaojing Huang3 and M. Kadim Suaidi4
1,4UniversitiTeknikalMalaysia
2Macquarie 3CSIROICTCentre
Melaka, University,
1,4Malaysia,
2,3Australia
1. Introduction
Cooperative Multiple-Input Multiple-Output (MIMO) has been proposed as a transmission
strategy to combat the fading problem in Wireless Sensor Networks (WSNs) to reduce the
retransmission probability and lower the transmission energy. Among the earliest work on
cooperative MIMO in WSNs is the analysis of the Space-Time Block Coding (STBC) scheme
to achieve lower Bit Error Rate (BER) and significant energy savings. The work is continued
with the implamentation of the Low-Energy Adaptive Clustering Hierarchy (LEACH)
Medium Access Control (MAC) protocol for clustered-based architectures. The combination
of STBC and the LEACH scheme resulted in a significant improvement in transmission
energy efficiency compared to the Single-Input Single Output (SISO) scheme.
Further study is conducted to compare the performance of STBC and various Spatial
Multiplexing (SM) schemes such as Vertical Bell Labs Layered Space-Time (V-BLAST) and
Diagonal BLAST. In this study, LEACH MAC was also utilized and lower transmission
energy and latency were achieved against the SISO scheme. However, the centralized
architecture leads to energy wastage and higher latency compared to a distributed
architecture. On the other hand, the implementation of a distributed architecture needs to
consider synchronisation issues. Thus a practical cooperative MIMO scheme for distributed
asynchronous WSNs is needed.
Moreover, a practical MAC that can suit cooperative transmission is required. A
combination of a practical MAC protocol and an efficient MIMO scheme for asynchronous
cooperative transmission leads to a more energy efficient and lower latency cooperative
MIMO system. A combination of a MAC protocol and a cooperative SM scheme for
cooperative MIMO transmission has been proposed in previous study where the combined
scheme achieves significant energy efficiency and lower latency.
Furthermore, a transmit Maximum Ratio Combiner (MRC) scheme is suggested to be more
tolerant to the jitter difference than the Alamouti STC scheme in network with imperfect
transmitting nodes synchronisation. In this chapter, we expand these studies to two other
cooperative MIMO schemes, namely Beamforming (BF) and STBC for both network
scenarios: perfect and imperfect transmitting nodes synchronisation. The optimal
cooperative MIMO scheme combined with an appropriate MAC protocol should lead to the
lowest energy consumption and lowest packet latency.
The rest of this chapter is organised as follows. Section 2 describes the system model
considered in this chapter. Section 3 and Section 4 model the system performance and are
followed by Section 5 presenting the analytical results for the three cooperative MIMO
schemes (BF, SM and STBC) in terms of total energy consumption and packet latency.
Finally the chapter is concluded in Section 6.
2. System Model
The baseline system for cooperative MIMO communication is equipped with a CMACON
protocol as proposed and evaluated in (Yang et. al., 2007). Sleep cycles are not implemented
in order to ensure that the cooperative nodes are always available to perform cooperative
transmission and reception. In order to avoid collision, we assume that during the
cooperative transmission and reception, other nodes in the vicinity that are not involved in
the transmission are put in the silent mode for the whole transmission duration. The
duration to remain silent is obtained from the Network Allocation Vector (NAV).
Also in this chapter we consider the impact of imperfect synchronisation caused by clock
jitter alone. Each cooperative transmitting node experiences clock jitter with the jitter around
m
a reference clock, T o denoted as Tj where 1  m  M . The worst case scenario is
considered here with only 2 cooperative transmitting nodes where the clock jitters are fixed
Tb T
at the extreme ends, T j 1   2
, T j   b where 0  Tb  Tb and Tb is the bit
2 2
1 2
duration. Thus the clock jitters difference is T j  T j  T j  Tb . The effect of imperfect
synchronisation can be modelled as a degrading function of the bit period which
consequently degrades the received bit energy.
The baseline network configurations for MISO BF and STBC are shown in Figures 1 and 2
while for MIMO SM it is shown in Figure 3. The network is assumed to be distributed
without any infrastructure and the nodes are fixed once they are deployed. A new node that
wants to join the network should broadcast a packet after powering up to acknowledge its
presence in the neighbourhood. A node checks its remaining energy regularly and when its
total remaining energy is below the threshold, which indicates that its death is near, it
informs the other nodes in the vicinity of the expected death time. Therefore the
neighbouring nodes will exclude this node from any future cooperative MIMO
transmission. The distance between the cooperating nodes either at the transmitting or
receiving side is assumed to be very small compared to the distance between the source
node and the destination node, d. We assume that there are M cooperative transmitting
nodes and one receiving node for the perfect synchronisation scenario and M = 2
cooperative transmitting nodes and one receiving node for the imperfect synchronisation
scenario. A special case for the spatial multiplexing scheme is used where the number of the
cooperative receivers is assumed to be N .
In this section, we introduce two kinds of network configurations. The first network
configuration involves data transmission from M cooperating transmitting nodes to one
destination node by utilizing either of the two MIMO schemes: BF or STBC. An RTS-CTS
handshaking method is performed as described in (Yang et. al., 2007) and the source node
broadcasts the original data packet to its M -1 neighbours.
RTS-CTS
2 w 2s
h1
s Destination Node
w 1s h2
1
s w 3s
. h3 RF ML
3 Chain
s . hM
.
M wM s
Fig. 1. A cooperative beamforming transmit diversity system with M transmit nodes and 1
destination.
2 x2 Destination Node
h1 ~
s Signal s
x1 h2 Combiner
1
s x3
. h3 ML
s .
3
hM ĥ
. Channel
Estimator
xM
M ĥ
Fig. 2. A cooperative STBC transmit diversity system with M transmit nodes and 1 destination.
2 3
s 2
1 s H
1
. 3
s . 4
. .
.
M N .
Fig. 3. A cooperative spatial multiplexing system with M transmit nodes and N receive nodes.
In the case of the BF scheme, the channel information is estimated and optimized from the
CTS packet by all the M nodes in order to weight the data packet. In the case of the STBC
scheme, all the M nodes encode the original data packet with the information supplied by
the source node in the broadcast packet. Both schemes utilize a Maximum Likelihood (ML)
detector and a coherent receiver is used. The second network configuration is the data
transmission from M cooperating transmitting nodes to N cooperating receiving nodes by
utilizing the concept of SM. The recovered data from N -1 nodes is forwarded to the
destination node.
3. Energy Consumption Performance Analysis

The energy consumed by a sensor node can be categorized into two major parts (Cui et. al,
2004; Nguyen et. al., 2007): energy expended during running the transceiver circuits, P c and
energy expended during packet transmission, P t. Therefore, both energy components must
be considered when comparing the total energy consumption of cooperative MIMO and
SISO transmission schemes. All the nodes in vicinity that are not involved in the
transmission and reception are assumed to be in the sleep mode. Also for simplicity, the
energy consumed during the transient mode from the sleep mode to the active mode and by
the digital signal processing blocks is neglected.
3.1 SISO System

To model transmission energy for the non-cooperative or SISO system, we start with the
power consumed by the power amplifier, P pa. As given in (Cui et. al., 2004; Nguyen et. al.,
2007), P pais dependent on the transmit power P tand can be approximated as:
Ppa  1   Pt (1)

where   1 with  denoting the drain efficiency of the Radio Frequency (RF) power

amplifier and  denoting the Peak-to-Average Ratio (PAR) which depends on the
modulation scheme and the associated constellation size. The total circuit power is given by:
Pc  M  Pct   N  Pcr  (2)
where Pcr  PLNA  Pmix  PIFA  Pfilr  PADC  Psyn and Pct  PDAC  Pmix  Pfilt  Psyn are
values for the power consumption of the Digital-to-Analogue Converter (DAC), mixer, Low
Noise Amplifier (LNA), Intermediate Frequency Amplifier (IFA), active filters at the
transmitter and the receiver, Analogue-to-Digital Converter (ADC) and frequency
synthesizer whose values and a detailed block diagram are given in (Cui et. al., 2004;
Nguyen et. al., 2007). Therefore, the total energy consumption per bit Ebt for the SISO
system can be obtained as:
P
pa  Pc 
(3)
Ebt 
Rb
when M = N = 1. Equations (1) and (2) can be used to model the cooperative BF, STBC and
SM systems with an arbitrary number of M and N . For the traditional Carrier Sense Multiple
Access with Collision Avoidance (CSMA-CA) protocol, the energy consumed for an
unsuccessful transmission attempt is given as:
Eu _ siso  Erts  Ects  Edata _ siso (4)
and that for a successful attempt is given as:
E s _ siso  Erts  Ects  E data _ siso  Eack (5)
where E rts , Ects , E data _ siso , E ack are the energy consumed while sending Ready-to-Send
(RTS), Clear-to-Send (CTS), SISO data and Acknowledgment (ACK). Given the size of each
packet as N rts
N, cts and N ack, Equations (4) and (5) can be rewritten as:
,N data_siso
Eu _ siso  Ebt N rts  N cts  N data _ siso  (6)
Es _ siso  Ebt N rts  N cts  N data _ siso  N ack  . (7)
The expected total energy consumption is given as:
 P 
Esiso   psiso  Eu _ siso  Es _ siso (8)
1 P 
 psiso 
where Ppsiso is the packet error probability of the SISO system which can be obtained in
(Ahmad et. al., 2008)
3.2 Cooperative MIMO System

In this sub-section, we consider two scenarios where the first scenario involves transmission
from M cooperating transmitting nodes to 1 destination node with a local exchange of
information at the transmitting side. This scenario applies to the cooperative MISO BF and
STBC schemes. The second scenario deals with transmissions from M cooperating
transmitting nodes to N receiving nodes with local exchanges at both the transmitting and
receiving sides. This scenario applies to the cooperative MIMO SM scheme.
To model transmission energy for the first scenario, we start with the power consumed by
the power amplifier, P paBsduring a local exchange between the source node and its
cooperating neighbours. P paBsis dependent on the local exchange transmitted power P tmand
can be approximated as:
PpaBs  1   Ptm . (9)

The total circuit power for the local exchange is given by:
PcBs  Pct  M  1  Pcr . (10)

Therefore the total energy consumption per bit EbtBs for the local exchange can be obtained
as:
PpaBs  PcBs 
E btBs  . (11)
Rb
The energy consumed for an unsuccessful BF and STBC transmissions attempt is given as:
Eu _ M  E rts  Ects  E Bs  M  E data _ M (12)
Es _ M  Eu _ M  Eack (13)
where E Bs and Edata _ M are the amounts of energy consumed during packet broadcasting
from the source node to its neighbours and the energy consumed for Cooperative BF or
STBC data transmission. Given the size of each packet as N rts , N Bs, N data_Mand N ack,,
, N cts
Equations (12) and (13) can be rewritten as:
Eu _ M  Ebt N rts  N cts   Ebt Bs N Bs  M  Ebt data _ M N data _ M (14)
Es _ M  Eu _ M  Ebt  N ack . (15)
 P 
EM   pM  Eu _ M  Es _ M (16)
1 P 
 pM 
where PpM is the packet error probability for BF or STBC which can be obtained in (Ahmad
et. al., 2008). To model transmission energy for the second scenario, we start with the power
consumed by the power amplifier, P paBrfrom the destination node to its cooperating
receiving nodes and P paColfrom N-1 receiving nodes to the destination node. P paBrand P paCol
are dependent on the local exchange transmit power P tmand can be approximated as:
PpaBr  1   Ptm (17)

PpaCol  1   Ptm  N  1 . (18)
The total circuit power for the former case is given by:
PcBr  Pct   N  1  Pcr (19)

and the total circuit power for the latter case is given by:
PcCol   N  1  Pct  Pcr . (20)
Therefore the total energy consumption per bit EbtBr and EbtCol for both cases can be
obtained as:
P paBr  PcBr 
EbtBr  (21)
Rb

E btCol
P P
paCol .
cCol
(22)
Rb
The energy consumed for an unsuccessful SM transmission attempt is given as:
Eu _ SM  E rts  E Br  Ects  E Bs  M  Edata _ SM  N  1  ECol (23)
Es _ SM  Eu _ SM  Eack (24)
where E Br , ECol and Edata _ SM are the energy consumed during packet broadcasting from
the destination node to its neighbours, the energy consumed by N-1 cooperating receiving
nodes to the destination node and the energy consumed for the cooperative SM data
transmission. Given the size of each packet as N rts , N Bs, N data_SMand N ack, Equations (23)
, N cts
and (24) can be rewritten as:
Eu _ SM  Ebt  N rts  N cts   Ebt Br N Br  Ebt Bs N Bs 

(25)
M  Ebt data _ SM N data _ SM   N  1Ebt Col N Col
E s _ SM  Eu _ SM  Ebt  N ack  . (26)
 P 
ESM   pSM  Eu _ SM  Es _ SM (27)
1 P 
 pSM 
where PpSM is the packet error probability of cooperative MIMO with spatial multiplexing
which can be obtained in (Yang et. al., 2007). The values of the system parameters used in
Figures 4 to 7 are listed in Table 1 (Cui et. al., 2004; Yang et. al., 2007).
Symbol Quantity
N rts 65 bits
N cts 55 bits
N ack 54 bits
N Bs 1300 bits
N Br 120 bits
N data=N Col 1024 bits
P mix 30.3mW
P syn 50mW
P filt
=P filr 2.5mW
P ADC 9.85mW
P DAC 15.48mW
P LNA 20mW
P IFA 3mW
Table 1. system parameter for energy consumption modeling
4. Packet Latency Performance Analysis

As we noted earlier, each packet transmission in cooperative transmission requires more
steps which introduces more overhead. These steps may increase packet delays. However,
the reduction of PER as the diversity gain increases from the cooperative MIMO exploitation
can reduce the retransmissions rates which in turn can reduce packet latency. Sub-section
4.1 models packet latency performance for the non-cooperative SISO system. Comparison is
then made with the models developed for the cooperative MIMO systems in Sub-section 4.2.
The performance results are discussed in Section 5.
4.1 SISO System

For SISO communication, T rts
, T cts and T ackare the transmission periods for the RTS,
, T data
CTS, DATA and ACK packets. The period with a successful transmission attempt is given
as:
Ts _ siso  Trts  Tcts  Tdata  Tack (28)
and the period with an unsuccessful transmission attempt is given as:
Tu _ siso  Trts  Tcts  Tdata  Twait (29)
where Twait is the duration for which the sender waits for an ACK packet. The packet
transmission delay is then given as:
 P 
Td _ SISO   psiso Tu _ siso  Ts _ siso . (30)
1 P 
 psiso 
4.2 Cooperative MIMO System

In addition to the delay incurred as calculated in the previous section, the broadcast packet
transmission from the source node to its neighbours introduces a broadcast transmission
period, T Bs in cooperative BF, STBC and SM transmissions. The transmission period of
cooperative BF, STBC and SM data packets is the same as that for the SISO system due to the
fact that the packet size and the modulation scheme are the same. The duration of a
successful transmission attempt is given as:
Ts _ M  Trts  Tcts  TBs  Tdata  Tack (31)
Tu _ M  Trts  Tcts  TBs  Tdata  Twait . (32)
The expected packet transmission delay is then given by:
 P 
Td _ M   pM Tu _ M  Ts _ M . (33)
1 P 
 pM 
For the case of cooperative MIMO SM, we introduce the delay for the broadcast
transmission time of a recruitment message sent by the destination node, TBr and the delay
for the time required by the cooperating receiving nodes (N-1) to send the data to the
destination, Tcol . The duration of a successful transmission attempt is given as:
Ts _ SM  Ts _ M  TBr  Tcol (34)
Tu _ SM  Ts _ SM  Twait  Tack . (35)
The expected packet transmission delay is then given by:
 PpSM 
Td _ SM   Tu _ SM  Ts _ SM . (36)
1 P 
 pSM 
The values of the system parameters used in Figures 8 to 11 are as follows: T rts = 0.52ms, T cts
= 0.44ms, T ack= 0.432ms, T Bs= 10.4ms, T Br= 0.96ms, T data
= 8.192ms, T col= 22.3ms (Nguyen et.
al., 2007), and T wait= 70ms (Yang et. al., 2007).
5. Performance Results and Discussions

As shown in Figure 4, SISO is more energy efficient than the cooperative schemes at
transmission powers above 100mW with any number of M and N nodes. The cooperative
SM scheme suffers more in terms of energy efficiency because the total energy consumption
is increasing as the diversity gain and the number of nodes M increases. The cooperative BF
and STBC schemes suffer only with the increasing of the diversity gain.
As we noted earlier the cooperative schemes are more energy efficient when the
transmission power is below 100mW. We can see in Figure 5, that the cooperative BF and
STBC schemes outperform the cooperative SM scheme and that the cooperative BF scheme
is more energy efficient than the cooperative STBC scheme with two transmitting nodes.
For imperfect synchronisation scenarios, as shown in Figure 6, in the case of equal diversity
gain for all the schemes, the cooperative BF scheme is more energy efficient than the other
schemes. However, as the diversity gain of the cooperative SM scheme is increased, as
shown in Figure 7, cooperative SM outperforms the other schemes in terms of energy
efficiency at and above 0.8T b in the region of operating transmission power for WSNs
(common operating transmission power is between 20mW to 60mW (Polastre et. al., 2004;
Kohvakka et. Al., 2006; Kuorilehto et. al., 2007). These results indicate that if we allow some
delays to occur within a particular range during transmission, the cooperative SM scheme
can achieve a significant energy saving. However, by relaxing the synchronisation algorithm
with 0.4T b jitters tolerance, the cooperative BF scheme can achieve the highest energy saving
among the other schemes.
As shown in Figures 8 and 9, the SISO scheme outperforms the cooperative schemes at the
transmission power region above 800mW. At the lower transmission power region, the
three cooperative schemes outperform the SISO scheme. The cooperative SM scheme enjoys
a lower transmission delay when the diversity gain is increasing with any arbitrary number
of transmitting nodes with one condition that the number of cooperative SM receiving N
nodes must be greater than the number of M nodes in cooperative BF and STBC. It also
important to note that cooperative BF outperforms cooperative STBC when M = 2.
For imperfect synchronisation scenarios, as shown in Figures 10 and 11, at the lower
transmission power region, the three cooperative schemes outperform the SISO scheme. The
cooperative BF scheme enjoys lower packet latency and outperforms the other schemes even
when the diversity gain of the cooperative SM scheme is increased.
12
SISO
2x1 BF
Total Eerngy Consum ption in m J
10 2x1 STBC
2x8 SM
4x1 BF
8
4x1 STBC
4x8 SM
6
6x1 BF
6x1 STBC
6x8 SM
4 7x1 BF
7x1 STBC
7x8 SM
2
8x8 SM
0
100 150 200 250 300 350 400 450 500
Fig. 4. Total energy consumption vs. transmission power for various schemes with M = 2, 4,
6, 7, 8 and N = 1 (Cooperative BF and STBC) and N = 8 (Cooperative SM).
-3
x 10
1.8
1.6
Total Energy C onsum ption, E in J
1.4
2x1 BF
1.2 2x1 STBC
2x6 SM
1 4x1 BF
4x1 STBC
0.8 4x6 SM
6x1 BF
0.6 6x1 STBC
6x6 SM
0.4
0.2
0
0 10 20 30 40 50 60 70 80 90
Fig. 5. Total energy consumption vs. transmission power for various schemes with M = 2, 4,
6 and N = 1 (Cooperative BF and STBC) and N = 6 (Cooperative SM).
-3
x 10
2
1.8
Total Energy Consum ption, E in Joule
1.6
1.4 2x1 BF, 0Tb

2x1 STBC, 0Tb
1.2 2x2 SM, 0Tb
2x1 BF, 0.4Tb
1 2x1 STBC, 0.4Tb
2x2 SM, 0.4Tb
0.8
2x1 BF, 0.8Tb
2x1 STBC, 0.8Tb
0.6
2x2 SM, 0.8Tb
0.4
0.2
0
10 20 30 40 50 60 70 80 90 100
Fig. 6. Total energy consumption vs. transmission power (lower region) for various
imperfect synchronisation schemes with M = 2 and N = 1 (Cooperative BF and STBC) and
N = 2 (Cooperative SM).
-3
x 10
2
1.8
Total Energy Consumption, E in Joule
1.6
1.4 2x1 BF, 0Tb

2x1 STBC, 0Tb
1.2 2x4 SM, 0Tb
2x1 BF, 0.4Tb
1 2x1 STBC, 0.4Tb
2x4 SM, 0.4Tb
0.8
2x1 BF, 0.8Tb
2x1 STBC, 0.8Tb
0.6
2x4 SM, 0.8Tb
0.4
0.2
0
10 20 30 40 50 60 70 80 90 100
Fig. 7. Total energy consumption vs. transmission power (lower region) for various
imperfect synchronisation schemes with M = 2 and N = 1 (Cooperative BF and STBC) and
N = 4 (Cooperative SM).
100
SISO
90 2x1 BF
2x1 STBC
80 2x2 SM
Packet Latency in msec
70
60
50
40
30
20
10
0.5 1 1.5 2 2.5 3 3.5 4 4.5 5
Transmitted Power, Pt in W
Fig. 8. Packet latency vs. transmission power for various schemes with M = 2 and N = 1
(Cooperative BF and STBC) and N = 2 (Cooperative SM).
180
2x1 BF
160 2x1 STBC
2x6 SM
140 4x1 BF
4x1 STBC
4x6 SM
120
6x1 BF
6x1 STBC
100
6x6 SM
80
60
40
20
20 30 40 50 60 70 80 80 100
Fig. 9. Packet latency vs. transmission power for various schemes with M = 2, 4, and 6 and
N = 1 (Cooperative BF and STBC) and N = 6 (Cooperative SM).
100
90
80
2x1 BF, 0Tb
2x1 STBC, 0Tb

70
2x2 SM, 0Tb
2x1 BF, 0.4Tb
60 2x1 STBC, 0.4Tb
2x2 SM, 0.4Tb
50 2x1 BF, 0.8Tb
2x1 STBC, 0.8Tb
40
2x2 SM, 0.8Tb
30
20
10 20 30 40 50 60 70 80 90 100
Fig. 10. Packet latency vs. transmission power (lower region) for various imperfect
synchronisation schemes with M = 2 and N = 1 (Cooperative BF and STBC) and N = 2
(Cooperative SM).
100
90
80
70
60
50
40
30
20
10 20 30 40 50 60 70 80 90 100
6. Conclusion
This chapter presents a comparison study of three cooperative MIMO schemes in WSNs. We
have developed analytical models for BER and PER to estimate retransmission rates from
PER in (Ahmad et. al., 2008) and these are used to evaluate the total energy consumption
and packet latency of the cooperative systems in this chapter. We show that the SISO
scheme is more energy efficient and has lower latency at higher regions of transmission
power while the three cooperative MIMO schemes are more energy efficient and
outperform the SISO scheme at lower regions. Clearly, at the higher transmission power
region, the SISO scheme enjoys lower transceiver circuit energy consumption and no energy
cost at all on establishing a cooperative mechanism compared to the cooperative MIMO
schemes. These results provide a constraint on the optimal transmission power or
equivalently the optimal distance that should be used when implementing cooperative
MIMO transmission in WSNs.
From the analysis we can conclude that at the lower transmission power region, the
cooperative optimal BF scheme outperforms both the cooperative SM and STBC schemes in
terms of energy efficiency and packet latency for both perfect and imperfect synchronisation
scenarios. Also we note that the cooperative BF scheme with M = 2 nodes is an efficient
cooperative system. Further work will involve development of MAC protocols optimised
for the cooperative transmission schemes and with the aim of creating an optimal
cooperative transmission mechanism for use in distributed WSNs.
7. References
Ahmad, M.R.; Dutkiewicz, E.; & Huang, X. (2008). Performance Analysis of Cooperative
MIMO Transmission Schemes in WSN, to be presented at the IEEE International
Symposium on Personal, Indoor and Mobile Radio Communications (PIMRC),
Cannes, France.
Cui, S.; Goldsmith, A.J.; & Bahai, A. (2004). Energy-efficient of MIMO and Cooperative
MIMO Techniques in Sensor Networks, IEEE Journal on Selected Areas in
Communications, vol. 22, issue 6, pp. 1089-1098.
Kohvakka, M.; Kuorilehto, M.; Hannikainen, M. & Hamalainen, T.D. (2006). Performance
Analysis of IEEE 802.15.4 and Zigbee for Large-scale Wireless Sensor Network
Applications, Proceedings of ACM International Workshop on Performance Evaluation
WirelessAdhoc,Sensor,andUbiquitous, Networks, pp. 1-6, Malaga, Spain.
Kuorilehto, M.; Kohvakka, M.; Suhonen, J.; Hamalainen, P.; Hannikainen, M. & Hamalainen,
T.D. (2007). MAC Protocols, In: Ultra-Low Energy Wireless Sensor Networks in
Practice, pp. 73-88, John Wiley & Sons, 978-0-470-05786-5, West Sussex, England.
Nguyen, T.D.; Berder, O.; & Sentieys, O. (2007). Cooperative MIMO Schemes Optimal
Selection for Wireless Sensor Networks, presented at IEEE Vehicular Technology
Conference (VTC2007), Baltimore, MD, USA.
Polastre, J.; Hill, J.; & Culler, D. (2004). Versatile Low Power Media Access for Wireless
Sensor Networks, presented at The ACM Conference on Embedded Networked
Sensor Systems (Sensys), Baltimore, Maryland, USA.
Yang, H.; Shen, H.-Y. & Sikdar, B. (2007). A MAC Protocol for Cooperative MIMO
Transmissions in Sensor Networks, Proceedings of IEEE Global Telecommunications
ConferenceGLOBECOM)
( , pp. 636-640, Washington, USA.
Single/Multi-User MIMO Differential Capacity 167
Single/Multi-User MIMO Differential Capacity

Daniel Castanheira and Atílio Gameiro
UniversityofAveiro
(InstitutodeTelecomunicações)
Portugal
1. Introduction
This chapter will be structured around two contributions by the authors on the topic,
(Castanheira & Gameiro 2008) and (Castanheira & Gameiro 2009).
The provision of broadband services to everyone is considered one of the key components
for enabling the so-called information society. It is more or less consensual that to attain the
high-rates envisioned by IMT-2000 (I. R. R. M. M1645 2003) of providing around 1Gbit/s for
pedestrian and 100Mbit/s for high mobility, will require the use of multiple antennas at the
transceivers, to exploit the scattering properties of the wireless medium. Nevertheless, due
to the physical size limitations of the transceivers, the number of antennas cannot be high
and the space between them is limited, which implies that the degree of channel
independence achieved is not sufficient to attain the high capacities envisioned, in most
scenarios. One possible solution to cope with this problem is to have the mobiles
simultaneously communicating with a group of geographically distributed antennas with
perfect cooperation between them. The key to achieve perfect cooperation is to have the
radio signals transparently transmitted / received to / from a central unit (CU), where all
the signal processing is done (FUTON 2008). Considering the high capacities envisioned
optical fiber, due to its low attenuation and enormous bandwidth is the obvious technology
of choice to build these transparent interconnections. However the joint processing of a
group of antennas and the remote transmission of their signals to the CU will require
additional processing power and will imply additional costs to the overall network. It is also
expected, by the law of diminishing returns, that as more and more antennas are jointly
processed the improvement in throughput will not increase linearly with the added
complexity. Thus a tradeoff must be made between the costs/complexity and the number of
antennas deployed. One possible way to ease the complexity problem is to use low
complexity/sub-optimal schemes, like Zero-Forcing (ZF) (Caire & Shamai 2003) or Block-
Diagonalization (BD) (Seijoon Shim et al. 2008) at the CU. Following this line of thought, in
(Jindal 2005) and (Juyul Lee & Jindal 2007), the authors study the incurred losses, in terms of
power/rate offsets, between ZF/BD and the optimal scheme, which is well known to be
Dirty Paper Coding (DPC). The authors conclude that the losses are higher when the
number of transmit antennas is close to the number of aggregate receive antennas.
Thereafter the analysis of the gains introduced by the connection of the system users to
more transmit antennas is of special importance, either to estimate a “reasonable” number of
antennas that should be jointly processed, either to give a measure for the network to check
if, in a distributed antenna system (DAS), it is worthwhile to provide an additional
connection to the mobiles, or to provide some guidelines for the deployment of the
distributed antennas or even to compare sub-optimal schemes to the optimal scheme. In that
context we define Differential CAPacity (DCAP) as the increase in ergodic sum-capacity
when one additional transmit antenna is connected to the system users, to quantify the gains
provided by the processing of one additional transmit antenna at the CU and analyze its
behavior.
This chapter is organized as follows: section 2 presents the channel model; section 3
describes the single-user and multi-user channel capacities. In section 4 and 5 the DCAP, for
the single-user and multi-user scenarios, is studied. In section 6 numerical results are
provided and finally in section 7 some conclusions are drawn.
2. Channel Model
Throughout this chapter a Distributed Multiple-Input Multiple-Output (MIMO) Broadcast
channel with K users, each with N receive antennas, and M transmit antennas is considered.
A broadcast channel (BC) is a communication channel in which there is one sender and two
or more receivers (Cover 1972). For a broadcast channel, the user k received signal,
can be modeled by:
(1)
where is the user channel matrix, is the transmitted signal

vector with power constraint is complex white Gaussian
noise with zero mean and unit variance per vector component
denotes the concatenation of all channels and is matrix-variate
complex Gaussian distributed (Shin & Lee 2003), with zero mean and covariance
. Through the chapter it is assumed that the receiver has perfect
knowledge of its own channel and the transmitter has perfect knowledge of all channels, for
each channel realization. Since a DAS is considered all channel gains are independent ( is
diagonal).
Notations: Boldface letter denote matrix-vector quantities. The operation , and

represents the trace, the Hermitian transpose and the determinant of a matrix, respectively.
By and we denote that is positive definite and an identity matrix. The
notation and is used to denote that the column vector is
distributed as p-variate complex Gaussian, with zero mean and covariance , and
to denote that matrix is complex Wishart distributed, with degrees of freedom
and with mean . The notation denotes that matrix and are identically
distributed.
3. Single/Multi-User Sum-Capacity
3.1 Single-User case
For the single-user case, if only one receive antenna is considered, the system reduces to the
classical multiple input single output (MISO) system. As a consequence, the resulting
capacity expression are much simpler to analyze than for the general case of a multi-user
system. Thus for this case, the system capacity and the corresponding DCAP is described in
more detail, either to introduce the reader to the topic, either to provide some baseline work
to be used in the multi-user case.
According to (Liu & Li 2005a) and (Goldsmith 2005), if only one user with one receive
antenna is considered ( ), the ergodic channel capacity can be expressed by:1
(2)
where is the channel matrix, is the transmit signal covariance matrix

and is the noise variance. Subject to a power constrain , , must be
circularly symmetric complex Gaussian (Telatar 1999) and its correlation matrix, , must be
diagonal2 (Visotsky & Madhow 2001), which is equivalent to independent transmit signals,
for the ergodic channel capacity, , to be maximal.
Taking into account that must be a diagonal matrix, , where

is the signal mean transmit power, and knowing that a single antenna user is considered,
, the ergodic capacity formula reduce to:
(3)
From the previous expression it is easy to identify as the link SNR , then the
capacity expression can be put into the following equivalent format:3
(4)
where is a random variable (RV) corresponding to the sum of exponential distributed

RVs with mean each, which follow the pdf:
(5)
1 We consider for now on that the capacity units are nats/s/Hz, when omitted.
2 The covariance matrix of the channel gains is diagonal, because the channel gains are
independent. So their unitary singular value decomposition matrices are equal to the
identity matrix.
3 According to (Goldsmith 2005) the random variable is exponential distributed and
consequently .
obtain an approximation for the DCAP, for , we need also to know that for ,
see equation (23). For the increase in capacity obtained by the connection of a new
user to the system, can be approximated taking into account the DCAP expression for the
case, equation (24) line , and making an extrapolation of equation (26). From
those assumptions the capacity increase by the connection of one additional user to the
system can be approximated by:
(27)
and as can be confirmed in sub-section 6.2 (Fig. 3(a)), this approximation is close to the
values obtained by numerical simulations and is better for a high number of users.
6.1 Single-User DCAP analysis
In this sub-section, the DCAP, for a grid antenna placement, as show in Fig. 3 (a), is
analyzed. First the exact DCAP values are compared to the obtained bounds, to access their
tightness. Next we analyze the sensitivity of the DCAP to the variation of the mean SNR and
finally we perform a DCAP analysis of a representative area covered by the antennas.
In all numerical analysis presented in this chapter we consider that the mean SNR is only
dependent on the signal path loss (Simplified Path Loss Model, (Goldsmith 2005)), that6
and that . It is also considered that the new connected antenna is
always the one with highest mean SNR from the group of unconnected antennas and that
. By circular ring we mean the group of antennas with the same SNR.
1.5 1.5
Lower Bound Lower Bound
Upper Bound 1/(M-1) bound
1/(M-1) bound 1/(M + M - 1) bound
1/(M + M - 1) bound Exact
 CM-1 (bit/s/Hz)
 CM-1 (bit/s/Hz)
1 Exact 1
M
0.5 0.5
0 0
2 4 6 8 10 12 14 16 2 4 6 8 10 12 14 16
M M
(a) Central Point . (b) Central Point .

Fig. 1. Differential capacity by the connection to one more antenna, exact, upper and lower
bounds.
6 is a unit less constant which depends on the antenna characteristics and on the average
channel attenuation, is a reference distance for the antenna far‐field, is the path loss
exponent and is the transmitted power at a distance .
(8)
From this affine approximation it is easy to see that the sum-capacity scales linearly with the
number of aggregate receive antennas and that this scaling factor is not dependent on the
individual values of or , in the high SNR regime. However, for the case of more
aggregate receive than transmit antennas, , equal power allocation ceases to be
asymptotic optimal, since even for the standard degraded BC the throughput is
maximized by transmitting only to the best user (Caire & Shamai 2003).
4. Single-User DCAP
4.1 Exact Expression
In this sub-section we obtain an exact expression for the DCAP considering the the single-
user scenario. During our analysis of the DCAP behavior it will be observed that the DCAP
tends to have a small decay relatively to the previous DCAP value, when the new added
antenna has the same mean SNR of a previously connected antenna. In that context we also
derive an expression for the DCAP that contains explicitly the value of the previous DCAP
to, latter on, do an analysis of the sensitivity of the DCAP values versus the mean SNR of the
new connected antenna, as more and more antennas are connected to the system.
From the definition of the RV it is possible to prove that:
(9)
with which a recursive algorithm for the calculation of the coefficients, which are central
to the calculation of the capacity values, equation (6), can be obtained, and the following
formula can be derived:
(10)
From equation (10), after multiplying by and integrating from zero to plus
infinity the variable, at both sides of the equation, the DCAP can be expressed by:
(11)
With some more mathematical manipulations, the differential capacity can also be
equivalently expressed by:
(12)
where is equal to for and equal to for . This expression will be

used in the following sections to analyze the DCAP sensitivity with the number of transmit
antennas and the mean SNR’s of the new connected antennas.
4.2 DCAP Bounds

Even if the previous expressions obtained for the DCAP, namely equation (11) and (12), are
exact they are not easy to analyze, mostly because of the presence of the coefficients.
Therefore the derivation of some simple upper/lower bounds to the DCAP is an important
step to analyze its behavior. For doing that, we will rely mostly on the second integral
definition of the DCAP, from equation (11). From that expression one can easily see that if
some bounds for equation are available, they can be easily used to bound
also the DCAP. One simple bound for equation is , for all in . As a
consequence the DCAP can be upper bounded by:
(13)
and the channel capacity can be upper bounded by:
(14)
Another possible bound for equation can be obtained from , for all in
, by the Taylor series expansion of the exponential function around zero. Thus
, for all in , with a maximum difference of ( )4, so:
(15)
and as a byproduct:
(16)
where is the RV moment generating function.
From bound (13) and (16) follows that in the low SNR regime the DCAP can be
approximated by and the channel capacity by:
(17)
The previous bounds although simple can provide some interesting information on the
behavior of the DCAP, as will be seen in section 6.1. However they do not give any idea on
the SNR vector that attain the maximum of the DCAP. This vector can give information on
how the distributed antennas should be geographically positioned to attain most of the
gains provided by the connection of additional antennas to the system. In the following
paragraphs we answer this question, and prove that the SNR vector that attains the highest
4
Obtained numerically, and knowing that this maximum is global.
DCAP value is the SNR vector where all elements are equal. In other words the highest
DCAP is obtained when all transmit antennas are co-located, have the same link SNR
From equation (11) one can see that for and . Therefore if we
assume that the new connected antenna is always the one with the highest mean SNR, the
DCAP will be upper bounded by the one where all mean SNRs are equal to the mean SNR
of the new connected antenna:
(18)
When the mean SNR tends to infinity the maximum of the DCAP is obtained and is equal to:
(19)
Thus the DCAP in general is upper bounded by the co-located transmit antennas case and
its maximal value is only dependent on the number of transmit antennas and not on the
actual values of the mean SNR of the links. However this limit is not attainable in practice.
But, how far is it from a real scenario? It can be shown that for a moderate SNR of 17 dB, for
all links, the difference is lower than 0.1 bps/Hz when we pass from one connected antenna
to two and even lower for the other connected antennas.
Since the previous bound is a limit bound it can be not very tight. However a tighter bound
can be developed considering only a slight higher degree of information:
(20)
From the previous expression one can also upper bound the capacity increase by the
connection of new antennas, considering that the user is actually only connected to one
antenna and that antenna has the greatest SNR of all of them, by:
(21)
where is the Euler constant and . Consequently as the number

of connected antennas increase the DCAP decreases. However, if the number of antennas
grows to infinity so does the capacity, but of course, in practice, one can only have a limited
number of antennas. Thus a tradeoff between the cost and capacity gains must be made
when we choose the number of antennas to be deployed in a real system.
5. Multi-User DCAP
In the previous section we have analyzed the DCAP for the single-user case and have
provided an exact closed form expression for it, that is valid for all mean SNR’s. We have
started with the single user case due to its simplicity and also because the analysis of that
simpler case can give some insight for the analysis of the more difficult case of a multi-user
system. However for the more general case of a multi-user system the analysis of the DCAP
over all mean SNR’s is intractable. Thus, for this case, we consider only the DCAP in the
high SNR regime. In this section we first define the DCAP and after derive a closed form
expression for it, for the co-located antenna case and for . For the DCAP
is analyzed numerically.
According to equation (8), for , the DCAP can be expressed by:
(22)
where is the channel matrix for transmit antennas. Thus for the
multiuser scenario the DCAP is only dependent on the matrix distribution. On the other
hand, for , the DCAP can be expressed by:
(23)
From which we can see that the connection of the system users to a new transmit antenna is
equal to , the multiplexing gain, plus a factor, that is equal to
plus the power offset gain provided when we pass from to
transmit antennas and users5, which can be closely approximated by , for any
number of transmit antennas. Hence, the significance of the factor versus the multiplexing
gain must be analyzed to have a clear picture of the full obtained gain, please refer to section
5.2 and 6.2.
5.1 DCAP for

In this sub-section the DCAP for the optimal scheme, DPC, and for two sub-optimal
schemes, namely ZF and BD is analyzed. A closed form expression and an approximation
for it are also derived. Unfortunately, the general case of a distributed antenna system is
difficult to analyze mathematically, thus for this case we will relie on numerical analysis,
which are provided in section 6.2. In the analysis, contained in this sub-section, we consider
that all transmit antennas are co-located.
For DPC, as shown in appendix (sub-section 9.1), the DCAP when we pass from to
transmit antennas is given by:
5
(24)
where and is the Euler constant. For the linear precoding

schemes a similar expression can be obtained, by noting that for BD the DCAP is equal to
times the DPC one, for a channel, from the BC equivalent point to
point MIMO interpretation, in the high SNR regime (Juyul Lee & Jindal 2007). For that
reason, the DCAP for ZF/BD can be show to be equal to:
(25)
From the previous expression one can see that asymptotically in the number of transmit
antennas, , the DCAP behaves like , for both DPC and ZF/BD,
implying that the difference between the gains of the optimal scheme and the sub-optimal
schemes is very small, when a high number of transmit antennas is considered.
Nevertheless for a finite number of transmit antennas the DCAP gains provided by DPC are
lower and the difference increases with , as can be seen by equations (24) and (25). But
how it scales with , the number of users? It can be seen that for a constant the
DCAP increases logarithmically with for DPC and linearly with for ZF/BD. Thus one
can conclude that even if the optimal scheme has a higher complexity, it will need a much
lower number of transmit antennas than the sub-optimal schemes, which require a lower
complexity, for the same target capacity. Thus the number of antennas to be jointly
processed must be carefully chosen , taking into account the tradeoff between complexity,
network costs and obtained benefits.
5.2 DCAP for

In this sub-section the DCAP is analyzed for the case of more aggregate receive than
transmit antennas. For this case, since equal power allocation stop to be asymptotically
optimal and no closed form solution exists for the power allocation problem (Jindal et al.
2005), a mathematical study to attain a closed form expression for the DCAP is not possible.
Nevertheless an approximation for this DCAP can be obtained. But how? It is know that for
the case of only one transmit antenna, , it is optimal to transmit only to the best user
(Caire & Shamai 2003) and as proven in the appendix (sub-section 9.2), for this case the
maximum increase in capacity, when one additional user is connected to the system, is
attained when all users are co-located i.e. when they have equal mean link SNR’s. For
and and for the high SNR regime this maximum is equal to, see appendix
(sub-section 9.2):
(26)
At this point we have only obtained an expression for the capacity increase by the
connection of one additional system user, for the case of only one transmit antenna. But to
obtain an approximation for the DCAP, for , we need also to know that for ,
see equation (23). For the increase in capacity obtained by the connection of a new
user to the system, can be approximated taking into account the DCAP expression for the
case, equation (24) line , and making an extrapolation of equation (26). From
those assumptions the capacity increase by the connection of one additional user to the
system can be approximated by:
(27)
and as can be confirmed in sub-section 6.2 (Fig. 3(a)), this approximation is close to the
values obtained by numerical simulations and is better for a high number of users.
6.1 Single-User DCAP analysis
In this sub-section, the DCAP, for a grid antenna placement, as show in Fig. 3 (a), is
analyzed. First the exact DCAP values are compared to the obtained bounds, to access their
tightness. Next we analyze the sensitivity of the DCAP to the variation of the mean SNR and
finally we perform a DCAP analysis of a representative area covered by the antennas.
In all numerical analysis presented in this chapter we consider that the mean SNR is only
dependent on the signal path loss (Simplified Path Loss Model, (Goldsmith 2005)), that6
and that . It is also considered that the new connected antenna is
always the one with highest mean SNR from the group of unconnected antennas and that
. By circular ring we mean the group of antennas with the same SNR.
1.5 1.5
Lower Bound Lower Bound
Upper Bound 1/(M-1) bound
1/(M-1) bound 1/(M + M - 1) bound
1/(M + M - 1) bound Exact
 CM-1 (bit/s/Hz)
 CM-1 (bit/s/Hz)
1 Exact 1
M
0.5 0.5
0 0
2 4 6 8 10 12 14 16 2 4 6 8 10 12 14 16
M M
(a) Central Point . (b) Central Point .

Fig. 1. Differential capacity by the connection to one more antenna, exact, upper and lower
bounds.
6 is a unit less constant which depends on the antenna characteristics and on the average
channel attenuation, is a reference distance for the antenna far‐field, is the path loss
exponent and is the transmitted power at a distance .
In Fig. 1 we plot the exact DCAP values and respective bounds for the central point of the
grid antenna placement and do that for two different inter-antenna distances, namely
and . From this figure one can see that if the new connected antenna is in
the same circular ring as a previously connected antenna then its DCAP value, , will
be approximately the same as the previous one, . On the other hand if the new
connected antenna is far away from the circular ring of a previously connected antenna then
the DCAP value decreases a lot. This approximation is better when a high number of
transmit antennas is considered. Thus one can conclude intuitively that if all antennas are in
the same circular ring, have the same mean SNR, their DCAP values will be maximum.
Indeed this is true as shown in section 4.2. This behavior can be explained by the fact that
is dependent on by a factor of , an approximation that due to its
importance will be analyzed in the following paragraphs. It can also be explained by the fact
that in equation (16), tends to zero as we connect to more antennas and as a
consequence the bound becomes independent off all SNR’s except the new one, which for a
circular ring is constant. Thus, since the central point in the grid antenna placement is the
one with highest symmetry we can say that if a target increase in capacity is pre-established
then the user will be connected to or or or or or antennas, depending on the
target increase in capacity defined.
Concerning the bounds tightness we can see from that figure that for low SNR’s the upper
bound provided by equation (16) is the better one, but for high SNR’s the bound from
equation (20) is more accurate. Although, the user will only connect to a small number of
antennas in a real system, due to the diminishing returns expected as the number of
antennas increase. Thus, taking into account this fact, the most important bounds will be the
ones that are tighter for a low number of transmit antennas, and in this case the winner is
the bound from equation (20).
1 1
0.8 0.8
M-1
CM-1/CM-2
M-1
CM-1/CM-2
0.6 0.6
M
0.4 0.4
Exact d = 0.1 Exact d = 0.1

0.2 Exact d = 1 0.2 Exact d = 1
Exact Equal and High SNR Exact Equal and High SNR
Bound M-1/M Bound M-1/M
0 0
2 4 6 8 10 12 14 16 2 4 6 8 10 12 14 16
M M
(a) Central Point. (b) Point.

Fig. 2. Differential capacity sensitivity with respect to and for different inter-antenna
distances.
From Fig. 1 (a) and Fig. 1 (b) one can also see that the exact values of the DCAP, the dashed
and solid blue lines, converge to the same value, for a high number of transmit antennas in
the case of SNR vectors that are multiple among themselves. This fact can be easily proven
together with the fact that as the SNR’s get higher this convergence occurs at a smaller
value. In the case of high SNR’s the DCAP cannot be higher than a given value, having as
critical value , in the case of equal SNR’s. As a consequence, if all transmit
antennas are put closer to the user terminal, by a given factor, the only antennas for which
the DCAP increases are the ones near the terminal.
As previously verified the DCAP in a circular ring tends to be constant. But how close to the
previous value and when this approximation can be made? In the following paragraphs we
propose to analyze this fact using a bound for the ratio . If the ratio is close
to one then the value will be close to . In )LJ (a) and )LJ (b) we plot this
ratio for two different points in the grid antenna placement, namely for the central point and
for . In each figure the exact ratio value for two inter-antenna distances
( and ), the bound from equation (12) and the exact value for equal and high
SNR’s are plotted. From those results one can view that as the number of connected transmit
antennas increase the ratio sensitivity to the SNR vector variation decreases and that its
exact value becomes closer to the bound. One can also see that when the new connected
antenna is not from the same ring as the previously connected antenna, the bound is indeed
very close to the actual value of the respective ratio and the ratio value as a step decrease
that is lower as more connected antennas are considered. As a consequence the DCAP tends
to be constant in a circular ring, as previously observed, and the approximation is better
when a high number of transmit antennas is considered.
0.5
M=1
M=2
0.4 M=3
M=4
0.3
dy
0.2
0.1
0
0 0.1 0.2 0.3 0.4 0.5
dx
(a) Geographical antenna placement. (b) Number of antennas where DCAP
is higher than 0.1 bps/Hz.
Fig. 3. Single-User differential capacity analysis, for a regular antenna placement.
Thus far, we have only analyzed the DCAP values for a single point, but how it will behave
if we consider a given area. The analysis of the DCAP values for a given area cannot be
easily visualized in a two dimensional plot, thus we have relied on a different measure to
analyze its behavior. The metric considered was the maximum number of transmit antennas
that achieve a target DCAP value or more. In this study we have considered a target DCAP
value of 0.1 bps/Hz. The geographical area considered was the one shown in figure )LJ
(a) in light blue, and we have also considered . The results of this study are
presented on figure )LJ (b). To explain the information co ntained into this figure it is
easier to give an example. Thus, let’s assume, for example, the point (0.4, 0.4). For this point
, thus when we connect the second, third and fourth antennas the DCAP value is
higher than the target, but departing from that number of transmit antennas, 4, the DCAP
will be lower than the target.
For the considered scenario and for a target DCAP of 0.1bps/Hz only the first four antennas
will be connected to the user, or equivalent the antennas presented in the first ring. This
figure also shows a circular pattern in the number of necessary antennas. This is related to
the fact that at a given distance the link mean SNR is only dependent on the distance from
the user to the considered transmit antenna.
6.2 Multi-User DCAP analysis

In this sub-section the DCAP of DPC is compared to the one of ZF/BD, for the co-located
antennas scenario. For the DAS, as stated before, we rely on numerical simulations to access
the capacity gains provided by the connection of additional transmit antennas. For this
numerical analysis we consider a scenario with 4 transmit antennas and 2 users, each with
only one receive antenna. Finally for we analyze the DCAP for the equal mean
SNR’s case. But to do that we first analyze the capacity gain provided by the connection of
an additional user to the system and after that we investigate the gain provided by in
comparison to , see equation (23).
4.5 9
ZF KN=1
4 ZF KN=2 8
ZF KN=3
3.5 DPC KN=1 7
DPC KN=2
3
DCAP(bps/Hz)
DCAP(bps/Hz)
DPC KN=3 6
2.5 DPC KN=1 approx.
DPC KN=3 approx. 5
2 DPC KN=3 approx.
4
1.5
3
1
0.5 2 DPC
ZF
0 1
1 2 3 4 5 6 7 8 1 2 3 4 5 6
M (Nº of Tx antennas) M (Nº of Tx antennas)
(a) DPC versus ZF. (b) DCAP for .

Fig. 4. Differential capacity for different number of user's and receive antennas.
As can be seen in Fig. 4 (a), were we present a plot of the DPC and ZF DCAP versus the
number of connected transmit antennas, for the co-located scenario, the logarithm
approximation although simple is very tight. Concerning the benefits of the connection of
new transmit antennas, ZF has higher gains than DPC and the difference increase with .
To see how this difference scales with we plot in Fig. 4 (b) the DCAP values versus for
. From this figure one can view that the DCAP scales logarithmically with the
number of users for DPC and linearly for ZF, as already seen from equations (24) and (25).
Thus even if the implementation complexity of DPC is much higher than the one for ZF, the
sub-optimal scheme, ZF, will need a higher number of transmit antennas than the optimal
scheme for a target system capacity. Thus a tradeoff must be made between the cost of
additional processing power and the cost of additional physical resources, namely transmit
antennas and respective connection to the central unit. As a result of this analysis, as a
scheme approaches the optimal one, the gains obtained by the connection of additional
antennas at the transmitter side decrease. But, remember that the optimal scheme will have
always a higher sum-capacity. As seen before, from equations (24) and (25), the DCAP for
the two precoding schemes converge to the same value for a high number of transmit
antennas and the convergence point increases with .
1.4 1
co-located DCAP value Symmetry lines
3,2,1 0.8 Transmitt Antennas
1.2 2,2,1 probability distribution
0.6 Users Positions
) (bps/Hz)
1
0.4
0.8 0.2
f C3,2,1(C3,2,1
2,2,1
y
0.6
-0.2
2,2,1
0.4 -0.4
0.2 -0.6
-0.8
0
0 0.5 1 1.5 2 2.5 -1
3,2,1 -1 -0.5 0 0.5 1
C2,2,1 (bps/Hz) x
(a) DCAP distribution for .

(b) User’s position, where distributed
DCAP is higher than co-located DCAP.
Fig. 5. Differential capacity distribution for a uniform user distribution and user's positions
where the distributed DCAP is higher than the co-located DCAP. Each pair of equal black
markers represent the positions of each user.
For the DAS we will only analyze the DCAP for DPC and that analysis will rely on
numerical simulations. For this simulation a scenario with 4 transmit antennas and two
users, each with only one receive antenna, is considered. To model the channel we have
considered the Simplified Path Loss Model again, (Goldsmith 2005), with a path loss
exponent equal to 3, and have considered Rayleigh multipath fading. The transmit antennas
were positioned as shown in Fig. 5 (b) and the user positions were randomly generated with
an uniform distribution in . Ten thousand user’s positions were generated
and the DCAP was averaged over 10000 trials. The new connected antennas were always
chosen to be the one that will imply the highest DCAP from the group of unconnected
antennas. As a result of this simulation, we have Fig. 5, where the co-located DCAP,
equation (24), value is represented by a red star. From that figure we see that most of the
user’s positions have a smaller DCAP than the co-located case. However for a small elite of
users, 7 in this specific case, the DCAP is higher than the one obtained by the co-located
scenario. Thus one can conclude that even if the DCAP for a DAS can be higher than the one
for the co-located case the difference will not be high and the number of user positions in
that condition will be low. To have a better understanding of the positions that attain the
maximum DCAP gains for a DAS, in Fig. 5 (b), we show the 7 positions that have a higher
DCAP than the co-located case. From that figure it is possible to infer that these positions
DUHDOOYHU\FORVHWRWKHV\VWHPV\PPHWU\OLQHVGHILQHGE\WKHIRXUWUDQVPLWDQWHQQ
WKDW ZD\ DV LQ WKH FRORFDWHG FDVH WKH QHZ FRQQHFWHG DQWHQQD ZLOO KDYH WKH
GLVWULEXWLRQDVDSUHYLRXVRQH7KXVV\VWHPV\PPHWU\SOD\VDQLPSRUWDQWUROHLQREW
PRVWRIWKH’&$3JDLQVHLWKHUIRUWKHVL
QJOHXVHUHLWKHUIRUWKHPXOWLXVHUFDVHV
2.5 1.4
M = 1 (Sim) K = 250
M = 1 (Approx) 1.2 K = 500
2 M = 2 (Sim) K = 750
M = 2 (Approx) 1 K = 1000
M = 3 (Sim)
DCAP (bps/Hz)
M = 3 (Approx) 0.8
1.5
(bps/Hz)
M = 4 (Sim)
M = 4 (Approx) 0.6
M = 5 (Sim)
1 M = 5 (Approx) 0.4
0.2
0.5
0
0 -0.2
0 5 10 15 20 1 2 3 4 5 6
Nº of Users M (Nº of Tx antennas)
D
&DSDFLW\LQFUHDVHE\WKHFRQQHFWLRQ E IDFWRUYDOXH1
RIDQHZXVHU1
)LJ &DSDFLW\LQFUHDVHE\WKHFRQQHFWLRQRIRQHPRUHWUDQVPLWDQWHQQDWRWKHV\VWH
:KHQ PRUH DJJUHJDWH UHFHLYH WKDQ WUDQVPLW

DQWHQQDV DUH FRQVLGHUHG LW LV LPSRUWDQW W
DQDO\]HWKHIDFWRULQHTXDWLRQWRVHHLILW KDVVRPHVLJQLILFDQFHLQUHODWLRQWR
7KXV ZHPXVWHYDOXDWHWKHFDSDFLW\LQFUHDVH
REWDLQHGE\WKHFRQQHFWLRQRIDQDGGLWLRQDO
XVHU WR WKH V\VWHP 7R GR WKDW ZH KDYH
G RQ
UHOLH
QXPHULFDO VLPXODWLRQV FRQVLGHULQJ WKDW
DQGWKDWHDFKFROXPQRI LV GLVWULEXWHG6LQFHIRUWKLVFDVHQRFORVHGIRUP
H[SUHVVLRQ H[LVWV IRU WKH SRZHU DOORFDWLRQ
KDYH UHOLHG
ZH RQ WKH DOJRULWKP SUHVHQWHG LQ
-LQGDOHWDO WRRSWLPDOO\DOORFDWHSRZHUDPRQJXVHUV$VDUHVXOW)LJDWKHHUJRGLF
FDSDFLW\LQFUHDVHZKHQDQHZXVHULVFRQQHFWHG
WRWKHV\VWHPZDVREWDLQHG,Q)LJDLW
LV DOVR VKRZQ WKH HTXDWLRQ
DSSUR[LPDWLRQ ZKLFK LV YHU\ FORVH WR WKH REWDLQHG
QXPHULFDOUHVXOWV7DNLQJLQWRDFFRXQWWKDWDSSUR[LPDWLRQLQRXUDQDO\VLVLQ)LJ E
SORWWKH IDFWRUYDOXHVIRUGLIIHUHQWQXPEHU HUV)URPWKRVHYDOXHVRQHFDQVHHWKDW
RIXV
HYHQ IRU XVHUV WKLV IDFWRU LV VPDOO QRW
HU WKDQ
KLJK ESV+] 7KXV LW KDV QR
VLJQLILFDQFHLQUHODWLRQWR VLQFHZHDUHFRQVLGHULQJWKHKLJK615UHJLPH)URPWKDW
ILJXUHZHFDQDOVRFRQILUPWKDW LQFUHDVHVIHZHUDVLQFUHDVHV&RQVHTXHQWO\WRREWDLQD
YDOXHZLWKWKHVDPHRUGHURIPDJQLWXGHRI DYHU\KLJKQXPEHURIXVHUVVKRXOGEH
MRLQWO\SURFHVVHGZKLFKZLOOUHVXOWLQDYHU\KLJKSURFHVVLQJFRPSOH[LW\DWWKH&8Z
PXVWEHDYRLGHG
7. Conclusion
,Q WKLV FKDSWHU ZH KDYH DQDO\]HG WKH
WVEHQHIL
RI WKH FRQQHFWLRQ RI DGGLWLRQDO WUDQVPL
DQWHQQDVWRWKHV\VWHPXVHUVRID%&7KHVLQJOH
XVHUFDVHKDVEHHQDQDO\]HGLQPRUHGHWDLO
IRUDOO615•VDQGWKHPXOWLXVHUFDVHKDVEHHQ
DQDO\]HGRQO\LQWKH KLJK 615UHJLPH )RU
the single user case we have obtained a closed-form expression for DCAP that is valid for all
SNR’s and have analyzed its variation/sensitivity with respect to . From that analysis we
have seen that symmetry is the most important system property to obtain most of the gains
provided by the connection of additional transmit antennas. It was also shown that the
maximum increase in capacity is obtained when all links have the same mean SNR’s and not
when they are different. Thus showing that the DCAP maximum is obtained when all
antennas are co-located and not distributed.
For the multi-user case a closed form expression for the co-located transmit antennas DCAP,
for , has been obtained. For the general case of a DAS the DCAP was analyzed
numerically, considering a scenario with 4 transmit antennas and 2 users, each with only
one receive antenna. For this case even if the DCAP can be superior to the co-located case it
is not much higher and most of the user’s positions will have a lower DCAP than the co-
located case. For the positions that attain a higher DCAP than the co-located case we have
shown that this positions are very close to the system symmetry lines, defined by the
transmit antennas, and in that way symmetry plays again a important role in the multi-user
case. For the case of more aggregate receive than transmit antennas we have verified that the
DCAP will be power dependent and will be given by the multiplexing gain plus a factor.
However this factor is much smaller than the multiplexing gain even for one thousand
users. Thus for this case the most significant gain is the multiplexing gain, .
8. References
Caire, G. & Shamai, S., 2003. On the achievable throughput of a multiantenna Gaussian
broadcast channel. InformationTheory,IEEETransactions , 49(7),on
1691-1706.
Castanheira, D. & Gameiro, A., 2008. Distributed MISO system capacity over Rayleigh flat
fading channels’. In Personal, Indoor and Mobiledio Communications,
Ra 208. PIMRC
208.IEEE19thIntern ationalSymposium. pp.
on1-5.
Castanheira, D. & Gameiro, A., 2009. High SNR Broadcast Channel Differential Capacity. In
ICT-MobileSummit209 . Santander, Spain, pp. 1-5.
Cover, T., 1972. Broadcast channels. InformationTheory,IEEETransactions , 18(1),on
2-14.
Dohler, M., Gkelias, A. & Aghvami, A., 2006. Capacity of distributed PHY-layer sensor
networks. VehicularTechnology,IEEETransactions , 55(2), 622-639.
on
FUTON, 2008. FUTON - Fibre-Optic Networks for Distributed Extendible Heterogeneous
Radio Architectures and Service Provisioning. Available at: http://www.ict-
futon.eu/.
Ghosh, S., 2005. NetworkTheory, PHI Learning Pvt. Ltd.
Goldsmith, A., 2005. Wirelesscommunications , Cambridge University Press.
Gradshteyn, I.S. & Ryzhik, I.M., 1994. TableofIntegrals,Series,and5th ed., Academic
Products
Press.
Gubner, J.A., 2006. Probability and Random Processes Electrical
for and Computer Engineers 1st
ed., Cambridge University Press.
I. R. R. M. M1645, 2003. M.1645 : Framework and overall objectives of the future
development of IMT-2000 and systems beyond IMT-2000. Available at:
http://www.itu.int/rec/R-REC-M.1645/e.
Jindal, N., 2005. High SNR analysis of MIMO broadcast channels. In Information Theory,
205.ISIT.Proceedings.
205 InternationalSymposium . pp. 2310-2314.
on
Jindal, N. et al., 2005. Sum power iterative water-filling for multi-antenna Gaussian
broadcast channels. InformationTheory,IEEETransactions , 51(4),on
1570-1580.
Juyul Lee & Jindal, N., 2007. High SNR Analysis for MIMO Broadcast Channels: Dirty Paper
Coding Versus Linear Precoding. Information Theory, IEEE Transactions , 53(12),
on
4787-4792
Liu, H. & Li, G., 2005a. OFDM-Based Broadband Wireless tworks:
Ne Design and Optimization ,
Wiley-Interscience.
Liu, H. & Li, G., 2005b. OFDM-Based Broadband Wireless Networks: Design and Optimization ,
Wiley-Interscience.
Seijoon Shim et al., 2008. Block diagonalization for multi-user MIMO with other-cell
interference. WirelessCommunications, IEEETransactions , 7(7),
on 2671-2681.
Shamai, S. & Verdu, S., 2001. The impact of frequency-flat fading on the spectral efficiency of
CDMA. InformationTheory,IEEETransactions , 47(4),on
1302-1327.
Shin, H. & Lee, J.H., 2003. Capacity of multiple-antenna fading channels: spatial fading
correlation, double scattering, and keyhole. Information Theory, IEEE Transactions
on, 49(10), 2636-2647.
Telatar, I.E., 1999. Capacity of multi-antenna Gaussian channels. Available at:
http://eprints.kfupm.edu.sa/29242/.
Vishwanath, S., Jindal, N. & Goldsmith, A., 2003. Duality, achievable rates, and sum-rate
capacity of Gaussian MIMO broadcast channels. Information Theory, IEEE
Transactions , 49(10),
on 2658-2668.
Visotsky, E. & Madhow, U., 2001. Space-time transmit precoding with imperfect feedback.
InformationTheory,IEEETransactions , 47(6),on
2632-2639.
Viswanath, P. & Tse, D., 2003. Sum capacity of the vector Gaussian broadcast channel and
uplink-downlink duality. InformationTheory,IEEETransactions , 49(8),on
1912-1921.
Wei Yu & Cioffi, J., 2004. Sum capacity of Gaussian vector broadcast channels. Information
Theory,IEEETransactions , 50(9),
on 1875-1892.
Weingarten, H., Steinberg, Y. & Shamai, S., 2006. The Capacity Region of the Gaussian
Multiple-Input Multiple-Output Broadcast Channel. Information Theory, IEEE
Transactions , 52(9),
on3936-3964.
9. Appendix
9.1 DCAP derivation
For the co-located transmit antennas case is distributed.
But since , where is distributed, the DCAP can be
expressed by:
(28)
where , is the Euler’s digamma function. The result from the second line
comes from the determinant property and the result from the third line
from lemma 1 of (Juyul Lee & Jindal 2007) . To obtain an
approximation to the partial sum of the harmonic series we have used
which has a maximum error of .
9.1 Capacity Increase by the connection of one more user, for and
For and the broadcast channel ergodic sum-capacity, equation (7), in the high
SNR regime, simplifies to:
(29)
where is the average SNR of user , and . It is easy to

show that for and that the opposite happens for .
Thus the maximum of is obtained when for all users, if the new connected
user is always the one which implies the highest increase in the ergodic sum-capacity. It can
also be proven that the same value for is obtained for any . Hence, for simplicity
we consider that . Thus has cumulative distribution function ,
(Liu & Li 2005b), where is exponential distributed with mean . Thus the
capacity increase by the connection of an additional user to the system can be expressed by7:
(30)
which after the integral evaluation gives the result shown in equation (26).
7 The term cancel out from the expression since it appears in and .
Low Dimensional MIMO Systems with Finite Sized Constellation Inputs 185
Low Dimensional MIMO Systems with

Finite Sized Constellation Inputs
Rizwan Ghaffar and Raymond Knopp
Eurecom
FRANCE
1. Introduction
The seminal works in (Foschini & Gans, 1998) and (Telatar, 1999) on multiple antenna ele-
ments at the transmitter and the receiver show a huge increase in the throughput of this point-
to-point channel referred to also as multiple input multiple output (MIMO) system. These
promising results of high spectral efficiency and enhanced reliability shifted the focus of re-
search on multi antenna communications and motivated the introduction of multiple antenna
elements in the future communication systems. Researchers persist to strive for finding space
time codes (STC) with reduced decoding complexity. These codes take into account both the
spatial and temporal dimensions of the MIMO channel. Orthogonal Space-Time Block Codes
(OSTBCs) (Larsson & Stoica, 2003) are widely used because they are easy to encode and de-
code. For the case of two transmit antennas, the OSTBC is known as Alamouti code (Alamouti,
1998). OSTBCs are repetition codes that only provide diversity gain. In order to approach the
capacity limit they have to be used in concatenation with an outer code. Remarkable coding
gains can be obtained if a capacity achieving temporal encoder, such as turbo or Low-Density
Parity Check (LDPC) code is used in concatenation with a STC (Gonzalez-Lopez et al., 2006).
Recently it has been shown for the ergodic channels that the complex concatenation of the STC
and the outer codes can be replaced with temporally coded and spatially multiplexed streams
(coded spatial streams) for nearing capacity (Ghaffar & Knopp, 2008a). Each spatial stream
can also be independently coded using temporal encoders as convolutional, turbo or LDPC
codes whereas at the receiver, standard off-the-shelf decoders are used after the demodulator.
To combat the frequency selectivity of MIMO wireless channels with low complexity equal-
ization at the receivers, MIMO OFDM is the appropriate alternative. To contest the inherent
fading of MIMO OFDM wireless channels, improved code diversity of bit interleaved coded
modulation (BICM) for fading channels is rendering it the preferred option. Consequently the
future wireless systems shall be based on BICM MIMO OFDM systems. However the requi-
site antenna spacing combined with the complexity constraints at the receiver are restricting
the future MIMO based communication systems to the maximum of 4 spatial streams whereas
it is reduced to 2 spatial streams in most scenarios. The existing and forthcoming standards as
IEEE 802.11n (802.11n, 2006), IEEE 802.16m (802.16m, 2007) and Third Generation Partnership
Project Long Term Evolution (3GPP LTE) (LTE, 2006) substantiate this argument.
This chapter therefore focuses on low dimensional spatially multiplexed time coded BICM
MIMO OFDM systems with first part being devoted to the transmission strategies and cor-
responding receiver structures for such systems in the broadcast scenario while second part
deliberates on interference suppression in such systems in the cellular scenario. This chapter
particularly takes into account the finite sized constellation inputs and departs from the cus-
tomary idealistic Gaussian assumption for the codewords. Each part is also accompanied by
relevant information theoretic analysis and by simulation results under the settings of upcom-
ing wireless standards.
2. MIMO Broadcast scenario

This part deliberates on the broadcast scenario of BICM MIMO OFDM system though the dis-
cussion also remains valid for the point-to-point MIMO systems. We consider the transmis-
sion strategy in which each spatial stream is independently encoded and modulated. We focus
on the case of uniform power and nonuniform rate spatial streams (Ghaffar & Knopp, 2008a)
and the case of uniform rate and nonuniform power distribution (Ghaffar & Knopp, 2008b)
between these spatial streams. In such a broadcast scenario, receiver consequently views a
multiple access channel (MAC). Shamai (Shamai & Steiner, 2003) termed the approach of sin-
gle code layer at each transmit antenna as MAC-outage approach. The reception is consequently
based on successive interference cancellation (SIC) i.e. sequential decoding and subtraction
(stripping) of spatial streams which introduces unequal error protection (UEP). This can be
coarsely regarded as MMSE DFE as described in (Varanasi & Guess, 1997). The idea of mul-
tiple data streams with UEP adds flexibility to the system which can be exploited for having
prioritized users or advanced services in MIMO broadcast systems and in multimedia broad-
cast multicast services (MBMS). For instance it can be the broadcast of multimedia streams
with different rates (quality) of the same data and the users decoding the stream depending
on the received SNR. It can also be the broadcast of low and high rate streams (as audio and
video) with prioritized or high SNR users decoding both audio and video streams while low
SNR users decoding only the low rate audio stream. It is also applicable to high-definition TV
(HDTV) scenario where low priority/quality users are able to receive standard-definition TV
(SDTV) transmission while high priority/quality users access HDTV. This idea has limited
similarity to superposition codes (Liu et al., 2002) whose signal space has a cloud/satellite
topology. Cloud centers because of relatively higher distance amongst them carry information
for low quality receivers whereas better receivers having larger noise tolerance can resolve up
to the actual transmitted satellite symbol within the cloud.
For coded spatial streams (also for the STC), the well-known data model after appropriate
filtering and sampling is y = Hx + z (to be made precise in the subsequent sections) where
y is the received data, H is the channel matrix, x is the symbol vector with the elements from
finite constellations and z is the noise. The problem is then to detect some or all elements
of x from y. Essentially the same problem occurs in multiuser detection for CDMA (Verdu,
1998) and for single-carrier transmission over channels that induce intersymbol interference.
In these cases, the matrix H usually has a specific structure.
The problem of detection of x from y has stimulated a large body of research (Verdu, 1998)
and references therein. One can easily show that if the noise z is Gaussian then obtaining
the maximum-likelihood (ML) solution for some or all elements of x is equivalent to mini-
mizing the Euclidean distance y − Hx2 with respect to x over the finite set spanned by all
possible combinations of constellation points that can constitute the vector x. For ML soft
MIMO detection, the demodulator calculates the log-likelihood ratios (LLRs) for all bits that
constitute the desired elements of x by summing the Euclidean distances for the values of x
for which that particular bit of the desired element of x is one and zero thereby amounting to
2log( M1 )+···+log( Mnr ) terms where Mk is the modulation alphabet of the k-th spatial stream and
nr is the total number of spatial streams (Larsson & Jalden, 2008). In many cases of practical
interest, one resorts to the approximation of replacing the sums with the largest term which
is equivalent to minimizing the Euclidean distance and is termed as max log MAP approach.
Unfortunately this problem is NP-hard for general H and y (Verdu, 1989) which implies that
there are no known efficient (i.e. polynomial-time) solutions. Many sophisticated methods as
lattice reduction and sphere decoding (Hochwald & Brink, 2003) exist which find the ML so-
lution with high probability, but these methods are in general still computationally complex.
This is true also in an average sense if H is random (i.e. for a fading channel). The popular
"sphere decoding" method is much more efficient than a brute-force search, but it still admits
an average complexity that is exponential in the dimension of x.
Naive solutions, like neglecting the integer constraint coupled with the Gaussian assumption
for the alphabets and then subsequently projecting the so-obtained solution onto the finite set
of permissible x [linear receivers as LMMSE and zero forcing ZF)], in general work poorly
especially at lower SNRs. Standard linear detection approaches are further based on ignoring
the spatial color at the output of linear detectors which results in the decoupling of spatial
streams thereby fundamentally reducing the complexity of detection. These disregards prolif-
erate the suboptimality of linear receivers which exhibit degraded performance especially at
lower SNRs.
Standard receiver solutions for spatially multiplexed broadcast schemes including V-BLAST
(Wolniansky et al., 1998) (Golden et al., 1999) use stripping decoders which incorporate sub-
optimal linear minimum mean square error (MMSE) filters (Medvedev et al., 2006) against
the yet undecoded streams at each successive cancellation stage. MMSE because of its relative
improved performance in the family of linear detectors is the preferred choice. Its optimal-
ity for power constrained Gaussian alphabets is well known but it is suboptimal for finite
size constellations. Gaussian assumption of the post detection interference is open to discus-
sion. Its behavior is close to Gaussian under various asymptotic conditions which include
large SNRs and large number of transmit and receive antennas (Poor & Verdu, 1997). But the
fidelity of Gaussian assumption in a low dimensional system at moderate SNRs is question-
able. Degradation of the performance due to the suboptimality combined with the complexity
in the calculation of linear equalizers at each frequency tone (in OFDM based system) renders
their real-time implementation debatable especially in fast fading wideband environments.
2.1 System Model

Before deliberating further on these receiver structures, we discuss the system model. As
the overall system is based on BICM MIMO OFDM, it is imperative to first understand the
significance and the implication of using BICM.
2.1.1 BICM SISO System

BICM because of its improved code diversity for fading channels and its flexibility to vari-
able transmission rates, is a likely choice for future wireless systems as IEEE 802.11n (802.11n,
2006), IEEE 802.16m (802.16m, 2007) and 3GPP LTE (LTE, 2006). The landmark paper of Caire
(Caire et al., 1998) on BICM showed that on some channels, the separation of demodulation
and decoding is beneficial, provided that the encoder output is interleaved bit wise and a
suitable soft decision metric is used in the Viterbi decoder. Code diversity, and therefore the
reliability of coded modulation over a Rayleigh channel, can be improved this way. The code
diversity in this case is equal to the smallest number of distinct bits along any error event.
This leads to a better coding gain over a fading channel when compared to other coded mod-
Encoder π µ, χ Channel Demodulator π −1 Decoder
Fig. 1. Block Diagram of BICM system. π denotes denotes a bit interleaver.
ulation schemes as Trellis Coded Modulation (TCM). BICM increases considerably Hamming
distance while reducing (often marginally) Euclidean distance so BICM outperforms TCM
over Rayleigh fading channel while suffering a moderate loss of performance over AWGN
channel. If the channel model is nonstationary, in the sense that the propagation environment
changes during transmission, then BICM provides a robust coding scheme.
The main idea of BICM is therefore to transform the channel generated by the multilevel con-
stellation χ into parallel and independent binary channels. For transmission of complex mod-
ulation, channel is not binary but after bit interleaving, any transmission of a multilevel signal
from χ with |χ| = 2m , can actually be thought of as taking place over m parallel channels,
each carrying one binary symbol from the signal label. However, these channels are generally
not independent, due to the constellation structure. To make them independent, binary sym-
bols are interleaved over infinite length before being used as signal labels. The maximum-
likelihood decoding (MLD) of BICM requires combined demodulation/decoding, which is
often too complicated to implement. As a result, MLD is separated at the receiver, concatenat-
ing soft-metrics computation, deinterleaving and decoding. BICM block diagram is shown in
fig. 1 which is the concatenation of an encoder for a code C with an interleaver π followed
by a modulator (µ, χ). In the decoder, the metrics reflect the fact of bits separation. Suppose
that the code word to be transmitted is c. After interleaving and modulation, we transmit the
codeword
x = ( x1 , x2 , ......., xn )
and we receive y at the output of a stationary memoryless channel. With symbol interleaving,
we decode by maximizing the metric
n
log p (y|x) = ∑ log p (yk | xk ) (1)
k =1
with respect to x.

The bit interleaver can be seen as a one-to-one correspondence π : k → (k, i ), where k de-
notes the original ordering of the coded bits ck , k denotes the time ordering of the signals
xk transmitted, and i indicates the position of the bit ck in the symbol xk . Let χib denote the
subset of all signals x ∈ χ whose label has the value b ∈ {0, 1} in position i. Then the ML bit
metric is given as

λi yk , ck = log ∑ p (yk | x ) where ck ∈ [0, 1] and i = 1, 2, ...., log |χ| (2)
x ∈χic
k

So in case of BICM, it is the summation of bit metrics λi yk , ck instead of the symbol metrics
log p (yk | xk ) for decoding. i.e.

ĉ = arg max ∑ λi yk , ck (3)
c∈C
k
x1 OFDM
Stream-1 π1 µ 1 , χ1 (IFFT + CP
Encoder-1
insertion)
2
OFDM
Stream-2 π2 µ 2 , χ2 x2 (IFFT + CP
Encoder-2
insertion)
Bit Spatial
Stream Spreading
. . . .
. . . . 3
. . . .
.
OFDM .
π nr µ nr , χ nr x nr
Encoder-nr (IFFT + CP .
Stream-nr
insertion)
nt
Fig. 2. Block diagram of Transmitter of nt × nr BICM MIMO OFDM system. π1 denotes

random interleaver, µ1 labeling map, χ1 signal set and x1 complex symbols vector for stream-
1.
The bit metrics (2) may be computationally too complex for implementation. Suboptimal sim-
plified branch metric can be obtained by the log-sum approximation log ∑ j z j ≈ max j log z j .
This yields

λi yk , ck = max log p (yk | x ) = min |yk − hk x |2 (4)
x ∈χic
x ∈χic
k k
where hk denotes the Rayleigh coefficient.
2.1.2 BICM MIMO OFDM System

We consider a MIMO broadcast system (without CSIT) which is a nt × nr (nt ≥ nr ) BICM
MIMO OFDM system with nr spatial streams as shown in figs. 2 and 3. We effectively reduce
this to nr × nr system by antenna cycling at the transmitter (Foschini & Gans, 1998) with each
stream being transmitted by one antenna in any dimension. The antenna used by a particular
stream is randomly assigned per dimension so that each stream sees all degrees of freedom
of the channel. Let the spatial streams be x1 , · · · , xnr . xl is the symbol of xl over a signal set
χl ⊆ C with a Gray labeling map µl : {0, 1}log2 |χl | → χl . During the transmission of l-th
spatial stream, the code sequence cl is interleaved by πl and then is mapped onto the signal

sequence xl ∈ χl . Bit interleaver for the l-th stream can be modeled as πl : k → (k, i ) where

k denotes the original ordering of the coded bits ck of the l-th stream, k denotes the time
ordering of the signal xl,k and i indicates the position of the bit ck in the symbol xl,k .
We assume that the frequency reuse factor is one and cyclic prefix (CP) of appropriate length
is added to the OFDM symbols. Cascading IFFT at the transmitter and FFT at the receiver
Standard
SIMO Soft Input
1 πn−1 Stream-nr
Demodulator r
Decoder
for
Stream-nr
OFDM
(FFT+CP
Removal)
2
Demodulator
Soft Input
for π2−1 Stream-2
y Stream-2
Decoder
OFDM
(FFT+CP
Removal)
Demodulator
Soft Input
for
π1−1 Stream-1
- - Stream-1
Decoder
nr
x1 Convolutional
h1 µ 1 , χ1 π1
Encoder-1
OFDM
(FFT+CP
Removal)
To be subtracted for subsequent detection of Stream-3
Fig. 3. Block diagram of SIC Receiver of BICM MIMO OFDM system. π1−1 denotes deinter-
leaver and h1 denotes the channel seen by stream-1.
with CP extension, transmission at the k-th frequency tone can be expressed as:-
yk = h1,k x1 + h2,k x2 + · · · + hnr ,k xnr + zk , k = 1, 2, · · · , T (5)

= Hk xk + zk
T
where Hk = h1,k · · · hnr ,k i.e. the channel at the k-th frequency tone, xk = x1,k , · · · , xnr ,k
and (.) T indicates the transpose operation. Each subcarrier corresponds to a symbol x from a
constellation map χ1 , · · · χnr . yk , zk ∈ C nr are the vectors of received symbols and circularly
symmetric complex white Gaussian noise of double-sided power spectral density N0 /2 at the
nr receive antennas. hl,k ∈ C nr is the vector characterizing flat fading channel response from
l-th transmitting antenna to nr receive antennas at k-th subcarrier. This vector
has complex-

valued multivariate Gaussian distribution with E hl,k = 0 and E hl,k h†l,k = I. The antennas
at the transmitter are also assumed to be sufficiently spaced and therefore are uncorrelated.
The complex symbols x1,k , · · · , xnr ,k of the spatial streams are assumed to be independent with
variances σ12 , · · · , σn2r respectively. The channels at different subcarriers are also assumed to
be independent. Bit metric for the bit ck at the i-th location of the symbol xl,k is given as

i
1 2
λl yk , ck = log ∑ · · · ∑ · · · ∑ exp − y − Hk x
x1 ∈ χ1 i x nr ∈ χ nr N0 k
xl ∈χl,c

k
Applying log-sum approximation we have:-

λil yk , ck ≈ min yk − Hk x 2 (6)
x1 ∈χ1 ··· xl ∈χil,c ··· xnr ∈χnr

k
2.2 Information Theoretic View

We now calculate the mutual information of this system for the cases of Gaussian and finite
sized constellation inputs.
2.2.1 Gaussian Inputs

The system equation ignoring the frequency index takes the form:-
y = h1 x 1 + h2 x 2 + · · · + h n r x n r + z (7)
Since the receiver knows the realization of H, the channel output is the pair (y; H) =
(Hx + z; H). The mutual information between input and output is then Telatar (1999)
I (x; (y, H)) = I (x; H) + I (x; y|H)

= I (x; y|H)
= EH I (x; y|H = H )
For the Gaussian inputs, we consider the following two cases:-

1. Spatial streams of uniform power and non-uniform rate.
2. Spatial streams of uniform rate and non-uniform power.
For Gaussian inputs, channel capacity of the system as per the chain rule (Foschini & Gans,
1998) is
I ( x1 , x2 · · · x nr ; y) = I ( x1 ; y) + I ( x2 ; y| x1 ) + · · · + I ( x nr ; y| x1 , x2 · · · x nr −1 )
The terms in the summation represent the channel capacities of each spatial stream once they
are detected in the successive subtractive cancellation way. Conditioned on the channel, these
terms can be written as:-
−1
I ( x1 ; y|H) = log2 det I + σ12 h1 h1† N0 I + σ22 h2 h2† + · · +σn2r hnr h†nr
−1
I ( x2 ; y|H, x1 ) = log2 det I + σ22 h2 h2† N0 I + σ32 h3 h3† + · · +σn2r hnr h†nr
and
σn2
I ( xnr ; y|H, x1 , x2 · · · , xnr −1 ) = log2 1 + r hnr 2
N0
where H = [h1 h2 · · · hnr ] is the channel matrix. Fig. 4 shows the ergodic capacity for the case
of 2 × 2 system with spatial streams of uniform power and nonuniform rate. Note that SNR
σ2 + σ2
is the received SNR per antenna i.e. SNR = 1N0 2 . It is evident that the stream to be detected
first has lower capacity as compared to the stream to be detected last which enjoys higher
diversity.
Fig. 5 compares two cases of spatial streams with uniform power and nonuniform rate and
spatial streams with uniform rate and nonuniform power for 2 × 2, 3 × 3 and 4 × 4 systems.
Key to the optimality of stripping is the use of Gaussian inputs as long as the stripping de-
coders incorporate MMSE filters against yet undecoded streams at each successive cancella-
tion stage. Successive stripping requires that each stream must be transmitted at a different
rate with uniform power. We investigate a slightly suboptimal solution where we guarantee
equal rate with nonuniform powers on each stream. Numerical optimization revealed that
uniform rate and nonuniform power distribution leads to negligible suboptimality as shown
in fig. 5.
12
I(x1 , x2 ; y)
I(x1 ; y|x2 )
10
I(x2 ; y)
Channel Capacity (bits/sec/Hz)
0
−5 0 5 10 15 20
SNR
Fig. 4. Capacity of 2 × 2 system for Gaussian alphabets for the case of uniform power and
nonuniform rate spatial streams.
25
20
Channel Capacity (bits/sec/Hz)
4x4 System
15
3x3 System
10
2x2 System
5
0
−5 0 5 10 15 20
SNR
Fig. 5. Capacity of 2 × 2, 3 × 3 and 4 × 4 systems for Gaussian alphabets for the cases of spatial
streams of uniform power and nonuniform rate and the spatial streams of uniform rate and
nonuniform power. Note that the circles indicate the case of uniform power and nonuniform
rate spatial streams while crosses indicate the case of uniform rate and nonuniform power
spatial streams.
2.2.2 Finite Sized Constellation Inputs

To reduce the complexity and enhance the understanding of mutual information for finite
sized constellation inputs, we restrict to the case of dual stream transmission. The system
equation ignoring the frequency index takes the form:-
y = h1 x1 + h2 x2 + z (8)
Mutual information expression for the dual streams from the chain rule (Foschini & Gans,
1998) is given as
I (y; x1 , x2 ) = I (y; x1 ) + I (y; x2 | x1 ) (9)
For equal power distribution, I (y; x1 ) < I (y; x2 | x1 ) dictating rate of first stream being less
than rate of second stream (R1 < R2 ). For finite size QAM constellation with x1 ∈ M1 and x2 ∈
M2 , the mutual information expression conditioned on the channel takes the form (Ghaffar &
Knopp, 2008a)
I (y; x1 |H) = H ( x1 |H) − H ( x1 |y, H)

= log M1 − H ( x1 |y, H) (10)
where H (.) = − E log p (.) is the entropy function. Second term of eq. (10) is given as:-

1
H ( x1 |y, H) = ∑ p ( x1 , y, H) log dydH
x1 y H p ( x1 |y, H)

p (y, H)
=∑ p ( x1 , y, H) log dydH
x1 y H p ( x1 , y, H)

∑ x ∑ x p y| x1 , x2 , H
= ∑∑ p ( x1 , x2 , y, H) log 1 2
dydH (11)
x1 x2 y H ∑ x p y| x1 , x2 , H
2
For our purposes, it suffices to note that for each choice of x1 and x2 , there are two sources of
randomness in the choices of channel and noise. The above quantities can be easily approxi-
mated numerically using sampling (Monte-Carlo) methods with Nz realizations of noise and
NH realizations of the channel i.e.

1 2
NH Nz ∑ ∑ exp − N0 y − h x
1 1 − h x
2 2
1 x1 x2
H ( x1 |y, H) = ∑ ∑ ∑ ∑ log
M1 M2 Nz NH x1 x2 H z 2
∑ x exp − N10 y − h1 x1 − h2 x2
2

2
NH Nz ∑ x ∑ x exp − N10 h1 x1 + h2 x2 + z − h1 x1 − h2 x2
1 1 2
=
M1 M2 Nz NH ∑ ∑ ∑ ∑ log

2
(12)
x1 x2 H z
∑ x exp − N10 h2 x2 + z − h2 x2
2
Similarly the mutual information of second stream conditioned on the channel when first
stream has been detected is given by:-
I (y; x2 | x1 , H) = H ( x2 | x1 , H) − H ( x2 |y, x1 , H)

1
= log M2 − ∑ ∑ p ( x1 , x2 , y, H) log dydH
x1 x2 y H p ( x2 |y, x1 , H)

p (y, x1 , H)
= log M2 − ∑ ∑ p ( x1 , x2 , y, H) log dydH
x1 x2 y H p ( x1 , x2 , y, H)

∑ x p y| x1 , x2 , H
= log M2 − ∑ ∑ p ( x1 , x2 , y, H) log 2
dydH (13)
x1 x2 y H p (y| x1 , x2 , H)
Estimation of this quantity using Monte-Carlo simulation

2
∑ x exp − N10 y − h1 x1 − h2 x2
NH Nz
1 2
I (y; x2 | x1 , H) = log M2 − ∑ ∑ ∑ log
M1 M2 Nz NH ∑
x1 x2 H z exp − N10 y − h1 x1 − h2 x2 2

2
NH Nz ∑ x exp − N10 h2 x2 + z − h2 x2
1 2
= log M2 − ∑ ∑ ∑ log
M1 M2 Nz NH ∑
x1 x2 H z exp − N10 z2
(14)
Fig. 6 shows the capacity of first stream once second stream is not yet decoded for different
combinations of finite constellation alphabets. For moderate values of SNR, the capacity of
first stream is a function of the yet undetected second stream and this capacity decreases as
the rate (constellation size) of second stream increases. This degradation is not observed at
low and high values of SNR as at low SNR, two streams are orthogonal while at high SNR,
second stream can be perfectly stripped off leading to detection of first stream. Rate of first
stream being a function of the rate of second stream leads to nonuniform rates in uniform
power dual stream scenario and this leads to the following proposed broadcast strategy.
2.3 Broadcast Strategy

We restrict to dual stream scenario for the broadcast case. The broadcast approach in dual
stream scenario based on UEP (MAC-outage (Shamai & Steiner, 2003)) is motivated by the
capacity of a Gaussian broadcast channel with two users i.e.
C = I ( x1 ; y1 ) + I ( x2 ; y2 | x1 ) (15)
where user 2 sees a better channel and so is able to decode and strip off the interference.
The broadcast strategy (Ghaffar & Knopp, 2008a) incorporates the transmission of two spatial
streams of uniform power and nonuniform rate and incorporates two levels of performance.
The reliably decoded information rate depends on the state of the channel which is determined
by monitoring the received SNR being above or below a certain threshold. Transmitter is
operating at a constant power and data rate but the limited adaptability of the system helps
receivers to gear up to a higher data rate as the channel conditions improve.
5
x1=QAM64
4
I(x1;y) x1=QAM16
3
2
x =QPSK
1
1
x x =0
2
O x =QPSK
2 * x2=QAM 16 + x2=QAM 64
0
−5 0 5 10 15 20 25 30 35
SNR
Fig. 6. Capacity of first stream in dual-stream broadcast approach for finite size alphabets
once second stream is not known. Both streams have equal power. x2 = 0 indicates the
special case when second stream has been decoded and stripped off. Note that SNR includes
power of both streams.
Low priority/quality users are able to decode low rate stream x1 while high priority/quality
users are able to decode both low and high rate streams i.e. x1 and x2 by successive stripping.
The rates of two streams are
R1 ≤ I (y; x1 ) (16)
and
R2 ≤ I (y; x2 | x1 ) (17)
The notion of priority/quality is typically the received SNR and/or stream decoupling. The
users are divided into two groups i.e. near-in users and far-out users based on their received
SNR. The lower rate stream x1 is designed for a lower value of SNR i.e. SNR1 while the higher
rate stream x2 is designed for higher value of SNR i.e. SNR2 . The received SNR of a particular
user dictates two decoding options.
1. If SNR2 >SNR≥SNR1 , the user decodes x1 .
2. If SNR≥SNR2 , the user decodes both streams i.e. x1 and x2 . The user first decodes low
rate stream x1 , strips it out and then decodes high rate stream x2 .
This leads us to SIC detection based MIMO broadcast scenario with uniform power and
nonuniform rate spatial streams. We now discuss the detectors for such broadcast scenario.
2.4 Detectors
The detectors discussed in this section are valid not only for spatially multiplexed MIMO
systems but may be extended to other types of STC systems. We discuss two types of detectors
as MMSE detector and low complexity max log MAP detector (Ghaffar & Knopp, 2009b).
2.4.1 MMSE
The frequency domain MMSE filter for x1,k is given as
−1
MMSE † −1 †
h1,k = h1,k R1,k h1,k + σ1−2 h1,k −1
R1,k (18)
† †
where R1,k = σ22 h2,k h2,k + σ32 h3,k h3,k + · · · + σn2r hnr ,k h†n,k + N0 I. After the application of MMSE
filter we get
yk = αk x1,k + zk (19)
where zk is assumed to be zero mean complex Gaussian random variable with variance
MMSE MMSE† MMSE
Nk = h1,k R1,k h1,k and αk = h1,k h1,k . Gaussianity has been assumed for post
detection interference which increases the suboptimality of MMSE in the case of less number
of interferers. Bit metric for the bit ck on first stream is given as:-

1
λ1i yk , ck ≈ min | y k − α k x1 |2 (20)
i
x1 ∈χ1,c Nk

k
i
where χ1,c denotes the subset of the signal set x1 ∈ χ1 whose labels have the value ck ∈

k
{0, 1} in the position i. This metric has computational complexity O (|χ1 |).
2.4.2 Low complexity max log MAP Detector

The max log MAP bit metric as per (6) is given as

λ1i yk , ck ≈ min y − h1,k x1 − · · · − hn ,k xn 2 (21)
i k r r
x1 ∈χ1,c
,x2 ∈χ2 ,··· ,xnr ∈χnr
k
which has computational complexity O (|χ1 | · · · |χnr |). For brevity we drop the frequency

index k and the bit position index k i.e.
λ1i (y, c) ≈ min y − h1 x 1 − · · · − h n r x n r 2

i ,x ∈ χ ,··· ,x ∈ χ
x1 ∈χ1,c 2 2 nr nr
 
 nr 2 nr −1 nr † nr 
2 † ∗
= min y + ∑ h j x j + 2 ∑ ∑ h j x j (hl xl )− 2 ∑ h j y x j
i ,x ∈ χ ,··· ,x ∈ χ
x1 ∈χ1,c 2 2 nr nr
 j =1 j =1 l = j +1 j =1


 nr −1 2 nr −1 nr −1 nr −1

= min y2 + ∑ h j x j + 2 ∑ ∑ p jl x ∗j xl − 2 ∑ y j x ∗j
i ,x ∈ χ ,··· ,x ∈ χ
x1 ∈χ1,c 2 2 nr nr
 j =1 j =1 l = j +1 j =1

nr −1 
+2 ∑ p jnr x ∗j xnr − 2ynr xn∗r + hnr xnr 2 (22)
j =1

where yk = h†k y be the matched filter (MF) output for k-th stream and pkm = h†k hm be the
cross correlation between k-th and m-th channel. Breaking some of the terms in their real and
imaginary parts with subscripts (.) R and (.) I indicating real and imaginary parts of a complex
number, we have

 nr −1 2 nr −1 nr −1 nr −1

λ1i (y, c) = min ∑ h j x j + 2 ∑ ∑ p jl x ∗j xl − 2 ∑ y j x ∗j
i ··· x ∈ χ
x1 ∈χ1,c nr nr
 j =1 j =1 l = j +1 j =1
 
nr −1
+ 2 ∑ p jnr ,R x j,R + p jnr ,I x j,I − 2ynr ,R  xnr ,R + hnr 2 x2nr ,R
j =1
  
nr −1 
+ 2 ∑ p jnr ,R x j,I − p jnr ,I x j,R − 2ynr ,I  xnr ,I + hnr 2 x2nr ,I (23)
j =1

This equation reduces one complex dimension of the system. For xnr belonging to equal en-
ergy alphabets, the bit metric is written as

 nr −1 2 nr −1 nr −1 nr −1

λ1i (y, c) = min ∑ h j x j + 2 ∑ ∑ p jl x ∗j xl − 2 ∑ y j x ∗j
i ··· x
x1 ∈χ1,c nr −1 ∈ χ nr −1
 j =1 j =1 l = j +1 j =1

nr −1 nr −1 

− 2 ∑ p jnr ,R x j,R + p jnr ,I x j,I − 2ynr ,R | xnr ,R | − 2 ∑ p jnr ,R x j,I − p jnr ,I x j,R − 2ynr ,I | xnr ,I |
j =1 j =1 
For xnr belonging to non-equal energy alphabets, it’s real and imaginary part which minimizes
(23) are given as

n −1
∑ j=r 1 p jnr ,R x j,R + p jnr ,I x j,I − ynr ,R
xnr ,R → −
hnr 2

n −1
∑ j=r 1 p jnr ,R x j,I − p jnr ,I x j,R − ynr ,I
xnr ,I → − (24)
hnr 2
where → indicates the quantization process in which amongst the finite available points, the
point closest to the calculated continuous value is selected.
This bit metric implies reduction in the complexity to O (|χ1 | · · · |χnr −1 |). Reduction of one
complex dimension without any additional processing is a fundamental result of significant
importance for lower dimensional systems. Additionally this bit metric is based on MF out-
puts and channel correlations and is therefore simpler for fixed point implementations. The
intricacy in the practical implementation of a higher dimensional MIMO system due to space
(requisite antenna spacing) and technology constraints underlines the significance of complex-
ity reduction algorithms for lower dimensional systems. MMSE based demodulators involve
computationally complex operations of matrix inversions which are very hard for fixed point
implementations. Moreover MMSE demodulator additionally needs the knowledge of noise
variance.
0
10
−1
10
FER QAM16
−2
10
QAM64
QPSK
Low Complexity Max Log MAP
MMSE SIC
−3 MMSE PIC
10
4 6 8 10 12 14 16 18 20 22 24 26
SNR
Fig. 7. 2 × 2 system with uniform rate and nonuniform power spatial streams. For QPSK
σ12 = 0.63PT , σ22 = 0.37PT , for QAM 16 σ12 = 0.67PT , σ22 = 0.33PT while for QAM64 σ12 =
0.70PT , σ22 = 0.30PT .
0
10
10
−1 x1=QAM16
1
FER of x
−2 x1=QPSK
10
−3
10 O x2=QPSK
* x2=QAM 16 + x2=QAM 64
4 6 8 10 12 14 16 18
SNR
Fig. 8. Performance of lower rate stream in 2 × 2 BICM MIMO OFDM system using 802.11n
convolutional code. Continuous lines indicate low complexity max log MAP detector while
dashed lines indicate MMSE detector.
2.5 Simulations
We consider a 2 × 2 BICM MIMO OFDM system using the de facto standard, 64 state rate-1/2
convolutional encoder of 802.11n standard (802.11n, 2006) and rate-1/2 punctured turbo code
0
10
−1
10
−2
10
1
FER of x
−3
10 x1=QAM64
−4
10 x =QAM16
1
−5 x =QPSK
10 1
−6
10
O x2=QPSK
* x2=QAM 16 + x2=QAM 64
4 6 8 10 12 14 16 18 20
SNR
Fig. 9. Performance of lower rate stream in 2 × 2 BICM MIMO OFDM system using 3GPP LTE
turbo code. Continuous lines indicate low complexity max log MAP detector while dashed
lines indicate MMSE detector. Block length of the lower rate stream is 1296 bits while number
of decoding iterations are 5.
proposed for 3GPP LTE (LTE, 2006) 1 . MIMO channel has iid Gaussian matrix entries with
unit variance. The channel is independently generated for each time instant and perfect CSI
at the receiver is assumed. Furthermore, all mappings of coded bits to QAM symbols use
Gray encoding. We consider MMSE and low complexity max log MAP detector. There are
two scenarios.
In first scenario, spatial streams of uniform rate and nonuniform power are transmitted in 2 ×
2 MIMO broadcast system. The upcoming WLAN standard 802.11n (802.11n, 2006) supports
the codeword sizes of 648, 1296, and 1944 bits. For our purposes, we selected the codeword
size of 1296 bits and coding scheme of convolutional coding. We focus on the frame error
rates (FER) of the system. We consider the low complexity max log MAP and MMSE SIC
approach in which the higher power stream is detected first and is subsequently stripped
off leading to the detection of lower power stream. With PT being the total power available,
the power distribution between two streams is optimized to equate their rates in the desired
PT
SNR region where SNR is defined as the received SNR per antenna i.e. N 0
. As a reference,
MMSE parallel interference cancellation (PIC) has also been simulated in which two streams
are independently detected using MMSE filters and two streams have equal power. Fig. 7
shows the improved performance of low complexity max log MAP approach with respect to
both MMSE SIC and MMSE PIC approach. The gap widens as the constellation proliferates
i.e QAM 16 and QAM64 which is attributed to the higher suboptimality of MMSE for larger
sized constellations.
In second scenario, spatial streams of uniform power and nonuniform rate are transmitted in
2 × 2 MIMO broadcast system. We focus on the FER of first stream (lower rate) as subsequent
1 LTE turbo decoder design was performed using the coded modulation library
www.iterativesolutions.com
to stripping, the detection of second stream (higher rate) is trivial (using SIMO detectors). The
frame length of first stream is fixed to 1296 information bits as per 802.11n (802.11n, 2006).
Figs. 8 and 9 compare the performance of low complexity max log MAP detector with MMSE
detector. The max log MAP detector performs significantly better than the MMSE detector.
Degradation of the performance for first stream as the rate (constellation size) of second stream
increases confirms the earlier result of sec. 2.2.2 that rate of first stream is a function of the rate
of second stream.
3. Interference Suppression for future Wireless Systems

To cope with the ever-increasing demands on the higher spectral efficiency, appendage of
spatial dimension (MIMO) needs to be coupled with a tight frequency reuse as is advocated
in the future wireless communication systems as 3GPP LTE (LTE, 2006) and LTE-Advanced
(LTE-A, 2008). Adaptive modulation and coding schemes will be supported in the next gener-
ation wireless systems which combined with the diversified data services will lead to variable
transmission rate streams. These system characteristics will overall lead to an interference-
limited system. Most state-of-the-art wireless systems deal with the interference either by
orthogonalizing the communication links in time or frequency (Gesbert et al., 2007) or allow
the communication links to share the same degrees of freedom but model the interference as
additive Gaussian random process (Russell & Stuber, 1995). Both of these approaches may be
suboptimal as first approach entails an a priori loss of the degrees of freedom in both links, in-
dependent of the interference strength while second approach treats the interference as pure
noise while it actually carries information and has the structure that can be potentially ex-
ploited in mitigating its effect.
3GPP LTE (LTE, 2006) has chosen orthogonal frequency division multiple access (OFDMA)
technology for the downlink in order to provide multiple access and eliminate the intracell
interference. However frequency reuse factor being 1 will lead to intercell interference impair-
ments among neighboring cells. Intercell interference coordination techniques (Gesbert et al.,
2007) are studied to minimize the interference level while spatial interference cancellation fil-
ters are the focus of attention to cancel the interferers which will be 1 in most cases (near cell
boundaries) and 2 in rare cases (near cell corners). Different spatial interference cancellation
techniques involving equalization and subtractive cancellation (Bladsjö et al., 1999) (Debbah
et al., 2000) have been proposed in the literature. Amongst them, MMSE linear detectors are
being considered as likely candidates for 3GPP LTE (Dahlman et al., 2006). The suboptimality
of MMSE for non Gaussian alphabets in low dimensional systems (less number of interfer-
ers) has already been discussed and simulated in the previous sections and moreover MMSE
detection being based on interference attenuation is void of exploiting the interference struc-
ture in mitigating its effect. Though not optimal, but their low complexity still makes them
attractive for practical systems.
Optimal strategy for treating the interference in the regime of very strong (Carleial, 1975) and
very weak interference is well known however if the interference is in the moderate region, no
optimal strategy is known but partial decoding of interference can significantly improve per-
formance (Han & Kobayashi, 1981). This part of the chapter discusses a low complexity spatial
interference cancellation algorithm for single frequency reuse synchronized cellular networks
in the presence of one strong interferer. This algorithm is based on the low complexity max
log MAP detector and benefits from its ability to exploit interference structure in mitigating
its effect. The algorithm encompasses two strategies for interference mitigation i.e. interfer-
ence suppression and interference cancellation and their selection in the receiver is dictated
x1 x2
h1 h2
BS-1 BS-2
MS
h3
x3
BS-3
Fig. 10. Interference cancellation in single frequency cellular network. x1 is the desired signal
while x2 and x3 are the interference signals.
by the relative strength and the rate of interfering stream. In the scenario of interfering stream
being weak or of higher rate relative to the desired stream, thereby making it unfeasible to
be decoded, the mobile station (MS) resorts to the strategy of interference suppression. It can
be interpreted as partial decoding of the interference which is the recommended strategy in
the regime of moderate interference (Han & Kobayashi, 1981). When the interfering stream
is relatively stronger or is of lower rate thereby making it feasible to be decoded, MS adopts
interference cancellation strategy (subtractive cancellation) which is the optimal strategy in
the case of strong interference (Carleial, 1975).
3.1 System Model

The system model as shown in fig. 10 remains same as described in the sec.2.1.2 with 3 spatial
streams. However these streams arriving at the receiver (MS) are now from three different
base stations (BS) thereby ensuring independent channels. The MS has receive diversity with
nr receive antennas. All the BSs are assumed to be synchronous.
3.2 Information Theoretic view

For better understanding of the effect of strength and rate (alphabet size) of interference, the
case of one strong interference is considered in this section. The focus is on the mutual in-
formation of the desired stream in the presence of one strong interferer (Ghaffar & Knopp,
2009b).

1 ∑ x1 p (y| x1 )
I (y; x1 ) = log M1 − ∑ p (y| x1 ) log dy (25)
M1 x1 y p (y| x1 )
Fig. 11 shows the mutual information of the desired stream in the presence of the interference
stream. We define the term α = σ22 /σ12 . Mutual information of the desired stream is a func-
tion of the rate as well as the strength of the interference stream. For moderate values of α
and when the interference has a lower rate (smaller constellation size) relative to the desired
stream, as the interference strength increases, the mutual information of the desired stream
Fig. 11. Mutual information of the desired stream x1 in the presence of the interference stream
x2 for different constellations. SNR is 4.5 dB for x1 =QPSK, 11 dB for x1 =QAM16 and 13 dB for
x1 =QAM64. Note that the flash sign indicates a discontinuity of abscissa.
increases. However when the interference stream has a higher rate as compared to the rate of
the desired stream, this behavior is observed for higher values of α. This can be interpreted
as the decoding capability of the MS of the interference in the presence of the desired stream.
Once the interference strength and its rate relative to the strength and the rate of the desired
stream permits the decoding of the interference, we observe an increase in the mutual infor-
mation of the desired stream with the increase of α. Fig. 11 also authenticates the well known
result of Gaussian being the worst case interference however the gap decreases as the rate of
the interference stream increases. This diminution of gap may be related to the proximity of
the behavior of large size constellations to Gaussianity as both are characterized by high peak
to average power ratios.
3.3 Interference Mitigation Strategies

Based on the low complexity max log MAP detector, an interference mitigation strategy (Ghaf-
far & Knopp, 2009a) is discussed which is based on the partial decoding of the interference
in the regime when interference because of its relative rate or strength is undecodable and
subtractive cancellation when the interference is quite strong and is decodable. This strategy
is based on exploiting the structure of the interference in mitigating its effect once subtractive
cancellation is not possible and resorting to subtractive cancellation otherwise. So there are
two options for interference mitigation.
1. In the regime when interference has higher rate or is weaker in strength relative to the
desired stream thereby rendering the absolute decoding of interference unfeasible, tar-
get stream is decoded using the low complexity max log MAP detector which takes into
account the effect of interference and can be termed as the partial decoding of interfer-
ence or partial joint decoding. This approach is termed as interference suppression.
2. In the regime when interference has lower rate or is stronger in strength relative to the
desired stream thereby rendering the absolute decoding of the interference feasible, the
interference stream is decoded using low complexity max log MAP detector, stripping
it off and then decoding the desired stream. This approach is termed as interference
cancellation.
The factors that will decide the strategy to be adopted will be the relative rate and the strength
of the interference stream comparative to the desired stream. The requisites for this algorithm
are the knowledge of interference channel and the modulation and coding scheme (MCS) of
interfering stream. The BSs need to be synchronous with pilot signals from the adjacent BSs
to be orthogonal to meet these requisites.
3.4 Performance Analysis

This section deliberates on the performance analysis of two detectors for detecting the desired
stream in the presence of interfering stream (Ghaffar & Knopp, 2009c).
3.4.1 PEP Analysis - Max Log MAP Detector

The conditional PEP i.e. P (c1 → ĉ1 |H) of max log MAP detector is given as

1
y − h1,k x1 − h2,k x2 2 ≥
P (c1 → ĉ1 |H) = P ∑ min k
x ∈χ i ,x ∈χ N
k 1 1,ck 2 2 0

1 2
∑ x ∈χimin,x ∈χ N0 yk − h1,k x1 − h2,k x2  (26)
1 2 2
k
1,ĉ
k

where H = [H1 · · ·HK ] i.e. the complete channel for the transmission of the codeword c1
and Hk = h1,k h2,k i.e. the channel at k-th frequency tone. For the worst case scenario once
d (c1 − ĉ1 ) = d f ree , the inequality on the right hand side of (26) shares the same terms on all
but d f ree summation points for which ĉk = c̄k where (¯.) denotes the binary complement. Let
1
y − h1,k x1 − h2,k x2 2
x̃1,k , x̃2,k = arg min k
i
x1 ∈χ1,c ,x2 ∈χ2 N0

k
1
y − h1,k x1 − h2,k x2 2
x̂1,k , x̂2,k = arg min k (27)
i
x1 ∈χ1, c̄ ,x2 ∈χ2 N0
k
2
As x1,k and x2,k are the transmitted symbols so yk − h1,k x1,k − h2,k x2,k ≥

y − h1,k x̃1,k − h2,k x̃2,k 2 . The conditional PEP is given as
k
 
1 2 1 2
P (c1 → ĉ1 |H) ≤ P  ∑ y − h1,k x1,k − h2,k x2,k ≥ ∑ y − h1,k x̂1,k − h2,k x̂2,k 
k k
N
k,d f ree 0
N
k,d f ree 0
 

 1 
= Q  ∑ Hk (x̂k − xk )2 
k,d
2N 0
f ree

1 †
† †
=Q vec H ∆ vec H (28)
2N0
T
where H = H1 · · · Hk,d f ree , x̂k = x̂1,k x̂2,k and ∆ = Inr ⊗ DD† while

D = diag x̂1 − x1 , x̂2 − x2 , · · · , x̂k,d f ree − xk,d f ree . Q is the Gaussian Q-function i.e. Q (y) =
∞ − x2 /2
√1 dx and vec indicates vectorization of a matrix. For a Hermitian quadratic form
2πy e
in complex Gaussian random variable q = m† Am where A is a Hermitian matrix and col-
umn vector m is a circularly
symmetric complex Gaussian vector i.e. m ∼ N C (µ , ∑ ) with
µ = E [m] and ∑ = E mm† − µµ † , the moment generating function (MGF) is

exp −tµµ † A (I + t∑ A)−1 µ
†
E exp −tm Am = (29)
det (I + t∑ A)
2
Using Chernoff bound Q ( x ) ≤ 12 exp −2x and the MGF, PEP is upper bounded as
1
P (c1 → ĉ1 ) ≤
1 ∆
2 det I + 4N0 I∆
1
= d f ree
nr (30)
2 ∏k= 1 1+ 1
4N0 x̂k − xk 2
x̂k − xk 2 ≥ d21,min + d22,min if x̂2,k = x2,k and x̂k − xk 2 ≥ d21,min if x̂2,k = x2,k . There ex-
T
ists 2d f ree possible vectors of x̂2,1 , · · · , x̂2,d f ree basing on the binary criteria that x̂2,k is equal
or not equal to x2,k . Taking into account all these cases combined with their corresponding
probabilities, the PEP is upper bounded as
 
nr d f ree d j d f ree − j 
 f ree 1 − P x̂2,k = x2,k
1 4N0  d f ree P x̂2,k = x2,k 
P (c1 → ĉ1 |H) ≤  ∑ C 
2 σ12 d˘21,min  j =0 j σ22 d˘22,min
jnr 
1 + 2 ˘2
σ1 d1,min
(31)
where d2j,min = σj2 d˘2j,min with d˘2j,min being the normalized minimum distance of the constella-
d
tion χ j for j = {1, 2} and C j f ree is the binomial coefficient. P x̂2,k = x2,k has been derived in
the following section.

3.4.2 P x̂2,k = x2,k
Considering (27), P x̂2,k = x2,k |h1,k , h2,k , x1,k is

P x̂2,k = x2,k |h1,k , h2,k , x1,k
†

2
= P −2 h1,k x1,k − x1 + zk h2,k x2,k − x2 ≥
h2,k x2,k − x2
|Hk , x1,k


 † 

h2,k x2,k − x2
2 2 h1,k x1,k − x1 h2,k x2,k − x2
= Q + 


2N0 N0
h2,k x2,k − x2
2

Using the relation Q ( a + b) ≤ Q ( amin − |bmax |) and a† b̂ ≤ a where b̂ is the unit vector
we get
 
1 h2,k 2 d2 h1,k 2 d2 h2,k h1,k d2,min d1,max
2,min 1,max
P x̂2,k = x2,k |h1,k , h2,k ≤ exp − − + 
2 4N0 N0 N0
(32)
Conditioned
on the norm of h1,k we make two non-overlapping regions as
h2,k ≥ h1,k h1,k and h2,k < h1,k h1,k with the corresponding probabilities
2
as Ph<1 and Ph>1 . Note that in first region h2,k h1,k ≤ h2,k while for second region

h2,k h1,k < h1,k 2 . So
 nr  
1 4N h1,k 2 d2
0 1,max 
P x̂2,k = x2,k ≤ Eh1  2 exp − Eh2 |h1 Ph<1
2 d2,min − 4d2,min d1,max N0
nr
4N0 2 d21,max − d2,min d1,max
+ exp − h1,k Eh2 |h1 Ph1 >
d22,min N0
 
nr nr
1 4N0 N0  1 1 
≤  nr + nr 
2 σ22 d˘22,min σ12 d˘21,max 4σ1 d˘1,max σ2 d˘2,min
1− ˘ 1− ˘
σ2 d2,min σ1 d1,max
(33)

where we upper bound Eh2 |h1 Ph<1 and Eh2 |h1 Ph>1 by 1.

P x̂2,k = x2,k → 0 as σ22 → ∞ while P x̂2,k = x2,k increases as σ22 increases. Eq. (31) demon-
strates a significant result of achieving full diversity by the low complexity max log MAP
detector and converging to the performance of single stream using maximum ratio combin-
ing in the case of very weak and strong interference. In the moderate region, as the strength
of interference increases, P x̂2,k = x2,k reduces and there is a coding gain for the detection of
desired stream contrary to the case of MMSE where there is a coding loss as the interference
gets stronger (shown in the next section).
3.4.3 PEP Analysis - MMSE Detector

3.5 Gaussian Assumption
Conditional PEP for MMSE basing on Gaussian assumption of post detection interference (20)
is given as
 
2 2
| y k − α k x 1 | | y k − α k x 1 |
P (c1 → ĉ1 |H) = P ∑ min ≥ ∑ min  (34)
i
x1 ∈ χ1,c Nk i
x1 ∈ χ1,ĉ Nk
k k
k k

Let
| y k − α k x1 |2 | y k − α k x1 |2
x̃1,k = arg min , x̂1,k = arg min
i
x1 ∈χ1,c Nk i
x1 ∈χ1, c̄
Nk

k k
Considering the worst case scenario d (c1 − ĉ1 ) = d f ree and using the fact that

1
2 2
Nk yk − αk x1,k ≥ 1 yk − αk x̃1,k , the conditional PEP is upper bounded as
Nk
 

α2k 2
P (c1 → ĉ1 |H) ≤ Q  ∑ x̂ − x1,k  (35)
k,d f ree
2Nk 1,k
2
Bounding x̂1,k − x1,k ≥ d21,min and using the Chernoff bound
 
1 d21,min † −1
P (c1 → ĉ1 |H) ≤ exp − ∑ h1,k R2,k h1,k  (36)
2 4 k,d f ree
where the summation in (36) can be written as

T
† −1 † † −1 −1 T T
∑ h1,k R2,k h1,k = h1,1 , · · · , h1,d f ree
diag R2,1 , · · · , R2,d h1,1 , · · · , h1,d f ree
f ree
k,d f ree
−1
The eigenvalues of R2,k are
 −1
 2
σ22 h2,k + N0 , l=1
λl = (37)
 N −1 , l = 2, · · · , nr
0

Using the MGF (29), PEP conditioned on h2 = h2,1 , · · · , h2,d f ree is upper bounded as
d f ree (nr −1) d f ree d f ree

1 4N0 4 2
P c1 → ĉ1 |h2 ≤ ∏ σ22 h2,l + N0
2 d21,min d21,min l =1
Channel independence at each subcarrier yields

d f ree (nr −1) d f ree
1 4N0 4 d f ree
P (c1 → ĉ1 ) ≤ nr σ22 + N0 (38)
2 σ2 d˘2
1 1,min σ2 d˘2
1 1,min
which not only demonstrates the well known result of the loss of one diversity order in MMSE
in the presence of one interferer (Winters, 1984) but also exhibits a coding loss as interference
gets stronger.
3.6 Simulation Results

Moderate and high SNR regime in the interference-limited scenario demands more attention
as when the noise is small, interference will have a significant impact on the performance. Low
SNR regime is less interesting since here the performance is noise-limited and interference is
not having a significant effect. For simulations, we have restricted ourselves to the case of
one strong interference. These simulations have been performed in moderate and high SNR
region while the interference strength is being varied.
We consider 2 BSs each using BICM OFDM system for downlink transmission using the de
facto standard, 64 state (133, 171) rate-1/2 convolutional encoder of 802.11n standard (802.11n,
−1
10
−2
1
FER of x 10
−3
10
O x2=QPSK
* x =QAM 16
2 + x2=QAM 64
−4
10
−5 −4 −3 −2 −1 0 1 2 3 4
INR (dBs)
Fig. 12. Desired stream x1 is QPSK while interference stream x2 is from QPSK, QAM16 and
QAM64. SNR is 4.5 dB. Continuous lines indicate interference suppression while dashed lines
indicate interference cancellation. Dotted lines indicates detection of x1 by MMSE detector.
64−state, rate 1/2 Convolutional Code is used. Note that SNR is with respect to the desired
stream
0
10
−1
10
1
FER of x
−2
10
−3
10
+
−4
10
O x2=QPSK
* x =QAM 16
2
x2=QAM 64
1 2 3 4 5 6 7 8 9 10 11
INR (dBs)
Fig. 13. Desired stream x1 is QAM 16 while interference stream x2 is from QPSK, QAM16 and
QAM64. SNR is 11 dB. Continuous lines indicate interference suppression while dashed lines
64−state, rate 1/2 Convolutional Code is used.
2006) and the punctured rate 1/2 turbo code of 3GPP LTE (LTE, 2006)2 . Each BS has multiple
antennas and employs antenna cycling. MS has two antennas. We consider an ideal OFDM
2 The LTE turbo decoder design was performed using the coded modulation library
www.iterativesolutions.com
0
10
−1
10
−2
10
−3
1
FER of x 10
−4
10
−5
10
−6
10
O x2=QPSK
* x =QAM 16
2 + x2=QAM 64
−7
10
3 4 5 6 7 8 9 10 11 12 13
INR (dBs)
Fig. 14. Desired stream x1 is QAM 64 while interference stream x2 is from QPSK, QAM16 and
QAM64. SNR is 13 dB. Continuous lines indicate interference suppression while dashed lines
Punctured rate 1/2 3GPP turbo code is used with 5 decoding iterations.
based system (no ISI) and analyze the system in frequency domain. Due to bit interleaving
followed by OFDM, this can be termed as frequency interleaving. Therefore SIMO channel at
each sub carrier from BS to MS has iid Gaussian matrix entries with unit variance. Perfect CSI
is assumed at the receiver. Furthermore, all mappings of coded bits to QAM symbols use Gray
encoding. We consider interference suppression and interference cancellation approaches us-
ing low complexity max log MAP detector. For comparison we also consider interference
suppression using MMSE detector.
Figs. 12, 13 and 14 show the FERs of target stream in the presence of one interference stream.
These simulation results show that the dependence of the performance for MMSE detection is
insignificant on the rate of the interference stream but its dependence on interference strength
is substantial. This can be interpreted as a consequence of the attenuation of interference
strength at the output of MMSE filter and the subsequent assumption of Gaussianity for its
behavior. For the low complexity max log MAP detector, a significant improvement is ob-
served in the performance as the rate of interference stream decreases which is in conformity
with the earlier results of mutual information analysis (fig. 11). It is observed that for a given
interference level, the performance is generally degraded as the rate (constellation size) of the
interfering stream increases. The performance gap with respect to MMSE decreases as the
desired and the interference streams grow in constellation size which can be attributed to the
proximity to the Gaussianity of these larger constellations due to their high peak to average
power ratio and to the optimality of MMSE for Gaussian alphabets.
4. References
802.11n (2006). Enhancements for Higher Throughput, IEEE 802.11 WG. IEEE 802.11n/D1.0 Draft
Amendment.
802.16m (2007). Draft IEEE 802.16m Evaluation Methodology, IEEE 802.16m-07/037r1.
Alamouti, S. (1998). A simple transmit diversity technique for wireless communications, IEEE
Journal on Selected Areas in Communications Vol. 16(No. 8): 1451–1458.
Bladsjö, D., Furuskär, A., Jäverbring, S. & Larsson, E. (1999). Interference cancellation using
antenna diversity for EDGE - Enhanced data rates in GSM and TDMA/136, IEEE
Vehicular Technology Conference Proceedings, 1999. VTC 1999-Fall, Vol. 4, pp. 1956–1960.
Caire, G., Taricco, G. & Biglieri, E. (1998). Bit-interleaved coded modulation, Vol. 44(No.
3): 927–946.
Carleial, A. (1975). A case where Interference does not reduce Capacity, IEEE Transactions on
Information Theory Vol. 21(No. 5): 569–570.
Dahlman, E., Ekstrom, H., Furuskar, A., Jading, Y., Karlsson, J., Lundevall, M. & Parkvall,
S. (2006). The 3G long-term evolution - radio interface concepts and performance
evaluation, IEEE 63rd Vehicular Technology Conference VTC-Spring, Vol. 1, pp. 137–141.
Debbah, M., Muquet, B., de Courville, M., Muck, M., Simoens, S. & Loubaton, P. (2000).
A MMSE successive interference cancellation scheme for a new adjustable hybrid
spread OFDM system, IEEE 51st Vehicular Technology Conference Proceedings, 2000.
VTC 2000-Spring Tokyo., Vol. 2, pp. 745–749 vol.2.
Foschini, G. J. & Gans, M. J. (1998). On limits of wireless communication in a fading envi-
ronment when using multiple antennas, Wireless Personal Communications. Vol. 6(No.
3): 311–335.
Gesbert, D., Kiani, S., Gjendemsj, A. & Oien, G. (2007). Adaptation, coordination, and dis-
tributed resource allocation in interference-limited wireless networks, Proceedings of
the IEEE Vol. 95(No. 12): 2393–2409.
Ghaffar, R. & Knopp, R. (2008a). Dual-antenna BICM reception with applications to MIMO
broadcast and single frequency cellular system, IEEE 19-th International Symposium
on Personal, Indoor and Mobile Radio Communications (PIMRC 2008), Cannes.
Ghaffar, R. & Knopp, R. (2008b). Low complexity BICM demodulation for MIMO transmis-
sion, 9th IEEE Workshop on Signal Processing Advances for Wireless Communications,
SPAWC 2008, Recife.
Ghaffar, R. & Knopp, R. (2009a). Interference Suppression for Next Generation Wireless Sys-
tems, IEEE 69-th Vehicular Technology Conference VTC-Spring 2009, Barcelona.
Ghaffar, R. & Knopp, R. (2009b). Spatial Interference Cancellation Algorithm, Proc. IEEE Wire-
less Communications and Networking Conference WCNC 2009, Budapest, Budapest, Hun-
gary.
Ghaffar, R. & Knopp, R. (2009c). Spatial Interference Cancellation and Pairwise Error Proba-
bility Analysis, IEEE International Conference on Communications, ICC 2009.
Golden, G., Foschini, C., Valenzuela, R. & Wolniansky, P. (1999). Detection algorithm and
initial laboratory results using v-blast space-time communication architecture, Elec-
tronics Letters Vol. 35(No. 1): 14–16.
Gonzalez-Lopez, M., Vazquez-Araujo, F., Castedo, L. & Garcia-Frias, J. (2006). Optimized
serially-concatenated LDGM and Alamouti codes for approaching MIMO Capacity,
IEEE 17th International Symposium on Personal, Indoor and Mobile Radio Communica-
tions, PIMRC 2006, pp. 1–5.
Han, T. & Kobayashi, K. (1981). A new achievable rate region for the interference channel,
IEEE Transactions on Information Theory Vol. 27(No. 1): 49–60.
Hochwald, B. & Brink, S. T. (2003). Achieving near-capacity on a multiple-antenna channel,
IEEE Transactions on Communications, Vol. 51(No. 3): 389–399.
Larsson, E. G. & Stoica, P. (2003). Space-Time Block Coding for Wireless Communications, Cam-
bridge University Press, Cambridge, U.K.
Larsson, E. & Jalden, J. (2008). Fixed-complexity soft MIMO detection via partial marginaliza-
tion, IEEE Transactions on Signal Processing, Vol. 56(No. 8): 3397–3407.
Liu, Y., Lau, K., Takeshita, O. & Fitz, M. (2002). Optimal rate allocation for superposition
coding in quasi-static fading channels, IEEE International Symposium on Information
Theory, ISIT 2002, pp. 111–.
LTE (2006). Requirements for Evolved UTRA (E-UTRA) and Evolved UTRAN (E-UTRAN), 3GPP
TR 25.913 v.7.3.0,.
LTE-A (2008). Requirements for Further Advancements for EUTRA, 3GPP TR 36.913,.
Medvedev, I., Bjerke, B., Walton, R., Ketchum, J., Wallace, M. & Howard, S. (2006). A compar-
ison of MIMO receiver structures for 802.11N WLAN - Performance and complex-
ity, IEEE 17th International Symposium on Personal Indoor and Mobile Radio Communica-
tions,, pp. 1–5.
Poor, H. & Verdu, S. (1997). Probability of error in MMSE multiuser detection, IEEE Transac-
tions on Information Theory Vol. 43(No. 3): 858–871.
Russell, M. & Stuber, G. (1995). Interchannel interference analysis of OFDM in a mobile envi-
ronment, IEEE 45th Vehicular Technology Conference, Vol. 2, pp. 820–824.
Shamai, S. & Steiner, A. (2003). A broadcast approach for a single-user slowly fading MIMO
channel, IEEE Transactions on Information Theory Vol. 49(No. 10): 2617–2635.
Telatar, I. E. (1999). Capacity of multi-antenna Gaussian channels, European Transactions on
Telecommunications Vol. 10(No. 6): 585–595.
Varanasi, M. K. & Guess, T. (1997). Optimum decision feedback multiuser equalization with
successive decoding achieves the total capacity of the Gaussian MAC channel, Asilo-
mar Conference on Signals, Systems and Computers.
Verdu, S. (1989). Computational complexity of multiuser detection, Algorithmica,, Vol. 4,
pp. 303–312.
Verdu, S. (1998). Multiuser Detection., Cambridge University Press, Cambridge, U.K.
Winters, J. (1984). Optimum combining in digital mobile radio with cochannel interference,
IEEE Journal on Selected Areas in Communications Vol. 2(No. 4): 528–539.
Wolniansky, P., Foschini, G., Golden, G. & Valenzuela, R. (1998). V-blast: an architecture for
realizing very high data rates over the rich-scattering wireless channel, International
Symposium on Signals, Systems, and Electronics, 1998. ISSSE 98., pp. 295–300.
Advanced Hybrid–ARQ Receivers for Broadband MIMO Communications 211
Advanced Hybrid–ARQ Receivers

for Broadband MIMO Communications
Tarik Ait-Idir1,2 , Houda Chafnaji1,2 , Samir Saoudi2
and Athanasios Vasilakos3
1Communications Systems Department, INPT, Madinat Al-Irfane, Rabat,
Morocco
2 Signal and Communications Department, Institut Telecom/Telecom Bretagne, Brest, and
Université Européenne de Bretagne (UEB),
France
3 University of Western Macedonia,
Greece
1. Introduction
Multiple input multiple output (MIMO) and hybrid–automatic repeat request (ARQ) mecha-
nisms play a key role in the evolution of current wireless communications systems towards
high data rate wireless packet access. In MIMO techniques, the spatial dimension of the
MIMO channel is exploited through the use of multiple antennas at both the transmitter and
receiver sides. This translates into an improvement in the spectrum efficiency and/or the link
quality Wolniansky et al. (1998). Hybrid–ARQ protocols provide an important source of time
diversity through the combination of channel coding and ARQ. This is performed with the aid
of packet combining techniques where erroneous data packets are kept in the receiver to help
detect/decode retransmitted frames.
In broadband MIMO communications, the MIMO wireless link suffers from intersymbol in-
terference (ISI) caused by multipath propagation. This effect can be mitigated using channel
equalization and/or Hybrid–ARQ. In this chapter, we focus on the joint design of the packet
combiner and the channel equalizer for MIMO ARQ transmission over the broadband wireless
channel. We start the chapter by reviewing the various approaches for joint packet combining
and equalization. We introduce the considered broadband MIMO ARQ transmission scheme.
Then, we derive the structure of the optimal maximum a posteriori (MAP) turbo packet com-
biner and study its outage performance. Finally, we introduce a new class of low-complexity
minimum mean square error (MMSE)-based turbo packet combiners and analyze their imple-
mentation requirements and block error rate (BLER) performance.
2. Advanced Receivers for MIMO ARQ

In the last few years, a special interest has been paid to the design of advanced MIMO ARQ re-
ceivers where packet combining and signal processing, i.e., detection/equalization, are jointly
performed. The concept of integrated equalization (IEQ) has been proposed in the framework
of MIMO systems with flat fading for joint multiple antenna interference (MAI) suppression
and packet combining (see for instance Onggosanusi et al. (2003) and Samra & Ding (2006)).
Turbo coded ARQ schemes with iterative MMSE frequency domain equalization (FDE) for
MIMO code division multiple access (CDMA) has been proposed in Garg & Adachi (2006).
Recently, we have introduced a new family of packet combining techniques for broadband
MIMO ARQ systems, where the decoding of a data packet is performed with the aid of an
iterative (turbo) processing between the soft combiner, i.e., joint packet combining and equal-
ization unit, and the soft input soft output (SISO) decoder (see Ait-Idir & Saoudi (2009)). The
following sections of this chapter focus on this new class of broadband MIMO ARQ tech-
niques. Our notation is introduced in the next section followed by the communication model
of the considered MIMO ARQ system.
3. Notation
In this chapter, scalars are denoted by small-case letters, vectors by small-case boldface letters,
and matrices by upper-case boldface letters. Superscripts and H denote the transpose and
transpose conjugate, respectively. I N is the N × N identity matrix, and 0 N × M is the N × M
zeros matrix. ⊗ is the Kronecker product, and Pr {.} denotes the probability of a given event.
4. Broadband MIMO ARQ Transceiver Scheme

4.1 MIMO ARQ Transmission
Let us consider a broadband multi-antenna, i.e., multiple input multiple output (MIMO),
system operating over a frequency selective fading channel and using an ARQ protocol at
the upper layer. The transmitter and the receiver are equipped with NT transmit (index
t = 1, · · · , NT ) and NR receive (index r = 1, · · · , NR ) antennas, respectively. The MIMO
channel suffers from intersymbol interference (ISI) and is composed of L symbol-spaced taps
(index l = 0, · · · , L − 1). Each data stream is first encoded with the aid of a ρ-rate chan-
nel encoder, interleaved using a semi-random interleaver Π, then modulated and space–time
multiplexed over the NT transmit antennas. This transmission scheme corresponds to the
so-called space–time bit interleaved coded modulation (STBICM). Let S denote the constel-
lation set, and M = log2 {S} its cardinality. A sequence of M coded and interleaved bits
b1,t,i , · · · , b M,t,i available on antenna t is transmitted at discrete-time instant i = 0, · · · , T − 1
over symbol st,i ∈ S according to the mapping function Ψ : {0, 1} M → S . The symbol vector
to be transmitted at time i is denoted

si s1,i , · · · , s NT ,i ∈ S NT . (1)
The rate of this transmission scheme is therefore R = ρMNT . The transmit symbol energy
is
normalized to one. Assuming infinitely deep space–time interleaving we get, E si siH = I NT .
At the receiver side, a positive/negative acknowledgment ACK/NACK message is sent back
to the transmitter upon the decoding of the information block. When the transmitter receives
a NACK message due to an erroneously decoded packet, subsequent transmission rounds
occur until the packet is correctly received or a preset maximum number of rounds K (index
k = 1, · · · , K) is reached. Parameter K is called the ARQ delay. Reception of a ACK message
indicates a successful decoding and the transmitter moves on to the next packet. We assume
an error-free feedback channel (the signaling channel carrying the ACK/NACK message),
and perfect packet error detection (using a cyclic redundancy check –CRC code). We focus on
Chase-type ARQ, i.e., the entire information block is retransmitted using the same STBICM
Noise
s 0, ..., sT−1
(1) (1)
ST Tx Transmission y0, ..., yT−1
encoder Π mapper buffer channel #1
ACK/NACK
(from receiver)
Noise
(k ) (k )
Transmission y0, ..., yT−1
channel # k
Fig. 1. STBICM ARQ Transmission over a short-term quasi-static MIMO channel
code. To prevent inter-block interference (IBI), we use either zero padding (ZP) or cyclic prefix
(CP)-aided transmission.
The MIMO-ISI channel is assumed to be short-term quasi-static block fading, i.e., constant
during one ARQ round and independently changes from round to round. The long-term
dynamic corresponds to the case when the channel is constant during all ARQ rounds cor-
responding to the transmission of the same information packet (see El Gamal et al. (2006)).
The short-term assumption is justified by the fact ARQ protocols are mainly used to improve
the link quality in the case of delay-tolerant applications where the processing delay is not a
(k)
major constraint. Let Hl denote the NR × NT complex matrix of the lth tap connecting the
(k)
transmitter and the receiver at ARQ round k. The elements of Hl are zero-mean circularly
(k) (k)
symmetric Gaussian random variables, i.e., hr,t,l ∼ CN 0, σl2 , where hr,t,l denotes the (r, t)th
(k)
element of matrix Hl and σl2 is the energy of tap l. The total channel energy is normalized
as ∑lL=−01 σl2 = 1. The discrete baseband signal received by antenna r at ARQ round k and time
instant i is given by
L−1 NT
(k) (k) (k)
yr,i = ∑ ∑ hr,t,l st,i−l + nr,i , (2)
l =0 t =1

(k) (k) (k)
where ni n1,i , · · · , n NR ,i ∼ CN 0 NR ×1 , σ2 I NR is the thermal noise at the NR receive
antennas. The STBICM ARQ transmission scheme over a short-term MIMO channel, i.e., k
distinct MIMO channels, is depicted in Fig. 1.
4.2 MIMO ARQ Turbo Receiver

In Ait-Idir & Saoudi (2009), turbo packet combining has been introduced as an efficient tech-
nique for combining multiple transmissions in the case of broadband MIMO ARQ systems.
In turbo packet combining, the decoding of a data packet is performed in an iterative (turbo)
fashion through the exchange of soft (extrinsic) information between the SISO packet com-
biner and the SISO decoder. The soft combiner exploits signals received at multiple ARQ
rounds to compute extrinsic log-likelihood ratios (LLR)s. Note that in conventional LLR-level
(1) (1)
y0, ...,yT−1
ARQ round #1 Π
Soft Packet
SISO LLRs info bits
Combiner Decoder
CRC
(k ) (k )
y0, ...,yT−1 Π −1
ARQ round # k ACK/NACK
(to transmitter)
Fig. 2. Block diagram of the turbo packet combing-aided MIMO ARQ receiver
combining techniques, the soft outputs obtained at different transmissions are simply added
together before channel decoding.
The general block diagram of the turbo packet combining-aided MIMO ARQ receiver is de-
picted in Fig. 2. Let

a a a
φt,i φ1,t,i , · · · , φ M,t,i (3)
denote the M × 1 real vector of a priori LLRs corresponding to coded and interleaved bits
b1,t,i , · · · , b M,t,i , and available at the input of the soft combiner at a certain iteration of ARQ
round k. Using a priori information φ1,0 a , · · · , φa
NT ,T −1 and signals received at rounds 1, · · · , k,
the soft packet combiner computes extrinsic LLR vectors

φet,i φ1,t,i
e
, · · · , φeM,t,i , (4)
which are de-interleaved and sent to the SISO decoder to obtain a posteriori LLRs about useful
bits and extrinsic information about coded bits. The generated extrinsic LLR values are then
interleaved and fed back to the soft combiner to help compute soft information during the next
turbo iteration of the same ARQ round. After a preset number of iterations, the decision about
the data packet is performed, and the ACK/NACK message is sent back to the transmitter
accordingly. Note that during the first iteration, a priori information corresponds to the soft
information available from the last iteration of previous ARQ round k − 1.
5. Information-Theoretic Issues
In this section, we derive the optimal maximum a posteriori (MAP) turbo packet combining
receiver for broadband MIMO ARQ transmission, and investigate its outage performance.
We first show that optimal turbo packet combining can be formulated as a MIMO-ISI turbo
equalization problem. We then obtain the outage probability of the broadband MIMO ARQ
channel, and analyze the impact of the ARQ delay, i.e., maximum number of ARQ rounds K,
on the outage performance.
5.1 Optimal Turbo Packet Combining

To derive the optimal maximum a posteriori (MAP) turbo packet combiner at ARQ round k, let
us consider the following signal vector that groups signals corresponding to all ARQ rounds
1, · · · , k,

(1) (k ) (1) (k )
y ( k ) y T −1 , · · · , y T −1 , · · · , y0 , · · · , y0 ∈ C kNR T , (5)
where
(u) (u) (u)
yi y1,i , · · · , y NR ,i ∈ C NR (6)
is the vector of received signals at ARQ round u = 1, · · · , k and time instant i.
The formulation in (5) is of a great importance because it allows us to view each ARQ round
as a set of virtual NR receive antennas. Note that in this section we assume that all signals and
channel matrices corresponding to previous ARQ rounds are available at the receiver side at
round k. In Section 6, we will present an optimized turbo packet combining technique that
makes use of all signals and channel matrices without being required to be explicitly stored
in the receiver. With respect to (2) and (5) and assuming a ZP-aided transmission strategy,
the signal vector y(k) corresponding to the transmission of the entire symbol frame over k
MIMO-ISI channels can be expressed as,
y(k) = H(k) s + n(k) , (7)

where H(k) is a block Toeplitz matrix given as
 
(1) (1)
H0 H L −1
 
 .. ··· .. 
 . . 
 
 (k) (k) 
 H0 H L −1 
 
k)  . .. 
H 
( .. .  , (8)
 
 (1) (1) 

 H0 H L −1 


 .. ··· .. 

 . . 
(k) (k)
H0 H L −1
kNR T × NT T
and vectors s and n(k) are defined as,

s s
T −1 , · · · , s0 ∈ S NT T , (9)

(1) (k ) (1) (k )
n ( k ) n T −1 , · · · , n T −1 , · · · , n0 , · · · , n0 ∈ C kNR T . (10)
The communication model (7) corresponds to a MIMO-ISI equalization problem with NT

transmit and kNR virtual receive antennas. It allows for jointly (over all ARQ rounds) can-
celing both multiple antenna interference (MAI) and ISI, while exploiting all the diversities
available in the MIMO-ISI ARQ channel. Using the MAP criterion and given a priori LLRs, the
extrinsic information about coded and interleaved bit bm,t,i can be expressed as,

(1) (k)
Pr y(k) | bm,t,i = 1 ; H0 , · · · , H L−1 , a priori LLRs
e
φm,t,i = log (1) (k)
. (11)
Pr y(k) | bm,t,i = 0 ; H0 , · · · , H L−1 , a priori LLRs
e
Then, by invoking the multi-round communication model (7), φm,t,i can be obtained as,
 
 2 
(k) −1 a
∑ exp − 2σ1 2 y − H(k) s + ∑ Ψm st ,i φm ,t ,i
1
 
s∈Sm,t,i (m ,t ,i )=(m,t,i)
e
φm,t,i = log  , (12)
 2 
(k) −1 a
∑ exp − 2σ1 2 y − H(k) s + ∑ Ψm st ,i φm ,t ,i
0  
s∈Sm,t,i (m ,t ,i )=(m,t,i)
b b

−1 s

where the subset Sm,t,i is defined as Sm,t,i s ∈ S NT T | Ψm t,i = b , b = 0, 1.
5.2 Outage Performance Analysis

Outage probability is a useful tool that allows to analyze the performance of non-ergodic
channels, i.e., quasi-static channels. It provides a lower bound on the BLER, and is generally
defined as the probability that the mutual information, as a function of the channel realization
and the average signal to noise ratio (SNR) γ per receive antenna, is below the transmission
rate R Tse & Viswanath (2005).
The outage probability of an ARQ protocol can be derived using the renewal theory (see Wolff
(1989)) as in Caire & Tuninetti (2001) and El Gamal et al. (2006). In the case of a broadband
MIMO system with Chase-type ARQ, perfect packet error detection, and error-free ACK/
NACK feedback, the outage probability can be derived using the multi-round communication
model (7) and the renewal theory (see Ait-Idir & Saoudi (2009)) as

R 1 (K ) ( K)

Pout (γ) = Pr I s; y | H , γ < R, Ā1 , · · · , ĀK −1 , (13)
K
where K is the ARQ delay, and Ak denotes the event that an ACK message is fed back at
round
k, i.e., the data
packet is positively acknowledged at ARQ round k. The quantity
1
KI s; y(K ) | H(K ) , γ denotes the mutual information rate at the last ARQ round K, and is
expressed in the case of a MIMO broadband channel with Gaussian inputs as

1 (K ) 1 T −1 γ (K ) (K ) H
I s; y | H(K ) , γ = ∑ log2 det IKNR + Λi Λi , (14)
K KT i=0 NT
(K )
where Λi is the discrete Fourier transform (DFT) of the KNR × NT virtual MIMO-ISI channel
at round K and frequency bin i, i.e.,
 (1) 
Hl
L −1  
Λi
(K )
∑  ..  exp − j 2π il . (15)
 .  T
l =0 (K )
Hl
Note that the factor K1 appearing in the outage probability (13) is due to the fact that a transmis-
sion scheme with Chase-type ARQ, and an ARQ delay K is equivalent to a repetition coding
scheme where K parallel sub-channels are used to transmit the same symbol frame.
In the following, we investigate the impact of the ARQ delay K on the outage performance.
We consider a MIMO-ISI channel with L = 2 taps having equal powers, i.e., σ02 = σ12 = 12 .

(k) (k)
At each ARQ round k, the mutual information rate 1k I s; y(k) | H0 , H1 , γ of the MIMO
ARQ system after k rounds is evaluated similarly to (14). If the target rate R is not reached
and k < K, the system moves on to the next round k + 1. An ARQ process is stopped and
an another is started, either because of system outage, i.e., the mutual rate after K rounds is
below R, or the rate R is achieved at a certain round k ≤ K. In all scenarios, we take T = 256
discrete time instants.
N =N =2, L=2, R=2

T R
0
10
−1
10
Outage Probability
−2
10
−3
10
ARQ with 1 round

ARQ with 2 rounds
−4
ARQ with 3 rounds
10
1 2 3 4 5 6
SNR (dB)
Fig. 3. Outage probability performance for two transmit and two receive antennas
In Fig. 3, we plot the outage probability performance for the NT = NR = 2 antenna configura-
tion, with a target rate R = 2. We observe that the ARQ diversity gain, due to the short-term
static channel dynamic, clearly appears when the ARQ delay is set to K = 2. It offers a signif-
icant SNR gain compared with the non-ARQ case (K = 1). When, the ARQ delay is increased
to K = 3, the outage performance is similar to that of K = 2. This means that if the system is in
outage in the second ARQ round, then it will almost be in outage in the third round. In Fig. 4,
we investigate the outage performance when the number of transmit antennas is increased to
NT = 4, while NR = 2. The target rate is set to R = 4. As in the previous configuration, both
ARQ delays K = 2 and K = 3 provide almost the same outage performance, while the overall
diversity gain is more important than that corresponding to NT = NR = 2. This can be seen
from the steeper slopes of outage curves. Note that the ARQ diversity due to multiple trans-
missions does not completely translate into a receive diversity gain (related to the NR virtual
receive antennas at each ARQ round). This is due to the fact that the target rate R has to be
maintained, as it can be seen from the expression of the outage probability (13). This means
that the diversity gain does not linearly increase with increase of the ARQ delay K. This issue
has been addressed by El Gamal et al. (2006) in the case of flat fading MIMO ARQ systems.
N =4, N =2, L=2, R=4

T R
0
10
−1
10
Outage Probability
−2
10
−3
10
ARQ with 1 round

ARQ with 2 rounds
−4
ARQ with 3 rounds
10
5 5.5 6 6.5 7 7.5 8 8.5 9 9.5
SNR (dB)
Fig. 4. Outage probability performance for four transmit and two receive antennas
6. MMSE-Based Turbo Packet Combining

The optimal MAP turbo packet combining algorithm we have presented so far in Subsection
5.1 has a computational complexity that exponentially increases with the increase in the num-
ber of ARQ rounds. In addition, all signals and channel matrices have to be stored in the
receiver to perform combining at each ARQ round. In this section, we present alternative
MMSE-based sub-optimal techniques (see Ait-Idir & Saoudi (2009)) that allow for performing
turbo packet combining with reduced computational complexity and memory requirements.
We first introduce the so-called signal-level turbo combining technique where packet combining
is performed at the signal level similarly to MAP combining but based on MMSE processing.
Then, we present the symbol-level turbo combining algorithm where packets are combined at the
input of the soft demapper after performing MMSE channel equalization separately for each
transmission. We also provide a brief presentation of the conventional LLR-level combining
technique. For all combining schemes, we assume a ZP-aided block transmission. The multi-
round block communication model is therefore described by a block Toeplitz channel matrix
as in (7).
6.1 Signal-Level Turbo Combining

The signal-level turbo packet technique is a low-complexity MMSE-based combining algo-
rithm that performs packet combining jointly with MAI and ISI cancellation using a block
length equal to ε = ε 1 + ε 2 + 1 T, where ε 1 and ε 2 are the lengths of the forward and
backward filters, respectively. Packet combining at ARQ round k is therefore performed with
respect to the following ε-length block communication model

(k)
y(k ) = H (k ) si + ni , (16)
i
where
(1) (k ) (1) (k )
y(k ) yi +ε 1 , · · · , yi +ε 1 , · · · , yi −ε 2 , · · · , yi −ε 2 ∈ C kNR ε , (17)
i

(k) (1) (k ) (1) (k )
ni ni +ε 1 , · · · , ni +ε 1 , · · · , ni −ε 2 , · · · , ni −ε 2 ∈ C kNR ε , (18)

si si+ε 1 , · · · , si−ε 2 − L+1 ∈ S NT (ε+ L−1) , (19)
and H(k) is a block Toeplitz matrix which is defined similarly to (8) but using ε block rows and
(ε + L − 1) block columns.
First of all, the soft packet combiner computes conditional means and variances of transmitted
symbols using a priori information vectors φt,i a , where t ∈ {1, · · · , N }, and i ∈ {0, · · · ,
T
T − 1}, available from the previous iteration. Note that, as it has been mentioned before, the
soft information generated by the SISO decoder during the last iteration of ARQ round k − 1
serves as a priori information at the first iteration of round k. Conditional symbol expectations,
also called soft symbols, serve for regenerating soft interference caused by multiple antenna
transmission and multipath propagation, i.e., MAI plus ISI. Conditional symbol variances and
channel matrices corresponding to ARQ rounds 1, · · · , k are used for computing the multi-
round MMSE filter that serves for performing signal combining. At each iteration of ARQ
round k, the signal-level turbo packet combiner exploits the communication model (16) to
cancel soft MAI plus ISI (estimated at the current iteration of round k) from all signals received
at rounds 1, · · · , k. Then, it combines the resulting soft interference-free signals to generate
soft decisions about transmitted
symbols.

a 2
2
Let s̃t,i E st,i | φ
t,i and σ̃ E st,i − s̃t,i | φ a denote the conditional mean and vari-
t,i t,i
ance of symbol st,i . By invoking the special block Toeplitz structure of matrix H(k) , and the
structure of the multi-round signal vector (17), the signal-level turbo packet combiner com-
(k)
putes, at a particular iteration of ARQ round k, soft decision ξ t,i about symbol st,i using the
following forward–backward filtering structure,
(k) (k) (k) (k)
ξ t,i = Ft zi |Bt s̃i|t , (20)
where s̃i|t is the NT (ε + L − 1)-length soft symbol vector corresponding to (19) with zero at
(k) (k)
the (ε 1 NT + t)th position corresponding to soft symbol s̃t,i . Ft and Bt are the forward and
(k)
backward filters related to antenna t. zi is a vector that contains properly weighed and
combined copies of signals received at rounds 1, · · · , k. It does not change in the course of
turbo iterations corresponding to the same ARQ round, and is produced at round k according
to the following recursion
(k) ( k −1) H (k)
zi = zi + H (k ) yi ,
(0) (21)
zi = 0 NT (ε+ L−1)×1 .
(k)
The signal vector yi and the channel matrix H(k) correspond to the ε-length block commu-
nication model of ARQ round k, and are given as,
 (k) (k) 
H0 · · · H L −1
 .. .. 
H(k) 
 . .

 , (22)
(k) (k)
H0 ··· H L −1 NR ε× NT (ε+ L−1)

(k) (k ) (k)
yi yi +ε 1 , · · · , yi −ε 2 ∈ C NR ε . (23)
(k) (k)
Filters Ft and Bt are iteration-dependent, and are computed as
−1
(k)
Ft = σ2 + 1 − σ̃t2 e (k) (k)
t Λ Υ et e (k)
t Λ , (24)
(k) (k)
= Ft Υ(k) . Bt (25)
where et is a NT (ε + L − 1)-length vector of zeros with one at the (ε 1 NT + t)th position, i.e.,
et [ 0, · · · , 0 , 1, 0, · · · , 0 ] . (26)

ε 1 NT +t−1 (ε 2 + L) NT −t
σ̃t2 is the unconditional variance of symbols transmitted over antenna t, and is computed as
2 , · · · , σ̃2 2 1 T −1 2
the time average of conditional variances σ̃t,0 t,T −1 , i.e., σ̃t = T ∑i =0 σ̃t,i . The square
matrix Λ(k) is updated at each turbo iteration according to,
−1
Λ(k) = I NT (ε+ L−1) − Υ(k) Υ(k) + σ2 Ξ−1 , (27)
where  
σ̃12
 ..  N ε+ L−1)× NT (ε+ L−1)
Ξ = I ε + L −1 ⊗  .  ∈ C T( , (28)
2
σ̃N T
and Υ(k) is recursively computed as,

H
Υ ( k ) = Υ ( k −1) + H ( k ) H ( k ) ,
(29)
Υ(0) = 0 NT (ε+ L−1)× NT (ε+ L−1) .
Details regarding the derivation of the forward–backward filtering structure (20) can be
found in Ait-Idir & Saoudi (2009)
e
Extrinsic information φm,t,i about coded and interleaved bit bm,t,i , m = 1, · · · , M, can be ob-
(k)
tained using the decision statistic ξ t,i of (20) as,

1 (k) ( k ) 2 −1 a
∑ exp − ( k )2
ξ t,i − α t s + ∑ Ψ
m =m m
( s ) φm ,t,i

s∈Sm1 2δt
e
φm,t,i = log , (30)
1 (k) ( k ) 2 −1 a
∑ exp − ( k )2
ξ t,i − αt s + ∑m =m Ψm ( s ) φm ,t,i
s∈Sm0 2δt
(k) (k)
where it is assumed that conditional soft demapper input ξ t,i | st,i is Gaussian with mean αt
2
(k)
and variance δt , and are given by,

α(k) (k)
= Bt et
t
2 (31)
δ(k) = 1 − αt
(k) (k)
αt ,
t

b is defined as S b s ∈ S | Ψ−1 ( s ) = b for b = 0, 1.
and the set Sm m m
6.2 Symbol-Level Turbo Combining

In symbol-level turbo packet combining, MMSE turbo equalization is separately performed
for each ARQ round k based on the ε-length block communication model of round k,
(k) (k)
yi = H (k ) si + ni , (32)
(k)
where si , H(k) , and yi are given by (19), (22), and (33), respectively, and

(k) (k ) (k )
ni ni +ε 1 , · · · , ni −ε 2 ∈ C NR ε (33)
is the spatially and temporally white Gaussian noise at the input of the equalizer at ARQ
(k)
round k, i.e., ni ∼ CN 0 NR ε×1 , σ2 I NR ε . Soft combining is iteratively performed at the level
of unconditional MMSE filter outputs by combining the output at each iteration of ARQ round
k with those obtained at the last iteration of previous rounds 1, · · · , k − 1.
(k) (k)
The forward and backward filters Ft and Bt in the case of symbol-level turbo packet com-
bining can easily be derived using the equations provided in the previous subsection with
(k)
k = 1. Now, let ξ̆ t,i denote the MMSE filter output corresponding to symbol st,i at a specific
(k)
turbo iteration of round k. The conditional decision statistic ξ̆ t,i | st,i is Gaussian, with mean
2
(k) (k) (u)
ᾰt and variance δ̆t given similarly to (31). For ARQ rounds u = 1, · · · , k − 1, ξ̆ t,i denotes
the decision statistic obtained at the last iteration of round u. Therefore, the soft combiner
e
provides extrinsic information φm,t,i about coded and interleaved bit bm,t,i as
(k) (k)
( k ) H ( k ) −1 (k) −1
∑ exp − 12 ξ̆ t,i − sᾰt ∆t ξ̆ t,i − sᾰt + ∑ m =m Ψ m ( s ) φ a
m ,t,i
e s∈Sm1
φm,t,i = log (k) (k) ,
( k ) H ( k ) −1 (k) −1
∑ exp − 12 ξ̆ t,i − sᾰt ∆t ξ̆ t,i − sᾰt + ∑ m =m Ψ m ( s ) φ a
m ,t,i
s∈Sm0
(34)
where vector
(k) (1) (k)
ξ̆ t,i ξ̆ t,i , · · · , ξ̆ t,i ∈ Ck (35)
(k) (k)
gathers MMSE filter soft outputs corresponding to all rounds 1, · · · , k. ᾰt and ∆t are the
(k)
mean and covariance of the conditional Gaussian vector ξ̆ t,i | st,i , and are given by,

(k) (1) (k)
ᾰt ᾰt , · · · , ᾰt ∈ Ck , (36)
 
(1)2
 δ̆t 
(k)  .. 
∆t =  .  . (37)
 
( k )2
δ̆t k×k
6.3 LLR-Level Turbo Combining

In conventional LLR-level combining, LLR values of transmitted bits obtained at multiple
ARQ rounds are stored in the receiver and simply added together to update the LLRs at each
ARQ round. In our framework, LLR-level turbo combining is carried out by separately per-
forming MMSE turbo equalization for multiple transmissions using the the ε-length block
communication model (32). Extrinsic LLRs φm,t,i e corresponding to coded and interleaved bits
bm,t,i ∀ m, t, i are then computed at each iteration of ARQ round k using only decision statistics
(k)
ξ̆ t,i introduced in the previous subsection. LLR values obtained at multiple ARQ rounds are
then added together to produce the soft LLR outputs
k
(k) e (u)
LLRm,t,i = ∑ φm,t,i , ∀ m, t, i, (38)
u =1
e (u)
which are de-interleaved and fed back to the SISO decoder. Note that in (38), φm,t,i denotes
the extrinsic LLR at the last iteration of ARQ round u = 1, · · · , k.
6.4 Implementation Issues

(k)
In signal-level turbo combining, the computation of the forward and backward filters Ft and
(k)
Bt involves, at each turbo iteration of ARQ round k, one inversion of the NT (ε + L − 1) ×
(k) (k)
NT (ε + L− 1) matrix
Υ(k) + σ2 Ξ−1 . The cost of computing Ft and Bt is therefore in the or-
3 3
der of O NT ε complex operations. This indicates that the computational complexity of the
signal-level combining scheme is less sensitive to k. The number of ARQ rounds only influ-
ences the number of additions required for performing recursions (21) and (29), which is in the
order of NT2 (ε + L − 1)2 + NR εT complex additions. Note that the operations of computing
H H (k)
H(k) H(k) and H(k) yi are also required in the case of symbol-level combining. Therefore,
the computational cost of forward and backward filters is almost the same for both signal and
symbol-level turbo combining schemes. The LLR-level turbo combining algorithm approxi-
mately involves the same amount of operations as symbol-level combining.
In the case of signal-level combining, memory requirements are determined by (21) and (29),
where two NT (ε + L − 1) × NT (ε + L − 1) and NT (ε + L − 1) × T complex matrices are re-
H (k) (k)
quired to accumulate channel matrices H(k) H(k) , and signal vectors z0 , · · · , z T −1 , respec-
tively. Note that the two recursions (21) and (29) play a key role in signal-level turbo combin-
ing since they avoid the storage of all signals and channel matrices as in MAP turbo combin-
ing. In symbol-level combining, only NT complex matrices of size K × T and two K × NT com-
plex matrices are required to store filter outputs and their corresponding parameters given by
(35), (36), and (37). Therefore, signal-level combining requires slightly more memory than its
symbol-level counterpart. Finally, in LLR-level turbo combining, a real vector of size NT MT
is required to combine extrinsic values. Therefore, the three combining strategies have similar
implementation requirements. They slightly differ in the number of additions and storage
memory.
NT=NR=2, CC(1338,1718), QPSK, L=2, K=2

0
10
−1
10
BLER
−2
10
−3 round 1
10 MFB, round 1
LLR level, round 2
Symbol Level, round 2
Signal Level, round 2
MFB, round 2
−4
Outage, 2x2, R=2, K=2
10
−5 −4 −3 −2 −1 0 1 2 3 4 5
SNR (dB)
Fig. 5. BLER performance comparison for NT = NR = 2 and QPSK.
6.5 BLER Performance

In this subsection, we provide simulated BLER performance for the packet combining strate-
gies studied in the previous subsections. The main focus of the analysis we provide is to show
that signal-level turbo combining has better ISI cancellation capability and diversity gain com-
pared with the other combining schemes.
We consider an STBICM transmitter with a 64-state 12 -rate convolutional code whose polyno-
mial generators are (1338 , 1718 ). The length of the code frame is 1800 bits. The modulation
scheme is quadrature phase shift keying (QPSK). The MIMO-ISI channel has the same profile
as in Subsection 5.2, i.e., two equal power taps. The ARQ delay is chosen K = 2 according
to the theoretic analysis in Subsection 5.2. In all figures, the BLER is per ARQ round, and the
SNR is per symbol per receive antenna.
We compare the resulting BLER performance with the outage probability and the matched fil-
ter bound (MFB). Note that for the purpose of fair comparison, the computation of the outage
probability does not take into account the rate distortion as in (14). The MFB curves are ob-
tained for each transmission assuming perfect ISI cancellation and maximum ratio combining
(MRC) of all time, space, multipath, and delay diversity branches.
In Fig. 5, we consider an STBICM code with NT = NR = 2 transmit and receive antennas.
This corresponds to a rate R = 2. The filter length is chosen equal to ε = 9 (ε 1 = ε 2 = 4)
for all combining schemes. A quick inspection of Fig. 5, shows that both signal and symbol-
level turbo combining offer a significant performance improvement after the second ARQ
round compared with LLR-level combining. The signal-level scheme has better ISI cancel-
lation capability compared with symbol-level combining. It almost achieves the MFB, while
the symbol-level scheme presents a gap of approximately 1dB compared with the MFB. Also,
note that both signal and symbol-level combining achieve the asymptotic slope of the outage
probability.
In Fig. 6, we evaluate the BLER performance of a ST-BICM code with NT = 4. This cor-
responds to a rate R = 4. The number of receive antennas is NR = 2. Note that this
N =4, N =2, CC(133 ,171 ), QPSK, L=2, K=2

T R 8 8
0
10
−1
10
BLER
−2
10
round 1
MFB, round 1
LLR Level, round 2
Symbol Level, round 2
Signal Level, round 2
MFB, round 2
−3
Outage, 4x2, R=4, K=2
10
−2 0 2 4 6 8 10
SNR (dB)
Fig. 6. BLER performance comparison for NT = 4, NR = 2 and QPSK.
type of unbalanced MIMO configurations where the transmitter is equipped with more an-
tennas than the receiver is suitable for the forward link. The filter length is increased to
ε = 13 (ε 1 = ε 2 = 6) for all combining schemes. The signal-level combining technique is
shown to achieve BLER performance close to the MFB with a gap less than 0.5dB, while both
the LLR-level and the symbol-level techniques have degraded BLER performance. The signal-
level combining manifests itself in almost achieving the diversity gain of the MIMO-ISI ARQ
channel, while it is shown that symbol-level combining fails to do so. This is because in the
second ARQ round, the signal-level scheme constructs a 4 × 4 virtual MIMO-ISI channel for
ISI cancellation and symbol detection, while the MIMO configuration remains unbalanced in
the case of symbol and LLR-level combining.
7. Conclusion
In this chapter, we considered the design of efficient iterative turbo packet combining algo-
rithms for broadband MIMO systems with Chase-type ARQ. We derived the structure of the
optimal MAP turbo packet combining technique that exploits all the diversities available in
the MIMO-ISI ARQ channel to perform packet combining, and analyzed its outage probabil-
ity. As optimal MAP turbo packet combining has a huge computational cost and memory
requirements, we introduced a new class of low-complexity MMSE-based turbo packet com-
bining schemes. In MMSE-based signal-level combining, each ARQ round is viewed as a set
of virtual receive antennas, and packet combining is jointly performed with ISI cancellation
at the signal level. In MMSE-based symbol-level combining, multiple transmissions are sep-
arately turbo equalized, and combining is performed at the level of filter outputs. The sim-
ulation results provided in this chapter indicate that signal-level combining provides better
BLER performance than that of symbol-level and conventional LLR-level combining.
8. References
Ait-Idir, T. & Saoudi, S. (2009) Turbo packet combining strategies for the MIMO-ISI ARQ chan-
nel, IEEE Transactions on Communications vol. 57, no. 12, December 2009, 3782-3793
Caire, G. & Tuninetti, D. (2001). ARQ protocols for the Gaussian collision channel, IEEE Trans-
actions on Information Theory, Vol.47, No.4, July -2001, 1971–1988
El Gamal, H.; Caire, G. & Damen, M. O. (2006) The MIMO ARQ channel diversity-
multiplexing-delay tradeoff, IEEE Transactions on Information Theory, Vol.52, No.8,
August -2006, 3601–3621
Garg, D. & Adachi, F. (2006) Packet access using DS-CDMA with frequency-domain equaliza-
tion IEEE Journal of Selected Areas in Communications, vol. 24, no. 1, Jan. 2006
Onggosanusi, E. N.; Dabak, A. G.; Hui, Y. & Jeong, G. (2003) Hybrid ARQ transmission and
combining for MIMO systems, Proceedings of IEEE International Conference on Commu-
nications, pp. 3205-3209, ISBN 0-7803-7802-4, Anchorage, May 2003
Samra, H. & Ding, Z. (2006) New MIMO ARQ protocols and joint detection via sphere decod-
ing, IEEE Transactions on Signal Processing, vol. 54, no. 2, pp. 473-482, Feb. 2006
Tse, D. & Viswanath, P. (2005). Fundamentals of Wireless Communication, Cambridge University
Press, ISBN 978-0-521-84527-4
Wolff, R. (1989). Stochastic Modeling and the Theory of Queues, Upper Saddle River, NJ: Prentice-
Hall, 1989
Wolniansky, P. W.; Foschini, G. J. & Valenzuela, R. A. (1998). V-BLAST: an architecture for
realizing very high data rates over the rich scattering wireless channel, in Proc. Int.
Symp. Signals, Systems, Electron., Pisa, Italy, Sep. 1998
Cooperative ARQ: A Medium Access Control (MAC) Layer Perspective 227
Cooperative ARQ: A Medium Access Control

(MAC) Layer Perspective
Jesús Alonso-Zárate*, Elli Kartsakli**, Luis Alonso**
and Christos Verikoukis*
*CentreTecnològicdeTelecomunicacionsdeCatalunya(C
Universitat
* PolitècnicadeCatalunyaUPC-
( EPSC)
1. Introduction
Cooperative Automatic Retransmission reQuest (C-ARQ) schemes have become a very
active research topic over the last years. C-ARQ schemes constitute a practical way of
executing cooperation in wireless networks with already existing equipment. C-ARQ
schemes exploit feedback from the receiver, i.e. cooperation is only executed when needed,
and thus are sometimes referred to as cooperation on-demand cooperative schemes.
In short, the idea of C-ARQ is to exploit the fact that, due to the broadcast nature of the
wireless channel, any transmission can be received by any of the stations in the transmission
range of the transmitter. What has been traditionally considered as interference, is exploited
in C-ARQ schemes to attain spatial diversity. Upon a transmission error, a retransmission
can be requested from any (or some) of the stations which overheard the original
transmission, which can act as spontaneous helpers (or relays). The result is that the
destination of a packet can receive different copies of the same information arriving via
statistically independent transmission paths, i.e., space diversity.
C-ARQ schemes have been already studied in the literature from a theoretical point of view
and there is no doubt that, under some conditions, they can dramatically boost the
performance of wireless communications compared to traditional ARQ, where
retransmissions are performed only from the source. However, involving a number of users
in a communication link requires coordination. To this end, efficient Medium Access
Control (MAC) protocols are necessary to get the maximum efficiency of the
communications. In this chapter we emphasize the important role of the MAC layer in this
context of C-ARQ.
Along the chapter, we first review in Section 2 the motivation and operation of C-ARQ
schemes into detail. We go through the parameters that affect the performance of these
schemes and we point out the role of the MAC layer. Taking into account the specific
requirements of the MAC layer in this kind of schemes, we present in Section 3 a novel high-
performance MAC protocol specifically tailored for this purpose. Computer-based
simulations are presented to evaluate the performance of the protocol. Finally, Section 4
concludes the chapter.
2. Cooperative ARQ (C-ARQ)

2.1 Background and Motivation
Traditionally, ARQ schemes have been used in communication networks to guarantee the
reliable delivery of data packets. Upon the reception of a packet with errors, retransmissions
are requested from the source (and along the same channel) until either the packet can be
properly decoded or it is discarded for the benefit of the backlogged data.
Several variations of ARQ schemes have been proposed in the past to improve the
performance of communications. These schemes perform well in wired networks where
there is no correlation between consecutive packet error probabilities, i.e., packet errors are
random and sparse. However, their performance in wireless networks is compromised by
phenomena such as the shadowing and fading of the radio channel. In wireless channels,
packet errors might come into bursts, and thus if a packet is received with errors, the
immediate retransmissions will be also received with errors with high probability if they are
performed through the same channel (Zorzi et al., 1997).
C-ARQ schemes constitute a practical solution to combat this fading nature of the wireless
channel. Their operation is described in the following section.
2.2 Description of C-ARQ

Consider a wireless network formed by an arbitrary number of stations equipped with half-
duplex radio frequency transceivers. In order to be able to execute a C-ARQ scheme, all the
stations must listen to (overhear) every ongoing transmission in order to be able to
cooperate if required. In addition, they should keep a copy of any received data packet
(regardless of its destination address) until it is acknowledged (positively or negatively) by
the destination. This packet is discarded whenever the destination successfully decodes the
original packet.
It is assumed that, although both error detection and Forward Error Correction (FEC) bits
are attached to all the transmitted data packets, errors can still occur due to the severe
wireless channel impairments. Whenever a destination receives a data packet with
unrecoverable errors, it broadcasts a retransmission request in the form of a control packet.
This packet is referred to as the Call for Cooperation (CFC) packet. A cooperation phase is
then initiated.
A subset of the stations which overheard both the original transmission from the source and
the CFC from the destination, become active relays or helpers. As it will be further
discussed later, some relay selection criteria can be attached to the CFC in order to activate
the most appropriate subset of stations to act as helpers. Orthogonally in time (TDMA),
frequency (FDMA or OFDMA), or code (CDMA), these active relays attempt to retransmit a
copy of the original packet to assist in the failed transmission. For the sake of clarity in the
explanation and without loss of generality, the data packets retransmitted by the relays will
be referred to as cooperative packets.
Eventually, the destination might either receive a correct copy of the original packet from a
relay or may be able to properly combine the different retransmissions from the relays to
successfully decode the original packet. Otherwise, if the destination is not able to recover
the data packet after some predefined time (cooperation time-out), it discards it. In any of
the two cases, the cooperation phase is finished.
Although slight different variations to this general operation can be found in the literature,
most of the proposed C-ARQ schemes follow this description. It is worth mentioning that
the CFC has sometimes received the name of Negative ACK (NACK) in the literature
(Dianati et al. 2006). However, this name falls short in describing the real function of the
CFC. Besides informing the Negative ACK, it also calls for cooperation and, indeed, it could
attach some relay selection criteria, among other control information required for the
execution of a cooperative technique.
An example of operation of a C-ARQ mechanism is illustrated in Fig. 1. Therein, the
communication between a source and a destination stations is assisted by an arbitrary
number (N ) of relays. In this particular example, the relays retransmit data orthogonally in
time until the destination station can send the ACK.
The performance of a C-ARQ scheme might be mainly influenced by the following four
parameters:
1) The relay selection criteria; as it could be expected, the number of potential helpers and
the “quality” of those helpers will have a direct impact on the efficiency of the C-ARQ
scheme. For this reason, there are several works focused on the design of efficient
techniques to select either the best or a subset of the best potential helpers to act as relays
(Gómez et al., 2007; Biswas & Morris, 2005).
2) The PHY forwarding technique executed by the relays (Nosratinia et al., 2004):
a. Amplify and forward techniques , when the relays transmit an amplified version of the
original received signal, without demodulating or decoding it.
b. Compress and forward techniques , when the relays transmit a compressed version of the
original transmitted signal, without decoding it.
c. Decode and forward techniques, when the relays transmit coded copies
re of the original
message. Note that using decode and forward, the recoding process can be done on the
basis of repeating the original codification, recoding the original data (or only a relevant part
of it), or using more sophisticated Space-Time Codes (STC) (Fitzek & Katz, 2006).
3) The number of required retransmissions necessary to decode a packet which can mainly
depend on:
Source DATA0
Destination CFC ACK
Relay 1 DATA1
Relay 2 DATA2
...
Relay N DATAN
Time+
Fig. 1. C-ARQ Scheme with Time-Orthogonal Relays
a. The channel conditions between the source and the destination, the source and the relays,
and the relays and the destination (Gómez & Pérez-Neira, 2006; Pfletschinger & Navarro,
2008).
b. The transmission scheme, which includes the forwarding technique executed by the
relays and the combination technique executed by the destination station to combine the
different retransmissions received from independent paths. The approach of combining
different erroneous copies of a same packet to decode the original packet has been tackled in
the past (Charaborty et al., 2005; Morillo-Pozo & García-Vidal, 2007).
4) The MAC protocol which is necessary to tackle with the contention among the relays. Just
as an example, the ideal scheduling among the relays represented in Fig. 1 is impossible to
attain in fully distributed networks without a central coordinator. Therefore, the set of active
relays should contend for the channel in order to retransmit the packets. Efficient MAC
protocols are necessary to execute a C-ARQ scheme in order to exploit the benefits of
cooperation in wireless networks.
2.3 Motivation and Contributions of the Chapter

C-ARQ schemes have been so far analyzed from a fundamental point of view and mainly
with emphasis on the PHY layer (Dianati et al. 2006; Zimmermann et al., 2004; Zimmermann
et al., 2005; Gupta et al., 2004; Cerruti et al., 2008; Morillo-Pozo et al., 2005). These previous
works put in evidence that C-ARQ schemes can yield an improvement in performance,
lower energy consumption, and interference, as well as an extended coverage area by
allowing communication at low Signal to Noise Ratios (SNRs). However, all of these
contributions assume simplified topologies (with one or very few relays) and perfect
scheduling among the relays at the MAC level. This scheduling might be difficult to attain in
the fully decentralized scenario represented by the cloud of relays without infrastructure.
Therefore, both the design of efficient MAC protocols and the evaluation of the actual
performance of C-ARQ techniques considering the MAC overhead are mandatory if C-ARQ
schemes are to find real application. Indeed, this is the main motivation for this chapter.
The focus in this chapter is on time-orthogonal C-ARQ schemes, which might be the easiest
approach to implement with already existing off-the-shelf equipment. By slightly modifying
the wireless controller (or driver), existing wireless cards could implement a C-ARQ
scheme. The emphasis is on the design and analysis of a novel MAC protocol to deal with
the unique characteristics of the contention process that takes place among the active relays
within a cooperation phase. Note that in the considered C-ARQ schemes, upon the
initialization of the cooperation phase, the network has the three following unique
characteristics:
1) The spontaneous “sub-network” formed by the active relays is ad hoc and thus there is no
infrastructure responsible for managing the access to the channel.
2) This sub-network formed by the active relays surrounding the node calling for
cooperation is suddenly (sharply) set into saturationnditions co whenever the cooperation
phase is initiated. Upon the transmission of a CFC packet, all the active relays have a data
packet ready to transmit in order to assist the failed transmission. Therefore, heavy
contention takes place in a previously idle network.
3) Opposite to general communications systems, now fairness is not a major issue to
achieve. Indeed, the main goal is to attempt to assist the failed transmission as fast and
reliable as possible, minimizing the use of the radio resources.
These three characteristics determine the way MAC protocols should be designed within the
context of C-ARQ schemes in wireless networks. Considering the aforementioned
characteristics, we present in this chapter the design and performance evaluation of a novel
high-performance MAC protocol for C-ARQ schemes, named DQCOOP.
an erroneous packet transmission between any other pair of source-destination stations

decides to cooperate by retransmitting a copy of the overheard transmission as long as the
received packet has no errors. A random backoff mechanism with a constant backoff
window is applied to avoid collisions among different helpers. The size of the contention
window of the helpers has to be very small in comparison to the contention window of the
source in order to ensure that helpers retransmit their copy before the original source
retransmits on its own the failed packet. Each helper transmits the copy of the packet at
most once, to ensure that all available helpers cooperate and thus the benefits of diversity
are obtained. On the other hand, FCMAC extends the operation of CMAC by fragmenting
data packets into smaller blocks. Each block contains its own inner FEC field and the whole
packet contains an outer FEC. Upon error detection of a whole packet, only a predefined
number of randomly selected blocks among those received without errors are retransmitted.
If the retransmitted blocks are those that were received with errors at destination, then the
performance is improved. Otherwise, the increased overhead becomes useless. A possible
solution consists in adding a negative acknowledgement (NACK) sent out by the
destination upon error detection, indicating which are the blocks received with errors.
However, the use of NACK in CMAC would imply higher overhead and again, it would
require hardware modifications, thus breaking with the claimed backwards compatibility.
The main limitation of CMAC and FCMAC is that they rely on the fact that helpers can learn
whether other transmissions between any pair of source and destination are successful or
not only by overhearing the radio channel.
In (Wang & Yang, 2005), the Cooperative Diversity Medium Access with Collision
Avoidance (CD-MACA) protocol is proposed within the context of wireless ad hoc networks
operating over the CSMA/CA protocol. Whenever a source terminal fails to receive the CTS
packet, all those stations that had properly received it, take the place of the source terminal
and retransmit the data packet. An analytical model based on Markov chain theory is
proposed to obtain the achievable throughput of the system considering cooperation.
Although the general idea of CD-MACA is rather interesting, the definition in (Wang &
Yang, 2005) is quite general and several implementation details are not considered.
From an energy-efficient perspective, another cooperative MAC protocol is also presented
within the context of ad hoc networks in (Azing et al., 2005). This proposal integrates
cooperative diversity into two different wireless routing protocols by embedding a
distributed cooperative MAC. The initial path establishment performed by the routing
protocol can be done either considering cooperation or not. Cooperation is then achieved by
forcing all the stations to act as a distributed virtual antenna, through which simultaneous
transmissions are separated with CDMA.
In (Sadek et al., 2006) a cooperative MAC protocol was presented within the context of a
mesh network formed by an access point, a number of regular stations, and one fixed
wireless router (relay). A fixed TDMA scheme is applied and empty slots are used for
cooperative relaying. The relay station keeps a copy of all those packets that are not
properly received by the Access Point (AP). At the beginning of each time slot, the relay
listens to the channel. If the channel is idle, it retransmits the packet at the head of its queue.
Based on this main idea, two specific algorithms are proposed to exploit the benefits of
cross-layer design between the PHY and MAC layers.
All these MAC protocols have been designed to achieve an improvement in the network
performance by transmitting through faster multi-hop routes. However, none of them takes
into account the unique characteristics of the C-ARQ schemes for their implementation in
on-demand cooperative schemes. DQCOOP is presented in the next section as a novel MAC
protocol that has been tailored to meet the requirements of the C-ARQ scenario. It
constitutes the adaptation of the high-performance DQMAN protocol (Alonso-Zárate et al.,
2008a) to this kind of scenarios.
3. DQMAN for C-ARQ: DQCOOP

The aim of this section is to present DQCOOP as an extension and adaptation of the high-
performance DQMAN (Alonso-Zárate et al., 2008a) to match the unique requirements posed
by the C-ARQ schemes. DQMAN, in its turn, is the extension of the infrastructure-based
DQCA protocol (Alonso-Zárate et al., 2008b) for wireless ad hoc networks. The new
resultant protocol is called DQCOOP. The rules of DQMAN and DQCA will not be
described into detail in this chapter as they can be found in (Alonso-Zárate et al., 2008a) and
(Alonso-Zárate et al., 2008b), respectively.
In short, the basic idea of DQMAN is that any idle station with data to transmit listens to the
channel for a randomized period of time before establishing its cluster. This Clear Channel
Assessment (CCA) period gets the name of Master Selection Phase (MSP). If the channel is
idle for the whole MSP, then a cluster is established. The station becomes master and starts
broadcasting a periodical clustering beacon (CB) that allows neighbor stations to get
synchronized and become slaves. The master operates as such for as long as there is data
activity within its cluster. Therefore, the cluster structure changes along time as a function of
the aggregate traffic load of the network. Once the cluster is established, the master station
transmits its own data and it acts as the AP of a WLAN wherein DQCA can be executed. For
completeness, we review the basic protocol rules of DQCA in the next section.
3.1 DQCA Overview

The purpose of this section is to highlight the basic features of DQCA. As demonstrated in
(Alonso-Zárate et al., 2008b), DQCA outperforms the widely commercially spread
Distributed Coordination Function (DCF) of the IEEE 802.11 Standard and remains stable
even when the traffic load occasionally exceeds the channel capacity.
DQCA is a MAC protocol designed to manage the access to the channel in the uplink of an
infrastructure WLAN. Time is divided into MAC frames, and each frame is divided in three
parts separated by a Short Inter Frame Space (SIFS) necessary to tolerate propagation delays,
turnaround times, and processing delays. The three parts, depicted in Fig. 2, are:
i) A Contention Window (CW) further divided into m access minislots wherein the nodes
can send a short chip sequence named Access Request Sequence (ARS) to request access to
the channel. An ARS is a short chip sequence that contains no explicit information but has a
specific and predefined pattern that allows the AP to distinguish between an idle minislot,
the presence of just one ARS, or the occurrence of a collision between two or more
simultaneous ARS.
ii) A data slot reserved for the transmission of data packets.
iii) A feedback part wherein the AP broadcastsa Feedback Packet (FBP) that contains the
data acknowledgment, the state of the each of the minislots of the CW for contention
resolution algorithm, and a ‘final message bit’ that is enabled (set to one) by the AP to
identify the last data packet (fragment) of a message. Of course, nodes must also include a
‘final message bit’ in their data packet transmissions in order to advertise the transmission
of the final fragment of each message.
All the nodes execute three sets of simple rules at the end of each MAC frame. By simply
using the feedback information attached to the FBP, they can update the state of two
distributed queues (explained below) to execute the access algorithm. According to the
protocol rules, DQCA operates as a random access protocol when the traffic load is low (an
immediate access rule of the protocol allows a station to get access to the channel
immediately if the distributed queues are empty), and it switches smoothly and
automatically to a reservation protocol as the traffic load increases. Therefore, it attains the
better of the access methods.
The protocol operation is based on two concatenated distributed queues, the Collision
Resolution Queue (CRQ) and the Data Transmission Queue (DTQ). The CRQ is responsible
for the resolution of collisions among ARS and the DTQ handles the transmission of data.
The number of occupied positions (or elements) in each queue is represented by an integer
counter (RQ and TQ for the CRQ and the DTQ, respectively). Both counters have the same
value for all the nodes in the system and are updated according to a set of rules at the end of
each frame. Each node must also maintain and update another set of counters that reveal its
position in the queue (pRQ and pTQ for the CRQ and the DTQ, respectively). By the term
“position” it is meant the relative order of arrival (or age) of the node in the respective
queue. In the CRQ, each position (or element) is occupied by a set of nodes that suffered an
ARS collision (i.e. attempted an ARS transmission in the same access minislot of the same
CW). The DTQ contains the nodes that successfully reserved the channel through an ARS
and therefore each queue element corresponds to exactly one node.
Fig. 2. DQCA Frame Structure
3.2 Motivation and Problem Statement

The intuitive idea behind DQCOOP is that the destination asking for cooperation gets the
role of masterand coordinates the retransmissions from the relays, which become slaves, as
in DQMAN. Then, a temporary cluster is established around the destination and a variation
of DQCA can be executed. This is represented in Fig. 3.
SLAVE
Relay 1
SLAVE
Relay 2
SLAVE
Relay 3
destination
MASTER
Time+
Master FBP FBP ACK
Slave 1 ReTX
Slave 2 ReTX
Slave 3
cooperation phase
Fig. 3. Master-Slave Architecture of DQCOOP (simplified example)
The master, i.e., the destination, initiates the periodic broadcast of the FBP and creates a
temporary cluster. A cooperation phase is initiated. The slaves, i.e., the relays, request access
to the channel to retransmit their cooperative packet (retransmissions of the original source
transmissions) by executing a variation of the DQCA rules. It is assumed that the relays
attempt to retransmit persistently until the cooperation phase is finished. Whenever the
cooperation phase is finished either an ACK or a NACK packet is transmitted, indicating
either the successful or unsuccessful recovery of the data packet originally received with
errors, respectively.
However, DQMAN, as defined in (Alonso-Zárate et al., 2008a), would be inefficient in
managing the access to the channel in a C-ARQ scheme. This is mainly due to the fact that
upon cooperation request (broadcast by the destination), the group of active relays forms an
ad hoc network wherein all the active stations suddenly have a data packet ready to be
transmitted. This turns temporarily the network from idle to saturation conditions. This
idle-to-saturation sharp transition would cause DQMAN to spend a non-negligible start-up
time before attaining its high performance, mainly due to:
1) The simultaneous channel access requests from the active relays in the first frame
immediately after the transmission of the CFC would have a high probability of collision.
Therefore, some empty frames would be needed until the first collision could be solved and
data retransmissions could actually start.
2) Upon the transmission of the first FBP, all the active relays (slaves) would retransmit in
the following frame by executing the immediate access rule of DQCA (Slotted ALOHA
access for low traffic loads). All these transmissions would collide, causing a waste of
resources for the duration of a complete MAC frame.
3) Even with the immediate access rule disabled, an empty frame would be present when
the collision resolution process starts due to the MAC frame structure with the feedback
broadcast at the end of the frame.
Therefore, it is necessary to expand and adapt the DQMAN operation to take into
consideration the aforementioned issues that may potentially degrade its performance in C-
ARQ schemes. DQCOOP is presented in the next section with the goal of attaining the near-
optimum performance of DQMAN within the context of the considered C-ARQ scheme.
3.3 Protocol Description

The core operation of DQCOOP is highly based on DQMAN. However, the clustering
algorithm and the MAC protocol (frame structure and protocol rules) are modified to meet
the requirements of the C-ARQ scheme. Their descriptions are presented in the next two
sections.
3.3.1 Clustering Algorithm

In DQCOOP, the clustering algorithm of DQMAN is modified as it follows:
1) The destination, and not the transmitter as in DQMAN, takes the master role when a
cooperation phase is initiated with the transmission of a CFC packet. Some of the relays
which received the original data packet (received with errors by the destination to trigger a
cooperation phase) and also receive the CFC transmitted by the destination become active
relays. These active relays get the role of slaves. A cluster is then established. The master
periodically broadcasts a FBP, in the same way as in DQMAN, to provide the slaves with
the minimum feedback information necessary to execute the protocol rules at the end of
each frame.
2) There is no CCA prior to the establishment of the cluster. This means that the destination
station does not have to contend with other users to get access to the channel. Therefore the
contention within the MSP associated with DQMAN is avoided with DQCOOP. This can be
actually performed as the CFC is transmitted instead of the ACK when receiving a packet
with errors. ACK packets are usually given priority over all kind of traffic (in wireless
networks), and thus there is no need for contention in this case.
3) The cluster is broken up whenever the master either manages to decode the original
packet or discards the packet. The cooperation phase is ended with the transmission of the
ACK packet. Otherwise, if a maximum time-out expires and the original packet cannot be
decoded, a NACK packet is transmitted and the cluster is broken up as well. That is, in fact
all the stations become idle upon the transmission of either the ACK or the NACK by the
master.
3.3.2 The MAC Protocol: Frame Structure and Protocol Rules

When a cooperation phase is initiated, time is divided into five parts as represented in Fig. 4.
Upon the transmission/reception of each FBP, all the stations execute the protocol rules of
DQCA. The five parts of a cooperation phase within the context of DQCOOP are:
1) A CFC transmission. The cooperation phase is initiated when a CFC is broadcast by the
destination station upon the reception of a data packet with errors. This CFC takes the form
of a special FBP and indicates that immediate access is forbidden.
2) An initial contention window composed of m0 minislots follows the CFC transmission
wherein every active relay station randomly selects (with equal probability) one out of the
m 0 minislots where to send an Access Request Sequence (ARS).
3) A FBP transmission. A FBP is broadcast by the master station with the feedback
information regarding the state of each of the m0 previous minislots. As in DQMAN, for
each minislot, this information can have one out of three values. It can be empty (E), i.e., no
ARS transmitted, success (S), i.e., exactly one ARS transmitted, or collision (C), i.e., more
than one ARS transmitted in the same minislot (no matter how many).
4) A number of regular DQMAN consecutive MAC frames follow this first FBP until the
cooperation phase is ended. The rules of DQMAN, with the exception of the immediate
access rule, are executed to manage the data retransmissions and the resolution of the
collisions. The contention window of these frames has m minislots, where in general
m <m 0 , although this is not a mandatory condition.
5) An ACK or NACK transmission. Whenever the destination is able to successfully
decode the original packet, it broadcasts an ACK packet indicating the end of the
cooperation phase. A NACK is transmitted if the packet cannot be decoded at some point in
time.
Short Inter Frame Spaces (SIFS) are left between each of the parts of the cooperation phase to
compensate for non-negligible propagation and data processing delays and turnaround
times to switch the radio transceiver from receiving to transmitting mode.
It is worth mentioning that the value of m 0 must be tuned according to the expected number
of active relays. The higher the number of active relays, the higher the value of m 0 in order
to reduce the probability that all the access requests collide in the first frame. However, a
high value for m 0 has a cost in terms of control overhead. On the other hand, as long as at
least one access request is successful, the data transmission process can be initiated from the
first MAC frame, avoiding thus the loss of resources.
Time+
Cooperation Phase (5 parts)

1 2 3 4 5
Source DATA
Destination CFC FBP FBP ACK
Relay 1 Coop_DATA
Relay n Coop_DATA
SIFS Initial Contention SIFS

Contention Window
Window m minislots Control
m0 minislots, Minislots
with m0>m Feedback
(S,C,E)
Fig. 4. DQCOOP MAC Frame Structure
3.3.3 Operational Example

A simple network layout with six stations is considered, all of them in the transmission
range of each other. A source station (S) transmits to a destination station (D ) with the
support of relays R1, R2, R3, and R4. TQ and RQ represent the size of the DTQ and the CRQ,
respectively, and pTQ i and pCRi represent the position of the ithuser in the DTQ and CRQ,
respectively.
The cooperation phase is represented in Fig. 5 and explained as follows:
1) Upon the reception of the data packet with errors, D initiates a cooperation phase by
broadcasting a CFC. This packet sets the start of frame 0.
2) Frame 0 contains 5 access minislots (m 0 =5). The set of relays {R1, R2, R3, R4 } select the set
of minislots {3, 1, 5, 5}.
3) At the end of frame 0, D broadcasts the FBP with the following feedback information
regarding the state of the minislots, i.e., {Success,Empty,Success,Empty,Collision }.
4)Upon the execution of the protocol rules, R2 gets the first position of DTQ, R1 gets the second
position of DTQ, and both R3 and R4 get the first position of CRQ. In terms of the four integer
number representing the queues, this can be written as {pTQ1 , pTQ2 , pTQ3 , pTQ4 }={2,1,0,0} and
{pRQ1 , pRQ2 , pRQ3 , pRQ4 }={0, 0, 1, 1}. On the other hand, TQ =2 and RQ =1.
5) During frame 1, both the data transmission and the collision resolution work in parallel.
At the beginning of the frame, containing 2 access minislots (m=2), R3 and R4 attempt to
solve their collision. They reselect an access minislot where to send an ARS. In this case, they
select minislots 1 and 2 respectively, and thus they successfully solve their collision.
6) On the other hand, R2, which is at the first position of DTQ, transmits data (a
retransmission of the original packet).
7) At the end of frame 1, the FBP broadcast by D indicates that a transmission has been successful
and the next station in DTQ should transmit in the following frame. In addition, the feedback
information on the state of the minislots allows R3 and R4 to queue, orderly in time, in DTQ.
Time+
Cooperation Phase
Frame 0 Frame 1 Frame 2
DATA Wait for cooperation

S, Source
CFC FBP FBP ACK
D, Destination
Coop_DATA
R1, Relay 1
Coop_DATA
R2, Relay 2 Access Request
Squence (ARS)
R3, Relay 3
R4, Relay 4
CRQ CRQ
3,4
DTQ DTQ
1 2 4 3 1
Fig. 5. DQCOOP Example of Operation
8) In frame 2, there are no collisions to be solved and thus the minislots are empty. R1
transmits data.
9) Upon the reception of the retransmission from R1, D is able to successfully decode the
original packet. Therefore, it transmits an ACK packet indicating the end of the cooperation
phase. All the relays discard the buffered cooperative packet.
3.4 Performance Evaluation

The performance of DQCOOP is evaluated in this section through a C++ custom-made
simulator. In order to focus on the evaluation of the cooperation phases, a single-hop
network wherein all the data transmissions from a fixed source to a fixed destination are
received with errors is considered. That is, the destination always broadcasts a CFC packet
upon the reception of every original data packet received from the source station. Moreover,
the source has always a packet ready to be transmitted to the destination.
In this performance evaluation, we will measure the average packet transmission delay
defined as the period of time elapsed from the moment that a packet is first transmitted
from the source until it can be decoded at destination after receiving K retransmissions. For
all the experiments, we assume that a constant number of relays are activated within each
cooperation phase. Furthermore, and without loss of generality, the destination is
considered to require a constant number K of retransmissions from the relay set to decode
the original packet.
The simulation parameters are summarized in Table 1.
Parameter Value Parameter Value

Control Rate 6 Mbps MAC header 34 bytes
Data Rate
24 Mbps PHY preamble 96 μs
(Source)
Data Rate ACK, CFC, FBP
54 Mbps 14 bytes
(Relays) length
Packet Length 1500 bytes SlotTime 10 μs
ARS 10 μs SIFS 10 μs
Table 1. Simulation Parameters
3.4.1 Number of Minislots in the Cooperation Phase

The average packet transmission delay as a function of the number of active relays in a
cooperative phase is represented in Fig. 6 when K=3 and for different values of m 0 and m ,
which in addition accomplish that m 0 =m . Each curve represents the results obtained with
different number of access minislots (m ).
For low values of m , the average packet transmission delay gets lower as the value of m
increases thanks to the faster collision resolution process. However, increasing the number
of access minislots also increases the MAC overhead. The addition of an extra minislot
entails an extension of the frame duration (devoted to overhead) and also enlarges the size
of the FBP that contains the state of each one of the minislots. Therefore, as it can be seen in
the figure, for high values of m , e.g., m=10, the fact that the collision resolution becomes
shorter in time does not pay off the increase in the protocol overhead when the number of
active relays is low and thus the average packet transmission delay gets higher. This can be
better appreciated in Fig. 7 where the average packet transmission delay for the scenario
with 5 relays is plotted as a function of the number of access minislots m=m 0 . In this curve it
is easier to see that, for low number of access minislots, an increase in the number of
minislots leads to lower average packet transmission delays. However, over a given
threshold, the faster resolution of collisions due to the longer contention window does not
compensate for the MAC overhead and the average packet transmission delay increases
with the number of access minislots. For this reason, it is necessary to find a good
compromise between the faster collision resolution and the protocol overhead. This tradeoff
will be further discussed later in the next section.
3.9
m0=m=2
3.8
Average Packet Transmission Delay (ms)
m0=m=3
3.7 m0=m=4
m0=m=5
3.6 m0=m=10
3.5
3.4
3.3
3.2
3.1
K=3
3
2 3 4 5 6 7 8 9 10 11 12 13 14 15
Active Relays
Fig. 6. Average Packet Transmission Delays for Different Values of m 0 =m
3.4
Active Relays=5
Average Packet Transmission Delay
3.35
3.3
(ms)
3.25
3.2
3.15
2 3 4 5 6 7 8 9 10
Number of minislots (m=m 0)
Fig. 7. Average Packet Transmission Delay for Different Values of m 0= m
Getting back to the results in Fig. 6, they show that the average packet transmission delay
drops remarkably when the number of access minislots is at least equal to 3. Higher values
of m do not result in any substantial reduction of this time. Therefore, as it happens with the
DQCA protocol (Alonso-Zárate et al., 2008b), a good operational point for DQCOOP is to set
m=3.
It is interesting to evaluate whether this discussion is still valid for any arbitrary number of
required retransmissions (K). The average packet tr ansmission delay is plotted in Fig. 8 as a
function of the value of K and for different values of m 0 =m when the number of active relays
is 15. In all cases, there is a considerable reduction of the average packet transmission delay
when shifting from 2 to 3 minislots. However, there is no much interest in increasing the
number of access minislots to higher values than 3, at least in terms of packet transmission
delay. Therefore, it is important to reinforce the already known argument that the number
of access minislots should be set to 3 in any DQCA-like protocol.
However, it seems reasonable to think that the value of m 0 (the number of access minislots
within the very first frame after the transmission of the CFC) could be set to a higher value
than m in order to absorb the first multiple access request arrival from all the active relays.
Note that the first frame is the one that receives the maximum number of simultaneous
access requests. In subsequent frames, the requests are split into smaller groups according to
the m -ary tree-splitting collision resolution operation of DQMAN.
In the next section, m is set to 3 and the performance of the protocol is evaluated for
different values of m 0 . The aim is to evaluate the reduction of the average packet
transmission delay for m 0 >m .
8
m=m0=2
7 m=m0=3
m=m0=4
m=m0=5
6
Active Relays=10
2
1 2 3 4 5 6 7 8 9 10
Number of Required Retransmissions
Fig. 8. Average Packet Transmission Times for Different Values of K
3.4.2 Number of Minislots in the Start-up Phase (m0)

The performance of DQCOOP for different values of m 0 is evaluated in this section. As
discussed before, an increase in the number of minislots of the first frame reduces the
probability of collision in the first access requests upon initialization of a cooperation phase
and, therefore, it should yield a lower average packet transmission delay. However, it also
entails an increase of the protocol overhead (frame length and amount of required feedback
information).
In order to quantify this tradeoff, first note that the duration of a cooperation phase can be
decomposed as the sum of time devoted to the transmission of data and the overhead due to
the necessary MAC protocol. This overhead time includes silent intervals as well as the time
devoted to the transmission of control packets. Considering this, the relative overhead is
defined as the ratio between the overhead time in the cases that m 0 > 1 and the overhead time
when m 0 = 1 (this latter case is the worst case in terms of overhead since all the relays collide
in the first access request with probability one). This definition allows plotting the curves
with different values of K in the same vertical axis and also makes the results independent
of the absolute values of the transmission rates used for the simulation and the numerical
evaluation.
The relative overhead is plotted in Fig. 9 as a function of the value of m 0 , for different
number of required retransmissions (K), and considering a total number of 5 active relays.
The first observation is that there is a close relationship between the overhead of the
protocol and the value ofm 0 . The value of the relative overhead is very sensitive to the value
of m 0 if the number of required retransmissions is low. This means that if the value of K is
low, the accurate tuning of the value of m 0 has a remarkable effect on the performance of the
C-ARQ scheme. All the curves show a local minimum of the relative overhead for any pair
of values of m 0 and K. However, on the other hand, the higher the values of K, the more flat
the curves become. This means that if the number of required retransmission is high, the
value of m 0 becomes a non-critical parameter on the performance of DQCOOP.
The main reason for this behavior is that when the number of required retransmissions is
high and thus the duration of the cooperation phase is long, the impact of the overhead of
the first frame on the performance of DQCOOP is low. Note that if K retransmissions are
needed, at least K frames are necessary.
On the other hand, it seems reasonable to believe that the selection of the value of m 0 should
depend on the number of active relays (which request access simultaneously in the first
frame). In order to evaluate this relationship, the average packet transmission delay is
plotted in Fig. 10 for K=3. Different curves are plotted for different number of active relays
and as a function of the value of m 0 .
1
Active Relays=5
K=1 K=2 K=3
K=4 K=5 K=6

0.95
K=7
Relative Overhead
0.9
0.85
0.8
0.75
1 2 3 4 5 6 7 8 9 10 11 12
Number of Minislots First Frame (m 0)
Fig. 9. Protocol Relative Overhead (DQCOOP)

3.5
Active Relays=5

3.45
Active Relays=10
3.4 Active Relays=15
3.35
3.3
3.25
3.2
3.15
K=3
3.1
2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20
Number of Access Minislots First Frame (m 0)
Fig. 10. Average Packet Transmission Delay as a Function of m 0
It is worth noting that when m 0  10 the three curves almost overlap. This means that, if this
condition is fulfilled, the average packet transmission delay is almost equal and
independent of the number of active relays. In addition, the value of the average packet
transmission delay at m 0 =10 is not substantially bigger than the one at the respective
minimum values that can be found for m 0 =6 (for 5 active relays), m 0 =7 (for 10 active relays),
and m 0 =10 (for 15 active relays). This constitutes a worthwhile design guideline since by
setting m 0 = 10 the average packet transmission delay for any value of K can be predicted
with reliable accuracy regardless of the number of active relays in each cooperation phase
(considering a practical situation with no more than 15 active relays). In addition, this fact
relaxes the configuration requirements of the network, which is of remarkable interest when
operating in fully decentralized and spontaneous networks.
4. Conclusions
In this chapter we have highlighted the important role of the MAC layer in the performance
of C-ARQ schemes. Typically, these kinds of schemes have been evaluated from
fundamental points of view and assuming perfect scheduling among the relays. However,
we have shown that efficient MAC protocols are necessary to fulfill the specific
requirements posed by C-ARQ schemes and to get the most of their potential to increase the
efficient of wireless communications.
In addition, we have presented the DQCOOP protocol as an extension and adaptation of
DQMAN to efficiently coordinate the contention among the relays in a C-ARQ scheme. It
has been necessary to redesign the initialization phase of a DQMAN cluster so as to manage
the idle-to-sharp traffic transition that takes place upon the transmission of a CFC. Since the
active relays attempt to help simultaneously, the first contention window of DQMAN has to
be resized. In addition, the protocol frame structure and the protocol rules have been also
modified to optimize the performance of DQMAN in the context of C-ARQ schemes.
The performance of the protocol has been evaluated with computer simulations. Results
show that the performance of DQCOOP can be independent of number of active relays and
the number of access minislots. This is a desirable characteristic in fully decentralized
networks, as is the case of ad hoc networks, where there might be no previous knowledge of
the network topology and configuration. Results also show that this independency can be
simply accomplished by setting the number of access minislots to 3 (attaining a faster
resolution of collisions compared to the transmission of data) and properly dimensioning
the number of access minislots in the very first frame, which is also modified to avoid an
otherwise certain empty data field. This last modification aims at absorbing the first
simultaneous access request by all the active relays. In fact, results show that the number of
access minislots in the very first frame can be overdimensioned at almost no cost, and thus
the performance of DQCOOP can be independent of the number of relays. The cost of
increasing by one unit the number of access minislots in terms of overhead pays off the
reduced probability of collision in the first access request.
5. References
Alonso-Zárate J., Gómez J., Verikoukis C., Alonso L., & Pérez-Neira A. (2006). Performance
Evaluation of a Cooperative Scheme for Wireless Networks, Proceedings of the IEEE
PIMRC , pp. 1-5, ISBN 1-4244-0329-4, Helsinki, Finland, September 2006
Alonso-Zárate J., Kartsakli E., Skianis C., Verikoukis C., & Alonso L. (2008a), Saturation
Throughput Analysis of a Cluster-based Medium Access Control Protocol for
Single-hop Ad Hoc Wireless Networks, Simulation: Transactions of the Society fo
ModelingandSimulationInternational , Vol. 84, No. 12, 619-633, ISSN 0037-5497
Alonso-Zárate J., Kartsakli E., Cateura A., Verikoukis C., & Alonso L. (2008b). A Near-
Optimum Cross-Layered Distributed Queueing Protocol for Wireless LAN, IEEE
Wireless Communications Magazine, Special Issue on MAC protocols , Vol. 15,for WLAN
No. 2, (February 2008) 48-55, ISSN 1536-1284
Azgin A., Altunbasak Y., & AlRebig G. (2005). Cooperative MAC and Routing Protocols for
Wireless Ad Hoc Networks, ProceedingsoftheIEEEGLOBECOM205 , ISBN: 0-7803-
9414-3, December 2005
Biswas S. & Morris R. (2005). ExOR: Opportunistic Multi-hop Routing for Wireless
Networks, Proceedings of the SIGCOMM '05 , pp. 133-144, ISBN 1-59593-009-4 , New
York, NY, USA, 2005.
Campbell G. et al. (2002), Method and apparatus for detecting collisions and controlling
access to a communications channel, USPatentno.US609B1 48 , June 2002.
Cerruti I., Fumagalli A., & Gupta P. (2008). Delay Model of Single-Relay Cooperative ARQ
Protocols in Slotted Radio Networks Poisson Frames Arrivals, IEEE/ACM Trans. on
Networking , Vol. 16, No. 2, (April 2008) 371-382, ISSN 1063-6692
Chakraborty S. S., Liinaharja M., & Ruttik K. (2005). Diversity and packet combining in
Rayleigh fading channels, IEE Proceedings-Communications , Vol. 152, No. 3, (June
2005) 353 – 356, ISSN 1350-2425
Dianati, M., Ling X., Naik K., & Shen X. (2006). A Node-Cooperative ARQ Scheme for
Wireless Ad Hoc Networks, IEEE Trans. on Vehicular Technology, Vol. 55, No. 3,
(May 2006) 1032-1044, ISSN 00189545
Fitzek F. H. P. & Katz M. D. (2006). Cooperation in Wireless Networks: Principles

Applications , Ed. Springer, ISBN 978-1-4020-4710-7, The Netherlands
García-Vidal J., Guerrero-Zapata M., Morillo J., & Fusté D. (2007). A Protocol Stack for
Cooperative Wireless Networks, In: Wireless Systems and Mobility in Next Generatio
Internet, in Lecture Notes in Computer Science , 62-73, Springer,
LNCS)
( ISBN: 978-3-
540-70968-8, Springer, (2007)
Goldsmith A. (2005). Wireless Communications , Cambridge University Press, ISBN
0521837162
Gómez J. & Pérez-Neira A. (2006). Average Rate Behavior for Cooperative Diversity in
Wireless Networks, Proceedings of the IEEE Internatio nal Symposium on Circuits and
SystemsISCAS)
( , pp. 5309-5402, ISBN 0-7803-9389-9, Kos, Greece, May 2006
Gómez J., Alonso-Zárate J., Verikoukis C., Pérez-Neira A., & Alonso L. (2007). Cooperation
On Demand Protocols for Wireless Networks, Proceedings of the IEEE PIMRC , pp. 5,
ISBN: 978-1-4244-1144-3, Athens, Greece, September 2007
Gupta P., Cerruti I., & Fumagalli A. (2004). Three Transmission Scheduling Policies for a
Cooperative ARQ Protocol in Radio Networks, Proceedings of the WNCG Conference ,
Austin, October 2004
IEEE 802.11 Standard, Part 11: Wireless LAN Medium Access Control (MAC) and Physical
Layer (PHY) Specifications, IEEE Std. 802.-11-99, Revision 2007
Korakis T., Natayanan S., Bagri A., & Panwar S. (2006). Implementing a Cooperative MAC
Protocol for Wireless LAN, Proceedings of the IEEE International Conference
Communications(ICC’0 , pp.
6) 4805-4810, ISBN: 1-4244-0355-3, June 2006
Liu P., Tao Z., Lin Z., Erkip E., & Panwar S. (2006), Cooperative Wireless Communications:
A Cross-Layer Approach, IEEE Wireless Communications Magazine, Special Issue on
AdvancesinSmartAntennas , Vol. 13, No. 4, (August 2006) 84-92, , ISSN 1536-1284
Liu P., Tao Z., & Panwar S. (2007). CoopMAC: A Cooperative MAC for Wireless LANs, IEEE
Journal on Selected Areas on Communications , Vol. 25, No.2, (February 2007) 340-354,
ISSN 0733-8716
Morillo-Pozo J., García-Vidal J., & Pérez-Neira A. I. (2005). Collaborative ARQ in Wireless
Energy-Constrained Networks, Proceedings of the 205 Join t Workshop on Foundations
ofMobileComputing205(DIAL-POM’05), ISBN 1-59593-092-2
Morillo-Pozo J. D., & García-Vidal J. (2007). A Low Coordination Overhead C-ARQ Protocol
with Frame Combining, Proceedings of the IEEE PIMRC , pp. 1-5, ISBN 978-1-4244-
1144-3, Athens, Greece, September 2007
Nosratinia, A., Hunter, T. E., & Hedayat A. (2004). Cooperative Communications in Wireless
Networks, IEEE Communications Magazine , Vol. 42, No. 10, (October 2004) 74-80,
ISSN 0163-6804
Pfletschinger S. & Navarro M. (2008). Link Adaptation with Retransmissions for Partial
Channel State Information, Proceedings of the IEEE GLOBECOM 208 , pp. 1-6, ISBN
978-1-4244-2324-8, New Orleans, Louisiana, USA, November 2008
Tao Z., Korakis T., Slutskiy Y., Panwar S., & Tassiulas L. (2007). Cooperation and
Directionality: A Co-opdirectional MAC for Wireless Ad Hoc Networks,
ProceedingsoftheWiOpt , pp.
207 1-8, ISBN: 978-1-4244-0960-0, April 2007
Shankar S., Chou C., & Ghosh M. (2005). Cooperative Communication MAC (CMAC) – A
new MAC protocol for Next Generation Wireless LANs, Proceedings of the IEEE
International Conference on Wireless Networks, Communications, and Mobil
205, pp. 1-6, ISBN: 0-7803-9305-8, June 2005
Sadek B., Ray Liu K. J., & Ephremides A. (2006). Collaborative Multiple-Access Protocols for
Wireless Networks, Proceedings of the IEEE International Conference
Communications(ICC’0 , pp.
6) 4495-4500, ISBN: 1-4244-0355-3, June 2006
Wang X. & Yang C. (2005). A MAC Protocol Supporting Cooperative Diversity for
Distributed Wireless Ad Hoc Networks, Proceedings of the IEEE PIMRC , pp. 1396-
1400, ISBN: 9783800729098, Berlin, Germany, September 2005
Zimmermann M. D., Herhold P., & Fettweis G. (2004). The Impact of Cooperation on
Diversity-Exploiting Protocols, Proceedings of the 59th IEEE Vehicular Technology
Conference , pp. 410-414, ISBN 0-7803-8255-2, Milan (Italy), May 2004
Zimmermann E., Herhold P., & Fettweis G. (2005). On the Performance of Cooperative
Relaying Protocols in Wireless Networks, European Trans. on Telecommunications ,
Vol. 16, No. 1, (January 2005) 5-16, ISSN 1541-8251
Zorzi M., Rao R. R., & Milstein L.B. (1997). ARQ error control for fading mobile radio
channels, IEEE Trans. on Vehicular Technology , vol. 46, no. 2, (May 1997) 445-455,
ISSN 00189545
A Hybrid Feedback Mechanism to Exploit Multiuser Diversity in Wireless Networks 247
A Hybrid Feedback Mechanism to Exploit

Multiuser Diversity in Wireless Networks
Yahya S. Al-Harthi
Electrical Engineering Department
King Fahd University of Petroleum and Minerals
Dhahran 31261, Saudi Arabia
Abstract
One of the most promising approaches to boost the communication efficiency in wireless sys-
tems is the use of multiuser diversity (MUDiv), where the fading of channels is exploited.
The mechanism of scheduling the user with the best channel condition is called opportunis-
tic scheduling (OS). In this paper we propose a joint polling and contention based feedback
(JPCF) algorithm that exploits MUDiv while reducing the feedback load. The guard time,
which is between bursts, is divided into minislots that alternate between polling-based feed-
back minislot (p-minislot) and contention-based feedback minislot (c-minislot). During the
minislot, users feedback their channel qualities if above a predetermined threshold. We ana-
lyze the scheduling algorithm under slow Rayleigh fading assumption and derive the closed-
form expressions of the feedback load as well as the system capacity. We also consider the
delay resulting from the time needed to schedule a user and derive the system throughput.
The scheduling algorithm is compared with other scheduling algorithms.
1. Introduction
With the emerging of new multimedia applications and the huge demand and growth for such
applications, the need to provide high-speed high-rate transmission techniques is demanding.
One technique that has the capability of supporting such applications is multiuser diversity
(MUDiv), where the fading of channels is exploited (1). Assume that a reasonably large num-
ber of users are actively transmitting/receiving packets in a given cell, and they experience
independent time-varying fading conditions. By transmitting data to only the instantaneous
“on-peak" user, opportunistic scheduling (OS) can efficiently utilize the wireless resources and
thus dramatically improve the overall system throughput (2), (3). In order to schedule the
best user, in terms of channel quality, each user measures his instantaneous signal-to-noise
ratio (SNR) and feeds it back to the network scheduler. Currently Qualcomm’s high data rate
(HDR) system require similar scheme (4). As the number of active users becomes high, re-
sources are wasted in carrying this amount of feedback, which lead to inefficient use of the
spectrum. This issue motivated researchers to propose new techniques to reduce the feedback
load while exploiting MUDiv.
Many investigations have been conducted to exploit multiuser diversity while keeping the
overhead as minimum as possible. For instance, the impact of the degree of quantization of
the SNR measurements on the throughput of constant rate transmission was investigated in
(5). It was shown that reducing the feedback rate (few quantization levels) yields a good per-
formance compared to the unquantized feedback. In single carrier systems, (6) proposed a
discrete rate switch-based multiuser diversity (DSMUDiv) algorithm that reduced the feed-
back load and feedback rate while preserving the performance of OS. The scheduling algo-
rithm relied on a probing mechanism. One drawback of this algorithm is that as the average
SNR decreases the need for full probing increases, therefore, the algorithm suffers from high
feedback load at low average SNR values. The work was extended to multi-carrier systems in
(7), where further reduction in the feedback load was shown. Similarly, in (8) users compare
their sum-rate on all sub-carriers to a predefined threshold value. In (9), the threshold value
was optimized to meet a specified outage probability. Users with channel quality above the
threshold are allowed to feedback their SNR measurements, while all others remain silent. In
case no feedback, a random user is selected, which leads to loss in capacity. Also, the feedback
values are unquantized (analog values), which increases the feedback rate. In (10), the work
was extended, where the scheduler requests full feedback if none of the users’ channel quali-
ties are above the threshold. Although the loss in capacity was compensated by full feedback
as opposed to (9), the feedback load was increased. In (11), multiple feedback thresholds are
used in order to reduce the feedback load and exploit multiuser diversity. The feedback load
was reduced at an expense of scheduling delay. In addition, a set of switched-based multiuser
access schemes were proposed in (12) in order to reduce the feedback load. In (13), (14), the
feedback rate was reduced to a one bit feedback per user. The scheduler uses these feedback
bits to partition all users into two sets and assigns the channel to one user belonging to the
set experiencing favorable channel conditions. Although this reduces the feedback load and
feedback rate but some loss is expected due to the low resolution of the quantized feedback
(one bit). Other work have considered multi-carrier systems, for instants, in (15) a clustering
scheme was proposed to reduce the feedback load while slightly degrading the performance.
Each user feeds back a figure-of-merit listing its strongest clusters of subcarriers. Similarly, in
(16) a fixed number of subcarriers for each user is fed back instead of a full feedback. These
subcarriers are either the best K out of M subcarriers or K predetermined subcarriers. All pre-
viously mentioned work considered a polling threshold-based scheduling schemes to reduce
the feedback load.
Other work considered random access algorithms to exploit multiuser diversity while reduc-
ing the overhead. In (17) a distributed multiaccess scheme was proposed. Based on a channel
quality threshold, users report the best m channels with quality above the threshold value.
Optimization was performed to find the best threshold and m values to maximize the system
capacity. In (18), a random access threshold-based feedback scheme was proposed to exploit
diversity gain while reducing the feedback load. Parameter optimization is performed to al-
low only users with high channel gains to feedback, which maximizes the system capacity.
In (19), a medium access control protocol was designed based on a splitting algorithms to re-
solve collisions over a sequence of minislots, and determine the user with the best channel.
Contention resolution algorithms while exploiting the diversity gain were proposed in (20).
Finally, reduced feedback overhead algorithms were studied in (21). Static splitting was con-
sidered for the best effort traffic scenario. Whereas, for the traffic mixture scenario combine
contention and polling based feedback was considered to maintain quality of service.
While Polling-based algorithms, like the DSMUDiv, guarantee best user selection, it suffers
from scheduling delay. On the other hand, although contention-based algorithms reduces the
overhead more then centralized scheduling it suffers from loss in capacity. In this chapter
we introduce an opportunistic scheduling algorithm that is not only a polling-based feedback
algorithm like in (6), but it is a joint polling and contention based feedback (JPCF) algorithm.
During the guard time, which is divided into minislots, users feedback their channel qualities,
based on a predetermined channel threshold, either in a polling-based feedback minislot (p-
minislot) or/and a contention-based feedback minislot (c-minislot). When a the best user is
found data transmission begins. The scheduling algorithm is analyzed under slow Rayleigh
fading assumption and closed-form expressions of both the feedback load and the system
capacity are derived. We also look at the effect of the delay on the throughput, where we
derive the system throughput. The scheduling algorithm is compared with the DSMUDiv (6)
and the optimal (full feedback) selective diversity scheduling algorithms.
This chapter is organized as follows. Section 2 introduces the system model. Section 3 and
4 present the scheduling algorithm and the mathematical analysis of the JPCF algorithm, re-
spectively. In Section 5 we present some numerical examples. Finally, Section 6 ends the paper
with some concluding remarks.
2. System Model
We consider a single free interference cell in a wireless network with K active users commu-
nicating with a base station (BS). We assume downlink scheduling, where only one user is
allowed to receive data transmission in each time slot. The communication is based on time
division duplex (TDD), where downlink and uplink channels are reciprocal.
2.1 Downlink Transmission Model

Let,
yi ( T ) = hi ( T ) · x ( T ) + ni ( T ); i = 1, 2, 3, ....., K (1)
be the baseband channel model, where x ( T ) ∈ C is the transmitted signal in time slot T
and yi ( T ) ∈ C is the received signal of user i in time slot T. The noise processes ni ( T ) are
independent and identical distributed (i.i.d.) sequences of zero mean complex Gaussian noise
with variance σn2 . The fading channel gain from the BS to the ith user in time slot T is hi ( T ). We
adopt a quasi-static fading channel model where hi ( T ) is i.i.d. from burst to burst but remains
constant over each burst. We consider a flat Rayleigh fading model, assuming the fading
coefficients of all users are i.i.d. Therefore, hi ( T ) is a zero-mean complex Gaussian random
variable. The amplitude of hi ( T ), αi ( T ) = |hi ( T )| is Rayleigh distributed with the probability
density function (PDF) given by,
2αi α2
f αi ( α i ) = exp(− i ) (2)
Ωi Ωi
where Ωi = E[α2i ] is the average fading power of the ith user.
2.2 Quantized Feedback

In this work we assume that users adapt their modulation level based on the instantaneous
channel condition. Using this transmission strategy, which is called adaptive modulation
(AM), if the channel is strong at a given time, the transmission can occur with a higher con-
stellation size. Otherwise, a lower constellation size has to be used. Note that the switching
thresholds of the adaptive transmission modes are functions of the modulation scheme and
target error performance. We consider adaptive multilevel quadrature amplitude modulation
(M-QAM) scheme (22). Specifically, we consider a transmission scheme employing uncoded
adaptive discrete rate M-QAM schemes with constellation sizes M = { Mn : Mn < Mn+1 , 0 ≤
n ≤ N }, where M0 = 1 is the user outage (deep fade), and M N is the highest modulation level.
We assume perfect channel estimation and negligible time delay between channel estimation
and signal set adaptation, and as such, the rate adaptation can happen instantaneously. We
also assume error free feedback channels. If we denote the target average bit error probability
(BEP) by BEPo , thresholds or switching thresholds can be obtained according to (22, eq.(30)):
(1)
γth = [erfc−1 (2 · BEPo )]2 ,
(n) 2
γth = − (2n − 1) ln(5 · BEPo ); n = 2, 3, ....., N, (3)
3
( N +1)
γth = +∞,
where erfc−1 (·) denotes the inverse complementary error function.

Similar to (6), with the assumption of discrete rates, a user estimates his SNR and instead of
feeding back the analog value it is mapped to a quantized value that represent the modulation
level, which can be supported, and then this quantized value is fed back.
Define the set of quantized values (binary bits) Q = {q(n) : q(n) < q(n+1) , 0 ≤ n ≤ N } that
represents the modulation levels, (q(n) →map Mn ). If we assume n modulation levels, then
each quantized value q(m) , 0 ≤ m ≤ n, will be log2 n bits in length. To illustrate the idea of the
SNR mapping, assume γi is the estimated SNR of the ith user, then the quantized value is,

(1)
q(0) if γi < γth (outage),


(n) ( n +1)
qi = q(n) if γth ≤ γi < γth (4)


q( N ) if γ ≥ γ( N ) .
i th
2.3 Uplink Feedback Structure

Fig. 1 shows the feedback channel structure, which consists of minislots. The number of
minislots constructing the feedback channel can vary from 1 minislot, in the case the first
user is granted the channel access, to 2K − 1 minislots, where K is the total number of users,
in the case all users feed back their channel qualities. This variation in length assumes that
data transmission can begin at different time periods. Such assumption can be possible if
dynamic spectrum access is implemented. The JPCF scheduling algorithm assumes that the
feedback is polling-based or contention-based. The feedback channel is divided into polling-
based feedback minislots (p-minislot) and contention-based feedback minislots (c-minislot),
where each p-minislot is followed by a c-minislot, except the last p-minislot. The contention
protocol is based on a modified slotted ALOHA protocol (23). As any random multiple access
protocol one out of three possible outcomes will occur: no access, one access, and a collision.
Multiple users feeding back the requested information will cause signals to interfere at the
receiver which we consider as a collision. When a collision occurs no resolution is consider
and the information is discarded.
3. Scheduling Algorithm
In this section, we will introduce our proposed algorithm, the JPCF scheduling algorithm.
Similar to the optimal selective scheduling algorithm, the JPCF algorithm always guarantees
that the best user is granted the channel access. The difference between the two is in the se-
lection process. In the JPCF algorithm decisions on whom to schedule are based on threshold
test. In simple terms, the BS compares the channel quality of a user to a predetermined thresh-
old value and selects the user if his channel quality exceeds the threshold. In case no user is
found with channel quality exceeding the threshold, then the user with the highest channel
quality among all is selected.
Let us define the following:
• L = {( D (i ), l ) : 1 ≤ l ≤ K is the probing order , and D (i ) is the ID of user i }. The set is
generated randomly each time slot.
(N) (N)
• The predetermined threshold value is set to γth . According to (3), γth is the SNR
threshold of the highest modulation level and q( N ) is the quantized threshold value.
• The probability of contention in the c-minislot is ρ.
The feedback mechanism of the JPCF algorithm is based on both polling period and contention
period. Each period is a minislot during. Fig. 1 shows the feedback channel structure, which
consists of p-minislots and c-minislots. The following describes the scheduling algorithm (Fig.
2 shows the flowchart):
(i) Through the broadcast channel all users know their polling order (L).
(ii) The feedback channel always begins with a p-minislot.
(iii) During the p-minislot the BS polls the channel state information of the user (his SNR)
who is in order.
(iv) The feedback channel is terminated and the data transmission begins if the polled user
has ql = q( N ) . By knowing (γl ) and according to (4) ql is determined.
(v) If ql = q( N ) , then his information is stored and the feedback channel continues with a
c-minislot following the previous p-minislot and step (vi) is performed.
(vi) During the c-minislot, which is a contention based minislot, users feedback their SNRs,
with probability ρ, if it satisfy q = q( N ) . Otherwise, they keep silent.
(vii) If the contention is successful, then the feedback channel is terminated and the data
transmission begins for that successful user. Otherwise, the feedback channel continues
with a p-minislot following the previous c-minislot and step (iii) is performed.
(viii) In case all users are polled, the BS picks the user with the highest channel quality among
them and the data transmission begins for that user.
(ix) In case a tie occurs, a random pick is performed.
The performance of the JPCF algorithm is evaluated in this section. Basically, we concentrate
on the feedback load, the system capacity, and the scheduling delay. Closed form expressions
for all three performance measures are also derived.
4.1 Feedback Load

We define the average feedback load (AFL) as the average number of responses until a user is
scheduled. The feedback response includes the response on the p-minislot, and the response
on the c-minislot. In the c-minislot, a contention resulting a success or a collision is counted
as one response, and zero response is considered in case no contention occurs.
Consider two consecutive minislots (p-minislot and c-minislot) and denote the discrete ran-
dom variable η ∈ [1, 2] by the total number of feedback during them. Let U be the set of events
where a successful search occurs during the two minislots. Note that a successful search occurs
when a user with q = q( N ) is found. Therefore, the occurrence of a successful search is associ-
ated with one of the following events: (i) The polled user during the p-minislot has q = q( N ) ,
or (ii) The polled user during the p-minislot has q < q( N ) and a successful contention occurs
in the following c-minislot.
The probability of U given l probes is:
i K − l −i
K −l
U (N) (N) K−l (N) (N)
Γ (l, ρ) = 1 − Fγ (γth ) + Fγ (γth ) ∑ 1 − Fγ (γth ) Fγ (γth )
i =1
i

× iρ(1 − ρ)i−1 ,
(5)
where
−ζ
Fγ (ζ ) = P[γ < ζ ] = (1 − e γ̄ ), (6)
is the cumulative distribution function (CDF) of the SNR (γ). The first part of the right hand
side of (5) refers to the success in the p-minislot and the second part refers in the success in
the c-minislot.
let X represent the set of events where no feedback occurs on the c-minislot and Y to be the set
of events where a feedback occurs on the c-minislot. Conditioning on l, the probability of X
and the probability of Y, respectively, are:
i K − l −i
K −l
X (N) (N) K−l (N) (N)
Γ (l, ρ) = 1 − Fγ (γth ) + Fγ (γth ) ∑ 1 − Fγ (γth ) Fγ (γth )
i =0
i

× (1 − ρ ) i ,
(7)
and
i K − l −i
K −l
Y (N) K−l (N) (N) i
Γ (l, ρ) = Fγ (γth ) ∑ 1− Fγ (γth ) Fγ (γth ) 1 − (1 − ρ ) . (8)
i =0
i
The conditioned expected value of η is:
η (l, ρ) = 1 · ΓX (l, ρ) + 2 · ΓY (l, ρ). (9)

Therefore, the AFL is:

K −1 l l −1
AFL(ρ) = ∑ ∑ η (i, ρ) ∏ 1 − ΓU (i, ρ) ΓU (l, ρ)
l =1 i =1 i =1
(10)
K −1 K −1
U
+ ∑ η (l, ρ) +1 ∏ 1 − Γ (l, ρ) .
l =1 l =1
4.2 System Capacity

The average spectral efficiency (ASE) is defined as the average transmitted data rate per unit
bandwidth in bits/sec/Hz for specified power and target error performance. In this work
we are considering discrete rates and using quantized SNR values (q). For example: if two
(i ) ( i +1) (i ) ( i +1)
users with γ1 ∈ [γth , γth ), γ2 ∈ [γth , γth ), and γ1 = γ2 , then according to (4) both
users feedback q(i) . Therefore, scheduling either one will result a transmission rate of log2 Mi
bps/Hz.
Based on the algorithm’s description in Section 3, it is clearly seen that both the DSMUDiv
scheduling algorithm proposed in (6) and the JPCF scheduling algorithm will always select
the best user. Therefore, resulting a similar performance in terms of spectral efficiency.
It has been shown in (6) that with i users in the system, scheduling the best user yields the
following average spectral efficiency:

(1)
R(i ) = bo [ Fγ (γth )]i

N −1
( n +1) (n)
+ ∑ bn [ Fγ (γth )]i − [ Fγ (γth )]i (11)
n =1

(N)
+ bN 1 − [ Fγ (γth )]i
where bn = log2 Mn is the number of bits per constellation.

Therefore, the ASE of the JPCF algorithm is:
ASE = R(K ). (12)
In the above expression (12) it is assumed that the guard time duration is negligible, meaning
that the feedback rate will not degrade the total system spectral efficiency. Practically, this
is not valid. Therefore, the amount of bits transmitted as feedback has to be counted and
at the end it will influence the performance of the system. To look into this issue, which is
spectral efficiency degradation caused by the feedback traffic, we define a the system capacity
as [bits/channel use]. Assuming N modulation levels, then we need log2 N bits to represent
them. Therefore, the system capacity is:
AFL · log2 N
Csys = R(K ) − , (13)
S
where S is the number of symbols transmitted in the data transmission time slot.
In (13), the last term takes into account the amount of bits transmitted as feedback. Also, in the
same expression the time delay is not taken into account, which is the guard time duration.
The only consideration is the feedback rate, or the amount of bits transmitted as feedback. In
the next section we investigate the effect of delay on the system performance.
4.3 Scheduling Delay

In this section we investigate the scheduling delay and its effect on the system performance.
This time delay is part of the system resources and it is important to identify the amount of
resources consumed when performing the JPCF algorithm.
4.3.1 Guard time

The scheduling process will take place during the gaud time (τg ), which is between bursts.
The delay resulted from this scheduling process is measured as the time needed to schedule a
user, which is simply the time duration of the minislots used (idle minislots are counted) until
data transmission is allowed. By looking at Fig. 1, we can see that τ f ≤ τg ≤ (2K − 1)τ f . For
simplicity, we assume that both p-minislot and c-minislot have the same time length (τ f ).
Assuming a successful search at the lth p-minislot, then the guard time is:
τg (l ) = τ f (2l − 1), (14)
with probability:
l −1
( p) U (N)
P (l ) = ∏ 1 − Γ (i, ρ) · 1 − Fγ (γth ) . (15)
i =1
On the other hand, if the successful search occurs at the lth c-minislot, then the guard time is:
τg (l ) = 2lτ f , (16)
with probability:

l −1
(N)
P(c) (l ) = ∏ 1 − ΓU (i, ρ) · ΓU (l, ρ) − 1 − Fγ (γth ) . (17)
i =1
Therefore, the average guard time is:

K −1
τ= ∑ τ f (2l − 1) P( p) (l ) + 2lτ f P(c) (l )
l =1
K −1 (18)
+ (2K − 1)τ f · ∏ 1 − ΓU (i, ρ) .
i =1
4.3.2 System throughput

The system throughput (STH) is the amount of data bits transmitted per time, where this
time includes the data transmission time (Td ) and the guard time (τg ). In (12), we looked at
the amount of bits per data transmission time, where the effect of the scheduling delay was
not included. To have a better insight, we derive the average system throughput (ASTH) by
taking into account the effect of the guard time duration.
The average system throughput is:

K −1 Td − τ f (2l − 1) Td − 2lτ f
ASTH = ∑ P( p) (l ) + P ( c ) ( l ) · R( K )
l =1
Td Td
(19)
Td − (2K − 1)τ f K −1
+ · R(K ) · ∏ 1 − ΓU (i, ρ) .
Td i =1
4.4 Parameters Optimizations

The algorithm’s objective is to strictly schedule the best user, therefore the search process will
(N)
last until a user with q = q( N ) is found, which depends on γth and ρ. In this section we
investigate the optimization of ρ with the objective to minimize the feedback load:
(N)
{ρ} =arg min AFL(γth , ρ, K ),
{ρ} (20)
subject to 0 ≤ ρ ≤ 1.
Similarly, the optimization solution is well defined by the equation:

∂AFL(ρ,K )
= 0,
∂ρ (21)
0 ≤ ρ ≤ 1.
Note that the optimization may not be convex. The derivation of (10) is involved so we apply
an exhaustive search method to find an optimal value. Table 3 shows the optimal values with
different K given the parameters in Section 5. It is clearly seen from Table 3 that for a given K,
ρ has two optimal values at the two average SNR regions. This is due to the collision. At low
average SNR values, the value of ρ is increased to encourage good users to compete, whereas,
at higher average SNR values the value of ρ is decreased to lower the possibility of collision.
In this section we elaborate on the the performance of the JPCF algorithm by presenting some
numerical examples. We compare its performance with both the DSMUDiv and the optimal
selective diversity scheduling schemes. Parameters values are found in Tables 1 and 2. Fig. 3
shows the normalized average feedback load (i.e., the average feedback load divided by the
number of users K) of the JPCF algorithm for different values of ρ, which is the probability
of contention. As the value of the average SNR changes from low to high, the optimal value
of ρ, at which the feedback is minimized, changes from ρ = 1 to ρ = 0.13, for a given value
of K. Table 3 shows the optimal values of ρ for different values of K. The reason for the
change of the optimal value is the threshold value. For a given threshold value, the percent of
users with channel quality exceeding it increases as the average SNR increases, which leads
to higher probability of collision. As more users contend the feedback load increases, which
forces the value of ρ to change to maintain the minimization in (20). The point at which this
(N)
change happens depend on the value of γth and the BEPo .
The comparison of the JPCF algorithm with both the DSMUDiv and the optimal selective di-
versity scheduling algorithms is presented in Fig. 4. In the figure, although the DSMUDiv
algorithm has decreased the feedback load compared to the optimal algorithm, the JPCF al-
gorithm has decreased the feedback load even more. This extra reduction comes from the
introduction of the contention-based feedback (c-minislot), where not all users need to be
polled as opposed to the DSMUDiv algorithm, which is a pure polling-based feedback algo-
rithm. In terms of spectral efficiency, as shown in Section 4.2, both the JPCF algorithm and the
DSMUDiv algorithm have the same spectral efficiency (here the feedback rate is not included).
In (6), it has been proven that the DSMUDiv algorithm maintain the performance of the full
feedback algorithm in terms of spectral efficiency, which means that the JPCF algorithm also
maintain the same performance. The difference will come when you include the feedback rate
(log2 N), which is defined as [bits/channel use]. From (13), we can see that as the feedback
load increases the system capacity decreases, therefore, we deduce that the JPCF algorithm
has the highest system capacity compared to the other two algorithms, which can be seen in
Fig. 5 and 6. As a benchmark we include the average spectral efficiency (ASE) that does not
include the feedback rate. The gap between the system capacity and the ASE shrinks with the
increase of number of symbols transmitted in the data transmission time slot.
Although the number of minislots of the JPCF algorithm varies from 1 to 2K − 1, which is
much greater than the DSMUDiv algorithm, where the minislots varies from 1 to K, the aver-
age guard time of the JPCF algorithm is much smaller than the guard time of the DSMUDiv
algorithm as depicted in Fig. 7. Such advantage occurs at medium to high average SNR. As
the average SNR decreases this advantage diminishes as seen in Fig. 8. The reason is that the
JPCF algorithm has contention minislots which creates an additional delay on top of the delay
created by the polling minislots. This delay increase as the probability of finding a user with
channel gain exceeding the threshold value decreases, which is the case for low average SNR
values. Fig. 9 shows the effect of different data transmission time durations on the system
throughput. Obviously, higher values give better performance.
6. Conclusions
In this chapter we proposed a scheduling algorithm that maximizes the spectral efficiency
while reducing the feedback load. The algorithm, called joint polling and contention based
feedback (JPCF) algorithm, collects channel quality information of the users either in a polling
form or in a contention form. Compared to the optimal (full feedback) algorithm, the JPCF
algorithm has a similar spectral efficiency and a higher system capacity, which takes into ac-
count the effect of the feedback rate. Also, the JPCF algorithm shows more reduction in feed-
back load compared to the DSMUDiv algorithm. One drawback of the JPCF algorithm is the
high delay compared to the DSMUDiv algorithm as the average SNR decreases, which affects
the performance. As the average SNR increases, the delay encountered when using the JPCF
algorithm drops below the delay created by using the DSMUDiv algorithm, which improves
the system performance. The work presented in this paper includes analysis of the JPCF algo-
rithm under slow Rayleigh fading assumption. Closed-form expressions of the feedback load,
system capacity and scheduling delay are also presented in this paper.
Modulation Level Switching Threshold (dB)

(1)
BPSK γth = 4.8
(2)
4-QAM γth = 7.8
(3)
16-QAM γth = 15
(4)
64-QAM γth = 20
Table 1. A List of Selected Modulation Levels (BEPo = 10−2 )
Parameter Value
N 4 modulations
K 30 users
τf 154 µsec (based on reference (24))
Td 5 msec (based on reference (24))
Table 2. A List of Parameters
K=5 K=30 K=50 K=100

ρ = 0.2, γ > 19 dB ρ = 0.13, γ > 15 dB ρ = 0.12, γ > 15 dB ρ = 0.1, γ > 14 dB
ρ = 1, γ ≤ 19 dB ρ = 1, γ ≤ 15 dB ρ = 1, γ ≤ 15 dB ρ = 1, γ ≤ 14 dB
Table 3. The Parameters Optimizations
Fig. 1. The framing structure of a TDD system. Each polling-based feedback minislot (p-minislot) is
followed by a contention-based feedback minislot (c-minislot), where they carry the feedback information.
Fig. 2. A flowchart of the JPCF algorithm.

Evaluating if a mobile goes through a critical period should not only focus on the classical
mean delay and jitter analysis. Indeed, a meaningful constraint regarding delay is the
limitation of the occurrences of large values. Accordingly, (Gueguen & Baey (c), 2008)
defines the concept of delay outage by analogy with the concept of outage used in system
coverage planning. A mobile k is in delay outage (in critical period) when its packets
experience a delay greater than a given threshold T k defined by the mobile application
requirements. The delay experienced by each mobile is tracked all along the lifetime of its
connection. At each transmission of a packet of mobile k, the ratio of the total number of
packets whose delay exceeded the threshold divided by the total number of packets
transmitted since the beginning of the connection is computed. The result is called Packet
Delay Outage Ratio (PDOR) of mobile k and is denoted PDOR k. This measure is
representative of the emergency for the mobile k to be served. Fig. 3 illustrates an example
cumulative distribution of the packet delay of a mobile at a given time instant. A mobile can
be considered as satisfied when, at the end of its connection, its delay constraint is met, i.e.
its experienced PDOR is less than the application specific PDOR target.
In WFO scheduling, the required QoS, the experienced QoS and the transmission conditions
are jointly considered in an extended cross-layer approach. The scheduling principle is to
allocate a Resource Unit n to the mobile jwhich has the greatest WFO parameter value
WFO k,n such as:
j  arg max k (WFOk ,n ), k  1,..., K , (3)
whereWFO k,n is defined by:
WFOk ,n  mk ,n  f ( PDORk ), (4)
with f a strictly increasing polynomial function (Gueguen & Baey (d), 2008; Gueguen &
Baey, 2009):
f ( x)  1    x  , (5)
The exponent parameter ´ allows being sensitive and reactive to PDOR fluctuations which
guarantees fairness at a short time scale.  is a normalization parameter that ensures that
f(PDOR k) and m k,n are in the same order of magnitude.
With this scheduling, physical layer information (represented through the factor m k,n) are
used in order to take advantage of the time, frequency and multiuser diversity and
maximize the system capacity. Higher layer information (represented through the factor
f(PDOR k) are exploited in order to introduce dyna mic priorities between flows for ensuring
the same QoS level to all mobiles. With this original weighted system that introduces
dynamic priorities between the flows, WFO keeps a maximum number of flows active
across time but with relatively low traffic backlogs. This results in a well-balanced resource
allocation. Preserving the multiuser diversity allows to continuously take a maximal benefit
6
Average spectral efficiency (proposed algorithm)
Proposed scheduling algorithm
DSMUDiv scheduling algorithm
5.5 Optimal selective diversity scheduling algorithm
Round Robin scheduling scheme algorithm
5
System Capacity [bits/channel use]
4.5
3.5
2.5
5 10 15 20 25 30 35 40 45 50
Number of users (K)
Fig. 5. System capacity of: (i) the proposed (JPCF) scheduling algorithm, (ii) the DSMUDiv scheduling
algorithm, and (iii) the optimal (full feedback) scheduling algorithm. Setting ρ = 1, 100 symbols are
transmitted, and γ = 15dB.
6
Average spectral efficiency (proposed algorithm)
Proposed scheduling algorithm
5.5 Optimal selective diversity scheduling algorithm
Round Robin scheduling scheme algorithm
5
System Capacity [bits/channel use]
4.5
3.5
2.5
5 10 15 20 25 30 35 40 45 50
Number of users (K)
Fig. 6. System capacity of: (i) the proposed (JPCF) scheduling algorithm, (ii) the DSMUDiv scheduling
algorithm, and (iii) the optimal (full feedback) scheduling algorithm. Setting ρ = 1, 500 symbols are
transmitted, and γ = 15dB.
3.5
JPCF scheduling algorithm
3
Average guard time [milliseconds]
2.5
1.5
0.5
10 15 20 25 30 35 40 45 50
Number of users (K)
Fig. 7. Average guard time of: (i) the JPCF scheduling algorithm, and (ii) the DSMUDiv scheduling
algorithm. Setting γ = 15dB.
16
JPCF scheduling algorithm
14
12
Average guard time [milliseconds]
10
0
10 15 20 25 30 35 40 45 50
Number of users (K)
Fig. 8. Average guard time of: (i) the JPCF scheduling algorithm, and (ii) the DSMUDiv scheduling
algorithm. Setting γ = 5dB.
0.95
System throughput as percentage of the ASE
0.9
0.85
0.8
Td = 5 msec
Td = 50 msec
0.75
10 15 20 25 30 35 40 45 50
Number of users (K)
Fig. 9. System throughput as percentage of the average spectral efficiency of the JPCF scheduling algo-
rithm. Setting γ = 15dB, and 1 symbol is transmitted.
7. References
[1] P. Viswanath, D. N. C. Tse, and R. Laroia, “Opportunistic beamforming using dumb
antenna,” IEEE Transactions on Information Theory, vol. 48, no. 6, pp. 1277–1294, June
2002.
[2] R. Knopp and P. Humblet, “Information capacity and power control in single-cell mul-
tiuser communications,” in Proc. IEEE International Communication Conference (ICC’95),
Paris, France, June 1995, pp. 331–335.
[3] D. N. C. Tse, “Optimal power allocation over parallel Gaussian broadcast channels,” in
Proc. IEEE International Symposium Information Theory (ISIT’97), Ulm, Germany, June 1997,
p. 27.
[4] P. Bender, P. Black, M. Grob, R. Padovani, N. Sindhushyana, and S. Viterbi,
“CDMA/HDR: A bandwidth efficient high speed wireless data service for nomadic
users,” IEEE Communication Magazine, vol. 38, no. 7, pp. 70–77, July 2000.
[5] F. Floren, O. Edfors, and B. A Molin, “The effect of feedback quantization on the through-
put of a multiuser diversity scheme,” in Proc. IEEE Global Telecommunication Conference
(GLOBECOM’03), San Francisco, California, December 2003, pp. 497–501.
[6] Y. S. Al-Harthi, A. Tewfik, and M. S. Alouini, “Multiuser diversity with quantized feed-
back,” IEEE Transactions on Wireless Communication, vol. 6, no. 1, pp. 330–337, January
2007.
[7] Y. Al-Harthi, A. Tewfik, and M. S. Alouini, “Multiuser diversity-enhanced equal access
with quantized feedback in multicarrier OFDM systems,” in Proc. IEEE Vehicular Technol-
ogy Conference (VTC-Fall’05), Dallas, TX, September 2005, pp. 568–572.
[8] M. A. Regaeig, N. Hamdi, and M. S. Alouini, “Switched-based reduced feedback OFDM
multi-user opportunistic scheduling,” in Proc. IEEE International Symposium on Personal
and Indoor and Mobile Radio Communication (PIMRC ’05), Berlin, Germany, September
2005, pp. 2495–2499.
[9] D. Gesbert and M. S. Alouini, “How much feedback is multi-user diversity really
worth?,” in Proc. IEEE International Communication Conference (ICC’04), Paris, France,
June 2004, pp. 234–238.
[10] V. Hassel, M. S. Alouini, G. E. Øien, and D. Gesbert, “Rate-optimal multiuser scheduling
with reduced feedback load and analysis of delay effects,” EURASIP Journal on Wireless
Communications and Networking, vol. 2006, no. 36424, pp. 7, 2006.
[11] V. Hassel, M.-S Alouini, D. Gesbert, and G. E. Øien, “Exploiting multiuser diversity
using multiple feedback thresholds,” in Proc. IEEE Vehicular Technology Conference (VTC-
Spring’05), Stockholm, Sweden, 30 May-1 June 2005, pp. 1302–1306.
[12] B. Holter, M. S. Alouini, G. E. Øien, and H-C. Yang, “Multiuser switched diversity trans-
mission,” in Proc. IEEE Vehicular Technology Conference (VTC-Fall’04), Los Angeles, CA,
September 2004, pp. 2038–2043.
[13] Y. Xue and T. Kaiser, “Exploiting multiuser diversity with imperfect one-bit channel state
feedback,” IEEE Transactions on Vehicular Technology, vol. 56, no. 1, pp. 183–193, January
2007.
[14] S. Sanayei and A. Nosratinia, “Exploiting multiuser diversity with only 1-bit feedback,”
in Proc. IEEE Wireless Communications and Networking Conference (WCNC’05), New Or-
leans, LA, March 2005, pp. 978–983.
[15] P. Svedman, S. K. Wilson, Jr. L. J. Cimini, and B. Ottersten, “A simplified opportunistic
feedback and scheduling scheme for OFDM,” in Proc. IEEE Vehicular Technology Confer-
ence (VTC-Spring’04), Milan, Italy, May 2004, pp. 1878–1882.
[16] Z. H. Han and Y. H. Lee, “Opportunistic scheduling with partial channel information in
OFDMA/FDD systems,” in Proc. IEEE Vehicular Technology Conference (VTC-Fall’04), Los
Angeles, CA, September 2004, pp. 511–514.
[17] K. Bai and Junshan Zhang, “Opportunistic multichannel Aloha: distributed multiaccess
control scheme for OFDMA wireless networks,” IEEE Transactions on Vehicular Technol-
ogy, vol. 55, no. 3, pp. 848–855, May 2006.
[18] T. Tang and R. W. Heath, “Opportunistic feedback for downlink multiuser diversity,”
IEEE Communications Letters, vol. 9, no. 10, pp. 948–950, October 2005.
[19] X. Qin and R. Berry, “Opportunistic splitting algorithms for wireless networks.,” in Proc.
IEEE Conference on Computer Communications (INFOCOM’04), Hong Kong, China, March
2004, pp. 1662–1672.
[20] H. Koubaa, V. Hassel, and G. E. Øien, “Contention-less feedback for multiuser diver-
sity scheduling,” in Proc. IEEE Vehicular Technology Conference (VTC-Fall’05), Dallas, TX,
September 2005, vol. 3, pp. 1574–1578.
[21] S. Patil and G. de Veciana, “Reducing feedback for opportunistic scheduling in wireless
systems,” IEEE Transactions on Wireless Communications, vol. 6, no. 12, pp. 4227–4232,
December 2007.
[22] M. S. Alouini and A. J. Goldsmith, “Adaptive modulation over Nakagami fading chan-
nels,” Kluwer Journal on Wireless Communications, vol. 13, pp. 119–143, May 2000.
[23] D. Bertsekas and R. Gallager, Data Networks, Prentice Hall, 1992.
[24] IEEE Standards Department, “IEEE Std 802.11. part 11: Wireless LAN medium access
control (MAC) and physical layer (PHY) specifications,” Technical Report, IEEE, NJ,
September 1999.
Fig. 11. Analysis of the respect of QoS constraints for different targeted QoS (Mobile
dissatisfaction if PDORtarget = 5 % on the left, if PDORtarget = 10 % on the right).
Highly unfair, MaxSNR fully satisfies the required QoS of close mobiles at the expense of
the satisfaction of far mobiles. Indeed, only 54.5 percents of these latter experience a final
PDOR inferior to a PDOR target of 5 % (cf. Fig. 9a). Unnecessary priorities are given to close
mobiles who easily respect their QoS constraints while more attention should be given to the
farther. This inadequate priority management dramatically increases the global mobile
dissatisfaction, which reaches 23 % as shown on Fig. 9a and Fig. 10 (on the left).
PF brings more fairness and allocates more priority to far mobiles. Compared to MaxSNR,
PF offers a QoS support improvement with only 12.8 % of dissatisfied mobiles (cf. Fig. 9b
and Fig. 10 (on the right)). Fairness is still not total since the farther mobiles have a lower
spectral efficiency than the closer ones due to pathloss. All mobiles do not all benefit of an
equal average throughput despite they all obtain an equal share of bandwidth. This induces
heterogeneous delays and unequal QoS. This fairness improvement compared to MaxSNR
indicates however that some flows can be slightly delayed to the benefit of others without
significantly affecting their QoS.
WFO was built on this idea. The easy satisfaction of close mobiles (with better spectral
efficiency) offers a degree of freedom which ideally should be exploited in order to help the
farther ones. WFO allocates to each mobile the accurate share of bandwidth required for the
satisfaction of its QoS constraints, whatever its position. With WFO, only 0.8 percents of the
mobiles are dissatisfied (cf. Fig. 9c and Fig. 10 (on the left)). Additionally, compared to Fig.
9a and Fig. 9b, Fig. 9c exhibits superimposed curves, which prove the WFO high fairness,
included at short term.
Fig. 10 shows that WFO brings the largest level of satisfaction. Indeed, for a tight PDOR
target of 5 % (see on the left), the dissatisfaction ratio with a high traffic load of 1120 Kbps is
equal to 18 % with WFO versus 29.7 % with PF, the best of the other scheduling schemes. If
we set the PDOR target to 10 %, the dissatisfaction ratio with a high traffic load of 1120 Kbps
is 0 % with WFO versus 13.8 % with the best of the other scheduling schemes (PF).
Opportunistic Access Schemes for Multiuser OFDM Wireless Networks 277
Fig. 12 a. Spectral efficiency. Fig. 12 b. Multiuser diversity

Fig. 12. Bandwidth usage efficiency.
We finally studied the system capacity offered by the four scheduling algorithms. Fig. 11a
shows the average number of bits carried on a used subcarrier by each tested scheduler
under various traffic loads. As expected, the non opportunistic Round Robin scheduling
provides a constant spectral efficiency, i.e. an equal bit rate per subcarrier whatever the
traffic load since it does not take advantage of the multiuser diversity. The three other tested
schedulers show better results. In contrast with RR, with the opportunistic schedulers
(MaxSNR, PF, WFO), we observe an interesting inflection of the spectral efficiency curve
when the traffic load increases. The join analysis of Fig. 11a and Fig. 11b shows that the
spectral efficiency of opportunistic scheduling is an increasing function of the number of
active mobiles, thanks to the exploitation of this supplementary multiuser diversity.
Consequently MaxSNR, PF and WFO increase their spectral efficiency with the traffic load
and the system capacity is highly extended compared to networks which use classical
scheduling algorithms. With these three schedulers, all mobiles are served even at the
highest traffic load of 1280 Kbps.
The performance of the four schedulers can be further qualified by computing the
theoretical maximal system throughput. Considering the Rayleigh distribution, it can be
noticed that ´ 2 k,n is greater or equal to 8 with a probability of only 0.002. In these ideal
situations, close mobiles can transmit/receive 6 bits per RU while far mobiles may
transmit/receive 4 bits per RU. If the scheduler always allocated the RUs to the mobiles in
these ideal situations, an overall efficiency of 5 bits per RU would be obtained which yields
a theoretical maximal system throughput of 1600 Kbps. Comparing this value to the highest
traffic load in Fig. 11a (1280 Kbps) further demonstrates the good efficiency obtained with
the opportunistic schedulers that nearly always serve the mobiles when their channel
conditions are very good. This result also shows that the WFO scheduling has slightly better
performances than the two other opportunistic schedulers. Keeping more mobiles active (cf.
Fig.11b) but with a relatively lower traffic backlog (cf. Fig.8a), the WFO scheme preserves
multiuser diversity and takes more advantage of it obtaining a slightly higher bit rate per
subcarrier (cf. Fig. 11a).
c d ac RU a d c d d .I
c b a , ac b d c a c bac a d
a c a Q S a a d a . T c
b a a db b ab c c .T a c
a c a d c c a b b a d acc
a d c b d (Va d B a ., 1999). T , ac a a a
a b d c a . Add a a b a a b d
a .
)UHTXHQF\
&KDQQHOVWDWHXSGDWHV7KURXJKSXW
5HVRXUFH8QLW
UHIUHVKLQJ
7LPHVORW
6XEFDUULHU
0RELOH$
%DQGZLGWK 0RELOH%
’RZQOLQN ’RZQOLQNGDWD 8SOLQNGDWD

8SOLQNFRQWHQWLRQ
IHHGEDFNVXEIUDPH VXEIUDPH VXEIUDPH VXEIUDPH
)UDPHQ
F . 2. F a c TDD d .
5EJGFWNKPI6GEJPKSWGUKP1(&/9KTGNGUU0GVYQTMU
T MAC c cd ca a a a a d
a d d d ca a a c . T c a acc
d R d R b (RR) a d Ra d Acc (RA) a ada d
a d d .M c ac
a b d c c d OFDM ba d a d
ca c c d c ab a ca c ac
b ( ) a ab c a c d a a . T c d
a b a d c d d a
. A c d d
a c .T a c d c a d: Ma S a - -N
Ra (Ma SNR), P a Fa (PF) a d c , W d Fa O c
(WFO).
across time but with relatively low traffic backlogs, WFO is designed for best profiting of the
multi-user diversity taking advantage of the dynamics of the multiplexed traffics.
Preserving the multiuser diversity, WFO takes a maximal benefit of the opportunistic
scheduling technique and maximizes the system capacity. Additionally, this also achieves a
time uniform fair allocation of the resource units to the flows ensuring short term fairness.
This higher layers/MAC/PHY cross-layer approach better conceals the system capacity
maximization, fairness objectives and the full support of multimedia services with adequate
QoS.
6. References
Hoymann, C. (2005). Analysis and performance evaluation of the OFDM-based
metropolitan area network IEEE 802.16. Computer Networks, ol. 49,
V No. 3, pp. 341-
363, ISSN: 1389-1286
Andrews, M.; Kumaran, K.; Ramanan, K. Stolvar, A. & whiting P. (2001). Providing quality
of service over a shared wireless link. IEEECommunicationsMagazine , Vol. 39, No.2,
pp. 150-154, ISSN: 0163-6804
Van de Beek, J.-J. ; Borjesson, P.O. ; Boucheret, M.-J ; Landstrom, D. ; Arenas, J.-M. Odling,
P.; Ostberg, C.; Wahlgvist, M. & Wilson, S.K. (1999). A time and frequency
synchronization scheme for multiuser OFDM. IEEE J.Sel. Areas Commun , Vol.17,
No.11, pp 1900-1914, ISSN: 0733-8716
Li, Y.G; Seshadri, N. & Ariyavisitakul, S. (1999). Channel estimation for ofdm systems with
transmitter diversity in mobile wireless channels. IEEE J.Sel. Areas Commun , Vol.17,
No.3, pp 461-471, ISSN: 0733-8716
Truman, T.E. & Brodersen, R.W (1997). A measurement-based characterization of the time
variation of an indoor wireless channel, proceedings of Int. Universal Personal
Communications Record (ICUPC) , pp. 25-32, ISBN: 0-7803-3777-8, San Diego, CA,
USA, October 1997
Knopp, R. & Humblet, P. (1995). Choi, J. (1996). Information capacity and power control in
single-cell multiuser communications, proceedings of IEEE Conference on
Communications,(ICC) pp 331-335, ISBN: 0-7803-2486-2, Seattle, WA, USA, june 1995
Wong, C.Y.; Cheng, R.S.; Lataief, K.B. & Murch, R.D. (1999). Multiuser OFDM with
adaptative subcarrier, bit and power allocation. IEEE J.Sel. Areas Commun , Vol.17,
No.10, pp 1747-1758, ISSN: 0733-8716
Wang, X. & Xiang, Y. (2006). An OFDM-TDMA/SA MAC protocol with QoS constaints for
broadband wireless LANs. ACM/Springer Wireless Networks , Vol. 12, No. 2, pp 159-
170, ISSN: 1022-0038
Viswanath, P.; Tse, D.N.C. & Laroia, R. (2002). Opportunistic beamforming using dumb
antennas. IEEE Transactions on Information , Vol. 48, No.6, pp. 1277–1294,
Theory
ISSN: 0018-9448
Kim, H.; Kim, K.; Han, Y. & Lee, J. (2002). An efficient scheduling algorithm for QoS in
wireless packet data transmission, proceedings of IEEE Int. Symposium on Personal,
Indoor and Mobile Radio Communications , ppPIMRC)
( 2244-2248, ISBN: 0-7803-7589-0,
Lisboa, Portugal, September 2002
Anchun, W.; Liang, X.; Xjiin, S.X. & Yan, Y. (2003). Dynamic resource management in the
fourth generation wireless systems, proceedings of IEEE Int. Conference on
Communication Technology (ICCT) pp , 1095-1098, ISBN: 7-5635-0686-1, Beijing,
China, April 2003
Svedman, P.; Wilson, K. & Ottersen, B. (2004). A QoS-aware proportional fair scheduler for
opportunistic OFDM, proceedings of IEEE Int. Vehicular Technology Conference , VTC)
(
pp 558-562, ISBN: 0-7803-8521-7, Los angeles, CA, USA, September 2004
Kim, H.; Kim, K.; Han, Y. & Yun, S. (2004). A proportional fair scheduling for multicarrier
transmission systems, proceedings of IEEEInt.Vehicu larTechnologyConference , (VTC)
pp. 409-413, ISBN: 0-7803-8521-7, Los angeles, CA, USA, September 2004
Choi, J.-G. & Bahk, S. (2007). Cell-throughput analysis of the proportional fair scheduler in
the single-cell environment. IEEETransactionsonVehicularTechnology , vol. 56, No.2,
pp. 766-778, ISSN: 0018-9545
Gueguen, C. & Baey, S. (2008). Compensated proportional fair scheduling in multiuser
OFDM wireless networks, proceedings of IEEE Wireless and Mobile Computing,
Networking and Communications (WIMOB), pp119-125, ISBN: 978-0-7695-3393-3,
Avignon, France, October 2008
Holtzman, J. (2001). Asymptotic analysis of proportional fair algorithm, proceedings of IEEE
Int.SymposiumonPersonal,Indoor dMobile
an RadioCommunicationsPIMRC) ( , pp. 33-
37, ISBN: 0-7803-7244-1, San Diego, CA, USA, October 2001
Gueguen, C. & Baey, S. (2008). Weighted fair opportunistic scheduling for multimedia QoS
support in multiuser OFDM wireless networks, proceedings of IEEE Int. Symposium
on Personal, Indoor and Mobile Radio Communications , pp. 1-6,
( ISBN: 978-1-
PIMRC)
4244-2643-0, Cannes, France, September 2008
Gueguen, C. & Baey, S. (2008). An efficient and fair scheduling scheme for multiuser OFDM
wireless networks, proceedings of IEEE Int.Wireless Communications and Networki
Conference WCNC) ( , pp. 1610-1615, ISBN: 978-1-4244-1997-5, Las Vegas, NV, USA,
April 2008
Gueguen, C. & Baey, S. (2008). Scheduling in OFDM wireless networks without tradeoff
between fairness and throughput, proceedings of IEEE Int. Vehicular Technology
Conference VTC) ( , pp. 1-5, ISBN: 978-1-4244-1721-6, Calgary, Canada, September
2008
Gueguen, C. & Baey, S. (2009). A Fair Opportunistic Access Scheme for Multiuser OFDM
Wireless Networks. EURASIP Journal on Wireless Communications and Networking.
Special issue on Fairness" in Radio Resource Management for Wireless Net
Volume 2009 (2009), Article ID 726495, pp. 70-83
Parsons, J.D (1992). TheMobileRadioPropagationChannel, Wiley, ISBN: 978-0-471-98857-1
Baey, S. (2004). Modeling MPEG4 video traffic based on a customization of the DBMAP,
proceedings of Int. Symposium on Performance Evaluation of Computer
Telecommunication Systems (SPECTS) pp. , 705-714, ISBN: 1-56555-284-9, San Jose,
California, USA, July 2004
Brady, P. (1969). A model for generating on-off speech patterns in two-conversation. Bell
SystemTechnicalJournal , vol. 48, No.1, pp. 2445-2472
Evaluating if a mobile goes through a critical period should not only focus on the classical
mean delay and jitter analysis. Indeed, a meaningful constraint regarding delay is the
limitation of the occurrences of large values. Accordingly, (Gueguen & Baey (c), 2008)
defines the concept of delay outage by analogy with the concept of outage used in system
coverage planning. A mobile k is in delay outage (in critical period) when its packets
experience a delay greater than a given threshold T k defined by the mobile application
requirements. The delay experienced by each mobile is tracked all along the lifetime of its
connection. At each transmission of a packet of mobile k, the ratio of the total number of
packets whose delay exceeded the threshold divided by the total number of packets
transmitted since the beginning of the connection is computed. The result is called Packet
Delay Outage Ratio (PDOR) of mobile k and is denoted PDOR k. This measure is
representative of the emergency for the mobile k to be served. Fig. 3 illustrates an example
cumulative distribution of the packet delay of a mobile at a given time instant. A mobile can
be considered as satisfied when, at the end of its connection, its delay constraint is met, i.e.
its experienced PDOR is less than the application specific PDOR target.
In WFO scheduling, the required QoS, the experienced QoS and the transmission conditions
are jointly considered in an extended cross-layer approach. The scheduling principle is to
allocate a Resource Unit n to the mobile jwhich has the greatest WFO parameter value
WFO k,n such as:
j  arg max k (WFOk ,n ), k  1,..., K , (3)
whereWFO k,n is defined by:
WFOk ,n  mk ,n  f ( PDORk ), (4)
with f a strictly increasing polynomial function (Gueguen & Baey (d), 2008; Gueguen &
Baey, 2009):
f ( x)  1    x  , (5)
The exponent parameter ´ allows being sensitive and reactive to PDOR fluctuations which
guarantees fairness at a short time scale.  is a normalization parameter that ensures that
f(PDOR k) and m k,n are in the same order of magnitude.
With this scheduling, physical layer information (represented through the factor m k,n) are
used in order to take advantage of the time, frequency and multiuser diversity and
maximize the system capacity. Higher layer information (represented through the factor
f(PDOR k) are exploited in order to introduce dyna mic priorities between flows for ensuring
the same QoS level to all mobiles. With this original weighted system that introduces
dynamic priorities between the flows, WFO keeps a maximum number of flows active
across time but with relatively low traffic backlogs. This results in a well-balanced resource
allocation. Preserving the multiuser diversity allows to continuously take a maximal benefit
of opportunistic scheduling and thus maximize the bandwidth usage efficiency. When the
frequency diversity is sufficient, WFO better conceals the system capacity maximization,
QoS support and fairness objectives than PF and MaxSNR schemes.
4. Performance Evaluation
In this section we compare the most acknowledged schedulers, the Round Robin (RR), the
MaxSNR and the PF schemes with the most promising, the Weighted Fair Opportunistic
(WFO) scheduling. Each is implemented with subcarrier by subcarrier allocation.
Performance evaluation results are obtained using OPNET discrete event simulations. We
focus on the main scheduling problem: maximize the system capacity while ensuring high
fairness between mobiles localised at heterogeneous spatial positions in the cell.
In the simulations we assume 128 subcarriers and 5 time slots in a frame. The channel gain
model on each subcarrier considers free space path loss and multipath Rayleigh fading
(Parsons, 1992). We introduce a reference distance dreffor which the free space attenuation
equals aref. As a result the channel gain is given by:
3.5
d 
ak ,n  aref   ref    k2,n , (6)
 dk 
where dk is the distance to the access point of the mobile kand ´ 2 k,n represents the flat fading
experienced by this mobile kif it transmits or receives on subcarrier n. In the following, ´ 2 k,n
is Rayleigh distributed with an expectancy equal to unity.
The maximum transmit power satisfies:
P  T 
10log10  max s  aref  31 dB , (7)
 N 0 
where T sis the time duration of an OFDM symbol, P max is the maximum achievable transmit
power and N 0 is the single-sided power spectral density of noise. The BER target is taken
equal to 10-3. With this setting, the value of m k,n for the mobiles situated at the reference
distance is 6 bits when ´ 2 k,n equals unity.
We assume all mobiles run the same videoconference application. This demanding type of
application generates a high volume of data with high sporadicity and requires tight delay
constraints, which substantially complicate the task of the scheduler. Each mobile has only
one service flow with traffic composed of an MPEG-4 video stream (Baey, 2004) and an
AMR voice stream (Brady, 1969).
In these extended simulations, we analyzed the behaviour of the schedulers when mobiles
occupy different geographical positions. The objective is to clearly exhibit the ability of the
opportunistic schedulers to provide fairness whatever the respective position of the mobiles.
We first study a general context that includes mobility. We constitute two groups of 7
mobiles that both move straight across the cell, following the pattern described in Fig. 4 and
Fig. 5. Each mobile has a speed of 3 km/h and the cell radius is taken equal to 5 km (3 dref).
When a group of mobiles comes closer to the access point, the other group simultaneously
goes farther away. Additionally, the threshold time T k is fixed to the value 80 ms in order to
consider real time constraints and the PDOR target is 5 %.
Fig. 4. Mobility pattern.
Considering the path loss, the Rayleigh fading and this mobility model, we have computed
in Fig. 5 the evolution of the mean number of bits that may be transmitted per Resource Unit
for each group of mobiles, averaging over all the Resource Units of a frame. This shows the
impact of the mobile position on the mean m k,n values.
Regarding fairness, in wireless networks, it is well known that the closest mobiles to the
access point generally obtain better QoS than mobiles more distant thanks to their higher
spectral efficiency. Fig. 6 reports the mean PDOR experienced by each group of mobiles
across the time. MaxSNR is highly unfair. Indeed, as soon as the mobiles move away from
the access point, they experience high delays with a high number of packets in delay outage.
PF offers better results. It brings more fairness and globally attenuates the delay peaks of the
critical periods. However, we observe that WFO is the one that best smoothes these peaks. It
adequately and continuously allocates the adequate priorities between the mobiles reacting
to their relative movement across the cell. Providing a totally fair allocation of the
bandwidth resources, the WFO scheduling smoothes the delay experienced by each mobile
across time. Consequently, it further enhances the PF performances and the PDOR values
are further decreased. WFO results in a very fair resource allocation that fully satisfies the
delay constraints whatever the movement of the mobile.
Fig. 5. Position of the mobiles across time (for mobiles of group 1 on the left and for mobiles
of group 2 on the right).
Fig. 6. Mean number of bit(s) per Resource Unit for each group of mobiles (for mobiles of
group 1 on the left and for mobiles of group 2 on the right).
Fig. 7. PDOR fluctuation experienced by each group of mobiles (for mobiles of group 1 on
the left and for mobiles of group 2 on the right).
In order to further underline the advantage of opportunistic schedulers compared to the

classical Round Robin, we now study precisely the performance of the algorithms in a sub-
scenario where all mobiles are static. A first half of mobiles are situated close to the access
point and a second half 1.5 farther. The other parameters are identical for all the mobiles as
described in Table 1. The total number of mobiles sets the traffic load.
Group Distance dk Delay threshold T k Data rate

1 2 dref 80 ms 80 Kbps
2 3 dref 80 ms 80 Kbps
Table 1. Scenario setup with static mobiles.
First we focus on the fairness provided by each scheduler. Fig. 7a, 7b, 7c and 7d display the
overall PDOR for different traffic loads considering the influence of the distance on the
scheduling.
Fig. 8 a. With RR. Fig. 8 b. With MaxSNR.
Fig. 8 c. With PF. Fig. 8 d. With WFO.

Fig. 8. Measured QoS with respect to distance.
The classical RR fails to ensure the same PDOR to all mobiles. Actually, the RR fairly
allocates the RUs to the mobiles without taking in consideration that far mobiles have a
much lower spectral efficiency than closer ones. Moreover, the RR does not take benefit of
multiuser diversity which results in a bad utilization of the bandwidth and in turn, poor
system throughput. Consequently, an acceptable PDOR target of 5 % is exceeded even with
relatively low traffic loads. Based on opportunistic scheduling, the three other schemes
globally show better QoS performances supporting a higher traffic load. However, MaxSNR
and PF still show severe fairness deficiencies. Close mobiles easily respect their delay
requirement while far mobiles experience much higher delays and go past the 5 % PDOR
target when the traffic load increases. In contrast, WFO provides the same QoS level to all
mobiles whatever their respective position. WFO is the only one to guarantee a totally fair
allocation. This allows reaching higher traffic loads with an acceptable PDOR for all mobiles.
Additionally, looking at the overall PDOR for all mobiles at different traffic loads shows
that, besides fairness, WFO provides a better overall QoS level as well.
Fig. 9 a. Mean buffer Fig. 9 b. Mean packet delay. Fig. 9 c. Packet jitter.
occupancy for close mobiles
(solid lines) and far mobiles
(dashed lines).
Fig. 9. Buffer occupancy, delay and jitter.
Observing the mean buffer occupancy in Fig. 8a, WFO clearly limits the buffer occupancy to
a same and reasonable value whatever the position of the mobile. This allows staying under
the PDOR target for any traffic load. With its system of weights, WFO dynamically adjusts
the relative priority of the flows according to their experienced delay. With this approach,
sparingly delaying the closer mobiles, WFO builds on the breathing space offered by the
easy respect of the delay constraints of the closer mobiles (with better spectral efficiency) for
helping the farther ones. The WFO interesting performance results are corroborated in Fig.
8b and 8c where the overall values of the mean packet delay and jitter obtained using WFO
are smaller.
Fig. 10 a. CDF of end cycle Fig. 10 b. CDF of end cycle Fig. 10 c. CDF of end cycle
PDOR with MaxSNR. PDOR with PF PDOR with WFO
Fig. 10. Perceived QoS with different allocation schemes.
We then had a look at the QoS satisfaction level that each mobile perceives across the
lifetime of a connection. We divided the connection of each mobile in cycles of five minutes
and measured the PDOR at the end of each cycle. Fig. 9 shows the CDF of end cycle PDOR
values for a traffic load of 960 Kbps, using respectively the MaxSNR, the PF and the WFO
schemes (RR performances are not presented here since it is not able to support this high
traffic load.). We also estimated the mobile dissatisfaction ratio. We checked if at the end of
each cycle the delay constraint is met or not. We then computed the mobile dissatisfaction
ratio defined as the number of times that the mobiles are not satisfied (experienced PDOR ≥
PDORtarget) divided by the total number of cycles (cf. Fig. 10).
Fig. 11. Analysis of the respect of QoS constraints for different targeted QoS (Mobile
dissatisfaction if PDORtarget = 5 % on the left, if PDORtarget = 10 % on the right).
Highly unfair, MaxSNR fully satisfies the required QoS of close mobiles at the expense of
the satisfaction of far mobiles. Indeed, only 54.5 percents of these latter experience a final
PDOR inferior to a PDOR target of 5 % (cf. Fig. 9a). Unnecessary priorities are given to close
mobiles who easily respect their QoS constraints while more attention should be given to the
farther. This inadequate priority management dramatically increases the global mobile
dissatisfaction, which reaches 23 % as shown on Fig. 9a and Fig. 10 (on the left).
PF brings more fairness and allocates more priority to far mobiles. Compared to MaxSNR,
PF offers a QoS support improvement with only 12.8 % of dissatisfied mobiles (cf. Fig. 9b
and Fig. 10 (on the right)). Fairness is still not total since the farther mobiles have a lower
spectral efficiency than the closer ones due to pathloss. All mobiles do not all benefit of an
equal average throughput despite they all obtain an equal share of bandwidth. This induces
heterogeneous delays and unequal QoS. This fairness improvement compared to MaxSNR
indicates however that some flows can be slightly delayed to the benefit of others without
significantly affecting their QoS.
WFO was built on this idea. The easy satisfaction of close mobiles (with better spectral
efficiency) offers a degree of freedom which ideally should be exploited in order to help the
farther ones. WFO allocates to each mobile the accurate share of bandwidth required for the
satisfaction of its QoS constraints, whatever its position. With WFO, only 0.8 percents of the
mobiles are dissatisfied (cf. Fig. 9c and Fig. 10 (on the left)). Additionally, compared to Fig.
9a and Fig. 9b, Fig. 9c exhibits superimposed curves, which prove the WFO high fairness,
included at short term.
Fig. 10 shows that WFO brings the largest level of satisfaction. Indeed, for a tight PDOR
target of 5 % (see on the left), the dissatisfaction ratio with a high traffic load of 1120 Kbps is
equal to 18 % with WFO versus 29.7 % with PF, the best of the other scheduling schemes. If
we set the PDOR target to 10 %, the dissatisfaction ratio with a high traffic load of 1120 Kbps
is 0 % with WFO versus 13.8 % with the best of the other scheduling schemes (PF).
Fig. 12 a. Spectral efficiency. Fig. 12 b. Multiuser diversity

Fig. 12. Bandwidth usage efficiency.
We finally studied the system capacity offered by the four scheduling algorithms. Fig. 11a
shows the average number of bits carried on a used subcarrier by each tested scheduler
under various traffic loads. As expected, the non opportunistic Round Robin scheduling
provides a constant spectral efficiency, i.e. an equal bit rate per subcarrier whatever the
traffic load since it does not take advantage of the multiuser diversity. The three other tested
schedulers show better results. In contrast with RR, with the opportunistic schedulers
(MaxSNR, PF, WFO), we observe an interesting inflection of the spectral efficiency curve
when the traffic load increases. The join analysis of Fig. 11a and Fig. 11b shows that the
spectral efficiency of opportunistic scheduling is an increasing function of the number of
active mobiles, thanks to the exploitation of this supplementary multiuser diversity.
Consequently MaxSNR, PF and WFO increase their spectral efficiency with the traffic load
and the system capacity is highly extended compared to networks which use classical
scheduling algorithms. With these three schedulers, all mobiles are served even at the
highest traffic load of 1280 Kbps.
The performance of the four schedulers can be further qualified by computing the
theoretical maximal system throughput. Considering the Rayleigh distribution, it can be
noticed that ´ 2 k,n is greater or equal to 8 with a probability of only 0.002. In these ideal
situations, close mobiles can transmit/receive 6 bits per RU while far mobiles may
transmit/receive 4 bits per RU. If the scheduler always allocated the RUs to the mobiles in
these ideal situations, an overall efficiency of 5 bits per RU would be obtained which yields
a theoretical maximal system throughput of 1600 Kbps. Comparing this value to the highest
traffic load in Fig. 11a (1280 Kbps) further demonstrates the good efficiency obtained with
the opportunistic schedulers that nearly always serve the mobiles when their channel
conditions are very good. This result also shows that the WFO scheduling has slightly better
performances than the two other opportunistic schedulers. Keeping more mobiles active (cf.
Fig.11b) but with a relatively lower traffic backlog (cf. Fig.8a), the WFO scheme preserves
multiuser diversity and takes more advantage of it obtaining a slightly higher bit rate per
subcarrier (cf. Fig. 11a).
Fig. 13a. Measured QoS for close mobiles Fig. 13b. Spectral efficiency.
(solid lines) and far mobiles (dashed lines).
Fig. 13. Performances of schedulers with fixed multiuser diversity.
In the results described so far, the traffic load was varied by increasing or decreasing the
number of mobiles in the system, which modified the multiuser diversity. This exhibited the
opportunistic behaviour of the schedulers and especially their ability to take advantage of
the multiuser diversity brought with the increase of the number of mobiles. We also studied
the ability of each scheduler to take profit of the multiuser diversity brought by a given
number of users. In Fig.12, we provide complementary results obtained in a context where
the traffic load variation is done through just increasing the mobile bit rate requirement and
keeping a constant number of users (10 mobiles). The results in Fig. 12a show that, as
previously, WFO outperforms the other scheduling schemes. With its weighted algorithm,
WFO dynamically adjusts the priorities of the mobiles and ensures a completely fair
allocation. WFO is the only one which allows reaching higher traffic loads with an
acceptable PDOR for all mobiles. Additionally, even if the traffic load increases without
variation in the number of mobiles, WFO keeps more mobiles active across the time than the
other schemes and takes better advantage of the multiuser diversity. The analysis of Fig. 12b
confirms that WFO maximizes the average bit rate per subcarrier.
5. Conclusion
Opportunistic schedulers take benefit of multiuser and frequency diversity. They preferably
allocate the resources to the active mobile(s) with the most favourable channel conditions at
a given time. This maximizes the system throughput of OFDM wireless networks. Three
major algorithms have emerged: MaxSNR, PF and more recently WFO. However, in spite of
their high performances in terms of system throughput maximization, both MaxSNR and PF
suffer of severe fairness deficiencies owing to unequal spatial positioning of the mobiles.
This issue is resolved with WFO which appears as the best current opportunistic scheduler.
WFO jointly considers the transmission conditions, the currently measured/experienced
QoS and the QoS targets of the mobiles in the bandwidth allocation process. With an
original weighted system that introduces dynamic priorities between the flows, it
dynamically favors the flows that go through a critical period and always attributes the
adequate priorities for improved QoS support. Keeping a maximum number of flows active
across time but with relatively low traffic backlogs, WFO is designed for best profiting of the
multi-user diversity taking advantage of the dynamics of the multiplexed traffics.
Preserving the multiuser diversity, WFO takes a maximal benefit of the opportunistic
scheduling technique and maximizes the system capacity. Additionally, this also achieves a
time uniform fair allocation of the resource units to the flows ensuring short term fairness.
This higher layers/MAC/PHY cross-layer approach better conceals the system capacity
maximization, fairness objectives and the full support of multimedia services with adequate
QoS.
6. References
Hoymann, C. (2005). Analysis and performance evaluation of the OFDM-based
metropolitan area network IEEE 802.16. Computer Networks, ol. 49,
V No. 3, pp. 341-
363, ISSN: 1389-1286
Andrews, M.; Kumaran, K.; Ramanan, K. Stolvar, A. & whiting P. (2001). Providing quality
of service over a shared wireless link. IEEECommunicationsMagazine , Vol. 39, No.2,
pp. 150-154, ISSN: 0163-6804
Van de Beek, J.-J. ; Borjesson, P.O. ; Boucheret, M.-J ; Landstrom, D. ; Arenas, J.-M. Odling,
P.; Ostberg, C.; Wahlgvist, M. & Wilson, S.K. (1999). A time and frequency
synchronization scheme for multiuser OFDM. IEEE J.Sel. Areas Commun , Vol.17,
No.11, pp 1900-1914, ISSN: 0733-8716
Li, Y.G; Seshadri, N. & Ariyavisitakul, S. (1999). Channel estimation for ofdm systems with
transmitter diversity in mobile wireless channels. IEEE J.Sel. Areas Commun , Vol.17,
No.3, pp 461-471, ISSN: 0733-8716
Truman, T.E. & Brodersen, R.W (1997). A measurement-based characterization of the time
variation of an indoor wireless channel, proceedings of Int. Universal Personal
Communications Record (ICUPC) , pp. 25-32, ISBN: 0-7803-3777-8, San Diego, CA,
USA, October 1997
Knopp, R. & Humblet, P. (1995). Choi, J. (1996). Information capacity and power control in
single-cell multiuser communications, proceedings of IEEE Conference on
Communications,(ICC) pp 331-335, ISBN: 0-7803-2486-2, Seattle, WA, USA, june 1995
Wong, C.Y.; Cheng, R.S.; Lataief, K.B. & Murch, R.D. (1999). Multiuser OFDM with
adaptative subcarrier, bit and power allocation. IEEE J.Sel. Areas Commun , Vol.17,
No.10, pp 1747-1758, ISSN: 0733-8716
Wang, X. & Xiang, Y. (2006). An OFDM-TDMA/SA MAC protocol with QoS constaints for
broadband wireless LANs. ACM/Springer Wireless Networks , Vol. 12, No. 2, pp 159-
170, ISSN: 1022-0038
Viswanath, P.; Tse, D.N.C. & Laroia, R. (2002). Opportunistic beamforming using dumb
antennas. IEEE Transactions on Information , Vol. 48, No.6, pp. 1277–1294,
Theory
ISSN: 0018-9448
Kim, H.; Kim, K.; Han, Y. & Lee, J. (2002). An efficient scheduling algorithm for QoS in
wireless packet data transmission, proceedings of IEEE Int. Symposium on Personal,
Indoor and Mobile Radio Communications , ppPIMRC)
( 2244-2248, ISBN: 0-7803-7589-0,
Lisboa, Portugal, September 2002
Anchun, W.; Liang, X.; Xjiin, S.X. & Yan, Y. (2003). Dynamic resource management in the
fourth generation wireless systems, proceedings of IEEE Int. Conference on
Communication Technology (ICCT) pp , 1095-1098, ISBN: 7-5635-0686-1, Beijing,
China, April 2003
Svedman, P.; Wilson, K. & Ottersen, B. (2004). A QoS-aware proportional fair scheduler for
opportunistic OFDM, proceedings of IEEE Int. Vehicular Technology Conference , VTC)
(
pp 558-562, ISBN: 0-7803-8521-7, Los angeles, CA, USA, September 2004
Kim, H.; Kim, K.; Han, Y. & Yun, S. (2004). A proportional fair scheduling for multicarrier
transmission systems, proceedings of IEEEInt.Vehicu larTechnologyConference , (VTC)
pp. 409-413, ISBN: 0-7803-8521-7, Los angeles, CA, USA, September 2004
Choi, J.-G. & Bahk, S. (2007). Cell-throughput analysis of the proportional fair scheduler in
the single-cell environment. IEEETransactionsonVehicularTechnology , vol. 56, No.2,
pp. 766-778, ISSN: 0018-9545
Gueguen, C. & Baey, S. (2008). Compensated proportional fair scheduling in multiuser
OFDM wireless networks, proceedings of IEEE Wireless and Mobile Computing,
Networking and Communications (WIMOB), pp119-125, ISBN: 978-0-7695-3393-3,
Avignon, France, October 2008
Holtzman, J. (2001). Asymptotic analysis of proportional fair algorithm, proceedings of IEEE
Int.SymposiumonPersonal,Indoor dMobile
an RadioCommunicationsPIMRC) ( , pp. 33-
37, ISBN: 0-7803-7244-1, San Diego, CA, USA, October 2001
Gueguen, C. & Baey, S. (2008). Weighted fair opportunistic scheduling for multimedia QoS
support in multiuser OFDM wireless networks, proceedings of IEEE Int. Symposium
on Personal, Indoor and Mobile Radio Communications , pp. 1-6,
( ISBN: 978-1-
PIMRC)
4244-2643-0, Cannes, France, September 2008
Gueguen, C. & Baey, S. (2008). An efficient and fair scheduling scheme for multiuser OFDM
wireless networks, proceedings of IEEE Int.Wireless Communications and Networki
Conference WCNC) ( , pp. 1610-1615, ISBN: 978-1-4244-1997-5, Las Vegas, NV, USA,
April 2008
Gueguen, C. & Baey, S. (2008). Scheduling in OFDM wireless networks without tradeoff
between fairness and throughput, proceedings of IEEE Int. Vehicular Technology
Conference VTC) ( , pp. 1-5, ISBN: 978-1-4244-1721-6, Calgary, Canada, September
2008
Gueguen, C. & Baey, S. (2009). A Fair Opportunistic Access Scheme for Multiuser OFDM
Wireless Networks. EURASIP Journal on Wireless Communications and Networking.
Special issue on Fairness" in Radio Resource Management for Wireless Net
Volume 2009 (2009), Article ID 726495, pp. 70-83
Parsons, J.D (1992). TheMobileRadioPropagationChannel, Wiley, ISBN: 978-0-471-98857-1
Baey, S. (2004). Modeling MPEG4 video traffic based on a customization of the DBMAP,
proceedings of Int. Symposium on Performance Evaluation of Computer
Telecommunication Systems (SPECTS) pp. , 705-714, ISBN: 1-56555-284-9, San Jose,
California, USA, July 2004
Brady, P. (1969). A model for generating on-off speech patterns in two-conversation. Bell
SystemTechnicalJournal , vol. 48, No.1, pp. 2445-2472
Bidirectional Cooperative Relaying 281
1
Bidirectional Cooperative Relaying
Prabhat Kumar Upadhyay and Shankar Prakriya
Indian Institute of Technology Delhi
India
1. Introduction
Modern radio communication systems aim to enhance throughput and reliability in wireless
networks with limited resources. Wireless mobile communication over a radio channel is lim-
ited by multipath, fading, path loss, shadowing, and interference. Spatial diversity techniques
are widely adopted to combat fading and other channel impairments. Cooperative commu-
nications have been recently developed to harness spatial diversity even with single-antenna
terminals. The distributed terminals cooperate by relaying each other’s message in order to
realize a virtual antenna array and achieve cooperative diversity. Cooperative relaying has
become a promising technique for enhancing coverage, reliability and throughput of wireless
networks with stringent spectrum and power constraints. They have found applications in
wireless cellular, ad hoc/sensor networks, WiFi/WiMAX, etc.
Dual-hop three-terminal channels wherein a relay terminal assists in the communication be-
tween source and destination terminals through some cooperation protocol are of particular
interest. Relaying can be performed in either full-duplex or half-duplex mode. Full-duplex
relaying allows the radios to receive and transmit simultaneously using the same frequency
channel and hence achieves higher spectral efficiency. However, the large difference in power
levels of the transmit and receive signals (typically 100-150 dB) makes its implementation
practically difficult. In half-duplex mode, the reception and transmission at the radios are
performed in time/frequency/code division orthogonal channels. Half-duplex systems are
therefore practically feasible. The major drawback of half-duplex relaying is a substantial loss
in spectral efficiency. This is because half of the channel resources are allocated to the relay for
cooperation, which reduces the overall data rate.
While much research has focused on exploiting cooperative diversity, little effort has been di-
rected towards improving spectral efficiency under half-duplex constraints. The authors in
(Rankov & Wittneben, 2007) propose a new two-phase two-way relaying protocol where a
bidirectional connection between two terminals is established with one half-duplex relay. Un-
der this scheme, two connections are realized in the same physical channel, thereby improving
the spectral efficiency. Referred to as the Two-Way Relay Channel (TWRC) in literature, this
pragmatic approach has become a focus of extensive recent research. An example of a TWRC
is the downlink and uplink in wireless mobile networks whereby both the base station and the
mobile station need to communicate via an assisting relay station due to the lack of a reliable
direct link. This is advantageous in the case when the mobile is highly shadowed or near the
cell edge. It is important to note that even when a direct link of sufficient quality is available,
it cannot be utilized in two-phase TWRC because otherwise both terminals need to transmit
and receive simultaneously in the same phase.
In a separate remarkable development, the emergence of network coding (Ahlswede et al.,

2000) has changed the way communication networks are designed. Network coding allows
the intermediate nodes to combine and code the data from multiple sources in order to en-
hance the overall network throughput. Originally proposed for wired communication net-
works, there has recently been much interest in applying network coding to wireless relay
networks (Hao et al., 2007). In view of the spectral efficiency loss due to half-duplex mode,
a coded bidirectional relaying scheme with three transmission phases has been proposed in-
dependently in (Wu et al., 2005) and (Larsson et al., 2006). The authors in (Kim et al., 2008)
compared and analyzed the performance of various half-duplex bidirectional relaying proto-
cols. The idea of network coding has been further exploited for the bidirectional cooperation
in (Hausl & Hagenauer, 2006); (Baik & Chung, 2008); (Cui et al., 2008a); (Cui et al., 2008b). It
has been shown in (Katti et al., 2007b) that wireless two-way relaying coupled with network
coding achieves higher data rates.
Two-way or bidirectional relaying is flexible to allow various physical-layer transmission tech-
niques. A lot of research is in progress on topics like TWRC capacity region or achievable rate
region (Oechtering et al., 2008), channel estimation (Zhao et al., 2008); (Gao et al., 2008), multi-
hop relaying (Vaze & Heath Jr., 2008), resource allocation (Agustin et al., 2008), distributed
space-time coding (Cui et al., 2008c), distributed relay selection (Ding et al., 2009) and the like,
using various physical layer signalling techniques, OFDM for example (Ho et al., 2008); (Jit-
vanichphaibool et al., 2008). Also, the bidirectional relaying scheme has been extended to the
multi-user scenario (Chen & Yener, 2008); (Esli & Wittneben, 2008). Multiple-Input Multiple-
Output (MIMO) bidirectional relaying (Unger & Klein, 2007); (Gunduz et al., 2008) is a hot
research area and is often envisioned to further improve the link reliability and bidirectional
throughput of wireless systems.
The aim of this chapter is to present, in a unified fashion, the state-of-the-art in this new area
of bidirectional cooperative communication, to elaborate on the recent analytical findings and
their significance, to support them with various simulation results, and to discuss future areas
of research.
2. Cooperative Communications
Cooperative communication systems seek to enhance the link capacity and transmission reli-
ability through cooperation between distributed radios. They exploit the broadcasting nature
of the wireless medium and allow single-antenna terminals to cooperate through relaying.
The conventional form of cooperation is multi-hopping, where a source communicates with a
destination via a series of dedicated relays. It is mainly used to combat signal attenuation in
long-range communication and it does not provide any diversity advantages. The key issue in
cooperative communications is resource sharing among network nodes. A three-terminal net-
work acts as a fundamental unit in cooperative communication and has been widely studied
in the literature.
The three-terminal relay channel model, introduced in (Meulen, 1971), comprises a source T1 ,
a destination T2 , and a dedicated relay R (as shown in Figure 1). The relay aids in commu-
nicating information from source to destination without actually being an information source
or sink. It was assumed that all nodes operate in the full-duplex mode, so the system can be
viewed as a Broadcast Channel (BC) at the source, and a Multiple Access Channel (MAC) at
the destination. The upper and lower bounds on the capacity of the non-faded relay chan-
Fig. 1. The wireless relay channel.
recent developments are motivated by the user cooperation (Sendonaris et al., 2003) and co-
operative diversity (Laneman et al., 2004) in a fading channel. The authors in (Sendonaris et
al., 2003) introduced user cooperation by allowing the relay to transmit its own independent
information. Cooperative diversity introduced in (Laneman et al., 2004) is realized by relaying
and user cooperation. They proposed different cooperative diversity protocols and analyzed
their performance in terms of outage probability. The terms Amplify-&-Forward (AF) and
Decode-&-Forward (DF) were introduced in their work.
2.1 Cooperative Relaying Protocols

Consider a three-terminal wireless network as shown in Figure 1 in which terminal T1 wants
to transmit data to terminal T2 with the help of a relay terminal R. In view of cellular network,
T1 and R might be mobile stations and T2 might be a base station. Under cooperative relaying
strategy, T1 and a suitable R can share their resources, such as power and bandwidth, to
transmit the information of T1 . This cooperation might provide diversity because, even if
the direct link between T1 and T2 is severely faded, the information might be successfully
transmitted via R. It is assumed that the relay node operates in half-duplex mode and has no
The source and relay nodes can share their resources on the basis of some cooperation strate-
gies to achieve the highest throughput possible for any given coding scheme. Based on dif-
ferent signal processing schemes employed at the relays, the cooperative relaying methods
are classified into fixed relaying and adaptive relaying (Laneman et al., 2004). For fixed re-
laying, the relay can amplify its received signal subject to its power constraint, or decode,
re-encode, and then retransmit the messages, referred to (respectively) as Amlify-&-Forward
(AF) or Decode-&-Forward (DF). This scheme has the advantage of easy implementation, but
the disadvantage of low spectral efficiency. This is because half of the channel resources are
allocated to the relay for transmission. Adaptive relaying schemes build upon fixed relaying
and adapt based upon Channel State Information (CSI) between cooperating terminals (selec-
tive relaying) or upon limited feedback from the destination (incremental relaying). Selective
relaying allows transmitting terminals to select a suitable cooperative or non-cooperative ac-
tion based on the measured SNR between them. If the received SNR at the relay exceeds a
certain threshold, the relay performs DF operation on the message. Otherwise, if the channel
between T1 and R has severe fading such that SNR falls below the threshold, the relay idles.
Incremental relaying improves upon the spectral efficiency of both fixed and selective relaying
by exploiting limited feedback from the destination and relaying only when direct link from
source to destination has an SNR below a threshold.
2.2 Outage Analysis and Diversity Gain

When the channel is time-varying, the channel capacity has different notions depending on
the different fading states. Ergodic (Shannon) capacity is an appropriate capacity metric for
channels that vary quickly, or where the channel is ergodic over the time period of interest. It
can be evaluated by averaging the mutual information over all possible channel realizations.
An alternate outage capacity notion is suitable for applications where the data rate cannot
depend on channel variations (except in outage states, where no data are transmitted). It is
a measure of data rate that can be supported by a system with a certain error probability. To
investigate the diversity gain, the performance of relaying protocols is characterised in terms
of outage probability. Assume frequency flat slow fading channel with CSI knowledge at the
receivers only. Perfect synchronization among the terminals is also assumed. Considering
a baseband-equivalent discrete-time channel model, the transmissions in time slot k can be
expressed as
yr [ k ] = h1 [ k ] x1 [ k ] + n1 [ k ] (1)
y2 [ k ] = h3 [ k ] x1 [ k ] + n3 [ k ] (2)
where x1 [k] is the transmitted signal from T1 , yr [k] and y2 [k] are the received signals at the
relay and T2 respectively, hi captures the effects of path-loss, shadowing and frequency nonse-
lective fading, ni ∼ CN (0, σ2 ) is the Additive White Gaussian Noise (AWGN) which captures
the effects of receiver noise and other forms of interference in the system, where i ∈ {1, 2, 3}.
Throughout the chapter, we use hi [k] and hi interchangeably for brevity. The relay processes
yr [k] and relays the information by transmitting xr [k]. The signal received at T2 in time slot
k+1
y2 [ k + 1] = h2 [ k + 1] xr [ k ] + n2 [ k + 1]. (3)
As a function of the fading coefficients hi (modeled as zero-mean, independent, circularly
symmetric complex Gaussian random variables with variances σhi 2 ), the mutual information
for a protocol is a random variable I. For a target rate R, I < R denotes the outage event
and Pr[I < R] denotes the outage probability (Laneman et al., 2004). The maximum average
mutual information between input and output in direct transmission, achieved by indepen-
dent identically distributed (i.i.d.) zero-mean, circularly symmetric complex Gaussian inputs,
is given by

ID = log 1 + γ| h3 |2 (4)
where γ = P1 /σ2 is defined as SNR without fading and P1 is the average transmit power of
terminal T1 . For Rayleigh fading, | h3 |2 is exponentially distributed with parameter 1/σh3
2 , the
outage probability derived in (Laneman et al., 2004) is given as

2R − 1 1 2R − 1
Pr[ ID < R] = 1 − exp − 2
∼ 2 . (5)
γσh3 σh3 γ
The direct transmission does not achieve any diversity gain as is obvious from γ−1 depen-
dence of outage probability in Equation (5).
2.3 AF Relaying
In this protocol, the relay amplifies the received signal in the first time slot according to its
available average transmit power and forwards a scaled signal in the second time slot to the
destination terminal. To remain within its power constraint, an amplifying relay must use
gain

Pr
g[k] ≤ (6)
P1 | h1 |2 + σ2
which is inversely proportional to the received power. Thus the relay transmits the signal
xr [k + 1] = g[k]yr [k] with the power Pr in the second time slot. This scheme can be viewed
as repetitive coding from two distributed transmitters T1 and R, except that the relay R am-
plifies the noise in its received signal. The destination T2 can decode its received signal y2 by
suitably combining the signals from the two time slots. This protocol produces an equivalent
one-input two-output complex Gaussian noise channel with different noise levels in the out-
puts. The SNR received at the destination is the sum of the SNRs from T1 and R links. The
maximum average mutual information between the input and the two outputs, achieved by
i.i.d. complex Gaussian inputs, is given by

1 2 γ| h2 gh1 |2
I AF = log 1 + γ| h3 | + . (7)
2 (1 + | h2 g |2 )
Note that g is a function of h1 . Notations g[k] and g are used interchangeably throughout for
brevity. The outage probability can be approximated at high SNR (Laneman et al., 2004) as
2
2 + σ2
σh1 h2 22 R − 1
Pr[ I AF < R] ∼
2 σ2 σ2
. (8)
2σh3 h1 h2
γ
The pre-log factor 1/2 in Equation (7) is due to half-duplex relaying which needs two channel
uses to transmit the information from source to destination. The outage behavior decays as
γ−2 , which indicates that fixed AF protocol offers diversity gain of 2.
2.4 DF Relaying
In this scheme the relay processes its received signal yr [k] in the first time slot to obtain an
estimate x̂1 [k] of the source transmitted signal. Under a repetition-coded scheme, the relay
transmits the signal xr [k + 1] = x̂1 [k] in the second time slot. Although fixed DF relaying has
the advantage over AF relaying in reducing the effects of additive noise at the relay, it entails
the possibility of forwarding erroneously detected symbols to the destination. Therefore it is
required that both the relay and destination decode the entire codeword without error. This
leads to the expression of maximum average mutual information IDF between T1 and T2 as the
minimum of the two maximum rates, one at which the relay R can reliably decode the source
message, and the other at which the destination T2 can reliably decode the source message
given repeated transmissions from the source and relay. This implies that
1
IDF = min log 1 + γ| h1 |2 , log 1 + γ| h3 |2 + γ| h2 |2 . (9)
2
Here it is obvious that the performance of this system is limited by the worst link among the
T1 -T2 and T1 -R. The outage probability can be obtained for high SNR (Laneman et al., 2004)
as
1 22R − 1
Pr[ IDF < R] ∼ 2
. (10)
σh1 γ
The γ−1 behavior in Equation (10) indicates that fixed DF protocol does not provide diversity
gain for large SNR.
2.5 Numerical Results

We compare the outage analysis results of AF and DF relaying protocols with direct transmis-
sion. We consider the case of statistically symmetric networks in which the Rayleigh fading
2 = 1. The noise variance σ2 is assumed to be unity. We
channel variances are identical i.e., σhi
realized 10000 random channels using Monte Carlo simulation. Figure 3 shows the outage
probabilities versus SNR in dB for low spectral efficiency (roughly 2 bps/Hz). The diversity
order of 2 achieved by AF protocol is clear from the steeper curve slope in Figure 3. Also the
fixed DF relaying curve indicates no diversity gain and hence does not have any advantage
Fig. 3. Outage probabilities versus SNR in the low spectral efficiency regime.
Fig. 4. Outage probabilities versus spectral efficiency for high SNR.
over direct transmission. In Figure 4, the outage probabilities are depicted as functions of
spectral efficiency R for a fixed SNR of 35 dB. It is clear from Figure 4 that the performance
of fixed AF and DF protocols generally degrade with increasing rate R. It degrades faster
for AF scheme because of the inherent loss in spectral efficiency. Again, fixed DF protocol
does not have any diversity advantage over direct transmission. At sufficiently high rate R,
direct transmission becomes more efficient than cooperative relay communication. So we can
conclude that half-duplex operation requires double channel resources compared to direct
transmission for a given rate and hence leads to larger effective SNR losses for increasing rate.
However the performance enhancements in low spectral efficiency regime can be translated
into decreased transmit power for the same reliability.
3. Bidirectional Relaying
Although unidirectional or one-way communication has been extensively considered in the
literature, there is a lot of interest in recent years on bidirectional or two-way communication.
In two-way communication, two terminals simultaneously transmit their messages to each
other and the messages interfere with each other. The Two-Way Communication Channel
(TWC) was first studied by Shannon, who derived inner and outer bounds on the capacity
region (Shannon, 1961). He used a restricted two-way channel in which the encoders of both
terminals do not cooperate, and the transmitted symbols at one terminal only depend on the
message to be transmitted at that terminal (and not on the previously received symbols). He
showed that the inner bound coincides with the capacity region of the restricted two-way
channel. Later, the two-way communication problem was investigated for the full-duplex
relay channel, and the achievable rate regions were derived in (Rankov & Wittneben, 2006),
(Avestimehr et al., 2008), (Nam et al., 2008) and references therein. Further TWC has been
exploited in (Rankov & Wittneben, 2007), as TWRC, in order to mitigate the spectral efficiency
loss of cooperative protocols under half-duplex relaying. Recall that cooperative protocols
can provide higher outage capacity but not ergodic capacity because of use of orthogonal
time slots for relaying. Our goal here is to analyze spectrally efficient (measured in bits per
channel use) transmission schemes for the half-duplex bidirectional relay channel. Presently
the TWRC protocol has drawn much interest from both academic and industrial communities
owing to its potential application in wireless networks.
3.1 One-Way Relay Channel (OWRC)

Consider a wireless channel in which two nodes T1 and T2 wish to exchange independent
messages with the help of a relay node R. Once again, we assume that all terminals operate in
half-duplex fashion. Therefore the relay terminal cannot receive and transmit simultaneously
on the same channel resource; it receives a signal on a first hop, applies signal processing
and retransmits the signal on a second hop. More importantly, there is no reliable direct link
between T1 and T2 due to shadowing, large separation between them, or use of low power
signaling. This is feasible in practice when the users are geographically separated, and the
signals received from each other are very weak. This is the case when two distant land stations
communicate with a satellite, or two mobile users located on opposite sides of a building
communicate with the same base station on top of the building. When there is no direct
connection between the two wireless terminals, relays are essential to enable communication.
where h1 is the complex channel gain between source and relay (first hop), x1 ∼ CN (0, P1 ) is
the transmit symbol of the source, and nr ∼ CN (0, σr2 ) is the AWGN at the relay. The relay
scales yr [k] by

Pr
g[k] = (12)
P1 | h1 [k]|2 + σr2
where Pr is the average transmit power of the relay. Depending on the amount of channel
knowledge at the relay, different choices for the relay gain are possible. In time slot k+1, the
destination receives
y2 [ k + 1] = h2 [ k + 1] g [ k ] h1 [ k ] x1 [ k ] + h2 [ k + 1] g [ k ] nr [ k ] + n2 [ k + 1] (13)
where h2 is the complex channel gain between relay and destination (second hop) and n2 ∼
CN (0, σ22 ) is the AWGN at the destination. The information rate of this scheme for i.i.d. fading
channels h1 [k] and h2 [k] is given by (Rankov & Wittneben, 2007)

1 P1 | h2 gh1 |2
I AF = E log 1 + 2 (14)
2 σ2 + σr2 | h2 g|2
where E {·} denotes the expectation with respect to the channels h1 and h2 . The pre-log factor
1/2 follows because of the two channel uses needed to transmit the information from T1 to T2 .
3.1.2 DF-OWRC
In this scheme, the relay decodes the message sent by the source, re-encodes it (by using the
same or a different codebook), and forwards the message to the destination. In time slot k, the
relay receives
yr [ k ] = h1 [ k ] x1 [ k ] + nr [ k ]. (15)
After decoding and retransmission, the destination receives in time slot k+1
y2 [ k + 1] = h2 [ k + 1] xr [ k + 1] + n2 [ k + 1] (16)
where xr ∼ CN (0, Pr ) is the transmit symbol of the relay. The information rate of this scheme
for i.i.d. fading channels h1 [k] and h2 [k] is given by (Rankov & Wittneben, 2007)

1 P1 | h1 |2 Pr | h2 |2
IDF = min E log 1 + , E log 1 + . (17)
2 σr2 σ22
This rate is exactly the ergodic capacity of the conventional half-duplex cooperative relay
channel with no direct connection. Compared to a bidirectional communication between T1
and T2 without two-hop relaying, the number of required resources is doubled. Therefore this
protocol is spectrally inefficient and does not take full advantage of the broadcast nature of
the wireless channel.
3.2 Two-Way Relay Channel (TWRC)

The two-way relaying protocol (Rankov & Wittneben, 2007) is an effective means to increase
the spectral efficiency of a half-duplex relay network. As illustrated in Figure 6, messages
of nodes T1 and T2 are delivered to nodes T2 and T1 respectively in two phases, named the
Multiple Access Channel (MAC) and Broadcast Channel (BC) phase. In the first (MAC) phase,
T1 and T2 transmit their signals to the relay node at the same time. After receiving the signals,
the relay node performs appropriate signal processing and broadcasts the resulting signal
to both nodes T1 and T2 in the second (BC) phase. At each node, its symbols contribute self
interference but can clearly be canceled (because they are known). The channels in the forward
direction are assumed to be the same as in the backward direction i.e., channel reciprocity is
assumed.
The transmission in each direction suffers still from the pre-log factor 1/2. However, the half-
duplex constraint can here be exploited to establish a bidirectional connection between two
terminals and to increase the sum rate of the network.
3.2.2 DF-TWRC
Consider now a two-way communication between terminals T1 and T2 via a half-duplex DF
relay R. In time slot k both terminals T1 and T2 transmit their symbols to relay R. In this MAC
phase, the relay receives
yr [ k ] = h1 [ k ] x1 [ k ] + h2 [ k ] x2 [ k ] + nr [ k ], (23)

decodes the symbols x1 [k] and x2 [k] and transmits xr [k + 1] = βx1 [k ] + 1 − βx2 [k] in the
next time slot (BC phase). The received signals at T2 and T1 are
y2 [ k + 1] = h2 [ k + 1] xr [ k + 1] + n2 [ k + 1] (24)
y1 [ k + 1] = h1 [ k + 1] xr [ k + 1] + n1 [ k + 1]. (25)
The relay uses an average transmit power of βPr for the forward direction and (1 − β) Pr
for the backward direction. Since T1 knows x1 [k] and T2 knows x2 [k], these symbols (back-
propagating self-interference) can be subtracted at the respective terminals prior to decoding
of the symbol transmitted by the partner terminal. We assume that the relay decodes x1 [k]
and x2 [k] without errors. The sum-rate is given (Rankov & Wittneben, 2007) by
IDF(sum) = max min( I MA , I1 ( β) + I2 (1 − β)) (26)

β

1 P1 | h1 |2 + P2 | h2 |2
where I MA = C
2 σr2

1 P1 | h1 |2 βPr | h2 |2
I1 ( β) = min C ,C
2 σr2 σ22

1 P2 | h2 |2 (1 − β) Pr |h1 |2
I2 (1 − β) = min C ,C
2 σr2 σ12
where C ( x ) = E {log (1 + x )} .
In the absence of CSI knowledge in the BC phase, β = 12 is used by the relay. Note that in
fast fading channels, the channel coefficients change from phase to phase, and reliable CSI
may not be available. In other case β may be optimally chosen to maximize the sum rate. The
choice of β will depend on the amount of channel knowledge available (CSI or its statistics),
and applicable path losses in the links.
3.3 Simulation Results

We compute the achievable rates of the one-way and two-way relaying schemes by Monte
Carlo simulations. We consider a fixed symmetric network in which the relay is equidistant
from the two terminals. The Rayleigh fading channel gains are modeled as hi ∼ CN (0, 1). The
AWGN variances are chosen as σ12 = σ22 = σr2 = σ2 and the transmit powers P1 = P2 = P/2
and Pr = P such that the network consumes in each time slot an average power of P. The SNR
is defined as the ratio P/σ2 . Over 10000 random channels were used to average the rates in
Figures 7 and 8.
Fig. 7. Sum rate for two-way half-duplex AF relaying protocol.

4. Resource Allocation
The performance of wireless relay networks can be significantly improved by efficient man-
agement of available radio resources. Mostly, resource management via power allocation is
employed. We have discussed bidirectional relaying in the previous section, static resource
allocation was assumed where all transmission phases are of same duration and all terminals
have individual power constraints with balanced rates. Dynamic resource allocation has been
investigated in (Agustin et al., 2008) in terms of phase durations, individual and sum-average
power, and data rate. The system model employs a DF relay that applies superposition coding
and takes into account the traffic asymmetry. It is assumed that the transmission is performed
in frames of length υ with N channel uses and normalized bandwidth of unity. The duration
of the two phases (MAC and BC) are denoted by υ1 and υ2 respectively. The two power con-
straints considered are maximum power and sum-average power (both denoted by P). The
first constraint assumes that all terminals transmit with power P, whereas in second case the
total average power used by the three terminals is considered to be P. The mutual information
of different links assuming equal noise power at all terminals is given by

P | h |2
I1r ( P1 ) = N log 1 + 1 21 (27)
σ

P | h |2
I2r ( P2 ) = N log 1 + 2 22 (28)
σ

P1 | h1 |2 P2 | h2 |2
I MAC ( P1 , P2 ) = N log 1 + + (29)
σ2 σ2
where I1r and I2r represent mutual information of T1 − R and T2 − R links respectively, and
I MAC is the mutual information at the relay when both terminals transmit simultaneously in
the MAC phase. The signal received by the relay in MAC phase is given by

h1 [k] x1 [k] + h2 [k] x2 [k] + nr [k] for 0 ≤ k ≤ υ1 N
yr [ k ] = (30)
0 for υ1 N ≤ k ≤ N.
Under superposition coding, the DF relay forwards one signal xr [k] intended to each destina-
tion by distributing the total power between them as

β 1 Pr β 2 Pr
xr = x + x (31)
P1 1 P2 2
where β 1 and β 2 indicate the fraction of power allocated to each signal. The signal received
by each destination in second phase is given by

0 for 0 ≤ k ≤ υ1 N
y1 [ k ] = (32)
h1 [k] xr [k] + n1 [k] for υ1 N ≤ k ≤ N

0 for 0 ≤ k ≤ υ1 N
y2 [ k ] = (33)
h2 [k] xr [k] + n2 [k] for υ1 N ≤ k ≤ N.
The optimal selection of phase duration, data rate of each terminal are found as the maximiza-
tion of the following problem (Agustin et al., 2008):

( R1 , R2 ) ∈ ρ(υ ) for 0 ≤ (υ ) ≤ 1
arg max ϑ1 R1 + ϑ2 R2 s.t. (34)
υ,R1 ,R2 ,P1 ,P2 ,Pr ϕ( P1 , P2 , Pr ) ≤ P for ζ ( R1 , R2 ) = κ
where R1 , R2 represents the rate transmitted by terminal T1 , T2 respectively, υ is a vector that

contains the duration of different phases, function (υ ) defines the linear connection between
duration of phases, ρ(υ ) denotes the achievable rate region for a given υ, function ϕ( P1 , P2 , Pr )
represents a combination of the transmitted power by the terminals considering the power
constraints, function ζ ( R1 , R2 ) indicates a linear dependence between data rates R1 and R2 .
The achievable rate region boundary can be attained with optimum phase and rate selection
(by adjusting the parameters ϑ1 and ϑ2 ).
The achievable rate region ρ(υ ) for the two-way DF protocol, described under MAC (Cover
& Thomas, 1991), is given by

 R1 ≤ min {υ1 I1r ( P1 ), υ2 I2r ( β 1 Pr )}
ρ ( υ1 , υ2 ) = R ≤ min {υ1 I2r ( P2 ), υ2 I1r ( β 2 Pr )} (35)
 2
R1 + R2 ≤ υ1 I MAC ( P1 , P2 )
with (υ1 , υ2 ) = υ1 + υ2 . For terminals transmitting with their maximum power P, the maxi-
mum power constraint can be expressed as
ϕmax ( P1 , P2 , Pr ) = { P1 = P, P2 = P, Pr = P} . (36)
The power distribution at the relay satisfies 0 ≤ β 1 + β 2 ≤ 1. Hence the power allocation
can be optimized at the relay only. For sum-average power constraint, each terminal uses a
fraction of total power P which is controlled by variables δ and β as follows

δ P δ P P
ϕ avg ( P1 , P2 , Pr ) = P1 = 1 , P2 = 2 , Pr = . (37)
υ1 υ1 υ2
It has been shown in (Agustin et al., 2008) that the sum-average power constraint must satisfy
P1 υ1 + P2 υ1 + ςPr υ2 = P (38)
where ς = β 1 + β 2 so that δ1 + δ2 + β 1 + β 2 = 1. The data rates achieved on each link are
connected through
ζ ( R1 , R2 ) = R1 − κR2 ≤ 0 (39)
where κ is a positive real number accounts for the traffic asymmetry.
Under the sum-average power constraint the optimization problem for resource allocation is
convex and has a unique solution. However for maximum power constraint, the problem has
to be transformed into a convex one by introducing some auxiliary variables [see (Agustin et
al., 2008) and references therein].
5. Coded Bidirectional Relaying

So far we have discussed the TWRC from cooperative communication perspectives with a
major objective being compensation to make up for for the half-duplex loss. In a separate
but significant development, the authors in (Ahlswede et al., 2000) have proposed the concept
of network coding in which intermediate network nodes are allowed not only to route but
also to mix and code the incoming data from multiple links. This reduces the amount of data
transmissions in the network (thus improving the overall network throughput). Originally,
the network coding concept was proposed for wired communication networks. Later it was
applied to wireless communication networks by exploiting the broadcasting nature of wireless
medium [it was used for relay networks for the first time in (Hao et al., 2007)]. Network
coding has been proven to be a very effective solution to overcome the interuser interference in
wireless networks because of its ability to combining the different signals instead of separating
them from a traditional viewpoint.
Traditionally simultaneous transmission from T1 and T2 was avoided in order to simplify the
medium access control, and to avoid the interference at the relay R. Thereby four phases
were required to perform one round of information exchange between T1 and T2 through R.
However, by applying the idea of network coding, the authors in (Wu et al., 2005) proposed
a scheme to reduce the number of required phases from four to three as illustrated in Fig-
ure 9. In this scheme, T1 first transmits during first phase the message x1 to R consisting of
bits b1 (1), ..., b1 ( N ) with N denoting the message length in bits, which are decoded. During
the second phase, T2 transmits to R the message x2 consisting of bits b2 (1), ..., b2 ( N ), which
R decodes. In the third phase, R broadcasts to T1 and T2 a new message xr consisting of
bits br (n)’s, n = 1, ..., N, obtained by bit-wise exclusive-or (XOR) operation over b1 (n)’s and
b2 (n)’s i.e., br (n) = b1 (n) ⊕ b2 (n), ∀ n. Since T1 knows b1 (n)’s, T1 can recover its desired mes-
sage x2 by first decoding br (n) and then obtaining b2 (n)’s of x2 as b1 (n) ⊕ br (n), ∀n. Similarly,
T2 can recover x1 . The same type of three-phase coded bidirectional relaying scheme was
proposed independently in (Larsson et al., 2006). The resulting pre-log factor with respect to
the sum-rate of this three-phase coded scheme is thus 2/3 compared to 1/2 for conventional
half-duplex scheme. In this protocol, if a reliable direct link is possible, then the scheme may
gain in additional diversity and often better coverage as discussed in (Kim et al., 2008).
Fig. 9. Three-phase two-way relaying.
In (Popovski & Yomo, 2006); (Popovski & Yomo, 2007a); (Popovski & Yomo, 2007b), the au-
thors reduce the number of required phases from three to two by allowing T1 and T2 to trans-
mit simultaneously to R during the first phase, thereby eliminating the need for the second
phase. This corresponds to the MAC phase of DF-TWRC (Rankov & Wittneben, 2007). The
scheme proposed in (Katti et al., 2007a) is named as Analog Network Coding (ANC), while
that in (Zhang et al., 2006) is referred to as Physical-layer Network Coding (PNC). These
schemes differ in their relay operations, which are Amplify-and-Forward (AF) and Estimate-
and-Forward (EF), respectively. In ANC, R simply amplifies the mixed signal received simul-
taneously from T1 and T2 and then broadcasts it to both. By subtracting the back-propagating
self-interference, both T1 and T2 are able to receive their intended messages. Thus ANC
scheme is similar to the AF-TWRC (Rankov & Wittneben, 2007). Compared to ANC, PNC
(Zhang et al., 2006) performs more sophisticated operations than AF at R. Instead of decoding
messages x1 from T1 and x2 from T2 separately in two different phases like in (Wu et al., 2005),
the EF method estimates at R the bitwise XORs between b1 (n)’s and b2 (n)’s from the mixed
signal received, and re-encodes the decoded bits into a new broadcasting message xr . Each of
T1 and T2 then recovers the otherŠs message by the same decoding method discussed in (Wu
et al., 2005).
Although the schemes proposed in these works are similar to AF- and DF-TWRC, they are
inspired by network coding. The principle of network coding has been further investigated
for the TWRC in (Hausl & Hagenauer, 2006); (Baik & Chung, 2008); (Cui et al., 2008a); (Cui
et al., 2008b). It has been shown in (Katti et al., 2007b) that joint relaying and network cod-
ing achieves higher data rates as compared to routing at the relay. In (Kim et al., 2008), the
authors compared the different half-duplex bidirectional DF relaying protocols and derived
their performance bounds. Note that two-phase TWRC does not exploit spatial diversity ad-
vantage like conventional approach. Including the direct link would provide diversity gain
but at the cost of spectral efficiency. Even more recently, the authors in (Li et al., 2009) analyze
the outage performance of two-phase AF- and DF-TWRC under half-duplex constraint. They
derived the exact closed-form expressions for the outage probabilities by considering network
coding at the relay for DF case. Furthermore, they propose an adaptive bidirectional relaying
protocol which switches between AF and DF to minimize the outage probability of the sys-
tem. TWRC coupled with network coding is thus developing as a promising technology to
combat interference and to improve throughput in wireless networks.
6. MIMO Bidirectional Relaying

It is well known that Multiple-Input Multiple-Output (MIMO) communication systems have
the ability to enhance the channel capacity and link reliability without requiring an increase
in power or bandwidth. In (Unger & Klein, 2007) it is proposed to extend two-way relaying
to terminals with multiple antennas (leading to MIMO-TWRC). They investigate the average
performance of MIMO-TWRC by using multiple antennas at the ralay terminal only. The pro-
posed scheme exploits the fact that the relay R is a receiver as well as a transmitter in the
dual-hop case, and hence assumes CSI at the R is not unreasonable. Like in a Time Division
Duplex (TDD) system, CSI for receive and transmit processing can be obtained by directly esti-
mating the channel from T1 to the R and from T2 to the R in the first phase, and then exploiting
channel reciprocity in the second phase. Thereafter, the relay can perform spatial filtering to
its receive and transmit signal. In (Han et al., 2008) the average sum rate improvement of
two-way relaying is analyzed by deriving an upper and a lower bound for average sum rate
of two-way relaying which was not derived in (Rankov & Wittneben, 2007). They also extend
the work to the case when the source terminal and the destination terminal have two anten-
nas each and the relay has only one antenna (in order to implement Alamouti’s scheme). The
proposed scheme achieves higher average sum rate compared to the single antenna case, and
furthermore both the source and destination terminals achieve diversity of order two. The au-
thors in (Zhang et al., 2009) analyze the capacity region for of the ANC/AF-based TWRC with
linear processing (beamforming) at the relay with multiple antennas. They have also shown
that the ANC/AF-based TWRC have a capacity gain over the DF-based TWRC for sufficiently
large channel correlations and equal MAC and BC phase-durations.
In (Hammerstrom et al., 2007) the authors further extend the two-way relaying scheme of
(Rankov & Wittneben, 2007) to the case when multiple antennas are used at all terminals
(assuming the knowledge of transmit CSI at the DF relay). Figure 10 shows a set up for MIMO
two-way relaying where all the terminals are equipped with M > 1 antennas. It is assumed
Fig. 10. MIMO two-way relay channel.
that both T1 and T2 perfectly know (H1 , H2 , H1T and H2T ) in the receiving mode, but not in the
transmit mode. The relay R on the other hand has knowledge of both receive and transmit
CSI. This is a reasonable assumption since the relay has to estimate the channels (H1 and H2 )
for decoding in the first phase and exploting channel reciprocity in the second phase (like
in TDD system). Terminal T1 transmits vector x1 to T2 whereas T2 transmits vector x2 to T1
respectively in two phases. Frequency flat slow fading and perfect synchronization is assumed
between all terminals. The signal received at the relay in the first phase is given by
yr = H1 x1 + H2 x2 + nr (40)
where H1 and H2 are the M × M channel matrices (with each element being i.i.d. com-
plex Gaussian with zero mean and unit variance) that remain constant during the block
transmission, x1 and x2 are M × 1 symbol vectors with power P1 and P2 respectively, and
nr ∼ CN (0, σr2 I M ) is the M × 1 complex AWGN vector.
During first phase, the relay decodes the messages from both terminals T1 and T2 . Using
a Gaussian codebook, the achievable rates of both terminals are theoretically described by
the MIMO-MAC (Cover & Thomas, 1991), which imposes constraints on the individual first-
phase rates R1,I and R2,I , as well as the first-phase sum rate R1,I + R2,I for successful decoding
at the destination terminal:

P
R1,I ≤ I1,I = log I + 1 2 H1 H1H (41)
Mσr

P2 H

R2,I ≤ I2,I = log I + H2 H2 (42)
Mσr2

P1 H P2 H
M H1 H1 + M H2 H2
R1,I + R2,I ≤ I I = log I + . (43)
σr2
In the second phase, the relay applies bit-level XOR precoding on decoded messages. The
relay therefore broadcasts the vector xr with power Pr . The received signals at T1 and T2 are
given by
y1 = H1T xr + n1 (44)
y2 = H2T xr + n2 (45)
where n1 ∼ CN (0, σ12 I M ) and nr ∼ CN (0, σ22 I M ) are the complex AWGN M × 1 vectors at T1
and T2 respectively. Since both destinations have to be able to decode xr , the maximum data
rate in the second phase is given by
I I I = min { I1,I I , I2,I I } (46)

where

1 T
∗
I1,I I = log I + H Λ r H1 (47)
σ12 1

1 T
∗
I2,I I = log I + H Λ r H2 (48)
σ22 2

where Λr = E xr xrH and trace (Λr ) = Pr . The maximum sum-rate of this MIMO two-way
relaying scheme is given by (Hammerstrom et al., 2007)
1
Rsum = min { I I , min { I1,I , I2,I I } + min { I2,I , I1,I I }} . (49)
2
The above rate expression can be optimized by exploiting CSI knowledge at the relay subject
to the relay transmit power constraint. This is achieved by maximizing the data rate in the
second phase as follows
I I I,opt = max min { I1,I I , I2,I I } (50)

Λr
s.t. trace (Λr ) = Pr .
This optimization problem is independent of first phase data rates and can be solved by
semidefinite programming method by assuming Λr to be positive semidefinite [see (Ham-
merstrom et al., 2007) and references therein].
Fig. 11. Average sum-rate of two-antenna DF-TWRC with XOR precoding compared to one-
antenna case.
Figure 11 compares the average sum-rate obtained (assuming Gaussian codebook) for DF-
TWRC using XOR precoding with one and two antennas at the terminals T1 and T2 . Rayleigh
fading is assumed, and the elements of the channel matrices H1 and H2 are zero mean and
unit variance complex Gaussian random variables. All nodes use the same transmit power
P1 = P2 = Pr = P and are assumed to have the same noise variance σ12 = σ22 = σr2 = σ2 .
The SNR is defined as P/σ2 . We simulated 10000 random channels for each value in Figure
11. We observe that there is considerable improvement in sum-rate by using two antennas at
each node compared to the single antenna case.
Further, the authors in (Hammerstrom et al., 2007) compared two approaches of combining
the messages in the second phase at the relay (the superposition coding and the XOR pre-
coding) and showed that MIMO-TWRC achieves substantial improvement in spectral effi-
ciency compared to conventional relaying with or without transmit CSI at the relay. They also
showed that the difference in sum-rate compared to the case where no CSIT is used increases
with increasing ratio between number of relay antennas and number of node antennas. Also
XOR precoding always achieves higher minimum user rates than superposition coding if CSIT
is used. In (Oechtering & Boche, 2007) the authors propose transmit strategies in a MIMO
two-way DF relaying scenario with individual power constraints. The optimal relay transmit
strategy is given by two point-to-point water-filling solutions which are coupled by the re-
lay power distribution. The diversity-multiplexing trade-off analysis for the MIMO-TWRC is
dealt in (Gunduz et al., 2008). In (Yang & Chun, 2008), the transmission rate is improved by
using the generalized Schur decomposition-based MIMO-TWRC.
7. Bidirectional Relaying with Multiple Relays

This section extends the theory of single-relay two-way communication to one level up in
the network hierarchy by employing multiple relays. In two-way multiple relay channel two
terminals T1 and T2 exchange information with the help of M relay terminals in two phases.
Dedicated multiple relays can be utilized to relay copies of the transmitted information to
the destination such that each copy experiences independent channel fading, hence providing
diversity gain to the system. Such communication strategy is best suited for applications in
wireless ad-hoc networks, cellular scenarios, and wireless backhaul interconnections.
7.1 Distributed Space-Time Coding

The idea of space-time coding devised for MIMO systems can be applied to a wireless relay
network [see (Jing & Hassibi, 2006)] by having the relays that cooperate distributively. The
concept of distributed space-time coding (Jing & Hassibi, 2006) is investigated for two-way
multiple relay channel in (Cui et al., 2008c). The authors in (Cui et al., 2008c) propose a new
type of relaying scheme called partial DF for distributed TWRC where each relay removes
part of the noise before relaying information in the broadcast phase. They suggest two-way
relaying protocols using Linear Dispersion (LD) codes that operate over two time slots. In
this scheme, two terminals T1 and T2 communicate through multiple relays Ri , i = 1, ..., M.
Each half-duplex terminal is equipped with a single antenna. Terminal Tj , j ∈ [1, 2], transmits
the signal vector s j = [s j1 , ..., s jT ] T where s jt ∈ Am , t = 1, ..., T, Am is a finite constellation

with average power of unity, and T is the length of each time slot. Hence, E s jH s j = T.
The average power of terminal Tj is Pj and each relay has the equal power Pr /M so that the
total power of all the relays is Pr . The noise variance is assumed to be unity at every node.
During the first phase, both terminals T1 and T2 transmit their message to the relays. The
signal received by the ith relay is given by

yri = 2P1 h1i s1 + 2P2 h2i s2 + nri (51)
where h ji ∼ CN (0, 1) represents the channel gain between terminal Tj and relay Ri , nri is the
T × 1
vector representing the AWGN at the ith relay. The source transmit power is assumed
to be 2Pj because each source terminal transmits every two time slots. During second time
slot, the ith relay processes yri and transmits sri scaled by g to maintain average power Pr . The
signal received by jth terminal is given as
M
yj = ∑ gh ji sri + nj , j = 1, 2, (52)
i =1
where n j is the AWGN vector at the jth terminal.
7.1.1 2-AF
In this
scheme sri is obtained by precoding yri with a unitary matrix Wri and then scaled by
2Pr
g= M(2P1 +2P2 +1)
.The signal received at terminal T2 is given by

y2 = g 2P1 S1 h1− + 2P2 S2 h2− + z2 = g 2P1 H1 s1 + 2P2 H2 s2 + z2 (53)

where S j = [Wri s j , ..., WrM s j ], h1 = [ h11 h21 , ..., h1M h2M ] T , h2 = [h221 , ..., h22M ] T ,
z2 = g ∑iM M M 2
=1 h2i Wri nri + n2 , H1 = g ∑i =1 h1i h2i Wri , H2 = g ∑i =1 h2i Wri .
Since terminal T2 knows the back propagating signal s2 , the maximum-likelihood (ML)
decoding of s1 is obtained as (Cui et al., 2008c)
2

ŝ1 = arg min y2 − g 2P1 H1 s̃1 + 2P2 H2 s2 . (54)
s̃1 ∈A1M
Similarly, ML decoding of s2 is performed at terminal T1 . The disadvantage of this scheme is

that it amplifies the relay noise.
7.1.2 Partial DF I
This protocol overcomes the drawback of 2-AF scheme (Cui et al., 2008c) by allowing each
relay Ri to first decode s1 and s2 via the ML decoder
2

{ŝ1i , ŝ2i } = arg min yri − 2P1 h1i Wri s̃1 − 2P2 h2i Wri s̃2 . (55)
s̃1 ∈A1M ,s̃2 ∈A2M
The number of unknowns in the above equation is twice the number of equations, this making
the error probability high. For this reason it has been suggested in (Cui et al., 2008c) that
instead of sending ŝ1i and ŝ2i directly, each relay transmits

sri = Wri 2P1 h1i ŝ1i + 2P2 h2i ŝ2i (56)

Pr
scaled by g = M( P + P )
. Thus the relays remove noise from the received signal without
1 2
dealing with the channel effects. If Pr (∆s1i , ∆s2i ) represents the pairwise error probability at
the ith relay, where ∆s1i = s1 − ŝ1i and ∆s2i = s2 − ŝ2i , then the ML decoding at T2 is given as
(Cui et al., 2008c)
M 2
ŝ1 = arg max ∑ ∏ Pr(∆s1i , ∆s2i ) exp − y2 + y − y (57)
s̃1 ∈A1M ∆s1i ,∆s2i i =1
M √ √
where y = g ∑ Wri 2P1 h1i ∆s1i + 2P2 h2i ∆s2i and y = g 2P1 H1 s̃1 + 2P2 H2 s2 .
i =1
It is difficult to implement the above decoder directly when either number M or constellation
size is large. At high SNR, ∏iM =1 Pr ( ∆s1i , ∆s2i ) is dominated by ∆s1i = 0, ∆s2i = 0. Therefore
the ML decoding at terminal T2 can be approximated as follows
2

ŝ1 = arg min y2 − g 2P1 H1 s̃1 + 2P2 H2 s2 . (58)
s̃1 ∈A1M
Similarly, ML decoding at terminal T1 can be approximated (Cui et al., 2008c).
7.1.3 Partial DF II
In both AF and partial DF I schemes, a weighted sum of symbols is transmitted from two termi-
nals. This causes wastage of power since each destination knows the back-propagating signal.
In partial DF II (Cui et al., 2008c), components are superimposed via modular arithmetic. Let
the size of constellation A j be Zj with A j (q) representing the qth element of A j , where j = 1, 2
and q = 0, ..., Zj − 1. Consider u1 and u2 such that A1 (u1 ) = s1 and A2 (u2 ) = s2 . With the
setting Z = max { Z1 , Z2 }, it can be assume that Z1 ≥ Z2 without loss of generality. Under
this protocol, each relay obtains ŝ1i , ŝ2i from Equation (55) as in partial DF I. If A1 (û1i ) = ŝ1i
and A2 (û2i ) = ŝ2i then each relay transmits
sri = Wri A1 (mod (û1i + û2i , Z )) (59)

2Pr
where “mod” stand for the componentwise modular operation and g = gi = M . Since
fading channels are√considered, the probability
√ that there
√ exists a pair of vectors
√ { u1 , u2 } and
{û1 , û2 } such that 2P1 h1i A1 (u1 ) + 2P2 h2i A2 (u2 ) = 2P1 h1i A1 (û1 ) + 2P2 h2i A2 (û2 ) is
very small. It has been shownin (Cui et al., 2008) that the AF protocol achieves the diversity
loglogP
order min { M, T } 1 − logP , where P is the total power of the network whereas the partial
DF II protocol achieves a diversity order M when T ≥ M.
7.2 Distributed Relay Selection Scheme

A distributed relay selection strategy is proposed in (Ding et al., 2009) that selects the best
suited relay for realizing PNC in a dense relays network. In this transmission scheme, both
T1 and T2 broadcast their information to all M relays simultaneously during first phase. The
signal received at the ith relay Ri is given by
√ √
yri = Ph1i s1 + Ph2i s2 + nri (60)
where P is the source transmission power, s j represents the unit-power signal transmitted by
the jth source, and h ji represents the channel gain between the jth source and the ith relay. The
channel model for frequency flat Rayleigh fading is considered as
h̄ ji
h ji = (61)
dαji
where h̄ ji accounts for the channel fading characteristics due to the rich scattering environ-
ment, d ji represents the distance between the jth source and the ith relay, and α is the path loss
exponent. It is reasonable to assume that each relay terminal has its local channel information
under channel reciprocity condition. This local channel information can be exploited in realiz-
ing a distributed strategy of relay selection to improve the system performance. For instance,
consider that the relay Rb has been chosen as the best relay with corresponding channels h1b
and h2b . In the second phase the best relay Rb performs AF operation and transmits the mixed
signal given by
√ √
Ph1b s1 + Ph2b s2 + nrb √
srb = P (62)
P| h1b |2 + P| h2b |2 + σ2
to the two destinations. After removing the back-propagating self interference, the signal
received at jth terminal is given by
√
Ph jb √
yj = Phlb sl + nrb + n j . (63)
P| h1b |2 + P| h2b |2 + σ2
Therefore the mutual information between the lth source and jth destination is given by

γ2 | h1b |2 | h2b |2
Ijl = log 1 + , ∀ j = l & j, l ∈ [1, 2], (64)
2γ| h jb |2 + γ| hlb |2 + 1
where γ = P/σ2 represents the SNR. Relay selection (Ding et al., 2009) is performed in
medium access layer. It has been claimed that the two destinations have different prefer-
ences but they do not tend to contradict each other. The relay with channels yielding large I12
also has channels that give a large value for I21 , if not exactly the maximum. The best relay is
selected based on the following criterion
|h1b |2 |h2b |2
(65)
2γ| h1b |2 + γ| h2b |2 + 1
that maximizes the value of I12 . Then for this selected relay, the mutual information for sec-
ond source, I21 is determined. The relay selected in such a way is suboptimal for the second
source and hence some performance loss for the second source can be expected. The outage
probability for this scheme (Ding et al., 2009) at high SNR is
[(dαjb + 2dαlb )(2R − 1)] M

Pr [ Ijl < R] = . (66)
γM
It is clear from Equation (66) that the proposed transmission scheme in (Ding et al., 2009) has
the advantage of diversity of order M. While the authors in (Ding et al., 2009) dealt with the
relay selection scheme for the specific case of PNC-based TWRC, the problem is relevent in all
TWRC based links.
8. Summary and Future Directions

Cooperative relaying has evolved in recent years as a powerful tool to enhance the reliability
and throughput of wireless radio networks. The basic research challenge is to design spec-
trally efficient relaying schemes for better utilization of the available resources like power
and spectrum. In this chapter, we discussed several half-duplex cooperative relaying proto-
cols and their performances. Among these, two-way relaying is envisioned as a promising
protocol to save radio resources in wireless networks, whereby both up- and down-link are
transmitted on the same channel resources.
There are still many open issues related to the channels investigated so far. Mostly the slow
frequency flat fading scenarios has been considered in the literature, performance analysis for
fast as well as frequency selective fading two-way relay channels need to be addressed. The
theoretical capacity limits or achievable rate regions of TWRC still needs to be developed for
clustered and distributed scenarios. The reported protocols still suffer performance loss as
compared to the theoretical bounds. So better code designs with acceptable complexity need
to be urgently evolved to meet the above challenge. MIMO bidirectional relaying strategies
has already gained some momentum, but schemes like beamforming, distributed coding, and
relay selection still need to be explored. Perfect synchronization among multiple radios is
perhaps the most difficult task to perform for bidirectional traffic in a cooperative network.
Cooperative relaying techniques can be expected to be adopted in future wireless systems, as
it has been introduced in the IEEE 802.16j (WiMAX) standard. However, substantial research
efforts are needed to construct practical systems based on bidirectional cooperation for larger
wireless networks. Immense research interest is currently being focused to assess whether the
cooperation technology enables the implementation of cognitive radio.
9. References
Agustin, A.; Vidal, J. & Munoz, O. (2008). Protocols and resource allocation for the two-way
relay channel with half-duplex terminals. Submitted to IEEE Transactions on Wireless
Communications, Paper-TW-Oct-08-1427, Oct. 2008.
Ahlswede, R.; Cai, N.; Li, S.Y.R. & Yeung, R.W. (2000). Network information flow. IEEE Trans-
actions on Information Theory, Vol. 46, No. 4, July 2000, 1204-1216.
Avestimehr, A.S.; Sezgin, A. & Tse, D.N.C. (2008). Approximate capacity of the two-
way relay channel: A deterministic approach. Available at http://arxiv.org/PS-
cache/arxiv/pdf/0808/0808.3145v1.pdf, Aug. 2008.
Baik, I.J. & Chung, S.Y. (2008). Network coding for two-way relay channels using lattices.
Proceedings of IEEE International Conference on Communications (ICC), 3898-3902, June
2008.
Chen, M. & Yener, A. (2008). Multiuser two-way relaying for interference limited systems.
Proceedings of IEEE International Conference on Communications (ICC), 3883-3887, May
2008.
Cover, T.M. & Gamal, A.A.E. (1979). Capacity theorems for the relay channel. IEEE Transactions
on Information Theory, Vol. 25, No. 5, Sep. 1979, 572-584.
Cover, T.M. & Thomas, J.A. (1991). Elements of Information Theory, ISBN 0-471-06259-6, New
York: Wiley, 1991.
Cui, T.; Ho, T. & Kliewer, J. (2008a). Memoryless relay strategies for two-way relay channels:
Performance analysis and optimization. Proceedings of IEEE International Conference
on Communications (ICC), May 2008.
Cui, T.; Gao, F.F. & Tellambura, C. (2008b). Physical layer differential network coding for two-
way relay channels. Proceedings of IEEE Global Communications Conference (GLOBE-
COM), New Orleans, Nov. 2008, LA, USA.
Cui, T.; Gao, F.F.; Ho, T. & Nallanathan, A. (2008c). Distributed space-time coding for two-way
wireless relay networks. Proceedings of IEEE International Conference on Communica-
tions (ICC), Beijing, May 2008, China.
Ding, Z.; Leung, K.K.; Goeckel, D. & Towsley, D. (2009). On the study of network coding
with diversity. IEEE Transactions on Wireless Communications, Vol. 8, No. 3, Mar. 2009,
1247-1259.
Esli, C. & Wittneben, A. (2008). One- and two-way decode-and-forward relaying for wire-
less multiuser MIMO networks. Proceedings of IEEE Global Communications Conference
(GLOBECOM), New Orleans, Nov. 2008, LA, USA.
Gao, F.; Zhang, R. & Liang, Y.C. (2008). On channel estimation for amplify-and-forward two-
way relay networks. Proceedings of IEEE Global Communications Conference (GLOBE-
COM), New Orleans, Nov. 2008, LA, USA.
Gunduz, D.; Goldsmith, A. & Vincent Poor, H. (2008). MIMO two-way relay chan-
nel: diversity-multiplexing tradeoff analysis. Available at http://arxiv.org/PS-
cache/arxiv/pdf/0812/0812.3642v1.pdf, Dec. 2008.
Hammerstrom,I.; Kuhn, M.; Esli, C.; Zhao, J. ; Wittneben, A. & Bauch, G. (2007). MIMO two-
way relaying with transmit CSI at the relay. Proceedings of IEEE Workshop on Signal
Processing Advances in Wireless Communications (SPAW), June 2007, Finland.
Han, Y.; Ting, S.H.; Ho, C.K. & Chin, W.H. (2008). High rate two-way amplify-and-forward
half-duplex relaying with OSTBC. Proceedings of IEEE Vehicular Technology Conference
(VTC), Marina Bay, 2426-2430, May 2008, Singapore.
Hao, Y.; Goeckel, D.; Ding, Z.; Towsley, D.& Leung, K.K. (2007). Achievable rates for network
coding on the exchange channel. Proceedings of IEEE Military Communications Confer-
ence (MILCOM), Orlando, 371-382, Oct. 2007, Florida.
Hausl, C. & Hagenauer, J. (2006). Iterative network and channel decoding for the two-way
relay channel. Proceedings of IEEE International Conference on Communications (ICC),
1568-1573, June 2006.
Ho, C.K.; Zhang, R. & Liang, Y.C. (2008). Two-way relaying over OFDM: Optimized tone per-
mutation and power allocation. Proceedings of IEEE International Conference on Com-
munications (ICC), 3908-3912, May 2008.
Jing, Y. & Hassibi, B. (2006). Distributed space-time coding in wireless relay networks. IEEE
Transactions on Wireless Communications, Vol. 5, No. 12, Dec. 2006, 3524-3536, ISSN
1536-1276.
Jitvanichphaibool, K. ; Zhang, R. & Liang, Y.C. (2008). Optimal resource allocation for two-
way relay-assisted OFDMA. Proceedings of IEEE Global Communications Conference
Katti, S.; Gollakota, S. & Katabi, D. (2007a). Embracing wireless interference: Analog network
coding. Proceedings of Conference on Applications, Technologies, Architectures, and Proto-
cols for Computer Communications, 397-408, ISBN 978-1-59593-713-1, Kyoto, Feb. 2007,
Japan.
Katti, S.; Maric, I.; Goldsmith, A.; Katabi, D. & Medard, M. (2007b). Joint relaying and network
coding in wireless networks. Proceedings of IEEE International Symposium on Informa-
tion Theory, Nice, June 2007, France.
Kim, S.J.; Mitran, P. & Tarokh, V. (2008). Performance bounds for bi-directional coded coop-
eration protocols. IEEE Transactions on Information Theory, Vol. 54, No. 11, Nov. 2008 ,
5235-5241.
Laneman, J.N.; Wornell, G.W. & Tse, D.N.C. (2004). Cooperative diversity in wireless net-
works: Efficient protocols and outage behavior. IEEE Transactions on Information The-
ory, Vol. 50, No. 12, Dec. 2004, 3062-3080, ISSN 0018-9448.
Larsson, P.; Johansson, N. & Sunell, K.E. (2006). Coded bidirectional relaying. Proceedings of
IEEE Vehicular Technology Conference (VTC), 851-855, ISBN 1-7803-9392-9, Melbourne,
May 2006, Australia.
Li, Q.; Ting, S.H.; Pandharipande, A. & Han, Y. (2009). Adaptive two-way relaying and outage
analysis. IEEE Transactions on Wireless Communications, Vol. 8, No. 6, June 2009.
Meulen, E.C. (1971). Three-terminal communication channels. Adv. Appl. Probab., Vol. 3, No. 1,
120-154, 1971.
Nam, W.; Chung, S.Y. & Yong H.L. (2008). Capacity bounds for two-way relay channels. Pro-
ceedings of IEEE International Zurich Seminar on Communications, 2008.
Oechtering, T.J. & Boche, H. (2007). Optimal transmit strategies in multi-antenna bidirectional
relaying. Proceedings of IEEE Intern. Conf. on Acoustics, Speech, and Signal Processing
(ICASSP), Honolulu, Hawaii, Apr. 2007, USA.
Oechtering, T.J.; Schnurr, C.; Bjelakovic, I. & Boche, H. (2008). Broadcast capacity region of
two-phase bidirectional relaying. IEEE Transactions on Information Theory, Vol. 54, No.
1, Jan. 2008, 454-458.
Popovski, P. & Yomo, H. (2006). Bi-directional amplification of throughput in a wireless multi-
hop network. Proceedings of IEEE Vehicular Technology Conference (VTC), 588-593, ISBN
1-7803-9392-9, Melbourne, May 2006, Australia.
Popovski, P. & Yomo, H. (2007a). Wireless network coding by amplify-and-forward for bi-
directional traffic flows. Proc. IEEE Communication Letters, Vol. 11, No. 1, Jan. 2007,
16-18, ISSN 1089-7798.
Popovski, P. & Yomo, H. (2007b). Physical network coding in two-way wireless relay channels.
Proceedings of IEEE International Conference on Communications (ICC), 707-712, ISBN 1-
4244-0353-7, Glasgow, June 2007, Scotland.
Rankov, B. & Wittneben, A. (2006). Achievable rate regions for the two-way relay channel.
Proceedings of IEEE International Symposium on Information Theory, Seattle, 2006, WA.
Rankov, B. & Wittneben, A. (2007). Spectral efficient protocols for half-duplex relay channels.
IEEE Journal on Selected Areas in Communications, Vol. 25, No. 2, Feb. 1979, 379-389,
ISSN 0733-8716.
Sendonaris, A.; Erkip, E. & Aazhang, B.(2003). User cooperation diversity-Part I: System de-
scription. IEEE Transactions on Communications, Vol. 51, No. 11, Nov. 2003, 1927-1938,
ISSN 0090-6778.
Shannon, C.E. (1961). Two-way communication channels. Proc. 4th Berkeley Symp. Math. Stat.
and Prob., Vol. 1, 1961, 611-644, UC Press, Berkeley.
Unger, T. & Klein, A. (2007). Linear transceive filters for relay stations with multiple antennas
in the two-way relay channel. Proceedings of 16th IST Mobile and Wireless Communica-
tions Summit, Budapest, July 2007, Hungary.
Vaze,R. & Heath Jr., R.W. (2008). To code or not to code in multi-hop relay channels. available
at http://arxiv.org/abs/0805.3164, May, 2008.
Wu, Y.; Chou, P. & Kung, S.Y. (2005). Information exchange in wireless networks with net-
work coding and physical-layer broadcast. Proceedings of 39th Annual Conference on
Information Sciences and Systems (CISS), Baltimore, Mar. 2005, MD.
Yang, H.J. & Chun,J. (2008). Generalized Schur decomposition-based two-way relaying
for wireless MIMO systems. Proceedings of IEEE Global Communications Conference
Zhang, S.; Liew, S.C. & Lam, P.P. (2006). Physical-layer network coding. Proceedings of the 12th
Annual International Conference on Mobile Computing and Networking, 358-365, ISBN
1-59593-286-0, Los Angeles, CA, Sept. 2006, USA.
Zhang, R.; Liang, Y.C.; Chai, C.C. & Cui, S. (2009). Optimal beamforming for two-way multi-
antenna relay channel with analogue network coding. IEEE Journal on Selected Areas
in Communications, Vol. 27, No. 5, June 2009, 699-712.
Zhao, J.; Kuhn, M.; Wittneben, A. & Bauch, G. (2008). Self-interference aided channel estima-
tion in two-way relaying systems. Proceedings of IEEE Global Communications Confer-
ence (GLOBECOM), New Orleans, Nov. 2008, LA, USA.
A Novel Amplify-and-Forward Relay Channel Model
for Mobile-to-Mobile Fading Channels Under Line-of-Sight Conditions 307
A Novel Amplify-and-Forward Relay

Channel Model for Mobile-to-Mobile Fading
Channels Under Line-of-Sight Conditions
Batool Talha and Matthias Pätzold
University of Agder
Norway
Mobile-to-mobile (M2M) fading channels in cooperative networks can efficiently be modeled

using the multiple scattering concept. In this chapter, we propose a new second-order scat-
tering channel model referred to as the multiple-LOS second-order scattering (MLSS) channel
model1 for M2M fading channels in amplify-and-forward relay links under line-of-sight (LOS)
conditions, where the received signal comprises only the single and double scattered compo-
nents. In the proposed model, LOS components exist in the direct link between the source
mobile station and the destination mobile station as well as the link via the mobile relay. An-
alytical expressions are derived for the probability density function (PDF) of the envelope
and phase of M2M fading channels. It is shown mathematically that the proposed model in-
cludes as special cases double Rayleigh, double Rice, single-LOS double-scattering (SLDS),
non-line-of-sight (NLOS) second-order scattering (NLSS), and single-LOS second-order scat-
tering (SLSS) processes. The validity of all theoretical results is confirmed by simulations.
Our novel M2M channel model is important for the investigation of the overall system per-
formance in different M2M fading environments under LOS conditions.
1. Introduction
Among several emerging wireless technologies, M2M communications in cooperative
networks has gained considerable attention in recent years. The driving force behind
merging M2M communications and cooperative networks is its promise to provide a better
link quality (diversity gain), an improved network range, and an overall increase in the
system capacity. M2M cooperative wireless networks exploit the fact that single-antenna
mobile stations cooperate with each other to share their antennas in order to form a virtual
multiple-input multiple-output (MIMO) system in a multi-user scenario (Dohler, 2003). Thus,
in such networks, cooperative diversity (Laneman et al., 2004; Sendonaris et al., 2003a;b) is
achieved by relaying the signal transmitted from a mobile station to the final destination
using other mobile stations in the network. However, to cope with the problems faced within
the development of such systems, a solid knowledge of the underlying multipath fading
channel characteristics is essential. Therefore, the aim of this chapter is to develop a flexible
M2M fading channel model for relay-based cooperative networks and to analyze its statistical
1 The material in this chapter was presented in part at the 19th IEEE International Symposium on Per-
sonal, Indoor and Mobile Radio Communications, PIMRC 2008, Cannes, France, September 2008.
properties. This newly developed model would help communication system designers to
investigate the overall performance of cooperative communication systems.
So far, M2M amplify-and-forward relay fading channels have been modeled only for
some specific communication scenarios, either assuming NLOS or partial LOS propagation
conditions. It has been shown in (Patel et al., 2006) that under NLOS propagation conditions,
the M2M amplify-and-forward relay fading channel can be modeled as a double Rayleigh
fading channel (Erceg et al., 1997; Kovacs et al., 2002). Motivated by the studies of double
Rayleigh fading channels for keyhole channels (Almers et al., 2006), the so-called double
Nakagami-m fading channel model has also been proposed in (Shin & Lee, 2004). Further-
more, in amplify-and-forward relay environments, the M2M fading channel under LOS
conditions can be modeled as a double Rice fading channel (Talha & Pätzold, 2007a) and/or
as an SLDS fading channel (Talha & Pätzold, 2007b). In addition, M2M fading channels
in cooperative networks can efficiently be modeled using the multiple scattering concept
(Andersen, 2002).
In multiple scattering radio propagation environments, the received signal comprises a

sum of the single, double, or generally multiple scattered components. In this chapter, we
model the amplify-and-forward relay fading channel as a second-order scattering channel,
i.e., the sum of only the single and the double scattered components (Salo et al., 2006). The
novelty of this approach is that we have extended the NLSS (Salo et al., 2006) and the SLSS
(Salo et al., 2006) channel models to an MLSS channel model (Talha & Pätzold, 2008b) by
incorporating multiple LOS components in all transmission links, i.e., in the direct link
between the source mobile station and the destination mobile station as well as in the link
via the mobile relay. Furthermore, an important feature of the proposed MLSS channel
model for M2M fading channels is that it includes several other well-known channel models
as special cases, e.g., the double Rayleigh model, the double Rice model, the SLDS model,
the NLSS model, and the SLSS model. Here, we derive an analytical expression of the
PDF of the MLSS process, along with the PDF of the corresponding phase process. The
correctness of all theoretical results would be confirmed using a high-performance channel
simulator. Furthermore, all presented results provide evidence that the statistics of MLSS
fading channels are entirely different from the special cases discussed above.
The chapter is structured as follows: In Section 2, the reference model for the amplify-
and-forward MLSS fading channel is developed. Section 3 deals with the analysis of the
statistical properties of MLSS fading processes. Section 4 confirms the validity of the
analytical expressions presented in Section 3 by simulations. Finally, concluding remarks are
made in Section 5.
2. The MLSS Fading Channel

Under NLOS propagation conditions, the complex time-varying channel gain of the multiple
scattering radio propagation channel proposed in (Andersen, 2002) can be written as
χ ( t ) = α 1 µ (1) ( t ) + α 2 µ (2) ( t ) µ (3) ( t ) + α 3 µ (4) ( t ) µ (5) ( t ) µ (6) ( t ) + · · · (1)
where µ(i) (t) (i = 1, 2, 3, . . .) is a zero-mean complex Gaussian process that represents the
scattered component of the ith link and αi (i = 1, 2, 3, . . .) is a real-valued constant that
determines the contribution of ith scattered component. In (1), Gaussian processes µ(i) (t) are
mutually independent. However, when the fading channel is modeled by taking into account
only the first two terms of (1), the resulting channel is referred to as the NLSS channel (Salo
et al., 2006). Here, we are presenting an extension of the NLSS channel to the MLSS channel
by incorporating LOS components in a novel manner for M2M amplify-and-forward relay
fading channels. The considered communication scenario determined by a source mobile
station, a destination mobile station and mobile relay is shown in Fig. 1.
PU t
1
PU t PU t
2 3
Source Destination
mobile station mobile station
Mobile relay
Fig. 1. The propagation scenario behind MLSS fading channels.
(i )
Starting from (1), ignoring µ(i) (t) ∀ i ≥ 4, and replacing µ(i) (t) by µρ (t) for i = 1, 2, 3,
results in
(1) (2) (3)
χρ (t) = µρ (t) + AMR µρ (t) µρ (t) (2)
where α1 = 1 and α2 = AMR . The quantity AMR in (2) is referred to as the relay gain. Since we
have assumed fixed gain relays in our system, it follows that AMR is a real constant. Further-
(1) (2) (3)
more, in (2), µρ (t), µρ (t), and µρ (t) are statistically independent non-zero-mean complex
Gaussian processes, which model the individual M2M fading channel in the source mobile sta-
tion to the mobile relay, the mobile relay to the destination mobile station, and the source mo-
bile station to the destination mobile station links (see Fig. 1). Each complex Gaussian process
(i ) (i ) (i )
µρ (t) = µρ1 (t) + jµρ2 (t) represents the sum of the scattered component µ(i) (t) and the LOS
(i )
component m(i) (t), i.e., µρ (t) = µ(i) (t) + m(i) (t). The scattered component µ(i) (t) is still
(i ) (i )
modeled by a zero-mean complex Gaussian process µ(i) (t) = µ1 (t) + jµ2 (t) with variance
2σ2 . The LOS component m(i) (t) = ρ e j(2π f ρi t+θρi ) assumes a fixed amplitude ρ , a constant
i i i
Doppler frequency f ρi , and a constant phase θρi for i = 1, 2, 3. Furthermore, it is obvious that
(2) (3)
the product term µρ (t) µρ (t) in (2) is a non-zero-mean complex double Gaussian process,
(2) (3)
i.e., µρ (t) = ς ρ (t) = ς ρ1 (t) + jς ρ2 (t). Furthermore, ς ρ (t) models the overall fad-
(t) µρ
ing in the source mobile station to the destination mobile station link via the mobile relay. It
should also be noted here that the relay gain AMR would just scale the mean value and variance
(3) (3)
of the complex Gaussian process µ (t), i.e., m(3) (t) = E{ A µ (t)} = ρ
ρ e j(2π f ρ3 t+θρ3 )
MR ρ AMR
(3)
where ρ AMR = AMR ρ3 , and 2σA 2
MR
= Var{ AMR µρ (t)} = 2 ( AMR σ3 )2 . Finally, the overall fad-
ing process consisting of the direct link between the source mobile station and the destination
mobile station as well as the source mobile station to the destination mobile station link via the
mobile relay results in the complex process χρ (t) = χρ1 (t) + jχρ2 (t) given in (2). The abso-
lute value of χρ (t) defines the MLSS process, i.e., Ξ (t) = |χρ (t) |. Furthermore, the argument
of χρ (t) introduces the phase process Θ (t), i.e., Θ (t) = arg{χρ (t)}.
3. Statistical Analysis Of the MLSS Fading Channel

In this section, we derive the analytical expressions for the statistical properties of MLSS
fading channels introduced in Section 2. The starting point for the derivation of the an-
alytical expression of the PDF of MLSS fading channels, as well as the PDF of the corre-
sponding phase process is the computation of the joint PDF pχρ1 χρ2 χ̇ρ1 χ̇ρ2 (u1 , u2 , u̇1 , u̇2 ; t) of
the stochastic processes χρ1 (t), χρ2 (t), χ̇ρ1 (t), and χ̇ρ2 (t) at the same time t. Throughout
this chapter, the overdot indicates the time derivative. Equation (2) shows that the joint PDF
pχρ1 χρ2 χ̇ρ1 χ̇ρ2 (u1 , u2 , u̇1 , u̇2 ; t) can be written in terms of a 4-dimensional (4D) convolution in-
tegral as
∞ ∞ ∞ ∞
pχρ1 χρ2 χ̇ρ1 χ̇ρ2(u1 , u2 , u̇1 , u̇2 ; t) = p (1) (1) (1) (1) (u1 − y1 , u2 − y2 , u̇1 − ẏ1 , u̇2 − ẏ2 ; t)
µρ1 µρ2 µ̇ρ1 µ̇ρ2
−∞−∞−∞−∞
pς ρ1 ς ρ2 ς̇ ρ1 ς̇ ρ2(y1 , y2 , ẏ1 , ẏ2 ; t) dẏ2 dẏ1 dy2 dy1 (3)
(1) (1)
where p (1) (1) (1) (1) (u1 , u2 , u̇1 , u̇2 ; t) is the joint PDF of the processes µρ1 (t), µρ2 (t),
(1) (1)
µ̇ρ1 (t), and µ̇ρ2 (t)
at the same time t. Similarly, pς ρ1 ς ρ2 ς̇ ρ1 ς̇ ρ2(y1 , y2 , ẏ1 , ẏ2 ; t) in (3), rep-
resents the joint PDF of the processes ς ρ1 (t), ς ρ2 (t), ς̇ ρ1 (t), and ς̇ ρ2 (t) at the same time
(1) (1)
t. It is worth mentioning here that the processes µρi (t), µ̇ρi (t), ς ρi (t), and ς̇ ρi (t)
(1) (1)
(i = 1, 2) are uncorrelated in pairs. Furthermore, the process pairs {µρi (t) , µ̇ρi (t)}
and {ς ρi (t) , ς̇ ρi (t)} (i = 1, 2) are statistically independent, which allows us to ex-
press p (1) (1) (1) (1)( y1 , y2 , ẏ1 , ẏ2 , u1 , u2 , u̇1 , u̇2 ; t ) as given in (3). The joint PDF
ς ς ς̇ ς̇ µ µ µ̇ µ̇
ρ1 ρ2 ρ1 ρ2 ρ1 ρ2 ρ1 ρ2
p (1) (1) (1) (1)
(u1 , u2 , u̇1 , u̇2 ; t) can be obtained using the multivariate Gaussian distribution
(see, e.g., (Simon, 2002, Eq. (3.2))). The expression for the joint PDF pς ρ1 ς ρ2 ς̇ ρ1 ς̇ ρ2(y1 , y2 , ẏ1 , ẏ2 ; t)
is presented in (4), where the quantity β i (i = 2, 3) is the negative curvature of the autocorrela-
tion function of the inphase and quadrature components of µ(i) (t) (i = 2, 3). Under isotropic
scattering conditions, β i (i = 2, 3) can be expressed as (Akki, 1994; Pätzold, 2002)
pς ρ1 ς ρ2 ς̇ ρ1 ς̇ ρ2 (y1 , y2 , ẏ1 , ẏ2 ; t) =
2 2 2 2
(3) (3) 2 2 2 (3) (3) (2) (2)
z1 − m
1
( t ) + z2 − m2 ( t ) (3) (3) β 2 ( z2 + z2 ṁ
ṁ (t) + ṁ2 (t) 1 2) 1
(t) + ṁ2 (t) + β 3 (y21 +y22 ) ṁ1 (t)+ṁ2 (t)
1
∞ ∞ −
2σ2
AMR − 2β 3 2
e e 2β 2 β 3 (y2 +y22 )+ β 2 (z2 +z22 )
1 1
dz2 dz1
e
2
2
−∞ −∞ (2π )3 σ22 σA MR
β 3 y 1 + y 2 + β z2 + z2 2
2 2 1 2
2 2
(2) (2) (2) (2) (2) (2) (3) (2) (2)
y2 + y2 m ( t ) + m2 ( t )
1 2 1 (z21 +z22 ) 1
−2 y2 m2 (t)+y1 m1 (t) z1 −2 y2 m1 (t)−y1 m2 (t) z2 + z1 ṁ2 (t) −y2 ẏ1 +y1 ẏ2 −ṁ2 (t)y1 z1 +ṁ (t)y2 z1
− 2
2σ22 (z2 +z2 β 3 (y2 +y22 )+ β 2 (z2 +z22 )
1 2) e 1 1

×e
2 2
(2) (2) (2) (2) (2) (2) (3) (2) (2)
ẏ2 +ẏ2 z1 ṁ
1 2
ṁ (t) + ṁ2 (t) (z21 +z22 )
1 1
−2 ẏ2 ṁ2 (t)+ẏ1 ṁ1 (t) z1 −2 ẏ2 ṁ1 (t)−ẏ1 ṁ2 (t) z2 + (t) y1 ẏ1 +y2 ẏ2 −ṁ1 (t)y1 z1 +ṁ2 (t)y2 z1
2
2β 2 (z2 +z2
−
β 3 (y2 +y22 )+ β 2 (z2 +z22 )
1 2) 1 e 1

×e
(3) (2) (2) (3) (2) (2)
2β 2 z2 (z2 +z2
1 2) 1 1 1
ṁ (t) y2 ẏ1 −y1 ẏ2 +2ṁ2 (t)y1 z1 −2ṁ (t)y2 z1 +ṁ2 (t) y1 ẏ1 +y2 ẏ2 −2ṁ (t)y1 z1 −2ṁ2 (t)y2 z1 + β 3 (y21 +y22 )(ẏ21 +ẏ22 )

2 2 2 2 2 2 2
2β 2 (z +z2 ) β 3 (y +y2 )+ β 2 (z +z2 )
1 1 1

for Mobile-to-Mobile Fading Channels Under Line-of-Sight Conditions
×e
(2) (2) (2) (2)
(2) (3) (2) (3) (2) (3) (2) (3) β 3 ( y2 + y2
1 2)
−2ẏ2 ṁ2 (t)z1 +ṁ1 (t)z2 +2ẏ1 −ṁ1 (t)z1 +ṁ2 (t)z2
1 1 1 1
z22 ṁ (t)ṁ (t)y1 +ṁ2 (t)ṁ2 (t)y1 +ṁ2 (t)ṁ (t)y2 −ṁ (t)ṁ2 (t)y2
2 2
+ β z2 + z2 2β 2 (z2 +z2 β y2 + y2 + β z2 + z2
β 3 ( y2 + y2
1 2) 2( 1 2) 1 2) 3( 1 2) 2( 1 2)
×e e .
(4)
311

312
2 2 2 2 2 2
(2) (2) (3) (3) (1) (1)
x2 ṁ (t) + ṁ2 (t) ṁ (t) + ṁ2 (t) ṁ (t) + ṁ2 (t)
1 1 1
2σ2
−
1
− 2β 2 − 2β 3 − 2β 1
x2 e e e e
pΞΞ̇ΘΘ̇ x, ẋ, θ, θ̇; t =
2
(2π )4 σ12 σ22 σA MR
2 2 2 2
(3) (3) (1) (1)
1
z1 − m ( t ) + z2 − m2 ( t ) y1 + m ( t ) + y2 + m2 ( t )
1 (1) (1)
r y1 +m (t) cos θ +r y2 +m2 (t) sin θ
∞ ∞ ∞ ∞ 2σ2 1
−
2σ2
−
e AMR
e 1
σ2
dz2 dz1 dy2 dy1 e 1
2
−∞−∞−∞−∞ β 1 z21 + z22 + β 3 y21 + y22 + β 2 z21 + z22
2 2
2 2 (1) (1)
(2) (2) (2) (2) (2) (2) β 3 (y2 +y22 )
1
ṁ (t) + ṁ2 (t)
1
y2 + y2
1 2
m ( t ) + m2 ( t ) (z21 +z22 )
1
−2 y2 m2 (t)+y1 m1 (t) z1 −2 y2 m1 (t)−y1 m2 (t) z2 +
2
2β 1 β 1 (z2 +z22 )+ β 3 (y2 +y22 )+ β 2 (z2 +z22 )
1 1 1
−
2σ22 (z2 +z2
1 2) e

×e
2 2 2 2 2 2
2 (1) (1) (3) (3) (2) (2)
β22 (z2 +z2 β3 ṁ (t) + ṁ (t) + β1 ṁ (t) + ṁ (t) + β3 β1 ṁ (t) + ṁ (t)
1 2) 1 2 1 2 1 2 [ β3 (y21 +y22 )+ β1 (z21 +z22 )]

2
2β 3 β 2 β 1 β 1 (z2 +z2 + β y2 + y2 + β z2 + z2
1 2) 3( 1 2) 2( 1 2)

×e
(2) (3) (2) (3) (2) (3) (2) (3) (1) (2) (1) (2) (1) (2) (1) (2)
1 1 1 1 1 1 1 1
z22 ṁ (t)ṁ (t)y1 +ṁ2 (t)ṁ2 (t)y1 +ṁ2 (t)ṁ (t)y2 −ṁ (t)ṁ2 (t)y2 −z1 ṁ (t)ṁ (t)+ṁ2 (t)ṁ2 (t) −z2 ṁ2 (t)ṁ (t)−ṁ (t)ṁ2 (t)
2
β 1 ( z2 + z2 + β y2 + y2 + β z2 + z2
1 2) 3( 1 2) 2( 1 2)

×e
(2) (3) (2) (3) (2) (3) (2) (3) (1) (2) (1) (2) (1) (3) (1) (3)
1 1 1 1 1 1 1 1 1
z2 2z1 ṁ2 (t)ṁ (t)y1 −ṁ (t)ṁ2 (t)y1 −ṁ (t)ṁ (t)y2 −ṁ2 (t)ṁ2 (t)y2 −z2 ṁ2 (t)ṁ (t)−ṁ (t)ṁ2 (t) + ṁ2 (t)ṁ (t)−ṁ (t)ṁ2 (t) y1
2
β 1 ( z2 + z2 + β y2 + y2 + β z2 + z2
1 2) 3( 1 2) 2( 1 2)

×e
(2) (3) (2) (3) (2) (3) (2) (3) (1) (3) (1) (3) (1) (3) (1) (3)
1 1 1 1 1 1 1 1 1
z2 ṁ (t)ṁ2 (t)y2 −ṁ (t)ṁ (t)y1 −ṁ2 (t)ṁ2 (t)y1 −ṁ2 (t)ṁ (t)y2 −z1 ṁ (t)ṁ (t)y1 +ṁ2 (t)ṁ2 (t)y1 +ṁ2 (t)ṁ (t)y2 −ṁ (t)ṁ2 (t)y2
2
β 1 ( z2 + z2 + β y2 + y2 + β z2 + z2
1 2) 3( 1 2) 2( 1 2)

×e
(1) (2) (1) (2) (3) (3) (1) (2) (2)
1 1 1 1 1 1
z3 ṁ (t)ṁ (t)+ṁ2 (t)ṁ2 (t) + ṁ (t)(y2 z1 −y1 z2 )+ṁ2 (t)(y1 z1 +y2 z2 )+ ṁ2 (t)+ṁ2 (t)z1 +ṁ (t)z2 (z2 +z22 ) ( x θ̇ cos θ + ẋ sin θ )
2
+ β y2 + y2 + β z2 + z2
−
β 1 ( z2 + z2
1 2) 3( 1 2) 2( 1 2)

×e
(3) (3) (1) (2) (2) (1) (3) (1) (3)
1 1 1 1 2) ( 1 1
ṁ2 (t)(y1 z2 −y2 z1 )+ṁ (t)(y1 z1 +y2 z2 )+ ṁ (t)+ṁ (t)z1 −ṁ2 (t)z2 (z2 +z2 ẋ cos θ − x θ̇ sin θ )− ṁ2 (t)ṁ2 (t)+ṁ (t)ṁ (t) y2 z2
2
β 1 ( z2 + z2 + β y2 + y2 + β z2 + z2
1 2) 3( 1 2) 2( 1 2)

×e
2 2
(3) (3)
β 1 ( z2 + z2 ṁ (t) + ṁ2 (t)
(z21 +z22 ) ẋ2 +( xθ̇ )2 1 2) 1

2 2
−
{ }
2 β 1 ( z2 + z2 + β y2 + y2 + β z2 + z2
1 2) 3( 1 2) 2( 1 2)
2β 3 β 1 (z2 +z2 + β y2 + y2 + β z2 + z2
1 2) 3( 1 2) 2( 1 2)
×e e , x ≥ 0, | ẋ | < ∞, |θ | ≤ π, |θ̇ | < ∞ . (5)

β2 = 2 (σ2 π )2 f max
2
1
2
+ f max2
(5a)

β3 = 2 (σAMR π )2 f max
2
2
2
+ f max 3
. (5b)
The symbols f max1 , f max2 , and f max3 appearing in (5a) and (5b) denote the maximum Doppler
frequency caused by the motion of the source mobile station, the mobile relay, and the des-
tination mobile station, respectively. Substituting the joint PDF p (1) (1) (1) (1) (u1 , u2 , u̇1 , u̇2 )
and (4) in (3), applying the concept of transformation of random variables (Papoulis
& Pillai,
2002), and doing tedious algebraic manipulations, the joint PDF pΞΞ̇ΘΘ̇ x, ẋ, θ, θ̇; t of the pro-

cesses Ξ (t), Ξ̇ (t), Θ (t), and Θ̇ (t) can be derived. The resulting joint PDF pΞΞ̇ΘΘ̇ x, ẋ, θ, θ̇; t
is presented in (5), which is of fundamental importance, because it provides the basis for
the computation of the PDF, the level-crossing rate (LCR), and the average duration of fades
(ADF) of MLSS processes Ξ (t) as well as the PDF of the phase process Θ (t). Using (5), the
analytical expressions of the LCR and the ADF of the MLSS processes Ξ (t) have been derived
2
in (Talha & Pätzold, 2008a). In (5), the quantity β 1 is given by β 1 = 2 (σ1 π )2 f max 1
2
+ f max 3
(Akki, 1994; Pätzold, 2002).
3.1 PDF of the MLSS Process

The joint PDF pΞΘ ( x, θ; t) of the MLSS process Ξ (t) and the phase process Θ (t) can be ob-
tained by solving the integrals over the joint PDF pΞΞ̇ΘΘ̇ x, ẋ, θ, θ̇; t according to
∞ ∞
pΞΘ ( x, θ; t) = pΞΞ̇ΘΘ̇ x, ẋ, θ, θ̇; t dθ̇ d ẋ (6)
−∞ −∞
for x ≥ 0 and |θ | ≤ π. Solving (6) results in the following expression

x2 ν2 + ρ2
2 ∞ ∞π (ω/ν)2 +ρ2 AMR g (ω,ψ;t)
x e 2σ1 ω − 2σ2 2 − 2σA2 − 1 2 xg (ω,θ,ψ;t)
− 3 σ
pΞΘ ( x, θ; t) = 2
e 2 e MR e
2σ
1 e 1
(2π ) σ12 σ22 σA
2 ν
MR 0 0 −π

× I0 g2 (ω, ν, ψ; t) dψ dω dν, x ≥ 0, |θ | ≤ π (7)
where

g1(ω, ψ; t) = ω 2 + ρ21 + 2ρ1 ω cos ψ − 2π f ρ1 t − θρ1 (8a)
2 2
ρ2 ω ρ AMR ν 2ρ2 ρ ω
g2(ω, ν, ψ; t) = 2 + 2 + 2 A2MR
σ2 ν σAMR σ2 σAMR

cos ψ − 2π f ρ2+ f ρ3 t − θρ2+ θρ3 (8b)

ρ1 cos θ − 2π f ρ1 t − θρ1 + ω cos (θ − ψ)
g3(ω, θ, ψ; t) = .
σ1
(8c)
In (7), I0 (·) is the zeroth-order modified Bessel function of the first kind (Gradshteyn &
Ryzhik, 2000).
The PDF pΞ ( x ) of the MLSS fading process Ξ (t) can be obtained by integrating (7)
over θ in the interval [−π, π ]. Hence,
x2 ν2 + ρ2
− 2 ∞∞ (ω/ν)2 +ρ22 AMR
x e 2σ1 ω − −
2σ22 2σ2
pΞ ( x ) = dν dω e e AMR
2π σ12 σ22 σA
2 ν
MR 0 0
π

g4 (ω,ψ)
−
2σ2
x
dψ e 1 I0 g 4(ω, ψ) I0 g 5 (ω, ν, ψ) ,x≥0 (9)
σ12
−π
where
g4 (ω, ψ) = ω 2 + ρ21 + 2ρ1 ω cos (ψ) (10a)

2 2
ρ2 ω ρ AMR ν 2ρ2 ρ ω
g5 (ω, ν, ψ) = 2
+ 2
+ 2 A2MR cos (ψ). (10b)
σ2 ν σAMR σ2 σAMR
It is worth mentioning that the joint PDF pΞΘ ( x, θ; t) in (6) is dependent on time t. Neverthe-
less, the PDF pΞ ( x ) in (9) is independent of time t showing that MLSS processes Ξ (t) are first
order stationary. From the PDF pΞ ( x ) of MLSS fading processes Ξ (t), the following special
cases can be obtained.
Substituting AMR = 1, ρ1 = ρ2 = ρ3 = 0, and taking the limit σ12 → 0 in (9), reduces

the PDF of MLSS processes to the PDF of double Rayleigh processes (see, e.g., (Kovacs et al.,
2002; Patel et al., 2006))

x x

pΞ ( x ) AMR =1 = 2 2 K0 , x ≥ 0. (11)
ρ1 ,ρ2 ,ρ3 =0 σ2 σAMR σ2 σAMR
σ 2 →0
1
For the special case when AMR = 1, ρ1 = 0, and σ12 → 0, the PDF of MLSS processes given in
(9) reduces to the PDF of double Rice processes (Talha & Pätzold, 2007a)

ν2 + ρ2
∞ ( x/ν)2 +ρ22
− 2 MR
A
x 1 − 2 2σ xρ2 νρ AMR

pΞ ( x ) AMR =1 = 2 2 e 2σ2
e AMR
I0 I0 dν, x ≥ 0 . (12)
ρ1 =0 σ2 σAMR ν νσ22 2
σA
σ 2 →0 0 MR
1
Similarly, substituting AMR = 1, ρ2 = ρ3 = 0, and σ12 → 0 in (9) allows us to write the PDF of
SLDS processes in the form (Talha & Pätzold, 2007b)


 σ2 σx2 I0 σ2 σxA
 ρ
K0 σ2 σA1 , x < ρ1
pΞ ( x ) AMR =1 = 2 AMR MR MR
(13)
 σ2 σx2 K0 σ2 σxA
 ρ
ρ2 ,ρ3 =0 I0 σ2 σA1 , x ≥ ρ1 .
σ 2 →0 2 AMR MR MR
1
The PDF of NLSS processes (see, e.g., (Salo et al., 2006)) can be derived from the PDF of MLSS
processes by substituting ρ1 = ρ2 = ρ3 = 0 in (9), i.e.,
2
− x 2 ∞ 2
x e 2σ
1 − ω2 ω ω
pΞ ( x ) AMR =1 = 2 2 2 ω e 1 K0
2σ
I0 2 x dω, x ≥ 0 . (14)
ρ1 ,ρ2 ,ρ3 =0 σ1 σ2 σA σ2 σAMR σ1
MR 0
Finally, solving (9) for ρ2 = ρ3 = 0, the PDF of SLSS processes (Salo et al., 2006) is obtained as
∞π
x
2 g ( x,θ )
− ω2 − 5 2 ω ω

pΞ ( x ) AMR =1 = dθ dω ω e 2σ
1 e 2σ
1 K0 I0 2 g6(x, θ) , x ≥ 0 (15)
ρ2 ,ρ3 =0 2π σ12 σ22 σA
2 σ2 σAMR σ1
MR 0 −π
where
g6 ( x, θ ) = x2 + ρ21 − 2xρ1 cos θ . (16)
3.2 PDF of the Phase Process

Integrating (7) over x in the interval [ 0, ∞ ) results in the following expression for the PDF
pΘ (θ; t) of the phase process Θ (t)
ν2 + ρ2
(ω/ν)2 +ρ22 AMR
− −
∞∞π 2σ22 2σ2 g1 (ω,ψ;t)
ωe e AMR −
2σ2
π
pΘ (θ; t) = e 1 I0 g2 (ω, ν, ψ; t) 1+ g (ω, θ, ψ; t)
ν 2 2 2
(2π ) σ2 σAMR 2 3
0 0 −π

1 2 g (ω, θ, ψ; t)
× e 2 g3 (ω,θ,ψ;t) 1 + Φ 3 √ dψ dω dν, |θ | ≤ π (17)
2
where g1 (·, ·; t), g2 (·, ·, ·; t), and g3 (·, ·, ·; t) are defined in (8a), (8b), and (8c), respectively.
In (17), Φ (·) represents the error function (Gradshteyn & Ryzhik, 2000, Eq. (8.250.1)).
From (17), it is obvious that the phase process Θ (t) is not strict sense stationary because
the density pΘ (θ; t) is a function of time t. This time dependency is due the Doppler
frequency f ρi of the LOS component m(i) (t) (i = 1, 2, 3). However, for the special case when
f ρi = 0, ρi = 0 (i = 1, 2, 3), the phase process Θ (t) becomes first order stationary.
The PDF pΘ (θ; t) of the phase process Θ (t) given in (17) reduces to the PDF of the
phase process corresponding to double Rayleigh (Talha & Pätzold, 2007a), double Rice (Talha
& Pätzold, 2007a), SLDS (Talha & Pätzold, 2007b), NLSS, and SLSS processes by selecting ρ1 ,
ρ2 , ρ3 , and σ12 in a similar fashion as described in Subsection 3.1.
In this section, we will provide sufficient evidence to support the validity of the analytical
expressions presented in Section 3 with the help of simulations. Furthermore, for the sake
of completeness, a detailed comparison of the PDF pΞ ( x ) of MLSS processes Ξ (t) to that of
the special cases mentioned above will be presented. Similarly, a comparison of the PDF
pΘ (θ ) of the phase process Θ (t) with other phase PDFs will also be presented. The concept
of sum-of-sinusoids (Pätzold, 2002) was employed to simulate the underlying uncorrelated
Gaussian noise processes of the overall MLSS process Ξ (t). The number of sinusoids required
to simulate the inphase and quadrature components of Gaussian processes µ(i) (t) was
selected to be 20 and 21, respectively. Furthermore, the simulation model parameters were
computed using the generalized method of exact Doppler spread (GMEDS1 ) (Pätzold &
Hogstad, 2006). The maximum Doppler frequencies, i.e., f max1 , f max2 , and f max3 were set to
91 Hz, 75 Hz, and 110 Hz, respectively. Furthermore, for simplicity, the amplitudes of the
three LOS components ρ1 , ρ2 , and ρ3 are assumed to be equal, i.e., ρ1 = ρ2 = ρ3 = ρ. The
quantities σ12 , σ22 , σ32 , and the relay gain AMR are set to 1, unless stated otherwise.
The results presented in Figs. 2 and 3 show a good fitting between the analytical and
simulation results. In Fig. 2, the PDF pΞ ( x ) of MLSS processes Ξ (t) is compared to that
of classical Rayleigh, classical Rice, double Rayleigh, double Rice, SLDS, NLSS, and SLSS
processes for ρ = 1, where f ρ1 , f ρ2 , and f ρ3 , were set to 166 Hz, 185 Hz, and 201 Hz,
respectively. It can be observed that the maximum value of the PDF pΞ ( x ) of MLSS processes
Ξ (t) is lower than that of all the other processes under consideration for the same value of
ρ. However, the PDF pΞ ( x ) of MLSS processes Ξ (t) has a higher spread as compared to the
spreads of the above mentioned processes for the same value of ρ. The same trend can be
seen when different values of ρi (i = 1, 2, 3) are selected. Furthermore, increasing the value of
the relay gain AMR causes a decrease in the maximum value and an increase in the spread of
the PDF of MLSS processes Ξ (t).
0.7
Theory
Probability density function, pΞ (x)
Simulation
0.6 Classical Rayleigh
0.5 ρ = 1 (SLDS process)

ρ = 1 (Classical Rice) ρ1 = ρ2 = ρ3 = ρ
0.4 Double Rayleigh
ρ = 1 (SLSS process)
0.3
ρ = 1 (Double Rice process)
ρ = 0 (NLSS process)
0.2 ρ = 1; AMR = 1.5 (MLSS process)
ρ1 = ρ3 = ρ2 (MLSS process)
0.1 ρ = 1 (MLSS process)
0
0 1 2 3 4 5 6 7 8 9 10
Level, x
Fig. 2. A comparison of the PDF pΞ ( x ) of the MLSS process Ξ (t) with that of various other
stochastic processes.
Figure 3 presents a comparison of the PDF pΘ (θ ) of the phase process Θ (t) with that
of the phase processes corresponding to the classical Rayleigh and double Rayleigh processes,
the classical Rice and double Rice processes, the SLDS process, the NLSS process, and the
SLSS process for ρ = 1. It should be noted that the results presented in Fig. 3 are valid for the
case when f ρi (i = 1, 2, 3) is set to zero. It can be observed that the PDF of the SLDS phase
process has the highest peak. The PDF pΘ (θ ) of the phase process Θ (t) follows the same
trend in terms of the maximum value and the spread as that of the classical Rice process for
the same value of ρ. Furthermore, it is interesting to note that for the phase process Θ (t) with
ρi (i = 1, 2, 3) being selected as ρ1 = ρ3 = 0.5 and ρ2 = 1, the PDF pΘ (θ ) is almost the same
as that of the double Rice process for ρ = 1. Figure 3 also shows the impact of the relay gain
AMR on the PDF pΘ (θ ) of the phase process Θ (t).
0.7
Theory
Simulation
Probability density function, pΘ (θ)
0.6
ρ = 1 (SLDS process)
0.5
fρ1 = fρ2 = fρ3 = 0
ρ = 1; AMR = 1 (MLSS process)
ρ = 1; AMR = 1.5 (MLSS process) ρ = 1 (Classical Rice)
0.4
ρ = 1 (SLSS Process)
0.3 ρ1 = ρ3 = ρ2 ; AMR
(MLSS process) ρ = 1 (Double Rice)
0.2
0.1
Classical Rayleigh
Double Rayleigh
ρ = 0 (NLSS process)
0
−3.14 0 3.14
Phase, θ
Fig. 3. A comparison of the PDF pΘ (θ ) of the phase process Θ (t) with that of various other
various processes.
5. Conclusion
In this chapter, we have proposed a new flexible M2M amplify-and-forward relay fading
channel model under LOS propagation conditions. The novelty in the model is that we have
considered the LOS components in both transmission links, i.e., the direct link between the
source mobile station and the destination mobile station as well as the link via the mobile
relay. By analogy with multiple scattering radio propagation channels, we have developed
the MLSS fading channel as a second-order scattering channel, where the received signal
comprises the single and double scattered components. Furthermore, the flexibility of the
MLSS fading channel model comes from the fact that it can be reduced to double Rayleigh,
double Rice, SLDS, NLSS, and SLSS channel models under certain assumptions.
This chapter also presents a deep analysis of the statistical properties of MLSS fading
channels. The statistical properties studied, include the PDF of MLSS processes along with
the PDF of the corresponding phase processes. Accurate analytical expressions have been
derived for the above mentioned statistical quantities. The accuracy and validity of the
analytical expressions are confirmed by simulations. The excellent fitting of the theoretical
and simulation results verifies the correctness of the derived analytical expressions. The
presented results show that the statistical properties of MLSS channels are quite different
from those of the processes embedded in the MLSS channel model as special cases. It is
also evident from the illustrated results that the relay gain has a significant impact on the
statistical properties of MLSS channels.
The statistics of a fading channel dictate the choice of the transmitter and the receiver
techniques, including the detection, modulation, and coding schemes, etc. Therefore, the the-
oretical results presented in this chapter are quite useful for the designers of the physical layer
of M2M cooperative wireless networks. Furthermore, the developed M2M channel model can
be employed to investigate the overall system performance of M2M communication systems
under both NLOS and LOS propagation conditions.
6. References
Akki, A. S. (1994). Statistical properties of mobile-to-mobile land communication channels,
IEEE Trans. Veh. Technol. Vol. 43(No. 4): 826–831.
Almers, P., Tufvesson, F. & Molisch, A. F. (2006). Keyhole effect in MIMO wireless channels:
measurements and theory, IEEE Trans. Wireless Commun. Vol. 5(No. 12): 3596–3604.
Andersen, J. B. (2002). Statistical distributions in mobile communications using multiple scat-
tering, Proc. 27th URSI General Assembly, Maastricht, Netherlands.
Dohler, M. (2003). Virtual Antenna Arrays, Ph.D. dissertation, King’s College, London, United
Kingdom.
Erceg, V., Fortune, S. J., Ling, J., Rustako, Jr, A. J. & Valenzuela, R. A. (1997). Comparison
of a computer-based propagation prediction tool with experimental data collected in
urban microcelluar environment, IEEE J. Select. Areas Commun. Vol. 15(No. 4): 677–
684.
Gradshteyn, I. S. & Ryzhik, I. M. (2000). Table of Integrals, Series, and Products, 6th edn, New
York: Academic Press.
Kovacs, I. Z., Eggers, P. C. F., Olesen, K. & Petersen, L. G. (2002). Investigations of outdoor-to-
indoor mobile-to-mobile radio communication channels, Proc. IEEE 56th Veh. Technol.
Conf., VTC’02-Fall, Vol. 1, Vancouver BC, Canada, pp. 430–434.
Laneman, J. N., Tse, D. N. C. & Wornell, G. W. (2004). Cooperative diversity in wireless net-
works: efficient protocols and outage behavior, IEEE Trans. Inform. Theory Vol. 50(No.
12): 3062–3080.
Papoulis, A. & Pillai, S. U. (2002). Probability, Random Variables and Stochastic Processes, 4th edn,
New York: McGraw-Hill.
Patel, C. S., Stüber, G. L. & Pratt, T. G. (2006). Statistical properties of amplify and forward
relay fading channels, IEEE Trans. Veh. Technol. Vol. 55(No. 1): 1–9.
Pätzold, M. (2002). Mobile Fading Channels, Chichester: John Wiley & Sons.
Pätzold, M. & Hogstad, B. O. (2006). Two new methods for the generation of multiple un-
correlated Rayleigh fading waveforms, Proc. IEEE 63rd Semiannual Veh. Tech. Conf.,
VTC’06-Spring, Vol. 6, Melbourne, Australia, pp. 2782–2786.
Salo, J., El-Sallabi, H. M. & Vainikainen, P. (2006). Statistical analysis of the multiple scattering
radio channel, IEEE Trans. Antennas Propagat. Vol. 54(No. 11): 3114–3124.
Sendonaris, A., Erkip, E. & Aazhang, B. (2003a). User cooperation diversity — Part I: System
description, IEEE Trans. Commun. Vol. 51(No. 11): 1927–1938.
Sendonaris, A., Erkip, E. & Aazhang, B. (2003b). User cooperation diversity — Part II: Im-
plementation aspects and performance analysis, IEEE Trans. Commun. Vol. 51(No.
11): 1939–1948.
Shin, H. & Lee, J. H. (2004). Performance analysis of space-time block codes over keyhole
Nakagami-m fading channels, IEEE Trans. Veh. Technol. Vol. 53(No. 2): 351–362.
Simon, M. K. (2002). Probability Distributions Involving Gaussian Random Variables: A Handbook
for Engineers and Scientists, Dordrecht: Kluwer Academic Publishers.
Talha, B. & Pätzold, M. (2007a). On the statistical properties of double Rice channels, Proc.
10th International Symposium on Wireless Personal Multimedia Communications, WPMC
2007, Jaipur, India, pp. 517–522.
Talha, B. & Pätzold, M. (2007b). On the statistical properties of mobile-to-mobile fading chan-
nels in cooperative networks under line-of-sight conditions, Proc. 10th International
Symposium on Wireless Personal Multimedia Communications, WPMC 2007, Jaipur, In-
dia, pp. 388–393.
Talha, B. & Pätzold, M. (2008a). Level-crossing rate and average duration of fades of the
envelope of mobile-to-mobile fading channels in cooperative networks under line-
of-sight conditions, Proc. 51st IEEE Globecom 2008, New Orleans, USA, pp. 1–6. DOI
10.1109/GLOCOM.2008.ECP.860.
Talha, B. & Pätzold, M. (2008b). A novel amplify-and-forward relay channel model for mobile-
to-mobile fading channels under line-of-sight conditions, Proc. 19th IEEE Int. Symp.
on Personal, Indoor and Mobile Radio Communications, PIMRC 2008, Cannes, France,
pp. 1–6. DOI 10.1109/PIMRC.2008.4699733.
Resource Management with Limited Capability
of Fixed Relay Station in Multi-hop Cellular Networks 321
Resource Management with

Limited Capability of Fixed Relay Station
in Multi-hop Cellular Networks
Jemin Lee and Daesik Hong
YonseiUniversity
RepublicofKorea
1. Introduction
The purpose of this chapter is to develop a resource management technique to utilize
resource efficiently in multi-hop cellular networks. Multi-hop cellular networks have been
proposed as a way to enhance throughput and extend coverage (Cho & Haas, 2004). This
enhancement can in general be achieved by deploying relay stations in conventional cellular
networks. The advantage of multi-hop networks arises from the reduction in the overall
path loss achieved by using a relay station between a base station and a mobile station.
Moreover, deeply shadowed mobile stations can be supported by using relay stations to
bypass obstacles.
Even though the multi-hop transmission has advantages, it also carries a penalty: the need
for additional resources to transmit data in multi-hop manner (for example, in two-hop
transmission, two time slots or frequency channels for the base station-relay station link and
the relay station-mobile station link) (Lee et al.2008).
, Hence, this penalty of multi-hop
transmission is represented by ‘worms’, which devour resources (Ju etal. 2009).
,
The trade-offs associated with the multi-hop networks make it difficult to assess their
overall performance. For mobile stations with quality-of-service (QoS) requirements
guaranteed by single-hop transmission with few resources, multi-hop transmission could
end up wasting additional resources through multiple hops, even though it may provide
higher end-to-end data rates. In other words, a higher end-to-end data rate does not
guarantee higher efficiency in resource utilization. Hence, the amount of resources required
to guarantee QoS should be considered when assessing multi-hop transmission performance
and the performance can be different depending on a mobile station.
In addition, the infrastructure cost increases almost linearly with the equipment capability
(Johansson et al.2004),
, and it is impossible to change the capability of the installed
equipment flexibly according to the change of the required capability. Hence, a relay station
has the limited and determined capability. Due to the limited capability, some of mobile
stations cannot transmit in multi-hop if all capability of relay station has been already fully
used for the other mobile stations. Hence, the resource management for assigning the
limited capability to the mobile station, who can take more advantage of multi-hop
transmission, is required.
In this chapter, for utilizing the resources efficiently, the resource management considering
both the different multi-hop gains of each mobile stations and the limited capability of relay
station is provided. First of all, a brief explanation about multi-hop cellular networks is
provided in Section 2. Then, the transmission mode selection is discussed as a way to
determine whether multi-hop or single-hop transmission is the transmission mode most
appropriate for minimizing the resources used to guarantee a certain QoS in Section 3. The
multi-hop gain is defined as the amount of resources saved by using multi-hop transmission
instead of single-hop transmission, and the elements which affect the multi-hop gain are
discussed. Based on the affecting elements, two criteria for transmission mode selection are
provided and the performance of them is also verified.
In Section 4, the multi-hop user admission is discussed as a way to determine which mobile
station should be admitted or rejected to transmit data in multi-hop for maximizing the
achievable multi-hop gain using the limited capability of relay station. The multi-hop user
admission is formulated as a multi-dimensional knapsack problem, and two efficient
heuristic algorithms for multi-hop user admission are introduced and the performance of
those algorithms is discussed.
Finally, the structure of resource management including the transmission mode selection
and the multi-hop user admission is provided in Section 5.
2. Multi-hop Cellular Networks

The network under consideration in this chapter is a downlink orthogonal frequency
division multiple access (OFDMA) multi-hop cellular network. The multi-hop system
adopted here is the two-hop relaying system, which is known to be the most efficient multi-
hop system with respect to system capacity (Cho & Haas, 2004). In this system, data can be
transmitted from a base station to a mobile station in one of two transmission modes: single-
hop transmission or multi-hop transmission via a fixed relay station.
Fig. 1. Downlink multi-hop cellular networks

Fig. 1 shows an example of the connectivity for this system. In Fig. 1, the k th mobile station,
MSk , is connected with the m th base station, BSm , and MSk uses the i th fixed relay station
in the m th cell, RS mi , for multi-hop transmission. A number of fixed relay stations are
placed on the relay belt. All the fixed relay stations are regenerative relays, so they decode
data received from the base station and then forward it to the target mobile stations. In this
system, three kinds of links are formed. The base station-fixed relay station links and the
fixed relay station-mobile station links occurring with two-hop transmission are denoted by
the link for the first hop ( L1 ) and the link for the second hop ( L2 ), respectively. The link for
single-hop transmission ( LS ) also denotes the base station-mobile station link. Generally, it
is assumed that L1 has good channel condition for a line-of-sight (LOS) environment, and
L2 and LS are in a non line-of-sight (NLOS) environment (Liu et al.2006).
, The LOS
assumption can be satisfied by deploying fixed relay stations at selected locations, such as
on top of a blinding.
As QoS parameters which can be handled in the physical layer, the target data rate and the
target bit-error-rate (BER) can be considered. In the regenerative relay, errors generated at
each hop are propagated to the next hop. The sum of the target bit-error-rates for each hop
in two-hop transmission should therefore be equal to or less than the target bit-error-rate in
2
B
T
two-hop transmission, B T , as T
Li  B T where BLi is the target bit-error-rate on the link
i 1
of the i th hop (Boyer etal.

, 2004).
In addition, the end-to-end data rate from a base station to a mobile station is determined by
the minimum data rate among the rates in each hop (Jing etal. , 2005). Hence, the target data
rate for each hop should be equal to or greater than the target data rate, R T , as RLT  RT
i
where RLT is the target data rate on the link of the i th hop.
i
In OFDMA systems, each subcarrier can obtain a different channel gain. So, the received
signal to interference and noise ratio (SINR) on the n th subcarrier of the link L
( L {LS , L1 , L2} ) can be defined as
GL ( n)  PI
 L (n) = , L, (1)
 j (n)  GI (n)  PI  
jm
j
where  is the additive Gaussian noise power and GL (n) is the channel gain of link L on
the n th subcarrier. I j is the link between an interferer in the j th cell and the target node,
and PI is the transmission power of the transmitter on that link.
If it is assumed that every link in a cell utilizes different resources and resources used for
transmission on a link are also reused at the same link in other cells concurrently, then the
link BSm - RS mi and the link BS j - RS ij ( j  m ) use the same resources. Hence, in single-hop
transmission, L is the link BSm - MSk , I j is the link BS j - MSk , and PI = PBS , where PBS is
the transmission power of the base station in (1). In addition, in two-hop transmission, L is
the link BSm - RS mi , I j is the link BS j - RS mi , and PI = PBS on the first hop. L is the link RS mi -
MSk , I j is the link RS ij - MSk , and PI = PFRS on the second hop where PFRS is the
transmission power of the fixed relay station.
Some of subcarriers are not being used when the mobile stations do not require all of
subcarriers in OFDMA systems. Hence,  j ( n) in (1) is adopted to express the loading state
of the n the subcarrier in the j th cell.  j ( n) = 1 if the n th subcarrier is used to transfer data
in the j th cell, while  j ( n) = 0 otherwise. Therefore, the average loading state of the j th
cell,  j , is defined as
1 N
 j (n),
N n =1
j = (2)
where N is the total number of subcarriers in a cell. In addition,  j becomes one when
there is full loading.
With adaptive modulation coding (AMC), the throughput on the link L is expressed as the
function of  L (n) and BLT , as follows:
RL (n) = f ( BLT , L (n))
1  1.5  (3)
= log 2 1  T
  L (n)  , L, n,
Ts  ln(5  BL ) 
where Ts is the symbol duration (Qiu and Chawla, 1999). The number of subcarriers on link
L required to guarantee the QoS for MSk , Ck , L , is determined by R T and RL (n) . For
instance, when the channel gains of all subcarriers are equal and the average loading state
for the other cells is one, then RL (1) = RL (2) =  = RL , and Ck , L can be defined as
Ck , L =  RT / RL   k , L, (4)
where    is the least integer equal to or greater than  . Hence, the total numbers of
subcarriers of MSk required to transmit data with a QoS guarantee by single-hop
transmission and multi-hop transmission, CkSH and CkMH , are respectively defined as
2
CkSH = Ck , L , CkMH = Ck , L ,  k . (5)
S i
i =1
A fixed relay station has a limited capability due to the cost, the limitation of power
amplifier and so on. In this chapter, the number of supportable subcarriers is considered as
the capability of the fixed relay station. It means that the i th fixed relay station can support
up to N i , RX subcarriers for receiving on L1 (the link between the base station and the relay
station) and N i ,TX subcarriers for transmitting on L2 (the link between the relay station and
the mobile station) in a time. Hence, the total number of required subcarriers of each link
should be equal or fewer than the supportable number of subcarriers of a fixed relay station
in each link as follows:
 Ck , L  Ni , RX ,  Ck , L  Ni,TX ,
1 2
(6)
k MU i k MU i
where MU i is the set of mobile stations who want to use the i th fixed relay station in
transmission. Fixed relay stations which have higher capabilities can generally support
more subcarriers with higher total power.
3. Transmission Model Selection: Multi-hop vs. Single-hop

The multi-hop transmission needs for additional resources, but it achieves more reliable
transmission than the single-hop transmission. Due to this trade-off, the multi-hop
transmission cannot be better than the single-hop transmission for all users. Hence, the
elements, which affect the performance of multi-hop transmission, are investigated and the
transmission mode selection for more efficient resource utilization is discussed in this
section.
3.1Achievable Gain from Multi-hop Transmission

With respect to efficiency of resources, the gain associated with multi-hop transmission is
achieved when subcarriers are saved by transmitting in multi-hop instead of in single-hop.
Hence, the multi-hop gain, Sk , can be defined as the relative ratio of the amount of saved
subcarriers to the number of required subcarriers in single-hop transmission as follows (Lee
., 2008):
etal
C SH  C MH
S k = k SH k , k . (7)
Ck
This approach shows the amount of gain or loss with multi-hop transmission. Thus, if Sk
has a positive value, that means that the multi-hop transmission is saving subcarriers with
guaranteeing the QoS requirements for the k th mobile station. On the other hand, if the
value of Sk is negative, that implies that the multi-hop transmission is wasting subcarriers.
Fig. 2 presents the multi-hop gain in various environments with the assumption of the equal
average loading states of other cells as  j =  ,j  m in (2). The distance between the k th
mobile station and the base station is denoted by dk and the cell radius is d cell . System
parameters in Table 1 are used for simulations. Fig. 2 and the formulas from (1) to (5) show
that multi-hop gain is affected by three elements: loading state of other cells, location of the
mobile station, and QoS requirements. Due to the long transmission distance in single-hop
transmission, Ck , L is generally greater than Ck , L or Ck , L . However, for the multi-hop
S 1 2
transmission, subcarriers for multi-hops, the sum of Ck , L and Ck , L , should be used. By this
1 2
relation, the multi-hop gain is affected by the loading states of other cells and the location of
the mobile station. As the loading state of other cells increases or if the mobile station is
located near the cell boundary, the SINRs of the three links decrease. At a lower SINR, the
achievable data rate is more sensitive to variation of SINR due to the log function as in (3).
This means that the farther the SINR falls, the faster the number of subcarriers required for
guaranteeing QoS increases. Hence, in this environment, Ck , L increases more rapidly than
S
the sum of Ck , L and Ck , L , so that a bigger multi-hop gain can be achieved.

1 2
System Parameters Values

System bandwidth 5 MHz
Number of subcarriers 1024
Path loss exponent (LOS/NLOS) 2.35 / 3.76
Standard deviation of shadowing 3.4 dB / 8 dB
(LOS/NLOS)
Transmission power of base station 43 dBm / 40 dBm
/fixed relay station
Cell radius /Radius of the relay belt 500 m / 250 m
Number of cells 7
Number of fixed relay stations 6, Symmetrically located
per cell on the relay belt
Number of mobile stations per cell 200, Uniformly distributed
Modulation order BPSK, QPSK,
16-,64-,128- QAM
Power control Equal power allocation
Table 1. System parameters (for simulations of Fig. 2, Fig. 3, Fig. 5 and Fig. 6)
0.6
0.4
0.2
 =1.00
Multi-hop gain
-0.2
 =0.95
-0.4
 =0.90
QoS Requirements
-0.6 -6
6 4kbps/1 0
-6
 =0.85 1 28 kbps/10
-0.8 -3
1 28 kbps/10
0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0

Normalized location of mobile station, dk / dcell
Fig. 2. Multi-hop gain based on QoS requirements (target data rate/target BER), loading
state of other cells (  ) and mobile station location (the mobile station’s QoS parameters are
128kbps / 103 for all cases except for the case of  = 1.0.)
However, this increasing multi-hop gain begins to decrease when dk is above a certain
value (e.g., d k / dcell  0.7 when   0.95 in Fig. 2). The reason for this is that Ck , L is also
2
increasing rapidly due to larger interference as the mobile station moves to the cell edge,
causing the difference between CkSH and CkMH to get smaller.
On the other hand, as the mobile station approaches the base station or the loading states of
other cells decrease, the SINRs for all three links are increasing. When the SINR has
increased enough, the required number of subcarriers becomes small and it is no longer
sensitive to the variation in SINR. Hence, in this environment, all of Ck , L , Ck , L and Ck , L
S 1 2
are small, so that the multi-hop transmission would waste subcarriers because of the usage
of subcarriers for two hops. For this reason, the multi-hop transmission does not have any
gain as the loading states of other cells decrease or the mobile station approaches the base
station (e.g., dk / dcell  0.5 when   0.95 in Fig. 2).
In addition, comparing the upper three lines in Fig. 2 shows that the multi-hop gain has
different values depending on the target data rate and the target BER. Even the multi-hop
gains for these three cases become similar when the mobile station approaches the cell
boundary because of the large amount of interference, a lower target BER induces a higher
multi-hop gain because the reliable transmission is more important to mobile stations
requiring lower target BERs. The multi-hop gain also varies depending on the target data
rate due to the subcarrier allocation process.
In this subsection, the affecting elements on the multi-hop gain, the loading state of other
cells, the location of mobile station, and QoS requirement, have been discussed. Since the
multi-hop transmission can save (or waste) resources according to those elements, the multi-
hop transmission should be used selectively for efficient utilization of resources. Hence,
transmission mode selection is required to determine the appropriate transmission mode for
each mobile station: multi-hop transmission or single-hop transmission.
3.2 Mechanism for Transmission Mode Selection

The number of required subcarriers changes depending on the transmission mode, the QoS
requirements, and the channel condition. To save subcarriers, whichever transmission mode
requires fewer subcarriers should be the one selected. Hence, the subcarrier-based criterion,
 k , S , is defined as follows (Lee etal
., 2008):
 k , S = CkSH CkMH ,  k. (8)
If  k , S is greater than zero, then the selected transmission mode becomes the multi-hop
transmission; otherwise, single-hop transmission mode is selected.
In the case where the SINRs for each subcarrier are different, CkSH and CkMH cannot be
calculated before the transmission mode is determined because the frequency bands for
single-hop and multi-hop transmission may be different. The subcarrier allocation which
determines the number of required subcarriers should therefore be performed after
determination of the transmission mode. In this case, the average number of required
subcarriers Ck can be used instead of Ck to calculate  k , S where   SH , MH  . CkSH and
CkMH are respectively defined as CkSH =  RT RL  , and CkMH =  i 1  RT RL ,  k , where

2
 S   i 
N
RL = (1 / N )  RL (n).
n =1
As simpler way to select transmission mode, the distance based criterion can be used. As
discussed in Section 3.1, the multi-hop gain changes depending on the locations of mobile
stations. Positive gain could be obtained when a mobile station is located near the base
station. Hence, the transmission mode can be determined using the following criterion (Lee
etal., 2008):
 k , D = d k  d ref ,  k . (9)
In (9), dk is ( x0  xk )2  ( y0  yk )2 where ( x0 , y0 ) and ( xk , yk ) are the coordinates (x and y)

of a base station's location and the mobile station's location obtained by a positioning system
(e.g., global positioning system), respectively, and d ref is the reference distance for the
transmission mode selection. If dk is greater than d ref ,  k , D has a positive value and the
multi-hop transmission is selected. This means that the multi-hop transmission is applied
for mobile stations located outside of the circle with radius d ref .
The performance of the transmission mode selection can be expressed as blocking
probability, which is the ratio of the number of unsupportable mobile stations, KU , to the
total number of mobile stations which access the system, KT , or KU / KT as Fig. 3. In this
figure, TMS-S and TMS-D represent the transmission mode selections using the subcarrier-
based criterion (  k , S ) and the distance-based criterion (  k , D ), respectively.
Only Single-hop
Only Multi-hop
0.5
TMS-D
TMS-S
Blocking Probability
0.4
 =1.0
0.3
 =0.9
0.2
 =1.0
0.1
 =0.9
0
30 70 110 150 190 230 270 310 350
Total number of mobile stations in a cell
Fig. 3. Comparison of blocking probabilities based on total number of mobile stations in a cell
In Fig. 3, TMS-D and TMS-S demonstrate better performance, regardless of the loading
states of other cells, than the cases where either single-hop transmission or the multi-hop
transmission is applied without the selection process. This means that more mobile stations
can be supported in a cell with lower blocking probability when the transmission mode
selection is applied. Specifically, when the loading states of other cells are 1.0 and the
blocking probability is 0.1, only 45 mobile stations and 95 mobile stations can be supported
using conventional single-hop transmission and the multi-hop transmission, respectively.
The number of supportable mobile stations increases to 150 with TMS-D and 250 with TMS-S.
4. Multi-hop User Admission for Relay Stations with Limited Capabilities

Based on the transmission mode selection in Section 3.2, themulti-hopand userthesingle-hop
user, which denote the mobile stations which select the multi-hop transmission and the
single-hop transmission, respectively, are determined. Since all fixed relay stations have the
limited capabilities, a fixed relay station may not be able to support all of multi-hop users.
When the required subcarriers of multi-hop users are beyond the number of supportable
subcarriers in (6), some multi-hop users cannot receive data in multi-hop. Hence, to utilize
the limited capability of the fixed relay station efficiently, the multi-hop user admission is
required to allow the multi-hop users, which achieve high multi-hop gain, to use a fixed
relay station. Therefore, in this section, the problem of multi-hop user admission is
discussed and the multi-hop user admission algorithms are presented.
4.1 Formulation of Multi-hop User Admission

The multi-hop user admission strategy is formulated to determine the admitted multi-hop
users to use the fixed relay station among all multi-hop users for maximizing the total multi-
hop gains which can be obtained from the admitted multi-hop users as follows:
F
max 
i =1 k MU i
xk  Sk
s.t 1,i :  xk  Ck , L  N i , RX , i
1 (10)
k MU i
 2,i :  xk  Ck , L  N i ,TX , i
2
k MU i
3 : xk  {0,1}, k ,
where F is the total number of fixed relay stations in a cell, and the xk is the indicator of
admission. If MSk is admitted to transmit in multi-hop using the fixed relay station,
then xk = 1 , otherwise xk = 0 and MSk should transmit in single-hop.
For each fixed relay station, there are two constraints from the limited capabilities: 1,i and
 2,i based on the i th fixed relay station. The Lagrangian of multi-hop user admission
problem, Lag (u , u) , can be defined as follows:
F  
i =1  k MU i k MU i
1
 k MU i
 2
 
Lag (u , u) =    xk  S k  ui  xk  Ck , L  N i , RX  ui  xk  Ck , L  N i ,TX 
 (11)
F
= Lag (ui , ui' )
i =1
where ui and ui' are Lagrangian multipliers. Hence, it is shown that the primal problem can
be decomposed into sub-problems Z (ui , ui' ) where Z (ui , ui' ) = max Lag (ui , ui' ) . It means that
the parallel multi-hop user admissions for each fixed relay station can work independently
since the result of one fixed relay station's multi-hop user admission does not affect the
other fixed relay stations' multi-hop user admission. Therefore, we can deal with the multi-
hop user admission problem by decomposing into the independent multi-hop user
admission problems for each fixed relay station, which can be formulated as follows:
max x k  Sk
k MU i (12)
s.t 1,i ,  2,i , and 3 .
The problem in (12) has the equivalent form with the two-dimensional knapsack problem
(TDKP) which is a kind of multi-dimensional knapsack problem (MDKP) (Qiu & Chawla,
1999). The MDKP is a variant of the classical 0-1 knapsack problem (KP) with more than two
knapsacks.
4.2 Algorithms for Multi-hop User Admission

As shown in Section 4.1, the multi-hop user admission strategy can be represented as the
two-dimensional knapsack problem a kind of MDKP. The KP and the MDKP are proven to
be NP-hard, so the optimal solutions of them cannot be obtained in a polynomial time
(Akbar etal., 2005). Hence, a heuristic algorithm could be used for multi-hop user admission,
and two multi-hop user admission algorithms are presented in this section.
4. 2. 1 Balanced Link Multi-hop User Admission (BL-MUA) Algorithm

In the multi-hop user admission, the balance between the used resources on L1 and those on
L2 is important. The reason for this is that no additional multi-hop users can be admitted
when at least one of N i , RX on L1 and N i ,TX on L2 is fully occupied for other multi-hop
users. Hence, the balanced link multi-hop user admission (BL-MUA) algorithm has been
proposed in (Lee et al., 2007) considering the balance in admission by adopting the primal
effective gradient method (PEGM) (Toyoda, 1975).
The PEGM determines the priority of admission using the new measurement of the
aggregate resource. The aggregate resource is to penalize the multi-hop user which requires
many subcarriers in the more loaded link. The aggregateresource in multi-hop user admission
can be defined as Ak  (Ck , L  l1  Ck , L  l2 ) / l12  l22 where l1 and l2 are the total numbers of
1 2
used subcarriers on L1 and L2 by the currently admitted multi-hop users, respectively. The
priority function of the BL-MUA algorithm, U BL , k , is set by the multi-hop gain over the
aggregate resources, Sk / Ak , as follows:
S k l12  l22
U BL , k = . (11)
Ck , L  l1  Ck , L  l2
1 2
If at least one of N i ,TX and N i , RX is insufficient for supporting all multi-hop users in MU i ,
the multi-hop user with the lowest priority is rejected to use the fixed relay station one by
one. The priority function helps to balance between the loading states of two links while
rejecting the multi-hop user.
4. 2. 2 Focused Link Multi-hop User Admission (FL-MUA) Algorithm

The BL-MUA algorithm attaches importance to the balance of used resources in L1 and L2 .
However, if fixed relay stations are located in a LOS environment with the base station and
the environment of L2 is a NLOS (like a general assumption), then the number of required
subcarriers to guarantee the same target data rate in L2 is much more than that in L1 ,
Ck , L  Ck , L . This means that more multi-hop users could not be admitted in a fixed relay
1 2
station generally due to the full loading of L2 , not that of L1 (the supportable subcarriers in
L2 is exhausted quickly than that in L1 ).
Therefore, the multi-hop user admission considering both loading states of L1 and L2 can
be simplified to that considering only that of L2 . In this case, the multi-hop user admission
strategy becomes a simple knapsack problem, not the two-dimensional knapsack problem
anymore. Hence, the focused link multi-hop user admission (FL-MUA) algorithm is
proposed to focus only on the loading state of L2 (Lee etal.
, 2007). As the priority function in
the FL-MUA algorithm, the multi-hop gain per the average number of required subcarriers
in L2 is used as
U FL , k  Sk / Ck , L2 . (12)
When the supportable subcarriers of a fixed relay station are not sufficient, the multi-hop
users are excluded in a low-priority order one by one.
4. 2. 3 Procedure of multi-hop user admission algorithms

The process of the multi-hop user admission algorithms progresses independently for each
fixed relay station, and the overall procedure is shown by the flow chart in Fig. 4. If multi-
hop users which want to use the i th fixed relay station exist, the total number of required
subcarriers for supporting all multi-hop users in L1 and L2 , 1
generally unknown, except for some centralized scheduling-based MAC protocols like Time
Division Multiple Access (TDMA) where the problem finds a mathematical formulation.
The main finding in (Gupta, 2000) is that per-node capacity of a random wireless network
with n static nodes scales as  ( 1 ) . They assume a threshold-based link layer model in
n log n
which a packet transmission is successful if the received SNR at the receiver is greater than a
fixed threshold. Instead of this ideal link layer model, Mhatre et al. considere a probabilistic
lossy link model and show that the per-node throughput scales as only  ( 1 ) instead of
n
1 (Mahtre & Rosenberg, 2006).
( )
n log n
These asymptotic bounds are calculated under assumptions such as node homogeneity and
random communication patterns. Therefore, some researches try to relax some of these
assumptions on network configuration. Jain et al. focus on interference status among the
transmitters as one of the main limiting factors on routing performance (Jain et. al, 2003).
They propose to represent interference among wireless links using a conflict graph. A
conflict graph shows, which wireless links interfere with each other, such that each edge in
the connectivity graph is represented by a vertex and there exists an edge between two
vertices if the links interfere with each other. Thus, the throughput optimization problem is
posed as a linear programming problem in which upper and lower bounds of the maximum
throughput are obtained by finding the maximal clique and independent set in the conflict
graph.
Karnik et al. extend the conflict graph idea to a conflict set (Karnik et. al, 2007). Their
rational is that an interference model can not be binary since for a given link, generally there
is a subset of links that at least one of them should be silent when the given link is
transmitting. They propose a joint optimization of routing, scheduling and physical layer
parameters to achieve the highest throughput. Both these proposals bring valuable
achievement but as they investigate the highest capacity of a network, they have to assume
TDMA instead of contention-based algorithm, which leads to probabilistic results. Hence,
they implicitly assume that data transmissions are scheduled by a central entity. Therefore
they may not be applied easily to more practical networks such as IEEE 802.11 with random
access to the channel.
Computing the optimal throughput, despite of giving a good vision to the maximum
achievable throughput in the network, may not be implementable in a real network. There
are many complicated issues such as necessity of having a distributed routing/scheduling
protocol, random quality for the wireless links, limited allowable overhead to the network,
compatibility with MAC 802.11, etc., that motivate the researchers to find a practical
solution to achieve a good performance.
2.2 Maximum Throughput Routing

Most of the work done in this area relies on the broadcasting of extra probe packets to
estimate the channel quality (ex. Sivakumar et. al, 1999; De Couto et. al, 2005; Draves & Zill,
2004). However, since the quality of the wireless links depends significantly on physical
settings (such as transmit rate and packet size), the probes may not reflect the actual quality
of the links. The reason is that for preventing to throttle the entire channel capacity they
4.3 Performance of Multi-hop User Admission Algorithms

In this section, the performance of the BL-MUA algorithm and that of the FL-MUA
algorithm are evaluated with the assumption that the deployed fixed relay stations have the
same capabilities as N R and all capabilities for transmitting and receiving are equal as
N i , RX = N i ,TX = N R , i . The performances of the multi-hop user algorithms are verified with
two types of fixed relay stations: the fixed relay station with low capability (L-FRS) and that
with high capability (H-FRS). The numbers of supportable subcarriers per fixed relay station
are set to 16 for L-FRS and 64 for H-FRS, and the average loading state of other cells is set
to one. The other system parameters are the same as Table 1. In addition, the performances
of the multi-hop user admission algorithms are compared to the case where multi-hop users
are randomly admitted without an admission algorithm (w/o MUA in Fig. 5).
In both multi-hop user admission algorithms, with the higher priority, the supportable
subcarriers of a fixed relay station are used for the mobile stations which occupy fewer
subcarriers with higher multi-hop gain. Hence, the number of admitted multi-hop users in a
fixed relay station can be increased by the algorithms. Those are verified in Fig. 5.
Fig. 5 presents the number of admitted multi-hop users in a fixed relay station according to
two types of fixed relay stations. More multi-hop users can be supported in a fixed relay
station by the multi-hop user admission algorithms within the limited capability of fixed
relay station compared to the case without multi-hop user admission algorithm.
Moreover, the performance difference between the multi-hop user admission algorithms
and the case without the algorithm becomes more significant as the number of multi-hop
users increases. When the number of multi-hop users is small, the capability of fixed relay
station is generally sufficient for supporting all multi-hop users. Hence, the multi-hop user
admission is not actually required and the performance of the multi-hop user admission
algorithms is similar to the case without the algorithms. However, the multi-hop user
admission becomes meaningful when the fixed relay station cannot support all multi-hop
users because of the insufficient capability.
In addition, the performance of BL-MUA algorithm and that of FL-MUA algorithm are
similar. The reason for this is from the physical characteristics of L1 and L2 . The number of
required subcarriers in L2 is much more than that in L1 as Ck , L  Ck , L . In this case, the
2 1
priority function of the BL-MUA algorithm in (11) can be approximated to the priority
function of the FL-MUA algorithm in (12) as
U BL , k = S k l12  l22 C k , L1  l1  C k , L  l2
2
 S k Ck , L = U FL, k . Thus, the similar priority functions
2
are used in both algorithms, so the total numbers of admitted users of them do not have big
difference. It implies that the FL-MUA algorithm could obtain similar performance to the
BL-MUA algorithm with less complexity.
In addition, in the aspect of total capacity in a cell, more mobile stations can be supported
with guaranteeing their QoS requirements regardless of single-hop or multi-hop
transmissions using the multi-hop user admission algorithms. This can be verified by Fig. 6
which shows the blocking probability as the number of mobile stations per cell increases.
30
w/o MUA
Total number of admitted multi-hop users FL-MUA
25
BL-MUA
H-FRS
20
15
L-FRS
10
0
1 11 21 31 41 51 60
Number of multi-hop users after transmission mode selection
Fig. 5. The total multi-hop gain obtained from the admitted multi-hop users in a fixed relay
station according to two types of fixed relay stations
w/o MUA
FL-MUA L-FRS
0.5
BL-MUA
Blocking probability
0.4
0.3 H-FRS
0.2
0.1
0
120 140 160 180 200 220 240 260 280 300
Total number of mobile stations per cell
Fig. 6. The blocking probabilities according to two types of fixed relay stations
The blocking probabilities of the multi-hop user admission algorithms are smaller than that
of the case without the algorithm over all range. Specifically, within 0.1 blocking probability,
the case without the algorithm can support at most 135 mobile stations with L-FRSs and 185
mobile stations with H-FRSs. On the other hand, the numbers of supportable mobile stations
are increased up to 180 mobile stations with L-FRSs and 223 mobile stations with H-FRSs by
the multi-hop user admission algorithms. It implies that 33 % and 20 % of the supportable
mobile stations in a cell are increased by the multi-hop user admission algorithms with L-
FRSs and H-FRSs, respectively. Thus, more mobile stations can be supported in a cell with
low blocking probability by the multi-hop user admission algorithms.
5. Overall Structure of Resource Management in Multi-hop Cellular Networks

In Section 3 and Section 4, the transmission mode selection and the multi-hop user
admission have been discussed. The transmission mode determined through the
transmission mode selection can be changed after the multi-hop user admission. The reason
for this is that some mobile stations should transmit in single-hop due to the limitation of
the fixed relay station’s capability even though they select the multi-hop transmission in the
transmission mode selection. Therefore, the final transmission mode determination and the
resource allocation should be performed after the multi-hop user admission. The overall
process of the resource management including the transmission mode selection and the
multi-hop user admission is summarized as follows:
Fig. 7. The overall structure of resource management in multi-hop cellular networks

1) Once the mobile station set is defined and the required SINR information of the mobile
stations are collected in a base station, the resource management process can start. The
required SINR information is determined according to the criterion of the transmission
mode selection and it could include the received SINRs of L1 , L2 and LS .
2) During the transmission mode selection with a specific criterion (subcarrier-based or
o
distance-based criterion), the set of multi-hop users, MU i , and the set of single-hop users,
o
SU , are determined.
3) After determining the set of multi-hop users, the multi-hop user admission can be
performed. Through a multi-hop user admission algorithm (BL-MUA or FL-MUA
algorithm), the mobile stations who can finally receive data in multi-hop, MU i and the
mobile stations who should receive data in single-hop, SU , due to the lack of the capability
of the fixed relay station are determined.
o
4) The single-hop users in the sets of SU and SU are forwarded to the resource allocation
process based on SINR information of LS , and the multi-hop users in the set of
MU i undergoes the resource allocation process based on SINR information of L1 and L2 .
After all of those processes are progressed, the final transmission mode which can maximize
the multi-hop gain within the limited capability can be obtained.
According to the transmission mode selection criterion and the multi-hop user admission
algorithm, the complexity of this resource management could be changed. In the
transmission mode selection, the complexity of the transmission mode selection with the
subcarrier-based criterion is O ( KT N (log p ) p 2 ) where p is the required number of digits
used in the operations such as square root and multiplication. It is higher than the
complexity of transmission mode selection with the distance-based criterion, O ( KT p 2 ) (Lee
., 2008). Hence, the subcarrier-based selection criterion’s complexity is higher than the
et al
distance-based selection criterion, but it achieves better performance since it consider more
elements for determining an appropriate transmission mode as shown in Section 3.
In the multi-hop user admission, the FL-MUA algorithm is simpler than the BL-MUA
algorithm while the BL-MUA algorithm could achieve better performance as shown in
Section 4.
Therefore, designers can use appropriate selection criterion and multi-hop user admission
algorithm based on what the acceptable complexity level for that system is.
6. Conclusion
This chapter provides the resource management for efficient resource utilization in multi-
hop cellular networks. The resource management has two parts: the transmission mode
selection and the multi-hop user admission. The transmission mode selection is a way to
select an appropriate transmission mode between the multi-hop transmission and the
single-hop transmission for saving resources with guaranteeing QoS requirements of mobile
stations. Two kinds of selection criteria, subcarrier-based and distance-based criterion, are
provided after discussing the elements which affect the multi-hop gains such as the QoS
requirements, the mobile station’s location, and the loading states of other cells.
However, due to the limited capability of a relay station, some mobile stations cannot
transmit data in multi-hop even though they select the multi-hop transmission mode.
Hence, the multi-hop user admission is provided as a way to assign the limited capability of
a relay station to the mobile stations which can maximize the multi-hop gain. Since the
multi-hop user admission strategy is a NP-hard problem, two heuristic algorithms are
provided: the BL-MUA algorithm focused on the load balance between L1 and L2 and the
FL-MUA algorithm focused only on the load state of L2 . Through the transmission mode
selection and the multi-hop user admission, the resources can be used efficiently with
supporting more mobile stations with lower blocking probability.
7. Acknowledgement
This work was supported by Korea Science and Engineering Foundation through the NRL
Program (Grant R0A-2007-000-20043-0).
8. References
Akbar, M. M.; Rahman, M. S.; Kaykobaad M. & Manning, E. G. (2005). Solving the
Multidimensional Multiple-choice Knapsack Problem by Constructing Convex
Hulls. Computers&operations,research Vol. 33, No. 5, pp. 1259-1273, ISSN 0305-0548.
Boyer, J.; Falconer, D. D. & Yanikomeroglu, H. (2004). Multihop Diversity in Wireless
Relaying Channels. IEEE Transactions onCommunications , Vol. 52, No. 10, pp. 1820-
1830, ISSN 0090-6778.
Cho, J. & Haas, Z. J. (2004). On the Throughput Enhancement of the Downstream Channel in
Cellular Radio Networks through Multihop Relaying. IEEEJournalonSelectedAreas
inCommunications , Vol. 22, No. 7, pp. 1206-1219, ISSN 0733-8716.
Jing, S.; Zhao-yang, Z.; Pei-liang, Q. & Guan-ding, Y. (2005). Subcarrier and Power
Allocation for OFDMA-Based Regenerative Multihop Links. InternationalConference
onWirelessCommunications, orking
Netw andMobileComputing , Vol. 1, pp. 207-210.
Johansson, K.; Furuskar, A.; Karlsson, P. & and Zander, J. (2004). Relation between Base
Station Characteristics and Cost Structure in Cellular Systems. IEEE International
SymposiumonPersonal,IndoorandMobileRadioCommunications.
Ju, H.; Oh, E. & Hong, D. (2009). Catching Resource-Devouring Worms in Next-Generation
Wireless Relay Systems - Two-way Relay and Full-Duplex Relay. IEEE
CommunicationsMagazine , Vol. 47, No. 9, pp. 58-65, ISSN 0163-6804.
Lee, J.; Wang, H.; Lim, S. & Hong, D. (2007). A Multi-hop User Admission Algorithm for
Fixed Relay Stations with Limited Capabilities in OFDMA Cellular Networks. IEEE
th
81 InternationalSymposiumonPersonal, IndoorandMobileRadioCommunications , pp.
1-5, ISBN 978-1-4244-1144-3, Athens, Sept. 2007.
Lee, J.; Wang, H.; Seo, W. & Hong, D. (2008). QoS-guaranteed Transmission Mode Selection
for Efficient Resource Utilization in Multi-hop Cellular Networks. IEEETransactions
onWirelssCommunications, Vol. 7, Issue 10, pp. 3697-3701, ISSN 1536-1276.
Liu, T. etal. (2006). Radio Resource Allocation in Two-hop Cellular Relaying Network, IEEE
VehicularTechnologyConference , Vol. 1, pp. 91-95, ISBN 1-7803-9392-9, May 2006.
Qiu, X. & Chawla, K. (1999). On the Performance of Adaptive Modulation in Cellular

Systems. IEEE Transactions on Communications , Vol. 47, No. 6, pp. 884-895, ISSN
0090-6778.
Toyoda, Y. (1975). A Simplified Algorithm for Obtaining Approximate Solution to Zero-one
Programming Problems. Management Science , Vol. 21, No. 12, pp. 1417-1427, ISSN
0025-1909.
On Cross-layer Routing in Wireless Multi-Hop Networks 339
On Cross-layer Routing in
Wireless Multi-Hop Networks
Golnaz Karbaschi1, Anne Fladenmuller2 and Sébastien Baey2
1InstitutNationaldeRecherche
ique
enet
Informat
enAutomatique(INRIA)–Saclay
2UniversitéPierreetMarieCurie(UPMC)–Paris
France
1. Introduction
Wireless multi-hop networks represent a fundamental step in the evolution of wireless
communications. Several new applications of such networks have recently emerged
including community wireless networks, last-mile access for people, instant surveillance
systems and back-haul service for large-scale wireless sensor networks, local high-speed P2P
networking, or connectivity to rural/remote sites which was previously limited by cables.
Wireless multi-hop networks consist of computers and devices (nodes), which are connected
by wireless communication channel, denoted as links. Since a wireless communication has a
limited range, many pairs of node cannot communicate directly, and must forward data to
each other via one or more cooperating intermediate nodes. Thus, in a unicast routing the
source node transmits its packets to a neighboring node with which it can communicate
directly. The neighboring node in turn transmits the packets to one of its neighbors and so
on until the packets reach their final destination. Each node that forwards the packets are
referred to as a hop and the set of the links, which are selected to transfer the packets, are
called the route. Different routes from any nodes to any destinations are discovered by a
distributed routing protocol in the network. Figure 1 shows an example of a wireless mesh
network in which node S2 sends data traffic to the destination D via cooperation of
intermediate nodes R1 and R2, while the other source node S1 sends out its data traffic to
the gateway via the node A.
Wireless multi-hop networks are self-expanding networks; connectivity of the network is
only due to existence of the nodes, thus the network can be expanded or decreased simply
by adding or removing a node. In contrast, cellular networks are a much more expensive
infrastructure since they need at least one base station to provide connectivity. Moreover,
the capacity of the base station is limited and so not all the nodes in the coverage area can be
connected to the network. Therefore, wireless multi-hop networks are a promising solution
to expand the network easily as they allow flexibility and rapid deployment at low cost.
Fig. 1. An illustration of a Wireless Multi-Hop Network.
1.1 The Challenges of Wireless Multi-hop Routing in a Time Varying Environment

Routing is the fundamental issue for the multi-hop networks. A lot of routing protocols
have been proposed for the wired networks and some of them have been widely used such
as Routing Information Protocol (RIP) (Hedrick, 1998) and Open Shortest Path First
(OSPF)(Moy, 1998). Characteristics of wireless links differ extremely from wired links. Thus,
the existing routing protocols for wired networks can not work efficiently in the face of the
vagaries of the radio channels and limited battery life and processing power of the devices.
Moreover, the traditional routing metric of minimum-hop is not always the best solution for
routing in wireless network. In the sequel, the wireless links characteristics and limitations
of shortest path routing are described.
Wireless networks have intrinsic characteristics that affect intensely the performance of
transport protocols. These peculiarities, which distinguish themselves from conventional
wireline networks, can be summarized as follows:
 Wireless links have fundamentally low capacity. The upper band for the capacity of a
wireless link follows the Shannon capacity bound.
 Signal propagation experiences large scale and small scale attenuations. Mobility of the
nodes, path loss, shadowing and multi-path fading due to reflection, diffraction, scattering,
absorption lead to slow and fast variations in channel quality even within the milliseconds
scale (Proakis, 2004).
 The wireless medium is a broadcast medium. Therefore, in contrast to wired networks,
the interference caused by other in-range traffic can unlimitedly disturb a transmission. This
causes the wireless link capacity to depend also on the sensitivity of receivers in sensing the
environment as well as other links status in terms of their transmission range and power.
 Packet reception reliability over a link depends on several parameters such as
modulation, source/channel coding of that link, the sensitivity of the link and the length of
the packets.
Radio channels have some additional features such as asymmetrical nature and non-
isotropic connectivity (Ganesan et. al, 2002; Cerpa et. al, 2003; Zhou et. al, 2004). Asymmetry
of the channels means connectivity from node A to node B might differ significantly from B
to A and non-isotropic connectivity means nodes geographically far away from source may
get better connectivity than nodes that are geographically closer.
As a result of these characteristics, the radio cell is neither binary nor static. From the
perspective of a node, the set of other nodes it can hear and the loss probability to or from
these nodes vary abruptly over time with a large magnitude. This has been widely
conrmed in real platforms (Couto et al, 2002; Ganesan et. al, 2002; Cerpa et. al, 2003; Couto,
2004). These random variations induce much more complexity for wireless networks to
guarantee performance to transmit real-time or even critical data.
1.2 Limitations of Shortest Path Routing

Most of the existing routing algorithms use the shortest-path metric to find one or more
multi-hop paths between the node pairs (Perkins & Royer, 1994; Johnson & Maltz, 1994;
Park & Corson, 1997; Perkins & Belding-Royer, 2003; De Couto et. al, 2005). The advantage
of this metric is its simplicity and a low overhead to the network. Once the topology is
known, it is easy to find a path with a minimum number of hops between a source and a
destination without additional measurement and overhead. Recent researches show that
choosing the path with the smallest number of hops between nodes often leads to poor
performances (De Couto et. al, 2002; De Couto et. al, 2005; Yarvis et. al, 2002).
One of the limitations of shortest path routing is that it does not capture the variable nature
of wireless links. Instead, it assumes that the links between nodes either work well or do not
work at all. Figure 2 shows an illustration of the different assumptions made by minimum-
hop routing and link quality aware routing on the wireless links. This shows that the
arbitrary choice made by minimum hop-count is not likely to select the best path among the
same minimum length with widely varying qualities. Moreover, one of the current trends in
wireless communication is to enable devices to operate using many different transmission
rates to deal with changes in connectivity due to mobility and interference. In multi-rate
wireless networks, minimum-hop works even worse. Selecting the minimum-hop paths
leads to maximizing the distance travelled by each hop, but longer links are not robust
enough to operate at the higher rates. Therefore, shortest path routing results in selecting
the paths with the lowest rates, which degrades dramatically the overall throughput of the
network.
Fig. 2. Different assumptions for wireless link connectivity made by minimum-hop routing
and link quality aware routing.
Furthermore, transmitting the flow over the low-rate links degrades the performance of
other flows, which are transmitted over higher rate links. The main reason of this effect is
that slow-speed links require larger amount of medium time to transmit a packet over the
shared wireless medium and so block the other flows for a longer time. Heusse et al. denotes
this problem as Performance Anomaly of 802.11b (Heusse et. al, 2003). They show
analytically that a contending node with lower nominal bit rate degrades the throughput of
faster contenders to even a lower bit-rate than the slowest sender. (Mahtre et. al, 2007;
Razandralambo et. al, 2008; Choi et. al, 2005) have evaluated and shown the same effect.
Another effect of multi-rate option for a minimum hop routing is that in multi-rate networks
broadcast packets benefit from the longer range of low rate transmissions to reach farther
nodes and so are always sent at the lowest transmission rate. Therefore, hearing the
broadcast Hello messages from a node is not a good enough basis for determining that two
nodes are well connected for transferring data packets at high rates. Lundgren et al. have
referred to this effect as the gray-zone area (Lundgren et. al, 2002). A gray zone is the
maximum area, which is covered by the broadcast messages at low rate, but not all the
nodes in this area can forward the packets at high rates.
Choosing closer nodes with shorter-range links instead of minimum-hop routes can solve
this problem. Consequently, minimum-hop metric has no flexibility in dealing with random
quality fluctuations of the links. Link quality aware routing counters these limitations by
using observation of miscellaneous parameters such as frame delivery or signal strength to
select the good paths. In this approach, link quality metric is measured and observed in
order to predict the near future quality of the links. This estimation is then used to
determine the best route.
1.3 Cross Layer Interaction as a Solution

Typically, Open System Interconnection protocol stack (OSI) is divided into several layers
which are designed independently. The interactions between adjacent layers are defined by
some specific interfaces. Recently, in the quest of finding a link quality aware routing for
wireless multi-hop networks, numerous link quality aware metrics have been proposed,
which most of them are based on cross layer interactions between various layers of the
protocol stack. Lately, there are many research efforts which show that transferring the
status information between the layers can lead to a great improvement in network
performance (Conti et. al, 2004; Shakkottai et. al, 2003; Goldsmith & Wicker, 1998). Recent
activities of IEEE 802.11 task group in mesh networking have released IEEE 802.11s. It
extends the IEEE 802.11 Medium Access Control (MAC) standard by defining an
architecture and protocol that support both broadcast/multicast and unicast delivery using
radio-aware metrics over self-configuring multi-hop topologies. This evolution pushes
employing the cross layering technique in the real platforms in near future.
This chapter argues that the cross layering technique can be a promising solution in
providing flexibility to the wireless network changes. Nevertheless evaluating the benefits
of cross layer routing is often only based on the throughput, which is simplistic. Current
studies generally do not consider the impact of other criteria such as response time or route
flapping, which influence greatly applications performances in terms of throughput, but
also mean delay, jitter and packet loss.
In the next section, a state-of-the-art of the main cross- layering metrics that have been
proposed in the literature are presented. Then, the concept of reactivity for a link quality
aware routing as a mean to analyse the true benefits of the cross-layer routing is introduced.
Section 4 concludes this chapter.
2. Link Quality Aware Routing

Most of the primitive works in routing protocols for wireless multi-hop networks are
inherited from existing routing protocols in wired networks. They devise mostly on coping
with changing topology and mobile nodes (Perkins & Bhagwa, 1994; Perkins & Royer, 1999;
Johnson & Maltz , 1994) and traditionally find the possible routes to any destination in the
network with the minimum hop-count. As explained in Section 1.2, shortest path routing
has sub-optimal performance, as they tend to include wireless links between distant nodes
(De Couto et. al, 2002). A multitude of quality aware metrics have been proposed in the last
decade which deal with the strict bandwidth and variable quality of wireless links and try to
overcome the disadvantages of the minimum hop (MH) routing. Although most of them
have been designed with the objective of increasing the transport capacity, each of them
considers different QoS demands such as overall throughput, end-to-end delay, etc.
Therefore, the proposed approaches for link quality aware routing can be categorized
according to the aim of their design.
2.1 Wireless Network Capacity

The main purpose of efficient routing in mesh networks is improving the achieved capacity.
The notion of capacity for wireless ad hoc network was defined as the maximum obtainable
throughput from the network. It was first introduced by Gupta and Kumar in their seminal
work (Gupta & Kumar, 2000). Network capacity for wireless multi-hop networks is
generally unknown, except for some centralized scheduling-based MAC protocols like Time
Division Multiple Access (TDMA) where the problem finds a mathematical formulation.
The main finding in (Gupta, 2000) is that per-node capacity of a random wireless network
with n static nodes scales as  ( 1 ) . They assume a threshold-based link layer model in
n log n
which a packet transmission is successful if the received SNR at the receiver is greater than a
fixed threshold. Instead of this ideal link layer model, Mhatre et al. considere a probabilistic
lossy link model and show that the per-node throughput scales as only  ( 1 ) instead of
n
1 (Mahtre & Rosenberg, 2006).
( )
n log n
These asymptotic bounds are calculated under assumptions such as node homogeneity and
random communication patterns. Therefore, some researches try to relax some of these
assumptions on network configuration. Jain et al. focus on interference status among the
transmitters as one of the main limiting factors on routing performance (Jain et. al, 2003).
They propose to represent interference among wireless links using a conflict graph. A
conflict graph shows, which wireless links interfere with each other, such that each edge in
the connectivity graph is represented by a vertex and there exists an edge between two
vertices if the links interfere with each other. Thus, the throughput optimization problem is
posed as a linear programming problem in which upper and lower bounds of the maximum
throughput are obtained by finding the maximal clique and independent set in the conflict
graph.
Karnik et al. extend the conflict graph idea to a conflict set (Karnik et. al, 2007). Their
rational is that an interference model can not be binary since for a given link, generally there
is a subset of links that at least one of them should be silent when the given link is
transmitting. They propose a joint optimization of routing, scheduling and physical layer
parameters to achieve the highest throughput. Both these proposals bring valuable
achievement but as they investigate the highest capacity of a network, they have to assume
TDMA instead of contention-based algorithm, which leads to probabilistic results. Hence,
they implicitly assume that data transmissions are scheduled by a central entity. Therefore
they may not be applied easily to more practical networks such as IEEE 802.11 with random
access to the channel.
Computing the optimal throughput, despite of giving a good vision to the maximum
achievable throughput in the network, may not be implementable in a real network. There
are many complicated issues such as necessity of having a distributed routing/scheduling
protocol, random quality for the wireless links, limited allowable overhead to the network,
compatibility with MAC 802.11, etc., that motivate the researchers to find a practical
solution to achieve a good performance.
2.2 Maximum Throughput Routing

Most of the work done in this area relies on the broadcasting of extra probe packets to
estimate the channel quality (ex. Sivakumar et. al, 1999; De Couto et. al, 2005; Draves & Zill,
2004). However, since the quality of the wireless links depends significantly on physical
settings (such as transmit rate and packet size), the probes may not reflect the actual quality
of the links. The reason is that for preventing to throttle the entire channel capacity they
have to use small-sized probes at low transmission rate. Therefore the quality experienced
by larger data packets at variable transmission rate is not the same as probe packets.
De Couto et al proposes a simple and effective routing metric called the expected
transmission count (ETX) for 802.11-based radios employing link-layer retransmissions to
recover frame losses (De Couto et. al, 2005). ETX of a wireless link is defined as the average
number of transmissions necessary to transfer a packet successfully over a link. For
estimating the expected number of transmission of the links, each node broadcasts
periodically fixed-size probe packets. This enables every node to estimate the frame loss
ratio pf to each of its neighbors over a window time, and obtain an estimate pr of the reverse
direction from its neighbors. Then, assuming uniform distribution of error-rate over each
link, the node can estimate the expected transmission count as 1 .The ETX of
(1  p f )(1  p r )
a path is obtained by summing up the ETX of its links. Therefore, each node picks the path
that has the smallest ETX value from a set of choices. ETX has several drawbacks. First, its
measurement scheme by using identical small-sized probe packets does not reflect the actual
error-rate that the data packets experience over each link. The accuracy of the measurement
scheme of a link quality metric has a great impact on its functionality (Karbaschi et al. 2008).
Furthermore, ETX does not account the link layer abandon after a certain threshold of
retransmissions. This may induce to select a path, which contains links with high loss rate.
Koksal et al. introduce another version of ETX, called ENT (Effective Number of
Transmissions), which deals with this problem (Koksal & Balakrishnan, 2006). ENT takes
into account the probability that the number of transmissions exceeds a certain threshold
and then calculates the effective transmission count based on an application requirement
parameter, which limits this probability. Moreover, ETX by taking an inversion from the
delivery rate to get the expected number of transmissions implicitly assumes uniform
distribution for the bit error rate (BER) of the channel, which may not be correct.
The Expected Transmission Time (ETT), proposed by Draves et al. improves ETX by
considering differences in link transmission rates and data packet sizes (Draves & Zill,
2004). The ETT of a link l is defined as the expected MAC layer latency to transfer
successfully a packet over link l. The relation between ETT and ETX of a link lis expressed
as:
sl
ETTl  ETX (1)
rl
where rlis the transmission rate of link land slis the data packet size transmitted over that
link. The weight of a path is simply the summation of the ETT’s of the links of that path. The
drawback of ETT is, as it is based on ETX, it may choose the paths, which contain the links
with high loss rates. (Draves & Zill, 2004) proposes also another new metric based on ETT,
which is called Weighted Cumulative Expected Transmission Time (WCETT). The purpose
of this metric is finding the minimum weight path in a multi-radio network. The WCETT is
motivated by observing that, enabling the nodes with multi-radio capability reduces the
intra-flow interference. This interference is caused by the nodes of a path of a given flow
competing with each other for channel bandwidth. For a path p, WCETT is defined as:
WCETT ( p)  (1   ). ETTl   . Max ( X j ) (2)

lP 1 j k
where X j is the number of times channel jis used along path p and  is an adjustable
parameter for the moving average subject to 0    1 . (Yang et. al, 2005) shows that
WCETT is not non-isotonic and thus it is not a loop-free metric.
One of the characteristics of wireless links that can be observed is the received signal
strength. It is very attractive if link quality can be reliably inferred by simply measuring the
received signal strength from each received packet. Theoretically, the BER is expected to
have a direct correlation with the received signal-to-noise ratio (SNR) of the packet, and the
packet error rate is a function of the BER and coding. Therefore, the SNR level of the
received packets has been widely used as a predictor for the loss rate of the wireless links
(ex. Goff et. al, 2001; Dube et. al, 1997). (Aguayo et. al, 2004; Woo, 2004) through collecting
experimental data have shown that although SNR has an impact on the delivery probability,
lower values of SNR has a weak correlation with the loss rate of the links. Thus, it can not
predict the quality of the links easily. In addition, (Woo, 2004) illustrates that where traffic
load interference happens, collisions can affect link quality even though the received signal
is very strong. The main reason is that prediction of the link quality by observing the SNR
samples of the packets, counts only on the packets which are received successfully. This
leads to ignoring the congestion status of the links.
A number of proposed wireless routing algorithms collect per-link signal strength
information and apply a threshold to avoid links with high loss ratios (Goff et. al, 2001;
Yarvis et. al, 2002). In the case that there is only one lossy route to the destination, this
approach may eliminate links that are necessary to maintain the network connectivity.
2.3 Minimum Delay Routing

Some existing link quality metrics focus on finding the best paths based on the end-to-end
delay associated to each path. The rational for minimizing the paths latency is that in a fixed
transmission power scenario, packets latency for reaching successfully to the other end of a
link can provide an estimation of the quality of that link. The average round trip time (RTT)
of the packets over each link is one of the delay based parameters representing the link
quality. (Adya et. al, 2004) for instance proposes this metric. To calculate RTT, a node sends
periodically a probe packet carrying its time stamp to each of its neighbors. Each neighbor
immediately responds to the received probe with a probe acknowledgment which echoes its
time stamp. This enables the sending node to calculate the RTT to each of its neighbors. Each
node keeps an average of the measured RTT to each of its neighbors. If a probe or a response
probe is lost, the average is increased to reflect this loss. A path with the least sum of RTTs is
selected between any node pair.
The RTT reflects several factors, which have impact on the quality of a link. First, if a link
between the nodes is lossy its average RTT is increased to give a higher weight to that link.
Second, either if the sender or the neighbor is busy, the probe or its response is delayed due
to queuing delay which leads to higher RTT. Third, if other nodes in the transmission range
of the sender are busy, the probes experience higher delay to access the channel again
resulting in higher RTT. Concisely, RTT measures the contention status and error rate of a
link.
However, the small probe packets in comparison to larger data packets are rarely dropped
over a lossy channel. This hides the actual bandwidth of the links. Moreover, (Draves et.
al, 2004) shows that RTT can be very load-sensitive which leads to unnecessary route
instability. Load-dependency of a metric is a well-known problem in wired networks

(Khanna& Zinky, 1989). To suppress the queuing delay in the RTT, Keshav, 91, proposed the
packet-pair technique to measure delay of a link. In this approach to calculate the per-hop
delay, a node sends periodically two probe packets back to back to each of its neighbors
such that the first probe is small and the next one is large. The neighbor upon receiving the
probes calculates the delay between them and then reports this delay back to the sender.
The sender keeps an average from the delay samples of each neighbor and paths with lower
cumulated delay are selected. This technique, by using larger packet for the second probe,
reflects more accurately the actual bandwidth of the links, although it has higher overhead
than RTT. Draves et. al, 2004 discusses again that packet-pair measurement is not
completely free of self interference between the neighbors, although less severe than RTT.
Awerbuch et. al 2004 proposes the Medium Time Metric (MTM) which assigns a weight to
each link proportional to the amount of medium time consumed by transmitting a packet on
the link. It takes the inverse of the nominal rate of the links to estimate the medium time.
The variable rate of the links is determined by an auto-rate algorithm employed by the
networks, such as ARF or RBAR. Existing shortest path protocols will then discover the path
that minimizes the total transmission time. This metric only handles the transmission rate of
the links and does not account the medium access contention and retransmission of packets
at the MAC layer. Zhao et al. 2005 introduces a cross layer metric called PARMA, which
aims to minimize end-to-end delay which includes the transmission delay, access delay and
the queuing delay.
They consider a low saturated system with ignorable queuing delay. A passive estimation is
used for the channel access delay and the transmission delay on each link is calculated as the
ratio of packet length to the link speed. This metric has a good insight into estimating the
total delay of each link but has simplified very much the problem of the delay calculation.
For instance, it assumes that the links are error free and no packet retransmissions occur
over them.
2.4 Load Balancing

In order to make routing efficiently and increase network utilization, some researchers have
proposed congestion aware routing with the aim of load balancing in the network. One of
the methods for spreading the traffic is using multiple non-overlapping channels. Kyasanur
& Vaidya, 2006 propose a Multi-Channel Routing protocol (MCR) with the assumption that
the number of interfaces per node is smaller than the number of channels. The purpose of
their protocol is choosing paths with channel diversity in order to reduce the self
interference between the node pairs. Moreover, they take into account the cost of interface
switching latency. MCR is based on on-demand routing in a multi-channel network. While
the on-demand route discovery provides strong resistance to mobility-caused link breaks,
the long expected lifetimes of links in mesh networks make on-demand route discovery
redundant and expensive in terms of control message overhead. Therefore, this protocol is
not totally appropriate for mesh networks. Yang et al. focus more on mesh networks and
propose another path weight function called Metric of Interference and Channel-switching
(MIC). A routing scheme, called Load and Interference Balanced Routing Algorithm
(LIBRA) is also presented to provide load balancing (Yang, 2005). MIC includes both
interference and channel switching cost.
2.5 Routing with Controlling Transmission Power

Numerous works in efficient routing in multi-hop networks has focused on power control
routing. The problem of power control has been investigated in two main research
directions: energy-aware and interference-aware routing. In energy-aware routing
approaches the objective is to find power values and routing strategy, which minimizes the
consumption of power in order to maximize the battery lifetime of mobile devices.
Therefore these works are suitable for sensor and ad hoc networks as in wireless mesh
networks power is not a restricted constraint. Power control in interference-aware routing
aims to find the optimal transmission power which gives the higher throughput or the
lowest end-to-end delay. Therefore, transmission power of the nodes is controlled in order
to reduce interference while preserving the connectivity. There are a lot of researches in this
area. For instance Iannone et al. propose Mesh Routing Strategy (MRS) in which
transmission rate, PER and interference of each link are taken into account (Iannone &
Fdida, 2006). The interference is calculated based on the transmission power and number of
reachable neighbors with that power level. The disadvantage of their approach is that they
do not consider that links with different transmission rate have different sensibility for being
disturbed by the neighbors’ transmission. This effect has been taken into account in (Karnik
et. al., 2008) where the authors propose a network configuration to have an optimal
throughput.
The foreseen alternatives to minimum hop metric consist in establishing high quality paths,
by tracking various link quality metrics in order to significantly improve the routing
performances. Thus, the challenge lies in selecting goodpaths, based on a relevant link
quality metric. However, the stability issue of link quality aware routing which can be
extremely important specially in providing quality of service for jitter sensitive applications
has not been addressed by the existing research efforts. In the next section, a quantitative
tool to investigate the routing reactivity and its impact on applications performances is
introduced.
3. Reactivity of Link Quality Aware Routing

A more reactive routing responds faster to link quality changes. This leads to detect the
lossy channel faster and so, to converge to the higher quality path in a shorter time.
Meanwhile, fast reacting to channel variations may produce higher path flapping and
consequently higher jitter level. Therefore, there is a trade-off between providing ensured
stability in selecting the paths and obtaining a high possible throughput from the network.
The frequency of link quality changes may be very different for distinct wireless links (due
to some factors such as fast or slow fading, nodes mobility, etc.) (Koskal & Balakrishnan,
2006; Aguayo et. al, 2004). In order to track as much as possible all the link changes and
always choose a high quality path, the routing should respond accurately and as fast as
possible to these changes. Response time refers to the time required by the routing agent to
take into account the new link quality status.
The reactivity of the routing depends on the updating frequency of the routing tables and
the sensitivity degree of the routing metric to channel variations. The updating frequency of
the routing tables defines the rate at which the shortest paths are recalculated based on the
current value of the link quality metrics. Although the updating period of the routing is
generally longer than the time-scale of link quality variations, a shorter update period is able
to respond faster to link breakage or quality degradation and in turn will lead to a higher
throughput. However, frequent changes of the selected route induce packet reordering and
jitter issues. Moreover, reducing the updating period of the metric obviously produces a
higher amount of routing overhead. This may overload the network and could severely
degrade the network performance. The sensitivity degree of the routing metric to link
quality variations is the other parameter which obviously has a great impact on the routing
response time. The sensitivity degree depends on the way the set of measured parameters
(frame loss, delay, SNR, ...) are mapped onto the metric. Sensitivity degree S of a link quality
metric is defined as the norm of the gradient of the defined metric function with respect to
the set of parameters that measures the link quality. Let q be the set of measured parameters
and m(q) the calculated metric based on q. The sensitivity of the metric is:
S m (q )  m(q ) (3)
With a highly sensitive metric, the variations of link conditions are intensified. The path
metric, which aggregates the link metrics along a path, fluctuates faster and the probability
of changing the selected route increases. The possibly resulting route flapping may cause
higher jitter which for some applications is harmful as reordered or delayed packets may be
considered as lost ones. This section focuses on investigating the impact of a more sensitive
metric on route flapping, control overhead and real-time application performances.
3.1 Impact of the Sensitivity of a Link Quality Metric

In (Karbaschi et. al, 2008) it is shown that measurement scheme and obviously the relevance
of the measured parameters have a great impact on the measurement accuracy and thus on
the final result. Numerous link quality routing have been proposed in the last decade.
However, in order to compare the impact of the sensitivity of two link quality metrics on the
routing performance, their measurement scheme and the parameters they measure for
estimating the quality of the links should be the same. No such two link quality metrics with
identical observed parameters and measurement scheme can be found in the literature.
Therefore, to conduct the study, two comparable and realistic link quality metrics are
introduced in the following.
ARQ mechanism in 802.11b with retransmission of frames over lossy channels wastes
bandwidth and causes higher end-to-end delay and interference to the other existing
traffics. Therefore, the number of frame retransmissions at the MAC layer has been widely
used as an estimator of the link quality (De Couto et. al, 2005; Koskal & Balakrishnan, 2006;
Karbaschi et. al, 2008). Therefore, this section presents two comparable link quality metrics
based on this retransmissions number.
The first link quality metric, called m 1 , is based on the FTE metric introduced in (Karbaschi
et. al. 2005). Assuming that RTS/CTS is enabled for solving the hidden terminal problem,
the quality and interference status of the adjacent links of a sender can be estimated by
measuring the average number of required retransmissions of data and RTS frames at the
MAC layer to transfer a unicast packet across a link. Therefore, each node measures the m 1
by keeping the retransmissions count of RTS and data frames over the neighbor links as
follows.
Let k xy (i ) - respectively l xy (i ) - be the number of transmissions (including retransmissions)

of the i data packet – respectively i th RTS packet - over the x to y link. Thus the set of
th

measured parameters ( q ) over this link will be defined as q xy (i )  k xy (i ), l xy (i ) .The 
i
success rate in delivering two frames of data and RTS from x to y denoted by m1 (qxy ) is
computed as:
i 2 (4)
m1 (qxy )
k xy (i )  lxy (i)
Referring to Equation 4, increasing the number of retransmissions of RTS or data frames

reduces the value of m1 and for a perfectly efficient link i
m1 (qxy ) is equal to unity. If the
number of retransmissions reaches a predefined threshold, the sender gives up sending the
i
frame. In this case, m1 (qxy ) is set to zero which degrades the overall average link quality
very much. m1 can be interpreted as an estimation of the success rate of transmissions over
a link. Another metric, called m2 is defined based on the same measured parameters.
Assuming that the failure in transmission of data and RTS frames are independent from
each other, the success rate of transmission over the link x to y can be calculated by
multiplying the success probability of sending RTS and data frames as follows:
i 1 1
m2 (q xy )  (5)
k xy (i ) l xy (i )
i
With the same argument, if the sender gives up sending RTS or data frames, m2 (q xy ) is set
to zero.
Figure 3 illustrates the calculated success rate returned by m1 and m2 as a function of the
number of data and RTS retransmissions for each sent packet over a given link (Equation 4
and 5). As shown in this figure, an interesting property of both m1 and m2 is that their
variations over the range of RTS and data retransmission numbers is not uniform. Indeed,
both metrics are much more sensitive to a given variation of its arguments k xy and l xy when
these parameters ranges between 1 and 4 than for values ranging between 6 to 10. In other
words, the quality variations of a poor link are far less reflected in the metric than the
variations of a high quality link. This is desirable since if a link does not work well and there
is no alternate much higher quality link, it is not worth changing the selected path. In
counterparts, quality variations of a good links have a much greater impact on the
throughput of that link. As a result, good quality links should be more prone to changes.
Fig. 3. Measured link quality using m1 and m2 function of the number of Data and RTS
frames retransmissions.
From this point of view, the two metrics differ. Indeed, m2 differentiates better than m1 a
small degradedness from a former high quality measured value. As clearly shown in
Figure 3, both functions return 1 when no retransmission occurs which confirms the value of
100% success for the transmission while by lessening in link quality, m2 drops more sharply
than m1 . For example, an increase in the number of data transmissions from 3 to 4 causes
32% decrease in m2 and about 15 % in m1 . Therefore, m 2 is called as Faster FTE (FFTE).
The variations of the metric with respect to quality changes can be evaluated using the
sensitivity degree (Equation 3). Figure 4 compares the sensitivity degree of m1 and m2 using
the difference S m1 ( q xy )  S m 2( q xy ) . We see that for all the variations range of k xy (i ) and
l xy (i ) , S m 2 is larger or equal to S m1 .
Consequently, m2 has an even greater sensitivity in detecting changes in the estimated link
quality than m1 . m2 obliges the routing agent to be more reactive and changes the selected
route more often than m1 .
Fig. 4. Comparison of the sensitivity of the metrics m1 and m2 .
3.2 Simulation Study

This section presents simulation results to illustrate the performance of the link quality
metrics compared to the Minimum Hop (MH) metric as the reference. Both m 1 and m 2 have
been employed in DSDV (Perkins, 1994). The efficiency of the routes is estimated by
multiplying the EWMA of m 1 values (or m 2 ) along the path towards the destination. This
estimation of the link quality is piggy backed into the Hello messages that are sent in a
periodic manner.
The simulations are performed under ns2.28 with the enriching the simulator in order to
contain wireless channel fading effects, time variable link quality for wireless links, signal to
interference and noise ratio, etc (cf. Karbaschi, 2008).
In order to show the impact of the link quality aware routing on the quality of service for a
jitter-sensible flow, VoIP traffic of is modelled and multiple random connections are set in a
30-nodes random topology. VoIP is basically UDP packets encapsulating RTP packets which
contain the voice data. For accurately modelling the bursty VoIP traffic, Pareto On/Off
traffic is used (Dang et. al, 2004), with different transmission rates corresponding to the
widely used ITU voice coders.
Firstly, the impact of the sensitivity of the metric on the performance of the DSDV is
evaluated and then the resultant instability and the generated jitter are investigated. T (resp.
T0) are used as the routing update period used in the case that m1 and m2 (resp. MH) are
implemented
In order to unify the impact of the update period on the reactivity of the routing, T and T0
are set to 15 s. The three metrics in terms of received throughput, defined as the average
number of data bits received per second, are compared in Figure 5. The result of one
connection confirms that both the link quality metrics are able to transfer more bits in
comparison to the MH metric in a time duration of 2000 s. Figure 5 also shows that m 2
outperforms the two other metrics. This confirms that m 2 , the more sensitive metric, is able
to find a higher throughput path faster than m 1 and so reduces the packet drops.
Fig. 5. Average received throughput with same routing update period (T = T 0 = 15s)
Comparing in Figure 6 the number of times that one dedicated flow flaps per second reveals
that a link quality metric leads the routing to change the selected path more frequently. This
effect is even greater when the sensitivity of the metric increases (m 2 compared to m 1 ). To
show the impact of the metrics on the routing overhead, the average bit rate of control
messages for the three metrics are compared in Figure 7. This reveals that the overhead
generated by m 1 and m 2 are nearly the same and both more than MH’s overhead. The reason
is that the higher sensitivity of m 1 and m 2 generates more paths changes than MH. This
obliges the nodes to piggy back more neighbors entries into their broadcast message and
makes it larger, thus raising the routing overhead. Another cause of overhead increase is
that a link quality metric needs a larger field in the control message than a hop metric (32
bits compared 16 bits). This enlarges the overall size of the control messages.
Fig. 6. Number of path changes per second per flow.
Fig. 7. Routing overhead for different VoIP coder types.
To see the efficiency of functionality of the link quality metrics, the evaluation is repeated by
adjusting the amount of overhead to an identical value for the three metrics by tuning the
value of T to 30 s. As explained in Section 3 this may reduce the throughput of m 1 and m 2
due to lower update rate of the metric. However, the average throughput comparison shows
that m 2 still brings a much higher throughput (Figure 8).
Fig. 8. Average received throughput with same overhead amount (T = 30 s, T 0 = 15 s) for

different VoIP coder types.
Flapping the selected route may cause the consecutive packets to be routed through
different routes. The subsequent instability of the selected path may cause a higher jitter
level. Figure 9 illustrates the measured jitter per packet for the three metrics during a sample
interval of a VoIP connection. Table 1 gives the mean and standard deviation of the jitter
measured using the three metrics, which gives an idea of the spreading of the jitter
distribution.
MH m1 m2
mean (ms) 9.5 9.6 9.43
std (ms) 20 25 30.2
Table 1. Jitter statistics
The jitter experienced using m 2 is much greater than the jitter observed with m 1 and MH.
High jitter levels can have a great impact on the perceived quality in a voice conversation
and as a result, many service providers now account for maximum jitter levels.
Fig. 9. Comparison of the jitter per received packet in a VoIP connection.
Most of the VoIP end-devices use a de-jitter buffer to compensate the jitter transforming the
variable delay into a fixed delay (Khasnabish, 2003). Thus, high levels of jitter increase the
network latency and cause a large number of packets to be discarded by the receiver. This
may result in severe degradation in call quality. Therefore, real-time applications may not
benefit from the higher throughput obtained with a more sensitive metric.
4. Conclusion
Wireless multi-hop networks are a promising technology to provide flexibility and rapid
deployment for connecting the users at low cost. This chapter has examined the issues of
link quality aware routing for wireless multi hop networks. Since the quality of wireless
communications depends on many different parameters, it can vary dramatically over time
and with even slight environmental changes. The peculiarity of wireless links and strong
fluctuations in their quality lead to challenges in designing wireless multi hop routings.
Therefore the necessity of having more efficient routing rather than the ones proposed for
wired networks has arisen. Traditional hop count shortest-path routing protocols fail to
provide reliable and high performance because of their blindness to under layer status. This
draws lots research efforts to improve the routing performance through choosing good
paths via transferring link status information from under layers.
This chapter addressed the stability issues of link quality aware routing which can be
extremely important specially in providing quality of service for jitter sensitive applications.
It was argued that having a reactive routing to cope with random changes of wireless links
is essential. A quantitative tool for estimating the sensitivity of a link quality metric was
introduced which indicated how strongly the metric reflects the quality changes. It was
shown that the sensitivity has a great impact on the routing adaptivity. To illustrate this,
two comparable link quality metrics (FTE and FFTE) with different sensitivity were
introduced and the routing performance observed with these metrics was compared by
simulation. It was shown that having a sensitive metric can improve the routing
functionality in terms of transferring a higher number of data packets through the network.
However, one resulting side-effect is more oscillation in path selection. This leads to higher
jitter level which a delicate application such as VoIP may not tolerate.
5. References
Adya, A. ; Bahl, P. ; Padhye, J. ; Wolman, A. & Zhou, L (2004). A multi-radio unication
protocol for ieee 802.11wireless networks, FirstInternationalConferenceonBroadb
Networks (BROADNETS’04) , ISBN 0-7695-2221-1, pp. 344–354, October 2004, San
Jose, California, USA, IEEE Computer Society, 2004.
Aguayo, D. ; Bicket, J. ; Biswas, S. ; Judd, G. & Morris, R. (2004). Link-level measurements
from an 802.11b mesh network, Proceedings of ACM Sigcomm , Vol. 34, No. 4, pp.
121–132, Aug-Sept 2004, ACM, 1-58113-862-8/04/0008.
Awerbuch, B. ; Holmer, D. & Rubens, H. (2004). High throughput route selection in multi-
rate ad hoc wireless networks, Kluwer Mobile Networks and Applications (MONET)
JournalSpecialIssueon etIntern
WirelessAccess:1.208andBeyond , Vol. 2928, pp. 253–
270, ISBN 3-540-20790-2, January 2004.
Cerpa, A. ; Busek, N. & Estrin, D. (2003). Scale: A tool for simple connectivity assessment in
lossy environments, In CENS Technical Report 021 , September 2003, experimental
evaluation of wireless links.
Conti, M. ; Maselli, G. ; Turi, G. & Giordano, S. (2004). Cross-Layering in Mobile Ad Hoc
Network Design, Computer , Vol. 37, No. 2, pp. 48-51, Feb. 2004,
doi:10.1109/MC.2004.1266295
Choi, S. ; Park, K. & Kim, C. (2005). On the performance characteristics of WLANs:
Revisited. Proceedings of the 205 ACM SIGMETR ICS international conference on
Measurementandmodelingofcomputer , pp.systems
97-108, ISBN 1-59593-022-1, Banff,
Alberta, Canada, ACM, New York, USA.
Dang, T. D. ; Sonkoly, B. & Molnar, S. (2004). Fractal analysis and modelling of VoIP traffic,
In International Telecommunications Network Strategy andISBN: Planning Sympos
3-8007-2840-0, pp. 556–562, June 2004, DOI 10.1109/NETWKS.2004.1341826
De Couto, D. S. J. ; Aguayo, D. ; Chambers, B. A. & Morris, R. (2002). Performance of
multihop wireless networks: Shortest path is not enough. Proceedings of the First
Workshop on Hot Topics in Networks (HotNets- . Princeton,
I) New Jersey: ACM
SIGCOMM, October 2002.
De Couto, S. J. D. (2004). High-throughput routing for multi-hop wireless networks. Ph.D.
dissertation, Department of Electrical Engineering,and Massachusetts
Computer Science
Institute of Technology, June 2004.
De Couto, D. S. J., Aguayo, B ; Chambers, B. A. & Morris R. (2005). A high-throughput path
metric for multi-hop wireless routing, Wireless Networks , Volume 11 , Issue 4, July
2005, pp 419 – 434, ISSN:1022-0038.
Draves, R. ; Padhye, J. & Zill B. (2004). Routing in multi-radio, multi-hop wireless mesh
networks, Proceedings of the 10th annual international conference on Mobile
and Networking , pp. 114–128, ISBN:1-58113-868-7, International Conference on
Mobile Computing and Networking, Philadelphia, PA, USA, ACM, New York, NY,
USA, 2004.
Dube, R. ; Rais, C. D. ; Wang, K.-Y. & Tripathi, S. K. (1997). Signal stability-based adaptive
routing (ssa) for ad hoc mobile networks, IEEE Personal Communications , Vol 4, pp
36-45, February 1997, Digital Object Identifier 10.1109/98.575990
Ganesan, D. ; Krishnamachari, B. ; Woo, A. ; Culler, D. ; Estrin, D. & Wicker, S. (2002).
Complex behavior at scale: An experimental study, UCLA Computer Science, Tech.
Rep. TR 02-0013, 2002.
Goldsmith A. & Wicker S. (1998). Design challenges for energy-constrained ad hoc wireless
networks,IEEEWirelessCommunications , vol. 9, no. 4, pp. 8–27, 2002.
Goﬀ, T. ; Abu-Ghazaleh, N. B. ; Phatak, D. S. & Kahvecioglu, R. (2001). Preemptive routing
in ad hoc networks, ProceedingofACM/IEEEMobiCom, pp. 43–52, ISBN 1-58113-422-
3, July 2001, ACM SIGMOBILE, Rome, Italy, 2001.
Gupta P. & Kumar P. R. (2000). The capacity of wireless networks, IEEE Transactions on
InformationTheory , March 2000, Vol. 46, No. 2, pp. 388–404, Digital Object Identifier
10.1109/18.825799.
Hedrick, C. (1998). Routing information protocol. Network working group , RFC 1058, The
Internet Society, June 1988.
Heusse, M. ; Rousseau, F. ; Berger-Sabbatel, G. & Duda, A. (2003) . Performance anomaly of
802.1b , Proceedings of INFOCOM, Twenty-Second Annual Joint Conference of the
IEEE Computer and Communications Societies. Vol. 2, pp. 836–843, 2003, ISBN: 0-
7803-7752-4, ISSN: 0743-166X.
Iannone, L.& Fdida, S. (2006). Evaluating a cross-layer approach for routing in wireless
mesh networks, Telecommunication Systems Journal Springer) ( Special issue: N
Generation Networks - Architectures, Protocols, , vol. 31, no. 2–3, pp. 173–
Performance
193, March 2006.
Jain, K. ; Padhye, J. ; Padmanabhan, V. & Qiu L. (2006). Impact of interference on multi-hop
wireless network performance , MobiCom '03: Proceedings of the 9th annual
international conference on Mobile computing and networking, pp. 66--80, ACM
Press, New York, NY, USA, 2003.
Johnson D. B. & Maltz, D. A. (1994). Dynamic source routing in ad-hoc wireless network
Proceedings of the Workshop on Mobile Computing Systems and Application, pp.
158–163, December 1994.
Karbaschi, G. (2008). Link Quality Aware Routing in IEEE 802.11 Multi Hop Networks, PhD
Thesisdissertation,of PierreetMarie,Curie
University April 2008.
Karbaschi, G. & Fladenmuller, A. (2005). A link-quality and congestion-aware cross layer
metric for multi-hop wireless routing, Proceedings of IEEE Mobile Ad hoc and Sensor
SystemsMASS) ( , Nov. 2005.
Karbaschi, G. ; Fladenmuller, A. & Wolnger, B. E. (2008) Link-quality measurement
enhancement for routing in wireless mesh , Proceedings
networksof International
Symposium on a World of Wireless, Mobile and Multimedia tworks,
Ne 208 WoWMoM
(
208), ISBN: 978-1-4244-2099-5, pp. 1-9, Digital Object Identifier:
10.1109/WOWMOM.2008.4594842
Keshav, S. (1991). A control-theoretic approach  ow controlto, ACM SIGCOMM Computer
Communication Review, Vol. 21, no. 4, pp. 3–15, 1991, ISSN 0146-4833
Khanna, A. & Zinky, J. (1989). The revised arpanet routing , ACM SIGCOMM
metric Computer
Communication Review, Vol. 19, No. 4, pp. 45–56, ACM New York, NY, USA 1989,
ISSN 0146-4833.
Karnik, A. ; Iyer A. & Rosenberg C. (2008). Throughput-optimal  guration
con of wireless
networks, IEEE/ACM Transaction in Networking, vol 16, Issue 5, Oct. 2008, pp
1161-1174, Digital Object Identifier 10.1109/TNET.2007.909717.
Khasnabish, B. (2003), Implementing Voice over IP, 1st edition, ISBN-10: 0471216666, ISBN-
13: 978-0471216667, Wiley-Interscience publisher, 2003.
Koksal C. E. & Balakrishnan H. (2006). Quality-aware routing metrics for time-varying wirel
mesh networks , IEEE Journal onSelected Areas mmunication
of Co Special Issue on Multi-
HopWirelessMeshNetworks , vol. 24, no. 11, pp. 1984–1994, November 2006.
Kyasanur, P. & Vaidya, N. H. (2006). Routing and link-layer protocols for multi channel
multi-interface ad hoc wireless networks, ACM SIGMOBILE Mobile Computing and
CommunicationsReview , vol. 10, no. 1, pp. 31 –43, ISSN:1559-1662, Jan. 2006.
Lundgren, H. ; Nordstrom, E. & Tschudin, C. (2002). Coping with communication grey
zones in IEEE 802.11b based ad hoc networks. ProceedingsofIEEEIntlSymposiumon
a WorldofWireless, MobileandMultimedia NetworksWoWMoM) ( pp. 49–55,
, Atlanta,
Ga, USA, September 2002.
Mhatre, V. P. ; Lundgren, H. & Diot, C. (2007). Mac-aware routing in wireless mesh
networks. Proceedings of theFourth Annual Conference onWireless onDemand Net
Systems and Services WONS’0 ( ),7 pp. 46–49, Jan. 2007, ISBN: 1-4244-0860-1, Digital
Object Identifier 10.1109/WONS.2007.340461, 2007-04-02.
Mhatre V. & Rosenberg C. (2006). The impact of link layer model on the capacity of a
Next random ad hoc network, Proceedings of IEEE International Symposium on Information
Theory (ISIT), pp. 1688–1692, July 2006, Digital Object Identifier
10.1109/ISIT.2006.261642
Moy, J. (1998). OSPF version 2, network working group, RFC,238 The Internet Society, April
1998.
Park V. D. & Corson M. S. (1997). A highly adaptive distributed routing algorithm for
mobile wireless networks, Proceedings of IEEE Infocom , vol.3, pp.1405–1413, April
ks. 1997.
Perkins, C. E. & Bhagwa, P. (1994). Highly dynamic destination-sequenced distance-vector
routing (dsdv) for mobile computers,” ACM SIGCOMM Computer Communication
Review , pp. 234–244, 1994, ACM, New York, NY, USA, ISSN 0146-4833
Perkins C. E. & Belding-Royer E. (1999). Ad-hoc on-demand distance vector routing,
Proceedings of the nd 2 IEEE Workshop on Mobile Computing Systems and,Applicatio
pp. 90–100, New Orleans, LA, USA, Feb 1999.
Perkins C. E. & Belding-Royer E. (2003). Ad hoc On-demand Distance Vector (aodv) routing,
InternetRFC3561, The Internet Society, July 2003.
Perkins C. E. & Bhagwa, P. (1994). Highly dynamic destination-sequenced distance-vector
routing (dsdv) for mobile computers. Association for Computing Machinery’s Special
Interest GrouponDataCommunication’9 , Volume 24,
4 Issue 4 October 1994, ISBN: 0-
89791-686-7, pp. 234–244, ACM, New York, NY, USA
Proakis, J. (2001). Digital Communications, McGraw-Hill , Fourth edition, ISBN 0-07-232111-
3, New York.
Razandralambo, T. ; Guérin-Lassous, I. ; Iannone, L. & Fdida, S. (2008). Dynamic and

distributed packet aggregation to solve the performance anomaly in 802.11 wireless
networks. Elsevier Computer Networks journal, Special Issue: Wireless N
Performance , vol. 52, no. 1, pp. 77–95, 2008.
Shakkottai, S. ; Rappaport, T. S. & Karlsson, P. C. (2003). Cross-layer design for wireless
networks, IEEE Communications Magazine , Oct 2003, Vol. 41, o. 10, pp. 74–80, ISSN:
0163-6804.
Sivakumar R. ; Sinha P. & Bharghavan V. (1999). Cedar: a core-extraction distributed ad hoc
routing algorithm, CommunicationIEEEJournal , vol. 17, no. 8, Aug. 1999.
Woo A. (2004). A holistic approach to multihop routing in sensor networks, Ph.D.
dissertation,Universityof Berkeley , 2004
California,
Yang, Y. ; Wang, J. & Kravets, R. (2005). Interference-aware load balancing for multi-hop
wireless networks, Tech Report UIUCDCS-R-205-6, Department of Computer
Science, University of Illinois at Urbana-Champaign, 2005.
Yarvis, M. Conner, W. S. Krishnamurthy, L. Chhabra, J. Elliott, B. & MainWaring A. (2002).
Real-world experiences with an interactive ad hoc sensor network, Proceedings of
31st International Conference on ocessing
Parallel Workshops
Pr (ICPP 20 Workshops),
pp. 143–151, ISBN 0-7695-1680-7,20-23 August 2002, Vancouver, BC, Canada, IEEE
Computer Society, August 2002.
Zhao, S. ; Wu, Z. ; Acharya, A. & Raychaudhuri, D. (2005). PARMA: A Phy/Mac aware
routing metric for ad-hoc wireless networks with multi-rate radios. Proceedings of
IEEE International Sympos ium on a World of Wireless, Mobile and Multimedia
Networks(WoWMoM), pp. 286 – 292, ISBN: 0-7695-2342-0, June 2005.
Zhou, G. ; He, T. ; Krishnamurthy, S. & Stankovic, J. A. (2004). Impact of radio irregularity
on wireless sensor networks, Proceedings of the nd 2 international conference on M
systems, applications, and services (MobiSys), pp. 125–138, June 6-9, 2004, Hyatt
Harborside, Boston, Massachusetts, USA, USENIX, 2004.
Mobile WiMAX Performance Investigation 361
Mobile WiMAX Performance Investigation*

Alessandro Bazzi, Giacomo Leonardi, Gianni Pasolini and Oreste Andrisano
WiLab, IEIIT-BO/CNR, DEIS-University of Bologna
Italy
1. Introduction
The IEEE802.16-2004 Air Interface standard (IEEE Std 802.16-2004, 2004), which is the basis
of the WiMAX technology, is the most recent solution for the provision of fixed broadband
wireless services in a wide geographical scale and proved to be a real effective solution for the
establishment of wireless metropolitan area networks (WirelessMAN).
On February 2006, the IEEE802.16e-2005 amendment (IEEE Std 802.16e-2005, 2006) to the
IEEE802.16-2004 standard has been released, which introduced a number of features aimed
at supporting also users mobility, thus originating the so-called Mobile-WiMAX profile. Cur-
rently IEEE802.16 Task Group (TG) and WiMAX Forum are developing the next generation
Mobile-WiMAX that will be defined in the future IEEE802.16m standard (Ahmadi, 2009; Li
et al., 2009).
Although the Mobile-WiMAX technology is being deployed in the United States, Europe,
Japan, Korea, Taiwan and in the Mideast, there are still ongoing discussions about the poten-
tial of this technology. What is really remarkable, in fact, with regard to the Mobile-WiMAX
profile, is the high number of degrees of freedom that are left to manufacturers. The final deci-
sion on a lot of very basic and crucial aspects, such as, just to cite few of them, the bandwidth,
the frame duration, the duplexing scheme and the up/downlink traffic asymmetry, are left
to implementers. If follows that the performance of this technology is not clear yet, even to
network operators.
This consideration motivated our work, which is focused on the derivation of an analyti-
cal framework that, starting from system parameters and implementation choices, allows to
evaluate the performance level provided by this technology, carefully taking all aspects of
IEEE802.16e into account. In particular, the analysis starts from the choices to be made at the
physical layer, among those admitted by the specification, and "goes up" through the proto-
col pillar to finally express the application layer throughput and the number of supported
voice over IP (VoIP) users, carefully considering "along the way" all characteristics of the the
medium access control (MAC) layer, the resource allocation strategies, the overhead intro-
duced, the inherent inefficiencies, etc.
Let us remark that the analytical framework described in the following can be used not only as
a mean to gain an insight into the IEEE802.16e performance, but, above all, to drive the choices
of network operators in terms of system configuration. This is particularly true considering
that beside the model derivation, here we provide criteria, equations and algorithms to make
the best choices from the viewpoint of the system efficiency.
* Portions reprinted, with permission, from Proceedings of IEEE International Symposium on Personal,
Indoor and Mobile Radio Communications, 2007 (PIMRC 2007). ©2007 IEEE.
2. IEEE802.16 overview
Before starting our analysis let us introduce the most relevant characteristics of the IEEE802.16
technology, that are recalled hereafter.
The result of the IEEE802.16 TG/WiMAX Forum activity is a complete standard family
(IEEE Std 802.16-2004, 2004; IEEE Std 802.16e-2005, 2006) that specifies the air interface for
both fixed and mobile broadband wireless access systems, thus enabling the convergence of
mobile and fixed broadband networks through a common wide area broadband radio access
technology and a flexible network architecture.
The IEEE802.16 standard family supports four transmission schemes:
• WirelessMAN-SC, which has been mainly developed for back-hauling in line-of-sight
(LOS) conditions and operates in the 10 GHz - 66 GHz frequency range adopting a
single carrier modulation scheme;
• WirelessMAN-SCa, which has the same characteristics of WirelessMAN-SC but oper-
ates even in non-LOS conditions in frequency bands below 11 GHz;
• WirelessMAN-OFDM, which has been developed for fixed wireless access in non-LOS
conditions and adopts the orthogonal frequency division multiplexing (OFDM) modu-
lation scheme in frequency bands below 11 GHz;
• WirelessMAN-OFDMA, which has been conceived for mobile access and adopts the
orthogonal frequency division multiple access (OFDMA) scheme in the 2 GHz - 6 GHz
frequency range.
Since we are interested in the mobility enhancement provided by the IEEE802.16e amend-
ment, here we focus our attention on the WirelessMAN-OFDMA transmission scheme.
WirelessMAN-OFDMA is based on the OFDMA multiple-access/multiplexing technique
which is, on its turn, based on an NFFT subcarriers OFDM modulation scheme (Cimini, 1985;
Van Nee & Prasad, 2000) with NFFT equal to 128, 512, 1024 or 2048.
The NFFT subcarriers form an OFDM symbol and can be further divided into three main
groups:
• data subcarriers, used for data transmission;
• pilot subcarriers, used for estimation and synchronization purposes;
• null subcarriers, not used for transmission: guard subcarriers and DC subcarrier.
Considering sequences of OFDM symbols, it is easy to understand that transmission resources
are available both in the time domain, by means of groups of consecutive OFDM symbols,
and in the frequency domain, by means of groups of subcarriers (subchannels); it follows that
a given mobile station can be allocated one or more subchannels for a specified number of
symbols.
Several different schemes (in the following, permutation schemes) for subcarriers grouping
are provided by the specification, with different possibilities for the downlink and uplink
phases: among them we can cite DL-FUSC (downlink full usage of subchannels), DL-PUSC
(downlink partial usage of subchannels), DL-TUSC (downlink tile usage of subchannels), or
UL-PUSC (uplink partial usage of subchannels) (see the first column of table 1 for a complete
list).
The minimum OFDMA time-frequency resource that can be allocated is one OFDMA-slot,
which corresponds to 48 data subcarriers that can be accommodated in one, two or three
OFDMA symbols, depending on which kind of permutation scheme (DL-FUSC, DL-PUSC,
UL-PUSC, ...) is adopted; in particular:
Permutation scheme Available subchannels NCh NGS

NFFT NFFT NFFT NFFT
128 512 1024 2048
DOWNLINK
DL-FUSC 2 8 16 32 1
DL-PUSC 3 15 30 60 2
DL-OptFUSC 2 8 16 32 1
DL-TUSC1 4 17 35 70 3
DL-TUSC2 4 17 35 70 3
UPLINK
UL-PUSC 4 17 35 70 3
UL-OptPUSC 4 17 35 70 3
Table 1. Permutation schemes’ parameters.
• with DL-FUSC a subchannel is constituted by 48 subcarriers in each OFDM symbol

(hence, one OFDMA-slot covers one symbol);
• with DL-PUSC a subchannel is constituted by 24 subcarriers in each OFDM symbol
(hence, one OFDMA-slot covers two symbols);
• with UL-PUSC a subchannel is constituted by 12 subcarriers in the first OFDM symbol,
24 subcarriers in the second OFDM symbol, 12 subcarriers in the third OFDM symbol
and so on, according to the sequence 12-24-12-12-24-12....In this case one OFDMA-slot
covers three symbols.
Since seven fixed combinations of modulation scheme and coding rate Rc , hereafter denoted
as transmission modes, are provided by the IEEE802.16e physical layer, it follows that a single
OFDMA-slot allows to transmit differently sized payloads (see table B in figure 1).
Both time division duplex (TDD) and frequency division duplex (FDD) are supported. How-
ever, the initial release of Mobile-WiMAX certification profiles only includes TDD, since it
makes resource allocation more flexible (the downlink/uplink ratio can be easily adjusted to
support asymmetric DL/UL traffic); for this reason, here we only consider the TDD duplexing
scheme.
The TDD frame structure is depicted in the bottommost part of figure 1. Each TDD frame
is divided into downlink and uplink subframes, separated by transmit/receive and re-
ceive/transmit transition gaps (TTG and RTG).
Each subframe may include "multiple zones", which means that the permutation method can
be changed, thus moving, for instance, from DL-PUSC to DL-FUSC.
Focusing on the TDD frame, the first OFDM symbol of the DL subframe always carries a
preamble, while a number of subsequent OFDM symbols are necessarily allocated to accomo-
date MAC layer control messages (FCH, DL-MAP and UL-MAP) adopting the DL-PUSC per-
mutation scheme; similarly, a number of OFDM symbols are necessarily allocated in the UL
subframe to accomodate several common signalling channels (e.g. UL Ranging, UL CQICH,
UL ACK CH) adopting the UL-PUSC permutation scheme (see figure 1).
Finally, in order to correctly manage each data flow giving an acceptable quality of service to
the end user, IEEE802.16e-2005 provides five different scheduling services for traffic delivery:
Fig. 1. IEEE802.16e WirelessMAN-OFDMA data processing and parameters setting.
• Unsolicited Grant Service (UGS),

• Real-Time Polling Service (rtPS),
• Extended Real-Time Polling Service (ertPS),
• Non-Real-Time Polling Service (nrtPS),
• Best Effort (BE).
Each scheduling service is associated with a set of quality of services (QoS) parameters: (a)
maximum sustained rate, (b) minimum reserved rate, (c) maximum latency tolerance, (d) jitter
tolerance and (e) traffic priority. These are the basic inputs for the service scheduler placed
in the base station, which is aimed at fulfilling service specific QoS requirements. The main
differences among these services are on the uplink resource allocation; resource allocation
is, in fact, defined by the base station, which cannot have a perfect knowledge of all uplink
buffers in any instant. The interested reader can find a detailed description of the scheduling
services in (IEEE Std 802.16-2004, 2004) and (IEEE Std 802.16e-2005, 2006).
3. Transmission resources: OFDMA-slots

In this section the amount of resources that are available for data transmission is evaluated
as a function of all parameters that can be chosen by system implementers. In particular, the
OFDMA-slot, which is the minimum resource available at the physical layer for data allo-
cation, is focused and the amount of available OFDMA-slots is derived as a function of the
physical layer configuration.
In order to ease the reader’s task, the scheme reported in figure 2 summarizes the analytical
framework outlined in this section, which lead to the assessment of the amount of available
OFDMA-slots.
Fig. 2. Analytical framework for the derivation of the amount of OFDMA-slots per down-
link/uplink subframe.
3.1 OFDM symbol duration

In order to assess the amount of resources available for data transmission at the physical layer,
the OFDM symbol duration Ts must be obtained at first. Ts depends on the transmission
bandwidth BW (it is typically a multiple of 1.75 MHz or 1.25 MHz), the number NFFT of
OFDM subcarriers (equal to 128, 512, 1024 or 2048) and the normalized (to the useful symbol
duration) guard interval G (equal to 1/4, 1/8, 1/16 or 1/32).
Given BW, the value of an auxiliary parameter n (called sampling factor), introduced by the
specification, can be immediately derived: in particular, n = 28/25 if BW is a multiple of 1.25
MHz, 1.5 MHz, 2 MHz or 2.75 MHz, otherwise n = 8/7.
Once BW and n are known, we can derive the sampling frequency Fs , which is defined by the
specification as follows:
BW
Fs = n · · 8000, (1)
8000
having denoted with x the highest integer not greater than x.
Given the number NFFT of OFDM subcarriers, the subcarriers spacing ∆ f and the useful
OFDM symbol duration Tu can be immediately derived from the knowledge of Fs :
Fs 1
∆f = NFFT , Tu = ∆f . (2)
The guard time interval Tg and, finally, the OFDM symbol duration Ts follow:
Tg = G · Tu , Ts = Tu + Tg . (3)
3.2 Number of OFDM symbols per frame

Once TS has been obtained, the second step to derive the amount of OFDMA-slots available
at the physical layer is to assess the number of useful OFDM symbols in a frame.
Since we are interested in the TDD version of IEEE802.16e WirelessMAN-OFDMA, we have
to consider a frame structure consisting of two parts, that represent the downlink and uplink
subframes, separated by the TTG and RTG time intervals (see figure 1).
In order to derive the number of useful OFDM symbols in a frame, let us recall that the first
symbol of the frame is used to transmit the preamble (thus the number of preamble symbols is
NPr = 1) and that both TTG and RTG cannot be smaller than 5 µs (RTGmin = TTGmin = 5 µs).
Thus, once the frame duration TF has been chosen (possible values admitted by the specifi-
cation are 2, 2.5, 4, 5, 8, 10, 12.5 and 20 ms), and given the previously derived value of Ts , the
maximum number NSy MAX of OFDM symbols per frame (excluding the preamble) can be de-
rived:
MAX TF − TTGmin − RTGmin
NSy = − NPr , (4)
Ts
as depicted in figure 2.
3.3 Number of OFDMA-slots per frame

MAX , it is now possible to assess the number of OFDMA-slots that can be
Given the value of NSy
allocated in the downlink and uplink subframes. Since OFDMA-slots extend both in the time
and in the frequency domains, the derivation of their amount requires considerations on both
domains.
As for the time domain, let us recall that, depending on the adopted permutation scheme (DL-
FUSC, DL-PUSC, UL-PUSC, ...), an OFDMA-slot is spread over NGS = 1, 2 or 3 consecutive
OFDM symbols, as reported in the last column of table 1, whereas, as far as the frequency
domain is concerned, the amount of available subchannels NCh depends on the permutation
scheme and the number NFFT of OFDM subcarriers (as reported in the second column of table
1).
Please note that different permutation schemes are provided for the downlink and uplink
subframes and that more than one scheme can be used in a single subframe. Each permutation
scheme (denoted in the following with the superscript m) requires the allocation of a multiple
of an integer (1, 2 or 3, depending on the permutation scheme) number of OFDM symbols NSy m
m m
(NSydl in the downlink and NSyul in the uplink, respectively) and the sum NSy of uplink and
downlink symbols allocated for each permutation scheme is bounded by the above assessed
MAX :
NSy
Mdl −1 Mul −1
m m MAX
∑ NSy dl
+ ∑ NSy ul
= NSy ≤ NSy , (5)
m =0 m =0
where Mdl (Mul ) is the amount of permutation schemes adopted in the downlink (uplink)
subframe.
As recalled in section 2, at least two OFDM symbols are allocated with DL-PUSC in the down-
link subframe, in order to carry frame management messages, while three OFDM symbols are
allocated with UL-PUSC in the uplink subframe, in order to carry signalling common chan-
nels; here we assume that the entire first two OFDM symbols in downlink (with DL-PUSC)
and the entire first three OFDM symbols in uplink (with UL-PUSC) are used for this scope,
denoting the related overhead with NOSydl = 2 and NOSyul = 3.
Moreover, the rest of the sub-frame is supposed to be transmitted adopting only one permu-
tation scheme. Thus, the superscript correspondent to the adopted permutation scheme will
be omitted in the following and (5) is rearranged as follows:
MAX
NOSydl + NSydl + NOSyul + NSyul = NSy ≤ NSy , (6)
where NSydl and NSyul now represent the amount of downlink/uplink symbols available for
user data.
Let us observe that the choice of NSydl and NSyul is not only constrained to fulfill (6), but is
also a consequence of the desired asymmetry between the downlink and uplink phases of the
TDD frame, hereafter referred to as “desired asymmetry factor” and denoted as AFin .
Let’s keep in mind, in this regard, that the minimum resource that can be allocated is the
OFDMA-slot and that the number of slots in a subframe is related not only to the number of
OFDM symbols within the frame, but also to the adopted permutation scheme; as an exam-
ple, with NFFT = 2048 subcarriers two OFDM symbols adopting the DL-PUSC permutation
scheme carry 60 slots while three OFDM symbols adopting UL-PUSC carry 70 slots (refer to
(8) and table 1). Thus, defining AFSl as the asymmetry factor in terms of ratio between the
amounts of downlink and uplink slots:
NSldl
AFSl = , (7)
NSlul
and deriving the amount of OFDMA-slots available for data transmission in the down-
link/uplink subframe through the equation (where dl/ul denotes downlink or uplink as al-
ternatives):
NSydl/ul
NSldl/ul = NChdl/ul · , (8)
NGSdl/ul
the desired asymmetry AFin can be approached finding the values NSydl and NSyul that make
AFSl as near as possible to AFin . In general, a perfect matching between AFin and AFSl will be
not possible, due to the system constrains.
The detection of NSydl and NSyul in such a way to minimize the resource wasting (that
is, OFDM symbols within the frame that are unused because unable to accomodate entire
OFDMA-slots) for a given AFin can be carried out by means of the algorithm provided in
appendix I.
Equation (8) represent the final outcome of this section since, jointly with the constraints given
by the desired asymmetry factor and system choices (bandwidth, guard interval, number of
subcarriers, frame duration, ....) accounted for by the previous equations, allows to derive the
amount of OFDMA-slots available in each subframe for data transmissions.
Please refer to figure 2 for a pictorial representation of the whole methodology described in
this section.
4. From application layer packets to subcarriers allocation

In the previous section the amount NSldl and NSlul of OFDMA-slots available for data alloca-
tion have been derived, taking into account all the physical and MAC layers parameters. The
next step is to understand how packets to be transmitted, coming from the higher protocol
layers, are mapped onto these resources. A brief overview on the packet processing is given
hereafter, followed by an analytical evaluation of the number of OFDMA-slots that are finally
needed to allocate each packet.
4.1 Packet processing overview

The data mapping process, starting from the application layer data unit down to the physical
layer, is illustrated step by step hereafter. Please refer to figure 1, where each step is depicted,
to better understand the whole process.
Let us denote as ASDU the application level data fragment of S AS bytes that is allocated into
the payload of a TCP/IP packet. Each ASDU is firstly added with O HL bytes, where O HL
represents the overhead added from the application to the network layer, and then mapped,
at the MAC layer, onto a MAC service data unit (MSDU) of S MS bytes. Each MSDU is then
partitioned into fragments of S Fmax bytes, whose value is negotiated during the connection
setup phase; obviously the last fragment of each MSDU may be smaller (S Flast bytes). If the
ARQ mechanism is active fragments are also called ARQ blocks.
One or more fragments are then allocated into a MAC protocol data unit (MPDU), with some
overhead: in particular, a MAC header will be added plus either (a) one fragmentation sub-
header if all fragments are contiguous and related to the same MSDU or (b) a packetization
subheader per each group of contiguous fragments belonging to the same MSDU; a CRC
(cyclic redundancy check) tail of 32 bits will be added at the end of the MPDU, in order to
check its integrity at the receiver side.
At the physical layer, MPDUs are partitioned into groups of bytes that are subject to the for-
ward error correction coding process, giving birth to a certain number of codewords. One
or more OFDMA-slots can be combined in order to convey each codeword. Adjacent slots,
both in the time and subchannels domain, are grouped into OFDMA data regions, which are
two-dimensional (squared o rectangular) allocations of a group of contiguous subchannels in
a group of contiguous OFDM symbols (see figure 1).
4.2 From application layer data to MPDUs

After the data processing overview provided above, in this subsection we analytically de-
rive the amount of OFDMA-slots needed to deliver an ASDU. Having derived (section 3)
the amount of OFDMA-slots available in the uplink/downlink subframes of a TDD Mobile-
WiMAX system, this is the second step along the path that leads to the Mobile-WiMAX per-
formance assessment.
Fig. 3. Diagram of the calculation from the packet size to the number of OFDMA-slots that are
needed to accomodate it.
Also in this case, in order to help the reader, the analytical framework outlined in the following
has been summarized in a pictorial fashion, reported in figure 3.
Let us consider the aforementioned ASDUs of S AS bytes; as represented in figure 1 they even-
tually arrive at the MAC layer with the addition of the higher layers overheads of O HL bytes,
thus originating MAC layer service data units (MSDUs) of S MS bytes:
S MS = S AS + O HL . (9)
The MSDUs are then fragmented into a fixed number NF← MS of fragments, each of them of
size S Fmax except, in case, the last one. This one has a size of S Flast ≤ S Fmax bytes; thus:

S MS
NF← MS = , (10)
S Fmax
S Flast = S MS − [( NF← MS − 1) · S Fmax ] (11)
where x indicates the lowest integer not less than x.
It follows that each MSDU is carried at the MAC layer by:
• (NF← MS − 1) fragments of size S Fmax ,
• 1 fragment of size S Flast .
Of course, if S MS is a multiple of S Fmax , then S Flast =S Fmax .
Let us assume now, for the sake of simplicity, that no packetization is performed at the MAC
layer; each fragment is therefore mapped onto one MPDU with the addition of the MAC
layer overhead of O M bytes. It follows that the number NMP← MS of MPDUs needed to carry
a single MSDU is equal to NF← MS . Since all but (in case) the last MPDU have the same size,
we have:
• (NMP← MS − 1) MPDUs of size S MPmax ,

• 1 MPDU of size S MPlast ,
where:
NMP← MS = NF← MS ,
S MPmax = S Fmax + O M , (12)
S MPlast = S Flast + O M .
Of course, if NMP← MS = 1, each MPDU carries a complete MSDU and its size is S MPlast .
4.3 MPDUs into OFDMA-slots

Starting from the results obtained in subsection 4.2, we can now derive the number of
OFDMA-slots needed to carry any MPDU and, as a consequence, any MSDU.
( j)
Let us recall that every transmission mode j can convey a different amount SSl of data bytes
into a single OFDMA-slot (see table B in figure 1); since there are two possible sizes for
MPDUs (S MPmax and S MPlast ), we can derive, for every transmission mode j, the minimum
number of slots needed to carry each of them:

( j) S MPmax
NSl ← MPmax = ( j) ,
SSl (13)
( j)
NSl ← MPlast = S MPlast
( j) .
SSl
These equations show that when the MPDU size is not a multiple of the amount of bytes
carried by a single OFDMA-slot, some padding bits have to be added in order to fill the last
slot, wasting some resources.
Thus, a single MSDU is transmitted with the generic transmission mode j through:
• ( NMP← MS − 1) MPDUs of size S MPmax , accommodated into
( j)
( NMP← MS − 1) · NSl ← MPmax slots,
( j)
• 1 MPDU of size S MPlast , accommodated into NSl ← MPlast slots.
( j)
Assuming no resource wastage due to data regions’ allocations 1 , the number NSl ← MS of
OFDMA-slots needed to carry a complete MSDU adopting transmission mode j is given by:
( j) ( j) ( j)
NSl ← MS = (( NMP← MS − 1) · NSl ← MPmax ) + NSl ← MPlast . (14)
Recalling that the scope of the analysis reported in this section was to derive the amount of
OFDMA-slots needed to accomodate an ASDU, we can state that (14) is the final outcome of
this section since, jointly with the equations reported in subsection 4.2, it achieves our end.
5. System performance: throughput of a TCP connection

Having derived the resources (that is, OFDMA-slots) available for data allocation at the phys-
ical layer (in section 3) and the amount of resources needed to carry each ASDU (in section 4),
we can now assess the performance level provided by IEEE802.16e for a given configuration.
1 Data regions must be squared or rectangular.

5.1 Numerical results

In this section some numerical results obtained through (19) are given. ASDUs of S AS = 1460
bytes were chosen, since this is the payload size of a typical TCP/IP packet. Considering
20 bytes for the IP overhead, 20 bytes for the TCP overhead and neglecting the overhead
introduced by the upper layers, we assumed that each MSDU has a size of S MS = S AS +
O HL = 1500 bytes.
A further overhead of O M = 10 bytes is introduced by the MAC layer, following the assump-
tion of no packetization.
The OFDM modulation parameters were set to NFFT = 2048, BW = 7 MHz, G = 1/32;
moreover a frame duration of TF = 10 ms has been chosen and RTG = TTG = 116 µs were
considered; all other physical layer parameters are consequently derived (e.g. TS = 264 µs).
The impact of the remaining parameters affecting the throughput will be investigated in the
following. In particular, different values of S Fmax and all transmission modes and permuta-
tions schemes will be considered.
In figure 4, a comparison between physical layer and application layer throughput is given
varying the fragments maximum size S Fmax . The physical layer throughput Thr P has been
evaluated considering the total amount of bits carried over the medium by all available
OFDMA-slots, as follows:
NSl · 8 · SSl [bytes]

Thr P [bit/s] = , (20)
TF [s]
where the same notation introduced in section 4.2 has been adopted (please note that the pre-
vious equation considers only those resources available for data transmission, thus excluding
the preamble symbol and the subcarriers used for signalling and control messages).
12
10
ThrA
8
Throughput [Mbps]
ThrP
0
0 500 1000 1500
S [Byte]
Fmax
Fig. 4. Comparison between physical layer and application layer throughput (Thr P and Thr A )
varying the fragment (ARQ block) maximum size S Fmax . Transmission modes 0 and 6.
DL-PUSC has been considered in the downlink and UL-PUSC in the uplink, with AFin =
1 and AFSl = 1.37 (following the equations reported in section 3.3, we obtain NSydl = 16,
NSyul = 15, NSldl = 480 and NSlul = 350). Transmission modes 0 and 6 are considered.
The comparison between the dashed and solid curves highlights the reduction of throughput
due to both the allocation procedure of ASDUs and the overhead. As can be noted, small
variations in the choice of S Fmax may affect the system performance, due to the slot granularity
in the physical resource allocation and the impossibility to further divide a fragment (i.e., an
ARQ block).
These curves also show that a too small value of S Fmax should not be chosen (this is mainly a
consequence of the presence of the overheads O HL and O M ). However, although considering
error prone transmissions is out of the scope of the present work, it should be clear that a large
value of S Fmax should be avoided too, since each ARQ block must be entirely retransmitted if
not correctly received.
Figure 5 deepens the previous results focusing the attention on the application layer through-
put and considering all transmission modes.
A direct comparison of the throughput perceived adopting the different transmission modes
as a function of S Fmax shows that the choice of S Fmax is a tricky task, since there is no optimal
value providing the maximum throughput for all transmission modes.
As can be observed, a number of choices for S Fmax are highlighted through vertical lines and
the correspondent throughput values with circles. These values are somehow suboptimal and
have been chosen according to the following steps:
1. for each value of S Fmax in the range [1, 500 bytes] the throughput values achieved by
each transmission mode normalized to the peak value for that mode were summed into
SU MThr (S Fmax );
2. the values of S Fmax that brought to relative maximum of SUMThr (S Fmax ) were found,
neglecting those values of S Fmax that do not give an absolute value of SUMThr (S Fmax )
higher than the previous one.
The values of S Fmax derived as previously described (S Fmax = 98, 125, 134, 152, 206, 254, 422
bytes) allow to reduce resource wasting when a single fragment (ARQ block) is transmitted in
a single MPDU.
In figure 6 the value of AFSl is compared to AFApp , which is defined as the ratio between the
maximum application layer throughput in downlink and the one in uplink. The DL-PUSC
and UL-PUSC permutation schemes have been considered in the downlink and in the uplink,
respectively; AFin = 1, AFin = 2 and AFin = 3 have been assumed as desired asymmetry
factors. Transmission mode 6 only.
As can be noted, a good match between AFSl and AFApp is achieved for all the considered
AFin (avoiding to consider too large values for S Fmax ). On the contrary, it is quite hard to
exactly respect the desired AFin with no wasting: note, in fact, that the cases AFin = 1 and
AFin = 2 bring to the same result (that is, the need to minimize the resource wasting brings,
in both cases, to the same choice of NSydl and NSyul ).
In figure 7 the throughput is shown as a function of the number of OFDM symbols available
for data transmission in the downlink subframe for all possible permutation schemes (refer
to table 1). In this case, NSydl is set and the correspondent AFSl follows as a consequence (see
section 3.3). Transmission mode 6 and S Fmax = 206bytes have been considered. This figure
also highlights that an increase (reduction) in NSydl has an effect only if it involves at least NGS
symbols.
10
9 Mode 6
8
Mode 5
7
Throughput [Mbps]
6
Modes 3,4
5
4
Mode 2
3
Mode 1
2
Mode 0
1
0
0 100 200 300 400 500
SFmax [byte]
Fig. 5. Application layer throughput (Thr A ) varying the fragments (ARQ blocks) maximum
size S Fmax for all transmission modes. The values of S Fmax that allow to have a good occupa-
tion adopting any possible transmission mode are marked with vertical lines and small circles
(o); they correspond to S Fmax = 98, 125, 134, 152, 206, 254, 422 bytes
4
AF =3
in
3.5
2.5
Asymmetry Factor
AF =1,2
in
1.5
AF
App
0.5 AFSl
AF
App
AFSl
0
0 500 1000 1500
S [byte]
Fmax
Fig. 6. Comparison between AFSl and AFApp , given AFin = 1, AFin = 2 and AFin = 3.
Transmission mode 6.
6. System performance: VoIP capacity on UGS or ertPS

In this section the maximum number of users performing a VoIP call that can be served by
IEEE802.16e is evaluated, following the analysis described in sections 3 and 4.
20
DL−FUSC/DL−OptFUSC
18 DL−PUSC
DL−TUSC1/DL−TUSC2
16 UL−PUSC/UL−OptPUSC
14
Throughput [Mbps]
12
10
0
0 5 10 15 20 25 30
NSy
dl
Fig. 7. Comparison of application layer throughput (Thr A ) adopting the various permuta-
tion schemes, varying the number of downlink useful symbols NSydl . Transmission mode 6.
S Fmax = 206 bytes.
A description of the considered VoIP codecs and scheduling services is given before entering
into the details of the analytical model.
6.1 UGS and ertPS scheduling services

As already mentioned in section 2, five scheduling services are provided by the IEEE802.16e
specification for traffic delivery. Since our attention is now focused on real-time VoIP traffic,
UGS and ertPS are the only possible choices, due to latency constraints, and are therefore
considered in the following:
• Unsolicited grant service (UGS) is designed to support real-time uplink service flows
that generate transport fixed-size data packets on a periodic basis, such as T1/E1 and
VoIP without silence suppression. The service offers fixed size grants on a real-time
periodic basis, which eliminate the overhead and latency of user’s requests and assure
that grants are available to meet the flowŠs real-time needs.
• Extended real-time polling service (ertPS) (Lee et al., 2006) improves UGS when the
application layer rate varies in time. The base station (BS) shall provide unicast grants
in an unsolicited manner like in UGS, thus saving the latency of a bandwidth request.
However, whereas UGS allocations are fixed in size, ertPS allocations are dynamic. The
BS may provide periodic uplink allocations that may be used for requesting the band-
width as well as for data transfer. By default, size of allocations corresponds to current
value of maximum sustained traffic Rate at the connection. Users may request chang-
ing the size of the uplink allocation by either using an extended piggyback request field
of the grant management subheader or using BR field of the MAC signaling headers,
or sending a specific codeword over the signalling channel CQICH. The BS shall not
change the size of uplink allocations until receiving another bandwidth change request
from the user.
6.2 VoIP codecs

The most important and mainly adopted voice codecs have been considered:
1. ITU G.711 (ITU-T Rec. G.711, 1988), the well known constant bit rate PCM at 64 kbps;
this is the codec used in PSTN networks, with no compression, neither during the
speech nor during silences of a conversation;
2. ITU G.729 (ITU-T Rec. G.729, 1996a), the most used codec for VoIP, at 8 kbps; when an
active speech period is detected, it produces one packet of 80 bits every 10 ms, but more
than one packet may be concatenated in order to reduce protocols overheads (Goode,
2002); obviously, this process enlarges the average delivery delay of packets. Hereafter,
we will consider the concatenation of a couple of packets (Goode, 2002) and we will
denote this codec as G.729. 160 bits packets are thus generated every 20 ms.
3. AMR (adaptive multi rate (3GPP TS 26.071, 2008)), standardized by 3GPP and used for
voice in second and third generation cellular radio access; it generates one packet every
20 ms with a variable data rate, going from a minimum of 4.75 kbps (95 data bits plus
18 overhead bits for each packet) to a maximum of 12.2 kbps (244 data bits plus 18 over-
head bits for each packet). The variation of the data rate is given in order to better select
the appropriate tradeoff between resource usage and speech quality (obviously, a data
rate reduction leads to a quality degradation). In order to investigate the capacity of
the IEEE802.16e WirelessMAN-OFDMA system with the AMR codec, we will consider
both the minimum data rate (4.75 kbps, denoted as AMR4.75) and the maximum data
rate (12.2 kbps, denoted as AMR12.2).
ITU G.729 and AMR codecs have been designed, in particular, with the specific goal to reduce
the resources occupation: when no voice activity is detected the silence suppression procedure
is activated. As a consequence small and less frequent packets are transmitted, which convey
the information for a “comfortable noise” generation at the receiver side.
In particular, adopting the AMR codec, the detection of a silence period (3GPP TS 26.092, 2008)
gives rise to the following steps:
• eight full voice packets are normally transmitted in the first interval, called hangover
period;
• then, a SID (silence insert descriptor) packet (36 data bits plus 18 overhead bits) is trans-
mitted after 60 ms;
• further SID packets are transmitted every 160 ms until a new speech activity is detected.
This process is depicted in the topmost part of figure 8.
As far as the G.729 codec is concerned, the silence suppression procedure is defined in the
Annex B (Benyassine et al., 1997; ITU-T Rec. G.729, 1996b). In this case, each SID packet has a
length of 15 bits. However, the time interval between two successive SID packets is not fixed:
in fact, for a good quality at the receiver, a lower or a greater transmission rate may be needed
depending on the specific background noise observed in each environment.
In order to allow the derivation of meaningful results, SID packets are here supposed to be
generated with a fixed rate during silences, as with AMR. In particular, the rate of G.729 SID
packets is assumed to be twice the one adopted by AMR, following the results provided in
(Estepa et al., 2005).
In the following, the activity factor ν is defined as the ratio between the time during which
full packets are generated and the total duration of the conversation. Thus, in particular, we
Codec B ρ BSID ρSID ν Maximum number of

allowed users
[bits] [pack/s] [bits] [pack/s] Mode 0 Mode 6
G.711 1280 50 - - 1 20 / __ 87 / __
G.729 160 50 - - 1 58 / __ 233 / __
G.729ss 160 50 15 12.5 0.45 58 / 90 233 / 482
AMR4.75 113 50 - - 1 63 / __ 233 / __
AMR4.75ss 113 50 54 6.25 0.45 63 / 111 233 / 506
AMR12.2 262 50 - - 1 50 / __ 175 / __
AMR12.2ss 262 50 54 6.25 0.45 50 / 90 175 / 374
Table 2. Codec parameters and analytical calculation of maximum number of allowed users.
Multiple values refer to UGS/ertPS.
will adopt ν = 1 when no silence suppression is activated and ν < 1 in the case of silence
suppression.
In the latter case, in order to derive a realistic value of ν, we simulated the dynamic of a con-
versation according to a detailed model that takes into account also periods of simultaneous
talks or silences of the two parties, and the short gaps through the speeches (Stern et al., 1996).
This model gives a 33% of voice activity over the total conversation duration for each of the
two parties. It must be noted, however, that since 8 full packets are still transmitted dur-
ing the hangover period, the activity factor ν is approximately 0.45; please note that this value
corresponds to both our simulations and the experimental results given in (Estepa et al., 2005).
As for the less sophisticated G.711 codec, here it was considered only as a reference, and no
silence suppression is introduced.
The parameters of all considered codecs, with and without silence suppression, are given in
the first four columns of table 2. In particular: the first column indicates the considered codecs
(the subscript ss indicates that silence suppression is considered); the second column defines
the number of bits B of full packets and the number of full packets per second ρ that are
transmitted when the voice is detected; similarly, the third column defines the number of bits
BSID of SID packets and the number ρSID of SID packets per second that are transmitted when
silences are detected; finally, the fourth column represents the activity factor ν; the rest of the
table will be illustrated in the following.
6.3 Amount of supported VoIP users

In this section, the number of VoIP users that can be served will be evaluated as a function of
the adopted codec x, the scheduling service k and the transmission mode j (all users are sup-
posed to be served adopting the same mode). Also in this case we need to consider the whole
packet processing from the application layer down to the transmission over the medium.
Before being transmitted, each packet generated by the codec must pass through the whole
protocol pillar, thus increasing its size owing to the overheads introduced by each protocol
layer. In particular, the RTP, UDP, IP and MAC layers overheads are added, which are, re-
spectively, ORTP = 12 bytes, OUDP = 8 bytes, O IP = 20 bytes and O M = 10 bytes (thus,
O HL = ORTP + OUDP + O IP = 40 bytes). Assuming that no fragmentation is carried out from
the application to the MAC layer, the size (in bytes) of full packets and SID packets generated
by the codec x is given by:
(x)
S MS = B( x) /8 + O HL + O M , (21)
Fig. 8. Topmost part: AMR codec full packets and SID packets generation, with ρ = 50pack/s.
Rest of the figure: buffer state and resource allocation of a generic uplink voice traffic flow
with UGS and ertPS, following the packets generation depicted in the topmost part.
(x) (x)
S MSSID = BSID /8 + O HL + O M . (22)
Of course, the ARQ mechanism is assumed inactive, since we are considering a real time
service.
Let us recall (from section 6.2), now, that for a given codec x:
Time with f ull packets

ν( x) = (23)
Total conversation duration
is the activity factor, while
Number o f f ull packets
ρ( x) = (24)
Time with f ull packets
is the rate of transmission of full packets (of B( x) bits) during voice activity and
(x) Number o f SID packets

ρSID = (25)
Time without f ull packets
(x)
is the rate of transmission of SID packets (of BSID bits) during silences, if silence suppression
is considered.
As already recalled each packet to be transmitted is mapped onto OFDMA-slots, thus, in
order to assess the maximum number of users that can be served, the number of slots that
( j)
are needed for each packet must be firstly calculated, also considering that the number SSl of
bytes carried by one slot depends on the adopted mode j (see table B of figure 1).
( x,j)
In particular, for a given codec x and a given transmission mode j, the number NSl ← MS of
( x,j)
slots needed to carry a full packet and the number NSl ← MS of slots needed to carry a SID
SID
packet are:  
(x)
( x,j) S MS
NSl ← MS = 
 ( j)  , (26)
 SSl 
 
(x)
( x,j) S MSSID
NSl ← MS
SID
=
 ( j)  .
 (27)
 SSl 
Depending on the considered scheduling service k and the specific VoIP codec x, the average
number of slots required in a given (DL/UL) subframe by a single user adopting mode j, is
given by:

( x,k,j) ( x,j) ( x,j)
MSl ←U = f ( x,k) NSl ← MS , NSl ← MS , (28)
SID
where the analytical expression of f ( x,k) (·)

will be provided in the following for both the con-
sidered scheduling services.
As a general consideration, please note that all difficulties in resource allocation are on the
uplink, since the base station has a perfect knowledge of all buffers in the downlink and no
resource requests are needed. For this reason, and assuming that AFsl ≥ 1 (that is, we have
more downlink slots than uplink slots), all evaluations will be done focusing on the uplink di-
rection. Thus, given an amount NSlul of available slots (in the uplink subframe), the maximum
number NUMAX of users can finally be evaluated:
 
 
( x,k,j)  NSlul 
NUMAX =  . (29)
( x,k,j)
MSl ←U
Performance on UGS. Since UGS resources are statically allocated, as negotiated during con-
nection setup, the silence suppression procedure does not provide any benefit in the uplink
direction.
Denoting with TF the frame duration, (28) becomes:

( x,UGS,j) ( x,j)
MSl ←U = TF ρ( x) · NSl ← MS , (30)
Please note that there is no dependence on the activity factor (some resources will be wasted
if ν( x) < 1).
( x,UGS,j)
Combining (30) and (29) the maximum number NUMAX of VoIP users supported by the UGS
scheduling service can be easily derived.
Performance on ertPS. In the case of ertPS scheduling service, besides the adopted codec x
and transmission mode j, also the activity factor ν( x) must be considered in order to derive
the amount of supported VoIP users. After the hangover period, in fact, the transmission rate
is reduced and a single slot is allocated (NSl ← MH = 1) in order to allow the transmission of
a stand-alone MAC signaling header for a quick modification of the resources request (see
figure 8). Please note, by the way, that, although reduced, this allocation entails a resource
wasting, since no transmission is performed in the most of cases. Resource wasting will occur
also at any rate reduction (refer to figure 8): before a rate decreases, in fact, the request is sent
in the uplink using an oversized resource.
Let us observe, furthermore, that following a rate increase request, the first (larger) resource is
allocated in the subsequent frame, without respecting the normal rate of allocation, in order
to reduce the latency.
( x,ertPS,j)
In the case of ertPS scheduling service, therefore, in order to calculate MSl ←U we must take
into account:
• the slots needed for full packets, that are generated with rate ρ( x) during active voice pe-
( x,j)
riods: R · NSl ← MS average slots per second, where R = ν( x) · ρ( x) indicates the average
number of full packets per second;
(x)
• the slots needed for SID packets, that are generated with rate ρSID during silent periods:
( x,j) (x)
RSID · NSl ← MS average slots per second, where RSID = (1 − ν( x) ) · ρSID indicates the
SID
average number of SID packets per second;
• the slots needed for stand-alone MAC headers, allocated with rate ρ( x) when neither full
packets nor SID packets are generated: R MH · NSl ← MH average slots per second, where
R MH = 1 − R − RSID ; indicates the average number of slots left for MAC headers per
second, transmitted at the same rate;
• the slots that are wasted during variations from full packets allocation to stand-alone
( x,j)
MAC headers allocations: RCFG · NSl ← MS average slots per second, where RCFG =
1
av + T av
TCFG indicates the average number of uninterrupted periods of full packets per
CSS
second, that depends on the average duration of a period of continuous full packets
av ) and on the average duration of a period of silence sup-
generation by the codec (TCFG
av );
pression with SID packets generation (TCSS
• the slots that are wasted during variations from SID packets to stand-alone MAC head-
( x,j)
ers: RSID · NSl ← MS average slots per second.
SID
( x,ertPS,j)
Thus, in this case, MSl ←U is given by:

( x,ertPS,j) ( x,j) ( x,j)
MSl ←U = TF ( R + RCFG ) · NSl ← MS + 2 · RSID · NSl ← MS + R MH · NSl ← MH , (31)
SID
Concerning the parameters TCFG av and T av , please note that they are not only related to the
CSS
adopted codec, but also to the characteristics of the specific conversation, also including, as an
example, the language; in order to derive meaningful results, hereafter we adopted the values
av
reported in (Stern et al., 1996) (TCFG = 0.17s and TCSSav = 0.3428s), although they are related
to the voice-silence intervals rather than to codec full generation-codec silence suppression
(they do not consider the hangover periods); for this reason, a slight underestimation of the
maximum number of VoIP users is expected in the numerical results.
( x,ertPS,j)
Finally, combining (31) and (29), the maximum number NUMAX of VoIP users supported by
the ertPS scheduling service can be easily derived.
6.4 Numerical results

For the numerical results derivation, the amount of OFDMA-slots available for data transmis-
sion in the downlink subframe has been assumed equal to NSlul = 350; this value is a con-
sequence of the same assumption reported in section 5.1: BW = 7 MHz, NFFT =2048 OFDM
subcarriers, normalized OFDM guard interval G = 1/32, frame duration TF = 10 ms (which
is the most suited value for VoIP traffic allocation); 15 of the 31 OFDM symbols of each frame
UGS
600
G.711
550 G.729
AMR4.75
500
AMR12.2
450
Maximum number of users

400
350
300
250
200
150
100
50
0
0 1 2 3 4 5 6
Mode
Fig. 9. Maximum number of users using UGS scheduling service. All modes. All VoIP codecs
without silence suppression.
ertPS
600
G.729SS
AMR4.75SS
500 AMR12.2SS
Maximum number of users
400
300
200
100
0
0 1 2 3 4 5 6
Mode
Fig. 10. Maximum number of users using ertPS scheduling services. All modes. All VoIP
codecs with silence suppression.
are left for uplink data, adopting UL-PUSC (the rest of the symbols are used for the uplink
common channels and the downlink subframe).
In the last column of table 2 the maximum number of VoIP users that can be served with
UGS and ertPS (separated by the symbol “/”) are reported for each codec; modes 0 and 6 are
considered. Two underscores are typed when the adoption of that scheduling service makes
no sense with that codec (i.e., ertPS with no silence suppression).
The maximum amount of VoIP users that can be supported is shown for all modes in figures 9
and 10 focusing on UGS and ertPS, respectively. Obviously, the VoIP capacity adopting UGS
with and without silence suppression is always the same; for this reason the results related to
the case of silence suppression are not reported in figure 9.
It can be observed that, as expected, VoIP capacity is strictly related to the average data rate
generated by the codec.
Comparing figures 9 and 10 we can also appreciate the significant benefit provided by the
adoption of the ertPS instead of UGS.
7. Conclusions
The performance of IEEE8021.16e WirelessMAN-OFDMA depends on a large number of sys-
tem parameters and implementation choices, such as, among the others, the available band-
width, the frame duration, the uplink/downlink traffic asymmetry and the fragmentation
policy.
In the previous sections we provided an analytical framework that allows to evaluate the
throughput achievable with TCP/IP connections as well as the amount of supported VoIP
users as a function of the most significant parameters characterizing this technology. Beside
the analytical model derivation, here we provided criteria, equations and algorithms to make
the best choices from the viewpoint of system efficiency.
Furthermore, some numerical results were given, showing the impact of some specific pa-
rameters on the system performance. With reference to TCP/IP connections, for instance, the
troublesome choice of the maximum size of ARQ blocks has been discussed as well as the
potential resource wastage entailed by a wrong choice of the asymmetry factor between the
downlink and uplink subframes. The main outcome of this analysis is given by a set of criteria
to be followed in order to maximize the throughput provided to the final user.
As far as VoIP connections are concerned, here we assessed the maximum number of users
that can be supported, carefully considering the voice codec characteristics and the adopted
scheduling service. The outcomes of this investigation provide an indication about the capac-
ity of this technology to be alternative to other technologies such as UMTS and LTE for the
provision of the voice service.
As a final remark, let us observe that the analytical framework proposed in this chapter pro-
vides a tool to evaluate the upper limits of the throughput and the maximum amount of VoIP
users that can be supported by IEEE802.16e WirelessMAN-OFDMA. However, in order to in-
vestigate the actual performance of such a complex technology in a given scenario considering
the degradation due, for instance, to fading, shadowing, noise and interference, the only feasi-
ble way is to adopt a simulation tool able to carefully reproduce all aspects of communications,
with particular reference to the physical layer behavior.
This kind of investigation has been carried out at WiLab (Italy) by means of the simulation
platform SHINE, that has been developed in the last years to assess the performance of wire-
less networks in realistic scenarios. The interested reader may refer to (Andrisano et al., 2007;
2009; Bazzi et al., 2006).
Appendix I
Here we illustrate the algorithm that allows to maximize symbols usage starting from the
number of useful symbols NUSy = NSy − NOSydl − NOSyul .
The following steps must be followed:
1. derive J = Resul = mod( NUSy , NGSul );
2. find, if possible:
a = min
NUSy
so that mod( J + a · NGSul , NGSdl ) = 0. (32)
a∈ 0, N
GSul
3. if a was not found, then reduce J by one and return to step 2, else exit.
The obtained value a and the parameter J allow the derivation of the minimum value for NSydl
MI N ) that maximally reduces the symbols wasting:
(NSy dl
MI N
NSy dl
= J + a · NGSul , (33)
All possible solutions will be:

OPT MI N
NSy dl
(b) = NSy dl
+ b · mcm( NGSul , NGSdl ), (34)
MI N

NUSy − NSy
where b ∈ 0, NGSdl
dl
and mcm( x, y) is the minimum common multiplier of x and y.
It follows:
OPT OPT
NSy ul
(b) = NUSy − ( Resul − J ) − NSy dl
( b ). (35)
and
OPT ( b )
NSl
AFSl (b) = dl
OPT ( b )
. (36)
NSl
ul
Finally, we can choose the value of b that brings to the AFSl (b) nearer to AFin .
8. References
3GPP TS 26.071 (2008). Mandatory speech codec speech processing functions; amr speech
codec; general description.
3GPP TS 26.092 (2008). Mandatory speech codec speech processing functions; adaptive multi-
rate (amr) speech codec; comfort noise aspects.
Ahmadi, S. (2009). An overview of next-generation mobile wimax technology, IEEE Commu-
nications Magazine Vol.47(n.6): pp.84–98.
Andrisano, O., Bazzi, A., Leonardi, G. & Pasolini, G. (2007). Ieee802.16e best effort perfor-
mance investigation, Proceedings of IEEE International Conference on Communications,
2007 (ICC 2007), IEEE, Glasgow, Scotland, pp. 4837–4842.
Andrisano, O., Bazzi, A., Leonardi, G. & Pasolini, G. (2009). Ieee802.16e simulation issues,
Proceedings of IEEE Mobile WiMAX Symposium 2009 (MWS 2009), IEEE, Napa Valley,
California, pp. –.
Bazzi, A., Gambetti, C. & Pasolini, G. (2006). Shine: Simulation platform for heterogeneous in-
terworking networks, Proceedings of IEEE International Conference on Communications,
2006 (ICC 2006), IEEE, Istanbul, Turkey, pp. 5534–5539.
Benyassine, A., Shlomot, E. & Su, H. (1997). Itu-t recommendation g.729 annex b: a silence
compression scheme for use with g.729 optimized for v.70 digital simultaneous voice
and data applications., IEEE Communications Magazine Vol.35(Issue 9): pp.64–73.
Cimini, J. (1985). Analysis and simulation of a digital mobile channel using orthogonal fre-
quency division multiplexing., IEEE Trans. Comm. Vol.COM-33(n.7): pp.665–675.
Estepa, R., Vozmediano, J. & Estepa, A. (2005). Accurate prediction of voip traffic mean bit
rate., ELECTRONICS LETTERS Vol.41(n.17): pp.985–987.
Goode, B. (2002). Voice over internet protocol (voip), Proceedings of the IEEE
Vol.90(n.9): pp.1495–1517.
IEEE Std 802.16-2004 (2004). Ieee standard for local and metropolitan area networks part 16:
Air interface for fixed broadband wireless access systems.
IEEE Std 802.16e-2005 (2006). Ieee std 802.16eŹ-2005 and ieee std 802.16Ź-2004/cor1-2005 ieee
standard for local and metropolitan area networks part 16: Air interface for fixed
and mobile broadband wireless access systems amendment 2: Physical and medium
access control layers for combined fixed and mobile operation in licensed bands and
corrigendum 1.
ITU-T Rec. G.711 (1988). Pulse code modulation (pcm) of voice frequencies.
ITU-T Rec. G.729 (1996a). Coding of speech at 8kbit/s using conjugate-structure algebraic-
code-excited linear-predictive (cs-acelp)coding.
ITU-T Rec. G.729, A. (1996b). A silence compression scheme for g.729 optimized for terminals
conforming to itu-t v.70.
Lee, H., Kwon, T., Cho, D., Lim, G. & Chang, Y. (2006). Performance analysis of scheduling
algorithms for voip services in ieee 802.16e systems, Proceedings of IEEE Vehicular
Technology Conference (VTC 2006-Spring)., IEEE, Melbourne, Australia, pp. 1231–1235.
Li, Q., Lin, X., Zhang, J. & Roh, W. (2009). Advancement of mimo technology in wimax: from
ieee 802.16d/e/j to 802.16m, IEEE Communications Magazine Vol.47(n.6): pp.100–107.
Stern, H., Mahmoud, S. & Wong, K. (1996). A comprehensive model for voice activity in
conversational speech-development and application to performance analysis of new-
generation wireless communication systems, Wireless Networks Vol.2(n.4): pp.359–
367.
Van Nee, R. & Prasad, R. (2000). OFDM for Wireless Multimedia Communications, Artech House.
Throughput-Enhanced Communication Approach for
Subscriber Stations in IEEE 802.16 Point-to-Multipoint Networks 385
Throughput-Enhanced Communication
Approach for Subscriber Stations in
IEEE 802.16 Point-to-Multipoint Networks
Chung-Hsien Hsu and Kai-Ten Feng
Department of Electrical Engineering, National Chiao Tung University
Taiwan, R.O.C.
1. Introduction
The IEEE 802.16 standard for wireless metropolitan area networks (WMANs) is designed to
satisfy various demands for high capacity, high data rate, and advanced multimedia services
(Abichar et al., 2006). The medium access control (MAC) layer of IEEE 802.16 networks sup-
ports both point-to-multipoint (PMP) and mesh modes for packet transmission (IEEE Std.
802.16-2004, 2004). Based on the application requirements, it is suggested in the standard that
only one of the modes can be exploited by the network components within the considered
time intervals, and the PMP mode is considered the well-adopted one. In the PMP mode,
packet transmission is coordinated by a base station (BS) which is responsible for controlling
the communication with multiple subscriber stations (SSs) in both downlink (DL) and up-
link (UL) directions. All the traffic within an IEEE 802.16 PMP network can be categorized
into two types, including inter-cell traffic and intra-cell traffic. For the inter-cell traffic, the
source-destination pair of each traffic flow are located in different cells. On the other hand,
the intra-cell traffic is defined if they are situated within the same cell. The inefficiency within
the PMP mode occurs while two SSs are intended to conduct packet transmission, i.e., the
intra-cell traffic between the SSs. It is required for the data packets between the SSs to be
forwarded by the BS even though the SSs are adjacent with each others. Due to the packet
rerouting process, the communication bandwidth is wasted which consequently increases the
packet-rerouting delay.
In order to alleviate the drawbacks resulted from the indirect transmission, a directly com-
municable mechanism between SSs should be considered in IEEE 802.16 networks. Several
direct communication approaches have been proposed for different types of networks. The
direct-link setup (DLS) protocol is standardized in the IEEE 802.11z draft standard to support
direct communication between two SSs in wireless local networks (IEEE P802.11zTM /D5.0,
2009). However, the DLS protocol is designed as a contention-based mechanism, which does
not guarantee the access of direct link setup and data exchanges between two SSs. The dy-
namic slot assignment (DSA) scheme for Bluetooth networks is proposed in (Zhang et al.,
2002) and (Cordeiro et al., 2003), which is primarily implemented based on the characteris-
tics of the Bluetooth standard. Since frame structures and medium access mechanisms are
different among these wireless communication technologies, both the DLS protocol and DSA
scheme cannot be directly applied to IEEE 802.16 networks.
In this book chapter, a point-to-point direct communication (PDC) approach is proposed for
achieving direct transmission between two SSs. The PDC approach is designed as a flexible
and contention-free scheme especially for time division duplexing based IEEE 802.16 PMP
networks. The BS is coordinating and arranging specific time intervals for the two SSs that are
actively involved in packet transmission. Both the relative locations and channel conditions
among the BS and SSs are utilized as constraints for determining if the direct communication
should be adopted. The advantage of exploiting the PDC approach is that both the required
bandwidth for packet transmission and packet-rerouting delay for intra-cell traffic can be sig-
nificantly reduced. The effectiveness of the proposed PDC approach can be observed via the
simulation results, which demonstrate that the PDC approach outperforms the conventional
IEEE 802.16 transmission mechanism in terms of user throughput.
The remainder of this book chapter is organized as follows. Section 2 briefly reviews the
MAC frame structure and packet transmission mechanism in IEEE 802.16 PMP networks. The
proposed PDC approach, consisting of management structures, an admission control scheme,
and direct communication procedures, is described in Section 3. The performance of the PDC
approach is evaluated in Section 4. Section 5 draws the conclusions.
2. IEEE 802.16 PMP Networks

The PMP mode is considered the well-adopted network configuration in IEEE 802.16 net-
works wherein the BS is responsible for controlling all the communication among SSs. Two
duplexing techniques are supported for the SSs to share common channels, i.e., time division
duplexing (TDD) and frequency division duplexing. The MAC protocol is structured to sup-
port multiple physical (PHY) layer specifications in the IEEE 802.16 standard. In this book
chapter, the WirelessMAN-OFDM PHY, utilizing the orthogonal frequency division multi-
plexing (OFDM), with TDD mode is exploited for the design of the proposed PDC approach.
Both the frame structure and packet transmission mechanism of IEEE 802.16 PMP networks
are described in the following subsections.
2.1 Frame Structure

Fig. 1 illustrates the schematic diagram of the IEEE 802.16 PMP OFDM frame structure with
TDD mode. It can be observed that each frame consists of a DL subframe and a UL subframe.
The DL subframe contains only one DL PHY protocol data unit (PDU), which starts with a
long preamble for PHY synchronization. The preamble is followed by a frame control header
(FCH) burst and several DL bursts. A DL frame prefix (DLFP), which is contained in the FCH,
specifies the burst profile and length for the first DL burst (at most four) via the information
element (IE). It is noted that each DL burst may contain an optional preamble and more than
one MAC PDUs that are destined for the same or different SSs. The first MAC PDU followed
by the FCH is the DL-MAP message, which employs DL-MAP IEs to describe the remaining
DL bursts. The DL-MAP message can be excluded in the case that the DL subframe consists
of less than five bursts; nevertheless, it must still be sent out periodically to maintain syn-
chronization. A UL-MAP message immediately following the DL-MAP message denotes the
usage of UL bursts via UL-MAP IEs. An interval usage code, corresponding to a burst profile,
describes a set of transmission parameters, e.g., the modulation and coding type, and the for-
ward error correction type. The DL interval usage code (DIUC) and UL interval usage code
(UIUC) are specified in the DL channel descriptor (DCD) and UL channel descriptor (UCD)
messages respectively. The BS broadcasts both the DCD and UCD messages periodically to
define the characteristics of the DL and UL physical channels respectively.
time
Frame n-1 Frame n Frame n+1 Frame n+2
DL Subframe UL Subframe
T R
T Initial Bandwidth UL PHY UL PHY UL PHY UL PHY UL PHY T
DL PHY PDU
G ranging request PDU #1 PDU #k-j PDU #k-j+1 PDU #k-1 PDU #k G
Pre- DL burst DL burst DL burst DL burst DL burst DL burst Pre-

FCH UL burst
amble #1 #5 #m-i #m-i+1 #m-1 #m amble
DLFP DL-MAP UL-MAP DCD UCD MAC PDUs MAC PDUs MAC PDUs
UL Subframe
Initial Bandwidth UL PHY UL PHY UL PHY UL PHY UL PHY

ranging request PDU #1 PDU #k-j PDU #k-j+1 PDU #k-1 PDU #k
IE IE IE IE IE IE IE IE IE IE IE IE IE IE IE IE
IE DLFP IE
Pre- DL burst DL burst DL burst DL burst DL burst DL burst IE DL-MAP IE

FCH
amble #1 #5 #m-i #m-i+1 #m-1 #m
IE UL-MAP IE
DL Subframe
Fig. 1. Schematic diagram of IEEE 802.16 PMP OFDM frame structure with TDD mode.
On the other hand, as can be seen from Fig. 1, the UL subframe starts with the contention
intervals that are specified for both initial ranging and bandwidth request. It is noted that
more than one UL PHY PDU can be transmitted after the contention intervals. Each UL PHY
PDU consists of a short preamble and a UL burst, where the UL burst transports the MAC
PDUs for each specific SS. Moreover, a transmit-to-receive transition gap (TTG) and a receive-
to-transmit transition gap (RTG) are inserted in between the DL and the UL subframes and
at the end of each frame respectively. These two gaps provide the required time for the BS to
switch from the transmit to receive mode and vice versa.
2.2 Packet Transmission Mechanism

A connection in IEEE 802.16 PMP networks is defined as a unidirectional mapping between
the BS and an MS, which is identified by a 16-bit connection identifier (CID). Two kinds of
connections, including management connections and transport connections, are defined in the
IEEE 802.16 standard. The management connections are utilized for delivering MAC manage-
ment messages; while the transport connections are employed to transmit user data. During
the initial ranging of a SS, a pair of UL and DL basic connections are established, which be-
long to a type of the management connections. It is noted that a single Basic CID is assigned
to a pair of UL and DL basic connections, which is served as the identification number for
the corresponding SS. Thus the SS uses the individual transport CID to request bandwidth for
each transport connection while the BS arranges the accumulated transmission opportunity
by addressing the Basic CID of the SS.
An exemplified network topology that consists of one BS and two neighboring SSs is shown
in Fig. 2. Two types of traffic exist in the network: inter-cell traffic and intra-cell traffic. For
the inter-cell traffic, the source and the destination for each traffic flow are located in different
cells, e.g., the traffic flow of SS2 for accessing the Internet. On the other hand, the intra-cell
traffic is defined while the source and destination are situated within the same cell network,
SS1
Intra-cell traffic
Internet
BS (or other networks)
SS2 Inter-cell traffic
Wirelses connection Wired connection
Fig. 2. Example of IEEE 802.16 PMP network topology.
Frame n Frame n+1

DL subframe UL subframe DL subframe UL subframe
The jth packet: The jth packet:

from SS1 to BS from BS to SS2
Packet-rerouting delay
Target intra-cell data packet Other data packets
Fig. 3. Schematic diagram of IEEE 802.16 packet transmission mechanism in time sequence.
such as the traffic flow between SS1 and SS2 in Fig. 2. Considering the scenario that SS1
intends to communicate with its neighboring station SS2 , two transport connections are re-
quired to be established via the service flow management mechanism for the intra-cell traffic,
i.e., the UL transport connection from SS1 to the BS and the DL transport connection from the
BS to SS2 . Fig. 3 illustrates the conventional transmission mechanism of IEEE 802.16 PMP
networks in time sequence. In the most ideal case, the jth intra-cell packet, transmitted from
SS1 to the BS in the nth frame, will be forwarded to SS2 in the (n+1)th frame by the BS. The
rerouting process apparently requires twice of communication bandwidth for achieving the
intra-cell packet transmission, which consequently increases control overhead by duplicating
the corresponding data packet. Moreover, the delay time for packet-rerouting can be more
than one half of a frame duration while the packet transmission from the BS to SS2 is post-
poned to a latter DL subframe.
3. Point-to-point Direct Communication (PDC) Approach

The objective of the proposed PDC approach is to provide a directly communicable mecha-
nism for SSs within IEEE 802.16 PMP networks such that both the communication bandwidth
and packet-rerouting delay of intra-cell traffic are reduced. The PDC approach is designed as
a flexible and contention-free scheme wherein the establishment of direct link is conducted
along with packet transmission. Based on the channel conditions among the BS and SSs, the
BS coordinates and arranges specific time intervals for the two SSs that are actively involved
in packet transmission. It is worthwhile to mention that the PDC approach is carried out after
the establishment of the original transmission path, which is compatible and can be directly
integrated with the existing protocols defined in the IEEE 802.16 standard. In the following
subsections, the proposed architecture and management structures will be described in Sub-
section 3.1; while an admission control scheme for direct link establishment is explained in
Subsection 3.2. The direct communication procedures of the PDC approach are given in Sub-
section 3.3.
3.1 Architecture and Management Structure

For the purpose of providing time intervals for direct transmission between SSs, a point-to-
point direct link (PDL) subframe is proposed in the PDC approach. A PDL subframe that
consists of one or more PDL PHY PDUs is designed as a subset of a DL or UL subframe. Each
PDL PHY PDU starts with a short preamble followed by a PDL burst, which is designed to
transport the MAC PDUs for each specific SS. Furthermore, in order to be compatible with
the existing IEEE 802.16 standard, three categories of management structures are proposed,
which are detailed as follows:
• DL-PDL IE and UL-PDL IE. The proposed DL-PDL IE and UL-PDL IE are designed to
depict burst profiles and lengths of their corresponding PDLs in the DL and UL sub-
frames respectively. The DL-PDL IE is a new type of the extended DIUC dependent IE
within the OFDM DL-MAP IE; while the UL-PDL IE is a new type of the UL extended
IE that is contained in the OFDM UL-MAP IE. It is noted that the formats of both the
proposed DL-PDL IE and UL-PDL IE are designed to conform to the formats of the
DL-MAP dummy IE and UL-MAP dummy IE, specified in the IEEE 802.16 standard,
respectively.
• PDL Subheader. The PDL subheader is designed for implementing the request, re-
sponse, announcement, and termination of the direct communication. It is a new type
of per-PDU subheader, which can be inserted in the MAC PDUs immediately followed
by the generic MAC header in both the DL and UL directions. For different purposes,
the DL subheader carries various types of information, including MAC addresses, CIDs,
and location information.
• PBPC-REQ and PBPC-REP Messages. In the IEEE 802.16 standard, the adaptive mod-
ulation and coding (AMC) is exploited as the link adaption technique to improve the
network performance on time-varying channels. The BS selects an adequate modula-
tion and coding scheme (MCS) for a SS based on the reported signal-to-interference and
noise ratio (SINR) value. Moreover, the BS permits the changes in MCS that are sug-
gested by the SS via the burst profile change request message. Similarly, both the pro-
posed PDL burst profile change request (PBPC-REQ) and response (PBPC-REP) mes-
sages are designed to change the MCS applied in a direct link. The PBPC-REQ message
is utilized to request the adjustment of assigned MCS for PDL burst. The BS will re-
spond with the PBPC-REP message for either confirming or denying the alternation in
the suggested MCS.
3.2 Admission Control Procedure

In the PDC approach, some criteria should be exploited to determine the execution of direct
communication between two SSs. A two-tiered admission control scheme for a BS and two at-
tached SSs is presented in this subsection. In wireless communication system, the data trans-
mission range for each station is proportional to its corresponding transmission power. In
order to avoid potential interference introduced by adopting the PDC approach, the distance
factor is considered as the first-tiered constraint (C1 ), which is defined as
C1 : D (SSs , SSd ) ≤ D (SSs , BS),
where D ( x, y) denotes the relative distance between x and y; while the source SS and desti-
nation SS of a intra-cell traffic is represented as SSs and SSd respectively. In other words, the
transmission power utilized by SSs for achieving direct transmission is adjusted to be equal
to or less than that as specified in the conventional IEEE 802.16 mechanism.
On the other hand, for the purpose of enhancing the efficiency for data transmission, channel
conditions among the BS and (SSs , SSd ) pair should be taken into account. Different MCSs
associated with various number of data bits are adopted for data transmission under different
channel conditions. Based on channel states and the corresponding MCSs, the second-tiered
constraint (C2 ) is defined as
C2 : TPDC (SSs , SSd ) ≥ TConv (SSs , SSd ),
where T (SSs , SSd ) represents the raw user throughput defined as "number of bits per second that
is received by the destination SSd while the source is SSs ". In other words, the raw user throughput
resulted from the PDC approach (TPDC ) should be at least equal to or higher than that from the
conventional IEEE 802.16 mechanism (TConv ). The values of both TPDC and TConv are derived
as the description in the following paragraph.
MCS index Modulation Coding rate Coded block size (byte) Receiver SNR (dB)
0 BPSK 1/2 24 3.0
1 QPSK 1/2 48 6.0
2 QPSK 3/4 48 8.5
3 16-QAM 1/2 96 11.5
4 16-QAM 3/4 96 15.0
5 64-QAM 2/3 144 19.0
6 64-QAM 3/4 144 21.0
Table 1. OFDM Modulation and Coding Schemes
Table 1 shows the supported MCSs that are specified within the IEEE 802.16 standard. In the
considered OFDM system, the raw data rate Rd of a MCS with index ξ is represented as
Bu [ξ ]
Rd [ξ ] = , (1)
Ts
where Ts is the OFDM symbol duration. The notation Bu [ξ ] indicates the number of uncoded
bits per OFDM symbol of a MCS with index ξ, which is obtained as
Bu [ξ ] = Nd · log2 M · Rc [ξ ], (2)
where Nd denotes the number of data subcarriers and Rc [ξ ] is the coding rate of a MCS with
index ξ. The value of the parameter M depends on the adopted MCS, i.e., M = 2 for BPSK,
M = 4 for QPSK, M = 16 for 16-QAM, and M = 64 for 64-QAM. Moreover, the OFDM
symbol duration Ts can be acquired as
1+G
Ts = Tb + Tg = Tb + G · Tb = , (3)
f
where Tb and Tg represent the useful symbol time and the cyclic prefix (CP) time respectively.
The notation G denotes the ratio of Tg to Tb . The subcarrier spacing f is obtained as

Fs 8000 n · BW
f = = · , (4)
Ns Ns 8000
where Ns indicates the number of total subcarriers. The notation Fs represents the sampling
frequency with its value specified by the IEEE 802.16 standard as in (4), where n is the sam-
pling factor and BW is the channel bandwidth. By substituting (4) into (3), the OFDM symbol
time can be approximated as
Ns Ns
Ts = · (1 + G ) ≈ · (1 + G ). (5)
Fs n · BW
With (2) and (5), the raw data rate Rd of a MCS with index ξ in (1) becomes
Nd · log2 M · Rc [ξ ] · n · BW
Rd [ξ ] ≈ . (6)
Ns · (1 + G )
Based on (6), the raw user throughput by adopting the PDC approach is acquired as
TPDC (SSs , SSd ) = Rd [ξ (s,d) ], (7)
where ξ (s,d) represents the index of a MCS that will be assigned to the direct link of the (SSs ,
SSd ) pair. On the other hand, the raw user throughput in the conventional IEEE 802.16 mech-
anism is constrained by the two-hop transmission, i.e., from SSs to BS and from BS to SSd .
Thus the TConv can be obtained as
1
TConv (SSs , SSd ) = R [φ ], (8)
2 d (s,d)
where
φ(s,d) = min ξ (s,BS) , ξ ( BS,d) . (9)
The notation ξ (s,BS) denotes the index of the MSC utilized in the link between the SSs and BS;
while that is assigned to the link between the BS and SSd is represented as ξ ( BS,d) .
3.3 Direct Communication Procedures

Based on the aforementioned management structures and admission control scheme, the di-
rect communication procedures of the PDC approach are explained in this subsection. Con-
sidering a basic IEEE 802.16 PMP network that consists of a BS and two SSs, an intra-cell
traffic flow is existed between the SSs. Two transport connections are established for packet
transmission, i.e., a UL transport connection from the source station SSs to the BS and a DL
transport connection from the BS to the destination station SSd . The initialization of direct
communication is achieved by conducting the link request and information collection. The
source-destination pair (SSs , SSd ) anticipating to establish the direct link are required to pro-
vide their location information and channel conditions to the BS. The collected information is
utilized in the admission control scheme mentioned above.
Fig. 4 illustrates an exemplified message flows of the SS-initiated procedure for direct commu-
nication. In the case that SSs intends to conduct direct communication with SSd , it attaches
a PDL subheader to a data packet that will be delivered to the BS; meanwhile, the location
SSs BS SSd
Data packet
DL/UL-MAPs (excludes PDL_IEs)
Data packet
Data + PDL subheader (request)
Data + PDL subheader (request)
SINR detection message
PDL subheader (response)
Fig. 4. Schematic diagram of SS-initiated procedure for direct communication.
SSs BS SSd
Data packet
Data packet
PDL subheader (request)
SINR detection message
Data + PDL subheader (response)
PDL subheader (response)
Fig. 5. Schematic diagram of BS-initiated procedure for direct communication.
information of SSs will be filled into the PDL subheader. As the BS receives the request PDL
subheader from the SSs , the BS will attach a PDL subheader to the data packet and conduct
the transmission to SSd . Moreover, the BS will arrange a DL burst for SSs with the assignment
in the corresponding DL-MAP message. SSs will transmit an SINR detection message to SSd
with the BPSK−1/2 MCS for estimating the channel state of the direct link. After receiving the
PDC subheader and the SINR detection message from the BS and SSs respectively, SSd will
transmit a response PDL subheader associated with the calculated SINR value. It is noted that
the location information of SSd is carried in the response PDL subheader if it is required by
the BS. On the other hand, the BS-initiated direct communication procedure is shown in Fig.
5. Contrary to the SS-initiated procedure, the BS actively announces the link request along
with the PDL subheader to the specific SSs, i.e, SSs and SSd . As SSs receives the requesting
PDL subheader from the BS, it will utilize the response PDL subheader to provide the location
information that is requested by the BS. The remaining steps of the BS-initiated procedure are
similar to that of the SS-initiated case, such as the SINR detection and SSd response.
The BS executes the admission control procedure after it received the response PDL subheader
transmitted from SSd . Based on the collected information, the aforementioned two-tiered con-
trol scheme is exploited by the BS to either confirm or deny the direct communication request
between SSs and SSd . If the request is rejected, the BS will broadcast a denying announce-
ment along with the PDL subheader. On the other hand, a confirming announcement will be
transmitted if the request is granted. Consequently, the BS will arrange the PDL bursts for the
direct link in the subsequent frames.
After receiving the confirmation announcement, the considered SSs will activate the proce-
dure of direct communication. According to the received MAPs associated with the PDL IEs,
SSs will conduct packet transmission directly to SSd within the PDL bursts. Moreover, SSd
will continuously observe and evaluate the cannel condition for the direct link with the adap-
tation to an appropriate MCS. The calculated SINR is compared with the receiver SNR range
of the current MCS (as listed in Table 1) by SSd . If the existing MCS is observed to be im-
proper for the current channel condition, SSd will initiate a PBPC-REQ message to the BS for
suggesting an appropriate MCS. Consequently, the BS will respond a PBPC-REP message with
a recommended MCS.
SSs BS SSd
DL/UL-MAPs (includes PDL_IEs)

(provides bandwidth request opportunity for polling-based service)
Data packet
(non-polling-based service)
Bandwidth request
(polling-based service)
DL/UL-MAPs (includes PDL_IEs)
Data packet
(non-polling-based service)
Data packet
(polling-based service)
Fig. 6. Schematic diagram of bandwidth request procedure in PDC approach.
It is worthwhile to mention that bandwidth requests are conducted by an SS based on indi-

vidual transport connection; while bandwidth grants from the BS is executed according to
the accumulated requests from the SS. In other words, the bandwidth grant is addressed to
the Basic CID of the corresponding SS, not to the individual transport CIDs. As a result, the
CID specified for the PDL burst becomes the Basic CID of SSs . Furthermore, in order to in-
tegrate with the existing specification, the procedures of bandwidth requests and allocations
specified in the IEEE 802.16 standard are implemented within the proposed PDC approach.
Fig. 6 illustrates the bandwidth request procedure while the PDC approach is adopted. It can
be observed that the BS preserves the PDL burst for non-polling based service periodically.
Furthermore, the BS will continue to provide unicast bandwidth request opportunity for the
polling-based services based on the original transport CIDs of SSs . The unicast bandwidth
grant of those services will consequently be assigned to the PDL burst based on the Basic CID
of SSs .
The procedure for the link termination occurs as one of the following two conditions is satis-
fied: (i) the channel condition of the direct link is becoming worse than that from the indirect
channels (i.e., via the BS); (ii) the direct communication is determined to be ceased. It is noted
that the link termination can be initiated by either the BS or SS. In the SS-initiated termination
procedure, the SS will transmit a termination PDL subheader to the BS. As the message is re-
ceived by the BS, it will broadcast an announcement along with a PDL subheader to both SSs
and SSd regarding the termination of the direct link. On the other hand, for the BS-initiated
termination procedure, the termination information is actively announced by the BS. As a
result, the BS and the associated SSs will return to adopt the original packet transmission
mechanism as defined in the IEEE 802.16 standard.
The performance of the proposed PDC approach is evaluated and compared with the conven-
tional packet transmission mechanism in IEEE 802.16 PMP networks via simulations. A single
BS with 12 SSs uniformly distributed within the BS’s coverage are considered as the simulation
layout. The OFDM modulation and coding schemes listed in Table 1 are adopted in the simu-
lation. The occurring frequencies for both inter-cell traffic and intra-cell traffic are considered
uniformly distributed. The packet lengths are selected to follow the exponential distribution;
while the Poisson distribution is adopted for packets arrival time. Since scheduling algorithm
is not specified in the IEEE 802.16 standard, the direct round robin (DRR) (Shreedhar & Vargh-
ese, 1996) and weighted round robin (WRR) (Katevenis et al., 1991) algorithms are selected as
the BS’s DL and UL schedulers respectively. The DRR algorithm is also utilized by the SS to
share the UL grants that are provided by the BS among their connections. The parameters
adopted in the simulations are listed in Table 2.
Parameter Value
Channel bandwidth (BW) 7 MHz
Number of total subcarriers (Ns ) 256
Number of data subcarriers (Nd ) 192
Sampling factor (n) 8/7
Sampling frequency (Fs ) 8 MHz
Useful symbol time (Tb ) 32 µs
CP time (Tg ) 2 µs
The ratio of CP time and useful time (G) 1/16
OFDM symbol duration (Ts ) 34 µs
Maps modulation BPSK
Data modulation QPSK, 16-QAM, 64-QAM
Frame duration 5 ms, 10 ms
SSTTG/SSRTG 35 µs
Initial ranging interval 5 OFDM symbols
Bandwidth request interval 5 OFDM symbols
Average packet size 200 bytes
Simulation time 1 sec
Fig. 7 shows the comparison of the user throughput with an increasing number of intra-cell
traffic flows ranging from 10 to 100 (frame duration = 5 and 10 ms). As can be expected that
the user throughput increases as the number of intra-cell traffic flows is augmented. It can be
observed that the proposed PDC approach outperforms the conventional IEEE 802.16 scheme
5
IEEE 802.16 w/o PDC (5 ms)
4.5 IEEE 802.16 w/ PDC (5 ms)
IEEE 802.16 w/ PDC (10 ms)
4
User Throughput (Mb/s)
3.5
2.5
1.5
0.5
10 20 30 40 50 60 70 80 90 100
Number of Intra−cell Traffic Flows
Fig. 7. Performance comparison: user throughput versus number of intra-cell traffic flows.
3.5
3
2.5
1.5

1 IEEE 802.16 w/ PDC (5 ms)
IEEE 802.16 w/ PDC (10 ms)
0.5
10 20 30 40 50 60 70 80 90 100
Traffic Load (λ)
Fig. 8. Performance comparison: user throughput versus traffic load.
with higher user throughput under different frame durations. In the conventional mechanism,
it is required for the intra-cell traffic to be forwarded by the BS. Consequently, more than twice
6.5
5.5
4.5
3.5
3 IEEE 802.16 w/o PDC (5 ms)

IEEE 802.16 w/ PDC (5 ms)
2.5 IEEE 802.16 w/o PDC (10 ms)
IEEE 802.16 w/ PDC (10 ms)
2
0 20 40 60 80 100
Percentage of Intra−cell Traffic Flows
Fig. 9. Performance comparison: user throughput versus percentage of intra-cell traffic flows.
of the communication bandwidth is necessitate for the packet transmission. By adopting the
proposed PDC approach, the intra-cell traffic can be directly transmitted from the source sta-
tion to the destination station, which resulted in saved bandwidth. Moreover, the longer frame
duration can achieve higher user throughput owing to the reason that less control overheads
are required within the transmission. The comparison of the user throughput under differ-
ent traffic load (λ) is illustrated in Fig. 8, wherein there are 50 intra-cell traffic flows. Similar
performance benefits can be observed by adopting the proposed PDC approach.
In order to evaluate the influence from the inter-cell traffic, the user throughput with an in-
creasing number of inter-cell traffic flows ranging from 10 to 100 is shown in Fig. 9 (with
the number of total traffic flows is equal to 100). It is noticed that the inter-cell traffic can
be considered as a particular type of direct communication within the cell since the packets
are passed from the BS to SS directly. Consequently, the user throughput is decreased as the
percentage of the intra-cell traffic is augmented since there are increasing amounts of indi-
rect links within the network. Nevertheless, the PDC approach can still provide comparably
higher user throughput under different percentages of intra-cell traffic flows. The merits of
the proposed PDC scheme can be observed.
5. Conclusions
In this book chapter, a flexible and contention-free point-to-point direct communication (PDC)
approach is proposed to achieve direct transmission between SSs within IEEE 802.16 PMP net-
works. With the considerations of both relative locations and channel conditions among the
BS and SSs, a two-tiered admission control scheme is proposed to determine the establishment
of direct link between the SSs in the PDC approach. While adapting the PDC approach, the
BS arranges specific time intervals for the two SSs that are actively involved in direct trans-
mission. The advantage of exploiting the PDC approach is that both the required bandwidth
for packet transmission and packet-rerouting delay for intra-cell traffic can be significantly
reduced. Furthermore, the design of the PDC approach is compatible and can be directly in-
tegrated with the existing protocols defined in the IEEE 802.16 standard. The effectiveness of
the proposed PDC approach can be observed via the simulation results, which demonstrate
that the PDC approach outperforms the conventional IEEE 802.16 transmission mechanism in
terms of user throughput.
6. Acknowledgments
This work was in part funded by the Aiming for the Top University and Elite Research Center
Development Plan, NSC 96-2221- E-009-016, NSC 98-2221-E-009- 065, the MediaTek research
center at National Chiao Tung University, the Universal Scientific Industrial (USI) Co., and the
Telecommunication Laboratories at Chunghwa Telecom Co. Ltd, Taiwan.
7. References
Abichar, Z., Peng, Y. & Chang, J. M. (2006). WiMAX: The emergence of wireless broadband,
IEEE IT Prof. 8(4): 44–48.
Cordeiro, C., Abhyankar, S. & Agrawal, D. P. (2003). A dynamic slot assignment scheme
for slave-to-slave and multicast-like communication in Bluetooth personal area net-
works, Proc. IEEE Global Telecommunications Conf. (GLOBECOM), San Francisco, CA,
pp. 4127–4132.
IEEE P802.11zTM /D5.0 (2009). Draft Standard for 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, Amendment 6: Extensions to Direct Link Setup (DLS), IEEE, 3 Park
Avenue, New York, NY 10016-5997, USA.
IEEE Std. 802.16-2004 (2004). IEEE Standard for Local and Metropolitan Area Networks- Part 16:
Air Interference for Fixed Broadband Wireless Access Systems, IEEE, 3 Park Avenue, New
York, NY 10016-5997, USA.
Katevenis, M., Sidiropoulos, S. & Courcoubetis, C. (1991). Weighted round-robin cell multi-
plexing in a general-purpose atm switch chip, IEEE J. Sel. Areas Commun. 9(8): 1265–
1279.
Shreedhar, M. & Varghese, G. (1996). Efficient fair queueing using deficit round robin,
IEEE/ACM Trans. Netw. 4(3): 375–385.
Zhang, W., Zhu, H. & Cao, G. (2002). Improving Bluetooth network performance through a
time-slot leasing approach, Proc. IEEE Wireless Communications and Networking Conf.
(WCNC), Orlando, FL, pp. 592–596.
Holdoff Algorithms for IEEE 802.16 Mesh Mode in Multi-hop Wireless Mesh Networks 399
Holdoff Algorithms for IEEE 802.16 Mesh Mode

in Multi-hop Wireless Mesh Networks
Bong Chan Kim1 and Hwang Soo Lee2
1 SamsungElectronics
2 KAIST
RepublicofKorea
1. Introduction
1
Multi-hop wireless mesh networks (M-WMNs) [Akyildiz, I. F., 2005] are one of the key
features of beyond 3G systems because of their flexibility and low-cost deployment. So far,
most of existing studies on multi-hop wireless mesh networks have been accomplished
based on the IEEE 802.11 ad hoc mode. The IEEE 802.16 working group (WG) specified the
IEEE 802.16-2004 standard [IEEE Std. 802.16-2004, 2004] in October 2004 and the standard
defined two modes: the point-to-multi-point (PMP) mode and the mesh mode. The IEEE
802.16 mesh standard defines three mechanisms to schedule the data transmission:
centralized scheduling (CSCH) [Morge, P. S., 2007], [Han, B., 2007], coordinated distributed
scheduling (C-DSCH) [Morge, P. S., 2007], and uncoordinated distributed scheduling (Un-
DSCH). In the IEEE 802.16 mesh mode with the CSCH, C-DSCH, and Un-DSCH, multi-hop
communication is possible between nodes such as mesh base stations (MeshBSs) and mesh
subscriber stations (MeshSSs) because all nodes are peers and each node can act as routers to
support multi-hop packet forwarding. In particular, in the IEEE 802.16 mesh mode with the
C-DSCH, every node competes for channel access using a distributed election algorithm
(DEA) based on the scheduling information of the extended neighborhoods (one-hop and
two-hop neighbors) in a completely distributed manner and reserves radio resource by a
three-way handshaking mechanism in which nodes request, grant, and confirm available
radio resource using mesh distributed scheduling (MSH-DSCH) message. Like this, because
the IEEE 802.16 mesh mode with the C-DSCH has good flexibility and scalability, it is
suitable as an alternative medium access control (MAC) protocol for establishing M-WMNs.
For M-WMNs to serve as a wireless network infrastructure, the protocol design for M-WMN
should target a high network throughput. In the IEEE 802.16 mesh mode with the C-DSCH,
after occupying radio resource, a node cannot transmit any MSH-DSCH message for a
holdoff time in order to share radio resource with other nodes in M-WMN. If nodes get a
short holdoff time in a heavily loaded network situation, the competition between nodes
will happens severe and thus they will experience long contention times before reserving
radio resource. On the other hand, if nodes get a long holdoff time in a lightly loaded
1 This work was performed while the first author was a Ph.D. student.
network situation, radio resource will be wasted unnecessarily. Like this, network
throughput has a close relationship with the performance of holdoff algorithm.
Fig. 1. Frame structure of IEEE 802.16 mesh mode
This chapter deals with holdoff algorithms for the IEEE 802.16 mesh mode with the C-DSCH
in multi-hop wireless mesh networks and is structured as follows. In Section 2, we present
the overview of the IEEE 802.16 mesh mode with the C-DSCH. In Section 3, we first explain
a static holdoff algorithm, which is defined in the IEEE 802.16 mesh standard mesh
standard, and introduce its limitations with the respect to network throughput. Next, we
describe existing dynamic holdoff algorithm and introduce its advantages and
disadvantages. Lastly, we propose an adaptive holdoff algorithm based on node state. In the
adaptive holdoff algorithm, nodes collect neighborhood information in a distributed manner
and calculate a metric to reflect the current network situation around them. And, nodes
decide appropriate holdoff times on the basis of the calculated metric in order to improve
network throughput. In Section 4, some simulation results are given. Simulation results
show that it is required for nodes to adaptively adjust their holdoff times according to
current network situations in order to improve the performance of M-WMN based on the
IEEE 802.16 mesh mode with the C-DSCH. In Section 5, we make a conclusion.
2. Overview of IEEE 802.16 mesh mode with C-DSCH

In this section, the overview of the IEEE 802.16 mesh mode with the C-DSCH is given in
four aspects: frame structure, MSH-DSCH message format, bandwidth reservation by three-
way handshaking, and distributed election algorithm.
2.1 Frame structure of IEEE 802.16 mesh mode

Figure 1 shows the frame structure of the IEEE 802.16 mesh mode. As shown in the figure,
the frame of the IEEE 802.16 mesh mode is based on the time division multiple access
(TDMA) frame structure and consists of the control and data subframes.
The data subframe is used to transmit data packets and the control subframe is used to
transmit MAC management messages. The control and data subframes are composed of
multiple transmission opportunities (TOs) and minislots, respectively. As one of network
parameters, the number of TOs consisting of the control subframe (MSH_CTRL_LEN) is set
to one value between 0 and 15. Each TO in the control subframe consists of seven OFDM
symbols and can carry only one MAC management message.
The control subframe consists of two types of subframes: the network control subframe and
the schedule control subframe. The network control subframe enables new nodes to join
mesh network. In a mesh network, nodes broadcast network entry and network
configuration information and this enables a new node to get synchronization and initial
network entry into the mesh network. The schedule control subframe is used to transmit
MAC management messages in order to reserve minislots in the data subframe and it is
divided into two parts, as shown in Fig. 1. The first part is used for the MSH-CSCH and
MSH-CCFG messages transmissions in the CSCH mechanism and the second part is used
for the MSH-DSCH message transmission in the C-DSCH mechanism.
The number of TOs (MSH_DSCH_NUM) consisting of the C-DSCH part is selected in a
range between 0 and 15 and thus, the length of the CSCH part is equal to MSH_CTRL_LEN
– MSH_DSCH_NUM.
2.2 MSH-DSCH message format

In this chapter, because we focus on holdoff algorithms for the IEEE 802.16 mesh mode with
the C-DSCH, only the MSH-DSCH message format related with the C-DSCH is given in
detail. Every node sends its available resource information to neighbor nodes via MSH-
DSCH messages. The request, grant, and confirmation of resource are also accomplished by
exchanging the MSH-DSCH messages between a pair of nodes.
As shown in Fig. 2, in the C-DSCH, the MSH-DSCH message contains the following
information elements (IE): [Morge, P. S., 2007]
 MSH-DSCH_Scheduling_IE includes the next MSH-DSCH transmission times and

holdoff exponents of a node and its neighbor nodes.
 MSH-DSCH_Request_IE is used by a node to specify its bandwidth demand for a
specific link.
 MSH-DSCH_Availability_IE is used by a node to convey its own status for individual
minislots to its neighbors.
 MSH-DSCH_Grant_IE is used by a node to send bandwidth grant in response to a
bandwidth request as well as to send a grant confirmation for a received bandwidth
grant.
Fig. 2. MSH-DSHC message format

Fig. 3. Three-way handshaking
2.3 Bandwidth reservation by three-way handshaking

Based on the four IEs above, in the C-DSCH, nodes perform the bandwidth reservation
process by three-way handshaking: bandwidth request, bandwidth grant, and bandwidth
grant confirmation, as shown in Fig. 3.
In the bandwidth reservation procedure, node A (requester) uses the Link ID to uniquely
indentify a link for which node A needs bandwidth, and it sends an MSH-DSCH message
that contains a set of MSH-DSCH_Request_IEs and a set of MSH-DSCH_Availability_IEs.
The MSH-DSCH message with an MSH-DSCH_Request_IE is received by all the neighbors
around node A. After receiving the bandwidth request message, node B (granter) looks up
the set of available minislots in order to select a subset of minislots that is available for data
receptions from node A. Node B chooses an available range of minislots for a bandwidth
grant and sends the MSH-DSCH message with the MSH-DSCH_Grant_IE, which contains
the set of minislots for the bandwidth grant. All the neighbors around node B receive the
bandwidth grant message and they update their availability status by reflecting the
scheduled data reception specified in the bandwidth grant message.
In the bandwidth grant confirmation, node A transmits an MSH-DSCH message with the
MSH-DSCH_Grant_IE in order to inform all the neighbors around node A of the scheduled
data transmission information. The neighbors update their availabilities by reflecting the
newly scheduled data transmission. Data transmission is accomplished only over the
reserved minislots, after the successful transmission of the bandwidth grant confirmation
message.
2.4 Distributed election algorithm

In the three-way handshaking process, nodes independently select their own transmission
times of MSH-DSCH messages using the distributed election algorithm. The DEA enables
nodes to forward MSH-DSCH messages in an M-WMN in a completely distributed manner.
A node can transmit an MSH-DSCH message without message collision at a selected
transmission time by the DEA. In the C-DSCH, the transmission time corresponds to a
specific TO in the schedule control subframe shown in Fig. 1. In the DEA, a node performs
two functions: collecting the next transmission times (Next_Xmt_Times) of all nodes within
its extended neighborhood (one-hop and two-hop neighbors) and selecting its
Next_Xmt_Time.
1) CollectingtheNext_Xmt_Times nodes
ofall to select a collision-
withinextendedneighborhood:
free TO for the MSH-DSCH transmission, every node should collect the Next_Xmt_Times of
all nodes within its extended neighborhood. So, every node must inform its neighbors of the
next MSH-DSCH transmission time of its neighbors as well as itself. In the DEA, nodes
broadcast not an exact Next_Xmt_Time but the next transmission time interval
(Next_Xmt_Time_Interval) information, which is expressed by two parameters of next
transmission maximum (Next_Xmt_Mx) and transmission holdofff exponent
(Xmt_Holdoff_Exp) as follows:
2 *Next_Xmt_Mx<Next_Xmt_Time
Xmt_Holdoff_Exp
2< Next_Xmt_Mx
*(
Xmt_Holdoff_Exp )+1 (1)
The Next_Xmt_Time_Interval is a series of one or more C-DSCH TOs. Every node

broadcasts MSH-DSCH message containing the next transmission time interval information
(Next_Xmt_Mx and Xmt_Holdoff_Exp) of its one-hop neighbors as well as itself. Using the
next transmission interval information received from neighbors via MSH-DSCH messages,
every node can calculate the Next_Xmt_Time_Intervals of all nodes within its extended
neighborhood. In the IEEE 802.16 mesh standard, a node first selects its own
Next_Xmt_Time at a current MSH-DSCH transmission time (Current_Xmt_Time) and then,
it calculates a Next_Xmt_Mx value using equation (1) and a current Xmt_Holdoff_Exp value.
Next, the node broadcasts an MSH-DSCH message containing the Next_Xmt_Mx and
Xmt_Holdoff_Exp, which represents its Next_Xmt_Time_Interval, in order to inform
neighbors of its next transmission time information.
2) Selecting a Next_Xmt_Time: in the IEEE 802.16 mesh standard, after an MSH-DSCH
message transmission, a node is not eligible to transmit an MSH-DSCH message for
transmission holdoff time (Xmt_Holdoff_Time) in order to share radio resource with other
nodes in an M-WMN. Therefore, when a node selects a Next_Xmt_Time in the DEA, it first
sets the temporary transmission time (Temp_Xmt_Time) as follows:
Temp_Xmt_Time=Current_Xmt_Time
Xmt_Holdoff_Time
+ +1 (2)
Next, the node should determine the set of eligible competing nodes related to the
Temp_Xmt_Time from its neighbor table. This set will include neighbors that meet at least
one of the following conditions: [Bayer, N., 2006]
 The Next_Xmt_Time_Interval of neighbor includes the Temp_Xmt_Time,

 The earliest subsequent transmission time (Earliest_Subsequent_Xmt_Time) of
neighbor is equal to or smaller than Temp_Xmt_Time, or
 The Next_Xmt_Time of neighbor is not known.
In the conditions above, the Earliest_Subsequent_Xmt_Time is the earliest time that the node
is eligible to transmit a MSH-DSCH message after the Next_Xmt_Time, as shown in
equation (3):
Earliest_Subsequent_Xmt_Time=Next_Xmt_Time+Xmt_Holdoff_Exp
+2 *Next_Xmt_Mx
Xmt_Holdoff_Exp (3)
The DEA is performed based on this set of eligible competing nodes. With this set, a node
checks whether any competing nodes do not use a specific TO corresponding to the
Temp_Xmt_Time, or not. If any competing nodes of the node do not use the specific TO, the
node is the winner of the distributed election and the Next_Xmt_Time is set equal to the
Temp_Xmt_Time. If a node does not win the distributed election, the Temp_Xmt_Time is set
to Temp_Xmt_Time + 1 and the above process is performed again in order to select a
collision-free Next_Xmt_Time.
3. Holdoff algorithms for IEEE 802.16 mesh mode

In the IEEE 802.16 mesh standard with the C-DSCH, a node is not eligible to transmit any
MSH-DSCH messages for at least holdoff time after an MSH-DSCH message transmission.
The reason to stop an MSH-DSCH message transmission for a holdoff time is for nodes to
share radio resource with other nodes in M-WMN. However, a holdoff mechanism can
result in resource waste by holding MSH-DSCH message transmission even in a lightly
loaded network situation. Some holdoff algorithms have been proposed for the IEEE 802.16
mesh mode with the C-DSCH in order to improve the network throughput and to share the
radio resource.
3.1 Static holdoff algorithm

In the IEEE 802.16 mesh standsard, the Xmt_Holdoff_Time is calculated based on the static
Xmt_Holdoff_Exp and Xmt_Holdoff_Exp_Base values, as shown in equation (4):
Xmt_Holdoff_Time(Xmt_Holdoff_Exp
=2 +Xm t_Holdoff_Exp_Base) (4)
The IEEE 802.16 mesh standard defines static holdoff algorithm. In the static holdoff
algorithm, the Xmt_Holdoff_Exp_Base is fixed to 4 for resource sharing between nodes in an
M-WMN. The Xmt_Holdoff_Exp is also set to a specific value between 0 and 7 in initial
node configuration procedure. That is, all nodes in M-WMN have an identical transmission
holdoff time regardless of current network situation. Thus, when Xmt_Holdoff_Exp = 0, the
competition between nodes happen severe and a node experiences long contention times
before reserving resource. On the other hand, as the Xmt_Holdoff_Exp value increases, the
contention between nodes becomes less competitive; however, nodes get a longer holdoff
time and thus transmission interval becomes longer [Cao, M., 2005]. In particular, if the
Xmt_Holdoff_Exp is set to an unnecessarily large value in a lightly loaded network situation,
the waste of resource happens. Hence, the static holdoff algorithm can result in network
throughput degradation regardless of an assigned Xmt_Holdoff_Exp value.
3.2 Dynamic holdoff algorithm

In [Cao, M., 2005], the authors performed the modelling and performance analysis of
distributed scheduler in the IEEE 802.16 mesh mode, and they concluded that the capacity of
IEEE 802.16 mesh network can be optimized by assigning appropriate Xmt_Holdoff_Exp
values to nodes in the network.
In [Bayer, N., 2007], the dynamic exponent (DynExp) algorithm was proposed. In the
DynExp algorithm, nodes that are currently not sending, receiving, or forwarding data
packets use large Xmt_Holdoff_Exp values and thus get large Xmt_Holdoff_Time values.
Nodes that transmit, receive, and forward data packets or nodes that have been selected by
the routing protocol as potential forwarding nodes use small Xmt_Holdoff_Exp values and
thus get small Xmt_Holdoff_Time values. For this purpose, nodes are classified as follows:
 Mesh base station (M-BS) is a normal mesh base station.

 Active node (ACT) is a node that is part of an active route and sends, receives, or
forwards data packets.
 Sponsor node (SN) is a node that is not part of an active route but has been selected
as a potential forwarding node by at least one of its neighbors.
 Inactive node (IN-ACT) is a node that is not part of an active route and does not send,
receive or forward data packets.
According to the node types above, different Xmt_Holdoff_Exp values are defined as
follows:
0<Xmt_Holdoff_Exp M-BS <Xmt_Holdoff_Exp ACT

<Xmt_Holdoff_Exp SN
<Xmt_Holdoff_Exp IN-ACT ’
By using different Xmt_Holdoff_Exp values according to current node types, the DyExp
algorithm improves network capacity. However, because the operation of DynExp
algorithm depends on information from the routing layer, the DynExp algorithm cannot be
operated independently without cooperation with routing layer.
3.3 Adaptive holdoff algorithm

In this section, we propose an adaptive holdoff algorithm. In the adaptive holdoff algorithm,
nodes adaptively adjust the Xmt_Holdoff_Exp value according to current node state.
3.3.1 Definitions
Before proposing the adaptive holdoff algorithm, the neighbors of a node are classified into
four types: tx-neighbor, rx-neighbor, tx/rx-neighbor, and null-neighbor. Each classification
is defined as follows:
 A tx-neighbor of a node is a neighbor that has a data packet to send to the node.
 An rx-neighbor of a node is a neighbor to which the node has a data packet to send.
 A tx/rx-neighbor of a node is a neighbor that is both tx-neighbor and rx-neighbor of
the node.
 A null-neighbor of a node is a neighbor that has no data packet to send to or receive

from the node.
Fig. 4. Classification of neighbor nodes: A solid line between a pair of nodes indicates that
there is at least one data packet to transmit between the two nodes, with an arrow showing
the direction of the pending data transmission. A dotted line indicates that there is no data
waiting for transmission between the pair of nodes
In Fig. 4, node I has four types of neighbors. Nodes A, B, C, and D correspond to the tx-
neighbor, rx-neighbor, tx/rx-neighbor, and null-neighbor of node I, respectively.
3.3.2 Adaptive holdoff algorithms based on node state

In this section, an adaptive holdoff algorithm based on node state is presented. The adaptive
holdoff algorithm does not depend on information from higher layer such as routing layer
and it operates independently. In addition, it maintains backward compatibility with the
IEEE 802.16 mesh standard.
Fig. 5. Example topology: A solid line between a pair of nodes indicates that there is at least
one data packet to transmit between the two nodes, with an arrow showing the direction of
the pending data transmission
In Fig. 5, node I has two links (Links 1 and 2) for communication with nodes A and B,
respectively. As shown in Fig. 5, node I should reserve each minislot for data reception from
node A and data transmission to node B. On the other hand, node A needs to reserve only
minislots for data transmission to node I, and node B needs to reserve only minislots for
data reception from node I. Therefore, node I should be able to access schedule control
subframe at a higher rate than nodes A and B in order to prevent node I from being a
bottleneck.
Fig. 6. Neighbor table of node I in Fig. 4
Based on the features above, the Weightof a node is first defined for the adaptive holdoff
algorithm as follows:
Weight=tx_exist*1+rx_exist*2 (5)
In equation (5), the tx_exist represents whether a node has at least one tx-neighbor or not.
And, the rx_exist indicates whether a node has at least one rx-neighbor or not. Because a
node transmits only one MSH-DSCH message (bandwidth grant message) to tx-neighbor
over the TO of the schedule control subframe during the three-way handshaking, the
tx_exist is multiplied by 1 in equation (5). On the other hand, because a node transmits two
MSH-DSCH messages (bandwidth request and bandwidth grant confirmation messages) to
rx-neighbor during the three-way handshaking, the rx_exist is multiplied by 2. In addition,
as shown in Fig. 2, the MSH-DSCH message can contains multiple MSH-
DSCH_Request_IE()s, MSH-DSCH-Availability_IE()s, and MSH-DSCH_Grants_IE()s at one
time. Namely, a node can simultaneously request, grant, and confirm resources for multiple
links via only one MSH-DSCH message. Therefore, the Weight depends on only the tx_exist
and rx_exist in equation (5).
Next, a neighbor table is presented in order to explain how to obtain the Weightof a node. In
the adaptive holdoff algorithm, nodes store neighbor information in their neighbor table, as
shown in Fig. 6. Each entry in the neighbor table contains Node_ID, Type_Tx/Rx,
Expire_Tx/Rx, and Expire. Node_ID denotes the identifier of a neighbor. Type_Tx/Rx is
used to distinguish the neighbor types: tx-neighbor, rx-neighbor, tx/rx-neighbor, or null-
neighbor. Expire_Tx/Rx indicates the expiration time of a neighbor as a tx-neighbor or as an
rx-neighbor. Expire represents the expiration time of a neighbor table entry.
In the adaptive holdoff algorithm, the neighbor table is managed in the following way:
1) Setting Type_Tx/Rx, Expire_Tx/Rx, and Expire: Whenever a node sends or receives a data
packet to or from a neighbor over minislots of data subframe, it sets the values of
Type_Tx/Rx, Expire_Tx/Rx, and Expire in the neighbor table entry for the neighbor. When
a node receives a data packet from a neighbor, it finds the neighbor table entry for that
neighbor in its neighbor table and sets the values of Type_Tx, Expire_Tx, and Expire as
follows: the Type_Tx is set to 1. The Expire_Tx and Expire are set to EXPIRE_TIME, which is
a pre-defined constant value. When a node sends a data packet to a neighbor, it finds the
neighbor table entry for that neighbor in its neighbor table and sets Type_Rx, Expire_Rx, and
Expire as follows: the Type_Rx is set to 1. The Expire_Rx and Expire are set to EXPIRE_TIME.
)2 Maintaining/removing neighbor table Maintenance
entries:and removal of a neighbor table
entry is performed using the neighbor timer. The neighbor timer expires periodically; when
this occurs, a node compares the current time with each values of Expire_Tx/Rx and Expire in
a neighbor table entry and updates Type_Tx/Rx, Expire_Tx/Rx, and Expire in the following
way: if the values of Expire_Tx/Rx are higher than the current time, the values of Type_Tx/Rx
and Expire_Tx/Rx are not changed; otherwise, they are set to 0. If the value of Expire is less
than the current time, the neighbor table entry is removed; otherwise, the value of Expire is not
changed. This procedure is repeated for all the entries in the neighbor table.
According to the neighbor table management above, node I in Fig. 4 has the neighbor table
as shown in Fig. 6. In the adaptive holdoff algorithm, a node can calculate its Weightusing
its neighbor table. For example, node I can calculate the Weightby checking the values of
Type_Tx/Rx in its neighbor table entries as follows: if a node has at least one neighbor entry
with the Type_Tx value of 1 in its neighbor table, the tx_exist is set to 1; otherwise, the
tx_exist is set to 0. In a similar way, if a node has at least one neighbor entry with the
Type_Rx value of 1 in its neighbor table, the rx_exist is set to 1; otherwise, the rx_exist is set
to 0. Therefore, node I in Fig. 4 has the Weightvalue of 3 (= 1 * 1 + 1 * 2).
In the adaptive holdoff algorithm, the Weightof a node represents its access rate required to
reserve a shared resource for data transmission or reception. Therefore, a node with a large
Weightvalue should be able to access the TOs of the schedule control subframe at a high rate,
and a node with a small Weightvalue should access the TOs of the schedule control
subframe at a low rate.
To achieve this goal, in the adaptive holdoff algorithm, current transmission holdoff
exponent (Cur_Xmt_Holdoff_Exp) is defined as follows:
Cur_Xmt_Holdoff_Exp=Max_Xmt_Holdoff_Exp*(1Weight
- / 3) (6)
In equation (6), the Max_Xmt_Holdoff_Exp is the maximum value of transmission exponent.

The Max_Xmt_Holdoff_Exp of a node is set to a constant in initial node configuration
process.
And then, a node selects its Xmt_Holdoff_Time as follows:
Xmt_Holdoff_TimeXmt_Holdoff_Base
( =2 +Cur_Xmt_Holdoff_Exp) (7)
Algorithm 1 shows the operation of the adaptive holdoff algorithm. In Algorithm 1, it is

assumed that the Xmt_Holdoff_Exp_Base is set to 4 and the Max_Xmt_Holdoff_Exp to 3.
As shown in Algorithm 1, if a node has a large Weightvalue, the small Xmt_Holdoff_Time
value is obtained by setting the Cur_Xmt_Holdoff_Exp to a small value. Thus, the node can
access the TOs of the schedule control subframe for resource reservation at a high rate. On
the other hand, if a node has a small Weightvalue, the large Xmt_Holdoff_Time value is
obtained by setting the Cur_Xmt_Holdoff_Exp to a large value. Thus, the node can access
the TOs of the schedule control subframe for resource reservation at a low rate. In this way,
nodes can adjust their access rate to the TOs of the schedule control subframe according to
their current state.
Algorithm 1 AdaptiveHoldoffAlgorith
mbasedonNodeState
Notations Used:
= indicate whether a node currently has at least one tx-neighbor
tx_exist
= indicate whether a node currently has at least one rx-neighbor
rx_exist
Weight= access rate of a node
Current_Xmt_Holdoff_Exp = current transmission holdoff exponent
Xmt_Holdoff_Exp_Base = transmission holdoff exponent base (= 4)
Max_Xmt_Holdoff_Exp = maximum value of transmission holdoff exponent (= 3)
Xmt_Holdoff_Time = transmission holdoff time
01: Set the tx_exist

by cheking the Type_Txs of neighbor table entries
02: Set the rx_exist
by cheking the Type_Rxs of neighbor table entries
03: Compute Weight=tx_exist * 1 +rx_exist *2
04: if (Weight== 0)
05: Current_Xmt_Holdoff_Exp=Max_Xmt_Holdoff_Exp ;
06: end if
07: else
08: Current_Xmt_Holdoff_Exp = Max_Xmt_Holdoff_ExpWeight
-1( * )3 / ;
09: end else
10: Compute Xmt_Holdoff_Time = 2 Xmt_Holdoff_Exp_Base
( Current_Xmt_Holdoff_Exp)
+
4. Simulation Evaluation
Computer simulations were performed to evaluate the performance of the holdoff
algorithms. To show the importance of adjusting holdoff time according to current network
situation from the perspective of performance improvement, the performance of the
adaptive holdoff algorithm based on node state (AHA) is compared with that of the static
holdoff algorithm (SHA), which is described in the IEEE 802.16 mesh standard.
4.1 Simulation environments

NCTUns-4.0 [Wang, S. Y., 2007-(a)], [Wang, S. Y., 2007-(b)], [Wang, S. Y., 2007-(c)] is used to
evaluate the performance of the holdoff algorithms for the IEEE 802.16 mesh mode with the
C-DSCH. Table 1 shows the parameters used in the simulation. For all other parameters, the
default values provided in NCTUns-4.0 are used. Figure 7 shows the network topology used
to evaluate the performance of the holdoff algorithms. The network consists of 16 nodes.
Among 16 nodes, one node acts as MeshBS and the others as MeshSSs. Node 11 corresponds
to the MeshBS and other nodes to MeshSSs. All nodes remain stationary for a simulation
time of 300s. The number of traffic flows is varied from 1 to 6 to investigate the performance
variation in different offered loads. Each traffic source generates user datagram protocol
(UDP)-based data packets with the size of 512 bytes at a rate of 2Mbits/s.
Parameters Value
Max. transmission range MeshSS: 400m, MeshBS: 400m
Bytes per OFDM symbol 108 (64-QAM_3/4)
Xmt_Holdoff_Exp_Base 4
Xmt_Holdoff_Exp 3 (only static holdoff algorithm)
Distance between MeshSSs 300m
Table 1. Simulation parameters for holdoff algorithms
Simulations are performed in two scenarios: multi-hop peer-to-peer and multi-hop Internet
access scenarios. Figure 8 shows traffic flows in the multi-hop peer-to-peer scenario. The
multi-hop Internet access scenarios are classified into upload and download patterns. The
traffic flows for upload and download patterns are shown in Figs. 9 and 10, respectively.
To evaluate the performances of two holdoff algorithms (SHA and AHA), the following
performance metrics are used:
 The average per-flow throughput is the average packet throughput at a receiver.
 The total network throughput is the sum of the average packet throughputs at all
receivers.
4.2 Simulation results

In the multi-hop peer-to-peer scenario where traffic is distributed throughout entire network,
the performance of AHA is compared with that of SHA. The average per-flow throughput and
total network throughput are shown in Figs. 11 and 12, respectively. In the SHA, because all
nodes have an identical Xmt_Holdoff_Exp value (= 3), they have an identical
Fig. 7. Simulation network topology

Fig. 8. Multi-hop peer-to-peer
Fig. 9. Multi-hop Internet access with upload pattern

Fig. 10. Multi-hop Internet access with download pattern
holdoff time regardless of their current state. However, in the AHA, because nodes adjust
their Xmt_Holdoff_Exp in a range between 0 and 3 according to their current state, the radio
resource can be used efficiently.
Fig. 11. Average per-flow throughput in multih-hop peer-to-peer

Fig. 12. Total network throughput in multi-hop peer-to-peer
Fig. 13. Average per-flow throughput in multi-hop Internet access - upload

Fig. 14. Total network throughput in multi-hop Internet access - upload
Fig. 15. Average per-flow throughput in multi-hop Internet access - download

Fig. 16. Total Total network throughput in multi-hop Internet access - download
For example, because a node with the Weight= 0 currenlty has no data communication with
neighbors, even though the node uses the Xmt_Holdoff_Exp of 3 in the holdoff process, the
communication problem does not happen. Furthermore, that a node with a small Weight
value reduces the access rate to the schedule control subframe enables other nodes with a
large Weightvalue to access the schedule control subframe at a high rate.
A node with the Weight= 3 can access the schedule control subframe at a high rate by setting
the Xmt_Holdoff_Exp to 0. By adaptively adjusting the Xmt_Holdoff_Exp value according
to the current node state, the AHA performs better than the SHA in terms of average per-
flow throughput and total network throughput, as shown in Figs. 11 and 12.
In general, access to the Internet through MeshBS is desirable for MeshSSs to obtain
necessary service. In the multi-hop Internet access scenario, it frequently happens that
MeshSSs upload files to Internet through MeshBS and MeshSSs download files provided in
the Internet through MeshBS. As illustrated in Figs. 9 and 10, computer simulations are also
performed in the multi-hop Internet access scenario with upload and download patterns.
Figures 13 and 14 show the average per-flow throughput and total network throughput in
the multi-hop Internet access scenario with upload pattern, respectively. In addition, Figs. 15
and 16 show the average per-flow throughput and total network throughput in the multi-
hop Internet access scenario with download pattern, respectively. In the multi-hop Internet
access scenario, it frequently happens that nodes act as one of sender or receiver. For
example, node 12 serves only as sender in Fig. 9, and only as receiver in Fig. 10. In the AHA,
because nodes set their Xmt_Holdoff_Exp to an appropriate value according to their current
role, the AHA outperforms the SHA, as shown in Figs. 13, 14, 15, and 16.
5. Conclusion
Multi-hop wireless mesh networks are one of the key features of beyond 3G system because
of their flexibility and low-cost deployment. The IEEE 802.16 mesh mode with the C-DSCH
has recently emerged as an alternative MAC protocol for establishing M-WMNs. For the
IEEE 802.16 mesh mode to serve as a MAC protocol for M-WMN, it should get a high
network throughput. However, the holdoff algorithm of the IEEE 802.16 mesh standard has
a limitation to the performance improvement of M-WMN. In this chapter, we dealt with
existing holdoff algorithms such as static holdoff algorithm and dynamic holdoff algorithm
and introduced their limitations. In addition, we proposed an adaptive holdoff algorithm
based on node state for the IEEE 802.16 mesh mode with the C-DSCH. The adaptive holdoff
algorithm assigns an appropriate Xmt_Holdoff_Exp value according to current network
situation and it maintains the backward compatibility with the IEEE 802.16 mesh standard.
Simulation results show that it is required for nodes to adaptively adjust their holdoff time
according to current network situations in order to improve the performance of an IEEE
802.16 mesh system.
6. Acknowledgments
This research was supported by the MIC (Ministry of Information and Communication),
Korea, under the ITRC (Information Technology Research Center) and MMPC (Mobile
Media Platform Center) support programs supervised by the IITA (Institute of Information
Technology Advancement)
7. References
Han, B., Jia, W., and Lin, L. (2007). Performace Evaluation of Scheduling in IEEE 802.16
based Wireless Mesh Networks, Elsevier Computer Communications, Vol. 30, No.
4, Feb. 2007, pp. 782-792, ISSN 0140-3664
Akyildiz, I. F., Wang, X., and Wang W. (2005). Wireless Mesh Networks : A Suervey,
Elsevier Computer Networks, Vol. 47, No. 4, Mar. 2005, pp. 445-487, ISSN 1389-
1286
IEEE Std. 802.16-2004 (2004). IEEE Standard for Local and Metropolitan Area Networks: Part
16: Air Interface for Fixed Broadband Wireless Access Systems, Oct. 2004
Cao, M.; Ma, W.; Zhang, Q.; Wang, X. & Zhu, W. (2005). Modeling and Performance
Analysis of the Distributed Scheduler in IEEE 802.16 Mesh Mode, Proceedings of
ACM MobiHoc , pp. 78-89, ISBN 1-59593-004-3, Urbana-Champaign, May 2005,
ACM, Illinois, USA
Bayer, N.; Xu, B. ; Rakocevic, V. & Habermann, J. (2007). Improving the Performance of the
Distributed Scheduler in IEEE 802.16 Mesh Networks, Proceedings of IEEE VTC
Spring , pp. 1193-1197, ISBN 1-4244-0266-2, Burlington Hotel, Apr. 2007, IEEE,
Dublin, Ireland
Bayer, N. ; Sivchenko, D.; Xu, B.; Rakocevic, V. & Habermann, J. (2006). Transmission
Timing of Signaling Messages in IEEE 802.16 based Mesh Networks, Proceedings of
European Wireless, ISBN 978-3-8007-2961-6, National Technical University of
Athens, Apr. 2006, Athens, Greece
Morge, P. S.; Hollick, M. & Steinmetz, R. (2007). The IEEE 802.16-2004 MeSH Mode
Explained Technical Report, KOM-TR-2006-08, Feb. 2007
Wang, S. Y.; Huang, C. H.; Lin, C. C.; Chou, C. L. & Liao, K. C. (2007-(a)). The Protocol
Developer Manual for the NCTUns 4.0 Network Simulator and Emulator, July
2007.
Wang, S. Y.; Chou, C. L. & Lin, C. C. (2007-(b)). NCTUns Hompage. [Online]. Available:
http://nsl10.csie.nctu.edu.tw/
Wang, S. Y.; Chou, C. L. & Lin, C. C. (2007-(c)). The GUI User Manual for the NCTUns 4.0
Network Simulator and Emulator, July 2007
Call Admission Control Algorithms based on
Random Waypoint Mobility for IEEE802.16e Networks 419
Call Admission Control Algorithms based

on Random Waypoint Mobility
for IEEE802.16e Networks 1
Khalil Ibrahimia,b, Rachid El-Azouzia, Thierry Peyrea

and El Houssine Bouyakhfb
aLIA/CERI,UniversityofAvignon
93chemindesMeinajarièsB.P.281 1cedex
9 4 8, Avignon
-9 -France
bLIMIARF/FSR,Mohammed V-AgdalUniversity
Avenue
4 IbnBattoutaB.P.Agdal 410 Rabat
- Morocco
-
1. Introduction
The next generation network WiMAX (Worldwide Interoperability for Microwave Access),
has become synonymous with the IEEE802.16 Wireless Metropolitan Area Network (MAN)
air interface standard. In its original release the 802.16 standard addressed applications in
licensed bands in the 10 to 66 GHz frequency range. Subsequent amendments have
extended the 802.16 air interface standard to cover non-line of sight (NLOS) applications in
licensed and unlicensed bands from 2 to 11 GHz bands. These 802.16 networks are able to
provide high data rates and are preferably based, for NLOS applications, on Orthogonal
Frequency Division Multiple Access (OFDMA) (Piggin, 2004). In OFDMA, modulation
and/or coding can be chosen differently for each sub-carrier, and can also change with time.
Indeed, in the IEEE802.16 standard, coherent modulation schemes are used starting from
low efficiency modulations (BPSK with coding rate 1/2) to very high efficiency ones (64-
QAM with coding rate 3/4) depending on the SNR (Signal-to-Noise Ratio). It has been
shown that systems using adaptive modulation perform better than systems whose
modulation and coding are fixed (Yaghoobi, 2004). Adaptive modulation increases data
transmission throughput and the system reliability by using different constellation sizes on
different sub-carriers.
The authors in (Peyre et al., 2008) introduce a new Quality of Service (QoS) for real-time
calls in IEEE802.16e Multi-class Capacity including AMC scheme and QoS Differentiation
for Initial and Bandwidth Request Ranging (Seo et al., 2004). The QoS defined in this work is
to maintain a same bit rate for RT calls independently of user position in the cell. The
authors use a Discrete Time Markov Chain (DTMC) to model the system over a
decomposition of IEEE802.16e cell. The authors took account the mobility of users among
1This work was supported by a research contract with Maroc Telecom R&D No.
10510005458.06 PI.
the regions of the cell. So, the analysis of wireless systems, as IEEE802.16e WiMAX, often
requires to model the effect of user mobility. The mobility model is a critical element for any
study about the radio communications. Therefore we choose to use the Random Waypoint
(RWP) mobility model. This model have been studied largely in Ad-hoc networks
(Johnson & Maltz, 1996) and briefly in wireless networks (Hyyti & Virtamo, 2007). In
particular we consider (Johnson & Maltz, 1996), wherein the authors introduced the RWP
mobility model to find the mean arrival and departure rates into a concentric cell as well as
the mean sojourn time in the cellular networks.
Call Admission Control schemes play an important part in radio resource management
(Kobbane et al., 2007), (Ibrahimi et al., 2008) and (Ibrahimi et al., 2009). Their aims are to
maintain an acceptable QoS to different calls by limiting the number of ongoing calls in the
system, minimize the call blocking and dropping probabilities and in the same time
efficiently utilize the available resources (Niyato & Hossain, 2005). Our study is motivated
by an attempt to find the better CAC in IEEE802.16e WiMAX, for both traffics RT and BE
that handles the intra or inter cell mobility issue in the downlink of IEEE802.16e WiMAX
with AMC technic. Based on the RWP mobility model for an IEEE802.16e cell, we model the
system capacity through a Continuous Time Markov Chain (CTMC). For this reason we
propose to study two promising CAC algorithms that guarantee a QoS. Our propositions
allow to Internet Service Providers (ISP) to choice a CAC dependently of its purpose to
manage their networks.
The rest of this paper is organized as follows: In Section 2, we describe the system and
mobility properties as well as both CAC algorithms. In section 3, we develop the system and
RWP mobility model. Based on it, the sections 4 and 5, develop the analysis for the first and
second CAC schemes respectively. In Section 6 we define the performance metrics used to
achieve the performance match. The section 7 provides some numerical results before
conclude in section 8.
2. Framework
2.1. IEEE802.16 basis principles
The IEEE802.16e Physical layer uses an OFDMA sub-carrier allocation policy for the data
transmission. The uplink and downlink sub-frames divide the time and frequency space
into sub-carriers. The minimum frequency-time unit of sub-channelization is one slot, and a
frame is constructed by a number of slots. Different sub-carriers are allocated to a mobile
transmission in function of the resource requested by the mobile. Moreover, a sub-channel
can be used periodically by different mobiles due to theirs classes of traffic. Once a mobile is
granted to transmit by a bandwidth response in the DL-MAP, base station assigns one or
more subcarriers and hence defines the sub-channels that the mobile will be able to use for
its data transmission.
An IEEE802.16e cell is organized in region. Each region use specific modulation and coding
technics. Users belongs to a region in function of theirs Signal-to-Noise Ratio SNR. The
number of subcarriers allocated to a mobile directly depends on the available modulation,
the type of traffic and the requested bandwidth.
2.2 System description

In this paper we consider a continuous time performance model for a single IEEE802.16e
cell. The cell is decomposed into several regions. Mobiles are uniformly distributed on the
whole cell. The cell population is dispatched between the regions according with theirs area
coverages. Each region is characterized by the modulation used for data transmissions. Due
to the AMC scheme described in the previous section, the calls use a modulation chosen in
function of the receiver SNR. We consider the Adaptive Modulation and Coding (AMC)
with pathloss only, consequently, the SNR depends only on the distance between the base
station and the calling mobile. The four classes of services defined in the standard are:
• Unsolicited Grant Service (UGS) that caters for real-time, fixed-size data packets with
constant bit rate (CBR). This service is granted at regular time intervals without a request or
polls.
• Real-Time Polling Service (rtPS) that caters for real time data packets that vary in size
generated at periodic intervals, such as MPEG video and VoIP with silence suppression,
where packet sizes are variable.
• Non-Real-Time Polling Service (nrtPS), designed for connections that do not have delay
requirements. nrtPS is similar to BE. It only differs from BE in that it guarantees a minimum
bandwidth, e.g., FTP.
• Best-Effort (BE) that service makes no guarantee of service. To guarantee a minimum
bandwidth the connection must subscribe to the nrtPS service, e.g., web browsing data.
We gather these classes of traffic into two types: Real Time (RT), corresponding to UGS or
rtPS classes, and Best Effort (BE), corresponding to classes nrtPS and BE. Thus the RT class
gathers the non delay-tolerant calls.
The calls can change of roamed region on the basis of the Random Waypoint (RWP) model.
We use the RWP model to determine the mobility behavior of a call over a convex area. This
area (i.e. the cell) is decomposed into several concentric regions. The RWP model helps us to
determine the incoming/migrating rates for each region (intra-mobility). Moreover we
extends our theoretical results to obtain the handover rate to/from external cell (inter-
mobility).
2.3 Connection Admission Control

In Fig. 1, we describe the first CAC algorithm for a new call of class-c. In this first algorithm,
the calls receive the same bit rate depending on its type of traffic (without the priority
between RT and BE calls). Consequently, the system allocate a number of subcarriers in
function of the location of the call. A new class- c call arriving in region i is accepted if the
required resources are available for it. Else, the call is blocked. If a call did not finish its
service in the region i and moves to the neighbor region j = i  1 , the call is accepted in
region j if the system accepts its modulation changing. The available resources of the
system could be not enough to accept a bandwidth increase. In this case, the migrating call
is dropped.
The Fig. 2 represents our last CAC algorithm for a new call of class- c . In this other CAC
algorithm the BE calls have no bandwidth requirement. Since the BE calls tolerate
throughput variation, they will use the sub-carriers left by the RT call occupancy. In fact, the
BE calls are never blocked and received the same resources according with the Processor
Sharing (PS) (Benameur et al., 2001). Thus, the bit rate of a BE call depends on its region (i.e
modulation). For the RT calls only, the CAC algorithm follows the same scheme than
previously: a new RT call arriving in region i is accepted if enough resources are available
for it. And since a RT call did not terminate its service in region i before migrating to a
region j , j = i  1 , the call will be able to remain in the system if enough resources are
available to afford the modulation change.
Class-c call moving

New class-c call in region i
to region j
no Required yes no Modulation

resources
yes
changing
available? accepted?
New class-c New class-c

call is call Class-c call Class-c call
blocked is accepted is blocked is accepted
Fig. 1. First CAC Algorithm decision.
RT call moving to
New class-c call in region i
region j
New BE call Type of New RT call

yes Enough resources
no
class-c call?
with mobility
reservation?
New class-c
call is
Required
n yes
resources
available?
New RT New RT RT call RT call

call call is dropped is accepted
Fig. 2. Second CAC Algorithm decision.

Finally, remarks that for both CAC algorithms, the RT calls are independent of the consume
resources: the RT-call remaining-time only depends on the behavior of the user. Conversely
the BE calls remain in the system in function of the consumed resources: the more sub-
carriers a BE call have, the faster it leaves the system.
In addition, our CAC algorithms seek to reduce the dropping probability: the probability
that an on-progress service is dropped due to its mobility. As explained above, the call
consumes bandwidth in function of the used modulation. By migrating to an outer region, a
call may require additional resources. Thus, this call might undergo a drop due to lack of
available resources. To prevent from these drops, our CAC algorithms introduce a reserved
part of bandwidth. This reservation aims to satisfy the need of additional resources
demanded in case of outer migration.
3. Model
3.1 Cell decomposition and instantaneous throughput
We consider without loss of generality, the Adaptive Modulation and Coding (AMC) with
pathloss only. Then, the OFDMA cell is decomposed into r regions according to the AMC
value corresponding to a certain value of SNR as depected in Fig. 3. Let R i (i=1, …,r) be the
radius of the i-th region and Si represents the corresponding surface. Each region corresponds
to a specific modulation order (see Table 1). In OFDMA scheme, the total number N of sub-
carriers is divided into L sub-channels (or groups) each containing ksub-carriers, such k=N/L .
In our study, we consider the multi-services WiMAX/OFDMA system with two types of
traffics real-time (RT) and best-effort (BE). Also, we define the instantaneous bit rate (radio
interface rate) for a call of class-c located in the region i as follows:
dci = Lic  k  B  ei , (1)
where K is the number of sub-carriers assigned to each sub-channel; B is the baud rate
(symbol/sec); ei is the modulation efficiency (bits/symbol) and Lci is the sub-channels
allowed for class-c call in region i . The above bit rate can be degraded by the error channel
due to collision, shadow fading effect, as defined in (Tarhini & Chahid, 2007),
Rci = dci  (1  BLERi ), (2)
where BLERi is the BLock Error Rate in region i. The Table 1 indicates the modulations and
codings used in a IEEE802.16e cell as function of the user SNR. The SNR requirement for a
BLER less than 10-6 depends on the modulation type as specified in the standard (Standard
IEEE802.16, 2004). Then, we have γ1= 24.4 dB, γ2= 18.2 dB, γ3= 9.4 dB, γ4= 6.4 dB and γ0 =∞.
Modulation Coding rate Received SNR (dB) Cell ratio (%)

64-QAM 3/4 [γ1, γ0) 1.74
16-QAM 3/4 [γ2, γ1) 5.14
QPSK 1/2 [γ3, γ2) 20.75
BPSK 1/2 [γ4, γ3) 39.4
Table 1. IEEE802.16e AMC settings.
Fig. 3. OFDMA Cell decomposed into concentric regions.
3.2 System state and transitions

Our model of the system is based on the Continuous Time Markov chain (CTMC) technic.
The different transition rates within the space of feasible states defined in the next sections
are caused by one of the following events: arrival of a new call of class-c to region i;
migration of an ongoing call of class-c from region ito j ; termination of an ongoing call of
class-c in the region i. Furthermore, we consider in our analysis the following assumptions:
1. The arrival process of new calls of class-c in region iis Poisson with rate c0,i ;
2. The service time of a class-c call is exponentially distributed with mean 1/µc;
3. The mean dwell time or sojourn time in region iis exponentially distributed with mean
1/ ic ;
4. The mean arrival rate of migrating call of class-c from the region ito region jis ci , j .
Let nci (t ) be the number of calls of class-c in progress at time tin region i. The state of the
 r r
system at time t is defined by:n(t )  (n1RT (t ), , nRT (t ), n1BE (t ), , nBE (t )). Then, we

model the process {n(t ), t > 0} as 2r-dimension quasi-birth and death Markov chain. In
the steady-state it has an unique stationary distribution, with:
 r r
n = (n1RT , , nRT , n1BE , , nBE ).
3.3 User mobility behaviour

We compute the arrival migration rates using RWP model. In the RWP model a node moves
2
in a convex domain   R along a straight line segment from one waypoint to another.
The waypoint, denoted by Pi, are uniformly distributed in Ω, Pi U(Ω). Transition from Pi-1
to Pi is referred to as the i-th leg, and the velocity of the node on i-th leg is given by random
variable vi. In particular in the RWP model, it is assumed that Pi and vi are all independent
and vi are uniformly distributed. Here, the domain Ω corresponds to one cell and leg to path
between both waypoints Pi and Pi-1 . Also, the node corresponds to a mobile or user moving
in the cell. With this notation the RWP process (for a single node) is defined by an infinite
sequence of triples (Bettstetter et al., 2004),
{( P0 , P1 , v1 ),( P1 , P2 , v2 ), ....}.
We note that the process RWP is time reversible. This means that the arrival rates across any
line segment or border are equal in both directions. In other words, the average rates of calls
moving from region ito region jper time unit is equal to the number of calls moving from
region jto region iper time unit, i. e., ci , j = cj ,i as proved in (Norris, 1999).
migration rates c . As the velocity of the user
i, j
Our aim is to compute the (node) is
assumed to have an uniform distribution from vmin, (vmin>0) to vmax, denoted by fv(v), where
 1
 , if v   vmin , vmax  ;
f v (v) =  vmax  vmin
0, otherwise.

li
Let T i denote the transition time on i-th leg (path) defined as Ti = , where
vi
li =| Pi  Pi 1 | . The variables li and vi are independent random variables. The average time
from one waypoint to another is given by:
1 ln (vmax /vmin )
E[T ] = l  f v (v)dv = l = l.E[1/v]. (3)
v v vmax  vmin
We consider the area Ai of each concentric cell of radius Ri as a convex disk of same
radius in which the mobiles move according to the RWP model. Our aim is to compute the
arrival rate into a cell of radius Ri 1 . As introduced in (Hyyti & Virtamo, 2007), let
a1 = a1 ( Ri 1 ,  ) denote the distance from point Ri 1 =| (0, Ri 1 ) | Ai to the border of
Ai in direction  (angle anti-clockwise away from the tangent at point
Ri 1 =| (0, Ri 1 ) | ). a2 = a1 ( Ri 1 ,    ) denotes the distance to the border in the
opposite direction (see Fig. 4). Also, we note the specific flux at Ri 1 in direction  by
1
 ( Ri 1 , ) = .a1a2 (a1  a2 ),
2Cv (4)
where Cv = lAi2 .E[1/ v]. We recall that the i  th disk surface is Ai =  Ri2 and the
mean length of a leg in this disk is
1 
l=
Ri2  a a (a
0
1 2 1  a2 )d . (5)
According with the RWP model developed in cellular network context (Hyyti & Virtamo,
2007) , the arrival rate for one user to region i  1 over all contour of disk of the radius Ri 1
is given by

 ( Ri 1 ) = 2Ri 1  sin ( ) ( Ri 1 ,  )d . (6)
0
Fig. 4. RWP domains (disk of radius Ri) and Rz=2Rr-Rr-1.
From the Fig. 6, we deduce easily the distances a1 and a2 as follow:
a1 ( Ri 1 ,  ) = Ri2  Ri21 cos 2 ( )  Ri 1 sin ( ),

a2 ( Ri 1 ,  ) = Ri2  Ri21 cos 2 ( )  Ri 1 cos ( ).
As the user mobility behaviors are independent, the total migration rate from region i
(i = 1, , r ) to region j ( j = i  1 ) of class- c is given by
ci , j =  ( R j ).nci . (7)
Finally, the handover arrival rate is given by
cho = cr 1,r =  ( Rr ).ncr . (8)
Now we can compute the mean sojourn time ic of one mobile that proceed to class-c call in
region i . This mobile can arrives from region j with rate cj ,i ( i = j  1 ) or from outside
as new call with rate c0,i . Let pci be the probability of finding a call of class-c in region i
(i = 1, , r ). Then the mean sojourn time is given by (Hyyti & Virtamo, 2007):
pci
i
 =
c , with c0,1 = 0, (9)
 i 1,i
c  i 1,i
c 0
c ,i
where

pci = P (nci (t )  1) =

 k (n). (10)
n , nci 1
where k is the probability distribution computed in the next system analysis andk=1,2 .
4. System analysis for the first CAC algorithm

4.1 Bandwidth occupancy
As described above in the Fig. 1, the interest QoS is to guarantee for a call of class-c a same
bit rate independently of its position in the cell. In fact, we allocate to it the needed sub-
channels by using equation (2) as
Rc
Lic = . (11)
k  B  ei  (1  BLERi )
We recall that the mean call duration of BE calls in the system depends on the transmitting
RBE
payload in bits, i.e.,  BE = , where E ( Pay ) is the mean file size (Downey,
E ( Pay )
2001). Thus, we define the space of the admissible states as follows
 r
E = {n  N 2 r | {nRT
i
LiRT  nBE
i
LiBE }  L}. (12)
i =1
Let Lm be the reserved capacity for migrating or handoff calls of class-c and L0 denotes

the remaining capacity given by L0 = L  Lm . Let B1 (n) be the occupancy bandwidth

when system state is n , with
 r
B1 (n)  {nRT
i
LiRT  nBE
i
LiBE }. (13)
i =1
4.2 Equilibrium distribution

The call of class-c can come as new call or migrating/handoff call in region i of the cell. For

n the current system state, we define the arrival rate of call in region i , as

c0,i  ci 1,i  ci 1,i , if B1 (n) < L0 ;
 
ci (n) =  (14)
 i 1,i 

 c   i 1,i
c , if L0  B1 ( n ) < L.
We have two classes of services RT and BE. Each class-c call in region i ( i = 1, 2,, r )
i
requires the effective bandwidth L c . Then, we have 2r classes in the cell and the
equilibrium distribution is given by BCMP theorem (Chao et al., 2001) for multiple classes

ci (n)
with possible class changes, with  = i
i
c as
 c  c
i i
 1 r (  RT
i n
) RT (  BE
i n
) BE
 1 ( n) =  i i
, (15)
G i =1 nRT ! nBE !

where n E and G is the normalizing constant given by
ni ni
r
(  RT
i
) RT (  BE
i
) BE
G =  i i
.

nE i =1 nRT ! nBE !
So, this probability depends of the mean sojourn time ic which depends itself on the
i
probability pc in equation (10). Also the latter depends of the above distribution and vice
versa. We can use the fixed point theorem to resolve this problem as follow:
Algorithm : Probability convergence algorithm

1: Initialize the probability in (9): pci ,old = pci = 0.1.
2: Compute the mean sojourn time in (9).

3: Calculate the steady-state probability  1 ( n) from (15).
4: Derive the new value of probability from (10), denoted by pci ,new .
5: Check the convergence of the probability between the old and the new values. if
| pci ,new  pci ,old |<  , where  is a very small positive number, then the new probability is
used to compute the performance metrics. Otherwise, go to step 2 with new value as initial
value. The iterations are continued until to reach the convergence of probability.
5. System analysis for the second CAC algorithm

5.1 Bit rates per class and feasible system states
Here, we consider the CAC algorithm follows the scheme described in the Fig. 2. Consider
there is a minimum capacity reserved for BE calls denoted by LBE and reserved LmRT sub-
channels for RT calls mobility. We denote by LRT the remaining capacity for RT calls given
by:
LRT = L  LBE  LmRT . (16)
 
Let B2 (n) be the bandwidth occupied by RT calls when system sate is n , with
 r
B2 (n) = nRT
i
LiRT , (17)
i =1
where
RRT
LiRT = .
k  B  ei  (1  BLERi )
In this section, we guarantee for RT calls a QoS in terms to maintain a same bit rate
everywhere in the covered area by the cell. Whereas a BE call receives the instantaneous bit
i

rate denoted by RBE in region i . The dynamic capacity C (n) shared fairly among all BE
calls simultaneously in progress in the system is
 
  L  B (n)  Lm , if B (n) < L ;
C ( n) =  2 RT 2 RT
(18)
 LBE , otherwise.
The BE calls share the reserved capacity with PS manner. The number of sub-channels LiBE
allocated to a BE call in region i with PS policy is


i C ( n)
L (n) =| r
BE |, (19)
nBE i
i =1
where | x| x . Then the BE call

indicates the largest integer that is less than or equal to
i
 i

receives in region i the bit rate RBE ( n) = LBE ( n)  k  B  ei  (1  BLERi ). Therefore,
i

R ( n )
the mean BE call duration is given by  BE =
i BE
. Since the system accept without
E ( Pay )
limit the BE calls, the space of admissible states is
 r
F = {n  N 2 r | nRT
i
LiRT  LRT }. (20)
i =1
We define the indication function as
1, if X is true;
 (X ) = 
0, otherwise.
5.2 Transition rates


The transition rates from the state n to other ones are introduced as described in the sequel.
c
Let ni  be the state when a new class-c call is arrived and we denote this transition by
c
q  c . Let ni  be the state when a class-c call in region i terminates its service or changes
( n , ni  )
c
its modulation order and we denote this transition by q  c . Let ni, j ( j = i  1 ) be the
( n , ni  )
state when a class-c call in region i moves to the neighbor region j and we denote this
transition by q   c . Then, we have
( n , ni , j )

q   RT =  ( B2 (n)  LiRT  LRT )RT
0
,i ,
( n , ni  )
q   BE = BE
0
,i ,
( n , ni  )

q   RT =  ( B2 ( n)   iRT
,j
 LRT  LmRT )RT
i, j
,
( n , ni , j )
q   BE = BE
i, j
,
( n , ni , j )
q   RT = nRTi
(  RT  iRT ),
( n , ni  )

q   BE = nBEi
(  BE
i
(n)  iBE ),
( n , ni  )
where ic is computed in Algorithm 1 by replacing the the probability 1 by  2 , and

i, j i j
 RT = L L . RT RT

Let Q be the matrix of the possible transitions, where Q = (q( n ,n') ) for nF and
  
n'  F . The transition rate from state n to n' is denoted by q( n ,n ') . Its value must be
obtained as the sum of all terms in each line in matrix Q is equal to zero for RT call and BE
one as well as i = 1, , r .
5.3 Steady-state distribution


 2 (n) denotes the steady-state probability when system is in the state
Now, we recall that
     
n ( n  F ) and by  the steady-state distribution vector, where  = { 2 (n) | n  F }.
The steady-state probability vector is solution of the following system of equations:

 Q = 0, (21)

 1 = 1. (22)
where 1 is a column vector of ones and 0 is a row vector of zeros.

6. Performances metric
Once the equilibrium distribution probabilities 4.2 and 5.3 are calculated, we compute many
interesting metrics of the system. In this section, we provide explicit expressions for various
metrics like dropping probabilities, blocking probabilities, average sojourn time and average
throughput.
6.1 First scheme

6.1.1 Blocking probabilities
A new call of class-c in region i is blocked with probability:

Bci =

 (n), i = 1, , r ,
1
nEci
(23)
 
where Eci = {n  E | B1 (n)  Lic > L0 }.
6.1.2 Dropping probabilities

Migrating call dropping probability in region i . The migrating call of class-c from region i
to region j is dropped with probability

Dci =

 1 (n), i = 2, , r , (24)
nEci , j
 
where Eci , j = {n  E | B1 (n)  Lcj  Lic > L}.
6.1.3 Average throughput

The average throughput in the system is
 r i
Th1 = 

 1 ( n )(nRT RRT  nBE
i
RBE ). (25)
nE i =1
6.2 Second scheme

6.2.1 Blocking probabilities
A new call of class-RT in region i is blocked with probability

i
BRT =   2 (n), i = 1, , r ,
 i
(26)
nFRT
 
where FRTi = {n  F | B2 (n)  LiRT > LRT }.
6.2.2 Dropping probabilities

Migrating call dropping probability in region i . The migrating call of class-RT from region
i to region j is dropped with probability

i
DRT = 
 i, j
2 (n), i = 2, , r , (27)
nFRT
 
where FRTi , j = {n  F | B2 (n)   RT
j ,i
> LRT  LmRT }.
6.2.3 Average throughput

The average throughput in the system is
 r i 
Th2 = 

 2 ( n )  ( n RT RRT  n i
BE R i
BE ( n )). (28)
nF i =1
7. Numerical applications
The following parameters and assumptions are used in our numerical applications.
Considering OFDMA cell system with an FFT (fast Fourier transform ) size of 2048 sub-
carriers. The cell is decomposed into two regions (r = 2), R 1 = 300 m and R 2 = 600 m with
AMC scheme: 16-QAM 3/4 (e1 = 3 bits/symbol) and QPSK 1/2 (e2 = 1 bit/symbol). The
baud value B = 2666 symbols/sec, BLER i = 0, and K = 48. These parameters correspond to
the transmission modes with conventionally coded modulation (Liu & Zhou, 2005); The bit
rate R RT is equal to 128 Kbps and R BE is 384 Kbps (Tarhini & Chahed, 2007); The total
bandwidth L is 10 sub-channels; The mean call duration for RT calls is equal to 120 sec and
the download of files with mean size for BE calls is equal to E(Pay) = 5 Mbits. We assume a
mobile moves according to the RWP model on a convex disk of radius R z = 900 m. It
randomly chooses a new speed in each waypoint from an uniform
distribution between [vmin, v max ], where vmin=3km/h (low mobility) or
vmin= 20km/h (high mobility), vmax = 90 km/h.
7.1 Impact of first scheme

The Fig. 5 presents the blocking probabilities for each burst profile (i.e. modulation) in terms
of reserved resources for mobility. As expected, the probabilities increase as the reserved
threshold Lm increases and the modulation efficiency decreases. Moreover, an appreciable
difference exists between class RT and class BE blocking probability. This is due to the
required capacity per class type, the BE class call in our numerical environment requires
more bandwidth than the RT class call. So, when threshold increases, the blocking
probability increases due to the CAC mechanism which gives the priority to
migrating/handoff call than the new call. So the blocking probability mainly depends on the
required bandwidth associated with the call modulation efficiency. We also observe that as
the reserved bandwidth Lm increases, the calls are more blocked as the incoming region is
away from the base station and as the calls demand an high bandwidth. In particular, we
observe on the figure the blocking probabilities for two type of arrivals : BE calls in 16QAM
and RT calls in QPSK. The blocking probabilities are exactly the same because the product
between the required bandwidth and the modulation efficiency are the same for both.
0.2
0.18
0.16
Mean speed = 46.5 km/h
0.14
Bocking probabilities
0.12
0.1
scheme 1 : class-RT 16-QAM
0.08 scheme 1 : class-RT QPSK
scheme 1 : class-BE 16-QAM
0.06 scheme 1 : class-BE QPSK
0.04
0.02
0
0 0.02 0.04 0.06 0.08 0.1 0.12 0.14 0.16 0.18
Reserved portion for mobility Lm
Fig. 5. Blocking probabilities versus threshold Lm and mean speed for
 0
RT ,i = 0
BE ,i = 0.3 call/sec.
The Fig. 6 shows the average throughput of the whole cell versus the reserved threshold
mobility. On this figure two singular behaviors have to be studied. The first main
observation deals with the throughput increasing because of an higher mobility. As a mobile
moves faster, it increases its probability to change region (and thus modulation) per unit of
time. This fact implies also more calls dropped due to lack of resources. In fact the system
will implicitly coerce the system calls to use better modulations. This last deduction explains
how an higher mobility allows to reach a better throughput. Note that this remark is
confirmed in (Peyre & Elazouzi, 2009) and also observed for the ad-hoc networks
(Grossglause & Tse, 2002). Moreover, we can criticize on the Fig. 6, the impact of resource
reservation to ease the call mobility. For any mobility behavior, the average cell throughput
decreases as the reservation share increases. But the throughput fall depends on the mobility
behavior. In fact, by introducing a resource reservation, we helps more and more migrating
calls to remain in the border regions. Theses calls use more subchannels to reach the same
bit rate than previously. Consequently, the system reaches a lower throughput by
prioritizing the user mobility management.
1.3
scheme 1 : mean speed = 46.5 km/h
1.2 scheme 1 : mean speed = 55 km/h
1.1
Average cell throughput-[Mbps]

1
0.9
0.8
0.7
0.6
0.5
0 0.02 0.04 0.06 0.08 0.1 0.12 0.14 0.16 0.18
Fig. 6. Average throughput versus threshold Lm and mean speed for RT
0
,i = BE ,i = 0.3
0
call/sec.
The Fig. 7 represents the dropping probabilities versus reserved threshold Lm . We plot the
results for both types of traffic in the border region (i.e. QPSK modulation), and for two
mobility behaviors. The figure helps to appreciate the great impact of the resource
reservation on the mobility management efficiency. Concerning the real-time traffics, we
observe that an higher mobility causes an twenty-times dropping increase. To fight against
this effect, our CAC algorithm permits to tune the resource reservation in order to decreases
the call drops under a desired value. For examples, a reservation lower than ten percents of
the total bandwidth decreases the number of dropped calls under a one-percent probability.
For the Best Effort traffic, we observe exactly the same behavior. Nevertheless, the dropping
probability for the BE traffic remains higher than for the RT because the BE calls require an
higher bit rate. As they ask for more bandwidth, they undergo more drops. For this type of
traffic, our CAC allows also to greatly reduce the dropping probability of the BE calls by
increasing the resources reservation Lm .
7.2 Impact of second scheme

The Fig. 8 shows the blocking probabilities for the RT traffics in both regions versus the
bandwidth reservation for the Best Effort call. We plot the results obtained for two resource
reservation profiles for the user mobility management. This figure clearly shows the impact
-7
x 10
1.6
scheme 2 : Lm
RT
=1%, mean speed = 46.5 km/h, 16-QAM
1.4
scheme 2 : Lm
RT
=1%, mean speed = 46.5 km/h, QPSK
1.2 scheme 2 : Lm
RT
=3%, mean speed = 46.5 km/h, 16-QAM
RT call blocking probabilities
scheme 2 : Lm
RT
=3%, mean speed = 46.5 km/h, QPSK
1
0.8
0.6
0.4
0.2
0
0 0.05 0.1 0.15
Minimum reserved portion LBE
Fig. 7. Blocking probabilities versus threshold Lm and mean speed for
 0
RT ,i = 0
BE ,i = 0.3 call/sec.
0.5
0.45
scheme 1: mean speed = 46.5 km/h, class-RT
Dropping probabilities in region 2 (QPSK)
0.4 scheme 1: mean speed = 45 km/h, class-RT

scheme 1: mean speed = 46.5 km/h, class-BE
0.35
scheme 1: mean speed = 45 km/h, class-BE
0.3
0.25
0.2
0.15
0.1
0.05
0
0 0.02 0.04 0.06 0.08 0.1 0.12 0.14 0.16 0.18
Fig. 8. Dropping probabilities versus threshold LBE and LmRT for RT
0
,i = BE ,i = 0.3
0
call/sec.
of LBE and LmRT on the blocking probabilities experienced in each region. In general, the
blocking probabilities are heavily increased by the BE bandwidth reservation. In addition,
this drawback is amplify by the increase of the mobility management resource reservation.
From this statement, we could compute the the value ranges for LBE and LmRT which
satisfy a maximum blocking probability threshold.
The Fig. 9 shows the dropping probabilities for the RT traffics in the border region versus
the bandwidth reservation for the Best Effort call. We plot the results obtained for two
resource reservation profiles for the user mobility management. The figure shows how the
resource reservation LmRT fights against the dropping due to the BE bandwidth reservation.
On this figure, the dropping probability is reduced as the BE reservation is greater than ten
percents. So we can determine the possible values for LmRT from a desired maximum
dropping probability.
The Fig. 10 shows the total cell throughput. This figure presents the impact of the different
bandwidth reservation to increase the BE call throughput and to ease the call mobility. On
the Fig. 10 we can appreciate the great impact of the BE bandwidth reservation LBE .
Indeed, by reservation 15% of the total bandwidth, we double the total cell throughput. In
addition, to increase the mobility-purpose reservation slightly decrease the total throughput.
This observation leads to think that a small quantity of subchannels for Best effort calls
allows to reach very higher throughput. In fact, this by reserving few subchannels to the BE
calls, we also increase the blocking and dropping probabilities for the RT calls. Therefore,
these amount of resources freed by the dropped call or let free by the blocking one allow to
the BE calls to use even more resources.
0.24
scheme 2 : Lm
RT
=1%, mean speed = 46.5 km/h
0.22 scheme 2 : Lm =3%, mean speed = 46.5 km/h
RT
Average cell throughput-[Mbps]
0.2
0.18
0.16
0.14
0.12
0.1
0 0.05 0.1 0.15
Fig. 9. Mean throughput versus LBE and Lm for RT

0
,i = BE ,i = 0.3 call/sec.
0
-8
x 10
3.5
Dropping probabilities in region 2 (QPSK)

3 scheme 2 : Lm
RT
scheme 2 : Lm
RT
2.5
1.5
0.5
0
0 0.02 0.04 0.06 0.08 0.1 0.12 0.14 0.16
Fig. 10. RT dropping probabilities versus threshold LBE and LmRT for RT
0
,i = BE ,i = 0.3
0
call/sec.
8. Conclusion
Recently many works have been introduced to study the WiMAX performances in order to
improve some QoS for users. In this sense our work contributes in order to improve the QoS
of users. In fact, by considering both traffics RT and BE, we proposed two strategies of QoS
management. The first defines constant bit rates (CBR) for RT and BE calls. We also
introduce a resource reservation to ease the mobility of the users. The second scheme
replaces the CBR policy of the BE calls by a Processor Sharing (PS), and we add an other
bandwidth reservation to improve the minimum Best Effort call throughput. Moreover, we
define a realistic mobility model through the accurate Random Waypoint model. Based on
these propositions we develop a continuous time Markov chain which determines the
steady state of the system. From the model, we provide a large range of performance
metrics. We conclude our analysis by criticize the impact of each CAC algorithm parameters
on the system performances. By gathering all our results, we can easily choose one of the
two proposed CAC algorithms to meet with the desired traffic prioritization policy. In
addition, we analyzed the possible ways to tune the parameters of each CAC algorithms in
order to specify the main thresholds (average throughput, blocking and dropping
probabilities). As future works we seek to introduce thinking times in our RWP model. Our
main objective is also to improve the CAC algorithms by determining the best way to roam
the mobile user to a region, in function of its speed and the expected time spend in the next
region.
9. References
Benameur, N.; Ben Fredj, S.; Delcoigne, F.; Oueslati-boulahia, F.; Roberts, J. W.; &
Moulineaux, I. L. (2001). Integrated admission control for streaming and elastic
traffic. Journal Lecture Notes in Computer , Vol.Science
2156, (February, 2001), pp. 69-
82, ISSN 0302-9347.
Bettstetter, C.; Hartenstein, H. & Perez-Costa, X. (2004). Stochastic properties of the random
waypoint mobility model. ACM/KluwerWireless Networks: Special Issue on Modeling
andAnalysisofMobileNetworks , Vol. 10, No. 5, (September, 2004).
Chao, X.; Miyazawa, M.; Pinedo, M. & Atkinson, B (2001). Queuing networks: Customers,
signals and product form solutions. The Journal of the Operational Research Socie
Vol. 52, No. 5, (May, 2001), pages number (600-601).
Downey, A.B. (2001). The structural cause of file size distributions. IGMETRICS/Performance,
Vol. 29, No. 1, (June, 2001), pages number (328–329), ISSN 0163-5999.
Grossglauser, M. & Tse, D. (2002). Mobility increases the capacity of ad-hoc wireless
networks. IEEE/ACM Transactions on Networking, Vol. 10, (August, 2002) pages
number (477-486), ISSN 1063-6692.
Hyyti, E. & Virtamo, J. (2007). Random waypoint mobility model in cellular networks.
Wireless Network , Vol. 13, No. 2, (Avril, 2007), pages number (177-188), ISSN 1022-
0038.
Ibrahimi, K.; El-Azouzi, R.; & Bouyakhf, E.H. (2008). Uplink Call Admission Control in
Multi-services W-CDMA Network. Proceedings of 13th IEEE Symposium on
ComputersandCommunications, ISCC' ( July )80 2008, Marrakech, Morocco.
Ibrahimi, K.; El-Azouzi, R.; S.K. Samanta; & Bouyakhf, E.H. (2009). Adaptive Modulation
and Coding scheme with intra- and inter-cell mobility for HSDPA system. To apear
in the Sixth International Conference on Broadband Communications, Networks,
and Systems (IEEE/ICST BROADNETS), Septembre 2009, Madrid, Spain.
Johnson, D.B. & Maltz, D.A. (1996). Dynamic source routing in ad hoc wireless networks. In
thebook MobileComputing , pages number (153–181). Kluwer Academic Publishers.
Kobbane, A.; El-Azouzi, R.; Ibrahimi, K. & Bouyakhf, E.H (2007). Capacity Evaluation of
multi-service W-CDMA : A Spectral Analysis Approach. Proceedings of the Sixth
IASTED International Conference on unication
Comm Systems and Networks, CSN,
August 2007, Palma of Mallorca, Spain.
Liu, G.O. & Zhou, Z. (2005). Queuing with adaptive modulation and coding over wireless
links: Cross-layer analysis and design. IEEE TRANSACTIONS ON WIRELESS
COMMUNICATIONS , Vol. 4, No. 3, (May 2005).
Niyato, D. & Hossain, E. (2005). Call admission control for qos provisioning in 4G wireless
networks: issues and approaches. IEEE Network, Vol. 19, No. 5, (2005) pages
number (5–11), ISSN 0890-8044.
Norris, J.R. (1999). Markovchains. Cambridge University ess,PrISBN = 0-521-68181-3.
Peyre, T. & El-Azouzi, R. (2009). IEEE802.16 multi-class capacity including AMC scheme,
mobility and QoS differentiation for initial and bandwidth request ranging. To
appearintheproceedings, of Hungary,
WCNC April 2009, Budapest.
Peyre, T.; Ibrahimi, K. & El-Azouzi, R. (2008). IEEE802.16 multi-class capacity including
AMC scheme and QoS differentiation for initial and bandwidth request ranging.
ProceedingsofValuetools , Greece, October 2008, Athens.
Piggin, P. & Depth, W. (2004). Broadband wireless access. IEEE Communications Engineer,
Vol. 2, Issue 5, (Octobre 2004), pages number (36-39).
Seo, H.H.; Ryu, B.H.; Hwang, E.S.; Cho, C.H. & Lee, N.W. (2004). A study of code
partioning scheme of efficient random access in OFDMA-CDMA ranging
subsystem. Proceedings of JCCI, pages 262, April 2004.
Standard IEEE802.16. (2004). IEEEstandardforlocalandmetropol itanareanetworks,part16:Air
interfaceforfixedbroadbandwirelessaccesssystems, New York, USA.
Tarhini, C. & Chahed, T. (2007). On capacity of OFDMA-based ieee802.16 WiMAX including
adaptative modulation and coding and inter-cell intereference. Proceedings of the
15 thIEEEWorkshoponLocalandMetr opolitanAreaNetwork(LANMAN) , Princeton NJ,
June 2007, USA.
Yaghoobi, H. (2004). Scalable OFDMA physical layer in IEEE 802.16 Wirelessman. Intel
TechnologyJournal , Vol. 8, (August 2004), pages number (201–212).
Queueing-Model-Based Analysis for IEEE802.11 Wireless LANs with Non-Saturated Nodes 441
Queueing-Model-Based Analysis for IEEE802.11

Wireless LANs with Non-Saturated Nodes
Shigeo Shioda and Mayumi Komatsu
Chiba University
Japan
1. Introduction
The IEEE 802.11 protocol has gained widespread popularity as a standard MAC-layer proto-
col for wireless local area networks (WLANs). The IEEE 802.11 standard defines Distributed
Coordination Function (DCF) as a contention-based MAC mechanism, but it does not have
quality-of-service (QoS) functionality. The IEEE 802.11 standard group has approved the
802.11e standard for MAC layer QoS enhancements to the former 802.11 standard, where the
Enhanced Distributed Channel Access (EDCA) function of 802.11e is a QoS enhancement of
the DCF.
While the IEEE 802.11e claims to support the QoS, several challenging problems still remain
on the support of real-time applications with strict QoS requirements. Under the DCF, the
real-time bidirectional applications like Voice over IP (VoIP) cannot efficiently utilize the
bandwidth of WLANs. The inefficient bandwidth utilization is mainly caused by the up-
link/downlink unfairness problem in WLANs (Cai et al., 2006). The DCF assigns the same
number of access opportunities to each individual mobile terminal as well as the access point
(AP), but each mobile terminal serves one uplink flow while the AP needs to serve all down-
link flows. Thus, a downlink flow necessarily gets comparatively lower bandwidth than an
uplink flow gets. The unfairness between uplink and downlink flows likely builds up the
queue at the access point (AP) and causes packet loss of downlink flows even at moderate
load, where unused bandwidth still remains for uplink flows in a WLAN. This implies that,
under the use of bidirectional applications, the AP is likely to become performance bottleneck
over the standard WLANs. Note that the occupancy of the AP buffer strongly depends on
the throughput of the WLAN, at which rate the AP can successfully transfer frames over the
WLAN. Thus, the performance of bidirectional applications over WLANs needs to be ana-
lyzed by taking into account both the occupancy of the AP buffer and the throughput of the
IEEE 802.11 DCF (or EDCA).
In this article, we present a mathematical model for evaluating the MAC-layer performance
such as per-flow throughput as well as the network-layer performance such as packet loss at
each station. Our proposal combines a Markov chain model for evaluating the throughput of
the IEEE 802.11 DCF and a queueing model for analyzing the network-layer performance of
each station. The Markov chain model used in our proposal is primarily based on the model
by Malone (Malone et al., 2007), which allows us to analyze the throughput of IEEE 802.11
under unsaturated nodes, but we have made an important extension of their model in order
to consider the effect that arriving IP packets are queued in the buffer of a station when the
station has frames to transmit. For analyzing the network-layer performance, we apply the
GI/M/1 model, where the service time corresponds to the MAC-layer packet service time,
which is the time interval between the instant that a packet reaches the head of the queue and
the instant that the packet is successfully transferred. It was shown in (Zhai et al., 2004) that
the exponential distribution is a good approximation for the MAC-layer packet service time,
the mean of which can be evaluated throughput the analysis of the IEEE 802.11 DCF.
Through extensive simulations using the simulator ns2, we show that our model can accu-
rately predict how many VoIP conversations can be multiplexed over a WLAN without loss
of any packets. Our model allows us to evaluate the IEEE 802.11 DCF with contention win-
dow (CW) differentiation, which is a service differentiation scheme provided by EDCA. By
using this feature of our proposal, in this article, we also investigate how much the CW differ-
entiation could improve the bandwidth utilization by making contention window of the AP
smaller than mobile terminals.
This article is organized as follows: in Section 2 we present review on related work. In Section
3, we propose an analytical model to evaluate the performance of the IEEE 802.11 DCF under
non-saturated conditions. In Section 4, we present a queueing model to analyze the queueing
delay and packet loss ratio at the buffer of the AP or mobile terminals. In Section 5, we show
the results of simulation experiments to show the accuracy of the proposed model. In Section
6, we conclude the article with a few remarks.
2. Related Work
The performance of the IEEE802.11 has been widely studied in the literature. Bianchi (Bianchi,
2000) proposed a two-dimensional Markov chain model to analyze the performance of the
IEEE 802.11 DCF under the so-called saturation condition, in which all stations always have
data to send. Robinson et al. (Robinson & T.S.Randhawa, 2004) and Xiao (Xiao, 2005) extended
Bianchi’s DCF model to analyze the performance of the EDCA function of IEEE802.11e under
the saturation condition.
Since persistent saturation continues only during a short time period in actual operation, it is
important to evaluate the performance of IEEE 802.11 under non-saturation conditions. Ergen
et al. (Ergena & Varaiya, 2005) proposed an extension of the Bianchi’s DCF model by introduc-
ing additional states to the Bianchi’s Markov chain to represent idle states of a station. Malone
et al. (Malone et al., 2007) developed a different extension of the Bianchi’s DCF model; their
model allows stations to have different packet-arrival rates. Daneshgaran et al. (Daneshgaran
et al., 2008) proposed an analytical model for non-saturated conditions in order to account for
packet transmission failures due to errors caused by propagation through the channel. Foh et
al. (Foh et al., 2007) proposes to use a queueing model to evaluate the performance of IEEE
802.11 under non-saturated conditions. In their queueing model, customers in the system
represent active stations, where being “active" means having frames to send. The Zhao et al.
(Zhao et al., 2008) proposed approximating the attempt rate, at which a station attempts to
send a frame, in non-saturated setting by scaling the attempt rate of saturated setting with the
probability that a packet arrives.
As we have explained, in the use of bidirectional applications, packets are likely to be delayed
and dropped in the buffer of the AP. The queueing delay and the packet loss in the buffer
of the AP would largely affect the performance of real-time applications. All of the studies
mentioned in the above, however, could not analyze the queueing delay and the packet loss
at the buffer of the AP or each station. Several proposals have been made to conduct cross-
layer analysis where the performance of the network layer such as queueing delay or packet
loss at stations is jointly evaluated with the MAC-layer performance such as the throughput
(Cheng et al., 2007; Tickoo & Sikdar, 2004; Xiang et al., 2007; Zhai et al., 2004). For example,
Zhai et al. (Zhai et al., 2004) integrated the Bianchi’s Markov-Chain with a queueing model.
Tickoo et al. (Tickoo & Sikdar, 2004) proposed a similar model where a simplified Bianchi’s
model was used. The proposal by Xiang et al. (Xiang et al., 2007) corresponds to the exten-
sion of the Zhai’s model to non-saturated conditions. The existing proposals concerning the
cross-layer analysis approximate the Bianchi’s Markov-chain by a simplified model. Our ana-
lytical model, which is categorized into the cross-layer analysis, attempts to directly integrate
Bianchi’s (or Malone’s) Markov-chain with the queueing model.
3. Model of Non-saturated Stations

In this section, we present a bi-dimensional Markov model for evaluating the performance
of IEEE 802.11 DCF under non-saturated conditions. We represent the state of each station
by a pair of integers (s(t), b(t)), where s(t) and b(t) respectively denote the back-off stage
and counter of a given station (say station A) at time t. We also let {t1 , t2 , . . . , } denote state
transition instants of station A. Note that {(s(t), b(t)), t ≥ 0} is not a continuous-time Markov
process because the inter-state-transition time is not exponentially distributed. The state at
state-transition instants {(s(tn ), b(tn )), n ≥ 1}, however, would define a Markov chain, where
{tn }n∈N form imbedded Markovian points. In the following, we focus on the state transitions
on imbedded Markovian points and simply represent the state of a station by (s, b), omitting
the time parameter t.
3.1 Per-station Markov Model

Assume that there are n stations (one access point and n − 1 terminals) in the system. The
back-off stage starts at 0 at the first attempt to transmit a packet and increases by 1 every
time a transmission attempt results in a collision up to the maximum value. We denote the
maximum back-off stage of station l (l = 1, . . . , n) by ml . The maximum back-off stage is
related to CWmax through 2ml W0 = CWmax + 1 where W0 = CWmin + 1. The probability that a
transmission attempt of station l results in a collision is assumed to be pl . The back-off stage
is reset at 0 after a successful transmission. At the back-off stage s, the back-off counter is
initially chosen uniformly between [0, Ws − 1], where Ws = 2s W0 . The counter decreases by
one at the start of every time slot when the medium is sensed idle. Note that the back-off
counter is suspended when the medium is busy due to the transmission (or collision) by other
stations. When the back-off counter reaches zero, the station attempts to transmit a frame at
the start of the next time slot.
When the back-off stage of station l reaches the maximum value ml , it remains ml even if the
station consecutively fails to send frames. Note that the frame is discarded and the back-off
stage is reset at 0 when the number of consecutive-frame-retransmission exceeds the retry
limit. In this article, however, we do not consider the influence on the frame discard due to
consecutive transmission failures because the frame discard due to the consecutive retrans-
mission failures rarely occur in usual cases. This simplification was also used in Bianchi
(Bianchi, 2000) and Malone (Malone et al., 2007).
In non-saturated conditions, a station may not have a frame to transmit just after transmitting
a frame and resetting the back-off stage and timer. In this paper, such a station is referred to
as being “post-backoff". As used in Malone (Malone et al., 2007), we introduce notation (0, k )e
for k ∈ [0, W0 − 1] to represent a post-backoff station with back-off timer k. A station in state
(0, k)e makes a transition into (0, k − 1) at the start of the next time slot if (at least) one frame
has arrived during the current time slot; otherwise it enters (0, k − 1)e . We assume that the
transition probability from state (0, k )e to state (0, k − 1) of station l is ql .
A station in state (k, 0) (0 ≤ k ≤ ml ) attempts to transmit a frame at the beginning of the
next time slot. In the case of a successful transmission, it makes a transition into one of post-
backoff states ((0, k )e , k = 0, . . . , W0 − 1) with probability 1 − rl , and it makes a transition into
one of backoff states with stage 0 ((0, k ), k = 0, . . . , W0 − 1) with probability rl . In the case of
a collision, it enters one of states with back-off stage k + 1 (when 0 ≤ k < ml ) or ml (when
k = ml ). More precisely,
P[(0, l )e |(k, 0)] = (1 − rl )(1 − pl )/W0 ,
P[(0, l )|(k, 0)] = rl (1 − pl )/W0 ,
P[(k + 1, l )|(k, 0)] = rl (1 − pl )/Wk+1 , for 0 ≤ k < ml
P[(m, 0)|(m, 0)] = rl (1 − pl )/Wml . (1)
Parameter rl is the probability that station l has at least one frame after frame transmission.
If the back-off counter of the station in post-backoff state reaches 0 but it has no frame, it
remains in post-backoff state (0, 0)e . A station in state (0, 0)e receives at least one frame with
probability ql during the current time slot. If it receives at least one frame during the current
time slot and the medium is sensed idle, it attempts to transmit a frame at the start of the
next time slot. In the case of a successful transmission, it makes a transition into one of post-
backoff states ((0, k )e , k = 0, . . . , W0 − 1) with probability 1 − ql , and it makes a transition into
one of backoff states with stage 0 ((0, k ), k = 0, . . . , W0 − 1) with probability ql . In the case of
a collision, it enters one of states with back-off stage 1. If a station in state (0, 0)e receives a
frame during the current time slot but the medium is sensed busy at the start of the next time
slot, it enters one of backoff-states with stage 0. More precisely,
ql (1 − pl ) Pidle
P[(0, 0)e |(0, 0)e ] = 1 − ql + ,
W0
P[(0, k )e |(0, 0)e ] = ql (1 − pl ) Pidle /W0 , for k > 0
P[(0, k )|(0, 0)e ] = ql (1 − Pidle )/W0 , for k ≥ 0
P[(1, k )|(0, 0)e ] = ql pl Pidle /W1 , for k ≥ 0.
3.2 Analysis of the Markov Chain

Figure 1 shows the state transition diagram of the Markov chain. Fortunately, the stationary
distribution of the Markov chain can be analytically obtained (see Appendix A). To show this,
let b(i, k ) denote the stationary probability of being in state (i, k ), and let b(i, k)e denote the
stationary probability of being in (i, k)e . We can show that the stationary distribution of the
state (0, 0)e , b(0, 0)e , is given through the following equation:

ql (W0 + 1) ql W0
1/b(0, 0)e = 1 − ql +
2(1 − r l ) 1 − (1 − ql )W0
+(1 − Pidle )(1 − rl ) − rl Pidle (1 − pl ))

pl q2l W0 Pidle (1 − pl )rl
+ −
2(1 − pl )(1 − rl ) 1 − (1 − ql )W0 ql

1 − pl − pl (2pl ) m − 1
× 1 + 2W0 , (2)
1 − 2pl
  

 

  



   

    









 
     
 










   
 








   
   
Fig. 1. State transition diagram.
where Pidle is the probability that the medium is idle when the station in state (0, 0)e attempts
to transfer a frame. Malone et al. assumed that Pidle = 1 − pl , and we use this assumption in
this article. We can explicitly obtain the stationary distribution of other states.
A station in state (k, 0) (0 ≤ k ≤ m) attempts to transmit a frame when the medium is idle at
the beginning of the next time slot. A station in state (0, 0)e also attempts transmission at the
beginning of the next time slot if (at least) one frame arrives during the current time slot. The
probability that station l attempts transmission, τl , is then given by
τl = ql Pidle b(0, 0)e + ∑ b(i, 0)

i ≥0

q2l W0 qrP
= b(0, 0)e − l l idle . (3)
(1 − pl )(1 − rl )(1 − (1 − ql )W0 ) 1 − rl
As shown in (2), the stationary distribution of each state contains unknown parameter pl , ql ,
and rl . If packets arrive at station l according to a Poisson process with mean rate λl , we can
estimate pl and ql through the following equations:
pl = 1 − ∏(1 − τj ),
j=l
   
ql = ∏(1 − τj ) (1 − e−λl Ts ) + 1 − ∏(1 − τj ) (1 − e−λl Tc ).

j j
(4)
To obtain rl , we assume that the station l can be modeled as a queue with infinite-buffer, and
observe that the mean inter-arrival time of packets at station l should be equal to the mean
frame-transmission interval of station l if the queue is stable. Now assume that a station
enters one of post-backoff or backoff states via absorbing state (0, 0) a after successful frame
transmission. (The sojourn time in (0, 0) a is assumed to be zero.) Note that the mean return
time to (0, 0) a is equal to the mean frame-transmission interval. With denoting Es the expected
time spent per state, it follows from the fact b(0, 0) a = τl (1 − pl ) that
Es Es
mean frame-transmission interval = = .
b(0, 0) a τl (1 − pl )
Since the mean inter-arrival time of packets is 1/λ,
1 Es
= , (5)
λ τl (1 − pl )
from which we obtain

ql (W0 + 1)(1 − Pidle ) TF − 1/λl q2l W0
rl = 1 − q + +
2 Es 1 − (1 − q)W0

q (W + 1)(1 − Pidle ) T − 1/λl
1−q+ l 0 + F (1 − p)qPidle , (6)
2 Es
where TF is the mean MAC-layer packet service time, which is defined as the time interval
between the instant that a packet reaches the head of the queue and the instant that the packet
is successfully transferred, and it is approximately represented by (8). If the right hand side of
(6) exceeds 1, we set rl = 1. Note that the right hand side of (6) exceeds 1 only when station l
is congested and thus the frame loss frequently occurs due to the buffer overflow at station l.
The expected time spent per state Es is given as follows:

Es = ∏(1 − τi ) Ts + 1 − ∏(1 − τi ) Tc ,
i i
where Ts is the length of time slot, and Tc is the expected time taken for a collision. In this
article, we assume that RTS/CTS is disenabled and thus
ACK + 2 × PHY DATA
Tc = + + SIFS + DIFS,
Rb Rd
where
– SIFS: SIFS duration
– DIFS: DIFS duration
– ACK: length of ACK frame (without physical header)
– PHY: length of physical header
– DATA: length of date frame (without physical header)
– Rb : basic rate
– Rd : data rate
Equations (3), (4), and (6) are simultaneous equations concerning p j , q j , r j , τj for j = 1, . . . , n
which can be numerically solved by iterative substitution.
The connection switching process is performed in the following steps. When detecting the
need for switching the current link based on its policy (e.g. signal strength, transmission rate,
missed beacon number, retransmission number, etc), a client looks up its preferred AP list
and selects one except its currently associated AP in the same station as a target AP. Then, it
sets up its interface to the desired channel and PHY type, transmits an AuthenticationRequest
frame to the target AP, and waits for an Authentication Response during
MinChannelTime ,a
Basic rate 1Mbps

Date rate 11Mbps
PHY header 192bits
MAC header 288bits
ACK length 112bits + PHY header
SIFS 10µs
DIFS 50µs
Slot time 20µs
Table 1. DCF parameters used in the numerical examples
5. Numerical Experiments: Evaluation of the Admissible Limits of Voice Flows

5.1 Conditions of Numerical Experiments
In this section, we see the accuracy of the proposed analytical model by comparing numerical
analysis and computer simulation results. We used network simulation tools ns2 to obtain
simulation results. In the simulation, there were n mobile terminals in an IEEE 802.11b-based
wireless LAN. Each mobile terminal conducted a bidirectional voice conversation through the
AP with a node outside the WLAN, and thus there were n uplink and n downlink voice flows
under n mobile terminals. Each voice flow generated G.711-codec traffic; 200 byte packets
(160-byte data and 40-byte RTP/UDP/IP header) were generated every 20 ms in a voice flow.
The parameters of the DCF used in the numerical examples is depicted in Table 1. The buffer-
sizes of the AP and mobile terminals were all set at 30 in packet.
In the experiments, we evaluated the throughput and packet-loss ratio of each flow. We also
investigated how many voice conversations could be multiplexed in the wireless LAN with-
out having packet loss, which we refer to as the “multiplexable limit of voice conversations"
and denote by Nmax in this article. As mentioned in Section 1, the uplink/downlink unfair-
ness in WLANs makes the AP the performance bottleneck under the standard IEEE 802.11
DCF. The CW differentiation between the AP and mobile terminals would provision a fair
resource sharing between the uplink and downlink traffic. In the experiments, we investigate
how much the CW differentiation enhances the multiplexable limit of voice flows.
5.2 Results of Numerical Experiments

5.2.1 Throughput
We first evaluated the throughput of uplink and downlink voice flows when the contention
window parameters of all stations were set at (CWmin , CWmax ) = (31, 1023), which are the
default setting of the IEEE 802.11. Figure 2 compares analytical and simulation results con-
cerning the total throughputs of uplink flows as well as the total throughputs of downlink
flows. For reference, we also show the results evaluated by the analytical model of Malone
(Malone et al., 2007). The result was given in terms of application level throughput defined
by (7), where we exclude the lengths of PHY, MAC, IP, UDP, and RTP headers from the length
of data frame. The throughput estimated by our analytical model agrees well with simulation
results. Figure 2 shows that the uplink flows obtained larger throughput than the downlink,
indicating that the AP was the performance bottleneck.
Figure 3 shows the result when the contention window parameters of the AP were set at
(CWmin , CWmax ) = (7, 1023). Note that parameter setting (CWmin , CWmax ) = (7, 1023) gives
higher priority to the AP over mobile terminals and thus, under this parameter setting, the
unfairness between uplink and downlink flows should be improved. Actually, the differ-
ence between uplink and downlink flows in the total throughput became smaller than the
case when (CWmin , CWmax ) = (31, 1023). The throughput estimated by our analytical model
agrees well with simulation results when the number of mobile terminals was less than 13,
but some discrepancy was observed when the number of mobile terminals was larger than 15.
This discrepancy may come from (5) where we neglect the packet loss at the buffer of stations.
We also evaluated the throughput when the contention window parameters of the AP were set
at (CWmin , CWmax ) = (3, 7). The results are shown in Figure 4. In this parameter setting, the
downlink flows obtained larger throughput than the uplink, indicating that mobile terminals
were the performance bottleneck.
Simulation Our model Malone model

1.4
1.2
Throughput [Mbps]
1
Uplink
0.8
0.6
0.4
Downlink
0.2
0
0 5 10 15 20 25 30
Number of Mobile Terminals
Fig. 2. Throughput versus the number of voice flows: (CWmin , CWmax ) = (31, 1023) at the AP.
Simulation Our model Malone model

1.2
1
Throughput [Mbps]
0.8
Uplink
0.6
Downlink
0.4
0.2
0
0 5 10 15 20 25 30
Fig. 3. Throughput versus the number of voice flows: (CWmin , CWmax ) = (7, 1023) at the AP.
Increasing the Time Connected to Already
Deployed 802.11 Wireless Networks while Traveling by Subway 467
From the two tables, we can observe that the combination of selective passive scan and
AuthScan achieves the lowest interrupted connection time. The delay obtained by selective
passive scan is 94.4% smaller than the one obtained by standard passive scan when
establishing terminated wireless connections. The delay obtained by AuthScan is 94.2%
smaller than the one obtained by active scan when switching wireless connection to the next
AP. The increased connected time will become larger in proportion to the number of
wireless connection establishment and switching while traving by subway. Besides the
aforementioned parameters, there are hardware induced delays such as interface setup time.
These delays are not considered in the previous analysis, because they vary from maker to
maker and, therefore, are unsettled. In fact, the real delay observed in experiments is even
larger than the values obtained in the analysis.
Fig. 6. Experimental setup
5. Implementation and Experiments

We have implemented a prototype of our proposed method on an IBM Thinkpad X31 (CPU
Pentium M 1.7GHz, 1GB RAM) with an Atheros AR5212-based wireless interface. It runs
Debian Linux 4.0 Etch with a 2.6.18-5 kernel, modified madwifi (MadWifi, [Online]) as the
wireless interface driver, and modified wpa_supplicant (Linux WPA/WPA2/IEEE 802.1X
Supplicant, [Online]) as the application software. We also implemented an active scan
enabled client6 to scan all 18 channels in an active manner for a comparison purpose. Since
the delay of selective passive scan in Testablish is probabilistically determined between 0 and
100 msec as shown in Section 3, our experiments focus on the delay to switch wireless
connection Tswitch of our prototype.
In order to evaluate the performance of our prototype in an actual network, we set up the
following experimental environment. We build six overlapping BSSs in an office7: AP0(11g,
ch.11), AP1(11g, ch.6), AP2(11g, ch.11), AP3(11g, ch.1), AP4(11g, ch.6), and AP5(11g, ch.1), as
described in Figure 6. The APs and a target host (IBM Thinkpad X31) are connected by 100
6Because of regulatory reasons, wpa_supplicant is implemented to passively scan channel

12, 13, 14, 34, 38, 42 and 46.
7There are 19 802.11 a/b/g APs on seven channels sharing the same medium with our
experimental APs.
↓ ↑
The unbalance between Nad and Nad when (CWmin , CWmax ) = (31, 1023) comes from the
uplink/downlink unfairness in WLANs. The table shows that the discrepancy was resolved
as the congestion window parameters of the AP became smaller. The multiplexable limit
of voice conversation, however, did not increase so much even when the uplink/downlink
unfairness was improved.
5.2.3 Packet Loss Ratio

We also evaluated the packet loss ratios of uplink and downlink voice flows by our analytical
model and simulation. Results were depicted in Figure 5 when the contention window pa-
rameters of the AP were (CWmin , CWmax ) = (31, 1023), in Figure 6 when (CWmin , CWmax ) =
(7, 1023), and in Figure 7 when (CWmin , CWmax ) = (3, 7). These figures indicate that results
by our analytical model agree well with the simulation results. The discrepancy between ana-
lytical results and simulation may come from that assumption that the mobile terminals have
large buffer to temporarily keep frames, which is not satisfied in the setting of ns2.
simulation
0.8
our model
Packet Loss Ratio
0.6
Downlink
0.4
Uplink
0.2
0
0 5 10 15 20 25 30
Fig. 5. Packet loss ratio: CWmin=31, CWmax=1023.
simulation
0.8
our model
Packet Loss Ratio
0.6
Downlink
0.4
0.2
Uplink
0
0 5 10 15 20 25 30
Fig. 6. Packet loss ratio: CWmin=7, CWmax=1023.
Asymmetric carrier sense in heterogeneous medical networks environment 473
Asymmetric carrier sense in heterogeneous

medical networks environment
Bin Zhen, Huan-Bang Li, Shinsuke Hara† and Ryuji Kohno††
NationalInstituteofInformation
mmunications
andCo Technology,Hikarino- ,4-3 oka,
Yokosuka,,Japan
7480-932
† OsakaCityUniversity,Su
831- gimoto,Osaka,530-1Japan
† YokohamaNationalUniversity,
okiwadai,
5T-97 Yokohama,Japan ,1058- 42
Summary: Complementary WLAN and WPAN technologies, as well as other wireless

technologies will play a fundamental role to support ubiquitous healthcare delivery. This
chapter investigates energy based clear channel assessment (CCA) of IEEE WLAN (802.11b)
and WPAN (802.15.4b) system when they coexist in a close space. We derive closed-form
expressions of energy based, qualify the asymmetric CCA in both AWGN channel and
fading channels, and show the impact of noise uncertainty on CCA operation. In the
heterogeneous medical networks environment, WPAN is oversensitive to the 802.11b
signals and WLAN is insensitive to the 802.15.4b signals. The asymmetric CCA issue in
heterogeneous networks is different from the traditional “hidden node” or “exposed node”
issues in homogeneous network. Energy based CCA can effectively avoid possible packet
collisions when they are close within the “heterogeneous exclusive CCA range”. However,
beyond this range, WPAN can still sense 802.11b signals, but WLAN lose its sense to
802.15.4b signals. This leads to WPAN traffic in a position secondary to the WLAN traffic. A
two-band CCA scheme, with an additional CCA detector in auxiliary channel, is proposed
to combat the asymmetric CCA issue in the heterogeneous networks.
1. Introduction
Integration of heterogeneous wireless technologies is required to for revolutionary
healthcare delivery in hospital, small clinic, residential care center, and home [1-4]. The
medical environment is a diverse workspace, which encompasses everything from the
patient admission process, to examination, diagnosis, therapy, and management of all these
procedures. The concept of “wireless hospital” combines all medical, diagnostic and clinical
data together whenever needed through wireless integration [4]. There is desire to use IEEE
version of wireless local area networks (WLAN) and wireless personal area networks
(WPAN) technologies in the unlicensed industrial, scientific and medical (ISM) bands as a
common communication infrastructure [2, 3]. The WLAN technology is typically used for
office oriented applications and patient connection to the outside world, while the WPAN
and thus
W0 −1
W0 q
∑ b(0, k )e = b(0, 0)e . (19)
k =0 1 − (1 − q)W0
Next we consider the stationary probability of state (0, k ). The balance equation concerning
state (0, W0 − 1) yields
(1 − p )r q(1 − Pidle )
b(0, W0 − 1) = ∑ b(k, 0) + b(0, 0)e
k ≥0
W 0 W0

b(0, 0) p + b(0, 0)e qpPidle (1 − p)r q(1 − Pidle )
= b(0, 0) + + b(0, 0)e
1− p W0 W0

b(0, 0) + b(0, 0)e qpPidle (1 − p)r q(1 − Pidle )
= + b(0, 0)e
1− p W0 W0
r q
= b(0, 0) + b(0, 0)e {1 − (1 − pr ) Pidle } . (20)
W0 W0
It comes from the balance equation concerning state (0, k ) that for W0 − 1 > k ≥ 0
b(0, k ) = b(0, k + 1) + b(0, W0 − 1) + qb(0, k + 1)e

= b(0, k + 1) + b(0, W0 − 1) + b(0, W0 − 1)e (1 − (1 − q)W0 −k−1 )
W0 −1−k
= (W0 − k)b(0, W0 − 1) + b(0, W0 − 1)e ∑ {1 − (1 − q ) n }
n =1
1 − (1 − q)W0 −k
= (W0 − k)(b(0, W0 − 1) + b(0, W0 − 1)e ) − b(0, W0 − 1)e . (21)
q
Combining (14), (20), and (21) yields

W0 −1
W0 + 1 1 − r (1 − q)(1 − (1 − q)W0 )(1 − r )
∑ b(0, k) = b(0, 0) 2
−
q
+
q2 W0
k =0

q(W0 + 1) Pidle (1 − q)(1 − rp)(1 − (1 − q)W0 )
+ b(0, 0)e − (1 − rp) Pidle + (22).
2 qW0
By representing b(0, 0) in (22) in terms of b(0, 0)e through (18), we obtain

W0 −1
q qW0
∑ b(0, k ) = b(0, 0)e − Pidle (1 − rp)
k =0
1−r 1 − (1 − q)W0

W0 + 1 1 − r (1 − q)(1 − (1 − q)W0 )(1 − r )
× − +
2 q q2 W0

q(W0 + 1) Pidle (1 − q)(1 − rp)(1 − (1 − q)W0 )
+ b(0, 0)e − (1 − rp) Pidle +
2 qW0

qW0 q(W0 + 1)
= b(0, 0)e 1 − q − +
1 − (1 − q)W0 2(1 − r )

qW0
× + ( 1 − Pidle )( 1 − r ) − rPidle ( 1 − p ) .
(1 − (1 − q)W0 )
(23)
Since b(i, k) = (Wi − k)/Wi b(i, 0) for i > 0, it follows that

Wi −1
Wi + 1
∑ b(i, k) = b(i, 0) ,
k =0
2
from which we have

Wi −1 m ∞
Wi + 1 Wm + 1
∑ ∑ b(i, k ) = ∑ b(i, 0) 2
+ ∑ b(i, 0)
2
i =1 k =0 i =1 i = m +1

m ∞
b(1, 0) i −1 i i −1 m
= ∑p (W0 2 + 1) + ∑ p (W0 2 + 1)
2 i =1 i = m +1

b(1, 0) 1 2W0 (1 − (2p)m )
W (2p)m
= + + 0
2 1− p 1 − 2p 1− p

b(1, 0) 1 − p − p(2p)m−1
= 1 + 2W0 . (24)
2(1 − p ) 1 − 2p
It comes from (13) and (18) that

pq2 W0 P (1 − p )r
b(1, 0) = b(0, 0)e − idle . (25)
1−r 1 − (1 − q)W0 q
By substituting (19), (23), (24), and (25) into the normalization condition
Wi −1 Wi −1
∑ ∑ b(i, k) + ∑ b(0, k )e = 1,
i =0 k =0 k =0
we finally have

q(W0 + 1) qW0
1/b(0, 0)e = 1 − q + + ( 1 − Pidle )( 1 − r ) − rPidle ( 1 − p )
2(1 − r ) (1 − (1 − q)W0 )

pq2 W0 Pidle (1 − p)r 1 − p − p(2p)m−1
+ − 1 + 2W0 .(26)
2(1 − p)(1 − r ) 1 − (1 − q)W0 q 1 − 2p
Once we have obtained b(0, 0)e , the stationary probabilities of other states are easy to calculate.

Deployed 802.11 Wireless Networks while
Traveling by Subway
Jaeouk Ok, Pedro Morales, Masateru Minami and Hiroyuki Morikawa
TheUniversityofTokyo
Japan
1. Introduction
Recently, an increasing number of people retrieve various contents via the Internet using a
publish/subscribe service such as podcasts. Typical usage of those applications is to
subscribe to as many favorite sites as possible, and selectively enjoy the automatically
downloaded latest episodes. Newly available portable multimedia players enable people to
easily enjoy downloaded video/audio contents in a variety of places, and the added
communication functionality such as 3G or 802.11 (IEEE Standard 802.11, 1999) makes it
possible to retrieve the latest episodes as soon as they are published on the web.
Service like podcasts are especially beneficial to people traveling by subway1 considering
their idle time in a subway train. However, most of subway tunnels in Tokyo are
unfortunately covered by neither 3G nor 802.11 as of December 2008. Wireless connection
can be established when a subway train stays under coverage areas at a station2, but it will
be repeatedly interrupted each time the subway train passes through non-coverage areas in
the tunnels while traveling along a railroad. This interruption limits the time under
coverage areas, which reduces the maximum possible connected time. This time is further
reduced by current implementation exploiting the intermittent connectivity poorly. In this
chapter, we focus on the 802.11 wireless connection management while traveling by subway
because of its higher throughput, lower subscription cost, and larger variety of 802.11-
enabled portable devices than 3G.
We aim to increase the time connected to already deployed 802.11 wireless networks for
podcast-like applications while traveling by subway in Tokyo. To understand the target
environment, we investigated the commercial 802.11 HOTSPOT networks (NTT
Communications HOTSPOT, [Online]) deployed in Tokyo Metro. One of the findings
1Tokyo Metro carries average 6.22 million passengers per day as of 2007 (Tokyo Metro,
[Online])
2In Tokyo, for example, approximately 97% of subway stations are densely covered by three
different service providers as of December 2008 (NTT Communications HOTSPOT; NTT

DoCoMo Mzone; NTT EAST FLET'S SPOT, [Online])
against the common belief regarding long distance mobility is that the main factor to the
diminishment of available connected time is link layer connection management, not IP layer
mobility support because of the deployed VLANs across the target networks. We propose
an optimized solution for this subway environment to increase the connected time by
reducing the following two types of delay. The one delay is experienced when establishing
the wireless connection after coming out of non-coverage area in the tunnel, and the other
when switching the wireless connection to the next AP while crossing overlapping coverage
areas at stations.
Our method reduces the establishment delay by building a chain that links the last AP in the
previous station before the tunnel with the first AP in the next station after the tunnel, called
border APs in this chapter. By referring to this chain when leaving a station, a client can
reduce the delay to establish the connection through the use of passive scan only on the
channel corresponding to the upcoming border AP. The switching delay is reduced by
2. Target Environment Investigation

In order to find out the factors that decrease the possible connected time, we investigate the
commercial 802.11 HOTSPOT network in Tokyo Metro. All experiments were performed
with a windows XP machine while walking at stations and moving by subway in November
2008. NetStumbler (NetStumbler, [Online]), Wireshark (Wireshark, [Online]), and built-in
Wireless Auto Configuration were used to monitor beacon frames, analyze IP packets and
manage 802.11 wireless network connections. The findings are classified by the points of our
interest: link layer handoff, IP mobility support, and restrictions on application layer. This
section discusses each of them in detail.
Fig. 2. Waseda (T4) station structure (Tokyo Metro, [Online]) There are six 802.11g APs
installed on channel 1, 6 and 11 to collaboratively cover the station including entrances,
ticket gates, platforms, etc.
2.1 Link Layer Handoff

To find the necessity of link layer handoff in 802.11 HOTSPOT3, we studied coverage areas
at stations by measuring the averaged SNR values of beacon frames at different locations in
10 stations 4 . In each station, multiple APs ranging from four to seven are installed to
collaboratively cover entrances, ticket gates, platforms, passages for transfer, etc. Figure 1
shows an example of the measured results in Waseda (T4) station on Tozai line, whose
structure is as shown in Figure 2. From the figure we observe that: 1) the coverage area of a
single AP is not large enough to cover the entire station, and 2) each location is under
coverage areas of multiple APs. Therefore, while a client moves around at stations, it is
necessary to switch the wireless connection across overlapping wireless coverage areas. For
example, let us assume that a subway train comes in from right in Figure 2. The clients in
the train get associated to AP2 according to Figure 1. For some clients staying in the train,
the established wireless connection will be switched from AP2 to AP3, when the train moves
to the left for the next station. For others getting off the subway train at Lower 1 on the
platform and walk out B1F Exit, the wireless connection will be switched from AP2 to AP1.
Unlike stations with overlapping coverage areas, there are non-coverage areas in each
tunnel between stations. To show non-coverage areas in tunnels, we measured the time
3In Tokyo Metro, NTT Communications HOTSPOT provides 173 subway stations among
179 with IEEE 802.11 b/g wireless access service. (Tokyo Metro, [Online])
4 Yoyogi-uehara (C1), Yoyogi-koen (C2), Meiji-jingumae (C3), Omote-sando (C4),
Takatanobaba (T3), Waseda (T4), Kagurazaka (T5), Kasumigaseki (M15), Ginza (M16),
Tokyo (M17), and Otemachi (M18) station
stamps of received beacon frames while moving by subway through seven stations from
Nihombashi (T10) to Waseda (T4) station on Tozai line. The measured results are shown in
Table 1. From the table we observe that: 1) there exist non-coverage areas in each tunnel,
and 2) the time length under non-coverage areas is approximately one third of the total
moving time on a railroad. This is explained by different speeds of subway when passing
above two areas: high speed within large non-coverage areas in the tunnels and low speed
within small coverage areas in the stations. Though there is large period of time spent under
coverage of 802.11 wireless networks even while traveling by subway, the wireless
connection is repeatedly interrupted due to non-coverage areas in the tunnel. Therefore, it is
necessary to establish disconnected wireless connection whenever the client enters the
following coverage areas.
Station Coverage Area Non-coverage Area

Nihombashi (T10) 83 sec 20 sec
Otemachi (T9) 66 sec 56 sec
Takebashi (T8) 69 sec 36 sec
Kudanshita (T7) 70 sec 29 sec
Iidabashi (T6) 81 sec 53 sec
Kagurazaka (T5) 81 sec 55 sec
Table 1. Non-coverage areas in the tunnels
2.2 IP Mobility Support

To find the necessity of IP mobility support, we studied IP network configuration by
dumping and analyzing IP packets. After a successful association using the proper ESSID
(i.e., 0033) and WEP key, a global IP address is assigned via Dynamic Host Configuration
Protocol (DHCP) (Droms, 1997). The DHCP server has multiple ranges of address pool. It is
possible to have various addresses assigned with the same client and AP at different times.
To show the range of DHCP address pool, we collected the obtained TCP/IP address
configuration with a single AP at different times in Yoyogi-uehara (C1) station on Chiyoda
line. Part of the collected information is shown in Table 2.
Host address Subnet Mask Default Gateway

210.162.9.65 255.255.254.0 210.162.9.1
211.0.159.54 255.255.254.0 211.0.159.1
61.127.100.37 255.255.254.0 61.127.100.1
Table 2. The range of DHCP address pool
Once an IP address is assigned in the beginning of a session, the same address is repeatedly
assigned not only from different APs in the same station, but also from the APs in other
stations by another DHCP request during the DHCP lease length. We confirmed this fact by
starting a session at Yoyogi-uehara (C1) station, and receiving the same IP address from two
different APs in the same station and also receiving it in other stations such as Yoyogi-koen
(C2), Meiji-jingumae (C3), Omote-sando (C4), etc.
From above two experiments, we observe that HOTSPOT implements Virtual Local Area
Network (VLAN) (IEEE Standard 802.1Q-2003, 2003) to accommodate multiple subnets in
HDFK SK\VLFDO /$1 DQG WKHUHIRUH ,3 OD\HU

PRELOLW\ VXSSRUW LV XQQHFHVVDU\ 7KH LQLWLDOO\
DVVLJQHG ,3 DGGUHVV GRHV QRW QHHG WR EH FKDQJHG DV ORQJ DV LW SHUIRUPV ’+&3 OHDV
UHQHZDOHYHU\OHDVHOHQJWKLH PLQXWHV,Q RUGHUWRSUHYHQW’+&3UHQHZDOIURPIDLOLQJ
LQ WKH QRQFRYHUDJH DUHDV LQ WKH WXQQHOV LW LV QHFHVVDU\ WR SHUIRUP ’+&3 UHQHZDO Z
VKRUWHU LQWHUYDO OHDVH OHQJWK WKH P WLPH
PD[LPX
VSHQW LQ WKH QRQFRYHUDJH DUHDV $V
VKRZQLQ7DEOHWKHWLPHVSHQWLQQRQFRYHUD JHDUHDVLQWKHWXQQHOVDUHPXFKVKRUWHUWKDQ
WKHOHDVHOHQJWKWKHUHIRUHZHFDQSHUIRUPWKH’+&3GLVFRYHU\RQO\RQFHLQWKHEHJLQQLQJ
RIWKHVHVVLRQDQGVNLSLWODWHURQE\UH
QHZLQJ’+&3ZLWKWKHDERYHVKRUWHULQWHUYDO
4GUVTKEVKQPUQP#RRNKECVKQP.C[GT
177&RPPXQLFDWLRQV+276327LPSOHPHQWVZHE EDVHGDXWKHQWLFDWLRQWRFRPSOHPHQWWKH
RSHQV\VWHPDXWKHQWLFDWLRQ’HVSLWHDVXFFH
VVIXO7&3,3DGGUHVVFRQILJXUDWLRQYLD’+&3
DFOLHQWLVXQDEOHWRVHQGDQGUHFHLYHGDWDWUD
IILFRXWVLGHWKHQHWZRUN7RJDLQ,QWHUQHWDFFHVV
RXWVLGH WKH QHWZRUN XVHUV DUH UHTXLUHG
HQWHU DWR
XVHUQDPH DQG SDVVZRUG LQ DQ
DXWKHQWLFDWLRQZHESDJHWRZKLFKWKHILUVWDWWH
PSWWRDFFHVVDQ\ZHESDJHDIWHUODXQFKLQJD
ZHEEURZVHULVDXWRPDWLFDOO\UHGLUHFWHG:HDO
VRIRXQGWKDWWKHZHEDXWKHQWLFDWLRQH[SLUH
ZKHQWKHUHLVQRWUDIILFVHQWRUUHFHLYHGGXUL
QJ’+&3OHDVHOHQJWKL
H PLQXWHV$JDLQ
DVWKHWLPHP
Usually, NP criteria is adopted in CCA because a miss detection of a busy channel is riskier
than a false alarm of a free channel. Eq. (7) can be re-written as
 Q 1 ( Pfa )  2 N  
Pd  Q . (11)
 1  2 
 
As expected, P f is independent of γ since there is no signal under H0. When the channel is
varying due to fading and shadowing, Eq. (12) gives a CCA performance conditioned on the
instantaneous SNR. The average CCA performance can be derived by averaging Eq. (11)
over fading statistics
  Q 1 ( Pfa )  2 N  
Pd   Q  f ( )d , (12)
0  1  2 
 
where f ( ) is the probability of distribution function (PDF) of SNR under fading.
The medium-scale variance of SNR can be characterized by log-normal distribution [22]. The
log-normal shadowing is usually described in-term of its dB-spread, dB, which is related to
by
   dB ln(10) / 10 . (13)
Under Rayleigh fading, the SNR γ has an exponential PDF
1 
f ( )  exp( )   0 , (14)
 
where  denotes to average SNR. If the SNR follows a Rician distribution, the PDF of γ
becomes
K 1 ( K  1) K ( K  1)
f ( )  exp(  K  ) I 0 (2 )   0, (15)
  
where K is the Rician factor and I 0 (.) is the modified Bessel function with order zero.
Because it is difficult to have close-form expressions of Eq. (12) over fading channels, we
evaluated them numerically in this chapter.
Figure 4 plots ROCs of energy based CCA over AWGN, log-normal shadowing, Rayleigh
fading, and Rician fading channels. The asymmetric CCA abilities of WPAN and WLAN
remain the same in fading channels. WLAN systems are insensitive to WPAN signals, while
WPAN systems are oversensitive to WLAN signals. Comparing with the AWGN curves, we
observe that channel fading degrades the performance of energy based CCA, and the
degradations are closely related with the CCA parameters and SNR. In other words,
meeting the desired performance demands a longer CCA window. Especially Rayleigh
fading and Rician fading degrade the CCA performance of all systems significantly.
0
10
(a)
Probability of detection
-1
10
15.4b CCA (15.4b signals)

15.4b CCA (11b signals)
11b CCA (15.4b signals)
11b CCA (11b signals)
-2
10
-3 -2 -1 0
10 10 10 10
Probability of false alarm
0
10
(b)
-1
10

-2
10
-3 -2 -1 0
10 10 10 10
The connection switching process is performed in the following steps. When detecting the
need for switching the current link based on its policy (e.g. signal strength, transmission rate,
missed beacon number, retransmission number, etc), a client looks up its preferred AP list
and selects one except its currently associated AP in the same station as a target AP. Then, it
sets up its interface to the desired channel and PHY type, transmits an AuthenticationRequest
frame to the target AP, and waits for an Authentication Response during
MinChannelTime ,a
The delay of AuthScan is as follows. Assuming at least one of the target APs is available and can
fulfill the handoff policy, it takes M * RTT + (N - M) * MinChannelTime , where N is the number of
target APs that exclude the associated AP before handoff in the preferred AP list at each station,
and M is the number of AuthenticationResponse frames received. The total connection switching
delay Tswitch will be composed of M * RTT + (N - M) * MinChannelTime due to the scanning
process, one RTT for authentication, and one RTT for association phase. This achieves channel
scanning with lowest delay among the approaches to work under subway's intermittent
connectivity without modifying deployed APs as shown in the next section.
4. Increased Connected Time Comparison

In this section, we introduce the related work and analyze the increased connected time by
our proposed method in a subway mobility scenario.
4.1 Limitations of Related Work

Many approaches have been proposed to address the channel scanning issues, and can be
classified into two groups as below. The first group tries to eliminate the necessity for the
channel scanning phase. Some approaches decouple the time-consuming channel scan from
the actual handoff phase and scan earlier for maintaining a list of candidate APs with their
handoff metrics before the connection to the current AP is terminated (Ramani & Savage,
2005; Wu et al., 2007). Other approaches enable scanning while communicating with the
currently associated AP utilizing multiple NICs (Brik et al, 2005; Ok et al, 2007). Though the
total handoff delay of these approaches in the first group is shorter than that of our proposed
method, none of the above related work can be applied in our subway environment. (Ramani
& Savage, 2005; Wu et al., 2007; Brik et al, 2005) can not generate a list of candidate APs with
their handoff metrics before the connection to the current AP is terminated under subway's
intermittent connectivity. (Ramani & Savage, 2005; Ok et al, 2007) require modification to APs,
and (Brik et al, 2005; Ok et al, 2007) require extra hardware on a client.
The second group tries to improve the efficiency of the channel scanning phase. The first
way to do this is to reduce the number of channels that are effectively going to be scanned
by the client. This can be achieved through various methods such as a cache (Shin et al.,
2004), Neighbor Graph (NG) (Mishra et al., 2004), sensor overlay network (Waharte et al.,
2004), etc. Another way to improve the efficiency is by reducing the time waiting at each
channel. In order to do this, a client is provided with an AP list, and only scans target APs in
a unicast fashion to reduce the time to wait at each channel (Ok et al., 2008; Kim et al., 2004;
Jeong et al., 2003; Huang et al., 2006). The performance of these approaches depend on the
number of channels to scan or deployed APs, but this is not an issue in our target
environment, where there are limited number of APs available. Among these, AuthScan
achieves the lowest handoff delay, one RTT less than selective unicast scan (Kim et al., 2004)
without modifying standard.
4.2 Performance Comparison

To show the increased connected time by our method, we estimate the interrupted time under
coverage areas (Testablish and Tswitch) by various methods. Firstly, we compare Testablish of passive
scan and selective passive scan, which do not generate excessive management traffic under
non-coverage area in the tunnel. Assuming that a client starts to scan at the benining of
coverage area, total delay to establish wireless connection of each method is as follows.
 Passive Scan: 185 * BeaconInterval + 2 * RTT
 Selective Passive Scan: 1 *BeaconInterval + 2 * RTT
For example, in the case of T4 station where a single border AP exists on channel 1, as
depicted in Figure 3, Testablish of above two methods are compared in Table 3, where RTT is
0.6 msec, beacon interval is 100 msec.
Method Testablish
Passive Scan 18 * 100 + 2 * 0.6 = 1801.2 msec
Selective Passive Scan 1 * 100 + 2 * 0.6 = 101.2 msec
Table 3. Connection establishment delay
Secondly, we compare Tswitch of active scan, selective active scan, selective unicast scan, and
AuthScan, which expedite scanning process by generating extra management traffic across
overlapping coverage area at station. Assuming that there are M target APs on N channels
and all of them response to requests, total delay to switch wireless connection of each
methods are as follows, where MaxChannelTime is a time parameter involved in active scan
long enough to guarantee the reception of the Probe Response frames from multiple APs
available in the same channel.
 Active Scan: MaxChannelTime * N + MinChannelTime * ( 18 - N ) + 2 * RTT

 Selective Active Scan: N * MaxChannelTime + 2 * RTT
 Selective Unicast Scan: M * RTT + 0 * MinChannelTime + 2 * RTT
 AuthScan: M * RTT + 0 * MinChannelTime + 1 * RTT
For example, in the case of T4 station where five target APs exist on three channels as
depicted in Figure 3, Tswitch of above four methods are compared in Table 4, where RTT is 0.6
msec, beacon interval is 100 msec, MaxChannelTime is 15 msec, and MinChannelTime is 1024
μsec (Jeong et al., 2003).
Method Tswitch
Active Scan 3*15+15*1.024+2*0.6=61.56 msec
Selective Active Scan 3 * 15 + 2 * 0.6 = 46.2 msec
Selective Unicast Scan 5 * 0.6 + 2 * 0.6 = 4.2 msec
AuthScan 5 * 0.6 + 1 * 0.6 = 3.6 msec
Table 4. Connection switching delay
5Additional 4 channels (52, 56, 60, and 64ch) in 5.3GHz (W53) and 11 channels (100, 104, 108,
112, 116, 120, 124, 128, 132, 136, and 140ch) in 5.6GHz (56W) were added to the conventional
4 channels in 5.2 GHz (52W) for 11a in 2005 and 2007, respectively. Therefore, the total
number of channels to scan sums up 33 channels. However, we focus on the conventional 18
channels, which most of devices in Japan support, in this chapter.
From the two tables, we can observe that the combination of selective passive scan and
AuthScan achieves the lowest interrupted connection time. The delay obtained by selective
passive scan is 94.4% smaller than the one obtained by standard passive scan when
establishing terminated wireless connections. The delay obtained by AuthScan is 94.2%
smaller than the one obtained by active scan when switching wireless connection to the next
AP. The increased connected time will become larger in proportion to the number of
wireless connection establishment and switching while traving by subway. Besides the
aforementioned parameters, there are hardware induced delays such as interface setup time.
These delays are not considered in the previous analysis, because they vary from maker to
maker and, therefore, are unsettled. In fact, the real delay observed in experiments is even
larger than the values obtained in the analysis.
Fig. 6. Experimental setup
5. Implementation and Experiments

We have implemented a prototype of our proposed method on an IBM Thinkpad X31 (CPU
Pentium M 1.7GHz, 1GB RAM) with an Atheros AR5212-based wireless interface. It runs
Debian Linux 4.0 Etch with a 2.6.18-5 kernel, modified madwifi (MadWifi, [Online]) as the
wireless interface driver, and modified wpa_supplicant (Linux WPA/WPA2/IEEE 802.1X
Supplicant, [Online]) as the application software. We also implemented an active scan
enabled client6 to scan all 18 channels in an active manner for a comparison purpose. Since
the delay of selective passive scan in Testablish is probabilistically determined between 0 and
100 msec as shown in Section 3, our experiments focus on the delay to switch wireless
connection Tswitch of our prototype.
In order to evaluate the performance of our prototype in an actual network, we set up the
following experimental environment. We build six overlapping BSSs in an office7: AP0(11g,
ch.11), AP1(11g, ch.6), AP2(11g, ch.11), AP3(11g, ch.1), AP4(11g, ch.6), and AP5(11g, ch.1), as
described in Figure 6. The APs and a target host (IBM Thinkpad X31) are connected by 100
6Because of regulatory reasons, wpa_supplicant is implemented to passively scan channel

12, 13, 14, 34, 38, 42 and 46.
7There are 19 802.11 a/b/g APs on seven channels sharing the same medium with our
experimental APs.
Base-T cable, and all APs are working as a bridge between the wireless and wired network
in link layer level under open system authentication.
We evaluate our prototype's performance by measuring 1) its average delay to switch
wireless connection on the application level, and 2) RTTs during handoff in comparison
with active scan in the following experiment scenario. A client with an IEEE 802.11 a/b/g
NIC is associated to AP0. The client moves towards AP5 while using pingcommand to
transmit ICMP Echo Request frames to the target host in the same subnet. We set the ICMP
frame size as 480 bytes, and the interval between frames as 10 msec. Then, we reduce the
transmission power of AP0, while increasing that of AP5 to emulate a mobility scenario in a
limited space. As the client gets closer to AP5, the degradation of signal strength from AP0
triggers handoff to AP5. Figure 7 shows the average delay to switch wireless connection
from ten runs of the handoff scenario since the sending of the Authenticationframe to
Request
the first AP scanned (2AQ in the graph). The x-axis shows the steps in the authentication
scanning process. They correspond to the sending of the Authenticationframe (AQ),
Request
reception of the Authentication Response frame (AS), sending of the
Reassociation Request
frame (RQ) and reception of the Reassociation Response frame (RS). The number before each
of them is the actual AP name being checked. The y-axis is the time delay in milliseconds
measured in the application in order to get the handoff delay from the user's perspective.
We added a checkpoint right before calling the driver ioctl , in the case of the AQ and RQ,
and right after receiving the driver's informational event for the AS and RS.
50
average delay
40
30
delay (msec)
20
10
0
2AQ 2AS 5AQ 5AS 3AQ 3AS 1AQ 1AS 4AQ 4AS 5RQ 5RS
handoff steps
Fig. 7. The average handoff delay checking five APs
The total handoff delay in the application level obtained when checking five APs with open
system authentication is 37.84 msec in average. This includes hardware induced delay such
as interface setup time, delay introduced by system calls and events that flow between
userland and kernel space, and 1 RTT for Authenticationand Authentication.Response
Request
It takes 3.60 msec in average to check APs on the same channel (from AP 0 to AP2, from AP5
to AP3, from AP1 to AP4), while it takes 8.46 msec in average to check APs on the different
channel(from AP 2 to AP5, from AP3 to AP1). Therefore, the AuthScan client saves 4.86
msec to setup the interface into the desired channel, each time checking APs on the same
channel consecutively in our experiment.
Figure 8 shows one example of the way RTT changes during handoff from AP0 to AP5. The
upper graph corresponds to one run by an AuthScan client, and the lower by an active scan client.
The x-axis shows the ICMP sequence number and the y-axis shows the RTT in milliseconds.
Handoff takes place between the 682nd and 686th frames (A) in the case of the AuthScan client
(i.e., 3 frames dropped, 30 msec disrupted), and between the 3637th and 3924th frames (B) for the
active scan client (i.e., 286 frames dropped, 2860 msec disrupted). Therefore, the AuthScan client
drops approximately 98.95% fewer frames than the active scan client in the experiment.
6. Conclusion
In order to increase the time connected to already deployed 802.11 wireless networks while
traveling by subway in Tokyo, we have developed a system equipped with two scanning
modes: 1) passively scanning on a selected channel, and 2) scanning with multiple open
authentication. Through analysis and experiments, we have shown that our method
increases the time connected to 802.11 wireless networks by establishing wireless connection
when coming out of non-coverage area in the tunnel and switching its wireless connection
across overlapping coverage area at station with less delay.
Fig. 8. Profile of RTTs during handoff from AP5 to AP0
The main contribution of this chapter is two-fold:
 We investigated the commercial 802.11 HOTSPOT wireless networks deployed in Tokyo

Metro, and clarified the main factor to the diminishment of available connected time.
 We proposed an optimized solution for the subway's intermittent connectivity
environment, and analyzed the increase connected time by our method. In addition,
we showed the effectiveness of our system through experiments in comparison with
standard active scan.
Our proposed method will work under similar subway 802.11 wireless network
environments in any other cities. Our future efforts will be oriented to build a more
sophisticated chain of border APs, and list of preferred APs. A chain of border APs
including interrupted time under non-coverage area in the tunnels can save power by
sleeping the interface before performing selective passive scan. A list of preferred APs built
per AP at each station can save unicast scanning time even further.
Multi-Agent Design for the Physical Layer of a Distributed Base Station Network 493
Multi-Agent Design for the Physical Layer

of a Distributed Base Station Network
Université Laval
Canada
1. Introduction
As wireless networks are becoming more omnipresent and pervasive, appropriate resource
allocation and organization becomes an increasingly pressing challenge. There exist on the
consumer market two important types of wireless network technologies. On the one hand,
cellular mobile networks are highly centralized and hierarchical. By contrast, wireless local
area networks (WLANs) are deployed in an ad-hoc unstructured manner, thus avoiding the
need for elaborate and costly planning. However, WLANs such as those falling under the
highly successful 802.11 standard do not manage interference effectively and tend to collapse
at high offered traffic loads. It can be seen that cellular and WLAN represent two radically
different approaches in radio resource management, characterized by different sets of advan-
tages and drawbacks.
The purpose of this chapter is to demonstrate the feasibility of a connection-oriented self-
organized wireless system which offers efficient radio resource management and provides
the best aspects of both cellular (reliable, connection-oriented operation even at high of-
fered loads) and WLANs (ad-hoc deployment and distributed intelligence). This is achieved
based on the multi-agent concept and local synergistic micro interaction (between neighbor-
ing transceivers) from which a global organization emerges.
The notion of Multiple Agent (MA) considered is of the “ant” variety, whereby small min-
imalist agents sense their environment and react to it in an interdependent manner. Social
insects and mostly ants or bees are the most cited biological examples. In the literature
such approaches have already been used to solve many combinatorial/optimization problems
(Beongku et al., 2003; Brueckner & Parunak, 2003; Muraleedharan & Osadciw, 2003).
This design philosophy differs from more traditional approaches which consist in postulating
criteria expressed by equations and models in order to formulate the problem in such a way
that an optimal solution is derived within the defined context. In the agent approach, precise
mathematical formulation of the problem is neither required nor very useful. The approach
thus becomes attractive for tackling complex multidimensional problems which would oth-
erwise be intractable. Therefore, our goal is not to demonstrate an optimal design, but to
illustrate how a Multi Agent System (MAS) can be empirically designed and fine-tuned to fit
a specific application. Moreover, it will be seen that such a dynamically adaptive solution, in
spite of its empirical nature, offers many advantages over a rigid analytically-derived coun-
terpart.
We will focus herein on Parunak’s methodology (Parunak, 1997) because it offers an intuitive
modeling framework, which is well suited to the empirical design approach.
Considering wireless networks, this chapter describes a flexible distributed base station (DBS)
framework which removes many limitations of current networks in order to augment the so-
lution space. For example, a plurality of DBS can simultaneously provide a network link
to the same mobile, thus leveraging macrodiversity to improve link quality and/or achieve
power savings. These DBS are designed with auto-organization in mind, such that the net-
work structures itself autonomously. This is where MAS come in, offering the desired dis-
tributed intelligence, adaptability, scalability and auto configuration properties. However, the
DBS architecture is challenging in at least three aspects:
1. It requires the continuously-updated solving of a large combinatorial problem, namely
finding a good allocation of DBS resources to mobiles requiring service.
2. Interference must be handled in a transparent way so that mobiles can gain the best
benefit of macrodiversity without being restrained by interfering mobiles.
3. Power control is an important aspect for both energy consumption and network capac-
ity given that it is tightly-coupled with interference patterns.
These three aspects are entangled together such that an optimal allocation is a complex combi-
natorial problem. It is NP hard unless some heavy simplifying assumptions are made (on the
geometry, on propagation, or other aspects). Moreover, in the context of mobility, an optimal
solution at one point in time is not optimal if it cannot easily adapt to changing parameters
(mobiles’ positions, fading, etc.).
Yet, this complex context is well suited for a MAS design. Indeed, MA need an active environ-
ment in which to generate interaction. And each event of allocating power, channel or connec-
tions to mobiles, that a DBS generate, has consequences on other mobiles’ links. This creates
the required active environment in which agents can sense parameters such as the received
power, interference and link quality, and where decisions can be made locally to generate new
actions. In turn, the effect of these actions are sensed by other agents. The next section de-
scribes the challenges of the proposed DBS architecture. Then, MA design concepts used in
this study are described. The fourth section details the proposed design of three categories of
interacting agents respectively for :
1. macrodiversity connection management,
2. channel allocation, and
3. power level control.
Finally, the system is emulated. Results, including simulation of complex cases with randomly
distributed DBS and mobile traffic, show first the resource allocation quality that can be ob-
tained, and second the effectiveness of MAS design in terms of auto-configuration/scalability
and dynamic adaptation properties.
A final brief discussion will extend Parunak’s agent design principles to summarize the
lessons learned from designing MAS for the application at hand.
Asymmetric carrier sense in heterogeneous

medical networks environment
Bin Zhen, Huan-Bang Li, Shinsuke Hara† and Ryuji Kohno††
NationalInstituteofInformation
mmunications
andCo Technology,Hikarino- ,4-3 oka,
Yokosuka,,Japan
7480-932
† OsakaCityUniversity,Su
831- gimoto,Osaka,530-1Japan
† YokohamaNationalUniversity,
okiwadai,
5T-97 Yokohama,Japan ,1058- 42
Summary: Complementary WLAN and WPAN technologies, as well as other wireless

technologies will play a fundamental role to support ubiquitous healthcare delivery. This
chapter investigates energy based clear channel assessment (CCA) of IEEE WLAN (802.11b)
and WPAN (802.15.4b) system when they coexist in a close space. We derive closed-form
expressions of energy based, qualify the asymmetric CCA in both AWGN channel and
fading channels, and show the impact of noise uncertainty on CCA operation. In the
heterogeneous medical networks environment, WPAN is oversensitive to the 802.11b
signals and WLAN is insensitive to the 802.15.4b signals. The asymmetric CCA issue in
heterogeneous networks is different from the traditional “hidden node” or “exposed node”
issues in homogeneous network. Energy based CCA can effectively avoid possible packet
collisions when they are close within the “heterogeneous exclusive CCA range”. However,
beyond this range, WPAN can still sense 802.11b signals, but WLAN lose its sense to
802.15.4b signals. This leads to WPAN traffic in a position secondary to the WLAN traffic. A
two-band CCA scheme, with an additional CCA detector in auxiliary channel, is proposed
to combat the asymmetric CCA issue in the heterogeneous networks.
1. Introduction
Integration of heterogeneous wireless technologies is required to for revolutionary
healthcare delivery in hospital, small clinic, residential care center, and home [1-4]. The
medical environment is a diverse workspace, which encompasses everything from the
patient admission process, to examination, diagnosis, therapy, and management of all these
procedures. The concept of “wireless hospital” combines all medical, diagnostic and clinical
data together whenever needed through wireless integration [4]. There is desire to use IEEE
version of wireless local area networks (WLAN) and wireless personal area networks
(WPAN) technologies in the unlicensed industrial, scientific and medical (ISM) bands as a
common communication infrastructure [2, 3]. The WLAN technology is typically used for
office oriented applications and patient connection to the outside world, while the WPAN
technology is usually used for wearable sensors around patients to collect vital information
for ubiquitous healthcare service [2-6].
The use of complementary heterogeneous WLANs and WPANs in the shared ISM band
results in coexistence, interference and spectrum utilization issues. The coexistence of
wireless technologies in ISM band has been a hot topic [6-9]. Adaptive frequency hopping
was proposed for Bluetooth devices to avoid interference from WLAN [7]. A model for
analyzing the effect of 802.15.4 on 802.11b performance was provided by Howitt and
Gutierrez [8]. The degradation of WLAN performance is small given that the WPAN activity
is low. However, the high duty cycle of WLAN traffic can drastically affect the WPAN
performance [6]. A distributed adaptation strategy for WPAN based on Q-learning has been
proposed to minimize the impact of interference from 802.11b [9]. However, the spatial reuse
issue in the heterogeneous networks has not drawn much attention. Some researches have
shown that the spatial reuse and aggregate throughput in the homogeneous WLAN mesh
network is closely related to physical channel sensing. Yang and Vaidya showed that the
aggregate throughput can suffer significant loss with an inappropriate choice of carrier
sense threshold [10]. Ma etal , by means of Markov chain model, evaluated how carrier sense
threshold affects the throughput and packet collision [11]. Zhai and Fang found that the
optimal carrier sensing threshold for one-hop flows does not work for multihop flows [12].
Zhu etal reported that a tunable sensing threshold can effectively leverage the spatial reuse
and demonstrated it through testbed measurement [13, 14]. In [15], Zhu et al proposed a
heuristic algorithm to adaptively tune the threshold of carrier sense to enhance throughput
per user. Jamieson found carrier sense can be inefficient at low data rate when capture effect
is most prevalent [16]. Ramachandran and Roy showed cross-layer dependence between
carrier sense and system performance [17]. Simulators widely used for performance
evaluation, like NS-2 and OPNET, do not contain detailed physical layer module like carrier
sense. For the lack of carrier sensing knowledge between WPAN and WLAN, Golmie et al
simply simulated two carrier sensing cases: the WPAN can only detect packets of its own
type and the WPAN can also detect WLAN’s transmission, in their coexistence study for
medical applications [6].
In this chapter we study the coexistence issue in the heterogeneous medical networks
environment from carrier sensing point. The remainder of the chapter is organized as
follows. Section II briefly reviews various carrier sense methods and the considered WLAN
and WPAN systems. In section III, a mathematical analysis of energy based carrier sense in
both AWGN channel and fading channel is presented. Section IV describes the impact of
asymmetric carrier sense in heterogeneous networks environment and presents a two-band
carrier sense to combat the asymmetry. Section V finally concludes the chapter.
2. Review of systems and clear channel assessment

A.Wirelessmedicalsensornetworks
Wireless medical sensor networks can be considered as a special part of general wireless
sensor networks (WSN), which are mainly implemented by low-rate WPAN technologies.
As compared in Table I, both share some common features which include limited resources
(e.g. computation power, memory, battery, bandwidth), low/modest duty cycle, energy
efficiency, plug-and-play, diverse coexistence environments, and heterogeneous device
ability. But we can also find significant differences between them in the sensor device,
dependability, networking, traffic pattern and channel.
Firstly medical sensors consider safety, quality and reliability as top priority, while general
WSN are cost sensitive for market reason. The safety to human/animal body is therefore the
first factor taken into considered. Thus medical sensors must be conscious of specific
absorption ratio (SAR) to protect human tissue. Wearable IEEE WPAN devices are
suggested to be separated at least 30cm distance from human body. Safe to human is the top
priority of medical sensors. The radio emission should be as weak as possible. And medical
sensors should be lightweight and small to achieve non-invasive and unobtrusive monitor.
This limits the available resource which includes memory, battery power and computation
ability in the medical sensors. The requirement is more stringent than the general WSN.
Secondly the medical sensor networks have more frequency bands to select than general
WSN, which usually work in ISM band. Although the specific medical bands are less noisy,
they are narrow band and conditional license. For example, the wireless medical telemetry
service (WMTS) band can only be used in the licensed hospital and clinic, but not at home.
On the contrary, the wideband ubiquitous ISM is somehow noisy since the frequency band
should be shared with other systems.
Thirdly, the traffic pattern in medical sensor networks is featured by periodical real time
data (e.g. EEG and ECG) and some top priority burst data (e.g. alarm and alert) [17]. In
contrast, general WSN typically consider versatile traffic. The medical information,
especially the alarm notification, have very strict requirement in terms of Quality-of-Service
(QoS) since they are life critical. The transmission of vital signal has a life or dead meaning.
This means more stringent QoS requirement than general WSN.
Fourthly, security of data is traditionally utmost important. Patient data needs to be
protected in all stages of data acquiring, data transmission and data storage. It therefore
importance to secure data at physical layer and MAC lay. However, security is not free,
extensive sources are needed to secure data at the link layer. In the resource limited WSN,
security becomes an overhead of existing network QoS. Because both of them are
paramount in the healthcare service, the balance of security and QoS is a new issue. The
general WSN do not require strict QoS and security simultaneously.
Fifthly, to improve reliability, general WSN tend to distribute redundant sensors as backup
for sensing, transmission and forwarding. In contrast, there is little redundancy in medical
WSN for medical reasons. For example, vital signals, like EEG (Electroencephalography)
and ECG (Electrocardiogram), are location dependent and can only be measured by
deterministic location. Therefore it is difficult to allocate redundant sensors in the limited
area. Especially, it makes no sense to allocate sensors outside of the interest/effect area.
In summary, the lack of redundancy, priority traffic, dominant periodical data and balance
of guaranteed QoS and security in versatile coexistence environment challenge the
reliability design of wireless medical sensor networks.
Medical wireless sensor General wireless sensor networks

networks
Common Limited resources: battery, computation, memory, energy efficiency
features Diversity coexistence environment
low/modest data rate, low/modest duty cycle
Dynamic network scale, plug-and-play, heterogeneous devices ability,
dense distribution
Sensor/ Single-function device Multi-function device
actuator Fast relative movement in Rare or slow movement in large range
small range
device lifetime, network lifetime and device lifetime,
days, <10 years (implant months, <10 years
sensor)
Safe (low SAR) and quality first Cost sensitive
Dependability Reliability (first), guaranteed expected QoS, redundancy-based
QoS reliability
Strongly security (except Required security
emergency)
Networking Small scale star network Large scale hierarchical network
No redundancy in device redundant distribution
Deterministic node distribution Random node distribution
Traffic Periodical RT (dominant), Burst (dominant), periodical
burst (priority)
Uni-directional traffic Uni-directional or bi-directional traffic
M:1 communication M:1 or point-point communication
channel Specific medical channel, ISM ISM band
band
Body surface or through body Obstacle is unknown
Table 1. Comparison between wireless medical sensor network and general wireless sensor
networks
B.Systemsoverview
We consider IEEE 802.15.4b and 802.11b as examples for the ubiquitous medical services [2, 3,
6]. The former is a good candidate technology for low data rate and low cost medical
sensors. Both systems operate in the unlicensed 2.4GHz ISM band, and both are based on
carrier sense multiple access with collision avoidance (CSMA/CA) protocol. Carrier sense is
more generally known as clear channel assessment (CCA) in the standards. The physical
CCA can be either energy based, or feature based, or a combination of two. As shown in Fig.
1, there are only 4 WPAN channels locate in the guard bands of WLAN. Table II lists the
parameters of both systems [18, 19]. The bit error rate (BER) of WLAN systems with AWGN
channel is give by [18]
128 6
BER11b    M 1l  Q ( M 2l  SNR ) , (1)
255 l 1
where SNR is the signal-to-noise ratio, Q(.) is Gaussian Q-function, M1=[24167 241] ,
and M2=[410246]
8 . The BER of WPAN systems is given by [19]
8 1 16 k 16 20 SNR (1/ k 1))

BER15.4b      1 ( )e . (2)
15 16 k 2 k
The radio path loss for indoor channels in the working frequency band is given by
 pl  40.2  20 log10 d d 8
 , (3)
 pl  58.5  33 log10 (d / 8) d  8
where d is separation distance between transmitter and receiver.
802.11b channels
25 MHz 22 MHz
1 6 11
2400 MHz 2483.5 MHz
802.15.4b channels
5 MHz 2 MHz
11 12 ….. ….. 26
2400 MHz 2483.5 MHz

Fig. 1. Channel plan for IEEE 802.11b and 802.15.4b in 2.4GHz ISM band
Parameters 802.15.4b 802.11b
Transmission power (dBm) 0 16
Channel bandwidth (MHz) 2 22
Adjacent channel separation (MHz)1 5 25
Background noise (dBm) 2 -94.9 -84.6
Spread code (chips) 32/4 bits 11
Data rate (Mbps) 0.25 11
CCA window (µs) 120 15
Table 2. System parameters of IEEE WLAN and WPAN
C.Clearchannelassessment
Several IEEE WLAN and WPAN standards adopt CSMA protocol for channel access. CCA
is a physical layer activity and is an essential element of the CSMA protocol. The concept of
CCA was first proposed as an enhancement to the ALOHA protocol. The CCA detects an
incoming packet and ensure a free medium before transmission. The CCA module
processes received radio signals in a suitable time duration termed CCA window. It then
reports the medium state, either busy or idle, by comparing the detection with a threshold.
The energy based CCA integrates signal strength from radio front end during the CCA
window. The feature based CCA looks for the known features, e.g. the modulation and
spreading characteristics, of the signal over the channel. Modulated signals are in general
coupled with sine wave carriers, pulse trains, repeating spreading, or cyclic prefixes, which
result in built-in periodicity. This periodicity can be used to detect signal of a particular
modulation type.
The feature based CCA performs far better than the energy based CCA. However, a prior
knowledge of the signal characteristic is necessary. And the CCA module would need a
dedicated detector for every potential coexistence signal class. The main advantages of
energy based CCA are its simplicity, generality, and low power consumption. It is a
universal mechanism that can be deployed in all systems. Unlike feature based CCA, there
is no need for waiting time for the specific features of the signal and synchronization [17].
The downside of energy based CCA is that it is prone to false detection.
We applied energy based CCA to 802.15.4b and 802.11b systems to deliver ubiquitous
healthcare service in this heterogeneous networks environment. There are several reasons
for this. First, there have been nearly 10 wireless technologies with different modulations,
band plans, and transmission powers in the 2.4GHz ISM bands due to its global availability.
1Distance between the central frequencies of non-overlapped adjacent channel.

2We assumed -174dBm/MHz thermal noise, 8dB implementation losses, and 8dB radio
noise figure.
Note that the error floor depends on the number of symbol chips and the SNR. When
SNR<<1, Eq. (9) can be approximated as
 
PCCA _ ef  Q N  . (10)
 2
A linear decrease in SNR requires a quadratic increase in N to maintain the same error floor.
A. Asymmetric energy based CCA
Table III lists the numbers of signal chips in the CCA windows of IEEE WLAN and WPAN.
Figure 3 shows the CCA error floor in the heterogeneous networks environment when the
noise is known. Per Eq. (9), the error floors decrease with increment in signal chips in the
CCA window. Given the defined CCA windows, the CCA abilities, e.g. sensitivity and
range, to determine the channel state are different, this is termed asymmetric CCA. Under
the same SNR conditions, the lowest error floor is where WPAN is used to sense 802.11b
signals; the highest error floor is where WLAN is used to sense 802.15.4b signals. The
performance difference is nearly 10dB.
The CCA asymmetry can be attributed to differences in the underlying signals over channel
(power, symbol rate and background noise) and CCA window. In physics, a higher data
rate and a longer CCA window means more signal pulses in baseband can be collected.
Better CCA performance is a natural result. Asymmetric CCA can be further reinforced by
other factors. For example, the difference in transmission powers which is usually stronger
for WLAN, and the difference in channel bandwidth, which are 22MHz and 2MHz for the
WLAN and WPAN, respectively. For both WPAN and WLAN, the performances to detect
the signal of its own type are similar. There is not big difference in the numbers of symbols
in the CCA window.
Table IV compares communication with CCA when both have an error probability of 1‰
with AWGN channel. As expected, the CCA range is larger than the communication range.
For WPAN, sensing 802.11b signals has 4dB greater link margin compared to sensing the
signals of its own type. In contrast, for WLAN sensing 802.15.4b signals requires a 4.8dB
higher SNR.
Asymmetric CCA makes channel sensing insensitive or oversensitive to other signals in the
mixed WLAN and WPAN environment. The asymmetric CCA in the heterogeneous
networks is different from the traditional “hidden node” or “exposed node” issues in the
homogeneous network. In the homogeneous network, two devices belong to the same
system are reciprocal in ability to sense each other (we do not consider the minor difference
due to implementation.). However, in the heterogeneous networks, the sensing abilities of
different systems are unequal and depend on the underlying signals over channel and the
separation distances. As shown in Fig. 3, WLAN signals are well sensed by both of them,
but WPAN signals could be ignored by the WLAN systems when they are separated by
enough space.
Sensed signals 802.15.4b 802.11b

signals signals
Device
802.15.4b 32*8 120*11
802.11b 15*2 15*11
Table 3. Number of signal chips in the CCA window
Devices 802.15.4b 802.11b
Signal
Communication -0.8 5.6
802.15.4b CCA -3.2 2.6
802.11b CCA -7.2 -2.2
Table 4. SNRs (dB) to achieve 1‰ communication BER and CCA error floors in AWGN channel
0
10
-1
10
Energy based CCA error floor
-2
10
-3
10
-4
10 15.4b CCA (11b signals)
-20 -15 -10 -5 0 5

EcN0 (dB)
Fig. 3. Error floor of energy based CCA with AWGN channel in heterogeneous networks
Usually, NP criteria is adopted in CCA because a miss detection of a busy channel is riskier
than a false alarm of a free channel. Eq. (7) can be re-written as
 Q 1 ( Pfa )  2 N  
Pd  Q . (11)
 1  2 
 
As expected, P f is independent of γ since there is no signal under H0. When the channel is
varying due to fading and shadowing, Eq. (12) gives a CCA performance conditioned on the
instantaneous SNR. The average CCA performance can be derived by averaging Eq. (11)
over fading statistics
  Q 1 ( Pfa )  2 N  
Pd   Q  f ( )d , (12)
0  1  2 
 
where f ( ) is the probability of distribution function (PDF) of SNR under fading.
The medium-scale variance of SNR can be characterized by log-normal distribution [22]. The
log-normal shadowing is usually described in-term of its dB-spread, dB, which is related to
by
   dB ln(10) / 10 . (13)
Under Rayleigh fading, the SNR γ has an exponential PDF
1 
f ( )  exp( )   0 , (14)
 
where  denotes to average SNR. If the SNR follows a Rician distribution, the PDF of γ
becomes
K 1 ( K  1) K ( K  1)
f ( )  exp(  K  ) I 0 (2 )   0, (15)
  
where K is the Rician factor and I 0 (.) is the modified Bessel function with order zero.
Because it is difficult to have close-form expressions of Eq. (12) over fading channels, we
evaluated them numerically in this chapter.
Figure 4 plots ROCs of energy based CCA over AWGN, log-normal shadowing, Rayleigh
fading, and Rician fading channels. The asymmetric CCA abilities of WPAN and WLAN
remain the same in fading channels. WLAN systems are insensitive to WPAN signals, while
WPAN systems are oversensitive to WLAN signals. Comparing with the AWGN curves, we
observe that channel fading degrades the performance of energy based CCA, and the
degradations are closely related with the CCA parameters and SNR. In other words,
meeting the desired performance demands a longer CCA window. Especially Rayleigh
fading and Rician fading degrade the CCA performance of all systems significantly.
0
10
(a)
-1
10

-2
10
-3 -2 -1 0
10 10 10 10
0
10
(b)
-1
10

-2
10
-3 -2 -1 0
10 10 10 10
-3
10
The background noise power fluctuates from time to time due to changes of environment
and mobility of device. Another source of uncertainty is the error in quantization which us
usually implemented by A/D convertor. Assume the noise estimation is expressed as

 2  k 2 , k  0 . (17)
When 0  k  1 , the noise is underestimated; when k  1 , it is overestimated. The Γ

biases to the desired value due to error in noise estimation. An overestimation of noise
decreases P fa at expense of boosting P d. It is viceversa
for an underestimation of noise level.
We can obtain the total CCA error by 1-P dP+ fa using Eq. (7). Figure 5 shows the impact of
noise uncertainty with 3 dB error. We used an 802.15.4 device to sense the 802.15.4 signals.
The x-axis is the true SNR condition of CCA. As expected, both overestimation and
underestimation deteriorate the CCA performance because of an un-optimal threshold. As
shown in Fig. 5 and other numeric results, the performance loss of energy based CCA can
 10 * lg(k ) 0  k  1
be approximated as  in both cases when SNR is high. In
 10 * lg(k ) k 1
practical, we are usually more interested in noise underestimation. In order to guarantee the
P fa the Γ is purposely biased. This increases the probability of miss detection of CCA and
therefore the probability of packet collision. In other words, the noise level estimation of
free channel also plays an important role in the CCA operation.
0
10
-1
10
-2
10
CCA error floor
-3
10
underestimation -3 (dB)
-4 0 (dB)
10
overestimation 3 (dB)
-20 -15 -10 -5 0 5

EcN0 (dB)
Fig. 5. Total error of energy based CCA with AWGN channel in the case of a 3dB noise
uncertainty
4. Two-band clear channel assessment

A.ImpactofasymmetricCCAinheterogeneousnetworks
Table V lists the required minimum SNRs and their corresponding distances to achieve
reliable energy based CCA (P FA <1% and P D >90%) over AWGN channel. The corresponding
distances were computed using Eq. 1 to Eq. 3 and the parameters listed in Table I. For WPAN,
the sensing of 802.11b signals is reliable at an SNR as low as -9.25dB. This SNR is 9.65dB lower
than the critical SNR which is the least SNR to achieve BER<0.1‰ for communication. The
CCA range is 180 meters longer than the communication range. In contrast, sensing 802.15.4b
signal by WLAN requires a high SNR up to 9.75dB, which is 3.15dB more than the critical SNR
for communication. The CCA range is 42 meters shorter than the communication range. In the
fading channels, all distances decreases depending on the fading condition.
We can define a “heterogeneous exclusive CCA range” (HECR), in which systems in the
heterogeneous environment can reliably sense the activities of each other. In the considered
scenario, the HECR is the maximum distance that WLAN can sense 802.15.4b signals. Given
the system parameters and assumptions, the HECR for IEEE WLAN and WPAN is 25m in
AWGN channel. Peaceful and fair coexistence between them can be expected when they are
located within the HECR. However, it becomes different when they are separated beyond
the HECR. For WPAN systems, the CCA range of WLAN signals is more than twice as long
as the communication range. And it is longer than the CCA range of its own signal type.
That is, the WPAN is oversensitive to the WLAN signals. It can even sense a WLAN packet
that is outside of the keep-out range of receiver in the worst case. 3 Although the
oversensitive CCA avoids the ‘hidden node’ issue, it suffers from the ‘exposed node’ issue.
This results in poor spatial reuse of frequency channels and low aggregation throughput
since WPAN sometimes unnecessarily withdraw packet before transmission. As simulated
in [11], the threshold optimized to maximize aggregate throughput is higher than the
optimal threshold for a single hop. For WLAN systems, the CCA range of WPAN signals is
about a quarter of the communication range. Packet collision may occur when WLAN traffic
occurs immediately after the WPAN traffic.
802.11b CCA 802.15.4b CCA
WPAN signals WLAN signals WPAN signals WLAN

signals
SNR (dB) 9.75 -4.5 -6 -9.25
Distance 25 200 155 280

(m)
Table 5. Minimum SNRs (dB) and corresponding distances (m) to achieve CCA (P FA <1%,
P D >90%) with AWGN channel
3 The keep-out range denotes to the minimum separation which WPAN and WLAN do not
interfere each other.
Although the HECR of 25 meters is not sufficient for outdoor applications, it seems to be
good enough for most indoor applications. Typical bedside medical applications define by
IEEE 1073 are within this range [24]. This is different from those of most coexistence studies
in which it is usually assumed that WLAN cannot sense the activities of WPAN [6, 8, 9]. The
HECR over fading channels is expected to be shorter than 25 meters depending on the
fading parameters. But the asymmetric CCA issue still exists. Putting the oversensitive and
insensitive CCAs together results in an unfair share of channel between WLAN and WPAN
when they are separated beyond HECR. There is a preferential treatment of WLAN traffic.
The WLAN is over-protected, while the WPAN is vulnerable.
B.Two-bandclearchannelassessment
The asymmetric CCA issue in heterogeneous medical networks environment must be

solved. It is needed to recognize the signal source, WPAN signals or WLAN signals, when
the medium is busy. When the channel is occupied by WLAN signals, an 802.15.4b device
should increase its CCA threshold to lower the CCA sensitivity. On the other hands, when
the channel is busy in WPAN signals, an 802.11b device should lower its CCA threshold to
heighten the CCA sensitivity.
Figure 6 illustrates the mechanism of a two-band CCA. The 22 MHz WLAN channel depicted
in dash-dotted line and the 2 MHz WPAN channel depicted in dotted line overlap each other.
There are about 3 MHz space between adjacent WPAN channels, which are inside the WLAN
channel. As shown in red-solid line in Fig. 7, we termed it auxiliary channel in this chapter.
We added a new energy based CCA detector in the auxiliary channel to provide additional
information. For example, the auxiliary CCA detector can be the same as the 802.15.4b CCA
detector except that the working frequency is tuned to the auxiliary channel. For both WLAN
systems and WPAN systems, there are two energy based CCA detectors tuned in its original
channel and in the auxiliary channel, respectively. The two detectors have the same ability to
conduct carrier sensing given the same configuration. Besides, there is a little increase in the
complexity of device and the power consumption of CCA.
The two-band CCA in the WPAN/WLAN device conduct channel sensing simultaneously.
When the channel is free, both CCA detectors indicate a free channel. Table VI lists the
output states of the two-band CCA detector when the channel is busy. The channel state
indications are the same in both systems. When the channel is occupied by WLAN signals,
both CCA detectors indicate a busy channel. In contrast, when the channel is occupied by
WPAN signals, the auxiliary CCA detector indicates a free channel. Therefore, the signal
source can be easily distinguished.
In the two-band CCA, the performance of every energy detector can be analyzed as in
Section III. The total performance can be expected as the AND of the two CCA detector.
WLAN channel
WPAN channel 5 MHz
Auxiliary channel Frequency
Fig. 6. Mechanism of two-band clear channel assessment
802.11b 802.15.4b
CCAch_11b CCAch_aux CCAch_15.4 CCAch_aux

b
WPAN signals √ × √ ×
WLAN signals √ √ √ √
Table 6. Output states of the two-band CCA detector when the channel is busy
5. Conclusion
In this chapter, we have investigated the coexistence issue in the heterogeneous medical
networks environment for ubiquitous health services from network access point. The
energy based CCA was considered because the 2.4GHz ISM band is too crowded to apply
feature based CCA for simple medical sensors.
Using central limit theorem, we have derived closed-form expressions for energy based
CCA. We have shown and qualified impact of noise uncertainty on CCA and the
asymmetric CCA in AWGN channel and fading channels. In the considered heterogeneous
medical networks environment for ubiquitous healthcare purposes, WPAN is oversensitive
to 802.11b signals and WLAN is insensitive to 802.15.4b signals. When WPAN and WLAN
are located within the HERC, energy based CCA can effectively avoid possible packet
collisions. The HERC is sufficient for most indoor medical applications. However, when
they are farther apart, WLAN lose its sense to 802.15.4b signals. The asymmetric CCA puts
WPAN traffic in a secondary position in the heterogeneous networks. We have proposed a
two-band energy-based CCA with an additional CCA detector tuned in the auxiliary
channel to combat the asymmetric CCA issue.
6. References
R. Istepanian, E. Jovano and Y Zhang, “Guest editorial introduction to the special section on
M-Health: beyond seamless mobility and global wireless healthcare connectivity,”
IEEETrans.InformationTechnologyinBiomedicine , vol.8, no.4, p.405-414, 2004.
A. Soomro and D. Cavalcanti, “Opportunities and challenges in using WPAN and WLAN
technologies in medical environments,” IEEECommunicationMagazine , vol.45, no.2,
p.114-122, 2007.
D. Cypher, N. Chevrollier, N. Montavont, and N. Golmie, “Prevailing over wires in
healthcare environments: benefits and challenges,” IEEE Communication Magazine ,
vol. 44, no. 4, p. 56-63, 2006.
http://www.wilho.net/
A. Lymberis “Smart wearable systems for personalized health management: current R&D
and future challenges,” Inter.Conf.ofIEEE, vol.4, p.3716-3719, 2003.
N. Golmie, D. Cypher, and O. Rebala, “Performance analysis of low rate wireless
technologies for medical application,” Computer Communication , vol.28, no.10,
p.1255-1275, 2005.
N. Golmie, N. Chevrollier, and O. Rebala, “Bluetooth and WLAN coexistence: challenges
and solutions,” IEEEWirelessCommunicationMagazine , vol.10, no.6, p.22-29, 2003.
I. Howitt and J.A. Gutierrez, “IEEE 802.15.4 low rate –wireless personal area network
coexistence issues,” IEEE Wireless Communication & Network, Conf. vol.3, p.1481-
1486, 2003.
S. Pollin, M. Ergen, and A. Dejonghe, “Distributed cognitive coexistence of 802.15.4 with
802.11,” Inter. Conf. on Cognitive Radio Oriented Wireless Networks
Communications , p.1-5, 2006.
Y. Xiao and N.H. Vaidya, “On physical carrier sensing in wireless ad hoc networks,” IEEE
Conf.onComputerCommunications , vol.4, p.2525-2535, 2005.
H. Ma, H. Alazemi and S. Roy, “A stochastic model for optimizing physical carrier sensing
and spatial reuse in wireless ad hoc networks,” IEEE Conf. on Mobile adhoc and
Sensorsystems , 2005.
H. Zhai and Y. Fang, “Physical carrier sensing and spatial reuse in multirate and multihop
ad hoc network,” IEEEConf.onComputerCommunications , p.276-285, 2006.
J. Zhu, S. Roy, X. Guo, and W.S. Conner, “Leveraging spatial reuse in 802.11 mesh networks
with enhanced physical carrier sensing,” IEEE Conf. on Communications , vol.7, p.
4004-4011, 2004.
J. Zhu, B. Metzler, X. Guo, and Y. Liu, “Adaptive CSMA for scalable network capacity in
high-density WLAN: a hardware prototyping approach,” IEEE Conf. on Computer
Communications , p.1-10, 2006.
Y. Zhu, Q. Zhang, Z. Niu and J. Zhu, “On optimal physical carrier sensing: theoretical
analysis and protocol design,”IEEEConf.onComputerCommunications , p.2351-2355,
2007.
K. Jamieson, B. Hull, A. Miu and H Balakrishnan, “Understanding the real-world
performance of carrier sense,” ACM SIGCOMM workshop on Experimental Approach
toWirelessNetworkDesignand, Analysis p. 52-57, 2005.
I. Ramachandran and S. Roy, “On the impact of clear channel assessment on MAC
performance,” IEEEGlobalCommunications, Conf p. 1-5, 2006.
5.4 Power Control

This section begins with a description of two known PC algorithms adapted for the DBS con-
text and used for benchmarking purpose. These are then compared through numerical exper-
iments with the multi-agent-based power control (MAPC) method described in 4.4.
5.4.1 Centralized Power Control (CPC)

Grandhi’s centralized power control (CPC) algorithm (Grandhi et al., 1993) is applied in the
DBS architecture by considering the M master DBS for the M mobiles on a given channel with
gij (1 ≤ i ≤ M, 1 ≤ j ≤ M) denoting the gain of the link from mobile i to DBS j, with DBS i
being mobile i’s master connection. Matrix A is defined as
Aij = gij /gii if i = j, (15)

Aii = 0. (16)
And the SIR at the master DBS is defined as

pi
γi = . (17)
∑ jM
=1 Aij pij
The power level for each mobile is then given by the eigenvector associated with the largest
positive eigenvalue of A.
Note that this algorithm is not trying to maximize each mobile’s SIR. Rather, it finds a set of
power levels which maximizes the lowest SIR, thus leading to each mobile’s SIR being equal
to the minimum (maximized) SIR. Also, the obtained power levels are proportional to at least
the eigenvector and thus need to be scaled to fit inside the mobiles’ power level range. This
is where the instability of this algorithm becomes apparent, since under certain interference
conditions, if a mobile is very close to its master DBS, its power level will be very low. Yet,
since its SIR is forced to be equal to the other mobiles’ SIR, proportionally, the noise at the
receiver will have a much stronger impact leading to very poor SINR. Supposing a mobile
faces 1W interference power and 0.1W noise power, and emits 10W to obtain a SIR of 10dB, it
has an SINR of 9.6dB. In contrast, consider a mobile faced with .1W of interference; it emits 1W
to obtain the same SIR of 10dB, but has an SINR of 7dB, hence an effective penalty of half. In
order to minimize this effect, the minimum power level should be high enough so that noise
remains as much as possible negligible. Hence, the power levels will be scaled such that the
maximum power level evaluated is set to the maximum mobile’s range.
On the other hand, a more complex evaluation of such effects would make it possible to lower
the maximum power level, keeping it as low as possible, and hence, maximizing the efficiency
of the link quality versus the power used per mobile.
5.4.2 SIR-balanced macro power control (SBMPC)

Yanikomeroglu’s SIR-balanced macro power control (SBMPC) (Yanikomeroglu & Sousa, 1998)
proposes an interesting algorithm for CDMA distributed antennas using macrodiversity. The
algorithm aims to balance, over all mobiles, each mobile’s aggregate (sum) SIR over all an-
tennas. This is a valid approach in the context of Rayleigh fading. However, as mentioned
in the introduction, in Rice fading, two similar average SIR values can lead to two different
BER figures given each may not have the same K factor (i.e. fading impact is more or less
severe). As such, balancing SIR does not balance BER with the relative importance of line of
sight components varying depending on mobiles’ locations.
Multi-Agent Design for the Physical Layer

of a Distributed Base Station Network
Université Laval
Canada
1. Introduction
As wireless networks are becoming more omnipresent and pervasive, appropriate resource
allocation and organization becomes an increasingly pressing challenge. There exist on the
consumer market two important types of wireless network technologies. On the one hand,
cellular mobile networks are highly centralized and hierarchical. By contrast, wireless local
area networks (WLANs) are deployed in an ad-hoc unstructured manner, thus avoiding the
need for elaborate and costly planning. However, WLANs such as those falling under the
highly successful 802.11 standard do not manage interference effectively and tend to collapse
at high offered traffic loads. It can be seen that cellular and WLAN represent two radically
different approaches in radio resource management, characterized by different sets of advan-
tages and drawbacks.
The purpose of this chapter is to demonstrate the feasibility of a connection-oriented self-
organized wireless system which offers efficient radio resource management and provides
the best aspects of both cellular (reliable, connection-oriented operation even at high of-
fered loads) and WLANs (ad-hoc deployment and distributed intelligence). This is achieved
based on the multi-agent concept and local synergistic micro interaction (between neighbor-
ing transceivers) from which a global organization emerges.
The notion of Multiple Agent (MA) considered is of the “ant” variety, whereby small min-
imalist agents sense their environment and react to it in an interdependent manner. Social
insects and mostly ants or bees are the most cited biological examples. In the literature
such approaches have already been used to solve many combinatorial/optimization problems
(Beongku et al., 2003; Brueckner & Parunak, 2003; Muraleedharan & Osadciw, 2003).
This design philosophy differs from more traditional approaches which consist in postulating
criteria expressed by equations and models in order to formulate the problem in such a way
that an optimal solution is derived within the defined context. In the agent approach, precise
mathematical formulation of the problem is neither required nor very useful. The approach
thus becomes attractive for tackling complex multidimensional problems which would oth-
erwise be intractable. Therefore, our goal is not to demonstrate an optimal design, but to
illustrate how a Multi Agent System (MAS) can be empirically designed and fine-tuned to fit
a specific application. Moreover, it will be seen that such a dynamically adaptive solution, in
spite of its empirical nature, offers many advantages over a rigid analytically-derived coun-
terpart.
We will focus herein on Parunak’s methodology (Parunak, 1997) because it offers an intuitive
modeling framework, which is well suited to the empirical design approach.
Considering wireless networks, this chapter describes a flexible distributed base station (DBS)
framework which removes many limitations of current networks in order to augment the so-
lution space. For example, a plurality of DBS can simultaneously provide a network link
to the same mobile, thus leveraging macrodiversity to improve link quality and/or achieve
power savings. These DBS are designed with auto-organization in mind, such that the net-
work structures itself autonomously. This is where MAS come in, offering the desired dis-
tributed intelligence, adaptability, scalability and auto configuration properties. However, the
DBS architecture is challenging in at least three aspects:
1. It requires the continuously-updated solving of a large combinatorial problem, namely
finding a good allocation of DBS resources to mobiles requiring service.
2. Interference must be handled in a transparent way so that mobiles can gain the best
benefit of macrodiversity without being restrained by interfering mobiles.
3. Power control is an important aspect for both energy consumption and network capac-
ity given that it is tightly-coupled with interference patterns.
These three aspects are entangled together such that an optimal allocation is a complex combi-
natorial problem. It is NP hard unless some heavy simplifying assumptions are made (on the
geometry, on propagation, or other aspects). Moreover, in the context of mobility, an optimal
solution at one point in time is not optimal if it cannot easily adapt to changing parameters
(mobiles’ positions, fading, etc.).
Yet, this complex context is well suited for a MAS design. Indeed, MA need an active environ-
ment in which to generate interaction. And each event of allocating power, channel or connec-
tions to mobiles, that a DBS generate, has consequences on other mobiles’ links. This creates
the required active environment in which agents can sense parameters such as the received
power, interference and link quality, and where decisions can be made locally to generate new
actions. In turn, the effect of these actions are sensed by other agents. The next section de-
scribes the challenges of the proposed DBS architecture. Then, MA design concepts used in
this study are described. The fourth section details the proposed design of three categories of
interacting agents respectively for :
1. macrodiversity connection management,
2. channel allocation, and
3. power level control.
Finally, the system is emulated. Results, including simulation of complex cases with randomly
distributed DBS and mobile traffic, show first the resource allocation quality that can be ob-
tained, and second the effectiveness of MAS design in terms of auto-configuration/scalability
and dynamic adaptation properties.
A final brief discussion will extend Parunak’s agent design principles to summarize the
lessons learned from designing MAS for the application at hand.
Fig. 1. Illustration of a DBS architecture exploiting 2 channels, and showing macrodiversity

relay connections.
2. Challenges of the Distributed Base Station Network

2.1 Macrodiversity Potential
In a perfectly geometrical network with homogeneous traffic and symmetric propagation con-
ditions, each DBS needs only to connect to the closest mobiles to maximize the provided qual-
ity of service. However, traffic is never homogeneous and varies across time and space in ac-
cordance with the users’ schedules and patterns of usage. Moreover, propagation conditions
are highly dependent on location, with varying availability of lines of sight and saturation
of the frequency band due to heavy traffic. In such a context, there is a need for a simple,
scalable, dynamic system to allocate relay links to mobiles and to continuously adapt the al-
location pattern to changing conditions.
DBS can choose to relay mobiles far away from themselves in order to provide them with more
macrodiversity, and thus better balance resource allocation. However, this choice involves a
trade off. The exponentially-decaying link quality with the mobile-DBS distance could lead a
remote mobile to consume many valuable relay links while deriving only marginal benefits,
whereas closer mobiles would obtain much higher macrodiversity benefits from those same
resources. Also, macrodiversity links provide not only enhanced overall quality links, but also
reliability against network disconnection when undergoing severe fading, and it facilitates
handover for mobiles moving outside from the range of some DBS to others. Therefore, a
single criterion such as maximizing the minimum QoS for all mobiles would fail in certain
conditions where enough resources would be available to provide the majority of mobiles
with decent QoS, because of a few mobiles consuming much of these resources while deriving
marginal benefits.
Such situations reveal the perils of pursuing a global solution based on a single perhaps overly
simplistic quality criterion. In fact, many Pareto equilibrium solutions exist, in which no mo-
bile can gain quality of service without stranding another user. And all these possible so-
lutions present multiple compromises on connection reliability and distribution of QoS. e.g.
some solutions could favor maximizing overall signal strength for high transfer rate, others
by distributing relaying links differently could prevent disconnections due to sudden strong
fading or interference because mobiles would in general enjoy higher probability of being
assigned multiple relay connections.
As such, it is not necessarily meaningful to define a priori goals for the search of a solution,
as it is not known beforehand what are the benefits and drawbacks of each possible Pareto
solution. This solution space is moreover hardly tractable due to the discrete nature of the
problem, with a finite but large number of link resources to attribute. It is limited by phys-
ical conditions where some links may not be feasible due to the weakness of the considered
signals. Also each link brings an increment of additional quality to the mobile’s overall link
quality, whose importance heavily depends on local propagation conditions that can vary
continuously (with slow fading and changing mobile-DBS distances), or abruptly given the
arrival of new connections in frequency allocation or strong fading situations. In this context,
no analytically tractable mathematical framework exists leading to an optimum solution tak-
ing into account all the dimensions of the problem. More specifically, one must consider that
an optimal solution, at a given state of the network and a given frame in time, could be too
heavily specialized to that particular situation, such that a sudden change (strong fading, new
mobiles joining the network) would make it ineffective. By analogy, it is known in biology
that a species too well adapted to its environment is heavily endangered due to its limited
capacity to adapt to environmental changes. Hence, a good solution is not an optimal one, but
a good enough one that provides margins for adaptation in time to face changes.
There is a strong need for distributed techniques which are flexible enough to be tuned to the
desired compromises while being able to handle unexpected events.
2.2 Channel allocation

Channel allocation faces the same propagation issues as connection management. To under-
stand the implications of channel management, we introduce the concept of channel footprint.
In a given situation (mobiles and DBS positions and relative densities, available channels,
power allocations, etc.), a mobile’s channel footprint can be understood as the space it oc-
cupies to maintain all its relaying links to DBS at a sufficient quality level. Hence, a second
mobile, if emitting on the same channel inside this space, would affect some or all of the first
mobile’s connections.
In the cellular context, it is assumed that each mobile enjoys the same channel footprint which
is controlled by the cell division of the space and an appropriate interference level threshold
to allow or prevent the reuse of a channel across cells. In the 802.11 protocol, it is a handshake
mechanism (the RTS/CTS exchange) which alerts neighboring transceivers that the channel
will be in use, in order to control, to some extent, this channel footprint by preventing neigh-
bors from reusing the channel in the vicinity, thus minimizing the hidden terminal effect (Ware
et al., 2001).
In the DBS architecture, it would be appropriate that mobiles be offered varying channel foot-
prints to adjust availability of channel resources and support the various needs of mobiles for
macrodiversity. Indeed, mobiles needing more macrodiversity would require a larger foot-
print. Moreover, mobiles close to all of their relaying DBS should allow other mobiles to reuse
the same channel at a closer range, compared to mobiles far from all DBS. This holds since
these mobiles can support higher interference power and still maintain a good signal to inter-
ference plus noise ratio (SINR). This aspect of channel allocation was taken into consideration
any gains by reducing its power level, a non cooperative strategy is not a good choice for a
game-theoretic approach to power level adaptation.
Necessarily, some mobiles will have to “accept” to reduce their power level in order to al-
low other mobiles in need to enjoy better QoS by reducing interference and enabling them
to reach more DBS for macrodiversity. Yet, and due to the non linearity of the propagation
environment, there necessarily is a point of diminishing returns for mobiles to reduce their
power level. While any reduction necessarily implies a reduction in interference, the potential
gain for other mobiles does not necessarily offset or compensate (given a compromise choice
at a global scale) the loss in QoS for this mobile. It is to be understood here that there exist
trade-offs for an infinity of Pareto solutions. Therefore, and again, postulating one global uni-
dimensional criteria (e.g. as is done in traditional algorithms (Grandhi et al., 1993) ) to derive
a power allocation method would not allow assessment of the potential benefits of different
trade-offs. Indeed, the results of the proposed design will show how the traditional approach
to power control (which consists in maximizing the minimum SINR for all mobiles) in spite of
offering interesting capabilities in some situations, also prevents most mobiles from achieving
their QoS potential.
2.4 Complexity, Dynamics and Scalability

2.4.1 Complexity
In existing types of networks, the complexity is constrained by simplifying the hypotheses.
For example, in cellular networks, channel allocation is simplified by segregating channels
given an interference power threshold in order to guarantee a minimum SINR for all mobiles
in a cell. This assumption simplifies the evaluation of provided QoS, as it guarantees a mini-
mum QoS for connected mobiles, and avoids the hidden terminal effect, such that there only
remains to evaluate the probability of a connection being blocked (when all channels in a cell
or sector are occupied).
In the considered architecture, such assumptions are not made a priori as the purpose of the
DBS architecture is to maximize flexibility. And considering the number of possible combina-
tions of connections, or channels or even power levels, it is obvious that an exhaustive search
to find all Pareto solutions is pointless. Even considering an exhaustive search in the case of a
very simple scenario with only a few mobiles is pointless, since in such cases, the non-linear
effects and interactions of large networks would not apply and the obtained results would be
too limited to draw meaningful conclusions.
Also, postulating a unidimensional criterion and over-simplifying the non-linear effects in-
volved, in order to provide a tractable mathematical framework would limit the solution space
and therefore restrict the possibilities of such an architecture.
MA offer interesting properties to cope with complexity. The approach involves segmenting
the problem into mulitple subproblems where each is tackled by its own agent class. Heavy
calculations for evaluating and selecting combinations are also avoided. Instead, specific com-
binations are attempted and modified by agents’ actions through local interactions.
2.4.2 Dynamics and Scalability

One particular aspect to consider is the fact that a given resource allocation solution must
necessarily adapt to changes in a mobile wireless network. Such a solution must also adapt
to unexpected events, such as the failure of a DBS. And finally it must scale, such that adding
DBS locally will seamlessly, without any need for configuration, increase the capacity of the
network in terms of either provided QoS or number of provided connections.
3. Multi-Agent Design
To solve the resource allocation problem, with the previously described considerations, mul-
tiple agents or bio-inspired optimization seems appropriate, as such approaches provide the
most important sought-after characteristics, namely
• scalability ;
• dynamic adaptation ;
• auto-configuration ;
• reliability facing unexpected events.
Following Parunak’s (Parunak, 1997) design principles, three main characteristics need to be
provided in a MA design: coupling, auto-catalysis, and function. Coupling implies that each
MA process is coupled directly or not to the others and their environment (e.g indirectly using
pheromones via an environment). Auto-catalysis implies that the agents’ actions taken in the
right direction1 , by the nature of the agents’ processes, favor similar actions leading the system
to converge to a desirable state (positive feedback reinforcing the convergence towards the
solution). And finally, the system must be such that a useful global function emerges out of
the induced local interactions.
3.1 Coupling
To achieve coupling, Parunak explains that we first need an active environment. The radio
propagation medium constitutes just such an environment, as each mobile emitting on a given
channel influences the others due to interference. Hence, a mobile’s movement changes the
interference patterns for all others in its immediate vicinity. Additionally, mobiles are entities
which strive to acquire connections and in so doing, they necessarily broadcast information
to inform neighboring DBS of their presence and of their link quality. This forms an active
environment in which information is exchanged to sustain coupled processes.
We emphasize the fact that DBS and mobiles do form appropriate entities to host agents that
are small in size and scope. In particular, DBS, compared to central cellular base stations, are
specifically meant to be small, and will necessarily have small scope as they can only relay a
(smaller) limited number of mobiles in their vicinity.
As a final criterion related to coupling, agents should be mapped as entities, not functions since
an agent does not implement a complete function. That is, the function optimizing resource
allocation should be the result of the interaction of the agents and not be implemented as the
output of one agent. Indeed, an ant (in ant colonies) does not find a shortest path alone.
In the proposed system, the agents are mapped to either mobiles or DBS. Their actions will
then be to either allocate or deallocate a channel, or a connection, or modify a mobile’s power
level. Necessarily, all processes which modify resource allocation are all coupled since each
agent’s actions will not only influence the concerned mobile (changing channel, obtaining a
new relay connection or changing its power level), but also influence the neighboring mo-
biles, modifying their own channel footprint, their QoS, hence influencing other agents, and
coupling each agent’s processes together indirectly.
1 Since a priori goals are not explicitly defined, neither is the concept of a ”right direction”. Rather, a
behavior is designed, tuned and retained because its auto-catalysis properties happen to converge to a
solution which satisfies the needs of the system. Therefore, such a design allows wide exploration of
the solution space rather then restricting to predefined goals by not including all the effects involved in
the multidimensional problem. The design represents a certain creative process.
3.2 Auto-catalysis
3.2.1 Flows
For agents to maintain their interactions, they must be designed to let the process evolve
continuously. Therefore, agents should not be designed based on discrete state transitions,
leading to pauses in the processes because of unverified conditions. That is why we must
favor flows instead of transitions. One way to achieve this is for agents to use volatile markers
(i.e. permanent and non-obstructive source of information which dissipate in time as they
become irrelevant –e.g. pheromones in ant colonies) to inform other agents on their particular
state, so that the agents’ processes continuously evolve rather than stop and wait for specific
conditions.
It is a design choice that no explicit information exchange is performed concerning the positions
of mobiles and DBS, available resources, etc. As mentioned, the available information stems
from what DBS and mobiles can sense locally (mobiles’ needs and QoS), which represents
our volatile markers. These bits of information are by nature volatile, as they only stay in
the environment as long as they are broadcasted by the mobiles, and hence are necessarily
current.
Since agents should not wait for predefined conditions to take actions, it is a comparative basis
that will trigger a corresponding action of:
1. allocating/deallocating a macrodiversity connection;
2. changing a mobile’s channel (frequency hopping);
3. increasing/decreasing a mobile’s power level to a certain amount.
3.2.2 Homeostasis
The notion stems directly from biology in which systems always strive to maintain an equi-
librium or homeostasis point, e.g. the blood sugar concentration is maintained (mainly) by two
different hormones which have opposite effects to balance the concentration.
This point of equilibrium must be sustained by an ongoing flow to ensure the system continu-
ously explores the solution space and does not get stuck in a deadend. This flow is analogous
to the variations of a stock market title whose value is influenced (at a macro level) by the
traders’ actions of selling and buying. In turn, at the micro level, the variations of the values
influences the traders’ decisions.
The corresponding aspect of our system is created by forcing DBS to continuously create and
destroy connections, continuously change channels (via channel hopping), and continuously
adjust power levels. Each of these actions — at the macro level of agents — influences the
status of mobiles, and these changes are in turn sensed by surrounding mobiles and DBS.
In effect, the flow of actions makes the system converge to a homeostasis point. This point
will be dependent on the the state of the network (traffic, available resources, etc.) due to
the comparative basis that triggers actions. As long as there exists a bias observed by the
agents that will trigger an action, the system will converge or oscillate to its homeostasis point.
These variations are important, since without them, and if there is no other change in the
system (e.g., induced by mobile motion), the sensed QoS of mobiles would never change,
never trigger actions, and the system might simply stop short of an optimal state.
3.2.3 Amplification and limitation

Together, amplification and limitation constitute an other important aspect to generate the
convergence to a homeostasis point. Amplification implies a positive feedback mechanism
such that convergence (to a solution) is favored. In other words, the actions of an agent which
lead the system in a desirable global direction should be favored and should also influence
the surrounding agents to act in the same direction.
In effect, an MA system is comparable to a Genetic Algorithm (Goldberg, 1989) preserving
"genes" that seem to provide the best fitness and hence are part of an optimal solution. The
difference is that there are no external observing entities that measure via a metric the fitness
of candidate solutions. Rather, it is the interactions between agents and their environment —
the propagation medium — that must provide the natural selection function.
Limitation also implies preventing the whole system from focusing on one point (exacerbating
the convergence of actions to a local minimum) and thus miss a better solution. Moreover, lim-
itation can favor convergence by dampening the effect of amplification to prevent the system
from going past a solution or oscillating around it without converging.
3.3 Function
Coupling may be trivial to obtain and auto-catalysis somewhat more involved, but if the pro-
cess as a whole does not realize a useful function, then it is irrelevant. Function implies that
the homeostasis point described previously is useful for the system, e.g. in biology the home-
ostasis point for the blood sugar concentration is such that enough sugar is available to fuel
the cells, but not too much to avoid excessive sugar loss through the kidneys.
In our system, the sought-after function consists in
• maximizing the potential usage of the resources;
• and balancing them to offer a good compromise of quality across all mobiles, while not
hindering the overall system performance.
Most often, function is obtained through a utility function which translates the flow of varia-
tions (of QoS) sensed into rational decisions. That is, it converts a multi-dimensional problem
into a one-dimensional quantity upon which decisions for actions are based.
In spite of the fact that many frameworks attempt to provide mathematical support to de-
rive such utility functions (such as game theory (Mackenzie & Wicker, 2001) or COIN theory
(Tumer & Wolpert, 2004)), these frameworks mostly consider intelligent agents having the
ability to learn (eventually using reinforcement learning techniques), which is not the nature
of the proposed design. Ultimately, defining simple agent behavior to obtain an intended
global behavior still relies on intuition and art such as in Conway’s “game of life” (Elwyn R.
Berlekamp et al., 1982), or with Wolfram’s cellular automatons (Wolfram, 2002). Therefore, no
systematic procedure is known which derives the locally-applicable utility function from the
desired global behavior.
Function can also be sustained (especially if a utility function is not found) with
• behavior diversity and
• randomness.
Randomness can be helpful to introduce alternative solutions, that will or not be kept in time
given how effective they are. Behavior diversity can be obtained by forcing neighboring
agents to act differently so as to provide different reactions and experiments given identi-
cal stimulus. These properties support the function property by breaking the symmetry so as to
prevent the system from entering any deterministic patterns which might hinder convergence.
In the following section it is described how auto-catalysis and function are obtained for each
class of agents.
Inter-RAT Handover Between UMTS And WiMAX 5.
1. If a mobile has no connection, it must be favored, since basic connectivity should take
precedence.
2. If a mobile has only one connection, the DBS should not disconnect it.
Furthermore, there are two complementary compromises involved in the DBS’ decision pro-
cess:
1. either to remove a link because the mobile already enjoys sufficient QoS,
2. or to maintain it because it is the main DBS providing it;
and,
1. either to connect a mobile because it is in need,
2. versus not connecting it because the additional diversity brought to this mobile would
be low (compared to other possible connections).
Finally, the function must provide a natural ordering to classify the compromises in order to
take a decision.
The following function addresses all the characteristics discussed above:
C (m, b) = Fneed (m) × log ( Fdiv (m, b)) . (5)
This function is necessarily positive or null. It is null if the mobile has no connection, since,
if it were connected to the DBS, it would have PT (m) = Pe (m, b) which implies Fdiv = 1
giving a null value of the logarithm. Likewise, it is null if considered for disconnection and
the mobile’s only link is to the considered DBS. Hence, if this utility function is null, the
agent will either privilege this mobile for connection or not disconnect it to keep the mobile’s
existing connection active.
The evaluation of the compromise is obtained by the multiplication of the two terms. Hence,
the more the DBS provides diversity, or the higher is the current QoS enjoyed by the mobile,
the higher is the function’s value.
The choice of compromise itself comes derives from “shaping function” used prior to the
multiplication of the two metrics. Simulation showed that the optimization happens most
efficiently if the shaping function of the second term is concave (naturally, it should be strictly
increasing), hence the use of the logarithm, which also provides the necessary null value for a
mobile with a single link.
4.2.4 Limitation and amplification

Limitation and amplification is naturally obtained with the environment propagation proper-
ties. Indeed, a poor signal quality will favor multiple connections (amplification), but distance
(mobile to DBS) and the infrastructure link capacity of DBS will restrict excessive connection
growth (limitation).
Also, this amplification (or attraction of macrodiversity links) and limitation sustains the
homeostatic behavior where mobiles in need get more links up to an equilibrium point where
additional links to these mobiles would overwhelmingly affect an otherwise well-served mo-
bile.
4.3 Channel allocation

4.3.1 Flow
The flow of actions in the channel allocation agents naturally consists of the changes in channel
allocation, or channel hopping which modifies mobiles’ QoS and interference patterns which in
turn should trigger other changes.
For this flow to be generated properly, appropriate actions are specified in the following.
4.3.2 Coupling
Following Parunak’s principles, the sought-after function (optimizing the allocation) is di-
vided into independent actions whose interactions should lead to the other two properties
(auto-catalysis and function).
First, given the macrodiversity context, a mobile will choose one of its relaying DBS to be its
“master” connection, which implies one type of action and one agent (to select the master)
mapped at each mobile.
Second, DBS will choose mobiles (from their master links) and change their channels as is done
in cellular systems. Except that here, the change, or channel hopping, will not be triggered by
specified conditions (e.g. a mobile SINR falling below a threshold, or a mobile changing cell).
Instead, the flow of channel hopping will be sustained by having DBS choose a mobile at each
agent activation and change its channel. Channel allocation agents will activate in the same
way that the connection agents do. Two types of actions must be defined:
1. choosing a mobile, and
2. choosing a channel.
Mapping these actions at the DBS level, rather than letting the mobile decide when to change
channel makes sense in that DBS can gather information most effectively on the different chan-
nels in use, thus preventing mobiles from having to continuously scan channels.
4.3.2.1 Sensing
In addition to the mobile’s sensed link quality, DBS can sense
1. the received power of surrounding mobiles pr (m, b);
2. and the interference level on various channels p I (b, c) (for channel c at DBS b).
4.3.3 Function
Maximizing the channel usage constitutes, in a sense, an effort against the second law of ther-
modynamics. Indeed, the channel allocation, if optimal at some point in time, will necessarily
deteriorate with mobility as two mobiles transmitting on the same channel get closer to a point
where the interference will degrade the offered QoS, such that resources are not balanced any-
more. Considering this aspect, and rather than trying to solve an NP-complete problem, load
balancing is obtained by always attempting to change the channels of mobiles in need such
that they enjoy better SINR.
Three utility functions need to be designed taking as input what the agents can sense, and
yielding a chosen parameter value as output.
a. Mobile m will choose a master DBS (among its relaying DBS) on activation (where its
activation follows a Poisson law) based on the DBS from which it obtains the highest
link quality:
b = arg max{ Pe (b, m)}. (6)
b
b. DBS b will choose a mobile (among mobiles connected as master to b), that is the most
in need, i.e.
m = arg min{ Fneed (m)}. (7)
m
c. Ideally, the DBS should try to use the channel with the lowest interference power level:
c = arg min{ p I (b, c)}. (8)

c
However, it may be overwhelming for a DBS to systematically sense channels to maintain up-
to-date information on interference levels on all channels, and this behavior (utility function
c.) is therefore only used as a benchmark.
Akaiwa & Andoh (1993) suggested a selection mechanism which is used herein with some
modifications. DBS b will scan channels in the order of a given priority list it maintains, and
determine if a channel can be assigned according to
• whether the resulting SINR will be above an SINR threshold ;
• and (in addition to Akaiwa’s method) whether it will also be above the actual SINR the
mobile enjoys.
The SINR threshold represents a mean to control the hidden terminal effect. It is a studied pa-
rameter in order to observe to which extend it prevents HTE while not limiting the flexibility
of the system.
For Akaiwa’s segregation algorithm, the priority list is obtained dynamically given the ratio
for each channel of previous assignments versus previous assignment attempts.
Finally, a random priority list is proposed as a simple, yet effective (as we will see) alternative
to the segregation algorithm approach.
4.3.4 Limitation
DBS will only test a limited number of channels given by the Chmax parameter, before giving
up. Indeed, there is no guarantee that the DBS will find a channel that will suit the chosen
mobile. Therefore, and instead of letting it scan all channels, it is forced to limit its search.
Eventually, it will try again, or another DBS will, thus providing behavior diversity as well.
In effect, the DBS are only trying to maintain channel assignments in a working state by “up-
grading” the solution iteratively in an opportunistic fashion given the eventual availability of
channels. It is the effect of a new channel allocation that will cause other DBS to also react
and change channels for the mobiles that will see their QoS affected by the new neighboring
interference. As this flow of action is sustained, the channel allocation remains functional and
should adapt to changes.
4.3.5 Channel availability

An additional functionality is provided for channel availability. A few spare channels are
reserved for the initial connection or reconnection of stranded mobiles (instead of using chan-
nels from the main pool). Then, a master DBS which has mobiles on these spare channels
will attempt to change their channels as a priority instead of choosing another mobile. Such
spare channels allow rapid network entry, providing higher availability as well as some time
margin for the DBS to find free channels in the main pool. It therefore eases the process and
the flow of channel hopping.
3.2.2.3 Handover from WiMAX to UMTS
Fig. 5. Handover signalling procedure from WiMAX to UMTS
The inter-RAT handover from WiMAX to UMTS is described in Fig. 5.

1) After the scanning interval, the MS sends scanning report to WiMAX serving BS by
message MON_SCN-REP that contains physical information such as mean RSSI.
2) The source WiMAX MAC sends CMacBSHOInd primitive to inform the IW sublayer of
target cell id. The IW then sends CPdcpBuffInfoReq primitive to the target RRC of the
UMTS network. The RRC shall return the CPdcpBuffInfoCnf primitive to inform the IW
sublayer of buffer size and buffer occupation. According to this information, the IW
adjusts its retransmission window size.
3) The IW sublayer sends a CRrcRelocReq primitive to the target RRC to apply for
resource allocation. The result is returned in CRrcRelocCnf primitive by the target RRC.
4) Upon receipt of the CRrcRelocCnf, the IW suspends and buffers data packets that
require delivery order.
5) IW sends CMacBSHOReq primitive to inform source MAC that the target network is
ready.
6) The MS performs handover to one of BSs specified in MOB_BSHO-REQ and responds
with a MOB_HO-IND message.
7) MS performs normal UMTS hard handover.
8) After the MS successfully finishes UMTS radio link setup, the target RRC shall send the
CRrcRelocCmpInd primitive to the IW, and the IW restarts data packet forwarding.
maximized with S = 0.8, and this has been shown to hold in many different conditions of
traffic, mobile speeds in (9) and available resources. .
The exponential in (9) is a shaping function which also naturally affects the dynamics of the
system. In effect, it affects mobiles’ convergence speed differently given their needs, and this
translates into behavior diversity as no mobile will react in a precisely proportional manner.
The proposed function is of course not the only possible choice, but it has proved stable and
effective. Again, for MAS, effectiveness does not lie in the mathematical exactness of the
function, but in the interactions it will generate.
4.4.3 Homeostasis
Finally, this Need factor must be converted to a delta (step) value to adjust the power level.
Homeostasis is obtained by comparing the Need value to the current power level the mobile
uses to transmit. Hence, the delta value is in the form of Needm − pm . The mobile will then try
to converge to a Needm value which depends on local interactions given itself and neighbor-
ing mobiles’ Fneed values (given that these are indirectly linked via interference). Eventually,
a non-linear concave function helps convergence so that with Needm and pm close, the gen-
erated delta is kept small to slow down variations and help stabilize the convergence. We
postulate
∆m = β sign(Needm − pm ) (|Needm − pm |)1.5 , (10)
where the β factor is used to modify the dynamics of the system to attain the proper compro-
mise between convergence speed and stability.
4.4.4 Limitation
Experience shows that this function is too unstable with high values of β. Still, it can be sta-
bilized with additional scaling parameters, while maintaining fast adaptation in time with
large values of β ≥ 5, which is important for mobility (β = 5 is used in the presented simu-
lations). Therefore, it is proposed that ∆ be scaled according to the current power level and
also the desired power level (the Need value). That way, if these values are small, ∆ is also
kept small to prevent strong changes in the system that would otherwise suddenly generate
exaggerated interference. Indeed, such changes would lead to complications such as breaking
existing links or simply propagating exaggerated reactions throughout the system. Building
upon (10), the following function is used :
∆m = pm × |Needm | × βsign(Needm − pm )(|Needm − pm |)1.5 . (11)
Finally, the delta value is constrained to not exceed the power level range:
− pm
∆m < 0 ⇒ ∆m = max{∆m , } (12)
2
1
∆m > 0 ⇒ ∆m = min{∆m , (1 − pm )}. (13)
2
As the mobile’s PC agent activates, its power level is adjusted as follows:
( ν +1) (ν)
pm = pm + ∆m . (14)
5. Evaluation
5.1 Simulation platform
For the channel and power agents, simulations are based on the following platform. The
results shown for the connection agents are based on a simpler scenario (detailed in the ap-
propriate subsection), in order to isolate the effect of connection management and observe its
convergence, while not confusing it with the effect of interference and power-level manage-
ment.
Physical parameters
A square field of 25 square kilometers is considered, in which 1000 mobiles evolve and 100
DBS are scattered randomly. Hence, the traffic’s and network resources’ geometry are not
uniform, thus generating good and bad coverage of different areas. A mobile moves in a ran-
dom direction at a random speed taken (at the start of a scenario) out of a uniform distribution
over [0, Vmax ]. DBS can relay 25 mobiles each, such that the mean number of macrodiversity
links per mobile is 2.5. A mobile’s maximum transmit power is 1W at 1 meter of its antenna,
and the propagation exponent is 4 (gij ∼ 1/d−4 ). Rayleigh fading is considered, except near
a DBS (closer than 100m) where a line of sight component is added with Rice factor K = 5
dB. Thermal noise at the receiver is considered for a bandwidth of 30kHz at a temperature of
20◦ C, hence N0 = −129 dBW. The number of available channels is denoted Ch.
Agents’ emulation
Simulations are run for 1000 seconds and repeated 10 times with different initializations of
the geometry (DBS positions and mobiles’ initial position, directions and speeds). Time is dis-
cretized with a time step of 1 second. At each time step, physical parameters are evaluated
(mobile’s position, propagation, interference, BER, connection outage). Agents activate ran-
domly given a Poisson distribution to estimate the next activation time with parameter λ = 3
time steps. At each time step, the agents which activate evaluate their local state and take
actions accordingly (adjust the power level, hop to a new channel, change connections of the
concerned mobile, etc.).
Results
At each time step, the set of QoS indexes (total BER level given on a logarithmic scale
PT (m) = log10 ( BER(m))) for each mobile are sorted, thus providing a snapshot in time of
the distribution of the network’s resources across all mobiles. These sorted distributions are
then averaged for all the time steps of the simulation. Given this information, it is then pos-
sible to compare how each algorithm distributes resources. The same is done for the power
level allocation. Also, to verify the stability in time (considering the dynamic properties) of
the algorithm, two factors are interesting to observe to understand how the system handles
outage :
1. the mean number N d of mobiles that loose all connections to the network per second,
and
2. the mean time tr it takes for the network to reconnect a mobile after it has been discon-
nected.
The latter also provides insight on how well the system is able to provide resources to mobiles
with high availability.
5.2 Connection agents

In order to show that the connection agents are indeed optimizing the connections to balance
resources, a simple centralized algorithm based on heuristics is proposed.
Algorithm 1 Centralized connection allocation algorithm

Considering initially that all DBS provide connections to all mobiles, and as long as there exist
DBS with more than N maximum connections :
A1 eliminate the connections with smallest Pe (m, b) as long as m has PT (m) < Pd (m) and
provided that DBS b has more than N connections ;
A2 (compromise on QoS) remove the ones with smallest Pe (m, b) as long as m remains con-
nected (another DBS is providing it a connection);
A3 (compromise on connectivity) finally remove connections with the smallest Pe (m, b) until
b has N maximum connections.
This algorithm is optimum at maximizing the sum of QoS ∑m PT (m), given it only removes
the smallest values. However, and given its limited ability to make compromises, it will not be
efficient at balancing resources for mobiles and preventing disconnections of some mobiles if
the network is resources-constrained. Necessarily, it offers a different trade-off than the agent
algorithm provides.
Three cases are observed:
1. there are not enough resources (Fig. 3(a)),
2. there are enough resources for connections, but not enough headroom / margin and the
agent system is not able to achieve swarming and converge (Fig. 3(b)),
3. there are enough resources to connect all mobiles and provide sufficient QoS, i.e. the
connection agent is efficient (Fig. 3(c)).
In practice, only the third case should be relevant provided that the network is appropriately
scaled for the needs of the users.
For the results shown in Figs. 3 and 4, a trellis of 19 DBS is used with 200 mobiles. Channel
management is not considered and each mobile has its own channel.
In the third case (Fig. 3(c)), three successive phases in time can be observed :
1. a connection stage, where connections are established to the closest mobiles;
2. a connection optimization stage, where connectivity is maximized, and
3. a connection rearrangement stage, where QoS is maximized.
Figure 4 depicts a sort of the QoS PT (m) of mobiles to provide insight on how well resources
are balanced. In this simulation, two classes of QoS are created, each comprising 100 mobiles.
Compared to the centralized algorithm, it is obvious that some load balancing is performed
by the agent system.
Due to lack of space, figures for the dynamic behavior are not shown herein. However, it is im-
portant to note that, without requiring any information centralization or excessive signaling
(which would generate delays), and based only on local interactions induced by connection
and disconnection actions, the agent system is able to keep up (maintain the connection al-
location in a relatively optimal state) fast enough to sustain mobiles moving at speeds of 50
km/h with connection agents activating in the mean only once every 3 seconds. Above that
200 200
number of mobiles not connected

150 150
number of mobiles
number of mobiles
number of mobiles without number of mobiles without
requested QoS requested QoS
100 100
number of mobiles not connected
50 50
0 0
0 100 200 300 400 500 0 100 200 300 400 500
time in seconds time in seconds
(a) Case 1 (b) Case 2
200
stage 1 : connection to closest mobiles
150 stage 2 : connections to all mobiles

number of mobiles
stage 3 : optimization, QoS and

ressource balancing
100
number of mobiles without

50 requested QoS
number of unconnected mobiles
0
0 100 200 300 400 500
time in seconds
(c) Case 2
Fig. 3. Convergence of the connection management agents (horizontal dashed lines show
results of the centralized heuristic algorithm).
speed, performance in terms of QoS degrades smoothly as the agents are not able to converge
fast enough to the optimal state. The system still provides much headroom as activation of
agents could be much faster.
Given that the proposed design is bio-inspired, it is most interesting to observe “health pa-
rameters” (analogous to e.g. blood pressure in the human body) which give us insight on the
capacity of the agents to achieve their function. For the connection agents, the mean number
of connections should be close to the mean number of disconnections. This indicates that the
agents have sufficient headroom (when faced with changes in the network) to actually swap
connections for optimization. If there are more connections than disconnections, it means that
the system is not able to keep up with changes so that some of the relay links are disconnected
for physical reasons (e.g. loss of signal quality) instead of explicit decisions by the agents.
−2
−4
−6
−8
−10
−12
−14
−16
−18
20 40 60 80 100
mobiles in class 1
mean number of blocked attempt to reconnect

1.8 1000
mean time for reconnection (in seconds)

mean number of HTE for 1000 mobiles
mean number of connected mobiles

90
100 995
1.6 mean number of
80
connected mobiles 990
1.4 70 80
(mean time to 985
(HTE) 60
1.2
reconnect) 980
60
50
975
1 40
970
40
0.8 30 965
20
blocked (re)connections rate
20 960
0.6
10 955
0.4 0
-25 -20 -15 -10 -5 0 5 10 15 20 -25 -20 -15 -10 -5 0 5 10 15 20
SINR threshold in dB SINR threshold in dB
(a) Comparing the number of lost connections (b) Comparing the mean number of connected
and the mean time before reconnection mobiles and the block rate
Fig. 5. Effect of the SINR threshold (Vmax = 5 m/s, Ch = 40, N = −129dBW).
to prevent momentary disconnection. But the efficiency of reconnection is so much improved

that in the mean, disconnection only occurs for less than 1 second (.8 seconds), and given
the exponential distribution of reconnection time, 90% of the disconnections have smaller
reconnection times. On the other hand, disconnection time is on the order of 5 seconds with
no reserved channels.
1000
with no reserved channels

with 2 reserved channels
number of mobiles
100
10
1
0.1 1 10 100
elapsed time before reconnection (seconds)
Fig. 6. Mean reconnection time with and without reserved channels. The Y axis represents the
number of mobiles remaining unconnected since their disconnection (for 1000 mobiles during
a 1000 seconds scenario length (Vmax = 5 m/s, Ch = 40, N = −129dBW).
Comparing the three different methods to select channels, little difference was observed in
terms of the mean number of connected mobiles. Differences appear as trade–offs between
mean number of disconnected mobiles and mean time of reconnection. The segregation
method seems more efficient at minimizing disconnection, but takes more time to reconnect,
compared to the random method. Differences appear most explicitly when looking at the
785
num ber of m obiles with QoS

780
775
770
765
760
755
750
Random
745
Segregation
740
-25 -20 -15 -10 -5 0 5 10 15 20

SINR threshold in dB
Fig. 7. Comparison of the performance of the three different methods for choosing channels
(Vmax = 5 m/s, Ch = 40, N = −129dBW).
provided QoS (Fig. 7). Choosing the channel with lowest interference power yields the best
results, followed by the random choice and the segregation method. We notice an increased
number of mobiles with their QoS demand met when the SINR threshold increases slightly.
This is simply due to the fact that fewer mobiles are connected, generating less interference
and higher QoS. However, this fact does not hold true for long, ince a further increase in the
threshold leads to a rapid increase of the agents’ block rate, showing that they are unable to
keep up with changes, and are not finding free channels to adapt the allocation pattern. This
leads to a fall in QoS (above 5-7 dB).
Finally, the maximum number of channel scans per allocation attempt Chmax , without power
management, reveals an exponential gain which saturates at around 5-6 channels scanned per
allocation attempt. However, combined with the effect of power management, no significant
gain is observed. It appears that with power management, mobiles adjust their footprint thus
offering more channel availability, such that only one channel test per allocation attempt is
sufficient. And, optimization of the channel allocation occurs naturally through the many dif-
ferent attempts from all surrounding DBS in time. Therefore, complexity is kept to a minimum
by having DBS only test and eventually allocate one channel per agent activation.
Connection management remains efficient as long as there are sufficient channel resources to
provide a large enough channel footprint for the macrodiversity links. However, this mini-
mum number of channels is low as macrodiversity links only require a small SINR to provide
sufficient QoS after macrodiversity combination.
In the current simulated scenarios, 25 channels2 shared by a 100 DBS for 1000 mobiles, includ-
ing 2 reserved channels (for (re)connection), are enough to provide sufficient flexibility to the
connection agents for them to be able to swap links for optimization (this is with the synergis-
tic effect of power level management). Below this threshold, there is too much interference for
the connectivity potential of DBS to be fully exploited for macrodiversity, and those resources
are left unused.
2 A cellular system with a channel reuse pattern of 7 hexagonal cells, would require 70 channels for 100
pico cells, without offering the flexibility the current architecture provides.
5.4 Power Control

This section begins with a description of two known PC algorithms adapted for the DBS con-
text and used for benchmarking purpose. These are then compared through numerical exper-
iments with the multi-agent-based power control (MAPC) method described in 4.4.
5.4.1 Centralized Power Control (CPC)

Grandhi’s centralized power control (CPC) algorithm (Grandhi et al., 1993) is applied in the
DBS architecture by considering the M master DBS for the M mobiles on a given channel with
gij (1 ≤ i ≤ M, 1 ≤ j ≤ M) denoting the gain of the link from mobile i to DBS j, with DBS i
being mobile i’s master connection. Matrix A is defined as
Aij = gij /gii if i = j, (15)

Aii = 0. (16)
And the SIR at the master DBS is defined as

pi
γi = . (17)
∑ jM
=1 Aij pij
The power level for each mobile is then given by the eigenvector associated with the largest
positive eigenvalue of A.
Note that this algorithm is not trying to maximize each mobile’s SIR. Rather, it finds a set of
power levels which maximizes the lowest SIR, thus leading to each mobile’s SIR being equal
to the minimum (maximized) SIR. Also, the obtained power levels are proportional to at least
the eigenvector and thus need to be scaled to fit inside the mobiles’ power level range. This
is where the instability of this algorithm becomes apparent, since under certain interference
conditions, if a mobile is very close to its master DBS, its power level will be very low. Yet,
since its SIR is forced to be equal to the other mobiles’ SIR, proportionally, the noise at the
receiver will have a much stronger impact leading to very poor SINR. Supposing a mobile
faces 1W interference power and 0.1W noise power, and emits 10W to obtain a SIR of 10dB, it
has an SINR of 9.6dB. In contrast, consider a mobile faced with .1W of interference; it emits 1W
to obtain the same SIR of 10dB, but has an SINR of 7dB, hence an effective penalty of half. In
order to minimize this effect, the minimum power level should be high enough so that noise
remains as much as possible negligible. Hence, the power levels will be scaled such that the
maximum power level evaluated is set to the maximum mobile’s range.
On the other hand, a more complex evaluation of such effects would make it possible to lower
the maximum power level, keeping it as low as possible, and hence, maximizing the efficiency
of the link quality versus the power used per mobile.
5.4.2 SIR-balanced macro power control (SBMPC)

Yanikomeroglu’s SIR-balanced macro power control (SBMPC) (Yanikomeroglu & Sousa, 1998)
proposes an interesting algorithm for CDMA distributed antennas using macrodiversity. The
algorithm aims to balance, over all mobiles, each mobile’s aggregate (sum) SIR over all an-
tennas. This is a valid approach in the context of Rayleigh fading. However, as mentioned
in the introduction, in Rice fading, two similar average SIR values can lead to two different
BER figures given each may not have the same K factor (i.e. fading impact is more or less
severe). As such, balancing SIR does not balance BER with the relative importance of line of
sight components varying depending on mobiles’ locations.
Inter-RAT Handover Between UMTS And WiMAX 541
Fig. 14. Control plane protocol stacks of tight coupling architecture
In Fig. 13 and Fig. 14, the user and control planes of the proposed tight coupling architecture
are illustrated. W-RNC is assumed to cover the same Routing Area (RA) like the RNC. The
IW sublayer on the W-RNC communicates with its counterpart entity on the RNC in order
to execute inter-RAT handover to/from its control area. The main contents of the
communication between them are as follows:
 GTP-U sequence numbers as well as GTP packets that need to be forwarded by the W-
RNC for PDP contexts requiring delivery order.
 IW ARQ parameters, such as windows size, queue length, retransmission timer period,
and retransmission count.
 The IW blocks stored in local retransmission buffer.
There are two reasons why add IW ARQ mechanism to W-RNC in addition to SRNS
(Serving RNS) context transfer of conventional RNC:
 When an inter-RAT handover takes place, there may exist packet sequence number
asynchronization between the source RNC and the target WiMAX BS. It is necessarily
that there exists a common context transfer mechanism for these two systems to assure
a lossless handover.
 The second reason is that the WiMAX supports cell reselection initiated by MS for
active traffics (like dedicated mode in UMTS), which is not the case in UMTS. Hence,
the packets that are lost during the cell reselection from WiMAX to UMTS, cannot be
retransmitted by the target network.
0 1
No PC
CPC
SBMPC
-2 MAPC
0.8
-4
0.6
Power levels
-6
0.4
-8
0.2
-10 No PC
CPC
SBMPC
MAPC
-12 0
0 200 400 600 800 1000 0 200 400 600 800 1000
mobiles mobiles
(a) Sorted BER profile averaged in time (b) Sorted power level ratio profile averaged
(meant (sortm ( PT (m)))) in time (meant (sorti ( pi i )) )
Fig. 8. Ch = 40, N = 0, Vmax = 0.
No PC CPC SBMPC MAPC

N̄d (×10−6 ) 570 8.9 130 190
t¯r (seconds) 24.2 1 2.0 2.65
Table 1. Outage behavior without mobility and noise with Ch = 40.
Figure 8(b) reveals how both CPC and SBMPC offer similar distributions of the power levels.
On the contrary, the MAPC power level allocation is radically different.
Faced with higher interference levels (Ch = 25), it can be seen that the centralized algorithms
break down (Fig. 9(a)). Indeed, in high interference levels, maximizing the minimum SIR
leads to very poor SIRs for all mobiles. In turn, this generates many disconnections. Figure
9(a) clearly shows that the traditional algorithms are here inefficient and even worse than
without PC. Still the MAPC algorithm manages to provide acceptable levels of QoS, while
still connecting more mobiles.
The situation deteriorates even more when noise is introduced. Figure 9(b) reveals how noise,
as explained previously, renders the traditional algorithms unstable, generating lots of dis-
connections. Indeed, the centralized algorithms, by maximizing the minimum SIR, force most
mobile power levels to be extremely low (c.f. Fig.8(b)) supposing thermal noise is not an im-
portant factor. This may be valid for a regular hexagonal cell geometry with homogeneous
traffic and high guaranteed SIR. However, it is not the case here, leading to very poor SINR,
and also significantly exacerbating the HTE as such mobiles’ presence on channels will not be
sensed by other DBS, rendering the PC algorithm completely inefficient.
Also, facing important mobility (Figure 9(c)), the CPC algorithm loses its strength (of mini-
mizing the outage probability) as Table 2 reveals. Indeed, with mobility, more interference is
present because mobiles do not obtain an optimal reallocation of channels at each iteration.
This implies far too many very low power levels with the CPC. This conflicts with the channel
agents trying to reorganize the channel allocation as it generates important hidden terminal
effects. This also shows that the CPC algorithm loses much of its capacity with even small
0 0
-2 -2
-4 -4
-6 -6
-8 -8
-10 No PC -10 No PC
CPC CPC
SBMPC SBMPC
MAPC MAPC
-12 -12
0 200 400 600 800 1000 0 200 400 600 800 1000
mobiles mobiles
(a) Low channel resources, Ch = 25, N = 0,

Vmax = 0.
No PC CPC SBMPC MAPC

N̄d (×10−4 ) 9.8 19.1 3.8 1.8
t¯r (seconds) 5.13 2.4 2.2 1.38
Table 2. Outage behavior with mobility (Vmax = 5 m/s) Ch = 50.
0 1
No PC
CPC
SBMPC
-2 MAPC
0.8
-4
0.6
Power levels
-6
0.4
-8
0.2
-10 No PC
CPC
SBMPC
MAPC
-12 0
0 200 400 600 800 1000 0 200 400 600 800 1000
mobiles mobiles
(a) Sorted BER profile averaged in time (b) Sorted power level ratio profile averaged in
(meant (sortm ( PT (m)))) time(meant (sorti ( pi i )))
Fig. 10. Ch = 40, N = 0, Vmax = 0, discrete power level.
centralized schemes by preventing them from assigning excessively low power levels, yet also
preventing them from achieving any proper resource balancing as the MAPC does.
Figure 10(b) shows the allocated power level over all mobiles in the discrete case.
Considering dynamics, it was found that tuning the different parameters mentioned in the
design (most specifically β and S) makes the agent system stable enough to prevent erratic
changes of the power levels, while still allowing fast adaptation in the wake of abrupt (dis-
crete) changes in interference levels.
It is also noteworthy that the power management proposed provides additional benefits to
the channel agents (lowering connection loss rate, block rate and mean time for reconnection)
and connection agents (offering them more headroom to swap links for optimization). Despite
being designed independently, the various described agents remain stable working together
in a synergistic way, such that one change induced by one type of agent (e.g. a connection
change) is correctly compensated by the other types of agents (e.g. an adaptation of the power
level) and not overcompensated, which would otherwise lead to erratic behaviors.
One of the great advantages of this approach is that each dimension of the problem can be
tackled independently, by its own class of agents, without affecting the performance of the
other classes. No explicit interaction mechanism between the agent classes needs to be de-
signed, yet they synergetically cooperate to achieve desirable results with a surprisingly low
implementation complexity.
6. Conclusion
The proposed proof-of-concept design described herein demonstrates that minimalist multi-
agent systems do provide all expected qualities: scalability, dynamic properties, efficiency,
simplicity, adaptability and auto-configuration. Moreover, it represents a novel and surpris-
ingly simple solution to resource allocation. It is most obvious with the power allocation
scheme which results in drastically lower spatial power distributions when compared with
traditional algorithms.
The multi agent design approach, based on heuristics appears effective. It requires an under-
standing of the underlying mechanisms and compromises within the context of the problem
at hand, from which insights and intuition can be drawn and used to design the agents.
While it is unclear a priori to which end result the system will converge to, it should be noted
that it is also unclear a priori to which it should. Indeed, the current context is far different
from formal frameworks such as information theory which are often characterized by a single
uni-dimension criterion, e.g. the channel capacity. In our multi-dimensional context, infor-
mation theory remains too limited at the time to model and grasp the many possibilities and
compromises facing a multitude of mobiles and DBS with macrodiversity where limited re-
sources lead to interference. And in such a context, the proposed MA approach has the virtue
of demonstrating via simulation that some novel allocation solutions (which should be under-
stood as compromises) can lead to much higher efficiency of resource usage (where efficiency
is necessarily also a notion of compromise).
The design in itself is not so complicated, and one should keep in mind the fuzziness of such
an approach. Indeed, the utility functions proposed could have a variety of alternatives. What
matters is not their exactness, but that they provide certain properties that will sustain the in-
teractions of agents. These properties remain to be understood and studied to provide insights
on the inner workings of the agent system. For example, the shaping function in the utility
function of the connection agents uses a logarithm which could be replaced by a first degree
approximation (x − 1) and still converge, but a concave function with similar properties (e.g.
( x − 1)2 , null for x = 1 and strictly increasing for x > 1) would, despite providing some de-
gree of convergence, fall short of a more balanced solution. The shaping function is therefore
crucial to converge to certain Pareto solutions, and this remains to be studied in detail.
Concerning MA design, we considered more specifically the notion of homeostasis which is
not explicitly mentioned in Parunak’s methodology. The proposed design shows how the
search for such an equilibrium helps in designing and tuning the properties of the agents’
behavior to obtain the desired global function.
Considering future work, the proposed MA design and DBS architecture offers malleability
and vast margins for tuning, enhancing, or providing additional functionality. It was studied
in (Leroux et al., 2008) that the reuse of channels could be enhanced by pairing mobiles to
cooperate in exploiting a single channel while multiplying their diversity gain. Interesting
results have been found in this study. Yet, coupled with power-level control, the management
of cooperation between mobiles revealed counter synergistic effects. To date, finding a way to
have the cooperation and power-control agents interoperate in a synergetic manner remains
an open problem.
Another additional functionality to be studied is beamforming. Channel allocation agents
would need to be improved to account for dynamically-created directive beams and provide
network-wide gains by minimizing interference .
Channelization also needs to be further studied, including a model to implement the IEEE
802.11 shared random access mechanism (CSMA/CA on conjunction with the so-called dis-
tributed coordination function) which is now globally deployed, rather then relying only on
orthogonal channels (whether it be time/frequency/ or code division) as is the case in this
study. Finally, practical implementations should be tested, as discussed in Leroux (2008)
where it was shown that macrodiversity could be obtained through minimal terminals syn-
chronized at the packet level. It would then be possible to implement the proposed MA strate-
gies using consumer Wi-Fi terminals and perhaps connect such terminals to a wired network,
make them work in synergy and thus offer much more reliable and efficient connections.
The proposed system is therefore not a simple exercise for MA design. It represents a mean-
ingful starting point for a new design paradigm of mobile wireless networking. It offers vast
potential for improvements, new designs and additional functionality.
7. References
Akaiwa, Y. & Andoh, H. (1993). Channel segregation-a self-organized dynamic channel al-
location method: application to TDMA/FDMA microcellular system, IEEE J. Select.
Areas Commun. 11(6): 949–954.
Beongku, A., Dohyeon, K. & Innho, J. (2003). A modeling framework for supporting QoS in
mobile ad-hoc networks, Vehicular Technology Conference, 2003. VTC 2003-Spring. The
57th IEEE Semiannual 2: 935–939.
Brueckner, S. & Parunak, H. V. D. (2003). Self-organizing MANET management, in G. D. M.
Serugendo, A. Karageorgos, O. F. Rana & F. Zambonelli (eds), Engineering Self-
Organising Systems, Vol. 2977 of Lecture Notes in Computer Science, Springer, pp. 20–35.
Elwyn R. Berlekamp, John H. Conway & Richard K. Guy (1982). Winning Ways for your Math-
ematical Plays, New York: Academic Press.
Furukawa, H. & Akaiwa, Y. (1994). A microcell overlaid with umbrella cell system, Vehicular
Technology Conference, 1994 IEEE 44th, Stockholm, pp. 1455–1459.
Goldberg, D. E. (1989). Genetic Algorithms in Search, Optimization and Machine Learning, Kluwer
Academic Publishers, Boston, MA.
Grandhi, S., Vijayan, R., Goodman, D. & Zander, J. (1993). Centralized power control in cellu-
lar radio systems, Vehicular Technology, IEEE Transactions on 42(4): 466–468.
Leroux, P. (2008). Architecture d’un système de stations de base distribuées, PhD thesis, Université
Laval.
Leroux, P., Roy, S. & Chouinard, J.-Y. (2006). The Performance of Soft Macrodiversity Based
on Maximal-Ratio Combining in Uncorrelated Rician Fading, 17th annual IEEE inter-
national symposium PIMRC’06, Helsinki, Finland.
Leroux, P., Roy, S. & Chouinard, J.-Y. (2008). Synergetic cooperation in a distributed base
station system, Personal, Indoor and Mobile Radio Communications, 2008. PIMRC 2008.
IEEE 19th International Symposium on, pp. 1–6.
Mackenzie, A. & Wicker, S. (2001). Game Theory and the Design of Self Configuring, Adaptive
Wireless Networks, IEEE Commun. Mag. .
Muraleedharan, R. & Osadciw, L. A. (2003). Balancing the Performance of a Sensor Network
Using an Ant System, 37th Annual Conference on Information Sciences and Systems (CISS
2003), Baltimore, MD. .
Parunak, H. V. D. (1997). “Go to the Ant” : Engineering Principles from Natural Agent Sys-
tems, Annals of Operations Research .
Tumer, K. & Wolpert, D. (2004). Collectives and the Design of Complex Systems, Springer-Verlag,
NY.
Inter-RAT Handover Between UMTS And WiMAX

Bin LIU and Philippe Martins
TélécomParisTechÉcole
- NationaleSupérieuredesTélécommunica
France
Philippe Bertin and Abed Ellatif Samhat
FranceTelecomResearchandDevelopment
France
1. Introduction
The future beyond third generation (B3G) or fourth generation (4G) systems will consist of
different radio access technologies, such as GSM/GPRS, UMTS, WiFi, and WiMAX. Many
intensive efforts have been made to identify the unsolved issues about the future mobile
systems, and one important issue is what the future vertical handover management solution
will be. A variety of mobility management solutions have been proposed, such as
MIPv6/FMIPv6 (D. Johnson, et al., 2004; R. Koodli, 2005), SCTP (M. Afif, et al., 2006), inter-
RAT (Radio Access Technologies) handover of 3GPP (3GPP TS 43.129; 3GPP TR 25.931).
Among these solutions, the layer 2 inter-RAT handover solution of 3GPP is a promising way
for its high reliable handover procedure. Unfortunately, the 3GPP inter-RAT solutions only
support inter-RAT handover between cellular networks, and do not support inter-RAT
handover between WiMAX (Worldwide Interoperability for Microwave Access) and UMTS
(Universal Mobile Telecommunications System).
Another important issue is the interworking architecture and the coupling scenario that are
used to provide an efficient inter-RAT handover management. Depending on where is the
coupling point, there are several interworking architectures: no coupling, loose coupling,
tight coupling, very tight coupling (integrated coupling) (G. Lanpropoulos, et al., 2005). The
loose coupling and tight coupling architectures often use Mobile IP or part of Mobile IP as
the handover management protocol. So these two kinds of coupling architectures require
less complicated modifications to the existing protocol stacks and are more flexible than
integrated coupling. However, they often suffer from longer handover latency varying from
some hundreds of milliseconds to some seconds. The integrated coupling generally achieves
better handover performance at expense of adding complex modification to existing
network protocol stacks.
In recent years, the 3GPP and IEEE organizations have proposed their respective
interworking solutions for convergence of heterogeneous networks. For instance, the
ongoing 3GPP standard for interworking between UMTS and WiFi (3GPP TS 23.234) only
focuses on the control plane, and defines the interworking topologies, access gateways,
AAA, charging, interfaces and so on. It does not provide a scheme to resolve the handover
problems in the user plane, like packet loss, long handover latency. Another promising
vertical handover solution is IEEE 802.21 (IEEE 802.21, 2006). This standard defines a generic
link layer to mask the heterogeneities of various RATs as well as three kinds of services.
How to resolve handover problems is still manufactures’ task. For these reasons, in this
chapter, we will only focus on the user plane instead of control plane, and propose an inter-
RAT handover solution to resolve some typical handover problems. Besides, our inter-RAT
handover solution can be applied to a variety of interworking scenarios in addition to the
integrated and tight coupling architectures. This solution is a novel solution for
interworking between UMTS and WiMAX, which has not been stated by other references or
standards.
In section 2, we firstly summarize the features of 3GPP Packet Switched (PS) network/cell
switch procedures, i.e. reselection and handover, and then we get some guides to the design
of inter-RAT handover mechanism between UMTS and WiMAX. The PS handover is
introduced in order to support real-time packet-switched traffics with strict QoS
requirements on low latency and packet loss. For one thing, the handover reduces the
service interruption of the user plane information when cell changes compared to the cell
reselection; for another, it enables buffer handling of user plane data in order to reduce
packet loss when cell changes. For unreal-time services with loose QoS requirements on
packet loss, or Mobile Station (MS) is not in dedicated state, the PS cell reselection is
introduced. Compared to PS handover, the cell reselection suffers from uncertain packet
loss, but benefits from the reduced signaling and resource overhead between cells or RATs.
Next, in section 3 and 4 respectively, based on the requirements of inter-RAT handover
between UMTS and WiMAX, we propose a novel layer 2 inter-RAT handover scheme by
introducing a novel common sublayer named IW (InterWorking) sublayer and SR ARQ
(Selective Repeat ARQ) mechanism in the integrated and tight coupling architectures to
resolve several typical inter-RAT handover problems, such as packet loss, long handover
latency. The better handover performance is validated by the simulation results carried out
on the NS2 emulator. In addition, this novel IW sublayer scheme also eliminates the false
fast retransmission, which is due to packet loss or out-of-order packet arrivals during a
handover period.
Finally, we come to our conclusions in section 5.
2. 3GPP Handover Features
This section mainly describes procedures that are used by the MS to select a suitable cell to
be connected to in the 3GPP cellular networks. For details, please refer to references (3GPP
TS 25.331; 3GPP TS 43.129; 3GPP TS 44.060; 3GPP TR 25.931; J.P Romero, et al., 2005).
When a MS is switched on, it must first select a PLMN (Public Land Mobile Network) and a
RAT (Radio Access Technology) automatically or manually. It searches for the most
adequate cells so as to camp on a suitable cell. Then the MS will register its presence in the
registration area of the chosen cell if necessary. As long as the MS remains in idle mode, it
will continuously execute the cell reselection procedures in order to choose and camp on the
most suitable cell of the selected PLMN. The events triggering the cell reselection may due
to high path loss, downlink signaling failure and so on. If the new cell is in a different
registration area, a Location Registration (LR) request is performed (3GPP TS 23.122).
When the MS is in the packet transfer or dedicated mode, the network use PS Handover to
command a MS to move from its source cell to a new cell, and continue the ongoing PS
service operation in the new cell. The handover trigger conditions could be serving cell
resource limitation, measurement reports from a MS, and the cell change notification from a
MS. This handover procedure between cells of the same RAT is also referred to as intra-RAT
handover.
Depending on the PLMN availability and network configuration, it is possible that cell
reselection and handover procedures involve a cell switch from one RAT to another RAT
(e.g., from 2G GSM/GPRS/EDGE to 3G UMTS networks). The handover procedure
between cells of different RATs is referred to as inter-RAT handover in 3GPP standards or
verticalhandover in IETF protocols.
In general, in contrast with the cell reselection, the intra-RAT handover and inter-RAT
handover of 3GPP have the following distinct features:
 The network makes the handover decision depending on the measurement reports,
network states and negotiation with target base station or cell.
 The network controls the whole handover procedure, including message transfer,
handover timing, handover target, resource allocation, and context transfer.
 Only packet transfer mode or dedicated mode needs the handover procedure due to
handover overheads.
 The handover can be considered as a kind of specific QoS-guaranteed cell reselection
procedure.
It should be stressed that Mobile IPv6 and its extensions (D. Johnson, et. al., 2004; R.Koodli,
2005) support both network-initiated and terminal-initiated handover procedures. In this
chapter, we only consider 3GPP inter-RAT handover procedures initiated by the network.
The conventional 3GPP inter-RAT handover procedure must involve the SGSN, whether for
handover from GSM/GPRS to UMTS or from UMTS to GSM/GPRS. This is because the
Link Control Sublayer (LLC) terminates at SGSN, which is in charge of making lossless
packets forwarding during the MS mobility thanks to its retransmission mechanism. This is
not a problem for GSM/GPRS or UMTS core network, because the SGSN is their common
network entity. But for inter-RAT handover between 3GPP cellular network and IEEE
wireless IP network like WiFi/WiMAX, it becomes a challenge for the packet lossless RAT
switch procedure due to the lack of SGSN in IEEE wireless network. In the following
sections, we utilize the 3GPP inter-RAT handover procedure, messages and signaling to
resolve this problem for two typical coupling network architectures: integrated coupling
and tight coupling.
3. Inter-RAT Handover between UMTS and WiMAX in Integrated

Coupling Architecture
In order to realize a seamless inter-RAT handover for future B3G or 4G mobile networks, a
variety of interworking architectures and inter-RAT handover mobility managements have
been proposed. Based on the integrated architecture, in this section, a novel common
interworking sublayer (IW sublayer) is proposed at Layer 2 on RNC and MS to provide a
seamless PS inter-RAT handover between UMTS and WiMAX systems. This IW sublayer
scheme focuses on eliminating packet loss and reducing handover latency that are common
problems for most inter-RAT handover scenarios. Compared with other context transfer
schemes, the simulation results show the IW sublayer with ARQ mechanism can achieve a
lossless and prompt handover procedure. In addition, this IW sublayer scheme can
eliminate false fast retransmission of TCP traffics that is usually caused by packet losses or
out-of-order packet arrivals. In what follows, the IW sublayers in integrated and tight
coupling architectures are specified in this section and section 4 respectively.
3.1 Context Transfer

The problems during the inter-RAT handover period have been extensively studied in (G.
Lanpropoulos, et al., 2005; S.L. Tsao, et al., 2002; N. Dailly, et al., 2006; J. Sachs, et al., 2006;
H. Inaura, et al., 2003; H. Rutagemwa, et al., 2007), such as long handover latency, BDP
(Bandwidth Delay Product) mismatch, delay spikes, packet losses, premature timeout, false
fast retransmission. Among these problems, the packet losses and long handover latency are
in particular not desirable for real-time and throughput-sensitive traffics. The most common
solution is applying the context transfer mechanism (J. Sachs, et al., 2006; R. Koodli, 2005) or
retransmission (3GPP TS 25.323; N. Dailly, et al., 2006) to accelerate the handover process or
reduce the amount of lost packets. There exist the following typical context transfer and
retransmission schemes: PDCP Synchronization, buffering-and-forwarding (B&F), SDU
Reconstruction, R-LLC.
PDCP Synchronization: In 3GPP UMTS network (3GPP TS 25.323), the PDCP sublayer is
applied to guarantee reliable data transmission service during Service Radio Network
Subsystem (SRNS) relocation. For this purpose, PDCP maintains PDCP sequence numbers
to avoid any data losses during SRNS relocation. After the successful relocation, the data
transmission starts from the (first) unconfirmed SDU having a sequence number equal to the
next expected sequence number by the PDCP entity. For instance, in the uplink, if some
transmitted SDUs are still left unacknowledged, the data transmission is resumed by
retransmission of the SDU with the “uplink send” sequence number equal to the uplink
receive sequence number. Otherwise, the data transmission is resumed with the
transmission of the first unsent SDU. Moreover, when the RLC entities mapped to a lossless
PDCP entity are reset or reestablished for reasons other than SRNS relocation, then it is the
PDCP’s responsibility that the peer lossless PDCP entities do not go out of synchronization
by the means of “PDCP sequence number synchronization procedure” (3GPP TS 25.323).
The retransmission mechanism of PDCP works well when a MS performs SRNS relocation
during data transmission in the domain of UMTS. Unfortunately, in the scenario of inter-
RAT handover, the PDCP will not take effect any more because:
 The other heterogeneous network system usually does not have the similar mechanism,
especially IEEE 802 RATs such as WiMAX or WiFi.
 In addition, the WiMAX system has its own IP packet header compression mechanism
rather than ROHC (3GPP TS 25.323) in UMTS. If the packets or frames stored in the
source system with their particular headers and control signaling parts are forwarded
to the target systems directly, the target system may discard these unreadable packets,
which will induce sequence number asynchronization and break down current
communication connection.
In a word, the sequence number synchronization, header compression and retransmission
mechanism of respective RAT complicate the inter-RAT handover procedure instead.
Buffering-and-Forwarding: R. Koodli (2005) propose to utilize buffering-and-forwarding
mechanism (B&F) to forward unsent data packets from previous access router to new access
Inter-RAT Handover Between UMTS And WiMAX 5.
 Consider the UMTS as the center of the integrated system and preserve its signaling
and control procedures as many as possible.
 In WiMAX access network, additional network components and new signaling and
primitives could be added.
 The two integrated systems should guarantee seamless service continuity, and execute
mobility management processes (e.g., connection establishment, handover) as fast as
possible in order to maintain the required QoS.
Fig. 1. IW sublayer working mechanism of integrated coupling
As stated above, our inter-RAT scheme is first based on the integrated coupling architecture.
We assume UMTS to be the master home network with roaming privileges to WiMAX
network. A novel common network entity named interworking sublayer (IW) is introduced
on the top of PDCP (Packet Data Convergence Protocol) sublayer of UMTS and the Medium
Access Control (MAC) CS sublayer of 802.16e on the RNC and MS, as shown in Fig.1. The
WiMAX BS is integrated with the RNC (Radio Network Controller) through Iub interface.
The IW takes the role of LLC sublayer of conventional cellular networks, such as
retransmission mechanism and handover support. The main functions of IW sublayer are:
 Determination of a suitable target network.
 Primitive mapping between the IW and the UMTS network, or between the IW and the
WiMAX network in case of inter-RAT handover.
 Support SR ARQ (Selective Repeat ARQ) mechanism, including packet segmentation
and re-sequencing, retransmission, and retransmission window size adjustment.
In Fig. 2 and Fig. 3, the user and control planes of the proposed architectures are illustrated.
It should be stressed that the SR ARQ retransmission mechanism is realized in the user
plane. While in the control plane, IW sublayer shall translate handover related signaling
between source and target networks. When an inter-RAT handover is made, the IW
sublayer is activated according to the QoS requirements of a PDP (Packet Data Protocol). In
the control plane, we prefer to reuse the RRC protocol functionality in the MS and RNC
respectively, instead of building them from scratch. What is actually needed is to enhance
RRC protocol entities in order to forward inter-RAT handover primitives to IW sublayer. In
order to minimize the modifications to respective systems, IW sublayer also realizes some
essential WiMAX-related primitives and makes RNC act as another WiMAX BS to the
WiMAX network.
Fig. 2. User plane protocol stacks in integrated coupling architecture
Fig. 3. Control plane protocol stacks in integrated coupling architecture

3.2.2 Signaling and Primitives
3.2.2.1 Overview
In order to have insight into IW sublayer working mechanism, this sub-clause describes the
inter-RAT handover signaling procedures and primitives among IW, PDCP, RRC (Radio
Resource Control) and WiMAX MAC. Some newly added cross-layer primitives are
complemented to the conventional inter-RAT handover signaling procedures of 3GPP (3GPP
TS 43.129; 3GPP TR 25.931). We suggest the future WiMAX and UMTS standards should
support these primitives and parameters for the seamless and smooth inter-RAT handover.
Generally, the inter-RAT handover consists of handover preparation phase and handover
execution phase. In the case of a handover from UMTS to WiMAX, when the inter-RAT
handover conditions e.g. low RSSI or load increase, are met, the MS is instructed by the
RNC to switch on its WiMAX transceiver. Then MS seeks and monitors the neighbor
WiMAX BSs given in System Information Block (SIB) on BCCH of serving cell. After the
WiMAX scanning intervals (IEEE 802.16e, 2005), the MS provides the network with its
measurement results of the target networks using Measurement Reports message.
Meanwhile, other important wireless link parameters, such as round trip time (RTT), BDP
are also calculated by the RNC. After that, the inter-RAT handover will enter into execution
phase if the RNC makes a positive handover decision.
3.2.2.2 Handover from UMTS to WiMAX
Fig. 4. Handover signaling procedure from UMTS to WiMAX

Fig. 4 describes the inter-RAT handover from UMTS to WiMAX and shows the exchanged
messages and primitives.
1) Based on measurement reports and knowledge of the RAN topology, the RNC, more
precisely source RRC decides to initiate an inter-RAT PS handover.
2) The source RRC sends the CRrcRelocInd primitive (contains target WiMAX cell id) to
the IW sublayer.
3) Then the IW sends the CMacBuffInfoReq primitive to the target WiMAX MAC to
request the buffer characteristics. The WiMAX MAC shall return the CMacBuffInfoCnf
primitive to inform the IW of the buffer size in its MAC sublayer. According to this
information, the IW adjusts its retransmission window size. (Note that current WiMAX
MAC does not support this interface, so the IW may adjust its retransmission window
size to a default value).
4) At this stage, the IW sends the CMacBSSynchReq primitive to the WiMAX MAC to
negotiate the location of the dedicated initial ranging transmission opportunity for the
MS. This information is returned by primitive CMacBSSynchCnf.
5) After that, the IW begins to buffer data packets that require delivery order and sends a
CRrcRelocReq primitive (including Transparent Container (MOB_BSHO-REQ)) to the
source RRC.
6) The RRC sends the Handover from UTRAN Command message to the MS, which
includes a MOB_BSHO-REQ.
7) The MS performs hard handover and normal WiMAX network entry procedure.
8) After the provisioned service flow is activated (IEEE 802.16e, 2005), the target WiMAX
MAC sends CMacBSHOCmpInd primitive as a Link_Up (LU) trigger to the IW
sublayer. On this trigger, the IW shall restart data packet forwarding.
Fig. 5. Handover signalling procedure from WiMAX to UMTS

target cell id. The IW then sends CPdcpBuffInfoReq primitive to the target RRC of the
UMTS network. The RRC shall return the CPdcpBuffInfoCnf primitive to inform the IW
sublayer of buffer size and buffer occupation. According to this information, the IW
adjusts its retransmission window size.
3) The IW sublayer sends a CRrcRelocReq primitive to the target RRC to apply for
resource allocation. The result is returned in CRrcRelocCnf primitive by the target RRC.
4) Upon receipt of the CRrcRelocCnf, the IW suspends and buffers data packets that
require delivery order.
5) IW sends CMacBSHOReq primitive to inform source MAC that the target network is
ready.
CRrcRelocCmpInd primitive to the IW, and the IW restarts data packet forwarding.
Note that primitive CMacBSHOCmpInd and primitive CRrcRelocCmpInd are defined as the
Link_Up (LU) triggers for handover from UMTS to WiMAX and for handover from WiMAX
to UMTS respectively.
3.2.2.4 IW ARQ Mechanism

For the sake of achieving lossless inter-RAT handover, a modified Selective Repeat ARQ (SR
ARQ) mechanism is applied to the IW sublayer during the handover period, which is
renamed IW ARQ. The IW ARQ is an error control mechanism that involves error detection
and retransmission of lost or corrupted packets. When a packet is accepted from upper
layer, it is segmented into smaller IW blocks, each of which is assigned a sequence number
(see Fig. 6). This new IW sub-header is used for block loss detection and block re-sequencing
in the receiver to guarantee in-sequence delivery. Afterward, each IW block is transmitted
through the UMTS or the WiMAX interface. These IW blocks are also queued in the
retransmission buffer in order to be scheduled for retransmission. The IW ARQ transmitter
maintains an adaptive window size that is set to target network buffer size. When an IW
block is received by the receiver, a positive or negative acknowledgement (ACK/NACK) is
sent back immediately for the purpose of reducing handover latency. In addition, in order to
avoid dead lock due to IW ACK/NACK losses during a handover period, a status report
timer is set when the receiver sends an ACK/NACK. When this timer expires, the receiver
sends back a status report (ARQ feedback bitmap) providing the receipt status. This status
report is map of the acknowledgement (ACK) or negative acknowledge (NACK) of each IW
block within the window. Compared with conventional SR ARQ mechanism of RLC, the IW
ARQ has the following features:
 Receiver-Driven scheme: the received status and ACK/NACK are sent back on receipt
of an IW block initiatively without transmitter’s polling message.
 Support Link_Up (LU) trigger: when a handover is finished, the target network will
signal the IW sublayer with a Link_UP trigger. On receipt of this trigger, the IW
sublayer will retransmit blocks in retransmission buffer to avoid unnecessary waiting
for an expiration of the status report timer.
 Adaptive Window Size: In order to avoid any buffer overflow in the target network
when the packets are retransmitted by the IW sublayer after a handover, the IW ARQ
window size is adaptively set to buffer size of the target network.
In Fig. 6, an example of the IW ARQ mechanism when the window size is 12 is depicted.
The right parts are two retransmission mechanisms: IW ARQ and R-LLC. In this figure, the
difference between them is in that the lost blocks are retransmitted when status report timer
expires in R-LLC scheme, while IW ARQ retransmits unacknowledged blocks not only on
the expiration of this timer but also on the Link_Up trigger.
Fig. 6. IW ARQ and R-LLC protocol: an example of time evolution
3.3 Simulation Environment
Fig. 7. Simulation topologies: integrated coupling (left) and tight coupling (right)
In order to analyze the performance of the IW sublayer during inter-RAT handover between
UMTS and WiMAX, network-level simulations are carried out using NS2. Several
extensions are made to this simulator, UMTS and WiMAX models, IW sublayer, multi-
channel model, IW ARQ mechanism and new signaling and primitives. The simulation
topologies for integrated and tight coupling scenarios are illustrated in Fig. 7. There is only
one MS with two transceivers and no other background traffics in this “clean” scenario. The
MS always has enough bandwidth to send packet whether it is in WiMAX region or in
UMTS region. Note that in this topology, the transmission delay in the wired network is set
very small deliberately to minimize its influence to handover procedure. An FTP session is
examined, with the CN designated as the sender and the MS designated as the receiver. In
UMTS module, a drop-tail policy is applied to radio network queues in PDCP and this
queue length is set to 25 IW blocks. As to the WiMAX module, the queue length is set to 50
IW blocks, which considers the fact that generally the bandwidth of WiMAX is much higher.
Other important simulation parameters are summarized in Table 1.
Parameter Value Parameter Value
Fragment Switch OFF TTI (ms) 10

UMTS Frame 10
PHY Duration(ms)
Max retransmit count 10
IW BLER 1e-6
Allocated data rate unlimit
Default Windows size 30 ed
(block)
Queue length 50
WiMAX
MAC
Status Report Timer (s) 2.5 Payload Header no
Suppression
TCP/IP Header no
compression,and
Retransmission Frame duration 4
(ms)
PDCP
Allocated data rate 64kb/s
Modulation OFDM
Queue length 25
WiMAX
PHY Interleaving 50
RLC Mode AM
interval (frames)
Windows size (Blocks) 500
FFT 256
RLC Block size (Bytes) 20 Number of 200
subcarrier used
maxDAT 20 Variant Reno
Ack timerout period 50 TCP/IP MSS (bytes) 512
(ms)
Default cwnd 32
3.4 Simulation Results in the Integrated Coupling Scenario
3.4.1 Handover form UMTS to WiMAX

For the simulation of inter-RAT handover from UMTS network to the WiMAX, an FTP
session starts at 0.4 sec., and the MS starts to perform handover at about 4 sec. after it enters
into the coverage region of WiMAX. The handover type is hard handover. At about 4.035
sec. the WiMAX network entry procedure is finished and the IW sublayer on the RNC
receives a Link_Up trigger. Fig. 8 shows the packet flows of three kinds of context transfer
schemes: R-LLC, SDU Reconstruction and IW ARQ.
The R-LLC scheme does not support Link_Up trigger, so it retransmits the last
unacknowledged data packets on the expiration of status report timer. During this period,
the TCP timer expires and the congestion window shrinks to one, as shown in Fig. 9. There
is a retransmitted TCP segment at about 5.7 sec.
Fig. 8. TCP segment number comparison (umts->wimax, sender side)
Fig. 9. TCP congestion window (umts->wimax)

The SDU Reconstruction scheme reconstructs the RLC PDUs stored in the RLC
retransmission buffer. However, if one PDU of a SDU is successfully transmitted, this PDU
is deleted from retransmission buffer and the remaining PDUs of this SDU cannot be
reconstructed and are discarded locally. The remaining RLC SDUs (TCP packets here) are
forwarded to WiMAX network after handover on RNC. These arrivals of out-of-order
packets generate several duplicate ACK and trigger TCP fast retransmission process. The
TCP congestion window size shrinks to half of congestion window of steady state, and the
average throughput is also reduced.
The IW ARQ scheme adjusts its retransmission window according to the target network’s
queue size and forwards the IW blocks in its retransmission buffer on receipt of Link_Up
trigger. After handover is over, there are no packet losses and the TCP ACK arrivals are not
as bursty as those of SDU Reconstruction scheme thanks to the IW ARQ window
mechanism (see Fig. 8 between 4 and 4.1 sec.).
3.4.2 Handover from WiMAX to UMTS
Fig. 10. TCP segment number comparison (wimax->umts, sender side)
A typical problem during handover process from high bandwidth data network WiMAX to
relative low bandwidth network UMTS is buffer overflow, which is caused by BDP
mismatch between these two networks. The UMTS network is likely to undergo buffer
overflow when a TCP congestion window for WiMAX is much larger than the buffer
allocation per MS in UMTS RNC.
For SDU Reconstruction scheme, even though TCP congestion window is not larger than
buffer size of UMTS, the buffered packet forwarded from WiMAX to UMTS still may have
the probability to overflow the UMTS queue, because queue in WiMAX may buffer more
packets than the queue size of UMTS due to inflated transmission time. For SDU
Reconstruction scheme, in Fig.10, the buffer overflow in UMTS after handover leads to TCP
retransmission starting at about 6.0 sec. The corresponding TCP window shrinks, as shown
in Fig. 11.
Fig. 11. TCP congestion window (wimax->umts)
For R-LLC scheme, the long status report period leads to TCP RTO and a segment is
retransmitted by the TCP sender three times before a status report timer expires. The period
of status report timer is set to 2.5 sec. in this scenario.
Whereas for IW ARQ scheme, the support of Link_Up trigger accelerates handover response
time, and the adaptive IW ARQ window size effectively eliminates buffer overflow in the
target UMTS network. It can be seen that the lossless handover of IW ARQ mechanism has a
“side effect”: eliminate the false fast retransmission caused by packet losses or out-of-order
packet arrivals during a handover.
4. Inter-RAT Handover between UMTS and WiMAX

in Tight Coupling Architecture
4.1 The IW Sublayer in the Tight Coupling Architecture
4.1.1 IW Sublayer Description
Fig. 12. IW sublayer working mechanism of tight coupling
In the tight coupling scenario, the WiMAX network may emulate a RNC (Radio Network
Controller) or a SGSN (Serving GPRS Support Node). We only consider RNC emulation in
this chapter. Thus, we introduce a new network component called RNC emulator for
WiMAX (W-RNC) in the WiMAX access network, which connects with the UMTS CN (Core
Network) at the Iu-PS interface, as shown in Fig. 12. Actually, the W-RNC is an enhanced
WiMAX BS with a novel sublayer named IW sublayer, which lies on the top of WiMAX
MAC (Medium Access Control) sublayer. The W-RNC with the IW sublayer has the
following functions:
 Realize Iu-PS interface.
 Primitive mapping between the IW and the UTRAN network or between the IW and
the WiMAX network in case of an inter-RAT handover.
 When an inter-RAT handover takes place, the IW sublayer functions as the LLC
sublayer of conventional cellular networks by enabling the SR ARQ (Selective Repeat
ARQ) mechanism that includes packet segmentation, re-sequencing, retransmission,
and retransmission window size adjustment.
 When a handover takes place, the IW sublayer transfers context to target RNC or W-
RNC where the counterpart sublayer locates. In order to provide a seamless inter-RAT
handover between UMTS and WiMAX, a peer IW sublayer shall also be realized on the
top of the PDCP sublayer on the conventional RNC. While on the MS, the IW sublayer
is a common sublayer on the top of the PDCP sublayer of UMTS and the MAC sublayer
of WiMAX.
Fig. 13. User plane protocol stacks of tight coupling architecture

Fig. 14. Control plane protocol stacks of tight coupling architecture
In Fig. 13 and Fig. 14, the user and control planes of the proposed tight coupling architecture
are illustrated. W-RNC is assumed to cover the same Routing Area (RA) like the RNC. The
IW sublayer on the W-RNC communicates with its counterpart entity on the RNC in order
to execute inter-RAT handover to/from its control area. The main contents of the
communication between them are as follows:
 GTP-U sequence numbers as well as GTP packets that need to be forwarded by the W-
RNC for PDP contexts requiring delivery order.
 IW ARQ parameters, such as windows size, queue length, retransmission timer period,
and retransmission count.
 The IW blocks stored in local retransmission buffer.
There are two reasons why add IW ARQ mechanism to W-RNC in addition to SRNS
(Serving RNS) context transfer of conventional RNC:
 When an inter-RAT handover takes place, there may exist packet sequence number
asynchronization between the source RNC and the target WiMAX BS. It is necessarily
that there exists a common context transfer mechanism for these two systems to assure
a lossless handover.
 The second reason is that the WiMAX supports cell reselection initiated by MS for
active traffics (like dedicated mode in UMTS), which is not the case in UMTS. Hence,
the packets that are lost during the cell reselection from WiMAX to UMTS, cannot be
retransmitted by the target network.
4.1.2 Signaling and Primitives
4.1.2.1 Handover from UMTS to WiMAX

This sub-clause describes the inter-RAT handover signaling procedures and primitives
among IW, PDCP, RRC (Radio Resource Control) and WiMAX MAC in the tight coupling
architecture. In the following figures, IW/RNC means the function combination of IW
sublayer and RNC, so are the IW/W-RNC and MAC/W-RNC. Some newly added cross-
layer primitives are augmented to the conventional inter-RAT handover signaling
procedures of 3GPP (3GPP TS 43.129; 3GPP TR 25.931). We suggest the future WiMAX and
UMTS standards should support these primitives and parameters for the smooth and
seamless inter-RAT handover. The handover preparation period is similar to that of
integrated coupling architecture and is omitted in this sub-clause.
Fig. 15. Signaling procedure of the handover from UMTS to WiMAX
Fig. 15 describes the inter-RAT handover from UMTS to WiMAX and shows the exchanged
messages.
1) Based on measurement reports and knowledge of the RAN topology, the RNC, more
precisely source RRC decides to initiate an inter-RAT PS handover.
2) The source RNC sends a Relocation Request (contains target WiMAX cell id) message to
the SGSN. The SGSN forwards Relocation Request message to target W-RNC.
3) Then the IW of target W-RNC sends the CMacBuffInfoReq primitive to the WiMAX
MAC to request the buffer characteristics. The WiMAX MAC returns the
CMacBuffInfoCnf primitive to inform the IW of the buffer size in its MAC sublayer.
According to this information, the target IW adjusts its retransmission window size. It
should be mentioned at this point that current WiMAX MAC does not support this
interface, so the IW may adjust its retransmission window size to a default value.
4) At this stage, the target IW sends the CMacBSSynchReq primitive to the WiMAX MAC
to negotiate the location of the dedicated initial ranging transmission opportunity for
the MS. This information is returned by primitive CMacBSSynchCnf.
5) The target W-RNC sends the Relocation Request Acknowledge message to SGSN, and
the SGSN continues the handover by sending a Relocation Command to source RNC
(including Transparent Container (MOB_BSHO-REQ)).
6) Upon receipt of Relocation Command message, the IW of source RNC will forward IW
context to target IW of W-RNC. The IW context consists of IW ARQ parameters,
received IW ACK and remaining IW blocks that have not been transmitted successfully.
7) The RRC of the source RNC sends the Handover from UTRAN Command message to
the MS.
8) The MS performs hard handover and normal network entry procedure.
9) After the provisioned service flow is activated, the target WiMAX MAC sends
CMacBSHOCmpInd primitive as a Link_UP (LU) trigger to the IW sublayer. On this
trigger, the IW starts data packet forwarding.
Fig. 16. Signaling procedure of the handover from WiMAX to UMTS


handover and target cell id. Then, the source W-RNC sends a Relocation Request
(contains target cell id) message to the SGSN. The SGSN forwards Relocation Request
message to target RNC.
3) The IW of target RNC sends CPdcpBuffInfoReq primitive to the RRC sublayer to
request the buffer characteristics of the PDCP sublayer, and RRC returns the
CPdcpBuffInfoCnf primitive to inform the IW of buffer size. According to this
information, the IW adjusts its retransmission window size.
4) The target IW sends a CRrcRelocReq primitive to the target RRC to apply for resource
allocation. The result is returned in CRrcRelocCnf primitive by the target RRC.
5) The target RNC sends the Relocation Request Acknowledge message (contains target
RNC to source W-RNC transparent Container) to SGSN. The SGSN continues the
handover by sending a Relocation Command to source W-RNC.
6) On receipt of Relocation Command message, the IW of source W-RNC will forward IW
context to the IW of target RNC. The IW context consists of IW ARQ parameters,
received IW ACK, and remaining IW blocks that have not been transmitted
successfully.
7) The source IW sends CMacBSHOReq primitive to inform MAC that the target network
is ready.
CRrcRelocCmpInd primitive to the IW, and the IW starts data packet forwarding.
Note that primitive CMacBSHOCmpInd and primitive CRrcRelocCmpInd are defined as the
Link_Up (LU) triggers for handover from UMTS to WiMAX, and for handover from
WiMAX to UMTS respectively.
4.1.3 Buffering-and-Forwarding (B&F) in Tight Coupling Architecture

We have mentioned that: in FMIPv6 protocol (R. Koodli, 2005), in order to make a handover
lossless, previous access router (PAR) will forward buffered packets destined for the MS
duringm l outer u CMacF CM orwF te Ar r r I
2QR 1 .LPXUD 7 )XMLL7 $6W XG\ RQ $XWRQRPRXV 1HLJKERXU $FFHVV 5RXWHU

’LVFRYHU\ IRU 0RELOH ,3
Fig. 18. TCP congestion window (umts->wimax)
The IW ARQ scheme adjusts its retransmission window according to the target network’s
queue size, and sends the IW blocks that are forwarded from the source IW on receipt of
Link_Up trigger. After the handover, there will be no packet losses and TCP congestion
window does not shrink thanks to the retransmission mechanism.
4.2.2 Handover from WiMAX to UMTS
Fig. 19. TCP segment number comparison (wimax->umts, sender side)

Fig. 20. TCP congestion window (wimax->umts)
When a handover from WiMAX to UMTS happens, there exist some TCP segments and IW
blocks, which are forwarded from W-RNC to RNC through the tunnel for two schemes, as
shown in Fig.20. For B&F scheme, the arrivals of tunneled segments (about 8 segments)
trigger the fast retransmissions twice at time 4.31 sec. and 4.51 sec. for the lost segments
during the handover (see Fig. 19), and then congestion window size reduces significantly.
From then on, the TCP sender retransmits the segments numbering from the first lost
segment to the tunneled segments. The receiver will acknowledge again those segments that
have been tunneled before at about 6.18 sec, which herein trigger the bursty segment
arrivals. Furthermore, those retransmitted segments that have been tunneled during
handover procedure delay the ACK feedback of new segments, and in consequence lead to a
retransmission caused by RTO at 6.58 sec.. We can see that, the B&F scheme degrades the
handover performance instead of improving it for TCP traffics due to the lack of a
mechanism that recovers the lost packets.
For IW ARQ scheme, there is no packet loss during the handover. The support of Link_Up
trigger accelerates handover response time, and the adaptive IW ARQ window size
effectively eliminates buffer overflow in the target UMTS network. The only price for this
lossless handover procedure is that the IW sender may retransmit a couple of IW blocks that
possibly have been received by IW receiver but the corresponding ACKs are lost in the air
during the handover period.
5. Conclusion
This chapter focuses on introduction of proposed inter-RAT handover solution for
interworking UMTS with WiMAX. First, the 3GPP cell reselection and handover mechanism
are outlined in the first section. This section gives us main guideline for designing a new
mechanism to deal with typical problems of inter-RAT handover between UMTS and
WiMAX. Then, a novel Layer 2 inter-RAT handover scheme on basis of the integrated
coupling and tight coupling architectures for the seamless roaming between UMTS and
WiMAX networks are elaborated. For instance, in integrated coupling architecture, a new
common sublayer named IW sublayer that lies on the RNC and MS is added on the top of
PDCP (UMTS) and MAC (WiMAX) sublayer. At this novel sublayer, a new retransmission
scheme called IW ARQ is also proposed to eliminate packet losses during handover
procedure and accelerate handover procedure. Compared with other context transfer
mechanisms, such as R-LLC, SDU Reconstruction and buffering-and-forwarding, IW ARQ
can achieve lossless and prompt handover procedure for TCP traffics thanks to the
introduction of SR ARQ mechanism. The better handover performance is validated by the
simulation results carried out on NS2 emulator. In addition, this novel IW sublayer scheme
also can eliminate the false fast retransmission due to packet loss or out-of-order packet
arrivals during a handover period. It also provides a suitable framework to solve the TCP
problem of BDP mismatch, premature RTO and so on.
6. References
3GPP TS 04.18. Technical Specification Group GSM/EDGE Radio Access Network; mobile
radio interface layer 3 specification; radio resource control protocol
3GPP TS 05.08. Technical Specification Group GSM/EDGE Radio Access Network; Radio
subsystem link control (release 99)
3GPP TS 23.060. Technical Specification Group Services and System Aspects; General Packet
Radio Service (GPRS); Service description, stage 2 (Release 7), v7.4.0
3GPP TS 23.122, “Technical Specification Group Core Network and Terminals; Non-Access-
Stratum (NAS) functions related to Mobile Station (MS) in idle mode”, (Release 7),
V7.2.0
3GPP TS 23.234. 3GPP system to Wireless Local Area Network (WLAN), interworking;
System description (Release 6), V6.5.0
3GPP TS 25.304. User Equipment (UE) procedures in Idle Mode and Procedures for Cell
Reselection in Connected Mode
3GPP TS 25.323. Technical Specification Group Radio Access Network; Packet Data
Convergence Protocol (PDCP) specification, (Release 7), V7.5.0
3GPP TS 25.331. Technical Specification Group Radio Access Network; Radio Resource
Control (RRC); Protocol Specification (Release 7), v7.5
3GPP TR 25.922. Technical Specification Group Radio Access Network; Radio resource
management strategies (Release 7), V7.10
3GPP TR 25.931. Technical Specification Group RAN; UTRAN functions, examples on
signaling procedures, (Realease 7), V7.4.0
3GPP TS 43.022, “Technical Specification Group GSM/EDGE; Radio Access Network;
Functions related to Mobile Station (MS) in idle mode and group receive mode”,
(Release 7), V7.2.0
3GPP TS 43.129. Technical Specification Group GSM/EDGE” Radio Access Network;
Packet-switched handover for GERAN A/Gb mode; Stage 2 (Release 7), V7.2.0
3GPP TS 44.060. Technical Specification Group GSM/EDGE Radio Access Network; General
Packet Radio Service (GPRS), Radio Link Control/Medium Access Control
(RLC/MAC) protocol (Release 7), V7.9.0
3GPP TS 45.008. Technical Specification Group GSM/EDGE; Radio Access Network; Radio
subsystem link control (Release 7), V7.8.0
C. Johnson, R. Cuny & N. Wimolpitayarat. (2005). Inter-System Handover for Packet

Switched Services”, 205 6th IEE International Co nference on G3 and Beyond, pp: 1-5,
7-9 Nov. 2005
D. Johnson, C. Perkins, & J. Arkko. (2004). IP Mobility Support in IPv6,
http://www.ietf.org/rfc/rfc3775.txt, IETF, June 2004
E. Seurre, P. Savelli & P.J. Pietri. (2003). GPRSforMobileInternet, Artech House Ltd, 2003
G. Lanpropoulos, N. Passas, L. Merakos & A. Kaloxylos. (2005). Handover Management
Architectures in Integrated WLAN/Cellular Networks. IEEECommunicationSurvey
&Tutorials. Vol.7, No.4, pp: 30-44, Fourth Quarter 2005
H. Inaura, G. Montenegro, R. Ludwig, A. Gurtov & F. Khafizov. (2003). TCP over Second
(2.5) and Third (3G) Generation Wireless Networks, IETF, RFC 3481
H. Rutagemwa, S. Park, X.M. Shen & J.W. Mark. (2007). Robust Cross-layer Design of
Wireless Profiled TCP Mobile Receiver for Vertical Handover, IEEE Tans. On
VehicularTechnology, Vol.56, No. 6, pp: 3899-3911, Nov. 2007
Http://www.isi.edu/nsnam/ns
IEEE 802.16e (2005). IEEE Standard for Local and metropolitan area networks Part 16: Air
Interface for Fixed and Mobile Broadband Wireless Access Systems Amendment 2:
Physical and Medium Access Control Layers for Combined Fixed and Mobile
Operation in Licensed Bands, 2005
IEEE 802.21 (2006). Standard and Metropolitan Area Networks: Media Independent
Handover Services, Draft P802.21/D00.05, January 2006
J.P Romero, O. Sallent, R. Agusti & A.D.G. Miguel. (2005). Radio Resource Management
Strategiesin,UMTS John Wiley & Sons, Ltd, 2005
J. Sachs, B. S. Khurana & P. Mahonen. (2006). Evaluation of Handover Performance for TCP
Traffic Based on Generic Link Layer Context Transfer, IEEEInternationalSymposium
on Personal, Indoor and Mobile dio Communications,
Ra 206(PIMRC’06), pp: 1-5, Sept.
2006
M. Afif, P. Martins, S. Tabbane & P. Godlewski. (2006). SCTP Extension for EGPRS/WLAN
Handover Data, st 13 IEEEConferenceonLocalComputerNetworks , pp: 746-750, Nov.
2006
N. Dailly, P. Martins & P. Godlewski. (2006). Performance evaluation of L2 handover
Mechanisms for Inter-Radio Access Networks, IEEEVehicularTechnologyConference,
206(VTC2-Spring)
06 , pp: 491-495, May 7-10, 2006
N. Vulic, I. Niemegeers & S.H de Groot. (2004). Architectural Options for the WLAN
Integration at the UMTS Radio Access level, IEEE Vehicular Technology Conference,
204(VTC4-Spring)
20 , pp: 3009-3013, May 17-19, 2004
R. Koodli. (2005). Fast Handovers for Mobile IPv6, http://www.ietf.org/rfc/rfc4068.txt,
IETF, July 2005
S.L. Tsao & C.C. Lin. (2002). Design and Evaluation of UMTS/WLAN interworking
Strategies, IEEE Vehicular Technology Conference, 20 (VTC202-Fall), pp: 777-781,
2002
MHD-CAR: A Distributed Cross-Layer Solution
for Augmenting Seamless Mobility Management Protocols 551
MHD-CAR: A Distributed Cross-Layer

Solution for Augmenting Seamless
Mobility Management Protocols
Faqir Zarrar Yousaf, Christian Müller and Christian Wietfeld
CommunicationNetworksInstitute
DortmundUniversityofTechnology(TUDortmund),
Germany
1. Introduction
The Next Generation Network (NGN) architecture is evolving into a highly heterogeneous
infrastructure composed of a variety of Wireless Access Technologies (WAT). In such a
technologically diverse network; one of the key challenges would be to ensure ubiquitous
communication services to mobile entities, with varying mobility patterns and speed,
irrespective of their location and/or the underlying WAT. This calls for devising efficient
mobility management solutions that would provide location management and handover
management services to mobile entities. The location management service is responsible for
keeping track of the location of the mobile entities whereas the handover management
service enables the mobile entity to change its point of connection in the Internet. The
essential mandate of any efficient mobility management solution is to provide effective and
fast location monitoring and updating services while enabling seamless inter-WAT
handover. The notion of seamless handover implies handovers with minimum latency and
packet losses.
Providing seamless handover is an imposing challenge because the location update takes
place after the successful execution of the handover. The handover process is executed at
both the data link layer (L2) and at the network layer (L3) based on prescribed rules at these
respective layers. The L2 handover process enables the mobile node (MN) to switch its link
connectivity from its serving access point (AP) or base station to the new one. After
successfully establishing link connectivity, the MN will then need to perform L3 handover
process to make the MN IP capable on the new link. Till the completion of the handover,
the MN is practically disconnected from the network resulting in loss of data. The amount of
data lost depends on the handover latency, which in turn is a sum of delay incurred by
handover procedure prescribed at L2 and L3 respectively. This implies that the latency of
the L3 handover process is directly dependent on the latency of the L2 handover process
and also on the timing of the provisioning of L2 triggers that will initiate the L3 handover
process.
Thus to develop seamless handover methodology, it is important to take into account the
effect of L2 handover process on the L3 specified handover operations. This calls for
devising cross-layer mobility management that will enable inter-layer communication
solution
(i.e., between L2 and L3) of critical information that will enhance the performance of the
overall handover process.
In this chapter we present the details of one such cross-layer solution called Multi-Hop
Discovery of Candidate Access Router MHD- ( that
CAR)not only optimises the standard
Candidate Access Router Discovery (CARD) protocol (Liebsch et al., 2005) but it also offers
inherent cross-layer capabilities that can potentially contribute towards achieving low
latency handovers and hence enhance the operational efficiency of seamless mobility
management protocols in general.
The details of this novel solution will be presented in the context of Fast Mobile IPv6
protocol. A portion of the work presented in this chapter is based on our previous efforts
that has been recorded and published in (Yousaf et al., 2008(b)).
2. Technical Background
To provide mobility management to MNs, IETF has specified Mobile IPv6 (MIPv6) protocol
at L3 (Perkins et al., 2004). However, it is an established fact that the handover performance
of MIPv6 is not seamless, in that it incurs a high handover latency and packet delay. This is
because the MIPv6 operation is based on a break-before-make operation in which the MN will
initiate the MIPv6 protocol after it has disconnected from its serving AP as it moves out of
its coverage range. After the disconnection, the MN will search and connect to the
appropriate new in-range wireless AP. After establishing link connectivity (or performing
L2 handover), the MIPv6 handover process will be initiated which is based on a sequential
execution of a series of sub-processes; namely Care-of-Address configuration, Duplicate
Address Detection (DAD) test, Home Registration, Return Routability Test and
Correspondent Registration. Each of the sub-process will incur a finite amount of delay
(DAD test alone incurs a delay of 1 sec) contributing thereby to the total handover latency.
During the execution of the MIPv6 protocol, the MN remains disconnected from the Internet
and is unable to transmit or receive packets resulting in data losses. Hence, the MIPv6
handover latency, which is in excess of 1.5 seconds (Yousaf et al., 2008(c)), is unsuitable for
delay sensitive and throughput sensitive applications.
To provide seamless handover services, IETF has specified Fast Mobile IPv6 (FMIPv6)
protocol in RFC 5268 (Koodli, 2008) which extends the standard MIPv6 protocol. The main
operational concept of FMIPv6 is based on the ability of the MN to detect and negotiate a
handover with the New Access Router (NAR) in advance while the MN is still connected to
its Present Access Router (PAR). During handover negotiation with NAR, a bi-directional
tunnel is established between the PAR and NAR so that packets arriving at the PAR (and
destined towards the MN) are tunnelled towards NAR where they will get buffered. These
buffered packets will get forwarded to the MN soon after it establishes link connectivity
with the AP associated with the NAR and becomes IP capable on the NAR’s link. In other
words, FMIPv6 is based on a make-before-break concept which not only reduces handover
latency but also the packet loss by virtue of the tunnelling and buffering of packets.
However, the key to the success for the FMIPv6 handover operation is the ability of the MN
to detect and identify the presence of in-range Candidate Access Routers (CARs) and then
select an appropriate NAR from amongst the identified CARs. The FMIPv6 protocol
specification only specifies the handover operation by assuming that the MN has already
identified and selected a suitable NAR and hence does not provide any specific mechanism
to this end. The method of discovering CARs and selecting NAR is left to the discretion of
the user.
Candidate Access Router Discovery CARD) ( is one such standard solution (Liebsch et al., 2005)
the protocol facilities of which can be utilised by FMIPv6 to enable a MN to discover CARs
and hence select NAR.
The summary of the CARD protocol and its interaction with the protocol operation of
FMIPv6 is given in the following sub-section.
2.1 CARD Protocol

The CARD protocol is a standard IETF solution the operational details of which are
specified in RFC 4066. The protocol provides a generic mechanism that allows a MN to
acquire the necessary and relevant information about the ARs that are potential candidates
for the MN’s next handover. It is based on the exchange of a series of request and reply
messages between the MN and its serving AR and also amongst ARs as well. The messages
exchanged between the MN and its serving AR (i.e., PAR) is designated as MN-AR CARD
Request and MN-AR CARD Reply message, whereas those exch anged between the ARs are
termed as AR-AR CARD Request and AR-AR CARD Reply message. These messages are
transported as options inside the ICMPv6 message. The format of the CARD Request and
CARD Reply message is illustrated in Figure 1 and 2 respectively.
Fig. 1. Format of the CARD Request Message Carried as an Option in ICMPv6 Message
Fig. 2. Format of the CARD Reply Message Carried as an Option in ICMPv6 Message
The CARD protocol is designed to perform the following two functions namely;
1. Reverse Address Translation (RAT)
2. Discovery of CAR Capabilities (DCC)
The RAT function enables a MN to map the L2 identifiers (L2-IDs) (e.g., a MAC address for
802.11 networks) of one or more in-range APs to the IP address (L3-IDs) of the associated
CAR connected to it. The L2-IDs are typically discovered during the scan operation initiated
by the MN as a reaction to the link condition going below a certain specified threshold value
of SNR or RSSI.
The DCC function on the other hand, allows the MN to acquire the capability’s information
of the discovered CARs. The notion of capabilities implies of the various QoS aspects offered
by a CAR that would then be used as input to the MN’s target AR (TAR) selection algorithm
to make optimal handover decisions. The DCC function will prevent the MN to make
inaccurate network selections and hence connections to the wrong one in case of many
available CARs.
Central to the CARD operation is a L2-L3 address mapping Table called a CARTable , which
is managed and maintained inside each AR. The information content of the CAR Tableis
used to resolve the L2-IDs of a Candidate AP (CAP) to the IP address and capabilities of the
associated CAR.
RFC 4066 suggests the use of a central entity called a CARD Serveras one of the strategy to
populate the CAR Table . During boot up time all ARs within the administrative domain of
the CARD Server will register their IP addresses and th e L2-IDs of the associated CAPs with
the CARDServer . The CARDServerwill then be queried during the RAT process.
The functionality of the CARD protocol extends many benefits which are outlined in
(Trossen et al., 2002). For example, the information that the MN acquires due to the CARD
operation will enable it to select and connect to an appropriate network that would provide
the necessary service to the MN based on its application requirements. This can be beneficial
to multi-homed MNs as it may also enable the MN to select the least-cost network and
was first presented in (Funato et al., 2002) which has been adopted by RFC 4066 as one of
the probable method.
The current AR, depending on the status of the C-flag (capabilities request flag) in the MN-
AR CARD Request message, will then directly contact the resolved CAR(s) and perform
capabilities discovery via AR-AR CARD Request/Reply message pair and then send the
resolved identities and capabilities of the CAR(s) to the MN via a MN-AR CARD Reply
message. It may be mentioned that the identity and capabilities information are carried in
specified message sub-options and containers the details and format of which is given in
(Liebsch et al., 2005). Based on the capabilities information and preset criteria, the MN will
select an appropriate TAR to which it will perform a handover with. This process will ensue
every time the handover is imminent.
Fig. 3. The CARD Protocol Operation
Upon connecting to the target AR the MN will send a wildcard MN-AR CARD Requestto
obtain the information of its CAR Table in order to improve the prospects of the next CAR
discovery. It is observed that the maintenance and management of local CAR Tables is
critical to the effectiveness of CARD protocol in terms of assisting seamless and fast
handover by way of quick address resolution and capabilities discovery.
2.2 FMIPv6 Operation with CARD Protocol

Figure 4 illustrates the FMIPv6 protocol operation in conjunction with the CARD protocol.
The only difference is that the MN-AR CARD Request and MN-AR CARD Reply message are
piggybacked on the FMIPv6 protocol specified Router Solicitation for Proxyand (RtSolPr)
Proxy Router Advertisement (PrRtAdv) messages respectively (Liebsch et al., 2005). The
RtSolPr and PrRtrAdv are ICMPv6 type messages the format of which is specified in (Koodli,
2008).
The MN will start scanning for in-range APs in response to deteriorating link conditions. The
MN will then send the L2-ID(s) of the scanned AP(s) to the upper layer (i.e., L3) in the form of L2
trigger which will initiate the CARD operation as described in Section 2.1.1. The PAR after
acquiring the identities and capabilities of the available CAR(s) against the L2-ID(s) provided by
the MN in the RtSolPr message will send this information as [L2-ID, AR-Info]1 tuple appended to
the PrRtAdv message. The MN will then select a suitable NAR from amongst the discovered
CAR(s) based on some TAR selection criteria which is beyond the scope of this chapter.
Based on the identity information of NAR, the MN will auto-configure (Thomson et al.,
2007) a prospective New Care of Address (pNCoA) and will send this in a Fast Binding
Update Message FBU) ( message to the PAR indicating the NAR to which the MN wishes to
handover its connection. The PAR will then notify the NAR of the pCoA and request for a
handover by sending a Handover Initiate message.
HI)
( The NAR in response will
acknowledge the handover request by transmitting a Handover Acknowledge (HAck) message
towards the PAR. The PAR upon receiving the HAck will immediately set up a forwarding
tunnel with the NAR (referred to as PAR-NAR tunnel) and will start tunnelling subsequent
packets destined for the MN towards NAR where they will get buffered. In the meantime
the PAR, upon processing the HAck, will inform the MN of the NAR’s decision via a Fast
Binding Acknowledgement (FBAck) message. It should be noted that all this operation is
executed while the MN is still connected to the PAR.
Fig. 4. The FMIPv6 Predictive Handover Operation Utilizing the CARD Protocol for NAR
Discovery
1 The AR-Info corresponds to the identity and capabilities of an AR.

After the MN moves out of the communication range of PAR, it will establish link
connectivity with the New Access Point (NAP) and will announce its presence to the NAR
by sending an UnsolicitedNeighbourAdvertisement message(UNA)
(Narten et al., 2007) to the
NAR. The NAR will immediately forward all the buffered packets, and the subsequent
packets arriving via the PAR-NAR tunnel, towards the MN. The MN will then inform the
HA and the CN(s) of its new location as per the MIPv6 protocol rules (Perkins et al., 2004),
after the completion of which the handover is said to be complete.
It should be noted that the FMIPv6 protocol specifies two handover modes namely Predictive
Handover Mode and Reactive Handover , depending
Mode on whether the MN receives the
FBAck on the PAR’s link or not. Of the two modes, the Predictive Handover Mode is more
seamless and hence is the preferred and default mode.
Besides FMIPv6, the CARD protocol functionality can also be used by the MIPv6 protocol to
facilitate the MN’s decision to select the appropriate AR for handover that would best serve
its QoS requirements. For further details see (Liebsch et al., 2005).
3. Problem Statement
As described previously, central to the success of the FMIPv6 protocol is the ability of the
MN to discover and select NAR using the CARD protocol facilities. However the process to
discover and select NAR depends on two discovery processes namely:
1. CAP discovery, and
2. CAR discovery.
Both of these processes will influence the overall discovery process in consideration of the
inherent process limitations described below.
Candidate Access Point (CAP) Discovery Delay:

The CAP discovery process is undertaken by the technique specified for the underlying
WAT, whereas the CAR discovery process is carried out by the CARD protocol after the L2
provides it with the identities of the discovered CAP(s) using specific L2 constructs such as
triggers, events, hints etc.
Since the CARD protocol, and hence the FMIPv6 protocol, rely on the timely and accurate
provisioning of these L2 constructs, therefore the performance inadequacies of the L2
specified operations will have adverse consequences on the performance of the CARD and
thus the FMIPv6 protocol. For example, in reference to the IEEE 802.11 WLAN, the MN is
required to perform scan operations (active or passive) to determine the presence of in-
range CAP(s). However, as pointed out in (Mishra et al., 2003), the scan operation accounts
for almost 90% of the L2 handover delay and can be approximately up to 400 ms. It should
also be noted that during the scan operation, the MN is unable to transmit or receive packets
resulting in data loss. Besides packet loss, the CAP discovery delay will translate into the
delayed provisioning of L2 triggers which in turn will delay the initiation of the CARD
protocol.
Thus in order to reduce the packet loss and ensure the timely initiation of the CARD
protocol, the duration of the scan operation must be reduced.
Candidate Access Router (CAR) Discovery Delay:

The CAR discovery process is undertaken by the L3 specified CARD protocol after receiving
L2 triggers. As described, the CARD protocol performs RAT and DCC function by the
exchange of CARDRequest/Replymessages. This incurs a high signalling cost especially over
the error prone radio links. Besides signalling cost, each function incurs a finite amount of
delay which is also influenced by the location of the CARD Server and its topological
distance from the PAR. Besides influencing the delay, the CARD Server is a central network
entity that must be managed and maintained thereby increasing the cost of network
management. The CARD Server also introduces a potential single-point-of-failure, the
failure of which will result in a failed handover.
In consideration of the above performance issues, a unified solution approach is desired that
must incorporate the following recommendations;
1. Minimize the duration of the L2 scan operation, and hence the CAP discovery delay.
2. Ensure the timely delivery of L2 triggers to initiate the CARD process.
3. Remove the dependence of the CARD protocol on a central CARD Server.
4. Reduce the signalling load of the CARD protocol, especially over the error
prone wireless link.
This implies a tightly coupled liaison between L2 and L3 operations and calls for a cross-
layer management solution that must incorporate the above recommendations to enable a
MN to discover NAR in the shortest possible time with low signalling latency and high
probability of success.
4. Related Work
Two approaches for discovering CARs worth mentioning is Push-Mode-Multicast based
Candidate Access Router Discovery (PMM CARD) (Dario et al., 2006(a)) and Access Router
Information Protocol (ARIP) (Kwon et al., 2005). PMM CARD introduces added complexity of
maintaining and managing multicast groups and addresses and its performance is restricted
within a single operator's domain, whereas ARIP does not scale to complex network
architectures and is not dynamic. Beside introducing additional signalling messages, ARIP
requires the ARs to maintain identity information of the adjacent AR's but , similar to (Ono et
al., 2003), suggests manual set up by the network administrator or by the automatic learning of
the AR's from the handover information offered by the MNs. This makes ARIP unscalable to
complex network architectures. Also both the above proposals do not provide any cross-layer
management capabilities and have not been designed keeping in view the requirements and
dynamics of a fast moving MN in the context of NGN.
In this chapter we present an enhanced and scalable mechanism for discovering CARs that
can enable a fast moving MN to discover the identity and capabilities of the CARs which
may not be geographically adjacent and/or directly linked to the PAR. We term this new
approach as Multi-hop Discovery of Candid ate Access Routers (MHD-CAR), which is a simple
approach that does away with the complexity and limitations of both PMM CARD and
ARIP. The MHD-CAR provides an inherent scalable cross-layer mobility management
solution that eliminates the performance issues discussed earlier by incorporating the
solution recommendations.
The protocol details of the MHD-CAR protocol is submitted to the IETF as an Internet Draft
(Yousaf, Wietfeld, 2008(a)) and the proof of concept presented in (Yousaf et al., 2008(b)).
5. MHD-CAR Operation Summary

MHD-CAR is a distributed mechanism proposed to enhance the operational reliability and
robustness of seamless and fast handover protocols without introducing any additional
message(s) and/or relying on any central server.
In MHD-CAR ARs dynamically update their local CAR Table with the identity information
of not only the neighbouring ARs but also of CARs located multiple wireless-hops away
through an iterative exchange of unsolicited AR-AR CARD Reply message and without
relying on a CARD Server.
The CAR Table information of the current AR is then transferred to a MN on the fly where it
updates/refreshes a local cache called New Access Network (NAN) Cache allowing the MN to
resolve the CAR(s) locally with minimum exchange of Request/Reply messages over the
error prone radio link.
The MHD-CAR protocol operation is depicted in Figure 5 and the functional details are
discussed in the subsequent sub-sections.
Fig. 5. MHD-CAR Protocol Message Sequence Diagram
5.1 CAR Table Initialization

The composition of the CAR Table in the context of MHD-CAR is different from the one
proposed in (Liebsch et al., 2005) in that it offers more detailed information content. Table 1
shows the conceptual design of CAR Table in the context of MHD-CAR that contains not
only the identities of the CAP/CAR but also information regarding the type of wireless
access technology and the channel number in use by a CAP. It also informs about the
capabilities of the CAP in terms of supported bit-rate and SSID (in case of 802.11), and most
importantly the 'Distance' parameter, which is a measure of the distance of a CAR, in terms
of the number of wireless-hops, with reference to the local AR maintaining the CAR Table.
At initialization, each AR will populate its CAR Table with its own (and associated AP(s))
identity and capabilities information and set the Distance'
' parameter to zero, indicating
local AR information.
For the MHD-CAR operation, it is imperative that each AR should be aware of the identity of
the neighbouring AR and the associated AP(s) and store this information in its local CAR
Table with a Distance value set to 1. The neighbour association can be established
dynamically using the handover information of a bootstrapping MN as an input to establish
a neighbour relationship. The first handover between any two neighbouring ARs will serve
as a bootstrapping handover that would invoke the discovery process between the two ARs.
This idea was first presented in (Shim, Gitlin, 2000), (Trossen et al., 2003) and also endorsed
by the official CARD protocol standard (Liebsch et al., 2005) , which is adopted by the
MHD-CAR operation for CAR Table initialisation and described as follows.
CAR Table
AP Information AR Information Capabilities
macaddrMAC Address IP Address
ipaddr Bitrate AP
double
int L2 Type Network Prefix
prefix SSID
string
int Channel Number int Prefix Length int Distance
User Defined QoS AV Pair
avpair
Table 1. Conceptual Design of the CAR Table
When some MN performs an inter-AR handover, it will inform its current AR about the
identity of the previous AR using a Router Identity message option appended to the MN-AR
CARD Requestmessage (Liebsch et al., 2005). The serving AR will acknowledge this with a
MN-ARCARDReply message and will store the identity of the MN’s previous AR in its CAR
Table and indicate it as its immediate neighbour. The serving AR will then send an
unsolicited AR-AR CARD Reply message to the previous AR informing it of its own identity
and identifying itself as its neighbour. The previous AR will thus store the identity
information of the MN’s current AR in its local CAR Table as an immediate neighbour
indicated by the Distance value of 1. In this way, all the ARs along the motion path of the
MN will bootstrap their local CAR Tables with the identity of the neighbouring ARs.
Besides the identity information, the two ARs must also exchange the capabilities
information with each other. This process will also eliminate the reliance on maintaining a
CARD Server. It may be noted that the identity information contains both the IP address of
the AR and the L2-Id(s) of the associated AP(s).
5.2 CAR Table Distribution

After the ARs are bootstrapped with the identity and capabilities of the neighbouring ARs,
each AR will exchange its local CAR Table information with the neighbouring AR(s)
through the iterative exchange of unsolicitedAR-ARCARDReplyMessage , where the number
of iterations is equal to the specified maximum distance, in wireless-hops, corresponding to
the maximum entries an AR will maintain in its CAR Table.
In the first iteration, the ARs will exchange their local CAR Table information (see Table 1)
with their neighbouring ARs, which will add this new information to their local CAR Tables
and increment the Distance value by 1. Now each AR will have information about the AR(s)
two wireless-hops away and this will be indicated by the Distance

value of 2. It should be noted
thattheARsdonotforwardthereceived messagesinter-
inorder
AR topreventnetwork . flooding
During the second iteration, the above process will be repeated and the receiving ARs will
compare the new information with their present CAR Table entries and if no match is found,
it will add the new CAR information to its local CAR Table by incrementing the distance
parameter by 1.
Fig. 6. Conceptual Representation of the Iterative CAR Table Distribution Process in MHD-
CAR
This new entry will thus be marked with the Distance value of 3, indicating that the
corresponding CAR is at distance of three wireless-hops.
The above inter-exchange of unsolicited inter-AR Reply messages will continue until each
AR has the information about CAR, which is at the specified maximum distance from the
current AR, after which the CAR Tables are said to have converged. The maximum distance
for which an AR is supposed to maintain CAR information is a constant that depends on the
network topology and can be specified by the administrator. The ARs will then periodically
exchange their local CAR Table information with their neighbouring AR(s) after every 60
seconds, or when there is some change in the contents of it (for example, change in
capabilities information). During CAR Table initialisation d distribution
an process, the ARs will
transmit the unsolicited inter-AR uniformly
messages distributed
at a random time between 0 an
0msec
1 .
The iterative CAR Table distribution process is illustrated in Figure 6, which shows the
contents of the CAR Table of an AR for each iteration of the distribution process. The
distribution takes place for a Distance value of 4, i.e., the iterative distribution process will
continue till each AR has information about ARs which are up to 4 wireless hops away.
Figure 6 illustrates the process for the reference topology shown in Figure 7 composed of 12
Access Networks (AN) (from A to L ), where the domain of each AN is defined by an AR and
an associated AP. For the sake of demonstration and simplicity only the AN Identity and
Distance parameters of a CAR Table are considered in Figure 6. The identity (Id) is
characterised by the [L3-ID, L2-ID] pair and is denoted by the AN identifier, whereas the
Distance (D) signifies the topological distance of a CAR/CAP from the local AR in terms of
wireless hops. The entries indicated in red signify the new entries that get stored in the CAR
Table during the particular distribution iteration. Iteration # 0 signifies the contents of the
CAR Table during the initialisation process as explained above.
Figure 7 clearly shows how the ARs (e.g., F) acquire the information of CARs (for instance B
& J) that are not immediate neighbours through the iterative distribution of the CAR Tables.
NAN Cache
CAP Information CAR Information
boolReachability
macaddr MAC Address
ipaddrIP Address
int L2 Type
Network Prefix
prefix
int Channel Number
int Prefix Length
doubleBitrate
int Distance
doubleLast Received Beacon time
doubleRSSI
SSID
string
Table 2. Conceptual Design of New Access Network (NAN) Cache
5.3 Mobile Node Operation

The CARD protocol specification (Liebsch et al., 2005) suggests a MN to maintain address
and capability information of CAR(s) discovered during previous CARD operation in a local
cache to avoid requesting the same information repeatedly and to select an appropriate TAR
as quickly as possible when a handover is imminent, but it does not specify the conceptual
design of such a cache. Besides, this proposal will only improve the CAR selection if the MN
is revisiting some previously visited CAR domain, a situation not very much likely in case of
high speed MNs.
MHD-CAR proposes a MN to maintain a local cache called New Access Network NAN) ( cache
(see Table 2) that maintains the identity and capabilities information of not only the
neighbouring CAR(s) and associated CAP(s) but those located multiple wireless-hops away.
This will allow the MN to perform RAT and DCC functions locally without the exchange of
any MN-AR CARD Request/Reply message pair with its current AR, or without involving a
CARD server, every time handover is imminent.
The information content of the NAN cache, that is expected to enhance the MN's TAR
selection process and handover related decision tasks, is derived mostly from the CAR Table
that is usually pulled by the MN on the fly from its current AR (via a wildcard MN-AR
CARD RequestMessage) when the MN senses that it is moving away from its current AR.
The MN will then add and/or refresh the relevant entries of its cache. The NAN cache also
derives some of the AP related information from the periodic beacon signals received from
the in-range wireless APs.
Fig. 7. Simulation Topology
In this Section we present the results of the simulation experiments comparing and
analyzing the performance of the proposed MHD-CAR mechanism to that of the IETF's
CARD protocol using the CARD server. Both the protocols are modeled in our mobility
management framework (Yousaf et al., 2008(c)) developed in OMNeT++ (OMNeT++) and
using realistic message structures and timer implementations.
The simulation experiments and its analysis is similar to the one presented by us in
(Yousaf et al., 2008(b)) with the difference that in this chapter we have extended our
simulation model to realise a more realistic hierarchical network instead of a flat
topology
networktopology in which all the ARs were directly connected to their immediate neighbours
via Ethernet links. In a flat network topology, the ARs will be able to derive the identity of
their neighbouring ARs by simply sending relevant messages directly and hence populate
their local CAR Tables. In contrast, to realise the MHD-CAR protocols in the hierarchical
network topology, we have extended our simulation framework with the bootstrapping
mechanism (see Section 4.1) that would enable the ARs to discover the IP address of the
neighbouring ARs. With this major difference, the same experiments were repeated and
found the results to match those presented previously in (Yousaf et al., 2008(b)).
The simulation network topology is shown in Figure 7 and the experiments are carried in a
homogeneous 802.11b wireless environment using a free space propagation model at 2.4
GHz for a radio channel over a total coverage area of 800m x 800m. To initialise the CAR
Tables, in ARs a bootstrapping MN is moved across the reference network at the beginning
of the simulation enabling each AR to become aware of the IP address of the previous
neighbouring AR. The results expressed in this Section will also apply equally to a
heterogeneous environment because the MHD-CAR operation is defined at the network
layer.
(a) AR-AR Message Throughput in CARD
(b) AR-AR Message Throughput in MHD-CAR

Fig. 8. Inter AR Message Throughput in (a) CARD, and (b) MHD-CAR
A single simulation run consists of a MN moving across 12 ARs, starting from ARA and
undergoing 11 handover instances, discovering and resolving the next CAP(s)/CAR(s)
while it is still connected to its current AR. The experiments are repeated 100 times for each
RIWKH VHHGYDOXHVJHQHUDWHGXVLQJWKH201H7V VHGWRRO

WKHUHE\FRQVWLWXWLQJDWRWDORI
UXQV HDFK IRU VLPXODWLQJ &$5’ DQG 0+’&$5 SURWRFRO 7KH UHVXOWV DUH WKHQ
H[SUHVVHGDVDYHUDJHVXPRIWKHPHDVXUHGSDUDPHWHUV
6.1 Impact of MHD-CAR on Signaling Load

7KHDYHUDJHVLJQDOOLQJORDGRYHUWKHZLUHOH VVOLQN01$50HVVDJHVDQGWKHZLUHGOLQNV
$5$5 0HVVDJHV IRU ERWK WKH &$5’ DQG 0+ ’&$5 LV FRPSDUHG LQ )LJXUH DQG WKH
UHVXOWV WDEXODWHG LQ 7DEOH 7KH $5V LQ ERWK &$5’ DQG 0+’&$5 SHULRGLFDOO\ XSGDWH
WKHLU&$57DEOHVDIWHUVHFRQGVDQGLQRXUVLPX ODWLRQWKLVXSGDWHWDNHVSODFHWKUHHWLPH
6.1.1 Signaling load over the Inter-AR links

,Q &$5’ HDFK $5 GXULQJ ERRWXS WLPH UHJLVWHU LWV LGHQWLW\ LQIRUPDWLRQ ZLWK WKH &$5’
VHUYHUDQGNHHSVWKH&$5’VHUYHUXSGDWHGE\VHQGLQJ XQVROLFLWHG$56HUYHU5HSO\0HVVDJHV
SHULRGLFDOO\ DIWHU HYHU\ VHFRQGV 7KH QXPEHU XQVROLFLWHG
RI $56HUYHU 5HSO\ 0HVVDJHV
WUDQVPLWWHG GXULQJ ERRWXS WLPH QJ DQGSHULRGLF
GXUL XSGDWHV UHPDLQV HTXDO DQG
FRUUHVSRQGV WR WKH QXPEHU RI $5V ,Q EHWZHHQ WKH SHULRGLF XSGDWHV WKH FXUUHQW $5 ZLOO
H[FKDQJH$5$5 &$5’ 5HTXHVW5HSO\ PHVVDJH SDLU ZLWK WKH 5’&$
6HUYHU IRU SHUIRUPLQJ
5$7IXQFWLRQXSRQ01VUHTXHVWDQG $5$5&$5’5HTXHVW5HSO\ PHVVDJHSDLUZLWKWKH
UHVROYHG &$5V IRU SHUIRUPLQJ WKH ’&& IXQFWLRQ 7KLV LV VHHQ LQ )LJXUH D ZKHUH ZH
VKRZWKHRFFXUUHQFHRIWKHDYHUDJHLQWHU$5 6HUYHU&$5UHSO\PHVVDJHWKURXJKSXW
,Q0+’&$5KRZHYHUVLQFHWKH01LVDEOHWRORFDOO\UHVROYHWKHLGHQWLW\DQGFDSDELOLW
RI&$5VEDVHGRQWKHLQIRUPDWLRQLQ1$1&DFK HVHH6HFWLRQWKH$5GRHVQRWQHHGWR
H[FKDQJHDQ\UHTXHVWUHSO\PHVVDJHSDLUZLWKVRPHVHUYHURUQHLJKERXULQJ$5VWKHUHE
UHVXOWLQJLQ UHGXFWLRQRILQWHU$5VLJQD OOLQJORDGUHODWHGWRGLVFRYHULQJ&$5V7KLV
LV GHSLFWHG LQ )LJXUHZKHUH E WKH $5V RQO\ H[FKDQJHXQVROLFLWHG $5$5 5HSO\ PHVVDJHV
ZLWK WKHLU QHLJKERXULQJ $5V DV SDUW SHULRGLF
RI WKH XSGDWH RI WKH ORFDO &$5 7DEOH DIWHU
HYHU\VHFRQGV
Messages
MN-AR AR-AR 2
Requests Replies Requests Replies 3
CARD
MHD-CAR
Message Load (%)
Total Load Reduction (%)
7DEOH01$5$5$50HVVDJH/RDG&RPSDULVRQ
6LQFH WKHUH LV QR FHQWUDO VHUYHU WKH

XSGDWH
LQWHU
PHVVDJH
$5 WKURXJKSXW LQ 0+’&$5 LV
KLJKHU WKDQ &$5’ EXW RYHUDOO UHFRUGLQJ D UHGXFWLRQ LQ WKHSRVW ERRWXS
H[FKDQJHRILQWHU$5PHVVDJHVHH7DEOH
7KHLQWHU$5PHVVDJHVDUHLQFOXVLYHRI
WKH$56HUYHUPHVVDJHVIRUWKH&$5’RSHUDWLRQ
7KH $5$5 5HSO\ PHVVDJHV WDNH LQWR DFFRXQW
RQO\ WKH UHSOLHV PHVVDJHV H[FKDQJHG DIWHU W
LQLWLDOL]DWLRQDQGGLVWULEXWLRQSURFHVV
Nasir, A., & Mah-Rukh. (2006). Internet Mobility using SIP and MIP. In Proceedings of the
Third International Conference on Information Technology: (ppNew 334 Generation
-
339).
Perkins, C. (1997). Mobile IP. IEEECommunicationsMagazine , 35 (5), 84-99.
Perkins, C. (1998). Mobile Networking through Mobile IP. IEEEInternetComputing, (1),2 58-
69.
Perkins, C. (2002). IP Mobility Support for IPv4. Request for Comments RFC) ( 43 , Available
at: www.ietf.org/rfc/rfc3344.txt.
Prado, R., Zagari, E., Cardozo, E., Magalhaes, M., Badan, T., Carrilho, J., Pinto, R.,
Berenguel, A., Barboza, D., Moraes, D., Johnson, T., & Westberg, L. (2008). A
Reference Architecture for Micro-Mobility Support in IP Networks, Proceedings of
theThirteenthIEEESymposiumonComputersandCommunications (pp. 624 – 630).
Rekhter, Y., & Rosen, E. (2001). Carrying Label Information in BGP-4. Request for Comment
(RFC3107) , Available at: http://www.ietf.org/rfc/rfc3107.txt.
Ramjee, R., Porta, T. L., Thuel, S., Varadhan, K., & Wang S. (1999). Hawaii: A Domain-
Based Approach for Supporting Mobility in Wide-Area Wireless Network,
Proceedings the IEEEInternationalConferenceonNetworkProtocols (pp. 283-292).
of
Ren, Z., Tham, C.-K., Foo, C.-C., & Ko, C.-C. (2001). Integration of Mobile IP and Multi-
Protocol Label Switching. Proceedings of the IEEE International Conference o
Communications (pp. 2123–2127).
Rosen, E.,Viswanathan, A., & Callon, R. (2001). Multiprotocol Label Switching Architecture.
Request For Comments (RFC) 301 , Available at:
http://www.ietf.org/rfc/rfc3031.txt
Saha, D., Mukherjee, A., Misra, I., Chakraborty, M., & Subhash, N. (2004). Mobility Support
in IP: a Survey of Related Protocols. IEEENetwork, 18 (6), 34–40.
Salsano, S., Polidoro, A., Mingardi, C., Niccolini, S., & Veltri, L. (2008). Sip-based Mobility
Management in Next Generation Networks. IEEE WirelessCommunications , 15(2),
92–99.
Samprakou, I., Bouras, C., & Karoubalis, T. (2004). Fast and Efficient IP Handover in IEEE
802.11 Wireless LANs. Proceedings theof InternationalConferenceonWirelessNet
(pp. 249-255).
Samprakou, I., Bouras, C., & Karoubalis, T. (2007). Improvements on IP IAPP: A Fast IP
Handover Protocol for IEEE 802.11 Wireless and Mobile Clients. Wireless Network ,
13 (4), 497–510.
Schulzrinne, H., & Wedlund, E. (2000). Application-layer Mobility Using SIP. SIGMOBILE
MobileComputingandCommunications , 4 (3),Review
47–57.
Snoeren, A. C., & Balakrishnan, H. (2000). An End-to-end Approach to Host Mobility,
Proceedings the
of6th ACM/IEEE International Conference on Mobile Computing and
Networking (pp. 155-166).
Snoeren, A. C., Balakrishnan, H., & Kaashoek, M. F. (2002). The Migrate Approach to
Internet Mobility,ProceedingsofStudent the OxygenWorkshop (pp. 14-17).
Valko, A. G. (1999). Cellular IP: A New Approach to Internet Host Mobility. SIGCOMM
ComputerCommunication, Review 29 (1), 50-65.
Yang, T., & Makrakis, D. (2001). Hierarchical Mobile MPLS: Supporting Delay Sensitive
Applications Over Wireless Internet. Proceedings the
of International Conferences o
Info-techandInfo-(vol. net2, pp. 453-458).
number of MN-AR messages exchanged over the wireless link for both the protocols and
depicts the percentage reduction of message load induced by MHD-CAR with reference to
the scenario depicted in Figure 7.
6.2 Impact of MHD-CAR on Scanning Delay

It is a known fact that the L2 handover delay is a major delay component adding to the
overall handover latency for MIPv6 (Yousaf et al., 2008(c)) and thus a major impediment to
the performance of higher layer mobility management schemes. The main contributing
factor to the L2 handover latency is the delay incurred by the channel scan operation as part
of the CAP discovery process (see Section 3). Typically a MN after losing its connectivity
with its current AP will perform the 802.11 all-channel(active/passive)
scan on, in our case,
13 frequency channels. This scan-delay is certainly not suitable for attaining seamless
handover performance for fast moving MNs and various methods have been proposed in
this respect. One of the consensus solutions to reduce the L2 handover delay is to perform
scanning on selective channels (Park et al., 2004) or perform a pro-active scan, i.e., before a
MN loses its connection with its current AP (Haito et al., 2007).
Fig. 10. Signalling Load Comparison of CARD and MHD-CAR
The MHD-CAR based on the information available in the NAN Cache proposes to reduce
the scan delay by enabling a MN to perform targetscanning on selective frequency channels.
The MN, instead of waiting to disconnect from the current AP, will start to scan for only
those channels that correspond to the nearest CAP(s) as specified by the Distance parameter
in the NAN Cache, and if the CAP(s) are not located (or are out of range) will proceed to
select and scan the next set of frequency channels corresponding to next farther CAP(s). The
MN will typically start the scan process when the RSSI from the current AP falls below a
certain threshold. Figure 11 compares the performance of MHD-CAR's target scanning with
the 802.11 all-channel scan.
From the Figure 11 it is evident that the performance of MHD-CAR's target scanning incurs
less delay than the full channel scan, however the delay performance is a function of the size
in the form of defining and developing new SAPs and primitives that would interface with
the generic SAP and primitives defined for the IEEE 802.21 before a MN can take advantage
of the MIH services (Eastwood & Migaldi, 2008). This would entail a major re-engineering
effort involving the upgrade of the whole network infrastructure and protocol standards.
Besides, the IEEE 802.21 model diverges from the standard ISO/OSI reference protocol
model by introducing an intermediate layer between L2 and L3.
(a) (b)
Fig. 12. Conceptual Models of (a) The IEEE 802.21 MIH Service Model, and (b) The MHD-
CAR Protocol
Similar to IEEE 802.21, the operational scope of MHD-CAR is to provide a mechanism to

enable a multi-homed MN to undergo inter-RAT handover by providing the requisite
information content that would enable a MN to choose the best network that would suit its
application service requirements. However, in sharp contrast to IEEE 802.21 service model
depicted in Figure 12(a) which is implemented as a protocol stack, the MHD-CAR is based
on managing and maintaining a NAN Cache inside the MN that can be accessed by the
mobility functions defined at the network layer and the data link layer. This translates into
defining simpler interfaces at L2 and L3 for interaction with the NAN cache rather than
demanding major revisions from the access technologies as in the case of IEEE 802.21
highlighted above. The NAN Cache thus provides cross-layer capabilities.
In contrast to IEEE 802.21, the MHD-CAR protocol is simply an optimised version of the
existing CARD protocol without introducing any new messages, interfaces and/or network
entities, or deviating from the reference ISO/OSI reference model making it scalable and
deployable and without any burden on the network itself.
8. Conclusions
In this chapter we have provided operational and functional details of a proposed protocol
called MHD-CAR that has been designed in view of the stringent performance requirements
imposed by fast moving MNs in terms of seamless and fast handovers in a heterogeneous
wireless network environment. MHD-CAR optimises the standard CARD protocol in
enabling a MN to discover on the fly the identity and capabilities of not only the
neighbouring CARs but also CARs that may be located multiple hops away, and all this is
achieved with minimum reliance on the network. This is expected to augment the
performance of seamless handover protocols like FMIPv6 by ensuring accurate selection of
NAR with minimum discovery latency.
The MHD-CAR is a distributed mechanism in which the ARs are able to inter-communicate
their identities and capabilities information to neighbouring ARs and to ARs that are located
multiple wireless-hops away without relying on maintaining and managing a central CARD
server. Each AR stores this information in their local CAR Tables which are then
communicated to the MN upon request. Due to the distributed mechanism, MHD-CAR is
more efficient, reliable, survivable and scalable protocol than the CARD protocol. Since the
MHD-CAR does not introduce any new protocol messages it can therefore be easily
integrated into the present deployment infrastructure.
It exhibits far better performance over the IETF’s CARD protocol in terms of the substantial
reduction of signalling load over both the inter-AR links (by 17.5%) and crucially over the
error prone wireless link (by 48%), while utilizing the CARD protocol messages. This
reduction in signalling load is achieved because the MN is able to perform RAT and DCC
functions locally, based on the information content of the NAN cache, and without relying
on the network.
Another very important aspect of the MHD-CAR scheme is that it provides cross-layer
liaison between L2 and L3 mobility function. This is achieved by having a NAN cache in the
MN, which provides the MN with a topological snapshot of the identity and capabilities of
the access networks that may be multiple hops away from its present point of attachment.
This enables a MN to perform target scanning on selected channels greatly reducing the
CAP discovery latency and enhancing the accuracy of the TAR selection process. This alone
will have a direct impact on the overall handover latency and fast moving MNs will greatly
benefit from it.
In contrast to the IEEE 802.21 standard, it is observed that the MHD-CAR is a light weight
and much simpler alternative solution that provides the main functional services of the
802.21 MIHS. Although MHD-CAR has not been designed as an alternative to 802.21 but it
does share its motivational, operational and functional scope. The IEEE 802.21 WG was
developed to provide a unified global mechanism by defining a common MIH layer
sandwiched between the Network Layer and the Data Link Layer and defined common
triggers that would be generated independent of the underlying access technology. The
motivation was to enable the MN to make accurate selection of the network and to provide
triggers that would aid the IP mobility protocols like FMIPv6. However all this is being
introduced at the cost of high complexity while deviating from the base ISO/OSI prescribed
layered approach by introducing a new layer between the L2 and L3. Also it would mandate
changes to the existing access technologies to confirm to the MIHS scheme of signalling. For
example different SAPs are required to be defined for each of the access technology. It
would also involve the exchange of signalling message between the network and the MN,
even over the air interface.
MHD-CAR therefore provides the same conceptual functionalities defined for MIHS but
without transgressing the functional boundaries of the standard OSI/ISO protocol reference
model and with a much simpler and scalable approach.
9. References
Dario, D.; Femminella, M.; Piacentini, L. & Reali, G. (2006(a)). Performance Evaluation of the
Push-Mode-Multicast based Candidate Access Router Discovery (PMM CARD),
Computer Networks , Volume 50, No. 3, (February 2006) page numbers (367-397),
ISSN 1389-1286.
Dario, D.; Femminella, M.; Piacentini, L. & Reali, G. (2006(b)). Target Access Router
Selection in Advanced Mobility Scenarios. Computer Communications , Volume 29,
No. 3, (February 2006), page numbers (337-357), ISSN 0140-3664.
Eastwood, L. & Migaldi, S. (2008). Mobility Using IEEE 802.21 in a Heterogeneous IEEE
802.16/802.11 –Based, IMT-Advanced (4G) Network, IEEEWirelessCommunications ,
Vol. 15, No. 2, (April 2008) pp. 26-34, ISSN: 1536-1284.
Funato, D.; He, X.; Williams, C.; Takeshita, A.; Ali, M. & Nakfour, J. (2002). Geographically
Adjacent Access Router Discovery Protocol. draft-funato-seamoby-gaard-01.txt,
Internet Draft (IETF), June 2002. Work in Progress.
George, L.; Salkintzis, A. & Passas, N. (2008). Media-Independent Handover for Seamless
Service Provision in Heterogeneous Networks. IEEE Communication Magazine , Vol.
46, No. 1, (January 2008) pp. 64-71, ISSN 0163-6804.
Haito, W.; Tan, K.; Zhang, Y. & Zhang, Q. (2007). Proactive Scan: Fast Handoff with Smart
Triggers for 802.11 Wireless LAN, Proceedings of th 62 IEEE International Conference
on Computer Communications INFOCOM) ( , pp. 749-757, ISBN 1-4244-1047-9, Alaska
(USA), May 2007, IEEE, USA.
IEEE Std 802.21. Draft D9.0, IEEE Standard for Local & Metropolitan Area Networks: Media
Independent Handover Services.
Koodli, R. (2008). Mobile IPv6 Fast Handovers, Request for Comments (Proposed Standard)
5268, Internet Engineering Task Force, June 2008.
Kwon, D.; Kim, Y.; Bae, K. & Suh, Y. (2005). Access Router Information Protocol with
FMIPv6 for Efficient Handovers and their Implementations. Proceedings of Global
Telecommunications nference
Co , pp. 3814-3819, ISBN 0-7803-9414-3,
GlobeCom)
(
St. Louis (Missouri), IEEE, December 2005, USA.
Liebsch, M.; Singh A.; Chaskar, H.; Finato, D. & Shim, E. (2005). Candidate Access Router
Discovery Protocol (CARD), Request for Comments roposed
(P Standard) , Internet
406
Engineering Task Force, July 2005.
Mishra, A.; Shin, M. & Arbaugh, W (2003). An Empirical Analysis of the IEEE 802.11 MAC
Layer Handoff Process, Proceedings ACM
of SIGCOMM Computer Communication
Review , Volume 33, No. 2, pp. 93-102, ISSN 0146-4833, April 2003.
Narten, T.; Nordmark, E., Simpson, W. & H. Soliman, (2007). Neighbor Discovery for IP
Version 6 (IPv6)”, RequestforComments(DraftStandard) , September
4861 2007.
OMNeT++. Community Site, http://www.omnetpp.org, July 2009.
2QR 1 .LPXUD 7 )XMLL7 $6W XG\ RQ $XWRQRPRXV 1HLJKERXU $FFHVV 5RXWHU

’LVFRYHU\ IRU 0RELOH ,3
Positioning in Indoor Mobile Systems 601
5) Based on the received signal strength (Received Signal Strength Indication – RSSI) – The
free space signal propagation is characterised with predictable attenuation dependent on
the distance from the source. Moreover, in real conditions, the attenuation also largely
depends on the obstacles and the configuration of the propagation path. That is why there
are various mathematical models which describe the wave propagation for diverse
surroundings and, ultimately, estimate the signal attenuation for the observed environment.
This approach grants the distance of the entity whose position is being determined, to one
or more transmitters.
6) Based on the fingerprint of the location (Database Correlation or Location Fingerprinting)
– With this approach, the certain, location dependant, information is acquired in as many
Reference Points (RPs) across the coverage area of the technique. This data is stored into so
called Location Fingerprints Database. Afterwards, when the actual position determination
process takes place, the information gathered at the unknown location is compared with the
pre-stored data and the entity’s position is estimated at a location of a pre-stored fingerprint
from the database whose data are “closest” to the measured data.
Most often, the estimated position with TOA and RSSI approaches is determined by
lateration. The process of lateration consists of determining the position of the entity when
the distance between the entity and one or more points with identified positions (i.e.
refernce points) is known. To uniquely laterate the position in N-dimensional space, the
distances to N+1 reference points ought to be known. With TDOA approach, the estimated
position is obtained as a cross-section of two or more hyperbolas in two-dimensional space,
or three or more hyperbolic surfaces in case of three-dimensional space. The process of
angulation is employed with AOA and DOA approaches. This process estimates the
location of a user as a cross-section of at least two rays (half-lines) originating at known
locations. The lateration and angulation processes are depicted in Fig. 1. As for the Location
Fingerprinting approach, the estimated location is obtained by utilizing the correlation
algorithm of some sort. This algorithm determines, following a certain metric, the
“closeness” of the gathered data to the pre-stored samples from the location fingerprinting
database.
Apart from these, basic, approaches, there are a number of other choices and hybrid
techniques that combine the aforementioned approaches when determining the estimated
position of the user.
implementation on small mobile devices such as cell phones and handhelds unfeasible. In
fact, manufactures of such devices never considered supporting the present solutions
(Zagari et al, 2008).
Fig. 1. Current IP address not valid in a visiting network
After MIP, new protocols for mobility between networks in the same domain were
proposed. Micromobility protocols aim to improve localized mobility by reducing handover
overheads. Other approaches to allow seamless mobility also include mobility provided by
transport and session layer schemes.
The objective of this chapter is to provide a major review on mobility protocols and
architectures. The protocols and architectures will be classified according their mobility
range (intra and inter domain), layer in which it operates (from the link layer to the
application layer), and the support required from the mobile node. We will place emphasis
on the solutions called “network-centered”, that is, solutions where mobility is handled
entirely by the network without the need of installation of mobility protocols on the mobile
nodes. The protocols and architectures discussed in this chapter are being proposed by
standardization bodies, e.g., IETF, by industry-driven forums, e.g., 3GPP, by academy and
by the industry.
This chapter is divided into 10 more sections. Section 2 presents an overview of mobility
issues and a classification schema, which will be used in the protocols sections. Mobile IP is
presented in Section 3. Section 4 shows link layer based protocols. Section 5 presents
mobility solutions based on L2½ protocols. Section 6 presents network layer protocols.
Section 7 presents transport layer protocols. Section 8 presents mobility using application
layer protocols. Section 9 presents the Mobility Plane Architecture. Section 10 presents a
general classification of the mobility solutions seen in this chapter and future work related
to mobility in IP networks. Finally, Section 11 concludes this chapter.
Mobility in IP Networks: From Link Layer to Application Layer Protocols and Architectures 575
2. Mobility Issues
The mobility process starts with the mobile node’s attachment to a local wireless network.
The node attachment process happens when a MN enters in the coverage of a wireless
access point. In this point a L2 (Layer 2) attachment process is performed. After that, the
MN must acquire its IP address from the network. After obtaining its new address, the
network is able to route packets to/from the mobile node (Johnson et al, 2008).
When the MN moves away from the current access point, it may detect another wireless
access point. Handover is “the process by which an active MN changes its point of
attachment to the network, or when such a change is attempted. The access network may
provide features to minimize the interruption to sessions in progress" (Manner & Cojo,
2004) by preserving the transport (or higher layers) connections such a way the packets are
forwarded to the MN via the new access point.
Mobility management is the key to enable this seamless mobility. It enables wireless or
mobile networks to search and locate mobile devices for network communications and to
maintain network/applications connections as the MN moves into a new service area. The
mobility management is composed of mainly two services: location management and
handover management.
Location management consists of two operations: registration or location update and
paging, to enable a network to discover the current point of attachment of an MN for
information delivery (Saha et al, 2004). Location update is used in support of idle users, and
paging is used in support of active communications (Campbell et al, 2002).
Location update procedures need the MN to periodically inform the system to update
relevant location databases with its up-to-date location information (Akyildiz et al, 2004).
Paging is the ability to track idle mobile hosts. For protocols using this kind of tracking, idle
MNs do not have to register if they move within the same paging area, but only if they
change paging area (Campbell & Gomez-Castellanos, 2000).
Handover management enables the network to maintain a MN’s connection as it continues
to move and change its access point to the network (Saha et al, 2004). There are many types
of handover, among which are:
 horizontal and vertical: horizontal handover occurs between wireless cells of the
same technology; vertical handover occurs between two different networks of
different technologies. This chapter is dedicated to study horizontal handover
only.
 mobile and network initiated: in MN initiated handovers the MN is responsible for
initiating handover requests, while in network initiated handover the network is
responsible for indicating that a handover must occur.
 MN and network controlled: in MN controlled handover, the MN must participate
in the handover process, while in the network controlled handover the network
handles the entire process.
 fast: fast handover tries to reduce the latency during a handover.
 seamless: change the MN’s point of attachment to an IP-based network, without
losing ongoing connections and without disruptions in the communication.
The basic terminology for mobility is (Perkins, 2002; Manner & Kojo, 2004):
 home network: a network having a network prefix matching that of a mobile
node's permanent address;
 home address: the IP address acquired when registering in its home network; a
stable address that belongs to the mobile node and is used by correspondent nodes
to reach mobile nodes;
 home agent (HA): A router located on the home network that acts on behalf of the
mobile node while away from the home network;
 correspondent node (CN): Any node that communicates with the mobile node;
 foreign network: Any network (other than the home network) visited by a mobile
node;
 foreign agent (FA): A router located on the foreign network that acts on behalf of
the MN in this network;
 care-of-address (CoA): An address that is assigned to the mobile node when
located in a foreign link.
2.1 Classification parameters

Existing proposals for mobility can be broadly classified into different types, based on many
parameters. We will employ a taxonomy based on 4 axis: Mobility Range, Mobility Routing,
Mobility Signaling and Mobility Layer. In this section we will briefly introduce each one of
them.
2.1.1 Mobility Range

In this work we adopted the definitions by Manner & Kojo (2004) for the mobility scope.
Micromobility
Also called intradomain mobility or local mobility (Kempf, 2007), is the process of mobility
over a small area. Usually this means mobility within an IP domain with an emphasis on
support for active mode using handover, although it may include idle mode procedures
also. Micromobility protocols exploit the locality of movement by confining movement
related changes and signaling to the access network.
Macromobility
Also called interdomain mobility or global mobility (Kempf, 2007), is the process of mobility
over a large area. This includes mobility support and associated address registration
procedures that are needed when a MN moves between IP domains. Interdomain
handovers typically involve macromobility protocols. MIP can be seen as a means to
provide macro mobility.
2.1.2 Mobility Routing

Defines how the MNs’ location information database is created and maintained.
Routing based
Routing based schemes aim to exploit the robustness of conventional IP forwarding. A
distributed mobile host location database is created and maintained within the network
domain. The database consists of individual flat mobile-specific address lookup tables and
is maintained by all the mobility agents within the domain (Chiussi et al, 2002).
Tunnel based
In tunnel based schemes, the location database is maintained in distributed form by a set of
foreign agents in the access network. Each foreign agent reads the incoming packet's
original destination address and searches its visitor list for a corresponding entry. If the
entry exists then it contains the address of next lower level foreign agent.
The sequence of visitor list entries corresponding to a particular mobile host constitutes the
host's location information and determines the route taken by its downlink packets. Entries
are created and maintained by registration messages transmitted by mobile hosts (Campbell
& Gomez-Castellanos, 2000). Tunnels may be IP-IP (IP over IP) or MPLS (Multi-protocol
Label Switching).
2.1.3 Mobility Signaling

Defines if the mobility signaling is carried out by the network alone or also needs the mobile
node participation in the signaling process.
Mobile Node Centric
In this approach, the MN must execute an instance of the mobility protocol, thus
participating actively in the mobility management process. But the requirement for
modification of MNs software may increase their complexity; considering these nodes have
less computational capacity, it may lead to performance degradation on the MN.
Network Centric
In network-based mobility management approach, the serving network handles the
mobility management on behalf of the MN; thus, the MN is not required to participate in
any mobility-related signaling. Contrary to the latter approach, the MN’s performance is not
degraded by processing signaling and mobility protocol management.
2.1.4 Mobility Layer

Defines the responsibility of each layer in the mobility management process.
Link Layer
This class includes mobility protocols that use link layer information, when the point of
attachment changes, to provide mobility management while the node preserves its network-
layer (L3) address. This can fulfill some of the attributes of a micromobility protocol.
Layer 2½
This class uses MPLS to provide mobility management and signaling. MPLS (Rosen et al,
2001) is a technology that substitutes conventional packet forwarding within a network, or
part of a network, with a fast operation of label lookup and switching. Each MPLS packet
has a label. Label swapping is done by associating labels with routes and using the label
value in the packet forwarding process.
In an MPLS cloud, switches are called Label Switching Routers (LSRs) and a connection or
tunnel between two endpoints is formed by the union of several LSRs along a route. It is
called a Label Switch Path (LSP). When a packet enters into an MPLS cloud, the egress LSR
classifies the packet accordingly to the rules defined in its Forwarding Equivalence Class
(FEC) and each FEC has an association with a particular LSP.
Through this mapping (FEC - LSP), a label is assigned to the package, which only identifies
the LSP to the downstream LSR in the LSP, so that they can continue this procedure until
reach the egress (edge) LSR. In core LSRs, the procedure is simpler, since the packet
reclassification is no longer required, but just forwarding it to the downstream LSR. Note
that the label has only local significance. Before a packet leaves an MPLS domain, its MPLS
label is removed (Ren et al, 2001).
Network Layer
All mobility management and signaling is carried out by L3 protocols, based or not in the
Mobile IP protocol.
Transport Layer
Mobility on transport layer intends to maintain TCP (Transmission Control Protocol)'s end-
to-end reliability and correctness semantics while allowing redirecting the endpoints of an
existing transport session (e.g., a TCP connection or a series of UDP – User Datagram
Protocol- packets) to arbitrary addresses (Maltz & Bhagwat, 1998).
Application Layer
Mobility provided by application layer protocols intends to allow communication end
systems to support mobility, heterogeneity, and multihoming. Terminal mobility also
allows a device to move between IP subnets, while continuing to be reachable for incoming
requests and maintaining sessions across subnet changes (Schulzrinne & Wedlud, 2000).
Session mobility also allows a user to maintain a media session even while changing
terminals. For example, a user may want to continue a session initiated on a MN on the
desktop PC when entering his/her office. IPv4 or IPv6 mobility does not directly support
such session mobility (Nasir & Mah-Rukh, 2006).
3. Mobile IP
The Mobile IP (MIP) (Perkins, 1997; Johnson et al, 2004) uses a stable IP address assigned to
mobile nodes. This home address is used to allow the MN to be reachable by having a stable
entry in the DNS service, and to hide the IP layer mobility from upper layers. A
consequence of keeping a stable address independently of the mobile node's location is that
all correspondent nodes try to reach the MN at that address, without knowing the actual
location of the mobile node. Therefore, if there are packets forwarded to the home address,
and the MN is not at its home network, its home agent is responsible for tunneling packets
to the MN’s new location.
MIPv4 (Mobile IP for IPv4 networks) solves the mobility problem by allowing the MN to
use a second IP address: the CoA. This address changes at each new point of attachment
and it indicates the network prefix, identifying the MN’s point of attachment with respect to
the network topology. The CoA is composed of a valid prefix in a foreign network. Thus,
the MN will have a home address and one or more CoAs when moving between networks.
MIPv4 works by the cooperation of three separable mechanisms (Perkins, 1998): discovering
the CoA, registering the CoA and tunneling to the CoA. The operation of Mobile IP protocol
can be briefly described by the following steps (Figure 2):
1. The mobility agents (HA and FA) announces their presence through messages called
Agent Advertisement (optionally, these messages can be requested by mobile agents
through messages called Agent Solicitation);
2. A MN receives these messages and determines whether it is on its home network or on
a foreign network;
3. When a MN detects it moved to a foreign network, it obtains a CoA in that network.
The CoA can be allocated by the foreign agent or some other address configuration
mechanism, such as DHCP (Dynamic Host Configuration Protocol);
4. When the MN is operating in the new network, it needs to register its CoA with its HA,
through the exchange of Registration Request and Registration Reply messages;
5. Datagrams sent to the MN’s home address by a CN are intercepted by the local HA and
tunneled to the MN’s CoA. The datagram is received at the exit of the tunnel, and
finally delivered to the mobile node in the new network;
6. Datagrams sent by the MN are generally delivered to the destination using standard
routing mechanisms, not necessarily through the HA.
The cooperation between MN, HA and CN is called triangular routing, as we can see in
Figure 2, which summarizes the MIPv4 operation.
The triangular routing generates a processing overhead on HA, in addition to this being a
single point of failure in the network. The MIPv6 solves this problem by optimizing the
route.
Mobile IPv6 (Johnson, 2004) is intended to provide mobility support in IPv6 networks. In
order to know where the MN is found, an association between home address and care-of
address should be performed (binding). This combination of CoA is made by the MN and
the HA. This association is achieved by a binding registration where the MN sends
messages called Binding Updates (BU) to HA, which responds with a message Binding
Acknowledgment (BA) (Figure 3).
The correspondent nodes may carry out route optimization, or they can store bindings
between MN’s home address and CoA. Thus, a MN can supply information about its
location to the corresponding nodes, through the Correspondent Binding Procedure, which
is a mechanism for authorizing the establishment of binding, called the return routability
procedure.
Fig. 2. MIPv4 Operation
Using the Route Optimization process, the CN must support MIPv6 and the MN must
register with the CN. In this case, the CN, before sending a package, looks for a cached
association between MN’s HA and CoA. If there is an association, the package will be
routed to the CoA of mobile node directly. This eliminates congestion at the home link and
the HA.
Fig. 3. MIPv6 Operation
MIPv4 and MIPv6 only define means of managing macromobility but do not address
micromobility separately. Indeed, it uses the same mechanism in both cases. So, this
protocol is not suited for micro mobility management, because of its high signaling load and
long handover delay (Habaebi, 2006), namely movement detection, new CoA configuration,
and Binding Update, is often unacceptable to real-time traffic such as Voice over IP (Koodli,
2008).
4. Link Layer related Micro-Mobility

FMIP - The Mobile IPv6 Fast Handovers (FMIPv6) protocol (Koodli, 2008) aims to reduce
MN movement detection latency and new MN’s CoA (Care-of Address) configuration
latency by providing information to the MN when it is still connected to its current subnet.
After discovering available access points, the MN requests subnet information from these
APs: prefix, IP address, and L2 address of their associated routers. If the MN eventually
attaches to one of the APs, the movement detection delay is reduced because the MN
doesn’t need to perform router discovery.
The MN formulates a new CoA (NCoA) based on the prefix of the new subnet and sends a
message to its current access router, Previous Access Router (PAR), which communicates
with the New Access Router (NAR) to determine whether the NCoA is unique. The PAR
also establishes a tunnel to redirect packets arriving for PCoA (Previous CoA) to NCoA.
After performing a handover the MN announces its attachment immediately with an
Unsolicited Neighbor Advertisement message (Narten et al, 2007) to circumvent the delay
associated to neighbor's address resolution. FMIPv6 also defines a different behavior when
the MN doesn’t receive acknowledge message prior to its handover. There is also an
adaptation of FMIPv6 to IPv4 networks (Koodli & Perkins, 2007).
IP-IAPP – The IP-IAPP proposal (Samprakou et al, 2004) extends 802.11f IAPP (IEEE, 2003)
to support inter-network handover via L2 specific methods. IP-IAPP defines the Home
Access Point (HAP) that is the AP to which the MN was last associated inside its home
network, similar to the Home Agent in MIP.
When the MN moves to a different network, it sends a modified IEEE 802.11
Reassociation.Request (IEEE, 1997) to an AP, the Foreign Access Point (FAP), informing IP
addresses of HAP, MN, and Previous FAP (PAP) and this message triggers the mobility
management procedure. The FAP communicates with the HAP to establish a bi-directional
HAP-FAP tunnel and the HAP starts mapping the MN IP address to the FAP IP address, the
Foreign Agent Care of Address (FACOA).
When the MN reassociates with a new FAP (NAP), the same procedure is performed with
the addition of a communication between the PAP e NAP to establish a temporary
unidirectional tunnel between them. The proposal has also been improved with the
provision of more advanced services: secure inter-AP IP-IAPP communications, zero
patching on the clients software, and support of clients which use a dynamic IP address
(Samprakou et al, 2007).
The IEEE 802 Executive Committee approved IAPP withdrawal in 2006, because “the trial
use period of 802.11F has expired, there has been no significant deployment of 802.11F
implementations and, the functionality provided by 802.11F is being addressed in other
standards fora“(IEEE P802.11, 2005).
5. Mobility with MPLS

All of the architectures discussed in this section consider that the MPLS cloud is surrounded
by the Internet cloud and that micromobility is to be applied in MPLS cloud while MIP is to
be applied in Internet cloud.
Mobile MPLS – The Mobile MPLS is a macro mobility protocol that borrows the
mechanisms defined in the MIP standard and applies it to MPLS networks, so that the IP-in-
IP tunnels are substituted by MPLS tunnels. The main objective in this migration is to
improve the delay time in the tunneling packets from the HA to the FA. Another objective is
to facilitate the use of QoS services that are native to MPLS networks (Ren et al, 2001).
In order to track the MN location, an entry in the LIB (Label Information Base) table at the
HA is created for each MN that is registered with it. When the MN arrives at a foreign
network, it registers itself with this FA and obtains a CoA from it. The FA sends this
information to the MN's HA and it establishes a new LSP for that FA.
After the completion of the LSP connection, the HA changes the LIB entry for that MN to
reflect the out label and port interface gathered from the previously created LSP. In doing
that, whatever packet that is sent to the MN’s home network will be tunneled to this LSP,
arriving at the FA in which the MN is actually connected. A lack of an entry about out label
and port interface for the MN at the LIB table at the MN's HA means that the MN returned
to its home network.
Three scenarios were discussed. The first one considered was that both FAs and HAs were
inside the same administrative MPLS domain. The second one considered was that HAs
and FAs were inside different administrative MPLS domains. In order to establish a tunnel
between them, Mobile MPLS suggests the use of a border protocol such as BGP (Border
Gateway Protocol) (Rekhter & Rosen, 2001). The third scenario considered was that HA and
FA were inside a different network tunneling technology, such as MPLS and IP clouds. LER
is responsible for de-tunneling packets from the MPLS cloud and re-tunneling it inside the
IP cloud.
H-MPLS - Hierarchical Mobile MPLS (Yang & Makrakis, 2001) extends Mobile MPLS,
which is a macro mobility protocol, in order to introduce into it micro mobility features. The
main objective is to reduce the signaling overhead in creating a LSP from HA to FA, due to
MN frequent handover in a small-size cells wireless environment.
To do that, H-MPLS introduces a new element, called FDA (Foreign Domain Agent) whose
function is between that defined for HA and FA, as described in the MIP standards. The role
of FDA is to track MN local mobility inside a MPLS domain. There is only one FDA per
MPLS domain and many FA per subnetworks inside this domain.
The dynamics of the protocol is as follows: whenever a MN enters a foreign MPLS domain
and it is its first registration, it acquires a CoA from its FA LSR and registers with it. This FA
sends a Registration Request to its FDA, which has an equivalent function of a FA, but its
scope is for a domain. This FDA will send back a Label Request message to FA and put as
its FEC, the MN's CoA. At meanwhile, FDA will send a Registration Request message to
HA, in the same way that was described in Mobile MPLS, but putting its IP address as a
FEC for this LSP. So, at this point, there will be two LSPs, one from HA to FDA and another
one from FDA to FA.
Now, if a MN does a handover, but stays in the same domain as the previous FDA, only the
LSP from the new FA to FDA needs to be established. To avoid FDA sending packets to MN
via the old FA, due to an out-of-date entry cache, the new FA sends a Binding Update
message to old FA instructing it to create a LSP to the new FA in order to tunnel packets
arriving at the MN’s old location.
LEMA - Label Edge Mobility Agent (Chiussi et al, 2002) is a tunnel-based micro mobility
architecture that uses MPLS as a network transport technology. This network is composed
of an overlay network whose nodes are called LEMA, an LER that has its function
augmented with LEMA features. This overlay network tracks MN location by building a set
of LEMAs nodes from the highest to the lowest LEMAs that compound a path.
Highest LEMAs are the ingress node which registers its address in the HA database, and
acts as FA in the MIP protocol; while the lowest LEMA is the access router, which remove
the MPLS tunnel and delivers messages to the AP which the MN is connected to. This kind
of scheme makes a hierarchical network, where only the LEMAs that compose a path to
track the MN need to be aware of it.
Others features that can be attributed to it are fast handover capability, scalable design, QoS
capability and gradual deployment. It is the MN’s role to define the set of LEMAs that
compose its path inside a LEMA network. This path is chosen based on a set of parameters,
such as: available bandwidth, mobility patterns, and so on. The algorithm employed to
choose a particular set is an open issue, and could be of high complexity. Finally, all LEMAs
are connected to themselves by pre-established LSPs.
MM-MPLS - Micro Mobile MPLS (Langar et al, 2004) extends mobile MPLS with the
principles employed by MIP-RR (L3 protocol) to support micro mobility on MPLS
networks. It introduces a new component, called Label Edge Router/Gateway (LER/GW)
that resides between HA and FA agents, as defined in the MIP protocol. It acts as a foreign
domain agent to HA and it is this address that the MN must register at HA database when
it first gets into the domain at which LER/GW is administrating.
So, there is a tunnel/LSP from HA to LER/GW, and an LSP from LER/GW to FA, with MN
CoA's address as a FEC. FA here is an AR (Access Router) that remove the MPLS tunnel and
delivers the packet to MN that is registered on it. Whenever MN moves to another FA,
which is under the same LER/GW, only a regional registration is required that will create a
new LSP from LER/GW to the new FA, using the new MN CoA's address as FEC. MM-
MPLS uses LDP (Label Distribution Protocol) as signaling protocol in MPLS cloud.
I-LIB - Intermediate Label Information Base (Fowler & Zeadally, 2006) maintains the same
idea of MM-MPLS architecture in general, where a FDA is placed between HA and FA.
FDA has the same role as described early, i.e., its CoA address is registered at HA database
and a tunnel connect HA to FDA. On the other side, FDA keeps track of MN by establishing
a LSP to FA that is directly connected to it. Whenever a MN does a handover and
establishes a new connection to a new FA, this new FA will try to establish a LSP to FDA,
sending a Registration Request.
Here is where this proposal differs from the previous ones. Instead of establishing a new
LSP from scratch, linking the new FA to the FDA, any segments that are common between
the new path and the old path will be preserved. As such, any LSR that already has an entry
in its LIB could preserve it and just update it to show the new configuration (the new
segment that connects the MN). In order to do that, a new LIB is proposed which
augmented the old ones with new fields to contemplate mobility issues. Among the fields
that are required, the previous and new MN's CoAs must be accounted for. It is necessary to
modify the packet since the FDA and HA know the MN by its old CoA, while the FA knows
the MN by its new CoA.
6. Network Layer
Cellular IP – The CIP (Valko, 1999) architecture is composed of different wireless access
networks (CIP access networks) connected to the Internet through a gateway. MIP manages
mobility between these CIP access networks, while Cellular IP handles mobility within one
domain. The IP address of the gateway is used as the MIP CoA. Thus, packets are first
routed to the host‘s HA and then tunneled to the gateway, which “detunnels” packets and
forwards them toward base stations, using host-specific routing path.
Base station (BS) components serve as wireless access points and also route IP packets, but
IP routing is replaced by Cellular IP routing and location management. Base stations cache
the path taken by uplink packets from MN to gateway for a period of time and use the
reverse path to route downlink packets. In order to route packets to idle MNs, Cellular IP
employs paging.
HAWAII - HAWAII divides the network into a hierarchy of domains. All issues related to
mobility management within one domain are handled by a gateway called a domain root
router, which uses a specialized path setup scheme which installs host-based forwarding
entries in specific routers to support intra-domain micromobility (Ramjee et al, 1999). While
moving inside its home domain, the MN maintains its stable IP address.
MIP mechanisms are used when the MN moves into a foreign domain. However, if the
foreign domain is also based on HAWAII, then the MN is assigned a co-located CoA from
its foreign domain to which packets for the MN are tunneled. The domain root router routes
the packets to the MN using the host-based routing entries. When the MN moves between
different subnets of the same domain, only the route from the domain root router to the BS
serving the MN is modified, and the remaining path remains the same, and connectivity is
maintained using dynamically established paths.
The protocol contains three different messages for establishing, updating and refreshing
host specific routes for the MN in the domain root router and any intermediate routers on
the path towards the mobile host. The protocol also has four different path setup schemes,
aiming to reduce disruption to the user traffic during a handoff, classified into two types
based on the way packets are delivered to MNs. In the first type, packets are forwarded
from the old base station to the new and, in the second type, they are diverted at the
crossover router.
MIP-RR - MIP-Regional Registration (Fogelstroem et al, 2007) is an optional extension to
the Mobile IPv4 protocol, and proposes a mean for mobile nodes to register locally within a
visited domain. By registering locally, the number of signaling messages to the home
network is kept to a minimum, and the signaling delay is reduced. This protocol introduces
a new network node called the Gateway Foreign Agent (GFA). Besides the regular MIP
Registration messages, a new pair of registration messages, Regional Registration
Requests/Replies, is used between MNs/FAs/GFAs.
There are two models of how the MN uses Regional Registration. In the first model, the FAs
in a visited domain advertise the address of the GFA, and, when a mobile node first arrives
at this visited domain, it performs a home registration. At this registration, the mobile node
registers the address of the GFA as its CoA with its HA. When moving between different
foreign agents within the same visited domain, the mobile node only needs to make a
regional registration to the GFA. In the second model, the FA can indicate that dynamic
assignment of GFA is to be used, if being the FA’s responsibility to choose the GFA after
receiving a Registration Request from the MN.
PMIP - The Proxy MIP (Gundavelli et al, 2008) is a network-centric micromobility approach
that relies on tunnels inside a domain to direct traffic to mobile nodes. PMIPv6 reuses many
concepts of MIPv6, like the HA functionality, and defines two new entities, the Mobile
Access Gateway (MAG) and the Local Mobility Anchor (LMA).
A MAG typically runs on the Access Router. It is responsible for detecting the mobile node
attachments, and, if security policies are fulfilled, establishes tunnels to the LMA for
directing traffic to the mobile nodes reached via this MAG. A MAG also emulates (via
Router Advertisements messages) the mobile node’s home network in such a way that the
mobile node may change the default router in a handover, but preserves the remaining L3
parameters. As the mobile node moves inside the domain, tunnels between MAG and LMA
and routes on the LMA are updated.
LMA maintains a binding cache entry for each currently registered MN, providing
reachability to the MN’s address. When the LMA receives a packet targeted to the mobile
node it forwards the packet via the tunnel ending on the MAG to where the node is
attached. PMIPv6 employs local binding update messages between MAG and LMA for
signaling purposes and only a single hierarchy of tunnels. Indeed, Proxy MIP considers
IPv4 support, but this requires an extension to the original protocol, since it uses some IPv6
features such as auto-configuration and extension headers.
HMIPv6 - To support local mobility, Hierarchical Mobile IPv6 (HMIPv6) extends Mobile
IPv6 and IPv6 Neighbor Discovery (Narten et al, 2007) and introduces Mobility Anchor
Point (MAP), a new Mobile IPv6 node. A mobile node entering an HMIP domain receives
Router Advertisements containing information about one or more local MAPs and
configures two CoAs: an on-link CoA (LCoA) and a Regional Care-of Address (RCoA).
The LCoA is configured on a mobile node's interface based on the prefix advertised by its
default router. It is a standard Mobile IP CoA and has a different name just to be
distinguished from RCoA. The RCoA is configured on the MAP's link and is obtained by
the MN from the MAP employing the address mechanisms described by RFC 4877
(Devarapalli & Dupon, 2007).
After configuration, the MN sends two binding update (BU) messages. The first is a local
BU to the MAP to bind the MN's RCoA to its LCoA and to establish a bi-directional tunnel
between them. The second is a BU to the home agent to bind the MN’s home address to its
RCoA. The MAP receives all packets on behalf of the mobile node it is serving and
encapsulates and forwards them directly to the mobile node's LCoA.
When the mobile node moves within the same MAP domain, it only needs to register its
new LCoA with its MAP, limiting the amount of Mobile IPv6 signaling outside the local
domain. The RCoA remains unchanged and the home agent (HA) or the correspondent
nodes (CNs) are not aware of the change in LCoA.
DMA - The Dynamic Mobility Agent (Misra et al, 2001) architecture uses the Intra-Domain
Mobility Management Protocol (IDMP) (Das et al, 2002) to manage intradomain mobility.
The architecture defines two entities to achieve mobility support: Mobility Agent (MA) that
acts as a domain-wide point for packet redirection, and the Subnet Agent (SA) that provides
subnet-specific mobility services. Two CoAs are associated with a MN: Global CoA (GCoA)
and Local CoA (LCoA).
The GCoA is the address used by macromobility protocols to redirect packets and remains
unchanged as long as the MN stays in the current domain. The LCoA is an address from the
subnet the MN is attached to. IDMP is used in the communication between the MN and the
SA, and between the MN and the MA.
When the MN first arrives at the domain, it obtains an LCoA and the SA assigns the MN a
MA. The MN registers its LCoA with the MA and obtains a GCoA. After the macromobility
updates process is performed by the MN, the packets from remote hosts, tunneled or
directly transmitted to the GCoA, are intercepted by the MA and tunneled to the MN’s
LCoA. When the MN moves to new subnet it obtains a new LCoA and informs its MA of
the new LCoA, updating the GCoA-LCoA mapping.
7. Transport Layer Mobility

MSOCKS is a split-connection proxy-based architecture that uses TCP Splice technique to
achieve the same end-to-end semantics as normal TCP connections (Maltz & Bhagwat,1998).
A special host, called a proxy, is placed in the communication path between a mobile node
and a correspondent node. An end-to-end TCP connection between a mobile node and a
correspondent node are split into two separated connections: one connection between the
mobile node and proxy and another between the proxy and the correspondent node. The
MSOCKS protocol extends the SOCKS protocol (Leech et al, 1996) to redirect TCP streams to
a mobile node’s changing location.
When the mobile node changes the address of its network interface it opens a new
connection to the proxy and sends an MSOCK message specifying the connection identifier
of the original connection. The proxy unsplices the old mobile-node-to-proxy connection
from the proxy-to-correspondent-node connection, and splices in the new mobile-node-to-

proxy connection. Only the proxy is aware of the mobile node migration and the
communication between the proxy and the correspondent node remains unchanged. This
technique also allows the mobile node to change the network interface used to communicate
with the proxy.
TCP Migrate – This mobility architecture (Snoeren & Balakrishnan, 2000) (Snoeren et al,
2002) allows an application running on mobile hosts to support transparent connectivity
across network address changes. As MNs change their network attachment point, new
addresses can be assigned through DHCP, manually or using an auto-configuration
protocol. To locate mobile hosts in the new network, Domain Name System (DNS) is used
and its ability to support secure dynamic updates.
Because most Internet applications resolve hostnames to an IP address at the beginning of a
transaction or connection, this approach is viable for initiating new sessions with mobile
hosts. When a host changes its network attachment point (IP address), it sends a secure
DNS update to one of the name servers in its home domain updating its current location.
The name-to-address mappings for these hosts are un-cacheable by other domains, so stale
bindings are eliminated.
Nevertheless, when a MN moves during a previously established connection, it may
suspend the open connection and reactivate it from the new address, sending a special
packet (Migrate SYN) to the correspondent node, which carries a token that identifies the
previous connection. This SYN packet signals the correspondent node to re-synchronize the
connection with the MN at the new point of attachment (new address). Thus, it is possible
to provide mobility support as an end-to-end service, according to the application’s specific
requirements, without changes in the network layer.
8. Application Layer Mobility

Mobility using SIP - The SIP (Session Initiation Protocol) is a signaling protocol, widely
used for setting up and tearing down multimedia communication sessions over the Internet.
It can be used in any application where session initiation is necessary.
The SIP registration mechanism is considered the application-layer equivalent of the MIP
registration mechanism. However, while mobile IP binds a permanent IP address
identifying a host to a temporary CoA, SIP binds a user-level identifier to a temporary IP
address or host name (Schulzrinne & Wedlund, 2000). An INVITE message is sent by a MN
to its CN to set up a communication session. The mechanism to provide MN mobility
during an active session foresees that the MN needs to send another INVITE message to the
CN to communicate the information about the new parameters of the communication
session after the handover, using the same call identifier as the original call setup.
This solution has some drawbacks (Salsano et al, 2008). The second INVITE is sent end-to-
end, and this could lead to high delays. Moreover, the handover procedure relies on the
capability of the CN to handle this procedure, thus increasing MN processing needs. An
auxiliary mechanism is necessary if the MN and CN move at the same time.
9. Mobility Plane Architecture

The Mobility Plane Architecture (MPA) (Zagari et al, 2008) is an instance of a reference
architecture for micromobility support in IP networks (Prado et al, 2008). The goal of MPA is
to speed up the handover process in order to minimize communication disruptions when the
mobile node changes its network point of attachment. One of the requirements of this
architecture is to place the burden demanded by micromobility on the network, not on the
mobile nodes. Another requirement is to use, ideally, only well established network protocols.
The key point of MPA is to employ an overlay network built above a transport network for
directing traffic to the mobile nodes. This overlay network is composed of network elements
called Mobility Aware Router (MAR), which are routers enhanced with MPA’s
functionalities. MPA employs point-to-multipoint (P2MP) tunnels in order to encapsulate
traffic directed to the mobile nodes and allows a gradual deployment once the architecture
elements are installed only at the MARs.
MPA addresses the following issues related to mobility in IP networks: tunnel management
(tunnel establishment, shutdown, and topology updating); secure mobile node attachment
and handover; tracking of mobile node actual point of attachment (location); routing on the
overlay network (decoupled from routing on the transport network); and quality and class
of service (QoS/CoS) offered to the mobile nodes.
9.1 Functional Description

The architecture defines the following basic elements:
 Transport network - an IP network from which the network operator wishes to
offer mobility services.
 Point-to-multipoint (P2MP) tunnel - a tunnel with a topology forming a tree.
Nodes in the tree are MARs and arcs are tunnel segments connecting MARs. The
tunnel has a single ingress (root) MAR, branch MARs (nodes with branching level
greater than one), and egress MARs (leaves of the tree). A packet being forwarded
through the tunnel may or may not be replicated at a branch MARs according to
the policies enforced by these MARs.
 Access router - a router (usually an egress MAR) connected to a wireless access point.
 Access network - an IP subnetwork formed by the access routers and access points.
 Overlay mobile network - logical network built with one or more P2MP tunnels
established through the transport network.
Figure 4 illustrates these basic elements. In addition to the basic elements, four functional
blocks (FB) are defined:
 Tunnel Management (TM) Functional Block: TM is the entity responsible for P2MP
tunnel establishment, shutdown, and re-routing. It must provide interfaces to the
network management system and to the human operator. Tunnel management is a
function carried out by MARs.
 Mobile Routing (MR) Functional Block: MR is the entity responsible for tracking
the mobile nodes actual point of attachment and for interacting with the MARs
forwarding engine in order to route traffic to the mobile nodes correct location.
Mobile routing is a function carried out by MARs.
 Address Configuration (AC) Functional Block: AC is the entity responsible for
supplying L3 addresses to the mobile nodes when they connects or reconnects to
the access network. Address configuration is a function carried out cooperatively

by MARs and mobile nodes.
 Handover Helper (HH) Functional Block: HH is the entity responsible for
facilitating the handover process with functions including L2 notification
(triggering), L2 re-association, secure node attachment, and handover-related
signaling. This function can be spread among MARs, network equipments (e.g.,
wireless switches), and mobile nodes.
Fig. 4. MPA overview
9.2 Operations Basics

When a packet targeted to a mobile node reaches an ingress MAR, it is tunneled until it
reaches an egress MAR able to route the packet to the mobile node. When the mobile node
performs a handover, the AC FB presented on the mobile node and on the network interact
in order to provide the mobile node with a new L3 address. If the address is identical to the
previous one the transport connections are not broken due the handover. The handover also
triggers a mobile node location update on the MR FB. The location updating process
updates the mobility routing tables on the MARs in such a way that when a packet is
targeted to the mobile node the packet is routed to the egress MAR serving the link the
mobile node is attached to.
Let us consider the mobile routing and the location update processes. Figure 5 shows a
mobile node attached via access router M4. When the mobile node moves to a link served
by M5, the entry related to this node on the mobile routing table at M2 must be updated
with a different tunnel segment (in this case from segment C to D). If the mobile node roams
to a link served by M7, the mobile routing table at M1 and M3 must be updated. Table
updates are performed as soon as the mobile routing protocol messages indicating the new
point of attachment are processed by the branch MARs.
The entries on the mobile routing table are soft state, meaning that the entries are dropped if
location update messages confirming them cease. Soft state is a clear way to drop routes to
mobile nodes when they no longer are reached through these routes. This scheme demands
that a mobile node perform network attachments periodically in order to generate location
update messages that will refresh the routes to it. A way to force the mobile nodes to perform
periodic attachments is to provide them L3 address with a short lease time. When the lease
time is close to expire, a mobile node performs address renewal that will trigger address
location update messages in its behalf.
Fig. 5. MN Routing in MPA
9.3 MPA advantages

The MPA architecture presents the following advantages:
1. The solution is not limited to IPv6, being deployable on both IPv4 and IPv6
networks. Since it relies on tunneling, mixed deployments with IPv4 on the access
network and IPv6 on the transport network, and vice-versa are possible.
2. The solution is not affected by middle boxes such as firewall and NAT boxes
placed anywhere on the access, transport, or backbone networks.
3. The solution demands no complex protocols such as MIPv6 on the mobile nodes.
Since it relies only on the standard IP protocol stack, the solution supports all the
commercially available mobile nodes based on, for instance, Windows Mobile,
Symbian, and PalmOne operating systems. The architecture does not forbid
enhancements deployed on the mobile nodes in order to improve handover speed

and security, for instance. Such enhancements can be installed in user space or in
the operating system kernel (e.g., as device drivers).
4. The solution preserves the L3 address of the mobile nodes when they roam inside
the access network, causing no disruption of transport connections maintained by
the mobile nodes.
5. The solution does not restrict the mobile node to employ macromobility protocols
such as MIPv6.
6. The solution complies with security mechanisms related to L2 (e.g. WPA), L3 (e.g.
IPSec), and L4+ (e.g., SSL, HTTPS).
7. The solution combines P2MP tunneling management, QoS/CoS management, and
mobile node location tracking into the same protocol (RSVP-TE), reducing
implementation and operating costs.
8. The solution does not interfere on services already deployed on the transport
network such as VPN and VoIP.
9. Only well standardized protocols are employed by the architecture. When
extensions to protocols are necessary they are introduced as opaque objects already
foreseen by these protocols.
For more details on MPA description, implementation (using IPv4 and IPv6) and
performance analysis, see Prado (2008), Zagari (2008), Johnson (2008), and Zagari (2009).
Badan et al (2009) details the MPA implementation using MPLS.
10. General Analysis and Research Directions

Table 1 summarizes all the protocols and solutions seen in this chapter.
Protocol Mobility Mobility Mobility Signaling Mobility

Range Routing Layer
CIP Micro Routing MN-centric L3
DMA Micro Tunnel MN-centric L3
FMIP Micro Tunnel MN-centric L2
Hawaii Micro Routing/Tunnel MN-centric L3
HMIP Micro Tunnel MN-centric L3
H-MPLS Micro Tunnel MN-centric L2½
I-LIB Micro Tunnel MN-centric L2½
IP-IAPP Macro/micro Tunnel MN-centric L2
LEMA Micro Tunnel MN-centric L2½
MPA Micro Tunnel Network-centric L3
MIP Macro Routing/Tunnel MN-centric L3
MIP-RR Micro Tunnel MN-centric L3
MM-MPLS Micro Tunnel MN-centric L2½
Mobile MPLS Macro Tunnel MN-centric L2½
MSOCKS Micro Routing MN-centric L4
PMIP Micro Tunnel Network-centric L3
SIP Macro Routing MN-centric L5
TCP Migrate Macro Routing MN-centric L4
Table 1. General Classification
As this work is a non-exhaustive selection of mobility solutions, our comments are limited
to these protocols revised in this chapter.
There are a majority of MN-centric solutions, since the beginning of mobility management
research, with the Mobile IP. Newer solutions, such as MPA and PMIP, work as network-
centric solutions, and it is our belief that this class of mobility signaling will have more
research focus and implementations, because of its advantages to the final user: they do not
need protocol installation of configuration on the MN, no overheads on devices to handle
mobility, and no intervention from the final user to adopt a particular mobility solution.
Another interesting issue is the network layer protocol. About 10 years ago nobody believed
in IPv4 for mobility, which is why both MIPv6 and its extensions appeared at that time.
Since then, however, the IPv6 protocol has not been massively adopted, so there is
reawakened interest in IPv4 (PMIP and MIPv4 raised again, for example). Now, 4G
researchers and professionals say IPv6 is inevitable because each cell phone and mobile
devices must have a fixed (stable) IP address. Solutions compatible only to IPv4 must
foresee this IPv6 adoption and implement a compatible version of their protocol, if intended
to be used in the next future.
Not mentioned in this chapter, solutions for mobility must adopt security mechanisms for
authentication, authorization and general security procedures. For example, for the MPA
architecture, the access points can be configured to authenticate mobile nodes based on
WPA2 employing Pre-Shared Keys (PSK) or RADIUS. PSK is easy to configure but is not as
secure as RADIUS-based authentication. RADIUS authentication can be strengthened by
using certificates installed on the mobile nodes. As RADIUS transactions take long time
(500ms in our testbed network), RADIUS-based authentication increases considerably the
handover overhead.
In order to speed up RADIUS-based authentication, a cache mechanism can be employed
such as PMK (Pairwise Master Key) caching (also called proactive key caching). In this
mechanism, once a mobile node completes successfully a RADIUS transaction, the access
point stores the PMK supplied by the RADIUS server in the cache. When the mobile node
connects to a new access point, the access point queries the cache (using the mobile node’s
MAC address as a search key) in order to recover the PMK assigned to the node. If an entry
is found, the access point accepts the mobile node without the need of a RADIUS
transaction. In this case, the PMK found on cache is used to secure the communication
between the mobile node and the access point.
As suggestions for future research directions in mobility management, we
can point out the development of architectures able to:
 integrate macro and micro mobility into a single mobility solution;
 support both vertical and horizontal handover (e.g., between WiFi and 4G
networks);
 support clean slate solution, by designing the network from the mobility
requirements (and not to incorporate mobility extensions over the existing
networks).
 These solutions decouple host identity from network address as suggested by HIP
(Host Identity Protocol) and other related solutions;
 restrict handover within the L2, employing, for instance, tunneling over Ethernet
(instead of over IP or MPLS), flat routing (based on MAC – Media Access Control -
address), etc.
11. Conclusion
This chapter was intended to present a major review on mobility architectures and
protocols. Many solutions were shown, ranging from link layer related solutions to
application layer solutions, including network, transport and intermediary layer protocols.
We also presented MPA, which is our solution for mobility management, using a network
centric paradigm.
Some of the reviewed solutions were abandoned, some were investigated till today. But
mobility management solution still needs investigation, to allow massive deployment and use.
Although the new mobility architectures and protocols are permanently under
investigation, all new solutions are constrained by factors such as: be deployable over
existing IPv4 (and future IPv6) fixed networks, operate without upgrading with the mobile
nodes already in the marked, comply with current network operation practices, be scalable
without degrade quality of service, and allow the introduction of new services with
stringent communication requirements such as media streaming (e.g., IP TV), location, and
entertainment services. The challenge in mobility for IP networks is to comply with these
factors that, unfortunately, are not restricted only to the technical issued commonly
addressed by the network architects.
12. References
Akyildiz, I., Xie, J., & Mohanty, S. (2004). A Survey of Mobility Management in Next-
Generation All-IP-Based Wireless Systems. IEEE Wireless Communications, (4),1
16–28.
Badan, T., Zagari, E., Prado, R., Cardozo, E., Magalhaes, M., Carrilho, J., Pinto, R.,
Berenguel, A., Moraes, D., Johnson, T., & Westberg, L. (2009). A Network
Architecture for Providing Micro-Mobility in MPLS/GMPLS Networks. Proceedings
ofIEEEWirelessCommunicationsandNetworkingConference pp. 1-6). (
Campbell, A. T., & Gomez-Castellanos, J. (2000). IP Micro-Mobility Protocols. ACM
SIGMOBILEMobileComputingandCommunicationsReview , 4 (4), 45 – 53.
Campbell, A. T., Gomez, J., Kim, S., Wan, C.-Y., Turany, Z. R., & Valko, A. G. (2002).
Comparison of IP Micromobility Protocols. IEEE Wireless Communications , 9 (1), 2–
12.
Chiussi, F. M., Khotimsky, D. A., & Krishnan, S. (2002). A Network Architecture for MPLS-
Based Micro-Mobility, Proceedings the of IEEE Wireless Communications and
NetworkingConference , vol. 2, pp. 549-555.
Chiussi, F., Khotimsky, D., & Krishnan, S. (2002). Mobility Management in Third-
Generation All-IP Networks. IEEECommunicationsMagazine , 40 (9), 124–135.
Das, S., Mcauley, A., Dutta, A., Misra, A., Chakraborty, K., & Das, S.K. (2002). IDMP: an
intradomain mobility management protocol for next-generation wireless
networks. IEEEWirelessCommunications, (3), 9 1536-1284.
Devarapalli, V., & Dupon, F. (2007). Mobile IPv6 Operation with IKEv2 and the Revised
IPsec Architecture. Request For Comment RFC) ( 784 , Available at:
www.faqs.org/rfcs/rfc4877.html.
Fogelstroem, E., Jonsson, A., & Perkins, C. (2007). Mobile IPv4 Regional Registration.
RequestForComments(RFC)4857 , Available at: www.ietf.org/rfc/rfc4857.txt.
Fowler, S., & Zeadally, S. (2006). Fast Handover over Micro-MPLS-based Wireless
Networks, Proceedings of the 1th Symposiu m on Computers and Communications (pp.
181-186).
Gundavelli, S., Leung, K., Devarapalli, V., Chowdhury, K., & Patil, B. (2008). Proxy Mobile
IPv6. Request For Comment (RFC) 5213, Available at:
http://tools.ietf.org/html/rfc5213.
Habaebi, M. H. (2006). Macro/Micro-Mobility Fast Handover in Hierarchical Mobile IPv6.
ComputerCommunications , ( 92 51), 611– 617.
IEEE Document P802.11/D6.1.97/5 Wireless LAN, MAC and Physical Specifications, June
1997.
IEEE Recommended Practice for Multi-Vendor Access Point Interoperability via an Inter-
Access Point Protocol Across Distribution Systems Supporting IEEE 802.11TM
Operation. Mar, 2003.
IEEE P802.11 Working Group. (2005). Approved Minutes of the IEEE P802.11 Full Working
Group. Available at: https://mentor.ieee.org/802.11/dcn/05/11-05-1136-00-0000-
minutes-working-group-nov-2005.doc.
Johnson, D., Perkins, C., & Arkko, J. (2004). Mobility Support in IPv6. Request For Comments
RFC)
( 3775, Available at: www.ietf.org/rfc/rfc3775.txt.
Johnson, T., Zagari, E., Prado, R., Badan, T., Cardozo, E., & Westberg, L. (2008). Performance
Analysis of a New Architecture for Mobility Support in IP Networks, Proceedingsof
the International Wireless Communications and Mobile Computing pp. 706- Confere
711).
Kempf, E. (2007). Problem Statement for Network-Based Localized Mobility Management.
Request For Comments RFC)
( 4830, Available at
http://www.ietf.org/rfc/rfc4830.txt.
Koodli, R., & Perkins, C. (2007). Mobile IPv4 Fast Handovers. Request For Comment RFC) (
498, Available at: www.faqs.org/rfcs/rfc4988.html.
Koodli, R. (2008). Mobile IPv6 Fast Handovers. Request For Comment (RFC) 5268 , Available
at: www.faqs.org/rfcs/rfc5268.html.
Langar, R., Le Grand, G., & Tohme, S. (2004). Micro Mobile MPLS Protocol in Next
Generation Wireless Access Networks, Proceedings of the IEEE Symposium on
ComputerandCommunication pp. 14-17).
(
Leech, M., Ganis, M., Lee, Y., Kuris, R., Koblas, D., & Jones, L. (1996). SOCKS Protocol
Version 5. Request For Comment (RFC) 1928 , Available at:
ww.faqs.org/rfcs/rfc1928.html.
Maltz, D.A., & Bhagwat, P. (1998). Msocks: an Architecture for Transport Layer Mobility,
Proceedingsthe of Seventeenth Annual erence Joint Conf of the IEEE Computer and
CommunicationsSocieties, vol.3, pp. 1037-1045.
Manner, J., & Kojo, M. (2004). Mobility Related Terminology. Request For Comment (RFC)
375, Available at: www.faqs.org/rfcs/rfc3753.html.
Misra, A., Das, S., McAuley, A. & Das, S.K. (2001). Autoconfiguration, Registration and
Mobility Management for Pervasive Computing. IEEE Personal Communications
Magazine,8 (4), 24-31.
Narten, T., Nordmark, E., Simpson, W., & Soliman, H. (2007). Neighbor Discovery for IP
version 6 (IPv6). Request For Comment RFC) ( ,1684 Available at:
www.faqs.org/rfcs/rfc4861.html.
Nasir, A., & Mah-Rukh. (2006). Internet Mobility using SIP and MIP. In Proceedings of the
Third International Conference on Information Technology: (ppNew 334 Generation
-
339).
Perkins, C. (1997). Mobile IP. IEEECommunicationsMagazine , 35 (5), 84-99.
Perkins, C. (1998). Mobile Networking through Mobile IP. IEEEInternetComputing, (1),2 58-
69.
Perkins, C. (2002). IP Mobility Support for IPv4. Request for Comments RFC) ( 43 , Available
at: www.ietf.org/rfc/rfc3344.txt.
Prado, R., Zagari, E., Cardozo, E., Magalhaes, M., Badan, T., Carrilho, J., Pinto, R.,
Berenguel, A., Barboza, D., Moraes, D., Johnson, T., & Westberg, L. (2008). A
Reference Architecture for Micro-Mobility Support in IP Networks, Proceedings of
theThirteenthIEEESymposiumonComputersandCommunications (pp. 624 – 630).
Rekhter, Y., & Rosen, E. (2001). Carrying Label Information in BGP-4. Request for Comment
(RFC3107) , Available at: http://www.ietf.org/rfc/rfc3107.txt.
Ramjee, R., Porta, T. L., Thuel, S., Varadhan, K., & Wang S. (1999). Hawaii: A Domain-
Based Approach for Supporting Mobility in Wide-Area Wireless Network,
Proceedings the IEEEInternationalConferenceonNetworkProtocols (pp. 283-292).
of
Ren, Z., Tham, C.-K., Foo, C.-C., & Ko, C.-C. (2001). Integration of Mobile IP and Multi-
Protocol Label Switching. Proceedings of the IEEE International Conference o
Communications (pp. 2123–2127).
Rosen, E.,Viswanathan, A., & Callon, R. (2001). Multiprotocol Label Switching Architecture.
Request For Comments (RFC) 301 , Available at:
http://www.ietf.org/rfc/rfc3031.txt
Saha, D., Mukherjee, A., Misra, I., Chakraborty, M., & Subhash, N. (2004). Mobility Support
in IP: a Survey of Related Protocols. IEEENetwork, 18 (6), 34–40.
Salsano, S., Polidoro, A., Mingardi, C., Niccolini, S., & Veltri, L. (2008). Sip-based Mobility
Management in Next Generation Networks. IEEE WirelessCommunications , 15(2),
92–99.
Samprakou, I., Bouras, C., & Karoubalis, T. (2004). Fast and Efficient IP Handover in IEEE
802.11 Wireless LANs. Proceedings theof InternationalConferenceonWirelessNet
(pp. 249-255).
Samprakou, I., Bouras, C., & Karoubalis, T. (2007). Improvements on IP IAPP: A Fast IP
Handover Protocol for IEEE 802.11 Wireless and Mobile Clients. Wireless Network ,
13 (4), 497–510.
Schulzrinne, H., & Wedlund, E. (2000). Application-layer Mobility Using SIP. SIGMOBILE
MobileComputingandCommunications , 4 (3),Review
47–57.
Snoeren, A. C., & Balakrishnan, H. (2000). An End-to-end Approach to Host Mobility,
Proceedings the
of6th ACM/IEEE International Conference on Mobile Computing and
Networking (pp. 155-166).
Snoeren, A. C., Balakrishnan, H., & Kaashoek, M. F. (2002). The Migrate Approach to
Internet Mobility,ProceedingsofStudent the OxygenWorkshop (pp. 14-17).
Valko, A. G. (1999). Cellular IP: A New Approach to Internet Host Mobility. SIGCOMM
ComputerCommunication, Review 29 (1), 50-65.
Yang, T., & Makrakis, D. (2001). Hierarchical Mobile MPLS: Supporting Delay Sensitive
Applications Over Wireless Internet. Proceedings the
of International Conferences o
Info-techandInfo-(vol. net2, pp. 453-458).
Zagari, E., Prado, R., Cardozo, E., Magalhaes, M., Badan, T., Carrilho, J., Pinto, R.,
Berenguel, A., Barboza, D., Moraes, D., Johnson, T., & Westberg, L. (2008). MPA: a
Network-Centric Proposal for Micro-Mobility Support in IP Networks, Proceedings
of Theth
6 Annual Communication Networks and Services Research Conference (
616).
Zagari, E., Prado, R., Badan, T., Cardozo, E., Magalhaes, M., Carrilho, J., Berenguel, A.,
Moraes, D., Dolphine, T., Johnson, T., & Westberg, L. (2009). Design and
Implementation of a Network-Centric Micro-Mobility Architecture. IEEE Wireless
CommunicationsandNetworkingConference (pp. 1-6).
Positioning in Indoor Mobile Systems

Miloš Borenović and Aleksandar Nešković
SchoolofElectricalEngineering,UniversityofBelg
Serbia
1. Introduction
At present times people travel far greater distances on daily bases than our not so distanced
ancestors had travelled in their lifetimes. Technological revolution had brought human race
in an excited state and steered it towards globalization. Nevertheless, the process of
globalization is not all about new and faster means of transportation or about people
covering superior distances. Immense amount of information, ubiquitous and easily
accessible, formulate the essence of this process. Consequently, ways through which the
information flows are getting too saturated for free usage so, for example, frequency
spectrum had become a vital natural resource with a price tagged on its lease. However, the
price of not having the information is usually much higher. By employing various wireless
technologies we are trying to make the most efficient use of frequency spectrum. These new
technologies have brought along the inherent habit of users to be able to exchange
information regardless of their whereabouts. Higher uncertainty of the user’s position has
produced increase in the amount of information contained in its position. As a result,
services built on the location awareness capabilities of the mobile devices and/or networks,
usually referred to as Location Based Services (LBS, also referred to as LoCation Services –
LCS), have been created. Example of services using the mobile location can be: location of
emergency calls, mobile yellow pages, tracking and monitoring, location sensitive billing,
commercials, etc. With the development of these services, more efforts are being pushed
into producing the maximum of location-dependent information from a wireless
technology. Simply, greater the amount of information available – more accurate the location
estimate is.
Whereas in outdoor environment the satellite-based positioning techniques, such as the
Global Positioning System (GPS), have considerable advantages in terms of accuracy, the
problem of position determination in an indoor environment is much farther from having a
unique solution. Cellular-based, Computer vision, IrDA (Infrared Data Association),
ultrasound, satellite-based (Indoor GPS) and RF (Radio Frequency) systems can be used to

Sometimes, in literature, the words position and location have different meaning. Most often, position
translates to the set of numerical values (such as geographical coordinates) which describe the user’s
placement, whereas the location usually refers to the descriptive information depicting the user’s
whereabouts (such as Picadilly Circus, London, UK). Nevertheless, this work treats both words
interchangeably.
obtain the user’s position indoors. Positioning technologies, specific for indoor
environment, such as computer vision, IrDA and ultrasound require deployment of
additional infrastructural elements. On the other hand, the performances of the satellite-
and cellular-based positioning technologies are often unsatisfactory for typical LBSs in an
indoor environment. Due to the proliferation of IEEE 802.11 clients and infrastructure
networks, and the fact that a broad scope of LBSs can be brought into an existing WLAN
network without the need for additional infrastructure, WLAN positioning techniques are
relevant and established subjects to intensive research.
2. Performance Parameters and Approaches to Positioning

The determination of user's location can be seen as a simple mechanism consisting in
calculating the whereabouts of the user. Those whereabouts could be descriptively
expressed or in terms of geographic or some other coordinates. Nonetheless, it is practically
impossible to obtain the exact location of a user, in 100% of the cases, regardless of the user
and its environment (Collomb, 2002). Therefore, it is only an estimate of the user’s location
that can be obtained and, it is very important to know how proximate the actual location
and location estimate are. To achieve that, it is necessary to characterise this location
estimate. On the other hand, it is also significant to describe the positioning technique itself
in terms of its practicality and viability. All this is generally done through a set of
performance parameters: Accuracy (Distance Error, Uncertainty, Confidence, and Distance
error’s Cumulative Distribution Function), Coverage and Availability, Latency, Direction
and Velocity, Scalability, Complexity and Cost effectiveness.
The first group of performance parameters is used to characterise the quality of a location
estimate.
Accuracy – This is undoubtedly the most important performance parameter as it illustrates
the essential characteristic of a positioning technique. This parameter enables to determine
whether the calculated position is close to the exact position. This parameter is composite
and consists of three different values that must be taken into account:
 Distance Error,
 Uncertainty, and
 Confidence.
The Distance Error corresponds to the difference between the exact location of the user (i.e.
of his/her terminal) and the calculated position, obtained through a position determination
method. It is also referred to as Location Error or Quadratic Error in terms of two-
dimensional positioning. Distance Error is generally expressed in units of length, such as
meters.
Determining the Distance Error can be very useful in depicting the particular position
determination cases. However, in order to express the positioning capabilities of a technique
it is usually much more suitable to exploit the Distance Error statistics via Uncertainty and
Confidence parameters.
Bearing in mind that the calculated user's location is not the exact location but is biased by
the Distance Error, it can be seen that the calculated position does not enable resolving the
single point at which the user is located, rather an area. Depending on the positioning
techniques used, this area may have different shapes (e.g. a circle, an ellipse, an annuli, etc).
For that reason, the Uncertainty value represents the distance from the "centre" of this area
to the edge of the furthest boundary of this area. In other words, the Uncertainty value can
be seen as the maximum potential Distance Error. The value of uncertainty is expressed
with the same unit as for the Distance Error.
However, the Uncertainty value is not sufficient to describe the Accuracy of a positioning
technique. The determination of the Uncertainty value goes through a statistical process and
does not enable to guarantee that 100% of the calculated positions have a Distance Error
lower than the Uncertainty value. That is the reason why the Uncertainty value is usually
associated with a Confidence value, which expresses the degree of confidence that one can
have into the position estimate. This degree of confidence is generally expressed in
percentage or as a value of probability.
As a consequence, it is the combination of Uncertainty and Confidence that validly describe
the accuracy of a positioning technique.
The other way of expressing the Accuracy, i.e. the performance or requirements associated
to location determination, is through the Distance error’s Cumulative Distribution Function
(CDF). This approach is more comprehensive and inclusive due to the fact that a particular
Uncertainty, Confidence pair can always be read of the graph for each and every
Confidence or Uncertainty value. When assessing the technique's suitability for LBSs,
expressing the Accuracy of a positioning technique through an Uncertainty, Confidence
pair might be descriptive enough for a certain LBS. On the other hand, stating a positioning
technique's CDF is more general and depicts the technique's accuracy for all potential LBSs.
Coverage and Availability – Accuracy is not the only parameter to be considered in order
to characterise a location estimate. Coverage and Availability must be considered too. These
two parameters are linked together:
 The Coverage area for a positioning method corresponds to the area in which the
location service is potentially available, and
 The Availability expresses the percentage of time during which the location service
is available in the coverage area and provides the required level of performance.
Latency – Location information makes sense only if it is obtained within a timeframe which
remains acceptable for the provision of the LBSs. Latency represents the period of time
between the position request and the provision of the location estimate and it is generally
expressed in seconds.
Direction and Velocity – Although the herein presented work is restrained to the initial
position determination algorithms, there are additional tracking algorithms that rely on
multiple sequential position determinations in order to estimate the speed vector of the
user. In such cases, two additional parameters have to be calculated: the Direction followed
by the user and his/hers Velocity. These parameters are generally expressed in degrees and
meters per second, respectively.
Scalability – The scalability is a desired and welcomed characteristic of a positioning
system. It represents the positioning system’s ability to readily respond to any
augmentation. The augmentation can be in terms of Coverage area, Availability, frequency
and total number of positioning requests, etc.
Complexity – There are many definitions for complexity depending on the domain of
application. Nevertheless, in terms of positioning systems, complexity is most often referred
to as the property that describes the difficulty of setting up the positioning system.
Cost effectiveness – This abstract characteristic of a positioning system is not entirely
independent of its other performance parameters (e.g. Complexity and Scalability).
For example, the greater the Complexity of the system, the lower the Cost effectiveness. One
of the ways of describing it is as a ratio between the benefits it provides (how broad range of
LBSs it enables) and the costs it induces for the user.
As can be seen from the aforementioned, the latter three parameters don’t have
standardized units and are usually of descriptive nature.
The approaches and metrics used in order to obtain the user’s position are also worth
discussing. There are a few fundamental methods of acquiring the user’s location:
1) Based on the identification of “base station” to which the user is associated (Cell-ID or
Cell of Origin – COO) – This simple approach assumes that the estimated location of a user
is equal to the location of a “base station” to which the user is associated. In other words,
the user is estimated to be in a location of the “nearest” node of the network. This method is
used both in indoor and outdoor environments (GSM, UMTS). Its popularity, despite
inferior performances, is due to the simplicity of implementation. Obviously, the accuracy is
proportional to the density of the network nods.
2) Based on the time of signal arrival (Time of Arrival – TOA) – Being that the waves
(electromagnetic, light and sound) are propagating through the free space at constant speed,
it is possible to asses the distance between the transmitter and a receiver based on the time
that the wave propagates in-between those two points. This approach assumes that the
receiver is informed of the exact time of signal’s departure. Being that this is not always
easily accomplished, the alternative approach takes into account the time needed for signal
to propagate in both directions (Round Trip Time – RTT). This way, one station is
transmitting the predefined sequence. The other station, upon receiving the sequence, after
a strictly defined time interval (used for allowing the stations of different processing power
to process the received information), resends the sequence. The station that initially sent the
sequence can now, by subtracting the known interval of time that the signal was delayed at
second station from the measured time interval, asses the time that signal propagated to the
other station and back and, consequently, the distance between the stations. This approach
is less dificoult to implement than TOA, since it does not require the stations to be
synchronised.
3) The distance between the stations can be measured based on the differences in times of
signal arrival (Time Difference of Arrival – TDOA) – With this approach, the problem of
precisely synchronised time in transmitter and receiver is resolved by using several
receivers that are synchronised whereas the transceiver, whose location is being
determined, does not have to be synchronised with the receivers. Upon receipt of the
transmitted signal, a network node computes the differences in times of the signal’s arrival
at different receivers. Based on that calculation, the user’s location is determined as a cross
section of two or more hyperboles. Owing to that, these techniques are often referred to as
hyperbolic techniques.
4) Based on the signal’s angle of arrival (Angle of Arrival – AOA or Direction of Arrival –
DOA) – The idea, with this approach, is to have directional antennas which can detect the
angle of arrival of the signal with the maximal strength or coherent phase. This procedure
grants the spatial angle to a point where the signal originated (and whose location is
determined). Vice versa, the mobile terminal can determine the angle of arrival of the signal
from the known reference transmitters. Being that this approach is often implemented
through the use of antenna arrays, the latter approach can have significant impact on the
mobile terminal and is, therefore, less commonly exercised.
5) Based on the received signal strength (Received Signal Strength Indication – RSSI) – The
free space signal propagation is characterised with predictable attenuation dependent on
the distance from the source. Moreover, in real conditions, the attenuation also largely
depends on the obstacles and the configuration of the propagation path. That is why there
are various mathematical models which describe the wave propagation for diverse
surroundings and, ultimately, estimate the signal attenuation for the observed environment.
This approach grants the distance of the entity whose position is being determined, to one
or more transmitters.
6) Based on the fingerprint of the location (Database Correlation or Location Fingerprinting)
– With this approach, the certain, location dependant, information is acquired in as many
Reference Points (RPs) across the coverage area of the technique. This data is stored into so
called Location Fingerprints Database. Afterwards, when the actual position determination
process takes place, the information gathered at the unknown location is compared with the
pre-stored data and the entity’s position is estimated at a location of a pre-stored fingerprint
from the database whose data are “closest” to the measured data.
Most often, the estimated position with TOA and RSSI approaches is determined by
lateration. The process of lateration consists of determining the position of the entity when
the distance between the entity and one or more points with identified positions (i.e.
refernce points) is known. To uniquely laterate the position in N-dimensional space, the
distances to N+1 reference points ought to be known. With TDOA approach, the estimated
position is obtained as a cross-section of two or more hyperbolas in two-dimensional space,
or three or more hyperbolic surfaces in case of three-dimensional space. The process of
angulation is employed with AOA and DOA approaches. This process estimates the
location of a user as a cross-section of at least two rays (half-lines) originating at known
locations. The lateration and angulation processes are depicted in Fig. 1. As for the Location
Fingerprinting approach, the estimated location is obtained by utilizing the correlation
algorithm of some sort. This algorithm determines, following a certain metric, the
“closeness” of the gathered data to the pre-stored samples from the location fingerprinting
database.
Apart from these, basic, approaches, there are a number of other choices and hybrid
techniques that combine the aforementioned approaches when determining the estimated
position of the user.
a) b)
r2
r1
d
r3
 
Fig. 1. The processes of estimating a user location: a) Lateration and b) Angulation (Green
circles represent the known positions and the red cross stands for the estimated location)
3. Classifications of Positioning Systems

There are more than a few classifications of positioning systems. While some of them are
very strict, others can be very arbitrary and overlapping. Without the need to judge or
justify any of them, the most common ones are given herein.
Regarding the type of provided information, positioning techniques can be split into two
main categories: Absolute and Relative positioning.
Absolute positioning methods consist in determining user location from scratch, generally
by using a receiver and a terrestrial or satellite infrastructure. A well-known example of
systems based on “absolute positioning” is the American GPS.
Relative positioning methods consist in determining user location by calculating the
movements made from an initial position which is known. These methods do not rely on an
external infrastructure, but require additional sensors (e.g. accelerometers, gyroscopes,
odometers, etc). Inertial Navigation Systems used in commercial and military aircrafts are a
good example of systems based on relative positioning.
LBSs currently offered by wireless telecommunication operators or by service providers are
all based on absolute positioning methods and not on relative positioning methods, since
these services are offered to users whose initial position is generally not known.
Within the “absolute positioning” family, the measurements and processing required for
determining user’s location can be performed in many different ways and rely on different
means. Thus, many different absolute positioning methods can be used for determining
user’s location. These methods can be clustered into different groups, depending on the
infrastructure used. Hence, the positioning techniques can be divided into:
 Satellite-based,
 PLMN-based (Public Land Mobile Network based), and
 Other (such as: WLAN, Bluetooth, RFID, UWB, etc).
The first group, which is known by the largest audience, is the “Satellite positioning” group.
This group relates to the positioning methods which are based on the use of orbiting
satellites, such as the GPS, Glonass or Galileo. Many applications and services based on
satellite positioning have been developed during the past years (e.g. in-vehicle navigation,
fleet management, tracking and tracing applications, etc). They generally require the use of
dedicated receivers. Today, more and more devices such as PDAs or mobile phones include
a satellite positioning capability, and this trend should persist in the future.
The second group, the “PLMN positioning” group, corresponds to the location techniques
which have been developed for public land mobile networks. Initially deployed in the US
under the pressure of the FCC mandate which forces US carriers to locate users placing calls
to the 911 emergency number, location technologies are now being implemented in most of
European wireless telecommunication networks for commercial purposes. Most of cellular
positioning methods are incorporated in mobile telecommunication standards
(2G/2.5G/3G/3.5G), but some solutions remain based on proprietary techniques.
The third and last group, the “Other positioning” group, corresponds to those technologies
which have not been developed specifically for positioning purposes, but that can be used,
in addition to their primary function, for determining user’s location. These technologies
encompass WLAN and Bluetooth for instance.
Another distinction can be made, depending on the “place” where the position calculation
is made. In some cases, the main processing is performed at the terminal level. In other
cases, the main processing is performed in the network. Therefore, the positioning
techniques can be classified into:
 Network-based (also referred to as mobile-assisted), and
 Terminal- or Mobile-based (also referred to as network-assisted).
Satellite technologies, as a rule, fit in the Terminal-based positioning techniques. As for the
positioning techniques from the PLMN and other groups, they can not be apriori associated
to either of the Terminal- or Network-based groups.
Finally, the positioning techniques can be classified according to the environment of their
coverage. Hence, the positioning techniques can be divided into:
 Outdoor, and
 Indoor.
Although there are intense research efforts to adopt the Satellite-based positioning
techniques for Indoor environment, thay are still considered to fit into Outdoor groop.
PLMN-based positioning can be implemented in both Indoor and Outdoor envornments,
whereas techniques from “Other” group usually fit Indoor environments.
Positioning techniques designed for a particular Indoor environment in most cases fit into
Relative positioning group.
Bearing in mind the ongoing convergence process of telecommunication systems and
numerous, newly developed, hybrid positioning techniques, the indoor/outdoor
categorization as well as other aforementioned classifications ought to be regarded more as
guidelines than as strict lines that divide techniques into disjoint sets.
4. Non-Radio Indoor Positioning Systems

This section contains a brief overview of the non-radio positioning systems most commonly
used in indoor environment.
4.1 IrDA Positioning Systems

IrDA technologies are based on devices with infrared light transceivers. This light occupies
the part of spectrum between the visible light and the radio-waves. Upon encountering an
obstacle, such as wall, the major part of the IR light’s energy is being absorbed. Therefore, in
order to communicate properly, two IR devices must have unobstructed Line of Sight (LoS)
path between them. This poses a limitation for employing this technology in positioning
purposes.
The most popular application of this technology for positioning use is the “Active Badge”
technique (Want et al., 1992). The person or entity, whose position is being determined,
possesses a device, badge alike, which periodically emits its ID code via IR transmitter. The
IR sensors must be deployed in the coverage area (building). The position of the user is then
determined based on the Cell-ID principle. With respect to the attributes of the IR light, the
sensors must be deployed in every room in which the positioning feature is needed.
Consequently, the accuracy of this technique is on a room level.
Other techniques based on this technology offer various accuracy and applications. The
systems with greater number of IR receivers and transmitters on each device are proposed
(Krohn et al., 2005). These systems are able to accurately estimate the position of a mobile
communication device (e.g. PDA, laptop, digital camera, etc.) in order to allow them to
automatically synchronise or perform other location dependent tasks. These activities are
supposed to be performed on a flat, table alike surface. The obtained distance error is less
than 20cm in more than 90% of the cases. On the other hand, there are systems that augment
the “Active Badge” technique by using more IR sensors, micro VGA display and,
optionally, video cameras. These systems provide so called Argumented Reality (Maeda et
al., 2003). The typical application of an Argumented Reality system would be for the
museum environment, where the visitor would be, via micro display (in eyeglasses, for
example), fed with the information related to the exhibit he is currently experiencing.
4.2 Ultrasound Positioning Systems

The term ultrasound is related to the high frequency sound waves, above the part of
spectrum perceivable to the human ear (20kHz). Although the ultrasound is most
frequently used in medicine, there are other areas of application such as: biomedicine,
industry (e.g. flow-meters), chemistry, military applications (sonic weapon), etc. As for the
positioning purposes, the greatest benefit of using the ultrasound positioning is the product
of a fact that ultrasound propagates through the air at limited speed, which is by far smaller
than the speed of light. Therefore, the implementation of techniques based on time of flight
(i.e. TOA, TDOA) of signal is very much facilitated. Moreover, the mechanic nature of
sound waves grants ultrasound positioning techniques immunity to electromagnetic
interference which could also be considered as an advantage. It ought to be pointed out that
ultrasound waves do not penetrate, but rather reflect of walls. Therefore, the ultrasound
receiver, in order to detect the signal, must be in the same room as transmitter but LoS is not
necessary.
Ultrasound positioning systems can be classified according to the number of ultrasound
“base stations” (transmitters and/or receivers) in each room (Dijk, 2004). The basic
ultrasound positioning technique comprises one receiver in each room, and a ultrasound
emitting tag which is worn by the entity that needs to be positioned. In this case, the
accuracy is on the level of the room. These systems are commercially available for some
time now.
More sophisticated ultrasound positioning systems invoke the use of a greater number of
transmitters in each room as well as the use of RF (seldom IR) signals for precise
determining the time delay (Fraser, 2006). In this case, the controlling unit, which is
connected to all the ultrasound emitters in one room as well as with RF transmitter,
determines the exact time when each of the transmitters is about to send its chirps.
Commonly, the RF signal is emitted first and then the chirps from all ultrasound
transmitters are emitted separated by known time intervals. The receiver, knowing the
separating time intervals and the propagation speed of RF and ultrasound waves, can now
calculate, based on the time it received each of the chirps, the distance to each of the
ultrasound emitters. The position is then determined by lateration. Consequently, for three-
dimensional positioning at least four transmitters per room are required. The accuracy is in
range of 10cm in 90% of the cases.
Furthermore, the system that eliminates the need for RF transmitter has been developed
(McCarthy & Muller, 2003). With this system, the processing power of the receiver can be
reduced, and the whole system is less complex. The transmitters are cyclically emitting
chirps in constant time intervals whereas the receiver is employing an extended Kalman
filter for resolving the chirp transmission and receipt times.
5. Indoor Radio Positioning Systems

The RF positioning techniques employ different parts of the frequency spectrum. Some are
implemented on existent short-range radio interfaces and serve as added services, while
others are especially developed for positioning. The most common RF technologies which,
through the use of these techniques, enable positioning are: RFID, UWB, Bluetooth and
WLAN.
5.1 RFID (Radio-Frequency IDentification) Positioning Systems

The beginnings of this technology go far back to the time of the Second World War Over
the recent years, due to the cheaper RFID components, the expansion of this technology is
occurring.
RFID system consists of tags, reader with antenna and accompanying software. The tags are
usually placed on entities whose position needs to be determined. The Line of Sight
between the tag and a reader is usually not necessary. The tags can contain additional
information apart from its ID code which broadens the usage this technology.
There are three types of RFID tags:
 Passive tags do not have their own power supply. In order to operate, they use the
energy, induced on their antenna, from the incoming radio wave from the reader.
Using that energy, the passive tag replays by emitting its ID code and, optionally,
additional information. Passive tags have very limited range (from a few cm up to
a couple of meters). Their advantage is within the scope of cheap construction,
compact size and cheap production.
 Active tags are encompassed with power supply witch enables them unrestrictive
signal emission. This kind of tags is more reliable and immune to highly polluted
RF environments. Their range can go up to a few hundreds of meters.
 Semi-active tags are equipped with battery power supply. Recent constructions
enable a battery life span of more than 10 years.
RFID devices can operate in different frequency bands: 100 – 500 kHz, 10 - 15 MHz, 850 –
900 MHz, and 2.4 – 5.8 GHz (Don Chon et al., 2004).
RFID positioning techniques are based on knowing the position of the reader. When the
tagged object enters the range of the reader, its position is assumed to be equal to the
position of the reader (similar to Cell-ID). Correspondingly, it is possible to deploy tags
across the coverage area. In that case the reader is mounted on the entity whose position is
being determined. The accuracy depends on the density of deployed objects (tags/readers)
across the coverage area. With active tags, the positioning accuracy can be upgraded with
the RSSI information.
Most common application areas of RFID technology are in replacing the barcode readers,
product tracking and management, personal documents identification, identification
implants for humans and animals, etc. It is interesting to mention that the latter
aforementioned application raises numerous ethical issues and there are organized groups
worldwide opposing the implementation of this technology.
5.2 Bluetooth Positioning Systems

Bluetooth is a short-range, low-consumption radio interface for data and voice
communication (Muller, 2001). Initially conceived in the mid 90s by the Ericsson Mobile
Communication as a technology that ought to replace the cable in personal
communications, Bluetooth shortly gained significant popularity. Ericsson was joined by
IBM, Microsoft, Nokia and Toshiba. They formed Bluetooth Special Interest Group (SIG)
with an aim to standardize Bluetooth specifications. Independent group, called Local
Positioning Working Group, had a goal of developing the Bluetooth profile which would
define the position calculation algorithm as well as the type and format of the messages that
would enable Bluetooth devices to exchange position information.
The basic Bluetooth specification does not support positioning services per se (Bluetooth
Special Interest Group Specification Volume 1 and 2, 2001). In absence of such support,
various research efforts have produced diverse solutions. Bahl and Padmanabhan used the
RSSI information for in-building locating and tracking (Bahl & Padmanabhan, 2000). Patil
introduced the concept of reference tags and readers (Patil, 2002). He also investigated
separately cases when Bluetooth supports and does not give support to RSSI parameter. On
the other hand, the research by Hallberg, Nilsson and Synnes goes to saying that RSSI
parameter is unreliable for positioning purposes and that its employment ought to be
avoided with Bluetooth positioning systems (Thapa & Case, 2003).
In addition, there are ideas of exploiting other parameters than RSSI for positioning
purposes. Link Quality and Bit Error Rate (BER) are most commonly referred in this
context. However, it should be stated that these solutions are still under development, and
that Link Quality is not uniquely defined and is therefore dependent on the equipment
manufacturer. Also, BER parameter is not defined in the basic Bluetooth specifications and
must be extrapolated from the message received as a response to echo command supported
at L2CAP layer. All in all, these parameters undoubtedly contain location dependent
information, but the extraction of that information is still subject to research.
The accuracy of Bluetooth positioning systems is decreasing with the increase in the
maximal range of the system (Hallberget al., 2003). That is, with the range increase, the
positioning system uncertainty is increased as well, therefore the accuracy is worsened. The
improvement of accuracy can be achieved through communicating with more than one
Bluetooth nodes and possibly utilizing some of the aforementioned parameters (RSSI, Link
Quality, BER). Finally, the major application of Bluetooth technology is expected in ad-hoc
networks and the positioning techniques and LBS should be conceived and designed
accordingly.
5.3 UWB (Ultra-WideBand) Positioning Systems

Ultra-wideband is a short-range high data throughput technology. The ultra-wideband
signal is defined (Harmer, 2004) as a radio-signal that occupies at least either 500MHz of
frequency spectrum or 20% of the central frequency of the band. There are many ways in
which the UWB signal can be generated. Two, most important from the positioning point of
view, are:
1) Impulse UWB – By generating very short impulses, with sub nanosecond duration, that
are mutually separated several tenths of nanoseconds. Clearly, this signal inherently
possesses very wide band.
2) Frequency Hopped UWB – By generating the typical DSSS (Direct-Sequence Spread
Spectrum) with the signal spectrum ranging from 10 to 20MHz which is then hopped
around 1GHz frequency, applying between 10 and 100 thousands of hops per second.
Unlike conventional radio-signals, the impulse UWB signals are practically immune to
multipath propagation problems. With conventional signals, the reflected component of the
signal is, in its large part, overlapped with the component that is travelling the direct path.
Hence, the direct and reflected component interfere at the receiver causing fading. Contrary
to that, with impulse UWB technology, due to the very short pulse duration, the reflected
component is most often arriving at the receiver after the direct component has been
completely received. With respect to this feature, the UWB positioning techniques utilising
high resolution TOA approach come as the logical choice. Typically, the position accuracy
of 1m in more than 95% of the cases is achievable.
Employing the mobile nods of the UWB network for accuracy improvement is also under
research. Computer simulation (Eltaher & Kaiser, 2005) shows that the positioning error
could be further reduced by employing a larger number of antennas with the beamforming
capabilities.
Bearing in mind the amount of research in this area, the wider scale commercialisation of
indoor UWB positioning systems can be expected in proximate future.
5.4 WLAN Positioning Systems

Positioning techniques in WLAN networks are growing in popularity. The reason for this
can be looked in-between the widespread of 802.11 networks and the fact that a broad scope
of LBSs can be brought into an existing network without the need for any additional
infrastructure. There are a number of approaches to the positioning problem in WLAN
networks. Unquestionably, the most popular ones are based on the Received Signal
Strength Information (RSSI). Nevertheless, there are other approaches that depend on
timing measurements or require additional hardware but offer superior accuracy and/or
faster implementation in return (Llombart et al., 2008; King et al., 2006; Sayrafian-Pour &
Kaspar, 2005).
Positioning with the use of RSSI parameter can be, in its essence, regarded as the path loss
estimation problem. The nature of the path loss prediction in an indoor environment is
extremely complex and dependent on a wide variety of assumptions (e.g. type of the
building, construction, materials, doors, windows, etc.)(Nešković et al, 2000). Even if these
basic parameters are known, precise estimation of the path loss remains a fairly complex
task.
Depending on the side on which the position calculation process takes place, positioning in
WLAN networks with the use of RSSI parameter can be either network-based or client-
based. Whereas the client-based solutions gather the RSSI vector from the radio-visible APs,
the network-based solutions have a central positioning engine which collects the client’s
signal strength vector from the APs and produces the position estimate. The network-based
solutions do not require clients to have a specific software installed which is of great essence
for security purposes. Moreover, the client does not need to be associated with the network
– the positioning can be done solely based on the probe requests the client sends (in case of
active scanning). Network-based solutions could also have an important advantage over the
client-based ones when used in WLAN networks employing the Automatic Radio
Management (ARM). This centralized mechanism is used to obtain the optimal radio
coverage by changing the channel assignment and adjusting the output power and/or
radiation pattern of the APs. Contrary to the client-based solutions, the network-based
positioning engine could take into account the changes made by ARM mechanism while
the ARM mechanism would present a setback for the client-based solutions. On the other
hand, client’s Network Interface Cards do not have to be consistent regarding the radiated
power which may, depending on the positioning algorithm used, present an analogue
problem for network-based solutions. In this work, for explanatory purposes, usually the
client-based solution will be presented. However, the reader should keep an open mind
towards the analogue network-based option.
Regarding the approach used to determine the user’s position, WLAN positioning
techniques can be categorised as: propagation model based, fingerprinting based or hybrid.
Propagation model based techniques rely on statistically derived mathematical expressions
that relate the distance of an AP with the client’s received signal strength. The estimated
position of the user is then obtained by lateration. Therefore, if there are less than three
radio-visible APs (for two-dimensional positioning) the estimated user’s position is
ambiguous. Also, the model derived for one specific indoor environment is usually not
applicable to other indoor environments.
Fingerprinting techniques are most commonly used for WLAN positioning. They are
conducted in two phases: the off-line or training phase, and the on-line or positioning
phase. The off-line phase comprises collecting the RSSI vectors from various APs and
storing them, along with the position of the measurement, into a fingerprinting database. In
the on-line phase, the estimate of the user’s position is determined by “comparing the
likeliness” of the RSSI vector measured during the on-line phase with the previously stored
vectors in the database. The fingerprinting process is shown in Fig. 2. These techniques have
yielded better performance than other positioning techniques, but are believed to have a
longer set-up time.
a) measurement
location fingerprint database
AP(1) client at known location (x,y)1 RSSI1,RSSI2,...

(RP i)
AP(2) (x,y)2 RSSI1,RSSI2,...
AP(3) store (x,y)3 RSSI1,RSSI2,...
AP(m) (x,y)n RSSI1,RSSI2,...
b)
(?,?) RSSI1,RSSI2,... algorithm (x,y)
user (client) at unknown position user’s location
Fig. 2. Two phases of positioning: a) training phase – mobile client is recording RSSI vectors
across RPs and stores them in fingerprint database, and b) positioning phase – based on the
measured RSSI vector and database access, the algorithm estimates the user’s location
Hybrid techniques combine features from both propagation modelling and fingerprinting
approaches, opting for better performances than propagation model techniques and shorter
set-up time than fingerprinting techniques (Wang & Jia, 2007).
The prospects of using RSSI parameter for indoor positioning were first systematically
analysed in “RADAR” (Bahl & Padmanabhan, 2000). According to this research, it is better
to use RSSI than SNR (Signal to Noise Ratio) for positioning purposes since the RSSI
parameter is much more dependent on the client’s position than SNR. Two algorithms to
establish the user’s location were proposed. The first one is the Nearest Neighbour (NN)
algorithm which compares the RSSI vector of a mobile client against the RSSI vectors
previously stored in the fingerprinting base. An extension to the proposed algorithm was
also considered: the estimated location is not identified as only one RP whose RSSI vector is
closest to the observed RSSI vector, but calculated as a “middle” point of kclosest RPs (kNN
algorithm). This analysis has shown that algorithm performance improved for k= 2 and k=
3. For larger k, the performance had started to decrease. The second algorithm is based on a
simple propagation model with Rician distribution assumed. It ought to be emphasized that
both approaches require a minimum of three radio visible access points (APs). The
measuring campaign comprised 70 RPs. At each RP measurements were made for four
orientations of the receiver, and each measurement was averaged from 20 samples.
To produce the maximum amount of information from the received RSSI vectors, the
Bayesian approach was proposed (Li et al., 2006). This concept yields better results than the
NN algorithm. The Bayes rule can be written as:

p l |o
t t
  p  ot | lt  p  lt  N (1)
where ltis location at time t,otis the observed RSSI vector at time t, whileN is a normalizing
factor that enables the sum of all probabilities to be equal to 1. In other words, at a given
time t, the probability that a client is at locationlt, if the received RSSI vector is ot, is equal to
the product of the probability to observe RSSI vector otat location ltand the probability that
the client can be found at location lt. The process of estimating client’s location is based on
calculating the conditional probability p  lt |ot  for each RP. The estimated client’s location is
equal to the RP with the greatest conditional probability. To accomplish this task, two terms
on the right hand side of Eq. (1) ought to be calculated. The first term, also referred to as the
likelihood function, can be calculated based on the RSSI map (for all RP) using any
approach that will yield probability density function of observation otfor all RPs. As for the
a priori probability p  lt  , it ought to be calculated according to the client’s habits. However,
for most cases the assumption of uniform distribution across all RPs is valid. The
measurements were made at 70 RPs. As with the previously discussed techniques, the
measurements were made for four orientations of a receiver, and each measurement was
averaged from 20 samples.
Another project, named Horus (Youssef & Agrawala, 2005; Eckert, 2005), had the goal of
providing high positioning accuracy with low computational demands. This is also a
probabilistic approach in which time series of the received signal strength are modelled
using Gaussian distributions. Due to the time dependence of the signal strength from an
observed AP, the authors of this project have shown that the time autocorrelation between
the time adjacent samples of signal strength can be as high as 0.9. To describe and benefit
from such behaviour, they have suggested the following autoregressive model:
s  s  1    , 0    1 (2)
t t 1 t
where t is the noise process and st is a stationary array of samples from the observed AP.
Throughout the off-line phase, the value of parameter a is assessed at each RP and stored
into the database along with Gaussian distribution parameters m and s . In the on-line
phase, Gaussian distribution is modified according to the corresponding values of a
retrieved from the fingerprinting database. Alike to the kNN algorithm, the Horus system
estimates the client’s location as a weight centre of kRPs with the highest probabilities. The
principal difference to the kNN algorithm is that, in case of Horus system, the kmost likely
RP are multiplied with their corresponding probabilities. For verification purposes, the
authors made measurements at 612 RPs, and each measurement was averaged from 110
samples.
More relevant information about the statistical modelling approach towards location
estimation can be found in (Roos et al., 2002) and in the references found therein.
Battiti et al. (2002) were the first to consider using Artificial Neural Networks (ANNs) for
positioning in WLAN networks. This approach does not insist upon a detailed knowledge
of the indoor structure, propagation characteristics, or the position of APs. A multilayer
feedforward network with two layers and one-step secant training function was used. The
number of units in the hidden layer was varied. No degradation in performance was
observed when the number of units grew above the optimal number. For verification
purposes, measurements were made at 56 RPs, and each measurement was averaged from
100 samples.
In most studies, WLAN positioning techniques are compared on the subject of their
accuracy while other attributes of a positioning technique such as latency, scalability, and
complexity are neglected. Another aspect that is seldom analyzed is size of the environment
in which the technique is implemented.
It also ought to be pointed out that averaging the RSSI vectors in the on-line phase has an
immense impact on the technique’s latency, so the scope of location based services that
could be utilized with such techniques is significantly narrowed. Moreover, bearing in mind
that all presented approaches require at least three radio-visible APs in each RP (which is
seldom the case in most WLAN installations), feasibility of sound frequency planning is
uncertain. Consequently, the degradation of packet data services is inevitable with respect
to positioning in larger indoor areas (i.e. large number of APs is required). Enabling the
radio-visibility of three APs across the indoor environment is usually constructively
irrational and economically unjustified. Hence, the presented techniques cannot be applied
to the majority of existing WLAN networks optimized for packet data services.
Finally, there are other studies that accompany the research for sophisticated positioning in
WLAN networks. Other relevant research efforts comprise the impact of Network Interface
Card on the RSSI parameter, compensation of small-scale variations of RSSI, clustering of
locations to reduce the computational cost of positioning, use of spatial and frequency
diversity, methods for generating a larger location fingerprinting database by interpolation,
and unequal fusing of RSSI from different APs (Kaemarungsi, 2006; Youssef & Agrawala,
2003; Ramachandran & Jagannathan, 2007; Li et al., 2005; Zhang et al.,2008).
6. Cascade-Connected ANN Structures for WLAN Positioning

The ANNs are an optimisation technique known to yield good results with noise polluted
processes (Hasoun, 1995). They are generally classified as a fingerprinting technique. In the
off-line phase, the set of collected RSSI fingerprints is used to train the network and set its
inner coefficients to perform the positioning function. In the on-line phase, the trained
network replaces trilateration and position determination processes.
Two basic concepts, a single ANN and a set of cascade-connected ANNs structures with
space partitioning, have been presented herein. These models were implemented in Matlab
and verified on a 147m x 67m test bed with eight APs. For training purposes, the traingda –
gradient descent training function with adaptive learning rate was selected. All neural units
had the hyperbolic tangent sigmoid transfer function. Being that the input probability
distribution function of RSSI values is near Gaussian, the Mean Square Error (MSE) was
selected as a criterion function (Hanson, 1988).
Regarding the purpose that ANN is intended for and, moreover, the nature of the problem,
it has been concluded that multilayer feedforward neural networks with error
backpropagation have substantial advantages in comparison to other structures (Nešković,
2000). The outer interfaces of the ANN must match the number of the APs on the input side
(i.e. eight inputs), and the number of coordinates as outputs (i.e. two outputs).
Multilayer feedforward networks can have one or more hidden layers with perceptron
units. The hidden layers with corresponding perceptron units form the inner structure of
the ANN. There is no exact analytical method for determining the optimal inner structure of
the network. However, there are algorithms that, starting with an intentionally oversized
network, reduce the number of units and converge to the optimal network structure. Also,
there are other algorithms such as the cascade correlation learning architecture (Fahlman &
Lebiere, 1990) that build the network towards the optimal structure during the training
process. However, being aware of the fact that these procedures can be complex and that
determining the most optimal structure was not the central scope of this research, we
intentionally slightly oversized our network’s inner structure knowing that an oversized
network will not yield degradation in performance. We also adopted that the first hidden
layer ought to have more perceptrons than the input layer so that the input information is
quantified and fragmented into smaller pieces (Shang & Wah, 1996). The number of
perceptrons in the following hidden layers ought to decrease, converging to the number of
perceptrons in the output layer. Bearing that in mind, the chosen structure for single ANN
(type 1) approach consisted of the input layer, three hidden layers and the output layer. The
number of perceptrons per layer was (from input to output) 8-15-9-5-2.
When utilizing space partitioning, the positioning process is split into two stages where
each stage could be implemented with the most suitable model. In this case, the two-step
space partitioning is implemented utilizing cascade-connected ANNs. The block scheme of
this system is shown in Fig. 2.
8 8 2 2
8 No. of
SubSps (n) 8 2
8 2
First Stage Second Stage

Fig. 2. Cascade-connected ANNs system structure (the input is the observed RSSI signal
vector, APs RSSI, and the position estimate vector, Pos Est, is the output)
In the first stage, an ANN (type 2) is used to determine the likeliness of a measured RSSI
vector belonging to one of the subspaces. This ANN (type 2) has 8 inputs and the number of
outputs is equal to the number of subspaces the environment is partitioned to. Each output
corresponds to the likeliness that a received RSSI vector originates from a particular
subspace. The outputs of the type 2 ANN, SubSp Ln, are connected to the Forwarding
block which, depending on the inputs, employs only one of the second stage networks by
forwarding the APs RSSI vector.
The inner structure of ANN (type 2) is designed using the same guidelines as with the
single ANN model. Therefore, it also has three hidden layers and the number of perceptron
units in those layers is varied to fit the different number of subspaces. The second stage
ANNs are type 1 networks with structure identical to the previously described ANN used
with the single ANN approach.
In the off-line phase, type 2 ANN is trained with the fingerprinting database that originates
from the whole environment. The targeted output vector has only one non-zero element
(equal to 1). The index of that element corresponds to the number of the subspace from which
the RSSI vector originates. Type 1 networks are trained following the training methodology
from the single ANN approach with the only difference being that each type 1 ANN is trained
with only the part of the fingerprinting database which originates from a particular subspace.
In the on-line phase, the first stage ANN estimates the likeliness that the received RSSI vector
originates from a particular subspace. The Forwarding block then determines the most likely
subspace by searching for the maximum value in the output vector from the ANN (type 2)
and forwards the APs RSSI vector only to the second stage ANN that correspond to that
subspace. The appropriate second stage ANN then determines the estimated position of the
user and, finally, the collecting block forwards that estimate to the structure output.
Several space separation patterns were chosen yielding a different number of subspaces ranging
from 4 to 44. The space partitioning patterns that have been employed are shown in Fig. 3.
a) b) c)
d) e) f)
g) h) i)
Fig. 3. Space partitioning patterns: a) no space partitioning (1x1), b) 2x2, c) 2x3, d) 2x4, e)
3x4, f) 4x6, g) x24, h) x32, and i) x44
The partitions with a smaller number of subspaces were made on geometrical bases.
However, with the increase in the number of subspaces, the subspace size decreased until it
came to a room size level. It was then worth to consider partitioning space in an other
manner. Starting with 24 subspaces (which was also portioned on geometrical bases), the
partitions were made on “logical” bases (i.e. x24, x32 and x44). This logical separation opted
for subspaces to be as homogeneous in the propagation manner as possible (e.g.
partitioning was made trough walls wherever possible). Note, the single ANN model is
herein referred to as 1x1 partitioning.
For the purpose of determining the optimal training parameters, as well as the optimal
training duration, the complete set of measurements was split into two subsets. The larger
subset was used to train the ANNs, while the smaller, containing measurements from a 100
randomly chosen RPs, was used to validate the obtained models.
The results obtained for different space partition patterns, for optimally trained ANNs, are
presented in Table 1.
From Table 1, it can be seen that, with geometrical partitioning, the overall median and
average distance errors decrease with the increase in number of subspaces. This behaviour
is even more emphasized with the distance errors in the correctly chosen subspace which
confirms the influence of environment size on positioning accuracy. When concerning the
logical partitioning, slightly better results are obtained for 24 subspaces (4x6 vs. x24) but,
with the further increase in the number of subspaces, the average distance error is starting
to rise again. Also, with the increase in the number of subspaces the probability of correct
subspace being chosen declines as expected while the probability of correct room estimation
rises from 26% for a 1x1 positioning to as much as 66% for a x24 configuration, after which
it starts declining a little.
Pattern 1x1 2x2 2x3 2x4 3x4 4x6 x24 x32 x44
Overall Average DEa [m] 9.26 9.00 8.97 8.91 8.54 8.28 8.14 8.58 9.11
Overall Median DEa [m] 7.75 7.49 6.87 5.86 5.59 5.10 4.57 4.70 4.44
Average DEa in ISb [m] - 21.3 22.7 21.2 19.0 18.0 18.4 19.5 19.2
Median DEa in ISb [m] - 15.4 17.4 15.3 16.3 14.7 17.5 15.8 16.1
Average DEa in CSc [m] 9.26 8.35 6.99 6.96 5.76 4.20 4.07 3.78 3.72
Median DEa in CSc [m] 7.75 7.33 6.13 5.52 4.40 3.87 3.56 3.39 3.32
Probability of CSEd 1.00 0.95 0.87 0.86 0.79 0.71 0.72 0.69 0.65
Probability of CREe 0.26 0.42 0.48 0.52 0.58 0.62 0.66 0.62 0.61
a Distance Error, b Incorrect Subspace, c Correct Subspace , d Correct Subspace
Estimation, e Correct Room Estimation
Table 1. Performance overview for different partitioning patterns
To better understand and discuss the performances of cascade-connected ANNs with space
partitioning, we observed and compared the distance error’s Cumulative Distribution
Function (CDF) of a single ANN approach with the cascade-connected ANNs. Fig. 4. shows
the obtained CDFs for representative space partitioning patterns.
a) b)
90 90
1x1 1x1
80 2x2 80 2x3
2x2H 2x3H
70 70
60 1x1 60
CDF [%]
CDF [%]
2x2
50 2x2 H 50
2x3
2x3 H
40 40
2x4
2x4 H
30 30
3x4
3x4 H
20 20
10 10
0 5 10 15 20 25 0 5 10 15 20 25
c) d)
Distance Error [m] Distance Error [m]
90 90
80 1x1 80 1x1
2x4 3x4
2x4H 3x4H
70 70
60 60
CDF [%]
CDF [%]
50 50
40 40
30 30
20 20
10 10
0 5 10 15 20 25 0 5 10 15 20 25
Distance Error [m] Distance Error [m]
e) f)
90 90
80 1x1 80 1x1
4x6 x24
4x6H x24H
70 70
60 60
CDF [%]
CDF [%]
50 50
40 40
30 30
20 20
10 10
0 5 10 15 20 25 0 5 10 15 20 25
g) Distance Error [m]
h) Distance Error [m]
1x1 1x1
x32 x44
x32H x44H
Fig. 4. Cumulative Distribution Function of distance error: a) 1x1and 2x2 partitioning and
correct subspace estimation – 2x2 H; b) 1x1and 2x3 partitioning and correct subspace
estimation – 2x3 H; c) 1x1and 2x4 partitioning and correct subspace estimation – 2x4 H; d)
1x1and 3x4 partitioning and correct subspace estimation – 3x4 H; e) 1x1and 4x6 partitioning
and correct subspace estimation – 4x6 H; f) 1x1and x24 partitioning and correct subspace
estimation – x24 H; g) 1x1and x32 partitioning and correct subspace estimation – x32 H; h)
1x1and x44 partitioning and correct subspace estimation – x44 H
The green filled areas on Fig. 4. could be considered as a partitioning gain in comparison to
1x1 positioning, while the red filled areas could be considered as partitioning loss. It can be
seen that, with geometrical partitioning, Fig. 4. a) – e), the gain areas are increasing with the
increase in the number of subspaces. When concerning logical space partitioning Fig. 4 f) –
h), it can be noticed that the best performances are obtained with x24 pattern – average
distance error 8.14m, median error 4.57m. With the further increase in the number of
subspaces, the benefit of decreasing the median error has faded, even though the median
error in correct subspace continues to decrease, whereas the average distance error is
starting to rise again due to the augmentation in probability of incorrectly chosen subspace.
In other words, with the further increase in the number of subspaces, the partitioning gain
surfaces are still expanding however, the partitioning loss surfaces are rising as well.
Furthermore, it should be noticed that with the increase in the number of subspaces, the
CDF is starting to create a knee roughly around 60th percentile. This has two effects: the
green surfaces are getting larger as discussed and the crossing angle between the space
partitioning model and 1x1 positioning is increasing while the crossing point between the
two is being pushed towards lower percentiles. The latter of the two effects has a negative
impact on positioning performances.
Finally, if the Average DE in correct subspace, from Table 1, is compared with the average
subspace area (the total area size divided by the number of subspaces), it can be seen that
with the increase in size of the subspaces the increase in average error is getting saturated.
So, given the constant RPs and APs density, the further increase in size of the test bed
should induce only the minor rise of the DE. This also goes to say that the chosen
verification environment was large enough to comprehensively explore the influence of the
test bed size on positioning accuracy.
7. References
Bahl, P. & Padmanabhan, V. (2000). „Radar: An in-building RF_based user location and
tracking system“, Proceedings of the IEEE Infocom 2000, Tel-Aviv, Israel, vol. 2,
Mar. 2000, pp. 775-784.
Battiti, R., Nhat, T. L., Villani, A. (2002). Location-aware Computing: A Neural Network
Model For Determining Location in Wireless LANs. Technical Report # DIT-02-
0083 (Feb. 2002)
Bluetooth Special Interest Group (2001). Specification Volume 1, Specification of the
Bluetooth System, Core. Version 1.1, February 22, 2001.
Bluetooth Special Interest Group (2001). Specification Volume 2, Specification of the
Bluetooth System, Profiles. Version 1.1, February 22, 2001.
Hae Don Chon, Sibum Jun, Heejae Jung, Sang Won An, (2004). „Using RFID for Accurate
Positioning“, Journal of Global Positioning Systems, 2004, Vol. 3, No. 1-2: 32-39
Collomb Frédéric, (2002). Location Service Study Report (Loc_Serv_Study_Rep_PU.doc),
Mobile and Vehicles Enhanced Services, 2002.
Eckert, K. (2005). Overview of Wireless LAN based Indoor Positioning Systems, Mobile
Bussines Seminar, University of Mannheim, Germany, (2005)
Eltaher A., Kaiser T., (2005). „A Novel Approach based on UWB Beamforming for Indoor
Positioning in None-Line-of-Sight Environments“, RadioTeCc, October 26-27, 2005,
Berlin, Germany.
Esko O. Dijk, (2004). „Indoor ultrasonic position estimation using a single base station“,
Technische Universiteit Eindhoven, 2004. October 6, 2004, p44-45
Fahlman, S., Lebiere, C. (1990). The cascade-correlation learning architecture. Advances in
Neural Information Processing Systems 2, pp. 524–532 (1990)
Fraser M., (2006). „Mobile and Ubiquitous Computing: Sensing Location Indoors.“,
COMSM0106, 2006.
Hallberg, Nilsson, Synnes, (2003). „Positioning with Bluetooth“, Telecommunications, 2003.
ICT 2003. 10th International Conference on, Volume 2, Issue , 23 Feb.-1 March 2003
Page(s): 954 - 958 vol.2
Harmer D., (2004).„Ultra Wide-Band (UWB) Indoor Positioning“, Thales Research and
Technology UK Ltd. ARTES 4 Project. ESTEC December 2004.
Hasoun H. M. (1995). Fundamentals of Artificial Neural Networks. Massachusetts Institute
of Technology (1995)
Hanson, S. J., Burr, D. J. (1988). Minkowski-r Backpropagation: Learning in Connectionist

Models with non-Euclidean Error Signals, Neural Information Processing Systems
(Denver, 1987), (editor Anderson D. Z.), pp. 348-357, American Institute of Physics,
New York (1988)
Kaemarungsi, K. (2006). Distribution of WLAN received signal strength indication for
indoor location determination. 1st International Symposium on Wireless Pervasive
Computing, 2006. pp.6 (Jan. 2006)
King, T., Kopf, S., Haenselmann, T., Lubberger, C., Effelsberg, W. (2006). COMPASS: A
Probabilistic Indoor Positioning System Based on 802.11 and Digital Compasses.
University of Mannheim, D-68159 Mannheim, Germany, TR-2006-012
Krohn, A. Beigl, M. Hazas, M. Gellersen, H.-W, (2005). „Using fine-grained infrared
positioning to support the surface-based activities of mobile users“, Distributed
Computing Systems Workshops, 2005. 25th IEEE International Conference on,
Telecooperation Office, Karlsruhe Univ., Germany.
Li, B., Salter, J., Dempster, A., Rizos, C. (2006). Indoor Positioning Techniques Based on
Wireless LAN. AusWireless ’06, Sydney (March 2006)
Li, B., Wang, Y., Lee, H.K., Dempster, A., Rizos, C. (2005). Method for yielding a database
of location fingerprints in WLAN. Communications, IEE Proceedings- Volume 152,
Issue 5, pp.580 - 586 (Oct. 2005)
Llombart, M., Ciurana, M., Barcelo-Arroyo, F. (2008). On the scalability of a novel WLAN
positioning system based on time of arrival measurements. 5th Workshop on
Positioning, Navigation and Communication, 2008. WPNC 2008, pp. 15 – 21
(March 2008)
Masaki Maeda, Takefumi Ogawa, Takashi Machida, Kiyoshi Kiyokawa, Haruo Takemura,
(2003). „Indoor Localization and Navigation using IR Markers for Augmented
Reality“, HCI International 2003 Interactive demo
Michael R. McCarthy, Henk L. Muller, (2003). „RF Free Ultrasonic Positioning“, University
of Bristol, 7th International Symposium on Wearable Computers, October 2003.
Muller N. (2001). „Bluetooth Demystified“, McGraw-Hill, New York, 2001.
Nešković A., Nešković N., Paunović Dj., (2000). „Indoor Electric Field Level Prediction
Model Based on the Artificial Neural Networks“, IEEE Communications Letters,
vol. 4, No. 6, June 2000
Patil, A. (2002). „Performance Of Bluetooth Technologies And Their Applications To
Location Sensing“, Michigan State University, 2002.
Ramachandran, A., Jagannathan, S. (2007). Spatial Diversity in Signal Strength based WLAN
Location Determination Systems. 32nd IEEE Conference on Local Computer
Networks, 2007. pp. 10 - 17 (Oct. 2007)
Ramachandran, A., Jagannathan, S. (2007). Use of Frequency Diversity in Signal Strength
based WLAN Location Determination Systems 32nd IEEE Conference on Local
Computer Networks, 2007. pp.117 - 124 (Oct. 2007)
Roos, T., Myllymaki, P., Tirri, H. (2002). A statistical modeling approach to location
estimation. Mobile Computing, IEEE Transactions on, Volume 1, Issue 1, pp.59 –
69, (First Quarter 2002)
Roy Want, Andy Hopper, Veronica Falcao, and Jon Gibbons, (1992). „The Active Badge
location system“, ACM Transactions on Information Systems, 10(1):91–102, January
1992.
Sayrafian-Pour, K., Kaspar, D. (2005). Indoor positioning using spatial power spectrum.
IEEE PIMRC 2005. Volume: 4, pp. 2722- 2726 (Sept. 2005)
Shang Y., Wah W. B. (1996). Global Optimization for Neural Network Training, IEEE
Computer Society, pp. 45-56 (March 1996)
Thapa K., Case S., (2003). „An indoor positioning service for Bluetooth Ad Hoc networks“,
in: MICS 2003, Duluth, MN, USA.
Wang, H., Jia, F. (2007). A Hybrid Modeling for WLAN Positioning System. International
Conference on Wireless Communications, Networking and Mobile Computing,
2007. pp.2152 - 2155 (Sept. 2007)
Youssef, M., Agrawala, A. (2005). The Horus WLAN Location Determination System. Int.
Conf. on Mobile Systems, Applications And Services, pp.205–218 (2005)
Youssef, M., Agrawala, A. (2003). Small-scale compensation for WLAN location
determination systems. Wireless Communications and Networking, 2003. Volume
3, 20-20 pp.1974 – 1978 (March 2003)
Youssef, M.A., Agrawala, A., Shankar, A. U. (2003). WLAN location determination via
clustering and probability distributions. Proceedings of the First IEEE International
Conference on Pervasive Computing and Communications. pp.143 – 150 (March
2003)
Zhang, M., Zhang, S., Cao, J. (2008). Fusing Received Signal Strength from Multiple Access
Points for WLAN User Location Estimation. International Conference on Internet
Computing in Science and Engineering, 2008. pp.173 - 180 (Jan. 2008)
Location in Ad Hoc Networks 619
Location in Ad Hoc Networks

Israel Martin-Escalona, Marc Ciurana and Francisco Barcelo-Arroyo
UniversitatPolitecnicadeCatalunyaU(
Spain
1. Introduction
An adhoc 1 network is defined as a decentralised wireless network that is set up on-the-fly for
a specific purpose. These networks were proposed years ago for military use, with the
purpose of communicating devices in a highly constrained scenario. Under such a network,
devices join and leave the network dynamically; thus, it cannot be expected to have any
kind of network infrastructure. This wish for decentralised on-the-fly networks has
subsequently expanded to cover several fields besides the military. Today, there are several
mobile services requiring the self-organising capabilities that ad hoc networks offer.
Examples include packet tracking, online-gaming, and measuring systems, among others.
Ad hoc networks have obvious benefits for mobile services, but they also introduce new
issues that regular network protocols cannot cope with, including optimum routing,
network fragmentation, reduced calculation power, energy-constrained terminals, etc.
In ad hoc networks, positioning takes a significant role, mainly due to the on-the-fly
condition. In fact, several services require nodes to know the position of the customers in
order to perform their duty properly. Wireless sensor networks concentrate most of the
services that need positioning to perform their duty. Such networks constitute a subset of ad
hoc networks involving dense topologies operating in an ad hoc fashion, and they are
composed of small, energy and computation constrained terminals. In ad hoc networks, and
especially in wireless sensor networks, nodes are spread over a certain area without a
precise knowledge about the topology. In fact, this topology is variable. Accordingly, there
are several unknowns (e.g., node density and coverage, network’s energy map, the presence
of shadowed zones, nodes’ placement in the network coverage area) that are likely to
constrain the performance of ad hoc services. Knowledge of the terminals’ locations can
substantially improve the service performance.
Positioning is not only important for the service provisioning; it is also crucial in the ad hoc
protocol stack development. Due to the changes in the topology and the lack of
communication infrastructure, ad hoc protocols have to address several issues not present in
regular cellular networks. Routing is one of the best examples of the dependence of ad hoc
networks on positioning. Studies such as (Stojmenovic, 2002) demonstrate that only
position-based routing protocols are scalable, i.e., able to cope with a higher density of
1 "Ad hoc" is actually a Latin phrase that means "to this (thing, purpose, end, etc.)"
nodes in the network. The same seems to apply to other management and operation tasks in
ad hoc networks.
1.1 The Location problem in ad hoc networks

Nodes in an ad hoc network can be grouped into three categories according to the
positioning capabilities: beacon nodes, settled nodes, and unknown nodes. Beaconnodes , also
known as anchors or landmarks, are those able to computer thei position on their own, i.e.,
without using an ad hoc location algorithm. Accordingly, they implement at least one
location technique (e.g., GPS, map matching), which can be used as standalone. Beacon
nodes usually constitute the reference frame necessary to set up a location algorithm.
Unknown nodes are those nodes that do not know their position yet. When an ad hoc
network is set up, all nodes except the beacon nodes are unknown. Settled nodes are
unknown nodes that are able to compute their position from the information that they
exchange with beacon nodes and/or other settled nodes. The purpose of the ad hoc location
system is thus to position as many nodes as possible, turning them from unknown to settled
nodes (Bourkerche et al., 2007).
Location systems in ad hoc networks function in two steps: local positioning and positioning
algorithm. The former is responsible for computing the position of an unknown node from the
metrics gathered. The second step consists of the positioning algorithm, which indicates how
the position information is managed in order to maximise the number of nodes being settled.
1.2 Measuring the performance of location solutions

Performance of location solutions in ad hoc networks can be computed according to several
parameters. The main ones are presented below.
1.2.1 Accuracy
Ad hoc positioning requires good accuracy since most of the networks in ad hoc mode are
deployed in constrained scenarios, often indoors. In such environments, accuracy is
especially relevant, since a few meters of error in the position may cause the node to be
identified in another room, floor, or even building. Furthermore, nodes are expected to be
very close (e.g., in medical applications), and inaccurate positions could hinder operation
and maintenance tasks or even prevent location-based applications from performing their
duty. Thus, location algorithms for ad hoc networks must produce positions for settled
nodes of the highest possible accuracy.
1.2.2 Latency
The location solutions must be able to converge, i.e., to produce as many settled nodes as
possible in the shortest time. Ideally, the location solutions for ad hoc networks should turn
all the unknown nodes into settled nodes in a defined time. However, optimality in terms of
accuracy (and many other factors) may collide with latency, which means that the
convergence (and hence the latency) of the location solution depends on the accuracy
requested, among many other factors. Latency is also modulated by the mobility of the
nodes in the network. The faster the nodes move, the shorter the convergence period needs
to be. It must be noted that, in the case that convergence is not achieved, estimated positions
would not involve the actual location of nodes, and this lack of accuracy would be spread by
the network according to the location algorithm used. Hence, it would degrade the accuracy
of the location solution.
1.2.3 Form factor of terminals

Ad hoc devices tend to be small, especially in the case of wireless sensor networks. The
requirement of small form constrains the capabilities of these devices, which prevents
sophisticated (and usually more complex) algorithms from being used. This, in turn,
prevents the best QoS from being reached.
1.2.4 Energy-efficient design

Due to restrictions on their size, ad hoc devices tend to include very-limited batteries.
Accordingly, location solutions must avoid using complex algorithms or sensing multiple
metrics in order to compute the position, since these would limit the lifetime of nodes and
hence of the entire network.
1.2.5 Self-organising design

Ad hoc devices are likely to move. Algorithms developed for positioning in such networks
should account for the overhead generated by the changes in the topology of the ad hoc
network. Changes in the topology are produced by nodes moving around the network area
or being added to and removed from the ad hoc network.
1.2.6 Random nature of the ad hoc network

Ad hoc devices can be added and removed from the network during its life cycle. Depending
on the scenario, these changes in the topology of the network can be noticeably intense.
Location solutions in ad hoc networks should be insensitive to the structure of the network.
1.2.7 Scalability
The algorithm should minimise the impact of adding a new terminal to the network. This
means that the amount of resources consumed by the network due to the addition of the
new node should be as low as possible. Scalability does not necessarily involve the use of
simple algorithms. However, it should allow as many ad hoc devices as possible to be
positioned with the same amount of resources.
1.2.8 Node density

In actual deployments of ad hoc networks, devices are not likely to be homogeneously
distributed along the layout. The density of terminals is variable in space and time, and
hence location algorithms should not assume isotropic conditions.
1.2.9 Beacon percentage

Beacons tend to be fixed nodes often plugged to wired power sources, which make them
more durable. Beacons are set up by the network operator, and, consequently, they collide
with the ad hoc philosophy. Moreover, it is difficult to ensure that the percentage of beacons
visible for an unknown node remains uniform. Accordingly, location algorithms should be
as insensitive to beacon percentage and beacon placement as possible.
2. Location metrics
There are several metrics than can be used as input for location techniques. Those metrics
are usually known as observables, since they refer to what can be observed and
subsequently measured. Timestamps, angles, and signal strength are metrics commonly
used by location techniques to compute the position of the nodes. The former usually
involves timestamps for the sending and receiving moments associated with one or several
signals. Precision of timestamps directly depends on the clock present in the ad hoc devices.
These clocks are usually low-profiled, mainly due to the small form-factor and the cost of
devices. Consequently, this impacts the accuracy of the time-based observables and
ultimately the positions fixed by the location solution. Furthermore, accessing the hardware
clock is rather difficult in most of the current devices and technologies. The same does not
apply to signal strength, which is usually available in most of the ad hoc technologies.
Consequently, there are several location techniques that use this metric for positioning.
Furthermore, this metric provides accurate observables, even though it does not mean that
positions computed from these observables achieve the same degree of accuracy.
Finally, angular information is proposed for positioning in several solutions. This metric
consists of measuring the angle or direction of arrival (AoA / DoA) of signals coming from
several nodes to the target one, or vice versa. Fig. 1 illustrates this metric, where the red-
coloured angles (i.e., ˘ 1 and ˘ 2 ) stand for the error produced in the angle-of-arrival estimate.
Fig. 1. Positioning according to the angle/direction of arrival

According to Fig. 1, the angles of arrival can be computed as
 y i  y jk
 k   tan  1  , (1)
 x i  x jk
 
where (x i, y i) is the position (in two dimensions) of the node to be positioned and (x jk ) is
, y jk
the position (in two dimensions) of the landmark k.
The use of this metric involves using arrays of antennas in order to capture the angle in
which the signal is being received. Furthermore, the positioning error derived from the
angle-estimation error depends on the distance between the pair of entities involved in the
angle estimation. Consequently, this metric is rarely used; the hardware is costly, the error is
range-dependent, and this metric often involves the customisation of network equipment.
3. Location techniques
Ad hoc networks use a subset of the location techniques proposed for other cellular
technologies (e.g., UMTS, IEEE 802.11, etc.). These techniques can be classified into two
main groups: ranging-based and angle-based. The following sections describe the
techniques in detail.
3.1 Ranging based on signal strength

Ranging-based techniques are based on computing the distance between two nodes (i.e.,
ranging) and then computing the position of the unknown node by using a multilateration
algorithm. Range estimations can be computed from several metrics, but two are preferably
used: signal strength and timestamp. Techniques based on the former estimate the range
between two nodes according to the received and transmitted power. Radio path models
depend on the distance according to a certain power, known as path-loss slope. This means
that distance can be computed from transmitted and received signal strength, which is
information easily accessible in the network. According to general knowledge on radio
propagation, received power can be expressed as
Prx  Ptx  P1m  10 log d , (2)
where P rx and P txare, respectively, the received and transmitted power in dB(m), P m1 stands for
the losses at 1 meter from the transmitter location, d is the distance in meters between the
transmitter and receiver placements (i.e., the ranging), and ´ is the path-loss gradient (or slope).
Modelling the radio path losses, such as those in Equation (2), is a difficult task. Obstacles in
the propagation path affect the signal in several ways, namely, reflection, diffraction, and
absorption. The consequence is that signals reach the receiver following more than a single
path (a phenomenon known as multi-path), and, consequently, the received signal strength
suffers random variations. Accordingly, different radio path models are proposed
depending on the scenario in which the network is going to be deployed. However, the
propagation conditions are likely to change (even dramatically) with time as new obstacles
appear. Hence, such models would need to be recalibrated periodically (constantly in the
worst case). Indoors is one of the most constrained scenarios, and, consequently, most of the
location solutions based on signal strength ranging are proposed for such an environment.
An example of a radio propagation model used in location is proposed in (Seidel & Rapport,
1992), where the free space model was adapted to indoor environments by adding several
parameters, such as the number of floors in the path or the number of walls. However, it is
demonstrated not to be a satisfactory approach since the number of obstacles is not known a
priori. Other approaches tried to improve radio signal propagation models for indoors
(Wang et al., 2003; Lassabe et al., 2005), but accurate distance estimates are not yet available.
Despite all these issues, several proposals are available for signal ranging. One of the first
approaches was presented in (Bahl & Padmanabhan, 2000), where several models
specifically addressed to ranging-based location solutions were proposed and tested
experimentally. Due to the randomness of the received signal, poor results were obtained with
all models, if compared with other location techniques based on signal fingerprinting. Better
results are reported in (Kotanen et al., 2003), where the authors propose processing the signal-
strength observables prior to position computation. This previous stage aims to reduce the
noise of the measurements so that more accurate positions can be fixed. Furthermore, an
extended Kalman filter is used to compute the position, which minimises the variance of the
distance estimation. This solution provides accuracy figures of less than three metres, even
though worse results are expected under arbitrary propagation conditions.
3.2 Ranging based on time measurements

Considering the number of solutions currently proposed, time-based ranging seems to be a
more appealing technology than solutions based on signal strength. This is because,
compared to signal strength, time measurements tend to be more stable and less sensitive to
environmental conditions. Time-based ranging solutions can be classified into two groups
according to the number of signals/paths under consideration: time of arrival and time-
difference of arrival. The former is based on estimating the distance between two nodes. It is
achieved by marking transmission (t tx) and reception (t
rx) times and then applying
d  c· t tx
rx  t (3)
to compute the distance, where c stands for the propagation time (usually the speed of
light). Equation (3) can only be applied if timestamps are taken under the same time line,
i.e., when all nodes in the network are time-synchronised. However, this is not the normal
case. Thus, the 2-way time-of-arrival approach was proposed to overcome this issue. This
approach computes the propagation time under a round-trip-time approach, i.e., measuring
the time spent by the signal in travelling the forward (Node1 to Node2 ) and backward (Node
2 to Node1 ) paths. Fig. 2 illustrates this procedure, which is explained in detail in (Ciurana et
al., 2007). Since all timestamps are taken using the same clock (in Node 1), propagation time
(Tprop) can be computed as half the time meas ured for both paths (RTT), as long as
processing time (Tproc ) is negligible (or calibrated). The position of the target node (i.e., the
one to be located) can be computed by multilateration once enough measurements are
achieved (e.g., 3 or more for 2D positioning). The accuracy of time of arrival directly
depends on the precision of the time estimations. Accordingly, there are several works
addressing improvements in the accuracy of time of arrival observables.
Some examples can be found in (Ibraheem & Schoebel, 2007) and (Reddy & Chandra, 2007),
which present approaches improving the traditional correlation-based methods.
Time-difference of arrival consists of observing ranging differences, rather than just
observing distances. Therefore, the node that is going to be positioned measures the ranging
difference between its position and the position of a pair of landmarks (or even settled
nodes). Time difference of arrival is also known as hyperbolic multilateration, since it
superposes several hyperbolas in order to fix the node’s position. It provides better accuracy
than time of arrival and hence is preferred in cellular networks (3GPP, 2002; 3GPP, 2004).
However, time-difference solutions present the same issue as time of arrival: nodes have to
be time-synchronised. Furthermore, the 2-way approach cannot easily be applied to the
time-difference of arrival technique, and hence synchronisation error has to be estimated (by
means of specific measurement devices) or removed (taking the location measurements with
a common clock).
Fig. 2. RTT estimation through 2-way time of arrival approach
3.3 Triangulation based on angular measurements

This location method uses the direction or angles of arrival of several signals as the metric to
compute the position. Therefore, the location techniques based on angulation are also
known as Angle of Arrival (AoA) or Direction or Arrival (DoA). The position of the user can
be computed according to several approaches. One of the simplest consist of intersecting the
lines computed as
yi y jk  
 x i  x jktan  k , (4)
where ´ k is computed according to Equation (1). There are several (and more complex)
proposals based on numerical approach and closed forms to carry out positioning with
angulation, as presented in (Pages-Zamora et al., 2002).
4. Location algorithms
Local positioning has been widely addressed for cellular networks, and the main methods
and procedures remain valid for ad hoc networks. Hence, positioning algorithms draw the
greatest amount of interest from the research community for location in ad hoc networks.
4.1 Taxonomy of location algorithms in ad hoc networks
4.1.1 Centralised vs. Distributed

Centralised algorithms rely on a network entity (e.g., location server) that gathers the location
information from all the unknown nodes and then computes their position. The main
advantage is that global optimisation can be performed, as the information of all nodes is
available in the location server. Centralised systems are often used in cellular networks such
as public land mobile networks (PLMNs), but they collide with the random nature of ad hoc
networks. The location server has to be significantly more powerful than regular nodes.
Moreover, the data gathered by the server must be synchronised: all measurements must be
performed at specific times. If synchronisation is not assured, optimality cannot be reached,
and the system may be degraded. Additionally, topology in such systems can be displayed
as a tree, with the root at the location server. Therefore, nodes near the server quickly run
out their batteries, since they concentrate most of the location traffic coming from unknown
nodes; this reduces the ad hoc network lifetime.
In distributed location algorithms, some or all nodes are able to compute their position (and
the position of other nodes depending on the specific algorithm). Thus, distributed location
is more robust to node failures. Distributed algorithms also converge faster than centric
solutions after topology changes and are usually insensitive to data synchronisation
requirements, since only local or regional data are accounted for. There are several degrees
of distributed algorithms. The most common are the localised or pure distributed
algorithms, where all unknown nodes are able to compute their position once the necessary
local metrics are available.
4.1.2 Incremental vs. concurrent

Incremental positioning algorithms start with only a few beacon nodes. Then, in each step,
the position of a reduced amount of unknown nodes is computed using the information
provided by settled and beacon nodes. The positions of such nodes are used in subsequent
iterations to compute the location of other unknown nodes. The advantage of iterative
algorithms is their simplicity. However, they tend to propagate their positioning error, since
they use metrics obtained from settled nodes for subsequent local positioning. Furthermore,
the convergence of incremental location algorithms is not always guaranteed. In concurrent
algorithms, all nodes are able to compute their position normally using local information.
Accordingly, they are more complex than iterative algorithms, but they can avoid error
propagation and hence achieve better accuracy results.
4.1.3 One hop vs. multi-hop

Location involves exchanging information to measure the metrics used to compute the
node’s position. Onehopalgorithms use only information (i.e., metrics) local to the unknown
nodes (e.g., ranging to the node’s neighbours). On the other hand, multi-hop algorithms use
information of all nodes that can be reached from the node in a certain number of hops,
usually two. Muti-hop techniques allow more accurate positions to be computed and fewer
beacon nodes to be deployed in the system (Savvides et al., 2001). The main drawbacks are
the overhead generated by the multi-hop estimation and the subsequent use of additional
resources in the nodes to store such data.
4.1.4 Beacon-Based, Mobile Beacon-Based and Beacon-Less

Ad hoc location algorithms can be classified according to the presence of beacon nodes in
three categories: localisation with beacons, localisation with moving beacons, and
localisation without beacons (Sun et al., 2005). Localisation-with-algorithms
beacons are those
in which a percentage of nodes are fixed beacons, i.e., beacons that do not change their
location. The major challenge of algorithms that rely on beacons is to maximise the accuracy
and coverage while at the same time minimising the number of landmarks in the network.
Localisation with moving beacons algorithms are similar to algorithms based on beacons, but
here the beacons are no longer fixed and move through the network. A moving beacon is
perceived by unknown nodes as different beacons (i.e., one per message exchanged, from
different positions of the mobile beacon). Fewer landmarks are necessary, and more accurate
positions can be achieved since beacon density is perceived as higher than it actually is. The
main drawback with mobile beacon algorithms is that mobile beacons have to cover the
entire ad hoc network and ensure that unknown nodes see the mobile beacons with a
suitable frequency, which is often difficult to achieve. The last category, known as beacon-free
location, involves those algorithms in which no node is aware of its position (i.e., all nodes
are unknown). Thus, all nodes work together to compute their position using only their
local information. This kind of algorithm usually works with a local coordinate system,
which may require translation of the achieved positions into a global coordinate system so
that they can be used by a location-based service or protocol.
4.1.5 Range-free vs. range-based

In range-based algorithms, the local position is computed according to ranging measurements
(i.e., distance or angle estimates). Accordingly, they involve multilateration techniques,
which are usually hardware-demanding and therefore energy-consuming. Accordingly,
range-based algorithms are suitable for ad hoc networks with powerful terminals (e.g., in
technologies such as IEEE 802.11). On the other hand, technologies with more constrained
terminals, such as those present in wireless sensor networks, favour the use of range-free
algorithms for positioning. Range-free algorithms do not rely on ranging to compute the
position of unknown nodes; rather, they consist of simpler approaches based on proximity.
4.2 State of the art

The first location solutions proposed for ad hoc networks used centralised algorithms
similar to techniques proposed for mobile networks. In (Doherty et al., 2001), the authors
propose to manage the location as a convex-optimisation problem: a mobile server is
defined to gather all the location data (e.g., distances, angles of arrival, etc.) from unknown
nodes in order to compute their positions. The advantage is the simplicity and optimality of
the positions computed. However, it involves delivering a significant amount of data to a
location server that must be powerful enough to handle complex data structures. Moreover,
the cost of the algorithm proposed for this technique is cubic in the number of connections,
which seriously constrains the scalability of the approach.

In order to overcome those drawbacks, distributed algorithms are present in many
solutions. The centroid algorithm (Bulusu et al., 2000) is a one-hop pure-distributed
positioning algorithm in which few beacons are spread in the ad hoc network forming a
grid. Unknown nodes compute their position by estimating their range to the three closest
beacons and a trilateration algorithm. The main benefit is that it is insensitive to the node
density and does not add significant overhead in the network. The main drawback of this
method is a larger error in the positioning. Another interesting example of a one-hop
algorithm is proposed in the Lighthouse project (Römer, 2003), in which a single base station
sees all sensors in the network. This full coverage is achieved by means of a beam that
rotates at known speed. Stations are able to compute their position knowing the rotation
speed, the width of the beam, and the signal time-of-flight. Although the accuracy might not
be suitable for many applications in ad hoc networks (it provides a bias up to 14 metres), it
represents an improvement in scalability.
Recent proposals emphasise distributed behaviour. The Ad hoc Localisation System (AhLOS)
presented in (Savvides et al., 2001) is an example of this trend. AhLOS is a one-hop pure-
distributed algorithm that uses three trilateration algorithms: atomic, iterative, and
collaborative. Atomic trilateration involves a one-hop scenario, where the nodes have three
or more beacons in sight so that they can compute their positions directly. The main
drawback of this approach is that it relies on a high density of beacons. Iterative trilateration
relaxes this assumption, considering all the nodes that compute their position by means of
atomic trilateration (i.e., settled nodes) as new beacons. It allows fewer beacons at the cost of
less accuracy. Despite covering most of the situations, these two trilateration approaches are
not sufficient to position all the nodes in the ad hoc network, since unknown nodes with
only one neighbour cannot be positioned. The authors of AhLOS followed a collaborative
multilateration approach to overcome this situation, consisting of identifying those
unknown nodes that cannot be handled by atomic and iterative algorithms and creating
groups that collaborate in order to compute the position of those nodes. This may involve
solving large nonlinear systems depending on the size of the groups created.
The Ad hoc Positioning System APS) ( presented in (Niculescu & Nath, 2003) combines two
concepts: beacon-based positioning and ad hoc propagation (i.e., multi-hop). The algorithm
proposed in APS consists of four stages. Firstly, some beacon nodes are spread by the network.
Secondly, nodes with some landmarks in sight measure their distance to the beacon nodes in
terms of some metric, such as propagation time, number of hops, etc. Then, the information
gathered by nodes in the neighbourhood of landmarks is propagated (and updated) using a
proper algorithm. Finally, once a node has the ranging information of three or more beacons, it
computes its position using a multilateration approach. The algorithms for propagating the
ranging information in the ad hoc network are the main contribution of (Niculescu & Nath,
2003): DV-Hop, DV-distance, Euclidean, and DV-Coordinate . In the first, all nodes build ranging
tables containing the position coordinates and the distance in hops to the landmarks. These
data are flooded in the network in a controlled way, so that all nodes know how far they are
from landmarks. On the other hand, landmarks use such data to compute the average distance
of one hop, according to the hops and the distance between them. This is achieved by:
 j i
x i  
 x j2  y i  y 
j
2
ci  , (5)
 h
j i j
where ci indicates the hop distance calculated by the landmark i, (x,y,z) are the coordinates
of a landmark and hjstands for the amount of hops from landmark ito j .
This average (i.e., ci) is then flooded in the network using a singleton approach: once a node
receives an announcement packet from a landmark containing the average value computed
by such a landmark, it discards any other further announcement packet. The average
distance is then stored in the nodes, which then use it to turn distances in hops into real
ranges. Finally, nodes compute their position using a multilateration approach. The
advantage of this approach is that it is insensitive to the ranging error, since ranging is
based on hop counting. However, it introduces more overhead than other algorithms.
Moreover, accuracy is degraded in non-dense networks.
The DV-distance approach is similar to DV-Hop but exchanges real distances instead of
distances in hops. Multi-hop distances are computed as the sum of the distances between
nodes involved in the path. Thus, this approach becomes more sensitive to ranging error.
Normally, the denser the ad hoc network is, the better the accuracy. On the other hand, it
improves the consistency of the DV-Hop approach, working similarly in isotropic and non-
isotropic networks.
The Euclidean approach involves a multi-hop algorithm, which gathers ranging information
up to 2 hops. Thus, the algorithm generates quadrilaterals involving one unknown node,
two neighbours, and a landmark, and it infers the distance from the node to the landmark
using trigonometric formulation. The advantage of this approach is that the ranging error
can be estimated, stored, and subsequently flooded together with the distance; this allows
distance-weighted approaches to be used and hence more accurate positions to be achieved.
The main drawback is that a 2-hop approach involves additional traffic in the network (even
more than in DV-Hop and DV-Distance algorithms) as well as more resources needed in the
node to store the additional information, resulting in more quickly depleting node batteries
running.
The last approach presented in (Niculescu & Nath, 2003) is the so-called DV-Coordinate ,
which is similar to the solution proposed in (Capkun et al., 2001). DV-Coordinate is based
on each node computing the position of its neighbourhood according to a local coordinate-
system. Then in a second step, called the registration stage, nodes exchange information to
build the transformation matrices, which allow coordinates of local systems to be
transformed from one local system to another. A global transformation matrix is necessary
to achieve global coherence. DV-Coordinate performs almost the same as Euclidean.
However, this approach impacts the scalability of the location system, since it depends on
the square of the nodes in the network and involves sending two pieces of data instead of
just a distance. In (Niculescu & Nath, 2003/2), the authors extend DV-Coordinate and
presented the Local Positioning System LPS) ( , which uses ranging and angle of arrival
information to compute the position of unknown nodes, in a fashion similar to the DV-
Coordinate. However, the LPS reduces the overhead of the DV-Coordinate systems, thus
improving the scalability and updating only a reduced number of nodes each time, not the
whole network.
The Amorphous Localisation algorithm (Nagpal et al., 2003) is similar to APS. It consists of
computing the distance from nodes to beacons in terms of hops. However, the hop-distance
is calculated offline according to the node density expected in the network. Then, a
multilateration approach is followed to compute the position. The main drawback of this
algorithm is the offline stage, which seriously constrains the scalability of the algorithm in
dynamic ad hoc networks.
A one-hop range-based concurrent pure-distributed algorithm is proposed in (Fu et al.,
2006) for networks based on DSSS, such as those based on IEEE 802.11. This algorithm is
based on propagating the clock from node to node so that the nodes involved in the
positioning work in a synchronised fashion. Then the ranging to neighbour nodes is
estimated, and a multilateration algorithm is applied. The synchronisation is achieved by
means of the pseudo-noise code used in the DSSS, and, consequently, the time-resolution
achieved matches with the code duration. Accordingly, the algorithm only works if times-
of-flight are much longer than the code duration, which is expected to be the usual case.
The Approximation of the Point- -Triangulation
In Testis presented in (He et al., 2003) as
(APIT)
another example of a one-hop range-based location algorithm. It is based on generating as
many triangles as possible involving three beacons. Then, the APIT algorithm evaluates
whether the unknown node is inside each triangle. Finally, it overlaps all the triangles,
reducing the final positioning error. The authors evaluate the algorithm through simulation
and conclude that this approach outperforms the centroid algorithm. Furthermore, this
approach achieves accuracy figures similar to those obtained in the APS and Amorphous
Localisation algorithms but requiring a lower node density and introducing less overhead.
On the other hand, it requires beacons with a radio range longer than that of regular nodes.
All these approaches to the ad hoc location rely on active multilateration; i.e., positioning an
unknown node involves a certain amount of location traffic in order to estimate the
distances to landmarks. Active location constrains the scalability of location algorithms in ad
hoc networks, in which topology and mobility are inherent to the network definition. The
next sections introduce a passive algorithm for location in collaborative networks (e.g., ad
hoc, wireless sensor networks, etc.), which aims to boost the scalability on positioning
systems.
5. Passive positioning
Recent advances in indoor positioning have led to proposals that time-of-arrival (TOA)
techniques for locating users are preferable to other techniques, such as fingerprinting.
Time-of-arrival solutions achieve accuracy figures that are similar to those obtained by other
techniques, but they do not require additional assistance for setup and maintenance.
Conversely, time-of-arrival techniques need to calculate a client’s range from at least three
receivers at known positions in order to obtain a 2D position. In addition, all the signal
transmitters involved in the TOA positioning system must be synchronised. Two-way TOA
techniques, such as those presented in (Ciurana et al., 2007) and (Yang et al., 2008), cope
with this issue by computing the range from the client (i.e., unknown node) to the base
station (i.e., landmark) using a round-trip-time (RTT) procedure. Since only the client's clock
is used to calculate the range, synchronisation between base stations is no longer necessary.
The drawback is that more traffic is generated on the network, thus reducing the available
throughput.
A recent proposal on ad hoc location presented in (Martin-Escalona & Barcelo-Arroyo, 2008)
extends the capabilities of time-of-arrival location techniques, allowing unknown nodes in a
network to position themselves in a passive fashion, i.e., without injecting traffic into the
network. The following sections explain this technique, namedpassiveTDOA, in detail.
5.1 Description of assisted passive-TDOA algorithm

The passive-TDOA algorithm listens to the access medium for messages that can be used to
compute time-difference of arrival (TDOA) figures. The only assumption of the algorithm is
that the nodes operate in a collaborative network. Note that this is the case for most of the
wireless local area networks, especially those based on ad hoc protocols. In this text, one
more assumption is taken only for explanatory purposes: the messages used to compute
TDOAs are generated by unknown nodes running a 2-way TOA technique.
Fig. 3. Operation of the passive-TDOA algorithm
The performance of the positioning algorithm is described in Fig. 3, which shows a network
with three anchors or landmarks (i.e., Anchor 1 to Anchor 3 ) and two regular nodes with
positioning capabilities (i.e., Node 1 and Node 2 ). At a given time, Node 1 begins a TOA
positioning process to locate itself. Thus, Node 1 sends data message (Data 1 ) to Anchor 1 at t
1 ,
which replies with an acknowledgement Ack ( 1 ), which reaches Node 1 at t 2 . The
corresponding RTT is hence calculated as t 2 – t 1 . Note, however, that other nodes in the
network also listen to all these messages, since it is a diffusion network. Thus, Node 2 hears
the Data 1 message at t 3 and the reply to that message, i.e., Ack1 , at t 4 . Therefore, a TDOA
measurement is generated as t 4 – t3 . The same process is followed by Node 1 to range with
Anchor 2 and Anchor 3 . Based on the assumption that Node 2 is only covered at Anchor 1 and
Anchor 2 , Node2 is able to calculate two TDOAs: t 4 –t 3 and t 7 –t 6 . These two measurements are
enough to position Node 2 using a multilateration TDOA algorithm. Note that the TDOA
position calculated at Node 2 involves hearing just two access points, which makes it possible
for positioning to take place where TOA techniques would be ineffective. The only datum
needed by MS 2 in addition to the TDOA measurements is the position of Node 1 . This
information can be supplied by the unknown node once it becomes settled (e.g., by
broadcasting), or it can be estimated in the passive-TDOA node. Simulation analysis
demonstrates that, under line-of-sight (i.e., visibility between nodes), the positioning error
achieved by this algorithm is often below 1.4 times the error achieved by the 2-way TOA.
Better results are achieved if the technique is deployed under non-line-of-sight conditions,
providing figures similar to those achieved by the 2-way TOA technique (Martin-Escalona &
Barcelo-Arroyo, 2008). This behaviour is especially relevant for location since non-line-of-
sight is the usual condition for location-system operation; hence, the best performance is
desired in such scenarios.
5.2 Autonomous passive TDOA: TOA position estimated

One of the most constraining requirements of the passive-TDOA algorithm is the position of
the 2-way TOA node. Supplying this information in location procedures where position has
been requested by a third party should not involve additional changes in the 2-way TOA
algorithm and could be considered the final step for the passive-TDOA algorithm. However,
the same does not apply to services in which the user requests his or her own position.
Although the impact of supplying the 2-way TOA positions on the capacity of the network
is expected to be small, it does become necessary to define a protocol that guarantees the
supply of TOA positions once they have been computed, so that passive-TDOA nodes can
figure out their own locations. This protocol could involve some modifications in the 2-way
TOA algorithm, which is not a desirable fact. OMA SUPL (OMA, 2008) can be used for such
purposes, but security issues need to be addressed before its implementation (e.g., positions
should not identify users).
The algorithm initially proposed for passive-TDOA has been modified to cope with TOA
position supplying. Accordingly, two operational modes have been defined: assisted and
autonomous . The former consists of the algorithm as defined in the previous sections. The
autonomous operational mode allows positions of TOA and passive-TDOA nodes to be
jointly-computed in the passive-TDOA node, in a passive fashion.
There are several benefits to computing the TOA and passive-TDOA positions jointly. The
first one is that the passive-TDOA algorithm does not depend on supplying the TOA
position. Accordingly, any 2-way TOA algorithm could be used together with the passive-
TDOA algorithm with only slight changes. Furthermore, the passive-TDOA nodes will
compute their own position and the position of the 2-way TOA nodes, which gives way to
approaches for improving accuracy. The autonomous passive-TDOA algorithm becomes
especially interesting in scenarios in which an application needs to locate all the users in the
network. Nodes report their positions, as well as the positions of 2-way TOA nodes
estimated in the passive-TDOA nodes, and then the application can use the redundancy of
positions to improve the accuracy of the 2-way TOA nodes.
The Autonomous mode of the passive-TDOA algorithm is based on a usual feature of 2-way
TOA algorithms: the redundancy on RTTs. These algorithms tend to measure several RTTs
involving the same landmark in order to reduce errors caused by the measurement system
and radio channel, hence improving the accuracy. Autonomous mode uses two consecutive
RTTs on the same landmark to estimate two TDOAs (as defined in the case of the normal
operational mode), as well as the RTT being measured at the TOA node. As expected, the
RTT estimate in the passive-TDOA node will be noisier than the ones made in the TOA
node, but it is expected to be accurate enough to allow the passive-TDOA algorithm to
compute its own position.
Fig. 4 shows the procedure that constitutes the autonomous operational mode of the passive-
TDOA algorithm. The explanation is based on the scenario proposed in Fig. 3, but reduced
to one 2-way TOA node (i.e., Node 1 ), a landmark (i.e., Landmark), and a passive-TDOA node
(i.e., Node 2 ). As explained for the case of the assistedoperational mode, Node 1 starts a ping-
fashion procedure to compute the range between the landmark and itself. As a result, the
RTT 1 (i.e., t 2 –t1 ) is measured. Consequently, TDOA 1 is deducted from the ping procedure as
t4 – t 3 . Until this point, the procedure is exactly the same as that presented in the case of
assisted mode.
Fig. 4. Flow diagram for the autonomous

operational mode of passive-TDOA algorithm
Then, it is assumed that Node 1 starts a new 2-way TOA procedure involving the same
entities (i.e., Landmarkand Node 2 ) after a predefined time ( ), which is known by all nodes in
the network. This new procedure provides new estimates for Node 1 and Node 2 , i.e., RTT 2 and
TDOA 2 , respectively. Furthermore, autonomous mode benefits from this redundancy in the
measurements by using it to estimate the ranging between Landmark and Node 1 in Node 2 .
This can be done by simply measuring a new time-difference in Node 2 : TDOA’ . This time-
difference corresponds to the difference between the arrival time to Node 2 of the first TOA
response and the second TOArequest messages, subtracting the time elapsed between the two
7 – t
ping processes (i.e., t 4 – in Fig. 4), which is assumed to be known by all nodes in the
network. This information, together with the TDOA 1 and TDOA 2 measurements, allows the
network to deduce the ranging information concerning Node 2 and the Landmark
. The
formulation starts from
, k  R i, k ,
T i, j  R i, j R j (6)
which computes the TDOA from the ranging information. R in Equation (6) computes the
distance between two nodes, T stands for the distance-difference, and subscripts i, j
, kstand
for the TOA node, the landmark, and the passive-TDOA node, respectively (as in the rest of
the document). According to Equation (6) and the scenario presented in Fig. 4, two TDOAs
are computed as
1 1 1 1
T i, j  R i, j  R j
, k  R i, k (7)
and
2 2 2 2
T i, j  R i, j  R j
, k  R i, k , (8)
where the superscript indicates the ping procedure involved in the measurement. These two
TDOAs (in distance) are then averaged (under the assumption of providing the same QoS)
as
1
2
 2 2
T i, j  T i, j  T i, j .  (9)
According to its definition, TDOA’ is computed as
1 2 1
TDOA ' T j
, i  R i, j  R i, k  R j
,k . (10)
Under the assumption of noiseless measurements, TOA ranging can be estimated in the
passive-TDOA node as
1
R i, j  T i, j T j
2

,i .  (11)
Finally, once R<i,j> is estimated, the same algorithm as used in the assisted mode is used to
compute the position of the passive-TDOA node.
Simulation results indicate that estimating the ranging of the 2-way TOA node results in less
accurate positions, as expected. However, under non-line-of-sight conditions, which are the
usual case, passive-TDOA provides positions with only 20% more error than the 2-way
TOA, with the benefit of no traffic injection. This is especially relevant for group location,
i.e., those applications that involve more than a single location process.
5.3 Applications of the passive-TDOA algorithm

The passive-TDOA algorithm has multiple applications in the field of location. The main
one has been discussed above and consists of allowing an unknown node to be positioned
without injecting traffic into the network. Therefore, the load due to positioning is reduced,
and the network throughput remains available for other services. This feature is essential for
location algorithms since it improves scalability, which is especially essential for the location
platform in the ad hoc environment.
Another application of passive TDOA is the capability of the algorithm to position unknown
nodes in environments where TOA techniques cannot. For instance, Fig. 3 shows how
passive-TDOA is able to position a customer who only has two access points in sight. Under
the same conditions, the TOA technique is not able to provide a location, since this
technique requires at least three transmitters (even more depending on the algorithm) to
perform a 2D trilateration. The passive-TDOA algorithm would be able to go further in
positioning under constrained scenarios. In fact, this technique would be able to compute
the position of a station with just one access point in sight, whenever enough settled nodes
are in sight and their positions are known. This makes the passive-TDOA algorithm a very
interesting solution for positioning under extreme conditions (e.g., scenarios in which there
is interference), eventually mitigating the impact of some access points being down (e.g.,
due to maintenance, fire damage, etc.). Since passive-TDOA can work with fewer
landmarks, it helps the system continue to offer location-based services in those
circumstances where TOA is not able to provide some positions.
Passive-TDOA can also be used to improve the accuracy of TOA positions. The passive-
TDOA node is able to estimate its position and the positions of other TOA nodes involved as
long as enough measurements are available. Subsequently, these positions can be coupled to
reduce the noise and improve the final accuracy. Furthermore, this operational mode can be
used to locate unknown nodes with no location capabilities at all, as explained in more
detail previously in this document.
All these applications of the passive-TDOA algorithm give rise to dramatic improvements in
the scalability of the system, since more customers can be located while only a few TOA
positioning processes are running. Note, however, that all the applications of the passive-
TOA algorithm depend on their expected accuracy, since a large error in the positions
computed by passive-TDOA will make these positions useless, and, therefore, system
scalability will not increase at all. This work analyses the accuracy expected from passive-
TDOA under several conditions and compares it with the positioning error achieved by a
regular 2-way TOA algorithm.
It must be noted that, even though the algorithms are addressed to ad hoc networks, they
can be implemented in other networks based on infrastructure, such as those operating
under the standard IEEE 802.11. In these networks, the anchors would be the access points,
and the terminal nodes would be the 802.11 clients. This demonstrates the capabilities of the
algorithm presented and the wide range of applications for which it can be used.
6. Conclusion
This chapter presents the positioning problem in the ad hoc context. According to the
current literature, ad hoc algorithms are predominantly focused on this topic, since location
techniques used for other cellular technologies remain valid in the ad hoc environment. The
main algorithms proposed for ad hoc positioning are presented, giving special attention to
the passive-TDOA. This algorithm is proposed to improve the scalability of 2-way TOA
solutions and while at the same time providing good accuracy figures. Two operational
modes are explained in detail, and the main applications for this algorithm are discussed.
7. References
Stojmenovic, I. (2002). Position-based routing in ad hoc networks, IEEE Communications
Magazine , vol. 40, July 2002, pp. 128-134, ISSN: 0163-6804.
Bourkerche, A.; Oliveira, H.; Nakamura, E. & Loureiro, A. (2007). Localization Systems for
Wireless Sensor Networks, IEEE Wireless Communications , Vol. 14, December 2007,
pp. 6-12, ISSN: 1536-1284.
Seidel, S.Y. & Rapport, T.S. (1992). 914 MHz path loss prediction Model for Indoor Wireless
Communications in Multi-floored buildings. IEEETrans.onAntennas&Propagation ,
vol. 40, no. 2, February 1992, pp. 207-217, ISSN: 0018-926X.
Wang, Y.; Jia, X. & Lee, H.K. (2003). An Indoors Wireless Positioning System Based on
Wireless Local Area Network Infrastructure, Proceedings of the thInternational
6
Symposium on Satellite Navigation , pp. 1-13, University of New South Wales,
Melbourne (Australia), July 2003.
Lassabe, F.; Canalda, P.; Chatonnay, P. & Spies, F. (2005). A Friis-based calibrated model for
WiFi terminals positioning, Proceedings th ofIEEE
6 International Symposium on a
World of Wireless Mobile and ltimedia
Mu Networks WOWMOM), ( pp. 382-387, ISBN:
0-7695-2342-0, Giardini Naxos, Messina (Italy), June 2005.
Bahl, P. & Padmanabhan, V. (2000). Radar: An In-Building RF-based User Location and
Tracking System, Proceedingsofthe th IEEEConference
91 onComputerCommunications
(INFOCOM) , vol. 2, pp. 775-784, ISBN: 0-7803-5880-5, Tel Aviv (Israel), March 2000.
Kotanen, A.; Hannikainen, M.; Leppakoski, H. & Hamalainen, T.D. (2003). Positioning with
IEEE 802.11b wireless LAN. Proceedings of the thIEEE 41 International Sympossium on
Personal, Indoor and Mobile Radio Communi cations PIMRC) ( , vol. 3, pp. 2218-2222,
ISBN: 0-7803-7822-9, Beijing (China), September 2003.
Ciurana, M.; Barcelo-Arroyo F. & Izquierdo, F. (2007). A ranging system with IEEE 802.11
data frames, Proceedings of the IEEE Radio
st 1 and Wireless Symposium , pp. 133-136,
ISBN: 1-4244-0445-2, Long Beach, California (USA), January 2007.
Ibraheem, A. & Schoebel, J. (2007). Time of Arrival Prediction for WLAN Systems Using Prony
Algorithm, Proceedingsofth Workshop
the 4 onPositioning,NavigationandCommunicatio
WPNC)
( , pp. 29-32, ISBN: 1-4244-0871-7, Hannover (Germany), March 2007.
Reddy, H. & Chandra, G. (2007). An improved Time-of-Arrival estimation for WLAN-based
local positioning, Proceedings of the ndInternational
2 Conference on Communicati
Systems software and middleware COMSWARE) ( , pp. 1-5, ISBN: 1-4244-0613-7,
Bangalore (India), January 2007.
3GPP. (2004). Functional stage 2 description of LCS; Technical Specification Group Services
and System Aspects, 3GPPTS23.71 , version 6.8.0, June 2004.
3GPP. (2002). Location Services (LCS); Functional description; Stage 2, 3GPP TS 03.71,
version 8.7.0, September 2002.
Pages-Zamora, A., Vidal, J. & Brooks, D.H. (2002). Closed-form solution for positioning based
on angle of arrival measurements. Proceedings of the IEEE International
th 31 Symposium
on Personal, Indoor and Mobile Radio Communications , vol. 4,PIMRC)
( pp. 1522-1526,
ISBN: 0-7803-7589-0, Pavilhão Atlântico, Lisbon (Portugal), September 2002.
Savvides, A.; Han, C.C. & Srivastava, M.B. (2001). Dynamic Fine-Grained Localization in
Ad-Hoc Wireless Sensor Networks. Proceedings of the thACM7 International
Conference on Mobile Computing and Networking , pp. 166-179, ISBN: 1-58113-422-3,
Rome (Italy), July 2001.
Sun, G.; Chen, J.; Guo, W. & Ray-Liu, K.J. (2005). Signal processing techniques in network-
aided positioning: a survey of state-of-the-art positioning designs, IEEE Signal
ProcessingMagazine , vol. 22, no. 4, July 2005, pp. 12-23, ISSN: 1053-5888.
Doherty, L.; Ghaoui, L.E. & Pister, K.S.J. (2001). Convex Positioning Estimation in Wireless
Sensor Networks.Proceedingsofthe IEEE02 ConferenceonComputerCommunications
th
(INFOCOM) . vol. 3, pp. 1655-1663, ISBN: 0-7803-7016-3, Anchorage, AK (USA),

April 2001.
Bulusu, N.; Heidemann, J. & Estrin, D. (2000). GPS-less low cost outdoor localization for
very small devices, IEEE Personal Communications, vol. 7, October 2000, pp. 28-34,
ISSN: 1070-9916.
Römer, K. (2003). The Lighthouse Location System for Smart Dust, Proceedings of the st 1
ACM/USENIXinternationalconferenceonMobilesystems,applications pp. ands
15-30, San Francisco, California (USA), May 2003.
Niculescu, D. & Nath, B. (2003). DV based positioning in ad hoc networks. SpringerJournalof
Telecommunication Systems , vol. 22, no. 1-4, January 2003, pp. 267-280, ISSN: 1018-
4864.
Capkun, S.; Hamdi, M. & Hubaux, J. (2001). GPS-free positioning in mobile ad-hoc
networks. Proceedings of the Hawaii International Conference on Systems
HICSS)
( , pp. 1-10, ISBN: 0-7695-0981-9, Maui, Hawaii (USA), January 2001.
Niculescu, D. & Nath, B. (2003). Localized Positioning in Ad Hoc Networks. Elsevier Ad Hoc
Networks , vol. 1, no. 2-3, September 2003, pp. 247-259.
Nagpal, R.; Shrobe, H. & Bachrach, J. (2003). Organizing a Global Coordinate System from
Local Information on an Ad Hoc Sensor Network. International Workshop on
Information Processing in Sensor Networks Published . as Lecture Notes in
IPSN)
(
Computer Science, LNCS 2634, pp. 333-348. ISSN 0302-9743.
Fu, Y.; Liu, H.; Qin, J. & Xing, T. (2006). The localization of wireless sensor networks nodes
based on DSSS. Proceedings of the IEEE Internationa l Conference on Electro/Information
Technology. pp. 465-469. ISBN: 0-7803-9592-1, East Lansing, MI (USA), May 2006.
He, T.; Huang, C.; Blum, B.; Stankovic, J. & Abdelzaher, T. (2003). Range-Free Location
Schemes for Large Scale Sensor Networks, Proceedings of the th 9 annual internation
conference on Mobile computing and networking, MOBICOM) ( pp. 81-95, ISBN: 1-
58113-753-2, San Diego, CA (USA), September 2003.
Yang, S; Kang, D.; Namgoong, Y.; Choi, S. & Shin, Y. (2008). A simple asynchronous UWB
position location algorithm based on single round-trip transmission. IEICE
Transactions on Fundamentals of Electronics, Communications and Comput
vol. E91-A, no. 1, January 2008, pp. 430-432, ISSN:0916-8508.
Martin-Escalona, I. & Barcelo-Arroyo, F. (2008). A New Time-Based Algorithm for
Positioning Mobile Terminals in Wireless Networks, EURASIP Journal on Advances
in Signal Processing , Hindawi Publishing Corporation, vol. 2008, pp. 1-10. ISSN:
1687-6172.
OMA (2008), Secure User Plane Location Architecture (SUPL), Open Mobile Alliance,
[Online], http://www.openmobilealliance.org.
Location Tracking Schemes for Broadband Wireless Networks 639
Location Tracking Schemes for

Broadband Wireless Networks
Po-Hsuan Tseng and Kai-Ten Feng
Department of Communication Engineering, National Chiao Tung University
Taiwan, R.O.C.
1. Introduction
In order to enable the delivery of last mile wireless broadband access, the IEEE 802.16-2004
standard (IEEE Std 802.16-2004, 2004) for the wireless metropolitan area networks (WMAN)
is designed to fulfill various demands for higher capacity, higher data rate, and advanced
multimedia services. Furthermore, the IEEE 802.16e standard (IEEE Std 802.16e-2005, 2006)
enhances the original IEEE 802.16-2004 specification by addressing the mobility issues for
the mobile stations (MSs). Recently, the IEEE 802.16-2009 standard (IEEE Std 802.16-2009,
2009) has been specified as an integrated version of the IEEE 802.16 specification by the IEEE
802.16 maintenance task group. The IEEE 802.16-2009 standard is as known as the revision of
IEEE 802.16-2004 and consolidates material from IEEE 802.16e-2005, IEEE 802.16-2004/Cor1-
2005, IEEE 802.16f-2005, and IEEE 802.16g-2007. In order to fulfill the requirement of the E911
phase II requirement advanced by Federal Communications Commission, the location-based
services (LBSs) (Perusco & Michael, 2007) are considered one of the key functions of the IEEE
802.16-2009 standard. Moreover, for fulfilling the resource management purpose, location
update is also essential to other numerous functions such as the paging processes.
Based on the IEEE 802.16 standard, it is required to provide satisfactory location estimation
performance under a wide-range of MS’s moving speeds. Location tracking is designed as
one of the options to provide feasible estimation performance in order to trace the MS’s mov-
ing behaviors. However, there are several issues required to be considered before enabling
the location tracking scheme within IEEE 802.16 system. It is noted that timing informa-
tion, i.e., the time-difference-of-arrival (TDOA) measurements, from at least four base stations
(BSs) is required to perform a two-dimensional location estimation and tracking for an MS.
With the stringent synchronization requirement for the IEEE 802.16 OFDMA-based system,
the frequent TDOA measurements with other neighbor (a.k.a. non-serving) BSs can be time-
consuming and impractical processes for location estimation and tracking. It is a waste of
the bandwidth to scan for the neighbor BSs frequently, especially under broadband wireless
communication.
In this book chapter, two location tracking schemes are proposed to alleviate the problem that
requires frequent connections between the MS and the neighbor BSs. The kinematics-assisted
location tracking (KLT) scheme adopts the kinematic relationship to estimate the MS’s loca-
tion at the time instant with unavailable neighbor BSs; while the geometry-assisted location
tracking (GLT) algorithm utilizes the geometric constraints for the prediction of MS’s posi-
tion. The two schemes are proposed to interpolate the location of an MS between two direct
location estimations from the MS to the neighbor BSs. It will be shown in the simulations that
both proposed schemes can provide feasible performance with significantly reduced commu-
nication overhead.
2. TDOA Measurements of IEEE 802.16 Network

To illustrate for the TDOA measurement scheme, the procedures related to the basics of IEEE
802.16 network operations are introduced first. The IEEE 802.16 adopts OFDMA based tech-
nique, which implies the MS and the serving BS should synchronize in both time and fre-
quency domain in order to receive the data correctly. Providing that an MS intends to join an
IEEE 802.16 network, it conducts initial ranging with the serving BSs to obtain the synchro-
nization parameter. In order to indicate which time slots for the MS to receive and transmit
data, the downlink map (DL-MAP) and the uplink map (UL-MAP) are designed respectively.
First the MS should listen to the BS’s broadcast message to capture the ranging opportunity
in the UL-MAP. The MS conducts contention based initial ranging to obtain related parameter
and then is granted to join the network. After the MS establishes the link with serving BS,
the MS performs periodic ranging in order to maintain the synchronization property while
the MS may move to different place or the channel condition may vary at different time. It
is noted that the periodic ranging is one of the routine process which is not considered as an
overhead in this work. The ranging scheme is conducted by the MS sending assigned CDMA
codes to serving BS to measure the distance related parameter, e.g. timing advance value. It
is noticed that the timing advance value is the round trip time between the MS and the serv-
ing BS. The timing advance information is utilized to reserve the proper timing for the uplink
transmission. By performing the ranging scheme, the serving BS measures the timing adjust-
ment according to previously recorded timing advance information to update the distance
between MS and serving BS. In order to support for the MS’s mobility, the scanning scheme
is specified for the MS to perform ranging with neighbor BSs. With the negotiation between
serving BS and neighbor BSs, the scanning scheme can directly obtain ranging opportunity
without contention. However, the MS is unavailable to serving BS while it performs scanning
scheme with the neighbor BSs.
In the IEEE 802.16-2009 standard, two types of TDOA are specified as the downlink TDOA
(D-TDOA) and the uplink TDOA (U-TDOA) where the measurements are performed at the
MS and the BS respectively. These two schemes are based on ranging and scanning scheme
to obtain time difference between the serving BS and the neighbor BSs. Due to the superior
timing resolution obtained from the U-TDOA measurement at the BS (i.e., around 25 to 50
nanoseconds measured at the BS compared to microseconds at the MS), the U-TDOA mea-
surement is chosen in this paper for achieving better location estimation accuracy. Moreover,
as described in the standard, the general U-TDOA method is adopted while the frequency
reuse factor is not equal to one which is considered as a more general case. Several assump-
tions are specified as follows: (a) both the serving and neighbor BSs are operating with the
same frame size; (b) the frames at both the serving and the neighbor BSs are synchronized;
and (c) the MS can communicate with both the serving and the neighbor BSs.
Fig. 1 illustrates the timing diagram of the general U-TDOA measurement in IEEE 802.16-
2009. The MS ranges sequentially with the serving BS and neighbor BSs. It is noted that the
second frame of the MS is operating on the same frequency with the serving BS and the third is
the same with the neighbor BS. The serving BS (i.e., the 1st BS) and the neighbor BS (i.e., the 2nd
BS) measure the timing adjustment t_adj1 and t_adj2 respectively, and the neighbor BS reports
t_adj2 to the serving BS. The timing advance value t_adv remains the same since the MS does
OFDMA Frames
Granted Slot #2
Neighbor
BS
UL Burst #2
t 1 = ( t_adj1 + t_adv) /2 received
Timing Adjustment t_adj 2
Granted Slot #1
Serving
BS
UL Burst #1
t 2 = ( t_adj2 + t_adv)/2
received
Timing Adjustment t_adj 1 t 21 = ( t_adj2 - t_adj 1 )/2
Delayed Granted Slot #1 Delayed Granted Slot #2

MS
UL Burst #1 UL Burst #2
Propagation transimitted transimitted
Delay
Timing Advance t_adv Timing Advance t_adv
Fig. 1. Timing Diagram of the general U-TDOA measurement proposed in the IEEE 802.16-
2009 standard.
not make any timing adjustments while conducting ranging with both the serving and the
neighbor BSs. Therefore, serving BS calculates the time difference t21 = (t_adj2 − t_adj1 )/2
and the difference of the MS’s distance to the serving BS and neighbor BS is obtained by
multiplying this difference by the speed of light.
Fig. 2 illustrates the message exchange sequences for the general U-TDOA measurement.
The serving BS requests neighbor BSs to assign a dedicated ranging opportunity for the MS.
The MS will first conduct ranging with the serving BS in order to perform the U-TDOA mea-
surement. Through the ranging response (RNG-RSP) message, the serving BS will assign a
Rendezvous time, a CDMA code, and a Tx opportunity offset for the MS. It is noted that the
Rendezvous time specifies the frame in which the BS transmits an UL-MAP containing the
definition of the dedicated ranging region. As the Rendezvous time is expired, the MS will
transmit the allocated CDMA code within the regular ranging region. The time-of-arrival
(TOA) measurement t1,k performed at the serving BS (BS1 ) is obtained as an average value of
both the timing adjustment t_adj1,k from the measurement and the timing advance t_advk ac-
quired from the periodic ranging, i.e., t1,k = (t_adj1,k + t_advk )/2 where the subscript denotes
the kth time step.
Moreover, the dedicated ranging process is repeated through the neighbor BS (e.g. BS2 ) by
sending the mobile scanning response (MOB_SCN-RSP) message. The neighbor BS measures
t_adj2,k and reports the value to the serving BS. The TOA measurement for BS2 can therefore be
acquired as t2,k = (t_adj2,k + t_advk )/2. As a result, the U-TDOA measurement calculated at
the serving BS is obtained as t21,k = (t_adj2,k − t_adj1,k )/2. Similar process can be performed
to obtain the timing information from the other neighbor BSs. Nevertheless, with the TDOA
measurements, at least four BSs should be involved to perform a 2-D location estimation.
There is a significant overhead to perform location tracking scheme as depicted in Fig. 2. In
the following section, the proposed location tracking schemes are proposed to alleviate the
overhead of frequent ranging with neighbor BSs.
MS BS#1 BS#2 BS#3

(Serving) (Neighbor) (Neighbor)
Request non-serving BS to
assign a dedicated ranging
Ranging Response opportunity for the MS
(RNG-RSP)
(status = continue, Rendezvous time,
CDMA code, Tx opportunity Offset)
Rendezvous time
Dedicated CDMA code
Measure t_adj1,k
t = ( t_adj + t_adv )/2
1,k 1,k k
Scanning Interval Allocation Response

(MOB_SCN-RSP)
(status = continue, Rendezvous time,
CDMA code, Tx opportunity Offset)
Rendezvous time
Dedicated CDMA code
Measure t_adj2,k
t = ( t_adj + t_adv )/2
2,k 2,k k
RNG-RSP
Return t_adj 2,k
t = ( t_adj - t_adj )/2

21,k 2,k 1,k
/ /
/ /
/ /
Fig. 2. The message exchange sequences of the general U-TDOA measurement method pro-
posed in the IEEE 802.16-2009 standard.
3. Proposed Location Tracking Schemes

3.1 Mathematical Modeling of Signal Measurements
In this subsection, the mathematical models for both the TOA and the TDOA measurements
are presented to facilitate the location estimation of the two-dimensional coordinates for the
MS. The measured relative distance (ri,k ) between the MS and the ith BS (obtained at the kth
time step) can be represented as
ri,k = c · ti,k = ζ i,k + mi,k + ni,k + si,k (1)
where ti,k denotes the TOA measurement obtained from the ith BS at the kth time step, and
c is the speed of light. ri,k is contaminated with the TOA measurement noise mi,k due to the
imprecision of the measuring device. The non-line-of-sight (NLOS) error ni,k is assumed to be
existent in the considered environments, which is inherently a parameter with positive value.
si,k denotes the unknown asynchronous clock time offset between the MS and the BS. The
noiseless relative distance ζ i,k between the MS and the ith BS can be obtained as
1
ζ i,k = [( xk◦ − xi,k )2 + (y◦k − yi,k )2 ] 2 (2)
where x◦k = [ xk◦ y◦k ] T represents the MS’s true position at the kth time step, and xi,k = [ xi,k yi,k ] T
is the location of the ith BS. On the other hand, the relative distance rij,k from the TDOA
measurement tij,k can be obtained by computing the time difference between the MS w.r.t. the
ith and the jth BSs as
rij,k = (ri,k − r j,k ) = c · (ti,k − t j,k ) = c · tij,k
= (ζ i,k − ζ j,k ) + (mi,k − m j,k ) + (ni,k − n j,k ) (3)
It is noted that in IEEE 802.16-2009 standard each BS equips a GPS in order to achieve time
synchronization. Therefore, the relationship si,k − s j,k = 0 is true since the frames at the serv-
ing and the neighbor BS are synchronized. As depicted in the previous section for the general
U-TDOA method, the TOA measurement (in (1)) is acquired for the purposed of obtaining the
TDOA measurement (in (3)).
3.2 Location Estimation and Tracking Algorithms

The main concept for the proposed schemes is to maintain the accuracy for location estimation
with a reduced number of dedicated ranging (i.e., the general U-TDOA measurement in Figs.
1 and 2) between the MS and the BSs. The increased number of dedicated ranging with the
neighbor BSs can result in unsatisfactory communication performance between the MS and its
serving BS, e.g. with degraded scheduling performance for realtime applications. It is noted
that the dedicated ranging state indicates the state for the MS to conduct the general U-TDOA
method with the serving and the neighbor BSs as shown in Fig. 2. Without communications
between the MS and the BSs, on the other hand, the non-dedicated ranging state (i.e., the gen-
eral U-TDOA measurement is not available) is defined as the state in which the MS’s location
information is estimated and predicted by the proposed KLT and GLT schemes, which will be
explained in the next two subsections.
KLT/GLT Scheme with T I = T E= 1 sec
1 sec
TI / TE
KLT/GLT Scheme with T E= 4 sec, T I = 1 sec
TI
TE
t=k t = k+1 t = k+2 t = k+3 t = k+4
Dedicated Ranging State

Non-Dedicated Ranging State
Fig. 3. The timing diagram for the relationship between the dedicated and non-dedicated
ranging states.
Fig. 3 illustrates the relationship between the dedicated and the non-dedicated ranging
states. The location estimation period (TE ) is defined as the time duration between two ded-
icated ranging states. On the other hand, the location information period (TI ) is designed as
TI = TE /m with m ≥ 1, which represents the time interval either between two non-dedicated
ranging states or between a dedicated and a non-dedicated states. In other words, TI is de-
fined as the interleaved period where the MS’s location information becomes available, either
obtained from the general U-TDOA method of the proposed KLT/GLT scheme. In other word,
the U-TDOA method is utilized at the dedicated ranging state; while the KLT/GLT scheme
interpolates the position information at the non-dedicated ranging state. Moreover, the sam-
pling time ∆t is denoted as the time interval between the kth and the (k − 1)th time steps as
in (1) to (3) It is selected the same as the location information period, i.e. ∆t = TI .
x1,k x2,k x3,k x4,k ...

t21,k t31,k t41,k ... 2-step LS Kalman Filter
(TDOA-Based)
Yes ( xk , yk )
Dedicated
Ranging
No
KLT / GLT
x1,k t1,k (for GLT) Scheme
Fig. 4. The schematic diagram of the proposed KLT and GLT algorithms.
Fig. 4 illustrates the schematic diagram of the proposed KLT and GLT algorithms. Either the
dedicated or the non-dedicated ranging can happen for obtaining the MS’s estimated position
x̂k = [ x̂k ŷk ] T . For the dedicated ranging case, the cascaded location tracking (CLT) scheme
as proposed in (Chen & Feng, 2005) is exploited. The CLT algorithm is cascaded by two
functional components, i.e. the two-step least square (LS) method for location estimation and
the Kalman filtering technique for location tracking. The two-step LS method obtains the
initial location estimation x̂kLS = [ x̂kLS ŷkLS ] T from the TDOA measurement input tij,k (in (3))
within two computing iterations. Furthermore, the Kalman filter is utilized to smooth out and
trace the estimation errors and finally acquires the MS’s estimated location x̂k = [ x̂k ŷk ] T . On
the other hand, since the TDOA measurements are not available during the non-dedicated
ranging state, two schemes are proposed to substitute the functionality of the two-step LS
method as follows.
3.2.1 Kinematics-Assisted Location Tracking (KLT) Scheme

With the unavailability of the TDOA measurements during the non-dedicated state, the KLT
scheme is utilized to adopt the predicted information from the output of the Kalman filter.
Considering a three-states linear model, the MS’s position, velocity, and acceleration can be
estimated via the Kalman filter as ẑk = [x̂k v̂k âk ] T , where x̂k = [ x̂k ŷk ] T , v̂k = [v̂ x,k v̂y,k ] T , and
âk = [ â x,k ây,k ] T . Assuming that the state vector ẑk is available either via the dedicated or
non-dedicated ranging, the next non-dedicated states ẑk+1 at the (k + 1)th time instant can be
acquired by utilizing the feedback information from the output of the Kalman filter at time k.
By adopting the updates from the kinematic relationship, the MS’s predicted position x̂k+1 at
the (k + 1)th time step can by acquired as
1
x̂kKLT
+1 = x̂k + v̂k · ∆t + · â · ∆t2 (4)
2 k
where ∆t is the sampling interval as ∆t = TI . The location estimation and tracking at the
non-dedicated ranging state can therefore be performed.
3.2.2 Geometry-Assisted Location Tracking (GLT) Scheme
MS BS#1 BS#2 BS#3

(Serving) (Neighbor) (Neighbor)
Initiated Dedicated Request non-serving BS to

from Timer Ranging assign a dedicated ranging
Ranging Response
(Every TE State opportunity for the MS
(RNG-RSP)
seconds)
Rendezvous time
Dedicated CDMA code
Measure t_adj1,k
Scanning Interval Allocation Response
(MOB_SCN-RSP)
Rendezvous time
Dedicated CDMA code
Measure t_adj2,k
RNG-RSP
Return t_adj 2,k
. .
. .
. .
BS estimates
MS’s position.
Initiated Non-dedicated UL-MAP

from Timer Ranging Dedicated CDMA code
(Every TI State Measure t_adj1,k
seconds) RNG-RSP
BS estimates
MS’s position.
Fig. 5. The message exchange sequences of propose GLT schemes.
Similar to the KLT algorithm as described in the previous subsection, the proposed GLT
scheme is utilized to provide location estimation during the non-dedicated ranging state. The
concept of the GLT algorithm is to utilize the frequent periodic ranging between the MS and
the serving BS. Based on the periodic ranging, the relative distance r1,k between the serving
BS and the MS can be obtained from the corresponding TOA measurements t1,k . Fig. 5 depicts
the message exchange sequences of the proposed GLT schemes within the IEEE 802.16-2009
network. It is noticed that the flowchart is the same as the general U-TDOA scheme as in Figs.
1 and 2 at the dedicated ranging state. However, at the non-dedicated ranging state, the TOA
measurement is obtained through periodic ranging scheme as shown in Fig. 5.
y=yk
r1,k
BS 1 ( x1,k , y1,k )
y
Geometric e ctor
Traj
Constrained ing
Mo v
Region MS’s
x=xk
MS
( xk , y k ) ( xGLT
k+1
, y GLT )
k+1
Fig. 6. The schematic diagram of the geometric constraints for the proposed GLT scheme.
As shown in Fig. 6, the circular region can be formed according to the center point x1,k =
[ x1,k y1,k ] T with radius of r1,k . Meanwhile, two additional linear equations can be acquired
from the feedback of the Kalman filter at the kth time instant, i.e. x = x̂k , y = ŷk . As a result,
the two linear and one circular equations can be utilized to provide the geometric constraints
for obtaining the MS’s position estimation. Based on the constrained region, the LS method is
employed to minimize the sum of the square errors for the MS’s position. Therefore, the MS’s
estimated position by adopting the GLT scheme is acquired as
−1 T
x̂kGLT T
+1 = G · (H H) H J (5)
where
1 0 0
G= (6)
0 1 0
 
−2x1,k −2y1,k 1
H= 1 0 0  (7)
0 1 0
 2 
r1,k − x1,k − y21,k
2
J= x̂k  (8)

ŷk
Simulations are performed to show the effectiveness of the proposed KLT and the GLT
schemes. Different noise models (Greenstein et al., 1997) are considered in the simulations
in order to represent various environments, including the urban, the suburban, and the rural
cases. In the cellular-based network, an exponential distribution is assumed for the NLOS
model with the distribution of pni,k (υ) as
−υ
1 υi,k
pni,k (υ) = υi,k e υ>0 (9)
0 υ≤0
where υi,k = c · τi,k = c · τm ζ i,k ε ω is the RMS delay spread between the ith BS to the MS at the
kth time step; τm is the median value of τi,k whose value depends on various environments, i.e.
τm = 0.4, 0.3, and 0.1 for urban, suburban, and rural respectively. ε is the path loss exponent
which is assumed to be 0.5. The shadow fading factor ω is a lognormal random variable
with zero mean and standard deviation σω chosen as 4 dB in the simulations. Moreover, the
measurement noises (i.e. mi,k in (1)) is considered Gaussian-distributed as N ∼(0, σm 2 ) with σ
m
= 10 m. The asynchronous offset (i.e. si,k in (1)) between the MS and the BS clock time is also
assumed to be Gaussian-distributed with σs = 7.5 m.
1500
BS
MS’s trajectory
1000
Y Axis (m)
500
0
−1500 −1000 −500 0 500 1000 1500
X Axis (m)
Fig. 7. The geometric layout of the simulation (Green Line: MS’s True Trajectory; Red Empty
Circles: the Position of the BSs)).
Fig. 7 illustrates the MS’s trajectory with the Manhattan street scenario in the simulation.
The MS’s true trajectory is illustrated via the green lines; while the locations of the BSs are
represented by the red empty circles as in Fig. 7. The acceleration is designed to vary at time
t = 21, 26, 31, 47, 63, 76, and 89 sec from ak = (a x,k , ay,k ) = (0, -1), (4, 0), (-4, 0), (0, 2), (0, -2), (-3,
0) to (3, 0) m/sec2 . The corresponding MS’s velocity lies between (0, 70) km/hr.
Fig. 8 shows the performance comparison between the proposed KLT and GLT schemes un-
der the urban environment (with TI = 1 sec and TE = 1, 2, and 4 sec). It is noted that the
position error is defined as Pe = x̂k − x◦k . The case with TI = TE = 1 sec indicates that
the non-dedicated ranging does not exist, which is served as the lower bound of the estima-
tion errors. As shown in the figure, comparably inferior performance is observed with larger
values of TE , which indicates that the dedicated ranging with the TDOA measurement is not
frequently available. Moreover, the proposed GLT algorithm outperforms the KLT scheme for
each specific case, e.g. around 50 m less of the position error under the case of TE = 4 sec with
67% of position errors.
100
90
Average Position Error <= Abscissa (%)
80
70
60
50
40
KLT / GLT (T=T =1sec)
30 I E
KLT (T =1sec. T =2sec)
I E
20 GLT (TI=1sec. TE=2sec)
10 I E
GLT (TI=1sec. TE=4sec)
0
0 100 200 300 400 500
Position Error (m)
Fig. 8. Performance comparison between the proposed KLT and GLT schemes under the urban
environment (TI = 1 sec).
Fig. 9 shows the performance comparison with 67% of position errors between the proposed
KLT and GLT schemes under the different noise environment. It is noticed that the KLT
scheme performs similar to GLT scheme at the smaller NLOS environment, e.g. τm = 0.1
(rural environment). However, the performance of KLT is comparably worse than the per-
formance of GLT scheme under the excessive NLOS errors. The effectiveness of adopting the
periodic ranging in GLT scheme can be observed in the case.
In order to evaluate the overhead of the frequent location tracking, serving BS unavailable
time (i.e. TU ) is defined for the percentage of the frame communicating between the MS and
the neighbor BSs. In the dedicating ranging period, the MS should synchronize with other
BSs and therefore the ongoing transmission should be buffered in serving BS. Although the
more frequent the dedicating ranging performed brings higher location estimation accuracy,
the total usage in the serving BS’s point of view would decrease. In the Fig. 2, the MS needs
to synchronize to the neighbor BSs first and then performs the ranging process. The following
parameter is defined to perform a dedicated ranging with the other BSs:
- Tsyn : average time to synchronize with the new BS
- Trng : average time to perform range process with a BS
The values for Tsyn and Trng are chosen as 20 millisecond (msec) and 30 msec in average as
reported in (Jiao et al., 2007). As the dedicated ranging specified for the LBS, the time of Trng
might be shorter. In terms of the Tsyn and Trng , the serving BS unavailable time counts in
percentage is:
3 ∗ ( Tsyn + Trng )
TU = (10)
TE
200
KLT/GLT (T=T =1sec)
I E
180 KLT (T =1sec. T =2sec)
I E
GLT (TI=1sec. TE=2sec)
160
I E
GLT (T =1sec. T =4sec)
67% Position Error (m)
140 I E
120
100
80
60
40
20
0.1 0.15 0.2 0.25 0.3 0.35 0.4
Median Value of NLOS Noise τ (µs)
m
Fig. 9. Performance comparison between the proposed KLT and GLT schemes under different
NLOS noise (TI = 1 sec).
18
KLT/GLT (TI=1 sec, TE=1 sec)
16 KLT/GLT (T=1 sec, T =2 sec)
I E
KLT/GLT (TI=1 sec, TE=4 sec)
14
Serving BS Unavailable Time (%)
12
10
0
5 10 15 20 25 30 35 40
T (msec)
rng
Fig. 10. Serving BS unavailable time via the average time to perform range process.
It is noted that at least three neighbor BSs participate a dedicated ranging. Fig. 10 shows the
percentage of the BS unavailable time via the Trng from 5 to 40 msec. As the curves shows
that the KLT/GLT scheme with TE = 2sec has half unavailable time than TE = 1sec. While
the overhead of the serving BS is considered as the important factor, comparing to the per-
formance of location estimation in Fig. 9, the KLT/GLT scheme with TE = 2sec is a better
solution with a tradeoff.
5. Conclusion
Two assisted location tracking schemes are proposed in this paper. The schemes are capable
of estimating the position, velocity, and acceleration of the MS during the dedicated rang-
ing state. With the non-dedicating ranging state, the assisted methods utilizing the tracking
information and the periodic ranging information are proposed. It is shown in the simula-
tion results that the proposed location tracking schemes provide consistent performance and
reduces the overhead of the serving BS.
6. References
Chen, C.-L. & Feng, K.-T. (2005). Hybrid location estimation and tracking system for mobile
devices, Proc. IEEE 61st Vehicular Technology Conferencec 2005-Spring, Vol. 4.
Greenstein, L. J., Erceg, V., Yeh, Y. S. & Clark, M. V. (1997). A new path-gain/ delay-spread
propagation model for digital cellular channels, IEEE Trans. Veh. Technol. 46: 477 –
485.
IEEE Std 802.16-2004 (2004). IEEE standard for local and metropolitan area networks - part
16: Air interference for fixed broadband wireless access systems.
IEEE Std 802.16-2009 (2009). IEEE Standard for Local and metropolitan area networks Part 16:
Air Interface for Broadband Wireless Access Systems, pp. 1–2082.
IEEE Std 802.16e-2005 (2006). IEEE Standard for Local and Metropolitan Area Networks - Part
16: Air Interference for Fixed Broadband Wireless Access Systems, Amendment 2:
Physical and Medium Access Control Layers for Combined Fixed and Mobile Oper-
ation in Licensed Bands and Corrigendum 1.
Jiao, W., Jiang, P. & Ma, Y. (2007). Fast handover scheme for real-time applications in mobile
wimax, Proc. IEEE International Conference on Communications, pp. 6038–6042.
Perusco, L. & Michael, K. (2007). Control, Trust, Privacy, and Security: Evaluating Location-
based Services, IEEE Technol. Soc. Mag. 26: 4–16.
Wireless Multi-hop Localization Games for Entertainment Computing 651
Wireless Multi-hop Localization Games

for Entertainment Computing
Tomoya Takenaka†, Hiroshi Mineno‡ and Tadanori Mizuno†
†Graduate School of Science and Technology, Shizuoka University, Japan
‡Faculty of Informatics, Shizuoka University, Japan
1. Introduction
Ad-hoc networking capabilities have provided the flexibility needed to construct various
types of networks without infrastructure base stations. Emerging products for sensor net-
works, such as Zigbee (1), use ad-hoc networking capabilities to construct networks. These
sensor nodes can construct the network such as in outdoor fields and inside buildings with-
out much effort on establishing the base stations, and monitor neighbor information on area
where people usually cannot stay for the monitoring. The technique of ad-hoc networking
has been discussed within the Internet Engineering Task Force (IETF) by the Mobile Ad-hoc
Network (MANET) Working Group (2). MANET is a promising technique to provide the
alternative network infrastructure such as in the disaster case that the existing network infras-
tructures are destroyed because of fires and earthquakes. The wireless terminals with radio
capabilities relay data and deliver to a desired destination. Recently, mobile game consoles
with ad-hoc networking capabilities have been produced by companies such as Nintendo (3)
and Sony Computer Entertainment (SCE) (4). Networking capabilities play an important role
in enabling multiple players to join together to play games. Since the ad-hoc networking
technique is independent of the infrastructure network, it is easy for a player to join a game
through a wireless network. To utilize this functionality, some games using ad-hoc network-
ing capabilities have been developed. However, the games released thus far only use ad-hoc
networking capabilities for joining the game.
We have developed two wireless multi-hop localization games with ad-hoc networking capa-
bilities, and have presented several initial results in (29). The proposed games, a war game
and a tag game, are based on classical field games. Players use mobile game consoles with
ad-hoc networking capabilities to move around a field. The games use wireless multi-hop lo-
calization to estimate node positions. Players on one team jointly establish an ad-hoc network
to estimate their positions and compete for positioning accuracy with the other team. We
used a previously developed multi-hop localization technique called ROULA (28). We used
simulation to evaluate the multi-hop localization games and analyze their characteristics. We
found that node velocity and obstruction position controlled the win rate for the games, and
maintaining connectivity and local rules led to higher win rates for the games. The results
revealed that the proposed games worked well as localization applications using the ad-hoc
networking capabilities.
The purpose of this paper is to present new localization applications using ad-hoc networking
capabilities. The concept behind multi-hop localization games is presented. Wireless multi-
hop localization games are evaluated to find out how well localization-based games with ad-
hoc networking capabilities perform in a simulation. These main results obtained from the
simulation are summarized as follows.
• The win rate for the games depends on node velocity and obstruction position.
• A higher connectivity constraint leads to a higher win rate for the games.
• Enforcing the local rule of the death penalty enables the win rate to be controlled for the
games.
We will next describe the multi-hop localization technique and the current state of mobile
games. The localization technique of ROULA is reviewed in Section 3. Section 4 describes our
wireless multi-hop localization games, and our evaluation of these is discussed in Section 5.
Section 6 concludes the paper with a brief summary and mentions future work.
2. Localization and mobile games

2.1 Multi-hop localization
Multi-hop localization techniques have been discussed for wireless multi-hop networks such
as sensor and ad-hoc networks. The motivation behind developing multi-hop localization is
wanting to know where the node position is in wireless multi-hop networks by using a small
fraction of the anchor nodes. An anchor node is one whose position is known in advance
through means such as global positioning system (GPS). Much research has been conducted
on how to estimate node positions in wireless multi-hop networks (9–15; 28). Most local-
ization techniques can be categorized into two types. The first is localization by using extra
ranging devices, such as ultra sound devices, and the second is localization without using
extra ranging devices.
In AHLoS (11), an iterative multilateration by using time-of-arrival (TOA) measurements was
proposed to estimate large numbers of node positions with a small number of anchor nodes.
The basic idea behind iterative multilateration is that at least three anchor nodes carry out
the multilateration to estimate unknown nodes. Once the positions for unknown nodes are
estimated by anchor nodes, the nodes are configured as pseudo-anchor nodes. Then, pseudo-
anchor nodes join to estimate unknown nodes that remain in the network. Another distance-
measurement approach have been extensively discussed in the literature. In (16), robust tri-
lateration using the rigidity of graph theory for flipping avoidance has been proposed. In
sweeps (17), algorithms to identify global rigidity were employed to estimate the node po-
sitions without flipping for sparse node networks. In (18), an error control algorithm was
formulated to mitigate against the error propagation for iterative localization. The distace-
measurement approach normally achieves precise positioning accuracy. However it requires
extra ranging devices, increasing the cost for all nodes.
The localization scheme without using extra ranging devices has been developed for large-
scale sensor networks, and it basically exploits connectivity information of multi-hop net-
works. In GPS-less (9), anchor nodes first flood beacon packets containing their anchor lo-
cation information, and unknown nodes estimate their positions by using anchor location
information with a Centroid formula. In DV-Hop (10), the positions for unknown nodes in
a network are estimated by using average hop-count distances from anchor nodes. First, an-
chor nodes flood their location information to all other anchor nodes, and calculate the aver-
age 1-hop distance. Next, anchor nodes carry out a trilateration to unknown nodes by using
hop-count distances. In AFL (15), the positions of unknown nodes are estimated without us-
ing anchor nodes. The basic idea behind AFL is to utilize reference nodes that represents the
relative axis in a network. The five reference nodes are automatically selected in the manner
described in (15), and they determine relative node positions based on the hop-counts from
their reference positions. In REP (19), the hop-count distance in a network with holes is cal-
culated by using boundary detection (20). Boundary detection is a technique that can detect
the network boundary with only information on network connectivity. In (21), boundary de-
tection and a Delaunay graph were jointly used to prevent node positions from flipping. The
localization scheme without using ranging devices enables nodes to estimate node positions
while only using the radio capabilities of a sensor node. Hence, it has great flexibility to enable
nodes to be applied to localization in the network.
We previously developed optimized link state routing-based localization (ROULA) (28).
ROULA does not require the use of extra ranging devices for any nodes and precisely esti-
mate the node distance by using multipoint relay (MPR) nodes. We thus used ROULA in
our proposed games to enable the nodes to estimate their positions. ROULA is described in
Section 3.
2.2 Mobile games

Let us now present a brief history of mobile games and discuss different aspects of the pro-
posed multi-hop localization games from these. A number of game consoles have been
developed for the entertainment computing market. These game consoles have two basic
types: home game consoles and mobile game consoles. The Nintendo Entertainment Sys-
tem™ (“Famicom” in Japan) is the iconic home game console and was introduced in 1983
by Nintendo (3). A user plays the games by using a wired hand-held controller. Famicon
supports capabilities for multiplayer games. Two users can play games using two controllers
connected by cables to the console.
Mobile game consoles have been developed with ad-hoc networking capabilities, such as Nin-
tendo DS™ in 2004 by Nintendo (3) and Play Station Portable (PSP)™ in 2004 by SCE (4).
Many video games on mobile game consoles have been released by game software compa-
nies. Hot shots golf™ (6) (“minna no golf” in Japan) is one of popular portable video games
for PSP. Hot shots golf supports multiple players by using ad-hoc networking capabilities.
However, hot shots golf only uses ad-hoc networking capabilities for joining the game.
Mobile games for ubiquitous computing environments have recently attracted a great deal of
attention (23; 26; 27). Many varieties of mobile games have been developed thus far. Geo-
caching (22) is a GPS-based treasure hunting game for outdoor environments. The basic idea
behind geocaching is that players hide and seek out containers called “geocaches”. A player
hides a geocache and registers the positions provided by the GPS receiver. Once the geo-
cache is registered and released on the geocaching web site, another player finds the geocache
based on the positions. Geocaching is being carried out in the actual field, and everyone can
get started by using a GPS receiver and mobile console with Internet capabilities.
Human Pacman (24) is a multiplayer field game using a GPS receiver and wireless networking
capabilities. Pacman is a video game for Famicon and was originally developed by Namco (5)
in 1980. Human Pacman is real field version of Pacman. Palyers are assigned to either the
Pacman team or the Ghost team. Each team has at least two helpers to assist its own team.
Virtual cookies are placed on the map and their positions correspond to a real field. The
goals are for the Pacman team to collect all virtual cookies and for the Ghost team to catch all
the Pacman players. Can You See Me Now? (CYSMN) (25) is a chase game based on location
Game names Scoring Player’s Networking Real No. of partici- Required equipment
metrics behavior for capabilities field use pants, Np
game win
Hot shots Golf game Individual Used to connect to No 1 ≤ NP ≤ 8 Mobile console
golf™ (6) score other players
Classical war Hitting with Individual Not used Yes 2 ≤ NP Model guns and gog-
game model gun gles
Classical tag Avoiding oni Individual Not used Yes 2 ≤ NP None
game and elapsed
time
Geocaching (22) Collecting Individual Used to display loca- Yes 1 ≤ NP Mobile console and
geocaches tions of geocaches GPS receiver
Human Pac- Collecting vir- Individual or Used to record Yes 8 ≤ NP HMD, mobile con-
man (24) tual cookies group player’s trajectory sole and GPS receiver
CYSMN (25) Avoiding Individual Used to display loca- Yes 4 ≤ NP Mobile console and
runners and tions of players GPS receiver
elapsed time
Proposed Increasing Group Used for ad-hoc con- Yes 20 NP Mobile console and
games positioning nections and position GPS receiver
accuracy and estimations
elapsed time
Table 1. Comparisons of conventional mobile games with proposed games.
654
information with GPS. Three runners are visible to players’ locations through an virtual online
map and they run through actual city streets. Players avoid the runners and compete for the
time elapsed since joining the game. If a runner gets within five virtual meters of a player, the
player is seen and is excluded from the game.
Table 1 summarizes the features of current mobile games and the proposed multi-hop local-
ization games. Our proposed game has novel distinct aspects from the other works. First,
our proposed games use ad-hoc networking capabilities. Conventional mobile games use net-
working capabilities such as wireless local area network (WLAN) to access the game server
that provide the players’ location information or a virtual map through the Internet. Our pro-
posed games have novel uses for ad-hoc networking capabilities to conduct multi-hop local-
ization in mobile games. Second, the scoring metric for the game is based on the positioning
accuracy of the multi-hop localization technique. In some of the literature, localization is de-
scribed as being “cooperative” (8). Nodes in the wireless network help to connect with one
another to establish their relative positions. The positioning accuracy depends on the number
of nodes in the network. Hence, nodes are required to cooperate to achieve higher positioning
accuracy. Players must cooperative with their own team in the games.
Finally, let us discuss the number of participants in the proposed games. The proposed games
require large numbers of participants compared with other games. This is because multi-hop
localization without ranging devices requires a large number of nodes to estimate the posi-
tions of the nodes (28). ROULA is guaranteed to estimate all node positions when connectiv-
ity is about 20 (28). Connectivity indicates how many nodes are connected to other nodes in
1-hop on average. The least number of participants can be reduced further by using localiza-
tion with ranging devices although improving the performance of the localization algorithm
is beyond the scope of this paper.
3. Optimized Link State Routing-based Localization

3.1 Overview of ROULA
In our two wireless multi-hop localization games, the players used ROULA (28) to enable
them to estimate their positions. Let us briefly describe the ROULA technique. A more de-
tailed description of ROULA and its performance are described in (28).
Figure 1 has a conceptual representation of ROULA in a non-convex network topology. The
basic idea behind ROULA is that each node matches regular triangles that form exactly convex
curves, and makes them into global coordinates by merging overlapping regular triangles iter-
atively. ROULA is independent of anchor nodes and can determine the correct node positions
in a non-convex network topology. In addition, ROULA is compatible with the optimized
link state routing (OLSR) network protocol (30) and uses the inherent distance characteristic
of MPR nodes.
A non-convex network topology can occur when nodes cannot be deployed in some areas
because of obstructions, e.g., buildings or natural features such as trees or mountains. A non-
convex network topology appears to be a non-convex curve if the network is seen from a
global point of view. However, if the network is viewed locally, each small set of the network
appears to be a convex curve. In other words, a non-convex network topology is composed of
partially convex curves. To find these convex curves, nodes in ROULA search for nodes that
are arranged into regular triangles.
Nodes in ROULA are assumed to use the OLSR protocol in the network layer. Using the OLSR
protocol has two advantages. First, the MPR selection used in OLSR has the inherent charac-
teristic of reducing distance errors in localization without using ranging devices. Second,
Fig. 1. Conceptual representation of ROULA in non-convex network topology.
nodes in OLSR always hold and update the latest 2-hop node information and MPR nodes in
a proactive action that periodically floods hello packets. Node in ROULA localize MPR nodes
as their 1-hop nodes without having to make any modifications to the MPR selection. Flood-
ing hello packets and the computational task of MPR selection can be integrated by using the
underlying network layer processes.
3.2 Algorithm
The four operations for ROULA are summarized below (28).
1. Estimating MPR node distances: Nodes flood hello packets containing their own 1-
hop nodes list to their 1-hop nodes. Once a node has a 2-hop nodes list, it selects MPR
nodes and estimates the distances between them.
2. Estimating farthest 2-hop node distances: Each node selects the farthest 2-hop node
for each MPR node and estimates the distances between them.
3. Estimating relative node positions on regular triangles: Nodes flood TRI_NOTICE
packets to their farthest 2-hop nodes with their farthest 2-hop nodes list. Then, nodes
that received TRI_NOTICE packets match regular triangles by using the received far-
thest 2-hop nodes lists. Nodes then obtain sets of local coordinates by merging their
overlapping regular triangles.
4. Estimating one set of relative coordinates of network: Sets of relative coordinates are
collected and merged into one set of relative coordinates for the network. After that,
nodes that have not estimated their positions estimate these by using the Centroid for-
mula (28). If at least three anchor nodes are in the network, the relative coordinates can
be converted into absolute coordinates that have the correct network orientation.
Fig. 2. Principle state transition diagram for multi-hop localization game.
We assume that nodes are deployed in a two-dimensional plane, and a sink node for the
network merges all sets of local coordinates in the network. Routing protocol operations are
assumed to be done without requiring additional time.
4. Wireless multi-hop localization games

4.1 Overview
The fundamental concept underlying wireless multi-hop localization games is that players on
a team establish an ad-hoc network to estimate their positions and then compete for position-
ing accuracy with the other team by using a multi-hop localization technique. The players
use mobile game consoles, called “nodes”, with ad-hoc networking capabilities and play the
game on a field.
Figure 2 presents a principle state transition diagram for a multi-hop localization game. Each
node is either in a “dead” or an “alive” state. The initial state is alive. Nodes periodically
send hello packets and update their positions by multi-hop localization. The condition for
transition to a dead state is based on positioning accuracy. The more accurate the positioning
obtained by a node, the lower the probability of it transitioning to a dead state. When an alive
node satisfies the condition to transition to a dead state, it makes the transition. The condition
for finalizing the game is different for each game’s goal.
Two objectives in using ad-hoc networking capabilities are to connect nodes within a limited
communications range and to estimate node positions. Once a node is connected to other
nodes, it can specify the number of 1-hop nodes. ROULA can estimate node positions by
using the number of 1-hop nodes.
Here, we have simplified the mobility characteristics of human motions to win a game as
random motions.
Fig. 3. Initial node placements of (a) war and (b) tag games. Teams 1, 2, 3, and 4 correspond to
circles, triangles, squares, and diamonds. Oni is represented by star. Arrows indicate locations
of enemy lines.
Algorithm 1 When node i senses hello packet of node j

1: if node i is on same team as node j then
2: receive packet.
3: else
4: drop packet.
5: if ei == e j && U (0, 100) ≤ 50 then
6: transition to dead state.
7: else if ei > e j then
8: transition to dead state.
4.2 War game

Let Ntm denote the number of teams and Ng denote the number of nodes required to finalize
the war game. Each team has an equal number, Nn , of players. The goal of the war game is for
Ng alive nodes on a team to reach the enemy line. Ntm | Ntm > 1, Ng | Ng > 0, and Nn | Nn ≥ Ng
can be varied. Ntm was consistently set to 2 and Ng was set to 1 for the war game discussed
here.
Let us consider the case of Nn = 80. Figure 3(a) shows the node placement at the beginning
of the war game. The field is divided in half, and each team initially occupies one of the two
areas. Each team has the same number of nodes. The nodes for team 1 are represented by
circles, and those for team 2 are represented by triangles. The arrows in Fig. 3(a) indicate the
locations of the enemy lines, which were set 10 [m] from the back end of each team’s area. The
mobility of each node was modeled as a random waypoint constrained to proceed toward the
enemy line. The velocity of each node was determined by using a random variable with an
Algorithm 2 When node i senses hello packet of oni

1: if U (0, 100) ≤ ei ∗ κ then
2: Node i transitions to dead state.
exponential distribution, E (v), with a mean of v. If a node reaches the enemy line, its team
is declared the winner. The nodes on each team periodically run multi-hop localization to
estimate their positions. The nodes cannot communicate with the nodes on the other team.
We used the positioning error obtained by multi-hop localization as the metric to determine
whether a node transitions to a dead state. In the proposed game, all nodes are assumed to
be anchor nodes to enable relative coordinates to be converted into absolute coordinates and
obtain the positioning error. The positioning error, ei , is normalized by the communication
range, R:

( x̂i − xi )2 + (ŷi − yi )2
ei = , (1)
R
i = 1...N,
where N is the total number of nodes, ( xi , yi ) represents the true position of node i, and ( x̂i , ŷi )
is the estimated position. The true position can be obtained by using GPS. When a node
cannot estimate its own position, it is assigned a positioning error of 100%. A node transitions
to a dead state depending on its positioning error. State transitions algorithm for node i is
described in Algorithm 1. In the proposed games, nodes periodically flood hello packets.
When a node senses a hello packet, the node receives the packet only if it has arrived from
a node on the same team. Otherwise, the node drops the packet and decides whether to
transition to the dead state on the basis of positioning errors (lines 5–8 of Algorithm 1). The
U ( a, b) represents a random variable with a uniform distribution in the interval [ a, b). If the
node has the same positioning error, it transitions to a dead state with a probability of 50%
(line 5 of Algorithm 1). Otherwise, the node transitions to a dead state if it has a greater
positioning error. Once the node transitions to the dead state, it stops moving and receiving
packets for localization. If all the nodes on a team transition to dead states, the other team is
declared the winner.
4.3 Tag game

Let Noni denote the number of demons (“oni” in Japanese). Each team has the same number of
players. The goal of the tag game is for players on each team to survive Noni oni attacks. Play-
ers belong to one of the Ntm teams. The Noni oni move around the field and never transition
to dead states. Ntm | Ntm > 0, Noni > 0, and Nn > 0 can be varied. Here, we have consistently
considered the case where Ntm = 4, Noni = 1, and Nn = 50 for the tag game.
Figure 3(b) shows the node placement at the beginning of the tag game. Four teams are located
in four equal areas. Teams 1, 2, 3, and 4 correspond to the circles, triangles, squares, and
diamonds. The oni is located at the center of the field, and is represented by the star. The
mobility of the oni and nodes was modeled as a random waypoint. The velocity of each node
was chosen by using a random variable with an exponential distribution, E (v). The nodes
periodically run multi-hop localization to estimate their positions.
Algorithm 2 gives the algorithm for node i to transition to a dead state. When node i receives
a hello packet from the oni, it transitions to a dead state according to a probability based on its
Field 500 × 500 [m]

Communication range 100 [m]
Node mobility Random waypoint constrained to advance
toward enemy line
Node velocity E (v), v = 1, 2, 4 [m/step]
Table 2. Simulation parameters for war game
own positioning error. The basic operation of this metric is that if the positioning error is 60%,
the probability to transition to the dead state is 60%. In line 1 of Algorithm 2, we introduced
design parameter κ to mitigate the impact of positioning error. Our evaluation of the impact
of κ is discussed in Section 5.3. We defined max_step, which denotes the maximum length of
time for the tag game. The conditions to finalize the tag game are cases where only one team
survives or max_step time is up. When max_step is out of time, the winner is the team that
has the maximum survival time. The survival time is defined as
max_step
Tsurvive = ∑ number of alive nodes. (2)
t =0
In our evaluation, we set max_step to 1000.
5. Evaluation
5.1 Simulation setting
The simulation environment we used was a discrete event simulation environment, OM-
NeT++ (31) with Mobility Framework (32). Existing network simulators do not have local-
ization functionality. We implemented localization functionality into OMNeT++. Our local-
ization simulation platform enables to test the localization performance with discrete event
simulation. The simulation trials were conducted 30 times with random seeds, and the results
were averaged.
5.2 War game

5.2.1 Impact of number of nodes
Table 2 lists the simulation parameters for the war game. The communication range was fixed
and we have ignored packet loss in this paper. Hello packets were periodically sent by 5 time
step.
Figure 4 shows snapshots of the war game at time steps of 500 and 1000 for 160 nodes. We
set the mean node velocity (v) of all nodes to 1 [m/step]. A cross on a node represents a dead
state. As we can see from Fig. 4(a), the nodes proceeded toward the enemy lines. As time
went by, the nodes got closer to the enemy lines, and more and more nodes died, as seen in
Fig. 4(b).
Figure 5 plots the positioning error, ratio of estimated nodes, and number of alive nodes
against the time step when each team (T) had 120 nodes. The variance in the positioning
error increased as the number of alive nodes decreased, which is consistent with the finding
that the number of nodes contributes to positioning accuracy in multi-hop localization (28).
As time went by, the number of nodes that could participate in localization decreased. Hence,
the variance in positioning accuracy increased. We defined the ratio of estimated nodes as the
Fig. 4. Snapshots of war games for time steps (a) 500 and (b) 1000 (N=160, v T1,T2 ={1,1}). Nodes
for team 1 are represented by circles, and those for team 2 are represented by triangles. Cross
on node represents dead state.
percentage of nodes that could estimate their positions out of all alive nodes. The number of
alive nodes decreased over time. A small number of nodes makes it difficult to estimate node
positions using multi-hop localization. Therefore, the ratio of estimated nodes decreased as
the number of nodes decreased. The number of alive nodes on both teams remained approx-
imately the same over time. This is because all the nodes had the same velocity and used the
same strategy to proceed to the enemy lines.
Figure 6 plots the results for 160 nodes. Compared with the results for 120 nodes, the ratio of
estimated nodes was better, confirming that the number of nodes contributed to the ratio of
estimated nodes.
Table 3 lists the number of wins by team for 30 trials with the different parameter settings. The
row for scenario A in Table 3 shows that the number of wins for teams 1 and 2 for 120 and 160
nodes, corresponded to 16 and 14, and 15 and 15. The number of nodes did not significantly
affect the number of wins.
Although the results presented here were for basic scenarios, we observed that the multi-hop
localization game using ROULA worked well as a game with ad-hoc networking capabilities.
5.2.2 Impact of velocity

We evaluated the impact of velocity by varying the velocities of nodes on each team. Figure 7
shows snapshots of war games at time steps of 250 and 500 for 160 nodes and the mean node
velocities of 1 [m/step] for team 1 and 4 [m/step] for team 2. As seen in Fig. 7(a), the nodes on
team 2 were closer to the enemy line than those on team 1. Figure 7(b) shows that many nodes
on team 2 died on their enemy’s side, and that many nodes on team 1 died on their own side.
T1 (v=1) T2 (v=1)
45 100 60
Ratio of estimated nodes (%)

50
Number of alive nodes

40
Positioning error (%)
80
40
35
60 30
30
20
40
25 10
20 20 0
0 500 1000 0 500 1000 0 500 1000
Time step Time step Time step
Fig. 5. Positioning error, ratio of estimated nodes, and number of alive nodes (N=120,
v T1,T2 ={1,1}).
T1 (v=1) T2 (v=1)
45 100 80

40
80 60
35
60 40
30
40 20
25
20 20 0
0 500 1000 0 500 1000 0 500 1000
v T1,T2 ={1,1}).
Figure 8 plots the positioning error, ratio of estimated nodes, and number of alive nodes when
the velocities of nodes on team 1 were 1 and those on team 2 were 2. The ratio of estimated
nodes of team 2 was slightly lower than that on team 1. This is because the nodes on team 2
were more spread out because they moved more quickly, making it more difficult to estimate
their node positions using multi-hop localization.
However, the goal of the war game was to reach the enemy line. The row for scenario B in
Table 3 indicates the number of wins for teams 1 and 2 for velocities of 1 and 2 [m/step].
Although the ratio of estimated nodes on team 2 was slightly lower than that for team 1,
team 2 had more wins. This is because the condition for finalizing the war game was reaching
the enemy line; the team with the higher average node velocity had the greater number of
wins.
Figure 9 plots the positioning error, ratio of estimated nodes, and the number of alive nodes
when the velocities of nodes on team 1 were 1 [m/step] and those on team 2 were 4 [m/step].
The ratio of estimated nodes on team 2 was lower than that on team 1. However, team 2 had
Team 1 Team 2
Scenario A: 120 16 14
N 160 15 15
Scenario B: {1,2} 12 18
v T1,T2 {1,4} 9 21
Scenario C: 250 15 15
y coord. of obst. 100 29 1
Table 3. Number of team wins in war game. In scenario A, number of nodes N was varied
and for v T1,T2 = {1, 1}. In scenario B, the velocity v T1,T2 was varied for N=160. In scenario C,
y coordinate of obstruction was varied for N=160 and v T1,T2 = {1, 1}.
more wins as can be seen from the scenario B results in Table 3. This result suggests that the
win rate for the war game depends on the node velocity of nodes.
5.2.3 Impact of obstruction position

We evaluated the impact of obstruction position by adding an obstruction to the field and
varying its position. Figure 10 shows snapshots of war games assuming that there is an ob-
struction at position (250, 100). The height and width of the obstruction were 200 and 200
[m], respectively. No node could enter the portion with the obstruction. Figure 10 shows that
nodes had trouble moving forward even though the velocities of the nodes on both teams
were the same.
To evaluate the impact of the obstruction’s position, we fixed its x-axis position at 250 and
varied its y-axis position. As seen in Fig. 11, the positioning accuracy and ratio of estimated
nodes were almost the same when the obstruction’s position was (250, 250). As we can see
from in Fig. 12, when the obstruction’s position was (250, 100), the ratio of estimated nodes
on team 2 was lower than that on team 1. This is because the obstruction made the network
topology non-convex, making it difficult to estimate the node positions using multi-hop lo-
calization (28). Although the positioning accuracy for team 2 was better than that for team 1,
the ratio of estimated nodes was lower. Hence, many nodes were assigned a 100% positioning
error. Consequently, the number of alive nodes on team 2 was less than that on team 1.
Not surprisingly, the row for scenario C in Table 3 reveals that the number of team wins
was closely related to the obstruction’s position. This is because the scoring metric is based
on positioning accuracy. The result proved that the win rate for the war game depends on
obstruction positions.
Since multi-hop localization increases the positioning error in a non-convex network, the char-
acteristics of an obstruction’s position can be considered in team strategies. For example, play-
ers can collaborate to move to avoid making a non-convex network in a team’s topology. The
characteristics of multi-hop localization open the door to creating various game strategies.
5.3 Tag game

5.3.1 Impact of local rule on connectivity constraint
We evaluated the impact of two local rules, i.e, connectivity constraint and death penalty, in
the tag game. The connectivity constraint is a rule where the nodes have to avoid situations
where the current connectivity becomes less than or equal to a specified connectivity. The
connectivity is defined by the number of nodes connected by 1-hop. Thus, nodes have to keep
Fig. 7. Snapshots of war games for time steps of (a) 250 and (b) 500 (N=160, v T1,T2 ={1,4}).
Nodes for team 1 are represented by circles, and those for team 2 are represented by triangles.
Cross on node represents dead state.
Field 700 × 700 [m]

Communication range 100 [m]
Number of nodes 200
Node mobility Random waypoint
Node velocity E (v), v = 1 [m/step]
Connectivity constraint CT1,T2,T3,T4 ={0 (not applied), 2, 5, 8}
Table 4. Simulation parameters for tag game.
moving to prevent the current connectivity from being violated. The rule of the connectivity
constraint can be easy to accomplish in an actual game, because each node can know the cur-
rent connectivity due to the use of mobile game consoles with ad-hoc networking capabilities.
First, we evaluated the impact of local rule of the connectivity constraint. Table 4 presents
the simulation parameters for the tag game. Figure 13 shows snapshots of a tag game with
the local rule of the connectivity constraint (C). Teams 1, 2, 3, and 4 are located at the top left,
bottom left, top right, and bottom right, respectively. The connectivity constraints correspond
to 0 (not applied), 2, 5, and 8. As shown in Fig. 13(a), the nodes on team 1 spread over the
field while those on team 4 are bunched together. Those on teams 3 and 4 spread gradually,
as shown in Fig. 13(b).
Figure 14 plots the results for positioning error, ratio of estimated nodes, and number of alive
nodes. The higher the connectivity constraint, the lower the positioning error. This is because
the positioning accuracy using multi-hop localization depends on the connectivity (28). The
greater the connectivity constraints, the higher the ratio of estimated nodes.
T1 (v=1) T2 (v=2)
50 100 80

45

80 60
40
35 60 40
30
40 20
25
20 20 0
0 100 200 300 400 500 0 100 200 300 400 500 0 100 200 300 400 500
v T1,T2 ={1,2}).
T1 (v=1) T2 (v=4)
50 100 80
45

80 60
40
35 60 40
30
40 20
25
20 20 0
0 100 200 300 400 500 0 100 200 300 400 500 0 100 200 300 400 500
v T1,T2 ={1,4}).
Table 5 summarizes the survival time for the tag game. As seen from the row for scenario D in
Table 5, team 1 with a connectivity constraint of 0, had the longest survival time even though
it had the lowest localization rate. This is because nodes with a lower connectivity constraint
could more readily move around the field. Since the probability of their encountering an oni
was lower, nodes with a lower connectivity constraint could survive longer. This result is not
suitable for playing the game, because there is no advantage to cooperate for localization, and
it does not support the fairness of the game. We thus examined the introduction of a design
parameter and a local rule to control the win rate for the game.
5.3.2 Impact of local rule on death penalty

We next evaluated the local rule of a death penalty to impose a penalty for moving alone. The
local rule of the death penalty was to impose transition to a dead state when the node could
not estimate its own position ω times, consecutively. In addition, we introduced a design pa-
rameter κ to mitigate transition to a dead state by using multi-hop localization. The parameter
κ encourages the longer survival time in the tag game. κ was introduced in Algorithm 2.
Fig. 10. Snapshots of war game for time steps (a) 200 and (b) 400 with obstruction at (250,
100) (N=160, v T1,T2 ={1,1}). Nodes for team 1 are represented by circles, and those for team 2
are represented by triangles. Cross on node represents dead state. Obstructions are drawn as
large gray squares.
Team 1 Team 2 Team 3 Team 4

Scenario D 6429.4 5460.9 5855.8 6001.9
Scenario E 2237.2 2739.4 2937.4 3454.3
Table 5. Survival time in tag game for CT1,T2,T3,T4 ={0,2,5,8}. Scenario D enabled the rule of
connectivity constraint. Scenario E enabled the rule of death penalty.
Figure 15 shows snapshots of the tag game with the local rule of the death penalty enabled.
Nodes on team 1 with a connectivity constraint of 0 are still widely spread out, however, their
death rate is higher due to the local rule of the death penalty.
Figure 16 plots the results for positioning error, ratio of estimated nodes, and number of alive
nodes with the local rule of the death penalty. The κ was set to 0.5, and ω was set to 2. The
number of alive nodes on team 1 decreased over time, because nodes with a lower connectivity
constraint had wider dispersion. As we can see from the row for scenario E in Table 5, the
higher the connectivity constraint, the longer the survival time. This result demonstrates that
the local rule of the death penalty and the design parameters κ, ω effectively maintained the
fairness in the tag game.
T1 (v=1) T2 (v=1)
50 100 80

45
80 60
40
60 40
35
40 20
30
25 20 0
0 500 1000 0 500 1000 0 500 1000
Fig. 11. Positioning error, ratio of estimated nodes, and number of alive nodes with obstruc-
tion at (250, 250) (N=160, v T1,T2 ={1,1}).
T1 (v=1) T2 (v=1)
50 100 80

45
80 60
40
60 40
35
40 20
30
25 20 0
0 500 1000 0 500 1000 0 500 1000
Fig. 12. Positioning error, ratio of estimated nodes, and number of alive nodes with obstruc-
tion at (250, 100) (N=160, v T1,T2 ={1,1}).
6. Conclusion
We developed two wireless multi-hop localization games, i.e., a war game and a tag game,
based on classical field games. The proposed games are played using mobile game consoles
with ad-hoc networking capabilities. The fundamental concept underlying a wireless multi-
hop localization game is that players on a team establish an ad-hoc network to estimate their
positions and then compete for positioning accuracy with other teams obtained using a multi-
hop localization technique. Using simulation, we found that node velocity and obstruction
positions were parameters to control the win rate for the war game. In the tag game, the higher
connectivity constraint led to be surviving longer. The simulations demonstrated that the win
rate for the proposed games depends on obstruction positions and connectivity constraint.
We also demonstrated that introducing a design parameter and enforcing local rules were
needed to control the win rate for the game. The results demonstrated that the proposed
games worked well as games with ad-hoc networking capabilities.
In this work, we simply assumed that nodes had random motion to investigate the primi-
tive operations of proposed wireless multi-hop localization games. In the real world, players
Fig. 13. Snapshots of tag game for time steps (a) 500 and (b) 1000 (N=200,
CT1,T2,T3,T4 ={0,2,5,8}). Teams 1, 2, 3, and 4 correspond to circles, triangles, squares, and di-
amonds. Oni is represented by star. Cross on node represents dead state.
would cooperate to minimize their positioning errors, or oni would employ a strategy to track
alive nodes. These motions can be embedded into node mobility in simulations to obtain
more realistic game results. Since the presented study only covered a range of application
proposals on combining ad-hoc networking and multi-hop localization, we suggest that there
is a need for further research in terms of appropriateness and effectiveness for the games.
Our future work includes detailed evaluations of games with various location-based strate-
T1 (c=0) T2 (c=2) T3 (c=5) T4 (c=8)

45 100 50
45
80
40
40
60
35 35
40
30
30
20 25
25 0 20
0 500 1000 0 500 1000 0 500 1000
Fig. 14. Positioning error, ratio of estimated nodes, and number of alive nodes (N=200)
Fig. 15. Snapshots of tag game for time steps (a) 500 and (b) 1000 with local rule of death
penalty enabled (N=200, CT1,T2,T3,T4 ={0,2,5,8}). Teams 1, 2, 3, and 4 correspond to circles, tri-
angles, squares, and diamonds. The oni is represented by a star. A cross on a node represents
a dead state.
gies with positioning errors, actual game testing in the real field, and verifying the degree of
user satisfactions when they actually play the proposed games.
T1 (c=0) T2 (c=2) T3 (c=5) T4 (c=8)

60 100 50
90 40
50
80 30
40
70 20
30
60 10
20 50 0
0 500 1000 0 500 1000 0 500 1000
Fig. 16. Positioning error, ratio of estimated nodes, and number of alive nodes with local rule
of death penalty (N=200, κ = 0.5 and ω = 2).
7. References
[1] Zigbee Alliance, http://www.zigbee.org/.
[2] MANET Working Group, http://www.ietf.org/html.charters/manet-charter.html.
[3] Nintendo Co.,Ltd, http://www.nintendo.com/.
[4] Sony Computer Entertainment Inc., http://www.scei.co.jp/index_e.html.
[5] Namco Limited, http://www.namco.co.jp/.
[6] Hot Shots Golf: Open Tee, http://www.us.playstation.com/PSP/Games/Hot_Shots_Golf_Open_Tee.
[7] J. Navas and T. Imielinski, “GeoCast - Geographic Addressing and Routing,” Proc.
ACM/IEEE Mobicom, pp. 66–76, 1997.
[8] 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
Processing Magazine, vol. 22, pp. 54–69, 2005.
[9] N. Bulusu, J. Heidemann, and D. Estrin, “GPS-less Low Cost Outdoor Localization For
Very Small Devices,” IEEE Personal Communications Magazine, vol. 7, no. 5, pp. 28–34, 2000.
[10] D. Niculescu and B. Nath, Ad Hoc Positioning System (APS), Proc. IEEE Globecom, vol. 5,
pp. 2926–2931, 2001.
[11] A. Savvides, C. Han, and M. B. Strivastava, “Dynamic Fine-grained Localization in Ad-
hoc Networks of Sensors,” Proc. ACM/IEEE Mobicom, pp. 166–179, 2001.
[12] C. Savarese, J. Rabaey, and K. Langendoen, “Robust Positioning Algorithms for Dis-
tributed Ad-hoc Wireless Sensor Networks,” Proc. USENIX Technical Annual Conference,
pp. 317–327, 2002.
[13] T. He, C. Huang, B. M. Blum, J. A. Stankovic, and T. Abdelzaher, “Range-free Localization
Schemes for Large Scale Sensor Networks,” Proc. ACM/IEEE Mobicom, pp. 81–95, 2003.
[14] D. Niculescu and B. Nath, Ad Hoc Positioning System (APS) Using AoA, Proc. IEEE
Infocom, vol. 3, pp. 1734–1743, 2003.
[15] N. B. Priyantha, H. Balakrishnan, E. Demaine, and S. Teller, “Anchor-free Distributed
Localization in Sensor Networks,” Technical Report TR-892, MIT LCS, 2003.
[16] D. Moore, J. Leonard, D. Rus, and S. Teller, “Robust Distributed Network Localization
with Noisy Range Measurements,” Proc. ACM Sensys, pp. 50–61, 2004.
[17] D. K. Goldenberg, P. Bihler, M. Cao, J. Fang, B. D. O. Anderson, A. S. Morse, and Y. R.
Yang, “Localization in Sparse Networks using Sweeps,” Proc. ACM Mobicom, pp. 110–121,
2006.
[18] J. Liu, Y. Zhang, and F. Zhao, “Robust Distributed Node Localization with Error Man-
agement,” Proc. ACM Mobihoc, 2006.
[19] M. Li and Y. Liu, “Rendered Path: Range-free Localization in Anisotropic Sensor Net-
works with Holes,” Proc. ACM Mobicom, pp. 51–62, 2007.
[20] Y. Wang, J. Gao, and J. S.B. Mitchell, “Boundary Recognition in Sensor Networks by
Topological Methods,” Proc. ACM Mobicom, pp. 122–133, 2006.
[21] S. Lederer, Y. Wang, and J. Gao, “Connectivity-based Localization of Large Scale Sensor
Networks with Complex Shape,” Proc. IEEE Infocom, pp. 13–18, 2008.
[22] Geocaching, http://www.geocaching.com/.
[23] S. Bjork, J. Holopainen, P. Ljungstrand, and K. Akesson, “Designing Ubiquitous Com-
puting Games – A Report from a Workshop Exploring Ubiquitous Computing Entertain-
ment,” Personal and Ubiquitous Computing, vol. 6, pp. 443–458, 2002.
[24] A. D. Cheok, K. H. Goh, W. Liu, F. Farbiz, S. W. Fong, S. L. Teo, Y. Li, and X. Yang,
“Human Pacman: A Mobile, Wide-area Entertainment System Based on Physical, Social,
and Ubiquitous Computing,” Personal and Ubiquitous Computing, vol. 8, pp. 71–81, 2004.
[25] S. Benford, R. Anastasi, M. Flintham, A. Drozd, A. Crabtree, C. Greenhalgh, N. Tanda-

vanitj, M. Adams, and J. Row-Farr, “Coping with Uncertainty in a Location-based Game,”
IEEE Pervasive Computing, vol. 2, no. 3, pp. 34–41, 2003.
[26] C. Magerkurth, A. D. Cheok, R. L. Mandryk, and T. Nilsen, “Pervasive Games: Bringing
Computer Entertainment Back to the Real World,” ACM Computers in Entertainment, vol. 3,
2005.
[27] C. Schlieder, P. Kiefer, S. Matyas, “Geogames: Designing Location-Based Games from
Classic Board Games,” IEEE Intelligent Systems, vol. 21, no. 5, pp. 40–46, 2006.
[28] T. Takenaka, H. Mineno, Y. Tokunaga, N. Miyauchi, and T. Mizuno, “Performance Anal-
ysis of Optimized Link State Routing-based Localization,“ Information Processing Society of
Japan (IPSJ) Journal, vol. 48, no. 9, pp. 3286–3299, 2007.
[29] T. Takenaka, H. Mineno, and T. Mizuno, “Evaluation of Wireless Multi-hop Localiza-
tion Game for Entertainment Computing,” Proc. IEEE International Symposium on Personal,
Indoor and Mobile Radio Communications (PIMRC), 2008.
[30] T. Clausen and P. Jacquet, “Optimized Link State Routing Protocol (OLSR),” IETF RFC
3626, 2003.
[31] OMNeT++ Descrete Event Simulation System, http://www.omnetpp.org/.
[32] Mobility Framework for OMNeT++, http://mobility-fw.sourceforge.net/.
Measuring Network Security 673
Measuring Network Security

Emmanouil Serrelis and Nikolaos Alexandris
UniversityofPiraeus
Greece
1. Introduction
The motive for this research has been the famous quote of Lord Kelvin “You can not
improve what you can not measure”. Today’s information era has given new interpretations
to this, expressing the need to measure abstract concepts such as Information Security.
There are multiple sources, ranging from academic research to industrial reports, such as
(Danahy, 2004), (Fisher, 2009) and (Sonnenreich et al., 2006) that share the same view and
highlight the importance of measuring security within the context of Information
Technologies.
This chapter provides the necessary information as well as the proper tools to measure the
security of both IT Systems and business services are based on IT Systems. The main target
of the methodologies that described is to provide a better way for managing the security of
IT Systems and Infrastructures.
The first section of this chapter covers the basic requirements of any security measurement
methodology. The second section of this chapter introduces a taxonomy of the existing
security measurement methodologies. The third section highlights the limitations of the
existing security measurement methodologies and supports that security should be
calculated instead. The fourth and final section of the chapter elevates the need of new
measurement methodologies for network security. New methodologies should be taking
into account business needs apart from the traditional information technology
requirements.
2. The need for security measurement

According to the American dictionary of Princeton the general definition of measurement is
“the act or process of assigning numbers to phenomena”. Measurement
according tohas a
a rule
very close relationship with metric which, according the same source, is a “asystemofrelated
measuresthatfacilitates ion
theofsome
quantificat ”.
particularcharacteristic
Focusing on the research area of IT (Maizlitsh, 2005) distinguishes between metrics that are
used to quantify values that act as a control of proper functionality and those that act as a
performance indicator. Security measurement is a topic that falls under the first category of
controlling the proper functionality of security processes.
Although IT Security measurement is very interesting and useful topic in both academic
and industrial environments, it deals with the quantification of an abstract concept.
Similarly to any other abstract concept, security measurements tend to have rather vague
implementations. This is justified by the fact that is hard to provide a figure that could
express the current level of security. Thus, the difficulty of measuring security leads to the
research of proper methodologies that could define the appropriate metrics as well as
describe the necessary measurement process.
The expected benefits from the measurement of security are:
 Enable business strategy: IT Security is essential for the development and the
support of trust between organizations, partners, customers and employees. This
implies that there measuring the current level of security can held the alignment of
business and technology strategies with security aspects and requirements.
 Support the daily business activities: Security measurement can accommodate the
increases of the value of information and thus the increased risk levels related to
each organizational asset.
 Facilitate risk management: The provision of a better toolset for managing risks can
improve decision making as well as taking advantage of business opportunities.
 Reduce costs: A limited understanding of security status could lead to a high-cost

operation of IT Systems and business processes, as well as to increased marketing
and promotion expenses in order to “protect” the reputation of the products and
the services of an organisation.
 Comply to regulatory and legal requirements: Being able to present and report the
current status of security and risks is the basic condition for compliance.
A standardized approach of security measurement should aim to enable the operation of an

organization without uncertainties or doubts, within a framework that could quantify the
probability of a threat occurring, estimate the cost a potential damage, depict the
performance overhead of the security processes and evaluate the effectiveness of security
measures.
3. Requirements of security measurement methodologies

Having described the expected benefits of security measurement, this paragraph presents
the necessary attributes of the security measurement methodologies, by interpreting
business requirements in terms of IT Security. The adoption of those requirements if
essential in order to make the measurement results utilizable.
A basic requirement is to enable envision to management in the section of security. Each

measurement methodology should primarily intend to provide information that would
depict the current status as well as the future trends from the security point of view.
Additionally, the analysis of the security measurement should:
 Aid an analyst to diagnose the issues that are related to security and evaluate the
performance of the existing mechanisms and processes.
 Quantify specific security characteristics and parameters.
 Easy the investigation of hypothetical and “before and after” scenarios.
 Focus the measurement interest to the causes, the media and the meaning of the
results instead of the methodologies that were used.
According to (Jaquith, 2007) each measurement methodology should have as many of the
following characteristics as possible:
 Consistently measured
 Cheap to gather
 Expressed as a number
 Uses at least one unit of measure
 Contextually specific
 Partial weight
 Repetitiveness
 Comparability
3.1 Consistently measured

The measurement methodologies provide reliability when they can be calculated with a
reliable way. Different persons should be in position to apply the method and result the
same answers using same set of data. The condition which is required to verify this, is
expressed as follows: “Will two different individuals in which is submitted the same
question give the same answer with regard to the measurement of some size?”. These
measurements should be differentiated from the “measurements” that they depend on the
subjective crises of researchers and analysts that are reported as classifications, gradations
or estimations.
A measurement methodology can ensure its consistence by recording the individual steps
of measurement using a way that will be transparent and explicit to the person that will be
asked to measure. Each measurement methodology should explain “how” each step should
be applied and “why” it is applied with this particular way.
A particularly efficient way of maintaining consistence is the usage of questions of partial

ignorance, that is to say questions that can be answered with “yes” or “no”. Another way is
the use of automated processes that would follow each time the same process of
measurement without procedural divergences.
3.2 Cheap to gather

Each measurement methodology needs time in order to calculate the results. All
measurement methodologies begin row data and afterwards, following the precise steps of
each model, generate some useful information. Hence, initially, somebody or something
should collect the data from a suitable source, convert them to the desirable form, and
finally calculate and format the results.
An efficient measurement methodology should collect those steps of transformation and

format using a unified and fast process. If the process of measurement is insufficient, the
method of collection of data can cost time and money to the organisation, which could have
been spent in the analysis of results.
The high cost of measurement methodology can be caused by a series of factors such as the
frequency of measurements, the complexity of process and its non automated nature.
It is therefore reasonable for a model of measurement to also make proposals on the most
optimal candidate sources of data, in the light of saving time and money
3.3 Expressed as a number

All measurements should be expressed as an absolute number or percentage, which
represents something that measures a quantity of size. Gradations such as “high,
intermediate, low” or “1, 2, 3” (from a third degree scale) represent relative grades but do
not also measure any size, therefore they cannot be used in a proper model of measurement.
Thus, “expressed as an absolute number” implies the number of total elements and not the
number that expresses the order of total elements.
Thus, measurement methodologies that are not expressed as numbers are not suitable for
the measurement of security. Indicators such as traffic lights with the three possible values
“red, yellow, green” they do not constitute some type measurement since they do not
include some kind of numerical scale.
It should be noted that the colors of traffic lights can be used as depiction or presentation of
the current state but in a more abstract level accompanying the necessary numerical data
that should remain the main objective of security measurement.
3.4 Uses at least one unit of measure

Another basic requirement of measurement models of security is that all the related
measurements should also include a relative unit of measurement, which will characterize
the sizes that are been measured. For instance, the measurement “number of natural
invasions in the IT building” uses as a unit of measurement the invasions. With the use of
units of measurement, the researcher knows how to express similar measurements using the
same way.
In certain cases it is better to use more than one measurement units aiming to facilitate the
comparison of different applications. In the previous example the more general
measurement unit can be also mentioned as the “number of individuals that tried to invade
in the IT building”, which is also another unit of measurement. The use of this unit can be
more suitable for the comparison with another measurable size, that of “total number of
individuals that enters in the IT building”.
Another requirement for the good measurement methodologies is that they mean
something to the persons that examine them. They could reveal issues of infrastructure or
service under review improving or demonstrating the value of persons and processes for
the organisation. Even if the close relation to the general context does not constitute a main
requirement for a good measurement, it helps to maintain measurement results inside the
framework of the organisation under discussion while making the results more useful. This
should benefit the end recipients (which are usually the management executives of
organisation) to comprehend the current security status and decide with rational way based
on results of the measurements.
As an example it can be mentioned the use of measurement as "the mean number of attacks"
for the entire organisation. This measurement can have the all above characteristics
(consistence, numerical price etc.) but it does not help anyone to improve his work. If this
measurement is differentiated and connected with the enterprising services that it offers, as
the servers of an electronic trade service, it will be a much more important tool for the
decision-making process of more specific sectors, such as the protection of specific servers
but also the physical protection of personnel.
3.5 Partial weight

The quantification of an individual factor that influences security is without a doubt very
useful. Nevertheless, another important issue is the effect of these individual factors to the
security of an entire organisation.
This characteristic is related with the previous paragraph (“Contextually Specific”) due of
the relativity that is implied between measurable sizes and the overall security of the
organisation. It differs however in the fact that the requirement of overall estimation
includes the way and the size with which a specific measurement influences the security
and the operation of organisation.
As an example, the measurement of “numbers of power failures” is precise and relative to

security. However, the way with which this measurement influences the operation of
organisation can become also the weight of this particular size. Thus, in the case where there
is no way to tackle a power failure (eg. A power generator) the weight of this measurement
concerning the total estimate of security should be a large figure.
It should also be clarified that the requirement of an overall estimate could also be
considered an extension of measurements, which could even require some form of
calculation.
3.6 Repetitiveness
The requirement of repetitiveness includes the measurement of same factors while applying
the same measurement methods in different time periods. This repetition aims in the
verification of previous measurements as well as in the observation and recording of the
evolution of a particular size.
Thus, this repetition should not constitute measurement from one only person but be also
verified from the measurements of different persons.
Regarding the evolution of factors that are measured, the measurements should be
performed periodically, in order to detect unexpected changes, or immediately after a
particular known change which will probably influence security.
So, the measurements should be calculated with a frequency proportional with the rate of
change of process. The methodologies that use samples at regular time intervals can help
the organisations to analyze the effectiveness of security in precise time intervals and
prepare them to be in the position to react in time in case of a new security incident. As
expected, in a decision for whether a measurement should be calculated often, the cost of
measurement should be taken into account in terms of time and money. Alternatively
measurements could be performed only before and after each change.
3.7 Comparability
Also, it is very important that one should measure and observe the improvement or the
deterioration of security as time advances. For this reason, the results of measurements
should comparable to corresponding results of other organisations or different situations for
the same organisation so that they can be contrasted to the current security status.
As reported above, a way to do that is to use common sizes and units of measurement.
Additionally, is possible to measure in equivalent or relevant time periods points that share
common characteristics, such as measurement of number of robberies during the last and
first day of each month.
4. Taxonomy of existing security methodologies

Currently, there are several approaches for Security measurement. Most of them tend to
emphasize on different aspects of security than objective measurements. There are very few
approaches that focus on providing the means for quantifying security. The most noticeable
solutions are:
 Solutions based on Vulnerability analysis
 Solutions based on Penetration testing
 Solutions based on Baseline comparison
 Solutions based on Best-practice and standards
 Solutions based on Risk management
4.1 Solutions based on vulnerability analysis

Solutions based on Vulnerability Analysis such as Microsoft Security Analyzer are
connecting the security status of a networked system to the number its network-related
vulnerabilities. Unfortunately vulnerability analyzers can not be used to measure security of
an entire organization because they do not take into account many factors such as
operational flaws and personnel security. Moreover, the results are not related at all to the
number of actual security incidents.
4.2 Solutions based on penetration testing

Solutions based on penetration testing such as Corsaire Testing, are following the exact
same patterns of the attackers without causing real damage to the systems. However the
presented outcomes of these approaches are more like subjective ratings and gratings than
objective measurements. Additionally, they always focus on the technology related aspects
of the organization and neglect other important factors such as operational and physical
security.
4.3 Solutions based on baseline comparison

The baseline comparison solutions contain standard security controls, which are applicable
to the great majority of IT systems providing basic security. The basis for the decision of
whether the organization or specific service fulfils the security requirements is based on the
Auditor’s personal judgment.
The main issues regarding this kind of approach are that it is very subjective and tends to
change every time the Auditor changes. Again the outcome is more like a rating than a
proper measurement.
4.4. Solutions based on best-practice and standards

These solutions (e.g. ISO/IEC 17799, BS 7799 and NIST SP 800-33) refer to several
suggestions for security countermeasures and controls to improve an organization’s
information security. Although these are approaches are quite thorough and explanatory,
they are more useful when developing new infrastructures and services. So far the aspects
of quantification and measurement have not been dealt with the same zeal.
4.5 Solutions based on risk management

These solutions assess security by describing, analyzing and evaluating single scenarios.
Again, since the estimation of the risk is based on the Auditor’s personal judgment, such
solutions tend to be very subjective.
4.6 Combining the security’s basic elements

Apart from the above categories two more categories of security measurement
methodologies can be proposed. The first is concerned with the combination of security’s
basic elements. These basic elements are Integrity, Availability and Confidentiality. Other
additional elements of IT Security are Non-Repudiation as well as Authentication.
An effort that could be included in this category is that of (Knorr, 2000). Within its
framework, is proposed a structured approach for the analysis of metrics of security and for
the quantification of the overall security of Electronic Business Applications. It uses a table
that represents “overall security” and divides it to smaller parts. These parts correspond to
sites, potential targets and mechanisms of application security and are connected with the
participating parts of an Electronic Business Application (customer, tradesman, means of
communication). This process aims to the calculation of a quantifier of an Electronic
Business Application, which functions as a means for the analysis, planning and
comparison tool of similar applications.
Another approach that falls under this category is the one by (Serrelis, 2007) that aims to
offer the foundation for a model that could help security analysts to quantify and measure
security. Comparing to the requirements that were initially set, the suggested model has
supported the consistency requirement throughout the document by the use of questions
with objective answers. The questions also aimed to answers which would be cheap to
gather, since the answers could come from automated systems such as IPSs. Additionally
the model has managed to express security as a number (percentage). Security calculation
has used at least one unit of measure (such as blocked spam emails) satisfying another
requirement of a proper quantification model. The last requirement has also been covered
since the level of security of the overall enterprise or the individual services makes a lot of
sense to management people. Thus the model can also claim to be contextually specific. On
the downside, it should be pointed out that the model is not considered so much with the
new products, but with the existing services. Other type of questions should be posed to
calculate the security level of new services and/or products.
4.7 Combining factors that are related to security

The second category of security measurement methodologies is concerned with the
combination of factors that are related indirectly to security. Approaches of this type, even if
of limited number, can be grouped together if they measure or calculate elements which are
not directly related to security unlike the previous category. They aim to describe the
relation between security and factors that are easily and objectively measured.
A typical example of this category is the approach that is presented in (Campbell, 2003),
where the economic impact of incidents of security it is examined. The economic impact is
translated in fall of the stock prices of an organisation, as a result of the negative image that
is created in the investment public. Thus the stock price constitutes a factor which can
indirectly be related to the level of security of organisation.
This approach has been an important factor for the development of the approach proposed
and presented in paragraph 6, which also aims in the development of methodology for the
objective measurement of security using factors that are related indirectly to security.
5. From measurement to calculation

As it is defined by Webster dictionary, calculation is “deliberate process for transforming
one or more inputs into one or more results, with variable change”. The term calculation is
used in numerous sciences, from the precisely defined arithmetic calculation to the
calculation of abstract concepts which is implemented with the use of special algorithms
and suitable combination of factors.
Another, alternative way of calculation of sizes is also the statistical analysis, eg. the
calculation of likely results of an electoral result.
In every case the calculation of a factor is advisable in cases where his direct measurement
or the quantification of its size is not feasible. These cases mainly include abstract concepts
that are not straightly measurable, with security being a very representative example.
The calculation of security is differentiated from the measurement of security. While the
measurement is based on the collection and representation of primary data, the calculation
uses primary data with a combinational way so that it produces a result which represents
security.
Because of the abstract nature of IT security, a direct measurement will not have real and
usable results since it will be based on the limitations of the methods reported in the
previou chapter. The calculation of however of security can be more efficient by overcoming
these limitations using measurable sizes.
There are various methods for the calculation of security. These can depend on the
judgments and estimates of researchers thus they are labelled as classifications or
gradations. A second category of methods can be based on statistical methods, which could
lead to an estimate of the level of security. Other methods can combine the measurement of
values with the co-calculation of which the value of security could be deduced.
The optimum method of calculation of security is differentiated depending on the specific

needs of each organisation, service, infrastructure or system. However, in all cases, it
should be selected with a process that will be based on well defined factors that should also
have well defined relations between them.
It should be also clarified that the methods of calculation and the methods of measurement
of security do not constitute alternative solutions from each other, but complementary.
Precise measurement of security should not be considered feasible, due to its abstract
nature, but due to the fact that calculation of security cannot be reliable if it is not based on
measurable factors. The researcher of security should therefore use measurement
methodologies that would result measurable factors. From the combination of these sizes
the value of security will also be deduced.
It becomes easily understood that the calculation of security can be realised with two ways.
These are the quantification of non measurable factors and abstract significances, such as
security and the use of measurable factors that are measured with the appropriate security
measurement methodologies. The latter involve measurable factors and are applied as a
type of quantitative analysis. Also, the level of security can result as an estimate of the
researchers which is applied as a type of qualitative analysis.
6. A new quantification methodology

Within the framework of the current research, a new approach of calculation of security was
also created, which manages security quantification methodology as a value which is
calculated with the combination of certain easily measurable factors. These factors are
related indirectly with security as well as with the level of security of specific business
services.
This approach is based on the principle that the abstract concepts can be calculated with the
combination of factors that is related indirectly with them and themselves can be easily
measured. At the same time it seeks the satisfaction of following general requirements:
 Appreciation of security as factor that is an important part of the business

production environment and not a collateral issue with minimal or no operational
interest.
 Usage and combination of factors that can be measured or be quantified with

objective ways.
 Connection of results of calculation of security with the business decisions.
For the application of the particular methodology of calculation of security, the sources of
primary data should be determined in order to be combined for security calculation. This
particular methodology considers that the value of security can result from the combination
of parameters that are related with it. The factors that were selected are five and they all
concern certain business, functional or commercial value. These five factors are mentioned
as CARLS from initial their names in the English language. These are:
 Compliance: It expresses the percentage of conformity with the legal and

regulatory framework that is applicable to the business service.
 Availability: It expresses the percentage of uptime of service in comparison its

mission time.
 Return: It expresses the size of profits that results from the particular service.
 Liabilities: It expresses the size of economic losses due to the particular service.
 Stock price: It expresses the price of stock of the company that offers the service.
All factors are essentially different aspects of business services and should be available in
order be composed in a value that would represent the level of security for each of the
business services. The usage of these specific factors has two basic advantages, which are
the main reasons for their choice in this model.
The first advantage is the fact that all factors are already have been measured by
organisations for reasons of operational evaluation. This means that there is no need for
additional effort in order to assemble the information required. The second advantage is the
fact that each factor provides an objective image of the organisation and its particular
business services, which objectivity can not disputed.
The measurement and monitoring of all factors is considered essential for the development,
viability and proper operation of each organisation. It is exceptionally usual, in the great
majority of organisations, to monitor and collect all the above factors. In many cases these
factors are also measured for each business service separately. This fact makes the collection
of the necessary elements a simple, easy and feasible process. Moreover, the CARLS factors
have a lot more meaning for the persons that are found in not technical positions and their
mentality is directed form the operational needs of their organisations. Compared to other
approaches this can be seen as an advantage because using this methodology the
understanding but also the usability of the value of security is facilitated which is based on
familiar notions and no on technical terms as “integrity” and “confidentiality”.
The main objective remains the calculation of the value of security of a specific service. This
is implemented with the determination, the quantification and the combination of factors
that can be measured easily and that is related indirectly with the security. This value
portrays the level of security of specific service. Τhe following paragraphs present the
arguments in favour the choice of the particular factors, sating why CARLS factors are
considered suitable for the calculation of security as well as an analysis for each one from
them.
Compliance: The impact of non-compliance is profound. While many small issuers can
operate with inconsistent compliance processes, problems eventually arise. Instead of
focusing on the regulatory and punitive aspects of incomplete or ineffective compliance, this
white paper will examine the functional impact of not remaining compliant with security
regulations. Compliance can impact liquidity, which can affect your ability to raise funds for
growth. The difficult aspect of compliance is knowing everything you have to do, when and
how. Ongoing compliance requires an investment – for the same reasons as the initial
compliance work you did when going public. The liquidity opportunities that initially
attracted your company to the publicly traded arena are the same reasons for remaining
compliant.
Availability: Government organizations and businesses of all sizes need to create and
implement comprehensive business and operations continuity plans. Most organizations
understand that they need to protect their data and systems -an activity known as disaster
recovery. The disaster recovery is only half the battle-enterprises need also be prepared to
quickly and seamlessly restart business processes in order to continue operations.
Return: The owners of a company and the company’s creditors share a similar goal: to
increase wealth. They are thus very concerned about profitability in all phases of operations.
Creditors are specifically concerned that the company use its resources profitably so that it
can pay interest and principal on its debt. Owners are concerned that the company be
profitable so that stock values will increase. Company managers must show they can
manage the owners’ investment and produce the profits that owners and creditors demand.
Because top management must meet the profit expectations of company owners, it passes
down to the lower levels of management those profitability goals, which are then spread
throughout the company. All managers, therefore, are expected to meet profitability goals,
which are often increased and tightened as each level of management seeks a margin of
security.
Liabilities: Most health related businesses would agree that securing insurance is one of the
basic costs of doing business. As responsible business owners, they have budgeted for the
appropriate coverages as a precautionary measure in the event of a loss— especially a
catastrophic loss. However, there are a few optimistic souls who think of business insurance
as an option. These risk takers appear perfectly content to operate their salons with little or
no coverage in place. Unfortunately, most financial experts agree that this is a very
dangerous practice, as those who gamble and lose usually pay a much higher price in the
long run. The bottom line is, while none of us ever expects to get sued, we’ve all got to
accept that even the most adept operator may face litigation for any number of reasons.
While there are a wide variety of insurance coverages available for protecting yourself and
your business, one of the most essential is liability. Liability is an especially important issue
for those in the tanning industry, whose business is providing customers a service which
may pose some risk of injury to them. Liability risks come in many more forms than might
be expected. In addition to liability arising specifically from the use of tanning equipment,
salon owners also may be held accountable for a variety of other kinds of business liabilities,
such as a customer slipping and falling.
Stock Market: By using a stock market return framework to examine the economic
implications of information security breaches, [3] study contributes to the literature
examining the economic effects of information security breaches. We find there evidence of
an overall negative stock market reaction to announcements of information security. The
economic cost of publicly announced information security breaches 445 breaches in major
newspapers, although this finding is not robust across all specifications. Nevertheless, these
results provide some support for the argument that information security breaches adversely
affect the future economic performance of affected firms.
The choice of the above factors satisfies the two of the three basic requirements that had
been placed within framework of the current approach of for the calculation of security, that
is to say to appreciate security as a factor that is part of the business production
environment as well as to use and combine measurable factors that can be quantified with
objectively.
Based on the ideas described in the previous paragraph, a figure that would represent the
level of security for a specific service should take into account all security factors presented.
In order to mathematically express a formula that calculates security, several assumptions
have been made.
Firstly, the notion of Target level for each factor is introduced. The target level is set by the
upper management who is responsible for the overall operating constrains, such as security,
of each business service offered. Within this context, the target level of Compliance,
Availability, Returns and Stock Price are set. The target level of compliance can be the
compliance to a specific industry directive or governmental law or even an international
standard. Similarly, the target level of Availability should be defined taking into account the
business needs of each service.
The Current level of Compliance will be represented as a “Yes” or “No” factor in order to
keep the model as simple as possible. A future extension to that model can express the
current compliance level as a percentage, signifying that a service does or does not cover all
compliance needs (e.g. covers a law but not a specific international standard).
So, the Security figure for each service can expressed as a function of independent variables
by the use of following formula:
C A RL S
SS     (1)
CT AT RT ST
Where:
Ss = Security level of a Specific Business Service

C = Current Level of Compliance [0|1]
CT = Target Level of Compliance (0-1]
A = Current Level of Availability [0…1]
AT = Target Level of Availability [0…1]
R = Current Return of the Service (in a monetary value)
L = Current Liabilities of the Service (in a monetary value)
RT = Target Return of the Service (in a monetary value)
S = Current Stock Price (Represents the company brand) (in a monetary value)
ST = Target Stock Price (in a monetary value)
Having expressed security using the formula above, a proper usage of the results could be
to a financial motivated one. Our target is to balance the spending in Security for each
Business Service in order to maximize the organisation’s Return on Investment:
max ROI ( S1 , S 2 ,..., S n ) (2)
subject to
 (R
i 1
i  I i  Li )  0 (3)
Where:
Sn = Security level of a Business Service

Ri = Revenues of a Business Service
Ii = Total Investments in a Business Service (Part of which is Security spending-investments)
Li = Total Liabilities of a Business Service
Using the last formula another objective is achieved. This is the connection of results of
calculation of security with the business decisions that are related to security investments.
In other words this formula aims to answer the question whether a security investment
would cost more than the expected benefits.
7. Conclusion
The question that was analyzed in this chapter is whether and how the principles of the
security measurement methodologies can be applied so that the objective measurement of
security of business services can be achieved. The motives that support this question are
focused in the justification of expenses and investments that are related with to security.
Thus, although the management of security is closely related to technical and organisational
level it is often difficult to define a quantified “version” of security that would be more
comprehensible and usable in operational level.
This chapter also presents a critical evaluation and categorisation of the requirements of
measurement and calculation methods of security, on which the restrictions of approaches
that exist are based. Additionally a new security calculation approach is developed that
attempts the quantification of security with the use of factors that are related indirectly to
security.
The basic principle that was followed is that the research focus should be moved from the
measurement of security to the calculation of security. Other principles were:
 Appreciation of security as factor that is an important part of the business

production environment and not a collateral issue with minimal or no operational
interest.
 Usage and combination of factors that can be measured or be quantified with

objective ways.
 Connection of results of calculation of security with the business decisions.
The approach for the quantification of security is implemented via calculation. The variables
that are used for the calculation are the CARLS factors, that is to say Compliance C with the
legal and regulatory framework, Availability A of business services, Return R of each
business service, Liabilities L due to specific services and the Stock price S of the
organisation that reflects its fame and public image. The methodology supports that the
security is mirrored in each one from these factors and hence the factors are related
indirectly to that. This connection is expressed with a mathematic formula through the use
of which the factors are considered equivalent.
An important advantage of the methodology is that through its use the management
executives can comprehend more immediately the values that are produced by this and
evaluate better where they should focus and support the security investments. This is
particularly important because Security is an abstract concept which it is not easy to eb
expressed as a measurable value.
One of the restrictions of method is the fact that the all factors they are considered
equivalent during the calculation of security. A future research could investigate further the
degree that security is influenced from every parameter as well as how this is altered in
terms of service, organisation or market type.
8. References
Danahy, J. (2004) The Need for Metrics and Measurement in Application Security, OWASP
Metrics and Measurement Standards
Fisher, D. (2009) Experts call for better measurement of security, Blog, recovered: 14/6/09,
http://www.threatpost.com/blogs/experts-call-better-measurement-security
Sonnenreich, W.; Albanese, J. & Stout, B. (2006) Return On Security Investment (ROSI) – A
Practical Quantitative Model, Journal of Research and Practice in Information
Technology, Vol. 38, No. 1, February 2006
Maizlitsh, B. & Handler, R. (2005) IT Portfolio Management: Step by Step, John Wiley &
Sons, ISBN: 978-0-471-64984-7, US
Sommerville, I. (1996) Software Engineering, Fifth Edition, Addison-Westley, ISBN: 978-
0201427653, UK
Tipton, H. & Krause, M. (2008) Information Security Management Handbook, Sixth edition,
Auerbach Publications, ISBN 978-1420067088
Olzak, T. (2007) The Pros and Cons of Security Risk Management, Tech Republic, recovered:
14/6/09, http://blogs.techrepublic.com.com/security/?p=180
Parker, D. (2007) Risks of risk-based security, Communications of the ACM, Volume 50,
Issue 3 , March 2007, pp 120.
Jaquith, A. (2007) Security Metrics: Replacing Fear, Uncertainty, and Doubt, Addison-
Wesley Professional, ISBN 978-0321349989
Campbell, K.; Gordon, L. A.; Loeb, M. P. & Zhou L. (2003) The economic cost of publicly
announced information security breaches: empirical evidence from the stock
market, Journal of Computer Security, Volume 11, Issue 3 (March 2003) pp 431–
448
Serrelis, Em. & Alexandris, N (2007) An Empirical Model for Quantifying Security Based on
Services, IEEE Computer Society, Proceedings of the International Multi-
Conference on Computing in the Global Information Technology, pp 30
Knorr, K.; Rohrig, S. (2000) Security of Electronic Business Applications: Structure and
Quantification, Proceedings of the 1st International Conference on Electronic
Commerce and Web Technologies EC-Web 2000, pp 25-37
A testing process for Interoperability and Conformance of secure Web Services 689
A testing process for Interoperability and

Conformance of secure Web Services
Spyridon Papastergiou and Despina Polemi
UniversityDepartmentof atics,
Inform UniversityofPiraeus
80KaraoliDimitriou
& Str,34Piraeus,
185 Hellas
1. Introduction
The design, development and implementation of electronic (e-) services relying on XML and
Web Service (WS)-based technologies is the current trend in achieving interoperability. E-
services can be offered either as autonomous Web Services or embedded in Service Oriented
Architectures (SOAs) (High et al., 2005).
In this context, despite the fact that applications with similar business goals adopt the same
technical standards, quite often their interactions capabilities are extremely limited. Thus,
application developers show an increasing concern for evaluating interoperability between
common services which are offered either autonomously or through a SOA. The creation of
a proper framework (EIF) has a significant importance in the evaluation of interoperability
of such services and is accomplished by the precise definition of the applied standards and
guidelines which guarantee the interaction of the services. Existing testing methodologies
developed by various organizations (e.g ISO/IEC 9646, ESTI) treat the interoperability of
services as a generic problem. They merely provide guidelines and describe high level
testing procedures that can be applied to test interoperability of various telecommunication
as well as software and data communication systems. Most Web Service-oriented
methodologies (i.e. WS-I, ebXML IIC framework) demonstrate weaknesses as they are not
capable of testing all the required aspects that compose an interoperability framework and
mostly the security aspects of the message content.
Additionally, in literature, specific testing types (Saglietti et al., 2008) have been presented
defining diverse testing approaches that treat the applications under test either as white
boxes having full knowledge of the software or as black boxes without any understanding of
their internal behaviour or even as grey boxes with limited knowledge of their internal
architecture. The nature the WSs (e.g. geographic distribution of the examined WSs and
dependencies with external trusted third parties) plays an important role in the adoption of
the most appropriate testing type as they raise specific challenges that should be underlined
and taken into account.
Therefore, there is a specific need for targeted methodologies and frameworks that check
and guarantee the end-to-end application interaction capabilities of common Web Services
and follow and deploy the most appropriate testing strategies covering all WSs aspects.
Identifying this need, this paper proposes a well-formed grey box testing methodology
entitled ICoM, able to test whether various services achieve communication effectively
based on the adopted standards. It defines the precise structure of the involved parties,
specifies distinct steps to follow, describes concrete tests that should be applied, and enables
the execution of specific testing suites. ICoM has been applied in order to evaluate the
interoperability of the existing autonomous SELIS e-invoicing service (Kaliontzoglou et al.,
2006) and the SWEB e-invoicing service embedded in a SOA-based platform (SWEB),
(Karantjias et al., 2008).
2. Prior Work
This section presents the existing testing methodologies and frameworks illustrating their
weaknesses and indicating the need for a more holistic methodological framework. Widely
used types of testing are also described identifying the most appropriate method that
should be applied to WSs due to their inherent characteristics.
2.1 Existing Testing Methodologies and Frameworks

Traditionally, interoperability testing methodologies for the Internet Protocols have been
used extensively in the telecommunication industry and in the Internet world. For example
the Open Systems Interconnection - Conformance Testing Methodology and Framework
(ISO/IEC 9646) (OSI, 1997) is a widely spread and successfully applied conformance testing
methodology which has evolved over the years. Nevertheless, it is considered as overly
generic framework that allows a high degree of freedom and gives little practical guidance
(ETSI, 1998).
The European Telecommunications Standards Institute (ETSI), acknowledging the
importance of the testing methodologies, has also contributed towards this direction. ETSI
defined a more integrated framework that consists of two primitive types of tests, the
conformance and the interoperability testing. The ETSI conformance testing (Moseley et al.,
2003), (ETSI, 1995) is based on the ISO/IEC 9646 using its principles as a basis, but not as
strict guidelines. It focuses mostly on making easier, more applicable and more readable the
use of test suites in the proposed methodology. Therefore, ETSI defined a core language the
Testing and Test Control Notation TTCN-3 (ETSI, 2002), (Dibuz & Kremer, 2003) that can be
used for the specification of test suites which are independent of test methods, layers and
protocols.
On the other hand, ETSI interoperability testing (ETSI, 2007) constitutes a generic approach
merely providing guidance on the specification and execution of the interoperability tests.
These guidelines are in the form of recommendations rather than strict rules. The TTCN-3
can also be used in this kind of testing offering a higher level of flexibility.
Despite the independent operation of both types of testing, they are closely connected
satisfying different objectives (Kulvatunyou et al., 2003). Conformance Testing checks to
what extent a solution conforms to the corresponding specification or standard.
Interoperability Testing proves the end-to-end functionality between the solutions. It should
be noted that the use of either type of test does not guarantee interoperability.
A significant limitation of both abovementioned methodologies is that they treat the
interoperability of the WS-based services as a generic problem providing generic testing
practices.
Currently, there are only two widely used WSs testing methodologies, WS-I and ebXML:
The Web Service Interoperability Standardization Organization (WS-I) (Seely & Lauzon,
2005), (Ehnebuske, 2003) is an industry consortium chartered to promote WS
interoperability across platforms, operating systems, and programming languages. It has
released a number of profiles and testing tools that compose a scalable testing environment.
The shortcomings of this methodology are the following:
1. WS-I does not support the definition of specific and discrete test cases that should be
followed during the tests.
2. WS-I tools achieve to monitor only the message flow, without being able to control the
testing execution.
3. WS-I tools do not achieve to support a wide range of evaluation criteria
(interoperability areas or multiple types of testing). They achieve to test only the
conformance of the exchanged messages and the defined services’ descriptions against
the appropriate standards. They fail to cover criteria regarding the interoperability and
the conformance of the applied security features of the message content and the
transformation of the exchanged documents between different formats.
OASIS ebXML (ebXML, 2001) is an end-to-end B2B XML framework that provides concrete
specifications for dynamic B2B collaborations. It has specified the Implementation,
Interoperability, and Conformance (IIC) test framework (OASIS, 2003), (Lee, 2005), (Kim &
Yun, 2003) describing the required architecture and providing the necessary test material to
be processed by the architecture, a mark-up language and format for representing test
requirements, and test suites (set of Test Cases). This approach has the following
shortcomings:
1. ebXML IIC is intended to support conformance and interoperability testing only for
ebXML specifications and implementations.
2. ebXML IIC does not impute the responsibility of a communication failure to the
corresponding system.
3. ebXML IIC does not cover criteria regarding the interoperability and the conformance
of the applied security features of the message content and the transformation of the
exchanged documents between different formats.
Despite their weaknesses, the above frameworks can be used as the basis for the
development of enhanced and more targeted methodologies to test and guarantee the inter-
working of common WS-based applications.
2.2 Testing Types Existing Testing

Software testing, including interoperability testing, may take place at different levels of
depth, depending on the actual known technical details of the application under test. The
bibliography acknowledges three main types of testing: black box, white box and grey box
testing (Peyton et al., 2008).
 Black box testing treats the software as a black-box where only the inputs and outputs
of the black box are tested without any understanding of the internal behaviour or the
adopted specifications. It aims to test the functionality according to the requirements.
Thus, the tester inputs data and only sees the output from the test object based on the
objects published and known interfaces.
 In White box testing, the tester has access and knowledge of the internal data structures,
code, and algorithms of the software. A White Box tester typically analyzes source code,
derives the corresponding test cases and targets specific code paths to achieve a certain
level of code coverage.
 In recent years the term Grey Box testing has come into common usage and it refers to a
technique of testing the system with limited knowledge of its internal architecture. The
tester usually has access to specification documents (beyond simple requirements) and
generates tests based on information such as state-based models or architecture
diagrams.
The selection of the most appropriate testing type relies on the nature of the application
under test. Web Services (WSs) are composed by a set of related and integrated services
(High et al., 2005). The main characteristics of these services are the following (Rizwan &
Mamoon, 2007):
 They can be distributed in the sense that they are located in different geographic areas,
having independent capabilities and are accessible only via specific interfaces that are
described by WSDL documents.
 They can be implemented in diverse programming languages and support independent
operating systems.
 They can be chained with dependencies on other trusted third parties such as PKI and
timestamp authorities.
The business design of WSs is essentially a composition of these repeatable services which
represents and forms the desired business logic. The above features of the WSs introduce
several challenges with respect to testing. White box testing especially in a SOA
environment is quite impractical to perform due to the nature of the WSs that make access to
source code or binaries very difficult. On the other hand, the WS interoperability testing
methodologies (WS-I, ebXML IIC) adopted in the frameworks analyzed in the previous
paragraph only enable black box testing of WS, leading to limited and inefficient test
coverage due to the “blind” nature of that type of testing.
The distributed nature of Web Services makes Grey Box testing ideal for detecting
interoperability flaws on the communication channels between WSs, mostly by leveraging
the rich information contained in the descriptions of the services interfaces (WSDL
documents). A Grey Box Tester is able to identify at a high level the internal structure of the
tested WSs, defining the composed services and accessing the provided interfaces having
limited or even no access to the actual code. Additionally, several test cases regarding the
deployment of the security features and the communication protocols can be performed
covering all the aspects of the WSs. Therefore grey box testing has been adopted as the most
appropriate type to use in the methodology proposed in this paper.
3. The ICoM methodology

In this Section, we will describe the structure of the proposed methodology. Before
presenting its main features and implementation steps, it is imperative to present the
requirements it satisfies which overpass the weaknesses of the existing frameworks. The
requirements that ICoM satisfies are:
 Clarity: the methodology specifies concisely the evaluated entities and the required
information items for the testing process (e.g. test cases and test data).
 Adaptability & Extensibility: the methodology is extensible in the sense it may easily
evaluate new aspects of the WSs and adopt and integrate new testing tools and
libraries.
 Flexibility: the methodology is parameterizable in the sense that different parts of the
methodology can be adopted for the realization of specific sets of test suites.
 Structural: the methodology is structured and comprises a definite and precise set of
implementation steps.
 Independency & scalability: the methodology offers a high level of independency from
testing technologies, the number of entities involved and the platforms hosting systems
under test. This fact will offer the possibility of testing interoperability on several
different WSs.
In order for the methodology to accomplish its objective goals comprises of four district
phases:
 Phase 1“Entity identification and setting”: the entities involved are identified and their
specific setting and order for the tests is defined. The involved entities are the Systems
under Test (SuTs) being evaluated for interoperability and conformance and the Test
Coordination Infrastructure (TCI) which monitors the testing suites applied by the SuTs
in order to identify any erroneous behaviour.
 Phase 2“Entity structure definition”: the structure of the involved entities is identified.
This includes the SuTs, which consist of the actual WS under evaluation and the specific
structural additions of the testing infrastructure that are required in the testing
procedure. The internal structure of the WS is immutable and an analysis of the
available services is carried out. This phase also includes the specification of the precise
structure of the TCI.
 Phase 3 “Conformance testing”: the definition and generation of the executed
conformance test cases for each SuT, based on which the parameterization of the SuTs
structural additions and the TCI is completed. Each SuT is evaluated against the
adopted standards.
 Phase 4 “Interoperability testing”: the formulation and the derivation of the
interoperability test cases, performing the necessary parameterization of the the SuTs
structural additions and the TCI. The actual interoperability testing between all SuTs’
communications takes place during this phase.
The following section describes these four phases in detail.
3.1 Phase 1: Entity Identification and Setting

As shown in the Figure 1, the initial step is the identification of the exact number, order and
setting of all entities participating in the testing suite. The full methodology deployment
demands the execution of the following process:
 Entities (SuTs) definition: defining the entities 1… (SuT
SuT N (Figure 1), TCI) that
participate in the test.
 WebService(WS)declaration : declaring the Web Service that will be tested.
x 7HVWLQJ6FHQDULR
x 5ROH GHWHUPLQDWLRQ
Example 1
acting as the Invoice Issuer (role: Issuer) that invokes SWEB SuT, which acts as the Invoice
Recipient (role: Recipient) handling the dispatched invoice and responding with an
acknowledgment. This scenario covers specific aspects of an invoicing transaction such as
the issuance, dispatch and receipt of an invoice document. The TCI triggers, monitors the
execution of the applied testing suites.
3.2 Phase 2: Entity Structure Definition

During this phase, the actual structure of the involved entities, TCI and SuTs, should be
defined and shaped precisely by the responsible operators, indicating their constituent parts.
The final form of each entity depends upon the tests that will be performed and the suites
that will be followed.
3.2.1 Definition System under Test (SuT)

A SuT primarily contains the WS to be tested (WSuT). A common WSuT, as depicted in
Figure 3, includes a set of primary services {S1,…,Sn} that are combined to form and execute
the WS’s business logic. Each service encapsulates an explicit function having a number of
Inputs and producing a specific Output. It may also interact with other services in order to
complete its objective goal.
Fig. 3. Web Service under Test (WSuT)
A representative example of an advanced WSuT (Papastergiou et al., 2009), (Karantjias et al.,

2008) may include a set of primary services such as the following:
 communication with legacy systems and proper data transformation,
 document management and application of digital signatures and / or encryption as a
secure mechanism on the business level, and
 messaging formulation, processing and application of security mechanisms on the
messaging level.
All these services, following concrete procedures, manage and derive data that are based on
specific standards. The factual conformance of these data to the individual requirements
specified by the corresponding specifications and the ability of other systems to handle these
data correctly constitute factors that are able to leverage and infer the interoperability and
the conformance capabilities of a WSuT. Take for example the case where the security
service of a “WSuT A” signs a XML document according to the W3C XML Digital Signature
standard. Initially, the produced signed document should be tested for compliance with the
principles of this standard and then there should be test that another “WSuT B” is able to
successfully verify that signature. The conclusion that comes from this example is that the
“WSuT A” produces signed documents that conform to the W3C XML Digital Signature
standard and is able to interoperate successfully with the “WSuT B” on the application of
signature at the business level.
Thus, in order for a WSuT to derive the required evaluation data, it should perform a certain
number of activities as represented and indicated by specific test cases. A test case comprises
a set of conditions and variable and depicts the test control logic which is sufficient for the
execution of a testing suite.
ICoM deals with the execution of test cases and the collection of the respective data defining
two structural components the 6HUYLFH2UFKHVWUDWLRQ(QJLQH and the 5HSRUW(QJLQH as seen on
the following figure. These components are part of the testing infrastructure and along with
the WSuT compose an integrated SuT.
Fig. 4. System under Test (SuT)
The Service Orchestration Engine is responsible to coordinate the available WSuT’s services
to perform specific workflows. Each workflow is represented by a predefined BPEL process
which imprints the logic of an executed test case. The BPEL test cases that will be adopted
and executed are designed during phases 3 and 4 where the actual conformance and
Concluding, ICoM defines four major steps that lead to the definition of a SuT.
1. serviceidentification: Identification of the primary services of the WSuT.
2. interface analysis
: Analysis of the services’ interfaces and data types used per service,
extracting the messages types of the inputs and outputs as described in the WSDL
documents. This may include an extensive analysis of several WSDL documents.
3. standardsdefinition
: definition of the standards that each WSuT adopts and implements.
4. testing components adaptation: adaptation of the two structural testing components, the
Service Orchestration Engine and the Report Engine to the WSuT.
It should be noted that the above steps are completely independent of the underlying
infrastructure and the WSuT implementation language and do not bind ICoM to a specific
technological solution. The application of these steps in our demonstrated case study is
depicted in the following example:
Example 2
The SELIS and SWEB SuTs are setup following the aforementioned four basic steps. In
SELIS, we have identified three specific areas of services with diverse functions. These
services are divided into:
 Basic services, which provide the basic functions used to perform primitive tasks. SELIS
includes document management services, message and document transformation
services, message forwarding services, publication and query services and notification
services.
 Security mechanisms and services that address the security requirements of SELIS-
invoicing. It supports the following security mechanisms: digital signatures, advanced
electronic signatures, encryption, timestamping and credential management.
 Infrastructuresupportservices, which manage the connection with the back-office system
such as databases, wrapper software on top of legacy systems, existing ERPs etc.
Each of these services has specific known interfaces as described by the corresponding
WSDL documents having concise inputs and giving a concrete output. The standards that
adopted in the services implementations are the following:
 the XML common business library version 4.0 (xCBL 4.0) and the Exact ERP XML
schema, for the representation of the invoice information,
 the Extensible Stylesheet Language (XSL) Transformations, for the transformation of the
invoice documents,
 several security standards such as W3C XML Encryption, W3C XML Digital Signature
and XML Advanced Electronic Signature (XAdES) for the application of the appropriate
security features on the business level and
 SOAP and the WS-Security standard for Web Services invocation and
 the Web Services Description Language (WSDL) for the description of invoked services.
Fig. 5. SELIS SuT
Figure 5 depicts the overall form of the defined SELIS SuT which is composed by the two
proposed structural testing components, the Service Orchestration Engine and the Report
Engine and the SELIS WSuT. The process is similar in the case of SWEB.
3.2.2 Test Coordination Infrastructure (TCI)

The logical and organizational structure of the TCI is independent of the nature of the SuTs.
On the contrary, technologically, it offers significant flexibility and scalability allowing the
upgrade of already adopted testing tools and libraries and the integration of new more
advanced ones. This upgrade enables the testing of different systems in various aspects.
Fig. 6. Test Coordination Infrastructure
The general structure and functionality of the TCI is depicted in the Figure 6. Three
fundamental layers, the Testing Layer
, the Interoperability andLayer
the Conformance Layer
compose the TCI at a high level.
Fig. 7. Testing Layer
The Testing Layer (Figure 7) consists of the Test Data Repository and the Test Executor . The
Test Executor orchestrates the execution of the deployed BPEL test cases in ICoM phases 3
and 4, based on a predefined schedule formed during these phases. Following this schedule,
the Executor invokes sequentially the interfaces provided by the SuT Service Orchestration
Engine in order to initiate the execution of the corresponding BPEL process. The invocation
of the interfaces may require as input specific parameters (test data) which are retrieved
from the Test Data Repository. As we will see in the following Section, the required test data
rely on the nature of the deployed BPEL processes and are created along with the definition
of the test cases.
Fig. 8. Conformance Layer
The next layer of the TCI is the Conformance Layer (Figure 8) its main responsibility is to
evaluate the conformance of each SuT against the adopted standards in phase 3. The
evaluation process is performed based on the data produced by the SuT during the
execution of the BPEL test cases. This layer contains two sub-components, the Conformance
Engineand the ConformanceTester .
The Conformance Tester consists of a set of testing tools and libraries that actually assess the
conformance of these data to the corresponding standards. The exact tools and libraries that
should be adopted depend on the SuT’s standards that have been defined during the
formulation of a SuT (step 3). These range from tools that verify the conformance of a XML
document to the corresponding XML schema and libraries that validate the security features
of the XML documents to tools that check the conformance of the exchanged SOAP
messages and message descriptions against the respective standards. Therefore, ICoM as
methodology is quite extensible allowing the integration of new more updated
tools/libraries in the Conformance Tester.
In the current phase, the steps that should be performed are the following:
 the identification and the integration of the required tools in the Conformance Tester,
 the implementation of the appropriate tools’ interfaces that enable Conformance Engine
to invoke them. In case that the required interfaces can not be implemented due to the
nature of the testing tools, a manual process is performed for the communication of
these components.
In this layer, the Conformance Engine acts as the recipient and the analyzer of the log files
that are produced by the SuTs in order to specify the tools that should be used for the
evaluated data respectively.
Fig. 9. Interoperability Layer
The Interoperability Layer (Figure 9) is the last but equally important layer comprising the
Message Handler and Interoperability . The Message Handler operates as the
Engine
intermediate node during Interoperability testing (phase 4) that intercepts and delegates the
exchanged messages to the SuTs. It comprises the Monitor Engine which intercepts the
SOAP messages exchanged among the SuTs and the Analyzer Engine that determines
whether these messages conform to their corresponding message descriptions.
The second subcomponent of the layer, the Interoperability Engine, is the actual
interoperability consolidation point. It obtains as input the log files that are produced by the
SuTs and the Analyzer Engine’s conformance results providing them to the test operator.
Based on these results, the test operator is able to do the following:
 notify the SuTs’ interaction possibilities,
 detect the inaccurate points and
 impute the responsibility of the communication failure to the corresponding SuT.
Deductions are made by carrying out a process that consists of three main steps. The first
one is to verify that the SuTs possess and handle the evaluated data without the detection of
any erroneous behaviour. This means that the evaluated data are valid for all the SuTs. The
second step includes a comparison process. During this process the exchanged data (e.g.
exchanged SOAP messages and documents) are compared in order to acknowledge that the
SuTs possess the same data. In the last step, the test operator confirms that the exchanged
messages are part of the business processes that the SuTs have defined in their service
descriptions. These steps enable the operator to identify whether the SuTs are interoperable
and to what extent during the last phase of the proposed methodology.
In the following example we present the form of the TCI in our working example.
Example 3
The TCI has been shaped as presented in Figure 10. All the layer’s components,
Conformance Engine, Test Executor and Interoperability Engine have been implemented in
Java technology. In the Testing Layer, a native XML Database, eXist has been adopted as the
Test Data Repository enabling the storage of XML-based test data. The Conformance Tester
of the Conformance Layer is shaped based on the SELIS and SWEB standards adopting
testing tools that perform the conformance evaluation. Representative tools include testing
environment and libaries for XML signature validation (IBM XML security suite (IBM XML),
IAIK XML signature library (IAIK XML), IAIK XAdES toolkit (IAIK XAdES)) and tools for
XML Schema document conformance (Altova XMLSpy).
The WS-I Tools (Brittenham, 2003), monitor (Brittenham et al., 2005) and analyzer
(Brittenham, 2005), have been used in order to operate as the two subcomponents of the
Interoperability Layer’s Message Handler, Monitor and Analyzer Engine correspondingly.
These tools provide an unobtrusive and automated way to log and analyze Web Service
messages producing the respective conformance results that will be used for the
interoperability evaluation.
Fig. 10. TCI Working Example
Based on the above mentioned form of the TCI, the SELIS and SWEB SuT are able to be
assessed taking into account the corresponding evaluation results that are derived from the
execution of the defined test case in phase 3 and 4 of the proposed methodology.
3.3 Phase 3: Conformance Testing

Conformance testing has the primary goal of testing a WS against the standard it
implements. This testing type involves a single SuT plus the TCI. The main steps that ICoM
proposes are:
 Definition of Conformance Test : In Section
Cases 0, we specified that ICoM adopts a BPEL
representation of the deployed test cases. Our decision was based on the identification
of a number of limitations (Pentafronimos et al., 2008) that the existing test case
languages present and the nature and features of the BPEL language.
Generally, BPEL allows the definition and the representation of specific business flows
in a XML format. It has unique features in both syntax (e.g. flow with activity
synchronization, join condition) and semantics (e.g. dead-path-elimination) that make

BPEL a highly compact and expressive language. In literature, there is intensive
research on providing precise semantics for BPEL and verification of BPEL models
(Fostre et al., 2006), (Xu et al., 2006). Additionally, there exist frameworks and models
(Zheng et al., 2007), (Sinha & Paradkar, 2006), (Yuan et al., 2006), (Yan et al., 2006) that
enable the automatic generation and definition of BPEL-based test cases.
Currently, most of the derived BPEL test cases are generated to test whether the
implementation of a WS conforms to the BPEL behaviour and WSDL interface models. In
ICoM, the BPEL processes specified and adopted depict activity flows that produce the
appropriate evaluated data based on which the conformance and the interoperability
capabilities of a WSuT are assessed.
Fig. 11. BPEL Test Case
As depicted in Figure 11, a BPEL test case is a partially-ordered list of basic activities that
that should be executed during a specific test run. The structural definition of a BPEL test
case is as follows:
BPEL test case = {Activityj, Sj, Inputj, Outputj}.
o Activityj is a set of activities that should be performed.

o Sj is a set of primary WSuT services that are invoked and orchestrated for the
execution of a test case. Each service is associated with the realization of a
specific activity.
o Outputj is the result that each Sj derives. These results are gathered as the
evaluated data of the SuTs.
o Inputj is set of parameters that each service Sj requires according to the
provided functionality in order to complete a defined process. The inputs can
be:
 test data that are fed by the TCI during the initiation of a BPEL process,
 outputs of other primary services,
 parameters that have been defined and are included in the BPEL
process.
Fig. 12. Test Case Activity Graph Diagram
An activity graph diagram is also able to provide a visualization of the BPEL test cases
(Figure 12). It illustrates and reflects the activity flow of the BPEL process in an effective
manner. A test case activity diagram is composed by {N, E, Ns, Nf} where
N is a set of nodes {N1, …, Nn} that correspond to the applied activities,
E is a set of edges {E1, …, En} that are the implicit sequence concatenations as defined by the
performed workflow,
Ns is the Start Node and Nf is the Final Node of the workflow.
In conformance testing, the designed (conformance) test cases include only internal
activities. This means that merely the services of the examined SuT interact with each other
in order to derive the evaluated data. Usually, in this type of test these data come from the
final node of the test case.
The main objective of the current step focuses on the definition of the deployed BPEL
processes based on the standards evaluated against. In this sense, the ICoM is quite
adaptable since it is able to evaluate different aspects of the WSuT.
The test operator is able to create these processes utilizing existing BPEL test case generation
frameworks or any other BPEL creation engine (ActiveBPEL, 2006). The processes are
embedded in the SuT Service Orchestration Engine and the appropriate interfaces are
implemented. During the BPEL generation phase, the required test data for each test case
are specified and produced by the conformance evaluated tools of the TCI or by any other
process.
Additionally, the test operator constructs a test case execution schedule taking into account
the complexity of the applied processes. The schedule is a XML document that consists of
the URLs of the BPEL processes interfaces and the corresponding required test data. This
schedule along with the test data is stored in the Test Data Repository of the TCI. The steps
that follow present and include the actual conformance testing sequence of the proposed
methodology.
Fig. 13. Conformance Testing suite
 Execution of test : In this step, the TCI Test Executor automatically initiates the
cases
testing procedure by invoking the interface of the Service Orchestration Engine
following the execution schedule. This action triggers the execution of the
corresponding BPEL process running the defined test case. In Figure 13, the Ss is the
initial service that is invoked and Si is a set of services that contribute to the derivation
of the evaluated data by the Sf.
 Collection of :results At the end of the test case, the Service Orchestration Engine has
accumulated the evaluated data that are produced as outputs by the involved service of
the WSuT. The Report Engine collects and consolidates these data in log files which are
prepared for evaluation.
 Resultsanalysis . Initially, the Conformance Engine analyzes the log files and extracts the
produced data. Then, the Engine specifies the tools (tool1, … ,tooli) of the Conformance
Tester that should be used based on the nature of these data. The Engine feeds the data
via the implemented interfaces to the corresponding tools, which in turn infer the
conformance of the implementation to the respective standard. Suggestions concerning
any corrections to the implementation are also provided to the SuT, enabling the
deployment of a fine-tuning process that will correct discrepancies.
 Corrective actions : Corrective actions by the WSuTs’ developers may include both re-
design and re-implementation or only updates of specific services within the WSuT.
 Re-executionoffailed : When
teststhe corrective actions are complete, the SuT undergoes a
new test round to check if the previously failed tests are now successfully passed.
At the end of the last step, the process begins again from “Execution of test cases” step in
subsequent iterations until all tests are successful.
In the example that follows, it is illustrated an instance of the execution of the current
phase’s steps in our demonstrated case study.
Example 4
Fig. 14. Activity Diagram of the XAdES conformance test case
Figure 14 depicts an activity diagram representing an example test case for the conformance
testing of the SELIS XAdES implementation against the XAdES standard. The actions that
compose the test case are the following:
 creation of a XML Invoice Document based on the xCBL standard,
 application of Advanced Digital Signature on the produced document. This process
includes four separate sub-processes which occur transparently:
o certificate retrieval of the certificate that will be used for the signature process,
o digital signing of the invoice and certain other properties according to the W3C
XML Digital Signature standard,
o time stamping by requesting, and embedding a time stamp token on the
generated signature according to the IETF 3161 standard, and finally
o revocation information inclusion by embedding certificate revocation status data
after requesting them on-line from the server of the Certification Authority that
has issued the signer’s certificate.
The above test case is to be used as follows: initially the TCI Test Executor initiates the
execution of the corresponding BPEL process of the SELIS Service Orchestration Engine
invoking the appropriate interface. The choreography of the required WSuT services derives
an xCBL invoice document signed according to the XAdES standard as illustrated in Figure
15. The Report Engine embeds the signed document to a log file which is retrieved by the
TCI’s Conformance Engine. The Engine delegates the document to the integrated IAIK
toolkit to validate the applied signature.
Fig. 15. SELIS Conformance XAdES Testing Suite
During our testing effort, in the actual first execution of this test case the toolkit indicated an
incompatibility of the produced signature to the corresponding XML schema. This
discrepancy was pointed out to the developers of the SELIS-invoicing which performed the
necessary corrections. The re-execution of the test case denoted the conformance of SELIS
with the XAdES standard.
3.4 Phase 4: Interoperability Testing

Interoperability testing is a more complex process than conformance testing because it
usually involves at least two SuTs wishing to intercommunicate (Figure 16). The
interoperability steps proposed by ICoM are not different from the corresponding
conformance steps described in the previous section with regards to the logic and the
sequence of the performed steps. On the contrary, the objective goals and the operation of
these steps present significant differentiations. Thus, interoperability testing includes:
 Definition of Conformance Test : TheCasesnature and the structure of the
(interoperability) test cases that are designed and used in this phase are almost similar
with the conformance ones, presented in Section 0. These test cases apart from internal
activities include also and external activities. This means that the primary services of a
SuT do not interact only with each other, but also with the services of other SuTs
(Figure 16). Thus, the complexity of the BPEL processes that should be adopted to
represent the logic of a test case is increased at a significant level.
The definition of the interoperability test cases takes into consideration two parameters.
The first one is the testing scenario and the role assignment to the SuTs which was
defined in the phase I of ICoM. The second parameter is the standards that the SuTs
have adopted and are evaluated against. The specified interoperability test cases should
constitute instance of this scenario taking into account the variants of this scenario
based on the adopted standards.
BPEL processes should be formulated merely for the SuTs that initiate the execution of a
testing suite. For example, the specified scenario, as depicted in Figure 16, is that the “SuT
A” interacts with the “SuT B” retrieving a response. Therefore, based on a defined test case,
the “WSuT A” orchestrates its corresponding services e.g. As and Ak to execute a specific
process that enables service Ai to communicate with the service Bj of the “WSuT B”
expecting a reply which is handled it appropriately. The BPEL process that corresponds to
the above test case includes merely the actions which are executed by the “WSuT A”.
“WSuT B” reacts to this interaction creating the reply according to the logic that is included
in the implementation.
Identically to the conformance testing, the BPEL processes are created via BPEL generation
framework or any other available BPEL engine and are embedded in the Service
Orchestration Engine of the respective SuT. The execution of the test cases is defined by a
testing schedule that is prepared.
Fig. 16. Interoperability Testing Suite
 Execution of test : As presented in Figure 16, this step is completed following a

cases
similar process with the corresponding of the conformance testing. The execution of the
external interactions among the involved WSuTs introduces an additional factor that
should be taken into account and concerns the conformance of the exchanged messages
to the defined business process. Thus, the TCI’s Message Handler participates in the
external interactions intercepting and analyzing the exchanged messages, and checking
that they belong to the proper message domain.
 Collection of:results
The Service Orchestration Engines of the SuTs taken part in the test
case accumulate the whole of the evaluated data that are produced by all the involved
WSuTs’ services. Then the corresponding Report Engines collect and consolidate them
in log files.
 Results analysis. The test results are analyzed and conclusions are made on the
interoperability capabilities of the SuTs, as presented in Section 0.
 : Corrective actions may include both re-design and re-implementation
Corrective actions
or only updates of specific areas within the SuT.
 Re-execution of failed : As soon as the corrective actions have finished, the SuTs
tests
undergo a new test round to check if the previously failed tests are now successfully
passed.
At the end of last step, the process begins again from second one in subsequent iterations
until all tests are successful. In the Example 5 that follows the operation of the
interoperability testing is described in detail.
Example 5
Figure 17 represents the activity diagram of an interoperability test case that is an instance of
our working demonstration scenario, as implemented by the SELIS and SWEB SuTs. This
test case includes only the actions performed by SELIS which is the actual initiator of the
testing procedure.
Fig. 17. Activity diagram of an Interoperability Test Case
As depicted in Figure 18 the process begins from the service A1 of SELIS WSuT that creates a
xCBL invoicing document (I1). The document is signed by service A2 according to the
XAdES standard (SI1) gathering the required time stamps and related certificate revocation
status information data from their respective sources. Service A4 packages the signed invoice
in a SOAP message (M1) where the WS Security features are applied (SM1) using the service
A3. Finally, the service A5 dispatches the message to the to the SWEB.
Fig. 18. SELIS and SWEB-invoicing Interoperability Testing Suite
SWEB handles the received SOAP message according to its predefined internal processes
without following any steps specified by the test operator. Thus, SWEB receives the SOAP
message, decrypts it and verifies the WS Security features. Then, the invoice document is
extracted and the XAdES signature is verified (along with the rest of the cryptographic
information it encompasses like any time stamps). A notification (N1) is created that is
packaged in a new SOAP message (M2) where the WS-Security extensions are applied (M2)
and is sent to SELIS.
The latter as soon as receives the response, handles it according to the test case actions.
Service A5 retrieves the SOAP message and using the service A3 validates the applied
security features. Finally, the service A6 manages the notification.
It should be noted that, the exchanged SOAP messages (SM1 and SM2) are intercepted by the
TCI Message Handler. The Handler analyzes them checking that these messages conform to
the defined SuTs’ WSDL documents and produces the appropriate conformance results.
The SuT Report Engines of both SuTs collect all the evaluation results and import them in
log files. These consist of the exchanged SOAP messages (M1, SM1, M2, and SM2), the
exchanged documents (I1, SI1, N1) and any erroneous behaviour that was reported during
the whole process e.g. the application and validation of the applied security features of the
invoice document.
The TCI Interoperability Engine receives the SuTs’ log files and the Message Handler
conformance results that are provided to the test operator which is responsible to analyze
Papastergiou, S.; Karantjias, A. & Polemi, D. (2007a). Innovative, Secure and Interoperable
E/M-Governmental Invoicing. 18th Annual IEEE International Symposium on
Personal,IndoorandMobileRadioCommunications , Athens.
PIMRC)
(
Secure, interoperable, cross border m-services contributing towards a trustful European
cooperation with the non-EU member Western Balkan countries (SWEB), IST-2006-
2.6.5, 6th Framework Programme, Priority 2, Information Society Technologies
(2007), www.sweb-project.org/.
OSI - Open System Interconnection, Conformance testing methodology and framework
(1997), ISO/IEC 9646.
European Telecommunications Standards Institute (ETSI), available at
http://portal.etsi.org/mbs/Testing/testing.htm
Moseley, S.; Randall, S. & Wiles, A. (2003). Experience within ETSI of the combined roles of
conformance testing and interoperability testing. Standardization and Innovation in
InformationTechnology , pp 177- 189, IEEE Press.
ETSI 300 406 (1995). Methods for testing and Specification (MTS); Protocol and profile
conformance testing specifications; Standardization methodology.
ETSI ES 201 873 (2002): The Testing and Test Control Notation version 3, v2.2.0.
Dibuz, S. & Kremer, P. (2003). Interoperability Testing - Framework and Model for
Automated Interoperability Test and Its Application to ROHC. TestCom203 , LNCS
vol 2644, pp. 243–257, Computer Science.
ETSI EG 202 237 V1.1.2 (2007-04): Methods for Testing and Specification (MTS); Internet
Protocol Testing (IPT); Generic approach to interoperability testing.
Kulvatunyou, B.; Ivezic, N.; Martin, M. & Jones, A.T. (2003). A Business-to-Business
Interoperability Testbed: An Overview. The Fifth International Conference on
ElectronicCommerence , Pittsburg, PA October.
Seely, S. & Lauzon, D. (2005). WS-I monitor tool functional specification , ver. 1.1. Technical
report, WS-I.
Ehnebuske, D. et al. (2003). WS-I Overview. Available at www.ws-
i.org/docs/20021003.wsi.introduction.pdf
ebXML Technical Architecture Project Team, (2001). ebXML Technical Architecture
Specification v1.0.4, February 16.
OASIS ebXML Implementation, Interoperability and Conformance Technical Committee:,
(2003). ebXML Test Framework v0.3, 07 March.
Lee, Y. (2005). ebXML Test Framework on Dually Coupled Asynchronous MSH. Fourth
Annual ACIS International Conference on Computer and Information , Science IC
(
pp. 198-203, IEEE Press.
Kim, D. & Yun, J.H. (2003): Development -of an ebXML Conformance Test System for e-
Business Solutions. EC-Web 203 , LNCS, vol 2738, pp. 145-154, Springer-Verlag
Berlin Heidelberg
IBM XML security suite, available at http://www.alphaworks.ibm.com/tech/
xmlsecuritysuite.
IAIK XML signature library, available at http://jce.iaik.tugraz.at/sic/products/
xml_security.
IAIK XAdES library, http://jce.iaik.tugraz.at/sic/products/xml_security/xades
Papastergiou, S.; Kaliontzoglou, A. & Polemi, D. (2007b) Interoperability issues of a Secure

ELectronic Invoicing Service (SELIS). 18th Annual IEEE International Symposium on
Personal, Indoor and Mobile dio Communications
Ra (PIMRC 207) , 3-7 September,
Athens.
ETSI Technical Report 266 (1996): Methods for testing and Specification (MTS); Test Purpose
style guide.
Saglietti, F.; Oster, N. & Pinte, F. (2008).White and grey-box verification and validation
approaches for safety- and security-critical software systems. Information Security
Technical Report, Vol. 13, ElsevierAdvancedTechnology , ISSN 1363-4127, p. 10 – 16.
Peyton L.; Stepien, B. & Seguin, P. (2008). Integration Testing of Composite Applications,
Proceedings of the Proceedings of the st 14 Annual Hawaii International Con
SystemSciences , IEEE Computer Society.
Rizwan M.; Mamoon Y., (2007). SOA Testing using Black, White and Gray Box Techniques,
White paper Crosscheck Networks, available at http://www.crosschecknet.com/
resources/articles/soatesting-v2.htm.
Brittenham, P. (2003). Understanding the WS-I Test Tool, available at http://www-
128.ibm.com/developerworks/webservices/library/ws-wsitest/.
Brittenham, P.; Durand, J.; Kleijkers, L. & Stobie, K. (2005a). WS-I Monitor Tool Functional
Specification, WS-I Testing Work Group.
Brittenham, P., et al. (2005b). WS-I Analyzer Tool Functional Specification WS-I Testing
Work Group.
Pentafronimos G.; Papastergiou S. & Polemi D. (2008). Definition and representation of test
cases for e-government Web Services, nd 2 International Conference on Theory and
PracticeofElectronic Governance(ICEGOV208) , 1 - 4 December, 2008, Cairo, Egypt.
Fostre, H.; Uchitel, S.; Magee, J. & Kramer, J. (2006). Model based analysis of obligations in
web service choreography, Proc.ofAICT-ICIW . IEEE Computer Society.
Xu, K.; Liu, Y. & Pu, G., (2006). Formalization, verification and restructuring of bpel models
with pi calculus and model checking, IBM, Tech. Rep.
Zheng, Y.; Zhou, J. & Krause, P., (2007). An Automatic Test Case Generation Framework for
Web Services. JournalofSoftware , vol. 2(3), pp. 64-77.
Sinha, A. & Paradkar, A. (2006). “Model based functional conformance testing of web
services operating on persistent data,” in Proc. of TAV-WEB. ACM Press, pp. 17–22.
Yuan, Y.; Li, Z. & Sun, W., (2006). A graph-search based approach to bpel4ws test
generation, Proc.ofICSEA. IEEE Computer Society, p. 14.
Yan, J.; Li, Z.; Yuan, Y.; Sun, W & Zhang, J. (2006). Bpel4ws unit testing: Test case generation
using a concurrent path analysis approach, Proc. of ISSRE. IEEE Computer Society ,
2006, pp. 75–84.
ActiveBPEL Engine Architecture (2006), available at http://www.activebpel.org/
docs/architecture.html.
Karantjias A.; Papastergiou S. & Polemi D. (2008). Holistic Electronic & Mobile Government
Platform Architecture. th 8 Joint Conference on Knowledge - Based Software Engineeri
208(JCKBSE8)0 , IOS Press, August 25-28, Athens, Greece.
Papastergiou, S.; Polemi, D. & Douligeris, C., (2009). SWEB: An advanced mobile Residence
Certificate Service, rd
3 International Conference on e-Democracy Next " Generati
Society: Technological and Legal Issues" , Lecture Notes of ICST (LNICST), Springer,
ISBN 978-963-9799-54-7, Athens, Greece.

Radio Communications

Uploaded by

Copyright:

Available Formats

Radio Communications

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Radio Communications

Uploaded by

Copyright:

Available Formats

I

First published April 2010

Technical Editor: Martina Peric

scheduling, advanced admission control, handover management, systems performance

1. Radio-communications architectures 001

2. Analytical SIR for Cross Layer Channel Model 037

3. The Impact of Fixed and Moving Scatterers on the Statistics

4. Planar Antenna Array Hybrid Beamforming for SDMA

5. A Distributed Multilayer Software Architecture for MIMO Testbeds 077

6. Recent Developments in Channel Estimation and Detection

7. Cooperative MIMO Systems in Wireless Sensor Networks 123

8. Optimal Cooperative MIMO Scheme in Wireless Sensor Networks 151

9. Single/Multi-User MIMO Differential Capacity 167

11. Advanced Hybrid–ARQ Receivers for Broadband MIMO Communications 211

13. A Hybrid Feedback Mechanism to Exploit Multiuser Diversity

15. Bidirectional Cooperative Relaying 281

16. A Novel Amplify-and-Forward Relay Channel Model for Mobile-to-Mobile

17. Resource Management with Limited Capability of Fixed Relay Station

18. On Cross-layer Routing in Wireless Multi-Hop Networks 339

19. Mobile WiMAX Performance Investigation 361

20. Throughput-Enhanced Communication Approach for Subscriber Stations

21. Holdoff Algorithms for IEEE 802.16 Mesh Mode in Multi-hop

22. Call Admission Control Algorithms based on Random Waypoint Mobility

23. Queueing-Model-Based Analysis for IEEE802.11 Wireless LANs

25. Asymmetric carrier sense in heterogeneous medical networks environment 473

27. Inter-RAT Handover Between UMTS And WiMAX 523

28. MHD-CAR: A Distributed Cross-Layer Solution for Augmenting Seamless

29. Mobility in IP Networks: From Link Layer to Application Layer Protocols

30. Positioning in Indoor Mobile Systems 597

31. Location in Ad Hoc Networks 619

32. Location Tracking Schemes for Broadband Wireless Networks 639

33. Wireless Multi-hop Localization Games for Entertainment Computing 651

34. Measuring Network Security 673

2. Transceiver aspects for radio-communications

PAPR due to shape filtering

CW PAPR due to multi-carrier

Mean power depending on

Fig. 1. Main types of radio-communications signals

1) NarrowBand CW (NB-CW) systems, like GSM (GMSK), EDGE (D-QPSK), BLUETOOTH,

2) WideBand CW (WB-CW) systems, like: UMTS (W-CDMA), WiMAX (OFDM enhanced),

Fig. 2. Telecommunications spectrum sharing and power limitations up to 6 GHz

2.2 Basic elements and their imperfections

Bi : Emitted Ai : Ideal Ei : Vectorial Error P0 Padj

Fig. 3. EVM and ACPR definitions

  1 - a.cos θ 2   2.a.sin θ 2 

θbref θbvco Synthesized frequency

Fig. 5. PLL functional blocks and frequency synthesizer phase noise

- RF filters are essential in communications chains for information selectivity (interferences,

Imax I/V characteristic of the transistor

Pout1dB IP3out IP3

Pin (dBm) 3 IP3in

- DDR (f,θ,φ) - DDR (f,θ,φ)

Bandwidth allowed Bandwidth allowed

Fig. 8. Antenna system characteristics in radio-communications

performance of the radio-communications link: spatial and frequency variations of the

3.1 Classic architectures for RF transmitters

Fig. 9. Classic architectures for RF transmitters

3.2 Linearization techniques for the transmitter

Vin Vout A2 : modelisation of A0

Fig. 10. Correction (A and B) and anticipation (C) linearization techniques

rzkl (τ ) = ρkl,k l (0, 0, τ )

ρkl,k l (δT , δR ) = ρkl,k l (δT , δR , 0)