Channel Estimation for OFDM in Fast Fading Channels

by

Ping Wan
B.Eng., Zhengzhou Institute of Technology, 1990
M.Sc., Chinese Aeronautical Establishment, 1993

A Dissertation Submitted in Partial Fulfillment of the

Requirements for the Degree of

DOCTOR OF PHILOSOPHY

in the Department of Electrical and Computer Engineering

c Ping Wan, 2011

°
University of Victoria

All rights reserved. This dissertation may not be reproduced in whole or in part, by
photocopying
or other means, without the permission of the author.
 Library and Archives Bibliothèque et
Canada Archives Canada
Published Heritage Direction du
Branch Patrimoine de l'édition

395 Wellington Street 395, rue Wellington

Ottawa ON K1A 0N4 Ottawa ON K1A 0N4
Canada Canada
Your file Votre référence
ISBN: 978-0-494-82438-2

Our file Notre référence

ISBN: 978-0-494-82438-2

NOTICE: AVIS:
The author has granted a non- L'auteur a accordé une licence non exclusive
exclusive license allowing Library and permettant à la Bibliothèque et Archives
Archives Canada to reproduce, Canada de reproduire, publier, archiver,
publish, archive, preserve, conserve, sauvegarder, conserver, transmettre au public
communicate to the public by par télécommunication ou par l'Internet, prêter,
telecommunication or on the Internet, distribuer et vendre des thèses partout dans le
loan, distrbute and sell theses monde, à des fins commerciales ou autres, sur
worldwide, for commercial or non- support microforme, papier, électronique et/ou
commercial purposes, in microform, autres formats.
paper, electronic and/or any other
formats.

The author retains copyright L'auteur conserve la propriété du droit d'auteur

ownership and moral rights in this et des droits moraux qui protege cette thèse. Ni
thesis. Neither the thesis nor la thèse ni des extraits substantiels de celle-ci
substantial extracts from it may be ne doivent être imprimés ou autrement
printed or otherwise reproduced reproduits sans son autorisation.
without the author's permission.

In compliance with the Canadian Conformément à la loi canadienne sur la

Privacy Act some supporting forms protection de la vie privée, quelques
may have been removed from this formulaires secondaires ont été enlevés de
thesis. cette thèse.

While these forms may be included Bien que ces formulaires aient inclus dans
in the document page count, their la pagination, il n'y aura aucun contenu
removal does not represent any loss manquant.
of content from the thesis.
 ii

Channel Estimation for OFDM in Fast Fading Channels

by

Ping Wan
B.Eng., Zhengzhou Institute of Technology, 1990
M.Sc., Chinese Aeronautical Establishment, 1993

Supervisory Committee

Dr. Michael McGuire, Co-Supervisor

(Department of Electrical and Computer Engineering)

Dr. Xiaodai Dong, Co-Supervisor

(Department of Electrical and Computer Engineering)

Dr. Aaron Gulliver, Departmental Member

(Department of Electrical and Computer Engineering)

Dr. Mihai Sima, Departmental Member

(Department of Electrical and Computer Engineering)

Dr. Kui Wu, Outside Member

(Department of Computer Science)
 iii

Supervisory Committee

Dr. Michael McGuire, Co-Supervisor

(Department of Electrical and Computer Engineering)

Dr. Xiaodai Dong, Co-Supervisor

(Department of Electrical and Computer Engineering)

Dr. Aaron Gulliver, Departmental Member

(Department of Electrical and Computer Engineering)

Dr. Mihai Sima, Departmental Member

(Department of Electrical and Computer Engineering)

Dr. Kui Wu, Outside Member

(Department of Computer Science)

ABSTRACT

The increasing demand for high data rate transmission over broadband radio chan-
nels has imposed significant challenges in wireless communications. Accurate channel
estimation has a major impact on the whole system performance. Specifically, reliable
estimate of the channel state information (CSI) is more challenging for orthogonal
frequency division multiplexing (OFDM) systems in doubly selective fading channels
than for the slower fading channels over which OFDM has been deployed traditionally.
With the help of a basis expansion model (BEM), a novel multivariate autoregressive
 iv

(AR) process is developed to model the time evolution of the fast fading channel.
Relying on pilot symbol aided modulation (PSAM), a novel Kalman smoothing algo-
rithm based on a second-order dynamic model is exploited, where the mean square
error (MSE) of the channel estimator is near to that of the optimal Wiener filter.
To further improve the performance of channel estimation, a novel low-complexity
iterative joint channel estimation and symbol detection procedure is developed for
fast fading channels with a small number of pilots and low pilot power to achieve the
bit error rate (BER) performance close to when the CSI is known perfectly. The new
channel estimation symbol detection technique is robust to variations of the radio
channel from the design values and applicable to multiple modulation and coding
types. By use of the extrinsic information transfer (EXIT) chart, we investigate the
convergence behavior of the new algorithm and analyze the modulation, pilot density,
and error correction code selection for good system performance for a given power
level. The algorithms developed in this thesis improve the performance of the whole
system requiring only low ratios of pilot to data for excellent performance in fast
fading channels.
 v

Contents

Supervisory Committee ii

Abstract iii

Table of Contents v

List of Abbreviations ix

List of Symbols xii

List of Figures xvi

Acknowledgements xviii

Dedication xix

1 Introduction 1
1.1 Background . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.2 History of Channel Estimation . . . . . . . . . . . . . . . . . . . . . . 2
1.3 Technical Challenges of Channel Estimation . . . . . . . . . . . . . . 3
1.4 Contributions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
1.5 Thesis Outline . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7

2 Literature Review of Channel Estimation for OFDM Systems 9

2.1 OFDM . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
 vi

2.2 Channel Estimation Techniques . . . . . . . . . . . . . . . . . . . . . 13

2.2.1 Channel Estimation in Slow Fading Channels . . . . . . . . . 15
2.2.2 Channel Estimation in Fast Fading Channels . . . . . . . . . . 18
2.2.3 Joint Channel Estimation Data Detection . . . . . . . . . . . 22

3 System Model 25
3.1 Time Blocks for Channel Processing . . . . . . . . . . . . . . . . . . . 25
3.1.1 Time Domain and Frequency Domain Model for a Single OFDM
Block . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
3.1.2 Time Domain Descriptions for a Transmission Block . . . . . . 30
3.1.3 Coding Block . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
3.2 Three Measurement Models Introduction . . . . . . . . . . . . . . . . 34
3.2.1 Measurements for Data Given Channel . . . . . . . . . . . . . 34
3.2.2 Measurements for Channel Given Pilots . . . . . . . . . . . . 35
3.2.3 Measurements for Channel Given Data . . . . . . . . . . . . . 37
3.3 Time Domain Evolution of the Channel . . . . . . . . . . . . . . . . . 39

4 Second-Order Kalman Filtering Channel Estimation in Fast Fading 43

4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
4.2 Second-Order Kalman Filtering Algorithm . . . . . . . . . . . . . . . 44
4.3 Wiener Bound . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48
4.4 Simulation Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
4.5 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52

5 Joint Channel Estimation and Data Detection 57

5.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 58
5.2 Iterative receivers for OFDM systems in Fast fading . . . . . . . . . . 62
5.2.1 Iterative Channel Estimation . . . . . . . . . . . . . . . . . . 63
 vii

5.2.2 Symbol Detection . . . . . . . . . . . . . . . . . . . . . . . . . 66

5.3 Simulation Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67
5.3.1 BER Performance . . . . . . . . . . . . . . . . . . . . . . . . 68
5.3.2 Robustness Analysis . . . . . . . . . . . . . . . . . . . . . . . 69
5.4 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77

6 EXIT Chart Analysis 78

6.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 78
6.1.1 Input of Transfer Function . . . . . . . . . . . . . . . . . . . . 79
6.1.2 Output of Transfer Function . . . . . . . . . . . . . . . . . . . 81
6.2 EXIT Chart for detection with channel estimation in Fast Fading chan-
nels . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 82
6.2.1 EXIT Chart of Decoder . . . . . . . . . . . . . . . . . . . . . 82
6.2.2 EXIT Chart of Decision-Directed Channel Estimation with De-
tector . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 83
6.2.3 EXIT Chart of Trajectory . . . . . . . . . . . . . . . . . . . . 86
6.3 Simulation Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . 86
6.4 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 90

7 Conclusions and Future Work 93

7.1 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 93
7.2 Future Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 95

A Basic Time Domain Channel Impulse Response Description 96

B Basis Expansion Model 99

B.1 Complex Exponential Basis Expansion Model . . . . . . . . . . . . . 100
B.2 Discrete Prolate Spheroidal Basis Expansion Model . . . . . . . . . . 102
 viii

B.3 Karhunen-Loève Basis Expansion Model . . . . . . . . . . . . . . . . 103

C Whittle-Wiggins-Robinson Algorithm 105

D Least Squares and Linear Minimum Mean Square Error Estimation 109

E Basic Kalman Filter Equations 111

Bibliography 113
 ix

List of Abbreviations

Acronym Definition
1-D one-dimensional
2-D two-dimensional
APP a posteriori probability
AR autoregressive
AWGN additive white Gaussian noise
BCH Bose-Chaudhuri-Hocquenghem
BCJR Bahl-Cocke-Jelinek-Raviv
BEM basis expansion model
BER bit error rate
BICM bit-interleaved coded modulation
CDMA code division multiple access
CE channel estimation
CE-BEM complex-exponential basis expansion model
CFO carrier frequency offset
CIR channel impulse response
CSI channel state information
CP cyclic prefix
CRB Cramér-Rao Bound
 x

DAB digital audio broadcasting

DFE decision-feedback equalizer
DFT discrete Fourier transform
DPS discrete prolate spheroidal
DSL digital subscriber line
DVB digital video broadcasting
EXIT extrinsic information transfer
FFT fast Fourier transform
HIPERLAN/2 high performance local area network type 2
ICI intercarrier interference
LDPC low-density parity-check
IDFT inverse discrete Fourier transform
IFFT inverse fast Fourier transform
ISI intersymbol interference
KL Karhunen-Loève
LMMSE linear minimum mean square error
LS least squares
LLR log-likelihood ratio
LTE long-term extension
MAP maximum a posteriori probability
MC multicarrier
MIMO multiple-input multiple-output
ML maximum-likelihood
MMSE minimum mean square error
MSE mean square error
OB OFDM block
 xi

OFDM orthogonal frequency division multiplexing

PAN personal area network
PAPR peak-to-average-power ratio
PB pilot block
PDF probability density function
PN pseudo-noise
P/S parallel-to-serial
PSAM pilot symbol aided modulation
PSWF prolate spheroidal wave function
QPSK quadrature phase-shift keying
QAM quadrature amplitude modulation
RS Reed-Solomon
SC single-carrier
SER symbol error rate
SISO soft input soft output
SNR signal-to-noise ratio
S/P serial-to-parallel
SVD singular value decomposition
TB transmission block
UWB ultra-wideband
WLAN wireless local area network
WMAN wireless metropolitan area network
WSSUS wide-sense stationary uncorrelated scattering
WWRA Whittle-Wiggins-Robinson Algorithm
ZP zero padding
 xii

List of Symbols

Acronym Definition
A multivariate AR coefficient matrix related to x[n]
Al multivariate AR coefficient matrix related to xl [n]
Al multivariate AR coefficient matrix related to xl [n]
ck the kth bit
Ck transition matrix between ytk and x
CP transition matrix between ytP and x
d pilot vector in the time domain
dp the pth pilot block in the time domain
D(z) diagonal matrix with vector z on its diagonal
E[x] expected value of the random variable x
E basis function matrix for the transmission block
Ek basis function matrix for the kth OFDM block
Eb /N0 bit-to-noise ratio
En (q) the qth basis function value at sample n
f frequency (hertz)
fc carrier frequency
fd maximum Doppler frequency
F fast Fourier transform matrix
 xiii

FH inverse fast Fourier transform matrix

G[n] measurement matrix
h(n, l) channel impulse response in the discrete time
h(t, τ ) channel impulse response
hl channel gain vector of the lth path
hkl channel gain vector of the lth path for the kth OFDM block
H Hermitian transpose
Hk channel matrix in the frequency domain for the kth OFDM block
Htk channel matrix in the time domain for the kth OFDM block
J0 (· ) the zeroth-order Bessel function of the first kind
K order of multivariate AR process
Kb number of data blocks
Kp number of pilot blocks
L channel order
LA a priori log-likelihood ratio
LD a posteriori log-likelihood ratio
LE extrinsic log-likelihood ratio
M number of samples in a transmission block
MD transition matrix between y D and x
Mk transition matrix between yk and x
N number of subcarriers
Ncp length of cyclic prefix
Ns length of OFDM block
N0 power spectrum density of AWGN
P [n] error covariance matrix of the state vector x[n]
P (x) probability of x
 xiv

Q BEM order
Q covariance matrix of noise
Qlf covariance matrix of noise vector wl
rt (∆t) spaced-time autocorrelation function
R correlation matrix of the measurement noise
R covariance matrix of x
Rx covariance matrix of xl
Rh covariance matrix of hl
sk transmitted signal vector in the frequency domain for the kth OFDM block
SH (f ) Doppler power spectrum of the channel
Sv (f ) power spectrum of the noise
Sl circularly shifting identity matrix IN down with l samples
t time variable
T OFDM symbol duration
T transpose
Tcp guard interval
Ts sampling period
u transmitted signal vector in the time domain
uk transmitted signal vector in the time domain for the kth OFDM block
up the pth data block in the time domain
ũ transmitted signal vector of the transmission block in the time domain
v vehicle speed
vt (n) AWGN in the time domain
vk AWGN vector in the frequency domain for the kth OFDM block
v[n] state noise vector
ṽt AWGN noise vector in the transmission block
 xv

vtD AWGN vector at the data position in the time domain

vtk AWGN vector in the time domain for the kth OFDM block
vtP AWGN vector at the pilot position in the time domain
W bandwidth
w[n] measurement vector
x basis coefficient vector
xl (q) the basis coefficient for path l corresponding to the qth basis function
x[n] state vector
xl basis coefficient vector of the lth path
xl [n] containing xl for blocks n − K + 1 to n
x[n] containing x for blocks n − K + 1 to n
yt (n) received signal in the time domain at sample n
yD received signal vector corresponding to data in the frequency domain
y[n] measurement vector at ‘time’ n
ỹt received signal vector of the transmission block in the time domain
ytk received signal vector of the kth OFDM block in the time domain
ytD received signal vector corresponding to data in the time domain
ytP received signal vector corresponding to pilots in the time domain
d
ytp received signal vector corresponding to up
p
ytp received signal vector corresponding to dp
∆f subcarrier spacing
Φ state transition matrix
σv2 variance of noise
τ propagation delay
τmax multipath delay spread
 xvi

List of Figures

Figure 2.1 A discrete-time baseband OFDM system . . . . . . . . . . . . 12

Figure 2.2 Structure of training and pilots. (a) Training symbols. (b)
Pilot symbols. . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
Figure 2.3 Structure of input and output of the channel in the time domain. 20
Figure 2.4 Structure of input and output in the frequency domain . . . . 21

Figure 3.1 Block structure with κ = 1 . . . . . . . . . . . . . . . . . . . . 27

Figure 3.2 Structure of Hk matrix . . . . . . . . . . . . . . . . . . . . . . 30
Figure 3.3 Discrete-time structure in OFDM systems . . . . . . . . . . . 30
Figure 3.4 Relationship between the transmitted data and the received
signal in the time domain for single transmission block . . . . 32

Figure 4.1 Channel estimation MSE for fd T = 0.05 . . . . . . . . . . . . 51

Figure 4.2 Channel estimation MSE for fd T = 0.1 . . . . . . . . . . . . . 52
Figure 4.3 Channel estimation BER for fd T = 0.05 . . . . . . . . . . . . 53
Figure 4.4 BER resulting from estimated channel state for fd T = 0.1 . . 54
Figure 4.5 Channel MSE for SN R = 10 dB and designed fd T = 0.05 . . 55
Figure 4.6 Channel MSE for fd T = 0.05 and designed SN R = 15 dB . . 56

Figure 5.1 Block diagram of the transmitter . . . . . . . . . . . . . . . . 59

Figure 5.2 Block diagram of the receiver . . . . . . . . . . . . . . . . . . 59
Figure 5.3 Coded OFDM transmitter . . . . . . . . . . . . . . . . . . . . 62
 xvii

Figure 5.4 Coded OFDM system . . . . . . . . . . . . . . . . . . . . . . 62

Figure 5.5 MSE resulting from estimated CSI for fd T = 0.1 with QPSK . 70
Figure 5.6 BER resulting from estimated CSI for fd T = 0.1 with QPSK . 71
Figure 5.7 BER resulting from estimated CSI for fd T = 0.15 with QPSK 72
Figure 5.8 BER resulting from estimated CSI for fd T = 0.1 with 16-QAM 73
Figure 5.9 BER resulting from estimated CSI with Kp = 10 for fd T = 0.1
with 64-QAM . . . . . . . . . . . . . . . . . . . . . . . . . . . 74
Figure 5.10 BER resulting for designed fd T = 0.1 with QPSK . . . . . . . 75
Figure 5.11 BER resulting for fd T = 0.1 and true L = 3 with QPSK . . . 76

Figure 6.1 Model for the EXIT chart analysis . . . . . . . . . . . . . . . 79

Figure 6.2 Model for the EXIT chart of detector . . . . . . . . . . . . . . 83
Figure 6.3 EXIT chart . . . . . . . . . . . . . . . . . . . . . . . . . . . . 84
Figure 6.4 EXIT chart using Kp = 10 and Kt = 10 for fd T = 0.1 at
Eb /No = 7dB . . . . . . . . . . . . . . . . . . . . . . . . . . . 88
Figure 6.5 EXIT chart using Kp = 5 and Kt = 10 for fd T = 0.1 at
Eb /N0 = 7dB . . . . . . . . . . . . . . . . . . . . . . . . . . . 89
Figure 6.6 EXIT chart using Kp = 5 and Kt = 100 for fd = 0.1 at Eb /N0 =
7dB . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 90
Figure 6.7 EXIT chart using Kp = 5 and Kt = 10 for fd T = 0.1 at
Eb /N0 = 8dB . . . . . . . . . . . . . . . . . . . . . . . . . . . 91
Figure 6.8 EXIT chart using Kp = 10 and Kt = 10 for fd T = 0.1 at
Eb /No = 7dB . . . . . . . . . . . . . . . . . . . . . . . . . . . 92
 xviii

ACKNOWLEDGEMENTS

First and foremost I would like to thank my supervisor, Dr. Michael McGuire,
for his valuable guidance, continuous encouragement and insightful technical advice
throughout my study. Dr. McGuire and my co-supervisor Dr. Dong have offered
great assistance to help me navigate through the bumps in the road. This thesis
could not have been completed without their support and help.
I would also like to thank Dr. Aaron Gulliver, Dr. Mihai Sima and Dr. Kui Wu
for the valuable suggestions on revising my thesis.
Thanks to many of my colleagues and friends at University of Victoria for being
so nice and helpful, which makes my stay a great pleasure. While I cannot write
about them all individually, each of them has been important to me at various stages
throughout my time at University of Victoria.
Special thanks to Steve, Kevin, Erik, Vicky, Moneca and Mary-Anne for their
patience and constant help.
Lastly, I would like to thank my family, for their boundless love, understanding,
and constant support in all that I do. This thesis would certainly not have existed
without them.
 xix

DEDICATION

To my family
Chapter 1

Introduction

1.1 Background

The increasing demand for high data rate transmission over broadband radio chan-
nels has imposed significant challenges in wireless communications. High data rate
transmission and high mobility of transmitters and/or receivers result in frequency-
selective and time-selective, i.e., doubly selective, fading channels for future mobile
broadband wireless systems. Mitigating such doubly selective fading effects is critical
for efficient data transmission. Moreover, perfect channel state information (CSI) is
not available at the receiver. Thus accurate estimate of the CSI has a major impact
on the whole system performance [1]. This motivates an extensive channel estimation
study of doubly selective fading channels for future wireless communication systems.
To extend the applications of orthogonal frequency division multiplexing (OFDM)
in the future mobile broadband wireless communication systems such as mobile Mi-
MAX and long-term extension (LTE), channel estimation techniques for OFDM sys-
tems in doubly selective channels are the topic in this thesis. For high data rate
transmission systems, employing OFDM converts a wide bandwidth channel into sev-
 2

eral narrow band subchannels (subcarriers). Due to its high bandwidth efficiency, its
simple implementation and its robustness over frequency-selective channels, OFDM
has been widely applied in wireless communication systems.
For conventional coherent detection, accurate CSI is needed for the receiver pro-
cessing [2]. Although channel estimation can be avoided by using differential modula-
tion techniques, these techniques will fail catastrophically in the fast fading channel,
where the channel impulse response (CIR) varies significantly within the symbol du-
ration [3]. In fact, differential modulation techniques assume that the channel is
stationary over the period of two OFDM symbols which is not true for the fast fad-
ing channels of greatest interest in this thesis [4]. Thus channel estimation is an
integral part of the receiver for fast fading channels. In such a case, the receiver
needs to perform channel estimation for each OFDM symbol. Moreover, in fast fad-
ing channels, the impulse response of the channel for each propagation path changes
from the beginning to the end of each OFDM symbol. The orthogonality among the
subcarriers is destroyed and intercarrier interference (ICI) is created, which, if left
uncompensated, can cause high bit error rates (BERs). Generally, the compensation
for the ICI due to the fast fading channel is based on more complex equalizers such
as minimum mean-square error (MMSE) equalizers [5, 6], which need not only the
individual subcarrier frequency responses but also the interference among subcarriers
in each OFDM symbol. Hence, channel estimation is more challenging for OFDM
systems in fast fading channels than in slow fading systems.

1.2 History of Channel Estimation

Of different classes channel estimation techniques that have been developed in the
literature, one class of techniques is based on pilot symbols which are known a priori
 3

to the receiver. In this case, a standard approach is pilot symbol aided modulation
(PSAM) [7] or so-called pilot assisted transmission [8], where known pilot signals mul-
tiplexed with information symbols are sent through the channel at regular intervals so
the radio receiver can make direct measurements of the channel variations created by
the propagation environment and terminal mobility. The CSI corresponding to pilot
symbols is estimated first and the CSI corresponding to data symbols is then obtained
by interpolation. In OFDM systems, the existing standards add pilot symbols in the
frequency domain.
Another class of channel estimation techniques is blind channel estimation which
depends only on the received symbols without inserting pilots [9, 10, 11]. In this case,
the CSI is obtained from the received signal by using high-order statistical methods
which require a large amount of data and high computational complexity [2].

1.3 Technical Challenges of Channel Estimation

As the case of the fast fading channel, the channel gain at each time sample of one
OFDM symbol block is needed for data detection, and as a consequence, the number
of unknown channel parameters is larger than the number of measurements if standard
channel modelling is employed. In such a case, an underdetermined system occurs
as the number of parameters exceeds the number of available measurements created
by the pilot symbols [12]. To reduce the number of parameters to be estimated, an
alternative approach is modelling the channel using a basis expansion model (BEM)
over a time period, called a transmission block, where the channel state is expressed
as a superposition of known basis functions weighted by unknown basis coefficients.
Thus, the channel estimation problem is converted to the estimation of a limited
number of basis coefficients.
 4

Due to Gibbs phenomenon and the bandlimited nature of the fading process,
the previous BEM-based channel modelling was not accurate when a short OFDM
period was used as the BEM period. To solve this problem, we choose several OFDM
blocks as the BEM period or transmission block length. With the use of a short
OFDM symbol block, a low cost data detection is available. Therefore, higher order
modulation can be used for fast fading channels without an exponential increase in
data detection cost.
Due to a lack of modelling of the time evolution of the fading channel, the previ-
ous BEM-based channel estimation techniques did not exploit the information from
OFDM blocks other than one currently being processed, and hence, such estimation
methods did not achieve optimal mean square error (MSE) performance. To model
the time evolution over a transmission block consisting of multiple OFDM symbols,
a multivariate autoregressive (AR) Gauss-Markov model is exploited to characterize
the dependency of the channel state between OFDM blocks for fast fading channels
for the purposes of improving the channel estimation ability.
To simplify the channel estimation algorithm for coded OFDM systems, time
blocks of three different durations are used at the receiver proposed in this thesis:
the OFDM block is used as the basis for detection; the transmission block is used for
channel estimation; and the interleaving block, which contains several transmission
blocks, is used for decoding. To develop an iterative scheme at the receiver, we derive
different measurement models, which include measurement model for data in terms
of channel for detection, measurement model for channel in terms of data symbols for
channel estimation, and measurement model for channel in terms of pilot symbols for
channel estimation.
In the case of pilot-aided channel estimation, the time evolution of basis coeffi-
cients is described as a second-order multivariate AR model, which can character-
 5

ize the channel more accurately than a first-order AR model [13]. Subsequently, a
Kalman filter is developed to track the basis coefficients from the pilot measurements.
A Kalman smoother, which incorporates the following measurements, gives a more
refined estimate than the previously presented channel estimation methods. It will
be shown that the MSE of the channel estimation is not significantly reduced when
a higher order (≥ 3) Kalman filter is used. To reduce the complexity of channel
estimation in fast fading channels, new pilot aided channel estimation in the presence
of ICI with low complexity is proposed in this thesis.
As the fading rate increases, the severity of the ICI also increases, and thus,
more channel state values in each OFDM symbol are needed. Under this condition,
more pilot symbols are required to obtain sufficiently accurate channel estimates for
reliable data detection. Even with the use of BEMs, the number of channel parameters
required to model the channel increases which, in turn, increases the needed pilot to
data ratio to achieve acceptable system performance. This reduces the overall data
rate during fast fading to unacceptable levels for many applications. The combination
of high fading rate and low pilot to data ratio necessitates the channel estimator in
OFDM systems and a good tradeoff is needed between the performance of the channel
estimation and the pilot to data ratio.
To further improve the performance of channel estimation with low pilot to data
ratio, a novel low-complexity iterative joint channel estimation and symbol detection
procedure is proposed. The time evolution of the basis coefficients between each
transmission block is modelled as a first-order multivariate AR process. An initial
channel estimate is obtained based on pilot-aided channel estimation and used for an
initial data detection. The detected data is then used to re-estimate the channel by
a decision-directed method and the data is detected again. The detected data is used
as ‘virtual’ pilot signal which is dense in the time domain. The error of the channel
 6

estimation is reduced with more ‘virtual’ pilots being available. As a result, the
BER performance of the proposed joint channel estimation data detection technique
is much better than that of the conventional PSAM approach. It will be shown that
the BER performance is close to that when ideal CSI is available.

1.4 Contributions

The main contributions of the thesis for OFDM system modelling :

• The introduction of a novel multivariate AR Gauss-Markov model for basis

coefficients over transmission blocks to model the time evolution of the fast
fading channel.

• The introduction of time blocks of three different durations for channel process-
ing in coded OFDM systems to both reduce channel estimation and detection
computation cost and error levels.

• The introduction of different measurement models for an iterative receiver in

fast fading channels.

The main contributions of the thesis to the subject of channel estimation based
on pilots are :

• The introduction of a novel Kalman smoothing algorithm based on a second-

order dynamic model with a MSE near that of the optimal Wiener filter and
a computational cost on the same order as the previously proposed algorithms
for channel estimation and improves the BER performance of wireless commu-
nications systems.

• The demonstration of the robustness of the new technique when the Doppler
 7

frequency or signal-to-noise ratio (SNR) is varied from the value used to design
the channel estimation filter.

The main contributions of the thesis in the areas of joint channel estimation and
data detection are :

• The introduction of a novel joint channel estimation/symbol detection algorithm

for fast fading channels requiring only a small number of pilots and low pilot
power to achieve performance close to when the CSI is known perfectly.

• The demonstration that this new channel estimation symbol detection technique
is robust to variations of the radio channel from the design values.

• The demonstration that the new approach is efficient when applied to different
modulation and coding types.

• The demonstration of how the convergence behavior of the new algorithm is

analyzed by the extrinsic information transfer (EXIT) chart technique.

• The introduction of the EXIT chart is used to select the modulation, pilot
density, and error correction code for good system performance for a given
power level.

1.5 Thesis Outline

The structure of the thesis is described as follows. In Chapter 2, we present a survey of

the existing techniques for estimating the CSI with pilot symbols. In Chapter 3, first,
we introduce three time blocks at the receiver; then, we derive different measurement
models for the proposed iterative receiver; finally, we develop a multivariate AR
model to characterize the time domain evolution of the channel. Assuming access
 8

to pilot symbols, we propose a second-order Kalman filter algorithm using BEMs of

the channel gains to accurately estimate fast fading channels in Chapter 4. A joint
channel estimation and data detection approach is discussed in Chapter 5, which
reduces the required number of pilot symbols for acceptable error performance in fast
fading channels to the levels needed by previous channel estimation algorithms with
slow fading channels. Using EXIT chart to investigate how the proposed iterative
receiver works well is discussed in Chapter 6. Conclusions and future work will be
present in Chapter 7.
 9

Chapter 2

Literature Review of Channel

Estimation for OFDM Systems

This chapter presents a literature survey of channel estimation methods for orthogonal
frequency division multiplexing (OFDM) system in fast fading channels. As described
in Chapter 1, we will focus on pilot-aided channel estimation schemes. In general, two
aspects affect the performance of the pilot-aided channel estimation techniques: the
design of pilot symbols and the estimation scheme. We begin with the description
of OFDM system. Different pilot-aided channel estimation schemes are treated in
Section 2.2.

2.1 OFDM

The use of OFDM techniques can be traced back to the late 1950’s and early 1960’s for
military high frequency radio systems. With the availability of simple and cheap im-
plementations of the discrete Fourier transform (DFT) and the inverse DFT (IDFT),
the DFT can generate data signal in parallel form and OFDM became popular [14].
In particular, coded OFDM has been adopted by standards and major manufacturers
 10

for a wide range of applications. Now, it has been used in digital audio broadcasting
(DAB) and digital video broadcasting (DVB) systems in Europe [15], digital sub-
scriber line (DSL) standards, and wireless local area network (WLAN) standards
such as IEEE Std.802.11a/g (WiFi) [16] or high performance local area network type
2 (HIPERLAN/2) in Europe [17], and wireless metropolitan area network (WMAN)
standards such as IEEE Std.802.16 (WiMAX) [18], also in ultra-wideband (UWB)
personal area network (PAN) (IEEE 802.15.3a). OFDM is also considered in IEEE
Std.802.11n in multiple-input multiple-output (MIMO) systems. For future mobile
broadband wireless communications, mobile WiMAX and long-term extension (LTE)
use OFDM-based modulation too.
A well-known approach to efficiently utilize available channel bandwidth is the
multicarrier (MC) transmission scheme first proposed by Chang [19]. OFDM is a
type of MC modulation [20], which converts a wide bandwidth W into N narrow-
band subchannels (subcarriers) over which data is transmitted in parallel [21, 22].
To obtain high spectral efficiency, these subchannels are set to be overlapping and
orthogonal under ideal propagation conditions. The subcarrier spacing, denoted as
∆f = W/N , provides the minimum frequency separation required to maintain orthog-
onality between subcarriers. Consequently, the following equation should be satisfied
R T 2π
over the OFDM symbol duration T , i.e., 0 ej T (m−n)t dt = 0, for different subcarriers
m and n. In other words, the OFDM symbol duration is T = N Ts = 1/∆f , where
Ts is the sampling period.
One problem created by multipath propagation is inter-symbol interference (ISI).
To manage the ISI, OFDM symbols use either a cyclic prefix (CP) or a zero padding
(ZP) guard interval between OFDM blocks which is longer than the delay spread of
the channel which is the difference between the maximum and minimum propagation
delay of the radio channel. A CP is the repetition of the last Ncp samples of the
 11

transmitted signal which are inserted at the beginning of each OFDM symbol. With
the used of CP, the transmitted signal is cyclically extended in the guard interval
Tcp = Ncp Ts . At the receiver, after removing the first Ncp signals corrupted by ISI,
the ISI is completely discarded. However, due to adding the CP at the transmitter,
Tcp
the proportional loss of the useful transmission energy is T +Tcp
. Instead of using the
CP, the ZP with Ncp zeros can be inserted at the end of each OFDM symbol without
energy loss. Unfortunately, this method increases the receiver complexity compared
to the traditional CP-based OFDM (CP-OFDM) [23]. Therefore, in this and the
following chapters we consider CP-OFDM systems.
The major advantage of OFDM lies in processing frequency-selective channels as
multiple flat-fading sub-channels. If the channel is time invariant over the period of
an OFDM symbol block, a condition known as slow fading, the orthogonality prop-
erty is maintained between the subcarriers. In such a case, channel estimation or
data detection is simple since each subcarrier is equalized with a single-tap equal-
izer. However, when the channel is time-varying over one OFDM symbol period, the
orthogonality among subcarriers is destroyed, resulting in ICI, which degrades the
bit error rate (BER) performance compared to the slow fading channels if it is not
properly compensated for. The ICI may occur due to the presence of the fast fading
channel or the presence of a carrier frequency offset (CFO) between the transmitter
and receiver caused by imperfect synchronization. CFO can be estimated by using
a maximum likelihood (ML) estimation algorithm [24, 25, 26]. The potential perfor-
mance degradation of OFDM caused by fading channels is a function of the fading
rate, with faster fading channels requiring more significant mitigation methods to
achieve the same error performance as slow fading channels. Furthermore, in the
presence of ICI due to fast fading, the channel estimation is more challenging since
both the individual subcarrier and the interference created by each subcarrier to its
 12

neighbouring subcarriers need to be estimated. Therefore, more sophisticated detec-

tion procedures or channel estimation schemes are required [27, 4, 28]. To further
advance the applications of OFDM in future standards, robust channel estimation
schemes for fast fading channels are discussed in this thesis.
s0
Symbol .. .. CP
S/P . IFFT . P/S
Mapper Insertion
sN −1
y0
Detection .. .. CP
P/S . FFT . S/P
&CE Removal
yN −1

Figure 2.1: A discrete-time baseband OFDM system

The block diagram of discrete-time baseband OFDM system is depicted in Fig-

ure 2.1. In discrete-time OFDM systems, the input data is first modulated and then
passed through a serial-to-parallel (S/P) converter, whose output in the frequency
domain is expressed in vector form as s = [s0 , ..., sN −1 ]T , where N is the number of
subcarriers. Using the inverse fast Fourier transform (IFFT) algorithm, a discrete-
time baseband OFDM symbol is converted into time-domain samples as

N −1
1 X nk
u(n) = √ sk ej2π N , 0≤n≤N − 1 (2.1)
N k=0

After inserting CP, the resulting transmitted signal becomes u = {u(N −Ncp ), ..., u(N −
2), u(N − 1), u(0), u(1), ..., u(N − 1)}. This signal is then serially transmitted through
a multipath radio propagation channel which is subject to additive white Gaussian
noise (AWGN) v(t) with variance σv2 = N0 /2, where N0 is power spectral density.
At the receiver, after removing the CP and processing by a parallel-to-serial (P/S)
converter, the received signal in the time domain is then converted back to the fre-
quency domain, which is implemented by using the fast Fourier transform (FFT)
 13

algorithm. Due to the use of CP, the interblock interference between contiguous
OFDM blocks in the frequency domain is eliminated so each OFDM block can be
processed independently, provided that the length of the CP is equal to or larger
than the delay spread of the channel.
The main drawbacks of OFDM systems are high peak-to-average-power ratio
(PAPR) and high sensitivity to CFO. Moreover, OFDM does not obtain frequency
diversity. If a deep fade occurs close to the frequency of a subcarrier, reliable data
detection carried by these faded subcarriers becomes difficult [29, 30]. To solve this
problem, an alternative method is to employ error-control codes. Therefore, the
diversity loss can be circumvented by incorporating error-control coding in conjunc-
tion with interleaving. The typical codes are block codes (e.g., Reed-Solomon (RS)
or Bose-Chaudhuri-Hocquenghem (BCH)), convolutional codes, trellis codes, turbo
codes, and low-density parity-check (LDPC) codes. Another one is combining multi-
carrier coding and code division multiple access (CDMA) techniques, where MC-
CDMA splits a wide band signal into narrowband signals and also exploits multipath
diversity [31]. The new channel estimation techniques which will be discussed later
in this thesis can be extended to MC-CDMA.

2.2 Channel Estimation Techniques

When the channel is unknown a priori to the receiver, pilot symbol aided modulation
(PSAM), where known pilot signals are periodically sent during the transmission, can
simplify the channel estimation. In general, the performance of channel estimation
depends on the number, the location, and the power of pilot symbols inserted into
OFDM blocks. Consider a fading multipath channel with the multipath delay spread
τmax and the maximum Doppler frequency fd . To recover the channel state informa-
 14

tion (CSI), the spaces between pilot symbols in the time and frequency domain must
satisfy two-dimensional (2-D) sampling theorem, that is,

fd T dt ≤ 1/2 (2.2)

and
τmax ∆f df ≤ 1 (2.3)

where T is the OFDM block duration, ∆f is the subcarrier spacing; dt and df are the
numbers of samples between pilot symbols in the time domain and frequency domain,
respectively [32]. Within the OFDM symbol duration, the number of pilot symbols in
the frequency domain is related to the delay spread; on the other hand, the number
of pilot symbols in the time domain is related to the normalized Doppler frequency
fd T . Based on 2-D arrangement of pilot symbols, 2-D channel estimators are too
complex in practice [33]. Therefore, channel estimation is exploited in one-dimension
(1-D) for OFDM systems in general.
In practice, for OFDM systems, the channel estimation techniques can be per-
formed by using either frequency domain samples or time domain samples. Thus,
pilot symbols can be added either in the frequency domain where the pilot symbols
are frequency multiplexed with the data, or in the time domain where the pilot sym-
bols are time multiplexed with the data. Since single-carrier (SC) modulation, in
which data symbols are transmitted in serial fashion, is a form of generalized MC
transmission [34, 35], methods for SC channel estimation will also be examined for
their applicability to OFDM channel estimation. In the following subsections, we
discuss different channel estimation techniques employing pilot schemes in slow and
fast fading channels, respectively.
 15

2.2.1 Channel Estimation in Slow Fading Channels

When the channel slowly changes over a number of OFDM symbol blocks, channel
estimation can be based on pilot symbols, which are inserted into all subcarriers of
OFDM symbol blocks within a specific period, as shown in Figure 2.2 (a), where such
pilot symbols are usually called training symbols. Then a batch of OFDM symbols
follows the training symbols. For the channel estimation based on training symbols,
the CSI corresponding to training symbols are first estimated, the CSI corresponding
to the subsequent data symbols can be tracked and further improved by decision
directed channel estimation [28]. If the channel varies slowly over OFDM blocks,
the estimated CSI based on previous training symbols are generally reliable so such
estimated channel state may be used in data detection. As the channel varies fast
over time, the training symbols must be sent more frequently to get reliable channel
estimates, and hence, the overall system efficiency is reduced [4]. In such a case,
channel estimation can be based on pilot symbols, which are periodically inserted
into different subcarriers for each OFDM symbol block, as shown in Figure 2.2 (b),
and will be discussed later.
In addition, to improve bandwidth efficiency, the superimposed pilot scheme was
proposed for flat-selective fading channels [36], where a pseudo-noise (PN) sequence is
synchronously added to the information symbols at the transmitter prior to modula-
tion transmission. Based on the first-order statistics method, this superimposed pilot
approach for channel estimation was discussed for frequency selective fading channels
[37, 38, 39], and for doubly selective fading channels [40, 41]. Although the use of
superimposed pilots can improve the bandwidth efficiency, the performance based on
superimposed pilot aided channel estimation is worse than that of traditional PSAM
[41, 42]. Therefore, we focus our attention on channel estimation based on traditional
PSAM in this thesis.
 16

F req.
x x h h ··· h F req.
6 x x h h x h h x ··· h
··· h 6 h h h h h
···
.. .. .. .. ..
. . . . ··· . .. .. .. .. ..
. . . . ··· .
x x h h ··· h
x h h x ··· h
|{z} | {z } - T ime
training data xpilots hdata
- T ime
(b)
(a)

Figure 2.2: Structure of training and pilots. (a) Training symbols. (b) Pilot symbols.

As in the case of the channel varying from one OFDM block to another, pilot
symbols are often inserted in every OFDM symbol in general. The pilot spacing
in the frequency domain is the main criteria for the pilot placement. When the
channel is only frequency selective, the channel gain is different for different frequency
components [43]. In order to capture the variation of the channel in the frequency
domain, the pilot spacing in the frequency domain should satisfy the sampling criteria
(2.3). For slow fading channels, different pilot designs have been discussed in [44, 45,
46, 47, 48, 49, 50], where optimally the pilots are equally spaced in the frequency
domain. Consider a fading multipath channel with L + 1 paths, where L is called
as channel order and τmax = LTs . To minimize the mean-square error (MSE) of the
channel estimation, the L + 1 pilots are equally spaced in each OFDM symbol in the
frequency domain in order to estimate the channel condition. If N/(L + 1) is not an
integer, the design of optimal pilot symbols is discussed in [48], where the L + 1 pilots
are uniformly distributed across subcarriers to minimize symbol error rate (SER).
The existing standards such as WiMAX choose frequency multiplexed pilots scheme
since these schemes operate only over slow-fading channels. However, more recent
innovations allow OFDM to operate over fast-fading channels if sufficiently accurate
channel state information is available. Therefore, the majority of this thesis is to
dedicate channel estimation in OFDM for fast fading channels.
 17

From the description given above, we now turn our attention to the channel esti-
mation. Several types of estimation techniques are exploited by using least squares
(LS), ML, minimum mean square error (MMSE) or linear minimum mean square error
(LMMSE) methods. Under the condition that the channel impulse response (CIR)
is deterministic and the channel statistics is unknown at the receiver, if the additive
noise is unknown at the receiver, LS approach can be used for pilot-aided channel
estimation; if the noise is AWGN, ML channel estimation will be optimal and achieves
the Cramér-Rao Bound (CRB), and furthermore, LS estimation is equivalent to ML
estimation [51]. Whereas in the case that the CIR is random and channel statistics
and signal-to-noise ratio (SNR) are known a priori, MMSE or LMMSE channel esti-
mation is exploited to find minimize the MSE of the channel for better performance
with higher computational complexity compared with LS or ML method. Moreover,
if the measurement model is linear without Gaussian assumption, the LMMSE chan-
nel estimation can be used. On the other hand, if the measurement model is linear
with Gaussian assumption, the LMMSE technique is same as the MMSE method. In
practice, LS method is used to get initial channel estimates at the pilot symbols, and
further improvement is based on MMSE or LMMSE method [4].
For slow fading channels as discussed above, depending on the pilot arrangement,
estimation schemes are performed in the frequency domain using either training sym-
bols or pilot symbols. When the channel is estimated based on training symbols or
detected symbols, the CSI may be estimated by using LS algorithm or LMMSE al-
gorithm [52, 53, 54]. When the channel is estimated based on pilot symbols which
are usually inserted in the frequency domain, the CSI of the frequency domain at the
known pilot subcarriers is first estimated by using LS algorithm [55] or ML algorithm
[56], then interpolation is performed between these channel estimates to get the CSI
at the data subcarriers via different methods such as the piecewise constant and linear
 18

interpolation methods [22, 57] or high-order polynomial interpolation [58, 59]. When
the information of SNR and channel statistics is obtained, MMSE(LMMSE) algo-
rithm is exploited for channel estimation with pilot symbols [60, 61]. Compared with
the LS algorithm, since the LMMSE algorithm exploiting channel statistics, LMMSE
estimation is more accurate than LS estimation; however, LMMSE estimation has
higher computational complexity.
Since data detection ignores ICI for slow fading channels, simple channel estima-
tion techniques are required for the receiver. However, in order to compensate for
the ICI in fast fading channels, more sophisticated channel estimation algorithms are
required for data detection techniques, which is discussed below.

2.2.2 Channel Estimation in Fast Fading Channels

When the channel varies significantly over one OFDM symbol block, the orthogo-
nality among the OFDM subcarriers is lost and the ICI is created. The severity of
ICI depends on the normalized Doppler frequency fd T [5]. In the presence of ICI,
the amount of channel states that need to be estimated for reliable data detection
increases. Not only the individual subcarrier frequency responses but also the in-
terference among subcarriers in each OFDM symbol are to be estimated. In such a
case, an underdetermined system occurs if standard channel modelling is employed
since the number of unknowns are more than the number of measurements (pilot
symbols). In order to reduce the number of unknown channel parameters, simpli-
fication approaches are exploited for channel estimation. One approach is that the
channel is approximated to a piece-wise linear model over one or two adjacent OFDM
blocks [62, 63]. However, this modelling approach degrades the performance of the
channel estimation at high normalized Doppler frequencies such as 10% [64]. Another
approach is to model the channel by a basis expansion model (BEM), where the sam-
 19

ples of the channel state are characterized as a linear combination of a finite number
of known basis functions weighted by unknown basis coefficients, which is discussed in
detail in Appendix B, as well as Appendix A for channel impulse response description.
Investigations into channel estimation for SC and OFDM modulation in [1, 65, 66],
have shown that pilot symbols placed in the time domain maximize the lower bound
on channel capacity and minimize the channel estimation mean square error for multi-
carrier communication systems such as OFDM. In order to mitigate interference be-
tween the pilot block and data block, which is introduced due to multipath propaga-
tion of the channel, each pilot block has the same form as [01×L 1 01×L ], where the
size of 01×L is 1 by L. In the case of the fixed data power and pilot power, based on
lower bound on the average channel capacity with the LMMSE channel estimation,
the optimal pilot symbols are equi-powered and equi-spaced in the time domain [1].
In addition, in order to minimize the MSE of the channel estimator, the optimal
number of pilot (data) blocks should be equal to the number of basis coefficients for
a given transmission block [1]. From a performance viewpoint, to minimize the total
MSE of the estimator which includes the BEM modelling error, the number of pilot
(data) blocks may be larger than the number of basis coefficients [65].
Similar to the time-domain pilot scheme, the pilot block in the form of [01×Q
1 01×Q ] is equally placed between data blocks in the frequency domain, where Q
is the number of basis coefficients, defined as Q ≥ 2dfd N Ts e and d e denotes the
integer ceiling [67, 68]. The affect of ICI caused by each subcarrier is spread into
its neighbouring subcarriers, where the length of ICI caused by each subcarrier is
equal to 2Q. In this case, the pilot symbols are equi-powered and equi-spaced in the
frequency domain, and the optimal number of pilot blocks is equal to L + 1.
Figure 2.3 (a) illustrates the transmitted signal in the time domain, where up
and dp are data block and pilot block for p ∈ [0, Kp − 1], respectively, and Kp is the
 20

u0 d0 u1 d1 uKp −1 dKp −1

data 0L 0L data 0L 0L ··· data 0L 0L

T ime
(a) Input

···

| {z } | {z } | {z }
r0 r1 rKp −1
T ime
(b) Output

Figure 2.3: Structure of input and output of the channel in the time domain.

number of pilot blocks. Figure 2.3 (b) illustrates the received signal in the the time
domain without noise. Due to the multipath spread, the output vector of the channel,
rp , corresponding to data block up can be spilt into two parts. The rectangular
part represents the desired output of the channel. The black triangular part is the
interference overlapping into the adjacent pilot block from the preceding data block,
where the length of the interference block is equal to L. Similarly, the output vector
of the channel corresponding to the pilot block has two parts. From Figure 2.3, we
can see that, as long as the pilot block has the form as described above, there is no
inter-block interference between pilot block and data block, and hence linear channel
estimation techniques can be used in the time domain. Therefore, time multiplexed
pilots scheme can temporally separate pilot blocks from data blocks, and also be used
for the purpose of timing synchronization.
Figure 2.4 (a) illustrates the transmitted signal in the frequency domain, where
sm and bm are data block and pilot block for m ∈ [0, M − 1], respectively, and M
is the number of pilot blocks. Figure 2.4 (b) illustrates the received signal in the
frequency domain without noise. Due to the Doppler spread, the output vector of
p
the channel, rm , corresponding to pilot block bpm can be spilt into three parts, the
rectangular part and the two triangular parts. As shown in Figure 2.4 (b), there
 21

s0 b0 s1 b1 sM−1 bM−1

data 0Q 0Q data 0Q 0Q ··· data 0Q 0Q

F req.
(a) Input

···

| {z } | {z } | {z }
p
r0p r1p rM−1
F req.
(b) Output

Figure 2.4: Structure of input and output in the frequency domain

p
is the inter-block interference (black portions) between rm and the adjacent output
corresponding to data block [67]. In this case, time multiplexed pilots scheme has
more advantage than the frequency multiplexed pilots scheme for channel estimation.
Therefore, in the following chapters, we focus our attention on channel estimation
based on time multiplexed pilots method.
From the use of a BEM, the channel gains for a finite duration block are repre-
sented by vectors of basis coefficients. The advantage of the use of basis coefficients
as the state vector is that the length of the basis coefficient vectors for a block is less
than the length of the channel gain vectors so the number of unknowns to characterize
the channel is reduced. Based on PSAM, basis coefficients can be estimated by using
LS method [69, 70, 71], MMSE method [72, 73] or LMMSE method [1, 74]
Since single block BEM channel estimation techniques described above do not
use information contained within blocks other than the current block, the previous
schemes do not achieve optimal MSE performance. By taking the advantage of au-
toregressive (AR) model that can capture the channel dynamics [75], a new dynamic
model is derived to describe the time evolution of the basis coefficients as a multivari-
ate AR process in Chapter 3. Relying on a second-order multivariate AR model, a
Kalman smoother is developed to track basis coefficients from the pilot measurements
 22

in Chapter 4.
As the fading rate increases, more pilot symbols are required to obtain sufficiently
accurate channel estimates for reliable detection. This reduces the overall data rate
during fast fading to unacceptable levels for many applications. To improve the per-
formance in communication systems, according to turbo principle which was originally
proposed for channel decoding [76, 77], the iterative processing has been received con-
siderable attention for slow fading channels, and furthermore it is being taken into
account for fast fading channels. We will discuss this technique in detail below.

2.2.3 Joint Channel Estimation Data Detection

By using the turbo processing principle, the performance advantage of the turbo
equalizer has motivated a significant amount of research on methods of the iterative
receiver for slow fading channels, where the CSI is unknown a priori to the receiver.
To improve the performance for frequency-selective fading channels, an alternative
method is to use iterative channel estimation and symbol detection in the presence
of ISI. Channel estimation can be performed with detection simultaneously based on
blind methods [78, 79, 80] or based on pilot symbols [81]. For the trellis-based equal-
izer, those methods have high computational complexities. To reduce the complexity
of equalizer, channel estimation is separated from data detection in general, where
PSAM is used to obtain an initial channel estimate.
The early work on the design of a reduced complexity iterative receiver focused on
methods that add pilot symbols in both the transmitter and the receiver, the detected
data and pilot symbols at the receiver are then used to re-estimate the channel and
the data is detected again in SC systems for flat fading channels [82, 83, 84], for
frequency-selective fading channels [85], for time-varying frequency-selective fading
channels [86], for block-fading channels [87], and for MC-CDMA in [88].
 23

To simplify the receiver structure in slow fading channels discussed above, only
detected data values are then used to re-estimate the channel in decision directed
estimation approach and the data is detected again in practice. With good initial
channel estimates, joint channel estimation and symbol detection method has been
shown to perform well with OFDM in slow fading channels, where the channel state is
time invariant over an OFDM symbol period [89, 90, 91, 92]. In order to reduce error
propagation due to decision directed channel estimation, a threshold is exploited
for the detected data, and hence, only accurate detected data is used for channel
estimation in [90, 91]. For doubly selective channels, this technique has been proposed
for SC in [93, 94, 95, 96] with BER performance worse than the case of the ideal CSI.
Using pilots inserted in the frequency domain for OFDM systems, joint channel
estimation and data detection method was proposed in [68] with high complexity of
detection and long symbol duration, or [71] with high pilot to data ratio and perfor-
mance degradation. Within a single OFDM symbol period, LS or MMSE algorithm
was discussed for channel estimation in [71, 68]. In order to detect data efficiently
in the presence of interference between the sub-carriers, more costly detection al-
gorithms such as maximum a posteriori probability (MAP) are employed, which is
implemented using Bahl-Cocke-Jelinek-Raviv (BCJR) technique [97]. By choosing
the transmission block of the CE-BEM exactly as the OFDM symbol period in [68],
this CE-BEM has high modelling error due to Gibb’s phenomenon [98]. Since the sam-
pling signal over the transmission block is equivalent to a signal OFDM block which
is truncated by a rectangular window, spectral leakage exists [98]. To increase spec-
tral resolution, the pre-existing iterative channel estimation/joint detection methods
extend the OFDM symbol period to a longer period.
However, the main cost of detection is mitigating ICI, which increases with the
normalized Doppler frequency. For a given Doppler frequency, the ICI is related to
 24

OFDM symbol duration. Although short OFDM symbol blocks have little ICI, the
channel estimation does not work well because of the truncation problem. On the
other hand, long OFDM symbol blocks require expensive detection as the channel
variation over the time period of the OFDM block causes the compensation for ICI
to become more difficult to handle.
Furthermore, the cost of detection algorithms increases exponentially with the
order of modulation, making the previous algorithms unsuitable for OFDM systems
using higher order modulation in fast fading channels, limiting the practical use of
the previous iterative channel estimation/joint detection algorithms in fast fading to
lower order modulation such as quaternary phase-shift keying (QPSK). Additionally,
these methods also require a high pilot to data ratio for effective operation based on
a single OFDM block.
To overcome the drawbacks described above, it is necessary to devise the chan-
nel estimation based on a number of OFDM symbol blocks. This motivates a new
low-complexity joint channel estimation/symbol detection scheme for OFDM subject
to fast fading with a low pilot-to-data ratio, where the BEM channel coefficients
are found for transmission block consisting of multiple OFDM symbols. Moreover,
the new method can be used for coded OFDM systems which are widely applied in
standards. The detailed discussion will be described in Chapter 5.
 25

Chapter 3

System Model

In this chapter, we focus our attention on system model for fast fading channels.
Unlike the previous time blocks for channel processing, three time blocks are discussed
in this chapter, which are OFDM block, transmission block and interleaving block.
To simplify the channel estimation, we derive different measurement models, which
include measurement model for data in terms of channel for detection, measurement
model for channel in terms of data symbols for channel estimation, and measurement
model for channel in terms of pilot symbols for channel estimation. To model the
time evolution over the transmission block, the unknown parameters of the channel
are characterized as a multivariate autoregressive (AR) Gauss-Markov model.

3.1 Time Blocks for Channel Processing

The proposed iterative scheme includes three parts, which are data detection, channel
estimation and decoding. This motivates the use of three durations of time blocks at
the receiver, as shown in Figure 3.1: the OFDM block (OB), which contains Ncp cyclic
prefix (CP) and N subcarriers, is used as the basis for detection; the transmission
block (TB), which includes Kb OFDM blocks and Kp pilot blocks (PBs), is used
 26

for channel estimation; and the interleaving block, which contains Kt transmission
blocks, is used for decoding.
For fast fading channels, the channel is modelled as the basis expansion model
(BEM), where the channel gains are characterized as a linear combination of known
basis functions weighted by unknown basis coefficients. Due to Gibbs phenomenon,
when a short duration OFDM period is used as the BEM period, the channel will
not be modelled accurately. To solve this problem, we propose to make the BEM
period consisting of multiple OFDM symbols. In such a case, a short OFDM symbol
block can still be used with a longer duration BEM period and hence, a low cost data
detection method is required. As a consequence, higher order modulation can used
for fast fading channels without an exponential increase in data detection cost.
The computational complexity of the channel estimator for a transmission block
of M samples is related to the number of the BEM coefficients, Q + 1, which is
bounded by Q ≥ 2dfd M Ts e, where fd is the maximum Doppler frequency, Ts is the
sampling interval that equals to the data symbol period [1]. Therefore, for a fixed
Doppler frequency, the longer the transmission block, the larger the number of the
BEM coefficients and hence the larger the cost of the channel estimator. Because of
truncation effects, the shorter the transmission block, the more spectral leakage will
be seen as the bandlimited fading process is imperfectly modelled for the short time
duration of the transmission block creating significant modelling errors. Therefore,
we propose a suitable size for transmission block in channel estimation. In order to
deal with bursts of errors due to temporary deep fading, a large interleaver is used to
enable the convolutional code to correct spread data bits over both times when the
channel is in a good condition and in a bad condition [99].
In the following sections, we describe these three types of blocks, their applications,
and how fast fading affects the different time blocks for processing.
 27

OFDM Block (OB)

CP OFDM symbol
| {z }| {z }
Ncp N
Transmission Block (TB)

OB0 P B0 ··· OB(Kb −1) P B(Kp −1)

| {z }
κ=1

T B0 ··· T B(Kt −1)

Interleaving Block

Figure 3.1: Block structure with κ = 1

3.1.1 Time Domain and Frequency Domain Model for a Sin-

gle OFDM Block

The fading channel in wireless communications is often modelled as a wide-sense

stationary uncorrelated scattering (WSSUS) process, which is discussed in detail in
Appendix A. Given a sampling interval Ts , the impulse response of the radio channel
in the discrete time is denoted as h(n, l) for propagation path l at sample n and
h(n, l) = 0 for l < 0 or l > L, where L is the channel order, the received signal in the
time domain at sample n is given by

L
X
yt (n) = h(n, l)u(n − l) + vt (n) (3.1)
l=0

where yt (n) is the received signal at sample n, u(n) is the transmitted data at the
nth sample, vt (n) is an additive white Gaussian noise (AWGN) with zero mean and
variance σv2 .
Consider an OFDM block with the length of Ns = N + Ncp . For notational
convenience, we choose n to denote the sample index and k to denote the index of
the OFDM block so that n = kNs + Ncp + i for i ∈ [0, N − 1] being the index of
the sample within the considered OFDM block. In order to express the received
 28

signal in matrix form, we define a number of vectors and matrices that are needed
in this development. First, the sequences for the kth OFDM block are denoted as
ytk (i) = y(kNs + Ncp + i), uk (i) = u(kNs + Ncp + i), hk (i, l) = h(kNs + Ncp + i, l) and
vtk (i) = v(kNs + Ncp + i). Then, we define ytk , uk , hkl and vtk as the received vector,
the data vector, the channel gain vector and the noise vector of the lth path for block
k, respectively, where hkl = [hk (0, l), ..., hk (N − 1, l)]T , ytk = [ytk (0), ..., ytk (N − 1)]T ,
uk = [uk (0), ..., uk (N − 1)]T and vtk = [vtk (0), ..., vtk (N − 1)]T . After removing CP, the
received signal can be rewritten in matrix-vector form as

ytk = Htk uk + vtk (3.2)

where
 
k k k
 h (0, 0) 0 ··· 0 h (0, L) · · · h (0, 1) 
 
 hk (1, 1) hk (1, 0) 0 ··· 0 ··· h (1, 2) 
k
 
 
 .. ... ... ... ... ... .. 
 . . 
k 
Ht =  
.. .. .. .. 
hk (L, L) hk (L, L − 1) . . . . 0 
 
 .. .. .. .. .. .. 
 . . . . . . 0 
 
 
0 0 hk (N − 1, L) ··· ··· ··· hk (N − 1, 0)
(3.3)
or
L
X
Htk = D(hkl )Sl (3.4)
l=0

where D(z) denotes a diagonal matrix with vector z on its diagonal and all other
entries being zero, Sl is given by circularly shifting identity matrix IN down with the
delay of l samples. A circular matrix shift is the operation of rearranging the matrix,
circularly shifting rows down (upper) or column left (right). The size of Htk is N by
N . When the channel is time invariant over the period of an OFDM symbol, and the
 29

length of CP is larger than the channel order L, then all of hk (i, l)’s are not a function
of i for all subcarriers i where i ∈ [0, N − 1], i.e., hk (i, l) is same for all i’s, resulting
in Htk being a circulant matrix. However, when the channel is time varying over one
OFDM block, hk (i, l) has different value for each i and Htk is not a circulant matrix.
In the frequency domain, the received signal at the kth OFDM block is then
expressed as

yk = F Htk F H sk + F vtk
(3.5)
= Hk sk + vk

where F is the fast Fourier transform (FFT) matrix with the size of N by N and
√
the (m, n)th entry Fm,n = 1/ N e−j2π(m−1)(n−1)/N , F H is the inverse fast Fourier
transform (IFFT) matrix. yk is an N by 1 vector defined as yk = F ytk , Hk =
F Htk F H and vk = F vtk . sk is the transmitted signal vector in the frequency domain
at the kth OFDM block, the size is N by 1.
In the special case that Htk is a circulant matrix, the matrix Hk is diagonal, and
hence, no interference between subcarriers exist. Therefore, no intercarrier interface
(ICI) occurs in slow fading channels. However, when the channel is time varying over
the OFDM symbol duration, the sample of the channel gain for each propagation
path changes from the beginning to the end of each OFDM symbol. Therefore, the
matrix Hk is not diagonal and becomes a banded matrix, the gray part as shown in
Figure 3.2. Cross-terms indicating interference between subcarriers are created and
result in ICI. In this case, the channel gain at each time sample of the OFDM block is
needed for data detection. In fact, the ICI on a subcarrier mainly comes from several
neighboring subcarriers. If the Doppler frequency increases, more symbol energy leaks
to neighboring subcarriers as shown in Figure 3.2.
In the case that the channel impulse response (CIR) is known, low-complexity de-
 30

fd T ↑
Hk Hk

Figure 3.2: Structure of Hk matrix

tection methods have been proposed to mitigate ICI. Whereas, in the case of initially
unknown channels, we propose joint channel estimation data detection which will be
discussed in detail in Chapter 5.

3.1.2 Time Domain Descriptions for a Transmission Block

vt

s u add ũ N ỹt
ifft add CP h(n, l)
pilot

Figure 3.3: Discrete-time structure in OFDM systems

When modelling a fast fading channel, we recall that the transmission block con-
sists of Kb OFDM blocks and Kp pilot blocks. On the one hand, short transmission
block will increase the error of the channel model, on the other hand, a long trans-
mission block will increase the complexity of the channel estimation. Therefore, there
is a tradeoff between the complexity and the channel modelling error. The choices of
Kb and Kp for the transmission block will be discussed in detail in Chapter 6.
As described in Chapter 2, the time multiplexed pilots scheme is preferable to
channel estimation in fast fading channels due to the advantage of temporal separation
from the data symbols. With a pilot block in the form of [01×L 1 01×L ], as shown in
Figure 2.3, inter-block interference between pilot block and OFDM symbol block can
 31

be avoided. Within a transmission block, Kp pilot blocks are uniformly inserted in

Kb OFDM blocks, where Kb = κKp , κ is the number of OFDM blocks between two
adjacent pilot blocks and κ ≥ 1 is an integer. To minimize the total channel of mean
square error (MSE), we will show that the optimal κ is 1 in Chapter 5, i.e., 1 pilot
block for every OFDM block, and hence, the transmitted signal u in time domain is
expressed as
· ¸T
ũ = uT0 dT0 · · · uTKp −1 dTKp −1 (3.6)

where up and dp for p = 0, ..., Kp − 1, are data and pilot block, respectively; u =
[uT0 , ..., uTKp −1 ]T , and d = [dT0 , ..., dTKp −1 ]T .
After adding pilot symbols, the received signal vector for the transmission block
in the time domain is then expressed as

 
d
yt0
 
 
 yt0 p 
 
 . 
ỹt = 

.. 
 (3.7)
 
 d 
yt(Kp −1) 
 
p
yt(Kp −1)

d p
where ytp and ytp for p = 0, ..., Kp − 1, are the received signal vector corresponding
to up and dp , respectively.
Similar to (3.2), we express ỹt as

ỹt = H̃t ũ + ṽt (3.8)

where H̃t is the channel matrix of transmission block, ṽt is the AWGN vector in the
transmission block.
The relationship between H̃t and the received signal ỹt is illustrated in Figure 3.4,
 32

u0 d0 u1 d1 u2 d2

d
yt0

p H̃t
yt0

d
yt1

p
yt1

d
yt2

p
yt2
received
signal transmitted
ỹt data
ũ

Figure 3.4: Relationship between the transmitted data and the received signal in the
time domain for single transmission block

with the transmitted index on the X-axis and the received index on the Y-axis. Since
H̃t is a banded matrix, H̃t is split into three parts: Ht , HtP and Htsd . In the figure,
Ht is the white central diagonal part, corresponding to the relationship between the
transmitted signal ui and the received signal ytid ; HtP is the gray part, corresponding
to the relationship between the transmitted signal di and the received signal ytip ; and
Htsd is the black triangular part, corresponding to the interblock interference that
emerges due to the channel delay spread. As a result, the received signal, ỹt , in the
transmission block can be separated into two portions, which are the received signal
at the data position, ytD , and the received signal at the pilot position, ytP . Therefore,
Equation (3.7) can be separated as

 
d
 yt0
 .. 
ytD 
= .  = Ht u + Htsd d˜ + vtD (3.9)

 
d
yt(K p −1)
 33

 
p
 yt0 
 .. 
ytP 
= .  = HtP d + vtP (3.10)

 
p
yt(Kp −1)

where d˜ = [d˜T0 , ..., d˜TKp −1 ]T and d˜p contains the first and last L entries of dp for
p = 0, ..., Kp − 1. vtD and vtP are the corresponding noise vectors. The term Htsd d˜
corresponds to the interference of the pilot blocks to their adjacent data blocks. By
use of the pilot form discussed above, d˜p = 0 for p = 0, ..., Kp − 1, and hence, the
interference part can be mitigated. In such a case, Equation (3.9) becomes

 
d
 yt0 
 . 
ytD =

..  = Ht u + vtD
 (3.11)
 
d
yt(K p −1)

ytD and ytP will be discussed in detail in Section 3.2.1 and Section 3.2.2, respectively.

3.1.3 Coding Block

Since the channel is time-variant and frequency-selective, the channel exhibits bursty
error characteristics [2, 99]. In order to deal with temporary deep fades, an effective
method is to exploit coding with a interleaver. After interleaving, errors are spread
out so that error events for two adjacent bits in the output data stream are inde-
pendent [2]. The major drawback with large interleavers is decoding delay and the
computational complexity in the iterative decoding scheme.
The size of the interleaver is chosen based on the Doppler frequency. Lower
Doppler frequency requires a longer interleaver to obtain the same time diversity as
a higher Doppler frequency and vice versa. On the other hand, in order to keep the
complexity of the channel estimation low, the channel is estimated over a transmis-
 34

sion block and interleaving is exploited over a coding block which contains several
transmission blocks. In Chapter 6, we will show how to use the extrinsic information
transfer (EXIT) chart to select the interleaver size for good performance.

3.2 Three Measurement Models Introduction

As mentioned before, the proposed iterative receiver includes data detection, channel
estimation and decoding. This motivates the need for three different measurement
models. Due to the separation between the received signal ỹt , we can develop the
three measurements models: measurements for data given channel, measurements
for channel given pilots, and measurements for channel given data, which will be
discussed in this section.

3.2.1 Measurements for Data Given Channel

For OFDM systems, if a transmission block contains Kp pilot blocks,where each pilot
block has the form of [01×L 1 01×L ], and Kp OFDM blocks, where each OFDM block
contains Ns = N + Ncp samples, the index n in Equations (3.2) and (3.3) becomes
n = kNs + Ncp + p(2L + 1) + i for k ∈ [0, Kb − 1] p ∈ [0, Kp − 1] and i ∈ [0, N − 1],
the size of ytD in Equation (3.11) is Kp N by 1 and Ht is given by

 
0
Ht 
 ... 
Ht = 


 (3.12)
 
(Kp −1)
Ht

(Kp −1) K −1
or Ht = diag{Ht0 , ..., Ht p
} = ⊕k=0 Htk , and Htk is the channel matrix in the
time domain of the kth OFDM block, expressed as Equation (3.3). Therefore, the
 35

received signal in the frequency domain is expressed as

 
 y0d

 . 
D
y = . 
.
 = Hs + v
D
(3.13)
 
d
y(Kp −1)

K −1 K −1
where y D = ⊕k=0
p
F ytD , H = diag{H0 , ..., H(Kp −1) } = ⊕k=0
p
Hk , s is the transmit-
ted data in the frequency domain, s = [sT0 , ..., sTKp −1 ]T ; v D is AWGN vector in the
frequency domain.
For single-carrier (SC) systems, if a transmission block contains Kp pilot blocks,where
each pilot block has the form of [0L 1 0L ], and Kb data blocks, where each data block
contains N samples, the index n in Equation (3.1) becomes n = kN + p(2L + 1) + i
for k ∈ [0, Kb − 1] p ∈ [0, Kp − 1] and i ∈ [0, N − 1]. Based on the description of Equa-
tion (3.1), for the nth sample of a transmission block, the received signal is defined as
yt (n) = hTtn Un +vt (n) where htn = [h(n, 0), ..., h(n, L)]T , Un = [u(n), ..., u(n−L+1)]T .

3.2.2 Measurements for Channel Given Pilots

As indicated in Section 3.1.1, when the channel varies significantly over one OFDM
duration, we need N (L+1) channel gains to obtain the channel matrix Htk in Equation
(3.2). Unfortunately, the number of unknowns is larger than that of measurements.
To overcome this difficulty, the estimation of basis coefficients for a basis expansion
model (BEM), which is discussed in detail in Appendix B, is instead. Therefore, the
channel gain for the transmission block of the lth path is hl = Exl , where E is
the basis matrix and xl specifies the basis coefficients for path l for the transmission
block, defined as xl = [xl (0), ..., xl (Q)]T , Q is the number of BEM order denoted by
Q ≥ 2dfd M Ts e, where Ts is the sampling time and M is the number of samples in
each transmission block; xl (q) is the coefficient for path l of the qth basis functions,
 36

all defined in Appendix B. With the number of pilot symbols 2L + 1 for each pilot
block, the required number of pilot blocks Kp for each transmission block should be
satisfied Kp ≥ Q + 1.
Define x as the basis coefficient vector,


x
 0  (0)
 . 
 .. 
   


 
x
 0  x 0 (Q) 
 .   . 

x= .  .. 
 . = . (3.14)
   


xL  
 xL (0) 
 . 
 .. 
 
 
xL (Q)

Based on the discussion of Section 3.1.2, the received signal corresponding to pilot
symbols is obtained as
   
p
 yt0   C0p

 .   . 
ytP = 

..  =  ..  x + vtP
  
    (3.15)
p p
yt(K p −1)
CKp −1

= C P x + vtP ,

where vtP is the AWGN vector, the length of ytP is Kp (L+1), C P = [C0p T , ..., CK
p
p −1
T T
] ,
and  
 epp 0Q+1 · · · 0Q+1 
 . .. .. .. 
Cpp = 

.. . . .  ,
 
0Q+1 · · · · · · epp +L
p
where pp is the first index of ytp , defined as pp = (k+1)Ns +k(2L+1) for k ∈ [0, Kp −1];
ei = [Ei (0) · · · Ei (Q)] is a basis function vector at sample i with the size of 1 by (Q+1),
 37

and Cpp is a block matrix, where the size of Cpp is (L+1) by (L+1)(Q+1). In the case
that the number of pilot blocks Kp is larger than the number of OFDM blocks Kb , as
long as the pilot form is the same as the described in Section 3.1.2, we can still use
Equation (3.15) with Kb instead of Kp . As a consequence, pp = (k + 1)Ns + p(2L + 1)
for k ∈ [0, Kb −1] and p ∈ [0, Kp −1]. Hence, Equation (3.15) denotes the measurement
model for channel in terms of pilot symbols for channel estimation in the time domain.

3.2.3 Measurements for Channel Given Data

In iterative joint channel estimation/data detection schemes, previously detected val-

ues for the transmitted data are used in conjunction with the measurements to es-
timate the channel state. This requires a model relating the channel state to the
measurements given the detected data values. This model is developed in this sec-
tion. After the initial channel estimation with pilot symbols, the measurements in the
frequency domain y D taken from the detected symbols are used to calculate the BEM
coefficients. In this section, first we turn our attention to develop the relationship
between the received signal and the BEM coefficients for each OFDM block. Then we
exploit the measurement signals for channel estimation based on the detected data
over a transmission block.
With the use of a BEM, the channel gain vector of the kth OFDM block in a
transmission block for the lth path are then expressed as hkl = Ek xl , where Ek =
E[kNs + Ncp + p(2L + 1) : kNs + Ncp + p(2L + 1) + N − 1, :)], k ∈ [0, Kb − 1] and
 38

p ∈ [0, Kp − 1]. In such a case, the time domain receiver for the kth OFDM block is

L
X
ytk = D (Ek xl ) Sl F H sk + vtk
l=0
XL
¡ ¢ (3.16)
= D Sl F H sk Ek xl + vtk
l=0

= Ck x + vtk

where x is basis coefficient vector defined in Equation (3.14), and the matrix Ck is a
block matrix, expressed as

· ¸
Ck = C0k · · · CLk (3.17)

¡ ¢
with Clk = D Sl F H sk Ek .
The received signal in the frequency domain is then expressed as

yk = F Ck x + vk
(3.18)
= Mk x + vk

where Mk = [M0k · · · MLk ] with Mlk = F Clk = F D(Sl F H sk )Ek . When the
transmitted data is given, Equation (3.18) can be used for channel estimation based
on the decision-directed method.
For each transmission block, the measurement in the frequency domain y D of
 39

Equation (3.13) is rewritten as

     
 y0d   M0   v0 
 ..    ..   
yD =   x +  ... 
 . = .   
     
d
y(Kp −1) MKp −1 vKp −1
 
(3.19)
 M00 ··· ML0 
 .. .. 
= . ··· .  x + vD

 
M0(Kp −1) · · · ML(Kp −1)

= MD x + v D ,

When the data in the transmission block is detected, Equation (3.19) is then used
to estimate the basis coefficients for data-aided channel estimation, which will be
discussed in detail in Chapter 5.
In summary, we have provided different system models that find applications
in fast fading channels. Of the models described, Equation (3.5) is used for data
detection when the channel gain is available. As the case of pilot aided channel
estimation, Equation (3.15) is used for the time domain channel estimation. As
the case of decision-directed channel estimation, Equation (3.19) is used for channel
estimation based on the detected data.

3.3 Time Domain Evolution of the Channel

Of the different channel models, an autoregressive (AR) Gauss-Markov model is ex-

ploited to describe the time-varying channel in general [75]. Since the channel is
bandlimited, the evolution of the BEM coefficients for each propagation path in a
fast fading channel between transmission blocks can be modelled as a multivariate
AR process. This AR model allows the estimated BEM coefficients for the previous
 40

transmission block to be used to estimate the BEM coefficients for the current trans-
mission block, reducing the need for pilot symbols and improving channel estimation
accuracy for the initial iteration of the channel estimation/data detection systems.
In order to exploit channel information contained within transmission blocks other
than current transmission block, a novel channel model is derived, where the unknown
basis coefficients are characterized by a multivariate AR Gauss-Markov model, dis-
cussed in detail below.
When the channel is modelled by a BEM for each transmission block, we assume
that the basis coefficient vector for each propagation path is a multivariate AR process
of order K, the basis coefficient vectors for each propagation path at transmission
block n are given by

K
X
xl [n] = − Al [i]xl [n − i] + wl [n − 1] (3.20)
i=1

where the Al [i]’s are multivariate AR coefficient matrices with the size of (Q + 1) by
(Q + 1), and wl [n] is an AWGN vector process with zero mean and a covariance of
Qlf . The bandlimited nature of the fading process makes the actual xl [n] process of
infinite order. However, in practice, the process is well modelled as an AR process of
finite order K with the approximation becoming more accurate with increasing K.
To calculate the multivariate AR coefficients and the covariance Qlf , it is necessary
to create a state vector xl [n] to contain basis coefficients for blocks n − K + 1 to n:

 
 xl [n] 
 .. 

xl [n] =  . . (3.21)

 
xl [n − K + 1]
 41

Equation (3.20) can be rewritten as

   
−Al [1] −Al [2] · · · −Al [K] I 
   
 I I · · · 0  0
   
xl [n] =  . ..  x l [n − 1] +  .  wl [n − 1] (3.22)
 .. ..   .. 
 . .   
   
0 ··· I 0 0

where I and 0 are (Q + 1) by (Q + 1) identity matrices and (Q + 1) by (Q + 1)

zero matrices, respectively. The values of the AR coefficient matrices satisfy the
multichannel Yule-Walker equation [100, 101]:

· ¸
Al Rl = Qlf 0 · · · 0 (3.23)

where
· ¸
Al = I Al [1] · · · Al [K] , (3.24)

and  
 Rx (0) ··· Rx (K) 
 
 Rx (−1) · · · Rx (K − 1)
 
Rl =  .. .. . (3.25)
 ... 
 . . 
 
Rx (−K) ··· Rx (0)

These matrices can be efficiently calculated by the use of Whittle-Wiggins-Robinson

Algorithm (WWRA) [102, 103], which is discussed in Appendix C. The matrix Rl has
a Toeplitz-block structure with Rx (i) = E[xl [n + i]xl [n]H ], and Rx (−i) = RxH (i).
 42

To calculate Rx (i), we get the covariance matrix of each channel path as

Rh (i) = E[hl [n + i]hl [n]H ]

= E[(Exl [n + i])(Exl [n])H ] (3.26)

= ERx (i)E H .

Based on Jakes’ model, the correlation function of each channel path is expressed as
rl [k] = σl2 J0 (2πfd Ts k), where J0 is the zeroth order Bessel function[104] and σl2 is the
variance of the lth path. The average power of the channel is normalized to one, i.e.,
PL 2
l=0 σl = 1. So the cross-covariance matrix Rh (i) is given by

 
 rl [M i] · · · r l [M (i − 1) + 1]
 
 rl [M i + 1] · · · rl [M (i − 1) + 2]
 
Rh (i) =  . . . (3.27)
 . ... . 
 . . 
 
rl [M (i + 1) − 1] · · · rl [M i]

Since E H E = I, then we calculate Rx (i) from Equations (3.26) and (3.27) as Rx (i) =
E H Rh (i)E. A difficulty encountered when computing the coefficients for higher order
models is that Rl is ill-conditioned due to the band-limited process of the channel
making the ratio of the magnitude of the smallest to largest singular value of Rl very
large. This problem can be overcome by adding small positive values to the diagonal
entries of Rx (0) or by using other techniques [105, 106]. If Al and Qlf are stationary,
which means that the Doppler frequency does not vary over the time, we can calculate
these matrices off-line. However, when the Doppler frequency changes over the time,
the calculation of Al and Qlf should be on-line.
 43

Chapter 4

Second-Order Kalman Filtering

Channel Estimation in Fast Fading

4.1 Introduction

In this chapter, we consider the problem of channel estimation in the presence of fast
fading. The channel state is characterized by basis coefficients over a fixed period
of time. Unlike the previous Kalman filter based channel estimation methods which
focus on the time variation within one OFDM block, the proposed algorithm takes
the time variation within each transmission block into account.
By use of a basis expansion model (BEM), a second-order multivariate autore-
gressive (AR) model is exploited to characterize the unknown channel state over each
transmission block. In this chapter a Kalman smoother is proposed, which uses more
measurements and is more accurate compared to the Kalman filtering [100]. It will
be shown that the proposed technique provides channel mean square error (MSE) is
near to that of the optimal minimum mean square error (MMSE) Wiener filter. The
computational cost of this new algorithm is on the same order as existing channel
 44

estimation algorithms. In addition, the proposed approach is robust to variations

of the radio channel parameters from the design values. For data detection, a low-
complexity MMSE and decision-feedback equalizer (DFE) algorithm [5] is used to
suppress intercarrier interference (ICI).

4.2 Second-Order Kalman Filtering Algorithm

Unlike the previous Kalman filter methods proposed for fast fading channel estimation
[13, 107], where a Kalman filter was developed in single-carrier (SC) systems, the time
evolution of basis coefficients over each transmission block is described as a second-
order multivariate AR model for OFDM systems. The second-order Kalman filter
will be shown to track the basis coefficients from the pilot measurements requiring
only a low pilot to data ratio. Moreover, a Kalman smoother, which incorporates the
following measurements, gives a more refined estimate than the previously presented
channel estimation methods. The channel state information (CSI) is estimated based
on pilot symbols which are multiplexed in the time domain. The basis coefficients are
first estimated using pilot symbols, then the CSI is calculated from basis matrices.
For Kalman filtering algorithms, the state space model is used for estimation,
which includes the state model and the measurement model. As discussed in Chapter
3, each transmission block includes Kb OFDM blocks (OBs) and Kp pilot blocks,
which are uniformly inserted in between the OFDM blocks. To exploit the state model
of the basis coefficients over the transmission block, we consider that a multipath
channel is modelled as a BEM with Q + 1 basis coefficients over the transmission
block and has L + 1 paths. From the description of the state model given in Chapter
 45

3, the state vector of the second-order, K = 2, is given by

 
 x[n] 
x[n] =   (4.1)
x[n − 1]

where x[n] is the basis coefficient vector with the size of (L + 1)(Q + 1) by 1 defined
in Equation (3.14) at the nth transmission block, that is, xl [n] is the vector of the
Q basis coefficients for the lth channel path for the nth transmission block. Consider
an OFDM symbol with N subcarriers, Ncp cyclic prefix (CP) samples and each pilot
block with the form [01×L 1 01×L ], the nth transmission block will include M =
(N + Ncp )Kb + (2L + 1)Kp samples for each channel path, defined in vector form as

 
 h(n, l) 
 .. 

hl [n] =  .  = Exl [n] (4.2)

 
h(n + M − 1, l)

where E is the basis function matrix defined in Equation B.2.

The state model is described by
     
 x[n]  x[n − 1] I 
  = Φ  +   w[n − 1] (4.3)
x[n − 1] x[n − 2] 0

where Φ is the state transition matrix given by

 
Φ11 Φ12 
Φ= . (4.4)
Φ21 Φ22

where Φij for i, j = 1, 2 is a block matrix. To simplify the equation, we use ⊕ to

specify the direct sum of matrices: ⊕Ll=0 Al = diag{A0 , ..., AL } where diag is the
 46

block diagonal operator [108]. Therefore, Φ11 = − ⊕Ll=0 Al [1], Φ12 = − ⊕Ll=0 Al [2],
Φ21 = − ⊕Ll=0 I, and Φ22 = − ⊕Ll=0 0, where I is (Q + 1) × (Q + 1) identity matrix, and
the Al [i]’s are multivariate AR coefficient matrices of the propagation path l with the
size of (Q + 1) by (Q + 1), respectively. w[n] is an AWGN process with a covariance
matrix Q given by  
Q1 0
Q=  (4.5)
0 0

where Q1 = ⊕Ll=0 Qlf , Qlf is the covariance matrix of the state vector for path l.
Let σl2 be the variance of the lth path. The average power of the channel is
P
normalized to one, i.e., Ll=0 σl2 = 1. In Appendix C, it is shown how to solve for
the A[i] and Qf matrices using the Whittle-Wiggins-Robinson Algorithm (WWRA)
[102, 103]. From calculations of A[i] and Qf , we obtain Al [i] = A[i] and Ql = σl2 Qlf .
Now we turn our attention to the measurement model over the nth transmission
block, which is expressed as

y[n] = G[n]x[n] + v[n] (4.6)

where G[n] is the measurement matrix and v[n] is the measurement noise with each
covariance σv2 = N0 /2, where N0 is power spectral density. As described in Chapter
3, the CSI is estimated based on pilot symbols which are multiplexed in the time
domain. For time multiplexed pilots, the measurement model of a transmission block
is given by Equation (3.15), i.e., ytP = C P x + vtP , where ytP is the received signal
vector corresponding to pilot symbols, vtP is the AWGN vector. Therefore, for the
nth transmission block, we obtain y[n] = ytP [n], v[n] = vtP [n] and G[n] expressed as

· ¸
G[n] = C P [n] 0 (4.7)
 47

The measurement noise process v[n] is an AWGN vector process independent of w[n],
and the variance matrix R[n] is a diagonal matrix with all entries on the main diagonal
given by variance σv2 .
The Kalman filter equations for a single block of the channel measurements are
given in Appendix E. The estimate of x[n] given all measurements y[n] up to block
m is denoted by x̂[n|m] with the corresponding error covariance matrix denoted as
P [n|m]. ŷ[n|n − 1] is the predicted measurement for block n, z[n] is the so-called
innovation sequence with a covariance matrix given by M [n|n − 1], and K[n] is the
Kalman gain for block n. Given an initial estimate of the state x̂[0] = 0 with error
covariance P [0] = ⊕Ll=0 Pl [0], where

  H   
 xl [0]   xl [0]    Rx [0] Rx [1]
Pl [0] = E    = ,
xl [−1] xl [−1] Rx [−1] Rx [0]

and Rx [i] is calculated based on Equation (3.26), the Kalman filter, which is calcu-
lated based on Appendix E, estimates the channel state at block n given measurements
up to y[n].
The Kalman filter is easily modified into a Kalman smoother [100]. The param-
eter estimation xk of the Kalman filter only takes into account the information in
[0, k]. However, by incorporating the future measurements relative to xk , that is, the
measurements between [0, n] for n > k, we can obtain a more refined estimate. If
the radio receiver can store the channel samples and wait until after the reception of
block n before decoding the data symbols for block n − 1, a substantial improvement
in channel estimation MSE and, thus, bit error rate (BER) is obtained. An esti-
mate of the channel BEM coefficients x̂ [n − 1|n] can be obtained from a sub-vector
of the state vector x̂ [n|n] from a standard Kalman filter. If the Kalman filter uses
an AR model of order K ≥ 2, this does not require any more computation and the
 48

MSE of x̂ [n − 1|n] is less than that x̂ [n − 1|n − 1] [100]. The addtional delay is a
transmission block length, which is tolerate compared to the interleaver block length.
Now we analyze the computational cost of channel estimation. For a given SNR
and fd , Kalman filter calculation requires the inversion of the square covariance ma-
trix of order (Q+1)(L+1) requiring O([(Q+1)(L+1)]3 ) operations. Comparing with
LMMSE estimation discussed in Appendix D, it can be seen that Kalman filter cal-
culation has the same cost as the computation of the LMMSE coefficient estimation
within one OFDM symbol block. By taking the advantage of second-order Kalman
smoother, it will be seen in Section 4.4 that the Kalman filter provides superior
channel estimation error to LMMSE estimation.

4.3 Wiener Bound

For any estimation procedure, it is useful to know what the performance of the best
possible estimation procedure of that type is in the sense of MSE. This can be difficult
to derive for discrete time estimation procedures, in particular for bandlimited pro-
cesses because the optimal filter is neither causal or of finite memory [109]. However,
it is fairly simple to calculate a bound on the estimation error for the continuous time
channel estimation problem via standard Wiener filter theory. Based on sampling
theory, estimation bounds for continuous time can be converted to the estimation
bounds for discrete time. In this section, we derive the MSE for the continuous time
Wiener filter to estimate the radio channel. Since the channel fading process is ban-
dlimited, it satisfies the Nyquist sampling theorem perfectly and the channel can be
perfectly reconstructed from its samples. Therefore, there exists an infinite memory
discrete time system that can perfectly reproduce the results of the continuous time
channel estimation system.
 49

In the continuous time domain, the estimated channel error is ε(t) = h(t) − ĥ(t).
According to [109], the MSE of the optimal filter is

Z ∞
2 SH (f )Sv (f )
E[ε (t)]min = df (4.8)
−∞ SH (f ) + Sv (f )

where SH (f ) is the Doppler power spectrum, given by Equation (A.4) in Appendix

A. Sv (f ) is the power spectrum of the noise, which is N̄0 . Due to insertion of pilot
symbols, N̄0 becomes
fs
N̄0 = N0 (4.9)
fp

where N0 is the power density of AWGN at the input of the receiver, fs is the sample
rate, and fp is the pilot rate. Therefore, the noise density N̄0 is proportional to fs .
It is not possible to reduce N̄0 by adding a low pass filter to prevent aliasing prior to
the channel estimation filter since this will destroy the time separability of the data
signals and pilot signals. If N0 is fixed and more pilot signals are sent, increasing fp ,
the Wiener bound is lowered indicating better channel estimation is possible. The
drawback is that the bandwidth efficiency of the radio channel decreases when more
pilot signals are used for channel estimations.
Since the Wiener filter is the optimal filter in terms of MSE, the Wiener filter error
gives us the lower bound on the MSE for all possible pilot-based channel estimation
procedures for a given pilot symbol density and SNR. We will show in the next section
that our new Kalman smoother comes close to Wiener filter MSE.

4.4 Simulation Results

We consider an OFDM system with N = 128 subcarriers and Ncp = N/8 CP for
each OFDM symbol. Each transmission block includes Kb = 10 OFDM symbols. We
 50

choose Q = 4, and P = 5 pilot symbols are inserted into each transmission block. We
simulate a one-path Rayleigh fading channel with the carrier frequency fc = 3.5GHz,
the sampling frequency fs = 1.4M Hz, and the symbol duration T = 91.4µs [18].
The system is designed for velocities v = 168.8km/h and v = 337.6km/h, which
result in fd T = 0.05 and fd T = 0.1, respectively. The ‘DPS-BEM’ and ‘Oversampled
CE-BEM’ are the results from using LS estimators to do channel estimation based
on the discrete prolate spheroidal BEM (DPS-BEM) [72] and oversampled complex
exponential BEM (CE-BEM) [110], respectively. The ‘KL-BEM’ is the results where
we use MMSE estimator to do channel estimation based on Karhunen-Loève BEM
(KL-BEM) [69]. The MSE results of the channel estimation are shown in Figure 4.1
for fd T = 0.05 and Figure 4.2 for fd T = 0.1. The Wiener bound is calculated from
Equation (4.8). We see that Kalman channel estimation improves MSE performance.
The small gap between the Wiener bound and the Kalman smoother shows that
the Kalman smoother improves the MSE performance significantly and provides near
optimal performance. It was found in simulation studies that for these Doppler fre-
quencies, increasing the order of the Kalman filter to K > 2 did not significantly
reduce the channel estimation MSE.
For our BER simulations, we use a DFE with MMSE described in [5], where
the ICI power on a subcarrier comes from several neighboring subcarriers. In the
simulation ,we assume that the ICI power for each subcarrier focuses on 2 neighboring
subcarriers. Quaternary phase-shift keying (QPSK) modulation is used for BER
simulation. The BER results for different channel estimations are shown in Figure 4.3
for fd T = 0.05 and Figure 4.4 for fd T = 0.1. The ‘Perfect CSI’ is the resulting BER
for the known CSI at the receiver, and the simulated BERs for different channel
estimators are shown in these figures. The results show that both Kalman filter and
smoother can improve the BER performance compared to previous channel estimation
 51

methods. It can be seen that for fd T = 0.1 and Eb /N0 around 30 dB the Kalman
smoother gives an improvement of about 1 dB for BER performance than the use of
BEM channel estimation alone.

0
10
DPS−BEM
oversampled CE−BEM
KL−BEM
Kalman Filter
−1
10 Kalman Smoother
Wiener Bound
MSE

−2
10

−3
10

−4
10
0 5 10 15 20 25 30
Eb/N0

Figure 4.1: Channel estimation MSE for fd T = 0.05

The new Kalman channel estimation is robust when the designed fd T or SN R

used to create the dynamic model and measurement model equations for Kalman
filter equations are not same as the true values. Simulation results are shown in
Figure 4.5 for the designed fd T = 0.05 when SN R = 10dB and Figure 4.6 for the
designed SN R = 15dB when fd T = 0.05. We can see that when the assumed fd T
is lower than the true value, Kalman channel estimation still works well. When the
assumed fd T is higher than the true value, the proposed method will fail. If this
 52

0
10
DPS−BEM
oversampled CE−BEM
KL−BEM
Kalman Filter
−1
10 Kalman Smoother
Wiener Bound
MSE

−2
10

−3
10

−4
10
0 5 10 15 20 25 30
Eb/N0

Figure 4.2: Channel estimation MSE for fd T = 0.1

algorithm is used in a radio receiver, a conservative design choice would be to set the
assumed Doppler frequency of the Kalman smoother to the highest expected level.

4.5 Conclusions

In this chapter, a time evolution model for basis coefficients in fast fading radio chan-
nels is introduced. This model allows channel state information measured in adjacent
time intervals to be combined together to improve the overall channel estimation
accuracy. A Kalman filter algorithm based on the derived dynamic model for basis
coefficients was introduced in this chapter. Simulation results show that Kalman fil-
 53

0
10
DPS−BEM
oversampled CE−BEM
KL−BEM
−1
10 Kalman Filter
Kalman Smoother
perfect CSI

−2
10
BER

−3
10

−4
10

−5
10
0 5 10 15 20 25 30
Eb/N0

Figure 4.3: Channel estimation BER for fd T = 0.05

ter/smoother improves the MSE and BER performance and is robust to changes of
Doppler frequency and SNR levels from the design values.
 54

0
10
DPS−BEM
oversampled CE−BEM
KL−BEM
−1
10 Kalman Filter
Kalman Smoother
perfect CSI

−2
10
BER

−3
10

−4
10

−5
10
0 5 10 15 20 25 30
Eb/N0

Figure 4.4: BER resulting from estimated channel state for fd T = 0.1
 55

0
10
DPS−BEM
oversampled CE−BEM
KL−BEM
Kalman Filter
Kalman Smoother
MSE

−1
10

−2
10
0.01 0.02 0.03 0.04 0.05 0.06 0.07 0.08 0.09 0.1
fdT

Figure 4.5: Channel MSE for SN R = 10 dB and designed fd T = 0.05

 56

0
10
DPS−BEM
oversampled CE−BEM
KL−BEM
Kalman Filter
Kalman Smoother
−1
10
MSE

−2
10

−3
10
0 5 10 15 20 25 30
Eb/N0

Figure 4.6: Channel MSE for fd T = 0.05 and designed SN R = 15 dB

 57

Chapter 5

Joint Channel Estimation and

Data Detection

In Chapter 4, the design of a second-order Kalman filtering method for fast fading
channels based on pilot symbol aided modulation (PSAM) was introduced. However,
as the fading rate increases, more pilot symbols are required to obtain sufficiently
accurate channel estimates for reliable data detection. This reduces the overall data
rate during fast fading to unacceptable levels for many applications. To improve
the performance of PSAM, it is necessary to exploit iterative channel estimation and
symbol detection approach. In this method, the detected data is used to estimate
the channel state information (CSI) and remove the need for dense pilot signalling
for accurate channel estimates.
An iterative receiver combining joint iterative channel estimation with symbol
detection for coded orthogonal frequency division multiplexing (OFDM) systems in
fast fading channels with Doppler frequencies up to 15% of the OFDM symbol rate
is considered in this chapter. The initial rough estimate of the channel using the
pilot signals provides a low enough error rate for the detected data from the first
 58

iteration to be used as a ‘virtual’ pilot signal which is dense in the time domain.
Data is detected and decoded before the channel is estimated again based on the
detected data. Traditional decision directed channel estimation is sensitive to errors
in previously detected data which can cause elevated error levels in data detected at
later times, this is known as error propagation. In order to reduce error propagation
due to decision directed channel estimation, the initial values for Kalman filtering
used for the decision directed channel estimation for each transmission block are only
from the channel estimation values obtained from pilot based channel estimation of
the previous transmission blocks. The receiver exchanges information between the
channel estimator and detector in an iterative fashion to obtain accurate estimates of
the CSI, and furthermore, the performance of the detection is improved. Therefore,
the proposed approach does not suffer from error propagation and the complexity of
decision directed channel estimation is comparable to that of PSAM while reducing
the pilot-to-data ratio.
We begin with the design of iterative receivers with perfect CSI. Then, we treat the
subject of iterative receiver design for coded OFDM systems in fast fading channels.

5.1 Introduction

In practice, communication systems utilize error correction coding to provide reliable

delivery on fading channels [2]. When the coded signal is transmitted, the soft infor-
mation, typically represented in terms of the log-likelihood ratio (LLR), is exchanged
between the detector and the decoder in an iterative fashion [111]. The iterative
processing principle, called turbo-decoding or detection, was originally proposed for
channel decoding [76, 77]. The iterative processing has received considerable attention
due to the improvement of the receiver’s performance. The use of turbo-processing
 59

for equalization and detection in fast fading channels with known CSI has been shown
to provide good performance [112, 113]. To help with later discussion, we focus our
attention on the conventional iterative structure for time-varying known channels in
this section.
awgn
c̃ c s y
Encoder Π Mapper Channel

Figure 5.1: Block diagram of the transmitter

y Ldet
D Ldet
E Ldec
A Ldec
D Hard
Detector Π−1 Decoder
− decision

−
Π
Ldet
A Ldec
E

Figure 5.2: Block diagram of the receiver

We consider a coded communication system depicted in Figure 5.1 for the trans-
mitter structure and Figure 5.2 for the iterative receiver structure, where Π and Π−1
are denoted the interleaver and the deinterleaver, respectively. At the transmitter, the
information bits are first encoded as c̃ and then interleaved into coded bits c = Π(c̃),
which are mapped to M −ary complexity-valued code symbols s. The encoded data
c̃ is reordered by the interleaver and transmitted over the channel. At the receiver,
the deinterleaver puts the LLR in proper sequence and passes it to the decoder.
The a priori LLR for the bit ck is given by

P (ck = +1)
LA (ck ) = ln . (5.1)
P (ck = −1)

where P (ck = +1) and P (ck = −1) are probabilities for bit ck = +1 and ck = −1,
respectively. Given the received signal y, the a posteriori LLR of the coded bit ck is
 60

expressed as
P (ck = +1|y)
LD (ck ) = ln . (5.2)
P (ck = −1|y)

With perfect CSI, the general iterative receiver consists of a soft-input soft-output
(SISO) detector, a SISO decoder, a bit interleaver and a deinterleaver as shown in
Figure 5.2. The SISO detector takes the received signal y and the a priori LLR, Ldet
A ,

and outputs the a posteriori LLR, Ldet

D . In most cases, for the initial detection step,

no the a priori information is available and hence Ldet

A = 0. The extrinsic information

of the detector Ldet det det

E = LD − LA is passed through the bit-deinterleaver to become

the a priori LLR Ldec

A of the decoder. The output of decoder, the a posteriori LLR

Ldec
D , is then passed through the hard decision device and the estimated bits are

obtained. The extrinsic information Ldec dec dec

E = LD − LA is fed back to the detector

and becomes the a priori input after bit-interleaving (Π). The detector uses this soft
information to obtain more accurate soft output Ldet det
D , which is then passed as LE

to decoder for further iterations. This cycle is repeated and more reliable values are
exchanged between the detector and decoder [113]. Therefore, the BER performance
of the receiver is improved.
The maximum a posteriori (MAP) algorithm proposed by Bahl et al. [114], called
the Bahl-Cocke-Jelinek-Raviv (BCJR) algorithm, is used to estimate the a posteriori
probabilities (APPs) for each bit. With Bayes’ rule, the a posteriori LLR of the
coded bit ck in Equation (5.2) is rewritten as

p(y|ck = +1)P (ck = +1)/p(y)

LD (ck |y) = ln
p(y|ck = −1)P (ck = −1)/p(y)
P (ck = +1) p(y|ck = +1)
= ln + ln (5.3)
P (ck = −1) p(y|ck = −1)
p(y|ck = +1)
= LA (ck ) + ln .
p(y|ck = −1)

To obtain the p (y|c), we sum over all symbols with the specified bit value: Ck,±1 =
 61

{c|ck = ±1},
X
p (y|ck = ±1) = p(y|c)P (c|ck ). (5.4)
Ck,±1

Equation (5.3) is modified as

P
Ck,+1 p(y|c)P (c|ck )
LD (ck |y) = LA (ck ) + ln P . (5.5)
Ck,−1 p(y|c)P (c|ck )

Using the max-log-MAP approximation in [115], that is,

ln(eδ1 + · · · + eδn ) ≈ max(δ1 , ..., δn ) (5.6)

Equation (5.5) is then simplified as

LD (ck |y) ≈ LA (ck ) + max {ln p(y|c) + ln P (c|ck )}

Ck,+1
(5.7)
− max {ln p(y|c) + ln P (c|ck )}
Ck,−1

Since (5.7) is derived based on (5.6), only the maximum likelihood is considered
when calculating this probability. Therefore, the value LD (ck |y) actually gives the
probability of the most likely path through the trellis [113].
In general, the complexity of the optimal BCJR algorithm is proportional to M L ,
and grows exponentially with the channel order, L. In order to reduce the complexity
of detection, we propose a low-complexity detection below.
 62

awgn
conv. c̃ c add ỹt
Π mapper s ifft add cp
ũ
channel
code pilot

Figure 5.3: Coded OFDM transmitter

Pilot signal y t
P Pilot-based Data-based
Channel Channel
s^ c^
Mapper P Encoder
~
yt Estimation Estimation
^
H
D
D y
Data signal y Remove LD Hard
P-1
t
FFT Detector +- Decoder
CP Decision

LA -
P +
LE

Figure 5.4: Coded OFDM system

5.2 Iterative receivers for OFDM systems in Fast

fading

Consider the coded OFDM system depicted in Figures 5.3-5.4. At the receiver, using
the initial channel estimates from pilot symbols, data is detected and then decoded.
The decoded data is then fed back into the data-based channel estimation block for the
estimation of the channel coefficients in the next iteration. The channel estimation,
data detection, and data decoding are repeated until convergence is reached or the
maximum allowable number of iterations has been performed.
The proposed iterative receiver uses three different time intervals as shown in
Fig. 3.1: the OFDM block (OB) is used as the basis for detection, the transmission
block (TB) for channel estimation, and the interleaving block for decoding. The
channel estimation for an interleaving block is first performed by using the pilot
symbols and then the detected data in subsequent iterations using a first order Kalman
filter iterating over all transmission blocks, which is described in detail in Section 5.2.1.
The data detection (based on OFDM blocks) and decoder (based on interleaving
 63

blocks) are described in Section 5.2.2.

5.2.1 Iterative Channel Estimation

Channel estimation is based on a transmission block, which includes Kb OFDM

blocks. The state model of basis coefficients is characterized as a first-order AR
model. From the description in Section 3.3, the dynamic model of the first-order for
the evolution of the basis coefficient vector from transmission block n − 1 to n is given
by
x[n] = Φx[n − 1] + w[n − 1] (5.8)

where x[n] is the state vector at transmission block n, the state transition matrix is
given as Φ = − ⊕Ll=0 Al where ⊕ is the direct sum of matrices, and w[n] is an AWGN
vector with the covariance matrix Q = ⊕Ll=0 Qlf , Qlf is the covariance matrix of the
state vector for the lth path.
For the initial channel estimation of a transmission block, the pilot measurements
in the time domain y P are used to estimate the basis coefficients. Similar to Equation
(4.6), we express the measurement model at transmission block n as

ytP [n] = C P [n]x[n] + vtP [n] (5.9)

where ytP [n] is the received signal corresponding to pilot symbols at transmission
block n, C P [n] is the matrix corresponding to basis functions corresponding to pilot
symbols at transmission block n, vtP [n] is an AWGN vector corresponding to pilot
symbols at transmission block n.
Since the size of unknown state vector x[n] is (Q + 1)(L + 1) by 1, we choose
Kp ≥ (Q + 1) to be able to get a minimum of one measurement per channel basis
coefficient. The tradeoff between the complexity and the performance exists in the
 64

iterative scheme. From the point of minimizing the total channel of mean square
error (MSE), we select Kp > (Q + 1). In such a case, (5.9) is overdetermined, and
the modified equation is given by

ỹP [n] = C̃P [n]x[n] + ṽP [n], (5.10)

where ỹP = (C P )H y P , C̃P = (C P )H C P , ṽP = (C P )H vtP with the variance matrix

RP = σv2 (C P )H C P . Using Equations (5.8) and (5.10), the basis coefficients are
estimated with a Kalman filter for the initial iteration.
The initial conditions of the state vector, x̂[0] and error covariance matrix P [0]
are taken from the estimated state vector and error covariance matrix for the last
transmission block of the preceding interleaving block estimated using the pilot signals
only. Using the pilot signals for initial conditions of the Kalman filters prevents
error propagation from the decision feedback propagating beyond a single interleaving
block. At system initialization, x̂[0] = 0 and P [0] = ⊕Li=0 σl2 Rx [0], where Rx [i] =
E[xl [n + i]xH
l [n]] and xl [n] is the basis vector for the lth path defined in Appendix

B. From the estimated channel coefficients for a given transmission block, x̂[n], the
estimated channel matrix for the kth OFDM symbol is calculated as

L
X
Ĥk = F D(Ek x̂l )Sl F H , k ∈ [0, Kb − 1]. (5.11)
l=0

where D(Ek x̂l ) denotes a diagonal matrix with vector Ek x̂l on its diagonal and all
other entries being zero, Sl is given by circularly shifting identity matrix IN down with
the delay of l samples. A circular shift of a matrix is the operation of rearranging the
matrix, circularly shifting rows down (upper) or column left (right). F is the Fourier
transform matrix.
After the initial channel estimation, the measurements in the frequency domain
 65

ỹ D , expressed as Equation (3.19), taken from the detected symbols are used to cal-
culate basis coefficients. Hence, the measurement model at transmission block n is

ỹ D [n] = M̂D [n]x[n] + v D [n] (5.12)

where ỹ D [n] is the received signal vector at transmission block n corresponding to

detected data, M̂D [n] is the block matrix corresponding to detected data at trans-
T T
mission block n denoted as M̂D = [M̂0 , ..., M̂Kp −1 ]T , where M̂k = [M̂0k · · · M̂Lk ],
M̂lk = F E lk , where E lk = D(Sl F H ŝ)Ek and ŝ is the detected OFDM symbol; v D is
an AWGN vector.
Similarly, when (5.12) is overdetermined, a modified measurement equation, given
below, is used:
ỹD [n] = M̃D x[n] + ṽD [n] (5.13)

H H H
where ỹD = M̂D ỹ D , M̃D = M̂D M̂D , ṽD = M̂D v D with the variance matrix
H
RD = σv2 M̂D M̂D . For the second and following iterations, the Kalman filter is rerun
over all transmission blocks and basis coefficients are re-estimated using Equations
(5.8) and (5.13).
We now turn our attention to computational complexity calculations. The most
expensive portions of the channel estimation algorithm is the computation of the
measurement matrices and the matrix inversion required in the Kalman filter iteration
calculations. For the initial channel estimation from pilots, the Kalman filtering
calculation requires the inversion of the square covariance matrix of order (Q+1)(L +
1) requiring O([(Q + 1)(L + 1)]3 ) operations. This is of the same order or operations
as the standard MMSE coefficient estimation within one transmission block.
For the iterative data-based channel estimation, the major computation is the cal-
culation of the measurement matrix M̂D , where the operation is evaluated from the
 66

number of element-element multiplications required for the matrix multiplications

as well as the FFT operation counts. To support this, calculation of E lk requires
O(N (Q + 1)) operations and FFT of each column of E lk requires O(N log2 (N )) op-
erations, hence M̂lk requires O(N (Q + 1) log2 (N )) operations. Since calculation of
H
M̂l1k1 M̂l2k2 requires O(N (Q + 1)2 ), calculation M̃D for each OFDM block requires
O(N (Q + 1)2 (L + 1)). Therefore, the total computational complexity for data-based
channel estimation in one OFDM block is O(N (Q + 1)2 (L + 1)). In the next sub-
section, it will be shown that this is of the same order as detection, indicating that
this channel estimation method, while more expensive than purely pilot-based ap-
proaches, will not require an increase in the order of the number of operations for the
channel estimation and symbol detection.

5.2.2 Symbol Detection

As mentioned before, the complexity of the optimal detection is very high. In order
to reduce the complexity of data detection, we use a suboptimal detection scheme in
this section.
We express the estimated frequency matrix Ĥk of Equation (5.11) for a single
OFDM block as Ĥk = Ĥk1 + Ĥk2 , where Ĥk1 is the diagonal matrix which has its
only non-zero values being the main diagonal elements of Ĥk , Ĥk2 is equal to Ĥk
with the exception of its main diagonal elements being zero [116].
For the initial detection, symbols are detected by ignoring ICI, using only the
1
elements on the main diagonal of the estimated channel transfer matrix, Ĥk1 , and
the received signal yk . To suppress ICI efficiently, we estimate the ICI from the
estimated symbol ŝi−1
k and remove the estimated ICI from the received signal yk ,
that is,
i i−1
yki = yk − Ĥk2 ŝk (5.14)
 67

where ŝik is the symbol estimate of the ith iteration, Ĥk2

i
is the estimated Hk2 of the
ith iteration, and ŝ0k = 0 for the first iteration. The values yki and Ĥk1
i
are used to
estimate the symbol sik for the ith iteration.
The detector and the decoder exchange the extrinsic information iteratively in each
detection scheme. In each detection loop, the detector uses the estimated channel Ĥki ,
the received signal yki and the a priori LLR, Ldet
A (c), and then the a posteriori LLR,

Ldet i
D (c|y ) is obtained by

P [ck = +1|y i , Ĥ i ]
Ldet i
D (ck |y ) = ln (5.15)
P [ck = −1|y i , Ĥ i ]

where Ĥ i = ⊕K b Kt −1
k=0 Hk is the estimated channel in the frequency domain over an
interleaving block, where Kb is the number of OFDM blocks for each transmission
block and Kt is the number of transmission blocks for each interleaver block.
In each detection loop, the detector uses the estimated channel Ĥ 1 , the received
signal y i and the a priori LLR, Ldet
A (ck ) supplied by a SISO decoder, to obtain

the a posteriori LLR [117, 118]. The extrinsic LLR obtained from the decoder,
Ldec dec i dec
E (c̃k ) = LD (c̃k |y ) − LA (c̃k ), is interleaved and then sent back to the detector

as the a priori LLR, Ldet

A . After several iterations, the LLRs pass through a hard

decision to get the estimated information bits.

For a single OFDM block, the cost of the SISO detection/decoding is O(N log2 N ),
with the interference calculation in Equation (5.14) requiring O(N 2 ) operations mak-
ing the overall cost O(N 2 ) per OFDM block.

5.3 Simulation Results

In this section, the performance of the proposed algorithm is tested in a simulated

wireless OFDM system transmitted over a radio channel. We simulate a 4-path
 68

WSSUS channel with mean propagation powers of 8/15, 4/15, 2/15 and 1/15 for
propagation delays of 0 to 3 samples respectively. Each path is subject to independent
Rayleigh fading generated using the method from [119] with Jakes’ model specifying
the autocorrelation function. We consider an OFDM system with N = 128 samples
for each OFDM symbol. The length of CP is Ncp = N/8. Gray-coded QPSK and 16
or 64-quadrature amplitude modulation (QAM) constellation are used in simulations.
The normalized Doppler frequency is set to fd T = 0.1 and 0.15 respectively. For a
WiMAX system operating at the carrier frequency of 5.0 GHz and the subcarrier
space is approximately 10 kHz, these Doppler frequencies map to radio terminal
velocities of 216 km/h and 324 km/h respectively. Based on the existing standards,
the simulations employ the convolutional code with 1/2 rate, constraint length 7 and
the generator polynomials G = [133 171] in octal form.

5.3.1 BER Performance

The channel is estimated using a first-order Kalman filter. Each transmission frame
consists of Kb = 10 OFDM blocks, so the length of the frame M = 1510 samples.
Each interleaving block consists of Kt = 10 transmission blocks; each transmission
block consists of Kb = 10 OFDM blocks and Kp = 10 pilot blocks. The pilot to
data ratio is low to 7/144 and the percentage of total power used on pilots is 1/145,
respectively. We choose Q = 4 basis functions for fd T = 0.1 and Q = 6 basis functions
for fd T = 0.15. Simulation results of coded OFDM are shown in Figures 5.5-5.9
for different normalized Doppler frequencies respectively. ‘CE ith’ shown in figures
represents ‘channel estimation ith iteration’. The mean square error (MSE) for the
channel estimation is shown in Figure 5.5. In Figure 5.6, the BER results of the
proposed channel estimation scheme are compared with [68], where the channel is
estimated using MMSE algorithm in the frequency domain and the pilot-to-data
 69

ratio is 20/128. The basis functions used in [68] are Fourier basis functions, which is
subject to Gibbs phenomenon resulting in channel modelling error. Specifically, the
algorithm in [68] uses a BEM over a single OFDM symbol whereas the new algorithms
extends the model over several OFDM symbols. It can be seen in Figure 5.6 that
for the new method when fd T = 0.1, the BER results using the estimated channel
approach the case of the ideal CSI after the third iteration. For fd T = 0.15, BER
results are shown in Figure 5.7.
The proposed algorithm is also tested with the higher order modulation constella-
tions of 16 and 64-QAM as shown in Figures 5.8 and 5.9, respectively. In these cases,
it can be seen that performance close to the ideal CSI case is obtained for higher
Eb /N0 values. Moreover, the cheap detection caused by short OFDM blocks allows
us to scale up to even higher order modulation if necessary.

5.3.2 Robustness Analysis

For the approach to be useful in field implementations, it is necessary for it to be

robust to variations of the true channel parameters from the design values. Robustness
to variations of Doppler frequency, number of propagation paths, and mean path
power will be demonstrated here. The simulation BER results shown in Fig. 5.10
are for Eb /N0 = 8 dB with the designed normalized Doppler frequency of fd T = 0.1.
When the designed fd T is lower than the true value, the proposed method will diverge.
If this algorithm is used in a radio receiver, a conservative design choice would be
to set the designed normalized Doppler frequency of the Kalman filter to the highest
expected level. In order to reduce the complexity cost of the channel estimation, the
designed Doppler frequency should be chosen close to the true values.
The approach is also robust to variations of the number of paths in the channel
estimator from the true L. The simulation BER results shown in Fig. 5.11 are for
 70

0
10
pilot−aided CE
data−aided CE 1st
data−aided CE 2nd

−1
10
MSE

−2
10

−3
10
6 6.5 7 7.5 8 8.5 9 9.5 10
Eb/N0

Figure 5.5: MSE resulting from estimated CSI for fd T = 0.1 with QPSK

Eb /N0 = 8 dB with the true L = 3. The designed mean power for each path is chosen
as 1. When the designed L is lower than the true value, the proposed method will
diverge. If this algorithm is used in the receiver, a conservative design choice would
be to set the designed number of paths of the Kalman filter to the highest expected
level. The channel estimator does not require knowledge of the mean power level of
each propagation path. Only the relative delays and number of propagation paths
are needed to obtain good channel estimation and data detection/decoding results.
 71

−1
10

−2
10

−3
10
BER

−4
10

−5
pilot−aided CE[68]
10 data−aided CE 1st[68]
data−aided CE 2nd[68]
−6
pilot−aided CE
10 data−aided CE 1st
data−aided CE 2nd
ideal CSI
−7
10
6 6.5 7 7.5 8 8.5 9 9.5 10
Eb/N0

Figure 5.6: BER resulting from estimated CSI for fd T = 0.1 with QPSK
 72

0
10

−1
10

−2
10

−3
10
BER

−4
10

−5
10
pilot−aided CE
−6 data−aided CE 1st iteration
10
data−aided CE 2nd iteration
ideal CSI
−7
10
6 6.5 7 7.5 8 8.5 9 9.5 10
Eb/N0

Figure 5.7: BER resulting from estimated CSI for fd T = 0.15 with QPSK
 73

0
10

−1
10

−2
10
BER

−3
10

−4
pilot−aided CE
10 data−aided CE 1st iteration
data−aided CE 2nd iteration
ideal CSI
−5
10
6 6.5 7 7.5 8 8.5 9 9.5 10
Eb/No

Figure 5.8: BER resulting from estimated CSI for fd T = 0.1 with 16-QAM
 74

0
10

−1
10

−2
10
BER

−3
10

−4 pilot−aided CE
10
data−aided CE 1st iteration
data−aided CE 2nd iteration
ideal CSI
−5
10
9 9.5 10 10.5 11 11.5 12 12.5 13
Eb/N0

Figure 5.9: BER resulting from estimated CSI with Kp = 10 for fd T = 0.1 with
64-QAM
 75

0
10
pilot−aided CE
data−aided CE 1st iteration
−1
10 data−aided CE 2nd iteration
ideal CSI

−2
10
BER

−3
10

−4
10

−5
10

−6
10
0.05 0.1 0.15
true fdT

Figure 5.10: BER resulting for designed fd T = 0.1 with QPSK

 76

0
10
pilot−aided CE
data−aided CE 1st iteration
−1 data−aided CE 2nd iteration
10
data−aided CE with true L and power

−2
10
BER

−3
10

−4
10

−5
10

−6
10
0 1 2 3 4 5
designed L

Figure 5.11: BER resulting for fd T = 0.1 and true L = 3 with QPSK
 77

5.4 Conclusion

We propose a new channel estimation/symbol detection methodology for fast fading

radio channels. The pilot-to-data ratio is low to 7/144. This method requires low
pilot overhead and computational complexity. The new method provides BER results
nearly as good as the detection/decoding with perfect knowledge of the channel state
information. This technique is robust to variations of the channel conditions from
the designed values.
 78

Chapter 6

EXIT Chart Analysis

This chapter deals with the convergence behavior of the iterative receiver with esti-
mated channel state information (CSI) in fast fading channels. By using the Extrinsic
Information Transfer (EXIT) chart, we can analyze and predict the performance of
the previous chapters without simulating the whole iterative receiver. The EXIT
chart analysis assumes that all log-likehood ratios (LLRs) are independent. There-
fore, it is an asymptotic analysis of the convergence behavior of iterative receivers.
The EXIT chart demonstrates that the proposed technique to achieve a bit error rate
(BER) nearly as good as when ideal CSI is available. Moreover, the EXIT chart can
be used to select the modulation, pilot density, and error correction code for good
performance of the proposed iterative receiver.

6.1 Introduction

In order to exploit the EXIT chart in fast fading channels, we give a brief introduction
in this section. The EXIT chart is a semi-analytical technique to investigate the
convergence behavior of iterative receiver schemes. It was first introduced by Stephan
ten Brink [120, 121] to analyze the behavior of turbo codes. The EXIT chart shows
 79

the mapping of the mutual information between the input and output of each system
component and the transfer characteristics are plotted into a single diagram which is
referred to as the EXIT chart. Now, the EXIT chart has been applied to investigate
iterative receivers such as turbo equalizer in multipath channels with time-invariant
impulse response [122] and time-variant impulse response [123], optimal mapping
in bit-interleaved coded modulation (BICM) [124] in AWGN channels, noncoherent
turbo detection [125] in fast fading channels, or [126] in multicarrier interleave division
multiple access.
The EXIT chart has also been proposed in [127] for adaptive turbo equalization,
[128, 129] for OFDM systems and [130] for BICM. In this chapter, we will show how
the EXIT chart can be used to predict the performance of the iterative receiver with
the imperfect CSI for OFDM systems in fast fading channels. Figure 6.1 illustrates
the analysis of the iterative processing, where the black line is used in the full system,
and the dotted line is used for EXIT chart analysis.

Ldet
A Ldet
E
Detector Π−1

Decoder Π
Ldec
A Ldec
E

Figure 6.1: Model for the EXIT chart analysis

[122]

6.1.1 Input of Transfer Function

The assumption of using a large interleaver assures that the a priori LLRs Ldec
A of

the decoder and Ldet

A of the detector are modelled as independent Gaussian random
 80

variables, respectively. For simplicity, they are denoted as LA , expresses as [121].

L A = µA · c + n A (6.1)

where c is the transmitted bit and c ∈ ±1. nA is an additive white Gaussian noise
(AWGN) with the variance σA2 . µA = σA2 /2. The conditional probability density
function (pdf) of LA is

2 2 2
e−((l−(σA /2)c) )/(2σA )
pA (l|c = c) = √ (6.2)
2πσA2

The mutual information, IA = I(c, LA ), between the transmitted bits c and the
a priori LLR LA , is expressed as [121]

Z ∞
1 X
IA = pA (l|c = c)
2 c=−1,1 −∞
(6.3)
2pA (l|c = c)
· log2 dl, 0 ≤ IA ≤ 1
pA (l|c = −1) + pA (l|c = 1)

or
Z ∞ 2 2 2
e−((l−(σA /2)c) )/(2σA )
IA = 1 − √ log2 [1 + e−l ]dl (6.4)
−∞ 2πσA2

Equation (6.4) is one-to-one relationship between σA and IA , that is, IA is a monotonic

function of σA , denoted as IA = J(σA ) with lim IA = 0 and lim IA = 1 for σA > 0,
σA →0 σA →∞

which correspond to no and perfect a priori LLRs respectively.

For a given mutual information IA , the variance is obtained from σA = J −1 (IA )
[121]. The corresponding a priori LLR Ldet dec
A (ck ) or LA (c̃k ) is then generated according

to Equation (6.4).
 81

6.1.2 Output of Transfer Function

Similar to the a priori LLR LA , the extrinsic LLRs of the detector and decoder
are denoted as LE for simplicity. The extrinsic information transfer function is then
defined as
IE = T (IA , Eb /N0 ) (6.5)

where N0 is the power spectral density of the AWGN, and Eb is the bit energy.
Similar to Subsection 6.1.1, mutual information of the extrinsic information is also
denoted as IE = I(c, LE ), expressed as[121]

Z ∞
1 X
IE = pE (l|c = c)
2 c=−1,1 −∞
(6.6)
2pE (l|c = c)
· log2 dl, 0 ≤ IE ≤ 1
pE (l|c = −1) + pE (l|c = 1)

where pE (l|c) is the conditional pdf of LE . Since no Gaussian distribution is assumed

to pE (l|c), pE (l|c) is simulated using Monte Carlo method (histogram method) [121].
pE (l|c) is obtained into two parts corresponding to c = −1 and c = +1, respectively.
IE is then obtained using the histogram method according to Equation (6.6). In
practice, pE (l|c) of the detectors cannot be approximated as Gaussian for fading
channels [123, 129].
For the EXIT chart of trajectory, the initial value of the mutual information of
the detector starts at the intersection of the detector curve with the IAdet = 0 line.
The system mutual information moves horizontally to intersect the decoder EXIT
curve. In each iteration, the trajectory moves vertically to intersect the detector line
and then horizontally to intersect the decoder line.
 82

6.2 EXIT Chart for detection with channel esti-

mation in Fast Fading channels

In a standard EXIT chart for detector/decoder systems, two curves are plotted. One
curve is plotted for the detector with the prior mutual information on the X-axis and
output extrinsic mutual information on the Y-axis. Another curve is plotted for the
decoder, with the a priori mutual information on the Y-axis and output extrinsic
mutual information on the X-axis. The performance of the system is a function of
the distance between these two curves and where these two curves intersect [121].
For all following iterations, the standard EXIT curve drawing rules are employed
based on the EXIT curves of the detector with data-aided channel estimation and
the decoder; the trajectory moving vertically to the EXIT curve of the detector with
data-aided channel estimation and horizontally to the EXIT curve of the decoder.

6.2.1 EXIT Chart of Decoder

For the decoder, transfer function (6.5) is only dependent on the a priori input IAdec
values, that is IEdec = T dec (IAdec ). Figure 6.3 depicts EXIT charts of the decoder.
The decoder uses a convolutional code with 1/2 rate, constraint length 7 and the
generator polynomials G = [133 171] in octal form or constraint length 3 and G = [7
5], respectively.
To obtain the EXIT curve, 25600 randomly equiprobable coded bits c̃ are used
for simulations. Figure 6.3 shows that the stronger the code the steeper the transfer
function. When IAdec < 0.5, the code G = [7 5] requires a lower IAdec than the code
G = [133 171] for the same IEdec . However, when IAdec ≥ 0.5, the code G = [133
171] requires a lower IAdec than the code G = [7 5] to approach IEdec = 1, indicating
that the stronger the code, the lower error floor after convergence and the faster the
 83

convergence rate.

6.2.2 EXIT Chart of Decision-Directed Channel Estimation

with Detector

yP
ŝ

y data- Pilot-
aided aided
CE CE
Ĥ

Ldet
A Ldet
E
Detector Π−1

Figure 6.2: Model for the EXIT chart of detector

For the detector, transfer function (6.5) is not only dependent on the a priori input
IAdet values but also Eb /N0 , CSI and modulation, etc; that is IEdet = T det (IAdet , Eb /N0 , H).
The model of the detector for the EXIT chart analysis is depicted in Figure 6.2.
For the channel estimation (CE) stage, a first-order Kalman filter is used for two
different iterative CE methods discussed in Chapter 4. The initial channel estimate
is provided solely from the measurements of the pilots y P . The detected data ŝ is
then fed back into the data-aided channel estimation block for the estimation of the
channel coefficients in the next iteration.
The EXIT chart of the detector is characterized by three EXIT curves: the EXIT
curve of the detector using channel estimates obtained using pilot-signals, the EXIT
curve of the detector using data-aided channel estimation, the EXIT curve of the
detector using perfect CSI. Figure 6.3 shows these curves for a typical fast fading
channel with gray-coded quadrature phase-shift keying (QPSK) modulation of gray
mapping and the normalized Doppler frequency is set to fd T = 0.1, where T is OFDM
 84

symbol period. We choose a 4-path wide-sense stationary uncorrelated scattering

(WSSUS) channel with mean propagation powers of 8/15, 4/15, 2/15 and 1/15 for
propagation delays of 0 to 3 samples respectively.

0.9

0.8

0.7

0.6
Idet,Idec
A

0.5
E

0.4
G=[133 171]
8
0.3 ideal CSI
B
data−aided CE
0.2
true system trajectory
A
pilot−aided CE
0.1
G=[5 7]8
0
0 0.2 0.4 0.6 0.8 1
Idet
A E
,Idec

Figure 6.3: EXIT chart

25600 randomly chosen equiprobable coded bits are generated for simulations. For
the ideal CSI case, the proposed detector only needs two iterations to reach IEdec ≈ 1.
In the detector curves for imperfect CSI, the EXIT curve of the detector based on
the data-aided CE is very close to that with the ideal CSI when the perfect a priori
mutual information IAdet = 1. Therefore, the degradation based on the data-estimated
channel estimation is very low and the BER performance is close to that of the ideal
CSI when using a large interleaver.
For the EXIT chart of detector based on data-aided CE, when IAdet is low, the
iterative receiver based on data-aided CE cannot approach to a large value IEdec with
 85

the code G = [133 177]8 or G = [5 7]8 and gets stuck at the intersection A or B
respectively, shown in Figure 6.3. After this bottleneck area, the detector based on
data-estimated case can improve the performance for a large IAdet . When the EXIT
curve of detector based on data-aided CE with the code G = [133 177]8 gets stuck,
there is still a path through the bottleneck using the code G = [5 7]8 . However,
the EXIT curve of detector based on data-aided CE with the code G = [133 177]8
approaches IEdec = 1 faster than the code G = [5 7]8 , indicating that a lower error
floor is obtained after convergence. Therefore, it is better to construct a decoder to
obtain values IEdec as high as possible after several iterations.
On the other hand, when IAdet is low, the value IEdet based on data-aided CE is
lower than that based on pilot-aided CE, indicating that an unreliable detection
is obtained comparing to the pilot-estimated case. However, the EXIT curve of the
detector based on pilot case is more flat than that of detector based on data-estimated
case, it will stop improving the performance and get stuck at the intersection of the
EXIT curves between the detector and the decoder, indicating that the performance
based on pilot-estimated channel cannot converge and has a high BER.
Therefore, the EXIT chart depicted in Figure 6.3 indicates that it is better to
combine the detector based on pilot-estimated case and data-estimated case together,
which is discussed in Chapter 5. With this joint channel estimation scheme, we
combine the good starting value IEdet of the pilot-based case and the good ending
value IEdet of the data-based case. For the initial iteration, the detector based on
pilot-estimated case can overcome the poor channel estimation based on the data-
estimated case; then it is switched to data-estimated case for the successive iterations
which can approach the performance of the ideal CSI.
 86

6.2.3 EXIT Chart of Trajectory

The mutual information trajectory of iterative receiver proposed in Chapter 5 is also

shown in Figure 6.3. The detector in the initial iteration uses the pilot-aided channel
estimator so the point of IEdet for IAdet = 0 is for the pilot-aided channel estimation.
The trajectory then moves horizontally to the decoders EXIT curve.
The proposed detector will only provide good final BER performance if the EXIT
curve for the detector with data-aided channel estimation is above the curve for the
detector using pilot-aided channel estimation at the IEdec value of the intersection of
the EXIT curves of the decoder curve and the detector using pilot-aided channel esti-
mation. That is, there must be enough information provided by the initial detection
using pilot-aided channel estimation to allow the error correction code to provide
some correction capability. This condition will always be satisfied if the Eb /N0 value
is large, as higher Eb /N0 values raise the EXIT curves of all detectors without chang-
ing the EXIT curve of the decoder. The required Eb /N0 value is a function of the
modulation and error correction code used. For this reason, it recommended that
proper operation be verified with EXIT chart analysis. The proposed method will
fail at low Eb /N0 values or if the slope of the error correction codes EXIT curve is
very steep for IAdec values lower than the initial IEdet provided by the detector. In these
cases, it is possible to get some BER improvement by running the detector with pilot-
aided channel estimation for more than one iteration so the interference cancellation
can provide some information improvement.

6.3 Simulation Results

In this section, the performance of the proposed algorithm is tested in a simulated

wireless OFDM system with a simulated multipath fast-fading radio channel. We
 87

simulate a 4-path WSSUS channel with mean propagation powers of 8/15, 4/15,
2/15 and 1/15 for propagation delays of 0 to 3 samples respectively. Each path is
subject to independent Rayleigh fading generated using the method from [119] with
Jakes’ model specifying the autocorrelation function. We select an OFDM system
with N = 128 samples for each OFDM symbol. The length of CP is Ncp = N/8. The
convolutional code used in simulations is with 1/2 rate, constraint length 7 and the
generator polynomials G = [133 171] in octal form.
The channel is estimated using a first-order Kalman filter. Each transmission
frame consists of Kb = 10 OFDM blocks. A length of Kt = 10 transmission frames is
used for the interleaving block. The joint channel estimation, data detection and de-
coding method described in Chapter 5 is used. Initial channel estimation is performed
using Kp = 35 or Kp = 70 pilots per each transmission block.
We analyze EXIT charts of the decoder and the detector with different estimated
channel schemes, where gray-coded QPSK constellation is used in simulations. The
normalized Doppler frequency is set to fd T = 0.1, where T is OFDM symbol period.
We choose Q = 4 basis functions, the corresponding number of pilot blocks for each
transmission is Kp ≥ 5.
To test the predicted performance, the true trajectory is simulated by calculating
the mutual information of the extrinsic LLR for the detector and decoder after each
iteration. The EXIT charts of the iterative process are shown in Figures 6.4-6.8 for
fd T = 0.1 at Eb /N0 = 7dB and Eb /N0 = 8dB.
Figures 6.4-6.6 depict the EXIT charts with Kt = 10 for each interleaving block
at Eb /N0 = 7dB for different number of pilot blocks. Figure 6.4 shows that the true
trajectory needs only three iterations to reach IEdec ≈ 1 when Kp = 10. However, for a
reduced value Kp = 5, Figure 6.6 shows that the trajectory gets stuck at IEdec ≈ 0.96
after three iterations. The receiver’s EXIT trajectory depicted in Figure 6.6 doesn’t
 88

0.9

0.8

0.7

0.6
Idet, Idec
A

0.5
E

0.4
detector with ideal CSI
0.3
detector with data−aided CE
0.2 detector with pilot−aided CE
trajectory
0.1 decoder G=[133 171]8

0
0 0.2 0.4 0.6 0.8 1
Idet
A
, Idec
E

Figure 6.4: EXIT chart using Kp = 10 and Kt = 10 for fd T = 0.1 at Eb /No = 7dB

match the transfer characteristics predicted by the EXIT chart due to dependencies
of the LLR between the detector and the decoder. Increasing the interleaver length
to Kt = 100 decreases these dependencies and the simulated trajectory matches the
transfer characteristics as can be seen in Figure 6.6. Enlarging Kt increases the com-
putational complexity and memory requirements of the channel estimation. There-
fore, finding the optimal number of pilot blocks is a tradeoff between the performance
of the error floor and the complexity cost at low SNRs. At this system, we choose
Kp = 10 for each transmission frame and Kt = 10 since the increase in pilot power
for Kp = 10 over Kp = 5 is only a small percentage of total power and the memory
requirement in the receiver is reduced by an order of magnitude. It was noted that
when Eb /N0 > 7dB, Kp = 5 provided good BER performance for Kt = 10.
The EXIT curves using different pilot schemes are shown in Figures 6.7 and 6.8
 89

0.9

0.8

0.7

0.6
Idet, Idec
A

0.5
E

0.4

0.3 detector with ideal CSI

detector with data−aided CE
0.2 detector with pilot−aided CE
trajectory
0.1 decoder G=[133 171]8

0
0 0.2 0.4 0.6 0.8 1
Idet
A
, Idec
E

Figure 6.5: EXIT chart using Kp = 5 and Kt = 10 for fd T = 0.1 at Eb /N0 = 7dB

with Kt = 10 at Eb /No = 8dB. The simulated trajectories are shown that after two
iterations based on data-estimated case, the trajectories are close to the EXIT curves
of the ideal CSI for different pilot schemes. Therefore, the whole system using Kp = 5
for initial detection has the same error floor as that using Kp = 10 at a high SNR.
The EXIT chart predicts that using Kp = 5 is enough for a high SNR.
From the analysis above, the EXIT chart predicts that the detector needs two
iterations after the initial detection and the detection based on data-aided channel
estimation comes very close to the case of perfect CSI. Therefore, the degradation of
the system performance due to channel estimation is very low. In order to make the
system converge at low SNR and have a low complexity, we Kt = 10 and Kp = 10 for
the whole iterative system.
 90

0.9

0.8

0.7

0.6
Idet, Idec
A

0.5
E

0.4
detector with ideal CSI
0.3 detector with data−aided CE
detector with pilot−aided CE
0.2
trajectory
decoder G=[133 171]8
0.1

0
0 0.2 0.4 0.6 0.8 1
Idet
A
, Idec
E

Figure 6.6: EXIT chart using Kp = 5 and Kt = 100 for fd = 0.1 at Eb /N0 = 7dB

6.4 Conclusion

The EXIT chart technique predicts the convergence behavior of the proposed iterative
scheme. EXIT chart analysis can be used to assist with the selection of pilot signals,
modulation type, and error correction code as well as the required Eb /N0 value for
the algorithm to converge.
 91

0.9

0.8

0.7

0.6
Idet, Idec
A

0.5
E

0.4

detector with ideal CSI

0.3
detector with data−aided CE
0.2 detector with pilot−aided CE
trajectory
0.1 decoder G=[133 171]8

0
0 0.2 0.4 0.6 0.8 1
Idet, Idec
A E

Figure 6.7: EXIT chart using Kp = 5 and Kt = 10 for fd T = 0.1 at Eb /N0 = 8dB
 92

0.9

0.8

0.7

0.6
Idet, Idec
A

0.5
E

0.4

detector with ideal CSI

0.3
detector with data−aided CE
0.2 detector with pilot−aided CE
trajectory
0.1 decoder G=[133 171]8

0
0 0.2 0.4 0.6 0.8 1
Idet, Idec
A E

Figure 6.8: EXIT chart using Kp = 10 and Kt = 10 for fd T = 0.1 at Eb /No = 7dB
 93

Chapter 7

Conclusions and Future Work

7.1 Summary

Reliable estimate of the channel state information (CSI) for orthogonal frequency
division multiplexing (OFDM) systems in doubly selective channels has been studied
in this thesis. The algorithms developed in this thesis have improved the performance
of the whole system requiring only low ratios of pilot to data for excellent performance
in fast fading channels.
The main contributions of the thesis are :

• To enable low cost detection and decoding of OFDM in fast fading, a new
structure for channel processing has been used at the receiver in coded OFDM
systems, where the OFDM block has been used as the basis for data detection;
the transmission block which contains several OFDM blocks has been used for
channel estimation; and the interleaving block which contains multiple trans-
mission blocks has been used for decoding. To capture the channel dynamics,
a novel multivariate autoregressive (AR) process over transmission blocks has
been developed to model the time evolution of the fast fading channel with the
 94

help of a basis expansion model (BEM). To develop the iterative scheme at the
receiver, three measurement models have been discussed in this thesis, i.e., the
measurement model for data detection based on the estimated or known CSI;
the measurement model for channel estimation based on pilot symbols and the
measurement model for channel estimation based on the detected data.

• A novel Kalman smoothing algorithm based on a second-order dynamic model

has been exploited, where the mean square error (MSE) of the channel is near
to that of the optimal Wiener filter and the computational complexity is same
as the previous algorithms. Moreover, the proposed method can improve the bit
error rate (BER) performance of wireless communication systems. Also, this
new technique is robust when the Doppler frequency or signal-to-noise ratio
(SNR) is varied from the value used to design the channel estimation filter.

• A novel joint channel estimation data detection algorithm has been developed
for fast fading channels with a small number of pilots and low pilot power to
achieve the BER performance close to when the CSI is known perfectly. The
new channel estimation symbol detection technique is robust to variations of
the radio channel from the design values and applicable to multiple modulation
and coding types.

• A useful analysis technique has been investigated to analyze the convergence

behavior of the new algorithm and select the modulation, pilot density, and
error correction code for good system performance for a given power level by
the extrinsic information transfer (EXIT) chart technique.
 95

7.2 Future Work

Future work should be done to integrate efficient data detection algorithm into joint
channel estimation and data detection techniques for coded OFDM systems in fast
fading channels. In particular, the new data detection method should be investigated
to see if alternate methods can be found to improve data detection without greatly
increasing the complexity.
It will also be interesting to investigate the performance of channel estimation
when used for higher Doppler frequencies. Currently, each transmission block con-
tains 10 OFDM symbol blocks. When Doppler frequency increases, the structure
of time blocks for channel processing will need to be adjusted to maintain efficient
performance for low pilot to data ratios. The required form of the signalling will be
investigated in future research.
In addition, the multivariate AR model can be applied to other systems such as
single-carrier (SC) systems or multicarrier-code division multiple access (MC-CDMA)
systems for fast fading channels.
Future work in the context of channel estimation and symbol detection as well as
decoding relies on EXIT chart analysis to achieve lower BER performance over fast
fading channels at a low possible complexity.
Appendix A

Basic Time Domain Channel

Impulse Response Description

In wireless communications, the radio channel is often assumed to be wide-sense

stationary uncorrelated scattering (WSSUS), where the time-variant impulse response
h(t, τ ) is modelled as a zero-mean complex Gaussian process for a fixed delay τ , and
uncorrelated for different delays [131, 2]. In such a case, the autocorrelation function
of h(t, τ ) is given by

E[h(t, τ1 )h∗ (t + ∆t, τ2 )] = rh (∆t, τ1 )δ(τ1 − τ2 ) (A.1)

If the arrival angle of the received waveform is a uniformly distributed random variable
over [−π, π], rh (∆t; τ ) can be separated in time and delay [104], i.e.,

rh (∆t; τ ) = rt (∆t)rτ (τ ) (A.2)

 97

where rτ (τ ) is the multipath intensity profile, and the Fourier transform of the auto-
correlation function rt (∆t) is

rt (∆t) = J0 (2πfd ∆t) (A.3)

where J0 (· ) is the zeroth-order Bessel function of the first kind and fd = vfc /c is
the maximum Doppler frequency, where v is the vehicle speed in meters per second
(m/s), fc is the carrier frequency, and c is the speed of light (3 × 108 m/s) [2]. In the
case of outdoor radio channels, a vertical receive antenna with constant azimuthal
gain, the angles of arrival of the radio waves are uniformly distributed over (−π, π),
the normalized Jakes’ Doppler power spectrum, SH (f ), is given by [99]




 1
πfd
√ 1
, if |f | ≤ fd
1−(f /fd )2
SH (f ) = (A.4)


0, otherwise

Suppose the multipath delay spread is τmax , the tapped delay line model for the
channel can be truncated at L = bτmax W c taps, where W is bandwidth, b·c is the
integer floor function, the impulse response is expressed as

L
X
h(t, τ ) = h(t, τl )δ(τ − τl ) (A.5)
l=0

where τl = l/W and τ0 < τ1 < · · · < τL for l = 0, 1, ..., L. The h(t, τl )’s represent the
tap weights corresponding to the L different delays τl . δ(t) is Dirac’s delta function.
Since h(t, τl ) is characterized as WSSUS, h(t, τl )’s are mutually uncorrelated with
variance σl2 = rτ (τl ). In such a case, when the h(t, τl )’s are Gaussian random process,
they are statistically independent. Usually, the total variances of the channel taps
P
are assumed to be normalized to unity, i.e., Ll=0 σl2 = 1. In deed, power is often
 98

normalized in communication systems [99]. Moreover, if the envelope |h(t; τ )| at time

t is Rayleigh-distributed, the channel is a Rayleigh fading channel.
Appendix B

Basis Expansion Model

This appendix provides a summary of basis expansion model (BEM). Beginning with
the papers by Giannakis et al. [132], Borah et al. [133] and Martone [134], BEMs have
been widely developed to model doubly selective channels in wireless communication,
where the time variant impulse responses h(n, l) are expressed by the sum of time-
varying basis functions, weighted by time-invariant channel coefficients within a fixed
time duration, which is called the transmission block. A general BEM expansion of
h(n, l) can be written as
Q
X
h(n, l) = En (q)xl (q) (B.1)
q=0

where En (q) gives the value at time n of the qth basis function and xl (q) is the
coefficient for path l corresponding to the qth basis function. Due to the bandlimited
nature of Equation (A.4), Q is BEM order denoted by Q = 2dfd M Ts e, where Ts is
the sampling period and M is the number of samples in each transmission block [1],
and Q + 1 is the number of basis coefficients.
It is convenient to express the BEM in matrix form. Hence, we define the basis
 100

function matrix as  
 E0 (0) · · · E0 (Q) 
 .. ... .. 
E=
 . . 
 (B.2)
 
EM −1 (0) · · · EM −1 (Q)

Let xl denote the coefficient vector for the lth path for the current transmission block,
i.e.,  
 xl (0) 
 . 
xl = 
 . 
.  (B.3)
 
xl (Q)

Thus, the channel gain in each transmission block for the lth path is expressed as

hl = Exl . (B.4)

For a given basis function E, xl vector gives all the necessary information about the
path l for a given block. Different BEM matrices have been performed on the criterion
for optimizing the BEM with various robustness properties [72, 110, 69]. Below, we
discuss these matrices.

B.1 Complex Exponential Basis Expansion Model

The complex exponential basis expansion model (CE-BEM) or Fourier BEM comes
from [132]. As Q increases, the CE-BEM approximates the random fading coefficient
model in mobile communication channels [132].
In [1], the Fourier basis functions over a window of M samples are expressed as

En (q) = ejωq n , q ∈ [0, Q] (B.5)

 101

where ωq = 2π(q −Q/2)/M . In this case, a discrete-time baseband equivalent matrix-

vector channel model is given as

 
 1 ··· 1 
 
 e −j 2πQ/2
··· j
e M
2πQ/2

 M

E= . .  (B.6)
 .
. · · · .
. 
 
 
2πQ/2(M −1) 2πQ/2(M −1)
e−j M · · · ej M

This model approximates doubly selective channels as a linear combination of a

finite number of complex exponentials. Since this sampling period of the Doppler
spectrum is equal to the symbol period Ts , sampling signal over a finite transmis-
sion block is equivalent to the signal truncated by a rectangular window. Then the
truncated spectrum by using the discrete Fourier transform (DFT) is spread out over
the whole frequency range, which is so-called spectral leakage [98]. As a result, the
modelling error between the true channel and the CE-BEM is quite high.
The basic limitation of the CE-BEM to cope with modelling error has motivated
a considerable amount of research into other BEMs. As an alternative, we may over-
sample the Doppler spectrum, which is discussed in [110], where ωq in equation(B.5)
becomes ωq = 2π(q − Q/2)/(kM ) with Q = 2dfd kM Ts e and k = 2. In fact, this
choice of Q may be increased compared to critical sampling case.
Instead of rectangular window, other windowing functions can be used in order
to reduce leakage. However, the mainlobe levels of the general windowing functions
such as Hamming, Hanning, etc. are wider than the rectangular window, but the
sidelobe levels are lower [98]. The optimal windowing function is the time limited
sequence whose energy is most concentrated in a finite frequency interval, and hence,
is related to the zeroth discrete prolate spheroidal (DPS) sequence [135, 136, 137],
which is described in detail in the following section.
 102

B.2 Discrete Prolate Spheroidal Basis Expansion

Model

Another approach is to use prolate spheroidal basis functions, which come from the
the prolate spheroidal wave functions (PSWFs). The PSWFs provide an orthogonal
set of the bandlimited real functions, with each concentration being each eigenvalue,
and are most concentrated in the bandwidth [138, 139]. In the discrete time, these
basis functions are the DPS sequences [72]. For a given transmission block with M
samples and the Doppler frequency fd , the CSI, h(n, l) of each path l, can be stated
in terms of PSWFs, DPS sequences and their associated eigenvalues λ’s.
The DPS-BEM functions for channel modelling are bandlimited to the Doppler
frequency and are concentrated on the time interval [0, M − 1]. Since the extension
of the finite sequence for the time duration has least energy [138], it can overcome
the drawback of windowing in the CE-BEM.
For a given transmission block with M samples and the Doppler frequency fd , the
DPS sequences em (q)’s can be concentrated on M samples and bandlimited to fd Ts ,
where Ts is the sampling period, [138]

M
X −1
sin[2πfd T (i − m)]
em (q) = λq ei (q), i ∈ [−∞, ∞] (B.7)
m=0
π(i − m)

where λq is the qth eigenvalue, which measures the energy concentration of the qth
DPS sequence.
The DPS sequences are orthogonal on both the infinite interval and the finite
 103

interval [0, M − 1], that is [138]

M
X −1 ∞
X
em (i)em (j) = λi em (i)em (j)
i=0 −∞ (B.8)
= δij , i, j ∈ [0, M − 1]

The eigenvalue λi is close to 1 for i ≤ d2fd Ts M e while λi is close to 0 for i > d2fd Ts M e
[138]. Hence, for a large M , a small number of d2fd Ts M e is sufficient to describe the
DPS sequence [138].
Let A be the M × M matrix with the element

sin[2πfd Ts (m − n)]
Am,n = , m, n ∈ [0, ..., M − 1]. (B.9)
π(m − n)

Equation (B.7) is rewritten as

Ae(q) = λq e(q), q ∈ [0, ..., M − 1]. (B.10)

Therefore, the DPS sequences e(q) is the eigenvector of A corresponding to the

eigenvalue λq .
In general, for a given interval M and Doppler frequency fd , define the basis
expansion number Q = d2fd Ts kM e, where k = 1 or k = 2 [72, 67]. In practice, when
the true Doppler frequency is unavailable, a conservative design would be to set the
assumed Doppler frequency the highest expected level.

B.3 Karhunen-Loève Basis Expansion Model

In the MSE criterion, Karhunen-Loève (KL) expansion model, which was proposed in
[140, 141, 142, 69], is optimal in minimizing the channel modeling MSE. For a given
 104

length M , the correlation matrix of hl from Equation (B.4) is

Rl = E[hl hH H H H
l ] = E[Exl (Exl ) ] = EE[xl xl ]E , for l ∈ [0, L] (B.11)

where Rl is the size of M by M . Using singular value decomposition (SVD), Rl is

expressed as
Rl = Ul Vl UlH (B.12)

where Ul is a unitary matrix with the size of M by M , Vl is a diagonal matrix with

M nonnegative real values called eigenvalues in non-increasing form.
Let us choose Q most significant eigenvalues. Corresponding to Vl with Q diagonal
entries, Ul becomes M by Q. In order to minimize the channel modeling error, KL
basis expansion functions E choose Ul with the size of M by Q.
Based on Jakes’ model, the correlation function of the channel impulse response
h(n, l) is expressed as

Rl (m, n) = E[h(m, l)hH (n, l)] = σl2 J0 (2πfd Ts |m − n|), for m, n ∈ [0, M ] (B.13)

where J0 is the zeroth order Bessel function [104], σl2 is the variance of the lth path.
Let Rl = σl2 R, where R is the normalized correlation matrix. Similarly, We apply
SVD on RU V U H , where U = Ul , Vl = σl2 V . Therefore, the KL basis functions are
chosen as E = U .
Appendix C

Whittle-Wiggins-Robinson
Algorithm

In this appendix, we provide the calculation of the values for a multivariate au-
toregressive (AR) process based on Whittle-Wiggins-Robinson Algorithm (WWRA)
[102, 103].
Consider an unknown vector of size q for time index n expressed as

· ¸T
x[n] = x1 [n] x2 [n] · · · xq [n] (C.1)

with the covariance matrix defined as

Rx [j] = E{x[n + j]x[n]H }

 
Rx1 x1 [j] Rx1 x2 [j] ··· Rx1 xq [j]
 
Rx x [j] Rx x [j] ··· Rx2 xq [j] (C.2)
 2 1 2 2 
= . .. .. 
 .. .. 
 . . . 
 
Rxq x1 [j] Rxq x2 [j] ··· Rxq xq [j]
 106

A Kth-order multivariate AR process is modelled as

K
X
x[n] = − A[i]x[n − i] + ef [n] (C.3)
i=1

where A[i], i = 1, 2, ..., K are q by q matrices, and ef [n] is an additive white

Gaussian noise (AWGN) vector process with zero mean and a covariance of Qf =
E{ef [n]ef [n]H }. High-order AR model can provide a more accurate model with high
complexity. However, low-order AR model is more practically used to approximate
discrete-time random process [75, 143].
Define
· ¸T
x[n] = x [n] · · · x (n − K + 1)
T T (C.4)

The covariance matrix of x[n] is a positive semi-definite block matrix defined as

Rk = E{x[n]x[n]H }
 
H
 E{x[n]x[n] } E{x[n]x[n − 1]H } ··· E{x[n]x[n − K]H }

 
 E{x[n − 1]x[n]H } E{x[n − 1]x[n − 1]H } · · · E{x[n − 1]x[n − K] }  H
 
= .. .. . 
 .. .. 
 . . . 
 
E{x[n − K]x[n]H } E{x[n − K]x[n − 1]H } · · · E{x[n − K]x[n − K]H }
 
R
 x [0] R x [1] · · · R x [K] 
 
 Rx [−1] Rx [0] · · · Rx [1 − K]
 
= . . . 
 . . ... . 
 . . . 
 
Rx [−K] Rx [1 − K] · · · Rx [0]

(C.5)

where Rk has both a Hermitian and a block-Toeplitz structure, that is Rx [n] =

Rx [−n]H , but the Rx [n] are not Hermitian (Rx [n] 6= RxH [n]).
 107

Equation (C.3) is then expressed as

Ak [n]x[n] = ef [n] (C.6)

where
· ¸
Ak = I A[1] · · · A[K] , (C.7)

The values of the AR coefficient matrices satisfy the multichannel Yule-Walker

equation [100, 101]:
· ¸
Ak R k = Q f
0 ··· 0 (C.8)

This is a multivariate forward linear prediction filter with order K. The multivariate
backward linear filter process is expressed as

B k [n]x[n] = eb [n] (C.9)

where
· ¸
B k = B[1] · · · B[K] I , (C.10)

The multichannel Yule-Walker equations for the backward coefficients of a stationary

AR process are written by

· ¸
B k Rk = 0 0 · · · Q b (C.11)

where Qb = E{eb [n]eb [n]H } is the covariance matrix of the noise process for the
backward AR process.
The values for the multivariate AR model can be solved efficiently with the
WWRA [102, 103]. The WWRA is the extension to the multivariate case of the
Levinson-Durbin algorithm. The forward and backward coefficient matrices for the
 108

(k + 1)th order are calculated by

Ak+1 (k + 1) = −∆k+1 (Qbk )−1 (C.12)

Bk+1 (k + 1) = −∇k+1 (Qfk )−1 (C.13)

where
K
X
∆k+1 = Ak (i)R(k + 1 − i) (C.14)
i=0

∇k+1 = ∆H
k+1 (C.15)

The update recursion are

· ¸ · ¸
Ak+1 = Ak 0 + Ak+1 (k + 1) 0 Bk (C.16)

· ¸ · ¸
Bk+1 = 0 Bk + Bk+1 (k + 1) Ak 0 (C.17)

The update error covariance matrices are

Qfk+1 = Qfk + Ak+1 (k + 1)∇k+1 (C.18)

Qbk+1 = Qbk + Bk+1 (k + 1)∆k+1 (C.19)

The initial conditions are

Qf0 = Qb0 = R(0) (C.20)

A0 = B0 = I (C.21)
 109

Appendix D

Least Squares and Linear

Minimum Mean Square Error
Estimation

Consider a linear system mode as

y = Cx + v (D.1)

where y is a known vector, x is an unknown vector, C is a known transition matrix,

and v is an additive white Gaussian noise (AWGN) with variance σv2 .
Least squares (LS) estimation is [51]

x̂ = (C T C)−1 C T x (D.2)

Linear minimum mean square error (LMMSE) approach is optimal in minimizing

the MSE of the channel estimates in the presence of AWGN. LMMSE estimation
exploits the channel statistics and the signal-to-noise ratio (SNR) values. LMMSE of
 110

the variable x is given by [1]

1 −1 1
x̂ = [R x + (C)H C]−1 (C)H y (D.3)
σv2 σv2

where Rx = E[xxH ], and σv2 are known to the receiver.

Appendix E

Basic Kalman Filter Equations

Consider a state space model expressed as

x[n] = Φx[n − 1] + w[n − 1] (E.1)

y[n] = G[n]x[n] + v[n] (E.2)

where x[n] is the state vector at time n with the covariance matrix P [n], y[n] is the
measurement vector; Φ and G[n] are known transition matrix; w[n] is the state noise
with zero mean and covariance Q[n], and v[n] is the measurement noise with zero
mean and covariance R[n]. The purpose of the Kalman filter is to use measurements
and the states to estimate the values which is better than the estimate obtained by
using measurements alone.
Let us choose the initial values at time n = 0 for Equation (E.1). From the practi-
cal point of view, the true value of the random constant has a standard normal prob-
ability distribution, so we assume the initial state x[0] = 0. Similarly, we choose the
initial value of the covariance of the state vector, P [0] = E([x[0]−x[0]][x[0]−x[0]]H ]),
where x[0] = E[x[0]]. Given the initial values x[0] and P [0], basic Kalman filter equa-
tions for one iteration are [100]:
 112

• Predict:

x̂ [n|n − 1] = Φx̂ [n − 1|n − 1] (E.3)

P [n|n − 1] = ΦP [n − 1|n − 1] ΦH + Q (E.4)

• Update:

ŷ [n|n − 1] = G[n]x̂ [n|n − 1] (E.5)

M [n|n − 1] = G[n]P [n|n − 1] GH [n] + R [n] (E.6)

z [n] = y [n] − ŷ [n|n − 1] (E.7)

K [n] = P [n|n − 1] GH [n] (M [n|n − 1])−1 (E.8)

x̂ [n|n] = x̂ [n|n − 1] + K [n] z [n] (E.9)

P [n|n] = (I − K [n] G[n]) P [n|n − 1] (E.10)

 113

Bibliography

[1] X. Ma, G. B. Giannakis, and S. Ohno, “Optimal training for block transmission
over doubly selective wireless fading channels,” IEEE Trans. Signal Processing,
vol. 51, no. 5, pp. 1351–1366, May 2003.

[2] J. G. Proakis, Digital Communications, 3rd ed. New York: McGraw-Hill, 1995.

[3] S. Haykin and M. Moher, Modern Wireless Communication. Upper Saddle

River, N.J.: Prentice Hall, 2005.

[4] M. K. Ozdemir and H. Arslan, “Channel estimation for wireless OFDM sys-
tems,” in IEEE Commun. Surveys Tutorials, vol. 9, no. 2, 2nd Quarter 2007,
pp. 18–48.

[5] X. Cai and G. B. Giannakis, “Bounding performance and suppressing intercar-

rier interference in wireless mobile OFDM,” IEEE Trans. Commun., vol. 51,
no. 12, pp. 2047–2056, Dec. 2003.

[6] P. Schniter, “Low-complexity equalization of OFDM in doubly selective chan-

nels,” IEEE Trans. Signal Processing, vol. 52, no. 4, pp. 1002–1011, April 2004.

[7] J. K. Cavers, “An analysis of pilot symbol assisted modulation for Rayleigh
fading channels,” IEEE Trans. Veh. Technol., vol. 40, pp. 686–693, Nov. 1991.
 114

[8] L. Tong, B. M. Sadler, and M. Dong, “Pilot-assisted wireless transmissions:

general model, design criteria, and signal processing,” IEEE Signal Processing
Magazine, vol. 21, no. 6, pp. 12–25, Nov. 2004.

[9] C. Tepedelenlioğlu and G. B. Giannakis, “Transmitter redundancy for blind

estimation and equalization of time- and frequency-selective channels,” IEEE
Trans. Signal Processing, vol. 48, no. 7, pp. 2029–2043, July 2000.

[10] H. Bölcskei, J. R. W. Heath, and A. J. Paulraj, “Blind channel identification

and equalization in OFDM-based multiantenna ssytems,” IEEE Trans. Signal
Processing, vol. 50, no. 1, pp. 99–109, Jan. 2002.

[11] T. Cui and C. Tellambura, “Blind receiver design for OFDM systems over
doubly selective channels,” IEEE Trans. Commun., vol. 55, no. 5, pp. 906–917,
May 2007.

[12] S. Haykin, Adaptive Filter Theory, 4th ed. Upper Saddle River, N.J.: Prentice
Hall, 2002.

[13] M. McGuire and P. Wan, “Analysis of joint channel estimation and equalization
using a Kalman filter,” in Proc. CCECE’06 conf., 2006, pp. 912–915.

[14] S. B. Weinstein, “The history of orthogonal frequency-division multiplexing,”

IEEE Commun. Mag., vol. 47, no. 11, pp. 26–35, Nov. 2009.

[15] H. Sari, G. Karam, and I. Jeanclaude, “Transmission techniques for digital

terrestrial TV broadcasting,” IEEE Commun. Mag., vol. 33, no. 2, pp. 100–
109, Feb. 1995.

[16] “Part 11: Wireless LAN medium access control (MAC) and physical layer
(PHY) specifications: High speed physical layer in the 5 GHz brand,” IEEE
Std 802.11, 1999 Edition (R2003).
 115

[17] “Broadband radio access networks (BRAN); HIPERLAN type 2; physical

(PHY) layer, HIPERLAN II,” 2001.

[18] “Local and metropolitan area networks part 16: Air interface for broadband
wireless access systems,” IEEE Std 802.16-2004.

[19] R. W. Chang, “Synthesis of band limited orthogonal signals for multichannel

data transmission,” Bell Syst. Tech. J., vol. 45, pp. 1775–1797, Dec. 1966.

[20] J. A. C. Bingham, “Multicarrier modulation for data transmission: An idea

whose time has come,” IEEE Commun. Mag., vol. 28, pp. 5–14, May 1990.

[21] S. Weinstein and P. Ebert, “Data transmission by frequency-division multiplex-

ing using the discrete Fourier transform,” no. 5, pp. 628–634, 1971.

[22] L. J. C. Jr., “Analysis and simulation of a digital mobile channel using orthogo-
nal frequency division multiplexing,” IEEE Commun. Mag., no. 7, pp. 665–675,
1985.

[23] B. Muquest, Z. Wang, G. B. Giannakis, M. de Courville, and P. Duhamel,

“Cyclic prefixing or zeros padding for wireless multicarrier transmissions?”
IEEE Trans. Commun., vol. 50, no. 12, pp. 2136–2148, Dec. 2002.

[24] P. H. Moose, “A technique for orthogonal frequency division multiplexing fre-

quency offset correction,” IEEE Trans. Commun., vol. 42, no. 10, pp. 2908–
2914, Oct. 1994.

[25] T. M. Schmidl and D. C. Cox, “Robust frequency and timing synchronization

for OFDM,” IEEE Trans. Commun., vol. 45, no. 12, pp. 1613–1621, Dec. 1997.
 116

[26] Y. Li, H. Minn, N. Al-Dhahir, and A. R. Calderbank, “Pilot designs for consis-
tent frequency-offset estimation in OFDM systems,” IEEE Trans. Commun.,
vol. 55, no. 5, pp. 864–877, May 2007.

[27] T. Keller and L. Hanzo, “Adaptive multicarrier modulation: a convenient frame-

work for time-frequency processing in wireless communications,” Pro.IEEE,
vol. 88, no. 5, pp. 611–640, May 2000.

[28] T. Hwang, C. Yang, G. Wu, S. Li, and G. Y. Li, “OFDM and its wireless
applications: A survey,” IEEE Trans. Veh. Technol., vol. 58, no. 4, pp. 1673–
1694, May 2009.

[29] Z. Wang and G. B. Giannakis, “Complex-field coding for OFDM over fading
wireless channels,” IEEE Trans. Inform. Theory, vol. 49, no. 3, pp. 707–720,
Mar. 2003.

[30] X. Cai, S. Zhou, and G. B. Giannakis, “Group-orthogonal multicarrier CDMA,”

IEEE Trans. Commun., vol. 52, no. 1, pp. 90–99, Jan. 2004.

[31] M. McGuire and M. Sima, “Parallel detection of MC-CDMA in fast fading,”

IEEE Trans. Wireless Commun., vol. 8, no. 12, pp. 5916–5927, Dec. 2009.

[32] P. Hoeher, S. Kaiser, and P. Robertson, “Two-dimensional pilot-symbol-aided

channel estimation by Wiener filering,” in Proc. ICASSP97, Munich, German,
vol. 3, Apr. 1997, pp. 1845–1848.

[33] Y. Shen and E. Martinez, “Channel estimation in OFDM systems,” Freescale

Semiconductor, pp. 1–15, 2006.

[34] Z. Wang and G. B. Giannakis, “Wireless multicarrier transmissions: Where

Fourier meets Shannon,” IEEE Signal Processing Mag., vol. 17, pp. 29–48,
May 2000.
 117

[35] N. Benvenuto, R. Dinis, D. Falconer, and S. Tomasin, “Single carrier modulation

with nonlinear frequency domain equalization: an idea whose time has come -
again,” Proceedings of the IEEE, vol. 98, no. 1, pp. 68–96, Jan. 2010.

[36] P. Hoeher and F. Tufvesson, “Channel estimation with superimposed pilot se-
quence,” in Proc. Globalcom99, Rio de Janeiro, Brazil, vol. 4, 1999, pp. 2162–
2166.

[37] G. T. Zhou, M. Viberg, and T. McKelvey, “A first-order statistical method for

channel estimation,” IEEE Signal Processing Lett., vol. 10, no. 6, pp. 57–60,
Mar. 2003.

[38] M. Ghogho, D. McLernon, E. Alameda-Hernandez, and A. Swami, “Channel

estimation and symbol detection for block transmission using data-dependent
superimposed training,” IEEE Signal Processing Lett., vol. 12, no. 3, pp. 226–
229, Mar. 2005.

[39] J. K. Tugnait and W. Luo, “On channel estimation using superimposed training
and first-order statistics,” IEEE Commun. Lett., vol. 7, no. 9, pp. 413–415, 2003.

[40] J. K. Tugnait, X. Meng, and S. He, “Doubly selective channel estimation using
superimposed training and exponential bases models,” EURASIP Journal on
Applied Signal Processing, pp. 1–11, 2006.

[41] S. He and J. K. Tugnait, “On doubly slective channels estimation using su-
perimposed training and discrete prolate shperoidal sequences,” IEEE Trans.
Signal Processing, vol. 56, no. 7, pp. 3214–3228, July 2008.

[42] W.-C. Huang, C.-P. Li, and H.-J. Li, “On the power allocation and system
capacity of OFDM sytems using superimposed training schemes,” IEEE Trans.
Veh. Technol., vol. 58, no. 4, pp. 1731–1740, May 2009.
 118

[43] T. S. Rappaport, Wireless Communications: Principles and Practice, 2nd ed.

Upper Saddle River, N.J: Prentice Hall PTR, 2002.

[44] R. Negi and J. Cioffi, “Pilot tone selection for channel estimation in a mobile
OFDM system,” IEEE Trans. Consum. Electron., vol. 44, no. 3, pp. 1122–1128,
Aug. 1998.

[45] S. Adireddy, L. Tong, and H. Viswanathan, “Optimal placement of training

for frequency-selective block-fading channels,” IEEE Trans. Inform. Theory,
vol. 48, no. 8, pp. 2338–2353, Aug. 2002.

[46] M. Dong and L. Tong, “Optimal design and placement of pilot symbols for
channel estimation,” IEEE Trans. Signal Processing, vol. 50, no. 12, pp. 3055–
3069, Dec. 2002.

[47] S. Ohno and G. B. Giannakis, “Capacity maximizing MMSE-optimal pilots for

wireless OFDM over frequency-selective block Rayleigh-fading channels,” IEEE
Trans. Inf. Theory, vol. 50, no. 9, pp. 2138–2145, Sept. 2004.

[48] X. Cai and G. B. Giannakis, “Error probability minimizing pilots for OFDM
with M-PSK modulation over Rayleigh-fading channels,” IEEE Trans. Veh.
Technol., vol. 53, no. 1, pp. 146–155, Jan. 2004.

[49] M. Dong, L. Tong, and B. M. Sadler, “Optimal insertion of pilot symbols for
transmissions over time-varying flat fading channels,” IEEE Trans. Signal Pro-
cessing, vol. 52, no. 5, pp. 1403–1418, May 2004.

[50] I. Barhumi, G. Leus, and M. Moonen, “Optimal training design for MIMO
OFDM systems in mobile wireless channels,” IEEE Trans. Signal Processing,
vol. 51, no. 6, pp. 1615–1634, June 2003.
 119

[51] S. M. Kay, Fundamentals of statistical signal processing: Estimation Theory.

Englewood Cliffs, N.J.: Prentice-Hall, 1993.

[52] J.-J. van de Beek, O. Edfors, M. Sandell, S. K. Wilson, and P. O. Börjesson,

“On channel estimation in OFDM systems,” in Proc. VTC, Sep. 1995.

[53] O. Edfors, M. Sandell, J.-J. van de Beek, S. K. Wilson, and P. O. Börjesson,

“OFDM channel estimation by singular value decomposition,” IEEE Trans.
Commun., vol. 46, no. 7, pp. 931–939, 1998.

[54] M. R. Raghavendra and K. Giridhar, “Improving chcannel estimation in OFDM

systems for sparse multipath channels,” IEEE Signal Processing Lett., vol. 12,
no. 1, pp. 52–55, Jan. 2005.

[55] W. G. Jeon, K. H. Paik, and Y. S. Cho, “An efficient channel estimation tech-
nique for OFDM systems with tranmitter diversity,” in Proc. IEEE PIMRC
2000, Sept. 2000, pp. 1246–1250.

[56] M. Morelli and U. Mengali, “A comparison of pilot-aided chanel estimation

methods for OFDM systems,” IEEE Trans. Signal Processing, vol. 49, no. 12,
pp. 3065–3073, Dec. 2001.

[57] S. Coleri, M. Ergen, A. Puri, and A. Bahai, “Channel estimation techniques

based on pilot arrangement in OFDM systems,” IEEE Trans. Broadcast.,
vol. 48, no. 3, pp. 223–229, Sep. 2002.

[58] Y. Zhao and A. Huang, “A novel channel estimation method for OFDM mobile
communication systems based on pilot signals and transform-domain process-
ing,” in Proc. IEEE VTC, vol. 3, May 1997, pp. 2089–2093.
 120

[59] X. Wang and K. J. R. Liu, “An adaptive channel estimation algorithm using
time-frequency polynomial model for OFDM with fading multipath channels,”
EURASIP Journal on Applied Signal Processing, no. 8, pp. 818–830, Aug. 2002.

[60] Y. Li, L. J. Cimini, Jr., and N. R. Sollenberger, “Robust channel estimation for
OFDM systems with rapid dispersive fading channels,” IEEE Trans. Commun.,
vol. 46, no. 7, July 1998.

[61] Y. Li, “Pilot-symbol-aided channel estimation for OFDM in wireless systems,”

IEEE Trans. Veh. Technol., vol. 49, no. 4, pp. 1207–1215, July 2000.

[62] S. Chen and T. Yao, “Intercarrier interference suppression and channel estima-
tion for OFDM systems in time-varying frequency-selective fading channels,”
IEEE Trans. Consumer Electron., vol. 50, no. 2, pp. 429–435, May 2004.

[63] Y. Mostofi and D. C. Cox, “ICI mitigation for pilot-aided OFDM mobile sys-
tems,” IEEE Trans. Wireless Commun., vol. 4, no. 2, pp. 765–774, 2005.

[64] T. Y. Al-Naffouri, K. M. Z. Islam, N. Al-Dhahir, and S. Lu, “A model reduction

approach for OFDM channel estimation under high mobility conditions,” IEEE
Trans. Signal Processing, vol. 58, no. 4, pp. 2181–2193, Apr. 2010.

[65] T. Whitworth, M. Ghogho, and D. McLernon, “Optimal training and basis

expansion model parameters for doubly-selective channel estimation,” IEEE
Trans. Wireless Commun., vol. 8, no. 3, pp. 1490–1498, 2009.

[66] A. P. Kannu and P. Schniter, “Design and analysis of MMSE pilot-aided cyclic-
prefixed block tranmissions for doubly selective channels,” IEEE Trans. Signal
Processing, vol. 56, no. 3, pp. 1148–1160, 2008.
 121

[67] Z. Tang, R. Cannizzaro, G. Leus, and P. Baneli, “Pilot-assisted time-varying

chnnel estimation for OFDM systems,” IEEE Trans. Signal Processing, vol. 55,
no. 5, pp. 2226–2238, May 2007.

[68] I. Barhumi and M. Moonen, “MLSE and MAP equalization for transmission
over doubly selective channels,” IEEE Trans. Veh. Technol., vol. 58, no. 8, pp.
4120–4128, Oct. 2009.

[69] K. A. D. Teo and S. Ohno, “Optimal MMSE finite parameter model for doubly-
selective channels,” in IEEE GLOBECOM, Dec. 2005, pp. 3503–3507.

[70] W. Qin and Q.-C. Peng, “Time-varying channel estimation and symbol detec-
tion using superimposed training in OFDM systems,” in Wireless Pers Com-
mun, vol. 47, Oct. 2008, pp. 293–301.

[71] H. Hijazi and L. Ros, “Polynomial estimation of time-varying multipath gains

with intercarrier interference mitigation in OFDM systems,” IEEE Trans. Veh.
Technol., vol. 58, no. 1, pp. 140–151, Jan. 2009.

[72] T. Zemen and C. F. Mecklenbräuker, “Time-variant channel estimation using

discrete prolate spheroidal sequences,” IEEE Trans. Signal Processing, vol. 53,
no. 9, pp. 3597–3607, Sept. 2005.

[73] I. Barhumi, G. Leus, and M. Moonen, “Estimation and direct equalization of

doubly selective channels,” EURASIP Jour. on Applied Signal Processing, pp.
1–15, 2006.

[74] T. Cui, C. Tellambura, and Y. Wu, “Low-complexity pilot-aided channel esti-

mation for OFDM sywtems over doubly-selective channels,” in Proc. ICC’05
Conf., vol. 3, May 2005, pp. 1980–1984.
 122

[75] C. Komninakis, C. Fragouli, A. Sayed, and R. Wesel, “Mulit-input multi-output

fading channel tracking snd equalization using kalman estimation,” IEEE Trans.
Signal Processing, vol. 50, no. 2, pp. 1065–1076, May 2002.

[76] C. Berrou, A. Glavieux, and P. Thitimajshima, “Near Shannon limit error-

correcting coding and decoding: turbo codes,” in Proc. ICC’93 Conf., 1993,
pp. 1064–1070.

[77] J. Hagenauer, E. Offer, and L. Papke, “Iterative deocding of binary block and
convolutional codes,” IEEE Trans. Inform. Theory.

[78] A. Anastasopoulos and K. M. Chugg, “Adaptive soft-input soft-output algo-

rithms for iterative detection with parametric uncertainty,” IEEE Trans. Com-
mun., vol. 48, no. 10, pp. 3686–3691, Oct. 2000.

[79] J. Garcia-Frias and J. D. Villasenor, “Combined turbo detection and decoding

for unknown ISI channels,” IEEE Trans. Commun., vol. 51, no. 1, pp. 79–85,
Jan. 2003.

[80] A. Hansson and T. Aulin, “Generalized APP detection of continuous phase

modulation over unknown ISI channels,” IEEE Trans. Commun., vol. 53, no. 10,
pp. 1615–1619, Oct. 2005.

[81] L. Davis, I. Collings, and P. Hoeher, “Joint MAP equalization and channel es-
timation for frequency-selective and frequency-flat fast fading channels,” IEEE
Trans. Commun., vol. 49, no. 12, pp. 2106–2114, Oct. 2001.

[82] M. C. Valenti and B. D. Woerner, “Iterative channel estimation and decoding

of pilot symbol assited turbo codes over flat-fading channels,” IEEE J. Select.
Areas Commun., vol. 19, no. 9, pp. 1697 – 1705, Sept. 2001.
 123

[83] B. Mielczarek and A. Svensson, “Improved iterative channel estimation and

turbo decoding over flat-fading channels,” in Proc. IEEE VTC2002-fall, vol. 2,
2002, pp. 975–980.

[84] ——, “Modeling fading channel-estimation errors in pilot-symbol-assisted sys-

tems, with application to turbo codes,” IEEE Trans. Commun., vol. 53, no. 11,
pp. 1822 – 1832, Nov. 2005.

[85] R. Otnes and M. Tücheal, “Iterative channel estimation for turbo eqalization
of time-varying frequency-selective channels,” IEEE Trans. Wireless Commun.,
vol. 3, no. 6, pp. 1918–1923, Nov. 2004.

[86] M. F. Flanagan and A. D. Fagan, “Iterative channel estimation, equalization,

and decoding for pilot-symbol assited modulation over frequency selective fast
fading channels,” IEEE Trans. Veh. Technol., vol. 56, no. 4, pp. 1661 – 1670,
July 2007.

[87] M. Nicoli, S. Ferrara, and U. Spagnolini, “Soft-iterative channel estimation:

methods and performance analysis,” IEEE Trans. Signal Processing, vol. 55,
no. 6, pp. 2993 – 3006, June 2007.

[88] J. W. Thomas Zeman, Christoph F. Mecklenbra and R. R.Muller, “Iterative

joint time-variant channel estimation and multi-user detection for MC-CDMA,”
IEEE Trans. Wireless Commun., vol. 5, no. 6, pp. 1469–1478, June 2006.

[89] F. Sanzi, S. Jelting, and J. Speidel, “A comparative study iterative channel es-
timations for mobile OFDM systems,” IEEE Trans. Wireless Commun., vol. 2,
no. 5, pp. 849–859, Sept. 2003.

[90] G. Auer and J. Bonnet, “Threshold controlled iterative channel estimation for
coded OFDM,” in Proc. VTC-Spring Conf., 22-25 April 2007, pp. 1737–1741.
 124

[91] I. Nevat and J. Yun, “Error propagation mitigation for iterative channel track-
ing, detection and decoding of BICM-OFDM systems,” in Proc. ISWCS Conf.,
17-19 Oct. 2007, pp. 75–80.

[92] S. Stefanatos and A. k. Katsaggelos, “Joint data detection and channel tracking
for OFDM systems with phase noise,” IEEE Trans. Signal Processing, vol. 56,
no. 9, pp. 4230–4243, Sept. 2008.

[93] M. Qaisrani and S. Lambotharan, “An iterative (turbo) channel estimation and
symbol detection technique for doubly selective channels,” in Proc. VTC-2007
Spring Conf., 22-25 April 2007, pp. 2253–2256.

[94] X. Li and T. F. Wong, “Turbo equalization with nolinear Kalman filtering

for time-varying frequency-selective fading channels,” IEEE Trans. Wireless
Commun., vol. 60, no. 2, pp. 691–700, Feb. 2007.

[95] Q. Guo, L. Ping, and D. D. Huang, “A low-complexity iterative channel es-

timation and detection technique for double selective channels,” IEEE Trans.
Wireless Commun., vol. 8, no. 8, pp. 4340–4349, Aug. 2009.

[96] H. Kim and J. K. Tugnait, “Turbo equalization for doubly-selective fading chan-
nels using nolinear Kalman filtering and basis expansion models,” IEEE Trans.
Wireless Commun., vol. 9, no. 6, pp. 2076–2087, June 2010.

[97] L. R. Bahl, J. Cocke, F. Jelinek, and J. Raviv, “Optimal decoding of linear

codes for minimizing symbol error rate,” IEEE Trans. Inform. Theory, vol. 20,
no. 2, pp. 284–287, Mar. 1974.

[98] J. G. Proakis and D. G. Manolakis, Digital Signal Processing, 4th ed. New
Jersey: Pearson Prentice Hall, 2006.
 125

[99] B. Sklar, Digital Communications Fundamentals and applications, 2nd ed. Up-
per Saddle River, N.J.: Prentice-Hall, 2000.

[100] B. D. O. Anderson and J. B. Moore, Optimal Filtering. Englewood Cliffs, N.J.:

Prentice-Hall, 1979.

[101] S. L. Marple Jr., Digital Spectral Analysis with Applications. Englewood Cliffs,
N.J.: Prentice Hall, 1987.

[102] P. Whittle, “On the fitting of multivariate autoregressions, and the approximate
canonical factorization of a spectral density matrix,” Biometrika, vol. 50(1-2),
pp. 129–134, 1963.

[103] R. A. Wiggins and E. A. Robinson, “Recursive solution to the multichannel

fitering problem,” J.Geophys.Res., vol. 70, pp. 1885–1891, April 1965.

[104] W. C. Jakes, Microwave Mobile Communications. Piscataway, NJ: IEEE Press,

1994.

[105] K. E. Baddour and N. C. Beaulieu, “Accurate simulation of multiple cross-

correlated rician fading channels,” IEEE Trans. Commun., vol. 52, no. 11, pp.
1980–1987, Nov. 2004.

[106] M. McGuire and M. Sima, “Low-order Kalman filters for channel estimation,”
in PACRIM 2005, Aug. 2005, pp. 165–168.

[107] Y. Li, H. Minn, N. Al-Dhahir, and A. R. Calderbank, “Pilot designs for consis-
tent frequency-offset estimation in OFDM systems,” IEEE Trans. Commun.,
vol. 55, no. 5, pp. 864–877, May 2007.

[108] R. A. Horn and C. R. Johnson, Matrix Analysis. Cambridge University Press,

1985.
 126

[109] P. Z. Peebles Jr., Probability, Random Variables, and Random Signal Principles,
4th ed. McGraw Hill, 2000.

[110] I. Barhumi, G. Leus, and M. Moonen, “Time-domain and frequency-domain

per-tone equalization for OFDM over doubly selective channels,” Signal Pro-
cessing, vol. 84, no. 11, pp. 2055–2066, Nov. 2004.

[111] J. Choi, Adaptive and iterative signal processing in communicstions. Cambridge

University Press, 2006.

[112] C. Douillard, A. Picart, M. JéZéquel, P. Didier, C. Berrou, and A. Glavieux, “It-

erative correction of itersymbol interference: turbo-equalization,” Eur. Trans.
Commun., vol. 6, pp. 507–511, 1995.

[113] L. Hanzo, J. P. Woodard, and P. Robertson, “Turbo decoding and detection

for wireless applications,” Pro. IEEE, vol. 95, no. 6, pp. 1178–1200, June 2007.

[114] L. R. Bahl, J. Cocke, F. Jelinek, and J. Raviv, “Optimal decoding of linear

codes for minimising symbol error rate,” IEEE Trans. Inform. Theory, vol. 20,
no. 3, pp. 284–287, Mar. 1974.

[115] P. Robertson, E. Villebrun, and P. Hoeher, “A comparison of optimal and

sub-optimalMAP decoding algorithms operating in the log domain,” in Proc.
ICC’95 Conf., vol. 2, Jun. 1995, pp. 1009–1013.

[116] L. Zou, Q. Chang, C. Xiu, and Q. Zhang, “Channel estimation and ICI
cancellation for OFDM systems in fast time-varying environments,” IEICE
Trans.Commun., vol. E91-B, no. 4, pp. 1203–1206, April 2008.

[117] J. Choi, Adaptive and iterative signal processing in communicstions. Cambridge

University Press, 2006.
 127

[118] Coded Modulation Library. http://www.iterativesolutions.com.

[119] D. J. Young and N. C.Beaulieu, “The generation of correlated Rayleigh random

variates by inverse discrete Fourier trnasform,” IEEE Commun. Mag., vol. 48,
no. 7, pp. 1114–1127, July 2000.

[120] S. ten Brink, “Designing iterative decoding schemes with the extrinsic infor-
mation transfer chart,” in AEÜ Int. J. Electron.Commun., vol. 54, no. 6, Nov.
2000, pp. 389–398.

[121] ——, “Convergence behavior of iteratively decoded parallel concatenated

codes,” IEEE Trans. Commun., vol. 49, no. 10, pp. 1727–1737, Oct. 2001.

[122] R. K. Michael Tüchler and A. C. Singer, “Turbo equalization: Principles and

new results,” IEEE Trans. Commun., vol. 50, no. 5, pp. 754–76, May 2002.

[123] K. Li and X. Wang, “EXIT chart analysis of turbo multiuser detection,” IEEE
Trans. Wireless Commun., vol. 4, no. 1, pp. 300–311, Jan. 2005.

[124] F. Schreckenbach, N. Görtz, and J. Hagenauer, “Optimized symbol mappings

for bit-interleaved coded modulation with iterative decoding,” IEEE Commun.
Lett., vol. 7, no. 12, pp. 593–595, Dec. 2003.

[125] M. Wetz, D. Huber, W. G.Teich, and J. Lindner, “Robust transmission over

frequency selective fast fading channels with noncoherent turbo detection,” in
VTC2009-Fall, Sept. 2009.

[126] R. Zhang and L. Hanzo, “Three design aspects of multicarrier interleave division
multiple access,” IEEE Trans. Veh. Technol., vol. 57, no. 6, pp. 3607–3617, Nov.
2008.
 128

[127] R. Otnes and M. Tüchler, “EXIT chart analysis applied to adaptive turbo
equalization,” in Proc. Nordic Signal Processing Symposium, 2002.

[128] F. Sanzi, S. Jelting, and J. Speidel, “A comparative study iterative channel es-
timations for mobile OFDM systems,” IEEE Trans. Wireless Commun., vol. 2,
no. 5, pp. 849–859, Sept. 2003.

[129] S. Sand, C. Mensing, and A. Dammann, “Transfer chart analysis of itera-

tive OFDM receivers with data aided channel esitamtion,” in Processings 3rd
COST389 workshop - enabling technologies for B3G systems, July 2006, pp.
1009–1013.

[130] Y. Huang and J. A. Ritcey, “Joint iterative channel estimation and decoding
for bit-interleaved coded modulation over correlated fading channels,” IEEE
Trans. Wireless Commun., vol. 4, no. 5, pp. 2549–2558, Sept. 2005.

[131] P. A. Bello, “Characterization of randomly time-variant linear channels,” IEEE

Trans. Commun., vol. 11, no. 4, pp. 360–393, Dec. 1963.

[132] G. B. Giannakis and C. Tepedelenlioğlu, “Basis expansion models and diversity

techniques for blind identification and equalization of time-varying channels,”
Pro. IEEE, vol. 86, no. 10, pp. 1969–1986, Oct. 1998.

[133] D. K. Borah and B. D. Hart, “Frequency-selective fading channel estimation

with a polynomial time-varying channel model,” IEEE Trans. Commn., vol. 47,
no. 6, pp. 862–873, June 1999.

[134] M. Martone, “Wavelet-based separating kernels for sequence estimation with

unknown rapidly time-varying channels,” IEEE Trans. Lett., vol. 3, no. 3, pp.
78–80, March 1999.
 129

[135] T. Verma, S. Bilbao, and T. H. Y. Meng, “The digital prolate spheroidal win-
dow,” in ICASSP, vol. 3, May 1996, pp. 1351– 1354.

[136] Z. Lin, R. W. McCallum, and H. Wang, “Computation and performance of the

prolate-spheroidal wave function window in spectral estimation,” in ICASSP,
vol. 5, May 1996, pp. 2976– 2978.

[137] D. V. D. Ville, W. Philips, and I. Lemahieu, “On the n-dimensional extension

of the discrete prolate spheroidal window,” IEEE Signal Processing Lett., vol. 9,
no. 3, pp. 89–91, Mar. 2002.

[138] D. Slepian, “Prolate spheriodal wave functions, Fourier analysis and

uncertainty–v: The discrete case,” Bell Syst. Tech. J., vol. 57, pp. 1371–1430,
May-Aug. 1978.

[139] ——, “Some comments on fourier analysis, uncertainty and modeling,” JSTOR:
SIAM Review, vol. 25, no. 3, pp. 379–393, July 1983.

[140] M. Visintin, “Karhunen-loève expansion of a fast Rayleigh fading process.”

[141] K.-W. Yip and T.-S. Ng, “Karhunen-loève expansion of the WSSUS channel
output and its application to efficient simulation,” IEEE J. Select. Areas Com-
mun., vol. 15, no. 4, pp. 640–646, May 1997.

[142] H. Senol, H. A. Cirpan, and E. Panayirci, “A low-complexity KL expansion-

based channel estimation for OFDM systems,” EURASIP Journal on Wireless
Commun. Netw., no. 2, pp. 163–174, Feb. 2005.

[143] H. S. Wang and P.-C. Chang, “On verifying the first-order Markovian assump-
tion for a Rayleigh fading channel model,” IEEE Trans. Veh. Technol., vol. 45,
no. 2, pp. 353–357, May 1996.