No d’ordre : 4379 Année 2011

THÈSE

présentée
pour obtenir le titre de

DOCTEUR DE L’UNIVERSITÉ DE BORDEAUX

Spécialité : Automatique, Productique, Signal et Image

par Héctor Enrique Poveda Poveda

Techniques d’Estimation de Canal et de

Décalage de Fréquence Porteuse pour Systèmes
Sans-fil Multiporteuses en Liaison Montante

Soutenue le 14 décembre 2011 devant le jury composé de :

Jean-Pierre Cances Professeur des Universités, ENSIL

Rapporteurs
Nicolaı̈ Christov Professeur des Universités, Université de Lille
Christophe Jego Professeur des Universités, IPB-ENSEIRB-MATMECA
Sébastien Houcke Maı̂tre des Conférences, Telecom Bretagne
Eric Grivel Professeur des Universités, IPB-ENSEIRB-MATMECA Directeur de thèse
Guillaume Ferré Maı̂tre des Conférences, IPB-ENSEIRB-MATMECA Co-encadrant

Préparée à l’Université de Bordeaux

351 avenue de la Libération - 33405 Talence Cedex Laboratoire d’accueil :
Laboratoire IMS
351 avenue de la Libération - 33405 Talence Cedex
Acknowledgments
First of all, I would like to thank my friends all over the world who helped me
during this period of my life.
I would like to express a huge appreciation to Prof. Eric Grivel and Dr. Guil-
laume Ferré for their guidance, advices and wise comments; without them this
PhD. dissertation could have never been possible.
I would like to thank Prof. Mohamed Najim and Prof. Yannick Berthoumieux,
directors of the signal and image processing team, while I was there, for giving
me the opportunity to join this group.
I would like to express my gratefulness to the reading committee: Prof. Nicolai
Christov and Prof. Jean-Pierre Cances for the careful reading of this thesis and
for their helpful comments. I would like to express my gratitude to the exam-
ination committee: Prof. Christophe Jego and Dr. Sebastien Houcke for their
constructive comments to my thesis.
I would also like to thank all the members of the signal and image processing
team group especially those of room 623, for their friendship, support, advices
and helpful suggestions.
I wish to thank Prof. Olivier Lavialle and Dr. Fernando Merchan for the help
they granted me at the beginning of this period.
I would like to express my gratitude to the Panamanian government for offering
me a scholarship to do a PhD. I also want to express my deep appreciation to
my friend Reinaldo Mclean and my aunt Elizabeth Poveda for their help in the
scholarship process.
Last but not least, I want to thank all my family in Panama especially my parents
and my sisters for the support they have always given me.
Abstract
Multicarrier modulation is the common feature of high-data rate mobile wireless
systems. In that case, two phenomena disturb the symbol detection. Firstly,
due to the relative transmitter-receiver motion and a difference between the local
oscillator (LO) frequency at the transmitter and the receiver, a carrier frequency
offset (CFO) affects the received signal. This leads to an intercarrier interference
(ICI). Secondly, several versions of the transmitted signal are received due to the
wireless propagation channel. These unwanted phenomena must be taken into
account when designing a receiver. As estimating the multipath channel and the
CFO is essential, this PhD deals with several CFO and channel estimation meth-
ods based on optimal filtering.
Firstly, as the estimation issue is nonlinear, we suggest using the extended Kalman
filter (EKF). It is based on a local linearization of the equations around the last
state estimate. However, this approach requires a linearization based on calcula-
tions of Jacobians and Hessians matrices and may not be a sufficient description
of the nonlinearity. For these reasons, we can consider the sigma-point Kalman
filter (SPKF), namely the unscented Kalman Filter (UKF) and the central differ-
ence Kalman filter (CDKF). The UKF is based on the unscented transformation
whereas the CDKF is based on the second order Sterling polynomial interpola-
tion formula. Nevertheless, the above methods require an exact and accurate a
priori system model as well as perfect knowledge of the additive measurement-
noise statistics. Therefore, we propose to use the H∞ filtering, which is known to
be more robust to uncertainties than Kalman filtering. As the state-space repre-
sentation of the system is non-linear, we first evaluate the “extended H∞ filter”,
which is based on a linearization of the state-space equations like the EKF. As an
alternative, the “unscented H∞ filter”, which has been recently proposed in the
literature, is implemented by embedding the unscented transformation into the
“extended H∞ filter” and carrying out the filtering by using the statistical linear
error propagation approach.
The above techniques have been implemented in different multicarrier contexts:
Firstly, we address the estimation of the multiple CFOs and channels by means
of a control data in an uplink orthogonal frequency division multiple access
(OFDMA) system. To reduce the amount of control data, the optimal filtering
techniques are combined in an iterative way with the so-called minimum mean
square error successive detector (MMSE-SD) to obtain an estimator that does
not require pilot subcarriers.
Then, as the SPKF gives a better compromise between computational complex-
ity and estimation performances when the noise characteristics are available, we
use this filter in a dynamic bandwidth allocation context. In that case, network
resources are dynamically adjusted to give the appropriate bandwidth to each
user at any time. This plays a key role in cognitive radio systems where the spec-
trum that is not used by licensed (primary) users must be detected in order to
be shared with unlicensed (secondary) users. However, harmful narrow band in-
terferences (NBI) produced by a cognitive radio system may appear. Therefore,
we propose to combine the estimation of the CFOs and channels by means of
SPKF with a statistical test to check whether there are interferences or not. Two
statistical tests, namely the binary hypothesis test (BHT) and the cumulative
sum (CUSUM) test, are studied.
Secondly, we deal with a system that combines orthogonal frequency division
multiplexing (OFDM) and the interleaved division multiple access (IDMA). The
OFDM-IDMA systems achieve a very high single-user capacity. However, the
complexity in the correction of the CFO effects is higher than in OFDMA sys-
tems. For that reason, we propose to modify the iterative architecture of the
IDMA receiver by using the estimations performed by a SPKF. In addition, an
iterative space-time block coding (STBC) OFDM-IDMA receiver is proposed.
The above optimal-filtering based architectures in the OFDMA and OFDM-
IDMA are compared with existing estimation methods. Simulation results clearly
show the efficiency of the proposed algorithms in terms of CFO estimation, chan-
nel estimation and bit error rate performances.

Keywords: Kalman filtering , EKF, CDKF, UKF, H∞ filtering, “extended H∞

filter”, “unscented H∞ filter”, multicarrier, OFDMA, OFDM-IDMA, CFO esti-
mation, channel estimation, cognitive radio.
Contents

Acronyms and Abbreviations v

Notations xv

Introduction 1

1 Multicarrier High-Data Rate Mobile Wireless Systems 7

1.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
1.2 Wireless Communication Systems . . . . . . . . . . . . . . . . . . 8
1.2.1 Conventional Systems . . . . . . . . . . . . . . . . . . . . 9
1.2.2 Cognitive Radio Systems . . . . . . . . . . . . . . . . . . . 9
1.3 Propagation Channel Model . . . . . . . . . . . . . . . . . . . . . 11
1.3.1 Rician and Rayleigh Fading Channels . . . . . . . . . . . . 13
1.3.2 Frequency-Selective and Time-Varying Fading Channels . . 15
1.4 Multiple Access Techniques for Mobile Systems . . . . . . . . . . 19
1.4.1 About Multicarrier Multiple Access Techniques . . . . . . 22
1.4.1.1 About Orthogonal Frequency Division Multiple
Access (OFDMA) . . . . . . . . . . . . . . . . . . 24
1.4.1.2 About Single-Carrier Frequency Division Multiple
Access (SC-FDMA) . . . . . . . . . . . . . . . . 25
1.4.1.3 About Orthogonal Frequency Division Multiplexing-
Interleaved Division Multiple Access
(OFDM-IDMA) . . . . . . . . . . . . . . . . . . . 26
1.4.2 OFDMA System . . . . . . . . . . . . . . . . . . . . . . . 27
1.4.2.1 OFDMA Uplink Transmitter . . . . . . . . . . . 27
1.4.2.2 OFDMA Uplink Receiver . . . . . . . . . . . . . 33
1.4.3 OFDM-IDMA System . . . . . . . . . . . . . . . . . . . . 36

i
CONTENTS

1.4.3.1 OFDM-IDMA Uplink Transmitter . . . . . . . . 37

1.4.3.2 OFDM-IDMA Uplink Receiver . . . . . . . . . . 37
1.5 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43

2 Frequency Synchronization and Channel Equalization of OFDMA

Uplink Systems 45
2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47
2.2 State of the Art on the OFDMA Uplink Systems - Carrier Fre-
quency Offset (CFO) and Channel Estimation . . . . . . . . . . . 47
2.3 A Joint CFO/Channel Estimator for OFDMA Systems . . . . . . 53
2.3.1 State-Space Representation of the System . . . . . . . . . 54
2.3.2 Kalman Filtering in the Non-Linear Case . . . . . . . . . . 56
2.3.3 H∞ Filtering in the Non-Linear Case . . . . . . . . . . . . 57
2.3.4 Simulation Results . . . . . . . . . . . . . . . . . . . . . . 58
2.3.4.1 Estimation Performance . . . . . . . . . . . . . . 59
2.3.4.2 Computational Complexity . . . . . . . . . . . . 62
2.3.5 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . 65
2.4 An OFDMA Non-Pilot Aided Iterative CFO Estimator . . . . . . 66
2.4.1 Problem Formulation . . . . . . . . . . . . . . . . . . . . . 67
2.4.2 User Detection - The Minimum Mean Square Error Succes-
sive Detector (MMSE-SD) . . . . . . . . . . . . . . . . . . 68
2.4.3 CFO Estimation with the MMSE-SD Preambles . . . . . . 71
2.4.3.1 Multiple Access Interference Cancellation Process 72
2.4.4 Simulation Results . . . . . . . . . . . . . . . . . . . . . . 72
2.4.4.1 Test 1: Recursive Estimation Using Perfectly Es-
timated MMSE-SD Preambles . . . . . . . . . . . 73
2.4.4.2 Test 2: Non-Pilot Aided Estimation . . . . . . . 74
2.4.4.3 Test 3: Influence of the CIR Estimations . . . . . 76
2.4.4.4 Computational Complexity . . . . . . . . . . . . 78
2.4.5 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . 78
2.5 An OFDMA Robust CFO Estimator . . . . . . . . . . . . . . . . 79
2.5.1 System Description . . . . . . . . . . . . . . . . . . . . . . 80
2.5.2 Joint Disturbance Detection and CFO Estimation . . . . . 84

ii
 CONTENTS

2.5.2.1 Combining SPKF and Binary Hypothesis Test

(SPKF-BHT) . . . . . . . . . . . . . . . . . . . . 88
2.5.2.2 Combining SPKF and CUSUM Test (SPKF-CT) 88
2.5.2.3 Improving the Disturbance Detection . . . . . . . 91
2.5.3 Simulation Results . . . . . . . . . . . . . . . . . . . . . . 92
2.5.3.1 How to Improve the Disturbance Detection? . . . 94
2.5.3.2 Comparative Study: Influence of the Disturbance
Power over the CFO Estimation and Channel Es-
timation Performance . . . . . . . . . . . . . . . 96
2.6 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 99

3 Frequency Synchronization and Channel Equalization of OFDM-

IDMA Uplink Systems 101
3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 102
3.2 OFDM-IDMA Modified Receiver . . . . . . . . . . . . . . . . . . 105
3.2.1 Multiple CFO and Channel Estimation . . . . . . . . . . . 107
3.2.2 Modified IDMA Receiver . . . . . . . . . . . . . . . . . . . 109
3.2.3 Simulation Results . . . . . . . . . . . . . . . . . . . . . . 111
3.2.4 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . 115
3.3 Space-Time Block Code OFDM-IDMA Receiver . . . . . . . . . . 115
3.3.1 System Description . . . . . . . . . . . . . . . . . . . . . . 116
3.3.2 Multiple CFO and Channel Estimation . . . . . . . . . . . 119
3.3.3 Maximum Ratio Combining Operations . . . . . . . . . . . 119
3.3.4 Modified Elementary Signal Estimator . . . . . . . . . . . 120
3.3.5 Simulation Results . . . . . . . . . . . . . . . . . . . . . . 122
3.4 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 124

Conclusions and Perspectives 127

Appendices

A Kalman Filter - Linear Case 131

B H∞ Filter - Linear Case 135

iii
CONTENTS

C Kalman Filter vs H∞ Filter 139

D Extended Kalman Filter Approaches 143

E Sigma-Point Kalman Filter 151

F Extended and Unscented H∞ Filter 157

References 161

iv
Acronyms and Abbreviations

1,2,3,4G 1st, 2nd, 3rd, 4th Generation Mobile Wireless Systems

3GPP 3rd Generation Partnership Project
3GPP2 3rd Generation Partnership Project 2
A/D Analog-Digital
ADD CP ADDition of a Cyclic Prefix
AF Amplify-and-Forward
AMPS Advanced Mobile Phone Service
AWGN Additive White Gaussian Noise
BHT Binary Hypothesis Test
BPF Band-Pass Filter
BPSK Binary Phase-Shift Keying
BS Base Station
CAS Carrier Allocation Strategy
CDKF Central Difference Kalman Filter
CDMA Code Division Multiple Access
CDMA EVDO CDMA EVolution-Data Optimized
CFO Carrier Frequency Offset
CIR Channel Impulse Response
CP Cyclic Prefix
CR Cognitive Radio
CR-NBI Cognitive Radio Narrow-Band Interference
CSI Channel State Information
CUSUM CUmulative SUM
D/A Digital-Analog
DAB Digital Audio Broadcasting
DDF Divided Difference Filter
DEC DECoder
DF Decode-and-Forward
DFT Discrete Fourier Transform
DVB-T Digital Video Broadcasting Terrestrial
EDGE Enhanced Data rates for GSM Evolution
EDGEev Enhanced Data rates for GSM Evolution evolved
EKF Extended Kalman Filter

v
Acronyms and Abbreviations

EM Expectation-Maximization
ENC low-data rate ENCoder
EQU channel EQUalizer
ESE Elementary Signal Estimator
FCC Federal Communications Commission
FDMA Frequency Division Multiple Access
FFT Fast Fourier Transform
GPRS General Packet Radio Service
GRV Gaussian Random Variable
GSM Global System for Mobile Communications
HSDPA High-Speed Downlink Packet Access
HSPA+ High-Speed Packet Access Plus
HSUPA High-Speed Uplink Packet Access
IBI Inter Block Interference
IC Interference Cancellation
ICI Inter Carrier Interference
IDFT Inverse Discrete Fourier Transform
IDMA Interleaved Division Multiple Access
IEEE Institute of Electrical and Electronics Engineers
IEKF Iterative Extended Kalman Filter
IFFT Inverse Fast Fourier Transform
ISI Inter Symbol Interference
ITU International Telecommunication Union
KF Kalman Filter
LLR Logarithm Likelihood Ratio
LMS Least Mean Squares
LO Local Oscillator
LS Least-Squares
LTE Long Term Evolution
LTE-A LTE Advanced
M-PSK M-ary Phase-Shift Keying
MAI Multiple Access Interference
MAP Maximum A Posteriori
MC-CDMA Multi Carrier CDMA
MC-DS-CDMA Multi Carrier Direct Sequence CDMA
MIMO Multiple-Input Multiple-Output
ML Maximum Likelihood
MMSE Minimum Mean Square Error
MMSE-SD Minimum Mean Square Error Successive Detection
MRC Maximum Ratio Combining
NAMTS Nippon Automatic Mobile Telephone System

vi
 Acronyms and Abbreviations

NBI Narrow-Band Interference

OFDM Orthogonal Frequency Division Multiplexing
OFDMA Orthogonal Frequency Division Multiple Access
OQAM Offset Quadrature Amplitude Modulation
P/S Parallel-Serial converter
PDC Personal Digital Cellular
PU Primary User
QKF Quadrature Kalman Filter
QPSK Quadrature Phase-Shift Keying
RF Radio Frequency
RLS Recursive Least Squares
SAGE Space Alternating Generalized Expectation-maximization
SC-FDMA Single Carrier Frequency Division Multiple Access
SIR Signal-to-Interference Ratio
SNR Signal-to-Noise Ratio
SOEKF Second-Order Kalman Filter
SPKF Sigma-Point Kalman Filter
SPKF-BHT Sigma-Point Kalman Filter combined with a BHT
SPKF-CT Sigma-Point Kalman Filter combined with a CUSUM Test
SR-UKF Square-Root UKF
STBC Space Time Block Code
TACS Total Access Communication System
TDMA Time Division Multiple Access
UKF Unscented Kalman Filter
UMB Ultra Mobile Broadband
UMTS Universal Mobile Telecommunication System
WCDMA Wideband CDMA
WiMAX Worldwide Interoperability of Microwave Access
WLAN Wireless Local Area Network

vii
List of Figures

1.1 Cognition cycle . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10

1.2 Multipath propagation channel . . . . . . . . . . . . . . . . . . . 11
1.3 Channel impulse response . . . . . . . . . . . . . . . . . . . . . . 12
1.4 Rician distribution . . . . . . . . . . . . . . . . . . . . . . . . . . 14
1.5 Rayleigh distribution . . . . . . . . . . . . . . . . . . . . . . . . . 15
1.6 Frequency-selective fading channel . . . . . . . . . . . . . . . . . . 17
1.7 FDMA . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
1.8 TDMA . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
1.9 CDMA . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
1.10 OFDM . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
1.11 OFDMA . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
1.12 Subband CAS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
1.13 Interleaved CAS . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
1.14 Generalized CAS . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
1.15 Cyclic prefix . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
1.16 OFDMA transmitter . . . . . . . . . . . . . . . . . . . . . . . . . 32
1.17 OFDMA spectrum . . . . . . . . . . . . . . . . . . . . . . . . . . 32
1.18 OFDMA uplink receiver . . . . . . . . . . . . . . . . . . . . . . . 36
1.19 OFDM-IDMA transmitter . . . . . . . . . . . . . . . . . . . . . . 38
1.20 OFDM-IDMA uplink receiver . . . . . . . . . . . . . . . . . . . . 42

2.1 OFDMA spectrum when using subband CAS . . . . . . . . . . . . 48

2.2 Single-user detector . . . . . . . . . . . . . . . . . . . . . . . . . 50
2.3 CFO correction by circular convolution . . . . . . . . . . . . . . . 50
2.4 CFO estimation performance for a joint CFO/channel estimation 60
2.5 Channel estimation performance for a joint CFO/channel estimation 60

ix
LIST OF FIGURES

2.6 Recursive CFO estimation for a joint CFO/channel estimation us-

ing optimal filtering . . . . . . . . . . . . . . . . . . . . . . . . . . 61
2.7 Comparison between the Kalman filtering based approaches and
the H∞ filtering based approaches in terms of convergence speed . 62
2.8 Example of an OFDMA frame composed of M OFDMA symbols
with a preamble composed of two OFDMA symbols and a deter-
mined number of pilot subcarriers . . . . . . . . . . . . . . . . . . 66
2.9 Example of an OFDMA frame composed of M OFDMA symbols
with a preamble composed of two OFDMA symbols and no pilot
subcarriers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67
2.10 Proposed OFDMA non-pilot aided iterative receiver using optimal
filtering . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68
2.11 Iterative MMSE-SD using a Kalman filtering based estimator and
a MAI suppression for BPSK modulation . . . . . . . . . . . . . . 69
2.12 CDKF based approach for CFO estimation with a known preamble 73
2.13 MMSE-SD combined with an UKF based approach for CFO esti-
mation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 74
2.14 BER performance when a non-pilot aided estimation is considered 75
2.15 CFO estimation performance for a non-pilot aided CFO estimation 76
2.16 Comparison between the SPKF and the “unscented H∞ filter” in
terms of convergence speed . . . . . . . . . . . . . . . . . . . . . . 77
2.17 BER performance when the channel is estimated and a non-pilot
aided CFO estimation is considered . . . . . . . . . . . . . . . . . 78
2.18 PU and CR system spectrum . . . . . . . . . . . . . . . . . . . . 81
2.19 Time representation of the PU and CR-NBI signals . . . . . . . . 85
2.20 Time-frequency representation of PU-OFDMA received symbol . . 85
2.21 Time representation of the innovation energy . . . . . . . . . . . . 87
2.22 SPKF-BHT algorithm . . . . . . . . . . . . . . . . . . . . . . . . 89
2.23 Time representation of the CUSUM test . . . . . . . . . . . . . . 90
+ +
2.24 Cmean (p + 1) and C (n) . . . . . . . . . . . . . . . . . . . . . . . 92
2.25 CR-NBI detection performance . . . . . . . . . . . . . . . . . . . 93
2.26 PU detection performance . . . . . . . . . . . . . . . . . . . . . . 94
2.27 CR-NBI detection performance for different values of Pf a . . . . . 96

x
 LIST OF FIGURES

2.28 PU detection performance for different values of Pf a . . . . . . . . 97

2.29 Robust CFO estimation performance for several durations of the
CR-NBI . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 97
2.30 Robust CFO estimation performance for different values of SNR . 98
2.31 Robust channel estimation performance for different values of SNR 99

3.1 Proposed OFDM-IDMA uplink receiver . . . . . . . . . . . . . . . 110

3.2 SPKF convergence speed for an OFDM-IDMA system . . . . . . . 112
3.3 CFO estimation performance of the OFDM-IDMA modified receiver113
3.4 BER performance of the OFDM-IDMA modified receiver 1 . . . . 114
3.5 BER performance of the OFDM-IDMA modified receiver 2 . . . . 114
3.6 OFDM-IDMA-STBC transmitter . . . . . . . . . . . . . . . . . . 116
3.7 Proposed STBC-OFDM-IDMA uplink receiver . . . . . . . . . . . 122
3.8 CFO estimation performance of the STBC-OFDM-IDMA modified
receiver . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 123
3.9 BER performance of the STBC-OFDM-IDMA receiver 1 . . . . . 124
3.10 BER performance of the STBC-OFDM-IDMA receiver 2 . . . . . 124

A.1 Kalman filter - system model . . . . . . . . . . . . . . . . . . . . 132

B.1 Transfer operator from disturbances to estimation error for H∞ -

norm based estimation . . . . . . . . . . . . . . . . . . . . . . . . 136

D.1 EKF . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 145

D.2 IEKF . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 149

E.1 Gaussian distribution approximated by the sigma-points . . . . . 151

E.2 Non-linear transformation for a random vector of size 2 . . . . . . 152

xi
List of Tables

1.1 Fading channel classification . . . . . . . . . . . . . . . . . . . . . 18

1.2 Mobile standards evolution from 2G to 4G. . . . . . . . . . . . . . 23
1.3 Physical layer parameters of IEEE 802.16 Wireless MAN . . . . . 32
1.4 CFO values for different values of LO tolerance . . . . . . . . . . 33
1.5 CFO values for different relative speeds . . . . . . . . . . . . . . . 34

2.1 Gaussian non-linear estimation methods . . . . . . . . . . . . . . 57

2.2 KF vs H∞ filter in a non-linear case . . . . . . . . . . . . . . . . . 58
2.3 BER performance when a joint channel/CFO estimation is consid-
ered . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62
2.4 Number of arithmetic operations performed by the EM joint chan-
nel/CFO estimator . . . . . . . . . . . . . . . . . . . . . . . . . . 63
2.5 Number of arithmetic operations performed by the SPKF joint
channel/CFO estimator . . . . . . . . . . . . . . . . . . . . . . . . 64
2.6 Number of Go/s performed by the EM joint channel/CFO estima-
tor for different grid search precisions . . . . . . . . . . . . . . . . 65
2.7 Number of Go/s performed by the EM and the SPKF joint chan-
nel/CFO estimators . . . . . . . . . . . . . . . . . . . . . . . . . . 65
2.8 PU system parameters for the uth user . . . . . . . . . . . . . . . 82
2.9 CR system parameters . . . . . . . . . . . . . . . . . . . . . . . . 83
2.10 Relations between PU and CR system parameters . . . . . . . . . 83
2.11 SPKF-BHT CFO estimation performance for several durations of
the CR-NBI and different values of β . . . . . . . . . . . . . . . . 95
2.12 SPKF-CT CFO estimation performance for several durations of
the CR-NBI and different values of β . . . . . . . . . . . . . . . . 95
2.13 Robust CFO estimation performance for different values of SIR . 98

xiii
Notations

A OFDMA symbol vector after the convolution with the channel

B Additive white Gaussian noise vector
Eu CFO matrix of the uth user
ǫu Normalized CFO of the uth user
ǫ̂u Estimation of ǫu
ǫpu
u Normalized CFO of the uth user in the PU-OFDMA system
ǫ̂pu
u Estimation of ǫpu
u
F Fourier matrix
FH Inverse Fourier matrix
hu CSI vector of the uth user
ĥu Estimation of hu
hupu CSI vector of the uth user in the PU-OFDMA system
ĥupu Estimation of hupu
Ix Identity matrix of size x × x
K Number of subcarriers
K pu Number of subcarriers of the PU-OFDMA system
K cr Number of subcarriers of the CR system
K(n) Kalman gain at time n
K∞ (n) H∞ gain at time n
L Number of paths of the channel
Lu Number of paths of the channel associated to the uth user
Lpuu Number of paths of the channel associated to the uth user
of the PU-OFDMA system
P(n|n − 1) Covariance matrix of the a priori estimation error, when
using Kalman filtering
P(n|n) Covariance matrix of the a posteriori estimation error, when
using Kalman filtering
P∞ (n|n) Recursive estimate of the Ricatti equation solution in a
H∞ setting

xv
Notations

Pf a Probability of false alarm

R(n) nth sample of the received OFDMA or OFDM-IDMA symbol
R(mpu , n) nth sample of the received mpu th PU-OFDMA symbol
S Symbol (M-PSK symbol) vector
T OFDMA symbol time
Ts Symbol time
Tsamp Sampling time, Tsamp = Ts in the OFDM case
Tc Channel coherence time
X OFDMA or OFDM-IDMA modulated symbol vector
x(n) State vector containing the channels and the CFOs
W Transmitted-signal bandwidth
Wc Channel coherence bandwidth
W pu PU-OFDMA transmitted signal bandwidth
W cr CR transmitted signal bandwidth
Y(n, x(n)) Observation vector, storing the real and the imaginary parts
of R(n) or R(mpu , n)
Isd Number of iterations when using the OFDMA non-pilot
receiver
Iidma Number of iterations when using the OFDM-IDMA
receiver
Ξ2 Prescribed noise attenuation level
β Factor improving the detection of the disturbance
Pb {.} Probability distribution of {.}
E{.} Expectation of {.}
Var{.} Variance of {.}
∝ Proportional to
diag{.} Diagonal matrix with entries in the main diagonal of {.}
Im{.} Imaginary part of {.}
Re{.} Real part of {.}
{.}H Hermitian matrix of {.}
{.}T Transpose matrix of {.}
{.}∗ Conjugate of {.}
tr{.} Trace of {.}
||.|| Euclidean norm of {.}
⌊.⌋ Integer part of {.}

xvi
Introduction
For more than 25 years, a great deal of interest has been paid to mobile communi-
cation systems and the design of new schemes to transmit and receive information.
Engineers and researchers have worked much and exchanged on that topic.
Several PhD in the signal group at the UMR CNRS 5218 IMS have dealt with
channel estimation and symbol detection in multicarrier systems involving sev-
eral users. In [Jamo 07b], the channel was assumed to be an autoregressive (AR)
process. Jamoos suggested estimating both the channel and its AR parameters
by using training sequences and optimal filtering. To avoid an approach dedi-
cated to the non-linear estimation issue, mutually interactive optimal filters were
used. One filter aimed at estimating the channel, whereas the other updated the
estimation of the AR parameters. H∞ and Kalman filtering were tested and the
resulting approaches were studied in a multicarrier direct-sequence code division
multiple access system (MC-DS-CDMA). In [Grol 07], the channel was modeled
by a sinusoidal stochastic process or a low-pass filter version of an AR process.
Then, both the symbols and the Rayleigh fading channel were jointly estimated in
MC-DS-CDMA systems by using a Rao-Blackwellized particle filter cross-coupled
with a Kalman filter.
Despite the great success of the CDMA as a multiple access technique, the cur-
rent trend is to use the orthogonal frequency division multiple access (OFDMA)
and the orthogonal frequency division multiplexing-interleaved division multi-
ple access (OFDM-IDMA). In that case, the input data stream is split into a
number of streams that are transmitted in parallel over a large number of orthog-
onal subcarriers. The frequency-selective fading over the entire bandwidth of the
transmitted signal has hence the advantage of being converted in frequency-flat
fading over each subcarrier. It should be noted that these schemes are partic-
ularly well adapted to mobile wireless communication to provide high-data rate
services, e.g. [3GPP 09], [IEEE 06].
However, there are two unwanted phenomena:
1/ the channel remains unknown and hence has to be estimated.

1
Introduction

2/ The relative transmitter-receiver motion and a difference between the local-

oscillator (LO) frequencies at the transmitter and the receiver lead to a carrier
frequency offset (CFO). These CFOs that affect the received signal no longer
guarantee orthogonality between subcarriers. To avoid the resulting intercarrier
interference (ICI), a CFO estimation/correction step must be introduced at the
receiver.
In the literature, some authors focus their attentions on the CFO estimation
[More 04], [Zhao 06], [Jing 08], while others address the joint estimation of the
CFOs and the channels. More specifically, a conventional expectation-maxi-
mization (EM) algorithm is proposed in [Pun 04a] for an OFDMA uplink sys-
tem. Then, to reduce the computational cost of the EM algorithm, the authors
use the space alternating generalized expectation-maximization (SAGE) algo-
rithm. Instead of addressing a multidimensional optimization issue, the so-called
alternating projection estimator is used in [Pun 06]. This method consists in iter-
atively estimating the CFO of one user, by means of an exhaustive grid search over
the possible range of the CFO value and by setting the other CFOs to their last
updated values. In [Fu 06], Fu et al. propose two iterative estimation approaches
using the SAGE method. In [Sezg 08], Sezginer et al. propose an iterative sub-
optimal method. It is based on an approximation of a maximum likelihood (ML)
estimator to reduce the computational complexity of the EM-based algorithms.
Nevertheless, an initialization step is required in the above algorithms. Another
drawback of those methods is the high computational cost due to the iterative
estimation and the exhaustive grid search.

In this PhD dissertation, our contribution is twofold:

1/ from the estimation point of view: as the joint estimation of the CFO and
the channel is a non-linear estimation problem, we propose to use variants of
the Kalman filter and H∞ filter. Based on a comparative study, we evaluate the
performance of the approaches in two uplink multicarrier contexts: the OFDMA
and the OFDM-IDMA systems. Various criteria are taken into account, such as
the minimum mean square error (MMSE) on the CFO and the channel estimates,
and the computational cost.

2
 Introduction

2/ From the digital-communication system point of view: we propose to combine

the above estimation methods with:

• the minimum mean square error successive detector (MMSE-SD) [Hou 08]
to increase the data rate for an OFDMA system [Pove 09a], [Pove 09b],
[Pove 11c].

• a statistical test, e.g. the binary hypothesis test (BHT) or the cumulative
sum (CUSUM) test, to detect the beginning and the end of the disturbances
induced by another system [Pove 11b], [Pove 11d]. This happens with cog-
nitive radio (CR) systems where the spectrum that is not used by licensed
(primary) users must be detected in order to be shared with unlicensed
(secondary) users1 .

• a modified version of the iterative IDMA receiver [Ping 02a] to design a new
receiver scheme for an OFDM-IDMA system [Pove 12].

• a space time block code (STBC) and a new variant of the iterative IDMA
receiver [Pove 11a].

The PhD dissertation is organized as follows:

In the first chapter, we present different multiuser scenarios and then focus our
attention on those based on a multicarrier modulation. Thus, we look at two mo-
bile wireless communication systems: the conventional ones [IEEE 99], [IEEE 06]
and the cognitive radio (CR) systems [Mito 00]. The brief presentation of these
wireless systems allows us to introduce both the multipath propagation channel,
its statistical properties and the Doppler shifts. We also define the received faded
signal that is a superposition of several delayed and attenuated copies of the
transmitted signal.
Then, we present the multicarrier multiple access schemes that can be used in
these systems. Among them, we give details about multicarrier multiple ac-
cess schemes. We first present the OFDMA, which results from the combina-
tion between the frequency division multiple access (FDMA) and the OFDM.
1
In that case, physical-layer parameters are dynamically adjusted to give the appropriate
bandwidth to each user at any time.

3
Introduction

Then, we look at the OFDM-IDMA scheme initially proposed by Mahafeno et al.

[Maha 06].
Whatever the multicarrier systems, the transmitted signals are subject to a CFO
and the multipath propagation channel. Therefore, we have to design new re-
ceivers taking into account both unwanted phenomena.

In chapter 2, we propose and compare several CFO and channel estimation meth-
ods based on optimal filtering techniques [Pove 10]. As the relation between the
CFO and the received signal is non-linear, a linearized version of the Least-
Squares (LS) approach can be first used to estimate the CFO. More generally, we
suggest using an extended Kalman filter (EKF). The EKF consists in analytically
propagating the estimation through the system dynamics, by means of a first-
order Taylor expansion of the functions defining the state-space representation
of the system. However, as the approximation may not be sufficient to describe
the non-linearity, the EKF may sometimes diverge. To solve this problem, a
second-order linearization can be considered and leads to the second-order EKF
(SOEKF) [Bar 01]. Another solution is to use the iterative extended Kalman
filter (IEKF). In the IEKF, the measurement model is linearized around the up-
dated state vector, instead of the predicted state vector. Then, the process is
iterated until the state vector estimate does not change much.
As the above approaches require the computation of the Jacobians and the Hes-
sians matrices for the first-order and the second-order linearizations respectively,
we also look at the sigma-point Kalman filter (SPKF) [VdMe 04a], namely the
unscented Kalman Filter (UKF) and the central difference Kalman filter (CDKF).
In that case, the state distribution is approximated by a Gaussian distribution,
which is characterized by the so-called sigma points. Then, the sigma points are
propagated through the non-linear system. A weighted combination of the re-
sulting values makes it possible to estimate the mean and the covariance matrix
of the transformed random variable. The UKF is based on the unscented trans-
formation whereas the CDKF is based on the second-order Sterling polynomial
interpolation formula.
Nevertheless, Kalman algorithms require an accurate system model as well as
perfect knowledge of the noise statistics. Therefore, we also propose to evaluate

4
 Introduction

the performance of the H∞ filtering, which has been very popular in the field of
control and is more and more used in signal processing (see for instance [Shen 99],
[Laba 05] and [Gire 09]). This approach is designed to be robust against uncer-
tainties. In addition, no Gaussian assumption on the additive noise and the
model noise in the state-space representation of the system is required. Here, we
analyze the relevance of the “extended H∞ filter”, based on Taylor expansion like
the EKF and the “unscented H∞ filter” in which an unscented transformation is
embedded into the “extended H∞ filter” [Li 10].
Then, the above techniques are combined with different processing blocks to de-
sign new receiver schemes:
Firstly, the optimal filtering techniques are combined with the so-called
MMSE-SD [Hou 08] in an iterative way to obtain a CFO estimator that does
not require pilot subcarriers [Pove 09a], [Pove 09b], [Pove 11c].
Secondly, we use the SPKF in a dynamic bandwidth allocation context. The
SPKF is combined with a statistical test, either the BHT or the CUSUM test.
This approach allows the CFOs and the channels to be jointly estimated and the
beginning and the end of a narrow-band interference (NBI) produced by a CR
system to be detected [Pove 11b], [Pove 11d].

In chapter 3, we address the CFO and channel estimation problems in an OFDM-

IDMA context. Although this access technique has the potential to solve the
constant growth on the user density per cell, CFO and channel estimations have
not been widely addressed. The IDMA conventional receiver consists of an el-
ementary signal estimator (ESE) and several a posteriori probability decoders
(DECs) for the different users in the system. The ESE operation is based on
the constraint of the a priori knowledge of the multipath channels. However,
few papers deal with the channel estimation in IDMA systems. The reader may
refer to [Suya 08] and [Rehm 08] for channel estimation in OFDM-IDMA sys-
tems, where the system is assumed to be perfectly frequency synchronized. To
our knowledge, the frequency synchronization issue has never been addressed in
an OFDM-IDMA system.
Thus, we propose a new scheme, which operates in two steps. Firstly, CFOs
and channels are estimated for each user in the system, by using control data

5
Introduction

of one OFDM-IDMA symbol and a SPKF. Secondly, the OFDM demodulation

is performed without any CFO correction on the received signal [Pove 12]. The
resulting signal is inserted in a modified version of the iterative IDMA receiver,
initially proposed by [Ping 02a]. The second part of the chapter 3 deals with a
STBC-OFDM-IDMA receiver. In that case, we show how to modify the IDMA
iterative receiver by taking advantage of the spatial diversity introduced by the
multiple inputs [Pove 11a].

In each case, simulation results confirm the efficiency of the proposed architec-
tures in terms of CFO estimation, channel estimation, computational complexity
and BER performance.

Finally, conclusions and perspectives are given.

To help the reader, we also present six appendices:

• In the first one, we recall the main results about the Kalman filter in the
linear case.

• In the second appendix, the H∞ filtering in the linear case is presented.

• In the third one, we present a theoretical comparison between the Kalman

and the H∞ filters.

• In the fourth appendix, we introduce the EKF, the SOEKF and the IEKF.

• The fifth appendix is about the SPKF.

• Finally, the last appendix deals with the H∞ filtering for non-linear cases.
The “extended H∞ filter” and the “unscented H∞ filter” are presented.

6
Chapter 1
Multicarrier High-Data Rate Mobile
Wireless Systems

Contents

1.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . 8
1.2 Wireless Communication Systems . . . . . . . . . . . 8
1.2.1 Conventional Systems . . . . . . . . . . . . . . . . . . 9
1.2.2 Cognitive Radio Systems . . . . . . . . . . . . . . . . 9
1.3 Propagation Channel Model . . . . . . . . . . . . . . . 11
1.3.1 Rician and Rayleigh Fading Channels . . . . . . . . . 13
1.3.2 Frequency-Selective and Time-Varying Fading Channels 15
1.4 Multiple Access Techniques for Mobile Systems . . . 19
1.4.1 About Multicarrier Multiple Access Techniques . . . . 22
1.4.2 OFDMA System . . . . . . . . . . . . . . . . . . . . . 27
1.4.3 OFDM-IDMA System . . . . . . . . . . . . . . . . . . 36
1.5 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . 43

7
1. Multicarrier High-Data Rate Mobile Wireless Systems

1.1 Introduction
The purpose of this chapter is first to give general information about wireless
communication systems. Then, we present the unwanted phenomena that may
appear when transmitting wireless information. More particularly, the channel
plays a key role. Its statistical properties may also vary and depend on the exis-
tence of a line-of-sight (LoS) between the transmitter and the receiver. Notions of
frequency-selective fading or flat-fading channels are recalled. In the second part
of this chapter, we introduce the various access schemes. After recalling what
FDMA, time division multiple access (TDMA) and code division multiple access
(CDMA) are all about, OFDMA and OFDM-IDMA are presented.

1.2 Wireless Communication Systems

A great deal of interest has been paid to wireless communications for more than
hundred years. At the end of the 19th and the beginning of the 20th century,
Marconi was at the origin of the first radio transmission [Hanz 98], and wireless
communications began.
During the second half of the 20th century, successive modifications in wireless
communications were performed, and correspond to the development of the data
transmission over telephone networks, the creation of wireless local area networks
(WLANs) and the evolution of mobile wireless communications systems.
Today, these transformations in the field of wireless communications still con-
tinue. Conventional wireless systems, (e.g. WLANs or cellular networks) that
are usually based on a static frequency allocation are evolving to an intelligent
concept, where a dynamic frequency allocation can be allowed. Intelligent wireless
systems known as cognitive radio (CR) systems [Mito 99] are solutions proposed
to improve wireless communications.
This PhD dissertation thesis is devoted to two kinds of wireless communication
systems: conventional systems and CR systems. In the following subsections,
both types of systems are defined.

8
 1.2 Wireless Communication Systems

1.2.1 Conventional Systems

A conventional system corresponds to a current wireless system. Common ex-
amples of these systems can be WLANs, digital video broadcasting terrestrial
(DVB-T) or 3G cellular networks. For more details on the conventional wireless
systems, the reader is referred to [Rapp 02] and [Stal 05].
These types of systems use a static and predetermined frequency band allocation.
Moreover, each terminal is programmed to a limited number of tasks.
Even if conventional systems provide high-data rate services, the exponential in-
crease of data rates and the number of users have produced an overload of the
frequency spectrum. For instance, in December 2008, there were 4 billion of mo-
bile subscriptions [ITU 08] worldwide. In 2013, the number of 4G subscribers
worldwide is expected to exceed 90 million. To keep on providing high-data rate
services, developers and researchers have to find alternative technologies. The
next subsection deals with CR systems. This is a technology proposed as a solu-
tion to correct this frequency spectrum overload.

1.2.2 Cognitive Radio Systems

Mitola was the first one to employ the concept of CR [Mito 99] and defined it
as: “a radio that employs model-based reasoning to achieve a specified level of
competence in radio-related domains”.
CR systems have emerged as a new technology to improve the utilization of the
limited radio bandwidth. The key features of those systems are their awareness
and intelligence, which are achieved through learning via a cognition cycle (ob-
serve, decide and act). This intelligence allows CR systems to tune the system
parameters such as power, carrier frequency, and modulation at the physical layer,
and higher-layer protocol parameters to improve their utilization.
Nowadays, there is an overload of the frequency spectrum assigned to the con-
ventional systems. However, at the same time there is an underutilization of this
assigned frequency spectrum [Mito 00]. CR may be a solution to this issue. It can
be achieved by a spectrum shared by primary and secondary users. The sharing
techniques can be classified into underlay and overlay spectrum sharing. On the
one hand, when using underlay systems, the primary and secondary users share

9
1. Multicarrier High-Data Rate Mobile Wireless Systems

Figure 1.1: Cognition cycle

the same frequency spectrum. However, secondary users should transmit the sig-
nal with a low power that does not exceed an interference threshold [Pali 10]. On
the other hand, in overlay systems the secondary users have to find out the bands
that are not used by the primary users (PUs). Then, secondary users can utilize
these unused portions of the spectrum. For this purpose, the secondary users
need information about the spectrum allocation of the PUs by regularly perform-
ing spectrum sensing techniques. Various spectrum sensing techniques have been
proposed [Arsl 07]. Some of them aim at identifying the characteristics of the
transmission whereas others distinguish the signal type. The most common are
the matched filter, the cyclostationary feature detection and the energy detection
[Arsl 07]. For more details about CR systems, the reader is referred to [Pali 10].
Understanding that CR systems have the potential to exploit the underutilized
conventional system frequency bands by spectrum sharing, the Federal Commu-
nications Commission (FCC) released a second report in which unlicensed devices
are allowed to operate in the unused portions of the TV-spectrum in 2008.
CR systems will be considered in this PhD dissertation, in chapter 2, section 2.5.
In the following section, we present one of the more challenging tasks when de-
signing a wireless system, taking into account the influence of the propagation
channel.

10
 1.3 Propagation Channel Model

1.3 Propagation Channel Model

In mobile wireless communication systems, the transmitted signal arrives at the
receiver from different propagation paths. Such phenomenon occurs due to the
obstacles such as buildings, mountains, trees, etc., between the transmitter and
the receiver, as shown in figure 1.2.

Figure 1.2: Multipath propagation channel

Due to these obstacles, the transmitted signal is subject to:

• reflection: it arises when the plane waves are incident upon a surface with
dimensions that are very large compared to the wavelength.

• diffraction: it occurs when there is an obstruction between the transmitter

and receiver antennas. Secondary waves are then generated behind the
obstruction.

• scattering: it happens when the plane waves are incident upon an object,
the dimensions of which are of the order of a wavelength or less, and causes
the energy to be redirected in many directions.

Given these three mechanisms, a wireless propagation can be roughly character-

ized by some independent phenomena: path loss variation, shadowing and mul-
tipath fading. Mathematically, the path loss is only distance dependent, whereas

11
1. Multicarrier High-Data Rate Mobile Wireless Systems

the other two phenomena are statistically described. In the following, let us focus
our attention on the multipath effect.
The received signal is a superposition of several delayed and attenuated copies
of the transmitted signal. By considering a stationary propagation, the channel
propagation model h(τ ) at time τ can be mathematically expressed as:
L−1
X
h(τ ) = A(l)δ(τ − τl ) (1.1)
l=0

where L is defined as the number of paths, A(l) and τl = lTsamp are the amplitude
and the time delay associated to the lth path, respectively. In addition, Tsamp is
the sampling time.
Thus, the coefficients of the discrete-time channel impulse response (CIR) can be
stored in the following vector:

h = [h(0), h(1), . . . , h(l), . . . , h(L − 1)] (1.2)

Figure 1.3 shows an example of a CIR, where τmax represents the maximum chan-
nel delay spread in seconds.
h(τ )

0 1 2 3 4 5 6 7 8
Tsamp τ
τmax

Figure 1.3: Channel impulse response with L = 9

12
 1.3 Propagation Channel Model

The fading that modifies the transmitted signal depends on the characteristics
of the channel and the nature of the transmitted signal. Moreover, the statis-
tical model of the channel is slightly different if the transmitter is in LoS with
the receiver. The multipath fading channel can be hence classified by looking at
its probability density distribution and its frequency response in the transmitted
signal bandwidth.
These classifications of fading are investigated in the next subsections.

1.3.1 Rician and Rayleigh Fading Channels

Due to the existence of a great variety of propagation environments, several sta-
tistical distributions have been proposed for channel modeling. However, in the
following let us consider the two most commonly used distributions: the Rayleigh
and Rician distributions.
In wireless systems, a predominant component of the transmitted signal is some-
times present at the receiver. If this predominant component can be the LoS wave
for instance, each coefficient of the CIR follows a Rician distribution [Jake 74].
Let us look at this case more carefully. When the received signal consists of a
large number of plane waves with different phases, it can be treated as a complex
Gaussian random process αr (n) = αI (n) + jαQ (n), where αI (n) and αQ (n) are
Gaussian random variables (GRVs) with non-zero means µI and µQ , respectively.
The processes are assumed to be uncorrelated and the GRVs have the same vari-
ance σr2 . Then, the magnitude of the received signal has the following Rician
distribution: !
2 2
x − x +µr xµr
Pdf (x) = 2 e 2σr2 I0 x≥0 (1.3)
σr σr2
 
xµr
where µ2r = µ2I + µ2Q is called the non-centrality parameter and I0 σr2
is the
zero-order modified Bessel function1 of the first kind [Stub 02].
1
The zero-order modified Bessel function of the first kind is defined as:
Z 2π
∆ 1
I0 (y) = e−ycos(z) dz
2π 0

13
1. Multicarrier High-Data Rate Mobile Wireless Systems

The received signal distribution can be rewritten by using the Rice factor defined
by:
Krice = µ2r /2σr2 (1.4)

and the average envelope power

E{|αr (t)|2 } = Ωr = µ2r + 2σr2 (1.5)

Therefore, the distribution can be expressed as follows1 :

 s 
2x(Krice + 1) −Krice − (KriceΩ +1)x2  Krice (Krice + 1) 
Pdf (x) = e r I0 2x x≥0
Ωr Ωr
(1.6)
Figure 1.4 shows examples of Rician distributions for differents values of Krice .
It should be noted that µ2r characterizes the LoS wave. Therefore, if there is non
LoS wave, µ2r = 0, i.e. Krice = 0. This is the case of a Rayleigh model.

Krice = 0
Krice = 3
1.8 Krice = 7
Krice = 10
1.6

1.4

1.2
Pdf (x)

0.8

0.6

0.4

0.2

0
0 0.5 1 1.5 2 2.5 3 3.5 4
x

Figure 1.4: Rician distribution

Rayleigh channel model is the most common assumption in wireless systems. In

this case, propagation path consists of a two dimensional isotropic scattering. The
1 Krice Ωr Ωr
It should be noted that: µ2r = Krice +1 and σr2 = 2(Krice +1)

14
 1.3 Propagation Channel Model

plane waves arrive from many directions with equal probability without a direct
LoS component. Then, each coefficient of the CIR follows a Rayleigh distribution
[Jake 74]. Therefore, given (1.3) and under these assumptions, the magnitude of
the received signal has the following Rayleigh distribution:
2x − Ωx2
Pdf (x) = e r x≥0 (1.7)
Ωr

where Ωr = E{|αr (t)|2 } = 2σr2 is the average envelope power [Stub 02].
Figure 1.5 shows examples of Rayleigh distribution for different values of Ωr .
In the following subsection, we present the frequency-selective and the time-
varying channels.

Ωr = 0.5
Ωr =1
0.8 Ωr =2
Ωr =4

0.7

0.6

0.5
Pdf (x)

0.4

0.3

0.2

0.1

0
0 1 2 3 4 5 6 7
x

Figure 1.5: Rayleigh distribution

1.3.2 Frequency-Selective and Time-Varying Fading Chan-

nels
The fading model can also be characterized by the nature of the transmitted
signal and the relative speed between the transmitter and the receiver. It can
be classified by the relationship between the transmitted signal bandwidth W ,

15
1. Multicarrier High-Data Rate Mobile Wireless Systems

that is proportional to the inverse of the symbol1 (M-PSK symbol) time Ts ,

1
(i.e. W ∝ Ts
) and the channel frequency response in this bandwidth.
Let us define the channel coherence bandwidth Wc as:
1
Wc = (1.8)
τmax
On the one hand, if all the multipaths arrive at the receiver within the symbol
duration, the fading channel is considered as a frequency non-selective fading or
flat-fading channel. Then, the channel coherence bandwidth is higher than the
transmitted signal bandwidth:
W << Wc (1.9)

i.e.
τmax << Ts (1.10)

At the receiver, without the noise contribution, the signal is an attenuated copy
of the transmitted signal:
r(n) = hs(n) (1.11)

where s(n) is the received symbol, h is the CIR coefficient and r(n) is the received
symbol.
On the other hand, if the multipaths are spread outside the symbol duration, i.e.
the maximum channel delay spread is higher than the symbol time, the fading
channel is considered as a frequency-selective fading channel. In this case, the
transmitted signal bandwidth is higher than the channel coherence bandwidth:

W >> Wc (1.12)

i.e.
τmax >> Ts (1.13)

Therefore, without the noise contribution, the received signal is a superposition

of several transmitted attenuated and delayed signals. Mathematically, it can be
1
In this PhD dissertation we consider M-ary phase shift keying (M-PSK) as the modulation
technique. It consists in changing the phase of the carrier frequency, where M is the number
of possible phases. For example, when using BPSK modulation the set of symbols is {−1, +1}
π 3π 5π 7π
and when using QPSK the set of symbols can be {ej 4 , ej 4 , ej 4 , ej 4 }.

16
 1.3 Propagation Channel Model

expressed as the convolution between the transmitted signal and the CIR:
L
X
r(n) = h(l)s(n − l) (1.14)
l=0

The received signal is spread in time and this leads to the so-called inter sym-
bol interference (ISI). Figure (1.6) shows the frequency response of a frequency-
selective fading channel.
In both cases, estimating and correcting the CIR, also called channel equaliza-
tion, is usually performed using a known signal at the receiver, called pilot signal.
Given (1.11) and (1.14), we can expect a higher complexity of the equalization
when dealing with frequency-selective channel.

−10
Power (dB)

frequency response of
the transmitted signal

−20

−30
−0.5 −0.25 0 0.25 0.5
Normalized frequency

Figure 1.6: Frequency-selective fading channel

Furthermore, there is sometimes a relative motion between the transmitter and

the receiver in wireless systems. This results in a time-varying channel. The
channel can be divided in two categories depending on the ratio between the
transmitted symbol time Ts and the channel coherence time Tc .
Let us define the channel coherence time Tc as follows:
1 c
Tc = DOP = (1.15)
f vfc

17
1. Multicarrier High-Data Rate Mobile Wireless Systems

where the Doppler shift f DOP = v fcc is the maximum measure in Hertz of a relative
frequency shift between the transmitted and the received signal, v is the relative
speed between the transmitter and the receiver, fc is the frequency carrier and c
is the speed of the light.
The Doppler shift is caused not only by the transmitter-receiver relative motion,
but also by the movement of surrounding objects. It leads to frequency offsets at
the receiver.
If the channel coherence time is higher than the transmitted symbol time:

Ts < Tc (1.16)

the channel is considered as a slow-fading channel.

If the transmitted symbol time is higher than channel coherence time:

Ts > Tc (1.17)

the channel is considered a fast-fading channel.

Table 1.1 shows a classification of the fading channels according to the nature
of the transmitted signal and the relative speed between the transmitter and the
receiver

slow-fading channel fast-fading channel

τmax << Ts τmax << Ts
flat-fading channel
Ts < Tc Ts > Tc
τmax >> Ts τmax >> Ts
frequency-selective channel
Ts < Tc Ts > Tc

Table 1.1: Fading channel classification

In both cases, at the receiver, the frequency offset caused by the Doppler shift
has to be estimated in order to recover the original transmitted signal.

The design of a mobile wireless system implies the above technical challenges.
In the following, we present some schemes that have been proposed during the
last years to take into account the impact of the fading effects.

18
 1.4 Multiple Access Techniques for Mobile Systems

1.4 Multiple Access Techniques for Mobile Sys-

tems
In mobile wireless systems, a big challenge is to choose the multiple access tech-
nology that will efficiently share the available bandwidth among a large number
of users and that will be robust against the channel propagation effects. For the
past decades, the mobile wireless communication industry has searched for dif-
ferent techniques to allocate the communication resources to the different users.
The first public cellular radio system, known as advanced mobile phone service
(AMPS), was introduced at the end of the ’70s in the United States, shortly
followed by the Nordic mobile telephone system and the total access communica-
tion system (TACS) in Europe. At the same time, the Japaneses introduced the
Nippon automatic mobile telephone system (NAMTS) [Hanz 98]. These systems
were analog and are known as the 1st generation mobile wireless systems (1G).
All of them used FDMA as their multiple access scheme. In FDMA, the allocated
spectrum is divided into several frequency bands where each band is assigned to
one user, i.e. each user can communicate at the same time. See figure 1.7. Mul-
tiple users using separate frequency bands can access system on the same time
without significant interference from other users simultaneously operating in the
system [Stub 02].

Power

user 1 user 2

Frequency
FD
M
A
Fr
ma
e
du
ra
tio
n

User 1
Time bandwidth

Available bandwidth

Figure 1.7: FDMA

19
1. Multicarrier High-Data Rate Mobile Wireless Systems

At the beginning of the ’90s, the 2nd generation mobile wireless systems (2G)
were developed, such as the digital AMPS in the United States, the global sys-
tem for mobile communications (GSM) in Europe and the personal digital cellular
(PDC) system in Japan [Hanz 98]. These systems employed TDMA as their mul-
tiple access scheme. When using TDMA, the whole bandwidth is assigned and
the time-domain transmission frame is divided into time slots, each assigned to
one user to transmit the data information [Stub 02]. See figure 1.8. TDMA is
used in the evolution of the 2G GSM standard, namely the general packet radio
service (GPRS) and the enhanced data rates for GSM Evolution (EDGE) sys-
tems.

Power
user 1 user 2

user 3 user U-1

user U

Frequency
U
TD

se
M

r 3
A

tim
Fr
am

e
sl
e

ot
du
ra
tio
n

Available
Time bandwidth

Figure 1.8: TDMA

Nevertheless in FDMA and TDMA, the number of frequency bands or time slots
is fixed for a given system, and one frequency band/time slot is assigned to one
user during the whole period of communications. This guarantees the service
quality for real-time and constant-bit-rate voice telephony. However, as the num-
ber of services is increasing from simple voice to multimedia traffic with different
requirements, fixed frequency band or time slot assignments have shown their
limitations, especially with the increasing number of users in the system. For
that reason, CDMA scheme, based on spread spectrum technology, has emerged.
In CDMA systems, the relatively narrow-band users information is spread into a
much wider spectrum using a high chip rate spreading code. When using different

20
 1.4 Multiple Access Techniques for Mobile Systems

codes, multiple user information can be transmitted on the same frequency band
at the same time. See figure 1.9. The spreading code of each user is orthogonal1
to the codes of all other users to minimize the multiple access interference (MAI)
produced by other users [Schu 05]. The 2G American system IS-95 was the first
mobile cellular communication system to use CDMA technology followed by the
CDMA 2000 technology [Stub 02].

user 1 user user U

Power

Frequency
C
D dur
M at
A io
Fr n
am
e

Available bandwidth

Time

Figure 1.9: CDMA

The digital 2G has shown higher transmission capacity and better voice quality
than the analog 1G. Like the 1G, 2G was primarily designed to support voice
communication. In the last releases of these standards, capabilities were intro-
duced to support data transmission. However, the data rates were generally lower
than those supported by the existent bandwidth. Thus, an initiative of the inter-
national telecommunication union (ITU) laid the way for evolution to 3rd genera-
tion mobile wireless systems (3G). Requirements such as a specific high-data rate
and support for vehicular mobility were established. Both the GSM and CDMA
camps formed their own separate 3G partnership projects, the 3GPP and 3GPP2,
respectively. Within the 3GPP evolution track, the 2G GSM/GPRS/EDGE fam-
ily is based on TDMA and FDMA, whereas the 2G IS-95/CDMA 2000 is based
on TDMA and CDMA in the 3GPP2. See table 1.2.
1
P∞
Two vectors x(n) and y(n) are orthogonal if their dot product hx(n), y(n)i= −∞ x(n)y(n)
is equal to zero.

21
1. Multicarrier High-Data Rate Mobile Wireless Systems

The 3G in the 3GPP was first referred to the universal mobile telecommunica-
tion system wideband CDMA (UMTS WCDMA). Then, it evolved in the high-
speed downlink and uplink packet access (HSDPA and HSUPA) enhancements
and in the high-speed packet access plus (HSPA+) enhancement. Meanwhile in
the 3GPP2, the 3G was known as the CDMA evolution-data optimized (CDMA
EVDO) [Sesi 09]. All these systems use CDMA as their multiple access tech-
nique. The number of orthogonal codes used for uplink transmission1 from one
user is limited to a few codes or complex techniques are used to limit the uplink
signal peakiness and to improve the low noise amplifier efficiency2 . In addition,
CDMA usually employs a rake receiver [Proa 95] technique to suppress multipath
effects and the related cost increases with the number of paths. Therefore, the
complexity of the receiver can be very high for a high-data rate mobile wireless
systems.

1.4.1 About Multicarrier Multiple Access Techniques

One solution to transmit the signal over a multipath frequency-selective and time-
varying fading channel, without ISI, is to choose a transmitted signal bandwidth
very higher than the Doppler shift and much lower than the channel coherence
bandwidth, fd << W << Wc , i.e τmax << Ts << Tc . See section 1.3.2.
This hypothesis is true for low data rate and low mobility systems. When a high
data rate system is considered usually, one has:

Ts << τmax << Tc (1.18)

To avoid the ISI, a solution could be to divide the entire bandwidth W in several
subbands of size ∆f << Wc and to send a large number of narrow-band signals
over several parallel subcarriers in the frequency band assigned to the transmis-
sion. This is the concept of a multicarrier modulation.
The preferred case of multicarrier modulation is the one that uses an orthogonal
1
In the uplink transmission, the transmitter is the user and the receiver is the base station.
2
For example, when using CDMA for an uplink transmission, a complex successive inter-
ference cancellation method is required to improve the performance of the receiver.

22
 1.4 Multiple Access Techniques for Mobile Systems

Peak Data Rates

Family Generation Standard Radio Access
Downlink (Mbits/s) Uplink (Mbits/s)
GSM 0.04 0.01
GPRS 0.17 0.13 FDMA,
2G
EDGE 0.47 0.36 TDMA
EDGEev 1.89 1.42
3GPP UMTS
0.38 0.38
WCMA
HSDPA 14.4 — CDMA,
3G
HSUPA — 5.76 OFDMA,
HSPA+ 42.2 11.5 SC-FDMA
4G LTE 300 50
LTE-A 1000 500
WiMAX
128 56
802.16e
3G
IEEE WiMAX OFDMA
300 135
802.16m
4G WiMAX 2 1000 500
IS-95 0.115 0.115
2G CDMA
0.307 0.307 TDMA,
3GPP2 2000
CDMA
CDMA
3G 73.5 27
EVDO

Table 1.2: Mobile standards evolution from 2G to 4G.

basis, namely the orthogonal frequency division multiplexing (OFDM)1 [Bing 90].
In this scheme, the input symbols are transmitted at the same time over orthogo-
nal subcarriers [Wein 71]. See figure 1.10. The main idea of OFDM is to convert a
frequency-selective channel in the time domain into a collection of frequency-flat
channels in the frequency domain.
OFDM increases robustness against multipath distortions, making the system ro-
bust against ISI. In addition, OFDM systems use a cyclic prefix (CP) to combat
interblock interference (IBI). The CP consists in prefixing the OFDM symbol
with the end of it2 . The above advantages allow the channel equalization to be
easily performed in the frequency and time domains through a bank of one-tap
multipliers [Cimi 85]. Furthermore, OFDM exploits the spectral diversity and
allows an independent selection over each subcarrier of resources, such as power,
constellation size and necessary bandwidth [Kell 00], in order to maximize the
1


Orthogonality between two frequencies fk , fk′ is defined as ej2πfk t , ej2πfk′ t = δk,k′ , where
δk,k′ = 0 if k = k ′ , and δk,k′ = 1 if k 6= k.
2
The CP is detailed in subsection 1.4.2.1.

23
1. Multicarrier High-Data Rate Mobile Wireless Systems

Power

pilots

user

Frequency
O
FD
M
O
FD

sy
m
M

bo
fra

ld
m

ur
e

at
du

io
ra

n
tio
n

Subcarrier spacing
Time

Figure 1.10: OFDM

link efficiency.
OFDM has been adopted in several communication standards such as digital
audio broadcasting (DAB) [ETSI 95], DVB-T [ETSI 97], and the WLAN IEEE
802.11a [IEEE 99].
The OFDM concept has been extended to multiuser communication scenarios. In
the following subsections three different multiple access schemes based on OFDM
are presented:

• OFDMA,

• single carrier frequency division multiple access (SC-FDMA),

• and OFDM-IDMA.

1.4.1.1 About OFDMA

This scheme was originally suggested for cable TV (CATV) networks [Sari 98]
and in the uplink communication of the Interaction Channel for Digital Terres-
trial Television (DVB-RCT) [ETSI 01]. The institute of electrical and electronics
engineers (IEEE) 3G project, called the mobile worldwide interoperability of mi-
crowave access (WiMAX) [IEEE 06] uses ODFMA technology. The last releases
of the 3GPP and 3GPP2, known as the long term evolution (LTE) and ultra

24
 1.4 Multiple Access Techniques for Mobile Systems

mobile broadband (UMB) respectively, are based on OFDMA1 . Note that LTE
uses OFDMA in the downlink communication2 [3GPP 09].
In this scheme, unlike the OFDM case where all subcarriers are assigned to a
single user, subcarriers are divided in several mutually exclusive subchannels and
they are exclusively assigned to a particular user in an OFDMA network. See
figure 1.11.
OFDMA inherits from OFDM the flexibility for simultaneous transmissions and
frequency allocation algorithms aim at exploiting the spectral diversity to allo-
cate the communication resources to the different users. In addition, OFDMA
inherits the ability to compensate channel distortions in the frequency domain
without computationally demanding time domain equalizers.
In subsection 1.4.2, more details about the OFDMA system are presented.

pilots
Power user 1

user 2

user 3

user 4

Frequency
O
FD
M
O
FD

A
sy
M

m
A

bo
fra

ld
m

ur
e

at
du

io
ra

n
tio
n

Subcarrier spacing
Time

Figure 1.11: OFDMA

1.4.1.2 About SC-FDMA

SC-FDMA is a modified form of OFDMA with similar throughput performance

and complexity. It is used as multiple access technique in the LTE uplink com-
1
The 3GPP2 announced it was ending development of the UMB technology, favoring LTE
instead.
2
In the downlink transmission, the transmitter is the base station and the receiver is the
user.

25
1. Multicarrier High-Data Rate Mobile Wireless Systems

munication [3GPP 09]. In this multiple access scheme, the symbols pass through
a discrete Fourier transform before going through the standard OFDMA mod-
ulation. This is often viewed as a DFT-coded OFDM. Thus, SC-FDMA in-
herits all the advantages of OFDMA over other well-known techniques such as
TDMA and CDMA. SC-FDMA brings additional benefit of low peak-to-average
power ratio (PAPR) compared to OFDMA, making it suitable for uplink trans-
missions [Sesi 09]. The SC-FDMA transceiver has similar structure as a typical
OFDMA system except the addition of a new DFT block before subcarrier map-
ping. Hence, SC-FDMA can be considered as an OFDMA system with a DFT
mapper.

1.4.1.3 About OFDM-IDMA

When using OFDMA and SC-FDMA, an MAI free transmission can be achieved
by allocating different subcarriers to different users. Thus, if there are more than
one user in the system, the entire bandwidth has to be shared by all the users.
This may limit the data rate.
In 2002, Ping et al. [Ping 02a] proposed the interleave-Division Multiple Access
(IDMA) for asynchronous1 spread-spectrum mobile systems. Whereas transmit-
ters can be defined by using different orthogonal codes in CDMA systems, they
are distinguished by a different chip-level interleaver in IDMA systems. While
CDMA allows the MAI to be suppressed using the different codes, IDMA requires
an iterative process. Like CDMA, the entire bandwidth can be allocated to a sin-
gle user, achieving a very high single-user capacity when using IDMA.
IDMA systems require a method to suppress multipath effects and avoid the ISI.
One solution may be to use an ISI cancellation method as in [Ping 06]. However,
the corresponding computational cost of this method increases linearly with the
number of paths and may be unsuitable for high-data rate systems.
To solve this constraint, a scheme that combines OFDM and IDMA has been
proposed by Mahafeno et al. [Maha 06]. This architecture combines most of
the advantages of the OFDM and the IDMA and avoids their individual disad-
vantages. When an OFDM-IDMA is considered, ISI is resolved by an OFDM
1
The users of an IDMA system do not have to be time-synchronized one to another.

26
 1.4 Multiple Access Techniques for Mobile Systems

modulation and MAI is suppressed by the IDMA iterative reception.

In the next subsection, we focus our attention on the OFDMA system. Then, in
subsection 1.4.3, the OFDM-IDMA system will be presented in details.

1.4.2 OFDMA System

Basically, the OFDMA system is equivalent to an OFDM system. The difference
is that each OFDMA symbol simultaneously carries the information for multiple
users while OFDM system carries data of a single specific user.
Let us consider an OFDMA system consisting of a single base station (BS) and
U simultaneously users performing an uplink communication1 .
The U users share the bandwidth W , divided in K subcarriers to perform a trans-
mission. The users are numbered from 1 to U , i.e u ∈ {1, . . . U }. In the following
the subscript u denotes the information associated to the uth user. In addition,
the overall subcarriers are numbered from 0 to K − 1, i.e k ∈ {0, . . . K − 1}. The
subcarriers are grouped in K subchannels. One or more subchannels may be al-
located to the same user depending on its requested data rate. As the maximum
number of users that the system can simultaneously support is limited to K, it is
assumed that U ≤ K.

1.4.2.1 OFDMA Uplink Transmitter

This subsection describes the OFDMA transmitter model. After channel coding
and modulation, the symbols are grouped into blocks of length Ku < K, where
PU
u=1 Ku = K.

Carrier Allocation Strategy (CAS)

The OFDMA transmitter performs the CAS. Three possible strategies [Wang 04]
to distribute subcarriers among the active users have been proposed:
1
In the OFDMA uplink communication, the receiver is the BS. Then, the received signal
is a superposition of the signals transmitted by each user. Thus, the synchronization and the
equalization in this case are more difficult than in the downlink case, where the user only
receives the signal transmitted by the BS.

27
1. Multicarrier High-Data Rate Mobile Wireless Systems

• Subband CAS: each subchannel is composed by a group of Ku adjacent

subcarriers. The main drawback of this scheme is that it does not exploit
the frequency diversity of the multipath channel since a large fading might
strike a substantial number of subcarriers for a given user. See figure 1.12.

Power user

user 2

user 3

user 4

Frequency
O
FD
M
O
FD

A
sy
M

m
A

bo
fra

ld
m

ur
e

at
du

io
ra

n
t
io
n

Subcarrier spacing
Time

Figure 1.12: Subband CAS

• Interleaved CAS: the subcarriers of each user are uniformly spaced over the
signal bandwidth at a distance K from each other. This method can exploit
the channel frequency diversity. See figure 1.13.

Power user

user 2

user 3

user 4

Frequency
O
FD
M
O
FD

A
sy
M

m
A

bo
fra

l
m

du
e

ra
du

tio
ra

n
tio
n

Subcarrier spacing
Time

Figure 1.13: Interleaved CAS

28
 1.4 Multiple Access Techniques for Mobile Systems

• Generalized CAS: each user can select the best subcarriers that are currently
available, e.g. those with the highest signal-to-noise ratios (SNRs). In this
allocation strategy, there is no rigid association between subcarriers and
users; the generalized CAS allows dynamic resource allocation and provides
more flexibility than the other CAS. When using generalized CAS, a priori
knowledge of the propagation channel is necesssary, i.e. the transmitter
needs a channel feedback. See figure 1.14.

Power user

user 2

user 3

user 4

Frequency
O
FD
M
O
FD

A
sy
M

m
A

bo
fra

ld
m

ur
e

at
du

io
ra

n
tio
n

Subcarrier spacing
Time

Figure 1.14: Generalized CAS

Then, the CAS maps the Ku symbols (M-PSK symbols) of each block to the
subcarriers assigned to the uth user. This operation is easily performed by ex-
tending the block with the insertion of K −Ku zeros. The resulting column vector
of dimension K is defined as follows:
∆
Su = [Su (0), Su (1), . . . , Su (k), . . . Su (K − 1)]T (1.19)

according to the frequency allocation of each user, Su (k) is non-zero if the kth
subcarrier is allocated to the uth user. Su (k) is the symbol associated to the kth
subcarrier. The resulting block has a time duration T = KTs . In the following
let T be the OFDMA symbol time duration.

29
1. Multicarrier High-Data Rate Mobile Wireless Systems

Inverse Fast Fourier Transform (IFFT)

To obtain the OFDMA symbol, the symbols are sent in parallel over orthogonal
subcarriers by using a K-dimensional IFFT step [Wein 71] as follows:
K−1
X kn
Xu (n) = Su (k)ej2π K 0≤n≤K −1 (1.20)
k=0

Then, the IFFT output samples are gathered into a vector, resulting in the
OFDMA symbol:
∆
Xu = [Xu (0), Xu (1), . . . , Xu (n), . . . , Xu (K − 1)]T
(1.21)
= FH Qu Su

where FH is the K × K IFFT matrix and F is the K × K fast Fourier transform

(FFT) matrix defined as1 :
 
1 1 1 ... 1
 j2π j4π j2(K−1)π 
1
 e− K e− K ... e− K 

j4π j8π j4(K−1)π
e− K e− K e− K
 
F=

1 ... 
 (1.22)
 .. .. .. ... .. 
 . . . . 
 
j2(K−1)π j4(K−1)π j2(K−1)2 π
1 e− K e− K . . . e− K

In addition, Qu is the CAS diagonal matrix where the kth coefficient of the main
diagonal is given by: (
1 if Su (k) 6= 0
Qu (k, k) = (1.23)
0 elsewhere
Adding a Cyclic Prefix (CP)
At the receiver due to the multipaths, the lth copy of a determined OFDMA
symbol2 arrives with a time delay τl , where l ∈ {0, . . . Lu − 1} and Lu is number
of paths of the channel. See section 1.3. Then, some samples of the OFDMA
symbol interfere with the adjacent OFDMA symbol causing IBI.
The next step of the OFDMA transmitter consists in introducing the cyclic prefix
(CP), that is a repetition of the Ng last samples of the vector Xu at the beginning
of it. The CP makes the OFDMA symbol robust to IBI. By inserting a CP of
1
It should be noted that FFH = IK .
2
The users are assumed to be time-synchronized one another.

30
 1.4 Multiple Access Techniques for Mobile Systems

Figure 1.15: Cyclic prefix, where m denotes the OFDMA symbol index

size Ng ≥ max
u
(Lu ), the interferences between adjacent OFDMA symbols are
avoided. To maintain acceptable data throughput, the CP length must be chosen
just greater than the CIR length. See figure 1.15 which shows the advantage of
adding the CP.
Thus, a CP of size Ng is inserted to Xu such that:

Xu (n) = Xu (n + K) for − Ng ≤ n < −1 (1.24)

Then, the OFDMA symbol to be transmitted of size KT = K + Ng is expressed

as:
∆ T
XCP
u = [Xu (K − Ng ), . . . , Xu (K − 1), Xu (0), . . . , Xu (K − 1)] (1.25)

Parallel/Serial Converter (P/S)

The OFDMA symbols pass through a parallel-serial converter (P/S) and are
stored to compose the OFDMA frame to be transmitted. A simplified OFDMA
transmitter scheme is shown in figure 1.16.
The OFDMA frame is then inserted in a digital/analog (D/A) converter and
the output signal is modulated to the carrier frequency of the local oscillator
(LO). Finally, the signal passes through the power amplifier and it is transmitted
over the channel. Figure 1.17 shows the OFDMA spectrum. It should be noted
that the distance in the frequency domain between two adjacent subcarriers is
1
∆f = KT s
= T1 .
As an example, table 1.3 shows the IEEE 802.16 OFDMA Wireless MAN system

31
1. Multicarrier High-Data Rate Mobile Wireless Systems

parameters, which uses OFDMA for the uplink transmisssion.

In the following, let us look at the OFDMA uplink receiver.

CP-OFDMA modulation
symbols to be
transmitted Su Xu XCP
u
ADD P/S
CAS IFFT
CP

Figure 1.16: Simplified OFDMA transmitter for the uth user, where Xu is the
OFDMA symbol

1
∆f = T
Power
User 1
User 2
User 3

f0 f1 f2 f3 fk fk+1 fk+2 Frequency

Figure 1.17: OFDMA spectrum, where fk is the subcarrier frequency

System parameters Mode 1 Mode 2 Mode 3 Mode 4 Mode 5

System bandwidth (MHz) 1.25 2.5 5.0 10 20
Sampling frequency (MHz) 1.429 2.857 5.714 11.429 22.857
Sampling period (ns) 700 350 175 88 44
FFT size 128 256 512 1024 2048
Subcarrier spacing 11.16 kHz
CP duration 11.2 µs
OFDMA symbol duration 100.8 µs

Table 1.3: Physical layer parameters of IEEE 802.16 Wireless MAN

32
 1.4 Multiple Access Techniques for Mobile Systems

LO
Tolerance= fufc (ppm) fuLO (kHz) fuDOP + fuLO (kHz) ǫu
4 10.40 11.12 0.9966
3 7.80 8.52 0.7636
2 5.20 5.92 0.5306
1 2.60 3.32 0.2977
0.5 1.30 2.02 0.1812

Table 1.4: CFO values for different values of LO tolerance, using the IEEE 802.16
Wireless MAN system parameters, vu = 300 km/h and fc =2.6 GHz

1.4.2.2 OFDMA Uplink Receiver

After down-conversion and low-pass filtering, the received signal passes into an
1
analog-to-digital (A/D) converter, where it is sampled at the frequency Fs = Ts
.
The signal received by the BS is a superposition of the contributions from the
U active users. After the time synchronization, the received OFDMA symbol of
size KT can be expressed as:
U
X
RCP = RuCP + BCP (1.26)
u=1

where BCP is a zero-mean AWGN vector with variance σB2 IK T and

∆
RuCP = [Ru (K − Ng ), . . . , Ru (K − 1), Ru (0), . . . , Ru (K − 1)]T (1.27)

At the BS, the signal transmitted by each user has been affected each propagation
channel, thus:
LX
u −1
ǫu n
RuCP (n) = ej2π K hu (l)Xu (n − l) − Ng ≤ n ≤ K − 1 (1.28)
l=0

where hu (l) is the lth coefficient of the CIR, ǫu = K(fuDOP + fuOSC )Ts is the carrier
frequency offset (CFO) normalized to the subcarrier spacing, f DOP is the Doppler
shift and fuOSC is the frequency difference in Hertz between the carrier frequencies
of the LOs . Tables 1.4 and 1.5 show examples of the CFO magnitudes.
Then, at the OFDMA uplink receiver a CFO estimation/correction is necessary
(See chapter 2). Let us consider that this step is performed perfectly. Then,

33
1. Multicarrier High-Data Rate Mobile Wireless Systems

relative speed (km/h) fuDOP (Hz) ǫu

300 722.22 0.0647
200 481.48 0.0431
100 240.74 0.0216
80 192.59 0.0173
60 144.44 0.0129

Table 1.5: CFO values for different relative speeds, using the IEEE 802.16 Wire-
less MAN system parameters and fuLO = 0 Hz

removing the CP, the contribution from each user to the nth sample of the received
OFDMA symbol can be expressed as:

LX
u −1

Ru (n) = hu (l)Xu (n − l) 0 ≤ n ≤ K − 1 (1.29)

l=0

Then, the contribution from each user to the received OFDMA symbol at the
BS satisfies:

∆
Ru = [Ru (0), Ru (1), . . . , Ru (n), . . . , Ru (K − 1)]T
(1.30)
= hu Xu

where hu is a K × K circulant matrix1 defined as:

 
hu (0) 0 ... 0 hu (Lu − 1) . . . hu (1)
 ... ... 

 hu (1) hu (0) 0 0 hu (2)  

 .. ... ... ... ... ... ..  


. .  
hu = 
 hu (Lu − 1) hu (Lu − 2) ... ... 
 hu (0) 0 0  
 ... ... .. 

 0 hu (Lu − 1) hu (1) hu (0) .  
 .. ... ... ... ... ... .. 

 . . 
0 0 ... hu (Lu − 1) hu (Lu − 2) . . . hu (0)
(1.31)

1
Due to the CP, at the reception, the convolution between the transmitted signal and the
channel can be expressed a circular convolution.

34
 1.4 Multiple Access Techniques for Mobile Systems

Finally, given (1.30) the received OFDMA symbol can be expressed as:
∆
R = [R(0), R(1), . . . , R(n) . . . , R(K − 1)]T
U
X
= Ru + B
u=1
(1.32)
XU
= hu Xu + B
u=1

where B = [B(0), B(1), . . . , B(n) . . . , B(K − 1)]T is a zero-mean AWGN vector

with variance σB2 IK .
Then, to recover the transmitted symbols, an FFT step is performed over the
received OFDMA symbol R as follows:
∆
r = [r(0), r(1), . . . , r(k), . . . , r(m, K − 1)]T
= FR
(1.33)
U
X
= Fhu Xu + b
u=1

where b = FB.
Then, given (1.21), (1.22) and (1.23), (1.33) can be rewritten as:
U
X
r= Fhu F Qu Su + b
| {z }
u=1
Hu
(1.34)
U
X
= Hu Qu Su + b
u=1

where Hu = diag {[Hu (0), hu (1), . . . , Hu (K − 1)]} is the channel frequency re-
sponse matrix, the entries of which are the channel frequency responses associated
to the kth subcarrier:
L−1
X kl
Hu (k) = hu (l)e−j2π K (1.35)
l=0

Then, r(k) can be expressed as:

U
X
r(k) = Hu (k)Su (k) + b(k) (1.36)
u=1

where Su (k) 6= 0 if the kth subcarrier is allocated to the uth user and
P −j2π nk
b(k) = K−1
n=0 B(n)e
K .

35
1. Multicarrier High-Data Rate Mobile Wireless Systems

After subcarrier selection, the symbols r(k) pass through channel equalizers
(EQU) to obtain the transmitted symbols. It should be noted that due to the
diagonal scheme of the matrix Hu the channel equalization can “easily” be per-
formed. The OFDMA uplink receiver block scheme is shown in figure 1.18.

CFO R r subcarrier
Removing selection
estimation/ FFT
CP Qu
correction
RCP CP-OFDMA demodulation
Time
synchronization EQUs

Figure 1.18: OFDMA uplink receiver

In this subsection, we assume a perfect CFO estimation/correction. In addi-

tion, we do not give details about the channel equalization. The CFO estima-
tion/correction and channel equalization are usually performed using a preamble
and/or pilot subcarriers. The preamble is a training sequence known by the re-
ceiver, whereas the pilot subcarriers are frequency subcarriers with high power
that carry a sequence known by the receiver. Those important issues have to
be considered in the design of an OFDMA mobile wireless system. In chapter
2, we propose several techniques for the frequency synchronization and channel
equalization.
In the following subsection, the OFDM-IDMA scheme is presented.

1.4.3 OFDM-IDMA System

The main difference between OFDM-IDMA and OFDMA is that the entire band-
width of the system can be allocated to all users at the same time. Therefore,
CAS is not necessary, reducing the complexity of the OFDM-IDMA transmitter.
Let us define an OFDM-IDMA system consisting of a single BS and U simultane-
ously users performing an uplink communication, as defined in section 1.4.2. The
U users share the bandwidth W , divided in K subcarriers to perform a transmis-
sion. The users are numbered from 1 to U , i.e u ∈ {1, . . . U }. In the following

36
 1.4 Multiple Access Techniques for Mobile Systems

the subscript u denotes the information associated to the uth user. In addition,
the overall subcarriers are numbered from 0 to K − 1, i.e k ∈ {0, . . . K − 1}.

1.4.3.1 OFDM-IDMA Uplink Transmitter

IDMA transmitter
The bits of the uth user to be transmitted are encoded by a low-rate encoder
(ENC), which consists of an association of channel coding1 and spreading code of
Q
size S. The encoded bits pass through the interleaver u and then are modulated.
The symbols are divided into blocks of length K; the resulting IDMA modulated
block is:
∆
Su = [Su (0), Su (1), . . . , Su (k), . . . Su (K − 1)]T (1.37)

CP-OFDM Modulation
An IFFT is performed over the IDMA modulated symbol as in (1.20) resulting
in the OFDM-IDMA symbol:

∆
Xu = [Xu (0), Xu (1), . . . , Xu (k), . . . Xu (K − 1)]T (1.38)

At that time, a CP is inserted as in (1.24). The OFDM-IDMA symbols pass

through a P/S and are stored to compose the OFDM-IDMA frame to be trans-
mitted. The basic principles of an OFDM-IDMA transmitter is shown in figure
1.19. Finally, as for the OFDMA transmitter, the OFDM-IDMA frame is inserted
in a D/A converter, modulated to the carrier frequency of the LO, amplified and
transmitted over the channel.

1.4.3.2 OFDM-IDMA Uplink Receiver

As in the OFDMA uplink receiver, after down-conversion and low-pass filtering,

the received signal passes into an A/D converter. At the BS, it has been affected
by the propagation channel. Thus, the transmitted OFDM-IDMA symbol is
affected by the different CFOs and multipath channels. As presented in subsection
1
Depending on the targeted system performance, the channel coding can be inserted or not.

37
 1. Multicarrier High-Data Rate Mobile Wireless Systems

IDMA transmitter
bits to be
transmitted Q Su CP-OFDM
Xu
ENC u modulation modulation

Q
Figure 1.19: OFDM-IDMA transmitter for the uth user, where u represents
the interleaver

1.4.2.2, the OFDM-IDMA received symbol after time-synchronization, frequency

synchronization and CP removing can be expressed as follows:
U
X
R= Ru + B
u=1
U
(1.39)
X
= hu Xu + B
u=1

where B = [B(0), B(1), . . . , B(n) . . . , B(K −1)]T is zero-mean AWGN vector with
variance σB2 IK and hu is the channel circulant matrix as is expressed in (1.31).
Thus, after the FFT step the received symbol can be expressed as:
∆
r = [r(0), r(1), . . . , r(k), . . . , r(K − 1)]T
= FR
(1.40)
U
X
= Fhu Xu + b
u=1

with r(k) defined in (1.36).

The receiver consists of an ESE and U a posteriori probability DECs, one for each
user. The two steps are considered separately. The results are then combined
by using an iterative process. The DECs produce hard decisions to obtain the
estimations of the transmitted bits during the final iteration1 .
Let us denote the a priori extrinsic logarithm likelihood ratios (LLRs) about
Su (k) by: !
∆ Pb {Su (k) = +1}
gu (k) = log (1.41)
Pb {Su (k) = −1}
1
It should be noted that the IDMA receiver may be applied to real and complex signals
[Maha 07]. However, for the sake of simplicity, let us consider a BPSK modulation.

38
 1.4 Multiple Access Techniques for Mobile Systems

where Pb {.} is the probability distribution of {.}.

In the following let us distinguish guESE (k) and guDEC (k) the a priori LLRs used by
the ESE and DECs, respectively.
The ESE uses r(k) and guESE (k) as inputs and only the channel frequency re-
sponse Hu = [Hu (0), . . . , Hu (k), . . . , Hu (K − 1)] is considered. Knowing Hu , the
a posteriori LLRs about Su (k) are defined as1 :
! !
Pb {Su (k) = +1|r(k), Hu } Pb {Su (k) = +1|Hu }
log = log
Pb {Su (k) = −1|r(k), Hu } Pb {Su (k) = −1|Hu }
ESE (1.42)
+ gu (k)
= gESE ESE
u (k) + gu (k)
where gESE
u (k) is the extrinsic LLR about Su (k) based on the channel observation
and the a priori information of the other users, excluding the uth treated user.
Then, if guDEC = [guDEC (0), . . . , guDEC (k), . . . , guDEC (K −1)] are the inputs of the uth
user DEC, the a posteriori LLRs about Su (k) are generated based on the code
ENC as:
! !
Pb {Su (k) = +1|ENC, guDEC } Pb {Su (k) = +1|g̃uDEC }
log = log + guDEC (k)
Pb {Su (k) = −1|ENC, guDEC } Pb {Su (k) = −1|g̃uDEC }
= gDEC
u (k) + guDEC (k)
(1.43)
where g̃u is obtained by setting gu (k) = 0 in gu . The outputs of the DEC
DEC DEC DEC

for the uth user consists of the extrinsic LLRs gDEC

u (k).
During the iterative process, the extrinsic information generated by the ESE
u (k) after deinterleaving is used as the a priori information in the DEC:
gESE
Q−1
ESE u
gu (k) ⇒ guDEC (k) (1.44)
Q−1
Q
where ⇒ u
represents the deinterleaving process and u−1 is the deinterleaver.
Furthermore, the extrinsic information generated by the DEC gDEC
u (k) after in-
terleaving is used as the a priori information in the ESE:
Q
guDEC
(k) ⇒u guESE (k) (1.45)
1
In Bayesian statistics, the a posteriori probability distribution of a random variable x
based on the observations of y is expressed as follows: Pb {x|y} = Pb {y|x}P
Pb {y}
b {x}
, where Pb {x} is
the so-called a priori probability distribution of x. As the denominator does not depend on x,
the so-called a posteriori probability distribution can be rewritten as: Pb {x|y} = Pb {y|x}Pb {x}

39
1. Multicarrier High-Data Rate Mobile Wireless Systems

Q
where ⇒u represents the interleaving process. As there is any a priori information
at the first iteration, the initial values of all guESE (k) are set to zeros.
On the one hand, the a posteriori probability DEC is a standard function that
depends on the ENC, such as convolutional or turbo codes [Ping 01], [Berr 93].
The reader may refer to [Ping 01] and [Berr 93] for more details about the DECs.
On the other hand, let us present details about the ESE. The ESE generates coarse
estimates of the symbol Su (k). The symbol is treated as a random variable and
the a priori LLR guESE (k) about Su (k) is obtained by updating gDEC
u (k) as shown
in (1.45). Given (1.41) we have [Maha 06]:
ESE !
egu (k) − 1
∆ guESE (k)
µu (k) = E{Su (k)} = gESE (k) = tanh ∀u, k (1.46)
eu +1 2
∆
υu (k) = Var{Su (k)} = 1 − (µu (k))2 ∀u, k (1.47)

where E{.} and Var{.} represents the mean and the variance of {.}, respectively.
For the initialization process, guESE (k) = 0 for i = 1, where i denotes the iteration
number with i ∈ {1, . . . , Iidma }, where Iidma is the maximum number of iterations.
Let us recall the expression of the received symbol r(k) done for the OFDMA
uplink receiver in (1.36):
U
X
r(k) = Hu (k)Su (k) + b(k) (1.48)
u=1

where Su (k) 6= 0 ∀k.

Let us rewrite (1.48) as:

r(k) = Hu (k)Su (k) + ςu (k) (1.49)

PU
where ςu (k) = u′ =1 Hu′ (k)Su′ (k) + b(k) represents the interference-plus-noise
u′ 6=u
component in r(k).
Then, assuming that Su are independent and identically distributed random vari-
ables, and applying the central limit theorem1 , a Gaussian approximation for ςu (k)
1
The central limit theorem states that the mean of a very large number of independent
random variables, each one with finite mean and variance, can be approximated by a Gaussian
distribution.

40
 1.4 Multiple Access Techniques for Mobile Systems

and r(k) can be considered. Thus, ςu (k) can be completely characterized by its
mean and variance:

E{ςu (k)} = E{r(k)} − Hu (k)µu (k) ∀u, k (1.50)

Var{ςu (k)} = V ar{r(k)} − (Hu (k))2 υu (k) ∀u, k (1.51)

where
U
X
E{r(k)} = Hu (k)µu (k) ∀u, k (1.52)
u=1
U
X
Var{r(k)} = σB2 + (Hu (k))2 υu (k) ∀u, k (1.53)
u=1

Then, the value of gESE

u (k) can be obtained as follows [Maha 06]:

r(k) − E{ςu (k)}

gESE
u (k) = 2Hu (k) × ∀u, k (1.54)
Var{ςu (k)}
The ESE is summarized in algorithm 1.
Then, gESE
u (k) is updated as gu
DEC
(k) and inserted in the DEC. The DEC provides
gDEC
u (k), this value is updated as guESE (k) and the ESE process is again performed.
During the final iteration the DECs produce hard decisions and the transmitted

Algorithm 1 ESE
Require: r(k), σB2 , Hu (k), Iidma
Initialization step:
guESE (k) = 0 ∀u, k
Iterative process
for i = 1 to Iidma do  ESE 
µu (k) = E{Su (k)} = tanh gu 2 (k) ∀u, k
2
υu (k) = V ar{Su (k)} = 1 − (µu (k)) ∀u, k
P
E{r(k)} = Uu=1 Hu (k)µu (k) ∀u, k
P
V ar{r(k)} = σB2 + Uu=1 (Hu (k))2 υu (k) ∀u, k
E{ςu (k)} = E{r(k)} − Hu (k)µu (k) ∀u, k
V ar{ςu (k)} = V ar{r(k)} − (Hu (k))2 υu (k) ∀u, k
r(k)−E{ςu (k)}
u (k) = 2Hu (k) × V ar{ςu (k)}
gESE ∀u, k
end for

41
1. Multicarrier High-Data Rate Mobile Wireless Systems

bits are estimated1 .

The block scheme of an OFDM-IDMA uplink receiver is shown in figure 1.20.
It should be noted that again we considered a perfectly frequency synchronized.
However, this is a very strong hypothesis for a real wireless communication sys-
tem. The effects of the propagation channel on the OFDM-IDMA received signal
are the same as the ones presented for an OFDMA system in subsection 1.18.
Therefore, the estimation of the CFOs can be implemented as in an OFDMA sys-
tem, but not the CFO correction. In an OFDM-IDMA system, all the users use
all the subcarriers at the same time. Thus, the CFO correction of one user may
destroy the orthogonality of the other users. In addition the IDMA receiver needs
the a priori knowledge of the CIRs. Hence, a channel estimation is required.

iterative IDMA receiver

user decoding 1st user

ESE
g1 (k) Q DEC
g1 (k) estimated
−1
u bits
ESE
g1 (k) DEC
g1 (k) DEC
Q
u
r(k) U th user
ESE estimated
gESE
U (k) Q−1 gUDEC(k)
u bits
DEC
CP-OFDM gUESE(k) Q gDEC
U (k)
demodulation u

2
σB Hu
R

Figure 1.20: OFDM-IDMA uplink receiver [Maha 06], where ESE is the elemen-
tary signal estimator and DEC is the a posteriori probability decoder

1
In this PhD dissertation we consider a parallel processing of the users in the IDMA receiver.

42
 1.5 Conclusions

1.5 Conclusions
In this chapter, our purpose was to present the multicarrier high-data rate mo-
bile wireless systems. We recall various properties of the system such as: 1/ the
influence of the propagation channel leading to various kinds of channels (slow
or fast fading channel, flat-fading or frequency-selective channel), 2/ the various
multiple access techniques that have been derived: FDMA, TDMA, CDMA, SC-
FDMA, OFDMA and OFDM-IDMA.
Then, we focus our attention on two systems. We first present in detail the
OFDMA uplink transmitter and receiver, by giving the expression of the input
and the output of each block in the processing chain. We also look at the OFDM-
IDMA system and present the uplink receiver proposed in [Maha 06]
In both cases, we point out the importance of the estimation of the CFO and the
channel.
In chapter 2, we are going to propose several methods based on Kalman Filtering
and H∞ filtering to carry out the frequency synchronization and channel equal-
ization.
In chapter 3, we will propose some modifications of the conventional OFDM-
IDMA uplink receiver to counteract the effect of the multiple CFOs and the
channel in the received OFDM-IDMA signal.

43
Chapter 2
Frequency Synchronization and Channel
Equalization of OFDMA Uplink Systems

Contents

2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . 47
2.2 State of the Art on the OFDMA Uplink Systems -
Carrier Frequency Offset (CFO) and Channel Esti-
mation . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47
2.3 A Joint CFO/Channel Estimator for OFDMA Systems 53
2.3.1 State-Space Representation of the System . . . . . . . 54
2.3.2 Kalman Filtering in the Non-Linear Case . . . . . . . 56
2.3.3 H∞ Filtering in the Non-Linear Case . . . . . . . . . . 57
2.3.4 Simulation Results . . . . . . . . . . . . . . . . . . . . 58
2.3.5 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . 65
2.4 An OFDMA Non-Pilot Aided Iterative CFO Estimator 66
2.4.1 Problem Formulation . . . . . . . . . . . . . . . . . . 67
2.4.2 User Detection - The Minimum Mean Square Error
Successive Detector (MMSE-SD) . . . . . . . . . . . . 68
2.4.3 CFO Estimation with the MMSE-SD Preambles . . . 71
2.4.4 Simulation Results . . . . . . . . . . . . . . . . . . . . 72
2.4.5 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . 78
2.5 An OFDMA Robust CFO Estimator . . . . . . . . . . 79
2.5.1 System Description . . . . . . . . . . . . . . . . . . . . 80
2.5.2 Joint Disturbance Detection and CFO Estimation . . 84

45
2. Frequency Synchronization and Channel Equalization of OFDMA
Uplink Systems

2.5.3 Simulation Results . . . . . . . . . . . . . . . . . . . . 92

2.6 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . 99

46
 2.1 Introduction

2.1 Introduction
As mentioned in chapter 1, in OFDMA uplink systems, estimating CFO and
channel plays a key role. This chapter deals with the frequency synchronization
and the channel equalization, and presents our contributions in that field.
The chapter is organized as follows: a state of the art dedicated to the CFO
and channel estimation in OFDMA uplink systems is given. We first focus our
attention on the estimation of the CFO for systems with interleaved and gener-
alized CAS [More 04], [Zhao 06], [Jing 08]. The estimation can be done in the
frequency domain or in the time domain. Then, we present the existing methods
dealing with the joint estimation of the CFO and the channel [Pun 04a], [Pun 06],
[Sezg 08]. At that stage, we suggest a recursive estimation of the CFO and the
channel by using optimal filtering, i.e. Kalman filtering and H∞ filtering (and
their variants). These approaches are derived:

• in combination with an MMSE-SD to increase the data rate in the system;

• in combination with a statistical test, i.e. the BHT and the CUSUM test,
to design a robust CFO estimator that detects the beginning and the end
of the disturbances induced by another system.

2.2 State of the Art on the OFDMA Uplink Sys-

tems - Carrier Frequency Offset (CFO) and
Channel Estimation
In an OFDMA uplink system, the signal received at the BS is affected by sev-
eral CFOs and multipath channels. Indeed, the signal transmitted by each user
propagates through different propagation channels. Estimating/correcting the
CFO and the channel is necessary. More particularly, without CFO estima-
tions/corrections, orthogonality between subcarriers is no longer satisfied. It
results in ICI as well as MAI. These multiple frequency offsets make the CFO

47
2. Frequency Synchronization and Channel Equalization of OFDMA
Uplink Systems

estimation more “complex” than the one in a downlink communication1 .

The CFO estimation problem for OFDMA uplink transmissions using control
data2 has been recently addressed in several papers. Thus, in [More 07], Morelli
et al. present a tutorial that illustrates various schemes for uplink CFO estima-
tion among different CAS.
More specifically, when using subband CAS (See chapter 1 in section 1.4.2.1),
CFO can be estimated by inserting unused subcarriers among subchannels to
provide frequency guard intervals as shown in figure 2.1.

0
10
User 1
User 2

−1
10

guard interval
Power (dB)

−2
10

−3
10

−4
10

−10 −8 −6 −4 −2 0 2 4 6 8 10
Frequency (MHz)

Figure 2.1: OFDMA spectrum when using subband CAS

If the CFOs are smaller than the guard intervals, the received signals can be sepa-
rated at the BS by passing through a bank of digital band-pass filters (BPFs), each
one selecting a subband. Therefore, CFO-estimation methods used in OFDM or
OFDMA downlink transmission can be used. For example, the method proposed
in [Moos 94] to estimate the CFO in an OFDM transmission can be performed.
1
Indeed, in an OFDMA downlink communication, the receiver is the user. Thus, one user
has to estimate only a single CFO and a single channel. Then, the CFO estimation can be
performed by using the methods proposed for OFDM single-user communications [Moos 94],
[VdBe 99].
2
The control data are composed by known training sequences called “preambles” and/or
pilot subcarriers.

48
 2.2 State of the Art on the OFDMA Uplink Systems - Carrier
Frequency Offset (CFO) and Channel Estimation

The CFO is estimated by calculating a phase shift in the frequency domain be-
tween two successive identical OFDM symbols. The repetitive OFDM symbol
remains identical after passing through the channel except for a phase shift pro-
duced by the CFO. In [Huan 10a], the authors use control data having a tile
structure in the frequency domain. Then, a high-resolution subspace method al-
lows the CFO to be estimated. In [VdBe 99], Van de Beek et al. take advantage
of the CP and estimate the CFOs by calculating the phase shifts in the time
domain between the Ng last samples of the received OFDMA symbols and the
CP of each user.
In OFDMA uplink systems with interleaved or generalized CAS, it is impossible
to take advantage of the frequency guard intervals. In those systems the CFO
estimation is a challenging task. When using control data, several methods have
been proposed. In [More 04], the CFO is estimated in the ML sense. However,
it requires an exhaustive grid search over all the possible values of the CFO,
leading to a high computational cost. For that reason, Morelli also proposes a
reduced-complexity frequency estimator, that compares the phase shifts of sev-
eral identical received OFDMA symbols in the frequency domain. Nevertheless,
the author assumes that all the active users, except the first one, are already
synchronized in frequency. In [Zhao 06], Zhao et al. use an EKF to estimate the
CFO. In addition, Jing et al. propose an extended H∞ filter to combat intercar-
rier interference for an OFDM system [Jing 08]. However, in both approaches,
the channels are preliminary estimated and this assumption cannot be necessar-
ily satisfied in real cases. Indeed, during the uplink CFO-synchronization stage,
channel state information1 (CSI) is not available, and hence has to be estimated
either jointly with the CFO or after CFO compensation.
In order to correct the MAI, the CFO has to be compensated. In the single-user
detector method, shown in figure 2.2, the CFOs are compensated by multiplying
the complex envelope of the signal before the CP-OFDMA demodulation at the
receiver. However, the number of FFTs increases linearly with the number of
users as well as the computational complexity of the receiver. As an alternative,
new correction methods have been proposed. More particularly in [Dai 07], the
1
The CSI refers to known channel properties of a communication link. It is usually associ-
ated to the CIR.

49
2. Frequency Synchronization and Channel Equalization of OFDMA
Uplink Systems

to the 1st EQU

1st CFO CP-OFDMA r subcarrier r1
Estimation/ selection
demodulation Q1
Correction
RCP
to the U th EQU
U th CFO r subcarrier rU
CP-OFDMA selection
Estimation/
Correction demodulation QU

Figure 2.2: Single-user detector

CFO is compensated by mitigating the effects of the major lobes and the side
lobes of the frequency response of the signals received from each user. The CFO
estimation can be obtained by calculating the phase shifts of two identical re-
ceived OFDMA symbols in the time domain. In [Choi 00], CFO compensation is
based on a circular convolution after the CP-OFDMA demodulation1 , as shown
in figure 2.3. The authors suggest using the method presented in [Moos 94] to
estimate the CFOs. However, both methods only work for subband CAS. Con-
sequently, in [Huan 05], the authors extend the concept proposed in [Choi 00]
and use it in a system with interleaved or generalized CAS. In that case, the

ǫ̂1
to the 1st EQU
BPF r1 BPF Ẑ1
circular
Q1 convolution Q1

R r ǫ̂U
FFT
to the U th EQU
rU ẐU
BPF circular BPF
QU convolution QU

Figure 2.3: CFO correction by means of a circular convolution after the FFT

1
Due to the CP, the received signal without CFO correction after the CP-OFDMA demod-
PU
ulation can be expressed as: r = u=1 FEu ⊗ Zu + b, where Zu = Fhu Xu , Eu is the CFO
matrix and ⊗ represents a circular convolution.

50
 2.2 State of the Art on the OFDMA Uplink Systems - Carrier
Frequency Offset (CFO) and Channel Estimation

Algorithm 2 Circular convolution + IC [Choi 00]

PU
Require: r = u=1 FEu ⊗ Zu + b, Zu = Fhu Xu , Êu , ⊗: circular convolution
Initialization step:
Ẑu = Qu {Qu r ⊗ FÊH u}
Circular convolution + IC step:
for i = 1 to Imax do
P
Z̃u = r − Uu′ =1 Ẑu ⊗ FÊu and
u′ 6=u
Ẑu = Qu {Qu Z̃u ⊗ FÊH
u}
end for

CFOs are corrected via a circular convolution after the CP-OFDMA demodula-
tion, but an iterative interference-cancellation (IC) step is required, as shown in
algorithm 2. The authors suggest again using the method presented in [Moos 94].
In [Cao 04a], the CFO is corrected by means of a linear detection, whereas the
CFO is estimated using a high-resolution subspace method, i.e. the multiple sig-
nal classification (MUSIC) method [Ther 92]. More recently, Hou et al. have
proposed a minimum mean square error successive detector (MMSE-SD) to sup-
press the MAI, but the CFO is assumed to be known [Hou 08].

In the above approaches, channel estimation is not included. The joint esti-
mation of the CFO and the channel has been also investigated. In [Pun 04a] and
[Pun 06], the authors study how to obtain the joint ML estimations of the chan-
nels and the CFOs of multiple users. Thus in [Pun 04a], a conventional EM is
first proposed: during the E-step, the received signals transmitted by each user,
namely the “complete data”, are estimated. During the M-step, all the CFOs and
the channels are jointly estimated by using these complete data. To simplify the
optimization issue, the value of the channel is replaced by its expression depend-
ing on the CFO in the criterion to be minimized. Therefore, only the CFO of
each user has to be estimated. Even if the criterion is explicitly given, the authors
do not mention the estimation method they used. For instance, exhaustive grid
search could be considered as suggested by the same authors in [Pun 04b]. To
reduce the computational cost, the authors in [Pun 04a] use the SAGE algorithm.

51
2. Frequency Synchronization and Channel Equalization of OFDMA
Uplink Systems

In that case, instead of simultaneously estimating every-user parameters, one it-

eration of the EM algorithm is dedicated to one user. Instead of addressing a
multidimensional optimization issue, the so-called alternating projection estima-
tor is used in [Pun 06]. This method consists in iteratively estimating the CFO
of one user, by means of an exhaustive grid search over the possible range of the
CFO value and by setting the other CFOs to their last updated values. In [Fu 06],
Fu et al. propose two iterative estimation approaches using the SAGE method.
In addition, in [Saem 07] Saemi et al. propose two EM-based algorithms. The
first one estimate the CFO without the CSI whereas the second one estimate
jointly the CFO and the channel. Nevertheless, the EM-based algorithms do not
necessarily converge to the global extremum. Like any EM algorithm, an ini-
tialization step is required. Another drawback of the above methods is the high
computational cost due to the iterative estimation and the exhaustive grid search.
For a large number of subcarriers, the EM-based algorithms are impractical. For
that reason, Sezginer et al. [Sezg 08] proposed an iterative suboptimal method.
It is based on an approximation of an ML estimator to reduce the computational
complexity of the EM-based algorithms. However, it is still based on an exhaus-
tive grid search and an iterative architecture.
To reduce the amount of control data, alternative estimation techniques have
been proposed. If a subband CAS OFDMA uplink system is considered, the
method presented in [Cibl 04] can be considered. Thus, Ciblat et al. have devel-
oped a blind1 CFO estimation for an OFDM single-user system based on offset
quadrature amplitude modulation (OQAM). The CFO is estimated by searching
the maximum of the periodogram of the received signal. Nevertheless, this is a
single-parameter estimation method and cannot be implemented for interleaved
and/or generalized CAS. For that reason, in [Cao 04b], the authors propose a
subspace method based blind CFO estimation algorithm for an interleaved CAS
OFDMA uplink system. It consists of a high-resolution signal-processing tech-
nique to estimate the CFO without control data. Nevertheless, more subchannels
than users are required. In addition when the values of the different CFOs are
close one another, the grid search approach leads to a suboptimal estimation. In
1
A blind CFO estimation is an estimation of the CFO that does not require control data.

52
 2.3 A Joint CFO/Channel Estimator for OFDMA Systems

[Mova 08], a blind method to estimate CFOs based on a linear precoder is pro-
posed. Using two OFDMA symbols, the idea is to find a time correlation using
a precoder which gives a second-order moment based CFO estimation for each
user.
Several approaches using ML estimation of the CFO and channel estimation in
multicarrier systems have been proposed. However, only few approaches are
based on optimal recursive filtering to perform the CFO and channel estimation
in OFDMA uplink systems. One feature is that this multiparameter estimation
solution can be implemented recursively. In particular, each updated estimate
is computed from the previous estimate and the current observation. Therefore,
only the previous estimation has to be stored. The optimal filtering is hence
computationally more efficient than algorithms that compute the estimation us-
ing the entire past observed data at each step of the estimation process.
In the following, different optimal filtering techniques aiming at estimating the
CFO and the channel are proposed for conventional and CR environments.

2.3 A Joint CFO/Channel Estimator for OFDMA

Systems

Due to the propagation channels and the radio frequency (RF) effects, the uplink
signals at the BS are plagued by synchronization and estimation errors. There-
fore at the BS, synchronization and equalization process has to be done. These
processes usually require control data. Commonly, the synchronization process is
performed before the CIR estimation. However, if a time-synchronized OFDMA
uplink system is considered, the CFO and the channel can be jointly estimated.
Then, for a time-synchronized system the nth sample of the received OFDMA
symbol at the BS after CP removing is:

U LX
X u −1
ǫu n
R(n) = ej2π K hu (l)Xu (n − l) + B(n) 0 ≤ n ≤ K − 1 (2.1)
u=1 l=0

53
2. Frequency Synchronization and Channel Equalization of OFDMA
Uplink Systems

Then, the received OFDMA symbol can be expressed as follows:

∆
R = [R(0), R(1), . . . , R(n), . . . , R(K − 1)]T
U
X (2.2)
= Eu hu Xu +B
u=1
| {z }
Ru

nh io
where Eu = diag 1, ej2πǫu /K , . . . , ej2π(K−1)ǫu /K (See chapter 1 in section 1.4.2.2).
In order to restore orthogonality among each user subcarrier, both the CFO-
synchronization error vector ǫ and the CSI vector h have to be estimated at the
BS.
PU
Let us define the vector ǫ of size U and the vector h of size L = u=1 Lu by
storing the CFOs and the CSIs of the U users in the system as follows:

ǫ = [ǫ1 , ǫ2 , . . . , ǫu , . . . , ǫU ] (2.3)
h = [h1 , h2 , . . . , hu , . . . , hU ] (2.4)

where hu = [hu (0), hu (1), . . . , hu (l), . . . , h(Lu − 1)].

In this section, a joint estimation of the CFOs and the CSIs using Kalman and H∞
filtering is proposed [Pove 10]. The Kalman filter requires an exact and accurate
system model as well as perfect knowledge of the noise statistics. In contrast,
the H∞ filtering is more robust against model uncertainty and does not need
any a priori knowledge of the noise statistics. It should be noted that the work
presented in this section is complementary to the study done in [Kang 07], that
presents a joint CFO/channel estimation in a single-user scenario using an UKF.
As both Kalman and the H∞ filtering are based on a state-space representation
of the system, the next subsection briefly recalls the equation updating the state
vector and the observation equation.

2.3.1 State-Space Representation of the System

The state-space representation of the system (2.1)-(2.2) is the representation of
what happens during the OFDMA frame1 .
1
The OFDMA frame has been introduced in chapter 1 in section 1.4.2.1; it is a block
composed of several OFDMA symbols.

54
 2.3 A Joint CFO/Channel Estimator for OFDMA Systems

Let us define the vector x(n) of size U = U + 2L, which is the state vector of the
system (2.1)-(2.2):
h iT
x(n) = ǫ(n) Re {h(n)} Im {h(n)} (2.5)

In order to improve the performance of the proposed algorithm [Kang 07], a

separate estimation of the real and the imaginary parts of the state-vector is
proposed1 . Let Y be the 2 × K observation matrix that stores the real and the
imaginary parts of the received OFDMA symbol2 R:

Y(ǫ, h) = C(ǫ, h) + B
(2.6)
= [Y(0, ǫ, h), Y(1, ǫ, h), . . . , Y(n, ǫ, h), . . . , Y(K − 1, ǫ, h)]

where C(ǫ, h) is a 2 × K matrix that stores the real and the imaginary parts of
the contributions of alls users to the received OFDMA symbol:
 nP oT 
U
Re u=1 Ru
C(ǫ, h) = 
 nP

oT 
Im U
u=1 Ru (2.7)
= [C(0, ǫ, h), C(1, ǫ, h), . . . , C(n, ǫ, h), . . . , C(K − 1, ǫ, h)]

and " #
Re{B(n)}T
B= = [B(0), B(1), . . . , B(K − 1)] (2.8)
Im{B(n)}T
B(n) is an AWGN vector with zero-mean and covariance matrix σ2B I2 .
Now, let us introduce the state-space representation of the system (2.1)-(2.2) to
estimate the CFO and the channel:
State equation:

x(n) = x(n − 1) + w(n) ∀ n ∈ [0, K − 1] (2.9)

Measurement equation:

Y(n) = C(n, x(n)) + B(n) ∀ n ∈ [0, K − 1] (2.10)

1
As an optimal filtering estimation is proposed, one solution could be to use a complex
Kalman filter [Dash 00] or a complex H∞ filter [Nish 99]. However, in order to guarantee the
stability of the filter [Dini 10], we have chosen to separate the estimation of the real and the
imaginary parts.
2
C(ǫ, h) indicates that the matrix C depends on these vectors.

55
2. Frequency Synchronization and Channel Equalization of OFDMA
Uplink Systems

where w(n) is an AWGN matrix with zero-mean and covariance matrix σw2 IU . It
should be noted that σw2 is very low and can even be equal to 0. In this latter
case, the CFO and the channel are assumed to be constant during one OFDMA
symbol.

2.3.2 Kalman Filtering in the Non-Linear Case

When dealing with the non-linear state-space representation (2.9)-(2.10) of the

system and if there is no model noise in the state-equation, methods such as the
linearized recursive least squares (RLS) can be considered [Hayk 96]. More gen-
erally, if there is a model noise in equation (2.9) the EKF [Hayk 96] presented in
appendix D can be used. However, due to the first-order approximation, the EKF
may sometimes diverge when evaluating the mean and the covariance matrix of
the random variable that undergoes the non-linear transform. To address this
approximation issue, a second-order linearization can be considered and leads to
the SOEKF [Bar 01] presented as well in appendix D. In that case, the Jacobian
matrix must be computed for the first-order linearization, but also the Hessian
matrix for the second-order linearization. As an alternative, the IEKF presented
in appendix D linearizes the measurement model around the updated state, in-
stead of the predicted state. At each time step, this linearization around the
estimate can be done several times (e.g. 2 to 5 times). It should be noted that in
[Grew 10], another kind of EKF is presented to address numerical issues of the
EKF. However, these derivations may be unstable and do not necessarily guaran-
tee the convergence of the filter. For that reason, the SPKF presented in appendix
E, namely the UKF or the CDKF, are considered in this thesis [VdMe 04a]. The
UKF is based on the unscented transformation [Wan 01], whereas the CDKF is
based on the second-order Sterling polynomial [Ito 00]. The main advantages of
the SPKF over the EKF is that they do not require calculations of Jacobians or
Hessians. See table 2.1.
In this section, these five types of Kalman filters (KFs) are used to recursively
estimate the state-vector.

56
 2.3 A Joint CFO/Channel Estimator for OFDMA Systems

Non-linear Estimation
No model noise Model noise
Linearized RLS Kalman filter
Linearization Gaussian approximation

EKF UKF 
SPKF
SOEKF CDKF
IEKF

Table 2.1: Gaussian non-linear estimation methods

2.3.3 H∞ Filtering in the Non-Linear Case

In the above approaches, a priori knowledge on the model noise w and the mea-
surement noise B characteristics is required. To relax the assumption on the
model noise and additive measurement noise, we suggest investigating the rele-
vance of the H∞ filter. H∞ filter is known to be more robust to uncertainties
than Kalman filtering-based estimation [Hass 99]. The H∞ theory purpose is to
minimize the worst possible effects of the disturbances on the estimation error.
No statistical assumptions have to be made on the model and the measurement
noises. They are just assumed to have finite energies.
Let us introduce a third state-space equation to focus on a linear combination of
the state-vector components:

z(n) = Lx(n) (2.11)

where L = IU . The reader is referred to appendix B, where the state-space

representation is linear and to appendix F, where the state-space representation
is non-linear.
In this section, two types of H∞ filters are proposed to estimate the state vector
in a recursive way:

• like the EKF, the“extended H∞ filter” consists of a first-order expansion of

the state-space functions around the last available estimation of the state
vector [Burl 98];

• the unscented H∞ filter [Li 10], which is based on the unscented transfor-
mation embedded into the “extended H∞ filter” architecture.

57
2. Frequency Synchronization and Channel Equalization of OFDMA
Uplink Systems

The reader is referred to table 2.2. In addition, appendix C deals with a compar-
ison between the equations that define Kalman filtering and H∞ filtering.

No statistical
No linearization
assumptions
EKF x x
SPKF X x
Extended
x X
H∞ filter
Unscented
X X
H∞ filter

Table 2.2: The KF vs the H∞ filter in a non-linear case, Xrepresents a fulfilled

option whereas x represents a not fulfilled option

In the following subsection, simulation results show the efficiency of the proposed
optimal filtering algorithms.

2.3.4 Simulation Results

In the following, a comparative study is carried out between:

• the EKF, the SOEKF, the IEKF, presented in appendix D;

• the UKF, the CDKF, presented in appendix E;

• the “extended H∞ filter”, the “unscented H∞ filter” presented in appendix

F;

• and the methods presented in [Pun 04a] and [Pun 06] where a grid search
approach is used to update the CFO estimation.

Simulation protocol:
An OFDMA IEEE 802.16 WirelessMANTM uplink system composed of U = 4
users sharing K = 128 subcarriers and with a cyclic prefix Ng = K/8 ≥ max
u
(Lu )
is considered. A transmission over a Rayleigh slow-fading frequency-selective

58
 2.3 A Joint CFO/Channel Estimator for OFDMA Systems

channel composed of Lu = 3 ∀u multipaths is supposed. BPSK is used to modu-

late the information bits. The carrier frequency is at fc = 2.6GHz and the band-
width is set to W = 20MHz. The users’ normalized CFO errors are randomly and
2
uniformly generated in the interval [-1,1]. Then, let us define SNR= 10log( σσ2u ),
B
where σu2 is the mean power of the received signal from the uth user.

2
σB
When using H∞ filtering: Ξ = 102 , V = I
2 2
and W = IU σw2 . One assumes
that there is a state noise w(n) with very small variance, e.g. σw2 = 10−3 . For
the CFO, the initialization parameters of the algorithm is ǫ̂u (0) = 0 ∀u.

The grid search algorithms proposed in [Pun 04a] and [Pun 06] are based on
Iem = 20 iterations and a grid search precision equal to 10−4 ; this means that
2
b = 10−4
+ 1 = 20001 values of CFO are studied in the grid search algorithm.
The estimations are obtained by using control data of one OFDMA symbol.

2.3.4.1 Estimation Performance

We focus our attention on the first user in the system. First of all, our approaches
provide similar results when UKF and CDKF are used. Therefore, in the follow-
ing, we make reference to SPKF filtering performance.
Figure 2.4 and 2.5 show the results in terms of CFO and channel MMSE consid-
ering a perfect knowledge of the noise statistics. When the noise characteristics
are available, the EKF provides quite similar results in comparison with the “ex-
tended H∞ filter” and there is no real difference between the SPKF and the
“unscented H∞ filter”. For that reason, in figure 2.4 and 2.5, we only show the
results of the Kalman approaches. In addition, the algorithms in [Pun 04a] and
[Pun 06] provide similar results after a given number of iterations. Thus, figure
2.4 and 2.5 show only the performance of the EM-based algorithm presented in
[Pun 04a].
The IEKF gives the best estimation performance. However, due to the calcula-
tion of the Jacobian matrix at each iteration, the computational complexity is
the highest among the proposed approaches. The SPKF algorithms give better
CFO and channel estimations than the SOEKF and the EKF.

59
2. Frequency Synchronization and Channel Equalization of OFDMA
Uplink Systems

−2
10
EKF
SOEKF
SPKF
IEKF
EM

−3
10
MMSE

−4
10

0 2 4 6 8 10 12 14 16
SNR (dB)

Figure 2.4: CFO estimation performance for a joint channel/CFO estimation

when the noise statistics are available

0.11
EKF
SOEKF
0.1 SPKF
IEKF
EM
0.09

0.08

0.07

0.06
MMSE

0.05

0.04

0.03

0.02

0.01

0 2 4 6 8 10 12 14 16
SNR (dB)

Figure 2.5: Channel estimation performance for a joint channel/CFO estimation

when the noise statistics are available

Figure 2.6 shows that the IEKF converges faster than the other Kalman ap-
proaches. In addition, the SPKF seem to require less observations than the EKF

60
 2.3 A Joint CFO/Channel Estimator for OFDMA Systems

CFO real value

−0.05
IEKF
SNR = 10 dB
SPKF

−0.1
CFO value

SOEKF
EKF
−0.15

−0.2

−0.25

20 40 60 80 100 120
Sample Index

Figure 2.6: Recursive CFO estimation for a joint CFO/channel estimation using
optimal filtering, when the noise statistics are available

and the SOEKF to estimate the CFOs.

Let us now look at the H∞ based approaches. Figure 2.7 shows the robustness of
the H∞ filtering to uncertainties. For an error of 5dB over the variance, the H∞
filtering approaches converge faster than the Kalman ones. However, the com-
putational complexity of the H∞ filtering approaches is higher than the SPKF.
It should be noted that when using the extended H∞ and the unscented H∞ ap-
proaches, the choice of the noise attenuation level Ξ plays a key role. If it is set
to a high value, there is no real difference between the Kalman algorithm and
the H∞ based approach, whereas no H∞ solution may exist if Ξ is set to a small
value.
Based on the above considerations, we think that among our algorithms the
SPKF gives the best compromise in terms of estimation performance, computa-
tional complexity and number of parameters to be a priori tuned.
The grid search algorithms proposed in [Pun 04a] and [Pun 06] provide better
performance in terms of MMSE. Nevertheless, the estimation error of the SPKF
is small enough to guarantee bit error rates (BERs) that are similar to the ones
obtained when the grid search algorithms are used, as shown in table 2.3.

61
2. Frequency Synchronization and Channel Equalization of OFDMA
Uplink Systems

0.1

SNR = 10 dB
0.08
5 dB noise variance error
EKF
SPKF
0.06
extended H∞ filter
unscented H∞ filter
CFO value

0.04

0.02

0
CFO real value

−0.02
20 40 60 80 100 120
Sample Index

Figure 2.7: Comparison between the Kalman filtering approaches and the H∞
filtering approaches in terms of convergence speed, when the noise statistics are
not available

SNR (dB) theoretical BER BER (grid search) BER (SPKF)

0 1.464 × 10−1 1.469 × 10−1 1.584 × 10−1
5 6.418 × 10−2 6.452 × 10−2 6.681 × 10−2
−2 −2
10 2.327 × 10 2.361 × 10 2.395 × 10−2

Table 2.3: BER performance when a joint channel/CFO estimation is considered

2.3.4.2 Computational Complexity

One of the goals in the design of wireless communication systems is to reduce the
energy consumption of the system. In current wireless communication systems,
only a few Go/s are dedicated to the channel estimation and the synchroniza-
tion. So, in the following let us have a look at the computational cost of the EM
algorithm and the SPKF algorithm. These approaches provide similar results in
terms of BER, but the EM-based algorithm have a higher computational com-
plexity due to the exhaustive grid search over the possible range of CFOs and the
iterative estimation.
Table 2.4 shows the number of arithmetic operations performed by the EM; N ob

62
 2.3 A Joint CFO/Channel Estimator for OFDMA Systems

denotes the number of observations for the algorithms and it is set to N ob = K.

The computational cost of the EM approach depends on the number of Iem it-
erations and on the number of b tested values. The M-step is based on the
inversions of U matrices of size Lu × Lu . In addition, the decision test has to be
done to decide which is the best value of the CFO. Table 2.5 shows the number

EM algorithm
E-step
additions and
Iem Nob U + Iem U (2Lu + 6) 1.1200 × 104
subtractions
multiplications
Iem Nob U + Iem U (U + 2Lu ) 1.1040 × 104
and divisions
M-Step
add./sub. for
Iem bN ob U Lu + Iem bU Lu (1 + Lu ) − Iem bU 6.2271 × 107
the grid search
mult./div. for Iem bU Nob (3U Lu + Lu )
8.0040 × 107
the grid search + Iem bU (L2u + 15U + 2Lu + 1)
other
Iem N ob U Lu + Iem U L2u 1.1366 × 106
add./sub.
other
Iem N ob U (3U Lu + Lu ) + Iem U (L2u + 15U + Lu ) 3.9329 × 106
mult./div.
Total arithmetic operations 1.4740 × 108

Table 2.4: Number of arithmetic operations performed by the EM joint chan-

nel/CFO estimator, 20001 tested CFO values and 20 iterations.

of arithmetic operations performed by the SPKF. The first step of the algorithm,
corresponding to the selection of the sigma points requires the Cholesky decom-
position1 of a matrix of size U × U. As the observation vector length is equal to
2, the measurement update also requires the inversion of a 2 × 2 matrix.
Table 2.6 shows the number of Giga-operations per second (Go/s) performed by
the EM for different grid search precisions.
Table 2.7 shows the number of Go/s for different numbers of users in the system
with a channel composed of 3 multi-paths. In addition, table 2.7 shows the num-
ber of Go/s for different number of channel multipaths with 4 users in the system.
1
When using the SPKF and by considering U the length of the state vector, a U×U Cholesky
decomposition is necessary to obtain the “square root” of the error covariance matrix.

63
2. Frequency Synchronization and Channel Equalization of OFDMA
Uplink Systems

SPKF algorithm
step 1: calculation of the sigma points
additions and
2N ob U 2 (4L2u + 4Lu + 1) 2.0070 × 105
subtractions
multiplications
2N ob U 2 (4L2u + 4Lu + 1) 2.0070 × 105
and divisions
step 2: estimation update
additions and 2N ob U (8U 2 L3u + 12U 2 L2u + 6U 2 Lu
6.2290 × 106
subtractions 12U L2u + U 2 + 12U Lu + 3U + 2Lu +1)
multiplications N ob U (16U 2 L3u + 24U 2 L2u + 12U 2 Lu
6.1286 × 106
and divisions 20U L2u + 2U 2 + 20U Lu + 5U + 4Lu + 2)
step 3: measurement update
additions and N ob U (44U Lu + 19U + 13)
1.8348 × 105
subtractions +N ob U (28U L2u + 96Lu )
multiplications 2N ob U (12U Lu + 7U +27) + 2N ob
2.9651 × 105
and divisions +2N ob U (4U L2u + 50Lu )
Total arithmetic operations 1.3239 × 107

Table 2.5: Number of arithmetic operations performed by the SPKF joint chan-
nel/CFO estimator

The computational complexity of the EM increases faster when the number of

users (or the number of channel multipaths) increases. Indeed the EM-based al-
gorithms work in blocks in an iterative way whereas the SPKF is recursive.
It is clearly seen by the results, that the computational complexity of the EM
is higher than the one of the SPKF. When 4 users in the system and a channel
composed of 3 multipaths are considered, the SPKF algorithm requires only 9%
of the number of operations required by the EM.
The approach presented in [Pun 06] required less iterations than the one pro-
posed in [Pun 04a]. In addition, an improvement to those iterative architectures
has been proposed [Sezg 08]. Here, the authors propose to reduce the number
of the iterations and the complexity of the matrix inversions. However, if only
one iteration is considered using a grid search precision of 10−4 , the grid search
architecture required 7.1378 × 107 Go/s, the value is still higher than the one pre-
sented in table 2.5. Due to the grid search, the above approaches implementation
is relatively difficult in real environments;

64
 2.3 A Joint CFO/Channel Estimator for OFDMA Systems

grid search precision b Go/s

−1
10 21 0.0066
−2
10 201 0.0194
−3
10 2001 0.1474
−4
10 20001 1.3239

Table 2.6: Number of Go/s performed by the EM joint channel/CFO estimator

for different grid search precisions

4 users in the system 3 multi-paths

Multipaths Go/s(EM) Go/s(SPKF) Number of users Go/s(EM) Go/s(SPKF)
1 0.5349 0.0130 2 0.6751 0.0192
2 0.9780 0.0516 3 1.0417 0.0588
3 1.4276 0.1324 4 1.4276 0.1324
4 1.8835 0.2712 5 1.8328 0.2505
5 2.3459 0.4838 6 2.2573 0.4238
6 2.8147 0.7859 7 2.7011 0.6626
7 3.2899 1.1931 8 3.1643 0.9777
8 3.7715 1.7213 9 3.6468 1.3794
9 4.2596 2.3861 10 4.1486 1.8784
10 4.7540 3.2033 11 4.6698 2.4852
11 5.2549 4.1885 12 5.2102 3.2103

Table 2.7: Number of Go/s performed by the EM and the SPKF joint chan-
nel/CFO estimators

2.3.5 Conclusions
The architectures proposed in this section require control data to perform the
estimation, but they do not need an initialization step. The robustness to un-
certainties of the H∞ filtering has been confirmed, but the selection of the noise
attenuation level remains quite difficult. When the noise characteristics are avail-
able, the best estimation performance have been obtained by the IEKF among
the proposed estimators. However, the computational complexity is the high-
est among the proposed algorithms. SPKF gives the best compromise between
computational complexity and estimation performance. They have lower compu-
tational cost than the grid search methods. This advantage is crucial, because
one of the goals in the design of wireless systems is to reduce the energy con-

65
2. Frequency Synchronization and Channel Equalization of OFDMA
Uplink Systems

sumption of the system. In current wireless communication systems, only a few

Go/s are dedicated to the channel estimation and the synchronization. These
proposed approaches are completely applicable to practical environments, unlike
the EM-based methods.
In the next subsection, to reduce the amount of control data in the transmitted
signal, we propose to combine the optimal filtering techniques with the so-called
MMSE-SD in an iterative way.

2.4 An OFDMA Non-Pilot Aided Iterative CFO

Estimator
As an alternative to this classic scheme, shown in figure 2.8 and in order to reduce
the amount of control data, a non-pilot aided estimator based on an iterative
architecture [Pove 11c] is proposed in this section.
In an OFDMA transmission, the OFDMA frame duration may be higher than
the coherence time Tc of the channel. In this case, the CSIs need to be updated
by using pilot subcarriers. In this section, we suggest using a shorter OFDMA
frame so that the CSIs do not change. The subcarriers are no longer necessary
to update the CSI and hence they can be removed. The data rate is increased,
but the CFO may vary from one symbol to another. It should be noted that the
preamble is kept in order to perform the time synchronization, to estimate the

: data sub-carriers
: pilot sub-carriers
Tc : coherence time of the channel

Tofdma : time duration of the OFDMA frame Fourier transform of the

frequency (Hz)
pth OFDMA symbol
Information
known training sequence
to be retrieved

1 st OFDMA Symbol 2 nd OFDMA Symbol 3 rd OFDMA Symbol (M − 1)th OFDMA Symbol M th OFDMA Symbol

time (s)
Tofdma << Tc

Figure 2.8: Example of an OFDMA frame composed of M OFDMA symbols

with a preamble composed of two OFDMA symbols and a determined number of
pilot subcarriers

66
 2.4 An OFDMA Non-Pilot Aided Iterative CFO Estimator

CSI for the whole frame and to provide an initial estimation of the CFO. See
figure 2.9.

2.4.1 Problem Formulation

: data sub-carriers

Tc : coherence time of the channel

Fourier transform of the frequency (Hz)
Tofdma : time duration of the OFDMA frame pth OFDMA symbol
Information
known training sequence to be retrieved

1 st OFDMA Symbol 2 nd OFDMA Symbol R(2) R(M − 2) R(M − 1)

time (s)
Tofdma << Tc

Figure 2.9: Example of an OFDMA frame composed of M OFDMA symbols

with a preamble composed of two OFDMA symbols and no pilot subcarriers, where
R is the received OFDMA symbol

The idea of the proposed receiver is to correct the MAI and to estimate the
signals sent from all users in the system by using an MMSE-SD. These estimated
signals called the “MMSE-SD preambles” play the role of the preambles and
are used to estimate the CFOs by means of the optimal filter. The iterative
architecture is shown in figure 2.10.
The frequency synchronization in an uplink communication can be seen as a
multiparameter estimation problem. A solution is to distinguish one user from
the others before starting the synchronization process. After the estimation of the
U “MMSE-SD preambles”, each one is inserted in U optimal filtering processes,
i.e. U Kalman filters or U H∞ filters. See figure 2.11. Finally, the CFOs between
the BS and each user are estimated independently, thanks to the optimal filter
and an additional MAI cancellation process.

67
2. Frequency Synchronization and Channel Equalization of OFDMA
Uplink Systems

^
A ^
1,1 1,1
R Optimal filter
MMSE-SD ^ ^
iteration 1 A U,1 based
U,1
CFO estimation

^ ^
A 1,2 1,2
MMSE-SD Optimal filter
^
A based ^
iteration 2 U,2
U,2
CFO estimation

^ ^
A 1,i 1,i
MMSE-SD Optimal filter
^
A based ^
iteration i U,i
U,i
CFO estimation

Figure 2.10: Proposed OFDMA non-pilot aided iterative receiver using optimal
filtering

2.4.2 User Detection - The Minimum Mean Square Error

Successive Detector (MMSE-SD)
The first step of the iterative architecture is to estimate the preambles by using
the MMSE-SD. According to [Hou 08], the MMSE-SD is robust against near-
far effect1 , produced by the strong MAI. These phenomena are induced by the
difference that may exist between two users in terms of the propagation loss.
Instead of a joint multiuser decoding, a combination between an MMSE pre-
detection scheme with an ordered successive detection is proposed. The detection
of the transmitted OFDMA signal components operates in two major steps:

• interference cancellation (IC): during this step, the previous ”detected”

OFDMA signal components are subtracted out of the received signal. In-
deed, let Ŝũ,i be the estimation of the signal sent by the ũth user at the ith
1
The near-far effect is a condition in which a strong power received signal does not allow
the receiver to detect a weak power received signal.

68
 2.4 An OFDMA Non-Pilot Aided Iterative CFO Estimator

MMSE-SD iteration i
S^ 1,i ^ ^
R A 1,i 1,i
W1,i Q1 sign G1,i
+ -
^
Y2,i S^ 2,i ^
A 2,i Kalman 2,i
W2,i Q2 sign G2,i based
+ -
CFO
estimation
+ -
^
S ^
A
^
YU,i U,i U,i U,i
WU,i QU sign GU,i

Figure 2.11: Iterative MMSE-SD using a Kalman filtering based estimator and
a MAI suppression for BPSK modulation

iteration of the MMSE-SD, where 1 ≤ i ≤ Isd and Isd denotes the max-
imum iteration number. The user’s decoding order is denoted as ũ where
ũ ∈ {1, . . . , U }; when ũ = 1, it is associated with the maximum signal to
interference-plus-noise ratio1 (SINR) among the users whereas ũ = U rep-
resents the user with the lowest SINR. In addition the so-called (ũ + 1)th
order MMSE-SD residual at the ith iteration Yũ+1,i ∀ũ 6= 1 is the differ-
ence between the received signal R and the components transmitted by the
detected users2 (namely those corresponding to the ũ highest SINRs) and
Y1,i = R ∀i. See figure 2.11.
The IC steps can be summarized as follows:

Yũ+1,i = Yũ,i − Ĝũ,i Ŝũ,i (2.12)

ũ
X
= R− Ĝl,i Ŝl,i (2.13)
l=1

( P
U
ũ=1 Êũ,i Hũ Qũ if ũ = 1
with Ĝũ,i = (2.14)
Ĝũ−1,i (IK − Qũ−1 ) if 2 ≤ ũ ≤ U
1
As the BS calculates the different SINRs, the signal can be decoded in an ordered way.
2
Given (1.21), (1.22) and (1.23), the received OFDMA symbol can be expressed as
PU
R = u=1 Gu Su + B, where Gu = Eu FH Hu Qu , See chapter 1 in section 1.4.2.2.

69
2. Frequency Synchronization and Channel Equalization of OFDMA
Uplink Systems

where Hũ = FH Hũ , Hũ is the channel frequency response matrix associated
to the ũth user1 , Qũ is the CAS matrix associated to the ũth user defined in
(1.23), and Ŝũ,i is the vector at the ith iteration that contains the estimated
symbols sent by the ũth user:
h i
Ŝũ,i = Ŝũ,i (0), Ŝũ,i (1), . . . , Ŝũ,i (k), . . . , Ŝũ,i (K − 1) (2.15)
nh io
Êũ,i = diag 1, ej2πǫ̂ũ,i−1 /K , . . . , ej2π(K−1)ǫ̂ũ,i−1 /K and ǫ̂ũ,i is the estimation
of the CFO associated with the ũth user at the ith iteration. A new iteration
begins when all the users have been processed and when the estimation ǫ̂ũ,i ,
using the recursive estimator approach proposed in subsection 2.4.3, has
been done.

• interference suppression (IS): this step aims at removing the interference

stemming from the as-yet undecoded components. The purpose of this step
is hence to filter the ũth order MMSE-SD residual Yũ,i . By denoting σs2 the
signal power allocated on each of the subcarriers, the suppression weight
matrix Wũ,i for the selected ũth user at the ith iteration satisfies [Hou 08]:
!−1
σB2
Wũ,i = 2
IK + G H
ũ,i Gũ,i GH
ũ,i (2.16)
σs

Then, (2.16) is used to decode the selected user and to obtain the estimated
signal sent by uth user4 .

Ŝũ,i (k) = argmin kTũ,i (k) − Ωm k2 (2.17)

Ωm ∈Ω

with : Tũ,i = Qũ Wũ,i R ũ = 1

Tũ,i = Qũ Wũ,i Yũ,i ũ > 1 (2.18)
Tũ,i = [Tũ,i (0), Tũ,i (1), . . . , Tũ,i (k), . . . , Tũ,i (K − 1)]
where Ω = {Ω1 , Ω2 , . . . Ωm , . . . ΩM } is the modulation constellation. At
that stage, (2.17) makes it possible to obtain the estimated signal of the
ũth user.
h i
Âũ,i = Hũ Qũ Ŝũ,i = Âũ,i (0), Âũ,i (1) . . . Âũ,i (K − 1) (2.19)
1
The matrix Hũ is assumed to be known.
4
For example in BPSK modulation Ŝũ,i = sign (Qũ Wũ,i Yũ,i )

70
 2.4 An OFDMA Non-Pilot Aided Iterative CFO Estimator

In the next subsection, those estimated signals the “MMSE-SD preambles” are
used to estimate the CFO of each user.

2.4.3 CFO Estimation with the MMSE-SD Preambles

In order to restore orthogonality among each user subcarrier, the synchronization
error vector ǫ defined in (2.4) has to be estimated. Here, we suggest a recursive
method. However, each CFO ǫu is estimated independently. In each recursion,
the results of these independent estimations are used together for MAI cancella-
tion.
Let us introduce the state-space model to estimate the CFO, that is the rep-
resentation of what happens during one OFDMA symbol. It is defined by the
following state-space model:
State equation:
ǫũ,i (n) = ǫũ,i (n − 1) + w(n) (2.20)

Measurement equation:
ǫũ,i n
R̂ũ,i (n) = Âũ,i (n)ej2π K + Bũ (n)
(2.21)
= f (ǫũ,i (n)) + Bũ (n)

where R̂ũ,i (n) is the estimated received signal from each user (see subsection
2.4.3.1). In addition, w and Bũ are white zero-mean gaussian noises with variances
assumed to be σw2 and σũ2 respectively. It should be noted that σw2 is very low and
can even be equal to 0. In this latter case, the CFO is assumed to be constant
during one OFDMA symbol.
Let us introduce a third state-space equation that is used in the general case
for H∞ filtering to focus on a linear combination of the state-vector components.
Here, as the dimension of the state vector is equal to 1, one has:

zũ,i (n) = Lǫũ,i (n) (2.22)

where L is equal to 1.
As the state-space model is non-linear due to (2.21), the study of several kinds
of algorithms is proposed:

71
2. Frequency Synchronization and Channel Equalization of OFDMA
Uplink Systems

• the EKF approaches including the EKF, the SOEKF and the IEKF, pre-
sented in appendix D,

• the SPKF, including the UKF and the CDKF, presented in appendix E,

• the “extended H∞ filter” and the “unscented H∞ filter” presented in ap-

pendix F.

2.4.3.1 Multiple Access Interference Cancellation Process

To improve the estimation, an MAI cancellation strategy as in [Zhao 06] is pro-

posed. The results from the (n−1)th recursion are used to estimate and eliminate
different users’ signals in the nth recursion.

(est)
MAI estimation: R̂ũ,i (n) = Âũ,i (n)ej2πnǫ̂ũ,i (n−1)/K (2.23)
U
X (est)
MAI correction: R̂ũ,i (n) = R(n) − R̂ũ,i (n) (2.24)
j=1,j6=ũ

After some recursions, the algorithm makes its possible to estimate the value of
the uth user CFO.

2.4.4 Simulation Results

In the following, a comparative study is carried out. Three kinds of tests are
considered to evaluate the performance of the algorithms.

Simulation protocol: An OFDMA IEEE 802.16 WirelessMANTM uplink sys-

tem, which involves U = 4 users sharing K = 512 subcarriers and with a cyclic
prefix set to Ng = K/8 ≥ max
u
(Lu ) is considered. The carrier frequency is at
fc = 2.6GHz and the bandwidth is set to W = 20MHz. Tof dma << Tc , where Tc
is the coherence time of the channel and Tof dma is the OFDMA frame duration.
A transmission over a Rayleigh slow-fading frequency-selective channel composed
of Lu = 3 ∀u multipaths is supposed. QPSK is used to modulate the information
2
bits. Then, Eb /No = 10log( σσu2 ) where σu2 is the mean power of the received signal
ũ
from the ũth user.

72
 2.4 An OFDMA Non-Pilot Aided Iterative CFO Estimator

When using H∞ filtering: Ξ = 102 , V = σũ2 and W = σw2 . One assumes that
there is a state noise w(n) with very small variance, e.g. and σw2 = 10−3 . It is
assumed that the channel estimation and a CFO pre-estimation have been per-
formed using a preamble. The algorithm estimates the “MMSE-SD preambles”
of the 4 users, and then estimates the CFOs for each OFDMA symbol so that the
coherent detection can be robust against variations over all the OFDMA frame.
1.2
Eb/N0 = 10 dB Doppler User 4

0.8
Eb/N0 = 0 dB
Estimated doppler

0.6

Doppler User 1
0.4

Doppler User 3
0.2

0 Doppler User 2

−0.2
0 50 100 150 200 250 300 350 400 450 500
Sample index

Figure 2.12: CDKF based approach for CFO estimation with a known preamble

2.4.4.1 Test 1: Recursive Estimation Using Perfectly Estimated MMSE-

SD Preambles

In this first case, it is assumed that the MMSE-SD preambles have been perfectly
estimated at the receiver. The CFO estimation algorithms for users with different
CFO are tested. The users’ CFO estimation errors are considered fixed during
one OFDMA symbol. 500 montecarlo runs are performed.
Based on preliminary tests, it was noticed that the Kalman filtering approaches,
“the extended H∞ filter” and the “unscented H∞ filter” provide very similar
results. In the following, we choose the SPKF and the unscented H∞ filter to
perform the CFO estimation, in order to avoid the linearization step. Therefore,
in figure 2.12, we only show the results obtained when using a CDKF. According

73
2. Frequency Synchronization and Channel Equalization of OFDMA
Uplink Systems

0.8

0.6
Iteration 5
Estimated doppler

CFO
0.4 Iteration 4
0.2

0
Iteration 3
−0.2 Iteration 2
Iteration 1
−0.4

−0.6

0 50 100 150 200 250 300 350 400 450 500

Sample index

Figure 2.13: MMSE-SD combined with an UKF based approach for CFO esti-
mation

to figure 2.12, our approach makes it possible to accurately estimate the CFO
recursively. It should be noted that this is the recursive CFO estimation over
only one OFDMA symbol.

2.4.4.2 Test 2: Non-Pilot Aided Estimation

In this second case, the receiver estimates the MMSE-SD preambles. The simu-
lation results deal with the first user of the system. The users’ CFO errors vary
between the different OFDMA symbols (but they are considered fixed during one
OFDMA symbol). They are modeled as independent zero-mean Gaussian ran-
2
dom variables with a variance of σcf o . Figure 2.13 shows the results in terms of
CFO estimation over one OFDMA symbol. The proposed algorithm provides an
estimation of the CFO in an iterative way. As expected, the first iteration leads
to poor performance, but iterating our approach especially up to 5 times leads to
good performance.
In figure 2.14, the performance of our algorithm in terms of BER are shown; the
estimation is done by using either the Kalman filtering with an error of 3 dB over
the variance or the H∞ filtering. In the following, we choose “the unscented H∞

74
 2.4 An OFDMA Non-Pilot Aided Iterative CFO Estimator

−1
10

BER

H∞ filter - iteration 1
H∞ filter - iteration 2
H∞ filter - iteration 3
H∞ filter - iteration 4
H∞ filter - iteration 5
−2
10 No CFO
KF - iteration 5
CFO perfectly estimated

4 6 8 10 12 14 16
Eb/N0 (dB)

Figure 2.14: BER performance, uniformly distributed CFO, 5 iterations, when a

non-pilot aided estimation is considered

filter” to perform the CFO estimation, in order to avoid the linearization step. In
addition, we show the results when a single-user detector is used by considering
a perfectly estimated CFOs. According to figure 2.14, the results of the proposed
method are better after performing the fourth iteration. A gain of around 1 dB for
a BER of 10−2 is obtained when using the proposed method with the “unscented
H∞ filter” at the fifth iteration in comparison with the single-user detector. In
[Huan 05] and [Cao 04a], the algorithms tend to the theoretical value. However,
in [Huan 05] more subchannels than users are required. In addition, in [Cao 04a]
a training sequence is necessary.
In figure 2.15, the results in terms of MMSE are shown. Based on previous tests,
it was seen that the results in [Mova 08] are better than the results in [Cao 04b]
in terms of CFO estimation performance. So, we propose a comparative study
with the multiuser interference resilient (MUI) approach [Cao 04b]. Note that
the algorithm presented in [Cao 04b] needs two OFDMA symbols for the estima-
tion. A gain of 4 dB for an MMSE of 10−4 is obtained between the MUI method
and the proposed method at the third iteration when using a SPKF. In addition,
figure 2.16 compares the SPKF and the “unscented H∞ filter” in terms of con-
vergence speed, for a Eb /No = 10 dB, for different errors over the noise variance

75
2. Frequency Synchronization and Channel Equalization of OFDMA
Uplink Systems

−3
10

KF iteration 3
KF iteration 4
KF iteration 5
MUI

−4
10
MMSE

−5
10

−6
10
4 6 8 10 12 14 16
Eb/N0 (dB)

Figure 2.15: CFO estimation performance for a non-pilot aided CFO estimation

when using the SPKF and for different values of Ξ when using the H∞ filtering.
Firstly, one considers an error over the variance of 3dB and 5dB; secondly, no
error over the variance is assumed. It can be seen that the convergence speed of
the SPKF is affected by the variance noise error. For the lowest values of Ξ, the
convergence may be slightly faster. If the value of Ξ increases, the “unscented
H∞ ” filter performance tend to be the same as the SPKF ones.

2.4.4.3 Test 3: Influence of the CIR Estimations

In this third case, we propose to analyze the influence of the CSIs variations over
the receiver performance. The CFO estimation is performed by a SPKF, provided
that the noise characteristics are known. The simulation results are focused on
the first user of the system. It is assumed that the CSIs have small variations in
time over the OFDMA frame. ĥu denotes the estimated channel impulse response
for each user using the preamble and hũ is the value of the channel response for
each user. In a first test, the following normalized random error over each channel
is considered |ĥu − hũ | < ILu ×1 10−2 , where ILu ×1 is a Lu × 1 vector composed of
ones. In the second test |ĥu − hũ | < ILu ×1 10−1 . The results obtained at the fifth
iteration for the different BER with a different error over the channels are shown

76
 2.4 An OFDMA Non-Pilot Aided Iterative CFO Estimator

in figure 2.17. One can notice that the performance of the architecture are not

really affected by small variations on the channel.

0.4

0.35

0.3

SPKF-3dB noise variance error

Estimated doppler

0.25 SPKF-5dB noise variance error

SPKF
H∞ filter Ξ = 20
0.2 H∞ filter Ξ = 102
CFO

0.15

0.1

0.05

0
0 50 100 150 200 250 300 350 400 450 500
Sample Index

(a)
0.4

0.35

CFO H∞ filter Ξ = 20

H∞ filter Ξ = 102
Estimated doppler

0.3
SPKF with a
an estimation error over the
SPKF with a noise variance of 5 dB
an estimation error over the
noise variance of 3 dB
0.25

SPKF with a
an estimation error over the
0.2 noise variance of 0 dB

20 40 60 80 100 120 140 160 180 200 220

Sample Index

(b)

Figure 2.16: Comparison between the SPKF and the “unscented H∞ filter” in
terms of convergence speed

77
2. Frequency Synchronization and Channel Equalization of OFDMA
Uplink Systems

−1
10

Theoretical value
Channel perfectly estimated
Channel estimation error < 10−2
Channel estimation error < 10−1
BER

−2
10

4 6 8 10 12 14 16
Eb /No (dB)

Figure 2.17: BER performance when the channel is estimated and a non-pilot
aided CFO estimation is considered

2.4.4.4 Computational Complexity

Now let us look at the complexity of the frequency synchronization scheme. Con-
cerning with the MMSE-SD step, at each iteration and for each user, the MMSE-
SD is based on the inversion of a matrix of size K × K in order to calculate the
suppression weight matrix Wũ,i . This matrix inversion requires O(K 3 ) opera-
tions. Thus, the overall complexity of the MMSE-SD step is O(Isd U K 3 ).
Then, concerning the estimation step, as the dimension of the state vector is equal
to 1, the optimal filtering step requires the computation of O(1) operations for
each user and for each iteration. Thus, the overall complexity of the estimation
step is O(Isd U ).
Therefore, we can conclude that the overall complexity of the proposed iterative
architecture is O(Isd U K 3 ).

2.4.5 Conclusions
The non-pilot aided MAI suppression scheme for an OFDMA uplink transmission
followed by the CFO estimation algorithm based on optimal filtering can estimate

78
 2.5 An OFDMA Robust CFO Estimator

and detect each user’s CFO and frame respectively, with no need of pilot sub-
carriers. This leads to a maximum transmission data rate.
In addition, simulation results confirm that the proposed scheme can effectively
suppress the MAI caused by a relatively large CFO, sufficiently robust to CFO
variations. The decoding of OFDMA signal is an ordered serial processing that
combines interference suppression and interference cancellation techniques. The
iterative decoding applies simple hard interference cancellation techniques, result-
ing in moderate complexity. The CFO estimation performance of the different
approaches give quite similar results. On the one hand, one advantage of the
SPKF is that it does not need the calculations of Jacobians or Hessians. In those
cases, the computational complexity is lower. On the other hand, when using the
H∞ filter, no information about the noise statistics is required, but the selection
of the noise attenuation level may be a problem.

2.5 An OFDMA Robust CFO Estimator

The above algorithms provide accurate estimation as long as the received signal
is only disturbed by the propagation channel and an AWGN, whereas significant
degradations are expected in the presence of an NBI [Mare 06].
When CR systems are used, the PU has to be detected. When dealing with mul-
ticarrier transmissions and if the cognitive devices have no a priori information
about the PU signal, energy detection has been suggested as the spectrum sensing
technique [Ferr 10], [Quan 08], [Cabr 04]. Nevertheless, when an energy detector
is used, deep PU fading must be counteracted to guarantee the PU detection
[Cabr 04]. Otherwise, the PU signal may be disturbed by an undesirable NBI
caused by the CR that shares the same frequency band. This may lead to an
inaccurate estimation of the CFO and hence increases the BER. Therefore, NBI
has to be detected and taken into account to deduce the CFO.
The NBI detection has been widely studied [Broe 03], [Li 07]. In particular, the
approach proposed in [Li 07] consists of a repetitive pilot block in the time domain
to detect an NBI localized in time. Hence, this method cannot be used in the
presence of CFO. Since the two duplicated parts remain identical after passing
through the channel, their subtraction at the receiver provides the contribution

79
2. Frequency Synchronization and Channel Equalization of OFDMA
Uplink Systems

of the noise-plus-interference, which is then easily detected.

To our knowledge, the CFO estimation in presence of an NBI has not been widely
addressed. In [More 08], Morelli et al. propose a novel scheme that jointly esti-
mates the CFO and the interference power on each subcarrier. They take advan-
tage of an IFFT property to estimate the CFO, by transmitting pilot symbols
over some specific subcarriers while setting the others to zero. By performing
the IFFT step, a block composed of several identical parts is obtained and used
to estimate the CFO. However to perform the frequency error estimation, an ex-
haustive grid search is considered over the possible range of the CFO, leading to
a high computational cost. In addition, the CFO estimation is very sensitive to
the number of subcarriers disturbed by the NBI.
This section shows how to jointly estimate the PU-CFOs and to detect the cog-
nitive radio-NBI (CR-NBI) in an OFDMA PU-system, damaged by a CR-NBI
localized in time [Pove 11d], [Pove 11b]. The approach is again based on the
SPKF, with an additional step using the innovation energy to detect a variation
of the measurement noise covariance matrix [Foki 09], [Bian 05]. Knowing the
innovation distribution, two algorithms are proposed to detect the CR-NBI. The
first one is based on a binary hypothesis test (BHT), whereas the second con-
sists in comparing the derivate of a cumulative sum (CUSUM) to a threshold. In
addition, an alternative way to use these tests is proposed.

2.5.1 System Description

First, let us consider a multicarrier CR system network with an available band-
width W cr divided among K cr subcarriers. In addition, the uplink PU-OFDMA
system consists of a single BS and U simultaneously independent users. Then, the
available bandwidth W pu = αW cr of the PU-OFDMA system is divided among
K pu
K pu ≤ K cr subcarriers, with 0 < α < K cr
≤ 1. The PU and CR bandwidths are
shown in figure 2.18.
In the following, let us associate the subscript u to the uth user in the PU-
OFDMA system, with u ∈ {1, . . . , U }. Let Supu (mpu , k pu ) be the emitted PU-
symbol with a time duration T pu at the k pu th subcarrier associated to the mpu th
PU-OFDMA symbol, where M pu is the number of emitted PU-OFDMA symbols.

80
 2.5 An OFDMA Robust CFO Estimator

Figure 2.18: PU and CR system spectrum, where PUA and PUB denote the
frequency bands used by other PUs

Each PU-OFDMA symbol is composed of K pu PU-symbols and has a time dura-

pu pu
tion1 Tof pu pu
dma . See figure 2.19. In addition, Su,kpu (m ) = 0 if the k th subcarrier
is not assigned to the uth user.
Remark: All the subcarriers are assumed to be used. For this reason, the num-
ber of samples in one OFDMA symbol corresponds to the number of subcarriers.

The PU-channel impulse response is defined as follows:

hupu=[hpu pu pu pu pu pu
u (0), hu (1), . . . , hu (l ), . . . , hu (Lu − 1)] (2.25)

where Lpu
u is the length of the maximum channel delay spread.
To help the reader, the PU system parameters are summarized in table 2.8.
At the receiver, due to the propagation conditions, time offset and CFO are
induced in the signal. The nth sample of the mth PU-OFDMA received symbol
after time synchronization and cyclic prefix removal can be written as follows
(See chapter 1 in section 1.4.2.2):
pu −1
KX
ǫu n kpu n
pu j2π K
Ru (m , n) = e pu
Supu (mpu , k pu )Hupu (k pu )ej2π K pu (2.26)
kpu =0

1 pu
It should be noted that Tof dma is the time duration of one PU-OFDMA symbol and is not
the time duration of one PU-OFDMA frame.

81
2. Frequency Synchronization and Channel Equalization of OFDMA
Uplink Systems

Description Parameter
Bandwidth W pu
Symbol time T = W1pu
pu

Number of subcarriers K pu
Number of emitted OFDMA symbols M pu
Subcarrier index k ∈ {0, . . . K pu − 1}
pu

OFDMA emitted symbol index mpu ∈ {0, . . . M pu − 1}

Emitted symbol at the k pu th subcarrier pu pu
Su,k pu (m )
associated to the mpu th PU-OFDMA symbol
pu pu pu
OFDMA symbol time Tof dma = K T
Channel delay spread index lupu∈ {0, . . . , Lpu
u − 1}
lpu th coefficient of
hpu
u,lpu
the channel impulse response
Maximum channel delay spread length Lpu
u
pu
Channel frequency response Hu,k pu

Table 2.8: PU system parameters for the uth user

where n ∈ {0, 1, . . . , K pu − 1}, ǫu is the normalized PU-CFO and

PLpu pu pu
u −1 pu pu −j2π l K pu
k
Hupu (k pu ) = lpu =0 hu (l )e is the channel frequency response associ-
pu
ated to the k th subcarrier.
The normalized PU-CFOs row vector ǫ is defined in (2.4). Let us now define the
row vector that contains the PU-channel impulse responses of each user, respec-
tively:
hpu = [h1pu , h2pu , . . . , hupu , . . . , hUpu ] (2.27)

Concerning the CR-signal, the analog baseband signal emitted I(t) by the
CR-antenna, without cyclic prefix, at time t is given by:
cr −1
KX
cr (t−mcr T cr
I(t) = S cr (mcr , k cr )ej2πfk of dm ) (2.28)
kcr =0

where Skcrcr (mcr ) is the emitted CR-symbol with a time duration T cr at the k cr th
subcarrier associated to the mcr th CR-OFDM symbol, where M cr is the number
of emitted CR-OFDM symbols. Moreover, Skcrcr (mcr ) = 0 if the spectrum sensing
decides that the k cr th subcarrier is busy and mcr Tof
cr cr cr
dm ≤ t < (m + 1)Tof dm .
cr cr
Tof dm is the CR-OFDM symbol time and fk is the baseband frequency of the
k cr th subcarrier.
At the PU-BS antenna, the CR-received signal after the convolution with the

82
 2.5 An OFDMA Robust CFO Estimator

Description Parameter
Bandwidth W cr
Symbol time T = W1cr
cr

Number of subcarriers K cr
Number of emitted OFDM symbols M cr
Subcarrier index k cr ∈ {0, . . . K cr − 1}
OFDM emitted symbol index mcr ∈ {0, . . . M cr − 1}
cr
Emitted symbol at the k th subcarrier
Skcrcr (mcr )
associated to the mcr th CR-OFDM symbol
cr cr cr
OFDM symbol time Tof dm = K T
Baseband frequency of the subcarrier cr k
fk = T cr = K crkT cr
of dm

Channel delay spread index lcr ∈ {0, . . . , Lcr − 1}

lcr th coefficient of
hcr
lcr
the channel impulse response
Maximum channel delay spread length Lcr
Channel frequency response Hkcrcr

Table 2.9: CR system parameters

Description Parameter
Bandwidth W cr = W pu /α
Symbol time T cr = αT pu
pu pu
OFDM symbol time Tof dm = γTof dma = α(K cr /K pu )Tof
cr
dma

Table 2.10: Relations between PU and CR system parameters

channel can be expressed as:

cr −1
LX cr −1
KX
cr (t−mcr T cr cr pu )
V (t) = cr
h (l )cr
S cr (mcr , k cr )ej2πfk of dm −l T (2.29)
lcr =0 kcr =0

where hcr
lcr are the coefficients of the CR-channel impulse response,
lcr ∈ {0, 1, . . . , Lcr − 1} and Lcr is the length of the maximum channel delay
spread.
Once again, to help the reader, the CR system parameters are summarized in
table 2.9. In addition, table 2.10 shows the relations between the PU and the
CR-system parameters.
The received signal is sampled at the sampling frequency:

1
fs = (2.30)
T pu

83
2. Frequency Synchronization and Channel Equalization of OFDMA
Uplink Systems

Then, using (2.29), (2.30) and tables 2.8-2.10, the unknown downsampled
CR-NBI1 produced by a fault of spectrum sensing detection can be expressed
as follows:
LX −1 cr KX −1 cr
n cr n cr pu
V (mcr , n) = Z(mcr Tof
cr
dm + )= hcr (lcr ) S cr (mcr , k cr )ej2πfk ( fs −l T )
fs lcr =0 kcr =0
cr −1
KX cr −1
LX
kcr lcr T pu kcr nT pu
cr cr cr
= S (m , k ) hcr (lcr )e−j2π K cr T cr ej2π K cr T cr
kcr =0 lcr =0
cr −1
KX
kcr n
= S cr (mcr , k cr )Hkcrcr ej2π αK cr
kcr =0
(2.31)
PLcr −1 cr cr
cr cr cr cr −j2π kαKlcr
where H (k ) = lcr =0 h (l )e represents the unknown CR-channel
frequency response associated to the k cr th subcarrier.
Thus, the PU-OFDMA received signal at the BS satisfies:
(
f (mpu , n, ǫ, hpu ) + V (mcr , n) + B(mpu , n) if n1 ≤ n ≤ n2
R(mpu , n) =
f (mpu , n, ǫ, hpu ) + B(mpu , n) otherwise
(2.32)
PU
where f (mpu , n, ǫu , hupu ) = u=1 Ru (mpu , n) and B(mpu , n) is a zero-mean com-
plex AWGN with variance σB2 . In addition, n1 ∈ {0, 1, . . . , K pu (1 − γ)} is the
first sample of the CR-NBI and n2 = n1 + γK pu − 1 is the last sample of the
CR-NBI, with 0 ≤ γ ≤ 1. The time-domain representation of the signals is shown
in figure 2.19. To have a clearer idea, a time-frequency domain representation of
the received signal at the BS is shown in figure 2.20.

2.5.2 Joint Disturbance Detection and CFO Estimation

Now, we suggest jointly detecting the CR-NBI and estimating both the PU-CFOs
and channels. A preamble of one OFDMA symbol is used for the channel/CFOs
estimation. It is assumed that the system is synchronized in time and that the
PU-channels and the PU-CFOs do not vary over the OFDMA symbols of the
OFDMA frame.

1
CRs may likely use wide bands. However, due to the fault of spectrum sensing, a CR-NBI
is produced over some subcarriers allocated to the PU.

84
 2.5 An OFDMA Robust CFO Estimator

Figure 2.19: Time representation of the PU and CR-NBI signals

Figure 2.20: Time-frequency representation of the mpu th PU-OFDMA received

symbol at the BS

85
2. Frequency Synchronization and Channel Equalization of OFDMA
Uplink Systems

First of all, let us define the following state vector:

h iT
x(n) = ǫ(n) Re {hpu (n)} Im {hpu (n)} (2.33)

that satisfies the following state equation:

x(n) = x(n − 1) ∀n (2.34)

Secondly, let Y(n) be the observation column vector that stores the real and the
imaginary parts of the PU-OFDMA received signal R(mpu , n):
(
C(n, x(n)) + V(n) + B(n) if n1 ≤ n ≤ n2
Y(n) = (2.35)
C(n, x(n)) + B(n) otherwise

where
h
C(n, x(n)) = Re {f (n, ǫ, hpu )} Im {f (n, ǫ, hpu )} ]T ,
h
B(n) = Re {B(n)} Im {B(n)} ]T is a 2 × 1 white Gaussian noise vector with co-
h
variance matrix (σB2 /2)I2 and V(n) = Re {V (n)} Im {V (n)} ]T is the
2 × 1 zero-mean Gaussian CR-NBI vector with covariance matrix (σV2 /2)I2 . The
CR-NBI signal can be considered as a zero-mean Gaussian vector due to the large
number of subcarriers K cr [Huan 10b].
The SPKF provide “accurate” estimations as long as the received signal is only
disturbed by the AWGN [Pove 10], whereas significant degradations are expected
in the presence of the CR-NBI. To maintain the estimation performance, it is
proposed to jointly detect the CR-NBI and estimate the PU-CFOs.
The approach operates for the nth sample in three steps:

1. estimating the state vector x̂(n|n) by means of SPKF from the observations
{Y(0), Y(1), . . . , Y(K pu − 1)}

2. calculating the innovation energy ||Ỹ(n)||2 where Ỹ(n) = Y(n) − Ŷ(n),

h n o n o iT
Ŷ(n) = Re R̂(mpu , n) Im R̂(mpu , n) = C(n, x̂(n|n)) is the estima-
tion of Y(n),

3. testing whether there is CR-NBI or not by using the innovation energy.

Indeed, ||Ỹ(n)||2 increases much when there is CR-NBI. See figure 2.21.
This gap is due to the received signal variance that jumps from σB2 to
σB2 + σV2 .

86
 2.5 An OFDMA Robust CFO Estimator

2
In the following, let us define the SNR expressed in dB as SNR= 10 log( σσ2u ),
B
pu 2
where E{|R(m , n)| } = σu2 is the mean power of the received signal from
the uth user. In addition, we define the signal-to-interference ratio (SIR)
2
expressed in dB as SIR = 10 log( σσ2u ).
V

Remark: If the CR-NBI is assumed to appear at the nth sample, the esti-
mated state vector at time n − 1 and its covariance matrix are kept stored:
x̂nu = x̂(n − 1|n − 1) and Pnu = P(n − 1|n − 1). They will be used as a new
starting point for the CFO and channel estimation once the CR-NBI is considered
to have disappeared. During the CR-NBI perturbation, SPKF provides an esti-
mation of the state vector that is not reliable, but useful to detect the presence
of the CR-NBI.
In the next subsections, two tests to detect the CR-NBI are presented.

40

35

30

25
n1 : beginning n2 : end
||Ỹ(n)||2

of the CR-NBI of the CR-NBI

20

15

λ(n): the threshold

10

0
50 100 150 200 250 300 350 400 450 500
Sample Index

Figure 2.21: Time representation of the innovation energy

87
2. Frequency Synchronization and Channel Equalization of OFDMA
Uplink Systems

2.5.2.1 Combining SPKF and Binary Hypothesis Test

(SPKF-BHT)

To decide whether there is CR-NBI in the OFDMA symbol or not, the following
binary hypothesis is tested:
(
H0 : C(n, x(n)) + V(n) + B(n)
(2.36)
H1 : C(n, x(n)) + B(n)

The complex residual R̃(mpu , n) is a zero-mean Gaussian process with variance

n o
tr PYY (n) where PYY (n) denotes the innovation covariance matrix obtained
by the SPKF at the nth sample. Then, the probability of false alarm can be
defined as:

Pf a = P (|R̃(n)|2 > λ(n)|H1 ) = P (||Ỹ(n)||2 > λ(n)|H1 )

q
= 2P (||Ỹ(n)|| > λ(n)|H1 ) (2.37)
q
= 2(1 − P (||Ỹ(n)|| < λ(n)|H1 ))
q
= 2(1 − cdf ( λ(n)))

where λ(n) is a threshold at the nth sample set by the practitioner and cdf (.) is
the cumulative density function (CDF) of (.).
Using ||Ỹ(n)|| the properties and the CDF of the Rayleigh distribution1 , one has:
n o 2
λ(n) = tr PYY (n) ln( ) (2.38)
Pf a

If ||Ỹ(n)||2 ≤ λ(n), no CR-NBI is assumed to be present. When ||Ỹ(n)||2 > λ(n),

the CR-NBI is supposed to be at the nth sample. The CR-NBI detection and
the PU estimation algorithm is summarized in figure 2.22.

2.5.2.2 Combining SPKF and CUSUM Test (SPKF-CT)

As an alternative to the BHT, a change detection algorithm known as the CUSUM

test [Bass 93] is also proposed to be combined with the SPKF. The CUSUM test
is an accumulative sum of a data sequence. It aims at detecting abrupt changes in
q
1
R̃(n) is Gaussian distributed, so ||Ỹ(n)|| = Re{R̃(n)}2 +Im{R̃(n)}2 is Rayleigh dis-
λ(n)
p −
tributed. Then cdf ( λ(n)) = 1 − e tr{PYY (n)} .

88
 2.5 An OFDMA Robust CFO Estimator

BHT inputs:

2
SPKF x̂(n) Measurement Ŷ(n) 2
||Ỹ(n)|| OK OK
Y(n) estimation
||.|| < λ(n) T update
+ -
NOT OK NOT OK
2
T : ||Ỹ(n − 1)|| < λ(n − 1)

update
Figure 2.22: SPKF-BHT algorithm

the mean of the sequence. When using the CUSUM, if the mean increases much,
the cumulative sum also increases. It has been used in various applications such
as information sciences [Huan 11] and biomedicine [Yang 10].
To decide wheter the CR-NBI is present or not, we now suggest using the upper
control limit of the CUSUM test defined as follows:
C + (n) = max(0, C + (n − 1)) + ||Ỹ(n)||2 − tr {PỸỸ (n)}
(2.39)
= max(0, C + (n − 1)) + ỸH (n)Ỹ(n) − E{ỸH (n)Ỹ(n)}

Given (2.39) and fixing a threshold, the CUSUM test makes it possible to detect
when the CR-NBI begins, but does not give information when it ends. See figure
2.23.
Thus, it is proposed to look at the slope of the curve C + (n):
δ(C + (n))
= C + (n) − C + (n − 1) (2.40)
δn

Two cases can be considered.

1/ Let us assume that C + (n − 1) > 0; if the PU-OFDMA symbol is disturbed by
the CR-NBI, one has:
n o
tr PYY (n) << ||Ỹ(n)||2 (2.41)
Given (2.39) and (2.40), this leads to:
δ(C + (n))
≈ ||Ỹ(n)||2 (2.42)
δn
Otherwise, if the PU-OFDMA is not disturbed by the CR-NBI:
n o
tr PYY (n) ≈ ||Ỹ(n)||2 (2.43)

89
2. Frequency Synchronization and Channel Equalization of OFDMA
Uplink Systems

hence one has:

δ(C + (n)) n o
= ||Ỹ(n)||2 − tr PYY (n) ≈ 0 (2.44)
δn

2/ Let us now assume that C + (n − 1) ≤ 0. Then if the PU-OFMDA symbol is

disturbed by the CR-NBI and given (2.39), (2.40) and (2.41) one has:

δ(C + (n))
= ||Ỹ(n)||2 − C + (n − 1) > ||Ỹ(n)||2 (2.45)
δn

Otherwise if there is no CR-NBI distortion, given (2.39), (2.40) and (2.43) one
has:
δ(C + (n)) n o
= ||Ỹ(n)||2 − tr PYY (n) + C + (n − 1)
δn (2.46)
≈ C + (n − 1)

δ(C + (n))
Therefore, one has to compare δn
to a threshold in order to detect the end of
the CR-NBI in the PU-OFDMA symbol. The CR-NBI is assumed to be present
δ(C + (n))
provided that δn
is higher than the threshold.
Concerning the choice of the threshold, we suggest the one defined in (2.38).

2000

1500

n2 : end
n1 : beginning of the CR-NBI
C + (n)

of the CR-NBI
1000

500

0
50 100 150 200 250 300 350 400 450 500
Sample Index

Figure 2.23: Time representation of C in the CUSUM test

90
 2.5 An OFDMA Robust CFO Estimator

2.5.2.3 Improving the Disturbance Detection

In this subsection, the goal is to reduce the gap between the CR-NBI probability
of detection Pcr 2
d = P (||Ỹ(n)|| > λ(n)|H0 ) and the PU probability of detection
Ppu 2
d = P (||Ỹ(n)|| < λ(n)|H1 ), in order to obtain better estimation performance
when using the SPKF-BHT and the SPKF-CT. If the CR-NBI probability of
pu
detection Pcr
d increases, the PU probability of detection Pd decreases. Therefore,
there are less samples available to perform the CFO estimation. This hence
may lead to poor estimation of the CFO. The PU probability of detection Pdpu
determines the values that are not disturbed by the CR-NBI and that are going
to be used in the estimation task.
When looking at figure 2.21, the values of ||Ỹ(n)||2 are not always higher than the
threshold between n1 and n2 sample indexes. In addition, the values of ||Ỹ(n)||2
may sometimes exceed the threshold outside the interval [n1 , n2 ]. To improve the
CR-NBI and the PU detections when using the BHT, we propose to take into
account the evolution of ||Ỹ(n)||2 over several samples. Therefore, we suggest
defining a new residual value ||Ỹmean (p)||2 as follows:

1 β(p+1)−1
X
||Ỹmean (p)||2 =
||Ỹ(q)||2 (2.47)
j pu k β q=βp
where p ∈ {0, . . . P − 1}, P = Kβ and ⌊.⌋ is the integer part of {.}.
j k
K pu K pu K pu
If β
6= β
, i.e. β
∈
/ Z, the last sample of the new residual value is:
KX −1 pu
2 1
||Ỹmean (P)|| = pu ||Ỹ(q)||2 (2.48)
K − βP q=βP

If the value of ||Ỹmean (p)||2 exceeds the value of the threshold, the algorithm
decides that the CR-NBI is present during the corresponding samples. See algo-
rithm 3.
Based on the same idea, to improve the detection performance when using the
j pu
k
CUSUM test, we propose to keep only P − P − Kβ (i.e P or P + 1) samples
as follows:
+
Cmean (p) = C + (βp − 1) (2.49)
+ K pu
with p ∈ {1, . . . P − 1}, Cmean (0) = C + (0) and if β
∈
/ Z then
+
Cmean (P) = C + (K pu − 1). See figure 2.24.

91
2. Frequency Synchronization and Channel Equalization of OFDMA
Uplink Systems

Algorithm 3 Calculate ||Ỹmean (p)||2

Require: ||Ỹ(q)||2 , K pu , β
K pu
for p = 0 to β
− 1 do
Pβ(p+1)−1
||Ỹmean (p)||2= β1 q=βp ||Ỹ(q)||2
end for j k
pu pu
if Kβ 6= Kβ then
j k PK pu −1
K pu
||Ỹmean ( )||2= K pu −β1 K pu q=β ⌊ Kβ
pu ||Ỹ(q)||2
β ⌊ β ⌋ ⌋
else
end if

+ K pu
Figure 2.24: Cmean (p + 1) and C + (n) with β ∈Z

+
δ(Cmean (p)) C + (p)−C + (p−1)
If the value of δp
= β
exceeds the value of the threshold, the
algorithm decides that the CR-NBI is present in the corresponding samples.
In the next subsection, simulation results show the relevance of our algorithms
and their improvements to jointly detect the CR-NBI and estimate the PU-CFOs.

2.5.3 Simulation Results

This section presents computer simulations to validate the performance of the
SPKF-BHT and the SPKF-CT estimators. Firstly, the simulation protocol is
presented. Secondly, simulations present how to choose the parameter β pre-
sented in subsection 2.5.2.3. Finally, the last subsection shows a comparative
study in terms of CFO estimation with already existing estimators [More 08], the
influence of CR-NBI power on the CFO estimation performance and the channel
estimation performance of the proposed algorithms.

92
 2.5 An OFDMA Robust CFO Estimator

Simulation protocol: A PU-OFDMA WirelessMANTM system, which is com-

posed of U = 4 users sharing K pu = 512 subcarriers and with a cyclic prefix
Ng = K pu /32 ≥ max
u
(Lpu
u ) is considered. QPSK is used to modulate the infor-

mation bits. The carrier frequency is at fc = 2.6GHz and the bandwidth is set
to W pu = 5MHz. A CR-NBI affects K nbi = 40 subcarriers of the PU-OFDMA
system bandwidth. It is supposed a transmission over a normalized Rayleigh
slow-fading frequency-selective channel composed of Lpu
u = 7 ∀u multipaths.

For very low values of SIR, as ||Ỹ(n)||2 >> 0 during the CR-NBI presence, its
detection can be easy. In addition for high values of SIR, i.e SIR> 0dB the dif-
ference between σB2 and (σB2 + σV2 ) may be negligible and would not affect the
Kalman filter estimation performance. Therefore the case when SIR= −10 dB is
studied. The CR-NBI can begin anywhere in the PU-OFDMA received symbol.
For comparison reasons with [More 08], the CR-NBI is supposed to affect only
the subcarriers assigned to one PU.

SPKF-BHT
SPKF-CT

0.8
CR-NBI Probability of detection

0.75

0.7

0.65

1 2 3 4 5 6 7 8
β

Figure 2.25: CR-NBI detection, SNR= 0 dB, SIR= −10 dB, Pf a = 0.05 and
γ = 0.5

93
2. Frequency Synchronization and Channel Equalization of OFDMA
Uplink Systems

2.5.3.1 How to Improve the Disturbance Detection?

This subsection presents how to select a “well suited” value of β, which is the
parameter presented in subsection 2.5.2.3, to reduce the gap between the CR-NBI
pu
probability of detection Pcr
d and the PU probability of detection Pd .
Firstly the impact of β in the CFO estimation performance is shown. The SNR
is set to 0 dB and the Pf a is set to 0.05.
Let us consider that the CR-NBI disturbs half of the PU-OFDMA symbol
(γ = 0.5). Figure 2.25 shows the CR-NBI probability of detection Pcr
d for dif-
ferent values of β. We clearly see that if β increases, the CR-NBI probability of
detection Pcr
d increases and the CFO estimation should normally be improved.
However, when looking at figure 2.26 we can see that the PU probability of de-
tection Ppu
d does not increase with the value of β as the CR-NBI probability of
detection Pcr
d . For both proposed algorithms, there are “well suited” values of β
in terms of PU probability of detection. For the SPKF-BHT, the “well suited”
value of β is between 4 and 6 whereas the “well suited” value of β is between 2
and 3 when using the SPKF-CT. The CR-NBI probability of detection Pcr
d and
the PU probability of detection Ppu
d has been increased, then better estimation
performance should be obtained by using this optimal β parameter.

1
SPKF-BHT
SPKF-CT

0.995
PU Probability of detection

0.99

0.985

0.98

1 2 3 4 5 6 7 8
β

Figure 2.26: PU detection, SNR= 0 dB, SIR= −10 dB, Pf a = 0.05 and γ = 0.5

94
 2.5 An OFDMA Robust CFO Estimator

Tables 2.11 and 2.12 show the CFO estimation performance in terms of MMSE
for different values of β and duration of the CR-NBI over the symbol. It should
be noted that the duration of the CR-NBI over the PU-OFDMA symbol does not
change the influence of β in the algorithm. For different durations of the CR-NBI
over the symbol, we can see that better performance in terms of CFO estimation
are obtained for values β between 4 and 6 for the SPKF-BHT. In addition, it can
be seen that the “well suited” value of β in terms of probability of detection is
between 2 and 3 for the SPKF-CT.

SNR = 0 dB SIR = −10 dB

γ 0.47060 0.5333 0.6154
−3 −3
β =1 1.124 × 10 1.257 × 10 2.103 × 10−3
β =2 8.517 × 10−4 9.526 × 10−4 1.593 × 10−3
−4 −4
β =3 7.949 × 10 8.891 × 10 1.487 × 10−3
β =4 5.735 × 10−4 6.414 × 10−4 1.073 × 10−3
β =5 5.962 × 10−4 6.668 × 10−4 1.115 × 10−3
−4 −4
β =6 5.678 × 10 6.350 × 10 1.062 × 10−3
β =7 6.814 × 10−4 7.621 × 10−4 1.274 × 10−3
β =8 7.382 × 10−4 8.256 × 10−4 1.381 × 10−3

Table 2.11: SPKF-BHT CFO estimation performance for different values of γ

and β, Pf a = 0.05

SNR = 0 dB SIR = −10 dB

γ 0.47060 0.5333 0.6154
−3 −3
β = 1 1.569 × 10 1.755 × 10 2.934 × 10−3
β = 2 7.471 × 10−4 8.356 × 10−4 1.397 × 10−3
β = 3 8.218 × 10−4 9.191 × 10−4 1.537 × 10−3
−3 −3
β = 4 1.494 × 10 1.671 × 10 2.795 × 10−3
β = 5 1.644 × 10−3 1.838 × 10−3 3.074 × 10−3
β = 6 1.718 × 10−3 1.922 × 10−3 3.214 × 10−3
−3 −3
β = 7 1.793 × 10 2.005 × 10 3.353 × 10−3
β = 8 1.868 × 10−3 2.089 × 10−3 3.493 × 10−3

Table 2.12: SPKF-CT CFO estimation performance for different values of γ and
β, Pf a = 0.05

95
2. Frequency Synchronization and Channel Equalization of OFDMA
Uplink Systems

In figure 2.27, we can see the performance in terms of CR-NBI probability of

detection. Like the previous tests, better performance are obtained when the
BHT is used. When β = 1 and Pf a = 0.05, the CR-NBI probability of detection
for the SPKF-BHT is higher than the CR-NBI probability of detection for the
SPKF-CT. The difference is around 0.08. In addition, when Pf a = 0.05, the
CR-NBI probability of detection for the SPKF-BHT β = 6 is higher than the
CR-NBI probability of detection for SPKF-CT β = 2. The difference is around
0.15. Again it can be seen that the detection is improved when using β 6= 1; it
confirms the results shown in tables 2.11 and 2.12. Figure 2.28 shows the de-
tection performance in terms of PU probability of detection. In this test, better
detection performance are obtained with the CUSUM test.
It should be noted that the choice of an appropriate Pf a is a key point in the
development of the algorithms.

0.85

0.8
SIR = −10 dB
SNR = 0 dB
0.75
CR-NBI probability of detection

0.7

0.65

0.6

0.55
Improved SPKF-CT β = 1
Improved SPKF-BHT β = 1
0.5
Improved SPKF-CT β = 2
Improved SPKF-BHT β = 6
0.45

0.01 0.02 0.03 0.04 0.05 0.06 0.07 0.08 0.09

Probability of false alarm

Figure 2.27: CR-NBI detection, SNR= 0 dB, SIR= −10 dB and γ = 0.5

2.5.3.2 Comparative Study: Influence of the Disturbance Power over

the CFO Estimation and Channel Estimation Performance

Figure 2.29 shows the estimation performance in terms of MMSE for different
durations of the CR-NBI over the PU-OFDMA symbol. The SNR is again set to

96
 2.5 An OFDMA Robust CFO Estimator

0.995

0.99

PU probability of detection
0.985

0.98
SIR = −10 dB
SNR = 0 dB
0.975

0.97
Improved SPKF-CT β = 1
Improved SPKF-BHT β = 1
0.965
Improved SPKF-CT β = 2
Improved SPKF-BHT β = 6
0.96

0.955
0.01 0.02 0.03 0.04 0.05 0.06 0.07 0.08 0.09
Probability of false alarm

Figure 2.28: PU detection, SNR= 0 dB, SIR= −10 dB and γ = 0.5

MMLE[11]
−1
10 SPKF-BHT β = 1
SPKF-BHT β = 6
MMSE

−2
10

−3
10

−4
10
0.3 0.4 0.5 0.6 0.7 0.8
Pf a=0.05 γ
SIR= −10dB
SNR= 0 dB
MMLE[11]
−1 SPKF-CT β = 1
10
SPKF-CT β = 2
MMSE

−2
10

−3
10

−4
10
0.3 0.4 0.5 0.6 0.7 0.8
γ

Figure 2.29: Robust CFO estimation performance, SIR= −10 dB, SNR= 0 dB
and Pf a = 0.05

0 dB. We clearly see that the proposed algorithms provide better CFO estimation
performance since γ < 0.6 in comparison with [More 08].
Figure 2.30 shows the CFO estimation performance in terms of MMSE. The CR-
NBI disturbs half of the PU-OFDMA symbol (γ = 0.5).

97
2. Frequency Synchronization and Channel Equalization of OFDMA
Uplink Systems

In comparison with the modified ML estimator (MMLE) proposed in [More 08],

an improvement of 2 dB is obtained using the SPKF-CT using β = 1 and 7 dB
is obtained when using the SPKF-BHT using β = 6, for an MMSE of 10−3 .

Normal KF

SIR = −10 dB
Pf a = 0.05
MMLE[12]
−3
10
MMSE

−4
10

SPKF-CT β = 1
SPKF-BHT β = 1
SPKF-CT β = 2
−5 SPKF-BHT β = 6
10

0 2 4 6 8 10 12 14 16 18
SNR

Figure 2.30: Robust CFO estimation performance, SIR= −10 dB, Pf a = 0.05
and γ = 0.5

Table 2.13 shows the estimation performance in terms of MMSE for different
values of SIR. It should be noted that the proposed algorithms are not disturbed
by the power of the CR-NBI.

MMSE
SNR= 0 dB
SIR −30 dB −20 dB −10 dB 0 dB
SPKF 0.43520 0.08150 0.003557 0.00172
SPKF-CT β = 1 0.0028500
SPKF-BT β = 1 0.0012670
SPKF-CT β = 2 0.0008217
SPKF-BT β = 6 0.0003958

Table 2.13: Robust CFO estimation performance, Pf a = 0.05 and γ = 0.5

Since ĥpu is determined by the values ǫ̂, one can conclude that the proposed
estimators achieve accurate channel estimation. Figure 2.31 confirms the channel

98
 2.6 Conclusions

estimation performance of the algorithm. We can see a gain of 12 dB between

the SPKF-CT using β = 1 and the SPKF-BHT using β = 6.

−3
x 10
6.5
SPKF-CT β = 1
6 SPKF-BHT β = 1
SPKF-CT β = 2
5.5 SPKF-BHT β = 6

4.5
MMSE

3.5

2.5
SIR = −10 dB
Pf a = 0.05
2

1.5
0 2 4 6 8 10 12 14 16 18
SNR

Figure 2.31: Robust channel estimation performance, SIR= −10 dB, Pf a = 0.05
and γ = 0.5

2.6 Conclusions
In this chapter we proposed to jointly estimate the CFOs and the channels for
an OFDMA uplink system by using a recursive approach. On the one hand, the
EKF, the SOEKF, the IEKF and the SPKF including the UKF and the CDKF
were considered. On the other hand, to relax the assumptions on the model and
measurement noise, the “extended H∞ filter” and the “unscented H∞ filter” filter
were studied.
Then, the above techniques have been combined with an MMSE-SD to obtain a
non-pilot aided CFO estimator for an OFDMA uplink. In addition, as the SPKF
give the best compromise between computational complexity and estimation per-
formance, it has been chosen to be combined with two statistical tests: the BHT
and the CUSUM test, to design two OFDMA robust CFO estimators.
Both SPKF-based algorithms are used to estimate the CFOs and the channels of
different users in a PU-OFDMA system disturbed by a localized in time CR-NBI.

99
2. Frequency Synchronization and Channel Equalization of OFDMA
Uplink Systems

The proposed estimators operate in the time domain unlike other approaches that
estimate the CFO in presence of an NBI in the frequency domain [More 08].
By using a SPKF and without increasing the computational complexity, we take
advantage of the innovation to detect a sharp variation of the measurement noise
covariance matrix when there is CR-NBI. The SPKF enable the state vector to
be estimated and the innovation energy to be calculated. Then, a test detects
the variation of the measurement-noise covariance matrix by comparing the in-
novation energy with a threshold set by the practitioner.
When using the BHT, better performance are obtained in terms of the CR-NBI
probability of detection. However, when using the CUSUM, better performance
in terms of PU probability of detection are obtained. To reduce the gap between
the CR-NBI probability of detection and the PU probability of detection, a tech-
nique is then proposed. It consists in improving both probabilities of detection
by watching the evolution of the innovation energy.
Our method outperforms other existing methods [More 08]. However, the pro-
posed estimators are sensitive to the duration in time of the CR-NBI.
The next chapter still deals with the constraints addressed in this chapter, i.e.
frequency synchronization and channel equalization, but in the context of the
uplink OFDM-IDMA systems.

100
Chapter 3
Frequency Synchronization and Channel
Equalization of OFDM-IDMA Uplink
Systems

Contents

3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . 102

3.2 OFDM-IDMA Modified Receiver . . . . . . . . . . . . 105
3.2.1 Multiple CFO and Channel Estimation . . . . . . . . 107
3.2.2 Modified IDMA Receiver . . . . . . . . . . . . . . . . 109
3.2.3 Simulation Results . . . . . . . . . . . . . . . . . . . . 111
3.2.4 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . 115
3.3 Space-Time Block Code OFDM-IDMA Receiver . . 115
3.3.1 System Description . . . . . . . . . . . . . . . . . . . . 116
3.3.2 Multiple CFO and Channel Estimation . . . . . . . . 119
3.3.3 Maximum Ratio Combining Operations . . . . . . . . 119
3.3.4 Modified Elementary Signal Estimator . . . . . . . . . 120
3.3.5 Simulation Results . . . . . . . . . . . . . . . . . . . . 122
3.4 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . 124

101
3. Frequency Synchronization and Channel Equalization of
OFDM-IDMA Uplink Systems

3.1 Introduction
The OFDM-IDMA technique initially proposed by Mahafeno et al. [Maha 06]
has been introduced in chapter 1.
Firstly, OFDM-IDMA can be seen as a special case of multicarrier CDMA
(MC-CDMA). IDMA inherits many advantages from CDMA, in particular, its
robustness against different kinds of interferences. In addition, the autocorrela-
tion properties of the IDMA transmitted signal make it possible to take advantage
of the multipath fading channel diversity1 . The key principle of IDMA is that
the spreading code is the same for all users, unlike CDMA2 . The user distinction
is done by using different interleavers. They are generated independently and
randomly, allowing BER to be near the theoretical values of single-user systems.
Secondly, OFDM-IDMA can be seen as a special case of OFDMA. OFDM-IDMA
inherits the robustness to ISI from OFDMA. In addition, the multicarrier scheme
allows the frequency allocation algorithms to improve the spectral efficiency by
exploiting the spectral diversity. The main difference between these techniques
is that the entire bandwidth can be allocated to a single user for all the OFDM-
IDMA frame duration. This may lead to a very high single-user capacity. Indeed,
if we assume an OFDMA system with U users and K subcarriers all used, each
K
user can transmit U
symbols at the same time. Now if one also consider an
OFDM-IDMA system with a spreading factor3 S = log2 (U ), each user can trans-
K K
mit S
= log2 (U )
symbols at the same time. We clearly see that log2 (U ) < U ; thus
the users in the OFDM-IDMA system can transmit more symbols at the same
time than in an OFDMA system.
Furthermore, an OFDM-IDMA scheme may improve the duty cycle of a single-
user transmission4 in comparison with an OFDMA system. Indeed, in an OFDMA
1
In chapter 1 in section 1.3, we show that the received signal is a superposition of several
delayed and attenuated copies of the transmitted signal, due to the multipath fading channel.
This diversity may be exploited by a combination of the different copies using a rake receiver
[Proa 95].
2
In chapter 1 in section 1.4.3.1, we present that the ENC of the OFDM-IDMA transmitter
incorporates a spreading code.
3
S = log2 (U ) is the lowest spreading factor to allow U users transmit at the same time.
4
The duty cycle of a single-user transmission is defined as the ratio between the duration
of one user frame and the time between two successive frames.

102
 3.1 Introduction

system, the maximum number of users allowed by the system to transmit at the
same time is U = K. In that case, each user transmits one symbol at the same
time. In an OFDM-IDMA system, the maximum number of users allowed by
the system to transmit at the same time is U = 2S . In addition, to transmit
one symbol at the same time, it is necessary that K = S. Then, the maximum
number of users allowed by an OFDM-IDMA system to transmit one symbol at
the same time is U = 2K . We clearly see that 2K > K; the OFDM-IDMA enables
more users to perform a transmission at the same time.
Today, the OFDM-IDMA scheme is not used as the multiple access technique by
mobile wireless standards, but for the above reasons, it can be seen as a promising
technique for the future mobile wireless standards.
To become one of the multiple access schemes of future mobile standards, several
constraints have to be addressed, such as the channel equalization. Nevertheless,
few papers deal with this issue in IDMA systems. Thus, Zhou et al. develop an
iterative channel estimation process that uses pilot subcarriers [Zhou 07]. The
authors propose two different approaches: the first one is based on an unbiased LS
estimator whereas the second is based on the maximum ratio combining (MRC)
method. In [Song 09], the authors propose an EM estimation method to obtain
the CSIs. More specifically, in [Suya 08], Suyama et al. insert a least mean
square (LMS) estimator in the iterative process of an OFDM-IDMA receiver. In
addition, the authors in [Rehm 08] present a low-complexity uplink channel es-
timation. By taking advantage of the channel coding, the transmitter creates a
training sequence in order to estimate the channel through a linear detection at
the receiver.
However, in the above approaches, a perfectly frequency synchronized system
is considered. In wireless communication systems with multicarrier modulation,
this assumption is usually not true. Therefore, a CFO estimation/correction has
to be considered at the receiver.
Several approaches of the frequency synchronization in multicarrier systems have
been proposed in the last years. More particularly, in MC-CDMA systems, Chien
et al. present a blind recursive estimation of the CFOs by using an EM algo-
rithm [Chie 07]. In [Thia 07], the authors take advantage of the knowledge of the

103
3. Frequency Synchronization and Channel Equalization of
OFDM-IDMA Uplink Systems

second-order statistics of the received signal. They suggest a two-stage CFO esti-
mation. The first one involves a grid search to obtain the initialization of the CFO
estimations, whereas the second consists of a CFO estimation based on the LMS
algorithm. In [Chen 10], the CFOs are estimated by searching the peaks of the
correlation function of the received signal. In addition, the reader can look at the
frequency synchronization techniques for OFDMA systems, [Pun 04a], [Pun 06]
and [Sezg 08] presented in chapter 2.
In the above methods, the CFO correction can be performed before or after the
CP-OFDM demodulation. The CFO correction methods can take advantage of
the orthogonal codes in MC-CDMA systems, whereas the CFO correction meth-
ods usually utilize the CAS1 in OFDMA systems. The CFO estimation/correction
issue in the particular OFDM-IDMA scheme becomes a challenging task2 . In
[Cao 04a], Cao et al. propose to correct the CFO in an OFDMA uplink system
after the CP-OFDMA demodulation through a linear detection3 , without taking
advantage of the CAS. This CFO correction method can be completely applied
to OFDM-IDMA systems. However, CFOs and CSIs are assumed to be perfectly
known. Although frequency synchronization has been addressed in MC-CDMA
and OFDMA, it has never been studied in an OFDM-IDMA system to our knowl-
edge.
Given the above considerations, our contribution in this chapter is twofold:

1. We propose two SPKF based receivers: the first one is a modification of

the OFDM-IDMA receiver proposed by [Maha 06], whereas the second uses
a STBC to take advantage of the space diversity. The proposed OFDM-
IDMA receivers aim at jointly estimating the CFO and the channel of each
user in the system, by using a preamble of one OFDM-IDMA symbol.
1
In chapter 2 in section 2.2, we present several CFO correction methods that take advantage
of the CAS such as the single-user detector method and the methods presented in [Choi 00]
and [Huan 05].
2
Let us recall that in an OFDM-IDMA system all users may use the same spreading code
and a subcarrier is allocated to all users at the same time.
3
In subsection 3.2, we show that after the CP-OFDM demodulation, the received OFDM-
IDMA symbol can be expressed as a multiplication between a function depending on the CFO
and a function depending on the symbols and channel frequency responses.

104
 3.2 OFDM-IDMA Modified Receiver

2. Both receivers operate in two steps. Firstly, the OFDM demodulation is

performed without any correction over the received signal. Then, the re-
sulting signal is inserted in the new versions of the IDMA receiver. Finally,
taking into account the interference produced by the CFO and knowing all
the CFOs and the CSIs, the IDMA iterative process is able to cancel both
the ICI and the MAI, and to recover the sent symbols.

3.2 OFDM-IDMA Modified Receiver

Like the OFDMA uplink system, the signal received at the BS is affected by
several CFOs and CSIs in an OFDM-IDMA uplink system. The nth sample of
received OFDM-IDMA symbol after time-synchronization and CP removing can
be expressed as:
U
X LX
u −1
ǫu n
R(n) = ej2π K hu (l)Xu (n − l) + B(n) 0 ≤ n ≤ K − 1 (3.1)
u=1 l=0

where we recall that hu (l) is the lth coefficient of the CIR, ǫu is the normalized
CFO, Xu (n) is the nth sample of the transmitted OFDM-IDMA symbol and B(n)
is a zero-mean AWGN with variance σB2 .
Then, the received OFDM-IDMA symbol can be expressed as follows:
∆
R = [R(0), R(1), . . . , R(n), . . . , R(K − 1)]T
U
X (3.2)
= Eu hu Xu +B
u=1
| {z }
Ru

nh io
where Eu = diag 1, ej2πǫu /K , . . . , ej2π(K−1)ǫu /K is the CFO matrix, hu is the
channel circulant matrix defined in (1.31), Xu is the transmitted OFDM-IDMA
symbol defined in (1.21) and B is a zero-mean AWGN vector with covariance
matrix σB2 IK (See chapter 1 in section 1.4.3.2).
Given (1.20), it should be noted that (3.1) can be rewritten as follows:

U K−1
X ǫu n X k′ n
R(n) = ej2π K Hu (k ′ )Su (k ′ )ej2π K + B(n) 0 ≤ n ≤ K − 1 (3.3)
u=1 k′ =0

105
3. Frequency Synchronization and Channel Equalization of
OFDM-IDMA Uplink Systems

PL−1 k′ l
where Hu (k ′ ) = l=0 hu (l)e−j2π K is the channel frequency response associated
to the k ′ th subcarrier and Su (k ′ ) is the symbol transmitted over the k ′ th subcar-
rier.
Given (3.3), although the effect of the CFO of one user or another can be elim-
inated, they cannot be suppressed all at the same time before the CP-OFDM
demodulation. Whatever the strategy chosen by the receiver to compensate the
CFO, there are still some interferences. Therefore, we suggest introducing the
CFO correction after the CP-OFDM demodulation.
After the CP-OFDM demodulation, the received symbol r(k) can be expressed
as follows:
K−1
X kn
r(k) = R(n)e−j2π K + b(k)
n=0
K−1
X X U K−1
X (k′ −k)n
j2π ǫK
un
= e Hu (k ′ )Su (k ′ )ej2π K + b(k)
n=0 u=1 k′ =0
U K−1
(3.4)
X X K−1
X (k′ −k+ǫu )n
j2π
= e K Hu (k ′ )Su (k ′ ) + b(k)
u=1 k′ =0
n=0
XU K−1
X
= F(k ′ − k + ǫu )Hu (k ′ )Su (k ′ ) + b(k)
u=1 k′ =0

where1
sin(xπ) jπ x(K−1)
F(x) = e K (3.5)
sin( xπ
K
)

and
K−1
X kn
b(k) = B(n)e−j2π K (3.6)
n=0

Then, (3.4) can be rewritten as:

U
X U K−1
X X
r(k) = F(ǫu )Hu (k)Su (k) + F(ǫu + k ′ − k)Hu (k ′ )Su (k ′ ) +b(k)
u=1 u=1 k′ =0 (3.7)
k′ 6=k
| {z }
ICI(k)

1
PK−1 nx 1−ej2πx sin(xπ) jπ x(K−1)
Let us recall that n=0 ej2π K = j2π x = sin( xπ e K .
1−e K K )

106
 3.2 OFDM-IDMA Modified Receiver

If we look at the received signal from the uth user point of view, (3.7) can be
expressed as a summation of a useful term and an interference term as follows:

U
X
r(k) = F(ǫu )Hu (k)Su (k) + F(ǫu′ )Hu′ (k)Su′ (k) + ICI(k) + b(k)
u′ =1
u′ 6=u
U
X (eq) (3.8)
= Hu(eq) (k)Su (k) + Hu′ (k)Su′ (k) + ICI(k) + b(k)
u′ =1
u′ 6=u

= Hu(eq) (k)Su (k) + ςu (k)

where
Hu(eq) (k) = F(ǫu )Hu (k) ∀u ∈ {1, . . . , U } (3.9)
U
X (eq)
ςu (k) = Hu′ (k)Su′ (k) + ICI(k) + b(k) (3.10)
u′ =1
u′ 6=u

2
and ICI(k) is a zero-mean ICI with variance σICI = E{|ICI(ǫ, k)|2 }, caused by
the CFOs [Moos 94]. This interference can be considered as a Gaussian vector due
to the large number of subcarriers [Huan 10b]. It should be noted that the ICI
is included in ςu (k), the characteristics of which are used by the IDMA receiver
to estimate the transmitted bits (See chapter 1 in section 1.4.3.2).
In the following, we will show that the transmitted bits can be recovered by taking
into account the ICI characteristics in the IDMA receiver. In addition, as r(k)
(eq)
in (3.8) is a function of ǫu′ and Hu′ (k), it is necessary to know the CFOs and
the CSIs values at the IDMA receiver. In the next subsection, the estimations of
these vectors are presented.

3.2.1 Multiple CFO and Channel Estimation

PU
Let us define the vector ǫ of size U and the vector h of size L = u=1 Lu by
storing the CFOs and the CSIs of the U users in the system as follows:

ǫ = [ǫ1 , ǫ2 , . . . , ǫu , . . . , ǫU ] (3.11)
h = [h1 , h2 , . . . , hu , . . . , hU ] (3.12)

107
3. Frequency Synchronization and Channel Equalization of
OFDM-IDMA Uplink Systems

where hu = [hu (0), hu (1), . . . , hu (l), . . . , h(Lu − 1)].

In order to restore orthogonality among each user subcarrier, both the CFO-
synchronization error vector ǫ and the CSI vector h have to be estimated at the
BS.
The state-space representation of the system (3.1)-(3.2) is the representation of
what happens during one OFDM-IDMA frame1 .
Let us define the vector x(n) of size U = U + 2L, which is the state vector of the
system (3.1)-(3.2):
h iT
x(n) = ǫ(n) Re {h(n)} Im {h(n)} (3.13)

Let Y be the 2 × K observation matrix that stores the real and the imaginary
parts of the received OFDM-IDMA symbol R:

Y(ǫ, h) = C(ǫ, h) + B
(3.14)
= [Y(0, ǫ, h), Y(1, ǫ, h), . . . , Y(n, ǫ, h), . . . , Y(K − 1, ǫ, h)]

where C(ǫ, h) is a 2 × K matrix that stores the real an the imaginary parts of
the contribution of all users to the received OFDM-IDMA symbol:
 nP oT 
U
Re u=1 Ru
C(ǫ, h) = 
 nP

oT 
Im U
u=1 Ru (3.15)
= [C(0, ǫ, h), C(1, ǫ, h), . . . , C(n, ǫ, h), . . . , C(K − 1, ǫ, h)]

and " #
Re{B(n)}T
B=
Im{B(n)}T (3.16)
= [B(0), B(1), . . . , B(K − 1)]
2
σB
B(n) is an AWGN vector with zero-mean and covariance matrix I.
2 2
Now let us introduce the state-space representation of the system to estimate the
CFO and the channel:
State equation:

x(n) = x(n − 1) + w(n) ∀ n ∈ [0, K − 1] (3.17)

1
The OFDM-IDMA frame is a block composed of several OFDM-IDMA symbols.

108
 3.2 OFDM-IDMA Modified Receiver

Measurement equation:

Y(n) = C(n, x(n)) + B(n) ∀ n ∈ [0, K − 1] (3.18)

where w(n) is an AWGN matrix with zero-mean and covariance matrix σw2 IU . It
should be noted that σw2 is very low and can even be equal to 0.
The SPKF detailed in appendix E is used to estimate the CFOs and the channels,
after the time synchronization and CP removal.Then, the OFDM demodulation
is performed without any CFO correction and the resulting samples are used in
the modified IDMA receiver presented in the next subsection.

3.2.2 Modified IDMA Receiver

At the IDMA receiver, an a priori knowledge of Hu(eq) (k) is required. In addition,
it is necessary to characterize the interference term ςu (k) and the noise b(k).
Given (3.9) and the estimated values of the CFO ǫ̂u and the CSI ĥu , we can
obtain Ĥu(eq) (k) which is an estimation of Hu(eq) (k).
Then, given the output r(k) of the OFDM demodulation, the values of Ĥu(eq) (k),
the additive-noise variance σB2 , the ICI variance σICI
2
and guESE (k), the IDMA
reception can be performed1 .
Let us view the transmitted symbol Su (k) as a random variable. Given (1.41),
the modified ESE computes the mean and the covariance of Su (k) as in (1.46) and
(1.47) respectively. For the initialization process, guESE (k) = 0 for i = 1, where i
denotes the iteration number with i ∈ {1, . . . , Iidma } and Iidma is the maximum
number of iterations. Then, assuming that Su are independent and identically
distributed, and applying the central limit theorem, we can consider a Gaussian
approximation for ςu (k) and r(k). Thus, they can be completely characterized by
their means and variances.
Given (3.8) and by replacing Hu(eq) (k) by its estimate, one has:

E(ςu (k)) = E(r(k)) − Ĥu(eq) (k)µu (k) ∀u, k

(3.19)
V ar(ςu (k)) = V ar(r(k)) − (Ĥu(eq) (k))2 υu (k) ∀u, k
1
The proposed IDMA receiver may be applied to real and complex signals. For the sake of
simplicity, let us consider a BPSK modulation.

109
3. Frequency Synchronization and Channel Equalization of
OFDM-IDMA Uplink Systems

It should be noted that (3.8) can be rewritten as:

U
X
r(k) = Hu(eq) (k)Su (k) + ICI(k) + b(k) (3.20)
u=1

Then, given (3.20) and by replacing Hu(eq) (k) by its estimate, one has:
U
X
E(r(k)) = Ĥu(eq) (k)µu (k) ∀u, k
u=1
U
(3.21)
X
V ar(r(k)) = σZ2 + (Ĥu(eq) (k))2 υu (k) ∀u, k
u=1
with σZ2 = σICI
2
+ σB2 .

Unlike a conventional IDMA receiver, the interference-plus-noise is now a combi-

nation of the MAI, the ICI produced by the CFOs and the Gaussian noise.
The modified ESE process for the uth user at the ith iteration generates the value
of the a posteriori LLR as follows (See chapter 1 in section 1.4.3.2):

(eq) r(k) − E(ςu (k))

gESE
u (k) = 2Ĥu (k) × ∀u, k (3.22)
V ar(ςu (k))

The outputs of the modified ESE are inserted in the user decoding process (See
1st user
gESE
1 (k) estimated
bits
g1ESE(k)

r(k) user U th user

modified
ESE gESE
U (k) decoding estimated
bits
ESE(k)
OFDM gU
demodulation

Ĥu
(eq) 2 + σ2
σz2 = σB ICI

ĥ, ǫ̂
R SPKF initialization step

Figure 3.1: Proposed OFDM-IDMA uplink receiver, where the user decoding is
the same as in figure 1.20

110
 3.2 OFDM-IDMA Modified Receiver

Algorithm 4 Modified ESE

Require: r(k), σB2 , σICI
2
, Ĥu(eq) (k), Iidma
Initialization step:
guESE (k) = 0 ∀u, k
Iterative process
for i = 1 to Iidma do  ESE 
µu (k) = E(Su (k)) = tanh gu 2 (k) ∀u, k
2
υu (k) = V ar(Su (k)) = 1 − (µu (k)) ∀u, k
P
E(r(k)) = Uu=1 Ĥu(eq) (k)µu (k) ∀u, k
P
V ar(r(k)) = σZ2 + Uu=1 (Ĥu(eq) (k))2 υu (k) ∀u, k
E(ςu (k)) = E(r(k)) − Ĥu(eq) (k)µu (k) ∀u, k
V ar(ςu (k)) = V ar(r(k)) − (Ĥu(eq) (k))2 υu (k) ∀u, k
(eq) r(k)−E(ςu (k))
u (k) = 2Ĥu (k) × V ar(ςu (k))
gESE ∀u, k
end for

chapter 1 in section 1.4.3.2). Then, at the ith iteration for i < Iidma , the outputs
of the decoding process are the values of guESE (k), that are reinserted in the ESE.
The ESE process is again performed. Finally, the decoding process produces
hard decisions to obtain the estimations of the transmitted bits during the final
iteration Iidma . See algorithm 4, where the modified ESE is summarized. Figure
3.1 shows the modified OFDM-IDMA uplink receiver.

3.2.3 Simulation Results

Simulation protocol:
We consider an OFDM-IDMA uplink system composed of U = 4, 8, 16 users over
K = 1024 subcarriers. The spreading factor is S = 64. BPSK is used to modulate
the information bits. The carrier frequency is at fc = 2.6GHz and the channel
bandwidth is set to W = 10MHz. We consider that the ENC consists only of the
2
insertion of a spreading code. We define the SNR= 10log( σσ2u ), where σu2 is the
B
2
mean power of the received signal from the uth user, and the SIR= 10log( σσ2 u ).
ICI
We consider a transmission over a normalized Rayleigh slow-fading frequency se-
lective channel channel composed of Lu = 1, 3, 5 ∀u. The CFOs are modeled as

111
3. Frequency Synchronization and Channel Equalization of
OFDM-IDMA Uplink Systems

random variables with a normal distribution with zero-mean and variance 0.1.
The estimations are obtained by using control data of one OFDM-IDMA symbol.
Figure 3.2 shows the convergence speed of the SPKF for a different number of
users in the system. We can see that the number of users in the system does
not really affect the estimation performance of the SPKF. This is a key point
in OFDM-IDMA systems that allow a large number of users to transmit at the
same time. However, the computational complexity of the SPKF increases with
the number of users in the system. Then, we fix the number of user to U = 4.
Figure 3.3 shows the CFO estimation performance in terms of MMSE of one user
in the system. The SPKF allows the CFO and the channel of the users to be
estimated. The MMSE for an AWGN channel and for channels with different
length are presented. The channel estimation for different channel lengths give
the same estimation performance, around 10−2 for a SNR= 0dB. Previous studies
in [Ping 02a] and [Maha 06] show the performance of the IDMA receiver for dif-
ferent number of multipaths, for different number of users and different number
of iterations. It has been already proved that the performance of the IDMA re-
ceiver in terms of BER are better when the number of multipaths and iterations
increases and also when the number of users decreases [Ping 02a], [Maha 06]. For

0.2

U=4
0.15 U=8
U=16
CFO value

0.1

0.05

−0.05
0 100 200 300 400 500 600 700 800 900 1000
Sample Index

Figure 3.2: SPKF convergence speed

112
 3.2 OFDM-IDMA Modified Receiver

SPKF Lu = 3
10
−2 SPKF Lu = 1
SPKF AWGN

MMSE

−3
10

−4
10
0 2 4 6 8 10 12 14 16 18
SNR (dB)

Figure 3.3: CFO estimation performance in terms of MMSE, SIR= 0dB

this reason, we keep the number of users equal to U = 4, and the number of iter-
ations to 3. The following results are the average of 5000 OFDM-IDMA symbols.
Figure 3.4 shows the BER for a transmission over a normalized Rayleigh slow-
fading frequency selective channel channel composed of ∀u Lu = 1 and Lu = 3
multipaths. The efficiency of the algorithm is confirmed. Indeed, we can see
that for Lu = 1 and a BER = 10−1 the difference between the results of the
proposed scheme and the perfect case is lower than 1dB. In addition, we can
notice that for Lu = 3 and a BER= 10−2 the difference between the results of the
proposed scheme and the perfect case is less than 1 dB. In both cases (Lu = 1 and
Lu = 3), that the proposed scheme gives a gain of around 3dB for BER= 10−1
in comparison with the linear detection method presented in [Cao 04a]. Figure
3.5 shows the BER for a transmission over a normalized Rayleigh slow-fading
frequency selective channel channel composed of ∀u Lu = 5 multipaths. We can
see that the difference between the results of the proposed scheme and the perfect
case is lower than 1 dB. In addition, the proposed scheme gives a gain of around
2dB for BER= 10−2 in comparison with the linear detection method presented
in [Cao 04a]. It should be noticed that the performance for Lu = 5 are better
than for Lu = 3 and Lu = 1. The IDMA receiver takes advantage of the diversity
introduced by the propagation channel.

113
3. Frequency Synchronization and Channel Equalization of
OFDM-IDMA Uplink Systems

−1 Lu = 1
10

Lu = 3
−2
10
BER

−3
10
No CFO - Channel Perfectly estimated
SPKF-CFO estimated - Channel perfectly estimated
SPKF-CFO/channel estimated
Linear detection-CFO and Channel perfectly estimated
−4
10
0 2 4 6 8 10 12 14
SNR (dB)

Figure 3.4: BER performance for a channel with Lu = 1 and Lu = 3

0
10

−1
10
BER

−2
10

−3
10
No CFO - Channel Perfectly estimated
SPKF-CFO estimated - Channel perfectly estimated
SPKF-CFO/channel estimated
Linear detection-CFO and Channel perfectly estimated
−4
10
0 2 4 6 8 10 12 14
SNR (dB)

Figure 3.5: BER performance for a channel with Lu = 5

114
 3.3 Space-Time Block Code OFDM-IDMA Receiver

3.2.4 Conclusions
In this section, a modified OFDM-IDMA receiver is proposed. The receiver is
robust against ISI due to the OFDM modulation and it is robust against the
MAI due to the IDMA scheme. Unlike common approaches, the IDMA scheme
allows the CFO correction step to be skipped. After the OFDM demodulation,
the CFOs and the channels are jointly estimated by using a SPKF. Then, the
OFDM demodulation is performed without any CFO correction. Finally, the
modified IDMA receiver accurately estimates the transmitted bits of the uth
user, by taking into account the estimations of the SPKF and the interference
produced by the CFO. The proposed IDMA receiver is able to cancel the ICI
produced by the CFOs , as well as the MAI produced by the other users in the
system. Simulation results clearly show the efficiency of the proposed OFDM-
IDMA receiver. In addition, we show that the proposed receiver outperforms
existent CFO correction methods such as [Cao 04a].
In the next section, we show that the proposed scheme can be extended to the
multiple-input multiple-output (MIMO) case.

3.3 Space-Time Block Code OFDM-IDMA Re-

ceiver
Space-time coding is an effective method to reach the capacity of multiple input
multiple output (MIMO) wireless channels. It can supply transmit diversity and
power gain without limiting the bandwidth efficiency [Taro 97]. STBC is a kind
of space-time coding with simple encoding structure. It allows a simple ML de-
coding algorithm based on linear signal processing at the receiver [Taro 99].
Given the advantages of the STBC, it has been combined with IDMA. More
specifically, in [Wu 03] and [Ping 02b], interleaving is used to separate signals
from different antennas. If P denotes the number of transmit antennas, U × P
interleavers are required by the transmitter. Then, to reduce the complexity of
the transmitter, Yang et al. [Yang 08] propose a STBC-IDMA system which uses
only U interleavers. However, the authors do not aim at estimating the channel.
In [Shik 10], Suyama et al. propose a MIMO-OFDM IDMA system that uses

115
3. Frequency Synchronization and Channel Equalization of
OFDM-IDMA Uplink Systems

Su,1 Xu,1
CP-OFDM
bits to be modulation
transmitted Su
IDMA Alamouti
transmitter encoder
Xu,2
CP-OFDM
Su,2 modulation

Figure 3.6: OFDM-IDMA-STBC transmitter for the uth user, the IDMA trans-
mitter is the same as in figure 1.19, Su,p are the IDMA symbols and Xu,p are the
CP-OFDM-IDMA symbols

U interleavers and address the channel estimation by inserting an LMS in the

iterative architecture of the IDMA receiver. Nevertheless, the authors consider a
frequency synchronized system.
To our knowledge the frequency synchronization in a STBC-OFDM-IDMA scheme
has never been addressed. Therefore, in this section we suggest a joint CFO and
channel estimation by using a SPKF in a STBC-OFDM-IDMA system.

3.3.1 System Description

In the following, let us consider an STBC-OFDM-IDMA system consisting of U
simultaneously users who transmit information toward a single BS. The available
bandwidth W is divided among K subcarriers. We assume that the transmitter
uses P = 2 antennas to transmit the information.
As the CSIs are unknown and in order to increase the space diversity, each
couple of antennas transmits information from one user by using the Alam-
outi scheme [Alam 98]. In the following, let us associate the subscript (u, p)
to the pth antenna and to the uth user in the system, with p ∈ {1, 2} and
u ∈ {1, . . . U }. As shown in figure 3.6, the STBC-OFDM-IDMA transmitter
works as follows:
The IDMA modulated symbols after the IDMA transmitter are denoted as:

∆
Su = [Su (0), Su (1), . . . , Su (K − 1)]T (3.23)

116
 3.3 Space-Time Block Code OFDM-IDMA Receiver

The Alamouti code is implemented. Then, the OFDM modulation is performed.

The resulting STBC-OFDM-IDMA symbols are:

∆
Xu,1 = FH Su,1 = FH [Su (0), −Su∗ (1), . . . Su (K − 2), −Su∗ (K − 1)]T
∆
(3.24)
Xu,2 = FH Su,2 = FH [Su (1), Su∗ (0), . . . Su (K − 1), Su∗ (K − 2)]T

Let us define the CSI as follows:

hu,p = [hu,p (0), hu,p (1), . . . , hu,p (l), . . . , hu,p (Lu,p − 1)] (3.25)

where Lu,p is the length of the channel.

Finally, a CP of size Ng ≥ max
u,p
(Lu,p ) is inserted and the STBC-OFDM-IDMA
symbol is transmitted.
Moreover, let us define the row vector of size 2U × 1 that contains the normalized
P2 PU
CFOs of each user and the row vector of size L = p=1 u=1 Lu,p that contains
the CSIs of each user, respectively:

ǫ = [ǫ1 , ǫ2 , . . . , ǫu , . . . , ǫU ]
h = [h1 , h2 , . . . , hu , . . . , hS ] (3.26)

where ǫu = [ǫu,1 , ǫu,2 ] and hu = [hu,1 , hu,2 ].

At the receiver, due to the propagation conditions, the multipath channels and
the CFOs affect the transmitted signal. Without the noise contribution, after
time-synchronization and CP removing, the received STBC-OFDM-IDMA sym-
bol from each user should be written as follows:

∆
Ru = [Ru (0), . . . , Ru (n) . . . , Ru (K − 1)]
= Eu,1 hu,1 Xu,1 + Eu,2 hXu,2 (3.27)

h i
where Eu,p = diag 1, ej2πǫu,p /K , . . . , ej2π(K−1)ǫu,p /K , ǫu,p is the normalized CFO
and hu,p is the channel circulant matrix defined as follows (See chapter 1 in

117
3. Frequency Synchronization and Channel Equalization of
OFDM-IDMA Uplink Systems

section 1.4.2.2):
 
hu,p (0) 0
. . . hu,p (Lu,p − 1) . . . hu,p (1)
 ... ... 

 hu,p (1) hu (0) 0 hu,p (2) 


 .. ... ... ... ... .. 



. . 

hu,p =  hu,p (Lu − 1) hu,p (Lu,p − 2) . . .

0
...
0

 (3.28)
 
 ... ... .. 

 0 hu,p (Lu − 1) hu,p (0) . 

 .
.. . .. . .. ... ... .. 

 . 

0 0 . . . hu,p (Lu,p − 2) . . . hu,p (0)

Finally, the received signal at the BS is a superposition of the contributions from

the U active users:
∆
R = [R(0, . . . , R(n) . . . , R(K − 1)]
U
X
= Ru + B(n) (3.29)
u=1

where B(n) is a zero-mean AWGN noise vector with covariance matrix IK σB2 . As
in section 3.2, the CP-OFDM is performed without any CFO correction. Then,
after some manipulations (See section 3.2), the received samples r(k) and r(k +1)
can be expressed can be expressed as:
U
X (eq) (eq)
r(k) = {Hu,1 (k)Su (k) + Hu,2 (k)Su (k + 1)} + ICI(k) + b(k)
u=1
XU
(eq) (eq)
r(k + 1) = {−Hu,1 (k + 1)Su∗ (k + 1) + Hu,2 (k + 1)Su∗ (k)}
u=1
+ ICI(k + 1) + b(k + 1)
(eq)
where Hu,p (k) = F(ǫu,p )Hu,p (k) for p ∈ {1, 2} and ICI(k) is a zero-mean Gaus-
2
sian ICI with variance σICI = E{|ICI(ǫ, k)|2 } caused by all the CFOs. In addi-
tion,
PK−1 j2πnk
b(k) = n=0 B(n)e− K , F(ǫu,p ) is defined in (3.5) and
Ls,p −1
X j2πlk
Hu,p (k) = hu,p (l)e− K forp ∈ {1, 2} (3.30)
l=0

In the following, we show that the transmitted bits can be recovered by taking
into account the ICI characteristics in the IDMA receiver. As it is necessary to

118
 3.3 Space-Time Block Code OFDM-IDMA Receiver

know the values of CFOs and the CSIs at the IDMA receiver, the estimation of
these vectors is presented in the next subsection.

3.3.2 Multiple CFO and Channel Estimation

Let us define the vector x(n) of size U = 2(U + L), which is the state vector of
the system (3.29):
h iT
x(n) = ǫ(n) Re {h(n)} Im {h(n)} (3.31)

Now let us introduce the state-space representation1 of the system estimate the
CFO and the channel:
State equation:

x(n) = x(n − 1) + w(n) ∀ n ∈ [0, K − 1] (3.32)

Measurement equation:

Y(n) = C(n, x(n)) + B(n) ∀ n ∈ [0, K − 1] (3.33)

where w(n) is an AWGN matrix with zero-mean and covariance matrix σw2 IU ,
Y(n) is the observation 2 × 1 vector that stores the real and the imaginary parts
of the STBC-OFDM-IDMA symbol R(n), B(n) is an AWGN matrix with zero-
mean and covariance matrix σB2 I2 .
The SPKF detailed in appendix E is used to estimate the CFOs and the chan-
nels. After the time synchronization and CP removing, the CFOs and channels
are estimated. Then, the OFDM demodulation is performed without any CFO
correction and the obtained samples are inserted in the MRC receiver presented
in the next subsection.

3.3.3 Maximum Ratio Combining Operations

Given the estimates of the CFOs and the channels, we can obtain the value of
(eq)
Ĥu,p . Assume that Su (k) and Su (k + 1) are two successive symbols. Then, Ŝu (k)
1
As in section 3.2.1, the state-space representation represents what happens during one
STBC-OFDM-IDMA frame.

119
3. Frequency Synchronization and Channel Equalization of
OFDM-IDMA Uplink Systems

and Ŝu (k + 1) are the decisions obtained by the MRC of two successive received
symbols r(k) and r(k + 1), and the channel frequency responses Ĥu(eq) [Alam 98]:

U
X (eq)∗ (eq) (eq) (eq)∗
Ŝu (k) = {Ĥu,1 (k)Ĥu′ ,1 (k) + Ĥu,2 (k + 1)Ĥu′ ,2 (k + 1)}Su (k)
u′ =1
XU
(eq)∗ (eq) (eq) (eq)∗
+ {Ĥu,1 (k)Ĥu′ ,2 (k) − Ĥu,2 (k + 1)Ĥu′ ,1 (k + 1)}Su (k + 1)
u′ =1
(eq)∗ (eq)
+ Ĥu,1 (k){ICI(ǫ, k) + b(k)} + Ĥu,2 (k + 1){ICI ∗ (ǫ, k + 1) + b∗ (k + 1)}
(3.34)
U
X (eq) (eq)∗ (eq)∗ (eq)
Ŝu (k + 1) = {Ĥu,1 (k + 1)Ĥu′ ,1 (k + 1) + Ĥu,2 (k)Ĥu′ ,2 (k)}Su (k + 1)
u′ =1
XU
(eq)∗ (eq) (eq) (eq)∗
+ {Ĥu,2 (k)Ĥu′ ,1 (k) − Ĥu,1 (k + 1)Ĥu′ ,2 (k + 1)}Su (k)
u′ =1
(eq)∗
+ Ĥu,2 (k){ICI(ǫ, k) + b(k)}
(eq)
+ Ĥu,1 (k + 1){ICI ∗ (ǫ, k + 1) + b∗ (k + 1)}∗
(3.35)
Then, the values Ŝu (k) and Ŝu (k + 1) are inserted in the ESE presented in the
next subsection.

3.3.4 Modified Elementary Signal Estimator

Then, given the outputs of the MRC Ŝ(k) and Ŝ(k + 1), the values of Ĥu(eq) (k),
the additive-noise variance σB2 , the ICI variance σICI
2
and guESE (k), the IDMA
reception is performed.
It should be noted that (3.34) and (3.35) can be rewritten as:

(eq) (eq)
Ŝu (k) = {||Ĥu,1 (k)||2 + ||Ĥu,2 (k + 1)||2 }Su (k) + ςu (k)
(eq) (eq)
Ŝu (k + 1) = {||Ĥu,1 (k + 1)||2 + ||Ĥu,2 (k)||2 }Su (k + 1)

+ςu (k + 1) (3.36)

120
 3.3 Space-Time Block Code OFDM-IDMA Receiver

and the distortions ςu (k) and ςu (k +1) contained in the output values of the MRC
can be expressed as:
U
X (eq)∗ (eq) (eq) (eq)∗
ςu (k) = {Ĥu,1 (k)Ĥu′ ,1 (k) + Ĥu,2 (k + 1)Hu′ ,2 (k + 1)}Su (k)
u′ =1
u′ 6=u
U
X (eq)∗ (eq) (eq) (eq)∗
+ {Ĥu,1 (k)Hu′ ,2 (k) − Ĥu,2 (k + 1)Hu′ ,1 (k + 1)}Su (k + 1) (3.37)
u′ =1
u′ 6=u
(eq)∗
+ Ĥu,1 (k){ICI(ǫ, k) + b(k)}
(eq)
+ Ĥu,2 (k + 1){ICI ∗ (ǫ, k + 1) + b∗ (k + 1)}
U
X (eq) (eq)∗ (eq)∗ (eq)
ςu (k + 1) = {Ĥu,1 (k + 1)Hu′ ,1 (k + 1) + Ĥu,2 (k)Ĥu′ ,2 (k)}Su (k + 1)
u′ =1
u′ 6=u
U
X (eq)∗ (eq) (eq) (eq)∗
+ {Ĥu,2 (k)Hu′ ,1 (k) − Ĥu,1 (k + 1)Hu′ ,2 (k + 1)}Su (k)
u′ =1
u′ 6=u
(eq)∗
+ Ĥu,2 (k){ICI(ǫ, k) + b(k)}
(eq)
+ Ĥu,1 (k + 1){ICI ∗ (ǫ, k + 1) + b∗ (k + 1)}
(3.38)
Then, assuming that Su are independent and identically distributed, and applying
the central limit theorem, a Gaussian approximation for ςu (k) and Ŝu (k) can
be considered. Thus, they can be completely characterized by their means and
variances. The modified ESE process for the uth user at the ith iteration generates
the value of the a posteriori LLR as follows:

2 (eq) 2 (eq) Ŝu (k) − E(ςu (k))

gESE
u (k) = 2{||Ĥu,1 (k)|| + ||Ĥu,2 (k)|| } × ∀u (3.39)
V ar(ςu (k))

For the initialization process guESE (k) = 0 for i = 1, where i denotes the iteration
number with i ∈ {1, . . . , Iidma }, with Iidma the maximum number of iterations.
The outputs of the modified ESE are inserted in the user decoding process pre-
sented in chapter 1 in section 1.4.3.2. Then, at the ith iteration for i < Iidma ,
the outputs of the decoding process are the values of guESE (k), that are reinserted
in the ESE. The ESE process is again performed. Finally, the decoding process
produces hard decisions to obtain the estimations of the transmitted bits during

121
3. Frequency Synchronization and Channel Equalization of
OFDM-IDMA Uplink Systems

the final iteration Iidma . Figure 3.7 shows the proposed STBC-OFDM-IDMA
scheme. 1st user
gESE
1 (k) estimated
bits
g1ESE(k)
Ŝ1 (k)
r(k) user U th user
modified
MRC estimated
ESE gESE
U (k) decoding
ŜU(k) bits
ESE(k)
OFDM gU
demodulation

(eq) σz2 = σB
2 + σ2
ICI
Ĥu
ĥ, ǫ
R SPKF initialization step

Figure 3.7: Proposed STBC-OFDM-IDMA uplink receiver, where the user de-
coding is the same as in figure 1.20

3.3.5 Simulation Results

Simulation protocol:
We consider an OFDM-IDMA uplink system composed of U = 4 users over
K = 1024 subcarriers. The spreading factor is S = 64. BPSK is used to modulate
the information bits. The carrier frequency is at fc = 2.6GHz and the channel
bandwidth is set to W = 10MHz. We consider that the ENC consists only
2
of the insertion of a spreading code. We define the SNR= 10log( σσ2u ), where
B
σu2 is the mean power of the received signal from the uth user, and the SIR=
2
10log( σσ2 u ). We consider a transmission over a normalized Rayleigh slow-fading
ICI
frequency selective channel composed of ∀u, p Lu,p = 1, 3 multipaths. The CFOs
are modeled as random variables with a normal distribution with zero-mean and
variance 0.1. The number of iterations is set to 5. The estimations are obtained
using control data of one OFDM-IDMA symbol.
Figure 3.8 shows the CFO estimation performance in terms of MMSE of one user
in the system. The SPKF allows the CFO and the channel of the users to be

122
 3.3 Space-Time Block Code OFDM-IDMA Receiver

estimated. The MMSE for an AWGN channel and for channels with different
length are presented. The channel estimation for different channel lengths give
the same estimation performance, around 10−2 for a SNR= 0dB. The following

−2
10
SPKF Lu = 3
SPKF Lu = 1
SPKF AWGN
MMSE

−3
10

−4
10
0 2 4 6 8 10 12 14 16
SNR (dB)

Figure 3.8: CFO estimation performance in terms of MMSE, SIR= 0dB

results are the average of 5000 OFDM-IDMA symbols. Figure 3.9 and 3.10 show
the BER for a transmission over a normalized Rayleigh slow-fading frequency
selective channel with Lu,p = 1 and Lu,p = 3 multipaths ∀u, p respectively. The
efficiency of the OFDM-IDMA-STBC proposed scheme is confirmed, we can see
that for a BER of 10−2 there are less than 1 dB between the curve without CFO
and the curve of the proposed estimator. In addition, we can see a gain of around
3dB of the proposed scheme in comparison with the linear detection method
presented in [Cao 04a].

123
3. Frequency Synchronization and Channel Equalization of
OFDM-IDMA Uplink Systems

−1
10
BER

−2
10
No CFO - Channel perfectly estimated
SPKF-CFO estimated - Channel perfectly estimated
SPKF-CFO/channel estimated
Linear detection - CFO and Channel perfectly estimated

0 2 4 6 8 10 12 14
SNR (dB)

Figure 3.9: BER performance for a channel Lu = 1

−1
10

−2
10
BER

−3
10

No CFO - Channel perfectly estimated

SPKF-CFO estimated - Channel perfectly estimated
SPKF-CFO/channel estimated
Linear detection-CFO and Channel perfectly estimated
−4
10
0 2 4 6 8 10 12 14
SNR(dB)

Figure 3.10: BER performance for a channel Lu = 3

3.4 Conclusions
In this section, two OFDM-IDMA receivers are proposed. A joint CFO/channel
estimation is performed using a SPKF. Then, the estimated parameters are in-

124
 3.4 Conclusions

serted in an IDMA iterative receiver. The IDMA receiver accurately estimates

the transmitted symbols from all the users; the ICI produced by the CFOs is
canceled, as well as the MAI. The receiver is robust against ISI thanks to the
OFDM modulation and robust against the MAI due to the IDMA scheme. Un-
like common approaches, the IDMA schemes also allow the CFO correction step
to be skipped.
OFDM-IDMA systems allow a large number of users to transmit at the same
time. We show that the proposed receiver outperforms existing CFO correction
methods such as [Cao 04a].
In chapter 1, we mentioned that CR systems may be one solution to the problem
induced by the exponential growth of the number of users in the system. As an
alternative, OFDM-IDMA could be a solution to the overload produced by the
constant increase of the user density per cell in cellular systems.

125
Conclusions and Perspectives
Estimating the CFOs and channels plays a key role in the design of mobile wire-
less systems. In this PhD, our purpose was to create new receiver schemes that
take into account both phenomena. We have proposed to jointly estimate the
CFOs and the channels for an OFDMA uplink system by using a recursive ap-
proach [Pove 10]. On the one hand, the EKF, the SOEKF, the IEKF and the
SPKF including the UKF and the CDKF were considered. On the other hand,
to relax the Gaussian assumptions on the model noise and measurement noise,
the “extended H∞ filter” and the “unscented H∞ filter” were studied.

From the estimation point of view, we first compared the different Kalman tech-
niques. We can draw the following conclusions:
1/ The SOEKF gives slightly “better” estimates of the CFO and the channels
than the EKF in terms of estimation accuracy. Nevertheless, the computational
cost of the SOEKF is higher than the EKF one, due to the calculation of the
Hessian matrix for the second-order expansion.
2/ Concerning the SPKF, the UKF and the CDKF give the same results in terms
of estimation accuracy. In addition, their computational costs are quite similar.
They converge faster than the EKF and the SOEKF. They have the advantage
of avoiding the calculation of Jacobians and Hessians matrices, but the Cholesky
factor of the autocorrelation matrices must be computed.
3/ The IEKF gives the “best” estimation performance in terms of estimation
accuracy and convergence speed. It needs less samples to converge to the true
value than the other Kalman approaches. However, due to the calculation of the
Jacobian matrix at each iteration, the computational complexity is the highest
among the proposed approaches.
Let us now look at the relevance of the H∞ approaches.
1/ the “unscented H∞ filter” gives better estimation performance than the “ex-
tended H∞ filter”. In addition, its convergence is faster than the one of the
“extended H∞ filter”.
2/ when the noise characteristics are available, the “extended H∞ filter” and the

127
Conclusions and Perspectives

EKF give similar results. In addition, there is no real difference between the “un-
scented H∞ filter” and the SPKF. If the noise variances are a priori unknown,
the H∞ approaches converge faster than the Kalman algorithms. Nevertheless,
the noise attenuation level and the weighting matrices have to be chosen by the
practitioner. If the noise attenuation level is set to a value that is too small, a
solution to the H∞ issue may not exist. If the noise attenuation level is set to a
high value to guarantee the existence of a solution, the resulting H∞ filter “looks
like” the Kalman filter.
3/ The computational complexity of the “extended H∞ filter” is slightly higher
than the EKF while the computational cost of the “unscented H∞ filter” is slightly
higher than the SPKF.

Then, the above techniques have been compared with the EM-based algorithms
proposed in [Pun 04a] and in [Pun 06] in terms of estimation accuracy and com-
putational complexity. Even if the EM-based estimator gives better estimation
performance, its computational complexity is much higher than the costs of the
SPKF and the “unscented H∞ filter”. Our methods are therefore very useful for
practical cases.

Based on the above considerations, we think that the SPKF gives the best com-
promise in terms of estimation performance, computational complexity and a
priori information to be tuned.

From the digital communication point of view, the above optimal filtering tech-
niques have been combined with an MMSE-SD to obtain a non-pilot aided CFO
estimator for an OFDMA uplink [Pove 09a], [Pove 09b], [Pove 11c]. The MAI
suppression scheme followed by the CFO estimation algorithm based on optimal
filtering can jointly estimate and detect each user’s CFO and symbols respectively.
According to the simulations we carried out, the proposed scheme can effectively
suppress the MAI caused by a relatively large CFO, and it is sufficiently robust
to CFO variations. The iterative decoding applies simple hard interference can-
cellation techniques, resulting in a moderate complexity.
Then, the SPKF has been chosen to be combined with two statistical tests: the

128
 Conclusions and Perspectives

BHT and the CUSUM test [Pove 11b], [Pove 11d] to detect the beginning and
the end of a CR-NBI disturbing the estimation of the CFOs and the channels of
different users in a PU-OFDMA system. The proposed estimator operates in the
time domain unlike other approaches such as [More 08] that estimate the CFO
in presence of an NBI in the frequency domain. By using a SPKF and without
increasing the computational complexity, we take advantage of the innovation in
the Kalman filter to detect the CR-NBI. When using the BHT, better perfor-
mance are obtained in terms of the CR-NBI probability of detection. However,
when using the CUSUM, better performance in terms of PU probability of detec-
tion are obtained. The results show that the SPKF without the statistical tests is
not able to estimate the CFO of each user. The proposed estimator outperforms
the existing CFO estimation methods such as [More 08] in presence of NBI and
it is robust to the NBI power. Nevertheless, the architecture is sensitive to the
duration in time of the CR-NBI.
We have also proposed to combine the SPKF with a modified version of an itera-
tive IDMA receiver to design a new OFDM-IDMA receiver [Pove 12]. Simulation
results show that the SPKF provides “accurate” estimations of the CFO and the
channel of each user. In addition, the modified OFDM-IDMA receiver accurately
estimates the transmitted bits of each user, by taking into account the estima-
tions of the SPKF and the interference produced by the CFO. According to the
simulation tests we carried out, our receiver outperforms CFO correction meth-
ods proposed for other multicarrier systems such as [Cao 04a].
Our last contribution consists in combining the SPKF with a STBC and a new
variant of an iterative IDMA receiver to take advantage of the space diversity
[Pove 11a]. The simulation results confirm that the SPKF provides accurate es-
timations of the CFO and the channel. In addition, the STBC-OFDM-IDMA
receiver can accurately estimate the transmitted bits of each user. We show
that the proposed receiver outperforms existing CFO correction methods such as
[Cao 04a].

Currently, we are working on the sensitivity analysis of the optimal filtering

techniques in the non-linear case. We are studying the local sensitivity of the
Kalman filtering and the H∞ filtering regarding perturbations. For this purpose

129
Conclusions and Perspectives

we take advantage of appendix C.

In the future, we propose to address the estimation issue by using a quadra-
ture Kalman filter (QKF) [Ito 00], which avoids the linearization step by using
the Gauss-Hermite numerical integration rule. In addition, the relevance of the
“second-order H∞ filter” [Hu 11] could be investigated.
In this PhD, some assumptions have been made. For instance, we suppose that
the radio-frequency (RF) power amplifier (PA) is used below its saturation point
to guarantee a linear behavior of the PA. However, this may lead to a poor power
efficiency, especially when the transmitted signal exhibits high PAPR1 . One so-
lution would be to use the PA closer to its saturation point, but this induces
nonlinear distortion of the transmitted signal. In [Roli 09], non-linear models for
the PA such as the Volterra model can be considered. In the future, we could
design new architectures to compensate the influence of these non-linearities.
In addition, in the future we propose to investigate an OFDM-IDMA with a suf-
ficiently long CP that allows a transmission in an asynchronous way. We could
study this non-optimal characteristic from a point of view of the spectral efficiency
to allow all the users to use the same interleaver.

1
Due to the large number of independent modulated subcarriers that are added up coher-
ently, the transmitted signal exhibits a high PAPR.

130
Appendix A
Kalman Filter - Linear Case

Firstly presented in [Kalm 60] by Rudolf E. Kalman, the Kalman filter (KF) is
based on a state-space representation of the system. It has been used in a wide
range of applications, from speech enhancement to time-varying autoregressive
parameters tracking, from biomedical applications to mobile communications.
It uses two equations to describe the system. When the state-space equations
are linear and the noises are additive white zero-mean Gaussian processes, they
satisfy:
State equation:
x(n) = Φ(n − 1)x(n − 1) + Γw(n) (A.1)
Measurement equation:
y(n) = Ψ(n)x(n) + b(n) (A.2)

where x(n) is the state vector1 of size U at time n and y(n) is the measurement
vector of size K at time n. The model noise w and the observation noise b are
uncorrelated white zero-mean Gaussian vectors with covariance matrices assumed
to be Q = σw2 IU and R = σb2 IK respectively. In addition, Φ(n−1) is the transition
matrix of size U × U from time n − 1 to n, Γ is the input gain matrix of size
U × U and Ψ(n) is the measurement matrix K × U at time n.
The KF operates in two steps: the prediction step and the filtering step.
In the prediction step, the KF uses the estimated state at the previous instant to
estimate the current state, without taking into account the current observation.
This is the so-called a priori estimation of the state vector defined as follows:

x̂(n|n − 1) = E [x(n)|y(0), y(1), . . . , y(n − 1)] (A.3)

1
The state vector stores the set of parameters of the system, necessary for prediction when
the input is known. In practical cases, it usually contains the unknown quantities to be esti-
mated.

131
A. Kalman Filter - Linear Case Appendix A

Figure A.1: Kalman filter - system model

Given (A.1) and (A.3) and as w is a zero-mean AWGN, the a priori estimation
of the state vector can be expressed as follows:

x̂(n|n − 1) = Φ(n − 1)x̂(n − 1|n − 1) (A.4)

where x̂(n − 1|n − 1) is the estimation of the state vector at time n − 1 given the
set of observations [y(0), y(1), . . . , y(n − 1)].
Then, let us introduce the a priori estimation error:

x̃(n|n − 1) = x(n) − x̂(n|n − 1) (A.5)

and its corresponding covariance matrix P(n|n − 1) defined by:

h i
P(n|n − 1) = E x̃(n|n − 1)x̃H (n|n − 1) (A.6)

Given (A.1) and (A.4), the a priori estimation error of the state vector can be
rewritten as:
x̃(n|n − 1) = Φ(n − 1)x(n − 1) + Γw(n) − Φ(n − 1)x̂(n − 1|n − 1)
(A.7)
= Φ(n − 1)x̃(n − 1|n − 1) + Γw(n)
Then, the covariance matrix of x̃(n|n − 1) satisfies:
h i
P(n|n − 1) = Φ(n − 1)E x̃(n − 1|n − 1)x̃H (n − 1|n − 1) ΦH (n − 1)
h i
+ ΓE w̃(n)w̃H (n) ΓH (A.8)
= Φ(n − 1)P(n − 1|n − 1)ΦH (n − 1) + ΓQΓH

132
A. Kalman Filter - Linear Case

At that stage, the filtering step uses the current observation to correct the a
priori estimated state vector, in order to obtain the a posteriori estimation. To
estimate the current state and by following the kind of update equation obtained
in the RLS algorithm, one has:

x̂(n|n) = x̂(n|n − 1) + K(n)ỹ(n) (A.9)

where K(n) is the Kalman gain and ỹ(n) is the innovation defined as:

ỹ(n) = y(n) − ŷ(n|n − 1)

(A.10)
= y(n) − Ψ(n)x̂(n|n − 1)

with ŷ(n|n − 1) the prediction of y(n) based on x̂(n|n − 1).

The Kalman gain is defined to obtain the MMSE estimation of the state vector.
This means that:
 h i
∂ E xH (n|n)x(n|n) ∂tr(P(n|n))
= =0 (A.11)
∂K(n) ∂K(n)

and it satisfies:
K(n) = {Pxy (n)}{Pyy (n)}−1 (A.12)

where Pyy (n) is the innovation covariance matrix defined as follows:

Pyy (n) = Ψ(n)P(n|n − 1)ΨH (n) + R (A.13)

and Pxy (n) is the cross-covariance matrix between y(n) and the state prediction
x̂(n|n − 1) defined as:

Pxy (n) = P(n|n − 1)ΨH (n) (A.14)

Finally, by combining (A.9) and (A.12) it can be easily shown that the error
estimation covariance matrix is updated as follows:

P(n|n) = P(n|n − 1) − K(n)Ψ(n)P(n|n − 1)

(A.15)
= {IU − K(n)Ψ(n)}P(n|n − 1)

Given (A.8) adjusted to the set of observations {y(1), . . . , y(n)} instead of

{y(1), . . . , y(n − 1)}, (A.12) and (A.15), the error estimation covariance matrix

133
A. Kalman Filter - Linear Case Appendix A

satisfies the following Riccati equation1 for the Kalman filter:

P(n + 1|n) = Φ(n)P(n|n)ΦH (n) + ΓQΓH

(A.8)

= Φ(n){IU − K(n)Ψ(n)}P(n|n − 1)ΦH (n) + ΓQΓH

(A.15)

= Φ(n){IU − P(n|n − 1)ΨH (n){Ψ(n)P(n|n − 1)ΨH (n) + R}−1

(A.12)

× Ψ(n)}P(n|n − 1)ΦH (n) + ΓQΓH

(A.16)
The Kalman filter can also be derived from a Bayesian MAP perspective [VdMe 04a].
Indeed, (A.3) can be expressed as:
Z
x̂(n|n − 1) = x(n)P{x(n)|Y(n − 1)}dx(n) (A.17)

where P{x(n)|Y(n − 1)} = P{x(n)|y(0), y(1), . . . , y(n − 1)} is the a priori prob-
ability density of the state vector. As it is Gaussian, it can be completely char-
acterized by its mean and its covariance matrix. Then, by using the Bayes rule
the a posteriori probability density can be obtained as follows:

P{x(n)|Y(n)} = P{x(n)|y(0), y(1), . . . , y(n)}

P{Y(n)|x(n)}P{x(n)}
=
P{Y(n)}
P{y(n), Y(n − 1)|x(n)}P{x(n)}
=
P{y(n), Y(n − 1)}
P{y(n)|Y(n − 1), x(n)}P{Y(n − 1)|x(n)}P{x(n)}
=
P{y(n)|Y(n − 1)}P{Y(n − 1)}
P{y(n)|Y(n − 1), x(n)}P{x(n)|Y(n − 1)}P{Y(n − 1)}P{x(n)}
=
P{y(n)|Y(n − 1)}P{Y(n − 1)}P{x(n)}
P{y(n)|x(n)}P{x(n)|Y(n − 1)}
=
P{y(n)|Y(n − 1)}
(A.18)
Then, maximizing P{x(n)|Y(n)} consists in maximizing the numerator of (A.18).

1
The Riccati equation is defined as: P = A{I − PBH (BPBH + R)−1 B}PAH + Q.

134
Appendix B
H∞ Filter - Linear Case

The H∞ approach was introduced in the field of control in 1981 [Zame 81]. A
solution of the H∞ estimation problem can be based on polynomial decompo-
sition techniques [Grim 90]. Nevertheless, they lead to equations with high-
computational cost that cannot be used in practical situations. As an alternative
solution, state-space approaches have emerged. In [Gero 99], the author presents
a solution of the H∞ estimation constraint, based on the resolution of a convex
optimization problem under linear matrix inequality constraints approaches. In
[Shak 92], Shaked et al. present a state-space approach solution based on the
resolution of a quadratic Riccati-type equation. This algorithm is easy to imple-
ment and has a lower computational cost than the above approaches.
The H∞ filter has been widely used in the field of control. For the last 12 years,
several studies have been conducted by the signal processing community.
In [Shen 99], instead of using a Kalman filtering, Shen et al. suggest using an
H∞ filter to enhance a speech signal disturbed by an additive noise and recorded
from a single microphone. For this purpose, the signal is assumed to be modeled
by an AR process. However, the AR parameters are unknown and hence need
to be estimated. Shen et al. propose to estimate them directly from the noisy
observations by using a second H∞ filter. Therefore, the resulting AR parameter
estimates are biased, as pointed out by Labarre et al. in [Laba 05], who inves-
tigated the relevance of the H∞ filtering for speech enhancement. To avoid the
problem of biased AR parameter estimates that can be obtained in [Shen 99],
Labarre et al. suggest estimating both the AR model and its parameters. Al-
though this leads to a non-linear estimation issue, they have developed a structure
based on two mutually interactive standard H∞ filters [Laba 07]. The first one
aims at estimating the AR model, while the second one updates the estimation
of the AR parameters. In addition, in [Jamo 07a], the authors take advantage

135
B. H∞ Filter - Linear Case Appendix B

of the two mutually interactive H∞ filter based approach to jointly estimate the
fading channel and its AR parameters. However, the authors do not obtain better
performance than a KF based method.

Let us now look more carefully at the H∞ filtering approach. Given the state-
space model in (A.1) and (A.2), let us introduce a third state-space equation to
focus on a linear combination of the state-vector components:

z(n) = Lx(n) (B.1)

where L is a linear transformation operator that can be either a matrix of size

U × U or a row vector of size U . z(n) is hence a vector or a scalar.
Given (A.1), (A.2), (B.1) and figure B.1, the H∞ filtering provides the estimation
of the state vector, by minimizing the H∞ norm of the transfer operator that
maps the discrete-time noise disturbances1 to the estimation error, as follows:
PN ob −1
∞ n=0 ke(n)k2
J = sup −1 PN ob −1 P ob −1 (B.2)
V n=0 kb(n)k2 + W−1 Nn=0 kw(n)k2 + eH (n)P(0)e(n)

where N ob denotes the number of available observations, e(n) = z(n) − ẑ(n), and
V and W are positive weighting matrices tuned by the practitioner to achieve
performance requirements.

1
V− 2 (n)b(n)
1
e(n) = z(n) − ẑ(n)
W− 2 (n)w(n) Transfer operator
P(0)e(0)

Figure B.1: Transfer operator from disturbances to estimation error for H∞ -norm
based estimation

1
Namely w(n), b(n) and the initial conditions on the state vector.

136
B. H∞ Filter - Linear Case

Since the minimization of (B.2) is often impossible, the following sub-optimal

H∞ problem is usually considered:

J ∞ < Ξ2 (B.3)

where Ξ2 is the prescribed noise attenuation level.

At that stage, P∞ (n + 1|n) satisfies the following Riccati equation for the H∞
filter:

P∞ (n + 1|n) = Φ(n)P∞ (n|n)ΦH (n) + ΓWΓH (B.4)

h i
= Φ(n)P∞ (n|n − 1){IU − ΨH (n) LH
" #
−1 Ψ(n)
×M P∞ (n|n − 1)}ΦH (n) + ΓWΓH
L

where
" # " #
V 0 Ψ(n) h i
M= + P∞ (n|n − 1) ΨH (n) LH
0 −Ξ2 IU L
" # (B.5)
Ψ(n)P∞ (n|n − 1)ΨH (n) + V Ψ(n)P∞ (n|n − 1)LH
=
LP∞ (n|n − 1)ΨH (n) LP∞ (n|n − 1)LH − Ξ2 IU

Remark: if L = IU , one has:

" #
Ψ(n)P∞ (n|n − 1)ΨH (n) + V Ψ(n)P∞ (n|n − 1)
M= (B.6)
P∞ (n|n − 1)ΨH (n) P∞ (n|n − 1) − Ξ2 IU

When using the H∞ filter, the state vector can be estimated recursively as follows:

x̂(n|n) = x̂(n|n − 1) + K∞ (n)ỹ(n) (B.7)

where K∞ (n) is the H∞ filter gain defined as:

K∞ (n) = {Pxy∞ (n)}{Pyy∞ (n)}−1 (B.8)

where Pyy∞ (n) is defined as follows:

Pyy∞ (n) = Ψ(n)P∞ (n|n − 1)ΨH (n) + V (B.9)

and Pxy∞ (n) is:

Pxy∞ (n) = P∞ (n|n − 1)ΨH (n) (B.10)

137
B. H∞ Filter - Linear Case Appendix B

(B.4) is true provided that:

P∞ (n + 1|n)−1 + ΨH (n)Ψ(n) − Ξ−2 LH L > 0 (B.11)

It should be noted that the purpose of the H∞ filter is to minimize the peak error
power in the frequency domain whereas the Kalman filter minimizes the average
error power [Grim 90].

138
Appendix C
Kalman Filter vs H∞ Filter

In this appendix, our purpose is to compare the H∞ filter and the KF in the linear
case when estimating model parameters, by comparing the Ricatti equations of
both algorithms. To our knowledge, there is no work dealing with this kind of
comparison.
Let us first assume that:

P∞ (n|n − 1) = P(n|n − 1) (C.1)

and
L = IU (C.2)

In addition, as it is often done when dealing with H∞ filter in signal processing,

let us set W and V to Q and R respectively. This implies that:

K∞ (n) = K(n) (C.3)

and
Pyy∞ (n) = Pyy (n) (C.4)

Then, let us compare P∞ (n + 1|n) and P(n + 1|n). For this purpose, the matrix
M defined in (B.6) must be inverted in (B.4). Among the approaches that could
be considered such as the matrix inversion lemma, we suggest using the one based
on the Schur complement [Brez 88] for the sake of simplicity1 . Indeed by defining
" # " #" #" #
−1 −1
1 A B I K −A B A 0 I K 0
M= , then M−1 =
C D 0 IU 0 S−1 −CA−1 IU
where A is matrix of size K × K, B is matrix of size K × U , C is matrix of size U × K, D is
matrix of size U × U , S = D − CA−1 B is the Schur complement of A in M and 0 is a zero
matrix.

139
C. Kalman Filter vs H∞ Filter Appendix C

Υ(n) the Schur complement of Pyy∞ (n) in M as follows:

Υ(n) = {P∞ (n|n − 1) − Ξ2 IU }

− P∞ (n|n − 1)ΨH (n){Pyy∞ (n)}−1 Ψ(n)P∞ (n|n − 1) (C.5)

Given (A.12), (A.13), (A.15), (C.1) and (C.4), one obtains:

Υ(n) = P(n|n − 1)
(C.1),(C.4)

− {P(n|n − 1)ΨH (n){Pyy (n)}−1 }Ψ(n)P(n|n − 1) − Ξ2 IU

= P(n|n − 1) − K(n)Ψ(n)P(n|n − 1) − Ξ2 IU
(A.12),(A.13)

= P(n|n) − Ξ2 IU (C.6)
(A.15)

Then given the previous footnote, M−1 can be expressed as the product of three
matrices, the coefficients of which are defined from the coefficients of M and the
Schur complement Υ(n) as follows:
" #" #" #
−1 IK −KH (n) {Pyy (n)}−1 0 IK 0
M =
0 IU 0 Υ−1 −K(n) IU
= W(n)X(n)Y(n) (C.7)

At that stage, given (C.7), we can rewrite the Riccati recursion (B.4) as:

P∞ (n + 1|n) = Φ(n)P(n|n − 1)ΦH (n) + ΓQΓH (C.8)

" #
h i Ψ(n)
−Φ(n)P(n|n − 1) ΨH (n) IU W(n)X(n)Y(n) P(n|n − 1)ΦH (n)
IU
h i
= Φ(n)P(n|n − 1)ΦH (n) + ΓQΓH − Φ(n) P(n|n − 1)ΨH (n) P(n|n)
" #" #
{Pyy (n)}−1 0 Ψ(n)P(n|n − 1)
× ΦH (n)
0 Υ−1 P(n|n)
Using (A.12), (A.15) and (C.6), this leads to:

P∞ (n + 1|n) = Φ(n)P(n|n − 1)ΦH (n) + ΓQΓH (C.9)

" #
h i Ψ(n)P(n|n − 1)
−Φ(n) P(n|n − 1)ΨH (n){Pyy (n)}−1 P(n|n)Υ−1 ΦH (n)
P(n|n)
= Φ(n)P(n|n − 1)ΦH (n) − Φ(n)K(n)Ψ(n)P(n|n − 1)ΦH (n)
(A.12)

−Φ(n)P(n|n)Υ−1 P(n|n)ΦH (n) + ΓQΓH

= Φ(n)P(n|n)ΦH (n) − Φ(n)P(n|n){P(n|n) − Ξ2 IU }−1 P(n|n)ΦH (n) + ΓQΓH
(A.15),(C.6)

140
C. Kalman Filter vs H∞ Filter

Hence, the solution of the Riccati equation when using the H∞ filter satisfies:

P∞ (n + 1|n) = Φ(n)P(n|n)ΦH (n) + QΞ (n) + ΓQΓH =Φ(n)P(n|n)ΦH (n) + QΞw

where

QΞ (n) = −Φ(n)P(n|n){P(n|n) − Ξ2 IU }−1 P(n|n)ΦH (n) (C.10)

Given, (A.16) and (C.10), one has

P∞ (n + 1|n) = P(n + 1|n) + QΞ (n) (C.11)

Therefore, H∞ filtering can be seen as a Kalman filtering with a model-noise

covariance matrix equal to QΞw = QΞ (n) + ΓQΓH .
For parameter tracking, the larger the coefficients of the state-noise covariance
matrix are, the easier it is to track the parameter variations, especially when the
parameters are subject to abrupt variations. Nevertheless, the larger they are,
the larger the variance of the estimated parameters over time is.
At that stage, let us introduce the eigenvalue decomposition of P(n|n):
 
λ1 0 · · · · · · 0

 0 λ2 0 · · · 0 

 .. . . . . . . .. 
 −1
P(n|n) = G 
 . 0 . G

 .. .. . . . .


 . . . . 0 
0 0 · · · 0 λU
where {λu }u∈{1,2,...,U } are the eigenvalues of P(n|n). Hence, one has:
 
λ1 − Ξ2 0 ··· ··· 0
2

 0 λ2 − Ξ 0 ··· 0 

 .. ... ... .. 
P(n|n) − Ξ2 IU = G
 . 0 .  −1
G

 .. .. ... ...


 . . 0 
0 0 · · · 0 λU − Ξ2
Therefore:
 2

− λ1λ−Ξ 1
2 0 ··· ··· 0
 
λ2 2

 0 − λ2 −Ξ2 0 · · · 0 

Ξ
 .. ... ... .. 
Q (n) = Φ(n)G   (Φ(n)G)H



. 0 . 
(C.12)
 .. .. ... ... 

 . . 0 

λU 2
0 0 · · · 0 − λU −Ξ2

141
C. Kalman Filter vs H∞ Filter Appendix C

2 Ξ
where {− λuλ−Ξ
u
2 }u∈{1,2,...,U } are the eigenvalues of Q (n).

It should be noted that when Ξ tends to +∞, QΞ (n) tends to be a zero matrix and
P∞ (n + 1|n) tends to P(n + 1|n).
If Ξ2 > λu , for u ∈ {1, 2, . . . , U }, QΞ (n) is a positive-definite matrix. Therefore,
the solution to the Ricatti equation for the H∞ filter can be seen as an upper
bound of the Kalman a priori error covariance matrix.

142
Appendix D
Extended Kalman Filter Approaches

When considering a non-linear system, the EKF consists in analytically propa-

gating the estimation through the system dynamics, by means of a first-order
linearization using Taylor expansion [Hayk 96].
The non-linear state-space equations that describe the system are:
State equation:
x(n) = Φn (x(n − 1)) + Γw(n) (D.1)

Measurement equation:

y(n) = Ψn (x(n)) + b(n) (D.2)

where Φn and Ψn are non-linear functions. As in appendix A, x(n) is the state

vector of size U at time n and y(n) is the measurement vector of size K at time
n. The noises w(n) and b(n) are additive white zero-mean Gaussian processes
with covariance matrices Q = σw2 IU and R = σb2 IK respectively. In addition, Γ is
the input gain matrix of size U × U .
Then, the first-order Taylor expansion of (D.1) around x̂(n − 1|n − 1) is:

Φn (x(n − 1)) ≈ Φn (x̂(n − 1|n − 1))

+ ∇Φn |x̂(n−1|n−1) {x(n − 1) − x̂(n − 1|n − 1)}
(D.3)
≈ Φn (x̂(n − 1|n − 1)) + ∇Φn |x̂(n−1|n−1) x(n − 1)
− ∇Φn |x̂(n−1|n−1) x̂(n − 1|n − 1)

where ∇Φn |x̂(n−1|n−1) is the Jacobian matrix of Φn evaluated for x̂(n − 1|n − 1).
Given (D.3), (D.1) can be rewritten as follows:

x(n) = Φn (x̂(n − 1|n − 1)) + ∇Φn |x̂(n−1|n−1) x(n − 1)

(D.4)
− ∇Φn |x̂(n−1|n−1) x̂(n − 1|n − 1) + Γw(n)

143
D. Extended Kalman Filter Approaches Appendix D

Given (A.3) and (D.4), the a priori estimation of the state vector is defined as
follows:
x̂(n|n − 1) = Φn (x̂(n − 1|n − 1))
+ ∇Φn |x̂(n−1|n−1) E [x(n − 1)|y(0), y(1), . . . , y(n − 1)]
(D.5)
− ∇Φn |x̂(n−1|n−1) x̂(n − 1|n − 1)}
+ ΓE [w(n)|y(0), y(1), . . . , y(n − 1)]

As w(n) is a white zero-mean Gaussian process, (D.5) can be rewritten as:

x̂(n|n − 1) = Φn (x̂(n − 1|n − 1))
+ ∇Φn |x̂(n−1|n−1) x̂(n − 1|n − 1) (D.6)
− ∇Φn |x̂(n−1|n−1) x̂(n − 1|n − 1)
So, the a priori estimation of the state vector satisfies:

x̂(n|n − 1) = Φn (x̂(n − 1|n − 1)) (D.7)

Thus, given (D.4) and (D.7), the a priori error estimation can be defined as:
x̃(n|n − 1) = x(n) − x̂(n|n − 1)
= Φn (x̂(n − 1|n − 1)) + ∇Φ|x̂(n−1|n−1) x(n − 1)
− ∇Φn |x̂(n−1|n−1) x̂(n − 1|n − 1) + Γw(n) − Φn (x̂(n − 1|n − 1))
= ∇Φn |x̂(n−1|n−1) x̃(n − 1|n − 1) + Γw(n)
(D.8)
and its corresponding covariance matrix P(n|n − 1) is defined by:

P(n|n − 1) = ∇Φn |x̂(n−1|n−1) P(n − 1|n − 1)∇ΦH

n |x̂(n−1|n−1) + ΓQΓ
H
(D.9)

It should be noted the a priori estimation of the state vector (D.7) does not
depend on the Jacobian matrix ∇Φn |x̂(n−1|n−1) . However, ∇Φn |x̂(n−1|n−1) is re-
quired to calculate the a priori error estimation covariance matrix P(n|n − 1).

Now, let us focus our attention on the a posteriori estimation of the state vector.
For this purpose, it is necessary to calculate the first-order Taylor expansion of
(D.2) around x̂(n|n − 1) as follows:

Ψn (x(n)) ≈ Ψn (x̂(n|n − 1)) + ∇Ψn |x̂(n|n−1) {x(n) − x̂(n|n − 1)} (D.10)

144
D. Extended Kalman Filter Approaches

where ∇Ψn |x̂(n|n−1) is the Jacobian matrix of Ψn evaluated for x̂(n|n − 1).
Then, given (D.2) and (D.10), the innovation can be expressed as:

y(n) − Ψn (x̂(n|n − 1)) ≈ Ψn (x(n)) + b(n)

− Ψn (x(n)) + ∇Ψn |x̂(n|n−1) {x(n) − x̂(n|n − 1)} (D.11)
≈ ∇Ψn |x̂(n|n−1) {x(n) − x̂(n|n − 1)} + b(n)

When looking at the linear case and more particularly (A.10), the innovation is
defined as follows:

y(n) − Ψ(n)x̂(n|n − 1) = Ψ(n){x(n) − x̂(n|n − 1)} + b(n) (D.12)

A similarity can be noticed between (D.11) and (D.12). The EKF can be hence
easily derived. The way the state-vector estimation can be updated and the
definition of the Kalman gain can be obtained similarly as in appendix A, by
replacing Ψ(n) by ∇Ψn |x̂(n|n−1) .
Thus, to obtain the a posteriori estimation of the state vector, one has:

x̂(n|n) = x̂(n|n − 1) + K(n){y(n) − Ψn (x̂(n|n − 1))} (D.13)

where K(n) is the Kalman gain defined by:

K(n) = {Pxy (n)}{Pyy (n)}−1 (D.14)

with Pyy (n) the innovation covariance matrix defined as follows:

Pyy (n) = ∇Ψn |x̂(n|n−1) P(n|n − 1)∇ΨH

n |x̂(n|n−1) + R (D.15)

x̂(n − 1|n − 1) x̂(n|n − 1) x̂(n|n)

(D.7) (D.13)

y(n) K(n)

Figure D.1: EKF

145
D. Extended Kalman Filter Approaches Appendix D

and Pxy (n) the cross-covariance matrix between y(n) and the state prediction
x̂(n|n − 1) defined as:

Pxy (n) = P(n|n − 1)∇ΨH

n |x̂(n|n−1) (D.16)

Figure D.1 shows the block scheme of the EKF.

Finally, by combining (D.13) and (D.14), it can be easily shown that the estima-
tion error covariance matrix is updated as follows:

P(n|n) = P(n|n − 1) − K(n)∇Ψn |x̂(n|n−1) (n)P(n|n − 1)

(D.17)
= {IU − K(n)∇Ψn |x̂(n|n−1) (n)}P(n|n − 1)

The EKF may sometimes diverge if the first-order approximation is not sufficient
to describe the non-linearity. Therefore, a second-order approximation can be
used and leads to the SOEKF [Bar 01]. It is still based on a Taylor expansion,
but a second-order expansion is considered.
Given the state-space equations in (D.1) and (D.2), let us look at the second-order
Taylor expansion of (D.1) around x̂(n − 1|n − 1):

Φn (x(n − 1)) ≈ Φn (x̂(n − 1|n − 1))

1
+ ∇Φn |x̂(n−1|n−1) {x(n − 1) − x̂(n − 1|n − 1)} + Φn
2
1
≈ Φn (x̂(n − 1|n − 1)) + ∇Φn |x̂(n−1|n−1) x̃(n − 1|n − 1) + Φn
2
(D.18)
where ∇Φn |x̂(n−1|n−1) is the Jacobian matrix of Φn evaluated at x̂(n − 1|n − 1)
and
U
X ∂ 2 Φn,u
Φn = ΘΦ
u {x(n − 1) − x̂(n − 1|n − 1)}
H
|x̂(n−1|n−1)
u=1 ∂x2 (n)
× {x(n − 1) − x̂(n − 1|n − 1)} (D.19)
U
X
= ΘΦ
u Φn,u
u=1

where Φn,u is the uth element of Φn (x(n)) and ΘΦ

u is a U × 1 vector with zeros
everywhere except for the uth element which is equal to 1.

146
D. Extended Kalman Filter Approaches

Remark 1: The uth quadratic term of Φn can expressed as follows:

∂ 2 Φn,u
Φn,u = tr{ |x̂(n−1|n−1)
∂x2 (n)
× {x(n − 1) − x̂(n − 1|n − 1)}{x(n − 1) − x̂(n − 1|n − 1)}H }
(D.20)
= tr{{x(n − 1) − x̂(n − 1|n − 1)}{x(n − 1) − x̂(n − 1|n − 1)}H
∂ 2 Φn,u
× 2 |x̂(n−1|n−1) }
∂x (n)

Remark 2: one can approximate the expectation of Φn,u given the set of obser-
vations {y(0), . . . , y(n − 1)} by:
mean
Φn,u = E{Φn,u |y(0), . . . , y(n − 1)}
∂ 2 Φn,u (D.21)
≈ tr{P(n − 1|n − 1) |x̂(n−1|n−1) }
∂x2 (n)
mean
The above approximation of Φn,u has the advantage of being “easily” computed
∂ 2 Φn,u
by using P(n − 1|n − 1) and |
∂x2 (n) x̂(n−1|n−1)
.
Given (D.19) and (D.21), one has:
mean
Φn = E{Φn |y(0), . . . , y(n − 1)}
U
X (D.22)
mean
= ΘΦ
u Φn,u
u=1

By following the same development as in (D.5) and (D.6), the a priori estimation
of the state vector can be obtained as follows:
1 mean
x̂(n|n − 1) = Φn (x̂(n − 1|n − 1)) + Φn (D.23)
2
mean
By assuming that Φn ≈ Φn and by combining (D.18) and (D.23), the a priori
estimation error x̃(n|n − 1) can be expressed as follows:

x̃(n|n − 1) = ∇Φn |x̂(n−1|n−1) x̃(n − 1|n − 1) + Γw(n) (D.24)

Therefore, the covariance matrix P(n|n − 1) satisfies:

P(n|n − 1) = ∇Φn |x̂(n−1|n−1) P(n − 1|n − 1)∇ΦH

n |x̂(n−1|n−1) + ΓQΓ
H
(D.25)

147
D. Extended Kalman Filter Approaches Appendix D

Then, the second-order Taylor expansion of (D.2) around x̂(n|n − 1) is:

Ψn (x(n)) ≈ Ψn (x̂(n|n − 1))

1 (D.26)
+ ∇Ψn |x̂(n|n−1) {x(n) − x̂(n|n − 1)} + Ψn
2
where ∇Ψn |x̂(n|n−1) is the Jacobian matrix of Ψn evaluated at x̂(n|n − 1) and
K
X ∂ 2 Ψn,k
Ψn = ΘΨ
k {x(n) − x̂(n|n − 1)}
H
|x̂(n|n−1) {x(n) − x̂(n|n − 1)} (D.27)
k=1 ∂x2 (n)

where Ψn,k is the kth element of Ψn (x(n)) and ΘΨ

k is a K × 1 vector with zeros
everywhere except for the kth element which is equal to 1.
Remark 3: Once again, it should be noted that the kth element of Ψn can be
expressed as follows:

∂ 2 Ψn,k
Ψn,k = tr{ |x̂(n|n−1)
∂x2 (n)
× {x(n) − x̂(n|n − 1)}{x(n) − x̂(n|n − 1)}H }
(D.28)
= tr{{x(n) − x̂(n|n − 1)}{x(n) − x̂(n|n − 1)}H
∂ 2 Ψn,k
|x̂(n|n−1) }
∂x2 (n)

Then, given (D.2) and (D.26) the innovation can be expressed as follows:
1
y(n) − Ψn (x̂(n|n − 1)) ≈ ∇Ψn |x̂(n|n−1) {x(n) − x̂(n|n − 1)} + Ψn + b(n)
2
(D.29)

Unlike the Kalman filter in the linear case or the EKF, the innovation depends
on Ψn . Therefore, to obtain the a posteriori estimation of the state vector, the
following relation can be considered:
1
x̂(n|n) = x̂(n|n − 1) + K(n){y(n) − Ψn (x̂(n|n − 1)) − Ψn } (D.30)
2
where K(n) is the Kalman gain defined in (D.14) and the error estimation co-
variance matrix is updated as in (D.17).

148
D. Extended Kalman Filter Approaches

However, in practical case x(n) − x̂(n|n − 1) is not available and hence Ψn cannot
be computed. For this reason, instead of Ψn one suggests using:
K
X mean
Ψn = ΘΨ
k Ψn,k (D.31)
k=1

mean ∂2Ψ
n,k
where Ψn,k = tr{P(n|n − 1) ∂x2 (n) |x̂(n−1|n−1) }.
As consequence, the state vector is updated as follows:
1
x̂(n|n) = x̂(n|n − 1) + K(n){y(n) − Ψn (x̂(n|n − 1)) − Ψn } (D.32)
2

The IEKF is another alternative that aims at improving the EKF [Gelb 74]. The
purpose is to linearize the measurement model around the updated state x̂(n|n),
instead of the predicted state x̂(n|n − 1). Indeed, the Bayesian approach makes
it possible to separate the Kalman filter in two steps: the prediction step and the
filtering step. The nearer the linearization point is to the real value, the lower
the linearization error should be. In addition, after the filtering step, x̂(n|n) is
expected to have a lower error variance than x̂(n|n − 1). For the above reasons,
using the IEKF can be of interest.
Given the state-space equations in (D.1) and (D.2), the a priori estimation of
the state vector and its covariance matrix can be obtained as in (D.7) and (D.9)
respectively.
To obtain the a posteriori estimation of the state vector, an iterative process
is introduced. At the i + 1th iteration, with i ∈ {0, 1, 2, . . . , Ikf }, where Ikf is
T : |x̂i (n|n) − x̂i+1 (n|n)| < αcvg Iu×1
x̂(n|n − 1) x̂(n|n)
x̂(n − 1|n − 1) = x̂0 (n|n) x̂i+1 (n|n) = x̂i+1 (n|n)
(D.7) (D.33) i ≥1 T
OK OK
y(n)
Not OK Not OK
Ki (n) i = i+1 i = i+1

∇Ψn |x̂i (n|n)

P(n|n − 1)
(D.34) Computing ∇Ψn |x̂i (n|n)
R

Figure D.2: IEKF

149
D. Extended Kalman Filter Approaches Appendix D

the number of iterations, the (i + 1)th a posteriori estimate of the state vector
satisfies:

x̂i+1 (n|n) = x̂(n|n − 1) + Ki (n){y(n) − Ψn (x̂(n|n − 1))} (D.33)

where
Ki (n) = {Pxy yy
i (n)}{Pi (n)}
−1
(D.34)

Pyy H
i (n) = ∇Ψn |x̂i (n|n) P(n|n − 1)∇Ψn |x̂i (n|n) + R (D.35)

Pxy H
i (n) = P(n|n − 1)∇Ψn |x̂i (n|n) (D.36)

and x̂0 (n|n) = x̂(n|n − 1). Figure D.2 shows the block scheme of the IEKF.
The estimation-error covariance matrix has to be updated as follows:

Pi+1 (n|n) = P(n|n − 1) − Ki (n)∇ΨH

n |x̂i (n|n) P(n|n − 1) (D.37)

Then, having a new value of the estimate, the procedure is repeated until the
difference between two consecutive estimates is lower than a predefined value
αcvg :
|x̂i (n|n) − x̂i+1 (n|n)| < αcvg IU ×1 αcvg > 0 (D.38)

where IU ×1 is a U × 1 vector composed of ones.

The relations (D.33)-(D.37) define the IEKF. Nevertheless, it is again a local
approximation and the convergence of the estimate is not guaranteed as in the
EKF.

150
Appendix E
Sigma-Point Kalman Filter

When dealing with the SPKF, the state distribution is still approximated by a
Gaussian distribution, which is now characterized by a set of points lying along
the main eigenaxes of the random variable covariance matrix, as shown in figure
E.1.

−3
x 10
5

4
Probability density

0
10
5 10
0 5
0
−5
−5
x2 −10 −10 x1

Figure E.1: Gaussian distribution approximated by the sigma-points

These so-called sigma-points are propagated through the non-linear system (D.2),
as shown in figure E.2. A weighted combination of the resulting values makes
it possible to estimate the mean and the covariance matrix of the transformed
random vector, i.e. the Gaussian random variable that undergoes the non-linear
transformation.
On the one hand, the UKF is based on the unscented transformation [Wan 01].

151
E. Sigma-Point Kalman Filter Appendix E

sigma-points

Non-linear
transformation
y(n) = Ψn (x(n))

real probability densities

x(n), P(n|n − 1) y(n), Pyy (n)
Figure E.2: Non-linear transformation for a random vector of size 2

When the density is odd, the weights are chosen to provide the exact second-order
Taylor expansion around the mean of the random variable. The UKF was pro-
posed by their authors as an alternative to the EKF to avoid the linearization step.
Nevertheless, Lefebvre et al. [Lefe 02] showed that the sigma-point approach cor-
responds to a weighted statistical linear regression (WSLR). In [VdMe 01], a
square-root UKF is presented. It has better numerical properties and guarantees
the positive semi-definiteness of the underlying state covariance.
On the other hand, the CDKF is based on the second-order Sterling polynomial
interpolation1 formula [Ito 00]. The linearization is based on Sterling polynomial
interpolation around the last available estimate of the state vector. The CDKF
can be seen as a Bayesian approach of a second-order divided difference filter
(DDF). According to Nörgaard et al. [Norg 00], if the last available estimate is
“far” from the real value, the Sterling polynomial interpolation approximation of
the non-linear function is better than a Taylor expansion approximation.
Let us consider the state-space representation done in (D.1) and (D.2). Thus,
as the size of the state vector is U , there are 2U + 1 sigma-points stored in the
following matrix:
h i
X(n − 1|n − 1) = X0 (n − 1|n − 1) Xa (n − 1|n − 1) Xb (n − 1|n − 1) (E.1)

1
Sterling polynomial interpolation [Boas 64] is the approximation in an infinite sum of terms
of a function yi = f (xi )|i∈Z = f (x0 + hu) around a value x0 , where u = xi −xh
0
and h is the
distance between two points chosen arbitrarily. Thus, the second-order Sterling formula of the
2
function f (xi ) is yi = f (xi ) ≈ y0 + u2 (y1 − y−1 ) + u2 (y1 + y−1 − 2y0 ).

152
E. Sigma-Point Kalman Filter

where

X0 (n − 1|n − 1) = x̂(n − 1|n − 1)

q 
⋆
Xa (n − 1|n − 1) = x̂(n − 1|n − 1) + Λ P(n − 1|n − 1) m ∈ {1, . . . , U }
m

q 
⋆
Xb (n − 1|n − 1) = x̂(n − 1|n − 1) − Λ P(n − 1|n − 1) m ∈ {U, . . . , 2U }
m
(E.2)
q q
( P(n − 1|n − 1))m is the mth column of the matrix square-root, (.) is the so-
⋆ ⋆

called matrix square-root that can be obtained by using the lower triangular
Cholesky decomposition and Λ is a scalar scaling factor that determines the
spread of the sigma-points. These points are deterministically chosen so that
they completely capture the true mean and the covariance matrix of the Gaussian
random variable.
The difference between CDKF and UKF stands in the way the mean and the
covariance matrix of the transformed Gaussian random variable are calculated.
The CDKF uses only a single scalar scaling parameter Λ, as opposed to three
√
required by the UKF. When using the CDKF, Λ = 3 [VdMe 04b]. If the UKF
is used, one has:

• Λ = ι2 (U + κ) + U , with 10−4 ≤ ι ≤ 1,

• κ is the secondary scaling parameter usually set to 0 or 3 − U [Juli 95].

• η is the third scaling parameter used to incorporate prior knowledge of the

distribution. For example, when dealing with Gaussian distributions η = 2
is the optimal value [Wan 01].

Let us define the weight scalar parameters for the UKF [Wan 01] as follows:
η
w0(c0) =
η+U
η
w0(c1) = + (1 − ι2 + η) (E.3)
η+U
1
wm(c0) = wm(c1) =
2(U + Λ)

153
E. Sigma-Point Kalman Filter Appendix E

and for the CDKF [VdMe 04b] as:

Λ2 − U
w0(c0) =
Λ2
1
wm(c0) =
2Λ2 (E.4)
1
wm(c1) =
4Λ2
Λ2 − 1
wm(c2) =
4Λ4
The sigma-points are propagated through the non-linear Φn as follows:

X(n|n − 1) = Φn (X(n − 1|n − 1)) (E.5)

Then, the a priori mean is obtained as follows:

2U
X
−
x̂ (n|n − 1) = wm(c0) Xm (n|n − 1) (E.6)
m=0

while the a priori estimation error covariance matrix when using the UKF satis-
fies:
2U
X h i⋆
P(n|n − 1) = wm(c1) Xm (n|n − 1) − x̂− (n|n − 1) +Q (E.7)
m=0

where [(.)]⋆ = [(.)] × [(.)]H .

When using the CDKF, one has:
U
X
P(n|n − 1) = c1
{wm [Xm (n|n − 1) − Xm+U (n|n − 1)]⋆
m=0 (E.8)
+ wmc2 [Xm (n|n − 1) + Xm+U (n|n − 1) − 2X0 (n|n − 1)] } + Q ⋆

Then, the a priori sigma points are calculated as follows:

h i
X(n|n − 1) = X0 (n|n − 1) Xa (n|n − 1) Xb (n|n − 1) (E.9)

where:
X0 (n|n − 1) = x̂(n|n − 1)
q 
⋆
Xa (n|n − 1) = x̂(n|n − 1) + Λ P(n|n − 1) m ∈ {1, . . . , U } (E.10)
q m
⋆
Xb (n − 1|n − 1) = x̂(n|n − 1) − Λ P(n|n − 1) m ∈ {U, . . . , 2U }
m

154
E. Sigma-Point Kalman Filter

Once the a priori sigma points are calculated, they are propagated through the
non-linear function:
Y(n|n − 1) = Ψn (X(n|n − 1)) (E.11)
The mean and the covariance of y(n) are approximated by using the posterior
sigma points as follows:
2U
X
ŷ− (n) = wmc0 Ym (n|n − 1) (E.12)
m=0

where Ym (n|n − 1) = Φn (X(n|n − 1)).

Then, the state vector is estimated recursively as follows:

x̂(n|n) = x̂(n|n − 1) + K(n){y(n) − ŷ− (n)} (E.13)

where K(n) is the Kalman gain defined as:

K(n) = {Pxy (n)}{Pyy (n)}−1 (E.14)

On the one hand, when using the UKF, the covariance matrix Pyy (n) of the
innovation ỹ(n) = y(n) − ŷ− (n) and the cross-covariance matrix Pxy (n) between
y(n) and the state prediction error satisfy:
2U
X h i⋆
yy
P (n) = wmc1 Ym (n|n − 1) − ŷ− (n) +R (E.15)
m=0
2U
X h i h i
xy
P (n) = wmc1 Xm (n|n − 1) − x̂− (n) × Ym (n|n − 1) − ŷ− (n) (E.16)
m=0
On the other hand, when using the CDKF, the covariance matrix of the innova-
tion and the cross-covariance matrix between y(n) and the state prediction error
satisfy:
U
X
yy
P (n) = {wmc1 [Ym (n|n − 1) − Ym+U (n|n − 1)]⋆
m=0 (E.17)
+ wmc2 [Ym (n|n − 1) + Ym+U (n|n − 1) − 2Y0 (n|n − 1)] } + R ⋆

q
Pxy (n) = ⋆ wmc1 P(n|n − 1) [Y1:U (n|n − 1) − YU +1:2U (n|n − 1)]H (E.18)
Finally, the estimation error covariance matrix is updated as follows:

P(n|n) = P(n|n − 1) − K(n)Pyy (n)KH (n) (E.19)

155
Appendix F
Extended and Unscented H∞ Filter

When a non-linear estimation issue is addressed, alternatives to the Kalman fil-

tering approaches are the so-called “extended H∞ filter” and the “unscented H∞
filter”.
Indeed, initial works about the “extended H∞ filter” were conducted by Burl
[Burl 98]. Like the EKF, it consists of a first-order linearization around the last
available estimation of the state vector. Since, various authors have used the
“extended H∞ filter. See for instance, Giremus et al. [Gire 09] who study the
relevance of H∞ filtering in the field of the global positioning system navigation.
However, the authors do not obtain noticeable improvements in terms of posi-
tioning error in comparison with the Kalman filtering.
Let us consider the equations (D.1), (D.2) and (B.1), that define the state-space
of the system. The matrix P∞ (n + 1|n) satisfies the following Riccati equation
for the “extended H∞ filter”:

P∞ (n + 1|n) = ∇ΦH ∞ H
n+1 |x̂(n|n) P (n|n)∇Φn+1 |x̂(n|n) + ΓWΓ
H
(F.1)
h i
H
= ∇Φn+1 |x̂(n|n) P∞ (n|n − 1){IU − ∇ΨH
n |x̂(n|n−1) L
" #
−1 ∇Ψn |x̂(n|n−1)
×M P∞ (n|n − 1)}∇ΦH
n+1 |x̂(n|n)
L
+ΓWΓH (F.2)

where ∇ΦH
n+1 |x̂(n|n) is the Jacobian matrix of Φn+1 evaluated for x̂(n|n), ∇Ψn |x̂(n|n−1)
is the Jacobian matrix of Ψn evaluated for x̂(n|n − 1) and
" # " #
V 0 ∇Ψn |x̂(n|n−1) h
H
i
M= + P∞ (n|n − 1) ∇ΨH
n |x̂(n|n−1) L
0 −Ξ2 IU L
" H
#
∇Ψn |x̂(n|n−1) P∞ (n|n − 1)∇ΨH ∞
n |x̂(n|n−1) + V ∇Ψn |x̂(n|n−1) P (n|n − 1)L
=
LP∞ (n|n − 1)∇ΨH n |x̂(n|n−1) LP∞ (n|n − 1)LH − Ξ2 IU
(F.3)

157
F. Extended and Unscented H∞ Filter Appendix F

where Ξ2 is the prescribed noise attenuation level given in (B.3).

When using the “extended H∞ filter”, the state vector can be estimated recur-
sively as follows:

x̂(n|n) = x̂(n|n − 1) + K∞ (n){y(n) − Ψn (x̂(n|n − 1))} (F.4)

where K∞ (n) is the filter gain defined as:

K∞ (n) = {Pxy∞ (n)}{Pyy∞ (n)}−1 (F.5)

where Pyy∞ (n) is defined as follows:

Pyy∞ (n) = ∇Ψn |x̂(n|n−1) P∞ (n|n − 1)∇ΨH

n |x̂(n|n−1) + V (F.6)

and Pxy∞ (n) is:

Pxy∞ (n) = P∞ (n|n − 1)∇ΨH
n |x̂(n|n−1) (F.7)

Remark 1: when Ξ2 tends to +∞, the “extended H∞ filter” reduces to the EKF
[Hass 99].
Remark 2: the computational cost of the “extended H∞ filter” is slightly higher
than the EKF.

The “unscented H∞ filter” was recently proposed by Li et al. in [Li 10]. It is

implemented by embedding the unscented transformation into the “extended H∞
filter” architecture. According to the authors, the unscented H∞ filtering can be
carried out by using the statistical linear error propagation approach [Sibl 06].
Like the UKF, the unscented H∞ filter avoids the linearization step by using an
unscented transformation. In [Wang 10], an unscented H∞ filter is used to pro-
vide an initial alignment of an inertial navigation system. The authors show that
the H∞ filtering approach would be an effective method when the measurement
noise is colored.
According to [Li 10] and given (D.1), (D.2) and (B.1), when using the “unscented
H∞ filter”, the state vector can be estimated as follows:

x̂(n|n) = x̂(n|n − 1) + K∞ (n){y(n) − ŷ− (n)} (F.8)

158
F. Extended and Unscented H∞ Filter

where ŷ− (n) is obtained as in (E.12) by performing the same process as in the
UKF, but by replacing the matrix Q by W. The gain K∞ (n) is obtained as in
(F.5), but Pxy∞ (n) and Pyy∞ (n) are now approximated as follows:
2U
X h i⋆
Pyy∞ (n) ≈ wmc1 Ym (n|n − 1) − ŷ− (n) +V (F.9)
m=0

and
2U
X h i h i
Pxy∞ (n) ≈ wmc1 Xm (n|n − 1) − x̂− (n) × Ym (n|n − 1) − ŷ− (n) (F.10)
m=0

where the wmc1 , Xm (n|n − 1), x̂− (n) and Ym (n|n − 1) are obtained as in (E.3),
(E.5), (E.6) and (E.11) respectively, by the using the UKF process.
Considering L = IU , given (F.6), (F.7), (F.9) and (F.10) the matrix M in (F.3)
can be rewritten as:
" #
Pyy∞ (n) {Pxy∞ (n)}H
M= (F.11)
Pxy∞ (n) P∞ (n|n − 1) − Ξ2 IU

Finally, the matrix P∞ (n|n) can be obtained as:

h i
P∞ (n|n) = P∞ (n|n − 1) − Pxy∞ (n) P∞ (n|n − 1)
" #
−1 {Pxy∞ (n)}H
×M (F.12)
P∞ (n|n − 1)

Remark 1: the unscented transform is used to approximate the matrices that

are involved in the definition of the H∞ gain.
Remark 2: concerning the work proposed by Li et al. in [Li 10], we have decided
to look at it and evaluate its performance even if we are not totally convinced by
the way the authors have motivated their approach and justified it. On the one
hand, in their paper [Li 10], they speak of covariance matrices and define some
quantities using the expectation. This should not be done as they are working in
an H∞ setting, where there is not statistical mean to be considered. On the other
hand, according to appendix C, we have seen that the equations describing the
H∞ filter could be seen as a kind of Kalman filter. Therefore, we can understand
that the authors in [Li 10] consider that the solution of the Ricatti equation in
the H∞ setting can be viewed as a covariance matrix.

159
References

[3GPP 09] 3GPP Standard. Technical Specification Group Radio Access Net-
work; Evolved Universal Terrestrial Radio Access (E-UTRA); Physi-
cal layer procedures (Release 8). 3GPP TS 36.213 V8.8.0, September
2009. 1, 25, 26

[Alam 98] S. M. Alamouti. A Simple Transmit Diversity Technique for Wireless

Communications. IEEE Journal on Selected Areas in Communica-
tions, Vol. 16, No. 8, pp. 1451–1458, October 1998. 116, 120

[Arsl 07] H. Arslan. Cognitive Radio, Software Defined Radio, and Adaptive
Wireless Systems. Springer, 2007. 10

[Bar 01] Y. Bar-Shalom, X. R. Li, and T. Kirubarajan. Estimation With

Applications to Tracking and Navigation. Wiley, 2001. 4, 56, 146

[Bass 93] M. Basseville and I. V. Nikiforov. Detection of Abrupt Changes:

Theory and Application. Prentice Hall, 1993. 88

[Berr 93] C. Berrou, A. Glavieux, and P. Thitimajshima. Near Shannon Limit

Error-Correcting Coding and Decoding: Turbo-codes. Proceedings of
the IEEE International Conference on Communications (ICC ’93),
Vol. 2, pp. 1064–1070, May 1993. 40

[Bian 05] H. Bian, Z. Jin, and W. Tian. Study on GPS Attitude Determination
System Aided INS Using Adaptive Kalman Filter. Measurement
Science and Technology, Vol. 6, No. 10, pp. 2072–2079, September
2005. 80

[Bing 90] J. Bingham. Multicarrier Modulation for Data Transmission: An

Idea Whose Time Has Come. IEEE Communication Magazine,
Vol. 28, pp. 5–14, May 1990. 23

161
References References

[Boas 64] R. P. Boas Jr. and R. C. Buck. Polynomial Expansions of Analytic

Functions. Springer-Verlag, 1964. 152

[Brez 88] C. Brezinski. Other Manifestations of the Schur Complement. Linear

Algebra and its Applications, Vol. 111, pp. 231–247, December 1988.
139

[Broe 03] L. Broetje, S. Vogeler, and K. D. Kammeyer. Blind Bluetooth Inter-

ference Detection and Suppression for OFDM Transmission in the
ISM Band. Proceedings of the IEEE Asilomar Conference on Sig-
nals, Systems and Computers (ACSSC ’03), Vol. 1, pp. 703–707,
November 2003. 79

[Burl 98] J. B. Burl. H∞ Estimation for Nonlinear Systems. IEEE Signal

Processing Letters, Vol. 5, No. 8, pp. 199–202, August 1998. 57, 157

[Cabr 04] D. Cabric, S. M. Mishra, and R. W. Brodersen. Implementation

Issues in Spectrum Sensing for Cognitive Radios. Proceedings of
the IEEE Asilomar Conference on Signals, Systems and Computers
(ACSSC ’04), Vol. 1, pp. 772–776, November 2004. 79

[Cao 04a] Z. Cao, U. Tureli, Y. D. Yao, and P. Honan. Frequency Synchroniza-

tion for Generalized OFDMA Uplink. Proceedings of the IEEE Global
Communications Conference (GLOBECOM ’04), pp. 1071–1075,
December 2004. 51, 75, 104, 113, 115, 123, 125, 129

[Cao 04b] Z. Cao, U. Tureli, and Y. D. Yao. Deterministic Multiuser Carrier-

Frequency Offset Estimation for Interleaved OFDMA Uplink. IEEE
Transactions on Communications, Vol. 52, No. 9, pp. 1585–1594,
September 2004. 52, 75

[Chen 10] C. J. Chen, T. T. Lin, and T. C. Chen. Blind Joint Frequency Off-
set and Channel Estimation for MC-CDMA Systems over Multipath
Fading Channels. Proceedings of the Mechanical and Electronics En-
gineering (ICMEE ’10), pp. 137–141, August 2010. 104

162
References References

[Chie 07] F. T. Chien and C. C. J. Kuo. Blind Recursive Tracking of

Carrier Frequency Offset (CFO) Vector in MC-CDMA Systems.
IEEE Transactions on Wireless Communications, Vol. 6, No. 4,
pp. 1246–1255, April 2007. 103

[Choi 00] J. H. Choi, Y. H. Lee, C. G. Lee, and H. W. Jung. Carrier Fre-

quency Offset Compensation for Uplink of OFDM-FDMA Systems.
Proceedings of the IEEE International Conference on Communica-
tions (ICC ’00), Vol. 1, pp. 425–429, June 2000. 50, 51, 104

[Cibl 04] P. Ciblat and E. Serpedin. A Fine Blind Frequency Offset Estimator
for OFDM/OQAM Systems. IEEE Transactions on Signal Process-
ing, Vol. 52, No. 1, pp. 291–296, January 2004. 52

[Cimi 85] L. J. Cimini. Analysis and Simulation of a Digital Mobile Channel

Using Orthogonal Frequency Division Multiplexing. IEEE Transac-
tions on Communications, Vol. COM-33, No. 7, pp. 665–675, July
1985. 23

[Dai 07] X. Dai. Carrier Frequency Offset Estimation and Correction

for OFDMA Uplink. IET Commununications, Vol. 1, No. 2,
pp. 273–281, April 2007. 49

[Dash 00] P. K. Dash, R. K. Jena, G. G. Panda, and A. Routray. An Extended

Complex Kalman Filter for Frequency Measurement of Distorted
Signals. IEEE Transactions on Instrumentation and Measurement,
Vol. 49, No. 4, pp. 746–753, August 2000. 55

[Dini 10] D. H. Dini and D. P. Mandic. Analysis of the Widely Linear Complex
Kalman Filter. Proceedings of the UDRC Sensor Signal Processing
for Defense (SSPD ’10), pp. 1–4, September 2010. 55

[ETSI 01] ETSI Standard. Interaction Channel for Digital Terrestrial Televi-
sion (RCT) Incorporating Multiple Access OFDM. ETSI DVB RCT,
March 2001. 24

163
References References

[ETSI 95] ETSI Standard. Radio Broadcasting Systems: Digital Audio Broad-
casting to Mobile, Portable and Fixed Receivers. ETS 300 401,
European Telecommunications Standard, 1995. 24

[ETSI 97] ETSI Standard. Digital Video Broadcasting (DVB-T); Frame Struc-
ture, Channel Coding, Modulation for Digital Terrestrial Television.
ETS 300 744, European Telecommunications Standard, 1997. 24

[Ferr 10] G. Ferré and M. Raoult. Flexible Distributed Wideband Cognitive

Radio Network with Double Threshold Energy Detector Combining
Cooperative and Spatial Diversity. Proceedings of the EURASIP Eu-
ropean Signal Processing Conference (EUSIPCO ’10), pp. 880–884,
August 2010. 79

[Foki 09] L. Fokin and A. G. Shchipitsyn. Innovation-Based Adaptive

Kalman Filter Derivation. Proceedings of the IEEE International
Sierian Conference on Control and Communications (SIBCON ’09),
pp. 318–323, March 2009. 80

[Fu 06] X. Y. Fu, H. Minn, and C. Cantrell. Two Novel Iterative Joint
Frequency-Offset and Channel Estimation Methods for OFDMA Up-
link. Proceedings of the IEEE Global Communications Conference
(GLOBECOM ’06), pp. 1–6, November 2006. 2, 52

[Gelb 74] A. Gelb. Applied Optimal Estimation. MIT Press, 1974. 149

[Gero 99] J. C. Geromel. Optimal Linear Filtering Under Parameter Un-

certainty. IEEE Transactions on Signal Processing, Vol. 47,
pp. 168–175, January 1999. 135

[Gire 09] A. Giremus, E. Grivel, and F. Castanié. Is H∞ Filter Relevant for

Correlated Noises in GPS Navigation. Proceedings of the IEEE Inter-
national Conference on Digital Signal Processing (DSP ’09), pp. 1–6,
July 2009. 5, 157

164
References References

[Grew 10] M. S. Grewal and J. Kain. Kalman Filter Implementation with

Improved Numerical Properties. IEEE Transactions on Automatic
Control, Vol. 55, No. 9, pp. 2058–2068, September 2010. 56

[Grim 90] M. J. Grimble and A. E. Sayed. Solution of the H∞ Optimal Linear

Filtering Problem for Discrete-time Systems. IEEE Transactions on
Acoustics, Speech, and Signal Processing, Vol. 38, pp. 1092–1104,
July 1990. 135, 138

[Grol 07] J. Grolleau. Modelès et Estimateurs de Canaux de Rayleigh pour la

Detéction de Symboles dans un Système Multiporteuse à Étalement
de Spectre. PhD thesis, under the supervision of E. Grivel and M. Na-
jim, Université Bordeaux 1, Ecole Doctorale des Sciences Physiques
et de l’Ingénieur, December 2007. In French. 1

[Hanz 98] L. Hanzo. Bandwidth-Efficient Wireless Multimedia Communica-

tions. Proceedings of the IEEE, Vol. 86, pp. 1342–1382, July 1998.
8, 19, 20

[Hass 99] B. Hassibi, A. H. Sayed, and T. Kailath. Indefinite-Quadratic Es-

timation and Control, a Unified Approach to H2 and H∞ Theories.
Society for Industrial and Applied Mathematics (SIAM), 1999. 57,
158

[Hayk 96] S. Haykin. Adaptative Filter Theory. Prentice Hall, 1996. 56, 143

[Hou 08] S. W. Hou and C. C. Ko. Intercarrier Interference Suppression for

OFDMA Uplink in Time and Frequency Selective Rayleigh Fading
Channels. Proceedings of the IEEE Vehicular Technology Conference
(VTC ’08), pp. 1438–1442, May 2008. 3, 5, 51, 68, 70

[Hu 11] J. S. Hu and C. H. Yang. Second-Order Extended H∞ Filter for Non-

linear Discrete-Time Systems Using Quadratic Error Matrix Approx-
imation. IEEE Transactions on Signal Processing, Vol. 59, No. 7,
pp. 3110–3119, July 2011. 130

165
References References

[Huan 05] D. Huang and K. B. Letaief. An Interference Cancellation Scheme

for Carrier Frequency Offsets Correction in OFDMA Systems. IEEE
Transactions on Communications, Vol. 53, No. 7, pp. 1155–1165,
July 2005. 50, 75, 104

[Huan 10a] Q. F. Huang, M. Ghogho, J. Wei, and P. Ciblat. Practical Tim-

ing and Frequency Synchronization for OFDM-Based Cooperative
Systems. IEEE Transactions on Signal Processing, Vol. 58, No. 7,
pp. 3706–3716, July 2010. 49

[Huan 10b] W. C. Huang, C. P. Li, and H. J. Li. An Investigation into the

Noise Variance and the SNR Estimators in Imperfectly-Synchronized
OFDM Systems. IEEE Transactions on Wireless Communications,
Vol. 9, No. 3, pp. 1159–1167, March 2010. 86, 107

[Huan 11] Y. Huang, H. Li, K. A. Campbell, and Z. Han. Defending False Data
Injection Attack on Smart Grid Network Using Adaptive CUSUM
Test. Proceedings of the IEEE Conference on Information Sciences
and Systems (CISS ’11), pp. 1–6, March 2011. 89

[IEEE 06] IEEE Standard. Part 16: Air Interface for Fixed and Mobile Broad-
band Wireless Access System; Amendment 2: Physical and Medium
Access Control Layers for Combined Fixed and Mobile Operation in
Licensed Bands. IEEE Std. 802.16e-2005, February 2006. 1, 3, 24

[IEEE 99] IEEE Standard. Part 11: Wireless LAN Medium Access Con-
trol (MAC) and Physical Layer (PHY) Specifications; Higher-Speed
Physical Layer Extension in the 5 GHz Band. IEEE Std. 802.11a,
December 1999. 3, 24

[Ito 00] K. Ito and K. Xiong. Gaussian Filters for Nonlinear Filtering Prob-
lems. IEEE Transactions on Automatic Control, Vol. 5, No. 5,
pp. 910–927, May 2000. 56, 130, 152

[ITU 08] ITU Report. ITU Corporate Annual Report 2008. pp. 38–40, June
2008. 9

166
References References

[Jake 74] W. C. Jakes. Microwave mobile communications. Wiley-IEEE Press,

1974. 13, 15

[Jamo 07a] A. Jamoos, J. Grolleau, E. Grivel, H. A. Nour, and M. Najim.

Kalman vs H∞ Algorithms for MC-DS-CDMA Channel Estimation
with or without a priori AR Modeling. Lecture Notes in Electrical
Engineering, 2007. 135

[Jamo 07b] A. Jamoos. Contribution à la Mise en Œuvre de Récepteurs et

de Techniques d’Estimation de Canal pour les Systèmes Mobiles
DS-CDMA Multi-porteuse. PhD thesis, under the supervision of
E. Grivel and M. Najim, Université Bordeaux 1, Ecole Doctorale
des Sciences Physiques et de l’Ingénieur, June 2007. 1

[Jing 08] Y. Jing, F. L. Yin, and Z. Chen. An H∞ Filter Based Approach to

Combat Inter-Carrier Interference for OFDM Systems. IEEE Com-
munications Letters, Vol. 12, No. 6, pp. 453–455, June 2008. 2, 47,
49

[Juli 95] S. J. Julier, J. K. Uhlmann, and H. F. Durrant-Whyte. A New Ap-

proach for Filtering Nonlinear Systems. Proceedings of the IEEE
American Control Conference (ACC ’95), Vol. 3, pp. 1628–1632,
June 1995. 153

[Kalm 60] R. E. Kalman. A New Approach to Linear Filtering and Prediction

Problems. Transactions of the ASME - Journal of Basic Engineering,
pp. 34–45, March 1960. 131

[Kang 07] T. H. Kang and R. A. Iltis. Iterative Decoding, Offset and Chan-
nel Estimation for OFDM using the Unscented Kalman Filter.
Proceedings of the IEEE Asilomar Conference on Signals, Systems
and Computers (ACSSC ’07), pp. 1828–1732, November 2007. 54,
55

167
References References

[Kell 00] T. Keller and L. Hanzo. Adaptive Multicarrier Modulation: A Con-

venient Framework for Time-Frequency Processing in Wireless Com-
munications. Proceedings of the IEEE, Vol. 88, No. 5, pp. 611–640,
May 2000. 23

[Laba 05] D. Labarre, E. Grivel, M. Najim, and N. Christov. Relevance of

H∞ Filtering for Speech Enhancement. Proceedings of the IEEE In-
ternational Conference on Acoustics, Speech, and Signal Processing
(ICASSP ’05), Vol. 4, pp. 169–172, March 2005. 5, 135

[Laba 07] D. Labarre, E. Grivel, M. Najim, and N. Christov. Dual H∞

Algorithms for Signal Processing-Application to Speech Enhance-
ment. IEEE Transactions on Signal Processing, Vol. 55, No. 11,
pp. 5195–5208, November 2007. 135

[Lefe 02] T. Lefebvre, H. Bruyninckx, and J. D. Schutter. Comment on a New

Method for the Nonlinear Transformation of Means and Covariances
in Filters and Estimators. IEEE Transactions on Automatic Control,
Vol. 7, No. 8, pp. 1406–1408, August 2002. 152

[Li 07] T. Li, W. H. Mow, V. K. N. Lau, M. H. Siu, R. S. Cheng, and

R. D. Murch. Robust Joint Interference Detection and Decoding for
OFDM-Based Cognitive Radio Systems with Unknown Interference.
IEEE Journal on Selected Areas Communications, Vol. 25, No. 3,
pp. 566–575, April 2007. 79

[Li 10] W. L. Li and Y. M. Jia. H∞ Filtering for a Class of Nonlinear

Discrete-Time Systems based on Unscented Transform. Signal Pro-
cessing, Vol. 90, No. 12, pp. 3301–3307, December 2010. 5, 57, 158,
159

[Maha 06] I. Mahafeno, C. Langlais, and C. Jego. OFDM-IDMA Versus IDMA

with ISI Cancellation for Quasi-Static Rayleigh Fading Multipath
Channels. Proceedings of the VDE International Symposium on
Turbo Codes in Connection with International ITG-Conference on

168
References References

Source and Channel Coding, April 2006. 4, 26, 40, 41, 42, 43, 102,
104, 112

[Maha 07] I. Mahafeno. Etude de la Technique d’Accès Multiple IDMA (Inter-

leave Division Multiple Access). PhD thesis, under the supervision
of C. Berrou, Ecole Nationale Supérieure des Télécommunications
de Bretagne, May 2007. In French. 38

[Mare 06] H. Marey and M. Steendam. The Effect of Narrowband Interfer-

ence on Frequency Ambiguity Resolution for OFDM. Proceedings
of the IEEE Vehicular Technology Conference (VTC ’06), pp. 1–5,
September 2006. 79

[Mito 00] J. Mitola III. An Integrated Agent Architecture for Software Defined
Radio. PhD thesis, under the supervision of G. Q Maguire, Royal
Institute of Technology (KTH), May 2000. 3, 9

[Mito 99] J. Mitola III and G. Q. Maguire Jr. Cognitive Radio: Making Soft-
ware Radios More Personal. IEEE Personal Communications, Vol. 6,
No. 4, pp. 13–18, August 1999. 8, 9

[Moos 94] P. H. Moose. A Technique for Orthogonal Frequency Division Mul-

tiplexing Frequency Offset Correction. IEEE Transactions on Com-
munications, Vol. 42, No. 10, pp. 2908–2914, October 1994. 48, 50,
51, 107

[More 04] M. Morelli. Timing and Frequency Synchronization for the Uplink
of an OFDMA System. IEEE Transactions on Communications,
Vol. 52, No. 2, pp. 296–306, February 2004. 2, 47, 49

[More 07] M. Morelli, C. C. J. Kuo, and M. O. Pun. Synchronization

Techniques for Orthogonal Frequency Division Multiple Access
(OFDMA): A Tutorial Review. Proceedings of the IEEE, Vol. 95,
No. 7, pp. 1394–1427, July 2007. 48

169
References References

[More 08] M. Morelli and M. Moretti. Robust Frequency Synchronization

for OFDM-Based Cognitive Radio Systems. IEEE Transactions on
Wireless Communications, Vol. 7, No. 12, pp. 5346–5355, December
2008. 80, 92, 93, 97, 98, 100, 129

[Mova 08] M. Movahedian, M. Yi, and R. Tafazolli. An MUI Resilient Ap-

proach for Blind CFO Estimation in OFDMA Uplink. Proceedings
of the IEEE International Symposium on Personal, Indoor and Mo-
bile Radio Communications (PIMRC ’08), pp. 1–5, September 2008.
53, 75

[Nish 99] K. Nishiyama. Robust Estimation of a Single Complex Sinusoid in

White Noise - H∞ Filtering Approach. IEEE Transactions on Signal
Processing, Vol. 47, No. 10, pp. 2853–2856, October 1999. 55

[Norg 00] M. Nørgard, N. K. Poulsen, and O. Ravn. New Developments

in State Estimation for Nonlinear Systems. Automatica, Vol. 36,
pp. 1627–1638, November 2000. 152

[Pali 10] J. Palicot. De la Radio Logicielle à la Radio Intelligente. Hermes

Science, 2010. In French. 10

[Ping 01] L. Ping, W. K. Leung, and K. Y. Wu. Low Rate Turbo-Hadamard

Codes. Proceedings of the IEEE International Symposium on Infor-
mation Theory, p. 211, August 2001. 40

[Ping 02a] L. Ping, K. Y. Wu, L. H. Liu, and W. K. Leung. A Simple Uni-

fied Approach to Nearly Optimal Multiuser Detection and Space-
Time Coding. Proceedings of the IEEE Information Theory Work-
shop (ITW ’02), pp. 53–56, October 2002. 3, 6, 26, 112

[Ping 02b] L. Ping, K. Y. Wu, L. Liu, and W. K. Leung. A Simple, Unified

Approach to Nearly Optimal Multiuser Detection and Space-Time
Coding. Proceedings of the IEEE Information Theory Workshop
(ITW ’02), pp. 53–56, October 2002. 115

170
References References

[Ping 06] L. Ping, L. H. Liu, K. Y. Wu, and W. K. Leung. Interleave-Division

Multiple-Access. IEEE Transactions on Communications, Vol. 5,
No. 4, pp. 938–947, April 2006. 26

[Pove 09a] H. Poveda, G. Ferré, and E. Grivel. Estimation Aveugle Itérative par
Filtrage H∞ des Décalages Doppler d’un Système OFDMA Entrelacé
en Liaison Montante. Actes de la Conférence Groupe de Recherche et
d’Etudes du Traitement du Signal (GRETSI ’09), September 2009.
In French. 3, 5, 128

[Pove 09b] H. Poveda, G. Ferré, E. Grivel, and P. Ramos. A Blind Itera-

tive Carrier Frequency Offset Estimator based on a Kalman Ap-
proach for an Interleaved OFDMA Uplink System. Proceedings of the
EURASIP European Signal Processing Conference (EUSIPCO ’10),
pp. 378–382, August 2009. 3, 5, 128

[Pove 10] H. Poveda, G. Ferré, and E. Grivel. Joint Channel and Frequency
Offset Estimation Using Sigma-Point Kalman Filter for an OFDMA
Uplink System. Proceedings of the EURASIP European Signal Pro-
cessing Conference (EUSIPCO ’10), pp. 855–859, August 2010. 4,
54, 86, 127

[Pove 11a] H. Poveda, G. Ferré, N. Bouny, and E. Grivel. Estimation Jointe

de Multiples Décalages de Fréquence dans un Système Coopératif à
Relais OFDM-IDMA. Actes de la Conférence Groupe de Recherche
et d’Etudes du Traitement du Signal (GRETSI ’11), September 2011.
In French. 3, 6, 129

[Pove 11b] H. Poveda, G. Ferré, and E. Grivel. Frequency Offset and Channel
Estimations for an OFDMA Uplink Primary User System Despite
a Cognitive Radio Interference. To be submitted IEEE Transactions
on Wireless Communications, 2011. 3, 5, 80, 129

[Pove 11c] H. Poveda, G. Ferré, and E. Grivel. A Non-Pilot Aided Iterative

Carrier Frequency Offset Estimator Using Optimal Filtering for an

171
References References

Interleaved OFDMA Uplink System. Accepted to Wireless Personal

Communications, 2011. 3, 5, 66, 128

[Pove 11d] H. Poveda, G. Ferré, and E. Grivel. Robust Frequency Synchroniza-

tion for an OFDMA Uplink System Disturbed by a Cognitive Radio
System Interference. Proceedings of the IEEE International Con-
ference on Acoustics, Speech, and Signal Processing (ICASSP ’11),
pp. 3552–3555, May 2011. 3, 5, 80, 129

[Pove 12] H. Poveda, G. Ferré, and E. Grivel. Frequency Synchronization

and Channel Equalization for an OFDM-IDMA Uplink system.
Submitted to IEEE International Conference on Acoustics, Speech,
and Signal Processing (ICASSP ’12), March 2012. 3, 6, 129

[Proa 95] J. G. Proakis. Digital Communications. McGraw-Hill, 1995. 22, 102

[Pun 04a] M. O. Pun, S. H. Tsai, and C. C. J. Kuo. An EM-Based Joint

Maximum Likelihood Estimation of Carrier Frequency Offset and
Channel for Uplink OFDMA Systems. Proceedings of the IEEE Ve-
hicular Technology Conference (VTC ’04), pp. 598–602, September
2004. 2, 47, 51, 58, 59, 61, 64, 104, 128

[Pun 04b] M. O. Pun, S. H. Tsai, and C. C. J. Kuo. Joint Maximum Likeli-

hood Estimation of Carrier Frequency Offset and Channel for Uplink
OFDMA Systems. Proceedings of the IEEE Global Communications
Conference (GLOBECOM ’04), Vol. 6, pp. 3748–3752, November
2004. 51

[Pun 06] M. O. Pun, M. Morelli, and C. C. J. Kuo. Maximum-Likelihood

Synchronization and Channel Estimation for OFDMA Uplink Trans-
missions. IEEE Transactions on Communications, Vol. 54, No. 4,
pp. 726–736, April 2006. 2, 47, 51, 52, 58, 59, 61, 64, 104, 128

[Quan 08] Z. Quan, S. G. Cui, A. H. Sayed, and H. Poor. Optimal Multi-

band Joint Detection for Spectrum Sensing in Cognitive Radio Net-
works. IEEE Transactions on Signal Processing, Vol. 57, No. 3,
pp. 1128–1140, March 2008. 79

172
References References

[Rapp 02] T. S. Rappaport. Wireless Communications: Principles and Prac-

tice. Prentice Hall, 2002. 9

[Rehm 08] H. U. Rehman, I. Zaka, S. Shah, and J. Ahmad. Combined Equaliza-

tion and Channel Estimation for MC-IDMA Uplink Transmissions.
Proceedings of the IEEE Networking and Communications Confer-
ence (INCC ’08), pp. 34–38, May 2008. 5, 103

[Roli 09] C. A. Rolim Fernandes. Nonlinear MIMO Communication Systems:

Channel Estimation and Information Recovery using Volterra Mod-
els. PhD thesis, under the supervision of G. Favier and J. C. Mota,
Université de Nice Sophia Antipolis, Ecole Doctorale Sciences et
Technologies de l’Information et de la Communication, July 2009.
130

[Saem 07] A. Saemi, J. P. Cances, and V. Meghdadi. Synchronization Algo-

rithms for MIMO OFDMA Systems. IEEE Transactions on Wireless
Communications, Vol. 6, No. 12, pp. 4441–4451, December 2007. 52

[Sari 98] H. Sari and G. Karam. Orthogonal Frequency Division Multiple Ac-
cess and its Application to CATV Networks. European Transactions
on Telecommunications, Vol. 45, pp. 507–516, November 1998. 24

[Schu 05] H. Schulze and C. Lüders. Theory and Applications of OFDM and
CDMA. Wiley, 2005. 21

[Sesi 09] S. Sesia, I. Toufik, and M. Baker. LTE, The UMTS Long Term
Evolution: From Theory to Practice. Wiley, 2009. 22, 26

[Sezg 08] S. Sezginer and P. Bianchi. Asymptotically Efficient Reduced Com-

plexity Frequency Offset and Channel Estimators for Uplink MIMO-
OFDMA Systems. IEEE Transactions on Signal Processing, Vol. 56,
No. 3, pp. 964–979, March 2008. 2, 47, 52, 64, 104

[Shak 92] U. Shaked and Y. Theodor. H∞ - Optimal Estimation: A Tutorial.

Proceedings of IEEE Conference on Decision and Control, Vol. 2,
pp. 2278–2286, August 1992. 135

173
References References

[Shen 99] X. Shen and L. Deng. A Dynamic System Approach to Speech En-
hancement Using the H∞ Filtering Algorithm. IEEE Transactions
on Speech and Audio Processing, Vol. 7, pp. 391–399, July 1999. 5,
135

[Shik 10] J. Shikida, S. Suyama, H. Suzuki, and K. Fukawa. Iterative Re-

ceiver Employing Multiuser Detection and Channel Estimation for
MIMO-OFDM IDMA. Proceedings of the IEEE Vehicular Technol-
ogy Conference (VTC ’10), pp. 1–5, May 2010. 115

[Sibl 06] G. Sibley, G. S. Sukhatme, and L. Matthies. The Iterated Sigma-

Point Kalman Filter with Applications to Long Range Stereo. Pro-
ceedings of the Robotics: Science and Systems Conference, pp. 16–19,
August 2006. 158

[Song 09] J. Song, J. H. Hu, and X. Z. Xiong. EM Estimation Algorithm for

TDR-IDMA Systems in Complex Multipath Channel. Proceedings of
the IEEE International Conference on Communications Technology
and Applications (ICCTA ’09), pp. 671–675, October 2009. 103

[Stal 05] W. Stallings. Wireless Communications and Networks,. Prentice

Hall, 2005. 9

[Stub 02] G. Stüber. Principles of Mobile Communication. Kluwer Academic

Publishers, 2002. 13, 15, 19, 20, 21

[Suya 08] S. Suyama, L. Zhang, H. Suzuki, and K. Fukawa. Performance of It-

erative Multiuser Detection with Channel Estimation for MC-IDMA
and Comparison with Chip-Interleaved MC-CDMA. Proceedings of
the IEEE Global Communications Conference (GLOBECOM ’08),
pp. 1–5, December 2008. 5, 103

[Taro 97] V. Tarokh, N. Seshadri, and A. R. Calderbank. Space-Time Codes

For High Data Rate Wireless Communication: Performance Criteria.
Proceedings of the IEEE International Conference on Communica-
tions (ICC ’97), Vol. 1, pp. 299–303, June 1997. 115

174
References References

[Taro 99] V. Tarokh, N. Seshadri, and A. R. Calderbank. Space-Time Block

Codes from Orthogonal Designs. IEEE Transactions on Information
Theory, Vol. 45, No. 5, pp. 1456–1467, July 1999. 115

[Ther 92] C. W. Therrien. Discrete Random Signals and Statistical Signal Pro-
cessing. Prentice-Hall, 1992. 51

[Thia 07] L. B. Thiagarajan, S. Attallah, and Y. C. Liang. Two-Stage Fre-

quency Synchronization for Uplink MC-CDMA System. Proceedings
of the IEEE Wireless Communications and Networking Conference
(WCNC ’07), pp. 2441–2445, March 2007. 103

[VdBe 99] J. J. Van de Beek, P. O. Börjesson, M. L. Boucheret, D. Landström,

J. M. Arenas, O. Ödling, C. Östberg, M. Wahlqvist, and S. K. Wil-
son. Time and Frequency Synchronization Scheme for Multiuser
OFDM. IEEE Journal on Selected Areas Communications, Vol. 17,
No. 11, pp. 1900–1914, November 1999. 48, 49

[VdMe 01] R. van der Merwe and E. A. Wan. The Square-Root Unscented
Kalman Filter for State and Parameter-Estimation. Proceedings of
the IEEE International Conference on Acoustics, Speech, and Signal
Processing (ICASSP ’01), Vol. 6, pp. 3461–3464, May 2001. 152

[VdMe 04a] R. van der Merwe. Sigma-Point Kalman Filters for Probabilistic
Inference in Dynamic State-space Models. PhD thesis, under the
supervision of E. A. Wan, OGI School of Science and Engineering at
Oregon Health and Science University, April 2004. 4, 56, 134

[VdMe 04b] R. van der Merwe, E. A. Wan, and S. J. Julier. Sigma-Point Kalman
Filters for Nonlinear Estimation and Sensor-Fusion: Applications to
Integrated Navigation. Proceedings of the AIAA Guidance, Navi-
gation and Control Conference (GNC ’04), pp. 1–30, August 2004.
153, 154

[Wan 01] E. A. Wan and R. van der Merwe. Chapter 7, The Unscented
Kalman Filter. Adaptive and Learning Systems for Signal Process-
ing, Communications and Control, Wiley, 2001. 56, 151, 153

175
References References

[Wang 04] H. Wang and B. Chen. A Comparison of the Subcarrier Allocation

Schemes for Uplink OFDMA Systems. Proceedings of the IEEE Con-
ference on Information, Sciences, and Systems (CISS ’04), March
2004. 27

[Wang 10] J. H. Wang, C. L. Song, X. T. Yao, and J. B. Chen. Sigma-Point H∞

Filter for Initial Alignment in Marine Strapdown Inertial Navigation
System. Proceedings of the IEEE International Conference on Signal
Processing Systems (ICSPS ’10), Vol. 1, No. 6, pp. 580–584, July
2010. 158

[Wein 71] S. B. Weinstein and P. M. Ebert. Data Transmission by Fre-

quency Division Multiplexing Using the Discrete Fourier Trans-
form. IEEE Transactions on Communications, Vol. COM-19, No. 10,
pp. 628–634, October 1971. 23, 30

[Wu 03] K. Y. Wu, W. K. Leung, and L. Ping. A Simple Approach to

Near-Optimal Multiple Transmit Antenna Space-Time Codes. Pro-
ceedings of the IEEE International Conference on Communications
(ICC ’03), Vol. 4, pp. 2603–2607, May 2003. 115

[Yang 08] S. S. Yang, W. K. Xu, L. Wang, and Q. F. Wei. Performance of

STBC-IDMA System over Quasi-static Rayleigh Fading Channel.
Proceedings of the IEEE International Conference on Communica-
tions, Circuits and Systems(ICCCAS ’08), pp. 44–47, May 2008. 115

[Yang 10] P. Yang, G. Dumont, and J. M. Ansermino. A Cusum-Based Mul-

tilevel Alerting Method for Physiological Monitoring. IEEE Trans-
actions on Information Technology in Biomedicine, Vol. 14, No. 4,
pp. 1046–1052, July 2010. 89

[Zame 81] G. Zames. Feedback and Optimal Sensitivity: Model Reference

Transformations, Multiplicative Seminorms, and Approximate In-
verses. IEEE Transactions on Automatic Control, Vol. 26, No. 2,
pp. 301–320, April 1981. 135

176
References References

[Zhao 06] P. K. Zhao, L. L. Kuang, and J. H. Lu. Carrier Frequency Offset Es-
timation Using Extended Kalman Filter in Uplink OFDMA Systems.
Proceedings of the IEEE International Conference on Communica-
tions (ICC ’06), Vol. 6, pp. 2870–2874, June 2006. 2, 47, 49, 72

[Zhou 07] X. Y. Zhou, Z. N. Shi, and M. C. Reed. Iterative Channel Estima-

tion for IDMA Systems in Time-Varying Channels. Proceedings of
the IEEE Global Communications Conference (GLOBECOM ’07),
pp. 4020–4024, November 2007. 103

177