ESTIMATION AND SEPARATION OF

LINEAR FREQUENCY–MODULATED SIGNALS

IN WIRELESS COMMUNICATIONS USING
TIME–FREQUENCY SIGNAL PROCESSING

Nguyen Linh–Trung
B. Eng. (Elect. & Comp.), QUT, Australia

Signal Processing Research Centre

Queensland University of Technology
2 George street, QLD 4000, Brisbane, Australia

Document submitted as a requirement for the PhD degree at

Queensland University of Technology.

October 2004
To my loving parents
Contents

Keywords iii

Abstract v

Preface vii

Authorship ix

Acknowledgments xi

List of Figures xiii

List of Abbreviations xvii

List of Symbols xxi

Author’s Publications xxi

1 Thesis Introduction 1
1.1 Problem statement . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.2 Research aim . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
1.3 Research objectives . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
1.4 Research contributions . . . . . . . . . . . . . . . . . . . . . . . . . . 5
1.5 Thesis organization . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7

2 Overview 11
2.1 Time–frequency signal processing . . . . . . . . . . . . . . . . . . . . 12
2.2 Characteristics of wireless communications . . . . . . . . . . . . . . . 22
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
i
ii CONTENTS

3 Instantaneous Frequency Estimation in Multiplicative Noise 39

3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
3.2 Literature review for multiplicative noise . . . . . . . . . . . . . . . . 41
3.3 Signal model and assumptions . . . . . . . . . . . . . . . . . . . . . . 43
3.4 IF estimation based on TFD . . . . . . . . . . . . . . . . . . . . . . . 44
3.5 TFD–based IF estimation with MUSIC . . . . . . . . . . . . . . . . . 59
3.6 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67
Appendix 3A: Scattering function estimation . . . . . . . . . . . . . . . . . 68
Appendix 3B: IF estimation in wideband multiplicative noise . . . . . . . . . 74
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79

4 Underdetermined Blind Source Separation 83

4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 83
4.2 Signal model and assumptions . . . . . . . . . . . . . . . . . . . . . . 86
4.3 Spatial time–frequency distributions . . . . . . . . . . . . . . . . . . . 89
4.4 TF-UBSS algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . 93
4.5 Experiments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 105
4.6 TF-UBSS algorithm using MWVD . . . . . . . . . . . . . . . . . . . . 109
4.7 TF-UBSS algorithm with component extraction . . . . . . . . . . . . . 115
4.8 Numerical performance evaluation . . . . . . . . . . . . . . . . . . . . 120
4.9 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 127
Appendix 4A: Image–based Road Extraction . . . . . . . . . . . . . . . . . 129
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 133

5 Thesis Conclusions and Future Research 137

5.1 Thesis conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . 137
5.2 Directions for future research . . . . . . . . . . . . . . . . . . . . . . 138

Bibiliography 139
Keywords

signal processing; wireless mobile communications; time–frequency distribution;

nonstationary signal; linear frequency–modulation; multiplicative noise (random
amplitude modulation); instantaneous frequency estimation; time–frequency distri-
bution; reduced–interference distribution; array processing; underdetermined blind
source separation; instantaneous linear mixture; spatial time–frequency distribu-
tion; vector clustering; Rayleigh fading; narrowband fading channel; wideband
fading channel; linear time–varying channel; wide–sense stationary uncorrelated
scattering; scattering function; estimation; separation; simulation.

iii
Abstract

Signal processing has been playing a key role in providing solutions to key prob-
lems encountered in communications, in general, and in wireless communications,
in particular. Time–Frequency Signal Processing (TFSP) provides effective tools
for analyzing nonstationary signals where the frequency content of signals varies in
time as well as for analyzing linear time-varying systems. This research aimed at
exploiting the advantages of TFSP, in dealing with nonstationary signals, into the
fundamental issues of signal processing, namely the signal estimation and signal
separation. In particular, it has investigated the problems of (i) the Instantaneous
Frequency (IF) estimation of Linear Frequency–Modulated (LFM) signals corrupted
in complex–valued zero–mean Multiplicative Noise (MN), and (ii) the Underdeter-
mined Blind Source Separation (UBSS) of LFM signals, while focusing onto the
fast–growing area of Wireless Communications (WCom).
A common problem in the issue of signal estimation is the estimation of the
frequency of Frequency–Modulated signals which are seen in many engineering and
real–life applications. Accurate frequency estimation leads to accurate recovery
of the true information. In some applications, the random amplitude modulation
shows up when the medium is dispersive and/or when the assumption of point tar-
get is not valid; the original signal is considered to be corrupted by an MN process
thus seriously affecting the recovery of the information–bearing frequency. The IF
estimation of nonstationary signals corrupted by complex–valued zero–mean MN
was investigated in this research. We have proposed a Second–Order Statistics ap-
proach, rather than a Higher–Order Statistics approach, for IF estimation using
Time–Frequency Distributions (TFDs). The main assumption was that the auto-
correlation function of the MN is real–valued but not necessarily positive (i.e. the
spectrum of the MN is symmetric but does not necessary has the highest peak at
zero frequency). The estimation performance was analyzed in terms of bias and
variance, and compared between four different TFDs: Wigner–Ville Distribution,
Spectrogram, Choi–Williams Distribution and Modified B Distribution. To further
improve the estimation, we proposed to use the Multiple Signal Classification al-
gorithm and showed its better performance. It was shown that the Modified B

v
vi Abstract

Distribution performance was the best for Signal–to–Noise Ratio less than 10dB.
In the issue of signal separation, a new research direction called Blind Source
Separation (BSS) has emerged over the last decade. BSS is a fundamental technique
in array signal processing aiming at recovering unobserved signals or sources from
observed mixtures exploiting only the assumption of mutual independence between
the signals. The term “blind” indicates that neither the structure of the mixtures
nor the source signals are known to the receivers. Applications of BSS are seen
in, for example, radar and sonar, communications, speech processing, biomedical
signal processing. In the case of nonstationary signals, a TF structure forcing
approach was introduced by Belouchrani and Amin by defining the Spatial Time–
Frequency Distribution (STFD), which combines both TF diversity and spatial
diversity. The benefit of STFD in an environment of nonstationary signals is the
direct exploitation of the information brought by the nonstationarity of the signals.
A drawback of most BSS algorithms is that they fail to separate sources in situations
where there are more sources than sensors, referred to as UBSS. The UBSS of
nonstationary signals was investigated in this research. We have presented a new
approach for blind separation of nonstationary sources using their TFDs. The
separation algorithm is based on a vector clustering procedure that estimates the
source TFDs by grouping together the TF points corresponding to “closely spaced”
spatial directions. Simulations illustrate the performances of the proposed method
for the underdetermined blind separation of FM signals. The method developed in
this research represents a new research direction for solving the UBSS problem.
The successful results obtained in the research development of the above two
problems has led to a conclusion that TFSP is useful for WCom. Future research
directions were also proposed.
Preface

I started the association with the Signal Processing Research Centre (SPRC), di-
rected by Professor Boualem Boashash, for my undergraduate final–year project
since I found interested in signal processing. The centre is known internationally
for its expertise in the field of TFSP built up by Professor Boashash. After my
undergraduate studies, I was fortunate to be offered a PhD research program in the
centre. Consequently, I have come to learn more about TFSP.
Is TFSP useful for wireless communications? On the one hand, a fundamental
characteristic of a WCom system is that its channel exhibits a linear time-varying
behavior due to the multipath propagation phenomenon and Doppler effect. Ob-
viously, TFSP provides tools for this problem. On the other hand, some current
drawbacks of TFSP, e.g. in terms of unwanted cross–term appearance and slow
processing time, would discourage practical use of TFSP for WCom requiring fast
computation. From the research point of view, one might hope to bring out some
interesting results for future development of both TFSP and WCom. Standing on
this basis, I carried out my research by investigating possible applications of TFSP
in WCom.
I began with the first application in estimating the instantaneous frequency
of frequency–modulated signals affected by multiplicative noise using TFSP. This
problem was pioneered by Professor Boashash in 1993. Later on, it was brought to
the field of WCom by Professor Boashash and Dr Senadji in 1997. This comes from
the fact that the multipath phenomenon results in a random fading process which
can be represented as multiplicative noise. Further development of this application
formed the first part of my thesis.
In searching for further applications of TFSP in WCom, I was brought into a
new research area in array signal processing, namely BSS, by working with Professor
Karim Abed–Meraim (ENST–Paris). The first idea of applying TFSP in BSS was
introduced by Belouchrani and Amin in 1996. Being a well–known method in array
signal processing, BSS has become an interesting solution for WCom. Especially
when smart antenna technology has opened new directions for WCom over the last

vii
viii Preface

few years. Developing BSS using TFSP then became the second part of my thesis.
Lastly, I hope the small contribution of this research could help accumulate the
understanding of the particular applications of TFSP presented in this thesis, thus
provide some more evidence of the usefulness of TFSP in WCom. It should be
noted that, the purpose of this thesis is not to provide performance comparison
against existing “non time–frequency” approaches. Instead, it aims to propose new
directions in order to show the usefulness of TFSP in WCom.
Authorship

“The work contained in this thesis has not been previously submitted for a de-
gree or diploma at this or any other higher education institution. To the best of
my knowledge and belief, the thesis contains no materials previously published or
written by another person except where due reference is made.”

Signature:

Date:

ix
Acknowledgments

I would like to express my sincere gratitude to Professor Boualem Boashash (QUT,

Australia), who is my PhD principal supervisor. Professor Boashash has given me
invaluable guidance for the last four years of my research for this thesis. Being an
open–minded person, he has always encouraged me to take initiative on whatever
I did. Being an internationally recognized expert in the field of signal processing,
he has always supported me with appropriate opportunities to collaborate with
international researchers. Apart from technical aspects, he was always there to
listen to my problems, of downs in research and life, and to give me sincere advice
and comfort. I was fortunate to have worked with Professor Boashash and to have
learnt substantially from him, a competent researcher and a fine person.
I am grateful to Dr Bouchra Senadji (QUT, Australia), who is my associate
supervisor for the first part of the thesis. She has contributed greatly to my under-
standing of wireless mobile communications and provided valuable technical assis-
tance and suggestions, especially in the first year of my research.
I owe a grateful appreciation to Professor Karim Abed–Meraim (ENST–Paris,
France), who is also another my associate supervisor for the second part of the
thesis. I have had wonderful and enjoyable times working with Professor Abed–
Meraim at the ENST. His brilliance in research and kindness in personality have
indeed brought me into another view of the world of scientific research, as the phrase
“international collaboration” exactly means.
I would like to also thank Dr Robert Iskander (QUT Australia) for his kindness
in helping me step–by–step to write my very first technical paper, and subsequent
assistance over the time; Professor Adel Belouchrani (Algeria) for his stimulating
discussions at ISSPA’2001 (Malaysia) and subsequent email communications; and
the members of the SPRC: Dr Mostefa Mesbah, Dr Zahir Hussain, Victor Sucic,
Mark Keir, Ghasem Azemi, Hamid Hassanpour, for their valuable technical com-
ments, discussions and encouragement, especially to Victor Sucic who has proofread
most of my thesis in details.

xi
xii Acknowledgments

I greatly appreciate the following organizations that have offered me financial

supports (scholarships, top–ups, teaching assistantship) and resources, thus leading
to the completion of the research in this thesis: Signal Processing Research Centre
(SPRC), the school of Electrical and Electronic Systems Engineering (EESE), the
faculty of Built Environment and Engineering (BEE) at the Queensland University
of Technology (QUT, Australia), and department of Signal and Image Processing
at the École Nationale Supérieure des Télécommunications (ENST–Paris, France).
Finally, I would like to thank the examiners (internal committee and external
anonymous reviewers) who have kindly given me invaluable comments and sugges-
tions for improving the quality of this thesis, both technically and presentatively.
List of Figures

2.1 Examples of nonstationary signals.

An engineering application is shown in (a) for a linear FM signal
(plotted using the Wigner–Ville distribution). Real–life applications
are shown in (b–d) for a whale signal, an electroencephalogram signal,
and a bat signal, respectively (all plotted using the B distribution). 14
2.2 Quadratic representations corresponding to the WVD.
ρwvd
z (t, f ), Az (τ, ν), Kz (t, τ ) and Dz (ν, f ) are respectively the WVD,
AF, time–lag signal kernel and the Doppler–frequency signal kernel
of the analytic signal z(t). . . . . . . . . . . . . . . . . . . . . . . . 20
2.3 Dual domains of general signal quadratic representations.
γ(t, f ), Γ (τ, ν), G(t, τ ) and G(ν, f ) are the TFD time–frequency,
Doppler–lag, time–lag and Doppler–frequency kernel, respectively.
ρz (t, f ) and Az (τ, ν) are the general quadratic TFD and the GAF of
the analytic signal z(t). . . . . . . . . . . . . . . . . . . . . . . . . . 21
2.4 Multipath phenomenon. . . . . . . . . . . . . . . . . . . . . . . . . 23
2.5 Typical profile of the received signals Rayleigh fading envelope.
The signal strength is severely faded. . . . . . . . . . . . . . . . . . 24
2.6 Channel classifications. . . . . . . . . . . . . . . . . . . . . . . . . . 26
2.7 Plots of Jakes power spectral density and its autocorrelation. . . . . 28
2.8 Relationship among Bello’s system functions of an LTV channel.
h(t, τ ), T (t, f ), H(ν, f ) and U (τ, ν) are the channel input–delay–
spread (time–varying impulse response) function, the time–varying
transfer function, output–Doppler–spread function and delay–Doppler–
spread function, respectively. A Fourier Transform (FT) operation
from one to another follows the arrow and with respect to the dif-
ferent variables between the underlying functions, for example FT of
h(t, τ ) from t to ν gives U (τ, ν). . . . . . . . . . . . . . . . . . . . . 30
xiii
xiv List of Figures

2.9 Practical QWSSUS wireless channel.

Tm and νD are the maximum multipath–delay and Doppler–shift for
a particular mobile environment, respectively. . . . . . . . . . . . . 33

3.1 Simulated spectrum Sξ (f ) . . . . . . . . . . . . . . . . . . . . . . . 50

3.2 Simulation results for an LFM signal in instantaneous multiplicative
noise.
This example corresponds to a Rayleigh fading narrowband channel.
The WVS achieves energy concentration along the IF of the signal
for 10 and more realizations. . . . . . . . . . . . . . . . . . . . . . . 50
3.3 Simulation results for an LFM signal in instantaneous multiplicative
noise.
The spectrogram (top), and the Choi–Williams distribution (bottom)
are used to reveal the signal IF in the TF domain. . . . . . . . . . . 51
3.4 (a) A general sketch of the colored MN spectrum with a symmetric
shape, and (b) the TF representation of the observed signal in such
a spectrum. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53
3.5 Block diagram of the proposed TFD–based IF estimator. . . . . . . 55
3.6 TFD spectra of the observed signal in MN and the corresponding IF
estimates using the proposed IF estimator. The figures on the left–
hand side are the spectra estimated from: WVD, SPEC, CWD and
MBD. The figures on the right–hand side show the true IF (dotted
lines) and the estimated ones (solid lines). . . . . . . . . . . . . . . 62
3.7 Plot of bias and variance, for the proposed TFD IF estimator, versus
the SNR using Monte Carlo simulation with 1000 runs. . . . . . . . 63
3.8 Block diagram of proposed TFD–based estimator using MUSIC al-
gorithm. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64
3.9 TFD spectra of the observed signal in MN and the corresponding IF
estimates using the proposed MUSIC–TFD IF estimator. The figures
on the left–hand side are the spectra estimated from: WVD, SPEC,
CWD and MBD. The figures on the right–hand side show the true
IF (dotted lines) and the estimated ones (solid lines). Improvement
on the IF estimation are seen (compared with Figure 3.6). . . . . . 65
3.10 Plot of bias and variance, for the proposed MUSIC–TFD IF estima-
tor, versus the SNR using Monte Carlo simulation with 1000 runs.
Improvement on the performance of IF estimation are seen (com-
pared with Figure 3.7). . . . . . . . . . . . . . . . . . . . . . . . . . 66
List of Figures xv

4.1 UBSS: Schematic diagram.

There are n unobserved source signals si (t), i = 1, . . . , n, to be sepa-
rated from m observed mixed signals xj (t), j = 1, . . . , m, corrupted
by AWGN. The underdetermined case corresponds to n > m. . . . . 87

4.2 TF orthogonality.
The TF supports of two sources are disjoint in the TF domain. . . . 88

4.3 TF-UBSS algorithm: Procedures. . . . . . . . . . . . . . . . . . . . 94

4.4 TF-UBSS algorithm: Schematic diagram. . . . . . . . . . . . . . . . 94

4.5 TF quasi–orthogonality.
Small overlapping of the two TF supports is allowable (Ω1 ∩ Ω2 ≈ ∅);
i.e. most of the energy of one source is localized in the TF region
disjoint from the TF support of all other sources. . . . . . . . . . . 101

4.6 TF-UBSS algorithm using notch filters: Schematic diagram. . . . . 104

4.7 Experiment 1: TF-UBSS algorithm with TF orthogonality.

Three monocomponent LFM signals s1 (t), s2 (t) and s3 (t) (a–c), being
the source signals, were tested. The recovered source signals shown
in (m–o) indicated the success of the UBSS. . . . . . . . . . . . . . 107

4.8 Experiment 2: TF-UBSS algorithm with TF quasi–orthogonality.

A mixture of two monocomponent and one multicomponent LFM
signals s1 (t), s2 (t) and s3 (t) (a–c), being the source signals, were
tested. s1 (t) and s2 (t) overlap in TF domain. Source s3 (t) was not
falsely separated into two monocomponent sources. . . . . . . . . . 108

4.9 TF-UBSS algorithm using MWVD: Schematic diagram. . . . . . . . 111

4.10 Experiment 3: TF-UBSS algorithm using MWVD.

A mixture of two monocomponent and one multicomponent LFM
signals s1 (t), s2 (t) and s3 (t) being the source signals with their WVD
shown in (a–c), allowing TF quasi–orthogonality. By using the MWVD
to first reduce the cross–source points in the TFD then apply any of
these two methods for separating the cross–source points from the
auto–source points, we obtained a much better performance (Com-
pared (g) with Figure 4.11) . . . . . . . . . . . . . . . . . . . . . . 114

4.11 Simulated comparison of auto–source selection methods.

Two WVD–based methods were used: (a) approximation projection
and (b) TF orthogonality. Similar performance was observed. . . . . 115

4.12 TF-UBSS algorithm with component extraction: Procedures. . . . . 116

xvi List of Figures

4.13 Experiment 4: TF-UBSS using component extraction for LFM sig-

nals.
(a–c) WVD of s1 (t), s2 (t), s3 (t); (d,e) spatial–averaged TFD of the
mixture outputs using WVD and MWVD; (f) convert STFD mix-
ture to image; (g-h) extraction of source components using image
processing; (i) auto–source points of known components; (j–l) TFD
estimates of the sources; (m–o) TFD of estimated sources after TF
synthesis. SNR = 10 dB. . . . . . . . . . . . . . . . . . . . . . . . . 120
4.14 Experiment 5: TF-UBSS using component extraction for non-LFM
signals.
(a–c) WVD of s1 (t), s2 (t), s3 (t); (d,e) spatial–averaged TFD of the
mixture outputs using WVD and MWVD; (f) convert STFD mix-
ture to image; (g-h) extraction of source components using image
processing; (i) auto–source points of known components; (j–l) TFD
estimates of the sources. SNR = 10 dB. . . . . . . . . . . . . . . . . 121
4.15 Signals used for performance evaluations. . . . . . . . . . . . . . . . 123
4.16 Error performance on mixing matrix estimation. . . . . . . . . . . . 124
4.17 Performance on auto–source selection.
(a)– estimation error on A, (b)–number of auto–source points se-
lected (over the total: 128 × 128 = 16384 points). . . . . . . . . . . 125
4.18 Error performance on IF estimation. . . . . . . . . . . . . . . . . . 126
4.19 Error performance on source waveform estimation. . . . . . . . . . . 127
4.20 Line detection for road extraction problem.
Diagram showing regions associated with pixel x0 and direction dk . 130
List of Abbreviations

AF Ambiguity Function

AIC Akaike Information Criterion

AWGN Additive White Gaussian Noise

BD B Distribution

BJD Born–Jordan Distribution

BS Base Station

BSS Blind Source Separation

CRB Cramé–Rao Lower Bound

CWD Choi–Williams Distribution

2D-FT Two–Dimensional Fourier Transform

2D-IFT Two–Dimensional Inverse Fourier Transform

EEG Electroencephalogram

FM Frequency–Modulated

FT Fourier Transform

GAF Generalized Ambiguity Function

HOS Higher–Order Statistics

ICA Independent Component Analysis

IF Instantaneous Frequency

IFT Inverse Fourier Transform

xvii
xviii List of Abbreviations

ISI Intersymbol Interference

LFM Linear Frequency–Modulated

LOS Line–Of–Sight

LTV Linear Time–Varying

MBD Modified B Distribution

MDL Minimum Description Length

MN Multiplicative Noise

MS Mobile Station

MSE Mean Squared Error

MUSIC Multiple Signal Classification

MWVD Masked Wigner–Ville Distribution

PDF Probability Density Function

QWSSUS Quasi WSSUS

RID Reduced–Interference Distribution

SF Scattering Function

SINR Signal–to–Interference Noise Ratio

SNR Signal–to–Noise Ratio

SOS Second–Order Statistics

SPEC Spectrogram

STFD Spatial Time–Frequency Distribution

STFT Short–Time Fourier Transform

SVD Singular Value Decomposition

TF Time–Frequency

TFD Time–Frequency Distribution

TFSP Time–Frequency Signal Processing

TF–UBSS Time–Frequency–based Underdetermined Blind Source Separation

List of Abbreviations xix

TV Time–Varying

UBSS Underdetermined Blind Source Separation

WMN Wideband Multiplicative Noise

WCom Wireless Communications

WSSUS Wide–Sense Stationary Uncorrelated Scattering

WVD Wigner–Ville Distribution

WVS Wigner–Ville Spectrum

XWVD Crossed Wigner–Ville Distribution

Author’s Publications

Papers in Conference Proceedings

[1] L.-T. Nguyen and B. Senadji, “Analysis of nonlinear signals in the presence of Rayleigh
fading,” in Proceedings of the Fifth International Symposium on Signal Processing and
its Applications (ISSPA’99), vol. 1, pp. 411–414, Brisbane, Australia, Aug. 1999.
[2] L.-T. Nguyen and B. Senadji, “Detection of frequency modulated signals in Rayleigh
fading channels based on time–frequency distributions,” in Proceedings of the Interna-
tional Conference on Acoustics, Speech, and Signal Processing, (ICASSP’00), vol. II,
pp. 729–732, Istanbul, Turkey, June 2000.
[3] L.-T. Nguyen, B. Senadji, and B. Boashash, “Time–frequency based estimators of
scattering function over WSSUS channels,” in Third Australasian Workshop on Signal
Processing and Applications, (WoSPA’00), Brisbane, Australia, Dec. 2000.
[4] L.-T. Nguyen, B. Senadji, and B. Boashash, “Scattering function and time–frequency
signal processing,” in Proceedings of the International Conference on Acoustics,
Speech, and Signal Processing,(ICASSP’01), Salt Lake city, Utah, USA, June 2001.
[5] L.-T. Nguyen, A. Belouchrani, K. Abed-Meraim, and B. Boashash, “Separating
more sources than sensors using time–frequency distributions,” in Proceedings of the
Sixth International Symposium on Signal Processing and its Applications, (ISSPA’01),
Kuala Lumpur, Malaysia, Aug. 2001.
[6] K. Abed–Meraim, N. Linh–Trung, V. Sucic, F. Tupin, and B. Boashash , “An image–
processing approach for underdetermined blind separation of nonstationary sources,”
in Proceedings of the International Symposium on Image and Signal Processing and
Analysis, (ISPA’03), Rome, Italy, Sep. 2003.

Papers in Journal/Bookchapter

[7] K. Abed–Meraim, N. Linh–Trung, and B. Boshash, “Underdetermined blind separa-

tion of FM–like signals.”, in Time–Frequency Signal Analysis and Processing: Methods
and Applications, B. Boashash ed., Elsevier, Oxford, 2003.
xxi
xxii Author’s Publications

[8] N. Linh–Trung, A. Belouchrani, K. Abed-Meraim, and B. Boashash, “Separating more

sources than sensors using time–frequency distributions.” submitted to EURASIP
Journal of Applied Signal Processing, July 2004.

[9] B. Boashash, A. Belouchrani, K. Abed–Meraim and N. Linh–Trung, “Time–frequency

signal processing for wireless communications, in Signal Processing for Mobile Com-
munications Handbook, M. Ibnkahla ed., CRC Press, (in press), 2004.
 Chapter 1

Thesis Introduction

1.1 Problem statement

In order to provide more insight into the nature of nonstationary signals, a new
field of science and engineering has emerged: TFSP [1]. The introduction of TFSP
has led to new tools to represent and characterize the Time–Varying (TV) contents
of nonstationary signals using TFDs. The essential characteristic of TFSP is that it
comprises a set of signal processing methods, techniques and algorithms in which the
two natural variables, time and frequency, are used concurrently [2]. This contrasts
with traditional signal processing methods in which time and frequency variables
are used exclusively and independently. It is envisaged to see how TFSP can be
applied into the two fundamental issues in signal processing (i) signal estimation [3]
and (ii) signal separation [4] of nonstationary signals, in the context of WCom.

In the issue of signal estimation, the estimation of the frequency of Frequency–

1
2 Chapter 1. Thesis Introduction

Modulated (FM) signals. FM signals are used in many engineering applications,

such as in radar, sonar, acoustic emission and communications [3]. Such signals
contain the intended information in the frequency content. Let us consider, as an
example, the application in telecommunications where an information–bearing sig-
nal is sent through a communication channel. Depending on the real–life condition
of the communication channel, the transmitted FM signal is often attenuated and
corrupted by different noise processes (e.g. additive noise, impulsive noise) and
interference agents (other signals in the same communication channel or from other
channels). At the receiving end, one needs to recover the intended information
from the observed signal having been corrupted; for the case of an FM signal in
particular, that is the frequency content. Accurate frequency estimation leads to
accurate recovery of the true information. In many cases, it is assumed that the
received FM signal has a deterministic amplitude over the duration of observation.
While this is a valid assumption in a wide range of scenarios, there are important
applications under which this assumption is inappropriate; the received signal has,
instead, a random amplitude. This phenomenon appears in applications such as
radar imaging, weather radar, underwater acoustic applications, wireless mobile
communications. In these applications, the random amplitude modulation shows
up when the medium is dispersive and/or when the assumption of point target is
not valid [5, 6]. In the literature, the original signal is often said to be corrupted
by a MN process. Different underlying media result in different structures of the
MN. Anyhow, the recovery process of the information–bearing frequency becomes
more difficult. In the case of nonstationary signals, a time–frequency approach
using Higher–Order Statistics (HOS) of TFD was proposed by Boashash and Ris-
tic [7, 8] for real–valued MN. Many other approach for real–valued model are seen
in [9–19]. In contrast, the complex–valued MN model has just been considered
recently in [20–22]. Estimation of the IF of nonstationary signals corrupted by
complex–valued zero–mean MN is investigated in this research.

A new research direction in the issue of signal separation is the BSS. BSS is a
fundamental technique in array signal processing aiming at recovering unobserved
1.1. Problem statement 3

signals or sources from observed mixtures (typically, the outputs of a multisensor

array), exploiting only the assumption of mutual independence between the sig-
nals [4]. The term “blind” indicates that neither the structure of the mixtures nor
the source signals are known to the receivers. This technique is of importance when
modeling the transfer from the sources to the sensors is difficult and/or when no
a priori information is available about the mixtures. Applications of BSS are seen
in, for example, radar and sonar, communications, speech processing, biomedical
signal processing. In the case of nonstationary signals, a Time–Frequency (TF)
structure forcing approach was introduced by Belouchrani and Amin [23, 24] by
defining the STFD, which combines both TF diversity and spatial diversity. The
benefits of STFD in an environment of nonstationary signals is the direct exploita-
tion of the information brought by the nonstationarity of the signals. In contrast
to BSS approaches using Second–Order Statistics (SOS) and HOS, this approach
allows the separation of Gaussian sources with identical spectral shape but with
different TF localization properties. Moreover, the effects of spreading the noise
power, while localizing the source energy in the TF domain, amounts to increasing
the Signal–to–Noise Ratio (SNR) [25]. A drawback of most BSS algorithms is that
they fail to separate sources in situations where there are more sources than sen-
sors [26]. This challenging problem, known as the UBSS, has recently been studied
in [26–32]. The UBSS of nonstationary signals is investigated in this research. The
TF structure forcing approach is chosen in order to take advantage of TF signal
processing, over the classical time–only and frequency–only signal processing, to
separate and recover nonstationary source signals.

In both problems addressed above, we will focus on a particular type of non-

stationary signals, the LFM signal. Though presenting the simplest form of FM
nonstationarity, LFM signals are both fundamental for research as well as practical
for applications such as in radar [33] and in communications [34, 35].
4 Chapter 1. Thesis Introduction

1.2 Research aim

This research aims to exploit the advantages of TFSP, in dealing with nonstationary
signals, into the problems of

1. the instantaneous frequency estimation of LFM signals corrupted in MN, and

2. the UBSS of LFM signals,

while focusing onto the fast–growing area of WCom.

1.3 Research objectives

To achieve the above aim, the main objectives of this research are set out as follows:

1. performing in–depth reviews on the fundamental concepts of: TFSP and

WCom (addressed in Chapter 2).

2. reviewing existing methods for IF estimation of nonstationary signals cor-

rupted by MN (addressed in Chapter 3).

3. investigating the IF estimation of LFM signals using TFSP in the context of

WCom so as to propose a new estimator for extracting the IF (addressed in
Chapter 3).

4. studying the statistical performance of the proposed estimator (addressed in

Chapter 3).

5. reviewing existing methods for BSS (addressed in Chapter 4).

6. investigating the UBSS of LFM signals so as to propose a new algorithm for

UBSS (addressed in Chapter 4).
1.4. Research contributions 5

7. studying the robustness of the proposed UBSS algorithm (addressed in Chap-

ter 4).

8. drawing a conclusion on the applicability and usefulness of TFSP in WCom

and proposing future research directions (addressed in Chapter 5).

1.4 Research contributions

To the best of the author’s knowledge, the research reported in this thesis has
achieved the above objectives. In particular, it has:

1. showed that the TFD–based IF estimation of LFM signals in MN can be

achieved by an SOS approach rather than an HOS one [Section 3.4.1].

2. proposed a TFD–based IF estimator for LFM signals in complex–valued zero–

mean MN under the main assumption that the MN autocorrelation is real (but
not necessary positive), and analyzed its statistical performance in terms of
bias and variance among 4 different TFDs (Wigner–Ville Distribution (WVD),
Spectrogram (SPEC), Choi–Williams Distribution (CWD) and Modified B
Distribution (MBD)). The estimator is based on peak–detection of the TFD
spectrum of the observed signal [Section 3.4.2].

3. proposed a performance–improved version of the above estimator by using

MUSIC algorithm instead of peak–detecting the TF representation. Both
versions of the estimator can still estimate the IF in the low SNR (less than
10dB) scenario in contrast to the approach using first–order moment of TFD
[Section 3.5].

4. proposed, with only preliminary analyses, two TFD–based estimators for the
Scattering Function (SF) [Appendix 3A], and a TFD–based IF estimator for
LFM signals in wideband MN [Appendix 3B ].
6 Chapter 1. Thesis Introduction

5. proposed a Time–Frequency–based Underdetermined Blind Source Separation

(TF–UBSS) algorithm for both monocomponent and multicomponent LFM
signals using mainly the assumption of TF orthogonality, and TF quasi–
orthogonality. The algorithm is based on a vector clustering procedure that
estimates the source TFDs by grouping together the TF points corresponding
to “closely spaced” spatial directions [Section 4.4 and Section 4.5].

6. proposed the use of Masked Wigner–Ville Distribution (MWVD) to achieve

the robustness of the TF–UBSS algorithm, especially in terms of auto–source
TF point selection. [Section 4.6].

7. proposed the use of image processing to also enhance the auto–source TF

point selection procedure, by using a component–extraction procedure in the
proposed TF–UBSS algorithm [Section 4.7].

8. analysed the statistical performance of the TF–UBSS algorithm with MWVD,

especially on the estimation of mixing matrix, IF and source signal [Sec-
tion 4.8].

1.5 Thesis organization

The thesis consists, in total, of 5 chapters (including this introductory chapter)

as follows. Chapter 2 serves as a review on fundamentals of TFSP and WCom,
upon which our research developments have been built in the subsequent chapters.
Chapter 3 reviews existing methods for IF estimation of nonstationary signals cor-
rupted by MN, proposes a TFD–based estimator for the IF of LFM signals and
studies the statistical performance comparison among different TFDs in term of
bias and variance. Chapter 4, including the literature review of BSS, proposes a
TFD–based UBSS algorithm for LFM signals and provides further improvements
on its robustness. Chapter 5 is for conclusions and future research directions.
References

[1] B. Boashash and V. Sucic, “High performance time–frequency distributions for prac-
tical applications,” in Wavelets and Signal Processing (L. Debnath, ed.), Birkhauser,
Boston, New York: Springer–Verlag, 2002.
[2] B. Boashash, ed., Time Frequency Signal Analysis and Processing: Method and Ap-
plications. Oxford: Elsevier, 2003.
[3] B. Boashash, “Estimating and interpreting the instantaneous frequency of a signal-
Part 1: Fundamentals,” Proceedings of the IEEE, vol. 80, pp. 519–538, Apr. 1992.
[4] J. F. Cardoso, “Blind signal separation: Statistical principles,” Proceedings of the
IEEE, vol. 9, pp. 2009–2025, Oct. 1998.
[5] R. Dwyer, “Fourth–order spectra of Gaussian amplitude modulated sinusoids,” The
Journal of the Acoustical Society of America, vol. 90, pp. 916–926, 1991.
[6] S. Haykin, Communication Systems. New York: Wiley, 3rd ed., 1994.
[7] B. Boashash and B. Ristic, “Analysis of FM signals affected by Gaussian AM using re-
duced Wigner–Ville trispectrum,” in International Conference on Acoustics, Speech,
and Signal Processing, ICASSP’93, vol. IV, (Minneapolis), pp. 408–411, 1993.
[8] B. Boashash and B. Ristic, “Polynomial time-frequency distributions and time-
varying higher order spectra: Application to the analysis of multicomponent FM
signals and to the treatment of multiplicative noise,” Signal Processing, vol. 67,
pp. 1–23, 1998.
[9] O. Besson and F. Castanié, “On estimating the frequency of a sinusoid in autore-
gressive multiplicative noise,” Signal Processing, vol. 30, pp. 65–83, Jan. 1993.
[10] A. Swami, “Multiplicative noise models: parameter estimation using cumulants,”
Signal Processing, vol. 36, pp. 355–373, Apr. 1994.
[11] G. Zhou and G. B. Giannakis, “On estimating random amplitude-modulated har-
monics using higher order spectra,” IEEE Journal of Oceanic Engineering, vol. 19,
pp. 529–539, Oct. 1994.
[12] S. Shamsunder and G. Giannakis, “Estimating random amplitude polynomial phase
signals: a cyclostationary approach,” IEEE Transaction on Signal Processing, vol. 43,
pp. 492–505, Feb. 1995.
7
8 Chapter 1. Thesis Introduction

[13] G. Zhou and G. Giannakis, “Harmonics in Gaussian multiplicative and additive noise:
Cramer-Rao bounds,” IEEE Transactions on Signal Processing, vol. 43, pp. 1217–
1231, May 1995.
[14] B. Friedlander and J. Francos, “Estimation of amplitude and phase parameters of
multicomponent signals,” IEEE Transactions on Signal Processing, vol. 43, pp. 917–
926, Apr. 1995.
[15] J. Francos and B. Friedlander, “Bounds for estimation of multicomponent signals
with random amplitude and deterministic phase,” IEEE Transactions on Signal Pro-
cessing, vol. 43, pp. 1161–1172, May 1995.
[16] A. Swami, “Cramer-Rao bounds for deterministic signals in additive and multiplica-
tive noise,” Signal Processing, vol. 53, pp. 231–244, 1996.
[17] O. Besson, N. Ghogho, and A. Swami, “Parameter estimation for random amplitude
chirp signals,” IEEE Transactions on Signal Processing, vol. 47, pp. 3208–3219, Dec.
1999.
[18] M. Ghogho, A. K. Nandi, and A. Swami, “Cramer-Rao bounds and maximum like-
lihood estimation for random amplitude phase–modulated signals,” IEEE Transac-
tions on Signal Processing, vol. 47, pp. 2905–2916, Nov. 1999.
[19] M. R. Morelande and A. M. Zoubir, “Model selection of random amplitude polyno-
mial phase signals,” IEEE Transactions on Signal Processing, vol. 50, pp. 578–589,
Mar. 2002.
[20] M. Ghogho, A. Swami, and T. S. Durrani, “Frequency estimation in the presence
of Doppler spread: performance analysis,” IEEE Transactions on Signal Processing,
vol. 49, pp. 777–789, Apr. 2001.
[21] B. Barkat, “Instantaneous frequency estimation of nonlinear frequency–modulated
signals in the presence of multiplicative and additive noise,” IEEE Transactions on
Signal Processing, vol. 49, pp. 2214–2222, Oct. 2001.
[22] G. Azemi, B. Senadji, and B. Boashash, “Instantaneous frequency estimation of
frequency modulated signals in the presence of multiplicative and additive noise:
application to mobile communication systems.” to appear in EUSIPCO2002.
[23] A. Belouchrani and M. G. Amin, “A new approach for blind source separation using
time-frequency distributions,” in Proceedings SPIE conference on Advanced algo-
rithms and Architectures for Signal Processing, (Denver, Colorado), 1996.
[24] A. Belouchrani and M. G. Amin, “Blind source separation based on time-frequency
signal representations,” IEEE Transactions on Signal Processing, vol. 46, pp. 2888–
2897, Nov. 1998.
[25] B. Boashash, ed., Time-Frequency Signal Analysis: Methods and Applications. Mel-
bourne, Australia: Longman Cheshire, 1992.
[26] P. Comon, “Blind channel identification and extraction of more sources than sensors,”
in Proceedings of the SPIE, vol. 3461, (San Diego), July 1998.
[27] A. Belouchrani and J.-F. Cardoso, “Maximum likelihood source separation for dis-
crete sources,” in Proceedings EUSIPCO, 1994.
References 9

[28] A. Taleb and C. Jutten, “On underdetermined source separation,” in ICASSP’99,

vol. 3, pp. 1445–1448, 1999.
[29] T.-W. Lee, M. S. Lewicki, M. Girolami, and T. J. Sejnowski, “Blind source separation
of more sources than mixtures using overcomplete representations,” IEEE Signal
Processing Letters, vol. 6, pp. 87–90, Apr. 1999.
[30] P. Comon and O. Grellier, “Nonlinear inversion of underdetermined mixtures,” in
ICA’99, (Aussois, France), pp. 461–465, Jan. 1999.
[31] K. I. Diamantaras, “Blind separation of multiple binary sources using a single lin-
ear mixture,” in International Conference on Acoustics, Speech, and Signal Process-
ing, ICASSP’2000, vol. V, (Istanbul, Turkey), pp. 2657–2660, June 2000.
[32] A. Jourjine, S. Rickard, and O. Yilmaz, “Blind separation of disjoint orthogonal sig-
nals: demixing n sources from 2 mixtures,” in International Conference on Acoustics,
Speech, and Signal Processing, ICASSP’2000, vol. 5, (Istanbul, Turkey), pp. 2985–
2988, June 2000.
[33] A. Rihaczek, Principles of High-Resolution Radar. Peninsula Publishing, 1985.
[34] M. G. Amin, “Interference mitigation in spread spectrum communication systems
using time-frequency distributions,” IEEE Transactions on Signal Processing, vol. 45,
pp. 90–101, Jan. 1997.
[35] C. Gupta and A. Papandreou-Suppappola, “Wireless CDMA communications using
time–varying signals,” in Proceedings of the Sixth International Symposium on Signal
Processing and its Applications, ISSPA’01, vol. 1, pp. 242–245, 2001.
 Chapter 2

Overview

TFSP provides effective tools for analyzing nonstationary signals, whose frequency
content varies in time, as well as linear time–varying systems. TFSP is a natural
extension of both the time domain and the frequency domain processing, that in-
volves representing signals in a two–dimensional space, and so reveals “complete”
information about the signal. Such a representation is intended to provide a dis-
tribution of signal energy versus time and frequency simultaneously. More details
and advances of TFSP can be found in [1–5].

In the field of telecommunications, WCom is growing at an explosive rate, stim-

ulated by a host of important emerging applications ranging from third-generation
mobile telephony, wireless personal communications, and wireless subscriber loops,
to radio and infrared indoor communications, nomadic computing and wireless tac-
tical military communications [6]. More details and advances of WCom can be
found in [6–11].

Signal processing has been playing an important role in providing solutions

11
12 Chapter 2. Overview

to some key problems encountered in communications in general, and in wireless

communications in particular [6]. The aim of this thesis (see Chapter 1) is to see
how TFSP can be applied into some existing theoretical issues of signal processing
as well as practical applications, such as WCom. This chapter, therefore, provides
a brief background on the fundamental concepts of TFSP and WCom that will be
used in the subsequent chapters.

2.1 Time–frequency signal processing

The magnitude spectrum (frequency representation) of a signal gives no indication

as to how the frequency content of the signal changes with time, an important
information when one deals with FM signals. TFSP, being a natural extension of
both the time domain and the frequency domain processing, preserves and reveals
this information about the signal. TFSP involves, and is intended to provide a
distribution of signal energy versus both time and frequency. For this reason, the
TF representation is commonly referred to as a TFD [5].

2.1.1 Nonstationarity and FM signals

We recall now some important definitions.

Definition 2.1 (Analytic signal [5]).

Let s(t) be a real FM signal of the general form:

s(t) = A(t) · cos[θ(t)], (2.1)

with the assumption that the spectra of the amplitude A(t) and phase θ(t) are sep-
arated (nonoverlapped) in frequency, i.e. the signal approaches a narrowband con-
dition [12].
2.1. Time–frequency signal processing 13

Let H[·] denote the Hilbert transform of the signal, such that
∞
s(t − τ )
Z
∆
H[s(t)] = p.v. dτ (2.2)
−∞ πτ

where p.v. is the Cauchy principle value of the integral.

A signal z(t) defined as

∆
z(t) = s(t) + jH[s(t)] ≈ A(t) ejθ(t) (2.3)

is called the analytic signal of the real signal s(t). The approximation is valid for
the above narrowband condition.

The definition of the analytic signal is important to define the IF of signal s(t).

Definition 2.2 (Instantaneous frequency [5]).

Let z(t) be an analytic signal given in the form

z(t) = Az (t) ejθz (t) (2.4)

The Instantaneous Frequency (IF) of signal z(t) is then defined as

∆ 1 dθz (t)
fin (t) = (2.5)
2π dt

The IF, fin (t), presents a measure of the localization in time of “that” frequency
at time t. In this sense, a signal is said to be nonstationary if its IF varies in time.
We can observe in Figure 2.1 the TV behavior of an engineering signal (linear FM
signal, used in radar and military applications) and real–life signals (whale song,
electroencephalogram signal, bat signal).

Note that, Definition 2.2 is applicable to monocomponent signals only, such as

the signal illustrated in Figure 2.1.a. When more than one “ridge” appears in the
signal TF representation, the signal is said to be multicomponent, e.g. the signals
in Figure 2.1.b–d. The importance of the IF and its applications is represented by
14 Chapter 2. Overview

Boashash in [1, 12, 13].

Fs=1Hz N=128 Fs=1Hz N=7000

WHALE SIGNAL
Time−res=4 Time−res=120
7000
120

6000

100
5000

80
Time (secs)

Time (secs)
4000

60
3000

40
2000

20 1000

0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5
Frequency (Hz) Frequency (Hz)

(a) Linear FM signal (b) Whale signal

Fs=20Hz N=600 Fs=1Hz N=400
b−Distribution BAT SIGNAL
Time−res=5 Time−res=8
30 400

350
25

300

20
250
Time (seconds)

Time (secs)

15 200

150
10

100

5
50

1 2 3 4 5 6 7 8 9 10 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5
Frequency (Hz) Frequency (Hz)

(c) Electroenphalogram signal (d) Bat signal

Figure 2.1: Examples of nonstationary signals.

An engineering application is shown in (a) for a linear FM signal (plotted using
the Wigner–Ville distribution). Real–life applications are shown in (b–d) for a
whale signal, an electroencephalogram signal, and a bat signal, respectively (all
plotted using the B distribution).

Definition 2.3 (LFM signal [5]).

Considering the typical FM transmission in communications systems, a narrowband
FM signal is commonly defined as [12]:
 Z t 
∆
s(t) = A(t) · cos 2πfc t + 2π m(τ ) dτ , (2.6)
−∞

where fc is the center frequency. When m(t) is a linear function of t, i.e. m(t) = αt,
signal s(t) is called a Linear Frequency–Modulated (LFM) signal. In addition,
2.1. Time–frequency signal processing 15

if A(t) is a rectangular function1 of duration T , as denoted by rect(t), the signal is

called a “chirp”. A chirp signal can be expressed as [14]:

α 2
schirp (t) = rectT (t) cos[2π(fc t + t )] (2.7)
2

The analytic signal associated with schirp (t) is then given by

α 2
z chirp (t) = rectT (t) ejθ(t) = rectT (t) ej2π(fc t+ 2 t )
(2.8)

and its IF is

1 dθ(t)
finchirp (t) = = fc + αt. (2.9)
2π dt

The chirp signal defined in (2.8) is of practical importance. It is the basic signal
used in radar applications, and can be easily generated [14]. It is also used in
military communication applications where the chirp is sent out as a hostile signal
to destroy other communications [8, 15, 16]. In this thesis, we will refer to the chirp
signal as an LFM signal (i.e. the rectangular amplitude is implicit).

Based on the above concepts of analytic signals and instantaneous frequency

for nonstationary signals, we now see how they evolve around the fundamentals of
TFSP.

2.1.2 SPEC, WVD, and quadratic TFDs

To study the spectral properties of the signal at time t, an intuitive approach is to,
first, take a slice of the signal by applying a moving window centered at time t to

1
Definition of rectangular function of duration T :
(
1 if t ≤ T,
rectT (t) =
0 otherwise.
16 Chapter 2. Overview

the signal, and then calculate the magnitude spectrum of the windowed signal. In
the literature, this method is referred to as the SPEC [1, 2]. It is mathematically
expressed as
Z ∞
2
 stft 2  −j2πf τ

ρspec
z (t, f ) = ρz (t, f ) =  z(τ )h(τ − t)e dτ  . (2.10)
−∞

where ρstft
z (t, f ) is the Short–Time Fourier Transform (STFT), and h(t) is some

window function. By varying t, one can obtain the spectral density as a function
of t.

The SPEC is a simple, popular and robust method in the analysis of nonstation-
ary signals. It is a proper energy distribution in the sense that it is positive. On
the other hand, the SPEC has an inherent limitation: the frequency resolution is
dependent on the length (and the type) of the analysis window; too short windows
cause a decrease in frequency resolution, and too long windows cause a decrease in
time resolution, thus an inherent trade–off between time and frequency resolution
in the SPEC for a particular window.

It was argued that since a signal has a spectral structure at any given time,
there should exist the notion of an “instantaneous spectrum” which has the physical
attributes of an energy density. Based on this argument, the WVD was derived,
and is defined for an analytic signal z(t) as [1]
Z ∞
∆ τ τ
ρwvd
z (t, f ) = z(t + )z ∗ (t − ) e−j2πf τ dτ. (2.11)
−∞ 2 2

It can be observed from (2.11) that the WVD is the Fourier Transform (FT)2 of
Kz (t, τ ) from τ to f , where

∆ τ τ
Kz (t, τ ) = z(t + )z ∗ (t − ) (2.12)
2 2

2
Convention of FT and Inverse Fourier Transform (IFT) operations: an FT operation will
transform a function either from t to ν domain, or from τ to f domain; inversely, an IFT operation
goes either from ν back to t, or from f back to τ .
2.1. Time–frequency signal processing 17

is called the time–lag signal kernel.

The WVD is the most widely studied TFD. It achieves maximum energy con-
centration in the TF plane about the IF for LFM signals[ref]. However, it is in
general non–positive and it introduces cross–terms when multiple frequency laws
(e.g. two LFM components) exist in the signals.

A general class of quadratic TFDs can be obtained by smoothing/filtering the

WVD in t and f , and is expressed as [17]
ZZZ ∞
∆ τ τ
ρz (t, f ) = ej2πν(u−t) Γ (τ, ν) z(u + )z ∗ (u − ) e−j2πf τ dν du dτ (2.13)
−∞ 2 2

where Γ (τ, ν) is a two–dimensional function in the Doppler–lag domain, (τ, ν), and
is called the TFD Doppler–lag kernel. The kernel determines the TFD and its
properties. We can obtain the TFDs with certain desired properties by properly
constraining the Γ (τ, ν) function.

Equation (2.13) can be simplified as [5]:

ρz (t, f ) = γ(t, f ) ? ? ρwvd

z (t, f ). (2.14)
t f

The notation ? ? 3 in (2.14) represents a convolution in both t and f directions, and

tf
γ(t, f ) is the time–frequency kernel obtained through a Two–Dimensional Fourier
Transform (2D-FT)4 operation on Γ (τ, ν) as:
ZZ ∞
γ(t, f ) = Γ (τ, ν) e−j2πf τ e+j2πtν dτ dν (2.15)
−∞

3
Later, the notations ? and ? will denote the convolution in time and frequency, respectively.
t f
4
Convention of 2D-FT and Two–Dimensional Inverse Fourier Transform (2D-IFT) operations:
a 2D-FT operation, transforming a function of two variables (t, f ) to another function of (τ, ν),
contains one FT operation from t to ν and one IFT operation from f to τ , and the FT and IFT
are interchangeable; inversely, a 2D-IFT operation, transforming a function of two variables (τ, ν)
back to (t, f ), contains one IFT operation from ν to t and one FT operation from τ to f , and
these IFT and FT operations are also interchangeable.
18 Chapter 2. Overview

2.1.3 Reduced interference distributions

The problem of cross–terms introduced by WVD when applying it to a multicompo-

nent signal can be dealt with by selecting a suitable kernel Γ (τ, ν) which minimizes
the cross–terms effectively. The corresponding TFDs to such kernels are known as
Reduced–Interference Distributions (RIDs). Following are some examples of RIDs
and their corresponding Doppler–lag kernels:

• Born–Jordan Distribution (BJD) [17]:

|τ |
Z ∞ Z t+
1 2 τ τ
ρBJD
z (t, f ) = z(u + )z ∗ (u − ) du e−j2πf τ dτ (2.16a)
−∞ |τ | t−
|τ |
2
2 2
sin(πντ )
Γ(τ, ν) = (2.16b)
πντ

• CWD [18]:
ZZ r
πσ −π2 σ(u−t)2 /τ 2 τ τ
ρcwd
z (t, f ) = 2
e z(t + ) z ∗ (t − ) e−j2πf τ du dτ (2.17a)
τ 2 2
2 τ 2 /σ
Γ (τ, ν) = e−v (2.17b)

where σ is a constant.

• MBD [19]:

Γ(2α)/22α−1 Γ2 (α)
  Z
τ τ
ρmbd
z (t, f )= ? z(t + ) z ∗ (t − ) e−j2πf τ dτ (2.18a)
cosh2α (t) t 2 2
|Γ(α + jπν)|2
Γ (τ, ν) = (2.18b)
Γ2 (α)

where α is a real positive number, Γ stands for the Gamma function.

2.1. Time–frequency signal processing 19

2.1.4 The WVD and AF

By taking the 2D-FT of the WVD, we obtain the symmetrical Ambiguity Function
(AF), also called Sussman AF

Az (τ, ν) = FT FT −1 ρwvd

z (t, f )
t→ν f →τ
ZZ ∞
j2πf τ −j2πνt
= ρwvd
z (t, f ) e e dt df
−∞
Z ∞
= Kz (t, τ ) e−j2πνt dt. (2.19)
−∞

Slightly different definitions of AF have been used by different authors, however,

they are all related to the symmetrical form Az (t, τ ) in (2.19) [20].

A nonstationary signal, therefore, can be analyzed in either the time–frequency

domain (t, f ) or the ambiguity domain (τ, ν), also called Doppler–lag domain.
There, also, exists a relationship between the WVD and the AF via the Radon
transform [21], that is, the FT of the Radon–transformed WVD yields the AF in
polar coordinates [22].

The concept of AF has been used as a very effective tool in the design of radar
signals [1, 23]. This function is the basis in modern radar technology.

Ambiguity domain directly represents the effects of multipath–delay spread and

Doppler–shift spread which characterize the TV nature of wireless communica-
tions [24]. Thus, the AF can be effectively used to analyze the behavior of signals in
wireless communications and, in turns, control the above two effects of spreading,
e.g., using windowing and smoothing.

2.1.5 Relationships among dual domains

The relationship between dual domain pairs, time–frequency and Doppler–delay,

and, time–delay and Doppler–frequency, can be represented as in Figure 2.2 [1,
20 Chapter 2. Overview

5] through FTs and IFTs with respect to variables. Each arrow in Figure 2.2
represents a FT from one variable to the other, the inverse direction represents an
IFT operation.


 

FT

FT


 

FT

FT


 
Figure 2.2: Quadratic representations corresponding to the WVD.
ρwvd
z (t, f ), Az (τ, ν), Kz (t, τ ) and Dz (ν, f ) are respectively the WVD, AF, time–
lag signal kernel and the Doppler–frequency signal kernel of the analytic signal
z(t).

Moreover, for the general quadratic class of TFDs in (2.13), the above relation-
ship is illustrated in Figure 2.3 [1, 5], the Az (τ, ν) is the Generalized Ambiguity
Function (GAF).

Note that, there is a strong coherence between quadratic TF signal representa-

tions and Linear Time–Varying (LTV) systems [20, 25]. A quadratic TF analysis of
a LTV system can be based on the linear relation between the WVDs [26] or mod-
ified WVDs [25] of both the input and the output. This input-output relationship
can be, in general, described as TFDs [27–30]

E {ρx (t, f )} = ρs (t, f ) ? ? Ψh (t, f ) (2.20)

t f

where ρs (t, f ) is a TFD of the input s(t); Ψh (t, f ) is the scattering function which
2.1. Time–frequency signal processing 21

+ ,
#! -$/.(
#! 0 1 +,24 63 5
#!

FT

FT


 




 !

FT

FT
" 
# ! $&%
  ('*) 
 
Figure 2.3: Dual domains of general signal quadratic representations.
γ(t, f ), Γ (τ, ν), G(t, τ ) and G(ν, f ) are the TFD time–frequency, Doppler–lag,
time–lag and Doppler–frequency kernel, respectively. ρz (t, f ) and Az (τ, ν) are
the general quadratic TFD and the GAF of the analytic signal z(t).

is related to the random LTV channel impulse response h(t, ν); and E {ρx (t, f )} is
the expected value of a TFD of the output x(t).

2.1.6 Time–frequency signal synthesis

Opposite to the TF signal analysis whereby the analysis algorithms are used to
analyze the TV frequency behavior of signals, TF signal synthesis algorithms are
used to synthesize, or estimate, signals from their TFDs. Mathematically, assuming
that z(t) is a signal of interest with ρz (t, f ) being its TFD in the general quadratic
class, the synthesis problem can be formulated as: find the analytic signal ẑ(t)
whose estimate TFD, ρẑ (t, f ), best approximates ρz (t, f ). Consequently, ẑ(t) gives
the best estimate of z(t). Seminal to the problem of TF signal synthesis is the
algorithm by Boudreaux–Bartels in [31] using WVD. The basis for the solution is
22 Chapter 2. Overview

the inversion property of the WVD [1]

Z ∞
1 t
z(t) = ∗ ρwvd
z ( , f) e
j2πf t
df (2.21)
z (0) −∞ 2

implying that the signal may be reconstructed to within a complex exponential

constant ejα = z ∗ (0)/|z(0)| given |z(0)| =
6 0. Other TF synthesis algorithms can be
found in [32–36].

2.1.7 IF estimation

There are two major existing approaches for IF estimation using TFDs. The first is
built on the first–order moment of TFDs [32]. The first–order moment of the WVD
yields the IF [32, 37], while others yield approximations of the IF [1]. However it
fails to estimate multicomponent signals due to the presence of cross–terms.

The second approach is built on utilizing the fact that all TFDs have peaks
around the IF laws of signals. The peaks of the WVD was used for IF estimation
and applied to many problems [1]. For better performance at lower SNRs, the
Crossed Wigner–Ville Distribution (XWVD) was proposed [38]. Other algorithms
of TFD–based peak estimation can be found in, for examples, [1, 39, 40]. Like
the first approach, this approach also suffers from the presence of cross–terms in
multicomponent signals which results in poor estimation.

Upon the desired to design high resolution RIDs, B Distribution (BD) was then
proposed in [41], and MBD was developed in [19], both with adaptive algorithms
for IF estimation of multicomponent signals.

2.2 Characteristics of wireless communications

Wireless communications (WCom) is growing at an explosive rate, stimulated by

a host of important emerging applications ranging from third–generation mobile
2.2. Characteristics of wireless communications 23

telephony, wireless personal communications, and wireless subscriber loops, to ra-

dio and infrared indoor communications, nomadic computing and wireless tactical
military communications. Signal processing has been playing a key role in provid-
ing solutions to key problems encountered in communications, in general, and in
wireless communications, in particular [6].

2.2.1 Multipath propagation and fading

WCom takes place between a fixed Base Station (BS) and a number of roaming
Mobile Stations (MSs) [7, 10]. Let consider a BS transmitting a signal which per-
vades the coverage area in which a MS is traveling. The MS does not receive one
version of the transmitted signal, but a number of versions through different signal
paths which have been reflected and diffracted by buildings and other parapher-
nalia. This phenomenon results in a common terminology called multipath radio
propagation [7] that is illustrated in Figure 2.4.



non line-o
f-sight

line
-of-
s ight



Base Station
(Transmitter)
Mobile Station
(Receiver)

Figure 2.4: Multipath phenomenon.

Each version of the transmitted signal received by the MS is subjected to a

24 Chapter 2. Overview

specific time–delay, amplitude attenuation, and phase distortion and Doppler–shift

depending on its path from the BS to the MS [11]. The received signal is a sum
of the transmitted signal versions arriving via many paths. A serious condition
occurs when the multipath signals sum to a small value. The received signal is
said to be in a fade, and the phenomenon is called multipath fading. Thus, the
communication channel is not only effected by the common Additive White Gaus-
sian Noise (AWGN) but also the multipath fading. Signal fading due to multipath
radio propagation is a dominant source of impairment in wireless communication
systems, often severely degrading performance [42, 43].

It is commonly assumed that each multipath component, in the received sig-

nal is independent then the Probability Density Function (PDF) of its envelope is
commonly modeled to be Rayleigh [7, 9, 44], when there is no Line–Of–Sight (LOS)
path5 . A typical signal fading envelope is shown in Figure 2.5.

−5

−10
Signal strength (dB)

−15

−20

−25

−30
0 0.5 1 1.5 2 2.5 3
Time (sec)

Figure 2.5: Typical profile of the received signals Rayleigh fading envelope.
The signal strength is severely faded.

If all the significant multipath components arrive at the receiver within a symbol
5
If there exists the LOS path, the PDF becomes Rician. This situation is not dealt with in
this research.
2.2. Characteristics of wireless communications 25

duration T , these components manifest themselves in a bunch with negligible delays

among them, and the mobile radio communication channel is called a “flat” Rayleigh
fading channel, now referred to as the narrowband channel. The impulse response
of the narrowband channel then consists of a single delta function whose weight has
a Rayleigh PDF [11].

If, on the other hand, the spread in delay among the multipath components is
not negligible (i.e. the symbol rate is sufficiently high), each symbol then spread
over adjacent symbols causing Intersymbol Interference (ISI). In such a situation,
the communication channel is referred to as a wideband channel [11].

2.2.2 Small–scale fading classification

Most challenges in WCom are prone to the situation of small–scale fading [7, 9,
42]. This fading represents changes in signal amplitude and phase that can be
experienced as a result of small changes (as small as a half–wavelength) in the spatial
separation between a receiver and a transmitter6 . Small–scale fading manifests itself
in two mechanisms: time–spreading (measured by delay time), and time–varying
(measured by transmission time), respectively [42].

The time–spreading behavior of the underlying digital pulses within the signal is
characterized by the delay resulting from the “non–optimum” impulse response of
the fading channel, through the notion of maximum delay τm in the delay domain, or
of channel coherence bandwidth Bc in the frequency domain. If signal bandwidth B
satisfies B  Bc ), the fading channel is said to be frequency–nonselective, otherwise,
frequency–selective [42].

The time–varying behavior of the channel is attributed to motion, where trans-

mission time is related to the antenna motion or spatial changes, through the notion
of channel coherence time Tc in the time domain or of Doppler spread νD in the
Doppler domain. If signal duration T satisfies T  Tc , the fading channel is said
6
Large–scale fading phenomenon such as path–loss is not dealt with in this research
26 Chapter 2. Overview

to be time–nonselective, otherwise, time–selective [42].

The classification of fading channels can be sketched as in Figure 2.6 [11]. Note
that, there are regions of distortion, in either time or frequency or both, in which
the underlying fading channel cannot be classified clearly.
Signal bandwidth

Time−nonselective Time−selective
Frequency−selective Frequency−selective




distortion


CO

Time−nonselective Time−selective
MM

Frequency−nonselective Frequency−nonselective
UNI
CAT
ION

IMPOS
SIBLE


 Signal duration


Figure 2.6: Channel classifications.

2.2.3 Mathematical analysis of wireless channels

Consider the transmission of a signal s̃(t), having bandwidth B and center fre-
quency fc , through a mobile radio propagation channel (assumed that B  fc ).
Due to multipath propagation, each component x̃p (t) of the received signal r̃(t) fol-
lowing the pth path will then be a replica of the transmitted waveform, delayed by
2.2. Characteristics of wireless communications 27

τp (t) seconds, attenuated by a factor Ap (t), and phase retarded (due to reflections
and diffractions) by θp (t) radians. Summing the received components over all the
propagation paths supported by the channel yields the total received signal [11]

P
( P )
X X
x̃(t) = x̃p (t) = < Ap (t) s̃(t − τp (t)) ej[2πfc (t−τp (t))−θp (t)] + η̃(t), (2.22)
p=1 p=1

where P is the number of paths comprising the channel, η̃(t) is the corrupting
bandpass AWGN, and < {·} denotes the real part.

Let s(t) be the equivalent complex lowpass signal of the transmitted signal s̃(t),
the equivalent complex lowpass received signal in (2.22) is then given by [11]:

P
X
x(t) = Ap (t) s(t − τp (t)) ej[2πfc (t−τp (t))−θp (t)] + η(t) (2.23)
p=1

where η(t) is the equivalent circular complex AWGN. where η(t) is a zero–mean
complex lowpass AWGN, having the variance of ση2 .

Model 2.1 (Narrowband channel).

Assume that, within a symbol duration T , τp (t) varies slowly (i.e. all τp (t) cluster
themselves around a constant delay τ ), the received signal in the general multipath
equation in (2.23) is approximated as in the following (multiplicative) model [7]:

x(t) = µ(t) · s(t) + η(t) = A(t)ejθ(t) · s(t) + η(t) (0 < t ≤ T ), (2.24)

where the random7 amplitude process A(t) is Rayleigh distributed, the random phase
variable θ(t) is uniformly distributed over [0, 2π).

The fading process µ(t) has the autocorrelation and corresponding spectrum
as shown in Figure 2.7 This model corresponds to the TF nonselective fading as
characterized in Section 2.2.2. Its impulse response becomes a single delta function
with Rayleigh weight as mentioned in Section 2.2.1.
7
All random terms are, hereafter, underlined to distinguish from deterministic terms.
28 Chapter 2. Overview

2.5

2
Power spectral density Sµ(f)

1.5

0.5

0
−1 −0.5 0 0.5 1
Normalised frequency (f/fm)

(a) Power spectral density

1
Autocorrelation Rµ(τ)

0.5

−0.5
0 1 2 3 4 5
Time lag τ [s]
(b) Autocorrelation

Figure 2.7: Plots of Jakes power spectral density and its autocorrelation.

If, on the other hand, the variation of delay is not negligible, one needs to find a
way to approximate the general multipath model in (2.23). In a mobile environment,
it is normally impossible to establish the exact value of P . The summation over the
number of paths in (2.23) is therefore replaced by the integral over all the possible
2.2. Characteristics of wireless communications 29

delays. This leads to the following model [11] characterizing the wideband channel:

Model 2.2 (Wideband channel).

Z ∞
x(t) ≈ h(t, τ ) s(t − τ ) dτ + η(t), (2.25)
0

where h(t, τ ) is the time–varying impulse response defined as:

∆
h(t, τ ) = aτ (t) e−j[2πfc τ +θτ (t)] (0 < t ≤ T ), (2.26)

with
   
 X  X
 jθ (t)

aτ (t) =  ap (t) e p ,
 and θτ (t) = arg  ap (t) ejθp (t)  . (2.27)
τp (t)=τ  τp (t)=τ

Equation (2.25) of Model 2.2 indicates that the fading channel exhibits a LTV
characteristics. Note that h(t, τ ) is physically interpreted as the response of the
LTV channel at time t due to a unit impulse input τ seconds in the past.

Applying the FT to both sides of (2.25) gives:

ZZ
x(t) = U (τ, ν)s(t − τ ) ej2πνt dτ dν + η(t) (2.28)

where ν denotes the Doppler–shift variable; U (τ, ν), obtained by Fourier transform-
ing h(t, τ ) from t to ν, is called the delay–Doppler–spread function of channel. Note
that, in practice, the double integral is bounded by the finite ranges of multipath–
delay τ and Doppler–shift ν variables, however, without loss of generality, a full
range of (−∞, ∞) is used for integration in both τ and ν, hence dropped out for
short notation.

By applying the Fourier transform among the variables t, f , τ and ν, a set of

system functions was laid out in a framework by Bello [11,24] with their relationship
illustrated in Figure 2.8: in which h(t, τ ), T (t, f ), H(ν, f ) and U (τ, ν) are the input–
delay–spread (time–varying impulse response) function, the time–varying transfer
30 Chapter 2. Overview

function, output–Doppler–spread function and delay–Doppler–spread function of

the channel, respectively.




FT

FT



 
FT

FT


 

Figure 2.8: Relationship among Bello’s system functions of an LTV channel.

h(t, τ ), T (t, f ), H(ν, f ) and U (τ, ν) are the channel input–delay–spread (time–
varying impulse response) function, the time–varying transfer function, output–
Doppler–spread function and delay–Doppler–spread function, respectively. A
Fourier Transform (FT) operation from one to another follows the arrow and
with respect to the different variables between the underlying functions, for ex-
ample FT of h(t, τ ) from t to ν gives U (τ, ν).

The wideband channel is commonly considered to be random [11]. The exam-

ination of the SOS of the channel involves the determination of the correlation,
for example the correlation function of the random input–delay–spread function is
given by:
∆
Rh (t1 , t2 ; τ1 , τ2 ) = E {h∗1 (t1 , τ1 ) h2 (t, τ2 )} . (2.29)

In practice, the random wideband channel is commonly assumed to be a zero–

mean Wide–Sense Stationary Uncorrelated Scattering (WSSUS) Gaussian random
process [11,24]. That is, Rh (t1 , t2 ; τ1 , τ2 ) is not only a wide–sense stationary in time
2.2. Characteristics of wireless communications 31

domain, as such


Rh (t1 , t2 ; τ1 , τ2 )WSS ≡ Rh (∆t; τ1 , τ2 ) (∆t = t2 − t1 ), (2.30)

but also have uncorrelated scattering in the delay domain, as such


Rh (t1 , t2 ; τ1 , τ2 )US ≡ Rh (t1 , t2 ; τ2 )δ(τ2 − τ1 ). (2.31)

Therefore


Rh (t1 , t2 ; τ1 , τ2 )WSSUS ≡ Rh (∆t; τ2 )δ(τ2 − τ1 ) (∆t = t2 − t1 ). (2.32)

The WSSUS assumption can also be expressed through the correlation of the ran-
dom delay–Doppler–spread function U (τ, ν) as:


RU (τ1 , τ2 ; ν1 , ν2 )WSSUS = RU (τ2 ; ν2 ) δ(τ2 − τ1 )δ(ν2 − ν1 ). (2.33)

It follows that the wideband channel under the WSSUS assumption may be repre-
sented as a collection of non–scintillating uncorrelated scatterers which cause both
multipath–delays and Doppler–shifts. The function RU (τ, ν) is called SF, and we
denote it now as Ψ(τ, ν) since it will be of great use later.

2.2.4 Practical assumptions of wireless channels

Restrictions of practical channels on time duration, bandwidth, fading rate, delay-

spread, Doppler–spread, etc. allows a simplified representation of LTV channels
in terms of canonical elements, hence the terms canonical channel models [24]. A
common class of canonical channel representations is the class of sampling models
which are applied when a system function vanishes for values of an independent
variable (t, f , τ , or ν) outside some interval or when the input or output time
function is time–limited or band–limited. Various sampling models can be found
in [24]. We, however, consider here the situation wherein the input signal s(t)
32 Chapter 2. Overview

Bi Bi
is band–limited in [fi − 2
, fi + 2
] (Bi is the input bandwidth) and the output
To To
signal x(t) (noise–free) is time–limited in [to − 2
, to + 2
] (To is the output time
duration). It is analytically [24] to say the equivalence that T (t, f ) vanishes outside
To To Bi Bi
the intervals [to − 2
, to + 2
] and [fi − 2
, fi + 2
].

Applying the sampling theorem on both t and f according two the above time
and frequency constraints, the received signal may be expressed as [24]:
 
XX m
x(t) = Umn s t − ej2π(n/To )(t−m/Bi ) , (2.34)
m n
Bi

where  
1 m n
Umn = · Usm , , (2.35)
Bi To Bi To
with U mn (τ, ν) being a smoothed version8 of U (τ, ν).

When the channel is random, the coefficients Umn are considered to be random
variables, now denoted by U mn . Under the WSSUS assumption, the correlation of
U mn can be expressed in terms of the SF Ψ(τ, ν). Additionally, if the SF varies very
little for changes in τ of the order 1/Bi or ν of the order 1/To , this correlation is
approximated as [24]:
  
1 m n

 ·Ψ , for m = n, r = s
E {U ∗mn U rs } = Bi To Bi To (2.36)

0 otherwise

which means that the delay–Doppler coefficients U mn are uncorrelated at different

values of multipath–delays and Doppler–shifts. The channel is called Quasi WSSUS
(QWSSUS) [24].

In practice, the multipath-delay τ vanishes outside [0, Tm ] and Doppler–shift

ZZ
U mn (τ, ν) = U (ζ, λ) · ej2πfi (τ −ζ) e−j2πto (ν−λ) sinc[Bi (τ − ζ)] sinc[To (ν − λ)] dλ dζ.
2.2. Characteristics of wireless communications 33

vanishes outside [−νD , νD ], where Tm and νD are the maximum multipath–delay and
maximum Doppler–shift, respectively. Consequently, the practical representation of
the QWSSUS channel through the above correlation may be sketched on the delay–
Doppler plane as shown in Figure 2.9 [45].

Doppler





Delay
 





Figure 2.9: Practical QWSSUS wireless channel.

Tm and νD are the maximum multipath–delay and Doppler–shift for a particular
mobile environment, respectively.
References

[1] B. Boashash, ed., Time-Frequency Signal Analysis: Methods and Applications. Mel-
bourne, Australia: Longman Cheshire, 1992.

[2] L. Cohen, Time-Frequency Analysis. Englewood Cliffs, NJ: Prentice-Hall, 1995.

[3] P. Flandrin, Time-Frequency / Time-Scale Analysis, Volume 10. Academic Press,

1998.

[4] F. Hlawatsch, Time-Frequency Analysis and Synthesis of Linear Signal Spaces:

Time-Frequency Filters, Signal Detection and Estimation, and Range-Doppler Es-
timation. USA: Kluwer Academic Publishers, 1998.

[5] B. Boashash, ed., Time Frequency Signal Analysis and Processing: Method and Ap-
plications. Oxford: Elsevier, 2003.

[6] H. V. Poor and G. W. Wornell, eds., Wireless Communications: Signal Processing

Perspectives. New Jersey: Prentice-Hall, 1998.

[7] W. Jakes, ed., Microwave Mobile Communications. IEEE Press, 1974.

[8] J. G. Proakis, Digital Communications. McGraw–Hill, 3rd. ed., 1995.

[9] T. S. Rappaport, Wireless Communications: Principles and Practice. New Jersey:

Prentice-Hall, 1996.

[10] W. Y. Lee, Mobile communications engineering. McGraw-Hill, 2nd. ed., 1998.

[11] R. Steele and L. Hanzo, eds., Mobile Radio Communications: Second and Third
Generation Cellular and WATM Systems. England: John Wiley & Sons, 2nd ed.,
1999.

[12] B. Boashash, “Estimating and interpreting the instantaneous frequency of a signal-

Part 1: Fundamentals,” Proceedings of the IEEE, vol. 80, pp. 519–538, Apr. 1992.

[13] B. Boashash, “Estimating and interpreting the instantaneous frequency of a signal-

Part 2: Algorithms and applications,” Proceedings of the IEEE, vol. 80, pp. 539–569,
Apr. 1992.

[14] A. Rihaczek, Principles of High-Resolution Radar. Peninsula Publishing, 1985.

34
References 35

[15] L. B. Milstein, “Interference rejection techniques in spread spectrum communica-

tions,” Proceedings of the IEEE, pp. 657–671, June 1988.

[16] M. G. Amin, “Interference mitigation in spread spectrum communication systems

using time-frequency distributions,” IEEE Transactions on Signal Processing, vol. 45,
pp. 90–101, Jan. 1997.

[17] L. Cohen, “Generalized phase-space distribution functions,” Journal of Mathematical

Physics, vol. 7, pp. 781–786, May 1966.

[18] H. Choi and W. Williams, “Improved time–frequency representation of multicom-

ponent signals using exponential kernels,” IEEE Transactions on Signal Processing,
vol. 37, pp. 862–871, June 1989.

[19] Z. M. Hussain, Adaptive instantaneous frequency estimation: Techniques and algo-

rithms. PhD thesis, Queensland University of Technology, Brisbane, Australia, 2002.

[20] G. Matz and F. Hlawatsch, “Time-frequency transfer function calculus (symbolic

calculus) of linear time-varying systems (linear operators) based on a generalized
underspread theory,” Journal of Mathematical Physics, vol. 39, pp. 4041–4070, Aug.
1998.

[21] A. K. Jain, Fundamentals of Digital Image Processing. Englewood Cliffs, New York:
Prentice-Hall, 1989.

[22] B. Ristic, Some aspects of signal dependent and higher-order time-frequency and
time-scale analysis of non-stationary signals. PhD thesis, Signal Processing Research
Centre, Queensland University of Technology, Brisbane, Australia, 1995.

[23] C. E. Cook and M. Bernfeld, Radar Signals: An Introduction to Theory and Appli-
cation. Norwood, MA: Artech House, 1993.

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

actions on Communication Systems, vol. COM-11, pp. 360–393, Dec. 1963.

[25] F. Hlawatsch and G. Matz, “Quadratic time-frequency analysis of linear time-varying

systems,” in Wavelet Transforms and Time-Frequency Signal Analysis (L. Debnath,
ed.), ch. 9, Boston (MA): Birkhäuser, to appear in September 2000.

[26] W. Mecklenbräuker and F. Hlawatsch, eds., The Wigner Distribution – Theory and
Applications in Signal Processing. Amsterdam, Netherlands: Elsevier, 1997.

[27] N. T. Gaarder, “Scattering function estimation,” IEEE Transactions on Information

Theory, vol. IT-14, pp. 684–693, Sept. 1968.

[28] R. A. Altes, “Detection, estimation, and classification with spectrograms,” The Jour-
nal of the Acoustical Society of America, vol. 67, pp. 1232–1246, Apr. 1980.

[29] P. Flandrin, “A time-frequency formulation of optimum detection,” International

Conference on Acoustics, Speech, and Signal Processing, ICASSP’98, vol. 36,
pp. 1377–1384, Sept. 1988.
36 Chapter 2. Overview

[30] L.-T. Nguyen, B. Senadji, and B. Boashash, “Scattering function and time-frequency
signal processing,” in International Conference on Acoustics, Speech, and Signal
Processing, ICASSP’2001, vol. VI, (Salt Lake city, Utah, USA), pp. 3597–3600, June
2001.

[31] G. F. Boudreaux-Bartels and T. W. Marks, “Time-varying filtering and signal esti-

mation using Wigner distributions,” IEEE Transactions on Acoustics, Speech, and
Signal Processing, vol. 34, pp. 422–430, Mar. 1986.

[32] B. Boashash, “Time-frequency signal analysis,” in Advances in Spectrum Analysis

and Array Processing (S. Haykin, ed.), vol. I, ch. 9, pp. 418–517, Englewood Cliffs,
New Jersey: Prentice-Hall, 1991.

[33] T. J. McHale and Boudreaux-Bartels, “An algorithm for synthesizing signals from
partial time-frequency models using the cross Wigner distribution,” IEEE Transac-
tions on Signal Processing, vol. 41, pp. 1986–1990, May 1993.

[34] J. C. Wood and D. T. Barry, “Linear signal synthesis using the Radon-Wigner trans-
form,” IEEE Transactions on Signal Processing, vol. 42, pp. 2105–2111, Aug. 1994.

[35] F. Hlawatsch and W. Krattenthaler, “Signal synthesis algorithms for bilinear time-
frequency signal representations,” in The Wigner Distribution - Theory and Applica-
tions in Signal Processing (W. Mecklenbräuker and F. Hlawatsch, eds.), pp. 135–209,
Amsterdam, Netherlands: Elsevier, 1997.

[36] A. Francos and M. Porat, “Analysis and synthesis of multicomponent signals us-
ing positive time-frequency distributions,” IEEE Transactions on Signal Processing,
vol. 47, pp. 493–504, Feb. 1999.

[37] L. V. White and B. Boashash, “On estimating the instantaneous frequency of a

Gaussian random signal by use of the Wigner–Ville distribution,” IEEE Transactions
on Acoustics, Speech, and Signal Processing, vol. 36, pp. 417–420, Mar. 1988.

[38] B. Boashash and P. O’Shea, “Use of the cross Wigner-Wille distribution for estima-
tion of instantaneous frequency,” IEEE Transactions on Signal Processing, vol. 41,
pp. 1439–1445, Mar. 1993.

[39] L. J. Stankovic and V. Katkovnik, “Algorithm for the instantaneous frequency esti-
mation using the time–frequency distributions with adaptive window length,” IEEE
Signal Processing Letters, vol. 5, Sept. 1998.

[40] V. Katkovnik and L. J. Stankovic, “Instantaneous frequency estimation using the

Wigner distribution with varying and data driven window length,” IEEE Transac-
tions on Signal Processing, vol. 46, pp. 2315–2325, Sept. 1998.

[41] B. Barkat, Design, estimation, and performance of time–frequency distributions. PhD

thesis, Queensland University of Technology, Brisbane, Australia, 2000.

[42] B. Sklar, “Rayleigh fading channels in mobile digital communication systems, part
I: Characterization,” IEEE Communications Magazine, pp. 90–100, July 1997.

[43] B. Sklar, “Rayleigh fading channels in mobile digital communication systems, part
II: Mitigation,” IEEE Communications Magazine, pp. 102–109, July 1997.
References 37

[44] S. Rice, “Mathematical analysis of random noise,” Bell System Technical Journal,
vol. 23, pp. 282–332, July 1944.
[45] A. M. Sayeed, A. Sendonaris, and B. Aazhang, “Multiuser detection in fast-fading
multipath environments,” IEEE Journal on Selected Areas in Communications,
vol. 16, pp. 1691–1701, Dec. 1998.
 Chapter 3

Instantaneous Frequency Estimation

in Multiplicative Noise

3.1 Introduction

FM signals are used in many engineering applications, such as in radar, sonar,

acoustic emission and communications [1]. Such signals contain the intended infor-
mation in the frequency content. Let us consider, as an example, the application
in telecommunications where an information–bearing signal is sent through a com-
munication channel. Depending on the real–life conditions of the communication
channel, the transmitted FM signal is often attenuated and corrupted by different
noise processes (e.g. additive noise, impulsive noise) and interference agents (other
signals in the same communication channel or from other channels). At the receiv-
ing end, one needs to recover the intended information from the observed signal
having been corrupted; for the case of an FM signal in particular, that is the fre-
39
40 Chapter 3. Instantaneous Frequency Estimation in Multiplicative Noise

quency content. Accurate frequency estimation leads to accurate recovery of the

true information.

In many cases, it is assumed that the received FM signal has a deterministic

amplitude over the duration of observation. While this is a valid assumption in a
wide range of scenarios, there are important applications under which this assump-
tion is inappropriate; the received signal has, instead, a random amplitude. This
phenomenon appears in radar imaging due to the presence of speckle noise [2], in
weather radar [3] and underwater acoustic [4, 5] applications due to the fluctua-
tion of the medium, in wireless mobile communications due to multipath propaga-
tion [6, 7], in power applications due to random plane wave field driving electrical
transmission line [8]. In these applications, the random amplitude modulation shows
up when the medium is dispersive and/or when the assumption of point target is
not valid [4, 9]. In the literature, the original signal is often said to be corrupted
by an MN process. Different underlying media result in different structures of the
MN. Anyhow, the recovery process of the information–bearing frequency becomes
more difficult.

This chapter aims to investigate the problem of frequency estimation of FM

signals corrupted by such an MN process. The chapter is organized as follows.
Section 3.2 gives a literature review on the existing methods of frequency estimation
in MN and addresses the motivation of this research. Section 3.3 mathematically
describes the underlying signal model and states assumptions that will be followed
in the subsequent sections. Section 3.4 analyses the MN in two different cases
concerning its phase and proposes a TFD–based IF estimator in a general situation.
Section 3.5 proposes a solution to improve the proposed estimator by the use of
the Multiple Signal Classification (MUSIC) algorithm. Simulations are provided
to analyze the statistical performance of the proposed estimator and its improved
version in terms of bias and variance while comparing them between different TFDs.
Section 3.6 is for conclusions and future directions.
3.2. Literature review for multiplicative noise 41

3.2 Literature review for multiplicative noise

The frequency estimation of FM signals observed in MN has drawn much research

attention in the last decade. Various approaches and techniques have been devel-
oped resulting in various estimators. We may categorize the research literature
on the MN problem in two directions based on the main characteristic of the MN
being a: (i) real –valued process, or (ii) complex –valued process. The real–valued
and complex–valued model of the MN models differ fundamentally in how the noise
affect the frequency estimation [10].

Existing studies focus mainly on the assumption that the underlying MN is a

real–valued random process. Dwyer [4,5] uses the fourth–order cumulant to estimate
a sinusoid in white Gaussian MN. Boashash and Ristic [11, 12] also address the
white Gaussian MN and use a nonparametric approach based on Time–Frequency
Distributions (TFDs) (in particular, the cumulant of Wigner–Ville Trispectrum, i.e.
a HOS approach) to estimate the frequency of nonstationary signals. Besson and
Castanié [13] use on an autoregressive model for the MN. Shamshunder et al. [14]
propose a parametric approach to the estimation of polynomial phase signals based
on higher-order ambiguity function (HAF) and cyclostationarity. Approaches using
HOS are also shown in [15] by Swami, and in [16, 17] by Zhou and Giannakis.
Friedlander and Francos [18, 19] investigate multicomponent harmonics. Besson
et al. [20] estimate the IF of chirp signals observed in time–varying MN using
nonlinear least square (NLS) approach, and combined HAF–NLS approach. Further
development on the abovementioned nonparametric approach are seen in [21] by
Senadji and Boashash when the MN process is Rayleigh in the context of mobile
communications using the HOS of TFDs. Morelande and Zoubir investigate the
model selection [22] of the MN in the case of polynomial phase signals, and further
assess the performance of cyclic–moment based estimators [23]. The Cramé–Rao
Lower Bounds (CRBs) for harmonic estimation in MN are addressed in [14, 17, 24],
and closed–form expressions are derived by Ghogho et al. [25] for exact and large
42 Chapter 3. Instantaneous Frequency Estimation in Multiplicative Noise

sample FM signals. It can be seen that the real–valued MN model has been treated
in great detail.

On the other hand, the complex–valued MN model is only considered recently

by Ghogho et al in [10] in which the CRBs for harmonic estimation are derived in
the effect of Doppler spread. Barkat [26] proposed the use of polynomial Wigner–
Ville distribution to estimate the frequency of nonstationary signals while explicitly
dealing with the assumption that the complex MN process has a non–zero mean.
It is noted that the zero–mean and nonzero–mean Gaussian MN cases should be
treated differently, as seen in [17]. In the nonzero–mean case the frequency content
of the signal is coupled in the mean of the observed signal which results in significant
energy concentration along the true frequency content of the signals, as shown
in [26]. However, in the zero–mean case the frequency content could be “lost” by the
smearing of the MN spectrum due to the Doppler effect. This more challenging zero–
mean case is addressed by Azemi et al. in [27] in which a TFD–based first–order
moment estimator is proposed while the MN spectrum follows the Jakes shape [28].
This method, however, fails when the SNR is below 10 dB.

We are, thus, motivated to investigate the frequency estimation of FM signals

when they are corrupted by a complex–valued zero–mean MN process. Our focus
is for the case of nonstationary FM signals, in particular the LFM signals1 . Here,
the notion of frequency estimation becomes that of IF estimation since the fre-
quency content of a nonstationary FM signal varies in time. By focusing on the
nonstationary case, one can explicitly take advantage of the powerful tool of TFSP.
Therefore, the underlying research follows up the nonparametric TFD–based ap-
proach initiated by Boashash and Ristic [11] above. Importantly, we assume that
the autocorrelation function of the MN is real–valued but not necessarily positive,
i.e. the spectrum of the MN is symmetric but does not necessary has the highest
peak at zero frequency. This assumption is less restrictive than that made in [10].

1
Though presenting the simplest form of FM nonstationarity, LFM is both fundamental for
research as well as practical for applications such as radar [29] and communications [30]. The
choice of LFM signals is motivated, but not limited.
3.3. Signal model and assumptions 43

This allows the study to be applied to various real–life scenarios, for example, in
mobile communications with Jakes spectrum and in ionospheric/satellite commu-
nications with Gaussian spectrum.

3.3 Signal model and assumptions

We now mathematically present the observed signal model and state the necessary
assumptions. Let s(t) be our original analytic LFM signal, µ(t) be the MN, and
η(t) be the additive noise, respectively. Under the effect of the noise, the observed
signal, denoted as x(t), is given by the following general model:

x(t) = µ(t) · s(t) + η(t), 0 < t ≤ T, (3.1)

where
n α o
s(t) = exp j2π(fo t + t2 )
2
µ(t) = Aµ (t)ejθµ (t) ,

Aµ (t) and θµ (t) being the amplitude and phase processes of the MN, respectively.
Our objective is to recover the IF, fin (t) = fo + αt, from the observed signal x(t). In
latter analyses, we will omit, except in the simulations, the effect of additive noise
since it is well–known that TFDs spread the additive noise power in TF domain
thus amounts to increasing the SNR [31].

The following assumptions are made:

As1) η(t) is assumed to be a zero–mean complex white Gaussian random process

with variance ση2 .

As2) µ(t) is assumed to be a zero–mean complex Gaussian random process with

variance σµ2 . Since µ(t) is zero–mean Gaussian process, the amplitude random
process Aµ (t) has a Rayleigh distribution, and the phase random process θµ (t)
44 Chapter 3. Instantaneous Frequency Estimation in Multiplicative Noise

is uniformly distributed over [0, 2π) [32]. Assume also that Aµ (t) and θµ (t)
are independent processes.

As3) µ(t) is wide–sense stationary, i.e. its autocorrelation function does not depend
on time; Rµ (t, τ ) ≡ Rµ (τ ).

As4) Rµ (τ ) is real–valued. Hence, the spectrum of the MN Sµ (f ), which is the

Fourier transform of Rµ (τ ), is symmetric.

3.4 IF estimation based on TFD

We recall here an important result found by Boashash and Ristic [11] expressing the
Wigner–Ville Spectrum (WVS) (i.e. the SOS of the WVD) of the observed noiseless
signal, Sxwvd (t, f ), in terms of the MN spectrum, Sµ (f ), when µ(t) is real–valued:

Sxwvd (t, f ) = Sµ (f − fin (t)), (3.2)

where
Z ∞
∆
n τ ∗ τ o −j2πf τ
Sxwvd (t, f ) = ρwvd

E x (t, f )
= E x(t + ) x (t − ) e dτ
−∞ 2 2
∆
n o n o
∆
n n τ τ oo
Sµ (f ) = FT Rµ (τ ) ≡ FT Rµ (t, τ ) = FT E µ(t + ) µ( t − ) .
τ →f τ →f τ →f 2 2

Equation (3.2) gives rise to the following remarks:

1. On the WVS representation of the observed signal, the MN spectrum follows

(varies in time according to) the IF law fin (t).

2. At each time instant, the IF is spread according to the entire MN spectrum,

thus the whole WVS is a representation of a time–varying MN spectrum
according to the IF.

3. From remarks (1) and (2), the possibility of the IF estimation from the WVS
3.4. IF estimation based on TFD 45

using the peak–detection over the entire TF domain relies on the structure of
the spectrum of the MN:

(a) If the MN spectrum is flat (i.e. white) the peak representing the IF at
any time instant is not resolvable; hence the estimation is impossible.

(b) If, on the other hand, the MN spectrum is colored, we may be able to
estimate the IF using peak–detection by exploiting the particular shape
of the MN spectrum.

Applying a similar approach in treating additive white noise, Boashash and Ris-
tic [11] assumed the MN to be white (as in the remark (3.a)). As a consequence,
the WVS failed to estimate the IF, hence the use of the cumulant of Wigner–Ville
trispectrum was proposed.

By observing the remark (3.b) in the context of mobile communications, we

explicitly exploit the “colored”ness of the MN and will show next that one is able
to achieve the IF estimation using an SOS approach (WVS) rather than the above
HOS one. As an example, if the highest peak of Sµ (f ) is sharp and centered at
frequency f = 0, then the WVS of the noise–corrupted signal x(t) will follow the
IF of the LFM signal s(t), thus facilitating the estimation of the IF using peak–
detection.

Following the perfective of this thesis research, we will focus our analysis in the
context of wireless communications. More precisely, we consider the Rayleigh flat
fading scenario (see the narrowband channel model in Section 2.1) whose model of
the received signal resembles our model of the observed signal corrupted by MN
in (3.1). In the following, we split the analysis into two cases in which the phase
of the MN is: (i) a random variable, and (ii) random process. The reason stems
from the fact that the former case, unlike the latter, proposes the same way to
achieve the IF estimation as in the case of real–valued MN. In other words, the
random phase, being a random variable, does not affect the estimation of the IF.
This reason will become clear later in our analysis.
46 Chapter 3. Instantaneous Frequency Estimation in Multiplicative Noise

3.4.1 Case I: Noise phase as a random variable

In the case where the phase of the MN is assumed to be a random variable, the
observed signal in Model (3.1) becomes:

x(t) = Aµ (t)ejθµ · s(t) + η(t), 0 < t ≤ T. (3.3)

Computing the WVS of x(t) in (3.3) above, leads to the following result:

Sxwvd (t, f ) = SAµ (f − fin (t)), (3.4)

where SAµ (f ) is the amplitude spectrum of the MN. Note that (3.4) is expressed
in terms of the MN amplitude spectrum whereas (3.2) is expressed in terms of the
total (i.e. amplitude and phase) MN spectrum.

Proof. (of Result (3.4))

The WVS of the LFM signal given in (3.1) is found to be:

Z ∞
∆
n τ τ o
Sxwvd (t, f ) = E x(t + ) x∗ (t − ) e−j2πf τ dτ
2 2
Z−∞
∞
= Rµ (t, τ ) e−j2π(f −fin (t))τ dτ,
−∞

where Rµ (t, τ ) is the MN autocorrelation expressed in the following form:

∆
n τ τ o
Rµ (t, τ ) = E µ(t + ) µ∗ (t − ) . (3.5)
2 2

Using (3.3) and applying Assumptions (As3, As4), we obtain:

n τ τ o
Rµ (t, τ ) = E Aµ (t + ) Aµ (t − ) = RAµ (t, τ ) ≡ RAµ (τ ). (3.6)
2 2
3.4. IF estimation based on TFD 47

Hence, applying the Wiener–Khintchine theorem [32] leads to the result:

Sxwvd (t, f ) = SAµ (f − fin (t)). (3.7)

We now apply the result in (3.4) to the Rayleigh narrowband environment where
the complex Gaussian fading process is the underlying MN; the amplitude process,
Aµ (t), follows the Rayleigh distribution. The mathematical model for the MN
spectrum, MN amplitude autocorrelation and MN amplitude spectrum are given in
the following expressions [28]:

• The MN spectrum:

b
Sµ (f ) = p [p(α)G(α) + p(−α)G(−α)], (3.8)
2 − (f − f )2
fm c

where fm is the maximum Doppler shift (fm = v/λ), α is the angle of the
incoming wave, p(α) is the incident power included in [α, α + dα] (p(α)dα
equals to the total incoming power within dα of angle α), G(α) is the antenna
power gain, b is the average power that would be received by an isotropic
1
antenna (G(α) = 1), and fc is the carrier frequency. When p(α) = 2π
, (−π ≤
α ≤ π) and the antenna gain G(α) = 1.5, we have the spectrum of the electric
field component given by:
   2 −1/2
 1.5b 1 − f −fc

|f | ≤ fm

πfm fm
Sµ (f ) = . (3.9)

0, |f | > fm


• The MN amplitude autocorrelation:

   
π 1 1 π 1
RAµ (τ ) = b0 F − ; − ; 1; ξ(τ ) ≈ b0 1 + ξ(τ ) . (3.10)
2 2 2 2 4

where b0 and ξ(τ ) are defined as in (3.12).

48 Chapter 3. Instantaneous Frequency Estimation in Multiplicative Noise

• The MN amplitude spectrum (of the electric field component):

s 
 2
π b0 f  = π b0 δ(f ) + π b0 Sξ (f ),
SAµ (f ) = b0 δ(f ) + K 1−
2 8πfm 2fm 2 8
(3.11)
where K[·] is the complete elliptic integral of the first kind. This spectrum is
symmetric and has its highest peak at frequency f = 0.

Therefore, from (3.4) and (3.11), Sxwvd (t, f ) represents an energy concentration along
the IF of the original signal s(t). It is this modeled structure of the Rayleigh
narrowband channel that suggests the estimation for the IF by peak–detecting the
WVS instead of using Wigner–Ville trispectrum as proposed in [11]. We will show
next some simulation results that confirm this remark. The estimation will be
deferred to in Section 3.4.2 in which we propose a TFD–based IF estimator for a
general case.

3.4.1.1 Simulations

The simulation of the MN process follows the theoretical model given in [28] as
below:

• The MN amplitude Aµ (t) is computed from:

q
Aµ (t) = T 2c (t) + T 2s (t)
N
X
T c (t) = Eo C i cos(ω i t + φi )
i=1
XN
T s (t) = Eo C i sin(ω i t + φi ).
i=1

T c (t) and T s (t) are random Gaussian processes corresponding to the in–phase
and quadrature components of Aµ (t), respectively. Eo2 /2 is the average mean
power of Aµ (t). The coefficients C i , ω i and φi are respectively the random
3.4. IF estimation based on TFD 49

attenuation following Rayleigh distribution, random Doppler shift uniformly

distributed over [−fm , fm ] (fm is the maximum Doppler–shift), and random
phase uniformly distributed over [0, 2π), of the ith path. N is the number of
paths arriving at the receiver. N was set equal to 50 to ensure the approxi-
mation of Gaussianity for T c and T s (according to the central limit theorem).
The mobile station was assumed to have a velocity of 60 km/h in an 850 MHz
mobile system.

• The function ξ(τ ) in (3.10) is formulated as:

1  2 2

ξ(τ ) = g (τ ) + h (τ ) , (3.12)
b20

where g(τ ) and h(τ ) being the autocorrelation of T c (t) and crosscorrelation
between T c (t) and T s (t), respectively, and bo = g(0).

The LFM signal (Figure 3.2(a)) was simulated using the following discrete form:
n αo o
s(n) = exp j2π(fo n + n2 ) , n = 1, . . . , T. (3.13)
2

The coefficients fo , αo were chosen so that the IF increases linearly from 10 Hz

to 40 Hz. The number of samples is T = 128 and the sampling frequency is
fs = 100 Hz.

The spectrum Sξ (f ) in (3.11) is simulated in Figure 3.1. The simulation confirms

that this spectrum is symmetric and has its highest peak at frequency f = 0.

The theoretical results in (3.4) are illustrated by the simulation in Figure 3.2(b)
using WVD (1 realization), and in Figure 3.2(c) using WVS (estimated over 10
realizations). The simulation shows that one is able to estimate the IF in MN using
the WVS.

We also apply other TFDs (rather then the WVD), e.g. the SPEC and the
CWD as shown in Figure 3.3. In comparison, the WVS is optimal, among the three
experimented TFDs, in the sense that it achieves the highest energy concentration
50 Chapter 3. Instantaneous Frequency Estimation in Multiplicative Noise

10

0
−50 −40 −30 −20 −10 0 10 20 30 40 50

Figure 3.1: Simulated spectrum Sξ (f )

Fs=1Hz N=128
Time−res=4

120

100

80
Time (secs)

60

40

20

0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5
Frequency (Hz)

(a) Test signal

Fs=100Hz N=128 Fs=100Hz N=128
Time−res=4 Time−res=4

1.2 1.2

1 1

0.8 0.8
Time (seconds)

Time (seconds)

0.6 0.6

0.4 0.4

0.2 0.2

5 10 15 20 25 30 35 40 45 50 5 10 15 20 25 30 35 40 45 50
Frequency (Hz) Frequency (Hz)

(b) WVD: 1 realization (c) WVS: 10 realizations

Figure 3.2: Simulation results for an LFM signal in instantaneous multiplicative

noise.
This example corresponds to a Rayleigh fading narrowband channel. The WVS
achieves energy concentration along the IF of the signal for 10 and more realiza-
tions.
3.4. IF estimation based on TFD 51

along the IF of the LFM signal in the TF plane, hence the optimal estimation of
the IF.

Fs=100Hz N=128 Fs=100Hz N=128

Time−res=4 Time−res=4

1.2 1.2

1 1

0.8 0.8
Time (seconds)

Time (seconds)
0.6 0.6

0.4 0.4

0.2 0.2

5 10 15 20 25 30 35 40 45 50 5 10 15 20 25 30 35 40 45 50
Frequency (Hz) Frequency (Hz)

(a) SPEC: 1 realization (b) SPEC: 10 realizations

Fs=100Hz N=128 Fs=100Hz N=128
Time−res=4 Time−res=4

1.2 1.2

1 1

0.8 0.8
Time (seconds)

Time (seconds)

0.6 0.6

0.4 0.4

0.2 0.2

5 10 15 20 25 30 35 40 45 50 5 10 15 20 25 30 35 40 45 50
Frequency (Hz) Frequency (Hz)

(c) CWD: 1 realization (d) CWD: 10 realizations

Figure 3.3: Simulation results for an LFM signal in instantaneous multiplicative

noise.
The spectrogram (top), and the Choi–Williams distribution (bottom) are used
to reveal the signal IF in the TF domain.

We have analyzed, in this section, the case when the MN phase is treated as a
random variable. This case could be physically interpreted as the mobile environ-
ment being “freezed”. Obviously, this situation is impossible in a practical wireless
mobile environment. However, the above analysis helps suggest the feasibility of
estimating the IF in the presence of the MN using the SOS of the TFDs. It is noted
that the MN phase was suppressed, thus, giving the same result as in the case of a
real–valued model of the MN. Next, in the second case, we will deal with a more
general case, which is practically valid, where the MN phase should be considered
as a random process instead.
52 Chapter 3. Instantaneous Frequency Estimation in Multiplicative Noise

3.4.2 Case II: Noise phase as a random process

When the MN phase, θµ (t), cannot be treated as a random variable (i.e. θµ ), the
analysis in Section 3.4.1 does not allow the suppression of this phase. Instead of
explicitly separating the amplitude Aµ (t) from the phase θµ (t), we now analyze this
general case in the full complex–valued form, µ(t).

The same result, as found by Boashash and Ristic [11] in (3.2) for the real–
valued model of the MN, can be easily derived for the complex–valued model of the
MN as given below:
Sxwvd (t, f ) = Sµ (f − fin (t)). (3.14)

However, a fundamental difference between (3.2) and (3.14) is that the MN spec-
trum in (3.14) does not necessarily have the highest peak at frequency f = 0 (an
example is shown in Figure 2.7.a). As a consequence, the true IF law could be
shifted to another location in the TF representation. Therefore, one could not es-
timate the IF simply by peak–detecting the WVS of the observed signal, as in the
real–valued case or even the alike–real–valued case in Section 3.4.1.

Based on the result in (3.14) we arrive to the following proposition.

Proposition 3.1. Let fp be the frequency at a peak of Sµ (f ), and ρx (t, f ) be a

TFD, which preserves the signal IF, of the observed signal x(t). Assume that the
noise spectrum Sµ (f ) changes negligibly in the observation duration 0 ≤ t ≤ T .
Then, the spectrum of the TFD, Sxρ (t, f ), represents an energy concentration along
(fin (t) + fp ).

Proof. Recall the expression of the quadratic class of TFDs in (2.14) for an analytic
signal x(t):
ρz (t, f ) = γ(t, f ) ? ? ρwvd
x (t, f ). (3.15)
t f
3.4. IF estimation based on TFD 53


 
  

IFe
Tru


    
         
(a) Peaks of Sµ (f ) (b) Shifted versions IF law

Figure 3.4: (a) A general sketch of the colored MN spectrum with a symmetric
shape, and (b) the TF representation of the observed signal in such a spectrum.

where γ(t, f ) is the TF TFD kernel. Applying (3.15) to (3.14) gives:

Sxρ (t, f ) = γ(t, f ) ? ? Sµ (f − fin (t)) (3.16)

t f

Corresponding to the time–frequency TFD kernel γ(t, f ) in the Doppler–lag do-

main is the Doppler–lag TFD kernel g(ν, τ ), which is obtained by doubly Fourier
transforming γ(t, f ).

In the simplest case when g(ν, τ ) = 1, ρx (t, f ) becomes ρwvd

x (t, f ), which is the

WVD; hence, γ(t, f ) = δ(t)δ(f ). Consequently, the above expression reduces to

Sxwvd (t, f ) ≡ Sxρ (t, f ) = Sµ (f − fin (t)) (3.17)

Obviously, Sxwvd (t, f ) follows the peak at fin (t) + fp .

In the general case, the Doppler–lag TFD kernel g(ν, τ ) will modify the structure
of Sµ (f − fin (t)) by means of double convolution in the TF domain. Since IF
estimation is of our concern in this study, the preservation of the IF through a
TFD transformation is desirable. Only a few TFDs satisfy the exact preservation
of the IF (e.g. WVD, CWD). Instead, others give a close approximation of the IF
while offering high TF resolution and/or reduced cross–terms. To make use of the
offerings of other TFDs, we relax ourselves to a subset of quadratic TFDs which
approximately preserve the IF of the signal.
54 Chapter 3. Instantaneous Frequency Estimation in Multiplicative Noise

With the stated subset of TFDs, the design of g(ν, τ ) allows the following ap-
proximation:
Sxρ (t, f ) ≈ Sµ0 (f − fin (t)) (3.18)

where Sµ0 (f − fin (t)) still preserves the IF but with a lower concentration of energy
(energy smearing) around the IF in comparison with the exact concentration at the
IF as in the case of WVD. Therefore, Sxρ (t, f ) follows the peak at fin (t) + fp .

Proposition 3.1 leads to the following result. Let f−P , . . . , fP be the frequencies
at each of which the MN spectrum has a peak. The TFD spectrum will then have
2P + 1 ridges corresponding to frequency–shifted versions of the IF law fin (t) with
the shifts being the frequency values f−P , . . . , fP , respectively. Note that if Sµ (f )
does not have a peak at f = 0 then Sxρ (t, f ) does not reveal the energy concentration
at the true IF law fin (t). This result is illustrated in Figure 3.4. We are now able
to propose a TFD–based IF estimator as shown in the next section.

3.4.2.1 IF estimator

Using Proposition 3.1 under Assumption (As4), ensuring the symmetry of Sµ (f ),

we propose the following general TFD–based IF estimator:

P P
1 X 1 X n ρ o
fˆin (t) = ˆ
fp (t) = arg max Ŝ x (t, f ) , (3.19)
2P + 1 p=−P 2P + 1 p=−P f

ρ
where Ŝ x (t, f ) is an estimate of the TFD spectrum over N realizations of the ob-
served signals x1 (t), . . . , xN (t):

N
ρ 1 X
Ŝ x (t, f ) = ρx (t, f ), (3.20)
N n=1 n

and the TFD used in the estimator satisfies the condition that it approximately
preserves the IF of the signal. Note that, in the above estimator, we have assumed
that the number of peaks are known a priori. This assumption is valid, for example,
3.4. IF estimation based on TFD 55

in the mobile channel where the Jakes model is used; in this case the number of
peaks is equal to two.

The block diagram of the IF estimator in (3.19) is illustrated in Figure 3.5.

 
  

IF Estimator block   

2  
 
 "! +-,  0/
1   
 

.

Average
Peak
#$%&
'( )*
Detection

 

Figure 3.5: Block diagram of the proposed TFD–based IF estimator.

3.4.2.2 Discussions

It is important to address the following issues concerning the performance of the

proposed TFD–based IF estimator:

• Spectral shape: The MN spectrum needs to have a symmetric shape, accord-

ing to Assumption (As4) in order for the proposed estimator to function. The
benefit of this estimator is that it takes into account any symmetric models
(with explicit peaks) of the MN spectrum, such as the Jakes spectrum in nar-
rowband mobile communications [28] (which has two peaks at f = −fm and
f = fm where fm is the Doppler–shift frequency), or the Gaussian spectrum
in ionospheric/satellite communications [33] (which has one peak at f = 0).
In other words, it does not depend on where the peaks are located. However,
it is obvious that the sharper the peaks, the better the performance of the
estimator.

• Choice of TFD: As stated in Proposition 3.1, we only consider a subset

of TFDs which approximately preserve the IF. Let fp be a peak on, for
56 Chapter 3. Instantaneous Frequency Estimation in Multiplicative Noise

example, the right–hand side of the symmetric MN spectrum. According to

Proposition 3.1, the signal IF is shifted to the right by an amount of fp .

If one is to estimate the signal IF by estimating the shifted IF then followed

with a shifting to the left, the performance of such an estimator depends on
how well the cross–terms are suppressed by the TFD and the high resolution
of the TFD if the estimator uses the peak-detection approach. Clearly, the
present of cross–terms may force the estimator to wrongly decide a peak at
the cross–terms as the peak of the true IF.

In addition, the higher the capability of suppressing the cross–terms leads

to the smaller the number of realizations needed. Theoretically, the TFD
spectrum Sxρ (t, f ) spreads the IF law in an exact accordance of the spectral
ρ
shape of the MN. However, in practice, the estimated TFD spectrum Ŝ x (t, f ),
as estimated in (3.20), depends on the number N of realizations available (N
is often small). As N decreases, the frequency resolution of the ridges on the
representation of the estimated TFD spectrum decreases.

Above are the reasoning for a need of a TFD with a good cross–term sup-
pression capability and high resolution if one was to use a one-side estimation
approach. However, as proposed in (3.19) and (3.20), our estimator averages
over the shifted versions of the IF on both sides. It will later shown that
the cross–term effect and high resolution may or may not problematic in the
estimation of the signal IF as they might seem to be. The result of this study
will illustrate this last statement.

3.4.2.3 Simulation

In the following simulation, we want to numerically analyse the bias and variance
performance of the proposed TFD–based IF estimator and compare the performance
in terms of four different TFDs: WVD, SPEC, CWD and MBD. The choice of TFDs
taken stems for the following reasons:
3.4. IF estimation based on TFD 57

• WVD has the highest energy concentration and is optimal for LFM signals,
however it suffers the cross–term problem.

• SPEC is robust and free of cross–terms, however it has low TF resolution.

• CWD reduces the cross–terms effectively and has a reasonalbly good resolu-
tion.

• MBD has a high resolution comparable to the WVD and reduces the cross–
terms effectively.

Note that, among the four TFDs chosen, WVD and CWD preserve the IF, whereas
the other provides an approximate preservation.

The MN process was simulated using the Jakes Doppler spectrum that models
the Rayleigh flat fading process in wireless narrowband mobile environment. The
MN spectrum Sµ (f ) of this model is expressed as follows [28]:

 2 −1
 " r #
σµ2 πfm 1 − f

, |f | ≤ fm

fm
Sµ (f ) = (3.21)


0, |f | > fm

where σµ2 is the variance of the MN. The corresponding MN autocorrelation is given
by:
Rµ (τ ) = σµ2 J0 (2πfm τ ) (3.22)

where J0 (·) denotes the zero–order Bessel function of the first kind. This MN
spectrum, shown in Figure 2.7, is symmetric and consists of 2 peaks allocated at
f = −fm and f = fm (there is no peak at f = 0). To simulate the above model,
we used the method of equal distances (MED) described in [34]. The maximum
Doppler–shift frequency was set equal to 10 Hz in all the simulations.

The original LFM signal was set up in the same way as in the simulation in Sec-
tion 3.4.1. The IF increases from 40 Hz to 60 Hz with the sampling frequency
58 Chapter 3. Instantaneous Frequency Estimation in Multiplicative Noise

fs = 200 Hz (or from 0.2 to 0.3 as normalized according to fs ). The number of

samples was T = 128.

Figure 3.6 show respectively the TF representation of the estimate TFD spectra
over 10 realizations and the corresponding IF estimates using the WVD, SPEC,
CWD and MBD. The dotted lines represent the true IF law and the solid lines
represent the estimated IF law. Figure 3.7 shows the bias and variance of the
proposed estimator versus the SNR ratio using Monte Carlo simulation with 1000
runs. The results were taken at the middle time–slide of the TFD spectra (we follow
the common way of analysing the performance of IF estimation at this time–slide).

Following are some remarks from the simulated results:

• For all different TFDs, it is seen that the estimated IF follows the true IF law.

• The “wiggling” pattern of the estimated IF indicates that the proposed esti-
mator needs to be improved. This is due to the frequency spreading of the
MN spectrum between the two peaks. The estimated TFD spectra were only
averaged over 10 realizations which did not give a close estimation of the true
TFD spectra as depicted in Proposition 3.1, hence the estimation of the IF.

• Nonetheless, it can be seen that the CWD and MBD give a clearer indication
of the shifted versions (on the most left and right sides of the representation)
of the true IF law. This comes from the reasons stated above in choosing the
four different TFDs for comparison.

• The performance of bias and variance is shown in Figure 3.7. A general

observation is that all TFDs are comparable for high SNR (>20dB). For
low SNR, the MBD performs the best in bias (Figure 3.7(a)) and the CWD
performs the best in variance. This result is not worth of analyzing since the
wiggling patterns of the estimated signal IF has indicated the need to improve
the estimator. Nonetheless, another result which is more interesting is that
the proposed estimator does allow the estimation of IF at low SNR whereas
3.5. TFD–based IF estimation with MUSIC 59

the method using the first–order moment in [27] fails when the SNR is less
than 10 dB.

Note that, the values of errors were normalized according to the sampling
frequency. This normalization does not indicate the absolute errors in order
to compare the proposed estimator with an existing estimator. Instead, it
is valid when comparing the performance of the proposed estimator among
different TFDs.

3.5 TFD–based IF estimation with MUSIC

It was noted from the simulation results in the previous section that the estimated
IF did not follow closely (i.e. wiggling pattern) the true IF. Expectedly, one would
achieve a better estimation by increasing the number of realizations to estimate
the TFD spectrum. However, in practice, we have a limited number of realizations
(or, often the case, just one realization). Thus, it is motivated to improve the
estimation of the IF while still using the same number of realizations. Instead
of directly estimating the peaks from the TFD spectrum, we propose to estimate
them using the well–known MUSIC algorithm [35] on each time slice of the TFD
spectrum. The block diagram of this modified estimator is illustrated in Figure 3.8.

Figure 3.9 respectively show the TF of the TFD spectrum estimates over 10
realizations and the corresponding IF estimates using the WVD, SPEC, CWD and
MBD. The dotted lines represent the true IF law and the solid lines represent the
estimated IF law. Figure 3.10 shows the bias and variance of the proposed estimator
versus the signal–to–additive–noise ratio using Monte Carlo simulation with 1000
runs. The results were taken at the middle time–slide of the TFD spectra.

Following are some remarks from the simulated results with the addition of the
MUSIC algorithm:
60 Chapter 3. Instantaneous Frequency Estimation in Multiplicative Noise

• Comparing the right–hand sides of Figure 3.6 and Figure 3.9 indicates that
the addition of the MUSIC algorithm gives a better estimation of the IF.

• The distinct improvement in terms of bias was achieved by all TFDs. However,
the WVD and the CWD showed some degradation in terms of variance for
low SNR (0–10dB). In line with the advantage of the proposed estimator,
compared to the first–order moment method in [27], that it is able to estimate
the IF at low SNR (0–10dB), we will focus the interpretation of the simulation
results within this range, as below.

• Bias performance: comparing the performance of the SPEC (no cross–term)

and the WVD (cross–term), the cross–term effect appeared not to be prob-
lematic since the WVD performed better than the SPEC. Comparing the
performance of the SPEC (low resolution, no cross–term) and the MBD (high
resolution, almost no cross–term), the low resolution was not problematic.
Thus, in terms of bias, both cross–term and low resolution do not constitute
any problem in our estimation. The CWD was comparable with the MBD
and both were worse than the SPEC.

• Variance performance: high resolution was not reflected as being problem-

atic since both the SPEC (low resolution) and the MBD (high resolution)
performed much better than the WVD (high resolution). Instead, the cross–
term effect was significant here; the SPEC (no cross–term) and the MBD
(almost no cross–term) performed the best among all TFDs.

• Based on the last two remarks above, if one is ready to consider the effect
of both the high resolution and the cross–terms, the SPEC or the MBD will
then be in favor rather than the WVD and the CWD.

• If we then compare between only the SPEC and the MBD, the SPEC was
always better than the MBD in terms of bias, but worse than the MBD in
terms of variance, except for very low SNR (0–5dB). Here, we can only say
that, the SPEC and the MBD are better than the WVD and the CWD for
3.5. TFD–based IF estimation with MUSIC 61

SNR ranging from 0db to 10dB. However, the performance analysis was taken
only at the midle time–slide and if we relook at the estimates of IF of these
two TFDs in Figure 3.6, the MBD was better than the SPEC on the whole
duration of the signal. Therefore, we can conclude that the MBD performs
the best.
62 Chapter 3. Instantaneous Frequency Estimation in Multiplicative Noise

WVD using GRAD WVD using GRAD

120 120

100 100

80 80
time [slice]

60 60

40 40

20 20

0 0
0 10 20 30 40 50 60 70 80 90 100 0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5
Normalised frequency [f/fs] Normalised frequency [f/fs]

(a) WVD (b) IF estimate

SPEC using GRAD SPEC using GRAD

120 120

100 100

80 80
time [slice]

60 60

40 40

20 20

0 0
0 10 20 30 40 50 60 70 80 90 100 0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5
Normalised frequency [f/fs] Normalised frequency [f/fs]

(c) SPEC (d) IF estimate

CWD using GRAD CWD using GRAD

120 120

100 100

80 80
time [slice]

60 60

40 40

20 20

0 0
0 10 20 30 40 50 60 70 80 90 100 0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5
Normalised frequency [f/fs] Normalised frequency [f/fs]

(e) CWD (f) IF estimate

MBD using GRAD MBD using GRAD

120 120

100 100

80 80
time [slice]

60 60

40 40

20 20

0 0
0 10 20 30 40 50 60 70 80 90 100 0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5
Normalised frequency [f/fs] Normalised frequency [f/fs]

(g) MBD (h) IF estimate

Figure 3.6: TFD spectra of the observed signal in MN and the corresponding
IF estimates using the proposed IF estimator. The figures on the left–hand side
are the spectra estimated from: WVD, SPEC, CWD and MBD. The figures on
the right–hand side show the true IF (dotted lines) and the estimated ones (solid
lines).
3.5. TFD–based IF estimation with MUSIC 63

−3
x 10
10

Bias normalised w.r.t sampling frequency

MBD−GRAD
CWD−GRAD
WVD−GRAD
SPEC−GRAD
−5
0 5 10 15 20 25 30 35 40
SNR [dB]

(a) Bias
−2
10

−3
10
Variance normalised w.r.t sampling frequency [dB]

−4
10

−5
10

−6
10

−7
10
MBD−GRAD
CWD−GRAD
WVD−GRAD
SPEC−GRAD
−8
10
0 5 10 15 20 25 30 35 40
SNR [dB]

(b) Variance

Figure 3.7: Plot of bias and variance, for the proposed TFD IF estimator, versus
the SNR using Monte Carlo simulation with 1000 runs.
64 Chapter 3. Instantaneous Frequency Estimation in Multiplicative Noise

0 
1 &
IF Estimator block   

/ &
2 
 (*) &-,
. MUSIC  
 

+

Average
!"
#$&%' Peak
Detection

 

Figure 3.8: Block diagram of proposed TFD–based estimator using MUSIC al-
gorithm.
3.5. TFD–based IF estimation with MUSIC 65

WVD using MUSIC WVD using MUSIC

120 120

100 100

time [slice] 80 80

60 60

40 40

20 20

0 0
0 10 20 30 40 50 60 70 80 90 100 0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5
Normalised frequency [f/fs] Normalised frequency [f/fs]

(a) WVD (b) IF estimate

SPEC using MUSIC SPEC using MUSIC

120 120

100 100

80 80
time [slice]

60 60

40 40

20 20

0 0
0 10 20 30 40 50 60 70 80 90 100 0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5
Normalised frequency [f/fs] Normalised frequency [f/fs]

(c) SPEC (d) IF estimate

CWD using MUSIC CWD using MUSIC

120 120

100 100

80 80
time [slice]

60 60

40 40

20 20

0 0
0 10 20 30 40 50 60 70 80 90 100 0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5
Normalised frequency [f/fs] Normalised frequency [f/fs]

(e) CWD (f) IF estimate

MBD using MUSIC MBD using MUSIC

120 120

100 100

80 80
time [slice]

60 60

40 40

20 20

0 0
0 10 20 30 40 50 60 70 80 90 100 0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5
Normalised frequency [f/fs] Normalised frequency [f/fs]

(g) MBD (h) IF estimate

Figure 3.9: TFD spectra of the observed signal in MN and the corresponding
IF estimates using the proposed MUSIC–TFD IF estimator. The figures on
the left–hand side are the spectra estimated from: WVD, SPEC, CWD and
MBD. The figures on the right–hand side show the true IF (dotted lines) and
the estimated ones (solid lines). Improvement on the IF estimation are seen
(compared with Figure 3.6).
66 Chapter 3. Instantaneous Frequency Estimation in Multiplicative Noise

−3
x 10
10
Bias normalised w.r.t sampling frequency

MBD−GRAD
MBD−MUSIC
CWD−GRAD
CWD−MUSIC
WVD−GRAD
WVD−MUSIC
SPEC−GRAD
SPEC−MUSIC
−5
0 5 10 15 20 25 30 35 40
SNR [dB]

(a) Bias
−2
10

−3
10
Variance normalised w.r.t sampling frequency [dB]

−4
10

−5
10

−6
10

MBD−GRAD
MBD−MUSIC
−7
CWD−GRAD
10 CWD−MUSIC
WVD−GRAD
WVD−MUSIC
SPEC−GRAD
SPEC−MUSIC
−8
10
0 5 10 15 20 25 30 35 40
SNR [dB]

(b) Variance

Figure 3.10: Plot of bias and variance, for the proposed MUSIC–TFD IF estima-
tor, versus the SNR using Monte Carlo simulation with 1000 runs. Improvement
on the performance of IF estimation are seen (compared with Figure 3.7).
3.6. Conclusions 67

3.6 Conclusions

This chapter investigated the IF estimation of nonstationary FM signals, in par-

ticular LFM signals when they are corrupted by a complex–valued zero–mean MN
process. Our focus is for the case of nonstationary FM signals, in particular the
LFM signals using TFDs. The main assumption (As4) was that the autocorrelation
function of the MN is real–valued but not necessarily positive (i.e. the spectrum
of the MN is symmetric but does not necessarily have the highest peak at zero
frequency). We have proposed a TFD–based estimator for the IF using second–
order spectra rather than higher–order spectra. The estimation performance was
analyzed in terms of bias and variance, and compared between four different TFDs:
WVD, SPEC, CWD and MBD. Further to improve the estimation, we proposed to
use the MUSIC algorithm and showed its better performance. It was shown that
the MBD performed the best for low SNR (less than 10 dB).

Further research on this direction may look at the following issues:

• Multiplicative noise with asymmetric power spectral density (e.g. Rice spec-
trum). This is important in wideband wireless communications since there
is a mixture of Doppler spectra with different shapes at different multipath
delays.

• Investigation of IF estimation for nonlinear FM signals.

• For wideband mobile channel, the MN is no longer wide–sense–stationary as

assumed in our signal model. Instead, one should use the well–known assump-
tion: WSSUS [36]. In the investigation into the wideband channel model, we
have achieved some preliminary theoretical results for the estimation of the
scattering function under the WSSUS assumption (see Appendix 3A) and
further analysis of the IF estimation for wideband MN. However, this went
beyond the scope defined for this thesis and was left for future development.
68 Chapter 3. Instantaneous Frequency Estimation in Multiplicative Noise

APPENDIX 3A: Scattering function estimation

3A-1. Introduction

In Section 2.2, the fading wideband channel was reviewed. This channel undergoes the
frequency–selective fading affected by the time–spreading behavior of multipath–delays.
When affected also by the TV behavior of Doppler–shifts, the channel is said to be TF
selective. The statistics of the fading wideband channel is completely characterized by
its SF under the assumption of WSSUS Gaussian process [36]. The SF, as a function of
both multipath–delay τ and Doppler–shifts ν, explicitly reveals the TF selective behavior
of the fading wideband channel. Its marginal in τ yields the power delay profile, and in
ν yields the Doppler spectrum [37] both of which give important parameters for receiver
designs when one deals with either frequency–selective or time–selective fading. In the
presence of TV selective fading, the marginals do not provide enough information since
the channel behavior is now expressed as a two dimensional function of both τ and ν. In
practice, it is difficult to estimate the SF, and instead, one measures it empirically for a
particular wireless environment [7].

The estimation of the SF are seen in [37–44] (and references therein). A common
approach is to relate it with the symmetric AF [38] or Woodward AF [41,42] of the input
signal. However, a classical problem faced in this approach is the division of zero. To get
around this problem, thresholding and its derivatives were introduced (review of this can
be found in [41]).

The appendix investigates the estimation of the SF using TFSP. We propose two
classes of TFD–based estimators that generalize the existing estimators while giving an
extra freedom according to different criteria wanted to be achieved in the estimation of the
scattering function. That is, instead of using Woodward or symmetric AFs, we employ
the GAF. The GAF was introduced in Section 2.1.
Appendix 3A: Scattering function estimation 69

3A-2. Analysis on scattering function

Given the WSSUS assumption to the channel, it is important to examine the relationship
of the transmitted and received signals through their SOS. Recall the wideband model of
the received signal given in (2.25):
Z ∞
x(t) ≈ h(t, τ ) s(t − τ ) dτ + η(t), (3.23)
0

where h(t, τ ) is the TV impulse response, or in (2.28)

ZZ
x(t) = U (τ, ν)s(t − τ ) ej2πνt dτ dν + η(t) (3.24)

The correlation of x(t) is expressed as:

 
∆ ∆t ∗ ∆t
Rxx (t, ∆t) = E x(t + ) · x (t − ) (3.25)
2 2

Taking Fourier transform from t to ∆f on both sides of the above, leads to the following
result:
E {Ax (∆f, ∆t)} = As (∆f, ∆t) · RT (∆f, ∆t) (3.26)

where RT (∆f, ∆t) is the double Fourier transform of the scattering function PU (ν 0 , τ 0 ).

Proof.
Express the correlation function of the received signal:
 
∆ ∆t ∗ ∆t
Rxx (t, ∆t) = E x(t + ) · x (t − )
2 2
(Z Z 
∆t j2πν(t+∆t/2)
=E U(ν, τ )s(t + − τ) e dτ dν
2
Z Z )
∆t 0
U ∗ (ν 0 , τ 0 )s∗ (t − − τ 0 ) e−j2πν (t−∆t/2) dτ 0 dν 0
2
ZZZZ
∆t
E U(ν, τ )U ∗ (ν 0 , τ 0 ) s(t +

= − τ)
2
∆t 0
× s∗ (t − − τ 0 ) · ej2π(ν(t+∆t/2)−ν (t−∆t/2)) dτ dν dτ 0 dν 0
2
(3.27)
70 Chapter 3. Instantaneous Frequency Estimation in Multiplicative Noise

Applying the expression in (2.33), and integrating over ν and τ , one can obtains
ZZ
∆t ∆t 0
Rxx (t, ∆t) = PU (ν 0 , τ 0 ) × s(t + − τ 0 )s∗ (t − − τ 0 ) ej2πν ∆t dτ 0 dν 0 (3.28)
2 2

Taking Fourier transform from t to ∆f on both sides of the above we then have2
ZZZ
∆t ∆t
E {Ax (∆f, ∆t)} = PU (ν 0 , τ 0 )s(t + − τ 0 )s∗ (t − − τ 0)
2 2
0
× ej2πν ∆t e−j2πt∆f dτ 0 dν 0 dt
ZZZ
0
= PU (ν 0 , τ 0 )Ks (t − τ 0 , ∆t)ej2πν ∆t e−j2πt∆f dτ 0 dν 0 dt
ZZ Z 
0 0 0 −j2πt∆f 0
= PU (ν , τ ) Ks (t − τ , ∆t)e dt ej2πν ∆t dτ 0 dν 0
ZZ
0 0
= PU (ν 0 , τ 0 ) · As (∆f, ∆t)e−j2πτ ∆f ej2πν ∆t dτ 0 dν 0

Fν−1 0 0
 
= As (∆f, ∆t) · 0FT 0 →∆t PU (ν , τ ) (3.30)
τ →∆f

where Ax (·) and Ks (·) are defined in Figure 2.2 for different signals (x(t) and s(t) instead
of z(t)) and different variables. Note that, the notations t, f , ν and τ used in Fourier
transforms are only for familiar convention in signal processing point of view, one can use
any others as long as they satisfy the definition of Fourier transform). It should be noted
that the result (3.26) appears also in [38] without proof.

Equation (3.26) represents the SOS relationship in the ambiguity domain based on
WSSUS assumption. By applying inverse double Fourier transform on both sides of (3.26),
we have another representation of the SOS relationship in the TF domain (see Figure 2.2)

n o
0 0 wvd 0 0 0 0
E ρwvd
x (ν , τ ) = ρs (ν , τ ) ?ν 0 ?τ 0 PU (ν , τ ) (3.31)

where ρwvd denotes the WVD. One often renames the variables (∆f , ∆t, ν 0 and τ 0 ) by
(ν, τ , t and f ), respectively [38]. We adopt this for the ease of visualization in the TF
2
Fourier transform operator and expected value operator are interchangeable under some spe-
cific conditions.
  
∆t ∗ ∆t
FT {Rxx (t, ∆t)} = E FT x(t + ) · x (t − ) = E {Ax (∆f, ∆t)} (3.29)
t→∆f t→∆f 2 2
Appendix 3A: Scattering function estimation 71

context, as a result, (3.26) and (3.31) can be rewritten as

E {Ax (ν, τ )} = As (ν, τ ) · RT (ν, τ ) (3.32)

n o
E ρwvd wvd
x (t, f ) = ρs (t, f ) ?t ?f PU (t, f ) (3.33)

By multiplying both sides of (3.32) with an arbitrary kernel g(ν, τ ), and taking the
inverse double Fourier transform of the result, we arrive to two general equations repre-
senting the SOS relationship in terms of the GAF A and the general quadratic TFD ρ
(see Figure 2.2), as

E {Ax (ν, τ )} = As (ν, τ ) · RT (ν, τ ) (3.34)

E {ρx (t, f )} = ρs (t, f ) ?t ?f PU (t, f ) (3.35)

It should be noted that the result given in [39,40] are the special cases (expressed in terms
of the SPEC and WVD) of the general case presented in (3.35).

Since the kernel g(ν, τ ), implicitly included in (3.34) and (3.35), is arbitrary, two gen-
eral classes of estimators for the SF are proposed: deconvolution and direct–implementa-
tion.

3A-3. Deconvolution–based estimation

The class of deconvolution estimators is defined based on the division of (3.34) by the
generalized ambiguity function of the input signal
  
(g,1) ∆ −1 E {Ax (ν, τ )}
P̂U (t, f ) = Fν→t FT (3.36)
τ →f As (ν, τ )

Similar to the approach in [41], zero–division problem in (3.36) is encountered. A

classical solution is to threshold the symmetric AF As (ν, τ ), or the Woodward ambiguity
function (when matched filtering is pre–applied at the output signal), at the points equal
to zero. Mathematically, it replaces As (ν, τ ) by As0 (ν, τ ) = As (ν, τ ) + λC(ν, τ ).

This replacement creates a problem in which the signal s0 (t) may not exist. One
72 Chapter 3. Instantaneous Frequency Estimation in Multiplicative Noise

could also use an interrogating signal (as in the case of image processing [41]) with some
well–behaved characteristics in order to achieve better estimation of PU (t, f ). However, in
communications, interrogating (pilot) signals are not encouraged since the communication
becomes more expensive with extra maneuver of the pilot signals.

On the other hand, by using the kernel g(ν, τ ), in turns, making use of the well–
defined GAF, the problems of the nonexistence of s0 (t) and using interrogating signals are
avoided. Also the kernel can be used to smooth As (ν, τ ) in the sense that it represents
the entire energy within a cell (see Figure 2.9) to a value at the center of the cell and the
estimates of the SF need only to be evaluated at these centers of the cells. This helps
partially minimize the zero division problem. Thresholding, not for zero division problem,
can be applied after smoothing in order to discard the cells that have negligible energy.
As a result, computational efficiency of post–processing in each cell for different purposes
(e.g. detection [45]) can be significantly improved.

3A-4. Direct–implementation–based estimation

One can choose the kernel g(ν, τ ) so that the TFD ρs (t, f ) in (3.35) is impulse–like (we
would ideally wish to have a delta representation in the TF plane, this, however, does
not exist due to the constraint of minimum TF bandwidth according to Heisenberg’s
uncertainty principle), the left–hand side of (3.35), then, approximates to PU (t, f ). Thus,
we define another class, namely direct–implementation, of estimators

(g,2) ∆
P̂U (t, f ) = E {ρx (t, f )} (3.37)

An example of this class can be obtained by choosing kernel such that ρs (t, f ) is approxi-
mated to that of Hermite functions known for having well–localized TF representation [46].

3A-V. Conclusion

We have proposed two classes of estimators for the SF of the TF dispersive fading mobile
channels with the use of the GAF familiarized in the context of TFSP. The degree of
freedom introduced by the arbitrary kernel g(ν, τ ) results in different estimators. This
Appendix 3A: Scattering function estimation 73

avoids or overcomes the problems encountered in the existing estimators. The selection
of optimum criteria for the estimators and, in turns, the performance of the estimators,
depends on the selection of the kernel with specific property.
74 Chapter 3. Instantaneous Frequency Estimation in Multiplicative Noise

APPENDIX 3B: IF estimation in wideband multiplica-

tive noise

3B-1. Introduction

One of the major effects on wideband transmission in mobile radio communications, due
to multipath propagation, is the TF dispersion as the results of time–delays over the
multipaths and Doppler–shifts from random motion of scatterers. This effect is known as
TF selective fading [7, 47]. The characteristics of Rayleigh faded wideband channel was
reviewed in Section 2.2 and formulated in Model 2.2. This can be considered as Wideband
Multiplicative Noise (WMN). This appendix investigates the problem of IF estimation of
LFM signals corrupted in WMN.

3B-2. Signal model and assumptions

Under the effect of both WMN µ(t) and additive noise η(t), we have the following general
model for the corrupted signal x(t):
Z ∞
x(t) = µ(t, λ) s(t − λ) dλ + η(t). (3.38)
−∞

where s(t) is the original LFM signal. The following assumptions are made:

As1) η(t) is assumed to be a zero–mean complex white Gaussian random process with
variance ση2 .

As2) µ(t, λ) is assumed to be a TV zero–mean complex Gaussian random process with

variance σµ2 .

As3) µ(t, λ) is also a TV WSSUS process, i.e. its correlation can be expressed as:

∆
Rµ (t1 , t2 ; λ1 , λ2 ) = E µ∗ (t1 , λ1 )µ(t2 , λ2 ) ≡ Rµ (∆t; λ2 ) δ(λ2 − λ1 ),

(3.39)

where ∆t = t2 − t1 .
Appendix 3B: IF estimation in wideband multiplicative noise 75

The received signal in (3.38) is equivalent to the received signal model in wideband channel
(Model 2.2).

3B-III. WVS analysis

Under the model in (3.38) and without considering the additive noise η(t), the WVS of
the corrupted signal x(t) is found to be:

Sxwvd (t, f ) = Sµ (t, f ) ? ? ρwvd

s (t, f ) (3.40)
t f

where ρwvd
s (t, f ) is the WVD of s(t) and Sµ (t, f ) is the TV spectrum of the WMN under

WSSUS assumption (As3).

Proof.
The correlation of the corrupted signal x(t) is expressed as

∆
n τ τ o
Rx (t, τ ) = E x(t + ) x∗ (t − ) =
(Z 2 2

τ τ
=E µ(t + , λ1 ) s(t + − λ1 ) dλ1
2 2
Z )
∗ τ ∗ τ
× µ (t − , λ2 ) s (t − − λ2 ) dλ2
2 2
ZZ n τ τ o
= E µ(t + , λ1 ) µ∗ (t − , λ2 )
2 2 (3.41)
τ ∗ τ
s(t + − λ1 ) s (t − − λ2 ) dλ1 dλ2
2 2

Then applying Assumption (As3) above, we obtain

ZZ
τ τ
Rx (t, τ ) = − λ1 ) s∗ (t − − λ2 ) dλ1 dλ2
Rµ (τ, λ2 ) δ(λ2 − λ1 ) s(t +
2 2
Z
τ τ
= Rµ (τ, λ)s(t + − λ) s∗ (t − − λ) dλ
2 2
Z
= Rµ (τ, λ) Ks (t − λ, τ ) dλ

= Rµ (t, τ ) ? Ks (t, τ ) (3.42)

t
76 Chapter 3. Instantaneous Frequency Estimation in Multiplicative Noise

The FT of the correlation Rx (t, τ ) gives us the WVS of x(t)

Z
Sxwvd (t, f ) = Rx (t, τ ) e−j2πf τ dτ
Z  
= Rµ (t, τ ) ? Ks (t, τ ) e−j2πf τ dτ
t

= Sµ (t, f ) ? ? Ws (t, f ) (3.43)

t f

where Sµ (t, f ) is the spectrum of Rµ (t, τ ).

Using the dual domain relationship for the WVD, as shown in Figure 2.2, we arrive
to an equivalent expression for the WVS of the corrupted signal x(t)

n o
FT FT −1 SxWVD (t, f ) = FT FT −1 Sµ (t, f ) · As (τ, ν)

(3.44)
t→ν f →τ t→ν f →τ

= Mµ (τ, ν) · As (τ, ν) (3.45)

In (3.45), As (τ, ν) is the AF of the signal s(t). Note that, the quantity Mµ (τ, ν) is indeed
the SF of the Rayleigh wideband channel as mentioned in Section 2.2.3.

Convolving both sides of (3.40) by a TFD kernel γ(t, f ) of a particular TFD, we

obtain the following expression for a general quadratic TFD spectrum Sxρ (t, f ) as:

Sxρ (t, f ) = Sµ (t, f ) ? ? ρTFD

s (t, f ), (3.46)
t f

where ρTFD
s (t, f ) is the quadratic TFD of s(t).

Equivalently, using the relationship between quadratic TFDs, as shown in Figure 2.3,
we may write:
FT FT −1 SxTFD (t, f ) = Mµ (τ, ν) · As (τ, ν).

(3.47)
t→ν f →τ

where As (τ, ν) is the GAF of the signal s(t).

Equation (3.45) suggests that if the structure of the WMN, expressed by Mµ (τ, ν), is
known, then the IF of the LFM signal s(t) can be estimated from the peak of:

FT −1 Sxwvd (t, f ) 
  
 FT
t→ν
−1 f →τ
ρ̂wvd
s (t, f ) = FT FT (3.48)
ν→t τ →f  Mµ (τ, ν) 
Appendix 3B: IF estimation in wideband multiplicative noise 77

or more general, using Equation (3.47), from the peak of:

FT −1 {Sxρ (t, f )} 
 
 FT
t→ν f →τ
ρ̂TFD
s (t, f ) = FT −1 FT (3.49)
ν→t τ →f  Mµ (τ, ν) 

Obviously, the above estimators fail at any point (τi , νi ) for which Mµ (τi , νi ) is equal to
zero. However, when applying to the Rayleigh wideband channel, Mµ (τi , νi ) has a special
structure upon which one is able to derive a more practical estimator. This will be shown
in the next section. The estimation of the lag–Doppler spectrum of the noise Mµ (τ, ν),
or in other words, the SF, was dealt with in Appendix 3A.

3B-3. TFD–based IF estimation in WMN

Previously noting that Mµ (τi , νi ) is the SF of the Rayleigh wideband channel, and under
the practical assumptions of wideband channel explained in Section 2.2.4, we can now
make an assumption on the WMN µ(t) that: If Mµ (τi , νi ) varies very little for changes in
τ of the order 1/B or for changes in ν of the order 1/T , this correlation is approximated
as [36]

 1 · Mµ m , n
 
for m ∈ [0, M ] and n ∈ [−N, N ]

Mµ (τ, ν) = BT B T (3.50)
0

otherwise

where M = dTm /Be and N = dνD /T e, and Tm is the delay–spread and νD is the maximum
Doppler–shift.

Based on the above assumption, the corrupting WMN can be decomposed into a set
of (M + 1) × (2N + 1) elements expressed as [36, 48]:

M N M N
X X 1 X X m n  m n m
x(t) = xmn (t) = Mµ , · s(t − ) ej2π T (t− B ) (3.51)
BT B T B
m=0 n=−N m=0 n=−N

where the corrupted signal x(t) under each version of WMN behaves as it has gone through
a Rayleigh fading narrowband channel [48].

Exploiting this structure of the channel decomposition, we now propose an estimator

78 Chapter 3. Instantaneous Frequency Estimation in Multiplicative Noise

for the IF of the underlying LFM signal as follows.

• First, apply a TF array–shifter of (M + 1) × (2N + 1) elements each shifted in time

by m/B and in frequency by n/T to obtain the shifted outputs:

m −j2π n (t+ m ) 1 m n 
x̂shifted
mn (t) = xmn (t + )e T B = Mµ , · s(t) (3.52)
B BT B T

• Next, obtain the IF estimate of x̂shifted

mn (t) using the previously TFD–based proposed
estimator (see Section 3.4).

n o
fˆmn (t) = arg max Sxρshifted (t, f ) (3.53)
f

• Finally, average over (M + 1) × (2N + 1) IF estimates to get the estimation:

M N
1 X X
fˆin (t) = fˆmn (t) (3.54)
(M + 1)(2N + 1)
m=0 n=−N

Further investigation should continue to confirm the success of this estimator and analyze
its performance. This is left for future research.
References

[1] B. Boashash, “Estimating and interpreting the instantaneous frequency of a signal-

Part 1: Fundamentals,” Proceedings of the IEEE, vol. 80, pp. 519–538, Apr. 1992.

[2] F. Henderson and A. Lewis, “Principle & applications of imaging radar,” in Manual
of Remote Sensing, vol. 2, New York: Wiley, 1998.

[3] R. J. Doviak and D. S. Zrnic, Doppler radar and weather observations. Orlando, FL:
Academic, 2nd ed., 1993.

[4] R. Dwyer, “Fourth–order spectra of Gaussian amplitude modulated sinusoids,” The

Journal of the Acoustical Society of America, vol. 90, pp. 916–926, 1991.

[5] R. Dwyer, “Range and Doppler information from fourth–order spectra,” IEEE Jour-
nal of Ocean Engineering, vol. 16, pp. 233–243, July 1991.

[6] J. G. Proakis, Digital Communications. McGraw–Hill, 3rd. ed., 1995.

[7] R. Steele and L. Hanzo, eds., Mobile Radio Communications: Second and Third
Generation Cellular and WATM Systems. England: John Wiley & Sons, 2nd ed.,
1999.

[8] D. Bellan and S. Pignari, “A probabilistic model for the response of an electrically
short two–conductor transmission line driven by a random plane wave field,” IEEE
Transactions on Electromagnetic compatibility, vol. 43, pp. 130–139, May 2001.

[9] S. Haykin, Communication Systems. New York: Wiley, 3rd ed., 1994.

[10] M. Ghogho, A. Swami, and T. S. Durrani, “Frequency estimation in the presence

of Doppler spread: performance analysis,” IEEE Transactions on Signal Processing,
vol. 49, pp. 777–789, Apr. 2001.

[11] B. Boashash and B. Ristic, “Analysis of FM signals affected by Gaussian AM using re-
duced Wigner–Ville trispectrum,” in International Conference on Acoustics, Speech,
and Signal Processing, ICASSP’93, vol. IV, (Minneapolis), pp. 408–411, 1993.

[12] B. Boashash and B. Ristic, “Polynomial time-frequency distributions and time-

varying higher order spectra: Application to the analysis of multicomponent FM
signals and to the treatment of multiplicative noise,” Signal Processing, vol. 67,
pp. 1–23, 1998.
79
80 Chapter 3. Instantaneous Frequency Estimation in Multiplicative Noise

[13] O. Besson and F. Castanié, “On estimating the frequency of a sinusoid in autore-
gressive multiplicative noise,” Signal Processing, vol. 30, pp. 65–83, Jan. 1993.
[14] S. Shamsunder and G. Giannakis, “Estimating random amplitude polynomial phase
signals: a cyclostationary approach,” IEEE Transaction on Signal Processing, vol. 43,
pp. 492–505, Feb. 1995.
[15] A. Swami, “Multiplicative noise models: parameter estimation using cumulants,”
Signal Processing, vol. 36, pp. 355–373, Apr. 1994.
[16] G. Zhou and G. B. Giannakis, “On estimating random amplitude-modulated har-
monics using higher order spectra,” IEEE Journal of Oceanic Engineering, vol. 19,
pp. 529–539, Oct. 1994.
[17] G. Zhou and G. Giannakis, “Harmonics in Gaussian multiplicative and additive noise:
Cramer-Rao bounds,” IEEE Transactions on Signal Processing, vol. 43, pp. 1217–
1231, May 1995.
[18] B. Friedlander and J. Francos, “Estimation of amplitude and phase parameters of
multicomponent signals,” IEEE Transactions on Signal Processing, vol. 43, pp. 917–
926, Apr. 1995.
[19] J. Francos and B. Friedlander, “Bounds for estimation of multicomponent signals
with random amplitude and deterministic phase,” IEEE Transactions on Signal Pro-
cessing, vol. 43, pp. 1161–1172, May 1995.
[20] O. Besson, N. Ghogho, and A. Swami, “Parameter estimation for random amplitude
chirp signals,” IEEE Transactions on Signal Processing, vol. 47, pp. 3208–3219, Dec.
1999.
[21] B. Senadji and B. Boashash, “A mobile communications application of time-varying
higher order spectra to FM signals affected by multiplicative noise,” in Proceedings of
ICICS, 1997 International Conference on Information, Communications, and Sig-
nal Processing. Theme: Trends in Information Systems Engineering and Wireless
Multimedia Communications, vol. 3, (New York, USA), pp. 1489–1492, 1997.
[22] M. R. Morelande and A. M. Zoubir, “Model selection of random amplitude polyno-
mial phase signals,” IEEE Transactions on Signal Processing, vol. 50, pp. 578–589,
Mar. 2002.
[23] M. R. Morelande and A. M. Zoubir, “On the performance of cyclic moments–based
parameter estimators of amplitude modulated polynomial phase signals,” IEEE
Transactions on Signal Processing, vol. 50, pp. 590–606, Mar. 2002.
[24] A. Swami, “Cramer-Rao bounds for deterministic signals in additive and multiplica-
tive noise,” Signal Processing, vol. 53, pp. 231–244, 1996.
[25] M. Ghogho, A. K. Nandi, and A. Swami, “Cramer-Rao bounds and maximum like-
lihood estimation for random amplitude phase–modulated signals,” IEEE Transac-
tions on Signal Processing, vol. 47, pp. 2905–2916, Nov. 1999.
[26] B. Barkat, “Instantaneous frequency estimation of nonlinear frequency–modulated
signals in the presence of multiplicative and additive noise,” IEEE Transactions on
Signal Processing, vol. 49, pp. 2214–2222, Oct. 2001.
References 81

[27] G. Azemi, B. Senadji, and B. Boashash, “Instantaneous frequency estimation of

frequency modulated signals in the presence of multiplicative and additive noise:
application to mobile communication systems.” to appear in EUSIPCO2002.
[28] W. Jakes, ed., Microwave Mobile Communications. IEEE Press, 1974.
[29] A. Rihaczek, Principles of High-Resolution Radar. Peninsula Publishing, 1985.
[30] C. Gupta and A. Papandreou-Suppappola, “Wireless CDMA communications using
time–varying signals,” in Proceedings of the Sixth International Symposium on Signal
Processing and its Applications, ISSPA’01, vol. 1, pp. 242–245, 2001.
[31] B. Boashash, ed., Time-Frequency Signal Analysis: Methods and Applications. Mel-
bourne, Australia: Longman Cheshire, 1992.
[32] P. Z. Peebles, Jr., Probability, Random Variables, and Random Signal Principles.
Singapore/New York: McGraw-Hill, 4th ed., 2001.
[33] M. Pätzold, Mobile fading channels. John Wiley, 2002.
[34] M. Pätzold, U. Killat, and F. Laue, “A deterministic digital simulation model for
Suzuki processes with application to shadowed Rayleigh land mobile radio channel,”
IEEE Transactions on Vehicular Technology, vol. 45, no. 2, pp. 318–331, 1996.
[35] S. M. Kay, Fundamentals of Statistical Signal Processing: Estimation Theory. A.V.
Oppenheim, series editor, Prentice-Hall Signal Processing Series, Englewood Cliffs,
New Jersey: Prentice-Hall, 1993.
[36] P. Bello, “Characterization of randomly time-variant linear channels,” IEEE Trans-
actions on Communication Systems, vol. COM-11, pp. 360–393, Dec. 1963.
[37] H. L. VanTrees, Detection, Estimation, and Modulation Theory. Radar-Sonar Signal
Processing and Gaussian Signals in Noise. Krieger Pub. Co., Malabar, Florida, 1992.
[38] N. T. Gaarder, “Scattering function estimation,” IEEE Transactions on Information
Theory, vol. IT-14, pp. 684–693, Sept. 1968.
[39] R. A. Altes, “Detection, estimation, and classification with spectrograms,” The Jour-
nal of the Acoustical Society of America, vol. 67, pp. 1232–1246, Apr. 1980.
[40] P. Flandrin, “A time-frequency formulation of optimum detection,” IEEE Trans-
actions on Acoustics, Speech, and Signal Processing, vol. 36, pp. 1377–1384, Sept.
1988.
[41] B. L. Johnson, D. W. Ricker, and J. R. Sacha, “The use of iterative deconvolution for
scattering function identification,” The Journal of the Acoustical Society of America,
vol. 91, pp. 2790–2798, May 1992.
[42] D. W. Ricker and M. J. Gustafson, “A low sidelobe technique for the direct mea-
surement of scattering functions,” IEEE Journal of Oceanic Engineering, vol. 21,
pp. 14–23, Jan. 1996.
[43] T. Wang, V. K. Dubey, and J. T. Ong, “Generation of scattering functions for mobile
communication channel: a computer simulation approach,” Wireless Information
Networks, vol. 4, no. 3, pp. 187–204, 1997.
82 Chapter 3. Instantaneous Frequency Estimation in Multiplicative Noise

[44] J. S. Sadowsky and V. Kafedziski, “On the correlation and scattering function of
the WSSUS channel for mobile communications,” IEEE Transactions on Vehicular
Technology, vol. 47, pp. 270–282, Feb. 1998.
[45] A. M. Sayeed, A. Sendonaris, and B. Aazhang, “Multiuser detection in fast-fading
multipath environments,” IEEE Journal on Selected Areas in Communications,
vol. 16, pp. 1691–1701, Dec. 1998.
[46] R. Haas and J.-C. Belfiore, “A time-frequency well-localized pulse for multiple carrier
transmission,” Wireless Personal Communications, vol. 5, pp. 1–18, 1997.
[47] B. Sklar, “Rayleigh fading channels in mobile digital communication systems, part
I: Characterization,” IEEE Communications Magazine, pp. 90–100, July 1997.
[48] A. M. Sayeed, “Canonical time-frequency processing for broadband signaling over
dispersive channels,” in Proceedings of the IEEE-SP International Symposium on
Time-Frequency and Time-Scale Analysis, (New York, USA), pp. 369–372, IEEE,
1998.
 Chapter 4

Underdetermined Blind Source

Separation

4.1 Introduction

BSS has its root in array signal processing. Signals from some particular sources
first pass through an intermediate medium (with possibly noise), which modifies
the original source signals, then arrive at an array of sensors. The observed output
of each sensor is a mixture of all the source signals. It is desired to recover the
unobserved source signals from the observed mixtures; this problem is known as
source/signal separation. If neither the structure of the medium transfer nor the
source signals are known, we are said to be in a “blind” context. One often assumes
in such a context that the source signals are mutually independent in order to
facilitate the separation [1]. BSS is important when modeling the transfer from the
sources to the sensors is difficult or when no a priori information is available about
83
84 Chapter 4. Underdetermined Blind Source Separation

the mixtures.

BSS is also known as: blind array processing, signal copy, Independent Com-
ponent Analysis (ICA), and waveform preserving estimation. It has emerged over
the past decade to become an important area of signal processing, being signified
by an ongoing series of dedicate conferences [2] and appearing as special sessions in
many signal processing conferences. Useful reviews of BSS theories and algorithms
can be found in [1, 3–7].

BSS has many applications in areas that involve the processing of signals from a
sensor array, which offers spatial diversity. Typical examples of BSS are seen in: (i)
radar and sonar applications (separation and recognition of sources from antenna
arrays, robust source localization from ill–calibrated arrays [8]), (ii) communications
(multiuser detection in communication systems [9]), (iii) speech processing (speaker
separation, also called the “cocktail party” problem; speech recorded in the presence
of background noise and/or competing speakers, automatic voice recognition in
noisy acoustic environments [10]), and (iv) biomedical signal processing (separation
of Electroencephalogram (EEG) signals [11, 12]).

BSS can be categorized into different classes according to the way the signal
structures are “forced”/conditioned using some particular criteria so that we can
restore the original structure of the source signals. These different classes are [7]:
probability structure forcing, spectral/time–coherence structure forcing, and TF
structure forcing.

When signals are nonstationary, the TF structure forcing approach was intro-
duced to achieve the separation, by Belouchrani and Amin [13, 14]. This approach
defines a STFD that combines both TF diversity and spatial diversity. The benefit
of using STFDs in an environment of nonstationary signals is the direct exploita-
tion of the information it offers due to the signal nonstationarity. In contrast to
BSS approaches using SOS and HOS (see [6] and references therein), this approach
allows the separation of Gaussian sources with identical spectral shape but with
different TF localization properties. Moreover, the effects of spreading the noise
4.1. Introduction 85

power, while localizing the source energy in the TF domain, amounts to increasing
the SNR [15]. Subsequent works have been carried out by Belouchrani, Amin and
their co–workers on the further development of this approach and its applications
to communications [16–18].

A drawback of most BSS algorithms is that they fail to separate sources in

situations where there are more sources than sensors [19]. Mathematically, the
invertability of mixing matrix, that is often used for separation, is no longer satis-
fied [20]. This challenging problem, known as the Underdetermined Blind Source
Separation (UBSS), has recently been studied in [20, 21] where the discrete sources
were treated, in [19,22–24] where a priori knowledge of the probability density func-
tions of the sources was needed; and in [10] where disjoint orthogonality displayed
by STFT was exploited.

UBSS for nonstationary signals is investigated in this chapter. The TF structure

forcing approach above is chosen in order to take advantage of TF signal processing,
over the classical time–only and frequency–only signal processing. We will propose
a TF–UBSS algorithm that uses the main assumption of TF orthogonality. In
particular, TF orthogonality facilitates the selection of TF points lying on the TF
supports (signatures) of all source signals, hence a clustering process allows the
identification/separation of the TF signatures. We make a distinction here regarding
the previous approach using TF information [13, 14, 16–18] that we are now in the
underdetermined situation in which the previous approach is inapplicable due to
the non–invertability of the mixing matrix.

The chapter is organized as follows. Section 4.2 presents the data model and
assumptions, especially the notion of TF orthogonality. Section 4.3 recalls the
definition and properties of STFD matrices. Section 4.4 proposes the TF–UBSS al-
gorithm. Section 4.5 provides an illustrative demonstration of the usefulness of the
algorithm by some simulated experiments. Section 4.6 presents an enhanced version
of the algorithm using MWVD to achieve better selection of TF points. Section 4.7
provides another method to enhance the selection of TF points using image com-
86 Chapter 4. Underdetermined Blind Source Separation

ponent extraction. Several measurements for numerical performance evaluation are

given in Section 4.8. The last section is for concluding remarks and perspectives.

4.2 Signal model and assumptions

Assume that an n–dimensional vector s(t) = [s1 (t), s2 (t), . . . , sn (t)]T ∈ C(n×1)
corresponds to n nonstationary complex source signals si (t), i = 1, . . . , n. The
source signals are transmitted through a medium so that an array of m sensors
picks up a set of mixed signals represented by an m–dimensional vector x(t) =
[x1 (t), x2 (t), . . . , xm (t)]T ∈ C(m×1) . Each observed signal xj (t), j = 1, . . . , m, at
each time instance t has been mixed by the transmission medium which may have
also been corrupted by AWGN η(t) = [η1 (t), η2 (t), . . . , ηm (t)]T ∈ C(m×1) . Consid-
ering the instantaneous linear mixture case, the observed signals can be modeled
as:
x(t) = As(t) + η(t), (4.1)

where A ∈ C(m×n) is called the mixing matrix. The instantaneity means that A
does not depend on t. The signal model is illustrated in Figure 4.1.

In the underdetermined situation, i.e. the UBSS problem, we have n > m. The
mixing matrix A is no longer invertible [20], thus any previous approaches in the
determined BSS problem (i.e. n ≤ m) is generally no longer applicable. Note that,
as m approaches infinity, the quantity (n − m) approaches zero (since n > m), thus,
UBSS becomes determined BSS. Therefore, one may approximately use the usual
methods in the determined BSS case to achieve the separation; in other words, this
happens when n − m is small compared to m.

We made the following two assumptions. The first assumption is usually made in
the context of BSS, and the second assumption is the main feature which facilitates
the proposal of our TF–UBSS algorithm.

As1) The column vectors of matrix A = [a1 , a2 , . . . , an ] are assumed to be pairwise

4.2. Signal model and assumptions 87

?A@ @ ? @
A
EG@ F
?C C ? C

E C F BLIND
?B 3;9 :<9=0> ? B
354687 9 : SOURCE

E F<H
H
SEPARATION
?D ? D


 
!#"  
  



$ %& '

$ %& '
&!-



 ( . 
1 )*+, )*2 0  )/0  

Figure 4.1: UBSS: Schematic diagram.

There are n unobserved source signals si (t), i = 1, . . . , n, to be separated from
m observed mixed signals xj (t), j = 1, . . . , m, corrupted by AWGN. The under-
determined case corresponds to n > m.

linearly independent, i.e., for any i, j ∈ 1, 2, . . . , n and i 6= j, ai and aj are

linearly independent. Obviously, if two sources, for example s1 (t) and s2 (t),
have linearly dependent vectors, i.e. a2 = αa1 , their separation is, then,
inherently impossible since we can write

x(t) = Ãs̃(t) + η(t), (4.2)

where à = [a1 , a3 , . . . , an ] and s̃(t) = [s1 (t) + αs2 (t), s3 (t), . . . , sn (t)]T . It
is also known that BSS is only possible up to an unknown scaling and an
unknown permutation [25]. We take the advantage of this indeterminacy to
assume, without loss of generality, that the column vectors of A have a unit–
norm, that is kai k = 1 for all i.

As2) The sources are assumed to have different structures and localization proper-
ties in the TF domain. More precisely, we assume the sources to be orthogonal
in the TF domain (Figure 4.2) as stated in the following definition:

Definition 4.1 (TF orthogonality).

88 Chapter 4. Underdetermined Blind Source Separation

Let S1 (t, f ) and S2 (t, f ) be TFDs of two source signals s1 (t) and s2 (t), respec-
tively. Let Ω1 and Ω2 be the corresponding TF supports of S1 and S2 , that
is 
S1 (t, f ) 6= 0

if and only if (t, f ) ∈ Ω1 ,
(4.3)
S2 (t, f ) 6= 0

if and only if (t, f ) ∈ Ω2 .

The sources s1 (t) and s2 (t) are said to be orthogonal in the TF domain
if the following satisfies:
Ω1 ∩ Ω2 = ∅. (4.4)

Time-frequen y orthogonality
t

1
time

2
frequen y f

Figure 4.2: TF orthogonality.

The TF supports of two sources are disjoint in the TF domain.

The above definition can be applied to any TFDs. It is clear that the TF
orthogonality is too restrictive and will almost never be satisfied exactly in practice.
Fortunately, only approximate orthogonality, said quasi–orthogonality, is needed to
achieve source separation, as will be shown in Section 4.5.2. Note that the source TF
orthogonality can be considered as a particular type of source sparse decomposition,
which can be used to achieve source separation [22, 26, 27].

A physical example of TF (quasi)–orthogonality is observed in a musical per-

formance; several musical instruments, e.g. a base–guitar and a lead–guitar, play
simultaneously but create musical sounds with different instantaneous frequency
laws. These laws can have some overlapping in the TF domain, representing the
4.3. Spatial time–frequency distributions 89

quasi situation. This happens when the very high (frequency) notes produced by
the base–guitar coincide in frequency with the very low (frequency) notes produced
by the lead–guitar, and these notes are played at the same duration of time.

Two arguments may be given here regarding the assumption of TF orthogonal-

ity, before going further introducing the TF–UBSS algorithm. Firstly, one would
think that if the sources are TF orthogonal then simple TF masking (if the source
TF signatures are known) and TF synthesis in the TF domain would be sufficient
to recover the source signals without using any sophisticated algorithm. However,
in the context of blind separation, source TF signatures are unknown. The pro-
posed algorithm will allow us to extract the source TF signatures from the spatial
information offered by the sensors; hence, the source signals. Secondly, one may
use an image processing technique to achieve a classification of different source TF
components (as distinguished to source TF signatures) as has been done in [28].
With the obtained TF components, this classification still fails to obtain the source
TF signatures because it is well possible that a source TF signature can have mul-
tiple TF components. Our algorithm provides necessary information to allow for
the determination of which TF components belong to one particular source, thus
allowing blind separation of multicomponent source signals.

4.3 Spatial time–frequency distributions

We provide here some definitions that will be used throughout the chapter.

Definition 4.2 (Spatial TFD [13]).

Let z(t) be a vector containing n signals z1 (t), . . ., zn (t); z(t) = [z1 (t), . . . , zn (t)]T ∈
C(n×1) . The Spatial Time–Frequency Distribution (STFD) matrix is math-
ematically defined as

∞ ∞
∆
X X
Dzz (t, f ) = φ(k, l)z(t + k + l)zH (t + k − l)e−j4πf l , (4.5)
l=−∞ k=−∞
90 Chapter 4. Underdetermined Blind Source Separation

where t and f represent the time index and the frequency index, respectively, the
superscript (H ) denotes the complex conjugate transpose operator, and φ(m, l) is a
TFD time–lag kernel. The matrix Dzz (t, f ) ∈ C(n×n) varies with respect to t and f .
Its (t–f ) elements are obtained from the TFD as:

[Dzz (t, f )]ij = Dzi zj (t, f )

∞
X ∞
X
= φ(k, l)zi (t + k + l)zj∗ (t + k − l)e−j4πf l , i, j = 1, 2, . . . , n (4.6)
l=−∞ k=−∞

with zj∗ being the complex conjugate of zj .

Note that Dzz (t, f ) is a matrix; when evaluated at a TF point (to , fo ), its ele-
ments are the values of Dzi zj (to , fo ) using (4.6).

Next, we will define the notion of cross– and auto–source STFDs, which are
slightly modified from those defined in [16] for more clarity. Before doing so, let
us recall the notions of “auto–term” and “cross–term” in the literature of TFSP.
Given a signal with multiple IF components, an auto–term TF point in the TF
representation of this signal represents the “true” energy concentration of the signal
at that point in time and frequency. A cross–term TF point, on the other hand,
represents a “ghost” energy concentration of the signal though the concentration
may visually appear high at this point the TF representation. This “ghost” effect
comes from the bilinearity of the TFD that applies on the signal among its IF
components [29].

Above, the TFD is applied on only one signal. In our context, we consider
several source signals, and each of which may have multiple IF components.

Definition 4.3 (Cross– and auto–source STFD).

Let z1 (t) and z2 (t) be two different source signals with possibly multiple IF compo-
nents, and that they be displayed on the TF representation by a TFD ρ(t, f ) through
the computation of the STFD Dzz (t, f ) where z(t) = [z1 (t)z2 (t)]T .

(a) An auto–source TF point (ta , fa ) of a source zi (t), i = 1, 2, is a point in

4.3. Spatial time–frequency distributions 91

the TF representation where the energy concentration of zi (t) is evaluated by

the auto–TFD ρzi zi (ta , fa ).

(b) A cross–source TF point (tc , fc ) between source z1 (t) and z2 (t) is a point
in the TF representation where the energy concentration is evaluated by the
cross–TFD ρz1 z2 (tc , fc )1 .

(c) For an auto–source TF point (ta , fa ), the STFD matrix computed at that point
is called an auto–source STFD matrix, denoted by Dzz (ta , fa ) .

(d) For a cross–source TF point (tc , fc ), the STFD matrix computed at that point
is called a cross–source STFD matrix, denoted by Dzz (tc , fc ).

A few remarks can be made according to the above definition. For simplicity,
hereafter, we use “point” to mean “TF point”.

• The energy concentration at an auto–source point can be “true” if zi (t) is

monocomponent but that can also be “ghost” if multicomponent. The lat-
ter means that the auto–source point coincides with the cross-term point if
the source is multicomponent. This will be illustrated in Experiment 4.5.2
(Figure 4.8.l).

• Since the diagonal elements of the matrix Dzz (t, f ) are evaluated by the auto–
TFD, this STFD matrix at an auto–source point, Dzz (ta , fa ), becomes an
auto–source STFD matrix and that it is quasi-diagonal (i.e. its diagonal
entries are close to one).

• Since the off–diagonal elements of the matrix Dzz (t, f ) are evaluated by the
cross–TFD, this STFD matrix at a cross–source point, Dzz (tc , fc ), becomes a
1
The cross–TFD is defined, similar to the auto–TFD as in (2.13), as below:
ZZZ ∞
∆ τ τ
ρz1 z2 (t, f ) = ej2πν(u−t) Γ (τ, ν) z1 (u + )z2∗ (u − ) e−j2πf τ dν du dτ (4.7)
−∞ 2 2

When z1 (t) = z2 (t), the cross–TFD becomes the auto–TFD.

92 Chapter 4. Underdetermined Blind Source Separation

cross–source STFD matrix and that it is quasi–off–diagonal. (i.e. its diagonal

entries are close to zero).

Applying (4.5) to the linear data model (4.1), assumed a noise–free environment,
leads to the following expression:

Dxx (t, f ) = ADss (t, f )AH , (4.8)

where Dss (t, f ) and Dxx (t, f ) are the source and mixture STFD matrices, respec-
tively. Further from the above remarks, since the sources are assumed to be TF
orthogonal, the diagonal entries of Dss (t, f ) are:

• all equal to zero except for one value, if the STFD matrix Dss (t, f ) is evaluated
at an auto–source point since only one source active at this point.

• all equal to zero, if the STFD matrix Dss (t, f ) is evaluated at a point other
than an auto–source point.

Therefore, if Ωi is the TF support of source signal si (t), the following is achieved:

Dxx (t, f ) = Dsi si (t, f )ai aH

i , ∀ (t, f ) ∈ Ωi . (4.9)

It is the particular structure in (4.9) that will be used for our TF–UBSS.

We also note that, if, on the other hand, the sources do not satisfy the TF
orthogonality assumption such that at an auto–source point there are k sources
active (i.e. there is an overlap, on the TF representation, of the TF signatures of
these sources), then among the diagonal entries of Dss (t, f ) there will have exactly
k values different from zero if k ≤ m, or at maximum m values different from zero
if k > m. This observation may be used to provide a test on TF orthogonality,
and further to analyse TF–nonorthogonality. However, detailed treatments of TF
non–orthogonality, e.g. the degree of acceptable non–orthogonality for successfully
achieving UBSS, is not carried out in this thesis and is subject to future research;
this issue is of importance when dealing with speech signals rather LFM signals.
4.4. TF-UBSS algorithm 93

4.4 TF-UBSS algorithm

Thanks to the structure in (4.9), the following observation is deduced for two auto–
source (t1 , f1 ) and (t2 , f2 ) corresponding to the same source si (t):

Dxx (t1 , f1 ) = Ds s (t1 , f1 )ai aH ,

i i i
(4.10)
Dxx (t2 , f2 ) = Dsi si (t2 , f2 )ai aH

i .

The above observation implies that Dxx (t1 , f1 ) and Dxx (t2 , f2 ) have the same prin-
cipal eigenvector ai . Therefore, all the auto–source points associated with the same
principal eigenvector belong to the TF support of one particular source signal.

This leads to the principal idea of our TF–UBSS algorithm as follows. We

first obtain only auto–source points from the TF representation. If we are able to
cluster auto–source points on the TF domain into different sets, each associating
with one principal eigenvector, then these sets represent different TF signatures
corresponding to different underlying source signals. And if each source signal can be
recovered from its set, we are able to achieve the UBSS. In particular, corresponding
to a principal eigenvector of one particular source, the estimated TFD values of
this source at its auto–source points are obtained as the principal eigenvalues of
the STFD matrices at those points. Hence, we can use a TF synthesis method to
recover the source waveform.

4.4.1 Separation algorithm

The proposed TF–UBSS algorithm includes four main procedures as shown in Fig-
ure 4.3 and its schematic diagram is illustrated in Figure 4.4. Details of these
procedures are given next.
94 Chapter 4. Underdetermined Blind Source Separation

TF–UBSS algorithm

Procedure 1: STFD computation and noise thresholding

Procedure 2: Auto–source TF point selection

Procedure 3: Clustering and source TFD estimation

Procedure 4: Source signal synthesis

Figure 4.3: TF-UBSS algorithm: Procedures.

Auto-term f(ta ; fa )g
Sele tion

Classier

C1 ^
D s1 s1 (t; f ) ^1 (t)
s
get TFD TF-Syn

x( ) = As( )
t t
STFD Dxx( t; f )
C2
get TFD
^
D s2 s2 (t; f )
TF-Syn
^2 (t)
s

WVD

Cn ^
D sn sn (t; f ) ^n (t)
s
get TFD TF-Syn

Figure 4.4: TF-UBSS algorithm: Schematic diagram.

4.4.1.1 STFD computation and noise thresholding

Given L observation vectors x(1), . . . , x(L), the STFD matrices Dxx (t, f ) defined
according to (4.5), can be estimated using time–lag domain discrete implementa-
tion [30] as below:

M
X M
X
D̂xx (l, k) = g(q − l, p) x(q + p)xH (q − p) e−j4πpk/L , (4.11)
p=−M q=−M
4.4. TF-UBSS algorithm 95

where g(l, p) is a discrete time–lag kernel, M = (L − 1)/2, and l = 1, . . . , L. The

elements of D̂xx (l, k) are obtained from the TFD as:

h i
D̂xx (l, k) = Dxi xj (l, k)
ij
M
X M
X
= g(q − l, p) xi (q + p)x∗j (q − p) e−j4πpk/L , i, j = 1, . . . , m. (4.12)
p=−M q=−M

In the later simulations (Experiment 1 and Experiment 2), we will use the WVD for
computing the STFD matrices. The WVD of an analytic signal x(t) is defined as
in (2.11). Its discrete implementation is of the form in (4.12) without the time–lag
kernel g(l, p).

These STFD matrices are next processed to extract the source signals. In order
to reduce the computational complexity, by processing only “significant” STFD
matrices, a noise thresholding step is then carried out for removing those points
with negligible energy. More precisely, a threshold 1 (typically, 1 = 0.05 of the
point with maximum energy) is used to keep only the points {(ts , fs )} with sufficient
energy:
If: kDxx (ts , fs )k > 1
(4.13)
then: keep (ts , fs )

4.4.1.2 Auto–source TF point selection

The second procedure of the algorithm consists of separating the auto–source points
from cross–source points using an appropriate testing criterion.

In the determined case, where the number of sensors is greater than or equal to
the number of sources and the mixing matrix A is of full–column rank, a selection
procedure that exploits the off–diagonal structure of the cross–source STFD matri-
ces has been proposed in [16]. This selection procedure proceeds through two steps
as follows:
96 Chapter 4. Underdetermined Blind Source Separation

– Data whitening: Let W denote an m × n matrix such that (WA)(WA)H =

UUH = I, i.e. WA is an n × n unitary matrix. Matrix W is referred to
as the whitening matrix since it whitens the signal part of the observations.
Pre– and post–multiplying the STFD matrices Dxx (t, f ) by W lead to the
whitened STFD matrices:

Dxx (t, f ) = WDxx (t, f )WH = UDss (t, f )UH (4.14)

In practice, W is often computed as an inverse squared root of the sample

estimate covariance matrix of the observation.

– Testing: Given a whitened cross–source STFD matrix Dxx (tc , fc ), we have:

trace {Dxx (tc , fc )} = trace UDss (tc , fc )UH



= trace {Dss (tc , fc )}

≈0 (4.15)

Based on this observation, the following test is given:

n o
trace D̂xx (tc , fc )
If: n o < 2
norm D̂xx (tc , fc ) (4.16)

then: (tc , fc ) is a cross–source point

where the threshold 2 is a positive scalar no greater than 1 (typically 2 =

0.8).

Contrary to the determined case explained above, the matrix U in the under-
determined case is non–square with more columns than rows, and consequently
UH U 6= I represents the projection matrix onto the row space of U. Therefore,
Eq. (4.15) becomes only an approximation; a good one if (m − n) is “small” as
observed in our simulation results (see Figure 4.7 of Experiment 1 and Figure 4.8
of Experiment 2 in Section 4.5).
4.4. TF-UBSS algorithm 97

Another method, alternative to the above approximation projection method,

consists of exploiting the sources TF orthogonality. Under this assumption, each
auto–source STFD matrix is of rank one, or at least has one “large” eigenvalue com-
pared to its other eigenvalues. Therefore, one can use rank selection criteria, such as
Minimum Description Length (MDL) or Akaike Information Criterion (AIC) [31],
to select auto–source points as those corresponding to STFD matrices of selected
rank equal to one. For simplicity, we use the following criterion (see Figure 4.11 of
Experiment 3 in Section 4.6):
 n o 
λ
 max D̂xx (t, f )


If: 
 n o − 1 > 2

 norm D̂xx (t, f )  (4.17)

then: (t, f ) is a cross–source point

where 2 is a small positive scalar (typically, 2 = 0.3), and λmax {·} represents the
largest eigenvalue of the matrix in the bracket.

Comparing the above two methods for auto–source point selection based on
approximation projection and TF orthogonality for the underdetermined case shows
a similar performance (see Figure 4.11).

4.4.1.3 Clustering and source TFD estimation

Once the auto–source points have been selected, a clustering procedure based on the
sources spatial directions/signatures is performed. This clustering is based on the
observation that two STFD matrices corresponding to the same source signal have
the same principal eigenvector. Moreover, the corresponding principal eigenvalues
are given by the desired source TFD. This implies that if we apply an appropriate
clustering procedure on the set auto–source points, we will be able to obtain the
separate TF signatures of the source signals. Specifically, we consider the following
steps:

– For each auto–source point, (ta , fa ), compute the main eigenvector, a(ta , fa ),
98 Chapter 4. Underdetermined Blind Source Separation

and its corresponding eigenvalue, λ(ta , fa ), of Dxx (ta , fa ).

– As the vectors {a(ta , fa )} are estimated up to a random phase ejφ , φ ∈ [0, 2π),
we force them to have, without loss of generality, their first entries real and
positive. These vectors are then clustered into different classes {Ci }. Mathe-
matically, a(ti , fi ) and a(tj , fj ) belong to the same class if:

d(a(ti , fi ), a(tj , fj )) < 3 (4.18)

where 3 is a properly chosen positive scalar and d is a distance measure

(different strategies for choosing the threshold 3 and the distance d or even
the clustering method can be found in [32]). As an example, we use a dis-
tance measure, in the simulated experiments in Section 4.5, according to their
angles:
d(a1 , a2 ) = arccos(ãTi ãj ), i 6= j (4.19)

where ã = [Re(a)T ; Im(a)T ]T and kãk = 1.

– Set the number of sources equal to the number of classes and, for each source
si (i.e. each class Ci ), estimate its TFD as:

λ(ta , fa ), if (t, f ) = (ta , fa ) ∈ Ci

D̂si si (t, f ) = . (4.20)
0,

otherwise

4.4.1.4 Source signal synthesis

Having obtained the source TFD estimates D̂si si , we then use an adequate source
synthesis procedure to estimate the source signals si (t) (i = 1, . . . , n). The recovery
of the waveform (in time) of a signal from its TFD is made possible thanks to the
following inversion property of the WVD [15]
Z ∞
1 t
x(t) = ∗ ρwvd
x ( , f) e
j2πf t
df , (4.21)
x (0) −∞ 2
4.4. TF-UBSS algorithm 99

which implies that the signal can be reconstructed to within a complex exponential
constant ejα = x∗ (0)/|x(0)| given |x(0)| =
6 0.

Some TF synthesis algorithms can be found in [15,33–35]. Among them, [33] pro-
vides a well–known synthesis algorithm recovering a signal from its WVD estimate.
Since we use WVD to compute our STFD matrices, we opt to use this synthesis
algorithm for recovering our original sources. Below, this algorithm is summarized
to assist the understanding of our UBSS algorithm (without any contribution of the
author of this thesis).

Given the TFD estimate of source s(t), denoted by D̂ss (t, f ), find the signal
ŝ(t) that its WVD, denoted by ρwvd
ŝ (t, f ), best approximates D̂ss (t, f ) in the least

square sense, i.e. minimizing the following:

Z ∞  2
J(s) = D̂ss (t, f ) − ρwvd (t, f ) df. (4.22)
 
ŝ
−∞

The above minimization leads to the discrete computation of the synthesized signal
ŝ(l), l = 0, . . . , L − 1, as below:

ŝ(2k) = se (k), for k = 0, . . . , Le − 1; Le = b(L + 1)/2c

, (4.23)
ŝ(2k − 1) = so (k), for k = 1, . . . , Lo ;

Lo = bL/2c

where se = [se (0)se (1) · · · se (Le − 1)]T and so = [so (1)so (2) · · · se (Lo )]T are the nor-
malized principal eigenvectors of the matrices Ce and Co , representing the even and
odd samples of ŝ(k). The elements of these matrices are computed as:



 ce (q + 1, p + 1) = y(q + p, q − p) + y ∗ (q + p, p − q),


for q, p = 0, . . . , Le − 1


, (4.24)



 co (q, p) = y(q + p + 1, q − p) + y ∗ (q + p + 1, p − q),


for q, p = 1, . . . , Lo


where y(l, p) is the discrete inverse Fourier transform of D̂ss (t, f ). If the phase of
100 Chapter 4. Underdetermined Blind Source Separation

the recovered signal is important, the phase can be corrected using the original
signal s(t) by computing:
 h nP o nP oi
αe = tan−1 < Le −1 ∗ Le −1 ∗
s(2k)s (k) /= s(2k)s (k)

k=0 e k=0 e
h nP o nP oi (4.25)
αo = tan−1 < Lo ∗ Lo ∗
s(2k − 1)s (k) /= s(2k − 1)s (k)

k=1 o k=1 o

then replacing se (k) and so (k) in (4.23) by se (k)ejαe and so (k)ejαo respectively.
Above, < {·} and = {·} denote the real part and imaginary part, respectively.

4.4.2 Discussion

It is essential to address the following issues regarding the above proposed algorithm
for UBSS.

4.4.2.1 Underdeterminacy

In the above description of the procedures involved in the proposed algorithm, we

do not use the information of the number of signals (n) and the number of sensors
(m). Therefore, our TF–UBSS algorithm is general in the sense that it is not only
specific to UBSS but it can also be used for determined BSS. However, in this work,
we only provide results for UBSS since it imposes a challenge in the area of BSS as
we have explained in the introductory section.

In the simulated experiments that will be shown later, we choose m = 2. Obvi-

ously, to exploit the spatial diversity offered by a sensor array, the minimum value
for m is two. This, however, is the most difficult case given a fixed number of
signals, contrast to an intuition that m = 2 is the simplest. This is due to the
fact that more sensors will provide more spatial diversity, hence more information.
On the other hand, we only use n = 3 source signals, as is the simplest case for
UBSS given that m = 2, in our simulation. This selection serves our purpose as
to illustrate the new approach, rather than to provide a very detailed performance
4.4. TF-UBSS algorithm 101

analysis on the approach.

4.4.2.2 TF orthogonality

It is important to have orthogonal sources in the TF domain in order to achieve

the blind separation of the sources. It is clear that this is too restrictive and
will almost never be satisfied exactly in practice. Nonetheless, as shown in the
simulation Section 4.5, it suffices that the source signals may need only to satisfy
a TF quasi–orthogonality condition for the signal separation to be achieved. The
term “quasi” implies that most of the energy of one source is localized in the TF
region disjoint from the TF regions of all other sources, as illustrated on Figure 4.5.

Time-frequen y quasi-orthogonality
t
time

1

2
frequen y f

Figure 4.5: TF quasi–orthogonality.

Small overlapping of the two TF supports is allowable (Ω1 ∩ Ω2 ≈ ∅); i.e. most
of the energy of one source is localized in the TF region disjoint from the TF
support of all other sources.

4.4.2.3 Choice of the TFD

We have chosen the WVD to compute the STFD matrices for our simulation. The
reason stems for, first, the fact that it is an invertible TFD up to a constant
phase [15]; and second, the WVD is the optimal TFD for LFM signals (used in
the simulations). In general, the choice of the TFD should be made according to
the nature of the application of interest and the properties desired in the TFD, as
102 Chapter 4. Underdetermined Blind Source Separation

explained in [15].

4.4.2.4 Noise thresholding

The threshold used for removing the noisy points can be chosen based on the SNR
and the possible structure of the mixed signals. The noise thresholding, however,
is used mainly for the benefit of reducing the computational complexity, and so is
not a critical factor in the proposed algorithm.

4.4.2.5 Auto–source point selection

We have proposed three selection criteria to separate the auto–source points from
the cross–source points in the TF plane. These criteria require a good choice of
the thresholding parameter as well as the signal TFD (a good choice of the TFD is
proposed in Section 4.6).

4.4.2.6 Vector clustering

A simple algorithm for vector clustering was used in the simulations in order to illus-
trate the feasibility of UBSS. More sophisticated algorithms (see [32] and references
therein) should be applied to achieve robust separation.

4.4.2.7 Number of sources

We have observed in the experiments that the number of classes, obtained from
the clustering procedure, was greater than the actual number of sources. Simple
thresholding scheme, based on energy leveling, was used to eliminate the classes with
insignificant energy compared to others. These classes may or may not be considered
as noise, depending on the nature of the sources in the particular application of
interest. At this stage, problems may arise if one or more sources have much
4.4. TF-UBSS algorithm 103

higher energy than others, in which the proposed UBSS algorithm may be used in
conjunction with a deflation technique [36].

4.4.2.8 TF synthesis

The source signatures, after a proper classification procedure, can be reconstructed

to obtain their original waveforms through the use of TF synthesis. We have applied
in our simulations a classical but seminal algorithm (without any TF masking),
proposed by Boudreaux–Bartels et al. [33]. Other synthesis algorithms can be found
in [15, 34, 35]. The successful recovery of original signal waveforms depends on the
signal type, choice of TFD, the robustness of vector clustering procedure, and the
performance of the TF synthesis algorithm itself.

On the other hand, instead of using TF synthesis, we may apply the time–
varying notched filter approach as sketched in Figure 4.6 in which selection block is
composed of all the steps from Procedure 4.4.1.1 to Procedure 4.4.1.3. Information
of notched filter design can be found in [37]. This approach is useful when the TF
synthesis algorithm corresponding to the TFD in used is not yet available.

4.4.2.9 Computational complexity

The total cost of computation is broken down into separate costs corresponding to
different procedures in the proposed algorithm. Major contributions to the total
cost Ctotal come from the computations of (i) STFD matrices (C1 ), (ii) the Singular
Value Decomposition (SVD) of the STFD matrices for separating auto–source points
from cross–source points (C2 ), (iii) clustering (C3 ), and of source synthesis (C4 ).
Note that, we use the values of SVD already obtained for the estimation of source
TFDs.
Ctotal ≈ C1 + C2 + C3 + C4 (4.26)

Denote n, m, L, Na , Nc the number of source signals, sensors, signal samples,

auto–source points, and cross–source points, respectively. CL is the cost for the
104 Chapter 4. Underdetermined Blind Source Separation

x1
s~1 + s~2 +


+ s~n s~1 Interpolation s^1
x2
Filter
xm
Time-varying
Notched filter s~2 +


+ s~n

s~2 Interpolation s^2

Filter

Time-varying
Notched filter s~1 + s~3 +


+ s~n

x(t) = As(t)

s~n Interpolation s^n

f1 (t)
SELECTION Filter
f(ta ; fa )g
f2 (t)
Tra e
Test

Classier

Time-varying
s~1 +


+ s~n 1
C1 ^
D s1 s1 (t; f )
get TFD

x( )
t
STFD
WVD
Dxx(
t; f )
C2
get TFD
^
D s2 s2 (t; f )
Notched filter
Cn ^
D s n sn (t; f )
get TFD

fn (t)
BLOCK

Figure 4.6: TF-UBSS algorithm using notch filters: Schematic diagram.

TFD computation of a signal of length L. Following are the associated costs:

m(m + 1)
C1 = CL × (4.27)
2
C2 = (Na + Nc ) × O(m3 ) (4.28)
Na (Na + 1)
C3 ≈ (4.29)
2
C4 ≈ n × O(L3 ) (4.30)

Note that the computation of CL depends on the TFD method, signal length and
the number of FFT points used. If a sophisticated clustering method, then C3 is
expected to increases. Overall, C2 and C3 are the most expensive computation due
to the high numbers of auto–source points and cross–source points present in the
TF representation; obviously, these numbers are dependent on the number of source
signals to be separated.
4.5. Experiments 105

4.5 Experiments

The algorithm in Section 4.4 is experimentally tested for the following two situa-
tions: (i ) with TF orthogonal sources, and (ii ) with TF almost orthogonal sources.
In both experiments, a uniform linear array of m = 2 sensors, having half wave-
length spacing, is used. It receives signals from n = 3 independent source signals,
each of length L = 128, in the presence of AWGN with SNR level of 20 dB. The
source signals arrive at different angles, 30◦ , 45◦ and 60◦ , respectively. The WVD
was used to compute the STFD matrices.

4.5.1 Experiment 1: TF orthogonal sources

The sources are chosen to be all monocomponent LFM signals (Figure 4.7.a–c) and
are well separated in the TF domain (Figure 4.7.d–f). The choice of LFM signals is
motivated, but not limited, from the practical application of such signals in radar
application [38] and communications [39]. The ‘noisy’ points appearing in the data
mixture (Figure 4.7.g) are first removed using energy thresholding (Figure 4.7.h;
there seemed to be no difference due to a visual effect, however, a significant num-
ber of points were indeed removed). The cross-source points are removed using
threshold 3 (Figure 4.7.i). After the vector classification procedure with 2 , three
classes containing three TF signatures representing the three original source sig-
nals are separated (Figure 4.7.j–l). Finally, estimates of the three original source
waveforms are obtained (Figure 4.7.m–o), and their corresponding TF representa-
tions (Figure 4.7.p–r), resembling the original sources (Figure 4.7.a–c) and their TF
representations (Figure 4.7.d–f), respectively.

By comparing the original with the estimates of source waveforms, it is concluded

that the proposed UBSS algorithm is successful. However, an amplitude fading at
the two ends of the recovered signals is due to the poor TFD energy concentration
in the vicinity of the TF support boundaries. In addition, though significant cross–
106 Chapter 4. Underdetermined Blind Source Separation

source points have been removed, there remain a number of them in the classified
TF signatures.

4.5.2 Experiment 2: TF quasi–orthogonal sources

As stated previously, we show here an example where the separation is achievable

in a quasi–orthogonal condition. A combination of 2 monocomponent LFM signals,
s1 (t) and s2 (t), and 1 multicomponent LFM signal, s3 (t), are used for the testing
source signals (Figure 4.8.a–c). Similar to Experiment 1, the auto–source points
are obtained, so are the separation of TF signatures, and finally the estimates of
source signal waveforms (Figure 4.8.d–r).
4.5. Experiments 107

signal s1(t) signal s2(t) signal s3(t)

1.5 1.5 1.5

1 1 1

0.5 0.5 0.5

0 0 0

−0.5 −0.5 −0.5

−1 −1 −1

−1.5 −1.5 −1.5

0 20 40 60 80 100 120 140 0 20 40 60 80 100 120 140 0 20 40 60 80 100 120 140

(a) s1 (t) (b) s2 (t) (c) s3 (t)

WVD of signal s1 WVD of signal s2 WVD of signal s3

120 120 120

100 100 100

80 80 80
time (sec)

time (sec)

time (sec)
60 60 60

40 40 40

20 20 20

0 0 0
0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5 0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5 0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5
frequency (Hz) frequency (Hz) frequency (Hz)

(d) WVD of s1 (t) (e) WVD of s2 (t) (f) WVD of s3 (t)

WVD of mixture from sensor 1 TFD of auto−terms and cross−terms only TFD of auto−terms only

120 120 120

100 100 100

80 80 80
time (sec)

time (sec)

time (sec)
60 60 60

40 40 40

20 20 20

0 0 0
0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5 0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5 0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5
frequency (Hz) frequency (Hz) frequency (Hz)

(g) WVD of x1 (t) (h) auto & cross points (i) auto–source points
TF signature of class 3 / 3 TF signature of class 1 / 3 TF signature of class 2 / 3

120 120 120

100 100 100

80 80 80
time (sec)

time (sec)

time (sec)

60 60 60

40 40 40

20 20 20

0 0 0
0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5 0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5 0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5
frequency (Hz) frequency (Hz) frequency (Hz)

(j) TF signature of s1 (t) (k) TF signature of s2 (t) (l) TF signature of s3 (t)

Synthesized signal from TF signature 3 / 3 Synthesized signal from TF signature 1 / 3 Synthesized signal from TF signature 2 / 3
1.5 1.5 1.5

1 1 1

0.5 0.5 0.5

0 0 0

−0.5 −0.5 −0.5

−1 −1 −1

−1.5 −1.5 −1.5

0 20 40 60 80 100 120 140 0 20 40 60 80 100 120 140 0 20 40 60 80 100 120 140

(m) ŝ1 (t) (n) ŝ2 (t) (o) ŝ3 (t)

WVD of synthesized signal from TF signature 3 / 3 WVD of synthesized signal from TF signature 1 / 3 WVD of synthesized signal from TF signature 2 / 3

120 120 120

100 100 100

80 80 80
time (sec)

time (sec)

time (sec)

60 60 60

40 40 40

20 20 20

0 0 0
0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5 0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5 0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5
frequency (Hz) frequency (Hz) frequency (Hz)

(p) WVD of ŝ1 (t) (q) WVD of ŝ2 (t) (r) WVD of ŝ3 (t)

Figure 4.7: Experiment 1: TF-UBSS algorithm with TF orthogonality.

Three monocomponent LFM signals s1 (t), s2 (t) and s3 (t) (a–c), being the source
signals, were tested. The recovered source signals shown in (m–o) indicated the
success of the UBSS.
108 Chapter 4. Underdetermined Blind Source Separation

signal s1(t) signal s2(t) signal s3(t)

1.5 1.5 1.5

1 1 1

0.5 0.5 0.5

0 0 0

−0.5 −0.5 −0.5

−1 −1 −1

−1.5 −1.5 −1.5

0 20 40 60 80 100 120 140 0 20 40 60 80 100 120 140 0 20 40 60 80 100 120 140

(a) s1 (t) (b) s2 (t) (c) s3 (t)

WVD of signal s1 WVD of signal s2 WVD of signal s3

120 120 120

100 100 100

80 80 80
time (sec)

time (sec)

time (sec)
60 60 60

40 40 40

20 20 20

0 0 0
0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5 0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5 0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5
frequency (Hz) frequency (Hz) frequency (Hz)

(d) WVD of s1 (t) (e) WVD of s1 (t) (f) WVD of s1 (t)

WVD of mixture from sensor 1 TFD of auto−terms and cross−terms only TFD of auto−terms only

120 120 120

100 100 100

80 80 80
time (sec)

time (sec)

time (sec)
60 60 60

40 40 40

20 20 20

0 0 0
0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5 0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5 0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5
frequency (Hz) frequency (Hz) frequency (Hz)

(g) WVD of x1 (t) (h) auto & cross points (i) auto–source points
TF signature of class 2 / 3 TF signature of class 1 / 3 TF signature of class 3 / 3

120 120 120

100 100 100

80 80 80
time (sec)

time (sec)

time (sec)

60 60 60

40 40 40

20 20 20

0 0 0
0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5 0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5 0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5
frequency (Hz) frequency (Hz) frequency (Hz)

(j) TF signature of s1 (t) (k) TF signature of s2 (t) (l) TF signature of s3 (t)

Synthesized signal from TF signature 2 / 3 Synthesized signal from TF signature 1 / 3 Synthesized signal from TF signature 3 / 3
1.5 1.5 1.5

1 1 1

0.5 0.5 0.5

0 0 0

−0.5 −0.5 −0.5

−1 −1 −1

−1.5 −1.5 −1.5

0 20 40 60 80 100 120 140 0 20 40 60 80 100 120 140 0 20 40 60 80 100 120 140

(m) ŝ1 (t) (n) ŝ2 (t) (o) ŝ3 (t)

WVD of synthesized signal from TF signature 2 / 3 WVD of synthesized signal from TF signature 1 / 3 WVD of synthesized signal from TF signature 3 / 3

120 120 120

100 100 100

80 80 80
time (sec)

time (sec)

time (sec)

60 60 60

40 40 40

20 20 20

0 0 0
0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5 0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5 0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5
frequency (Hz) frequency (Hz) frequency (Hz)

(p) WVD of ŝ1 (t) (q) WVD of ŝ2 (t) (r) WVD of ŝ3 (t)

Figure 4.8: Experiment 2: TF-UBSS algorithm with TF quasi–orthogonality.

A mixture of two monocomponent and one multicomponent LFM signals s1 (t),
s2 (t) and s3 (t) (a–c), being the source signals, were tested. s1 (t) and s2 (t) overlap
in TF domain. Source s3 (t) was not falsely separated into two monocomponent
sources.
4.6. TF-UBSS algorithm using MWVD 109

Under the TF quasi–orthogonality, the sources s1 (t) from s2 (t) were successfully
separated by the proposed BSS algorithm. In addition, the purpose of the third
source being multicomponent in this experiment is, as confirmed from the simulated
result, to distinguish the proposed BSS algorithm from any time–varying filtering
approach; the algorithm does not falsely separate s3 (t) into two monocomponent
LFM signals, whereas, a time–varying filtering would normally interpret this source
as two separate monocomponent sources.

4.6 TF-UBSS algorithm using MWVD

The simulated results in Experiment 1 and Experiment 2 show that the proposed
algorithm was successful in separating nonstationary signals in the underdetermined
case. However, as observed in both experiments, there were undesirable (cross–
source) points, in the TF signatures (Figure 4.8.k,l), present along with the desired
(auto–source) points for a particular source. Consequently, extra ridges appear in
the TFD of the recovered signals (Figure 4.8.q,r). The presence of these extra ridges
may lead to a wrong interpretation of the original signal, e.g. to have another IF
law. Thus, we need to seek for a more robust solution. In this section, we propose a
modified version of TF–UBSS algorithm which helps improve the auto–source point
selection procedure, hence, the performance of the separation.

It is essential to note that, apart from the observed problem in auto–source

selection, other problems are remained: (i) the thresholds used in these experiments
are ad–hoc and are subject for detailed analysis, and (ii) the extended treatment at
the overlapping of quasi–orthogonal sources. These issues need further development
which is out of the scope of this chapter.

4.6.1 Remarks

In improving the proposed algorithm, we first notice the following:

110 Chapter 4. Underdetermined Blind Source Separation

1. WVD is optimal for LFM signals, however, it suffers from the cross–source
problem [15]. There are other TFDs specifically designed for cross–source sup-
pression (see, for example, in Section 2.1.3). However, there remains some bias
in the IF law among different TFDs. We choose to use another distribution
called the MWVD [15], as defined below:

ρmwvd
x (t, f ) = ρwvd spec
x (t, f ) · ρx (t, f ) (4.31)

where SPEC, as defined in (2.10), is the square of STFT. This choice serves
two purposes: the WVD keeps the high resolution and the optimality for LFM
signals, and SPEC is free of cross–terms. In addition, the implementation of
the TF synthesis algorithm used in this chapter is based on WVD, thus we
still need to perform the original computation of STFD matrices using WVD.

2. Previously, the inputs of the clustering procedure are the selected set of auto–
source points and the WVD–based STFD matrices. However, as observed in
Experiment 2, there are points which are a superposition of both auto–source
and cross–source points. We propose a solution to this by applying STFT in
the clustering procedure. This is due to the fact that STFT is the square root
of SPEC, hence is free of cross–source points.

4.6.2 Algorithm

Based on the above discussion, we are now able to set up the steps of the TF–UBSS
algorithmusing MWVD for a refined auto–source point selection. The algorithm
consists of the same overall procedures as those in Figure 4.3. A diagram of the
algorithm is also shown in Figure 4.9.
4.6. TF-UBSS algorithm using MWVD 111

STFD Dxxmwvd (
( )
t; f )
Auto-term f(ta ; fa )g
MWVD Sele tion

STFD Dxxstft (
( )
t; f )

STFT

Classier

C1 ^
D s1 s1 (t; f ) ^1 (t)
s
get TFD TF-Syn

x( ) = As( )
t t
STFD Dxxwvd (
( )
t; f )
C2
get TFD
^
D s2 s2 (t; f )
TF-Syn
^2 (t)
s

WVD

Cn ^
D s n sn (t; f ) ^n (t)
s
get TFD TF-Syn

Figure 4.9: TF-UBSS algorithm using MWVD: Schematic diagram.

4.6.2.1 STFD computation and noise thresholding

We compute the STFD matrices of the observation vectors x(1), . . . , x(L) using both
WVD and STFT, denoted as Dwvd stft wvd
xx (t, f ) and Dxx (t, f ), respectively. Dxx (t, f ) is

computed as in Procedure 4.4.1.1. The off–diagonal elements of Dstft

xx (t, f ) are zeros

and its diagonal elements are computed as:


M
X
x(p)h(p − l) e−j2πpk/L , for i = j






 stft   p=−M
Dxx (l, k) ij = (4.32)





0, for i 6= j


where M = (L − 1)/2, and the rectangular window function h(l) of length Lw is

given by: 
 1 , for |l| ≤ (Lw − 1)/2

h(l) = Lw . (4.33)
0,

otherwise

The STFD matrices using MWVD is then obtained using the following expression

 stft 2
Dmwvd
xx (t, f ) = Dwvd
xx (t, f )  Dxx (t, f ) ,
  (4.34)
112 Chapter 4. Underdetermined Blind Source Separation

where  denotes the Hadamard product.

To reduce the complexity, among all the TF points in each time-slice of the
TFD, keep only those with sufficient energy, according to the point with maximum
energy along this time–slice, compared to a threshold 1 (typically, 1 = 0.05). More
precisely, along a particular time–slice ith

kDxx (ts(i) , fs )k
If: > 1
max kDxx (t(i) , f )k

f (4.35)
then: keep point (t(i)
s , fs ).

Note that by removing the low–energy point in each time–slice, rather than in the
entire TF domain as in Procedure 4.4.1.1, we are able to pick up the points in the
starting– and ending– time–slices thus improve from the previous experiments.

4.6.2.2 Auto–source TF point selection

This procedure is similar to that in Section 4.4.1.2 except that we use the MWVD
instead of the WVD. Using the MWVD results in a more robust selection of the
auto–source points (due to the reduced–interference property of this distribution).
Note that, we have tested the use of both methods for auto–source selection based
on approximation projection and TF orthogonality as proposed in Section 4.4.1.2,
however similar performance were obtained for WVD (see Figure 4.11). By using
the MWVD to first reduce the cross–source points in the TFD then apply any
of these two methods for separating the cross–source points from the auto–source
points, we obtained a much better performance (see Figure 4.10(g)).
4.6. TF-UBSS algorithm using MWVD 113

4.6.2.3 Vector clustering and source TFD estimation

For each selected auto–source point (ta , fa ), estimate the corresponding spatial di-
rection as:
diag Dstft

xx (ta , fa )
a(ta , fa ) = . (4.36)
kdiag Dstft

xx (ta , fa ) k
These vectors are then clustered into different classes using the clustering procedure
as in Section 4.4.1.3. The source TFD are estimated (up to a scalar constant) as:


trace Dwvd (ta , fa ) , if (t, f ) = (ta , fa ) ∈ Ci

xx
D̂si si (t, f ) = . (4.37)
0,

otherwise

4.6.2.4 Source signal synthesis

This procedure is the same as Procedure 4.4.1.4.

4.6.3 Experiment 3: TF-UBSS using MWVD

In this third experiment, we test the MWVD–based TF–UBSS algorithm algorithm

using, again, 2 monocomponent LFM signals (Figure 4.10.a–b) and 1 multicom-
ponent signal (Figure 4.10.c). Figure 4.10.d–f respectively show the TFD of the
mixtures using WVD, MWVD and STFT. It is seen that most cross–source points
were removed using MWVD compared to using WVD. The auto–source selection
procedure was then applied to further separate the cross–source points from auto–
source points (Figure 4.10.g–i). The direction vectors were found based on the
STFD of the STFT shown in Figure 4.10.i). These vectors were then clustered
into 3 different TF signatures where the energy was computed using WVD and
shown in Figure 4.10.j–l. Obviously, a much cleaner TF representation of the TFD
estimates were obtained using this modified algorithm.

As already mentioned previously, we provide here a simulation test (Figure 4.11)

114 Chapter 4. Underdetermined Blind Source Separation

WVD of signal s WVD of signal s WVD of signal s

1 2 3

120 120 120

100 100 100

80 80 80
time (sec)

time (sec)

time (sec)
60 60 60

40 40 40

20 20 20

0 0 0
0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5 0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5 0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5
frequency (Hz) frequency (Hz) frequency (Hz)

(a) WVD of source s1 (t) (b) WVD of source s2 (t) (c) WVD of source s3 (t)
WVD for mixture MWVD for mixture STFT for mixture

120 120 120

100 100 100

80 80 80
time (sec)

time (sec)

time (sec)
60 60 60

40 40 40

20 20 20

0 0 0
0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5 0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5 0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5
frequency (Hz) frequency (Hz) frequency (Hz)

(d) WVD of mixture (e) MWVD of mixture (f) STFT of mixture

WVD for auto−terms MWVD for auto−term STFT for auto−terms

120 120 120

100 100 100

80 80 80
time (sec)

time (sec)

time (sec)
60 60 60

40 40 40

20 20 20

0 0 0
0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5 0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5 0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5
frequency (Hz) frequency (Hz) frequency (Hz)

(g) Auto–source displayed by (h) Auto–source displayed (i) Auto–source displayed by

WVD by MWVD STFT
TF signature of class 2 / 3 TF signature of class 1 / 3 TF signature of class 3 / 3

120 120 120

100 100 100

80 80 80
time (sec)

time (sec)

time (sec)

60 60 60

40 40 40

20 20 20

0 0 0
0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5 0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5 0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5
frequency (Hz) frequency (Hz) frequency (Hz)

(j) TFD estimate of s1 (t) (k) TFD estimate of s2 (t) (l) TFD estimate of s3 (t)

Figure 4.10: Experiment 3: TF-UBSS algorithm using MWVD.

A mixture of two monocomponent and one multicomponent LFM signals s1 (t),
s2 (t) and s3 (t) being the source signals with their WVD shown in (a–c), allowing
TF quasi–orthogonality. By using the MWVD to first reduce the cross–source
points in the TFD then apply any of these two methods for separating the cross–
source points from the auto–source points, we obtained a much better perfor-
mance (Compared (g) with Figure 4.11)

showing the similar performance in comparing the two methods for the auto–source
selection procedure, namely: the approximation project and the TF orthogonality.
4.7. TF-UBSS algorithm with component extraction 115

Auto−term selection (WVD) using approximation projection Auto−term selection (WVD) using TF orthogonality

120 120

100 100

80 80
time (sec)

time (sec)
60 60

40 40

20 20

0 0
0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5 0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5
frequency (Hz) frequency (Hz)

(a) (b)

Figure 4.11: Simulated comparison of auto–source selection methods.

Two WVD–based methods were used: (a) approximation projection and (b) TF
orthogonality. Similar performance was observed.

4.7 TF-UBSS algorithm with component extraction

Alternative to using MWVD in the previous section for enhancing the auto–source
point selection procedure, we propose here another solution which is based on image
processing by using a component–extraction procedure. The underlying idea of this
solution is based on the observation that a monocomponent FM signal is represented
by a linear feature corresponding to the ‘energy concentration points’ in the TF
image. If we are able to obtain all the IF components of the sources from the mixture
TF image, then for each source we will be able to group its IF components into
source TF signatures appropriately using the clustering procedure in Section 4.4.1.3
of the proposed UBSS algorithm. As mention in Section 4.2, knowing only the IF
components would not allow us to separate the sources since sources can have IF
monocomponents. To be able to use the image processing approach presented next,
we must make another assumption (additional to those in Section 4.2) on the source
signals such that the source signals are well localized in the TF domain. Visually,
the sources should only show ridges on the TF domain.
116 Chapter 4. Underdetermined Blind Source Separation

4.7.1 Algorithm

The procedures of the TF–UBSS algorithm with image processing based component
extraction are shown in Figure 4.12. Note that, Procedure 2 as in Figure 4.3 is
replaced by Procedure 2∗ in this algorithm.

TF–UBSS algorithm

Procedure 1: STFD computation and noise thresholding

Procedure 2∗ : Image component extraction

Procedure 3: Clustering and source TFD estimation

Procedure 4: Source signal synthesis

Figure 4.12: TF-UBSS algorithm with component extraction: Procedures.

4.7.1.1 STFD computation and noise thresholding

This procedure is similar to that in Section 4.6.2.1.

In addition, we apply spatial averaging [40] that mitigates further the cross–
source points by a factor depending on their spatial signatures angle (see [40] for
more details). More precisely, we compute the spatially averaged TFD as:

m
X
Davg (t, f ) = Trace(Dxx (t, f )) = Dxl xl (t, f ). (4.38)
l=1

The image of this spatially averaged TFD will be used as the input for the image
component extraction procedure as will be described in Section 4.7.1.2.
4.7. TF-UBSS algorithm with component extraction 117

4.7.1.2 Image component extraction

A practical application of satellite image processing is to extract terrestrial roads

from satellite images [41]. We apply this so called “road network tracking” approach
to extract the TF component from the TF image. This procedure includes three
main steps: (i) preprocessing: because of the particularity of the TF image, a
preprocessing is needed before applying the component extraction (ii) line detection
(or local optimization) giving local binary detection of the potential linear structures
(segments) in the image, and (iii) road detection (or global optimization) giving a
set of labeled segments. Some mathematics regarding the last two steps can be
found in Appendix 4A, and further in [41], for the understanding of the method.

1) Preprocessing: First, the TF image is transformed to a real positive–valued

image by forcing to zero all negative values2 of the TFD and by using a gray scale
in the range [1, 256]. Also, line detectors are usually limited to a line width of 5
pixels. If the components being searched do not respect this limit (which is usually
the case for a TF image), an image subsampling by block–averaging is applied to
reduce the pixel size. Despite the blurring effect, this filter presents the advantage
of reducing the noise in the TF image. Moreover, as the TF image is unisotropic
(i.e., it contains horizontal lines as can be observed in Figure 4.13.e), this image
downsampling (see Figure 4.13.f) removes this particular feature of the TF image.

2) Line detection (Local optimization): A line detector is applied at each pixel

of the image. We use the detector proposed in [41] for radar image processing.
For a given direction, its response is based on the ratio of the means computed on
both sides of the suspected line and the mean of the line itself. Height directions
are studied and the best response is kept. The resulting image is then binarized
using a simple thresholding. If statistics on the image are available (noise distribu-
tion, additive or multiplicative noise, etc.), a statistical study of the line detector
2
Negative values correspond mainly to undesired cross–terms or noise.
118 Chapter 4. Underdetermined Blind Source Separation

performance can be made to choose the more adapted threshold (for instance the
threshold corresponding to a fixed false alarm rate in homogeneous areas).

3) Road detection (Global optimization): This step is a global step introducing

constraints on the shape of the linear features are introduced to the global opti-
mization to extract connected components and to suppress the false alarms [41]. It
works on segments extracted on the thresholded line response image by thinning
and linearization. The previously detected segments are connected depending on
proximity and alignment constraints (specially on the line curvature) to form coher-
ent components. Small isolated segments are suppressed. The algorithm depends
on the following thresholds: the maximum gap between two segments to connect
them, the allowable angular difference between the two segments, and the minimum
size of a component. The result of this step is a labeled image of components.

4.7.1.3 Clustering and source TFD estimation

This procedure is similar to that in Section 4.6.2.3. However, instead of clustering

over the set of spatial direction vectors corresponding to all the auto–source points,
we cluster only the vectors representing the spatial directions of the components
which have been obtained in the previous section. The spatial direction of a com-
ponent is estimated as the averaged value, over all the points in that component, of
the principal eigenvectors of the corresponding STFD matrices. More precisely, for
each extracted component C, one estimates the corresponding spatial direction as:

1 X
aC = a(ti , fi ) (4.39)
#IC i∈I
C

where IC denotes the set of points of component C, #IC denotes the number of
points in IC and a(ti , fi ) is the estimated principal eigenvector of the i–th compo-
nent point STFD matrix Dxx (ti , fi ).
4.7. TF-UBSS algorithm with component extraction 119

4.7.1.4 Source signal synthesis

This is carried out in the same way as that in Section 4.4.1.4.

4.7.2 Experiments

To illustrate the performance of the TF–UBSS algorithm using component extrac-

tion, we present here two simulation examples corresponding to the separation of
n = 3 multicomponent FM sources using m = 2 sensors. All monocomponents are
of constant amplitude equal to one and with SNR of 10 dB. Although no thorough
statistical analysis has been done so far, we can observe from these results very
good estimation performance. A conclusion can be drawn as such: due to the lin-
ear features of the TF image, the method using component extraction can give a
better performance in terms of the extraction of the TF component (hence source
separation) which are present in all the underlying sources under the assumption
that these components are FM–like signals.

4.7.2.1 Experiment 4

In this experiment, all sources are LFM signals; two of them are monocomponent
LFMs and the third one is a two–component LFM signal. The simulation results
are shown in Figure 4.13.

4.7.2.2 Experiment 5

In this experiment, we include a quadratic FM signal as well in order to show that

our algorithm does not only work for the limited case of LFM signals. The results
are shown in Figure 4.14.
120 Chapter 4. Underdetermined Blind Source Separation

WVD of signal s WVD of signal s WVD of signal s

1 2 3
500 500 500

450 450 450

400 400 400

350 350 350

300 300 300

time (sec)

time (sec)

time (sec)
250 250 250

200 200 200

150 150 150

100 100 100

50 50 50

0 0 0
0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5 0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5 0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5
frequency (Hz) frequency (Hz) frequency (Hz)

(a) (b) (c)

Mixture output using WVD Mixture output using MWVD 500
500 500

450
450 450

400
400 400

350
350 350

300
300 300

time (slices)
time (sec)

time (sec)

250
250 250

200 200 200

150 150 150

100 100 100

50 50 50

0 0 0
0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5 0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5 0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5
frequency (Hz) frequency (Hz) Normalised frequency (f/Fs)

(d) (e) (f)

500 500 Spatial averaged across 2 sensors
500

450 450
450

400 400
400

350 350
350

300 300
time (slices)

time (slices)

300

time (sec)
250 250
250

200 200 200

150 150 150

100 100 100

50 50 50

0 0 0
0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5 0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5 0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5
Normalised frequency (f/Fs) Normalised frequency (f/Fs) frequency (Hz)

(g) (h) (i)

TF signature of class 3 / 3 TF signature of class 2 / 3 TF signature of class 1 / 3
500 500 500

450 450 450

400 400 400

350 350 350

300 300 300

time (sec)

time (sec)

time (sec)

250 250 250

200 200 200

150 150 150

100 100 100

50 50 50

0 0 0
0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5 0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5 0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5
frequency (Hz) frequency (Hz) frequency (Hz)

(j) (k) (l)

WVD of synthesized signal from TF signature 3 / 3 WVD of synthesized signal from TF signature 2 / 3 WVD of synthesized signal from TF signature 1 / 3
500 500 500

450 450 450

400 400 400

350 350 350

300 300 300

time (sec)

time (sec)

time (sec)

250 250 250

200 200 200

150 150 150

100 100 100

50 50 50

0 0 0
0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5 0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5 0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5
frequency (Hz) frequency (Hz) frequency (Hz)

(m) (n) (o)

Figure 4.13: Experiment 4: TF-UBSS using component extraction for LFM sig-
nals.
(a–c) WVD of s1 (t), s2 (t), s3 (t); (d,e) spatial–averaged TFD of the mixture
outputs using WVD and MWVD; (f) convert STFD mixture to image; (g-h) ex-
traction of source components using image processing; (i) auto–source points of
known components; (j–l) TFD estimates of the sources; (m–o) TFD of estimated
sources after TF synthesis. SNR = 10 dB.

4.8 Numerical performance evaluation

There are several common performance criteria used for the evaluation of BSS
algorithms in practice, such as: Crosstalk (SNR, Signal–to–Interference Noise Ratio
4.8. Numerical performance evaluation 121

WVD of signal s WVD of signal s WVD of signal s

1 2 3
500 500 500

450 450 450

400 400 400

350 350 350

300 300 300

time (sec)

time (sec)

time (sec)
250 250 250

200 200 200

150 150 150

100 100 100

50 50 50

0 0 0
0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5 0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5 0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5
frequency (Hz) frequency (Hz) frequency (Hz)

(a) (b) (c)

Mixture output using WVD Mixture output using MWVD 500
500 500

450
450 450

400
400 400

350
350 350

300
300 300

time (slices)
time (sec)

time (sec)
250
250 250

200 200 200

150 150 150

100 100 100

50 50 50

0 0 0
0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5 0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5 0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5
frequency (Hz) frequency (Hz) Normalised frequency (f/Fs)

(d) (e) (f)

500 500 Spatial averaged across 2 sensors
500

450 450
450

400 400
400

350 350
350

300 300
time (slices)

time (slices)

300

time (sec)
250 250
250

200 200 200

150 150 150

100 100 100

50 50 50

0 0 0
0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5 0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5 0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5
Normalised frequency (f/Fs) Normalised frequency (f/Fs) frequency (Hz)

(g) (h) (i)

TF signature of class 1 / 3 TF signature of class 3 / 3 TF signature of class 2 / 3
500 500 500

450 450 450

400 400 400

350 350 350

300 300 300

time (sec)

time (sec)

time (sec)

250 250 250

200 200 200

150 150 150

100 100 100

50 50 50

0 0 0
0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5 0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5 0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5
frequency (Hz) frequency (Hz) frequency (Hz)

(j) (k) (l)

Figure 4.14: Experiment 5: TF-UBSS using component extraction for non-LFM

signals.
(a–c) WVD of s1 (t), s2 (t), s3 (t); (d,e) spatial–averaged TFD of the mixture
outputs using WVD and MWVD; (f) convert STFD mixture to image; (g-h)
extraction of source components using image processing; (i) auto–source points
of known components; (j–l) TFD estimates of the sources. SNR = 10 dB.

(SINR)), Distance to Diagonal Matrix, Rejection Level, Global Index, and Mean
Squared Error (MSE) (see [42] for a survey of these criteria). In our work, we apply
the MSE criterion defined as:

Nr
1 X kx̂k − xk2
εx = (4.40)
Nr k=1 kxk2
122 Chapter 4. Underdetermined Blind Source Separation

where the norm k·k is evaluated in the Frobenius sense [43] and Nr is the number of
Monte Carlo simulation runs. Nr = 100 was used in all the performance simulation.
The generic variable x in (4.40) represent the true value of the measure to be
analyzed shortly, being mixing matrix A, IF fin (t), or signal waveform si (t). The
estimate of x is denoted by x̂.

The experiment was set up as follows. A uniform linear array of m = 2 sensors,

having half wavelength spacing, was used. This sensor array received source signals
from n = 3 independent monocomponent LFM signals in the presence of AWGN.
All signals have the same length of L = 128 samples. The corresponding IFs, in pair
of starting and stopping normalized frequencies, were: [0.1, 0.05], [0.33, 0.3], and
[0.45, 0.35], respectively. The source signals arrived at the sensorarray at different
angles, 30◦ , 45◦ and 60◦ .

All the performance evaluations were done using the version of our TF–UBSS
algorithm that uses MWVD (see Section 4.6) since this version was shown to give
better results than that with WVD (see Section 4.5). Note that we have corrected
the permutation problem, inherent to BSS, in our simulation in order to run the
numerical performance analysis. The plots of the source signal waveforms and
their TFD (using WVD) are shown in Figure 4.15.a–f. In addition, Figure 4.15.g–i
and Figure 4.15.j–l represent the TFD estimates of the sources and their recovered
waveforms obtained by the algorithm.

4.8.1 On mixing matrix estimation

The first measure to be analyzed is the estimation of the mixing matrix. With the
given angles of arrival (i.e. θ1 = 30◦ , θ2 = 45◦ and θ3 = 60◦ ), we then have the
following true mixing matrix:
4.8. Numerical performance evaluation 123

1 1 1

0.8 0.8 0.8

0.6 0.6 0.6

0.4 0.4 0.4

0.2 0.2 0.2

0 0 0

−0.2 −0.2 −0.2

−0.4 −0.4 −0.4

−0.6 −0.6 −0.6

−0.8 −0.8 −0.8

−1 −1 −1
0 20 40 60 80 100 120 140 0 20 40 60 80 100 120 140 0 20 40 60 80 100 120 140

(a) (b) (c)

120 120 120

100 100 100

80 80 80
time (sec)

time (sec)

time (sec)
60 60 60

40 40 40

20 20 20

0 0 0
0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5 0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5 0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5
frequency (Hz) frequency (Hz) frequency (Hz)

(d) (e) (f)

120 120 120

100 100 100

80 80 80
time (sec)

time (sec)

time (sec)
60 60 60

40 40 40

20 20 20

0 0 0
0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5 0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5 0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5
frequency (Hz) frequency (Hz) frequency (Hz)

(g) (h) (i)

1.5 1.5 1.5

1 1 1

0.5 0.5 0.5

0 0 0

−0.5 −0.5 −0.5

−1 −1 −1

−1.5 −1.5 −1.5

0 20 40 60 80 100 120 140 0 20 40 60 80 100 120 140 0 20 40 60 80 100 120 140

(j) (k) (l)

Figure 4.15: Signals used for performance evaluations.

A = [a1 a2 · · · an ]
 
ejπ0 sin(θ1 ) ejπ0 sin(θ2 ) · · · ejπ0 sin(θn )
 
= ··· ··· ··· ··· 

ejπ(m−1) sin(θ1 ) ejπ(m−1) sin(θ2 ) · · · ejπ(m−1) sin(θn )
 
1 1 1
 
=

 (4.41)
0 + j −0.6057 + 0.7957j −0.9127 + 0.4086j
124 Chapter 4. Underdetermined Blind Source Separation

For a particular simulation run, each spatial direction ai , representing the source
si (t), was estimated as the average of all “closely spaced” spatial directions at the
auto–source points which belong to the obtained TF signature of si (t). Mathemat-
ically, in a similar manner of (4.39), this writes:

1 X
âi = âp (tp , fp ) (4.42)
#Ii p∈I
i

where Ii denotes the set of auto–source points of the clustered TF signature of

si (t), #Ii denotes the number of auto–source points in Ii , and hata(tp , fp ) is the
estimated principal eigenvector of the STFD matrix Dxx (tp , fp ) at the point (tp , fp ).

The performance of mixing matrix estimation was evaluated against different

values of SNR as shown in Figure 4.16. The result shows a very good performance.
In addition, the plot indicates that the estimation error decreases linearly as SNR
increases.
−2
10

−3
10
Mean squared error

−4
10

−5
10
10 11 12 13 14 15 16 17 18 19 20
SNR (dB)

Figure 4.16: Error performance on mixing matrix estimation.

4.8.2 On auto–source selection

To address the performance on auto–source selection, we may choose to evaluate the

performance of mixing matrix estimation as well as the number of selected points
with respect to the threshold 2 (see Equation (4.17)), while keeping the SNR
value at 20 dB. The performance results are plotted in Figure 4.17. It is observed
4.8. Numerical performance evaluation 125

that the estimation of A was not sensitive to 2 (Figure 4.17.a). Concerning the
number of selected auto–source points (Figure 4.17.b), it increases with an increase
of 2 , but approaches to a constant value of around 1500 points (over the total of
128 × 128 = 16384 points). Therefore, we may conclude that a typical value for 2
is 0.3.
−3
10 2000

Number of auto−source TF points

1500
Mean squared error

−4
10

1000

−5
10 500
0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0 0.05 0.1 0.15 0.2 0.25 0.3 0.35
Threshold Threshold

(a) (b)

Figure 4.17: Performance on auto–source selection.

(a)– estimation error on A, (b)–number of auto–source points selected (over the
total: 128 × 128 = 16384 points).

4.8.3 On IF estimation

The performance of IF estimation was, similarly, evaluated against different val-

ues of SNR. As usual, one measures the performance at a particular time in-
stance[Boashash book]. We did choose to do the same and evaluate the frequency
estimation at the middle time slide of the TF representation. Only the perfor-
mance for the second LFM signal, with starting and stopping frequency pair of
[0.33, 0.3], was shown here for the purpose of demonstration. The result, illustrated
in Figure 4.18.a, shows a good performance. Furthermore, it indicates that the IF
estimation was not affected by the AWGN (at least for SNR greater than 10 dB).
This comes from the fact that TFDs spread AWGN over the TF domain, and we
were measuring over a high range of SNR (10–20 dB).

In addition, since the underlying signals were LFMs, we can also use polynomial
126 Chapter 4. Underdetermined Blind Source Separation

1 1
10 10
fc
alpha

0 0
10 10
Mean squared error

Mean squared error

−1 −1
10 10

−2 −2
10 10

−3 −3
10 10

−4 −4
10 10
10 11 12 13 14 15 16 17 18 19 20 10 11 12 13 14 15 16 17 18 19 20
SNR (dB) SNR (dB)

(a) (b)

Figure 4.18: Error performance on IF estimation.

fitting in our estimation and, in turns, evaluate the estimated polynomial coeffi-
cients. More precisely, using the following form, recalled from (2.9), of IF of LFM
signal:
fin (t) = fc + αt, (4.43)

we measured the estimation errors on the center frequency fc and the sweeping
rate (slope) α, accordingly. The result (Figure 4.18.b) shows that the estimation
of the center frequency is very poor compared to that of the sweeping rate. The
poor estimation of center frequency is expected as such: since our underlying signal
(second signal) is almost parallel with the time axis (see Figure 4.15.e), a small
error in the sweeping rate causes a large error in the center frequency. Another
observation is that the error in the sweeping rate is higher than the error evaluated as
in Figure 4.18.a. This was also expected since the points collected at the boundary
of the TF representation are normally deviated from the true line of the IF (see
Figure 4.18.g–i), causing some bias through the use of polynomial fitting.

As a conclusion for the performance of IF estimation, the method comparing

the true and estimate IF at a time slide gives better indication of the IF estimation,
and this estimation was well performed.
4.9. Conclusions 127

4.8.4 On source waveform estimation

The performance of waveform estimation is shown on Figure 4.19. It indicates

that the estimation was poorer compared to the estimation of mixing matrix and
IF. This poorness was due to the boundary effect on the TF representation which
have caused the loss in obtaining the TF points around the two ends of the signal
(see Figure 4.15.g–i). Hence, the signal waveforms were poorly estimated (see the
two ends of estimated signals in Figure 4.15.g–i). The poor performance of source
waveform estimation gives an objective for future investigation on the clustering
procedure. This is due to the fact that some “small” clusters, thought of noisy
clusters, had been removed during the clustering procedure. A better clustering
procedure may be able to correctly assign the points in these small clusters to the
appropriate source TF signatures.
−1
10
s1
s
2
s
3
Mean squared error

−2
10
10 11 12 13 14 15 16 17 18 19 20
SNR (dB)

(a)

Figure 4.19: Error performance on source waveform estimation.

4.9 Conclusions

In this chapter we have presented a new approach for blind separation of nonsta-
tionary sources using their TFDs. The proposed TF–UBSS algorithm is based on
a vector clustering procedure that estimates the source TFDs by grouping together
the TF points corresponding to “closely spaced” spatial directions. Simulation ex-
128 Chapter 4. Underdetermined Blind Source Separation

amples illustrating the performance of the proposed algorithm blind separation of

LFM signals have been provided. The work in this chapter represents a new research
direction for solving the challenging UBSS problem. Still many problems remain
under investigation including: the improvement on the vector clustering procedure,
and extension to the convolutive mixture case. We note here that in the course of
our study, two other approaches, namely using the neural network [44] and Gap
statistics [45], have been proposed to enhance the clustering procedure from our
TF–UBSS algorithm.
Appendix 4A: Image–based Road Extraction 129

APPENDIX 4A: Image based Road Extraction

A practical application of satellite image processing is to extract terrestrial roads from

satellite images. The road detection method in [41] includes two main steps: (i) line detec-
tion (or local optimization) giving local binary detection of the potential linear structures
(segments) in the image, and (ii) road detection (or global optimization) giving a set of
labeled segments. Below, we briefly describe the mathematics of above steps from [41].

4A-1. Line detection: Local optimization

Line detection is done at the pixel level by determining whether a pixel belongs to a line
crossing it along a particular direction. Given a pixel x0 and a direction dk ∈ {d1 , . . . , dNd }
(Nd = 8, typically), three regions associated with x0 and dk are then set up as shown in
Figure 4.20 with µi being the averaged amplitude (in terms of intensity). The response
from regions i to j is defined through their contrasts cij = µi /µj as in (4.44a), or through
their cross–correlation coefficients as in (4.44b):

rij = 1 − min {cij , cji } , (4.44a)

!−1/2
ni γi2 c2ij + nj γj2
ρij = 1 + (ni + nj ) (4.44b)
ni nj (cij − 1)2

where, for region i, ni is the number of pixels and γi is the variation coefficient (ration of
standard deviation and mean). The detector is then defined by the minimum response of
the filter on both sides of the structure:

r = min {r01 , r02 } ≷ r (4.45a)

ρ = min {ρ01 , ρ02 } ≷ ρ (4.45b)

A line passing x0 along direction dk is detected when the filter response is higher than
the decision threshold r (or ρ ). In practice, the line detector defined by (4.45a) is less
accurate, whereas the one defined by (4.45b) is sensitive to the threshold. Therefore, a
130 Chapter 4. Underdetermined Blind Source Separation



  
 


Figure 4.20: Line detection for road extraction problem.

Diagram showing regions associated with pixel x0 and direction dk .

combined binary detector was proposed using an associative symmetrical sum below:

r̃ρ̃
σ(r̃, ρ̃) = ≷ 0.5, r̃, ρ̃ ∈ [0, 1] (4.46)
1 − r̃ − ρ̃ + 2r̃ρ

where r̃ and ρ̃ are the normalized, to the range of [0, 1], of r and ρ according to: r̃ :=
max {0, min {1, r + 0.5 − r }} (similarly for ρ̃). The detector in (4.46) is chosen because
it is indulgent disjunctive for r̃, ρ̃ > 0.5, and conjunctive for r̃, ρ̃ < 0.5.

Further processing on the line–detected pixels to form segments is performed by first

suppressing the isolated pixels, which are not closed (in terms of direction) to any other
neighboring pixels, and then linking pixels closed (in terms of direction and spatial dis-
tance) to each other.

4A–2. Road detection: global optimization

Given a set of locally detected segments found in the previous step, this step introduce
global constraints on the shape of the linear features to connect segments that correspond
spatially to a larger feature in the whole image, i.e. to connect parts of a true “road”,
while suppressing falsely detected segments. The connection scheme is globally optimized
using Markov random field (MRF)–based model for roads [41]. The underlying MRF
model is defined on a graph structure as follows [46].

Let G = {V, E} be a graph, where V = {s1 , . . . , sN } is the set of vertices (nodes), and
Appendix 4A: Image–based Road Extraction 131

E is the set of edges connecting them. Suppose that there exists a neighborhood system
N = {n(s1 ), . . . , n(sN )} on G where n(si ) is the set of all the nodes in V that are neighbors
of si such that (i) si ∈
/ n(si ), and (ii) sj ∈ n(si ) ⇔ si ∈ n(sj ). Let X = {x1 , . . . , xN } be a
family of random variables defined on V, then X is called a random field where xi is the
random variable associated with si . We say X is an MRF on G with respect to N iff3 :

1. p(x) > 0 for all realizations x ∈ Ω, and

2. p(xi |xj , sj 6= sj ) = p(xi |xj , sj ∈ n(si ))

A clique c is a subset of V for which every pair (si , sj ) ∈ c are neighbors. Denote C(G, N )
the collection of all the cliques of G with respect to N , the general functional form of the
pdf of the MRF can be expressed as the following Gibbs distribution:

1
p(x) = exp [−U(x)] (4.47)
z
P
where U(x) = c∈C Vc (x) is called the Gibbs’s energy function, Vc is called a potential
depending on c, and z is a normalizing constant.

Applying the above MRF model on graphs to the problem of road detection, the nodes
si are the detected segments. The set E contains “possible” connections. A possible con-
nection is verified by: (i) it links two end–points (eki , elj ; k, l ∈ {1, 2}) of two different
segments, (ii) the end–points are closed enough, and (iii) the alignment of the two seg-
n o
ments is acceptable. The neighborhood of si is n(si ) = sj : ∃(k, p), ekj = epi , j 6= i . The
cliques are all subsets of V sharing an extremity, including singletons and cycles of three
segments. Road detection consists in identifying the nodes that belong to a road, i.e. in
labeling the graph, resulting a label random field: L = {L1 , . . . , LN } (Li = 1 if si belongs
to a road, and Li = 0 otherwise). L takes its values in Ω, the set of all (2N ) possible
configurations (realizations).

The result of road detection is defined as the most probable configuration for L given
the observation D = {D1 , . . . , DN } for the segments of V. Note that Di is deduced
by averaging σ(r̃, ρ̃) of all pixels in segment si . The solution, then, corresponds to the
maximum of the conditional probability distribution of L given D, using Bayesian rule,
3
p(x) = P (X = x).
132 Chapter 4. Underdetermined Blind Source Separation

as:
p(D|L) p(L)
p(L|D) = (4.48)
p(D)

All the probability distributions p(D|L), p(L) and p(D) follow the Gibbs distribution
in (4.47). Details of their corresponding energy functions and clique potentials can be
found in [41].

The output of the road detection step is the set of roads with their associated segments.
References

[1] J. F. Cardoso, “Blind signal separation: Statistical principles,” Proceedings of the

IEEE, vol. 9, pp. 2009–2025, Oct. 1998.

[2] Series of Proceedings of the International Conferences on Independent Component

Analysis and Blind Signal Separation (ICA), 1999, 2000, 2001.

[3] P. Comon, “Independent component analysis: A new concept?,” Signal Processing,

vol. 36, pp. 287–314, 1994.

[4] X.-R. Cao and R.-W. Liu, “General approach to blind source separation,” IEEE
Transactions on Signal Processing, vol. 44, pp. 562–571, Mar. 1996.

[5] T.-W. Lee, Independent Component Analysis: Theory and Applications. Boston/
Dordrecht/ London: Kluwer Academic, 1998.

[6] V. Zarzoso and A. K. Nandi, “Blind source separation,” in Blind Estimation Using
Higher-Order Statistics (A. K. Nandi, ed.), pp. 167–252, Boston: Kluwer Academic
Publishers, 1999.

[7] A. Belouchrani, “Blind source separation: Concepts, approaches and applications,”

in ISSPA’2001 Tutorial, (Kuala–Lumpur, Malaysia), Aug. 2001.

[8] V. Varajarajan and J. L. Krolik, “Multichannel system identification methods for

sensor array calibration in uncertain multipath environments,” in Proceedings of the
11th IEEE Signal Processing Workshop on Statistical Signal Processing, pp. 297–300,
2001.

[9] K. Abed-Meraim, S. Attallah, T. J. Lim, and M. O. Damen, “A blind interference

canceller in DS–CDMA,” in IEEE Sixth International Symposium on Spread Spec-
trum and Applications, vol. 2, pp. 358–362, 2000.

[10] A. Jourjine, S. Rickard, and O. Yilmaz, “Blind separation of disjoint orthogonal sig-
nals: demixing n sources from 2 mixtures,” in International Conference on Acoustics,
Speech, and Signal Processing, ICASSP’2000, vol. 5, (Istanbul, Turkey), pp. 2985–
2988, June 2000.

[11] A. K. Nandi, ed., Blind Estimation Using Higher-Order Statistics. Boston: Kluwer
Academic Publishers, 1999.
133
134 Chapter 4. Underdetermined Blind Source Separation

[12] A. Rouxel, D. L. Guennec, and O. Macchi, “Unsupervised adaptive separation of

impulse signals applied to EEG analysis,” in International Conference on Acoustics,
Speech, and Signal Processing, ICASSP’2000, vol. 1, pp. 420–423, 2000.

[13] A. Belouchrani and M. G. Amin, “A new approach for blind source separation using
time-frequency distributions,” in Proceedings SPIE conference on Advanced algo-
rithms and Architectures for Signal Processing, (Denver, Colorado), 1996.

[14] A. Belouchrani and M. G. Amin, “Blind source separation based on time-frequency

signal representations,” IEEE Transactions on Signal Processing, vol. 46, pp. 2888–
2897, Nov. 1998.

[15] B. Boashash, ed., Time-Frequency Signal Analysis: Methods and Applications. Mel-
bourne, Australia: Longman Cheshire, 1992.

[16] A. Belouchrani, K. Abed-Meraim, M. G. Amin, and A. M. Zoubir, “Joint anti-

diagonalization for blind source separation,” in International Conference on Acous-
tics, Speech, and Signal Processing, ICASSP’2001, (Salt Lake city, Utah), May 2001.

[17] K. Abed-Meraim, A. Belouchrani, and R. Leyman, “Blind source separation using

time frequency distributions,” in Time–Frequency Signal Analysis and Processing,
Prentice-Hall, 2002. to appear.

[18] Y. Zhang and M. G. Amin, “Blind separation of sources based on their time-frequency
signatures,” in Proceedings. 2000 IEEE International Conference on Acoustics,
Speech, and Signal Processing, ICASSP ’00, vol. 5, (Istanbul, Turkey), June 2000.

[19] P. Comon, “Blind channel identification and extraction of more sources than sensors,”
in Proceedings of the SPIE, vol. 3461, (San Diego), July 1998.

[20] A. Belouchrani and J.-F. Cardoso, “Maximum likelihood source separation for dis-
crete sources,” in Proceedings EUSIPCO, 1994.

[21] A. Taleb and C. Jutten, “On underdetermined source separation,” in ICASSP’99,

vol. 3, pp. 1445–1448, 1999.

[22] T.-W. Lee, M. S. Lewicki, M. Girolami, and T. J. Sejnowski, “Blind source separation
of more sources than mixtures using overcomplete representations,” IEEE Signal
Processing Letters, vol. 6, pp. 87–90, Apr. 1999.

[23] P. Comon and O. Grellier, “Nonlinear inversion of underdetermined mixtures,” in

ICA’99, (Aussois, France), pp. 461–465, Jan. 1999.

[24] K. I. Diamantaras, “Blind separation of multiple binary sources using a single lin-
ear mixture,” in International Conference on Acoustics, Speech, and Signal Process-
ing, ICASSP’2000, vol. V, (Istanbul, Turkey), pp. 2657–2660, June 2000.

[25] A. Belouchrani, K. Abed-Meraim, J. F. Cardoso, and E. Moulines, “Blind source

separation using second order statistics,” IEEE Transactions on Signal Processing,
vol. 42, pp. 434–444, Feb. 1997.

[26] M. Zibulevsky and B. A. Pearlmutter, “Blind source separation by sparse decompo-

sition in a signal dictionary,” Neural Computations, vol. 13, no. 4, 2001.
References 135

[27] P. Bofill and M. Zibulevsky, “Underdetermined blind source separation using sparse
representations,” Signal Processing, vol. 81, pp. 2353–2362, 2001.

[28] K. Abed-Meraim and B. Brakat, “Blind source separation using the time–frequency
distribution of the mixture signal,” in Proceedings of the Second IEEE International
Symposium on Signal Processing and Information Technology (ISSPIT’02), vol. 2,
(Marrakech, Morocco), pp. 663–666, Dec. 2002.

[29] B. Boashash, ed., Time Frequency Signal Analysis and Processing: Method and Ap-
plications. Oxford: Elsevier, 2003.

[30] B. Boashash, “Time-frequency signal analysis,” in Advances in Spectrum Analysis

and Array Processing (S. Haykin, ed.), vol. I, ch. 9, pp. 418–517, Englewood Cliffs,
New Jersey: Prentice-Hall, 1991.

[31] M. Wax and T. Kailath, “Detection of signals by information theoretic crite-

ria,” IEEE Transactions on Acoustic, Speech and Signal Processing, vol. ASSP–33,
pp. 387–392, Apr. 1985.

[32] A. Gersho and R. M. Gray, Vector Quantization and Signal Compression. Kluwer
Academic Publishers, 1991.

[33] G. F. Boudreaux-Bartels and T. W. Marks, “Time-varying filtering and signal esti-

mation using Wigner distributions,” IEEE Transactions on Acoustics, Speech, and
Signal Processing, vol. 34, pp. 422–430, Mar. 1986.

[34] F. Hlawatsch and W. Krattenthaler, “Signal synthesis algorithms for bilinear time-
frequency signal representations,” in The Wigner Distribution - Theory and Applica-
tions in Signal Processing (W. Mecklenbräuker and F. Hlawatsch, eds.), pp. 135–209,
Amsterdam, Netherlands: Elsevier, 1997.

[35] A. Francos and M. Porat, “Analysis and synthesis of multicomponent signals us-
ing positive time-frequency distributions,” IEEE Transactions on Signal Processing,
vol. 47, pp. 493–504, Feb. 1999.

[36] A. Delfosse and P. Loubaton, “Adaptive separation of independent sources: a defla-

tion approach,” in International Conference on Acoustics, Speech, and Signal Pro-
cessing, ICASSP’94, vol. IV, pp. 41–44, 1994.

[37] M. G. Amin, “Interference mitigation in spread spectrum communication systems

using time-frequency distributions,” IEEE Transactions on Signal Processing, vol. 45,
pp. 90–101, Jan. 1997.

[38] A. Rihaczek, Principles of High-Resolution Radar. Peninsula Publishing, 1985.

[39] C. Gupta and A. Papandreou-Suppappola, “Wireless CDMA communications using

time–varying signals,” in Proceedings of the Sixth International Symposium on Signal
Processing and its Applications, ISSPA’01, vol. 1, pp. 242–245, 2001.

[40] M. Amin, W. Mu, and Y. Zhang, “Spatial and time–frequency signature estimation
of non–stationary sources,” in Proceedings of the 11th IEEE Workshop on Statistical
Signal Processing, (Singapore), pp. 313–316, Aug. 6–8 2001.
136 Chapter 4. Underdetermined Blind Source Separation

[41] F. Tupin, H. Maı̂tre, J.-F. Mangin, J.-M. Nicolas, and E. Pechersky, “Detection
of linear features in SAR images: Application to road network extraction,” IEEE
Transactions on Geoscience and Remote Sensing, vol. 36, no. 2, pp. 434–453, 1998.
[42] A. Mansour, M. Kawamoto, and N. Ohnishi, “A survey of performance indexes
of ICA algorithms,” in Proceedings of the 21st IASTED International Conference:
Modelling, Identification and Control (MIC2002), (Innsbruck, Austria), pp. 660–666,
Feb. 18–21 2002.
[43] G. H. Golub and C. F. V. Loan, Matrix Computations. Johns Hopkins University
Press, 3rd ed., 1996.
[44] Y. Luo and J. A. Chambers, “Active source selection using Gap statistics for un-
derdetermined blind source separation,” in Proceedings of the 7th IEEE Interna-
tional Symposium on Signal Processing and Applications (ISSPA’03), vol. 1, (Paris,
France), pp. 137–140, July 1–4 2003.
[45] Y. Luo and J. A. Chambers, “A modified underdetermined blind source separa-
tion method using competitive learning,” in Proceedings of the Third IEEE Inter-
national Symposium on Image and Signal Processing Analysis (ISPA’03), (Rome,
Italy), pp. 966–969, Sept. 18–20 2003.
[46] J. W. Modestino and J. Zhang, “A Markov random field model–based approach to
image interpretation,” IEEE Transactions on Pattern Analysis and Machine Intelli-
gence, vol. 40, pp. 56–66, June 1992.
 Chapter 5

Thesis Conclusions and Future

Research

5.1 Thesis conclusions

This research aimed at exploiting the advantages of TFSP, in dealing with nonsta-
tionary signals, into the fundamental issues of signal processing, namely the signal
estimation and signal separation. In particular, it has investigated the problems of
(i) the IF estimation of LFM signals corrupted in complex–valued zero–mean MN,
and (ii) the UBSS of LFM signals, while focusing onto the fast–growing area of
WCom.

In the first problem, we have proposed a TFD–based estimator for the IF us-
ing second–order spectra rather than higher–order spectra. The main assumption
was that the autocorrelation function of the MN is real–valued but not necessarily
137
138 Chapter 5. Thesis Conclusions and Future Research

positive (i.e. the spectrum of the MN is symmetric but does not necessary has
the highest peak at zero frequency). The estimation performance was analyzed in
terms of bias and variance, and compared between four different TFDs: WVD,
SPEC, CWD and MBD. Further to improve the estimation, we proposed to use the
MUSIC algorithm and showed its better performance. It was shown that the MBD
performed the best for low SNR (less than 10 dB).

In the second problem, we have presented a new approach for blind separation
of nonstationary sources using their TFDs. The separation algorithm is based on
a vector clustering procedure that estimates the source TFDs by grouping together
the TF points corresponding to “closely spaced” spatial directions. Simulations
illustrate the performances of the proposed method for the underdetermined blind
separation of FM signals. The method developed in this research represents a new
research direction for solving the UBSS problem.

The former problem was investigated directly in the context of WCom (Rayleigh
wireless mobile channel), while the latter, being a well–known problem in array
signal processing, is indirectly related to WCom (since smart antenna array has
recently been applied to WCom with promising perspectives). In addition, LFM
signals has been a well–known type of communication signals. Therefore, we can
conclude that TFSP is useful for WCom.

5.2 Directions for future research

• IF estimation: extensions to MN with asymmetric power spectral density (e.g.

Rice spectrum), to MN with WSSUS assumption (i.e. wideband model), and
to nonlinear FM signals.

• UBSS: auto–term selection, vector clustering procedure, a detailed perfor-

mance analysis to better assess the advantages and limits of the proposed
algorithm, and extensions to the convolutive mixture case.
Bibliography

[1] Series of Proceedings of the International Conferences on Independent Component

Analysis and Blind Signal Separation (ICA), 1999, 2000, 2001.
[2] K. Abed-Meraim, S. Attallah, T. J. Lim, and M. O. Damen. A blind interfer-
ence canceller in DS–CDMA. In IEEE Sixth International Symposium on Spread
Spectrum and Applications, volume 2, pages 358–362, 2000.
[3] K. Abed-Meraim, A. Belouchrani, and R. Leyman. Blind source separation using
time frequency distributions. In Time–Frequency Signal Analysis and Processing.
Prentice-Hall, 2002. to appear.
[4] K. Abed-Meraim and B. Brakat. Blind source separation using the time–frequency
distribution of the mixture signal. In Proceedings of the Second IEEE Interna-
tional Symposium on Signal Processing and Information Technology (ISSPIT’02),
volume 2, pages 663–666, Marrakech, Morocco, Dec. 2002.
[5] R. A. Altes. Detection, estimation, and classification with spectrograms. The
Journal of the Acoustical Society of America, 67(4):1232–1246, Apr. 1980.
[6] M. Amin, W. Mu, and Y. Zhang. Spatial and time–frequency signature estimation
of non–stationary sources. In Proceedings of the 11th IEEE Workshop on Statistical
Signal Processing, pages 313–316, Singapore, Aug. 6–8 2001.
[7] M. G. Amin. Interference mitigation in spread spectrum communication sys-
tems using time-frequency distributions. IEEE Transactions on Signal Processing,
45(1):90–101, Jan. 1997.
[8] G. Azemi, B. Senadji, and B. Boashash. Instantaneous frequency estimation of
frequency modulated signals in the presence of multiplicative and additive noise:
application to mobile communication systems. to appear in EUSIPCO2002.
[9] B. Barkat. Design, estimation, and performance of time–frequency distributions.
PhD thesis, Queensland University of Technology, Brisbane, Australia, 2000.
[10] B. Barkat. Instantaneous frequency estimation of nonlinear frequency–modulated
signals in the presence of multiplicative and additive noise. IEEE Transactions on
Signal Processing, 49(10):2214–2222, Oct. 2001.
[11] D. Bellan and S. Pignari. A probabilistic model for the response of an electrically
short two–conductor transmission line driven by a random plane wave field. IEEE
Transactions on Electromagnetic compatibility, 43(2):130–139, May 2001.
139
140 Bibliography

[12] P. Bello. Characterization of randomly time-variant linear channels. IEEE Trans-

actions on Communication Systems, COM-11:360–393, Dec. 1963.

[13] A. Belouchrani. Blind source separation: Concepts, approaches and applications.

In ISSPA’2001 Tutorial, Kuala–Lumpur, Malaysia, Aug. 2001.

[14] A. Belouchrani, K. Abed-Meraim, M. G. Amin, and A. M. Zoubir. Joint anti-

diagonalization for blind source separation. In International Conference on Acous-
tics, Speech, and Signal Processing, ICASSP’2001, Salt Lake city, Utah, May 2001.

[15] A. Belouchrani, K. Abed-Meraim, J. F. Cardoso, and E. Moulines. Blind source

separation using second order statistics. IEEE Transactions on Signal Processing,
42(2):434–444, Feb. 1997.

[16] A. Belouchrani and M. G. Amin. A new approach for blind source separation
using time-frequency distributions. In Proceedings SPIE conference on Advanced
algorithms and Architectures for Signal Processing, Denver, Colorado, 1996.

[17] A. Belouchrani and M. G. Amin. Blind source separation based on time-frequency

signal representations. IEEE Transactions on Signal Processing, 46(11):2888–2897,
Nov. 1998.

[18] A. Belouchrani and J.-F. Cardoso. Maximum likelihood source separation for dis-
crete sources. In Proceedings EUSIPCO, 1994.

[19] O. Besson and F. Castanié. On estimating the frequency of a sinusoid in autore-

gressive multiplicative noise. Signal Processing, 30(1):65–83, Jan. 1993.

[20] O. Besson, N. Ghogho, and A. Swami. Parameter estimation for random amplitude
chirp signals. IEEE Transactions on Signal Processing, 47(12):3208–3219, Dec.
1999.

[21] B. Boashash. Time-frequency signal analysis. In S. Haykin, editor, Advances

in Spectrum Analysis and Array Processing, volume I, chapter 9, pages 418–517.
Prentice-Hall, Englewood Cliffs, New Jersey, 1991.

[22] B. Boashash. Estimating and interpreting the instantaneous frequency of a signal-

Part 1: Fundamentals. Proceedings of the IEEE, 80(4):519–538, Apr. 1992.

[23] B. Boashash. Estimating and interpreting the instantaneous frequency of a signal-

Part 2: Algorithms and applications. Proceedings of the IEEE, 80(4):539–569, Apr.
1992.

[24] B. Boashash, editor. Time-Frequency Signal Analysis: Methods and Applications.

Longman Cheshire, Melbourne, Australia, 1992.

[25] B. Boashash, editor. Time Frequency Signal Analysis and Processing: Method and
Applications. Elsevier, Oxford, 2003.

[26] B. Boashash and P. O’Shea. Use of the cross Wigner-Wille distribution for es-
timation of instantaneous frequency. IEEE Transactions on Signal Processing,
41(3):1439–1445, Mar. 1993.
Bibliography 141

[27] B. Boashash and B. Ristic. Analysis of FM signals affected by Gaussian AM us-

ing reduced Wigner–Ville trispectrum. In International Conference on Acoustics,
Speech, and Signal Processing, ICASSP’93, volume IV, pages 408–411, Minneapolis,
1993.
[28] B. Boashash and B. Ristic. Polynomial time-frequency distributions and time-
varying higher order spectra: Application to the analysis of multicomponent FM
signals and to the treatment of multiplicative noise. Signal Processing, 67:1–23,
1998.
[29] B. Boashash and V. Sucic. High performance time–frequency distributions for
practical applications. In L. Debnath, editor, Wavelets and Signal Processing.
Birkhauser, Boston, New York: Springer–Verlag, 2002.
[30] P. Bofill and M. Zibulevsky. Underdetermined blind source separation using sparse
representations. Signal Processing, 81:2353–2362, 2001.
[31] G. F. Boudreaux-Bartels and T. W. Marks. Time-varying filtering and signal esti-
mation using Wigner distributions. IEEE Transactions on Acoustics, Speech, and
Signal Processing, 34:422–430, Mar. 1986.
[32] X.-R. Cao and R.-W. Liu. General approach to blind source separation. IEEE
Transactions on Signal Processing, 44(3):562–571, Mar. 1996.
[33] J. F. Cardoso. Blind signal separation: Statistical principles. Proceedings of the
IEEE, 9(1):2009–2025, Oct. 1998.
[34] H. Choi and W. Williams. Improved time–frequency representation of multicom-
ponent signals using exponential kernels. IEEE Transactions on Signal Processing,
37(6):862–871, June 1989.
[35] L. Cohen. Generalized phase-space distribution functions. Journal of Mathematical
Physics, 7(5):781–786, May 1966.
[36] L. Cohen. Time-Frequency Analysis. Prentice-Hall, Englewood Cliffs, NJ, 1995.
[37] P. Comon. Independent component analysis: A new concept? Signal Processing,
36:287–314, 1994.
[38] P. Comon. Blind channel identification and extraction of more sources than sensors.
In Proceedings of the SPIE, volume 3461, San Diego, July 1998.
[39] P. Comon and O. Grellier. Nonlinear inversion of underdetermined mixtures. In
ICA’99, pages 461–465, Aussois, France, Jan. 1999.
[40] C. E. Cook and M. Bernfeld. Radar Signals: An Introduction to Theory and Appli-
cation. Artech House, Norwood, MA, 1993.
[41] A. Delfosse and P. Loubaton. Adaptive separation of independent sources: a de-
flation approach. In International Conference on Acoustics, Speech, and Signal
Processing, ICASSP’94, volume IV, pages 41–44, 1994.
[42] K. I. Diamantaras. Blind separation of multiple binary sources using a single linear
mixture. In International Conference on Acoustics, Speech, and Signal Process-
ing, ICASSP’2000, volume V, pages 2657–2660, Istanbul, Turkey, June 2000.
142 Bibliography

[43] R. J. Doviak and D. S. Zrnic. Doppler radar and weather observations. Orlando,
FL: Academic, 2nd edition, 1993.
[44] R. Dwyer. Fourth–order spectra of Gaussian amplitude modulated sinusoids. The
Journal of the Acoustical Society of America, 90:916–926, 1991.
[45] R. Dwyer. Range and Doppler information from fourth–order spectra. IEEE Journal
of Ocean Engineering, 16:233–243, July 1991.
[46] P. Flandrin. A time-frequency formulation of optimum detection. International
Conference on Acoustics, Speech, and Signal Processing, ICASSP’98, 36(9):1377–
1384, Sept. 1988.
[47] P. Flandrin. A time-frequency formulation of optimum detection. IEEE Transac-
tions on Acoustics, Speech, and Signal Processing, 36(9):1377–1384, Sept. 1988.
[48] P. Flandrin. Time-Frequency / Time-Scale Analysis, Volume 10. Academic Press,
1998.
[49] A. Francos and M. Porat. Analysis and synthesis of multicomponent signals using
positive time-frequency distributions. IEEE Transactions on Signal Processing,
47(2):493–504, Feb. 1999.
[50] J. Francos and B. Friedlander. Bounds for estimation of multicomponent signals
with random amplitude and deterministic phase. IEEE Transactions on Signal
Processing, 43(5):1161–1172, May 1995.
[51] B. Friedlander and J. Francos. Estimation of amplitude and phase parameters of
multicomponent signals. IEEE Transactions on Signal Processing, 43(4):917–926,
Apr. 1995.
[52] N. T. Gaarder. Scattering function estimation. IEEE Transactions on Information
Theory, IT-14(5):684–693, Sept. 1968.
[53] A. Gersho and R. M. Gray. Vector Quantization and Signal Compression. Kluwer
Academic Publishers, 1991.
[54] M. Ghogho, A. K. Nandi, and A. Swami. Cramer-Rao bounds and maximum like-
lihood estimation for random amplitude phase–modulated signals. IEEE Transac-
tions on Signal Processing, 47(11):2905–2916, Nov. 1999.
[55] M. Ghogho, A. Swami, and T. S. Durrani. Frequency estimation in the presence
of Doppler spread: performance analysis. IEEE Transactions on Signal Processing,
49(4):777–789, Apr. 2001.
[56] G. H. Golub and C. F. V. Loan. Matrix Computations. Johns Hopkins University
Press, 3rd edition, 1996.
[57] C. Gupta and A. Papandreou-Suppappola. Wireless CDMA communications using
time–varying signals. In Proceedings of the Sixth International Symposium on Signal
Processing and its Applications, ISSPA’01, volume 1, pages 242–245, 2001.
[58] R. Haas and J.-C. Belfiore. A time-frequency well-localized pulse for multiple carrier
transmission. Wireless Personal Communications, 5:1–18, 1997.
Bibliography 143

[59] S. Haykin. Communication Systems. New York: Wiley, 3rd edition, 1994.

[60] F. Henderson and A. Lewis. Principle & applications of imaging radar. In Manual
of Remote Sensing, volume 2. New York: Wiley, 1998.

[61] F. Hlawatsch. Time-Frequency Analysis and Synthesis of Linear Signal Spaces:

Time-Frequency Filters, Signal Detection and Estimation, and Range-Doppler Es-
timation. Kluwer Academic Publishers, USA, 1998.

[62] F. Hlawatsch and W. Krattenthaler. Signal synthesis algorithms for bilinear time-
frequency signal representations. In W. Mecklenbräuker and F. Hlawatsch, editors,
The Wigner Distribution - Theory and Applications in Signal Processing, pages
135–209. Elsevier, Amsterdam, Netherlands, 1997.

[63] F. Hlawatsch and G. Matz. Quadratic time-frequency analysis of linear time-varying

systems. In L. Debnath, editor, Wavelet Transforms and Time-Frequency Signal
Analysis, chapter 9. Birkhäuser, Boston (MA), to appear in September 2000.

[64] Z. M. Hussain. Adaptive instantaneous frequency estimation: Techniques and al-

gorithms. PhD thesis, Queensland University of Technology, Brisbane, Australia,
2002.

[65] A. K. Jain. Fundamentals of Digital Image Processing. Prentice-Hall, Englewood

Cliffs, New York, 1989.

[66] W. Jakes, editor. Microwave Mobile Communications. IEEE Press, 1974.

[67] B. L. Johnson, D. W. Ricker, and J. R. Sacha. The use of iterative deconvolution for
scattering function identification. The Journal of the Acoustical Society of America,
91(5):2790–2798, May 1992.

[68] A. Jourjine, S. Rickard, and O. Yilmaz. Blind separation of disjoint orthogonal sig-
nals: demixing n sources from 2 mixtures. In International Conference on Acoustics,
Speech, and Signal Processing, ICASSP’2000, volume 5, pages 2985–2988, Istanbul,
Turkey, June 2000.

[69] V. Katkovnik and L. J. Stankovic. Instantaneous frequency estimation using the

Wigner distribution with varying and data driven window length. IEEE Transac-
tions on Signal Processing, 46(9):2315–2325, Sept. 1998.

[70] S. M. Kay. Fundamentals of Statistical Signal Processing: Estimation Theory. A.V.

Oppenheim, series editor, Prentice-Hall Signal Processing Series. Prentice-Hall, En-
glewood Cliffs, New Jersey, 1993.

[71] T.-W. Lee. Independent Component Analysis: Theory and Applications. Kluwer
Academic, Boston/ Dordrecht/ London, 1998.

[72] T.-W. Lee, M. S. Lewicki, M. Girolami, and T. J. Sejnowski. Blind source separation
of more sources than mixtures using overcomplete representations. IEEE Signal
Processing Letters, 6(4):87–90, Apr. 1999.

[73] W. Y. Lee. Mobile communications engineering. McGraw-Hill, 2nd. edition, 1998.

144 Bibliography

[74] Y. Luo and J. A. Chambers. Active source selection using Gap statistics for under-
determined blind source separation. In Proceedings of the 7th IEEE International
Symposium on Signal Processing and Applications (ISSPA’03), volume 1, pages
137–140, Paris, France, July 1–4 2003.
[75] Y. Luo and J. A. Chambers. A modified underdetermined blind source separation
method using competitive learning. In Proceedings of the Third IEEE International
Symposium on Image and Signal Processing Analysis (ISPA’03), pages 966–969,
Rome, Italy, Sept. 18–20 2003.
[76] A. Mansour, M. Kawamoto, and N. Ohnishi. A survey of performance indexes of
ICA algorithms. In Proceedings of the 21st IASTED International Conference: Mod-
elling, Identification and Control (MIC2002), pages 660–666, Innsbruck, Austria,
Feb. 18–21 2002.
[77] G. Matz and F. Hlawatsch. Time-frequency transfer function calculus (symbolic
calculus) of linear time-varying systems (linear operators) based on a generalized
underspread theory. Journal of Mathematical Physics, 39(8):4041–4070, Aug. 1998.
[78] T. J. McHale and Boudreaux-Bartels. An algorithm for synthesizing signals from
partial time-frequency models using the cross Wigner distribution. IEEE Transac-
tions on Signal Processing, 41(5):1986–1990, May 1993.
[79] W. Mecklenbräuker and F. Hlawatsch, editors. The Wigner Distribution – Theory
and Applications in Signal Processing. Elsevier, Amsterdam, Netherlands, 1997.
[80] L. B. Milstein. Interference rejection techniques in spread spectrum communica-
tions. Proceedings of the IEEE, pages 657–671, June 1988.
[81] J. W. Modestino and J. Zhang. A Markov random field model–based approach to
image interpretation. IEEE Transactions on Pattern Analysis and Machine Intel-
ligence, 40(1):56–66, June 1992.
[82] M. R. Morelande and A. M. Zoubir. Model selection of random amplitude polyno-
mial phase signals. IEEE Transactions on Signal Processing, 50(3):578–589, Mar.
2002.
[83] M. R. Morelande and A. M. Zoubir. On the performance of cyclic moments–based
parameter estimators of amplitude modulated polynomial phase signals. IEEE
Transactions on Signal Processing, 50(3):590–606, Mar. 2002.
[84] A. K. Nandi, editor. Blind Estimation Using Higher-Order Statistics. Boston:
Kluwer Academic Publishers, 1999.
[85] L.-T. Nguyen, B. Senadji, and B. Boashash. Scattering function and time-frequency
signal processing. In International Conference on Acoustics, Speech, and Signal
Processing, ICASSP’2001, volume VI, pages 3597–3600, Salt Lake city, Utah, USA,
June 2001.
[86] M. Pätzold. Mobile fading channels. John Wiley, 2002.
[87] M. Pätzold, U. Killat, and F. Laue. A deterministic digital simulation model for
Suzuki processes with application to shadowed Rayleigh land mobile radio channel.
IEEE Transactions on Vehicular Technology, 45(2):318–331, 1996.
Bibliography 145

[88] P. Z. Peebles, Jr. Probability, Random Variables, and Random Signal Principles.
McGraw-Hill, Singapore/New York, 4th edition, 2001.

[89] H. V. Poor and G. W. Wornell, editors. Wireless Communications: Signal Process-

ing Perspectives. Prentice-Hall, New Jersey, 1998.

[90] J. G. Proakis. Digital Communications. McGraw–Hill, 3rd. edition, 1995.

[91] T. S. Rappaport. Wireless Communications: Principles and Practice. Prentice-Hall,

New Jersey, 1996.

[92] S. Rice. Mathematical analysis of random noise. Bell System Technical Journal,
23:282–332, July 1944.

[93] D. W. Ricker and M. J. Gustafson. A low sidelobe technique for the direct measure-
ment of scattering functions. IEEE Journal of Oceanic Engineering, 21(1):14–23,
Jan. 1996.

[94] A. Rihaczek. Principles of High-Resolution Radar. Peninsula Publishing, 1985.

[95] B. Ristic. Some aspects of signal dependent and higher-order time-frequency and
time-scale analysis of non-stationary signals. PhD thesis, Signal Processing Re-
search Centre, Queensland University of Technology, Brisbane, Australia, 1995.

[96] A. Rouxel, D. L. Guennec, and O. Macchi. Unsupervised adaptive separation of

impulse signals applied to EEG analysis. In International Conference on Acoustics,
Speech, and Signal Processing, ICASSP’2000, volume 1, pages 420–423, 2000.

[97] J. S. Sadowsky and V. Kafedziski. On the correlation and scattering function of

the WSSUS channel for mobile communications. IEEE Transactions on Vehicular
Technology, 47(1):270–282, Feb. 1998.

[98] A. M. Sayeed. Canonical time-frequency processing for broadband signaling over

dispersive channels. In Proceedings of the IEEE-SP International Symposium on
Time-Frequency and Time-Scale Analysis, pages 369–372, New York, USA, 1998.
IEEE.

[99] A. M. Sayeed, A. Sendonaris, and B. Aazhang. Multiuser detection in fast-fading

multipath environments. IEEE Journal on Selected Areas in Communications,
16(9):1691–1701, Dec. 1998.

[100] B. Senadji and B. Boashash. A mobile communications application of time-varying

higher order spectra to FM signals affected by multiplicative noise. In Proceedings of
ICICS, 1997 International Conference on Information, Communications, and Sig-
nal Processing. Theme: Trends in Information Systems Engineering and Wireless
Multimedia Communications, volume 3, pages 1489–1492, New York, USA, 1997.

[101] S. Shamsunder and G. Giannakis. Estimating random amplitude polynomial phase

signals: a cyclostationary approach. IEEE Transaction on Signal Processing,
43(2):492–505, Feb. 1995.

[102] B. Sklar. Rayleigh fading channels in mobile digital communication systems, part
I: Characterization. IEEE Communications Magazine, pages 90–100, July 1997.
146 Bibliography

[103] B. Sklar. Rayleigh fading channels in mobile digital communication systems, part
II: Mitigation. IEEE Communications Magazine, pages 102–109, July 1997.
[104] L. J. Stankovic and V. Katkovnik. Algorithm for the instantaneous frequency esti-
mation using the time–frequency distributions with adaptive window length. IEEE
Signal Processing Letters, 5(9), Sept. 1998.
[105] R. Steele and L. Hanzo, editors. Mobile Radio Communications: Second and Third
Generation Cellular and WATM Systems. John Wiley & Sons, England, 2nd edition,
1999.
[106] A. Swami. Multiplicative noise models: parameter estimation using cumulants.
Signal Processing, 36:355–373, Apr. 1994.
[107] A. Swami. Cramer-Rao bounds for deterministic signals in additive and multiplica-
tive noise. Signal Processing, 53:231–244, 1996.
[108] A. Taleb and C. Jutten. On underdetermined source separation. In ICASSP’99,
volume 3, pages 1445–1448, 1999.
[109] F. Tupin, H. Maı̂tre, J.-F. Mangin, J.-M. Nicolas, and E. Pechersky. Detection
of linear features in SAR images: Application to road network extraction. IEEE
Transactions on Geoscience and Remote Sensing, 36(2):434–453, 1998.
[110] H. L. VanTrees. Detection, Estimation, and Modulation Theory. Radar-Sonar Signal
Processing and Gaussian Signals in Noise. Krieger Pub. Co., Malabar, Florida,
1992.
[111] V. Varajarajan and J. L. Krolik. Multichannel system identification methods for
sensor array calibration in uncertain multipath environments. In Proceedings of
the 11th IEEE Signal Processing Workshop on Statistical Signal Processing, pages
297–300, 2001.
[112] T. Wang, V. K. Dubey, and J. T. Ong. Generation of scattering functions for mobile
communication channel: a computer simulation approach. Wireless Information
Networks, 4(3):187–204, 1997.
[113] M. Wax and T. Kailath. Detection of signals by information theoretic criteria.
IEEE Transactions on Acoustic, Speech and Signal Processing, ASSP–33:387–392,
Apr. 1985.
[114] L. V. White and B. Boashash. On estimating the instantaneous frequency of a
Gaussian random signal by use of the Wigner–Ville distribution. IEEE Transactions
on Acoustics, Speech, and Signal Processing, 36(3):417–420, Mar. 1988.
[115] J. C. Wood and D. T. Barry. Linear signal synthesis using the Radon-Wigner
transform. IEEE Transactions on Signal Processing, 42(8):2105–2111, Aug. 1994.
[116] V. Zarzoso and A. K. Nandi. Blind source separation. In A. K. Nandi, editor,
Blind Estimation Using Higher-Order Statistics, pages 167–252. Boston: Kluwer
Academic Publishers, 1999.
[117] Y. Zhang and M. G. Amin. Blind separation of sources based on their time-frequency
signatures. In Proceedings. 2000 IEEE International Conference on Acoustics,
Speech, and Signal Processing, ICASSP ’00, volume 5, Istanbul, Turkey, June 2000.
Bibliography 147

[118] G. Zhou and G. Giannakis. Harmonics in Gaussian multiplicative and additive

noise: Cramer-Rao bounds. IEEE Transactions on Signal Processing, 43(5):1217–
1231, May 1995.
[119] G. Zhou and G. B. Giannakis. On estimating random amplitude-modulated harmon-
ics using higher order spectra. IEEE Journal of Oceanic Engineering, 19(4):529–539,
Oct. 1994.
[120] M. Zibulevsky and B. A. Pearlmutter. Blind source separation by sparse decompo-
sition in a signal dictionary. Neural Computations, 13(4), 2001.